diff --git a/DESCRIPTION b/DESCRIPTION
index 2b2c97dc..6ca927bf 100644
--- a/DESCRIPTION
+++ b/DESCRIPTION
@@ -1,7 +1,7 @@
 Package: audio.whisper
 Type: Package
 Title: Transcribe Audio Files using the "Whisper" Automatic Speech Recognition Model
-Version: 0.4.1
+Version: 0.5.0
 Maintainer: Jan Wijffels <jwijffels@bnosac.be>
 Authors@R: c(
     person('Jan', 'Wijffels', role = c('aut', 'cre', 'cph'), email = 'jwijffels@bnosac.be', comment = "R wrapper"), 
@@ -29,6 +29,5 @@ Suggests:
     audio.vadwebrtc (>= 0.2.0)
 LinkingTo: Rcpp
 SystemRequirements: GNU make
-RoxygenNote: 7.1.2
+RoxygenNote: 7.3.2
 Remotes: bnosac/audio.vadwebrtc
-
diff --git a/NEWS.md b/NEWS.md
index ba44cb48..c3b34e6b 100644
--- a/NEWS.md
+++ b/NEWS.md
@@ -1,3 +1,7 @@
+## CHANGES IN audio.whisper VERSION 0.5.0
+
+- Upgrade to whisper.cpp version v1.6.2 up to commit sha ede1718f6d45aa3f7ad4a1e169dfbc9d51570c4e
+
 ## CHANGES IN audio.whisper VERSION 0.4.1
 
 - Added function predict.whisper_transcription which allows to assign a transcription segment to either a left/right channel based on a Voice Activity Detection
diff --git a/R/RcppExports.R b/R/RcppExports.R
index 3c8e655b..595246ab 100644
--- a/R/RcppExports.R
+++ b/R/RcppExports.R
@@ -1,18 +1,6 @@
 # Generated by using Rcpp::compileAttributes() -> do not edit by hand
 # Generator token: 10BE3573-1514-4C36-9D1C-5A225CD40393
 
-whisper_load_model <- function(model, use_gpu = FALSE) {
-    .Call('_audio_whisper_whisper_load_model', PACKAGE = 'audio.whisper', model, use_gpu)
-}
-
-whisper_encode <- function(model, path, language, token_timestamps = FALSE, translate = FALSE, duration = 0L, offset = 0L, trace = 1L, n_threads = 1L, n_processors = 1L, entropy_thold = 2.40, logprob_thold = -1.00, beam_size = -1L, best_of = 5L, split_on_word = FALSE, max_context = -1L, prompt = "", print_special = FALSE, diarize = FALSE, diarize_percent = 1.1) {
-    .Call('_audio_whisper_whisper_encode', PACKAGE = 'audio.whisper', model, path, language, token_timestamps, translate, duration, offset, trace, n_threads, n_processors, entropy_thold, logprob_thold, beam_size, best_of, split_on_word, max_context, prompt, print_special, diarize, diarize_percent)
-}
-
-whisper_print_benchmark <- function(model, n_threads = 1L) {
-    invisible(.Call('_audio_whisper_whisper_print_benchmark', PACKAGE = 'audio.whisper', model, n_threads))
-}
-
 whisper_language_info <- function() {
     .Call('_audio_whisper_whisper_language_info', PACKAGE = 'audio.whisper')
 }
diff --git a/src/Makevars b/src/Makevars
index f910b057..e7f95a43 100644
--- a/src/Makevars
+++ b/src/Makevars
@@ -1,303 +1,14 @@
 PKG_LIBS = 
 CXX_STD = CXX11
 
-#CFLAGS   = -I.              -O3 -DNDEBUG -std=c11   -fPIC
-#CXXFLAGS = -I. -I./examples -O3 -DNDEBUG -std=c++11 -fPIC
-#LDFLAGS  =
+PKG_CPPFLAGS += -DSTRICT_R_HEADERS -I./dr_libs -I./whisper_cpp/ggml/include -I./whisper_cpp/src -I./whisper_cpp/include -I./whisper_cpp/examples
 
-# Enable O3 optimisations
-PKG_CFLAGS   += -O3
-PKG_CPPFLAGS += -O3
-
-ifndef UNAME_S
-UNAME_S := $(shell uname -s)
-endif
-
-ifndef UNAME_P
-UNAME_P := $(shell uname -p)
-endif
-
-ifndef UNAME_M
-UNAME_M := $(shell uname -m)
-endif
-
-ifndef NVCC_VERSION
-	ifeq ($(call,$(shell which nvcc))$(.SHELLSTATUS),0)
-		NVCC_VERSION := $(shell nvcc --version | egrep -o "V[0-9]+.[0-9]+.[0-9]+" | cut -c2-)
-	endif
-endif
-
-# Architecture specific
-# TODO: probably these flags need to be tweaked on some architectures
-#       feel free to update the Makefile for your architecture and send a pull request or issue
-ifeq ($(UNAME_M),$(filter $(UNAME_M),x86_64 i686 amd64))
-	ifeq ($(UNAME_S),Darwin)
-		CPUINFO_CMD := sysctl machdep.cpu.features machdep.cpu.leaf7_features
-	else ifeq ($(UNAME_S),Linux)
-		CPUINFO_CMD := cat /proc/cpuinfo
-	else ifneq (,$(filter MINGW32_NT% MINGW64_NT% MSYS_NT%,$(UNAME_S)))
-		CPUINFO_CMD := cat /proc/cpuinfo
-	else ifneq (,$(filter DragonFly FreeBSD,$(UNAME_S)))
-		CPUINFO_CMD := grep Features /var/run/dmesg.boot
-	else ifeq ($(UNAME_S),Haiku)
-		CPUINFO_CMD := sysinfo -cpu
-	endif
-
-	ifdef CPUINFO_CMD
-		AVX_M := $(shell $(CPUINFO_CMD) | grep -iwE 'AVX|AVX1.0')
-		ifneq (,$(AVX_M))
-			PKG_CFLAGS   += -mavx
-			PKG_CPPFLAGS += -mavx
-		endif
-
-		AVX2_M := $(shell $(CPUINFO_CMD) | grep -iw 'AVX2')
-		ifneq (,$(AVX2_M))
-			PKG_CFLAGS   += -mavx2
-			PKG_CPPFLAGS += -mavx2
-		endif
-		
-		AVX512F_M := $(shell $(CPUINFO_CMD) | grep -iw 'AVX512F')
-		ifneq (,$(AVX512F_M))
-			PKG_CFLAGS   += -mavx512f
-			PKG_CPPFLAGS += -mavx512f
-		endif
-
-		FMA_M := $(shell $(CPUINFO_CMD) | grep -iw 'FMA')
-		ifneq (,$(FMA_M))
-			PKG_CFLAGS   += -mfma
-			PKG_CPPFLAGS += -mfma
-		endif
-
-		F16C_M := $(shell $(CPUINFO_CMD) | grep -iw 'F16C')
-		ifneq (,$(F16C_M))
-			PKG_CFLAGS   += -mf16c
-			PKG_CPPFLAGS += -mf16c
-		endif
-
-		SSE3_M := $(shell $(CPUINFO_CMD) | grep -iwE 'PNI|SSE3')
-		ifneq (,$(SSE3_M))
-			PKG_CFLAGS   += -msse3
-			PKG_CPPFLAGS += -msse3
-		endif
-
-		SSSE3_M := $(shell $(CPUINFO_CMD) | grep -iw 'SSSE3')
-		ifneq (,$(SSSE3_M))
-			PKG_CFLAGS   += -mssse3
-			PKG_CPPFLAGS += -mssse3
-		endif
-	endif
-endif
-
-ifneq ($(filter ppc64%,$(UNAME_M)),)
-	POWER9_M := $(shell grep "POWER9" /proc/cpuinfo)
-	ifneq (,$(findstring POWER9,$(POWER9_M)))
-		PKG_CFLAGS += -mpower9-vector
-	endif
-	# Require c++23's std::byteswap for big-endian support.
-	ifeq ($(UNAME_M),ppc64)
-		PKG_CPPFLAGS += -std=c++23 -DGGML_BIG_ENDIAN
-	endif
-endif
-
-ifdef WHISPER_ACCELERATE
-	# Mac M1 - include Accelerate framework
-	ifeq ($(UNAME_S),Darwin)
-		PKG_CFLAGS  += -DGGML_USE_ACCELERATE
-		PKG_LIBS += -framework Accelerate
-	endif
-endif
-
-SOURCES_COREML = 
-OBJECTS_COREML = 
-ifdef WHISPER_COREML
-	PKG_CPPFLAGS += -DWHISPER_USE_COREML
-	PKG_LIBS  += -framework Foundation -framework CoreML
-
-ifdef WHISPER_COREML_ALLOW_FALLBACK
-	PKG_CPPFLAGS += -DWHISPER_COREML_ALLOW_FALLBACK
-endif
-	SOURCES_COREML = whisper_cpp/coreml/whisper-encoder.mm whisper_cpp/coreml/whisper-encoder-impl.m
-	OBJECTS_COREML = whisper_cpp/coreml/whisper-encoder.o  whisper_cpp/coreml/whisper-encoder-impl.o
-	PKG_CPPFLAGS += -fobjc-arc
-endif
-
-SOURCES_METAL = 
-OBJECTS_METAL = 
-ifdef WHISPER_METAL
-	ifeq ($(UNAME_S),Darwin)
-		WHISPER_METAL := 1
-
-		PKG_CFLAGS   += -DGGML_USE_METAL
-		PKG_CPPFLAGS += -DGGML_USE_METAL
-		PKG_LIBS  += -framework Foundation -framework Metal -framework MetalKit
-		
-		SOURCES_METAL = whisper_cpp/ggml-metal.m
-		OBJECTS_METAL = whisper_cpp/ggml-metal.o
-	endif
-endif
-
-ifdef WHISPER_OPENBLAS
-	PKG_CFLAGS  += -DGGML_USE_OPENBLAS $(shell pkg-config --libs openblas)
-	PKG_LIBS    += $(shell pkg-config --cflags openblas)
-endif
-
-ifdef WHISPER_CUBLAS
-	ifeq ($(shell expr $(NVCC_VERSION) \>= 11.6), 1)
-		CUDA_ARCH_FLAG ?= native
-	else
-		CUDA_ARCH_FLAG ?= all
-	endif
-	#CUDA_ARCH_FLAG = all
-	
-	ifndef CUDA_PATH
-		CUDA_PATH := /usr/local/cuda
-	endif
+SOURCES = whisper_cpp/ggml/src/ggml-quants.c whisper_cpp/ggml/src/ggml-backend.c whisper_cpp/ggml/src/ggml-alloc.c whisper_cpp/ggml/src/ggml-aarch64.c whisper_cpp/ggml/src/ggml.c 
+OBJECTS = whisper_cpp/ggml/src/ggml-quants.o whisper_cpp/ggml/src/ggml-backend.o whisper_cpp/ggml/src/ggml-alloc.o whisper_cpp/ggml/src/ggml-aarch64.o whisper_cpp/ggml/src/ggml.o 
 	
-	NVCC         = nvcc
-	PKG_CFLAGS   += -DGGML_USE_CUBLAS -I"$(CUDA_PATH)/include" -I"$(CUDA_PATH)/targets/$(UNAME_M)-linux/include"
-	PKG_CPPFLAGS += -DGGML_USE_CUBLAS -I"$(CUDA_PATH)/include" -I"$(CUDA_PATH)/targets/$(UNAME_M)-linux/include"
-	PKG_LIBS     += -lcuda -lcublas -lculibos -lcudart -lcublasLt -lpthread -ldl -lrt -L"$(CUDA_PATH)/lib64" -L/opt/cuda/lib64 -L"$(CUDA_PATH)/targets/$(UNAME_M)-linux/lib" -L/usr/lib/wsl/lib
-	NVCCFLAGS    = --forward-unknown-to-host-compiler -arch=$(CUDA_ARCH_FLAG)
-	OBJECTS_CUDA = whisper_cpp/ggml-cuda.o
-	
-endif
-
-## Note/TODO: removed sections on WHISPER_HIPBLAS / WHISPER_CLBLAST
-
-ifdef WHISPER_GPROF
-	PKG_CFLAGS   += -pg
-	PKG_CPPFLAGS += -pg
-endif
-
-ifneq ($(filter aarch64%,$(UNAME_M)),)
-	PKG_CFLAGS   += -mcpu=native
-	PKG_CPPFLAGS += -mcpu=native
-endif
-
-ifneq ($(filter armv6%,$(UNAME_M)),)
-	# 32-bit Raspberry Pi 1, 2, 3
-	PKG_CFLAGS += -mfpu=neon -mfp16-format=ieee -mno-unaligned-access
-endif
-
-ifneq ($(filter armv7%,$(UNAME_M)),)
-	# 32-bit ARM, for example on Armbian or possibly raspbian
-	#PKG_CFLAGS   += -mfpu=neon -mfp16-format=ieee -funsafe-math-optimizations -mno-unaligned-access
-	#PKG_CPPFLAGS += -mfpu=neon -mfp16-format=ieee -funsafe-math-optimizations -mno-unaligned-access
-
-	# 64-bit ARM on 32-bit OS, use these (TODO: auto-detect 64-bit)
-	PKG_CFLAGS   += -mfpu=neon-fp-armv8 -mfp16-format=ieee -funsafe-math-optimizations -mno-unaligned-access
-	PKG_CPPFLAGS += -mfpu=neon-fp-armv8 -mfp16-format=ieee -funsafe-math-optimizations -mno-unaligned-access
-endif
-
-ifneq ($(filter armv8%,$(UNAME_M)),)
-	# Raspberry Pi 4
-	PKG_CFLAGS   += -mfpu=neon-fp-armv8 -mfp16-format=ieee -funsafe-math-optimizations -mno-unaligned-access
-	PKG_CPPFLAGS += -mfpu=neon-fp-armv8 -mfp16-format=ieee -funsafe-math-optimizations -mno-unaligned-access
-endif
-
-
-##############################################################################################################
-##################### OVERWRITE PKG_CFLAGS / PKG_CPPFLAGS IF THE USER KNOWS WHAT TO DO #######################
-##############################################################################################################
-ifdef WHISPER_CFLAGS
-PKG_CFLAGS = $(WHISPER_CFLAGS) 
-endif
-ifdef WHISPER_CPPFLAGS
-PKG_CPPFLAGS = $(WHISPER_CPPFLAGS) 
-endif
-ifdef WHISPER_LIBS
-PKG_LIBS = $(WHISPER_LIBS) 
-endif
-
-PKG_CPPFLAGS += -DSTRICT_R_HEADERS -I./dr_libs -I./whisper_cpp 
-PKG_CFLAGS   += -D_XOPEN_SOURCE=600
-PKG_CPPFLAGS += -D_XOPEN_SOURCE=600
-
-# Somehow in OpenBSD whenever POSIX conformance is specified
-# some string functions rely on locale_t availability,
-# which was introduced in POSIX.1-2008, forcing us to go higher
-ifeq ($(UNAME_S),OpenBSD)
-	PKG_CFLAGS   += -U_XOPEN_SOURCE -D_XOPEN_SOURCE=700
-	PKG_CPPFLAGS += -U_XOPEN_SOURCE -D_XOPEN_SOURCE=700
-endif
-
-# Data types, macros and functions related to controlling CPU affinity
-# are available on Linux through GNU extensions in libc
-ifeq ($(UNAME_S),Linux)
-	PKG_CFLAGS   += -D_GNU_SOURCE
-	PKG_CPPFLAGS += -D_GNU_SOURCE
-endif
-
-# RLIMIT_MEMLOCK came in BSD, is not specified in POSIX.1,
-# and on macOS its availability depends on enabling Darwin extensions
-# similarly on DragonFly, enabling BSD extensions is necessary
-ifeq ($(UNAME_S),Darwin)
-	PKG_CFLAGS   += -D_DARWIN_C_SOURCE
-	PKG_CPPFLAGS += -D_DARWIN_C_SOURCE
-endif
-ifeq ($(UNAME_S),DragonFly)
-	PKG_CFLAGS   += -D__BSD_VISIBLE
-	PKG_CPPFLAGS += -D__BSD_VISIBLE
-endif
-
-# alloca is a non-standard interface that is not visible on BSDs when
-# POSIX conformance is specified, but not all of them provide a clean way
-# to enable it in such cases
-ifeq ($(UNAME_S),FreeBSD)
-	PKG_CFLAGS   += -D__BSD_VISIBLE
-	PKG_CPPFLAGS += -D__BSD_VISIBLE
-endif
-ifeq ($(UNAME_S),NetBSD)
-	PKG_CFLAGS   += -D_NETBSD_SOURCE
-	PKG_CPPFLAGS += -D_NETBSD_SOURCE
-endif
-ifeq ($(UNAME_S),OpenBSD)
-	PKG_CFLAGS   += -D_BSD_SOURCE
-	PKG_CPPFLAGS += -D_BSD_SOURCE
-endif
-
-# OS specific
-# TODO: support Windows
-ifeq ($(filter $(UNAME_S),Linux Darwin DragonFly FreeBSD NetBSD OpenBSD Haiku),$(UNAME_S))
-	PKG_CFLAGS   += -pthread
-	PKG_CPPFLAGS += -pthread
-endif
-
-#
-# Print build information
-#
-
-$(info I whisper.cpp build info: )
-$(info I UNAME_S:  $(UNAME_S))
-$(info I UNAME_P:  $(UNAME_P))
-$(info I UNAME_M:  $(UNAME_M))
-$(info I PKG_CFLAGS:   $(PKG_CFLAGS))
-$(info I PKG_CPPFLAGS: $(PKG_CPPFLAGS))
-$(info I PKG_LIBS:  $(PKG_LIBS))
-#$(info I CC:       $(CCV))
-#$(info I CXX:      $(CXXV))
-$(info )
-
-#PKG_CFLAGS = -mavx -mavx2 -mfma -mf16c
-#PKG_CFLAGS = -msse3
-#PKG_CLFAGS = -mcpu=native 
-#PKG_CFLAGS = -O3
-#PKG_CPPFLAGS += -mcpu=native
-
-SOURCES = whisper_cpp/ggml-quants.c whisper_cpp/ggml-backend.c whisper_cpp/ggml-alloc.c whisper_cpp/ggml.c whisper_cpp/whisper.cpp whisper_cpp/common-ggml.cpp whisper_cpp/common.cpp rcpp_whisper.cpp RcppExports.cpp
-OBJECTS = whisper_cpp/ggml-quants.o whisper_cpp/ggml-backend.o whisper_cpp/ggml-alloc.o whisper_cpp/ggml.o whisper_cpp/whisper.o   whisper_cpp/common-ggml.o   whisper_cpp/common.o   rcpp_whisper.o   RcppExports.o
-
-SOURCES += $(SOURCES_METAL)
-OBJECTS += $(OBJECTS_METAL)
-
-SOURCES += $(SOURCES_COREML)
-OBJECTS += $(OBJECTS_COREML)
-
-OBJECTS += $(OBJECTS_CUDA)
+SOURCES += whisper_cpp/src/whisper.cpp rcpp_whisper.cpp RcppExports.cpp
+OBJECTS += whisper_cpp/src/whisper.o   rcpp_whisper.o   RcppExports.o
 
 all: $(SHLIB)
 
 $(SHLIB): $(OBJECTS)
-
-whisper_cpp/ggml-cuda.o: whisper_cpp/ggml-cuda.cu whisper_cpp/ggml-cuda.h
-	$(NVCC) $(NVCCFLAGS) $(PKG_CPPFLAGS) -I"$(R_INCLUDE_DIR)" -fPIC -c $< -o $@
diff --git a/src/RcppExports.cpp b/src/RcppExports.cpp
index 5bb9bf53..db5c0ae2 100644
--- a/src/RcppExports.cpp
+++ b/src/RcppExports.cpp
@@ -5,59 +5,11 @@
 
 using namespace Rcpp;
 
-// whisper_load_model
-SEXP whisper_load_model(std::string model, bool use_gpu);
-RcppExport SEXP _audio_whisper_whisper_load_model(SEXP modelSEXP, SEXP use_gpuSEXP) {
-BEGIN_RCPP
-    Rcpp::RObject rcpp_result_gen;
-    Rcpp::RNGScope rcpp_rngScope_gen;
-    Rcpp::traits::input_parameter< std::string >::type model(modelSEXP);
-    Rcpp::traits::input_parameter< bool >::type use_gpu(use_gpuSEXP);
-    rcpp_result_gen = Rcpp::wrap(whisper_load_model(model, use_gpu));
-    return rcpp_result_gen;
-END_RCPP
-}
-// whisper_encode
-Rcpp::List whisper_encode(SEXP model, std::string path, std::string language, bool token_timestamps, bool translate, Rcpp::IntegerVector duration, Rcpp::IntegerVector offset, int trace, int n_threads, int n_processors, float entropy_thold, float logprob_thold, int beam_size, int best_of, bool split_on_word, int max_context, std::string prompt, bool print_special, bool diarize, float diarize_percent);
-RcppExport SEXP _audio_whisper_whisper_encode(SEXP modelSEXP, SEXP pathSEXP, SEXP languageSEXP, SEXP token_timestampsSEXP, SEXP translateSEXP, SEXP durationSEXP, SEXP offsetSEXP, SEXP traceSEXP, SEXP n_threadsSEXP, SEXP n_processorsSEXP, SEXP entropy_tholdSEXP, SEXP logprob_tholdSEXP, SEXP beam_sizeSEXP, SEXP best_ofSEXP, SEXP split_on_wordSEXP, SEXP max_contextSEXP, SEXP promptSEXP, SEXP print_specialSEXP, SEXP diarizeSEXP, SEXP diarize_percentSEXP) {
-BEGIN_RCPP
-    Rcpp::RObject rcpp_result_gen;
-    Rcpp::RNGScope rcpp_rngScope_gen;
-    Rcpp::traits::input_parameter< SEXP >::type model(modelSEXP);
-    Rcpp::traits::input_parameter< std::string >::type path(pathSEXP);
-    Rcpp::traits::input_parameter< std::string >::type language(languageSEXP);
-    Rcpp::traits::input_parameter< bool >::type token_timestamps(token_timestampsSEXP);
-    Rcpp::traits::input_parameter< bool >::type translate(translateSEXP);
-    Rcpp::traits::input_parameter< Rcpp::IntegerVector >::type duration(durationSEXP);
-    Rcpp::traits::input_parameter< Rcpp::IntegerVector >::type offset(offsetSEXP);
-    Rcpp::traits::input_parameter< int >::type trace(traceSEXP);
-    Rcpp::traits::input_parameter< int >::type n_threads(n_threadsSEXP);
-    Rcpp::traits::input_parameter< int >::type n_processors(n_processorsSEXP);
-    Rcpp::traits::input_parameter< float >::type entropy_thold(entropy_tholdSEXP);
-    Rcpp::traits::input_parameter< float >::type logprob_thold(logprob_tholdSEXP);
-    Rcpp::traits::input_parameter< int >::type beam_size(beam_sizeSEXP);
-    Rcpp::traits::input_parameter< int >::type best_of(best_ofSEXP);
-    Rcpp::traits::input_parameter< bool >::type split_on_word(split_on_wordSEXP);
-    Rcpp::traits::input_parameter< int >::type max_context(max_contextSEXP);
-    Rcpp::traits::input_parameter< std::string >::type prompt(promptSEXP);
-    Rcpp::traits::input_parameter< bool >::type print_special(print_specialSEXP);
-    Rcpp::traits::input_parameter< bool >::type diarize(diarizeSEXP);
-    Rcpp::traits::input_parameter< float >::type diarize_percent(diarize_percentSEXP);
-    rcpp_result_gen = Rcpp::wrap(whisper_encode(model, path, language, token_timestamps, translate, duration, offset, trace, n_threads, n_processors, entropy_thold, logprob_thold, beam_size, best_of, split_on_word, max_context, prompt, print_special, diarize, diarize_percent));
-    return rcpp_result_gen;
-END_RCPP
-}
-// whisper_print_benchmark
-void whisper_print_benchmark(SEXP model, int n_threads);
-RcppExport SEXP _audio_whisper_whisper_print_benchmark(SEXP modelSEXP, SEXP n_threadsSEXP) {
-BEGIN_RCPP
-    Rcpp::RNGScope rcpp_rngScope_gen;
-    Rcpp::traits::input_parameter< SEXP >::type model(modelSEXP);
-    Rcpp::traits::input_parameter< int >::type n_threads(n_threadsSEXP);
-    whisper_print_benchmark(model, n_threads);
-    return R_NilValue;
-END_RCPP
-}
+#ifdef RCPP_USE_GLOBAL_ROSTREAM
+Rcpp::Rostream<true>&  Rcpp::Rcout = Rcpp::Rcpp_cout_get();
+Rcpp::Rostream<false>& Rcpp::Rcerr = Rcpp::Rcpp_cerr_get();
+#endif
+
 // whisper_language_info
 Rcpp::DataFrame whisper_language_info();
 RcppExport SEXP _audio_whisper_whisper_language_info() {
@@ -70,9 +22,6 @@ END_RCPP
 }
 
 static const R_CallMethodDef CallEntries[] = {
-    {"_audio_whisper_whisper_load_model", (DL_FUNC) &_audio_whisper_whisper_load_model, 2},
-    {"_audio_whisper_whisper_encode", (DL_FUNC) &_audio_whisper_whisper_encode, 20},
-    {"_audio_whisper_whisper_print_benchmark", (DL_FUNC) &_audio_whisper_whisper_print_benchmark, 2},
     {"_audio_whisper_whisper_language_info", (DL_FUNC) &_audio_whisper_whisper_language_info, 0},
     {NULL, NULL, 0}
 };
diff --git a/src/rcpp_whisper.cpp b/src/rcpp_whisper.cpp
index d04fe630..1fb10927 100644
--- a/src/rcpp_whisper.cpp
+++ b/src/rcpp_whisper.cpp
@@ -3,491 +3,6 @@
 
 #include "whisper.h"
 
-#include <cmath>
-#include <fstream>
-#include <cstdio>
-#include <string>
-#include <thread>
-#include <vector>
-#include <cstring>
-
-#if defined(_MSC_VER)
-#pragma warning(disable: 4244 4267) // possible loss of data
-#endif
-
-// Terminal color map. 10 colors grouped in ranges [0.0, 0.1, ..., 0.9]
-// Lowest is red, middle is yellow, highest is green.
-const std::vector<std::string> k_colors = {
-    "\033[38;5;196m", "\033[38;5;202m", "\033[38;5;208m", "\033[38;5;214m", "\033[38;5;220m",
-    "\033[38;5;226m", "\033[38;5;190m", "\033[38;5;154m", "\033[38;5;118m", "\033[38;5;82m",
-};
-
-//  500 -> 00:05.000
-// 6000 -> 01:00.000
-std::string to_timestamp(int64_t t, bool comma = false) {
-    int64_t msec = t * 10;
-    int64_t hr = msec / (1000 * 60 * 60);
-    msec = msec - hr * (1000 * 60 * 60);
-    int64_t min = msec / (1000 * 60);
-    msec = msec - min * (1000 * 60);
-    int64_t sec = msec / 1000;
-    msec = msec - sec * 1000;
-
-    char buf[32];
-    snprintf(buf, sizeof(buf), "%02d:%02d:%02d%s%03d", (int) hr, (int) min, (int) sec, comma ? "," : ".", (int) msec);
-
-    return std::string(buf);
-}
-
-int timestamp_to_sample(int64_t t, int n_samples) {
-    return std::max(0, std::min((int) n_samples - 1, (int) ((t*WHISPER_SAMPLE_RATE)/100)));
-}
-
-// helper function to replace substrings
-void replace_all(std::string & s, const std::string & search, const std::string & replace) {
-    for (size_t pos = 0; ; pos += replace.length()) {
-        pos = s.find(search, pos);
-        if (pos == std::string::npos) break;
-        s.erase(pos, search.length());
-        s.insert(pos, replace);
-    }
-}
-
-// command-line parameters
-struct whisper_params {
-    int32_t n_threads    = std::min(4, (int32_t) std::thread::hardware_concurrency());
-    int32_t n_processors =  1;
-    int32_t offset_t_ms  =  0;
-    int32_t offset_n     =  0;
-    int32_t duration_ms  =  0;
-    int32_t progress_step =  5;
-    int32_t max_context  = -1;
-    int32_t max_len      =  0;
-    int32_t best_of      = whisper_full_default_params(WHISPER_SAMPLING_GREEDY).greedy.best_of;
-    int32_t beam_size    = whisper_full_default_params(WHISPER_SAMPLING_BEAM_SEARCH).beam_search.beam_size;
-
-    float word_thold    =  0.01f;
-    float entropy_thold =  2.40f;
-    float logprob_thold = -1.00f;
-
-    bool speed_up        = false;
-    bool debug_mode      = false;
-    bool translate       = false;
-    bool detect_language = false;
-    bool diarize         = false;
-    bool tinydiarize     = false;
-    bool split_on_word   = false;
-    bool no_fallback     = false;
-    bool output_txt      = false;
-    bool output_vtt      = false;
-    bool output_srt      = false;
-    bool output_wts      = false;
-    bool output_csv      = false;
-    bool output_jsn      = false;
-    bool output_jsn_full = false;
-    bool output_lrc      = false;
-    bool print_special   = false;
-    bool print_colors    = false;
-    bool print_progress  = false;
-    bool no_timestamps   = false;
-    bool log_score       = false;
-    bool use_gpu         = true;
-
-    std::string language  = "en";
-    std::string prompt;
-    std::string font_path = "/System/Library/Fonts/Supplemental/Courier New Bold.ttf";
-    std::string model     = "models/ggml-base.en.bin";
-
-    // [TDRZ] speaker turn string
-    std::string tdrz_speaker_turn = " [SPEAKER_TURN]"; // TODO: set from command line
-
-    std::string openvino_encode_device = "CPU";
-
-    std::vector<std::string> fname_inp = {};
-    std::vector<std::string> fname_out = {};
-};
-
-
-struct whisper_print_user_data {
-    const whisper_params * params;
-
-    const std::vector<std::vector<float>> * pcmf32s;
-    int progress_prev;
-};
-
-std::string estimate_diarization_speaker(std::vector<std::vector<float>> pcmf32s, int64_t t0, int64_t t1, bool id_only = false, float energy_higher_percent = 1.1) {
-    std::string speaker = "";
-    const int64_t n_samples = pcmf32s[0].size();
-
-    const int64_t is0 = timestamp_to_sample(t0, n_samples);
-    const int64_t is1 = timestamp_to_sample(t1, n_samples);
-
-    double energy0 = 0.0f;
-    double energy1 = 0.0f;
-
-    for (int64_t j = is0; j < is1; j++) {
-        energy0 += fabs(pcmf32s[0][j]);
-        energy1 += fabs(pcmf32s[1][j]);
-    }
-
-    if (energy0 > energy_higher_percent*energy1) {
-        speaker = "0";
-    } else if (energy1 > energy_higher_percent*energy0) {
-        speaker = "1";
-    } else {
-        speaker = "?";
-    }
-
-    //printf("is0 = %lld, is1 = %lld, energy0 = %f, energy1 = %f, speaker = %s\n", is0, is1, energy0, energy1, speaker.c_str());
-
-    if (!id_only) {
-        speaker.insert(0, "(speaker ");
-        speaker.append(")");
-    }
-
-    return speaker;
-}
-void whisper_print_progress_callback(struct whisper_context * /*ctx*/, struct whisper_state * /*state*/, int progress, void * user_data) {
-    int progress_step = ((whisper_print_user_data *) user_data)->params->progress_step;
-    int * progress_prev  = &(((whisper_print_user_data *) user_data)->progress_prev);
-    if (progress >= *progress_prev + progress_step) {
-        *progress_prev += progress_step;
-        Rprintf("%s: progress = %3d%%\n", __func__, progress);
-    }
-}
-
-void whisper_print_segment_callback(struct whisper_context * ctx, struct whisper_state * /*state*/, int n_new, void * user_data) {
-    const auto & params  = *((whisper_print_user_data *) user_data)->params;
-
-    const int n_segments = whisper_full_n_segments(ctx);
-
-    std::string speaker = "";
-
-    int64_t t0 = 0;
-    int64_t t1 = 0;
-
-    // print the last n_new segments
-    const int s0 = n_segments - n_new;
-
-    if (s0 == 0) {
-        if(params.print_progress){
-            Rprintf("\n");
-        }
-    }
-
-    for (int i = s0; i < n_segments; i++) {
-        if (!params.no_timestamps || params.diarize) {
-            t0 = whisper_full_get_segment_t0(ctx, i);
-            t1 = whisper_full_get_segment_t1(ctx, i);
-        }
-        const char * text = whisper_full_get_segment_text(ctx, i);
-        if(params.print_progress){
-          Rprintf("[%s --> %s]  %s%s\n", to_timestamp(t0).c_str(), to_timestamp(t1).c_str(), speaker.c_str(), text);  
-        }
-        Rcpp::checkUserInterrupt();
-    }
-}
-
-
-
-// Functionality to free the Rcpp::XPtr
-class WhisperModel {
-    public: 
-        struct whisper_context * ctx;
-        WhisperModel(std::string model, bool use_gpu = false){
-          struct whisper_context_params cparams;
-          cparams.use_gpu = use_gpu;
-          ctx = whisper_init_from_file_with_params(model.c_str(), cparams);
-        }
-        ~WhisperModel(){
-            whisper_free(ctx);
-        }
-};
-
-// [[Rcpp::export]]
-SEXP whisper_load_model(std::string model, bool use_gpu = false) {
-    // Load language model and return the pointer to be used by whisper_encode
-    //struct whisper_context * ctx = whisper_init(model.c_str());
-    //Rcpp::XPtr<whisper_context> ptr(ctx, false);
-    WhisperModel * wp = new WhisperModel(model, use_gpu);
-    Rcpp::XPtr<WhisperModel> ptr(wp, false);
-    return ptr;
-}
-    
-
-// [[Rcpp::export]]
-Rcpp::List whisper_encode(SEXP model, std::string path, std::string language, 
-                          bool token_timestamps = false, bool translate = false, Rcpp::IntegerVector duration = 0, Rcpp::IntegerVector offset = 0, int trace = 1,
-                          int n_threads = 1, int n_processors = 1,
-                          float entropy_thold = 2.40,
-                          float logprob_thold = -1.00,
-                          int beam_size = -1,
-                          int best_of = 5,
-                          bool split_on_word = false,
-                          int max_context = -1,
-                          std::string prompt = "",
-                          bool print_special = false,
-                          bool diarize = false,
-                          float diarize_percent = 1.1) {
-    float audio_duration=0;
-  
-    whisper_params params;
-    params.language = language;
-    params.translate = translate;
-    params.print_special = print_special;
-    params.duration_ms = duration[0];
-    params.offset_t_ms = offset[0];
-    params.fname_inp.push_back(path);
-    params.n_threads = n_threads;
-    params.n_processors = n_processors;
-    
-    params.entropy_thold = entropy_thold;
-    params.logprob_thold = logprob_thold;
-    params.beam_size = beam_size;
-    params.best_of = best_of;
-    params.split_on_word = split_on_word;
-    params.max_context = max_context;
-    params.prompt = prompt;
-    params.diarize = diarize;
-    if (params.fname_inp.empty()) {
-        Rcpp::stop("error: no input files specified");
-    }
-
-    if (params.language != "auto" && whisper_lang_id(params.language.c_str()) == -1) {
-        Rcpp::stop("Unknown language");
-    }
-    
-    // whisper init
-    Rcpp::XPtr<WhisperModel> whispermodel(model);
-    struct whisper_context * ctx = whispermodel->ctx;
-    //Rcpp::XPtr<whisper_context> ctx(model);
-    //struct whisper_context * ctx = whisper_init(params.model.c_str());
-
-    const auto fname_inp = params.fname_inp[0];
-    std::vector<float> pcmf32;               // mono-channel F32 PCM
-    std::vector<std::vector<float>> pcmf32s; // stereo-channel F32 PCM
-    
-    if (!::read_wav(fname_inp, pcmf32, pcmf32s, params.diarize)) {
-      Rprintf("error: failed to read WAV file '%s'\n", fname_inp.c_str());
-      Rcpp::stop("The input audio needs to be a 16-bit .wav file.");
-    }
-    
-    if(trace > 0){
-      Rprintf("system_info: n_threads = %d / %d | %s\n", params.n_threads*params.n_processors, std::thread::hardware_concurrency(), whisper_print_system_info());  
-    }
-    
-    {
-      if (!whisper_is_multilingual(ctx)) {
-        if (params.language != "en" || params.translate) {
-          params.language = "en";
-          params.translate = false;
-          Rcpp::warning("WARNING: model is not multilingual, ignoring language and translation options");
-        }
-      }
-      if(trace > 0){
-        Rcpp::Rcout << "Processing " << fname_inp << " (" << int(pcmf32.size()) << " samples, " << float(pcmf32.size())/WHISPER_SAMPLE_RATE << " sec)" << ", lang = " << params.language << ", translate = " << params.translate << ", timestamps = " << token_timestamps << ", beam_size = " << params.beam_size << ", best_of = " << params.best_of << "\n";
-      }
-    }
-    audio_duration = float(pcmf32.size())/WHISPER_SAMPLE_RATE;
-    
-    // Structures to get the data back in R
-    std::vector<int> segment_nr;
-    std::vector<int> segment_offset;
-    Rcpp::StringVector transcriptions(0);
-    Rcpp::StringVector transcriptions_from(0);
-    Rcpp::StringVector transcriptions_to(0);
-    Rcpp::StringVector transcriptions_speaker(0);
-    std::vector<int> token_segment_nr;
-    std::vector<int> token_segment_id;
-    std::vector<std::string> token_segment_text;
-    std::vector<float> token_segment_probability;
-    std::vector<std::string> token_segment_from;
-    std::vector<std::string> token_segment_to;
-    //Rcpp::StringVector token_speaker(0);
-    int n_segments;
-    
-    for (int f = 0; f < (int) offset.size(); ++f) {
-        // run the inference
-        {
-            whisper_full_params wparams = whisper_full_default_params(WHISPER_SAMPLING_GREEDY);
-            
-            wparams.strategy = params.beam_size > 1 ? WHISPER_SAMPLING_BEAM_SEARCH : WHISPER_SAMPLING_GREEDY;
-
-            wparams.print_realtime   = false;
-            wparams.print_progress   = false;
-            if(trace > 0){
-              wparams.print_progress = true;
-              wparams.print_realtime = true;
-            }
-            wparams.print_timestamps = !params.no_timestamps;
-            wparams.print_special    = params.print_special;
-            wparams.translate        = params.translate;
-            wparams.language         = params.language.c_str();
-            wparams.detect_language  = params.detect_language;
-            wparams.n_threads        = params.n_threads;
-            wparams.n_max_text_ctx   = params.max_context >= 0 ? params.max_context : wparams.n_max_text_ctx;
-            wparams.offset_ms        = (int) offset[f];
-            wparams.duration_ms      = (int) duration[f];
-
-            wparams.token_timestamps = token_timestamps;
-            wparams.thold_pt         = params.word_thold;
-            wparams.max_len          = params.output_wts && params.max_len == 0 ? 60 : params.max_len;
-            wparams.split_on_word    = params.split_on_word;
-
-            wparams.speed_up         = params.speed_up;
-            wparams.debug_mode       = params.debug_mode;
-
-            wparams.tdrz_enable      = params.tinydiarize; // [TDRZ]
-
-            wparams.initial_prompt   = params.prompt.c_str();
-            
-            
-
-            wparams.greedy.best_of        = params.best_of;
-            wparams.beam_search.beam_size = params.beam_size;
-
-            wparams.temperature_inc  = params.no_fallback ? 0.0f : wparams.temperature_inc;
-            wparams.entropy_thold    = params.entropy_thold;
-            wparams.logprob_thold    = params.logprob_thold;
-            
-            whisper_print_user_data user_data = { &params, &pcmf32s, 0 };
-            
-            // this callback is called on each new segment
-            if (!wparams.print_realtime) {
-                wparams.new_segment_callback           = whisper_print_segment_callback;
-                wparams.new_segment_callback_user_data = &user_data;
-            }
-            if(trace > 0 && offset.size() > 1){
-              Rcpp::Rcout << "Processing audio offset section " << f+1 << " (" << wparams.offset_ms << " ms - " << wparams.offset_ms+wparams.duration_ms << " ms)\n";
-            }
-            
-            if (whisper_full_parallel(ctx, wparams, pcmf32.data(), pcmf32.size(), params.n_processors) != 0) {
-                Rcpp::stop("failed to process audio");
-            }
-        }
-        n_segments = whisper_full_n_segments(ctx);
-        for (int i = 0; i < n_segments; ++i) {
-          segment_nr.push_back(segment_nr.size() + 1);
-          segment_offset.push_back(offset[f]);
-          const char * text = whisper_full_get_segment_text(ctx, i);
-          transcriptions.push_back(Rcpp::String(text));
-          int64_t t0 = whisper_full_get_segment_t0(ctx, i);
-          int64_t t1 = whisper_full_get_segment_t1(ctx, i);
-          transcriptions_from.push_back(Rcpp::String(to_timestamp(t0).c_str()));
-          transcriptions_to.push_back(Rcpp::String(to_timestamp(t1).c_str()));
-          Rcpp::String channel_speaker;
-          if (params.diarize && pcmf32s.size() == 2) {
-            channel_speaker = Rcpp::String(estimate_diarization_speaker(pcmf32s, t0, t1, true, diarize_percent));
-          }else{
-            channel_speaker = NA_STRING;
-          }    
-          transcriptions_speaker.push_back(channel_speaker);
-          
-          for (int j = 0; j < whisper_full_n_tokens(ctx, i); ++j) {
-            if (params.print_special == false) {
-              const whisper_token id = whisper_full_get_token_id(ctx, i, j);
-              if (id >= whisper_token_eot(ctx)) {
-                continue;
-              }
-            }
-            const char * text = whisper_full_get_token_text(ctx, i, j);
-            const float  p    = whisper_full_get_token_p   (ctx, i, j);
-            const int tokenid = whisper_full_get_token_id  (ctx, i, j);
-            token_segment_nr.push_back(i + 1);
-            token_segment_id.push_back(tokenid);
-            std::string str(text);
-            token_segment_text.push_back(str);
-            token_segment_probability.push_back(p);
-            if(token_timestamps){
-              whisper_token_data token = whisper_full_get_token_data(ctx, i, j);
-              t0 = token.t0;
-              t1 = token.t1;
-              token_segment_from.push_back(Rcpp::String(to_timestamp(t0).c_str()));
-              token_segment_to.push_back(to_timestamp(token.t1));
-            } 
-            //token_speaker.push_back(channel_speaker);
-          }
-        }
-    }
-    Rcpp::DataFrame tokens;
-    if(token_timestamps){
-        tokens = Rcpp::DataFrame::create(
-            Rcpp::Named("segment") = token_segment_nr, 
-            Rcpp::Named("token_id") = token_segment_id, 
-            Rcpp::Named("token") = token_segment_text, 
-            Rcpp::Named("token_prob") = token_segment_probability,
-            Rcpp::Named("token_from") = token_segment_from,
-            Rcpp::Named("token_to") = token_segment_to,
-            //Rcpp::Named("token_speaker") = token_speaker,
-            Rcpp::Named("stringsAsFactors") = false);
-    }else{
-        tokens = Rcpp::DataFrame::create(
-            Rcpp::Named("segment") = token_segment_nr, 
-            Rcpp::Named("token_id") = token_segment_id, 
-            Rcpp::Named("token") = token_segment_text, 
-            Rcpp::Named("token_prob") = token_segment_probability,
-            //Rcpp::Named("token_speaker") = token_speaker,
-            Rcpp::Named("stringsAsFactors") = false);
-    }
-    
-    //whisper_free(ctx);
-    Rcpp::List output = Rcpp::List::create(Rcpp::Named("n_segments") = segment_nr.size(),
-                                           Rcpp::Named("data") = Rcpp::DataFrame::create(
-                                               Rcpp::Named("segment") = segment_nr, 
-                                               Rcpp::Named("segment_offset") = segment_offset, 
-                                               Rcpp::Named("from") = transcriptions_from,
-                                               Rcpp::Named("to") = transcriptions_to,
-                                               Rcpp::Named("text") = transcriptions, 
-                                               Rcpp::Named("speaker") = transcriptions_speaker,
-                                               Rcpp::Named("stringsAsFactors") = false),
-                                           Rcpp::Named("tokens") = tokens,
-                                           Rcpp::Named("params") = Rcpp::List::create(
-                                               Rcpp::Named("audio") = path,
-                                               Rcpp::Named("audio_duration_seconds") = audio_duration,
-                                               Rcpp::Named("language") = params.language, 
-                                               Rcpp::Named("offset") = offset,
-                                               Rcpp::Named("duration") = duration,
-                                               Rcpp::Named("translate") = params.translate,
-                                               Rcpp::Named("token_timestamps") = token_timestamps,
-                                               Rcpp::Named("word_threshold") = params.word_thold,
-                                               Rcpp::Named("entropy_thold") = params.entropy_thold,
-                                               Rcpp::Named("logprob_thold") = params.logprob_thold,
-                                               Rcpp::Named("beam_size") = params.beam_size,
-                                               Rcpp::Named("best_of") = params.best_of,
-                                               Rcpp::Named("split_on_word") = params.split_on_word,
-                                               Rcpp::Named("diarize") = params.diarize,
-                                               Rcpp::Named("system_info") = Rcpp::List::create(
-                                                 Rcpp::Named("n_threads") = params.n_threads,
-                                                 Rcpp::Named("n_processors") = params.n_processors,
-                                                 Rcpp::Named("available_concurrency") = std::thread::hardware_concurrency(),
-                                                 Rcpp::Named("optimisations") = whisper_print_system_info())));
-    return output;
-}
-
-
-
-
-// [[Rcpp::export]]
-void whisper_print_benchmark(SEXP model, int n_threads = 1) {
-  whisper_params params;
-  params.n_threads = n_threads;
-  // whisper init
-  Rcpp::XPtr<WhisperModel> whispermodel(model);
-  struct whisper_context * ctx = whispermodel->ctx;
-  Rprintf("\n");
-  Rprintf("system_info: n_threads = %d / %d | %s\n", params.n_threads, std::thread::hardware_concurrency(), whisper_print_system_info());
-  const int n_mels = whisper_model_n_mels(ctx);
-  if (int ret = whisper_set_mel(ctx, nullptr, 0, n_mels)) {
-    Rprintf("error: failed to set mel: %d\n", ret);
-  }
-  if (int ret = whisper_encode(ctx, 0, params.n_threads) != 0) {
-    Rprintf("error: failed to encode model: %d\n", ret);
-  }
-  whisper_print_timings(ctx);
-}
-
-
-
 // [[Rcpp::export]]
 Rcpp::DataFrame whisper_language_info() {
   auto max_id = whisper_lang_max_id();
@@ -497,7 +12,7 @@ Rcpp::DataFrame whisper_language_info() {
   for (int i = 0; i <= max_id; ++i) {
     id.push_back(i);
     language.push_back(whisper_lang_str(i));
-    label.push_back(whisper_lang_str_full(i));
+    //label.push_back(whisper_lang_str_full(i));
   }
   return Rcpp::DataFrame::create(
     Rcpp::Named("id") = id, 
diff --git a/src/whisper_cpp/CMakeLists.txt b/src/whisper_cpp/CMakeLists.txt
new file mode 100644
index 00000000..cb17def7
--- /dev/null
+++ b/src/whisper_cpp/CMakeLists.txt
@@ -0,0 +1,185 @@
+cmake_minimum_required(VERSION 3.5) # for add_link_options and implicit target directories.
+project("whisper.cpp" C CXX)
+project("whisper.cpp" VERSION 1.6.2)
+include(CheckIncludeFileCXX)
+
+set(SOVERSION 1)
+
+#set(CMAKE_WARN_DEPRECATED YES)
+set(CMAKE_WARN_UNUSED_CLI YES)
+
+set(CMAKE_EXPORT_COMPILE_COMMANDS ON)
+
+if (NOT XCODE AND NOT MSVC AND NOT CMAKE_BUILD_TYPE)
+    set(CMAKE_BUILD_TYPE Release CACHE STRING "Build type" FORCE)
+    set_property(CACHE CMAKE_BUILD_TYPE PROPERTY STRINGS "Debug" "Release" "MinSizeRel" "RelWithDebInfo")
+endif()
+
+# Add path to modules
+list(APPEND CMAKE_MODULE_PATH "${CMAKE_CURRENT_SOURCE_DIR}/cmake/")
+
+set(CMAKE_RUNTIME_OUTPUT_DIRECTORY ${CMAKE_BINARY_DIR}/bin)
+
+if (CMAKE_SOURCE_DIR STREQUAL CMAKE_CURRENT_SOURCE_DIR)
+    set(WHISPER_STANDALONE ON)
+
+    include(git-vars)
+
+    # configure project version
+    configure_file(${CMAKE_SOURCE_DIR}/bindings/javascript/package-tmpl.json ${CMAKE_SOURCE_DIR}/bindings/javascript/package.json @ONLY)
+else()
+    set(WHISPER_STANDALONE OFF)
+endif()
+
+if (EMSCRIPTEN)
+    set(BUILD_SHARED_LIBS_DEFAULT OFF)
+
+    option(WHISPER_WASM_SINGLE_FILE "whisper: embed WASM inside the generated whisper.js" ON)
+
+    # TODO: without these, we get the following error:
+    #       wasm-ld: error: --shared-memory is disallowed by whisper.cpp.o because it was not compiled with 'atomics' or 'bulk-memory' features.
+    set(CMAKE_C_FLAGS   "${CMAKE_C_FLAGS}   -pthread -s TOTAL_STACK=5242880")
+    set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -pthread -s TOTAL_STACK=5242880")
+else()
+    if (MINGW)
+        set(BUILD_SHARED_LIBS_DEFAULT OFF)
+    else()
+        set(BUILD_SHARED_LIBS_DEFAULT ON)
+    endif()
+endif()
+
+option(BUILD_SHARED_LIBS "build shared libraries" ${BUILD_SHARED_LIBS_DEFAULT})
+
+#
+# option list
+#
+
+# general
+option(WHISPER_CCACHE "whisper: use ccache if available" ON)
+
+# debug
+option(WHISPER_ALL_WARNINGS           "whisper: enable all compiler warnings"                   ON)
+option(WHISPER_ALL_WARNINGS_3RD_PARTY "whisper: enable all compiler warnings in 3rd party libs" OFF)
+
+# build
+option(WHISPER_FATAL_WARNINGS "whisper: enable -Werror flag" OFF)
+
+# sanitizers
+option(WHISPER_SANITIZE_THREAD    "whisper: enable thread sanitizer"    OFF)
+option(WHISPER_SANITIZE_ADDRESS   "whisper: enable address sanitizer"   OFF)
+option(WHISPER_SANITIZE_UNDEFINED "whisper: enable undefined sanitizer" OFF)
+
+# extra artifacts
+option(WHISPER_BUILD_TESTS    "whisper: build tests"          ${WHISPER_STANDALONE})
+option(WHISPER_BUILD_EXAMPLES "whisper: build examples"       ${WHISPER_STANDALONE})
+option(WHISPER_BUILD_SERVER   "whisper: build server example" ${WHISPER_STANDALONE})
+
+# 3rd party libs
+option(WHISPER_CURL "whisper: use libcurl to download model from an URL" OFF)
+option(WHISPER_SDL2 "whisper: support for libSDL2" OFF)
+
+if (CMAKE_SYSTEM_NAME MATCHES "Linux")
+    option(WHISPER_FFMPEG "whisper: support building and linking with ffmpeg libs (avcodec, swresample, ...)" OFF)
+endif()
+
+option(WHISPER_COREML                "whisper: enable Core ML framework"  OFF)
+option(WHISPER_COREML_ALLOW_FALLBACK "whisper: allow non-CoreML fallback" OFF)
+option(WHISPER_OPENVINO              "whisper: support for OpenVINO"      OFF)
+
+# Required for relocatable CMake package
+include(${CMAKE_CURRENT_SOURCE_DIR}/cmake/build-info.cmake)
+
+# override ggml options
+set(GGML_CCACHE             ${WHISPER_CCACHE})
+set(GGML_SANITIZE_THREAD    ${WHISPER_SANITIZE_THREAD})
+set(GGML_SANITIZE_ADDRESS   ${WHISPER_SANITIZE_ADDRESS})
+set(GGML_SANITIZE_UNDEFINED ${WHISPER_SANITIZE_UNDEFINED})
+set(GGML_ALL_WARNINGS       ${WHISPER_ALL_WARNINGS})
+set(GGML_FATAL_WARNINGS     ${WHISPER_FATAL_WARNINGS})
+
+# transition helpers
+function (whisper_option_depr TYPE OLD NEW)
+    if (${OLD})
+        message(${TYPE} "${OLD} is deprecated and will be removed in the future.\nUse ${NEW} instead\n")
+        set(${NEW} ON)
+    endif()
+endfunction()
+
+whisper_option_depr(FATAL_ERROR WHISPER_CUBLAS              GGML_CUDA)
+whisper_option_depr(WARNING     WHISPER_CUDA                GGML_CUDA)
+whisper_option_depr(WARNING     WHISPER_KOMPUTE             GGML_KOMPUTE)
+whisper_option_depr(WARNING     WHISPER_METAL               GGML_METAL)
+whisper_option_depr(WARNING     WHISPER_METAL_EMBED_LIBRARY GGML_METAL_EMBED_LIBRARY)
+whisper_option_depr(WARNING     WHISPER_NATIVE              GGML_NATIVE)
+whisper_option_depr(WARNING     WHISPER_OPENMP              GGML_OPENMP)
+whisper_option_depr(WARNING     WHISPER_RPC                 GGML_RPC)
+whisper_option_depr(WARNING     WHISPER_SYCL                GGML_SYCL)
+whisper_option_depr(WARNING     WHISPER_SYCL_F16            GGML_SYCL_F16)
+
+#
+# build the library
+#
+
+if (NOT TARGET ggml)
+    add_subdirectory(ggml)
+    # ... otherwise assume ggml is added by a parent CMakeLists.txt
+endif()
+add_subdirectory(src)
+
+#
+# install
+#
+
+include(GNUInstallDirs)
+include(CMakePackageConfigHelpers)
+
+set(WHISPER_BUILD_NUMBER        ${BUILD_NUMBER})
+set(WHISPER_BUILD_COMMIT        ${BUILD_COMMIT})
+set(WHISPER_INSTALL_VERSION     ${CMAKE_PROJECT_VERSION})
+
+set(WHISPER_INCLUDE_INSTALL_DIR ${CMAKE_INSTALL_INCLUDEDIR} CACHE PATH "Location of header  files")
+set(WHISPER_LIB_INSTALL_DIR     ${CMAKE_INSTALL_LIBDIR}     CACHE PATH "Location of library files")
+set(WHISPER_BIN_INSTALL_DIR     ${CMAKE_INSTALL_BINDIR}     CACHE PATH "Location of binary  files")
+
+get_directory_property(WHISPER_TRANSIENT_DEFINES COMPILE_DEFINITIONS)
+
+set_target_properties(whisper PROPERTIES PUBLIC_HEADER ${CMAKE_CURRENT_SOURCE_DIR}/include/whisper.h)
+install(TARGETS whisper LIBRARY PUBLIC_HEADER)
+
+configure_package_config_file(
+        ${CMAKE_CURRENT_SOURCE_DIR}/cmake/whisper-config.cmake.in
+        ${CMAKE_CURRENT_BINARY_DIR}/whisper-config.cmake
+    INSTALL_DESTINATION ${CMAKE_INSTALL_LIBDIR}/cmake/whisper
+    PATH_VARS
+    WHISPER_INCLUDE_INSTALL_DIR
+    WHISPER_LIB_INSTALL_DIR
+    WHISPER_BIN_INSTALL_DIR )
+
+write_basic_package_version_file(
+    ${CMAKE_CURRENT_BINARY_DIR}/whisper-version.cmake
+    VERSION ${WHISPER_INSTALL_VERSION}
+    COMPATIBILITY SameMajorVersion)
+
+install(FILES ${CMAKE_CURRENT_BINARY_DIR}/whisper-config.cmake
+              ${CMAKE_CURRENT_BINARY_DIR}/whisper-version.cmake
+        DESTINATION ${CMAKE_INSTALL_LIBDIR}/cmake/whisper)
+
+configure_file(cmake/whisper.pc.in
+        "${CMAKE_CURRENT_BINARY_DIR}/whisper.pc"
+        @ONLY)
+
+install(FILES "${CMAKE_CURRENT_BINARY_DIR}/whisper.pc"
+        DESTINATION lib/pkgconfig)
+
+#
+# programs, examples and tests
+#
+
+if (WHISPER_BUILD_TESTS AND NOT CMAKE_JS_VERSION)
+    #include(CTest)
+    #add_subdirectory(tests)
+endif ()
+
+if (WHISPER_BUILD_EXAMPLES)
+    add_subdirectory(examples)
+endif()
diff --git a/src/whisper_cpp/LICENSE b/src/whisper_cpp/LICENSE
index 76f67efd..acb96ce7 100644
--- a/src/whisper_cpp/LICENSE
+++ b/src/whisper_cpp/LICENSE
@@ -1,6 +1,6 @@
 MIT License
 
-Copyright (c) 2023 Georgi Gerganov
+Copyright (c) 2023-2024 The ggml authors
 
 Permission is hereby granted, free of charge, to any person obtaining a copy
 of this software and associated documentation files (the "Software"), to deal
diff --git a/src/whisper_cpp/Makefile b/src/whisper_cpp/Makefile
new file mode 100644
index 00000000..3c69aa85
--- /dev/null
+++ b/src/whisper_cpp/Makefile
@@ -0,0 +1,1164 @@
+# Define the default target now so that it is always the first target
+BUILD_TARGETS = \
+	main \
+	bench \
+	quantize \
+	server
+
+# Binaries only useful for tests
+TEST_TARGETS = \
+	tests/test-c.o
+
+# Deprecation aliases
+ifdef WHISPER_CUBLAS
+$(error WHISPER_CUBLAS is removed. Use GGML_CUDA instead.)
+endif
+
+ifdef WHISPER_CUDA
+GGML_CUDA := 1
+DEPRECATE_WARNING := 1
+endif
+
+ifdef WHISPER_KOMPUTE
+GGML_KOMPUTE := 1
+DEPRECATE_WARNING := 1
+endif
+
+ifdef WHISPER_METAL
+GGML_METAL := 1
+DEPRECATE_WARNING := 1
+endif
+
+ifdef WHISPER_OPENMP
+GGML_OPENMP := 1
+DEPRECATE_WARNING := 1
+endif
+
+ifdef WHISPER_RPC
+GGML_RPC := 1
+DEPRECATE_WARNING := 1
+endif
+
+ifdef WHISPER_SYCL
+GGML_SYCL := 1
+DEPRECATE_WARNING := 1
+endif
+
+ifdef WHISPER_SYCL_F16
+GGML_SYCL_F16 := 1
+DEPRECATE_WARNING := 1
+endif
+
+ifdef WHISPER_OPENBLAS
+GGML_OPENBLAS := 1
+DEPRECATE_WARNING := 1
+endif
+
+ifdef WHISPER_OPENBLAS64
+GGML_OPENBLAS64 := 1
+DEPRECATE_WARNING := 1
+endif
+
+ifdef WHISPER_BLIS
+GGML_BLIS := 1
+DEPRECATE_WARNING := 1
+endif
+
+ifdef WHISPER_NO_WHISPERFILE
+GGML_NO_WHISPERFILE := 1
+DEPRECATE_WARNING := 1
+endif
+
+ifdef WHISPER_NO_ACCELERATE
+GGML_NO_ACCELERATE := 1
+DEPRECATE_WARNING := 1
+endif
+
+ifdef WHISPER_NO_OPENMP
+GGML_NO_OPENMP := 1
+DEPRECATE_WARNING := 1
+endif
+
+ifdef WHISPER_NO_METAL
+GGML_NO_METAL := 1
+DEPRECATE_WARNING := 1
+endif
+
+ifndef UNAME_S
+UNAME_S := $(shell uname -s)
+endif
+
+ifndef UNAME_P
+UNAME_P := $(shell uname -p)
+endif
+
+ifndef UNAME_M
+UNAME_M := $(shell uname -m)
+endif
+
+# In GNU make default CXX is g++ instead of c++.  Let's fix that so that users
+# of non-gcc compilers don't have to provide g++ alias or wrapper.
+DEFCC  := cc
+DEFCXX := c++
+ifeq ($(origin CC),default)
+CC  := $(DEFCC)
+endif
+ifeq ($(origin CXX),default)
+CXX := $(DEFCXX)
+endif
+
+# Mac OS + Arm can report x86_64
+# ref: https://github.com/ggerganov/whisper.cpp/issues/66#issuecomment-1282546789
+ifeq ($(UNAME_S),Darwin)
+	ifndef GGML_NO_METAL
+		GGML_METAL := 1
+	endif
+
+	GGML_NO_OPENMP := 1
+
+	ifneq ($(UNAME_P),arm)
+		SYSCTL_M := $(shell sysctl -n hw.optional.arm64 2>/dev/null)
+		ifeq ($(SYSCTL_M),1)
+			# UNAME_P := arm
+			# UNAME_M := arm64
+			warn := $(warning Your arch is announced as x86_64, but it seems to actually be ARM64. Not fixing that can lead to bad performance. For more info see: https://github.com/ggerganov/whisper.cpp/issues/66\#issuecomment-1282546789)
+		endif
+	endif
+endif
+
+ifdef GGML_METAL
+	GGML_METAL_EMBED_LIBRARY := 1
+endif
+
+ifdef GGML_RPC
+	BUILD_TARGETS += rpc-server
+endif
+
+ifeq ($(shell sdl2-config --cflags --libs 2>/dev/null),)
+else
+	BUILD_TARGETS += \
+		command \
+		stream \
+		lsp \
+		talk-llama
+	# talk (TODO: disalbed)
+endif
+
+default: $(BUILD_TARGETS)
+
+test: $(TEST_TARGETS)
+	@failures=0; \
+	for test_target in $(TEST_TARGETS); do \
+		echo "Running test $$test_target..."; \
+		./$$test_target; \
+		if [ $$? -ne 0 ]; then \
+			printf 'Test %s FAILED!\n\n' $$test_target; \
+			failures=$$(( failures + 1 )); \
+		else \
+			printf 'Test %s passed.\n\n' $$test_target; \
+		fi; \
+	done; \
+	failures=$$(( failures + $$? )); \
+	if [ $$failures -gt 0 ]; then \
+		printf '\n%s tests failed.\n' $$failures; \
+		exit 1; \
+	fi
+	@echo 'All tests passed.'
+
+all: $(BUILD_TARGETS) $(TEST_TARGETS)
+
+ifdef RISCV_CROSS_COMPILE
+	CC	:= riscv64-unknown-linux-gnu-gcc
+	CXX	:= riscv64-unknown-linux-gnu-g++
+endif
+
+#
+# Compile flags
+#
+
+# keep standard at C11 and C++11
+MK_CPPFLAGS  = -Iggml/include -Iggml/src -Iinclude -Isrc -Iexamples
+MK_CFLAGS    = -std=c11   -fPIC
+MK_CXXFLAGS  = -std=c++11 -fPIC
+MK_NVCCFLAGS = -std=c++11
+
+ifndef WHISPER_NO_CCACHE
+CCACHE := $(shell which ccache)
+ifdef CCACHE
+export CCACHE_SLOPPINESS = time_macros
+$(info I ccache found, compilation results will be cached. Disable with WHISPER_NO_CCACHE.)
+CC    := $(CCACHE) $(CC)
+CXX   := $(CCACHE) $(CXX)
+else
+$(info I ccache not found. Consider installing it for faster compilation.)
+endif # CCACHE
+endif # WHISPER_NO_CCACHE
+
+# clock_gettime came in POSIX.1b (1993)
+# CLOCK_MONOTONIC came in POSIX.1-2001 / SUSv3 as optional
+# posix_memalign came in POSIX.1-2001 / SUSv3
+# M_PI is an XSI extension since POSIX.1-2001 / SUSv3, came in XPG1 (1985)
+MK_CPPFLAGS += -D_XOPEN_SOURCE=600
+
+# Somehow in OpenBSD whenever POSIX conformance is specified
+# some string functions rely on locale_t availability,
+# which was introduced in POSIX.1-2008, forcing us to go higher
+ifeq ($(UNAME_S),OpenBSD)
+	MK_CPPFLAGS += -U_XOPEN_SOURCE -D_XOPEN_SOURCE=700
+endif
+
+# Data types, macros and functions related to controlling CPU affinity and
+# some memory allocation are available on Linux through GNU extensions in libc
+ifeq ($(UNAME_S),Linux)
+	MK_CPPFLAGS += -D_GNU_SOURCE
+endif
+
+# RLIMIT_MEMLOCK came in BSD, is not specified in POSIX.1,
+# and on macOS its availability depends on enabling Darwin extensions
+# similarly on DragonFly, enabling BSD extensions is necessary
+ifeq ($(UNAME_S),Darwin)
+	MK_CPPFLAGS += -D_DARWIN_C_SOURCE
+endif
+ifeq ($(UNAME_S),DragonFly)
+	MK_CPPFLAGS += -D__BSD_VISIBLE
+endif
+
+# alloca is a non-standard interface that is not visible on BSDs when
+# POSIX conformance is specified, but not all of them provide a clean way
+# to enable it in such cases
+ifeq ($(UNAME_S),FreeBSD)
+	MK_CPPFLAGS += -D__BSD_VISIBLE
+endif
+ifeq ($(UNAME_S),NetBSD)
+	MK_CPPFLAGS += -D_NETBSD_SOURCE
+endif
+ifeq ($(UNAME_S),OpenBSD)
+	MK_CPPFLAGS += -D_BSD_SOURCE
+endif
+
+ifdef GGML_SCHED_MAX_COPIES
+	MK_CPPFLAGS += -DGGML_SCHED_MAX_COPIES=$(GGML_SCHED_MAX_COPIES)
+endif
+
+ifdef WHISPER_DEBUG
+	MK_CFLAGS    += -O0 -g
+	MK_CXXFLAGS  += -O0 -g
+	MK_LDFLAGS   += -g
+	MK_NVCCFLAGS += -O0 -g
+
+	ifeq ($(UNAME_S),Linux)
+		MK_CPPFLAGS += -D_GLIBCXX_ASSERTIONS
+	endif
+else
+	MK_CPPFLAGS   += -DNDEBUG
+	MK_CFLAGS     += -O3
+	MK_CXXFLAGS   += -O3
+	MK_NVCCFLAGS  += -O3
+endif
+
+ifdef WHISPER_SANITIZE_THREAD
+	MK_CFLAGS   += -fsanitize=thread -g
+	MK_CXXFLAGS += -fsanitize=thread -g
+	MK_LDFLAGS  += -fsanitize=thread -g
+endif
+
+ifdef WHISPER_SANITIZE_ADDRESS
+	MK_CFLAGS   += -fsanitize=address -fno-omit-frame-pointer -g
+	MK_CXXFLAGS += -fsanitize=address -fno-omit-frame-pointer -g
+	MK_LDFLAGS  += -fsanitize=address -fno-omit-frame-pointer -g
+endif
+
+ifdef WHISPER_SANITIZE_UNDEFINED
+	MK_CFLAGS   += -fsanitize=undefined -g
+	MK_CXXFLAGS += -fsanitize=undefined -g
+	MK_LDFLAGS  += -fsanitize=undefined -g
+endif
+
+ifdef WHISPER_SERVER_VERBOSE
+	MK_CPPFLAGS += -DSERVER_VERBOSE=$(WHISPER_SERVER_VERBOSE)
+endif
+
+ifdef WHISPER_SERVER_SSL
+	MK_CPPFLAGS += -DCPPHTTPLIB_OPENSSL_SUPPORT
+	MK_LDFLAGS  += -lssl -lcrypto
+endif
+
+ifdef WHISPER_DISABLE_LOGS
+	MK_CPPFLAGS += -DLOG_DISABLE_LOGS
+endif # WHISPER_DISABLE_LOGS
+
+# warnings
+WARN_FLAGS = \
+	-Wall \
+	-Wextra \
+	-Wpedantic \
+	-Wcast-qual \
+	-Wno-unused-function
+
+MK_CFLAGS += \
+	$(WARN_FLAGS) \
+	-Wshadow \
+	-Wstrict-prototypes \
+	-Wpointer-arith \
+	-Wmissing-prototypes \
+	-Werror=implicit-int \
+	-Werror=implicit-function-declaration
+
+MK_CXXFLAGS += \
+	$(WARN_FLAGS) \
+	-Wmissing-declarations \
+	-Wmissing-noreturn
+
+ifeq ($(WHISPER_FATAL_WARNINGS),1)
+	MK_CFLAGS   += -Werror
+	MK_CXXFLAGS += -Werror
+endif
+
+# this version of Apple ld64 is buggy
+ifneq '' '$(findstring dyld-1015.7,$(shell $(CC) $(LDFLAGS) -Wl,-v 2>&1))'
+	MK_CPPFLAGS += -DHAVE_BUGGY_APPLE_LINKER
+endif
+
+# OS specific
+# TODO: support Windows
+ifneq '' '$(filter $(UNAME_S),Linux Darwin FreeBSD NetBSD OpenBSD Haiku)'
+	MK_CFLAGS   += -pthread
+	MK_CXXFLAGS += -pthread
+endif
+
+# detect Windows
+ifneq ($(findstring _NT,$(UNAME_S)),)
+	_WIN32 := 1
+endif
+
+# library name prefix
+ifneq ($(_WIN32),1)
+	LIB_PRE := lib
+endif
+
+# Dynamic Shared Object extension
+ifneq ($(_WIN32),1)
+	DSO_EXT := .so
+else
+	DSO_EXT := .dll
+endif
+
+# Windows Sockets 2 (Winsock) for network-capable apps
+ifeq ($(_WIN32),1)
+	LWINSOCK2 := -lws2_32
+endif
+
+ifdef WHISPER_GPROF
+	MK_CFLAGS   += -pg
+	MK_CXXFLAGS += -pg
+endif
+
+# Architecture specific
+# TODO: probably these flags need to be tweaked on some architectures
+#       feel free to update the Makefile for your architecture and send a pull request or issue
+
+ifndef RISCV
+
+ifeq ($(UNAME_M),$(filter $(UNAME_M),x86_64 i686 amd64))
+	# Use all CPU extensions that are available:
+	MK_CFLAGS     += -march=native -mtune=native
+	HOST_CXXFLAGS += -march=native -mtune=native
+
+	# Usage AVX-only
+	#MK_CFLAGS   += -mfma -mf16c -mavx
+	#MK_CXXFLAGS += -mfma -mf16c -mavx
+
+	# Usage SSSE3-only (Not is SSE3!)
+	#MK_CFLAGS   += -mssse3
+	#MK_CXXFLAGS += -mssse3
+endif
+
+ifneq '' '$(findstring mingw,$(shell $(CC) -dumpmachine))'
+	# The stack is only 16-byte aligned on Windows, so don't let gcc emit aligned moves.
+	# https://gcc.gnu.org/bugzilla/show_bug.cgi?id=54412
+	# https://github.com/ggerganov/llama.cpp/issues/2922
+	MK_CFLAGS   += -Xassembler -muse-unaligned-vector-move
+	MK_CXXFLAGS += -Xassembler -muse-unaligned-vector-move
+
+	# Target Windows 8 for PrefetchVirtualMemory
+	MK_CPPFLAGS += -D_WIN32_WINNT=0x602
+endif
+
+ifneq ($(filter aarch64%,$(UNAME_M)),)
+	# Apple M1, M2, etc.
+	# Raspberry Pi 3, 4, Zero 2 (64-bit)
+	# Nvidia Jetson
+	MK_CFLAGS   += -mcpu=native
+	MK_CXXFLAGS += -mcpu=native
+	JETSON_RELEASE_INFO = $(shell jetson_release)
+	ifdef JETSON_RELEASE_INFO
+		ifneq ($(filter TX2%,$(JETSON_RELEASE_INFO)),)
+			JETSON_EOL_MODULE_DETECT = 1
+			CC = aarch64-unknown-linux-gnu-gcc
+			cxx = aarch64-unknown-linux-gnu-g++
+		endif
+	endif
+endif
+
+ifneq ($(filter armv6%,$(UNAME_M)),)
+	# Raspberry Pi 1, Zero
+	MK_CFLAGS   += -mfpu=neon-fp-armv8 -mfp16-format=ieee -mno-unaligned-access
+	MK_CXXFLAGS += -mfpu=neon-fp-armv8 -mfp16-format=ieee -mno-unaligned-access
+endif
+
+ifneq ($(filter armv7%,$(UNAME_M)),)
+	# Raspberry Pi 2
+	MK_CFLAGS   += -mfpu=neon-fp-armv8 -mfp16-format=ieee -mno-unaligned-access -funsafe-math-optimizations
+	MK_CXXFLAGS += -mfpu=neon-fp-armv8 -mfp16-format=ieee -mno-unaligned-access -funsafe-math-optimizations
+endif
+
+ifneq ($(filter armv8%,$(UNAME_M)),)
+	# Raspberry Pi 3, 4, Zero 2 (32-bit)
+	MK_CFLAGS   += -mfp16-format=ieee -mno-unaligned-access
+	MK_CXXFLAGS += -mfp16-format=ieee -mno-unaligned-access
+endif
+
+ifneq ($(filter ppc64%,$(UNAME_M)),)
+	POWER9_M := $(shell grep "POWER9" /proc/cpuinfo)
+	ifneq (,$(findstring POWER9,$(POWER9_M)))
+		MK_CFLAGS   += -mcpu=power9
+		MK_CXXFLAGS += -mcpu=power9
+	endif
+endif
+
+ifneq ($(filter ppc64le%,$(UNAME_M)),)
+	MK_CFLAGS   += -mcpu=powerpc64le
+	MK_CXXFLAGS += -mcpu=powerpc64le
+	CUDA_POWER_ARCH = 1
+endif
+
+ifneq ($(filter loongarch64%,$(UNAME_M)),)
+	MK_CFLAGS   += -mlasx
+	MK_CXXFLAGS += -mlasx
+endif
+
+else
+	MK_CFLAGS   += -march=rv64gcv -mabi=lp64d
+	MK_CXXFLAGS += -march=rv64gcv -mabi=lp64d
+endif
+
+ifndef GGML_NO_ACCELERATE
+	# Mac OS - include Accelerate framework.
+	# `-framework Accelerate` works both with Apple Silicon and Mac Intel
+	ifeq ($(UNAME_S),Darwin)
+		MK_CPPFLAGS += -DGGML_USE_ACCELERATE -DGGML_USE_BLAS
+		MK_CPPFLAGS += -DACCELERATE_NEW_LAPACK
+		MK_CPPFLAGS += -DACCELERATE_LAPACK_ILP64
+		MK_LDFLAGS  += -framework Accelerate
+		OBJ_GGML    += ggml/src/ggml-blas.o
+	endif
+endif # GGML_NO_ACCELERATE
+
+ifndef GGML_NO_OPENMP
+	MK_CPPFLAGS += -DGGML_USE_OPENMP
+	MK_CFLAGS   += -fopenmp
+	MK_CXXFLAGS += -fopenmp
+endif # GGML_NO_OPENMP
+
+ifdef GGML_OPENBLAS
+	MK_CPPFLAGS += -DGGML_USE_BLAS $(shell pkg-config --cflags-only-I openblas)
+	MK_CFLAGS   += $(shell pkg-config --cflags-only-other openblas)
+	MK_LDFLAGS  += $(shell pkg-config --libs openblas)
+	OBJ_GGML    += ggml/src/ggml-blas.o
+endif # GGML_OPENBLAS
+
+ifdef GGML_OPENBLAS64
+	MK_CPPFLAGS += -DGGML_USE_BLAS $(shell pkg-config --cflags-only-I openblas64)
+	MK_CFLAGS   += $(shell pkg-config --cflags-only-other openblas64)
+	MK_LDFLAGS  += $(shell pkg-config --libs openblas64)
+	OBJ_GGML    += ggml/src/ggml-blas.o
+endif # GGML_OPENBLAS64
+
+ifdef GGML_BLIS
+	MK_CPPFLAGS += -DGGML_USE_BLAS -I/usr/local/include/blis -I/usr/include/blis
+	MK_LDFLAGS  += -lblis -L/usr/local/lib
+	OBJ_GGML    += ggml/src/ggml-blas.o
+endif # GGML_BLIS
+
+ifdef GGML_RPC
+	MK_CPPFLAGS += -DGGML_USE_RPC
+	OBJ_GGML    += ggml/src/ggml-rpc.o
+endif # GGML_RPC
+
+OBJ_CUDA_TMPL      = $(patsubst %.cu,%.o,$(wildcard ggml/src/ggml-cuda/template-instances/fattn-wmma*.cu))
+OBJ_CUDA_TMPL     += $(patsubst %.cu,%.o,$(wildcard ggml/src/ggml-cuda/template-instances/mmq*.cu))
+
+ifdef GGML_CUDA_FA_ALL_QUANTS
+	OBJ_CUDA_TMPL += $(patsubst %.cu,%.o,$(wildcard ggml/src/ggml-cuda/template-instances/fattn-vec*.cu))
+else
+	OBJ_CUDA_TMPL += $(patsubst %.cu,%.o,$(wildcard ggml/src/ggml-cuda/template-instances/fattn-vec*q4_0-q4_0.cu))
+	OBJ_CUDA_TMPL += $(patsubst %.cu,%.o,$(wildcard ggml/src/ggml-cuda/template-instances/fattn-vec*q8_0-q8_0.cu))
+	OBJ_CUDA_TMPL += $(patsubst %.cu,%.o,$(wildcard ggml/src/ggml-cuda/template-instances/fattn-vec*f16-f16.cu))
+endif # GGML_CUDA_FA_ALL_QUANTS
+
+ifdef GGML_CUDA
+	ifneq ('', '$(wildcard /opt/cuda)')
+		CUDA_PATH ?= /opt/cuda
+	else
+		CUDA_PATH ?= /usr/local/cuda
+	endif
+
+	#MK_CPPFLAGS  += -DGGML_USE_CUDA -I$(CUDA_PATH)/include -I$(CUDA_PATH)/targets/$(UNAME_M)-linux/include -DGGML_CUDA_USE_GRAPHS
+	#MK_LDFLAGS   += -lcuda -lcublas -lculibos -lcudart -lcufft -lcublasLt -lpthread -ldl -lrt -L$(CUDA_PATH)/lib64 -L/usr/lib64 -L$(CUDA_PATH)/targets/$(UNAME_M)-linux/lib -L$(CUDA_PATH)/lib64/stubs -L/usr/lib/wsl/lib
+	MK_CPPFLAGS  += -DGGML_USE_CUDA -I$(CUDA_PATH)/include -I$(CUDA_PATH)/targets/$(UNAME_M)-linux/include
+	MK_LDFLAGS   += -lcuda -lcublas -lculibos -lcudart -lcublasLt -lpthread -ldl -lrt -L$(CUDA_PATH)/lib64 -L/usr/lib64 -L$(CUDA_PATH)/targets/$(UNAME_M)-linux/lib -L$(CUDA_PATH)/lib64/stubs -L/usr/lib/wsl/lib
+	MK_NVCCFLAGS += -use_fast_math
+
+	OBJ_GGML += ggml/src/ggml-cuda.o
+	OBJ_GGML += $(patsubst %.cu,%.o,$(wildcard ggml/src/ggml-cuda/*.cu))
+	OBJ_GGML += $(OBJ_CUDA_TMPL)
+
+	#OBJ_WHISPER += src/whisper-mel-cuda.o
+
+ifdef WHISPER_FATAL_WARNINGS
+	MK_NVCCFLAGS += -Werror all-warnings
+endif # WHISPER_FATAL_WARNINGS
+
+ifndef JETSON_EOL_MODULE_DETECT
+	MK_NVCCFLAGS += --forward-unknown-to-host-compiler
+endif # JETSON_EOL_MODULE_DETECT
+
+ifdef WHISPER_DEBUG
+	MK_NVCCFLAGS += -lineinfo
+endif # WHISPER_DEBUG
+
+ifdef GGML_CUDA_DEBUG
+	MK_NVCCFLAGS += --device-debug
+endif # GGML_CUDA_DEBUG
+
+ifdef GGML_CUDA_NVCC
+	NVCC = $(CCACHE) $(GGML_CUDA_NVCC)
+else
+	NVCC = $(CCACHE) nvcc
+endif #GGML_CUDA_NVCC
+
+ifdef CUDA_DOCKER_ARCH
+	MK_NVCCFLAGS += -Wno-deprecated-gpu-targets -arch=$(CUDA_DOCKER_ARCH)
+else ifndef CUDA_POWER_ARCH
+	MK_NVCCFLAGS += -arch=native
+endif # CUDA_DOCKER_ARCH
+
+ifdef GGML_CUDA_FORCE_DMMV
+	MK_NVCCFLAGS += -DGGML_CUDA_FORCE_DMMV
+endif # GGML_CUDA_FORCE_DMMV
+
+ifdef GGML_CUDA_FORCE_MMQ
+	MK_NVCCFLAGS += -DGGML_CUDA_FORCE_MMQ
+endif # GGML_CUDA_FORCE_MMQ
+
+ifdef GGML_CUDA_DMMV_X
+	MK_NVCCFLAGS += -DGGML_CUDA_DMMV_X=$(GGML_CUDA_DMMV_X)
+else
+	MK_NVCCFLAGS += -DGGML_CUDA_DMMV_X=32
+endif # GGML_CUDA_DMMV_X
+
+ifdef GGML_CUDA_MMV_Y
+	MK_NVCCFLAGS += -DGGML_CUDA_MMV_Y=$(GGML_CUDA_MMV_Y)
+else ifdef GGML_CUDA_DMMV_Y
+	MK_NVCCFLAGS += -DGGML_CUDA_MMV_Y=$(GGML_CUDA_DMMV_Y) # for backwards compatibility
+else
+	MK_NVCCFLAGS += -DGGML_CUDA_MMV_Y=1
+endif # GGML_CUDA_MMV_Y
+
+ifdef GGML_CUDA_F16
+	MK_NVCCFLAGS += -DGGML_CUDA_F16
+endif # GGML_CUDA_F16
+
+ifdef GGML_CUDA_DMMV_F16
+	MK_NVCCFLAGS += -DGGML_CUDA_F16
+endif # GGML_CUDA_DMMV_F16
+
+ifdef GGML_CUDA_KQUANTS_ITER
+	MK_NVCCFLAGS += -DK_QUANTS_PER_ITERATION=$(GGML_CUDA_KQUANTS_ITER)
+else
+	MK_NVCCFLAGS += -DK_QUANTS_PER_ITERATION=2
+endif
+
+ifdef GGML_CUDA_PEER_MAX_BATCH_SIZE
+	MK_NVCCFLAGS += -DGGML_CUDA_PEER_MAX_BATCH_SIZE=$(GGML_CUDA_PEER_MAX_BATCH_SIZE)
+else
+	MK_NVCCFLAGS += -DGGML_CUDA_PEER_MAX_BATCH_SIZE=128
+endif # GGML_CUDA_PEER_MAX_BATCH_SIZE
+
+ifdef GGML_CUDA_NO_PEER_COPY
+	MK_NVCCFLAGS += -DGGML_CUDA_NO_PEER_COPY
+endif # GGML_CUDA_NO_PEER_COPY
+
+ifdef GGML_CUDA_CCBIN
+	MK_NVCCFLAGS += -ccbin $(GGML_CUDA_CCBIN)
+endif # GGML_CUDA_CCBIN
+
+ifdef GGML_CUDA_FA_ALL_QUANTS
+	MK_NVCCFLAGS += -DGGML_CUDA_FA_ALL_QUANTS
+endif # GGML_CUDA_FA_ALL_QUANTS
+
+ifdef JETSON_EOL_MODULE_DETECT
+define NVCC_COMPILE
+	$(NVCC) -I. -Icommon -D_XOPEN_SOURCE=600 -D_GNU_SOURCE -DNDEBUG -DGGML_USE_CUDA -I/usr/local/cuda/include -I/opt/cuda/include -I/usr/local/cuda/targets/aarch64-linux/include -std=c++11 -O3 $(NVCCFLAGS) $(CPPFLAGS) -Xcompiler "$(CUDA_CXXFLAGS)" -c $< -o $@
+endef # NVCC_COMPILE
+else
+define NVCC_COMPILE
+	$(NVCC) $(NVCCFLAGS) $(CPPFLAGS) -Xcompiler "$(CUDA_CXXFLAGS)" -c $< -o $@
+endef # NVCC_COMPILE
+endif # JETSON_EOL_MODULE_DETECT
+
+ggml/src/ggml-cuda/%.o: \
+	ggml/src/ggml-cuda/%.cu \
+	ggml/include/ggml.h \
+	ggml/src/ggml-common.h \
+	ggml/src/ggml-cuda/common.cuh
+	$(NVCC_COMPILE)
+
+ggml/src/ggml-cuda.o: \
+	ggml/src/ggml-cuda.cu \
+	ggml/include/ggml.h \
+	ggml/include/ggml-backend.h \
+	ggml/include/ggml-cuda.h \
+	ggml/src/ggml-backend-impl.h \
+	ggml/src/ggml-common.h \
+	$(wildcard ggml/src/ggml-cuda/*.cuh)
+	$(NVCC_COMPILE)
+
+#src/whisper-mel-cuda.o: src/whisper-mel-cuda.cu src/whisper-mel-cuda.hpp
+#	$(NVCC) $(NVCCFLAGS) $(CPPFLAGS) -Xcompiler "$(CUDA_CXXFLAGS)" -c $< -o $@
+
+endif # GGML_CUDA
+
+ifdef GGML_VULKAN
+	MK_CPPFLAGS += -DGGML_USE_VULKAN
+	MK_LDFLAGS  += -lvulkan
+	OBJ_GGML    += ggml/src/ggml-vulkan.o
+
+ifdef GGML_VULKAN_CHECK_RESULTS
+	MK_CPPFLAGS  += -DGGML_VULKAN_CHECK_RESULTS
+endif
+
+ifdef GGML_VULKAN_DEBUG
+	MK_CPPFLAGS  += -DGGML_VULKAN_DEBUG
+endif
+
+ifdef GGML_VULKAN_MEMORY_DEBUG
+	MK_CPPFLAGS  += -DGGML_VULKAN_MEMORY_DEBUG
+endif
+
+ifdef GGML_VULKAN_VALIDATE
+	MK_CPPFLAGS  += -DGGML_VULKAN_VALIDATE
+endif
+
+ifdef GGML_VULKAN_RUN_TESTS
+	MK_CPPFLAGS  += -DGGML_VULKAN_RUN_TESTS
+endif
+
+ggml/src/ggml-vulkan.o: \
+	ggml/src/ggml-vulkan.cpp \
+	ggml/include/ggml-vulkan.h
+	$(CXX) $(CXXFLAGS) -c $< -o $@
+endif # GGML_VULKAN
+
+ifdef GGML_HIPBLAS
+	ifeq ($(wildcard /opt/rocm),)
+		ROCM_PATH      ?= /usr
+		AMDGPU_TARGETS ?= $(shell $(shell which amdgpu-arch))
+	else
+		ROCM_PATH	?= /opt/rocm
+		AMDGPU_TARGETS ?= $(shell $(ROCM_PATH)/llvm/bin/amdgpu-arch)
+	endif
+
+	GGML_CUDA_DMMV_X       ?= 32
+	GGML_CUDA_MMV_Y        ?= 1
+	GGML_CUDA_KQUANTS_ITER ?= 2
+
+	MK_CPPFLAGS += -DGGML_USE_HIPBLAS -DGGML_USE_CUDA
+
+ifdef GGML_HIP_UMA
+	MK_CPPFLAGS += -DGGML_HIP_UMA
+endif # GGML_HIP_UMA
+
+	MK_LDFLAGS += -L$(ROCM_PATH)/lib -Wl,-rpath=$(ROCM_PATH)/lib
+	MK_LDFLAGS += -L$(ROCM_PATH)/lib64 -Wl,-rpath=$(ROCM_PATH)/lib64
+	MK_LDFLAGS += -lhipblas -lamdhip64 -lrocblas
+
+	HIPCC ?= $(CCACHE) $(ROCM_PATH)/bin/hipcc
+
+	HIPFLAGS += $(addprefix --offload-arch=,$(AMDGPU_TARGETS))
+	HIPFLAGS += -DGGML_CUDA_DMMV_X=$(GGML_CUDA_DMMV_X)
+	HIPFLAGS += -DGGML_CUDA_MMV_Y=$(GGML_CUDA_MMV_Y)
+	HIPFLAGS += -DK_QUANTS_PER_ITERATION=$(GGML_CUDA_KQUANTS_ITER)
+
+ifdef GGML_CUDA_FORCE_DMMV
+	HIPFLAGS += -DGGML_CUDA_FORCE_DMMV
+endif # GGML_CUDA_FORCE_DMMV
+
+ifdef GGML_CUDA_NO_PEER_COPY
+	HIPFLAGS += -DGGML_CUDA_NO_PEER_COPY
+endif # GGML_CUDA_NO_PEER_COPY
+
+	OBJ_GGML += ggml/src/ggml-cuda.o
+	OBJ_GGML += $(patsubst %.cu,%.o,$(wildcard ggml/src/ggml-cuda/*.cu))
+	OBJ_GGML += $(OBJ_CUDA_TMPL)
+
+ggml/src/ggml-cuda.o: \
+	ggml/src/ggml-cuda.cu \
+	ggml/include/ggml.h \
+	ggml/include/ggml-backend.h \
+	ggml/include/ggml-cuda.h \
+	ggml/src/ggml-backend-impl.h \
+	ggml/src/ggml-common.h \
+	$(wildcard ggml/src/ggml-cuda/*.cuh)
+	$(HIPCC) $(CXXFLAGS) $(HIPFLAGS) -x hip -c -o $@ $<
+
+ggml/src/ggml-cuda/%.o: \
+	ggml/src/ggml-cuda/%.cu \
+	ggml/include/ggml.h \
+	ggml/src/ggml-common.h \
+	ggml/src/ggml-cuda/common.cuh
+	$(HIPCC) $(CXXFLAGS) $(HIPFLAGS) -x hip -c -o $@ $<
+endif # GGML_HIPBLAS
+
+ifdef GGML_METAL
+	MK_CPPFLAGS += -DGGML_USE_METAL
+	MK_LDFLAGS  += -framework Foundation -framework Metal -framework MetalKit
+	OBJ_GGML	+= ggml/src/ggml-metal.o
+ifdef GGML_METAL_NDEBUG
+	MK_CPPFLAGS += -DGGML_METAL_NDEBUG
+endif
+
+ifdef GGML_METAL_EMBED_LIBRARY
+	MK_CPPFLAGS += -DGGML_METAL_EMBED_LIBRARY
+	OBJ_GGML    += ggml/src/ggml-metal-embed.o
+endif
+endif # GGML_METAL
+
+ifdef WHISPER_COREML
+	MK_CXXFLAGS += -DWHISPER_USE_COREML
+	LDFLAGS     += -framework Foundation -framework CoreML
+
+ifdef WHISPER_COREML_ALLOW_FALLBACK
+	MK_CXXFLAGS += -DWHISPER_COREML_ALLOW_FALLBACK
+endif
+endif
+
+# ===
+
+ifdef GGML_METAL
+ggml/src/ggml-metal.o: \
+	ggml/src/ggml-metal.m \
+	ggml/include/ggml-metal.h \
+	ggml/include/ggml.h
+	$(CC) $(CFLAGS) -c $< -o $@
+
+ifdef GGML_METAL_EMBED_LIBRARY
+ggml/src/ggml-metal-embed.o: \
+	ggml/src/ggml-metal.metal \
+	ggml/src/ggml-common.h
+	@echo "Embedding Metal library"
+	@sed -e '/#include "ggml-common.h"/r ggml/src/ggml-common.h' -e '/#include "ggml-common.h"/d' < ggml/src/ggml-metal.metal > ggml/src/ggml-metal-embed.metal
+	$(eval TEMP_ASSEMBLY=$(shell mktemp))
+	@echo ".section __DATA, __ggml_metallib"            >  $(TEMP_ASSEMBLY)
+	@echo ".globl _ggml_metallib_start"                 >> $(TEMP_ASSEMBLY)
+	@echo "_ggml_metallib_start:"                       >> $(TEMP_ASSEMBLY)
+	@echo ".incbin \"ggml/src/ggml-metal-embed.metal\"" >> $(TEMP_ASSEMBLY)
+	@echo ".globl _ggml_metallib_end"                   >> $(TEMP_ASSEMBLY)
+	@echo "_ggml_metallib_end:"                         >> $(TEMP_ASSEMBLY)
+	@$(AS) $(TEMP_ASSEMBLY) -o $@
+	@rm -f ${TEMP_ASSEMBLY}
+endif
+endif # GGML_METAL
+
+ifdef WHISPER_COREML
+src/coreml/whisper-encoder.o: src/coreml/whisper-encoder.mm src/coreml/whisper-encoder.h
+	$(CXX) -O3 -I . -fobjc-arc -c src/coreml/whisper-encoder.mm -o src/coreml/whisper-encoder.o
+
+src/coreml/whisper-encoder-impl.o: src/coreml/whisper-encoder-impl.m src/coreml/whisper-encoder-impl.h
+	$(CXX) -O3 -I . -fobjc-arc -c src/coreml/whisper-encoder-impl.m -o src/coreml/whisper-encoder-impl.o
+
+OBJ_WHISPER += src/coreml/whisper-encoder.o src/coreml/whisper-encoder-impl.o
+endif
+
+OBJ_GGML += \
+	ggml/src/ggml.o \
+	ggml/src/ggml-alloc.o \
+	ggml/src/ggml-backend.o \
+	ggml/src/ggml-quants.o \
+	ggml/src/ggml-aarch64.o
+
+OBJ_WHISPER += \
+	src/whisper.o
+
+OBJ_COMMON += \
+	examples/common.o \
+	examples/common-ggml.o \
+	examples/grammar-parser.o
+
+OBJ_SDL += \
+	examples/common-sdl.o
+
+OBJ_ALL = $(OBJ_GGML) $(OBJ_WHISPER) $(OBJ_COMMON) $(OBJ_SDL)
+
+LIB_GGML   = $(LIB_PRE)ggml$(DSO_EXT)
+LIB_GGML_S = $(LIB_PRE)ggml.a
+
+LIB_WHISPER   = $(LIB_PRE)whisper$(DSO_EXT)
+LIB_WHISPER_S = $(LIB_PRE)whisper.a
+
+LIB_COMMON   = $(LIB_PRE)common$(DSO_EXT)
+LIB_COMMON_S = $(LIB_PRE)common.a
+
+LIB_COMMON_SDL   = $(LIB_PRE)common-sdl$(DSO_EXT)
+LIB_COMMON_SDL_S = $(LIB_PRE)common-sdl.a
+
+LIB_ALL   = $(LIB_GGML)   $(LIB_WHISPER)   $(LIB_COMMON)   $(LIB_COMMON_SDL)
+LIB_ALL_S = $(LIB_GGML_S) $(LIB_WHISPER_S) $(LIB_COMMON_S) $(LIB_COMMON_SDL_S)
+
+GF_CC := $(CC)
+include scripts/get-flags.mk
+
+# combine build flags with cmdline overrides
+override CPPFLAGS  := $(MK_CPPFLAGS) $(CPPFLAGS)
+override CFLAGS    := $(CPPFLAGS) $(MK_CFLAGS) $(GF_CFLAGS) $(CFLAGS)
+BASE_CXXFLAGS      := $(MK_CXXFLAGS) $(CXXFLAGS)
+override CXXFLAGS  := $(BASE_CXXFLAGS) $(HOST_CXXFLAGS) $(GF_CXXFLAGS) $(CPPFLAGS)
+override NVCCFLAGS := $(MK_NVCCFLAGS) $(NVCCFLAGS)
+override LDFLAGS   := $(MK_LDFLAGS) $(LDFLAGS)
+
+# identify CUDA host compiler
+ifdef GGML_CUDA
+GF_CC := $(NVCC) $(NVCCFLAGS) 2>/dev/null .c -Xcompiler
+include scripts/get-flags.mk
+CUDA_CXXFLAGS := $(BASE_CXXFLAGS) $(GF_CXXFLAGS) -Wno-pedantic
+endif
+
+ifdef WHISPER_CURL
+override CXXFLAGS := $(CXXFLAGS) -DWHISPER_USE_CURL
+override LDFLAGS  := $(LDFLAGS) -lcurl
+endif
+
+#
+# Print build information
+#
+
+$(info I whisper.cpp build info: )
+$(info I UNAME_S:   $(UNAME_S))
+$(info I UNAME_P:   $(UNAME_P))
+$(info I UNAME_M:   $(UNAME_M))
+$(info I CFLAGS:    $(CFLAGS))
+$(info I CXXFLAGS:  $(CXXFLAGS))
+$(info I NVCCFLAGS: $(NVCCFLAGS))
+$(info I LDFLAGS:   $(LDFLAGS))
+$(info I CC:        $(shell $(CC)   --version | head -n 1))
+$(info I CXX:       $(shell $(CXX)  --version | head -n 1))
+ifdef GGML_CUDA
+$(info I NVCC:      $(shell $(NVCC) --version | tail -n 1))
+CUDA_VERSION := $(shell $(NVCC) --version | grep -oP 'release (\K[0-9]+\.[0-9])')
+ifeq ($(shell awk -v "v=$(CUDA_VERSION)" 'BEGIN { print (v < 11.7) }'),1)
+
+ifndef CUDA_DOCKER_ARCH
+ifndef CUDA_POWER_ARCH
+$(error I ERROR: For CUDA versions < 11.7 a target CUDA architecture must be explicitly provided via environment variable CUDA_DOCKER_ARCH, e.g. by running "export CUDA_DOCKER_ARCH=compute_XX" on Unix-like systems, where XX is the minimum compute capability that the code needs to run on. A list with compute capabilities can be found here: https://developer.nvidia.com/cuda-gpus )
+endif # CUDA_POWER_ARCH
+endif # CUDA_DOCKER_ARCH
+
+endif # eq ($(shell echo "$(CUDA_VERSION) < 11.7" | bc),1)
+endif # GGML_CUDA
+$(info )
+
+ifdef DEPRECATE_WARNING
+$(info !!! DEPRECATION WARNING !!!)
+$(info The following WHISPER_ options are deprecated and will be removed in the future. Use the GGML_ prefix instead)
+$(info   - WHISPER_CUDA)
+$(info   - WHISPER_METAL)
+$(info   - WHISPER_OPENMP)
+$(info   - WHISPER_RPC)
+$(info   - WHISPER_SYCL)
+$(info   - WHISPER_SYCL_F16)
+$(info   - WHISPER_OPENBLAS)
+$(info   - WHISPER_OPENBLAS64)
+$(info   - WHISPER_BLIS)
+$(info   - WHISPER_NO_LLAMAFILE)
+$(info   - WHISPER_NO_ACCELERATE)
+$(info   - WHISPER_NO_OPENMP)
+$(info   - WHISPER_NO_METAL)
+$(info )
+endif
+
+#
+# Build libraries
+#
+
+# ggml
+
+ggml/src/ggml.o: \
+	ggml/src/ggml.c \
+	ggml/include/ggml.h
+	$(CC)  $(CFLAGS)   -c $< -o $@
+
+ggml/src/ggml-alloc.o: \
+	ggml/src/ggml-alloc.c \
+	ggml/include/ggml.h \
+	ggml/include/ggml-alloc.h
+	$(CC)  $(CFLAGS)   -c $< -o $@
+
+ggml/src/ggml-backend.o: \
+	ggml/src/ggml-backend.c \
+	ggml/include/ggml.h \
+	ggml/include/ggml-backend.h
+	$(CC)  $(CFLAGS)   -c $< -o $@
+
+ggml/src/ggml-quants.o: \
+	ggml/src/ggml-quants.c \
+	ggml/include/ggml.h \
+	ggml/src/ggml-quants.h \
+	ggml/src/ggml-common.h
+	$(CC) $(CFLAGS)    -c $< -o $@
+
+ggml/src/ggml-aarch64.o: \
+	ggml/src/ggml-aarch64.c \
+	ggml/include/ggml.h \
+	ggml/src/ggml-aarch64.h \
+	ggml/src/ggml-common.h
+	$(CC) $(CFLAGS)    -c $< -o $@
+
+ggml/src/ggml-blas.o: \
+	ggml/src/ggml-blas.cpp \
+	ggml/include/ggml-blas.h
+	$(CXX) $(CXXFLAGS) -c $< -o $@
+
+ifdef GGML_LLAMAFILE
+ggml/src/sgemm.o: \
+	ggml/src/sgemm.cpp \
+	ggml/src/sgemm.h \
+	ggml/include/ggml.h
+	$(CXX) $(CXXFLAGS) -c $< -o $@
+endif # GGML_LLAMAFILE
+
+ifdef GGML_RPC
+ggml/src/ggml-rpc.o: \
+	ggml/src/ggml-rpc.cpp \
+	ggml/include/ggml-rpc.h
+	$(CXX) $(CXXFLAGS) -c $< -o $@
+endif # GGML_RPC
+
+$(LIB_GGML): \
+	$(OBJ_GGML)
+	$(CXX) $(CXXFLAGS) -shared -fPIC -o $@ $^ $(LDFLAGS)
+
+$(LIB_GGML_S): \
+	$(OBJ_GGML)
+	ar rcs $(LIB_GGML_S) $^
+
+# whisper
+
+src/whisper.o: \
+	src/whisper.cpp \
+	src/whisper-mel.hpp \
+	include/whisper.h \
+	ggml/include/ggml.h \
+	ggml/include/ggml-alloc.h \
+	ggml/include/ggml-backend.h \
+	ggml/include/ggml-cuda.h \
+	ggml/include/ggml-metal.h
+	$(CXX) $(CXXFLAGS) -c $< -o $@
+
+$(LIB_WHISPER): \
+	$(OBJ_WHISPER) \
+	$(LIB_GGML)
+	$(CXX) $(CXXFLAGS) -shared -fPIC -o $@ $^ $(LDFLAGS)
+
+$(LIB_WHISPER_S): \
+	$(OBJ_WHISPER) \
+	$(OBJ_GGML)
+	ar rcs $(LIB_WHISPER_S) $^
+
+# common
+
+examples/common.o: \
+	examples/common.cpp \
+	examples/common.h
+	$(CXX) $(CXXFLAGS) -c $< -o $@
+
+examples/common-ggml.o: \
+	examples/common-ggml.cpp \
+	examples/common-ggml.h
+	$(CXX) $(CXXFLAGS) -c $< -o $@
+
+$(LIB_COMMON): \
+	$(OBJ_COMMON)
+	$(CXX) $(CXXFLAGS) -shared -fPIC -o $@ $^ $(LDFLAGS)
+
+$(LIB_COMMON_S): \
+	$(OBJ_COMMON)
+	ar rcs $(LIB_COMMON_S) $^
+
+# common-sdl
+
+CFLAGS_SDL=$(shell sdl2-config --cflags)
+LDFLAGS_SDL=$(shell sdl2-config --libs)
+
+examples/common-sdl.o: \
+	examples/common-sdl.cpp \
+	examples/common-sdl.h
+	$(CXX) $(CXXFLAGS) $(CFLAGS_SDL) -c $< -o $@
+
+$(LIB_COMMON_SDL): \
+	$(OBJ_SDL)
+	$(CXX) $(CXXFLAGS) -shared -fPIC -o $@ $^ $(LDFLAGS) $(LDFLAGS_SDL)
+
+$(LIB_COMMON_SDL_S): \
+	$(OBJ_SDL)
+	ar rcs $(LIB_COMMON_SDL_S) $^
+
+clean:
+	rm -vrf *.dot $(BUILD_TARGETS) $(TEST_TARGETS)
+	rm -rvf src/*.o
+	rm -rvf src/coreml/*.o
+	rm -rvf tests/*.o
+	rm -rvf examples/*.o
+	rm -rvf *.a
+	rm -rvf *.dll
+	rm -rvf *.so
+	rm -rvf *.dot
+	rm -rvf ggml/*.a
+	rm -rvf ggml/*.dll
+	rm -rvf ggml/*.so
+	rm -vrf ggml/src/*.o
+	rm -vrf ggml/src/ggml-metal-embed.metal
+	rm -vrf ggml/src/ggml-cuda/*.o
+	rm -vrf ggml/src/ggml-cuda/template-instances/*.o
+	rm -rvf $(BUILD_TARGETS)
+	rm -rvf $(TEST_TARGETS)
+	find examples -type f -name "*.o" -delete
+
+#
+# Examples
+#
+
+# $< is the first prerequisite, i.e. the source file.
+# Explicitly compile this to an object file so that it can be cached with ccache.
+# The source file is then filtered out from $^ (the list of all prerequisites) and the object file is added instead.
+
+# Helper function that replaces .c, .cpp, and .cu file endings with .o:
+GET_OBJ_FILE = $(patsubst %.c,%.o,$(patsubst %.cpp,%.o,$(patsubst %.cu,%.o,$(1))))
+
+main: examples/main/main.cpp \
+	$(OBJ_GGML) $(OBJ_WHISPER) $(OBJ_COMMON)
+	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
+	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
+
+bench: examples/bench/bench.cpp \
+	$(OBJ_GGML) $(OBJ_WHISPER) $(OBJ_COMMON)
+	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
+	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
+
+quantize: examples/quantize/quantize.cpp \
+	$(OBJ_GGML) $(OBJ_WHISPER) $(OBJ_COMMON)
+	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
+	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
+
+server: examples/server/server.cpp \
+	$(OBJ_GGML) $(OBJ_WHISPER) $(OBJ_COMMON)
+	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
+	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS) $(LWINSOCK2)
+
+command: examples/command/command.cpp \
+	$(OBJ_GGML) $(OBJ_WHISPER) $(OBJ_COMMON) $(OBJ_SDL)
+	$(CXX) $(CXXFLAGS) $(CFLAGS_SDL) -c $< -o $(call GET_OBJ_FILE, $<)
+	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS) $(LDFLAGS_SDL)
+
+stream: examples/stream/stream.cpp \
+	$(OBJ_GGML) $(OBJ_WHISPER) $(OBJ_COMMON) $(OBJ_SDL)
+	$(CXX) $(CXXFLAGS) $(CFLAGS_SDL) -c $< -o $(call GET_OBJ_FILE, $<)
+	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS) $(LDFLAGS_SDL)
+
+lsp: examples/lsp/lsp.cpp \
+	$(OBJ_GGML) $(OBJ_WHISPER) $(OBJ_COMMON) $(OBJ_SDL)
+	$(CXX) $(CXXFLAGS) $(CFLAGS_SDL) -c $< -o $(call GET_OBJ_FILE, $<)
+	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS) $(LDFLAGS_SDL)
+
+# TODO: disabled until update
+#       https://github.com/ggerganov/whisper.cpp/issues/1818
+#talk: examples/talk/talk.cpp examples/talk/gpt-2.cpp \
+#	$(OBJ_GGML) $(OBJ_WHISPER) $(OBJ_COMMON) $(OBJ_SDL)
+#	$(CXX) $(CXXFLAGS) $(CFLAGS_SDL) -c $< -o $(call GET_OBJ_FILE, $<)
+#	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS) $(LDFLAGS_SDL)
+
+talk-llama: examples/talk-llama/talk-llama.cpp examples/talk-llama/llama.cpp examples/talk-llama/llama-vocab.cpp examples/talk-llama/llama-grammar.cpp examples/talk-llama/llama-sampling.cpp examples/talk-llama/unicode.cpp examples/talk-llama/unicode-data.cpp \
+	$(OBJ_GGML) $(OBJ_WHISPER) $(OBJ_COMMON) $(OBJ_SDL)
+	$(CXX) $(CXXFLAGS) $(CFLAGS_SDL) -c $< -o $(call GET_OBJ_FILE, $<)
+	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS) $(LDFLAGS_SDL)
+
+#
+# Tests
+#
+
+tests: $(TEST_TARGETS)
+
+tests/test-c.o: tests/test-c.c include/whisper.h
+	$(CC) $(CFLAGS) -c $(filter-out %.h,$^) -o $@
+
+#
+# Audio samples
+#
+
+# download a few audio samples into folder "./samples":
+.PHONY: samples
+samples:
+	@echo "Downloading samples..."
+	@mkdir -p samples
+	@wget --quiet --show-progress -O samples/gb0.ogg https://upload.wikimedia.org/wikipedia/commons/2/22/George_W._Bush%27s_weekly_radio_address_%28November_1%2C_2008%29.oga
+	@wget --quiet --show-progress -O samples/gb1.ogg https://upload.wikimedia.org/wikipedia/commons/1/1f/George_W_Bush_Columbia_FINAL.ogg
+	@wget --quiet --show-progress -O samples/hp0.ogg https://upload.wikimedia.org/wikipedia/en/d/d4/En.henryfphillips.ogg
+	@wget --quiet --show-progress -O samples/mm1.wav https://cdn.openai.com/whisper/draft-20220913a/micro-machines.wav
+	@wget --quiet --show-progress -O samples/a13.mp3 https://upload.wikimedia.org/wikipedia/commons/transcoded/6/6f/Apollo13-wehaveaproblem.ogg/Apollo13-wehaveaproblem.ogg.mp3
+	@wget --quiet --show-progress -O samples/diffusion2023-07-03.flac https://archive.org/download/diffusion2023-07-03/diffusion2023-07-03.flac
+	@echo "Converting to 16-bit WAV ..."
+	@ffmpeg -loglevel -0 -y -i samples/gb0.ogg -ar 16000 -ac 1 -c:a pcm_s16le samples/gb0.wav
+	@ffmpeg -loglevel -0 -y -i samples/gb1.ogg -ar 16000 -ac 1 -c:a pcm_s16le samples/gb1.wav
+	@ffmpeg -loglevel -0 -y -i samples/hp0.ogg -ar 16000 -ac 1 -c:a pcm_s16le samples/hp0.wav
+	@rm samples/*.ogg
+	@ffmpeg -loglevel -0 -y -i samples/mm1.wav -ar 16000 -ac 1 -c:a pcm_s16le samples/mm0.wav
+	@rm samples/mm1.wav
+	@ffmpeg -loglevel -0 -y -i samples/a13.mp3 -ar 16000 -ac 1 -c:a pcm_s16le -ss 00:00:00 -to 00:00:30 samples/a13.wav
+	@rm samples/a13.mp3
+	@ffmpeg -loglevel -0 -y -i samples/diffusion2023-07-03.flac -ar 16000 -ac 1 -c:a pcm_s16le samples/diffusion2023-07-03.wav
+	@rm samples/diffusion2023-07-03.flac
+
+#
+# Models
+#
+
+# if not already downloaded, the following targets download the specified model and
+# runs it on all samples in the folder "./samples":
+
+.PHONY: tiny.en
+.PHONY: tiny
+.PHONY: base.en
+.PHONY: base
+.PHONY: small.en
+.PHONY: small
+.PHONY: medium.en
+.PHONY: medium
+.PHONY: large-v1
+.PHONY: large-v2
+.PHONY: large-v3
+.PHONY: large-v3-turbo
+
+tiny.en tiny base.en base small.en small medium.en medium large-v1 large-v2 large-v3 large-v3-turbo: main
+	bash ./models/download-ggml-model.sh $@
+	@echo ""
+	@echo "==============================================="
+	@echo "Running $@ on all samples in ./samples ..."
+	@echo "==============================================="
+	@echo ""
+	@for f in samples/*.wav; do \
+		echo "----------------------------------------------" ; \
+		echo "[+] Running $@ on $$f ... (run 'ffplay $$f' to listen)" ; \
+	    echo "----------------------------------------------" ; \
+		echo "" ; \
+		./main -m models/ggml-$@.bin -f $$f ; \
+		echo "" ; \
+	done
diff --git a/src/whisper_cpp/examples/CMakeLists.txt b/src/whisper_cpp/examples/CMakeLists.txt
new file mode 100644
index 00000000..33e6249e
--- /dev/null
+++ b/src/whisper_cpp/examples/CMakeLists.txt
@@ -0,0 +1,148 @@
+# dependencies
+
+find_package(Threads REQUIRED)
+
+# third-party
+
+if (WHISPER_SDL2)
+    # SDL2
+    find_package(SDL2 REQUIRED)
+
+    string(STRIP "${SDL2_LIBRARIES}" SDL2_LIBRARIES)
+
+    message(STATUS "SDL2_INCLUDE_DIRS = ${SDL2_INCLUDE_DIRS}")
+    message(STATUS "SDL2_LIBRARIES    = ${SDL2_LIBRARIES}")
+endif()
+
+if (WHISPER_CLBLAST)
+    find_package(CLBlast REQUIRED)
+endif()
+
+# common
+
+set(TARGET common)
+
+unset(COMMON_EXTRA_LIBS)
+
+if (WHISPER_FFMPEG)
+    # As of cmake 3.27, there is no official cmake support for FindFFmpeg.
+    # Consequnelty we added a FindFFmpeg.cmake script the cmake subfolder:
+    # whisper.cpp does not need the full ffmpeg libs, just AVFORMAT AVCODEC AVUTIL SWRESAMPLE
+    # libswresample  performs highly optimized audio resampling, rematrixing and sample format conversion operations
+    # libavcodec provides a generic encoding/decoding framework and contains multiple decoders and encoders for audio, video and subtitle streams, and several bitstream filters.
+    # libavformat provides a generic framework for multiplexing and demultiplexing (muxing and demuxing) audio, video and subtitle streams.
+    find_package(FFmpeg REQUIRED)
+
+    if (NOT ${FFMPEG_FOUND})
+        message(FATAL_ERROR "Cannot find ffmpeg libs/headers")
+    endif()
+
+    message(STATUS "Found ffmpeg libs:       ${FFMPEG_LIBRARIES}")
+    message(STATUS "Found ffmpeg headers in: ${FFMPEG_INCLUDE_DIRS}")
+    message(STATUS "ffmpeg definitions:      ${FFMPEG_DEFINITIONS}")
+    message(STATUS "Found avformat           ${AVFORMAT_VERSION}")
+
+    include_directories(${FFMPEG_INCLUDE_DIRS})
+    add_compile_definitions(WHISPER_FFMPEG)
+
+    list(APPEND COMMON_EXTRA_LIBS ${FFMPEG_LIBRARIES})
+
+    set(COMMON_SOURCES_FFMPEG ffmpeg-transcode.cpp)
+endif()
+
+
+add_library(${TARGET} STATIC
+    common.h
+    common.cpp
+    common-ggml.h
+    common-ggml.cpp
+    grammar-parser.h
+    grammar-parser.cpp
+    ${COMMON_SOURCES_FFMPEG}
+    )
+
+include(DefaultTargetOptions)
+
+target_link_libraries(${TARGET} PRIVATE whisper ${COMMON_EXTRA_LIBS})
+
+set_target_properties(${TARGET} PROPERTIES POSITION_INDEPENDENT_CODE ON)
+set_target_properties(${TARGET} PROPERTIES FOLDER "libs")
+
+if (WHISPER_SDL2)
+    # common-sdl
+
+    set(TARGET common-sdl)
+
+    add_library(${TARGET} STATIC
+        common-sdl.h
+        common-sdl.cpp
+        )
+
+    include(DefaultTargetOptions)
+
+    target_include_directories(${TARGET} PUBLIC  ${SDL2_INCLUDE_DIRS})
+    target_link_libraries     (${TARGET} PRIVATE ${SDL2_LIBRARIES})
+
+    set_target_properties(${TARGET} PROPERTIES POSITION_INDEPENDENT_CODE ON)
+    set_target_properties(${TARGET} PROPERTIES FOLDER "libs")
+endif()
+
+# add json lib
+add_library(json_cpp INTERFACE)
+target_include_directories(json_cpp INTERFACE ${CMAKE_CURRENT_SOURCE_DIR})
+
+# examples
+
+include_directories(${CMAKE_CURRENT_SOURCE_DIR})
+
+if (EMSCRIPTEN)
+    add_subdirectory(whisper.wasm)
+    set_target_properties(libmain PROPERTIES FOLDER "libs")
+    add_subdirectory(stream.wasm)
+    set_target_properties(libstream PROPERTIES FOLDER "libs")
+    add_subdirectory(command.wasm)
+    set_target_properties(libcommand PROPERTIES FOLDER "libs")
+    #add_subdirectory(talk.wasm)
+    #set_target_properties(libtalk PROPERTIES FOLDER "libs")
+    add_subdirectory(bench.wasm)
+    set_target_properties(libbench PROPERTIES FOLDER "libs")
+elseif(CMAKE_JS_VERSION)
+    add_subdirectory(addon.node)
+    set_target_properties(addon.node PROPERTIES FOLDER "examples")
+else()
+    add_subdirectory(main)
+    set_target_properties(main PROPERTIES FOLDER "examples")
+if (WHISPER_SDL2)
+    add_subdirectory(stream)
+    set_target_properties(stream PROPERTIES FOLDER "examples")
+endif (WHISPER_SDL2)
+    add_subdirectory(server)
+    set_target_properties(server PROPERTIES FOLDER "examples")
+if (WHISPER_SDL2)
+    add_subdirectory(command)
+    set_target_properties(command PROPERTIES FOLDER "examples")
+endif (WHISPER_SDL2)
+    add_subdirectory(bench)
+    set_target_properties(bench PROPERTIES FOLDER "examples")
+    add_subdirectory(quantize)
+    set_target_properties(quantize PROPERTIES FOLDER "examples")
+if (WHISPER_SDL2)
+    # TODO: disabled until update
+    #       https://github.com/ggerganov/whisper.cpp/issues/1818
+    #add_subdirectory(talk)
+    #set_target_properties(talk PROPERTIES FOLDER "examples")
+    add_subdirectory(talk-llama)
+    set_target_properties(talk-llama PROPERTIES FOLDER "examples")
+    add_subdirectory(lsp)
+    set_target_properties(lsp PROPERTIES FOLDER "examples")
+    if (GGML_SYCL)
+        add_subdirectory(sycl)
+        set_target_properties(sycl PROPERTIES FOLDER "examples")
+    endif()
+endif (WHISPER_SDL2)
+endif()
+
+if (WHISPER_SDL2)
+    add_subdirectory(wchess)
+    set_target_properties(wchess PROPERTIES FOLDER "examples")
+endif (WHISPER_SDL2)
diff --git a/src/whisper_cpp/examples/bench/CMakeLists.txt b/src/whisper_cpp/examples/bench/CMakeLists.txt
new file mode 100644
index 00000000..f8a72ffd
--- /dev/null
+++ b/src/whisper_cpp/examples/bench/CMakeLists.txt
@@ -0,0 +1,6 @@
+set(TARGET bench)
+add_executable(${TARGET} bench.cpp)
+
+include(DefaultTargetOptions)
+
+target_link_libraries(${TARGET} PRIVATE whisper ${CMAKE_THREAD_LIBS_INIT})
diff --git a/src/whisper_cpp/examples/bench/README.md b/src/whisper_cpp/examples/bench/README.md
new file mode 100644
index 00000000..5b42cb4d
--- /dev/null
+++ b/src/whisper_cpp/examples/bench/README.md
@@ -0,0 +1,54 @@
+# bench
+
+A very basic tool for benchmarking the inference performance on your device. The tool simply runs the Encoder part of
+the transformer on some random audio data and records the execution time. This way we can have an objective comparison
+of the performance of the model for various setups.
+
+Benchmark results are tracked in the following Github issue: https://github.com/ggerganov/whisper.cpp/issues/89
+
+```bash
+# build the bench tool
+$ make bench
+
+# run it on the small.en model using 4 threads
+$ ./bench -m ./models/ggml-small.en.bin -t 4
+
+whisper_model_load: loading model from './models/ggml-small.en.bin'
+whisper_model_load: n_vocab       = 51864
+whisper_model_load: n_audio_ctx   = 1500
+whisper_model_load: n_audio_state = 768
+whisper_model_load: n_audio_head  = 12
+whisper_model_load: n_audio_layer = 12
+whisper_model_load: n_text_ctx    = 448
+whisper_model_load: n_text_state  = 768
+whisper_model_load: n_text_head   = 12
+whisper_model_load: n_text_layer  = 12
+whisper_model_load: n_mels        = 80
+whisper_model_load: f16           = 1
+whisper_model_load: type          = 3
+whisper_model_load: mem_required  = 1048.00 MB
+whisper_model_load: adding 1607 extra tokens
+whisper_model_load: ggml ctx size = 533.05 MB
+whisper_model_load: memory size =    68.48 MB 
+whisper_model_load: model size  =   464.44 MB
+
+whisper_print_timings:     load time =   240.82 ms
+whisper_print_timings:      mel time =     0.00 ms
+whisper_print_timings:   sample time =     0.00 ms
+whisper_print_timings:   encode time =  1062.21 ms / 88.52 ms per layer
+whisper_print_timings:   decode time =     0.00 ms / 0.00 ms per layer
+whisper_print_timings:    total time =  1303.04 ms
+
+system_info: n_threads = 4 | AVX2 = 0 | AVX512 = 0 | NEON = 1 | FP16_VA = 1 | WASM_SIMD = 0 | BLAS = 1 | 
+
+If you wish, you can submit these results here:
+
+  https://github.com/ggerganov/whisper.cpp/issues/89
+
+Please include the following information:
+
+  - CPU model
+  - Operating system
+  - Compiler
+
+```
diff --git a/src/whisper_cpp/examples/bench/bench.cpp b/src/whisper_cpp/examples/bench/bench.cpp
new file mode 100644
index 00000000..d19c8ac5
--- /dev/null
+++ b/src/whisper_cpp/examples/bench/bench.cpp
@@ -0,0 +1,175 @@
+#include "whisper.h"
+
+#include <cstdio>
+#include <cstring>
+#include <string>
+#include <thread>
+
+// command-line parameters
+struct whisper_params {
+    int32_t n_threads = std::min(4, (int32_t) std::thread::hardware_concurrency());
+    int32_t what = 0; // what to benchmark: 0 - whisper encoder, 1 - memcpy, 2 - ggml_mul_mat
+
+    std::string model = "models/ggml-base.en.bin";
+
+    bool use_gpu    = true;
+    bool flash_attn = false;
+};
+
+void whisper_print_usage(int argc, char ** argv, const whisper_params & params);
+
+static bool whisper_params_parse(int argc, char ** argv, whisper_params & params) {
+    for (int i = 1; i < argc; i++) {
+        std::string arg = argv[i];
+
+        if (arg == "-h" || arg == "--help") {
+            whisper_print_usage(argc, argv, params);
+            exit(0);
+        }
+        else if (arg == "-t"  || arg == "--threads")    { params.n_threads  = std::stoi(argv[++i]); }
+        else if (arg == "-m"  || arg == "--model")      { params.model      = argv[++i]; }
+        else if (arg == "-w"  || arg == "--what")       { params.what       = atoi(argv[++i]); }
+        else if (arg == "-ng" || arg == "--no-gpu")     { params.use_gpu    = false; }
+        else if (arg == "-fa" || arg == "--flash-attn") { params.flash_attn = true; }
+        else {
+            fprintf(stderr, "error: unknown argument: %s\n", arg.c_str());
+            whisper_print_usage(argc, argv, params);
+            exit(0);
+        }
+    }
+
+    return true;
+}
+
+void whisper_print_usage(int /*argc*/, char ** argv, const whisper_params & params) {
+    fprintf(stderr, "\n");
+    fprintf(stderr, "usage: %s [options]\n", argv[0]);
+    fprintf(stderr, "\n");
+    fprintf(stderr, "options:\n");
+    fprintf(stderr, "  -h,       --help        [default] show this help message and exit\n");
+    fprintf(stderr, "  -t N,     --threads N   [%-7d] number of threads to use during computation\n", params.n_threads);
+    fprintf(stderr, "  -m FNAME, --model FNAME [%-7s] model path\n",                                  params.model.c_str());
+    fprintf(stderr, "  -w N,     --what N      [%-7d] what to benchmark:\n",                          params.what);
+    fprintf(stderr, "  -ng,      --no-gpu      [%-7s] disable GPU\n",                                 params.use_gpu ? "false" : "true");
+    fprintf(stderr, "  -fa,      --flash-attn  [%-7s] enable flash attention\n",                      params.flash_attn ? "true" : "false");
+    fprintf(stderr, "                           %-7s  0 - whisper\n",                                 "");
+    fprintf(stderr, "                           %-7s  1 - memcpy\n",                                  "");
+    fprintf(stderr, "                           %-7s  2 - ggml_mul_mat\n",                            "");
+    fprintf(stderr, "\n");
+}
+
+static int whisper_bench_full(const whisper_params & params) {
+    // whisper init
+
+    struct whisper_context_params cparams = whisper_context_default_params();
+
+    cparams.use_gpu    = params.use_gpu;
+    cparams.flash_attn = params.flash_attn;
+
+    struct whisper_context * ctx = whisper_init_from_file_with_params(params.model.c_str(), cparams);
+
+    {
+        fprintf(stderr, "\n");
+        fprintf(stderr, "system_info: n_threads = %d / %d | %s\n", params.n_threads, std::thread::hardware_concurrency(), whisper_print_system_info());
+    }
+
+    if (ctx == nullptr) {
+        fprintf(stderr, "error: failed to initialize whisper context\n");
+        return 2;
+    }
+
+    const int n_mels = whisper_model_n_mels(ctx);
+
+    if (int ret = whisper_set_mel(ctx, nullptr, 0, n_mels)) {
+        fprintf(stderr, "error: failed to set mel: %d\n", ret);
+        return 3;
+    }
+    // heat encoder
+    if (int ret = whisper_encode(ctx, 0, params.n_threads) != 0) {
+        fprintf(stderr, "error: failed to encode: %d\n", ret);
+        return 4;
+    }
+
+    whisper_token tokens[512];
+    memset(tokens, 0, sizeof(tokens));
+
+    // prompt heat
+    if (int ret = whisper_decode(ctx, tokens, 256, 0, params.n_threads) != 0) {
+        fprintf(stderr, "error: failed to decode: %d\n", ret);
+        return 4;
+    }
+
+    // text-generation heat
+    if (int ret = whisper_decode(ctx, tokens, 1, 256, params.n_threads) != 0) {
+        fprintf(stderr, "error: failed to decode: %d\n", ret);
+        return 4;
+    }
+
+    whisper_reset_timings(ctx);
+
+    // actual run
+    if (int ret = whisper_encode(ctx, 0, params.n_threads) != 0) {
+        fprintf(stderr, "error: failed to encode: %d\n", ret);
+        return 4;
+    }
+
+    // text-generation
+    for (int i = 0; i < 256; i++) {
+        if (int ret = whisper_decode(ctx, tokens, 1, i, params.n_threads) != 0) {
+            fprintf(stderr, "error: failed to decode: %d\n", ret);
+            return 4;
+        }
+    }
+
+    // batched decoding
+    for (int i = 0; i < 64; i++) {
+        if (int ret = whisper_decode(ctx, tokens, 5, 0, params.n_threads) != 0) {
+            fprintf(stderr, "error: failed to decode: %d\n", ret);
+            return 4;
+        }
+    }
+
+    // prompt processing
+    for (int i = 0; i < 16; i++) {
+        if (int ret = whisper_decode(ctx, tokens, 256, 0, params.n_threads) != 0) {
+            fprintf(stderr, "error: failed to decode: %d\n", ret);
+            return 4;
+        }
+    }
+
+    whisper_print_timings(ctx);
+    whisper_free(ctx);
+
+    fprintf(stderr, "\n");
+    fprintf(stderr, "If you wish, you can submit these results here:\n");
+    fprintf(stderr, "\n");
+    fprintf(stderr, "  https://github.com/ggerganov/whisper.cpp/issues/89\n");
+    fprintf(stderr, "\n");
+    fprintf(stderr, "Please include the following information:\n");
+    fprintf(stderr, "\n");
+    fprintf(stderr, "  - CPU model\n");
+    fprintf(stderr, "  - Operating system\n");
+    fprintf(stderr, "  - Compiler\n");
+    fprintf(stderr, "\n");
+
+    return 0;
+}
+
+int main(int argc, char ** argv) {
+    whisper_params params;
+
+    if (whisper_params_parse(argc, argv, params) == false) {
+        return 1;
+    }
+
+    int ret = -1;
+
+    switch (params.what) {
+        case 0: ret = whisper_bench_full(params);                break;
+        case 1: ret = whisper_bench_memcpy(params.n_threads);       break;
+        case 2: ret = whisper_bench_ggml_mul_mat(params.n_threads); break;
+        default: fprintf(stderr, "error: unknown benchmark: %d\n", params.what); break;
+    }
+
+    return ret;
+}
diff --git a/src/whisper_cpp/examples/command/CMakeLists.txt b/src/whisper_cpp/examples/command/CMakeLists.txt
new file mode 100644
index 00000000..40f278c1
--- /dev/null
+++ b/src/whisper_cpp/examples/command/CMakeLists.txt
@@ -0,0 +1,9 @@
+if (WHISPER_SDL2)
+    # command
+    set(TARGET command)
+    add_executable(${TARGET} command.cpp)
+
+    include(DefaultTargetOptions)
+
+    target_link_libraries(${TARGET} PRIVATE common common-sdl whisper ${CMAKE_THREAD_LIBS_INIT})
+endif ()
diff --git a/src/whisper_cpp/examples/command/README.md b/src/whisper_cpp/examples/command/README.md
new file mode 100644
index 00000000..2cd27aa9
--- /dev/null
+++ b/src/whisper_cpp/examples/command/README.md
@@ -0,0 +1,51 @@
+# command
+
+This is a basic Voice Assistant example that accepts voice commands from the microphone.
+More info is available in [issue #171](https://github.com/ggerganov/whisper.cpp/issues/171).
+
+```bash
+# Run with default arguments and small model
+./command -m ./models/ggml-small.en.bin -t 8
+
+# On Raspberry Pi, use tiny or base models + "-ac 768" for better performance
+./command -m ./models/ggml-tiny.en.bin -ac 768 -t 3 -c 0
+```
+
+https://user-images.githubusercontent.com/1991296/204038393-2f846eae-c255-4099-a76d-5735c25c49da.mp4
+
+Web version: [examples/command.wasm](/examples/command.wasm)
+
+## Guided mode
+
+"Guided mode" allows you to specify a list of commands (i.e. strings) and the transcription will be guided to classify your command into one from the list. This can be useful in situations where a device is listening only for a small subset of commands.
+
+Initial tests show that this approach might be extremely efficient in terms of performance, since it integrates very well with the "partial Encoder" idea from #137.
+
+```bash
+# Run in guided mode, the list of allowed commands is in commands.txt
+./command -m ./models/ggml-base.en.bin -cmd ./examples/command/commands.txt
+
+# On Raspberry Pi, in guided mode you can use "-ac 128" for extra performance
+./command -m ./models/ggml-tiny.en.bin -cmd ./examples/command/commands.txt -ac 128 -t 3 -c 0
+```
+
+https://user-images.githubusercontent.com/1991296/207435352-8fc4ed3f-bde5-4555-9b8b-aeeb76bee969.mp4
+
+
+## Building
+
+The `command` tool depends on SDL2 library to capture audio from the microphone. You can build it like this:
+
+```bash
+# Install SDL2
+# On Debian based linux distributions:
+sudo apt-get install libsdl2-dev
+
+# On Fedora Linux:
+sudo dnf install SDL2 SDL2-devel
+
+# Install SDL2 on Mac OS
+brew install sdl2
+
+make command
+```
diff --git a/src/whisper_cpp/examples/command/command.cpp b/src/whisper_cpp/examples/command/command.cpp
new file mode 100644
index 00000000..11ed9ed6
--- /dev/null
+++ b/src/whisper_cpp/examples/command/command.cpp
@@ -0,0 +1,777 @@
+// Voice assistant example
+//
+// Speak short text commands to the microphone.
+// This program will detect your voice command and convert them to text.
+//
+// ref: https://github.com/ggerganov/whisper.cpp/issues/171
+//
+
+#include "common-sdl.h"
+#include "common.h"
+#include "whisper.h"
+#include "grammar-parser.h"
+
+#include <sstream>
+#include <cassert>
+#include <cstdio>
+#include <fstream>
+#include <mutex>
+#include <regex>
+#include <string>
+#include <thread>
+#include <vector>
+#include <map>
+
+// command-line parameters
+struct whisper_params {
+    int32_t n_threads  = std::min(4, (int32_t) std::thread::hardware_concurrency());
+    int32_t prompt_ms  = 5000;
+    int32_t command_ms = 8000;
+    int32_t capture_id = -1;
+    int32_t max_tokens = 32;
+    int32_t audio_ctx  = 0;
+
+    float vad_thold  = 0.6f;
+    float freq_thold = 100.0f;
+
+    float grammar_penalty = 100.0f;
+
+    grammar_parser::parse_state grammar_parsed;
+
+    bool translate     = false;
+    bool print_special = false;
+    bool print_energy  = false;
+    bool no_timestamps = true;
+    bool use_gpu       = true;
+    bool flash_attn    = false;
+
+    std::string language  = "en";
+    std::string model     = "models/ggml-base.en.bin";
+    std::string fname_out;
+    std::string commands;
+    std::string prompt;
+    std::string context;
+    std::string grammar;
+
+    // A regular expression that matches tokens to suppress
+    std::string suppress_regex;
+};
+
+void whisper_print_usage(int argc, char ** argv, const whisper_params & params);
+
+static bool whisper_params_parse(int argc, char ** argv, whisper_params & params) {
+    for (int i = 1; i < argc; i++) {
+        std::string arg = argv[i];
+
+        if (arg == "-h" || arg == "--help") {
+            whisper_print_usage(argc, argv, params);
+            exit(0);
+        }
+        else if (arg == "-t"   || arg == "--threads")       { params.n_threads     = std::stoi(argv[++i]); }
+        else if (arg == "-pms" || arg == "--prompt-ms")     { params.prompt_ms     = std::stoi(argv[++i]); }
+        else if (arg == "-cms" || arg == "--command-ms")    { params.command_ms    = std::stoi(argv[++i]); }
+        else if (arg == "-c"   || arg == "--capture")       { params.capture_id    = std::stoi(argv[++i]); }
+        else if (arg == "-mt"  || arg == "--max-tokens")    { params.max_tokens    = std::stoi(argv[++i]); }
+        else if (arg == "-ac"  || arg == "--audio-ctx")     { params.audio_ctx     = std::stoi(argv[++i]); }
+        else if (arg == "-vth" || arg == "--vad-thold")     { params.vad_thold     = std::stof(argv[++i]); }
+        else if (arg == "-fth" || arg == "--freq-thold")    { params.freq_thold    = std::stof(argv[++i]); }
+        else if (arg == "-tr"  || arg == "--translate")     { params.translate     = true; }
+        else if (arg == "-ps"  || arg == "--print-special") { params.print_special = true; }
+        else if (arg == "-pe"  || arg == "--print-energy")  { params.print_energy  = true; }
+        else if (arg == "-ng"  || arg == "--no-gpu")        { params.use_gpu       = false; }
+        else if (arg == "-fa"  || arg == "--flash-attn")    { params.flash_attn    = true; }
+        else if (arg == "-l"   || arg == "--language")      { params.language      = argv[++i]; }
+        else if (arg == "-m"   || arg == "--model")         { params.model         = argv[++i]; }
+        else if (arg == "-f"   || arg == "--file")          { params.fname_out     = argv[++i]; }
+        else if (arg == "-cmd" || arg == "--commands")      { params.commands      = argv[++i]; }
+        else if (arg == "-p"   || arg == "--prompt")        { params.prompt        = argv[++i]; }
+        else if (arg == "-ctx" || arg == "--context")       { params.context       = argv[++i]; }
+        else if (                 arg == "--grammar")       { params.grammar       = argv[++i]; }
+        else if (                 arg == "--grammar-penalty") { params.grammar_penalty = std::stof(argv[++i]); }
+        else if (                 arg == "--suppress-regex") { params.suppress_regex = argv[++i]; }
+        else {
+            fprintf(stderr, "error: unknown argument: %s\n", arg.c_str());
+            whisper_print_usage(argc, argv, params);
+            exit(0);
+        }
+    }
+
+    return true;
+}
+
+void whisper_print_usage(int /*argc*/, char ** argv, const whisper_params & params) {
+    fprintf(stderr, "\n");
+    fprintf(stderr, "usage: %s [options]\n", argv[0]);
+    fprintf(stderr, "\n");
+    fprintf(stderr, "options:\n");
+    fprintf(stderr, "  -h,         --help           [default] show this help message and exit\n");
+    fprintf(stderr, "  -t N,       --threads N      [%-7d] number of threads to use during computation\n", params.n_threads);
+    fprintf(stderr, "  -pms N,     --prompt-ms N    [%-7d] prompt duration in milliseconds\n",             params.prompt_ms);
+    fprintf(stderr, "  -cms N,     --command-ms N   [%-7d] command duration in milliseconds\n",            params.command_ms);
+    fprintf(stderr, "  -c ID,      --capture ID     [%-7d] capture device ID\n",                           params.capture_id);
+    fprintf(stderr, "  -mt N,      --max-tokens N   [%-7d] maximum number of tokens per audio chunk\n",    params.max_tokens);
+    fprintf(stderr, "  -ac N,      --audio-ctx N    [%-7d] audio context size (0 - all)\n",                params.audio_ctx);
+    fprintf(stderr, "  -vth N,     --vad-thold N    [%-7.2f] voice activity detection threshold\n",        params.vad_thold);
+    fprintf(stderr, "  -fth N,     --freq-thold N   [%-7.2f] high-pass frequency cutoff\n",                params.freq_thold);
+    fprintf(stderr, "  -tr,        --translate      [%-7s] translate from source language to english\n",   params.translate ? "true" : "false");
+    fprintf(stderr, "  -ps,        --print-special  [%-7s] print special tokens\n",                        params.print_special ? "true" : "false");
+    fprintf(stderr, "  -pe,        --print-energy   [%-7s] print sound energy (for debugging)\n",          params.print_energy ? "true" : "false");
+    fprintf(stderr, "  -ng,        --no-gpu         [%-7s] disable GPU\n",                                 params.use_gpu ? "false" : "true");
+    fprintf(stderr, "  -fa,        --flash-attn     [%-7s] flash attention\n",                             params.flash_attn ? "true" : "false");
+    fprintf(stderr, "  -l LANG,    --language LANG  [%-7s] spoken language\n",                             params.language.c_str());
+    fprintf(stderr, "  -m FNAME,   --model FNAME    [%-7s] model path\n",                                  params.model.c_str());
+    fprintf(stderr, "  -f FNAME,   --file FNAME     [%-7s] text output file name\n",                       params.fname_out.c_str());
+    fprintf(stderr, "  -cmd FNAME, --commands FNAME [%-7s] text file with allowed commands\n",             params.commands.c_str());
+    fprintf(stderr, "  -p,         --prompt         [%-7s] the required activation prompt\n",              params.prompt.c_str());
+    fprintf(stderr, "  -ctx,       --context        [%-7s] sample text to help the transcription\n",       params.context.c_str());
+    fprintf(stderr, "  --grammar GRAMMAR            [%-7s] GBNF grammar to guide decoding\n",              params.grammar.c_str());
+    fprintf(stderr, "  --grammar-penalty N          [%-7.1f] scales down logits of nongrammar tokens\n",   params.grammar_penalty);
+    fprintf(stderr, "  --suppress-regex REGEX       [%-7s] regular expression matching tokens to suppress\n", params.suppress_regex.c_str());
+    fprintf(stderr, "\n");
+}
+
+static std::string transcribe(
+                 whisper_context * ctx,
+            const whisper_params & params,
+        const std::vector<float> & pcmf32,
+               const std::string & grammar_rule,
+                           float & logprob_min,
+                           float & logprob_sum,
+                             int & n_tokens,
+                         int64_t & t_ms) {
+    const auto t_start = std::chrono::high_resolution_clock::now();
+
+    logprob_min = 0.0f;
+    logprob_sum = 0.0f;
+    n_tokens    = 0;
+    t_ms = 0;
+
+    //whisper_full_params wparams = whisper_full_default_params(WHISPER_SAMPLING_GREEDY);
+    whisper_full_params wparams = whisper_full_default_params(WHISPER_SAMPLING_BEAM_SEARCH);
+
+    wparams.print_progress   = false;
+    wparams.print_special    = params.print_special;
+    wparams.print_realtime   = false;
+    wparams.print_timestamps = !params.no_timestamps;
+    wparams.translate        = params.translate;
+    wparams.no_context       = true;
+    wparams.no_timestamps    = params.no_timestamps;
+    wparams.single_segment   = true;
+    wparams.max_tokens       = params.max_tokens;
+    wparams.language         = params.language.c_str();
+    wparams.n_threads        = params.n_threads;
+
+    wparams.audio_ctx = params.audio_ctx;
+
+    wparams.temperature     = 0.4f;
+    wparams.temperature_inc = 1.0f;
+    wparams.greedy.best_of  = 5;
+
+    wparams.beam_search.beam_size = 5;
+
+    wparams.initial_prompt = params.context.data();
+
+    wparams.suppress_regex = params.suppress_regex.c_str();
+
+    const auto & grammar_parsed = params.grammar_parsed;
+    auto grammar_rules = grammar_parsed.c_rules();
+
+    if (!params.grammar_parsed.rules.empty() && !grammar_rule.empty()) {
+        if (grammar_parsed.symbol_ids.find(grammar_rule) == grammar_parsed.symbol_ids.end()) {
+            fprintf(stderr, "%s: warning: grammar rule '%s' not found - skipping grammar sampling\n", __func__, grammar_rule.c_str());
+        } else {
+            wparams.grammar_rules   = grammar_rules.data();
+            wparams.n_grammar_rules = grammar_rules.size();
+            wparams.i_start_rule    = grammar_parsed.symbol_ids.at(grammar_rule);
+            wparams.grammar_penalty = params.grammar_penalty;
+        }
+    }
+
+    if (whisper_full(ctx, wparams, pcmf32.data(), pcmf32.size()) != 0) {
+        return "";
+    }
+
+    std::string result;
+
+    const int n_segments = whisper_full_n_segments(ctx);
+    for (int i = 0; i < n_segments; ++i) {
+        const char * text = whisper_full_get_segment_text(ctx, i);
+
+        result += text;
+
+        const int n = whisper_full_n_tokens(ctx, i);
+        for (int j = 0; j < n; ++j) {
+            const auto token = whisper_full_get_token_data(ctx, i, j);
+
+            if(token.plog > 0.0f) exit(0);
+            logprob_min = std::min(logprob_min, token.plog);
+            logprob_sum += token.plog;
+            ++n_tokens;
+        }
+    }
+
+    const auto t_end = std::chrono::high_resolution_clock::now();
+    t_ms = std::chrono::duration_cast<std::chrono::milliseconds>(t_end - t_start).count();
+
+    return result;
+}
+
+static std::vector<std::string> read_allowed_commands(const std::string & fname) {
+    std::vector<std::string> allowed_commands;
+
+    std::ifstream ifs(fname);
+    if (!ifs.is_open()) {
+        return allowed_commands;
+    }
+
+    std::string line;
+    while (std::getline(ifs, line)) {
+        line = ::trim(line);
+        if (line.empty()) {
+            continue;
+        }
+
+        std::transform(line.begin(), line.end(),line.begin(), ::tolower);
+        allowed_commands.push_back(std::move(line));
+    }
+
+    return allowed_commands;
+}
+
+static std::vector<std::string> get_words(const std::string &txt) {
+    std::vector<std::string> words;
+
+    std::istringstream iss(txt);
+    std::string word;
+    while (iss >> word) {
+        words.push_back(word);
+    }
+
+    return words;
+}
+
+// command-list mode
+// guide the transcription to match the most likely command from a provided list
+static int process_command_list(struct whisper_context * ctx, audio_async &audio, const whisper_params &params) {
+    fprintf(stderr, "\n");
+    fprintf(stderr, "%s: guided mode\n", __func__);
+
+    std::vector<std::string> allowed_commands = read_allowed_commands(params.commands);
+
+    if (allowed_commands.empty()) {
+        fprintf(stderr, "%s: error: failed to read allowed commands from '%s'\n", __func__, params.commands.c_str());
+        return 2;
+    }
+
+    int max_len = 0;
+
+    std::vector<std::vector<whisper_token>> allowed_tokens;
+
+    for (const auto & cmd : allowed_commands) {
+        whisper_token tokens[1024];
+        allowed_tokens.emplace_back();
+
+        for (int l = 0; l < (int) cmd.size(); ++l) {
+            // NOTE: very important to add the whitespace !
+            //       the reason is that the first decoded token starts with a whitespace too!
+            std::string ss = std::string(" ") + cmd.substr(0, l + 1);
+
+            const int n = whisper_tokenize(ctx, ss.c_str(), tokens, 1024);
+            if (n < 0) {
+                fprintf(stderr, "%s: error: failed to tokenize command '%s'\n", __func__, cmd.c_str());
+                return 3;
+            }
+
+            if (n == 1) {
+                allowed_tokens.back().push_back(tokens[0]);
+            }
+        }
+
+        max_len = std::max(max_len, (int) cmd.size());
+    }
+
+    fprintf(stderr, "%s: allowed commands [ tokens ]:\n", __func__);
+    fprintf(stderr, "\n");
+    for (int i = 0; i < (int) allowed_commands.size(); ++i) {
+        fprintf(stderr, "  - \033[1m%-*s\033[0m = [", max_len, allowed_commands[i].c_str());
+        for (const auto & token : allowed_tokens[i]) {
+            fprintf(stderr, " %5d", token);
+        }
+        fprintf(stderr, " ]\n");
+    }
+
+    std::string k_prompt = "select one from the available words: ";
+    for (int i = 0; i < (int) allowed_commands.size(); ++i) {
+        if (i > 0) {
+            k_prompt += ", ";
+        }
+        k_prompt += allowed_commands[i];
+    }
+    k_prompt += ". selected word: ";
+
+    // tokenize prompt
+    std::vector<whisper_token> k_tokens;
+    {
+        k_tokens.resize(1024);
+        const int n = whisper_tokenize(ctx, k_prompt.c_str(), k_tokens.data(), 1024);
+        if (n < 0) {
+            fprintf(stderr, "%s: error: failed to tokenize prompt '%s'\n", __func__, k_prompt.c_str());
+            return 4;
+        }
+        k_tokens.resize(n);
+    }
+
+    fprintf(stderr, "\n");
+    fprintf(stderr, "%s: prompt: '%s'\n", __func__, k_prompt.c_str());
+    fprintf(stderr, "%s: tokens: [", __func__);
+    for (const auto & token : k_tokens) {
+        fprintf(stderr, " %d", token);
+    }
+    fprintf(stderr, " ]\n");
+
+    fprintf(stderr, "\n");
+    fprintf(stderr, "%s: listening for a command ...\n", __func__);
+    fprintf(stderr, "\n");
+
+    bool is_running  = true;
+
+    std::vector<float> pcmf32_cur;
+    std::vector<float> pcmf32_prompt;
+
+    // main loop
+    while (is_running) {
+        // handle Ctrl + C
+        is_running = sdl_poll_events();
+
+        // delay
+        std::this_thread::sleep_for(std::chrono::milliseconds(100));
+
+        audio.get(2000, pcmf32_cur);
+
+        if (::vad_simple(pcmf32_cur, WHISPER_SAMPLE_RATE, 1000, params.vad_thold, params.freq_thold, params.print_energy)) {
+            fprintf(stdout, "%s: Speech detected! Processing ...\n", __func__);
+
+            const auto t_start = std::chrono::high_resolution_clock::now();
+
+            whisper_full_params wparams = whisper_full_default_params(WHISPER_SAMPLING_GREEDY);
+
+            wparams.print_progress   = false;
+            wparams.print_special    = params.print_special;
+            wparams.print_realtime   = false;
+            wparams.print_timestamps = !params.no_timestamps;
+            wparams.translate        = params.translate;
+            wparams.no_context       = true;
+            wparams.single_segment   = true;
+            wparams.max_tokens       = 1;
+            wparams.language         = params.language.c_str();
+            wparams.n_threads        = params.n_threads;
+
+            wparams.audio_ctx        = params.audio_ctx;
+
+            wparams.prompt_tokens    = k_tokens.data();
+            wparams.prompt_n_tokens  = k_tokens.size();
+
+            // run the transformer and a single decoding pass
+            if (whisper_full(ctx, wparams, pcmf32_cur.data(), pcmf32_cur.size()) != 0) {
+                fprintf(stderr, "%s: ERROR: whisper_full() failed\n", __func__);
+                break;
+            }
+
+            // estimate command probability
+            // NOTE: not optimal
+            {
+                const auto * logits = whisper_get_logits(ctx);
+
+                std::vector<float> probs(whisper_n_vocab(ctx), 0.0f);
+
+                // compute probs from logits via softmax
+                {
+                    float max = -1e9;
+                    for (int i = 0; i < (int) probs.size(); ++i) {
+                        max = std::max(max, logits[i]);
+                    }
+
+                    float sum = 0.0f;
+                    for (int i = 0; i < (int) probs.size(); ++i) {
+                        probs[i] = expf(logits[i] - max);
+                        sum += probs[i];
+                    }
+
+                    for (int i = 0; i < (int) probs.size(); ++i) {
+                        probs[i] /= sum;
+                    }
+                }
+
+                std::vector<std::pair<float, int>> probs_id;
+
+                double psum = 0.0;
+                for (int i = 0; i < (int) allowed_commands.size(); ++i) {
+                    probs_id.emplace_back(probs[allowed_tokens[i][0]], i);
+                    for (int j = 1; j < (int) allowed_tokens[i].size(); ++j) {
+                        probs_id.back().first += probs[allowed_tokens[i][j]];
+                    }
+                    probs_id.back().first /= allowed_tokens[i].size();
+                    psum += probs_id.back().first;
+                }
+
+                // normalize
+                for (auto & p : probs_id) {
+                    p.first /= psum;
+                }
+
+                // sort descending
+                {
+                    using pair_type = decltype(probs_id)::value_type;
+                    std::sort(probs_id.begin(), probs_id.end(), [](const pair_type & a, const pair_type & b) {
+                        return a.first > b.first;
+                    });
+                }
+
+                // print the commands and the respective probabilities
+                {
+                    fprintf(stdout, "\n");
+                    for (const auto & cmd : probs_id) {
+                        fprintf(stdout, "%s: %s%-*s%s = %f | ", __func__, "\033[1m", max_len, allowed_commands[cmd.second].c_str(), "\033[0m", cmd.first);
+                        for (int token : allowed_tokens[cmd.second]) {
+                            fprintf(stdout, "'%4s' %f ", whisper_token_to_str(ctx, token), probs[token]);
+                        }
+                        fprintf(stdout, "\n");
+                    }
+                }
+
+                // best command
+                {
+                    const auto t_end = std::chrono::high_resolution_clock::now();
+
+                    const float prob = probs_id[0].first;
+                    const int index = probs_id[0].second;
+
+                    fprintf(stdout, "\n");
+                    fprintf(stdout, "%s: detected command: %s%s%s | p = %f | t = %d ms\n", __func__,
+                            "\033[1m", allowed_commands[index].c_str(), "\033[0m", prob,
+                            (int) std::chrono::duration_cast<std::chrono::milliseconds>(t_end - t_start).count());
+                    fprintf(stdout, "\n");
+                }
+            }
+
+            audio.clear();
+        }
+    }
+
+    return 0;
+}
+
+// always-prompt mode
+// transcribe the voice into text after valid prompt
+static int always_prompt_transcription(struct whisper_context * ctx, audio_async & audio, const whisper_params & params) {
+    bool is_running = true;
+    bool ask_prompt = true;
+
+    float logprob_min = 0.0f;
+    float logprob_sum = 0.0f;
+    int   n_tokens    = 0;
+
+    std::vector<float> pcmf32_cur;
+
+    const std::string k_prompt = params.prompt;
+
+    const int k_prompt_length = get_words(k_prompt).size();
+
+    fprintf(stderr, "\n");
+    fprintf(stderr, "%s: always-prompt mode\n", __func__);
+
+    // main loop
+    while (is_running) {
+        // handle Ctrl + C
+        is_running = sdl_poll_events();
+
+        // delay
+        std::this_thread::sleep_for(std::chrono::milliseconds(100));
+
+        if (ask_prompt) {
+            fprintf(stdout, "\n");
+            fprintf(stdout, "%s: The prompt is: '%s%s%s'\n", __func__, "\033[1m", k_prompt.c_str(), "\033[0m");
+            fprintf(stdout, "\n");
+
+            ask_prompt = false;
+        }
+
+        {
+            audio.get(2000, pcmf32_cur);
+
+            if (::vad_simple(pcmf32_cur, WHISPER_SAMPLE_RATE, 1000, params.vad_thold, params.freq_thold, params.print_energy)) {
+                fprintf(stdout, "%s: Speech detected! Processing ...\n", __func__);
+
+                int64_t t_ms = 0;
+
+                // detect the commands
+                audio.get(params.command_ms, pcmf32_cur);
+
+                const auto txt = ::trim(::transcribe(ctx, params, pcmf32_cur, "", logprob_min, logprob_sum, n_tokens, t_ms));
+
+                const auto words = get_words(txt);
+
+                std::string prompt;
+                std::string command;
+
+                for (int i = 0; i < (int) words.size(); ++i) {
+                    if (i < k_prompt_length) {
+                        prompt += words[i] + " ";
+                    } else {
+                        command += words[i] + " ";
+                    }
+                }
+
+                const float sim = similarity(prompt, k_prompt);
+
+                //debug
+                //fprintf(stdout, "command size: %i\n", command_length);
+
+                if ((sim > 0.7f) && (command.size() > 0)) {
+                    fprintf(stdout, "%s: Command '%s%s%s', (t = %d ms)\n", __func__, "\033[1m", command.c_str(), "\033[0m", (int) t_ms);
+                }
+
+                fprintf(stdout, "\n");
+
+                audio.clear();
+            }
+        }
+    }
+
+    return 0;
+}
+
+// general-purpose mode
+// freely transcribe the voice into text
+static int process_general_transcription(struct whisper_context * ctx, audio_async & audio, const whisper_params & params) {
+    bool is_running  = true;
+    bool have_prompt = false;
+    bool ask_prompt  = true;
+
+    float logprob_min0 = 0.0f;
+    float logprob_min  = 0.0f;
+
+    float logprob_sum0 = 0.0f;
+    float logprob_sum  = 0.0f;
+
+    int n_tokens0 = 0;
+    int n_tokens  = 0;
+
+    std::vector<float> pcmf32_cur;
+    std::vector<float> pcmf32_prompt;
+
+    std::string k_prompt = "Ok Whisper, start listening for commands.";
+    if (!params.prompt.empty()) {
+        k_prompt = params.prompt;
+    }
+
+    fprintf(stderr, "\n");
+    fprintf(stderr, "%s: general-purpose mode\n", __func__);
+
+    // main loop
+    while (is_running) {
+        // handle Ctrl + C
+        is_running = sdl_poll_events();
+
+        // delay
+        std::this_thread::sleep_for(std::chrono::milliseconds(100));
+
+        if (ask_prompt) {
+            fprintf(stdout, "\n");
+            fprintf(stdout, "%s: Say the following phrase: '%s%s%s'\n", __func__, "\033[1m", k_prompt.c_str(), "\033[0m");
+            fprintf(stdout, "\n");
+
+            ask_prompt = false;
+        }
+
+        {
+            audio.get(2000, pcmf32_cur);
+
+            if (::vad_simple(pcmf32_cur, WHISPER_SAMPLE_RATE, 1000, params.vad_thold, params.freq_thold, params.print_energy)) {
+                fprintf(stdout, "%s: Speech detected! Processing ...\n", __func__);
+
+                int64_t t_ms = 0;
+
+                if (!have_prompt) {
+                    // wait for activation phrase
+                    audio.get(params.prompt_ms, pcmf32_cur);
+
+                    const auto txt = ::trim(::transcribe(ctx, params, pcmf32_cur, "prompt", logprob_min0, logprob_sum0, n_tokens0, t_ms));
+
+                    const float p = 100.0f * std::exp(logprob_min0);
+
+                    fprintf(stdout, "%s: Heard '%s%s%s', (t = %d ms, p = %.2f%%)\n", __func__, "\033[1m", txt.c_str(), "\033[0m", (int) t_ms, p);
+
+                    const float sim = similarity(txt, k_prompt);
+
+                    if (txt.length() < 0.8*k_prompt.length() || txt.length() > 1.2*k_prompt.length() || sim < 0.8f) {
+                        fprintf(stdout, "%s: WARNING: prompt not recognized, try again\n", __func__);
+                        ask_prompt = true;
+                    } else {
+                        fprintf(stdout, "\n");
+                        fprintf(stdout, "%s: The prompt has been recognized!\n", __func__);
+                        fprintf(stdout, "%s: Waiting for voice commands ...\n", __func__);
+                        fprintf(stdout, "\n");
+
+                        // save the audio for the prompt
+                        pcmf32_prompt = pcmf32_cur;
+                        have_prompt = true;
+                    }
+                } else {
+                    // we have heard the activation phrase, now detect the commands
+                    audio.get(params.command_ms, pcmf32_cur);
+
+                    //printf("len prompt:  %.4f\n", pcmf32_prompt.size() / (float) WHISPER_SAMPLE_RATE);
+                    //printf("len command: %.4f\n", pcmf32_cur.size() / (float) WHISPER_SAMPLE_RATE);
+
+                    // prepend 3 second of silence
+                    pcmf32_cur.insert(pcmf32_cur.begin(), 3.0f*WHISPER_SAMPLE_RATE, 0.0f);
+
+                    // prepend the prompt audio
+                    pcmf32_cur.insert(pcmf32_cur.begin(), pcmf32_prompt.begin(), pcmf32_prompt.end());
+
+                    const auto txt = ::trim(::transcribe(ctx, params, pcmf32_cur, "root", logprob_min, logprob_sum, n_tokens, t_ms));
+
+                    //const float p = 100.0f * std::exp((logprob - logprob0) / (n_tokens - n_tokens0));
+                    const float p = 100.0f * std::exp(logprob_min);
+
+                    //fprintf(stdout, "%s: heard '%s'\n", __func__, txt.c_str());
+
+                    // find the prompt in the text
+                    float best_sim = 0.0f;
+                    size_t best_len = 0;
+                    for (size_t n = 0.8*k_prompt.size(); n <= 1.2*k_prompt.size(); ++n) {
+                        if (n >= txt.size()) {
+                            break;
+                        }
+
+                        const auto prompt = txt.substr(0, n);
+
+                        const float sim = similarity(prompt, k_prompt);
+
+                        //fprintf(stderr, "%s: prompt = '%s', sim = %f\n", __func__, prompt.c_str(), sim);
+
+                        if (sim > best_sim) {
+                            best_sim = sim;
+                            best_len = n;
+                        }
+                    }
+
+                    fprintf(stdout, "%s:   DEBUG: txt = '%s', prob = %.2f%%\n", __func__, txt.c_str(), p);
+                    if (best_len == 0) {
+                        fprintf(stdout, "%s: WARNING: command not recognized, try again\n", __func__);
+                    } else {
+                        // cut the prompt from the decoded text
+                        const std::string command = ::trim(txt.substr(best_len));
+
+                        fprintf(stdout, "%s: Command '%s%s%s', (t = %d ms)\n", __func__, "\033[1m", command.c_str(), "\033[0m", (int) t_ms);
+                    }
+
+                    fprintf(stdout, "\n");
+                }
+
+                audio.clear();
+            }
+        }
+    }
+
+    return 0;
+}
+
+int main(int argc, char ** argv) {
+    whisper_params params;
+
+    if (whisper_params_parse(argc, argv, params) == false) {
+        return 1;
+    }
+
+    if (whisper_lang_id(params.language.c_str()) == -1) {
+        fprintf(stderr, "error: unknown language '%s'\n", params.language.c_str());
+        whisper_print_usage(argc, argv, params);
+        exit(0);
+    }
+
+    // whisper init
+
+    struct whisper_context_params cparams = whisper_context_default_params();
+
+    cparams.use_gpu    = params.use_gpu;
+    cparams.flash_attn = params.flash_attn;
+
+    struct whisper_context * ctx = whisper_init_from_file_with_params(params.model.c_str(), cparams);
+
+    // print some info about the processing
+    {
+        fprintf(stderr, "\n");
+        if (!whisper_is_multilingual(ctx)) {
+            if (params.language != "en" || params.translate) {
+                params.language = "en";
+                params.translate = false;
+                fprintf(stderr, "%s: WARNING: model is not multilingual, ignoring language and translation options\n", __func__);
+            }
+        }
+        fprintf(stderr, "%s: processing, %d threads, lang = %s, task = %s, timestamps = %d ...\n",
+                __func__,
+                params.n_threads,
+                params.language.c_str(),
+                params.translate ? "translate" : "transcribe",
+                params.no_timestamps ? 0 : 1);
+
+        fprintf(stderr, "\n");
+    }
+
+    // init audio
+
+    audio_async audio(30*1000);
+    if (!audio.init(params.capture_id, WHISPER_SAMPLE_RATE)) {
+        fprintf(stderr, "%s: audio.init() failed!\n", __func__);
+        return 1;
+    }
+
+    audio.resume();
+
+    // wait for 1 second to avoid any buffered noise
+    std::this_thread::sleep_for(std::chrono::milliseconds(1000));
+    audio.clear();
+
+    int  ret_val = 0;
+
+    if (!params.grammar.empty()) {
+        auto & grammar = params.grammar_parsed;
+        if (is_file_exist(params.grammar.c_str())) {
+            // read grammar from file
+            std::ifstream ifs(params.grammar.c_str());
+            const std::string txt = std::string((std::istreambuf_iterator<char>(ifs)), std::istreambuf_iterator<char>());
+            grammar = grammar_parser::parse(txt.c_str());
+        } else {
+            // read grammar from string
+            grammar = grammar_parser::parse(params.grammar.c_str());
+        }
+
+        // will be empty (default) if there are parse errors
+        if (grammar.rules.empty()) {
+            ret_val = 1;
+        } else {
+            fprintf(stderr, "%s: grammar:\n", __func__);
+            grammar_parser::print_grammar(stderr, grammar);
+            fprintf(stderr, "\n");
+        }
+    }
+
+    if (ret_val == 0) {
+        if (!params.commands.empty()) {
+            ret_val = process_command_list(ctx, audio, params);
+        } else if (!params.prompt.empty() && params.grammar_parsed.rules.empty()) {
+            ret_val = always_prompt_transcription(ctx, audio, params);
+        } else {
+            ret_val = process_general_transcription(ctx, audio, params);
+        }
+    }
+
+    audio.pause();
+
+    whisper_print_timings(ctx);
+    whisper_free(ctx);
+
+    return ret_val;
+}
diff --git a/src/whisper_cpp/examples/command/commands.txt b/src/whisper_cpp/examples/command/commands.txt
new file mode 100644
index 00000000..2653de65
--- /dev/null
+++ b/src/whisper_cpp/examples/command/commands.txt
@@ -0,0 +1,9 @@
+enable
+disable
+cat
+dog
+apple
+red
+blue
+green
+lightblue
diff --git a/src/whisper_cpp/common-ggml.cpp b/src/whisper_cpp/examples/common-ggml.cpp
similarity index 73%
rename from src/whisper_cpp/common-ggml.cpp
rename to src/whisper_cpp/examples/common-ggml.cpp
index aadded92..760cd1f4 100644
--- a/src/whisper_cpp/common-ggml.cpp
+++ b/src/whisper_cpp/examples/common-ggml.cpp
@@ -1,4 +1,3 @@
-#include <Rcpp.h>
 #include "common-ggml.h"
 
 #include <regex>
@@ -28,7 +27,7 @@ enum ggml_ftype ggml_parse_ftype(const char * str) {
     if (str[0] == 'q') {
         const auto it = GGML_FTYPE_MAP.find(str);
         if (it == GGML_FTYPE_MAP.end()) {
-            Rprintf("%s: unknown ftype '%s'\n", __func__, str);
+            fprintf(stderr, "%s: unknown ftype '%s'\n", __func__, str);
             return GGML_FTYPE_UNKNOWN;
         }
         ftype = it->second;
@@ -63,14 +62,27 @@ bool ggml_common_quantize_0(
         case GGML_FTYPE_ALL_F32:
         case GGML_FTYPE_MOSTLY_F16:
         case GGML_FTYPE_MOSTLY_Q4_1_SOME_F16:
+        case GGML_FTYPE_MOSTLY_IQ2_XXS:
+        case GGML_FTYPE_MOSTLY_IQ2_XS:
+        case GGML_FTYPE_MOSTLY_IQ2_S:
+        case GGML_FTYPE_MOSTLY_IQ3_XXS:
+        case GGML_FTYPE_MOSTLY_IQ3_S:
+        case GGML_FTYPE_MOSTLY_IQ1_S:
+        case GGML_FTYPE_MOSTLY_IQ4_NL:
+        case GGML_FTYPE_MOSTLY_IQ4_XS:
+        case GGML_FTYPE_MOSTLY_IQ1_M:
+        case GGML_FTYPE_MOSTLY_BF16:
+        case GGML_FTYPE_MOSTLY_Q4_0_4_4:
+        case GGML_FTYPE_MOSTLY_Q4_0_4_8:
+        case GGML_FTYPE_MOSTLY_Q4_0_8_8:
                 {
-                    Rprintf("%s: invalid model type %d\n", __func__, ftype);
+                    fprintf(stderr, "%s: invalid model type %d\n", __func__, ftype);
                     return false;
                 }
     };
 
     if (!ggml_is_quantized(qtype)) {
-        Rprintf("%s: invalid quantization type %d (%s)\n", __func__, qtype, ggml_type_name(qtype));
+        fprintf(stderr, "%s: invalid quantization type %d (%s)\n", __func__, qtype, ggml_type_name(qtype));
         return false;
     }
 
@@ -83,8 +95,6 @@ bool ggml_common_quantize_0(
     std::vector<ggml_fp16_t> data_f16;
     std::vector<float>       data_f32;
 
-    std::vector<int64_t> hist_all(1 << 4, 0);
-
     while (true) {
         int32_t n_dims;
         int32_t length;
@@ -108,7 +118,7 @@ bool ggml_common_quantize_0(
         std::string name(length, 0);
         finp.read (&name[0], length);
 
-        Rprintf("%64s - [%5d, %5d, %5d], type = %6s ", name.data(), ne[0], ne[1], ne[2], ggml_type_name((ggml_type) ttype));
+        printf("%64s - [%5d, %5d, %5d], type = %6s ", name.data(), ne[0], ne[1], ne[2], ggml_type_name((ggml_type) ttype));
 
         bool quantize = false;
 
@@ -133,7 +143,7 @@ bool ggml_common_quantize_0(
 
         if (quantize) {
             if (ttype != GGML_TYPE_F32 && ttype != GGML_TYPE_F16) {
-                Rprintf("%s: unsupported ttype %d (%s) for integer quantization\n", __func__, ttype, ggml_type_name((ggml_type) ttype));
+                fprintf(stderr, "%s: unsupported ttype %d (%s) for integer quantization\n", __func__, ttype, ggml_type_name((ggml_type) ttype));
                 return false;
             }
 
@@ -169,8 +179,6 @@ bool ggml_common_quantize_0(
             work.resize(nelements); // for quantization
 
             size_t cur_size = 0;
-            std::vector<int64_t> hist_cur(1 << 4, 0);
-
             switch ((ggml_type) ttype) {
                 case GGML_TYPE_Q4_0:
                 case GGML_TYPE_Q4_1:
@@ -183,18 +191,35 @@ bool ggml_common_quantize_0(
                 case GGML_TYPE_Q5_K:
                 case GGML_TYPE_Q6_K:
                     {
-                        cur_size = ggml_quantize_chunk((ggml_type) ttype, data_f32.data(), work.data(), 0, nelements, hist_cur.data());
+                        cur_size = ggml_quantize_chunk((ggml_type) ttype, data_f32.data(), work.data(), 0, nelements/ne[0], ne[0], nullptr);
                     } break;
                 case GGML_TYPE_F32:
                 case GGML_TYPE_F16:
                 case GGML_TYPE_I8:
                 case GGML_TYPE_I16:
                 case GGML_TYPE_I32:
+                case GGML_TYPE_I64:
+                case GGML_TYPE_F64:
                 case GGML_TYPE_Q8_1:
                 case GGML_TYPE_Q8_K:
+                case GGML_TYPE_IQ2_XXS:
+                case GGML_TYPE_IQ2_XS:
+                case GGML_TYPE_IQ2_S:
+                case GGML_TYPE_IQ3_XXS:
+                case GGML_TYPE_IQ3_S:
+                case GGML_TYPE_IQ1_S:
+                case GGML_TYPE_IQ4_NL:
+                case GGML_TYPE_IQ4_XS:
+                case GGML_TYPE_IQ1_M:
+                case GGML_TYPE_BF16:
+                case GGML_TYPE_Q4_0_4_4:
+                case GGML_TYPE_Q4_0_4_8:
+                case GGML_TYPE_Q4_0_8_8:
+                case GGML_TYPE_TQ1_0:
+                case GGML_TYPE_TQ2_0:
                 case GGML_TYPE_COUNT:
                     {
-                        Rprintf("%s: unsupported quantization type %d (%s)\n", __func__, ttype, ggml_type_name((ggml_type) ttype));
+                        fprintf(stderr, "%s: unsupported quantization type %d (%s)\n", __func__, ttype, ggml_type_name((ggml_type) ttype));
                         return false;
                     }
             }
@@ -202,17 +227,9 @@ bool ggml_common_quantize_0(
             fout.write(reinterpret_cast<char *>(work.data()), cur_size);
             total_size_new += cur_size;
 
-            Rprintf("size = %8.2f MB -> %8.2f MB | hist: ", nelements * sizeof(float)/1024.0/1024.0, cur_size/1024.0/1024.0);
-            for (int i = 0; i < (int) hist_cur.size(); ++i) {
-                hist_all[i] += hist_cur[i];
-            }
-
-            for (int i = 0; i < (int) hist_cur.size(); ++i) {
-                Rprintf("%5.3f ", hist_cur[i] / (float)nelements);
-            }
-            Rprintf("\n");
+            printf("size = %8.2f MB -> %8.2f MB\n", nelements * sizeof(float)/1024.0/1024.0, cur_size/1024.0/1024.0);
         } else {
-            Rprintf("size = %8.3f MB\n", data_u8.size()/1024.0/1024.0);
+            printf("size = %8.3f MB\n", data_u8.size()/1024.0/1024.0);
             fout.write(reinterpret_cast<char *>(data_u8.data()), data_u8.size());
             total_size_new += data_u8.size();
         }
@@ -220,21 +237,8 @@ bool ggml_common_quantize_0(
         total_size_org += nelements * sizeof(float);
     }
 
-    Rprintf("%s: model size  = %8.2f MB\n", __func__, total_size_org/1024.0/1024.0);
-    Rprintf("%s: quant size  = %8.2f MB | ftype = %d (%s)\n", __func__, total_size_new/1024.0/1024.0, ftype, ggml_type_name(qtype));
-
-    {
-        int64_t sum_all = 0;
-        for (int i = 0; i < (int) hist_all.size(); ++i) {
-            sum_all += hist_all[i];
-        }
-
-        Rprintf("%s: hist: ", __func__);
-        for (int i = 0; i < (int) hist_all.size(); ++i) {
-            Rprintf("%5.3f ", hist_all[i] / (float)sum_all);
-        }
-        Rprintf("\n");
-    }
+    printf("%s: model size  = %8.2f MB\n", __func__, total_size_org/1024.0/1024.0);
+    printf("%s: quant size  = %8.2f MB | ftype = %d (%s)\n", __func__, total_size_new/1024.0/1024.0, ftype, ggml_type_name(qtype));
 
     return true;
 }
diff --git a/src/whisper_cpp/common-ggml.h b/src/whisper_cpp/examples/common-ggml.h
similarity index 100%
rename from src/whisper_cpp/common-ggml.h
rename to src/whisper_cpp/examples/common-ggml.h
diff --git a/src/whisper_cpp/common-sdl.cpp b/src/whisper_cpp/examples/common-sdl.cpp
similarity index 99%
rename from src/whisper_cpp/common-sdl.cpp
rename to src/whisper_cpp/examples/common-sdl.cpp
index 5fc28e53..5329ec73 100644
--- a/src/whisper_cpp/common-sdl.cpp
+++ b/src/whisper_cpp/examples/common-sdl.cpp
@@ -219,7 +219,7 @@ bool sdl_poll_events() {
             case SDL_QUIT:
                 {
                     return false;
-                } break;
+                }
             default:
                 break;
         }
diff --git a/src/whisper_cpp/common-sdl.h b/src/whisper_cpp/examples/common-sdl.h
similarity index 100%
rename from src/whisper_cpp/common-sdl.h
rename to src/whisper_cpp/examples/common-sdl.h
diff --git a/src/whisper_cpp/common.cpp b/src/whisper_cpp/examples/common.cpp
similarity index 71%
rename from src/whisper_cpp/common.cpp
rename to src/whisper_cpp/examples/common.cpp
index 930635b6..11035736 100644
--- a/src/whisper_cpp/common.cpp
+++ b/src/whisper_cpp/examples/common.cpp
@@ -1,4 +1,3 @@
-#include <Rcpp.h>
 #define _USE_MATH_DEFINES // for M_PI
 
 #include "common.h"
@@ -20,14 +19,24 @@
 #pragma warning(disable: 4244 4267) // possible loss of data
 #endif
 
+#ifdef _WIN32
+#include <fcntl.h>
+#include <io.h>
+#endif
+
+#ifdef WHISPER_FFMPEG
+// as implemented in ffmpeg_trancode.cpp only embedded in common lib if whisper built with ffmpeg support
+extern bool ffmpeg_decode_audio(const std::string & ifname, std::vector<uint8_t> & wav_data);
+#endif
+
 // Function to check if the next argument exists
-std::string get_next_arg(int& i, int argc, char** argv, const std::string& flag, gpt_params& params) {
+static std::string get_next_arg(int& i, int argc, char** argv, const std::string& flag, gpt_params& params) {
     if (i + 1 < argc && argv[i + 1][0] != '-') {
         return argv[++i];
     } else {
-        Rprintf("error: %s requires one argument.\n", flag.c_str());
+        fprintf(stderr, "error: %s requires one argument.\n", flag.c_str());
         gpt_print_usage(argc, argv, params);
-        Rcpp::stop("whispercpp error");
+        exit(0);
     }
 }
 
@@ -72,12 +81,12 @@ bool gpt_params_parse(int argc, char ** argv, gpt_params & params) {
             params.interactive_port = std::stoi(get_next_arg(i, argc, argv, arg, params));
         } else if (arg == "-h" || arg == "--help") {
             gpt_print_usage(argc, argv, params);
-            Rcpp::stop("whispercpp error");
+            exit(0);
         } else if (arg == "-f" || arg == "--file") {
             get_next_arg(i, argc, argv, arg, params);
             std::ifstream file(argv[i]);
             if (!file) {
-                Rprintf("error: failed to open file '%s'\n", argv[i]);
+                fprintf(stderr, "error: failed to open file '%s'\n", argv[i]);
                 break;
             }
             std::copy(std::istreambuf_iterator<char>(file), std::istreambuf_iterator<char>(), back_inserter(params.prompt));
@@ -88,9 +97,9 @@ bool gpt_params_parse(int argc, char ** argv, gpt_params & params) {
             params.token_test = get_next_arg(i, argc, argv, arg, params);
         }
         else {
-            Rprintf("error: unknown argument: %s\n", arg.c_str());
+            fprintf(stderr, "error: unknown argument: %s\n", arg.c_str());
             gpt_print_usage(argc, argv, params);
-            Rcpp::stop("whispercpp error");
+            exit(0);
         }
     }
 
@@ -98,31 +107,31 @@ bool gpt_params_parse(int argc, char ** argv, gpt_params & params) {
 }
 
 void gpt_print_usage(int /*argc*/, char ** argv, const gpt_params & params) {
-    Rprintf("usage: %s [options]\n", argv[0]);
-    Rprintf("\n");
-    Rprintf("options:\n");
-    Rprintf("  -h, --help            show this help message and exit\n");
-    Rprintf("  -s SEED, --seed SEED  RNG seed (default: -1)\n");
-    Rprintf("  -t N, --threads N     number of threads to use during computation (default: %d)\n", params.n_threads);
-    Rprintf("  -p PROMPT, --prompt PROMPT\n");
-    Rprintf("                        prompt to start generation with (default: random)\n");
-    Rprintf("  -f FNAME, --file FNAME\n");
-    Rprintf("                        load prompt from a file\n");
-    Rprintf("  -tt TOKEN_TEST, --token_test TOKEN_TEST\n");
-    Rprintf("                        test tokenization\n");
-    Rprintf("  -n N, --n_predict N   number of tokens to predict (default: %d)\n", params.n_predict);
-    Rprintf("  --top_k N             top-k sampling (default: %d)\n", params.top_k);
-    Rprintf("  --top_p N             top-p sampling (default: %.1f)\n", params.top_p);
-    Rprintf("  --temp N              temperature (default: %.1f)\n", params.temp);
-    Rprintf("  --repeat-last-n N     last n tokens to consider for penalize (default: %d, 0 = disabled)\n", params.repeat_last_n);
-    Rprintf("  --repeat-penalty N    penalize repeat sequence of tokens (default: %.2f, 1.0 = disabled)\n", (double)params.repeat_penalty);
-    Rprintf("  -b N, --batch_size N  batch size for prompt processing (default: %d)\n", params.n_batch);
-    Rprintf("  -c N, --context N     context / KV cache size (default: %d)\n", params.n_ctx);
-    Rprintf("  --ignore-eos          ignore EOS token during generation\n");
-    Rprintf("  -ngl N, --gpu-layers N  number of layers to offload to GPU on supported models (default: %d)\n", params.n_gpu_layers);
-    Rprintf("  -m FNAME, --model FNAME\n");
-    Rprintf("                        model path (default: %s)\n", params.model.c_str());
-    Rprintf("\n");
+    fprintf(stderr, "usage: %s [options]\n", argv[0]);
+    fprintf(stderr, "\n");
+    fprintf(stderr, "options:\n");
+    fprintf(stderr, "  -h, --help            show this help message and exit\n");
+    fprintf(stderr, "  -s SEED, --seed SEED  RNG seed (default: -1)\n");
+    fprintf(stderr, "  -t N, --threads N     number of threads to use during computation (default: %d)\n", params.n_threads);
+    fprintf(stderr, "  -p PROMPT, --prompt PROMPT\n");
+    fprintf(stderr, "                        prompt to start generation with (default: random)\n");
+    fprintf(stderr, "  -f FNAME, --file FNAME\n");
+    fprintf(stderr, "                        load prompt from a file\n");
+    fprintf(stderr, "  -tt TOKEN_TEST, --token_test TOKEN_TEST\n");
+    fprintf(stderr, "                        test tokenization\n");
+    fprintf(stderr, "  -n N, --n_predict N   number of tokens to predict (default: %d)\n", params.n_predict);
+    fprintf(stderr, "  --top_k N             top-k sampling (default: %d)\n", params.top_k);
+    fprintf(stderr, "  --top_p N             top-p sampling (default: %.1f)\n", params.top_p);
+    fprintf(stderr, "  --temp N              temperature (default: %.1f)\n", params.temp);
+    fprintf(stderr, "  --repeat-last-n N     last n tokens to consider for penalize (default: %d, 0 = disabled)\n", params.repeat_last_n);
+    fprintf(stderr, "  --repeat-penalty N    penalize repeat sequence of tokens (default: %.2f, 1.0 = disabled)\n", (double)params.repeat_penalty);
+    fprintf(stderr, "  -b N, --batch_size N  batch size for prompt processing (default: %d)\n", params.n_batch);
+    fprintf(stderr, "  -c N, --context N     context / KV cache size (default: %d)\n", params.n_ctx);
+    fprintf(stderr, "  --ignore-eos          ignore EOS token during generation\n");
+    fprintf(stderr, "  -ngl N, --gpu-layers N  number of layers to offload to GPU on supported models (default: %d)\n", params.n_gpu_layers);
+    fprintf(stderr, "  -m FNAME, --model FNAME\n");
+    fprintf(stderr, "                        model path (default: %s)\n", params.model.c_str());
+    fprintf(stderr, "\n");
 }
 
 std::string gpt_random_prompt(std::mt19937 & rng) {
@@ -138,7 +147,6 @@ std::string gpt_random_prompt(std::mt19937 & rng) {
         case 7: return "He";
         case 8: return "She";
         case 9: return "They";
-        default: return "To";
     }
 
     return "The";
@@ -171,8 +179,8 @@ std::map<std::string, int32_t> json_parse(const std::string & fname) {
     {
         std::ifstream ifs(fname);
         if (!ifs) {
-            Rprintf("Failed to open %s\n", fname.c_str());
-            Rcpp::stop("whispercpp error");
+            fprintf(stderr, "Failed to open %s\n", fname.c_str());
+            exit(1);
         }
 
         json = std::string((std::istreambuf_iterator<char>(ifs)),
@@ -234,7 +242,7 @@ std::map<std::string, int32_t> json_parse(const std::string & fname) {
                     try {
                         result[str_key] = std::stoi(str_val);
                     } catch (...) {
-                        //Rprintf("%s: ignoring key '%s' with value '%s'\n", fname.c_str(), str_key.c_str(), str_val.c_str());
+                        //fprintf(stderr, "%s: ignoring key '%s' with value '%s'\n", fname.c_str(), str_key.c_str(), str_val.c_str());
 
                     }
                     str_key = "";
@@ -327,7 +335,7 @@ std::vector<gpt_vocab::id> gpt_tokenize(const gpt_vocab & vocab, const std::stri
                     break;
                 }
                 else if (j == i){ // word.substr(i, 1) has no matching
-                    Rprintf("%s: unknown token '%s'\n", __func__, word.substr(i, 1).data());
+                    fprintf(stderr, "%s: unknown token '%s'\n", __func__, word.substr(i, 1).data());
                     i++;
                 }
             }
@@ -337,7 +345,7 @@ std::vector<gpt_vocab::id> gpt_tokenize(const gpt_vocab & vocab, const std::stri
     return tokens;
 }
 
-std::vector<gpt_vocab::id> parse_tokens_from_string(const std::string& input, char delimiter) {
+static std::vector<gpt_vocab::id> parse_tokens_from_string(const std::string& input, char delimiter) {
     std::vector<gpt_vocab::id> output;
     std::stringstream ss(input);
     std::string token;
@@ -349,9 +357,9 @@ std::vector<gpt_vocab::id> parse_tokens_from_string(const std::string& input, ch
     return output;
 }
 
-std::map<std::string, std::vector<gpt_vocab::id>> extract_tests_from_file(const std::string & fpath_test){
+static std::map<std::string, std::vector<gpt_vocab::id>> extract_tests_from_file(const std::string & fpath_test){
     if (fpath_test.empty()){
-        Rprintf("%s : No test file found.\n", __func__);
+        fprintf(stderr, "%s : No test file found.\n", __func__);
         return std::map<std::string, std::vector<gpt_vocab::id>>();
     }
 
@@ -384,25 +392,25 @@ void test_gpt_tokenizer(gpt_vocab & vocab, const std::string & fpath_test){
             n_fails++;
 
             // print out failure cases
-            Rprintf("%s : failed test: '%s'\n", __func__, test.first.c_str());
-            Rprintf("%s : tokens in hf:   ", __func__);
+            fprintf(stderr, "%s : failed test: '%s'\n", __func__, test.first.c_str());
+            fprintf(stderr, "%s : tokens in hf:   ", __func__);
             for (const auto & t : test.second) {
-                Rprintf("%s(%d), ", vocab.id_to_token[t].c_str(), t);
+                fprintf(stderr, "%s(%d), ", vocab.id_to_token[t].c_str(), t);
             }
-            Rprintf("\n");
-            Rprintf("%s : tokens in ggml: ", __func__);
+            fprintf(stderr, "\n");
+            fprintf(stderr, "%s : tokens in ggml: ", __func__);
             for (const auto & t : tokens) {
-                Rprintf("%s(%d), ", vocab.id_to_token[t].c_str(), t);
+                fprintf(stderr, "%s(%d), ", vocab.id_to_token[t].c_str(), t);
             }
-            Rprintf("\n");
+            fprintf(stderr, "\n");
         }
     }
 
-    Rprintf("%s : %zu tests failed out of %zu tests.\n", __func__, n_fails, tests.size());
+    fprintf(stderr, "%s : %zu tests failed out of %zu tests.\n", __func__, n_fails, tests.size());
 }
 
 bool gpt_vocab_init(const std::string & fname, gpt_vocab & vocab) {
-    Rprintf("%s: loading vocab from '%s'\n", __func__, fname.c_str());
+    printf("%s: loading vocab from '%s'\n", __func__, fname.c_str());
 
     vocab.token_to_id = ::json_parse(fname);
 
@@ -410,11 +418,11 @@ bool gpt_vocab_init(const std::string & fname, gpt_vocab & vocab) {
         vocab.id_to_token[kv.second] = kv.first;
     }
 
-    Rprintf("%s: vocab size = %d\n", __func__, (int) vocab.token_to_id.size());
+    printf("%s: vocab size = %d\n", __func__, (int) vocab.token_to_id.size());
 
     // print the vocabulary
     //for (auto kv : vocab.token_to_id) {
-    //    Rprintf("'%s' -> %d\n", kv.first.data(), kv.second);
+    //    printf("'%s' -> %d\n", kv.first.data(), kv.second);
     //}
 
     return true;
@@ -490,9 +498,9 @@ gpt_vocab::id gpt_sample_top_k_top_p(
 
     //printf("\n");
     //for (int i = 0; i < (int) probs.size(); i++) {
-    //    Rprintf("%d: '%s' %f\n", i, vocab.id_to_token.at(logits_id[i].second).c_str(), probs[i]);
+    //    printf("%d: '%s' %f\n", i, vocab.id_to_token.at(logits_id[i].second).c_str(), probs[i]);
     //}
-    //Rcpp::stop("whispercpp error");
+    //exit(0);
 
     std::discrete_distribution<> dist(probs.begin(), probs.end());
     int idx = dist(rng);
@@ -603,10 +611,10 @@ gpt_vocab::id gpt_sample_top_k_top_p_repeat(
         }
     }
 
-//    Rprintf("\n");
+//    printf("\n");
 //    for (int i = 0; i < (int) probs.size(); i++) {
 //    for (int i = 0; i < 10; i++) {
-//        Rprintf("%d: '%s' %f\n", i, vocab.id_to_token.at(logits_id[i].second).c_str(), probs[i]);
+//        printf("%d: '%s' %f\n", i, vocab.id_to_token.at(logits_id[i].second).c_str(), probs[i]);
 //    }
 
     std::discrete_distribution<> dist(probs.begin(), probs.end());
@@ -616,12 +624,31 @@ gpt_vocab::id gpt_sample_top_k_top_p_repeat(
 
 }
 
+bool is_wav_buffer(const std::string buf) {
+    // RIFF ref: https://en.wikipedia.org/wiki/Resource_Interchange_File_Format
+    // WAV ref: https://www.mmsp.ece.mcgill.ca/Documents/AudioFormats/WAVE/WAVE.html
+    if (buf.size() < 12 || buf.substr(0, 4) != "RIFF" || buf.substr(8, 4) != "WAVE") {
+        return false;
+    }
+
+    uint32_t chunk_size = *reinterpret_cast<const uint32_t*>(buf.data() + 4);
+    if (chunk_size + 8 != buf.size()) {
+        return false;
+    }
+
+    return true;
+}
+
 bool read_wav(const std::string & fname, std::vector<float>& pcmf32, std::vector<std::vector<float>>& pcmf32s, bool stereo) {
     drwav wav;
-    std::vector<uint8_t> wav_data; // used for pipe input from stdin
+    std::vector<uint8_t> wav_data; // used for pipe input from stdin or ffmpeg decoding output
 
     if (fname == "-") {
         {
+            #ifdef _WIN32
+            _setmode(_fileno(stdin), _O_BINARY);
+            #endif
+
             uint8_t buf[1024];
             while (true)
             {
@@ -634,34 +661,55 @@ bool read_wav(const std::string & fname, std::vector<float>& pcmf32, std::vector
         }
 
         if (drwav_init_memory(&wav, wav_data.data(), wav_data.size(), nullptr) == false) {
-            Rprintf("error: failed to open WAV file from stdin\n");
+            fprintf(stderr, "error: failed to open WAV file from stdin\n");
             return false;
         }
 
-        Rprintf("%s: read %zu bytes from stdin\n", __func__, wav_data.size());
+        fprintf(stderr, "%s: read %zu bytes from stdin\n", __func__, wav_data.size());
+    }
+    else if (is_wav_buffer(fname)) {
+        if (drwav_init_memory(&wav, fname.c_str(), fname.size(), nullptr) == false) {
+            fprintf(stderr, "error: failed to open WAV file from fname buffer\n");
+            return false;
+        }
     }
     else if (drwav_init_file(&wav, fname.c_str(), nullptr) == false) {
-        Rprintf("error: failed to open '%s' as WAV file\n", fname.c_str());
+#if defined(WHISPER_FFMPEG)
+        if (ffmpeg_decode_audio(fname, wav_data) != 0) {
+            fprintf(stderr, "error: failed to ffmpeg decode '%s' \n", fname.c_str());
+            return false;
+        }
+        if (drwav_init_memory(&wav, wav_data.data(), wav_data.size(), nullptr) == false) {
+            fprintf(stderr, "error: failed to read wav data as wav \n");
+            return false;
+        }
+#else
+        fprintf(stderr, "error: failed to open '%s' as WAV file\n", fname.c_str());
         return false;
+#endif
     }
 
     if (wav.channels != 1 && wav.channels != 2) {
-        Rprintf("%s: WAV file '%s' must be mono or stereo\n", __func__, fname.c_str());
+        fprintf(stderr, "%s: WAV file '%s' must be mono or stereo\n", __func__, fname.c_str());
+        drwav_uninit(&wav);
         return false;
     }
 
     if (stereo && wav.channels != 2) {
-        Rprintf("%s: WAV file '%s' must be stereo for diarization\n", __func__, fname.c_str());
+        fprintf(stderr, "%s: WAV file '%s' must be stereo for diarization\n", __func__, fname.c_str());
+        drwav_uninit(&wav);
         return false;
     }
 
     if (wav.sampleRate != COMMON_SAMPLE_RATE) {
-        Rprintf("%s: WAV file '%s' must be %i kHz\n", __func__, fname.c_str(), COMMON_SAMPLE_RATE/1000);
+        fprintf(stderr, "%s: WAV file '%s' must be %i kHz\n", __func__, fname.c_str(), COMMON_SAMPLE_RATE/1000);
+        drwav_uninit(&wav);
         return false;
     }
 
     if (wav.bitsPerSample != 16) {
-        Rprintf("%s: WAV file '%s' must be 16-bit\n", __func__, fname.c_str());
+        fprintf(stderr, "%s: WAV file '%s' must be 16-bit\n", __func__, fname.c_str());
+        drwav_uninit(&wav);
         return false;
     }
 
@@ -740,7 +788,7 @@ bool vad_simple(std::vector<float> & pcmf32, int sample_rate, int last_ms, float
     energy_last /= n_samples_last;
 
     if (verbose) {
-        Rprintf("%s: energy_all: %f, energy_last: %f, vad_thold: %f, freq_thold: %f\n", __func__, energy_all, energy_last, vad_thold, freq_thold);
+        fprintf(stderr, "%s: energy_all: %f, energy_last: %f, vad_thold: %f, freq_thold: %f\n", __func__, energy_all, energy_last, vad_thold, freq_thold);
     }
 
     if (energy_last > vad_thold*energy_all) {
@@ -790,11 +838,11 @@ bool sam_params_parse(int argc, char ** argv, sam_params & params) {
             params.fname_out = argv[++i];
         } else if (arg == "-h" || arg == "--help") {
             sam_print_usage(argc, argv, params);
-            Rcpp::stop("whispercpp error");
+            exit(0);
         } else {
-            Rprintf("error: unknown argument: %s\n", arg.c_str());
+            fprintf(stderr, "error: unknown argument: %s\n", arg.c_str());
             sam_print_usage(argc, argv, params);
-            Rcpp::stop("whispercpp error");
+            exit(0);
         }
     }
 
@@ -802,17 +850,62 @@ bool sam_params_parse(int argc, char ** argv, sam_params & params) {
 }
 
 void sam_print_usage(int /*argc*/, char ** argv, const sam_params & params) {
-    Rprintf("usage: %s [options]\n", argv[0]);
-    Rprintf("\n");
-    Rprintf("options:\n");
-    Rprintf("  -h, --help            show this help message and exit\n");
-    Rprintf("  -s SEED, --seed SEED  RNG seed (default: -1)\n");
-    Rprintf("  -t N, --threads N     number of threads to use during computation (default: %d)\n", params.n_threads);
-    Rprintf("  -m FNAME, --model FNAME\n");
-    Rprintf("                        model path (default: %s)\n", params.model.c_str());
-    Rprintf("  -i FNAME, --inp FNAME\n");
-    Rprintf("                        input file (default: %s)\n", params.fname_inp.c_str());
-    Rprintf("  -o FNAME, --out FNAME\n");
-    Rprintf("                        output file (default: %s)\n", params.fname_out.c_str());
-    Rprintf("\n");
+    fprintf(stderr, "usage: %s [options]\n", argv[0]);
+    fprintf(stderr, "\n");
+    fprintf(stderr, "options:\n");
+    fprintf(stderr, "  -h, --help            show this help message and exit\n");
+    fprintf(stderr, "  -s SEED, --seed SEED  RNG seed (default: -1)\n");
+    fprintf(stderr, "  -t N, --threads N     number of threads to use during computation (default: %d)\n", params.n_threads);
+    fprintf(stderr, "  -m FNAME, --model FNAME\n");
+    fprintf(stderr, "                        model path (default: %s)\n", params.model.c_str());
+    fprintf(stderr, "  -i FNAME, --inp FNAME\n");
+    fprintf(stderr, "                        input file (default: %s)\n", params.fname_inp.c_str());
+    fprintf(stderr, "  -o FNAME, --out FNAME\n");
+    fprintf(stderr, "                        output file (default: %s)\n", params.fname_out.c_str());
+    fprintf(stderr, "\n");
+}
+
+//  500 -> 00:05.000
+// 6000 -> 01:00.000
+std::string to_timestamp(int64_t t, bool comma) {
+    int64_t msec = t * 10;
+    int64_t hr = msec / (1000 * 60 * 60);
+    msec = msec - hr * (1000 * 60 * 60);
+    int64_t min = msec / (1000 * 60);
+    msec = msec - min * (1000 * 60);
+    int64_t sec = msec / 1000;
+    msec = msec - sec * 1000;
+
+    char buf[32];
+    snprintf(buf, sizeof(buf), "%02d:%02d:%02d%s%03d", (int) hr, (int) min, (int) sec, comma ? "," : ".", (int) msec);
+
+    return std::string(buf);
+}
+
+int timestamp_to_sample(int64_t t, int n_samples, int whisper_sample_rate) {
+    return std::max(0, std::min((int) n_samples - 1, (int) ((t*whisper_sample_rate)/100)));
+}
+
+bool is_file_exist(const char *fileName)
+{
+    std::ifstream infile(fileName);
+    return infile.good();
+}
+
+bool speak_with_file(const std::string & command, const std::string & text, const std::string & path, int voice_id)
+{
+    std::ofstream speak_file(path.c_str());
+    if (speak_file.fail()) {
+        fprintf(stderr, "%s: failed to open speak_file\n", __func__);
+        return false;
+    } else {
+        speak_file.write(text.c_str(), text.size());
+        speak_file.close();
+        int ret = system((command + " " + std::to_string(voice_id) + " " + path).c_str());
+        if (ret != 0) {
+            fprintf(stderr, "%s: failed to speak\n", __func__);
+            return false;
+        }
+    }
+    return true;
 }
diff --git a/src/whisper_cpp/common.h b/src/whisper_cpp/examples/common.h
similarity index 79%
rename from src/whisper_cpp/common.h
rename to src/whisper_cpp/examples/common.h
index 54f0b00d..8c514c9f 100644
--- a/src/whisper_cpp/common.h
+++ b/src/whisper_cpp/examples/common.h
@@ -9,6 +9,7 @@
 #include <thread>
 #include <ctime>
 #include <fstream>
+#include <sstream>
 
 #define COMMON_SAMPLE_RATE 16000
 
@@ -21,7 +22,7 @@ struct gpt_params {
     int32_t n_threads    = std::min(4, (int32_t) std::thread::hardware_concurrency());
     int32_t n_predict    = 200;  // new tokens to predict
     int32_t n_parallel   = 1;    // number of parallel streams
-    int32_t n_batch      = 8;    // batch size for prompt processing
+    int32_t n_batch      = 32;   // batch size for prompt processing
     int32_t n_ctx        = 2048; // context size (this is the KV cache max size)
     int32_t n_gpu_layers = 0;    // number of layers to offlload to the GPU
 
@@ -135,7 +136,11 @@ gpt_vocab::id gpt_sample_top_k_top_p_repeat(
 // Audio utils
 //
 
+// Check if a buffer is a WAV audio file
+bool is_wav_buffer(const std::string buf);
+
 // Read WAV audio file and store the PCM data into pcmf32
+// fname can be a buffer of WAV data instead of a filename
 // The sample rate of the audio must be equal to COMMON_SAMPLE_RATE
 // If stereo flag is set and the audio has 2 channels, the pcmf32s will contain 2 channel PCM
 bool read_wav(
@@ -181,7 +186,7 @@ class wav_writer {
     // It is assumed that PCM data is normalized to a range from -1 to 1
     bool write_audio(const float * data, size_t length) {
         for (size_t i = 0; i < length; ++i) {
-            const int16_t intSample = data[i] * 32767;
+            const int16_t intSample = int16_t(data[i] * 32767);
             file.write(reinterpret_cast<const char *>(&intSample), sizeof(int16_t));
             dataSize += sizeof(int16_t);
         }
@@ -277,3 +282,62 @@ struct sam_params {
 bool sam_params_parse(int argc, char ** argv, sam_params & params);
 
 void sam_print_usage(int argc, char ** argv, const sam_params & params);
+
+//
+// Terminal utils
+//
+
+#define SQR(X)    ((X) * (X))
+#define UNCUBE(x) x < 48 ? 0 : x < 115 ? 1 : (x - 35) / 40
+
+/**
+ * Quantizes 24-bit RGB to xterm256 code range [16,256).
+ */
+static int rgb2xterm256(int r, int g, int b) {
+    unsigned char cube[] = {0, 0137, 0207, 0257, 0327, 0377};
+    int av, ir, ig, ib, il, qr, qg, qb, ql;
+    av = r * .299 + g * .587 + b * .114 + .5;
+    ql = (il = av > 238 ? 23 : (av - 3) / 10) * 10 + 8;
+    qr = cube[(ir = UNCUBE(r))];
+    qg = cube[(ig = UNCUBE(g))];
+    qb = cube[(ib = UNCUBE(b))];
+    if (SQR(qr - r) + SQR(qg - g) + SQR(qb - b) <=
+        SQR(ql - r) + SQR(ql - g) + SQR(ql - b))
+        return ir * 36 + ig * 6 + ib + 020;
+    return il + 0350;
+}
+
+static std::string set_xterm256_foreground(int r, int g, int b) {
+    int x = rgb2xterm256(r, g, b);
+    std::ostringstream oss;
+    oss << "\033[38;5;" << x << "m";
+    return oss.str();
+}
+
+// Lowest is red, middle is yellow, highest is green. Color scheme from
+// Paul Tol; it is colorblind friendly https://personal.sron.nl/~pault/
+const std::vector<std::string> k_colors = {
+    set_xterm256_foreground(220,   5,  12),
+    set_xterm256_foreground(232,  96,  28),
+    set_xterm256_foreground(241, 147,  45),
+    set_xterm256_foreground(246, 193,  65),
+    set_xterm256_foreground(247, 240,  86),
+    set_xterm256_foreground(144, 201, 135),
+    set_xterm256_foreground( 78, 178, 101),
+};
+
+//
+// Other utils
+//
+
+// convert timestamp to string, 6000 -> 01:00.000
+std::string to_timestamp(int64_t t, bool comma = false);
+
+// given a timestamp get the sample
+int timestamp_to_sample(int64_t t, int n_samples, int whisper_sample_rate);
+
+// check if file exists using ifstream
+bool is_file_exist(const char *fileName);
+
+// write text to file, and call system("command voice_id file")
+bool speak_with_file(const std::string & command, const std::string & text, const std::string & path, int voice_id);
diff --git a/src/whisper_cpp/examples/dr_wav.h b/src/whisper_cpp/examples/dr_wav.h
new file mode 100644
index 00000000..aabeff69
--- /dev/null
+++ b/src/whisper_cpp/examples/dr_wav.h
@@ -0,0 +1,8815 @@
+/*
+WAV audio loader and writer. Choice of public domain or MIT-0. See license statements at the end of this file.
+dr_wav - v0.13.16 - 2024-02-27
+
+David Reid - mackron@gmail.com
+
+GitHub: https://github.com/mackron/dr_libs
+*/
+
+/*
+Introduction
+============
+This is a single file library. To use it, do something like the following in one .c file.
+
+    ```c
+    #define DR_WAV_IMPLEMENTATION
+    #include "dr_wav.h"
+    ```
+
+You can then #include this file in other parts of the program as you would with any other header file. Do something like the following to read audio data:
+
+    ```c
+    drwav wav;
+    if (!drwav_init_file(&wav, "my_song.wav", NULL)) {
+        // Error opening WAV file.
+    }
+
+    drwav_int32* pDecodedInterleavedPCMFrames = malloc(wav.totalPCMFrameCount * wav.channels * sizeof(drwav_int32));
+    size_t numberOfSamplesActuallyDecoded = drwav_read_pcm_frames_s32(&wav, wav.totalPCMFrameCount, pDecodedInterleavedPCMFrames);
+
+    ...
+
+    drwav_uninit(&wav);
+    ```
+
+If you just want to quickly open and read the audio data in a single operation you can do something like this:
+
+    ```c
+    unsigned int channels;
+    unsigned int sampleRate;
+    drwav_uint64 totalPCMFrameCount;
+    float* pSampleData = drwav_open_file_and_read_pcm_frames_f32("my_song.wav", &channels, &sampleRate, &totalPCMFrameCount, NULL);
+    if (pSampleData == NULL) {
+        // Error opening and reading WAV file.
+    }
+
+    ...
+
+    drwav_free(pSampleData, NULL);
+    ```
+
+The examples above use versions of the API that convert the audio data to a consistent format (32-bit signed PCM, in this case), but you can still output the
+audio data in its internal format (see notes below for supported formats):
+
+    ```c
+    size_t framesRead = drwav_read_pcm_frames(&wav, wav.totalPCMFrameCount, pDecodedInterleavedPCMFrames);
+    ```
+
+You can also read the raw bytes of audio data, which could be useful if dr_wav does not have native support for a particular data format:
+
+    ```c
+    size_t bytesRead = drwav_read_raw(&wav, bytesToRead, pRawDataBuffer);
+    ```
+
+dr_wav can also be used to output WAV files. This does not currently support compressed formats. To use this, look at `drwav_init_write()`,
+`drwav_init_file_write()`, etc. Use `drwav_write_pcm_frames()` to write samples, or `drwav_write_raw()` to write raw data in the "data" chunk.
+
+    ```c
+    drwav_data_format format;
+    format.container = drwav_container_riff;     // <-- drwav_container_riff = normal WAV files, drwav_container_w64 = Sony Wave64.
+    format.format = DR_WAVE_FORMAT_PCM;          // <-- Any of the DR_WAVE_FORMAT_* codes.
+    format.channels = 2;
+    format.sampleRate = 44100;
+    format.bitsPerSample = 16;
+    drwav_init_file_write(&wav, "data/recording.wav", &format, NULL);
+
+    ...
+
+    drwav_uint64 framesWritten = drwav_write_pcm_frames(pWav, frameCount, pSamples);
+    ```
+
+Note that writing to AIFF or RIFX is not supported.
+
+dr_wav has support for decoding from a number of different encapsulation formats. See below for details.
+
+
+Build Options
+=============
+#define these options before including this file.
+
+#define DR_WAV_NO_CONVERSION_API
+  Disables conversion APIs such as `drwav_read_pcm_frames_f32()` and `drwav_s16_to_f32()`.
+
+#define DR_WAV_NO_STDIO
+  Disables APIs that initialize a decoder from a file such as `drwav_init_file()`, `drwav_init_file_write()`, etc.
+
+#define DR_WAV_NO_WCHAR
+  Disables all functions ending with `_w`. Use this if your compiler does not provide wchar.h. Not required if DR_WAV_NO_STDIO is also defined.
+
+
+Supported Encapsulations
+========================
+- RIFF (Regular WAV)
+- RIFX (Big-Endian)
+- AIFF (Does not currently support ADPCM)
+- RF64
+- W64
+
+Note that AIFF and RIFX do not support write mode, nor do they support reading of metadata.
+
+
+Supported Encodings
+===================
+- Unsigned 8-bit PCM
+- Signed 12-bit PCM
+- Signed 16-bit PCM
+- Signed 24-bit PCM
+- Signed 32-bit PCM
+- IEEE 32-bit floating point
+- IEEE 64-bit floating point
+- A-law and u-law
+- Microsoft ADPCM
+- IMA ADPCM (DVI, format code 0x11)
+
+8-bit PCM encodings are always assumed to be unsigned. Signed 8-bit encoding can only be read with `drwav_read_raw()`.
+
+Note that ADPCM is not currently supported with AIFF. Contributions welcome.
+
+
+Notes
+=====
+- Samples are always interleaved.
+- The default read function does not do any data conversion. Use `drwav_read_pcm_frames_f32()`, `drwav_read_pcm_frames_s32()` and `drwav_read_pcm_frames_s16()`
+  to read and convert audio data to 32-bit floating point, signed 32-bit integer and signed 16-bit integer samples respectively.
+- dr_wav will try to read the WAV file as best it can, even if it's not strictly conformant to the WAV format.
+*/
+
+#ifndef dr_wav_h
+#define dr_wav_h
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+#define DRWAV_STRINGIFY(x)      #x
+#define DRWAV_XSTRINGIFY(x)     DRWAV_STRINGIFY(x)
+
+#define DRWAV_VERSION_MAJOR     0
+#define DRWAV_VERSION_MINOR     13
+#define DRWAV_VERSION_REVISION  16
+#define DRWAV_VERSION_STRING    DRWAV_XSTRINGIFY(DRWAV_VERSION_MAJOR) "." DRWAV_XSTRINGIFY(DRWAV_VERSION_MINOR) "." DRWAV_XSTRINGIFY(DRWAV_VERSION_REVISION)
+
+#include <stddef.h> /* For size_t. */
+
+/* Sized Types */
+typedef   signed char           drwav_int8;
+typedef unsigned char           drwav_uint8;
+typedef   signed short          drwav_int16;
+typedef unsigned short          drwav_uint16;
+typedef   signed int            drwav_int32;
+typedef unsigned int            drwav_uint32;
+#if defined(_MSC_VER) && !defined(__clang__)
+    typedef   signed __int64    drwav_int64;
+    typedef unsigned __int64    drwav_uint64;
+#else
+    #if defined(__clang__) || (defined(__GNUC__) && (__GNUC__ > 4 || (__GNUC__ == 4 && __GNUC_MINOR__ >= 6)))
+        #pragma GCC diagnostic push
+        #pragma GCC diagnostic ignored "-Wlong-long"
+        #if defined(__clang__)
+            #pragma GCC diagnostic ignored "-Wc++11-long-long"
+        #endif
+    #endif
+    typedef   signed long long  drwav_int64;
+    typedef unsigned long long  drwav_uint64;
+    #if defined(__clang__) || (defined(__GNUC__) && (__GNUC__ > 4 || (__GNUC__ == 4 && __GNUC_MINOR__ >= 6)))
+        #pragma GCC diagnostic pop
+    #endif
+#endif
+#if defined(__LP64__) || defined(_WIN64) || (defined(__x86_64__) && !defined(__ILP32__)) || defined(_M_X64) || defined(__ia64) || defined (_M_IA64) || defined(__aarch64__) || defined(_M_ARM64) || defined(__powerpc64__)
+    typedef drwav_uint64        drwav_uintptr;
+#else
+    typedef drwav_uint32        drwav_uintptr;
+#endif
+typedef drwav_uint8             drwav_bool8;
+typedef drwav_uint32            drwav_bool32;
+#define DRWAV_TRUE              1
+#define DRWAV_FALSE             0
+/* End Sized Types */
+
+/* Decorations */
+#if !defined(DRWAV_API)
+    #if defined(DRWAV_DLL)
+        #if defined(_WIN32)
+            #define DRWAV_DLL_IMPORT  __declspec(dllimport)
+            #define DRWAV_DLL_EXPORT  __declspec(dllexport)
+            #define DRWAV_DLL_PRIVATE static
+        #else
+            #if defined(__GNUC__) && __GNUC__ >= 4
+                #define DRWAV_DLL_IMPORT  __attribute__((visibility("default")))
+                #define DRWAV_DLL_EXPORT  __attribute__((visibility("default")))
+                #define DRWAV_DLL_PRIVATE __attribute__((visibility("hidden")))
+            #else
+                #define DRWAV_DLL_IMPORT
+                #define DRWAV_DLL_EXPORT
+                #define DRWAV_DLL_PRIVATE static
+            #endif
+        #endif
+
+        #if defined(DR_WAV_IMPLEMENTATION) || defined(DRWAV_IMPLEMENTATION)
+            #define DRWAV_API  DRWAV_DLL_EXPORT
+        #else
+            #define DRWAV_API  DRWAV_DLL_IMPORT
+        #endif
+        #define DRWAV_PRIVATE DRWAV_DLL_PRIVATE
+    #else
+        #define DRWAV_API extern
+        #define DRWAV_PRIVATE static
+    #endif
+#endif
+/* End Decorations */
+
+/* Result Codes */
+typedef drwav_int32 drwav_result;
+#define DRWAV_SUCCESS                        0
+#define DRWAV_ERROR                         -1   /* A generic error. */
+#define DRWAV_INVALID_ARGS                  -2
+#define DRWAV_INVALID_OPERATION             -3
+#define DRWAV_OUT_OF_MEMORY                 -4
+#define DRWAV_OUT_OF_RANGE                  -5
+#define DRWAV_ACCESS_DENIED                 -6
+#define DRWAV_DOES_NOT_EXIST                -7
+#define DRWAV_ALREADY_EXISTS                -8
+#define DRWAV_TOO_MANY_OPEN_FILES           -9
+#define DRWAV_INVALID_FILE                  -10
+#define DRWAV_TOO_BIG                       -11
+#define DRWAV_PATH_TOO_LONG                 -12
+#define DRWAV_NAME_TOO_LONG                 -13
+#define DRWAV_NOT_DIRECTORY                 -14
+#define DRWAV_IS_DIRECTORY                  -15
+#define DRWAV_DIRECTORY_NOT_EMPTY           -16
+#define DRWAV_END_OF_FILE                   -17
+#define DRWAV_NO_SPACE                      -18
+#define DRWAV_BUSY                          -19
+#define DRWAV_IO_ERROR                      -20
+#define DRWAV_INTERRUPT                     -21
+#define DRWAV_UNAVAILABLE                   -22
+#define DRWAV_ALREADY_IN_USE                -23
+#define DRWAV_BAD_ADDRESS                   -24
+#define DRWAV_BAD_SEEK                      -25
+#define DRWAV_BAD_PIPE                      -26
+#define DRWAV_DEADLOCK                      -27
+#define DRWAV_TOO_MANY_LINKS                -28
+#define DRWAV_NOT_IMPLEMENTED               -29
+#define DRWAV_NO_MESSAGE                    -30
+#define DRWAV_BAD_MESSAGE                   -31
+#define DRWAV_NO_DATA_AVAILABLE             -32
+#define DRWAV_INVALID_DATA                  -33
+#define DRWAV_TIMEOUT                       -34
+#define DRWAV_NO_NETWORK                    -35
+#define DRWAV_NOT_UNIQUE                    -36
+#define DRWAV_NOT_SOCKET                    -37
+#define DRWAV_NO_ADDRESS                    -38
+#define DRWAV_BAD_PROTOCOL                  -39
+#define DRWAV_PROTOCOL_UNAVAILABLE          -40
+#define DRWAV_PROTOCOL_NOT_SUPPORTED        -41
+#define DRWAV_PROTOCOL_FAMILY_NOT_SUPPORTED -42
+#define DRWAV_ADDRESS_FAMILY_NOT_SUPPORTED  -43
+#define DRWAV_SOCKET_NOT_SUPPORTED          -44
+#define DRWAV_CONNECTION_RESET              -45
+#define DRWAV_ALREADY_CONNECTED             -46
+#define DRWAV_NOT_CONNECTED                 -47
+#define DRWAV_CONNECTION_REFUSED            -48
+#define DRWAV_NO_HOST                       -49
+#define DRWAV_IN_PROGRESS                   -50
+#define DRWAV_CANCELLED                     -51
+#define DRWAV_MEMORY_ALREADY_MAPPED         -52
+#define DRWAV_AT_END                        -53
+/* End Result Codes */
+
+/* Common data formats. */
+#define DR_WAVE_FORMAT_PCM          0x1
+#define DR_WAVE_FORMAT_ADPCM        0x2
+#define DR_WAVE_FORMAT_IEEE_FLOAT   0x3
+#define DR_WAVE_FORMAT_ALAW         0x6
+#define DR_WAVE_FORMAT_MULAW        0x7
+#define DR_WAVE_FORMAT_DVI_ADPCM    0x11
+#define DR_WAVE_FORMAT_EXTENSIBLE   0xFFFE
+
+/* Flags to pass into drwav_init_ex(), etc. */
+#define DRWAV_SEQUENTIAL            0x00000001
+#define DRWAV_WITH_METADATA         0x00000002
+
+DRWAV_API void drwav_version(drwav_uint32* pMajor, drwav_uint32* pMinor, drwav_uint32* pRevision);
+DRWAV_API const char* drwav_version_string(void);
+
+/* Allocation Callbacks */
+typedef struct
+{
+    void* pUserData;
+    void* (* onMalloc)(size_t sz, void* pUserData);
+    void* (* onRealloc)(void* p, size_t sz, void* pUserData);
+    void  (* onFree)(void* p, void* pUserData);
+} drwav_allocation_callbacks;
+/* End Allocation Callbacks */
+
+typedef enum
+{
+    drwav_seek_origin_start,
+    drwav_seek_origin_current
+} drwav_seek_origin;
+
+typedef enum
+{
+    drwav_container_riff,
+    drwav_container_rifx,
+    drwav_container_w64,
+    drwav_container_rf64,
+    drwav_container_aiff
+} drwav_container;
+
+typedef struct
+{
+    union
+    {
+        drwav_uint8 fourcc[4];
+        drwav_uint8 guid[16];
+    } id;
+
+    /* The size in bytes of the chunk. */
+    drwav_uint64 sizeInBytes;
+
+    /*
+    RIFF = 2 byte alignment.
+    W64  = 8 byte alignment.
+    */
+    unsigned int paddingSize;
+} drwav_chunk_header;
+
+typedef struct
+{
+    /*
+    The format tag exactly as specified in the wave file's "fmt" chunk. This can be used by applications
+    that require support for data formats not natively supported by dr_wav.
+    */
+    drwav_uint16 formatTag;
+
+    /* The number of channels making up the audio data. When this is set to 1 it is mono, 2 is stereo, etc. */
+    drwav_uint16 channels;
+
+    /* The sample rate. Usually set to something like 44100. */
+    drwav_uint32 sampleRate;
+
+    /* Average bytes per second. You probably don't need this, but it's left here for informational purposes. */
+    drwav_uint32 avgBytesPerSec;
+
+    /* Block align. This is equal to the number of channels * bytes per sample. */
+    drwav_uint16 blockAlign;
+
+    /* Bits per sample. */
+    drwav_uint16 bitsPerSample;
+
+    /* The size of the extended data. Only used internally for validation, but left here for informational purposes. */
+    drwav_uint16 extendedSize;
+
+    /*
+    The number of valid bits per sample. When <formatTag> is equal to WAVE_FORMAT_EXTENSIBLE, <bitsPerSample>
+    is always rounded up to the nearest multiple of 8. This variable contains information about exactly how
+    many bits are valid per sample. Mainly used for informational purposes.
+    */
+    drwav_uint16 validBitsPerSample;
+
+    /* The channel mask. Not used at the moment. */
+    drwav_uint32 channelMask;
+
+    /* The sub-format, exactly as specified by the wave file. */
+    drwav_uint8 subFormat[16];
+} drwav_fmt;
+
+DRWAV_API drwav_uint16 drwav_fmt_get_format(const drwav_fmt* pFMT);
+
+
+/*
+Callback for when data is read. Return value is the number of bytes actually read.
+
+pUserData   [in]  The user data that was passed to drwav_init() and family.
+pBufferOut  [out] The output buffer.
+bytesToRead [in]  The number of bytes to read.
+
+Returns the number of bytes actually read.
+
+A return value of less than bytesToRead indicates the end of the stream. Do _not_ return from this callback until
+either the entire bytesToRead is filled or you have reached the end of the stream.
+*/
+typedef size_t (* drwav_read_proc)(void* pUserData, void* pBufferOut, size_t bytesToRead);
+
+/*
+Callback for when data is written. Returns value is the number of bytes actually written.
+
+pUserData    [in]  The user data that was passed to drwav_init_write() and family.
+pData        [out] A pointer to the data to write.
+bytesToWrite [in]  The number of bytes to write.
+
+Returns the number of bytes actually written.
+
+If the return value differs from bytesToWrite, it indicates an error.
+*/
+typedef size_t (* drwav_write_proc)(void* pUserData, const void* pData, size_t bytesToWrite);
+
+/*
+Callback for when data needs to be seeked.
+
+pUserData [in] The user data that was passed to drwav_init() and family.
+offset    [in] The number of bytes to move, relative to the origin. Will never be negative.
+origin    [in] The origin of the seek - the current position or the start of the stream.
+
+Returns whether or not the seek was successful.
+
+Whether or not it is relative to the beginning or current position is determined by the "origin" parameter which will be either drwav_seek_origin_start or
+drwav_seek_origin_current.
+*/
+typedef drwav_bool32 (* drwav_seek_proc)(void* pUserData, int offset, drwav_seek_origin origin);
+
+/*
+Callback for when drwav_init_ex() finds a chunk.
+
+pChunkUserData    [in] The user data that was passed to the pChunkUserData parameter of drwav_init_ex() and family.
+onRead            [in] A pointer to the function to call when reading.
+onSeek            [in] A pointer to the function to call when seeking.
+pReadSeekUserData [in] The user data that was passed to the pReadSeekUserData parameter of drwav_init_ex() and family.
+pChunkHeader      [in] A pointer to an object containing basic header information about the chunk. Use this to identify the chunk.
+container         [in] Whether or not the WAV file is a RIFF or Wave64 container. If you're unsure of the difference, assume RIFF.
+pFMT              [in] A pointer to the object containing the contents of the "fmt" chunk.
+
+Returns the number of bytes read + seeked.
+
+To read data from the chunk, call onRead(), passing in pReadSeekUserData as the first parameter. Do the same for seeking with onSeek(). The return value must
+be the total number of bytes you have read _plus_ seeked.
+
+Use the `container` argument to discriminate the fields in `pChunkHeader->id`. If the container is `drwav_container_riff` or `drwav_container_rf64` you should
+use `id.fourcc`, otherwise you should use `id.guid`.
+
+The `pFMT` parameter can be used to determine the data format of the wave file. Use `drwav_fmt_get_format()` to get the sample format, which will be one of the
+`DR_WAVE_FORMAT_*` identifiers.
+
+The read pointer will be sitting on the first byte after the chunk's header. You must not attempt to read beyond the boundary of the chunk.
+*/
+typedef drwav_uint64 (* drwav_chunk_proc)(void* pChunkUserData, drwav_read_proc onRead, drwav_seek_proc onSeek, void* pReadSeekUserData, const drwav_chunk_header* pChunkHeader, drwav_container container, const drwav_fmt* pFMT);
+
+
+/* Structure for internal use. Only used for loaders opened with drwav_init_memory(). */
+typedef struct
+{
+    const drwav_uint8* data;
+    size_t dataSize;
+    size_t currentReadPos;
+} drwav__memory_stream;
+
+/* Structure for internal use. Only used for writers opened with drwav_init_memory_write(). */
+typedef struct
+{
+    void** ppData;
+    size_t* pDataSize;
+    size_t dataSize;
+    size_t dataCapacity;
+    size_t currentWritePos;
+} drwav__memory_stream_write;
+
+typedef struct
+{
+    drwav_container container;  /* RIFF, W64. */
+    drwav_uint32 format;        /* DR_WAVE_FORMAT_* */
+    drwav_uint32 channels;
+    drwav_uint32 sampleRate;
+    drwav_uint32 bitsPerSample;
+} drwav_data_format;
+
+typedef enum
+{
+    drwav_metadata_type_none                        = 0,
+
+    /*
+    Unknown simply means a chunk that drwav does not handle specifically. You can still ask to
+    receive these chunks as metadata objects. It is then up to you to interpret the chunk's data.
+    You can also write unknown metadata to a wav file. Be careful writing unknown chunks if you
+    have also edited the audio data. The unknown chunks could represent offsets/sizes that no
+    longer correctly correspond to the audio data.
+    */
+    drwav_metadata_type_unknown                     = 1 << 0,
+
+    /* Only 1 of each of these metadata items are allowed in a wav file. */
+    drwav_metadata_type_smpl                        = 1 << 1,
+    drwav_metadata_type_inst                        = 1 << 2,
+    drwav_metadata_type_cue                         = 1 << 3,
+    drwav_metadata_type_acid                        = 1 << 4,
+    drwav_metadata_type_bext                        = 1 << 5,
+
+    /*
+    Wav files often have a LIST chunk. This is a chunk that contains a set of subchunks. For this
+    higher-level metadata API, we don't make a distinction between a regular chunk and a LIST
+    subchunk. Instead, they are all just 'metadata' items.
+
+    There can be multiple of these metadata items in a wav file.
+    */
+    drwav_metadata_type_list_label                  = 1 << 6,
+    drwav_metadata_type_list_note                   = 1 << 7,
+    drwav_metadata_type_list_labelled_cue_region    = 1 << 8,
+
+    drwav_metadata_type_list_info_software          = 1 << 9,
+    drwav_metadata_type_list_info_copyright         = 1 << 10,
+    drwav_metadata_type_list_info_title             = 1 << 11,
+    drwav_metadata_type_list_info_artist            = 1 << 12,
+    drwav_metadata_type_list_info_comment           = 1 << 13,
+    drwav_metadata_type_list_info_date              = 1 << 14,
+    drwav_metadata_type_list_info_genre             = 1 << 15,
+    drwav_metadata_type_list_info_album             = 1 << 16,
+    drwav_metadata_type_list_info_tracknumber       = 1 << 17,
+
+    /* Other type constants for convenience. */
+    drwav_metadata_type_list_all_info_strings       = drwav_metadata_type_list_info_software
+                                                    | drwav_metadata_type_list_info_copyright
+                                                    | drwav_metadata_type_list_info_title
+                                                    | drwav_metadata_type_list_info_artist
+                                                    | drwav_metadata_type_list_info_comment
+                                                    | drwav_metadata_type_list_info_date
+                                                    | drwav_metadata_type_list_info_genre
+                                                    | drwav_metadata_type_list_info_album
+                                                    | drwav_metadata_type_list_info_tracknumber,
+
+    drwav_metadata_type_list_all_adtl               = drwav_metadata_type_list_label
+                                                    | drwav_metadata_type_list_note
+                                                    | drwav_metadata_type_list_labelled_cue_region,
+
+    drwav_metadata_type_all                         = -2,   /*0xFFFFFFFF & ~drwav_metadata_type_unknown,*/
+    drwav_metadata_type_all_including_unknown       = -1    /*0xFFFFFFFF,*/
+} drwav_metadata_type;
+
+/*
+Sampler Metadata
+
+The sampler chunk contains information about how a sound should be played in the context of a whole
+audio production, and when used in a sampler. See https://en.wikipedia.org/wiki/Sample-based_synthesis.
+*/
+typedef enum
+{
+    drwav_smpl_loop_type_forward  = 0,
+    drwav_smpl_loop_type_pingpong = 1,
+    drwav_smpl_loop_type_backward = 2
+} drwav_smpl_loop_type;
+
+typedef struct
+{
+    /* The ID of the associated cue point, see drwav_cue and drwav_cue_point. As with all cue point IDs, this can correspond to a label chunk to give this loop a name, see drwav_list_label_or_note. */
+    drwav_uint32 cuePointId;
+
+    /* See drwav_smpl_loop_type. */
+    drwav_uint32 type;
+
+    /* The byte offset of the first sample to be played in the loop. */
+    drwav_uint32 firstSampleByteOffset;
+
+    /* The byte offset into the audio data of the last sample to be played in the loop. */
+    drwav_uint32 lastSampleByteOffset;
+
+    /* A value to represent that playback should occur at a point between samples. This value ranges from 0 to UINT32_MAX. Where a value of 0 means no fraction, and a value of (UINT32_MAX / 2) would mean half a sample. */
+    drwav_uint32 sampleFraction;
+
+    /* Number of times to play the loop. 0 means loop infinitely. */
+    drwav_uint32 playCount;
+} drwav_smpl_loop;
+
+typedef struct
+{
+    /* IDs for a particular MIDI manufacturer. 0 if not used. */
+    drwav_uint32 manufacturerId;
+    drwav_uint32 productId;
+
+    /* The period of 1 sample in nanoseconds. */
+    drwav_uint32 samplePeriodNanoseconds;
+
+    /* The MIDI root note of this file. 0 to 127. */
+    drwav_uint32 midiUnityNote;
+
+    /* The fraction of a semitone up from the given MIDI note. This is a value from 0 to UINT32_MAX, where 0 means no change and (UINT32_MAX / 2) is half a semitone (AKA 50 cents). */
+    drwav_uint32 midiPitchFraction;
+
+    /* Data relating to SMPTE standards which are used for syncing audio and video. 0 if not used. */
+    drwav_uint32 smpteFormat;
+    drwav_uint32 smpteOffset;
+
+    /* drwav_smpl_loop loops. */
+    drwav_uint32 sampleLoopCount;
+
+    /* Optional sampler-specific data. */
+    drwav_uint32 samplerSpecificDataSizeInBytes;
+
+    drwav_smpl_loop* pLoops;
+    drwav_uint8* pSamplerSpecificData;
+} drwav_smpl;
+
+/*
+Instrument Metadata
+
+The inst metadata contains data about how a sound should be played as part of an instrument. This
+commonly read by samplers. See https://en.wikipedia.org/wiki/Sample-based_synthesis.
+*/
+typedef struct
+{
+    drwav_int8 midiUnityNote;   /* The root note of the audio as a MIDI note number. 0 to 127. */
+    drwav_int8 fineTuneCents;   /* -50 to +50 */
+    drwav_int8 gainDecibels;    /* -64 to +64 */
+    drwav_int8 lowNote;         /* 0 to 127 */
+    drwav_int8 highNote;        /* 0 to 127 */
+    drwav_int8 lowVelocity;     /* 1 to 127 */
+    drwav_int8 highVelocity;    /* 1 to 127 */
+} drwav_inst;
+
+/*
+Cue Metadata
+
+Cue points are markers at specific points in the audio. They often come with an associated piece of
+drwav_list_label_or_note metadata which contains the text for the marker.
+*/
+typedef struct
+{
+    /* Unique identification value. */
+    drwav_uint32 id;
+
+    /* Set to 0. This is only relevant if there is a 'playlist' chunk - which is not supported by dr_wav. */
+    drwav_uint32 playOrderPosition;
+
+    /* Should always be "data". This represents the fourcc value of the chunk that this cue point corresponds to. dr_wav only supports a single data chunk so this should always be "data". */
+    drwav_uint8 dataChunkId[4];
+
+    /* Set to 0. This is only relevant if there is a wave list chunk. dr_wav, like lots of readers/writers, do not support this. */
+    drwav_uint32 chunkStart;
+
+    /* Set to 0 for uncompressed formats. Else the last byte in compressed wave data where decompression can begin to find the value of the corresponding sample value. */
+    drwav_uint32 blockStart;
+
+    /* For uncompressed formats this is the byte offset of the cue point into the audio data. For compressed formats this is relative to the block specified with blockStart. */
+    drwav_uint32 sampleByteOffset;
+} drwav_cue_point;
+
+typedef struct
+{
+    drwav_uint32 cuePointCount;
+    drwav_cue_point *pCuePoints;
+} drwav_cue;
+
+/*
+Acid Metadata
+
+This chunk contains some information about the time signature and the tempo of the audio.
+*/
+typedef enum
+{
+    drwav_acid_flag_one_shot      = 1,  /* If this is not set, then it is a loop instead of a one-shot. */
+    drwav_acid_flag_root_note_set = 2,
+    drwav_acid_flag_stretch       = 4,
+    drwav_acid_flag_disk_based    = 8,
+    drwav_acid_flag_acidizer      = 16  /* Not sure what this means. */
+} drwav_acid_flag;
+
+typedef struct
+{
+    /* A bit-field, see drwav_acid_flag. */
+    drwav_uint32 flags;
+
+    /* Valid if flags contains drwav_acid_flag_root_note_set. It represents the MIDI root note the file - a value from 0 to 127. */
+    drwav_uint16 midiUnityNote;
+
+    /* Reserved values that should probably be ignored. reserved1 seems to often be 128 and reserved2 is 0. */
+    drwav_uint16 reserved1;
+    float reserved2;
+
+    /* Number of beats. */
+    drwav_uint32 numBeats;
+
+    /* The time signature of the audio. */
+    drwav_uint16 meterDenominator;
+    drwav_uint16 meterNumerator;
+
+    /* Beats per minute of the track. Setting a value of 0 suggests that there is no tempo. */
+    float tempo;
+} drwav_acid;
+
+/*
+Cue Label or Note metadata
+
+These are 2 different types of metadata, but they have the exact same format. Labels tend to be the
+more common and represent a short name for a cue point. Notes might be used to represent a longer
+comment.
+*/
+typedef struct
+{
+    /* The ID of a cue point that this label or note corresponds to. */
+    drwav_uint32 cuePointId;
+
+    /* Size of the string not including any null terminator. */
+    drwav_uint32 stringLength;
+
+    /* The string. The *init_with_metadata functions null terminate this for convenience. */
+    char* pString;
+} drwav_list_label_or_note;
+
+/*
+BEXT metadata, also known as Broadcast Wave Format (BWF)
+
+This metadata adds some extra description to an audio file. You must check the version field to
+determine if the UMID or the loudness fields are valid.
+*/
+typedef struct
+{
+    /*
+    These top 3 fields, and the umid field are actually defined in the standard as a statically
+    sized buffers. In order to reduce the size of this struct (and therefore the union in the
+    metadata struct), we instead store these as pointers.
+    */
+    char* pDescription;                 /* Can be NULL or a null-terminated string, must be <= 256 characters. */
+    char* pOriginatorName;              /* Can be NULL or a null-terminated string, must be <= 32 characters. */
+    char* pOriginatorReference;         /* Can be NULL or a null-terminated string, must be <= 32 characters. */
+    char  pOriginationDate[10];         /* ASCII "yyyy:mm:dd". */
+    char  pOriginationTime[8];          /* ASCII "hh:mm:ss". */
+    drwav_uint64 timeReference;         /* First sample count since midnight. */
+    drwav_uint16 version;               /* Version of the BWF, check this to see if the fields below are valid. */
+
+    /*
+    Unrestricted ASCII characters containing a collection of strings terminated by CR/LF. Each
+    string shall contain a description of a coding process applied to the audio data.
+    */
+    char* pCodingHistory;
+    drwav_uint32 codingHistorySize;
+
+    /* Fields below this point are only valid if the version is 1 or above. */
+    drwav_uint8* pUMID;                  /* Exactly 64 bytes of SMPTE UMID */
+
+    /* Fields below this point are only valid if the version is 2 or above. */
+    drwav_uint16 loudnessValue;         /* Integrated Loudness Value of the file in LUFS (multiplied by 100). */
+    drwav_uint16 loudnessRange;         /* Loudness Range of the file in LU (multiplied by 100). */
+    drwav_uint16 maxTruePeakLevel;      /* Maximum True Peak Level of the file expressed as dBTP (multiplied by 100). */
+    drwav_uint16 maxMomentaryLoudness;  /* Highest value of the Momentary Loudness Level of the file in LUFS (multiplied by 100). */
+    drwav_uint16 maxShortTermLoudness;  /* Highest value of the Short-Term Loudness Level of the file in LUFS (multiplied by 100). */
+} drwav_bext;
+
+/*
+Info Text Metadata
+
+There a many different types of information text that can be saved in this format. This is where
+things like the album name, the artists, the year it was produced, etc are saved. See
+drwav_metadata_type for the full list of types that dr_wav supports.
+*/
+typedef struct
+{
+    /* Size of the string not including any null terminator. */
+    drwav_uint32 stringLength;
+
+    /* The string. The *init_with_metadata functions null terminate this for convenience. */
+    char* pString;
+} drwav_list_info_text;
+
+/*
+Labelled Cue Region Metadata
+
+The labelled cue region metadata is used to associate some region of audio with text. The region
+starts at a cue point, and extends for the given number of samples.
+*/
+typedef struct
+{
+    /* The ID of a cue point that this object corresponds to. */
+    drwav_uint32 cuePointId;
+
+    /* The number of samples from the cue point forwards that should be considered this region */
+    drwav_uint32 sampleLength;
+
+    /* Four characters used to say what the purpose of this region is. */
+    drwav_uint8 purposeId[4];
+
+    /* Unsure of the exact meanings of these. It appears to be acceptable to set them all to 0. */
+    drwav_uint16 country;
+    drwav_uint16 language;
+    drwav_uint16 dialect;
+    drwav_uint16 codePage;
+
+    /* Size of the string not including any null terminator. */
+    drwav_uint32 stringLength;
+
+    /* The string. The *init_with_metadata functions null terminate this for convenience. */
+    char* pString;
+} drwav_list_labelled_cue_region;
+
+/*
+Unknown Metadata
+
+This chunk just represents a type of chunk that dr_wav does not understand.
+
+Unknown metadata has a location attached to it. This is because wav files can have a LIST chunk
+that contains subchunks. These LIST chunks can be one of two types. An adtl list, or an INFO
+list. This enum is used to specify the location of a chunk that dr_wav currently doesn't support.
+*/
+typedef enum
+{
+    drwav_metadata_location_invalid,
+    drwav_metadata_location_top_level,
+    drwav_metadata_location_inside_info_list,
+    drwav_metadata_location_inside_adtl_list
+} drwav_metadata_location;
+
+typedef struct
+{
+    drwav_uint8 id[4];
+    drwav_metadata_location chunkLocation;
+    drwav_uint32 dataSizeInBytes;
+    drwav_uint8* pData;
+} drwav_unknown_metadata;
+
+/*
+Metadata is saved as a union of all the supported types.
+*/
+typedef struct
+{
+    /* Determines which item in the union is valid. */
+    drwav_metadata_type type;
+
+    union
+    {
+        drwav_cue cue;
+        drwav_smpl smpl;
+        drwav_acid acid;
+        drwav_inst inst;
+        drwav_bext bext;
+        drwav_list_label_or_note labelOrNote;   /* List label or list note. */
+        drwav_list_labelled_cue_region labelledCueRegion;
+        drwav_list_info_text infoText;          /* Any of the list info types. */
+        drwav_unknown_metadata unknown;
+    } data;
+} drwav_metadata;
+
+typedef struct
+{
+    /* A pointer to the function to call when more data is needed. */
+    drwav_read_proc onRead;
+
+    /* A pointer to the function to call when data needs to be written. Only used when the drwav object is opened in write mode. */
+    drwav_write_proc onWrite;
+
+    /* A pointer to the function to call when the wav file needs to be seeked. */
+    drwav_seek_proc onSeek;
+
+    /* The user data to pass to callbacks. */
+    void* pUserData;
+
+    /* Allocation callbacks. */
+    drwav_allocation_callbacks allocationCallbacks;
+
+
+    /* Whether or not the WAV file is formatted as a standard RIFF file or W64. */
+    drwav_container container;
+
+
+    /* Structure containing format information exactly as specified by the wav file. */
+    drwav_fmt fmt;
+
+    /* The sample rate. Will be set to something like 44100. */
+    drwav_uint32 sampleRate;
+
+    /* The number of channels. This will be set to 1 for monaural streams, 2 for stereo, etc. */
+    drwav_uint16 channels;
+
+    /* The bits per sample. Will be set to something like 16, 24, etc. */
+    drwav_uint16 bitsPerSample;
+
+    /* Equal to fmt.formatTag, or the value specified by fmt.subFormat if fmt.formatTag is equal to 65534 (WAVE_FORMAT_EXTENSIBLE). */
+    drwav_uint16 translatedFormatTag;
+
+    /* The total number of PCM frames making up the audio data. */
+    drwav_uint64 totalPCMFrameCount;
+
+
+    /* The size in bytes of the data chunk. */
+    drwav_uint64 dataChunkDataSize;
+
+    /* The position in the stream of the first data byte of the data chunk. This is used for seeking. */
+    drwav_uint64 dataChunkDataPos;
+
+    /* The number of bytes remaining in the data chunk. */
+    drwav_uint64 bytesRemaining;
+
+    /* The current read position in PCM frames. */
+    drwav_uint64 readCursorInPCMFrames;
+
+
+    /*
+    Only used in sequential write mode. Keeps track of the desired size of the "data" chunk at the point of initialization time. Always
+    set to 0 for non-sequential writes and when the drwav object is opened in read mode. Used for validation.
+    */
+    drwav_uint64 dataChunkDataSizeTargetWrite;
+
+    /* Keeps track of whether or not the wav writer was initialized in sequential mode. */
+    drwav_bool32 isSequentialWrite;
+
+
+    /* A array of metadata. This is valid after the *init_with_metadata call returns. It will be valid until drwav_uninit() is called. You can take ownership of this data with drwav_take_ownership_of_metadata(). */
+    drwav_metadata* pMetadata;
+    drwav_uint32 metadataCount;
+
+
+    /* A hack to avoid a DRWAV_MALLOC() when opening a decoder with drwav_init_memory(). */
+    drwav__memory_stream memoryStream;
+    drwav__memory_stream_write memoryStreamWrite;
+
+
+    /* Microsoft ADPCM specific data. */
+    struct
+    {
+        drwav_uint32 bytesRemainingInBlock;
+        drwav_uint16 predictor[2];
+        drwav_int32  delta[2];
+        drwav_int32  cachedFrames[4];  /* Samples are stored in this cache during decoding. */
+        drwav_uint32 cachedFrameCount;
+        drwav_int32  prevFrames[2][2]; /* The previous 2 samples for each channel (2 channels at most). */
+    } msadpcm;
+
+    /* IMA ADPCM specific data. */
+    struct
+    {
+        drwav_uint32 bytesRemainingInBlock;
+        drwav_int32  predictor[2];
+        drwav_int32  stepIndex[2];
+        drwav_int32  cachedFrames[16]; /* Samples are stored in this cache during decoding. */
+        drwav_uint32 cachedFrameCount;
+    } ima;
+
+    /* AIFF specific data. */
+    struct
+    {
+        drwav_bool8 isLE;   /* Will be set to true if the audio data is little-endian encoded. */
+        drwav_bool8 isUnsigned; /* Only used for 8-bit samples. When set to true, will be treated as unsigned. */
+    } aiff;
+} drwav;
+
+
+/*
+Initializes a pre-allocated drwav object for reading.
+
+pWav                         [out]          A pointer to the drwav object being initialized.
+onRead                       [in]           The function to call when data needs to be read from the client.
+onSeek                       [in]           The function to call when the read position of the client data needs to move.
+onChunk                      [in, optional] The function to call when a chunk is enumerated at initialized time.
+pUserData, pReadSeekUserData [in, optional] A pointer to application defined data that will be passed to onRead and onSeek.
+pChunkUserData               [in, optional] A pointer to application defined data that will be passed to onChunk.
+flags                        [in, optional] A set of flags for controlling how things are loaded.
+
+Returns true if successful; false otherwise.
+
+Close the loader with drwav_uninit().
+
+This is the lowest level function for initializing a WAV file. You can also use drwav_init_file() and drwav_init_memory()
+to open the stream from a file or from a block of memory respectively.
+
+Possible values for flags:
+  DRWAV_SEQUENTIAL: Never perform a backwards seek while loading. This disables the chunk callback and will cause this function
+                    to return as soon as the data chunk is found. Any chunks after the data chunk will be ignored.
+
+drwav_init() is equivalent to "drwav_init_ex(pWav, onRead, onSeek, NULL, pUserData, NULL, 0);".
+
+The onChunk callback is not called for the WAVE or FMT chunks. The contents of the FMT chunk can be read from pWav->fmt
+after the function returns.
+
+See also: drwav_init_file(), drwav_init_memory(), drwav_uninit()
+*/
+DRWAV_API drwav_bool32 drwav_init(drwav* pWav, drwav_read_proc onRead, drwav_seek_proc onSeek, void* pUserData, const drwav_allocation_callbacks* pAllocationCallbacks);
+DRWAV_API drwav_bool32 drwav_init_ex(drwav* pWav, drwav_read_proc onRead, drwav_seek_proc onSeek, drwav_chunk_proc onChunk, void* pReadSeekUserData, void* pChunkUserData, drwav_uint32 flags, const drwav_allocation_callbacks* pAllocationCallbacks);
+DRWAV_API drwav_bool32 drwav_init_with_metadata(drwav* pWav, drwav_read_proc onRead, drwav_seek_proc onSeek, void* pUserData, drwav_uint32 flags, const drwav_allocation_callbacks* pAllocationCallbacks);
+
+/*
+Initializes a pre-allocated drwav object for writing.
+
+onWrite               [in]           The function to call when data needs to be written.
+onSeek                [in]           The function to call when the write position needs to move.
+pUserData             [in, optional] A pointer to application defined data that will be passed to onWrite and onSeek.
+metadata, numMetadata [in, optional] An array of metadata objects that should be written to the file. The array is not edited. You are responsible for this metadata memory and it must maintain valid until drwav_uninit() is called.
+
+Returns true if successful; false otherwise.
+
+Close the writer with drwav_uninit().
+
+This is the lowest level function for initializing a WAV file. You can also use drwav_init_file_write() and drwav_init_memory_write()
+to open the stream from a file or from a block of memory respectively.
+
+If the total sample count is known, you can use drwav_init_write_sequential(). This avoids the need for dr_wav to perform
+a post-processing step for storing the total sample count and the size of the data chunk which requires a backwards seek.
+
+See also: drwav_init_file_write(), drwav_init_memory_write(), drwav_uninit()
+*/
+DRWAV_API drwav_bool32 drwav_init_write(drwav* pWav, const drwav_data_format* pFormat, drwav_write_proc onWrite, drwav_seek_proc onSeek, void* pUserData, const drwav_allocation_callbacks* pAllocationCallbacks);
+DRWAV_API drwav_bool32 drwav_init_write_sequential(drwav* pWav, const drwav_data_format* pFormat, drwav_uint64 totalSampleCount, drwav_write_proc onWrite, void* pUserData, const drwav_allocation_callbacks* pAllocationCallbacks);
+DRWAV_API drwav_bool32 drwav_init_write_sequential_pcm_frames(drwav* pWav, const drwav_data_format* pFormat, drwav_uint64 totalPCMFrameCount, drwav_write_proc onWrite, void* pUserData, const drwav_allocation_callbacks* pAllocationCallbacks);
+DRWAV_API drwav_bool32 drwav_init_write_with_metadata(drwav* pWav, const drwav_data_format* pFormat, drwav_write_proc onWrite, drwav_seek_proc onSeek, void* pUserData, const drwav_allocation_callbacks* pAllocationCallbacks, drwav_metadata* pMetadata, drwav_uint32 metadataCount);
+
+/*
+Utility function to determine the target size of the entire data to be written (including all headers and chunks).
+
+Returns the target size in bytes.
+
+The metadata argument can be NULL meaning no metadata exists.
+
+Useful if the application needs to know the size to allocate.
+
+Only writing to the RIFF chunk and one data chunk is currently supported.
+
+See also: drwav_init_write(), drwav_init_file_write(), drwav_init_memory_write()
+*/
+DRWAV_API drwav_uint64 drwav_target_write_size_bytes(const drwav_data_format* pFormat, drwav_uint64 totalFrameCount, drwav_metadata* pMetadata, drwav_uint32 metadataCount);
+
+/*
+Take ownership of the metadata objects that were allocated via one of the init_with_metadata() function calls. The init_with_metdata functions perform a single heap allocation for this metadata.
+
+Useful if you want the data to persist beyond the lifetime of the drwav object.
+
+You must free the data returned from this function using drwav_free().
+*/
+DRWAV_API drwav_metadata* drwav_take_ownership_of_metadata(drwav* pWav);
+
+/*
+Uninitializes the given drwav object.
+
+Use this only for objects initialized with drwav_init*() functions (drwav_init(), drwav_init_ex(), drwav_init_write(), drwav_init_write_sequential()).
+*/
+DRWAV_API drwav_result drwav_uninit(drwav* pWav);
+
+
+/*
+Reads raw audio data.
+
+This is the lowest level function for reading audio data. It simply reads the given number of
+bytes of the raw internal sample data.
+
+Consider using drwav_read_pcm_frames_s16(), drwav_read_pcm_frames_s32() or drwav_read_pcm_frames_f32() for
+reading sample data in a consistent format.
+
+pBufferOut can be NULL in which case a seek will be performed.
+
+Returns the number of bytes actually read.
+*/
+DRWAV_API size_t drwav_read_raw(drwav* pWav, size_t bytesToRead, void* pBufferOut);
+
+/*
+Reads up to the specified number of PCM frames from the WAV file.
+
+The output data will be in the file's internal format, converted to native-endian byte order. Use
+drwav_read_pcm_frames_s16/f32/s32() to read data in a specific format.
+
+If the return value is less than <framesToRead> it means the end of the file has been reached or
+you have requested more PCM frames than can possibly fit in the output buffer.
+
+This function will only work when sample data is of a fixed size and uncompressed. If you are
+using a compressed format consider using drwav_read_raw() or drwav_read_pcm_frames_s16/s32/f32().
+
+pBufferOut can be NULL in which case a seek will be performed.
+*/
+DRWAV_API drwav_uint64 drwav_read_pcm_frames(drwav* pWav, drwav_uint64 framesToRead, void* pBufferOut);
+DRWAV_API drwav_uint64 drwav_read_pcm_frames_le(drwav* pWav, drwav_uint64 framesToRead, void* pBufferOut);
+DRWAV_API drwav_uint64 drwav_read_pcm_frames_be(drwav* pWav, drwav_uint64 framesToRead, void* pBufferOut);
+
+/*
+Seeks to the given PCM frame.
+
+Returns true if successful; false otherwise.
+*/
+DRWAV_API drwav_bool32 drwav_seek_to_pcm_frame(drwav* pWav, drwav_uint64 targetFrameIndex);
+
+/*
+Retrieves the current read position in pcm frames.
+*/
+DRWAV_API drwav_result drwav_get_cursor_in_pcm_frames(drwav* pWav, drwav_uint64* pCursor);
+
+/*
+Retrieves the length of the file.
+*/
+DRWAV_API drwav_result drwav_get_length_in_pcm_frames(drwav* pWav, drwav_uint64* pLength);
+
+
+/*
+Writes raw audio data.
+
+Returns the number of bytes actually written. If this differs from bytesToWrite, it indicates an error.
+*/
+DRWAV_API size_t drwav_write_raw(drwav* pWav, size_t bytesToWrite, const void* pData);
+
+/*
+Writes PCM frames.
+
+Returns the number of PCM frames written.
+
+Input samples need to be in native-endian byte order. On big-endian architectures the input data will be converted to
+little-endian. Use drwav_write_raw() to write raw audio data without performing any conversion.
+*/
+DRWAV_API drwav_uint64 drwav_write_pcm_frames(drwav* pWav, drwav_uint64 framesToWrite, const void* pData);
+DRWAV_API drwav_uint64 drwav_write_pcm_frames_le(drwav* pWav, drwav_uint64 framesToWrite, const void* pData);
+DRWAV_API drwav_uint64 drwav_write_pcm_frames_be(drwav* pWav, drwav_uint64 framesToWrite, const void* pData);
+
+/* Conversion Utilities */
+#ifndef DR_WAV_NO_CONVERSION_API
+
+/*
+Reads a chunk of audio data and converts it to signed 16-bit PCM samples.
+
+pBufferOut can be NULL in which case a seek will be performed.
+
+Returns the number of PCM frames actually read.
+
+If the return value is less than <framesToRead> it means the end of the file has been reached.
+*/
+DRWAV_API drwav_uint64 drwav_read_pcm_frames_s16(drwav* pWav, drwav_uint64 framesToRead, drwav_int16* pBufferOut);
+DRWAV_API drwav_uint64 drwav_read_pcm_frames_s16le(drwav* pWav, drwav_uint64 framesToRead, drwav_int16* pBufferOut);
+DRWAV_API drwav_uint64 drwav_read_pcm_frames_s16be(drwav* pWav, drwav_uint64 framesToRead, drwav_int16* pBufferOut);
+
+/* Low-level function for converting unsigned 8-bit PCM samples to signed 16-bit PCM samples. */
+DRWAV_API void drwav_u8_to_s16(drwav_int16* pOut, const drwav_uint8* pIn, size_t sampleCount);
+
+/* Low-level function for converting signed 24-bit PCM samples to signed 16-bit PCM samples. */
+DRWAV_API void drwav_s24_to_s16(drwav_int16* pOut, const drwav_uint8* pIn, size_t sampleCount);
+
+/* Low-level function for converting signed 32-bit PCM samples to signed 16-bit PCM samples. */
+DRWAV_API void drwav_s32_to_s16(drwav_int16* pOut, const drwav_int32* pIn, size_t sampleCount);
+
+/* Low-level function for converting IEEE 32-bit floating point samples to signed 16-bit PCM samples. */
+DRWAV_API void drwav_f32_to_s16(drwav_int16* pOut, const float* pIn, size_t sampleCount);
+
+/* Low-level function for converting IEEE 64-bit floating point samples to signed 16-bit PCM samples. */
+DRWAV_API void drwav_f64_to_s16(drwav_int16* pOut, const double* pIn, size_t sampleCount);
+
+/* Low-level function for converting A-law samples to signed 16-bit PCM samples. */
+DRWAV_API void drwav_alaw_to_s16(drwav_int16* pOut, const drwav_uint8* pIn, size_t sampleCount);
+
+/* Low-level function for converting u-law samples to signed 16-bit PCM samples. */
+DRWAV_API void drwav_mulaw_to_s16(drwav_int16* pOut, const drwav_uint8* pIn, size_t sampleCount);
+
+
+/*
+Reads a chunk of audio data and converts it to IEEE 32-bit floating point samples.
+
+pBufferOut can be NULL in which case a seek will be performed.
+
+Returns the number of PCM frames actually read.
+
+If the return value is less than <framesToRead> it means the end of the file has been reached.
+*/
+DRWAV_API drwav_uint64 drwav_read_pcm_frames_f32(drwav* pWav, drwav_uint64 framesToRead, float* pBufferOut);
+DRWAV_API drwav_uint64 drwav_read_pcm_frames_f32le(drwav* pWav, drwav_uint64 framesToRead, float* pBufferOut);
+DRWAV_API drwav_uint64 drwav_read_pcm_frames_f32be(drwav* pWav, drwav_uint64 framesToRead, float* pBufferOut);
+
+/* Low-level function for converting unsigned 8-bit PCM samples to IEEE 32-bit floating point samples. */
+DRWAV_API void drwav_u8_to_f32(float* pOut, const drwav_uint8* pIn, size_t sampleCount);
+
+/* Low-level function for converting signed 16-bit PCM samples to IEEE 32-bit floating point samples. */
+DRWAV_API void drwav_s16_to_f32(float* pOut, const drwav_int16* pIn, size_t sampleCount);
+
+/* Low-level function for converting signed 24-bit PCM samples to IEEE 32-bit floating point samples. */
+DRWAV_API void drwav_s24_to_f32(float* pOut, const drwav_uint8* pIn, size_t sampleCount);
+
+/* Low-level function for converting signed 32-bit PCM samples to IEEE 32-bit floating point samples. */
+DRWAV_API void drwav_s32_to_f32(float* pOut, const drwav_int32* pIn, size_t sampleCount);
+
+/* Low-level function for converting IEEE 64-bit floating point samples to IEEE 32-bit floating point samples. */
+DRWAV_API void drwav_f64_to_f32(float* pOut, const double* pIn, size_t sampleCount);
+
+/* Low-level function for converting A-law samples to IEEE 32-bit floating point samples. */
+DRWAV_API void drwav_alaw_to_f32(float* pOut, const drwav_uint8* pIn, size_t sampleCount);
+
+/* Low-level function for converting u-law samples to IEEE 32-bit floating point samples. */
+DRWAV_API void drwav_mulaw_to_f32(float* pOut, const drwav_uint8* pIn, size_t sampleCount);
+
+
+/*
+Reads a chunk of audio data and converts it to signed 32-bit PCM samples.
+
+pBufferOut can be NULL in which case a seek will be performed.
+
+Returns the number of PCM frames actually read.
+
+If the return value is less than <framesToRead> it means the end of the file has been reached.
+*/
+DRWAV_API drwav_uint64 drwav_read_pcm_frames_s32(drwav* pWav, drwav_uint64 framesToRead, drwav_int32* pBufferOut);
+DRWAV_API drwav_uint64 drwav_read_pcm_frames_s32le(drwav* pWav, drwav_uint64 framesToRead, drwav_int32* pBufferOut);
+DRWAV_API drwav_uint64 drwav_read_pcm_frames_s32be(drwav* pWav, drwav_uint64 framesToRead, drwav_int32* pBufferOut);
+
+/* Low-level function for converting unsigned 8-bit PCM samples to signed 32-bit PCM samples. */
+DRWAV_API void drwav_u8_to_s32(drwav_int32* pOut, const drwav_uint8* pIn, size_t sampleCount);
+
+/* Low-level function for converting signed 16-bit PCM samples to signed 32-bit PCM samples. */
+DRWAV_API void drwav_s16_to_s32(drwav_int32* pOut, const drwav_int16* pIn, size_t sampleCount);
+
+/* Low-level function for converting signed 24-bit PCM samples to signed 32-bit PCM samples. */
+DRWAV_API void drwav_s24_to_s32(drwav_int32* pOut, const drwav_uint8* pIn, size_t sampleCount);
+
+/* Low-level function for converting IEEE 32-bit floating point samples to signed 32-bit PCM samples. */
+DRWAV_API void drwav_f32_to_s32(drwav_int32* pOut, const float* pIn, size_t sampleCount);
+
+/* Low-level function for converting IEEE 64-bit floating point samples to signed 32-bit PCM samples. */
+DRWAV_API void drwav_f64_to_s32(drwav_int32* pOut, const double* pIn, size_t sampleCount);
+
+/* Low-level function for converting A-law samples to signed 32-bit PCM samples. */
+DRWAV_API void drwav_alaw_to_s32(drwav_int32* pOut, const drwav_uint8* pIn, size_t sampleCount);
+
+/* Low-level function for converting u-law samples to signed 32-bit PCM samples. */
+DRWAV_API void drwav_mulaw_to_s32(drwav_int32* pOut, const drwav_uint8* pIn, size_t sampleCount);
+
+#endif  /* DR_WAV_NO_CONVERSION_API */
+
+
+/* High-Level Convenience Helpers */
+
+#ifndef DR_WAV_NO_STDIO
+/*
+Helper for initializing a wave file for reading using stdio.
+
+This holds the internal FILE object until drwav_uninit() is called. Keep this in mind if you're caching drwav
+objects because the operating system may restrict the number of file handles an application can have open at
+any given time.
+*/
+DRWAV_API drwav_bool32 drwav_init_file(drwav* pWav, const char* filename, const drwav_allocation_callbacks* pAllocationCallbacks);
+DRWAV_API drwav_bool32 drwav_init_file_ex(drwav* pWav, const char* filename, drwav_chunk_proc onChunk, void* pChunkUserData, drwav_uint32 flags, const drwav_allocation_callbacks* pAllocationCallbacks);
+DRWAV_API drwav_bool32 drwav_init_file_w(drwav* pWav, const wchar_t* filename, const drwav_allocation_callbacks* pAllocationCallbacks);
+DRWAV_API drwav_bool32 drwav_init_file_ex_w(drwav* pWav, const wchar_t* filename, drwav_chunk_proc onChunk, void* pChunkUserData, drwav_uint32 flags, const drwav_allocation_callbacks* pAllocationCallbacks);
+DRWAV_API drwav_bool32 drwav_init_file_with_metadata(drwav* pWav, const char* filename, drwav_uint32 flags, const drwav_allocation_callbacks* pAllocationCallbacks);
+DRWAV_API drwav_bool32 drwav_init_file_with_metadata_w(drwav* pWav, const wchar_t* filename, drwav_uint32 flags, const drwav_allocation_callbacks* pAllocationCallbacks);
+
+
+/*
+Helper for initializing a wave file for writing using stdio.
+
+This holds the internal FILE object until drwav_uninit() is called. Keep this in mind if you're caching drwav
+objects because the operating system may restrict the number of file handles an application can have open at
+any given time.
+*/
+DRWAV_API drwav_bool32 drwav_init_file_write(drwav* pWav, const char* filename, const drwav_data_format* pFormat, const drwav_allocation_callbacks* pAllocationCallbacks);
+DRWAV_API drwav_bool32 drwav_init_file_write_sequential(drwav* pWav, const char* filename, const drwav_data_format* pFormat, drwav_uint64 totalSampleCount, const drwav_allocation_callbacks* pAllocationCallbacks);
+DRWAV_API drwav_bool32 drwav_init_file_write_sequential_pcm_frames(drwav* pWav, const char* filename, const drwav_data_format* pFormat, drwav_uint64 totalPCMFrameCount, const drwav_allocation_callbacks* pAllocationCallbacks);
+DRWAV_API drwav_bool32 drwav_init_file_write_w(drwav* pWav, const wchar_t* filename, const drwav_data_format* pFormat, const drwav_allocation_callbacks* pAllocationCallbacks);
+DRWAV_API drwav_bool32 drwav_init_file_write_sequential_w(drwav* pWav, const wchar_t* filename, const drwav_data_format* pFormat, drwav_uint64 totalSampleCount, const drwav_allocation_callbacks* pAllocationCallbacks);
+DRWAV_API drwav_bool32 drwav_init_file_write_sequential_pcm_frames_w(drwav* pWav, const wchar_t* filename, const drwav_data_format* pFormat, drwav_uint64 totalPCMFrameCount, const drwav_allocation_callbacks* pAllocationCallbacks);
+#endif  /* DR_WAV_NO_STDIO */
+
+/*
+Helper for initializing a loader from a pre-allocated memory buffer.
+
+This does not create a copy of the data. It is up to the application to ensure the buffer remains valid for
+the lifetime of the drwav object.
+
+The buffer should contain the contents of the entire wave file, not just the sample data.
+*/
+DRWAV_API drwav_bool32 drwav_init_memory(drwav* pWav, const void* data, size_t dataSize, const drwav_allocation_callbacks* pAllocationCallbacks);
+DRWAV_API drwav_bool32 drwav_init_memory_ex(drwav* pWav, const void* data, size_t dataSize, drwav_chunk_proc onChunk, void* pChunkUserData, drwav_uint32 flags, const drwav_allocation_callbacks* pAllocationCallbacks);
+DRWAV_API drwav_bool32 drwav_init_memory_with_metadata(drwav* pWav, const void* data, size_t dataSize, drwav_uint32 flags, const drwav_allocation_callbacks* pAllocationCallbacks);
+
+/*
+Helper for initializing a writer which outputs data to a memory buffer.
+
+dr_wav will manage the memory allocations, however it is up to the caller to free the data with drwav_free().
+
+The buffer will remain allocated even after drwav_uninit() is called. The buffer should not be considered valid
+until after drwav_uninit() has been called.
+*/
+DRWAV_API drwav_bool32 drwav_init_memory_write(drwav* pWav, void** ppData, size_t* pDataSize, const drwav_data_format* pFormat, const drwav_allocation_callbacks* pAllocationCallbacks);
+DRWAV_API drwav_bool32 drwav_init_memory_write_sequential(drwav* pWav, void** ppData, size_t* pDataSize, const drwav_data_format* pFormat, drwav_uint64 totalSampleCount, const drwav_allocation_callbacks* pAllocationCallbacks);
+DRWAV_API drwav_bool32 drwav_init_memory_write_sequential_pcm_frames(drwav* pWav, void** ppData, size_t* pDataSize, const drwav_data_format* pFormat, drwav_uint64 totalPCMFrameCount, const drwav_allocation_callbacks* pAllocationCallbacks);
+
+
+#ifndef DR_WAV_NO_CONVERSION_API
+/*
+Opens and reads an entire wav file in a single operation.
+
+The return value is a heap-allocated buffer containing the audio data. Use drwav_free() to free the buffer.
+*/
+DRWAV_API drwav_int16* drwav_open_and_read_pcm_frames_s16(drwav_read_proc onRead, drwav_seek_proc onSeek, void* pUserData, unsigned int* channelsOut, unsigned int* sampleRateOut, drwav_uint64* totalFrameCountOut, const drwav_allocation_callbacks* pAllocationCallbacks);
+DRWAV_API float* drwav_open_and_read_pcm_frames_f32(drwav_read_proc onRead, drwav_seek_proc onSeek, void* pUserData, unsigned int* channelsOut, unsigned int* sampleRateOut, drwav_uint64* totalFrameCountOut, const drwav_allocation_callbacks* pAllocationCallbacks);
+DRWAV_API drwav_int32* drwav_open_and_read_pcm_frames_s32(drwav_read_proc onRead, drwav_seek_proc onSeek, void* pUserData, unsigned int* channelsOut, unsigned int* sampleRateOut, drwav_uint64* totalFrameCountOut, const drwav_allocation_callbacks* pAllocationCallbacks);
+#ifndef DR_WAV_NO_STDIO
+/*
+Opens and decodes an entire wav file in a single operation.
+
+The return value is a heap-allocated buffer containing the audio data. Use drwav_free() to free the buffer.
+*/
+DRWAV_API drwav_int16* drwav_open_file_and_read_pcm_frames_s16(const char* filename, unsigned int* channelsOut, unsigned int* sampleRateOut, drwav_uint64* totalFrameCountOut, const drwav_allocation_callbacks* pAllocationCallbacks);
+DRWAV_API float* drwav_open_file_and_read_pcm_frames_f32(const char* filename, unsigned int* channelsOut, unsigned int* sampleRateOut, drwav_uint64* totalFrameCountOut, const drwav_allocation_callbacks* pAllocationCallbacks);
+DRWAV_API drwav_int32* drwav_open_file_and_read_pcm_frames_s32(const char* filename, unsigned int* channelsOut, unsigned int* sampleRateOut, drwav_uint64* totalFrameCountOut, const drwav_allocation_callbacks* pAllocationCallbacks);
+DRWAV_API drwav_int16* drwav_open_file_and_read_pcm_frames_s16_w(const wchar_t* filename, unsigned int* channelsOut, unsigned int* sampleRateOut, drwav_uint64* totalFrameCountOut, const drwav_allocation_callbacks* pAllocationCallbacks);
+DRWAV_API float* drwav_open_file_and_read_pcm_frames_f32_w(const wchar_t* filename, unsigned int* channelsOut, unsigned int* sampleRateOut, drwav_uint64* totalFrameCountOut, const drwav_allocation_callbacks* pAllocationCallbacks);
+DRWAV_API drwav_int32* drwav_open_file_and_read_pcm_frames_s32_w(const wchar_t* filename, unsigned int* channelsOut, unsigned int* sampleRateOut, drwav_uint64* totalFrameCountOut, const drwav_allocation_callbacks* pAllocationCallbacks);
+#endif
+/*
+Opens and decodes an entire wav file from a block of memory in a single operation.
+
+The return value is a heap-allocated buffer containing the audio data. Use drwav_free() to free the buffer.
+*/
+DRWAV_API drwav_int16* drwav_open_memory_and_read_pcm_frames_s16(const void* data, size_t dataSize, unsigned int* channelsOut, unsigned int* sampleRateOut, drwav_uint64* totalFrameCountOut, const drwav_allocation_callbacks* pAllocationCallbacks);
+DRWAV_API float* drwav_open_memory_and_read_pcm_frames_f32(const void* data, size_t dataSize, unsigned int* channelsOut, unsigned int* sampleRateOut, drwav_uint64* totalFrameCountOut, const drwav_allocation_callbacks* pAllocationCallbacks);
+DRWAV_API drwav_int32* drwav_open_memory_and_read_pcm_frames_s32(const void* data, size_t dataSize, unsigned int* channelsOut, unsigned int* sampleRateOut, drwav_uint64* totalFrameCountOut, const drwav_allocation_callbacks* pAllocationCallbacks);
+#endif
+
+/* Frees data that was allocated internally by dr_wav. */
+DRWAV_API void drwav_free(void* p, const drwav_allocation_callbacks* pAllocationCallbacks);
+
+/* Converts bytes from a wav stream to a sized type of native endian. */
+DRWAV_API drwav_uint16 drwav_bytes_to_u16(const drwav_uint8* data);
+DRWAV_API drwav_int16 drwav_bytes_to_s16(const drwav_uint8* data);
+DRWAV_API drwav_uint32 drwav_bytes_to_u32(const drwav_uint8* data);
+DRWAV_API drwav_int32 drwav_bytes_to_s32(const drwav_uint8* data);
+DRWAV_API drwav_uint64 drwav_bytes_to_u64(const drwav_uint8* data);
+DRWAV_API drwav_int64 drwav_bytes_to_s64(const drwav_uint8* data);
+DRWAV_API float drwav_bytes_to_f32(const drwav_uint8* data);
+
+/* Compares a GUID for the purpose of checking the type of a Wave64 chunk. */
+DRWAV_API drwav_bool32 drwav_guid_equal(const drwav_uint8 a[16], const drwav_uint8 b[16]);
+
+/* Compares a four-character-code for the purpose of checking the type of a RIFF chunk. */
+DRWAV_API drwav_bool32 drwav_fourcc_equal(const drwav_uint8* a, const char* b);
+
+#ifdef __cplusplus
+}
+#endif
+#endif  /* dr_wav_h */
+
+
+/************************************************************************************************************************************************************
+ ************************************************************************************************************************************************************
+
+ IMPLEMENTATION
+
+ ************************************************************************************************************************************************************
+ ************************************************************************************************************************************************************/
+#if defined(DR_WAV_IMPLEMENTATION) || defined(DRWAV_IMPLEMENTATION)
+#ifndef dr_wav_c
+#define dr_wav_c
+
+#ifdef __MRC__
+/* MrC currently doesn't compile dr_wav correctly with any optimizations enabled. */
+#pragma options opt off
+#endif
+
+#include <stdlib.h>
+#include <string.h>
+#include <limits.h> /* For INT_MAX */
+
+#ifndef DR_WAV_NO_STDIO
+#include <stdio.h>
+#ifndef DR_WAV_NO_WCHAR
+#include <wchar.h>
+#endif
+#endif
+
+/* Standard library stuff. */
+#ifndef DRWAV_ASSERT
+#include <assert.h>
+#define DRWAV_ASSERT(expression)           assert(expression)
+#endif
+#ifndef DRWAV_MALLOC
+#define DRWAV_MALLOC(sz)                   malloc((sz))
+#endif
+#ifndef DRWAV_REALLOC
+#define DRWAV_REALLOC(p, sz)               realloc((p), (sz))
+#endif
+#ifndef DRWAV_FREE
+#define DRWAV_FREE(p)                      free((p))
+#endif
+#ifndef DRWAV_COPY_MEMORY
+#define DRWAV_COPY_MEMORY(dst, src, sz)    memcpy((dst), (src), (sz))
+#endif
+#ifndef DRWAV_ZERO_MEMORY
+#define DRWAV_ZERO_MEMORY(p, sz)           memset((p), 0, (sz))
+#endif
+#ifndef DRWAV_ZERO_OBJECT
+#define DRWAV_ZERO_OBJECT(p)               DRWAV_ZERO_MEMORY((p), sizeof(*p))
+#endif
+
+#define drwav_countof(x)                   (sizeof(x) / sizeof(x[0]))
+#define drwav_align(x, a)                  ((((x) + (a) - 1) / (a)) * (a))
+#define drwav_min(a, b)                    (((a) < (b)) ? (a) : (b))
+#define drwav_max(a, b)                    (((a) > (b)) ? (a) : (b))
+#define drwav_clamp(x, lo, hi)             (drwav_max((lo), drwav_min((hi), (x))))
+#define drwav_offset_ptr(p, offset)        (((drwav_uint8*)(p)) + (offset))
+
+#define DRWAV_MAX_SIMD_VECTOR_SIZE         32
+
+/* Architecture Detection */
+#if defined(__x86_64__) || defined(_M_X64)
+    #define DRWAV_X64
+#elif defined(__i386) || defined(_M_IX86)
+    #define DRWAV_X86
+#elif defined(__arm__) || defined(_M_ARM)
+    #define DRWAV_ARM
+#endif
+/* End Architecture Detection */
+
+/* Inline */
+#ifdef _MSC_VER
+    #define DRWAV_INLINE __forceinline
+#elif defined(__GNUC__)
+    /*
+    I've had a bug report where GCC is emitting warnings about functions possibly not being inlineable. This warning happens when
+    the __attribute__((always_inline)) attribute is defined without an "inline" statement. I think therefore there must be some
+    case where "__inline__" is not always defined, thus the compiler emitting these warnings. When using -std=c89 or -ansi on the
+    command line, we cannot use the "inline" keyword and instead need to use "__inline__". In an attempt to work around this issue
+    I am using "__inline__" only when we're compiling in strict ANSI mode.
+    */
+    #if defined(__STRICT_ANSI__)
+        #define DRWAV_GNUC_INLINE_HINT __inline__
+    #else
+        #define DRWAV_GNUC_INLINE_HINT inline
+    #endif
+
+    #if (__GNUC__ > 3 || (__GNUC__ == 3 && __GNUC_MINOR__ >= 2)) || defined(__clang__)
+        #define DRWAV_INLINE DRWAV_GNUC_INLINE_HINT __attribute__((always_inline))
+    #else
+        #define DRWAV_INLINE DRWAV_GNUC_INLINE_HINT
+    #endif
+#elif defined(__WATCOMC__)
+    #define DRWAV_INLINE __inline
+#else
+    #define DRWAV_INLINE
+#endif
+/* End Inline */
+
+/* SIZE_MAX */
+#if defined(SIZE_MAX)
+    #define DRWAV_SIZE_MAX  SIZE_MAX
+#else
+    #if defined(_WIN64) || defined(_LP64) || defined(__LP64__)
+        #define DRWAV_SIZE_MAX  ((drwav_uint64)0xFFFFFFFFFFFFFFFF)
+    #else
+        #define DRWAV_SIZE_MAX  0xFFFFFFFF
+    #endif
+#endif
+/* End SIZE_MAX */
+
+/* Weird bit manipulation is for C89 compatibility (no direct support for 64-bit integers). */
+#define DRWAV_INT64_MIN ((drwav_int64) ((drwav_uint64)0x80000000 << 32))
+#define DRWAV_INT64_MAX ((drwav_int64)(((drwav_uint64)0x7FFFFFFF << 32) | 0xFFFFFFFF))
+
+#if defined(_MSC_VER) && _MSC_VER >= 1400
+    #define DRWAV_HAS_BYTESWAP16_INTRINSIC
+    #define DRWAV_HAS_BYTESWAP32_INTRINSIC
+    #define DRWAV_HAS_BYTESWAP64_INTRINSIC
+#elif defined(__clang__)
+    #if defined(__has_builtin)
+        #if __has_builtin(__builtin_bswap16)
+            #define DRWAV_HAS_BYTESWAP16_INTRINSIC
+        #endif
+        #if __has_builtin(__builtin_bswap32)
+            #define DRWAV_HAS_BYTESWAP32_INTRINSIC
+        #endif
+        #if __has_builtin(__builtin_bswap64)
+            #define DRWAV_HAS_BYTESWAP64_INTRINSIC
+        #endif
+    #endif
+#elif defined(__GNUC__)
+    #if ((__GNUC__ > 4) || (__GNUC__ == 4 && __GNUC_MINOR__ >= 3))
+        #define DRWAV_HAS_BYTESWAP32_INTRINSIC
+        #define DRWAV_HAS_BYTESWAP64_INTRINSIC
+    #endif
+    #if ((__GNUC__ > 4) || (__GNUC__ == 4 && __GNUC_MINOR__ >= 8))
+        #define DRWAV_HAS_BYTESWAP16_INTRINSIC
+    #endif
+#endif
+
+DRWAV_API void drwav_version(drwav_uint32* pMajor, drwav_uint32* pMinor, drwav_uint32* pRevision)
+{
+    if (pMajor) {
+        *pMajor = DRWAV_VERSION_MAJOR;
+    }
+
+    if (pMinor) {
+        *pMinor = DRWAV_VERSION_MINOR;
+    }
+
+    if (pRevision) {
+        *pRevision = DRWAV_VERSION_REVISION;
+    }
+}
+
+DRWAV_API const char* drwav_version_string(void)
+{
+    return DRWAV_VERSION_STRING;
+}
+
+/*
+These limits are used for basic validation when initializing the decoder. If you exceed these limits, first of all: what on Earth are
+you doing?! (Let me know, I'd be curious!) Second, you can adjust these by #define-ing them before the dr_wav implementation.
+*/
+#ifndef DRWAV_MAX_SAMPLE_RATE
+#define DRWAV_MAX_SAMPLE_RATE       384000
+#endif
+#ifndef DRWAV_MAX_CHANNELS
+#define DRWAV_MAX_CHANNELS          256
+#endif
+#ifndef DRWAV_MAX_BITS_PER_SAMPLE
+#define DRWAV_MAX_BITS_PER_SAMPLE   64
+#endif
+
+static const drwav_uint8 drwavGUID_W64_RIFF[16] = {0x72,0x69,0x66,0x66, 0x2E,0x91, 0xCF,0x11, 0xA5,0xD6, 0x28,0xDB,0x04,0xC1,0x00,0x00};    /* 66666972-912E-11CF-A5D6-28DB04C10000 */
+static const drwav_uint8 drwavGUID_W64_WAVE[16] = {0x77,0x61,0x76,0x65, 0xF3,0xAC, 0xD3,0x11, 0x8C,0xD1, 0x00,0xC0,0x4F,0x8E,0xDB,0x8A};    /* 65766177-ACF3-11D3-8CD1-00C04F8EDB8A */
+/*static const drwav_uint8 drwavGUID_W64_JUNK[16] = {0x6A,0x75,0x6E,0x6B, 0xF3,0xAC, 0xD3,0x11, 0x8C,0xD1, 0x00,0xC0,0x4F,0x8E,0xDB,0x8A};*/    /* 6B6E756A-ACF3-11D3-8CD1-00C04F8EDB8A */
+static const drwav_uint8 drwavGUID_W64_FMT [16] = {0x66,0x6D,0x74,0x20, 0xF3,0xAC, 0xD3,0x11, 0x8C,0xD1, 0x00,0xC0,0x4F,0x8E,0xDB,0x8A};    /* 20746D66-ACF3-11D3-8CD1-00C04F8EDB8A */
+static const drwav_uint8 drwavGUID_W64_FACT[16] = {0x66,0x61,0x63,0x74, 0xF3,0xAC, 0xD3,0x11, 0x8C,0xD1, 0x00,0xC0,0x4F,0x8E,0xDB,0x8A};    /* 74636166-ACF3-11D3-8CD1-00C04F8EDB8A */
+static const drwav_uint8 drwavGUID_W64_DATA[16] = {0x64,0x61,0x74,0x61, 0xF3,0xAC, 0xD3,0x11, 0x8C,0xD1, 0x00,0xC0,0x4F,0x8E,0xDB,0x8A};    /* 61746164-ACF3-11D3-8CD1-00C04F8EDB8A */
+/*static const drwav_uint8 drwavGUID_W64_SMPL[16] = {0x73,0x6D,0x70,0x6C, 0xF3,0xAC, 0xD3,0x11, 0x8C,0xD1, 0x00,0xC0,0x4F,0x8E,0xDB,0x8A};*/    /* 6C706D73-ACF3-11D3-8CD1-00C04F8EDB8A */
+
+
+static DRWAV_INLINE int drwav__is_little_endian(void)
+{
+#if defined(DRWAV_X86) || defined(DRWAV_X64)
+    return DRWAV_TRUE;
+#elif defined(__BYTE_ORDER) && defined(__LITTLE_ENDIAN) && __BYTE_ORDER == __LITTLE_ENDIAN
+    return DRWAV_TRUE;
+#else
+    int n = 1;
+    return (*(char*)&n) == 1;
+#endif
+}
+
+
+static DRWAV_INLINE void drwav_bytes_to_guid(const drwav_uint8* data, drwav_uint8* guid)
+{
+    int i;
+    for (i = 0; i < 16; ++i) {
+        guid[i] = data[i];
+    }
+}
+
+
+static DRWAV_INLINE drwav_uint16 drwav__bswap16(drwav_uint16 n)
+{
+#ifdef DRWAV_HAS_BYTESWAP16_INTRINSIC
+    #if defined(_MSC_VER)
+        return _byteswap_ushort(n);
+    #elif defined(__GNUC__) || defined(__clang__)
+        return __builtin_bswap16(n);
+    #else
+        #error "This compiler does not support the byte swap intrinsic."
+    #endif
+#else
+    return ((n & 0xFF00) >> 8) |
+           ((n & 0x00FF) << 8);
+#endif
+}
+
+static DRWAV_INLINE drwav_uint32 drwav__bswap32(drwav_uint32 n)
+{
+#ifdef DRWAV_HAS_BYTESWAP32_INTRINSIC
+    #if defined(_MSC_VER)
+        return _byteswap_ulong(n);
+    #elif defined(__GNUC__) || defined(__clang__)
+        #if defined(DRWAV_ARM) && (defined(__ARM_ARCH) && __ARM_ARCH >= 6) && !defined(DRWAV_64BIT)   /* <-- 64-bit inline assembly has not been tested, so disabling for now. */
+            /* Inline assembly optimized implementation for ARM. In my testing, GCC does not generate optimized code with __builtin_bswap32(). */
+            drwav_uint32 r;
+            __asm__ __volatile__ (
+            #if defined(DRWAV_64BIT)
+                "rev %w[out], %w[in]" : [out]"=r"(r) : [in]"r"(n)   /* <-- This is untested. If someone in the community could test this, that would be appreciated! */
+            #else
+                "rev %[out], %[in]" : [out]"=r"(r) : [in]"r"(n)
+            #endif
+            );
+            return r;
+        #else
+            return __builtin_bswap32(n);
+        #endif
+    #else
+        #error "This compiler does not support the byte swap intrinsic."
+    #endif
+#else
+    return ((n & 0xFF000000) >> 24) |
+           ((n & 0x00FF0000) >>  8) |
+           ((n & 0x0000FF00) <<  8) |
+           ((n & 0x000000FF) << 24);
+#endif
+}
+
+static DRWAV_INLINE drwav_uint64 drwav__bswap64(drwav_uint64 n)
+{
+#ifdef DRWAV_HAS_BYTESWAP64_INTRINSIC
+    #if defined(_MSC_VER)
+        return _byteswap_uint64(n);
+    #elif defined(__GNUC__) || defined(__clang__)
+        return __builtin_bswap64(n);
+    #else
+        #error "This compiler does not support the byte swap intrinsic."
+    #endif
+#else
+    /* Weird "<< 32" bitshift is required for C89 because it doesn't support 64-bit constants. Should be optimized out by a good compiler. */
+    return ((n & ((drwav_uint64)0xFF000000 << 32)) >> 56) |
+           ((n & ((drwav_uint64)0x00FF0000 << 32)) >> 40) |
+           ((n & ((drwav_uint64)0x0000FF00 << 32)) >> 24) |
+           ((n & ((drwav_uint64)0x000000FF << 32)) >>  8) |
+           ((n & ((drwav_uint64)0xFF000000      )) <<  8) |
+           ((n & ((drwav_uint64)0x00FF0000      )) << 24) |
+           ((n & ((drwav_uint64)0x0000FF00      )) << 40) |
+           ((n & ((drwav_uint64)0x000000FF      )) << 56);
+#endif
+}
+
+
+static DRWAV_INLINE drwav_int16 drwav__bswap_s16(drwav_int16 n)
+{
+    return (drwav_int16)drwav__bswap16((drwav_uint16)n);
+}
+
+static DRWAV_INLINE void drwav__bswap_samples_s16(drwav_int16* pSamples, drwav_uint64 sampleCount)
+{
+    drwav_uint64 iSample;
+    for (iSample = 0; iSample < sampleCount; iSample += 1) {
+        pSamples[iSample] = drwav__bswap_s16(pSamples[iSample]);
+    }
+}
+
+
+static DRWAV_INLINE void drwav__bswap_s24(drwav_uint8* p)
+{
+    drwav_uint8 t;
+    t = p[0];
+    p[0] = p[2];
+    p[2] = t;
+}
+
+static DRWAV_INLINE void drwav__bswap_samples_s24(drwav_uint8* pSamples, drwav_uint64 sampleCount)
+{
+    drwav_uint64 iSample;
+    for (iSample = 0; iSample < sampleCount; iSample += 1) {
+        drwav_uint8* pSample = pSamples + (iSample*3);
+        drwav__bswap_s24(pSample);
+    }
+}
+
+
+static DRWAV_INLINE drwav_int32 drwav__bswap_s32(drwav_int32 n)
+{
+    return (drwav_int32)drwav__bswap32((drwav_uint32)n);
+}
+
+static DRWAV_INLINE void drwav__bswap_samples_s32(drwav_int32* pSamples, drwav_uint64 sampleCount)
+{
+    drwav_uint64 iSample;
+    for (iSample = 0; iSample < sampleCount; iSample += 1) {
+        pSamples[iSample] = drwav__bswap_s32(pSamples[iSample]);
+    }
+}
+
+
+static DRWAV_INLINE drwav_int64 drwav__bswap_s64(drwav_int64 n)
+{
+    return (drwav_int64)drwav__bswap64((drwav_uint64)n);
+}
+
+static DRWAV_INLINE void drwav__bswap_samples_s64(drwav_int64* pSamples, drwav_uint64 sampleCount)
+{
+    drwav_uint64 iSample;
+    for (iSample = 0; iSample < sampleCount; iSample += 1) {
+        pSamples[iSample] = drwav__bswap_s64(pSamples[iSample]);
+    }
+}
+
+
+static DRWAV_INLINE float drwav__bswap_f32(float n)
+{
+    union {
+        drwav_uint32 i;
+        float f;
+    } x;
+    x.f = n;
+    x.i = drwav__bswap32(x.i);
+
+    return x.f;
+}
+
+static DRWAV_INLINE void drwav__bswap_samples_f32(float* pSamples, drwav_uint64 sampleCount)
+{
+    drwav_uint64 iSample;
+    for (iSample = 0; iSample < sampleCount; iSample += 1) {
+        pSamples[iSample] = drwav__bswap_f32(pSamples[iSample]);
+    }
+}
+
+
+static DRWAV_INLINE void drwav__bswap_samples(void* pSamples, drwav_uint64 sampleCount, drwav_uint32 bytesPerSample)
+{
+    switch (bytesPerSample)
+    {
+        case 1:
+        {
+            /* No-op. */
+        } break;
+        case 2:
+        {
+            drwav__bswap_samples_s16((drwav_int16*)pSamples, sampleCount);
+        } break;
+        case 3:
+        {
+            drwav__bswap_samples_s24((drwav_uint8*)pSamples, sampleCount);
+        } break;
+        case 4:
+        {
+            drwav__bswap_samples_s32((drwav_int32*)pSamples, sampleCount);
+        } break;
+        case 8:
+        {
+            drwav__bswap_samples_s64((drwav_int64*)pSamples, sampleCount);
+        } break;
+        default:
+        {
+            /* Unsupported format. */
+            DRWAV_ASSERT(DRWAV_FALSE);
+        } break;
+    }
+}
+
+
+
+DRWAV_PRIVATE DRWAV_INLINE drwav_bool32 drwav_is_container_be(drwav_container container)
+{
+    if (container == drwav_container_rifx || container == drwav_container_aiff) {
+        return DRWAV_TRUE;
+    } else {
+        return DRWAV_FALSE;
+    }
+}
+
+
+DRWAV_PRIVATE DRWAV_INLINE drwav_uint16 drwav_bytes_to_u16_le(const drwav_uint8* data)
+{
+    return ((drwav_uint16)data[0] << 0) | ((drwav_uint16)data[1] << 8);
+}
+
+DRWAV_PRIVATE DRWAV_INLINE drwav_uint16 drwav_bytes_to_u16_be(const drwav_uint8* data)
+{
+    return ((drwav_uint16)data[1] << 0) | ((drwav_uint16)data[0] << 8);
+}
+
+DRWAV_PRIVATE DRWAV_INLINE drwav_uint16 drwav_bytes_to_u16_ex(const drwav_uint8* data, drwav_container container)
+{
+    if (drwav_is_container_be(container)) {
+        return drwav_bytes_to_u16_be(data);
+    } else {
+        return drwav_bytes_to_u16_le(data);
+    }
+}
+
+
+DRWAV_PRIVATE DRWAV_INLINE drwav_uint32 drwav_bytes_to_u32_le(const drwav_uint8* data)
+{
+    return ((drwav_uint32)data[0] << 0) | ((drwav_uint32)data[1] << 8) | ((drwav_uint32)data[2] << 16) | ((drwav_uint32)data[3] << 24);
+}
+
+DRWAV_PRIVATE DRWAV_INLINE drwav_uint32 drwav_bytes_to_u32_be(const drwav_uint8* data)
+{
+    return ((drwav_uint32)data[3] << 0) | ((drwav_uint32)data[2] << 8) | ((drwav_uint32)data[1] << 16) | ((drwav_uint32)data[0] << 24);
+}
+
+DRWAV_PRIVATE DRWAV_INLINE drwav_uint32 drwav_bytes_to_u32_ex(const drwav_uint8* data, drwav_container container)
+{
+    if (drwav_is_container_be(container)) {
+        return drwav_bytes_to_u32_be(data);
+    } else {
+        return drwav_bytes_to_u32_le(data);
+    }
+}
+
+
+
+DRWAV_PRIVATE drwav_int64 drwav_aiff_extented_to_s64(const drwav_uint8* data)
+{
+    drwav_uint32 exponent = ((drwav_uint32)data[0] << 8) | data[1];
+    drwav_uint64 hi = ((drwav_uint64)data[2] << 24) | ((drwav_uint64)data[3] << 16) | ((drwav_uint64)data[4] <<  8) | ((drwav_uint64)data[5] <<  0);
+    drwav_uint64 lo = ((drwav_uint64)data[6] << 24) | ((drwav_uint64)data[7] << 16) | ((drwav_uint64)data[8] <<  8) | ((drwav_uint64)data[9] <<  0);
+    drwav_uint64 significand = (hi << 32) | lo;
+    int sign = exponent >> 15;
+
+    /* Remove sign bit. */
+    exponent &= 0x7FFF;
+
+    /* Special cases. */
+    if (exponent == 0 && significand == 0) {
+        return 0;
+    } else if (exponent == 0x7FFF) {
+        return sign ? DRWAV_INT64_MIN : DRWAV_INT64_MAX;    /* Infinite. */
+    }
+
+    exponent -= 16383;
+
+    if (exponent > 63) {
+        return sign ? DRWAV_INT64_MIN : DRWAV_INT64_MAX;    /* Too big for a 64-bit integer. */
+    } else if (exponent < 1) {
+        return 0;  /* Number is less than 1, so rounds down to 0. */
+    }
+
+    significand >>= (63 - exponent);
+
+    if (sign) {
+        return -(drwav_int64)significand;
+    } else {
+        return  (drwav_int64)significand;
+    }
+}
+
+
+DRWAV_PRIVATE void* drwav__malloc_default(size_t sz, void* pUserData)
+{
+    (void)pUserData;
+    return DRWAV_MALLOC(sz);
+}
+
+DRWAV_PRIVATE void* drwav__realloc_default(void* p, size_t sz, void* pUserData)
+{
+    (void)pUserData;
+    return DRWAV_REALLOC(p, sz);
+}
+
+DRWAV_PRIVATE void drwav__free_default(void* p, void* pUserData)
+{
+    (void)pUserData;
+    DRWAV_FREE(p);
+}
+
+
+DRWAV_PRIVATE void* drwav__malloc_from_callbacks(size_t sz, const drwav_allocation_callbacks* pAllocationCallbacks)
+{
+    if (pAllocationCallbacks == NULL) {
+        return NULL;
+    }
+
+    if (pAllocationCallbacks->onMalloc != NULL) {
+        return pAllocationCallbacks->onMalloc(sz, pAllocationCallbacks->pUserData);
+    }
+
+    /* Try using realloc(). */
+    if (pAllocationCallbacks->onRealloc != NULL) {
+        return pAllocationCallbacks->onRealloc(NULL, sz, pAllocationCallbacks->pUserData);
+    }
+
+    return NULL;
+}
+
+DRWAV_PRIVATE void* drwav__realloc_from_callbacks(void* p, size_t szNew, size_t szOld, const drwav_allocation_callbacks* pAllocationCallbacks)
+{
+    if (pAllocationCallbacks == NULL) {
+        return NULL;
+    }
+
+    if (pAllocationCallbacks->onRealloc != NULL) {
+        return pAllocationCallbacks->onRealloc(p, szNew, pAllocationCallbacks->pUserData);
+    }
+
+    /* Try emulating realloc() in terms of malloc()/free(). */
+    if (pAllocationCallbacks->onMalloc != NULL && pAllocationCallbacks->onFree != NULL) {
+        void* p2;
+
+        p2 = pAllocationCallbacks->onMalloc(szNew, pAllocationCallbacks->pUserData);
+        if (p2 == NULL) {
+            return NULL;
+        }
+
+        if (p != NULL) {
+            DRWAV_COPY_MEMORY(p2, p, szOld);
+            pAllocationCallbacks->onFree(p, pAllocationCallbacks->pUserData);
+        }
+
+        return p2;
+    }
+
+    return NULL;
+}
+
+DRWAV_PRIVATE void drwav__free_from_callbacks(void* p, const drwav_allocation_callbacks* pAllocationCallbacks)
+{
+    if (p == NULL || pAllocationCallbacks == NULL) {
+        return;
+    }
+
+    if (pAllocationCallbacks->onFree != NULL) {
+        pAllocationCallbacks->onFree(p, pAllocationCallbacks->pUserData);
+    }
+}
+
+
+DRWAV_PRIVATE drwav_allocation_callbacks drwav_copy_allocation_callbacks_or_defaults(const drwav_allocation_callbacks* pAllocationCallbacks)
+{
+    if (pAllocationCallbacks != NULL) {
+        /* Copy. */
+        return *pAllocationCallbacks;
+    } else {
+        /* Defaults. */
+        drwav_allocation_callbacks allocationCallbacks;
+        allocationCallbacks.pUserData = NULL;
+        allocationCallbacks.onMalloc  = drwav__malloc_default;
+        allocationCallbacks.onRealloc = drwav__realloc_default;
+        allocationCallbacks.onFree    = drwav__free_default;
+        return allocationCallbacks;
+    }
+}
+
+
+static DRWAV_INLINE drwav_bool32 drwav__is_compressed_format_tag(drwav_uint16 formatTag)
+{
+    return
+        formatTag == DR_WAVE_FORMAT_ADPCM ||
+        formatTag == DR_WAVE_FORMAT_DVI_ADPCM;
+}
+
+DRWAV_PRIVATE unsigned int drwav__chunk_padding_size_riff(drwav_uint64 chunkSize)
+{
+    return (unsigned int)(chunkSize % 2);
+}
+
+DRWAV_PRIVATE unsigned int drwav__chunk_padding_size_w64(drwav_uint64 chunkSize)
+{
+    return (unsigned int)(chunkSize % 8);
+}
+
+DRWAV_PRIVATE drwav_uint64 drwav_read_pcm_frames_s16__msadpcm(drwav* pWav, drwav_uint64 samplesToRead, drwav_int16* pBufferOut);
+DRWAV_PRIVATE drwav_uint64 drwav_read_pcm_frames_s16__ima(drwav* pWav, drwav_uint64 samplesToRead, drwav_int16* pBufferOut);
+DRWAV_PRIVATE drwav_bool32 drwav_init_write__internal(drwav* pWav, const drwav_data_format* pFormat, drwav_uint64 totalSampleCount);
+
+DRWAV_PRIVATE drwav_result drwav__read_chunk_header(drwav_read_proc onRead, void* pUserData, drwav_container container, drwav_uint64* pRunningBytesReadOut, drwav_chunk_header* pHeaderOut)
+{
+    if (container == drwav_container_riff || container == drwav_container_rifx || container == drwav_container_rf64 || container == drwav_container_aiff) {
+        drwav_uint8 sizeInBytes[4];
+
+        if (onRead(pUserData, pHeaderOut->id.fourcc, 4) != 4) {
+            return DRWAV_AT_END;
+        }
+
+        if (onRead(pUserData, sizeInBytes, 4) != 4) {
+            return DRWAV_INVALID_FILE;
+        }
+
+        pHeaderOut->sizeInBytes = drwav_bytes_to_u32_ex(sizeInBytes, container);
+        pHeaderOut->paddingSize = drwav__chunk_padding_size_riff(pHeaderOut->sizeInBytes);
+
+        *pRunningBytesReadOut += 8;
+    } else if (container == drwav_container_w64) {
+        drwav_uint8 sizeInBytes[8];
+
+        if (onRead(pUserData, pHeaderOut->id.guid, 16) != 16) {
+            return DRWAV_AT_END;
+        }
+
+        if (onRead(pUserData, sizeInBytes, 8) != 8) {
+            return DRWAV_INVALID_FILE;
+        }
+
+        pHeaderOut->sizeInBytes = drwav_bytes_to_u64(sizeInBytes) - 24;    /* <-- Subtract 24 because w64 includes the size of the header. */
+        pHeaderOut->paddingSize = drwav__chunk_padding_size_w64(pHeaderOut->sizeInBytes);
+        *pRunningBytesReadOut += 24;
+    } else {
+        return DRWAV_INVALID_FILE;
+    }
+
+    return DRWAV_SUCCESS;
+}
+
+DRWAV_PRIVATE drwav_bool32 drwav__seek_forward(drwav_seek_proc onSeek, drwav_uint64 offset, void* pUserData)
+{
+    drwav_uint64 bytesRemainingToSeek = offset;
+    while (bytesRemainingToSeek > 0) {
+        if (bytesRemainingToSeek > 0x7FFFFFFF) {
+            if (!onSeek(pUserData, 0x7FFFFFFF, drwav_seek_origin_current)) {
+                return DRWAV_FALSE;
+            }
+            bytesRemainingToSeek -= 0x7FFFFFFF;
+        } else {
+            if (!onSeek(pUserData, (int)bytesRemainingToSeek, drwav_seek_origin_current)) {
+                return DRWAV_FALSE;
+            }
+            bytesRemainingToSeek = 0;
+        }
+    }
+
+    return DRWAV_TRUE;
+}
+
+DRWAV_PRIVATE drwav_bool32 drwav__seek_from_start(drwav_seek_proc onSeek, drwav_uint64 offset, void* pUserData)
+{
+    if (offset <= 0x7FFFFFFF) {
+        return onSeek(pUserData, (int)offset, drwav_seek_origin_start);
+    }
+
+    /* Larger than 32-bit seek. */
+    if (!onSeek(pUserData, 0x7FFFFFFF, drwav_seek_origin_start)) {
+        return DRWAV_FALSE;
+    }
+    offset -= 0x7FFFFFFF;
+
+    for (;;) {
+        if (offset <= 0x7FFFFFFF) {
+            return onSeek(pUserData, (int)offset, drwav_seek_origin_current);
+        }
+
+        if (!onSeek(pUserData, 0x7FFFFFFF, drwav_seek_origin_current)) {
+            return DRWAV_FALSE;
+        }
+        offset -= 0x7FFFFFFF;
+    }
+
+    /* Should never get here. */
+    /*return DRWAV_TRUE; */
+}
+
+
+
+DRWAV_PRIVATE size_t drwav__on_read(drwav_read_proc onRead, void* pUserData, void* pBufferOut, size_t bytesToRead, drwav_uint64* pCursor)
+{
+    size_t bytesRead;
+
+    DRWAV_ASSERT(onRead != NULL);
+    DRWAV_ASSERT(pCursor != NULL);
+
+    bytesRead = onRead(pUserData, pBufferOut, bytesToRead);
+    *pCursor += bytesRead;
+    return bytesRead;
+}
+
+#if 0
+DRWAV_PRIVATE drwav_bool32 drwav__on_seek(drwav_seek_proc onSeek, void* pUserData, int offset, drwav_seek_origin origin, drwav_uint64* pCursor)
+{
+    DRWAV_ASSERT(onSeek != NULL);
+    DRWAV_ASSERT(pCursor != NULL);
+
+    if (!onSeek(pUserData, offset, origin)) {
+        return DRWAV_FALSE;
+    }
+
+    if (origin == drwav_seek_origin_start) {
+        *pCursor = offset;
+    } else {
+        *pCursor += offset;
+    }
+
+    return DRWAV_TRUE;
+}
+#endif
+
+
+#define DRWAV_SMPL_BYTES                    36
+#define DRWAV_SMPL_LOOP_BYTES               24
+#define DRWAV_INST_BYTES                    7
+#define DRWAV_ACID_BYTES                    24
+#define DRWAV_CUE_BYTES                     4
+#define DRWAV_BEXT_BYTES                    602
+#define DRWAV_BEXT_DESCRIPTION_BYTES        256
+#define DRWAV_BEXT_ORIGINATOR_NAME_BYTES    32
+#define DRWAV_BEXT_ORIGINATOR_REF_BYTES     32
+#define DRWAV_BEXT_RESERVED_BYTES           180
+#define DRWAV_BEXT_UMID_BYTES               64
+#define DRWAV_CUE_POINT_BYTES               24
+#define DRWAV_LIST_LABEL_OR_NOTE_BYTES      4
+#define DRWAV_LIST_LABELLED_TEXT_BYTES      20
+
+#define DRWAV_METADATA_ALIGNMENT            8
+
+typedef enum
+{
+    drwav__metadata_parser_stage_count,
+    drwav__metadata_parser_stage_read
+} drwav__metadata_parser_stage;
+
+typedef struct
+{
+    drwav_read_proc onRead;
+    drwav_seek_proc onSeek;
+    void *pReadSeekUserData;
+    drwav__metadata_parser_stage stage;
+    drwav_metadata *pMetadata;
+    drwav_uint32 metadataCount;
+    drwav_uint8 *pData;
+    drwav_uint8 *pDataCursor;
+    drwav_uint64 metadataCursor;
+    drwav_uint64 extraCapacity;
+} drwav__metadata_parser;
+
+DRWAV_PRIVATE size_t drwav__metadata_memory_capacity(drwav__metadata_parser* pParser)
+{
+    drwav_uint64 cap = sizeof(drwav_metadata) * (drwav_uint64)pParser->metadataCount + pParser->extraCapacity;
+    if (cap > DRWAV_SIZE_MAX) {
+        return 0;   /* Too big. */
+    }
+
+    return (size_t)cap; /* Safe cast thanks to the check above. */
+}
+
+DRWAV_PRIVATE drwav_uint8* drwav__metadata_get_memory(drwav__metadata_parser* pParser, size_t size, size_t align)
+{
+    drwav_uint8* pResult;
+
+    if (align) {
+        drwav_uintptr modulo = (drwav_uintptr)pParser->pDataCursor % align;
+        if (modulo != 0) {
+            pParser->pDataCursor += align - modulo;
+        }
+    }
+    
+    pResult = pParser->pDataCursor;
+
+    /*
+    Getting to the point where this function is called means there should always be memory
+    available. Out of memory checks should have been done at an earlier stage.
+    */
+    DRWAV_ASSERT((pResult + size) <= (pParser->pData + drwav__metadata_memory_capacity(pParser)));
+
+    pParser->pDataCursor += size;
+    return pResult;
+}
+
+DRWAV_PRIVATE void drwav__metadata_request_extra_memory_for_stage_2(drwav__metadata_parser* pParser, size_t bytes, size_t align)
+{
+    size_t extra = bytes + (align ? (align - 1) : 0);
+    pParser->extraCapacity += extra;
+}
+
+DRWAV_PRIVATE drwav_result drwav__metadata_alloc(drwav__metadata_parser* pParser, drwav_allocation_callbacks* pAllocationCallbacks)
+{
+    if (pParser->extraCapacity != 0 || pParser->metadataCount != 0) {
+        pAllocationCallbacks->onFree(pParser->pData, pAllocationCallbacks->pUserData);
+
+        pParser->pData = (drwav_uint8*)pAllocationCallbacks->onMalloc(drwav__metadata_memory_capacity(pParser), pAllocationCallbacks->pUserData);
+        pParser->pDataCursor = pParser->pData;
+
+        if (pParser->pData == NULL) {
+            return DRWAV_OUT_OF_MEMORY;
+        }
+
+        /*
+        We don't need to worry about specifying an alignment here because malloc always returns something
+        of suitable alignment. This also means pParser->pMetadata is all that we need to store in order
+        for us to free when we are done.
+        */
+        pParser->pMetadata = (drwav_metadata*)drwav__metadata_get_memory(pParser, sizeof(drwav_metadata) * pParser->metadataCount, 1);
+        pParser->metadataCursor = 0;
+    }
+
+    return DRWAV_SUCCESS;
+}
+
+DRWAV_PRIVATE size_t drwav__metadata_parser_read(drwav__metadata_parser* pParser, void* pBufferOut, size_t bytesToRead, drwav_uint64* pCursor)
+{
+    if (pCursor != NULL) {
+        return drwav__on_read(pParser->onRead, pParser->pReadSeekUserData, pBufferOut, bytesToRead, pCursor);
+    } else {
+        return pParser->onRead(pParser->pReadSeekUserData, pBufferOut, bytesToRead);
+    }
+}
+
+DRWAV_PRIVATE drwav_uint64 drwav__read_smpl_to_metadata_obj(drwav__metadata_parser* pParser, const drwav_chunk_header* pChunkHeader, drwav_metadata* pMetadata)
+{
+    drwav_uint8 smplHeaderData[DRWAV_SMPL_BYTES];
+    drwav_uint64 totalBytesRead = 0;
+    size_t bytesJustRead;
+
+    if (pMetadata == NULL) {
+        return 0;
+    }
+
+    bytesJustRead = drwav__metadata_parser_read(pParser, smplHeaderData, sizeof(smplHeaderData), &totalBytesRead);
+
+    DRWAV_ASSERT(pParser->stage == drwav__metadata_parser_stage_read);
+    DRWAV_ASSERT(pChunkHeader != NULL);
+
+    if (pMetadata != NULL && bytesJustRead == sizeof(smplHeaderData)) {
+        drwav_uint32 iSampleLoop;
+
+        pMetadata->type                                     = drwav_metadata_type_smpl;
+        pMetadata->data.smpl.manufacturerId                 = drwav_bytes_to_u32(smplHeaderData + 0);
+        pMetadata->data.smpl.productId                      = drwav_bytes_to_u32(smplHeaderData + 4);
+        pMetadata->data.smpl.samplePeriodNanoseconds        = drwav_bytes_to_u32(smplHeaderData + 8);
+        pMetadata->data.smpl.midiUnityNote                  = drwav_bytes_to_u32(smplHeaderData + 12);
+        pMetadata->data.smpl.midiPitchFraction              = drwav_bytes_to_u32(smplHeaderData + 16);
+        pMetadata->data.smpl.smpteFormat                    = drwav_bytes_to_u32(smplHeaderData + 20);
+        pMetadata->data.smpl.smpteOffset                    = drwav_bytes_to_u32(smplHeaderData + 24);
+        pMetadata->data.smpl.sampleLoopCount                = drwav_bytes_to_u32(smplHeaderData + 28);
+        pMetadata->data.smpl.samplerSpecificDataSizeInBytes = drwav_bytes_to_u32(smplHeaderData + 32);
+
+        /*
+        The loop count needs to be validated against the size of the chunk for safety so we don't
+        attempt to read over the boundary of the chunk.
+        */
+        if (pMetadata->data.smpl.sampleLoopCount == (pChunkHeader->sizeInBytes - DRWAV_SMPL_BYTES) / DRWAV_SMPL_LOOP_BYTES) {
+            pMetadata->data.smpl.pLoops = (drwav_smpl_loop*)drwav__metadata_get_memory(pParser, sizeof(drwav_smpl_loop) * pMetadata->data.smpl.sampleLoopCount, DRWAV_METADATA_ALIGNMENT);
+
+            for (iSampleLoop = 0; iSampleLoop < pMetadata->data.smpl.sampleLoopCount; ++iSampleLoop) {
+                drwav_uint8 smplLoopData[DRWAV_SMPL_LOOP_BYTES];
+                bytesJustRead = drwav__metadata_parser_read(pParser, smplLoopData, sizeof(smplLoopData), &totalBytesRead);
+
+                if (bytesJustRead == sizeof(smplLoopData)) {
+                    pMetadata->data.smpl.pLoops[iSampleLoop].cuePointId            = drwav_bytes_to_u32(smplLoopData + 0);
+                    pMetadata->data.smpl.pLoops[iSampleLoop].type                  = drwav_bytes_to_u32(smplLoopData + 4);
+                    pMetadata->data.smpl.pLoops[iSampleLoop].firstSampleByteOffset = drwav_bytes_to_u32(smplLoopData + 8);
+                    pMetadata->data.smpl.pLoops[iSampleLoop].lastSampleByteOffset  = drwav_bytes_to_u32(smplLoopData + 12);
+                    pMetadata->data.smpl.pLoops[iSampleLoop].sampleFraction        = drwav_bytes_to_u32(smplLoopData + 16);
+                    pMetadata->data.smpl.pLoops[iSampleLoop].playCount             = drwav_bytes_to_u32(smplLoopData + 20);
+                } else {
+                    break;
+                }
+            }
+
+            if (pMetadata->data.smpl.samplerSpecificDataSizeInBytes > 0) {
+                pMetadata->data.smpl.pSamplerSpecificData = drwav__metadata_get_memory(pParser, pMetadata->data.smpl.samplerSpecificDataSizeInBytes, 1);
+                DRWAV_ASSERT(pMetadata->data.smpl.pSamplerSpecificData != NULL);
+
+                drwav__metadata_parser_read(pParser, pMetadata->data.smpl.pSamplerSpecificData, pMetadata->data.smpl.samplerSpecificDataSizeInBytes, &totalBytesRead);
+            }
+        }
+    }
+
+    return totalBytesRead;
+}
+
+DRWAV_PRIVATE drwav_uint64 drwav__read_cue_to_metadata_obj(drwav__metadata_parser* pParser, const drwav_chunk_header* pChunkHeader, drwav_metadata* pMetadata)
+{
+    drwav_uint8 cueHeaderSectionData[DRWAV_CUE_BYTES];
+    drwav_uint64 totalBytesRead = 0;
+    size_t bytesJustRead;
+
+    if (pMetadata == NULL) {
+        return 0;
+    }
+
+    bytesJustRead = drwav__metadata_parser_read(pParser, cueHeaderSectionData, sizeof(cueHeaderSectionData), &totalBytesRead);
+
+    DRWAV_ASSERT(pParser->stage == drwav__metadata_parser_stage_read);
+
+    if (bytesJustRead == sizeof(cueHeaderSectionData)) {
+        pMetadata->type                   = drwav_metadata_type_cue;
+        pMetadata->data.cue.cuePointCount = drwav_bytes_to_u32(cueHeaderSectionData);
+
+        /*
+        We need to validate the cue point count against the size of the chunk so we don't read
+        beyond the chunk.
+        */
+        if (pMetadata->data.cue.cuePointCount == (pChunkHeader->sizeInBytes - DRWAV_CUE_BYTES) / DRWAV_CUE_POINT_BYTES) {
+            pMetadata->data.cue.pCuePoints    = (drwav_cue_point*)drwav__metadata_get_memory(pParser, sizeof(drwav_cue_point) * pMetadata->data.cue.cuePointCount, DRWAV_METADATA_ALIGNMENT);
+            DRWAV_ASSERT(pMetadata->data.cue.pCuePoints != NULL);
+
+            if (pMetadata->data.cue.cuePointCount > 0) {
+                drwav_uint32 iCuePoint;
+
+                for (iCuePoint = 0; iCuePoint < pMetadata->data.cue.cuePointCount; ++iCuePoint) {
+                    drwav_uint8 cuePointData[DRWAV_CUE_POINT_BYTES];
+                    bytesJustRead = drwav__metadata_parser_read(pParser, cuePointData, sizeof(cuePointData), &totalBytesRead);
+
+                    if (bytesJustRead == sizeof(cuePointData)) {
+                        pMetadata->data.cue.pCuePoints[iCuePoint].id                = drwav_bytes_to_u32(cuePointData + 0);
+                        pMetadata->data.cue.pCuePoints[iCuePoint].playOrderPosition = drwav_bytes_to_u32(cuePointData + 4);
+                        pMetadata->data.cue.pCuePoints[iCuePoint].dataChunkId[0]    = cuePointData[8];
+                        pMetadata->data.cue.pCuePoints[iCuePoint].dataChunkId[1]    = cuePointData[9];
+                        pMetadata->data.cue.pCuePoints[iCuePoint].dataChunkId[2]    = cuePointData[10];
+                        pMetadata->data.cue.pCuePoints[iCuePoint].dataChunkId[3]    = cuePointData[11];
+                        pMetadata->data.cue.pCuePoints[iCuePoint].chunkStart        = drwav_bytes_to_u32(cuePointData + 12);
+                        pMetadata->data.cue.pCuePoints[iCuePoint].blockStart        = drwav_bytes_to_u32(cuePointData + 16);
+                        pMetadata->data.cue.pCuePoints[iCuePoint].sampleByteOffset  = drwav_bytes_to_u32(cuePointData + 20);
+                    } else {
+                        break;
+                    }
+                }
+            }
+        }
+    }
+
+    return totalBytesRead;
+}
+
+DRWAV_PRIVATE drwav_uint64 drwav__read_inst_to_metadata_obj(drwav__metadata_parser* pParser, drwav_metadata* pMetadata)
+{
+    drwav_uint8 instData[DRWAV_INST_BYTES];
+    drwav_uint64 bytesRead;
+
+    if (pMetadata == NULL) {
+        return 0;
+    }
+
+    bytesRead = drwav__metadata_parser_read(pParser, instData, sizeof(instData), NULL);
+
+    DRWAV_ASSERT(pParser->stage == drwav__metadata_parser_stage_read);
+
+    if (bytesRead == sizeof(instData)) {
+        pMetadata->type                    = drwav_metadata_type_inst;
+        pMetadata->data.inst.midiUnityNote = (drwav_int8)instData[0];
+        pMetadata->data.inst.fineTuneCents = (drwav_int8)instData[1];
+        pMetadata->data.inst.gainDecibels  = (drwav_int8)instData[2];
+        pMetadata->data.inst.lowNote       = (drwav_int8)instData[3];
+        pMetadata->data.inst.highNote      = (drwav_int8)instData[4];
+        pMetadata->data.inst.lowVelocity   = (drwav_int8)instData[5];
+        pMetadata->data.inst.highVelocity  = (drwav_int8)instData[6];
+    }
+
+    return bytesRead;
+}
+
+DRWAV_PRIVATE drwav_uint64 drwav__read_acid_to_metadata_obj(drwav__metadata_parser* pParser, drwav_metadata* pMetadata)
+{
+    drwav_uint8 acidData[DRWAV_ACID_BYTES];
+    drwav_uint64 bytesRead;
+
+    if (pMetadata == NULL) {
+        return 0;
+    }
+
+    bytesRead = drwav__metadata_parser_read(pParser, acidData, sizeof(acidData), NULL);
+
+    DRWAV_ASSERT(pParser->stage == drwav__metadata_parser_stage_read);
+
+    if (bytesRead == sizeof(acidData)) {
+        pMetadata->type                       = drwav_metadata_type_acid;
+        pMetadata->data.acid.flags            = drwav_bytes_to_u32(acidData + 0);
+        pMetadata->data.acid.midiUnityNote    = drwav_bytes_to_u16(acidData + 4);
+        pMetadata->data.acid.reserved1        = drwav_bytes_to_u16(acidData + 6);
+        pMetadata->data.acid.reserved2        = drwav_bytes_to_f32(acidData + 8);
+        pMetadata->data.acid.numBeats         = drwav_bytes_to_u32(acidData + 12);
+        pMetadata->data.acid.meterDenominator = drwav_bytes_to_u16(acidData + 16);
+        pMetadata->data.acid.meterNumerator   = drwav_bytes_to_u16(acidData + 18);
+        pMetadata->data.acid.tempo            = drwav_bytes_to_f32(acidData + 20);
+    }
+
+    return bytesRead;
+}
+
+DRWAV_PRIVATE size_t drwav__strlen(const char* str)
+{
+    size_t result = 0;
+
+    while (*str++) {
+        result += 1;
+    }
+
+    return result;
+}
+
+DRWAV_PRIVATE size_t drwav__strlen_clamped(const char* str, size_t maxToRead)
+{
+    size_t result = 0;
+
+    while (*str++ && result < maxToRead) {
+        result += 1;
+    }
+
+    return result;
+}
+
+DRWAV_PRIVATE char* drwav__metadata_copy_string(drwav__metadata_parser* pParser, const char* str, size_t maxToRead)
+{
+    size_t len = drwav__strlen_clamped(str, maxToRead);
+
+    if (len) {
+        char* result = (char*)drwav__metadata_get_memory(pParser, len + 1, 1);
+        DRWAV_ASSERT(result != NULL);
+
+        DRWAV_COPY_MEMORY(result, str, len);
+        result[len] = '\0';
+
+        return result;
+    } else {
+        return NULL;
+    }
+}
+
+typedef struct
+{
+    const void* pBuffer;
+    size_t sizeInBytes;
+    size_t cursor;
+} drwav_buffer_reader;
+
+DRWAV_PRIVATE drwav_result drwav_buffer_reader_init(const void* pBuffer, size_t sizeInBytes, drwav_buffer_reader* pReader)
+{
+    DRWAV_ASSERT(pBuffer != NULL);
+    DRWAV_ASSERT(pReader != NULL);
+
+    DRWAV_ZERO_OBJECT(pReader);
+
+    pReader->pBuffer     = pBuffer;
+    pReader->sizeInBytes = sizeInBytes;
+    pReader->cursor      = 0;
+
+    return DRWAV_SUCCESS;
+}
+
+DRWAV_PRIVATE const void* drwav_buffer_reader_ptr(const drwav_buffer_reader* pReader)
+{
+    DRWAV_ASSERT(pReader != NULL);
+
+    return drwav_offset_ptr(pReader->pBuffer, pReader->cursor);
+}
+
+DRWAV_PRIVATE drwav_result drwav_buffer_reader_seek(drwav_buffer_reader* pReader, size_t bytesToSeek)
+{
+    DRWAV_ASSERT(pReader != NULL);
+
+    if (pReader->cursor + bytesToSeek > pReader->sizeInBytes) {
+        return DRWAV_BAD_SEEK;  /* Seeking too far forward. */
+    }
+
+    pReader->cursor += bytesToSeek;
+
+    return DRWAV_SUCCESS;
+}
+
+DRWAV_PRIVATE drwav_result drwav_buffer_reader_read(drwav_buffer_reader* pReader, void* pDst, size_t bytesToRead, size_t* pBytesRead)
+{
+    drwav_result result = DRWAV_SUCCESS;
+    size_t bytesRemaining;
+
+    DRWAV_ASSERT(pReader != NULL);
+    
+    if (pBytesRead != NULL) {
+        *pBytesRead = 0;
+    }
+
+    bytesRemaining = (pReader->sizeInBytes - pReader->cursor);
+    if (bytesToRead > bytesRemaining) {
+        bytesToRead = bytesRemaining;
+    }
+
+    if (pDst == NULL) {
+        /* Seek. */
+        result = drwav_buffer_reader_seek(pReader, bytesToRead);
+    } else {
+        /* Read. */
+        DRWAV_COPY_MEMORY(pDst, drwav_buffer_reader_ptr(pReader), bytesToRead);
+        pReader->cursor += bytesToRead;
+    }
+
+    DRWAV_ASSERT(pReader->cursor <= pReader->sizeInBytes);
+
+    if (result == DRWAV_SUCCESS) {
+        if (pBytesRead != NULL) {
+            *pBytesRead = bytesToRead;
+        }
+    }
+
+    return DRWAV_SUCCESS;
+}
+
+DRWAV_PRIVATE drwav_result drwav_buffer_reader_read_u16(drwav_buffer_reader* pReader, drwav_uint16* pDst)
+{
+    drwav_result result;
+    size_t bytesRead;
+    drwav_uint8 data[2];
+
+    DRWAV_ASSERT(pReader != NULL);
+    DRWAV_ASSERT(pDst != NULL);
+
+    *pDst = 0;  /* Safety. */
+
+    result = drwav_buffer_reader_read(pReader, data, sizeof(*pDst), &bytesRead);
+    if (result != DRWAV_SUCCESS || bytesRead != sizeof(*pDst)) {
+        return result;
+    }
+
+    *pDst = drwav_bytes_to_u16(data);
+
+    return DRWAV_SUCCESS;
+}
+
+DRWAV_PRIVATE drwav_result drwav_buffer_reader_read_u32(drwav_buffer_reader* pReader, drwav_uint32* pDst)
+{
+    drwav_result result;
+    size_t bytesRead;
+    drwav_uint8 data[4];
+
+    DRWAV_ASSERT(pReader != NULL);
+    DRWAV_ASSERT(pDst != NULL);
+
+    *pDst = 0;  /* Safety. */
+
+    result = drwav_buffer_reader_read(pReader, data, sizeof(*pDst), &bytesRead);
+    if (result != DRWAV_SUCCESS || bytesRead != sizeof(*pDst)) {
+        return result;
+    }
+
+    *pDst = drwav_bytes_to_u32(data);
+
+    return DRWAV_SUCCESS;
+}
+
+
+
+DRWAV_PRIVATE drwav_uint64 drwav__read_bext_to_metadata_obj(drwav__metadata_parser* pParser, drwav_metadata* pMetadata, drwav_uint64 chunkSize)
+{
+    drwav_uint8 bextData[DRWAV_BEXT_BYTES];
+    size_t bytesRead = drwav__metadata_parser_read(pParser, bextData, sizeof(bextData), NULL);
+
+    DRWAV_ASSERT(pParser->stage == drwav__metadata_parser_stage_read);
+    
+    if (bytesRead == sizeof(bextData)) {
+        drwav_buffer_reader reader;
+        drwav_uint32 timeReferenceLow;
+        drwav_uint32 timeReferenceHigh;
+        size_t extraBytes;
+
+        pMetadata->type = drwav_metadata_type_bext;
+
+        if (drwav_buffer_reader_init(bextData, bytesRead, &reader) == DRWAV_SUCCESS) {
+            pMetadata->data.bext.pDescription = drwav__metadata_copy_string(pParser, (const char*)drwav_buffer_reader_ptr(&reader), DRWAV_BEXT_DESCRIPTION_BYTES);
+            drwav_buffer_reader_seek(&reader, DRWAV_BEXT_DESCRIPTION_BYTES);
+
+            pMetadata->data.bext.pOriginatorName = drwav__metadata_copy_string(pParser, (const char*)drwav_buffer_reader_ptr(&reader), DRWAV_BEXT_ORIGINATOR_NAME_BYTES);
+            drwav_buffer_reader_seek(&reader, DRWAV_BEXT_ORIGINATOR_NAME_BYTES);
+
+            pMetadata->data.bext.pOriginatorReference = drwav__metadata_copy_string(pParser, (const char*)drwav_buffer_reader_ptr(&reader), DRWAV_BEXT_ORIGINATOR_REF_BYTES);
+            drwav_buffer_reader_seek(&reader, DRWAV_BEXT_ORIGINATOR_REF_BYTES);
+
+            drwav_buffer_reader_read(&reader, pMetadata->data.bext.pOriginationDate, sizeof(pMetadata->data.bext.pOriginationDate), NULL);
+            drwav_buffer_reader_read(&reader, pMetadata->data.bext.pOriginationTime, sizeof(pMetadata->data.bext.pOriginationTime), NULL);
+
+            drwav_buffer_reader_read_u32(&reader, &timeReferenceLow);
+            drwav_buffer_reader_read_u32(&reader, &timeReferenceHigh);
+            pMetadata->data.bext.timeReference = ((drwav_uint64)timeReferenceHigh << 32) + timeReferenceLow;
+
+            drwav_buffer_reader_read_u16(&reader, &pMetadata->data.bext.version);
+
+            pMetadata->data.bext.pUMID = drwav__metadata_get_memory(pParser, DRWAV_BEXT_UMID_BYTES, 1);
+            drwav_buffer_reader_read(&reader, pMetadata->data.bext.pUMID, DRWAV_BEXT_UMID_BYTES, NULL);
+
+            drwav_buffer_reader_read_u16(&reader, &pMetadata->data.bext.loudnessValue);
+            drwav_buffer_reader_read_u16(&reader, &pMetadata->data.bext.loudnessRange);
+            drwav_buffer_reader_read_u16(&reader, &pMetadata->data.bext.maxTruePeakLevel);
+            drwav_buffer_reader_read_u16(&reader, &pMetadata->data.bext.maxMomentaryLoudness);
+            drwav_buffer_reader_read_u16(&reader, &pMetadata->data.bext.maxShortTermLoudness);
+
+            DRWAV_ASSERT((drwav_offset_ptr(drwav_buffer_reader_ptr(&reader), DRWAV_BEXT_RESERVED_BYTES)) == (bextData + DRWAV_BEXT_BYTES));
+
+            extraBytes = (size_t)(chunkSize - DRWAV_BEXT_BYTES);
+            if (extraBytes > 0) {
+                pMetadata->data.bext.pCodingHistory = (char*)drwav__metadata_get_memory(pParser, extraBytes + 1, 1);
+                DRWAV_ASSERT(pMetadata->data.bext.pCodingHistory != NULL);
+
+                bytesRead += drwav__metadata_parser_read(pParser, pMetadata->data.bext.pCodingHistory, extraBytes, NULL);
+                pMetadata->data.bext.codingHistorySize = (drwav_uint32)drwav__strlen(pMetadata->data.bext.pCodingHistory);
+            } else {
+                pMetadata->data.bext.pCodingHistory    = NULL;
+                pMetadata->data.bext.codingHistorySize = 0;
+            }
+        }
+    }
+
+    return bytesRead;
+}
+
+DRWAV_PRIVATE drwav_uint64 drwav__read_list_label_or_note_to_metadata_obj(drwav__metadata_parser* pParser, drwav_metadata* pMetadata, drwav_uint64 chunkSize, drwav_metadata_type type)
+{
+    drwav_uint8 cueIDBuffer[DRWAV_LIST_LABEL_OR_NOTE_BYTES];
+    drwav_uint64 totalBytesRead = 0;
+    size_t bytesJustRead = drwav__metadata_parser_read(pParser, cueIDBuffer, sizeof(cueIDBuffer), &totalBytesRead);
+
+    DRWAV_ASSERT(pParser->stage == drwav__metadata_parser_stage_read);    
+
+    if (bytesJustRead == sizeof(cueIDBuffer)) {
+        drwav_uint32 sizeIncludingNullTerminator;
+
+        pMetadata->type = type;
+        pMetadata->data.labelOrNote.cuePointId = drwav_bytes_to_u32(cueIDBuffer);
+
+        sizeIncludingNullTerminator = (drwav_uint32)chunkSize - DRWAV_LIST_LABEL_OR_NOTE_BYTES;
+        if (sizeIncludingNullTerminator > 0) {
+            pMetadata->data.labelOrNote.stringLength = sizeIncludingNullTerminator - 1;
+            pMetadata->data.labelOrNote.pString      = (char*)drwav__metadata_get_memory(pParser, sizeIncludingNullTerminator, 1);
+            DRWAV_ASSERT(pMetadata->data.labelOrNote.pString != NULL);
+
+            drwav__metadata_parser_read(pParser, pMetadata->data.labelOrNote.pString, sizeIncludingNullTerminator, &totalBytesRead);
+        } else {
+            pMetadata->data.labelOrNote.stringLength = 0;
+            pMetadata->data.labelOrNote.pString      = NULL;
+        }
+    }
+
+    return totalBytesRead;
+}
+
+DRWAV_PRIVATE drwav_uint64 drwav__read_list_labelled_cue_region_to_metadata_obj(drwav__metadata_parser* pParser, drwav_metadata* pMetadata, drwav_uint64 chunkSize)
+{
+    drwav_uint8 buffer[DRWAV_LIST_LABELLED_TEXT_BYTES];
+    drwav_uint64 totalBytesRead = 0;
+    size_t bytesJustRead = drwav__metadata_parser_read(pParser, buffer, sizeof(buffer), &totalBytesRead);
+
+    DRWAV_ASSERT(pParser->stage == drwav__metadata_parser_stage_read);
+
+    if (bytesJustRead == sizeof(buffer)) {
+        drwav_uint32 sizeIncludingNullTerminator;
+
+        pMetadata->type                                = drwav_metadata_type_list_labelled_cue_region;
+        pMetadata->data.labelledCueRegion.cuePointId   = drwav_bytes_to_u32(buffer + 0);
+        pMetadata->data.labelledCueRegion.sampleLength = drwav_bytes_to_u32(buffer + 4);
+        pMetadata->data.labelledCueRegion.purposeId[0] = buffer[8];
+        pMetadata->data.labelledCueRegion.purposeId[1] = buffer[9];
+        pMetadata->data.labelledCueRegion.purposeId[2] = buffer[10];
+        pMetadata->data.labelledCueRegion.purposeId[3] = buffer[11];
+        pMetadata->data.labelledCueRegion.country      = drwav_bytes_to_u16(buffer + 12);
+        pMetadata->data.labelledCueRegion.language     = drwav_bytes_to_u16(buffer + 14);
+        pMetadata->data.labelledCueRegion.dialect      = drwav_bytes_to_u16(buffer + 16);
+        pMetadata->data.labelledCueRegion.codePage     = drwav_bytes_to_u16(buffer + 18);
+
+        sizeIncludingNullTerminator = (drwav_uint32)chunkSize - DRWAV_LIST_LABELLED_TEXT_BYTES;
+        if (sizeIncludingNullTerminator > 0) {
+            pMetadata->data.labelledCueRegion.stringLength = sizeIncludingNullTerminator - 1;
+            pMetadata->data.labelledCueRegion.pString      = (char*)drwav__metadata_get_memory(pParser, sizeIncludingNullTerminator, 1);
+            DRWAV_ASSERT(pMetadata->data.labelledCueRegion.pString != NULL);
+
+            drwav__metadata_parser_read(pParser, pMetadata->data.labelledCueRegion.pString, sizeIncludingNullTerminator, &totalBytesRead);
+        } else {
+            pMetadata->data.labelledCueRegion.stringLength = 0;
+            pMetadata->data.labelledCueRegion.pString      = NULL;
+        }
+    }
+
+    return totalBytesRead;
+}
+
+DRWAV_PRIVATE drwav_uint64 drwav__metadata_process_info_text_chunk(drwav__metadata_parser* pParser, drwav_uint64 chunkSize, drwav_metadata_type type)
+{
+    drwav_uint64 bytesRead = 0;
+    drwav_uint32 stringSizeWithNullTerminator = (drwav_uint32)chunkSize;
+
+    if (pParser->stage == drwav__metadata_parser_stage_count) {
+        pParser->metadataCount += 1;
+        drwav__metadata_request_extra_memory_for_stage_2(pParser, stringSizeWithNullTerminator, 1);
+    } else {
+        drwav_metadata* pMetadata = &pParser->pMetadata[pParser->metadataCursor];
+        pMetadata->type = type;
+        if (stringSizeWithNullTerminator > 0) {
+            pMetadata->data.infoText.stringLength = stringSizeWithNullTerminator - 1;
+            pMetadata->data.infoText.pString = (char*)drwav__metadata_get_memory(pParser, stringSizeWithNullTerminator, 1);
+            DRWAV_ASSERT(pMetadata->data.infoText.pString != NULL);
+
+            bytesRead = drwav__metadata_parser_read(pParser, pMetadata->data.infoText.pString, (size_t)stringSizeWithNullTerminator, NULL);
+            if (bytesRead == chunkSize) {
+                pParser->metadataCursor += 1;
+            } else {
+                /* Failed to parse. */
+            }
+        } else {
+            pMetadata->data.infoText.stringLength = 0;
+            pMetadata->data.infoText.pString      = NULL;
+            pParser->metadataCursor += 1;
+        }
+    }
+
+    return bytesRead;
+}
+
+DRWAV_PRIVATE drwav_uint64 drwav__metadata_process_unknown_chunk(drwav__metadata_parser* pParser, const drwav_uint8* pChunkId, drwav_uint64 chunkSize, drwav_metadata_location location)
+{
+    drwav_uint64 bytesRead = 0;
+
+    if (location == drwav_metadata_location_invalid) {
+        return 0;
+    }
+
+    if (drwav_fourcc_equal(pChunkId, "data") || drwav_fourcc_equal(pChunkId, "fmt ") || drwav_fourcc_equal(pChunkId, "fact")) {
+        return 0;
+    }
+
+    if (pParser->stage == drwav__metadata_parser_stage_count) {
+        pParser->metadataCount += 1;
+        drwav__metadata_request_extra_memory_for_stage_2(pParser, (size_t)chunkSize, 1);
+    } else {
+        drwav_metadata* pMetadata = &pParser->pMetadata[pParser->metadataCursor];
+        pMetadata->type                         = drwav_metadata_type_unknown;
+        pMetadata->data.unknown.chunkLocation   = location;
+        pMetadata->data.unknown.id[0]           = pChunkId[0];
+        pMetadata->data.unknown.id[1]           = pChunkId[1];
+        pMetadata->data.unknown.id[2]           = pChunkId[2];
+        pMetadata->data.unknown.id[3]           = pChunkId[3];
+        pMetadata->data.unknown.dataSizeInBytes = (drwav_uint32)chunkSize;
+        pMetadata->data.unknown.pData           = (drwav_uint8 *)drwav__metadata_get_memory(pParser, (size_t)chunkSize, 1);
+        DRWAV_ASSERT(pMetadata->data.unknown.pData != NULL);
+
+        bytesRead = drwav__metadata_parser_read(pParser, pMetadata->data.unknown.pData, pMetadata->data.unknown.dataSizeInBytes, NULL);
+        if (bytesRead == pMetadata->data.unknown.dataSizeInBytes) {
+            pParser->metadataCursor += 1;
+        } else {
+            /* Failed to read. */
+        }
+    }
+
+    return bytesRead;
+}
+
+DRWAV_PRIVATE drwav_bool32 drwav__chunk_matches(drwav_metadata_type allowedMetadataTypes, const drwav_uint8* pChunkID, drwav_metadata_type type, const char* pID)
+{
+    return (allowedMetadataTypes & type) && drwav_fourcc_equal(pChunkID, pID);
+}
+
+DRWAV_PRIVATE drwav_uint64 drwav__metadata_process_chunk(drwav__metadata_parser* pParser, const drwav_chunk_header* pChunkHeader, drwav_metadata_type allowedMetadataTypes)
+{
+    const drwav_uint8 *pChunkID = pChunkHeader->id.fourcc;
+    drwav_uint64 bytesRead = 0;
+
+    if (drwav__chunk_matches(allowedMetadataTypes, pChunkID, drwav_metadata_type_smpl, "smpl")) {
+        if (pChunkHeader->sizeInBytes >= DRWAV_SMPL_BYTES) {
+            if (pParser->stage == drwav__metadata_parser_stage_count) {
+                drwav_uint8 buffer[4];
+                size_t bytesJustRead;
+
+                if (!pParser->onSeek(pParser->pReadSeekUserData, 28, drwav_seek_origin_current)) {
+                    return bytesRead;
+                }
+                bytesRead += 28;
+
+                bytesJustRead = drwav__metadata_parser_read(pParser, buffer, sizeof(buffer), &bytesRead);
+                if (bytesJustRead == sizeof(buffer)) {
+                    drwav_uint32 loopCount = drwav_bytes_to_u32(buffer);
+                    drwav_uint64 calculatedLoopCount;
+
+                    /* The loop count must be validated against the size of the chunk. */
+                    calculatedLoopCount = (pChunkHeader->sizeInBytes - DRWAV_SMPL_BYTES) / DRWAV_SMPL_LOOP_BYTES;
+                    if (calculatedLoopCount == loopCount) {
+                        bytesJustRead = drwav__metadata_parser_read(pParser, buffer, sizeof(buffer), &bytesRead);
+                        if (bytesJustRead == sizeof(buffer)) {
+                            drwav_uint32 samplerSpecificDataSizeInBytes = drwav_bytes_to_u32(buffer);
+
+                            pParser->metadataCount += 1;
+                            drwav__metadata_request_extra_memory_for_stage_2(pParser, sizeof(drwav_smpl_loop) * loopCount, DRWAV_METADATA_ALIGNMENT);
+                            drwav__metadata_request_extra_memory_for_stage_2(pParser, samplerSpecificDataSizeInBytes, 1);
+                        }
+                    } else {
+                        /* Loop count in header does not match the size of the chunk. */
+                    }                    
+                }
+            } else {
+                bytesRead = drwav__read_smpl_to_metadata_obj(pParser, pChunkHeader, &pParser->pMetadata[pParser->metadataCursor]);
+                if (bytesRead == pChunkHeader->sizeInBytes) {
+                    pParser->metadataCursor += 1;
+                } else {
+                    /* Failed to parse. */
+                }
+            }
+        } else {
+            /* Incorrectly formed chunk. */
+        }
+    } else if (drwav__chunk_matches(allowedMetadataTypes, pChunkID, drwav_metadata_type_inst, "inst")) {
+        if (pChunkHeader->sizeInBytes == DRWAV_INST_BYTES) {
+            if (pParser->stage == drwav__metadata_parser_stage_count) {
+                pParser->metadataCount += 1;
+            } else {
+                bytesRead = drwav__read_inst_to_metadata_obj(pParser, &pParser->pMetadata[pParser->metadataCursor]);
+                if (bytesRead == pChunkHeader->sizeInBytes) {
+                    pParser->metadataCursor += 1;
+                } else {
+                    /* Failed to parse. */
+                }
+            }
+        } else {
+            /* Incorrectly formed chunk. */
+        }
+    } else if (drwav__chunk_matches(allowedMetadataTypes, pChunkID, drwav_metadata_type_acid, "acid")) {
+        if (pChunkHeader->sizeInBytes == DRWAV_ACID_BYTES) {
+            if (pParser->stage == drwav__metadata_parser_stage_count) {
+                pParser->metadataCount += 1;
+            } else {
+                bytesRead = drwav__read_acid_to_metadata_obj(pParser, &pParser->pMetadata[pParser->metadataCursor]);
+                if (bytesRead == pChunkHeader->sizeInBytes) {
+                    pParser->metadataCursor += 1;
+                } else {
+                    /* Failed to parse. */
+                }
+            }
+        } else {
+            /* Incorrectly formed chunk. */
+        }
+    } else if (drwav__chunk_matches(allowedMetadataTypes, pChunkID, drwav_metadata_type_cue, "cue ")) {
+        if (pChunkHeader->sizeInBytes >= DRWAV_CUE_BYTES) {
+            if (pParser->stage == drwav__metadata_parser_stage_count) {
+                size_t cueCount;
+
+                pParser->metadataCount += 1;
+                cueCount = (size_t)(pChunkHeader->sizeInBytes - DRWAV_CUE_BYTES) / DRWAV_CUE_POINT_BYTES;
+                drwav__metadata_request_extra_memory_for_stage_2(pParser, sizeof(drwav_cue_point) * cueCount, DRWAV_METADATA_ALIGNMENT);
+            } else {
+                bytesRead = drwav__read_cue_to_metadata_obj(pParser, pChunkHeader, &pParser->pMetadata[pParser->metadataCursor]);
+                if (bytesRead == pChunkHeader->sizeInBytes) {
+                    pParser->metadataCursor += 1;
+                } else {
+                    /* Failed to parse. */
+                }
+            }
+        } else {
+            /* Incorrectly formed chunk. */
+        }
+    } else if (drwav__chunk_matches(allowedMetadataTypes, pChunkID, drwav_metadata_type_bext, "bext")) {
+        if (pChunkHeader->sizeInBytes >= DRWAV_BEXT_BYTES) {
+            if (pParser->stage == drwav__metadata_parser_stage_count) {
+                /* The description field is the largest one in a bext chunk, so that is the max size of this temporary buffer. */
+                char buffer[DRWAV_BEXT_DESCRIPTION_BYTES + 1];
+                size_t allocSizeNeeded = DRWAV_BEXT_UMID_BYTES; /* We know we will need SMPTE umid size. */
+                size_t bytesJustRead;
+
+                buffer[DRWAV_BEXT_DESCRIPTION_BYTES] = '\0';
+                bytesJustRead = drwav__metadata_parser_read(pParser, buffer, DRWAV_BEXT_DESCRIPTION_BYTES, &bytesRead);
+                if (bytesJustRead != DRWAV_BEXT_DESCRIPTION_BYTES) {
+                    return bytesRead;
+                }
+                allocSizeNeeded += drwav__strlen(buffer) + 1;
+
+                buffer[DRWAV_BEXT_ORIGINATOR_NAME_BYTES] = '\0';
+                bytesJustRead = drwav__metadata_parser_read(pParser, buffer, DRWAV_BEXT_ORIGINATOR_NAME_BYTES, &bytesRead);
+                if (bytesJustRead != DRWAV_BEXT_ORIGINATOR_NAME_BYTES) {
+                    return bytesRead;
+                }
+                allocSizeNeeded += drwav__strlen(buffer) + 1;
+
+                buffer[DRWAV_BEXT_ORIGINATOR_REF_BYTES] = '\0';
+                bytesJustRead = drwav__metadata_parser_read(pParser, buffer, DRWAV_BEXT_ORIGINATOR_REF_BYTES, &bytesRead);
+                if (bytesJustRead != DRWAV_BEXT_ORIGINATOR_REF_BYTES) {
+                    return bytesRead;
+                }
+                allocSizeNeeded += drwav__strlen(buffer) + 1;
+                allocSizeNeeded += (size_t)pChunkHeader->sizeInBytes - DRWAV_BEXT_BYTES; /* Coding history. */
+
+                drwav__metadata_request_extra_memory_for_stage_2(pParser, allocSizeNeeded, 1);
+
+                pParser->metadataCount += 1;
+            } else {
+                bytesRead = drwav__read_bext_to_metadata_obj(pParser, &pParser->pMetadata[pParser->metadataCursor], pChunkHeader->sizeInBytes);
+                if (bytesRead == pChunkHeader->sizeInBytes) {
+                    pParser->metadataCursor += 1;
+                } else {
+                    /* Failed to parse. */
+                }
+            }
+        } else {
+            /* Incorrectly formed chunk. */
+        }
+    } else if (drwav_fourcc_equal(pChunkID, "LIST") || drwav_fourcc_equal(pChunkID, "list")) {
+        drwav_metadata_location listType = drwav_metadata_location_invalid;
+        while (bytesRead < pChunkHeader->sizeInBytes) {
+            drwav_uint8 subchunkId[4];
+            drwav_uint8 subchunkSizeBuffer[4];
+            drwav_uint64 subchunkDataSize;
+            drwav_uint64 subchunkBytesRead = 0;
+            drwav_uint64 bytesJustRead = drwav__metadata_parser_read(pParser, subchunkId, sizeof(subchunkId), &bytesRead);
+            if (bytesJustRead != sizeof(subchunkId)) {
+                break;
+            }
+
+            /*
+            The first thing in a list chunk should be "adtl" or "INFO".
+
+              - adtl means this list is a Associated Data List Chunk and will contain labels, notes
+                or labelled cue regions.
+              - INFO means this list is an Info List Chunk containing info text chunks such as IPRD
+                which would specifies the album of this wav file.
+
+            No data follows the adtl or INFO id so we just make note of what type this list is and
+            continue.
+            */
+            if (drwav_fourcc_equal(subchunkId, "adtl")) {
+                listType = drwav_metadata_location_inside_adtl_list;
+                continue;
+            } else if (drwav_fourcc_equal(subchunkId, "INFO")) {
+                listType = drwav_metadata_location_inside_info_list;
+                continue;
+            }
+
+            bytesJustRead = drwav__metadata_parser_read(pParser, subchunkSizeBuffer, sizeof(subchunkSizeBuffer), &bytesRead);
+            if (bytesJustRead != sizeof(subchunkSizeBuffer)) {
+                break;
+            }
+            subchunkDataSize = drwav_bytes_to_u32(subchunkSizeBuffer);
+
+            if (drwav__chunk_matches(allowedMetadataTypes, subchunkId, drwav_metadata_type_list_label, "labl") || drwav__chunk_matches(allowedMetadataTypes, subchunkId, drwav_metadata_type_list_note, "note")) {
+                if (subchunkDataSize >= DRWAV_LIST_LABEL_OR_NOTE_BYTES) {
+                    drwav_uint64 stringSizeWithNullTerm = subchunkDataSize - DRWAV_LIST_LABEL_OR_NOTE_BYTES;
+                    if (pParser->stage == drwav__metadata_parser_stage_count) {
+                        pParser->metadataCount += 1;
+                        drwav__metadata_request_extra_memory_for_stage_2(pParser, (size_t)stringSizeWithNullTerm, 1);
+                    } else {
+                        subchunkBytesRead = drwav__read_list_label_or_note_to_metadata_obj(pParser, &pParser->pMetadata[pParser->metadataCursor], subchunkDataSize, drwav_fourcc_equal(subchunkId, "labl") ? drwav_metadata_type_list_label : drwav_metadata_type_list_note);
+                        if (subchunkBytesRead == subchunkDataSize) {
+                            pParser->metadataCursor += 1;
+                        } else {
+                            /* Failed to parse. */
+                        }
+                    }
+                } else {
+                    /* Incorrectly formed chunk. */
+                }
+            } else if (drwav__chunk_matches(allowedMetadataTypes, subchunkId, drwav_metadata_type_list_labelled_cue_region, "ltxt")) {
+                if (subchunkDataSize >= DRWAV_LIST_LABELLED_TEXT_BYTES) {
+                    drwav_uint64 stringSizeWithNullTerminator = subchunkDataSize - DRWAV_LIST_LABELLED_TEXT_BYTES;
+                    if (pParser->stage == drwav__metadata_parser_stage_count) {
+                        pParser->metadataCount += 1;
+                        drwav__metadata_request_extra_memory_for_stage_2(pParser, (size_t)stringSizeWithNullTerminator, 1);
+                    } else {
+                        subchunkBytesRead = drwav__read_list_labelled_cue_region_to_metadata_obj(pParser, &pParser->pMetadata[pParser->metadataCursor], subchunkDataSize);
+                        if (subchunkBytesRead == subchunkDataSize) {
+                            pParser->metadataCursor += 1;
+                        } else {
+                            /* Failed to parse. */
+                        }
+                    }
+                } else {
+                    /* Incorrectly formed chunk. */
+                }
+            } else if (drwav__chunk_matches(allowedMetadataTypes, subchunkId, drwav_metadata_type_list_info_software, "ISFT")) {
+                subchunkBytesRead = drwav__metadata_process_info_text_chunk(pParser, subchunkDataSize,  drwav_metadata_type_list_info_software);
+            } else if (drwav__chunk_matches(allowedMetadataTypes, subchunkId, drwav_metadata_type_list_info_copyright, "ICOP")) {
+                subchunkBytesRead = drwav__metadata_process_info_text_chunk(pParser, subchunkDataSize,  drwav_metadata_type_list_info_copyright);
+            } else if (drwav__chunk_matches(allowedMetadataTypes, subchunkId, drwav_metadata_type_list_info_title, "INAM")) {
+                subchunkBytesRead = drwav__metadata_process_info_text_chunk(pParser, subchunkDataSize,  drwav_metadata_type_list_info_title);
+            } else if (drwav__chunk_matches(allowedMetadataTypes, subchunkId, drwav_metadata_type_list_info_artist, "IART")) {
+                subchunkBytesRead = drwav__metadata_process_info_text_chunk(pParser, subchunkDataSize,  drwav_metadata_type_list_info_artist);
+            } else if (drwav__chunk_matches(allowedMetadataTypes, subchunkId, drwav_metadata_type_list_info_comment, "ICMT")) {
+                subchunkBytesRead = drwav__metadata_process_info_text_chunk(pParser, subchunkDataSize,  drwav_metadata_type_list_info_comment);
+            } else if (drwav__chunk_matches(allowedMetadataTypes, subchunkId, drwav_metadata_type_list_info_date, "ICRD")) {
+                subchunkBytesRead = drwav__metadata_process_info_text_chunk(pParser, subchunkDataSize,  drwav_metadata_type_list_info_date);
+            } else if (drwav__chunk_matches(allowedMetadataTypes, subchunkId, drwav_metadata_type_list_info_genre, "IGNR")) {
+                subchunkBytesRead = drwav__metadata_process_info_text_chunk(pParser, subchunkDataSize,  drwav_metadata_type_list_info_genre);
+            } else if (drwav__chunk_matches(allowedMetadataTypes, subchunkId, drwav_metadata_type_list_info_album, "IPRD")) {
+                subchunkBytesRead = drwav__metadata_process_info_text_chunk(pParser, subchunkDataSize,  drwav_metadata_type_list_info_album);
+            } else if (drwav__chunk_matches(allowedMetadataTypes, subchunkId, drwav_metadata_type_list_info_tracknumber, "ITRK")) {
+                subchunkBytesRead = drwav__metadata_process_info_text_chunk(pParser, subchunkDataSize,  drwav_metadata_type_list_info_tracknumber);
+            } else if ((allowedMetadataTypes & drwav_metadata_type_unknown) != 0) {
+                subchunkBytesRead = drwav__metadata_process_unknown_chunk(pParser, subchunkId, subchunkDataSize, listType);
+            }
+
+            bytesRead += subchunkBytesRead;
+            DRWAV_ASSERT(subchunkBytesRead <= subchunkDataSize);
+
+            if (subchunkBytesRead < subchunkDataSize) {
+                drwav_uint64 bytesToSeek = subchunkDataSize - subchunkBytesRead;
+
+                if (!pParser->onSeek(pParser->pReadSeekUserData, (int)bytesToSeek, drwav_seek_origin_current)) {
+                    break;
+                }
+                bytesRead += bytesToSeek;
+            }
+
+            if ((subchunkDataSize % 2) == 1) {
+                if (!pParser->onSeek(pParser->pReadSeekUserData, 1, drwav_seek_origin_current)) {
+                    break;
+                }
+                bytesRead += 1;
+            }
+        }
+    } else if ((allowedMetadataTypes & drwav_metadata_type_unknown) != 0) {
+        bytesRead = drwav__metadata_process_unknown_chunk(pParser, pChunkID, pChunkHeader->sizeInBytes, drwav_metadata_location_top_level);
+    }
+
+    return bytesRead;
+}
+
+
+DRWAV_PRIVATE drwav_uint32 drwav_get_bytes_per_pcm_frame(drwav* pWav)
+{
+    drwav_uint32 bytesPerFrame;
+
+    /*
+    The bytes per frame is a bit ambiguous. It can be either be based on the bits per sample, or the block align. The way I'm doing it here
+    is that if the bits per sample is a multiple of 8, use floor(bitsPerSample*channels/8), otherwise fall back to the block align.
+    */
+    if ((pWav->bitsPerSample & 0x7) == 0) {
+        /* Bits per sample is a multiple of 8. */
+        bytesPerFrame = (pWav->bitsPerSample * pWav->fmt.channels) >> 3;
+    } else {
+        bytesPerFrame = pWav->fmt.blockAlign;
+    }
+
+    /* Validation for known formats. a-law and mu-law should be 1 byte per channel. If it's not, it's not decodable. */
+    if (pWav->translatedFormatTag == DR_WAVE_FORMAT_ALAW || pWav->translatedFormatTag == DR_WAVE_FORMAT_MULAW) {
+        if (bytesPerFrame != pWav->fmt.channels) {
+            return 0;   /* Invalid file. */
+        }
+    }
+
+    return bytesPerFrame;
+}
+
+DRWAV_API drwav_uint16 drwav_fmt_get_format(const drwav_fmt* pFMT)
+{
+    if (pFMT == NULL) {
+        return 0;
+    }
+
+    if (pFMT->formatTag != DR_WAVE_FORMAT_EXTENSIBLE) {
+        return pFMT->formatTag;
+    } else {
+        return drwav_bytes_to_u16(pFMT->subFormat);    /* Only the first two bytes are required. */
+    }
+}
+
+DRWAV_PRIVATE drwav_bool32 drwav_preinit(drwav* pWav, drwav_read_proc onRead, drwav_seek_proc onSeek, void* pReadSeekUserData, const drwav_allocation_callbacks* pAllocationCallbacks)
+{
+    if (pWav == NULL || onRead == NULL || onSeek == NULL) {
+        return DRWAV_FALSE;
+    }
+
+    DRWAV_ZERO_MEMORY(pWav, sizeof(*pWav));
+    pWav->onRead    = onRead;
+    pWav->onSeek    = onSeek;
+    pWav->pUserData = pReadSeekUserData;
+    pWav->allocationCallbacks = drwav_copy_allocation_callbacks_or_defaults(pAllocationCallbacks);
+
+    if (pWav->allocationCallbacks.onFree == NULL || (pWav->allocationCallbacks.onMalloc == NULL && pWav->allocationCallbacks.onRealloc == NULL)) {
+        return DRWAV_FALSE;    /* Invalid allocation callbacks. */
+    }
+
+    return DRWAV_TRUE;
+}
+
+DRWAV_PRIVATE drwav_bool32 drwav_init__internal(drwav* pWav, drwav_chunk_proc onChunk, void* pChunkUserData, drwav_uint32 flags)
+{
+    /* This function assumes drwav_preinit() has been called beforehand. */
+    drwav_result result;
+    drwav_uint64 cursor;    /* <-- Keeps track of the byte position so we can seek to specific locations. */
+    drwav_bool32 sequential;
+    drwav_uint8 riff[4];
+    drwav_fmt fmt;
+    unsigned short translatedFormatTag;
+    drwav_uint64 dataChunkSize = 0;             /* <-- Important! Don't explicitly set this to 0 anywhere else. Calculation of the size of the data chunk is performed in different paths depending on the container. */
+    drwav_uint64 sampleCountFromFactChunk = 0;  /* Same as dataChunkSize - make sure this is the only place this is initialized to 0. */
+    drwav_uint64 metadataStartPos;
+    drwav__metadata_parser metadataParser;
+    drwav_bool8 isProcessingMetadata = DRWAV_FALSE;
+    drwav_bool8 foundChunk_fmt  = DRWAV_FALSE;
+    drwav_bool8 foundChunk_data = DRWAV_FALSE;
+    drwav_bool8 isAIFCFormType = DRWAV_FALSE;   /* Only used with AIFF. */
+    drwav_uint64 aiffFrameCount = 0;
+
+    cursor = 0;
+    sequential = (flags & DRWAV_SEQUENTIAL) != 0;
+    DRWAV_ZERO_OBJECT(&fmt);
+
+    /* The first 4 bytes should be the RIFF identifier. */
+    if (drwav__on_read(pWav->onRead, pWav->pUserData, riff, sizeof(riff), &cursor) != sizeof(riff)) {
+        return DRWAV_FALSE;
+    }
+
+    /*
+    The first 4 bytes can be used to identify the container. For RIFF files it will start with "RIFF" and for
+    w64 it will start with "riff".
+    */
+    if (drwav_fourcc_equal(riff, "RIFF")) {
+        pWav->container = drwav_container_riff;
+    } else if (drwav_fourcc_equal(riff, "RIFX")) {
+        pWav->container = drwav_container_rifx;
+    } else if (drwav_fourcc_equal(riff, "riff")) {
+        int i;
+        drwav_uint8 riff2[12];
+
+        pWav->container = drwav_container_w64;
+
+        /* Check the rest of the GUID for validity. */
+        if (drwav__on_read(pWav->onRead, pWav->pUserData, riff2, sizeof(riff2), &cursor) != sizeof(riff2)) {
+            return DRWAV_FALSE;
+        }
+
+        for (i = 0; i < 12; ++i) {
+            if (riff2[i] != drwavGUID_W64_RIFF[i+4]) {
+                return DRWAV_FALSE;
+            }
+        }
+    } else if (drwav_fourcc_equal(riff, "RF64")) {
+        pWav->container = drwav_container_rf64;
+    } else if (drwav_fourcc_equal(riff, "FORM")) {
+        pWav->container = drwav_container_aiff;
+    } else {
+        return DRWAV_FALSE;   /* Unknown or unsupported container. */
+    }
+
+
+    if (pWav->container == drwav_container_riff || pWav->container == drwav_container_rifx || pWav->container == drwav_container_rf64) {
+        drwav_uint8 chunkSizeBytes[4];
+        drwav_uint8 wave[4];
+
+        if (drwav__on_read(pWav->onRead, pWav->pUserData, chunkSizeBytes, sizeof(chunkSizeBytes), &cursor) != sizeof(chunkSizeBytes)) {
+            return DRWAV_FALSE;
+        }
+
+        if (pWav->container == drwav_container_riff || pWav->container == drwav_container_rifx) {
+            if (drwav_bytes_to_u32_ex(chunkSizeBytes, pWav->container) < 36) {
+                /*
+                I've had a report of a WAV file failing to load when the size of the WAVE chunk is not encoded
+                and is instead just set to 0. I'm going to relax the validation here to allow these files to
+                load. Considering the chunk size isn't actually used this should be safe. With this change my
+                test suite still passes.
+                */
+                /*return DRWAV_FALSE;*/    /* Chunk size should always be at least 36 bytes. */
+            }
+        } else if (pWav->container == drwav_container_rf64) {
+            if (drwav_bytes_to_u32_le(chunkSizeBytes) != 0xFFFFFFFF) {
+                return DRWAV_FALSE;    /* Chunk size should always be set to -1/0xFFFFFFFF for RF64. The actual size is retrieved later. */
+            }
+        } else {
+            return DRWAV_FALSE; /* Should never hit this. */
+        }
+
+        if (drwav__on_read(pWav->onRead, pWav->pUserData, wave, sizeof(wave), &cursor) != sizeof(wave)) {
+            return DRWAV_FALSE;
+        }
+
+        if (!drwav_fourcc_equal(wave, "WAVE")) {
+            return DRWAV_FALSE;    /* Expecting "WAVE". */
+        }
+    } else if (pWav->container == drwav_container_w64) {
+        drwav_uint8 chunkSizeBytes[8];
+        drwav_uint8 wave[16];
+
+        if (drwav__on_read(pWav->onRead, pWav->pUserData, chunkSizeBytes, sizeof(chunkSizeBytes), &cursor) != sizeof(chunkSizeBytes)) {
+            return DRWAV_FALSE;
+        }
+
+        if (drwav_bytes_to_u64(chunkSizeBytes) < 80) {
+            return DRWAV_FALSE;
+        }
+
+        if (drwav__on_read(pWav->onRead, pWav->pUserData, wave, sizeof(wave), &cursor) != sizeof(wave)) {
+            return DRWAV_FALSE;
+        }
+
+        if (!drwav_guid_equal(wave, drwavGUID_W64_WAVE)) {
+            return DRWAV_FALSE;
+        }
+    } else if (pWav->container == drwav_container_aiff) {
+        drwav_uint8 chunkSizeBytes[4];
+        drwav_uint8 aiff[4];
+
+        if (drwav__on_read(pWav->onRead, pWav->pUserData, chunkSizeBytes, sizeof(chunkSizeBytes), &cursor) != sizeof(chunkSizeBytes)) {
+            return DRWAV_FALSE;
+        }
+
+        if (drwav_bytes_to_u32_be(chunkSizeBytes) < 18) {
+            return DRWAV_FALSE;
+        }
+
+        if (drwav__on_read(pWav->onRead, pWav->pUserData, aiff, sizeof(aiff), &cursor) != sizeof(aiff)) {
+            return DRWAV_FALSE;
+        }
+
+        if (drwav_fourcc_equal(aiff, "AIFF")) {
+            isAIFCFormType = DRWAV_FALSE;
+        } else if (drwav_fourcc_equal(aiff, "AIFC")) {
+            isAIFCFormType = DRWAV_TRUE;
+        } else {
+            return DRWAV_FALSE; /* Expecting "AIFF" or "AIFC". */
+        }
+    } else {
+        return DRWAV_FALSE;
+    }
+
+
+    /* For RF64, the "ds64" chunk must come next, before the "fmt " chunk. */
+    if (pWav->container == drwav_container_rf64) {
+        drwav_uint8 sizeBytes[8];
+        drwav_uint64 bytesRemainingInChunk;
+        drwav_chunk_header header;
+        result = drwav__read_chunk_header(pWav->onRead, pWav->pUserData, pWav->container, &cursor, &header);
+        if (result != DRWAV_SUCCESS) {
+            return DRWAV_FALSE;
+        }
+
+        if (!drwav_fourcc_equal(header.id.fourcc, "ds64")) {
+            return DRWAV_FALSE; /* Expecting "ds64". */
+        }
+
+        bytesRemainingInChunk = header.sizeInBytes + header.paddingSize;
+
+        /* We don't care about the size of the RIFF chunk - skip it. */
+        if (!drwav__seek_forward(pWav->onSeek, 8, pWav->pUserData)) {
+            return DRWAV_FALSE;
+        }
+        bytesRemainingInChunk -= 8;
+        cursor += 8;
+
+
+        /* Next 8 bytes is the size of the "data" chunk. */
+        if (drwav__on_read(pWav->onRead, pWav->pUserData, sizeBytes, sizeof(sizeBytes), &cursor) != sizeof(sizeBytes)) {
+            return DRWAV_FALSE;
+        }
+        bytesRemainingInChunk -= 8;
+        dataChunkSize = drwav_bytes_to_u64(sizeBytes);
+
+
+        /* Next 8 bytes is the same count which we would usually derived from the FACT chunk if it was available. */
+        if (drwav__on_read(pWav->onRead, pWav->pUserData, sizeBytes, sizeof(sizeBytes), &cursor) != sizeof(sizeBytes)) {
+            return DRWAV_FALSE;
+        }
+        bytesRemainingInChunk -= 8;
+        sampleCountFromFactChunk = drwav_bytes_to_u64(sizeBytes);
+
+
+        /* Skip over everything else. */
+        if (!drwav__seek_forward(pWav->onSeek, bytesRemainingInChunk, pWav->pUserData)) {
+            return DRWAV_FALSE;
+        }
+        cursor += bytesRemainingInChunk;
+    }
+
+
+    metadataStartPos = cursor;
+
+    /*
+    Whether or not we are processing metadata controls how we load. We can load more efficiently when
+    metadata is not being processed, but we also cannot process metadata for Wave64 because I have not
+    been able to test it. If someone is able to test this and provide a patch I'm happy to enable it.
+
+    Seqential mode cannot support metadata because it involves seeking backwards.
+    */
+    isProcessingMetadata = !sequential && ((flags & DRWAV_WITH_METADATA) != 0);
+
+    /* Don't allow processing of metadata with untested containers. */
+    if (pWav->container != drwav_container_riff && pWav->container != drwav_container_rf64) {
+        isProcessingMetadata = DRWAV_FALSE;
+    }
+
+    DRWAV_ZERO_MEMORY(&metadataParser, sizeof(metadataParser));
+    if (isProcessingMetadata) {
+        metadataParser.onRead = pWav->onRead;
+        metadataParser.onSeek = pWav->onSeek;
+        metadataParser.pReadSeekUserData = pWav->pUserData;
+        metadataParser.stage  = drwav__metadata_parser_stage_count;
+    }
+
+
+    /*
+    From here on out, chunks might be in any order. In order to robustly handle metadata we'll need
+    to loop through every chunk and handle them as we find them. In sequential mode we need to get
+    out of the loop as soon as we find the data chunk because we won't be able to seek back.
+    */
+    for (;;) {  /* For each chunk... */
+        drwav_chunk_header header;
+        drwav_uint64 chunkSize;
+
+        result = drwav__read_chunk_header(pWav->onRead, pWav->pUserData, pWav->container, &cursor, &header);
+        if (result != DRWAV_SUCCESS) {
+            break;
+        }
+
+        chunkSize = header.sizeInBytes;
+
+
+        /*
+        Always tell the caller about this chunk. We cannot do this in sequential mode because the
+        callback is allowed to read from the file, in which case we'll need to rewind.
+        */
+        if (!sequential && onChunk != NULL) {
+            drwav_uint64 callbackBytesRead = onChunk(pChunkUserData, pWav->onRead, pWav->onSeek, pWav->pUserData, &header, pWav->container, &fmt);
+
+            /*
+            dr_wav may need to read the contents of the chunk, so we now need to seek back to the position before
+            we called the callback.
+            */
+            if (callbackBytesRead > 0) {
+                if (drwav__seek_from_start(pWav->onSeek, cursor, pWav->pUserData) == DRWAV_FALSE) {
+                    return DRWAV_FALSE;
+                }
+            }
+        }
+
+
+        /* Explicitly handle known chunks first. */
+
+        /* "fmt " */
+        if (((pWav->container == drwav_container_riff || pWav->container == drwav_container_rifx || pWav->container == drwav_container_rf64) && drwav_fourcc_equal(header.id.fourcc, "fmt ")) ||
+            ((pWav->container == drwav_container_w64) && drwav_guid_equal(header.id.guid, drwavGUID_W64_FMT))) {
+            drwav_uint8 fmtData[16];
+
+            foundChunk_fmt = DRWAV_TRUE;
+
+            if (pWav->onRead(pWav->pUserData, fmtData, sizeof(fmtData)) != sizeof(fmtData)) {
+                return DRWAV_FALSE;
+            }
+            cursor += sizeof(fmtData);
+
+            fmt.formatTag      = drwav_bytes_to_u16_ex(fmtData + 0,  pWav->container);
+            fmt.channels       = drwav_bytes_to_u16_ex(fmtData + 2,  pWav->container);
+            fmt.sampleRate     = drwav_bytes_to_u32_ex(fmtData + 4,  pWav->container);
+            fmt.avgBytesPerSec = drwav_bytes_to_u32_ex(fmtData + 8,  pWav->container);
+            fmt.blockAlign     = drwav_bytes_to_u16_ex(fmtData + 12, pWav->container);
+            fmt.bitsPerSample  = drwav_bytes_to_u16_ex(fmtData + 14, pWav->container);
+
+            fmt.extendedSize       = 0;
+            fmt.validBitsPerSample = 0;
+            fmt.channelMask        = 0;
+            DRWAV_ZERO_MEMORY(fmt.subFormat, sizeof(fmt.subFormat));
+
+            if (header.sizeInBytes > 16) {
+                drwav_uint8 fmt_cbSize[2];
+                int bytesReadSoFar = 0;
+
+                if (pWav->onRead(pWav->pUserData, fmt_cbSize, sizeof(fmt_cbSize)) != sizeof(fmt_cbSize)) {
+                    return DRWAV_FALSE;    /* Expecting more data. */
+                }
+                cursor += sizeof(fmt_cbSize);
+
+                bytesReadSoFar = 18;
+
+                fmt.extendedSize = drwav_bytes_to_u16_ex(fmt_cbSize, pWav->container);
+                if (fmt.extendedSize > 0) {
+                    /* Simple validation. */
+                    if (fmt.formatTag == DR_WAVE_FORMAT_EXTENSIBLE) {
+                        if (fmt.extendedSize != 22) {
+                            return DRWAV_FALSE;
+                        }
+                    }
+
+                    if (fmt.formatTag == DR_WAVE_FORMAT_EXTENSIBLE) {
+                        drwav_uint8 fmtext[22];
+
+                        if (pWav->onRead(pWav->pUserData, fmtext, fmt.extendedSize) != fmt.extendedSize) {
+                            return DRWAV_FALSE;    /* Expecting more data. */
+                        }
+
+                        fmt.validBitsPerSample = drwav_bytes_to_u16_ex(fmtext + 0, pWav->container);
+                        fmt.channelMask        = drwav_bytes_to_u32_ex(fmtext + 2, pWav->container);
+                        drwav_bytes_to_guid(fmtext + 6, fmt.subFormat);
+                    } else {
+                        if (pWav->onSeek(pWav->pUserData, fmt.extendedSize, drwav_seek_origin_current) == DRWAV_FALSE) {
+                            return DRWAV_FALSE;
+                        }
+                    }
+                    cursor += fmt.extendedSize;
+
+                    bytesReadSoFar += fmt.extendedSize;
+                }
+
+                /* Seek past any leftover bytes. For w64 the leftover will be defined based on the chunk size. */
+                if (pWav->onSeek(pWav->pUserData, (int)(header.sizeInBytes - bytesReadSoFar), drwav_seek_origin_current) == DRWAV_FALSE) {
+                    return DRWAV_FALSE;
+                }
+                cursor += (header.sizeInBytes - bytesReadSoFar);
+            }
+
+            if (header.paddingSize > 0) {
+                if (drwav__seek_forward(pWav->onSeek, header.paddingSize, pWav->pUserData) == DRWAV_FALSE) {
+                    break;
+                }
+                cursor += header.paddingSize;
+            }
+
+            /* Go to the next chunk. Don't include this chunk in metadata. */
+            continue;
+        }
+
+        /* "data" */
+        if (((pWav->container == drwav_container_riff || pWav->container == drwav_container_rifx || pWav->container == drwav_container_rf64) && drwav_fourcc_equal(header.id.fourcc, "data")) ||
+            ((pWav->container == drwav_container_w64) && drwav_guid_equal(header.id.guid, drwavGUID_W64_DATA))) {
+            foundChunk_data = DRWAV_TRUE;
+            
+            pWav->dataChunkDataPos  = cursor;
+
+            if (pWav->container != drwav_container_rf64) {  /* The data chunk size for RF64 will always be set to 0xFFFFFFFF here. It was set to it's true value earlier. */
+                dataChunkSize = chunkSize;
+            }
+
+            /* If we're running in sequential mode, or we're not reading metadata, we have enough now that we can get out of the loop. */
+            if (sequential || !isProcessingMetadata) {
+                break;      /* No need to keep reading beyond the data chunk. */
+            } else {
+                chunkSize += header.paddingSize;    /* <-- Make sure we seek past the padding. */
+                if (drwav__seek_forward(pWav->onSeek, chunkSize, pWav->pUserData) == DRWAV_FALSE) {
+                    break;
+                }
+                cursor += chunkSize;
+
+                continue;   /* There may be some more metadata to read. */
+            }
+        }
+
+        /* "fact". This is optional. Can use this to get the sample count which is useful for compressed formats. For RF64 we retrieved the sample count from the ds64 chunk earlier. */
+        if (((pWav->container == drwav_container_riff || pWav->container == drwav_container_rifx || pWav->container == drwav_container_rf64) && drwav_fourcc_equal(header.id.fourcc, "fact")) ||
+            ((pWav->container == drwav_container_w64) && drwav_guid_equal(header.id.guid, drwavGUID_W64_FACT))) {
+            if (pWav->container == drwav_container_riff || pWav->container == drwav_container_rifx) {
+                drwav_uint8 sampleCount[4];
+                if (drwav__on_read(pWav->onRead, pWav->pUserData, &sampleCount, 4, &cursor) != 4) {
+                    return DRWAV_FALSE;
+                }
+
+                chunkSize -= 4;
+
+                /*
+                The sample count in the "fact" chunk is either unreliable, or I'm not understanding it properly. For now I am only enabling this
+                for Microsoft ADPCM formats.
+                */
+                if (pWav->translatedFormatTag == DR_WAVE_FORMAT_ADPCM) {
+                    sampleCountFromFactChunk = drwav_bytes_to_u32_ex(sampleCount, pWav->container);
+                } else {
+                    sampleCountFromFactChunk = 0;
+                }
+            } else if (pWav->container == drwav_container_w64) {
+                if (drwav__on_read(pWav->onRead, pWav->pUserData, &sampleCountFromFactChunk, 8, &cursor) != 8) {
+                    return DRWAV_FALSE;
+                }
+
+                chunkSize -= 8;
+            } else if (pWav->container == drwav_container_rf64) {
+                /* We retrieved the sample count from the ds64 chunk earlier so no need to do that here. */
+            }
+
+            /* Seek to the next chunk in preparation for the next iteration. */
+            chunkSize += header.paddingSize;    /* <-- Make sure we seek past the padding. */
+            if (drwav__seek_forward(pWav->onSeek, chunkSize, pWav->pUserData) == DRWAV_FALSE) {
+                break;
+            }
+            cursor += chunkSize;
+
+            continue;
+        }
+
+
+        /* "COMM". AIFF/AIFC only. */
+        if (pWav->container == drwav_container_aiff && drwav_fourcc_equal(header.id.fourcc, "COMM")) {
+            drwav_uint8 commData[24];
+            drwav_uint32 commDataBytesToRead;
+            drwav_uint16 channels;
+            drwav_uint32 frameCount;
+            drwav_uint16 sampleSizeInBits;
+            drwav_int64  sampleRate;
+            drwav_uint16 compressionFormat;
+
+            foundChunk_fmt = DRWAV_TRUE;
+
+            if (isAIFCFormType) {
+                commDataBytesToRead = 24;
+                if (header.sizeInBytes < commDataBytesToRead) {
+                    return DRWAV_FALSE; /* Invalid COMM chunk. */
+                }
+            } else {
+                commDataBytesToRead = 18;
+                if (header.sizeInBytes != commDataBytesToRead) {
+                    return DRWAV_FALSE; /* INVALID COMM chunk. */
+                }
+            }
+
+            if (drwav__on_read(pWav->onRead, pWav->pUserData, commData, commDataBytesToRead, &cursor) != commDataBytesToRead) {
+                return DRWAV_FALSE;
+            }
+
+            
+            channels         = drwav_bytes_to_u16_ex     (commData + 0, pWav->container);
+            frameCount       = drwav_bytes_to_u32_ex     (commData + 2, pWav->container);
+            sampleSizeInBits = drwav_bytes_to_u16_ex     (commData + 6, pWav->container);
+            sampleRate       = drwav_aiff_extented_to_s64(commData + 8);
+
+            if (sampleRate < 0 || sampleRate > 0xFFFFFFFF) {
+                return DRWAV_FALSE; /* Invalid sample rate. */
+            }
+
+            if (isAIFCFormType) {
+                const drwav_uint8* type = commData + 18;
+
+                if (drwav_fourcc_equal(type, "NONE")) {
+                    compressionFormat = DR_WAVE_FORMAT_PCM; /* PCM, big-endian. */
+                } else if (drwav_fourcc_equal(type, "raw ")) {
+                    compressionFormat = DR_WAVE_FORMAT_PCM;
+
+                    /* In my testing, it looks like when the "raw " compression type is used, 8-bit samples should be considered unsigned. */
+                    if (sampleSizeInBits == 8) {
+                        pWav->aiff.isUnsigned = DRWAV_TRUE;
+                    }
+                } else if (drwav_fourcc_equal(type, "sowt")) {
+                    compressionFormat = DR_WAVE_FORMAT_PCM; /* PCM, little-endian. */
+                    pWav->aiff.isLE = DRWAV_TRUE;
+                } else if (drwav_fourcc_equal(type, "fl32") || drwav_fourcc_equal(type, "fl64") || drwav_fourcc_equal(type, "FL32") || drwav_fourcc_equal(type, "FL64")) {
+                    compressionFormat = DR_WAVE_FORMAT_IEEE_FLOAT;
+                } else if (drwav_fourcc_equal(type, "alaw") || drwav_fourcc_equal(type, "ALAW")) {
+                    compressionFormat = DR_WAVE_FORMAT_ALAW;
+                } else if (drwav_fourcc_equal(type, "ulaw") || drwav_fourcc_equal(type, "ULAW")) {
+                    compressionFormat = DR_WAVE_FORMAT_MULAW;
+                } else if (drwav_fourcc_equal(type, "ima4")) {
+                    compressionFormat = DR_WAVE_FORMAT_DVI_ADPCM;
+                    sampleSizeInBits = 4;
+
+                    /*
+                    I haven't been able to figure out how to get correct decoding for IMA ADPCM. Until this is figured out
+                    we'll need to abort when we encounter such an encoding. Advice welcome!
+                    */
+                    return DRWAV_FALSE;
+                } else {
+                    return DRWAV_FALSE; /* Unknown or unsupported compression format. Need to abort. */
+                }
+            } else {
+                compressionFormat = DR_WAVE_FORMAT_PCM; /* It's a standard AIFF form which is always compressed. */
+            }
+
+            /* With AIFF we want to use the explicitly defined frame count rather than deriving it from the size of the chunk. */
+            aiffFrameCount = frameCount;
+
+            /* We should now have enough information to fill out our fmt structure. */
+            fmt.formatTag      = compressionFormat;
+            fmt.channels       = channels;
+            fmt.sampleRate     = (drwav_uint32)sampleRate;
+            fmt.bitsPerSample  = sampleSizeInBits;
+            fmt.blockAlign     = (drwav_uint16)(fmt.channels * fmt.bitsPerSample / 8);
+            fmt.avgBytesPerSec = fmt.blockAlign * fmt.sampleRate;
+
+            if (fmt.blockAlign == 0 && compressionFormat == DR_WAVE_FORMAT_DVI_ADPCM) {
+                fmt.blockAlign = 34 * fmt.channels;
+            }
+
+            /*
+            Weird one. I've seen some alaw and ulaw encoded files that for some reason set the bits per sample to 16 when
+            it should be 8. To get this working I need to explicitly check for this and change it.
+            */
+            if (compressionFormat == DR_WAVE_FORMAT_ALAW || compressionFormat == DR_WAVE_FORMAT_MULAW) {
+                if (fmt.bitsPerSample > 8) {
+                    fmt.bitsPerSample = 8;
+                    fmt.blockAlign = fmt.channels;
+                }
+            }
+
+            /* In AIFF, samples are padded to 8 byte boundaries. We need to round up our bits per sample here. */
+            fmt.bitsPerSample += (fmt.bitsPerSample & 7);
+            
+
+            /* If the form type is AIFC there will be some additional data in the chunk. We need to seek past it. */
+            if (isAIFCFormType) {
+                if (drwav__seek_forward(pWav->onSeek, (chunkSize - commDataBytesToRead), pWav->pUserData) == DRWAV_FALSE) {
+                    return DRWAV_FALSE;
+                }
+                cursor += (chunkSize - commDataBytesToRead);
+            }
+
+            /* Don't fall through or else we'll end up treating this chunk as metadata which is incorrect. */
+            continue;
+        }
+
+
+        /* "SSND". AIFF/AIFC only. This is the AIFF equivalent of the "data" chunk. */
+        if (pWav->container == drwav_container_aiff && drwav_fourcc_equal(header.id.fourcc, "SSND")) {
+            drwav_uint8 offsetAndBlockSizeData[8];
+            drwav_uint32 offset;
+
+            foundChunk_data = DRWAV_TRUE;
+
+            if (drwav__on_read(pWav->onRead, pWav->pUserData, offsetAndBlockSizeData, sizeof(offsetAndBlockSizeData), &cursor) != sizeof(offsetAndBlockSizeData)) {
+                return DRWAV_FALSE;
+            }
+
+            /* We need to seek forward by the offset. */
+            offset = drwav_bytes_to_u32_ex(offsetAndBlockSizeData + 0, pWav->container);
+            if (drwav__seek_forward(pWav->onSeek, offset, pWav->pUserData) == DRWAV_FALSE) {
+                return DRWAV_FALSE;
+            }
+            cursor += offset;
+
+            pWav->dataChunkDataPos = cursor;
+            dataChunkSize = chunkSize;
+
+            /* If we're running in sequential mode, or we're not reading metadata, we have enough now that we can get out of the loop. */
+            if (sequential || !isProcessingMetadata) {
+                break;      /* No need to keep reading beyond the data chunk. */
+            } else {
+                if (drwav__seek_forward(pWav->onSeek, chunkSize, pWav->pUserData) == DRWAV_FALSE) {
+                    break;
+                }
+                cursor += chunkSize;
+
+                continue;   /* There may be some more metadata to read. */
+            }
+        }
+
+
+
+        /* Getting here means it's not a chunk that we care about internally, but might need to be handled as metadata by the caller. */
+        if (isProcessingMetadata) {
+            drwav__metadata_process_chunk(&metadataParser, &header, drwav_metadata_type_all_including_unknown);
+
+            /* Go back to the start of the chunk so we can normalize the position of the cursor. */
+            if (drwav__seek_from_start(pWav->onSeek, cursor, pWav->pUserData) == DRWAV_FALSE) {
+                break;  /* Failed to seek. Can't reliable read the remaining chunks. Get out. */
+            }
+        }
+
+
+        /* Make sure we skip past the content of this chunk before we go to the next one. */
+        chunkSize += header.paddingSize;    /* <-- Make sure we seek past the padding. */
+        if (drwav__seek_forward(pWav->onSeek, chunkSize, pWav->pUserData) == DRWAV_FALSE) {
+            break;
+        }
+        cursor += chunkSize;
+    }
+
+    /* There's some mandatory chunks that must exist. If they were not found in the iteration above we must abort. */
+    if (!foundChunk_fmt || !foundChunk_data) {
+        return DRWAV_FALSE;
+    }
+
+    /* Basic validation. */
+    if ((fmt.sampleRate    == 0 || fmt.sampleRate    > DRWAV_MAX_SAMPLE_RATE    ) ||
+        (fmt.channels      == 0 || fmt.channels      > DRWAV_MAX_CHANNELS       ) ||
+        (fmt.bitsPerSample == 0 || fmt.bitsPerSample > DRWAV_MAX_BITS_PER_SAMPLE) ||
+        fmt.blockAlign == 0) {
+        return DRWAV_FALSE; /* Probably an invalid WAV file. */
+    }
+
+    /* Translate the internal format. */
+    translatedFormatTag = fmt.formatTag;
+    if (translatedFormatTag == DR_WAVE_FORMAT_EXTENSIBLE) {
+        translatedFormatTag = drwav_bytes_to_u16_ex(fmt.subFormat + 0, pWav->container);
+    }
+
+    /* We may have moved passed the data chunk. If so we need to move back. If running in sequential mode we can assume we are already sitting on the data chunk. */
+    if (!sequential) {
+        if (!drwav__seek_from_start(pWav->onSeek, pWav->dataChunkDataPos, pWav->pUserData)) {
+            return DRWAV_FALSE;
+        }
+        cursor = pWav->dataChunkDataPos;
+    }
+
+
+    /*
+    At this point we should have done the initial parsing of each of our chunks, but we now need to
+    do a second pass to extract the actual contents of the metadata (the first pass just calculated
+    the length of the memory allocation).
+
+    We only do this if we've actually got metadata to parse.
+    */
+    if (isProcessingMetadata && metadataParser.metadataCount > 0) {
+        if (drwav__seek_from_start(pWav->onSeek, metadataStartPos, pWav->pUserData) == DRWAV_FALSE) {
+            return DRWAV_FALSE;
+        }
+
+        result = drwav__metadata_alloc(&metadataParser, &pWav->allocationCallbacks);
+        if (result != DRWAV_SUCCESS) {
+            return DRWAV_FALSE;
+        }
+
+        metadataParser.stage = drwav__metadata_parser_stage_read;
+
+        for (;;) {
+            drwav_chunk_header header;
+            drwav_uint64 metadataBytesRead;
+
+            result = drwav__read_chunk_header(pWav->onRead, pWav->pUserData, pWav->container, &cursor, &header);
+            if (result != DRWAV_SUCCESS) {
+                break;
+            }
+
+            metadataBytesRead = drwav__metadata_process_chunk(&metadataParser, &header, drwav_metadata_type_all_including_unknown);
+
+            /* Move to the end of the chunk so we can keep iterating. */
+            if (drwav__seek_forward(pWav->onSeek, (header.sizeInBytes + header.paddingSize) - metadataBytesRead, pWav->pUserData) == DRWAV_FALSE) {
+                drwav_free(metadataParser.pMetadata, &pWav->allocationCallbacks);
+                return DRWAV_FALSE;
+            }
+        }
+
+        /* Getting here means we're finished parsing the metadata. */
+        pWav->pMetadata     = metadataParser.pMetadata;
+        pWav->metadataCount = metadataParser.metadataCount;
+    }
+
+
+    /* At this point we should be sitting on the first byte of the raw audio data. */
+
+    /*
+    I've seen a WAV file in the wild where a RIFF-ecapsulated file has the size of it's "RIFF" and
+    "data" chunks set to 0xFFFFFFFF when the file is definitely not that big. In this case we're
+    going to have to calculate the size by reading and discarding bytes, and then seeking back. We
+    cannot do this in sequential mode. We just assume that the rest of the file is audio data.
+    */
+    if (dataChunkSize == 0xFFFFFFFF && (pWav->container == drwav_container_riff || pWav->container == drwav_container_rifx) && pWav->isSequentialWrite == DRWAV_FALSE) {
+        dataChunkSize = 0;
+
+        for (;;) {
+            drwav_uint8 temp[4096];
+            size_t bytesRead = pWav->onRead(pWav->pUserData, temp, sizeof(temp));
+            dataChunkSize += bytesRead;
+
+            if (bytesRead < sizeof(temp)) {
+                break;
+            }
+        }
+    }
+
+    if (drwav__seek_from_start(pWav->onSeek, pWav->dataChunkDataPos, pWav->pUserData) == DRWAV_FALSE) {
+        drwav_free(pWav->pMetadata, &pWav->allocationCallbacks);
+        return DRWAV_FALSE;
+    }
+
+
+    pWav->fmt                 = fmt;
+    pWav->sampleRate          = fmt.sampleRate;
+    pWav->channels            = fmt.channels;
+    pWav->bitsPerSample       = fmt.bitsPerSample;
+    pWav->bytesRemaining      = dataChunkSize;
+    pWav->translatedFormatTag = translatedFormatTag;
+    pWav->dataChunkDataSize   = dataChunkSize;
+
+    if (sampleCountFromFactChunk != 0) {
+        pWav->totalPCMFrameCount = sampleCountFromFactChunk;
+    } else if (aiffFrameCount != 0) {
+        pWav->totalPCMFrameCount = aiffFrameCount;
+    } else {
+        drwav_uint32 bytesPerFrame = drwav_get_bytes_per_pcm_frame(pWav);
+        if (bytesPerFrame == 0) {
+            drwav_free(pWav->pMetadata, &pWav->allocationCallbacks);
+            return DRWAV_FALSE; /* Invalid file. */
+        }
+
+        pWav->totalPCMFrameCount = dataChunkSize / bytesPerFrame;
+
+        if (pWav->translatedFormatTag == DR_WAVE_FORMAT_ADPCM) {
+            drwav_uint64 totalBlockHeaderSizeInBytes;
+            drwav_uint64 blockCount = dataChunkSize / fmt.blockAlign;
+
+            /* Make sure any trailing partial block is accounted for. */
+            if ((blockCount * fmt.blockAlign) < dataChunkSize) {
+                blockCount += 1;
+            }
+
+            /* We decode two samples per byte. There will be blockCount headers in the data chunk. This is enough to know how to calculate the total PCM frame count. */
+            totalBlockHeaderSizeInBytes = blockCount * (6*fmt.channels);
+            pWav->totalPCMFrameCount = ((dataChunkSize - totalBlockHeaderSizeInBytes) * 2) / fmt.channels;
+        }
+        if (pWav->translatedFormatTag == DR_WAVE_FORMAT_DVI_ADPCM) {
+            drwav_uint64 totalBlockHeaderSizeInBytes;
+            drwav_uint64 blockCount = dataChunkSize / fmt.blockAlign;
+
+            /* Make sure any trailing partial block is accounted for. */
+            if ((blockCount * fmt.blockAlign) < dataChunkSize) {
+                blockCount += 1;
+            }
+
+            /* We decode two samples per byte. There will be blockCount headers in the data chunk. This is enough to know how to calculate the total PCM frame count. */
+            totalBlockHeaderSizeInBytes = blockCount * (4*fmt.channels);
+            pWav->totalPCMFrameCount = ((dataChunkSize - totalBlockHeaderSizeInBytes) * 2) / fmt.channels;
+
+            /* The header includes a decoded sample for each channel which acts as the initial predictor sample. */
+            pWav->totalPCMFrameCount += blockCount;
+        }
+    }
+
+    /* Some formats only support a certain number of channels. */
+    if (pWav->translatedFormatTag == DR_WAVE_FORMAT_ADPCM || pWav->translatedFormatTag == DR_WAVE_FORMAT_DVI_ADPCM) {
+        if (pWav->channels > 2) {
+            drwav_free(pWav->pMetadata, &pWav->allocationCallbacks);
+            return DRWAV_FALSE;
+        }
+    }
+
+    /* The number of bytes per frame must be known. If not, it's an invalid file and not decodable. */
+    if (drwav_get_bytes_per_pcm_frame(pWav) == 0) {
+        drwav_free(pWav->pMetadata, &pWav->allocationCallbacks);
+        return DRWAV_FALSE;
+    }
+
+#ifdef DR_WAV_LIBSNDFILE_COMPAT
+    /*
+    I use libsndfile as a benchmark for testing, however in the version I'm using (from the Windows installer on the libsndfile website),
+    it appears the total sample count libsndfile uses for MS-ADPCM is incorrect. It would seem they are computing the total sample count
+    from the number of blocks, however this results in the inclusion of extra silent samples at the end of the last block. The correct
+    way to know the total sample count is to inspect the "fact" chunk, which should always be present for compressed formats, and should
+    always include the sample count. This little block of code below is only used to emulate the libsndfile logic so I can properly run my
+    correctness tests against libsndfile, and is disabled by default.
+    */
+    if (pWav->translatedFormatTag == DR_WAVE_FORMAT_ADPCM) {
+        drwav_uint64 blockCount = dataChunkSize / fmt.blockAlign;
+        pWav->totalPCMFrameCount = (((blockCount * (fmt.blockAlign - (6*pWav->channels))) * 2)) / fmt.channels;  /* x2 because two samples per byte. */
+    }
+    if (pWav->translatedFormatTag == DR_WAVE_FORMAT_DVI_ADPCM) {
+        drwav_uint64 blockCount = dataChunkSize / fmt.blockAlign;
+        pWav->totalPCMFrameCount = (((blockCount * (fmt.blockAlign - (4*pWav->channels))) * 2) + (blockCount * pWav->channels)) / fmt.channels;
+    }
+#endif
+
+    return DRWAV_TRUE;
+}
+
+DRWAV_API drwav_bool32 drwav_init(drwav* pWav, drwav_read_proc onRead, drwav_seek_proc onSeek, void* pUserData, const drwav_allocation_callbacks* pAllocationCallbacks)
+{
+    return drwav_init_ex(pWav, onRead, onSeek, NULL, pUserData, NULL, 0, pAllocationCallbacks);
+}
+
+DRWAV_API drwav_bool32 drwav_init_ex(drwav* pWav, drwav_read_proc onRead, drwav_seek_proc onSeek, drwav_chunk_proc onChunk, void* pReadSeekUserData, void* pChunkUserData, drwav_uint32 flags, const drwav_allocation_callbacks* pAllocationCallbacks)
+{
+    if (!drwav_preinit(pWav, onRead, onSeek, pReadSeekUserData, pAllocationCallbacks)) {
+        return DRWAV_FALSE;
+    }
+
+    return drwav_init__internal(pWav, onChunk, pChunkUserData, flags);
+}
+
+DRWAV_API drwav_bool32 drwav_init_with_metadata(drwav* pWav, drwav_read_proc onRead, drwav_seek_proc onSeek, void* pUserData, drwav_uint32 flags, const drwav_allocation_callbacks* pAllocationCallbacks)
+{
+    if (!drwav_preinit(pWav, onRead, onSeek, pUserData, pAllocationCallbacks)) {
+        return DRWAV_FALSE;
+    }
+
+    return drwav_init__internal(pWav, NULL, NULL, flags | DRWAV_WITH_METADATA);
+}
+
+DRWAV_API drwav_metadata* drwav_take_ownership_of_metadata(drwav* pWav)
+{
+    drwav_metadata *result = pWav->pMetadata;
+
+    pWav->pMetadata     = NULL;
+    pWav->metadataCount = 0;
+
+    return result;
+}
+
+
+DRWAV_PRIVATE size_t drwav__write(drwav* pWav, const void* pData, size_t dataSize)
+{
+    DRWAV_ASSERT(pWav          != NULL);
+    DRWAV_ASSERT(pWav->onWrite != NULL);
+
+    /* Generic write. Assumes no byte reordering required. */
+    return pWav->onWrite(pWav->pUserData, pData, dataSize);
+}
+
+DRWAV_PRIVATE size_t drwav__write_byte(drwav* pWav, drwav_uint8 byte)
+{
+    DRWAV_ASSERT(pWav          != NULL);
+    DRWAV_ASSERT(pWav->onWrite != NULL);
+
+    return pWav->onWrite(pWav->pUserData, &byte, 1);
+}
+
+DRWAV_PRIVATE size_t drwav__write_u16ne_to_le(drwav* pWav, drwav_uint16 value)
+{
+    DRWAV_ASSERT(pWav          != NULL);
+    DRWAV_ASSERT(pWav->onWrite != NULL);
+
+    if (!drwav__is_little_endian()) {
+        value = drwav__bswap16(value);
+    }
+
+    return drwav__write(pWav, &value, 2);
+}
+
+DRWAV_PRIVATE size_t drwav__write_u32ne_to_le(drwav* pWav, drwav_uint32 value)
+{
+    DRWAV_ASSERT(pWav          != NULL);
+    DRWAV_ASSERT(pWav->onWrite != NULL);
+
+    if (!drwav__is_little_endian()) {
+        value = drwav__bswap32(value);
+    }
+
+    return drwav__write(pWav, &value, 4);
+}
+
+DRWAV_PRIVATE size_t drwav__write_u64ne_to_le(drwav* pWav, drwav_uint64 value)
+{
+    DRWAV_ASSERT(pWav          != NULL);
+    DRWAV_ASSERT(pWav->onWrite != NULL);
+
+    if (!drwav__is_little_endian()) {
+        value = drwav__bswap64(value);
+    }
+
+    return drwav__write(pWav, &value, 8);
+}
+
+DRWAV_PRIVATE size_t drwav__write_f32ne_to_le(drwav* pWav, float value)
+{
+    union {
+       drwav_uint32 u32;
+       float f32;
+    } u;
+
+    DRWAV_ASSERT(pWav          != NULL);
+    DRWAV_ASSERT(pWav->onWrite != NULL);
+
+    u.f32 = value;
+
+    if (!drwav__is_little_endian()) {
+        u.u32 = drwav__bswap32(u.u32);
+    }
+
+    return drwav__write(pWav, &u.u32, 4);
+}
+
+DRWAV_PRIVATE size_t drwav__write_or_count(drwav* pWav, const void* pData, size_t dataSize)
+{
+    if (pWav == NULL) {
+        return dataSize;
+    }
+
+    return drwav__write(pWav, pData, dataSize);
+}
+
+DRWAV_PRIVATE size_t drwav__write_or_count_byte(drwav* pWav, drwav_uint8 byte)
+{
+    if (pWav == NULL) {
+        return 1;
+    }
+
+    return drwav__write_byte(pWav, byte);
+}
+
+DRWAV_PRIVATE size_t drwav__write_or_count_u16ne_to_le(drwav* pWav, drwav_uint16 value)
+{
+    if (pWav == NULL) {
+        return 2;
+    }
+
+    return drwav__write_u16ne_to_le(pWav, value);
+}
+
+DRWAV_PRIVATE size_t drwav__write_or_count_u32ne_to_le(drwav* pWav, drwav_uint32 value)
+{
+    if (pWav == NULL) {
+        return 4;
+    }
+
+    return drwav__write_u32ne_to_le(pWav, value);
+}
+
+#if 0   /* Unused for now. */
+DRWAV_PRIVATE size_t drwav__write_or_count_u64ne_to_le(drwav* pWav, drwav_uint64 value)
+{
+    if (pWav == NULL) {
+        return 8;
+    }
+
+    return drwav__write_u64ne_to_le(pWav, value);
+}
+#endif
+
+DRWAV_PRIVATE size_t drwav__write_or_count_f32ne_to_le(drwav* pWav, float value)
+{
+    if (pWav == NULL) {
+        return 4;
+    }
+
+    return drwav__write_f32ne_to_le(pWav, value);
+}
+
+DRWAV_PRIVATE size_t drwav__write_or_count_string_to_fixed_size_buf(drwav* pWav, char* str, size_t bufFixedSize)
+{
+    size_t len;
+
+    if (pWav == NULL) {
+        return bufFixedSize;
+    }
+
+    len = drwav__strlen_clamped(str, bufFixedSize);
+    drwav__write_or_count(pWav, str, len);
+
+    if (len < bufFixedSize) {
+        size_t i;
+        for (i = 0; i < bufFixedSize - len; ++i) {
+            drwav__write_byte(pWav, 0);
+        }
+    }
+
+    return bufFixedSize;
+}
+
+
+/* pWav can be NULL meaning just count the bytes that would be written. */
+DRWAV_PRIVATE size_t drwav__write_or_count_metadata(drwav* pWav, drwav_metadata* pMetadatas, drwav_uint32 metadataCount)
+{
+    size_t bytesWritten = 0;
+    drwav_bool32 hasListAdtl = DRWAV_FALSE;
+    drwav_bool32 hasListInfo = DRWAV_FALSE;
+    drwav_uint32 iMetadata;
+
+    if (pMetadatas == NULL || metadataCount == 0) {
+        return 0;
+    }
+
+    for (iMetadata = 0; iMetadata < metadataCount; ++iMetadata) {
+        drwav_metadata* pMetadata = &pMetadatas[iMetadata];
+        drwav_uint32 chunkSize = 0;
+
+        if ((pMetadata->type & drwav_metadata_type_list_all_info_strings) || (pMetadata->type == drwav_metadata_type_unknown && pMetadata->data.unknown.chunkLocation == drwav_metadata_location_inside_info_list)) {
+            hasListInfo = DRWAV_TRUE;
+        }
+
+        if ((pMetadata->type & drwav_metadata_type_list_all_adtl) || (pMetadata->type == drwav_metadata_type_unknown && pMetadata->data.unknown.chunkLocation == drwav_metadata_location_inside_adtl_list)) {
+            hasListAdtl = DRWAV_TRUE;
+        }
+
+        switch (pMetadata->type) {
+            case drwav_metadata_type_smpl:
+            {
+                drwav_uint32 iLoop;
+
+                chunkSize = DRWAV_SMPL_BYTES + DRWAV_SMPL_LOOP_BYTES * pMetadata->data.smpl.sampleLoopCount + pMetadata->data.smpl.samplerSpecificDataSizeInBytes;
+
+                bytesWritten += drwav__write_or_count(pWav, "smpl", 4);
+                bytesWritten += drwav__write_or_count_u32ne_to_le(pWav, chunkSize);
+
+                bytesWritten += drwav__write_or_count_u32ne_to_le(pWav, pMetadata->data.smpl.manufacturerId);
+                bytesWritten += drwav__write_or_count_u32ne_to_le(pWav, pMetadata->data.smpl.productId);
+                bytesWritten += drwav__write_or_count_u32ne_to_le(pWav, pMetadata->data.smpl.samplePeriodNanoseconds);
+                bytesWritten += drwav__write_or_count_u32ne_to_le(pWav, pMetadata->data.smpl.midiUnityNote);
+                bytesWritten += drwav__write_or_count_u32ne_to_le(pWav, pMetadata->data.smpl.midiPitchFraction);
+                bytesWritten += drwav__write_or_count_u32ne_to_le(pWav, pMetadata->data.smpl.smpteFormat);
+                bytesWritten += drwav__write_or_count_u32ne_to_le(pWav, pMetadata->data.smpl.smpteOffset);
+                bytesWritten += drwav__write_or_count_u32ne_to_le(pWav, pMetadata->data.smpl.sampleLoopCount);
+                bytesWritten += drwav__write_or_count_u32ne_to_le(pWav, pMetadata->data.smpl.samplerSpecificDataSizeInBytes);
+
+                for (iLoop = 0; iLoop < pMetadata->data.smpl.sampleLoopCount; ++iLoop) {
+                    bytesWritten += drwav__write_or_count_u32ne_to_le(pWav, pMetadata->data.smpl.pLoops[iLoop].cuePointId);
+                    bytesWritten += drwav__write_or_count_u32ne_to_le(pWav, pMetadata->data.smpl.pLoops[iLoop].type);
+                    bytesWritten += drwav__write_or_count_u32ne_to_le(pWav, pMetadata->data.smpl.pLoops[iLoop].firstSampleByteOffset);
+                    bytesWritten += drwav__write_or_count_u32ne_to_le(pWav, pMetadata->data.smpl.pLoops[iLoop].lastSampleByteOffset);
+                    bytesWritten += drwav__write_or_count_u32ne_to_le(pWav, pMetadata->data.smpl.pLoops[iLoop].sampleFraction);
+                    bytesWritten += drwav__write_or_count_u32ne_to_le(pWav, pMetadata->data.smpl.pLoops[iLoop].playCount);
+                }
+
+                if (pMetadata->data.smpl.samplerSpecificDataSizeInBytes > 0) {
+                    bytesWritten += drwav__write_or_count(pWav, pMetadata->data.smpl.pSamplerSpecificData, pMetadata->data.smpl.samplerSpecificDataSizeInBytes);
+                }
+            } break;
+
+            case drwav_metadata_type_inst:
+            {
+                chunkSize = DRWAV_INST_BYTES;
+
+                bytesWritten += drwav__write_or_count(pWav, "inst", 4);
+                bytesWritten += drwav__write_or_count_u32ne_to_le(pWav, chunkSize);
+                bytesWritten += drwav__write_or_count(pWav, &pMetadata->data.inst.midiUnityNote, 1);
+                bytesWritten += drwav__write_or_count(pWav, &pMetadata->data.inst.fineTuneCents, 1);
+                bytesWritten += drwav__write_or_count(pWav, &pMetadata->data.inst.gainDecibels, 1);
+                bytesWritten += drwav__write_or_count(pWav, &pMetadata->data.inst.lowNote, 1);
+                bytesWritten += drwav__write_or_count(pWav, &pMetadata->data.inst.highNote, 1);
+                bytesWritten += drwav__write_or_count(pWav, &pMetadata->data.inst.lowVelocity, 1);
+                bytesWritten += drwav__write_or_count(pWav, &pMetadata->data.inst.highVelocity, 1);
+            } break;
+
+            case drwav_metadata_type_cue:
+            {
+                drwav_uint32 iCuePoint;
+
+                chunkSize = DRWAV_CUE_BYTES + DRWAV_CUE_POINT_BYTES * pMetadata->data.cue.cuePointCount;
+
+                bytesWritten += drwav__write_or_count(pWav, "cue ", 4);
+                bytesWritten += drwav__write_or_count_u32ne_to_le(pWav, chunkSize);
+                bytesWritten += drwav__write_or_count_u32ne_to_le(pWav, pMetadata->data.cue.cuePointCount);
+                for (iCuePoint = 0; iCuePoint < pMetadata->data.cue.cuePointCount; ++iCuePoint) {
+                    bytesWritten += drwav__write_or_count_u32ne_to_le(pWav, pMetadata->data.cue.pCuePoints[iCuePoint].id);
+                    bytesWritten += drwav__write_or_count_u32ne_to_le(pWav, pMetadata->data.cue.pCuePoints[iCuePoint].playOrderPosition);
+                    bytesWritten += drwav__write_or_count(pWav, pMetadata->data.cue.pCuePoints[iCuePoint].dataChunkId, 4);
+                    bytesWritten += drwav__write_or_count_u32ne_to_le(pWav, pMetadata->data.cue.pCuePoints[iCuePoint].chunkStart);
+                    bytesWritten += drwav__write_or_count_u32ne_to_le(pWav, pMetadata->data.cue.pCuePoints[iCuePoint].blockStart);
+                    bytesWritten += drwav__write_or_count_u32ne_to_le(pWav, pMetadata->data.cue.pCuePoints[iCuePoint].sampleByteOffset);
+                }
+            } break;
+
+            case drwav_metadata_type_acid:
+            {
+                chunkSize = DRWAV_ACID_BYTES;
+
+                bytesWritten += drwav__write_or_count(pWav, "acid", 4);
+                bytesWritten += drwav__write_or_count_u32ne_to_le(pWav, chunkSize);
+                bytesWritten += drwav__write_or_count_u32ne_to_le(pWav, pMetadata->data.acid.flags);
+                bytesWritten += drwav__write_or_count_u16ne_to_le(pWav, pMetadata->data.acid.midiUnityNote);
+                bytesWritten += drwav__write_or_count_u16ne_to_le(pWav, pMetadata->data.acid.reserved1);
+                bytesWritten += drwav__write_or_count_f32ne_to_le(pWav, pMetadata->data.acid.reserved2);
+                bytesWritten += drwav__write_or_count_u32ne_to_le(pWav, pMetadata->data.acid.numBeats);
+                bytesWritten += drwav__write_or_count_u16ne_to_le(pWav, pMetadata->data.acid.meterDenominator);
+                bytesWritten += drwav__write_or_count_u16ne_to_le(pWav, pMetadata->data.acid.meterNumerator);
+                bytesWritten += drwav__write_or_count_f32ne_to_le(pWav, pMetadata->data.acid.tempo);
+            } break;
+
+            case drwav_metadata_type_bext:
+            {
+                char reservedBuf[DRWAV_BEXT_RESERVED_BYTES];
+                drwav_uint32 timeReferenceLow;
+                drwav_uint32 timeReferenceHigh;
+
+                chunkSize = DRWAV_BEXT_BYTES + pMetadata->data.bext.codingHistorySize;
+
+                bytesWritten += drwav__write_or_count(pWav, "bext", 4);
+                bytesWritten += drwav__write_or_count_u32ne_to_le(pWav, chunkSize);
+
+                bytesWritten += drwav__write_or_count_string_to_fixed_size_buf(pWav, pMetadata->data.bext.pDescription, DRWAV_BEXT_DESCRIPTION_BYTES);
+                bytesWritten += drwav__write_or_count_string_to_fixed_size_buf(pWav, pMetadata->data.bext.pOriginatorName, DRWAV_BEXT_ORIGINATOR_NAME_BYTES);
+                bytesWritten += drwav__write_or_count_string_to_fixed_size_buf(pWav, pMetadata->data.bext.pOriginatorReference, DRWAV_BEXT_ORIGINATOR_REF_BYTES);
+                bytesWritten += drwav__write_or_count(pWav, pMetadata->data.bext.pOriginationDate, sizeof(pMetadata->data.bext.pOriginationDate));
+                bytesWritten += drwav__write_or_count(pWav, pMetadata->data.bext.pOriginationTime, sizeof(pMetadata->data.bext.pOriginationTime));
+
+                timeReferenceLow  = (drwav_uint32)(pMetadata->data.bext.timeReference & 0xFFFFFFFF);
+                timeReferenceHigh = (drwav_uint32)(pMetadata->data.bext.timeReference >> 32);
+                bytesWritten += drwav__write_or_count_u32ne_to_le(pWav, timeReferenceLow);
+                bytesWritten += drwav__write_or_count_u32ne_to_le(pWav, timeReferenceHigh);
+
+                bytesWritten += drwav__write_or_count_u16ne_to_le(pWav, pMetadata->data.bext.version);
+                bytesWritten += drwav__write_or_count(pWav, pMetadata->data.bext.pUMID, DRWAV_BEXT_UMID_BYTES);
+                bytesWritten += drwav__write_or_count_u16ne_to_le(pWav, pMetadata->data.bext.loudnessValue);
+                bytesWritten += drwav__write_or_count_u16ne_to_le(pWav, pMetadata->data.bext.loudnessRange);
+                bytesWritten += drwav__write_or_count_u16ne_to_le(pWav, pMetadata->data.bext.maxTruePeakLevel);
+                bytesWritten += drwav__write_or_count_u16ne_to_le(pWav, pMetadata->data.bext.maxMomentaryLoudness);
+                bytesWritten += drwav__write_or_count_u16ne_to_le(pWav, pMetadata->data.bext.maxShortTermLoudness);
+
+                DRWAV_ZERO_MEMORY(reservedBuf, sizeof(reservedBuf));
+                bytesWritten += drwav__write_or_count(pWav, reservedBuf, sizeof(reservedBuf));
+
+                if (pMetadata->data.bext.codingHistorySize > 0) {
+                    bytesWritten += drwav__write_or_count(pWav, pMetadata->data.bext.pCodingHistory, pMetadata->data.bext.codingHistorySize);
+                }
+            } break;
+
+            case drwav_metadata_type_unknown:
+            {
+                if (pMetadata->data.unknown.chunkLocation == drwav_metadata_location_top_level) {
+                    chunkSize = pMetadata->data.unknown.dataSizeInBytes;
+
+                    bytesWritten += drwav__write_or_count(pWav, pMetadata->data.unknown.id, 4);
+                    bytesWritten += drwav__write_or_count_u32ne_to_le(pWav, chunkSize);
+                    bytesWritten += drwav__write_or_count(pWav, pMetadata->data.unknown.pData, pMetadata->data.unknown.dataSizeInBytes);
+                }
+            } break;
+
+            default: break;
+        }
+        if ((chunkSize % 2) != 0) {
+            bytesWritten += drwav__write_or_count_byte(pWav, 0);
+        }
+    }
+
+    if (hasListInfo) {
+        drwav_uint32 chunkSize = 4; /* Start with 4 bytes for "INFO". */
+        for (iMetadata = 0; iMetadata < metadataCount; ++iMetadata) {
+            drwav_metadata* pMetadata = &pMetadatas[iMetadata];
+
+            if ((pMetadata->type & drwav_metadata_type_list_all_info_strings)) {
+                chunkSize += 8; /* For id and string size. */
+                chunkSize += pMetadata->data.infoText.stringLength + 1; /* Include null terminator. */
+            } else if (pMetadata->type == drwav_metadata_type_unknown && pMetadata->data.unknown.chunkLocation == drwav_metadata_location_inside_info_list) {
+                chunkSize += 8; /* For id string size. */
+                chunkSize += pMetadata->data.unknown.dataSizeInBytes;
+            }
+
+            if ((chunkSize % 2) != 0) {
+                chunkSize += 1;
+            }
+        }
+
+        bytesWritten += drwav__write_or_count(pWav, "LIST", 4);
+        bytesWritten += drwav__write_or_count_u32ne_to_le(pWav, chunkSize);
+        bytesWritten += drwav__write_or_count(pWav, "INFO", 4);
+
+        for (iMetadata = 0; iMetadata < metadataCount; ++iMetadata) {
+            drwav_metadata* pMetadata = &pMetadatas[iMetadata];
+            drwav_uint32 subchunkSize = 0;
+
+            if (pMetadata->type & drwav_metadata_type_list_all_info_strings) {
+                const char* pID = NULL;
+
+                switch (pMetadata->type) {
+                    case drwav_metadata_type_list_info_software:    pID = "ISFT"; break;
+                    case drwav_metadata_type_list_info_copyright:   pID = "ICOP"; break;
+                    case drwav_metadata_type_list_info_title:       pID = "INAM"; break;
+                    case drwav_metadata_type_list_info_artist:      pID = "IART"; break;
+                    case drwav_metadata_type_list_info_comment:     pID = "ICMT"; break;
+                    case drwav_metadata_type_list_info_date:        pID = "ICRD"; break;
+                    case drwav_metadata_type_list_info_genre:       pID = "IGNR"; break;
+                    case drwav_metadata_type_list_info_album:       pID = "IPRD"; break;
+                    case drwav_metadata_type_list_info_tracknumber: pID = "ITRK"; break;
+                    default: break;
+                }
+
+                DRWAV_ASSERT(pID != NULL);
+
+                if (pMetadata->data.infoText.stringLength) {
+                    subchunkSize = pMetadata->data.infoText.stringLength + 1;
+                    bytesWritten += drwav__write_or_count(pWav, pID, 4);
+                    bytesWritten += drwav__write_or_count_u32ne_to_le(pWav, subchunkSize);
+                    bytesWritten += drwav__write_or_count(pWav, pMetadata->data.infoText.pString, pMetadata->data.infoText.stringLength);
+                    bytesWritten += drwav__write_or_count_byte(pWav, '\0');
+                }
+            } else if (pMetadata->type == drwav_metadata_type_unknown && pMetadata->data.unknown.chunkLocation == drwav_metadata_location_inside_info_list) {
+                if (pMetadata->data.unknown.dataSizeInBytes) {
+                    subchunkSize = pMetadata->data.unknown.dataSizeInBytes;
+
+                    bytesWritten += drwav__write_or_count(pWav, pMetadata->data.unknown.id, 4);
+                    bytesWritten += drwav__write_or_count_u32ne_to_le(pWav, pMetadata->data.unknown.dataSizeInBytes);
+                    bytesWritten += drwav__write_or_count(pWav, pMetadata->data.unknown.pData, subchunkSize);
+                }
+            }
+
+            if ((subchunkSize % 2) != 0) {
+                bytesWritten += drwav__write_or_count_byte(pWav, 0);
+            }
+        }
+    }
+
+    if (hasListAdtl) {
+        drwav_uint32 chunkSize = 4; /* start with 4 bytes for "adtl" */
+
+        for (iMetadata = 0; iMetadata < metadataCount; ++iMetadata) {
+            drwav_metadata* pMetadata = &pMetadatas[iMetadata];
+
+            switch (pMetadata->type)
+            {
+                case drwav_metadata_type_list_label:
+                case drwav_metadata_type_list_note:
+                {
+                    chunkSize += 8; /* for id and chunk size */
+                    chunkSize += DRWAV_LIST_LABEL_OR_NOTE_BYTES;
+
+                    if (pMetadata->data.labelOrNote.stringLength > 0) {
+                        chunkSize += pMetadata->data.labelOrNote.stringLength + 1;
+                    }    
+                } break;
+
+                case drwav_metadata_type_list_labelled_cue_region:
+                {
+                    chunkSize += 8; /* for id and chunk size */
+                    chunkSize += DRWAV_LIST_LABELLED_TEXT_BYTES;
+
+                    if (pMetadata->data.labelledCueRegion.stringLength > 0) {
+                        chunkSize += pMetadata->data.labelledCueRegion.stringLength + 1;
+                    }
+                } break;
+
+                case drwav_metadata_type_unknown:
+                {
+                    if (pMetadata->data.unknown.chunkLocation == drwav_metadata_location_inside_adtl_list) {
+                        chunkSize += 8; /* for id and chunk size */
+                        chunkSize += pMetadata->data.unknown.dataSizeInBytes;
+                    }
+                } break;
+
+                default: break;
+            }
+
+            if ((chunkSize % 2) != 0) {
+                chunkSize += 1;
+            }
+        }
+
+        bytesWritten += drwav__write_or_count(pWav, "LIST", 4);
+        bytesWritten += drwav__write_or_count_u32ne_to_le(pWav, chunkSize);
+        bytesWritten += drwav__write_or_count(pWav, "adtl", 4);
+
+        for (iMetadata = 0; iMetadata < metadataCount; ++iMetadata) {
+            drwav_metadata* pMetadata = &pMetadatas[iMetadata];
+            drwav_uint32 subchunkSize = 0;
+
+            switch (pMetadata->type)
+            {
+                case drwav_metadata_type_list_label:
+                case drwav_metadata_type_list_note:
+                {
+                    if (pMetadata->data.labelOrNote.stringLength > 0) {
+                        const char *pID = NULL;
+
+                        if (pMetadata->type == drwav_metadata_type_list_label) {
+                            pID = "labl";
+                        }
+                        else if (pMetadata->type == drwav_metadata_type_list_note) {
+                            pID = "note";
+                        }
+
+                        DRWAV_ASSERT(pID != NULL);
+                        DRWAV_ASSERT(pMetadata->data.labelOrNote.pString != NULL);
+
+                        subchunkSize = DRWAV_LIST_LABEL_OR_NOTE_BYTES;
+
+                        bytesWritten += drwav__write_or_count(pWav, pID, 4);
+                        subchunkSize += pMetadata->data.labelOrNote.stringLength + 1;
+                        bytesWritten += drwav__write_or_count_u32ne_to_le(pWav, subchunkSize);
+
+                        bytesWritten += drwav__write_or_count_u32ne_to_le(pWav, pMetadata->data.labelOrNote.cuePointId);
+                        bytesWritten += drwav__write_or_count(pWav, pMetadata->data.labelOrNote.pString, pMetadata->data.labelOrNote.stringLength);
+                        bytesWritten += drwav__write_or_count_byte(pWav, '\0');
+                    }
+                } break;
+
+                case drwav_metadata_type_list_labelled_cue_region:
+                {
+                    subchunkSize = DRWAV_LIST_LABELLED_TEXT_BYTES;
+
+                    bytesWritten += drwav__write_or_count(pWav, "ltxt", 4);
+                    if (pMetadata->data.labelledCueRegion.stringLength > 0) {
+                        subchunkSize += pMetadata->data.labelledCueRegion.stringLength + 1;
+                    }
+                    bytesWritten += drwav__write_or_count_u32ne_to_le(pWav, subchunkSize);
+                    bytesWritten += drwav__write_or_count_u32ne_to_le(pWav, pMetadata->data.labelledCueRegion.cuePointId);
+                    bytesWritten += drwav__write_or_count_u32ne_to_le(pWav, pMetadata->data.labelledCueRegion.sampleLength);
+                    bytesWritten += drwav__write_or_count(pWav, pMetadata->data.labelledCueRegion.purposeId, 4);
+                    bytesWritten += drwav__write_or_count_u16ne_to_le(pWav, pMetadata->data.labelledCueRegion.country);
+                    bytesWritten += drwav__write_or_count_u16ne_to_le(pWav, pMetadata->data.labelledCueRegion.language);
+                    bytesWritten += drwav__write_or_count_u16ne_to_le(pWav, pMetadata->data.labelledCueRegion.dialect);
+                    bytesWritten += drwav__write_or_count_u16ne_to_le(pWav, pMetadata->data.labelledCueRegion.codePage);
+
+                    if (pMetadata->data.labelledCueRegion.stringLength > 0) {
+                        DRWAV_ASSERT(pMetadata->data.labelledCueRegion.pString != NULL);
+
+                        bytesWritten += drwav__write_or_count(pWav, pMetadata->data.labelledCueRegion.pString, pMetadata->data.labelledCueRegion.stringLength);
+                        bytesWritten += drwav__write_or_count_byte(pWav, '\0');
+                    }
+                } break;
+
+                case drwav_metadata_type_unknown:
+                {
+                    if (pMetadata->data.unknown.chunkLocation == drwav_metadata_location_inside_adtl_list) {
+                        subchunkSize = pMetadata->data.unknown.dataSizeInBytes;
+
+                        DRWAV_ASSERT(pMetadata->data.unknown.pData != NULL);
+                        bytesWritten += drwav__write_or_count(pWav, pMetadata->data.unknown.id, 4);
+                        bytesWritten += drwav__write_or_count_u32ne_to_le(pWav, subchunkSize);
+                        bytesWritten += drwav__write_or_count(pWav, pMetadata->data.unknown.pData, subchunkSize);
+                    }
+                } break;
+
+                default: break;
+            }
+
+            if ((subchunkSize % 2) != 0) {
+                bytesWritten += drwav__write_or_count_byte(pWav, 0);
+            }
+        }
+    }
+
+    DRWAV_ASSERT((bytesWritten % 2) == 0);
+
+    return bytesWritten;
+}
+
+DRWAV_PRIVATE drwav_uint32 drwav__riff_chunk_size_riff(drwav_uint64 dataChunkSize, drwav_metadata* pMetadata, drwav_uint32 metadataCount)
+{
+    drwav_uint64 chunkSize = 4 + 24 + (drwav_uint64)drwav__write_or_count_metadata(NULL, pMetadata, metadataCount) + 8 + dataChunkSize + drwav__chunk_padding_size_riff(dataChunkSize); /* 4 = "WAVE". 24 = "fmt " chunk. 8 = "data" + u32 data size. */
+    if (chunkSize > 0xFFFFFFFFUL) {
+        chunkSize = 0xFFFFFFFFUL;
+    }
+
+    return (drwav_uint32)chunkSize; /* Safe cast due to the clamp above. */
+}
+
+DRWAV_PRIVATE drwav_uint32 drwav__data_chunk_size_riff(drwav_uint64 dataChunkSize)
+{
+    if (dataChunkSize <= 0xFFFFFFFFUL) {
+        return (drwav_uint32)dataChunkSize;
+    } else {
+        return 0xFFFFFFFFUL;
+    }
+}
+
+DRWAV_PRIVATE drwav_uint64 drwav__riff_chunk_size_w64(drwav_uint64 dataChunkSize)
+{
+    drwav_uint64 dataSubchunkPaddingSize = drwav__chunk_padding_size_w64(dataChunkSize);
+
+    return 80 + 24 + dataChunkSize + dataSubchunkPaddingSize;   /* +24 because W64 includes the size of the GUID and size fields. */
+}
+
+DRWAV_PRIVATE drwav_uint64 drwav__data_chunk_size_w64(drwav_uint64 dataChunkSize)
+{
+    return 24 + dataChunkSize;        /* +24 because W64 includes the size of the GUID and size fields. */
+}
+
+DRWAV_PRIVATE drwav_uint64 drwav__riff_chunk_size_rf64(drwav_uint64 dataChunkSize, drwav_metadata *metadata, drwav_uint32 numMetadata)
+{
+    drwav_uint64 chunkSize = 4 + 36 + 24 + (drwav_uint64)drwav__write_or_count_metadata(NULL, metadata, numMetadata) + 8 + dataChunkSize + drwav__chunk_padding_size_riff(dataChunkSize); /* 4 = "WAVE". 36 = "ds64" chunk. 24 = "fmt " chunk. 8 = "data" + u32 data size. */
+    if (chunkSize > 0xFFFFFFFFUL) {
+        chunkSize = 0xFFFFFFFFUL;
+    }
+
+    return chunkSize;
+}
+
+DRWAV_PRIVATE drwav_uint64 drwav__data_chunk_size_rf64(drwav_uint64 dataChunkSize)
+{
+    return dataChunkSize;
+}
+
+
+
+DRWAV_PRIVATE drwav_bool32 drwav_preinit_write(drwav* pWav, const drwav_data_format* pFormat, drwav_bool32 isSequential, drwav_write_proc onWrite, drwav_seek_proc onSeek, void* pUserData, const drwav_allocation_callbacks* pAllocationCallbacks)
+{
+    if (pWav == NULL || onWrite == NULL) {
+        return DRWAV_FALSE;
+    }
+
+    if (!isSequential && onSeek == NULL) {
+        return DRWAV_FALSE; /* <-- onSeek is required when in non-sequential mode. */
+    }
+
+    /* Not currently supporting compressed formats. Will need to add support for the "fact" chunk before we enable this. */
+    if (pFormat->format == DR_WAVE_FORMAT_EXTENSIBLE) {
+        return DRWAV_FALSE;
+    }
+    if (pFormat->format == DR_WAVE_FORMAT_ADPCM || pFormat->format == DR_WAVE_FORMAT_DVI_ADPCM) {
+        return DRWAV_FALSE;
+    }
+
+    DRWAV_ZERO_MEMORY(pWav, sizeof(*pWav));
+    pWav->onWrite   = onWrite;
+    pWav->onSeek    = onSeek;
+    pWav->pUserData = pUserData;
+    pWav->allocationCallbacks = drwav_copy_allocation_callbacks_or_defaults(pAllocationCallbacks);
+
+    if (pWav->allocationCallbacks.onFree == NULL || (pWav->allocationCallbacks.onMalloc == NULL && pWav->allocationCallbacks.onRealloc == NULL)) {
+        return DRWAV_FALSE;    /* Invalid allocation callbacks. */
+    }
+
+    pWav->fmt.formatTag = (drwav_uint16)pFormat->format;
+    pWav->fmt.channels = (drwav_uint16)pFormat->channels;
+    pWav->fmt.sampleRate = pFormat->sampleRate;
+    pWav->fmt.avgBytesPerSec = (drwav_uint32)((pFormat->bitsPerSample * pFormat->sampleRate * pFormat->channels) / 8);
+    pWav->fmt.blockAlign = (drwav_uint16)((pFormat->channels * pFormat->bitsPerSample) / 8);
+    pWav->fmt.bitsPerSample = (drwav_uint16)pFormat->bitsPerSample;
+    pWav->fmt.extendedSize = 0;
+    pWav->isSequentialWrite = isSequential;
+
+    return DRWAV_TRUE;
+}
+
+
+DRWAV_PRIVATE drwav_bool32 drwav_init_write__internal(drwav* pWav, const drwav_data_format* pFormat, drwav_uint64 totalSampleCount)
+{
+    /* The function assumes drwav_preinit_write() was called beforehand. */
+
+    size_t runningPos = 0;
+    drwav_uint64 initialDataChunkSize = 0;
+    drwav_uint64 chunkSizeFMT;
+
+    /*
+    The initial values for the "RIFF" and "data" chunks depends on whether or not we are initializing in sequential mode or not. In
+    sequential mode we set this to its final values straight away since they can be calculated from the total sample count. In non-
+    sequential mode we initialize it all to zero and fill it out in drwav_uninit() using a backwards seek.
+    */
+    if (pWav->isSequentialWrite) {
+        initialDataChunkSize = (totalSampleCount * pWav->fmt.bitsPerSample) / 8;
+
+        /*
+        The RIFF container has a limit on the number of samples. drwav is not allowing this. There's no practical limits for Wave64
+        so for the sake of simplicity I'm not doing any validation for that.
+        */
+        if (pFormat->container == drwav_container_riff) {
+            if (initialDataChunkSize > (0xFFFFFFFFUL - 36)) {
+                return DRWAV_FALSE; /* Not enough room to store every sample. */
+            }
+        }
+    }
+
+    pWav->dataChunkDataSizeTargetWrite = initialDataChunkSize;
+
+
+    /* "RIFF" chunk. */
+    if (pFormat->container == drwav_container_riff) {
+        drwav_uint32 chunkSizeRIFF = 28 + (drwav_uint32)initialDataChunkSize;   /* +28 = "WAVE" + [sizeof "fmt " chunk] */
+        runningPos += drwav__write(pWav, "RIFF", 4);
+        runningPos += drwav__write_u32ne_to_le(pWav, chunkSizeRIFF);
+        runningPos += drwav__write(pWav, "WAVE", 4);
+    } else if (pFormat->container == drwav_container_w64) {
+        drwav_uint64 chunkSizeRIFF = 80 + 24 + initialDataChunkSize;            /* +24 because W64 includes the size of the GUID and size fields. */
+        runningPos += drwav__write(pWav, drwavGUID_W64_RIFF, 16);
+        runningPos += drwav__write_u64ne_to_le(pWav, chunkSizeRIFF);
+        runningPos += drwav__write(pWav, drwavGUID_W64_WAVE, 16);
+    } else if (pFormat->container == drwav_container_rf64) {
+        runningPos += drwav__write(pWav, "RF64", 4);
+        runningPos += drwav__write_u32ne_to_le(pWav, 0xFFFFFFFF);               /* Always 0xFFFFFFFF for RF64. Set to a proper value in the "ds64" chunk. */
+        runningPos += drwav__write(pWav, "WAVE", 4);
+    } else {
+        return DRWAV_FALSE; /* Container not supported for writing. */
+    }
+
+
+    /* "ds64" chunk (RF64 only). */
+    if (pFormat->container == drwav_container_rf64) {
+        drwav_uint32 initialds64ChunkSize = 28;                                 /* 28 = [Size of RIFF (8 bytes)] + [Size of DATA (8 bytes)] + [Sample Count (8 bytes)] + [Table Length (4 bytes)]. Table length always set to 0. */
+        drwav_uint64 initialRiffChunkSize = 8 + initialds64ChunkSize + initialDataChunkSize;    /* +8 for the ds64 header. */
+
+        runningPos += drwav__write(pWav, "ds64", 4);
+        runningPos += drwav__write_u32ne_to_le(pWav, initialds64ChunkSize);     /* Size of ds64. */
+        runningPos += drwav__write_u64ne_to_le(pWav, initialRiffChunkSize);     /* Size of RIFF. Set to true value at the end. */
+        runningPos += drwav__write_u64ne_to_le(pWav, initialDataChunkSize);     /* Size of DATA. Set to true value at the end. */
+        runningPos += drwav__write_u64ne_to_le(pWav, totalSampleCount);         /* Sample count. */
+        runningPos += drwav__write_u32ne_to_le(pWav, 0);                        /* Table length. Always set to zero in our case since we're not doing any other chunks than "DATA". */
+    }
+
+
+    /* "fmt " chunk. */
+    if (pFormat->container == drwav_container_riff || pFormat->container == drwav_container_rf64) {
+        chunkSizeFMT = 16;
+        runningPos += drwav__write(pWav, "fmt ", 4);
+        runningPos += drwav__write_u32ne_to_le(pWav, (drwav_uint32)chunkSizeFMT);
+    } else if (pFormat->container == drwav_container_w64) {
+        chunkSizeFMT = 40;
+        runningPos += drwav__write(pWav, drwavGUID_W64_FMT, 16);
+        runningPos += drwav__write_u64ne_to_le(pWav, chunkSizeFMT);
+    }
+
+    runningPos += drwav__write_u16ne_to_le(pWav, pWav->fmt.formatTag);
+    runningPos += drwav__write_u16ne_to_le(pWav, pWav->fmt.channels);
+    runningPos += drwav__write_u32ne_to_le(pWav, pWav->fmt.sampleRate);
+    runningPos += drwav__write_u32ne_to_le(pWav, pWav->fmt.avgBytesPerSec);
+    runningPos += drwav__write_u16ne_to_le(pWav, pWav->fmt.blockAlign);
+    runningPos += drwav__write_u16ne_to_le(pWav, pWav->fmt.bitsPerSample);
+
+    /* TODO: is a 'fact' chunk required for DR_WAVE_FORMAT_IEEE_FLOAT? */
+
+    if (!pWav->isSequentialWrite && pWav->pMetadata != NULL && pWav->metadataCount > 0 && (pFormat->container == drwav_container_riff || pFormat->container == drwav_container_rf64)) {
+        runningPos += drwav__write_or_count_metadata(pWav, pWav->pMetadata, pWav->metadataCount);
+    }
+
+    pWav->dataChunkDataPos = runningPos;
+
+    /* "data" chunk. */
+    if (pFormat->container == drwav_container_riff) {
+        drwav_uint32 chunkSizeDATA = (drwav_uint32)initialDataChunkSize;
+        runningPos += drwav__write(pWav, "data", 4);
+        runningPos += drwav__write_u32ne_to_le(pWav, chunkSizeDATA);
+    } else if (pFormat->container == drwav_container_w64) {
+        drwav_uint64 chunkSizeDATA = 24 + initialDataChunkSize;     /* +24 because W64 includes the size of the GUID and size fields. */
+        runningPos += drwav__write(pWav, drwavGUID_W64_DATA, 16);
+        runningPos += drwav__write_u64ne_to_le(pWav, chunkSizeDATA);
+    } else if (pFormat->container == drwav_container_rf64) {
+        runningPos += drwav__write(pWav, "data", 4);
+        runningPos += drwav__write_u32ne_to_le(pWav, 0xFFFFFFFF);   /* Always set to 0xFFFFFFFF for RF64. The true size of the data chunk is specified in the ds64 chunk. */
+    }
+
+    /* Set some properties for the client's convenience. */
+    pWav->container = pFormat->container;
+    pWav->channels = (drwav_uint16)pFormat->channels;
+    pWav->sampleRate = pFormat->sampleRate;
+    pWav->bitsPerSample = (drwav_uint16)pFormat->bitsPerSample;
+    pWav->translatedFormatTag = (drwav_uint16)pFormat->format;
+    pWav->dataChunkDataPos = runningPos;
+
+    return DRWAV_TRUE;
+}
+
+
+DRWAV_API drwav_bool32 drwav_init_write(drwav* pWav, const drwav_data_format* pFormat, drwav_write_proc onWrite, drwav_seek_proc onSeek, void* pUserData, const drwav_allocation_callbacks* pAllocationCallbacks)
+{
+    if (!drwav_preinit_write(pWav, pFormat, DRWAV_FALSE, onWrite, onSeek, pUserData, pAllocationCallbacks)) {
+        return DRWAV_FALSE;
+    }
+
+    return drwav_init_write__internal(pWav, pFormat, 0);               /* DRWAV_FALSE = Not Sequential */
+}
+
+DRWAV_API drwav_bool32 drwav_init_write_sequential(drwav* pWav, const drwav_data_format* pFormat, drwav_uint64 totalSampleCount, drwav_write_proc onWrite, void* pUserData, const drwav_allocation_callbacks* pAllocationCallbacks)
+{
+    if (!drwav_preinit_write(pWav, pFormat, DRWAV_TRUE, onWrite, NULL, pUserData, pAllocationCallbacks)) {
+        return DRWAV_FALSE;
+    }
+
+    return drwav_init_write__internal(pWav, pFormat, totalSampleCount); /* DRWAV_TRUE = Sequential */
+}
+
+DRWAV_API drwav_bool32 drwav_init_write_sequential_pcm_frames(drwav* pWav, const drwav_data_format* pFormat, drwav_uint64 totalPCMFrameCount, drwav_write_proc onWrite, void* pUserData, const drwav_allocation_callbacks* pAllocationCallbacks)
+{
+    if (pFormat == NULL) {
+        return DRWAV_FALSE;
+    }
+
+    return drwav_init_write_sequential(pWav, pFormat, totalPCMFrameCount*pFormat->channels, onWrite, pUserData, pAllocationCallbacks);
+}
+
+DRWAV_API drwav_bool32 drwav_init_write_with_metadata(drwav* pWav, const drwav_data_format* pFormat, drwav_write_proc onWrite, drwav_seek_proc onSeek, void* pUserData, const drwav_allocation_callbacks* pAllocationCallbacks, drwav_metadata* pMetadata, drwav_uint32 metadataCount)
+{
+    if (!drwav_preinit_write(pWav, pFormat, DRWAV_FALSE, onWrite, onSeek, pUserData, pAllocationCallbacks)) {
+        return DRWAV_FALSE;
+    }
+
+    pWav->pMetadata     = pMetadata;
+    pWav->metadataCount = metadataCount;
+
+    return drwav_init_write__internal(pWav, pFormat, 0);
+}
+
+
+DRWAV_API drwav_uint64 drwav_target_write_size_bytes(const drwav_data_format* pFormat, drwav_uint64 totalFrameCount, drwav_metadata* pMetadata, drwav_uint32 metadataCount)
+{
+    /* Casting totalFrameCount to drwav_int64 for VC6 compatibility. No issues in practice because nobody is going to exhaust the whole 63 bits. */
+    drwav_uint64 targetDataSizeBytes = (drwav_uint64)((drwav_int64)totalFrameCount * pFormat->channels * pFormat->bitsPerSample/8.0);
+    drwav_uint64 riffChunkSizeBytes;
+    drwav_uint64 fileSizeBytes = 0;
+
+    if (pFormat->container == drwav_container_riff) {
+        riffChunkSizeBytes = drwav__riff_chunk_size_riff(targetDataSizeBytes, pMetadata, metadataCount);
+        fileSizeBytes = (8 + riffChunkSizeBytes);   /* +8 because WAV doesn't include the size of the ChunkID and ChunkSize fields. */
+    } else if (pFormat->container == drwav_container_w64) {
+        riffChunkSizeBytes = drwav__riff_chunk_size_w64(targetDataSizeBytes);
+        fileSizeBytes = riffChunkSizeBytes;
+    } else if (pFormat->container == drwav_container_rf64) {
+        riffChunkSizeBytes = drwav__riff_chunk_size_rf64(targetDataSizeBytes, pMetadata, metadataCount);
+        fileSizeBytes = (8 + riffChunkSizeBytes);   /* +8 because WAV doesn't include the size of the ChunkID and ChunkSize fields. */
+    }
+
+    return fileSizeBytes;
+}
+
+
+#ifndef DR_WAV_NO_STDIO
+
+/* Errno */
+/* drwav_result_from_errno() is only used for fopen() and wfopen() so putting it inside DR_WAV_NO_STDIO for now. If something else needs this later we can move it out. */
+#include <errno.h>
+DRWAV_PRIVATE drwav_result drwav_result_from_errno(int e)
+{
+    switch (e)
+    {
+        case 0: return DRWAV_SUCCESS;
+    #ifdef EPERM
+        case EPERM: return DRWAV_INVALID_OPERATION;
+    #endif
+    #ifdef ENOENT
+        case ENOENT: return DRWAV_DOES_NOT_EXIST;
+    #endif
+    #ifdef ESRCH
+        case ESRCH: return DRWAV_DOES_NOT_EXIST;
+    #endif
+    #ifdef EINTR
+        case EINTR: return DRWAV_INTERRUPT;
+    #endif
+    #ifdef EIO
+        case EIO: return DRWAV_IO_ERROR;
+    #endif
+    #ifdef ENXIO
+        case ENXIO: return DRWAV_DOES_NOT_EXIST;
+    #endif
+    #ifdef E2BIG
+        case E2BIG: return DRWAV_INVALID_ARGS;
+    #endif
+    #ifdef ENOEXEC
+        case ENOEXEC: return DRWAV_INVALID_FILE;
+    #endif
+    #ifdef EBADF
+        case EBADF: return DRWAV_INVALID_FILE;
+    #endif
+    #ifdef ECHILD
+        case ECHILD: return DRWAV_ERROR;
+    #endif
+    #ifdef EAGAIN
+        case EAGAIN: return DRWAV_UNAVAILABLE;
+    #endif
+    #ifdef ENOMEM
+        case ENOMEM: return DRWAV_OUT_OF_MEMORY;
+    #endif
+    #ifdef EACCES
+        case EACCES: return DRWAV_ACCESS_DENIED;
+    #endif
+    #ifdef EFAULT
+        case EFAULT: return DRWAV_BAD_ADDRESS;
+    #endif
+    #ifdef ENOTBLK
+        case ENOTBLK: return DRWAV_ERROR;
+    #endif
+    #ifdef EBUSY
+        case EBUSY: return DRWAV_BUSY;
+    #endif
+    #ifdef EEXIST
+        case EEXIST: return DRWAV_ALREADY_EXISTS;
+    #endif
+    #ifdef EXDEV
+        case EXDEV: return DRWAV_ERROR;
+    #endif
+    #ifdef ENODEV
+        case ENODEV: return DRWAV_DOES_NOT_EXIST;
+    #endif
+    #ifdef ENOTDIR
+        case ENOTDIR: return DRWAV_NOT_DIRECTORY;
+    #endif
+    #ifdef EISDIR
+        case EISDIR: return DRWAV_IS_DIRECTORY;
+    #endif
+    #ifdef EINVAL
+        case EINVAL: return DRWAV_INVALID_ARGS;
+    #endif
+    #ifdef ENFILE
+        case ENFILE: return DRWAV_TOO_MANY_OPEN_FILES;
+    #endif
+    #ifdef EMFILE
+        case EMFILE: return DRWAV_TOO_MANY_OPEN_FILES;
+    #endif
+    #ifdef ENOTTY
+        case ENOTTY: return DRWAV_INVALID_OPERATION;
+    #endif
+    #ifdef ETXTBSY
+        case ETXTBSY: return DRWAV_BUSY;
+    #endif
+    #ifdef EFBIG
+        case EFBIG: return DRWAV_TOO_BIG;
+    #endif
+    #ifdef ENOSPC
+        case ENOSPC: return DRWAV_NO_SPACE;
+    #endif
+    #ifdef ESPIPE
+        case ESPIPE: return DRWAV_BAD_SEEK;
+    #endif
+    #ifdef EROFS
+        case EROFS: return DRWAV_ACCESS_DENIED;
+    #endif
+    #ifdef EMLINK
+        case EMLINK: return DRWAV_TOO_MANY_LINKS;
+    #endif
+    #ifdef EPIPE
+        case EPIPE: return DRWAV_BAD_PIPE;
+    #endif
+    #ifdef EDOM
+        case EDOM: return DRWAV_OUT_OF_RANGE;
+    #endif
+    #ifdef ERANGE
+        case ERANGE: return DRWAV_OUT_OF_RANGE;
+    #endif
+    #ifdef EDEADLK
+        case EDEADLK: return DRWAV_DEADLOCK;
+    #endif
+    #ifdef ENAMETOOLONG
+        case ENAMETOOLONG: return DRWAV_PATH_TOO_LONG;
+    #endif
+    #ifdef ENOLCK
+        case ENOLCK: return DRWAV_ERROR;
+    #endif
+    #ifdef ENOSYS
+        case ENOSYS: return DRWAV_NOT_IMPLEMENTED;
+    #endif
+    #ifdef ENOTEMPTY
+        case ENOTEMPTY: return DRWAV_DIRECTORY_NOT_EMPTY;
+    #endif
+    #ifdef ELOOP
+        case ELOOP: return DRWAV_TOO_MANY_LINKS;
+    #endif
+    #ifdef ENOMSG
+        case ENOMSG: return DRWAV_NO_MESSAGE;
+    #endif
+    #ifdef EIDRM
+        case EIDRM: return DRWAV_ERROR;
+    #endif
+    #ifdef ECHRNG
+        case ECHRNG: return DRWAV_ERROR;
+    #endif
+    #ifdef EL2NSYNC
+        case EL2NSYNC: return DRWAV_ERROR;
+    #endif
+    #ifdef EL3HLT
+        case EL3HLT: return DRWAV_ERROR;
+    #endif
+    #ifdef EL3RST
+        case EL3RST: return DRWAV_ERROR;
+    #endif
+    #ifdef ELNRNG
+        case ELNRNG: return DRWAV_OUT_OF_RANGE;
+    #endif
+    #ifdef EUNATCH
+        case EUNATCH: return DRWAV_ERROR;
+    #endif
+    #ifdef ENOCSI
+        case ENOCSI: return DRWAV_ERROR;
+    #endif
+    #ifdef EL2HLT
+        case EL2HLT: return DRWAV_ERROR;
+    #endif
+    #ifdef EBADE
+        case EBADE: return DRWAV_ERROR;
+    #endif
+    #ifdef EBADR
+        case EBADR: return DRWAV_ERROR;
+    #endif
+    #ifdef EXFULL
+        case EXFULL: return DRWAV_ERROR;
+    #endif
+    #ifdef ENOANO
+        case ENOANO: return DRWAV_ERROR;
+    #endif
+    #ifdef EBADRQC
+        case EBADRQC: return DRWAV_ERROR;
+    #endif
+    #ifdef EBADSLT
+        case EBADSLT: return DRWAV_ERROR;
+    #endif
+    #ifdef EBFONT
+        case EBFONT: return DRWAV_INVALID_FILE;
+    #endif
+    #ifdef ENOSTR
+        case ENOSTR: return DRWAV_ERROR;
+    #endif
+    #ifdef ENODATA
+        case ENODATA: return DRWAV_NO_DATA_AVAILABLE;
+    #endif
+    #ifdef ETIME
+        case ETIME: return DRWAV_TIMEOUT;
+    #endif
+    #ifdef ENOSR
+        case ENOSR: return DRWAV_NO_DATA_AVAILABLE;
+    #endif
+    #ifdef ENONET
+        case ENONET: return DRWAV_NO_NETWORK;
+    #endif
+    #ifdef ENOPKG
+        case ENOPKG: return DRWAV_ERROR;
+    #endif
+    #ifdef EREMOTE
+        case EREMOTE: return DRWAV_ERROR;
+    #endif
+    #ifdef ENOLINK
+        case ENOLINK: return DRWAV_ERROR;
+    #endif
+    #ifdef EADV
+        case EADV: return DRWAV_ERROR;
+    #endif
+    #ifdef ESRMNT
+        case ESRMNT: return DRWAV_ERROR;
+    #endif
+    #ifdef ECOMM
+        case ECOMM: return DRWAV_ERROR;
+    #endif
+    #ifdef EPROTO
+        case EPROTO: return DRWAV_ERROR;
+    #endif
+    #ifdef EMULTIHOP
+        case EMULTIHOP: return DRWAV_ERROR;
+    #endif
+    #ifdef EDOTDOT
+        case EDOTDOT: return DRWAV_ERROR;
+    #endif
+    #ifdef EBADMSG
+        case EBADMSG: return DRWAV_BAD_MESSAGE;
+    #endif
+    #ifdef EOVERFLOW
+        case EOVERFLOW: return DRWAV_TOO_BIG;
+    #endif
+    #ifdef ENOTUNIQ
+        case ENOTUNIQ: return DRWAV_NOT_UNIQUE;
+    #endif
+    #ifdef EBADFD
+        case EBADFD: return DRWAV_ERROR;
+    #endif
+    #ifdef EREMCHG
+        case EREMCHG: return DRWAV_ERROR;
+    #endif
+    #ifdef ELIBACC
+        case ELIBACC: return DRWAV_ACCESS_DENIED;
+    #endif
+    #ifdef ELIBBAD
+        case ELIBBAD: return DRWAV_INVALID_FILE;
+    #endif
+    #ifdef ELIBSCN
+        case ELIBSCN: return DRWAV_INVALID_FILE;
+    #endif
+    #ifdef ELIBMAX
+        case ELIBMAX: return DRWAV_ERROR;
+    #endif
+    #ifdef ELIBEXEC
+        case ELIBEXEC: return DRWAV_ERROR;
+    #endif
+    #ifdef EILSEQ
+        case EILSEQ: return DRWAV_INVALID_DATA;
+    #endif
+    #ifdef ERESTART
+        case ERESTART: return DRWAV_ERROR;
+    #endif
+    #ifdef ESTRPIPE
+        case ESTRPIPE: return DRWAV_ERROR;
+    #endif
+    #ifdef EUSERS
+        case EUSERS: return DRWAV_ERROR;
+    #endif
+    #ifdef ENOTSOCK
+        case ENOTSOCK: return DRWAV_NOT_SOCKET;
+    #endif
+    #ifdef EDESTADDRREQ
+        case EDESTADDRREQ: return DRWAV_NO_ADDRESS;
+    #endif
+    #ifdef EMSGSIZE
+        case EMSGSIZE: return DRWAV_TOO_BIG;
+    #endif
+    #ifdef EPROTOTYPE
+        case EPROTOTYPE: return DRWAV_BAD_PROTOCOL;
+    #endif
+    #ifdef ENOPROTOOPT
+        case ENOPROTOOPT: return DRWAV_PROTOCOL_UNAVAILABLE;
+    #endif
+    #ifdef EPROTONOSUPPORT
+        case EPROTONOSUPPORT: return DRWAV_PROTOCOL_NOT_SUPPORTED;
+    #endif
+    #ifdef ESOCKTNOSUPPORT
+        case ESOCKTNOSUPPORT: return DRWAV_SOCKET_NOT_SUPPORTED;
+    #endif
+    #ifdef EOPNOTSUPP
+        case EOPNOTSUPP: return DRWAV_INVALID_OPERATION;
+    #endif
+    #ifdef EPFNOSUPPORT
+        case EPFNOSUPPORT: return DRWAV_PROTOCOL_FAMILY_NOT_SUPPORTED;
+    #endif
+    #ifdef EAFNOSUPPORT
+        case EAFNOSUPPORT: return DRWAV_ADDRESS_FAMILY_NOT_SUPPORTED;
+    #endif
+    #ifdef EADDRINUSE
+        case EADDRINUSE: return DRWAV_ALREADY_IN_USE;
+    #endif
+    #ifdef EADDRNOTAVAIL
+        case EADDRNOTAVAIL: return DRWAV_ERROR;
+    #endif
+    #ifdef ENETDOWN
+        case ENETDOWN: return DRWAV_NO_NETWORK;
+    #endif
+    #ifdef ENETUNREACH
+        case ENETUNREACH: return DRWAV_NO_NETWORK;
+    #endif
+    #ifdef ENETRESET
+        case ENETRESET: return DRWAV_NO_NETWORK;
+    #endif
+    #ifdef ECONNABORTED
+        case ECONNABORTED: return DRWAV_NO_NETWORK;
+    #endif
+    #ifdef ECONNRESET
+        case ECONNRESET: return DRWAV_CONNECTION_RESET;
+    #endif
+    #ifdef ENOBUFS
+        case ENOBUFS: return DRWAV_NO_SPACE;
+    #endif
+    #ifdef EISCONN
+        case EISCONN: return DRWAV_ALREADY_CONNECTED;
+    #endif
+    #ifdef ENOTCONN
+        case ENOTCONN: return DRWAV_NOT_CONNECTED;
+    #endif
+    #ifdef ESHUTDOWN
+        case ESHUTDOWN: return DRWAV_ERROR;
+    #endif
+    #ifdef ETOOMANYREFS
+        case ETOOMANYREFS: return DRWAV_ERROR;
+    #endif
+    #ifdef ETIMEDOUT
+        case ETIMEDOUT: return DRWAV_TIMEOUT;
+    #endif
+    #ifdef ECONNREFUSED
+        case ECONNREFUSED: return DRWAV_CONNECTION_REFUSED;
+    #endif
+    #ifdef EHOSTDOWN
+        case EHOSTDOWN: return DRWAV_NO_HOST;
+    #endif
+    #ifdef EHOSTUNREACH
+        case EHOSTUNREACH: return DRWAV_NO_HOST;
+    #endif
+    #ifdef EALREADY
+        case EALREADY: return DRWAV_IN_PROGRESS;
+    #endif
+    #ifdef EINPROGRESS
+        case EINPROGRESS: return DRWAV_IN_PROGRESS;
+    #endif
+    #ifdef ESTALE
+        case ESTALE: return DRWAV_INVALID_FILE;
+    #endif
+    #ifdef EUCLEAN
+        case EUCLEAN: return DRWAV_ERROR;
+    #endif
+    #ifdef ENOTNAM
+        case ENOTNAM: return DRWAV_ERROR;
+    #endif
+    #ifdef ENAVAIL
+        case ENAVAIL: return DRWAV_ERROR;
+    #endif
+    #ifdef EISNAM
+        case EISNAM: return DRWAV_ERROR;
+    #endif
+    #ifdef EREMOTEIO
+        case EREMOTEIO: return DRWAV_IO_ERROR;
+    #endif
+    #ifdef EDQUOT
+        case EDQUOT: return DRWAV_NO_SPACE;
+    #endif
+    #ifdef ENOMEDIUM
+        case ENOMEDIUM: return DRWAV_DOES_NOT_EXIST;
+    #endif
+    #ifdef EMEDIUMTYPE
+        case EMEDIUMTYPE: return DRWAV_ERROR;
+    #endif
+    #ifdef ECANCELED
+        case ECANCELED: return DRWAV_CANCELLED;
+    #endif
+    #ifdef ENOKEY
+        case ENOKEY: return DRWAV_ERROR;
+    #endif
+    #ifdef EKEYEXPIRED
+        case EKEYEXPIRED: return DRWAV_ERROR;
+    #endif
+    #ifdef EKEYREVOKED
+        case EKEYREVOKED: return DRWAV_ERROR;
+    #endif
+    #ifdef EKEYREJECTED
+        case EKEYREJECTED: return DRWAV_ERROR;
+    #endif
+    #ifdef EOWNERDEAD
+        case EOWNERDEAD: return DRWAV_ERROR;
+    #endif
+    #ifdef ENOTRECOVERABLE
+        case ENOTRECOVERABLE: return DRWAV_ERROR;
+    #endif
+    #ifdef ERFKILL
+        case ERFKILL: return DRWAV_ERROR;
+    #endif
+    #ifdef EHWPOISON
+        case EHWPOISON: return DRWAV_ERROR;
+    #endif
+        default: return DRWAV_ERROR;
+    }
+}
+/* End Errno */
+
+/* fopen */
+DRWAV_PRIVATE drwav_result drwav_fopen(FILE** ppFile, const char* pFilePath, const char* pOpenMode)
+{
+#if defined(_MSC_VER) && _MSC_VER >= 1400
+    errno_t err;
+#endif
+
+    if (ppFile != NULL) {
+        *ppFile = NULL;  /* Safety. */
+    }
+
+    if (pFilePath == NULL || pOpenMode == NULL || ppFile == NULL) {
+        return DRWAV_INVALID_ARGS;
+    }
+
+#if defined(_MSC_VER) && _MSC_VER >= 1400
+    err = fopen_s(ppFile, pFilePath, pOpenMode);
+    if (err != 0) {
+        return drwav_result_from_errno(err);
+    }
+#else
+#if defined(_WIN32) || defined(__APPLE__)
+    *ppFile = fopen(pFilePath, pOpenMode);
+#else
+    #if defined(_FILE_OFFSET_BITS) && _FILE_OFFSET_BITS == 64 && defined(_LARGEFILE64_SOURCE)
+        *ppFile = fopen64(pFilePath, pOpenMode);
+    #else
+        *ppFile = fopen(pFilePath, pOpenMode);
+    #endif
+#endif
+    if (*ppFile == NULL) {
+        drwav_result result = drwav_result_from_errno(errno);
+        if (result == DRWAV_SUCCESS) {
+            result = DRWAV_ERROR;   /* Just a safety check to make sure we never ever return success when pFile == NULL. */
+        }
+
+        return result;
+    }
+#endif
+
+    return DRWAV_SUCCESS;
+}
+
+/*
+_wfopen() isn't always available in all compilation environments.
+
+    * Windows only.
+    * MSVC seems to support it universally as far back as VC6 from what I can tell (haven't checked further back).
+    * MinGW-64 (both 32- and 64-bit) seems to support it.
+    * MinGW wraps it in !defined(__STRICT_ANSI__).
+    * OpenWatcom wraps it in !defined(_NO_EXT_KEYS).
+
+This can be reviewed as compatibility issues arise. The preference is to use _wfopen_s() and _wfopen() as opposed to the wcsrtombs()
+fallback, so if you notice your compiler not detecting this properly I'm happy to look at adding support.
+*/
+#if defined(_WIN32)
+    #if defined(_MSC_VER) || defined(__MINGW64__) || (!defined(__STRICT_ANSI__) && !defined(_NO_EXT_KEYS))
+        #define DRWAV_HAS_WFOPEN
+    #endif
+#endif
+
+#ifndef DR_WAV_NO_WCHAR
+DRWAV_PRIVATE drwav_result drwav_wfopen(FILE** ppFile, const wchar_t* pFilePath, const wchar_t* pOpenMode, const drwav_allocation_callbacks* pAllocationCallbacks)
+{
+    if (ppFile != NULL) {
+        *ppFile = NULL;  /* Safety. */
+    }
+
+    if (pFilePath == NULL || pOpenMode == NULL || ppFile == NULL) {
+        return DRWAV_INVALID_ARGS;
+    }
+
+#if defined(DRWAV_HAS_WFOPEN)
+    {
+        /* Use _wfopen() on Windows. */
+    #if defined(_MSC_VER) && _MSC_VER >= 1400
+        errno_t err = _wfopen_s(ppFile, pFilePath, pOpenMode);
+        if (err != 0) {
+            return drwav_result_from_errno(err);
+        }
+    #else
+        *ppFile = _wfopen(pFilePath, pOpenMode);
+        if (*ppFile == NULL) {
+            return drwav_result_from_errno(errno);
+        }
+    #endif
+        (void)pAllocationCallbacks;
+    }
+#else
+	/*
+    Use fopen() on anything other than Windows. Requires a conversion. This is annoying because
+	fopen() is locale specific. The only real way I can think of to do this is with wcsrtombs(). Note
+	that wcstombs() is apparently not thread-safe because it uses a static global mbstate_t object for
+    maintaining state. I've checked this with -std=c89 and it works, but if somebody get's a compiler
+	error I'll look into improving compatibility.
+    */
+
+	/*
+	Some compilers don't support wchar_t or wcsrtombs() which we're using below. In this case we just
+	need to abort with an error. If you encounter a compiler lacking such support, add it to this list
+	and submit a bug report and it'll be added to the library upstream.
+	*/
+	#if defined(__DJGPP__)
+	{
+		/* Nothing to do here. This will fall through to the error check below. */
+	}
+	#else
+    {
+        mbstate_t mbs;
+        size_t lenMB;
+        const wchar_t* pFilePathTemp = pFilePath;
+        char* pFilePathMB = NULL;
+        char pOpenModeMB[32] = {0};
+
+        /* Get the length first. */
+        DRWAV_ZERO_OBJECT(&mbs);
+        lenMB = wcsrtombs(NULL, &pFilePathTemp, 0, &mbs);
+        if (lenMB == (size_t)-1) {
+            return drwav_result_from_errno(errno);
+        }
+
+        pFilePathMB = (char*)drwav__malloc_from_callbacks(lenMB + 1, pAllocationCallbacks);
+        if (pFilePathMB == NULL) {
+            return DRWAV_OUT_OF_MEMORY;
+        }
+
+        pFilePathTemp = pFilePath;
+        DRWAV_ZERO_OBJECT(&mbs);
+        wcsrtombs(pFilePathMB, &pFilePathTemp, lenMB + 1, &mbs);
+
+        /* The open mode should always consist of ASCII characters so we should be able to do a trivial conversion. */
+        {
+            size_t i = 0;
+            for (;;) {
+                if (pOpenMode[i] == 0) {
+                    pOpenModeMB[i] = '\0';
+                    break;
+                }
+
+                pOpenModeMB[i] = (char)pOpenMode[i];
+                i += 1;
+            }
+        }
+
+        *ppFile = fopen(pFilePathMB, pOpenModeMB);
+
+        drwav__free_from_callbacks(pFilePathMB, pAllocationCallbacks);
+    }
+	#endif
+
+    if (*ppFile == NULL) {
+        return DRWAV_ERROR;
+    }
+#endif
+
+    return DRWAV_SUCCESS;
+}
+#endif
+/* End fopen */
+
+
+DRWAV_PRIVATE size_t drwav__on_read_stdio(void* pUserData, void* pBufferOut, size_t bytesToRead)
+{
+    return fread(pBufferOut, 1, bytesToRead, (FILE*)pUserData);
+}
+
+DRWAV_PRIVATE size_t drwav__on_write_stdio(void* pUserData, const void* pData, size_t bytesToWrite)
+{
+    return fwrite(pData, 1, bytesToWrite, (FILE*)pUserData);
+}
+
+DRWAV_PRIVATE drwav_bool32 drwav__on_seek_stdio(void* pUserData, int offset, drwav_seek_origin origin)
+{
+    return fseek((FILE*)pUserData, offset, (origin == drwav_seek_origin_current) ? SEEK_CUR : SEEK_SET) == 0;
+}
+
+DRWAV_API drwav_bool32 drwav_init_file(drwav* pWav, const char* filename, const drwav_allocation_callbacks* pAllocationCallbacks)
+{
+    return drwav_init_file_ex(pWav, filename, NULL, NULL, 0, pAllocationCallbacks);
+}
+
+
+DRWAV_PRIVATE drwav_bool32 drwav_init_file__internal_FILE(drwav* pWav, FILE* pFile, drwav_chunk_proc onChunk, void* pChunkUserData, drwav_uint32 flags, const drwav_allocation_callbacks* pAllocationCallbacks)
+{
+    drwav_bool32 result;
+
+    result = drwav_preinit(pWav, drwav__on_read_stdio, drwav__on_seek_stdio, (void*)pFile, pAllocationCallbacks);
+    if (result != DRWAV_TRUE) {
+        fclose(pFile);
+        return result;
+    }
+    
+    result = drwav_init__internal(pWav, onChunk, pChunkUserData, flags);
+    if (result != DRWAV_TRUE) {
+        fclose(pFile);
+        return result;
+    }
+
+    return DRWAV_TRUE;
+}
+
+DRWAV_API drwav_bool32 drwav_init_file_ex(drwav* pWav, const char* filename, drwav_chunk_proc onChunk, void* pChunkUserData, drwav_uint32 flags, const drwav_allocation_callbacks* pAllocationCallbacks)
+{
+    FILE* pFile;
+    if (drwav_fopen(&pFile, filename, "rb") != DRWAV_SUCCESS) {
+        return DRWAV_FALSE;
+    }
+
+    /* This takes ownership of the FILE* object. */
+    return drwav_init_file__internal_FILE(pWav, pFile, onChunk, pChunkUserData, flags, pAllocationCallbacks);
+}
+
+#ifndef DR_WAV_NO_WCHAR
+DRWAV_API drwav_bool32 drwav_init_file_w(drwav* pWav, const wchar_t* filename, const drwav_allocation_callbacks* pAllocationCallbacks)
+{
+    return drwav_init_file_ex_w(pWav, filename, NULL, NULL, 0, pAllocationCallbacks);
+}
+
+DRWAV_API drwav_bool32 drwav_init_file_ex_w(drwav* pWav, const wchar_t* filename, drwav_chunk_proc onChunk, void* pChunkUserData, drwav_uint32 flags, const drwav_allocation_callbacks* pAllocationCallbacks)
+{
+    FILE* pFile;
+    if (drwav_wfopen(&pFile, filename, L"rb", pAllocationCallbacks) != DRWAV_SUCCESS) {
+        return DRWAV_FALSE;
+    }
+
+    /* This takes ownership of the FILE* object. */
+    return drwav_init_file__internal_FILE(pWav, pFile, onChunk, pChunkUserData, flags, pAllocationCallbacks);
+}
+#endif
+
+DRWAV_API drwav_bool32 drwav_init_file_with_metadata(drwav* pWav, const char* filename, drwav_uint32 flags, const drwav_allocation_callbacks* pAllocationCallbacks)
+{
+    FILE* pFile;
+    if (drwav_fopen(&pFile, filename, "rb") != DRWAV_SUCCESS) {
+        return DRWAV_FALSE;
+    }
+
+    /* This takes ownership of the FILE* object. */
+    return drwav_init_file__internal_FILE(pWav, pFile, NULL, NULL, flags | DRWAV_WITH_METADATA, pAllocationCallbacks);
+}
+
+#ifndef DR_WAV_NO_WCHAR
+DRWAV_API drwav_bool32 drwav_init_file_with_metadata_w(drwav* pWav, const wchar_t* filename, drwav_uint32 flags, const drwav_allocation_callbacks* pAllocationCallbacks)
+{
+    FILE* pFile;
+    if (drwav_wfopen(&pFile, filename, L"rb", pAllocationCallbacks) != DRWAV_SUCCESS) {
+        return DRWAV_FALSE;
+    }
+
+    /* This takes ownership of the FILE* object. */
+    return drwav_init_file__internal_FILE(pWav, pFile, NULL, NULL, flags | DRWAV_WITH_METADATA, pAllocationCallbacks);
+}
+#endif
+
+
+DRWAV_PRIVATE drwav_bool32 drwav_init_file_write__internal_FILE(drwav* pWav, FILE* pFile, const drwav_data_format* pFormat, drwav_uint64 totalSampleCount, drwav_bool32 isSequential, const drwav_allocation_callbacks* pAllocationCallbacks)
+{
+    drwav_bool32 result;
+
+    result = drwav_preinit_write(pWav, pFormat, isSequential, drwav__on_write_stdio, drwav__on_seek_stdio, (void*)pFile, pAllocationCallbacks);
+    if (result != DRWAV_TRUE) {
+        fclose(pFile);
+        return result;
+    }
+
+    result = drwav_init_write__internal(pWav, pFormat, totalSampleCount);
+    if (result != DRWAV_TRUE) {
+        fclose(pFile);
+        return result;
+    }
+
+    return DRWAV_TRUE;
+}
+
+DRWAV_PRIVATE drwav_bool32 drwav_init_file_write__internal(drwav* pWav, const char* filename, const drwav_data_format* pFormat, drwav_uint64 totalSampleCount, drwav_bool32 isSequential, const drwav_allocation_callbacks* pAllocationCallbacks)
+{
+    FILE* pFile;
+    if (drwav_fopen(&pFile, filename, "wb") != DRWAV_SUCCESS) {
+        return DRWAV_FALSE;
+    }
+
+    /* This takes ownership of the FILE* object. */
+    return drwav_init_file_write__internal_FILE(pWav, pFile, pFormat, totalSampleCount, isSequential, pAllocationCallbacks);
+}
+
+#ifndef DR_WAV_NO_WCHAR
+DRWAV_PRIVATE drwav_bool32 drwav_init_file_write_w__internal(drwav* pWav, const wchar_t* filename, const drwav_data_format* pFormat, drwav_uint64 totalSampleCount, drwav_bool32 isSequential, const drwav_allocation_callbacks* pAllocationCallbacks)
+{
+    FILE* pFile;
+    if (drwav_wfopen(&pFile, filename, L"wb", pAllocationCallbacks) != DRWAV_SUCCESS) {
+        return DRWAV_FALSE;
+    }
+
+    /* This takes ownership of the FILE* object. */
+    return drwav_init_file_write__internal_FILE(pWav, pFile, pFormat, totalSampleCount, isSequential, pAllocationCallbacks);
+}
+#endif
+
+DRWAV_API drwav_bool32 drwav_init_file_write(drwav* pWav, const char* filename, const drwav_data_format* pFormat, const drwav_allocation_callbacks* pAllocationCallbacks)
+{
+    return drwav_init_file_write__internal(pWav, filename, pFormat, 0, DRWAV_FALSE, pAllocationCallbacks);
+}
+
+DRWAV_API drwav_bool32 drwav_init_file_write_sequential(drwav* pWav, const char* filename, const drwav_data_format* pFormat, drwav_uint64 totalSampleCount, const drwav_allocation_callbacks* pAllocationCallbacks)
+{
+    return drwav_init_file_write__internal(pWav, filename, pFormat, totalSampleCount, DRWAV_TRUE, pAllocationCallbacks);
+}
+
+DRWAV_API drwav_bool32 drwav_init_file_write_sequential_pcm_frames(drwav* pWav, const char* filename, const drwav_data_format* pFormat, drwav_uint64 totalPCMFrameCount, const drwav_allocation_callbacks* pAllocationCallbacks)
+{
+    if (pFormat == NULL) {
+        return DRWAV_FALSE;
+    }
+
+    return drwav_init_file_write_sequential(pWav, filename, pFormat, totalPCMFrameCount*pFormat->channels, pAllocationCallbacks);
+}
+
+#ifndef DR_WAV_NO_WCHAR
+DRWAV_API drwav_bool32 drwav_init_file_write_w(drwav* pWav, const wchar_t* filename, const drwav_data_format* pFormat, const drwav_allocation_callbacks* pAllocationCallbacks)
+{
+    return drwav_init_file_write_w__internal(pWav, filename, pFormat, 0, DRWAV_FALSE, pAllocationCallbacks);
+}
+
+DRWAV_API drwav_bool32 drwav_init_file_write_sequential_w(drwav* pWav, const wchar_t* filename, const drwav_data_format* pFormat, drwav_uint64 totalSampleCount, const drwav_allocation_callbacks* pAllocationCallbacks)
+{
+    return drwav_init_file_write_w__internal(pWav, filename, pFormat, totalSampleCount, DRWAV_TRUE, pAllocationCallbacks);
+}
+
+DRWAV_API drwav_bool32 drwav_init_file_write_sequential_pcm_frames_w(drwav* pWav, const wchar_t* filename, const drwav_data_format* pFormat, drwav_uint64 totalPCMFrameCount, const drwav_allocation_callbacks* pAllocationCallbacks)
+{
+    if (pFormat == NULL) {
+        return DRWAV_FALSE;
+    }
+
+    return drwav_init_file_write_sequential_w(pWav, filename, pFormat, totalPCMFrameCount*pFormat->channels, pAllocationCallbacks);
+}
+#endif
+#endif  /* DR_WAV_NO_STDIO */
+
+
+DRWAV_PRIVATE size_t drwav__on_read_memory(void* pUserData, void* pBufferOut, size_t bytesToRead)
+{
+    drwav* pWav = (drwav*)pUserData;
+    size_t bytesRemaining;
+
+    DRWAV_ASSERT(pWav != NULL);
+    DRWAV_ASSERT(pWav->memoryStream.dataSize >= pWav->memoryStream.currentReadPos);
+
+    bytesRemaining = pWav->memoryStream.dataSize - pWav->memoryStream.currentReadPos;
+    if (bytesToRead > bytesRemaining) {
+        bytesToRead = bytesRemaining;
+    }
+
+    if (bytesToRead > 0) {
+        DRWAV_COPY_MEMORY(pBufferOut, pWav->memoryStream.data + pWav->memoryStream.currentReadPos, bytesToRead);
+        pWav->memoryStream.currentReadPos += bytesToRead;
+    }
+
+    return bytesToRead;
+}
+
+DRWAV_PRIVATE drwav_bool32 drwav__on_seek_memory(void* pUserData, int offset, drwav_seek_origin origin)
+{
+    drwav* pWav = (drwav*)pUserData;
+    DRWAV_ASSERT(pWav != NULL);
+
+    if (origin == drwav_seek_origin_current) {
+        if (offset > 0) {
+            if (pWav->memoryStream.currentReadPos + offset > pWav->memoryStream.dataSize) {
+                return DRWAV_FALSE; /* Trying to seek too far forward. */
+            }
+        } else {
+            if (pWav->memoryStream.currentReadPos < (size_t)-offset) {
+                return DRWAV_FALSE; /* Trying to seek too far backwards. */
+            }
+        }
+
+        /* This will never underflow thanks to the clamps above. */
+        pWav->memoryStream.currentReadPos += offset;
+    } else {
+        if ((drwav_uint32)offset <= pWav->memoryStream.dataSize) {
+            pWav->memoryStream.currentReadPos = offset;
+        } else {
+            return DRWAV_FALSE; /* Trying to seek too far forward. */
+        }
+    }
+
+    return DRWAV_TRUE;
+}
+
+DRWAV_PRIVATE size_t drwav__on_write_memory(void* pUserData, const void* pDataIn, size_t bytesToWrite)
+{
+    drwav* pWav = (drwav*)pUserData;
+    size_t bytesRemaining;
+
+    DRWAV_ASSERT(pWav != NULL);
+    DRWAV_ASSERT(pWav->memoryStreamWrite.dataCapacity >= pWav->memoryStreamWrite.currentWritePos);
+
+    bytesRemaining = pWav->memoryStreamWrite.dataCapacity - pWav->memoryStreamWrite.currentWritePos;
+    if (bytesRemaining < bytesToWrite) {
+        /* Need to reallocate. */
+        void* pNewData;
+        size_t newDataCapacity = (pWav->memoryStreamWrite.dataCapacity == 0) ? 256 : pWav->memoryStreamWrite.dataCapacity * 2;
+
+        /* If doubling wasn't enough, just make it the minimum required size to write the data. */
+        if ((newDataCapacity - pWav->memoryStreamWrite.currentWritePos) < bytesToWrite) {
+            newDataCapacity = pWav->memoryStreamWrite.currentWritePos + bytesToWrite;
+        }
+
+        pNewData = drwav__realloc_from_callbacks(*pWav->memoryStreamWrite.ppData, newDataCapacity, pWav->memoryStreamWrite.dataCapacity, &pWav->allocationCallbacks);
+        if (pNewData == NULL) {
+            return 0;
+        }
+
+        *pWav->memoryStreamWrite.ppData = pNewData;
+        pWav->memoryStreamWrite.dataCapacity = newDataCapacity;
+    }
+
+    DRWAV_COPY_MEMORY(((drwav_uint8*)(*pWav->memoryStreamWrite.ppData)) + pWav->memoryStreamWrite.currentWritePos, pDataIn, bytesToWrite);
+
+    pWav->memoryStreamWrite.currentWritePos += bytesToWrite;
+    if (pWav->memoryStreamWrite.dataSize < pWav->memoryStreamWrite.currentWritePos) {
+        pWav->memoryStreamWrite.dataSize = pWav->memoryStreamWrite.currentWritePos;
+    }
+
+    *pWav->memoryStreamWrite.pDataSize = pWav->memoryStreamWrite.dataSize;
+
+    return bytesToWrite;
+}
+
+DRWAV_PRIVATE drwav_bool32 drwav__on_seek_memory_write(void* pUserData, int offset, drwav_seek_origin origin)
+{
+    drwav* pWav = (drwav*)pUserData;
+    DRWAV_ASSERT(pWav != NULL);
+
+    if (origin == drwav_seek_origin_current) {
+        if (offset > 0) {
+            if (pWav->memoryStreamWrite.currentWritePos + offset > pWav->memoryStreamWrite.dataSize) {
+                offset = (int)(pWav->memoryStreamWrite.dataSize - pWav->memoryStreamWrite.currentWritePos);  /* Trying to seek too far forward. */
+            }
+        } else {
+            if (pWav->memoryStreamWrite.currentWritePos < (size_t)-offset) {
+                offset = -(int)pWav->memoryStreamWrite.currentWritePos;  /* Trying to seek too far backwards. */
+            }
+        }
+
+        /* This will never underflow thanks to the clamps above. */
+        pWav->memoryStreamWrite.currentWritePos += offset;
+    } else {
+        if ((drwav_uint32)offset <= pWav->memoryStreamWrite.dataSize) {
+            pWav->memoryStreamWrite.currentWritePos = offset;
+        } else {
+            pWav->memoryStreamWrite.currentWritePos = pWav->memoryStreamWrite.dataSize;  /* Trying to seek too far forward. */
+        }
+    }
+
+    return DRWAV_TRUE;
+}
+
+DRWAV_API drwav_bool32 drwav_init_memory(drwav* pWav, const void* data, size_t dataSize, const drwav_allocation_callbacks* pAllocationCallbacks)
+{
+    return drwav_init_memory_ex(pWav, data, dataSize, NULL, NULL, 0, pAllocationCallbacks);
+}
+
+DRWAV_API drwav_bool32 drwav_init_memory_ex(drwav* pWav, const void* data, size_t dataSize, drwav_chunk_proc onChunk, void* pChunkUserData, drwav_uint32 flags, const drwav_allocation_callbacks* pAllocationCallbacks)
+{
+    if (data == NULL || dataSize == 0) {
+        return DRWAV_FALSE;
+    }
+
+    if (!drwav_preinit(pWav, drwav__on_read_memory, drwav__on_seek_memory, pWav, pAllocationCallbacks)) {
+        return DRWAV_FALSE;
+    }
+
+    pWav->memoryStream.data = (const drwav_uint8*)data;
+    pWav->memoryStream.dataSize = dataSize;
+    pWav->memoryStream.currentReadPos = 0;
+
+    return drwav_init__internal(pWav, onChunk, pChunkUserData, flags);
+}
+
+DRWAV_API drwav_bool32 drwav_init_memory_with_metadata(drwav* pWav, const void* data, size_t dataSize, drwav_uint32 flags, const drwav_allocation_callbacks* pAllocationCallbacks)
+{
+    if (data == NULL || dataSize == 0) {
+        return DRWAV_FALSE;
+    }
+
+    if (!drwav_preinit(pWav, drwav__on_read_memory, drwav__on_seek_memory, pWav, pAllocationCallbacks)) {
+        return DRWAV_FALSE;
+    }
+
+    pWav->memoryStream.data = (const drwav_uint8*)data;
+    pWav->memoryStream.dataSize = dataSize;
+    pWav->memoryStream.currentReadPos = 0;
+
+    return drwav_init__internal(pWav, NULL, NULL, flags | DRWAV_WITH_METADATA);
+}
+
+
+DRWAV_PRIVATE drwav_bool32 drwav_init_memory_write__internal(drwav* pWav, void** ppData, size_t* pDataSize, const drwav_data_format* pFormat, drwav_uint64 totalSampleCount, drwav_bool32 isSequential, const drwav_allocation_callbacks* pAllocationCallbacks)
+{
+    if (ppData == NULL || pDataSize == NULL) {
+        return DRWAV_FALSE;
+    }
+
+    *ppData = NULL; /* Important because we're using realloc()! */
+    *pDataSize = 0;
+
+    if (!drwav_preinit_write(pWav, pFormat, isSequential, drwav__on_write_memory, drwav__on_seek_memory_write, pWav, pAllocationCallbacks)) {
+        return DRWAV_FALSE;
+    }
+
+    pWav->memoryStreamWrite.ppData = ppData;
+    pWav->memoryStreamWrite.pDataSize = pDataSize;
+    pWav->memoryStreamWrite.dataSize = 0;
+    pWav->memoryStreamWrite.dataCapacity = 0;
+    pWav->memoryStreamWrite.currentWritePos = 0;
+
+    return drwav_init_write__internal(pWav, pFormat, totalSampleCount);
+}
+
+DRWAV_API drwav_bool32 drwav_init_memory_write(drwav* pWav, void** ppData, size_t* pDataSize, const drwav_data_format* pFormat, const drwav_allocation_callbacks* pAllocationCallbacks)
+{
+    return drwav_init_memory_write__internal(pWav, ppData, pDataSize, pFormat, 0, DRWAV_FALSE, pAllocationCallbacks);
+}
+
+DRWAV_API drwav_bool32 drwav_init_memory_write_sequential(drwav* pWav, void** ppData, size_t* pDataSize, const drwav_data_format* pFormat, drwav_uint64 totalSampleCount, const drwav_allocation_callbacks* pAllocationCallbacks)
+{
+    return drwav_init_memory_write__internal(pWav, ppData, pDataSize, pFormat, totalSampleCount, DRWAV_TRUE, pAllocationCallbacks);
+}
+
+DRWAV_API drwav_bool32 drwav_init_memory_write_sequential_pcm_frames(drwav* pWav, void** ppData, size_t* pDataSize, const drwav_data_format* pFormat, drwav_uint64 totalPCMFrameCount, const drwav_allocation_callbacks* pAllocationCallbacks)
+{
+    if (pFormat == NULL) {
+        return DRWAV_FALSE;
+    }
+
+    return drwav_init_memory_write_sequential(pWav, ppData, pDataSize, pFormat, totalPCMFrameCount*pFormat->channels, pAllocationCallbacks);
+}
+
+
+
+DRWAV_API drwav_result drwav_uninit(drwav* pWav)
+{
+    drwav_result result = DRWAV_SUCCESS;
+
+    if (pWav == NULL) {
+        return DRWAV_INVALID_ARGS;
+    }
+
+    /*
+    If the drwav object was opened in write mode we'll need to finalize a few things:
+      - Make sure the "data" chunk is aligned to 16-bits for RIFF containers, or 64 bits for W64 containers.
+      - Set the size of the "data" chunk.
+    */
+    if (pWav->onWrite != NULL) {
+        drwav_uint32 paddingSize = 0;
+
+        /* Padding. Do not adjust pWav->dataChunkDataSize - this should not include the padding. */
+        if (pWav->container == drwav_container_riff || pWav->container == drwav_container_rf64) {
+            paddingSize = drwav__chunk_padding_size_riff(pWav->dataChunkDataSize);
+        } else {
+            paddingSize = drwav__chunk_padding_size_w64(pWav->dataChunkDataSize);
+        }
+
+        if (paddingSize > 0) {
+            drwav_uint64 paddingData = 0;
+            drwav__write(pWav, &paddingData, paddingSize);  /* Byte order does not matter for this. */
+        }
+
+        /*
+        Chunk sizes. When using sequential mode, these will have been filled in at initialization time. We only need
+        to do this when using non-sequential mode.
+        */
+        if (pWav->onSeek && !pWav->isSequentialWrite) {
+            if (pWav->container == drwav_container_riff) {
+                /* The "RIFF" chunk size. */
+                if (pWav->onSeek(pWav->pUserData, 4, drwav_seek_origin_start)) {
+                    drwav_uint32 riffChunkSize = drwav__riff_chunk_size_riff(pWav->dataChunkDataSize, pWav->pMetadata, pWav->metadataCount);
+                    drwav__write_u32ne_to_le(pWav, riffChunkSize);
+                }
+
+                /* The "data" chunk size. */
+                if (pWav->onSeek(pWav->pUserData, (int)pWav->dataChunkDataPos - 4, drwav_seek_origin_start)) {
+                    drwav_uint32 dataChunkSize = drwav__data_chunk_size_riff(pWav->dataChunkDataSize);
+                    drwav__write_u32ne_to_le(pWav, dataChunkSize);
+                }
+            } else if (pWav->container == drwav_container_w64) {
+                /* The "RIFF" chunk size. */
+                if (pWav->onSeek(pWav->pUserData, 16, drwav_seek_origin_start)) {
+                    drwav_uint64 riffChunkSize = drwav__riff_chunk_size_w64(pWav->dataChunkDataSize);
+                    drwav__write_u64ne_to_le(pWav, riffChunkSize);
+                }
+
+                /* The "data" chunk size. */
+                if (pWav->onSeek(pWav->pUserData, (int)pWav->dataChunkDataPos - 8, drwav_seek_origin_start)) {
+                    drwav_uint64 dataChunkSize = drwav__data_chunk_size_w64(pWav->dataChunkDataSize);
+                    drwav__write_u64ne_to_le(pWav, dataChunkSize);
+                }
+            } else if (pWav->container == drwav_container_rf64) {
+                /* We only need to update the ds64 chunk. The "RIFF" and "data" chunks always have their sizes set to 0xFFFFFFFF for RF64. */
+                int ds64BodyPos = 12 + 8;
+
+                /* The "RIFF" chunk size. */
+                if (pWav->onSeek(pWav->pUserData, ds64BodyPos + 0, drwav_seek_origin_start)) {
+                    drwav_uint64 riffChunkSize = drwav__riff_chunk_size_rf64(pWav->dataChunkDataSize, pWav->pMetadata, pWav->metadataCount);
+                    drwav__write_u64ne_to_le(pWav, riffChunkSize);
+                }
+
+                /* The "data" chunk size. */
+                if (pWav->onSeek(pWav->pUserData, ds64BodyPos + 8, drwav_seek_origin_start)) {
+                    drwav_uint64 dataChunkSize = drwav__data_chunk_size_rf64(pWav->dataChunkDataSize);
+                    drwav__write_u64ne_to_le(pWav, dataChunkSize);
+                }
+            }
+        }
+
+        /* Validation for sequential mode. */
+        if (pWav->isSequentialWrite) {
+            if (pWav->dataChunkDataSize != pWav->dataChunkDataSizeTargetWrite) {
+                result = DRWAV_INVALID_FILE;
+            }
+        }
+    } else {
+        drwav_free(pWav->pMetadata, &pWav->allocationCallbacks);
+    }
+
+#ifndef DR_WAV_NO_STDIO
+    /*
+    If we opened the file with drwav_open_file() we will want to close the file handle. We can know whether or not drwav_open_file()
+    was used by looking at the onRead and onSeek callbacks.
+    */
+    if (pWav->onRead == drwav__on_read_stdio || pWav->onWrite == drwav__on_write_stdio) {
+        fclose((FILE*)pWav->pUserData);
+    }
+#endif
+
+    return result;
+}
+
+
+
+DRWAV_API size_t drwav_read_raw(drwav* pWav, size_t bytesToRead, void* pBufferOut)
+{
+    size_t bytesRead;
+    drwav_uint32 bytesPerFrame;
+
+    if (pWav == NULL || bytesToRead == 0) {
+        return 0;   /* Invalid args. */
+    }
+
+    if (bytesToRead > pWav->bytesRemaining) {
+        bytesToRead = (size_t)pWav->bytesRemaining;
+    }
+
+    if (bytesToRead == 0) {
+        return 0;   /* At end. */
+    }
+
+    bytesPerFrame = drwav_get_bytes_per_pcm_frame(pWav);
+    if (bytesPerFrame == 0) {
+        return 0;   /* Could not determine the bytes per frame. */
+    }
+
+    if (pBufferOut != NULL) {
+        bytesRead = pWav->onRead(pWav->pUserData, pBufferOut, bytesToRead);
+    } else {
+        /* We need to seek. If we fail, we need to read-and-discard to make sure we get a good byte count. */
+        bytesRead = 0;
+        while (bytesRead < bytesToRead) {
+            size_t bytesToSeek = (bytesToRead - bytesRead);
+            if (bytesToSeek > 0x7FFFFFFF) {
+                bytesToSeek = 0x7FFFFFFF;
+            }
+
+            if (pWav->onSeek(pWav->pUserData, (int)bytesToSeek, drwav_seek_origin_current) == DRWAV_FALSE) {
+                break;
+            }
+
+            bytesRead += bytesToSeek;
+        }
+
+        /* When we get here we may need to read-and-discard some data. */
+        while (bytesRead < bytesToRead) {
+            drwav_uint8 buffer[4096];
+            size_t bytesSeeked;
+            size_t bytesToSeek = (bytesToRead - bytesRead);
+            if (bytesToSeek > sizeof(buffer)) {
+                bytesToSeek = sizeof(buffer);
+            }
+
+            bytesSeeked = pWav->onRead(pWav->pUserData, buffer, bytesToSeek);
+            bytesRead += bytesSeeked;
+
+            if (bytesSeeked < bytesToSeek) {
+                break;  /* Reached the end. */
+            }
+        }
+    }
+
+    pWav->readCursorInPCMFrames += bytesRead / bytesPerFrame;
+
+    pWav->bytesRemaining -= bytesRead;
+    return bytesRead;
+}
+
+
+
+DRWAV_API drwav_uint64 drwav_read_pcm_frames_le(drwav* pWav, drwav_uint64 framesToRead, void* pBufferOut)
+{
+    drwav_uint32 bytesPerFrame;
+    drwav_uint64 bytesToRead;   /* Intentionally uint64 instead of size_t so we can do a check that we're not reading too much on 32-bit builds. */
+    drwav_uint64 framesRemainingInFile;
+
+    if (pWav == NULL || framesToRead == 0) {
+        return 0;
+    }
+
+    /* Cannot use this function for compressed formats. */
+    if (drwav__is_compressed_format_tag(pWav->translatedFormatTag)) {
+        return 0;
+    }
+
+    framesRemainingInFile = pWav->totalPCMFrameCount - pWav->readCursorInPCMFrames;
+    if (framesToRead > framesRemainingInFile) {
+        framesToRead = framesRemainingInFile;
+    }
+
+    bytesPerFrame = drwav_get_bytes_per_pcm_frame(pWav);
+    if (bytesPerFrame == 0) {
+        return 0;
+    }
+
+    /* Don't try to read more samples than can potentially fit in the output buffer. */
+    bytesToRead = framesToRead * bytesPerFrame;
+    if (bytesToRead > DRWAV_SIZE_MAX) {
+        bytesToRead = (DRWAV_SIZE_MAX / bytesPerFrame) * bytesPerFrame; /* Round the number of bytes to read to a clean frame boundary. */
+    }
+
+    /*
+    Doing an explicit check here just to make it clear that we don't want to be attempt to read anything if there's no bytes to read. There
+    *could* be a time where it evaluates to 0 due to overflowing.
+    */
+    if (bytesToRead == 0) {
+        return 0;
+    }
+
+    return drwav_read_raw(pWav, (size_t)bytesToRead, pBufferOut) / bytesPerFrame;
+}
+
+DRWAV_API drwav_uint64 drwav_read_pcm_frames_be(drwav* pWav, drwav_uint64 framesToRead, void* pBufferOut)
+{
+    drwav_uint64 framesRead = drwav_read_pcm_frames_le(pWav, framesToRead, pBufferOut);
+
+    if (pBufferOut != NULL) {
+        drwav_uint32 bytesPerFrame = drwav_get_bytes_per_pcm_frame(pWav);
+        if (bytesPerFrame == 0) {
+            return 0;   /* Could not get the bytes per frame which means bytes per sample cannot be determined and we don't know how to byte swap. */
+        }
+
+        drwav__bswap_samples(pBufferOut, framesRead*pWav->channels, bytesPerFrame/pWav->channels);
+    }
+
+    return framesRead;
+}
+
+DRWAV_API drwav_uint64 drwav_read_pcm_frames(drwav* pWav, drwav_uint64 framesToRead, void* pBufferOut)
+{
+    drwav_uint64 framesRead = 0;
+
+    if (drwav_is_container_be(pWav->container)) {
+        /*
+        Special case for AIFF. AIFF is a big-endian encoded format, but it supports a format that is
+        PCM in little-endian encoding. In this case, we fall through this branch and treate it as
+        little-endian.
+        */
+        if (pWav->container != drwav_container_aiff || pWav->aiff.isLE == DRWAV_FALSE) {
+            if (drwav__is_little_endian()) {
+                framesRead = drwav_read_pcm_frames_be(pWav, framesToRead, pBufferOut);
+            } else {
+                framesRead = drwav_read_pcm_frames_le(pWav, framesToRead, pBufferOut);
+            }
+
+            goto post_process;
+        }
+    }
+
+    /* Getting here means the data should be considered little-endian. */
+    if (drwav__is_little_endian()) {
+        framesRead = drwav_read_pcm_frames_le(pWav, framesToRead, pBufferOut);
+    } else {
+        framesRead = drwav_read_pcm_frames_be(pWav, framesToRead, pBufferOut);
+    }
+
+    /*
+    Here is where we check if we need to do a signed/unsigned conversion for AIFF. The reason we need to do this
+    is because dr_wav always assumes an 8-bit sample is unsigned, whereas AIFF can have signed 8-bit formats.
+    */
+    post_process:
+    {
+        if (pWav->container == drwav_container_aiff && pWav->bitsPerSample == 8 && pWav->aiff.isUnsigned == DRWAV_FALSE) {
+            if (pBufferOut != NULL) {
+                drwav_uint64 iSample;
+
+                for (iSample = 0; iSample < framesRead * pWav->channels; iSample += 1) {
+                    ((drwav_uint8*)pBufferOut)[iSample] += 128;
+                }
+            }
+        }
+    }
+
+    return framesRead;
+}
+
+
+
+DRWAV_PRIVATE drwav_bool32 drwav_seek_to_first_pcm_frame(drwav* pWav)
+{
+    if (pWav->onWrite != NULL) {
+        return DRWAV_FALSE; /* No seeking in write mode. */
+    }
+
+    if (!pWav->onSeek(pWav->pUserData, (int)pWav->dataChunkDataPos, drwav_seek_origin_start)) {
+        return DRWAV_FALSE;
+    }
+
+    if (drwav__is_compressed_format_tag(pWav->translatedFormatTag)) {
+        /* Cached data needs to be cleared for compressed formats. */
+        if (pWav->translatedFormatTag == DR_WAVE_FORMAT_ADPCM) {
+            DRWAV_ZERO_OBJECT(&pWav->msadpcm);
+        } else if (pWav->translatedFormatTag == DR_WAVE_FORMAT_DVI_ADPCM) {
+            DRWAV_ZERO_OBJECT(&pWav->ima);
+        } else {
+            DRWAV_ASSERT(DRWAV_FALSE);  /* If this assertion is triggered it means I've implemented a new compressed format but forgot to add a branch for it here. */
+        }
+    }
+
+    pWav->readCursorInPCMFrames = 0;
+    pWav->bytesRemaining = pWav->dataChunkDataSize;
+
+    return DRWAV_TRUE;
+}
+
+DRWAV_API drwav_bool32 drwav_seek_to_pcm_frame(drwav* pWav, drwav_uint64 targetFrameIndex)
+{
+    /* Seeking should be compatible with wave files > 2GB. */
+
+    if (pWav == NULL || pWav->onSeek == NULL) {
+        return DRWAV_FALSE;
+    }
+
+    /* No seeking in write mode. */
+    if (pWav->onWrite != NULL) {
+        return DRWAV_FALSE;
+    }
+
+    /* If there are no samples, just return DRWAV_TRUE without doing anything. */
+    if (pWav->totalPCMFrameCount == 0) {
+        return DRWAV_TRUE;
+    }
+
+    /* Make sure the sample is clamped. */
+    if (targetFrameIndex > pWav->totalPCMFrameCount) {
+        targetFrameIndex = pWav->totalPCMFrameCount;
+    }
+
+    /*
+    For compressed formats we just use a slow generic seek. If we are seeking forward we just seek forward. If we are going backwards we need
+    to seek back to the start.
+    */
+    if (drwav__is_compressed_format_tag(pWav->translatedFormatTag)) {
+        /* TODO: This can be optimized. */
+
+        /*
+        If we're seeking forward it's simple - just keep reading samples until we hit the sample we're requesting. If we're seeking backwards,
+        we first need to seek back to the start and then just do the same thing as a forward seek.
+        */
+        if (targetFrameIndex < pWav->readCursorInPCMFrames) {
+            if (!drwav_seek_to_first_pcm_frame(pWav)) {
+                return DRWAV_FALSE;
+            }
+        }
+
+        if (targetFrameIndex > pWav->readCursorInPCMFrames) {
+            drwav_uint64 offsetInFrames = targetFrameIndex - pWav->readCursorInPCMFrames;
+
+            drwav_int16 devnull[2048];
+            while (offsetInFrames > 0) {
+                drwav_uint64 framesRead = 0;
+                drwav_uint64 framesToRead = offsetInFrames;
+                if (framesToRead > drwav_countof(devnull)/pWav->channels) {
+                    framesToRead = drwav_countof(devnull)/pWav->channels;
+                }
+
+                if (pWav->translatedFormatTag == DR_WAVE_FORMAT_ADPCM) {
+                    framesRead = drwav_read_pcm_frames_s16__msadpcm(pWav, framesToRead, devnull);
+                } else if (pWav->translatedFormatTag == DR_WAVE_FORMAT_DVI_ADPCM) {
+                    framesRead = drwav_read_pcm_frames_s16__ima(pWav, framesToRead, devnull);
+                } else {
+                    DRWAV_ASSERT(DRWAV_FALSE);  /* If this assertion is triggered it means I've implemented a new compressed format but forgot to add a branch for it here. */
+                }
+
+                if (framesRead != framesToRead) {
+                    return DRWAV_FALSE;
+                }
+
+                offsetInFrames -= framesRead;
+            }
+        }
+    } else {
+        drwav_uint64 totalSizeInBytes;
+        drwav_uint64 currentBytePos;
+        drwav_uint64 targetBytePos;
+        drwav_uint64 offset;
+        drwav_uint32 bytesPerFrame;
+
+        bytesPerFrame = drwav_get_bytes_per_pcm_frame(pWav);
+        if (bytesPerFrame == 0) {
+            return DRWAV_FALSE; /* Not able to calculate offset. */
+        }
+
+        totalSizeInBytes = pWav->totalPCMFrameCount * bytesPerFrame;
+        /*DRWAV_ASSERT(totalSizeInBytes >= pWav->bytesRemaining);*/
+
+        currentBytePos = totalSizeInBytes - pWav->bytesRemaining;
+        targetBytePos  = targetFrameIndex * bytesPerFrame;
+
+        if (currentBytePos < targetBytePos) {
+            /* Offset forwards. */
+            offset = (targetBytePos - currentBytePos);
+        } else {
+            /* Offset backwards. */
+            if (!drwav_seek_to_first_pcm_frame(pWav)) {
+                return DRWAV_FALSE;
+            }
+            offset = targetBytePos;
+        }
+
+        while (offset > 0) {
+            int offset32 = ((offset > INT_MAX) ? INT_MAX : (int)offset);
+            if (!pWav->onSeek(pWav->pUserData, offset32, drwav_seek_origin_current)) {
+                return DRWAV_FALSE;
+            }
+
+            pWav->readCursorInPCMFrames += offset32 / bytesPerFrame;
+            pWav->bytesRemaining        -= offset32;
+            offset                      -= offset32;
+        }
+    }
+
+    return DRWAV_TRUE;
+}
+
+DRWAV_API drwav_result drwav_get_cursor_in_pcm_frames(drwav* pWav, drwav_uint64* pCursor)
+{
+    if (pCursor == NULL) {
+        return DRWAV_INVALID_ARGS;
+    }
+
+    *pCursor = 0;   /* Safety. */
+
+    if (pWav == NULL) {
+        return DRWAV_INVALID_ARGS;
+    }
+
+    *pCursor = pWav->readCursorInPCMFrames;
+
+    return DRWAV_SUCCESS;
+}
+
+DRWAV_API drwav_result drwav_get_length_in_pcm_frames(drwav* pWav, drwav_uint64* pLength)
+{
+    if (pLength == NULL) {
+        return DRWAV_INVALID_ARGS;
+    }
+
+    *pLength = 0;   /* Safety. */
+
+    if (pWav == NULL) {
+        return DRWAV_INVALID_ARGS;
+    }
+
+    *pLength = pWav->totalPCMFrameCount;
+
+    return DRWAV_SUCCESS;
+}
+
+
+DRWAV_API size_t drwav_write_raw(drwav* pWav, size_t bytesToWrite, const void* pData)
+{
+    size_t bytesWritten;
+
+    if (pWav == NULL || bytesToWrite == 0 || pData == NULL) {
+        return 0;
+    }
+
+    bytesWritten = pWav->onWrite(pWav->pUserData, pData, bytesToWrite);
+    pWav->dataChunkDataSize += bytesWritten;
+
+    return bytesWritten;
+}
+
+DRWAV_API drwav_uint64 drwav_write_pcm_frames_le(drwav* pWav, drwav_uint64 framesToWrite, const void* pData)
+{
+    drwav_uint64 bytesToWrite;
+    drwav_uint64 bytesWritten;
+    const drwav_uint8* pRunningData;
+
+    if (pWav == NULL || framesToWrite == 0 || pData == NULL) {
+        return 0;
+    }
+
+    bytesToWrite = ((framesToWrite * pWav->channels * pWav->bitsPerSample) / 8);
+    if (bytesToWrite > DRWAV_SIZE_MAX) {
+        return 0;
+    }
+
+    bytesWritten = 0;
+    pRunningData = (const drwav_uint8*)pData;
+
+    while (bytesToWrite > 0) {
+        size_t bytesJustWritten;
+        drwav_uint64 bytesToWriteThisIteration;
+
+        bytesToWriteThisIteration = bytesToWrite;
+        DRWAV_ASSERT(bytesToWriteThisIteration <= DRWAV_SIZE_MAX);  /* <-- This is checked above. */
+
+        bytesJustWritten = drwav_write_raw(pWav, (size_t)bytesToWriteThisIteration, pRunningData);
+        if (bytesJustWritten == 0) {
+            break;
+        }
+
+        bytesToWrite -= bytesJustWritten;
+        bytesWritten += bytesJustWritten;
+        pRunningData += bytesJustWritten;
+    }
+
+    return (bytesWritten * 8) / pWav->bitsPerSample / pWav->channels;
+}
+
+DRWAV_API drwav_uint64 drwav_write_pcm_frames_be(drwav* pWav, drwav_uint64 framesToWrite, const void* pData)
+{
+    drwav_uint64 bytesToWrite;
+    drwav_uint64 bytesWritten;
+    drwav_uint32 bytesPerSample;
+    const drwav_uint8* pRunningData;
+
+    if (pWav == NULL || framesToWrite == 0 || pData == NULL) {
+        return 0;
+    }
+
+    bytesToWrite = ((framesToWrite * pWav->channels * pWav->bitsPerSample) / 8);
+    if (bytesToWrite > DRWAV_SIZE_MAX) {
+        return 0;
+    }
+
+    bytesWritten = 0;
+    pRunningData = (const drwav_uint8*)pData;
+
+    bytesPerSample = drwav_get_bytes_per_pcm_frame(pWav) / pWav->channels;
+    if (bytesPerSample == 0) {
+        return 0;   /* Cannot determine bytes per sample, or bytes per sample is less than one byte. */
+    }
+
+    while (bytesToWrite > 0) {
+        drwav_uint8 temp[4096];
+        drwav_uint32 sampleCount;
+        size_t bytesJustWritten;
+        drwav_uint64 bytesToWriteThisIteration;
+
+        bytesToWriteThisIteration = bytesToWrite;
+        DRWAV_ASSERT(bytesToWriteThisIteration <= DRWAV_SIZE_MAX);  /* <-- This is checked above. */
+
+        /*
+        WAV files are always little-endian. We need to byte swap on big-endian architectures. Since our input buffer is read-only we need
+        to use an intermediary buffer for the conversion.
+        */
+        sampleCount = sizeof(temp)/bytesPerSample;
+
+        if (bytesToWriteThisIteration > ((drwav_uint64)sampleCount)*bytesPerSample) {
+            bytesToWriteThisIteration = ((drwav_uint64)sampleCount)*bytesPerSample;
+        }
+
+        DRWAV_COPY_MEMORY(temp, pRunningData, (size_t)bytesToWriteThisIteration);
+        drwav__bswap_samples(temp, sampleCount, bytesPerSample);
+
+        bytesJustWritten = drwav_write_raw(pWav, (size_t)bytesToWriteThisIteration, temp);
+        if (bytesJustWritten == 0) {
+            break;
+        }
+
+        bytesToWrite -= bytesJustWritten;
+        bytesWritten += bytesJustWritten;
+        pRunningData += bytesJustWritten;
+    }
+
+    return (bytesWritten * 8) / pWav->bitsPerSample / pWav->channels;
+}
+
+DRWAV_API drwav_uint64 drwav_write_pcm_frames(drwav* pWav, drwav_uint64 framesToWrite, const void* pData)
+{
+    if (drwav__is_little_endian()) {
+        return drwav_write_pcm_frames_le(pWav, framesToWrite, pData);
+    } else {
+        return drwav_write_pcm_frames_be(pWav, framesToWrite, pData);
+    }
+}
+
+
+DRWAV_PRIVATE drwav_uint64 drwav_read_pcm_frames_s16__msadpcm(drwav* pWav, drwav_uint64 framesToRead, drwav_int16* pBufferOut)
+{
+    drwav_uint64 totalFramesRead = 0;
+
+    DRWAV_ASSERT(pWav != NULL);
+    DRWAV_ASSERT(framesToRead > 0);
+
+    /* TODO: Lots of room for optimization here. */
+
+    while (pWav->readCursorInPCMFrames < pWav->totalPCMFrameCount) {
+        DRWAV_ASSERT(framesToRead > 0); /* This loop iteration will never get hit with framesToRead == 0 because it's asserted at the top, and we check for 0 inside the loop just below. */
+
+        /* If there are no cached frames we need to load a new block. */
+        if (pWav->msadpcm.cachedFrameCount == 0 && pWav->msadpcm.bytesRemainingInBlock == 0) {
+            if (pWav->channels == 1) {
+                /* Mono. */
+                drwav_uint8 header[7];
+                if (pWav->onRead(pWav->pUserData, header, sizeof(header)) != sizeof(header)) {
+                    return totalFramesRead;
+                }
+                pWav->msadpcm.bytesRemainingInBlock = pWav->fmt.blockAlign - sizeof(header);
+
+                pWav->msadpcm.predictor[0]     = header[0];
+                pWav->msadpcm.delta[0]         = drwav_bytes_to_s16(header + 1);
+                pWav->msadpcm.prevFrames[0][1] = (drwav_int32)drwav_bytes_to_s16(header + 3);
+                pWav->msadpcm.prevFrames[0][0] = (drwav_int32)drwav_bytes_to_s16(header + 5);
+                pWav->msadpcm.cachedFrames[2]  = pWav->msadpcm.prevFrames[0][0];
+                pWav->msadpcm.cachedFrames[3]  = pWav->msadpcm.prevFrames[0][1];
+                pWav->msadpcm.cachedFrameCount = 2;
+            } else {
+                /* Stereo. */
+                drwav_uint8 header[14];
+                if (pWav->onRead(pWav->pUserData, header, sizeof(header)) != sizeof(header)) {
+                    return totalFramesRead;
+                }
+                pWav->msadpcm.bytesRemainingInBlock = pWav->fmt.blockAlign - sizeof(header);
+
+                pWav->msadpcm.predictor[0] = header[0];
+                pWav->msadpcm.predictor[1] = header[1];
+                pWav->msadpcm.delta[0] = drwav_bytes_to_s16(header + 2);
+                pWav->msadpcm.delta[1] = drwav_bytes_to_s16(header + 4);
+                pWav->msadpcm.prevFrames[0][1] = (drwav_int32)drwav_bytes_to_s16(header + 6);
+                pWav->msadpcm.prevFrames[1][1] = (drwav_int32)drwav_bytes_to_s16(header + 8);
+                pWav->msadpcm.prevFrames[0][0] = (drwav_int32)drwav_bytes_to_s16(header + 10);
+                pWav->msadpcm.prevFrames[1][0] = (drwav_int32)drwav_bytes_to_s16(header + 12);
+
+                pWav->msadpcm.cachedFrames[0] = pWav->msadpcm.prevFrames[0][0];
+                pWav->msadpcm.cachedFrames[1] = pWav->msadpcm.prevFrames[1][0];
+                pWav->msadpcm.cachedFrames[2] = pWav->msadpcm.prevFrames[0][1];
+                pWav->msadpcm.cachedFrames[3] = pWav->msadpcm.prevFrames[1][1];
+                pWav->msadpcm.cachedFrameCount = 2;
+            }
+        }
+
+        /* Output anything that's cached. */
+        while (framesToRead > 0 && pWav->msadpcm.cachedFrameCount > 0 && pWav->readCursorInPCMFrames < pWav->totalPCMFrameCount) {
+            if (pBufferOut != NULL) {
+                drwav_uint32 iSample = 0;
+                for (iSample = 0; iSample < pWav->channels; iSample += 1) {
+                    pBufferOut[iSample] = (drwav_int16)pWav->msadpcm.cachedFrames[(drwav_countof(pWav->msadpcm.cachedFrames) - (pWav->msadpcm.cachedFrameCount*pWav->channels)) + iSample];
+                }
+
+                pBufferOut += pWav->channels;
+            }
+
+            framesToRead    -= 1;
+            totalFramesRead += 1;
+            pWav->readCursorInPCMFrames += 1;
+            pWav->msadpcm.cachedFrameCount -= 1;
+        }
+
+        if (framesToRead == 0) {
+            break;
+        }
+
+
+        /*
+        If there's nothing left in the cache, just go ahead and load more. If there's nothing left to load in the current block we just continue to the next
+        loop iteration which will trigger the loading of a new block.
+        */
+        if (pWav->msadpcm.cachedFrameCount == 0) {
+            if (pWav->msadpcm.bytesRemainingInBlock == 0) {
+                continue;
+            } else {
+                static drwav_int32 adaptationTable[] = {
+                    230, 230, 230, 230, 307, 409, 512, 614,
+                    768, 614, 512, 409, 307, 230, 230, 230
+                };
+                static drwav_int32 coeff1Table[] = { 256, 512, 0, 192, 240, 460,  392 };
+                static drwav_int32 coeff2Table[] = { 0,  -256, 0, 64,  0,  -208, -232 };
+
+                drwav_uint8 nibbles;
+                drwav_int32 nibble0;
+                drwav_int32 nibble1;
+
+                if (pWav->onRead(pWav->pUserData, &nibbles, 1) != 1) {
+                    return totalFramesRead;
+                }
+                pWav->msadpcm.bytesRemainingInBlock -= 1;
+
+                /* TODO: Optimize away these if statements. */
+                nibble0 = ((nibbles & 0xF0) >> 4); if ((nibbles & 0x80)) { nibble0 |= 0xFFFFFFF0UL; }
+                nibble1 = ((nibbles & 0x0F) >> 0); if ((nibbles & 0x08)) { nibble1 |= 0xFFFFFFF0UL; }
+
+                if (pWav->channels == 1) {
+                    /* Mono. */
+                    drwav_int32 newSample0;
+                    drwav_int32 newSample1;
+
+                    newSample0  = ((pWav->msadpcm.prevFrames[0][1] * coeff1Table[pWav->msadpcm.predictor[0]]) + (pWav->msadpcm.prevFrames[0][0] * coeff2Table[pWav->msadpcm.predictor[0]])) >> 8;
+                    newSample0 += nibble0 * pWav->msadpcm.delta[0];
+                    newSample0  = drwav_clamp(newSample0, -32768, 32767);
+
+                    pWav->msadpcm.delta[0] = (adaptationTable[((nibbles & 0xF0) >> 4)] * pWav->msadpcm.delta[0]) >> 8;
+                    if (pWav->msadpcm.delta[0] < 16) {
+                        pWav->msadpcm.delta[0] = 16;
+                    }
+
+                    pWav->msadpcm.prevFrames[0][0] = pWav->msadpcm.prevFrames[0][1];
+                    pWav->msadpcm.prevFrames[0][1] = newSample0;
+
+
+                    newSample1  = ((pWav->msadpcm.prevFrames[0][1] * coeff1Table[pWav->msadpcm.predictor[0]]) + (pWav->msadpcm.prevFrames[0][0] * coeff2Table[pWav->msadpcm.predictor[0]])) >> 8;
+                    newSample1 += nibble1 * pWav->msadpcm.delta[0];
+                    newSample1  = drwav_clamp(newSample1, -32768, 32767);
+
+                    pWav->msadpcm.delta[0] = (adaptationTable[((nibbles & 0x0F) >> 0)] * pWav->msadpcm.delta[0]) >> 8;
+                    if (pWav->msadpcm.delta[0] < 16) {
+                        pWav->msadpcm.delta[0] = 16;
+                    }
+
+                    pWav->msadpcm.prevFrames[0][0] = pWav->msadpcm.prevFrames[0][1];
+                    pWav->msadpcm.prevFrames[0][1] = newSample1;
+
+
+                    pWav->msadpcm.cachedFrames[2] = newSample0;
+                    pWav->msadpcm.cachedFrames[3] = newSample1;
+                    pWav->msadpcm.cachedFrameCount = 2;
+                } else {
+                    /* Stereo. */
+                    drwav_int32 newSample0;
+                    drwav_int32 newSample1;
+
+                    /* Left. */
+                    newSample0  = ((pWav->msadpcm.prevFrames[0][1] * coeff1Table[pWav->msadpcm.predictor[0]]) + (pWav->msadpcm.prevFrames[0][0] * coeff2Table[pWav->msadpcm.predictor[0]])) >> 8;
+                    newSample0 += nibble0 * pWav->msadpcm.delta[0];
+                    newSample0  = drwav_clamp(newSample0, -32768, 32767);
+
+                    pWav->msadpcm.delta[0] = (adaptationTable[((nibbles & 0xF0) >> 4)] * pWav->msadpcm.delta[0]) >> 8;
+                    if (pWav->msadpcm.delta[0] < 16) {
+                        pWav->msadpcm.delta[0] = 16;
+                    }
+
+                    pWav->msadpcm.prevFrames[0][0] = pWav->msadpcm.prevFrames[0][1];
+                    pWav->msadpcm.prevFrames[0][1] = newSample0;
+
+
+                    /* Right. */
+                    newSample1  = ((pWav->msadpcm.prevFrames[1][1] * coeff1Table[pWav->msadpcm.predictor[1]]) + (pWav->msadpcm.prevFrames[1][0] * coeff2Table[pWav->msadpcm.predictor[1]])) >> 8;
+                    newSample1 += nibble1 * pWav->msadpcm.delta[1];
+                    newSample1  = drwav_clamp(newSample1, -32768, 32767);
+
+                    pWav->msadpcm.delta[1] = (adaptationTable[((nibbles & 0x0F) >> 0)] * pWav->msadpcm.delta[1]) >> 8;
+                    if (pWav->msadpcm.delta[1] < 16) {
+                        pWav->msadpcm.delta[1] = 16;
+                    }
+
+                    pWav->msadpcm.prevFrames[1][0] = pWav->msadpcm.prevFrames[1][1];
+                    pWav->msadpcm.prevFrames[1][1] = newSample1;
+
+                    pWav->msadpcm.cachedFrames[2] = newSample0;
+                    pWav->msadpcm.cachedFrames[3] = newSample1;
+                    pWav->msadpcm.cachedFrameCount = 1;
+                }
+            }
+        }
+    }
+
+    return totalFramesRead;
+}
+
+
+DRWAV_PRIVATE drwav_uint64 drwav_read_pcm_frames_s16__ima(drwav* pWav, drwav_uint64 framesToRead, drwav_int16* pBufferOut)
+{
+    drwav_uint64 totalFramesRead = 0;
+    drwav_uint32 iChannel;
+
+    static drwav_int32 indexTable[16] = {
+        -1, -1, -1, -1, 2, 4, 6, 8,
+        -1, -1, -1, -1, 2, 4, 6, 8
+    };
+
+    static drwav_int32 stepTable[89] = {
+        7,     8,     9,     10,    11,    12,    13,    14,    16,    17,
+        19,    21,    23,    25,    28,    31,    34,    37,    41,    45,
+        50,    55,    60,    66,    73,    80,    88,    97,    107,   118,
+        130,   143,   157,   173,   190,   209,   230,   253,   279,   307,
+        337,   371,   408,   449,   494,   544,   598,   658,   724,   796,
+        876,   963,   1060,  1166,  1282,  1411,  1552,  1707,  1878,  2066,
+        2272,  2499,  2749,  3024,  3327,  3660,  4026,  4428,  4871,  5358,
+        5894,  6484,  7132,  7845,  8630,  9493,  10442, 11487, 12635, 13899,
+        15289, 16818, 18500, 20350, 22385, 24623, 27086, 29794, 32767
+    };
+
+    DRWAV_ASSERT(pWav != NULL);
+    DRWAV_ASSERT(framesToRead > 0);
+
+    /* TODO: Lots of room for optimization here. */
+
+    while (pWav->readCursorInPCMFrames < pWav->totalPCMFrameCount) {
+        DRWAV_ASSERT(framesToRead > 0); /* This loop iteration will never get hit with framesToRead == 0 because it's asserted at the top, and we check for 0 inside the loop just below. */
+
+        /* If there are no cached samples we need to load a new block. */
+        if (pWav->ima.cachedFrameCount == 0 && pWav->ima.bytesRemainingInBlock == 0) {
+            if (pWav->channels == 1) {
+                /* Mono. */
+                drwav_uint8 header[4];
+                if (pWav->onRead(pWav->pUserData, header, sizeof(header)) != sizeof(header)) {
+                    return totalFramesRead;
+                }
+                pWav->ima.bytesRemainingInBlock = pWav->fmt.blockAlign - sizeof(header);
+
+                if (header[2] >= drwav_countof(stepTable)) {
+                    pWav->onSeek(pWav->pUserData, pWav->ima.bytesRemainingInBlock, drwav_seek_origin_current);
+                    pWav->ima.bytesRemainingInBlock = 0;
+                    return totalFramesRead; /* Invalid data. */
+                }
+
+                pWav->ima.predictor[0] = (drwav_int16)drwav_bytes_to_u16(header + 0);
+                pWav->ima.stepIndex[0] = drwav_clamp(header[2], 0, (drwav_int32)drwav_countof(stepTable)-1);    /* Clamp not necessary because we checked above, but adding here to silence a static analysis warning. */
+                pWav->ima.cachedFrames[drwav_countof(pWav->ima.cachedFrames) - 1] = pWav->ima.predictor[0];
+                pWav->ima.cachedFrameCount = 1;
+            } else {
+                /* Stereo. */
+                drwav_uint8 header[8];
+                if (pWav->onRead(pWav->pUserData, header, sizeof(header)) != sizeof(header)) {
+                    return totalFramesRead;
+                }
+                pWav->ima.bytesRemainingInBlock = pWav->fmt.blockAlign - sizeof(header);
+
+                if (header[2] >= drwav_countof(stepTable) || header[6] >= drwav_countof(stepTable)) {
+                    pWav->onSeek(pWav->pUserData, pWav->ima.bytesRemainingInBlock, drwav_seek_origin_current);
+                    pWav->ima.bytesRemainingInBlock = 0;
+                    return totalFramesRead; /* Invalid data. */
+                }
+
+                pWav->ima.predictor[0] = drwav_bytes_to_s16(header + 0);
+                pWav->ima.stepIndex[0] = drwav_clamp(header[2], 0, (drwav_int32)drwav_countof(stepTable)-1);    /* Clamp not necessary because we checked above, but adding here to silence a static analysis warning. */
+                pWav->ima.predictor[1] = drwav_bytes_to_s16(header + 4);
+                pWav->ima.stepIndex[1] = drwav_clamp(header[6], 0, (drwav_int32)drwav_countof(stepTable)-1);    /* Clamp not necessary because we checked above, but adding here to silence a static analysis warning. */
+
+                pWav->ima.cachedFrames[drwav_countof(pWav->ima.cachedFrames) - 2] = pWav->ima.predictor[0];
+                pWav->ima.cachedFrames[drwav_countof(pWav->ima.cachedFrames) - 1] = pWav->ima.predictor[1];
+                pWav->ima.cachedFrameCount = 1;
+            }
+        }
+
+        /* Output anything that's cached. */
+        while (framesToRead > 0 && pWav->ima.cachedFrameCount > 0 && pWav->readCursorInPCMFrames < pWav->totalPCMFrameCount) {
+            if (pBufferOut != NULL) {
+                drwav_uint32 iSample;
+                for (iSample = 0; iSample < pWav->channels; iSample += 1) {
+                    pBufferOut[iSample] = (drwav_int16)pWav->ima.cachedFrames[(drwav_countof(pWav->ima.cachedFrames) - (pWav->ima.cachedFrameCount*pWav->channels)) + iSample];
+                }
+                pBufferOut += pWav->channels;
+            }
+
+            framesToRead    -= 1;
+            totalFramesRead += 1;
+            pWav->readCursorInPCMFrames += 1;
+            pWav->ima.cachedFrameCount -= 1;
+        }
+
+        if (framesToRead == 0) {
+            break;
+        }
+
+        /*
+        If there's nothing left in the cache, just go ahead and load more. If there's nothing left to load in the current block we just continue to the next
+        loop iteration which will trigger the loading of a new block.
+        */
+        if (pWav->ima.cachedFrameCount == 0) {
+            if (pWav->ima.bytesRemainingInBlock == 0) {
+                continue;
+            } else {
+                /*
+                From what I can tell with stereo streams, it looks like every 4 bytes (8 samples) is for one channel. So it goes 4 bytes for the
+                left channel, 4 bytes for the right channel.
+                */
+                pWav->ima.cachedFrameCount = 8;
+                for (iChannel = 0; iChannel < pWav->channels; ++iChannel) {
+                    drwav_uint32 iByte;
+                    drwav_uint8 nibbles[4];
+                    if (pWav->onRead(pWav->pUserData, &nibbles, 4) != 4) {
+                        pWav->ima.cachedFrameCount = 0;
+                        return totalFramesRead;
+                    }
+                    pWav->ima.bytesRemainingInBlock -= 4;
+
+                    for (iByte = 0; iByte < 4; ++iByte) {
+                        drwav_uint8 nibble0 = ((nibbles[iByte] & 0x0F) >> 0);
+                        drwav_uint8 nibble1 = ((nibbles[iByte] & 0xF0) >> 4);
+
+                        drwav_int32 step      = stepTable[pWav->ima.stepIndex[iChannel]];
+                        drwav_int32 predictor = pWav->ima.predictor[iChannel];
+
+                        drwav_int32      diff  = step >> 3;
+                        if (nibble0 & 1) diff += step >> 2;
+                        if (nibble0 & 2) diff += step >> 1;
+                        if (nibble0 & 4) diff += step;
+                        if (nibble0 & 8) diff  = -diff;
+
+                        predictor = drwav_clamp(predictor + diff, -32768, 32767);
+                        pWav->ima.predictor[iChannel] = predictor;
+                        pWav->ima.stepIndex[iChannel] = drwav_clamp(pWav->ima.stepIndex[iChannel] + indexTable[nibble0], 0, (drwav_int32)drwav_countof(stepTable)-1);
+                        pWav->ima.cachedFrames[(drwav_countof(pWav->ima.cachedFrames) - (pWav->ima.cachedFrameCount*pWav->channels)) + (iByte*2+0)*pWav->channels + iChannel] = predictor;
+
+
+                        step      = stepTable[pWav->ima.stepIndex[iChannel]];
+                        predictor = pWav->ima.predictor[iChannel];
+
+                                         diff  = step >> 3;
+                        if (nibble1 & 1) diff += step >> 2;
+                        if (nibble1 & 2) diff += step >> 1;
+                        if (nibble1 & 4) diff += step;
+                        if (nibble1 & 8) diff  = -diff;
+
+                        predictor = drwav_clamp(predictor + diff, -32768, 32767);
+                        pWav->ima.predictor[iChannel] = predictor;
+                        pWav->ima.stepIndex[iChannel] = drwav_clamp(pWav->ima.stepIndex[iChannel] + indexTable[nibble1], 0, (drwav_int32)drwav_countof(stepTable)-1);
+                        pWav->ima.cachedFrames[(drwav_countof(pWav->ima.cachedFrames) - (pWav->ima.cachedFrameCount*pWav->channels)) + (iByte*2+1)*pWav->channels + iChannel] = predictor;
+                    }
+                }
+            }
+        }
+    }
+
+    return totalFramesRead;
+}
+
+
+#ifndef DR_WAV_NO_CONVERSION_API
+static unsigned short g_drwavAlawTable[256] = {
+    0xEA80, 0xEB80, 0xE880, 0xE980, 0xEE80, 0xEF80, 0xEC80, 0xED80, 0xE280, 0xE380, 0xE080, 0xE180, 0xE680, 0xE780, 0xE480, 0xE580,
+    0xF540, 0xF5C0, 0xF440, 0xF4C0, 0xF740, 0xF7C0, 0xF640, 0xF6C0, 0xF140, 0xF1C0, 0xF040, 0xF0C0, 0xF340, 0xF3C0, 0xF240, 0xF2C0,
+    0xAA00, 0xAE00, 0xA200, 0xA600, 0xBA00, 0xBE00, 0xB200, 0xB600, 0x8A00, 0x8E00, 0x8200, 0x8600, 0x9A00, 0x9E00, 0x9200, 0x9600,
+    0xD500, 0xD700, 0xD100, 0xD300, 0xDD00, 0xDF00, 0xD900, 0xDB00, 0xC500, 0xC700, 0xC100, 0xC300, 0xCD00, 0xCF00, 0xC900, 0xCB00,
+    0xFEA8, 0xFEB8, 0xFE88, 0xFE98, 0xFEE8, 0xFEF8, 0xFEC8, 0xFED8, 0xFE28, 0xFE38, 0xFE08, 0xFE18, 0xFE68, 0xFE78, 0xFE48, 0xFE58,
+    0xFFA8, 0xFFB8, 0xFF88, 0xFF98, 0xFFE8, 0xFFF8, 0xFFC8, 0xFFD8, 0xFF28, 0xFF38, 0xFF08, 0xFF18, 0xFF68, 0xFF78, 0xFF48, 0xFF58,
+    0xFAA0, 0xFAE0, 0xFA20, 0xFA60, 0xFBA0, 0xFBE0, 0xFB20, 0xFB60, 0xF8A0, 0xF8E0, 0xF820, 0xF860, 0xF9A0, 0xF9E0, 0xF920, 0xF960,
+    0xFD50, 0xFD70, 0xFD10, 0xFD30, 0xFDD0, 0xFDF0, 0xFD90, 0xFDB0, 0xFC50, 0xFC70, 0xFC10, 0xFC30, 0xFCD0, 0xFCF0, 0xFC90, 0xFCB0,
+    0x1580, 0x1480, 0x1780, 0x1680, 0x1180, 0x1080, 0x1380, 0x1280, 0x1D80, 0x1C80, 0x1F80, 0x1E80, 0x1980, 0x1880, 0x1B80, 0x1A80,
+    0x0AC0, 0x0A40, 0x0BC0, 0x0B40, 0x08C0, 0x0840, 0x09C0, 0x0940, 0x0EC0, 0x0E40, 0x0FC0, 0x0F40, 0x0CC0, 0x0C40, 0x0DC0, 0x0D40,
+    0x5600, 0x5200, 0x5E00, 0x5A00, 0x4600, 0x4200, 0x4E00, 0x4A00, 0x7600, 0x7200, 0x7E00, 0x7A00, 0x6600, 0x6200, 0x6E00, 0x6A00,
+    0x2B00, 0x2900, 0x2F00, 0x2D00, 0x2300, 0x2100, 0x2700, 0x2500, 0x3B00, 0x3900, 0x3F00, 0x3D00, 0x3300, 0x3100, 0x3700, 0x3500,
+    0x0158, 0x0148, 0x0178, 0x0168, 0x0118, 0x0108, 0x0138, 0x0128, 0x01D8, 0x01C8, 0x01F8, 0x01E8, 0x0198, 0x0188, 0x01B8, 0x01A8,
+    0x0058, 0x0048, 0x0078, 0x0068, 0x0018, 0x0008, 0x0038, 0x0028, 0x00D8, 0x00C8, 0x00F8, 0x00E8, 0x0098, 0x0088, 0x00B8, 0x00A8,
+    0x0560, 0x0520, 0x05E0, 0x05A0, 0x0460, 0x0420, 0x04E0, 0x04A0, 0x0760, 0x0720, 0x07E0, 0x07A0, 0x0660, 0x0620, 0x06E0, 0x06A0,
+    0x02B0, 0x0290, 0x02F0, 0x02D0, 0x0230, 0x0210, 0x0270, 0x0250, 0x03B0, 0x0390, 0x03F0, 0x03D0, 0x0330, 0x0310, 0x0370, 0x0350
+};
+
+static unsigned short g_drwavMulawTable[256] = {
+    0x8284, 0x8684, 0x8A84, 0x8E84, 0x9284, 0x9684, 0x9A84, 0x9E84, 0xA284, 0xA684, 0xAA84, 0xAE84, 0xB284, 0xB684, 0xBA84, 0xBE84,
+    0xC184, 0xC384, 0xC584, 0xC784, 0xC984, 0xCB84, 0xCD84, 0xCF84, 0xD184, 0xD384, 0xD584, 0xD784, 0xD984, 0xDB84, 0xDD84, 0xDF84,
+    0xE104, 0xE204, 0xE304, 0xE404, 0xE504, 0xE604, 0xE704, 0xE804, 0xE904, 0xEA04, 0xEB04, 0xEC04, 0xED04, 0xEE04, 0xEF04, 0xF004,
+    0xF0C4, 0xF144, 0xF1C4, 0xF244, 0xF2C4, 0xF344, 0xF3C4, 0xF444, 0xF4C4, 0xF544, 0xF5C4, 0xF644, 0xF6C4, 0xF744, 0xF7C4, 0xF844,
+    0xF8A4, 0xF8E4, 0xF924, 0xF964, 0xF9A4, 0xF9E4, 0xFA24, 0xFA64, 0xFAA4, 0xFAE4, 0xFB24, 0xFB64, 0xFBA4, 0xFBE4, 0xFC24, 0xFC64,
+    0xFC94, 0xFCB4, 0xFCD4, 0xFCF4, 0xFD14, 0xFD34, 0xFD54, 0xFD74, 0xFD94, 0xFDB4, 0xFDD4, 0xFDF4, 0xFE14, 0xFE34, 0xFE54, 0xFE74,
+    0xFE8C, 0xFE9C, 0xFEAC, 0xFEBC, 0xFECC, 0xFEDC, 0xFEEC, 0xFEFC, 0xFF0C, 0xFF1C, 0xFF2C, 0xFF3C, 0xFF4C, 0xFF5C, 0xFF6C, 0xFF7C,
+    0xFF88, 0xFF90, 0xFF98, 0xFFA0, 0xFFA8, 0xFFB0, 0xFFB8, 0xFFC0, 0xFFC8, 0xFFD0, 0xFFD8, 0xFFE0, 0xFFE8, 0xFFF0, 0xFFF8, 0x0000,
+    0x7D7C, 0x797C, 0x757C, 0x717C, 0x6D7C, 0x697C, 0x657C, 0x617C, 0x5D7C, 0x597C, 0x557C, 0x517C, 0x4D7C, 0x497C, 0x457C, 0x417C,
+    0x3E7C, 0x3C7C, 0x3A7C, 0x387C, 0x367C, 0x347C, 0x327C, 0x307C, 0x2E7C, 0x2C7C, 0x2A7C, 0x287C, 0x267C, 0x247C, 0x227C, 0x207C,
+    0x1EFC, 0x1DFC, 0x1CFC, 0x1BFC, 0x1AFC, 0x19FC, 0x18FC, 0x17FC, 0x16FC, 0x15FC, 0x14FC, 0x13FC, 0x12FC, 0x11FC, 0x10FC, 0x0FFC,
+    0x0F3C, 0x0EBC, 0x0E3C, 0x0DBC, 0x0D3C, 0x0CBC, 0x0C3C, 0x0BBC, 0x0B3C, 0x0ABC, 0x0A3C, 0x09BC, 0x093C, 0x08BC, 0x083C, 0x07BC,
+    0x075C, 0x071C, 0x06DC, 0x069C, 0x065C, 0x061C, 0x05DC, 0x059C, 0x055C, 0x051C, 0x04DC, 0x049C, 0x045C, 0x041C, 0x03DC, 0x039C,
+    0x036C, 0x034C, 0x032C, 0x030C, 0x02EC, 0x02CC, 0x02AC, 0x028C, 0x026C, 0x024C, 0x022C, 0x020C, 0x01EC, 0x01CC, 0x01AC, 0x018C,
+    0x0174, 0x0164, 0x0154, 0x0144, 0x0134, 0x0124, 0x0114, 0x0104, 0x00F4, 0x00E4, 0x00D4, 0x00C4, 0x00B4, 0x00A4, 0x0094, 0x0084,
+    0x0078, 0x0070, 0x0068, 0x0060, 0x0058, 0x0050, 0x0048, 0x0040, 0x0038, 0x0030, 0x0028, 0x0020, 0x0018, 0x0010, 0x0008, 0x0000
+};
+
+static DRWAV_INLINE drwav_int16 drwav__alaw_to_s16(drwav_uint8 sampleIn)
+{
+    return (short)g_drwavAlawTable[sampleIn];
+}
+
+static DRWAV_INLINE drwav_int16 drwav__mulaw_to_s16(drwav_uint8 sampleIn)
+{
+    return (short)g_drwavMulawTable[sampleIn];
+}
+
+
+
+DRWAV_PRIVATE void drwav__pcm_to_s16(drwav_int16* pOut, const drwav_uint8* pIn, size_t totalSampleCount, unsigned int bytesPerSample)
+{
+    size_t i;
+
+    /* Special case for 8-bit sample data because it's treated as unsigned. */
+    if (bytesPerSample == 1) {
+        drwav_u8_to_s16(pOut, pIn, totalSampleCount);
+        return;
+    }
+
+
+    /* Slightly more optimal implementation for common formats. */
+    if (bytesPerSample == 2) {
+        for (i = 0; i < totalSampleCount; ++i) {
+           *pOut++ = ((const drwav_int16*)pIn)[i];
+        }
+        return;
+    }
+    if (bytesPerSample == 3) {
+        drwav_s24_to_s16(pOut, pIn, totalSampleCount);
+        return;
+    }
+    if (bytesPerSample == 4) {
+        drwav_s32_to_s16(pOut, (const drwav_int32*)pIn, totalSampleCount);
+        return;
+    }
+
+
+    /* Anything more than 64 bits per sample is not supported. */
+    if (bytesPerSample > 8) {
+        DRWAV_ZERO_MEMORY(pOut, totalSampleCount * sizeof(*pOut));
+        return;
+    }
+
+
+    /* Generic, slow converter. */
+    for (i = 0; i < totalSampleCount; ++i) {
+        drwav_uint64 sample = 0;
+        unsigned int shift  = (8 - bytesPerSample) * 8;
+
+        unsigned int j;
+        for (j = 0; j < bytesPerSample; j += 1) {
+            DRWAV_ASSERT(j < 8);
+            sample |= (drwav_uint64)(pIn[j]) << shift;
+            shift  += 8;
+        }
+
+        pIn += j;
+        *pOut++ = (drwav_int16)((drwav_int64)sample >> 48);
+    }
+}
+
+DRWAV_PRIVATE void drwav__ieee_to_s16(drwav_int16* pOut, const drwav_uint8* pIn, size_t totalSampleCount, unsigned int bytesPerSample)
+{
+    if (bytesPerSample == 4) {
+        drwav_f32_to_s16(pOut, (const float*)pIn, totalSampleCount);
+        return;
+    } else if (bytesPerSample == 8) {
+        drwav_f64_to_s16(pOut, (const double*)pIn, totalSampleCount);
+        return;
+    } else {
+        /* Only supporting 32- and 64-bit float. Output silence in all other cases. Contributions welcome for 16-bit float. */
+        DRWAV_ZERO_MEMORY(pOut, totalSampleCount * sizeof(*pOut));
+        return;
+    }
+}
+
+DRWAV_PRIVATE drwav_uint64 drwav_read_pcm_frames_s16__pcm(drwav* pWav, drwav_uint64 framesToRead, drwav_int16* pBufferOut)
+{
+    drwav_uint64 totalFramesRead;
+    drwav_uint8 sampleData[4096] = {0};
+    drwav_uint32 bytesPerFrame;
+    drwav_uint32 bytesPerSample;
+    drwav_uint64 samplesRead;
+
+    /* Fast path. */
+    if ((pWav->translatedFormatTag == DR_WAVE_FORMAT_PCM && pWav->bitsPerSample == 16) || pBufferOut == NULL) {
+        return drwav_read_pcm_frames(pWav, framesToRead, pBufferOut);
+    }
+
+    bytesPerFrame = drwav_get_bytes_per_pcm_frame(pWav);
+    if (bytesPerFrame == 0) {
+        return 0;
+    }
+
+    bytesPerSample = bytesPerFrame / pWav->channels;
+    if (bytesPerSample == 0 || (bytesPerFrame % pWav->channels) != 0) {
+        return 0;   /* Only byte-aligned formats are supported. */
+    }
+
+    totalFramesRead = 0;
+
+    while (framesToRead > 0) {
+        drwav_uint64 framesToReadThisIteration = drwav_min(framesToRead, sizeof(sampleData)/bytesPerFrame);
+        drwav_uint64 framesRead = drwav_read_pcm_frames(pWav, framesToReadThisIteration, sampleData);
+        if (framesRead == 0) {
+            break;
+        }
+
+        DRWAV_ASSERT(framesRead <= framesToReadThisIteration);   /* If this fails it means there's a bug in drwav_read_pcm_frames(). */
+
+        /* Validation to ensure we don't read too much from out intermediary buffer. This is to protect from invalid files. */
+        samplesRead = framesRead * pWav->channels;
+        if ((samplesRead * bytesPerSample) > sizeof(sampleData)) {
+            DRWAV_ASSERT(DRWAV_FALSE);  /* This should never happen with a valid file. */
+            break;
+        }
+
+        drwav__pcm_to_s16(pBufferOut, sampleData, (size_t)samplesRead, bytesPerSample);
+
+        pBufferOut      += samplesRead;
+        framesToRead    -= framesRead;
+        totalFramesRead += framesRead;
+    }
+
+    return totalFramesRead;
+}
+
+DRWAV_PRIVATE drwav_uint64 drwav_read_pcm_frames_s16__ieee(drwav* pWav, drwav_uint64 framesToRead, drwav_int16* pBufferOut)
+{
+    drwav_uint64 totalFramesRead;
+    drwav_uint8 sampleData[4096] = {0};
+    drwav_uint32 bytesPerFrame;
+    drwav_uint32 bytesPerSample;
+    drwav_uint64 samplesRead;
+
+    if (pBufferOut == NULL) {
+        return drwav_read_pcm_frames(pWav, framesToRead, NULL);
+    }
+
+    bytesPerFrame = drwav_get_bytes_per_pcm_frame(pWav);
+    if (bytesPerFrame == 0) {
+        return 0;
+    }
+
+    bytesPerSample = bytesPerFrame / pWav->channels;
+    if (bytesPerSample == 0 || (bytesPerFrame % pWav->channels) != 0) {
+        return 0;   /* Only byte-aligned formats are supported. */
+    }
+
+    totalFramesRead = 0;
+
+    while (framesToRead > 0) {
+        drwav_uint64 framesToReadThisIteration = drwav_min(framesToRead, sizeof(sampleData)/bytesPerFrame);
+        drwav_uint64 framesRead = drwav_read_pcm_frames(pWav, framesToReadThisIteration, sampleData);
+        if (framesRead == 0) {
+            break;
+        }
+
+        DRWAV_ASSERT(framesRead <= framesToReadThisIteration);   /* If this fails it means there's a bug in drwav_read_pcm_frames(). */
+
+        /* Validation to ensure we don't read too much from out intermediary buffer. This is to protect from invalid files. */
+        samplesRead = framesRead * pWav->channels;
+        if ((samplesRead * bytesPerSample) > sizeof(sampleData)) {
+            DRWAV_ASSERT(DRWAV_FALSE);  /* This should never happen with a valid file. */
+            break;
+        }
+
+        drwav__ieee_to_s16(pBufferOut, sampleData, (size_t)samplesRead, bytesPerSample);    /* Safe cast. */
+
+        pBufferOut      += samplesRead;
+        framesToRead    -= framesRead;
+        totalFramesRead += framesRead;
+    }
+
+    return totalFramesRead;
+}
+
+DRWAV_PRIVATE drwav_uint64 drwav_read_pcm_frames_s16__alaw(drwav* pWav, drwav_uint64 framesToRead, drwav_int16* pBufferOut)
+{
+    drwav_uint64 totalFramesRead;
+    drwav_uint8 sampleData[4096] = {0};
+    drwav_uint32 bytesPerFrame;
+    drwav_uint32 bytesPerSample;
+    drwav_uint64 samplesRead;
+
+    if (pBufferOut == NULL) {
+        return drwav_read_pcm_frames(pWav, framesToRead, NULL);
+    }
+
+    bytesPerFrame = drwav_get_bytes_per_pcm_frame(pWav);
+    if (bytesPerFrame == 0) {
+        return 0;
+    }
+
+    bytesPerSample = bytesPerFrame / pWav->channels;
+    if (bytesPerSample == 0 || (bytesPerFrame % pWav->channels) != 0) {
+        return 0;   /* Only byte-aligned formats are supported. */
+    }
+
+    totalFramesRead = 0;
+
+    while (framesToRead > 0) {
+        drwav_uint64 framesToReadThisIteration = drwav_min(framesToRead, sizeof(sampleData)/bytesPerFrame);
+        drwav_uint64 framesRead = drwav_read_pcm_frames(pWav, framesToReadThisIteration, sampleData);
+        if (framesRead == 0) {
+            break;
+        }
+
+        DRWAV_ASSERT(framesRead <= framesToReadThisIteration);   /* If this fails it means there's a bug in drwav_read_pcm_frames(). */
+
+        /* Validation to ensure we don't read too much from out intermediary buffer. This is to protect from invalid files. */
+        samplesRead = framesRead * pWav->channels;
+        if ((samplesRead * bytesPerSample) > sizeof(sampleData)) {
+            DRWAV_ASSERT(DRWAV_FALSE);  /* This should never happen with a valid file. */
+            break;
+        }
+
+        drwav_alaw_to_s16(pBufferOut, sampleData, (size_t)samplesRead);
+
+        /*
+        For some reason libsndfile seems to be returning samples of the opposite sign for a-law, but only
+        with AIFF files. For WAV files it seems to be the same as dr_wav. This is resulting in dr_wav's
+        automated tests failing. I'm not sure which is correct, but will assume dr_wav. If we're enforcing
+        libsndfile compatibility we'll swap the signs here.
+        */
+        #ifdef DR_WAV_LIBSNDFILE_COMPAT
+        {
+            if (pWav->container == drwav_container_aiff) {
+                drwav_uint64 iSample;
+                for (iSample = 0; iSample < samplesRead; iSample += 1) {
+                    pBufferOut[iSample] = -pBufferOut[iSample];
+                }
+            }
+        }
+        #endif
+
+        pBufferOut      += samplesRead;
+        framesToRead    -= framesRead;
+        totalFramesRead += framesRead;
+    }
+
+    return totalFramesRead;
+}
+
+DRWAV_PRIVATE drwav_uint64 drwav_read_pcm_frames_s16__mulaw(drwav* pWav, drwav_uint64 framesToRead, drwav_int16* pBufferOut)
+{
+    drwav_uint64 totalFramesRead;
+    drwav_uint8 sampleData[4096] = {0};
+    drwav_uint32 bytesPerFrame;
+    drwav_uint32 bytesPerSample;
+    drwav_uint64 samplesRead;
+
+    if (pBufferOut == NULL) {
+        return drwav_read_pcm_frames(pWav, framesToRead, NULL);
+    }
+
+    bytesPerFrame = drwav_get_bytes_per_pcm_frame(pWav);
+    if (bytesPerFrame == 0) {
+        return 0;
+    }
+
+    bytesPerSample = bytesPerFrame / pWav->channels;
+    if (bytesPerSample == 0 || (bytesPerFrame % pWav->channels) != 0) {
+        return 0;   /* Only byte-aligned formats are supported. */
+    }
+
+    totalFramesRead = 0;
+
+    while (framesToRead > 0) {
+        drwav_uint64 framesToReadThisIteration = drwav_min(framesToRead, sizeof(sampleData)/bytesPerFrame);
+        drwav_uint64 framesRead = drwav_read_pcm_frames(pWav, framesToReadThisIteration, sampleData);
+        if (framesRead == 0) {
+            break;
+        }
+
+        DRWAV_ASSERT(framesRead <= framesToReadThisIteration);   /* If this fails it means there's a bug in drwav_read_pcm_frames(). */
+
+        /* Validation to ensure we don't read too much from out intermediary buffer. This is to protect from invalid files. */
+        samplesRead = framesRead * pWav->channels;
+        if ((samplesRead * bytesPerSample) > sizeof(sampleData)) {
+            DRWAV_ASSERT(DRWAV_FALSE);  /* This should never happen with a valid file. */
+            break;
+        }
+
+        drwav_mulaw_to_s16(pBufferOut, sampleData, (size_t)samplesRead);
+
+        /*
+        Just like with alaw, for some reason the signs between libsndfile and dr_wav are opposite. We just need to
+        swap the sign if we're compiling with libsndfile compatiblity so our automated tests don't fail.
+        */
+        #ifdef DR_WAV_LIBSNDFILE_COMPAT
+        {
+            if (pWav->container == drwav_container_aiff) {
+                drwav_uint64 iSample;
+                for (iSample = 0; iSample < samplesRead; iSample += 1) {
+                    pBufferOut[iSample] = -pBufferOut[iSample];
+                }
+            }
+        }
+        #endif
+
+        pBufferOut      += samplesRead;
+        framesToRead    -= framesRead;
+        totalFramesRead += framesRead;
+    }
+
+    return totalFramesRead;
+}
+
+DRWAV_API drwav_uint64 drwav_read_pcm_frames_s16(drwav* pWav, drwav_uint64 framesToRead, drwav_int16* pBufferOut)
+{
+    if (pWav == NULL || framesToRead == 0) {
+        return 0;
+    }
+
+    if (pBufferOut == NULL) {
+        return drwav_read_pcm_frames(pWav, framesToRead, NULL);
+    }
+
+    /* Don't try to read more samples than can potentially fit in the output buffer. */
+    if (framesToRead * pWav->channels * sizeof(drwav_int16) > DRWAV_SIZE_MAX) {
+        framesToRead = DRWAV_SIZE_MAX / sizeof(drwav_int16) / pWav->channels;
+    }
+
+    if (pWav->translatedFormatTag == DR_WAVE_FORMAT_PCM) {
+        return drwav_read_pcm_frames_s16__pcm(pWav, framesToRead, pBufferOut);
+    }
+
+    if (pWav->translatedFormatTag == DR_WAVE_FORMAT_IEEE_FLOAT) {
+        return drwav_read_pcm_frames_s16__ieee(pWav, framesToRead, pBufferOut);
+    }
+
+    if (pWav->translatedFormatTag == DR_WAVE_FORMAT_ALAW) {
+        return drwav_read_pcm_frames_s16__alaw(pWav, framesToRead, pBufferOut);
+    }
+
+    if (pWav->translatedFormatTag == DR_WAVE_FORMAT_MULAW) {
+        return drwav_read_pcm_frames_s16__mulaw(pWav, framesToRead, pBufferOut);
+    }
+
+    if (pWav->translatedFormatTag == DR_WAVE_FORMAT_ADPCM) {
+        return drwav_read_pcm_frames_s16__msadpcm(pWav, framesToRead, pBufferOut);
+    }
+
+    if (pWav->translatedFormatTag == DR_WAVE_FORMAT_DVI_ADPCM) {
+        return drwav_read_pcm_frames_s16__ima(pWav, framesToRead, pBufferOut);
+    }
+
+    return 0;
+}
+
+DRWAV_API drwav_uint64 drwav_read_pcm_frames_s16le(drwav* pWav, drwav_uint64 framesToRead, drwav_int16* pBufferOut)
+{
+    drwav_uint64 framesRead = drwav_read_pcm_frames_s16(pWav, framesToRead, pBufferOut);
+    if (pBufferOut != NULL && drwav__is_little_endian() == DRWAV_FALSE) {
+        drwav__bswap_samples_s16(pBufferOut, framesRead*pWav->channels);
+    }
+
+    return framesRead;
+}
+
+DRWAV_API drwav_uint64 drwav_read_pcm_frames_s16be(drwav* pWav, drwav_uint64 framesToRead, drwav_int16* pBufferOut)
+{
+    drwav_uint64 framesRead = drwav_read_pcm_frames_s16(pWav, framesToRead, pBufferOut);
+    if (pBufferOut != NULL && drwav__is_little_endian() == DRWAV_TRUE) {
+        drwav__bswap_samples_s16(pBufferOut, framesRead*pWav->channels);
+    }
+
+    return framesRead;
+}
+
+
+DRWAV_API void drwav_u8_to_s16(drwav_int16* pOut, const drwav_uint8* pIn, size_t sampleCount)
+{
+    int r;
+    size_t i;
+    for (i = 0; i < sampleCount; ++i) {
+        int x = pIn[i];
+        r = x << 8;
+        r = r - 32768;
+        pOut[i] = (short)r;
+    }
+}
+
+DRWAV_API void drwav_s24_to_s16(drwav_int16* pOut, const drwav_uint8* pIn, size_t sampleCount)
+{
+    int r;
+    size_t i;
+    for (i = 0; i < sampleCount; ++i) {
+        int x = ((int)(((unsigned int)(((const drwav_uint8*)pIn)[i*3+0]) << 8) | ((unsigned int)(((const drwav_uint8*)pIn)[i*3+1]) << 16) | ((unsigned int)(((const drwav_uint8*)pIn)[i*3+2])) << 24)) >> 8;
+        r = x >> 8;
+        pOut[i] = (short)r;
+    }
+}
+
+DRWAV_API void drwav_s32_to_s16(drwav_int16* pOut, const drwav_int32* pIn, size_t sampleCount)
+{
+    int r;
+    size_t i;
+    for (i = 0; i < sampleCount; ++i) {
+        int x = pIn[i];
+        r = x >> 16;
+        pOut[i] = (short)r;
+    }
+}
+
+DRWAV_API void drwav_f32_to_s16(drwav_int16* pOut, const float* pIn, size_t sampleCount)
+{
+    int r;
+    size_t i;
+    for (i = 0; i < sampleCount; ++i) {
+        float x = pIn[i];
+        float c;
+        c = ((x < -1) ? -1 : ((x > 1) ? 1 : x));
+        c = c + 1;
+        r = (int)(c * 32767.5f);
+        r = r - 32768;
+        pOut[i] = (short)r;
+    }
+}
+
+DRWAV_API void drwav_f64_to_s16(drwav_int16* pOut, const double* pIn, size_t sampleCount)
+{
+    int r;
+    size_t i;
+    for (i = 0; i < sampleCount; ++i) {
+        double x = pIn[i];
+        double c;
+        c = ((x < -1) ? -1 : ((x > 1) ? 1 : x));
+        c = c + 1;
+        r = (int)(c * 32767.5);
+        r = r - 32768;
+        pOut[i] = (short)r;
+    }
+}
+
+DRWAV_API void drwav_alaw_to_s16(drwav_int16* pOut, const drwav_uint8* pIn, size_t sampleCount)
+{
+    size_t i;
+    for (i = 0; i < sampleCount; ++i) {
+        pOut[i] = drwav__alaw_to_s16(pIn[i]);
+    }
+}
+
+DRWAV_API void drwav_mulaw_to_s16(drwav_int16* pOut, const drwav_uint8* pIn, size_t sampleCount)
+{
+    size_t i;
+    for (i = 0; i < sampleCount; ++i) {
+        pOut[i] = drwav__mulaw_to_s16(pIn[i]);
+    }
+}
+
+
+DRWAV_PRIVATE void drwav__pcm_to_f32(float* pOut, const drwav_uint8* pIn, size_t sampleCount, unsigned int bytesPerSample)
+{
+    unsigned int i;
+
+    /* Special case for 8-bit sample data because it's treated as unsigned. */
+    if (bytesPerSample == 1) {
+        drwav_u8_to_f32(pOut, pIn, sampleCount);
+        return;
+    }
+
+    /* Slightly more optimal implementation for common formats. */
+    if (bytesPerSample == 2) {
+        drwav_s16_to_f32(pOut, (const drwav_int16*)pIn, sampleCount);
+        return;
+    }
+    if (bytesPerSample == 3) {
+        drwav_s24_to_f32(pOut, pIn, sampleCount);
+        return;
+    }
+    if (bytesPerSample == 4) {
+        drwav_s32_to_f32(pOut, (const drwav_int32*)pIn, sampleCount);
+        return;
+    }
+
+
+    /* Anything more than 64 bits per sample is not supported. */
+    if (bytesPerSample > 8) {
+        DRWAV_ZERO_MEMORY(pOut, sampleCount * sizeof(*pOut));
+        return;
+    }
+
+
+    /* Generic, slow converter. */
+    for (i = 0; i < sampleCount; ++i) {
+        drwav_uint64 sample = 0;
+        unsigned int shift  = (8 - bytesPerSample) * 8;
+
+        unsigned int j;
+        for (j = 0; j < bytesPerSample; j += 1) {
+            DRWAV_ASSERT(j < 8);
+            sample |= (drwav_uint64)(pIn[j]) << shift;
+            shift  += 8;
+        }
+
+        pIn += j;
+        *pOut++ = (float)((drwav_int64)sample / 9223372036854775807.0);
+    }
+}
+
+DRWAV_PRIVATE void drwav__ieee_to_f32(float* pOut, const drwav_uint8* pIn, size_t sampleCount, unsigned int bytesPerSample)
+{
+    if (bytesPerSample == 4) {
+        unsigned int i;
+        for (i = 0; i < sampleCount; ++i) {
+            *pOut++ = ((const float*)pIn)[i];
+        }
+        return;
+    } else if (bytesPerSample == 8) {
+        drwav_f64_to_f32(pOut, (const double*)pIn, sampleCount);
+        return;
+    } else {
+        /* Only supporting 32- and 64-bit float. Output silence in all other cases. Contributions welcome for 16-bit float. */
+        DRWAV_ZERO_MEMORY(pOut, sampleCount * sizeof(*pOut));
+        return;
+    }
+}
+
+
+DRWAV_PRIVATE drwav_uint64 drwav_read_pcm_frames_f32__pcm(drwav* pWav, drwav_uint64 framesToRead, float* pBufferOut)
+{
+    drwav_uint64 totalFramesRead;
+    drwav_uint8 sampleData[4096] = {0};
+    drwav_uint32 bytesPerFrame;
+    drwav_uint32 bytesPerSample;
+    drwav_uint64 samplesRead;
+
+    bytesPerFrame = drwav_get_bytes_per_pcm_frame(pWav);
+    if (bytesPerFrame == 0) {
+        return 0;
+    }
+
+    bytesPerSample = bytesPerFrame / pWav->channels;
+    if (bytesPerSample == 0 || (bytesPerFrame % pWav->channels) != 0) {
+        return 0;   /* Only byte-aligned formats are supported. */
+    }
+
+    totalFramesRead = 0;
+
+    while (framesToRead > 0) {
+        drwav_uint64 framesToReadThisIteration = drwav_min(framesToRead, sizeof(sampleData)/bytesPerFrame);
+        drwav_uint64 framesRead = drwav_read_pcm_frames(pWav, framesToReadThisIteration, sampleData);
+        if (framesRead == 0) {
+            break;
+        }
+
+        DRWAV_ASSERT(framesRead <= framesToReadThisIteration);   /* If this fails it means there's a bug in drwav_read_pcm_frames(). */
+
+        /* Validation to ensure we don't read too much from out intermediary buffer. This is to protect from invalid files. */
+        samplesRead = framesRead * pWav->channels;
+        if ((samplesRead * bytesPerSample) > sizeof(sampleData)) {
+            DRWAV_ASSERT(DRWAV_FALSE);  /* This should never happen with a valid file. */
+            break;
+        }
+
+        drwav__pcm_to_f32(pBufferOut, sampleData, (size_t)samplesRead, bytesPerSample);
+
+        pBufferOut      += samplesRead;
+        framesToRead    -= framesRead;
+        totalFramesRead += framesRead;
+    }
+
+    return totalFramesRead;
+}
+
+DRWAV_PRIVATE drwav_uint64 drwav_read_pcm_frames_f32__msadpcm_ima(drwav* pWav, drwav_uint64 framesToRead, float* pBufferOut)
+{
+    /*
+    We're just going to borrow the implementation from the drwav_read_s16() since ADPCM is a little bit more complicated than other formats and I don't
+    want to duplicate that code.
+    */
+    drwav_uint64 totalFramesRead;
+    drwav_int16 samples16[2048];
+
+    totalFramesRead = 0;
+
+    while (framesToRead > 0) {
+        drwav_uint64 framesToReadThisIteration = drwav_min(framesToRead, drwav_countof(samples16)/pWav->channels);
+        drwav_uint64 framesRead = drwav_read_pcm_frames_s16(pWav, framesToReadThisIteration, samples16);
+        if (framesRead == 0) {
+            break;
+        }
+
+        DRWAV_ASSERT(framesRead <= framesToReadThisIteration);   /* If this fails it means there's a bug in drwav_read_pcm_frames(). */
+
+        drwav_s16_to_f32(pBufferOut, samples16, (size_t)(framesRead*pWav->channels));   /* <-- Safe cast because we're clamping to 2048. */
+
+        pBufferOut      += framesRead*pWav->channels;
+        framesToRead    -= framesRead;
+        totalFramesRead += framesRead;
+    }
+
+    return totalFramesRead;
+}
+
+DRWAV_PRIVATE drwav_uint64 drwav_read_pcm_frames_f32__ieee(drwav* pWav, drwav_uint64 framesToRead, float* pBufferOut)
+{
+    drwav_uint64 totalFramesRead;
+    drwav_uint8 sampleData[4096] = {0};
+    drwav_uint32 bytesPerFrame;
+    drwav_uint32 bytesPerSample;
+    drwav_uint64 samplesRead;
+
+    /* Fast path. */
+    if (pWav->translatedFormatTag == DR_WAVE_FORMAT_IEEE_FLOAT && pWav->bitsPerSample == 32) {
+        return drwav_read_pcm_frames(pWav, framesToRead, pBufferOut);
+    }
+
+    bytesPerFrame = drwav_get_bytes_per_pcm_frame(pWav);
+    if (bytesPerFrame == 0) {
+        return 0;
+    }
+
+    bytesPerSample = bytesPerFrame / pWav->channels;
+    if (bytesPerSample == 0 || (bytesPerFrame % pWav->channels) != 0) {
+        return 0;   /* Only byte-aligned formats are supported. */
+    }
+
+    totalFramesRead = 0;
+
+    while (framesToRead > 0) {
+        drwav_uint64 framesToReadThisIteration = drwav_min(framesToRead, sizeof(sampleData)/bytesPerFrame);
+        drwav_uint64 framesRead = drwav_read_pcm_frames(pWav, framesToReadThisIteration, sampleData);
+        if (framesRead == 0) {
+            break;
+        }
+
+        DRWAV_ASSERT(framesRead <= framesToReadThisIteration);   /* If this fails it means there's a bug in drwav_read_pcm_frames(). */
+
+        /* Validation to ensure we don't read too much from out intermediary buffer. This is to protect from invalid files. */
+        samplesRead = framesRead * pWav->channels;
+        if ((samplesRead * bytesPerSample) > sizeof(sampleData)) {
+            DRWAV_ASSERT(DRWAV_FALSE);  /* This should never happen with a valid file. */
+            break;
+        }
+
+        drwav__ieee_to_f32(pBufferOut, sampleData, (size_t)samplesRead, bytesPerSample);
+
+        pBufferOut      += samplesRead;
+        framesToRead    -= framesRead;
+        totalFramesRead += framesRead;
+    }
+
+    return totalFramesRead;
+}
+
+DRWAV_PRIVATE drwav_uint64 drwav_read_pcm_frames_f32__alaw(drwav* pWav, drwav_uint64 framesToRead, float* pBufferOut)
+{
+    drwav_uint64 totalFramesRead;
+    drwav_uint8 sampleData[4096] = {0};
+    drwav_uint32 bytesPerFrame;
+    drwav_uint32 bytesPerSample;
+    drwav_uint64 samplesRead;
+
+    bytesPerFrame = drwav_get_bytes_per_pcm_frame(pWav);
+    if (bytesPerFrame == 0) {
+        return 0;
+    }
+
+    bytesPerSample = bytesPerFrame / pWav->channels;
+    if (bytesPerSample == 0 || (bytesPerFrame % pWav->channels) != 0) {
+        return 0;   /* Only byte-aligned formats are supported. */
+    }
+
+    totalFramesRead = 0;
+
+    while (framesToRead > 0) {
+        drwav_uint64 framesToReadThisIteration = drwav_min(framesToRead, sizeof(sampleData)/bytesPerFrame);
+        drwav_uint64 framesRead = drwav_read_pcm_frames(pWav, framesToReadThisIteration, sampleData);
+        if (framesRead == 0) {
+            break;
+        }
+
+        DRWAV_ASSERT(framesRead <= framesToReadThisIteration);   /* If this fails it means there's a bug in drwav_read_pcm_frames(). */
+
+        /* Validation to ensure we don't read too much from out intermediary buffer. This is to protect from invalid files. */
+        samplesRead = framesRead * pWav->channels;
+        if ((samplesRead * bytesPerSample) > sizeof(sampleData)) {
+            DRWAV_ASSERT(DRWAV_FALSE);  /* This should never happen with a valid file. */
+            break;
+        }
+
+        drwav_alaw_to_f32(pBufferOut, sampleData, (size_t)samplesRead);
+
+        #ifdef DR_WAV_LIBSNDFILE_COMPAT
+        {
+            if (pWav->container == drwav_container_aiff) {
+                drwav_uint64 iSample;
+                for (iSample = 0; iSample < samplesRead; iSample += 1) {
+                    pBufferOut[iSample] = -pBufferOut[iSample];
+                }
+            }
+        }
+        #endif
+
+        pBufferOut      += samplesRead;
+        framesToRead    -= framesRead;
+        totalFramesRead += framesRead;
+    }
+
+    return totalFramesRead;
+}
+
+DRWAV_PRIVATE drwav_uint64 drwav_read_pcm_frames_f32__mulaw(drwav* pWav, drwav_uint64 framesToRead, float* pBufferOut)
+{
+    drwav_uint64 totalFramesRead;
+    drwav_uint8 sampleData[4096] = {0};
+    drwav_uint32 bytesPerFrame;
+    drwav_uint32 bytesPerSample;
+    drwav_uint64 samplesRead;
+
+    bytesPerFrame = drwav_get_bytes_per_pcm_frame(pWav);
+    if (bytesPerFrame == 0) {
+        return 0;
+    }
+
+    bytesPerSample = bytesPerFrame / pWav->channels;
+    if (bytesPerSample == 0 || (bytesPerFrame % pWav->channels) != 0) {
+        return 0;   /* Only byte-aligned formats are supported. */
+    }
+
+    totalFramesRead = 0;
+
+    while (framesToRead > 0) {
+        drwav_uint64 framesToReadThisIteration = drwav_min(framesToRead, sizeof(sampleData)/bytesPerFrame);
+        drwav_uint64 framesRead = drwav_read_pcm_frames(pWav, framesToReadThisIteration, sampleData);
+        if (framesRead == 0) {
+            break;
+        }
+
+        DRWAV_ASSERT(framesRead <= framesToReadThisIteration);   /* If this fails it means there's a bug in drwav_read_pcm_frames(). */
+
+        /* Validation to ensure we don't read too much from out intermediary buffer. This is to protect from invalid files. */
+        samplesRead = framesRead * pWav->channels;
+        if ((samplesRead * bytesPerSample) > sizeof(sampleData)) {
+            DRWAV_ASSERT(DRWAV_FALSE);  /* This should never happen with a valid file. */
+            break;
+        }
+
+        drwav_mulaw_to_f32(pBufferOut, sampleData, (size_t)samplesRead);
+
+        #ifdef DR_WAV_LIBSNDFILE_COMPAT
+        {
+            if (pWav->container == drwav_container_aiff) {
+                drwav_uint64 iSample;
+                for (iSample = 0; iSample < samplesRead; iSample += 1) {
+                    pBufferOut[iSample] = -pBufferOut[iSample];
+                }
+            }
+        }
+        #endif
+
+        pBufferOut      += samplesRead;
+        framesToRead    -= framesRead;
+        totalFramesRead += framesRead;
+    }
+
+    return totalFramesRead;
+}
+
+DRWAV_API drwav_uint64 drwav_read_pcm_frames_f32(drwav* pWav, drwav_uint64 framesToRead, float* pBufferOut)
+{
+    if (pWav == NULL || framesToRead == 0) {
+        return 0;
+    }
+
+    if (pBufferOut == NULL) {
+        return drwav_read_pcm_frames(pWav, framesToRead, NULL);
+    }
+
+    /* Don't try to read more samples than can potentially fit in the output buffer. */
+    if (framesToRead * pWav->channels * sizeof(float) > DRWAV_SIZE_MAX) {
+        framesToRead = DRWAV_SIZE_MAX / sizeof(float) / pWav->channels;
+    }
+
+    if (pWav->translatedFormatTag == DR_WAVE_FORMAT_PCM) {
+        return drwav_read_pcm_frames_f32__pcm(pWav, framesToRead, pBufferOut);
+    }
+
+    if (pWav->translatedFormatTag == DR_WAVE_FORMAT_ADPCM || pWav->translatedFormatTag == DR_WAVE_FORMAT_DVI_ADPCM) {
+        return drwav_read_pcm_frames_f32__msadpcm_ima(pWav, framesToRead, pBufferOut);
+    }
+
+    if (pWav->translatedFormatTag == DR_WAVE_FORMAT_IEEE_FLOAT) {
+        return drwav_read_pcm_frames_f32__ieee(pWav, framesToRead, pBufferOut);
+    }
+
+    if (pWav->translatedFormatTag == DR_WAVE_FORMAT_ALAW) {
+        return drwav_read_pcm_frames_f32__alaw(pWav, framesToRead, pBufferOut);
+    }
+
+    if (pWav->translatedFormatTag == DR_WAVE_FORMAT_MULAW) {
+        return drwav_read_pcm_frames_f32__mulaw(pWav, framesToRead, pBufferOut);
+    }
+
+    return 0;
+}
+
+DRWAV_API drwav_uint64 drwav_read_pcm_frames_f32le(drwav* pWav, drwav_uint64 framesToRead, float* pBufferOut)
+{
+    drwav_uint64 framesRead = drwav_read_pcm_frames_f32(pWav, framesToRead, pBufferOut);
+    if (pBufferOut != NULL && drwav__is_little_endian() == DRWAV_FALSE) {
+        drwav__bswap_samples_f32(pBufferOut, framesRead*pWav->channels);
+    }
+
+    return framesRead;
+}
+
+DRWAV_API drwav_uint64 drwav_read_pcm_frames_f32be(drwav* pWav, drwav_uint64 framesToRead, float* pBufferOut)
+{
+    drwav_uint64 framesRead = drwav_read_pcm_frames_f32(pWav, framesToRead, pBufferOut);
+    if (pBufferOut != NULL && drwav__is_little_endian() == DRWAV_TRUE) {
+        drwav__bswap_samples_f32(pBufferOut, framesRead*pWav->channels);
+    }
+
+    return framesRead;
+}
+
+
+DRWAV_API void drwav_u8_to_f32(float* pOut, const drwav_uint8* pIn, size_t sampleCount)
+{
+    size_t i;
+
+    if (pOut == NULL || pIn == NULL) {
+        return;
+    }
+
+#ifdef DR_WAV_LIBSNDFILE_COMPAT
+    /*
+    It appears libsndfile uses slightly different logic for the u8 -> f32 conversion to dr_wav, which in my opinion is incorrect. It appears
+    libsndfile performs the conversion something like "f32 = (u8 / 256) * 2 - 1", however I think it should be "f32 = (u8 / 255) * 2 - 1" (note
+    the divisor of 256 vs 255). I use libsndfile as a benchmark for testing, so I'm therefore leaving this block here just for my automated
+    correctness testing. This is disabled by default.
+    */
+    for (i = 0; i < sampleCount; ++i) {
+        *pOut++ = (pIn[i] / 256.0f) * 2 - 1;
+    }
+#else
+    for (i = 0; i < sampleCount; ++i) {
+        float x = pIn[i];
+        x = x * 0.00784313725490196078f;    /* 0..255 to 0..2 */
+        x = x - 1;                          /* 0..2 to -1..1 */
+
+        *pOut++ = x;
+    }
+#endif
+}
+
+DRWAV_API void drwav_s16_to_f32(float* pOut, const drwav_int16* pIn, size_t sampleCount)
+{
+    size_t i;
+
+    if (pOut == NULL || pIn == NULL) {
+        return;
+    }
+
+    for (i = 0; i < sampleCount; ++i) {
+        *pOut++ = pIn[i] * 0.000030517578125f;
+    }
+}
+
+DRWAV_API void drwav_s24_to_f32(float* pOut, const drwav_uint8* pIn, size_t sampleCount)
+{
+    size_t i;
+
+    if (pOut == NULL || pIn == NULL) {
+        return;
+    }
+
+    for (i = 0; i < sampleCount; ++i) {
+        double x;
+        drwav_uint32 a = ((drwav_uint32)(pIn[i*3+0]) <<  8);
+        drwav_uint32 b = ((drwav_uint32)(pIn[i*3+1]) << 16);
+        drwav_uint32 c = ((drwav_uint32)(pIn[i*3+2]) << 24);
+
+        x = (double)((drwav_int32)(a | b | c) >> 8);
+        *pOut++ = (float)(x * 0.00000011920928955078125);
+    }
+}
+
+DRWAV_API void drwav_s32_to_f32(float* pOut, const drwav_int32* pIn, size_t sampleCount)
+{
+    size_t i;
+    if (pOut == NULL || pIn == NULL) {
+        return;
+    }
+
+    for (i = 0; i < sampleCount; ++i) {
+        *pOut++ = (float)(pIn[i] / 2147483648.0);
+    }
+}
+
+DRWAV_API void drwav_f64_to_f32(float* pOut, const double* pIn, size_t sampleCount)
+{
+    size_t i;
+
+    if (pOut == NULL || pIn == NULL) {
+        return;
+    }
+
+    for (i = 0; i < sampleCount; ++i) {
+        *pOut++ = (float)pIn[i];
+    }
+}
+
+DRWAV_API void drwav_alaw_to_f32(float* pOut, const drwav_uint8* pIn, size_t sampleCount)
+{
+    size_t i;
+
+    if (pOut == NULL || pIn == NULL) {
+        return;
+    }
+
+    for (i = 0; i < sampleCount; ++i) {
+        *pOut++ = drwav__alaw_to_s16(pIn[i]) / 32768.0f;
+    }
+}
+
+DRWAV_API void drwav_mulaw_to_f32(float* pOut, const drwav_uint8* pIn, size_t sampleCount)
+{
+    size_t i;
+
+    if (pOut == NULL || pIn == NULL) {
+        return;
+    }
+
+    for (i = 0; i < sampleCount; ++i) {
+        *pOut++ = drwav__mulaw_to_s16(pIn[i]) / 32768.0f;
+    }
+}
+
+
+
+DRWAV_PRIVATE void drwav__pcm_to_s32(drwav_int32* pOut, const drwav_uint8* pIn, size_t totalSampleCount, unsigned int bytesPerSample)
+{
+    unsigned int i;
+
+    /* Special case for 8-bit sample data because it's treated as unsigned. */
+    if (bytesPerSample == 1) {
+        drwav_u8_to_s32(pOut, pIn, totalSampleCount);
+        return;
+    }
+
+    /* Slightly more optimal implementation for common formats. */
+    if (bytesPerSample == 2) {
+        drwav_s16_to_s32(pOut, (const drwav_int16*)pIn, totalSampleCount);
+        return;
+    }
+    if (bytesPerSample == 3) {
+        drwav_s24_to_s32(pOut, pIn, totalSampleCount);
+        return;
+    }
+    if (bytesPerSample == 4) {
+        for (i = 0; i < totalSampleCount; ++i) {
+           *pOut++ = ((const drwav_int32*)pIn)[i];
+        }
+        return;
+    }
+
+
+    /* Anything more than 64 bits per sample is not supported. */
+    if (bytesPerSample > 8) {
+        DRWAV_ZERO_MEMORY(pOut, totalSampleCount * sizeof(*pOut));
+        return;
+    }
+
+
+    /* Generic, slow converter. */
+    for (i = 0; i < totalSampleCount; ++i) {
+        drwav_uint64 sample = 0;
+        unsigned int shift  = (8 - bytesPerSample) * 8;
+
+        unsigned int j;
+        for (j = 0; j < bytesPerSample; j += 1) {
+            DRWAV_ASSERT(j < 8);
+            sample |= (drwav_uint64)(pIn[j]) << shift;
+            shift  += 8;
+        }
+
+        pIn += j;
+        *pOut++ = (drwav_int32)((drwav_int64)sample >> 32);
+    }
+}
+
+DRWAV_PRIVATE void drwav__ieee_to_s32(drwav_int32* pOut, const drwav_uint8* pIn, size_t totalSampleCount, unsigned int bytesPerSample)
+{
+    if (bytesPerSample == 4) {
+        drwav_f32_to_s32(pOut, (const float*)pIn, totalSampleCount);
+        return;
+    } else if (bytesPerSample == 8) {
+        drwav_f64_to_s32(pOut, (const double*)pIn, totalSampleCount);
+        return;
+    } else {
+        /* Only supporting 32- and 64-bit float. Output silence in all other cases. Contributions welcome for 16-bit float. */
+        DRWAV_ZERO_MEMORY(pOut, totalSampleCount * sizeof(*pOut));
+        return;
+    }
+}
+
+
+DRWAV_PRIVATE drwav_uint64 drwav_read_pcm_frames_s32__pcm(drwav* pWav, drwav_uint64 framesToRead, drwav_int32* pBufferOut)
+{
+    drwav_uint64 totalFramesRead;
+    drwav_uint8 sampleData[4096] = {0};
+    drwav_uint32 bytesPerFrame;
+    drwav_uint32 bytesPerSample;
+    drwav_uint64 samplesRead;
+
+    /* Fast path. */
+    if (pWav->translatedFormatTag == DR_WAVE_FORMAT_PCM && pWav->bitsPerSample == 32) {
+        return drwav_read_pcm_frames(pWav, framesToRead, pBufferOut);
+    }
+
+    bytesPerFrame = drwav_get_bytes_per_pcm_frame(pWav);
+    if (bytesPerFrame == 0) {
+        return 0;
+    }
+
+    bytesPerSample = bytesPerFrame / pWav->channels;
+    if (bytesPerSample == 0 || (bytesPerFrame % pWav->channels) != 0) {
+        return 0;   /* Only byte-aligned formats are supported. */
+    }
+
+    totalFramesRead = 0;
+
+    while (framesToRead > 0) {
+        drwav_uint64 framesToReadThisIteration = drwav_min(framesToRead, sizeof(sampleData)/bytesPerFrame);
+        drwav_uint64 framesRead = drwav_read_pcm_frames(pWav, framesToReadThisIteration, sampleData);
+        if (framesRead == 0) {
+            break;
+        }
+
+        DRWAV_ASSERT(framesRead <= framesToReadThisIteration);   /* If this fails it means there's a bug in drwav_read_pcm_frames(). */
+
+        /* Validation to ensure we don't read too much from out intermediary buffer. This is to protect from invalid files. */
+        samplesRead = framesRead * pWav->channels;
+        if ((samplesRead * bytesPerSample) > sizeof(sampleData)) {
+            DRWAV_ASSERT(DRWAV_FALSE);  /* This should never happen with a valid file. */
+            break;
+        }
+
+        drwav__pcm_to_s32(pBufferOut, sampleData, (size_t)samplesRead, bytesPerSample);
+
+        pBufferOut      += samplesRead;
+        framesToRead    -= framesRead;
+        totalFramesRead += framesRead;
+    }
+
+    return totalFramesRead;
+}
+
+DRWAV_PRIVATE drwav_uint64 drwav_read_pcm_frames_s32__msadpcm_ima(drwav* pWav, drwav_uint64 framesToRead, drwav_int32* pBufferOut)
+{
+    /*
+    We're just going to borrow the implementation from the drwav_read_s16() since ADPCM is a little bit more complicated than other formats and I don't
+    want to duplicate that code.
+    */
+    drwav_uint64 totalFramesRead = 0;
+    drwav_int16 samples16[2048];
+
+    while (framesToRead > 0) {
+        drwav_uint64 framesToReadThisIteration = drwav_min(framesToRead, drwav_countof(samples16)/pWav->channels);
+        drwav_uint64 framesRead = drwav_read_pcm_frames_s16(pWav, framesToReadThisIteration, samples16);
+        if (framesRead == 0) {
+            break;
+        }
+
+        DRWAV_ASSERT(framesRead <= framesToReadThisIteration);   /* If this fails it means there's a bug in drwav_read_pcm_frames(). */
+
+        drwav_s16_to_s32(pBufferOut, samples16, (size_t)(framesRead*pWav->channels));   /* <-- Safe cast because we're clamping to 2048. */
+
+        pBufferOut      += framesRead*pWav->channels;
+        framesToRead    -= framesRead;
+        totalFramesRead += framesRead;
+    }
+
+    return totalFramesRead;
+}
+
+DRWAV_PRIVATE drwav_uint64 drwav_read_pcm_frames_s32__ieee(drwav* pWav, drwav_uint64 framesToRead, drwav_int32* pBufferOut)
+{
+    drwav_uint64 totalFramesRead;
+    drwav_uint8 sampleData[4096] = {0};
+    drwav_uint32 bytesPerFrame;
+    drwav_uint32 bytesPerSample;
+    drwav_uint64 samplesRead;
+
+    bytesPerFrame = drwav_get_bytes_per_pcm_frame(pWav);
+    if (bytesPerFrame == 0) {
+        return 0;
+    }
+
+    bytesPerSample = bytesPerFrame / pWav->channels;
+    if (bytesPerSample == 0 || (bytesPerFrame % pWav->channels) != 0) {
+        return 0;   /* Only byte-aligned formats are supported. */
+    }
+
+    totalFramesRead = 0;
+
+    while (framesToRead > 0) {
+        drwav_uint64 framesToReadThisIteration = drwav_min(framesToRead, sizeof(sampleData)/bytesPerFrame);
+        drwav_uint64 framesRead = drwav_read_pcm_frames(pWav, framesToReadThisIteration, sampleData);
+        if (framesRead == 0) {
+            break;
+        }
+
+        DRWAV_ASSERT(framesRead <= framesToReadThisIteration);   /* If this fails it means there's a bug in drwav_read_pcm_frames(). */
+
+        /* Validation to ensure we don't read too much from out intermediary buffer. This is to protect from invalid files. */
+        samplesRead = framesRead * pWav->channels;
+        if ((samplesRead * bytesPerSample) > sizeof(sampleData)) {
+            DRWAV_ASSERT(DRWAV_FALSE);  /* This should never happen with a valid file. */
+            break;
+        }
+
+        drwav__ieee_to_s32(pBufferOut, sampleData, (size_t)samplesRead, bytesPerSample);
+
+        pBufferOut      += samplesRead;
+        framesToRead    -= framesRead;
+        totalFramesRead += framesRead;
+    }
+
+    return totalFramesRead;
+}
+
+DRWAV_PRIVATE drwav_uint64 drwav_read_pcm_frames_s32__alaw(drwav* pWav, drwav_uint64 framesToRead, drwav_int32* pBufferOut)
+{
+    drwav_uint64 totalFramesRead;
+    drwav_uint8 sampleData[4096] = {0};
+    drwav_uint32 bytesPerFrame;
+    drwav_uint32 bytesPerSample;
+    drwav_uint64 samplesRead;
+
+    bytesPerFrame = drwav_get_bytes_per_pcm_frame(pWav);
+    if (bytesPerFrame == 0) {
+        return 0;
+    }
+
+    bytesPerSample = bytesPerFrame / pWav->channels;
+    if (bytesPerSample == 0 || (bytesPerFrame % pWav->channels) != 0) {
+        return 0;   /* Only byte-aligned formats are supported. */
+    }
+
+    totalFramesRead = 0;
+
+    while (framesToRead > 0) {
+        drwav_uint64 framesToReadThisIteration = drwav_min(framesToRead, sizeof(sampleData)/bytesPerFrame);
+        drwav_uint64 framesRead = drwav_read_pcm_frames(pWav, framesToReadThisIteration, sampleData);
+        if (framesRead == 0) {
+            break;
+        }
+
+        DRWAV_ASSERT(framesRead <= framesToReadThisIteration);   /* If this fails it means there's a bug in drwav_read_pcm_frames(). */
+
+        /* Validation to ensure we don't read too much from out intermediary buffer. This is to protect from invalid files. */
+        samplesRead = framesRead * pWav->channels;
+        if ((samplesRead * bytesPerSample) > sizeof(sampleData)) {
+            DRWAV_ASSERT(DRWAV_FALSE);  /* This should never happen with a valid file. */
+            break;
+        }
+
+        drwav_alaw_to_s32(pBufferOut, sampleData, (size_t)samplesRead);
+
+        #ifdef DR_WAV_LIBSNDFILE_COMPAT
+        {
+            if (pWav->container == drwav_container_aiff) {
+                drwav_uint64 iSample;
+                for (iSample = 0; iSample < samplesRead; iSample += 1) {
+                    pBufferOut[iSample] = -pBufferOut[iSample];
+                }
+            }
+        }
+        #endif
+
+        pBufferOut      += samplesRead;
+        framesToRead    -= framesRead;
+        totalFramesRead += framesRead;
+    }
+
+    return totalFramesRead;
+}
+
+DRWAV_PRIVATE drwav_uint64 drwav_read_pcm_frames_s32__mulaw(drwav* pWav, drwav_uint64 framesToRead, drwav_int32* pBufferOut)
+{
+    drwav_uint64 totalFramesRead;
+    drwav_uint8 sampleData[4096] = {0};
+    drwav_uint32 bytesPerFrame;
+    drwav_uint32 bytesPerSample;
+    drwav_uint64 samplesRead;
+
+    bytesPerFrame = drwav_get_bytes_per_pcm_frame(pWav);
+    if (bytesPerFrame == 0) {
+        return 0;
+    }
+
+    bytesPerSample = bytesPerFrame / pWav->channels;
+    if (bytesPerSample == 0 || (bytesPerFrame % pWav->channels) != 0) {
+        return 0;   /* Only byte-aligned formats are supported. */
+    }
+
+    totalFramesRead = 0;
+
+    while (framesToRead > 0) {
+        drwav_uint64 framesToReadThisIteration = drwav_min(framesToRead, sizeof(sampleData)/bytesPerFrame);
+        drwav_uint64 framesRead = drwav_read_pcm_frames(pWav, framesToReadThisIteration, sampleData);
+        if (framesRead == 0) {
+            break;
+        }
+
+        DRWAV_ASSERT(framesRead <= framesToReadThisIteration);   /* If this fails it means there's a bug in drwav_read_pcm_frames(). */
+
+        /* Validation to ensure we don't read too much from out intermediary buffer. This is to protect from invalid files. */
+        samplesRead = framesRead * pWav->channels;
+        if ((samplesRead * bytesPerSample) > sizeof(sampleData)) {
+            DRWAV_ASSERT(DRWAV_FALSE);  /* This should never happen with a valid file. */
+            break;
+        }
+
+        drwav_mulaw_to_s32(pBufferOut, sampleData, (size_t)samplesRead);
+
+        #ifdef DR_WAV_LIBSNDFILE_COMPAT
+        {
+            if (pWav->container == drwav_container_aiff) {
+                drwav_uint64 iSample;
+                for (iSample = 0; iSample < samplesRead; iSample += 1) {
+                    pBufferOut[iSample] = -pBufferOut[iSample];
+                }
+            }
+        }
+        #endif
+
+        pBufferOut      += samplesRead;
+        framesToRead    -= framesRead;
+        totalFramesRead += framesRead;
+    }
+
+    return totalFramesRead;
+}
+
+DRWAV_API drwav_uint64 drwav_read_pcm_frames_s32(drwav* pWav, drwav_uint64 framesToRead, drwav_int32* pBufferOut)
+{
+    if (pWav == NULL || framesToRead == 0) {
+        return 0;
+    }
+
+    if (pBufferOut == NULL) {
+        return drwav_read_pcm_frames(pWav, framesToRead, NULL);
+    }
+
+    /* Don't try to read more samples than can potentially fit in the output buffer. */
+    if (framesToRead * pWav->channels * sizeof(drwav_int32) > DRWAV_SIZE_MAX) {
+        framesToRead = DRWAV_SIZE_MAX / sizeof(drwav_int32) / pWav->channels;
+    }
+
+    if (pWav->translatedFormatTag == DR_WAVE_FORMAT_PCM) {
+        return drwav_read_pcm_frames_s32__pcm(pWav, framesToRead, pBufferOut);
+    }
+
+    if (pWav->translatedFormatTag == DR_WAVE_FORMAT_ADPCM || pWav->translatedFormatTag == DR_WAVE_FORMAT_DVI_ADPCM) {
+        return drwav_read_pcm_frames_s32__msadpcm_ima(pWav, framesToRead, pBufferOut);
+    }
+
+    if (pWav->translatedFormatTag == DR_WAVE_FORMAT_IEEE_FLOAT) {
+        return drwav_read_pcm_frames_s32__ieee(pWav, framesToRead, pBufferOut);
+    }
+
+    if (pWav->translatedFormatTag == DR_WAVE_FORMAT_ALAW) {
+        return drwav_read_pcm_frames_s32__alaw(pWav, framesToRead, pBufferOut);
+    }
+
+    if (pWav->translatedFormatTag == DR_WAVE_FORMAT_MULAW) {
+        return drwav_read_pcm_frames_s32__mulaw(pWav, framesToRead, pBufferOut);
+    }
+
+    return 0;
+}
+
+DRWAV_API drwav_uint64 drwav_read_pcm_frames_s32le(drwav* pWav, drwav_uint64 framesToRead, drwav_int32* pBufferOut)
+{
+    drwav_uint64 framesRead = drwav_read_pcm_frames_s32(pWav, framesToRead, pBufferOut);
+    if (pBufferOut != NULL && drwav__is_little_endian() == DRWAV_FALSE) {
+        drwav__bswap_samples_s32(pBufferOut, framesRead*pWav->channels);
+    }
+
+    return framesRead;
+}
+
+DRWAV_API drwav_uint64 drwav_read_pcm_frames_s32be(drwav* pWav, drwav_uint64 framesToRead, drwav_int32* pBufferOut)
+{
+    drwav_uint64 framesRead = drwav_read_pcm_frames_s32(pWav, framesToRead, pBufferOut);
+    if (pBufferOut != NULL && drwav__is_little_endian() == DRWAV_TRUE) {
+        drwav__bswap_samples_s32(pBufferOut, framesRead*pWav->channels);
+    }
+
+    return framesRead;
+}
+
+
+DRWAV_API void drwav_u8_to_s32(drwav_int32* pOut, const drwav_uint8* pIn, size_t sampleCount)
+{
+    size_t i;
+
+    if (pOut == NULL || pIn == NULL) {
+        return;
+    }
+
+    for (i = 0; i < sampleCount; ++i) {
+        *pOut++ = ((int)pIn[i] - 128) << 24;
+    }
+}
+
+DRWAV_API void drwav_s16_to_s32(drwav_int32* pOut, const drwav_int16* pIn, size_t sampleCount)
+{
+    size_t i;
+
+    if (pOut == NULL || pIn == NULL) {
+        return;
+    }
+
+    for (i = 0; i < sampleCount; ++i) {
+        *pOut++ = pIn[i] << 16;
+    }
+}
+
+DRWAV_API void drwav_s24_to_s32(drwav_int32* pOut, const drwav_uint8* pIn, size_t sampleCount)
+{
+    size_t i;
+
+    if (pOut == NULL || pIn == NULL) {
+        return;
+    }
+
+    for (i = 0; i < sampleCount; ++i) {
+        unsigned int s0 = pIn[i*3 + 0];
+        unsigned int s1 = pIn[i*3 + 1];
+        unsigned int s2 = pIn[i*3 + 2];
+
+        drwav_int32 sample32 = (drwav_int32)((s0 << 8) | (s1 << 16) | (s2 << 24));
+        *pOut++ = sample32;
+    }
+}
+
+DRWAV_API void drwav_f32_to_s32(drwav_int32* pOut, const float* pIn, size_t sampleCount)
+{
+    size_t i;
+
+    if (pOut == NULL || pIn == NULL) {
+        return;
+    }
+
+    for (i = 0; i < sampleCount; ++i) {
+        *pOut++ = (drwav_int32)(2147483648.0f * pIn[i]);
+    }
+}
+
+DRWAV_API void drwav_f64_to_s32(drwav_int32* pOut, const double* pIn, size_t sampleCount)
+{
+    size_t i;
+
+    if (pOut == NULL || pIn == NULL) {
+        return;
+    }
+
+    for (i = 0; i < sampleCount; ++i) {
+        *pOut++ = (drwav_int32)(2147483648.0 * pIn[i]);
+    }
+}
+
+DRWAV_API void drwav_alaw_to_s32(drwav_int32* pOut, const drwav_uint8* pIn, size_t sampleCount)
+{
+    size_t i;
+
+    if (pOut == NULL || pIn == NULL) {
+        return;
+    }
+
+    for (i = 0; i < sampleCount; ++i) {
+        *pOut++ = ((drwav_int32)drwav__alaw_to_s16(pIn[i])) << 16;
+    }
+}
+
+DRWAV_API void drwav_mulaw_to_s32(drwav_int32* pOut, const drwav_uint8* pIn, size_t sampleCount)
+{
+    size_t i;
+
+    if (pOut == NULL || pIn == NULL) {
+        return;
+    }
+
+    for (i= 0; i < sampleCount; ++i) {
+        *pOut++ = ((drwav_int32)drwav__mulaw_to_s16(pIn[i])) << 16;
+    }
+}
+
+
+
+DRWAV_PRIVATE drwav_int16* drwav__read_pcm_frames_and_close_s16(drwav* pWav, unsigned int* channels, unsigned int* sampleRate, drwav_uint64* totalFrameCount)
+{
+    drwav_uint64 sampleDataSize;
+    drwav_int16* pSampleData;
+    drwav_uint64 framesRead;
+
+    DRWAV_ASSERT(pWav != NULL);
+
+    sampleDataSize = pWav->totalPCMFrameCount * pWav->channels * sizeof(drwav_int16);
+    if (sampleDataSize > DRWAV_SIZE_MAX) {
+        drwav_uninit(pWav);
+        return NULL;    /* File's too big. */
+    }
+
+    pSampleData = (drwav_int16*)drwav__malloc_from_callbacks((size_t)sampleDataSize, &pWav->allocationCallbacks); /* <-- Safe cast due to the check above. */
+    if (pSampleData == NULL) {
+        drwav_uninit(pWav);
+        return NULL;    /* Failed to allocate memory. */
+    }
+
+    framesRead = drwav_read_pcm_frames_s16(pWav, (size_t)pWav->totalPCMFrameCount, pSampleData);
+    if (framesRead != pWav->totalPCMFrameCount) {
+        drwav__free_from_callbacks(pSampleData, &pWav->allocationCallbacks);
+        drwav_uninit(pWav);
+        return NULL;    /* There was an error reading the samples. */
+    }
+
+    drwav_uninit(pWav);
+
+    if (sampleRate) {
+        *sampleRate = pWav->sampleRate;
+    }
+    if (channels) {
+        *channels = pWav->channels;
+    }
+    if (totalFrameCount) {
+        *totalFrameCount = pWav->totalPCMFrameCount;
+    }
+
+    return pSampleData;
+}
+
+DRWAV_PRIVATE float* drwav__read_pcm_frames_and_close_f32(drwav* pWav, unsigned int* channels, unsigned int* sampleRate, drwav_uint64* totalFrameCount)
+{
+    drwav_uint64 sampleDataSize;
+    float* pSampleData;
+    drwav_uint64 framesRead;
+
+    DRWAV_ASSERT(pWav != NULL);
+
+    sampleDataSize = pWav->totalPCMFrameCount * pWav->channels * sizeof(float);
+    if (sampleDataSize > DRWAV_SIZE_MAX) {
+        drwav_uninit(pWav);
+        return NULL;    /* File's too big. */
+    }
+
+    pSampleData = (float*)drwav__malloc_from_callbacks((size_t)sampleDataSize, &pWav->allocationCallbacks); /* <-- Safe cast due to the check above. */
+    if (pSampleData == NULL) {
+        drwav_uninit(pWav);
+        return NULL;    /* Failed to allocate memory. */
+    }
+
+    framesRead = drwav_read_pcm_frames_f32(pWav, (size_t)pWav->totalPCMFrameCount, pSampleData);
+    if (framesRead != pWav->totalPCMFrameCount) {
+        drwav__free_from_callbacks(pSampleData, &pWav->allocationCallbacks);
+        drwav_uninit(pWav);
+        return NULL;    /* There was an error reading the samples. */
+    }
+
+    drwav_uninit(pWav);
+
+    if (sampleRate) {
+        *sampleRate = pWav->sampleRate;
+    }
+    if (channels) {
+        *channels = pWav->channels;
+    }
+    if (totalFrameCount) {
+        *totalFrameCount = pWav->totalPCMFrameCount;
+    }
+
+    return pSampleData;
+}
+
+DRWAV_PRIVATE drwav_int32* drwav__read_pcm_frames_and_close_s32(drwav* pWav, unsigned int* channels, unsigned int* sampleRate, drwav_uint64* totalFrameCount)
+{
+    drwav_uint64 sampleDataSize;
+    drwav_int32* pSampleData;
+    drwav_uint64 framesRead;
+
+    DRWAV_ASSERT(pWav != NULL);
+
+    sampleDataSize = pWav->totalPCMFrameCount * pWav->channels * sizeof(drwav_int32);
+    if (sampleDataSize > DRWAV_SIZE_MAX) {
+        drwav_uninit(pWav);
+        return NULL;    /* File's too big. */
+    }
+
+    pSampleData = (drwav_int32*)drwav__malloc_from_callbacks((size_t)sampleDataSize, &pWav->allocationCallbacks); /* <-- Safe cast due to the check above. */
+    if (pSampleData == NULL) {
+        drwav_uninit(pWav);
+        return NULL;    /* Failed to allocate memory. */
+    }
+
+    framesRead = drwav_read_pcm_frames_s32(pWav, (size_t)pWav->totalPCMFrameCount, pSampleData);
+    if (framesRead != pWav->totalPCMFrameCount) {
+        drwav__free_from_callbacks(pSampleData, &pWav->allocationCallbacks);
+        drwav_uninit(pWav);
+        return NULL;    /* There was an error reading the samples. */
+    }
+
+    drwav_uninit(pWav);
+
+    if (sampleRate) {
+        *sampleRate = pWav->sampleRate;
+    }
+    if (channels) {
+        *channels = pWav->channels;
+    }
+    if (totalFrameCount) {
+        *totalFrameCount = pWav->totalPCMFrameCount;
+    }
+
+    return pSampleData;
+}
+
+
+
+DRWAV_API drwav_int16* drwav_open_and_read_pcm_frames_s16(drwav_read_proc onRead, drwav_seek_proc onSeek, void* pUserData, unsigned int* channelsOut, unsigned int* sampleRateOut, drwav_uint64* totalFrameCountOut, const drwav_allocation_callbacks* pAllocationCallbacks)
+{
+    drwav wav;
+
+    if (channelsOut) {
+        *channelsOut = 0;
+    }
+    if (sampleRateOut) {
+        *sampleRateOut = 0;
+    }
+    if (totalFrameCountOut) {
+        *totalFrameCountOut = 0;
+    }
+
+    if (!drwav_init(&wav, onRead, onSeek, pUserData, pAllocationCallbacks)) {
+        return NULL;
+    }
+
+    return drwav__read_pcm_frames_and_close_s16(&wav, channelsOut, sampleRateOut, totalFrameCountOut);
+}
+
+DRWAV_API float* drwav_open_and_read_pcm_frames_f32(drwav_read_proc onRead, drwav_seek_proc onSeek, void* pUserData, unsigned int* channelsOut, unsigned int* sampleRateOut, drwav_uint64* totalFrameCountOut, const drwav_allocation_callbacks* pAllocationCallbacks)
+{
+    drwav wav;
+
+    if (channelsOut) {
+        *channelsOut = 0;
+    }
+    if (sampleRateOut) {
+        *sampleRateOut = 0;
+    }
+    if (totalFrameCountOut) {
+        *totalFrameCountOut = 0;
+    }
+
+    if (!drwav_init(&wav, onRead, onSeek, pUserData, pAllocationCallbacks)) {
+        return NULL;
+    }
+
+    return drwav__read_pcm_frames_and_close_f32(&wav, channelsOut, sampleRateOut, totalFrameCountOut);
+}
+
+DRWAV_API drwav_int32* drwav_open_and_read_pcm_frames_s32(drwav_read_proc onRead, drwav_seek_proc onSeek, void* pUserData, unsigned int* channelsOut, unsigned int* sampleRateOut, drwav_uint64* totalFrameCountOut, const drwav_allocation_callbacks* pAllocationCallbacks)
+{
+    drwav wav;
+
+    if (channelsOut) {
+        *channelsOut = 0;
+    }
+    if (sampleRateOut) {
+        *sampleRateOut = 0;
+    }
+    if (totalFrameCountOut) {
+        *totalFrameCountOut = 0;
+    }
+
+    if (!drwav_init(&wav, onRead, onSeek, pUserData, pAllocationCallbacks)) {
+        return NULL;
+    }
+
+    return drwav__read_pcm_frames_and_close_s32(&wav, channelsOut, sampleRateOut, totalFrameCountOut);
+}
+
+#ifndef DR_WAV_NO_STDIO
+DRWAV_API drwav_int16* drwav_open_file_and_read_pcm_frames_s16(const char* filename, unsigned int* channelsOut, unsigned int* sampleRateOut, drwav_uint64* totalFrameCountOut, const drwav_allocation_callbacks* pAllocationCallbacks)
+{
+    drwav wav;
+
+    if (channelsOut) {
+        *channelsOut = 0;
+    }
+    if (sampleRateOut) {
+        *sampleRateOut = 0;
+    }
+    if (totalFrameCountOut) {
+        *totalFrameCountOut = 0;
+    }
+
+    if (!drwav_init_file(&wav, filename, pAllocationCallbacks)) {
+        return NULL;
+    }
+
+    return drwav__read_pcm_frames_and_close_s16(&wav, channelsOut, sampleRateOut, totalFrameCountOut);
+}
+
+DRWAV_API float* drwav_open_file_and_read_pcm_frames_f32(const char* filename, unsigned int* channelsOut, unsigned int* sampleRateOut, drwav_uint64* totalFrameCountOut, const drwav_allocation_callbacks* pAllocationCallbacks)
+{
+    drwav wav;
+
+    if (channelsOut) {
+        *channelsOut = 0;
+    }
+    if (sampleRateOut) {
+        *sampleRateOut = 0;
+    }
+    if (totalFrameCountOut) {
+        *totalFrameCountOut = 0;
+    }
+
+    if (!drwav_init_file(&wav, filename, pAllocationCallbacks)) {
+        return NULL;
+    }
+
+    return drwav__read_pcm_frames_and_close_f32(&wav, channelsOut, sampleRateOut, totalFrameCountOut);
+}
+
+DRWAV_API drwav_int32* drwav_open_file_and_read_pcm_frames_s32(const char* filename, unsigned int* channelsOut, unsigned int* sampleRateOut, drwav_uint64* totalFrameCountOut, const drwav_allocation_callbacks* pAllocationCallbacks)
+{
+    drwav wav;
+
+    if (channelsOut) {
+        *channelsOut = 0;
+    }
+    if (sampleRateOut) {
+        *sampleRateOut = 0;
+    }
+    if (totalFrameCountOut) {
+        *totalFrameCountOut = 0;
+    }
+
+    if (!drwav_init_file(&wav, filename, pAllocationCallbacks)) {
+        return NULL;
+    }
+
+    return drwav__read_pcm_frames_and_close_s32(&wav, channelsOut, sampleRateOut, totalFrameCountOut);
+}
+
+
+#ifndef DR_WAV_NO_WCHAR
+DRWAV_API drwav_int16* drwav_open_file_and_read_pcm_frames_s16_w(const wchar_t* filename, unsigned int* channelsOut, unsigned int* sampleRateOut, drwav_uint64* totalFrameCountOut, const drwav_allocation_callbacks* pAllocationCallbacks)
+{
+    drwav wav;
+
+    if (sampleRateOut) {
+        *sampleRateOut = 0;
+    }
+    if (channelsOut) {
+        *channelsOut = 0;
+    }
+    if (totalFrameCountOut) {
+        *totalFrameCountOut = 0;
+    }
+
+    if (!drwav_init_file_w(&wav, filename, pAllocationCallbacks)) {
+        return NULL;
+    }
+
+    return drwav__read_pcm_frames_and_close_s16(&wav, channelsOut, sampleRateOut, totalFrameCountOut);
+}
+
+DRWAV_API float* drwav_open_file_and_read_pcm_frames_f32_w(const wchar_t* filename, unsigned int* channelsOut, unsigned int* sampleRateOut, drwav_uint64* totalFrameCountOut, const drwav_allocation_callbacks* pAllocationCallbacks)
+{
+    drwav wav;
+
+    if (sampleRateOut) {
+        *sampleRateOut = 0;
+    }
+    if (channelsOut) {
+        *channelsOut = 0;
+    }
+    if (totalFrameCountOut) {
+        *totalFrameCountOut = 0;
+    }
+
+    if (!drwav_init_file_w(&wav, filename, pAllocationCallbacks)) {
+        return NULL;
+    }
+
+    return drwav__read_pcm_frames_and_close_f32(&wav, channelsOut, sampleRateOut, totalFrameCountOut);
+}
+
+DRWAV_API drwav_int32* drwav_open_file_and_read_pcm_frames_s32_w(const wchar_t* filename, unsigned int* channelsOut, unsigned int* sampleRateOut, drwav_uint64* totalFrameCountOut, const drwav_allocation_callbacks* pAllocationCallbacks)
+{
+    drwav wav;
+
+    if (sampleRateOut) {
+        *sampleRateOut = 0;
+    }
+    if (channelsOut) {
+        *channelsOut = 0;
+    }
+    if (totalFrameCountOut) {
+        *totalFrameCountOut = 0;
+    }
+
+    if (!drwav_init_file_w(&wav, filename, pAllocationCallbacks)) {
+        return NULL;
+    }
+
+    return drwav__read_pcm_frames_and_close_s32(&wav, channelsOut, sampleRateOut, totalFrameCountOut);
+}
+#endif /* DR_WAV_NO_WCHAR */
+#endif /* DR_WAV_NO_STDIO */
+
+DRWAV_API drwav_int16* drwav_open_memory_and_read_pcm_frames_s16(const void* data, size_t dataSize, unsigned int* channelsOut, unsigned int* sampleRateOut, drwav_uint64* totalFrameCountOut, const drwav_allocation_callbacks* pAllocationCallbacks)
+{
+    drwav wav;
+
+    if (channelsOut) {
+        *channelsOut = 0;
+    }
+    if (sampleRateOut) {
+        *sampleRateOut = 0;
+    }
+    if (totalFrameCountOut) {
+        *totalFrameCountOut = 0;
+    }
+
+    if (!drwav_init_memory(&wav, data, dataSize, pAllocationCallbacks)) {
+        return NULL;
+    }
+
+    return drwav__read_pcm_frames_and_close_s16(&wav, channelsOut, sampleRateOut, totalFrameCountOut);
+}
+
+DRWAV_API float* drwav_open_memory_and_read_pcm_frames_f32(const void* data, size_t dataSize, unsigned int* channelsOut, unsigned int* sampleRateOut, drwav_uint64* totalFrameCountOut, const drwav_allocation_callbacks* pAllocationCallbacks)
+{
+    drwav wav;
+
+    if (channelsOut) {
+        *channelsOut = 0;
+    }
+    if (sampleRateOut) {
+        *sampleRateOut = 0;
+    }
+    if (totalFrameCountOut) {
+        *totalFrameCountOut = 0;
+    }
+
+    if (!drwav_init_memory(&wav, data, dataSize, pAllocationCallbacks)) {
+        return NULL;
+    }
+
+    return drwav__read_pcm_frames_and_close_f32(&wav, channelsOut, sampleRateOut, totalFrameCountOut);
+}
+
+DRWAV_API drwav_int32* drwav_open_memory_and_read_pcm_frames_s32(const void* data, size_t dataSize, unsigned int* channelsOut, unsigned int* sampleRateOut, drwav_uint64* totalFrameCountOut, const drwav_allocation_callbacks* pAllocationCallbacks)
+{
+    drwav wav;
+
+    if (channelsOut) {
+        *channelsOut = 0;
+    }
+    if (sampleRateOut) {
+        *sampleRateOut = 0;
+    }
+    if (totalFrameCountOut) {
+        *totalFrameCountOut = 0;
+    }
+
+    if (!drwav_init_memory(&wav, data, dataSize, pAllocationCallbacks)) {
+        return NULL;
+    }
+
+    return drwav__read_pcm_frames_and_close_s32(&wav, channelsOut, sampleRateOut, totalFrameCountOut);
+}
+#endif  /* DR_WAV_NO_CONVERSION_API */
+
+
+DRWAV_API void drwav_free(void* p, const drwav_allocation_callbacks* pAllocationCallbacks)
+{
+    if (pAllocationCallbacks != NULL) {
+        drwav__free_from_callbacks(p, pAllocationCallbacks);
+    } else {
+        drwav__free_default(p, NULL);
+    }
+}
+
+DRWAV_API drwav_uint16 drwav_bytes_to_u16(const drwav_uint8* data)
+{
+    return ((drwav_uint16)data[0] << 0) | ((drwav_uint16)data[1] << 8);
+}
+
+DRWAV_API drwav_int16 drwav_bytes_to_s16(const drwav_uint8* data)
+{
+    return (drwav_int16)drwav_bytes_to_u16(data);
+}
+
+DRWAV_API drwav_uint32 drwav_bytes_to_u32(const drwav_uint8* data)
+{
+    return drwav_bytes_to_u32_le(data);
+}
+
+DRWAV_API float drwav_bytes_to_f32(const drwav_uint8* data)
+{
+    union {
+        drwav_uint32 u32;
+        float f32;
+    } value;
+
+    value.u32 = drwav_bytes_to_u32(data);
+    return value.f32;
+}
+
+DRWAV_API drwav_int32 drwav_bytes_to_s32(const drwav_uint8* data)
+{
+    return (drwav_int32)drwav_bytes_to_u32(data);
+}
+
+DRWAV_API drwav_uint64 drwav_bytes_to_u64(const drwav_uint8* data)
+{
+    return
+        ((drwav_uint64)data[0] <<  0) | ((drwav_uint64)data[1] <<  8) | ((drwav_uint64)data[2] << 16) | ((drwav_uint64)data[3] << 24) |
+        ((drwav_uint64)data[4] << 32) | ((drwav_uint64)data[5] << 40) | ((drwav_uint64)data[6] << 48) | ((drwav_uint64)data[7] << 56);
+}
+
+DRWAV_API drwav_int64 drwav_bytes_to_s64(const drwav_uint8* data)
+{
+    return (drwav_int64)drwav_bytes_to_u64(data);
+}
+
+
+DRWAV_API drwav_bool32 drwav_guid_equal(const drwav_uint8 a[16], const drwav_uint8 b[16])
+{
+    int i;
+    for (i = 0; i < 16; i += 1) {
+        if (a[i] != b[i]) {
+            return DRWAV_FALSE;
+        }
+    }
+
+    return DRWAV_TRUE;
+}
+
+DRWAV_API drwav_bool32 drwav_fourcc_equal(const drwav_uint8* a, const char* b)
+{
+    return
+        a[0] == b[0] &&
+        a[1] == b[1] &&
+        a[2] == b[2] &&
+        a[3] == b[3];
+}
+
+#ifdef __MRC__
+/* Undo the pragma at the beginning of this file. */
+#pragma options opt reset
+#endif
+
+#endif  /* dr_wav_c */
+#endif  /* DR_WAV_IMPLEMENTATION */
+
+/*
+REVISION HISTORY
+================
+v0.13.16 - 2024-02-27
+  - Fix a Wdouble-promotion warning.
+
+v0.13.15 - 2024-01-23
+  - Relax some unnecessary validation that prevented some files from loading.
+
+v0.13.14 - 2023-12-02
+  - Fix a warning about an unused variable.
+
+v0.13.13 - 2023-11-02
+  - Fix a warning when compiling with Clang.
+
+v0.13.12 - 2023-08-07
+  - Fix a possible crash in drwav_read_pcm_frames().
+
+v0.13.11 - 2023-07-07
+  - AIFF compatibility improvements.
+
+v0.13.10 - 2023-05-29
+  - Fix a bug where drwav_init_with_metadata() does not decode any frames after initializtion.
+
+v0.13.9 - 2023-05-22
+  - Add support for AIFF decoding (writing and metadata not supported).
+  - Add support for RIFX decoding (writing and metadata not supported).
+  - Fix a bug where metadata is not processed if it's located before the "fmt " chunk.
+  - Add a workaround for a type of malformed WAV file where the size of the "RIFF" and "data" chunks
+    are incorrectly set to 0xFFFFFFFF.
+
+v0.13.8 - 2023-03-25
+  - Fix a possible null pointer dereference.
+  - Fix a crash when loading files with badly formed metadata.
+
+v0.13.7 - 2022-09-17
+  - Fix compilation with DJGPP.
+  - Add support for disabling wchar_t with DR_WAV_NO_WCHAR.
+
+v0.13.6 - 2022-04-10
+  - Fix compilation error on older versions of GCC.
+  - Remove some dependencies on the standard library.
+
+v0.13.5 - 2022-01-26
+  - Fix an error when seeking to the end of the file.
+
+v0.13.4 - 2021-12-08
+  - Fix some static analysis warnings.
+
+v0.13.3 - 2021-11-24
+  - Fix an incorrect assertion when trying to endian swap 1-byte sample formats. This is now a no-op
+    rather than a failed assertion.
+  - Fix a bug with parsing of the bext chunk.
+  - Fix some static analysis warnings.
+
+v0.13.2 - 2021-10-02
+  - Fix a possible buffer overflow when reading from compressed formats.
+
+v0.13.1 - 2021-07-31
+  - Fix platform detection for ARM64.
+
+v0.13.0 - 2021-07-01
+  - Improve support for reading and writing metadata. Use the `_with_metadata()` APIs to initialize
+    a WAV decoder and store the metadata within the `drwav` object. Use the `pMetadata` and
+    `metadataCount` members of the `drwav` object to read the data. The old way of handling metadata
+    via a callback is still usable and valid.
+  - API CHANGE: drwav_target_write_size_bytes() now takes extra parameters for calculating the
+    required write size when writing metadata.
+  - Add drwav_get_cursor_in_pcm_frames()
+  - Add drwav_get_length_in_pcm_frames()
+  - Fix a bug where drwav_read_raw() can call the read callback with a byte count of zero.
+
+v0.12.20 - 2021-06-11
+  - Fix some undefined behavior.
+
+v0.12.19 - 2021-02-21
+  - Fix a warning due to referencing _MSC_VER when it is undefined.
+  - Minor improvements to the management of some internal state concerning the data chunk cursor.
+
+v0.12.18 - 2021-01-31
+  - Clean up some static analysis warnings.
+
+v0.12.17 - 2021-01-17
+  - Minor fix to sample code in documentation.
+  - Correctly qualify a private API as private rather than public.
+  - Code cleanup.
+
+v0.12.16 - 2020-12-02
+  - Fix a bug when trying to read more bytes than can fit in a size_t.
+
+v0.12.15 - 2020-11-21
+  - Fix compilation with OpenWatcom.
+
+v0.12.14 - 2020-11-13
+  - Minor code clean up.
+
+v0.12.13 - 2020-11-01
+  - Improve compiler support for older versions of GCC.
+
+v0.12.12 - 2020-09-28
+  - Add support for RF64.
+  - Fix a bug in writing mode where the size of the RIFF chunk incorrectly includes the header section.
+
+v0.12.11 - 2020-09-08
+  - Fix a compilation error on older compilers.
+
+v0.12.10 - 2020-08-24
+  - Fix a bug when seeking with ADPCM formats.
+
+v0.12.9 - 2020-08-02
+  - Simplify sized types.
+
+v0.12.8 - 2020-07-25
+  - Fix a compilation warning.
+
+v0.12.7 - 2020-07-15
+  - Fix some bugs on big-endian architectures.
+  - Fix an error in s24 to f32 conversion.
+
+v0.12.6 - 2020-06-23
+  - Change drwav_read_*() to allow NULL to be passed in as the output buffer which is equivalent to a forward seek.
+  - Fix a buffer overflow when trying to decode invalid IMA-ADPCM files.
+  - Add include guard for the implementation section.
+
+v0.12.5 - 2020-05-27
+  - Minor documentation fix.
+
+v0.12.4 - 2020-05-16
+  - Replace assert() with DRWAV_ASSERT().
+  - Add compile-time and run-time version querying.
+    - DRWAV_VERSION_MINOR
+    - DRWAV_VERSION_MAJOR
+    - DRWAV_VERSION_REVISION
+    - DRWAV_VERSION_STRING
+    - drwav_version()
+    - drwav_version_string()
+
+v0.12.3 - 2020-04-30
+  - Fix compilation errors with VC6.
+
+v0.12.2 - 2020-04-21
+  - Fix a bug where drwav_init_file() does not close the file handle after attempting to load an erroneous file.
+
+v0.12.1 - 2020-04-13
+  - Fix some pedantic warnings.
+
+v0.12.0 - 2020-04-04
+  - API CHANGE: Add container and format parameters to the chunk callback.
+  - Minor documentation updates.
+
+v0.11.5 - 2020-03-07
+  - Fix compilation error with Visual Studio .NET 2003.
+
+v0.11.4 - 2020-01-29
+  - Fix some static analysis warnings.
+  - Fix a bug when reading f32 samples from an A-law encoded stream.
+
+v0.11.3 - 2020-01-12
+  - Minor changes to some f32 format conversion routines.
+  - Minor bug fix for ADPCM conversion when end of file is reached.
+
+v0.11.2 - 2019-12-02
+  - Fix a possible crash when using custom memory allocators without a custom realloc() implementation.
+  - Fix an integer overflow bug.
+  - Fix a null pointer dereference bug.
+  - Add limits to sample rate, channels and bits per sample to tighten up some validation.
+
+v0.11.1 - 2019-10-07
+  - Internal code clean up.
+
+v0.11.0 - 2019-10-06
+  - API CHANGE: Add support for user defined memory allocation routines. This system allows the program to specify their own memory allocation
+    routines with a user data pointer for client-specific contextual data. This adds an extra parameter to the end of the following APIs:
+    - drwav_init()
+    - drwav_init_ex()
+    - drwav_init_file()
+    - drwav_init_file_ex()
+    - drwav_init_file_w()
+    - drwav_init_file_w_ex()
+    - drwav_init_memory()
+    - drwav_init_memory_ex()
+    - drwav_init_write()
+    - drwav_init_write_sequential()
+    - drwav_init_write_sequential_pcm_frames()
+    - drwav_init_file_write()
+    - drwav_init_file_write_sequential()
+    - drwav_init_file_write_sequential_pcm_frames()
+    - drwav_init_file_write_w()
+    - drwav_init_file_write_sequential_w()
+    - drwav_init_file_write_sequential_pcm_frames_w()
+    - drwav_init_memory_write()
+    - drwav_init_memory_write_sequential()
+    - drwav_init_memory_write_sequential_pcm_frames()
+    - drwav_open_and_read_pcm_frames_s16()
+    - drwav_open_and_read_pcm_frames_f32()
+    - drwav_open_and_read_pcm_frames_s32()
+    - drwav_open_file_and_read_pcm_frames_s16()
+    - drwav_open_file_and_read_pcm_frames_f32()
+    - drwav_open_file_and_read_pcm_frames_s32()
+    - drwav_open_file_and_read_pcm_frames_s16_w()
+    - drwav_open_file_and_read_pcm_frames_f32_w()
+    - drwav_open_file_and_read_pcm_frames_s32_w()
+    - drwav_open_memory_and_read_pcm_frames_s16()
+    - drwav_open_memory_and_read_pcm_frames_f32()
+    - drwav_open_memory_and_read_pcm_frames_s32()
+    Set this extra parameter to NULL to use defaults which is the same as the previous behaviour. Setting this NULL will use
+    DRWAV_MALLOC, DRWAV_REALLOC and DRWAV_FREE.
+  - Add support for reading and writing PCM frames in an explicit endianness. New APIs:
+    - drwav_read_pcm_frames_le()
+    - drwav_read_pcm_frames_be()
+    - drwav_read_pcm_frames_s16le()
+    - drwav_read_pcm_frames_s16be()
+    - drwav_read_pcm_frames_f32le()
+    - drwav_read_pcm_frames_f32be()
+    - drwav_read_pcm_frames_s32le()
+    - drwav_read_pcm_frames_s32be()
+    - drwav_write_pcm_frames_le()
+    - drwav_write_pcm_frames_be()
+  - Remove deprecated APIs.
+  - API CHANGE: The following APIs now return native-endian data. Previously they returned little-endian data.
+    - drwav_read_pcm_frames()
+    - drwav_read_pcm_frames_s16()
+    - drwav_read_pcm_frames_s32()
+    - drwav_read_pcm_frames_f32()
+    - drwav_open_and_read_pcm_frames_s16()
+    - drwav_open_and_read_pcm_frames_s32()
+    - drwav_open_and_read_pcm_frames_f32()
+    - drwav_open_file_and_read_pcm_frames_s16()
+    - drwav_open_file_and_read_pcm_frames_s32()
+    - drwav_open_file_and_read_pcm_frames_f32()
+    - drwav_open_file_and_read_pcm_frames_s16_w()
+    - drwav_open_file_and_read_pcm_frames_s32_w()
+    - drwav_open_file_and_read_pcm_frames_f32_w()
+    - drwav_open_memory_and_read_pcm_frames_s16()
+    - drwav_open_memory_and_read_pcm_frames_s32()
+    - drwav_open_memory_and_read_pcm_frames_f32()
+
+v0.10.1 - 2019-08-31
+  - Correctly handle partial trailing ADPCM blocks.
+
+v0.10.0 - 2019-08-04
+  - Remove deprecated APIs.
+  - Add wchar_t variants for file loading APIs:
+      drwav_init_file_w()
+      drwav_init_file_ex_w()
+      drwav_init_file_write_w()
+      drwav_init_file_write_sequential_w()
+  - Add drwav_target_write_size_bytes() which calculates the total size in bytes of a WAV file given a format and sample count.
+  - Add APIs for specifying the PCM frame count instead of the sample count when opening in sequential write mode:
+      drwav_init_write_sequential_pcm_frames()
+      drwav_init_file_write_sequential_pcm_frames()
+      drwav_init_file_write_sequential_pcm_frames_w()
+      drwav_init_memory_write_sequential_pcm_frames()
+  - Deprecate drwav_open*() and drwav_close():
+      drwav_open()
+      drwav_open_ex()
+      drwav_open_write()
+      drwav_open_write_sequential()
+      drwav_open_file()
+      drwav_open_file_ex()
+      drwav_open_file_write()
+      drwav_open_file_write_sequential()
+      drwav_open_memory()
+      drwav_open_memory_ex()
+      drwav_open_memory_write()
+      drwav_open_memory_write_sequential()
+      drwav_close()
+  - Minor documentation updates.
+
+v0.9.2 - 2019-05-21
+  - Fix warnings.
+
+v0.9.1 - 2019-05-05
+  - Add support for C89.
+  - Change license to choice of public domain or MIT-0.
+
+v0.9.0 - 2018-12-16
+  - API CHANGE: Add new reading APIs for reading by PCM frames instead of samples. Old APIs have been deprecated and
+    will be removed in v0.10.0. Deprecated APIs and their replacements:
+      drwav_read()                     -> drwav_read_pcm_frames()
+      drwav_read_s16()                 -> drwav_read_pcm_frames_s16()
+      drwav_read_f32()                 -> drwav_read_pcm_frames_f32()
+      drwav_read_s32()                 -> drwav_read_pcm_frames_s32()
+      drwav_seek_to_sample()           -> drwav_seek_to_pcm_frame()
+      drwav_write()                    -> drwav_write_pcm_frames()
+      drwav_open_and_read_s16()        -> drwav_open_and_read_pcm_frames_s16()
+      drwav_open_and_read_f32()        -> drwav_open_and_read_pcm_frames_f32()
+      drwav_open_and_read_s32()        -> drwav_open_and_read_pcm_frames_s32()
+      drwav_open_file_and_read_s16()   -> drwav_open_file_and_read_pcm_frames_s16()
+      drwav_open_file_and_read_f32()   -> drwav_open_file_and_read_pcm_frames_f32()
+      drwav_open_file_and_read_s32()   -> drwav_open_file_and_read_pcm_frames_s32()
+      drwav_open_memory_and_read_s16() -> drwav_open_memory_and_read_pcm_frames_s16()
+      drwav_open_memory_and_read_f32() -> drwav_open_memory_and_read_pcm_frames_f32()
+      drwav_open_memory_and_read_s32() -> drwav_open_memory_and_read_pcm_frames_s32()
+      drwav::totalSampleCount          -> drwav::totalPCMFrameCount
+  - API CHANGE: Rename drwav_open_and_read_file_*() to drwav_open_file_and_read_*().
+  - API CHANGE: Rename drwav_open_and_read_memory_*() to drwav_open_memory_and_read_*().
+  - Add built-in support for smpl chunks.
+  - Add support for firing a callback for each chunk in the file at initialization time.
+    - This is enabled through the drwav_init_ex(), etc. family of APIs.
+  - Handle invalid FMT chunks more robustly.
+
+v0.8.5 - 2018-09-11
+  - Const correctness.
+  - Fix a potential stack overflow.
+
+v0.8.4 - 2018-08-07
+  - Improve 64-bit detection.
+
+v0.8.3 - 2018-08-05
+  - Fix C++ build on older versions of GCC.
+
+v0.8.2 - 2018-08-02
+  - Fix some big-endian bugs.
+
+v0.8.1 - 2018-06-29
+  - Add support for sequential writing APIs.
+  - Disable seeking in write mode.
+  - Fix bugs with Wave64.
+  - Fix typos.
+
+v0.8 - 2018-04-27
+  - Bug fix.
+  - Start using major.minor.revision versioning.
+
+v0.7f - 2018-02-05
+  - Restrict ADPCM formats to a maximum of 2 channels.
+
+v0.7e - 2018-02-02
+  - Fix a crash.
+
+v0.7d - 2018-02-01
+  - Fix a crash.
+
+v0.7c - 2018-02-01
+  - Set drwav.bytesPerSample to 0 for all compressed formats.
+  - Fix a crash when reading 16-bit floating point WAV files. In this case dr_wav will output silence for
+    all format conversion reading APIs (*_s16, *_s32, *_f32 APIs).
+  - Fix some divide-by-zero errors.
+
+v0.7b - 2018-01-22
+  - Fix errors with seeking of compressed formats.
+  - Fix compilation error when DR_WAV_NO_CONVERSION_API
+
+v0.7a - 2017-11-17
+  - Fix some GCC warnings.
+
+v0.7 - 2017-11-04
+  - Add writing APIs.
+
+v0.6 - 2017-08-16
+  - API CHANGE: Rename dr_* types to drwav_*.
+  - Add support for custom implementations of malloc(), realloc(), etc.
+  - Add support for Microsoft ADPCM.
+  - Add support for IMA ADPCM (DVI, format code 0x11).
+  - Optimizations to drwav_read_s16().
+  - Bug fixes.
+
+v0.5g - 2017-07-16
+  - Change underlying type for booleans to unsigned.
+
+v0.5f - 2017-04-04
+  - Fix a minor bug with drwav_open_and_read_s16() and family.
+
+v0.5e - 2016-12-29
+  - Added support for reading samples as signed 16-bit integers. Use the _s16() family of APIs for this.
+  - Minor fixes to documentation.
+
+v0.5d - 2016-12-28
+  - Use drwav_int* and drwav_uint* sized types to improve compiler support.
+
+v0.5c - 2016-11-11
+  - Properly handle JUNK chunks that come before the FMT chunk.
+
+v0.5b - 2016-10-23
+  - A minor change to drwav_bool8 and drwav_bool32 types.
+
+v0.5a - 2016-10-11
+  - Fixed a bug with drwav_open_and_read() and family due to incorrect argument ordering.
+  - Improve A-law and mu-law efficiency.
+
+v0.5 - 2016-09-29
+  - API CHANGE. Swap the order of "channels" and "sampleRate" parameters in drwav_open_and_read*(). Rationale for this is to
+    keep it consistent with dr_audio and dr_flac.
+
+v0.4b - 2016-09-18
+  - Fixed a typo in documentation.
+
+v0.4a - 2016-09-18
+  - Fixed a typo.
+  - Change date format to ISO 8601 (YYYY-MM-DD)
+
+v0.4 - 2016-07-13
+  - API CHANGE. Make onSeek consistent with dr_flac.
+  - API CHANGE. Rename drwav_seek() to drwav_seek_to_sample() for clarity and consistency with dr_flac.
+  - Added support for Sony Wave64.
+
+v0.3a - 2016-05-28
+  - API CHANGE. Return drwav_bool32 instead of int in onSeek callback.
+  - Fixed a memory leak.
+
+v0.3 - 2016-05-22
+  - Lots of API changes for consistency.
+
+v0.2a - 2016-05-16
+  - Fixed Linux/GCC build.
+
+v0.2 - 2016-05-11
+  - Added support for reading data as signed 32-bit PCM for consistency with dr_flac.
+
+v0.1a - 2016-05-07
+  - Fixed a bug in drwav_open_file() where the file handle would not be closed if the loader failed to initialize.
+
+v0.1 - 2016-05-04
+  - Initial versioned release.
+*/
+
+/*
+This software is available as a choice of the following licenses. Choose
+whichever you prefer.
+
+===============================================================================
+ALTERNATIVE 1 - Public Domain (www.unlicense.org)
+===============================================================================
+This is free and unencumbered software released into the public domain.
+
+Anyone is free to copy, modify, publish, use, compile, sell, or distribute this
+software, either in source code form or as a compiled binary, for any purpose,
+commercial or non-commercial, and by any means.
+
+In jurisdictions that recognize copyright laws, the author or authors of this
+software dedicate any and all copyright interest in the software to the public
+domain. We make this dedication for the benefit of the public at large and to
+the detriment of our heirs and successors. We intend this dedication to be an
+overt act of relinquishment in perpetuity of all present and future rights to
+this software under copyright law.
+
+THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+AUTHORS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
+ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION
+WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
+
+For more information, please refer to <http://unlicense.org/>
+
+===============================================================================
+ALTERNATIVE 2 - MIT No Attribution
+===============================================================================
+Copyright 2023 David Reid
+
+Permission is hereby granted, free of charge, to any person obtaining a copy of
+this software and associated documentation files (the "Software"), to deal in
+the Software without restriction, including without limitation the rights to
+use, copy, modify, merge, publish, distribute, sublicense, and/or sell copies
+of the Software, and to permit persons to whom the Software is furnished to do
+so.
+
+THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+SOFTWARE.
+*/
\ No newline at end of file
diff --git a/src/whisper_cpp/examples/grammar-parser.cpp b/src/whisper_cpp/examples/grammar-parser.cpp
new file mode 100644
index 00000000..e282fc7b
--- /dev/null
+++ b/src/whisper_cpp/examples/grammar-parser.cpp
@@ -0,0 +1,423 @@
+#include "grammar-parser.h"
+#include <cstdint>
+#include <cwchar>
+#include <string>
+#include <utility>
+#include <stdexcept>
+#include <exception>
+
+namespace grammar_parser {
+    // NOTE: assumes valid utf8 (but checks for overrun)
+    // copied from whisper.cpp
+    static std::pair<uint32_t, const char *> decode_utf8(const char * src) {
+        static const int lookup[] = { 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 2, 2, 3, 4 };
+        uint8_t  first_byte = static_cast<uint8_t>(*src);
+        uint8_t  highbits   = first_byte >> 4;
+        int      len        = lookup[highbits];
+        uint8_t  mask       = (1 << (8 - len)) - 1;
+        uint32_t value      = first_byte & mask;
+        const char * end    = src + len; // may overrun!
+        const char * pos    = src + 1;
+        for ( ; pos < end && *pos; pos++) {
+            value = (value << 6) + (static_cast<uint8_t>(*pos) & 0x3F);
+        }
+        return std::make_pair(value, pos);
+    }
+
+    static uint32_t get_symbol_id(parse_state & state, const char * src, size_t len) {
+        uint32_t next_id = static_cast<uint32_t>(state.symbol_ids.size());
+        auto result = state.symbol_ids.insert(std::make_pair(std::string(src, len), next_id));
+        return result.first->second;
+    }
+
+    static uint32_t generate_symbol_id(parse_state & state, const std::string & base_name) {
+        uint32_t next_id = static_cast<uint32_t>(state.symbol_ids.size());
+        state.symbol_ids[base_name + '_' + std::to_string(next_id)] = next_id;
+        return next_id;
+    }
+
+    static void add_rule(
+            parse_state & state,
+            uint32_t      rule_id,
+            const std::vector<whisper_grammar_element> & rule) {
+        if (state.rules.size() <= rule_id) {
+            state.rules.resize(rule_id + 1);
+        }
+        state.rules[rule_id] = rule;
+    }
+
+    static bool is_word_char(char c) {
+        return ('a' <= c && c <= 'z') || ('A' <= c && c <= 'Z') || c == '-' || ('0' <= c && c <= '9');
+    }
+
+    static std::pair<uint32_t, const char *> parse_hex(const char * src, int size) {
+        const char * pos   = src;
+        const char * end   = src + size;
+        uint32_t     value = 0;
+        for ( ; pos < end && *pos; pos++) {
+            value <<= 4;
+            char c = *pos;
+            if ('a' <= c && c <= 'f') {
+                value += c - 'a' + 10;
+            } else if ('A' <= c && c <= 'F') {
+                value += c - 'A' + 10;
+            } else if ('0' <= c && c <= '9') {
+                value += c - '0';
+            } else {
+                break;
+            }
+        }
+        if (pos != end) {
+            throw std::runtime_error("expecting " + std::to_string(size) + " hex chars at " + src);
+        }
+        return std::make_pair(value, pos);
+    }
+
+    static const char * parse_space(const char * src, bool newline_ok) {
+        const char * pos = src;
+        while (*pos == ' ' || *pos == '\t' || *pos == '#' ||
+                (newline_ok && (*pos == '\r' || *pos == '\n'))) {
+            if (*pos == '#') {
+                while (*pos && *pos != '\r' && *pos != '\n') {
+                    pos++;
+                }
+            } else {
+                pos++;
+            }
+        }
+        return pos;
+    }
+
+    static const char * parse_name(const char * src) {
+        const char * pos = src;
+        while (is_word_char(*pos)) {
+            pos++;
+        }
+        if (pos == src) {
+            throw std::runtime_error(std::string("expecting name at ") + src);
+        }
+        return pos;
+    }
+
+    static std::pair<uint32_t, const char *> parse_char(const char * src) {
+        if (*src == '\\') {
+            switch (src[1]) {
+                case 'x': return parse_hex(src + 2, 2);
+                case 'u': return parse_hex(src + 2, 4);
+                case 'U': return parse_hex(src + 2, 8);
+                case 't': return std::make_pair('\t', src + 2);
+                case 'r': return std::make_pair('\r', src + 2);
+                case 'n': return std::make_pair('\n', src + 2);
+                case '\\':
+                case '"':
+                case '[':
+                case ']':
+                    return std::make_pair(src[1], src + 2);
+                default:
+                    throw std::runtime_error(std::string("unknown escape at ") + src);
+            }
+        } else if (*src) {
+            return decode_utf8(src);
+        }
+        throw std::runtime_error("unexpected end of input");
+    }
+
+    static const char * parse_alternates(
+            parse_state       & state,
+            const char        * src,
+            const std::string & rule_name,
+            uint32_t            rule_id,
+            bool                is_nested);
+
+    static const char * parse_sequence(
+            parse_state                        & state,
+            const char                         * src,
+            const std::string                  & rule_name,
+            std::vector<whisper_grammar_element> & out_elements,
+            bool                                 is_nested) {
+        size_t last_sym_start = out_elements.size();
+        const char * pos = src;
+        while (*pos) {
+            if (*pos == '"') { // literal string
+                pos++;
+                last_sym_start = out_elements.size();
+                while (*pos != '"') {
+                    auto char_pair = parse_char(pos);
+                         pos       = char_pair.second;
+                    out_elements.push_back({WHISPER_GRETYPE_CHAR, char_pair.first});
+                }
+                pos = parse_space(pos + 1, is_nested);
+            } else if (*pos == '[') { // char range(s)
+                pos++;
+                enum whisper_gretype start_type = WHISPER_GRETYPE_CHAR;
+                if (*pos == '^') {
+                    pos++;
+                    start_type = WHISPER_GRETYPE_CHAR_NOT;
+                }
+                last_sym_start = out_elements.size();
+                while (*pos != ']') {
+                    auto char_pair = parse_char(pos);
+                         pos       = char_pair.second;
+                    enum whisper_gretype type = last_sym_start < out_elements.size()
+                        ? WHISPER_GRETYPE_CHAR_ALT
+                        : start_type;
+
+                    out_elements.push_back({type, char_pair.first});
+                    if (pos[0] == '-' && pos[1] != ']') {
+                        auto endchar_pair = parse_char(pos + 1);
+                             pos          = endchar_pair.second;
+                        out_elements.push_back({WHISPER_GRETYPE_CHAR_RNG_UPPER, endchar_pair.first});
+                    }
+                }
+                pos = parse_space(pos + 1, is_nested);
+            } else if (is_word_char(*pos)) { // rule reference
+                const char * name_end    = parse_name(pos);
+                uint32_t     ref_rule_id = get_symbol_id(state, pos, name_end - pos);
+                pos = parse_space(name_end, is_nested);
+                last_sym_start = out_elements.size();
+                out_elements.push_back({WHISPER_GRETYPE_RULE_REF, ref_rule_id});
+            } else if (*pos == '(') { // grouping
+                // parse nested alternates into synthesized rule
+                pos = parse_space(pos + 1, true);
+                uint32_t sub_rule_id = generate_symbol_id(state, rule_name);
+                pos = parse_alternates(state, pos, rule_name, sub_rule_id, true);
+                last_sym_start = out_elements.size();
+                // output reference to synthesized rule
+                out_elements.push_back({WHISPER_GRETYPE_RULE_REF, sub_rule_id});
+                if (*pos != ')') {
+                    throw std::runtime_error(std::string("expecting ')' at ") + pos);
+                }
+                pos = parse_space(pos + 1, is_nested);
+            } else if (*pos == '*' || *pos == '+' || *pos == '?') { // repetition operator
+                if (last_sym_start == out_elements.size()) {
+                    throw std::runtime_error(std::string("expecting preceding item to */+/? at ") + pos);
+                }
+
+                // apply transformation to previous symbol (last_sym_start to end) according to
+                // rewrite rules:
+                // S* --> S' ::= S S' |
+                // S+ --> S' ::= S S' | S
+                // S? --> S' ::= S |
+                uint32_t sub_rule_id = generate_symbol_id(state, rule_name);
+                std::vector<whisper_grammar_element> sub_rule;
+                // add preceding symbol to generated rule
+                sub_rule.insert(
+                    sub_rule.end(), out_elements.begin() + last_sym_start, out_elements.end());
+                if (*pos == '*' || *pos == '+') {
+                    // cause generated rule to recurse
+                    sub_rule.push_back({WHISPER_GRETYPE_RULE_REF, sub_rule_id});
+                }
+                // mark start of alternate def
+                sub_rule.push_back({WHISPER_GRETYPE_ALT, 0});
+                if (*pos == '+') {
+                    // add preceding symbol as alternate only for '+' (otherwise empty)
+                    sub_rule.insert(
+                        sub_rule.end(), out_elements.begin() + last_sym_start, out_elements.end());
+                }
+                sub_rule.push_back({WHISPER_GRETYPE_END, 0});
+                add_rule(state, sub_rule_id, sub_rule);
+
+                // in original rule, replace previous symbol with reference to generated rule
+                out_elements.resize(last_sym_start);
+                out_elements.push_back({WHISPER_GRETYPE_RULE_REF, sub_rule_id});
+
+                pos = parse_space(pos + 1, is_nested);
+            } else {
+                break;
+            }
+        }
+        return pos;
+    }
+
+    static const char * parse_alternates(
+            parse_state       & state,
+            const char        * src,
+            const std::string & rule_name,
+            uint32_t            rule_id,
+            bool                is_nested) {
+        std::vector<whisper_grammar_element> rule;
+        const char * pos = parse_sequence(state, src, rule_name, rule, is_nested);
+        while (*pos == '|') {
+            rule.push_back({WHISPER_GRETYPE_ALT, 0});
+            pos = parse_space(pos + 1, true);
+            pos = parse_sequence(state, pos, rule_name, rule, is_nested);
+        }
+        rule.push_back({WHISPER_GRETYPE_END, 0});
+        add_rule(state, rule_id, rule);
+        return pos;
+    }
+
+    static const char * parse_rule(parse_state & state, const char * src) {
+        const char * name_end = parse_name(src);
+        const char * pos      = parse_space(name_end, false);
+        size_t       name_len = name_end - src;
+        uint32_t     rule_id  = get_symbol_id(state, src, name_len);
+        const std::string name(src, name_len);
+
+        if (!(pos[0] == ':' && pos[1] == ':' && pos[2] == '=')) {
+            throw std::runtime_error(std::string("expecting ::= at ") + pos);
+        }
+        pos = parse_space(pos + 3, true);
+
+        pos = parse_alternates(state, pos, name, rule_id, false);
+
+        if (*pos == '\r') {
+            pos += pos[1] == '\n' ? 2 : 1;
+        } else if (*pos == '\n') {
+            pos++;
+        } else if (*pos) {
+            throw std::runtime_error(std::string("expecting newline or end at ") + pos);
+        }
+        return parse_space(pos, true);
+    }
+
+    parse_state parse(const char * src) {
+        try {
+            parse_state state;
+            const char * pos = parse_space(src, true);
+            while (*pos) {
+                pos = parse_rule(state, pos);
+            }
+            return state;
+        } catch (const std::exception & err) {
+            fprintf(stderr, "%s: error parsing grammar: %s\n", __func__, err.what());
+            return parse_state();
+        }
+    }
+
+    static void print_grammar_char(FILE * file, uint32_t c) {
+        if (0x20 <= c && c <= 0x7f) {
+            fprintf(file, "%c", static_cast<char>(c));
+        } else {
+            // cop out of encoding UTF-8
+            fprintf(file, "<U+%04X>", c);
+        }
+    }
+
+    static bool is_char_element(whisper_grammar_element elem) {
+        switch (elem.type) {
+            case WHISPER_GRETYPE_CHAR:           return true;
+            case WHISPER_GRETYPE_CHAR_NOT:       return true;
+            case WHISPER_GRETYPE_CHAR_ALT:       return true;
+            case WHISPER_GRETYPE_CHAR_RNG_UPPER: return true;
+            default:                           return false;
+        }
+    }
+
+    static void print_rule_binary(FILE * file, const std::vector<whisper_grammar_element> & rule) {
+        for (auto elem : rule) {
+            switch (elem.type) {
+                case WHISPER_GRETYPE_END:            fprintf(file, "END");            break;
+                case WHISPER_GRETYPE_ALT:            fprintf(file, "ALT");            break;
+                case WHISPER_GRETYPE_RULE_REF:       fprintf(file, "RULE_REF");       break;
+                case WHISPER_GRETYPE_CHAR:           fprintf(file, "CHAR");           break;
+                case WHISPER_GRETYPE_CHAR_NOT:       fprintf(file, "CHAR_NOT");       break;
+                case WHISPER_GRETYPE_CHAR_RNG_UPPER: fprintf(file, "CHAR_RNG_UPPER"); break;
+                case WHISPER_GRETYPE_CHAR_ALT:       fprintf(file, "CHAR_ALT");       break;
+            }
+            switch (elem.type) {
+                case WHISPER_GRETYPE_END:
+                case WHISPER_GRETYPE_ALT:
+                case WHISPER_GRETYPE_RULE_REF:
+                    fprintf(file, "(%u) ", elem.value);
+                    break;
+                case WHISPER_GRETYPE_CHAR:
+                case WHISPER_GRETYPE_CHAR_NOT:
+                case WHISPER_GRETYPE_CHAR_RNG_UPPER:
+                case WHISPER_GRETYPE_CHAR_ALT:
+                    fprintf(file, "(\"");
+                    print_grammar_char(file, elem.value);
+                    fprintf(file, "\") ");
+                    break;
+            }
+        }
+        fprintf(file, "\n");
+    }
+
+    static void print_rule(
+            FILE     * file,
+            uint32_t   rule_id,
+            const std::vector<whisper_grammar_element> & rule,
+            const std::map<uint32_t, std::string>    & symbol_id_names) {
+        if (rule.empty() || rule.back().type != WHISPER_GRETYPE_END) {
+            throw std::runtime_error(
+                "malformed rule, does not end with WHISPER_GRETYPE_END: " + std::to_string(rule_id));
+        }
+        fprintf(file, "%s ::= ", symbol_id_names.at(rule_id).c_str());
+        for (size_t i = 0, end = rule.size() - 1; i < end; i++) {
+            whisper_grammar_element elem = rule[i];
+            switch (elem.type) {
+                case WHISPER_GRETYPE_END:
+                    throw std::runtime_error(
+                        "unexpected end of rule: " + std::to_string(rule_id) + "," +
+                        std::to_string(i));
+                case WHISPER_GRETYPE_ALT:
+                    fprintf(file, "| ");
+                    break;
+                case WHISPER_GRETYPE_RULE_REF:
+                    fprintf(file, "%s ", symbol_id_names.at(elem.value).c_str());
+                    break;
+                case WHISPER_GRETYPE_CHAR:
+                    fprintf(file, "[");
+                    print_grammar_char(file, elem.value);
+                    break;
+                case WHISPER_GRETYPE_CHAR_NOT:
+                    fprintf(file, "[^");
+                    print_grammar_char(file, elem.value);
+                    break;
+                case WHISPER_GRETYPE_CHAR_RNG_UPPER:
+                    if (i == 0 || !is_char_element(rule[i - 1])) {
+                        throw std::runtime_error(
+                            "WHISPER_GRETYPE_CHAR_RNG_UPPER without preceding char: " +
+                            std::to_string(rule_id) + "," + std::to_string(i));
+                    }
+                    fprintf(file, "-");
+                    print_grammar_char(file, elem.value);
+                    break;
+                case WHISPER_GRETYPE_CHAR_ALT:
+                    if (i == 0 || !is_char_element(rule[i - 1])) {
+                        throw std::runtime_error(
+                            "WHISPER_GRETYPE_CHAR_ALT without preceding char: " +
+                            std::to_string(rule_id) + "," + std::to_string(i));
+                    }
+                    print_grammar_char(file, elem.value);
+                    break;
+            }
+            if (is_char_element(elem)) {
+                switch (rule[i + 1].type) {
+                    case WHISPER_GRETYPE_CHAR_ALT:
+                    case WHISPER_GRETYPE_CHAR_RNG_UPPER:
+                        break;
+                    default:
+                        fprintf(file, "] ");
+                }
+            }
+        }
+        fprintf(file, "\n");
+    }
+
+    void print_grammar(FILE * file, const parse_state & state) {
+        try {
+            std::map<uint32_t, std::string> symbol_id_names;
+            for (auto kv : state.symbol_ids) {
+                symbol_id_names[kv.second] = kv.first;
+            }
+            for (size_t i = 0, end = state.rules.size(); i < end; i++) {
+                // fprintf(file, "%zu: ", i);
+                // print_rule_binary(file, state.rules[i]);
+                print_rule(file, uint32_t(i), state.rules[i], symbol_id_names);
+                // fprintf(file, "\n");
+            }
+        } catch (const std::exception & err) {
+            fprintf(stderr, "\n%s: error printing grammar: %s\n", __func__, err.what());
+        }
+    }
+
+    std::vector<const whisper_grammar_element *> parse_state::c_rules() const {
+        std::vector<const whisper_grammar_element *> ret;
+        for (const auto & rule : rules) {
+            ret.push_back(rule.data());
+        }
+        return ret;
+    }
+}
diff --git a/src/whisper_cpp/examples/grammar-parser.h b/src/whisper_cpp/examples/grammar-parser.h
new file mode 100644
index 00000000..47d019c3
--- /dev/null
+++ b/src/whisper_cpp/examples/grammar-parser.h
@@ -0,0 +1,29 @@
+// Implements a parser for an extended Backus-Naur form (BNF), producing the
+// binary context-free grammar format specified by whisper.h. Supports character
+// ranges, grouping, and repetition operators. As an example, a grammar for
+// arithmetic might look like:
+//
+// root  ::= expr
+// expr  ::= term ([-+*/] term)*
+// term  ::= num | "(" space expr ")" space
+// num   ::= [0-9]+ space
+// space ::= [ \t\n]*
+
+#pragma once
+#include "whisper.h"
+#include <vector>
+#include <map>
+#include <cstdint>
+#include <string>
+
+namespace grammar_parser {
+    struct parse_state {
+        std::map<std::string, uint32_t>                   symbol_ids;
+        std::vector<std::vector<whisper_grammar_element>> rules;
+
+        std::vector<const whisper_grammar_element *>      c_rules() const;
+    };
+
+    parse_state parse(const char * src);
+    void print_grammar(FILE * file, const parse_state & state);
+}
diff --git a/src/whisper_cpp/examples/json.hpp b/src/whisper_cpp/examples/json.hpp
new file mode 100644
index 00000000..4d1a37ad
--- /dev/null
+++ b/src/whisper_cpp/examples/json.hpp
@@ -0,0 +1,24596 @@
+//     __ _____ _____ _____
+//  __|  |   __|     |   | |  JSON for Modern C++
+// |  |  |__   |  |  | | | |  version 3.11.2
+// |_____|_____|_____|_|___|  https://github.com/nlohmann/json
+//
+// SPDX-FileCopyrightText: 2013-2022 Niels Lohmann <https://nlohmann.me>
+// SPDX-License-Identifier: MIT
+
+/****************************************************************************\
+ * Note on documentation: The source files contain links to the online      *
+ * documentation of the public API at https://json.nlohmann.me. This URL    *
+ * contains the most recent documentation and should also be applicable to  *
+ * previous versions; documentation for deprecated functions is not         *
+ * removed, but marked deprecated. See "Generate documentation" section in  *
+ * file docs/README.md.                                                     *
+\****************************************************************************/
+
+#ifndef INCLUDE_NLOHMANN_JSON_HPP_
+#define INCLUDE_NLOHMANN_JSON_HPP_
+
+#include <algorithm> // all_of, find, for_each
+#include <cstddef> // nullptr_t, ptrdiff_t, size_t
+#include <functional> // hash, less
+#include <initializer_list> // initializer_list
+#ifndef JSON_NO_IO
+    #include <iosfwd> // istream, ostream
+#endif  // JSON_NO_IO
+#include <iterator> // random_access_iterator_tag
+#include <memory> // unique_ptr
+#include <numeric> // accumulate
+#include <string> // string, stoi, to_string
+#include <utility> // declval, forward, move, pair, swap
+#include <vector> // vector
+
+// #include <nlohmann/adl_serializer.hpp>
+//     __ _____ _____ _____
+//  __|  |   __|     |   | |  JSON for Modern C++
+// |  |  |__   |  |  | | | |  version 3.11.2
+// |_____|_____|_____|_|___|  https://github.com/nlohmann/json
+//
+// SPDX-FileCopyrightText: 2013-2022 Niels Lohmann <https://nlohmann.me>
+// SPDX-License-Identifier: MIT
+
+
+
+#include <utility>
+
+// #include <nlohmann/detail/abi_macros.hpp>
+//     __ _____ _____ _____
+//  __|  |   __|     |   | |  JSON for Modern C++
+// |  |  |__   |  |  | | | |  version 3.11.2
+// |_____|_____|_____|_|___|  https://github.com/nlohmann/json
+//
+// SPDX-FileCopyrightText: 2013-2022 Niels Lohmann <https://nlohmann.me>
+// SPDX-License-Identifier: MIT
+
+
+
+// This file contains all macro definitions affecting or depending on the ABI
+
+#ifndef JSON_SKIP_LIBRARY_VERSION_CHECK
+    #if defined(NLOHMANN_JSON_VERSION_MAJOR) && defined(NLOHMANN_JSON_VERSION_MINOR) && defined(NLOHMANN_JSON_VERSION_PATCH)
+        #if NLOHMANN_JSON_VERSION_MAJOR != 3 || NLOHMANN_JSON_VERSION_MINOR != 11 || NLOHMANN_JSON_VERSION_PATCH != 2
+            #warning "Already included a different version of the library!"
+        #endif
+    #endif
+#endif
+
+#define NLOHMANN_JSON_VERSION_MAJOR 3   // NOLINT(modernize-macro-to-enum)
+#define NLOHMANN_JSON_VERSION_MINOR 11  // NOLINT(modernize-macro-to-enum)
+#define NLOHMANN_JSON_VERSION_PATCH 2   // NOLINT(modernize-macro-to-enum)
+
+#ifndef JSON_DIAGNOSTICS
+    #define JSON_DIAGNOSTICS 0
+#endif
+
+#ifndef JSON_USE_LEGACY_DISCARDED_VALUE_COMPARISON
+    #define JSON_USE_LEGACY_DISCARDED_VALUE_COMPARISON 0
+#endif
+
+#if JSON_DIAGNOSTICS
+    #define NLOHMANN_JSON_ABI_TAG_DIAGNOSTICS _diag
+#else
+    #define NLOHMANN_JSON_ABI_TAG_DIAGNOSTICS
+#endif
+
+#if JSON_USE_LEGACY_DISCARDED_VALUE_COMPARISON
+    #define NLOHMANN_JSON_ABI_TAG_LEGACY_DISCARDED_VALUE_COMPARISON _ldvcmp
+#else
+    #define NLOHMANN_JSON_ABI_TAG_LEGACY_DISCARDED_VALUE_COMPARISON
+#endif
+
+#ifndef NLOHMANN_JSON_NAMESPACE_NO_VERSION
+    #define NLOHMANN_JSON_NAMESPACE_NO_VERSION 0
+#endif
+
+// Construct the namespace ABI tags component
+#define NLOHMANN_JSON_ABI_TAGS_CONCAT_EX(a, b) json_abi ## a ## b
+#define NLOHMANN_JSON_ABI_TAGS_CONCAT(a, b) \
+    NLOHMANN_JSON_ABI_TAGS_CONCAT_EX(a, b)
+
+#define NLOHMANN_JSON_ABI_TAGS                                       \
+    NLOHMANN_JSON_ABI_TAGS_CONCAT(                                   \
+            NLOHMANN_JSON_ABI_TAG_DIAGNOSTICS,                       \
+            NLOHMANN_JSON_ABI_TAG_LEGACY_DISCARDED_VALUE_COMPARISON)
+
+// Construct the namespace version component
+#define NLOHMANN_JSON_NAMESPACE_VERSION_CONCAT_EX(major, minor, patch) \
+    _v ## major ## _ ## minor ## _ ## patch
+#define NLOHMANN_JSON_NAMESPACE_VERSION_CONCAT(major, minor, patch) \
+    NLOHMANN_JSON_NAMESPACE_VERSION_CONCAT_EX(major, minor, patch)
+
+#if NLOHMANN_JSON_NAMESPACE_NO_VERSION
+#define NLOHMANN_JSON_NAMESPACE_VERSION
+#else
+#define NLOHMANN_JSON_NAMESPACE_VERSION                                 \
+    NLOHMANN_JSON_NAMESPACE_VERSION_CONCAT(NLOHMANN_JSON_VERSION_MAJOR, \
+                                           NLOHMANN_JSON_VERSION_MINOR, \
+                                           NLOHMANN_JSON_VERSION_PATCH)
+#endif
+
+// Combine namespace components
+#define NLOHMANN_JSON_NAMESPACE_CONCAT_EX(a, b) a ## b
+#define NLOHMANN_JSON_NAMESPACE_CONCAT(a, b) \
+    NLOHMANN_JSON_NAMESPACE_CONCAT_EX(a, b)
+
+#ifndef NLOHMANN_JSON_NAMESPACE
+#define NLOHMANN_JSON_NAMESPACE               \
+    nlohmann::NLOHMANN_JSON_NAMESPACE_CONCAT( \
+            NLOHMANN_JSON_ABI_TAGS,           \
+            NLOHMANN_JSON_NAMESPACE_VERSION)
+#endif
+
+#ifndef NLOHMANN_JSON_NAMESPACE_BEGIN
+#define NLOHMANN_JSON_NAMESPACE_BEGIN                \
+    namespace nlohmann                               \
+    {                                                \
+    inline namespace NLOHMANN_JSON_NAMESPACE_CONCAT( \
+                NLOHMANN_JSON_ABI_TAGS,              \
+                NLOHMANN_JSON_NAMESPACE_VERSION)     \
+    {
+#endif
+
+#ifndef NLOHMANN_JSON_NAMESPACE_END
+#define NLOHMANN_JSON_NAMESPACE_END                                     \
+    }  /* namespace (inline namespace) NOLINT(readability/namespace) */ \
+    }  // namespace nlohmann
+#endif
+
+// #include <nlohmann/detail/conversions/from_json.hpp>
+//     __ _____ _____ _____
+//  __|  |   __|     |   | |  JSON for Modern C++
+// |  |  |__   |  |  | | | |  version 3.11.2
+// |_____|_____|_____|_|___|  https://github.com/nlohmann/json
+//
+// SPDX-FileCopyrightText: 2013-2022 Niels Lohmann <https://nlohmann.me>
+// SPDX-License-Identifier: MIT
+
+
+
+#include <algorithm> // transform
+#include <array> // array
+#include <forward_list> // forward_list
+#include <iterator> // inserter, front_inserter, end
+#include <map> // map
+#include <string> // string
+#include <tuple> // tuple, make_tuple
+#include <type_traits> // is_arithmetic, is_same, is_enum, underlying_type, is_convertible
+#include <unordered_map> // unordered_map
+#include <utility> // pair, declval
+#include <valarray> // valarray
+
+// #include <nlohmann/detail/exceptions.hpp>
+//     __ _____ _____ _____
+//  __|  |   __|     |   | |  JSON for Modern C++
+// |  |  |__   |  |  | | | |  version 3.11.2
+// |_____|_____|_____|_|___|  https://github.com/nlohmann/json
+//
+// SPDX-FileCopyrightText: 2013-2022 Niels Lohmann <https://nlohmann.me>
+// SPDX-License-Identifier: MIT
+
+
+
+#include <cstddef> // nullptr_t
+#include <exception> // exception
+#include <stdexcept> // runtime_error
+#include <string> // to_string
+#include <vector> // vector
+
+// #include <nlohmann/detail/value_t.hpp>
+//     __ _____ _____ _____
+//  __|  |   __|     |   | |  JSON for Modern C++
+// |  |  |__   |  |  | | | |  version 3.11.2
+// |_____|_____|_____|_|___|  https://github.com/nlohmann/json
+//
+// SPDX-FileCopyrightText: 2013-2022 Niels Lohmann <https://nlohmann.me>
+// SPDX-License-Identifier: MIT
+
+
+
+#include <array> // array
+#include <cstddef> // size_t
+#include <cstdint> // uint8_t
+#include <string> // string
+
+// #include <nlohmann/detail/macro_scope.hpp>
+//     __ _____ _____ _____
+//  __|  |   __|     |   | |  JSON for Modern C++
+// |  |  |__   |  |  | | | |  version 3.11.2
+// |_____|_____|_____|_|___|  https://github.com/nlohmann/json
+//
+// SPDX-FileCopyrightText: 2013-2022 Niels Lohmann <https://nlohmann.me>
+// SPDX-License-Identifier: MIT
+
+
+
+#include <utility> // declval, pair
+// #include <nlohmann/detail/meta/detected.hpp>
+//     __ _____ _____ _____
+//  __|  |   __|     |   | |  JSON for Modern C++
+// |  |  |__   |  |  | | | |  version 3.11.2
+// |_____|_____|_____|_|___|  https://github.com/nlohmann/json
+//
+// SPDX-FileCopyrightText: 2013-2022 Niels Lohmann <https://nlohmann.me>
+// SPDX-License-Identifier: MIT
+
+
+
+#include <type_traits>
+
+// #include <nlohmann/detail/meta/void_t.hpp>
+//     __ _____ _____ _____
+//  __|  |   __|     |   | |  JSON for Modern C++
+// |  |  |__   |  |  | | | |  version 3.11.2
+// |_____|_____|_____|_|___|  https://github.com/nlohmann/json
+//
+// SPDX-FileCopyrightText: 2013-2022 Niels Lohmann <https://nlohmann.me>
+// SPDX-License-Identifier: MIT
+
+
+
+// #include <nlohmann/detail/abi_macros.hpp>
+
+
+NLOHMANN_JSON_NAMESPACE_BEGIN
+namespace detail
+{
+
+template<typename ...Ts> struct make_void
+{
+    using type = void;
+};
+template<typename ...Ts> using void_t = typename make_void<Ts...>::type;
+
+}  // namespace detail
+NLOHMANN_JSON_NAMESPACE_END
+
+
+NLOHMANN_JSON_NAMESPACE_BEGIN
+namespace detail
+{
+
+// https://en.cppreference.com/w/cpp/experimental/is_detected
+struct nonesuch
+{
+    nonesuch() = delete;
+    ~nonesuch() = delete;
+    nonesuch(nonesuch const&) = delete;
+    nonesuch(nonesuch const&&) = delete;
+    void operator=(nonesuch const&) = delete;
+    void operator=(nonesuch&&) = delete;
+};
+
+template<class Default,
+         class AlwaysVoid,
+         template<class...> class Op,
+         class... Args>
+struct detector
+{
+    using value_t = std::false_type;
+    using type = Default;
+};
+
+template<class Default, template<class...> class Op, class... Args>
+struct detector<Default, void_t<Op<Args...>>, Op, Args...>
+{
+    using value_t = std::true_type;
+    using type = Op<Args...>;
+};
+
+template<template<class...> class Op, class... Args>
+using is_detected = typename detector<nonesuch, void, Op, Args...>::value_t;
+
+template<template<class...> class Op, class... Args>
+struct is_detected_lazy : is_detected<Op, Args...> { };
+
+template<template<class...> class Op, class... Args>
+using detected_t = typename detector<nonesuch, void, Op, Args...>::type;
+
+template<class Default, template<class...> class Op, class... Args>
+using detected_or = detector<Default, void, Op, Args...>;
+
+template<class Default, template<class...> class Op, class... Args>
+using detected_or_t = typename detected_or<Default, Op, Args...>::type;
+
+template<class Expected, template<class...> class Op, class... Args>
+using is_detected_exact = std::is_same<Expected, detected_t<Op, Args...>>;
+
+template<class To, template<class...> class Op, class... Args>
+using is_detected_convertible =
+    std::is_convertible<detected_t<Op, Args...>, To>;
+
+}  // namespace detail
+NLOHMANN_JSON_NAMESPACE_END
+
+// #include <nlohmann/thirdparty/hedley/hedley.hpp>
+
+
+//     __ _____ _____ _____
+//  __|  |   __|     |   | |  JSON for Modern C++
+// |  |  |__   |  |  | | | |  version 3.11.2
+// |_____|_____|_____|_|___|  https://github.com/nlohmann/json
+//
+// SPDX-FileCopyrightText: 2013-2022 Niels Lohmann <https://nlohmann.me>
+// SPDX-FileCopyrightText: 2016-2021 Evan Nemerson <evan@nemerson.com>
+// SPDX-License-Identifier: MIT
+
+/* Hedley - https://nemequ.github.io/hedley
+ * Created by Evan Nemerson <evan@nemerson.com>
+ */
+
+#if !defined(JSON_HEDLEY_VERSION) || (JSON_HEDLEY_VERSION < 15)
+#if defined(JSON_HEDLEY_VERSION)
+    #undef JSON_HEDLEY_VERSION
+#endif
+#define JSON_HEDLEY_VERSION 15
+
+#if defined(JSON_HEDLEY_STRINGIFY_EX)
+    #undef JSON_HEDLEY_STRINGIFY_EX
+#endif
+#define JSON_HEDLEY_STRINGIFY_EX(x) #x
+
+#if defined(JSON_HEDLEY_STRINGIFY)
+    #undef JSON_HEDLEY_STRINGIFY
+#endif
+#define JSON_HEDLEY_STRINGIFY(x) JSON_HEDLEY_STRINGIFY_EX(x)
+
+#if defined(JSON_HEDLEY_CONCAT_EX)
+    #undef JSON_HEDLEY_CONCAT_EX
+#endif
+#define JSON_HEDLEY_CONCAT_EX(a,b) a##b
+
+#if defined(JSON_HEDLEY_CONCAT)
+    #undef JSON_HEDLEY_CONCAT
+#endif
+#define JSON_HEDLEY_CONCAT(a,b) JSON_HEDLEY_CONCAT_EX(a,b)
+
+#if defined(JSON_HEDLEY_CONCAT3_EX)
+    #undef JSON_HEDLEY_CONCAT3_EX
+#endif
+#define JSON_HEDLEY_CONCAT3_EX(a,b,c) a##b##c
+
+#if defined(JSON_HEDLEY_CONCAT3)
+    #undef JSON_HEDLEY_CONCAT3
+#endif
+#define JSON_HEDLEY_CONCAT3(a,b,c) JSON_HEDLEY_CONCAT3_EX(a,b,c)
+
+#if defined(JSON_HEDLEY_VERSION_ENCODE)
+    #undef JSON_HEDLEY_VERSION_ENCODE
+#endif
+#define JSON_HEDLEY_VERSION_ENCODE(major,minor,revision) (((major) * 1000000) + ((minor) * 1000) + (revision))
+
+#if defined(JSON_HEDLEY_VERSION_DECODE_MAJOR)
+    #undef JSON_HEDLEY_VERSION_DECODE_MAJOR
+#endif
+#define JSON_HEDLEY_VERSION_DECODE_MAJOR(version) ((version) / 1000000)
+
+#if defined(JSON_HEDLEY_VERSION_DECODE_MINOR)
+    #undef JSON_HEDLEY_VERSION_DECODE_MINOR
+#endif
+#define JSON_HEDLEY_VERSION_DECODE_MINOR(version) (((version) % 1000000) / 1000)
+
+#if defined(JSON_HEDLEY_VERSION_DECODE_REVISION)
+    #undef JSON_HEDLEY_VERSION_DECODE_REVISION
+#endif
+#define JSON_HEDLEY_VERSION_DECODE_REVISION(version) ((version) % 1000)
+
+#if defined(JSON_HEDLEY_GNUC_VERSION)
+    #undef JSON_HEDLEY_GNUC_VERSION
+#endif
+#if defined(__GNUC__) && defined(__GNUC_PATCHLEVEL__)
+    #define JSON_HEDLEY_GNUC_VERSION JSON_HEDLEY_VERSION_ENCODE(__GNUC__, __GNUC_MINOR__, __GNUC_PATCHLEVEL__)
+#elif defined(__GNUC__)
+    #define JSON_HEDLEY_GNUC_VERSION JSON_HEDLEY_VERSION_ENCODE(__GNUC__, __GNUC_MINOR__, 0)
+#endif
+
+#if defined(JSON_HEDLEY_GNUC_VERSION_CHECK)
+    #undef JSON_HEDLEY_GNUC_VERSION_CHECK
+#endif
+#if defined(JSON_HEDLEY_GNUC_VERSION)
+    #define JSON_HEDLEY_GNUC_VERSION_CHECK(major,minor,patch) (JSON_HEDLEY_GNUC_VERSION >= JSON_HEDLEY_VERSION_ENCODE(major, minor, patch))
+#else
+    #define JSON_HEDLEY_GNUC_VERSION_CHECK(major,minor,patch) (0)
+#endif
+
+#if defined(JSON_HEDLEY_MSVC_VERSION)
+    #undef JSON_HEDLEY_MSVC_VERSION
+#endif
+#if defined(_MSC_FULL_VER) && (_MSC_FULL_VER >= 140000000) && !defined(__ICL)
+    #define JSON_HEDLEY_MSVC_VERSION JSON_HEDLEY_VERSION_ENCODE(_MSC_FULL_VER / 10000000, (_MSC_FULL_VER % 10000000) / 100000, (_MSC_FULL_VER % 100000) / 100)
+#elif defined(_MSC_FULL_VER) && !defined(__ICL)
+    #define JSON_HEDLEY_MSVC_VERSION JSON_HEDLEY_VERSION_ENCODE(_MSC_FULL_VER / 1000000, (_MSC_FULL_VER % 1000000) / 10000, (_MSC_FULL_VER % 10000) / 10)
+#elif defined(_MSC_VER) && !defined(__ICL)
+    #define JSON_HEDLEY_MSVC_VERSION JSON_HEDLEY_VERSION_ENCODE(_MSC_VER / 100, _MSC_VER % 100, 0)
+#endif
+
+#if defined(JSON_HEDLEY_MSVC_VERSION_CHECK)
+    #undef JSON_HEDLEY_MSVC_VERSION_CHECK
+#endif
+#if !defined(JSON_HEDLEY_MSVC_VERSION)
+    #define JSON_HEDLEY_MSVC_VERSION_CHECK(major,minor,patch) (0)
+#elif defined(_MSC_VER) && (_MSC_VER >= 1400)
+    #define JSON_HEDLEY_MSVC_VERSION_CHECK(major,minor,patch) (_MSC_FULL_VER >= ((major * 10000000) + (minor * 100000) + (patch)))
+#elif defined(_MSC_VER) && (_MSC_VER >= 1200)
+    #define JSON_HEDLEY_MSVC_VERSION_CHECK(major,minor,patch) (_MSC_FULL_VER >= ((major * 1000000) + (minor * 10000) + (patch)))
+#else
+    #define JSON_HEDLEY_MSVC_VERSION_CHECK(major,minor,patch) (_MSC_VER >= ((major * 100) + (minor)))
+#endif
+
+#if defined(JSON_HEDLEY_INTEL_VERSION)
+    #undef JSON_HEDLEY_INTEL_VERSION
+#endif
+#if defined(__INTEL_COMPILER) && defined(__INTEL_COMPILER_UPDATE) && !defined(__ICL)
+    #define JSON_HEDLEY_INTEL_VERSION JSON_HEDLEY_VERSION_ENCODE(__INTEL_COMPILER / 100, __INTEL_COMPILER % 100, __INTEL_COMPILER_UPDATE)
+#elif defined(__INTEL_COMPILER) && !defined(__ICL)
+    #define JSON_HEDLEY_INTEL_VERSION JSON_HEDLEY_VERSION_ENCODE(__INTEL_COMPILER / 100, __INTEL_COMPILER % 100, 0)
+#endif
+
+#if defined(JSON_HEDLEY_INTEL_VERSION_CHECK)
+    #undef JSON_HEDLEY_INTEL_VERSION_CHECK
+#endif
+#if defined(JSON_HEDLEY_INTEL_VERSION)
+    #define JSON_HEDLEY_INTEL_VERSION_CHECK(major,minor,patch) (JSON_HEDLEY_INTEL_VERSION >= JSON_HEDLEY_VERSION_ENCODE(major, minor, patch))
+#else
+    #define JSON_HEDLEY_INTEL_VERSION_CHECK(major,minor,patch) (0)
+#endif
+
+#if defined(JSON_HEDLEY_INTEL_CL_VERSION)
+    #undef JSON_HEDLEY_INTEL_CL_VERSION
+#endif
+#if defined(__INTEL_COMPILER) && defined(__INTEL_COMPILER_UPDATE) && defined(__ICL)
+    #define JSON_HEDLEY_INTEL_CL_VERSION JSON_HEDLEY_VERSION_ENCODE(__INTEL_COMPILER, __INTEL_COMPILER_UPDATE, 0)
+#endif
+
+#if defined(JSON_HEDLEY_INTEL_CL_VERSION_CHECK)
+    #undef JSON_HEDLEY_INTEL_CL_VERSION_CHECK
+#endif
+#if defined(JSON_HEDLEY_INTEL_CL_VERSION)
+    #define JSON_HEDLEY_INTEL_CL_VERSION_CHECK(major,minor,patch) (JSON_HEDLEY_INTEL_CL_VERSION >= JSON_HEDLEY_VERSION_ENCODE(major, minor, patch))
+#else
+    #define JSON_HEDLEY_INTEL_CL_VERSION_CHECK(major,minor,patch) (0)
+#endif
+
+#if defined(JSON_HEDLEY_PGI_VERSION)
+    #undef JSON_HEDLEY_PGI_VERSION
+#endif
+#if defined(__PGI) && defined(__PGIC__) && defined(__PGIC_MINOR__) && defined(__PGIC_PATCHLEVEL__)
+    #define JSON_HEDLEY_PGI_VERSION JSON_HEDLEY_VERSION_ENCODE(__PGIC__, __PGIC_MINOR__, __PGIC_PATCHLEVEL__)
+#endif
+
+#if defined(JSON_HEDLEY_PGI_VERSION_CHECK)
+    #undef JSON_HEDLEY_PGI_VERSION_CHECK
+#endif
+#if defined(JSON_HEDLEY_PGI_VERSION)
+    #define JSON_HEDLEY_PGI_VERSION_CHECK(major,minor,patch) (JSON_HEDLEY_PGI_VERSION >= JSON_HEDLEY_VERSION_ENCODE(major, minor, patch))
+#else
+    #define JSON_HEDLEY_PGI_VERSION_CHECK(major,minor,patch) (0)
+#endif
+
+#if defined(JSON_HEDLEY_SUNPRO_VERSION)
+    #undef JSON_HEDLEY_SUNPRO_VERSION
+#endif
+#if defined(__SUNPRO_C) && (__SUNPRO_C > 0x1000)
+    #define JSON_HEDLEY_SUNPRO_VERSION JSON_HEDLEY_VERSION_ENCODE((((__SUNPRO_C >> 16) & 0xf) * 10) + ((__SUNPRO_C >> 12) & 0xf), (((__SUNPRO_C >> 8) & 0xf) * 10) + ((__SUNPRO_C >> 4) & 0xf), (__SUNPRO_C & 0xf) * 10)
+#elif defined(__SUNPRO_C)
+    #define JSON_HEDLEY_SUNPRO_VERSION JSON_HEDLEY_VERSION_ENCODE((__SUNPRO_C >> 8) & 0xf, (__SUNPRO_C >> 4) & 0xf, (__SUNPRO_C) & 0xf)
+#elif defined(__SUNPRO_CC) && (__SUNPRO_CC > 0x1000)
+    #define JSON_HEDLEY_SUNPRO_VERSION JSON_HEDLEY_VERSION_ENCODE((((__SUNPRO_CC >> 16) & 0xf) * 10) + ((__SUNPRO_CC >> 12) & 0xf), (((__SUNPRO_CC >> 8) & 0xf) * 10) + ((__SUNPRO_CC >> 4) & 0xf), (__SUNPRO_CC & 0xf) * 10)
+#elif defined(__SUNPRO_CC)
+    #define JSON_HEDLEY_SUNPRO_VERSION JSON_HEDLEY_VERSION_ENCODE((__SUNPRO_CC >> 8) & 0xf, (__SUNPRO_CC >> 4) & 0xf, (__SUNPRO_CC) & 0xf)
+#endif
+
+#if defined(JSON_HEDLEY_SUNPRO_VERSION_CHECK)
+    #undef JSON_HEDLEY_SUNPRO_VERSION_CHECK
+#endif
+#if defined(JSON_HEDLEY_SUNPRO_VERSION)
+    #define JSON_HEDLEY_SUNPRO_VERSION_CHECK(major,minor,patch) (JSON_HEDLEY_SUNPRO_VERSION >= JSON_HEDLEY_VERSION_ENCODE(major, minor, patch))
+#else
+    #define JSON_HEDLEY_SUNPRO_VERSION_CHECK(major,minor,patch) (0)
+#endif
+
+#if defined(JSON_HEDLEY_EMSCRIPTEN_VERSION)
+    #undef JSON_HEDLEY_EMSCRIPTEN_VERSION
+#endif
+#if defined(__EMSCRIPTEN__)
+    #define JSON_HEDLEY_EMSCRIPTEN_VERSION JSON_HEDLEY_VERSION_ENCODE(__EMSCRIPTEN_major__, __EMSCRIPTEN_minor__, __EMSCRIPTEN_tiny__)
+#endif
+
+#if defined(JSON_HEDLEY_EMSCRIPTEN_VERSION_CHECK)
+    #undef JSON_HEDLEY_EMSCRIPTEN_VERSION_CHECK
+#endif
+#if defined(JSON_HEDLEY_EMSCRIPTEN_VERSION)
+    #define JSON_HEDLEY_EMSCRIPTEN_VERSION_CHECK(major,minor,patch) (JSON_HEDLEY_EMSCRIPTEN_VERSION >= JSON_HEDLEY_VERSION_ENCODE(major, minor, patch))
+#else
+    #define JSON_HEDLEY_EMSCRIPTEN_VERSION_CHECK(major,minor,patch) (0)
+#endif
+
+#if defined(JSON_HEDLEY_ARM_VERSION)
+    #undef JSON_HEDLEY_ARM_VERSION
+#endif
+#if defined(__CC_ARM) && defined(__ARMCOMPILER_VERSION)
+    #define JSON_HEDLEY_ARM_VERSION JSON_HEDLEY_VERSION_ENCODE(__ARMCOMPILER_VERSION / 1000000, (__ARMCOMPILER_VERSION % 1000000) / 10000, (__ARMCOMPILER_VERSION % 10000) / 100)
+#elif defined(__CC_ARM) && defined(__ARMCC_VERSION)
+    #define JSON_HEDLEY_ARM_VERSION JSON_HEDLEY_VERSION_ENCODE(__ARMCC_VERSION / 1000000, (__ARMCC_VERSION % 1000000) / 10000, (__ARMCC_VERSION % 10000) / 100)
+#endif
+
+#if defined(JSON_HEDLEY_ARM_VERSION_CHECK)
+    #undef JSON_HEDLEY_ARM_VERSION_CHECK
+#endif
+#if defined(JSON_HEDLEY_ARM_VERSION)
+    #define JSON_HEDLEY_ARM_VERSION_CHECK(major,minor,patch) (JSON_HEDLEY_ARM_VERSION >= JSON_HEDLEY_VERSION_ENCODE(major, minor, patch))
+#else
+    #define JSON_HEDLEY_ARM_VERSION_CHECK(major,minor,patch) (0)
+#endif
+
+#if defined(JSON_HEDLEY_IBM_VERSION)
+    #undef JSON_HEDLEY_IBM_VERSION
+#endif
+#if defined(__ibmxl__)
+    #define JSON_HEDLEY_IBM_VERSION JSON_HEDLEY_VERSION_ENCODE(__ibmxl_version__, __ibmxl_release__, __ibmxl_modification__)
+#elif defined(__xlC__) && defined(__xlC_ver__)
+    #define JSON_HEDLEY_IBM_VERSION JSON_HEDLEY_VERSION_ENCODE(__xlC__ >> 8, __xlC__ & 0xff, (__xlC_ver__ >> 8) & 0xff)
+#elif defined(__xlC__)
+    #define JSON_HEDLEY_IBM_VERSION JSON_HEDLEY_VERSION_ENCODE(__xlC__ >> 8, __xlC__ & 0xff, 0)
+#endif
+
+#if defined(JSON_HEDLEY_IBM_VERSION_CHECK)
+    #undef JSON_HEDLEY_IBM_VERSION_CHECK
+#endif
+#if defined(JSON_HEDLEY_IBM_VERSION)
+    #define JSON_HEDLEY_IBM_VERSION_CHECK(major,minor,patch) (JSON_HEDLEY_IBM_VERSION >= JSON_HEDLEY_VERSION_ENCODE(major, minor, patch))
+#else
+    #define JSON_HEDLEY_IBM_VERSION_CHECK(major,minor,patch) (0)
+#endif
+
+#if defined(JSON_HEDLEY_TI_VERSION)
+    #undef JSON_HEDLEY_TI_VERSION
+#endif
+#if \
+    defined(__TI_COMPILER_VERSION__) && \
+    ( \
+      defined(__TMS470__) || defined(__TI_ARM__) || \
+      defined(__MSP430__) || \
+      defined(__TMS320C2000__) \
+    )
+#if (__TI_COMPILER_VERSION__ >= 16000000)
+    #define JSON_HEDLEY_TI_VERSION JSON_HEDLEY_VERSION_ENCODE(__TI_COMPILER_VERSION__ / 1000000, (__TI_COMPILER_VERSION__ % 1000000) / 1000, (__TI_COMPILER_VERSION__ % 1000))
+#endif
+#endif
+
+#if defined(JSON_HEDLEY_TI_VERSION_CHECK)
+    #undef JSON_HEDLEY_TI_VERSION_CHECK
+#endif
+#if defined(JSON_HEDLEY_TI_VERSION)
+    #define JSON_HEDLEY_TI_VERSION_CHECK(major,minor,patch) (JSON_HEDLEY_TI_VERSION >= JSON_HEDLEY_VERSION_ENCODE(major, minor, patch))
+#else
+    #define JSON_HEDLEY_TI_VERSION_CHECK(major,minor,patch) (0)
+#endif
+
+#if defined(JSON_HEDLEY_TI_CL2000_VERSION)
+    #undef JSON_HEDLEY_TI_CL2000_VERSION
+#endif
+#if defined(__TI_COMPILER_VERSION__) && defined(__TMS320C2000__)
+    #define JSON_HEDLEY_TI_CL2000_VERSION JSON_HEDLEY_VERSION_ENCODE(__TI_COMPILER_VERSION__ / 1000000, (__TI_COMPILER_VERSION__ % 1000000) / 1000, (__TI_COMPILER_VERSION__ % 1000))
+#endif
+
+#if defined(JSON_HEDLEY_TI_CL2000_VERSION_CHECK)
+    #undef JSON_HEDLEY_TI_CL2000_VERSION_CHECK
+#endif
+#if defined(JSON_HEDLEY_TI_CL2000_VERSION)
+    #define JSON_HEDLEY_TI_CL2000_VERSION_CHECK(major,minor,patch) (JSON_HEDLEY_TI_CL2000_VERSION >= JSON_HEDLEY_VERSION_ENCODE(major, minor, patch))
+#else
+    #define JSON_HEDLEY_TI_CL2000_VERSION_CHECK(major,minor,patch) (0)
+#endif
+
+#if defined(JSON_HEDLEY_TI_CL430_VERSION)
+    #undef JSON_HEDLEY_TI_CL430_VERSION
+#endif
+#if defined(__TI_COMPILER_VERSION__) && defined(__MSP430__)
+    #define JSON_HEDLEY_TI_CL430_VERSION JSON_HEDLEY_VERSION_ENCODE(__TI_COMPILER_VERSION__ / 1000000, (__TI_COMPILER_VERSION__ % 1000000) / 1000, (__TI_COMPILER_VERSION__ % 1000))
+#endif
+
+#if defined(JSON_HEDLEY_TI_CL430_VERSION_CHECK)
+    #undef JSON_HEDLEY_TI_CL430_VERSION_CHECK
+#endif
+#if defined(JSON_HEDLEY_TI_CL430_VERSION)
+    #define JSON_HEDLEY_TI_CL430_VERSION_CHECK(major,minor,patch) (JSON_HEDLEY_TI_CL430_VERSION >= JSON_HEDLEY_VERSION_ENCODE(major, minor, patch))
+#else
+    #define JSON_HEDLEY_TI_CL430_VERSION_CHECK(major,minor,patch) (0)
+#endif
+
+#if defined(JSON_HEDLEY_TI_ARMCL_VERSION)
+    #undef JSON_HEDLEY_TI_ARMCL_VERSION
+#endif
+#if defined(__TI_COMPILER_VERSION__) && (defined(__TMS470__) || defined(__TI_ARM__))
+    #define JSON_HEDLEY_TI_ARMCL_VERSION JSON_HEDLEY_VERSION_ENCODE(__TI_COMPILER_VERSION__ / 1000000, (__TI_COMPILER_VERSION__ % 1000000) / 1000, (__TI_COMPILER_VERSION__ % 1000))
+#endif
+
+#if defined(JSON_HEDLEY_TI_ARMCL_VERSION_CHECK)
+    #undef JSON_HEDLEY_TI_ARMCL_VERSION_CHECK
+#endif
+#if defined(JSON_HEDLEY_TI_ARMCL_VERSION)
+    #define JSON_HEDLEY_TI_ARMCL_VERSION_CHECK(major,minor,patch) (JSON_HEDLEY_TI_ARMCL_VERSION >= JSON_HEDLEY_VERSION_ENCODE(major, minor, patch))
+#else
+    #define JSON_HEDLEY_TI_ARMCL_VERSION_CHECK(major,minor,patch) (0)
+#endif
+
+#if defined(JSON_HEDLEY_TI_CL6X_VERSION)
+    #undef JSON_HEDLEY_TI_CL6X_VERSION
+#endif
+#if defined(__TI_COMPILER_VERSION__) && defined(__TMS320C6X__)
+    #define JSON_HEDLEY_TI_CL6X_VERSION JSON_HEDLEY_VERSION_ENCODE(__TI_COMPILER_VERSION__ / 1000000, (__TI_COMPILER_VERSION__ % 1000000) / 1000, (__TI_COMPILER_VERSION__ % 1000))
+#endif
+
+#if defined(JSON_HEDLEY_TI_CL6X_VERSION_CHECK)
+    #undef JSON_HEDLEY_TI_CL6X_VERSION_CHECK
+#endif
+#if defined(JSON_HEDLEY_TI_CL6X_VERSION)
+    #define JSON_HEDLEY_TI_CL6X_VERSION_CHECK(major,minor,patch) (JSON_HEDLEY_TI_CL6X_VERSION >= JSON_HEDLEY_VERSION_ENCODE(major, minor, patch))
+#else
+    #define JSON_HEDLEY_TI_CL6X_VERSION_CHECK(major,minor,patch) (0)
+#endif
+
+#if defined(JSON_HEDLEY_TI_CL7X_VERSION)
+    #undef JSON_HEDLEY_TI_CL7X_VERSION
+#endif
+#if defined(__TI_COMPILER_VERSION__) && defined(__C7000__)
+    #define JSON_HEDLEY_TI_CL7X_VERSION JSON_HEDLEY_VERSION_ENCODE(__TI_COMPILER_VERSION__ / 1000000, (__TI_COMPILER_VERSION__ % 1000000) / 1000, (__TI_COMPILER_VERSION__ % 1000))
+#endif
+
+#if defined(JSON_HEDLEY_TI_CL7X_VERSION_CHECK)
+    #undef JSON_HEDLEY_TI_CL7X_VERSION_CHECK
+#endif
+#if defined(JSON_HEDLEY_TI_CL7X_VERSION)
+    #define JSON_HEDLEY_TI_CL7X_VERSION_CHECK(major,minor,patch) (JSON_HEDLEY_TI_CL7X_VERSION >= JSON_HEDLEY_VERSION_ENCODE(major, minor, patch))
+#else
+    #define JSON_HEDLEY_TI_CL7X_VERSION_CHECK(major,minor,patch) (0)
+#endif
+
+#if defined(JSON_HEDLEY_TI_CLPRU_VERSION)
+    #undef JSON_HEDLEY_TI_CLPRU_VERSION
+#endif
+#if defined(__TI_COMPILER_VERSION__) && defined(__PRU__)
+    #define JSON_HEDLEY_TI_CLPRU_VERSION JSON_HEDLEY_VERSION_ENCODE(__TI_COMPILER_VERSION__ / 1000000, (__TI_COMPILER_VERSION__ % 1000000) / 1000, (__TI_COMPILER_VERSION__ % 1000))
+#endif
+
+#if defined(JSON_HEDLEY_TI_CLPRU_VERSION_CHECK)
+    #undef JSON_HEDLEY_TI_CLPRU_VERSION_CHECK
+#endif
+#if defined(JSON_HEDLEY_TI_CLPRU_VERSION)
+    #define JSON_HEDLEY_TI_CLPRU_VERSION_CHECK(major,minor,patch) (JSON_HEDLEY_TI_CLPRU_VERSION >= JSON_HEDLEY_VERSION_ENCODE(major, minor, patch))
+#else
+    #define JSON_HEDLEY_TI_CLPRU_VERSION_CHECK(major,minor,patch) (0)
+#endif
+
+#if defined(JSON_HEDLEY_CRAY_VERSION)
+    #undef JSON_HEDLEY_CRAY_VERSION
+#endif
+#if defined(_CRAYC)
+    #if defined(_RELEASE_PATCHLEVEL)
+        #define JSON_HEDLEY_CRAY_VERSION JSON_HEDLEY_VERSION_ENCODE(_RELEASE_MAJOR, _RELEASE_MINOR, _RELEASE_PATCHLEVEL)
+    #else
+        #define JSON_HEDLEY_CRAY_VERSION JSON_HEDLEY_VERSION_ENCODE(_RELEASE_MAJOR, _RELEASE_MINOR, 0)
+    #endif
+#endif
+
+#if defined(JSON_HEDLEY_CRAY_VERSION_CHECK)
+    #undef JSON_HEDLEY_CRAY_VERSION_CHECK
+#endif
+#if defined(JSON_HEDLEY_CRAY_VERSION)
+    #define JSON_HEDLEY_CRAY_VERSION_CHECK(major,minor,patch) (JSON_HEDLEY_CRAY_VERSION >= JSON_HEDLEY_VERSION_ENCODE(major, minor, patch))
+#else
+    #define JSON_HEDLEY_CRAY_VERSION_CHECK(major,minor,patch) (0)
+#endif
+
+#if defined(JSON_HEDLEY_IAR_VERSION)
+    #undef JSON_HEDLEY_IAR_VERSION
+#endif
+#if defined(__IAR_SYSTEMS_ICC__)
+    #if __VER__ > 1000
+        #define JSON_HEDLEY_IAR_VERSION JSON_HEDLEY_VERSION_ENCODE((__VER__ / 1000000), ((__VER__ / 1000) % 1000), (__VER__ % 1000))
+    #else
+        #define JSON_HEDLEY_IAR_VERSION JSON_HEDLEY_VERSION_ENCODE(__VER__ / 100, __VER__ % 100, 0)
+    #endif
+#endif
+
+#if defined(JSON_HEDLEY_IAR_VERSION_CHECK)
+    #undef JSON_HEDLEY_IAR_VERSION_CHECK
+#endif
+#if defined(JSON_HEDLEY_IAR_VERSION)
+    #define JSON_HEDLEY_IAR_VERSION_CHECK(major,minor,patch) (JSON_HEDLEY_IAR_VERSION >= JSON_HEDLEY_VERSION_ENCODE(major, minor, patch))
+#else
+    #define JSON_HEDLEY_IAR_VERSION_CHECK(major,minor,patch) (0)
+#endif
+
+#if defined(JSON_HEDLEY_TINYC_VERSION)
+    #undef JSON_HEDLEY_TINYC_VERSION
+#endif
+#if defined(__TINYC__)
+    #define JSON_HEDLEY_TINYC_VERSION JSON_HEDLEY_VERSION_ENCODE(__TINYC__ / 1000, (__TINYC__ / 100) % 10, __TINYC__ % 100)
+#endif
+
+#if defined(JSON_HEDLEY_TINYC_VERSION_CHECK)
+    #undef JSON_HEDLEY_TINYC_VERSION_CHECK
+#endif
+#if defined(JSON_HEDLEY_TINYC_VERSION)
+    #define JSON_HEDLEY_TINYC_VERSION_CHECK(major,minor,patch) (JSON_HEDLEY_TINYC_VERSION >= JSON_HEDLEY_VERSION_ENCODE(major, minor, patch))
+#else
+    #define JSON_HEDLEY_TINYC_VERSION_CHECK(major,minor,patch) (0)
+#endif
+
+#if defined(JSON_HEDLEY_DMC_VERSION)
+    #undef JSON_HEDLEY_DMC_VERSION
+#endif
+#if defined(__DMC__)
+    #define JSON_HEDLEY_DMC_VERSION JSON_HEDLEY_VERSION_ENCODE(__DMC__ >> 8, (__DMC__ >> 4) & 0xf, __DMC__ & 0xf)
+#endif
+
+#if defined(JSON_HEDLEY_DMC_VERSION_CHECK)
+    #undef JSON_HEDLEY_DMC_VERSION_CHECK
+#endif
+#if defined(JSON_HEDLEY_DMC_VERSION)
+    #define JSON_HEDLEY_DMC_VERSION_CHECK(major,minor,patch) (JSON_HEDLEY_DMC_VERSION >= JSON_HEDLEY_VERSION_ENCODE(major, minor, patch))
+#else
+    #define JSON_HEDLEY_DMC_VERSION_CHECK(major,minor,patch) (0)
+#endif
+
+#if defined(JSON_HEDLEY_COMPCERT_VERSION)
+    #undef JSON_HEDLEY_COMPCERT_VERSION
+#endif
+#if defined(__COMPCERT_VERSION__)
+    #define JSON_HEDLEY_COMPCERT_VERSION JSON_HEDLEY_VERSION_ENCODE(__COMPCERT_VERSION__ / 10000, (__COMPCERT_VERSION__ / 100) % 100, __COMPCERT_VERSION__ % 100)
+#endif
+
+#if defined(JSON_HEDLEY_COMPCERT_VERSION_CHECK)
+    #undef JSON_HEDLEY_COMPCERT_VERSION_CHECK
+#endif
+#if defined(JSON_HEDLEY_COMPCERT_VERSION)
+    #define JSON_HEDLEY_COMPCERT_VERSION_CHECK(major,minor,patch) (JSON_HEDLEY_COMPCERT_VERSION >= JSON_HEDLEY_VERSION_ENCODE(major, minor, patch))
+#else
+    #define JSON_HEDLEY_COMPCERT_VERSION_CHECK(major,minor,patch) (0)
+#endif
+
+#if defined(JSON_HEDLEY_PELLES_VERSION)
+    #undef JSON_HEDLEY_PELLES_VERSION
+#endif
+#if defined(__POCC__)
+    #define JSON_HEDLEY_PELLES_VERSION JSON_HEDLEY_VERSION_ENCODE(__POCC__ / 100, __POCC__ % 100, 0)
+#endif
+
+#if defined(JSON_HEDLEY_PELLES_VERSION_CHECK)
+    #undef JSON_HEDLEY_PELLES_VERSION_CHECK
+#endif
+#if defined(JSON_HEDLEY_PELLES_VERSION)
+    #define JSON_HEDLEY_PELLES_VERSION_CHECK(major,minor,patch) (JSON_HEDLEY_PELLES_VERSION >= JSON_HEDLEY_VERSION_ENCODE(major, minor, patch))
+#else
+    #define JSON_HEDLEY_PELLES_VERSION_CHECK(major,minor,patch) (0)
+#endif
+
+#if defined(JSON_HEDLEY_MCST_LCC_VERSION)
+    #undef JSON_HEDLEY_MCST_LCC_VERSION
+#endif
+#if defined(__LCC__) && defined(__LCC_MINOR__)
+    #define JSON_HEDLEY_MCST_LCC_VERSION JSON_HEDLEY_VERSION_ENCODE(__LCC__ / 100, __LCC__ % 100, __LCC_MINOR__)
+#endif
+
+#if defined(JSON_HEDLEY_MCST_LCC_VERSION_CHECK)
+    #undef JSON_HEDLEY_MCST_LCC_VERSION_CHECK
+#endif
+#if defined(JSON_HEDLEY_MCST_LCC_VERSION)
+    #define JSON_HEDLEY_MCST_LCC_VERSION_CHECK(major,minor,patch) (JSON_HEDLEY_MCST_LCC_VERSION >= JSON_HEDLEY_VERSION_ENCODE(major, minor, patch))
+#else
+    #define JSON_HEDLEY_MCST_LCC_VERSION_CHECK(major,minor,patch) (0)
+#endif
+
+#if defined(JSON_HEDLEY_GCC_VERSION)
+    #undef JSON_HEDLEY_GCC_VERSION
+#endif
+#if \
+    defined(JSON_HEDLEY_GNUC_VERSION) && \
+    !defined(__clang__) && \
+    !defined(JSON_HEDLEY_INTEL_VERSION) && \
+    !defined(JSON_HEDLEY_PGI_VERSION) && \
+    !defined(JSON_HEDLEY_ARM_VERSION) && \
+    !defined(JSON_HEDLEY_CRAY_VERSION) && \
+    !defined(JSON_HEDLEY_TI_VERSION) && \
+    !defined(JSON_HEDLEY_TI_ARMCL_VERSION) && \
+    !defined(JSON_HEDLEY_TI_CL430_VERSION) && \
+    !defined(JSON_HEDLEY_TI_CL2000_VERSION) && \
+    !defined(JSON_HEDLEY_TI_CL6X_VERSION) && \
+    !defined(JSON_HEDLEY_TI_CL7X_VERSION) && \
+    !defined(JSON_HEDLEY_TI_CLPRU_VERSION) && \
+    !defined(__COMPCERT__) && \
+    !defined(JSON_HEDLEY_MCST_LCC_VERSION)
+    #define JSON_HEDLEY_GCC_VERSION JSON_HEDLEY_GNUC_VERSION
+#endif
+
+#if defined(JSON_HEDLEY_GCC_VERSION_CHECK)
+    #undef JSON_HEDLEY_GCC_VERSION_CHECK
+#endif
+#if defined(JSON_HEDLEY_GCC_VERSION)
+    #define JSON_HEDLEY_GCC_VERSION_CHECK(major,minor,patch) (JSON_HEDLEY_GCC_VERSION >= JSON_HEDLEY_VERSION_ENCODE(major, minor, patch))
+#else
+    #define JSON_HEDLEY_GCC_VERSION_CHECK(major,minor,patch) (0)
+#endif
+
+#if defined(JSON_HEDLEY_HAS_ATTRIBUTE)
+    #undef JSON_HEDLEY_HAS_ATTRIBUTE
+#endif
+#if \
+  defined(__has_attribute) && \
+  ( \
+    (!defined(JSON_HEDLEY_IAR_VERSION) || JSON_HEDLEY_IAR_VERSION_CHECK(8,5,9)) \
+  )
+#  define JSON_HEDLEY_HAS_ATTRIBUTE(attribute) __has_attribute(attribute)
+#else
+#  define JSON_HEDLEY_HAS_ATTRIBUTE(attribute) (0)
+#endif
+
+#if defined(JSON_HEDLEY_GNUC_HAS_ATTRIBUTE)
+    #undef JSON_HEDLEY_GNUC_HAS_ATTRIBUTE
+#endif
+#if defined(__has_attribute)
+    #define JSON_HEDLEY_GNUC_HAS_ATTRIBUTE(attribute,major,minor,patch) JSON_HEDLEY_HAS_ATTRIBUTE(attribute)
+#else
+    #define JSON_HEDLEY_GNUC_HAS_ATTRIBUTE(attribute,major,minor,patch) JSON_HEDLEY_GNUC_VERSION_CHECK(major,minor,patch)
+#endif
+
+#if defined(JSON_HEDLEY_GCC_HAS_ATTRIBUTE)
+    #undef JSON_HEDLEY_GCC_HAS_ATTRIBUTE
+#endif
+#if defined(__has_attribute)
+    #define JSON_HEDLEY_GCC_HAS_ATTRIBUTE(attribute,major,minor,patch) JSON_HEDLEY_HAS_ATTRIBUTE(attribute)
+#else
+    #define JSON_HEDLEY_GCC_HAS_ATTRIBUTE(attribute,major,minor,patch) JSON_HEDLEY_GCC_VERSION_CHECK(major,minor,patch)
+#endif
+
+#if defined(JSON_HEDLEY_HAS_CPP_ATTRIBUTE)
+    #undef JSON_HEDLEY_HAS_CPP_ATTRIBUTE
+#endif
+#if \
+    defined(__has_cpp_attribute) && \
+    defined(__cplusplus) && \
+    (!defined(JSON_HEDLEY_SUNPRO_VERSION) || JSON_HEDLEY_SUNPRO_VERSION_CHECK(5,15,0))
+    #define JSON_HEDLEY_HAS_CPP_ATTRIBUTE(attribute) __has_cpp_attribute(attribute)
+#else
+    #define JSON_HEDLEY_HAS_CPP_ATTRIBUTE(attribute) (0)
+#endif
+
+#if defined(JSON_HEDLEY_HAS_CPP_ATTRIBUTE_NS)
+    #undef JSON_HEDLEY_HAS_CPP_ATTRIBUTE_NS
+#endif
+#if !defined(__cplusplus) || !defined(__has_cpp_attribute)
+    #define JSON_HEDLEY_HAS_CPP_ATTRIBUTE_NS(ns,attribute) (0)
+#elif \
+    !defined(JSON_HEDLEY_PGI_VERSION) && \
+    !defined(JSON_HEDLEY_IAR_VERSION) && \
+    (!defined(JSON_HEDLEY_SUNPRO_VERSION) || JSON_HEDLEY_SUNPRO_VERSION_CHECK(5,15,0)) && \
+    (!defined(JSON_HEDLEY_MSVC_VERSION) || JSON_HEDLEY_MSVC_VERSION_CHECK(19,20,0))
+    #define JSON_HEDLEY_HAS_CPP_ATTRIBUTE_NS(ns,attribute) JSON_HEDLEY_HAS_CPP_ATTRIBUTE(ns::attribute)
+#else
+    #define JSON_HEDLEY_HAS_CPP_ATTRIBUTE_NS(ns,attribute) (0)
+#endif
+
+#if defined(JSON_HEDLEY_GNUC_HAS_CPP_ATTRIBUTE)
+    #undef JSON_HEDLEY_GNUC_HAS_CPP_ATTRIBUTE
+#endif
+#if defined(__has_cpp_attribute) && defined(__cplusplus)
+    #define JSON_HEDLEY_GNUC_HAS_CPP_ATTRIBUTE(attribute,major,minor,patch) __has_cpp_attribute(attribute)
+#else
+    #define JSON_HEDLEY_GNUC_HAS_CPP_ATTRIBUTE(attribute,major,minor,patch) JSON_HEDLEY_GNUC_VERSION_CHECK(major,minor,patch)
+#endif
+
+#if defined(JSON_HEDLEY_GCC_HAS_CPP_ATTRIBUTE)
+    #undef JSON_HEDLEY_GCC_HAS_CPP_ATTRIBUTE
+#endif
+#if defined(__has_cpp_attribute) && defined(__cplusplus)
+    #define JSON_HEDLEY_GCC_HAS_CPP_ATTRIBUTE(attribute,major,minor,patch) __has_cpp_attribute(attribute)
+#else
+    #define JSON_HEDLEY_GCC_HAS_CPP_ATTRIBUTE(attribute,major,minor,patch) JSON_HEDLEY_GCC_VERSION_CHECK(major,minor,patch)
+#endif
+
+#if defined(JSON_HEDLEY_HAS_BUILTIN)
+    #undef JSON_HEDLEY_HAS_BUILTIN
+#endif
+#if defined(__has_builtin)
+    #define JSON_HEDLEY_HAS_BUILTIN(builtin) __has_builtin(builtin)
+#else
+    #define JSON_HEDLEY_HAS_BUILTIN(builtin) (0)
+#endif
+
+#if defined(JSON_HEDLEY_GNUC_HAS_BUILTIN)
+    #undef JSON_HEDLEY_GNUC_HAS_BUILTIN
+#endif
+#if defined(__has_builtin)
+    #define JSON_HEDLEY_GNUC_HAS_BUILTIN(builtin,major,minor,patch) __has_builtin(builtin)
+#else
+    #define JSON_HEDLEY_GNUC_HAS_BUILTIN(builtin,major,minor,patch) JSON_HEDLEY_GNUC_VERSION_CHECK(major,minor,patch)
+#endif
+
+#if defined(JSON_HEDLEY_GCC_HAS_BUILTIN)
+    #undef JSON_HEDLEY_GCC_HAS_BUILTIN
+#endif
+#if defined(__has_builtin)
+    #define JSON_HEDLEY_GCC_HAS_BUILTIN(builtin,major,minor,patch) __has_builtin(builtin)
+#else
+    #define JSON_HEDLEY_GCC_HAS_BUILTIN(builtin,major,minor,patch) JSON_HEDLEY_GCC_VERSION_CHECK(major,minor,patch)
+#endif
+
+#if defined(JSON_HEDLEY_HAS_FEATURE)
+    #undef JSON_HEDLEY_HAS_FEATURE
+#endif
+#if defined(__has_feature)
+    #define JSON_HEDLEY_HAS_FEATURE(feature) __has_feature(feature)
+#else
+    #define JSON_HEDLEY_HAS_FEATURE(feature) (0)
+#endif
+
+#if defined(JSON_HEDLEY_GNUC_HAS_FEATURE)
+    #undef JSON_HEDLEY_GNUC_HAS_FEATURE
+#endif
+#if defined(__has_feature)
+    #define JSON_HEDLEY_GNUC_HAS_FEATURE(feature,major,minor,patch) __has_feature(feature)
+#else
+    #define JSON_HEDLEY_GNUC_HAS_FEATURE(feature,major,minor,patch) JSON_HEDLEY_GNUC_VERSION_CHECK(major,minor,patch)
+#endif
+
+#if defined(JSON_HEDLEY_GCC_HAS_FEATURE)
+    #undef JSON_HEDLEY_GCC_HAS_FEATURE
+#endif
+#if defined(__has_feature)
+    #define JSON_HEDLEY_GCC_HAS_FEATURE(feature,major,minor,patch) __has_feature(feature)
+#else
+    #define JSON_HEDLEY_GCC_HAS_FEATURE(feature,major,minor,patch) JSON_HEDLEY_GCC_VERSION_CHECK(major,minor,patch)
+#endif
+
+#if defined(JSON_HEDLEY_HAS_EXTENSION)
+    #undef JSON_HEDLEY_HAS_EXTENSION
+#endif
+#if defined(__has_extension)
+    #define JSON_HEDLEY_HAS_EXTENSION(extension) __has_extension(extension)
+#else
+    #define JSON_HEDLEY_HAS_EXTENSION(extension) (0)
+#endif
+
+#if defined(JSON_HEDLEY_GNUC_HAS_EXTENSION)
+    #undef JSON_HEDLEY_GNUC_HAS_EXTENSION
+#endif
+#if defined(__has_extension)
+    #define JSON_HEDLEY_GNUC_HAS_EXTENSION(extension,major,minor,patch) __has_extension(extension)
+#else
+    #define JSON_HEDLEY_GNUC_HAS_EXTENSION(extension,major,minor,patch) JSON_HEDLEY_GNUC_VERSION_CHECK(major,minor,patch)
+#endif
+
+#if defined(JSON_HEDLEY_GCC_HAS_EXTENSION)
+    #undef JSON_HEDLEY_GCC_HAS_EXTENSION
+#endif
+#if defined(__has_extension)
+    #define JSON_HEDLEY_GCC_HAS_EXTENSION(extension,major,minor,patch) __has_extension(extension)
+#else
+    #define JSON_HEDLEY_GCC_HAS_EXTENSION(extension,major,minor,patch) JSON_HEDLEY_GCC_VERSION_CHECK(major,minor,patch)
+#endif
+
+#if defined(JSON_HEDLEY_HAS_DECLSPEC_ATTRIBUTE)
+    #undef JSON_HEDLEY_HAS_DECLSPEC_ATTRIBUTE
+#endif
+#if defined(__has_declspec_attribute)
+    #define JSON_HEDLEY_HAS_DECLSPEC_ATTRIBUTE(attribute) __has_declspec_attribute(attribute)
+#else
+    #define JSON_HEDLEY_HAS_DECLSPEC_ATTRIBUTE(attribute) (0)
+#endif
+
+#if defined(JSON_HEDLEY_GNUC_HAS_DECLSPEC_ATTRIBUTE)
+    #undef JSON_HEDLEY_GNUC_HAS_DECLSPEC_ATTRIBUTE
+#endif
+#if defined(__has_declspec_attribute)
+    #define JSON_HEDLEY_GNUC_HAS_DECLSPEC_ATTRIBUTE(attribute,major,minor,patch) __has_declspec_attribute(attribute)
+#else
+    #define JSON_HEDLEY_GNUC_HAS_DECLSPEC_ATTRIBUTE(attribute,major,minor,patch) JSON_HEDLEY_GNUC_VERSION_CHECK(major,minor,patch)
+#endif
+
+#if defined(JSON_HEDLEY_GCC_HAS_DECLSPEC_ATTRIBUTE)
+    #undef JSON_HEDLEY_GCC_HAS_DECLSPEC_ATTRIBUTE
+#endif
+#if defined(__has_declspec_attribute)
+    #define JSON_HEDLEY_GCC_HAS_DECLSPEC_ATTRIBUTE(attribute,major,minor,patch) __has_declspec_attribute(attribute)
+#else
+    #define JSON_HEDLEY_GCC_HAS_DECLSPEC_ATTRIBUTE(attribute,major,minor,patch) JSON_HEDLEY_GCC_VERSION_CHECK(major,minor,patch)
+#endif
+
+#if defined(JSON_HEDLEY_HAS_WARNING)
+    #undef JSON_HEDLEY_HAS_WARNING
+#endif
+#if defined(__has_warning)
+    #define JSON_HEDLEY_HAS_WARNING(warning) __has_warning(warning)
+#else
+    #define JSON_HEDLEY_HAS_WARNING(warning) (0)
+#endif
+
+#if defined(JSON_HEDLEY_GNUC_HAS_WARNING)
+    #undef JSON_HEDLEY_GNUC_HAS_WARNING
+#endif
+#if defined(__has_warning)
+    #define JSON_HEDLEY_GNUC_HAS_WARNING(warning,major,minor,patch) __has_warning(warning)
+#else
+    #define JSON_HEDLEY_GNUC_HAS_WARNING(warning,major,minor,patch) JSON_HEDLEY_GNUC_VERSION_CHECK(major,minor,patch)
+#endif
+
+#if defined(JSON_HEDLEY_GCC_HAS_WARNING)
+    #undef JSON_HEDLEY_GCC_HAS_WARNING
+#endif
+#if defined(__has_warning)
+    #define JSON_HEDLEY_GCC_HAS_WARNING(warning,major,minor,patch) __has_warning(warning)
+#else
+    #define JSON_HEDLEY_GCC_HAS_WARNING(warning,major,minor,patch) JSON_HEDLEY_GCC_VERSION_CHECK(major,minor,patch)
+#endif
+
+#if \
+    (defined(__STDC_VERSION__) && (__STDC_VERSION__ >= 199901L)) || \
+    defined(__clang__) || \
+    JSON_HEDLEY_GCC_VERSION_CHECK(3,0,0) || \
+    JSON_HEDLEY_INTEL_VERSION_CHECK(13,0,0) || \
+    JSON_HEDLEY_IAR_VERSION_CHECK(8,0,0) || \
+    JSON_HEDLEY_PGI_VERSION_CHECK(18,4,0) || \
+    JSON_HEDLEY_ARM_VERSION_CHECK(4,1,0) || \
+    JSON_HEDLEY_TI_VERSION_CHECK(15,12,0) || \
+    JSON_HEDLEY_TI_ARMCL_VERSION_CHECK(4,7,0) || \
+    JSON_HEDLEY_TI_CL430_VERSION_CHECK(2,0,1) || \
+    JSON_HEDLEY_TI_CL2000_VERSION_CHECK(6,1,0) || \
+    JSON_HEDLEY_TI_CL6X_VERSION_CHECK(7,0,0) || \
+    JSON_HEDLEY_TI_CL7X_VERSION_CHECK(1,2,0) || \
+    JSON_HEDLEY_TI_CLPRU_VERSION_CHECK(2,1,0) || \
+    JSON_HEDLEY_CRAY_VERSION_CHECK(5,0,0) || \
+    JSON_HEDLEY_TINYC_VERSION_CHECK(0,9,17) || \
+    JSON_HEDLEY_SUNPRO_VERSION_CHECK(8,0,0) || \
+    (JSON_HEDLEY_IBM_VERSION_CHECK(10,1,0) && defined(__C99_PRAGMA_OPERATOR))
+    #define JSON_HEDLEY_PRAGMA(value) _Pragma(#value)
+#elif JSON_HEDLEY_MSVC_VERSION_CHECK(15,0,0)
+    #define JSON_HEDLEY_PRAGMA(value) __pragma(value)
+#else
+    #define JSON_HEDLEY_PRAGMA(value)
+#endif
+
+#if defined(JSON_HEDLEY_DIAGNOSTIC_PUSH)
+    #undef JSON_HEDLEY_DIAGNOSTIC_PUSH
+#endif
+#if defined(JSON_HEDLEY_DIAGNOSTIC_POP)
+    #undef JSON_HEDLEY_DIAGNOSTIC_POP
+#endif
+#if defined(__clang__)
+    #define JSON_HEDLEY_DIAGNOSTIC_PUSH _Pragma("clang diagnostic push")
+    #define JSON_HEDLEY_DIAGNOSTIC_POP _Pragma("clang diagnostic pop")
+#elif JSON_HEDLEY_INTEL_VERSION_CHECK(13,0,0)
+    #define JSON_HEDLEY_DIAGNOSTIC_PUSH _Pragma("warning(push)")
+    #define JSON_HEDLEY_DIAGNOSTIC_POP _Pragma("warning(pop)")
+#elif JSON_HEDLEY_GCC_VERSION_CHECK(4,6,0)
+    #define JSON_HEDLEY_DIAGNOSTIC_PUSH _Pragma("GCC diagnostic push")
+    #define JSON_HEDLEY_DIAGNOSTIC_POP _Pragma("GCC diagnostic pop")
+#elif \
+    JSON_HEDLEY_MSVC_VERSION_CHECK(15,0,0) || \
+    JSON_HEDLEY_INTEL_CL_VERSION_CHECK(2021,1,0)
+    #define JSON_HEDLEY_DIAGNOSTIC_PUSH __pragma(warning(push))
+    #define JSON_HEDLEY_DIAGNOSTIC_POP __pragma(warning(pop))
+#elif JSON_HEDLEY_ARM_VERSION_CHECK(5,6,0)
+    #define JSON_HEDLEY_DIAGNOSTIC_PUSH _Pragma("push")
+    #define JSON_HEDLEY_DIAGNOSTIC_POP _Pragma("pop")
+#elif \
+    JSON_HEDLEY_TI_VERSION_CHECK(15,12,0) || \
+    JSON_HEDLEY_TI_ARMCL_VERSION_CHECK(5,2,0) || \
+    JSON_HEDLEY_TI_CL430_VERSION_CHECK(4,4,0) || \
+    JSON_HEDLEY_TI_CL6X_VERSION_CHECK(8,1,0) || \
+    JSON_HEDLEY_TI_CL7X_VERSION_CHECK(1,2,0) || \
+    JSON_HEDLEY_TI_CLPRU_VERSION_CHECK(2,1,0)
+    #define JSON_HEDLEY_DIAGNOSTIC_PUSH _Pragma("diag_push")
+    #define JSON_HEDLEY_DIAGNOSTIC_POP _Pragma("diag_pop")
+#elif JSON_HEDLEY_PELLES_VERSION_CHECK(2,90,0)
+    #define JSON_HEDLEY_DIAGNOSTIC_PUSH _Pragma("warning(push)")
+    #define JSON_HEDLEY_DIAGNOSTIC_POP _Pragma("warning(pop)")
+#else
+    #define JSON_HEDLEY_DIAGNOSTIC_PUSH
+    #define JSON_HEDLEY_DIAGNOSTIC_POP
+#endif
+
+/* JSON_HEDLEY_DIAGNOSTIC_DISABLE_CPP98_COMPAT_WRAP_ is for
+   HEDLEY INTERNAL USE ONLY.  API subject to change without notice. */
+#if defined(JSON_HEDLEY_DIAGNOSTIC_DISABLE_CPP98_COMPAT_WRAP_)
+    #undef JSON_HEDLEY_DIAGNOSTIC_DISABLE_CPP98_COMPAT_WRAP_
+#endif
+#if defined(__cplusplus)
+#  if JSON_HEDLEY_HAS_WARNING("-Wc++98-compat")
+#    if JSON_HEDLEY_HAS_WARNING("-Wc++17-extensions")
+#      if JSON_HEDLEY_HAS_WARNING("-Wc++1z-extensions")
+#        define JSON_HEDLEY_DIAGNOSTIC_DISABLE_CPP98_COMPAT_WRAP_(xpr) \
+    JSON_HEDLEY_DIAGNOSTIC_PUSH \
+    _Pragma("clang diagnostic ignored \"-Wc++98-compat\"") \
+    _Pragma("clang diagnostic ignored \"-Wc++17-extensions\"") \
+    _Pragma("clang diagnostic ignored \"-Wc++1z-extensions\"") \
+    xpr \
+    JSON_HEDLEY_DIAGNOSTIC_POP
+#      else
+#        define JSON_HEDLEY_DIAGNOSTIC_DISABLE_CPP98_COMPAT_WRAP_(xpr) \
+    JSON_HEDLEY_DIAGNOSTIC_PUSH \
+    _Pragma("clang diagnostic ignored \"-Wc++98-compat\"") \
+    _Pragma("clang diagnostic ignored \"-Wc++17-extensions\"") \
+    xpr \
+    JSON_HEDLEY_DIAGNOSTIC_POP
+#      endif
+#    else
+#      define JSON_HEDLEY_DIAGNOSTIC_DISABLE_CPP98_COMPAT_WRAP_(xpr) \
+    JSON_HEDLEY_DIAGNOSTIC_PUSH \
+    _Pragma("clang diagnostic ignored \"-Wc++98-compat\"") \
+    xpr \
+    JSON_HEDLEY_DIAGNOSTIC_POP
+#    endif
+#  endif
+#endif
+#if !defined(JSON_HEDLEY_DIAGNOSTIC_DISABLE_CPP98_COMPAT_WRAP_)
+    #define JSON_HEDLEY_DIAGNOSTIC_DISABLE_CPP98_COMPAT_WRAP_(x) x
+#endif
+
+#if defined(JSON_HEDLEY_CONST_CAST)
+    #undef JSON_HEDLEY_CONST_CAST
+#endif
+#if defined(__cplusplus)
+#  define JSON_HEDLEY_CONST_CAST(T, expr) (const_cast<T>(expr))
+#elif \
+  JSON_HEDLEY_HAS_WARNING("-Wcast-qual") || \
+  JSON_HEDLEY_GCC_VERSION_CHECK(4,6,0) || \
+  JSON_HEDLEY_INTEL_VERSION_CHECK(13,0,0)
+#  define JSON_HEDLEY_CONST_CAST(T, expr) (__extension__ ({ \
+        JSON_HEDLEY_DIAGNOSTIC_PUSH \
+        JSON_HEDLEY_DIAGNOSTIC_DISABLE_CAST_QUAL \
+        ((T) (expr)); \
+        JSON_HEDLEY_DIAGNOSTIC_POP \
+    }))
+#else
+#  define JSON_HEDLEY_CONST_CAST(T, expr) ((T) (expr))
+#endif
+
+#if defined(JSON_HEDLEY_REINTERPRET_CAST)
+    #undef JSON_HEDLEY_REINTERPRET_CAST
+#endif
+#if defined(__cplusplus)
+    #define JSON_HEDLEY_REINTERPRET_CAST(T, expr) (reinterpret_cast<T>(expr))
+#else
+    #define JSON_HEDLEY_REINTERPRET_CAST(T, expr) ((T) (expr))
+#endif
+
+#if defined(JSON_HEDLEY_STATIC_CAST)
+    #undef JSON_HEDLEY_STATIC_CAST
+#endif
+#if defined(__cplusplus)
+    #define JSON_HEDLEY_STATIC_CAST(T, expr) (static_cast<T>(expr))
+#else
+    #define JSON_HEDLEY_STATIC_CAST(T, expr) ((T) (expr))
+#endif
+
+#if defined(JSON_HEDLEY_CPP_CAST)
+    #undef JSON_HEDLEY_CPP_CAST
+#endif
+#if defined(__cplusplus)
+#  if JSON_HEDLEY_HAS_WARNING("-Wold-style-cast")
+#    define JSON_HEDLEY_CPP_CAST(T, expr) \
+    JSON_HEDLEY_DIAGNOSTIC_PUSH \
+    _Pragma("clang diagnostic ignored \"-Wold-style-cast\"") \
+    ((T) (expr)) \
+    JSON_HEDLEY_DIAGNOSTIC_POP
+#  elif JSON_HEDLEY_IAR_VERSION_CHECK(8,3,0)
+#    define JSON_HEDLEY_CPP_CAST(T, expr) \
+    JSON_HEDLEY_DIAGNOSTIC_PUSH \
+    _Pragma("diag_suppress=Pe137") \
+    JSON_HEDLEY_DIAGNOSTIC_POP
+#  else
+#    define JSON_HEDLEY_CPP_CAST(T, expr) ((T) (expr))
+#  endif
+#else
+#  define JSON_HEDLEY_CPP_CAST(T, expr) (expr)
+#endif
+
+#if defined(JSON_HEDLEY_DIAGNOSTIC_DISABLE_DEPRECATED)
+    #undef JSON_HEDLEY_DIAGNOSTIC_DISABLE_DEPRECATED
+#endif
+#if JSON_HEDLEY_HAS_WARNING("-Wdeprecated-declarations")
+    #define JSON_HEDLEY_DIAGNOSTIC_DISABLE_DEPRECATED _Pragma("clang diagnostic ignored \"-Wdeprecated-declarations\"")
+#elif JSON_HEDLEY_INTEL_VERSION_CHECK(13,0,0)
+    #define JSON_HEDLEY_DIAGNOSTIC_DISABLE_DEPRECATED _Pragma("warning(disable:1478 1786)")
+#elif JSON_HEDLEY_INTEL_CL_VERSION_CHECK(2021,1,0)
+    #define JSON_HEDLEY_DIAGNOSTIC_DISABLE_DEPRECATED __pragma(warning(disable:1478 1786))
+#elif JSON_HEDLEY_PGI_VERSION_CHECK(20,7,0)
+    #define JSON_HEDLEY_DIAGNOSTIC_DISABLE_DEPRECATED _Pragma("diag_suppress 1215,1216,1444,1445")
+#elif JSON_HEDLEY_PGI_VERSION_CHECK(17,10,0)
+    #define JSON_HEDLEY_DIAGNOSTIC_DISABLE_DEPRECATED _Pragma("diag_suppress 1215,1444")
+#elif JSON_HEDLEY_GCC_VERSION_CHECK(4,3,0)
+    #define JSON_HEDLEY_DIAGNOSTIC_DISABLE_DEPRECATED _Pragma("GCC diagnostic ignored \"-Wdeprecated-declarations\"")
+#elif JSON_HEDLEY_MSVC_VERSION_CHECK(15,0,0)
+    #define JSON_HEDLEY_DIAGNOSTIC_DISABLE_DEPRECATED __pragma(warning(disable:4996))
+#elif JSON_HEDLEY_MCST_LCC_VERSION_CHECK(1,25,10)
+    #define JSON_HEDLEY_DIAGNOSTIC_DISABLE_DEPRECATED _Pragma("diag_suppress 1215,1444")
+#elif \
+    JSON_HEDLEY_TI_VERSION_CHECK(15,12,0) || \
+    (JSON_HEDLEY_TI_ARMCL_VERSION_CHECK(4,8,0) && defined(__TI_GNU_ATTRIBUTE_SUPPORT__)) || \
+    JSON_HEDLEY_TI_ARMCL_VERSION_CHECK(5,2,0) || \
+    (JSON_HEDLEY_TI_CL2000_VERSION_CHECK(6,0,0) && defined(__TI_GNU_ATTRIBUTE_SUPPORT__)) || \
+    JSON_HEDLEY_TI_CL2000_VERSION_CHECK(6,4,0) || \
+    (JSON_HEDLEY_TI_CL430_VERSION_CHECK(4,0,0) && defined(__TI_GNU_ATTRIBUTE_SUPPORT__)) || \
+    JSON_HEDLEY_TI_CL430_VERSION_CHECK(4,3,0) || \
+    (JSON_HEDLEY_TI_CL6X_VERSION_CHECK(7,2,0) && defined(__TI_GNU_ATTRIBUTE_SUPPORT__)) || \
+    JSON_HEDLEY_TI_CL6X_VERSION_CHECK(7,5,0) || \
+    JSON_HEDLEY_TI_CL7X_VERSION_CHECK(1,2,0) || \
+    JSON_HEDLEY_TI_CLPRU_VERSION_CHECK(2,1,0)
+    #define JSON_HEDLEY_DIAGNOSTIC_DISABLE_DEPRECATED _Pragma("diag_suppress 1291,1718")
+#elif JSON_HEDLEY_SUNPRO_VERSION_CHECK(5,13,0) && !defined(__cplusplus)
+    #define JSON_HEDLEY_DIAGNOSTIC_DISABLE_DEPRECATED _Pragma("error_messages(off,E_DEPRECATED_ATT,E_DEPRECATED_ATT_MESS)")
+#elif JSON_HEDLEY_SUNPRO_VERSION_CHECK(5,13,0) && defined(__cplusplus)
+    #define JSON_HEDLEY_DIAGNOSTIC_DISABLE_DEPRECATED _Pragma("error_messages(off,symdeprecated,symdeprecated2)")
+#elif JSON_HEDLEY_IAR_VERSION_CHECK(8,0,0)
+    #define JSON_HEDLEY_DIAGNOSTIC_DISABLE_DEPRECATED _Pragma("diag_suppress=Pe1444,Pe1215")
+#elif JSON_HEDLEY_PELLES_VERSION_CHECK(2,90,0)
+    #define JSON_HEDLEY_DIAGNOSTIC_DISABLE_DEPRECATED _Pragma("warn(disable:2241)")
+#else
+    #define JSON_HEDLEY_DIAGNOSTIC_DISABLE_DEPRECATED
+#endif
+
+#if defined(JSON_HEDLEY_DIAGNOSTIC_DISABLE_UNKNOWN_PRAGMAS)
+    #undef JSON_HEDLEY_DIAGNOSTIC_DISABLE_UNKNOWN_PRAGMAS
+#endif
+#if JSON_HEDLEY_HAS_WARNING("-Wunknown-pragmas")
+    #define JSON_HEDLEY_DIAGNOSTIC_DISABLE_UNKNOWN_PRAGMAS _Pragma("clang diagnostic ignored \"-Wunknown-pragmas\"")
+#elif JSON_HEDLEY_INTEL_VERSION_CHECK(13,0,0)
+    #define JSON_HEDLEY_DIAGNOSTIC_DISABLE_UNKNOWN_PRAGMAS _Pragma("warning(disable:161)")
+#elif JSON_HEDLEY_INTEL_CL_VERSION_CHECK(2021,1,0)
+    #define JSON_HEDLEY_DIAGNOSTIC_DISABLE_UNKNOWN_PRAGMAS __pragma(warning(disable:161))
+#elif JSON_HEDLEY_PGI_VERSION_CHECK(17,10,0)
+    #define JSON_HEDLEY_DIAGNOSTIC_DISABLE_UNKNOWN_PRAGMAS _Pragma("diag_suppress 1675")
+#elif JSON_HEDLEY_GCC_VERSION_CHECK(4,3,0)
+    #define JSON_HEDLEY_DIAGNOSTIC_DISABLE_UNKNOWN_PRAGMAS _Pragma("GCC diagnostic ignored \"-Wunknown-pragmas\"")
+#elif JSON_HEDLEY_MSVC_VERSION_CHECK(15,0,0)
+    #define JSON_HEDLEY_DIAGNOSTIC_DISABLE_UNKNOWN_PRAGMAS __pragma(warning(disable:4068))
+#elif \
+    JSON_HEDLEY_TI_VERSION_CHECK(16,9,0) || \
+    JSON_HEDLEY_TI_CL6X_VERSION_CHECK(8,0,0) || \
+    JSON_HEDLEY_TI_CL7X_VERSION_CHECK(1,2,0) || \
+    JSON_HEDLEY_TI_CLPRU_VERSION_CHECK(2,3,0)
+    #define JSON_HEDLEY_DIAGNOSTIC_DISABLE_UNKNOWN_PRAGMAS _Pragma("diag_suppress 163")
+#elif JSON_HEDLEY_TI_CL6X_VERSION_CHECK(8,0,0)
+    #define JSON_HEDLEY_DIAGNOSTIC_DISABLE_UNKNOWN_PRAGMAS _Pragma("diag_suppress 163")
+#elif JSON_HEDLEY_IAR_VERSION_CHECK(8,0,0)
+    #define JSON_HEDLEY_DIAGNOSTIC_DISABLE_UNKNOWN_PRAGMAS _Pragma("diag_suppress=Pe161")
+#elif JSON_HEDLEY_MCST_LCC_VERSION_CHECK(1,25,10)
+    #define JSON_HEDLEY_DIAGNOSTIC_DISABLE_UNKNOWN_PRAGMAS _Pragma("diag_suppress 161")
+#else
+    #define JSON_HEDLEY_DIAGNOSTIC_DISABLE_UNKNOWN_PRAGMAS
+#endif
+
+#if defined(JSON_HEDLEY_DIAGNOSTIC_DISABLE_UNKNOWN_CPP_ATTRIBUTES)
+    #undef JSON_HEDLEY_DIAGNOSTIC_DISABLE_UNKNOWN_CPP_ATTRIBUTES
+#endif
+#if JSON_HEDLEY_HAS_WARNING("-Wunknown-attributes")
+    #define JSON_HEDLEY_DIAGNOSTIC_DISABLE_UNKNOWN_CPP_ATTRIBUTES _Pragma("clang diagnostic ignored \"-Wunknown-attributes\"")
+#elif JSON_HEDLEY_GCC_VERSION_CHECK(4,6,0)
+    #define JSON_HEDLEY_DIAGNOSTIC_DISABLE_UNKNOWN_CPP_ATTRIBUTES _Pragma("GCC diagnostic ignored \"-Wdeprecated-declarations\"")
+#elif JSON_HEDLEY_INTEL_VERSION_CHECK(17,0,0)
+    #define JSON_HEDLEY_DIAGNOSTIC_DISABLE_UNKNOWN_CPP_ATTRIBUTES _Pragma("warning(disable:1292)")
+#elif JSON_HEDLEY_INTEL_CL_VERSION_CHECK(2021,1,0)
+    #define JSON_HEDLEY_DIAGNOSTIC_DISABLE_UNKNOWN_CPP_ATTRIBUTES __pragma(warning(disable:1292))
+#elif JSON_HEDLEY_MSVC_VERSION_CHECK(19,0,0)
+    #define JSON_HEDLEY_DIAGNOSTIC_DISABLE_UNKNOWN_CPP_ATTRIBUTES __pragma(warning(disable:5030))
+#elif JSON_HEDLEY_PGI_VERSION_CHECK(20,7,0)
+    #define JSON_HEDLEY_DIAGNOSTIC_DISABLE_UNKNOWN_CPP_ATTRIBUTES _Pragma("diag_suppress 1097,1098")
+#elif JSON_HEDLEY_PGI_VERSION_CHECK(17,10,0)
+    #define JSON_HEDLEY_DIAGNOSTIC_DISABLE_UNKNOWN_CPP_ATTRIBUTES _Pragma("diag_suppress 1097")
+#elif JSON_HEDLEY_SUNPRO_VERSION_CHECK(5,14,0) && defined(__cplusplus)
+    #define JSON_HEDLEY_DIAGNOSTIC_DISABLE_UNKNOWN_CPP_ATTRIBUTES _Pragma("error_messages(off,attrskipunsup)")
+#elif \
+    JSON_HEDLEY_TI_VERSION_CHECK(18,1,0) || \
+    JSON_HEDLEY_TI_CL6X_VERSION_CHECK(8,3,0) || \
+    JSON_HEDLEY_TI_CL7X_VERSION_CHECK(1,2,0)
+    #define JSON_HEDLEY_DIAGNOSTIC_DISABLE_UNKNOWN_CPP_ATTRIBUTES _Pragma("diag_suppress 1173")
+#elif JSON_HEDLEY_IAR_VERSION_CHECK(8,0,0)
+    #define JSON_HEDLEY_DIAGNOSTIC_DISABLE_UNKNOWN_CPP_ATTRIBUTES _Pragma("diag_suppress=Pe1097")
+#elif JSON_HEDLEY_MCST_LCC_VERSION_CHECK(1,25,10)
+    #define JSON_HEDLEY_DIAGNOSTIC_DISABLE_UNKNOWN_CPP_ATTRIBUTES _Pragma("diag_suppress 1097")
+#else
+    #define JSON_HEDLEY_DIAGNOSTIC_DISABLE_UNKNOWN_CPP_ATTRIBUTES
+#endif
+
+#if defined(JSON_HEDLEY_DIAGNOSTIC_DISABLE_CAST_QUAL)
+    #undef JSON_HEDLEY_DIAGNOSTIC_DISABLE_CAST_QUAL
+#endif
+#if JSON_HEDLEY_HAS_WARNING("-Wcast-qual")
+    #define JSON_HEDLEY_DIAGNOSTIC_DISABLE_CAST_QUAL _Pragma("clang diagnostic ignored \"-Wcast-qual\"")
+#elif JSON_HEDLEY_INTEL_VERSION_CHECK(13,0,0)
+    #define JSON_HEDLEY_DIAGNOSTIC_DISABLE_CAST_QUAL _Pragma("warning(disable:2203 2331)")
+#elif JSON_HEDLEY_GCC_VERSION_CHECK(3,0,0)
+    #define JSON_HEDLEY_DIAGNOSTIC_DISABLE_CAST_QUAL _Pragma("GCC diagnostic ignored \"-Wcast-qual\"")
+#else
+    #define JSON_HEDLEY_DIAGNOSTIC_DISABLE_CAST_QUAL
+#endif
+
+#if defined(JSON_HEDLEY_DIAGNOSTIC_DISABLE_UNUSED_FUNCTION)
+    #undef JSON_HEDLEY_DIAGNOSTIC_DISABLE_UNUSED_FUNCTION
+#endif
+#if JSON_HEDLEY_HAS_WARNING("-Wunused-function")
+    #define JSON_HEDLEY_DIAGNOSTIC_DISABLE_UNUSED_FUNCTION _Pragma("clang diagnostic ignored \"-Wunused-function\"")
+#elif JSON_HEDLEY_GCC_VERSION_CHECK(3,4,0)
+    #define JSON_HEDLEY_DIAGNOSTIC_DISABLE_UNUSED_FUNCTION _Pragma("GCC diagnostic ignored \"-Wunused-function\"")
+#elif JSON_HEDLEY_MSVC_VERSION_CHECK(1,0,0)
+    #define JSON_HEDLEY_DIAGNOSTIC_DISABLE_UNUSED_FUNCTION __pragma(warning(disable:4505))
+#elif JSON_HEDLEY_MCST_LCC_VERSION_CHECK(1,25,10)
+    #define JSON_HEDLEY_DIAGNOSTIC_DISABLE_UNUSED_FUNCTION _Pragma("diag_suppress 3142")
+#else
+    #define JSON_HEDLEY_DIAGNOSTIC_DISABLE_UNUSED_FUNCTION
+#endif
+
+#if defined(JSON_HEDLEY_DEPRECATED)
+    #undef JSON_HEDLEY_DEPRECATED
+#endif
+#if defined(JSON_HEDLEY_DEPRECATED_FOR)
+    #undef JSON_HEDLEY_DEPRECATED_FOR
+#endif
+#if \
+    JSON_HEDLEY_MSVC_VERSION_CHECK(14,0,0) || \
+    JSON_HEDLEY_INTEL_CL_VERSION_CHECK(2021,1,0)
+    #define JSON_HEDLEY_DEPRECATED(since) __declspec(deprecated("Since " # since))
+    #define JSON_HEDLEY_DEPRECATED_FOR(since, replacement) __declspec(deprecated("Since " #since "; use " #replacement))
+#elif \
+    (JSON_HEDLEY_HAS_EXTENSION(attribute_deprecated_with_message) && !defined(JSON_HEDLEY_IAR_VERSION)) || \
+    JSON_HEDLEY_GCC_VERSION_CHECK(4,5,0) || \
+    JSON_HEDLEY_INTEL_VERSION_CHECK(13,0,0) || \
+    JSON_HEDLEY_ARM_VERSION_CHECK(5,6,0) || \
+    JSON_HEDLEY_SUNPRO_VERSION_CHECK(5,13,0) || \
+    JSON_HEDLEY_PGI_VERSION_CHECK(17,10,0) || \
+    JSON_HEDLEY_TI_VERSION_CHECK(18,1,0) || \
+    JSON_HEDLEY_TI_ARMCL_VERSION_CHECK(18,1,0) || \
+    JSON_HEDLEY_TI_CL6X_VERSION_CHECK(8,3,0) || \
+    JSON_HEDLEY_TI_CL7X_VERSION_CHECK(1,2,0) || \
+    JSON_HEDLEY_TI_CLPRU_VERSION_CHECK(2,3,0) || \
+    JSON_HEDLEY_MCST_LCC_VERSION_CHECK(1,25,10)
+    #define JSON_HEDLEY_DEPRECATED(since) __attribute__((__deprecated__("Since " #since)))
+    #define JSON_HEDLEY_DEPRECATED_FOR(since, replacement) __attribute__((__deprecated__("Since " #since "; use " #replacement)))
+#elif defined(__cplusplus) && (__cplusplus >= 201402L)
+    #define JSON_HEDLEY_DEPRECATED(since) JSON_HEDLEY_DIAGNOSTIC_DISABLE_CPP98_COMPAT_WRAP_([[deprecated("Since " #since)]])
+    #define JSON_HEDLEY_DEPRECATED_FOR(since, replacement) JSON_HEDLEY_DIAGNOSTIC_DISABLE_CPP98_COMPAT_WRAP_([[deprecated("Since " #since "; use " #replacement)]])
+#elif \
+    JSON_HEDLEY_HAS_ATTRIBUTE(deprecated) || \
+    JSON_HEDLEY_GCC_VERSION_CHECK(3,1,0) || \
+    JSON_HEDLEY_ARM_VERSION_CHECK(4,1,0) || \
+    JSON_HEDLEY_TI_VERSION_CHECK(15,12,0) || \
+    (JSON_HEDLEY_TI_ARMCL_VERSION_CHECK(4,8,0) && defined(__TI_GNU_ATTRIBUTE_SUPPORT__)) || \
+    JSON_HEDLEY_TI_ARMCL_VERSION_CHECK(5,2,0) || \
+    (JSON_HEDLEY_TI_CL2000_VERSION_CHECK(6,0,0) && defined(__TI_GNU_ATTRIBUTE_SUPPORT__)) || \
+    JSON_HEDLEY_TI_CL2000_VERSION_CHECK(6,4,0) || \
+    (JSON_HEDLEY_TI_CL430_VERSION_CHECK(4,0,0) && defined(__TI_GNU_ATTRIBUTE_SUPPORT__)) || \
+    JSON_HEDLEY_TI_CL430_VERSION_CHECK(4,3,0) || \
+    (JSON_HEDLEY_TI_CL6X_VERSION_CHECK(7,2,0) && defined(__TI_GNU_ATTRIBUTE_SUPPORT__)) || \
+    JSON_HEDLEY_TI_CL6X_VERSION_CHECK(7,5,0) || \
+    JSON_HEDLEY_TI_CL7X_VERSION_CHECK(1,2,0) || \
+    JSON_HEDLEY_TI_CLPRU_VERSION_CHECK(2,1,0) || \
+    JSON_HEDLEY_MCST_LCC_VERSION_CHECK(1,25,10) || \
+    JSON_HEDLEY_IAR_VERSION_CHECK(8,10,0)
+    #define JSON_HEDLEY_DEPRECATED(since) __attribute__((__deprecated__))
+    #define JSON_HEDLEY_DEPRECATED_FOR(since, replacement) __attribute__((__deprecated__))
+#elif \
+    JSON_HEDLEY_MSVC_VERSION_CHECK(13,10,0) || \
+    JSON_HEDLEY_PELLES_VERSION_CHECK(6,50,0) || \
+    JSON_HEDLEY_INTEL_CL_VERSION_CHECK(2021,1,0)
+    #define JSON_HEDLEY_DEPRECATED(since) __declspec(deprecated)
+    #define JSON_HEDLEY_DEPRECATED_FOR(since, replacement) __declspec(deprecated)
+#elif JSON_HEDLEY_IAR_VERSION_CHECK(8,0,0)
+    #define JSON_HEDLEY_DEPRECATED(since) _Pragma("deprecated")
+    #define JSON_HEDLEY_DEPRECATED_FOR(since, replacement) _Pragma("deprecated")
+#else
+    #define JSON_HEDLEY_DEPRECATED(since)
+    #define JSON_HEDLEY_DEPRECATED_FOR(since, replacement)
+#endif
+
+#if defined(JSON_HEDLEY_UNAVAILABLE)
+    #undef JSON_HEDLEY_UNAVAILABLE
+#endif
+#if \
+    JSON_HEDLEY_HAS_ATTRIBUTE(warning) || \
+    JSON_HEDLEY_GCC_VERSION_CHECK(4,3,0) || \
+    JSON_HEDLEY_INTEL_VERSION_CHECK(13,0,0) || \
+    JSON_HEDLEY_MCST_LCC_VERSION_CHECK(1,25,10)
+    #define JSON_HEDLEY_UNAVAILABLE(available_since) __attribute__((__warning__("Not available until " #available_since)))
+#else
+    #define JSON_HEDLEY_UNAVAILABLE(available_since)
+#endif
+
+#if defined(JSON_HEDLEY_WARN_UNUSED_RESULT)
+    #undef JSON_HEDLEY_WARN_UNUSED_RESULT
+#endif
+#if defined(JSON_HEDLEY_WARN_UNUSED_RESULT_MSG)
+    #undef JSON_HEDLEY_WARN_UNUSED_RESULT_MSG
+#endif
+#if \
+    JSON_HEDLEY_HAS_ATTRIBUTE(warn_unused_result) || \
+    JSON_HEDLEY_GCC_VERSION_CHECK(3,4,0) || \
+    JSON_HEDLEY_INTEL_VERSION_CHECK(13,0,0) || \
+    JSON_HEDLEY_TI_VERSION_CHECK(15,12,0) || \
+    (JSON_HEDLEY_TI_ARMCL_VERSION_CHECK(4,8,0) && defined(__TI_GNU_ATTRIBUTE_SUPPORT__)) || \
+    JSON_HEDLEY_TI_ARMCL_VERSION_CHECK(5,2,0) || \
+    (JSON_HEDLEY_TI_CL2000_VERSION_CHECK(6,0,0) && defined(__TI_GNU_ATTRIBUTE_SUPPORT__)) || \
+    JSON_HEDLEY_TI_CL2000_VERSION_CHECK(6,4,0) || \
+    (JSON_HEDLEY_TI_CL430_VERSION_CHECK(4,0,0) && defined(__TI_GNU_ATTRIBUTE_SUPPORT__)) || \
+    JSON_HEDLEY_TI_CL430_VERSION_CHECK(4,3,0) || \
+    (JSON_HEDLEY_TI_CL6X_VERSION_CHECK(7,2,0) && defined(__TI_GNU_ATTRIBUTE_SUPPORT__)) || \
+    JSON_HEDLEY_TI_CL6X_VERSION_CHECK(7,5,0) || \
+    JSON_HEDLEY_TI_CL7X_VERSION_CHECK(1,2,0) || \
+    JSON_HEDLEY_TI_CLPRU_VERSION_CHECK(2,1,0) || \
+    (JSON_HEDLEY_SUNPRO_VERSION_CHECK(5,15,0) && defined(__cplusplus)) || \
+    JSON_HEDLEY_PGI_VERSION_CHECK(17,10,0) || \
+    JSON_HEDLEY_MCST_LCC_VERSION_CHECK(1,25,10)
+    #define JSON_HEDLEY_WARN_UNUSED_RESULT __attribute__((__warn_unused_result__))
+    #define JSON_HEDLEY_WARN_UNUSED_RESULT_MSG(msg) __attribute__((__warn_unused_result__))
+#elif (JSON_HEDLEY_HAS_CPP_ATTRIBUTE(nodiscard) >= 201907L)
+    #define JSON_HEDLEY_WARN_UNUSED_RESULT JSON_HEDLEY_DIAGNOSTIC_DISABLE_CPP98_COMPAT_WRAP_([[nodiscard]])
+    #define JSON_HEDLEY_WARN_UNUSED_RESULT_MSG(msg) JSON_HEDLEY_DIAGNOSTIC_DISABLE_CPP98_COMPAT_WRAP_([[nodiscard(msg)]])
+#elif JSON_HEDLEY_HAS_CPP_ATTRIBUTE(nodiscard)
+    #define JSON_HEDLEY_WARN_UNUSED_RESULT JSON_HEDLEY_DIAGNOSTIC_DISABLE_CPP98_COMPAT_WRAP_([[nodiscard]])
+    #define JSON_HEDLEY_WARN_UNUSED_RESULT_MSG(msg) JSON_HEDLEY_DIAGNOSTIC_DISABLE_CPP98_COMPAT_WRAP_([[nodiscard]])
+#elif defined(_Check_return_) /* SAL */
+    #define JSON_HEDLEY_WARN_UNUSED_RESULT _Check_return_
+    #define JSON_HEDLEY_WARN_UNUSED_RESULT_MSG(msg) _Check_return_
+#else
+    #define JSON_HEDLEY_WARN_UNUSED_RESULT
+    #define JSON_HEDLEY_WARN_UNUSED_RESULT_MSG(msg)
+#endif
+
+#if defined(JSON_HEDLEY_SENTINEL)
+    #undef JSON_HEDLEY_SENTINEL
+#endif
+#if \
+    JSON_HEDLEY_HAS_ATTRIBUTE(sentinel) || \
+    JSON_HEDLEY_GCC_VERSION_CHECK(4,0,0) || \
+    JSON_HEDLEY_INTEL_VERSION_CHECK(13,0,0) || \
+    JSON_HEDLEY_ARM_VERSION_CHECK(5,4,0) || \
+    JSON_HEDLEY_MCST_LCC_VERSION_CHECK(1,25,10)
+    #define JSON_HEDLEY_SENTINEL(position) __attribute__((__sentinel__(position)))
+#else
+    #define JSON_HEDLEY_SENTINEL(position)
+#endif
+
+#if defined(JSON_HEDLEY_NO_RETURN)
+    #undef JSON_HEDLEY_NO_RETURN
+#endif
+#if JSON_HEDLEY_IAR_VERSION_CHECK(8,0,0)
+    #define JSON_HEDLEY_NO_RETURN __noreturn
+#elif \
+    JSON_HEDLEY_INTEL_VERSION_CHECK(13,0,0) || \
+    JSON_HEDLEY_MCST_LCC_VERSION_CHECK(1,25,10)
+    #define JSON_HEDLEY_NO_RETURN __attribute__((__noreturn__))
+#elif defined(__STDC_VERSION__) && __STDC_VERSION__ >= 201112L
+    #define JSON_HEDLEY_NO_RETURN _Noreturn
+#elif defined(__cplusplus) && (__cplusplus >= 201103L)
+    #define JSON_HEDLEY_NO_RETURN JSON_HEDLEY_DIAGNOSTIC_DISABLE_CPP98_COMPAT_WRAP_([[noreturn]])
+#elif \
+    JSON_HEDLEY_HAS_ATTRIBUTE(noreturn) || \
+    JSON_HEDLEY_GCC_VERSION_CHECK(3,2,0) || \
+    JSON_HEDLEY_SUNPRO_VERSION_CHECK(5,11,0) || \
+    JSON_HEDLEY_ARM_VERSION_CHECK(4,1,0) || \
+    JSON_HEDLEY_IBM_VERSION_CHECK(10,1,0) || \
+    JSON_HEDLEY_TI_VERSION_CHECK(15,12,0) || \
+    (JSON_HEDLEY_TI_ARMCL_VERSION_CHECK(4,8,0) && defined(__TI_GNU_ATTRIBUTE_SUPPORT__)) || \
+    JSON_HEDLEY_TI_ARMCL_VERSION_CHECK(5,2,0) || \
+    (JSON_HEDLEY_TI_CL2000_VERSION_CHECK(6,0,0) && defined(__TI_GNU_ATTRIBUTE_SUPPORT__)) || \
+    JSON_HEDLEY_TI_CL2000_VERSION_CHECK(6,4,0) || \
+    (JSON_HEDLEY_TI_CL430_VERSION_CHECK(4,0,0) && defined(__TI_GNU_ATTRIBUTE_SUPPORT__)) || \
+    JSON_HEDLEY_TI_CL430_VERSION_CHECK(4,3,0) || \
+    (JSON_HEDLEY_TI_CL6X_VERSION_CHECK(7,2,0) && defined(__TI_GNU_ATTRIBUTE_SUPPORT__)) || \
+    JSON_HEDLEY_TI_CL6X_VERSION_CHECK(7,5,0) || \
+    JSON_HEDLEY_TI_CL7X_VERSION_CHECK(1,2,0) || \
+    JSON_HEDLEY_TI_CLPRU_VERSION_CHECK(2,1,0) || \
+    JSON_HEDLEY_IAR_VERSION_CHECK(8,10,0)
+    #define JSON_HEDLEY_NO_RETURN __attribute__((__noreturn__))
+#elif JSON_HEDLEY_SUNPRO_VERSION_CHECK(5,10,0)
+    #define JSON_HEDLEY_NO_RETURN _Pragma("does_not_return")
+#elif \
+    JSON_HEDLEY_MSVC_VERSION_CHECK(13,10,0) || \
+    JSON_HEDLEY_INTEL_CL_VERSION_CHECK(2021,1,0)
+    #define JSON_HEDLEY_NO_RETURN __declspec(noreturn)
+#elif JSON_HEDLEY_TI_CL6X_VERSION_CHECK(6,0,0) && defined(__cplusplus)
+    #define JSON_HEDLEY_NO_RETURN _Pragma("FUNC_NEVER_RETURNS;")
+#elif JSON_HEDLEY_COMPCERT_VERSION_CHECK(3,2,0)
+    #define JSON_HEDLEY_NO_RETURN __attribute((noreturn))
+#elif JSON_HEDLEY_PELLES_VERSION_CHECK(9,0,0)
+    #define JSON_HEDLEY_NO_RETURN __declspec(noreturn)
+#else
+    #define JSON_HEDLEY_NO_RETURN
+#endif
+
+#if defined(JSON_HEDLEY_NO_ESCAPE)
+    #undef JSON_HEDLEY_NO_ESCAPE
+#endif
+#if JSON_HEDLEY_HAS_ATTRIBUTE(noescape)
+    #define JSON_HEDLEY_NO_ESCAPE __attribute__((__noescape__))
+#else
+    #define JSON_HEDLEY_NO_ESCAPE
+#endif
+
+#if defined(JSON_HEDLEY_UNREACHABLE)
+    #undef JSON_HEDLEY_UNREACHABLE
+#endif
+#if defined(JSON_HEDLEY_UNREACHABLE_RETURN)
+    #undef JSON_HEDLEY_UNREACHABLE_RETURN
+#endif
+#if defined(JSON_HEDLEY_ASSUME)
+    #undef JSON_HEDLEY_ASSUME
+#endif
+#if \
+    JSON_HEDLEY_MSVC_VERSION_CHECK(13,10,0) || \
+    JSON_HEDLEY_INTEL_VERSION_CHECK(13,0,0) || \
+    JSON_HEDLEY_INTEL_CL_VERSION_CHECK(2021,1,0)
+    #define JSON_HEDLEY_ASSUME(expr) __assume(expr)
+#elif JSON_HEDLEY_HAS_BUILTIN(__builtin_assume)
+    #define JSON_HEDLEY_ASSUME(expr) __builtin_assume(expr)
+#elif \
+    JSON_HEDLEY_TI_CL2000_VERSION_CHECK(6,2,0) || \
+    JSON_HEDLEY_TI_CL6X_VERSION_CHECK(4,0,0)
+    #if defined(__cplusplus)
+        #define JSON_HEDLEY_ASSUME(expr) std::_nassert(expr)
+    #else
+        #define JSON_HEDLEY_ASSUME(expr) _nassert(expr)
+    #endif
+#endif
+#if \
+    (JSON_HEDLEY_HAS_BUILTIN(__builtin_unreachable) && (!defined(JSON_HEDLEY_ARM_VERSION))) || \
+    JSON_HEDLEY_GCC_VERSION_CHECK(4,5,0) || \
+    JSON_HEDLEY_PGI_VERSION_CHECK(18,10,0) || \
+    JSON_HEDLEY_INTEL_VERSION_CHECK(13,0,0) || \
+    JSON_HEDLEY_IBM_VERSION_CHECK(13,1,5) || \
+    JSON_HEDLEY_CRAY_VERSION_CHECK(10,0,0) || \
+    JSON_HEDLEY_MCST_LCC_VERSION_CHECK(1,25,10)
+    #define JSON_HEDLEY_UNREACHABLE() __builtin_unreachable()
+#elif defined(JSON_HEDLEY_ASSUME)
+    #define JSON_HEDLEY_UNREACHABLE() JSON_HEDLEY_ASSUME(0)
+#endif
+#if !defined(JSON_HEDLEY_ASSUME)
+    #if defined(JSON_HEDLEY_UNREACHABLE)
+        #define JSON_HEDLEY_ASSUME(expr) JSON_HEDLEY_STATIC_CAST(void, ((expr) ? 1 : (JSON_HEDLEY_UNREACHABLE(), 1)))
+    #else
+        #define JSON_HEDLEY_ASSUME(expr) JSON_HEDLEY_STATIC_CAST(void, expr)
+    #endif
+#endif
+#if defined(JSON_HEDLEY_UNREACHABLE)
+    #if  \
+        JSON_HEDLEY_TI_CL2000_VERSION_CHECK(6,2,0) || \
+        JSON_HEDLEY_TI_CL6X_VERSION_CHECK(4,0,0)
+        #define JSON_HEDLEY_UNREACHABLE_RETURN(value) return (JSON_HEDLEY_STATIC_CAST(void, JSON_HEDLEY_ASSUME(0)), (value))
+    #else
+        #define JSON_HEDLEY_UNREACHABLE_RETURN(value) JSON_HEDLEY_UNREACHABLE()
+    #endif
+#else
+    #define JSON_HEDLEY_UNREACHABLE_RETURN(value) return (value)
+#endif
+#if !defined(JSON_HEDLEY_UNREACHABLE)
+    #define JSON_HEDLEY_UNREACHABLE() JSON_HEDLEY_ASSUME(0)
+#endif
+
+JSON_HEDLEY_DIAGNOSTIC_PUSH
+#if JSON_HEDLEY_HAS_WARNING("-Wpedantic")
+    #pragma clang diagnostic ignored "-Wpedantic"
+#endif
+#if JSON_HEDLEY_HAS_WARNING("-Wc++98-compat-pedantic") && defined(__cplusplus)
+    #pragma clang diagnostic ignored "-Wc++98-compat-pedantic"
+#endif
+#if JSON_HEDLEY_GCC_HAS_WARNING("-Wvariadic-macros",4,0,0)
+    #if defined(__clang__)
+        #pragma clang diagnostic ignored "-Wvariadic-macros"
+    #elif defined(JSON_HEDLEY_GCC_VERSION)
+        #pragma GCC diagnostic ignored "-Wvariadic-macros"
+    #endif
+#endif
+#if defined(JSON_HEDLEY_NON_NULL)
+    #undef JSON_HEDLEY_NON_NULL
+#endif
+#if \
+    JSON_HEDLEY_HAS_ATTRIBUTE(nonnull) || \
+    JSON_HEDLEY_GCC_VERSION_CHECK(3,3,0) || \
+    JSON_HEDLEY_INTEL_VERSION_CHECK(13,0,0) || \
+    JSON_HEDLEY_ARM_VERSION_CHECK(4,1,0)
+    #define JSON_HEDLEY_NON_NULL(...) __attribute__((__nonnull__(__VA_ARGS__)))
+#else
+    #define JSON_HEDLEY_NON_NULL(...)
+#endif
+JSON_HEDLEY_DIAGNOSTIC_POP
+
+#if defined(JSON_HEDLEY_PRINTF_FORMAT)
+    #undef JSON_HEDLEY_PRINTF_FORMAT
+#endif
+#if defined(__MINGW32__) && JSON_HEDLEY_GCC_HAS_ATTRIBUTE(format,4,4,0) && !defined(__USE_MINGW_ANSI_STDIO)
+    #define JSON_HEDLEY_PRINTF_FORMAT(string_idx,first_to_check) __attribute__((__format__(ms_printf, string_idx, first_to_check)))
+#elif defined(__MINGW32__) && JSON_HEDLEY_GCC_HAS_ATTRIBUTE(format,4,4,0) && defined(__USE_MINGW_ANSI_STDIO)
+    #define JSON_HEDLEY_PRINTF_FORMAT(string_idx,first_to_check) __attribute__((__format__(gnu_printf, string_idx, first_to_check)))
+#elif \
+    JSON_HEDLEY_HAS_ATTRIBUTE(format) || \
+    JSON_HEDLEY_GCC_VERSION_CHECK(3,1,0) || \
+    JSON_HEDLEY_INTEL_VERSION_CHECK(13,0,0) || \
+    JSON_HEDLEY_ARM_VERSION_CHECK(5,6,0) || \
+    JSON_HEDLEY_IBM_VERSION_CHECK(10,1,0) || \
+    JSON_HEDLEY_TI_VERSION_CHECK(15,12,0) || \
+    (JSON_HEDLEY_TI_ARMCL_VERSION_CHECK(4,8,0) && defined(__TI_GNU_ATTRIBUTE_SUPPORT__)) || \
+    JSON_HEDLEY_TI_ARMCL_VERSION_CHECK(5,2,0) || \
+    (JSON_HEDLEY_TI_CL2000_VERSION_CHECK(6,0,0) && defined(__TI_GNU_ATTRIBUTE_SUPPORT__)) || \
+    JSON_HEDLEY_TI_CL2000_VERSION_CHECK(6,4,0) || \
+    (JSON_HEDLEY_TI_CL430_VERSION_CHECK(4,0,0) && defined(__TI_GNU_ATTRIBUTE_SUPPORT__)) || \
+    JSON_HEDLEY_TI_CL430_VERSION_CHECK(4,3,0) || \
+    (JSON_HEDLEY_TI_CL6X_VERSION_CHECK(7,2,0) && defined(__TI_GNU_ATTRIBUTE_SUPPORT__)) || \
+    JSON_HEDLEY_TI_CL6X_VERSION_CHECK(7,5,0) || \
+    JSON_HEDLEY_TI_CL7X_VERSION_CHECK(1,2,0) || \
+    JSON_HEDLEY_TI_CLPRU_VERSION_CHECK(2,1,0) || \
+    JSON_HEDLEY_MCST_LCC_VERSION_CHECK(1,25,10)
+    #define JSON_HEDLEY_PRINTF_FORMAT(string_idx,first_to_check) __attribute__((__format__(__printf__, string_idx, first_to_check)))
+#elif JSON_HEDLEY_PELLES_VERSION_CHECK(6,0,0)
+    #define JSON_HEDLEY_PRINTF_FORMAT(string_idx,first_to_check) __declspec(vaformat(printf,string_idx,first_to_check))
+#else
+    #define JSON_HEDLEY_PRINTF_FORMAT(string_idx,first_to_check)
+#endif
+
+#if defined(JSON_HEDLEY_CONSTEXPR)
+    #undef JSON_HEDLEY_CONSTEXPR
+#endif
+#if defined(__cplusplus)
+    #if __cplusplus >= 201103L
+        #define JSON_HEDLEY_CONSTEXPR JSON_HEDLEY_DIAGNOSTIC_DISABLE_CPP98_COMPAT_WRAP_(constexpr)
+    #endif
+#endif
+#if !defined(JSON_HEDLEY_CONSTEXPR)
+    #define JSON_HEDLEY_CONSTEXPR
+#endif
+
+#if defined(JSON_HEDLEY_PREDICT)
+    #undef JSON_HEDLEY_PREDICT
+#endif
+#if defined(JSON_HEDLEY_LIKELY)
+    #undef JSON_HEDLEY_LIKELY
+#endif
+#if defined(JSON_HEDLEY_UNLIKELY)
+    #undef JSON_HEDLEY_UNLIKELY
+#endif
+#if defined(JSON_HEDLEY_UNPREDICTABLE)
+    #undef JSON_HEDLEY_UNPREDICTABLE
+#endif
+#if JSON_HEDLEY_HAS_BUILTIN(__builtin_unpredictable)
+    #define JSON_HEDLEY_UNPREDICTABLE(expr) __builtin_unpredictable((expr))
+#endif
+#if \
+  (JSON_HEDLEY_HAS_BUILTIN(__builtin_expect_with_probability) && !defined(JSON_HEDLEY_PGI_VERSION)) || \
+  JSON_HEDLEY_GCC_VERSION_CHECK(9,0,0) || \
+  JSON_HEDLEY_MCST_LCC_VERSION_CHECK(1,25,10)
+#  define JSON_HEDLEY_PREDICT(expr, value, probability) __builtin_expect_with_probability(  (expr), (value), (probability))
+#  define JSON_HEDLEY_PREDICT_TRUE(expr, probability)   __builtin_expect_with_probability(!!(expr),    1   , (probability))
+#  define JSON_HEDLEY_PREDICT_FALSE(expr, probability)  __builtin_expect_with_probability(!!(expr),    0   , (probability))
+#  define JSON_HEDLEY_LIKELY(expr)                      __builtin_expect                 (!!(expr),    1                  )
+#  define JSON_HEDLEY_UNLIKELY(expr)                    __builtin_expect                 (!!(expr),    0                  )
+#elif \
+  (JSON_HEDLEY_HAS_BUILTIN(__builtin_expect) && !defined(JSON_HEDLEY_INTEL_CL_VERSION)) || \
+  JSON_HEDLEY_GCC_VERSION_CHECK(3,0,0) || \
+  JSON_HEDLEY_INTEL_VERSION_CHECK(13,0,0) || \
+  (JSON_HEDLEY_SUNPRO_VERSION_CHECK(5,15,0) && defined(__cplusplus)) || \
+  JSON_HEDLEY_ARM_VERSION_CHECK(4,1,0) || \
+  JSON_HEDLEY_IBM_VERSION_CHECK(10,1,0) || \
+  JSON_HEDLEY_TI_VERSION_CHECK(15,12,0) || \
+  JSON_HEDLEY_TI_ARMCL_VERSION_CHECK(4,7,0) || \
+  JSON_HEDLEY_TI_CL430_VERSION_CHECK(3,1,0) || \
+  JSON_HEDLEY_TI_CL2000_VERSION_CHECK(6,1,0) || \
+  JSON_HEDLEY_TI_CL6X_VERSION_CHECK(6,1,0) || \
+  JSON_HEDLEY_TI_CL7X_VERSION_CHECK(1,2,0) || \
+  JSON_HEDLEY_TI_CLPRU_VERSION_CHECK(2,1,0) || \
+  JSON_HEDLEY_TINYC_VERSION_CHECK(0,9,27) || \
+  JSON_HEDLEY_CRAY_VERSION_CHECK(8,1,0) || \
+  JSON_HEDLEY_MCST_LCC_VERSION_CHECK(1,25,10)
+#  define JSON_HEDLEY_PREDICT(expr, expected, probability) \
+    (((probability) >= 0.9) ? __builtin_expect((expr), (expected)) : (JSON_HEDLEY_STATIC_CAST(void, expected), (expr)))
+#  define JSON_HEDLEY_PREDICT_TRUE(expr, probability) \
+    (__extension__ ({ \
+        double hedley_probability_ = (probability); \
+        ((hedley_probability_ >= 0.9) ? __builtin_expect(!!(expr), 1) : ((hedley_probability_ <= 0.1) ? __builtin_expect(!!(expr), 0) : !!(expr))); \
+    }))
+#  define JSON_HEDLEY_PREDICT_FALSE(expr, probability) \
+    (__extension__ ({ \
+        double hedley_probability_ = (probability); \
+        ((hedley_probability_ >= 0.9) ? __builtin_expect(!!(expr), 0) : ((hedley_probability_ <= 0.1) ? __builtin_expect(!!(expr), 1) : !!(expr))); \
+    }))
+#  define JSON_HEDLEY_LIKELY(expr)   __builtin_expect(!!(expr), 1)
+#  define JSON_HEDLEY_UNLIKELY(expr) __builtin_expect(!!(expr), 0)
+#else
+#  define JSON_HEDLEY_PREDICT(expr, expected, probability) (JSON_HEDLEY_STATIC_CAST(void, expected), (expr))
+#  define JSON_HEDLEY_PREDICT_TRUE(expr, probability) (!!(expr))
+#  define JSON_HEDLEY_PREDICT_FALSE(expr, probability) (!!(expr))
+#  define JSON_HEDLEY_LIKELY(expr) (!!(expr))
+#  define JSON_HEDLEY_UNLIKELY(expr) (!!(expr))
+#endif
+#if !defined(JSON_HEDLEY_UNPREDICTABLE)
+    #define JSON_HEDLEY_UNPREDICTABLE(expr) JSON_HEDLEY_PREDICT(expr, 1, 0.5)
+#endif
+
+#if defined(JSON_HEDLEY_MALLOC)
+    #undef JSON_HEDLEY_MALLOC
+#endif
+#if \
+    JSON_HEDLEY_HAS_ATTRIBUTE(malloc) || \
+    JSON_HEDLEY_GCC_VERSION_CHECK(3,1,0) || \
+    JSON_HEDLEY_INTEL_VERSION_CHECK(13,0,0) || \
+    JSON_HEDLEY_SUNPRO_VERSION_CHECK(5,11,0) || \
+    JSON_HEDLEY_ARM_VERSION_CHECK(4,1,0) || \
+    JSON_HEDLEY_IBM_VERSION_CHECK(12,1,0) || \
+    JSON_HEDLEY_TI_VERSION_CHECK(15,12,0) || \
+    (JSON_HEDLEY_TI_ARMCL_VERSION_CHECK(4,8,0) && defined(__TI_GNU_ATTRIBUTE_SUPPORT__)) || \
+    JSON_HEDLEY_TI_ARMCL_VERSION_CHECK(5,2,0) || \
+    (JSON_HEDLEY_TI_CL2000_VERSION_CHECK(6,0,0) && defined(__TI_GNU_ATTRIBUTE_SUPPORT__)) || \
+    JSON_HEDLEY_TI_CL2000_VERSION_CHECK(6,4,0) || \
+    (JSON_HEDLEY_TI_CL430_VERSION_CHECK(4,0,0) && defined(__TI_GNU_ATTRIBUTE_SUPPORT__)) || \
+    JSON_HEDLEY_TI_CL430_VERSION_CHECK(4,3,0) || \
+    (JSON_HEDLEY_TI_CL6X_VERSION_CHECK(7,2,0) && defined(__TI_GNU_ATTRIBUTE_SUPPORT__)) || \
+    JSON_HEDLEY_TI_CL6X_VERSION_CHECK(7,5,0) || \
+    JSON_HEDLEY_TI_CL7X_VERSION_CHECK(1,2,0) || \
+    JSON_HEDLEY_TI_CLPRU_VERSION_CHECK(2,1,0) || \
+    JSON_HEDLEY_MCST_LCC_VERSION_CHECK(1,25,10)
+    #define JSON_HEDLEY_MALLOC __attribute__((__malloc__))
+#elif JSON_HEDLEY_SUNPRO_VERSION_CHECK(5,10,0)
+    #define JSON_HEDLEY_MALLOC _Pragma("returns_new_memory")
+#elif \
+    JSON_HEDLEY_MSVC_VERSION_CHECK(14,0,0) || \
+    JSON_HEDLEY_INTEL_CL_VERSION_CHECK(2021,1,0)
+    #define JSON_HEDLEY_MALLOC __declspec(restrict)
+#else
+    #define JSON_HEDLEY_MALLOC
+#endif
+
+#if defined(JSON_HEDLEY_PURE)
+    #undef JSON_HEDLEY_PURE
+#endif
+#if \
+  JSON_HEDLEY_HAS_ATTRIBUTE(pure) || \
+  JSON_HEDLEY_GCC_VERSION_CHECK(2,96,0) || \
+  JSON_HEDLEY_INTEL_VERSION_CHECK(13,0,0) || \
+  JSON_HEDLEY_SUNPRO_VERSION_CHECK(5,11,0) || \
+  JSON_HEDLEY_ARM_VERSION_CHECK(4,1,0) || \
+  JSON_HEDLEY_IBM_VERSION_CHECK(10,1,0) || \
+  JSON_HEDLEY_TI_VERSION_CHECK(15,12,0) || \
+  (JSON_HEDLEY_TI_ARMCL_VERSION_CHECK(4,8,0) && defined(__TI_GNU_ATTRIBUTE_SUPPORT__)) || \
+  JSON_HEDLEY_TI_ARMCL_VERSION_CHECK(5,2,0) || \
+  (JSON_HEDLEY_TI_CL2000_VERSION_CHECK(6,0,0) && defined(__TI_GNU_ATTRIBUTE_SUPPORT__)) || \
+  JSON_HEDLEY_TI_CL2000_VERSION_CHECK(6,4,0) || \
+  (JSON_HEDLEY_TI_CL430_VERSION_CHECK(4,0,0) && defined(__TI_GNU_ATTRIBUTE_SUPPORT__)) || \
+  JSON_HEDLEY_TI_CL430_VERSION_CHECK(4,3,0) || \
+  (JSON_HEDLEY_TI_CL6X_VERSION_CHECK(7,2,0) && defined(__TI_GNU_ATTRIBUTE_SUPPORT__)) || \
+  JSON_HEDLEY_TI_CL6X_VERSION_CHECK(7,5,0) || \
+  JSON_HEDLEY_TI_CL7X_VERSION_CHECK(1,2,0) || \
+  JSON_HEDLEY_TI_CLPRU_VERSION_CHECK(2,1,0) || \
+  JSON_HEDLEY_PGI_VERSION_CHECK(17,10,0) || \
+  JSON_HEDLEY_MCST_LCC_VERSION_CHECK(1,25,10)
+#  define JSON_HEDLEY_PURE __attribute__((__pure__))
+#elif JSON_HEDLEY_SUNPRO_VERSION_CHECK(5,10,0)
+#  define JSON_HEDLEY_PURE _Pragma("does_not_write_global_data")
+#elif defined(__cplusplus) && \
+    ( \
+      JSON_HEDLEY_TI_CL430_VERSION_CHECK(2,0,1) || \
+      JSON_HEDLEY_TI_CL6X_VERSION_CHECK(4,0,0) || \
+      JSON_HEDLEY_TI_CL7X_VERSION_CHECK(1,2,0) \
+    )
+#  define JSON_HEDLEY_PURE _Pragma("FUNC_IS_PURE;")
+#else
+#  define JSON_HEDLEY_PURE
+#endif
+
+#if defined(JSON_HEDLEY_CONST)
+    #undef JSON_HEDLEY_CONST
+#endif
+#if \
+    JSON_HEDLEY_HAS_ATTRIBUTE(const) || \
+    JSON_HEDLEY_GCC_VERSION_CHECK(2,5,0) || \
+    JSON_HEDLEY_INTEL_VERSION_CHECK(13,0,0) || \
+    JSON_HEDLEY_SUNPRO_VERSION_CHECK(5,11,0) || \
+    JSON_HEDLEY_ARM_VERSION_CHECK(4,1,0) || \
+    JSON_HEDLEY_IBM_VERSION_CHECK(10,1,0) || \
+    JSON_HEDLEY_TI_VERSION_CHECK(15,12,0) || \
+    (JSON_HEDLEY_TI_ARMCL_VERSION_CHECK(4,8,0) && defined(__TI_GNU_ATTRIBUTE_SUPPORT__)) || \
+    JSON_HEDLEY_TI_ARMCL_VERSION_CHECK(5,2,0) || \
+    (JSON_HEDLEY_TI_CL2000_VERSION_CHECK(6,0,0) && defined(__TI_GNU_ATTRIBUTE_SUPPORT__)) || \
+    JSON_HEDLEY_TI_CL2000_VERSION_CHECK(6,4,0) || \
+    (JSON_HEDLEY_TI_CL430_VERSION_CHECK(4,0,0) && defined(__TI_GNU_ATTRIBUTE_SUPPORT__)) || \
+    JSON_HEDLEY_TI_CL430_VERSION_CHECK(4,3,0) || \
+    (JSON_HEDLEY_TI_CL6X_VERSION_CHECK(7,2,0) && defined(__TI_GNU_ATTRIBUTE_SUPPORT__)) || \
+    JSON_HEDLEY_TI_CL6X_VERSION_CHECK(7,5,0) || \
+    JSON_HEDLEY_TI_CL7X_VERSION_CHECK(1,2,0) || \
+    JSON_HEDLEY_TI_CLPRU_VERSION_CHECK(2,1,0) || \
+    JSON_HEDLEY_PGI_VERSION_CHECK(17,10,0) || \
+    JSON_HEDLEY_MCST_LCC_VERSION_CHECK(1,25,10)
+    #define JSON_HEDLEY_CONST __attribute__((__const__))
+#elif \
+    JSON_HEDLEY_SUNPRO_VERSION_CHECK(5,10,0)
+    #define JSON_HEDLEY_CONST _Pragma("no_side_effect")
+#else
+    #define JSON_HEDLEY_CONST JSON_HEDLEY_PURE
+#endif
+
+#if defined(JSON_HEDLEY_RESTRICT)
+    #undef JSON_HEDLEY_RESTRICT
+#endif
+#if defined(__STDC_VERSION__) && (__STDC_VERSION__ >= 199901L) && !defined(__cplusplus)
+    #define JSON_HEDLEY_RESTRICT restrict
+#elif \
+    JSON_HEDLEY_GCC_VERSION_CHECK(3,1,0) || \
+    JSON_HEDLEY_MSVC_VERSION_CHECK(14,0,0) || \
+    JSON_HEDLEY_INTEL_VERSION_CHECK(13,0,0) || \
+    JSON_HEDLEY_INTEL_CL_VERSION_CHECK(2021,1,0) || \
+    JSON_HEDLEY_ARM_VERSION_CHECK(4,1,0) || \
+    JSON_HEDLEY_IBM_VERSION_CHECK(10,1,0) || \
+    JSON_HEDLEY_PGI_VERSION_CHECK(17,10,0) || \
+    JSON_HEDLEY_TI_CL430_VERSION_CHECK(4,3,0) || \
+    JSON_HEDLEY_TI_CL2000_VERSION_CHECK(6,2,4) || \
+    JSON_HEDLEY_TI_CL6X_VERSION_CHECK(8,1,0) || \
+    JSON_HEDLEY_TI_CL7X_VERSION_CHECK(1,2,0) || \
+    (JSON_HEDLEY_SUNPRO_VERSION_CHECK(5,14,0) && defined(__cplusplus)) || \
+    JSON_HEDLEY_IAR_VERSION_CHECK(8,0,0) || \
+    defined(__clang__) || \
+    JSON_HEDLEY_MCST_LCC_VERSION_CHECK(1,25,10)
+    #define JSON_HEDLEY_RESTRICT __restrict
+#elif JSON_HEDLEY_SUNPRO_VERSION_CHECK(5,3,0) && !defined(__cplusplus)
+    #define JSON_HEDLEY_RESTRICT _Restrict
+#else
+    #define JSON_HEDLEY_RESTRICT
+#endif
+
+#if defined(JSON_HEDLEY_INLINE)
+    #undef JSON_HEDLEY_INLINE
+#endif
+#if \
+    (defined(__STDC_VERSION__) && (__STDC_VERSION__ >= 199901L)) || \
+    (defined(__cplusplus) && (__cplusplus >= 199711L))
+    #define JSON_HEDLEY_INLINE inline
+#elif \
+    defined(JSON_HEDLEY_GCC_VERSION) || \
+    JSON_HEDLEY_ARM_VERSION_CHECK(6,2,0)
+    #define JSON_HEDLEY_INLINE __inline__
+#elif \
+    JSON_HEDLEY_MSVC_VERSION_CHECK(12,0,0) || \
+    JSON_HEDLEY_INTEL_CL_VERSION_CHECK(2021,1,0) || \
+    JSON_HEDLEY_ARM_VERSION_CHECK(4,1,0) || \
+    JSON_HEDLEY_TI_ARMCL_VERSION_CHECK(5,1,0) || \
+    JSON_HEDLEY_TI_CL430_VERSION_CHECK(3,1,0) || \
+    JSON_HEDLEY_TI_CL2000_VERSION_CHECK(6,2,0) || \
+    JSON_HEDLEY_TI_CL6X_VERSION_CHECK(8,0,0) || \
+    JSON_HEDLEY_TI_CL7X_VERSION_CHECK(1,2,0) || \
+    JSON_HEDLEY_TI_CLPRU_VERSION_CHECK(2,1,0) || \
+    JSON_HEDLEY_MCST_LCC_VERSION_CHECK(1,25,10)
+    #define JSON_HEDLEY_INLINE __inline
+#else
+    #define JSON_HEDLEY_INLINE
+#endif
+
+#if defined(JSON_HEDLEY_ALWAYS_INLINE)
+    #undef JSON_HEDLEY_ALWAYS_INLINE
+#endif
+#if \
+  JSON_HEDLEY_HAS_ATTRIBUTE(always_inline) || \
+  JSON_HEDLEY_GCC_VERSION_CHECK(4,0,0) || \
+  JSON_HEDLEY_INTEL_VERSION_CHECK(13,0,0) || \
+  JSON_HEDLEY_SUNPRO_VERSION_CHECK(5,11,0) || \
+  JSON_HEDLEY_ARM_VERSION_CHECK(4,1,0) || \
+  JSON_HEDLEY_IBM_VERSION_CHECK(10,1,0) || \
+  JSON_HEDLEY_TI_VERSION_CHECK(15,12,0) || \
+  (JSON_HEDLEY_TI_ARMCL_VERSION_CHECK(4,8,0) && defined(__TI_GNU_ATTRIBUTE_SUPPORT__)) || \
+  JSON_HEDLEY_TI_ARMCL_VERSION_CHECK(5,2,0) || \
+  (JSON_HEDLEY_TI_CL2000_VERSION_CHECK(6,0,0) && defined(__TI_GNU_ATTRIBUTE_SUPPORT__)) || \
+  JSON_HEDLEY_TI_CL2000_VERSION_CHECK(6,4,0) || \
+  (JSON_HEDLEY_TI_CL430_VERSION_CHECK(4,0,0) && defined(__TI_GNU_ATTRIBUTE_SUPPORT__)) || \
+  JSON_HEDLEY_TI_CL430_VERSION_CHECK(4,3,0) || \
+  (JSON_HEDLEY_TI_CL6X_VERSION_CHECK(7,2,0) && defined(__TI_GNU_ATTRIBUTE_SUPPORT__)) || \
+  JSON_HEDLEY_TI_CL6X_VERSION_CHECK(7,5,0) || \
+  JSON_HEDLEY_TI_CL7X_VERSION_CHECK(1,2,0) || \
+  JSON_HEDLEY_TI_CLPRU_VERSION_CHECK(2,1,0) || \
+  JSON_HEDLEY_MCST_LCC_VERSION_CHECK(1,25,10) || \
+  JSON_HEDLEY_IAR_VERSION_CHECK(8,10,0)
+#  define JSON_HEDLEY_ALWAYS_INLINE __attribute__((__always_inline__)) JSON_HEDLEY_INLINE
+#elif \
+  JSON_HEDLEY_MSVC_VERSION_CHECK(12,0,0) || \
+  JSON_HEDLEY_INTEL_CL_VERSION_CHECK(2021,1,0)
+#  define JSON_HEDLEY_ALWAYS_INLINE __forceinline
+#elif defined(__cplusplus) && \
+    ( \
+      JSON_HEDLEY_TI_ARMCL_VERSION_CHECK(5,2,0) || \
+      JSON_HEDLEY_TI_CL430_VERSION_CHECK(4,3,0) || \
+      JSON_HEDLEY_TI_CL2000_VERSION_CHECK(6,4,0) || \
+      JSON_HEDLEY_TI_CL6X_VERSION_CHECK(6,1,0) || \
+      JSON_HEDLEY_TI_CL7X_VERSION_CHECK(1,2,0) || \
+      JSON_HEDLEY_TI_CLPRU_VERSION_CHECK(2,1,0) \
+    )
+#  define JSON_HEDLEY_ALWAYS_INLINE _Pragma("FUNC_ALWAYS_INLINE;")
+#elif JSON_HEDLEY_IAR_VERSION_CHECK(8,0,0)
+#  define JSON_HEDLEY_ALWAYS_INLINE _Pragma("inline=forced")
+#else
+#  define JSON_HEDLEY_ALWAYS_INLINE JSON_HEDLEY_INLINE
+#endif
+
+#if defined(JSON_HEDLEY_NEVER_INLINE)
+    #undef JSON_HEDLEY_NEVER_INLINE
+#endif
+#if \
+    JSON_HEDLEY_HAS_ATTRIBUTE(noinline) || \
+    JSON_HEDLEY_GCC_VERSION_CHECK(4,0,0) || \
+    JSON_HEDLEY_INTEL_VERSION_CHECK(13,0,0) || \
+    JSON_HEDLEY_SUNPRO_VERSION_CHECK(5,11,0) || \
+    JSON_HEDLEY_ARM_VERSION_CHECK(4,1,0) || \
+    JSON_HEDLEY_IBM_VERSION_CHECK(10,1,0) || \
+    JSON_HEDLEY_TI_VERSION_CHECK(15,12,0) || \
+    (JSON_HEDLEY_TI_ARMCL_VERSION_CHECK(4,8,0) && defined(__TI_GNU_ATTRIBUTE_SUPPORT__)) || \
+    JSON_HEDLEY_TI_ARMCL_VERSION_CHECK(5,2,0) || \
+    (JSON_HEDLEY_TI_CL2000_VERSION_CHECK(6,0,0) && defined(__TI_GNU_ATTRIBUTE_SUPPORT__)) || \
+    JSON_HEDLEY_TI_CL2000_VERSION_CHECK(6,4,0) || \
+    (JSON_HEDLEY_TI_CL430_VERSION_CHECK(4,0,0) && defined(__TI_GNU_ATTRIBUTE_SUPPORT__)) || \
+    JSON_HEDLEY_TI_CL430_VERSION_CHECK(4,3,0) || \
+    (JSON_HEDLEY_TI_CL6X_VERSION_CHECK(7,2,0) && defined(__TI_GNU_ATTRIBUTE_SUPPORT__)) || \
+    JSON_HEDLEY_TI_CL6X_VERSION_CHECK(7,5,0) || \
+    JSON_HEDLEY_TI_CL7X_VERSION_CHECK(1,2,0) || \
+    JSON_HEDLEY_TI_CLPRU_VERSION_CHECK(2,1,0) || \
+    JSON_HEDLEY_MCST_LCC_VERSION_CHECK(1,25,10) || \
+    JSON_HEDLEY_IAR_VERSION_CHECK(8,10,0)
+    #define JSON_HEDLEY_NEVER_INLINE __attribute__((__noinline__))
+#elif \
+    JSON_HEDLEY_MSVC_VERSION_CHECK(13,10,0) || \
+    JSON_HEDLEY_INTEL_CL_VERSION_CHECK(2021,1,0)
+    #define JSON_HEDLEY_NEVER_INLINE __declspec(noinline)
+#elif JSON_HEDLEY_PGI_VERSION_CHECK(10,2,0)
+    #define JSON_HEDLEY_NEVER_INLINE _Pragma("noinline")
+#elif JSON_HEDLEY_TI_CL6X_VERSION_CHECK(6,0,0) && defined(__cplusplus)
+    #define JSON_HEDLEY_NEVER_INLINE _Pragma("FUNC_CANNOT_INLINE;")
+#elif JSON_HEDLEY_IAR_VERSION_CHECK(8,0,0)
+    #define JSON_HEDLEY_NEVER_INLINE _Pragma("inline=never")
+#elif JSON_HEDLEY_COMPCERT_VERSION_CHECK(3,2,0)
+    #define JSON_HEDLEY_NEVER_INLINE __attribute((noinline))
+#elif JSON_HEDLEY_PELLES_VERSION_CHECK(9,0,0)
+    #define JSON_HEDLEY_NEVER_INLINE __declspec(noinline)
+#else
+    #define JSON_HEDLEY_NEVER_INLINE
+#endif
+
+#if defined(JSON_HEDLEY_PRIVATE)
+    #undef JSON_HEDLEY_PRIVATE
+#endif
+#if defined(JSON_HEDLEY_PUBLIC)
+    #undef JSON_HEDLEY_PUBLIC
+#endif
+#if defined(JSON_HEDLEY_IMPORT)
+    #undef JSON_HEDLEY_IMPORT
+#endif
+#if defined(_WIN32) || defined(__CYGWIN__)
+#  define JSON_HEDLEY_PRIVATE
+#  define JSON_HEDLEY_PUBLIC   __declspec(dllexport)
+#  define JSON_HEDLEY_IMPORT   __declspec(dllimport)
+#else
+#  if \
+    JSON_HEDLEY_HAS_ATTRIBUTE(visibility) || \
+    JSON_HEDLEY_GCC_VERSION_CHECK(3,3,0) || \
+    JSON_HEDLEY_SUNPRO_VERSION_CHECK(5,11,0) || \
+    JSON_HEDLEY_INTEL_VERSION_CHECK(13,0,0) || \
+    JSON_HEDLEY_ARM_VERSION_CHECK(4,1,0) || \
+    JSON_HEDLEY_IBM_VERSION_CHECK(13,1,0) || \
+    ( \
+      defined(__TI_EABI__) && \
+      ( \
+        (JSON_HEDLEY_TI_CL6X_VERSION_CHECK(7,2,0) && defined(__TI_GNU_ATTRIBUTE_SUPPORT__)) || \
+        JSON_HEDLEY_TI_CL6X_VERSION_CHECK(7,5,0) \
+      ) \
+    ) || \
+    JSON_HEDLEY_MCST_LCC_VERSION_CHECK(1,25,10)
+#    define JSON_HEDLEY_PRIVATE __attribute__((__visibility__("hidden")))
+#    define JSON_HEDLEY_PUBLIC  __attribute__((__visibility__("default")))
+#  else
+#    define JSON_HEDLEY_PRIVATE
+#    define JSON_HEDLEY_PUBLIC
+#  endif
+#  define JSON_HEDLEY_IMPORT    extern
+#endif
+
+#if defined(JSON_HEDLEY_NO_THROW)
+    #undef JSON_HEDLEY_NO_THROW
+#endif
+#if \
+    JSON_HEDLEY_HAS_ATTRIBUTE(nothrow) || \
+    JSON_HEDLEY_GCC_VERSION_CHECK(3,3,0) || \
+    JSON_HEDLEY_INTEL_VERSION_CHECK(13,0,0) || \
+    JSON_HEDLEY_MCST_LCC_VERSION_CHECK(1,25,10)
+    #define JSON_HEDLEY_NO_THROW __attribute__((__nothrow__))
+#elif \
+    JSON_HEDLEY_MSVC_VERSION_CHECK(13,1,0) || \
+    JSON_HEDLEY_INTEL_CL_VERSION_CHECK(2021,1,0) || \
+    JSON_HEDLEY_ARM_VERSION_CHECK(4,1,0)
+    #define JSON_HEDLEY_NO_THROW __declspec(nothrow)
+#else
+    #define JSON_HEDLEY_NO_THROW
+#endif
+
+#if defined(JSON_HEDLEY_FALL_THROUGH)
+    #undef JSON_HEDLEY_FALL_THROUGH
+#endif
+#if \
+    JSON_HEDLEY_HAS_ATTRIBUTE(fallthrough) || \
+    JSON_HEDLEY_GCC_VERSION_CHECK(7,0,0) || \
+    JSON_HEDLEY_MCST_LCC_VERSION_CHECK(1,25,10)
+    #define JSON_HEDLEY_FALL_THROUGH __attribute__((__fallthrough__))
+#elif JSON_HEDLEY_HAS_CPP_ATTRIBUTE_NS(clang,fallthrough)
+    #define JSON_HEDLEY_FALL_THROUGH JSON_HEDLEY_DIAGNOSTIC_DISABLE_CPP98_COMPAT_WRAP_([[clang::fallthrough]])
+#elif JSON_HEDLEY_HAS_CPP_ATTRIBUTE(fallthrough)
+    #define JSON_HEDLEY_FALL_THROUGH JSON_HEDLEY_DIAGNOSTIC_DISABLE_CPP98_COMPAT_WRAP_([[fallthrough]])
+#elif defined(__fallthrough) /* SAL */
+    #define JSON_HEDLEY_FALL_THROUGH __fallthrough
+#else
+    #define JSON_HEDLEY_FALL_THROUGH
+#endif
+
+#if defined(JSON_HEDLEY_RETURNS_NON_NULL)
+    #undef JSON_HEDLEY_RETURNS_NON_NULL
+#endif
+#if \
+    JSON_HEDLEY_HAS_ATTRIBUTE(returns_nonnull) || \
+    JSON_HEDLEY_GCC_VERSION_CHECK(4,9,0) || \
+    JSON_HEDLEY_MCST_LCC_VERSION_CHECK(1,25,10)
+    #define JSON_HEDLEY_RETURNS_NON_NULL __attribute__((__returns_nonnull__))
+#elif defined(_Ret_notnull_) /* SAL */
+    #define JSON_HEDLEY_RETURNS_NON_NULL _Ret_notnull_
+#else
+    #define JSON_HEDLEY_RETURNS_NON_NULL
+#endif
+
+#if defined(JSON_HEDLEY_ARRAY_PARAM)
+    #undef JSON_HEDLEY_ARRAY_PARAM
+#endif
+#if \
+    defined(__STDC_VERSION__) && (__STDC_VERSION__ >= 199901L) && \
+    !defined(__STDC_NO_VLA__) && \
+    !defined(__cplusplus) && \
+    !defined(JSON_HEDLEY_PGI_VERSION) && \
+    !defined(JSON_HEDLEY_TINYC_VERSION)
+    #define JSON_HEDLEY_ARRAY_PARAM(name) (name)
+#else
+    #define JSON_HEDLEY_ARRAY_PARAM(name)
+#endif
+
+#if defined(JSON_HEDLEY_IS_CONSTANT)
+    #undef JSON_HEDLEY_IS_CONSTANT
+#endif
+#if defined(JSON_HEDLEY_REQUIRE_CONSTEXPR)
+    #undef JSON_HEDLEY_REQUIRE_CONSTEXPR
+#endif
+/* JSON_HEDLEY_IS_CONSTEXPR_ is for
+   HEDLEY INTERNAL USE ONLY.  API subject to change without notice. */
+#if defined(JSON_HEDLEY_IS_CONSTEXPR_)
+    #undef JSON_HEDLEY_IS_CONSTEXPR_
+#endif
+#if \
+    JSON_HEDLEY_HAS_BUILTIN(__builtin_constant_p) || \
+    JSON_HEDLEY_GCC_VERSION_CHECK(3,4,0) || \
+    JSON_HEDLEY_INTEL_VERSION_CHECK(13,0,0) || \
+    JSON_HEDLEY_TINYC_VERSION_CHECK(0,9,19) || \
+    JSON_HEDLEY_ARM_VERSION_CHECK(4,1,0) || \
+    JSON_HEDLEY_IBM_VERSION_CHECK(13,1,0) || \
+    JSON_HEDLEY_TI_CL6X_VERSION_CHECK(6,1,0) || \
+    (JSON_HEDLEY_SUNPRO_VERSION_CHECK(5,10,0) && !defined(__cplusplus)) || \
+    JSON_HEDLEY_CRAY_VERSION_CHECK(8,1,0) || \
+    JSON_HEDLEY_MCST_LCC_VERSION_CHECK(1,25,10)
+    #define JSON_HEDLEY_IS_CONSTANT(expr) __builtin_constant_p(expr)
+#endif
+#if !defined(__cplusplus)
+#  if \
+       JSON_HEDLEY_HAS_BUILTIN(__builtin_types_compatible_p) || \
+       JSON_HEDLEY_GCC_VERSION_CHECK(3,4,0) || \
+       JSON_HEDLEY_INTEL_VERSION_CHECK(13,0,0) || \
+       JSON_HEDLEY_IBM_VERSION_CHECK(13,1,0) || \
+       JSON_HEDLEY_CRAY_VERSION_CHECK(8,1,0) || \
+       JSON_HEDLEY_ARM_VERSION_CHECK(5,4,0) || \
+       JSON_HEDLEY_TINYC_VERSION_CHECK(0,9,24)
+#if defined(__INTPTR_TYPE__)
+    #define JSON_HEDLEY_IS_CONSTEXPR_(expr) __builtin_types_compatible_p(__typeof__((1 ? (void*) ((__INTPTR_TYPE__) ((expr) * 0)) : (int*) 0)), int*)
+#else
+    #include <stdint.h>
+    #define JSON_HEDLEY_IS_CONSTEXPR_(expr) __builtin_types_compatible_p(__typeof__((1 ? (void*) ((intptr_t) ((expr) * 0)) : (int*) 0)), int*)
+#endif
+#  elif \
+       ( \
+          defined(__STDC_VERSION__) && (__STDC_VERSION__ >= 201112L) && \
+          !defined(JSON_HEDLEY_SUNPRO_VERSION) && \
+          !defined(JSON_HEDLEY_PGI_VERSION) && \
+          !defined(JSON_HEDLEY_IAR_VERSION)) || \
+       (JSON_HEDLEY_HAS_EXTENSION(c_generic_selections) && !defined(JSON_HEDLEY_IAR_VERSION)) || \
+       JSON_HEDLEY_GCC_VERSION_CHECK(4,9,0) || \
+       JSON_HEDLEY_INTEL_VERSION_CHECK(17,0,0) || \
+       JSON_HEDLEY_IBM_VERSION_CHECK(12,1,0) || \
+       JSON_HEDLEY_ARM_VERSION_CHECK(5,3,0)
+#if defined(__INTPTR_TYPE__)
+    #define JSON_HEDLEY_IS_CONSTEXPR_(expr) _Generic((1 ? (void*) ((__INTPTR_TYPE__) ((expr) * 0)) : (int*) 0), int*: 1, void*: 0)
+#else
+    #include <stdint.h>
+    #define JSON_HEDLEY_IS_CONSTEXPR_(expr) _Generic((1 ? (void*) ((intptr_t) * 0) : (int*) 0), int*: 1, void*: 0)
+#endif
+#  elif \
+       defined(JSON_HEDLEY_GCC_VERSION) || \
+       defined(JSON_HEDLEY_INTEL_VERSION) || \
+       defined(JSON_HEDLEY_TINYC_VERSION) || \
+       defined(JSON_HEDLEY_TI_ARMCL_VERSION) || \
+       JSON_HEDLEY_TI_CL430_VERSION_CHECK(18,12,0) || \
+       defined(JSON_HEDLEY_TI_CL2000_VERSION) || \
+       defined(JSON_HEDLEY_TI_CL6X_VERSION) || \
+       defined(JSON_HEDLEY_TI_CL7X_VERSION) || \
+       defined(JSON_HEDLEY_TI_CLPRU_VERSION) || \
+       defined(__clang__)
+#    define JSON_HEDLEY_IS_CONSTEXPR_(expr) ( \
+        sizeof(void) != \
+        sizeof(*( \
+                  1 ? \
+                  ((void*) ((expr) * 0L) ) : \
+((struct { char v[sizeof(void) * 2]; } *) 1) \
+                ) \
+              ) \
+                                            )
+#  endif
+#endif
+#if defined(JSON_HEDLEY_IS_CONSTEXPR_)
+    #if !defined(JSON_HEDLEY_IS_CONSTANT)
+        #define JSON_HEDLEY_IS_CONSTANT(expr) JSON_HEDLEY_IS_CONSTEXPR_(expr)
+    #endif
+    #define JSON_HEDLEY_REQUIRE_CONSTEXPR(expr) (JSON_HEDLEY_IS_CONSTEXPR_(expr) ? (expr) : (-1))
+#else
+    #if !defined(JSON_HEDLEY_IS_CONSTANT)
+        #define JSON_HEDLEY_IS_CONSTANT(expr) (0)
+    #endif
+    #define JSON_HEDLEY_REQUIRE_CONSTEXPR(expr) (expr)
+#endif
+
+#if defined(JSON_HEDLEY_BEGIN_C_DECLS)
+    #undef JSON_HEDLEY_BEGIN_C_DECLS
+#endif
+#if defined(JSON_HEDLEY_END_C_DECLS)
+    #undef JSON_HEDLEY_END_C_DECLS
+#endif
+#if defined(JSON_HEDLEY_C_DECL)
+    #undef JSON_HEDLEY_C_DECL
+#endif
+#if defined(__cplusplus)
+    #define JSON_HEDLEY_BEGIN_C_DECLS extern "C" {
+    #define JSON_HEDLEY_END_C_DECLS }
+    #define JSON_HEDLEY_C_DECL extern "C"
+#else
+    #define JSON_HEDLEY_BEGIN_C_DECLS
+    #define JSON_HEDLEY_END_C_DECLS
+    #define JSON_HEDLEY_C_DECL
+#endif
+
+#if defined(JSON_HEDLEY_STATIC_ASSERT)
+    #undef JSON_HEDLEY_STATIC_ASSERT
+#endif
+#if \
+  !defined(__cplusplus) && ( \
+      (defined(__STDC_VERSION__) && (__STDC_VERSION__ >= 201112L)) || \
+      (JSON_HEDLEY_HAS_FEATURE(c_static_assert) && !defined(JSON_HEDLEY_INTEL_CL_VERSION)) || \
+      JSON_HEDLEY_GCC_VERSION_CHECK(6,0,0) || \
+      JSON_HEDLEY_INTEL_VERSION_CHECK(13,0,0) || \
+      defined(_Static_assert) \
+    )
+#  define JSON_HEDLEY_STATIC_ASSERT(expr, message) _Static_assert(expr, message)
+#elif \
+  (defined(__cplusplus) && (__cplusplus >= 201103L)) || \
+  JSON_HEDLEY_MSVC_VERSION_CHECK(16,0,0) || \
+  JSON_HEDLEY_INTEL_CL_VERSION_CHECK(2021,1,0)
+#  define JSON_HEDLEY_STATIC_ASSERT(expr, message) JSON_HEDLEY_DIAGNOSTIC_DISABLE_CPP98_COMPAT_WRAP_(static_assert(expr, message))
+#else
+#  define JSON_HEDLEY_STATIC_ASSERT(expr, message)
+#endif
+
+#if defined(JSON_HEDLEY_NULL)
+    #undef JSON_HEDLEY_NULL
+#endif
+#if defined(__cplusplus)
+    #if __cplusplus >= 201103L
+        #define JSON_HEDLEY_NULL JSON_HEDLEY_DIAGNOSTIC_DISABLE_CPP98_COMPAT_WRAP_(nullptr)
+    #elif defined(NULL)
+        #define JSON_HEDLEY_NULL NULL
+    #else
+        #define JSON_HEDLEY_NULL JSON_HEDLEY_STATIC_CAST(void*, 0)
+    #endif
+#elif defined(NULL)
+    #define JSON_HEDLEY_NULL NULL
+#else
+    #define JSON_HEDLEY_NULL ((void*) 0)
+#endif
+
+#if defined(JSON_HEDLEY_MESSAGE)
+    #undef JSON_HEDLEY_MESSAGE
+#endif
+#if JSON_HEDLEY_HAS_WARNING("-Wunknown-pragmas")
+#  define JSON_HEDLEY_MESSAGE(msg) \
+    JSON_HEDLEY_DIAGNOSTIC_PUSH \
+    JSON_HEDLEY_DIAGNOSTIC_DISABLE_UNKNOWN_PRAGMAS \
+    JSON_HEDLEY_PRAGMA(message msg) \
+    JSON_HEDLEY_DIAGNOSTIC_POP
+#elif \
+  JSON_HEDLEY_GCC_VERSION_CHECK(4,4,0) || \
+  JSON_HEDLEY_INTEL_VERSION_CHECK(13,0,0)
+#  define JSON_HEDLEY_MESSAGE(msg) JSON_HEDLEY_PRAGMA(message msg)
+#elif JSON_HEDLEY_CRAY_VERSION_CHECK(5,0,0)
+#  define JSON_HEDLEY_MESSAGE(msg) JSON_HEDLEY_PRAGMA(_CRI message msg)
+#elif JSON_HEDLEY_IAR_VERSION_CHECK(8,0,0)
+#  define JSON_HEDLEY_MESSAGE(msg) JSON_HEDLEY_PRAGMA(message(msg))
+#elif JSON_HEDLEY_PELLES_VERSION_CHECK(2,0,0)
+#  define JSON_HEDLEY_MESSAGE(msg) JSON_HEDLEY_PRAGMA(message(msg))
+#else
+#  define JSON_HEDLEY_MESSAGE(msg)
+#endif
+
+#if defined(JSON_HEDLEY_WARNING)
+    #undef JSON_HEDLEY_WARNING
+#endif
+#if JSON_HEDLEY_HAS_WARNING("-Wunknown-pragmas")
+#  define JSON_HEDLEY_WARNING(msg) \
+    JSON_HEDLEY_DIAGNOSTIC_PUSH \
+    JSON_HEDLEY_DIAGNOSTIC_DISABLE_UNKNOWN_PRAGMAS \
+    JSON_HEDLEY_PRAGMA(clang warning msg) \
+    JSON_HEDLEY_DIAGNOSTIC_POP
+#elif \
+  JSON_HEDLEY_GCC_VERSION_CHECK(4,8,0) || \
+  JSON_HEDLEY_PGI_VERSION_CHECK(18,4,0) || \
+  JSON_HEDLEY_INTEL_VERSION_CHECK(13,0,0)
+#  define JSON_HEDLEY_WARNING(msg) JSON_HEDLEY_PRAGMA(GCC warning msg)
+#elif \
+  JSON_HEDLEY_MSVC_VERSION_CHECK(15,0,0) || \
+  JSON_HEDLEY_INTEL_CL_VERSION_CHECK(2021,1,0)
+#  define JSON_HEDLEY_WARNING(msg) JSON_HEDLEY_PRAGMA(message(msg))
+#else
+#  define JSON_HEDLEY_WARNING(msg) JSON_HEDLEY_MESSAGE(msg)
+#endif
+
+#if defined(JSON_HEDLEY_REQUIRE)
+    #undef JSON_HEDLEY_REQUIRE
+#endif
+#if defined(JSON_HEDLEY_REQUIRE_MSG)
+    #undef JSON_HEDLEY_REQUIRE_MSG
+#endif
+#if JSON_HEDLEY_HAS_ATTRIBUTE(diagnose_if)
+#  if JSON_HEDLEY_HAS_WARNING("-Wgcc-compat")
+#    define JSON_HEDLEY_REQUIRE(expr) \
+    JSON_HEDLEY_DIAGNOSTIC_PUSH \
+    _Pragma("clang diagnostic ignored \"-Wgcc-compat\"") \
+    __attribute__((diagnose_if(!(expr), #expr, "error"))) \
+    JSON_HEDLEY_DIAGNOSTIC_POP
+#    define JSON_HEDLEY_REQUIRE_MSG(expr,msg) \
+    JSON_HEDLEY_DIAGNOSTIC_PUSH \
+    _Pragma("clang diagnostic ignored \"-Wgcc-compat\"") \
+    __attribute__((diagnose_if(!(expr), msg, "error"))) \
+    JSON_HEDLEY_DIAGNOSTIC_POP
+#  else
+#    define JSON_HEDLEY_REQUIRE(expr) __attribute__((diagnose_if(!(expr), #expr, "error")))
+#    define JSON_HEDLEY_REQUIRE_MSG(expr,msg) __attribute__((diagnose_if(!(expr), msg, "error")))
+#  endif
+#else
+#  define JSON_HEDLEY_REQUIRE(expr)
+#  define JSON_HEDLEY_REQUIRE_MSG(expr,msg)
+#endif
+
+#if defined(JSON_HEDLEY_FLAGS)
+    #undef JSON_HEDLEY_FLAGS
+#endif
+#if JSON_HEDLEY_HAS_ATTRIBUTE(flag_enum) && (!defined(__cplusplus) || JSON_HEDLEY_HAS_WARNING("-Wbitfield-enum-conversion"))
+    #define JSON_HEDLEY_FLAGS __attribute__((__flag_enum__))
+#else
+    #define JSON_HEDLEY_FLAGS
+#endif
+
+#if defined(JSON_HEDLEY_FLAGS_CAST)
+    #undef JSON_HEDLEY_FLAGS_CAST
+#endif
+#if JSON_HEDLEY_INTEL_VERSION_CHECK(19,0,0)
+#  define JSON_HEDLEY_FLAGS_CAST(T, expr) (__extension__ ({ \
+        JSON_HEDLEY_DIAGNOSTIC_PUSH \
+        _Pragma("warning(disable:188)") \
+        ((T) (expr)); \
+        JSON_HEDLEY_DIAGNOSTIC_POP \
+    }))
+#else
+#  define JSON_HEDLEY_FLAGS_CAST(T, expr) JSON_HEDLEY_STATIC_CAST(T, expr)
+#endif
+
+#if defined(JSON_HEDLEY_EMPTY_BASES)
+    #undef JSON_HEDLEY_EMPTY_BASES
+#endif
+#if \
+    (JSON_HEDLEY_MSVC_VERSION_CHECK(19,0,23918) && !JSON_HEDLEY_MSVC_VERSION_CHECK(20,0,0)) || \
+    JSON_HEDLEY_INTEL_CL_VERSION_CHECK(2021,1,0)
+    #define JSON_HEDLEY_EMPTY_BASES __declspec(empty_bases)
+#else
+    #define JSON_HEDLEY_EMPTY_BASES
+#endif
+
+/* Remaining macros are deprecated. */
+
+#if defined(JSON_HEDLEY_GCC_NOT_CLANG_VERSION_CHECK)
+    #undef JSON_HEDLEY_GCC_NOT_CLANG_VERSION_CHECK
+#endif
+#if defined(__clang__)
+    #define JSON_HEDLEY_GCC_NOT_CLANG_VERSION_CHECK(major,minor,patch) (0)
+#else
+    #define JSON_HEDLEY_GCC_NOT_CLANG_VERSION_CHECK(major,minor,patch) JSON_HEDLEY_GCC_VERSION_CHECK(major,minor,patch)
+#endif
+
+#if defined(JSON_HEDLEY_CLANG_HAS_ATTRIBUTE)
+    #undef JSON_HEDLEY_CLANG_HAS_ATTRIBUTE
+#endif
+#define JSON_HEDLEY_CLANG_HAS_ATTRIBUTE(attribute) JSON_HEDLEY_HAS_ATTRIBUTE(attribute)
+
+#if defined(JSON_HEDLEY_CLANG_HAS_CPP_ATTRIBUTE)
+    #undef JSON_HEDLEY_CLANG_HAS_CPP_ATTRIBUTE
+#endif
+#define JSON_HEDLEY_CLANG_HAS_CPP_ATTRIBUTE(attribute) JSON_HEDLEY_HAS_CPP_ATTRIBUTE(attribute)
+
+#if defined(JSON_HEDLEY_CLANG_HAS_BUILTIN)
+    #undef JSON_HEDLEY_CLANG_HAS_BUILTIN
+#endif
+#define JSON_HEDLEY_CLANG_HAS_BUILTIN(builtin) JSON_HEDLEY_HAS_BUILTIN(builtin)
+
+#if defined(JSON_HEDLEY_CLANG_HAS_FEATURE)
+    #undef JSON_HEDLEY_CLANG_HAS_FEATURE
+#endif
+#define JSON_HEDLEY_CLANG_HAS_FEATURE(feature) JSON_HEDLEY_HAS_FEATURE(feature)
+
+#if defined(JSON_HEDLEY_CLANG_HAS_EXTENSION)
+    #undef JSON_HEDLEY_CLANG_HAS_EXTENSION
+#endif
+#define JSON_HEDLEY_CLANG_HAS_EXTENSION(extension) JSON_HEDLEY_HAS_EXTENSION(extension)
+
+#if defined(JSON_HEDLEY_CLANG_HAS_DECLSPEC_DECLSPEC_ATTRIBUTE)
+    #undef JSON_HEDLEY_CLANG_HAS_DECLSPEC_DECLSPEC_ATTRIBUTE
+#endif
+#define JSON_HEDLEY_CLANG_HAS_DECLSPEC_ATTRIBUTE(attribute) JSON_HEDLEY_HAS_DECLSPEC_ATTRIBUTE(attribute)
+
+#if defined(JSON_HEDLEY_CLANG_HAS_WARNING)
+    #undef JSON_HEDLEY_CLANG_HAS_WARNING
+#endif
+#define JSON_HEDLEY_CLANG_HAS_WARNING(warning) JSON_HEDLEY_HAS_WARNING(warning)
+
+#endif /* !defined(JSON_HEDLEY_VERSION) || (JSON_HEDLEY_VERSION < X) */
+
+
+// This file contains all internal macro definitions (except those affecting ABI)
+// You MUST include macro_unscope.hpp at the end of json.hpp to undef all of them
+
+// #include <nlohmann/detail/abi_macros.hpp>
+
+
+// exclude unsupported compilers
+#if !defined(JSON_SKIP_UNSUPPORTED_COMPILER_CHECK)
+    #if defined(__clang__)
+        #if (__clang_major__ * 10000 + __clang_minor__ * 100 + __clang_patchlevel__) < 30400
+            #error "unsupported Clang version - see https://github.com/nlohmann/json#supported-compilers"
+        #endif
+    #elif defined(__GNUC__) && !(defined(__ICC) || defined(__INTEL_COMPILER))
+        #if (__GNUC__ * 10000 + __GNUC_MINOR__ * 100 + __GNUC_PATCHLEVEL__) < 40800
+            #error "unsupported GCC version - see https://github.com/nlohmann/json#supported-compilers"
+        #endif
+    #endif
+#endif
+
+// C++ language standard detection
+// if the user manually specified the used c++ version this is skipped
+#if !defined(JSON_HAS_CPP_20) && !defined(JSON_HAS_CPP_17) && !defined(JSON_HAS_CPP_14) && !defined(JSON_HAS_CPP_11)
+    #if (defined(__cplusplus) && __cplusplus >= 202002L) || (defined(_MSVC_LANG) && _MSVC_LANG >= 202002L)
+        #define JSON_HAS_CPP_20
+        #define JSON_HAS_CPP_17
+        #define JSON_HAS_CPP_14
+    #elif (defined(__cplusplus) && __cplusplus >= 201703L) || (defined(_HAS_CXX17) && _HAS_CXX17 == 1) // fix for issue #464
+        #define JSON_HAS_CPP_17
+        #define JSON_HAS_CPP_14
+    #elif (defined(__cplusplus) && __cplusplus >= 201402L) || (defined(_HAS_CXX14) && _HAS_CXX14 == 1)
+        #define JSON_HAS_CPP_14
+    #endif
+    // the cpp 11 flag is always specified because it is the minimal required version
+    #define JSON_HAS_CPP_11
+#endif
+
+#ifdef __has_include
+    #if __has_include(<version>)
+        #include <version>
+    #endif
+#endif
+
+#if !defined(JSON_HAS_FILESYSTEM) && !defined(JSON_HAS_EXPERIMENTAL_FILESYSTEM)
+    #ifdef JSON_HAS_CPP_17
+        #if defined(__cpp_lib_filesystem)
+            #define JSON_HAS_FILESYSTEM 1
+        #elif defined(__cpp_lib_experimental_filesystem)
+            #define JSON_HAS_EXPERIMENTAL_FILESYSTEM 1
+        #elif !defined(__has_include)
+            #define JSON_HAS_EXPERIMENTAL_FILESYSTEM 1
+        #elif __has_include(<filesystem>)
+            #define JSON_HAS_FILESYSTEM 1
+        #elif __has_include(<experimental/filesystem>)
+            #define JSON_HAS_EXPERIMENTAL_FILESYSTEM 1
+        #endif
+
+        // std::filesystem does not work on MinGW GCC 8: https://sourceforge.net/p/mingw-w64/bugs/737/
+        #if defined(__MINGW32__) && defined(__GNUC__) && __GNUC__ == 8
+            #undef JSON_HAS_FILESYSTEM
+            #undef JSON_HAS_EXPERIMENTAL_FILESYSTEM
+        #endif
+
+        // no filesystem support before GCC 8: https://en.cppreference.com/w/cpp/compiler_support
+        #if defined(__GNUC__) && !defined(__clang__) && __GNUC__ < 8
+            #undef JSON_HAS_FILESYSTEM
+            #undef JSON_HAS_EXPERIMENTAL_FILESYSTEM
+        #endif
+
+        // no filesystem support before Clang 7: https://en.cppreference.com/w/cpp/compiler_support
+        #if defined(__clang_major__) && __clang_major__ < 7
+            #undef JSON_HAS_FILESYSTEM
+            #undef JSON_HAS_EXPERIMENTAL_FILESYSTEM
+        #endif
+
+        // no filesystem support before MSVC 19.14: https://en.cppreference.com/w/cpp/compiler_support
+        #if defined(_MSC_VER) && _MSC_VER < 1914
+            #undef JSON_HAS_FILESYSTEM
+            #undef JSON_HAS_EXPERIMENTAL_FILESYSTEM
+        #endif
+
+        // no filesystem support before iOS 13
+        #if defined(__IPHONE_OS_VERSION_MIN_REQUIRED) && __IPHONE_OS_VERSION_MIN_REQUIRED < 130000
+            #undef JSON_HAS_FILESYSTEM
+            #undef JSON_HAS_EXPERIMENTAL_FILESYSTEM
+        #endif
+
+        // no filesystem support before macOS Catalina
+        #if defined(__MAC_OS_X_VERSION_MIN_REQUIRED) && __MAC_OS_X_VERSION_MIN_REQUIRED < 101500
+            #undef JSON_HAS_FILESYSTEM
+            #undef JSON_HAS_EXPERIMENTAL_FILESYSTEM
+        #endif
+    #endif
+#endif
+
+#ifndef JSON_HAS_EXPERIMENTAL_FILESYSTEM
+    #define JSON_HAS_EXPERIMENTAL_FILESYSTEM 0
+#endif
+
+#ifndef JSON_HAS_FILESYSTEM
+    #define JSON_HAS_FILESYSTEM 0
+#endif
+
+#ifndef JSON_HAS_THREE_WAY_COMPARISON
+    #if defined(__cpp_impl_three_way_comparison) && __cpp_impl_three_way_comparison >= 201907L \
+        && defined(__cpp_lib_three_way_comparison) && __cpp_lib_three_way_comparison >= 201907L
+        #define JSON_HAS_THREE_WAY_COMPARISON 1
+    #else
+        #define JSON_HAS_THREE_WAY_COMPARISON 0
+    #endif
+#endif
+
+#ifndef JSON_HAS_RANGES
+    // ranges header shipping in GCC 11.1.0 (released 2021-04-27) has syntax error
+    #if defined(__GLIBCXX__) && __GLIBCXX__ == 20210427
+        #define JSON_HAS_RANGES 0
+    #elif defined(__cpp_lib_ranges)
+        #define JSON_HAS_RANGES 1
+    #else
+        #define JSON_HAS_RANGES 0
+    #endif
+#endif
+
+#ifdef JSON_HAS_CPP_17
+    #define JSON_INLINE_VARIABLE inline
+#else
+    #define JSON_INLINE_VARIABLE
+#endif
+
+#if JSON_HEDLEY_HAS_ATTRIBUTE(no_unique_address)
+    #define JSON_NO_UNIQUE_ADDRESS [[no_unique_address]]
+#else
+    #define JSON_NO_UNIQUE_ADDRESS
+#endif
+
+// disable documentation warnings on clang
+#if defined(__clang__)
+    #pragma clang diagnostic push
+    #pragma clang diagnostic ignored "-Wdocumentation"
+    #pragma clang diagnostic ignored "-Wdocumentation-unknown-command"
+#endif
+
+// allow disabling exceptions
+#if (defined(__cpp_exceptions) || defined(__EXCEPTIONS) || defined(_CPPUNWIND)) && !defined(JSON_NOEXCEPTION)
+    #define JSON_THROW(exception) throw exception
+    #define JSON_TRY try
+    #define JSON_CATCH(exception) catch(exception)
+    #define JSON_INTERNAL_CATCH(exception) catch(exception)
+#else
+    #include <cstdlib>
+    #define JSON_THROW(exception) std::abort()
+    #define JSON_TRY if(true)
+    #define JSON_CATCH(exception) if(false)
+    #define JSON_INTERNAL_CATCH(exception) if(false)
+#endif
+
+// override exception macros
+#if defined(JSON_THROW_USER)
+    #undef JSON_THROW
+    #define JSON_THROW JSON_THROW_USER
+#endif
+#if defined(JSON_TRY_USER)
+    #undef JSON_TRY
+    #define JSON_TRY JSON_TRY_USER
+#endif
+#if defined(JSON_CATCH_USER)
+    #undef JSON_CATCH
+    #define JSON_CATCH JSON_CATCH_USER
+    #undef JSON_INTERNAL_CATCH
+    #define JSON_INTERNAL_CATCH JSON_CATCH_USER
+#endif
+#if defined(JSON_INTERNAL_CATCH_USER)
+    #undef JSON_INTERNAL_CATCH
+    #define JSON_INTERNAL_CATCH JSON_INTERNAL_CATCH_USER
+#endif
+
+// allow overriding assert
+#if !defined(JSON_ASSERT)
+    #include <cassert> // assert
+    #define JSON_ASSERT(x) assert(x)
+#endif
+
+// allow to access some private functions (needed by the test suite)
+#if defined(JSON_TESTS_PRIVATE)
+    #define JSON_PRIVATE_UNLESS_TESTED public
+#else
+    #define JSON_PRIVATE_UNLESS_TESTED private
+#endif
+
+/*!
+@brief macro to briefly define a mapping between an enum and JSON
+@def NLOHMANN_JSON_SERIALIZE_ENUM
+@since version 3.4.0
+*/
+#define NLOHMANN_JSON_SERIALIZE_ENUM(ENUM_TYPE, ...)                                            \
+    template<typename BasicJsonType>                                                            \
+    inline void to_json(BasicJsonType& j, const ENUM_TYPE& e)                                   \
+    {                                                                                           \
+        static_assert(std::is_enum<ENUM_TYPE>::value, #ENUM_TYPE " must be an enum!");          \
+        static const std::pair<ENUM_TYPE, BasicJsonType> m[] = __VA_ARGS__;                     \
+        auto it = std::find_if(std::begin(m), std::end(m),                                      \
+                               [e](const std::pair<ENUM_TYPE, BasicJsonType>& ej_pair) -> bool  \
+        {                                                                                       \
+            return ej_pair.first == e;                                                          \
+        });                                                                                     \
+        j = ((it != std::end(m)) ? it : std::begin(m))->second;                                 \
+    }                                                                                           \
+    template<typename BasicJsonType>                                                            \
+    inline void from_json(const BasicJsonType& j, ENUM_TYPE& e)                                 \
+    {                                                                                           \
+        static_assert(std::is_enum<ENUM_TYPE>::value, #ENUM_TYPE " must be an enum!");          \
+        static const std::pair<ENUM_TYPE, BasicJsonType> m[] = __VA_ARGS__;                     \
+        auto it = std::find_if(std::begin(m), std::end(m),                                      \
+                               [&j](const std::pair<ENUM_TYPE, BasicJsonType>& ej_pair) -> bool \
+        {                                                                                       \
+            return ej_pair.second == j;                                                         \
+        });                                                                                     \
+        e = ((it != std::end(m)) ? it : std::begin(m))->first;                                  \
+    }
+
+// Ugly macros to avoid uglier copy-paste when specializing basic_json. They
+// may be removed in the future once the class is split.
+
+#define NLOHMANN_BASIC_JSON_TPL_DECLARATION                                \
+    template<template<typename, typename, typename...> class ObjectType,   \
+             template<typename, typename...> class ArrayType,              \
+             class StringType, class BooleanType, class NumberIntegerType, \
+             class NumberUnsignedType, class NumberFloatType,              \
+             template<typename> class AllocatorType,                       \
+             template<typename, typename = void> class JSONSerializer,     \
+             class BinaryType>
+
+#define NLOHMANN_BASIC_JSON_TPL                                            \
+    basic_json<ObjectType, ArrayType, StringType, BooleanType,             \
+    NumberIntegerType, NumberUnsignedType, NumberFloatType,                \
+    AllocatorType, JSONSerializer, BinaryType>
+
+// Macros to simplify conversion from/to types
+
+#define NLOHMANN_JSON_EXPAND( x ) x
+#define NLOHMANN_JSON_GET_MACRO(_1, _2, _3, _4, _5, _6, _7, _8, _9, _10, _11, _12, _13, _14, _15, _16, _17, _18, _19, _20, _21, _22, _23, _24, _25, _26, _27, _28, _29, _30, _31, _32, _33, _34, _35, _36, _37, _38, _39, _40, _41, _42, _43, _44, _45, _46, _47, _48, _49, _50, _51, _52, _53, _54, _55, _56, _57, _58, _59, _60, _61, _62, _63, _64, NAME,...) NAME
+#define NLOHMANN_JSON_PASTE(...) NLOHMANN_JSON_EXPAND(NLOHMANN_JSON_GET_MACRO(__VA_ARGS__, \
+        NLOHMANN_JSON_PASTE64, \
+        NLOHMANN_JSON_PASTE63, \
+        NLOHMANN_JSON_PASTE62, \
+        NLOHMANN_JSON_PASTE61, \
+        NLOHMANN_JSON_PASTE60, \
+        NLOHMANN_JSON_PASTE59, \
+        NLOHMANN_JSON_PASTE58, \
+        NLOHMANN_JSON_PASTE57, \
+        NLOHMANN_JSON_PASTE56, \
+        NLOHMANN_JSON_PASTE55, \
+        NLOHMANN_JSON_PASTE54, \
+        NLOHMANN_JSON_PASTE53, \
+        NLOHMANN_JSON_PASTE52, \
+        NLOHMANN_JSON_PASTE51, \
+        NLOHMANN_JSON_PASTE50, \
+        NLOHMANN_JSON_PASTE49, \
+        NLOHMANN_JSON_PASTE48, \
+        NLOHMANN_JSON_PASTE47, \
+        NLOHMANN_JSON_PASTE46, \
+        NLOHMANN_JSON_PASTE45, \
+        NLOHMANN_JSON_PASTE44, \
+        NLOHMANN_JSON_PASTE43, \
+        NLOHMANN_JSON_PASTE42, \
+        NLOHMANN_JSON_PASTE41, \
+        NLOHMANN_JSON_PASTE40, \
+        NLOHMANN_JSON_PASTE39, \
+        NLOHMANN_JSON_PASTE38, \
+        NLOHMANN_JSON_PASTE37, \
+        NLOHMANN_JSON_PASTE36, \
+        NLOHMANN_JSON_PASTE35, \
+        NLOHMANN_JSON_PASTE34, \
+        NLOHMANN_JSON_PASTE33, \
+        NLOHMANN_JSON_PASTE32, \
+        NLOHMANN_JSON_PASTE31, \
+        NLOHMANN_JSON_PASTE30, \
+        NLOHMANN_JSON_PASTE29, \
+        NLOHMANN_JSON_PASTE28, \
+        NLOHMANN_JSON_PASTE27, \
+        NLOHMANN_JSON_PASTE26, \
+        NLOHMANN_JSON_PASTE25, \
+        NLOHMANN_JSON_PASTE24, \
+        NLOHMANN_JSON_PASTE23, \
+        NLOHMANN_JSON_PASTE22, \
+        NLOHMANN_JSON_PASTE21, \
+        NLOHMANN_JSON_PASTE20, \
+        NLOHMANN_JSON_PASTE19, \
+        NLOHMANN_JSON_PASTE18, \
+        NLOHMANN_JSON_PASTE17, \
+        NLOHMANN_JSON_PASTE16, \
+        NLOHMANN_JSON_PASTE15, \
+        NLOHMANN_JSON_PASTE14, \
+        NLOHMANN_JSON_PASTE13, \
+        NLOHMANN_JSON_PASTE12, \
+        NLOHMANN_JSON_PASTE11, \
+        NLOHMANN_JSON_PASTE10, \
+        NLOHMANN_JSON_PASTE9, \
+        NLOHMANN_JSON_PASTE8, \
+        NLOHMANN_JSON_PASTE7, \
+        NLOHMANN_JSON_PASTE6, \
+        NLOHMANN_JSON_PASTE5, \
+        NLOHMANN_JSON_PASTE4, \
+        NLOHMANN_JSON_PASTE3, \
+        NLOHMANN_JSON_PASTE2, \
+        NLOHMANN_JSON_PASTE1)(__VA_ARGS__))
+#define NLOHMANN_JSON_PASTE2(func, v1) func(v1)
+#define NLOHMANN_JSON_PASTE3(func, v1, v2) NLOHMANN_JSON_PASTE2(func, v1) NLOHMANN_JSON_PASTE2(func, v2)
+#define NLOHMANN_JSON_PASTE4(func, v1, v2, v3) NLOHMANN_JSON_PASTE2(func, v1) NLOHMANN_JSON_PASTE3(func, v2, v3)
+#define NLOHMANN_JSON_PASTE5(func, v1, v2, v3, v4) NLOHMANN_JSON_PASTE2(func, v1) NLOHMANN_JSON_PASTE4(func, v2, v3, v4)
+#define NLOHMANN_JSON_PASTE6(func, v1, v2, v3, v4, v5) NLOHMANN_JSON_PASTE2(func, v1) NLOHMANN_JSON_PASTE5(func, v2, v3, v4, v5)
+#define NLOHMANN_JSON_PASTE7(func, v1, v2, v3, v4, v5, v6) NLOHMANN_JSON_PASTE2(func, v1) NLOHMANN_JSON_PASTE6(func, v2, v3, v4, v5, v6)
+#define NLOHMANN_JSON_PASTE8(func, v1, v2, v3, v4, v5, v6, v7) NLOHMANN_JSON_PASTE2(func, v1) NLOHMANN_JSON_PASTE7(func, v2, v3, v4, v5, v6, v7)
+#define NLOHMANN_JSON_PASTE9(func, v1, v2, v3, v4, v5, v6, v7, v8) NLOHMANN_JSON_PASTE2(func, v1) NLOHMANN_JSON_PASTE8(func, v2, v3, v4, v5, v6, v7, v8)
+#define NLOHMANN_JSON_PASTE10(func, v1, v2, v3, v4, v5, v6, v7, v8, v9) NLOHMANN_JSON_PASTE2(func, v1) NLOHMANN_JSON_PASTE9(func, v2, v3, v4, v5, v6, v7, v8, v9)
+#define NLOHMANN_JSON_PASTE11(func, v1, v2, v3, v4, v5, v6, v7, v8, v9, v10) NLOHMANN_JSON_PASTE2(func, v1) NLOHMANN_JSON_PASTE10(func, v2, v3, v4, v5, v6, v7, v8, v9, v10)
+#define NLOHMANN_JSON_PASTE12(func, v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11) NLOHMANN_JSON_PASTE2(func, v1) NLOHMANN_JSON_PASTE11(func, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11)
+#define NLOHMANN_JSON_PASTE13(func, v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12) NLOHMANN_JSON_PASTE2(func, v1) NLOHMANN_JSON_PASTE12(func, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12)
+#define NLOHMANN_JSON_PASTE14(func, v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13) NLOHMANN_JSON_PASTE2(func, v1) NLOHMANN_JSON_PASTE13(func, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13)
+#define NLOHMANN_JSON_PASTE15(func, v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14) NLOHMANN_JSON_PASTE2(func, v1) NLOHMANN_JSON_PASTE14(func, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14)
+#define NLOHMANN_JSON_PASTE16(func, v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15) NLOHMANN_JSON_PASTE2(func, v1) NLOHMANN_JSON_PASTE15(func, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15)
+#define NLOHMANN_JSON_PASTE17(func, v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16) NLOHMANN_JSON_PASTE2(func, v1) NLOHMANN_JSON_PASTE16(func, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16)
+#define NLOHMANN_JSON_PASTE18(func, v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17) NLOHMANN_JSON_PASTE2(func, v1) NLOHMANN_JSON_PASTE17(func, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17)
+#define NLOHMANN_JSON_PASTE19(func, v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18) NLOHMANN_JSON_PASTE2(func, v1) NLOHMANN_JSON_PASTE18(func, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18)
+#define NLOHMANN_JSON_PASTE20(func, v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19) NLOHMANN_JSON_PASTE2(func, v1) NLOHMANN_JSON_PASTE19(func, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19)
+#define NLOHMANN_JSON_PASTE21(func, v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20) NLOHMANN_JSON_PASTE2(func, v1) NLOHMANN_JSON_PASTE20(func, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20)
+#define NLOHMANN_JSON_PASTE22(func, v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21) NLOHMANN_JSON_PASTE2(func, v1) NLOHMANN_JSON_PASTE21(func, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21)
+#define NLOHMANN_JSON_PASTE23(func, v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22) NLOHMANN_JSON_PASTE2(func, v1) NLOHMANN_JSON_PASTE22(func, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22)
+#define NLOHMANN_JSON_PASTE24(func, v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23) NLOHMANN_JSON_PASTE2(func, v1) NLOHMANN_JSON_PASTE23(func, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23)
+#define NLOHMANN_JSON_PASTE25(func, v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24) NLOHMANN_JSON_PASTE2(func, v1) NLOHMANN_JSON_PASTE24(func, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24)
+#define NLOHMANN_JSON_PASTE26(func, v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25) NLOHMANN_JSON_PASTE2(func, v1) NLOHMANN_JSON_PASTE25(func, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25)
+#define NLOHMANN_JSON_PASTE27(func, v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26) NLOHMANN_JSON_PASTE2(func, v1) NLOHMANN_JSON_PASTE26(func, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26)
+#define NLOHMANN_JSON_PASTE28(func, v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27) NLOHMANN_JSON_PASTE2(func, v1) NLOHMANN_JSON_PASTE27(func, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27)
+#define NLOHMANN_JSON_PASTE29(func, v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27, v28) NLOHMANN_JSON_PASTE2(func, v1) NLOHMANN_JSON_PASTE28(func, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27, v28)
+#define NLOHMANN_JSON_PASTE30(func, v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27, v28, v29) NLOHMANN_JSON_PASTE2(func, v1) NLOHMANN_JSON_PASTE29(func, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27, v28, v29)
+#define NLOHMANN_JSON_PASTE31(func, v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27, v28, v29, v30) NLOHMANN_JSON_PASTE2(func, v1) NLOHMANN_JSON_PASTE30(func, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27, v28, v29, v30)
+#define NLOHMANN_JSON_PASTE32(func, v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27, v28, v29, v30, v31) NLOHMANN_JSON_PASTE2(func, v1) NLOHMANN_JSON_PASTE31(func, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27, v28, v29, v30, v31)
+#define NLOHMANN_JSON_PASTE33(func, v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27, v28, v29, v30, v31, v32) NLOHMANN_JSON_PASTE2(func, v1) NLOHMANN_JSON_PASTE32(func, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27, v28, v29, v30, v31, v32)
+#define NLOHMANN_JSON_PASTE34(func, v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27, v28, v29, v30, v31, v32, v33) NLOHMANN_JSON_PASTE2(func, v1) NLOHMANN_JSON_PASTE33(func, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27, v28, v29, v30, v31, v32, v33)
+#define NLOHMANN_JSON_PASTE35(func, v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27, v28, v29, v30, v31, v32, v33, v34) NLOHMANN_JSON_PASTE2(func, v1) NLOHMANN_JSON_PASTE34(func, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27, v28, v29, v30, v31, v32, v33, v34)
+#define NLOHMANN_JSON_PASTE36(func, v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27, v28, v29, v30, v31, v32, v33, v34, v35) NLOHMANN_JSON_PASTE2(func, v1) NLOHMANN_JSON_PASTE35(func, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27, v28, v29, v30, v31, v32, v33, v34, v35)
+#define NLOHMANN_JSON_PASTE37(func, v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27, v28, v29, v30, v31, v32, v33, v34, v35, v36) NLOHMANN_JSON_PASTE2(func, v1) NLOHMANN_JSON_PASTE36(func, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27, v28, v29, v30, v31, v32, v33, v34, v35, v36)
+#define NLOHMANN_JSON_PASTE38(func, v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27, v28, v29, v30, v31, v32, v33, v34, v35, v36, v37) NLOHMANN_JSON_PASTE2(func, v1) NLOHMANN_JSON_PASTE37(func, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27, v28, v29, v30, v31, v32, v33, v34, v35, v36, v37)
+#define NLOHMANN_JSON_PASTE39(func, v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27, v28, v29, v30, v31, v32, v33, v34, v35, v36, v37, v38) NLOHMANN_JSON_PASTE2(func, v1) NLOHMANN_JSON_PASTE38(func, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27, v28, v29, v30, v31, v32, v33, v34, v35, v36, v37, v38)
+#define NLOHMANN_JSON_PASTE40(func, v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27, v28, v29, v30, v31, v32, v33, v34, v35, v36, v37, v38, v39) NLOHMANN_JSON_PASTE2(func, v1) NLOHMANN_JSON_PASTE39(func, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27, v28, v29, v30, v31, v32, v33, v34, v35, v36, v37, v38, v39)
+#define NLOHMANN_JSON_PASTE41(func, v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27, v28, v29, v30, v31, v32, v33, v34, v35, v36, v37, v38, v39, v40) NLOHMANN_JSON_PASTE2(func, v1) NLOHMANN_JSON_PASTE40(func, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27, v28, v29, v30, v31, v32, v33, v34, v35, v36, v37, v38, v39, v40)
+#define NLOHMANN_JSON_PASTE42(func, v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27, v28, v29, v30, v31, v32, v33, v34, v35, v36, v37, v38, v39, v40, v41) NLOHMANN_JSON_PASTE2(func, v1) NLOHMANN_JSON_PASTE41(func, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27, v28, v29, v30, v31, v32, v33, v34, v35, v36, v37, v38, v39, v40, v41)
+#define NLOHMANN_JSON_PASTE43(func, v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27, v28, v29, v30, v31, v32, v33, v34, v35, v36, v37, v38, v39, v40, v41, v42) NLOHMANN_JSON_PASTE2(func, v1) NLOHMANN_JSON_PASTE42(func, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27, v28, v29, v30, v31, v32, v33, v34, v35, v36, v37, v38, v39, v40, v41, v42)
+#define NLOHMANN_JSON_PASTE44(func, v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27, v28, v29, v30, v31, v32, v33, v34, v35, v36, v37, v38, v39, v40, v41, v42, v43) NLOHMANN_JSON_PASTE2(func, v1) NLOHMANN_JSON_PASTE43(func, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27, v28, v29, v30, v31, v32, v33, v34, v35, v36, v37, v38, v39, v40, v41, v42, v43)
+#define NLOHMANN_JSON_PASTE45(func, v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27, v28, v29, v30, v31, v32, v33, v34, v35, v36, v37, v38, v39, v40, v41, v42, v43, v44) NLOHMANN_JSON_PASTE2(func, v1) NLOHMANN_JSON_PASTE44(func, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27, v28, v29, v30, v31, v32, v33, v34, v35, v36, v37, v38, v39, v40, v41, v42, v43, v44)
+#define NLOHMANN_JSON_PASTE46(func, v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27, v28, v29, v30, v31, v32, v33, v34, v35, v36, v37, v38, v39, v40, v41, v42, v43, v44, v45) NLOHMANN_JSON_PASTE2(func, v1) NLOHMANN_JSON_PASTE45(func, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27, v28, v29, v30, v31, v32, v33, v34, v35, v36, v37, v38, v39, v40, v41, v42, v43, v44, v45)
+#define NLOHMANN_JSON_PASTE47(func, v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27, v28, v29, v30, v31, v32, v33, v34, v35, v36, v37, v38, v39, v40, v41, v42, v43, v44, v45, v46) NLOHMANN_JSON_PASTE2(func, v1) NLOHMANN_JSON_PASTE46(func, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27, v28, v29, v30, v31, v32, v33, v34, v35, v36, v37, v38, v39, v40, v41, v42, v43, v44, v45, v46)
+#define NLOHMANN_JSON_PASTE48(func, v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27, v28, v29, v30, v31, v32, v33, v34, v35, v36, v37, v38, v39, v40, v41, v42, v43, v44, v45, v46, v47) NLOHMANN_JSON_PASTE2(func, v1) NLOHMANN_JSON_PASTE47(func, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27, v28, v29, v30, v31, v32, v33, v34, v35, v36, v37, v38, v39, v40, v41, v42, v43, v44, v45, v46, v47)
+#define NLOHMANN_JSON_PASTE49(func, v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27, v28, v29, v30, v31, v32, v33, v34, v35, v36, v37, v38, v39, v40, v41, v42, v43, v44, v45, v46, v47, v48) NLOHMANN_JSON_PASTE2(func, v1) NLOHMANN_JSON_PASTE48(func, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27, v28, v29, v30, v31, v32, v33, v34, v35, v36, v37, v38, v39, v40, v41, v42, v43, v44, v45, v46, v47, v48)
+#define NLOHMANN_JSON_PASTE50(func, v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27, v28, v29, v30, v31, v32, v33, v34, v35, v36, v37, v38, v39, v40, v41, v42, v43, v44, v45, v46, v47, v48, v49) NLOHMANN_JSON_PASTE2(func, v1) NLOHMANN_JSON_PASTE49(func, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27, v28, v29, v30, v31, v32, v33, v34, v35, v36, v37, v38, v39, v40, v41, v42, v43, v44, v45, v46, v47, v48, v49)
+#define NLOHMANN_JSON_PASTE51(func, v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27, v28, v29, v30, v31, v32, v33, v34, v35, v36, v37, v38, v39, v40, v41, v42, v43, v44, v45, v46, v47, v48, v49, v50) NLOHMANN_JSON_PASTE2(func, v1) NLOHMANN_JSON_PASTE50(func, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27, v28, v29, v30, v31, v32, v33, v34, v35, v36, v37, v38, v39, v40, v41, v42, v43, v44, v45, v46, v47, v48, v49, v50)
+#define NLOHMANN_JSON_PASTE52(func, v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27, v28, v29, v30, v31, v32, v33, v34, v35, v36, v37, v38, v39, v40, v41, v42, v43, v44, v45, v46, v47, v48, v49, v50, v51) NLOHMANN_JSON_PASTE2(func, v1) NLOHMANN_JSON_PASTE51(func, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27, v28, v29, v30, v31, v32, v33, v34, v35, v36, v37, v38, v39, v40, v41, v42, v43, v44, v45, v46, v47, v48, v49, v50, v51)
+#define NLOHMANN_JSON_PASTE53(func, v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27, v28, v29, v30, v31, v32, v33, v34, v35, v36, v37, v38, v39, v40, v41, v42, v43, v44, v45, v46, v47, v48, v49, v50, v51, v52) NLOHMANN_JSON_PASTE2(func, v1) NLOHMANN_JSON_PASTE52(func, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27, v28, v29, v30, v31, v32, v33, v34, v35, v36, v37, v38, v39, v40, v41, v42, v43, v44, v45, v46, v47, v48, v49, v50, v51, v52)
+#define NLOHMANN_JSON_PASTE54(func, v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27, v28, v29, v30, v31, v32, v33, v34, v35, v36, v37, v38, v39, v40, v41, v42, v43, v44, v45, v46, v47, v48, v49, v50, v51, v52, v53) NLOHMANN_JSON_PASTE2(func, v1) NLOHMANN_JSON_PASTE53(func, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27, v28, v29, v30, v31, v32, v33, v34, v35, v36, v37, v38, v39, v40, v41, v42, v43, v44, v45, v46, v47, v48, v49, v50, v51, v52, v53)
+#define NLOHMANN_JSON_PASTE55(func, v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27, v28, v29, v30, v31, v32, v33, v34, v35, v36, v37, v38, v39, v40, v41, v42, v43, v44, v45, v46, v47, v48, v49, v50, v51, v52, v53, v54) NLOHMANN_JSON_PASTE2(func, v1) NLOHMANN_JSON_PASTE54(func, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27, v28, v29, v30, v31, v32, v33, v34, v35, v36, v37, v38, v39, v40, v41, v42, v43, v44, v45, v46, v47, v48, v49, v50, v51, v52, v53, v54)
+#define NLOHMANN_JSON_PASTE56(func, v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27, v28, v29, v30, v31, v32, v33, v34, v35, v36, v37, v38, v39, v40, v41, v42, v43, v44, v45, v46, v47, v48, v49, v50, v51, v52, v53, v54, v55) NLOHMANN_JSON_PASTE2(func, v1) NLOHMANN_JSON_PASTE55(func, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27, v28, v29, v30, v31, v32, v33, v34, v35, v36, v37, v38, v39, v40, v41, v42, v43, v44, v45, v46, v47, v48, v49, v50, v51, v52, v53, v54, v55)
+#define NLOHMANN_JSON_PASTE57(func, v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27, v28, v29, v30, v31, v32, v33, v34, v35, v36, v37, v38, v39, v40, v41, v42, v43, v44, v45, v46, v47, v48, v49, v50, v51, v52, v53, v54, v55, v56) NLOHMANN_JSON_PASTE2(func, v1) NLOHMANN_JSON_PASTE56(func, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27, v28, v29, v30, v31, v32, v33, v34, v35, v36, v37, v38, v39, v40, v41, v42, v43, v44, v45, v46, v47, v48, v49, v50, v51, v52, v53, v54, v55, v56)
+#define NLOHMANN_JSON_PASTE58(func, v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27, v28, v29, v30, v31, v32, v33, v34, v35, v36, v37, v38, v39, v40, v41, v42, v43, v44, v45, v46, v47, v48, v49, v50, v51, v52, v53, v54, v55, v56, v57) NLOHMANN_JSON_PASTE2(func, v1) NLOHMANN_JSON_PASTE57(func, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27, v28, v29, v30, v31, v32, v33, v34, v35, v36, v37, v38, v39, v40, v41, v42, v43, v44, v45, v46, v47, v48, v49, v50, v51, v52, v53, v54, v55, v56, v57)
+#define NLOHMANN_JSON_PASTE59(func, v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27, v28, v29, v30, v31, v32, v33, v34, v35, v36, v37, v38, v39, v40, v41, v42, v43, v44, v45, v46, v47, v48, v49, v50, v51, v52, v53, v54, v55, v56, v57, v58) NLOHMANN_JSON_PASTE2(func, v1) NLOHMANN_JSON_PASTE58(func, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27, v28, v29, v30, v31, v32, v33, v34, v35, v36, v37, v38, v39, v40, v41, v42, v43, v44, v45, v46, v47, v48, v49, v50, v51, v52, v53, v54, v55, v56, v57, v58)
+#define NLOHMANN_JSON_PASTE60(func, v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27, v28, v29, v30, v31, v32, v33, v34, v35, v36, v37, v38, v39, v40, v41, v42, v43, v44, v45, v46, v47, v48, v49, v50, v51, v52, v53, v54, v55, v56, v57, v58, v59) NLOHMANN_JSON_PASTE2(func, v1) NLOHMANN_JSON_PASTE59(func, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27, v28, v29, v30, v31, v32, v33, v34, v35, v36, v37, v38, v39, v40, v41, v42, v43, v44, v45, v46, v47, v48, v49, v50, v51, v52, v53, v54, v55, v56, v57, v58, v59)
+#define NLOHMANN_JSON_PASTE61(func, v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27, v28, v29, v30, v31, v32, v33, v34, v35, v36, v37, v38, v39, v40, v41, v42, v43, v44, v45, v46, v47, v48, v49, v50, v51, v52, v53, v54, v55, v56, v57, v58, v59, v60) NLOHMANN_JSON_PASTE2(func, v1) NLOHMANN_JSON_PASTE60(func, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27, v28, v29, v30, v31, v32, v33, v34, v35, v36, v37, v38, v39, v40, v41, v42, v43, v44, v45, v46, v47, v48, v49, v50, v51, v52, v53, v54, v55, v56, v57, v58, v59, v60)
+#define NLOHMANN_JSON_PASTE62(func, v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27, v28, v29, v30, v31, v32, v33, v34, v35, v36, v37, v38, v39, v40, v41, v42, v43, v44, v45, v46, v47, v48, v49, v50, v51, v52, v53, v54, v55, v56, v57, v58, v59, v60, v61) NLOHMANN_JSON_PASTE2(func, v1) NLOHMANN_JSON_PASTE61(func, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27, v28, v29, v30, v31, v32, v33, v34, v35, v36, v37, v38, v39, v40, v41, v42, v43, v44, v45, v46, v47, v48, v49, v50, v51, v52, v53, v54, v55, v56, v57, v58, v59, v60, v61)
+#define NLOHMANN_JSON_PASTE63(func, v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27, v28, v29, v30, v31, v32, v33, v34, v35, v36, v37, v38, v39, v40, v41, v42, v43, v44, v45, v46, v47, v48, v49, v50, v51, v52, v53, v54, v55, v56, v57, v58, v59, v60, v61, v62) NLOHMANN_JSON_PASTE2(func, v1) NLOHMANN_JSON_PASTE62(func, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27, v28, v29, v30, v31, v32, v33, v34, v35, v36, v37, v38, v39, v40, v41, v42, v43, v44, v45, v46, v47, v48, v49, v50, v51, v52, v53, v54, v55, v56, v57, v58, v59, v60, v61, v62)
+#define NLOHMANN_JSON_PASTE64(func, v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27, v28, v29, v30, v31, v32, v33, v34, v35, v36, v37, v38, v39, v40, v41, v42, v43, v44, v45, v46, v47, v48, v49, v50, v51, v52, v53, v54, v55, v56, v57, v58, v59, v60, v61, v62, v63) NLOHMANN_JSON_PASTE2(func, v1) NLOHMANN_JSON_PASTE63(func, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15, v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27, v28, v29, v30, v31, v32, v33, v34, v35, v36, v37, v38, v39, v40, v41, v42, v43, v44, v45, v46, v47, v48, v49, v50, v51, v52, v53, v54, v55, v56, v57, v58, v59, v60, v61, v62, v63)
+
+#define NLOHMANN_JSON_TO(v1) nlohmann_json_j[#v1] = nlohmann_json_t.v1;
+#define NLOHMANN_JSON_FROM(v1) nlohmann_json_j.at(#v1).get_to(nlohmann_json_t.v1);
+#define NLOHMANN_JSON_FROM_WITH_DEFAULT(v1) nlohmann_json_t.v1 = nlohmann_json_j.value(#v1, nlohmann_json_default_obj.v1);
+
+/*!
+@brief macro
+@def NLOHMANN_DEFINE_TYPE_INTRUSIVE
+@since version 3.9.0
+*/
+#define NLOHMANN_DEFINE_TYPE_INTRUSIVE(Type, ...)  \
+    friend void to_json(nlohmann::json& nlohmann_json_j, const Type& nlohmann_json_t) { NLOHMANN_JSON_EXPAND(NLOHMANN_JSON_PASTE(NLOHMANN_JSON_TO, __VA_ARGS__)) } \
+    friend void from_json(const nlohmann::json& nlohmann_json_j, Type& nlohmann_json_t) { NLOHMANN_JSON_EXPAND(NLOHMANN_JSON_PASTE(NLOHMANN_JSON_FROM, __VA_ARGS__)) }
+
+#define NLOHMANN_DEFINE_TYPE_INTRUSIVE_WITH_DEFAULT(Type, ...)  \
+    friend void to_json(nlohmann::json& nlohmann_json_j, const Type& nlohmann_json_t) { NLOHMANN_JSON_EXPAND(NLOHMANN_JSON_PASTE(NLOHMANN_JSON_TO, __VA_ARGS__)) } \
+    friend void from_json(const nlohmann::json& nlohmann_json_j, Type& nlohmann_json_t) { Type nlohmann_json_default_obj; NLOHMANN_JSON_EXPAND(NLOHMANN_JSON_PASTE(NLOHMANN_JSON_FROM_WITH_DEFAULT, __VA_ARGS__)) }
+
+/*!
+@brief macro
+@def NLOHMANN_DEFINE_TYPE_NON_INTRUSIVE
+@since version 3.9.0
+*/
+#define NLOHMANN_DEFINE_TYPE_NON_INTRUSIVE(Type, ...)  \
+    inline void to_json(nlohmann::json& nlohmann_json_j, const Type& nlohmann_json_t) { NLOHMANN_JSON_EXPAND(NLOHMANN_JSON_PASTE(NLOHMANN_JSON_TO, __VA_ARGS__)) } \
+    inline void from_json(const nlohmann::json& nlohmann_json_j, Type& nlohmann_json_t) { NLOHMANN_JSON_EXPAND(NLOHMANN_JSON_PASTE(NLOHMANN_JSON_FROM, __VA_ARGS__)) }
+
+#define NLOHMANN_DEFINE_TYPE_NON_INTRUSIVE_WITH_DEFAULT(Type, ...)  \
+    inline void to_json(nlohmann::json& nlohmann_json_j, const Type& nlohmann_json_t) { NLOHMANN_JSON_EXPAND(NLOHMANN_JSON_PASTE(NLOHMANN_JSON_TO, __VA_ARGS__)) } \
+    inline void from_json(const nlohmann::json& nlohmann_json_j, Type& nlohmann_json_t) { Type nlohmann_json_default_obj; NLOHMANN_JSON_EXPAND(NLOHMANN_JSON_PASTE(NLOHMANN_JSON_FROM_WITH_DEFAULT, __VA_ARGS__)) }
+
+
+// inspired from https://stackoverflow.com/a/26745591
+// allows to call any std function as if (e.g. with begin):
+// using std::begin; begin(x);
+//
+// it allows using the detected idiom to retrieve the return type
+// of such an expression
+#define NLOHMANN_CAN_CALL_STD_FUNC_IMPL(std_name)                                 \
+    namespace detail {                                                            \
+    using std::std_name;                                                          \
+    \
+    template<typename... T>                                                       \
+    using result_of_##std_name = decltype(std_name(std::declval<T>()...));        \
+    }                                                                             \
+    \
+    namespace detail2 {                                                           \
+    struct std_name##_tag                                                         \
+    {                                                                             \
+    };                                                                            \
+    \
+    template<typename... T>                                                       \
+    std_name##_tag std_name(T&&...);                                              \
+    \
+    template<typename... T>                                                       \
+    using result_of_##std_name = decltype(std_name(std::declval<T>()...));        \
+    \
+    template<typename... T>                                                       \
+    struct would_call_std_##std_name                                              \
+    {                                                                             \
+        static constexpr auto const value = ::nlohmann::detail::                  \
+                                            is_detected_exact<std_name##_tag, result_of_##std_name, T...>::value; \
+    };                                                                            \
+    } /* namespace detail2 */ \
+    \
+    template<typename... T>                                                       \
+    struct would_call_std_##std_name : detail2::would_call_std_##std_name<T...>   \
+    {                                                                             \
+    }
+
+#ifndef JSON_USE_IMPLICIT_CONVERSIONS
+    #define JSON_USE_IMPLICIT_CONVERSIONS 1
+#endif
+
+#if JSON_USE_IMPLICIT_CONVERSIONS
+    #define JSON_EXPLICIT
+#else
+    #define JSON_EXPLICIT explicit
+#endif
+
+#ifndef JSON_DISABLE_ENUM_SERIALIZATION
+    #define JSON_DISABLE_ENUM_SERIALIZATION 0
+#endif
+
+#ifndef JSON_USE_GLOBAL_UDLS
+    #define JSON_USE_GLOBAL_UDLS 1
+#endif
+
+#if JSON_HAS_THREE_WAY_COMPARISON
+    #include <compare> // partial_ordering
+#endif
+
+NLOHMANN_JSON_NAMESPACE_BEGIN
+namespace detail
+{
+
+///////////////////////////
+// JSON type enumeration //
+///////////////////////////
+
+/*!
+@brief the JSON type enumeration
+
+This enumeration collects the different JSON types. It is internally used to
+distinguish the stored values, and the functions @ref basic_json::is_null(),
+@ref basic_json::is_object(), @ref basic_json::is_array(),
+@ref basic_json::is_string(), @ref basic_json::is_boolean(),
+@ref basic_json::is_number() (with @ref basic_json::is_number_integer(),
+@ref basic_json::is_number_unsigned(), and @ref basic_json::is_number_float()),
+@ref basic_json::is_discarded(), @ref basic_json::is_primitive(), and
+@ref basic_json::is_structured() rely on it.
+
+@note There are three enumeration entries (number_integer, number_unsigned, and
+number_float), because the library distinguishes these three types for numbers:
+@ref basic_json::number_unsigned_t is used for unsigned integers,
+@ref basic_json::number_integer_t is used for signed integers, and
+@ref basic_json::number_float_t is used for floating-point numbers or to
+approximate integers which do not fit in the limits of their respective type.
+
+@sa see @ref basic_json::basic_json(const value_t value_type) -- create a JSON
+value with the default value for a given type
+
+@since version 1.0.0
+*/
+enum class value_t : std::uint8_t
+{
+    null,             ///< null value
+    object,           ///< object (unordered set of name/value pairs)
+    array,            ///< array (ordered collection of values)
+    string,           ///< string value
+    boolean,          ///< boolean value
+    number_integer,   ///< number value (signed integer)
+    number_unsigned,  ///< number value (unsigned integer)
+    number_float,     ///< number value (floating-point)
+    binary,           ///< binary array (ordered collection of bytes)
+    discarded         ///< discarded by the parser callback function
+};
+
+/*!
+@brief comparison operator for JSON types
+
+Returns an ordering that is similar to Python:
+- order: null < boolean < number < object < array < string < binary
+- furthermore, each type is not smaller than itself
+- discarded values are not comparable
+- binary is represented as a b"" string in python and directly comparable to a
+  string; however, making a binary array directly comparable with a string would
+  be surprising behavior in a JSON file.
+
+@since version 1.0.0
+*/
+#if JSON_HAS_THREE_WAY_COMPARISON
+    inline std::partial_ordering operator<=>(const value_t lhs, const value_t rhs) noexcept // *NOPAD*
+#else
+    inline bool operator<(const value_t lhs, const value_t rhs) noexcept
+#endif
+{
+    static constexpr std::array<std::uint8_t, 9> order = {{
+            0 /* null */, 3 /* object */, 4 /* array */, 5 /* string */,
+            1 /* boolean */, 2 /* integer */, 2 /* unsigned */, 2 /* float */,
+            6 /* binary */
+        }
+    };
+
+    const auto l_index = static_cast<std::size_t>(lhs);
+    const auto r_index = static_cast<std::size_t>(rhs);
+#if JSON_HAS_THREE_WAY_COMPARISON
+    if (l_index < order.size() && r_index < order.size())
+    {
+        return order[l_index] <=> order[r_index]; // *NOPAD*
+    }
+    return std::partial_ordering::unordered;
+#else
+    return l_index < order.size() && r_index < order.size() && order[l_index] < order[r_index];
+#endif
+}
+
+// GCC selects the built-in operator< over an operator rewritten from
+// a user-defined spaceship operator
+// Clang, MSVC, and ICC select the rewritten candidate
+// (see GCC bug https://gcc.gnu.org/bugzilla/show_bug.cgi?id=105200)
+#if JSON_HAS_THREE_WAY_COMPARISON && defined(__GNUC__)
+inline bool operator<(const value_t lhs, const value_t rhs) noexcept
+{
+    return std::is_lt(lhs <=> rhs); // *NOPAD*
+}
+#endif
+
+}  // namespace detail
+NLOHMANN_JSON_NAMESPACE_END
+
+// #include <nlohmann/detail/string_escape.hpp>
+//     __ _____ _____ _____
+//  __|  |   __|     |   | |  JSON for Modern C++
+// |  |  |__   |  |  | | | |  version 3.11.2
+// |_____|_____|_____|_|___|  https://github.com/nlohmann/json
+//
+// SPDX-FileCopyrightText: 2013-2022 Niels Lohmann <https://nlohmann.me>
+// SPDX-License-Identifier: MIT
+
+
+
+// #include <nlohmann/detail/abi_macros.hpp>
+
+
+NLOHMANN_JSON_NAMESPACE_BEGIN
+namespace detail
+{
+
+/*!
+@brief replace all occurrences of a substring by another string
+
+@param[in,out] s  the string to manipulate; changed so that all
+               occurrences of @a f are replaced with @a t
+@param[in]     f  the substring to replace with @a t
+@param[in]     t  the string to replace @a f
+
+@pre The search string @a f must not be empty. **This precondition is
+enforced with an assertion.**
+
+@since version 2.0.0
+*/
+template<typename StringType>
+inline void replace_substring(StringType& s, const StringType& f,
+                              const StringType& t)
+{
+    JSON_ASSERT(!f.empty());
+    for (auto pos = s.find(f);                // find first occurrence of f
+            pos != StringType::npos;          // make sure f was found
+            s.replace(pos, f.size(), t),      // replace with t, and
+            pos = s.find(f, pos + t.size()))  // find next occurrence of f
+    {}
+}
+
+/*!
+ * @brief string escaping as described in RFC 6901 (Sect. 4)
+ * @param[in] s string to escape
+ * @return    escaped string
+ *
+ * Note the order of escaping "~" to "~0" and "/" to "~1" is important.
+ */
+template<typename StringType>
+inline StringType escape(StringType s)
+{
+    replace_substring(s, StringType{"~"}, StringType{"~0"});
+    replace_substring(s, StringType{"/"}, StringType{"~1"});
+    return s;
+}
+
+/*!
+ * @brief string unescaping as described in RFC 6901 (Sect. 4)
+ * @param[in] s string to unescape
+ * @return    unescaped string
+ *
+ * Note the order of escaping "~1" to "/" and "~0" to "~" is important.
+ */
+template<typename StringType>
+static void unescape(StringType& s)
+{
+    replace_substring(s, StringType{"~1"}, StringType{"/"});
+    replace_substring(s, StringType{"~0"}, StringType{"~"});
+}
+
+}  // namespace detail
+NLOHMANN_JSON_NAMESPACE_END
+
+// #include <nlohmann/detail/input/position_t.hpp>
+//     __ _____ _____ _____
+//  __|  |   __|     |   | |  JSON for Modern C++
+// |  |  |__   |  |  | | | |  version 3.11.2
+// |_____|_____|_____|_|___|  https://github.com/nlohmann/json
+//
+// SPDX-FileCopyrightText: 2013-2022 Niels Lohmann <https://nlohmann.me>
+// SPDX-License-Identifier: MIT
+
+
+
+#include <cstddef> // size_t
+
+// #include <nlohmann/detail/abi_macros.hpp>
+
+
+NLOHMANN_JSON_NAMESPACE_BEGIN
+namespace detail
+{
+
+/// struct to capture the start position of the current token
+struct position_t
+{
+    /// the total number of characters read
+    std::size_t chars_read_total = 0;
+    /// the number of characters read in the current line
+    std::size_t chars_read_current_line = 0;
+    /// the number of lines read
+    std::size_t lines_read = 0;
+
+    /// conversion to size_t to preserve SAX interface
+    constexpr operator size_t() const
+    {
+        return chars_read_total;
+    }
+};
+
+}  // namespace detail
+NLOHMANN_JSON_NAMESPACE_END
+
+// #include <nlohmann/detail/macro_scope.hpp>
+
+// #include <nlohmann/detail/meta/cpp_future.hpp>
+//     __ _____ _____ _____
+//  __|  |   __|     |   | |  JSON for Modern C++
+// |  |  |__   |  |  | | | |  version 3.11.2
+// |_____|_____|_____|_|___|  https://github.com/nlohmann/json
+//
+// SPDX-FileCopyrightText: 2013-2022 Niels Lohmann <https://nlohmann.me>
+// SPDX-FileCopyrightText: 2018 The Abseil Authors
+// SPDX-License-Identifier: MIT
+
+
+
+#include <array> // array
+#include <cstddef> // size_t
+#include <type_traits> // conditional, enable_if, false_type, integral_constant, is_constructible, is_integral, is_same, remove_cv, remove_reference, true_type
+#include <utility> // index_sequence, make_index_sequence, index_sequence_for
+
+// #include <nlohmann/detail/macro_scope.hpp>
+
+
+NLOHMANN_JSON_NAMESPACE_BEGIN
+namespace detail
+{
+
+template<typename T>
+using uncvref_t = typename std::remove_cv<typename std::remove_reference<T>::type>::type;
+
+#ifdef JSON_HAS_CPP_14
+
+// the following utilities are natively available in C++14
+using std::enable_if_t;
+using std::index_sequence;
+using std::make_index_sequence;
+using std::index_sequence_for;
+
+#else
+
+// alias templates to reduce boilerplate
+template<bool B, typename T = void>
+using enable_if_t = typename std::enable_if<B, T>::type;
+
+// The following code is taken from https://github.com/abseil/abseil-cpp/blob/10cb35e459f5ecca5b2ff107635da0bfa41011b4/absl/utility/utility.h
+// which is part of Google Abseil (https://github.com/abseil/abseil-cpp), licensed under the Apache License 2.0.
+
+//// START OF CODE FROM GOOGLE ABSEIL
+
+// integer_sequence
+//
+// Class template representing a compile-time integer sequence. An instantiation
+// of `integer_sequence<T, Ints...>` has a sequence of integers encoded in its
+// type through its template arguments (which is a common need when
+// working with C++11 variadic templates). `absl::integer_sequence` is designed
+// to be a drop-in replacement for C++14's `std::integer_sequence`.
+//
+// Example:
+//
+//   template< class T, T... Ints >
+//   void user_function(integer_sequence<T, Ints...>);
+//
+//   int main()
+//   {
+//     // user_function's `T` will be deduced to `int` and `Ints...`
+//     // will be deduced to `0, 1, 2, 3, 4`.
+//     user_function(make_integer_sequence<int, 5>());
+//   }
+template <typename T, T... Ints>
+struct integer_sequence
+{
+    using value_type = T;
+    static constexpr std::size_t size() noexcept
+    {
+        return sizeof...(Ints);
+    }
+};
+
+// index_sequence
+//
+// A helper template for an `integer_sequence` of `size_t`,
+// `absl::index_sequence` is designed to be a drop-in replacement for C++14's
+// `std::index_sequence`.
+template <size_t... Ints>
+using index_sequence = integer_sequence<size_t, Ints...>;
+
+namespace utility_internal
+{
+
+template <typename Seq, size_t SeqSize, size_t Rem>
+struct Extend;
+
+// Note that SeqSize == sizeof...(Ints). It's passed explicitly for efficiency.
+template <typename T, T... Ints, size_t SeqSize>
+struct Extend<integer_sequence<T, Ints...>, SeqSize, 0>
+{
+    using type = integer_sequence < T, Ints..., (Ints + SeqSize)... >;
+};
+
+template <typename T, T... Ints, size_t SeqSize>
+struct Extend<integer_sequence<T, Ints...>, SeqSize, 1>
+{
+    using type = integer_sequence < T, Ints..., (Ints + SeqSize)..., 2 * SeqSize >;
+};
+
+// Recursion helper for 'make_integer_sequence<T, N>'.
+// 'Gen<T, N>::type' is an alias for 'integer_sequence<T, 0, 1, ... N-1>'.
+template <typename T, size_t N>
+struct Gen
+{
+    using type =
+        typename Extend < typename Gen < T, N / 2 >::type, N / 2, N % 2 >::type;
+};
+
+template <typename T>
+struct Gen<T, 0>
+{
+    using type = integer_sequence<T>;
+};
+
+}  // namespace utility_internal
+
+// Compile-time sequences of integers
+
+// make_integer_sequence
+//
+// This template alias is equivalent to
+// `integer_sequence<int, 0, 1, ..., N-1>`, and is designed to be a drop-in
+// replacement for C++14's `std::make_integer_sequence`.
+template <typename T, T N>
+using make_integer_sequence = typename utility_internal::Gen<T, N>::type;
+
+// make_index_sequence
+//
+// This template alias is equivalent to `index_sequence<0, 1, ..., N-1>`,
+// and is designed to be a drop-in replacement for C++14's
+// `std::make_index_sequence`.
+template <size_t N>
+using make_index_sequence = make_integer_sequence<size_t, N>;
+
+// index_sequence_for
+//
+// Converts a typename pack into an index sequence of the same length, and
+// is designed to be a drop-in replacement for C++14's
+// `std::index_sequence_for()`
+template <typename... Ts>
+using index_sequence_for = make_index_sequence<sizeof...(Ts)>;
+
+//// END OF CODE FROM GOOGLE ABSEIL
+
+#endif
+
+// dispatch utility (taken from ranges-v3)
+template<unsigned N> struct priority_tag : priority_tag < N - 1 > {};
+template<> struct priority_tag<0> {};
+
+// taken from ranges-v3
+template<typename T>
+struct static_const
+{
+    static JSON_INLINE_VARIABLE constexpr T value{};
+};
+
+#ifndef JSON_HAS_CPP_17
+    template<typename T>
+    constexpr T static_const<T>::value;
+#endif
+
+template<typename T, typename... Args>
+inline constexpr std::array<T, sizeof...(Args)> make_array(Args&& ... args)
+{
+    return std::array<T, sizeof...(Args)> {{static_cast<T>(std::forward<Args>(args))...}};
+}
+
+}  // namespace detail
+NLOHMANN_JSON_NAMESPACE_END
+
+// #include <nlohmann/detail/meta/type_traits.hpp>
+//     __ _____ _____ _____
+//  __|  |   __|     |   | |  JSON for Modern C++
+// |  |  |__   |  |  | | | |  version 3.11.2
+// |_____|_____|_____|_|___|  https://github.com/nlohmann/json
+//
+// SPDX-FileCopyrightText: 2013-2022 Niels Lohmann <https://nlohmann.me>
+// SPDX-License-Identifier: MIT
+
+
+
+#include <limits> // numeric_limits
+#include <type_traits> // false_type, is_constructible, is_integral, is_same, true_type
+#include <utility> // declval
+#include <tuple> // tuple
+
+// #include <nlohmann/detail/iterators/iterator_traits.hpp>
+//     __ _____ _____ _____
+//  __|  |   __|     |   | |  JSON for Modern C++
+// |  |  |__   |  |  | | | |  version 3.11.2
+// |_____|_____|_____|_|___|  https://github.com/nlohmann/json
+//
+// SPDX-FileCopyrightText: 2013-2022 Niels Lohmann <https://nlohmann.me>
+// SPDX-License-Identifier: MIT
+
+
+
+#include <iterator> // random_access_iterator_tag
+
+// #include <nlohmann/detail/abi_macros.hpp>
+
+// #include <nlohmann/detail/meta/void_t.hpp>
+
+// #include <nlohmann/detail/meta/cpp_future.hpp>
+
+
+NLOHMANN_JSON_NAMESPACE_BEGIN
+namespace detail
+{
+
+template<typename It, typename = void>
+struct iterator_types {};
+
+template<typename It>
+struct iterator_types <
+    It,
+    void_t<typename It::difference_type, typename It::value_type, typename It::pointer,
+    typename It::reference, typename It::iterator_category >>
+{
+    using difference_type = typename It::difference_type;
+    using value_type = typename It::value_type;
+    using pointer = typename It::pointer;
+    using reference = typename It::reference;
+    using iterator_category = typename It::iterator_category;
+};
+
+// This is required as some compilers implement std::iterator_traits in a way that
+// doesn't work with SFINAE. See https://github.com/nlohmann/json/issues/1341.
+template<typename T, typename = void>
+struct iterator_traits
+{
+};
+
+template<typename T>
+struct iterator_traits < T, enable_if_t < !std::is_pointer<T>::value >>
+            : iterator_types<T>
+{
+};
+
+template<typename T>
+struct iterator_traits<T*, enable_if_t<std::is_object<T>::value>>
+{
+    using iterator_category = std::random_access_iterator_tag;
+    using value_type = T;
+    using difference_type = ptrdiff_t;
+    using pointer = T*;
+    using reference = T&;
+};
+
+}  // namespace detail
+NLOHMANN_JSON_NAMESPACE_END
+
+// #include <nlohmann/detail/macro_scope.hpp>
+
+// #include <nlohmann/detail/meta/call_std/begin.hpp>
+//     __ _____ _____ _____
+//  __|  |   __|     |   | |  JSON for Modern C++
+// |  |  |__   |  |  | | | |  version 3.11.2
+// |_____|_____|_____|_|___|  https://github.com/nlohmann/json
+//
+// SPDX-FileCopyrightText: 2013-2022 Niels Lohmann <https://nlohmann.me>
+// SPDX-License-Identifier: MIT
+
+
+
+// #include <nlohmann/detail/macro_scope.hpp>
+
+
+NLOHMANN_JSON_NAMESPACE_BEGIN
+
+NLOHMANN_CAN_CALL_STD_FUNC_IMPL(begin);
+
+NLOHMANN_JSON_NAMESPACE_END
+
+// #include <nlohmann/detail/meta/call_std/end.hpp>
+//     __ _____ _____ _____
+//  __|  |   __|     |   | |  JSON for Modern C++
+// |  |  |__   |  |  | | | |  version 3.11.2
+// |_____|_____|_____|_|___|  https://github.com/nlohmann/json
+//
+// SPDX-FileCopyrightText: 2013-2022 Niels Lohmann <https://nlohmann.me>
+// SPDX-License-Identifier: MIT
+
+
+
+// #include <nlohmann/detail/macro_scope.hpp>
+
+
+NLOHMANN_JSON_NAMESPACE_BEGIN
+
+NLOHMANN_CAN_CALL_STD_FUNC_IMPL(end);
+
+NLOHMANN_JSON_NAMESPACE_END
+
+// #include <nlohmann/detail/meta/cpp_future.hpp>
+
+// #include <nlohmann/detail/meta/detected.hpp>
+
+// #include <nlohmann/json_fwd.hpp>
+//     __ _____ _____ _____
+//  __|  |   __|     |   | |  JSON for Modern C++
+// |  |  |__   |  |  | | | |  version 3.11.2
+// |_____|_____|_____|_|___|  https://github.com/nlohmann/json
+//
+// SPDX-FileCopyrightText: 2013-2022 Niels Lohmann <https://nlohmann.me>
+// SPDX-License-Identifier: MIT
+
+#ifndef INCLUDE_NLOHMANN_JSON_FWD_HPP_
+    #define INCLUDE_NLOHMANN_JSON_FWD_HPP_
+
+    #include <cstdint> // int64_t, uint64_t
+    #include <map> // map
+    #include <memory> // allocator
+    #include <string> // string
+    #include <vector> // vector
+
+    // #include <nlohmann/detail/abi_macros.hpp>
+
+
+    /*!
+    @brief namespace for Niels Lohmann
+    @see https://github.com/nlohmann
+    @since version 1.0.0
+    */
+    NLOHMANN_JSON_NAMESPACE_BEGIN
+
+    /*!
+    @brief default JSONSerializer template argument
+
+    This serializer ignores the template arguments and uses ADL
+    ([argument-dependent lookup](https://en.cppreference.com/w/cpp/language/adl))
+    for serialization.
+    */
+    template<typename T = void, typename SFINAE = void>
+    struct adl_serializer;
+
+    /// a class to store JSON values
+    /// @sa https://json.nlohmann.me/api/basic_json/
+    template<template<typename U, typename V, typename... Args> class ObjectType =
+    std::map,
+    template<typename U, typename... Args> class ArrayType = std::vector,
+    class StringType = std::string, class BooleanType = bool,
+    class NumberIntegerType = std::int64_t,
+    class NumberUnsignedType = std::uint64_t,
+    class NumberFloatType = double,
+    template<typename U> class AllocatorType = std::allocator,
+    template<typename T, typename SFINAE = void> class JSONSerializer =
+    adl_serializer,
+    class BinaryType = std::vector<std::uint8_t>>
+    class basic_json;
+
+    /// @brief JSON Pointer defines a string syntax for identifying a specific value within a JSON document
+    /// @sa https://json.nlohmann.me/api/json_pointer/
+    template<typename RefStringType>
+    class json_pointer;
+
+    /*!
+    @brief default specialization
+    @sa https://json.nlohmann.me/api/json/
+    */
+    using json = basic_json<>;
+
+    /// @brief a minimal map-like container that preserves insertion order
+    /// @sa https://json.nlohmann.me/api/ordered_map/
+    template<class Key, class T, class IgnoredLess, class Allocator>
+    struct ordered_map;
+
+    /// @brief specialization that maintains the insertion order of object keys
+    /// @sa https://json.nlohmann.me/api/ordered_json/
+    using ordered_json = basic_json<nlohmann::ordered_map>;
+
+    NLOHMANN_JSON_NAMESPACE_END
+
+#endif  // INCLUDE_NLOHMANN_JSON_FWD_HPP_
+
+
+NLOHMANN_JSON_NAMESPACE_BEGIN
+/*!
+@brief detail namespace with internal helper functions
+
+This namespace collects functions that should not be exposed,
+implementations of some @ref basic_json methods, and meta-programming helpers.
+
+@since version 2.1.0
+*/
+namespace detail
+{
+
+/////////////
+// helpers //
+/////////////
+
+// Note to maintainers:
+//
+// Every trait in this file expects a non CV-qualified type.
+// The only exceptions are in the 'aliases for detected' section
+// (i.e. those of the form: decltype(T::member_function(std::declval<T>())))
+//
+// In this case, T has to be properly CV-qualified to constraint the function arguments
+// (e.g. to_json(BasicJsonType&, const T&))
+
+template<typename> struct is_basic_json : std::false_type {};
+
+NLOHMANN_BASIC_JSON_TPL_DECLARATION
+struct is_basic_json<NLOHMANN_BASIC_JSON_TPL> : std::true_type {};
+
+// used by exceptions create() member functions
+// true_type for pointer to possibly cv-qualified basic_json or std::nullptr_t
+// false_type otherwise
+template<typename BasicJsonContext>
+struct is_basic_json_context :
+    std::integral_constant < bool,
+    is_basic_json<typename std::remove_cv<typename std::remove_pointer<BasicJsonContext>::type>::type>::value
+    || std::is_same<BasicJsonContext, std::nullptr_t>::value >
+{};
+
+//////////////////////
+// json_ref helpers //
+//////////////////////
+
+template<typename>
+class json_ref;
+
+template<typename>
+struct is_json_ref : std::false_type {};
+
+template<typename T>
+struct is_json_ref<json_ref<T>> : std::true_type {};
+
+//////////////////////////
+// aliases for detected //
+//////////////////////////
+
+template<typename T>
+using mapped_type_t = typename T::mapped_type;
+
+template<typename T>
+using key_type_t = typename T::key_type;
+
+template<typename T>
+using value_type_t = typename T::value_type;
+
+template<typename T>
+using difference_type_t = typename T::difference_type;
+
+template<typename T>
+using pointer_t = typename T::pointer;
+
+template<typename T>
+using reference_t = typename T::reference;
+
+template<typename T>
+using iterator_category_t = typename T::iterator_category;
+
+template<typename T, typename... Args>
+using to_json_function = decltype(T::to_json(std::declval<Args>()...));
+
+template<typename T, typename... Args>
+using from_json_function = decltype(T::from_json(std::declval<Args>()...));
+
+template<typename T, typename U>
+using get_template_function = decltype(std::declval<T>().template get<U>());
+
+// trait checking if JSONSerializer<T>::from_json(json const&, udt&) exists
+template<typename BasicJsonType, typename T, typename = void>
+struct has_from_json : std::false_type {};
+
+// trait checking if j.get<T> is valid
+// use this trait instead of std::is_constructible or std::is_convertible,
+// both rely on, or make use of implicit conversions, and thus fail when T
+// has several constructors/operator= (see https://github.com/nlohmann/json/issues/958)
+template <typename BasicJsonType, typename T>
+struct is_getable
+{
+    static constexpr bool value = is_detected<get_template_function, const BasicJsonType&, T>::value;
+};
+
+template<typename BasicJsonType, typename T>
+struct has_from_json < BasicJsonType, T, enable_if_t < !is_basic_json<T>::value >>
+{
+    using serializer = typename BasicJsonType::template json_serializer<T, void>;
+
+    static constexpr bool value =
+        is_detected_exact<void, from_json_function, serializer,
+        const BasicJsonType&, T&>::value;
+};
+
+// This trait checks if JSONSerializer<T>::from_json(json const&) exists
+// this overload is used for non-default-constructible user-defined-types
+template<typename BasicJsonType, typename T, typename = void>
+struct has_non_default_from_json : std::false_type {};
+
+template<typename BasicJsonType, typename T>
+struct has_non_default_from_json < BasicJsonType, T, enable_if_t < !is_basic_json<T>::value >>
+{
+    using serializer = typename BasicJsonType::template json_serializer<T, void>;
+
+    static constexpr bool value =
+        is_detected_exact<T, from_json_function, serializer,
+        const BasicJsonType&>::value;
+};
+
+// This trait checks if BasicJsonType::json_serializer<T>::to_json exists
+// Do not evaluate the trait when T is a basic_json type, to avoid template instantiation infinite recursion.
+template<typename BasicJsonType, typename T, typename = void>
+struct has_to_json : std::false_type {};
+
+template<typename BasicJsonType, typename T>
+struct has_to_json < BasicJsonType, T, enable_if_t < !is_basic_json<T>::value >>
+{
+    using serializer = typename BasicJsonType::template json_serializer<T, void>;
+
+    static constexpr bool value =
+        is_detected_exact<void, to_json_function, serializer, BasicJsonType&,
+        T>::value;
+};
+
+template<typename T>
+using detect_key_compare = typename T::key_compare;
+
+template<typename T>
+struct has_key_compare : std::integral_constant<bool, is_detected<detect_key_compare, T>::value> {};
+
+// obtains the actual object key comparator
+template<typename BasicJsonType>
+struct actual_object_comparator
+{
+    using object_t = typename BasicJsonType::object_t;
+    using object_comparator_t = typename BasicJsonType::default_object_comparator_t;
+    using type = typename std::conditional < has_key_compare<object_t>::value,
+          typename object_t::key_compare, object_comparator_t>::type;
+};
+
+template<typename BasicJsonType>
+using actual_object_comparator_t = typename actual_object_comparator<BasicJsonType>::type;
+
+///////////////////
+// is_ functions //
+///////////////////
+
+// https://en.cppreference.com/w/cpp/types/conjunction
+template<class...> struct conjunction : std::true_type { };
+template<class B> struct conjunction<B> : B { };
+template<class B, class... Bn>
+struct conjunction<B, Bn...>
+: std::conditional<static_cast<bool>(B::value), conjunction<Bn...>, B>::type {};
+
+// https://en.cppreference.com/w/cpp/types/negation
+template<class B> struct negation : std::integral_constant < bool, !B::value > { };
+
+// Reimplementation of is_constructible and is_default_constructible, due to them being broken for
+// std::pair and std::tuple until LWG 2367 fix (see https://cplusplus.github.io/LWG/lwg-defects.html#2367).
+// This causes compile errors in e.g. clang 3.5 or gcc 4.9.
+template <typename T>
+struct is_default_constructible : std::is_default_constructible<T> {};
+
+template <typename T1, typename T2>
+struct is_default_constructible<std::pair<T1, T2>>
+            : conjunction<is_default_constructible<T1>, is_default_constructible<T2>> {};
+
+template <typename T1, typename T2>
+struct is_default_constructible<const std::pair<T1, T2>>
+            : conjunction<is_default_constructible<T1>, is_default_constructible<T2>> {};
+
+template <typename... Ts>
+struct is_default_constructible<std::tuple<Ts...>>
+            : conjunction<is_default_constructible<Ts>...> {};
+
+template <typename... Ts>
+struct is_default_constructible<const std::tuple<Ts...>>
+            : conjunction<is_default_constructible<Ts>...> {};
+
+
+template <typename T, typename... Args>
+struct is_constructible : std::is_constructible<T, Args...> {};
+
+template <typename T1, typename T2>
+struct is_constructible<std::pair<T1, T2>> : is_default_constructible<std::pair<T1, T2>> {};
+
+template <typename T1, typename T2>
+struct is_constructible<const std::pair<T1, T2>> : is_default_constructible<const std::pair<T1, T2>> {};
+
+template <typename... Ts>
+struct is_constructible<std::tuple<Ts...>> : is_default_constructible<std::tuple<Ts...>> {};
+
+template <typename... Ts>
+struct is_constructible<const std::tuple<Ts...>> : is_default_constructible<const std::tuple<Ts...>> {};
+
+
+template<typename T, typename = void>
+struct is_iterator_traits : std::false_type {};
+
+template<typename T>
+struct is_iterator_traits<iterator_traits<T>>
+{
+  private:
+    using traits = iterator_traits<T>;
+
+  public:
+    static constexpr auto value =
+        is_detected<value_type_t, traits>::value &&
+        is_detected<difference_type_t, traits>::value &&
+        is_detected<pointer_t, traits>::value &&
+        is_detected<iterator_category_t, traits>::value &&
+        is_detected<reference_t, traits>::value;
+};
+
+template<typename T>
+struct is_range
+{
+  private:
+    using t_ref = typename std::add_lvalue_reference<T>::type;
+
+    using iterator = detected_t<result_of_begin, t_ref>;
+    using sentinel = detected_t<result_of_end, t_ref>;
+
+    // to be 100% correct, it should use https://en.cppreference.com/w/cpp/iterator/input_or_output_iterator
+    // and https://en.cppreference.com/w/cpp/iterator/sentinel_for
+    // but reimplementing these would be too much work, as a lot of other concepts are used underneath
+    static constexpr auto is_iterator_begin =
+        is_iterator_traits<iterator_traits<iterator>>::value;
+
+  public:
+    static constexpr bool value = !std::is_same<iterator, nonesuch>::value && !std::is_same<sentinel, nonesuch>::value && is_iterator_begin;
+};
+
+template<typename R>
+using iterator_t = enable_if_t<is_range<R>::value, result_of_begin<decltype(std::declval<R&>())>>;
+
+template<typename T>
+using range_value_t = value_type_t<iterator_traits<iterator_t<T>>>;
+
+// The following implementation of is_complete_type is taken from
+// https://blogs.msdn.microsoft.com/vcblog/2015/12/02/partial-support-for-expression-sfinae-in-vs-2015-update-1/
+// and is written by Xiang Fan who agreed to using it in this library.
+
+template<typename T, typename = void>
+struct is_complete_type : std::false_type {};
+
+template<typename T>
+struct is_complete_type<T, decltype(void(sizeof(T)))> : std::true_type {};
+
+template<typename BasicJsonType, typename CompatibleObjectType,
+         typename = void>
+struct is_compatible_object_type_impl : std::false_type {};
+
+template<typename BasicJsonType, typename CompatibleObjectType>
+struct is_compatible_object_type_impl <
+    BasicJsonType, CompatibleObjectType,
+    enable_if_t < is_detected<mapped_type_t, CompatibleObjectType>::value&&
+    is_detected<key_type_t, CompatibleObjectType>::value >>
+{
+    using object_t = typename BasicJsonType::object_t;
+
+    // macOS's is_constructible does not play well with nonesuch...
+    static constexpr bool value =
+        is_constructible<typename object_t::key_type,
+        typename CompatibleObjectType::key_type>::value &&
+        is_constructible<typename object_t::mapped_type,
+        typename CompatibleObjectType::mapped_type>::value;
+};
+
+template<typename BasicJsonType, typename CompatibleObjectType>
+struct is_compatible_object_type
+    : is_compatible_object_type_impl<BasicJsonType, CompatibleObjectType> {};
+
+template<typename BasicJsonType, typename ConstructibleObjectType,
+         typename = void>
+struct is_constructible_object_type_impl : std::false_type {};
+
+template<typename BasicJsonType, typename ConstructibleObjectType>
+struct is_constructible_object_type_impl <
+    BasicJsonType, ConstructibleObjectType,
+    enable_if_t < is_detected<mapped_type_t, ConstructibleObjectType>::value&&
+    is_detected<key_type_t, ConstructibleObjectType>::value >>
+{
+    using object_t = typename BasicJsonType::object_t;
+
+    static constexpr bool value =
+        (is_default_constructible<ConstructibleObjectType>::value &&
+         (std::is_move_assignable<ConstructibleObjectType>::value ||
+          std::is_copy_assignable<ConstructibleObjectType>::value) &&
+         (is_constructible<typename ConstructibleObjectType::key_type,
+          typename object_t::key_type>::value &&
+          std::is_same <
+          typename object_t::mapped_type,
+          typename ConstructibleObjectType::mapped_type >::value)) ||
+        (has_from_json<BasicJsonType,
+         typename ConstructibleObjectType::mapped_type>::value ||
+         has_non_default_from_json <
+         BasicJsonType,
+         typename ConstructibleObjectType::mapped_type >::value);
+};
+
+template<typename BasicJsonType, typename ConstructibleObjectType>
+struct is_constructible_object_type
+    : is_constructible_object_type_impl<BasicJsonType,
+      ConstructibleObjectType> {};
+
+template<typename BasicJsonType, typename CompatibleStringType>
+struct is_compatible_string_type
+{
+    static constexpr auto value =
+        is_constructible<typename BasicJsonType::string_t, CompatibleStringType>::value;
+};
+
+template<typename BasicJsonType, typename ConstructibleStringType>
+struct is_constructible_string_type
+{
+    // launder type through decltype() to fix compilation failure on ICPC
+#ifdef __INTEL_COMPILER
+    using laundered_type = decltype(std::declval<ConstructibleStringType>());
+#else
+    using laundered_type = ConstructibleStringType;
+#endif
+
+    static constexpr auto value =
+        conjunction <
+        is_constructible<laundered_type, typename BasicJsonType::string_t>,
+        is_detected_exact<typename BasicJsonType::string_t::value_type,
+        value_type_t, laundered_type >>::value;
+};
+
+template<typename BasicJsonType, typename CompatibleArrayType, typename = void>
+struct is_compatible_array_type_impl : std::false_type {};
+
+template<typename BasicJsonType, typename CompatibleArrayType>
+struct is_compatible_array_type_impl <
+    BasicJsonType, CompatibleArrayType,
+    enable_if_t <
+    is_detected<iterator_t, CompatibleArrayType>::value&&
+    is_iterator_traits<iterator_traits<detected_t<iterator_t, CompatibleArrayType>>>::value&&
+// special case for types like std::filesystem::path whose iterator's value_type are themselves
+// c.f. https://github.com/nlohmann/json/pull/3073
+    !std::is_same<CompatibleArrayType, detected_t<range_value_t, CompatibleArrayType>>::value >>
+{
+    static constexpr bool value =
+        is_constructible<BasicJsonType,
+        range_value_t<CompatibleArrayType>>::value;
+};
+
+template<typename BasicJsonType, typename CompatibleArrayType>
+struct is_compatible_array_type
+    : is_compatible_array_type_impl<BasicJsonType, CompatibleArrayType> {};
+
+template<typename BasicJsonType, typename ConstructibleArrayType, typename = void>
+struct is_constructible_array_type_impl : std::false_type {};
+
+template<typename BasicJsonType, typename ConstructibleArrayType>
+struct is_constructible_array_type_impl <
+    BasicJsonType, ConstructibleArrayType,
+    enable_if_t<std::is_same<ConstructibleArrayType,
+    typename BasicJsonType::value_type>::value >>
+            : std::true_type {};
+
+template<typename BasicJsonType, typename ConstructibleArrayType>
+struct is_constructible_array_type_impl <
+    BasicJsonType, ConstructibleArrayType,
+    enable_if_t < !std::is_same<ConstructibleArrayType,
+    typename BasicJsonType::value_type>::value&&
+    !is_compatible_string_type<BasicJsonType, ConstructibleArrayType>::value&&
+    is_default_constructible<ConstructibleArrayType>::value&&
+(std::is_move_assignable<ConstructibleArrayType>::value ||
+ std::is_copy_assignable<ConstructibleArrayType>::value)&&
+is_detected<iterator_t, ConstructibleArrayType>::value&&
+is_iterator_traits<iterator_traits<detected_t<iterator_t, ConstructibleArrayType>>>::value&&
+is_detected<range_value_t, ConstructibleArrayType>::value&&
+// special case for types like std::filesystem::path whose iterator's value_type are themselves
+// c.f. https://github.com/nlohmann/json/pull/3073
+!std::is_same<ConstructibleArrayType, detected_t<range_value_t, ConstructibleArrayType>>::value&&
+        is_complete_type <
+        detected_t<range_value_t, ConstructibleArrayType >>::value >>
+{
+    using value_type = range_value_t<ConstructibleArrayType>;
+
+    static constexpr bool value =
+        std::is_same<value_type,
+        typename BasicJsonType::array_t::value_type>::value ||
+        has_from_json<BasicJsonType,
+        value_type>::value ||
+        has_non_default_from_json <
+        BasicJsonType,
+        value_type >::value;
+};
+
+template<typename BasicJsonType, typename ConstructibleArrayType>
+struct is_constructible_array_type
+    : is_constructible_array_type_impl<BasicJsonType, ConstructibleArrayType> {};
+
+template<typename RealIntegerType, typename CompatibleNumberIntegerType,
+         typename = void>
+struct is_compatible_integer_type_impl : std::false_type {};
+
+template<typename RealIntegerType, typename CompatibleNumberIntegerType>
+struct is_compatible_integer_type_impl <
+    RealIntegerType, CompatibleNumberIntegerType,
+    enable_if_t < std::is_integral<RealIntegerType>::value&&
+    std::is_integral<CompatibleNumberIntegerType>::value&&
+    !std::is_same<bool, CompatibleNumberIntegerType>::value >>
+{
+    // is there an assert somewhere on overflows?
+    using RealLimits = std::numeric_limits<RealIntegerType>;
+    using CompatibleLimits = std::numeric_limits<CompatibleNumberIntegerType>;
+
+    static constexpr auto value =
+        is_constructible<RealIntegerType,
+        CompatibleNumberIntegerType>::value &&
+        CompatibleLimits::is_integer &&
+        RealLimits::is_signed == CompatibleLimits::is_signed;
+};
+
+template<typename RealIntegerType, typename CompatibleNumberIntegerType>
+struct is_compatible_integer_type
+    : is_compatible_integer_type_impl<RealIntegerType,
+      CompatibleNumberIntegerType> {};
+
+template<typename BasicJsonType, typename CompatibleType, typename = void>
+struct is_compatible_type_impl: std::false_type {};
+
+template<typename BasicJsonType, typename CompatibleType>
+struct is_compatible_type_impl <
+    BasicJsonType, CompatibleType,
+    enable_if_t<is_complete_type<CompatibleType>::value >>
+{
+    static constexpr bool value =
+        has_to_json<BasicJsonType, CompatibleType>::value;
+};
+
+template<typename BasicJsonType, typename CompatibleType>
+struct is_compatible_type
+    : is_compatible_type_impl<BasicJsonType, CompatibleType> {};
+
+template<typename T1, typename T2>
+struct is_constructible_tuple : std::false_type {};
+
+template<typename T1, typename... Args>
+struct is_constructible_tuple<T1, std::tuple<Args...>> : conjunction<is_constructible<T1, Args>...> {};
+
+template<typename BasicJsonType, typename T>
+struct is_json_iterator_of : std::false_type {};
+
+template<typename BasicJsonType>
+struct is_json_iterator_of<BasicJsonType, typename BasicJsonType::iterator> : std::true_type {};
+
+template<typename BasicJsonType>
+struct is_json_iterator_of<BasicJsonType, typename BasicJsonType::const_iterator> : std::true_type
+{};
+
+// checks if a given type T is a template specialization of Primary
+template<template <typename...> class Primary, typename T>
+struct is_specialization_of : std::false_type {};
+
+template<template <typename...> class Primary, typename... Args>
+struct is_specialization_of<Primary, Primary<Args...>> : std::true_type {};
+
+template<typename T>
+using is_json_pointer = is_specialization_of<::nlohmann::json_pointer, uncvref_t<T>>;
+
+// checks if A and B are comparable using Compare functor
+template<typename Compare, typename A, typename B, typename = void>
+struct is_comparable : std::false_type {};
+
+template<typename Compare, typename A, typename B>
+struct is_comparable<Compare, A, B, void_t<
+decltype(std::declval<Compare>()(std::declval<A>(), std::declval<B>())),
+decltype(std::declval<Compare>()(std::declval<B>(), std::declval<A>()))
+>> : std::true_type {};
+
+template<typename T>
+using detect_is_transparent = typename T::is_transparent;
+
+// type trait to check if KeyType can be used as object key (without a BasicJsonType)
+// see is_usable_as_basic_json_key_type below
+template<typename Comparator, typename ObjectKeyType, typename KeyTypeCVRef, bool RequireTransparentComparator = true,
+         bool ExcludeObjectKeyType = RequireTransparentComparator, typename KeyType = uncvref_t<KeyTypeCVRef>>
+using is_usable_as_key_type = typename std::conditional <
+                              is_comparable<Comparator, ObjectKeyType, KeyTypeCVRef>::value
+                              && !(ExcludeObjectKeyType && std::is_same<KeyType,
+                                   ObjectKeyType>::value)
+                              && (!RequireTransparentComparator
+                                  || is_detected <detect_is_transparent, Comparator>::value)
+                              && !is_json_pointer<KeyType>::value,
+                              std::true_type,
+                              std::false_type >::type;
+
+// type trait to check if KeyType can be used as object key
+// true if:
+//   - KeyType is comparable with BasicJsonType::object_t::key_type
+//   - if ExcludeObjectKeyType is true, KeyType is not BasicJsonType::object_t::key_type
+//   - the comparator is transparent or RequireTransparentComparator is false
+//   - KeyType is not a JSON iterator or json_pointer
+template<typename BasicJsonType, typename KeyTypeCVRef, bool RequireTransparentComparator = true,
+         bool ExcludeObjectKeyType = RequireTransparentComparator, typename KeyType = uncvref_t<KeyTypeCVRef>>
+using is_usable_as_basic_json_key_type = typename std::conditional <
+        is_usable_as_key_type<typename BasicJsonType::object_comparator_t,
+        typename BasicJsonType::object_t::key_type, KeyTypeCVRef,
+        RequireTransparentComparator, ExcludeObjectKeyType>::value
+        && !is_json_iterator_of<BasicJsonType, KeyType>::value,
+        std::true_type,
+        std::false_type >::type;
+
+template<typename ObjectType, typename KeyType>
+using detect_erase_with_key_type = decltype(std::declval<ObjectType&>().erase(std::declval<KeyType>()));
+
+// type trait to check if object_t has an erase() member functions accepting KeyType
+template<typename BasicJsonType, typename KeyType>
+using has_erase_with_key_type = typename std::conditional <
+                                is_detected <
+                                detect_erase_with_key_type,
+                                typename BasicJsonType::object_t, KeyType >::value,
+                                std::true_type,
+                                std::false_type >::type;
+
+// a naive helper to check if a type is an ordered_map (exploits the fact that
+// ordered_map inherits capacity() from std::vector)
+template <typename T>
+struct is_ordered_map
+{
+    using one = char;
+
+    struct two
+    {
+        char x[2]; // NOLINT(cppcoreguidelines-avoid-c-arrays,hicpp-avoid-c-arrays,modernize-avoid-c-arrays)
+    };
+
+    template <typename C> static one test( decltype(&C::capacity) ) ;
+    template <typename C> static two test(...);
+
+    enum { value = sizeof(test<T>(nullptr)) == sizeof(char) }; // NOLINT(cppcoreguidelines-pro-type-vararg,hicpp-vararg)
+};
+
+// to avoid useless casts (see https://github.com/nlohmann/json/issues/2893#issuecomment-889152324)
+template < typename T, typename U, enable_if_t < !std::is_same<T, U>::value, int > = 0 >
+T conditional_static_cast(U value)
+{
+    return static_cast<T>(value);
+}
+
+template<typename T, typename U, enable_if_t<std::is_same<T, U>::value, int> = 0>
+T conditional_static_cast(U value)
+{
+    return value;
+}
+
+template<typename... Types>
+using all_integral = conjunction<std::is_integral<Types>...>;
+
+template<typename... Types>
+using all_signed = conjunction<std::is_signed<Types>...>;
+
+template<typename... Types>
+using all_unsigned = conjunction<std::is_unsigned<Types>...>;
+
+// there's a disjunction trait in another PR; replace when merged
+template<typename... Types>
+using same_sign = std::integral_constant < bool,
+      all_signed<Types...>::value || all_unsigned<Types...>::value >;
+
+template<typename OfType, typename T>
+using never_out_of_range = std::integral_constant < bool,
+      (std::is_signed<OfType>::value && (sizeof(T) < sizeof(OfType)))
+      || (same_sign<OfType, T>::value && sizeof(OfType) == sizeof(T)) >;
+
+template<typename OfType, typename T,
+         bool OfTypeSigned = std::is_signed<OfType>::value,
+         bool TSigned = std::is_signed<T>::value>
+struct value_in_range_of_impl2;
+
+template<typename OfType, typename T>
+struct value_in_range_of_impl2<OfType, T, false, false>
+{
+    static constexpr bool test(T val)
+    {
+        using CommonType = typename std::common_type<OfType, T>::type;
+        return static_cast<CommonType>(val) <= static_cast<CommonType>((std::numeric_limits<OfType>::max)());
+    }
+};
+
+template<typename OfType, typename T>
+struct value_in_range_of_impl2<OfType, T, true, false>
+{
+    static constexpr bool test(T val)
+    {
+        using CommonType = typename std::common_type<OfType, T>::type;
+        return static_cast<CommonType>(val) <= static_cast<CommonType>((std::numeric_limits<OfType>::max)());
+    }
+};
+
+template<typename OfType, typename T>
+struct value_in_range_of_impl2<OfType, T, false, true>
+{
+    static constexpr bool test(T val)
+    {
+        using CommonType = typename std::common_type<OfType, T>::type;
+        return val >= 0 && static_cast<CommonType>(val) <= static_cast<CommonType>((std::numeric_limits<OfType>::max)());
+    }
+};
+
+
+template<typename OfType, typename T>
+struct value_in_range_of_impl2<OfType, T, true, true>
+{
+    static constexpr bool test(T val)
+    {
+        using CommonType = typename std::common_type<OfType, T>::type;
+        return static_cast<CommonType>(val) >= static_cast<CommonType>((std::numeric_limits<OfType>::min)())
+               && static_cast<CommonType>(val) <= static_cast<CommonType>((std::numeric_limits<OfType>::max)());
+    }
+};
+
+template<typename OfType, typename T,
+         bool NeverOutOfRange = never_out_of_range<OfType, T>::value,
+         typename = detail::enable_if_t<all_integral<OfType, T>::value>>
+struct value_in_range_of_impl1;
+
+template<typename OfType, typename T>
+struct value_in_range_of_impl1<OfType, T, false>
+{
+    static constexpr bool test(T val)
+    {
+        return value_in_range_of_impl2<OfType, T>::test(val);
+    }
+};
+
+template<typename OfType, typename T>
+struct value_in_range_of_impl1<OfType, T, true>
+{
+    static constexpr bool test(T /*val*/)
+    {
+        return true;
+    }
+};
+
+template<typename OfType, typename T>
+inline constexpr bool value_in_range_of(T val)
+{
+    return value_in_range_of_impl1<OfType, T>::test(val);
+}
+
+template<bool Value>
+using bool_constant = std::integral_constant<bool, Value>;
+
+///////////////////////////////////////////////////////////////////////////////
+// is_c_string
+///////////////////////////////////////////////////////////////////////////////
+
+namespace impl
+{
+
+template<typename T>
+inline constexpr bool is_c_string()
+{
+    using TUnExt = typename std::remove_extent<T>::type;
+    using TUnCVExt = typename std::remove_cv<TUnExt>::type;
+    using TUnPtr = typename std::remove_pointer<T>::type;
+    using TUnCVPtr = typename std::remove_cv<TUnPtr>::type;
+    return
+        (std::is_array<T>::value && std::is_same<TUnCVExt, char>::value)
+        || (std::is_pointer<T>::value && std::is_same<TUnCVPtr, char>::value);
+}
+
+}  // namespace impl
+
+// checks whether T is a [cv] char */[cv] char[] C string
+template<typename T>
+struct is_c_string : bool_constant<impl::is_c_string<T>()> {};
+
+template<typename T>
+using is_c_string_uncvref = is_c_string<uncvref_t<T>>;
+
+///////////////////////////////////////////////////////////////////////////////
+// is_transparent
+///////////////////////////////////////////////////////////////////////////////
+
+namespace impl
+{
+
+template<typename T>
+inline constexpr bool is_transparent()
+{
+    return is_detected<detect_is_transparent, T>::value;
+}
+
+}  // namespace impl
+
+// checks whether T has a member named is_transparent
+template<typename T>
+struct is_transparent : bool_constant<impl::is_transparent<T>()> {};
+
+///////////////////////////////////////////////////////////////////////////////
+
+}  // namespace detail
+NLOHMANN_JSON_NAMESPACE_END
+
+// #include <nlohmann/detail/string_concat.hpp>
+//     __ _____ _____ _____
+//  __|  |   __|     |   | |  JSON for Modern C++
+// |  |  |__   |  |  | | | |  version 3.11.2
+// |_____|_____|_____|_|___|  https://github.com/nlohmann/json
+//
+// SPDX-FileCopyrightText: 2013-2022 Niels Lohmann <https://nlohmann.me>
+// SPDX-License-Identifier: MIT
+
+
+
+#include <cstring> // strlen
+#include <string> // string
+#include <utility> // forward
+
+// #include <nlohmann/detail/meta/cpp_future.hpp>
+
+// #include <nlohmann/detail/meta/detected.hpp>
+
+
+NLOHMANN_JSON_NAMESPACE_BEGIN
+namespace detail
+{
+
+inline std::size_t concat_length()
+{
+    return 0;
+}
+
+template<typename... Args>
+inline std::size_t concat_length(const char* cstr, Args&& ... rest);
+
+template<typename StringType, typename... Args>
+inline std::size_t concat_length(const StringType& str, Args&& ... rest);
+
+template<typename... Args>
+inline std::size_t concat_length(const char /*c*/, Args&& ... rest)
+{
+    return 1 + concat_length(std::forward<Args>(rest)...);
+}
+
+template<typename... Args>
+inline std::size_t concat_length(const char* cstr, Args&& ... rest)
+{
+    // cppcheck-suppress ignoredReturnValue
+    return ::strlen(cstr) + concat_length(std::forward<Args>(rest)...);
+}
+
+template<typename StringType, typename... Args>
+inline std::size_t concat_length(const StringType& str, Args&& ... rest)
+{
+    return str.size() + concat_length(std::forward<Args>(rest)...);
+}
+
+template<typename OutStringType>
+inline void concat_into(OutStringType& /*out*/)
+{}
+
+template<typename StringType, typename Arg>
+using string_can_append = decltype(std::declval<StringType&>().append(std::declval < Arg && > ()));
+
+template<typename StringType, typename Arg>
+using detect_string_can_append = is_detected<string_can_append, StringType, Arg>;
+
+template<typename StringType, typename Arg>
+using string_can_append_op = decltype(std::declval<StringType&>() += std::declval < Arg && > ());
+
+template<typename StringType, typename Arg>
+using detect_string_can_append_op = is_detected<string_can_append_op, StringType, Arg>;
+
+template<typename StringType, typename Arg>
+using string_can_append_iter = decltype(std::declval<StringType&>().append(std::declval<const Arg&>().begin(), std::declval<const Arg&>().end()));
+
+template<typename StringType, typename Arg>
+using detect_string_can_append_iter = is_detected<string_can_append_iter, StringType, Arg>;
+
+template<typename StringType, typename Arg>
+using string_can_append_data = decltype(std::declval<StringType&>().append(std::declval<const Arg&>().data(), std::declval<const Arg&>().size()));
+
+template<typename StringType, typename Arg>
+using detect_string_can_append_data = is_detected<string_can_append_data, StringType, Arg>;
+
+template < typename OutStringType, typename Arg, typename... Args,
+           enable_if_t < !detect_string_can_append<OutStringType, Arg>::value
+                         && detect_string_can_append_op<OutStringType, Arg>::value, int > = 0 >
+inline void concat_into(OutStringType& out, Arg && arg, Args && ... rest);
+
+template < typename OutStringType, typename Arg, typename... Args,
+           enable_if_t < !detect_string_can_append<OutStringType, Arg>::value
+                         && !detect_string_can_append_op<OutStringType, Arg>::value
+                         && detect_string_can_append_iter<OutStringType, Arg>::value, int > = 0 >
+inline void concat_into(OutStringType& out, const Arg& arg, Args && ... rest);
+
+template < typename OutStringType, typename Arg, typename... Args,
+           enable_if_t < !detect_string_can_append<OutStringType, Arg>::value
+                         && !detect_string_can_append_op<OutStringType, Arg>::value
+                         && !detect_string_can_append_iter<OutStringType, Arg>::value
+                         && detect_string_can_append_data<OutStringType, Arg>::value, int > = 0 >
+inline void concat_into(OutStringType& out, const Arg& arg, Args && ... rest);
+
+template<typename OutStringType, typename Arg, typename... Args,
+         enable_if_t<detect_string_can_append<OutStringType, Arg>::value, int> = 0>
+inline void concat_into(OutStringType& out, Arg && arg, Args && ... rest)
+{
+    out.append(std::forward<Arg>(arg));
+    concat_into(out, std::forward<Args>(rest)...);
+}
+
+template < typename OutStringType, typename Arg, typename... Args,
+           enable_if_t < !detect_string_can_append<OutStringType, Arg>::value
+                         && detect_string_can_append_op<OutStringType, Arg>::value, int > >
+inline void concat_into(OutStringType& out, Arg&& arg, Args&& ... rest)
+{
+    out += std::forward<Arg>(arg);
+    concat_into(out, std::forward<Args>(rest)...);
+}
+
+template < typename OutStringType, typename Arg, typename... Args,
+           enable_if_t < !detect_string_can_append<OutStringType, Arg>::value
+                         && !detect_string_can_append_op<OutStringType, Arg>::value
+                         && detect_string_can_append_iter<OutStringType, Arg>::value, int > >
+inline void concat_into(OutStringType& out, const Arg& arg, Args&& ... rest)
+{
+    out.append(arg.begin(), arg.end());
+    concat_into(out, std::forward<Args>(rest)...);
+}
+
+template < typename OutStringType, typename Arg, typename... Args,
+           enable_if_t < !detect_string_can_append<OutStringType, Arg>::value
+                         && !detect_string_can_append_op<OutStringType, Arg>::value
+                         && !detect_string_can_append_iter<OutStringType, Arg>::value
+                         && detect_string_can_append_data<OutStringType, Arg>::value, int > >
+inline void concat_into(OutStringType& out, const Arg& arg, Args&& ... rest)
+{
+    out.append(arg.data(), arg.size());
+    concat_into(out, std::forward<Args>(rest)...);
+}
+
+template<typename OutStringType = std::string, typename... Args>
+inline OutStringType concat(Args && ... args)
+{
+    OutStringType str;
+    str.reserve(concat_length(std::forward<Args>(args)...));
+    concat_into(str, std::forward<Args>(args)...);
+    return str;
+}
+
+}  // namespace detail
+NLOHMANN_JSON_NAMESPACE_END
+
+
+
+NLOHMANN_JSON_NAMESPACE_BEGIN
+namespace detail
+{
+
+////////////////
+// exceptions //
+////////////////
+
+/// @brief general exception of the @ref basic_json class
+/// @sa https://json.nlohmann.me/api/basic_json/exception/
+class exception : public std::exception
+{
+  public:
+    /// returns the explanatory string
+    const char* what() const noexcept override
+    {
+        return m.what();
+    }
+
+    /// the id of the exception
+    const int id; // NOLINT(cppcoreguidelines-non-private-member-variables-in-classes)
+
+  protected:
+    JSON_HEDLEY_NON_NULL(3)
+    exception(int id_, const char* what_arg) : id(id_), m(what_arg) {} // NOLINT(bugprone-throw-keyword-missing)
+
+    static std::string name(const std::string& ename, int id_)
+    {
+        return concat("[json.exception.", ename, '.', std::to_string(id_), "] ");
+    }
+
+    static std::string diagnostics(std::nullptr_t /*leaf_element*/)
+    {
+        return "";
+    }
+
+    template<typename BasicJsonType>
+    static std::string diagnostics(const BasicJsonType* leaf_element)
+    {
+#if JSON_DIAGNOSTICS
+        std::vector<std::string> tokens;
+        for (const auto* current = leaf_element; current != nullptr && current->m_parent != nullptr; current = current->m_parent)
+        {
+            switch (current->m_parent->type())
+            {
+                case value_t::array:
+                {
+                    for (std::size_t i = 0; i < current->m_parent->m_value.array->size(); ++i)
+                    {
+                        if (&current->m_parent->m_value.array->operator[](i) == current)
+                        {
+                            tokens.emplace_back(std::to_string(i));
+                            break;
+                        }
+                    }
+                    break;
+                }
+
+                case value_t::object:
+                {
+                    for (const auto& element : *current->m_parent->m_value.object)
+                    {
+                        if (&element.second == current)
+                        {
+                            tokens.emplace_back(element.first.c_str());
+                            break;
+                        }
+                    }
+                    break;
+                }
+
+                case value_t::null: // LCOV_EXCL_LINE
+                case value_t::string: // LCOV_EXCL_LINE
+                case value_t::boolean: // LCOV_EXCL_LINE
+                case value_t::number_integer: // LCOV_EXCL_LINE
+                case value_t::number_unsigned: // LCOV_EXCL_LINE
+                case value_t::number_float: // LCOV_EXCL_LINE
+                case value_t::binary: // LCOV_EXCL_LINE
+                case value_t::discarded: // LCOV_EXCL_LINE
+                default:   // LCOV_EXCL_LINE
+                    break; // LCOV_EXCL_LINE
+            }
+        }
+
+        if (tokens.empty())
+        {
+            return "";
+        }
+
+        auto str = std::accumulate(tokens.rbegin(), tokens.rend(), std::string{},
+                                   [](const std::string & a, const std::string & b)
+        {
+            return concat(a, '/', detail::escape(b));
+        });
+        return concat('(', str, ") ");
+#else
+        static_cast<void>(leaf_element);
+        return "";
+#endif
+    }
+
+  private:
+    /// an exception object as storage for error messages
+    std::runtime_error m;
+};
+
+/// @brief exception indicating a parse error
+/// @sa https://json.nlohmann.me/api/basic_json/parse_error/
+class parse_error : public exception
+{
+  public:
+    /*!
+    @brief create a parse error exception
+    @param[in] id_       the id of the exception
+    @param[in] pos       the position where the error occurred (or with
+                         chars_read_total=0 if the position cannot be
+                         determined)
+    @param[in] what_arg  the explanatory string
+    @return parse_error object
+    */
+    template<typename BasicJsonContext, enable_if_t<is_basic_json_context<BasicJsonContext>::value, int> = 0>
+    static parse_error create(int id_, const position_t& pos, const std::string& what_arg, BasicJsonContext context)
+    {
+        std::string w = concat(exception::name("parse_error", id_), "parse error",
+                               position_string(pos), ": ", exception::diagnostics(context), what_arg);
+        return {id_, pos.chars_read_total, w.c_str()};
+    }
+
+    template<typename BasicJsonContext, enable_if_t<is_basic_json_context<BasicJsonContext>::value, int> = 0>
+    static parse_error create(int id_, std::size_t byte_, const std::string& what_arg, BasicJsonContext context)
+    {
+        std::string w = concat(exception::name("parse_error", id_), "parse error",
+                               (byte_ != 0 ? (concat(" at byte ", std::to_string(byte_))) : ""),
+                               ": ", exception::diagnostics(context), what_arg);
+        return {id_, byte_, w.c_str()};
+    }
+
+    /*!
+    @brief byte index of the parse error
+
+    The byte index of the last read character in the input file.
+
+    @note For an input with n bytes, 1 is the index of the first character and
+          n+1 is the index of the terminating null byte or the end of file.
+          This also holds true when reading a byte vector (CBOR or MessagePack).
+    */
+    const std::size_t byte;
+
+  private:
+    parse_error(int id_, std::size_t byte_, const char* what_arg)
+        : exception(id_, what_arg), byte(byte_) {}
+
+    static std::string position_string(const position_t& pos)
+    {
+        return concat(" at line ", std::to_string(pos.lines_read + 1),
+                      ", column ", std::to_string(pos.chars_read_current_line));
+    }
+};
+
+/// @brief exception indicating errors with iterators
+/// @sa https://json.nlohmann.me/api/basic_json/invalid_iterator/
+class invalid_iterator : public exception
+{
+  public:
+    template<typename BasicJsonContext, enable_if_t<is_basic_json_context<BasicJsonContext>::value, int> = 0>
+    static invalid_iterator create(int id_, const std::string& what_arg, BasicJsonContext context)
+    {
+        std::string w = concat(exception::name("invalid_iterator", id_), exception::diagnostics(context), what_arg);
+        return {id_, w.c_str()};
+    }
+
+  private:
+    JSON_HEDLEY_NON_NULL(3)
+    invalid_iterator(int id_, const char* what_arg)
+        : exception(id_, what_arg) {}
+};
+
+/// @brief exception indicating executing a member function with a wrong type
+/// @sa https://json.nlohmann.me/api/basic_json/type_error/
+class type_error : public exception
+{
+  public:
+    template<typename BasicJsonContext, enable_if_t<is_basic_json_context<BasicJsonContext>::value, int> = 0>
+    static type_error create(int id_, const std::string& what_arg, BasicJsonContext context)
+    {
+        std::string w = concat(exception::name("type_error", id_), exception::diagnostics(context), what_arg);
+        return {id_, w.c_str()};
+    }
+
+  private:
+    JSON_HEDLEY_NON_NULL(3)
+    type_error(int id_, const char* what_arg) : exception(id_, what_arg) {}
+};
+
+/// @brief exception indicating access out of the defined range
+/// @sa https://json.nlohmann.me/api/basic_json/out_of_range/
+class out_of_range : public exception
+{
+  public:
+    template<typename BasicJsonContext, enable_if_t<is_basic_json_context<BasicJsonContext>::value, int> = 0>
+    static out_of_range create(int id_, const std::string& what_arg, BasicJsonContext context)
+    {
+        std::string w = concat(exception::name("out_of_range", id_), exception::diagnostics(context), what_arg);
+        return {id_, w.c_str()};
+    }
+
+  private:
+    JSON_HEDLEY_NON_NULL(3)
+    out_of_range(int id_, const char* what_arg) : exception(id_, what_arg) {}
+};
+
+/// @brief exception indicating other library errors
+/// @sa https://json.nlohmann.me/api/basic_json/other_error/
+class other_error : public exception
+{
+  public:
+    template<typename BasicJsonContext, enable_if_t<is_basic_json_context<BasicJsonContext>::value, int> = 0>
+    static other_error create(int id_, const std::string& what_arg, BasicJsonContext context)
+    {
+        std::string w = concat(exception::name("other_error", id_), exception::diagnostics(context), what_arg);
+        return {id_, w.c_str()};
+    }
+
+  private:
+    JSON_HEDLEY_NON_NULL(3)
+    other_error(int id_, const char* what_arg) : exception(id_, what_arg) {}
+};
+
+}  // namespace detail
+NLOHMANN_JSON_NAMESPACE_END
+
+// #include <nlohmann/detail/macro_scope.hpp>
+
+// #include <nlohmann/detail/meta/cpp_future.hpp>
+
+// #include <nlohmann/detail/meta/identity_tag.hpp>
+//     __ _____ _____ _____
+//  __|  |   __|     |   | |  JSON for Modern C++
+// |  |  |__   |  |  | | | |  version 3.11.2
+// |_____|_____|_____|_|___|  https://github.com/nlohmann/json
+//
+// SPDX-FileCopyrightText: 2013-2022 Niels Lohmann <https://nlohmann.me>
+// SPDX-License-Identifier: MIT
+
+
+
+// #include <nlohmann/detail/abi_macros.hpp>
+
+
+NLOHMANN_JSON_NAMESPACE_BEGIN
+namespace detail
+{
+
+// dispatching helper struct
+template <class T> struct identity_tag {};
+
+}  // namespace detail
+NLOHMANN_JSON_NAMESPACE_END
+
+// #include <nlohmann/detail/meta/std_fs.hpp>
+//     __ _____ _____ _____
+//  __|  |   __|     |   | |  JSON for Modern C++
+// |  |  |__   |  |  | | | |  version 3.11.2
+// |_____|_____|_____|_|___|  https://github.com/nlohmann/json
+//
+// SPDX-FileCopyrightText: 2013-2022 Niels Lohmann <https://nlohmann.me>
+// SPDX-License-Identifier: MIT
+
+
+
+// #include <nlohmann/detail/macro_scope.hpp>
+
+
+#if JSON_HAS_EXPERIMENTAL_FILESYSTEM
+#include <experimental/filesystem>
+NLOHMANN_JSON_NAMESPACE_BEGIN
+namespace detail
+{
+namespace std_fs = std::experimental::filesystem;
+}  // namespace detail
+NLOHMANN_JSON_NAMESPACE_END
+#elif JSON_HAS_FILESYSTEM
+#include <filesystem>
+NLOHMANN_JSON_NAMESPACE_BEGIN
+namespace detail
+{
+namespace std_fs = std::filesystem;
+}  // namespace detail
+NLOHMANN_JSON_NAMESPACE_END
+#endif
+
+// #include <nlohmann/detail/meta/type_traits.hpp>
+
+// #include <nlohmann/detail/string_concat.hpp>
+
+// #include <nlohmann/detail/value_t.hpp>
+
+
+NLOHMANN_JSON_NAMESPACE_BEGIN
+namespace detail
+{
+
+template<typename BasicJsonType>
+inline void from_json(const BasicJsonType& j, typename std::nullptr_t& n)
+{
+    if (JSON_HEDLEY_UNLIKELY(!j.is_null()))
+    {
+        JSON_THROW(type_error::create(302, concat("type must be null, but is ", j.type_name()), &j));
+    }
+    n = nullptr;
+}
+
+// overloads for basic_json template parameters
+template < typename BasicJsonType, typename ArithmeticType,
+           enable_if_t < std::is_arithmetic<ArithmeticType>::value&&
+                         !std::is_same<ArithmeticType, typename BasicJsonType::boolean_t>::value,
+                         int > = 0 >
+void get_arithmetic_value(const BasicJsonType& j, ArithmeticType& val)
+{
+    switch (static_cast<value_t>(j))
+    {
+        case value_t::number_unsigned:
+        {
+            val = static_cast<ArithmeticType>(*j.template get_ptr<const typename BasicJsonType::number_unsigned_t*>());
+            break;
+        }
+        case value_t::number_integer:
+        {
+            val = static_cast<ArithmeticType>(*j.template get_ptr<const typename BasicJsonType::number_integer_t*>());
+            break;
+        }
+        case value_t::number_float:
+        {
+            val = static_cast<ArithmeticType>(*j.template get_ptr<const typename BasicJsonType::number_float_t*>());
+            break;
+        }
+
+        case value_t::null:
+        case value_t::object:
+        case value_t::array:
+        case value_t::string:
+        case value_t::boolean:
+        case value_t::binary:
+        case value_t::discarded:
+        default:
+            JSON_THROW(type_error::create(302, concat("type must be number, but is ", j.type_name()), &j));
+    }
+}
+
+template<typename BasicJsonType>
+inline void from_json(const BasicJsonType& j, typename BasicJsonType::boolean_t& b)
+{
+    if (JSON_HEDLEY_UNLIKELY(!j.is_boolean()))
+    {
+        JSON_THROW(type_error::create(302, concat("type must be boolean, but is ", j.type_name()), &j));
+    }
+    b = *j.template get_ptr<const typename BasicJsonType::boolean_t*>();
+}
+
+template<typename BasicJsonType>
+inline void from_json(const BasicJsonType& j, typename BasicJsonType::string_t& s)
+{
+    if (JSON_HEDLEY_UNLIKELY(!j.is_string()))
+    {
+        JSON_THROW(type_error::create(302, concat("type must be string, but is ", j.type_name()), &j));
+    }
+    s = *j.template get_ptr<const typename BasicJsonType::string_t*>();
+}
+
+template <
+    typename BasicJsonType, typename StringType,
+    enable_if_t <
+        std::is_assignable<StringType&, const typename BasicJsonType::string_t>::value
+        && is_detected_exact<typename BasicJsonType::string_t::value_type, value_type_t, StringType>::value
+        && !std::is_same<typename BasicJsonType::string_t, StringType>::value
+        && !is_json_ref<StringType>::value, int > = 0 >
+inline void from_json(const BasicJsonType& j, StringType& s)
+{
+    if (JSON_HEDLEY_UNLIKELY(!j.is_string()))
+    {
+        JSON_THROW(type_error::create(302, concat("type must be string, but is ", j.type_name()), &j));
+    }
+
+    s = *j.template get_ptr<const typename BasicJsonType::string_t*>();
+}
+
+template<typename BasicJsonType>
+inline void from_json(const BasicJsonType& j, typename BasicJsonType::number_float_t& val)
+{
+    get_arithmetic_value(j, val);
+}
+
+template<typename BasicJsonType>
+inline void from_json(const BasicJsonType& j, typename BasicJsonType::number_unsigned_t& val)
+{
+    get_arithmetic_value(j, val);
+}
+
+template<typename BasicJsonType>
+inline void from_json(const BasicJsonType& j, typename BasicJsonType::number_integer_t& val)
+{
+    get_arithmetic_value(j, val);
+}
+
+#if !JSON_DISABLE_ENUM_SERIALIZATION
+template<typename BasicJsonType, typename EnumType,
+         enable_if_t<std::is_enum<EnumType>::value, int> = 0>
+inline void from_json(const BasicJsonType& j, EnumType& e)
+{
+    typename std::underlying_type<EnumType>::type val;
+    get_arithmetic_value(j, val);
+    e = static_cast<EnumType>(val);
+}
+#endif  // JSON_DISABLE_ENUM_SERIALIZATION
+
+// forward_list doesn't have an insert method
+template<typename BasicJsonType, typename T, typename Allocator,
+         enable_if_t<is_getable<BasicJsonType, T>::value, int> = 0>
+inline void from_json(const BasicJsonType& j, std::forward_list<T, Allocator>& l)
+{
+    if (JSON_HEDLEY_UNLIKELY(!j.is_array()))
+    {
+        JSON_THROW(type_error::create(302, concat("type must be array, but is ", j.type_name()), &j));
+    }
+    l.clear();
+    std::transform(j.rbegin(), j.rend(),
+                   std::front_inserter(l), [](const BasicJsonType & i)
+    {
+        return i.template get<T>();
+    });
+}
+
+// valarray doesn't have an insert method
+template<typename BasicJsonType, typename T,
+         enable_if_t<is_getable<BasicJsonType, T>::value, int> = 0>
+inline void from_json(const BasicJsonType& j, std::valarray<T>& l)
+{
+    if (JSON_HEDLEY_UNLIKELY(!j.is_array()))
+    {
+        JSON_THROW(type_error::create(302, concat("type must be array, but is ", j.type_name()), &j));
+    }
+    l.resize(j.size());
+    std::transform(j.begin(), j.end(), std::begin(l),
+                   [](const BasicJsonType & elem)
+    {
+        return elem.template get<T>();
+    });
+}
+
+template<typename BasicJsonType, typename T, std::size_t N>
+auto from_json(const BasicJsonType& j, T (&arr)[N])  // NOLINT(cppcoreguidelines-avoid-c-arrays,hicpp-avoid-c-arrays,modernize-avoid-c-arrays)
+-> decltype(j.template get<T>(), void())
+{
+    for (std::size_t i = 0; i < N; ++i)
+    {
+        arr[i] = j.at(i).template get<T>();
+    }
+}
+
+template<typename BasicJsonType>
+inline void from_json_array_impl(const BasicJsonType& j, typename BasicJsonType::array_t& arr, priority_tag<3> /*unused*/)
+{
+    arr = *j.template get_ptr<const typename BasicJsonType::array_t*>();
+}
+
+template<typename BasicJsonType, typename T, std::size_t N>
+auto from_json_array_impl(const BasicJsonType& j, std::array<T, N>& arr,
+                          priority_tag<2> /*unused*/)
+-> decltype(j.template get<T>(), void())
+{
+    for (std::size_t i = 0; i < N; ++i)
+    {
+        arr[i] = j.at(i).template get<T>();
+    }
+}
+
+template<typename BasicJsonType, typename ConstructibleArrayType,
+         enable_if_t<
+             std::is_assignable<ConstructibleArrayType&, ConstructibleArrayType>::value,
+             int> = 0>
+auto from_json_array_impl(const BasicJsonType& j, ConstructibleArrayType& arr, priority_tag<1> /*unused*/)
+-> decltype(
+    arr.reserve(std::declval<typename ConstructibleArrayType::size_type>()),
+    j.template get<typename ConstructibleArrayType::value_type>(),
+    void())
+{
+    using std::end;
+
+    ConstructibleArrayType ret;
+    ret.reserve(j.size());
+    std::transform(j.begin(), j.end(),
+                   std::inserter(ret, end(ret)), [](const BasicJsonType & i)
+    {
+        // get<BasicJsonType>() returns *this, this won't call a from_json
+        // method when value_type is BasicJsonType
+        return i.template get<typename ConstructibleArrayType::value_type>();
+    });
+    arr = std::move(ret);
+}
+
+template<typename BasicJsonType, typename ConstructibleArrayType,
+         enable_if_t<
+             std::is_assignable<ConstructibleArrayType&, ConstructibleArrayType>::value,
+             int> = 0>
+inline void from_json_array_impl(const BasicJsonType& j, ConstructibleArrayType& arr,
+                                 priority_tag<0> /*unused*/)
+{
+    using std::end;
+
+    ConstructibleArrayType ret;
+    std::transform(
+        j.begin(), j.end(), std::inserter(ret, end(ret)),
+        [](const BasicJsonType & i)
+    {
+        // get<BasicJsonType>() returns *this, this won't call a from_json
+        // method when value_type is BasicJsonType
+        return i.template get<typename ConstructibleArrayType::value_type>();
+    });
+    arr = std::move(ret);
+}
+
+template < typename BasicJsonType, typename ConstructibleArrayType,
+           enable_if_t <
+               is_constructible_array_type<BasicJsonType, ConstructibleArrayType>::value&&
+               !is_constructible_object_type<BasicJsonType, ConstructibleArrayType>::value&&
+               !is_constructible_string_type<BasicJsonType, ConstructibleArrayType>::value&&
+               !std::is_same<ConstructibleArrayType, typename BasicJsonType::binary_t>::value&&
+               !is_basic_json<ConstructibleArrayType>::value,
+               int > = 0 >
+auto from_json(const BasicJsonType& j, ConstructibleArrayType& arr)
+-> decltype(from_json_array_impl(j, arr, priority_tag<3> {}),
+j.template get<typename ConstructibleArrayType::value_type>(),
+void())
+{
+    if (JSON_HEDLEY_UNLIKELY(!j.is_array()))
+    {
+        JSON_THROW(type_error::create(302, concat("type must be array, but is ", j.type_name()), &j));
+    }
+
+    from_json_array_impl(j, arr, priority_tag<3> {});
+}
+
+template < typename BasicJsonType, typename T, std::size_t... Idx >
+std::array<T, sizeof...(Idx)> from_json_inplace_array_impl(BasicJsonType&& j,
+        identity_tag<std::array<T, sizeof...(Idx)>> /*unused*/, index_sequence<Idx...> /*unused*/)
+{
+    return { { std::forward<BasicJsonType>(j).at(Idx).template get<T>()... } };
+}
+
+template < typename BasicJsonType, typename T, std::size_t N >
+auto from_json(BasicJsonType&& j, identity_tag<std::array<T, N>> tag)
+-> decltype(from_json_inplace_array_impl(std::forward<BasicJsonType>(j), tag, make_index_sequence<N> {}))
+{
+    if (JSON_HEDLEY_UNLIKELY(!j.is_array()))
+    {
+        JSON_THROW(type_error::create(302, concat("type must be array, but is ", j.type_name()), &j));
+    }
+
+    return from_json_inplace_array_impl(std::forward<BasicJsonType>(j), tag, make_index_sequence<N> {});
+}
+
+template<typename BasicJsonType>
+inline void from_json(const BasicJsonType& j, typename BasicJsonType::binary_t& bin)
+{
+    if (JSON_HEDLEY_UNLIKELY(!j.is_binary()))
+    {
+        JSON_THROW(type_error::create(302, concat("type must be binary, but is ", j.type_name()), &j));
+    }
+
+    bin = *j.template get_ptr<const typename BasicJsonType::binary_t*>();
+}
+
+template<typename BasicJsonType, typename ConstructibleObjectType,
+         enable_if_t<is_constructible_object_type<BasicJsonType, ConstructibleObjectType>::value, int> = 0>
+inline void from_json(const BasicJsonType& j, ConstructibleObjectType& obj)
+{
+    if (JSON_HEDLEY_UNLIKELY(!j.is_object()))
+    {
+        JSON_THROW(type_error::create(302, concat("type must be object, but is ", j.type_name()), &j));
+    }
+
+    ConstructibleObjectType ret;
+    const auto* inner_object = j.template get_ptr<const typename BasicJsonType::object_t*>();
+    using value_type = typename ConstructibleObjectType::value_type;
+    std::transform(
+        inner_object->begin(), inner_object->end(),
+        std::inserter(ret, ret.begin()),
+        [](typename BasicJsonType::object_t::value_type const & p)
+    {
+        return value_type(p.first, p.second.template get<typename ConstructibleObjectType::mapped_type>());
+    });
+    obj = std::move(ret);
+}
+
+// overload for arithmetic types, not chosen for basic_json template arguments
+// (BooleanType, etc..); note: Is it really necessary to provide explicit
+// overloads for boolean_t etc. in case of a custom BooleanType which is not
+// an arithmetic type?
+template < typename BasicJsonType, typename ArithmeticType,
+           enable_if_t <
+               std::is_arithmetic<ArithmeticType>::value&&
+               !std::is_same<ArithmeticType, typename BasicJsonType::number_unsigned_t>::value&&
+               !std::is_same<ArithmeticType, typename BasicJsonType::number_integer_t>::value&&
+               !std::is_same<ArithmeticType, typename BasicJsonType::number_float_t>::value&&
+               !std::is_same<ArithmeticType, typename BasicJsonType::boolean_t>::value,
+               int > = 0 >
+inline void from_json(const BasicJsonType& j, ArithmeticType& val)
+{
+    switch (static_cast<value_t>(j))
+    {
+        case value_t::number_unsigned:
+        {
+            val = static_cast<ArithmeticType>(*j.template get_ptr<const typename BasicJsonType::number_unsigned_t*>());
+            break;
+        }
+        case value_t::number_integer:
+        {
+            val = static_cast<ArithmeticType>(*j.template get_ptr<const typename BasicJsonType::number_integer_t*>());
+            break;
+        }
+        case value_t::number_float:
+        {
+            val = static_cast<ArithmeticType>(*j.template get_ptr<const typename BasicJsonType::number_float_t*>());
+            break;
+        }
+        case value_t::boolean:
+        {
+            val = static_cast<ArithmeticType>(*j.template get_ptr<const typename BasicJsonType::boolean_t*>());
+            break;
+        }
+
+        case value_t::null:
+        case value_t::object:
+        case value_t::array:
+        case value_t::string:
+        case value_t::binary:
+        case value_t::discarded:
+        default:
+            JSON_THROW(type_error::create(302, concat("type must be number, but is ", j.type_name()), &j));
+    }
+}
+
+template<typename BasicJsonType, typename... Args, std::size_t... Idx>
+std::tuple<Args...> from_json_tuple_impl_base(BasicJsonType&& j, index_sequence<Idx...> /*unused*/)
+{
+    return std::make_tuple(std::forward<BasicJsonType>(j).at(Idx).template get<Args>()...);
+}
+
+template < typename BasicJsonType, class A1, class A2 >
+std::pair<A1, A2> from_json_tuple_impl(BasicJsonType&& j, identity_tag<std::pair<A1, A2>> /*unused*/, priority_tag<0> /*unused*/)
+{
+    return {std::forward<BasicJsonType>(j).at(0).template get<A1>(),
+            std::forward<BasicJsonType>(j).at(1).template get<A2>()};
+}
+
+template<typename BasicJsonType, typename A1, typename A2>
+inline void from_json_tuple_impl(BasicJsonType&& j, std::pair<A1, A2>& p, priority_tag<1> /*unused*/)
+{
+    p = from_json_tuple_impl(std::forward<BasicJsonType>(j), identity_tag<std::pair<A1, A2>> {}, priority_tag<0> {});
+}
+
+template<typename BasicJsonType, typename... Args>
+std::tuple<Args...> from_json_tuple_impl(BasicJsonType&& j, identity_tag<std::tuple<Args...>> /*unused*/, priority_tag<2> /*unused*/)
+{
+    return from_json_tuple_impl_base<BasicJsonType, Args...>(std::forward<BasicJsonType>(j), index_sequence_for<Args...> {});
+}
+
+template<typename BasicJsonType, typename... Args>
+inline void from_json_tuple_impl(BasicJsonType&& j, std::tuple<Args...>& t, priority_tag<3> /*unused*/)
+{
+    t = from_json_tuple_impl_base<BasicJsonType, Args...>(std::forward<BasicJsonType>(j), index_sequence_for<Args...> {});
+}
+
+template<typename BasicJsonType, typename TupleRelated>
+auto from_json(BasicJsonType&& j, TupleRelated&& t)
+-> decltype(from_json_tuple_impl(std::forward<BasicJsonType>(j), std::forward<TupleRelated>(t), priority_tag<3> {}))
+{
+    if (JSON_HEDLEY_UNLIKELY(!j.is_array()))
+    {
+        JSON_THROW(type_error::create(302, concat("type must be array, but is ", j.type_name()), &j));
+    }
+
+    return from_json_tuple_impl(std::forward<BasicJsonType>(j), std::forward<TupleRelated>(t), priority_tag<3> {});
+}
+
+template < typename BasicJsonType, typename Key, typename Value, typename Compare, typename Allocator,
+           typename = enable_if_t < !std::is_constructible <
+                                        typename BasicJsonType::string_t, Key >::value >>
+inline void from_json(const BasicJsonType& j, std::map<Key, Value, Compare, Allocator>& m)
+{
+    if (JSON_HEDLEY_UNLIKELY(!j.is_array()))
+    {
+        JSON_THROW(type_error::create(302, concat("type must be array, but is ", j.type_name()), &j));
+    }
+    m.clear();
+    for (const auto& p : j)
+    {
+        if (JSON_HEDLEY_UNLIKELY(!p.is_array()))
+        {
+            JSON_THROW(type_error::create(302, concat("type must be array, but is ", p.type_name()), &j));
+        }
+        m.emplace(p.at(0).template get<Key>(), p.at(1).template get<Value>());
+    }
+}
+
+template < typename BasicJsonType, typename Key, typename Value, typename Hash, typename KeyEqual, typename Allocator,
+           typename = enable_if_t < !std::is_constructible <
+                                        typename BasicJsonType::string_t, Key >::value >>
+inline void from_json(const BasicJsonType& j, std::unordered_map<Key, Value, Hash, KeyEqual, Allocator>& m)
+{
+    if (JSON_HEDLEY_UNLIKELY(!j.is_array()))
+    {
+        JSON_THROW(type_error::create(302, concat("type must be array, but is ", j.type_name()), &j));
+    }
+    m.clear();
+    for (const auto& p : j)
+    {
+        if (JSON_HEDLEY_UNLIKELY(!p.is_array()))
+        {
+            JSON_THROW(type_error::create(302, concat("type must be array, but is ", p.type_name()), &j));
+        }
+        m.emplace(p.at(0).template get<Key>(), p.at(1).template get<Value>());
+    }
+}
+
+#if JSON_HAS_FILESYSTEM || JSON_HAS_EXPERIMENTAL_FILESYSTEM
+template<typename BasicJsonType>
+inline void from_json(const BasicJsonType& j, std_fs::path& p)
+{
+    if (JSON_HEDLEY_UNLIKELY(!j.is_string()))
+    {
+        JSON_THROW(type_error::create(302, concat("type must be string, but is ", j.type_name()), &j));
+    }
+    p = *j.template get_ptr<const typename BasicJsonType::string_t*>();
+}
+#endif
+
+struct from_json_fn
+{
+    template<typename BasicJsonType, typename T>
+    auto operator()(const BasicJsonType& j, T&& val) const
+    noexcept(noexcept(from_json(j, std::forward<T>(val))))
+    -> decltype(from_json(j, std::forward<T>(val)))
+    {
+        return from_json(j, std::forward<T>(val));
+    }
+};
+
+}  // namespace detail
+
+#ifndef JSON_HAS_CPP_17
+/// namespace to hold default `from_json` function
+/// to see why this is required:
+/// http://www.open-std.org/jtc1/sc22/wg21/docs/papers/2015/n4381.html
+namespace // NOLINT(cert-dcl59-cpp,fuchsia-header-anon-namespaces,google-build-namespaces)
+{
+#endif
+JSON_INLINE_VARIABLE constexpr const auto& from_json = // NOLINT(misc-definitions-in-headers)
+    detail::static_const<detail::from_json_fn>::value;
+#ifndef JSON_HAS_CPP_17
+}  // namespace
+#endif
+
+NLOHMANN_JSON_NAMESPACE_END
+
+// #include <nlohmann/detail/conversions/to_json.hpp>
+//     __ _____ _____ _____
+//  __|  |   __|     |   | |  JSON for Modern C++
+// |  |  |__   |  |  | | | |  version 3.11.2
+// |_____|_____|_____|_|___|  https://github.com/nlohmann/json
+//
+// SPDX-FileCopyrightText: 2013-2022 Niels Lohmann <https://nlohmann.me>
+// SPDX-License-Identifier: MIT
+
+
+
+#include <algorithm> // copy
+#include <iterator> // begin, end
+#include <string> // string
+#include <tuple> // tuple, get
+#include <type_traits> // is_same, is_constructible, is_floating_point, is_enum, underlying_type
+#include <utility> // move, forward, declval, pair
+#include <valarray> // valarray
+#include <vector> // vector
+
+// #include <nlohmann/detail/iterators/iteration_proxy.hpp>
+//     __ _____ _____ _____
+//  __|  |   __|     |   | |  JSON for Modern C++
+// |  |  |__   |  |  | | | |  version 3.11.2
+// |_____|_____|_____|_|___|  https://github.com/nlohmann/json
+//
+// SPDX-FileCopyrightText: 2013-2022 Niels Lohmann <https://nlohmann.me>
+// SPDX-License-Identifier: MIT
+
+
+
+#include <cstddef> // size_t
+#include <iterator> // input_iterator_tag
+#include <string> // string, to_string
+#include <tuple> // tuple_size, get, tuple_element
+#include <utility> // move
+
+#if JSON_HAS_RANGES
+    #include <ranges> // enable_borrowed_range
+#endif
+
+// #include <nlohmann/detail/abi_macros.hpp>
+
+// #include <nlohmann/detail/meta/type_traits.hpp>
+
+// #include <nlohmann/detail/value_t.hpp>
+
+
+NLOHMANN_JSON_NAMESPACE_BEGIN
+namespace detail
+{
+
+template<typename string_type>
+void int_to_string( string_type& target, std::size_t value )
+{
+    // For ADL
+    using std::to_string;
+    target = to_string(value);
+}
+template<typename IteratorType> class iteration_proxy_value
+{
+  public:
+    using difference_type = std::ptrdiff_t;
+    using value_type = iteration_proxy_value;
+    using pointer = value_type *;
+    using reference = value_type &;
+    using iterator_category = std::input_iterator_tag;
+    using string_type = typename std::remove_cv< typename std::remove_reference<decltype( std::declval<IteratorType>().key() ) >::type >::type;
+
+  private:
+    /// the iterator
+    IteratorType anchor{};
+    /// an index for arrays (used to create key names)
+    std::size_t array_index = 0;
+    /// last stringified array index
+    mutable std::size_t array_index_last = 0;
+    /// a string representation of the array index
+    mutable string_type array_index_str = "0";
+    /// an empty string (to return a reference for primitive values)
+    string_type empty_str{};
+
+  public:
+    explicit iteration_proxy_value() = default;
+    explicit iteration_proxy_value(IteratorType it, std::size_t array_index_ = 0)
+    noexcept(std::is_nothrow_move_constructible<IteratorType>::value
+             && std::is_nothrow_default_constructible<string_type>::value)
+        : anchor(std::move(it))
+        , array_index(array_index_)
+    {}
+
+    iteration_proxy_value(iteration_proxy_value const&) = default;
+    iteration_proxy_value& operator=(iteration_proxy_value const&) = default;
+    // older GCCs are a bit fussy and require explicit noexcept specifiers on defaulted functions
+    iteration_proxy_value(iteration_proxy_value&&)
+    noexcept(std::is_nothrow_move_constructible<IteratorType>::value
+             && std::is_nothrow_move_constructible<string_type>::value) = default;
+    iteration_proxy_value& operator=(iteration_proxy_value&&)
+    noexcept(std::is_nothrow_move_assignable<IteratorType>::value
+             && std::is_nothrow_move_assignable<string_type>::value) = default;
+    ~iteration_proxy_value() = default;
+
+    /// dereference operator (needed for range-based for)
+    const iteration_proxy_value& operator*() const
+    {
+        return *this;
+    }
+
+    /// increment operator (needed for range-based for)
+    iteration_proxy_value& operator++()
+    {
+        ++anchor;
+        ++array_index;
+
+        return *this;
+    }
+
+    iteration_proxy_value operator++(int)& // NOLINT(cert-dcl21-cpp)
+    {
+        auto tmp = iteration_proxy_value(anchor, array_index);
+        ++anchor;
+        ++array_index;
+        return tmp;
+    }
+
+    /// equality operator (needed for InputIterator)
+    bool operator==(const iteration_proxy_value& o) const
+    {
+        return anchor == o.anchor;
+    }
+
+    /// inequality operator (needed for range-based for)
+    bool operator!=(const iteration_proxy_value& o) const
+    {
+        return anchor != o.anchor;
+    }
+
+    /// return key of the iterator
+    const string_type& key() const
+    {
+        JSON_ASSERT(anchor.m_object != nullptr);
+
+        switch (anchor.m_object->type())
+        {
+            // use integer array index as key
+            case value_t::array:
+            {
+                if (array_index != array_index_last)
+                {
+                    int_to_string( array_index_str, array_index );
+                    array_index_last = array_index;
+                }
+                return array_index_str;
+            }
+
+            // use key from the object
+            case value_t::object:
+                return anchor.key();
+
+            // use an empty key for all primitive types
+            case value_t::null:
+            case value_t::string:
+            case value_t::boolean:
+            case value_t::number_integer:
+            case value_t::number_unsigned:
+            case value_t::number_float:
+            case value_t::binary:
+            case value_t::discarded:
+            default:
+                return empty_str;
+        }
+    }
+
+    /// return value of the iterator
+    typename IteratorType::reference value() const
+    {
+        return anchor.value();
+    }
+};
+
+/// proxy class for the items() function
+template<typename IteratorType> class iteration_proxy
+{
+  private:
+    /// the container to iterate
+    typename IteratorType::pointer container = nullptr;
+
+  public:
+    explicit iteration_proxy() = default;
+
+    /// construct iteration proxy from a container
+    explicit iteration_proxy(typename IteratorType::reference cont) noexcept
+        : container(&cont) {}
+
+    iteration_proxy(iteration_proxy const&) = default;
+    iteration_proxy& operator=(iteration_proxy const&) = default;
+    iteration_proxy(iteration_proxy&&) noexcept = default;
+    iteration_proxy& operator=(iteration_proxy&&) noexcept = default;
+    ~iteration_proxy() = default;
+
+    /// return iterator begin (needed for range-based for)
+    iteration_proxy_value<IteratorType> begin() const noexcept
+    {
+        return iteration_proxy_value<IteratorType>(container->begin());
+    }
+
+    /// return iterator end (needed for range-based for)
+    iteration_proxy_value<IteratorType> end() const noexcept
+    {
+        return iteration_proxy_value<IteratorType>(container->end());
+    }
+};
+
+// Structured Bindings Support
+// For further reference see https://blog.tartanllama.xyz/structured-bindings/
+// And see https://github.com/nlohmann/json/pull/1391
+template<std::size_t N, typename IteratorType, enable_if_t<N == 0, int> = 0>
+auto get(const nlohmann::detail::iteration_proxy_value<IteratorType>& i) -> decltype(i.key())
+{
+    return i.key();
+}
+// Structured Bindings Support
+// For further reference see https://blog.tartanllama.xyz/structured-bindings/
+// And see https://github.com/nlohmann/json/pull/1391
+template<std::size_t N, typename IteratorType, enable_if_t<N == 1, int> = 0>
+auto get(const nlohmann::detail::iteration_proxy_value<IteratorType>& i) -> decltype(i.value())
+{
+    return i.value();
+}
+
+}  // namespace detail
+NLOHMANN_JSON_NAMESPACE_END
+
+// The Addition to the STD Namespace is required to add
+// Structured Bindings Support to the iteration_proxy_value class
+// For further reference see https://blog.tartanllama.xyz/structured-bindings/
+// And see https://github.com/nlohmann/json/pull/1391
+namespace std
+{
+
+#if defined(__clang__)
+    // Fix: https://github.com/nlohmann/json/issues/1401
+    #pragma clang diagnostic push
+    #pragma clang diagnostic ignored "-Wmismatched-tags"
+#endif
+template<typename IteratorType>
+class tuple_size<::nlohmann::detail::iteration_proxy_value<IteratorType>>
+            : public std::integral_constant<std::size_t, 2> {};
+
+template<std::size_t N, typename IteratorType>
+class tuple_element<N, ::nlohmann::detail::iteration_proxy_value<IteratorType >>
+{
+  public:
+    using type = decltype(
+                     get<N>(std::declval <
+                            ::nlohmann::detail::iteration_proxy_value<IteratorType >> ()));
+};
+#if defined(__clang__)
+    #pragma clang diagnostic pop
+#endif
+
+}  // namespace std
+
+#if JSON_HAS_RANGES
+    template <typename IteratorType>
+    inline constexpr bool ::std::ranges::enable_borrowed_range<::nlohmann::detail::iteration_proxy<IteratorType>> = true;
+#endif
+
+// #include <nlohmann/detail/macro_scope.hpp>
+
+// #include <nlohmann/detail/meta/cpp_future.hpp>
+
+// #include <nlohmann/detail/meta/std_fs.hpp>
+
+// #include <nlohmann/detail/meta/type_traits.hpp>
+
+// #include <nlohmann/detail/value_t.hpp>
+
+
+NLOHMANN_JSON_NAMESPACE_BEGIN
+namespace detail
+{
+
+//////////////////
+// constructors //
+//////////////////
+
+/*
+ * Note all external_constructor<>::construct functions need to call
+ * j.m_value.destroy(j.m_type) to avoid a memory leak in case j contains an
+ * allocated value (e.g., a string). See bug issue
+ * https://github.com/nlohmann/json/issues/2865 for more information.
+ */
+
+template<value_t> struct external_constructor;
+
+template<>
+struct external_constructor<value_t::boolean>
+{
+    template<typename BasicJsonType>
+    static void construct(BasicJsonType& j, typename BasicJsonType::boolean_t b) noexcept
+    {
+        j.m_value.destroy(j.m_type);
+        j.m_type = value_t::boolean;
+        j.m_value = b;
+        j.assert_invariant();
+    }
+};
+
+template<>
+struct external_constructor<value_t::string>
+{
+    template<typename BasicJsonType>
+    static void construct(BasicJsonType& j, const typename BasicJsonType::string_t& s)
+    {
+        j.m_value.destroy(j.m_type);
+        j.m_type = value_t::string;
+        j.m_value = s;
+        j.assert_invariant();
+    }
+
+    template<typename BasicJsonType>
+    static void construct(BasicJsonType& j, typename BasicJsonType::string_t&& s)
+    {
+        j.m_value.destroy(j.m_type);
+        j.m_type = value_t::string;
+        j.m_value = std::move(s);
+        j.assert_invariant();
+    }
+
+    template < typename BasicJsonType, typename CompatibleStringType,
+               enable_if_t < !std::is_same<CompatibleStringType, typename BasicJsonType::string_t>::value,
+                             int > = 0 >
+    static void construct(BasicJsonType& j, const CompatibleStringType& str)
+    {
+        j.m_value.destroy(j.m_type);
+        j.m_type = value_t::string;
+        j.m_value.string = j.template create<typename BasicJsonType::string_t>(str);
+        j.assert_invariant();
+    }
+};
+
+template<>
+struct external_constructor<value_t::binary>
+{
+    template<typename BasicJsonType>
+    static void construct(BasicJsonType& j, const typename BasicJsonType::binary_t& b)
+    {
+        j.m_value.destroy(j.m_type);
+        j.m_type = value_t::binary;
+        j.m_value = typename BasicJsonType::binary_t(b);
+        j.assert_invariant();
+    }
+
+    template<typename BasicJsonType>
+    static void construct(BasicJsonType& j, typename BasicJsonType::binary_t&& b)
+    {
+        j.m_value.destroy(j.m_type);
+        j.m_type = value_t::binary;
+        j.m_value = typename BasicJsonType::binary_t(std::move(b));
+        j.assert_invariant();
+    }
+};
+
+template<>
+struct external_constructor<value_t::number_float>
+{
+    template<typename BasicJsonType>
+    static void construct(BasicJsonType& j, typename BasicJsonType::number_float_t val) noexcept
+    {
+        j.m_value.destroy(j.m_type);
+        j.m_type = value_t::number_float;
+        j.m_value = val;
+        j.assert_invariant();
+    }
+};
+
+template<>
+struct external_constructor<value_t::number_unsigned>
+{
+    template<typename BasicJsonType>
+    static void construct(BasicJsonType& j, typename BasicJsonType::number_unsigned_t val) noexcept
+    {
+        j.m_value.destroy(j.m_type);
+        j.m_type = value_t::number_unsigned;
+        j.m_value = val;
+        j.assert_invariant();
+    }
+};
+
+template<>
+struct external_constructor<value_t::number_integer>
+{
+    template<typename BasicJsonType>
+    static void construct(BasicJsonType& j, typename BasicJsonType::number_integer_t val) noexcept
+    {
+        j.m_value.destroy(j.m_type);
+        j.m_type = value_t::number_integer;
+        j.m_value = val;
+        j.assert_invariant();
+    }
+};
+
+template<>
+struct external_constructor<value_t::array>
+{
+    template<typename BasicJsonType>
+    static void construct(BasicJsonType& j, const typename BasicJsonType::array_t& arr)
+    {
+        j.m_value.destroy(j.m_type);
+        j.m_type = value_t::array;
+        j.m_value = arr;
+        j.set_parents();
+        j.assert_invariant();
+    }
+
+    template<typename BasicJsonType>
+    static void construct(BasicJsonType& j, typename BasicJsonType::array_t&& arr)
+    {
+        j.m_value.destroy(j.m_type);
+        j.m_type = value_t::array;
+        j.m_value = std::move(arr);
+        j.set_parents();
+        j.assert_invariant();
+    }
+
+    template < typename BasicJsonType, typename CompatibleArrayType,
+               enable_if_t < !std::is_same<CompatibleArrayType, typename BasicJsonType::array_t>::value,
+                             int > = 0 >
+    static void construct(BasicJsonType& j, const CompatibleArrayType& arr)
+    {
+        using std::begin;
+        using std::end;
+
+        j.m_value.destroy(j.m_type);
+        j.m_type = value_t::array;
+        j.m_value.array = j.template create<typename BasicJsonType::array_t>(begin(arr), end(arr));
+        j.set_parents();
+        j.assert_invariant();
+    }
+
+    template<typename BasicJsonType>
+    static void construct(BasicJsonType& j, const std::vector<bool>& arr)
+    {
+        j.m_value.destroy(j.m_type);
+        j.m_type = value_t::array;
+        j.m_value = value_t::array;
+        j.m_value.array->reserve(arr.size());
+        for (const bool x : arr)
+        {
+            j.m_value.array->push_back(x);
+            j.set_parent(j.m_value.array->back());
+        }
+        j.assert_invariant();
+    }
+
+    template<typename BasicJsonType, typename T,
+             enable_if_t<std::is_convertible<T, BasicJsonType>::value, int> = 0>
+    static void construct(BasicJsonType& j, const std::valarray<T>& arr)
+    {
+        j.m_value.destroy(j.m_type);
+        j.m_type = value_t::array;
+        j.m_value = value_t::array;
+        j.m_value.array->resize(arr.size());
+        if (arr.size() > 0)
+        {
+            std::copy(std::begin(arr), std::end(arr), j.m_value.array->begin());
+        }
+        j.set_parents();
+        j.assert_invariant();
+    }
+};
+
+template<>
+struct external_constructor<value_t::object>
+{
+    template<typename BasicJsonType>
+    static void construct(BasicJsonType& j, const typename BasicJsonType::object_t& obj)
+    {
+        j.m_value.destroy(j.m_type);
+        j.m_type = value_t::object;
+        j.m_value = obj;
+        j.set_parents();
+        j.assert_invariant();
+    }
+
+    template<typename BasicJsonType>
+    static void construct(BasicJsonType& j, typename BasicJsonType::object_t&& obj)
+    {
+        j.m_value.destroy(j.m_type);
+        j.m_type = value_t::object;
+        j.m_value = std::move(obj);
+        j.set_parents();
+        j.assert_invariant();
+    }
+
+    template < typename BasicJsonType, typename CompatibleObjectType,
+               enable_if_t < !std::is_same<CompatibleObjectType, typename BasicJsonType::object_t>::value, int > = 0 >
+    static void construct(BasicJsonType& j, const CompatibleObjectType& obj)
+    {
+        using std::begin;
+        using std::end;
+
+        j.m_value.destroy(j.m_type);
+        j.m_type = value_t::object;
+        j.m_value.object = j.template create<typename BasicJsonType::object_t>(begin(obj), end(obj));
+        j.set_parents();
+        j.assert_invariant();
+    }
+};
+
+/////////////
+// to_json //
+/////////////
+
+template<typename BasicJsonType, typename T,
+         enable_if_t<std::is_same<T, typename BasicJsonType::boolean_t>::value, int> = 0>
+inline void to_json(BasicJsonType& j, T b) noexcept
+{
+    external_constructor<value_t::boolean>::construct(j, b);
+}
+
+template < typename BasicJsonType, typename BoolRef,
+           enable_if_t <
+               ((std::is_same<std::vector<bool>::reference, BoolRef>::value
+                 && !std::is_same <std::vector<bool>::reference, typename BasicJsonType::boolean_t&>::value)
+                || (std::is_same<std::vector<bool>::const_reference, BoolRef>::value
+                    && !std::is_same <detail::uncvref_t<std::vector<bool>::const_reference>,
+                                      typename BasicJsonType::boolean_t >::value))
+               && std::is_convertible<const BoolRef&, typename BasicJsonType::boolean_t>::value, int > = 0 >
+inline void to_json(BasicJsonType& j, const BoolRef& b) noexcept
+{
+    external_constructor<value_t::boolean>::construct(j, static_cast<typename BasicJsonType::boolean_t>(b));
+}
+
+template<typename BasicJsonType, typename CompatibleString,
+         enable_if_t<std::is_constructible<typename BasicJsonType::string_t, CompatibleString>::value, int> = 0>
+inline void to_json(BasicJsonType& j, const CompatibleString& s)
+{
+    external_constructor<value_t::string>::construct(j, s);
+}
+
+template<typename BasicJsonType>
+inline void to_json(BasicJsonType& j, typename BasicJsonType::string_t&& s)
+{
+    external_constructor<value_t::string>::construct(j, std::move(s));
+}
+
+template<typename BasicJsonType, typename FloatType,
+         enable_if_t<std::is_floating_point<FloatType>::value, int> = 0>
+inline void to_json(BasicJsonType& j, FloatType val) noexcept
+{
+    external_constructor<value_t::number_float>::construct(j, static_cast<typename BasicJsonType::number_float_t>(val));
+}
+
+template<typename BasicJsonType, typename CompatibleNumberUnsignedType,
+         enable_if_t<is_compatible_integer_type<typename BasicJsonType::number_unsigned_t, CompatibleNumberUnsignedType>::value, int> = 0>
+inline void to_json(BasicJsonType& j, CompatibleNumberUnsignedType val) noexcept
+{
+    external_constructor<value_t::number_unsigned>::construct(j, static_cast<typename BasicJsonType::number_unsigned_t>(val));
+}
+
+template<typename BasicJsonType, typename CompatibleNumberIntegerType,
+         enable_if_t<is_compatible_integer_type<typename BasicJsonType::number_integer_t, CompatibleNumberIntegerType>::value, int> = 0>
+inline void to_json(BasicJsonType& j, CompatibleNumberIntegerType val) noexcept
+{
+    external_constructor<value_t::number_integer>::construct(j, static_cast<typename BasicJsonType::number_integer_t>(val));
+}
+
+#if !JSON_DISABLE_ENUM_SERIALIZATION
+template<typename BasicJsonType, typename EnumType,
+         enable_if_t<std::is_enum<EnumType>::value, int> = 0>
+inline void to_json(BasicJsonType& j, EnumType e) noexcept
+{
+    using underlying_type = typename std::underlying_type<EnumType>::type;
+    external_constructor<value_t::number_integer>::construct(j, static_cast<underlying_type>(e));
+}
+#endif  // JSON_DISABLE_ENUM_SERIALIZATION
+
+template<typename BasicJsonType>
+inline void to_json(BasicJsonType& j, const std::vector<bool>& e)
+{
+    external_constructor<value_t::array>::construct(j, e);
+}
+
+template < typename BasicJsonType, typename CompatibleArrayType,
+           enable_if_t < is_compatible_array_type<BasicJsonType,
+                         CompatibleArrayType>::value&&
+                         !is_compatible_object_type<BasicJsonType, CompatibleArrayType>::value&&
+                         !is_compatible_string_type<BasicJsonType, CompatibleArrayType>::value&&
+                         !std::is_same<typename BasicJsonType::binary_t, CompatibleArrayType>::value&&
+                         !is_basic_json<CompatibleArrayType>::value,
+                         int > = 0 >
+inline void to_json(BasicJsonType& j, const CompatibleArrayType& arr)
+{
+    external_constructor<value_t::array>::construct(j, arr);
+}
+
+template<typename BasicJsonType>
+inline void to_json(BasicJsonType& j, const typename BasicJsonType::binary_t& bin)
+{
+    external_constructor<value_t::binary>::construct(j, bin);
+}
+
+template<typename BasicJsonType, typename T,
+         enable_if_t<std::is_convertible<T, BasicJsonType>::value, int> = 0>
+inline void to_json(BasicJsonType& j, const std::valarray<T>& arr)
+{
+    external_constructor<value_t::array>::construct(j, std::move(arr));
+}
+
+template<typename BasicJsonType>
+inline void to_json(BasicJsonType& j, typename BasicJsonType::array_t&& arr)
+{
+    external_constructor<value_t::array>::construct(j, std::move(arr));
+}
+
+template < typename BasicJsonType, typename CompatibleObjectType,
+           enable_if_t < is_compatible_object_type<BasicJsonType, CompatibleObjectType>::value&& !is_basic_json<CompatibleObjectType>::value, int > = 0 >
+inline void to_json(BasicJsonType& j, const CompatibleObjectType& obj)
+{
+    external_constructor<value_t::object>::construct(j, obj);
+}
+
+template<typename BasicJsonType>
+inline void to_json(BasicJsonType& j, typename BasicJsonType::object_t&& obj)
+{
+    external_constructor<value_t::object>::construct(j, std::move(obj));
+}
+
+template <
+    typename BasicJsonType, typename T, std::size_t N,
+    enable_if_t < !std::is_constructible<typename BasicJsonType::string_t,
+                  const T(&)[N]>::value, // NOLINT(cppcoreguidelines-avoid-c-arrays,hicpp-avoid-c-arrays,modernize-avoid-c-arrays)
+                  int > = 0 >
+inline void to_json(BasicJsonType& j, const T(&arr)[N]) // NOLINT(cppcoreguidelines-avoid-c-arrays,hicpp-avoid-c-arrays,modernize-avoid-c-arrays)
+{
+    external_constructor<value_t::array>::construct(j, arr);
+}
+
+template < typename BasicJsonType, typename T1, typename T2, enable_if_t < std::is_constructible<BasicJsonType, T1>::value&& std::is_constructible<BasicJsonType, T2>::value, int > = 0 >
+inline void to_json(BasicJsonType& j, const std::pair<T1, T2>& p)
+{
+    j = { p.first, p.second };
+}
+
+// for https://github.com/nlohmann/json/pull/1134
+template<typename BasicJsonType, typename T,
+         enable_if_t<std::is_same<T, iteration_proxy_value<typename BasicJsonType::iterator>>::value, int> = 0>
+inline void to_json(BasicJsonType& j, const T& b)
+{
+    j = { {b.key(), b.value()} };
+}
+
+template<typename BasicJsonType, typename Tuple, std::size_t... Idx>
+inline void to_json_tuple_impl(BasicJsonType& j, const Tuple& t, index_sequence<Idx...> /*unused*/)
+{
+    j = { std::get<Idx>(t)... };
+}
+
+template<typename BasicJsonType, typename T, enable_if_t<is_constructible_tuple<BasicJsonType, T>::value, int > = 0>
+inline void to_json(BasicJsonType& j, const T& t)
+{
+    to_json_tuple_impl(j, t, make_index_sequence<std::tuple_size<T>::value> {});
+}
+
+#if JSON_HAS_FILESYSTEM || JSON_HAS_EXPERIMENTAL_FILESYSTEM
+template<typename BasicJsonType>
+inline void to_json(BasicJsonType& j, const std_fs::path& p)
+{
+    j = p.string();
+}
+#endif
+
+struct to_json_fn
+{
+    template<typename BasicJsonType, typename T>
+    auto operator()(BasicJsonType& j, T&& val) const noexcept(noexcept(to_json(j, std::forward<T>(val))))
+    -> decltype(to_json(j, std::forward<T>(val)), void())
+    {
+        return to_json(j, std::forward<T>(val));
+    }
+};
+}  // namespace detail
+
+#ifndef JSON_HAS_CPP_17
+/// namespace to hold default `to_json` function
+/// to see why this is required:
+/// http://www.open-std.org/jtc1/sc22/wg21/docs/papers/2015/n4381.html
+namespace // NOLINT(cert-dcl59-cpp,fuchsia-header-anon-namespaces,google-build-namespaces)
+{
+#endif
+JSON_INLINE_VARIABLE constexpr const auto& to_json = // NOLINT(misc-definitions-in-headers)
+    detail::static_const<detail::to_json_fn>::value;
+#ifndef JSON_HAS_CPP_17
+}  // namespace
+#endif
+
+NLOHMANN_JSON_NAMESPACE_END
+
+// #include <nlohmann/detail/meta/identity_tag.hpp>
+
+
+NLOHMANN_JSON_NAMESPACE_BEGIN
+
+/// @sa https://json.nlohmann.me/api/adl_serializer/
+template<typename ValueType, typename>
+struct adl_serializer
+{
+    /// @brief convert a JSON value to any value type
+    /// @sa https://json.nlohmann.me/api/adl_serializer/from_json/
+    template<typename BasicJsonType, typename TargetType = ValueType>
+    static auto from_json(BasicJsonType && j, TargetType& val) noexcept(
+        noexcept(::nlohmann::from_json(std::forward<BasicJsonType>(j), val)))
+    -> decltype(::nlohmann::from_json(std::forward<BasicJsonType>(j), val), void())
+    {
+        ::nlohmann::from_json(std::forward<BasicJsonType>(j), val);
+    }
+
+    /// @brief convert a JSON value to any value type
+    /// @sa https://json.nlohmann.me/api/adl_serializer/from_json/
+    template<typename BasicJsonType, typename TargetType = ValueType>
+    static auto from_json(BasicJsonType && j) noexcept(
+    noexcept(::nlohmann::from_json(std::forward<BasicJsonType>(j), detail::identity_tag<TargetType> {})))
+    -> decltype(::nlohmann::from_json(std::forward<BasicJsonType>(j), detail::identity_tag<TargetType> {}))
+    {
+        return ::nlohmann::from_json(std::forward<BasicJsonType>(j), detail::identity_tag<TargetType> {});
+    }
+
+    /// @brief convert any value type to a JSON value
+    /// @sa https://json.nlohmann.me/api/adl_serializer/to_json/
+    template<typename BasicJsonType, typename TargetType = ValueType>
+    static auto to_json(BasicJsonType& j, TargetType && val) noexcept(
+        noexcept(::nlohmann::to_json(j, std::forward<TargetType>(val))))
+    -> decltype(::nlohmann::to_json(j, std::forward<TargetType>(val)), void())
+    {
+        ::nlohmann::to_json(j, std::forward<TargetType>(val));
+    }
+};
+
+NLOHMANN_JSON_NAMESPACE_END
+
+// #include <nlohmann/byte_container_with_subtype.hpp>
+//     __ _____ _____ _____
+//  __|  |   __|     |   | |  JSON for Modern C++
+// |  |  |__   |  |  | | | |  version 3.11.2
+// |_____|_____|_____|_|___|  https://github.com/nlohmann/json
+//
+// SPDX-FileCopyrightText: 2013-2022 Niels Lohmann <https://nlohmann.me>
+// SPDX-License-Identifier: MIT
+
+
+
+#include <cstdint> // uint8_t, uint64_t
+#include <tuple> // tie
+#include <utility> // move
+
+// #include <nlohmann/detail/abi_macros.hpp>
+
+
+NLOHMANN_JSON_NAMESPACE_BEGIN
+
+/// @brief an internal type for a backed binary type
+/// @sa https://json.nlohmann.me/api/byte_container_with_subtype/
+template<typename BinaryType>
+class byte_container_with_subtype : public BinaryType
+{
+  public:
+    using container_type = BinaryType;
+    using subtype_type = std::uint64_t;
+
+    /// @sa https://json.nlohmann.me/api/byte_container_with_subtype/byte_container_with_subtype/
+    byte_container_with_subtype() noexcept(noexcept(container_type()))
+        : container_type()
+    {}
+
+    /// @sa https://json.nlohmann.me/api/byte_container_with_subtype/byte_container_with_subtype/
+    byte_container_with_subtype(const container_type& b) noexcept(noexcept(container_type(b)))
+        : container_type(b)
+    {}
+
+    /// @sa https://json.nlohmann.me/api/byte_container_with_subtype/byte_container_with_subtype/
+    byte_container_with_subtype(container_type&& b) noexcept(noexcept(container_type(std::move(b))))
+        : container_type(std::move(b))
+    {}
+
+    /// @sa https://json.nlohmann.me/api/byte_container_with_subtype/byte_container_with_subtype/
+    byte_container_with_subtype(const container_type& b, subtype_type subtype_) noexcept(noexcept(container_type(b)))
+        : container_type(b)
+        , m_subtype(subtype_)
+        , m_has_subtype(true)
+    {}
+
+    /// @sa https://json.nlohmann.me/api/byte_container_with_subtype/byte_container_with_subtype/
+    byte_container_with_subtype(container_type&& b, subtype_type subtype_) noexcept(noexcept(container_type(std::move(b))))
+        : container_type(std::move(b))
+        , m_subtype(subtype_)
+        , m_has_subtype(true)
+    {}
+
+    bool operator==(const byte_container_with_subtype& rhs) const
+    {
+        return std::tie(static_cast<const BinaryType&>(*this), m_subtype, m_has_subtype) ==
+               std::tie(static_cast<const BinaryType&>(rhs), rhs.m_subtype, rhs.m_has_subtype);
+    }
+
+    bool operator!=(const byte_container_with_subtype& rhs) const
+    {
+        return !(rhs == *this);
+    }
+
+    /// @brief sets the binary subtype
+    /// @sa https://json.nlohmann.me/api/byte_container_with_subtype/set_subtype/
+    void set_subtype(subtype_type subtype_) noexcept
+    {
+        m_subtype = subtype_;
+        m_has_subtype = true;
+    }
+
+    /// @brief return the binary subtype
+    /// @sa https://json.nlohmann.me/api/byte_container_with_subtype/subtype/
+    constexpr subtype_type subtype() const noexcept
+    {
+        return m_has_subtype ? m_subtype : static_cast<subtype_type>(-1);
+    }
+
+    /// @brief return whether the value has a subtype
+    /// @sa https://json.nlohmann.me/api/byte_container_with_subtype/has_subtype/
+    constexpr bool has_subtype() const noexcept
+    {
+        return m_has_subtype;
+    }
+
+    /// @brief clears the binary subtype
+    /// @sa https://json.nlohmann.me/api/byte_container_with_subtype/clear_subtype/
+    void clear_subtype() noexcept
+    {
+        m_subtype = 0;
+        m_has_subtype = false;
+    }
+
+  private:
+    subtype_type m_subtype = 0;
+    bool m_has_subtype = false;
+};
+
+NLOHMANN_JSON_NAMESPACE_END
+
+// #include <nlohmann/detail/conversions/from_json.hpp>
+
+// #include <nlohmann/detail/conversions/to_json.hpp>
+
+// #include <nlohmann/detail/exceptions.hpp>
+
+// #include <nlohmann/detail/hash.hpp>
+//     __ _____ _____ _____
+//  __|  |   __|     |   | |  JSON for Modern C++
+// |  |  |__   |  |  | | | |  version 3.11.2
+// |_____|_____|_____|_|___|  https://github.com/nlohmann/json
+//
+// SPDX-FileCopyrightText: 2013-2022 Niels Lohmann <https://nlohmann.me>
+// SPDX-License-Identifier: MIT
+
+
+
+#include <cstdint> // uint8_t
+#include <cstddef> // size_t
+#include <functional> // hash
+
+// #include <nlohmann/detail/abi_macros.hpp>
+
+// #include <nlohmann/detail/value_t.hpp>
+
+
+NLOHMANN_JSON_NAMESPACE_BEGIN
+namespace detail
+{
+
+// boost::hash_combine
+inline std::size_t combine(std::size_t seed, std::size_t h) noexcept
+{
+    seed ^= h + 0x9e3779b9 + (seed << 6U) + (seed >> 2U);
+    return seed;
+}
+
+/*!
+@brief hash a JSON value
+
+The hash function tries to rely on std::hash where possible. Furthermore, the
+type of the JSON value is taken into account to have different hash values for
+null, 0, 0U, and false, etc.
+
+@tparam BasicJsonType basic_json specialization
+@param j JSON value to hash
+@return hash value of j
+*/
+template<typename BasicJsonType>
+std::size_t hash(const BasicJsonType& j)
+{
+    using string_t = typename BasicJsonType::string_t;
+    using number_integer_t = typename BasicJsonType::number_integer_t;
+    using number_unsigned_t = typename BasicJsonType::number_unsigned_t;
+    using number_float_t = typename BasicJsonType::number_float_t;
+
+    const auto type = static_cast<std::size_t>(j.type());
+    switch (j.type())
+    {
+        case BasicJsonType::value_t::null:
+        case BasicJsonType::value_t::discarded:
+        {
+            return combine(type, 0);
+        }
+
+        case BasicJsonType::value_t::object:
+        {
+            auto seed = combine(type, j.size());
+            for (const auto& element : j.items())
+            {
+                const auto h = std::hash<string_t> {}(element.key());
+                seed = combine(seed, h);
+                seed = combine(seed, hash(element.value()));
+            }
+            return seed;
+        }
+
+        case BasicJsonType::value_t::array:
+        {
+            auto seed = combine(type, j.size());
+            for (const auto& element : j)
+            {
+                seed = combine(seed, hash(element));
+            }
+            return seed;
+        }
+
+        case BasicJsonType::value_t::string:
+        {
+            const auto h = std::hash<string_t> {}(j.template get_ref<const string_t&>());
+            return combine(type, h);
+        }
+
+        case BasicJsonType::value_t::boolean:
+        {
+            const auto h = std::hash<bool> {}(j.template get<bool>());
+            return combine(type, h);
+        }
+
+        case BasicJsonType::value_t::number_integer:
+        {
+            const auto h = std::hash<number_integer_t> {}(j.template get<number_integer_t>());
+            return combine(type, h);
+        }
+
+        case BasicJsonType::value_t::number_unsigned:
+        {
+            const auto h = std::hash<number_unsigned_t> {}(j.template get<number_unsigned_t>());
+            return combine(type, h);
+        }
+
+        case BasicJsonType::value_t::number_float:
+        {
+            const auto h = std::hash<number_float_t> {}(j.template get<number_float_t>());
+            return combine(type, h);
+        }
+
+        case BasicJsonType::value_t::binary:
+        {
+            auto seed = combine(type, j.get_binary().size());
+            const auto h = std::hash<bool> {}(j.get_binary().has_subtype());
+            seed = combine(seed, h);
+            seed = combine(seed, static_cast<std::size_t>(j.get_binary().subtype()));
+            for (const auto byte : j.get_binary())
+            {
+                seed = combine(seed, std::hash<std::uint8_t> {}(byte));
+            }
+            return seed;
+        }
+
+        default:                   // LCOV_EXCL_LINE
+            JSON_ASSERT(false); // NOLINT(cert-dcl03-c,hicpp-static-assert,misc-static-assert) LCOV_EXCL_LINE
+            return 0;              // LCOV_EXCL_LINE
+    }
+}
+
+}  // namespace detail
+NLOHMANN_JSON_NAMESPACE_END
+
+// #include <nlohmann/detail/input/binary_reader.hpp>
+//     __ _____ _____ _____
+//  __|  |   __|     |   | |  JSON for Modern C++
+// |  |  |__   |  |  | | | |  version 3.11.2
+// |_____|_____|_____|_|___|  https://github.com/nlohmann/json
+//
+// SPDX-FileCopyrightText: 2013-2022 Niels Lohmann <https://nlohmann.me>
+// SPDX-License-Identifier: MIT
+
+
+
+#include <algorithm> // generate_n
+#include <array> // array
+#include <cmath> // ldexp
+#include <cstddef> // size_t
+#include <cstdint> // uint8_t, uint16_t, uint32_t, uint64_t
+#include <cstdio> // snprintf
+#include <cstring> // memcpy
+#include <iterator> // back_inserter
+#include <limits> // numeric_limits
+#include <string> // char_traits, string
+#include <utility> // make_pair, move
+#include <vector> // vector
+
+// #include <nlohmann/detail/exceptions.hpp>
+
+// #include <nlohmann/detail/input/input_adapters.hpp>
+//     __ _____ _____ _____
+//  __|  |   __|     |   | |  JSON for Modern C++
+// |  |  |__   |  |  | | | |  version 3.11.2
+// |_____|_____|_____|_|___|  https://github.com/nlohmann/json
+//
+// SPDX-FileCopyrightText: 2013-2022 Niels Lohmann <https://nlohmann.me>
+// SPDX-License-Identifier: MIT
+
+
+
+#include <array> // array
+#include <cstddef> // size_t
+#include <cstring> // strlen
+#include <iterator> // begin, end, iterator_traits, random_access_iterator_tag, distance, next
+#include <memory> // shared_ptr, make_shared, addressof
+#include <numeric> // accumulate
+#include <string> // string, char_traits
+#include <type_traits> // enable_if, is_base_of, is_pointer, is_integral, remove_pointer
+#include <utility> // pair, declval
+
+#ifndef JSON_NO_IO
+    #include <cstdio>   // FILE *
+    #include <istream>  // istream
+#endif                  // JSON_NO_IO
+
+// #include <nlohmann/detail/iterators/iterator_traits.hpp>
+
+// #include <nlohmann/detail/macro_scope.hpp>
+
+
+NLOHMANN_JSON_NAMESPACE_BEGIN
+namespace detail
+{
+
+/// the supported input formats
+enum class input_format_t { json, cbor, msgpack, ubjson, bson, bjdata };
+
+////////////////////
+// input adapters //
+////////////////////
+
+#ifndef JSON_NO_IO
+/*!
+Input adapter for stdio file access. This adapter read only 1 byte and do not use any
+ buffer. This adapter is a very low level adapter.
+*/
+class file_input_adapter
+{
+  public:
+    using char_type = char;
+
+    JSON_HEDLEY_NON_NULL(2)
+    explicit file_input_adapter(std::FILE* f) noexcept
+        : m_file(f)
+    {
+        JSON_ASSERT(m_file != nullptr);
+    }
+
+    // make class move-only
+    file_input_adapter(const file_input_adapter&) = delete;
+    file_input_adapter(file_input_adapter&&) noexcept = default;
+    file_input_adapter& operator=(const file_input_adapter&) = delete;
+    file_input_adapter& operator=(file_input_adapter&&) = delete;
+    ~file_input_adapter() = default;
+
+    std::char_traits<char>::int_type get_character() noexcept
+    {
+        return std::fgetc(m_file);
+    }
+
+  private:
+    /// the file pointer to read from
+    std::FILE* m_file;
+};
+
+
+/*!
+Input adapter for a (caching) istream. Ignores a UFT Byte Order Mark at
+beginning of input. Does not support changing the underlying std::streambuf
+in mid-input. Maintains underlying std::istream and std::streambuf to support
+subsequent use of standard std::istream operations to process any input
+characters following those used in parsing the JSON input.  Clears the
+std::istream flags; any input errors (e.g., EOF) will be detected by the first
+subsequent call for input from the std::istream.
+*/
+class input_stream_adapter
+{
+  public:
+    using char_type = char;
+
+    ~input_stream_adapter()
+    {
+        // clear stream flags; we use underlying streambuf I/O, do not
+        // maintain ifstream flags, except eof
+        if (is != nullptr)
+        {
+            is->clear(is->rdstate() & std::ios::eofbit);
+        }
+    }
+
+    explicit input_stream_adapter(std::istream& i)
+        : is(&i), sb(i.rdbuf())
+    {}
+
+    // delete because of pointer members
+    input_stream_adapter(const input_stream_adapter&) = delete;
+    input_stream_adapter& operator=(input_stream_adapter&) = delete;
+    input_stream_adapter& operator=(input_stream_adapter&&) = delete;
+
+    input_stream_adapter(input_stream_adapter&& rhs) noexcept
+        : is(rhs.is), sb(rhs.sb)
+    {
+        rhs.is = nullptr;
+        rhs.sb = nullptr;
+    }
+
+    // std::istream/std::streambuf use std::char_traits<char>::to_int_type, to
+    // ensure that std::char_traits<char>::eof() and the character 0xFF do not
+    // end up as the same value, e.g. 0xFFFFFFFF.
+    std::char_traits<char>::int_type get_character()
+    {
+        auto res = sb->sbumpc();
+        // set eof manually, as we don't use the istream interface.
+        if (JSON_HEDLEY_UNLIKELY(res == std::char_traits<char>::eof()))
+        {
+            is->clear(is->rdstate() | std::ios::eofbit);
+        }
+        return res;
+    }
+
+  private:
+    /// the associated input stream
+    std::istream* is = nullptr;
+    std::streambuf* sb = nullptr;
+};
+#endif  // JSON_NO_IO
+
+// General-purpose iterator-based adapter. It might not be as fast as
+// theoretically possible for some containers, but it is extremely versatile.
+template<typename IteratorType>
+class iterator_input_adapter
+{
+  public:
+    using char_type = typename std::iterator_traits<IteratorType>::value_type;
+
+    iterator_input_adapter(IteratorType first, IteratorType last)
+        : current(std::move(first)), end(std::move(last))
+    {}
+
+    typename std::char_traits<char_type>::int_type get_character()
+    {
+        if (JSON_HEDLEY_LIKELY(current != end))
+        {
+            auto result = std::char_traits<char_type>::to_int_type(*current);
+            std::advance(current, 1);
+            return result;
+        }
+
+        return std::char_traits<char_type>::eof();
+    }
+
+  private:
+    IteratorType current;
+    IteratorType end;
+
+    template<typename BaseInputAdapter, size_t T>
+    friend struct wide_string_input_helper;
+
+    bool empty() const
+    {
+        return current == end;
+    }
+};
+
+
+template<typename BaseInputAdapter, size_t T>
+struct wide_string_input_helper;
+
+template<typename BaseInputAdapter>
+struct wide_string_input_helper<BaseInputAdapter, 4>
+{
+    // UTF-32
+    static void fill_buffer(BaseInputAdapter& input,
+                            std::array<std::char_traits<char>::int_type, 4>& utf8_bytes,
+                            size_t& utf8_bytes_index,
+                            size_t& utf8_bytes_filled)
+    {
+        utf8_bytes_index = 0;
+
+        if (JSON_HEDLEY_UNLIKELY(input.empty()))
+        {
+            utf8_bytes[0] = std::char_traits<char>::eof();
+            utf8_bytes_filled = 1;
+        }
+        else
+        {
+            // get the current character
+            const auto wc = input.get_character();
+
+            // UTF-32 to UTF-8 encoding
+            if (wc < 0x80)
+            {
+                utf8_bytes[0] = static_cast<std::char_traits<char>::int_type>(wc);
+                utf8_bytes_filled = 1;
+            }
+            else if (wc <= 0x7FF)
+            {
+                utf8_bytes[0] = static_cast<std::char_traits<char>::int_type>(0xC0u | ((static_cast<unsigned int>(wc) >> 6u) & 0x1Fu));
+                utf8_bytes[1] = static_cast<std::char_traits<char>::int_type>(0x80u | (static_cast<unsigned int>(wc) & 0x3Fu));
+                utf8_bytes_filled = 2;
+            }
+            else if (wc <= 0xFFFF)
+            {
+                utf8_bytes[0] = static_cast<std::char_traits<char>::int_type>(0xE0u | ((static_cast<unsigned int>(wc) >> 12u) & 0x0Fu));
+                utf8_bytes[1] = static_cast<std::char_traits<char>::int_type>(0x80u | ((static_cast<unsigned int>(wc) >> 6u) & 0x3Fu));
+                utf8_bytes[2] = static_cast<std::char_traits<char>::int_type>(0x80u | (static_cast<unsigned int>(wc) & 0x3Fu));
+                utf8_bytes_filled = 3;
+            }
+            else if (wc <= 0x10FFFF)
+            {
+                utf8_bytes[0] = static_cast<std::char_traits<char>::int_type>(0xF0u | ((static_cast<unsigned int>(wc) >> 18u) & 0x07u));
+                utf8_bytes[1] = static_cast<std::char_traits<char>::int_type>(0x80u | ((static_cast<unsigned int>(wc) >> 12u) & 0x3Fu));
+                utf8_bytes[2] = static_cast<std::char_traits<char>::int_type>(0x80u | ((static_cast<unsigned int>(wc) >> 6u) & 0x3Fu));
+                utf8_bytes[3] = static_cast<std::char_traits<char>::int_type>(0x80u | (static_cast<unsigned int>(wc) & 0x3Fu));
+                utf8_bytes_filled = 4;
+            }
+            else
+            {
+                // unknown character
+                utf8_bytes[0] = static_cast<std::char_traits<char>::int_type>(wc);
+                utf8_bytes_filled = 1;
+            }
+        }
+    }
+};
+
+template<typename BaseInputAdapter>
+struct wide_string_input_helper<BaseInputAdapter, 2>
+{
+    // UTF-16
+    static void fill_buffer(BaseInputAdapter& input,
+                            std::array<std::char_traits<char>::int_type, 4>& utf8_bytes,
+                            size_t& utf8_bytes_index,
+                            size_t& utf8_bytes_filled)
+    {
+        utf8_bytes_index = 0;
+
+        if (JSON_HEDLEY_UNLIKELY(input.empty()))
+        {
+            utf8_bytes[0] = std::char_traits<char>::eof();
+            utf8_bytes_filled = 1;
+        }
+        else
+        {
+            // get the current character
+            const auto wc = input.get_character();
+
+            // UTF-16 to UTF-8 encoding
+            if (wc < 0x80)
+            {
+                utf8_bytes[0] = static_cast<std::char_traits<char>::int_type>(wc);
+                utf8_bytes_filled = 1;
+            }
+            else if (wc <= 0x7FF)
+            {
+                utf8_bytes[0] = static_cast<std::char_traits<char>::int_type>(0xC0u | ((static_cast<unsigned int>(wc) >> 6u)));
+                utf8_bytes[1] = static_cast<std::char_traits<char>::int_type>(0x80u | (static_cast<unsigned int>(wc) & 0x3Fu));
+                utf8_bytes_filled = 2;
+            }
+            else if (0xD800 > wc || wc >= 0xE000)
+            {
+                utf8_bytes[0] = static_cast<std::char_traits<char>::int_type>(0xE0u | ((static_cast<unsigned int>(wc) >> 12u)));
+                utf8_bytes[1] = static_cast<std::char_traits<char>::int_type>(0x80u | ((static_cast<unsigned int>(wc) >> 6u) & 0x3Fu));
+                utf8_bytes[2] = static_cast<std::char_traits<char>::int_type>(0x80u | (static_cast<unsigned int>(wc) & 0x3Fu));
+                utf8_bytes_filled = 3;
+            }
+            else
+            {
+                if (JSON_HEDLEY_UNLIKELY(!input.empty()))
+                {
+                    const auto wc2 = static_cast<unsigned int>(input.get_character());
+                    const auto charcode = 0x10000u + (((static_cast<unsigned int>(wc) & 0x3FFu) << 10u) | (wc2 & 0x3FFu));
+                    utf8_bytes[0] = static_cast<std::char_traits<char>::int_type>(0xF0u | (charcode >> 18u));
+                    utf8_bytes[1] = static_cast<std::char_traits<char>::int_type>(0x80u | ((charcode >> 12u) & 0x3Fu));
+                    utf8_bytes[2] = static_cast<std::char_traits<char>::int_type>(0x80u | ((charcode >> 6u) & 0x3Fu));
+                    utf8_bytes[3] = static_cast<std::char_traits<char>::int_type>(0x80u | (charcode & 0x3Fu));
+                    utf8_bytes_filled = 4;
+                }
+                else
+                {
+                    utf8_bytes[0] = static_cast<std::char_traits<char>::int_type>(wc);
+                    utf8_bytes_filled = 1;
+                }
+            }
+        }
+    }
+};
+
+// Wraps another input apdater to convert wide character types into individual bytes.
+template<typename BaseInputAdapter, typename WideCharType>
+class wide_string_input_adapter
+{
+  public:
+    using char_type = char;
+
+    wide_string_input_adapter(BaseInputAdapter base)
+        : base_adapter(base) {}
+
+    typename std::char_traits<char>::int_type get_character() noexcept
+    {
+        // check if buffer needs to be filled
+        if (utf8_bytes_index == utf8_bytes_filled)
+        {
+            fill_buffer<sizeof(WideCharType)>();
+
+            JSON_ASSERT(utf8_bytes_filled > 0);
+            JSON_ASSERT(utf8_bytes_index == 0);
+        }
+
+        // use buffer
+        JSON_ASSERT(utf8_bytes_filled > 0);
+        JSON_ASSERT(utf8_bytes_index < utf8_bytes_filled);
+        return utf8_bytes[utf8_bytes_index++];
+    }
+
+  private:
+    BaseInputAdapter base_adapter;
+
+    template<size_t T>
+    void fill_buffer()
+    {
+        wide_string_input_helper<BaseInputAdapter, T>::fill_buffer(base_adapter, utf8_bytes, utf8_bytes_index, utf8_bytes_filled);
+    }
+
+    /// a buffer for UTF-8 bytes
+    std::array<std::char_traits<char>::int_type, 4> utf8_bytes = {{0, 0, 0, 0}};
+
+    /// index to the utf8_codes array for the next valid byte
+    std::size_t utf8_bytes_index = 0;
+    /// number of valid bytes in the utf8_codes array
+    std::size_t utf8_bytes_filled = 0;
+};
+
+
+template<typename IteratorType, typename Enable = void>
+struct iterator_input_adapter_factory
+{
+    using iterator_type = IteratorType;
+    using char_type = typename std::iterator_traits<iterator_type>::value_type;
+    using adapter_type = iterator_input_adapter<iterator_type>;
+
+    static adapter_type create(IteratorType first, IteratorType last)
+    {
+        return adapter_type(std::move(first), std::move(last));
+    }
+};
+
+template<typename T>
+struct is_iterator_of_multibyte
+{
+    using value_type = typename std::iterator_traits<T>::value_type;
+    enum
+    {
+        value = sizeof(value_type) > 1
+    };
+};
+
+template<typename IteratorType>
+struct iterator_input_adapter_factory<IteratorType, enable_if_t<is_iterator_of_multibyte<IteratorType>::value>>
+{
+    using iterator_type = IteratorType;
+    using char_type = typename std::iterator_traits<iterator_type>::value_type;
+    using base_adapter_type = iterator_input_adapter<iterator_type>;
+    using adapter_type = wide_string_input_adapter<base_adapter_type, char_type>;
+
+    static adapter_type create(IteratorType first, IteratorType last)
+    {
+        return adapter_type(base_adapter_type(std::move(first), std::move(last)));
+    }
+};
+
+// General purpose iterator-based input
+template<typename IteratorType>
+typename iterator_input_adapter_factory<IteratorType>::adapter_type input_adapter(IteratorType first, IteratorType last)
+{
+    using factory_type = iterator_input_adapter_factory<IteratorType>;
+    return factory_type::create(first, last);
+}
+
+// Convenience shorthand from container to iterator
+// Enables ADL on begin(container) and end(container)
+// Encloses the using declarations in namespace for not to leak them to outside scope
+
+namespace container_input_adapter_factory_impl
+{
+
+using std::begin;
+using std::end;
+
+template<typename ContainerType, typename Enable = void>
+struct container_input_adapter_factory {};
+
+template<typename ContainerType>
+struct container_input_adapter_factory< ContainerType,
+       void_t<decltype(begin(std::declval<ContainerType>()), end(std::declval<ContainerType>()))>>
+       {
+           using adapter_type = decltype(input_adapter(begin(std::declval<ContainerType>()), end(std::declval<ContainerType>())));
+
+           static adapter_type create(const ContainerType& container)
+{
+    return input_adapter(begin(container), end(container));
+}
+       };
+
+}  // namespace container_input_adapter_factory_impl
+
+template<typename ContainerType>
+typename container_input_adapter_factory_impl::container_input_adapter_factory<ContainerType>::adapter_type input_adapter(const ContainerType& container)
+{
+    return container_input_adapter_factory_impl::container_input_adapter_factory<ContainerType>::create(container);
+}
+
+#ifndef JSON_NO_IO
+// Special cases with fast paths
+inline file_input_adapter input_adapter(std::FILE* file)
+{
+    return file_input_adapter(file);
+}
+
+inline input_stream_adapter input_adapter(std::istream& stream)
+{
+    return input_stream_adapter(stream);
+}
+
+inline input_stream_adapter input_adapter(std::istream&& stream)
+{
+    return input_stream_adapter(stream);
+}
+#endif  // JSON_NO_IO
+
+using contiguous_bytes_input_adapter = decltype(input_adapter(std::declval<const char*>(), std::declval<const char*>()));
+
+// Null-delimited strings, and the like.
+template < typename CharT,
+           typename std::enable_if <
+               std::is_pointer<CharT>::value&&
+               !std::is_array<CharT>::value&&
+               std::is_integral<typename std::remove_pointer<CharT>::type>::value&&
+               sizeof(typename std::remove_pointer<CharT>::type) == 1,
+               int >::type = 0 >
+contiguous_bytes_input_adapter input_adapter(CharT b)
+{
+    auto length = std::strlen(reinterpret_cast<const char*>(b));
+    const auto* ptr = reinterpret_cast<const char*>(b);
+    return input_adapter(ptr, ptr + length);
+}
+
+template<typename T, std::size_t N>
+auto input_adapter(T (&array)[N]) -> decltype(input_adapter(array, array + N)) // NOLINT(cppcoreguidelines-avoid-c-arrays,hicpp-avoid-c-arrays,modernize-avoid-c-arrays)
+{
+    return input_adapter(array, array + N);
+}
+
+// This class only handles inputs of input_buffer_adapter type.
+// It's required so that expressions like {ptr, len} can be implicitly cast
+// to the correct adapter.
+class span_input_adapter
+{
+  public:
+    template < typename CharT,
+               typename std::enable_if <
+                   std::is_pointer<CharT>::value&&
+                   std::is_integral<typename std::remove_pointer<CharT>::type>::value&&
+                   sizeof(typename std::remove_pointer<CharT>::type) == 1,
+                   int >::type = 0 >
+    span_input_adapter(CharT b, std::size_t l)
+        : ia(reinterpret_cast<const char*>(b), reinterpret_cast<const char*>(b) + l) {}
+
+    template<class IteratorType,
+             typename std::enable_if<
+                 std::is_same<typename iterator_traits<IteratorType>::iterator_category, std::random_access_iterator_tag>::value,
+                 int>::type = 0>
+    span_input_adapter(IteratorType first, IteratorType last)
+        : ia(input_adapter(first, last)) {}
+
+    contiguous_bytes_input_adapter&& get()
+    {
+        return std::move(ia); // NOLINT(hicpp-move-const-arg,performance-move-const-arg)
+    }
+
+  private:
+    contiguous_bytes_input_adapter ia;
+};
+
+}  // namespace detail
+NLOHMANN_JSON_NAMESPACE_END
+
+// #include <nlohmann/detail/input/json_sax.hpp>
+//     __ _____ _____ _____
+//  __|  |   __|     |   | |  JSON for Modern C++
+// |  |  |__   |  |  | | | |  version 3.11.2
+// |_____|_____|_____|_|___|  https://github.com/nlohmann/json
+//
+// SPDX-FileCopyrightText: 2013-2022 Niels Lohmann <https://nlohmann.me>
+// SPDX-License-Identifier: MIT
+
+
+
+#include <cstddef>
+#include <string> // string
+#include <utility> // move
+#include <vector> // vector
+
+// #include <nlohmann/detail/exceptions.hpp>
+
+// #include <nlohmann/detail/macro_scope.hpp>
+
+// #include <nlohmann/detail/string_concat.hpp>
+
+
+NLOHMANN_JSON_NAMESPACE_BEGIN
+
+/*!
+@brief SAX interface
+
+This class describes the SAX interface used by @ref nlohmann::json::sax_parse.
+Each function is called in different situations while the input is parsed. The
+boolean return value informs the parser whether to continue processing the
+input.
+*/
+template<typename BasicJsonType>
+struct json_sax
+{
+    using number_integer_t = typename BasicJsonType::number_integer_t;
+    using number_unsigned_t = typename BasicJsonType::number_unsigned_t;
+    using number_float_t = typename BasicJsonType::number_float_t;
+    using string_t = typename BasicJsonType::string_t;
+    using binary_t = typename BasicJsonType::binary_t;
+
+    /*!
+    @brief a null value was read
+    @return whether parsing should proceed
+    */
+    virtual bool null() = 0;
+
+    /*!
+    @brief a boolean value was read
+    @param[in] val  boolean value
+    @return whether parsing should proceed
+    */
+    virtual bool boolean(bool val) = 0;
+
+    /*!
+    @brief an integer number was read
+    @param[in] val  integer value
+    @return whether parsing should proceed
+    */
+    virtual bool number_integer(number_integer_t val) = 0;
+
+    /*!
+    @brief an unsigned integer number was read
+    @param[in] val  unsigned integer value
+    @return whether parsing should proceed
+    */
+    virtual bool number_unsigned(number_unsigned_t val) = 0;
+
+    /*!
+    @brief a floating-point number was read
+    @param[in] val  floating-point value
+    @param[in] s    raw token value
+    @return whether parsing should proceed
+    */
+    virtual bool number_float(number_float_t val, const string_t& s) = 0;
+
+    /*!
+    @brief a string value was read
+    @param[in] val  string value
+    @return whether parsing should proceed
+    @note It is safe to move the passed string value.
+    */
+    virtual bool string(string_t& val) = 0;
+
+    /*!
+    @brief a binary value was read
+    @param[in] val  binary value
+    @return whether parsing should proceed
+    @note It is safe to move the passed binary value.
+    */
+    virtual bool binary(binary_t& val) = 0;
+
+    /*!
+    @brief the beginning of an object was read
+    @param[in] elements  number of object elements or -1 if unknown
+    @return whether parsing should proceed
+    @note binary formats may report the number of elements
+    */
+    virtual bool start_object(std::size_t elements) = 0;
+
+    /*!
+    @brief an object key was read
+    @param[in] val  object key
+    @return whether parsing should proceed
+    @note It is safe to move the passed string.
+    */
+    virtual bool key(string_t& val) = 0;
+
+    /*!
+    @brief the end of an object was read
+    @return whether parsing should proceed
+    */
+    virtual bool end_object() = 0;
+
+    /*!
+    @brief the beginning of an array was read
+    @param[in] elements  number of array elements or -1 if unknown
+    @return whether parsing should proceed
+    @note binary formats may report the number of elements
+    */
+    virtual bool start_array(std::size_t elements) = 0;
+
+    /*!
+    @brief the end of an array was read
+    @return whether parsing should proceed
+    */
+    virtual bool end_array() = 0;
+
+    /*!
+    @brief a parse error occurred
+    @param[in] position    the position in the input where the error occurs
+    @param[in] last_token  the last read token
+    @param[in] ex          an exception object describing the error
+    @return whether parsing should proceed (must return false)
+    */
+    virtual bool parse_error(std::size_t position,
+                             const std::string& last_token,
+                             const detail::exception& ex) = 0;
+
+    json_sax() = default;
+    json_sax(const json_sax&) = default;
+    json_sax(json_sax&&) noexcept = default;
+    json_sax& operator=(const json_sax&) = default;
+    json_sax& operator=(json_sax&&) noexcept = default;
+    virtual ~json_sax() = default;
+};
+
+
+namespace detail
+{
+/*!
+@brief SAX implementation to create a JSON value from SAX events
+
+This class implements the @ref json_sax interface and processes the SAX events
+to create a JSON value which makes it basically a DOM parser. The structure or
+hierarchy of the JSON value is managed by the stack `ref_stack` which contains
+a pointer to the respective array or object for each recursion depth.
+
+After successful parsing, the value that is passed by reference to the
+constructor contains the parsed value.
+
+@tparam BasicJsonType  the JSON type
+*/
+template<typename BasicJsonType>
+class json_sax_dom_parser
+{
+  public:
+    using number_integer_t = typename BasicJsonType::number_integer_t;
+    using number_unsigned_t = typename BasicJsonType::number_unsigned_t;
+    using number_float_t = typename BasicJsonType::number_float_t;
+    using string_t = typename BasicJsonType::string_t;
+    using binary_t = typename BasicJsonType::binary_t;
+
+    /*!
+    @param[in,out] r  reference to a JSON value that is manipulated while
+                       parsing
+    @param[in] allow_exceptions_  whether parse errors yield exceptions
+    */
+    explicit json_sax_dom_parser(BasicJsonType& r, const bool allow_exceptions_ = true)
+        : root(r), allow_exceptions(allow_exceptions_)
+    {}
+
+    // make class move-only
+    json_sax_dom_parser(const json_sax_dom_parser&) = delete;
+    json_sax_dom_parser(json_sax_dom_parser&&) = default; // NOLINT(hicpp-noexcept-move,performance-noexcept-move-constructor)
+    json_sax_dom_parser& operator=(const json_sax_dom_parser&) = delete;
+    json_sax_dom_parser& operator=(json_sax_dom_parser&&) = default; // NOLINT(hicpp-noexcept-move,performance-noexcept-move-constructor)
+    ~json_sax_dom_parser() = default;
+
+    bool null()
+    {
+        handle_value(nullptr);
+        return true;
+    }
+
+    bool boolean(bool val)
+    {
+        handle_value(val);
+        return true;
+    }
+
+    bool number_integer(number_integer_t val)
+    {
+        handle_value(val);
+        return true;
+    }
+
+    bool number_unsigned(number_unsigned_t val)
+    {
+        handle_value(val);
+        return true;
+    }
+
+    bool number_float(number_float_t val, const string_t& /*unused*/)
+    {
+        handle_value(val);
+        return true;
+    }
+
+    bool string(string_t& val)
+    {
+        handle_value(val);
+        return true;
+    }
+
+    bool binary(binary_t& val)
+    {
+        handle_value(std::move(val));
+        return true;
+    }
+
+    bool start_object(std::size_t len)
+    {
+        ref_stack.push_back(handle_value(BasicJsonType::value_t::object));
+
+        if (JSON_HEDLEY_UNLIKELY(len != static_cast<std::size_t>(-1) && len > ref_stack.back()->max_size()))
+        {
+            JSON_THROW(out_of_range::create(408, concat("excessive object size: ", std::to_string(len)), ref_stack.back()));
+        }
+
+        return true;
+    }
+
+    bool key(string_t& val)
+    {
+        JSON_ASSERT(!ref_stack.empty());
+        JSON_ASSERT(ref_stack.back()->is_object());
+
+        // add null at given key and store the reference for later
+        object_element = &(ref_stack.back()->m_value.object->operator[](val));
+        return true;
+    }
+
+    bool end_object()
+    {
+        JSON_ASSERT(!ref_stack.empty());
+        JSON_ASSERT(ref_stack.back()->is_object());
+
+        ref_stack.back()->set_parents();
+        ref_stack.pop_back();
+        return true;
+    }
+
+    bool start_array(std::size_t len)
+    {
+        ref_stack.push_back(handle_value(BasicJsonType::value_t::array));
+
+        if (JSON_HEDLEY_UNLIKELY(len != static_cast<std::size_t>(-1) && len > ref_stack.back()->max_size()))
+        {
+            JSON_THROW(out_of_range::create(408, concat("excessive array size: ", std::to_string(len)), ref_stack.back()));
+        }
+
+        return true;
+    }
+
+    bool end_array()
+    {
+        JSON_ASSERT(!ref_stack.empty());
+        JSON_ASSERT(ref_stack.back()->is_array());
+
+        ref_stack.back()->set_parents();
+        ref_stack.pop_back();
+        return true;
+    }
+
+    template<class Exception>
+    bool parse_error(std::size_t /*unused*/, const std::string& /*unused*/,
+                     const Exception& ex)
+    {
+        errored = true;
+        static_cast<void>(ex);
+        if (allow_exceptions)
+        {
+            JSON_THROW(ex);
+        }
+        return false;
+    }
+
+    constexpr bool is_errored() const
+    {
+        return errored;
+    }
+
+  private:
+    /*!
+    @invariant If the ref stack is empty, then the passed value will be the new
+               root.
+    @invariant If the ref stack contains a value, then it is an array or an
+               object to which we can add elements
+    */
+    template<typename Value>
+    JSON_HEDLEY_RETURNS_NON_NULL
+    BasicJsonType* handle_value(Value&& v)
+    {
+        if (ref_stack.empty())
+        {
+            root = BasicJsonType(std::forward<Value>(v));
+            return &root;
+        }
+
+        JSON_ASSERT(ref_stack.back()->is_array() || ref_stack.back()->is_object());
+
+        if (ref_stack.back()->is_array())
+        {
+            ref_stack.back()->m_value.array->emplace_back(std::forward<Value>(v));
+            return &(ref_stack.back()->m_value.array->back());
+        }
+
+        JSON_ASSERT(ref_stack.back()->is_object());
+        JSON_ASSERT(object_element);
+        *object_element = BasicJsonType(std::forward<Value>(v));
+        return object_element;
+    }
+
+    /// the parsed JSON value
+    BasicJsonType& root;
+    /// stack to model hierarchy of values
+    std::vector<BasicJsonType*> ref_stack {};
+    /// helper to hold the reference for the next object element
+    BasicJsonType* object_element = nullptr;
+    /// whether a syntax error occurred
+    bool errored = false;
+    /// whether to throw exceptions in case of errors
+    const bool allow_exceptions = true;
+};
+
+template<typename BasicJsonType>
+class json_sax_dom_callback_parser
+{
+  public:
+    using number_integer_t = typename BasicJsonType::number_integer_t;
+    using number_unsigned_t = typename BasicJsonType::number_unsigned_t;
+    using number_float_t = typename BasicJsonType::number_float_t;
+    using string_t = typename BasicJsonType::string_t;
+    using binary_t = typename BasicJsonType::binary_t;
+    using parser_callback_t = typename BasicJsonType::parser_callback_t;
+    using parse_event_t = typename BasicJsonType::parse_event_t;
+
+    json_sax_dom_callback_parser(BasicJsonType& r,
+                                 const parser_callback_t cb,
+                                 const bool allow_exceptions_ = true)
+        : root(r), callback(cb), allow_exceptions(allow_exceptions_)
+    {
+        keep_stack.push_back(true);
+    }
+
+    // make class move-only
+    json_sax_dom_callback_parser(const json_sax_dom_callback_parser&) = delete;
+    json_sax_dom_callback_parser(json_sax_dom_callback_parser&&) = default; // NOLINT(hicpp-noexcept-move,performance-noexcept-move-constructor)
+    json_sax_dom_callback_parser& operator=(const json_sax_dom_callback_parser&) = delete;
+    json_sax_dom_callback_parser& operator=(json_sax_dom_callback_parser&&) = default; // NOLINT(hicpp-noexcept-move,performance-noexcept-move-constructor)
+    ~json_sax_dom_callback_parser() = default;
+
+    bool null()
+    {
+        handle_value(nullptr);
+        return true;
+    }
+
+    bool boolean(bool val)
+    {
+        handle_value(val);
+        return true;
+    }
+
+    bool number_integer(number_integer_t val)
+    {
+        handle_value(val);
+        return true;
+    }
+
+    bool number_unsigned(number_unsigned_t val)
+    {
+        handle_value(val);
+        return true;
+    }
+
+    bool number_float(number_float_t val, const string_t& /*unused*/)
+    {
+        handle_value(val);
+        return true;
+    }
+
+    bool string(string_t& val)
+    {
+        handle_value(val);
+        return true;
+    }
+
+    bool binary(binary_t& val)
+    {
+        handle_value(std::move(val));
+        return true;
+    }
+
+    bool start_object(std::size_t len)
+    {
+        // check callback for object start
+        const bool keep = callback(static_cast<int>(ref_stack.size()), parse_event_t::object_start, discarded);
+        keep_stack.push_back(keep);
+
+        auto val = handle_value(BasicJsonType::value_t::object, true);
+        ref_stack.push_back(val.second);
+
+        // check object limit
+        if (ref_stack.back() && JSON_HEDLEY_UNLIKELY(len != static_cast<std::size_t>(-1) && len > ref_stack.back()->max_size()))
+        {
+            JSON_THROW(out_of_range::create(408, concat("excessive object size: ", std::to_string(len)), ref_stack.back()));
+        }
+
+        return true;
+    }
+
+    bool key(string_t& val)
+    {
+        BasicJsonType k = BasicJsonType(val);
+
+        // check callback for key
+        const bool keep = callback(static_cast<int>(ref_stack.size()), parse_event_t::key, k);
+        key_keep_stack.push_back(keep);
+
+        // add discarded value at given key and store the reference for later
+        if (keep && ref_stack.back())
+        {
+            object_element = &(ref_stack.back()->m_value.object->operator[](val) = discarded);
+        }
+
+        return true;
+    }
+
+    bool end_object()
+    {
+        if (ref_stack.back())
+        {
+            if (!callback(static_cast<int>(ref_stack.size()) - 1, parse_event_t::object_end, *ref_stack.back()))
+            {
+                // discard object
+                *ref_stack.back() = discarded;
+            }
+            else
+            {
+                ref_stack.back()->set_parents();
+            }
+        }
+
+        JSON_ASSERT(!ref_stack.empty());
+        JSON_ASSERT(!keep_stack.empty());
+        ref_stack.pop_back();
+        keep_stack.pop_back();
+
+        if (!ref_stack.empty() && ref_stack.back() && ref_stack.back()->is_structured())
+        {
+            // remove discarded value
+            for (auto it = ref_stack.back()->begin(); it != ref_stack.back()->end(); ++it)
+            {
+                if (it->is_discarded())
+                {
+                    ref_stack.back()->erase(it);
+                    break;
+                }
+            }
+        }
+
+        return true;
+    }
+
+    bool start_array(std::size_t len)
+    {
+        const bool keep = callback(static_cast<int>(ref_stack.size()), parse_event_t::array_start, discarded);
+        keep_stack.push_back(keep);
+
+        auto val = handle_value(BasicJsonType::value_t::array, true);
+        ref_stack.push_back(val.second);
+
+        // check array limit
+        if (ref_stack.back() && JSON_HEDLEY_UNLIKELY(len != static_cast<std::size_t>(-1) && len > ref_stack.back()->max_size()))
+        {
+            JSON_THROW(out_of_range::create(408, concat("excessive array size: ", std::to_string(len)), ref_stack.back()));
+        }
+
+        return true;
+    }
+
+    bool end_array()
+    {
+        bool keep = true;
+
+        if (ref_stack.back())
+        {
+            keep = callback(static_cast<int>(ref_stack.size()) - 1, parse_event_t::array_end, *ref_stack.back());
+            if (keep)
+            {
+                ref_stack.back()->set_parents();
+            }
+            else
+            {
+                // discard array
+                *ref_stack.back() = discarded;
+            }
+        }
+
+        JSON_ASSERT(!ref_stack.empty());
+        JSON_ASSERT(!keep_stack.empty());
+        ref_stack.pop_back();
+        keep_stack.pop_back();
+
+        // remove discarded value
+        if (!keep && !ref_stack.empty() && ref_stack.back()->is_array())
+        {
+            ref_stack.back()->m_value.array->pop_back();
+        }
+
+        return true;
+    }
+
+    template<class Exception>
+    bool parse_error(std::size_t /*unused*/, const std::string& /*unused*/,
+                     const Exception& ex)
+    {
+        errored = true;
+        static_cast<void>(ex);
+        if (allow_exceptions)
+        {
+            JSON_THROW(ex);
+        }
+        return false;
+    }
+
+    constexpr bool is_errored() const
+    {
+        return errored;
+    }
+
+  private:
+    /*!
+    @param[in] v  value to add to the JSON value we build during parsing
+    @param[in] skip_callback  whether we should skip calling the callback
+               function; this is required after start_array() and
+               start_object() SAX events, because otherwise we would call the
+               callback function with an empty array or object, respectively.
+
+    @invariant If the ref stack is empty, then the passed value will be the new
+               root.
+    @invariant If the ref stack contains a value, then it is an array or an
+               object to which we can add elements
+
+    @return pair of boolean (whether value should be kept) and pointer (to the
+            passed value in the ref_stack hierarchy; nullptr if not kept)
+    */
+    template<typename Value>
+    std::pair<bool, BasicJsonType*> handle_value(Value&& v, const bool skip_callback = false)
+    {
+        JSON_ASSERT(!keep_stack.empty());
+
+        // do not handle this value if we know it would be added to a discarded
+        // container
+        if (!keep_stack.back())
+        {
+            return {false, nullptr};
+        }
+
+        // create value
+        auto value = BasicJsonType(std::forward<Value>(v));
+
+        // check callback
+        const bool keep = skip_callback || callback(static_cast<int>(ref_stack.size()), parse_event_t::value, value);
+
+        // do not handle this value if we just learnt it shall be discarded
+        if (!keep)
+        {
+            return {false, nullptr};
+        }
+
+        if (ref_stack.empty())
+        {
+            root = std::move(value);
+            return {true, &root};
+        }
+
+        // skip this value if we already decided to skip the parent
+        // (https://github.com/nlohmann/json/issues/971#issuecomment-413678360)
+        if (!ref_stack.back())
+        {
+            return {false, nullptr};
+        }
+
+        // we now only expect arrays and objects
+        JSON_ASSERT(ref_stack.back()->is_array() || ref_stack.back()->is_object());
+
+        // array
+        if (ref_stack.back()->is_array())
+        {
+            ref_stack.back()->m_value.array->emplace_back(std::move(value));
+            return {true, &(ref_stack.back()->m_value.array->back())};
+        }
+
+        // object
+        JSON_ASSERT(ref_stack.back()->is_object());
+        // check if we should store an element for the current key
+        JSON_ASSERT(!key_keep_stack.empty());
+        const bool store_element = key_keep_stack.back();
+        key_keep_stack.pop_back();
+
+        if (!store_element)
+        {
+            return {false, nullptr};
+        }
+
+        JSON_ASSERT(object_element);
+        *object_element = std::move(value);
+        return {true, object_element};
+    }
+
+    /// the parsed JSON value
+    BasicJsonType& root;
+    /// stack to model hierarchy of values
+    std::vector<BasicJsonType*> ref_stack {};
+    /// stack to manage which values to keep
+    std::vector<bool> keep_stack {};
+    /// stack to manage which object keys to keep
+    std::vector<bool> key_keep_stack {};
+    /// helper to hold the reference for the next object element
+    BasicJsonType* object_element = nullptr;
+    /// whether a syntax error occurred
+    bool errored = false;
+    /// callback function
+    const parser_callback_t callback = nullptr;
+    /// whether to throw exceptions in case of errors
+    const bool allow_exceptions = true;
+    /// a discarded value for the callback
+    BasicJsonType discarded = BasicJsonType::value_t::discarded;
+};
+
+template<typename BasicJsonType>
+class json_sax_acceptor
+{
+  public:
+    using number_integer_t = typename BasicJsonType::number_integer_t;
+    using number_unsigned_t = typename BasicJsonType::number_unsigned_t;
+    using number_float_t = typename BasicJsonType::number_float_t;
+    using string_t = typename BasicJsonType::string_t;
+    using binary_t = typename BasicJsonType::binary_t;
+
+    bool null()
+    {
+        return true;
+    }
+
+    bool boolean(bool /*unused*/)
+    {
+        return true;
+    }
+
+    bool number_integer(number_integer_t /*unused*/)
+    {
+        return true;
+    }
+
+    bool number_unsigned(number_unsigned_t /*unused*/)
+    {
+        return true;
+    }
+
+    bool number_float(number_float_t /*unused*/, const string_t& /*unused*/)
+    {
+        return true;
+    }
+
+    bool string(string_t& /*unused*/)
+    {
+        return true;
+    }
+
+    bool binary(binary_t& /*unused*/)
+    {
+        return true;
+    }
+
+    bool start_object(std::size_t /*unused*/ = static_cast<std::size_t>(-1))
+    {
+        return true;
+    }
+
+    bool key(string_t& /*unused*/)
+    {
+        return true;
+    }
+
+    bool end_object()
+    {
+        return true;
+    }
+
+    bool start_array(std::size_t /*unused*/ = static_cast<std::size_t>(-1))
+    {
+        return true;
+    }
+
+    bool end_array()
+    {
+        return true;
+    }
+
+    bool parse_error(std::size_t /*unused*/, const std::string& /*unused*/, const detail::exception& /*unused*/)
+    {
+        return false;
+    }
+};
+
+}  // namespace detail
+NLOHMANN_JSON_NAMESPACE_END
+
+// #include <nlohmann/detail/input/lexer.hpp>
+//     __ _____ _____ _____
+//  __|  |   __|     |   | |  JSON for Modern C++
+// |  |  |__   |  |  | | | |  version 3.11.2
+// |_____|_____|_____|_|___|  https://github.com/nlohmann/json
+//
+// SPDX-FileCopyrightText: 2013-2022 Niels Lohmann <https://nlohmann.me>
+// SPDX-License-Identifier: MIT
+
+
+
+#include <array> // array
+#include <clocale> // localeconv
+#include <cstddef> // size_t
+#include <cstdio> // snprintf
+#include <cstdlib> // strtof, strtod, strtold, strtoll, strtoull
+#include <initializer_list> // initializer_list
+#include <string> // char_traits, string
+#include <utility> // move
+#include <vector> // vector
+
+// #include <nlohmann/detail/input/input_adapters.hpp>
+
+// #include <nlohmann/detail/input/position_t.hpp>
+
+// #include <nlohmann/detail/macro_scope.hpp>
+
+
+NLOHMANN_JSON_NAMESPACE_BEGIN
+namespace detail
+{
+
+///////////
+// lexer //
+///////////
+
+template<typename BasicJsonType>
+class lexer_base
+{
+  public:
+    /// token types for the parser
+    enum class token_type
+    {
+        uninitialized,    ///< indicating the scanner is uninitialized
+        literal_true,     ///< the `true` literal
+        literal_false,    ///< the `false` literal
+        literal_null,     ///< the `null` literal
+        value_string,     ///< a string -- use get_string() for actual value
+        value_unsigned,   ///< an unsigned integer -- use get_number_unsigned() for actual value
+        value_integer,    ///< a signed integer -- use get_number_integer() for actual value
+        value_float,      ///< an floating point number -- use get_number_float() for actual value
+        begin_array,      ///< the character for array begin `[`
+        begin_object,     ///< the character for object begin `{`
+        end_array,        ///< the character for array end `]`
+        end_object,       ///< the character for object end `}`
+        name_separator,   ///< the name separator `:`
+        value_separator,  ///< the value separator `,`
+        parse_error,      ///< indicating a parse error
+        end_of_input,     ///< indicating the end of the input buffer
+        literal_or_value  ///< a literal or the begin of a value (only for diagnostics)
+    };
+
+    /// return name of values of type token_type (only used for errors)
+    JSON_HEDLEY_RETURNS_NON_NULL
+    JSON_HEDLEY_CONST
+    static const char* token_type_name(const token_type t) noexcept
+    {
+        switch (t)
+        {
+            case token_type::uninitialized:
+                return "<uninitialized>";
+            case token_type::literal_true:
+                return "true literal";
+            case token_type::literal_false:
+                return "false literal";
+            case token_type::literal_null:
+                return "null literal";
+            case token_type::value_string:
+                return "string literal";
+            case token_type::value_unsigned:
+            case token_type::value_integer:
+            case token_type::value_float:
+                return "number literal";
+            case token_type::begin_array:
+                return "'['";
+            case token_type::begin_object:
+                return "'{'";
+            case token_type::end_array:
+                return "']'";
+            case token_type::end_object:
+                return "'}'";
+            case token_type::name_separator:
+                return "':'";
+            case token_type::value_separator:
+                return "','";
+            case token_type::parse_error:
+                return "<parse error>";
+            case token_type::end_of_input:
+                return "end of input";
+            case token_type::literal_or_value:
+                return "'[', '{', or a literal";
+            // LCOV_EXCL_START
+            default: // catch non-enum values
+                return "unknown token";
+                // LCOV_EXCL_STOP
+        }
+    }
+};
+/*!
+@brief lexical analysis
+
+This class organizes the lexical analysis during JSON deserialization.
+*/
+template<typename BasicJsonType, typename InputAdapterType>
+class lexer : public lexer_base<BasicJsonType>
+{
+    using number_integer_t = typename BasicJsonType::number_integer_t;
+    using number_unsigned_t = typename BasicJsonType::number_unsigned_t;
+    using number_float_t = typename BasicJsonType::number_float_t;
+    using string_t = typename BasicJsonType::string_t;
+    using char_type = typename InputAdapterType::char_type;
+    using char_int_type = typename std::char_traits<char_type>::int_type;
+
+  public:
+    using token_type = typename lexer_base<BasicJsonType>::token_type;
+
+    explicit lexer(InputAdapterType&& adapter, bool ignore_comments_ = false) noexcept
+        : ia(std::move(adapter))
+        , ignore_comments(ignore_comments_)
+        , decimal_point_char(static_cast<char_int_type>(get_decimal_point()))
+    {}
+
+    // delete because of pointer members
+    lexer(const lexer&) = delete;
+    lexer(lexer&&) = default; // NOLINT(hicpp-noexcept-move,performance-noexcept-move-constructor)
+    lexer& operator=(lexer&) = delete;
+    lexer& operator=(lexer&&) = default; // NOLINT(hicpp-noexcept-move,performance-noexcept-move-constructor)
+    ~lexer() = default;
+
+  private:
+    /////////////////////
+    // locales
+    /////////////////////
+
+    /// return the locale-dependent decimal point
+    JSON_HEDLEY_PURE
+    static char get_decimal_point() noexcept
+    {
+        const auto* loc = localeconv();
+        JSON_ASSERT(loc != nullptr);
+        return (loc->decimal_point == nullptr) ? '.' : *(loc->decimal_point);
+    }
+
+    /////////////////////
+    // scan functions
+    /////////////////////
+
+    /*!
+    @brief get codepoint from 4 hex characters following `\u`
+
+    For input "\u c1 c2 c3 c4" the codepoint is:
+      (c1 * 0x1000) + (c2 * 0x0100) + (c3 * 0x0010) + c4
+    = (c1 << 12) + (c2 << 8) + (c3 << 4) + (c4 << 0)
+
+    Furthermore, the possible characters '0'..'9', 'A'..'F', and 'a'..'f'
+    must be converted to the integers 0x0..0x9, 0xA..0xF, 0xA..0xF, resp. The
+    conversion is done by subtracting the offset (0x30, 0x37, and 0x57)
+    between the ASCII value of the character and the desired integer value.
+
+    @return codepoint (0x0000..0xFFFF) or -1 in case of an error (e.g. EOF or
+            non-hex character)
+    */
+    int get_codepoint()
+    {
+        // this function only makes sense after reading `\u`
+        JSON_ASSERT(current == 'u');
+        int codepoint = 0;
+
+        const auto factors = { 12u, 8u, 4u, 0u };
+        for (const auto factor : factors)
+        {
+            get();
+
+            if (current >= '0' && current <= '9')
+            {
+                codepoint += static_cast<int>((static_cast<unsigned int>(current) - 0x30u) << factor);
+            }
+            else if (current >= 'A' && current <= 'F')
+            {
+                codepoint += static_cast<int>((static_cast<unsigned int>(current) - 0x37u) << factor);
+            }
+            else if (current >= 'a' && current <= 'f')
+            {
+                codepoint += static_cast<int>((static_cast<unsigned int>(current) - 0x57u) << factor);
+            }
+            else
+            {
+                return -1;
+            }
+        }
+
+        JSON_ASSERT(0x0000 <= codepoint && codepoint <= 0xFFFF);
+        return codepoint;
+    }
+
+    /*!
+    @brief check if the next byte(s) are inside a given range
+
+    Adds the current byte and, for each passed range, reads a new byte and
+    checks if it is inside the range. If a violation was detected, set up an
+    error message and return false. Otherwise, return true.
+
+    @param[in] ranges  list of integers; interpreted as list of pairs of
+                       inclusive lower and upper bound, respectively
+
+    @pre The passed list @a ranges must have 2, 4, or 6 elements; that is,
+         1, 2, or 3 pairs. This precondition is enforced by an assertion.
+
+    @return true if and only if no range violation was detected
+    */
+    bool next_byte_in_range(std::initializer_list<char_int_type> ranges)
+    {
+        JSON_ASSERT(ranges.size() == 2 || ranges.size() == 4 || ranges.size() == 6);
+        add(current);
+
+        for (auto range = ranges.begin(); range != ranges.end(); ++range)
+        {
+            get();
+            if (JSON_HEDLEY_LIKELY(*range <= current && current <= *(++range)))
+            {
+                add(current);
+            }
+            else
+            {
+                error_message = "invalid string: ill-formed UTF-8 byte";
+                return false;
+            }
+        }
+
+        return true;
+    }
+
+    /*!
+    @brief scan a string literal
+
+    This function scans a string according to Sect. 7 of RFC 8259. While
+    scanning, bytes are escaped and copied into buffer token_buffer. Then the
+    function returns successfully, token_buffer is *not* null-terminated (as it
+    may contain \0 bytes), and token_buffer.size() is the number of bytes in the
+    string.
+
+    @return token_type::value_string if string could be successfully scanned,
+            token_type::parse_error otherwise
+
+    @note In case of errors, variable error_message contains a textual
+          description.
+    */
+    token_type scan_string()
+    {
+        // reset token_buffer (ignore opening quote)
+        reset();
+
+        // we entered the function by reading an open quote
+        JSON_ASSERT(current == '\"');
+
+        while (true)
+        {
+            // get next character
+            switch (get())
+            {
+                // end of file while parsing string
+                case std::char_traits<char_type>::eof():
+                {
+                    error_message = "invalid string: missing closing quote";
+                    return token_type::parse_error;
+                }
+
+                // closing quote
+                case '\"':
+                {
+                    return token_type::value_string;
+                }
+
+                // escapes
+                case '\\':
+                {
+                    switch (get())
+                    {
+                        // quotation mark
+                        case '\"':
+                            add('\"');
+                            break;
+                        // reverse solidus
+                        case '\\':
+                            add('\\');
+                            break;
+                        // solidus
+                        case '/':
+                            add('/');
+                            break;
+                        // backspace
+                        case 'b':
+                            add('\b');
+                            break;
+                        // form feed
+                        case 'f':
+                            add('\f');
+                            break;
+                        // line feed
+                        case 'n':
+                            add('\n');
+                            break;
+                        // carriage return
+                        case 'r':
+                            add('\r');
+                            break;
+                        // tab
+                        case 't':
+                            add('\t');
+                            break;
+
+                        // unicode escapes
+                        case 'u':
+                        {
+                            const int codepoint1 = get_codepoint();
+                            int codepoint = codepoint1; // start with codepoint1
+
+                            if (JSON_HEDLEY_UNLIKELY(codepoint1 == -1))
+                            {
+                                error_message = "invalid string: '\\u' must be followed by 4 hex digits";
+                                return token_type::parse_error;
+                            }
+
+                            // check if code point is a high surrogate
+                            if (0xD800 <= codepoint1 && codepoint1 <= 0xDBFF)
+                            {
+                                // expect next \uxxxx entry
+                                if (JSON_HEDLEY_LIKELY(get() == '\\' && get() == 'u'))
+                                {
+                                    const int codepoint2 = get_codepoint();
+
+                                    if (JSON_HEDLEY_UNLIKELY(codepoint2 == -1))
+                                    {
+                                        error_message = "invalid string: '\\u' must be followed by 4 hex digits";
+                                        return token_type::parse_error;
+                                    }
+
+                                    // check if codepoint2 is a low surrogate
+                                    if (JSON_HEDLEY_LIKELY(0xDC00 <= codepoint2 && codepoint2 <= 0xDFFF))
+                                    {
+                                        // overwrite codepoint
+                                        codepoint = static_cast<int>(
+                                                        // high surrogate occupies the most significant 22 bits
+                                                        (static_cast<unsigned int>(codepoint1) << 10u)
+                                                        // low surrogate occupies the least significant 15 bits
+                                                        + static_cast<unsigned int>(codepoint2)
+                                                        // there is still the 0xD800, 0xDC00 and 0x10000 noise
+                                                        // in the result, so we have to subtract with:
+                                                        // (0xD800 << 10) + DC00 - 0x10000 = 0x35FDC00
+                                                        - 0x35FDC00u);
+                                    }
+                                    else
+                                    {
+                                        error_message = "invalid string: surrogate U+D800..U+DBFF must be followed by U+DC00..U+DFFF";
+                                        return token_type::parse_error;
+                                    }
+                                }
+                                else
+                                {
+                                    error_message = "invalid string: surrogate U+D800..U+DBFF must be followed by U+DC00..U+DFFF";
+                                    return token_type::parse_error;
+                                }
+                            }
+                            else
+                            {
+                                if (JSON_HEDLEY_UNLIKELY(0xDC00 <= codepoint1 && codepoint1 <= 0xDFFF))
+                                {
+                                    error_message = "invalid string: surrogate U+DC00..U+DFFF must follow U+D800..U+DBFF";
+                                    return token_type::parse_error;
+                                }
+                            }
+
+                            // result of the above calculation yields a proper codepoint
+                            JSON_ASSERT(0x00 <= codepoint && codepoint <= 0x10FFFF);
+
+                            // translate codepoint into bytes
+                            if (codepoint < 0x80)
+                            {
+                                // 1-byte characters: 0xxxxxxx (ASCII)
+                                add(static_cast<char_int_type>(codepoint));
+                            }
+                            else if (codepoint <= 0x7FF)
+                            {
+                                // 2-byte characters: 110xxxxx 10xxxxxx
+                                add(static_cast<char_int_type>(0xC0u | (static_cast<unsigned int>(codepoint) >> 6u)));
+                                add(static_cast<char_int_type>(0x80u | (static_cast<unsigned int>(codepoint) & 0x3Fu)));
+                            }
+                            else if (codepoint <= 0xFFFF)
+                            {
+                                // 3-byte characters: 1110xxxx 10xxxxxx 10xxxxxx
+                                add(static_cast<char_int_type>(0xE0u | (static_cast<unsigned int>(codepoint) >> 12u)));
+                                add(static_cast<char_int_type>(0x80u | ((static_cast<unsigned int>(codepoint) >> 6u) & 0x3Fu)));
+                                add(static_cast<char_int_type>(0x80u | (static_cast<unsigned int>(codepoint) & 0x3Fu)));
+                            }
+                            else
+                            {
+                                // 4-byte characters: 11110xxx 10xxxxxx 10xxxxxx 10xxxxxx
+                                add(static_cast<char_int_type>(0xF0u | (static_cast<unsigned int>(codepoint) >> 18u)));
+                                add(static_cast<char_int_type>(0x80u | ((static_cast<unsigned int>(codepoint) >> 12u) & 0x3Fu)));
+                                add(static_cast<char_int_type>(0x80u | ((static_cast<unsigned int>(codepoint) >> 6u) & 0x3Fu)));
+                                add(static_cast<char_int_type>(0x80u | (static_cast<unsigned int>(codepoint) & 0x3Fu)));
+                            }
+
+                            break;
+                        }
+
+                        // other characters after escape
+                        default:
+                            error_message = "invalid string: forbidden character after backslash";
+                            return token_type::parse_error;
+                    }
+
+                    break;
+                }
+
+                // invalid control characters
+                case 0x00:
+                {
+                    error_message = "invalid string: control character U+0000 (NUL) must be escaped to \\u0000";
+                    return token_type::parse_error;
+                }
+
+                case 0x01:
+                {
+                    error_message = "invalid string: control character U+0001 (SOH) must be escaped to \\u0001";
+                    return token_type::parse_error;
+                }
+
+                case 0x02:
+                {
+                    error_message = "invalid string: control character U+0002 (STX) must be escaped to \\u0002";
+                    return token_type::parse_error;
+                }
+
+                case 0x03:
+                {
+                    error_message = "invalid string: control character U+0003 (ETX) must be escaped to \\u0003";
+                    return token_type::parse_error;
+                }
+
+                case 0x04:
+                {
+                    error_message = "invalid string: control character U+0004 (EOT) must be escaped to \\u0004";
+                    return token_type::parse_error;
+                }
+
+                case 0x05:
+                {
+                    error_message = "invalid string: control character U+0005 (ENQ) must be escaped to \\u0005";
+                    return token_type::parse_error;
+                }
+
+                case 0x06:
+                {
+                    error_message = "invalid string: control character U+0006 (ACK) must be escaped to \\u0006";
+                    return token_type::parse_error;
+                }
+
+                case 0x07:
+                {
+                    error_message = "invalid string: control character U+0007 (BEL) must be escaped to \\u0007";
+                    return token_type::parse_error;
+                }
+
+                case 0x08:
+                {
+                    error_message = "invalid string: control character U+0008 (BS) must be escaped to \\u0008 or \\b";
+                    return token_type::parse_error;
+                }
+
+                case 0x09:
+                {
+                    error_message = "invalid string: control character U+0009 (HT) must be escaped to \\u0009 or \\t";
+                    return token_type::parse_error;
+                }
+
+                case 0x0A:
+                {
+                    error_message = "invalid string: control character U+000A (LF) must be escaped to \\u000A or \\n";
+                    return token_type::parse_error;
+                }
+
+                case 0x0B:
+                {
+                    error_message = "invalid string: control character U+000B (VT) must be escaped to \\u000B";
+                    return token_type::parse_error;
+                }
+
+                case 0x0C:
+                {
+                    error_message = "invalid string: control character U+000C (FF) must be escaped to \\u000C or \\f";
+                    return token_type::parse_error;
+                }
+
+                case 0x0D:
+                {
+                    error_message = "invalid string: control character U+000D (CR) must be escaped to \\u000D or \\r";
+                    return token_type::parse_error;
+                }
+
+                case 0x0E:
+                {
+                    error_message = "invalid string: control character U+000E (SO) must be escaped to \\u000E";
+                    return token_type::parse_error;
+                }
+
+                case 0x0F:
+                {
+                    error_message = "invalid string: control character U+000F (SI) must be escaped to \\u000F";
+                    return token_type::parse_error;
+                }
+
+                case 0x10:
+                {
+                    error_message = "invalid string: control character U+0010 (DLE) must be escaped to \\u0010";
+                    return token_type::parse_error;
+                }
+
+                case 0x11:
+                {
+                    error_message = "invalid string: control character U+0011 (DC1) must be escaped to \\u0011";
+                    return token_type::parse_error;
+                }
+
+                case 0x12:
+                {
+                    error_message = "invalid string: control character U+0012 (DC2) must be escaped to \\u0012";
+                    return token_type::parse_error;
+                }
+
+                case 0x13:
+                {
+                    error_message = "invalid string: control character U+0013 (DC3) must be escaped to \\u0013";
+                    return token_type::parse_error;
+                }
+
+                case 0x14:
+                {
+                    error_message = "invalid string: control character U+0014 (DC4) must be escaped to \\u0014";
+                    return token_type::parse_error;
+                }
+
+                case 0x15:
+                {
+                    error_message = "invalid string: control character U+0015 (NAK) must be escaped to \\u0015";
+                    return token_type::parse_error;
+                }
+
+                case 0x16:
+                {
+                    error_message = "invalid string: control character U+0016 (SYN) must be escaped to \\u0016";
+                    return token_type::parse_error;
+                }
+
+                case 0x17:
+                {
+                    error_message = "invalid string: control character U+0017 (ETB) must be escaped to \\u0017";
+                    return token_type::parse_error;
+                }
+
+                case 0x18:
+                {
+                    error_message = "invalid string: control character U+0018 (CAN) must be escaped to \\u0018";
+                    return token_type::parse_error;
+                }
+
+                case 0x19:
+                {
+                    error_message = "invalid string: control character U+0019 (EM) must be escaped to \\u0019";
+                    return token_type::parse_error;
+                }
+
+                case 0x1A:
+                {
+                    error_message = "invalid string: control character U+001A (SUB) must be escaped to \\u001A";
+                    return token_type::parse_error;
+                }
+
+                case 0x1B:
+                {
+                    error_message = "invalid string: control character U+001B (ESC) must be escaped to \\u001B";
+                    return token_type::parse_error;
+                }
+
+                case 0x1C:
+                {
+                    error_message = "invalid string: control character U+001C (FS) must be escaped to \\u001C";
+                    return token_type::parse_error;
+                }
+
+                case 0x1D:
+                {
+                    error_message = "invalid string: control character U+001D (GS) must be escaped to \\u001D";
+                    return token_type::parse_error;
+                }
+
+                case 0x1E:
+                {
+                    error_message = "invalid string: control character U+001E (RS) must be escaped to \\u001E";
+                    return token_type::parse_error;
+                }
+
+                case 0x1F:
+                {
+                    error_message = "invalid string: control character U+001F (US) must be escaped to \\u001F";
+                    return token_type::parse_error;
+                }
+
+                // U+0020..U+007F (except U+0022 (quote) and U+005C (backspace))
+                case 0x20:
+                case 0x21:
+                case 0x23:
+                case 0x24:
+                case 0x25:
+                case 0x26:
+                case 0x27:
+                case 0x28:
+                case 0x29:
+                case 0x2A:
+                case 0x2B:
+                case 0x2C:
+                case 0x2D:
+                case 0x2E:
+                case 0x2F:
+                case 0x30:
+                case 0x31:
+                case 0x32:
+                case 0x33:
+                case 0x34:
+                case 0x35:
+                case 0x36:
+                case 0x37:
+                case 0x38:
+                case 0x39:
+                case 0x3A:
+                case 0x3B:
+                case 0x3C:
+                case 0x3D:
+                case 0x3E:
+                case 0x3F:
+                case 0x40:
+                case 0x41:
+                case 0x42:
+                case 0x43:
+                case 0x44:
+                case 0x45:
+                case 0x46:
+                case 0x47:
+                case 0x48:
+                case 0x49:
+                case 0x4A:
+                case 0x4B:
+                case 0x4C:
+                case 0x4D:
+                case 0x4E:
+                case 0x4F:
+                case 0x50:
+                case 0x51:
+                case 0x52:
+                case 0x53:
+                case 0x54:
+                case 0x55:
+                case 0x56:
+                case 0x57:
+                case 0x58:
+                case 0x59:
+                case 0x5A:
+                case 0x5B:
+                case 0x5D:
+                case 0x5E:
+                case 0x5F:
+                case 0x60:
+                case 0x61:
+                case 0x62:
+                case 0x63:
+                case 0x64:
+                case 0x65:
+                case 0x66:
+                case 0x67:
+                case 0x68:
+                case 0x69:
+                case 0x6A:
+                case 0x6B:
+                case 0x6C:
+                case 0x6D:
+                case 0x6E:
+                case 0x6F:
+                case 0x70:
+                case 0x71:
+                case 0x72:
+                case 0x73:
+                case 0x74:
+                case 0x75:
+                case 0x76:
+                case 0x77:
+                case 0x78:
+                case 0x79:
+                case 0x7A:
+                case 0x7B:
+                case 0x7C:
+                case 0x7D:
+                case 0x7E:
+                case 0x7F:
+                {
+                    add(current);
+                    break;
+                }
+
+                // U+0080..U+07FF: bytes C2..DF 80..BF
+                case 0xC2:
+                case 0xC3:
+                case 0xC4:
+                case 0xC5:
+                case 0xC6:
+                case 0xC7:
+                case 0xC8:
+                case 0xC9:
+                case 0xCA:
+                case 0xCB:
+                case 0xCC:
+                case 0xCD:
+                case 0xCE:
+                case 0xCF:
+                case 0xD0:
+                case 0xD1:
+                case 0xD2:
+                case 0xD3:
+                case 0xD4:
+                case 0xD5:
+                case 0xD6:
+                case 0xD7:
+                case 0xD8:
+                case 0xD9:
+                case 0xDA:
+                case 0xDB:
+                case 0xDC:
+                case 0xDD:
+                case 0xDE:
+                case 0xDF:
+                {
+                    if (JSON_HEDLEY_UNLIKELY(!next_byte_in_range({0x80, 0xBF})))
+                    {
+                        return token_type::parse_error;
+                    }
+                    break;
+                }
+
+                // U+0800..U+0FFF: bytes E0 A0..BF 80..BF
+                case 0xE0:
+                {
+                    if (JSON_HEDLEY_UNLIKELY(!(next_byte_in_range({0xA0, 0xBF, 0x80, 0xBF}))))
+                    {
+                        return token_type::parse_error;
+                    }
+                    break;
+                }
+
+                // U+1000..U+CFFF: bytes E1..EC 80..BF 80..BF
+                // U+E000..U+FFFF: bytes EE..EF 80..BF 80..BF
+                case 0xE1:
+                case 0xE2:
+                case 0xE3:
+                case 0xE4:
+                case 0xE5:
+                case 0xE6:
+                case 0xE7:
+                case 0xE8:
+                case 0xE9:
+                case 0xEA:
+                case 0xEB:
+                case 0xEC:
+                case 0xEE:
+                case 0xEF:
+                {
+                    if (JSON_HEDLEY_UNLIKELY(!(next_byte_in_range({0x80, 0xBF, 0x80, 0xBF}))))
+                    {
+                        return token_type::parse_error;
+                    }
+                    break;
+                }
+
+                // U+D000..U+D7FF: bytes ED 80..9F 80..BF
+                case 0xED:
+                {
+                    if (JSON_HEDLEY_UNLIKELY(!(next_byte_in_range({0x80, 0x9F, 0x80, 0xBF}))))
+                    {
+                        return token_type::parse_error;
+                    }
+                    break;
+                }
+
+                // U+10000..U+3FFFF F0 90..BF 80..BF 80..BF
+                case 0xF0:
+                {
+                    if (JSON_HEDLEY_UNLIKELY(!(next_byte_in_range({0x90, 0xBF, 0x80, 0xBF, 0x80, 0xBF}))))
+                    {
+                        return token_type::parse_error;
+                    }
+                    break;
+                }
+
+                // U+40000..U+FFFFF F1..F3 80..BF 80..BF 80..BF
+                case 0xF1:
+                case 0xF2:
+                case 0xF3:
+                {
+                    if (JSON_HEDLEY_UNLIKELY(!(next_byte_in_range({0x80, 0xBF, 0x80, 0xBF, 0x80, 0xBF}))))
+                    {
+                        return token_type::parse_error;
+                    }
+                    break;
+                }
+
+                // U+100000..U+10FFFF F4 80..8F 80..BF 80..BF
+                case 0xF4:
+                {
+                    if (JSON_HEDLEY_UNLIKELY(!(next_byte_in_range({0x80, 0x8F, 0x80, 0xBF, 0x80, 0xBF}))))
+                    {
+                        return token_type::parse_error;
+                    }
+                    break;
+                }
+
+                // remaining bytes (80..C1 and F5..FF) are ill-formed
+                default:
+                {
+                    error_message = "invalid string: ill-formed UTF-8 byte";
+                    return token_type::parse_error;
+                }
+            }
+        }
+    }
+
+    /*!
+     * @brief scan a comment
+     * @return whether comment could be scanned successfully
+     */
+    bool scan_comment()
+    {
+        switch (get())
+        {
+            // single-line comments skip input until a newline or EOF is read
+            case '/':
+            {
+                while (true)
+                {
+                    switch (get())
+                    {
+                        case '\n':
+                        case '\r':
+                        case std::char_traits<char_type>::eof():
+                        case '\0':
+                            return true;
+
+                        default:
+                            break;
+                    }
+                }
+            }
+
+            // multi-line comments skip input until */ is read
+            case '*':
+            {
+                while (true)
+                {
+                    switch (get())
+                    {
+                        case std::char_traits<char_type>::eof():
+                        case '\0':
+                        {
+                            error_message = "invalid comment; missing closing '*/'";
+                            return false;
+                        }
+
+                        case '*':
+                        {
+                            switch (get())
+                            {
+                                case '/':
+                                    return true;
+
+                                default:
+                                {
+                                    unget();
+                                    continue;
+                                }
+                            }
+                        }
+
+                        default:
+                            continue;
+                    }
+                }
+            }
+
+            // unexpected character after reading '/'
+            default:
+            {
+                error_message = "invalid comment; expecting '/' or '*' after '/'";
+                return false;
+            }
+        }
+    }
+
+    JSON_HEDLEY_NON_NULL(2)
+    static void strtof(float& f, const char* str, char** endptr) noexcept
+    {
+        f = std::strtof(str, endptr);
+    }
+
+    JSON_HEDLEY_NON_NULL(2)
+    static void strtof(double& f, const char* str, char** endptr) noexcept
+    {
+        f = std::strtod(str, endptr);
+    }
+
+    JSON_HEDLEY_NON_NULL(2)
+    static void strtof(long double& f, const char* str, char** endptr) noexcept
+    {
+        f = std::strtold(str, endptr);
+    }
+
+    /*!
+    @brief scan a number literal
+
+    This function scans a string according to Sect. 6 of RFC 8259.
+
+    The function is realized with a deterministic finite state machine derived
+    from the grammar described in RFC 8259. Starting in state "init", the
+    input is read and used to determined the next state. Only state "done"
+    accepts the number. State "error" is a trap state to model errors. In the
+    table below, "anything" means any character but the ones listed before.
+
+    state    | 0        | 1-9      | e E      | +       | -       | .        | anything
+    ---------|----------|----------|----------|---------|---------|----------|-----------
+    init     | zero     | any1     | [error]  | [error] | minus   | [error]  | [error]
+    minus    | zero     | any1     | [error]  | [error] | [error] | [error]  | [error]
+    zero     | done     | done     | exponent | done    | done    | decimal1 | done
+    any1     | any1     | any1     | exponent | done    | done    | decimal1 | done
+    decimal1 | decimal2 | decimal2 | [error]  | [error] | [error] | [error]  | [error]
+    decimal2 | decimal2 | decimal2 | exponent | done    | done    | done     | done
+    exponent | any2     | any2     | [error]  | sign    | sign    | [error]  | [error]
+    sign     | any2     | any2     | [error]  | [error] | [error] | [error]  | [error]
+    any2     | any2     | any2     | done     | done    | done    | done     | done
+
+    The state machine is realized with one label per state (prefixed with
+    "scan_number_") and `goto` statements between them. The state machine
+    contains cycles, but any cycle can be left when EOF is read. Therefore,
+    the function is guaranteed to terminate.
+
+    During scanning, the read bytes are stored in token_buffer. This string is
+    then converted to a signed integer, an unsigned integer, or a
+    floating-point number.
+
+    @return token_type::value_unsigned, token_type::value_integer, or
+            token_type::value_float if number could be successfully scanned,
+            token_type::parse_error otherwise
+
+    @note The scanner is independent of the current locale. Internally, the
+          locale's decimal point is used instead of `.` to work with the
+          locale-dependent converters.
+    */
+    token_type scan_number()  // lgtm [cpp/use-of-goto]
+    {
+        // reset token_buffer to store the number's bytes
+        reset();
+
+        // the type of the parsed number; initially set to unsigned; will be
+        // changed if minus sign, decimal point or exponent is read
+        token_type number_type = token_type::value_unsigned;
+
+        // state (init): we just found out we need to scan a number
+        switch (current)
+        {
+            case '-':
+            {
+                add(current);
+                goto scan_number_minus;
+            }
+
+            case '0':
+            {
+                add(current);
+                goto scan_number_zero;
+            }
+
+            case '1':
+            case '2':
+            case '3':
+            case '4':
+            case '5':
+            case '6':
+            case '7':
+            case '8':
+            case '9':
+            {
+                add(current);
+                goto scan_number_any1;
+            }
+
+            // all other characters are rejected outside scan_number()
+            default:            // LCOV_EXCL_LINE
+                JSON_ASSERT(false); // NOLINT(cert-dcl03-c,hicpp-static-assert,misc-static-assert) LCOV_EXCL_LINE
+        }
+
+scan_number_minus:
+        // state: we just parsed a leading minus sign
+        number_type = token_type::value_integer;
+        switch (get())
+        {
+            case '0':
+            {
+                add(current);
+                goto scan_number_zero;
+            }
+
+            case '1':
+            case '2':
+            case '3':
+            case '4':
+            case '5':
+            case '6':
+            case '7':
+            case '8':
+            case '9':
+            {
+                add(current);
+                goto scan_number_any1;
+            }
+
+            default:
+            {
+                error_message = "invalid number; expected digit after '-'";
+                return token_type::parse_error;
+            }
+        }
+
+scan_number_zero:
+        // state: we just parse a zero (maybe with a leading minus sign)
+        switch (get())
+        {
+            case '.':
+            {
+                add(decimal_point_char);
+                goto scan_number_decimal1;
+            }
+
+            case 'e':
+            case 'E':
+            {
+                add(current);
+                goto scan_number_exponent;
+            }
+
+            default:
+                goto scan_number_done;
+        }
+
+scan_number_any1:
+        // state: we just parsed a number 0-9 (maybe with a leading minus sign)
+        switch (get())
+        {
+            case '0':
+            case '1':
+            case '2':
+            case '3':
+            case '4':
+            case '5':
+            case '6':
+            case '7':
+            case '8':
+            case '9':
+            {
+                add(current);
+                goto scan_number_any1;
+            }
+
+            case '.':
+            {
+                add(decimal_point_char);
+                goto scan_number_decimal1;
+            }
+
+            case 'e':
+            case 'E':
+            {
+                add(current);
+                goto scan_number_exponent;
+            }
+
+            default:
+                goto scan_number_done;
+        }
+
+scan_number_decimal1:
+        // state: we just parsed a decimal point
+        number_type = token_type::value_float;
+        switch (get())
+        {
+            case '0':
+            case '1':
+            case '2':
+            case '3':
+            case '4':
+            case '5':
+            case '6':
+            case '7':
+            case '8':
+            case '9':
+            {
+                add(current);
+                goto scan_number_decimal2;
+            }
+
+            default:
+            {
+                error_message = "invalid number; expected digit after '.'";
+                return token_type::parse_error;
+            }
+        }
+
+scan_number_decimal2:
+        // we just parsed at least one number after a decimal point
+        switch (get())
+        {
+            case '0':
+            case '1':
+            case '2':
+            case '3':
+            case '4':
+            case '5':
+            case '6':
+            case '7':
+            case '8':
+            case '9':
+            {
+                add(current);
+                goto scan_number_decimal2;
+            }
+
+            case 'e':
+            case 'E':
+            {
+                add(current);
+                goto scan_number_exponent;
+            }
+
+            default:
+                goto scan_number_done;
+        }
+
+scan_number_exponent:
+        // we just parsed an exponent
+        number_type = token_type::value_float;
+        switch (get())
+        {
+            case '+':
+            case '-':
+            {
+                add(current);
+                goto scan_number_sign;
+            }
+
+            case '0':
+            case '1':
+            case '2':
+            case '3':
+            case '4':
+            case '5':
+            case '6':
+            case '7':
+            case '8':
+            case '9':
+            {
+                add(current);
+                goto scan_number_any2;
+            }
+
+            default:
+            {
+                error_message =
+                    "invalid number; expected '+', '-', or digit after exponent";
+                return token_type::parse_error;
+            }
+        }
+
+scan_number_sign:
+        // we just parsed an exponent sign
+        switch (get())
+        {
+            case '0':
+            case '1':
+            case '2':
+            case '3':
+            case '4':
+            case '5':
+            case '6':
+            case '7':
+            case '8':
+            case '9':
+            {
+                add(current);
+                goto scan_number_any2;
+            }
+
+            default:
+            {
+                error_message = "invalid number; expected digit after exponent sign";
+                return token_type::parse_error;
+            }
+        }
+
+scan_number_any2:
+        // we just parsed a number after the exponent or exponent sign
+        switch (get())
+        {
+            case '0':
+            case '1':
+            case '2':
+            case '3':
+            case '4':
+            case '5':
+            case '6':
+            case '7':
+            case '8':
+            case '9':
+            {
+                add(current);
+                goto scan_number_any2;
+            }
+
+            default:
+                goto scan_number_done;
+        }
+
+scan_number_done:
+        // unget the character after the number (we only read it to know that
+        // we are done scanning a number)
+        unget();
+
+        char* endptr = nullptr; // NOLINT(cppcoreguidelines-pro-type-vararg,hicpp-vararg)
+        errno = 0;
+
+        // try to parse integers first and fall back to floats
+        if (number_type == token_type::value_unsigned)
+        {
+            const auto x = std::strtoull(token_buffer.data(), &endptr, 10);
+
+            // we checked the number format before
+            JSON_ASSERT(endptr == token_buffer.data() + token_buffer.size());
+
+            if (errno == 0)
+            {
+                value_unsigned = static_cast<number_unsigned_t>(x);
+                if (value_unsigned == x)
+                {
+                    return token_type::value_unsigned;
+                }
+            }
+        }
+        else if (number_type == token_type::value_integer)
+        {
+            const auto x = std::strtoll(token_buffer.data(), &endptr, 10);
+
+            // we checked the number format before
+            JSON_ASSERT(endptr == token_buffer.data() + token_buffer.size());
+
+            if (errno == 0)
+            {
+                value_integer = static_cast<number_integer_t>(x);
+                if (value_integer == x)
+                {
+                    return token_type::value_integer;
+                }
+            }
+        }
+
+        // this code is reached if we parse a floating-point number or if an
+        // integer conversion above failed
+        strtof(value_float, token_buffer.data(), &endptr);
+
+        // we checked the number format before
+        JSON_ASSERT(endptr == token_buffer.data() + token_buffer.size());
+
+        return token_type::value_float;
+    }
+
+    /*!
+    @param[in] literal_text  the literal text to expect
+    @param[in] length        the length of the passed literal text
+    @param[in] return_type   the token type to return on success
+    */
+    JSON_HEDLEY_NON_NULL(2)
+    token_type scan_literal(const char_type* literal_text, const std::size_t length,
+                            token_type return_type)
+    {
+        JSON_ASSERT(std::char_traits<char_type>::to_char_type(current) == literal_text[0]);
+        for (std::size_t i = 1; i < length; ++i)
+        {
+            if (JSON_HEDLEY_UNLIKELY(std::char_traits<char_type>::to_char_type(get()) != literal_text[i]))
+            {
+                error_message = "invalid literal";
+                return token_type::parse_error;
+            }
+        }
+        return return_type;
+    }
+
+    /////////////////////
+    // input management
+    /////////////////////
+
+    /// reset token_buffer; current character is beginning of token
+    void reset() noexcept
+    {
+        token_buffer.clear();
+        token_string.clear();
+        token_string.push_back(std::char_traits<char_type>::to_char_type(current));
+    }
+
+    /*
+    @brief get next character from the input
+
+    This function provides the interface to the used input adapter. It does
+    not throw in case the input reached EOF, but returns a
+    `std::char_traits<char>::eof()` in that case.  Stores the scanned characters
+    for use in error messages.
+
+    @return character read from the input
+    */
+    char_int_type get()
+    {
+        ++position.chars_read_total;
+        ++position.chars_read_current_line;
+
+        if (next_unget)
+        {
+            // just reset the next_unget variable and work with current
+            next_unget = false;
+        }
+        else
+        {
+            current = ia.get_character();
+        }
+
+        if (JSON_HEDLEY_LIKELY(current != std::char_traits<char_type>::eof()))
+        {
+            token_string.push_back(std::char_traits<char_type>::to_char_type(current));
+        }
+
+        if (current == '\n')
+        {
+            ++position.lines_read;
+            position.chars_read_current_line = 0;
+        }
+
+        return current;
+    }
+
+    /*!
+    @brief unget current character (read it again on next get)
+
+    We implement unget by setting variable next_unget to true. The input is not
+    changed - we just simulate ungetting by modifying chars_read_total,
+    chars_read_current_line, and token_string. The next call to get() will
+    behave as if the unget character is read again.
+    */
+    void unget()
+    {
+        next_unget = true;
+
+        --position.chars_read_total;
+
+        // in case we "unget" a newline, we have to also decrement the lines_read
+        if (position.chars_read_current_line == 0)
+        {
+            if (position.lines_read > 0)
+            {
+                --position.lines_read;
+            }
+        }
+        else
+        {
+            --position.chars_read_current_line;
+        }
+
+        if (JSON_HEDLEY_LIKELY(current != std::char_traits<char_type>::eof()))
+        {
+            JSON_ASSERT(!token_string.empty());
+            token_string.pop_back();
+        }
+    }
+
+    /// add a character to token_buffer
+    void add(char_int_type c)
+    {
+        token_buffer.push_back(static_cast<typename string_t::value_type>(c));
+    }
+
+  public:
+    /////////////////////
+    // value getters
+    /////////////////////
+
+    /// return integer value
+    constexpr number_integer_t get_number_integer() const noexcept
+    {
+        return value_integer;
+    }
+
+    /// return unsigned integer value
+    constexpr number_unsigned_t get_number_unsigned() const noexcept
+    {
+        return value_unsigned;
+    }
+
+    /// return floating-point value
+    constexpr number_float_t get_number_float() const noexcept
+    {
+        return value_float;
+    }
+
+    /// return current string value (implicitly resets the token; useful only once)
+    string_t& get_string()
+    {
+        return token_buffer;
+    }
+
+    /////////////////////
+    // diagnostics
+    /////////////////////
+
+    /// return position of last read token
+    constexpr position_t get_position() const noexcept
+    {
+        return position;
+    }
+
+    /// return the last read token (for errors only).  Will never contain EOF
+    /// (an arbitrary value that is not a valid char value, often -1), because
+    /// 255 may legitimately occur.  May contain NUL, which should be escaped.
+    std::string get_token_string() const
+    {
+        // escape control characters
+        std::string result;
+        for (const auto c : token_string)
+        {
+            if (static_cast<unsigned char>(c) <= '\x1F')
+            {
+                // escape control characters
+                std::array<char, 9> cs{{}};
+                static_cast<void>((std::snprintf)(cs.data(), cs.size(), "<U+%.4X>", static_cast<unsigned char>(c))); // NOLINT(cppcoreguidelines-pro-type-vararg,hicpp-vararg)
+                result += cs.data();
+            }
+            else
+            {
+                // add character as is
+                result.push_back(static_cast<std::string::value_type>(c));
+            }
+        }
+
+        return result;
+    }
+
+    /// return syntax error message
+    JSON_HEDLEY_RETURNS_NON_NULL
+    constexpr const char* get_error_message() const noexcept
+    {
+        return error_message;
+    }
+
+    /////////////////////
+    // actual scanner
+    /////////////////////
+
+    /*!
+    @brief skip the UTF-8 byte order mark
+    @return true iff there is no BOM or the correct BOM has been skipped
+    */
+    bool skip_bom()
+    {
+        if (get() == 0xEF)
+        {
+            // check if we completely parse the BOM
+            return get() == 0xBB && get() == 0xBF;
+        }
+
+        // the first character is not the beginning of the BOM; unget it to
+        // process is later
+        unget();
+        return true;
+    }
+
+    void skip_whitespace()
+    {
+        do
+        {
+            get();
+        }
+        while (current == ' ' || current == '\t' || current == '\n' || current == '\r');
+    }
+
+    token_type scan()
+    {
+        // initially, skip the BOM
+        if (position.chars_read_total == 0 && !skip_bom())
+        {
+            error_message = "invalid BOM; must be 0xEF 0xBB 0xBF if given";
+            return token_type::parse_error;
+        }
+
+        // read next character and ignore whitespace
+        skip_whitespace();
+
+        // ignore comments
+        while (ignore_comments && current == '/')
+        {
+            if (!scan_comment())
+            {
+                return token_type::parse_error;
+            }
+
+            // skip following whitespace
+            skip_whitespace();
+        }
+
+        switch (current)
+        {
+            // structural characters
+            case '[':
+                return token_type::begin_array;
+            case ']':
+                return token_type::end_array;
+            case '{':
+                return token_type::begin_object;
+            case '}':
+                return token_type::end_object;
+            case ':':
+                return token_type::name_separator;
+            case ',':
+                return token_type::value_separator;
+
+            // literals
+            case 't':
+            {
+                std::array<char_type, 4> true_literal = {{static_cast<char_type>('t'), static_cast<char_type>('r'), static_cast<char_type>('u'), static_cast<char_type>('e')}};
+                return scan_literal(true_literal.data(), true_literal.size(), token_type::literal_true);
+            }
+            case 'f':
+            {
+                std::array<char_type, 5> false_literal = {{static_cast<char_type>('f'), static_cast<char_type>('a'), static_cast<char_type>('l'), static_cast<char_type>('s'), static_cast<char_type>('e')}};
+                return scan_literal(false_literal.data(), false_literal.size(), token_type::literal_false);
+            }
+            case 'n':
+            {
+                std::array<char_type, 4> null_literal = {{static_cast<char_type>('n'), static_cast<char_type>('u'), static_cast<char_type>('l'), static_cast<char_type>('l')}};
+                return scan_literal(null_literal.data(), null_literal.size(), token_type::literal_null);
+            }
+
+            // string
+            case '\"':
+                return scan_string();
+
+            // number
+            case '-':
+            case '0':
+            case '1':
+            case '2':
+            case '3':
+            case '4':
+            case '5':
+            case '6':
+            case '7':
+            case '8':
+            case '9':
+                return scan_number();
+
+            // end of input (the null byte is needed when parsing from
+            // string literals)
+            case '\0':
+            case std::char_traits<char_type>::eof():
+                return token_type::end_of_input;
+
+            // error
+            default:
+                error_message = "invalid literal";
+                return token_type::parse_error;
+        }
+    }
+
+  private:
+    /// input adapter
+    InputAdapterType ia;
+
+    /// whether comments should be ignored (true) or signaled as errors (false)
+    const bool ignore_comments = false;
+
+    /// the current character
+    char_int_type current = std::char_traits<char_type>::eof();
+
+    /// whether the next get() call should just return current
+    bool next_unget = false;
+
+    /// the start position of the current token
+    position_t position {};
+
+    /// raw input token string (for error messages)
+    std::vector<char_type> token_string {};
+
+    /// buffer for variable-length tokens (numbers, strings)
+    string_t token_buffer {};
+
+    /// a description of occurred lexer errors
+    const char* error_message = "";
+
+    // number values
+    number_integer_t value_integer = 0;
+    number_unsigned_t value_unsigned = 0;
+    number_float_t value_float = 0;
+
+    /// the decimal point
+    const char_int_type decimal_point_char = '.';
+};
+
+}  // namespace detail
+NLOHMANN_JSON_NAMESPACE_END
+
+// #include <nlohmann/detail/macro_scope.hpp>
+
+// #include <nlohmann/detail/meta/is_sax.hpp>
+//     __ _____ _____ _____
+//  __|  |   __|     |   | |  JSON for Modern C++
+// |  |  |__   |  |  | | | |  version 3.11.2
+// |_____|_____|_____|_|___|  https://github.com/nlohmann/json
+//
+// SPDX-FileCopyrightText: 2013-2022 Niels Lohmann <https://nlohmann.me>
+// SPDX-License-Identifier: MIT
+
+
+
+#include <cstdint> // size_t
+#include <utility> // declval
+#include <string> // string
+
+// #include <nlohmann/detail/abi_macros.hpp>
+
+// #include <nlohmann/detail/meta/detected.hpp>
+
+// #include <nlohmann/detail/meta/type_traits.hpp>
+
+
+NLOHMANN_JSON_NAMESPACE_BEGIN
+namespace detail
+{
+
+template<typename T>
+using null_function_t = decltype(std::declval<T&>().null());
+
+template<typename T>
+using boolean_function_t =
+    decltype(std::declval<T&>().boolean(std::declval<bool>()));
+
+template<typename T, typename Integer>
+using number_integer_function_t =
+    decltype(std::declval<T&>().number_integer(std::declval<Integer>()));
+
+template<typename T, typename Unsigned>
+using number_unsigned_function_t =
+    decltype(std::declval<T&>().number_unsigned(std::declval<Unsigned>()));
+
+template<typename T, typename Float, typename String>
+using number_float_function_t = decltype(std::declval<T&>().number_float(
+                                    std::declval<Float>(), std::declval<const String&>()));
+
+template<typename T, typename String>
+using string_function_t =
+    decltype(std::declval<T&>().string(std::declval<String&>()));
+
+template<typename T, typename Binary>
+using binary_function_t =
+    decltype(std::declval<T&>().binary(std::declval<Binary&>()));
+
+template<typename T>
+using start_object_function_t =
+    decltype(std::declval<T&>().start_object(std::declval<std::size_t>()));
+
+template<typename T, typename String>
+using key_function_t =
+    decltype(std::declval<T&>().key(std::declval<String&>()));
+
+template<typename T>
+using end_object_function_t = decltype(std::declval<T&>().end_object());
+
+template<typename T>
+using start_array_function_t =
+    decltype(std::declval<T&>().start_array(std::declval<std::size_t>()));
+
+template<typename T>
+using end_array_function_t = decltype(std::declval<T&>().end_array());
+
+template<typename T, typename Exception>
+using parse_error_function_t = decltype(std::declval<T&>().parse_error(
+        std::declval<std::size_t>(), std::declval<const std::string&>(),
+        std::declval<const Exception&>()));
+
+template<typename SAX, typename BasicJsonType>
+struct is_sax
+{
+  private:
+    static_assert(is_basic_json<BasicJsonType>::value,
+                  "BasicJsonType must be of type basic_json<...>");
+
+    using number_integer_t = typename BasicJsonType::number_integer_t;
+    using number_unsigned_t = typename BasicJsonType::number_unsigned_t;
+    using number_float_t = typename BasicJsonType::number_float_t;
+    using string_t = typename BasicJsonType::string_t;
+    using binary_t = typename BasicJsonType::binary_t;
+    using exception_t = typename BasicJsonType::exception;
+
+  public:
+    static constexpr bool value =
+        is_detected_exact<bool, null_function_t, SAX>::value &&
+        is_detected_exact<bool, boolean_function_t, SAX>::value &&
+        is_detected_exact<bool, number_integer_function_t, SAX, number_integer_t>::value &&
+        is_detected_exact<bool, number_unsigned_function_t, SAX, number_unsigned_t>::value &&
+        is_detected_exact<bool, number_float_function_t, SAX, number_float_t, string_t>::value &&
+        is_detected_exact<bool, string_function_t, SAX, string_t>::value &&
+        is_detected_exact<bool, binary_function_t, SAX, binary_t>::value &&
+        is_detected_exact<bool, start_object_function_t, SAX>::value &&
+        is_detected_exact<bool, key_function_t, SAX, string_t>::value &&
+        is_detected_exact<bool, end_object_function_t, SAX>::value &&
+        is_detected_exact<bool, start_array_function_t, SAX>::value &&
+        is_detected_exact<bool, end_array_function_t, SAX>::value &&
+        is_detected_exact<bool, parse_error_function_t, SAX, exception_t>::value;
+};
+
+template<typename SAX, typename BasicJsonType>
+struct is_sax_static_asserts
+{
+  private:
+    static_assert(is_basic_json<BasicJsonType>::value,
+                  "BasicJsonType must be of type basic_json<...>");
+
+    using number_integer_t = typename BasicJsonType::number_integer_t;
+    using number_unsigned_t = typename BasicJsonType::number_unsigned_t;
+    using number_float_t = typename BasicJsonType::number_float_t;
+    using string_t = typename BasicJsonType::string_t;
+    using binary_t = typename BasicJsonType::binary_t;
+    using exception_t = typename BasicJsonType::exception;
+
+  public:
+    static_assert(is_detected_exact<bool, null_function_t, SAX>::value,
+                  "Missing/invalid function: bool null()");
+    static_assert(is_detected_exact<bool, boolean_function_t, SAX>::value,
+                  "Missing/invalid function: bool boolean(bool)");
+    static_assert(is_detected_exact<bool, boolean_function_t, SAX>::value,
+                  "Missing/invalid function: bool boolean(bool)");
+    static_assert(
+        is_detected_exact<bool, number_integer_function_t, SAX,
+        number_integer_t>::value,
+        "Missing/invalid function: bool number_integer(number_integer_t)");
+    static_assert(
+        is_detected_exact<bool, number_unsigned_function_t, SAX,
+        number_unsigned_t>::value,
+        "Missing/invalid function: bool number_unsigned(number_unsigned_t)");
+    static_assert(is_detected_exact<bool, number_float_function_t, SAX,
+                  number_float_t, string_t>::value,
+                  "Missing/invalid function: bool number_float(number_float_t, const string_t&)");
+    static_assert(
+        is_detected_exact<bool, string_function_t, SAX, string_t>::value,
+        "Missing/invalid function: bool string(string_t&)");
+    static_assert(
+        is_detected_exact<bool, binary_function_t, SAX, binary_t>::value,
+        "Missing/invalid function: bool binary(binary_t&)");
+    static_assert(is_detected_exact<bool, start_object_function_t, SAX>::value,
+                  "Missing/invalid function: bool start_object(std::size_t)");
+    static_assert(is_detected_exact<bool, key_function_t, SAX, string_t>::value,
+                  "Missing/invalid function: bool key(string_t&)");
+    static_assert(is_detected_exact<bool, end_object_function_t, SAX>::value,
+                  "Missing/invalid function: bool end_object()");
+    static_assert(is_detected_exact<bool, start_array_function_t, SAX>::value,
+                  "Missing/invalid function: bool start_array(std::size_t)");
+    static_assert(is_detected_exact<bool, end_array_function_t, SAX>::value,
+                  "Missing/invalid function: bool end_array()");
+    static_assert(
+        is_detected_exact<bool, parse_error_function_t, SAX, exception_t>::value,
+        "Missing/invalid function: bool parse_error(std::size_t, const "
+        "std::string&, const exception&)");
+};
+
+}  // namespace detail
+NLOHMANN_JSON_NAMESPACE_END
+
+// #include <nlohmann/detail/meta/type_traits.hpp>
+
+// #include <nlohmann/detail/string_concat.hpp>
+
+// #include <nlohmann/detail/value_t.hpp>
+
+
+NLOHMANN_JSON_NAMESPACE_BEGIN
+namespace detail
+{
+
+/// how to treat CBOR tags
+enum class cbor_tag_handler_t
+{
+    error,   ///< throw a parse_error exception in case of a tag
+    ignore,  ///< ignore tags
+    store    ///< store tags as binary type
+};
+
+/*!
+@brief determine system byte order
+
+@return true if and only if system's byte order is little endian
+
+@note from https://stackoverflow.com/a/1001328/266378
+*/
+static inline bool little_endianness(int num = 1) noexcept
+{
+    return *reinterpret_cast<char*>(&num) == 1;
+}
+
+
+///////////////////
+// binary reader //
+///////////////////
+
+/*!
+@brief deserialization of CBOR, MessagePack, and UBJSON values
+*/
+template<typename BasicJsonType, typename InputAdapterType, typename SAX = json_sax_dom_parser<BasicJsonType>>
+class binary_reader
+{
+    using number_integer_t = typename BasicJsonType::number_integer_t;
+    using number_unsigned_t = typename BasicJsonType::number_unsigned_t;
+    using number_float_t = typename BasicJsonType::number_float_t;
+    using string_t = typename BasicJsonType::string_t;
+    using binary_t = typename BasicJsonType::binary_t;
+    using json_sax_t = SAX;
+    using char_type = typename InputAdapterType::char_type;
+    using char_int_type = typename std::char_traits<char_type>::int_type;
+
+  public:
+    /*!
+    @brief create a binary reader
+
+    @param[in] adapter  input adapter to read from
+    */
+    explicit binary_reader(InputAdapterType&& adapter, const input_format_t format = input_format_t::json) noexcept : ia(std::move(adapter)), input_format(format)
+    {
+        (void)detail::is_sax_static_asserts<SAX, BasicJsonType> {};
+    }
+
+    // make class move-only
+    binary_reader(const binary_reader&) = delete;
+    binary_reader(binary_reader&&) = default; // NOLINT(hicpp-noexcept-move,performance-noexcept-move-constructor)
+    binary_reader& operator=(const binary_reader&) = delete;
+    binary_reader& operator=(binary_reader&&) = default; // NOLINT(hicpp-noexcept-move,performance-noexcept-move-constructor)
+    ~binary_reader() = default;
+
+    /*!
+    @param[in] format  the binary format to parse
+    @param[in] sax_    a SAX event processor
+    @param[in] strict  whether to expect the input to be consumed completed
+    @param[in] tag_handler  how to treat CBOR tags
+
+    @return whether parsing was successful
+    */
+    JSON_HEDLEY_NON_NULL(3)
+    bool sax_parse(const input_format_t format,
+                   json_sax_t* sax_,
+                   const bool strict = true,
+                   const cbor_tag_handler_t tag_handler = cbor_tag_handler_t::error)
+    {
+        sax = sax_;
+        bool result = false;
+
+        switch (format)
+        {
+            case input_format_t::bson:
+                result = parse_bson_internal();
+                break;
+
+            case input_format_t::cbor:
+                result = parse_cbor_internal(true, tag_handler);
+                break;
+
+            case input_format_t::msgpack:
+                result = parse_msgpack_internal();
+                break;
+
+            case input_format_t::ubjson:
+            case input_format_t::bjdata:
+                result = parse_ubjson_internal();
+                break;
+
+            case input_format_t::json: // LCOV_EXCL_LINE
+            default:            // LCOV_EXCL_LINE
+                JSON_ASSERT(false); // NOLINT(cert-dcl03-c,hicpp-static-assert,misc-static-assert) LCOV_EXCL_LINE
+        }
+
+        // strict mode: next byte must be EOF
+        if (result && strict)
+        {
+            if (input_format == input_format_t::ubjson || input_format == input_format_t::bjdata)
+            {
+                get_ignore_noop();
+            }
+            else
+            {
+                get();
+            }
+
+            if (JSON_HEDLEY_UNLIKELY(current != std::char_traits<char_type>::eof()))
+            {
+                return sax->parse_error(chars_read, get_token_string(), parse_error::create(110, chars_read,
+                                        exception_message(input_format, concat("expected end of input; last byte: 0x", get_token_string()), "value"), nullptr));
+            }
+        }
+
+        return result;
+    }
+
+  private:
+    //////////
+    // BSON //
+    //////////
+
+    /*!
+    @brief Reads in a BSON-object and passes it to the SAX-parser.
+    @return whether a valid BSON-value was passed to the SAX parser
+    */
+    bool parse_bson_internal()
+    {
+        std::int32_t document_size{};
+        get_number<std::int32_t, true>(input_format_t::bson, document_size);
+
+        if (JSON_HEDLEY_UNLIKELY(!sax->start_object(static_cast<std::size_t>(-1))))
+        {
+            return false;
+        }
+
+        if (JSON_HEDLEY_UNLIKELY(!parse_bson_element_list(/*is_array*/false)))
+        {
+            return false;
+        }
+
+        return sax->end_object();
+    }
+
+    /*!
+    @brief Parses a C-style string from the BSON input.
+    @param[in,out] result  A reference to the string variable where the read
+                            string is to be stored.
+    @return `true` if the \x00-byte indicating the end of the string was
+             encountered before the EOF; false` indicates an unexpected EOF.
+    */
+    bool get_bson_cstr(string_t& result)
+    {
+        auto out = std::back_inserter(result);
+        while (true)
+        {
+            get();
+            if (JSON_HEDLEY_UNLIKELY(!unexpect_eof(input_format_t::bson, "cstring")))
+            {
+                return false;
+            }
+            if (current == 0x00)
+            {
+                return true;
+            }
+            *out++ = static_cast<typename string_t::value_type>(current);
+        }
+    }
+
+    /*!
+    @brief Parses a zero-terminated string of length @a len from the BSON
+           input.
+    @param[in] len  The length (including the zero-byte at the end) of the
+                    string to be read.
+    @param[in,out] result  A reference to the string variable where the read
+                            string is to be stored.
+    @tparam NumberType The type of the length @a len
+    @pre len >= 1
+    @return `true` if the string was successfully parsed
+    */
+    template<typename NumberType>
+    bool get_bson_string(const NumberType len, string_t& result)
+    {
+        if (JSON_HEDLEY_UNLIKELY(len < 1))
+        {
+            auto last_token = get_token_string();
+            return sax->parse_error(chars_read, last_token, parse_error::create(112, chars_read,
+                                    exception_message(input_format_t::bson, concat("string length must be at least 1, is ", std::to_string(len)), "string"), nullptr));
+        }
+
+        return get_string(input_format_t::bson, len - static_cast<NumberType>(1), result) && get() != std::char_traits<char_type>::eof();
+    }
+
+    /*!
+    @brief Parses a byte array input of length @a len from the BSON input.
+    @param[in] len  The length of the byte array to be read.
+    @param[in,out] result  A reference to the binary variable where the read
+                            array is to be stored.
+    @tparam NumberType The type of the length @a len
+    @pre len >= 0
+    @return `true` if the byte array was successfully parsed
+    */
+    template<typename NumberType>
+    bool get_bson_binary(const NumberType len, binary_t& result)
+    {
+        if (JSON_HEDLEY_UNLIKELY(len < 0))
+        {
+            auto last_token = get_token_string();
+            return sax->parse_error(chars_read, last_token, parse_error::create(112, chars_read,
+                                    exception_message(input_format_t::bson, concat("byte array length cannot be negative, is ", std::to_string(len)), "binary"), nullptr));
+        }
+
+        // All BSON binary values have a subtype
+        std::uint8_t subtype{};
+        get_number<std::uint8_t>(input_format_t::bson, subtype);
+        result.set_subtype(subtype);
+
+        return get_binary(input_format_t::bson, len, result);
+    }
+
+    /*!
+    @brief Read a BSON document element of the given @a element_type.
+    @param[in] element_type The BSON element type, c.f. http://bsonspec.org/spec.html
+    @param[in] element_type_parse_position The position in the input stream,
+               where the `element_type` was read.
+    @warning Not all BSON element types are supported yet. An unsupported
+             @a element_type will give rise to a parse_error.114:
+             Unsupported BSON record type 0x...
+    @return whether a valid BSON-object/array was passed to the SAX parser
+    */
+    bool parse_bson_element_internal(const char_int_type element_type,
+                                     const std::size_t element_type_parse_position)
+    {
+        switch (element_type)
+        {
+            case 0x01: // double
+            {
+                double number{};
+                return get_number<double, true>(input_format_t::bson, number) && sax->number_float(static_cast<number_float_t>(number), "");
+            }
+
+            case 0x02: // string
+            {
+                std::int32_t len{};
+                string_t value;
+                return get_number<std::int32_t, true>(input_format_t::bson, len) && get_bson_string(len, value) && sax->string(value);
+            }
+
+            case 0x03: // object
+            {
+                return parse_bson_internal();
+            }
+
+            case 0x04: // array
+            {
+                return parse_bson_array();
+            }
+
+            case 0x05: // binary
+            {
+                std::int32_t len{};
+                binary_t value;
+                return get_number<std::int32_t, true>(input_format_t::bson, len) && get_bson_binary(len, value) && sax->binary(value);
+            }
+
+            case 0x08: // boolean
+            {
+                return sax->boolean(get() != 0);
+            }
+
+            case 0x0A: // null
+            {
+                return sax->null();
+            }
+
+            case 0x10: // int32
+            {
+                std::int32_t value{};
+                return get_number<std::int32_t, true>(input_format_t::bson, value) && sax->number_integer(value);
+            }
+
+            case 0x12: // int64
+            {
+                std::int64_t value{};
+                return get_number<std::int64_t, true>(input_format_t::bson, value) && sax->number_integer(value);
+            }
+
+            default: // anything else not supported (yet)
+            {
+                std::array<char, 3> cr{{}};
+                static_cast<void>((std::snprintf)(cr.data(), cr.size(), "%.2hhX", static_cast<unsigned char>(element_type))); // NOLINT(cppcoreguidelines-pro-type-vararg,hicpp-vararg)
+                std::string cr_str{cr.data()};
+                return sax->parse_error(element_type_parse_position, cr_str,
+                                        parse_error::create(114, element_type_parse_position, concat("Unsupported BSON record type 0x", cr_str), nullptr));
+            }
+        }
+    }
+
+    /*!
+    @brief Read a BSON element list (as specified in the BSON-spec)
+
+    The same binary layout is used for objects and arrays, hence it must be
+    indicated with the argument @a is_array which one is expected
+    (true --> array, false --> object).
+
+    @param[in] is_array Determines if the element list being read is to be
+                        treated as an object (@a is_array == false), or as an
+                        array (@a is_array == true).
+    @return whether a valid BSON-object/array was passed to the SAX parser
+    */
+    bool parse_bson_element_list(const bool is_array)
+    {
+        string_t key;
+
+        while (auto element_type = get())
+        {
+            if (JSON_HEDLEY_UNLIKELY(!unexpect_eof(input_format_t::bson, "element list")))
+            {
+                return false;
+            }
+
+            const std::size_t element_type_parse_position = chars_read;
+            if (JSON_HEDLEY_UNLIKELY(!get_bson_cstr(key)))
+            {
+                return false;
+            }
+
+            if (!is_array && !sax->key(key))
+            {
+                return false;
+            }
+
+            if (JSON_HEDLEY_UNLIKELY(!parse_bson_element_internal(element_type, element_type_parse_position)))
+            {
+                return false;
+            }
+
+            // get_bson_cstr only appends
+            key.clear();
+        }
+
+        return true;
+    }
+
+    /*!
+    @brief Reads an array from the BSON input and passes it to the SAX-parser.
+    @return whether a valid BSON-array was passed to the SAX parser
+    */
+    bool parse_bson_array()
+    {
+        std::int32_t document_size{};
+        get_number<std::int32_t, true>(input_format_t::bson, document_size);
+
+        if (JSON_HEDLEY_UNLIKELY(!sax->start_array(static_cast<std::size_t>(-1))))
+        {
+            return false;
+        }
+
+        if (JSON_HEDLEY_UNLIKELY(!parse_bson_element_list(/*is_array*/true)))
+        {
+            return false;
+        }
+
+        return sax->end_array();
+    }
+
+    //////////
+    // CBOR //
+    //////////
+
+    /*!
+    @param[in] get_char  whether a new character should be retrieved from the
+                         input (true) or whether the last read character should
+                         be considered instead (false)
+    @param[in] tag_handler how CBOR tags should be treated
+
+    @return whether a valid CBOR value was passed to the SAX parser
+    */
+    bool parse_cbor_internal(const bool get_char,
+                             const cbor_tag_handler_t tag_handler)
+    {
+        switch (get_char ? get() : current)
+        {
+            // EOF
+            case std::char_traits<char_type>::eof():
+                return unexpect_eof(input_format_t::cbor, "value");
+
+            // Integer 0x00..0x17 (0..23)
+            case 0x00:
+            case 0x01:
+            case 0x02:
+            case 0x03:
+            case 0x04:
+            case 0x05:
+            case 0x06:
+            case 0x07:
+            case 0x08:
+            case 0x09:
+            case 0x0A:
+            case 0x0B:
+            case 0x0C:
+            case 0x0D:
+            case 0x0E:
+            case 0x0F:
+            case 0x10:
+            case 0x11:
+            case 0x12:
+            case 0x13:
+            case 0x14:
+            case 0x15:
+            case 0x16:
+            case 0x17:
+                return sax->number_unsigned(static_cast<number_unsigned_t>(current));
+
+            case 0x18: // Unsigned integer (one-byte uint8_t follows)
+            {
+                std::uint8_t number{};
+                return get_number(input_format_t::cbor, number) && sax->number_unsigned(number);
+            }
+
+            case 0x19: // Unsigned integer (two-byte uint16_t follows)
+            {
+                std::uint16_t number{};
+                return get_number(input_format_t::cbor, number) && sax->number_unsigned(number);
+            }
+
+            case 0x1A: // Unsigned integer (four-byte uint32_t follows)
+            {
+                std::uint32_t number{};
+                return get_number(input_format_t::cbor, number) && sax->number_unsigned(number);
+            }
+
+            case 0x1B: // Unsigned integer (eight-byte uint64_t follows)
+            {
+                std::uint64_t number{};
+                return get_number(input_format_t::cbor, number) && sax->number_unsigned(number);
+            }
+
+            // Negative integer -1-0x00..-1-0x17 (-1..-24)
+            case 0x20:
+            case 0x21:
+            case 0x22:
+            case 0x23:
+            case 0x24:
+            case 0x25:
+            case 0x26:
+            case 0x27:
+            case 0x28:
+            case 0x29:
+            case 0x2A:
+            case 0x2B:
+            case 0x2C:
+            case 0x2D:
+            case 0x2E:
+            case 0x2F:
+            case 0x30:
+            case 0x31:
+            case 0x32:
+            case 0x33:
+            case 0x34:
+            case 0x35:
+            case 0x36:
+            case 0x37:
+                return sax->number_integer(static_cast<std::int8_t>(0x20 - 1 - current));
+
+            case 0x38: // Negative integer (one-byte uint8_t follows)
+            {
+                std::uint8_t number{};
+                return get_number(input_format_t::cbor, number) && sax->number_integer(static_cast<number_integer_t>(-1) - number);
+            }
+
+            case 0x39: // Negative integer -1-n (two-byte uint16_t follows)
+            {
+                std::uint16_t number{};
+                return get_number(input_format_t::cbor, number) && sax->number_integer(static_cast<number_integer_t>(-1) - number);
+            }
+
+            case 0x3A: // Negative integer -1-n (four-byte uint32_t follows)
+            {
+                std::uint32_t number{};
+                return get_number(input_format_t::cbor, number) && sax->number_integer(static_cast<number_integer_t>(-1) - number);
+            }
+
+            case 0x3B: // Negative integer -1-n (eight-byte uint64_t follows)
+            {
+                std::uint64_t number{};
+                return get_number(input_format_t::cbor, number) && sax->number_integer(static_cast<number_integer_t>(-1)
+                        - static_cast<number_integer_t>(number));
+            }
+
+            // Binary data (0x00..0x17 bytes follow)
+            case 0x40:
+            case 0x41:
+            case 0x42:
+            case 0x43:
+            case 0x44:
+            case 0x45:
+            case 0x46:
+            case 0x47:
+            case 0x48:
+            case 0x49:
+            case 0x4A:
+            case 0x4B:
+            case 0x4C:
+            case 0x4D:
+            case 0x4E:
+            case 0x4F:
+            case 0x50:
+            case 0x51:
+            case 0x52:
+            case 0x53:
+            case 0x54:
+            case 0x55:
+            case 0x56:
+            case 0x57:
+            case 0x58: // Binary data (one-byte uint8_t for n follows)
+            case 0x59: // Binary data (two-byte uint16_t for n follow)
+            case 0x5A: // Binary data (four-byte uint32_t for n follow)
+            case 0x5B: // Binary data (eight-byte uint64_t for n follow)
+            case 0x5F: // Binary data (indefinite length)
+            {
+                binary_t b;
+                return get_cbor_binary(b) && sax->binary(b);
+            }
+
+            // UTF-8 string (0x00..0x17 bytes follow)
+            case 0x60:
+            case 0x61:
+            case 0x62:
+            case 0x63:
+            case 0x64:
+            case 0x65:
+            case 0x66:
+            case 0x67:
+            case 0x68:
+            case 0x69:
+            case 0x6A:
+            case 0x6B:
+            case 0x6C:
+            case 0x6D:
+            case 0x6E:
+            case 0x6F:
+            case 0x70:
+            case 0x71:
+            case 0x72:
+            case 0x73:
+            case 0x74:
+            case 0x75:
+            case 0x76:
+            case 0x77:
+            case 0x78: // UTF-8 string (one-byte uint8_t for n follows)
+            case 0x79: // UTF-8 string (two-byte uint16_t for n follow)
+            case 0x7A: // UTF-8 string (four-byte uint32_t for n follow)
+            case 0x7B: // UTF-8 string (eight-byte uint64_t for n follow)
+            case 0x7F: // UTF-8 string (indefinite length)
+            {
+                string_t s;
+                return get_cbor_string(s) && sax->string(s);
+            }
+
+            // array (0x00..0x17 data items follow)
+            case 0x80:
+            case 0x81:
+            case 0x82:
+            case 0x83:
+            case 0x84:
+            case 0x85:
+            case 0x86:
+            case 0x87:
+            case 0x88:
+            case 0x89:
+            case 0x8A:
+            case 0x8B:
+            case 0x8C:
+            case 0x8D:
+            case 0x8E:
+            case 0x8F:
+            case 0x90:
+            case 0x91:
+            case 0x92:
+            case 0x93:
+            case 0x94:
+            case 0x95:
+            case 0x96:
+            case 0x97:
+                return get_cbor_array(
+                           conditional_static_cast<std::size_t>(static_cast<unsigned int>(current) & 0x1Fu), tag_handler);
+
+            case 0x98: // array (one-byte uint8_t for n follows)
+            {
+                std::uint8_t len{};
+                return get_number(input_format_t::cbor, len) && get_cbor_array(static_cast<std::size_t>(len), tag_handler);
+            }
+
+            case 0x99: // array (two-byte uint16_t for n follow)
+            {
+                std::uint16_t len{};
+                return get_number(input_format_t::cbor, len) && get_cbor_array(static_cast<std::size_t>(len), tag_handler);
+            }
+
+            case 0x9A: // array (four-byte uint32_t for n follow)
+            {
+                std::uint32_t len{};
+                return get_number(input_format_t::cbor, len) && get_cbor_array(conditional_static_cast<std::size_t>(len), tag_handler);
+            }
+
+            case 0x9B: // array (eight-byte uint64_t for n follow)
+            {
+                std::uint64_t len{};
+                return get_number(input_format_t::cbor, len) && get_cbor_array(conditional_static_cast<std::size_t>(len), tag_handler);
+            }
+
+            case 0x9F: // array (indefinite length)
+                return get_cbor_array(static_cast<std::size_t>(-1), tag_handler);
+
+            // map (0x00..0x17 pairs of data items follow)
+            case 0xA0:
+            case 0xA1:
+            case 0xA2:
+            case 0xA3:
+            case 0xA4:
+            case 0xA5:
+            case 0xA6:
+            case 0xA7:
+            case 0xA8:
+            case 0xA9:
+            case 0xAA:
+            case 0xAB:
+            case 0xAC:
+            case 0xAD:
+            case 0xAE:
+            case 0xAF:
+            case 0xB0:
+            case 0xB1:
+            case 0xB2:
+            case 0xB3:
+            case 0xB4:
+            case 0xB5:
+            case 0xB6:
+            case 0xB7:
+                return get_cbor_object(conditional_static_cast<std::size_t>(static_cast<unsigned int>(current) & 0x1Fu), tag_handler);
+
+            case 0xB8: // map (one-byte uint8_t for n follows)
+            {
+                std::uint8_t len{};
+                return get_number(input_format_t::cbor, len) && get_cbor_object(static_cast<std::size_t>(len), tag_handler);
+            }
+
+            case 0xB9: // map (two-byte uint16_t for n follow)
+            {
+                std::uint16_t len{};
+                return get_number(input_format_t::cbor, len) && get_cbor_object(static_cast<std::size_t>(len), tag_handler);
+            }
+
+            case 0xBA: // map (four-byte uint32_t for n follow)
+            {
+                std::uint32_t len{};
+                return get_number(input_format_t::cbor, len) && get_cbor_object(conditional_static_cast<std::size_t>(len), tag_handler);
+            }
+
+            case 0xBB: // map (eight-byte uint64_t for n follow)
+            {
+                std::uint64_t len{};
+                return get_number(input_format_t::cbor, len) && get_cbor_object(conditional_static_cast<std::size_t>(len), tag_handler);
+            }
+
+            case 0xBF: // map (indefinite length)
+                return get_cbor_object(static_cast<std::size_t>(-1), tag_handler);
+
+            case 0xC6: // tagged item
+            case 0xC7:
+            case 0xC8:
+            case 0xC9:
+            case 0xCA:
+            case 0xCB:
+            case 0xCC:
+            case 0xCD:
+            case 0xCE:
+            case 0xCF:
+            case 0xD0:
+            case 0xD1:
+            case 0xD2:
+            case 0xD3:
+            case 0xD4:
+            case 0xD8: // tagged item (1 bytes follow)
+            case 0xD9: // tagged item (2 bytes follow)
+            case 0xDA: // tagged item (4 bytes follow)
+            case 0xDB: // tagged item (8 bytes follow)
+            {
+                switch (tag_handler)
+                {
+                    case cbor_tag_handler_t::error:
+                    {
+                        auto last_token = get_token_string();
+                        return sax->parse_error(chars_read, last_token, parse_error::create(112, chars_read,
+                                                exception_message(input_format_t::cbor, concat("invalid byte: 0x", last_token), "value"), nullptr));
+                    }
+
+                    case cbor_tag_handler_t::ignore:
+                    {
+                        // ignore binary subtype
+                        switch (current)
+                        {
+                            case 0xD8:
+                            {
+                                std::uint8_t subtype_to_ignore{};
+                                get_number(input_format_t::cbor, subtype_to_ignore);
+                                break;
+                            }
+                            case 0xD9:
+                            {
+                                std::uint16_t subtype_to_ignore{};
+                                get_number(input_format_t::cbor, subtype_to_ignore);
+                                break;
+                            }
+                            case 0xDA:
+                            {
+                                std::uint32_t subtype_to_ignore{};
+                                get_number(input_format_t::cbor, subtype_to_ignore);
+                                break;
+                            }
+                            case 0xDB:
+                            {
+                                std::uint64_t subtype_to_ignore{};
+                                get_number(input_format_t::cbor, subtype_to_ignore);
+                                break;
+                            }
+                            default:
+                                break;
+                        }
+                        return parse_cbor_internal(true, tag_handler);
+                    }
+
+                    case cbor_tag_handler_t::store:
+                    {
+                        binary_t b;
+                        // use binary subtype and store in binary container
+                        switch (current)
+                        {
+                            case 0xD8:
+                            {
+                                std::uint8_t subtype{};
+                                get_number(input_format_t::cbor, subtype);
+                                b.set_subtype(detail::conditional_static_cast<typename binary_t::subtype_type>(subtype));
+                                break;
+                            }
+                            case 0xD9:
+                            {
+                                std::uint16_t subtype{};
+                                get_number(input_format_t::cbor, subtype);
+                                b.set_subtype(detail::conditional_static_cast<typename binary_t::subtype_type>(subtype));
+                                break;
+                            }
+                            case 0xDA:
+                            {
+                                std::uint32_t subtype{};
+                                get_number(input_format_t::cbor, subtype);
+                                b.set_subtype(detail::conditional_static_cast<typename binary_t::subtype_type>(subtype));
+                                break;
+                            }
+                            case 0xDB:
+                            {
+                                std::uint64_t subtype{};
+                                get_number(input_format_t::cbor, subtype);
+                                b.set_subtype(detail::conditional_static_cast<typename binary_t::subtype_type>(subtype));
+                                break;
+                            }
+                            default:
+                                return parse_cbor_internal(true, tag_handler);
+                        }
+                        get();
+                        return get_cbor_binary(b) && sax->binary(b);
+                    }
+
+                    default:                 // LCOV_EXCL_LINE
+                        JSON_ASSERT(false); // NOLINT(cert-dcl03-c,hicpp-static-assert,misc-static-assert) LCOV_EXCL_LINE
+                        return false;        // LCOV_EXCL_LINE
+                }
+            }
+
+            case 0xF4: // false
+                return sax->boolean(false);
+
+            case 0xF5: // true
+                return sax->boolean(true);
+
+            case 0xF6: // null
+                return sax->null();
+
+            case 0xF9: // Half-Precision Float (two-byte IEEE 754)
+            {
+                const auto byte1_raw = get();
+                if (JSON_HEDLEY_UNLIKELY(!unexpect_eof(input_format_t::cbor, "number")))
+                {
+                    return false;
+                }
+                const auto byte2_raw = get();
+                if (JSON_HEDLEY_UNLIKELY(!unexpect_eof(input_format_t::cbor, "number")))
+                {
+                    return false;
+                }
+
+                const auto byte1 = static_cast<unsigned char>(byte1_raw);
+                const auto byte2 = static_cast<unsigned char>(byte2_raw);
+
+                // code from RFC 7049, Appendix D, Figure 3:
+                // As half-precision floating-point numbers were only added
+                // to IEEE 754 in 2008, today's programming platforms often
+                // still only have limited support for them. It is very
+                // easy to include at least decoding support for them even
+                // without such support. An example of a small decoder for
+                // half-precision floating-point numbers in the C language
+                // is shown in Fig. 3.
+                const auto half = static_cast<unsigned int>((byte1 << 8u) + byte2);
+                const double val = [&half]
+                {
+                    const int exp = (half >> 10u) & 0x1Fu;
+                    const unsigned int mant = half & 0x3FFu;
+                    JSON_ASSERT(0 <= exp&& exp <= 32);
+                    JSON_ASSERT(mant <= 1024);
+                    switch (exp)
+                    {
+                        case 0:
+                            return std::ldexp(mant, -24);
+                        case 31:
+                            return (mant == 0)
+                            ? std::numeric_limits<double>::infinity()
+                            : std::numeric_limits<double>::quiet_NaN();
+                        default:
+                            return std::ldexp(mant + 1024, exp - 25);
+                    }
+                }();
+                return sax->number_float((half & 0x8000u) != 0
+                                         ? static_cast<number_float_t>(-val)
+                                         : static_cast<number_float_t>(val), "");
+            }
+
+            case 0xFA: // Single-Precision Float (four-byte IEEE 754)
+            {
+                float number{};
+                return get_number(input_format_t::cbor, number) && sax->number_float(static_cast<number_float_t>(number), "");
+            }
+
+            case 0xFB: // Double-Precision Float (eight-byte IEEE 754)
+            {
+                double number{};
+                return get_number(input_format_t::cbor, number) && sax->number_float(static_cast<number_float_t>(number), "");
+            }
+
+            default: // anything else (0xFF is handled inside the other types)
+            {
+                auto last_token = get_token_string();
+                return sax->parse_error(chars_read, last_token, parse_error::create(112, chars_read,
+                                        exception_message(input_format_t::cbor, concat("invalid byte: 0x", last_token), "value"), nullptr));
+            }
+        }
+    }
+
+    /*!
+    @brief reads a CBOR string
+
+    This function first reads starting bytes to determine the expected
+    string length and then copies this number of bytes into a string.
+    Additionally, CBOR's strings with indefinite lengths are supported.
+
+    @param[out] result  created string
+
+    @return whether string creation completed
+    */
+    bool get_cbor_string(string_t& result)
+    {
+        if (JSON_HEDLEY_UNLIKELY(!unexpect_eof(input_format_t::cbor, "string")))
+        {
+            return false;
+        }
+
+        switch (current)
+        {
+            // UTF-8 string (0x00..0x17 bytes follow)
+            case 0x60:
+            case 0x61:
+            case 0x62:
+            case 0x63:
+            case 0x64:
+            case 0x65:
+            case 0x66:
+            case 0x67:
+            case 0x68:
+            case 0x69:
+            case 0x6A:
+            case 0x6B:
+            case 0x6C:
+            case 0x6D:
+            case 0x6E:
+            case 0x6F:
+            case 0x70:
+            case 0x71:
+            case 0x72:
+            case 0x73:
+            case 0x74:
+            case 0x75:
+            case 0x76:
+            case 0x77:
+            {
+                return get_string(input_format_t::cbor, static_cast<unsigned int>(current) & 0x1Fu, result);
+            }
+
+            case 0x78: // UTF-8 string (one-byte uint8_t for n follows)
+            {
+                std::uint8_t len{};
+                return get_number(input_format_t::cbor, len) && get_string(input_format_t::cbor, len, result);
+            }
+
+            case 0x79: // UTF-8 string (two-byte uint16_t for n follow)
+            {
+                std::uint16_t len{};
+                return get_number(input_format_t::cbor, len) && get_string(input_format_t::cbor, len, result);
+            }
+
+            case 0x7A: // UTF-8 string (four-byte uint32_t for n follow)
+            {
+                std::uint32_t len{};
+                return get_number(input_format_t::cbor, len) && get_string(input_format_t::cbor, len, result);
+            }
+
+            case 0x7B: // UTF-8 string (eight-byte uint64_t for n follow)
+            {
+                std::uint64_t len{};
+                return get_number(input_format_t::cbor, len) && get_string(input_format_t::cbor, len, result);
+            }
+
+            case 0x7F: // UTF-8 string (indefinite length)
+            {
+                while (get() != 0xFF)
+                {
+                    string_t chunk;
+                    if (!get_cbor_string(chunk))
+                    {
+                        return false;
+                    }
+                    result.append(chunk);
+                }
+                return true;
+            }
+
+            default:
+            {
+                auto last_token = get_token_string();
+                return sax->parse_error(chars_read, last_token, parse_error::create(113, chars_read,
+                                        exception_message(input_format_t::cbor, concat("expected length specification (0x60-0x7B) or indefinite string type (0x7F); last byte: 0x", last_token), "string"), nullptr));
+            }
+        }
+    }
+
+    /*!
+    @brief reads a CBOR byte array
+
+    This function first reads starting bytes to determine the expected
+    byte array length and then copies this number of bytes into the byte array.
+    Additionally, CBOR's byte arrays with indefinite lengths are supported.
+
+    @param[out] result  created byte array
+
+    @return whether byte array creation completed
+    */
+    bool get_cbor_binary(binary_t& result)
+    {
+        if (JSON_HEDLEY_UNLIKELY(!unexpect_eof(input_format_t::cbor, "binary")))
+        {
+            return false;
+        }
+
+        switch (current)
+        {
+            // Binary data (0x00..0x17 bytes follow)
+            case 0x40:
+            case 0x41:
+            case 0x42:
+            case 0x43:
+            case 0x44:
+            case 0x45:
+            case 0x46:
+            case 0x47:
+            case 0x48:
+            case 0x49:
+            case 0x4A:
+            case 0x4B:
+            case 0x4C:
+            case 0x4D:
+            case 0x4E:
+            case 0x4F:
+            case 0x50:
+            case 0x51:
+            case 0x52:
+            case 0x53:
+            case 0x54:
+            case 0x55:
+            case 0x56:
+            case 0x57:
+            {
+                return get_binary(input_format_t::cbor, static_cast<unsigned int>(current) & 0x1Fu, result);
+            }
+
+            case 0x58: // Binary data (one-byte uint8_t for n follows)
+            {
+                std::uint8_t len{};
+                return get_number(input_format_t::cbor, len) &&
+                       get_binary(input_format_t::cbor, len, result);
+            }
+
+            case 0x59: // Binary data (two-byte uint16_t for n follow)
+            {
+                std::uint16_t len{};
+                return get_number(input_format_t::cbor, len) &&
+                       get_binary(input_format_t::cbor, len, result);
+            }
+
+            case 0x5A: // Binary data (four-byte uint32_t for n follow)
+            {
+                std::uint32_t len{};
+                return get_number(input_format_t::cbor, len) &&
+                       get_binary(input_format_t::cbor, len, result);
+            }
+
+            case 0x5B: // Binary data (eight-byte uint64_t for n follow)
+            {
+                std::uint64_t len{};
+                return get_number(input_format_t::cbor, len) &&
+                       get_binary(input_format_t::cbor, len, result);
+            }
+
+            case 0x5F: // Binary data (indefinite length)
+            {
+                while (get() != 0xFF)
+                {
+                    binary_t chunk;
+                    if (!get_cbor_binary(chunk))
+                    {
+                        return false;
+                    }
+                    result.insert(result.end(), chunk.begin(), chunk.end());
+                }
+                return true;
+            }
+
+            default:
+            {
+                auto last_token = get_token_string();
+                return sax->parse_error(chars_read, last_token, parse_error::create(113, chars_read,
+                                        exception_message(input_format_t::cbor, concat("expected length specification (0x40-0x5B) or indefinite binary array type (0x5F); last byte: 0x", last_token), "binary"), nullptr));
+            }
+        }
+    }
+
+    /*!
+    @param[in] len  the length of the array or static_cast<std::size_t>(-1) for an
+                    array of indefinite size
+    @param[in] tag_handler how CBOR tags should be treated
+    @return whether array creation completed
+    */
+    bool get_cbor_array(const std::size_t len,
+                        const cbor_tag_handler_t tag_handler)
+    {
+        if (JSON_HEDLEY_UNLIKELY(!sax->start_array(len)))
+        {
+            return false;
+        }
+
+        if (len != static_cast<std::size_t>(-1))
+        {
+            for (std::size_t i = 0; i < len; ++i)
+            {
+                if (JSON_HEDLEY_UNLIKELY(!parse_cbor_internal(true, tag_handler)))
+                {
+                    return false;
+                }
+            }
+        }
+        else
+        {
+            while (get() != 0xFF)
+            {
+                if (JSON_HEDLEY_UNLIKELY(!parse_cbor_internal(false, tag_handler)))
+                {
+                    return false;
+                }
+            }
+        }
+
+        return sax->end_array();
+    }
+
+    /*!
+    @param[in] len  the length of the object or static_cast<std::size_t>(-1) for an
+                    object of indefinite size
+    @param[in] tag_handler how CBOR tags should be treated
+    @return whether object creation completed
+    */
+    bool get_cbor_object(const std::size_t len,
+                         const cbor_tag_handler_t tag_handler)
+    {
+        if (JSON_HEDLEY_UNLIKELY(!sax->start_object(len)))
+        {
+            return false;
+        }
+
+        if (len != 0)
+        {
+            string_t key;
+            if (len != static_cast<std::size_t>(-1))
+            {
+                for (std::size_t i = 0; i < len; ++i)
+                {
+                    get();
+                    if (JSON_HEDLEY_UNLIKELY(!get_cbor_string(key) || !sax->key(key)))
+                    {
+                        return false;
+                    }
+
+                    if (JSON_HEDLEY_UNLIKELY(!parse_cbor_internal(true, tag_handler)))
+                    {
+                        return false;
+                    }
+                    key.clear();
+                }
+            }
+            else
+            {
+                while (get() != 0xFF)
+                {
+                    if (JSON_HEDLEY_UNLIKELY(!get_cbor_string(key) || !sax->key(key)))
+                    {
+                        return false;
+                    }
+
+                    if (JSON_HEDLEY_UNLIKELY(!parse_cbor_internal(true, tag_handler)))
+                    {
+                        return false;
+                    }
+                    key.clear();
+                }
+            }
+        }
+
+        return sax->end_object();
+    }
+
+    /////////////
+    // MsgPack //
+    /////////////
+
+    /*!
+    @return whether a valid MessagePack value was passed to the SAX parser
+    */
+    bool parse_msgpack_internal()
+    {
+        switch (get())
+        {
+            // EOF
+            case std::char_traits<char_type>::eof():
+                return unexpect_eof(input_format_t::msgpack, "value");
+
+            // positive fixint
+            case 0x00:
+            case 0x01:
+            case 0x02:
+            case 0x03:
+            case 0x04:
+            case 0x05:
+            case 0x06:
+            case 0x07:
+            case 0x08:
+            case 0x09:
+            case 0x0A:
+            case 0x0B:
+            case 0x0C:
+            case 0x0D:
+            case 0x0E:
+            case 0x0F:
+            case 0x10:
+            case 0x11:
+            case 0x12:
+            case 0x13:
+            case 0x14:
+            case 0x15:
+            case 0x16:
+            case 0x17:
+            case 0x18:
+            case 0x19:
+            case 0x1A:
+            case 0x1B:
+            case 0x1C:
+            case 0x1D:
+            case 0x1E:
+            case 0x1F:
+            case 0x20:
+            case 0x21:
+            case 0x22:
+            case 0x23:
+            case 0x24:
+            case 0x25:
+            case 0x26:
+            case 0x27:
+            case 0x28:
+            case 0x29:
+            case 0x2A:
+            case 0x2B:
+            case 0x2C:
+            case 0x2D:
+            case 0x2E:
+            case 0x2F:
+            case 0x30:
+            case 0x31:
+            case 0x32:
+            case 0x33:
+            case 0x34:
+            case 0x35:
+            case 0x36:
+            case 0x37:
+            case 0x38:
+            case 0x39:
+            case 0x3A:
+            case 0x3B:
+            case 0x3C:
+            case 0x3D:
+            case 0x3E:
+            case 0x3F:
+            case 0x40:
+            case 0x41:
+            case 0x42:
+            case 0x43:
+            case 0x44:
+            case 0x45:
+            case 0x46:
+            case 0x47:
+            case 0x48:
+            case 0x49:
+            case 0x4A:
+            case 0x4B:
+            case 0x4C:
+            case 0x4D:
+            case 0x4E:
+            case 0x4F:
+            case 0x50:
+            case 0x51:
+            case 0x52:
+            case 0x53:
+            case 0x54:
+            case 0x55:
+            case 0x56:
+            case 0x57:
+            case 0x58:
+            case 0x59:
+            case 0x5A:
+            case 0x5B:
+            case 0x5C:
+            case 0x5D:
+            case 0x5E:
+            case 0x5F:
+            case 0x60:
+            case 0x61:
+            case 0x62:
+            case 0x63:
+            case 0x64:
+            case 0x65:
+            case 0x66:
+            case 0x67:
+            case 0x68:
+            case 0x69:
+            case 0x6A:
+            case 0x6B:
+            case 0x6C:
+            case 0x6D:
+            case 0x6E:
+            case 0x6F:
+            case 0x70:
+            case 0x71:
+            case 0x72:
+            case 0x73:
+            case 0x74:
+            case 0x75:
+            case 0x76:
+            case 0x77:
+            case 0x78:
+            case 0x79:
+            case 0x7A:
+            case 0x7B:
+            case 0x7C:
+            case 0x7D:
+            case 0x7E:
+            case 0x7F:
+                return sax->number_unsigned(static_cast<number_unsigned_t>(current));
+
+            // fixmap
+            case 0x80:
+            case 0x81:
+            case 0x82:
+            case 0x83:
+            case 0x84:
+            case 0x85:
+            case 0x86:
+            case 0x87:
+            case 0x88:
+            case 0x89:
+            case 0x8A:
+            case 0x8B:
+            case 0x8C:
+            case 0x8D:
+            case 0x8E:
+            case 0x8F:
+                return get_msgpack_object(conditional_static_cast<std::size_t>(static_cast<unsigned int>(current) & 0x0Fu));
+
+            // fixarray
+            case 0x90:
+            case 0x91:
+            case 0x92:
+            case 0x93:
+            case 0x94:
+            case 0x95:
+            case 0x96:
+            case 0x97:
+            case 0x98:
+            case 0x99:
+            case 0x9A:
+            case 0x9B:
+            case 0x9C:
+            case 0x9D:
+            case 0x9E:
+            case 0x9F:
+                return get_msgpack_array(conditional_static_cast<std::size_t>(static_cast<unsigned int>(current) & 0x0Fu));
+
+            // fixstr
+            case 0xA0:
+            case 0xA1:
+            case 0xA2:
+            case 0xA3:
+            case 0xA4:
+            case 0xA5:
+            case 0xA6:
+            case 0xA7:
+            case 0xA8:
+            case 0xA9:
+            case 0xAA:
+            case 0xAB:
+            case 0xAC:
+            case 0xAD:
+            case 0xAE:
+            case 0xAF:
+            case 0xB0:
+            case 0xB1:
+            case 0xB2:
+            case 0xB3:
+            case 0xB4:
+            case 0xB5:
+            case 0xB6:
+            case 0xB7:
+            case 0xB8:
+            case 0xB9:
+            case 0xBA:
+            case 0xBB:
+            case 0xBC:
+            case 0xBD:
+            case 0xBE:
+            case 0xBF:
+            case 0xD9: // str 8
+            case 0xDA: // str 16
+            case 0xDB: // str 32
+            {
+                string_t s;
+                return get_msgpack_string(s) && sax->string(s);
+            }
+
+            case 0xC0: // nil
+                return sax->null();
+
+            case 0xC2: // false
+                return sax->boolean(false);
+
+            case 0xC3: // true
+                return sax->boolean(true);
+
+            case 0xC4: // bin 8
+            case 0xC5: // bin 16
+            case 0xC6: // bin 32
+            case 0xC7: // ext 8
+            case 0xC8: // ext 16
+            case 0xC9: // ext 32
+            case 0xD4: // fixext 1
+            case 0xD5: // fixext 2
+            case 0xD6: // fixext 4
+            case 0xD7: // fixext 8
+            case 0xD8: // fixext 16
+            {
+                binary_t b;
+                return get_msgpack_binary(b) && sax->binary(b);
+            }
+
+            case 0xCA: // float 32
+            {
+                float number{};
+                return get_number(input_format_t::msgpack, number) && sax->number_float(static_cast<number_float_t>(number), "");
+            }
+
+            case 0xCB: // float 64
+            {
+                double number{};
+                return get_number(input_format_t::msgpack, number) && sax->number_float(static_cast<number_float_t>(number), "");
+            }
+
+            case 0xCC: // uint 8
+            {
+                std::uint8_t number{};
+                return get_number(input_format_t::msgpack, number) && sax->number_unsigned(number);
+            }
+
+            case 0xCD: // uint 16
+            {
+                std::uint16_t number{};
+                return get_number(input_format_t::msgpack, number) && sax->number_unsigned(number);
+            }
+
+            case 0xCE: // uint 32
+            {
+                std::uint32_t number{};
+                return get_number(input_format_t::msgpack, number) && sax->number_unsigned(number);
+            }
+
+            case 0xCF: // uint 64
+            {
+                std::uint64_t number{};
+                return get_number(input_format_t::msgpack, number) && sax->number_unsigned(number);
+            }
+
+            case 0xD0: // int 8
+            {
+                std::int8_t number{};
+                return get_number(input_format_t::msgpack, number) && sax->number_integer(number);
+            }
+
+            case 0xD1: // int 16
+            {
+                std::int16_t number{};
+                return get_number(input_format_t::msgpack, number) && sax->number_integer(number);
+            }
+
+            case 0xD2: // int 32
+            {
+                std::int32_t number{};
+                return get_number(input_format_t::msgpack, number) && sax->number_integer(number);
+            }
+
+            case 0xD3: // int 64
+            {
+                std::int64_t number{};
+                return get_number(input_format_t::msgpack, number) && sax->number_integer(number);
+            }
+
+            case 0xDC: // array 16
+            {
+                std::uint16_t len{};
+                return get_number(input_format_t::msgpack, len) && get_msgpack_array(static_cast<std::size_t>(len));
+            }
+
+            case 0xDD: // array 32
+            {
+                std::uint32_t len{};
+                return get_number(input_format_t::msgpack, len) && get_msgpack_array(conditional_static_cast<std::size_t>(len));
+            }
+
+            case 0xDE: // map 16
+            {
+                std::uint16_t len{};
+                return get_number(input_format_t::msgpack, len) && get_msgpack_object(static_cast<std::size_t>(len));
+            }
+
+            case 0xDF: // map 32
+            {
+                std::uint32_t len{};
+                return get_number(input_format_t::msgpack, len) && get_msgpack_object(conditional_static_cast<std::size_t>(len));
+            }
+
+            // negative fixint
+            case 0xE0:
+            case 0xE1:
+            case 0xE2:
+            case 0xE3:
+            case 0xE4:
+            case 0xE5:
+            case 0xE6:
+            case 0xE7:
+            case 0xE8:
+            case 0xE9:
+            case 0xEA:
+            case 0xEB:
+            case 0xEC:
+            case 0xED:
+            case 0xEE:
+            case 0xEF:
+            case 0xF0:
+            case 0xF1:
+            case 0xF2:
+            case 0xF3:
+            case 0xF4:
+            case 0xF5:
+            case 0xF6:
+            case 0xF7:
+            case 0xF8:
+            case 0xF9:
+            case 0xFA:
+            case 0xFB:
+            case 0xFC:
+            case 0xFD:
+            case 0xFE:
+            case 0xFF:
+                return sax->number_integer(static_cast<std::int8_t>(current));
+
+            default: // anything else
+            {
+                auto last_token = get_token_string();
+                return sax->parse_error(chars_read, last_token, parse_error::create(112, chars_read,
+                                        exception_message(input_format_t::msgpack, concat("invalid byte: 0x", last_token), "value"), nullptr));
+            }
+        }
+    }
+
+    /*!
+    @brief reads a MessagePack string
+
+    This function first reads starting bytes to determine the expected
+    string length and then copies this number of bytes into a string.
+
+    @param[out] result  created string
+
+    @return whether string creation completed
+    */
+    bool get_msgpack_string(string_t& result)
+    {
+        if (JSON_HEDLEY_UNLIKELY(!unexpect_eof(input_format_t::msgpack, "string")))
+        {
+            return false;
+        }
+
+        switch (current)
+        {
+            // fixstr
+            case 0xA0:
+            case 0xA1:
+            case 0xA2:
+            case 0xA3:
+            case 0xA4:
+            case 0xA5:
+            case 0xA6:
+            case 0xA7:
+            case 0xA8:
+            case 0xA9:
+            case 0xAA:
+            case 0xAB:
+            case 0xAC:
+            case 0xAD:
+            case 0xAE:
+            case 0xAF:
+            case 0xB0:
+            case 0xB1:
+            case 0xB2:
+            case 0xB3:
+            case 0xB4:
+            case 0xB5:
+            case 0xB6:
+            case 0xB7:
+            case 0xB8:
+            case 0xB9:
+            case 0xBA:
+            case 0xBB:
+            case 0xBC:
+            case 0xBD:
+            case 0xBE:
+            case 0xBF:
+            {
+                return get_string(input_format_t::msgpack, static_cast<unsigned int>(current) & 0x1Fu, result);
+            }
+
+            case 0xD9: // str 8
+            {
+                std::uint8_t len{};
+                return get_number(input_format_t::msgpack, len) && get_string(input_format_t::msgpack, len, result);
+            }
+
+            case 0xDA: // str 16
+            {
+                std::uint16_t len{};
+                return get_number(input_format_t::msgpack, len) && get_string(input_format_t::msgpack, len, result);
+            }
+
+            case 0xDB: // str 32
+            {
+                std::uint32_t len{};
+                return get_number(input_format_t::msgpack, len) && get_string(input_format_t::msgpack, len, result);
+            }
+
+            default:
+            {
+                auto last_token = get_token_string();
+                return sax->parse_error(chars_read, last_token, parse_error::create(113, chars_read,
+                                        exception_message(input_format_t::msgpack, concat("expected length specification (0xA0-0xBF, 0xD9-0xDB); last byte: 0x", last_token), "string"), nullptr));
+            }
+        }
+    }
+
+    /*!
+    @brief reads a MessagePack byte array
+
+    This function first reads starting bytes to determine the expected
+    byte array length and then copies this number of bytes into a byte array.
+
+    @param[out] result  created byte array
+
+    @return whether byte array creation completed
+    */
+    bool get_msgpack_binary(binary_t& result)
+    {
+        // helper function to set the subtype
+        auto assign_and_return_true = [&result](std::int8_t subtype)
+        {
+            result.set_subtype(static_cast<std::uint8_t>(subtype));
+            return true;
+        };
+
+        switch (current)
+        {
+            case 0xC4: // bin 8
+            {
+                std::uint8_t len{};
+                return get_number(input_format_t::msgpack, len) &&
+                       get_binary(input_format_t::msgpack, len, result);
+            }
+
+            case 0xC5: // bin 16
+            {
+                std::uint16_t len{};
+                return get_number(input_format_t::msgpack, len) &&
+                       get_binary(input_format_t::msgpack, len, result);
+            }
+
+            case 0xC6: // bin 32
+            {
+                std::uint32_t len{};
+                return get_number(input_format_t::msgpack, len) &&
+                       get_binary(input_format_t::msgpack, len, result);
+            }
+
+            case 0xC7: // ext 8
+            {
+                std::uint8_t len{};
+                std::int8_t subtype{};
+                return get_number(input_format_t::msgpack, len) &&
+                       get_number(input_format_t::msgpack, subtype) &&
+                       get_binary(input_format_t::msgpack, len, result) &&
+                       assign_and_return_true(subtype);
+            }
+
+            case 0xC8: // ext 16
+            {
+                std::uint16_t len{};
+                std::int8_t subtype{};
+                return get_number(input_format_t::msgpack, len) &&
+                       get_number(input_format_t::msgpack, subtype) &&
+                       get_binary(input_format_t::msgpack, len, result) &&
+                       assign_and_return_true(subtype);
+            }
+
+            case 0xC9: // ext 32
+            {
+                std::uint32_t len{};
+                std::int8_t subtype{};
+                return get_number(input_format_t::msgpack, len) &&
+                       get_number(input_format_t::msgpack, subtype) &&
+                       get_binary(input_format_t::msgpack, len, result) &&
+                       assign_and_return_true(subtype);
+            }
+
+            case 0xD4: // fixext 1
+            {
+                std::int8_t subtype{};
+                return get_number(input_format_t::msgpack, subtype) &&
+                       get_binary(input_format_t::msgpack, 1, result) &&
+                       assign_and_return_true(subtype);
+            }
+
+            case 0xD5: // fixext 2
+            {
+                std::int8_t subtype{};
+                return get_number(input_format_t::msgpack, subtype) &&
+                       get_binary(input_format_t::msgpack, 2, result) &&
+                       assign_and_return_true(subtype);
+            }
+
+            case 0xD6: // fixext 4
+            {
+                std::int8_t subtype{};
+                return get_number(input_format_t::msgpack, subtype) &&
+                       get_binary(input_format_t::msgpack, 4, result) &&
+                       assign_and_return_true(subtype);
+            }
+
+            case 0xD7: // fixext 8
+            {
+                std::int8_t subtype{};
+                return get_number(input_format_t::msgpack, subtype) &&
+                       get_binary(input_format_t::msgpack, 8, result) &&
+                       assign_and_return_true(subtype);
+            }
+
+            case 0xD8: // fixext 16
+            {
+                std::int8_t subtype{};
+                return get_number(input_format_t::msgpack, subtype) &&
+                       get_binary(input_format_t::msgpack, 16, result) &&
+                       assign_and_return_true(subtype);
+            }
+
+            default:           // LCOV_EXCL_LINE
+                return false;  // LCOV_EXCL_LINE
+        }
+    }
+
+    /*!
+    @param[in] len  the length of the array
+    @return whether array creation completed
+    */
+    bool get_msgpack_array(const std::size_t len)
+    {
+        if (JSON_HEDLEY_UNLIKELY(!sax->start_array(len)))
+        {
+            return false;
+        }
+
+        for (std::size_t i = 0; i < len; ++i)
+        {
+            if (JSON_HEDLEY_UNLIKELY(!parse_msgpack_internal()))
+            {
+                return false;
+            }
+        }
+
+        return sax->end_array();
+    }
+
+    /*!
+    @param[in] len  the length of the object
+    @return whether object creation completed
+    */
+    bool get_msgpack_object(const std::size_t len)
+    {
+        if (JSON_HEDLEY_UNLIKELY(!sax->start_object(len)))
+        {
+            return false;
+        }
+
+        string_t key;
+        for (std::size_t i = 0; i < len; ++i)
+        {
+            get();
+            if (JSON_HEDLEY_UNLIKELY(!get_msgpack_string(key) || !sax->key(key)))
+            {
+                return false;
+            }
+
+            if (JSON_HEDLEY_UNLIKELY(!parse_msgpack_internal()))
+            {
+                return false;
+            }
+            key.clear();
+        }
+
+        return sax->end_object();
+    }
+
+    ////////////
+    // UBJSON //
+    ////////////
+
+    /*!
+    @param[in] get_char  whether a new character should be retrieved from the
+                         input (true, default) or whether the last read
+                         character should be considered instead
+
+    @return whether a valid UBJSON value was passed to the SAX parser
+    */
+    bool parse_ubjson_internal(const bool get_char = true)
+    {
+        return get_ubjson_value(get_char ? get_ignore_noop() : current);
+    }
+
+    /*!
+    @brief reads a UBJSON string
+
+    This function is either called after reading the 'S' byte explicitly
+    indicating a string, or in case of an object key where the 'S' byte can be
+    left out.
+
+    @param[out] result   created string
+    @param[in] get_char  whether a new character should be retrieved from the
+                         input (true, default) or whether the last read
+                         character should be considered instead
+
+    @return whether string creation completed
+    */
+    bool get_ubjson_string(string_t& result, const bool get_char = true)
+    {
+        if (get_char)
+        {
+            get();  // TODO(niels): may we ignore N here?
+        }
+
+        if (JSON_HEDLEY_UNLIKELY(!unexpect_eof(input_format, "value")))
+        {
+            return false;
+        }
+
+        switch (current)
+        {
+            case 'U':
+            {
+                std::uint8_t len{};
+                return get_number(input_format, len) && get_string(input_format, len, result);
+            }
+
+            case 'i':
+            {
+                std::int8_t len{};
+                return get_number(input_format, len) && get_string(input_format, len, result);
+            }
+
+            case 'I':
+            {
+                std::int16_t len{};
+                return get_number(input_format, len) && get_string(input_format, len, result);
+            }
+
+            case 'l':
+            {
+                std::int32_t len{};
+                return get_number(input_format, len) && get_string(input_format, len, result);
+            }
+
+            case 'L':
+            {
+                std::int64_t len{};
+                return get_number(input_format, len) && get_string(input_format, len, result);
+            }
+
+            case 'u':
+            {
+                if (input_format != input_format_t::bjdata)
+                {
+                    break;
+                }
+                std::uint16_t len{};
+                return get_number(input_format, len) && get_string(input_format, len, result);
+            }
+
+            case 'm':
+            {
+                if (input_format != input_format_t::bjdata)
+                {
+                    break;
+                }
+                std::uint32_t len{};
+                return get_number(input_format, len) && get_string(input_format, len, result);
+            }
+
+            case 'M':
+            {
+                if (input_format != input_format_t::bjdata)
+                {
+                    break;
+                }
+                std::uint64_t len{};
+                return get_number(input_format, len) && get_string(input_format, len, result);
+            }
+
+            default:
+                break;
+        }
+        auto last_token = get_token_string();
+        std::string message;
+
+        if (input_format != input_format_t::bjdata)
+        {
+            message = "expected length type specification (U, i, I, l, L); last byte: 0x" + last_token;
+        }
+        else
+        {
+            message = "expected length type specification (U, i, u, I, m, l, M, L); last byte: 0x" + last_token;
+        }
+        return sax->parse_error(chars_read, last_token, parse_error::create(113, chars_read, exception_message(input_format, message, "string"), nullptr));
+    }
+
+    /*!
+    @param[out] dim  an integer vector storing the ND array dimensions
+    @return whether reading ND array size vector is successful
+    */
+    bool get_ubjson_ndarray_size(std::vector<size_t>& dim)
+    {
+        std::pair<std::size_t, char_int_type> size_and_type;
+        size_t dimlen = 0;
+        bool no_ndarray = true;
+
+        if (JSON_HEDLEY_UNLIKELY(!get_ubjson_size_type(size_and_type, no_ndarray)))
+        {
+            return false;
+        }
+
+        if (size_and_type.first != npos)
+        {
+            if (size_and_type.second != 0)
+            {
+                if (size_and_type.second != 'N')
+                {
+                    for (std::size_t i = 0; i < size_and_type.first; ++i)
+                    {
+                        if (JSON_HEDLEY_UNLIKELY(!get_ubjson_size_value(dimlen, no_ndarray, size_and_type.second)))
+                        {
+                            return false;
+                        }
+                        dim.push_back(dimlen);
+                    }
+                }
+            }
+            else
+            {
+                for (std::size_t i = 0; i < size_and_type.first; ++i)
+                {
+                    if (JSON_HEDLEY_UNLIKELY(!get_ubjson_size_value(dimlen, no_ndarray)))
+                    {
+                        return false;
+                    }
+                    dim.push_back(dimlen);
+                }
+            }
+        }
+        else
+        {
+            while (current != ']')
+            {
+                if (JSON_HEDLEY_UNLIKELY(!get_ubjson_size_value(dimlen, no_ndarray, current)))
+                {
+                    return false;
+                }
+                dim.push_back(dimlen);
+                get_ignore_noop();
+            }
+        }
+        return true;
+    }
+
+    /*!
+    @param[out] result  determined size
+    @param[in,out] is_ndarray  for input, `true` means already inside an ndarray vector
+                               or ndarray dimension is not allowed; `false` means ndarray
+                               is allowed; for output, `true` means an ndarray is found;
+                               is_ndarray can only return `true` when its initial value
+                               is `false`
+    @param[in] prefix  type marker if already read, otherwise set to 0
+
+    @return whether size determination completed
+    */
+    bool get_ubjson_size_value(std::size_t& result, bool& is_ndarray, char_int_type prefix = 0)
+    {
+        if (prefix == 0)
+        {
+            prefix = get_ignore_noop();
+        }
+
+        switch (prefix)
+        {
+            case 'U':
+            {
+                std::uint8_t number{};
+                if (JSON_HEDLEY_UNLIKELY(!get_number(input_format, number)))
+                {
+                    return false;
+                }
+                result = static_cast<std::size_t>(number);
+                return true;
+            }
+
+            case 'i':
+            {
+                std::int8_t number{};
+                if (JSON_HEDLEY_UNLIKELY(!get_number(input_format, number)))
+                {
+                    return false;
+                }
+                if (number < 0)
+                {
+                    return sax->parse_error(chars_read, get_token_string(), parse_error::create(113, chars_read,
+                                            exception_message(input_format, "count in an optimized container must be positive", "size"), nullptr));
+                }
+                result = static_cast<std::size_t>(number); // NOLINT(bugprone-signed-char-misuse,cert-str34-c): number is not a char
+                return true;
+            }
+
+            case 'I':
+            {
+                std::int16_t number{};
+                if (JSON_HEDLEY_UNLIKELY(!get_number(input_format, number)))
+                {
+                    return false;
+                }
+                if (number < 0)
+                {
+                    return sax->parse_error(chars_read, get_token_string(), parse_error::create(113, chars_read,
+                                            exception_message(input_format, "count in an optimized container must be positive", "size"), nullptr));
+                }
+                result = static_cast<std::size_t>(number);
+                return true;
+            }
+
+            case 'l':
+            {
+                std::int32_t number{};
+                if (JSON_HEDLEY_UNLIKELY(!get_number(input_format, number)))
+                {
+                    return false;
+                }
+                if (number < 0)
+                {
+                    return sax->parse_error(chars_read, get_token_string(), parse_error::create(113, chars_read,
+                                            exception_message(input_format, "count in an optimized container must be positive", "size"), nullptr));
+                }
+                result = static_cast<std::size_t>(number);
+                return true;
+            }
+
+            case 'L':
+            {
+                std::int64_t number{};
+                if (JSON_HEDLEY_UNLIKELY(!get_number(input_format, number)))
+                {
+                    return false;
+                }
+                if (number < 0)
+                {
+                    return sax->parse_error(chars_read, get_token_string(), parse_error::create(113, chars_read,
+                                            exception_message(input_format, "count in an optimized container must be positive", "size"), nullptr));
+                }
+                if (!value_in_range_of<std::size_t>(number))
+                {
+                    return sax->parse_error(chars_read, get_token_string(), out_of_range::create(408,
+                                            exception_message(input_format, "integer value overflow", "size"), nullptr));
+                }
+                result = static_cast<std::size_t>(number);
+                return true;
+            }
+
+            case 'u':
+            {
+                if (input_format != input_format_t::bjdata)
+                {
+                    break;
+                }
+                std::uint16_t number{};
+                if (JSON_HEDLEY_UNLIKELY(!get_number(input_format, number)))
+                {
+                    return false;
+                }
+                result = static_cast<std::size_t>(number);
+                return true;
+            }
+
+            case 'm':
+            {
+                if (input_format != input_format_t::bjdata)
+                {
+                    break;
+                }
+                std::uint32_t number{};
+                if (JSON_HEDLEY_UNLIKELY(!get_number(input_format, number)))
+                {
+                    return false;
+                }
+                result = conditional_static_cast<std::size_t>(number);
+                return true;
+            }
+
+            case 'M':
+            {
+                if (input_format != input_format_t::bjdata)
+                {
+                    break;
+                }
+                std::uint64_t number{};
+                if (JSON_HEDLEY_UNLIKELY(!get_number(input_format, number)))
+                {
+                    return false;
+                }
+                if (!value_in_range_of<std::size_t>(number))
+                {
+                    return sax->parse_error(chars_read, get_token_string(), out_of_range::create(408,
+                                            exception_message(input_format, "integer value overflow", "size"), nullptr));
+                }
+                result = detail::conditional_static_cast<std::size_t>(number);
+                return true;
+            }
+
+            case '[':
+            {
+                if (input_format != input_format_t::bjdata)
+                {
+                    break;
+                }
+                if (is_ndarray) // ndarray dimensional vector can only contain integers, and can not embed another array
+                {
+                    return sax->parse_error(chars_read, get_token_string(), parse_error::create(113, chars_read, exception_message(input_format, "ndarray dimentional vector is not allowed", "size"), nullptr));
+                }
+                std::vector<size_t> dim;
+                if (JSON_HEDLEY_UNLIKELY(!get_ubjson_ndarray_size(dim)))
+                {
+                    return false;
+                }
+                if (dim.size() == 1 || (dim.size() == 2 && dim.at(0) == 1)) // return normal array size if 1D row vector
+                {
+                    result = dim.at(dim.size() - 1);
+                    return true;
+                }
+                if (!dim.empty())  // if ndarray, convert to an object in JData annotated array format
+                {
+                    for (auto i : dim) // test if any dimension in an ndarray is 0, if so, return a 1D empty container
+                    {
+                        if ( i == 0 )
+                        {
+                            result = 0;
+                            return true;
+                        }
+                    }
+
+                    string_t key = "_ArraySize_";
+                    if (JSON_HEDLEY_UNLIKELY(!sax->start_object(3) || !sax->key(key) || !sax->start_array(dim.size())))
+                    {
+                        return false;
+                    }
+                    result = 1;
+                    for (auto i : dim)
+                    {
+                        result *= i;
+                        if (result == 0 || result == npos) // because dim elements shall not have zeros, result = 0 means overflow happened; it also can't be npos as it is used to initialize size in get_ubjson_size_type()
+                        {
+                            return sax->parse_error(chars_read, get_token_string(), out_of_range::create(408, exception_message(input_format, "excessive ndarray size caused overflow", "size"), nullptr));
+                        }
+                        if (JSON_HEDLEY_UNLIKELY(!sax->number_unsigned(static_cast<number_unsigned_t>(i))))
+                        {
+                            return false;
+                        }
+                    }
+                    is_ndarray = true;
+                    return sax->end_array();
+                }
+                result = 0;
+                return true;
+            }
+
+            default:
+                break;
+        }
+        auto last_token = get_token_string();
+        std::string message;
+
+        if (input_format != input_format_t::bjdata)
+        {
+            message = "expected length type specification (U, i, I, l, L) after '#'; last byte: 0x" + last_token;
+        }
+        else
+        {
+            message = "expected length type specification (U, i, u, I, m, l, M, L) after '#'; last byte: 0x" + last_token;
+        }
+        return sax->parse_error(chars_read, last_token, parse_error::create(113, chars_read, exception_message(input_format, message, "size"), nullptr));
+    }
+
+    /*!
+    @brief determine the type and size for a container
+
+    In the optimized UBJSON format, a type and a size can be provided to allow
+    for a more compact representation.
+
+    @param[out] result  pair of the size and the type
+    @param[in] inside_ndarray  whether the parser is parsing an ND array dimensional vector
+
+    @return whether pair creation completed
+    */
+    bool get_ubjson_size_type(std::pair<std::size_t, char_int_type>& result, bool inside_ndarray = false)
+    {
+        result.first = npos; // size
+        result.second = 0; // type
+        bool is_ndarray = false;
+
+        get_ignore_noop();
+
+        if (current == '$')
+        {
+            result.second = get();  // must not ignore 'N', because 'N' maybe the type
+            if (input_format == input_format_t::bjdata
+                    && JSON_HEDLEY_UNLIKELY(std::binary_search(bjd_optimized_type_markers.begin(), bjd_optimized_type_markers.end(), result.second)))
+            {
+                auto last_token = get_token_string();
+                return sax->parse_error(chars_read, last_token, parse_error::create(112, chars_read,
+                                        exception_message(input_format, concat("marker 0x", last_token, " is not a permitted optimized array type"), "type"), nullptr));
+            }
+
+            if (JSON_HEDLEY_UNLIKELY(!unexpect_eof(input_format, "type")))
+            {
+                return false;
+            }
+
+            get_ignore_noop();
+            if (JSON_HEDLEY_UNLIKELY(current != '#'))
+            {
+                if (JSON_HEDLEY_UNLIKELY(!unexpect_eof(input_format, "value")))
+                {
+                    return false;
+                }
+                auto last_token = get_token_string();
+                return sax->parse_error(chars_read, last_token, parse_error::create(112, chars_read,
+                                        exception_message(input_format, concat("expected '#' after type information; last byte: 0x", last_token), "size"), nullptr));
+            }
+
+            bool is_error = get_ubjson_size_value(result.first, is_ndarray);
+            if (input_format == input_format_t::bjdata && is_ndarray)
+            {
+                if (inside_ndarray)
+                {
+                    return sax->parse_error(chars_read, get_token_string(), parse_error::create(112, chars_read,
+                                            exception_message(input_format, "ndarray can not be recursive", "size"), nullptr));
+                }
+                result.second |= (1 << 8); // use bit 8 to indicate ndarray, all UBJSON and BJData markers should be ASCII letters
+            }
+            return is_error;
+        }
+
+        if (current == '#')
+        {
+            bool is_error = get_ubjson_size_value(result.first, is_ndarray);
+            if (input_format == input_format_t::bjdata && is_ndarray)
+            {
+                return sax->parse_error(chars_read, get_token_string(), parse_error::create(112, chars_read,
+                                        exception_message(input_format, "ndarray requires both type and size", "size"), nullptr));
+            }
+            return is_error;
+        }
+
+        return true;
+    }
+
+    /*!
+    @param prefix  the previously read or set type prefix
+    @return whether value creation completed
+    */
+    bool get_ubjson_value(const char_int_type prefix)
+    {
+        switch (prefix)
+        {
+            case std::char_traits<char_type>::eof():  // EOF
+                return unexpect_eof(input_format, "value");
+
+            case 'T':  // true
+                return sax->boolean(true);
+            case 'F':  // false
+                return sax->boolean(false);
+
+            case 'Z':  // null
+                return sax->null();
+
+            case 'U':
+            {
+                std::uint8_t number{};
+                return get_number(input_format, number) && sax->number_unsigned(number);
+            }
+
+            case 'i':
+            {
+                std::int8_t number{};
+                return get_number(input_format, number) && sax->number_integer(number);
+            }
+
+            case 'I':
+            {
+                std::int16_t number{};
+                return get_number(input_format, number) && sax->number_integer(number);
+            }
+
+            case 'l':
+            {
+                std::int32_t number{};
+                return get_number(input_format, number) && sax->number_integer(number);
+            }
+
+            case 'L':
+            {
+                std::int64_t number{};
+                return get_number(input_format, number) && sax->number_integer(number);
+            }
+
+            case 'u':
+            {
+                if (input_format != input_format_t::bjdata)
+                {
+                    break;
+                }
+                std::uint16_t number{};
+                return get_number(input_format, number) && sax->number_unsigned(number);
+            }
+
+            case 'm':
+            {
+                if (input_format != input_format_t::bjdata)
+                {
+                    break;
+                }
+                std::uint32_t number{};
+                return get_number(input_format, number) && sax->number_unsigned(number);
+            }
+
+            case 'M':
+            {
+                if (input_format != input_format_t::bjdata)
+                {
+                    break;
+                }
+                std::uint64_t number{};
+                return get_number(input_format, number) && sax->number_unsigned(number);
+            }
+
+            case 'h':
+            {
+                if (input_format != input_format_t::bjdata)
+                {
+                    break;
+                }
+                const auto byte1_raw = get();
+                if (JSON_HEDLEY_UNLIKELY(!unexpect_eof(input_format, "number")))
+                {
+                    return false;
+                }
+                const auto byte2_raw = get();
+                if (JSON_HEDLEY_UNLIKELY(!unexpect_eof(input_format, "number")))
+                {
+                    return false;
+                }
+
+                const auto byte1 = static_cast<unsigned char>(byte1_raw);
+                const auto byte2 = static_cast<unsigned char>(byte2_raw);
+
+                // code from RFC 7049, Appendix D, Figure 3:
+                // As half-precision floating-point numbers were only added
+                // to IEEE 754 in 2008, today's programming platforms often
+                // still only have limited support for them. It is very
+                // easy to include at least decoding support for them even
+                // without such support. An example of a small decoder for
+                // half-precision floating-point numbers in the C language
+                // is shown in Fig. 3.
+                const auto half = static_cast<unsigned int>((byte2 << 8u) + byte1);
+                const double val = [&half]
+                {
+                    const int exp = (half >> 10u) & 0x1Fu;
+                    const unsigned int mant = half & 0x3FFu;
+                    JSON_ASSERT(0 <= exp&& exp <= 32);
+                    JSON_ASSERT(mant <= 1024);
+                    switch (exp)
+                    {
+                        case 0:
+                            return std::ldexp(mant, -24);
+                        case 31:
+                            return (mant == 0)
+                            ? std::numeric_limits<double>::infinity()
+                            : std::numeric_limits<double>::quiet_NaN();
+                        default:
+                            return std::ldexp(mant + 1024, exp - 25);
+                    }
+                }();
+                return sax->number_float((half & 0x8000u) != 0
+                                         ? static_cast<number_float_t>(-val)
+                                         : static_cast<number_float_t>(val), "");
+            }
+
+            case 'd':
+            {
+                float number{};
+                return get_number(input_format, number) && sax->number_float(static_cast<number_float_t>(number), "");
+            }
+
+            case 'D':
+            {
+                double number{};
+                return get_number(input_format, number) && sax->number_float(static_cast<number_float_t>(number), "");
+            }
+
+            case 'H':
+            {
+                return get_ubjson_high_precision_number();
+            }
+
+            case 'C':  // char
+            {
+                get();
+                if (JSON_HEDLEY_UNLIKELY(!unexpect_eof(input_format, "char")))
+                {
+                    return false;
+                }
+                if (JSON_HEDLEY_UNLIKELY(current > 127))
+                {
+                    auto last_token = get_token_string();
+                    return sax->parse_error(chars_read, last_token, parse_error::create(113, chars_read,
+                                            exception_message(input_format, concat("byte after 'C' must be in range 0x00..0x7F; last byte: 0x", last_token), "char"), nullptr));
+                }
+                string_t s(1, static_cast<typename string_t::value_type>(current));
+                return sax->string(s);
+            }
+
+            case 'S':  // string
+            {
+                string_t s;
+                return get_ubjson_string(s) && sax->string(s);
+            }
+
+            case '[':  // array
+                return get_ubjson_array();
+
+            case '{':  // object
+                return get_ubjson_object();
+
+            default: // anything else
+                break;
+        }
+        auto last_token = get_token_string();
+        return sax->parse_error(chars_read, last_token, parse_error::create(112, chars_read, exception_message(input_format, "invalid byte: 0x" + last_token, "value"), nullptr));
+    }
+
+    /*!
+    @return whether array creation completed
+    */
+    bool get_ubjson_array()
+    {
+        std::pair<std::size_t, char_int_type> size_and_type;
+        if (JSON_HEDLEY_UNLIKELY(!get_ubjson_size_type(size_and_type)))
+        {
+            return false;
+        }
+
+        // if bit-8 of size_and_type.second is set to 1, encode bjdata ndarray as an object in JData annotated array format (https://github.com/NeuroJSON/jdata):
+        // {"_ArrayType_" : "typeid", "_ArraySize_" : [n1, n2, ...], "_ArrayData_" : [v1, v2, ...]}
+
+        if (input_format == input_format_t::bjdata && size_and_type.first != npos && (size_and_type.second & (1 << 8)) != 0)
+        {
+            size_and_type.second &= ~(static_cast<char_int_type>(1) << 8);  // use bit 8 to indicate ndarray, here we remove the bit to restore the type marker
+            auto it = std::lower_bound(bjd_types_map.begin(), bjd_types_map.end(), size_and_type.second, [](const bjd_type & p, char_int_type t)
+            {
+                return p.first < t;
+            });
+            string_t key = "_ArrayType_";
+            if (JSON_HEDLEY_UNLIKELY(it == bjd_types_map.end() || it->first != size_and_type.second))
+            {
+                auto last_token = get_token_string();
+                return sax->parse_error(chars_read, last_token, parse_error::create(112, chars_read,
+                                        exception_message(input_format, "invalid byte: 0x" + last_token, "type"), nullptr));
+            }
+
+            string_t type = it->second; // sax->string() takes a reference
+            if (JSON_HEDLEY_UNLIKELY(!sax->key(key) || !sax->string(type)))
+            {
+                return false;
+            }
+
+            if (size_and_type.second == 'C')
+            {
+                size_and_type.second = 'U';
+            }
+
+            key = "_ArrayData_";
+            if (JSON_HEDLEY_UNLIKELY(!sax->key(key) || !sax->start_array(size_and_type.first) ))
+            {
+                return false;
+            }
+
+            for (std::size_t i = 0; i < size_and_type.first; ++i)
+            {
+                if (JSON_HEDLEY_UNLIKELY(!get_ubjson_value(size_and_type.second)))
+                {
+                    return false;
+                }
+            }
+
+            return (sax->end_array() && sax->end_object());
+        }
+
+        if (size_and_type.first != npos)
+        {
+            if (JSON_HEDLEY_UNLIKELY(!sax->start_array(size_and_type.first)))
+            {
+                return false;
+            }
+
+            if (size_and_type.second != 0)
+            {
+                if (size_and_type.second != 'N')
+                {
+                    for (std::size_t i = 0; i < size_and_type.first; ++i)
+                    {
+                        if (JSON_HEDLEY_UNLIKELY(!get_ubjson_value(size_and_type.second)))
+                        {
+                            return false;
+                        }
+                    }
+                }
+            }
+            else
+            {
+                for (std::size_t i = 0; i < size_and_type.first; ++i)
+                {
+                    if (JSON_HEDLEY_UNLIKELY(!parse_ubjson_internal()))
+                    {
+                        return false;
+                    }
+                }
+            }
+        }
+        else
+        {
+            if (JSON_HEDLEY_UNLIKELY(!sax->start_array(static_cast<std::size_t>(-1))))
+            {
+                return false;
+            }
+
+            while (current != ']')
+            {
+                if (JSON_HEDLEY_UNLIKELY(!parse_ubjson_internal(false)))
+                {
+                    return false;
+                }
+                get_ignore_noop();
+            }
+        }
+
+        return sax->end_array();
+    }
+
+    /*!
+    @return whether object creation completed
+    */
+    bool get_ubjson_object()
+    {
+        std::pair<std::size_t, char_int_type> size_and_type;
+        if (JSON_HEDLEY_UNLIKELY(!get_ubjson_size_type(size_and_type)))
+        {
+            return false;
+        }
+
+        // do not accept ND-array size in objects in BJData
+        if (input_format == input_format_t::bjdata && size_and_type.first != npos && (size_and_type.second & (1 << 8)) != 0)
+        {
+            auto last_token = get_token_string();
+            return sax->parse_error(chars_read, last_token, parse_error::create(112, chars_read,
+                                    exception_message(input_format, "BJData object does not support ND-array size in optimized format", "object"), nullptr));
+        }
+
+        string_t key;
+        if (size_and_type.first != npos)
+        {
+            if (JSON_HEDLEY_UNLIKELY(!sax->start_object(size_and_type.first)))
+            {
+                return false;
+            }
+
+            if (size_and_type.second != 0)
+            {
+                for (std::size_t i = 0; i < size_and_type.first; ++i)
+                {
+                    if (JSON_HEDLEY_UNLIKELY(!get_ubjson_string(key) || !sax->key(key)))
+                    {
+                        return false;
+                    }
+                    if (JSON_HEDLEY_UNLIKELY(!get_ubjson_value(size_and_type.second)))
+                    {
+                        return false;
+                    }
+                    key.clear();
+                }
+            }
+            else
+            {
+                for (std::size_t i = 0; i < size_and_type.first; ++i)
+                {
+                    if (JSON_HEDLEY_UNLIKELY(!get_ubjson_string(key) || !sax->key(key)))
+                    {
+                        return false;
+                    }
+                    if (JSON_HEDLEY_UNLIKELY(!parse_ubjson_internal()))
+                    {
+                        return false;
+                    }
+                    key.clear();
+                }
+            }
+        }
+        else
+        {
+            if (JSON_HEDLEY_UNLIKELY(!sax->start_object(static_cast<std::size_t>(-1))))
+            {
+                return false;
+            }
+
+            while (current != '}')
+            {
+                if (JSON_HEDLEY_UNLIKELY(!get_ubjson_string(key, false) || !sax->key(key)))
+                {
+                    return false;
+                }
+                if (JSON_HEDLEY_UNLIKELY(!parse_ubjson_internal()))
+                {
+                    return false;
+                }
+                get_ignore_noop();
+                key.clear();
+            }
+        }
+
+        return sax->end_object();
+    }
+
+    // Note, no reader for UBJSON binary types is implemented because they do
+    // not exist
+
+    bool get_ubjson_high_precision_number()
+    {
+        // get size of following number string
+        std::size_t size{};
+        bool no_ndarray = true;
+        auto res = get_ubjson_size_value(size, no_ndarray);
+        if (JSON_HEDLEY_UNLIKELY(!res))
+        {
+            return res;
+        }
+
+        // get number string
+        std::vector<char> number_vector;
+        for (std::size_t i = 0; i < size; ++i)
+        {
+            get();
+            if (JSON_HEDLEY_UNLIKELY(!unexpect_eof(input_format, "number")))
+            {
+                return false;
+            }
+            number_vector.push_back(static_cast<char>(current));
+        }
+
+        // parse number string
+        using ia_type = decltype(detail::input_adapter(number_vector));
+        auto number_lexer = detail::lexer<BasicJsonType, ia_type>(detail::input_adapter(number_vector), false);
+        const auto result_number = number_lexer.scan();
+        const auto number_string = number_lexer.get_token_string();
+        const auto result_remainder = number_lexer.scan();
+
+        using token_type = typename detail::lexer_base<BasicJsonType>::token_type;
+
+        if (JSON_HEDLEY_UNLIKELY(result_remainder != token_type::end_of_input))
+        {
+            return sax->parse_error(chars_read, number_string, parse_error::create(115, chars_read,
+                                    exception_message(input_format, concat("invalid number text: ", number_lexer.get_token_string()), "high-precision number"), nullptr));
+        }
+
+        switch (result_number)
+        {
+            case token_type::value_integer:
+                return sax->number_integer(number_lexer.get_number_integer());
+            case token_type::value_unsigned:
+                return sax->number_unsigned(number_lexer.get_number_unsigned());
+            case token_type::value_float:
+                return sax->number_float(number_lexer.get_number_float(), std::move(number_string));
+            case token_type::uninitialized:
+            case token_type::literal_true:
+            case token_type::literal_false:
+            case token_type::literal_null:
+            case token_type::value_string:
+            case token_type::begin_array:
+            case token_type::begin_object:
+            case token_type::end_array:
+            case token_type::end_object:
+            case token_type::name_separator:
+            case token_type::value_separator:
+            case token_type::parse_error:
+            case token_type::end_of_input:
+            case token_type::literal_or_value:
+            default:
+                return sax->parse_error(chars_read, number_string, parse_error::create(115, chars_read,
+                                        exception_message(input_format, concat("invalid number text: ", number_lexer.get_token_string()), "high-precision number"), nullptr));
+        }
+    }
+
+    ///////////////////////
+    // Utility functions //
+    ///////////////////////
+
+    /*!
+    @brief get next character from the input
+
+    This function provides the interface to the used input adapter. It does
+    not throw in case the input reached EOF, but returns a -'ve valued
+    `std::char_traits<char_type>::eof()` in that case.
+
+    @return character read from the input
+    */
+    char_int_type get()
+    {
+        ++chars_read;
+        return current = ia.get_character();
+    }
+
+    /*!
+    @return character read from the input after ignoring all 'N' entries
+    */
+    char_int_type get_ignore_noop()
+    {
+        do
+        {
+            get();
+        }
+        while (current == 'N');
+
+        return current;
+    }
+
+    /*
+    @brief read a number from the input
+
+    @tparam NumberType the type of the number
+    @param[in] format   the current format (for diagnostics)
+    @param[out] result  number of type @a NumberType
+
+    @return whether conversion completed
+
+    @note This function needs to respect the system's endianness, because
+          bytes in CBOR, MessagePack, and UBJSON are stored in network order
+          (big endian) and therefore need reordering on little endian systems.
+          On the other hand, BSON and BJData use little endian and should reorder
+          on big endian systems.
+    */
+    template<typename NumberType, bool InputIsLittleEndian = false>
+    bool get_number(const input_format_t format, NumberType& result)
+    {
+        // step 1: read input into array with system's byte order
+        std::array<std::uint8_t, sizeof(NumberType)> vec{};
+        for (std::size_t i = 0; i < sizeof(NumberType); ++i)
+        {
+            get();
+            if (JSON_HEDLEY_UNLIKELY(!unexpect_eof(format, "number")))
+            {
+                return false;
+            }
+
+            // reverse byte order prior to conversion if necessary
+            if (is_little_endian != (InputIsLittleEndian || format == input_format_t::bjdata))
+            {
+                vec[sizeof(NumberType) - i - 1] = static_cast<std::uint8_t>(current);
+            }
+            else
+            {
+                vec[i] = static_cast<std::uint8_t>(current); // LCOV_EXCL_LINE
+            }
+        }
+
+        // step 2: convert array into number of type T and return
+        std::memcpy(&result, vec.data(), sizeof(NumberType));
+        return true;
+    }
+
+    /*!
+    @brief create a string by reading characters from the input
+
+    @tparam NumberType the type of the number
+    @param[in] format the current format (for diagnostics)
+    @param[in] len number of characters to read
+    @param[out] result string created by reading @a len bytes
+
+    @return whether string creation completed
+
+    @note We can not reserve @a len bytes for the result, because @a len
+          may be too large. Usually, @ref unexpect_eof() detects the end of
+          the input before we run out of string memory.
+    */
+    template<typename NumberType>
+    bool get_string(const input_format_t format,
+                    const NumberType len,
+                    string_t& result)
+    {
+        bool success = true;
+        for (NumberType i = 0; i < len; i++)
+        {
+            get();
+            if (JSON_HEDLEY_UNLIKELY(!unexpect_eof(format, "string")))
+            {
+                success = false;
+                break;
+            }
+            result.push_back(static_cast<typename string_t::value_type>(current));
+        }
+        return success;
+    }
+
+    /*!
+    @brief create a byte array by reading bytes from the input
+
+    @tparam NumberType the type of the number
+    @param[in] format the current format (for diagnostics)
+    @param[in] len number of bytes to read
+    @param[out] result byte array created by reading @a len bytes
+
+    @return whether byte array creation completed
+
+    @note We can not reserve @a len bytes for the result, because @a len
+          may be too large. Usually, @ref unexpect_eof() detects the end of
+          the input before we run out of memory.
+    */
+    template<typename NumberType>
+    bool get_binary(const input_format_t format,
+                    const NumberType len,
+                    binary_t& result)
+    {
+        bool success = true;
+        for (NumberType i = 0; i < len; i++)
+        {
+            get();
+            if (JSON_HEDLEY_UNLIKELY(!unexpect_eof(format, "binary")))
+            {
+                success = false;
+                break;
+            }
+            result.push_back(static_cast<std::uint8_t>(current));
+        }
+        return success;
+    }
+
+    /*!
+    @param[in] format   the current format (for diagnostics)
+    @param[in] context  further context information (for diagnostics)
+    @return whether the last read character is not EOF
+    */
+    JSON_HEDLEY_NON_NULL(3)
+    bool unexpect_eof(const input_format_t format, const char* context) const
+    {
+        if (JSON_HEDLEY_UNLIKELY(current == std::char_traits<char_type>::eof()))
+        {
+            return sax->parse_error(chars_read, "<end of file>",
+                                    parse_error::create(110, chars_read, exception_message(format, "unexpected end of input", context), nullptr));
+        }
+        return true;
+    }
+
+    /*!
+    @return a string representation of the last read byte
+    */
+    std::string get_token_string() const
+    {
+        std::array<char, 3> cr{{}};
+        static_cast<void>((std::snprintf)(cr.data(), cr.size(), "%.2hhX", static_cast<unsigned char>(current))); // NOLINT(cppcoreguidelines-pro-type-vararg,hicpp-vararg)
+        return std::string{cr.data()};
+    }
+
+    /*!
+    @param[in] format   the current format
+    @param[in] detail   a detailed error message
+    @param[in] context  further context information
+    @return a message string to use in the parse_error exceptions
+    */
+    std::string exception_message(const input_format_t format,
+                                  const std::string& detail,
+                                  const std::string& context) const
+    {
+        std::string error_msg = "syntax error while parsing ";
+
+        switch (format)
+        {
+            case input_format_t::cbor:
+                error_msg += "CBOR";
+                break;
+
+            case input_format_t::msgpack:
+                error_msg += "MessagePack";
+                break;
+
+            case input_format_t::ubjson:
+                error_msg += "UBJSON";
+                break;
+
+            case input_format_t::bson:
+                error_msg += "BSON";
+                break;
+
+            case input_format_t::bjdata:
+                error_msg += "BJData";
+                break;
+
+            case input_format_t::json: // LCOV_EXCL_LINE
+            default:            // LCOV_EXCL_LINE
+                JSON_ASSERT(false); // NOLINT(cert-dcl03-c,hicpp-static-assert,misc-static-assert) LCOV_EXCL_LINE
+        }
+
+        return concat(error_msg, ' ', context, ": ", detail);
+    }
+
+  private:
+    static JSON_INLINE_VARIABLE constexpr std::size_t npos = static_cast<std::size_t>(-1);
+
+    /// input adapter
+    InputAdapterType ia;
+
+    /// the current character
+    char_int_type current = std::char_traits<char_type>::eof();
+
+    /// the number of characters read
+    std::size_t chars_read = 0;
+
+    /// whether we can assume little endianness
+    const bool is_little_endian = little_endianness();
+
+    /// input format
+    const input_format_t input_format = input_format_t::json;
+
+    /// the SAX parser
+    json_sax_t* sax = nullptr;
+
+    // excluded markers in bjdata optimized type
+#define JSON_BINARY_READER_MAKE_BJD_OPTIMIZED_TYPE_MARKERS_ \
+    make_array<char_int_type>('F', 'H', 'N', 'S', 'T', 'Z', '[', '{')
+
+#define JSON_BINARY_READER_MAKE_BJD_TYPES_MAP_ \
+    make_array<bjd_type>(                      \
+    bjd_type{'C', "char"},                     \
+    bjd_type{'D', "double"},                   \
+    bjd_type{'I', "int16"},                    \
+    bjd_type{'L', "int64"},                    \
+    bjd_type{'M', "uint64"},                   \
+    bjd_type{'U', "uint8"},                    \
+    bjd_type{'d', "single"},                   \
+    bjd_type{'i', "int8"},                     \
+    bjd_type{'l', "int32"},                    \
+    bjd_type{'m', "uint32"},                   \
+    bjd_type{'u', "uint16"})
+
+  JSON_PRIVATE_UNLESS_TESTED:
+    // lookup tables
+    // NOLINTNEXTLINE(cppcoreguidelines-non-private-member-variables-in-classes)
+    const decltype(JSON_BINARY_READER_MAKE_BJD_OPTIMIZED_TYPE_MARKERS_) bjd_optimized_type_markers =
+        JSON_BINARY_READER_MAKE_BJD_OPTIMIZED_TYPE_MARKERS_;
+
+    using bjd_type = std::pair<char_int_type, string_t>;
+    // NOLINTNEXTLINE(cppcoreguidelines-non-private-member-variables-in-classes)
+    const decltype(JSON_BINARY_READER_MAKE_BJD_TYPES_MAP_) bjd_types_map =
+        JSON_BINARY_READER_MAKE_BJD_TYPES_MAP_;
+
+#undef JSON_BINARY_READER_MAKE_BJD_OPTIMIZED_TYPE_MARKERS_
+#undef JSON_BINARY_READER_MAKE_BJD_TYPES_MAP_
+};
+
+#ifndef JSON_HAS_CPP_17
+    template<typename BasicJsonType, typename InputAdapterType, typename SAX>
+    constexpr std::size_t binary_reader<BasicJsonType, InputAdapterType, SAX>::npos;
+#endif
+
+}  // namespace detail
+NLOHMANN_JSON_NAMESPACE_END
+
+// #include <nlohmann/detail/input/input_adapters.hpp>
+
+// #include <nlohmann/detail/input/lexer.hpp>
+
+// #include <nlohmann/detail/input/parser.hpp>
+//     __ _____ _____ _____
+//  __|  |   __|     |   | |  JSON for Modern C++
+// |  |  |__   |  |  | | | |  version 3.11.2
+// |_____|_____|_____|_|___|  https://github.com/nlohmann/json
+//
+// SPDX-FileCopyrightText: 2013-2022 Niels Lohmann <https://nlohmann.me>
+// SPDX-License-Identifier: MIT
+
+
+
+#include <cmath> // isfinite
+#include <cstdint> // uint8_t
+#include <functional> // function
+#include <string> // string
+#include <utility> // move
+#include <vector> // vector
+
+// #include <nlohmann/detail/exceptions.hpp>
+
+// #include <nlohmann/detail/input/input_adapters.hpp>
+
+// #include <nlohmann/detail/input/json_sax.hpp>
+
+// #include <nlohmann/detail/input/lexer.hpp>
+
+// #include <nlohmann/detail/macro_scope.hpp>
+
+// #include <nlohmann/detail/meta/is_sax.hpp>
+
+// #include <nlohmann/detail/string_concat.hpp>
+
+// #include <nlohmann/detail/value_t.hpp>
+
+
+NLOHMANN_JSON_NAMESPACE_BEGIN
+namespace detail
+{
+////////////
+// parser //
+////////////
+
+enum class parse_event_t : std::uint8_t
+{
+    /// the parser read `{` and started to process a JSON object
+    object_start,
+    /// the parser read `}` and finished processing a JSON object
+    object_end,
+    /// the parser read `[` and started to process a JSON array
+    array_start,
+    /// the parser read `]` and finished processing a JSON array
+    array_end,
+    /// the parser read a key of a value in an object
+    key,
+    /// the parser finished reading a JSON value
+    value
+};
+
+template<typename BasicJsonType>
+using parser_callback_t =
+    std::function<bool(int /*depth*/, parse_event_t /*event*/, BasicJsonType& /*parsed*/)>;
+
+/*!
+@brief syntax analysis
+
+This class implements a recursive descent parser.
+*/
+template<typename BasicJsonType, typename InputAdapterType>
+class parser
+{
+    using number_integer_t = typename BasicJsonType::number_integer_t;
+    using number_unsigned_t = typename BasicJsonType::number_unsigned_t;
+    using number_float_t = typename BasicJsonType::number_float_t;
+    using string_t = typename BasicJsonType::string_t;
+    using lexer_t = lexer<BasicJsonType, InputAdapterType>;
+    using token_type = typename lexer_t::token_type;
+
+  public:
+    /// a parser reading from an input adapter
+    explicit parser(InputAdapterType&& adapter,
+                    const parser_callback_t<BasicJsonType> cb = nullptr,
+                    const bool allow_exceptions_ = true,
+                    const bool skip_comments = false)
+        : callback(cb)
+        , m_lexer(std::move(adapter), skip_comments)
+        , allow_exceptions(allow_exceptions_)
+    {
+        // read first token
+        get_token();
+    }
+
+    /*!
+    @brief public parser interface
+
+    @param[in] strict      whether to expect the last token to be EOF
+    @param[in,out] result  parsed JSON value
+
+    @throw parse_error.101 in case of an unexpected token
+    @throw parse_error.102 if to_unicode fails or surrogate error
+    @throw parse_error.103 if to_unicode fails
+    */
+    void parse(const bool strict, BasicJsonType& result)
+    {
+        if (callback)
+        {
+            json_sax_dom_callback_parser<BasicJsonType> sdp(result, callback, allow_exceptions);
+            sax_parse_internal(&sdp);
+
+            // in strict mode, input must be completely read
+            if (strict && (get_token() != token_type::end_of_input))
+            {
+                sdp.parse_error(m_lexer.get_position(),
+                                m_lexer.get_token_string(),
+                                parse_error::create(101, m_lexer.get_position(),
+                                                    exception_message(token_type::end_of_input, "value"), nullptr));
+            }
+
+            // in case of an error, return discarded value
+            if (sdp.is_errored())
+            {
+                result = value_t::discarded;
+                return;
+            }
+
+            // set top-level value to null if it was discarded by the callback
+            // function
+            if (result.is_discarded())
+            {
+                result = nullptr;
+            }
+        }
+        else
+        {
+            json_sax_dom_parser<BasicJsonType> sdp(result, allow_exceptions);
+            sax_parse_internal(&sdp);
+
+            // in strict mode, input must be completely read
+            if (strict && (get_token() != token_type::end_of_input))
+            {
+                sdp.parse_error(m_lexer.get_position(),
+                                m_lexer.get_token_string(),
+                                parse_error::create(101, m_lexer.get_position(), exception_message(token_type::end_of_input, "value"), nullptr));
+            }
+
+            // in case of an error, return discarded value
+            if (sdp.is_errored())
+            {
+                result = value_t::discarded;
+                return;
+            }
+        }
+
+        result.assert_invariant();
+    }
+
+    /*!
+    @brief public accept interface
+
+    @param[in] strict  whether to expect the last token to be EOF
+    @return whether the input is a proper JSON text
+    */
+    bool accept(const bool strict = true)
+    {
+        json_sax_acceptor<BasicJsonType> sax_acceptor;
+        return sax_parse(&sax_acceptor, strict);
+    }
+
+    template<typename SAX>
+    JSON_HEDLEY_NON_NULL(2)
+    bool sax_parse(SAX* sax, const bool strict = true)
+    {
+        (void)detail::is_sax_static_asserts<SAX, BasicJsonType> {};
+        const bool result = sax_parse_internal(sax);
+
+        // strict mode: next byte must be EOF
+        if (result && strict && (get_token() != token_type::end_of_input))
+        {
+            return sax->parse_error(m_lexer.get_position(),
+                                    m_lexer.get_token_string(),
+                                    parse_error::create(101, m_lexer.get_position(), exception_message(token_type::end_of_input, "value"), nullptr));
+        }
+
+        return result;
+    }
+
+  private:
+    template<typename SAX>
+    JSON_HEDLEY_NON_NULL(2)
+    bool sax_parse_internal(SAX* sax)
+    {
+        // stack to remember the hierarchy of structured values we are parsing
+        // true = array; false = object
+        std::vector<bool> states;
+        // value to avoid a goto (see comment where set to true)
+        bool skip_to_state_evaluation = false;
+
+        while (true)
+        {
+            if (!skip_to_state_evaluation)
+            {
+                // invariant: get_token() was called before each iteration
+                switch (last_token)
+                {
+                    case token_type::begin_object:
+                    {
+                        if (JSON_HEDLEY_UNLIKELY(!sax->start_object(static_cast<std::size_t>(-1))))
+                        {
+                            return false;
+                        }
+
+                        // closing } -> we are done
+                        if (get_token() == token_type::end_object)
+                        {
+                            if (JSON_HEDLEY_UNLIKELY(!sax->end_object()))
+                            {
+                                return false;
+                            }
+                            break;
+                        }
+
+                        // parse key
+                        if (JSON_HEDLEY_UNLIKELY(last_token != token_type::value_string))
+                        {
+                            return sax->parse_error(m_lexer.get_position(),
+                                                    m_lexer.get_token_string(),
+                                                    parse_error::create(101, m_lexer.get_position(), exception_message(token_type::value_string, "object key"), nullptr));
+                        }
+                        if (JSON_HEDLEY_UNLIKELY(!sax->key(m_lexer.get_string())))
+                        {
+                            return false;
+                        }
+
+                        // parse separator (:)
+                        if (JSON_HEDLEY_UNLIKELY(get_token() != token_type::name_separator))
+                        {
+                            return sax->parse_error(m_lexer.get_position(),
+                                                    m_lexer.get_token_string(),
+                                                    parse_error::create(101, m_lexer.get_position(), exception_message(token_type::name_separator, "object separator"), nullptr));
+                        }
+
+                        // remember we are now inside an object
+                        states.push_back(false);
+
+                        // parse values
+                        get_token();
+                        continue;
+                    }
+
+                    case token_type::begin_array:
+                    {
+                        if (JSON_HEDLEY_UNLIKELY(!sax->start_array(static_cast<std::size_t>(-1))))
+                        {
+                            return false;
+                        }
+
+                        // closing ] -> we are done
+                        if (get_token() == token_type::end_array)
+                        {
+                            if (JSON_HEDLEY_UNLIKELY(!sax->end_array()))
+                            {
+                                return false;
+                            }
+                            break;
+                        }
+
+                        // remember we are now inside an array
+                        states.push_back(true);
+
+                        // parse values (no need to call get_token)
+                        continue;
+                    }
+
+                    case token_type::value_float:
+                    {
+                        const auto res = m_lexer.get_number_float();
+
+                        if (JSON_HEDLEY_UNLIKELY(!std::isfinite(res)))
+                        {
+                            return sax->parse_error(m_lexer.get_position(),
+                                                    m_lexer.get_token_string(),
+                                                    out_of_range::create(406, concat("number overflow parsing '", m_lexer.get_token_string(), '\''), nullptr));
+                        }
+
+                        if (JSON_HEDLEY_UNLIKELY(!sax->number_float(res, m_lexer.get_string())))
+                        {
+                            return false;
+                        }
+
+                        break;
+                    }
+
+                    case token_type::literal_false:
+                    {
+                        if (JSON_HEDLEY_UNLIKELY(!sax->boolean(false)))
+                        {
+                            return false;
+                        }
+                        break;
+                    }
+
+                    case token_type::literal_null:
+                    {
+                        if (JSON_HEDLEY_UNLIKELY(!sax->null()))
+                        {
+                            return false;
+                        }
+                        break;
+                    }
+
+                    case token_type::literal_true:
+                    {
+                        if (JSON_HEDLEY_UNLIKELY(!sax->boolean(true)))
+                        {
+                            return false;
+                        }
+                        break;
+                    }
+
+                    case token_type::value_integer:
+                    {
+                        if (JSON_HEDLEY_UNLIKELY(!sax->number_integer(m_lexer.get_number_integer())))
+                        {
+                            return false;
+                        }
+                        break;
+                    }
+
+                    case token_type::value_string:
+                    {
+                        if (JSON_HEDLEY_UNLIKELY(!sax->string(m_lexer.get_string())))
+                        {
+                            return false;
+                        }
+                        break;
+                    }
+
+                    case token_type::value_unsigned:
+                    {
+                        if (JSON_HEDLEY_UNLIKELY(!sax->number_unsigned(m_lexer.get_number_unsigned())))
+                        {
+                            return false;
+                        }
+                        break;
+                    }
+
+                    case token_type::parse_error:
+                    {
+                        // using "uninitialized" to avoid "expected" message
+                        return sax->parse_error(m_lexer.get_position(),
+                                                m_lexer.get_token_string(),
+                                                parse_error::create(101, m_lexer.get_position(), exception_message(token_type::uninitialized, "value"), nullptr));
+                    }
+
+                    case token_type::uninitialized:
+                    case token_type::end_array:
+                    case token_type::end_object:
+                    case token_type::name_separator:
+                    case token_type::value_separator:
+                    case token_type::end_of_input:
+                    case token_type::literal_or_value:
+                    default: // the last token was unexpected
+                    {
+                        return sax->parse_error(m_lexer.get_position(),
+                                                m_lexer.get_token_string(),
+                                                parse_error::create(101, m_lexer.get_position(), exception_message(token_type::literal_or_value, "value"), nullptr));
+                    }
+                }
+            }
+            else
+            {
+                skip_to_state_evaluation = false;
+            }
+
+            // we reached this line after we successfully parsed a value
+            if (states.empty())
+            {
+                // empty stack: we reached the end of the hierarchy: done
+                return true;
+            }
+
+            if (states.back())  // array
+            {
+                // comma -> next value
+                if (get_token() == token_type::value_separator)
+                {
+                    // parse a new value
+                    get_token();
+                    continue;
+                }
+
+                // closing ]
+                if (JSON_HEDLEY_LIKELY(last_token == token_type::end_array))
+                {
+                    if (JSON_HEDLEY_UNLIKELY(!sax->end_array()))
+                    {
+                        return false;
+                    }
+
+                    // We are done with this array. Before we can parse a
+                    // new value, we need to evaluate the new state first.
+                    // By setting skip_to_state_evaluation to false, we
+                    // are effectively jumping to the beginning of this if.
+                    JSON_ASSERT(!states.empty());
+                    states.pop_back();
+                    skip_to_state_evaluation = true;
+                    continue;
+                }
+
+                return sax->parse_error(m_lexer.get_position(),
+                                        m_lexer.get_token_string(),
+                                        parse_error::create(101, m_lexer.get_position(), exception_message(token_type::end_array, "array"), nullptr));
+            }
+
+            // states.back() is false -> object
+
+            // comma -> next value
+            if (get_token() == token_type::value_separator)
+            {
+                // parse key
+                if (JSON_HEDLEY_UNLIKELY(get_token() != token_type::value_string))
+                {
+                    return sax->parse_error(m_lexer.get_position(),
+                                            m_lexer.get_token_string(),
+                                            parse_error::create(101, m_lexer.get_position(), exception_message(token_type::value_string, "object key"), nullptr));
+                }
+
+                if (JSON_HEDLEY_UNLIKELY(!sax->key(m_lexer.get_string())))
+                {
+                    return false;
+                }
+
+                // parse separator (:)
+                if (JSON_HEDLEY_UNLIKELY(get_token() != token_type::name_separator))
+                {
+                    return sax->parse_error(m_lexer.get_position(),
+                                            m_lexer.get_token_string(),
+                                            parse_error::create(101, m_lexer.get_position(), exception_message(token_type::name_separator, "object separator"), nullptr));
+                }
+
+                // parse values
+                get_token();
+                continue;
+            }
+
+            // closing }
+            if (JSON_HEDLEY_LIKELY(last_token == token_type::end_object))
+            {
+                if (JSON_HEDLEY_UNLIKELY(!sax->end_object()))
+                {
+                    return false;
+                }
+
+                // We are done with this object. Before we can parse a
+                // new value, we need to evaluate the new state first.
+                // By setting skip_to_state_evaluation to false, we
+                // are effectively jumping to the beginning of this if.
+                JSON_ASSERT(!states.empty());
+                states.pop_back();
+                skip_to_state_evaluation = true;
+                continue;
+            }
+
+            return sax->parse_error(m_lexer.get_position(),
+                                    m_lexer.get_token_string(),
+                                    parse_error::create(101, m_lexer.get_position(), exception_message(token_type::end_object, "object"), nullptr));
+        }
+    }
+
+    /// get next token from lexer
+    token_type get_token()
+    {
+        return last_token = m_lexer.scan();
+    }
+
+    std::string exception_message(const token_type expected, const std::string& context)
+    {
+        std::string error_msg = "syntax error ";
+
+        if (!context.empty())
+        {
+            error_msg += concat("while parsing ", context, ' ');
+        }
+
+        error_msg += "- ";
+
+        if (last_token == token_type::parse_error)
+        {
+            error_msg += concat(m_lexer.get_error_message(), "; last read: '",
+                                m_lexer.get_token_string(), '\'');
+        }
+        else
+        {
+            error_msg += concat("unexpected ", lexer_t::token_type_name(last_token));
+        }
+
+        if (expected != token_type::uninitialized)
+        {
+            error_msg += concat("; expected ", lexer_t::token_type_name(expected));
+        }
+
+        return error_msg;
+    }
+
+  private:
+    /// callback function
+    const parser_callback_t<BasicJsonType> callback = nullptr;
+    /// the type of the last read token
+    token_type last_token = token_type::uninitialized;
+    /// the lexer
+    lexer_t m_lexer;
+    /// whether to throw exceptions in case of errors
+    const bool allow_exceptions = true;
+};
+
+}  // namespace detail
+NLOHMANN_JSON_NAMESPACE_END
+
+// #include <nlohmann/detail/iterators/internal_iterator.hpp>
+//     __ _____ _____ _____
+//  __|  |   __|     |   | |  JSON for Modern C++
+// |  |  |__   |  |  | | | |  version 3.11.2
+// |_____|_____|_____|_|___|  https://github.com/nlohmann/json
+//
+// SPDX-FileCopyrightText: 2013-2022 Niels Lohmann <https://nlohmann.me>
+// SPDX-License-Identifier: MIT
+
+
+
+// #include <nlohmann/detail/abi_macros.hpp>
+
+// #include <nlohmann/detail/iterators/primitive_iterator.hpp>
+//     __ _____ _____ _____
+//  __|  |   __|     |   | |  JSON for Modern C++
+// |  |  |__   |  |  | | | |  version 3.11.2
+// |_____|_____|_____|_|___|  https://github.com/nlohmann/json
+//
+// SPDX-FileCopyrightText: 2013-2022 Niels Lohmann <https://nlohmann.me>
+// SPDX-License-Identifier: MIT
+
+
+
+#include <cstddef> // ptrdiff_t
+#include <limits>  // numeric_limits
+
+// #include <nlohmann/detail/macro_scope.hpp>
+
+
+NLOHMANN_JSON_NAMESPACE_BEGIN
+namespace detail
+{
+
+/*
+@brief an iterator for primitive JSON types
+
+This class models an iterator for primitive JSON types (boolean, number,
+string). It's only purpose is to allow the iterator/const_iterator classes
+to "iterate" over primitive values. Internally, the iterator is modeled by
+a `difference_type` variable. Value begin_value (`0`) models the begin,
+end_value (`1`) models past the end.
+*/
+class primitive_iterator_t
+{
+  private:
+    using difference_type = std::ptrdiff_t;
+    static constexpr difference_type begin_value = 0;
+    static constexpr difference_type end_value = begin_value + 1;
+
+  JSON_PRIVATE_UNLESS_TESTED:
+    /// iterator as signed integer type
+    difference_type m_it = (std::numeric_limits<std::ptrdiff_t>::min)();
+
+  public:
+    constexpr difference_type get_value() const noexcept
+    {
+        return m_it;
+    }
+
+    /// set iterator to a defined beginning
+    void set_begin() noexcept
+    {
+        m_it = begin_value;
+    }
+
+    /// set iterator to a defined past the end
+    void set_end() noexcept
+    {
+        m_it = end_value;
+    }
+
+    /// return whether the iterator can be dereferenced
+    constexpr bool is_begin() const noexcept
+    {
+        return m_it == begin_value;
+    }
+
+    /// return whether the iterator is at end
+    constexpr bool is_end() const noexcept
+    {
+        return m_it == end_value;
+    }
+
+    friend constexpr bool operator==(primitive_iterator_t lhs, primitive_iterator_t rhs) noexcept
+    {
+        return lhs.m_it == rhs.m_it;
+    }
+
+    friend constexpr bool operator<(primitive_iterator_t lhs, primitive_iterator_t rhs) noexcept
+    {
+        return lhs.m_it < rhs.m_it;
+    }
+
+    primitive_iterator_t operator+(difference_type n) noexcept
+    {
+        auto result = *this;
+        result += n;
+        return result;
+    }
+
+    friend constexpr difference_type operator-(primitive_iterator_t lhs, primitive_iterator_t rhs) noexcept
+    {
+        return lhs.m_it - rhs.m_it;
+    }
+
+    primitive_iterator_t& operator++() noexcept
+    {
+        ++m_it;
+        return *this;
+    }
+
+    primitive_iterator_t operator++(int)& noexcept // NOLINT(cert-dcl21-cpp)
+    {
+        auto result = *this;
+        ++m_it;
+        return result;
+    }
+
+    primitive_iterator_t& operator--() noexcept
+    {
+        --m_it;
+        return *this;
+    }
+
+    primitive_iterator_t operator--(int)& noexcept // NOLINT(cert-dcl21-cpp)
+    {
+        auto result = *this;
+        --m_it;
+        return result;
+    }
+
+    primitive_iterator_t& operator+=(difference_type n) noexcept
+    {
+        m_it += n;
+        return *this;
+    }
+
+    primitive_iterator_t& operator-=(difference_type n) noexcept
+    {
+        m_it -= n;
+        return *this;
+    }
+};
+
+}  // namespace detail
+NLOHMANN_JSON_NAMESPACE_END
+
+
+NLOHMANN_JSON_NAMESPACE_BEGIN
+namespace detail
+{
+
+/*!
+@brief an iterator value
+
+@note This structure could easily be a union, but MSVC currently does not allow
+unions members with complex constructors, see https://github.com/nlohmann/json/pull/105.
+*/
+template<typename BasicJsonType> struct internal_iterator
+{
+    /// iterator for JSON objects
+    typename BasicJsonType::object_t::iterator object_iterator {};
+    /// iterator for JSON arrays
+    typename BasicJsonType::array_t::iterator array_iterator {};
+    /// generic iterator for all other types
+    primitive_iterator_t primitive_iterator {};
+};
+
+}  // namespace detail
+NLOHMANN_JSON_NAMESPACE_END
+
+// #include <nlohmann/detail/iterators/iter_impl.hpp>
+//     __ _____ _____ _____
+//  __|  |   __|     |   | |  JSON for Modern C++
+// |  |  |__   |  |  | | | |  version 3.11.2
+// |_____|_____|_____|_|___|  https://github.com/nlohmann/json
+//
+// SPDX-FileCopyrightText: 2013-2022 Niels Lohmann <https://nlohmann.me>
+// SPDX-License-Identifier: MIT
+
+
+
+#include <iterator> // iterator, random_access_iterator_tag, bidirectional_iterator_tag, advance, next
+#include <type_traits> // conditional, is_const, remove_const
+
+// #include <nlohmann/detail/exceptions.hpp>
+
+// #include <nlohmann/detail/iterators/internal_iterator.hpp>
+
+// #include <nlohmann/detail/iterators/primitive_iterator.hpp>
+
+// #include <nlohmann/detail/macro_scope.hpp>
+
+// #include <nlohmann/detail/meta/cpp_future.hpp>
+
+// #include <nlohmann/detail/meta/type_traits.hpp>
+
+// #include <nlohmann/detail/value_t.hpp>
+
+
+NLOHMANN_JSON_NAMESPACE_BEGIN
+namespace detail
+{
+
+// forward declare, to be able to friend it later on
+template<typename IteratorType> class iteration_proxy;
+template<typename IteratorType> class iteration_proxy_value;
+
+/*!
+@brief a template for a bidirectional iterator for the @ref basic_json class
+This class implements a both iterators (iterator and const_iterator) for the
+@ref basic_json class.
+@note An iterator is called *initialized* when a pointer to a JSON value has
+      been set (e.g., by a constructor or a copy assignment). If the iterator is
+      default-constructed, it is *uninitialized* and most methods are undefined.
+      **The library uses assertions to detect calls on uninitialized iterators.**
+@requirement The class satisfies the following concept requirements:
+-
+[BidirectionalIterator](https://en.cppreference.com/w/cpp/named_req/BidirectionalIterator):
+  The iterator that can be moved can be moved in both directions (i.e.
+  incremented and decremented).
+@since version 1.0.0, simplified in version 2.0.9, change to bidirectional
+       iterators in version 3.0.0 (see https://github.com/nlohmann/json/issues/593)
+*/
+template<typename BasicJsonType>
+class iter_impl // NOLINT(cppcoreguidelines-special-member-functions,hicpp-special-member-functions)
+{
+    /// the iterator with BasicJsonType of different const-ness
+    using other_iter_impl = iter_impl<typename std::conditional<std::is_const<BasicJsonType>::value, typename std::remove_const<BasicJsonType>::type, const BasicJsonType>::type>;
+    /// allow basic_json to access private members
+    friend other_iter_impl;
+    friend BasicJsonType;
+    friend iteration_proxy<iter_impl>;
+    friend iteration_proxy_value<iter_impl>;
+
+    using object_t = typename BasicJsonType::object_t;
+    using array_t = typename BasicJsonType::array_t;
+    // make sure BasicJsonType is basic_json or const basic_json
+    static_assert(is_basic_json<typename std::remove_const<BasicJsonType>::type>::value,
+                  "iter_impl only accepts (const) basic_json");
+    // superficial check for the LegacyBidirectionalIterator named requirement
+    static_assert(std::is_base_of<std::bidirectional_iterator_tag, std::bidirectional_iterator_tag>::value
+                  &&  std::is_base_of<std::bidirectional_iterator_tag, typename std::iterator_traits<typename array_t::iterator>::iterator_category>::value,
+                  "basic_json iterator assumes array and object type iterators satisfy the LegacyBidirectionalIterator named requirement.");
+
+  public:
+    /// The std::iterator class template (used as a base class to provide typedefs) is deprecated in C++17.
+    /// The C++ Standard has never required user-defined iterators to derive from std::iterator.
+    /// A user-defined iterator should provide publicly accessible typedefs named
+    /// iterator_category, value_type, difference_type, pointer, and reference.
+    /// Note that value_type is required to be non-const, even for constant iterators.
+    using iterator_category = std::bidirectional_iterator_tag;
+
+    /// the type of the values when the iterator is dereferenced
+    using value_type = typename BasicJsonType::value_type;
+    /// a type to represent differences between iterators
+    using difference_type = typename BasicJsonType::difference_type;
+    /// defines a pointer to the type iterated over (value_type)
+    using pointer = typename std::conditional<std::is_const<BasicJsonType>::value,
+          typename BasicJsonType::const_pointer,
+          typename BasicJsonType::pointer>::type;
+    /// defines a reference to the type iterated over (value_type)
+    using reference =
+        typename std::conditional<std::is_const<BasicJsonType>::value,
+        typename BasicJsonType::const_reference,
+        typename BasicJsonType::reference>::type;
+
+    iter_impl() = default;
+    ~iter_impl() = default;
+    iter_impl(iter_impl&&) noexcept = default;
+    iter_impl& operator=(iter_impl&&) noexcept = default;
+
+    /*!
+    @brief constructor for a given JSON instance
+    @param[in] object  pointer to a JSON object for this iterator
+    @pre object != nullptr
+    @post The iterator is initialized; i.e. `m_object != nullptr`.
+    */
+    explicit iter_impl(pointer object) noexcept : m_object(object)
+    {
+        JSON_ASSERT(m_object != nullptr);
+
+        switch (m_object->m_type)
+        {
+            case value_t::object:
+            {
+                m_it.object_iterator = typename object_t::iterator();
+                break;
+            }
+
+            case value_t::array:
+            {
+                m_it.array_iterator = typename array_t::iterator();
+                break;
+            }
+
+            case value_t::null:
+            case value_t::string:
+            case value_t::boolean:
+            case value_t::number_integer:
+            case value_t::number_unsigned:
+            case value_t::number_float:
+            case value_t::binary:
+            case value_t::discarded:
+            default:
+            {
+                m_it.primitive_iterator = primitive_iterator_t();
+                break;
+            }
+        }
+    }
+
+    /*!
+    @note The conventional copy constructor and copy assignment are implicitly
+          defined. Combined with the following converting constructor and
+          assignment, they support: (1) copy from iterator to iterator, (2)
+          copy from const iterator to const iterator, and (3) conversion from
+          iterator to const iterator. However conversion from const iterator
+          to iterator is not defined.
+    */
+
+    /*!
+    @brief const copy constructor
+    @param[in] other const iterator to copy from
+    @note This copy constructor had to be defined explicitly to circumvent a bug
+          occurring on msvc v19.0 compiler (VS 2015) debug build. For more
+          information refer to: https://github.com/nlohmann/json/issues/1608
+    */
+    iter_impl(const iter_impl<const BasicJsonType>& other) noexcept
+        : m_object(other.m_object), m_it(other.m_it)
+    {}
+
+    /*!
+    @brief converting assignment
+    @param[in] other const iterator to copy from
+    @return const/non-const iterator
+    @note It is not checked whether @a other is initialized.
+    */
+    iter_impl& operator=(const iter_impl<const BasicJsonType>& other) noexcept
+    {
+        if (&other != this)
+        {
+            m_object = other.m_object;
+            m_it = other.m_it;
+        }
+        return *this;
+    }
+
+    /*!
+    @brief converting constructor
+    @param[in] other  non-const iterator to copy from
+    @note It is not checked whether @a other is initialized.
+    */
+    iter_impl(const iter_impl<typename std::remove_const<BasicJsonType>::type>& other) noexcept
+        : m_object(other.m_object), m_it(other.m_it)
+    {}
+
+    /*!
+    @brief converting assignment
+    @param[in] other  non-const iterator to copy from
+    @return const/non-const iterator
+    @note It is not checked whether @a other is initialized.
+    */
+    iter_impl& operator=(const iter_impl<typename std::remove_const<BasicJsonType>::type>& other) noexcept // NOLINT(cert-oop54-cpp)
+    {
+        m_object = other.m_object;
+        m_it = other.m_it;
+        return *this;
+    }
+
+  JSON_PRIVATE_UNLESS_TESTED:
+    /*!
+    @brief set the iterator to the first value
+    @pre The iterator is initialized; i.e. `m_object != nullptr`.
+    */
+    void set_begin() noexcept
+    {
+        JSON_ASSERT(m_object != nullptr);
+
+        switch (m_object->m_type)
+        {
+            case value_t::object:
+            {
+                m_it.object_iterator = m_object->m_value.object->begin();
+                break;
+            }
+
+            case value_t::array:
+            {
+                m_it.array_iterator = m_object->m_value.array->begin();
+                break;
+            }
+
+            case value_t::null:
+            {
+                // set to end so begin()==end() is true: null is empty
+                m_it.primitive_iterator.set_end();
+                break;
+            }
+
+            case value_t::string:
+            case value_t::boolean:
+            case value_t::number_integer:
+            case value_t::number_unsigned:
+            case value_t::number_float:
+            case value_t::binary:
+            case value_t::discarded:
+            default:
+            {
+                m_it.primitive_iterator.set_begin();
+                break;
+            }
+        }
+    }
+
+    /*!
+    @brief set the iterator past the last value
+    @pre The iterator is initialized; i.e. `m_object != nullptr`.
+    */
+    void set_end() noexcept
+    {
+        JSON_ASSERT(m_object != nullptr);
+
+        switch (m_object->m_type)
+        {
+            case value_t::object:
+            {
+                m_it.object_iterator = m_object->m_value.object->end();
+                break;
+            }
+
+            case value_t::array:
+            {
+                m_it.array_iterator = m_object->m_value.array->end();
+                break;
+            }
+
+            case value_t::null:
+            case value_t::string:
+            case value_t::boolean:
+            case value_t::number_integer:
+            case value_t::number_unsigned:
+            case value_t::number_float:
+            case value_t::binary:
+            case value_t::discarded:
+            default:
+            {
+                m_it.primitive_iterator.set_end();
+                break;
+            }
+        }
+    }
+
+  public:
+    /*!
+    @brief return a reference to the value pointed to by the iterator
+    @pre The iterator is initialized; i.e. `m_object != nullptr`.
+    */
+    reference operator*() const
+    {
+        JSON_ASSERT(m_object != nullptr);
+
+        switch (m_object->m_type)
+        {
+            case value_t::object:
+            {
+                JSON_ASSERT(m_it.object_iterator != m_object->m_value.object->end());
+                return m_it.object_iterator->second;
+            }
+
+            case value_t::array:
+            {
+                JSON_ASSERT(m_it.array_iterator != m_object->m_value.array->end());
+                return *m_it.array_iterator;
+            }
+
+            case value_t::null:
+                JSON_THROW(invalid_iterator::create(214, "cannot get value", m_object));
+
+            case value_t::string:
+            case value_t::boolean:
+            case value_t::number_integer:
+            case value_t::number_unsigned:
+            case value_t::number_float:
+            case value_t::binary:
+            case value_t::discarded:
+            default:
+            {
+                if (JSON_HEDLEY_LIKELY(m_it.primitive_iterator.is_begin()))
+                {
+                    return *m_object;
+                }
+
+                JSON_THROW(invalid_iterator::create(214, "cannot get value", m_object));
+            }
+        }
+    }
+
+    /*!
+    @brief dereference the iterator
+    @pre The iterator is initialized; i.e. `m_object != nullptr`.
+    */
+    pointer operator->() const
+    {
+        JSON_ASSERT(m_object != nullptr);
+
+        switch (m_object->m_type)
+        {
+            case value_t::object:
+            {
+                JSON_ASSERT(m_it.object_iterator != m_object->m_value.object->end());
+                return &(m_it.object_iterator->second);
+            }
+
+            case value_t::array:
+            {
+                JSON_ASSERT(m_it.array_iterator != m_object->m_value.array->end());
+                return &*m_it.array_iterator;
+            }
+
+            case value_t::null:
+            case value_t::string:
+            case value_t::boolean:
+            case value_t::number_integer:
+            case value_t::number_unsigned:
+            case value_t::number_float:
+            case value_t::binary:
+            case value_t::discarded:
+            default:
+            {
+                if (JSON_HEDLEY_LIKELY(m_it.primitive_iterator.is_begin()))
+                {
+                    return m_object;
+                }
+
+                JSON_THROW(invalid_iterator::create(214, "cannot get value", m_object));
+            }
+        }
+    }
+
+    /*!
+    @brief post-increment (it++)
+    @pre The iterator is initialized; i.e. `m_object != nullptr`.
+    */
+    iter_impl operator++(int)& // NOLINT(cert-dcl21-cpp)
+    {
+        auto result = *this;
+        ++(*this);
+        return result;
+    }
+
+    /*!
+    @brief pre-increment (++it)
+    @pre The iterator is initialized; i.e. `m_object != nullptr`.
+    */
+    iter_impl& operator++()
+    {
+        JSON_ASSERT(m_object != nullptr);
+
+        switch (m_object->m_type)
+        {
+            case value_t::object:
+            {
+                std::advance(m_it.object_iterator, 1);
+                break;
+            }
+
+            case value_t::array:
+            {
+                std::advance(m_it.array_iterator, 1);
+                break;
+            }
+
+            case value_t::null:
+            case value_t::string:
+            case value_t::boolean:
+            case value_t::number_integer:
+            case value_t::number_unsigned:
+            case value_t::number_float:
+            case value_t::binary:
+            case value_t::discarded:
+            default:
+            {
+                ++m_it.primitive_iterator;
+                break;
+            }
+        }
+
+        return *this;
+    }
+
+    /*!
+    @brief post-decrement (it--)
+    @pre The iterator is initialized; i.e. `m_object != nullptr`.
+    */
+    iter_impl operator--(int)& // NOLINT(cert-dcl21-cpp)
+    {
+        auto result = *this;
+        --(*this);
+        return result;
+    }
+
+    /*!
+    @brief pre-decrement (--it)
+    @pre The iterator is initialized; i.e. `m_object != nullptr`.
+    */
+    iter_impl& operator--()
+    {
+        JSON_ASSERT(m_object != nullptr);
+
+        switch (m_object->m_type)
+        {
+            case value_t::object:
+            {
+                std::advance(m_it.object_iterator, -1);
+                break;
+            }
+
+            case value_t::array:
+            {
+                std::advance(m_it.array_iterator, -1);
+                break;
+            }
+
+            case value_t::null:
+            case value_t::string:
+            case value_t::boolean:
+            case value_t::number_integer:
+            case value_t::number_unsigned:
+            case value_t::number_float:
+            case value_t::binary:
+            case value_t::discarded:
+            default:
+            {
+                --m_it.primitive_iterator;
+                break;
+            }
+        }
+
+        return *this;
+    }
+
+    /*!
+    @brief comparison: equal
+    @pre The iterator is initialized; i.e. `m_object != nullptr`.
+    */
+    template < typename IterImpl, detail::enable_if_t < (std::is_same<IterImpl, iter_impl>::value || std::is_same<IterImpl, other_iter_impl>::value), std::nullptr_t > = nullptr >
+    bool operator==(const IterImpl& other) const
+    {
+        // if objects are not the same, the comparison is undefined
+        if (JSON_HEDLEY_UNLIKELY(m_object != other.m_object))
+        {
+            JSON_THROW(invalid_iterator::create(212, "cannot compare iterators of different containers", m_object));
+        }
+
+        JSON_ASSERT(m_object != nullptr);
+
+        switch (m_object->m_type)
+        {
+            case value_t::object:
+                return (m_it.object_iterator == other.m_it.object_iterator);
+
+            case value_t::array:
+                return (m_it.array_iterator == other.m_it.array_iterator);
+
+            case value_t::null:
+            case value_t::string:
+            case value_t::boolean:
+            case value_t::number_integer:
+            case value_t::number_unsigned:
+            case value_t::number_float:
+            case value_t::binary:
+            case value_t::discarded:
+            default:
+                return (m_it.primitive_iterator == other.m_it.primitive_iterator);
+        }
+    }
+
+    /*!
+    @brief comparison: not equal
+    @pre The iterator is initialized; i.e. `m_object != nullptr`.
+    */
+    template < typename IterImpl, detail::enable_if_t < (std::is_same<IterImpl, iter_impl>::value || std::is_same<IterImpl, other_iter_impl>::value), std::nullptr_t > = nullptr >
+    bool operator!=(const IterImpl& other) const
+    {
+        return !operator==(other);
+    }
+
+    /*!
+    @brief comparison: smaller
+    @pre The iterator is initialized; i.e. `m_object != nullptr`.
+    */
+    bool operator<(const iter_impl& other) const
+    {
+        // if objects are not the same, the comparison is undefined
+        if (JSON_HEDLEY_UNLIKELY(m_object != other.m_object))
+        {
+            JSON_THROW(invalid_iterator::create(212, "cannot compare iterators of different containers", m_object));
+        }
+
+        JSON_ASSERT(m_object != nullptr);
+
+        switch (m_object->m_type)
+        {
+            case value_t::object:
+                JSON_THROW(invalid_iterator::create(213, "cannot compare order of object iterators", m_object));
+
+            case value_t::array:
+                return (m_it.array_iterator < other.m_it.array_iterator);
+
+            case value_t::null:
+            case value_t::string:
+            case value_t::boolean:
+            case value_t::number_integer:
+            case value_t::number_unsigned:
+            case value_t::number_float:
+            case value_t::binary:
+            case value_t::discarded:
+            default:
+                return (m_it.primitive_iterator < other.m_it.primitive_iterator);
+        }
+    }
+
+    /*!
+    @brief comparison: less than or equal
+    @pre The iterator is initialized; i.e. `m_object != nullptr`.
+    */
+    bool operator<=(const iter_impl& other) const
+    {
+        return !other.operator < (*this);
+    }
+
+    /*!
+    @brief comparison: greater than
+    @pre The iterator is initialized; i.e. `m_object != nullptr`.
+    */
+    bool operator>(const iter_impl& other) const
+    {
+        return !operator<=(other);
+    }
+
+    /*!
+    @brief comparison: greater than or equal
+    @pre The iterator is initialized; i.e. `m_object != nullptr`.
+    */
+    bool operator>=(const iter_impl& other) const
+    {
+        return !operator<(other);
+    }
+
+    /*!
+    @brief add to iterator
+    @pre The iterator is initialized; i.e. `m_object != nullptr`.
+    */
+    iter_impl& operator+=(difference_type i)
+    {
+        JSON_ASSERT(m_object != nullptr);
+
+        switch (m_object->m_type)
+        {
+            case value_t::object:
+                JSON_THROW(invalid_iterator::create(209, "cannot use offsets with object iterators", m_object));
+
+            case value_t::array:
+            {
+                std::advance(m_it.array_iterator, i);
+                break;
+            }
+
+            case value_t::null:
+            case value_t::string:
+            case value_t::boolean:
+            case value_t::number_integer:
+            case value_t::number_unsigned:
+            case value_t::number_float:
+            case value_t::binary:
+            case value_t::discarded:
+            default:
+            {
+                m_it.primitive_iterator += i;
+                break;
+            }
+        }
+
+        return *this;
+    }
+
+    /*!
+    @brief subtract from iterator
+    @pre The iterator is initialized; i.e. `m_object != nullptr`.
+    */
+    iter_impl& operator-=(difference_type i)
+    {
+        return operator+=(-i);
+    }
+
+    /*!
+    @brief add to iterator
+    @pre The iterator is initialized; i.e. `m_object != nullptr`.
+    */
+    iter_impl operator+(difference_type i) const
+    {
+        auto result = *this;
+        result += i;
+        return result;
+    }
+
+    /*!
+    @brief addition of distance and iterator
+    @pre The iterator is initialized; i.e. `m_object != nullptr`.
+    */
+    friend iter_impl operator+(difference_type i, const iter_impl& it)
+    {
+        auto result = it;
+        result += i;
+        return result;
+    }
+
+    /*!
+    @brief subtract from iterator
+    @pre The iterator is initialized; i.e. `m_object != nullptr`.
+    */
+    iter_impl operator-(difference_type i) const
+    {
+        auto result = *this;
+        result -= i;
+        return result;
+    }
+
+    /*!
+    @brief return difference
+    @pre The iterator is initialized; i.e. `m_object != nullptr`.
+    */
+    difference_type operator-(const iter_impl& other) const
+    {
+        JSON_ASSERT(m_object != nullptr);
+
+        switch (m_object->m_type)
+        {
+            case value_t::object:
+                JSON_THROW(invalid_iterator::create(209, "cannot use offsets with object iterators", m_object));
+
+            case value_t::array:
+                return m_it.array_iterator - other.m_it.array_iterator;
+
+            case value_t::null:
+            case value_t::string:
+            case value_t::boolean:
+            case value_t::number_integer:
+            case value_t::number_unsigned:
+            case value_t::number_float:
+            case value_t::binary:
+            case value_t::discarded:
+            default:
+                return m_it.primitive_iterator - other.m_it.primitive_iterator;
+        }
+    }
+
+    /*!
+    @brief access to successor
+    @pre The iterator is initialized; i.e. `m_object != nullptr`.
+    */
+    reference operator[](difference_type n) const
+    {
+        JSON_ASSERT(m_object != nullptr);
+
+        switch (m_object->m_type)
+        {
+            case value_t::object:
+                JSON_THROW(invalid_iterator::create(208, "cannot use operator[] for object iterators", m_object));
+
+            case value_t::array:
+                return *std::next(m_it.array_iterator, n);
+
+            case value_t::null:
+                JSON_THROW(invalid_iterator::create(214, "cannot get value", m_object));
+
+            case value_t::string:
+            case value_t::boolean:
+            case value_t::number_integer:
+            case value_t::number_unsigned:
+            case value_t::number_float:
+            case value_t::binary:
+            case value_t::discarded:
+            default:
+            {
+                if (JSON_HEDLEY_LIKELY(m_it.primitive_iterator.get_value() == -n))
+                {
+                    return *m_object;
+                }
+
+                JSON_THROW(invalid_iterator::create(214, "cannot get value", m_object));
+            }
+        }
+    }
+
+    /*!
+    @brief return the key of an object iterator
+    @pre The iterator is initialized; i.e. `m_object != nullptr`.
+    */
+    const typename object_t::key_type& key() const
+    {
+        JSON_ASSERT(m_object != nullptr);
+
+        if (JSON_HEDLEY_LIKELY(m_object->is_object()))
+        {
+            return m_it.object_iterator->first;
+        }
+
+        JSON_THROW(invalid_iterator::create(207, "cannot use key() for non-object iterators", m_object));
+    }
+
+    /*!
+    @brief return the value of an iterator
+    @pre The iterator is initialized; i.e. `m_object != nullptr`.
+    */
+    reference value() const
+    {
+        return operator*();
+    }
+
+  JSON_PRIVATE_UNLESS_TESTED:
+    /// associated JSON instance
+    pointer m_object = nullptr;
+    /// the actual iterator of the associated instance
+    internal_iterator<typename std::remove_const<BasicJsonType>::type> m_it {};
+};
+
+}  // namespace detail
+NLOHMANN_JSON_NAMESPACE_END
+
+// #include <nlohmann/detail/iterators/iteration_proxy.hpp>
+
+// #include <nlohmann/detail/iterators/json_reverse_iterator.hpp>
+//     __ _____ _____ _____
+//  __|  |   __|     |   | |  JSON for Modern C++
+// |  |  |__   |  |  | | | |  version 3.11.2
+// |_____|_____|_____|_|___|  https://github.com/nlohmann/json
+//
+// SPDX-FileCopyrightText: 2013-2022 Niels Lohmann <https://nlohmann.me>
+// SPDX-License-Identifier: MIT
+
+
+
+#include <cstddef> // ptrdiff_t
+#include <iterator> // reverse_iterator
+#include <utility> // declval
+
+// #include <nlohmann/detail/abi_macros.hpp>
+
+
+NLOHMANN_JSON_NAMESPACE_BEGIN
+namespace detail
+{
+
+//////////////////////
+// reverse_iterator //
+//////////////////////
+
+/*!
+@brief a template for a reverse iterator class
+
+@tparam Base the base iterator type to reverse. Valid types are @ref
+iterator (to create @ref reverse_iterator) and @ref const_iterator (to
+create @ref const_reverse_iterator).
+
+@requirement The class satisfies the following concept requirements:
+-
+[BidirectionalIterator](https://en.cppreference.com/w/cpp/named_req/BidirectionalIterator):
+  The iterator that can be moved can be moved in both directions (i.e.
+  incremented and decremented).
+- [OutputIterator](https://en.cppreference.com/w/cpp/named_req/OutputIterator):
+  It is possible to write to the pointed-to element (only if @a Base is
+  @ref iterator).
+
+@since version 1.0.0
+*/
+template<typename Base>
+class json_reverse_iterator : public std::reverse_iterator<Base>
+{
+  public:
+    using difference_type = std::ptrdiff_t;
+    /// shortcut to the reverse iterator adapter
+    using base_iterator = std::reverse_iterator<Base>;
+    /// the reference type for the pointed-to element
+    using reference = typename Base::reference;
+
+    /// create reverse iterator from iterator
+    explicit json_reverse_iterator(const typename base_iterator::iterator_type& it) noexcept
+        : base_iterator(it) {}
+
+    /// create reverse iterator from base class
+    explicit json_reverse_iterator(const base_iterator& it) noexcept : base_iterator(it) {}
+
+    /// post-increment (it++)
+    json_reverse_iterator operator++(int)& // NOLINT(cert-dcl21-cpp)
+    {
+        return static_cast<json_reverse_iterator>(base_iterator::operator++(1));
+    }
+
+    /// pre-increment (++it)
+    json_reverse_iterator& operator++()
+    {
+        return static_cast<json_reverse_iterator&>(base_iterator::operator++());
+    }
+
+    /// post-decrement (it--)
+    json_reverse_iterator operator--(int)& // NOLINT(cert-dcl21-cpp)
+    {
+        return static_cast<json_reverse_iterator>(base_iterator::operator--(1));
+    }
+
+    /// pre-decrement (--it)
+    json_reverse_iterator& operator--()
+    {
+        return static_cast<json_reverse_iterator&>(base_iterator::operator--());
+    }
+
+    /// add to iterator
+    json_reverse_iterator& operator+=(difference_type i)
+    {
+        return static_cast<json_reverse_iterator&>(base_iterator::operator+=(i));
+    }
+
+    /// add to iterator
+    json_reverse_iterator operator+(difference_type i) const
+    {
+        return static_cast<json_reverse_iterator>(base_iterator::operator+(i));
+    }
+
+    /// subtract from iterator
+    json_reverse_iterator operator-(difference_type i) const
+    {
+        return static_cast<json_reverse_iterator>(base_iterator::operator-(i));
+    }
+
+    /// return difference
+    difference_type operator-(const json_reverse_iterator& other) const
+    {
+        return base_iterator(*this) - base_iterator(other);
+    }
+
+    /// access to successor
+    reference operator[](difference_type n) const
+    {
+        return *(this->operator+(n));
+    }
+
+    /// return the key of an object iterator
+    auto key() const -> decltype(std::declval<Base>().key())
+    {
+        auto it = --this->base();
+        return it.key();
+    }
+
+    /// return the value of an iterator
+    reference value() const
+    {
+        auto it = --this->base();
+        return it.operator * ();
+    }
+};
+
+}  // namespace detail
+NLOHMANN_JSON_NAMESPACE_END
+
+// #include <nlohmann/detail/iterators/primitive_iterator.hpp>
+
+// #include <nlohmann/detail/json_pointer.hpp>
+//     __ _____ _____ _____
+//  __|  |   __|     |   | |  JSON for Modern C++
+// |  |  |__   |  |  | | | |  version 3.11.2
+// |_____|_____|_____|_|___|  https://github.com/nlohmann/json
+//
+// SPDX-FileCopyrightText: 2013-2022 Niels Lohmann <https://nlohmann.me>
+// SPDX-License-Identifier: MIT
+
+
+
+#include <algorithm> // all_of
+#include <cctype> // isdigit
+#include <cerrno> // errno, ERANGE
+#include <cstdlib> // strtoull
+#ifndef JSON_NO_IO
+    #include <iosfwd> // ostream
+#endif  // JSON_NO_IO
+#include <limits> // max
+#include <numeric> // accumulate
+#include <string> // string
+#include <utility> // move
+#include <vector> // vector
+
+// #include <nlohmann/detail/exceptions.hpp>
+
+// #include <nlohmann/detail/macro_scope.hpp>
+
+// #include <nlohmann/detail/string_concat.hpp>
+
+// #include <nlohmann/detail/string_escape.hpp>
+
+// #include <nlohmann/detail/value_t.hpp>
+
+
+NLOHMANN_JSON_NAMESPACE_BEGIN
+
+/// @brief JSON Pointer defines a string syntax for identifying a specific value within a JSON document
+/// @sa https://json.nlohmann.me/api/json_pointer/
+template<typename RefStringType>
+class json_pointer
+{
+    // allow basic_json to access private members
+    NLOHMANN_BASIC_JSON_TPL_DECLARATION
+    friend class basic_json;
+
+    template<typename>
+    friend class json_pointer;
+
+    template<typename T>
+    struct string_t_helper
+    {
+        using type = T;
+    };
+
+    NLOHMANN_BASIC_JSON_TPL_DECLARATION
+    struct string_t_helper<NLOHMANN_BASIC_JSON_TPL>
+    {
+        using type = StringType;
+    };
+
+  public:
+    // for backwards compatibility accept BasicJsonType
+    using string_t = typename string_t_helper<RefStringType>::type;
+
+    /// @brief create JSON pointer
+    /// @sa https://json.nlohmann.me/api/json_pointer/json_pointer/
+    explicit json_pointer(const string_t& s = "")
+        : reference_tokens(split(s))
+    {}
+
+    /// @brief return a string representation of the JSON pointer
+    /// @sa https://json.nlohmann.me/api/json_pointer/to_string/
+    string_t to_string() const
+    {
+        return std::accumulate(reference_tokens.begin(), reference_tokens.end(),
+                               string_t{},
+                               [](const string_t& a, const string_t& b)
+        {
+            return detail::concat(a, '/', detail::escape(b));
+        });
+    }
+
+    /// @brief return a string representation of the JSON pointer
+    /// @sa https://json.nlohmann.me/api/json_pointer/operator_string/
+    JSON_HEDLEY_DEPRECATED_FOR(3.11.0, to_string())
+    operator string_t() const
+    {
+        return to_string();
+    }
+
+#ifndef JSON_NO_IO
+    /// @brief write string representation of the JSON pointer to stream
+    /// @sa https://json.nlohmann.me/api/basic_json/operator_ltlt/
+    friend std::ostream& operator<<(std::ostream& o, const json_pointer& ptr)
+    {
+        o << ptr.to_string();
+        return o;
+    }
+#endif
+
+    /// @brief append another JSON pointer at the end of this JSON pointer
+    /// @sa https://json.nlohmann.me/api/json_pointer/operator_slasheq/
+    json_pointer& operator/=(const json_pointer& ptr)
+    {
+        reference_tokens.insert(reference_tokens.end(),
+                                ptr.reference_tokens.begin(),
+                                ptr.reference_tokens.end());
+        return *this;
+    }
+
+    /// @brief append an unescaped reference token at the end of this JSON pointer
+    /// @sa https://json.nlohmann.me/api/json_pointer/operator_slasheq/
+    json_pointer& operator/=(string_t token)
+    {
+        push_back(std::move(token));
+        return *this;
+    }
+
+    /// @brief append an array index at the end of this JSON pointer
+    /// @sa https://json.nlohmann.me/api/json_pointer/operator_slasheq/
+    json_pointer& operator/=(std::size_t array_idx)
+    {
+        return *this /= std::to_string(array_idx);
+    }
+
+    /// @brief create a new JSON pointer by appending the right JSON pointer at the end of the left JSON pointer
+    /// @sa https://json.nlohmann.me/api/json_pointer/operator_slash/
+    friend json_pointer operator/(const json_pointer& lhs,
+                                  const json_pointer& rhs)
+    {
+        return json_pointer(lhs) /= rhs;
+    }
+
+    /// @brief create a new JSON pointer by appending the unescaped token at the end of the JSON pointer
+    /// @sa https://json.nlohmann.me/api/json_pointer/operator_slash/
+    friend json_pointer operator/(const json_pointer& lhs, string_t token) // NOLINT(performance-unnecessary-value-param)
+    {
+        return json_pointer(lhs) /= std::move(token);
+    }
+
+    /// @brief create a new JSON pointer by appending the array-index-token at the end of the JSON pointer
+    /// @sa https://json.nlohmann.me/api/json_pointer/operator_slash/
+    friend json_pointer operator/(const json_pointer& lhs, std::size_t array_idx)
+    {
+        return json_pointer(lhs) /= array_idx;
+    }
+
+    /// @brief returns the parent of this JSON pointer
+    /// @sa https://json.nlohmann.me/api/json_pointer/parent_pointer/
+    json_pointer parent_pointer() const
+    {
+        if (empty())
+        {
+            return *this;
+        }
+
+        json_pointer res = *this;
+        res.pop_back();
+        return res;
+    }
+
+    /// @brief remove last reference token
+    /// @sa https://json.nlohmann.me/api/json_pointer/pop_back/
+    void pop_back()
+    {
+        if (JSON_HEDLEY_UNLIKELY(empty()))
+        {
+            JSON_THROW(detail::out_of_range::create(405, "JSON pointer has no parent", nullptr));
+        }
+
+        reference_tokens.pop_back();
+    }
+
+    /// @brief return last reference token
+    /// @sa https://json.nlohmann.me/api/json_pointer/back/
+    const string_t& back() const
+    {
+        if (JSON_HEDLEY_UNLIKELY(empty()))
+        {
+            JSON_THROW(detail::out_of_range::create(405, "JSON pointer has no parent", nullptr));
+        }
+
+        return reference_tokens.back();
+    }
+
+    /// @brief append an unescaped token at the end of the reference pointer
+    /// @sa https://json.nlohmann.me/api/json_pointer/push_back/
+    void push_back(const string_t& token)
+    {
+        reference_tokens.push_back(token);
+    }
+
+    /// @brief append an unescaped token at the end of the reference pointer
+    /// @sa https://json.nlohmann.me/api/json_pointer/push_back/
+    void push_back(string_t&& token)
+    {
+        reference_tokens.push_back(std::move(token));
+    }
+
+    /// @brief return whether pointer points to the root document
+    /// @sa https://json.nlohmann.me/api/json_pointer/empty/
+    bool empty() const noexcept
+    {
+        return reference_tokens.empty();
+    }
+
+  private:
+    /*!
+    @param[in] s  reference token to be converted into an array index
+
+    @return integer representation of @a s
+
+    @throw parse_error.106  if an array index begins with '0'
+    @throw parse_error.109  if an array index begins not with a digit
+    @throw out_of_range.404 if string @a s could not be converted to an integer
+    @throw out_of_range.410 if an array index exceeds size_type
+    */
+    template<typename BasicJsonType>
+    static typename BasicJsonType::size_type array_index(const string_t& s)
+    {
+        using size_type = typename BasicJsonType::size_type;
+
+        // error condition (cf. RFC 6901, Sect. 4)
+        if (JSON_HEDLEY_UNLIKELY(s.size() > 1 && s[0] == '0'))
+        {
+            JSON_THROW(detail::parse_error::create(106, 0, detail::concat("array index '", s, "' must not begin with '0'"), nullptr));
+        }
+
+        // error condition (cf. RFC 6901, Sect. 4)
+        if (JSON_HEDLEY_UNLIKELY(s.size() > 1 && !(s[0] >= '1' && s[0] <= '9')))
+        {
+            JSON_THROW(detail::parse_error::create(109, 0, detail::concat("array index '", s, "' is not a number"), nullptr));
+        }
+
+        const char* p = s.c_str();
+        char* p_end = nullptr;
+        errno = 0; // strtoull doesn't reset errno
+        unsigned long long res = std::strtoull(p, &p_end, 10); // NOLINT(runtime/int)
+        if (p == p_end // invalid input or empty string
+                || errno == ERANGE // out of range
+                || JSON_HEDLEY_UNLIKELY(static_cast<std::size_t>(p_end - p) != s.size())) // incomplete read
+        {
+            JSON_THROW(detail::out_of_range::create(404, detail::concat("unresolved reference token '", s, "'"), nullptr));
+        }
+
+        // only triggered on special platforms (like 32bit), see also
+        // https://github.com/nlohmann/json/pull/2203
+        if (res >= static_cast<unsigned long long>((std::numeric_limits<size_type>::max)()))  // NOLINT(runtime/int)
+        {
+            JSON_THROW(detail::out_of_range::create(410, detail::concat("array index ", s, " exceeds size_type"), nullptr));   // LCOV_EXCL_LINE
+        }
+
+        return static_cast<size_type>(res);
+    }
+
+  JSON_PRIVATE_UNLESS_TESTED:
+    json_pointer top() const
+    {
+        if (JSON_HEDLEY_UNLIKELY(empty()))
+        {
+            JSON_THROW(detail::out_of_range::create(405, "JSON pointer has no parent", nullptr));
+        }
+
+        json_pointer result = *this;
+        result.reference_tokens = {reference_tokens[0]};
+        return result;
+    }
+
+  private:
+    /*!
+    @brief create and return a reference to the pointed to value
+
+    @complexity Linear in the number of reference tokens.
+
+    @throw parse_error.109 if array index is not a number
+    @throw type_error.313 if value cannot be unflattened
+    */
+    template<typename BasicJsonType>
+    BasicJsonType& get_and_create(BasicJsonType& j) const
+    {
+        auto* result = &j;
+
+        // in case no reference tokens exist, return a reference to the JSON value
+        // j which will be overwritten by a primitive value
+        for (const auto& reference_token : reference_tokens)
+        {
+            switch (result->type())
+            {
+                case detail::value_t::null:
+                {
+                    if (reference_token == "0")
+                    {
+                        // start a new array if reference token is 0
+                        result = &result->operator[](0);
+                    }
+                    else
+                    {
+                        // start a new object otherwise
+                        result = &result->operator[](reference_token);
+                    }
+                    break;
+                }
+
+                case detail::value_t::object:
+                {
+                    // create an entry in the object
+                    result = &result->operator[](reference_token);
+                    break;
+                }
+
+                case detail::value_t::array:
+                {
+                    // create an entry in the array
+                    result = &result->operator[](array_index<BasicJsonType>(reference_token));
+                    break;
+                }
+
+                /*
+                The following code is only reached if there exists a reference
+                token _and_ the current value is primitive. In this case, we have
+                an error situation, because primitive values may only occur as
+                single value; that is, with an empty list of reference tokens.
+                */
+                case detail::value_t::string:
+                case detail::value_t::boolean:
+                case detail::value_t::number_integer:
+                case detail::value_t::number_unsigned:
+                case detail::value_t::number_float:
+                case detail::value_t::binary:
+                case detail::value_t::discarded:
+                default:
+                    JSON_THROW(detail::type_error::create(313, "invalid value to unflatten", &j));
+            }
+        }
+
+        return *result;
+    }
+
+    /*!
+    @brief return a reference to the pointed to value
+
+    @note This version does not throw if a value is not present, but tries to
+          create nested values instead. For instance, calling this function
+          with pointer `"/this/that"` on a null value is equivalent to calling
+          `operator[]("this").operator[]("that")` on that value, effectively
+          changing the null value to an object.
+
+    @param[in] ptr  a JSON value
+
+    @return reference to the JSON value pointed to by the JSON pointer
+
+    @complexity Linear in the length of the JSON pointer.
+
+    @throw parse_error.106   if an array index begins with '0'
+    @throw parse_error.109   if an array index was not a number
+    @throw out_of_range.404  if the JSON pointer can not be resolved
+    */
+    template<typename BasicJsonType>
+    BasicJsonType& get_unchecked(BasicJsonType* ptr) const
+    {
+        for (const auto& reference_token : reference_tokens)
+        {
+            // convert null values to arrays or objects before continuing
+            if (ptr->is_null())
+            {
+                // check if reference token is a number
+                const bool nums =
+                    std::all_of(reference_token.begin(), reference_token.end(),
+                                [](const unsigned char x)
+                {
+                    return std::isdigit(x);
+                });
+
+                // change value to array for numbers or "-" or to object otherwise
+                *ptr = (nums || reference_token == "-")
+                       ? detail::value_t::array
+                       : detail::value_t::object;
+            }
+
+            switch (ptr->type())
+            {
+                case detail::value_t::object:
+                {
+                    // use unchecked object access
+                    ptr = &ptr->operator[](reference_token);
+                    break;
+                }
+
+                case detail::value_t::array:
+                {
+                    if (reference_token == "-")
+                    {
+                        // explicitly treat "-" as index beyond the end
+                        ptr = &ptr->operator[](ptr->m_value.array->size());
+                    }
+                    else
+                    {
+                        // convert array index to number; unchecked access
+                        ptr = &ptr->operator[](array_index<BasicJsonType>(reference_token));
+                    }
+                    break;
+                }
+
+                case detail::value_t::null:
+                case detail::value_t::string:
+                case detail::value_t::boolean:
+                case detail::value_t::number_integer:
+                case detail::value_t::number_unsigned:
+                case detail::value_t::number_float:
+                case detail::value_t::binary:
+                case detail::value_t::discarded:
+                default:
+                    JSON_THROW(detail::out_of_range::create(404, detail::concat("unresolved reference token '", reference_token, "'"), ptr));
+            }
+        }
+
+        return *ptr;
+    }
+
+    /*!
+    @throw parse_error.106   if an array index begins with '0'
+    @throw parse_error.109   if an array index was not a number
+    @throw out_of_range.402  if the array index '-' is used
+    @throw out_of_range.404  if the JSON pointer can not be resolved
+    */
+    template<typename BasicJsonType>
+    BasicJsonType& get_checked(BasicJsonType* ptr) const
+    {
+        for (const auto& reference_token : reference_tokens)
+        {
+            switch (ptr->type())
+            {
+                case detail::value_t::object:
+                {
+                    // note: at performs range check
+                    ptr = &ptr->at(reference_token);
+                    break;
+                }
+
+                case detail::value_t::array:
+                {
+                    if (JSON_HEDLEY_UNLIKELY(reference_token == "-"))
+                    {
+                        // "-" always fails the range check
+                        JSON_THROW(detail::out_of_range::create(402, detail::concat(
+                                "array index '-' (", std::to_string(ptr->m_value.array->size()),
+                                ") is out of range"), ptr));
+                    }
+
+                    // note: at performs range check
+                    ptr = &ptr->at(array_index<BasicJsonType>(reference_token));
+                    break;
+                }
+
+                case detail::value_t::null:
+                case detail::value_t::string:
+                case detail::value_t::boolean:
+                case detail::value_t::number_integer:
+                case detail::value_t::number_unsigned:
+                case detail::value_t::number_float:
+                case detail::value_t::binary:
+                case detail::value_t::discarded:
+                default:
+                    JSON_THROW(detail::out_of_range::create(404, detail::concat("unresolved reference token '", reference_token, "'"), ptr));
+            }
+        }
+
+        return *ptr;
+    }
+
+    /*!
+    @brief return a const reference to the pointed to value
+
+    @param[in] ptr  a JSON value
+
+    @return const reference to the JSON value pointed to by the JSON
+    pointer
+
+    @throw parse_error.106   if an array index begins with '0'
+    @throw parse_error.109   if an array index was not a number
+    @throw out_of_range.402  if the array index '-' is used
+    @throw out_of_range.404  if the JSON pointer can not be resolved
+    */
+    template<typename BasicJsonType>
+    const BasicJsonType& get_unchecked(const BasicJsonType* ptr) const
+    {
+        for (const auto& reference_token : reference_tokens)
+        {
+            switch (ptr->type())
+            {
+                case detail::value_t::object:
+                {
+                    // use unchecked object access
+                    ptr = &ptr->operator[](reference_token);
+                    break;
+                }
+
+                case detail::value_t::array:
+                {
+                    if (JSON_HEDLEY_UNLIKELY(reference_token == "-"))
+                    {
+                        // "-" cannot be used for const access
+                        JSON_THROW(detail::out_of_range::create(402, detail::concat("array index '-' (", std::to_string(ptr->m_value.array->size()), ") is out of range"), ptr));
+                    }
+
+                    // use unchecked array access
+                    ptr = &ptr->operator[](array_index<BasicJsonType>(reference_token));
+                    break;
+                }
+
+                case detail::value_t::null:
+                case detail::value_t::string:
+                case detail::value_t::boolean:
+                case detail::value_t::number_integer:
+                case detail::value_t::number_unsigned:
+                case detail::value_t::number_float:
+                case detail::value_t::binary:
+                case detail::value_t::discarded:
+                default:
+                    JSON_THROW(detail::out_of_range::create(404, detail::concat("unresolved reference token '", reference_token, "'"), ptr));
+            }
+        }
+
+        return *ptr;
+    }
+
+    /*!
+    @throw parse_error.106   if an array index begins with '0'
+    @throw parse_error.109   if an array index was not a number
+    @throw out_of_range.402  if the array index '-' is used
+    @throw out_of_range.404  if the JSON pointer can not be resolved
+    */
+    template<typename BasicJsonType>
+    const BasicJsonType& get_checked(const BasicJsonType* ptr) const
+    {
+        for (const auto& reference_token : reference_tokens)
+        {
+            switch (ptr->type())
+            {
+                case detail::value_t::object:
+                {
+                    // note: at performs range check
+                    ptr = &ptr->at(reference_token);
+                    break;
+                }
+
+                case detail::value_t::array:
+                {
+                    if (JSON_HEDLEY_UNLIKELY(reference_token == "-"))
+                    {
+                        // "-" always fails the range check
+                        JSON_THROW(detail::out_of_range::create(402, detail::concat(
+                                "array index '-' (", std::to_string(ptr->m_value.array->size()),
+                                ") is out of range"), ptr));
+                    }
+
+                    // note: at performs range check
+                    ptr = &ptr->at(array_index<BasicJsonType>(reference_token));
+                    break;
+                }
+
+                case detail::value_t::null:
+                case detail::value_t::string:
+                case detail::value_t::boolean:
+                case detail::value_t::number_integer:
+                case detail::value_t::number_unsigned:
+                case detail::value_t::number_float:
+                case detail::value_t::binary:
+                case detail::value_t::discarded:
+                default:
+                    JSON_THROW(detail::out_of_range::create(404, detail::concat("unresolved reference token '", reference_token, "'"), ptr));
+            }
+        }
+
+        return *ptr;
+    }
+
+    /*!
+    @throw parse_error.106   if an array index begins with '0'
+    @throw parse_error.109   if an array index was not a number
+    */
+    template<typename BasicJsonType>
+    bool contains(const BasicJsonType* ptr) const
+    {
+        for (const auto& reference_token : reference_tokens)
+        {
+            switch (ptr->type())
+            {
+                case detail::value_t::object:
+                {
+                    if (!ptr->contains(reference_token))
+                    {
+                        // we did not find the key in the object
+                        return false;
+                    }
+
+                    ptr = &ptr->operator[](reference_token);
+                    break;
+                }
+
+                case detail::value_t::array:
+                {
+                    if (JSON_HEDLEY_UNLIKELY(reference_token == "-"))
+                    {
+                        // "-" always fails the range check
+                        return false;
+                    }
+                    if (JSON_HEDLEY_UNLIKELY(reference_token.size() == 1 && !("0" <= reference_token && reference_token <= "9")))
+                    {
+                        // invalid char
+                        return false;
+                    }
+                    if (JSON_HEDLEY_UNLIKELY(reference_token.size() > 1))
+                    {
+                        if (JSON_HEDLEY_UNLIKELY(!('1' <= reference_token[0] && reference_token[0] <= '9')))
+                        {
+                            // first char should be between '1' and '9'
+                            return false;
+                        }
+                        for (std::size_t i = 1; i < reference_token.size(); i++)
+                        {
+                            if (JSON_HEDLEY_UNLIKELY(!('0' <= reference_token[i] && reference_token[i] <= '9')))
+                            {
+                                // other char should be between '0' and '9'
+                                return false;
+                            }
+                        }
+                    }
+
+                    const auto idx = array_index<BasicJsonType>(reference_token);
+                    if (idx >= ptr->size())
+                    {
+                        // index out of range
+                        return false;
+                    }
+
+                    ptr = &ptr->operator[](idx);
+                    break;
+                }
+
+                case detail::value_t::null:
+                case detail::value_t::string:
+                case detail::value_t::boolean:
+                case detail::value_t::number_integer:
+                case detail::value_t::number_unsigned:
+                case detail::value_t::number_float:
+                case detail::value_t::binary:
+                case detail::value_t::discarded:
+                default:
+                {
+                    // we do not expect primitive values if there is still a
+                    // reference token to process
+                    return false;
+                }
+            }
+        }
+
+        // no reference token left means we found a primitive value
+        return true;
+    }
+
+    /*!
+    @brief split the string input to reference tokens
+
+    @note This function is only called by the json_pointer constructor.
+          All exceptions below are documented there.
+
+    @throw parse_error.107  if the pointer is not empty or begins with '/'
+    @throw parse_error.108  if character '~' is not followed by '0' or '1'
+    */
+    static std::vector<string_t> split(const string_t& reference_string)
+    {
+        std::vector<string_t> result;
+
+        // special case: empty reference string -> no reference tokens
+        if (reference_string.empty())
+        {
+            return result;
+        }
+
+        // check if nonempty reference string begins with slash
+        if (JSON_HEDLEY_UNLIKELY(reference_string[0] != '/'))
+        {
+            JSON_THROW(detail::parse_error::create(107, 1, detail::concat("JSON pointer must be empty or begin with '/' - was: '", reference_string, "'"), nullptr));
+        }
+
+        // extract the reference tokens:
+        // - slash: position of the last read slash (or end of string)
+        // - start: position after the previous slash
+        for (
+            // search for the first slash after the first character
+            std::size_t slash = reference_string.find_first_of('/', 1),
+            // set the beginning of the first reference token
+            start = 1;
+            // we can stop if start == 0 (if slash == string_t::npos)
+            start != 0;
+            // set the beginning of the next reference token
+            // (will eventually be 0 if slash == string_t::npos)
+            start = (slash == string_t::npos) ? 0 : slash + 1,
+            // find next slash
+            slash = reference_string.find_first_of('/', start))
+        {
+            // use the text between the beginning of the reference token
+            // (start) and the last slash (slash).
+            auto reference_token = reference_string.substr(start, slash - start);
+
+            // check reference tokens are properly escaped
+            for (std::size_t pos = reference_token.find_first_of('~');
+                    pos != string_t::npos;
+                    pos = reference_token.find_first_of('~', pos + 1))
+            {
+                JSON_ASSERT(reference_token[pos] == '~');
+
+                // ~ must be followed by 0 or 1
+                if (JSON_HEDLEY_UNLIKELY(pos == reference_token.size() - 1 ||
+                                         (reference_token[pos + 1] != '0' &&
+                                          reference_token[pos + 1] != '1')))
+                {
+                    JSON_THROW(detail::parse_error::create(108, 0, "escape character '~' must be followed with '0' or '1'", nullptr));
+                }
+            }
+
+            // finally, store the reference token
+            detail::unescape(reference_token);
+            result.push_back(reference_token);
+        }
+
+        return result;
+    }
+
+  private:
+    /*!
+    @param[in] reference_string  the reference string to the current value
+    @param[in] value             the value to consider
+    @param[in,out] result        the result object to insert values to
+
+    @note Empty objects or arrays are flattened to `null`.
+    */
+    template<typename BasicJsonType>
+    static void flatten(const string_t& reference_string,
+                        const BasicJsonType& value,
+                        BasicJsonType& result)
+    {
+        switch (value.type())
+        {
+            case detail::value_t::array:
+            {
+                if (value.m_value.array->empty())
+                {
+                    // flatten empty array as null
+                    result[reference_string] = nullptr;
+                }
+                else
+                {
+                    // iterate array and use index as reference string
+                    for (std::size_t i = 0; i < value.m_value.array->size(); ++i)
+                    {
+                        flatten(detail::concat(reference_string, '/', std::to_string(i)),
+                                value.m_value.array->operator[](i), result);
+                    }
+                }
+                break;
+            }
+
+            case detail::value_t::object:
+            {
+                if (value.m_value.object->empty())
+                {
+                    // flatten empty object as null
+                    result[reference_string] = nullptr;
+                }
+                else
+                {
+                    // iterate object and use keys as reference string
+                    for (const auto& element : *value.m_value.object)
+                    {
+                        flatten(detail::concat(reference_string, '/', detail::escape(element.first)), element.second, result);
+                    }
+                }
+                break;
+            }
+
+            case detail::value_t::null:
+            case detail::value_t::string:
+            case detail::value_t::boolean:
+            case detail::value_t::number_integer:
+            case detail::value_t::number_unsigned:
+            case detail::value_t::number_float:
+            case detail::value_t::binary:
+            case detail::value_t::discarded:
+            default:
+            {
+                // add primitive value with its reference string
+                result[reference_string] = value;
+                break;
+            }
+        }
+    }
+
+    /*!
+    @param[in] value  flattened JSON
+
+    @return unflattened JSON
+
+    @throw parse_error.109 if array index is not a number
+    @throw type_error.314  if value is not an object
+    @throw type_error.315  if object values are not primitive
+    @throw type_error.313  if value cannot be unflattened
+    */
+    template<typename BasicJsonType>
+    static BasicJsonType
+    unflatten(const BasicJsonType& value)
+    {
+        if (JSON_HEDLEY_UNLIKELY(!value.is_object()))
+        {
+            JSON_THROW(detail::type_error::create(314, "only objects can be unflattened", &value));
+        }
+
+        BasicJsonType result;
+
+        // iterate the JSON object values
+        for (const auto& element : *value.m_value.object)
+        {
+            if (JSON_HEDLEY_UNLIKELY(!element.second.is_primitive()))
+            {
+                JSON_THROW(detail::type_error::create(315, "values in object must be primitive", &element.second));
+            }
+
+            // assign value to reference pointed to by JSON pointer; Note that if
+            // the JSON pointer is "" (i.e., points to the whole value), function
+            // get_and_create returns a reference to result itself. An assignment
+            // will then create a primitive value.
+            json_pointer(element.first).get_and_create(result) = element.second;
+        }
+
+        return result;
+    }
+
+    // can't use conversion operator because of ambiguity
+    json_pointer<string_t> convert() const&
+    {
+        json_pointer<string_t> result;
+        result.reference_tokens = reference_tokens;
+        return result;
+    }
+
+    json_pointer<string_t> convert()&&
+    {
+        json_pointer<string_t> result;
+        result.reference_tokens = std::move(reference_tokens);
+        return result;
+    }
+
+  public:
+#if JSON_HAS_THREE_WAY_COMPARISON
+    /// @brief compares two JSON pointers for equality
+    /// @sa https://json.nlohmann.me/api/json_pointer/operator_eq/
+    template<typename RefStringTypeRhs>
+    bool operator==(const json_pointer<RefStringTypeRhs>& rhs) const noexcept
+    {
+        return reference_tokens == rhs.reference_tokens;
+    }
+
+    /// @brief compares JSON pointer and string for equality
+    /// @sa https://json.nlohmann.me/api/json_pointer/operator_eq/
+    JSON_HEDLEY_DEPRECATED_FOR(3.11.2, operator==(json_pointer))
+    bool operator==(const string_t& rhs) const
+    {
+        return *this == json_pointer(rhs);
+    }
+
+    /// @brief 3-way compares two JSON pointers
+    template<typename RefStringTypeRhs>
+    std::strong_ordering operator<=>(const json_pointer<RefStringTypeRhs>& rhs) const noexcept // *NOPAD*
+    {
+        return  reference_tokens <=> rhs.reference_tokens; // *NOPAD*
+    }
+#else
+    /// @brief compares two JSON pointers for equality
+    /// @sa https://json.nlohmann.me/api/json_pointer/operator_eq/
+    template<typename RefStringTypeLhs, typename RefStringTypeRhs>
+    // NOLINTNEXTLINE(readability-redundant-declaration)
+    friend bool operator==(const json_pointer<RefStringTypeLhs>& lhs,
+                           const json_pointer<RefStringTypeRhs>& rhs) noexcept;
+
+    /// @brief compares JSON pointer and string for equality
+    /// @sa https://json.nlohmann.me/api/json_pointer/operator_eq/
+    template<typename RefStringTypeLhs, typename StringType>
+    // NOLINTNEXTLINE(readability-redundant-declaration)
+    friend bool operator==(const json_pointer<RefStringTypeLhs>& lhs,
+                           const StringType& rhs);
+
+    /// @brief compares string and JSON pointer for equality
+    /// @sa https://json.nlohmann.me/api/json_pointer/operator_eq/
+    template<typename RefStringTypeRhs, typename StringType>
+    // NOLINTNEXTLINE(readability-redundant-declaration)
+    friend bool operator==(const StringType& lhs,
+                           const json_pointer<RefStringTypeRhs>& rhs);
+
+    /// @brief compares two JSON pointers for inequality
+    /// @sa https://json.nlohmann.me/api/json_pointer/operator_ne/
+    template<typename RefStringTypeLhs, typename RefStringTypeRhs>
+    // NOLINTNEXTLINE(readability-redundant-declaration)
+    friend bool operator!=(const json_pointer<RefStringTypeLhs>& lhs,
+                           const json_pointer<RefStringTypeRhs>& rhs) noexcept;
+
+    /// @brief compares JSON pointer and string for inequality
+    /// @sa https://json.nlohmann.me/api/json_pointer/operator_ne/
+    template<typename RefStringTypeLhs, typename StringType>
+    // NOLINTNEXTLINE(readability-redundant-declaration)
+    friend bool operator!=(const json_pointer<RefStringTypeLhs>& lhs,
+                           const StringType& rhs);
+
+    /// @brief compares string and JSON pointer for inequality
+    /// @sa https://json.nlohmann.me/api/json_pointer/operator_ne/
+    template<typename RefStringTypeRhs, typename StringType>
+    // NOLINTNEXTLINE(readability-redundant-declaration)
+    friend bool operator!=(const StringType& lhs,
+                           const json_pointer<RefStringTypeRhs>& rhs);
+
+    /// @brief compares two JSON pointer for less-than
+    template<typename RefStringTypeLhs, typename RefStringTypeRhs>
+    // NOLINTNEXTLINE(readability-redundant-declaration)
+    friend bool operator<(const json_pointer<RefStringTypeLhs>& lhs,
+                          const json_pointer<RefStringTypeRhs>& rhs) noexcept;
+#endif
+
+  private:
+    /// the reference tokens
+    std::vector<string_t> reference_tokens;
+};
+
+#if !JSON_HAS_THREE_WAY_COMPARISON
+// functions cannot be defined inside class due to ODR violations
+template<typename RefStringTypeLhs, typename RefStringTypeRhs>
+inline bool operator==(const json_pointer<RefStringTypeLhs>& lhs,
+                       const json_pointer<RefStringTypeRhs>& rhs) noexcept
+{
+    return lhs.reference_tokens == rhs.reference_tokens;
+}
+
+template<typename RefStringTypeLhs,
+         typename StringType = typename json_pointer<RefStringTypeLhs>::string_t>
+JSON_HEDLEY_DEPRECATED_FOR(3.11.2, operator==(json_pointer, json_pointer))
+inline bool operator==(const json_pointer<RefStringTypeLhs>& lhs,
+                       const StringType& rhs)
+{
+    return lhs == json_pointer<RefStringTypeLhs>(rhs);
+}
+
+template<typename RefStringTypeRhs,
+         typename StringType = typename json_pointer<RefStringTypeRhs>::string_t>
+JSON_HEDLEY_DEPRECATED_FOR(3.11.2, operator==(json_pointer, json_pointer))
+inline bool operator==(const StringType& lhs,
+                       const json_pointer<RefStringTypeRhs>& rhs)
+{
+    return json_pointer<RefStringTypeRhs>(lhs) == rhs;
+}
+
+template<typename RefStringTypeLhs, typename RefStringTypeRhs>
+inline bool operator!=(const json_pointer<RefStringTypeLhs>& lhs,
+                       const json_pointer<RefStringTypeRhs>& rhs) noexcept
+{
+    return !(lhs == rhs);
+}
+
+template<typename RefStringTypeLhs,
+         typename StringType = typename json_pointer<RefStringTypeLhs>::string_t>
+JSON_HEDLEY_DEPRECATED_FOR(3.11.2, operator!=(json_pointer, json_pointer))
+inline bool operator!=(const json_pointer<RefStringTypeLhs>& lhs,
+                       const StringType& rhs)
+{
+    return !(lhs == rhs);
+}
+
+template<typename RefStringTypeRhs,
+         typename StringType = typename json_pointer<RefStringTypeRhs>::string_t>
+JSON_HEDLEY_DEPRECATED_FOR(3.11.2, operator!=(json_pointer, json_pointer))
+inline bool operator!=(const StringType& lhs,
+                       const json_pointer<RefStringTypeRhs>& rhs)
+{
+    return !(lhs == rhs);
+}
+
+template<typename RefStringTypeLhs, typename RefStringTypeRhs>
+inline bool operator<(const json_pointer<RefStringTypeLhs>& lhs,
+                      const json_pointer<RefStringTypeRhs>& rhs) noexcept
+{
+    return lhs.reference_tokens < rhs.reference_tokens;
+}
+#endif
+
+NLOHMANN_JSON_NAMESPACE_END
+
+// #include <nlohmann/detail/json_ref.hpp>
+//     __ _____ _____ _____
+//  __|  |   __|     |   | |  JSON for Modern C++
+// |  |  |__   |  |  | | | |  version 3.11.2
+// |_____|_____|_____|_|___|  https://github.com/nlohmann/json
+//
+// SPDX-FileCopyrightText: 2013-2022 Niels Lohmann <https://nlohmann.me>
+// SPDX-License-Identifier: MIT
+
+
+
+#include <initializer_list>
+#include <utility>
+
+// #include <nlohmann/detail/abi_macros.hpp>
+
+// #include <nlohmann/detail/meta/type_traits.hpp>
+
+
+NLOHMANN_JSON_NAMESPACE_BEGIN
+namespace detail
+{
+
+template<typename BasicJsonType>
+class json_ref
+{
+  public:
+    using value_type = BasicJsonType;
+
+    json_ref(value_type&& value)
+        : owned_value(std::move(value))
+    {}
+
+    json_ref(const value_type& value)
+        : value_ref(&value)
+    {}
+
+    json_ref(std::initializer_list<json_ref> init)
+        : owned_value(init)
+    {}
+
+    template <
+        class... Args,
+        enable_if_t<std::is_constructible<value_type, Args...>::value, int> = 0 >
+    json_ref(Args && ... args)
+        : owned_value(std::forward<Args>(args)...)
+    {}
+
+    // class should be movable only
+    json_ref(json_ref&&) noexcept = default;
+    json_ref(const json_ref&) = delete;
+    json_ref& operator=(const json_ref&) = delete;
+    json_ref& operator=(json_ref&&) = delete;
+    ~json_ref() = default;
+
+    value_type moved_or_copied() const
+    {
+        if (value_ref == nullptr)
+        {
+            return std::move(owned_value);
+        }
+        return *value_ref;
+    }
+
+    value_type const& operator*() const
+    {
+        return value_ref ? *value_ref : owned_value;
+    }
+
+    value_type const* operator->() const
+    {
+        return &** this;
+    }
+
+  private:
+    mutable value_type owned_value = nullptr;
+    value_type const* value_ref = nullptr;
+};
+
+}  // namespace detail
+NLOHMANN_JSON_NAMESPACE_END
+
+// #include <nlohmann/detail/macro_scope.hpp>
+
+// #include <nlohmann/detail/string_concat.hpp>
+
+// #include <nlohmann/detail/string_escape.hpp>
+
+// #include <nlohmann/detail/meta/cpp_future.hpp>
+
+// #include <nlohmann/detail/meta/type_traits.hpp>
+
+// #include <nlohmann/detail/output/binary_writer.hpp>
+//     __ _____ _____ _____
+//  __|  |   __|     |   | |  JSON for Modern C++
+// |  |  |__   |  |  | | | |  version 3.11.2
+// |_____|_____|_____|_|___|  https://github.com/nlohmann/json
+//
+// SPDX-FileCopyrightText: 2013-2022 Niels Lohmann <https://nlohmann.me>
+// SPDX-License-Identifier: MIT
+
+
+
+#include <algorithm> // reverse
+#include <array> // array
+#include <map> // map
+#include <cmath> // isnan, isinf
+#include <cstdint> // uint8_t, uint16_t, uint32_t, uint64_t
+#include <cstring> // memcpy
+#include <limits> // numeric_limits
+#include <string> // string
+#include <utility> // move
+#include <vector> // vector
+
+// #include <nlohmann/detail/input/binary_reader.hpp>
+
+// #include <nlohmann/detail/macro_scope.hpp>
+
+// #include <nlohmann/detail/output/output_adapters.hpp>
+//     __ _____ _____ _____
+//  __|  |   __|     |   | |  JSON for Modern C++
+// |  |  |__   |  |  | | | |  version 3.11.2
+// |_____|_____|_____|_|___|  https://github.com/nlohmann/json
+//
+// SPDX-FileCopyrightText: 2013-2022 Niels Lohmann <https://nlohmann.me>
+// SPDX-License-Identifier: MIT
+
+
+
+#include <algorithm> // copy
+#include <cstddef> // size_t
+#include <iterator> // back_inserter
+#include <memory> // shared_ptr, make_shared
+#include <string> // basic_string
+#include <vector> // vector
+
+#ifndef JSON_NO_IO
+    #include <ios>      // streamsize
+    #include <ostream>  // basic_ostream
+#endif  // JSON_NO_IO
+
+// #include <nlohmann/detail/macro_scope.hpp>
+
+
+NLOHMANN_JSON_NAMESPACE_BEGIN
+namespace detail
+{
+
+/// abstract output adapter interface
+template<typename CharType> struct output_adapter_protocol
+{
+    virtual void write_character(CharType c) = 0;
+    virtual void write_characters(const CharType* s, std::size_t length) = 0;
+    virtual ~output_adapter_protocol() = default;
+
+    output_adapter_protocol() = default;
+    output_adapter_protocol(const output_adapter_protocol&) = default;
+    output_adapter_protocol(output_adapter_protocol&&) noexcept = default;
+    output_adapter_protocol& operator=(const output_adapter_protocol&) = default;
+    output_adapter_protocol& operator=(output_adapter_protocol&&) noexcept = default;
+};
+
+/// a type to simplify interfaces
+template<typename CharType>
+using output_adapter_t = std::shared_ptr<output_adapter_protocol<CharType>>;
+
+/// output adapter for byte vectors
+template<typename CharType, typename AllocatorType = std::allocator<CharType>>
+class output_vector_adapter : public output_adapter_protocol<CharType>
+{
+  public:
+    explicit output_vector_adapter(std::vector<CharType, AllocatorType>& vec) noexcept
+        : v(vec)
+    {}
+
+    void write_character(CharType c) override
+    {
+        v.push_back(c);
+    }
+
+    JSON_HEDLEY_NON_NULL(2)
+    void write_characters(const CharType* s, std::size_t length) override
+    {
+        v.insert(v.end(), s, s + length);
+    }
+
+  private:
+    std::vector<CharType, AllocatorType>& v;
+};
+
+#ifndef JSON_NO_IO
+/// output adapter for output streams
+template<typename CharType>
+class output_stream_adapter : public output_adapter_protocol<CharType>
+{
+  public:
+    explicit output_stream_adapter(std::basic_ostream<CharType>& s) noexcept
+        : stream(s)
+    {}
+
+    void write_character(CharType c) override
+    {
+        stream.put(c);
+    }
+
+    JSON_HEDLEY_NON_NULL(2)
+    void write_characters(const CharType* s, std::size_t length) override
+    {
+        stream.write(s, static_cast<std::streamsize>(length));
+    }
+
+  private:
+    std::basic_ostream<CharType>& stream;
+};
+#endif  // JSON_NO_IO
+
+/// output adapter for basic_string
+template<typename CharType, typename StringType = std::basic_string<CharType>>
+class output_string_adapter : public output_adapter_protocol<CharType>
+{
+  public:
+    explicit output_string_adapter(StringType& s) noexcept
+        : str(s)
+    {}
+
+    void write_character(CharType c) override
+    {
+        str.push_back(c);
+    }
+
+    JSON_HEDLEY_NON_NULL(2)
+    void write_characters(const CharType* s, std::size_t length) override
+    {
+        str.append(s, length);
+    }
+
+  private:
+    StringType& str;
+};
+
+template<typename CharType, typename StringType = std::basic_string<CharType>>
+class output_adapter
+{
+  public:
+    template<typename AllocatorType = std::allocator<CharType>>
+    output_adapter(std::vector<CharType, AllocatorType>& vec)
+        : oa(std::make_shared<output_vector_adapter<CharType, AllocatorType>>(vec)) {}
+
+#ifndef JSON_NO_IO
+    output_adapter(std::basic_ostream<CharType>& s)
+        : oa(std::make_shared<output_stream_adapter<CharType>>(s)) {}
+#endif  // JSON_NO_IO
+
+    output_adapter(StringType& s)
+        : oa(std::make_shared<output_string_adapter<CharType, StringType>>(s)) {}
+
+    operator output_adapter_t<CharType>()
+    {
+        return oa;
+    }
+
+  private:
+    output_adapter_t<CharType> oa = nullptr;
+};
+
+}  // namespace detail
+NLOHMANN_JSON_NAMESPACE_END
+
+// #include <nlohmann/detail/string_concat.hpp>
+
+
+NLOHMANN_JSON_NAMESPACE_BEGIN
+namespace detail
+{
+
+///////////////////
+// binary writer //
+///////////////////
+
+/*!
+@brief serialization to CBOR and MessagePack values
+*/
+template<typename BasicJsonType, typename CharType>
+class binary_writer
+{
+    using string_t = typename BasicJsonType::string_t;
+    using binary_t = typename BasicJsonType::binary_t;
+    using number_float_t = typename BasicJsonType::number_float_t;
+
+  public:
+    /*!
+    @brief create a binary writer
+
+    @param[in] adapter  output adapter to write to
+    */
+    explicit binary_writer(output_adapter_t<CharType> adapter) : oa(std::move(adapter))
+    {
+        JSON_ASSERT(oa);
+    }
+
+    /*!
+    @param[in] j  JSON value to serialize
+    @pre       j.type() == value_t::object
+    */
+    void write_bson(const BasicJsonType& j)
+    {
+        switch (j.type())
+        {
+            case value_t::object:
+            {
+                write_bson_object(*j.m_value.object);
+                break;
+            }
+
+            case value_t::null:
+            case value_t::array:
+            case value_t::string:
+            case value_t::boolean:
+            case value_t::number_integer:
+            case value_t::number_unsigned:
+            case value_t::number_float:
+            case value_t::binary:
+            case value_t::discarded:
+            default:
+            {
+                JSON_THROW(type_error::create(317, concat("to serialize to BSON, top-level type must be object, but is ", j.type_name()), &j));
+            }
+        }
+    }
+
+    /*!
+    @param[in] j  JSON value to serialize
+    */
+    void write_cbor(const BasicJsonType& j)
+    {
+        switch (j.type())
+        {
+            case value_t::null:
+            {
+                oa->write_character(to_char_type(0xF6));
+                break;
+            }
+
+            case value_t::boolean:
+            {
+                oa->write_character(j.m_value.boolean
+                                    ? to_char_type(0xF5)
+                                    : to_char_type(0xF4));
+                break;
+            }
+
+            case value_t::number_integer:
+            {
+                if (j.m_value.number_integer >= 0)
+                {
+                    // CBOR does not differentiate between positive signed
+                    // integers and unsigned integers. Therefore, we used the
+                    // code from the value_t::number_unsigned case here.
+                    if (j.m_value.number_integer <= 0x17)
+                    {
+                        write_number(static_cast<std::uint8_t>(j.m_value.number_integer));
+                    }
+                    else if (j.m_value.number_integer <= (std::numeric_limits<std::uint8_t>::max)())
+                    {
+                        oa->write_character(to_char_type(0x18));
+                        write_number(static_cast<std::uint8_t>(j.m_value.number_integer));
+                    }
+                    else if (j.m_value.number_integer <= (std::numeric_limits<std::uint16_t>::max)())
+                    {
+                        oa->write_character(to_char_type(0x19));
+                        write_number(static_cast<std::uint16_t>(j.m_value.number_integer));
+                    }
+                    else if (j.m_value.number_integer <= (std::numeric_limits<std::uint32_t>::max)())
+                    {
+                        oa->write_character(to_char_type(0x1A));
+                        write_number(static_cast<std::uint32_t>(j.m_value.number_integer));
+                    }
+                    else
+                    {
+                        oa->write_character(to_char_type(0x1B));
+                        write_number(static_cast<std::uint64_t>(j.m_value.number_integer));
+                    }
+                }
+                else
+                {
+                    // The conversions below encode the sign in the first
+                    // byte, and the value is converted to a positive number.
+                    const auto positive_number = -1 - j.m_value.number_integer;
+                    if (j.m_value.number_integer >= -24)
+                    {
+                        write_number(static_cast<std::uint8_t>(0x20 + positive_number));
+                    }
+                    else if (positive_number <= (std::numeric_limits<std::uint8_t>::max)())
+                    {
+                        oa->write_character(to_char_type(0x38));
+                        write_number(static_cast<std::uint8_t>(positive_number));
+                    }
+                    else if (positive_number <= (std::numeric_limits<std::uint16_t>::max)())
+                    {
+                        oa->write_character(to_char_type(0x39));
+                        write_number(static_cast<std::uint16_t>(positive_number));
+                    }
+                    else if (positive_number <= (std::numeric_limits<std::uint32_t>::max)())
+                    {
+                        oa->write_character(to_char_type(0x3A));
+                        write_number(static_cast<std::uint32_t>(positive_number));
+                    }
+                    else
+                    {
+                        oa->write_character(to_char_type(0x3B));
+                        write_number(static_cast<std::uint64_t>(positive_number));
+                    }
+                }
+                break;
+            }
+
+            case value_t::number_unsigned:
+            {
+                if (j.m_value.number_unsigned <= 0x17)
+                {
+                    write_number(static_cast<std::uint8_t>(j.m_value.number_unsigned));
+                }
+                else if (j.m_value.number_unsigned <= (std::numeric_limits<std::uint8_t>::max)())
+                {
+                    oa->write_character(to_char_type(0x18));
+                    write_number(static_cast<std::uint8_t>(j.m_value.number_unsigned));
+                }
+                else if (j.m_value.number_unsigned <= (std::numeric_limits<std::uint16_t>::max)())
+                {
+                    oa->write_character(to_char_type(0x19));
+                    write_number(static_cast<std::uint16_t>(j.m_value.number_unsigned));
+                }
+                else if (j.m_value.number_unsigned <= (std::numeric_limits<std::uint32_t>::max)())
+                {
+                    oa->write_character(to_char_type(0x1A));
+                    write_number(static_cast<std::uint32_t>(j.m_value.number_unsigned));
+                }
+                else
+                {
+                    oa->write_character(to_char_type(0x1B));
+                    write_number(static_cast<std::uint64_t>(j.m_value.number_unsigned));
+                }
+                break;
+            }
+
+            case value_t::number_float:
+            {
+                if (std::isnan(j.m_value.number_float))
+                {
+                    // NaN is 0xf97e00 in CBOR
+                    oa->write_character(to_char_type(0xF9));
+                    oa->write_character(to_char_type(0x7E));
+                    oa->write_character(to_char_type(0x00));
+                }
+                else if (std::isinf(j.m_value.number_float))
+                {
+                    // Infinity is 0xf97c00, -Infinity is 0xf9fc00
+                    oa->write_character(to_char_type(0xf9));
+                    oa->write_character(j.m_value.number_float > 0 ? to_char_type(0x7C) : to_char_type(0xFC));
+                    oa->write_character(to_char_type(0x00));
+                }
+                else
+                {
+                    write_compact_float(j.m_value.number_float, detail::input_format_t::cbor);
+                }
+                break;
+            }
+
+            case value_t::string:
+            {
+                // step 1: write control byte and the string length
+                const auto N = j.m_value.string->size();
+                if (N <= 0x17)
+                {
+                    write_number(static_cast<std::uint8_t>(0x60 + N));
+                }
+                else if (N <= (std::numeric_limits<std::uint8_t>::max)())
+                {
+                    oa->write_character(to_char_type(0x78));
+                    write_number(static_cast<std::uint8_t>(N));
+                }
+                else if (N <= (std::numeric_limits<std::uint16_t>::max)())
+                {
+                    oa->write_character(to_char_type(0x79));
+                    write_number(static_cast<std::uint16_t>(N));
+                }
+                else if (N <= (std::numeric_limits<std::uint32_t>::max)())
+                {
+                    oa->write_character(to_char_type(0x7A));
+                    write_number(static_cast<std::uint32_t>(N));
+                }
+                // LCOV_EXCL_START
+                else if (N <= (std::numeric_limits<std::uint64_t>::max)())
+                {
+                    oa->write_character(to_char_type(0x7B));
+                    write_number(static_cast<std::uint64_t>(N));
+                }
+                // LCOV_EXCL_STOP
+
+                // step 2: write the string
+                oa->write_characters(
+                    reinterpret_cast<const CharType*>(j.m_value.string->c_str()),
+                    j.m_value.string->size());
+                break;
+            }
+
+            case value_t::array:
+            {
+                // step 1: write control byte and the array size
+                const auto N = j.m_value.array->size();
+                if (N <= 0x17)
+                {
+                    write_number(static_cast<std::uint8_t>(0x80 + N));
+                }
+                else if (N <= (std::numeric_limits<std::uint8_t>::max)())
+                {
+                    oa->write_character(to_char_type(0x98));
+                    write_number(static_cast<std::uint8_t>(N));
+                }
+                else if (N <= (std::numeric_limits<std::uint16_t>::max)())
+                {
+                    oa->write_character(to_char_type(0x99));
+                    write_number(static_cast<std::uint16_t>(N));
+                }
+                else if (N <= (std::numeric_limits<std::uint32_t>::max)())
+                {
+                    oa->write_character(to_char_type(0x9A));
+                    write_number(static_cast<std::uint32_t>(N));
+                }
+                // LCOV_EXCL_START
+                else if (N <= (std::numeric_limits<std::uint64_t>::max)())
+                {
+                    oa->write_character(to_char_type(0x9B));
+                    write_number(static_cast<std::uint64_t>(N));
+                }
+                // LCOV_EXCL_STOP
+
+                // step 2: write each element
+                for (const auto& el : *j.m_value.array)
+                {
+                    write_cbor(el);
+                }
+                break;
+            }
+
+            case value_t::binary:
+            {
+                if (j.m_value.binary->has_subtype())
+                {
+                    if (j.m_value.binary->subtype() <= (std::numeric_limits<std::uint8_t>::max)())
+                    {
+                        write_number(static_cast<std::uint8_t>(0xd8));
+                        write_number(static_cast<std::uint8_t>(j.m_value.binary->subtype()));
+                    }
+                    else if (j.m_value.binary->subtype() <= (std::numeric_limits<std::uint16_t>::max)())
+                    {
+                        write_number(static_cast<std::uint8_t>(0xd9));
+                        write_number(static_cast<std::uint16_t>(j.m_value.binary->subtype()));
+                    }
+                    else if (j.m_value.binary->subtype() <= (std::numeric_limits<std::uint32_t>::max)())
+                    {
+                        write_number(static_cast<std::uint8_t>(0xda));
+                        write_number(static_cast<std::uint32_t>(j.m_value.binary->subtype()));
+                    }
+                    else if (j.m_value.binary->subtype() <= (std::numeric_limits<std::uint64_t>::max)())
+                    {
+                        write_number(static_cast<std::uint8_t>(0xdb));
+                        write_number(static_cast<std::uint64_t>(j.m_value.binary->subtype()));
+                    }
+                }
+
+                // step 1: write control byte and the binary array size
+                const auto N = j.m_value.binary->size();
+                if (N <= 0x17)
+                {
+                    write_number(static_cast<std::uint8_t>(0x40 + N));
+                }
+                else if (N <= (std::numeric_limits<std::uint8_t>::max)())
+                {
+                    oa->write_character(to_char_type(0x58));
+                    write_number(static_cast<std::uint8_t>(N));
+                }
+                else if (N <= (std::numeric_limits<std::uint16_t>::max)())
+                {
+                    oa->write_character(to_char_type(0x59));
+                    write_number(static_cast<std::uint16_t>(N));
+                }
+                else if (N <= (std::numeric_limits<std::uint32_t>::max)())
+                {
+                    oa->write_character(to_char_type(0x5A));
+                    write_number(static_cast<std::uint32_t>(N));
+                }
+                // LCOV_EXCL_START
+                else if (N <= (std::numeric_limits<std::uint64_t>::max)())
+                {
+                    oa->write_character(to_char_type(0x5B));
+                    write_number(static_cast<std::uint64_t>(N));
+                }
+                // LCOV_EXCL_STOP
+
+                // step 2: write each element
+                oa->write_characters(
+                    reinterpret_cast<const CharType*>(j.m_value.binary->data()),
+                    N);
+
+                break;
+            }
+
+            case value_t::object:
+            {
+                // step 1: write control byte and the object size
+                const auto N = j.m_value.object->size();
+                if (N <= 0x17)
+                {
+                    write_number(static_cast<std::uint8_t>(0xA0 + N));
+                }
+                else if (N <= (std::numeric_limits<std::uint8_t>::max)())
+                {
+                    oa->write_character(to_char_type(0xB8));
+                    write_number(static_cast<std::uint8_t>(N));
+                }
+                else if (N <= (std::numeric_limits<std::uint16_t>::max)())
+                {
+                    oa->write_character(to_char_type(0xB9));
+                    write_number(static_cast<std::uint16_t>(N));
+                }
+                else if (N <= (std::numeric_limits<std::uint32_t>::max)())
+                {
+                    oa->write_character(to_char_type(0xBA));
+                    write_number(static_cast<std::uint32_t>(N));
+                }
+                // LCOV_EXCL_START
+                else if (N <= (std::numeric_limits<std::uint64_t>::max)())
+                {
+                    oa->write_character(to_char_type(0xBB));
+                    write_number(static_cast<std::uint64_t>(N));
+                }
+                // LCOV_EXCL_STOP
+
+                // step 2: write each element
+                for (const auto& el : *j.m_value.object)
+                {
+                    write_cbor(el.first);
+                    write_cbor(el.second);
+                }
+                break;
+            }
+
+            case value_t::discarded:
+            default:
+                break;
+        }
+    }
+
+    /*!
+    @param[in] j  JSON value to serialize
+    */
+    void write_msgpack(const BasicJsonType& j)
+    {
+        switch (j.type())
+        {
+            case value_t::null: // nil
+            {
+                oa->write_character(to_char_type(0xC0));
+                break;
+            }
+
+            case value_t::boolean: // true and false
+            {
+                oa->write_character(j.m_value.boolean
+                                    ? to_char_type(0xC3)
+                                    : to_char_type(0xC2));
+                break;
+            }
+
+            case value_t::number_integer:
+            {
+                if (j.m_value.number_integer >= 0)
+                {
+                    // MessagePack does not differentiate between positive
+                    // signed integers and unsigned integers. Therefore, we used
+                    // the code from the value_t::number_unsigned case here.
+                    if (j.m_value.number_unsigned < 128)
+                    {
+                        // positive fixnum
+                        write_number(static_cast<std::uint8_t>(j.m_value.number_integer));
+                    }
+                    else if (j.m_value.number_unsigned <= (std::numeric_limits<std::uint8_t>::max)())
+                    {
+                        // uint 8
+                        oa->write_character(to_char_type(0xCC));
+                        write_number(static_cast<std::uint8_t>(j.m_value.number_integer));
+                    }
+                    else if (j.m_value.number_unsigned <= (std::numeric_limits<std::uint16_t>::max)())
+                    {
+                        // uint 16
+                        oa->write_character(to_char_type(0xCD));
+                        write_number(static_cast<std::uint16_t>(j.m_value.number_integer));
+                    }
+                    else if (j.m_value.number_unsigned <= (std::numeric_limits<std::uint32_t>::max)())
+                    {
+                        // uint 32
+                        oa->write_character(to_char_type(0xCE));
+                        write_number(static_cast<std::uint32_t>(j.m_value.number_integer));
+                    }
+                    else if (j.m_value.number_unsigned <= (std::numeric_limits<std::uint64_t>::max)())
+                    {
+                        // uint 64
+                        oa->write_character(to_char_type(0xCF));
+                        write_number(static_cast<std::uint64_t>(j.m_value.number_integer));
+                    }
+                }
+                else
+                {
+                    if (j.m_value.number_integer >= -32)
+                    {
+                        // negative fixnum
+                        write_number(static_cast<std::int8_t>(j.m_value.number_integer));
+                    }
+                    else if (j.m_value.number_integer >= (std::numeric_limits<std::int8_t>::min)() &&
+                             j.m_value.number_integer <= (std::numeric_limits<std::int8_t>::max)())
+                    {
+                        // int 8
+                        oa->write_character(to_char_type(0xD0));
+                        write_number(static_cast<std::int8_t>(j.m_value.number_integer));
+                    }
+                    else if (j.m_value.number_integer >= (std::numeric_limits<std::int16_t>::min)() &&
+                             j.m_value.number_integer <= (std::numeric_limits<std::int16_t>::max)())
+                    {
+                        // int 16
+                        oa->write_character(to_char_type(0xD1));
+                        write_number(static_cast<std::int16_t>(j.m_value.number_integer));
+                    }
+                    else if (j.m_value.number_integer >= (std::numeric_limits<std::int32_t>::min)() &&
+                             j.m_value.number_integer <= (std::numeric_limits<std::int32_t>::max)())
+                    {
+                        // int 32
+                        oa->write_character(to_char_type(0xD2));
+                        write_number(static_cast<std::int32_t>(j.m_value.number_integer));
+                    }
+                    else if (j.m_value.number_integer >= (std::numeric_limits<std::int64_t>::min)() &&
+                             j.m_value.number_integer <= (std::numeric_limits<std::int64_t>::max)())
+                    {
+                        // int 64
+                        oa->write_character(to_char_type(0xD3));
+                        write_number(static_cast<std::int64_t>(j.m_value.number_integer));
+                    }
+                }
+                break;
+            }
+
+            case value_t::number_unsigned:
+            {
+                if (j.m_value.number_unsigned < 128)
+                {
+                    // positive fixnum
+                    write_number(static_cast<std::uint8_t>(j.m_value.number_integer));
+                }
+                else if (j.m_value.number_unsigned <= (std::numeric_limits<std::uint8_t>::max)())
+                {
+                    // uint 8
+                    oa->write_character(to_char_type(0xCC));
+                    write_number(static_cast<std::uint8_t>(j.m_value.number_integer));
+                }
+                else if (j.m_value.number_unsigned <= (std::numeric_limits<std::uint16_t>::max)())
+                {
+                    // uint 16
+                    oa->write_character(to_char_type(0xCD));
+                    write_number(static_cast<std::uint16_t>(j.m_value.number_integer));
+                }
+                else if (j.m_value.number_unsigned <= (std::numeric_limits<std::uint32_t>::max)())
+                {
+                    // uint 32
+                    oa->write_character(to_char_type(0xCE));
+                    write_number(static_cast<std::uint32_t>(j.m_value.number_integer));
+                }
+                else if (j.m_value.number_unsigned <= (std::numeric_limits<std::uint64_t>::max)())
+                {
+                    // uint 64
+                    oa->write_character(to_char_type(0xCF));
+                    write_number(static_cast<std::uint64_t>(j.m_value.number_integer));
+                }
+                break;
+            }
+
+            case value_t::number_float:
+            {
+                write_compact_float(j.m_value.number_float, detail::input_format_t::msgpack);
+                break;
+            }
+
+            case value_t::string:
+            {
+                // step 1: write control byte and the string length
+                const auto N = j.m_value.string->size();
+                if (N <= 31)
+                {
+                    // fixstr
+                    write_number(static_cast<std::uint8_t>(0xA0 | N));
+                }
+                else if (N <= (std::numeric_limits<std::uint8_t>::max)())
+                {
+                    // str 8
+                    oa->write_character(to_char_type(0xD9));
+                    write_number(static_cast<std::uint8_t>(N));
+                }
+                else if (N <= (std::numeric_limits<std::uint16_t>::max)())
+                {
+                    // str 16
+                    oa->write_character(to_char_type(0xDA));
+                    write_number(static_cast<std::uint16_t>(N));
+                }
+                else if (N <= (std::numeric_limits<std::uint32_t>::max)())
+                {
+                    // str 32
+                    oa->write_character(to_char_type(0xDB));
+                    write_number(static_cast<std::uint32_t>(N));
+                }
+
+                // step 2: write the string
+                oa->write_characters(
+                    reinterpret_cast<const CharType*>(j.m_value.string->c_str()),
+                    j.m_value.string->size());
+                break;
+            }
+
+            case value_t::array:
+            {
+                // step 1: write control byte and the array size
+                const auto N = j.m_value.array->size();
+                if (N <= 15)
+                {
+                    // fixarray
+                    write_number(static_cast<std::uint8_t>(0x90 | N));
+                }
+                else if (N <= (std::numeric_limits<std::uint16_t>::max)())
+                {
+                    // array 16
+                    oa->write_character(to_char_type(0xDC));
+                    write_number(static_cast<std::uint16_t>(N));
+                }
+                else if (N <= (std::numeric_limits<std::uint32_t>::max)())
+                {
+                    // array 32
+                    oa->write_character(to_char_type(0xDD));
+                    write_number(static_cast<std::uint32_t>(N));
+                }
+
+                // step 2: write each element
+                for (const auto& el : *j.m_value.array)
+                {
+                    write_msgpack(el);
+                }
+                break;
+            }
+
+            case value_t::binary:
+            {
+                // step 0: determine if the binary type has a set subtype to
+                // determine whether or not to use the ext or fixext types
+                const bool use_ext = j.m_value.binary->has_subtype();
+
+                // step 1: write control byte and the byte string length
+                const auto N = j.m_value.binary->size();
+                if (N <= (std::numeric_limits<std::uint8_t>::max)())
+                {
+                    std::uint8_t output_type{};
+                    bool fixed = true;
+                    if (use_ext)
+                    {
+                        switch (N)
+                        {
+                            case 1:
+                                output_type = 0xD4; // fixext 1
+                                break;
+                            case 2:
+                                output_type = 0xD5; // fixext 2
+                                break;
+                            case 4:
+                                output_type = 0xD6; // fixext 4
+                                break;
+                            case 8:
+                                output_type = 0xD7; // fixext 8
+                                break;
+                            case 16:
+                                output_type = 0xD8; // fixext 16
+                                break;
+                            default:
+                                output_type = 0xC7; // ext 8
+                                fixed = false;
+                                break;
+                        }
+
+                    }
+                    else
+                    {
+                        output_type = 0xC4; // bin 8
+                        fixed = false;
+                    }
+
+                    oa->write_character(to_char_type(output_type));
+                    if (!fixed)
+                    {
+                        write_number(static_cast<std::uint8_t>(N));
+                    }
+                }
+                else if (N <= (std::numeric_limits<std::uint16_t>::max)())
+                {
+                    std::uint8_t output_type = use_ext
+                                               ? 0xC8 // ext 16
+                                               : 0xC5; // bin 16
+
+                    oa->write_character(to_char_type(output_type));
+                    write_number(static_cast<std::uint16_t>(N));
+                }
+                else if (N <= (std::numeric_limits<std::uint32_t>::max)())
+                {
+                    std::uint8_t output_type = use_ext
+                                               ? 0xC9 // ext 32
+                                               : 0xC6; // bin 32
+
+                    oa->write_character(to_char_type(output_type));
+                    write_number(static_cast<std::uint32_t>(N));
+                }
+
+                // step 1.5: if this is an ext type, write the subtype
+                if (use_ext)
+                {
+                    write_number(static_cast<std::int8_t>(j.m_value.binary->subtype()));
+                }
+
+                // step 2: write the byte string
+                oa->write_characters(
+                    reinterpret_cast<const CharType*>(j.m_value.binary->data()),
+                    N);
+
+                break;
+            }
+
+            case value_t::object:
+            {
+                // step 1: write control byte and the object size
+                const auto N = j.m_value.object->size();
+                if (N <= 15)
+                {
+                    // fixmap
+                    write_number(static_cast<std::uint8_t>(0x80 | (N & 0xF)));
+                }
+                else if (N <= (std::numeric_limits<std::uint16_t>::max)())
+                {
+                    // map 16
+                    oa->write_character(to_char_type(0xDE));
+                    write_number(static_cast<std::uint16_t>(N));
+                }
+                else if (N <= (std::numeric_limits<std::uint32_t>::max)())
+                {
+                    // map 32
+                    oa->write_character(to_char_type(0xDF));
+                    write_number(static_cast<std::uint32_t>(N));
+                }
+
+                // step 2: write each element
+                for (const auto& el : *j.m_value.object)
+                {
+                    write_msgpack(el.first);
+                    write_msgpack(el.second);
+                }
+                break;
+            }
+
+            case value_t::discarded:
+            default:
+                break;
+        }
+    }
+
+    /*!
+    @param[in] j  JSON value to serialize
+    @param[in] use_count   whether to use '#' prefixes (optimized format)
+    @param[in] use_type    whether to use '$' prefixes (optimized format)
+    @param[in] add_prefix  whether prefixes need to be used for this value
+    @param[in] use_bjdata  whether write in BJData format, default is false
+    */
+    void write_ubjson(const BasicJsonType& j, const bool use_count,
+                      const bool use_type, const bool add_prefix = true,
+                      const bool use_bjdata = false)
+    {
+        switch (j.type())
+        {
+            case value_t::null:
+            {
+                if (add_prefix)
+                {
+                    oa->write_character(to_char_type('Z'));
+                }
+                break;
+            }
+
+            case value_t::boolean:
+            {
+                if (add_prefix)
+                {
+                    oa->write_character(j.m_value.boolean
+                                        ? to_char_type('T')
+                                        : to_char_type('F'));
+                }
+                break;
+            }
+
+            case value_t::number_integer:
+            {
+                write_number_with_ubjson_prefix(j.m_value.number_integer, add_prefix, use_bjdata);
+                break;
+            }
+
+            case value_t::number_unsigned:
+            {
+                write_number_with_ubjson_prefix(j.m_value.number_unsigned, add_prefix, use_bjdata);
+                break;
+            }
+
+            case value_t::number_float:
+            {
+                write_number_with_ubjson_prefix(j.m_value.number_float, add_prefix, use_bjdata);
+                break;
+            }
+
+            case value_t::string:
+            {
+                if (add_prefix)
+                {
+                    oa->write_character(to_char_type('S'));
+                }
+                write_number_with_ubjson_prefix(j.m_value.string->size(), true, use_bjdata);
+                oa->write_characters(
+                    reinterpret_cast<const CharType*>(j.m_value.string->c_str()),
+                    j.m_value.string->size());
+                break;
+            }
+
+            case value_t::array:
+            {
+                if (add_prefix)
+                {
+                    oa->write_character(to_char_type('['));
+                }
+
+                bool prefix_required = true;
+                if (use_type && !j.m_value.array->empty())
+                {
+                    JSON_ASSERT(use_count);
+                    const CharType first_prefix = ubjson_prefix(j.front(), use_bjdata);
+                    const bool same_prefix = std::all_of(j.begin() + 1, j.end(),
+                                                         [this, first_prefix, use_bjdata](const BasicJsonType & v)
+                    {
+                        return ubjson_prefix(v, use_bjdata) == first_prefix;
+                    });
+
+                    std::vector<CharType> bjdx = {'[', '{', 'S', 'H', 'T', 'F', 'N', 'Z'}; // excluded markers in bjdata optimized type
+
+                    if (same_prefix && !(use_bjdata && std::find(bjdx.begin(), bjdx.end(), first_prefix) != bjdx.end()))
+                    {
+                        prefix_required = false;
+                        oa->write_character(to_char_type('$'));
+                        oa->write_character(first_prefix);
+                    }
+                }
+
+                if (use_count)
+                {
+                    oa->write_character(to_char_type('#'));
+                    write_number_with_ubjson_prefix(j.m_value.array->size(), true, use_bjdata);
+                }
+
+                for (const auto& el : *j.m_value.array)
+                {
+                    write_ubjson(el, use_count, use_type, prefix_required, use_bjdata);
+                }
+
+                if (!use_count)
+                {
+                    oa->write_character(to_char_type(']'));
+                }
+
+                break;
+            }
+
+            case value_t::binary:
+            {
+                if (add_prefix)
+                {
+                    oa->write_character(to_char_type('['));
+                }
+
+                if (use_type && !j.m_value.binary->empty())
+                {
+                    JSON_ASSERT(use_count);
+                    oa->write_character(to_char_type('$'));
+                    oa->write_character('U');
+                }
+
+                if (use_count)
+                {
+                    oa->write_character(to_char_type('#'));
+                    write_number_with_ubjson_prefix(j.m_value.binary->size(), true, use_bjdata);
+                }
+
+                if (use_type)
+                {
+                    oa->write_characters(
+                        reinterpret_cast<const CharType*>(j.m_value.binary->data()),
+                        j.m_value.binary->size());
+                }
+                else
+                {
+                    for (size_t i = 0; i < j.m_value.binary->size(); ++i)
+                    {
+                        oa->write_character(to_char_type('U'));
+                        oa->write_character(j.m_value.binary->data()[i]);
+                    }
+                }
+
+                if (!use_count)
+                {
+                    oa->write_character(to_char_type(']'));
+                }
+
+                break;
+            }
+
+            case value_t::object:
+            {
+                if (use_bjdata && j.m_value.object->size() == 3 && j.m_value.object->find("_ArrayType_") != j.m_value.object->end() && j.m_value.object->find("_ArraySize_") != j.m_value.object->end() && j.m_value.object->find("_ArrayData_") != j.m_value.object->end())
+                {
+                    if (!write_bjdata_ndarray(*j.m_value.object, use_count, use_type))  // decode bjdata ndarray in the JData format (https://github.com/NeuroJSON/jdata)
+                    {
+                        break;
+                    }
+                }
+
+                if (add_prefix)
+                {
+                    oa->write_character(to_char_type('{'));
+                }
+
+                bool prefix_required = true;
+                if (use_type && !j.m_value.object->empty())
+                {
+                    JSON_ASSERT(use_count);
+                    const CharType first_prefix = ubjson_prefix(j.front(), use_bjdata);
+                    const bool same_prefix = std::all_of(j.begin(), j.end(),
+                                                         [this, first_prefix, use_bjdata](const BasicJsonType & v)
+                    {
+                        return ubjson_prefix(v, use_bjdata) == first_prefix;
+                    });
+
+                    std::vector<CharType> bjdx = {'[', '{', 'S', 'H', 'T', 'F', 'N', 'Z'}; // excluded markers in bjdata optimized type
+
+                    if (same_prefix && !(use_bjdata && std::find(bjdx.begin(), bjdx.end(), first_prefix) != bjdx.end()))
+                    {
+                        prefix_required = false;
+                        oa->write_character(to_char_type('$'));
+                        oa->write_character(first_prefix);
+                    }
+                }
+
+                if (use_count)
+                {
+                    oa->write_character(to_char_type('#'));
+                    write_number_with_ubjson_prefix(j.m_value.object->size(), true, use_bjdata);
+                }
+
+                for (const auto& el : *j.m_value.object)
+                {
+                    write_number_with_ubjson_prefix(el.first.size(), true, use_bjdata);
+                    oa->write_characters(
+                        reinterpret_cast<const CharType*>(el.first.c_str()),
+                        el.first.size());
+                    write_ubjson(el.second, use_count, use_type, prefix_required, use_bjdata);
+                }
+
+                if (!use_count)
+                {
+                    oa->write_character(to_char_type('}'));
+                }
+
+                break;
+            }
+
+            case value_t::discarded:
+            default:
+                break;
+        }
+    }
+
+  private:
+    //////////
+    // BSON //
+    //////////
+
+    /*!
+    @return The size of a BSON document entry header, including the id marker
+            and the entry name size (and its null-terminator).
+    */
+    static std::size_t calc_bson_entry_header_size(const string_t& name, const BasicJsonType& j)
+    {
+        const auto it = name.find(static_cast<typename string_t::value_type>(0));
+        if (JSON_HEDLEY_UNLIKELY(it != BasicJsonType::string_t::npos))
+        {
+            JSON_THROW(out_of_range::create(409, concat("BSON key cannot contain code point U+0000 (at byte ", std::to_string(it), ")"), &j));
+            static_cast<void>(j);
+        }
+
+        return /*id*/ 1ul + name.size() + /*zero-terminator*/1u;
+    }
+
+    /*!
+    @brief Writes the given @a element_type and @a name to the output adapter
+    */
+    void write_bson_entry_header(const string_t& name,
+                                 const std::uint8_t element_type)
+    {
+        oa->write_character(to_char_type(element_type)); // boolean
+        oa->write_characters(
+            reinterpret_cast<const CharType*>(name.c_str()),
+            name.size() + 1u);
+    }
+
+    /*!
+    @brief Writes a BSON element with key @a name and boolean value @a value
+    */
+    void write_bson_boolean(const string_t& name,
+                            const bool value)
+    {
+        write_bson_entry_header(name, 0x08);
+        oa->write_character(value ? to_char_type(0x01) : to_char_type(0x00));
+    }
+
+    /*!
+    @brief Writes a BSON element with key @a name and double value @a value
+    */
+    void write_bson_double(const string_t& name,
+                           const double value)
+    {
+        write_bson_entry_header(name, 0x01);
+        write_number<double>(value, true);
+    }
+
+    /*!
+    @return The size of the BSON-encoded string in @a value
+    */
+    static std::size_t calc_bson_string_size(const string_t& value)
+    {
+        return sizeof(std::int32_t) + value.size() + 1ul;
+    }
+
+    /*!
+    @brief Writes a BSON element with key @a name and string value @a value
+    */
+    void write_bson_string(const string_t& name,
+                           const string_t& value)
+    {
+        write_bson_entry_header(name, 0x02);
+
+        write_number<std::int32_t>(static_cast<std::int32_t>(value.size() + 1ul), true);
+        oa->write_characters(
+            reinterpret_cast<const CharType*>(value.c_str()),
+            value.size() + 1);
+    }
+
+    /*!
+    @brief Writes a BSON element with key @a name and null value
+    */
+    void write_bson_null(const string_t& name)
+    {
+        write_bson_entry_header(name, 0x0A);
+    }
+
+    /*!
+    @return The size of the BSON-encoded integer @a value
+    */
+    static std::size_t calc_bson_integer_size(const std::int64_t value)
+    {
+        return (std::numeric_limits<std::int32_t>::min)() <= value && value <= (std::numeric_limits<std::int32_t>::max)()
+               ? sizeof(std::int32_t)
+               : sizeof(std::int64_t);
+    }
+
+    /*!
+    @brief Writes a BSON element with key @a name and integer @a value
+    */
+    void write_bson_integer(const string_t& name,
+                            const std::int64_t value)
+    {
+        if ((std::numeric_limits<std::int32_t>::min)() <= value && value <= (std::numeric_limits<std::int32_t>::max)())
+        {
+            write_bson_entry_header(name, 0x10); // int32
+            write_number<std::int32_t>(static_cast<std::int32_t>(value), true);
+        }
+        else
+        {
+            write_bson_entry_header(name, 0x12); // int64
+            write_number<std::int64_t>(static_cast<std::int64_t>(value), true);
+        }
+    }
+
+    /*!
+    @return The size of the BSON-encoded unsigned integer in @a j
+    */
+    static constexpr std::size_t calc_bson_unsigned_size(const std::uint64_t value) noexcept
+    {
+        return (value <= static_cast<std::uint64_t>((std::numeric_limits<std::int32_t>::max)()))
+               ? sizeof(std::int32_t)
+               : sizeof(std::int64_t);
+    }
+
+    /*!
+    @brief Writes a BSON element with key @a name and unsigned @a value
+    */
+    void write_bson_unsigned(const string_t& name,
+                             const BasicJsonType& j)
+    {
+        if (j.m_value.number_unsigned <= static_cast<std::uint64_t>((std::numeric_limits<std::int32_t>::max)()))
+        {
+            write_bson_entry_header(name, 0x10 /* int32 */);
+            write_number<std::int32_t>(static_cast<std::int32_t>(j.m_value.number_unsigned), true);
+        }
+        else if (j.m_value.number_unsigned <= static_cast<std::uint64_t>((std::numeric_limits<std::int64_t>::max)()))
+        {
+            write_bson_entry_header(name, 0x12 /* int64 */);
+            write_number<std::int64_t>(static_cast<std::int64_t>(j.m_value.number_unsigned), true);
+        }
+        else
+        {
+            JSON_THROW(out_of_range::create(407, concat("integer number ", std::to_string(j.m_value.number_unsigned), " cannot be represented by BSON as it does not fit int64"), &j));
+        }
+    }
+
+    /*!
+    @brief Writes a BSON element with key @a name and object @a value
+    */
+    void write_bson_object_entry(const string_t& name,
+                                 const typename BasicJsonType::object_t& value)
+    {
+        write_bson_entry_header(name, 0x03); // object
+        write_bson_object(value);
+    }
+
+    /*!
+    @return The size of the BSON-encoded array @a value
+    */
+    static std::size_t calc_bson_array_size(const typename BasicJsonType::array_t& value)
+    {
+        std::size_t array_index = 0ul;
+
+        const std::size_t embedded_document_size = std::accumulate(std::begin(value), std::end(value), static_cast<std::size_t>(0), [&array_index](std::size_t result, const typename BasicJsonType::array_t::value_type & el)
+        {
+            return result + calc_bson_element_size(std::to_string(array_index++), el);
+        });
+
+        return sizeof(std::int32_t) + embedded_document_size + 1ul;
+    }
+
+    /*!
+    @return The size of the BSON-encoded binary array @a value
+    */
+    static std::size_t calc_bson_binary_size(const typename BasicJsonType::binary_t& value)
+    {
+        return sizeof(std::int32_t) + value.size() + 1ul;
+    }
+
+    /*!
+    @brief Writes a BSON element with key @a name and array @a value
+    */
+    void write_bson_array(const string_t& name,
+                          const typename BasicJsonType::array_t& value)
+    {
+        write_bson_entry_header(name, 0x04); // array
+        write_number<std::int32_t>(static_cast<std::int32_t>(calc_bson_array_size(value)), true);
+
+        std::size_t array_index = 0ul;
+
+        for (const auto& el : value)
+        {
+            write_bson_element(std::to_string(array_index++), el);
+        }
+
+        oa->write_character(to_char_type(0x00));
+    }
+
+    /*!
+    @brief Writes a BSON element with key @a name and binary value @a value
+    */
+    void write_bson_binary(const string_t& name,
+                           const binary_t& value)
+    {
+        write_bson_entry_header(name, 0x05);
+
+        write_number<std::int32_t>(static_cast<std::int32_t>(value.size()), true);
+        write_number(value.has_subtype() ? static_cast<std::uint8_t>(value.subtype()) : static_cast<std::uint8_t>(0x00));
+
+        oa->write_characters(reinterpret_cast<const CharType*>(value.data()), value.size());
+    }
+
+    /*!
+    @brief Calculates the size necessary to serialize the JSON value @a j with its @a name
+    @return The calculated size for the BSON document entry for @a j with the given @a name.
+    */
+    static std::size_t calc_bson_element_size(const string_t& name,
+            const BasicJsonType& j)
+    {
+        const auto header_size = calc_bson_entry_header_size(name, j);
+        switch (j.type())
+        {
+            case value_t::object:
+                return header_size + calc_bson_object_size(*j.m_value.object);
+
+            case value_t::array:
+                return header_size + calc_bson_array_size(*j.m_value.array);
+
+            case value_t::binary:
+                return header_size + calc_bson_binary_size(*j.m_value.binary);
+
+            case value_t::boolean:
+                return header_size + 1ul;
+
+            case value_t::number_float:
+                return header_size + 8ul;
+
+            case value_t::number_integer:
+                return header_size + calc_bson_integer_size(j.m_value.number_integer);
+
+            case value_t::number_unsigned:
+                return header_size + calc_bson_unsigned_size(j.m_value.number_unsigned);
+
+            case value_t::string:
+                return header_size + calc_bson_string_size(*j.m_value.string);
+
+            case value_t::null:
+                return header_size + 0ul;
+
+            // LCOV_EXCL_START
+            case value_t::discarded:
+            default:
+                JSON_ASSERT(false); // NOLINT(cert-dcl03-c,hicpp-static-assert,misc-static-assert)
+                return 0ul;
+                // LCOV_EXCL_STOP
+        }
+    }
+
+    /*!
+    @brief Serializes the JSON value @a j to BSON and associates it with the
+           key @a name.
+    @param name The name to associate with the JSON entity @a j within the
+                current BSON document
+    */
+    void write_bson_element(const string_t& name,
+                            const BasicJsonType& j)
+    {
+        switch (j.type())
+        {
+            case value_t::object:
+                return write_bson_object_entry(name, *j.m_value.object);
+
+            case value_t::array:
+                return write_bson_array(name, *j.m_value.array);
+
+            case value_t::binary:
+                return write_bson_binary(name, *j.m_value.binary);
+
+            case value_t::boolean:
+                return write_bson_boolean(name, j.m_value.boolean);
+
+            case value_t::number_float:
+                return write_bson_double(name, j.m_value.number_float);
+
+            case value_t::number_integer:
+                return write_bson_integer(name, j.m_value.number_integer);
+
+            case value_t::number_unsigned:
+                return write_bson_unsigned(name, j);
+
+            case value_t::string:
+                return write_bson_string(name, *j.m_value.string);
+
+            case value_t::null:
+                return write_bson_null(name);
+
+            // LCOV_EXCL_START
+            case value_t::discarded:
+            default:
+                JSON_ASSERT(false); // NOLINT(cert-dcl03-c,hicpp-static-assert,misc-static-assert)
+                return;
+                // LCOV_EXCL_STOP
+        }
+    }
+
+    /*!
+    @brief Calculates the size of the BSON serialization of the given
+           JSON-object @a j.
+    @param[in] value  JSON value to serialize
+    @pre       value.type() == value_t::object
+    */
+    static std::size_t calc_bson_object_size(const typename BasicJsonType::object_t& value)
+    {
+        std::size_t document_size = std::accumulate(value.begin(), value.end(), static_cast<std::size_t>(0),
+                                    [](size_t result, const typename BasicJsonType::object_t::value_type & el)
+        {
+            return result += calc_bson_element_size(el.first, el.second);
+        });
+
+        return sizeof(std::int32_t) + document_size + 1ul;
+    }
+
+    /*!
+    @param[in] value  JSON value to serialize
+    @pre       value.type() == value_t::object
+    */
+    void write_bson_object(const typename BasicJsonType::object_t& value)
+    {
+        write_number<std::int32_t>(static_cast<std::int32_t>(calc_bson_object_size(value)), true);
+
+        for (const auto& el : value)
+        {
+            write_bson_element(el.first, el.second);
+        }
+
+        oa->write_character(to_char_type(0x00));
+    }
+
+    //////////
+    // CBOR //
+    //////////
+
+    static constexpr CharType get_cbor_float_prefix(float /*unused*/)
+    {
+        return to_char_type(0xFA);  // Single-Precision Float
+    }
+
+    static constexpr CharType get_cbor_float_prefix(double /*unused*/)
+    {
+        return to_char_type(0xFB);  // Double-Precision Float
+    }
+
+    /////////////
+    // MsgPack //
+    /////////////
+
+    static constexpr CharType get_msgpack_float_prefix(float /*unused*/)
+    {
+        return to_char_type(0xCA);  // float 32
+    }
+
+    static constexpr CharType get_msgpack_float_prefix(double /*unused*/)
+    {
+        return to_char_type(0xCB);  // float 64
+    }
+
+    ////////////
+    // UBJSON //
+    ////////////
+
+    // UBJSON: write number (floating point)
+    template<typename NumberType, typename std::enable_if<
+                 std::is_floating_point<NumberType>::value, int>::type = 0>
+    void write_number_with_ubjson_prefix(const NumberType n,
+                                         const bool add_prefix,
+                                         const bool use_bjdata)
+    {
+        if (add_prefix)
+        {
+            oa->write_character(get_ubjson_float_prefix(n));
+        }
+        write_number(n, use_bjdata);
+    }
+
+    // UBJSON: write number (unsigned integer)
+    template<typename NumberType, typename std::enable_if<
+                 std::is_unsigned<NumberType>::value, int>::type = 0>
+    void write_number_with_ubjson_prefix(const NumberType n,
+                                         const bool add_prefix,
+                                         const bool use_bjdata)
+    {
+        if (n <= static_cast<std::uint64_t>((std::numeric_limits<std::int8_t>::max)()))
+        {
+            if (add_prefix)
+            {
+                oa->write_character(to_char_type('i'));  // int8
+            }
+            write_number(static_cast<std::uint8_t>(n), use_bjdata);
+        }
+        else if (n <= (std::numeric_limits<std::uint8_t>::max)())
+        {
+            if (add_prefix)
+            {
+                oa->write_character(to_char_type('U'));  // uint8
+            }
+            write_number(static_cast<std::uint8_t>(n), use_bjdata);
+        }
+        else if (n <= static_cast<std::uint64_t>((std::numeric_limits<std::int16_t>::max)()))
+        {
+            if (add_prefix)
+            {
+                oa->write_character(to_char_type('I'));  // int16
+            }
+            write_number(static_cast<std::int16_t>(n), use_bjdata);
+        }
+        else if (use_bjdata && n <= static_cast<uint64_t>((std::numeric_limits<uint16_t>::max)()))
+        {
+            if (add_prefix)
+            {
+                oa->write_character(to_char_type('u'));  // uint16 - bjdata only
+            }
+            write_number(static_cast<std::uint16_t>(n), use_bjdata);
+        }
+        else if (n <= static_cast<std::uint64_t>((std::numeric_limits<std::int32_t>::max)()))
+        {
+            if (add_prefix)
+            {
+                oa->write_character(to_char_type('l'));  // int32
+            }
+            write_number(static_cast<std::int32_t>(n), use_bjdata);
+        }
+        else if (use_bjdata && n <= static_cast<uint64_t>((std::numeric_limits<uint32_t>::max)()))
+        {
+            if (add_prefix)
+            {
+                oa->write_character(to_char_type('m'));  // uint32 - bjdata only
+            }
+            write_number(static_cast<std::uint32_t>(n), use_bjdata);
+        }
+        else if (n <= static_cast<std::uint64_t>((std::numeric_limits<std::int64_t>::max)()))
+        {
+            if (add_prefix)
+            {
+                oa->write_character(to_char_type('L'));  // int64
+            }
+            write_number(static_cast<std::int64_t>(n), use_bjdata);
+        }
+        else if (use_bjdata && n <= (std::numeric_limits<uint64_t>::max)())
+        {
+            if (add_prefix)
+            {
+                oa->write_character(to_char_type('M'));  // uint64 - bjdata only
+            }
+            write_number(static_cast<std::uint64_t>(n), use_bjdata);
+        }
+        else
+        {
+            if (add_prefix)
+            {
+                oa->write_character(to_char_type('H'));  // high-precision number
+            }
+
+            const auto number = BasicJsonType(n).dump();
+            write_number_with_ubjson_prefix(number.size(), true, use_bjdata);
+            for (std::size_t i = 0; i < number.size(); ++i)
+            {
+                oa->write_character(to_char_type(static_cast<std::uint8_t>(number[i])));
+            }
+        }
+    }
+
+    // UBJSON: write number (signed integer)
+    template < typename NumberType, typename std::enable_if <
+                   std::is_signed<NumberType>::value&&
+                   !std::is_floating_point<NumberType>::value, int >::type = 0 >
+    void write_number_with_ubjson_prefix(const NumberType n,
+                                         const bool add_prefix,
+                                         const bool use_bjdata)
+    {
+        if ((std::numeric_limits<std::int8_t>::min)() <= n && n <= (std::numeric_limits<std::int8_t>::max)())
+        {
+            if (add_prefix)
+            {
+                oa->write_character(to_char_type('i'));  // int8
+            }
+            write_number(static_cast<std::int8_t>(n), use_bjdata);
+        }
+        else if (static_cast<std::int64_t>((std::numeric_limits<std::uint8_t>::min)()) <= n && n <= static_cast<std::int64_t>((std::numeric_limits<std::uint8_t>::max)()))
+        {
+            if (add_prefix)
+            {
+                oa->write_character(to_char_type('U'));  // uint8
+            }
+            write_number(static_cast<std::uint8_t>(n), use_bjdata);
+        }
+        else if ((std::numeric_limits<std::int16_t>::min)() <= n && n <= (std::numeric_limits<std::int16_t>::max)())
+        {
+            if (add_prefix)
+            {
+                oa->write_character(to_char_type('I'));  // int16
+            }
+            write_number(static_cast<std::int16_t>(n), use_bjdata);
+        }
+        else if (use_bjdata && (static_cast<std::int64_t>((std::numeric_limits<std::uint16_t>::min)()) <= n && n <= static_cast<std::int64_t>((std::numeric_limits<std::uint16_t>::max)())))
+        {
+            if (add_prefix)
+            {
+                oa->write_character(to_char_type('u'));  // uint16 - bjdata only
+            }
+            write_number(static_cast<uint16_t>(n), use_bjdata);
+        }
+        else if ((std::numeric_limits<std::int32_t>::min)() <= n && n <= (std::numeric_limits<std::int32_t>::max)())
+        {
+            if (add_prefix)
+            {
+                oa->write_character(to_char_type('l'));  // int32
+            }
+            write_number(static_cast<std::int32_t>(n), use_bjdata);
+        }
+        else if (use_bjdata && (static_cast<std::int64_t>((std::numeric_limits<std::uint32_t>::min)()) <= n && n <= static_cast<std::int64_t>((std::numeric_limits<std::uint32_t>::max)())))
+        {
+            if (add_prefix)
+            {
+                oa->write_character(to_char_type('m'));  // uint32 - bjdata only
+            }
+            write_number(static_cast<uint32_t>(n), use_bjdata);
+        }
+        else if ((std::numeric_limits<std::int64_t>::min)() <= n && n <= (std::numeric_limits<std::int64_t>::max)())
+        {
+            if (add_prefix)
+            {
+                oa->write_character(to_char_type('L'));  // int64
+            }
+            write_number(static_cast<std::int64_t>(n), use_bjdata);
+        }
+        // LCOV_EXCL_START
+        else
+        {
+            if (add_prefix)
+            {
+                oa->write_character(to_char_type('H'));  // high-precision number
+            }
+
+            const auto number = BasicJsonType(n).dump();
+            write_number_with_ubjson_prefix(number.size(), true, use_bjdata);
+            for (std::size_t i = 0; i < number.size(); ++i)
+            {
+                oa->write_character(to_char_type(static_cast<std::uint8_t>(number[i])));
+            }
+        }
+        // LCOV_EXCL_STOP
+    }
+
+    /*!
+    @brief determine the type prefix of container values
+    */
+    CharType ubjson_prefix(const BasicJsonType& j, const bool use_bjdata) const noexcept
+    {
+        switch (j.type())
+        {
+            case value_t::null:
+                return 'Z';
+
+            case value_t::boolean:
+                return j.m_value.boolean ? 'T' : 'F';
+
+            case value_t::number_integer:
+            {
+                if ((std::numeric_limits<std::int8_t>::min)() <= j.m_value.number_integer && j.m_value.number_integer <= (std::numeric_limits<std::int8_t>::max)())
+                {
+                    return 'i';
+                }
+                if ((std::numeric_limits<std::uint8_t>::min)() <= j.m_value.number_integer && j.m_value.number_integer <= (std::numeric_limits<std::uint8_t>::max)())
+                {
+                    return 'U';
+                }
+                if ((std::numeric_limits<std::int16_t>::min)() <= j.m_value.number_integer && j.m_value.number_integer <= (std::numeric_limits<std::int16_t>::max)())
+                {
+                    return 'I';
+                }
+                if (use_bjdata && ((std::numeric_limits<std::uint16_t>::min)() <= j.m_value.number_integer && j.m_value.number_integer <= (std::numeric_limits<std::uint16_t>::max)()))
+                {
+                    return 'u';
+                }
+                if ((std::numeric_limits<std::int32_t>::min)() <= j.m_value.number_integer && j.m_value.number_integer <= (std::numeric_limits<std::int32_t>::max)())
+                {
+                    return 'l';
+                }
+                if (use_bjdata && ((std::numeric_limits<std::uint32_t>::min)() <= j.m_value.number_integer && j.m_value.number_integer <= (std::numeric_limits<std::uint32_t>::max)()))
+                {
+                    return 'm';
+                }
+                if ((std::numeric_limits<std::int64_t>::min)() <= j.m_value.number_integer && j.m_value.number_integer <= (std::numeric_limits<std::int64_t>::max)())
+                {
+                    return 'L';
+                }
+                // anything else is treated as high-precision number
+                return 'H'; // LCOV_EXCL_LINE
+            }
+
+            case value_t::number_unsigned:
+            {
+                if (j.m_value.number_unsigned <= static_cast<std::uint64_t>((std::numeric_limits<std::int8_t>::max)()))
+                {
+                    return 'i';
+                }
+                if (j.m_value.number_unsigned <= static_cast<std::uint64_t>((std::numeric_limits<std::uint8_t>::max)()))
+                {
+                    return 'U';
+                }
+                if (j.m_value.number_unsigned <= static_cast<std::uint64_t>((std::numeric_limits<std::int16_t>::max)()))
+                {
+                    return 'I';
+                }
+                if (use_bjdata && j.m_value.number_unsigned <= static_cast<std::uint64_t>((std::numeric_limits<std::uint16_t>::max)()))
+                {
+                    return 'u';
+                }
+                if (j.m_value.number_unsigned <= static_cast<std::uint64_t>((std::numeric_limits<std::int32_t>::max)()))
+                {
+                    return 'l';
+                }
+                if (use_bjdata && j.m_value.number_unsigned <= static_cast<std::uint64_t>((std::numeric_limits<std::uint32_t>::max)()))
+                {
+                    return 'm';
+                }
+                if (j.m_value.number_unsigned <= static_cast<std::uint64_t>((std::numeric_limits<std::int64_t>::max)()))
+                {
+                    return 'L';
+                }
+                if (use_bjdata && j.m_value.number_unsigned <= (std::numeric_limits<std::uint64_t>::max)())
+                {
+                    return 'M';
+                }
+                // anything else is treated as high-precision number
+                return 'H'; // LCOV_EXCL_LINE
+            }
+
+            case value_t::number_float:
+                return get_ubjson_float_prefix(j.m_value.number_float);
+
+            case value_t::string:
+                return 'S';
+
+            case value_t::array: // fallthrough
+            case value_t::binary:
+                return '[';
+
+            case value_t::object:
+                return '{';
+
+            case value_t::discarded:
+            default:  // discarded values
+                return 'N';
+        }
+    }
+
+    static constexpr CharType get_ubjson_float_prefix(float /*unused*/)
+    {
+        return 'd';  // float 32
+    }
+
+    static constexpr CharType get_ubjson_float_prefix(double /*unused*/)
+    {
+        return 'D';  // float 64
+    }
+
+    /*!
+    @return false if the object is successfully converted to a bjdata ndarray, true if the type or size is invalid
+    */
+    bool write_bjdata_ndarray(const typename BasicJsonType::object_t& value, const bool use_count, const bool use_type)
+    {
+        std::map<string_t, CharType> bjdtype = {{"uint8", 'U'},  {"int8", 'i'},  {"uint16", 'u'}, {"int16", 'I'},
+            {"uint32", 'm'}, {"int32", 'l'}, {"uint64", 'M'}, {"int64", 'L'}, {"single", 'd'}, {"double", 'D'}, {"char", 'C'}
+        };
+
+        string_t key = "_ArrayType_";
+        auto it = bjdtype.find(static_cast<string_t>(value.at(key)));
+        if (it == bjdtype.end())
+        {
+            return true;
+        }
+        CharType dtype = it->second;
+
+        key = "_ArraySize_";
+        std::size_t len = (value.at(key).empty() ? 0 : 1);
+        for (const auto& el : value.at(key))
+        {
+            len *= static_cast<std::size_t>(el.m_value.number_unsigned);
+        }
+
+        key = "_ArrayData_";
+        if (value.at(key).size() != len)
+        {
+            return true;
+        }
+
+        oa->write_character('[');
+        oa->write_character('$');
+        oa->write_character(dtype);
+        oa->write_character('#');
+
+        key = "_ArraySize_";
+        write_ubjson(value.at(key), use_count, use_type, true,  true);
+
+        key = "_ArrayData_";
+        if (dtype == 'U' || dtype == 'C')
+        {
+            for (const auto& el : value.at(key))
+            {
+                write_number(static_cast<std::uint8_t>(el.m_value.number_unsigned), true);
+            }
+        }
+        else if (dtype == 'i')
+        {
+            for (const auto& el : value.at(key))
+            {
+                write_number(static_cast<std::int8_t>(el.m_value.number_integer), true);
+            }
+        }
+        else if (dtype == 'u')
+        {
+            for (const auto& el : value.at(key))
+            {
+                write_number(static_cast<std::uint16_t>(el.m_value.number_unsigned), true);
+            }
+        }
+        else if (dtype == 'I')
+        {
+            for (const auto& el : value.at(key))
+            {
+                write_number(static_cast<std::int16_t>(el.m_value.number_integer), true);
+            }
+        }
+        else if (dtype == 'm')
+        {
+            for (const auto& el : value.at(key))
+            {
+                write_number(static_cast<std::uint32_t>(el.m_value.number_unsigned), true);
+            }
+        }
+        else if (dtype == 'l')
+        {
+            for (const auto& el : value.at(key))
+            {
+                write_number(static_cast<std::int32_t>(el.m_value.number_integer), true);
+            }
+        }
+        else if (dtype == 'M')
+        {
+            for (const auto& el : value.at(key))
+            {
+                write_number(static_cast<std::uint64_t>(el.m_value.number_unsigned), true);
+            }
+        }
+        else if (dtype == 'L')
+        {
+            for (const auto& el : value.at(key))
+            {
+                write_number(static_cast<std::int64_t>(el.m_value.number_integer), true);
+            }
+        }
+        else if (dtype == 'd')
+        {
+            for (const auto& el : value.at(key))
+            {
+                write_number(static_cast<float>(el.m_value.number_float), true);
+            }
+        }
+        else if (dtype == 'D')
+        {
+            for (const auto& el : value.at(key))
+            {
+                write_number(static_cast<double>(el.m_value.number_float), true);
+            }
+        }
+        return false;
+    }
+
+    ///////////////////////
+    // Utility functions //
+    ///////////////////////
+
+    /*
+    @brief write a number to output input
+    @param[in] n number of type @a NumberType
+    @param[in] OutputIsLittleEndian Set to true if output data is
+                                 required to be little endian
+    @tparam NumberType the type of the number
+
+    @note This function needs to respect the system's endianness, because bytes
+          in CBOR, MessagePack, and UBJSON are stored in network order (big
+          endian) and therefore need reordering on little endian systems.
+          On the other hand, BSON and BJData use little endian and should reorder
+          on big endian systems.
+    */
+    template<typename NumberType>
+    void write_number(const NumberType n, const bool OutputIsLittleEndian = false)
+    {
+        // step 1: write number to array of length NumberType
+        std::array<CharType, sizeof(NumberType)> vec{};
+        std::memcpy(vec.data(), &n, sizeof(NumberType));
+
+        // step 2: write array to output (with possible reordering)
+        if (is_little_endian != OutputIsLittleEndian)
+        {
+            // reverse byte order prior to conversion if necessary
+            std::reverse(vec.begin(), vec.end());
+        }
+
+        oa->write_characters(vec.data(), sizeof(NumberType));
+    }
+
+    void write_compact_float(const number_float_t n, detail::input_format_t format)
+    {
+#ifdef __GNUC__
+#pragma GCC diagnostic push
+#pragma GCC diagnostic ignored "-Wfloat-equal"
+#endif
+        if (static_cast<double>(n) >= static_cast<double>(std::numeric_limits<float>::lowest()) &&
+                static_cast<double>(n) <= static_cast<double>((std::numeric_limits<float>::max)()) &&
+                static_cast<double>(static_cast<float>(n)) == static_cast<double>(n))
+        {
+            oa->write_character(format == detail::input_format_t::cbor
+                                ? get_cbor_float_prefix(static_cast<float>(n))
+                                : get_msgpack_float_prefix(static_cast<float>(n)));
+            write_number(static_cast<float>(n));
+        }
+        else
+        {
+            oa->write_character(format == detail::input_format_t::cbor
+                                ? get_cbor_float_prefix(n)
+                                : get_msgpack_float_prefix(n));
+            write_number(n);
+        }
+#ifdef __GNUC__
+#pragma GCC diagnostic pop
+#endif
+    }
+
+  public:
+    // The following to_char_type functions are implement the conversion
+    // between uint8_t and CharType. In case CharType is not unsigned,
+    // such a conversion is required to allow values greater than 128.
+    // See <https://github.com/nlohmann/json/issues/1286> for a discussion.
+    template < typename C = CharType,
+               enable_if_t < std::is_signed<C>::value && std::is_signed<char>::value > * = nullptr >
+    static constexpr CharType to_char_type(std::uint8_t x) noexcept
+    {
+        return *reinterpret_cast<char*>(&x);
+    }
+
+    template < typename C = CharType,
+               enable_if_t < std::is_signed<C>::value && std::is_unsigned<char>::value > * = nullptr >
+    static CharType to_char_type(std::uint8_t x) noexcept
+    {
+        static_assert(sizeof(std::uint8_t) == sizeof(CharType), "size of CharType must be equal to std::uint8_t");
+        static_assert(std::is_trivial<CharType>::value, "CharType must be trivial");
+        CharType result;
+        std::memcpy(&result, &x, sizeof(x));
+        return result;
+    }
+
+    template<typename C = CharType,
+             enable_if_t<std::is_unsigned<C>::value>* = nullptr>
+    static constexpr CharType to_char_type(std::uint8_t x) noexcept
+    {
+        return x;
+    }
+
+    template < typename InputCharType, typename C = CharType,
+               enable_if_t <
+                   std::is_signed<C>::value &&
+                   std::is_signed<char>::value &&
+                   std::is_same<char, typename std::remove_cv<InputCharType>::type>::value
+                   > * = nullptr >
+    static constexpr CharType to_char_type(InputCharType x) noexcept
+    {
+        return x;
+    }
+
+  private:
+    /// whether we can assume little endianness
+    const bool is_little_endian = little_endianness();
+
+    /// the output
+    output_adapter_t<CharType> oa = nullptr;
+};
+
+}  // namespace detail
+NLOHMANN_JSON_NAMESPACE_END
+
+// #include <nlohmann/detail/output/output_adapters.hpp>
+
+// #include <nlohmann/detail/output/serializer.hpp>
+//     __ _____ _____ _____
+//  __|  |   __|     |   | |  JSON for Modern C++
+// |  |  |__   |  |  | | | |  version 3.11.2
+// |_____|_____|_____|_|___|  https://github.com/nlohmann/json
+//
+// SPDX-FileCopyrightText: 2008-2009 Björn Hoehrmann <bjoern@hoehrmann.de>
+// SPDX-FileCopyrightText: 2013-2022 Niels Lohmann <https://nlohmann.me>
+// SPDX-License-Identifier: MIT
+
+
+
+#include <algorithm> // reverse, remove, fill, find, none_of
+#include <array> // array
+#include <clocale> // localeconv, lconv
+#include <cmath> // labs, isfinite, isnan, signbit
+#include <cstddef> // size_t, ptrdiff_t
+#include <cstdint> // uint8_t
+#include <cstdio> // snprintf
+#include <limits> // numeric_limits
+#include <string> // string, char_traits
+#include <iomanip> // setfill, setw
+#include <type_traits> // is_same
+#include <utility> // move
+
+// #include <nlohmann/detail/conversions/to_chars.hpp>
+//     __ _____ _____ _____
+//  __|  |   __|     |   | |  JSON for Modern C++
+// |  |  |__   |  |  | | | |  version 3.11.2
+// |_____|_____|_____|_|___|  https://github.com/nlohmann/json
+//
+// SPDX-FileCopyrightText: 2009 Florian Loitsch <https://florian.loitsch.com/>
+// SPDX-FileCopyrightText: 2013-2022 Niels Lohmann <https://nlohmann.me>
+// SPDX-License-Identifier: MIT
+
+
+
+#include <array> // array
+#include <cmath>   // signbit, isfinite
+#include <cstdint> // intN_t, uintN_t
+#include <cstring> // memcpy, memmove
+#include <limits> // numeric_limits
+#include <type_traits> // conditional
+
+// #include <nlohmann/detail/macro_scope.hpp>
+
+
+NLOHMANN_JSON_NAMESPACE_BEGIN
+namespace detail
+{
+
+/*!
+@brief implements the Grisu2 algorithm for binary to decimal floating-point
+conversion.
+
+This implementation is a slightly modified version of the reference
+implementation which may be obtained from
+http://florian.loitsch.com/publications (bench.tar.gz).
+
+The code is distributed under the MIT license, Copyright (c) 2009 Florian Loitsch.
+
+For a detailed description of the algorithm see:
+
+[1] Loitsch, "Printing Floating-Point Numbers Quickly and Accurately with
+    Integers", Proceedings of the ACM SIGPLAN 2010 Conference on Programming
+    Language Design and Implementation, PLDI 2010
+[2] Burger, Dybvig, "Printing Floating-Point Numbers Quickly and Accurately",
+    Proceedings of the ACM SIGPLAN 1996 Conference on Programming Language
+    Design and Implementation, PLDI 1996
+*/
+namespace dtoa_impl
+{
+
+template<typename Target, typename Source>
+Target reinterpret_bits(const Source source)
+{
+    static_assert(sizeof(Target) == sizeof(Source), "size mismatch");
+
+    Target target;
+    std::memcpy(&target, &source, sizeof(Source));
+    return target;
+}
+
+struct diyfp // f * 2^e
+{
+    static constexpr int kPrecision = 64; // = q
+
+    std::uint64_t f = 0;
+    int e = 0;
+
+    constexpr diyfp(std::uint64_t f_, int e_) noexcept : f(f_), e(e_) {}
+
+    /*!
+    @brief returns x - y
+    @pre x.e == y.e and x.f >= y.f
+    */
+    static diyfp sub(const diyfp& x, const diyfp& y) noexcept
+    {
+        JSON_ASSERT(x.e == y.e);
+        JSON_ASSERT(x.f >= y.f);
+
+        return {x.f - y.f, x.e};
+    }
+
+    /*!
+    @brief returns x * y
+    @note The result is rounded. (Only the upper q bits are returned.)
+    */
+    static diyfp mul(const diyfp& x, const diyfp& y) noexcept
+    {
+        static_assert(kPrecision == 64, "internal error");
+
+        // Computes:
+        //  f = round((x.f * y.f) / 2^q)
+        //  e = x.e + y.e + q
+
+        // Emulate the 64-bit * 64-bit multiplication:
+        //
+        // p = u * v
+        //   = (u_lo + 2^32 u_hi) (v_lo + 2^32 v_hi)
+        //   = (u_lo v_lo         ) + 2^32 ((u_lo v_hi         ) + (u_hi v_lo         )) + 2^64 (u_hi v_hi         )
+        //   = (p0                ) + 2^32 ((p1                ) + (p2                )) + 2^64 (p3                )
+        //   = (p0_lo + 2^32 p0_hi) + 2^32 ((p1_lo + 2^32 p1_hi) + (p2_lo + 2^32 p2_hi)) + 2^64 (p3                )
+        //   = (p0_lo             ) + 2^32 (p0_hi + p1_lo + p2_lo                      ) + 2^64 (p1_hi + p2_hi + p3)
+        //   = (p0_lo             ) + 2^32 (Q                                          ) + 2^64 (H                 )
+        //   = (p0_lo             ) + 2^32 (Q_lo + 2^32 Q_hi                           ) + 2^64 (H                 )
+        //
+        // (Since Q might be larger than 2^32 - 1)
+        //
+        //   = (p0_lo + 2^32 Q_lo) + 2^64 (Q_hi + H)
+        //
+        // (Q_hi + H does not overflow a 64-bit int)
+        //
+        //   = p_lo + 2^64 p_hi
+
+        const std::uint64_t u_lo = x.f & 0xFFFFFFFFu;
+        const std::uint64_t u_hi = x.f >> 32u;
+        const std::uint64_t v_lo = y.f & 0xFFFFFFFFu;
+        const std::uint64_t v_hi = y.f >> 32u;
+
+        const std::uint64_t p0 = u_lo * v_lo;
+        const std::uint64_t p1 = u_lo * v_hi;
+        const std::uint64_t p2 = u_hi * v_lo;
+        const std::uint64_t p3 = u_hi * v_hi;
+
+        const std::uint64_t p0_hi = p0 >> 32u;
+        const std::uint64_t p1_lo = p1 & 0xFFFFFFFFu;
+        const std::uint64_t p1_hi = p1 >> 32u;
+        const std::uint64_t p2_lo = p2 & 0xFFFFFFFFu;
+        const std::uint64_t p2_hi = p2 >> 32u;
+
+        std::uint64_t Q = p0_hi + p1_lo + p2_lo;
+
+        // The full product might now be computed as
+        //
+        // p_hi = p3 + p2_hi + p1_hi + (Q >> 32)
+        // p_lo = p0_lo + (Q << 32)
+        //
+        // But in this particular case here, the full p_lo is not required.
+        // Effectively we only need to add the highest bit in p_lo to p_hi (and
+        // Q_hi + 1 does not overflow).
+
+        Q += std::uint64_t{1} << (64u - 32u - 1u); // round, ties up
+
+        const std::uint64_t h = p3 + p2_hi + p1_hi + (Q >> 32u);
+
+        return {h, x.e + y.e + 64};
+    }
+
+    /*!
+    @brief normalize x such that the significand is >= 2^(q-1)
+    @pre x.f != 0
+    */
+    static diyfp normalize(diyfp x) noexcept
+    {
+        JSON_ASSERT(x.f != 0);
+
+        while ((x.f >> 63u) == 0)
+        {
+            x.f <<= 1u;
+            x.e--;
+        }
+
+        return x;
+    }
+
+    /*!
+    @brief normalize x such that the result has the exponent E
+    @pre e >= x.e and the upper e - x.e bits of x.f must be zero.
+    */
+    static diyfp normalize_to(const diyfp& x, const int target_exponent) noexcept
+    {
+        const int delta = x.e - target_exponent;
+
+        JSON_ASSERT(delta >= 0);
+        JSON_ASSERT(((x.f << delta) >> delta) == x.f);
+
+        return {x.f << delta, target_exponent};
+    }
+};
+
+struct boundaries
+{
+    diyfp w;
+    diyfp minus;
+    diyfp plus;
+};
+
+/*!
+Compute the (normalized) diyfp representing the input number 'value' and its
+boundaries.
+
+@pre value must be finite and positive
+*/
+template<typename FloatType>
+boundaries compute_boundaries(FloatType value)
+{
+    JSON_ASSERT(std::isfinite(value));
+    JSON_ASSERT(value > 0);
+
+    // Convert the IEEE representation into a diyfp.
+    //
+    // If v is denormal:
+    //      value = 0.F * 2^(1 - bias) = (          F) * 2^(1 - bias - (p-1))
+    // If v is normalized:
+    //      value = 1.F * 2^(E - bias) = (2^(p-1) + F) * 2^(E - bias - (p-1))
+
+    static_assert(std::numeric_limits<FloatType>::is_iec559,
+                  "internal error: dtoa_short requires an IEEE-754 floating-point implementation");
+
+    constexpr int      kPrecision = std::numeric_limits<FloatType>::digits; // = p (includes the hidden bit)
+    constexpr int      kBias      = std::numeric_limits<FloatType>::max_exponent - 1 + (kPrecision - 1);
+    constexpr int      kMinExp    = 1 - kBias;
+    constexpr std::uint64_t kHiddenBit = std::uint64_t{1} << (kPrecision - 1); // = 2^(p-1)
+
+    using bits_type = typename std::conditional<kPrecision == 24, std::uint32_t, std::uint64_t >::type;
+
+    const auto bits = static_cast<std::uint64_t>(reinterpret_bits<bits_type>(value));
+    const std::uint64_t E = bits >> (kPrecision - 1);
+    const std::uint64_t F = bits & (kHiddenBit - 1);
+
+    const bool is_denormal = E == 0;
+    const diyfp v = is_denormal
+                    ? diyfp(F, kMinExp)
+                    : diyfp(F + kHiddenBit, static_cast<int>(E) - kBias);
+
+    // Compute the boundaries m- and m+ of the floating-point value
+    // v = f * 2^e.
+    //
+    // Determine v- and v+, the floating-point predecessor and successor if v,
+    // respectively.
+    //
+    //      v- = v - 2^e        if f != 2^(p-1) or e == e_min                (A)
+    //         = v - 2^(e-1)    if f == 2^(p-1) and e > e_min                (B)
+    //
+    //      v+ = v + 2^e
+    //
+    // Let m- = (v- + v) / 2 and m+ = (v + v+) / 2. All real numbers _strictly_
+    // between m- and m+ round to v, regardless of how the input rounding
+    // algorithm breaks ties.
+    //
+    //      ---+-------------+-------------+-------------+-------------+---  (A)
+    //         v-            m-            v             m+            v+
+    //
+    //      -----------------+------+------+-------------+-------------+---  (B)
+    //                       v-     m-     v             m+            v+
+
+    const bool lower_boundary_is_closer = F == 0 && E > 1;
+    const diyfp m_plus = diyfp(2 * v.f + 1, v.e - 1);
+    const diyfp m_minus = lower_boundary_is_closer
+                          ? diyfp(4 * v.f - 1, v.e - 2)  // (B)
+                          : diyfp(2 * v.f - 1, v.e - 1); // (A)
+
+    // Determine the normalized w+ = m+.
+    const diyfp w_plus = diyfp::normalize(m_plus);
+
+    // Determine w- = m- such that e_(w-) = e_(w+).
+    const diyfp w_minus = diyfp::normalize_to(m_minus, w_plus.e);
+
+    return {diyfp::normalize(v), w_minus, w_plus};
+}
+
+// Given normalized diyfp w, Grisu needs to find a (normalized) cached
+// power-of-ten c, such that the exponent of the product c * w = f * 2^e lies
+// within a certain range [alpha, gamma] (Definition 3.2 from [1])
+//
+//      alpha <= e = e_c + e_w + q <= gamma
+//
+// or
+//
+//      f_c * f_w * 2^alpha <= f_c 2^(e_c) * f_w 2^(e_w) * 2^q
+//                          <= f_c * f_w * 2^gamma
+//
+// Since c and w are normalized, i.e. 2^(q-1) <= f < 2^q, this implies
+//
+//      2^(q-1) * 2^(q-1) * 2^alpha <= c * w * 2^q < 2^q * 2^q * 2^gamma
+//
+// or
+//
+//      2^(q - 2 + alpha) <= c * w < 2^(q + gamma)
+//
+// The choice of (alpha,gamma) determines the size of the table and the form of
+// the digit generation procedure. Using (alpha,gamma)=(-60,-32) works out well
+// in practice:
+//
+// The idea is to cut the number c * w = f * 2^e into two parts, which can be
+// processed independently: An integral part p1, and a fractional part p2:
+//
+//      f * 2^e = ( (f div 2^-e) * 2^-e + (f mod 2^-e) ) * 2^e
+//              = (f div 2^-e) + (f mod 2^-e) * 2^e
+//              = p1 + p2 * 2^e
+//
+// The conversion of p1 into decimal form requires a series of divisions and
+// modulos by (a power of) 10. These operations are faster for 32-bit than for
+// 64-bit integers, so p1 should ideally fit into a 32-bit integer. This can be
+// achieved by choosing
+//
+//      -e >= 32   or   e <= -32 := gamma
+//
+// In order to convert the fractional part
+//
+//      p2 * 2^e = p2 / 2^-e = d[-1] / 10^1 + d[-2] / 10^2 + ...
+//
+// into decimal form, the fraction is repeatedly multiplied by 10 and the digits
+// d[-i] are extracted in order:
+//
+//      (10 * p2) div 2^-e = d[-1]
+//      (10 * p2) mod 2^-e = d[-2] / 10^1 + ...
+//
+// The multiplication by 10 must not overflow. It is sufficient to choose
+//
+//      10 * p2 < 16 * p2 = 2^4 * p2 <= 2^64.
+//
+// Since p2 = f mod 2^-e < 2^-e,
+//
+//      -e <= 60   or   e >= -60 := alpha
+
+constexpr int kAlpha = -60;
+constexpr int kGamma = -32;
+
+struct cached_power // c = f * 2^e ~= 10^k
+{
+    std::uint64_t f;
+    int e;
+    int k;
+};
+
+/*!
+For a normalized diyfp w = f * 2^e, this function returns a (normalized) cached
+power-of-ten c = f_c * 2^e_c, such that the exponent of the product w * c
+satisfies (Definition 3.2 from [1])
+
+     alpha <= e_c + e + q <= gamma.
+*/
+inline cached_power get_cached_power_for_binary_exponent(int e)
+{
+    // Now
+    //
+    //      alpha <= e_c + e + q <= gamma                                    (1)
+    //      ==> f_c * 2^alpha <= c * 2^e * 2^q
+    //
+    // and since the c's are normalized, 2^(q-1) <= f_c,
+    //
+    //      ==> 2^(q - 1 + alpha) <= c * 2^(e + q)
+    //      ==> 2^(alpha - e - 1) <= c
+    //
+    // If c were an exact power of ten, i.e. c = 10^k, one may determine k as
+    //
+    //      k = ceil( log_10( 2^(alpha - e - 1) ) )
+    //        = ceil( (alpha - e - 1) * log_10(2) )
+    //
+    // From the paper:
+    // "In theory the result of the procedure could be wrong since c is rounded,
+    //  and the computation itself is approximated [...]. In practice, however,
+    //  this simple function is sufficient."
+    //
+    // For IEEE double precision floating-point numbers converted into
+    // normalized diyfp's w = f * 2^e, with q = 64,
+    //
+    //      e >= -1022      (min IEEE exponent)
+    //           -52        (p - 1)
+    //           -52        (p - 1, possibly normalize denormal IEEE numbers)
+    //           -11        (normalize the diyfp)
+    //         = -1137
+    //
+    // and
+    //
+    //      e <= +1023      (max IEEE exponent)
+    //           -52        (p - 1)
+    //           -11        (normalize the diyfp)
+    //         = 960
+    //
+    // This binary exponent range [-1137,960] results in a decimal exponent
+    // range [-307,324]. One does not need to store a cached power for each
+    // k in this range. For each such k it suffices to find a cached power
+    // such that the exponent of the product lies in [alpha,gamma].
+    // This implies that the difference of the decimal exponents of adjacent
+    // table entries must be less than or equal to
+    //
+    //      floor( (gamma - alpha) * log_10(2) ) = 8.
+    //
+    // (A smaller distance gamma-alpha would require a larger table.)
+
+    // NB:
+    // Actually this function returns c, such that -60 <= e_c + e + 64 <= -34.
+
+    constexpr int kCachedPowersMinDecExp = -300;
+    constexpr int kCachedPowersDecStep = 8;
+
+    static constexpr std::array<cached_power, 79> kCachedPowers =
+    {
+        {
+            { 0xAB70FE17C79AC6CA, -1060, -300 },
+            { 0xFF77B1FCBEBCDC4F, -1034, -292 },
+            { 0xBE5691EF416BD60C, -1007, -284 },
+            { 0x8DD01FAD907FFC3C,  -980, -276 },
+            { 0xD3515C2831559A83,  -954, -268 },
+            { 0x9D71AC8FADA6C9B5,  -927, -260 },
+            { 0xEA9C227723EE8BCB,  -901, -252 },
+            { 0xAECC49914078536D,  -874, -244 },
+            { 0x823C12795DB6CE57,  -847, -236 },
+            { 0xC21094364DFB5637,  -821, -228 },
+            { 0x9096EA6F3848984F,  -794, -220 },
+            { 0xD77485CB25823AC7,  -768, -212 },
+            { 0xA086CFCD97BF97F4,  -741, -204 },
+            { 0xEF340A98172AACE5,  -715, -196 },
+            { 0xB23867FB2A35B28E,  -688, -188 },
+            { 0x84C8D4DFD2C63F3B,  -661, -180 },
+            { 0xC5DD44271AD3CDBA,  -635, -172 },
+            { 0x936B9FCEBB25C996,  -608, -164 },
+            { 0xDBAC6C247D62A584,  -582, -156 },
+            { 0xA3AB66580D5FDAF6,  -555, -148 },
+            { 0xF3E2F893DEC3F126,  -529, -140 },
+            { 0xB5B5ADA8AAFF80B8,  -502, -132 },
+            { 0x87625F056C7C4A8B,  -475, -124 },
+            { 0xC9BCFF6034C13053,  -449, -116 },
+            { 0x964E858C91BA2655,  -422, -108 },
+            { 0xDFF9772470297EBD,  -396, -100 },
+            { 0xA6DFBD9FB8E5B88F,  -369,  -92 },
+            { 0xF8A95FCF88747D94,  -343,  -84 },
+            { 0xB94470938FA89BCF,  -316,  -76 },
+            { 0x8A08F0F8BF0F156B,  -289,  -68 },
+            { 0xCDB02555653131B6,  -263,  -60 },
+            { 0x993FE2C6D07B7FAC,  -236,  -52 },
+            { 0xE45C10C42A2B3B06,  -210,  -44 },
+            { 0xAA242499697392D3,  -183,  -36 },
+            { 0xFD87B5F28300CA0E,  -157,  -28 },
+            { 0xBCE5086492111AEB,  -130,  -20 },
+            { 0x8CBCCC096F5088CC,  -103,  -12 },
+            { 0xD1B71758E219652C,   -77,   -4 },
+            { 0x9C40000000000000,   -50,    4 },
+            { 0xE8D4A51000000000,   -24,   12 },
+            { 0xAD78EBC5AC620000,     3,   20 },
+            { 0x813F3978F8940984,    30,   28 },
+            { 0xC097CE7BC90715B3,    56,   36 },
+            { 0x8F7E32CE7BEA5C70,    83,   44 },
+            { 0xD5D238A4ABE98068,   109,   52 },
+            { 0x9F4F2726179A2245,   136,   60 },
+            { 0xED63A231D4C4FB27,   162,   68 },
+            { 0xB0DE65388CC8ADA8,   189,   76 },
+            { 0x83C7088E1AAB65DB,   216,   84 },
+            { 0xC45D1DF942711D9A,   242,   92 },
+            { 0x924D692CA61BE758,   269,  100 },
+            { 0xDA01EE641A708DEA,   295,  108 },
+            { 0xA26DA3999AEF774A,   322,  116 },
+            { 0xF209787BB47D6B85,   348,  124 },
+            { 0xB454E4A179DD1877,   375,  132 },
+            { 0x865B86925B9BC5C2,   402,  140 },
+            { 0xC83553C5C8965D3D,   428,  148 },
+            { 0x952AB45CFA97A0B3,   455,  156 },
+            { 0xDE469FBD99A05FE3,   481,  164 },
+            { 0xA59BC234DB398C25,   508,  172 },
+            { 0xF6C69A72A3989F5C,   534,  180 },
+            { 0xB7DCBF5354E9BECE,   561,  188 },
+            { 0x88FCF317F22241E2,   588,  196 },
+            { 0xCC20CE9BD35C78A5,   614,  204 },
+            { 0x98165AF37B2153DF,   641,  212 },
+            { 0xE2A0B5DC971F303A,   667,  220 },
+            { 0xA8D9D1535CE3B396,   694,  228 },
+            { 0xFB9B7CD9A4A7443C,   720,  236 },
+            { 0xBB764C4CA7A44410,   747,  244 },
+            { 0x8BAB8EEFB6409C1A,   774,  252 },
+            { 0xD01FEF10A657842C,   800,  260 },
+            { 0x9B10A4E5E9913129,   827,  268 },
+            { 0xE7109BFBA19C0C9D,   853,  276 },
+            { 0xAC2820D9623BF429,   880,  284 },
+            { 0x80444B5E7AA7CF85,   907,  292 },
+            { 0xBF21E44003ACDD2D,   933,  300 },
+            { 0x8E679C2F5E44FF8F,   960,  308 },
+            { 0xD433179D9C8CB841,   986,  316 },
+            { 0x9E19DB92B4E31BA9,  1013,  324 },
+        }
+    };
+
+    // This computation gives exactly the same results for k as
+    //      k = ceil((kAlpha - e - 1) * 0.30102999566398114)
+    // for |e| <= 1500, but doesn't require floating-point operations.
+    // NB: log_10(2) ~= 78913 / 2^18
+    JSON_ASSERT(e >= -1500);
+    JSON_ASSERT(e <=  1500);
+    const int f = kAlpha - e - 1;
+    const int k = (f * 78913) / (1 << 18) + static_cast<int>(f > 0);
+
+    const int index = (-kCachedPowersMinDecExp + k + (kCachedPowersDecStep - 1)) / kCachedPowersDecStep;
+    JSON_ASSERT(index >= 0);
+    JSON_ASSERT(static_cast<std::size_t>(index) < kCachedPowers.size());
+
+    const cached_power cached = kCachedPowers[static_cast<std::size_t>(index)];
+    JSON_ASSERT(kAlpha <= cached.e + e + 64);
+    JSON_ASSERT(kGamma >= cached.e + e + 64);
+
+    return cached;
+}
+
+/*!
+For n != 0, returns k, such that pow10 := 10^(k-1) <= n < 10^k.
+For n == 0, returns 1 and sets pow10 := 1.
+*/
+inline int find_largest_pow10(const std::uint32_t n, std::uint32_t& pow10)
+{
+    // LCOV_EXCL_START
+    if (n >= 1000000000)
+    {
+        pow10 = 1000000000;
+        return 10;
+    }
+    // LCOV_EXCL_STOP
+    if (n >= 100000000)
+    {
+        pow10 = 100000000;
+        return  9;
+    }
+    if (n >= 10000000)
+    {
+        pow10 = 10000000;
+        return  8;
+    }
+    if (n >= 1000000)
+    {
+        pow10 = 1000000;
+        return  7;
+    }
+    if (n >= 100000)
+    {
+        pow10 = 100000;
+        return  6;
+    }
+    if (n >= 10000)
+    {
+        pow10 = 10000;
+        return  5;
+    }
+    if (n >= 1000)
+    {
+        pow10 = 1000;
+        return  4;
+    }
+    if (n >= 100)
+    {
+        pow10 = 100;
+        return  3;
+    }
+    if (n >= 10)
+    {
+        pow10 = 10;
+        return  2;
+    }
+
+    pow10 = 1;
+    return 1;
+}
+
+inline void grisu2_round(char* buf, int len, std::uint64_t dist, std::uint64_t delta,
+                         std::uint64_t rest, std::uint64_t ten_k)
+{
+    JSON_ASSERT(len >= 1);
+    JSON_ASSERT(dist <= delta);
+    JSON_ASSERT(rest <= delta);
+    JSON_ASSERT(ten_k > 0);
+
+    //               <--------------------------- delta ---->
+    //                                  <---- dist --------->
+    // --------------[------------------+-------------------]--------------
+    //               M-                 w                   M+
+    //
+    //                                  ten_k
+    //                                <------>
+    //                                       <---- rest ---->
+    // --------------[------------------+----+--------------]--------------
+    //                                  w    V
+    //                                       = buf * 10^k
+    //
+    // ten_k represents a unit-in-the-last-place in the decimal representation
+    // stored in buf.
+    // Decrement buf by ten_k while this takes buf closer to w.
+
+    // The tests are written in this order to avoid overflow in unsigned
+    // integer arithmetic.
+
+    while (rest < dist
+            && delta - rest >= ten_k
+            && (rest + ten_k < dist || dist - rest > rest + ten_k - dist))
+    {
+        JSON_ASSERT(buf[len - 1] != '0');
+        buf[len - 1]--;
+        rest += ten_k;
+    }
+}
+
+/*!
+Generates V = buffer * 10^decimal_exponent, such that M- <= V <= M+.
+M- and M+ must be normalized and share the same exponent -60 <= e <= -32.
+*/
+inline void grisu2_digit_gen(char* buffer, int& length, int& decimal_exponent,
+                             diyfp M_minus, diyfp w, diyfp M_plus)
+{
+    static_assert(kAlpha >= -60, "internal error");
+    static_assert(kGamma <= -32, "internal error");
+
+    // Generates the digits (and the exponent) of a decimal floating-point
+    // number V = buffer * 10^decimal_exponent in the range [M-, M+]. The diyfp's
+    // w, M- and M+ share the same exponent e, which satisfies alpha <= e <= gamma.
+    //
+    //               <--------------------------- delta ---->
+    //                                  <---- dist --------->
+    // --------------[------------------+-------------------]--------------
+    //               M-                 w                   M+
+    //
+    // Grisu2 generates the digits of M+ from left to right and stops as soon as
+    // V is in [M-,M+].
+
+    JSON_ASSERT(M_plus.e >= kAlpha);
+    JSON_ASSERT(M_plus.e <= kGamma);
+
+    std::uint64_t delta = diyfp::sub(M_plus, M_minus).f; // (significand of (M+ - M-), implicit exponent is e)
+    std::uint64_t dist  = diyfp::sub(M_plus, w      ).f; // (significand of (M+ - w ), implicit exponent is e)
+
+    // Split M+ = f * 2^e into two parts p1 and p2 (note: e < 0):
+    //
+    //      M+ = f * 2^e
+    //         = ((f div 2^-e) * 2^-e + (f mod 2^-e)) * 2^e
+    //         = ((p1        ) * 2^-e + (p2        )) * 2^e
+    //         = p1 + p2 * 2^e
+
+    const diyfp one(std::uint64_t{1} << -M_plus.e, M_plus.e);
+
+    auto p1 = static_cast<std::uint32_t>(M_plus.f >> -one.e); // p1 = f div 2^-e (Since -e >= 32, p1 fits into a 32-bit int.)
+    std::uint64_t p2 = M_plus.f & (one.f - 1);                    // p2 = f mod 2^-e
+
+    // 1)
+    //
+    // Generate the digits of the integral part p1 = d[n-1]...d[1]d[0]
+
+    JSON_ASSERT(p1 > 0);
+
+    std::uint32_t pow10{};
+    const int k = find_largest_pow10(p1, pow10);
+
+    //      10^(k-1) <= p1 < 10^k, pow10 = 10^(k-1)
+    //
+    //      p1 = (p1 div 10^(k-1)) * 10^(k-1) + (p1 mod 10^(k-1))
+    //         = (d[k-1]         ) * 10^(k-1) + (p1 mod 10^(k-1))
+    //
+    //      M+ = p1                                             + p2 * 2^e
+    //         = d[k-1] * 10^(k-1) + (p1 mod 10^(k-1))          + p2 * 2^e
+    //         = d[k-1] * 10^(k-1) + ((p1 mod 10^(k-1)) * 2^-e + p2) * 2^e
+    //         = d[k-1] * 10^(k-1) + (                         rest) * 2^e
+    //
+    // Now generate the digits d[n] of p1 from left to right (n = k-1,...,0)
+    //
+    //      p1 = d[k-1]...d[n] * 10^n + d[n-1]...d[0]
+    //
+    // but stop as soon as
+    //
+    //      rest * 2^e = (d[n-1]...d[0] * 2^-e + p2) * 2^e <= delta * 2^e
+
+    int n = k;
+    while (n > 0)
+    {
+        // Invariants:
+        //      M+ = buffer * 10^n + (p1 + p2 * 2^e)    (buffer = 0 for n = k)
+        //      pow10 = 10^(n-1) <= p1 < 10^n
+        //
+        const std::uint32_t d = p1 / pow10;  // d = p1 div 10^(n-1)
+        const std::uint32_t r = p1 % pow10;  // r = p1 mod 10^(n-1)
+        //
+        //      M+ = buffer * 10^n + (d * 10^(n-1) + r) + p2 * 2^e
+        //         = (buffer * 10 + d) * 10^(n-1) + (r + p2 * 2^e)
+        //
+        JSON_ASSERT(d <= 9);
+        buffer[length++] = static_cast<char>('0' + d); // buffer := buffer * 10 + d
+        //
+        //      M+ = buffer * 10^(n-1) + (r + p2 * 2^e)
+        //
+        p1 = r;
+        n--;
+        //
+        //      M+ = buffer * 10^n + (p1 + p2 * 2^e)
+        //      pow10 = 10^n
+        //
+
+        // Now check if enough digits have been generated.
+        // Compute
+        //
+        //      p1 + p2 * 2^e = (p1 * 2^-e + p2) * 2^e = rest * 2^e
+        //
+        // Note:
+        // Since rest and delta share the same exponent e, it suffices to
+        // compare the significands.
+        const std::uint64_t rest = (std::uint64_t{p1} << -one.e) + p2;
+        if (rest <= delta)
+        {
+            // V = buffer * 10^n, with M- <= V <= M+.
+
+            decimal_exponent += n;
+
+            // We may now just stop. But instead look if the buffer could be
+            // decremented to bring V closer to w.
+            //
+            // pow10 = 10^n is now 1 ulp in the decimal representation V.
+            // The rounding procedure works with diyfp's with an implicit
+            // exponent of e.
+            //
+            //      10^n = (10^n * 2^-e) * 2^e = ulp * 2^e
+            //
+            const std::uint64_t ten_n = std::uint64_t{pow10} << -one.e;
+            grisu2_round(buffer, length, dist, delta, rest, ten_n);
+
+            return;
+        }
+
+        pow10 /= 10;
+        //
+        //      pow10 = 10^(n-1) <= p1 < 10^n
+        // Invariants restored.
+    }
+
+    // 2)
+    //
+    // The digits of the integral part have been generated:
+    //
+    //      M+ = d[k-1]...d[1]d[0] + p2 * 2^e
+    //         = buffer            + p2 * 2^e
+    //
+    // Now generate the digits of the fractional part p2 * 2^e.
+    //
+    // Note:
+    // No decimal point is generated: the exponent is adjusted instead.
+    //
+    // p2 actually represents the fraction
+    //
+    //      p2 * 2^e
+    //          = p2 / 2^-e
+    //          = d[-1] / 10^1 + d[-2] / 10^2 + ...
+    //
+    // Now generate the digits d[-m] of p1 from left to right (m = 1,2,...)
+    //
+    //      p2 * 2^e = d[-1]d[-2]...d[-m] * 10^-m
+    //                      + 10^-m * (d[-m-1] / 10^1 + d[-m-2] / 10^2 + ...)
+    //
+    // using
+    //
+    //      10^m * p2 = ((10^m * p2) div 2^-e) * 2^-e + ((10^m * p2) mod 2^-e)
+    //                = (                   d) * 2^-e + (                   r)
+    //
+    // or
+    //      10^m * p2 * 2^e = d + r * 2^e
+    //
+    // i.e.
+    //
+    //      M+ = buffer + p2 * 2^e
+    //         = buffer + 10^-m * (d + r * 2^e)
+    //         = (buffer * 10^m + d) * 10^-m + 10^-m * r * 2^e
+    //
+    // and stop as soon as 10^-m * r * 2^e <= delta * 2^e
+
+    JSON_ASSERT(p2 > delta);
+
+    int m = 0;
+    for (;;)
+    {
+        // Invariant:
+        //      M+ = buffer * 10^-m + 10^-m * (d[-m-1] / 10 + d[-m-2] / 10^2 + ...) * 2^e
+        //         = buffer * 10^-m + 10^-m * (p2                                 ) * 2^e
+        //         = buffer * 10^-m + 10^-m * (1/10 * (10 * p2)                   ) * 2^e
+        //         = buffer * 10^-m + 10^-m * (1/10 * ((10*p2 div 2^-e) * 2^-e + (10*p2 mod 2^-e)) * 2^e
+        //
+        JSON_ASSERT(p2 <= (std::numeric_limits<std::uint64_t>::max)() / 10);
+        p2 *= 10;
+        const std::uint64_t d = p2 >> -one.e;     // d = (10 * p2) div 2^-e
+        const std::uint64_t r = p2 & (one.f - 1); // r = (10 * p2) mod 2^-e
+        //
+        //      M+ = buffer * 10^-m + 10^-m * (1/10 * (d * 2^-e + r) * 2^e
+        //         = buffer * 10^-m + 10^-m * (1/10 * (d + r * 2^e))
+        //         = (buffer * 10 + d) * 10^(-m-1) + 10^(-m-1) * r * 2^e
+        //
+        JSON_ASSERT(d <= 9);
+        buffer[length++] = static_cast<char>('0' + d); // buffer := buffer * 10 + d
+        //
+        //      M+ = buffer * 10^(-m-1) + 10^(-m-1) * r * 2^e
+        //
+        p2 = r;
+        m++;
+        //
+        //      M+ = buffer * 10^-m + 10^-m * p2 * 2^e
+        // Invariant restored.
+
+        // Check if enough digits have been generated.
+        //
+        //      10^-m * p2 * 2^e <= delta * 2^e
+        //              p2 * 2^e <= 10^m * delta * 2^e
+        //                    p2 <= 10^m * delta
+        delta *= 10;
+        dist  *= 10;
+        if (p2 <= delta)
+        {
+            break;
+        }
+    }
+
+    // V = buffer * 10^-m, with M- <= V <= M+.
+
+    decimal_exponent -= m;
+
+    // 1 ulp in the decimal representation is now 10^-m.
+    // Since delta and dist are now scaled by 10^m, we need to do the
+    // same with ulp in order to keep the units in sync.
+    //
+    //      10^m * 10^-m = 1 = 2^-e * 2^e = ten_m * 2^e
+    //
+    const std::uint64_t ten_m = one.f;
+    grisu2_round(buffer, length, dist, delta, p2, ten_m);
+
+    // By construction this algorithm generates the shortest possible decimal
+    // number (Loitsch, Theorem 6.2) which rounds back to w.
+    // For an input number of precision p, at least
+    //
+    //      N = 1 + ceil(p * log_10(2))
+    //
+    // decimal digits are sufficient to identify all binary floating-point
+    // numbers (Matula, "In-and-Out conversions").
+    // This implies that the algorithm does not produce more than N decimal
+    // digits.
+    //
+    //      N = 17 for p = 53 (IEEE double precision)
+    //      N = 9  for p = 24 (IEEE single precision)
+}
+
+/*!
+v = buf * 10^decimal_exponent
+len is the length of the buffer (number of decimal digits)
+The buffer must be large enough, i.e. >= max_digits10.
+*/
+JSON_HEDLEY_NON_NULL(1)
+inline void grisu2(char* buf, int& len, int& decimal_exponent,
+                   diyfp m_minus, diyfp v, diyfp m_plus)
+{
+    JSON_ASSERT(m_plus.e == m_minus.e);
+    JSON_ASSERT(m_plus.e == v.e);
+
+    //  --------(-----------------------+-----------------------)--------    (A)
+    //          m-                      v                       m+
+    //
+    //  --------------------(-----------+-----------------------)--------    (B)
+    //                      m-          v                       m+
+    //
+    // First scale v (and m- and m+) such that the exponent is in the range
+    // [alpha, gamma].
+
+    const cached_power cached = get_cached_power_for_binary_exponent(m_plus.e);
+
+    const diyfp c_minus_k(cached.f, cached.e); // = c ~= 10^-k
+
+    // The exponent of the products is = v.e + c_minus_k.e + q and is in the range [alpha,gamma]
+    const diyfp w       = diyfp::mul(v,       c_minus_k);
+    const diyfp w_minus = diyfp::mul(m_minus, c_minus_k);
+    const diyfp w_plus  = diyfp::mul(m_plus,  c_minus_k);
+
+    //  ----(---+---)---------------(---+---)---------------(---+---)----
+    //          w-                      w                       w+
+    //          = c*m-                  = c*v                   = c*m+
+    //
+    // diyfp::mul rounds its result and c_minus_k is approximated too. w, w- and
+    // w+ are now off by a small amount.
+    // In fact:
+    //
+    //      w - v * 10^k < 1 ulp
+    //
+    // To account for this inaccuracy, add resp. subtract 1 ulp.
+    //
+    //  --------+---[---------------(---+---)---------------]---+--------
+    //          w-  M-                  w                   M+  w+
+    //
+    // Now any number in [M-, M+] (bounds included) will round to w when input,
+    // regardless of how the input rounding algorithm breaks ties.
+    //
+    // And digit_gen generates the shortest possible such number in [M-, M+].
+    // Note that this does not mean that Grisu2 always generates the shortest
+    // possible number in the interval (m-, m+).
+    const diyfp M_minus(w_minus.f + 1, w_minus.e);
+    const diyfp M_plus (w_plus.f  - 1, w_plus.e );
+
+    decimal_exponent = -cached.k; // = -(-k) = k
+
+    grisu2_digit_gen(buf, len, decimal_exponent, M_minus, w, M_plus);
+}
+
+/*!
+v = buf * 10^decimal_exponent
+len is the length of the buffer (number of decimal digits)
+The buffer must be large enough, i.e. >= max_digits10.
+*/
+template<typename FloatType>
+JSON_HEDLEY_NON_NULL(1)
+void grisu2(char* buf, int& len, int& decimal_exponent, FloatType value)
+{
+    static_assert(diyfp::kPrecision >= std::numeric_limits<FloatType>::digits + 3,
+                  "internal error: not enough precision");
+
+    JSON_ASSERT(std::isfinite(value));
+    JSON_ASSERT(value > 0);
+
+    // If the neighbors (and boundaries) of 'value' are always computed for double-precision
+    // numbers, all float's can be recovered using strtod (and strtof). However, the resulting
+    // decimal representations are not exactly "short".
+    //
+    // The documentation for 'std::to_chars' (https://en.cppreference.com/w/cpp/utility/to_chars)
+    // says "value is converted to a string as if by std::sprintf in the default ("C") locale"
+    // and since sprintf promotes floats to doubles, I think this is exactly what 'std::to_chars'
+    // does.
+    // On the other hand, the documentation for 'std::to_chars' requires that "parsing the
+    // representation using the corresponding std::from_chars function recovers value exactly". That
+    // indicates that single precision floating-point numbers should be recovered using
+    // 'std::strtof'.
+    //
+    // NB: If the neighbors are computed for single-precision numbers, there is a single float
+    //     (7.0385307e-26f) which can't be recovered using strtod. The resulting double precision
+    //     value is off by 1 ulp.
+#if 0
+    const boundaries w = compute_boundaries(static_cast<double>(value));
+#else
+    const boundaries w = compute_boundaries(value);
+#endif
+
+    grisu2(buf, len, decimal_exponent, w.minus, w.w, w.plus);
+}
+
+/*!
+@brief appends a decimal representation of e to buf
+@return a pointer to the element following the exponent.
+@pre -1000 < e < 1000
+*/
+JSON_HEDLEY_NON_NULL(1)
+JSON_HEDLEY_RETURNS_NON_NULL
+inline char* append_exponent(char* buf, int e)
+{
+    JSON_ASSERT(e > -1000);
+    JSON_ASSERT(e <  1000);
+
+    if (e < 0)
+    {
+        e = -e;
+        *buf++ = '-';
+    }
+    else
+    {
+        *buf++ = '+';
+    }
+
+    auto k = static_cast<std::uint32_t>(e);
+    if (k < 10)
+    {
+        // Always print at least two digits in the exponent.
+        // This is for compatibility with printf("%g").
+        *buf++ = '0';
+        *buf++ = static_cast<char>('0' + k);
+    }
+    else if (k < 100)
+    {
+        *buf++ = static_cast<char>('0' + k / 10);
+        k %= 10;
+        *buf++ = static_cast<char>('0' + k);
+    }
+    else
+    {
+        *buf++ = static_cast<char>('0' + k / 100);
+        k %= 100;
+        *buf++ = static_cast<char>('0' + k / 10);
+        k %= 10;
+        *buf++ = static_cast<char>('0' + k);
+    }
+
+    return buf;
+}
+
+/*!
+@brief prettify v = buf * 10^decimal_exponent
+
+If v is in the range [10^min_exp, 10^max_exp) it will be printed in fixed-point
+notation. Otherwise it will be printed in exponential notation.
+
+@pre min_exp < 0
+@pre max_exp > 0
+*/
+JSON_HEDLEY_NON_NULL(1)
+JSON_HEDLEY_RETURNS_NON_NULL
+inline char* format_buffer(char* buf, int len, int decimal_exponent,
+                           int min_exp, int max_exp)
+{
+    JSON_ASSERT(min_exp < 0);
+    JSON_ASSERT(max_exp > 0);
+
+    const int k = len;
+    const int n = len + decimal_exponent;
+
+    // v = buf * 10^(n-k)
+    // k is the length of the buffer (number of decimal digits)
+    // n is the position of the decimal point relative to the start of the buffer.
+
+    if (k <= n && n <= max_exp)
+    {
+        // digits[000]
+        // len <= max_exp + 2
+
+        std::memset(buf + k, '0', static_cast<size_t>(n) - static_cast<size_t>(k));
+        // Make it look like a floating-point number (#362, #378)
+        buf[n + 0] = '.';
+        buf[n + 1] = '0';
+        return buf + (static_cast<size_t>(n) + 2);
+    }
+
+    if (0 < n && n <= max_exp)
+    {
+        // dig.its
+        // len <= max_digits10 + 1
+
+        JSON_ASSERT(k > n);
+
+        std::memmove(buf + (static_cast<size_t>(n) + 1), buf + n, static_cast<size_t>(k) - static_cast<size_t>(n));
+        buf[n] = '.';
+        return buf + (static_cast<size_t>(k) + 1U);
+    }
+
+    if (min_exp < n && n <= 0)
+    {
+        // 0.[000]digits
+        // len <= 2 + (-min_exp - 1) + max_digits10
+
+        std::memmove(buf + (2 + static_cast<size_t>(-n)), buf, static_cast<size_t>(k));
+        buf[0] = '0';
+        buf[1] = '.';
+        std::memset(buf + 2, '0', static_cast<size_t>(-n));
+        return buf + (2U + static_cast<size_t>(-n) + static_cast<size_t>(k));
+    }
+
+    if (k == 1)
+    {
+        // dE+123
+        // len <= 1 + 5
+
+        buf += 1;
+    }
+    else
+    {
+        // d.igitsE+123
+        // len <= max_digits10 + 1 + 5
+
+        std::memmove(buf + 2, buf + 1, static_cast<size_t>(k) - 1);
+        buf[1] = '.';
+        buf += 1 + static_cast<size_t>(k);
+    }
+
+    *buf++ = 'e';
+    return append_exponent(buf, n - 1);
+}
+
+}  // namespace dtoa_impl
+
+/*!
+@brief generates a decimal representation of the floating-point number value in [first, last).
+
+The format of the resulting decimal representation is similar to printf's %g
+format. Returns an iterator pointing past-the-end of the decimal representation.
+
+@note The input number must be finite, i.e. NaN's and Inf's are not supported.
+@note The buffer must be large enough.
+@note The result is NOT null-terminated.
+*/
+template<typename FloatType>
+JSON_HEDLEY_NON_NULL(1, 2)
+JSON_HEDLEY_RETURNS_NON_NULL
+char* to_chars(char* first, const char* last, FloatType value)
+{
+    static_cast<void>(last); // maybe unused - fix warning
+    JSON_ASSERT(std::isfinite(value));
+
+    // Use signbit(value) instead of (value < 0) since signbit works for -0.
+    if (std::signbit(value))
+    {
+        value = -value;
+        *first++ = '-';
+    }
+
+#ifdef __GNUC__
+#pragma GCC diagnostic push
+#pragma GCC diagnostic ignored "-Wfloat-equal"
+#endif
+    if (value == 0) // +-0
+    {
+        *first++ = '0';
+        // Make it look like a floating-point number (#362, #378)
+        *first++ = '.';
+        *first++ = '0';
+        return first;
+    }
+#ifdef __GNUC__
+#pragma GCC diagnostic pop
+#endif
+
+    JSON_ASSERT(last - first >= std::numeric_limits<FloatType>::max_digits10);
+
+    // Compute v = buffer * 10^decimal_exponent.
+    // The decimal digits are stored in the buffer, which needs to be interpreted
+    // as an unsigned decimal integer.
+    // len is the length of the buffer, i.e. the number of decimal digits.
+    int len = 0;
+    int decimal_exponent = 0;
+    dtoa_impl::grisu2(first, len, decimal_exponent, value);
+
+    JSON_ASSERT(len <= std::numeric_limits<FloatType>::max_digits10);
+
+    // Format the buffer like printf("%.*g", prec, value)
+    constexpr int kMinExp = -4;
+    // Use digits10 here to increase compatibility with version 2.
+    constexpr int kMaxExp = std::numeric_limits<FloatType>::digits10;
+
+    JSON_ASSERT(last - first >= kMaxExp + 2);
+    JSON_ASSERT(last - first >= 2 + (-kMinExp - 1) + std::numeric_limits<FloatType>::max_digits10);
+    JSON_ASSERT(last - first >= std::numeric_limits<FloatType>::max_digits10 + 6);
+
+    return dtoa_impl::format_buffer(first, len, decimal_exponent, kMinExp, kMaxExp);
+}
+
+}  // namespace detail
+NLOHMANN_JSON_NAMESPACE_END
+
+// #include <nlohmann/detail/exceptions.hpp>
+
+// #include <nlohmann/detail/macro_scope.hpp>
+
+// #include <nlohmann/detail/meta/cpp_future.hpp>
+
+// #include <nlohmann/detail/output/binary_writer.hpp>
+
+// #include <nlohmann/detail/output/output_adapters.hpp>
+
+// #include <nlohmann/detail/string_concat.hpp>
+
+// #include <nlohmann/detail/value_t.hpp>
+
+
+NLOHMANN_JSON_NAMESPACE_BEGIN
+namespace detail
+{
+
+///////////////////
+// serialization //
+///////////////////
+
+/// how to treat decoding errors
+enum class error_handler_t
+{
+    strict,  ///< throw a type_error exception in case of invalid UTF-8
+    replace, ///< replace invalid UTF-8 sequences with U+FFFD
+    ignore   ///< ignore invalid UTF-8 sequences
+};
+
+template<typename BasicJsonType>
+class serializer
+{
+    using string_t = typename BasicJsonType::string_t;
+    using number_float_t = typename BasicJsonType::number_float_t;
+    using number_integer_t = typename BasicJsonType::number_integer_t;
+    using number_unsigned_t = typename BasicJsonType::number_unsigned_t;
+    using binary_char_t = typename BasicJsonType::binary_t::value_type;
+    static constexpr std::uint8_t UTF8_ACCEPT = 0;
+    static constexpr std::uint8_t UTF8_REJECT = 1;
+
+  public:
+    /*!
+    @param[in] s  output stream to serialize to
+    @param[in] ichar  indentation character to use
+    @param[in] error_handler_  how to react on decoding errors
+    */
+    serializer(output_adapter_t<char> s, const char ichar,
+               error_handler_t error_handler_ = error_handler_t::strict)
+        : o(std::move(s))
+        , loc(std::localeconv())
+        , thousands_sep(loc->thousands_sep == nullptr ? '\0' : std::char_traits<char>::to_char_type(* (loc->thousands_sep)))
+        , decimal_point(loc->decimal_point == nullptr ? '\0' : std::char_traits<char>::to_char_type(* (loc->decimal_point)))
+        , indent_char(ichar)
+        , indent_string(512, indent_char)
+        , error_handler(error_handler_)
+    {}
+
+    // delete because of pointer members
+    serializer(const serializer&) = delete;
+    serializer& operator=(const serializer&) = delete;
+    serializer(serializer&&) = delete;
+    serializer& operator=(serializer&&) = delete;
+    ~serializer() = default;
+
+    /*!
+    @brief internal implementation of the serialization function
+
+    This function is called by the public member function dump and organizes
+    the serialization internally. The indentation level is propagated as
+    additional parameter. In case of arrays and objects, the function is
+    called recursively.
+
+    - strings and object keys are escaped using `escape_string()`
+    - integer numbers are converted implicitly via `operator<<`
+    - floating-point numbers are converted to a string using `"%g"` format
+    - binary values are serialized as objects containing the subtype and the
+      byte array
+
+    @param[in] val               value to serialize
+    @param[in] pretty_print      whether the output shall be pretty-printed
+    @param[in] ensure_ascii If @a ensure_ascii is true, all non-ASCII characters
+    in the output are escaped with `\uXXXX` sequences, and the result consists
+    of ASCII characters only.
+    @param[in] indent_step       the indent level
+    @param[in] current_indent    the current indent level (only used internally)
+    */
+    void dump(const BasicJsonType& val,
+              const bool pretty_print,
+              const bool ensure_ascii,
+              const unsigned int indent_step,
+              const unsigned int current_indent = 0)
+    {
+        switch (val.m_type)
+        {
+            case value_t::object:
+            {
+                if (val.m_value.object->empty())
+                {
+                    o->write_characters("{}", 2);
+                    return;
+                }
+
+                if (pretty_print)
+                {
+                    o->write_characters("{\n", 2);
+
+                    // variable to hold indentation for recursive calls
+                    const auto new_indent = current_indent + indent_step;
+                    if (JSON_HEDLEY_UNLIKELY(indent_string.size() < new_indent))
+                    {
+                        indent_string.resize(indent_string.size() * 2, ' ');
+                    }
+
+                    // first n-1 elements
+                    auto i = val.m_value.object->cbegin();
+                    for (std::size_t cnt = 0; cnt < val.m_value.object->size() - 1; ++cnt, ++i)
+                    {
+                        o->write_characters(indent_string.c_str(), new_indent);
+                        o->write_character('\"');
+                        dump_escaped(i->first, ensure_ascii);
+                        o->write_characters("\": ", 3);
+                        dump(i->second, true, ensure_ascii, indent_step, new_indent);
+                        o->write_characters(",\n", 2);
+                    }
+
+                    // last element
+                    JSON_ASSERT(i != val.m_value.object->cend());
+                    JSON_ASSERT(std::next(i) == val.m_value.object->cend());
+                    o->write_characters(indent_string.c_str(), new_indent);
+                    o->write_character('\"');
+                    dump_escaped(i->first, ensure_ascii);
+                    o->write_characters("\": ", 3);
+                    dump(i->second, true, ensure_ascii, indent_step, new_indent);
+
+                    o->write_character('\n');
+                    o->write_characters(indent_string.c_str(), current_indent);
+                    o->write_character('}');
+                }
+                else
+                {
+                    o->write_character('{');
+
+                    // first n-1 elements
+                    auto i = val.m_value.object->cbegin();
+                    for (std::size_t cnt = 0; cnt < val.m_value.object->size() - 1; ++cnt, ++i)
+                    {
+                        o->write_character('\"');
+                        dump_escaped(i->first, ensure_ascii);
+                        o->write_characters("\":", 2);
+                        dump(i->second, false, ensure_ascii, indent_step, current_indent);
+                        o->write_character(',');
+                    }
+
+                    // last element
+                    JSON_ASSERT(i != val.m_value.object->cend());
+                    JSON_ASSERT(std::next(i) == val.m_value.object->cend());
+                    o->write_character('\"');
+                    dump_escaped(i->first, ensure_ascii);
+                    o->write_characters("\":", 2);
+                    dump(i->second, false, ensure_ascii, indent_step, current_indent);
+
+                    o->write_character('}');
+                }
+
+                return;
+            }
+
+            case value_t::array:
+            {
+                if (val.m_value.array->empty())
+                {
+                    o->write_characters("[]", 2);
+                    return;
+                }
+
+                if (pretty_print)
+                {
+                    o->write_characters("[\n", 2);
+
+                    // variable to hold indentation for recursive calls
+                    const auto new_indent = current_indent + indent_step;
+                    if (JSON_HEDLEY_UNLIKELY(indent_string.size() < new_indent))
+                    {
+                        indent_string.resize(indent_string.size() * 2, ' ');
+                    }
+
+                    // first n-1 elements
+                    for (auto i = val.m_value.array->cbegin();
+                            i != val.m_value.array->cend() - 1; ++i)
+                    {
+                        o->write_characters(indent_string.c_str(), new_indent);
+                        dump(*i, true, ensure_ascii, indent_step, new_indent);
+                        o->write_characters(",\n", 2);
+                    }
+
+                    // last element
+                    JSON_ASSERT(!val.m_value.array->empty());
+                    o->write_characters(indent_string.c_str(), new_indent);
+                    dump(val.m_value.array->back(), true, ensure_ascii, indent_step, new_indent);
+
+                    o->write_character('\n');
+                    o->write_characters(indent_string.c_str(), current_indent);
+                    o->write_character(']');
+                }
+                else
+                {
+                    o->write_character('[');
+
+                    // first n-1 elements
+                    for (auto i = val.m_value.array->cbegin();
+                            i != val.m_value.array->cend() - 1; ++i)
+                    {
+                        dump(*i, false, ensure_ascii, indent_step, current_indent);
+                        o->write_character(',');
+                    }
+
+                    // last element
+                    JSON_ASSERT(!val.m_value.array->empty());
+                    dump(val.m_value.array->back(), false, ensure_ascii, indent_step, current_indent);
+
+                    o->write_character(']');
+                }
+
+                return;
+            }
+
+            case value_t::string:
+            {
+                o->write_character('\"');
+                dump_escaped(*val.m_value.string, ensure_ascii);
+                o->write_character('\"');
+                return;
+            }
+
+            case value_t::binary:
+            {
+                if (pretty_print)
+                {
+                    o->write_characters("{\n", 2);
+
+                    // variable to hold indentation for recursive calls
+                    const auto new_indent = current_indent + indent_step;
+                    if (JSON_HEDLEY_UNLIKELY(indent_string.size() < new_indent))
+                    {
+                        indent_string.resize(indent_string.size() * 2, ' ');
+                    }
+
+                    o->write_characters(indent_string.c_str(), new_indent);
+
+                    o->write_characters("\"bytes\": [", 10);
+
+                    if (!val.m_value.binary->empty())
+                    {
+                        for (auto i = val.m_value.binary->cbegin();
+                                i != val.m_value.binary->cend() - 1; ++i)
+                        {
+                            dump_integer(*i);
+                            o->write_characters(", ", 2);
+                        }
+                        dump_integer(val.m_value.binary->back());
+                    }
+
+                    o->write_characters("],\n", 3);
+                    o->write_characters(indent_string.c_str(), new_indent);
+
+                    o->write_characters("\"subtype\": ", 11);
+                    if (val.m_value.binary->has_subtype())
+                    {
+                        dump_integer(val.m_value.binary->subtype());
+                    }
+                    else
+                    {
+                        o->write_characters("null", 4);
+                    }
+                    o->write_character('\n');
+                    o->write_characters(indent_string.c_str(), current_indent);
+                    o->write_character('}');
+                }
+                else
+                {
+                    o->write_characters("{\"bytes\":[", 10);
+
+                    if (!val.m_value.binary->empty())
+                    {
+                        for (auto i = val.m_value.binary->cbegin();
+                                i != val.m_value.binary->cend() - 1; ++i)
+                        {
+                            dump_integer(*i);
+                            o->write_character(',');
+                        }
+                        dump_integer(val.m_value.binary->back());
+                    }
+
+                    o->write_characters("],\"subtype\":", 12);
+                    if (val.m_value.binary->has_subtype())
+                    {
+                        dump_integer(val.m_value.binary->subtype());
+                        o->write_character('}');
+                    }
+                    else
+                    {
+                        o->write_characters("null}", 5);
+                    }
+                }
+                return;
+            }
+
+            case value_t::boolean:
+            {
+                if (val.m_value.boolean)
+                {
+                    o->write_characters("true", 4);
+                }
+                else
+                {
+                    o->write_characters("false", 5);
+                }
+                return;
+            }
+
+            case value_t::number_integer:
+            {
+                dump_integer(val.m_value.number_integer);
+                return;
+            }
+
+            case value_t::number_unsigned:
+            {
+                dump_integer(val.m_value.number_unsigned);
+                return;
+            }
+
+            case value_t::number_float:
+            {
+                dump_float(val.m_value.number_float);
+                return;
+            }
+
+            case value_t::discarded:
+            {
+                o->write_characters("<discarded>", 11);
+                return;
+            }
+
+            case value_t::null:
+            {
+                o->write_characters("null", 4);
+                return;
+            }
+
+            default:            // LCOV_EXCL_LINE
+                JSON_ASSERT(false); // NOLINT(cert-dcl03-c,hicpp-static-assert,misc-static-assert) LCOV_EXCL_LINE
+        }
+    }
+
+  JSON_PRIVATE_UNLESS_TESTED:
+    /*!
+    @brief dump escaped string
+
+    Escape a string by replacing certain special characters by a sequence of an
+    escape character (backslash) and another character and other control
+    characters by a sequence of "\u" followed by a four-digit hex
+    representation. The escaped string is written to output stream @a o.
+
+    @param[in] s  the string to escape
+    @param[in] ensure_ascii  whether to escape non-ASCII characters with
+                             \uXXXX sequences
+
+    @complexity Linear in the length of string @a s.
+    */
+    void dump_escaped(const string_t& s, const bool ensure_ascii)
+    {
+        std::uint32_t codepoint{};
+        std::uint8_t state = UTF8_ACCEPT;
+        std::size_t bytes = 0;  // number of bytes written to string_buffer
+
+        // number of bytes written at the point of the last valid byte
+        std::size_t bytes_after_last_accept = 0;
+        std::size_t undumped_chars = 0;
+
+        for (std::size_t i = 0; i < s.size(); ++i)
+        {
+            const auto byte = static_cast<std::uint8_t>(s[i]);
+
+            switch (decode(state, codepoint, byte))
+            {
+                case UTF8_ACCEPT:  // decode found a new code point
+                {
+                    switch (codepoint)
+                    {
+                        case 0x08: // backspace
+                        {
+                            string_buffer[bytes++] = '\\';
+                            string_buffer[bytes++] = 'b';
+                            break;
+                        }
+
+                        case 0x09: // horizontal tab
+                        {
+                            string_buffer[bytes++] = '\\';
+                            string_buffer[bytes++] = 't';
+                            break;
+                        }
+
+                        case 0x0A: // newline
+                        {
+                            string_buffer[bytes++] = '\\';
+                            string_buffer[bytes++] = 'n';
+                            break;
+                        }
+
+                        case 0x0C: // formfeed
+                        {
+                            string_buffer[bytes++] = '\\';
+                            string_buffer[bytes++] = 'f';
+                            break;
+                        }
+
+                        case 0x0D: // carriage return
+                        {
+                            string_buffer[bytes++] = '\\';
+                            string_buffer[bytes++] = 'r';
+                            break;
+                        }
+
+                        case 0x22: // quotation mark
+                        {
+                            string_buffer[bytes++] = '\\';
+                            string_buffer[bytes++] = '\"';
+                            break;
+                        }
+
+                        case 0x5C: // reverse solidus
+                        {
+                            string_buffer[bytes++] = '\\';
+                            string_buffer[bytes++] = '\\';
+                            break;
+                        }
+
+                        default:
+                        {
+                            // escape control characters (0x00..0x1F) or, if
+                            // ensure_ascii parameter is used, non-ASCII characters
+                            if ((codepoint <= 0x1F) || (ensure_ascii && (codepoint >= 0x7F)))
+                            {
+                                if (codepoint <= 0xFFFF)
+                                {
+                                    // NOLINTNEXTLINE(cppcoreguidelines-pro-type-vararg,hicpp-vararg)
+                                    static_cast<void>((std::snprintf)(string_buffer.data() + bytes, 7, "\\u%04x",
+                                                                      static_cast<std::uint16_t>(codepoint)));
+                                    bytes += 6;
+                                }
+                                else
+                                {
+                                    // NOLINTNEXTLINE(cppcoreguidelines-pro-type-vararg,hicpp-vararg)
+                                    static_cast<void>((std::snprintf)(string_buffer.data() + bytes, 13, "\\u%04x\\u%04x",
+                                                                      static_cast<std::uint16_t>(0xD7C0u + (codepoint >> 10u)),
+                                                                      static_cast<std::uint16_t>(0xDC00u + (codepoint & 0x3FFu))));
+                                    bytes += 12;
+                                }
+                            }
+                            else
+                            {
+                                // copy byte to buffer (all previous bytes
+                                // been copied have in default case above)
+                                string_buffer[bytes++] = s[i];
+                            }
+                            break;
+                        }
+                    }
+
+                    // write buffer and reset index; there must be 13 bytes
+                    // left, as this is the maximal number of bytes to be
+                    // written ("\uxxxx\uxxxx\0") for one code point
+                    if (string_buffer.size() - bytes < 13)
+                    {
+                        o->write_characters(string_buffer.data(), bytes);
+                        bytes = 0;
+                    }
+
+                    // remember the byte position of this accept
+                    bytes_after_last_accept = bytes;
+                    undumped_chars = 0;
+                    break;
+                }
+
+                case UTF8_REJECT:  // decode found invalid UTF-8 byte
+                {
+                    switch (error_handler)
+                    {
+                        case error_handler_t::strict:
+                        {
+                            JSON_THROW(type_error::create(316, concat("invalid UTF-8 byte at index ", std::to_string(i), ": 0x", hex_bytes(byte | 0)), nullptr));
+                        }
+
+                        case error_handler_t::ignore:
+                        case error_handler_t::replace:
+                        {
+                            // in case we saw this character the first time, we
+                            // would like to read it again, because the byte
+                            // may be OK for itself, but just not OK for the
+                            // previous sequence
+                            if (undumped_chars > 0)
+                            {
+                                --i;
+                            }
+
+                            // reset length buffer to the last accepted index;
+                            // thus removing/ignoring the invalid characters
+                            bytes = bytes_after_last_accept;
+
+                            if (error_handler == error_handler_t::replace)
+                            {
+                                // add a replacement character
+                                if (ensure_ascii)
+                                {
+                                    string_buffer[bytes++] = '\\';
+                                    string_buffer[bytes++] = 'u';
+                                    string_buffer[bytes++] = 'f';
+                                    string_buffer[bytes++] = 'f';
+                                    string_buffer[bytes++] = 'f';
+                                    string_buffer[bytes++] = 'd';
+                                }
+                                else
+                                {
+                                    string_buffer[bytes++] = detail::binary_writer<BasicJsonType, char>::to_char_type('\xEF');
+                                    string_buffer[bytes++] = detail::binary_writer<BasicJsonType, char>::to_char_type('\xBF');
+                                    string_buffer[bytes++] = detail::binary_writer<BasicJsonType, char>::to_char_type('\xBD');
+                                }
+
+                                // write buffer and reset index; there must be 13 bytes
+                                // left, as this is the maximal number of bytes to be
+                                // written ("\uxxxx\uxxxx\0") for one code point
+                                if (string_buffer.size() - bytes < 13)
+                                {
+                                    o->write_characters(string_buffer.data(), bytes);
+                                    bytes = 0;
+                                }
+
+                                bytes_after_last_accept = bytes;
+                            }
+
+                            undumped_chars = 0;
+
+                            // continue processing the string
+                            state = UTF8_ACCEPT;
+                            break;
+                        }
+
+                        default:            // LCOV_EXCL_LINE
+                            JSON_ASSERT(false); // NOLINT(cert-dcl03-c,hicpp-static-assert,misc-static-assert) LCOV_EXCL_LINE
+                    }
+                    break;
+                }
+
+                default:  // decode found yet incomplete multi-byte code point
+                {
+                    if (!ensure_ascii)
+                    {
+                        // code point will not be escaped - copy byte to buffer
+                        string_buffer[bytes++] = s[i];
+                    }
+                    ++undumped_chars;
+                    break;
+                }
+            }
+        }
+
+        // we finished processing the string
+        if (JSON_HEDLEY_LIKELY(state == UTF8_ACCEPT))
+        {
+            // write buffer
+            if (bytes > 0)
+            {
+                o->write_characters(string_buffer.data(), bytes);
+            }
+        }
+        else
+        {
+            // we finish reading, but do not accept: string was incomplete
+            switch (error_handler)
+            {
+                case error_handler_t::strict:
+                {
+                    JSON_THROW(type_error::create(316, concat("incomplete UTF-8 string; last byte: 0x", hex_bytes(static_cast<std::uint8_t>(s.back() | 0))), nullptr));
+                }
+
+                case error_handler_t::ignore:
+                {
+                    // write all accepted bytes
+                    o->write_characters(string_buffer.data(), bytes_after_last_accept);
+                    break;
+                }
+
+                case error_handler_t::replace:
+                {
+                    // write all accepted bytes
+                    o->write_characters(string_buffer.data(), bytes_after_last_accept);
+                    // add a replacement character
+                    if (ensure_ascii)
+                    {
+                        o->write_characters("\\ufffd", 6);
+                    }
+                    else
+                    {
+                        o->write_characters("\xEF\xBF\xBD", 3);
+                    }
+                    break;
+                }
+
+                default:            // LCOV_EXCL_LINE
+                    JSON_ASSERT(false); // NOLINT(cert-dcl03-c,hicpp-static-assert,misc-static-assert) LCOV_EXCL_LINE
+            }
+        }
+    }
+
+  private:
+    /*!
+    @brief count digits
+
+    Count the number of decimal (base 10) digits for an input unsigned integer.
+
+    @param[in] x  unsigned integer number to count its digits
+    @return    number of decimal digits
+    */
+    inline unsigned int count_digits(number_unsigned_t x) noexcept
+    {
+        unsigned int n_digits = 1;
+        for (;;)
+        {
+            if (x < 10)
+            {
+                return n_digits;
+            }
+            if (x < 100)
+            {
+                return n_digits + 1;
+            }
+            if (x < 1000)
+            {
+                return n_digits + 2;
+            }
+            if (x < 10000)
+            {
+                return n_digits + 3;
+            }
+            x = x / 10000u;
+            n_digits += 4;
+        }
+    }
+
+    /*!
+     * @brief convert a byte to a uppercase hex representation
+     * @param[in] byte byte to represent
+     * @return representation ("00".."FF")
+     */
+    static std::string hex_bytes(std::uint8_t byte)
+    {
+        std::string result = "FF";
+        constexpr const char* nibble_to_hex = "0123456789ABCDEF";
+        result[0] = nibble_to_hex[byte / 16];
+        result[1] = nibble_to_hex[byte % 16];
+        return result;
+    }
+
+    // templates to avoid warnings about useless casts
+    template <typename NumberType, enable_if_t<std::is_signed<NumberType>::value, int> = 0>
+    bool is_negative_number(NumberType x)
+    {
+        return x < 0;
+    }
+
+    template < typename NumberType, enable_if_t <std::is_unsigned<NumberType>::value, int > = 0 >
+    bool is_negative_number(NumberType /*unused*/)
+    {
+        return false;
+    }
+
+    /*!
+    @brief dump an integer
+
+    Dump a given integer to output stream @a o. Works internally with
+    @a number_buffer.
+
+    @param[in] x  integer number (signed or unsigned) to dump
+    @tparam NumberType either @a number_integer_t or @a number_unsigned_t
+    */
+    template < typename NumberType, detail::enable_if_t <
+                   std::is_integral<NumberType>::value ||
+                   std::is_same<NumberType, number_unsigned_t>::value ||
+                   std::is_same<NumberType, number_integer_t>::value ||
+                   std::is_same<NumberType, binary_char_t>::value,
+                   int > = 0 >
+    void dump_integer(NumberType x)
+    {
+        static constexpr std::array<std::array<char, 2>, 100> digits_to_99
+        {
+            {
+                {{'0', '0'}}, {{'0', '1'}}, {{'0', '2'}}, {{'0', '3'}}, {{'0', '4'}}, {{'0', '5'}}, {{'0', '6'}}, {{'0', '7'}}, {{'0', '8'}}, {{'0', '9'}},
+                {{'1', '0'}}, {{'1', '1'}}, {{'1', '2'}}, {{'1', '3'}}, {{'1', '4'}}, {{'1', '5'}}, {{'1', '6'}}, {{'1', '7'}}, {{'1', '8'}}, {{'1', '9'}},
+                {{'2', '0'}}, {{'2', '1'}}, {{'2', '2'}}, {{'2', '3'}}, {{'2', '4'}}, {{'2', '5'}}, {{'2', '6'}}, {{'2', '7'}}, {{'2', '8'}}, {{'2', '9'}},
+                {{'3', '0'}}, {{'3', '1'}}, {{'3', '2'}}, {{'3', '3'}}, {{'3', '4'}}, {{'3', '5'}}, {{'3', '6'}}, {{'3', '7'}}, {{'3', '8'}}, {{'3', '9'}},
+                {{'4', '0'}}, {{'4', '1'}}, {{'4', '2'}}, {{'4', '3'}}, {{'4', '4'}}, {{'4', '5'}}, {{'4', '6'}}, {{'4', '7'}}, {{'4', '8'}}, {{'4', '9'}},
+                {{'5', '0'}}, {{'5', '1'}}, {{'5', '2'}}, {{'5', '3'}}, {{'5', '4'}}, {{'5', '5'}}, {{'5', '6'}}, {{'5', '7'}}, {{'5', '8'}}, {{'5', '9'}},
+                {{'6', '0'}}, {{'6', '1'}}, {{'6', '2'}}, {{'6', '3'}}, {{'6', '4'}}, {{'6', '5'}}, {{'6', '6'}}, {{'6', '7'}}, {{'6', '8'}}, {{'6', '9'}},
+                {{'7', '0'}}, {{'7', '1'}}, {{'7', '2'}}, {{'7', '3'}}, {{'7', '4'}}, {{'7', '5'}}, {{'7', '6'}}, {{'7', '7'}}, {{'7', '8'}}, {{'7', '9'}},
+                {{'8', '0'}}, {{'8', '1'}}, {{'8', '2'}}, {{'8', '3'}}, {{'8', '4'}}, {{'8', '5'}}, {{'8', '6'}}, {{'8', '7'}}, {{'8', '8'}}, {{'8', '9'}},
+                {{'9', '0'}}, {{'9', '1'}}, {{'9', '2'}}, {{'9', '3'}}, {{'9', '4'}}, {{'9', '5'}}, {{'9', '6'}}, {{'9', '7'}}, {{'9', '8'}}, {{'9', '9'}},
+            }
+        };
+
+        // special case for "0"
+        if (x == 0)
+        {
+            o->write_character('0');
+            return;
+        }
+
+        // use a pointer to fill the buffer
+        auto buffer_ptr = number_buffer.begin(); // NOLINT(llvm-qualified-auto,readability-qualified-auto,cppcoreguidelines-pro-type-vararg,hicpp-vararg)
+
+        number_unsigned_t abs_value;
+
+        unsigned int n_chars{};
+
+        if (is_negative_number(x))
+        {
+            *buffer_ptr = '-';
+            abs_value = remove_sign(static_cast<number_integer_t>(x));
+
+            // account one more byte for the minus sign
+            n_chars = 1 + count_digits(abs_value);
+        }
+        else
+        {
+            abs_value = static_cast<number_unsigned_t>(x);
+            n_chars = count_digits(abs_value);
+        }
+
+        // spare 1 byte for '\0'
+        JSON_ASSERT(n_chars < number_buffer.size() - 1);
+
+        // jump to the end to generate the string from backward,
+        // so we later avoid reversing the result
+        buffer_ptr += n_chars;
+
+        // Fast int2ascii implementation inspired by "Fastware" talk by Andrei Alexandrescu
+        // See: https://www.youtube.com/watch?v=o4-CwDo2zpg
+        while (abs_value >= 100)
+        {
+            const auto digits_index = static_cast<unsigned>((abs_value % 100));
+            abs_value /= 100;
+            *(--buffer_ptr) = digits_to_99[digits_index][1];
+            *(--buffer_ptr) = digits_to_99[digits_index][0];
+        }
+
+        if (abs_value >= 10)
+        {
+            const auto digits_index = static_cast<unsigned>(abs_value);
+            *(--buffer_ptr) = digits_to_99[digits_index][1];
+            *(--buffer_ptr) = digits_to_99[digits_index][0];
+        }
+        else
+        {
+            *(--buffer_ptr) = static_cast<char>('0' + abs_value);
+        }
+
+        o->write_characters(number_buffer.data(), n_chars);
+    }
+
+    /*!
+    @brief dump a floating-point number
+
+    Dump a given floating-point number to output stream @a o. Works internally
+    with @a number_buffer.
+
+    @param[in] x  floating-point number to dump
+    */
+    void dump_float(number_float_t x)
+    {
+        // NaN / inf
+        if (!std::isfinite(x))
+        {
+            o->write_characters("null", 4);
+            return;
+        }
+
+        // If number_float_t is an IEEE-754 single or double precision number,
+        // use the Grisu2 algorithm to produce short numbers which are
+        // guaranteed to round-trip, using strtof and strtod, resp.
+        //
+        // NB: The test below works if <long double> == <double>.
+        static constexpr bool is_ieee_single_or_double
+            = (std::numeric_limits<number_float_t>::is_iec559 && std::numeric_limits<number_float_t>::digits == 24 && std::numeric_limits<number_float_t>::max_exponent == 128) ||
+              (std::numeric_limits<number_float_t>::is_iec559 && std::numeric_limits<number_float_t>::digits == 53 && std::numeric_limits<number_float_t>::max_exponent == 1024);
+
+        dump_float(x, std::integral_constant<bool, is_ieee_single_or_double>());
+    }
+
+    void dump_float(number_float_t x, std::true_type /*is_ieee_single_or_double*/)
+    {
+        auto* begin = number_buffer.data();
+        auto* end = ::nlohmann::detail::to_chars(begin, begin + number_buffer.size(), x);
+
+        o->write_characters(begin, static_cast<size_t>(end - begin));
+    }
+
+    void dump_float(number_float_t x, std::false_type /*is_ieee_single_or_double*/)
+    {
+        // get number of digits for a float -> text -> float round-trip
+        static constexpr auto d = std::numeric_limits<number_float_t>::max_digits10;
+
+        // the actual conversion
+        // NOLINTNEXTLINE(cppcoreguidelines-pro-type-vararg,hicpp-vararg)
+        std::ptrdiff_t len = (std::snprintf)(number_buffer.data(), number_buffer.size(), "%.*g", d, x);
+
+        // negative value indicates an error
+        JSON_ASSERT(len > 0);
+        // check if buffer was large enough
+        JSON_ASSERT(static_cast<std::size_t>(len) < number_buffer.size());
+
+        // erase thousands separator
+        if (thousands_sep != '\0')
+        {
+            // NOLINTNEXTLINE(readability-qualified-auto,llvm-qualified-auto): std::remove returns an iterator, see https://github.com/nlohmann/json/issues/3081
+            const auto end = std::remove(number_buffer.begin(), number_buffer.begin() + len, thousands_sep);
+            std::fill(end, number_buffer.end(), '\0');
+            JSON_ASSERT((end - number_buffer.begin()) <= len);
+            len = (end - number_buffer.begin());
+        }
+
+        // convert decimal point to '.'
+        if (decimal_point != '\0' && decimal_point != '.')
+        {
+            // NOLINTNEXTLINE(readability-qualified-auto,llvm-qualified-auto): std::find returns an iterator, see https://github.com/nlohmann/json/issues/3081
+            const auto dec_pos = std::find(number_buffer.begin(), number_buffer.end(), decimal_point);
+            if (dec_pos != number_buffer.end())
+            {
+                *dec_pos = '.';
+            }
+        }
+
+        o->write_characters(number_buffer.data(), static_cast<std::size_t>(len));
+
+        // determine if we need to append ".0"
+        const bool value_is_int_like =
+            std::none_of(number_buffer.begin(), number_buffer.begin() + len + 1,
+                         [](char c)
+        {
+            return c == '.' || c == 'e';
+        });
+
+        if (value_is_int_like)
+        {
+            o->write_characters(".0", 2);
+        }
+    }
+
+    /*!
+    @brief check whether a string is UTF-8 encoded
+
+    The function checks each byte of a string whether it is UTF-8 encoded. The
+    result of the check is stored in the @a state parameter. The function must
+    be called initially with state 0 (accept). State 1 means the string must
+    be rejected, because the current byte is not allowed. If the string is
+    completely processed, but the state is non-zero, the string ended
+    prematurely; that is, the last byte indicated more bytes should have
+    followed.
+
+    @param[in,out] state  the state of the decoding
+    @param[in,out] codep  codepoint (valid only if resulting state is UTF8_ACCEPT)
+    @param[in] byte       next byte to decode
+    @return               new state
+
+    @note The function has been edited: a std::array is used.
+
+    @copyright Copyright (c) 2008-2009 Bjoern Hoehrmann <bjoern@hoehrmann.de>
+    @sa http://bjoern.hoehrmann.de/utf-8/decoder/dfa/
+    */
+    static std::uint8_t decode(std::uint8_t& state, std::uint32_t& codep, const std::uint8_t byte) noexcept
+    {
+        static const std::array<std::uint8_t, 400> utf8d =
+        {
+            {
+                0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, // 00..1F
+                0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, // 20..3F
+                0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, // 40..5F
+                0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, // 60..7F
+                1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, 9, // 80..9F
+                7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, 7, // A0..BF
+                8, 8, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, // C0..DF
+                0xA, 0x3, 0x3, 0x3, 0x3, 0x3, 0x3, 0x3, 0x3, 0x3, 0x3, 0x3, 0x3, 0x4, 0x3, 0x3, // E0..EF
+                0xB, 0x6, 0x6, 0x6, 0x5, 0x8, 0x8, 0x8, 0x8, 0x8, 0x8, 0x8, 0x8, 0x8, 0x8, 0x8, // F0..FF
+                0x0, 0x1, 0x2, 0x3, 0x5, 0x8, 0x7, 0x1, 0x1, 0x1, 0x4, 0x6, 0x1, 0x1, 0x1, 0x1, // s0..s0
+                1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 0, 1, 1, 1, 1, 1, 0, 1, 0, 1, 1, 1, 1, 1, 1, // s1..s2
+                1, 2, 1, 1, 1, 1, 1, 2, 1, 2, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 2, 1, 1, 1, 1, 1, 1, 1, 1, // s3..s4
+                1, 2, 1, 1, 1, 1, 1, 1, 1, 2, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 3, 1, 3, 1, 1, 1, 1, 1, 1, // s5..s6
+                1, 3, 1, 1, 1, 1, 1, 3, 1, 3, 1, 1, 1, 1, 1, 1, 1, 3, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1 // s7..s8
+            }
+        };
+
+        JSON_ASSERT(byte < utf8d.size());
+        const std::uint8_t type = utf8d[byte];
+
+        codep = (state != UTF8_ACCEPT)
+                ? (byte & 0x3fu) | (codep << 6u)
+                : (0xFFu >> type) & (byte);
+
+        std::size_t index = 256u + static_cast<size_t>(state) * 16u + static_cast<size_t>(type);
+        JSON_ASSERT(index < 400);
+        state = utf8d[index];
+        return state;
+    }
+
+    /*
+     * Overload to make the compiler happy while it is instantiating
+     * dump_integer for number_unsigned_t.
+     * Must never be called.
+     */
+    number_unsigned_t remove_sign(number_unsigned_t x)
+    {
+        JSON_ASSERT(false); // NOLINT(cert-dcl03-c,hicpp-static-assert,misc-static-assert) LCOV_EXCL_LINE
+        return x; // LCOV_EXCL_LINE
+    }
+
+    /*
+     * Helper function for dump_integer
+     *
+     * This function takes a negative signed integer and returns its absolute
+     * value as unsigned integer. The plus/minus shuffling is necessary as we can
+     * not directly remove the sign of an arbitrary signed integer as the
+     * absolute values of INT_MIN and INT_MAX are usually not the same. See
+     * #1708 for details.
+     */
+    inline number_unsigned_t remove_sign(number_integer_t x) noexcept
+    {
+        JSON_ASSERT(x < 0 && x < (std::numeric_limits<number_integer_t>::max)()); // NOLINT(misc-redundant-expression)
+        return static_cast<number_unsigned_t>(-(x + 1)) + 1;
+    }
+
+  private:
+    /// the output of the serializer
+    output_adapter_t<char> o = nullptr;
+
+    /// a (hopefully) large enough character buffer
+    std::array<char, 64> number_buffer{{}};
+
+    /// the locale
+    const std::lconv* loc = nullptr;
+    /// the locale's thousand separator character
+    const char thousands_sep = '\0';
+    /// the locale's decimal point character
+    const char decimal_point = '\0';
+
+    /// string buffer
+    std::array<char, 512> string_buffer{{}};
+
+    /// the indentation character
+    const char indent_char;
+    /// the indentation string
+    string_t indent_string;
+
+    /// error_handler how to react on decoding errors
+    const error_handler_t error_handler;
+};
+
+}  // namespace detail
+NLOHMANN_JSON_NAMESPACE_END
+
+// #include <nlohmann/detail/value_t.hpp>
+
+// #include <nlohmann/json_fwd.hpp>
+
+// #include <nlohmann/ordered_map.hpp>
+//     __ _____ _____ _____
+//  __|  |   __|     |   | |  JSON for Modern C++
+// |  |  |__   |  |  | | | |  version 3.11.2
+// |_____|_____|_____|_|___|  https://github.com/nlohmann/json
+//
+// SPDX-FileCopyrightText: 2013-2022 Niels Lohmann <https://nlohmann.me>
+// SPDX-License-Identifier: MIT
+
+
+
+#include <functional> // equal_to, less
+#include <initializer_list> // initializer_list
+#include <iterator> // input_iterator_tag, iterator_traits
+#include <memory> // allocator
+#include <stdexcept> // for out_of_range
+#include <type_traits> // enable_if, is_convertible
+#include <utility> // pair
+#include <vector> // vector
+
+// #include <nlohmann/detail/macro_scope.hpp>
+
+// #include <nlohmann/detail/meta/type_traits.hpp>
+
+
+NLOHMANN_JSON_NAMESPACE_BEGIN
+
+/// ordered_map: a minimal map-like container that preserves insertion order
+/// for use within nlohmann::basic_json<ordered_map>
+template <class Key, class T, class IgnoredLess = std::less<Key>,
+          class Allocator = std::allocator<std::pair<const Key, T>>>
+                  struct ordered_map : std::vector<std::pair<const Key, T>, Allocator>
+{
+    using key_type = Key;
+    using mapped_type = T;
+    using Container = std::vector<std::pair<const Key, T>, Allocator>;
+    using iterator = typename Container::iterator;
+    using const_iterator = typename Container::const_iterator;
+    using size_type = typename Container::size_type;
+    using value_type = typename Container::value_type;
+#ifdef JSON_HAS_CPP_14
+    using key_compare = std::equal_to<>;
+#else
+    using key_compare = std::equal_to<Key>;
+#endif
+
+    // Explicit constructors instead of `using Container::Container`
+    // otherwise older compilers choke on it (GCC <= 5.5, xcode <= 9.4)
+    ordered_map() noexcept(noexcept(Container())) : Container{} {}
+    explicit ordered_map(const Allocator& alloc) noexcept(noexcept(Container(alloc))) : Container{alloc} {}
+    template <class It>
+    ordered_map(It first, It last, const Allocator& alloc = Allocator())
+        : Container{first, last, alloc} {}
+    ordered_map(std::initializer_list<value_type> init, const Allocator& alloc = Allocator() )
+        : Container{init, alloc} {}
+
+    std::pair<iterator, bool> emplace(const key_type& key, T&& t)
+    {
+        for (auto it = this->begin(); it != this->end(); ++it)
+        {
+            if (m_compare(it->first, key))
+            {
+                return {it, false};
+            }
+        }
+        Container::emplace_back(key, std::forward<T>(t));
+        return {std::prev(this->end()), true};
+    }
+
+    template<class KeyType, detail::enable_if_t<
+                 detail::is_usable_as_key_type<key_compare, key_type, KeyType>::value, int> = 0>
+    std::pair<iterator, bool> emplace(KeyType && key, T && t)
+    {
+        for (auto it = this->begin(); it != this->end(); ++it)
+        {
+            if (m_compare(it->first, key))
+            {
+                return {it, false};
+            }
+        }
+        Container::emplace_back(std::forward<KeyType>(key), std::forward<T>(t));
+        return {std::prev(this->end()), true};
+    }
+
+    T& operator[](const key_type& key)
+    {
+        return emplace(key, T{}).first->second;
+    }
+
+    template<class KeyType, detail::enable_if_t<
+                 detail::is_usable_as_key_type<key_compare, key_type, KeyType>::value, int> = 0>
+    T & operator[](KeyType && key)
+    {
+        return emplace(std::forward<KeyType>(key), T{}).first->second;
+    }
+
+    const T& operator[](const key_type& key) const
+    {
+        return at(key);
+    }
+
+    template<class KeyType, detail::enable_if_t<
+                 detail::is_usable_as_key_type<key_compare, key_type, KeyType>::value, int> = 0>
+    const T & operator[](KeyType && key) const
+    {
+        return at(std::forward<KeyType>(key));
+    }
+
+    T& at(const key_type& key)
+    {
+        for (auto it = this->begin(); it != this->end(); ++it)
+        {
+            if (m_compare(it->first, key))
+            {
+                return it->second;
+            }
+        }
+
+        JSON_THROW(std::out_of_range("key not found"));
+    }
+
+    template<class KeyType, detail::enable_if_t<
+                 detail::is_usable_as_key_type<key_compare, key_type, KeyType>::value, int> = 0>
+    T & at(KeyType && key)
+    {
+        for (auto it = this->begin(); it != this->end(); ++it)
+        {
+            if (m_compare(it->first, key))
+            {
+                return it->second;
+            }
+        }
+
+        JSON_THROW(std::out_of_range("key not found"));
+    }
+
+    const T& at(const key_type& key) const
+    {
+        for (auto it = this->begin(); it != this->end(); ++it)
+        {
+            if (m_compare(it->first, key))
+            {
+                return it->second;
+            }
+        }
+
+        JSON_THROW(std::out_of_range("key not found"));
+    }
+
+    template<class KeyType, detail::enable_if_t<
+                 detail::is_usable_as_key_type<key_compare, key_type, KeyType>::value, int> = 0>
+    const T & at(KeyType && key) const
+    {
+        for (auto it = this->begin(); it != this->end(); ++it)
+        {
+            if (m_compare(it->first, key))
+            {
+                return it->second;
+            }
+        }
+
+        JSON_THROW(std::out_of_range("key not found"));
+    }
+
+    size_type erase(const key_type& key)
+    {
+        for (auto it = this->begin(); it != this->end(); ++it)
+        {
+            if (m_compare(it->first, key))
+            {
+                // Since we cannot move const Keys, re-construct them in place
+                for (auto next = it; ++next != this->end(); ++it)
+                {
+                    it->~value_type(); // Destroy but keep allocation
+                    new (&*it) value_type{std::move(*next)};
+                }
+                Container::pop_back();
+                return 1;
+            }
+        }
+        return 0;
+    }
+
+    template<class KeyType, detail::enable_if_t<
+                 detail::is_usable_as_key_type<key_compare, key_type, KeyType>::value, int> = 0>
+    size_type erase(KeyType && key)
+    {
+        for (auto it = this->begin(); it != this->end(); ++it)
+        {
+            if (m_compare(it->first, key))
+            {
+                // Since we cannot move const Keys, re-construct them in place
+                for (auto next = it; ++next != this->end(); ++it)
+                {
+                    it->~value_type(); // Destroy but keep allocation
+                    new (&*it) value_type{std::move(*next)};
+                }
+                Container::pop_back();
+                return 1;
+            }
+        }
+        return 0;
+    }
+
+    iterator erase(iterator pos)
+    {
+        return erase(pos, std::next(pos));
+    }
+
+    iterator erase(iterator first, iterator last)
+    {
+        if (first == last)
+        {
+            return first;
+        }
+
+        const auto elements_affected = std::distance(first, last);
+        const auto offset = std::distance(Container::begin(), first);
+
+        // This is the start situation. We need to delete elements_affected
+        // elements (3 in this example: e, f, g), and need to return an
+        // iterator past the last deleted element (h in this example).
+        // Note that offset is the distance from the start of the vector
+        // to first. We will need this later.
+
+        // [ a, b, c, d, e, f, g, h, i, j ]
+        //               ^        ^
+        //             first    last
+
+        // Since we cannot move const Keys, we re-construct them in place.
+        // We start at first and re-construct (viz. copy) the elements from
+        // the back of the vector. Example for first iteration:
+
+        //               ,--------.
+        //               v        |   destroy e and re-construct with h
+        // [ a, b, c, d, e, f, g, h, i, j ]
+        //               ^        ^
+        //               it       it + elements_affected
+
+        for (auto it = first; std::next(it, elements_affected) != Container::end(); ++it)
+        {
+            it->~value_type(); // destroy but keep allocation
+            new (&*it) value_type{std::move(*std::next(it, elements_affected))}; // "move" next element to it
+        }
+
+        // [ a, b, c, d, h, i, j, h, i, j ]
+        //               ^        ^
+        //             first    last
+
+        // remove the unneeded elements at the end of the vector
+        Container::resize(this->size() - static_cast<size_type>(elements_affected));
+
+        // [ a, b, c, d, h, i, j ]
+        //               ^        ^
+        //             first    last
+
+        // first is now pointing past the last deleted element, but we cannot
+        // use this iterator, because it may have been invalidated by the
+        // resize call. Instead, we can return begin() + offset.
+        return Container::begin() + offset;
+    }
+
+    size_type count(const key_type& key) const
+    {
+        for (auto it = this->begin(); it != this->end(); ++it)
+        {
+            if (m_compare(it->first, key))
+            {
+                return 1;
+            }
+        }
+        return 0;
+    }
+
+    template<class KeyType, detail::enable_if_t<
+                 detail::is_usable_as_key_type<key_compare, key_type, KeyType>::value, int> = 0>
+    size_type count(KeyType && key) const
+    {
+        for (auto it = this->begin(); it != this->end(); ++it)
+        {
+            if (m_compare(it->first, key))
+            {
+                return 1;
+            }
+        }
+        return 0;
+    }
+
+    iterator find(const key_type& key)
+    {
+        for (auto it = this->begin(); it != this->end(); ++it)
+        {
+            if (m_compare(it->first, key))
+            {
+                return it;
+            }
+        }
+        return Container::end();
+    }
+
+    template<class KeyType, detail::enable_if_t<
+                 detail::is_usable_as_key_type<key_compare, key_type, KeyType>::value, int> = 0>
+    iterator find(KeyType && key)
+    {
+        for (auto it = this->begin(); it != this->end(); ++it)
+        {
+            if (m_compare(it->first, key))
+            {
+                return it;
+            }
+        }
+        return Container::end();
+    }
+
+    const_iterator find(const key_type& key) const
+    {
+        for (auto it = this->begin(); it != this->end(); ++it)
+        {
+            if (m_compare(it->first, key))
+            {
+                return it;
+            }
+        }
+        return Container::end();
+    }
+
+    std::pair<iterator, bool> insert( value_type&& value )
+    {
+        return emplace(value.first, std::move(value.second));
+    }
+
+    std::pair<iterator, bool> insert( const value_type& value )
+    {
+        for (auto it = this->begin(); it != this->end(); ++it)
+        {
+            if (m_compare(it->first, value.first))
+            {
+                return {it, false};
+            }
+        }
+        Container::push_back(value);
+        return {--this->end(), true};
+    }
+
+    template<typename InputIt>
+    using require_input_iter = typename std::enable_if<std::is_convertible<typename std::iterator_traits<InputIt>::iterator_category,
+            std::input_iterator_tag>::value>::type;
+
+    template<typename InputIt, typename = require_input_iter<InputIt>>
+    void insert(InputIt first, InputIt last)
+    {
+        for (auto it = first; it != last; ++it)
+        {
+            insert(*it);
+        }
+    }
+
+private:
+    JSON_NO_UNIQUE_ADDRESS key_compare m_compare = key_compare();
+};
+
+NLOHMANN_JSON_NAMESPACE_END
+
+
+#if defined(JSON_HAS_CPP_17)
+    #include <any>
+    #include <string_view>
+#endif
+
+/*!
+@brief namespace for Niels Lohmann
+@see https://github.com/nlohmann
+@since version 1.0.0
+*/
+NLOHMANN_JSON_NAMESPACE_BEGIN
+
+/*!
+@brief a class to store JSON values
+
+@internal
+@invariant The member variables @a m_value and @a m_type have the following
+relationship:
+- If `m_type == value_t::object`, then `m_value.object != nullptr`.
+- If `m_type == value_t::array`, then `m_value.array != nullptr`.
+- If `m_type == value_t::string`, then `m_value.string != nullptr`.
+The invariants are checked by member function assert_invariant().
+
+@note ObjectType trick from https://stackoverflow.com/a/9860911
+@endinternal
+
+@since version 1.0.0
+
+@nosubgrouping
+*/
+NLOHMANN_BASIC_JSON_TPL_DECLARATION
+class basic_json // NOLINT(cppcoreguidelines-special-member-functions,hicpp-special-member-functions)
+{
+  private:
+    template<detail::value_t> friend struct detail::external_constructor;
+
+    template<typename>
+    friend class ::nlohmann::json_pointer;
+    // can be restored when json_pointer backwards compatibility is removed
+    // friend ::nlohmann::json_pointer<StringType>;
+
+    template<typename BasicJsonType, typename InputType>
+    friend class ::nlohmann::detail::parser;
+    friend ::nlohmann::detail::serializer<basic_json>;
+    template<typename BasicJsonType>
+    friend class ::nlohmann::detail::iter_impl;
+    template<typename BasicJsonType, typename CharType>
+    friend class ::nlohmann::detail::binary_writer;
+    template<typename BasicJsonType, typename InputType, typename SAX>
+    friend class ::nlohmann::detail::binary_reader;
+    template<typename BasicJsonType>
+    friend class ::nlohmann::detail::json_sax_dom_parser;
+    template<typename BasicJsonType>
+    friend class ::nlohmann::detail::json_sax_dom_callback_parser;
+    friend class ::nlohmann::detail::exception;
+
+    /// workaround type for MSVC
+    using basic_json_t = NLOHMANN_BASIC_JSON_TPL;
+
+  JSON_PRIVATE_UNLESS_TESTED:
+    // convenience aliases for types residing in namespace detail;
+    using lexer = ::nlohmann::detail::lexer_base<basic_json>;
+
+    template<typename InputAdapterType>
+    static ::nlohmann::detail::parser<basic_json, InputAdapterType> parser(
+        InputAdapterType adapter,
+        detail::parser_callback_t<basic_json>cb = nullptr,
+        const bool allow_exceptions = true,
+        const bool ignore_comments = false
+                                 )
+    {
+        return ::nlohmann::detail::parser<basic_json, InputAdapterType>(std::move(adapter),
+                std::move(cb), allow_exceptions, ignore_comments);
+    }
+
+  private:
+    using primitive_iterator_t = ::nlohmann::detail::primitive_iterator_t;
+    template<typename BasicJsonType>
+    using internal_iterator = ::nlohmann::detail::internal_iterator<BasicJsonType>;
+    template<typename BasicJsonType>
+    using iter_impl = ::nlohmann::detail::iter_impl<BasicJsonType>;
+    template<typename Iterator>
+    using iteration_proxy = ::nlohmann::detail::iteration_proxy<Iterator>;
+    template<typename Base> using json_reverse_iterator = ::nlohmann::detail::json_reverse_iterator<Base>;
+
+    template<typename CharType>
+    using output_adapter_t = ::nlohmann::detail::output_adapter_t<CharType>;
+
+    template<typename InputType>
+    using binary_reader = ::nlohmann::detail::binary_reader<basic_json, InputType>;
+    template<typename CharType> using binary_writer = ::nlohmann::detail::binary_writer<basic_json, CharType>;
+
+  JSON_PRIVATE_UNLESS_TESTED:
+    using serializer = ::nlohmann::detail::serializer<basic_json>;
+
+  public:
+    using value_t = detail::value_t;
+    /// JSON Pointer, see @ref nlohmann::json_pointer
+    using json_pointer = ::nlohmann::json_pointer<StringType>;
+    template<typename T, typename SFINAE>
+    using json_serializer = JSONSerializer<T, SFINAE>;
+    /// how to treat decoding errors
+    using error_handler_t = detail::error_handler_t;
+    /// how to treat CBOR tags
+    using cbor_tag_handler_t = detail::cbor_tag_handler_t;
+    /// helper type for initializer lists of basic_json values
+    using initializer_list_t = std::initializer_list<detail::json_ref<basic_json>>;
+
+    using input_format_t = detail::input_format_t;
+    /// SAX interface type, see @ref nlohmann::json_sax
+    using json_sax_t = json_sax<basic_json>;
+
+    ////////////////
+    // exceptions //
+    ////////////////
+
+    /// @name exceptions
+    /// Classes to implement user-defined exceptions.
+    /// @{
+
+    using exception = detail::exception;
+    using parse_error = detail::parse_error;
+    using invalid_iterator = detail::invalid_iterator;
+    using type_error = detail::type_error;
+    using out_of_range = detail::out_of_range;
+    using other_error = detail::other_error;
+
+    /// @}
+
+
+    /////////////////////
+    // container types //
+    /////////////////////
+
+    /// @name container types
+    /// The canonic container types to use @ref basic_json like any other STL
+    /// container.
+    /// @{
+
+    /// the type of elements in a basic_json container
+    using value_type = basic_json;
+
+    /// the type of an element reference
+    using reference = value_type&;
+    /// the type of an element const reference
+    using const_reference = const value_type&;
+
+    /// a type to represent differences between iterators
+    using difference_type = std::ptrdiff_t;
+    /// a type to represent container sizes
+    using size_type = std::size_t;
+
+    /// the allocator type
+    using allocator_type = AllocatorType<basic_json>;
+
+    /// the type of an element pointer
+    using pointer = typename std::allocator_traits<allocator_type>::pointer;
+    /// the type of an element const pointer
+    using const_pointer = typename std::allocator_traits<allocator_type>::const_pointer;
+
+    /// an iterator for a basic_json container
+    using iterator = iter_impl<basic_json>;
+    /// a const iterator for a basic_json container
+    using const_iterator = iter_impl<const basic_json>;
+    /// a reverse iterator for a basic_json container
+    using reverse_iterator = json_reverse_iterator<typename basic_json::iterator>;
+    /// a const reverse iterator for a basic_json container
+    using const_reverse_iterator = json_reverse_iterator<typename basic_json::const_iterator>;
+
+    /// @}
+
+
+    /// @brief returns the allocator associated with the container
+    /// @sa https://json.nlohmann.me/api/basic_json/get_allocator/
+    static allocator_type get_allocator()
+    {
+        return allocator_type();
+    }
+
+    /// @brief returns version information on the library
+    /// @sa https://json.nlohmann.me/api/basic_json/meta/
+    JSON_HEDLEY_WARN_UNUSED_RESULT
+    static basic_json meta()
+    {
+        basic_json result;
+
+        result["copyright"] = "(C) 2013-2022 Niels Lohmann";
+        result["name"] = "JSON for Modern C++";
+        result["url"] = "https://github.com/nlohmann/json";
+        result["version"]["string"] =
+            detail::concat(std::to_string(NLOHMANN_JSON_VERSION_MAJOR), '.',
+                           std::to_string(NLOHMANN_JSON_VERSION_MINOR), '.',
+                           std::to_string(NLOHMANN_JSON_VERSION_PATCH));
+        result["version"]["major"] = NLOHMANN_JSON_VERSION_MAJOR;
+        result["version"]["minor"] = NLOHMANN_JSON_VERSION_MINOR;
+        result["version"]["patch"] = NLOHMANN_JSON_VERSION_PATCH;
+
+#ifdef _WIN32
+        result["platform"] = "win32";
+#elif defined __linux__
+        result["platform"] = "linux";
+#elif defined __APPLE__
+        result["platform"] = "apple";
+#elif defined __unix__
+        result["platform"] = "unix";
+#else
+        result["platform"] = "unknown";
+#endif
+
+#if defined(__ICC) || defined(__INTEL_COMPILER)
+        result["compiler"] = {{"family", "icc"}, {"version", __INTEL_COMPILER}};
+#elif defined(__clang__)
+        result["compiler"] = {{"family", "clang"}, {"version", __clang_version__}};
+#elif defined(__GNUC__) || defined(__GNUG__)
+        result["compiler"] = {{"family", "gcc"}, {"version", detail::concat(
+                    std::to_string(__GNUC__), '.',
+                    std::to_string(__GNUC_MINOR__), '.',
+                    std::to_string(__GNUC_PATCHLEVEL__))
+            }
+        };
+#elif defined(__HP_cc) || defined(__HP_aCC)
+        result["compiler"] = "hp"
+#elif defined(__IBMCPP__)
+        result["compiler"] = {{"family", "ilecpp"}, {"version", __IBMCPP__}};
+#elif defined(_MSC_VER)
+        result["compiler"] = {{"family", "msvc"}, {"version", _MSC_VER}};
+#elif defined(__PGI)
+        result["compiler"] = {{"family", "pgcpp"}, {"version", __PGI}};
+#elif defined(__SUNPRO_CC)
+        result["compiler"] = {{"family", "sunpro"}, {"version", __SUNPRO_CC}};
+#else
+        result["compiler"] = {{"family", "unknown"}, {"version", "unknown"}};
+#endif
+
+
+#if defined(_MSVC_LANG)
+        result["compiler"]["c++"] = std::to_string(_MSVC_LANG);
+#elif defined(__cplusplus)
+        result["compiler"]["c++"] = std::to_string(__cplusplus);
+#else
+        result["compiler"]["c++"] = "unknown";
+#endif
+        return result;
+    }
+
+
+    ///////////////////////////
+    // JSON value data types //
+    ///////////////////////////
+
+    /// @name JSON value data types
+    /// The data types to store a JSON value. These types are derived from
+    /// the template arguments passed to class @ref basic_json.
+    /// @{
+
+    /// @brief default object key comparator type
+    /// The actual object key comparator type (@ref object_comparator_t) may be
+    /// different.
+    /// @sa https://json.nlohmann.me/api/basic_json/default_object_comparator_t/
+#if defined(JSON_HAS_CPP_14)
+    // use of transparent comparator avoids unnecessary repeated construction of temporaries
+    // in functions involving lookup by key with types other than object_t::key_type (aka. StringType)
+    using default_object_comparator_t = std::less<>;
+#else
+    using default_object_comparator_t = std::less<StringType>;
+#endif
+
+    /// @brief a type for an object
+    /// @sa https://json.nlohmann.me/api/basic_json/object_t/
+    using object_t = ObjectType<StringType,
+          basic_json,
+          default_object_comparator_t,
+          AllocatorType<std::pair<const StringType,
+          basic_json>>>;
+
+    /// @brief a type for an array
+    /// @sa https://json.nlohmann.me/api/basic_json/array_t/
+    using array_t = ArrayType<basic_json, AllocatorType<basic_json>>;
+
+    /// @brief a type for a string
+    /// @sa https://json.nlohmann.me/api/basic_json/string_t/
+    using string_t = StringType;
+
+    /// @brief a type for a boolean
+    /// @sa https://json.nlohmann.me/api/basic_json/boolean_t/
+    using boolean_t = BooleanType;
+
+    /// @brief a type for a number (integer)
+    /// @sa https://json.nlohmann.me/api/basic_json/number_integer_t/
+    using number_integer_t = NumberIntegerType;
+
+    /// @brief a type for a number (unsigned)
+    /// @sa https://json.nlohmann.me/api/basic_json/number_unsigned_t/
+    using number_unsigned_t = NumberUnsignedType;
+
+    /// @brief a type for a number (floating-point)
+    /// @sa https://json.nlohmann.me/api/basic_json/number_float_t/
+    using number_float_t = NumberFloatType;
+
+    /// @brief a type for a packed binary type
+    /// @sa https://json.nlohmann.me/api/basic_json/binary_t/
+    using binary_t = nlohmann::byte_container_with_subtype<BinaryType>;
+
+    /// @brief object key comparator type
+    /// @sa https://json.nlohmann.me/api/basic_json/object_comparator_t/
+    using object_comparator_t = detail::actual_object_comparator_t<basic_json>;
+
+    /// @}
+
+  private:
+
+    /// helper for exception-safe object creation
+    template<typename T, typename... Args>
+    JSON_HEDLEY_RETURNS_NON_NULL
+    static T* create(Args&& ... args)
+    {
+        AllocatorType<T> alloc;
+        using AllocatorTraits = std::allocator_traits<AllocatorType<T>>;
+
+        auto deleter = [&](T * obj)
+        {
+            AllocatorTraits::deallocate(alloc, obj, 1);
+        };
+        std::unique_ptr<T, decltype(deleter)> obj(AllocatorTraits::allocate(alloc, 1), deleter);
+        AllocatorTraits::construct(alloc, obj.get(), std::forward<Args>(args)...);
+        JSON_ASSERT(obj != nullptr);
+        return obj.release();
+    }
+
+    ////////////////////////
+    // JSON value storage //
+    ////////////////////////
+
+  JSON_PRIVATE_UNLESS_TESTED:
+    /*!
+    @brief a JSON value
+
+    The actual storage for a JSON value of the @ref basic_json class. This
+    union combines the different storage types for the JSON value types
+    defined in @ref value_t.
+
+    JSON type | value_t type    | used type
+    --------- | --------------- | ------------------------
+    object    | object          | pointer to @ref object_t
+    array     | array           | pointer to @ref array_t
+    string    | string          | pointer to @ref string_t
+    boolean   | boolean         | @ref boolean_t
+    number    | number_integer  | @ref number_integer_t
+    number    | number_unsigned | @ref number_unsigned_t
+    number    | number_float    | @ref number_float_t
+    binary    | binary          | pointer to @ref binary_t
+    null      | null            | *no value is stored*
+
+    @note Variable-length types (objects, arrays, and strings) are stored as
+    pointers. The size of the union should not exceed 64 bits if the default
+    value types are used.
+
+    @since version 1.0.0
+    */
+    union json_value
+    {
+        /// object (stored with pointer to save storage)
+        object_t* object;
+        /// array (stored with pointer to save storage)
+        array_t* array;
+        /// string (stored with pointer to save storage)
+        string_t* string;
+        /// binary (stored with pointer to save storage)
+        binary_t* binary;
+        /// boolean
+        boolean_t boolean;
+        /// number (integer)
+        number_integer_t number_integer;
+        /// number (unsigned integer)
+        number_unsigned_t number_unsigned;
+        /// number (floating-point)
+        number_float_t number_float;
+
+        /// default constructor (for null values)
+        json_value() = default;
+        /// constructor for booleans
+        json_value(boolean_t v) noexcept : boolean(v) {}
+        /// constructor for numbers (integer)
+        json_value(number_integer_t v) noexcept : number_integer(v) {}
+        /// constructor for numbers (unsigned)
+        json_value(number_unsigned_t v) noexcept : number_unsigned(v) {}
+        /// constructor for numbers (floating-point)
+        json_value(number_float_t v) noexcept : number_float(v) {}
+        /// constructor for empty values of a given type
+        json_value(value_t t)
+        {
+            switch (t)
+            {
+                case value_t::object:
+                {
+                    object = create<object_t>();
+                    break;
+                }
+
+                case value_t::array:
+                {
+                    array = create<array_t>();
+                    break;
+                }
+
+                case value_t::string:
+                {
+                    string = create<string_t>("");
+                    break;
+                }
+
+                case value_t::binary:
+                {
+                    binary = create<binary_t>();
+                    break;
+                }
+
+                case value_t::boolean:
+                {
+                    boolean = static_cast<boolean_t>(false);
+                    break;
+                }
+
+                case value_t::number_integer:
+                {
+                    number_integer = static_cast<number_integer_t>(0);
+                    break;
+                }
+
+                case value_t::number_unsigned:
+                {
+                    number_unsigned = static_cast<number_unsigned_t>(0);
+                    break;
+                }
+
+                case value_t::number_float:
+                {
+                    number_float = static_cast<number_float_t>(0.0);
+                    break;
+                }
+
+                case value_t::null:
+                {
+                    object = nullptr;  // silence warning, see #821
+                    break;
+                }
+
+                case value_t::discarded:
+                default:
+                {
+                    object = nullptr;  // silence warning, see #821
+                    if (JSON_HEDLEY_UNLIKELY(t == value_t::null))
+                    {
+                        JSON_THROW(other_error::create(500, "961c151d2e87f2686a955a9be24d316f1362bf21 3.11.2", nullptr)); // LCOV_EXCL_LINE
+                    }
+                    break;
+                }
+            }
+        }
+
+        /// constructor for strings
+        json_value(const string_t& value) : string(create<string_t>(value)) {}
+
+        /// constructor for rvalue strings
+        json_value(string_t&& value) : string(create<string_t>(std::move(value))) {}
+
+        /// constructor for objects
+        json_value(const object_t& value) : object(create<object_t>(value)) {}
+
+        /// constructor for rvalue objects
+        json_value(object_t&& value) : object(create<object_t>(std::move(value))) {}
+
+        /// constructor for arrays
+        json_value(const array_t& value) : array(create<array_t>(value)) {}
+
+        /// constructor for rvalue arrays
+        json_value(array_t&& value) : array(create<array_t>(std::move(value))) {}
+
+        /// constructor for binary arrays
+        json_value(const typename binary_t::container_type& value) : binary(create<binary_t>(value)) {}
+
+        /// constructor for rvalue binary arrays
+        json_value(typename binary_t::container_type&& value) : binary(create<binary_t>(std::move(value))) {}
+
+        /// constructor for binary arrays (internal type)
+        json_value(const binary_t& value) : binary(create<binary_t>(value)) {}
+
+        /// constructor for rvalue binary arrays (internal type)
+        json_value(binary_t&& value) : binary(create<binary_t>(std::move(value))) {}
+
+        void destroy(value_t t)
+        {
+            if (t == value_t::array || t == value_t::object)
+            {
+                // flatten the current json_value to a heap-allocated stack
+                std::vector<basic_json> stack;
+
+                // move the top-level items to stack
+                if (t == value_t::array)
+                {
+                    stack.reserve(array->size());
+                    std::move(array->begin(), array->end(), std::back_inserter(stack));
+                }
+                else
+                {
+                    stack.reserve(object->size());
+                    for (auto&& it : *object)
+                    {
+                        stack.push_back(std::move(it.second));
+                    }
+                }
+
+                while (!stack.empty())
+                {
+                    // move the last item to local variable to be processed
+                    basic_json current_item(std::move(stack.back()));
+                    stack.pop_back();
+
+                    // if current_item is array/object, move
+                    // its children to the stack to be processed later
+                    if (current_item.is_array())
+                    {
+                        std::move(current_item.m_value.array->begin(), current_item.m_value.array->end(), std::back_inserter(stack));
+
+                        current_item.m_value.array->clear();
+                    }
+                    else if (current_item.is_object())
+                    {
+                        for (auto&& it : *current_item.m_value.object)
+                        {
+                            stack.push_back(std::move(it.second));
+                        }
+
+                        current_item.m_value.object->clear();
+                    }
+
+                    // it's now safe that current_item get destructed
+                    // since it doesn't have any children
+                }
+            }
+
+            switch (t)
+            {
+                case value_t::object:
+                {
+                    AllocatorType<object_t> alloc;
+                    std::allocator_traits<decltype(alloc)>::destroy(alloc, object);
+                    std::allocator_traits<decltype(alloc)>::deallocate(alloc, object, 1);
+                    break;
+                }
+
+                case value_t::array:
+                {
+                    AllocatorType<array_t> alloc;
+                    std::allocator_traits<decltype(alloc)>::destroy(alloc, array);
+                    std::allocator_traits<decltype(alloc)>::deallocate(alloc, array, 1);
+                    break;
+                }
+
+                case value_t::string:
+                {
+                    AllocatorType<string_t> alloc;
+                    std::allocator_traits<decltype(alloc)>::destroy(alloc, string);
+                    std::allocator_traits<decltype(alloc)>::deallocate(alloc, string, 1);
+                    break;
+                }
+
+                case value_t::binary:
+                {
+                    AllocatorType<binary_t> alloc;
+                    std::allocator_traits<decltype(alloc)>::destroy(alloc, binary);
+                    std::allocator_traits<decltype(alloc)>::deallocate(alloc, binary, 1);
+                    break;
+                }
+
+                case value_t::null:
+                case value_t::boolean:
+                case value_t::number_integer:
+                case value_t::number_unsigned:
+                case value_t::number_float:
+                case value_t::discarded:
+                default:
+                {
+                    break;
+                }
+            }
+        }
+    };
+
+  private:
+    /*!
+    @brief checks the class invariants
+
+    This function asserts the class invariants. It needs to be called at the
+    end of every constructor to make sure that created objects respect the
+    invariant. Furthermore, it has to be called each time the type of a JSON
+    value is changed, because the invariant expresses a relationship between
+    @a m_type and @a m_value.
+
+    Furthermore, the parent relation is checked for arrays and objects: If
+    @a check_parents true and the value is an array or object, then the
+    container's elements must have the current value as parent.
+
+    @param[in] check_parents  whether the parent relation should be checked.
+               The value is true by default and should only be set to false
+               during destruction of objects when the invariant does not
+               need to hold.
+    */
+    void assert_invariant(bool check_parents = true) const noexcept
+    {
+        JSON_ASSERT(m_type != value_t::object || m_value.object != nullptr);
+        JSON_ASSERT(m_type != value_t::array || m_value.array != nullptr);
+        JSON_ASSERT(m_type != value_t::string || m_value.string != nullptr);
+        JSON_ASSERT(m_type != value_t::binary || m_value.binary != nullptr);
+
+#if JSON_DIAGNOSTICS
+        JSON_TRY
+        {
+            // cppcheck-suppress assertWithSideEffect
+            JSON_ASSERT(!check_parents || !is_structured() || std::all_of(begin(), end(), [this](const basic_json & j)
+            {
+                return j.m_parent == this;
+            }));
+        }
+        JSON_CATCH(...) {} // LCOV_EXCL_LINE
+#endif
+        static_cast<void>(check_parents);
+    }
+
+    void set_parents()
+    {
+#if JSON_DIAGNOSTICS
+        switch (m_type)
+        {
+            case value_t::array:
+            {
+                for (auto& element : *m_value.array)
+                {
+                    element.m_parent = this;
+                }
+                break;
+            }
+
+            case value_t::object:
+            {
+                for (auto& element : *m_value.object)
+                {
+                    element.second.m_parent = this;
+                }
+                break;
+            }
+
+            case value_t::null:
+            case value_t::string:
+            case value_t::boolean:
+            case value_t::number_integer:
+            case value_t::number_unsigned:
+            case value_t::number_float:
+            case value_t::binary:
+            case value_t::discarded:
+            default:
+                break;
+        }
+#endif
+    }
+
+    iterator set_parents(iterator it, typename iterator::difference_type count_set_parents)
+    {
+#if JSON_DIAGNOSTICS
+        for (typename iterator::difference_type i = 0; i < count_set_parents; ++i)
+        {
+            (it + i)->m_parent = this;
+        }
+#else
+        static_cast<void>(count_set_parents);
+#endif
+        return it;
+    }
+
+    reference set_parent(reference j, std::size_t old_capacity = static_cast<std::size_t>(-1))
+    {
+#if JSON_DIAGNOSTICS
+        if (old_capacity != static_cast<std::size_t>(-1))
+        {
+            // see https://github.com/nlohmann/json/issues/2838
+            JSON_ASSERT(type() == value_t::array);
+            if (JSON_HEDLEY_UNLIKELY(m_value.array->capacity() != old_capacity))
+            {
+                // capacity has changed: update all parents
+                set_parents();
+                return j;
+            }
+        }
+
+        // ordered_json uses a vector internally, so pointers could have
+        // been invalidated; see https://github.com/nlohmann/json/issues/2962
+#ifdef JSON_HEDLEY_MSVC_VERSION
+#pragma warning(push )
+#pragma warning(disable : 4127) // ignore warning to replace if with if constexpr
+#endif
+        if (detail::is_ordered_map<object_t>::value)
+        {
+            set_parents();
+            return j;
+        }
+#ifdef JSON_HEDLEY_MSVC_VERSION
+#pragma warning( pop )
+#endif
+
+        j.m_parent = this;
+#else
+        static_cast<void>(j);
+        static_cast<void>(old_capacity);
+#endif
+        return j;
+    }
+
+  public:
+    //////////////////////////
+    // JSON parser callback //
+    //////////////////////////
+
+    /// @brief parser event types
+    /// @sa https://json.nlohmann.me/api/basic_json/parse_event_t/
+    using parse_event_t = detail::parse_event_t;
+
+    /// @brief per-element parser callback type
+    /// @sa https://json.nlohmann.me/api/basic_json/parser_callback_t/
+    using parser_callback_t = detail::parser_callback_t<basic_json>;
+
+    //////////////////
+    // constructors //
+    //////////////////
+
+    /// @name constructors and destructors
+    /// Constructors of class @ref basic_json, copy/move constructor, copy
+    /// assignment, static functions creating objects, and the destructor.
+    /// @{
+
+    /// @brief create an empty value with a given type
+    /// @sa https://json.nlohmann.me/api/basic_json/basic_json/
+    basic_json(const value_t v)
+        : m_type(v), m_value(v)
+    {
+        assert_invariant();
+    }
+
+    /// @brief create a null object
+    /// @sa https://json.nlohmann.me/api/basic_json/basic_json/
+    basic_json(std::nullptr_t = nullptr) noexcept // NOLINT(bugprone-exception-escape)
+        : basic_json(value_t::null)
+    {
+        assert_invariant();
+    }
+
+    /// @brief create a JSON value from compatible types
+    /// @sa https://json.nlohmann.me/api/basic_json/basic_json/
+    template < typename CompatibleType,
+               typename U = detail::uncvref_t<CompatibleType>,
+               detail::enable_if_t <
+                   !detail::is_basic_json<U>::value && detail::is_compatible_type<basic_json_t, U>::value, int > = 0 >
+    basic_json(CompatibleType && val) noexcept(noexcept( // NOLINT(bugprone-forwarding-reference-overload,bugprone-exception-escape)
+                JSONSerializer<U>::to_json(std::declval<basic_json_t&>(),
+                                           std::forward<CompatibleType>(val))))
+    {
+        JSONSerializer<U>::to_json(*this, std::forward<CompatibleType>(val));
+        set_parents();
+        assert_invariant();
+    }
+
+    /// @brief create a JSON value from an existing one
+    /// @sa https://json.nlohmann.me/api/basic_json/basic_json/
+    template < typename BasicJsonType,
+               detail::enable_if_t <
+                   detail::is_basic_json<BasicJsonType>::value&& !std::is_same<basic_json, BasicJsonType>::value, int > = 0 >
+    basic_json(const BasicJsonType& val)
+    {
+        using other_boolean_t = typename BasicJsonType::boolean_t;
+        using other_number_float_t = typename BasicJsonType::number_float_t;
+        using other_number_integer_t = typename BasicJsonType::number_integer_t;
+        using other_number_unsigned_t = typename BasicJsonType::number_unsigned_t;
+        using other_string_t = typename BasicJsonType::string_t;
+        using other_object_t = typename BasicJsonType::object_t;
+        using other_array_t = typename BasicJsonType::array_t;
+        using other_binary_t = typename BasicJsonType::binary_t;
+
+        switch (val.type())
+        {
+            case value_t::boolean:
+                JSONSerializer<other_boolean_t>::to_json(*this, val.template get<other_boolean_t>());
+                break;
+            case value_t::number_float:
+                JSONSerializer<other_number_float_t>::to_json(*this, val.template get<other_number_float_t>());
+                break;
+            case value_t::number_integer:
+                JSONSerializer<other_number_integer_t>::to_json(*this, val.template get<other_number_integer_t>());
+                break;
+            case value_t::number_unsigned:
+                JSONSerializer<other_number_unsigned_t>::to_json(*this, val.template get<other_number_unsigned_t>());
+                break;
+            case value_t::string:
+                JSONSerializer<other_string_t>::to_json(*this, val.template get_ref<const other_string_t&>());
+                break;
+            case value_t::object:
+                JSONSerializer<other_object_t>::to_json(*this, val.template get_ref<const other_object_t&>());
+                break;
+            case value_t::array:
+                JSONSerializer<other_array_t>::to_json(*this, val.template get_ref<const other_array_t&>());
+                break;
+            case value_t::binary:
+                JSONSerializer<other_binary_t>::to_json(*this, val.template get_ref<const other_binary_t&>());
+                break;
+            case value_t::null:
+                *this = nullptr;
+                break;
+            case value_t::discarded:
+                m_type = value_t::discarded;
+                break;
+            default:            // LCOV_EXCL_LINE
+                JSON_ASSERT(false); // NOLINT(cert-dcl03-c,hicpp-static-assert,misc-static-assert) LCOV_EXCL_LINE
+        }
+        JSON_ASSERT(m_type == val.type());
+        set_parents();
+        assert_invariant();
+    }
+
+    /// @brief create a container (array or object) from an initializer list
+    /// @sa https://json.nlohmann.me/api/basic_json/basic_json/
+    basic_json(initializer_list_t init,
+               bool type_deduction = true,
+               value_t manual_type = value_t::array)
+    {
+        // check if each element is an array with two elements whose first
+        // element is a string
+        bool is_an_object = std::all_of(init.begin(), init.end(),
+                                        [](const detail::json_ref<basic_json>& element_ref)
+        {
+            return element_ref->is_array() && element_ref->size() == 2 && (*element_ref)[0].is_string();
+        });
+
+        // adjust type if type deduction is not wanted
+        if (!type_deduction)
+        {
+            // if array is wanted, do not create an object though possible
+            if (manual_type == value_t::array)
+            {
+                is_an_object = false;
+            }
+
+            // if object is wanted but impossible, throw an exception
+            if (JSON_HEDLEY_UNLIKELY(manual_type == value_t::object && !is_an_object))
+            {
+                JSON_THROW(type_error::create(301, "cannot create object from initializer list", nullptr));
+            }
+        }
+
+        if (is_an_object)
+        {
+            // the initializer list is a list of pairs -> create object
+            m_type = value_t::object;
+            m_value = value_t::object;
+
+            for (auto& element_ref : init)
+            {
+                auto element = element_ref.moved_or_copied();
+                m_value.object->emplace(
+                    std::move(*((*element.m_value.array)[0].m_value.string)),
+                    std::move((*element.m_value.array)[1]));
+            }
+        }
+        else
+        {
+            // the initializer list describes an array -> create array
+            m_type = value_t::array;
+            m_value.array = create<array_t>(init.begin(), init.end());
+        }
+
+        set_parents();
+        assert_invariant();
+    }
+
+    /// @brief explicitly create a binary array (without subtype)
+    /// @sa https://json.nlohmann.me/api/basic_json/binary/
+    JSON_HEDLEY_WARN_UNUSED_RESULT
+    static basic_json binary(const typename binary_t::container_type& init)
+    {
+        auto res = basic_json();
+        res.m_type = value_t::binary;
+        res.m_value = init;
+        return res;
+    }
+
+    /// @brief explicitly create a binary array (with subtype)
+    /// @sa https://json.nlohmann.me/api/basic_json/binary/
+    JSON_HEDLEY_WARN_UNUSED_RESULT
+    static basic_json binary(const typename binary_t::container_type& init, typename binary_t::subtype_type subtype)
+    {
+        auto res = basic_json();
+        res.m_type = value_t::binary;
+        res.m_value = binary_t(init, subtype);
+        return res;
+    }
+
+    /// @brief explicitly create a binary array
+    /// @sa https://json.nlohmann.me/api/basic_json/binary/
+    JSON_HEDLEY_WARN_UNUSED_RESULT
+    static basic_json binary(typename binary_t::container_type&& init)
+    {
+        auto res = basic_json();
+        res.m_type = value_t::binary;
+        res.m_value = std::move(init);
+        return res;
+    }
+
+    /// @brief explicitly create a binary array (with subtype)
+    /// @sa https://json.nlohmann.me/api/basic_json/binary/
+    JSON_HEDLEY_WARN_UNUSED_RESULT
+    static basic_json binary(typename binary_t::container_type&& init, typename binary_t::subtype_type subtype)
+    {
+        auto res = basic_json();
+        res.m_type = value_t::binary;
+        res.m_value = binary_t(std::move(init), subtype);
+        return res;
+    }
+
+    /// @brief explicitly create an array from an initializer list
+    /// @sa https://json.nlohmann.me/api/basic_json/array/
+    JSON_HEDLEY_WARN_UNUSED_RESULT
+    static basic_json array(initializer_list_t init = {})
+    {
+        return basic_json(init, false, value_t::array);
+    }
+
+    /// @brief explicitly create an object from an initializer list
+    /// @sa https://json.nlohmann.me/api/basic_json/object/
+    JSON_HEDLEY_WARN_UNUSED_RESULT
+    static basic_json object(initializer_list_t init = {})
+    {
+        return basic_json(init, false, value_t::object);
+    }
+
+    /// @brief construct an array with count copies of given value
+    /// @sa https://json.nlohmann.me/api/basic_json/basic_json/
+    basic_json(size_type cnt, const basic_json& val)
+        : m_type(value_t::array)
+    {
+        m_value.array = create<array_t>(cnt, val);
+        set_parents();
+        assert_invariant();
+    }
+
+    /// @brief construct a JSON container given an iterator range
+    /// @sa https://json.nlohmann.me/api/basic_json/basic_json/
+    template < class InputIT, typename std::enable_if <
+                   std::is_same<InputIT, typename basic_json_t::iterator>::value ||
+                   std::is_same<InputIT, typename basic_json_t::const_iterator>::value, int >::type = 0 >
+    basic_json(InputIT first, InputIT last)
+    {
+        JSON_ASSERT(first.m_object != nullptr);
+        JSON_ASSERT(last.m_object != nullptr);
+
+        // make sure iterator fits the current value
+        if (JSON_HEDLEY_UNLIKELY(first.m_object != last.m_object))
+        {
+            JSON_THROW(invalid_iterator::create(201, "iterators are not compatible", nullptr));
+        }
+
+        // copy type from first iterator
+        m_type = first.m_object->m_type;
+
+        // check if iterator range is complete for primitive values
+        switch (m_type)
+        {
+            case value_t::boolean:
+            case value_t::number_float:
+            case value_t::number_integer:
+            case value_t::number_unsigned:
+            case value_t::string:
+            {
+                if (JSON_HEDLEY_UNLIKELY(!first.m_it.primitive_iterator.is_begin()
+                                         || !last.m_it.primitive_iterator.is_end()))
+                {
+                    JSON_THROW(invalid_iterator::create(204, "iterators out of range", first.m_object));
+                }
+                break;
+            }
+
+            case value_t::null:
+            case value_t::object:
+            case value_t::array:
+            case value_t::binary:
+            case value_t::discarded:
+            default:
+                break;
+        }
+
+        switch (m_type)
+        {
+            case value_t::number_integer:
+            {
+                m_value.number_integer = first.m_object->m_value.number_integer;
+                break;
+            }
+
+            case value_t::number_unsigned:
+            {
+                m_value.number_unsigned = first.m_object->m_value.number_unsigned;
+                break;
+            }
+
+            case value_t::number_float:
+            {
+                m_value.number_float = first.m_object->m_value.number_float;
+                break;
+            }
+
+            case value_t::boolean:
+            {
+                m_value.boolean = first.m_object->m_value.boolean;
+                break;
+            }
+
+            case value_t::string:
+            {
+                m_value = *first.m_object->m_value.string;
+                break;
+            }
+
+            case value_t::object:
+            {
+                m_value.object = create<object_t>(first.m_it.object_iterator,
+                                                  last.m_it.object_iterator);
+                break;
+            }
+
+            case value_t::array:
+            {
+                m_value.array = create<array_t>(first.m_it.array_iterator,
+                                                last.m_it.array_iterator);
+                break;
+            }
+
+            case value_t::binary:
+            {
+                m_value = *first.m_object->m_value.binary;
+                break;
+            }
+
+            case value_t::null:
+            case value_t::discarded:
+            default:
+                JSON_THROW(invalid_iterator::create(206, detail::concat("cannot construct with iterators from ", first.m_object->type_name()), first.m_object));
+        }
+
+        set_parents();
+        assert_invariant();
+    }
+
+
+    ///////////////////////////////////////
+    // other constructors and destructor //
+    ///////////////////////////////////////
+
+    template<typename JsonRef,
+             detail::enable_if_t<detail::conjunction<detail::is_json_ref<JsonRef>,
+                                 std::is_same<typename JsonRef::value_type, basic_json>>::value, int> = 0 >
+    basic_json(const JsonRef& ref) : basic_json(ref.moved_or_copied()) {}
+
+    /// @brief copy constructor
+    /// @sa https://json.nlohmann.me/api/basic_json/basic_json/
+    basic_json(const basic_json& other)
+        : m_type(other.m_type)
+    {
+        // check of passed value is valid
+        other.assert_invariant();
+
+        switch (m_type)
+        {
+            case value_t::object:
+            {
+                m_value = *other.m_value.object;
+                break;
+            }
+
+            case value_t::array:
+            {
+                m_value = *other.m_value.array;
+                break;
+            }
+
+            case value_t::string:
+            {
+                m_value = *other.m_value.string;
+                break;
+            }
+
+            case value_t::boolean:
+            {
+                m_value = other.m_value.boolean;
+                break;
+            }
+
+            case value_t::number_integer:
+            {
+                m_value = other.m_value.number_integer;
+                break;
+            }
+
+            case value_t::number_unsigned:
+            {
+                m_value = other.m_value.number_unsigned;
+                break;
+            }
+
+            case value_t::number_float:
+            {
+                m_value = other.m_value.number_float;
+                break;
+            }
+
+            case value_t::binary:
+            {
+                m_value = *other.m_value.binary;
+                break;
+            }
+
+            case value_t::null:
+            case value_t::discarded:
+            default:
+                break;
+        }
+
+        set_parents();
+        assert_invariant();
+    }
+
+    /// @brief move constructor
+    /// @sa https://json.nlohmann.me/api/basic_json/basic_json/
+    basic_json(basic_json&& other) noexcept
+        : m_type(std::move(other.m_type)),
+          m_value(std::move(other.m_value))
+    {
+        // check that passed value is valid
+        other.assert_invariant(false);
+
+        // invalidate payload
+        other.m_type = value_t::null;
+        other.m_value = {};
+
+        set_parents();
+        assert_invariant();
+    }
+
+    /// @brief copy assignment
+    /// @sa https://json.nlohmann.me/api/basic_json/operator=/
+    basic_json& operator=(basic_json other) noexcept (
+        std::is_nothrow_move_constructible<value_t>::value&&
+        std::is_nothrow_move_assignable<value_t>::value&&
+        std::is_nothrow_move_constructible<json_value>::value&&
+        std::is_nothrow_move_assignable<json_value>::value
+    )
+    {
+        // check that passed value is valid
+        other.assert_invariant();
+
+        using std::swap;
+        swap(m_type, other.m_type);
+        swap(m_value, other.m_value);
+
+        set_parents();
+        assert_invariant();
+        return *this;
+    }
+
+    /// @brief destructor
+    /// @sa https://json.nlohmann.me/api/basic_json/~basic_json/
+    ~basic_json() noexcept
+    {
+        assert_invariant(false);
+        m_value.destroy(m_type);
+    }
+
+    /// @}
+
+  public:
+    ///////////////////////
+    // object inspection //
+    ///////////////////////
+
+    /// @name object inspection
+    /// Functions to inspect the type of a JSON value.
+    /// @{
+
+    /// @brief serialization
+    /// @sa https://json.nlohmann.me/api/basic_json/dump/
+    string_t dump(const int indent = -1,
+                  const char indent_char = ' ',
+                  const bool ensure_ascii = false,
+                  const error_handler_t error_handler = error_handler_t::strict) const
+    {
+        string_t result;
+        serializer s(detail::output_adapter<char, string_t>(result), indent_char, error_handler);
+
+        if (indent >= 0)
+        {
+            s.dump(*this, true, ensure_ascii, static_cast<unsigned int>(indent));
+        }
+        else
+        {
+            s.dump(*this, false, ensure_ascii, 0);
+        }
+
+        return result;
+    }
+
+    /// @brief return the type of the JSON value (explicit)
+    /// @sa https://json.nlohmann.me/api/basic_json/type/
+    constexpr value_t type() const noexcept
+    {
+        return m_type;
+    }
+
+    /// @brief return whether type is primitive
+    /// @sa https://json.nlohmann.me/api/basic_json/is_primitive/
+    constexpr bool is_primitive() const noexcept
+    {
+        return is_null() || is_string() || is_boolean() || is_number() || is_binary();
+    }
+
+    /// @brief return whether type is structured
+    /// @sa https://json.nlohmann.me/api/basic_json/is_structured/
+    constexpr bool is_structured() const noexcept
+    {
+        return is_array() || is_object();
+    }
+
+    /// @brief return whether value is null
+    /// @sa https://json.nlohmann.me/api/basic_json/is_null/
+    constexpr bool is_null() const noexcept
+    {
+        return m_type == value_t::null;
+    }
+
+    /// @brief return whether value is a boolean
+    /// @sa https://json.nlohmann.me/api/basic_json/is_boolean/
+    constexpr bool is_boolean() const noexcept
+    {
+        return m_type == value_t::boolean;
+    }
+
+    /// @brief return whether value is a number
+    /// @sa https://json.nlohmann.me/api/basic_json/is_number/
+    constexpr bool is_number() const noexcept
+    {
+        return is_number_integer() || is_number_float();
+    }
+
+    /// @brief return whether value is an integer number
+    /// @sa https://json.nlohmann.me/api/basic_json/is_number_integer/
+    constexpr bool is_number_integer() const noexcept
+    {
+        return m_type == value_t::number_integer || m_type == value_t::number_unsigned;
+    }
+
+    /// @brief return whether value is an unsigned integer number
+    /// @sa https://json.nlohmann.me/api/basic_json/is_number_unsigned/
+    constexpr bool is_number_unsigned() const noexcept
+    {
+        return m_type == value_t::number_unsigned;
+    }
+
+    /// @brief return whether value is a floating-point number
+    /// @sa https://json.nlohmann.me/api/basic_json/is_number_float/
+    constexpr bool is_number_float() const noexcept
+    {
+        return m_type == value_t::number_float;
+    }
+
+    /// @brief return whether value is an object
+    /// @sa https://json.nlohmann.me/api/basic_json/is_object/
+    constexpr bool is_object() const noexcept
+    {
+        return m_type == value_t::object;
+    }
+
+    /// @brief return whether value is an array
+    /// @sa https://json.nlohmann.me/api/basic_json/is_array/
+    constexpr bool is_array() const noexcept
+    {
+        return m_type == value_t::array;
+    }
+
+    /// @brief return whether value is a string
+    /// @sa https://json.nlohmann.me/api/basic_json/is_string/
+    constexpr bool is_string() const noexcept
+    {
+        return m_type == value_t::string;
+    }
+
+    /// @brief return whether value is a binary array
+    /// @sa https://json.nlohmann.me/api/basic_json/is_binary/
+    constexpr bool is_binary() const noexcept
+    {
+        return m_type == value_t::binary;
+    }
+
+    /// @brief return whether value is discarded
+    /// @sa https://json.nlohmann.me/api/basic_json/is_discarded/
+    constexpr bool is_discarded() const noexcept
+    {
+        return m_type == value_t::discarded;
+    }
+
+    /// @brief return the type of the JSON value (implicit)
+    /// @sa https://json.nlohmann.me/api/basic_json/operator_value_t/
+    constexpr operator value_t() const noexcept
+    {
+        return m_type;
+    }
+
+    /// @}
+
+  private:
+    //////////////////
+    // value access //
+    //////////////////
+
+    /// get a boolean (explicit)
+    boolean_t get_impl(boolean_t* /*unused*/) const
+    {
+        if (JSON_HEDLEY_LIKELY(is_boolean()))
+        {
+            return m_value.boolean;
+        }
+
+        JSON_THROW(type_error::create(302, detail::concat("type must be boolean, but is ", type_name()), this));
+    }
+
+    /// get a pointer to the value (object)
+    object_t* get_impl_ptr(object_t* /*unused*/) noexcept
+    {
+        return is_object() ? m_value.object : nullptr;
+    }
+
+    /// get a pointer to the value (object)
+    constexpr const object_t* get_impl_ptr(const object_t* /*unused*/) const noexcept
+    {
+        return is_object() ? m_value.object : nullptr;
+    }
+
+    /// get a pointer to the value (array)
+    array_t* get_impl_ptr(array_t* /*unused*/) noexcept
+    {
+        return is_array() ? m_value.array : nullptr;
+    }
+
+    /// get a pointer to the value (array)
+    constexpr const array_t* get_impl_ptr(const array_t* /*unused*/) const noexcept
+    {
+        return is_array() ? m_value.array : nullptr;
+    }
+
+    /// get a pointer to the value (string)
+    string_t* get_impl_ptr(string_t* /*unused*/) noexcept
+    {
+        return is_string() ? m_value.string : nullptr;
+    }
+
+    /// get a pointer to the value (string)
+    constexpr const string_t* get_impl_ptr(const string_t* /*unused*/) const noexcept
+    {
+        return is_string() ? m_value.string : nullptr;
+    }
+
+    /// get a pointer to the value (boolean)
+    boolean_t* get_impl_ptr(boolean_t* /*unused*/) noexcept
+    {
+        return is_boolean() ? &m_value.boolean : nullptr;
+    }
+
+    /// get a pointer to the value (boolean)
+    constexpr const boolean_t* get_impl_ptr(const boolean_t* /*unused*/) const noexcept
+    {
+        return is_boolean() ? &m_value.boolean : nullptr;
+    }
+
+    /// get a pointer to the value (integer number)
+    number_integer_t* get_impl_ptr(number_integer_t* /*unused*/) noexcept
+    {
+        return is_number_integer() ? &m_value.number_integer : nullptr;
+    }
+
+    /// get a pointer to the value (integer number)
+    constexpr const number_integer_t* get_impl_ptr(const number_integer_t* /*unused*/) const noexcept
+    {
+        return is_number_integer() ? &m_value.number_integer : nullptr;
+    }
+
+    /// get a pointer to the value (unsigned number)
+    number_unsigned_t* get_impl_ptr(number_unsigned_t* /*unused*/) noexcept
+    {
+        return is_number_unsigned() ? &m_value.number_unsigned : nullptr;
+    }
+
+    /// get a pointer to the value (unsigned number)
+    constexpr const number_unsigned_t* get_impl_ptr(const number_unsigned_t* /*unused*/) const noexcept
+    {
+        return is_number_unsigned() ? &m_value.number_unsigned : nullptr;
+    }
+
+    /// get a pointer to the value (floating-point number)
+    number_float_t* get_impl_ptr(number_float_t* /*unused*/) noexcept
+    {
+        return is_number_float() ? &m_value.number_float : nullptr;
+    }
+
+    /// get a pointer to the value (floating-point number)
+    constexpr const number_float_t* get_impl_ptr(const number_float_t* /*unused*/) const noexcept
+    {
+        return is_number_float() ? &m_value.number_float : nullptr;
+    }
+
+    /// get a pointer to the value (binary)
+    binary_t* get_impl_ptr(binary_t* /*unused*/) noexcept
+    {
+        return is_binary() ? m_value.binary : nullptr;
+    }
+
+    /// get a pointer to the value (binary)
+    constexpr const binary_t* get_impl_ptr(const binary_t* /*unused*/) const noexcept
+    {
+        return is_binary() ? m_value.binary : nullptr;
+    }
+
+    /*!
+    @brief helper function to implement get_ref()
+
+    This function helps to implement get_ref() without code duplication for
+    const and non-const overloads
+
+    @tparam ThisType will be deduced as `basic_json` or `const basic_json`
+
+    @throw type_error.303 if ReferenceType does not match underlying value
+    type of the current JSON
+    */
+    template<typename ReferenceType, typename ThisType>
+    static ReferenceType get_ref_impl(ThisType& obj)
+    {
+        // delegate the call to get_ptr<>()
+        auto* ptr = obj.template get_ptr<typename std::add_pointer<ReferenceType>::type>();
+
+        if (JSON_HEDLEY_LIKELY(ptr != nullptr))
+        {
+            return *ptr;
+        }
+
+        JSON_THROW(type_error::create(303, detail::concat("incompatible ReferenceType for get_ref, actual type is ", obj.type_name()), &obj));
+    }
+
+  public:
+    /// @name value access
+    /// Direct access to the stored value of a JSON value.
+    /// @{
+
+    /// @brief get a pointer value (implicit)
+    /// @sa https://json.nlohmann.me/api/basic_json/get_ptr/
+    template<typename PointerType, typename std::enable_if<
+                 std::is_pointer<PointerType>::value, int>::type = 0>
+    auto get_ptr() noexcept -> decltype(std::declval<basic_json_t&>().get_impl_ptr(std::declval<PointerType>()))
+    {
+        // delegate the call to get_impl_ptr<>()
+        return get_impl_ptr(static_cast<PointerType>(nullptr));
+    }
+
+    /// @brief get a pointer value (implicit)
+    /// @sa https://json.nlohmann.me/api/basic_json/get_ptr/
+    template < typename PointerType, typename std::enable_if <
+                   std::is_pointer<PointerType>::value&&
+                   std::is_const<typename std::remove_pointer<PointerType>::type>::value, int >::type = 0 >
+    constexpr auto get_ptr() const noexcept -> decltype(std::declval<const basic_json_t&>().get_impl_ptr(std::declval<PointerType>()))
+    {
+        // delegate the call to get_impl_ptr<>() const
+        return get_impl_ptr(static_cast<PointerType>(nullptr));
+    }
+
+  private:
+    /*!
+    @brief get a value (explicit)
+
+    Explicit type conversion between the JSON value and a compatible value
+    which is [CopyConstructible](https://en.cppreference.com/w/cpp/named_req/CopyConstructible)
+    and [DefaultConstructible](https://en.cppreference.com/w/cpp/named_req/DefaultConstructible).
+    The value is converted by calling the @ref json_serializer<ValueType>
+    `from_json()` method.
+
+    The function is equivalent to executing
+    @code {.cpp}
+    ValueType ret;
+    JSONSerializer<ValueType>::from_json(*this, ret);
+    return ret;
+    @endcode
+
+    This overloads is chosen if:
+    - @a ValueType is not @ref basic_json,
+    - @ref json_serializer<ValueType> has a `from_json()` method of the form
+      `void from_json(const basic_json&, ValueType&)`, and
+    - @ref json_serializer<ValueType> does not have a `from_json()` method of
+      the form `ValueType from_json(const basic_json&)`
+
+    @tparam ValueType the returned value type
+
+    @return copy of the JSON value, converted to @a ValueType
+
+    @throw what @ref json_serializer<ValueType> `from_json()` method throws
+
+    @liveexample{The example below shows several conversions from JSON values
+    to other types. There a few things to note: (1) Floating-point numbers can
+    be converted to integers\, (2) A JSON array can be converted to a standard
+    `std::vector<short>`\, (3) A JSON object can be converted to C++
+    associative containers such as `std::unordered_map<std::string\,
+    json>`.,get__ValueType_const}
+
+    @since version 2.1.0
+    */
+    template < typename ValueType,
+               detail::enable_if_t <
+                   detail::is_default_constructible<ValueType>::value&&
+                   detail::has_from_json<basic_json_t, ValueType>::value,
+                   int > = 0 >
+    ValueType get_impl(detail::priority_tag<0> /*unused*/) const noexcept(noexcept(
+                JSONSerializer<ValueType>::from_json(std::declval<const basic_json_t&>(), std::declval<ValueType&>())))
+    {
+        auto ret = ValueType();
+        JSONSerializer<ValueType>::from_json(*this, ret);
+        return ret;
+    }
+
+    /*!
+    @brief get a value (explicit); special case
+
+    Explicit type conversion between the JSON value and a compatible value
+    which is **not** [CopyConstructible](https://en.cppreference.com/w/cpp/named_req/CopyConstructible)
+    and **not** [DefaultConstructible](https://en.cppreference.com/w/cpp/named_req/DefaultConstructible).
+    The value is converted by calling the @ref json_serializer<ValueType>
+    `from_json()` method.
+
+    The function is equivalent to executing
+    @code {.cpp}
+    return JSONSerializer<ValueType>::from_json(*this);
+    @endcode
+
+    This overloads is chosen if:
+    - @a ValueType is not @ref basic_json and
+    - @ref json_serializer<ValueType> has a `from_json()` method of the form
+      `ValueType from_json(const basic_json&)`
+
+    @note If @ref json_serializer<ValueType> has both overloads of
+    `from_json()`, this one is chosen.
+
+    @tparam ValueType the returned value type
+
+    @return copy of the JSON value, converted to @a ValueType
+
+    @throw what @ref json_serializer<ValueType> `from_json()` method throws
+
+    @since version 2.1.0
+    */
+    template < typename ValueType,
+               detail::enable_if_t <
+                   detail::has_non_default_from_json<basic_json_t, ValueType>::value,
+                   int > = 0 >
+    ValueType get_impl(detail::priority_tag<1> /*unused*/) const noexcept(noexcept(
+                JSONSerializer<ValueType>::from_json(std::declval<const basic_json_t&>())))
+    {
+        return JSONSerializer<ValueType>::from_json(*this);
+    }
+
+    /*!
+    @brief get special-case overload
+
+    This overloads converts the current @ref basic_json in a different
+    @ref basic_json type
+
+    @tparam BasicJsonType == @ref basic_json
+
+    @return a copy of *this, converted into @a BasicJsonType
+
+    @complexity Depending on the implementation of the called `from_json()`
+                method.
+
+    @since version 3.2.0
+    */
+    template < typename BasicJsonType,
+               detail::enable_if_t <
+                   detail::is_basic_json<BasicJsonType>::value,
+                   int > = 0 >
+    BasicJsonType get_impl(detail::priority_tag<2> /*unused*/) const
+    {
+        return *this;
+    }
+
+    /*!
+    @brief get special-case overload
+
+    This overloads avoids a lot of template boilerplate, it can be seen as the
+    identity method
+
+    @tparam BasicJsonType == @ref basic_json
+
+    @return a copy of *this
+
+    @complexity Constant.
+
+    @since version 2.1.0
+    */
+    template<typename BasicJsonType,
+             detail::enable_if_t<
+                 std::is_same<BasicJsonType, basic_json_t>::value,
+                 int> = 0>
+    basic_json get_impl(detail::priority_tag<3> /*unused*/) const
+    {
+        return *this;
+    }
+
+    /*!
+    @brief get a pointer value (explicit)
+    @copydoc get()
+    */
+    template<typename PointerType,
+             detail::enable_if_t<
+                 std::is_pointer<PointerType>::value,
+                 int> = 0>
+    constexpr auto get_impl(detail::priority_tag<4> /*unused*/) const noexcept
+    -> decltype(std::declval<const basic_json_t&>().template get_ptr<PointerType>())
+    {
+        // delegate the call to get_ptr
+        return get_ptr<PointerType>();
+    }
+
+  public:
+    /*!
+    @brief get a (pointer) value (explicit)
+
+    Performs explicit type conversion between the JSON value and a compatible value if required.
+
+    - If the requested type is a pointer to the internally stored JSON value that pointer is returned.
+    No copies are made.
+
+    - If the requested type is the current @ref basic_json, or a different @ref basic_json convertible
+    from the current @ref basic_json.
+
+    - Otherwise the value is converted by calling the @ref json_serializer<ValueType> `from_json()`
+    method.
+
+    @tparam ValueTypeCV the provided value type
+    @tparam ValueType the returned value type
+
+    @return copy of the JSON value, converted to @tparam ValueType if necessary
+
+    @throw what @ref json_serializer<ValueType> `from_json()` method throws if conversion is required
+
+    @since version 2.1.0
+    */
+    template < typename ValueTypeCV, typename ValueType = detail::uncvref_t<ValueTypeCV>>
+#if defined(JSON_HAS_CPP_14)
+    constexpr
+#endif
+    auto get() const noexcept(
+    noexcept(std::declval<const basic_json_t&>().template get_impl<ValueType>(detail::priority_tag<4> {})))
+    -> decltype(std::declval<const basic_json_t&>().template get_impl<ValueType>(detail::priority_tag<4> {}))
+    {
+        // we cannot static_assert on ValueTypeCV being non-const, because
+        // there is support for get<const basic_json_t>(), which is why we
+        // still need the uncvref
+        static_assert(!std::is_reference<ValueTypeCV>::value,
+                      "get() cannot be used with reference types, you might want to use get_ref()");
+        return get_impl<ValueType>(detail::priority_tag<4> {});
+    }
+
+    /*!
+    @brief get a pointer value (explicit)
+
+    Explicit pointer access to the internally stored JSON value. No copies are
+    made.
+
+    @warning The pointer becomes invalid if the underlying JSON object
+    changes.
+
+    @tparam PointerType pointer type; must be a pointer to @ref array_t, @ref
+    object_t, @ref string_t, @ref boolean_t, @ref number_integer_t,
+    @ref number_unsigned_t, or @ref number_float_t.
+
+    @return pointer to the internally stored JSON value if the requested
+    pointer type @a PointerType fits to the JSON value; `nullptr` otherwise
+
+    @complexity Constant.
+
+    @liveexample{The example below shows how pointers to internal values of a
+    JSON value can be requested. Note that no type conversions are made and a
+    `nullptr` is returned if the value and the requested pointer type does not
+    match.,get__PointerType}
+
+    @sa see @ref get_ptr() for explicit pointer-member access
+
+    @since version 1.0.0
+    */
+    template<typename PointerType, typename std::enable_if<
+                 std::is_pointer<PointerType>::value, int>::type = 0>
+    auto get() noexcept -> decltype(std::declval<basic_json_t&>().template get_ptr<PointerType>())
+    {
+        // delegate the call to get_ptr
+        return get_ptr<PointerType>();
+    }
+
+    /// @brief get a value (explicit)
+    /// @sa https://json.nlohmann.me/api/basic_json/get_to/
+    template < typename ValueType,
+               detail::enable_if_t <
+                   !detail::is_basic_json<ValueType>::value&&
+                   detail::has_from_json<basic_json_t, ValueType>::value,
+                   int > = 0 >
+    ValueType & get_to(ValueType& v) const noexcept(noexcept(
+                JSONSerializer<ValueType>::from_json(std::declval<const basic_json_t&>(), v)))
+    {
+        JSONSerializer<ValueType>::from_json(*this, v);
+        return v;
+    }
+
+    // specialization to allow calling get_to with a basic_json value
+    // see https://github.com/nlohmann/json/issues/2175
+    template<typename ValueType,
+             detail::enable_if_t <
+                 detail::is_basic_json<ValueType>::value,
+                 int> = 0>
+    ValueType & get_to(ValueType& v) const
+    {
+        v = *this;
+        return v;
+    }
+
+    template <
+        typename T, std::size_t N,
+        typename Array = T (&)[N], // NOLINT(cppcoreguidelines-avoid-c-arrays,hicpp-avoid-c-arrays,modernize-avoid-c-arrays)
+        detail::enable_if_t <
+            detail::has_from_json<basic_json_t, Array>::value, int > = 0 >
+    Array get_to(T (&v)[N]) const // NOLINT(cppcoreguidelines-avoid-c-arrays,hicpp-avoid-c-arrays,modernize-avoid-c-arrays)
+    noexcept(noexcept(JSONSerializer<Array>::from_json(
+                          std::declval<const basic_json_t&>(), v)))
+    {
+        JSONSerializer<Array>::from_json(*this, v);
+        return v;
+    }
+
+    /// @brief get a reference value (implicit)
+    /// @sa https://json.nlohmann.me/api/basic_json/get_ref/
+    template<typename ReferenceType, typename std::enable_if<
+                 std::is_reference<ReferenceType>::value, int>::type = 0>
+    ReferenceType get_ref()
+    {
+        // delegate call to get_ref_impl
+        return get_ref_impl<ReferenceType>(*this);
+    }
+
+    /// @brief get a reference value (implicit)
+    /// @sa https://json.nlohmann.me/api/basic_json/get_ref/
+    template < typename ReferenceType, typename std::enable_if <
+                   std::is_reference<ReferenceType>::value&&
+                   std::is_const<typename std::remove_reference<ReferenceType>::type>::value, int >::type = 0 >
+    ReferenceType get_ref() const
+    {
+        // delegate call to get_ref_impl
+        return get_ref_impl<ReferenceType>(*this);
+    }
+
+    /*!
+    @brief get a value (implicit)
+
+    Implicit type conversion between the JSON value and a compatible value.
+    The call is realized by calling @ref get() const.
+
+    @tparam ValueType non-pointer type compatible to the JSON value, for
+    instance `int` for JSON integer numbers, `bool` for JSON booleans, or
+    `std::vector` types for JSON arrays. The character type of @ref string_t
+    as well as an initializer list of this type is excluded to avoid
+    ambiguities as these types implicitly convert to `std::string`.
+
+    @return copy of the JSON value, converted to type @a ValueType
+
+    @throw type_error.302 in case passed type @a ValueType is incompatible
+    to the JSON value type (e.g., the JSON value is of type boolean, but a
+    string is requested); see example below
+
+    @complexity Linear in the size of the JSON value.
+
+    @liveexample{The example below shows several conversions from JSON values
+    to other types. There a few things to note: (1) Floating-point numbers can
+    be converted to integers\, (2) A JSON array can be converted to a standard
+    `std::vector<short>`\, (3) A JSON object can be converted to C++
+    associative containers such as `std::unordered_map<std::string\,
+    json>`.,operator__ValueType}
+
+    @since version 1.0.0
+    */
+    template < typename ValueType, typename std::enable_if <
+                   detail::conjunction <
+                       detail::negation<std::is_pointer<ValueType>>,
+                       detail::negation<std::is_same<ValueType, std::nullptr_t>>,
+                       detail::negation<std::is_same<ValueType, detail::json_ref<basic_json>>>,
+                                        detail::negation<std::is_same<ValueType, typename string_t::value_type>>,
+                                        detail::negation<detail::is_basic_json<ValueType>>,
+                                        detail::negation<std::is_same<ValueType, std::initializer_list<typename string_t::value_type>>>,
+#if defined(JSON_HAS_CPP_17) && (defined(__GNUC__) || (defined(_MSC_VER) && _MSC_VER >= 1910 && _MSC_VER <= 1914))
+                                                detail::negation<std::is_same<ValueType, std::string_view>>,
+#endif
+#if defined(JSON_HAS_CPP_17)
+                                                detail::negation<std::is_same<ValueType, std::any>>,
+#endif
+                                                detail::is_detected_lazy<detail::get_template_function, const basic_json_t&, ValueType>
+                                                >::value, int >::type = 0 >
+                                        JSON_EXPLICIT operator ValueType() const
+    {
+        // delegate the call to get<>() const
+        return get<ValueType>();
+    }
+
+    /// @brief get a binary value
+    /// @sa https://json.nlohmann.me/api/basic_json/get_binary/
+    binary_t& get_binary()
+    {
+        if (!is_binary())
+        {
+            JSON_THROW(type_error::create(302, detail::concat("type must be binary, but is ", type_name()), this));
+        }
+
+        return *get_ptr<binary_t*>();
+    }
+
+    /// @brief get a binary value
+    /// @sa https://json.nlohmann.me/api/basic_json/get_binary/
+    const binary_t& get_binary() const
+    {
+        if (!is_binary())
+        {
+            JSON_THROW(type_error::create(302, detail::concat("type must be binary, but is ", type_name()), this));
+        }
+
+        return *get_ptr<const binary_t*>();
+    }
+
+    /// @}
+
+
+    ////////////////////
+    // element access //
+    ////////////////////
+
+    /// @name element access
+    /// Access to the JSON value.
+    /// @{
+
+    /// @brief access specified array element with bounds checking
+    /// @sa https://json.nlohmann.me/api/basic_json/at/
+    reference at(size_type idx)
+    {
+        // at only works for arrays
+        if (JSON_HEDLEY_LIKELY(is_array()))
+        {
+            JSON_TRY
+            {
+                return set_parent(m_value.array->at(idx));
+            }
+            JSON_CATCH (std::out_of_range&)
+            {
+                // create better exception explanation
+                JSON_THROW(out_of_range::create(401, detail::concat("array index ", std::to_string(idx), " is out of range"), this));
+            }
+        }
+        else
+        {
+            JSON_THROW(type_error::create(304, detail::concat("cannot use at() with ", type_name()), this));
+        }
+    }
+
+    /// @brief access specified array element with bounds checking
+    /// @sa https://json.nlohmann.me/api/basic_json/at/
+    const_reference at(size_type idx) const
+    {
+        // at only works for arrays
+        if (JSON_HEDLEY_LIKELY(is_array()))
+        {
+            JSON_TRY
+            {
+                return m_value.array->at(idx);
+            }
+            JSON_CATCH (std::out_of_range&)
+            {
+                // create better exception explanation
+                JSON_THROW(out_of_range::create(401, detail::concat("array index ", std::to_string(idx), " is out of range"), this));
+            }
+        }
+        else
+        {
+            JSON_THROW(type_error::create(304, detail::concat("cannot use at() with ", type_name()), this));
+        }
+    }
+
+    /// @brief access specified object element with bounds checking
+    /// @sa https://json.nlohmann.me/api/basic_json/at/
+    reference at(const typename object_t::key_type& key)
+    {
+        // at only works for objects
+        if (JSON_HEDLEY_UNLIKELY(!is_object()))
+        {
+            JSON_THROW(type_error::create(304, detail::concat("cannot use at() with ", type_name()), this));
+        }
+
+        auto it = m_value.object->find(key);
+        if (it == m_value.object->end())
+        {
+            JSON_THROW(out_of_range::create(403, detail::concat("key '", key, "' not found"), this));
+        }
+        return set_parent(it->second);
+    }
+
+    /// @brief access specified object element with bounds checking
+    /// @sa https://json.nlohmann.me/api/basic_json/at/
+    template<class KeyType, detail::enable_if_t<
+                 detail::is_usable_as_basic_json_key_type<basic_json_t, KeyType>::value, int> = 0>
+    reference at(KeyType && key)
+    {
+        // at only works for objects
+        if (JSON_HEDLEY_UNLIKELY(!is_object()))
+        {
+            JSON_THROW(type_error::create(304, detail::concat("cannot use at() with ", type_name()), this));
+        }
+
+        auto it = m_value.object->find(std::forward<KeyType>(key));
+        if (it == m_value.object->end())
+        {
+            JSON_THROW(out_of_range::create(403, detail::concat("key '", string_t(std::forward<KeyType>(key)), "' not found"), this));
+        }
+        return set_parent(it->second);
+    }
+
+    /// @brief access specified object element with bounds checking
+    /// @sa https://json.nlohmann.me/api/basic_json/at/
+    const_reference at(const typename object_t::key_type& key) const
+    {
+        // at only works for objects
+        if (JSON_HEDLEY_UNLIKELY(!is_object()))
+        {
+            JSON_THROW(type_error::create(304, detail::concat("cannot use at() with ", type_name()), this));
+        }
+
+        auto it = m_value.object->find(key);
+        if (it == m_value.object->end())
+        {
+            JSON_THROW(out_of_range::create(403, detail::concat("key '", key, "' not found"), this));
+        }
+        return it->second;
+    }
+
+    /// @brief access specified object element with bounds checking
+    /// @sa https://json.nlohmann.me/api/basic_json/at/
+    template<class KeyType, detail::enable_if_t<
+                 detail::is_usable_as_basic_json_key_type<basic_json_t, KeyType>::value, int> = 0>
+    const_reference at(KeyType && key) const
+    {
+        // at only works for objects
+        if (JSON_HEDLEY_UNLIKELY(!is_object()))
+        {
+            JSON_THROW(type_error::create(304, detail::concat("cannot use at() with ", type_name()), this));
+        }
+
+        auto it = m_value.object->find(std::forward<KeyType>(key));
+        if (it == m_value.object->end())
+        {
+            JSON_THROW(out_of_range::create(403, detail::concat("key '", string_t(std::forward<KeyType>(key)), "' not found"), this));
+        }
+        return it->second;
+    }
+
+    /// @brief access specified array element
+    /// @sa https://json.nlohmann.me/api/basic_json/operator%5B%5D/
+    reference operator[](size_type idx)
+    {
+        // implicitly convert null value to an empty array
+        if (is_null())
+        {
+            m_type = value_t::array;
+            m_value.array = create<array_t>();
+            assert_invariant();
+        }
+
+        // operator[] only works for arrays
+        if (JSON_HEDLEY_LIKELY(is_array()))
+        {
+            // fill up array with null values if given idx is outside range
+            if (idx >= m_value.array->size())
+            {
+#if JSON_DIAGNOSTICS
+                // remember array size & capacity before resizing
+                const auto old_size = m_value.array->size();
+                const auto old_capacity = m_value.array->capacity();
+#endif
+                m_value.array->resize(idx + 1);
+
+#if JSON_DIAGNOSTICS
+                if (JSON_HEDLEY_UNLIKELY(m_value.array->capacity() != old_capacity))
+                {
+                    // capacity has changed: update all parents
+                    set_parents();
+                }
+                else
+                {
+                    // set parent for values added above
+                    set_parents(begin() + static_cast<typename iterator::difference_type>(old_size), static_cast<typename iterator::difference_type>(idx + 1 - old_size));
+                }
+#endif
+                assert_invariant();
+            }
+
+            return m_value.array->operator[](idx);
+        }
+
+        JSON_THROW(type_error::create(305, detail::concat("cannot use operator[] with a numeric argument with ", type_name()), this));
+    }
+
+    /// @brief access specified array element
+    /// @sa https://json.nlohmann.me/api/basic_json/operator%5B%5D/
+    const_reference operator[](size_type idx) const
+    {
+        // const operator[] only works for arrays
+        if (JSON_HEDLEY_LIKELY(is_array()))
+        {
+            return m_value.array->operator[](idx);
+        }
+
+        JSON_THROW(type_error::create(305, detail::concat("cannot use operator[] with a numeric argument with ", type_name()), this));
+    }
+
+    /// @brief access specified object element
+    /// @sa https://json.nlohmann.me/api/basic_json/operator%5B%5D/
+    reference operator[](typename object_t::key_type key)
+    {
+        // implicitly convert null value to an empty object
+        if (is_null())
+        {
+            m_type = value_t::object;
+            m_value.object = create<object_t>();
+            assert_invariant();
+        }
+
+        // operator[] only works for objects
+        if (JSON_HEDLEY_LIKELY(is_object()))
+        {
+            auto result = m_value.object->emplace(std::move(key), nullptr);
+            return set_parent(result.first->second);
+        }
+
+        JSON_THROW(type_error::create(305, detail::concat("cannot use operator[] with a string argument with ", type_name()), this));
+    }
+
+    /// @brief access specified object element
+    /// @sa https://json.nlohmann.me/api/basic_json/operator%5B%5D/
+    const_reference operator[](const typename object_t::key_type& key) const
+    {
+        // const operator[] only works for objects
+        if (JSON_HEDLEY_LIKELY(is_object()))
+        {
+            auto it = m_value.object->find(key);
+            JSON_ASSERT(it != m_value.object->end());
+            return it->second;
+        }
+
+        JSON_THROW(type_error::create(305, detail::concat("cannot use operator[] with a string argument with ", type_name()), this));
+    }
+
+    // these two functions resolve a (const) char * ambiguity affecting Clang and MSVC
+    // (they seemingly cannot be constrained to resolve the ambiguity)
+    template<typename T>
+    reference operator[](T* key)
+    {
+        return operator[](typename object_t::key_type(key));
+    }
+
+    template<typename T>
+    const_reference operator[](T* key) const
+    {
+        return operator[](typename object_t::key_type(key));
+    }
+
+    /// @brief access specified object element
+    /// @sa https://json.nlohmann.me/api/basic_json/operator%5B%5D/
+    template<class KeyType, detail::enable_if_t<
+                 detail::is_usable_as_basic_json_key_type<basic_json_t, KeyType>::value, int > = 0 >
+    reference operator[](KeyType && key)
+    {
+        // implicitly convert null value to an empty object
+        if (is_null())
+        {
+            m_type = value_t::object;
+            m_value.object = create<object_t>();
+            assert_invariant();
+        }
+
+        // operator[] only works for objects
+        if (JSON_HEDLEY_LIKELY(is_object()))
+        {
+            auto result = m_value.object->emplace(std::forward<KeyType>(key), nullptr);
+            return set_parent(result.first->second);
+        }
+
+        JSON_THROW(type_error::create(305, detail::concat("cannot use operator[] with a string argument with ", type_name()), this));
+    }
+
+    /// @brief access specified object element
+    /// @sa https://json.nlohmann.me/api/basic_json/operator%5B%5D/
+    template<class KeyType, detail::enable_if_t<
+                 detail::is_usable_as_basic_json_key_type<basic_json_t, KeyType>::value, int > = 0 >
+    const_reference operator[](KeyType && key) const
+    {
+        // const operator[] only works for objects
+        if (JSON_HEDLEY_LIKELY(is_object()))
+        {
+            auto it = m_value.object->find(std::forward<KeyType>(key));
+            JSON_ASSERT(it != m_value.object->end());
+            return it->second;
+        }
+
+        JSON_THROW(type_error::create(305, detail::concat("cannot use operator[] with a string argument with ", type_name()), this));
+    }
+
+  private:
+    template<typename KeyType>
+    using is_comparable_with_object_key = detail::is_comparable <
+        object_comparator_t, const typename object_t::key_type&, KeyType >;
+
+    template<typename ValueType>
+    using value_return_type = std::conditional <
+        detail::is_c_string_uncvref<ValueType>::value,
+        string_t, typename std::decay<ValueType>::type >;
+
+  public:
+    /// @brief access specified object element with default value
+    /// @sa https://json.nlohmann.me/api/basic_json/value/
+    template < class ValueType, detail::enable_if_t <
+                   !detail::is_transparent<object_comparator_t>::value
+                   && detail::is_getable<basic_json_t, ValueType>::value
+                   && !std::is_same<value_t, detail::uncvref_t<ValueType>>::value, int > = 0 >
+    ValueType value(const typename object_t::key_type& key, const ValueType& default_value) const
+    {
+        // value only works for objects
+        if (JSON_HEDLEY_LIKELY(is_object()))
+        {
+            // if key is found, return value and given default value otherwise
+            const auto it = find(key);
+            if (it != end())
+            {
+                return it->template get<ValueType>();
+            }
+
+            return default_value;
+        }
+
+        JSON_THROW(type_error::create(306, detail::concat("cannot use value() with ", type_name()), this));
+    }
+
+    /// @brief access specified object element with default value
+    /// @sa https://json.nlohmann.me/api/basic_json/value/
+    template < class ValueType, class ReturnType = typename value_return_type<ValueType>::type,
+               detail::enable_if_t <
+                   !detail::is_transparent<object_comparator_t>::value
+                   && detail::is_getable<basic_json_t, ReturnType>::value
+                   && !std::is_same<value_t, detail::uncvref_t<ValueType>>::value, int > = 0 >
+    ReturnType value(const typename object_t::key_type& key, ValueType && default_value) const
+    {
+        // value only works for objects
+        if (JSON_HEDLEY_LIKELY(is_object()))
+        {
+            // if key is found, return value and given default value otherwise
+            const auto it = find(key);
+            if (it != end())
+            {
+                return it->template get<ReturnType>();
+            }
+
+            return std::forward<ValueType>(default_value);
+        }
+
+        JSON_THROW(type_error::create(306, detail::concat("cannot use value() with ", type_name()), this));
+    }
+
+    /// @brief access specified object element with default value
+    /// @sa https://json.nlohmann.me/api/basic_json/value/
+    template < class ValueType, class KeyType, detail::enable_if_t <
+                   detail::is_transparent<object_comparator_t>::value
+                   && !detail::is_json_pointer<KeyType>::value
+                   && is_comparable_with_object_key<KeyType>::value
+                   && detail::is_getable<basic_json_t, ValueType>::value
+                   && !std::is_same<value_t, detail::uncvref_t<ValueType>>::value, int > = 0 >
+    ValueType value(KeyType && key, const ValueType& default_value) const
+    {
+        // value only works for objects
+        if (JSON_HEDLEY_LIKELY(is_object()))
+        {
+            // if key is found, return value and given default value otherwise
+            const auto it = find(std::forward<KeyType>(key));
+            if (it != end())
+            {
+                return it->template get<ValueType>();
+            }
+
+            return default_value;
+        }
+
+        JSON_THROW(type_error::create(306, detail::concat("cannot use value() with ", type_name()), this));
+    }
+
+    /// @brief access specified object element via JSON Pointer with default value
+    /// @sa https://json.nlohmann.me/api/basic_json/value/
+    template < class ValueType, class KeyType, class ReturnType = typename value_return_type<ValueType>::type,
+               detail::enable_if_t <
+                   detail::is_transparent<object_comparator_t>::value
+                   && !detail::is_json_pointer<KeyType>::value
+                   && is_comparable_with_object_key<KeyType>::value
+                   && detail::is_getable<basic_json_t, ReturnType>::value
+                   && !std::is_same<value_t, detail::uncvref_t<ValueType>>::value, int > = 0 >
+    ReturnType value(KeyType && key, ValueType && default_value) const
+    {
+        // value only works for objects
+        if (JSON_HEDLEY_LIKELY(is_object()))
+        {
+            // if key is found, return value and given default value otherwise
+            const auto it = find(std::forward<KeyType>(key));
+            if (it != end())
+            {
+                return it->template get<ReturnType>();
+            }
+
+            return std::forward<ValueType>(default_value);
+        }
+
+        JSON_THROW(type_error::create(306, detail::concat("cannot use value() with ", type_name()), this));
+    }
+
+    /// @brief access specified object element via JSON Pointer with default value
+    /// @sa https://json.nlohmann.me/api/basic_json/value/
+    template < class ValueType, detail::enable_if_t <
+                   detail::is_getable<basic_json_t, ValueType>::value
+                   && !std::is_same<value_t, detail::uncvref_t<ValueType>>::value, int > = 0 >
+    ValueType value(const json_pointer& ptr, const ValueType& default_value) const
+    {
+        // value only works for objects
+        if (JSON_HEDLEY_LIKELY(is_object()))
+        {
+            // if pointer resolves a value, return it or use default value
+            JSON_TRY
+            {
+                return ptr.get_checked(this).template get<ValueType>();
+            }
+            JSON_INTERNAL_CATCH (out_of_range&)
+            {
+                return default_value;
+            }
+        }
+
+        JSON_THROW(type_error::create(306, detail::concat("cannot use value() with ", type_name()), this));
+    }
+
+    /// @brief access specified object element via JSON Pointer with default value
+    /// @sa https://json.nlohmann.me/api/basic_json/value/
+    template < class ValueType, class ReturnType = typename value_return_type<ValueType>::type,
+               detail::enable_if_t <
+                   detail::is_getable<basic_json_t, ReturnType>::value
+                   && !std::is_same<value_t, detail::uncvref_t<ValueType>>::value, int > = 0 >
+    ReturnType value(const json_pointer& ptr, ValueType && default_value) const
+    {
+        // value only works for objects
+        if (JSON_HEDLEY_LIKELY(is_object()))
+        {
+            // if pointer resolves a value, return it or use default value
+            JSON_TRY
+            {
+                return ptr.get_checked(this).template get<ReturnType>();
+            }
+            JSON_INTERNAL_CATCH (out_of_range&)
+            {
+                return std::forward<ValueType>(default_value);
+            }
+        }
+
+        JSON_THROW(type_error::create(306, detail::concat("cannot use value() with ", type_name()), this));
+    }
+
+    template < class ValueType, class BasicJsonType, detail::enable_if_t <
+                   detail::is_basic_json<BasicJsonType>::value
+                   && detail::is_getable<basic_json_t, ValueType>::value
+                   && !std::is_same<value_t, detail::uncvref_t<ValueType>>::value, int > = 0 >
+    JSON_HEDLEY_DEPRECATED_FOR(3.11.0, basic_json::json_pointer or nlohmann::json_pointer<basic_json::string_t>) // NOLINT(readability/alt_tokens)
+    ValueType value(const ::nlohmann::json_pointer<BasicJsonType>& ptr, const ValueType& default_value) const
+    {
+        return value(ptr.convert(), default_value);
+    }
+
+    template < class ValueType, class BasicJsonType, class ReturnType = typename value_return_type<ValueType>::type,
+               detail::enable_if_t <
+                   detail::is_basic_json<BasicJsonType>::value
+                   && detail::is_getable<basic_json_t, ReturnType>::value
+                   && !std::is_same<value_t, detail::uncvref_t<ValueType>>::value, int > = 0 >
+    JSON_HEDLEY_DEPRECATED_FOR(3.11.0, basic_json::json_pointer or nlohmann::json_pointer<basic_json::string_t>) // NOLINT(readability/alt_tokens)
+    ReturnType value(const ::nlohmann::json_pointer<BasicJsonType>& ptr, ValueType && default_value) const
+    {
+        return value(ptr.convert(), std::forward<ValueType>(default_value));
+    }
+
+    /// @brief access the first element
+    /// @sa https://json.nlohmann.me/api/basic_json/front/
+    reference front()
+    {
+        return *begin();
+    }
+
+    /// @brief access the first element
+    /// @sa https://json.nlohmann.me/api/basic_json/front/
+    const_reference front() const
+    {
+        return *cbegin();
+    }
+
+    /// @brief access the last element
+    /// @sa https://json.nlohmann.me/api/basic_json/back/
+    reference back()
+    {
+        auto tmp = end();
+        --tmp;
+        return *tmp;
+    }
+
+    /// @brief access the last element
+    /// @sa https://json.nlohmann.me/api/basic_json/back/
+    const_reference back() const
+    {
+        auto tmp = cend();
+        --tmp;
+        return *tmp;
+    }
+
+    /// @brief remove element given an iterator
+    /// @sa https://json.nlohmann.me/api/basic_json/erase/
+    template < class IteratorType, detail::enable_if_t <
+                   std::is_same<IteratorType, typename basic_json_t::iterator>::value ||
+                   std::is_same<IteratorType, typename basic_json_t::const_iterator>::value, int > = 0 >
+    IteratorType erase(IteratorType pos)
+    {
+        // make sure iterator fits the current value
+        if (JSON_HEDLEY_UNLIKELY(this != pos.m_object))
+        {
+            JSON_THROW(invalid_iterator::create(202, "iterator does not fit current value", this));
+        }
+
+        IteratorType result = end();
+
+        switch (m_type)
+        {
+            case value_t::boolean:
+            case value_t::number_float:
+            case value_t::number_integer:
+            case value_t::number_unsigned:
+            case value_t::string:
+            case value_t::binary:
+            {
+                if (JSON_HEDLEY_UNLIKELY(!pos.m_it.primitive_iterator.is_begin()))
+                {
+                    JSON_THROW(invalid_iterator::create(205, "iterator out of range", this));
+                }
+
+                if (is_string())
+                {
+                    AllocatorType<string_t> alloc;
+                    std::allocator_traits<decltype(alloc)>::destroy(alloc, m_value.string);
+                    std::allocator_traits<decltype(alloc)>::deallocate(alloc, m_value.string, 1);
+                    m_value.string = nullptr;
+                }
+                else if (is_binary())
+                {
+                    AllocatorType<binary_t> alloc;
+                    std::allocator_traits<decltype(alloc)>::destroy(alloc, m_value.binary);
+                    std::allocator_traits<decltype(alloc)>::deallocate(alloc, m_value.binary, 1);
+                    m_value.binary = nullptr;
+                }
+
+                m_type = value_t::null;
+                assert_invariant();
+                break;
+            }
+
+            case value_t::object:
+            {
+                result.m_it.object_iterator = m_value.object->erase(pos.m_it.object_iterator);
+                break;
+            }
+
+            case value_t::array:
+            {
+                result.m_it.array_iterator = m_value.array->erase(pos.m_it.array_iterator);
+                break;
+            }
+
+            case value_t::null:
+            case value_t::discarded:
+            default:
+                JSON_THROW(type_error::create(307, detail::concat("cannot use erase() with ", type_name()), this));
+        }
+
+        return result;
+    }
+
+    /// @brief remove elements given an iterator range
+    /// @sa https://json.nlohmann.me/api/basic_json/erase/
+    template < class IteratorType, detail::enable_if_t <
+                   std::is_same<IteratorType, typename basic_json_t::iterator>::value ||
+                   std::is_same<IteratorType, typename basic_json_t::const_iterator>::value, int > = 0 >
+    IteratorType erase(IteratorType first, IteratorType last)
+    {
+        // make sure iterator fits the current value
+        if (JSON_HEDLEY_UNLIKELY(this != first.m_object || this != last.m_object))
+        {
+            JSON_THROW(invalid_iterator::create(203, "iterators do not fit current value", this));
+        }
+
+        IteratorType result = end();
+
+        switch (m_type)
+        {
+            case value_t::boolean:
+            case value_t::number_float:
+            case value_t::number_integer:
+            case value_t::number_unsigned:
+            case value_t::string:
+            case value_t::binary:
+            {
+                if (JSON_HEDLEY_LIKELY(!first.m_it.primitive_iterator.is_begin()
+                                       || !last.m_it.primitive_iterator.is_end()))
+                {
+                    JSON_THROW(invalid_iterator::create(204, "iterators out of range", this));
+                }
+
+                if (is_string())
+                {
+                    AllocatorType<string_t> alloc;
+                    std::allocator_traits<decltype(alloc)>::destroy(alloc, m_value.string);
+                    std::allocator_traits<decltype(alloc)>::deallocate(alloc, m_value.string, 1);
+                    m_value.string = nullptr;
+                }
+                else if (is_binary())
+                {
+                    AllocatorType<binary_t> alloc;
+                    std::allocator_traits<decltype(alloc)>::destroy(alloc, m_value.binary);
+                    std::allocator_traits<decltype(alloc)>::deallocate(alloc, m_value.binary, 1);
+                    m_value.binary = nullptr;
+                }
+
+                m_type = value_t::null;
+                assert_invariant();
+                break;
+            }
+
+            case value_t::object:
+            {
+                result.m_it.object_iterator = m_value.object->erase(first.m_it.object_iterator,
+                                              last.m_it.object_iterator);
+                break;
+            }
+
+            case value_t::array:
+            {
+                result.m_it.array_iterator = m_value.array->erase(first.m_it.array_iterator,
+                                             last.m_it.array_iterator);
+                break;
+            }
+
+            case value_t::null:
+            case value_t::discarded:
+            default:
+                JSON_THROW(type_error::create(307, detail::concat("cannot use erase() with ", type_name()), this));
+        }
+
+        return result;
+    }
+
+  private:
+    template < typename KeyType, detail::enable_if_t <
+                   detail::has_erase_with_key_type<basic_json_t, KeyType>::value, int > = 0 >
+    size_type erase_internal(KeyType && key)
+    {
+        // this erase only works for objects
+        if (JSON_HEDLEY_UNLIKELY(!is_object()))
+        {
+            JSON_THROW(type_error::create(307, detail::concat("cannot use erase() with ", type_name()), this));
+        }
+
+        return m_value.object->erase(std::forward<KeyType>(key));
+    }
+
+    template < typename KeyType, detail::enable_if_t <
+                   !detail::has_erase_with_key_type<basic_json_t, KeyType>::value, int > = 0 >
+    size_type erase_internal(KeyType && key)
+    {
+        // this erase only works for objects
+        if (JSON_HEDLEY_UNLIKELY(!is_object()))
+        {
+            JSON_THROW(type_error::create(307, detail::concat("cannot use erase() with ", type_name()), this));
+        }
+
+        const auto it = m_value.object->find(std::forward<KeyType>(key));
+        if (it != m_value.object->end())
+        {
+            m_value.object->erase(it);
+            return 1;
+        }
+        return 0;
+    }
+
+  public:
+
+    /// @brief remove element from a JSON object given a key
+    /// @sa https://json.nlohmann.me/api/basic_json/erase/
+    size_type erase(const typename object_t::key_type& key)
+    {
+        // the indirection via erase_internal() is added to avoid making this
+        // function a template and thus de-rank it during overload resolution
+        return erase_internal(key);
+    }
+
+    /// @brief remove element from a JSON object given a key
+    /// @sa https://json.nlohmann.me/api/basic_json/erase/
+    template<class KeyType, detail::enable_if_t<
+                 detail::is_usable_as_basic_json_key_type<basic_json_t, KeyType>::value, int> = 0>
+    size_type erase(KeyType && key)
+    {
+        return erase_internal(std::forward<KeyType>(key));
+    }
+
+    /// @brief remove element from a JSON array given an index
+    /// @sa https://json.nlohmann.me/api/basic_json/erase/
+    void erase(const size_type idx)
+    {
+        // this erase only works for arrays
+        if (JSON_HEDLEY_LIKELY(is_array()))
+        {
+            if (JSON_HEDLEY_UNLIKELY(idx >= size()))
+            {
+                JSON_THROW(out_of_range::create(401, detail::concat("array index ", std::to_string(idx), " is out of range"), this));
+            }
+
+            m_value.array->erase(m_value.array->begin() + static_cast<difference_type>(idx));
+        }
+        else
+        {
+            JSON_THROW(type_error::create(307, detail::concat("cannot use erase() with ", type_name()), this));
+        }
+    }
+
+    /// @}
+
+
+    ////////////
+    // lookup //
+    ////////////
+
+    /// @name lookup
+    /// @{
+
+    /// @brief find an element in a JSON object
+    /// @sa https://json.nlohmann.me/api/basic_json/find/
+    iterator find(const typename object_t::key_type& key)
+    {
+        auto result = end();
+
+        if (is_object())
+        {
+            result.m_it.object_iterator = m_value.object->find(key);
+        }
+
+        return result;
+    }
+
+    /// @brief find an element in a JSON object
+    /// @sa https://json.nlohmann.me/api/basic_json/find/
+    const_iterator find(const typename object_t::key_type& key) const
+    {
+        auto result = cend();
+
+        if (is_object())
+        {
+            result.m_it.object_iterator = m_value.object->find(key);
+        }
+
+        return result;
+    }
+
+    /// @brief find an element in a JSON object
+    /// @sa https://json.nlohmann.me/api/basic_json/find/
+    template<class KeyType, detail::enable_if_t<
+                 detail::is_usable_as_basic_json_key_type<basic_json_t, KeyType>::value, int> = 0>
+    iterator find(KeyType && key)
+    {
+        auto result = end();
+
+        if (is_object())
+        {
+            result.m_it.object_iterator = m_value.object->find(std::forward<KeyType>(key));
+        }
+
+        return result;
+    }
+
+    /// @brief find an element in a JSON object
+    /// @sa https://json.nlohmann.me/api/basic_json/find/
+    template<class KeyType, detail::enable_if_t<
+                 detail::is_usable_as_basic_json_key_type<basic_json_t, KeyType>::value, int> = 0>
+    const_iterator find(KeyType && key) const
+    {
+        auto result = cend();
+
+        if (is_object())
+        {
+            result.m_it.object_iterator = m_value.object->find(std::forward<KeyType>(key));
+        }
+
+        return result;
+    }
+
+    /// @brief returns the number of occurrences of a key in a JSON object
+    /// @sa https://json.nlohmann.me/api/basic_json/count/
+    size_type count(const typename object_t::key_type& key) const
+    {
+        // return 0 for all nonobject types
+        return is_object() ? m_value.object->count(key) : 0;
+    }
+
+    /// @brief returns the number of occurrences of a key in a JSON object
+    /// @sa https://json.nlohmann.me/api/basic_json/count/
+    template<class KeyType, detail::enable_if_t<
+                 detail::is_usable_as_basic_json_key_type<basic_json_t, KeyType>::value, int> = 0>
+    size_type count(KeyType && key) const
+    {
+        // return 0 for all nonobject types
+        return is_object() ? m_value.object->count(std::forward<KeyType>(key)) : 0;
+    }
+
+    /// @brief check the existence of an element in a JSON object
+    /// @sa https://json.nlohmann.me/api/basic_json/contains/
+    bool contains(const typename object_t::key_type& key) const
+    {
+        return is_object() && m_value.object->find(key) != m_value.object->end();
+    }
+
+    /// @brief check the existence of an element in a JSON object
+    /// @sa https://json.nlohmann.me/api/basic_json/contains/
+    template<class KeyType, detail::enable_if_t<
+                 detail::is_usable_as_basic_json_key_type<basic_json_t, KeyType>::value, int> = 0>
+    bool contains(KeyType && key) const
+    {
+        return is_object() && m_value.object->find(std::forward<KeyType>(key)) != m_value.object->end();
+    }
+
+    /// @brief check the existence of an element in a JSON object given a JSON pointer
+    /// @sa https://json.nlohmann.me/api/basic_json/contains/
+    bool contains(const json_pointer& ptr) const
+    {
+        return ptr.contains(this);
+    }
+
+    template<typename BasicJsonType, detail::enable_if_t<detail::is_basic_json<BasicJsonType>::value, int> = 0>
+    JSON_HEDLEY_DEPRECATED_FOR(3.11.0, basic_json::json_pointer or nlohmann::json_pointer<basic_json::string_t>) // NOLINT(readability/alt_tokens)
+    bool contains(const typename ::nlohmann::json_pointer<BasicJsonType>& ptr) const
+    {
+        return ptr.contains(this);
+    }
+
+    /// @}
+
+
+    ///////////////
+    // iterators //
+    ///////////////
+
+    /// @name iterators
+    /// @{
+
+    /// @brief returns an iterator to the first element
+    /// @sa https://json.nlohmann.me/api/basic_json/begin/
+    iterator begin() noexcept
+    {
+        iterator result(this);
+        result.set_begin();
+        return result;
+    }
+
+    /// @brief returns an iterator to the first element
+    /// @sa https://json.nlohmann.me/api/basic_json/begin/
+    const_iterator begin() const noexcept
+    {
+        return cbegin();
+    }
+
+    /// @brief returns a const iterator to the first element
+    /// @sa https://json.nlohmann.me/api/basic_json/cbegin/
+    const_iterator cbegin() const noexcept
+    {
+        const_iterator result(this);
+        result.set_begin();
+        return result;
+    }
+
+    /// @brief returns an iterator to one past the last element
+    /// @sa https://json.nlohmann.me/api/basic_json/end/
+    iterator end() noexcept
+    {
+        iterator result(this);
+        result.set_end();
+        return result;
+    }
+
+    /// @brief returns an iterator to one past the last element
+    /// @sa https://json.nlohmann.me/api/basic_json/end/
+    const_iterator end() const noexcept
+    {
+        return cend();
+    }
+
+    /// @brief returns an iterator to one past the last element
+    /// @sa https://json.nlohmann.me/api/basic_json/cend/
+    const_iterator cend() const noexcept
+    {
+        const_iterator result(this);
+        result.set_end();
+        return result;
+    }
+
+    /// @brief returns an iterator to the reverse-beginning
+    /// @sa https://json.nlohmann.me/api/basic_json/rbegin/
+    reverse_iterator rbegin() noexcept
+    {
+        return reverse_iterator(end());
+    }
+
+    /// @brief returns an iterator to the reverse-beginning
+    /// @sa https://json.nlohmann.me/api/basic_json/rbegin/
+    const_reverse_iterator rbegin() const noexcept
+    {
+        return crbegin();
+    }
+
+    /// @brief returns an iterator to the reverse-end
+    /// @sa https://json.nlohmann.me/api/basic_json/rend/
+    reverse_iterator rend() noexcept
+    {
+        return reverse_iterator(begin());
+    }
+
+    /// @brief returns an iterator to the reverse-end
+    /// @sa https://json.nlohmann.me/api/basic_json/rend/
+    const_reverse_iterator rend() const noexcept
+    {
+        return crend();
+    }
+
+    /// @brief returns a const reverse iterator to the last element
+    /// @sa https://json.nlohmann.me/api/basic_json/crbegin/
+    const_reverse_iterator crbegin() const noexcept
+    {
+        return const_reverse_iterator(cend());
+    }
+
+    /// @brief returns a const reverse iterator to one before the first
+    /// @sa https://json.nlohmann.me/api/basic_json/crend/
+    const_reverse_iterator crend() const noexcept
+    {
+        return const_reverse_iterator(cbegin());
+    }
+
+  public:
+    /// @brief wrapper to access iterator member functions in range-based for
+    /// @sa https://json.nlohmann.me/api/basic_json/items/
+    /// @deprecated This function is deprecated since 3.1.0 and will be removed in
+    ///             version 4.0.0 of the library. Please use @ref items() instead;
+    ///             that is, replace `json::iterator_wrapper(j)` with `j.items()`.
+    JSON_HEDLEY_DEPRECATED_FOR(3.1.0, items())
+    static iteration_proxy<iterator> iterator_wrapper(reference ref) noexcept
+    {
+        return ref.items();
+    }
+
+    /// @brief wrapper to access iterator member functions in range-based for
+    /// @sa https://json.nlohmann.me/api/basic_json/items/
+    /// @deprecated This function is deprecated since 3.1.0 and will be removed in
+    ///         version 4.0.0 of the library. Please use @ref items() instead;
+    ///         that is, replace `json::iterator_wrapper(j)` with `j.items()`.
+    JSON_HEDLEY_DEPRECATED_FOR(3.1.0, items())
+    static iteration_proxy<const_iterator> iterator_wrapper(const_reference ref) noexcept
+    {
+        return ref.items();
+    }
+
+    /// @brief helper to access iterator member functions in range-based for
+    /// @sa https://json.nlohmann.me/api/basic_json/items/
+    iteration_proxy<iterator> items() noexcept
+    {
+        return iteration_proxy<iterator>(*this);
+    }
+
+    /// @brief helper to access iterator member functions in range-based for
+    /// @sa https://json.nlohmann.me/api/basic_json/items/
+    iteration_proxy<const_iterator> items() const noexcept
+    {
+        return iteration_proxy<const_iterator>(*this);
+    }
+
+    /// @}
+
+
+    //////////////
+    // capacity //
+    //////////////
+
+    /// @name capacity
+    /// @{
+
+    /// @brief checks whether the container is empty.
+    /// @sa https://json.nlohmann.me/api/basic_json/empty/
+    bool empty() const noexcept
+    {
+        switch (m_type)
+        {
+            case value_t::null:
+            {
+                // null values are empty
+                return true;
+            }
+
+            case value_t::array:
+            {
+                // delegate call to array_t::empty()
+                return m_value.array->empty();
+            }
+
+            case value_t::object:
+            {
+                // delegate call to object_t::empty()
+                return m_value.object->empty();
+            }
+
+            case value_t::string:
+            case value_t::boolean:
+            case value_t::number_integer:
+            case value_t::number_unsigned:
+            case value_t::number_float:
+            case value_t::binary:
+            case value_t::discarded:
+            default:
+            {
+                // all other types are nonempty
+                return false;
+            }
+        }
+    }
+
+    /// @brief returns the number of elements
+    /// @sa https://json.nlohmann.me/api/basic_json/size/
+    size_type size() const noexcept
+    {
+        switch (m_type)
+        {
+            case value_t::null:
+            {
+                // null values are empty
+                return 0;
+            }
+
+            case value_t::array:
+            {
+                // delegate call to array_t::size()
+                return m_value.array->size();
+            }
+
+            case value_t::object:
+            {
+                // delegate call to object_t::size()
+                return m_value.object->size();
+            }
+
+            case value_t::string:
+            case value_t::boolean:
+            case value_t::number_integer:
+            case value_t::number_unsigned:
+            case value_t::number_float:
+            case value_t::binary:
+            case value_t::discarded:
+            default:
+            {
+                // all other types have size 1
+                return 1;
+            }
+        }
+    }
+
+    /// @brief returns the maximum possible number of elements
+    /// @sa https://json.nlohmann.me/api/basic_json/max_size/
+    size_type max_size() const noexcept
+    {
+        switch (m_type)
+        {
+            case value_t::array:
+            {
+                // delegate call to array_t::max_size()
+                return m_value.array->max_size();
+            }
+
+            case value_t::object:
+            {
+                // delegate call to object_t::max_size()
+                return m_value.object->max_size();
+            }
+
+            case value_t::null:
+            case value_t::string:
+            case value_t::boolean:
+            case value_t::number_integer:
+            case value_t::number_unsigned:
+            case value_t::number_float:
+            case value_t::binary:
+            case value_t::discarded:
+            default:
+            {
+                // all other types have max_size() == size()
+                return size();
+            }
+        }
+    }
+
+    /// @}
+
+
+    ///////////////
+    // modifiers //
+    ///////////////
+
+    /// @name modifiers
+    /// @{
+
+    /// @brief clears the contents
+    /// @sa https://json.nlohmann.me/api/basic_json/clear/
+    void clear() noexcept
+    {
+        switch (m_type)
+        {
+            case value_t::number_integer:
+            {
+                m_value.number_integer = 0;
+                break;
+            }
+
+            case value_t::number_unsigned:
+            {
+                m_value.number_unsigned = 0;
+                break;
+            }
+
+            case value_t::number_float:
+            {
+                m_value.number_float = 0.0;
+                break;
+            }
+
+            case value_t::boolean:
+            {
+                m_value.boolean = false;
+                break;
+            }
+
+            case value_t::string:
+            {
+                m_value.string->clear();
+                break;
+            }
+
+            case value_t::binary:
+            {
+                m_value.binary->clear();
+                break;
+            }
+
+            case value_t::array:
+            {
+                m_value.array->clear();
+                break;
+            }
+
+            case value_t::object:
+            {
+                m_value.object->clear();
+                break;
+            }
+
+            case value_t::null:
+            case value_t::discarded:
+            default:
+                break;
+        }
+    }
+
+    /// @brief add an object to an array
+    /// @sa https://json.nlohmann.me/api/basic_json/push_back/
+    void push_back(basic_json&& val)
+    {
+        // push_back only works for null objects or arrays
+        if (JSON_HEDLEY_UNLIKELY(!(is_null() || is_array())))
+        {
+            JSON_THROW(type_error::create(308, detail::concat("cannot use push_back() with ", type_name()), this));
+        }
+
+        // transform null object into an array
+        if (is_null())
+        {
+            m_type = value_t::array;
+            m_value = value_t::array;
+            assert_invariant();
+        }
+
+        // add element to array (move semantics)
+        const auto old_capacity = m_value.array->capacity();
+        m_value.array->push_back(std::move(val));
+        set_parent(m_value.array->back(), old_capacity);
+        // if val is moved from, basic_json move constructor marks it null, so we do not call the destructor
+    }
+
+    /// @brief add an object to an array
+    /// @sa https://json.nlohmann.me/api/basic_json/operator+=/
+    reference operator+=(basic_json&& val)
+    {
+        push_back(std::move(val));
+        return *this;
+    }
+
+    /// @brief add an object to an array
+    /// @sa https://json.nlohmann.me/api/basic_json/push_back/
+    void push_back(const basic_json& val)
+    {
+        // push_back only works for null objects or arrays
+        if (JSON_HEDLEY_UNLIKELY(!(is_null() || is_array())))
+        {
+            JSON_THROW(type_error::create(308, detail::concat("cannot use push_back() with ", type_name()), this));
+        }
+
+        // transform null object into an array
+        if (is_null())
+        {
+            m_type = value_t::array;
+            m_value = value_t::array;
+            assert_invariant();
+        }
+
+        // add element to array
+        const auto old_capacity = m_value.array->capacity();
+        m_value.array->push_back(val);
+        set_parent(m_value.array->back(), old_capacity);
+    }
+
+    /// @brief add an object to an array
+    /// @sa https://json.nlohmann.me/api/basic_json/operator+=/
+    reference operator+=(const basic_json& val)
+    {
+        push_back(val);
+        return *this;
+    }
+
+    /// @brief add an object to an object
+    /// @sa https://json.nlohmann.me/api/basic_json/push_back/
+    void push_back(const typename object_t::value_type& val)
+    {
+        // push_back only works for null objects or objects
+        if (JSON_HEDLEY_UNLIKELY(!(is_null() || is_object())))
+        {
+            JSON_THROW(type_error::create(308, detail::concat("cannot use push_back() with ", type_name()), this));
+        }
+
+        // transform null object into an object
+        if (is_null())
+        {
+            m_type = value_t::object;
+            m_value = value_t::object;
+            assert_invariant();
+        }
+
+        // add element to object
+        auto res = m_value.object->insert(val);
+        set_parent(res.first->second);
+    }
+
+    /// @brief add an object to an object
+    /// @sa https://json.nlohmann.me/api/basic_json/operator+=/
+    reference operator+=(const typename object_t::value_type& val)
+    {
+        push_back(val);
+        return *this;
+    }
+
+    /// @brief add an object to an object
+    /// @sa https://json.nlohmann.me/api/basic_json/push_back/
+    void push_back(initializer_list_t init)
+    {
+        if (is_object() && init.size() == 2 && (*init.begin())->is_string())
+        {
+            basic_json&& key = init.begin()->moved_or_copied();
+            push_back(typename object_t::value_type(
+                          std::move(key.get_ref<string_t&>()), (init.begin() + 1)->moved_or_copied()));
+        }
+        else
+        {
+            push_back(basic_json(init));
+        }
+    }
+
+    /// @brief add an object to an object
+    /// @sa https://json.nlohmann.me/api/basic_json/operator+=/
+    reference operator+=(initializer_list_t init)
+    {
+        push_back(init);
+        return *this;
+    }
+
+    /// @brief add an object to an array
+    /// @sa https://json.nlohmann.me/api/basic_json/emplace_back/
+    template<class... Args>
+    reference emplace_back(Args&& ... args)
+    {
+        // emplace_back only works for null objects or arrays
+        if (JSON_HEDLEY_UNLIKELY(!(is_null() || is_array())))
+        {
+            JSON_THROW(type_error::create(311, detail::concat("cannot use emplace_back() with ", type_name()), this));
+        }
+
+        // transform null object into an array
+        if (is_null())
+        {
+            m_type = value_t::array;
+            m_value = value_t::array;
+            assert_invariant();
+        }
+
+        // add element to array (perfect forwarding)
+        const auto old_capacity = m_value.array->capacity();
+        m_value.array->emplace_back(std::forward<Args>(args)...);
+        return set_parent(m_value.array->back(), old_capacity);
+    }
+
+    /// @brief add an object to an object if key does not exist
+    /// @sa https://json.nlohmann.me/api/basic_json/emplace/
+    template<class... Args>
+    std::pair<iterator, bool> emplace(Args&& ... args)
+    {
+        // emplace only works for null objects or arrays
+        if (JSON_HEDLEY_UNLIKELY(!(is_null() || is_object())))
+        {
+            JSON_THROW(type_error::create(311, detail::concat("cannot use emplace() with ", type_name()), this));
+        }
+
+        // transform null object into an object
+        if (is_null())
+        {
+            m_type = value_t::object;
+            m_value = value_t::object;
+            assert_invariant();
+        }
+
+        // add element to array (perfect forwarding)
+        auto res = m_value.object->emplace(std::forward<Args>(args)...);
+        set_parent(res.first->second);
+
+        // create result iterator and set iterator to the result of emplace
+        auto it = begin();
+        it.m_it.object_iterator = res.first;
+
+        // return pair of iterator and boolean
+        return {it, res.second};
+    }
+
+    /// Helper for insertion of an iterator
+    /// @note: This uses std::distance to support GCC 4.8,
+    ///        see https://github.com/nlohmann/json/pull/1257
+    template<typename... Args>
+    iterator insert_iterator(const_iterator pos, Args&& ... args)
+    {
+        iterator result(this);
+        JSON_ASSERT(m_value.array != nullptr);
+
+        auto insert_pos = std::distance(m_value.array->begin(), pos.m_it.array_iterator);
+        m_value.array->insert(pos.m_it.array_iterator, std::forward<Args>(args)...);
+        result.m_it.array_iterator = m_value.array->begin() + insert_pos;
+
+        // This could have been written as:
+        // result.m_it.array_iterator = m_value.array->insert(pos.m_it.array_iterator, cnt, val);
+        // but the return value of insert is missing in GCC 4.8, so it is written this way instead.
+
+        set_parents();
+        return result;
+    }
+
+    /// @brief inserts element into array
+    /// @sa https://json.nlohmann.me/api/basic_json/insert/
+    iterator insert(const_iterator pos, const basic_json& val)
+    {
+        // insert only works for arrays
+        if (JSON_HEDLEY_LIKELY(is_array()))
+        {
+            // check if iterator pos fits to this JSON value
+            if (JSON_HEDLEY_UNLIKELY(pos.m_object != this))
+            {
+                JSON_THROW(invalid_iterator::create(202, "iterator does not fit current value", this));
+            }
+
+            // insert to array and return iterator
+            return insert_iterator(pos, val);
+        }
+
+        JSON_THROW(type_error::create(309, detail::concat("cannot use insert() with ", type_name()), this));
+    }
+
+    /// @brief inserts element into array
+    /// @sa https://json.nlohmann.me/api/basic_json/insert/
+    iterator insert(const_iterator pos, basic_json&& val)
+    {
+        return insert(pos, val);
+    }
+
+    /// @brief inserts copies of element into array
+    /// @sa https://json.nlohmann.me/api/basic_json/insert/
+    iterator insert(const_iterator pos, size_type cnt, const basic_json& val)
+    {
+        // insert only works for arrays
+        if (JSON_HEDLEY_LIKELY(is_array()))
+        {
+            // check if iterator pos fits to this JSON value
+            if (JSON_HEDLEY_UNLIKELY(pos.m_object != this))
+            {
+                JSON_THROW(invalid_iterator::create(202, "iterator does not fit current value", this));
+            }
+
+            // insert to array and return iterator
+            return insert_iterator(pos, cnt, val);
+        }
+
+        JSON_THROW(type_error::create(309, detail::concat("cannot use insert() with ", type_name()), this));
+    }
+
+    /// @brief inserts range of elements into array
+    /// @sa https://json.nlohmann.me/api/basic_json/insert/
+    iterator insert(const_iterator pos, const_iterator first, const_iterator last)
+    {
+        // insert only works for arrays
+        if (JSON_HEDLEY_UNLIKELY(!is_array()))
+        {
+            JSON_THROW(type_error::create(309, detail::concat("cannot use insert() with ", type_name()), this));
+        }
+
+        // check if iterator pos fits to this JSON value
+        if (JSON_HEDLEY_UNLIKELY(pos.m_object != this))
+        {
+            JSON_THROW(invalid_iterator::create(202, "iterator does not fit current value", this));
+        }
+
+        // check if range iterators belong to the same JSON object
+        if (JSON_HEDLEY_UNLIKELY(first.m_object != last.m_object))
+        {
+            JSON_THROW(invalid_iterator::create(210, "iterators do not fit", this));
+        }
+
+        if (JSON_HEDLEY_UNLIKELY(first.m_object == this))
+        {
+            JSON_THROW(invalid_iterator::create(211, "passed iterators may not belong to container", this));
+        }
+
+        // insert to array and return iterator
+        return insert_iterator(pos, first.m_it.array_iterator, last.m_it.array_iterator);
+    }
+
+    /// @brief inserts elements from initializer list into array
+    /// @sa https://json.nlohmann.me/api/basic_json/insert/
+    iterator insert(const_iterator pos, initializer_list_t ilist)
+    {
+        // insert only works for arrays
+        if (JSON_HEDLEY_UNLIKELY(!is_array()))
+        {
+            JSON_THROW(type_error::create(309, detail::concat("cannot use insert() with ", type_name()), this));
+        }
+
+        // check if iterator pos fits to this JSON value
+        if (JSON_HEDLEY_UNLIKELY(pos.m_object != this))
+        {
+            JSON_THROW(invalid_iterator::create(202, "iterator does not fit current value", this));
+        }
+
+        // insert to array and return iterator
+        return insert_iterator(pos, ilist.begin(), ilist.end());
+    }
+
+    /// @brief inserts range of elements into object
+    /// @sa https://json.nlohmann.me/api/basic_json/insert/
+    void insert(const_iterator first, const_iterator last)
+    {
+        // insert only works for objects
+        if (JSON_HEDLEY_UNLIKELY(!is_object()))
+        {
+            JSON_THROW(type_error::create(309, detail::concat("cannot use insert() with ", type_name()), this));
+        }
+
+        // check if range iterators belong to the same JSON object
+        if (JSON_HEDLEY_UNLIKELY(first.m_object != last.m_object))
+        {
+            JSON_THROW(invalid_iterator::create(210, "iterators do not fit", this));
+        }
+
+        // passed iterators must belong to objects
+        if (JSON_HEDLEY_UNLIKELY(!first.m_object->is_object()))
+        {
+            JSON_THROW(invalid_iterator::create(202, "iterators first and last must point to objects", this));
+        }
+
+        m_value.object->insert(first.m_it.object_iterator, last.m_it.object_iterator);
+    }
+
+    /// @brief updates a JSON object from another object, overwriting existing keys
+    /// @sa https://json.nlohmann.me/api/basic_json/update/
+    void update(const_reference j, bool merge_objects = false)
+    {
+        update(j.begin(), j.end(), merge_objects);
+    }
+
+    /// @brief updates a JSON object from another object, overwriting existing keys
+    /// @sa https://json.nlohmann.me/api/basic_json/update/
+    void update(const_iterator first, const_iterator last, bool merge_objects = false)
+    {
+        // implicitly convert null value to an empty object
+        if (is_null())
+        {
+            m_type = value_t::object;
+            m_value.object = create<object_t>();
+            assert_invariant();
+        }
+
+        if (JSON_HEDLEY_UNLIKELY(!is_object()))
+        {
+            JSON_THROW(type_error::create(312, detail::concat("cannot use update() with ", type_name()), this));
+        }
+
+        // check if range iterators belong to the same JSON object
+        if (JSON_HEDLEY_UNLIKELY(first.m_object != last.m_object))
+        {
+            JSON_THROW(invalid_iterator::create(210, "iterators do not fit", this));
+        }
+
+        // passed iterators must belong to objects
+        if (JSON_HEDLEY_UNLIKELY(!first.m_object->is_object()))
+        {
+            JSON_THROW(type_error::create(312, detail::concat("cannot use update() with ", first.m_object->type_name()), first.m_object));
+        }
+
+        for (auto it = first; it != last; ++it)
+        {
+            if (merge_objects && it.value().is_object())
+            {
+                auto it2 = m_value.object->find(it.key());
+                if (it2 != m_value.object->end())
+                {
+                    it2->second.update(it.value(), true);
+                    continue;
+                }
+            }
+            m_value.object->operator[](it.key()) = it.value();
+#if JSON_DIAGNOSTICS
+            m_value.object->operator[](it.key()).m_parent = this;
+#endif
+        }
+    }
+
+    /// @brief exchanges the values
+    /// @sa https://json.nlohmann.me/api/basic_json/swap/
+    void swap(reference other) noexcept (
+        std::is_nothrow_move_constructible<value_t>::value&&
+        std::is_nothrow_move_assignable<value_t>::value&&
+        std::is_nothrow_move_constructible<json_value>::value&&
+        std::is_nothrow_move_assignable<json_value>::value
+    )
+    {
+        std::swap(m_type, other.m_type);
+        std::swap(m_value, other.m_value);
+
+        set_parents();
+        other.set_parents();
+        assert_invariant();
+    }
+
+    /// @brief exchanges the values
+    /// @sa https://json.nlohmann.me/api/basic_json/swap/
+    friend void swap(reference left, reference right) noexcept (
+        std::is_nothrow_move_constructible<value_t>::value&&
+        std::is_nothrow_move_assignable<value_t>::value&&
+        std::is_nothrow_move_constructible<json_value>::value&&
+        std::is_nothrow_move_assignable<json_value>::value
+    )
+    {
+        left.swap(right);
+    }
+
+    /// @brief exchanges the values
+    /// @sa https://json.nlohmann.me/api/basic_json/swap/
+    void swap(array_t& other) // NOLINT(bugprone-exception-escape)
+    {
+        // swap only works for arrays
+        if (JSON_HEDLEY_LIKELY(is_array()))
+        {
+            using std::swap;
+            swap(*(m_value.array), other);
+        }
+        else
+        {
+            JSON_THROW(type_error::create(310, detail::concat("cannot use swap(array_t&) with ", type_name()), this));
+        }
+    }
+
+    /// @brief exchanges the values
+    /// @sa https://json.nlohmann.me/api/basic_json/swap/
+    void swap(object_t& other) // NOLINT(bugprone-exception-escape)
+    {
+        // swap only works for objects
+        if (JSON_HEDLEY_LIKELY(is_object()))
+        {
+            using std::swap;
+            swap(*(m_value.object), other);
+        }
+        else
+        {
+            JSON_THROW(type_error::create(310, detail::concat("cannot use swap(object_t&) with ", type_name()), this));
+        }
+    }
+
+    /// @brief exchanges the values
+    /// @sa https://json.nlohmann.me/api/basic_json/swap/
+    void swap(string_t& other) // NOLINT(bugprone-exception-escape)
+    {
+        // swap only works for strings
+        if (JSON_HEDLEY_LIKELY(is_string()))
+        {
+            using std::swap;
+            swap(*(m_value.string), other);
+        }
+        else
+        {
+            JSON_THROW(type_error::create(310, detail::concat("cannot use swap(string_t&) with ", type_name()), this));
+        }
+    }
+
+    /// @brief exchanges the values
+    /// @sa https://json.nlohmann.me/api/basic_json/swap/
+    void swap(binary_t& other) // NOLINT(bugprone-exception-escape)
+    {
+        // swap only works for strings
+        if (JSON_HEDLEY_LIKELY(is_binary()))
+        {
+            using std::swap;
+            swap(*(m_value.binary), other);
+        }
+        else
+        {
+            JSON_THROW(type_error::create(310, detail::concat("cannot use swap(binary_t&) with ", type_name()), this));
+        }
+    }
+
+    /// @brief exchanges the values
+    /// @sa https://json.nlohmann.me/api/basic_json/swap/
+    void swap(typename binary_t::container_type& other) // NOLINT(bugprone-exception-escape)
+    {
+        // swap only works for strings
+        if (JSON_HEDLEY_LIKELY(is_binary()))
+        {
+            using std::swap;
+            swap(*(m_value.binary), other);
+        }
+        else
+        {
+            JSON_THROW(type_error::create(310, detail::concat("cannot use swap(binary_t::container_type&) with ", type_name()), this));
+        }
+    }
+
+    /// @}
+
+    //////////////////////////////////////////
+    // lexicographical comparison operators //
+    //////////////////////////////////////////
+
+    /// @name lexicographical comparison operators
+    /// @{
+
+    // note parentheses around operands are necessary; see
+    // https://github.com/nlohmann/json/issues/1530
+#define JSON_IMPLEMENT_OPERATOR(op, null_result, unordered_result, default_result)                       \
+    const auto lhs_type = lhs.type();                                                                    \
+    const auto rhs_type = rhs.type();                                                                    \
+    \
+    if (lhs_type == rhs_type) /* NOLINT(readability/braces) */                                           \
+    {                                                                                                    \
+        switch (lhs_type)                                                                                \
+        {                                                                                                \
+            case value_t::array:                                                                         \
+                return (*lhs.m_value.array) op (*rhs.m_value.array);                                     \
+                \
+            case value_t::object:                                                                        \
+                return (*lhs.m_value.object) op (*rhs.m_value.object);                                   \
+                \
+            case value_t::null:                                                                          \
+                return (null_result);                                                                    \
+                \
+            case value_t::string:                                                                        \
+                return (*lhs.m_value.string) op (*rhs.m_value.string);                                   \
+                \
+            case value_t::boolean:                                                                       \
+                return (lhs.m_value.boolean) op (rhs.m_value.boolean);                                   \
+                \
+            case value_t::number_integer:                                                                \
+                return (lhs.m_value.number_integer) op (rhs.m_value.number_integer);                     \
+                \
+            case value_t::number_unsigned:                                                               \
+                return (lhs.m_value.number_unsigned) op (rhs.m_value.number_unsigned);                   \
+                \
+            case value_t::number_float:                                                                  \
+                return (lhs.m_value.number_float) op (rhs.m_value.number_float);                         \
+                \
+            case value_t::binary:                                                                        \
+                return (*lhs.m_value.binary) op (*rhs.m_value.binary);                                   \
+                \
+            case value_t::discarded:                                                                     \
+            default:                                                                                     \
+                return (unordered_result);                                                               \
+        }                                                                                                \
+    }                                                                                                    \
+    else if (lhs_type == value_t::number_integer && rhs_type == value_t::number_float)                   \
+    {                                                                                                    \
+        return static_cast<number_float_t>(lhs.m_value.number_integer) op rhs.m_value.number_float;      \
+    }                                                                                                    \
+    else if (lhs_type == value_t::number_float && rhs_type == value_t::number_integer)                   \
+    {                                                                                                    \
+        return lhs.m_value.number_float op static_cast<number_float_t>(rhs.m_value.number_integer);      \
+    }                                                                                                    \
+    else if (lhs_type == value_t::number_unsigned && rhs_type == value_t::number_float)                  \
+    {                                                                                                    \
+        return static_cast<number_float_t>(lhs.m_value.number_unsigned) op rhs.m_value.number_float;     \
+    }                                                                                                    \
+    else if (lhs_type == value_t::number_float && rhs_type == value_t::number_unsigned)                  \
+    {                                                                                                    \
+        return lhs.m_value.number_float op static_cast<number_float_t>(rhs.m_value.number_unsigned);     \
+    }                                                                                                    \
+    else if (lhs_type == value_t::number_unsigned && rhs_type == value_t::number_integer)                \
+    {                                                                                                    \
+        return static_cast<number_integer_t>(lhs.m_value.number_unsigned) op rhs.m_value.number_integer; \
+    }                                                                                                    \
+    else if (lhs_type == value_t::number_integer && rhs_type == value_t::number_unsigned)                \
+    {                                                                                                    \
+        return lhs.m_value.number_integer op static_cast<number_integer_t>(rhs.m_value.number_unsigned); \
+    }                                                                                                    \
+    else if(compares_unordered(lhs, rhs))\
+    {\
+        return (unordered_result);\
+    }\
+    \
+    return (default_result);
+
+  JSON_PRIVATE_UNLESS_TESTED:
+    // returns true if:
+    // - any operand is NaN and the other operand is of number type
+    // - any operand is discarded
+    // in legacy mode, discarded values are considered ordered if
+    // an operation is computed as an odd number of inverses of others
+    static bool compares_unordered(const_reference lhs, const_reference rhs, bool inverse = false) noexcept
+    {
+        if ((lhs.is_number_float() && std::isnan(lhs.m_value.number_float) && rhs.is_number())
+                || (rhs.is_number_float() && std::isnan(rhs.m_value.number_float) && lhs.is_number()))
+        {
+            return true;
+        }
+#if JSON_USE_LEGACY_DISCARDED_VALUE_COMPARISON
+        return (lhs.is_discarded() || rhs.is_discarded()) && !inverse;
+#else
+        static_cast<void>(inverse);
+        return lhs.is_discarded() || rhs.is_discarded();
+#endif
+    }
+
+  private:
+    bool compares_unordered(const_reference rhs, bool inverse = false) const noexcept
+    {
+        return compares_unordered(*this, rhs, inverse);
+    }
+
+  public:
+#if JSON_HAS_THREE_WAY_COMPARISON
+    /// @brief comparison: equal
+    /// @sa https://json.nlohmann.me/api/basic_json/operator_eq/
+    bool operator==(const_reference rhs) const noexcept
+    {
+#ifdef __GNUC__
+#pragma GCC diagnostic push
+#pragma GCC diagnostic ignored "-Wfloat-equal"
+#endif
+        const_reference lhs = *this;
+        JSON_IMPLEMENT_OPERATOR( ==, true, false, false)
+#ifdef __GNUC__
+#pragma GCC diagnostic pop
+#endif
+    }
+
+    /// @brief comparison: equal
+    /// @sa https://json.nlohmann.me/api/basic_json/operator_eq/
+    template<typename ScalarType>
+    requires std::is_scalar_v<ScalarType>
+    bool operator==(ScalarType rhs) const noexcept
+    {
+        return *this == basic_json(rhs);
+    }
+
+    /// @brief comparison: not equal
+    /// @sa https://json.nlohmann.me/api/basic_json/operator_ne/
+    bool operator!=(const_reference rhs) const noexcept
+    {
+        if (compares_unordered(rhs, true))
+        {
+            return false;
+        }
+        return !operator==(rhs);
+    }
+
+    /// @brief comparison: 3-way
+    /// @sa https://json.nlohmann.me/api/basic_json/operator_spaceship/
+    std::partial_ordering operator<=>(const_reference rhs) const noexcept // *NOPAD*
+    {
+        const_reference lhs = *this;
+        // default_result is used if we cannot compare values. In that case,
+        // we compare types.
+        JSON_IMPLEMENT_OPERATOR(<=>, // *NOPAD*
+                                std::partial_ordering::equivalent,
+                                std::partial_ordering::unordered,
+                                lhs_type <=> rhs_type) // *NOPAD*
+    }
+
+    /// @brief comparison: 3-way
+    /// @sa https://json.nlohmann.me/api/basic_json/operator_spaceship/
+    template<typename ScalarType>
+    requires std::is_scalar_v<ScalarType>
+    std::partial_ordering operator<=>(ScalarType rhs) const noexcept // *NOPAD*
+    {
+        return *this <=> basic_json(rhs); // *NOPAD*
+    }
+
+#if JSON_USE_LEGACY_DISCARDED_VALUE_COMPARISON
+    // all operators that are computed as an odd number of inverses of others
+    // need to be overloaded to emulate the legacy comparison behavior
+
+    /// @brief comparison: less than or equal
+    /// @sa https://json.nlohmann.me/api/basic_json/operator_le/
+    JSON_HEDLEY_DEPRECATED_FOR(3.11.0, undef JSON_USE_LEGACY_DISCARDED_VALUE_COMPARISON)
+    bool operator<=(const_reference rhs) const noexcept
+    {
+        if (compares_unordered(rhs, true))
+        {
+            return false;
+        }
+        return !(rhs < *this);
+    }
+
+    /// @brief comparison: less than or equal
+    /// @sa https://json.nlohmann.me/api/basic_json/operator_le/
+    template<typename ScalarType>
+    requires std::is_scalar_v<ScalarType>
+    bool operator<=(ScalarType rhs) const noexcept
+    {
+        return *this <= basic_json(rhs);
+    }
+
+    /// @brief comparison: greater than or equal
+    /// @sa https://json.nlohmann.me/api/basic_json/operator_ge/
+    JSON_HEDLEY_DEPRECATED_FOR(3.11.0, undef JSON_USE_LEGACY_DISCARDED_VALUE_COMPARISON)
+    bool operator>=(const_reference rhs) const noexcept
+    {
+        if (compares_unordered(rhs, true))
+        {
+            return false;
+        }
+        return !(*this < rhs);
+    }
+
+    /// @brief comparison: greater than or equal
+    /// @sa https://json.nlohmann.me/api/basic_json/operator_ge/
+    template<typename ScalarType>
+    requires std::is_scalar_v<ScalarType>
+    bool operator>=(ScalarType rhs) const noexcept
+    {
+        return *this >= basic_json(rhs);
+    }
+#endif
+#else
+    /// @brief comparison: equal
+    /// @sa https://json.nlohmann.me/api/basic_json/operator_eq/
+    friend bool operator==(const_reference lhs, const_reference rhs) noexcept
+    {
+#ifdef __GNUC__
+#pragma GCC diagnostic push
+#pragma GCC diagnostic ignored "-Wfloat-equal"
+#endif
+        JSON_IMPLEMENT_OPERATOR( ==, true, false, false)
+#ifdef __GNUC__
+#pragma GCC diagnostic pop
+#endif
+    }
+
+    /// @brief comparison: equal
+    /// @sa https://json.nlohmann.me/api/basic_json/operator_eq/
+    template<typename ScalarType, typename std::enable_if<
+                 std::is_scalar<ScalarType>::value, int>::type = 0>
+    friend bool operator==(const_reference lhs, ScalarType rhs) noexcept
+    {
+        return lhs == basic_json(rhs);
+    }
+
+    /// @brief comparison: equal
+    /// @sa https://json.nlohmann.me/api/basic_json/operator_eq/
+    template<typename ScalarType, typename std::enable_if<
+                 std::is_scalar<ScalarType>::value, int>::type = 0>
+    friend bool operator==(ScalarType lhs, const_reference rhs) noexcept
+    {
+        return basic_json(lhs) == rhs;
+    }
+
+    /// @brief comparison: not equal
+    /// @sa https://json.nlohmann.me/api/basic_json/operator_ne/
+    friend bool operator!=(const_reference lhs, const_reference rhs) noexcept
+    {
+        if (compares_unordered(lhs, rhs, true))
+        {
+            return false;
+        }
+        return !(lhs == rhs);
+    }
+
+    /// @brief comparison: not equal
+    /// @sa https://json.nlohmann.me/api/basic_json/operator_ne/
+    template<typename ScalarType, typename std::enable_if<
+                 std::is_scalar<ScalarType>::value, int>::type = 0>
+    friend bool operator!=(const_reference lhs, ScalarType rhs) noexcept
+    {
+        return lhs != basic_json(rhs);
+    }
+
+    /// @brief comparison: not equal
+    /// @sa https://json.nlohmann.me/api/basic_json/operator_ne/
+    template<typename ScalarType, typename std::enable_if<
+                 std::is_scalar<ScalarType>::value, int>::type = 0>
+    friend bool operator!=(ScalarType lhs, const_reference rhs) noexcept
+    {
+        return basic_json(lhs) != rhs;
+    }
+
+    /// @brief comparison: less than
+    /// @sa https://json.nlohmann.me/api/basic_json/operator_lt/
+    friend bool operator<(const_reference lhs, const_reference rhs) noexcept
+    {
+        // default_result is used if we cannot compare values. In that case,
+        // we compare types. Note we have to call the operator explicitly,
+        // because MSVC has problems otherwise.
+        JSON_IMPLEMENT_OPERATOR( <, false, false, operator<(lhs_type, rhs_type))
+    }
+
+    /// @brief comparison: less than
+    /// @sa https://json.nlohmann.me/api/basic_json/operator_lt/
+    template<typename ScalarType, typename std::enable_if<
+                 std::is_scalar<ScalarType>::value, int>::type = 0>
+    friend bool operator<(const_reference lhs, ScalarType rhs) noexcept
+    {
+        return lhs < basic_json(rhs);
+    }
+
+    /// @brief comparison: less than
+    /// @sa https://json.nlohmann.me/api/basic_json/operator_lt/
+    template<typename ScalarType, typename std::enable_if<
+                 std::is_scalar<ScalarType>::value, int>::type = 0>
+    friend bool operator<(ScalarType lhs, const_reference rhs) noexcept
+    {
+        return basic_json(lhs) < rhs;
+    }
+
+    /// @brief comparison: less than or equal
+    /// @sa https://json.nlohmann.me/api/basic_json/operator_le/
+    friend bool operator<=(const_reference lhs, const_reference rhs) noexcept
+    {
+        if (compares_unordered(lhs, rhs, true))
+        {
+            return false;
+        }
+        return !(rhs < lhs);
+    }
+
+    /// @brief comparison: less than or equal
+    /// @sa https://json.nlohmann.me/api/basic_json/operator_le/
+    template<typename ScalarType, typename std::enable_if<
+                 std::is_scalar<ScalarType>::value, int>::type = 0>
+    friend bool operator<=(const_reference lhs, ScalarType rhs) noexcept
+    {
+        return lhs <= basic_json(rhs);
+    }
+
+    /// @brief comparison: less than or equal
+    /// @sa https://json.nlohmann.me/api/basic_json/operator_le/
+    template<typename ScalarType, typename std::enable_if<
+                 std::is_scalar<ScalarType>::value, int>::type = 0>
+    friend bool operator<=(ScalarType lhs, const_reference rhs) noexcept
+    {
+        return basic_json(lhs) <= rhs;
+    }
+
+    /// @brief comparison: greater than
+    /// @sa https://json.nlohmann.me/api/basic_json/operator_gt/
+    friend bool operator>(const_reference lhs, const_reference rhs) noexcept
+    {
+        // double inverse
+        if (compares_unordered(lhs, rhs))
+        {
+            return false;
+        }
+        return !(lhs <= rhs);
+    }
+
+    /// @brief comparison: greater than
+    /// @sa https://json.nlohmann.me/api/basic_json/operator_gt/
+    template<typename ScalarType, typename std::enable_if<
+                 std::is_scalar<ScalarType>::value, int>::type = 0>
+    friend bool operator>(const_reference lhs, ScalarType rhs) noexcept
+    {
+        return lhs > basic_json(rhs);
+    }
+
+    /// @brief comparison: greater than
+    /// @sa https://json.nlohmann.me/api/basic_json/operator_gt/
+    template<typename ScalarType, typename std::enable_if<
+                 std::is_scalar<ScalarType>::value, int>::type = 0>
+    friend bool operator>(ScalarType lhs, const_reference rhs) noexcept
+    {
+        return basic_json(lhs) > rhs;
+    }
+
+    /// @brief comparison: greater than or equal
+    /// @sa https://json.nlohmann.me/api/basic_json/operator_ge/
+    friend bool operator>=(const_reference lhs, const_reference rhs) noexcept
+    {
+        if (compares_unordered(lhs, rhs, true))
+        {
+            return false;
+        }
+        return !(lhs < rhs);
+    }
+
+    /// @brief comparison: greater than or equal
+    /// @sa https://json.nlohmann.me/api/basic_json/operator_ge/
+    template<typename ScalarType, typename std::enable_if<
+                 std::is_scalar<ScalarType>::value, int>::type = 0>
+    friend bool operator>=(const_reference lhs, ScalarType rhs) noexcept
+    {
+        return lhs >= basic_json(rhs);
+    }
+
+    /// @brief comparison: greater than or equal
+    /// @sa https://json.nlohmann.me/api/basic_json/operator_ge/
+    template<typename ScalarType, typename std::enable_if<
+                 std::is_scalar<ScalarType>::value, int>::type = 0>
+    friend bool operator>=(ScalarType lhs, const_reference rhs) noexcept
+    {
+        return basic_json(lhs) >= rhs;
+    }
+#endif
+
+#undef JSON_IMPLEMENT_OPERATOR
+
+    /// @}
+
+    ///////////////////
+    // serialization //
+    ///////////////////
+
+    /// @name serialization
+    /// @{
+#ifndef JSON_NO_IO
+    /// @brief serialize to stream
+    /// @sa https://json.nlohmann.me/api/basic_json/operator_ltlt/
+    friend std::ostream& operator<<(std::ostream& o, const basic_json& j)
+    {
+        // read width member and use it as indentation parameter if nonzero
+        const bool pretty_print = o.width() > 0;
+        const auto indentation = pretty_print ? o.width() : 0;
+
+        // reset width to 0 for subsequent calls to this stream
+        o.width(0);
+
+        // do the actual serialization
+        serializer s(detail::output_adapter<char>(o), o.fill());
+        s.dump(j, pretty_print, false, static_cast<unsigned int>(indentation));
+        return o;
+    }
+
+    /// @brief serialize to stream
+    /// @sa https://json.nlohmann.me/api/basic_json/operator_ltlt/
+    /// @deprecated This function is deprecated since 3.0.0 and will be removed in
+    ///             version 4.0.0 of the library. Please use
+    ///             operator<<(std::ostream&, const basic_json&) instead; that is,
+    ///             replace calls like `j >> o;` with `o << j;`.
+    JSON_HEDLEY_DEPRECATED_FOR(3.0.0, operator<<(std::ostream&, const basic_json&))
+    friend std::ostream& operator>>(const basic_json& j, std::ostream& o)
+    {
+        return o << j;
+    }
+#endif  // JSON_NO_IO
+    /// @}
+
+
+    /////////////////////
+    // deserialization //
+    /////////////////////
+
+    /// @name deserialization
+    /// @{
+
+    /// @brief deserialize from a compatible input
+    /// @sa https://json.nlohmann.me/api/basic_json/parse/
+    template<typename InputType>
+    JSON_HEDLEY_WARN_UNUSED_RESULT
+    static basic_json parse(InputType&& i,
+                            const parser_callback_t cb = nullptr,
+                            const bool allow_exceptions = true,
+                            const bool ignore_comments = false)
+    {
+        basic_json result;
+        parser(detail::input_adapter(std::forward<InputType>(i)), cb, allow_exceptions, ignore_comments).parse(true, result);
+        return result;
+    }
+
+    /// @brief deserialize from a pair of character iterators
+    /// @sa https://json.nlohmann.me/api/basic_json/parse/
+    template<typename IteratorType>
+    JSON_HEDLEY_WARN_UNUSED_RESULT
+    static basic_json parse(IteratorType first,
+                            IteratorType last,
+                            const parser_callback_t cb = nullptr,
+                            const bool allow_exceptions = true,
+                            const bool ignore_comments = false)
+    {
+        basic_json result;
+        parser(detail::input_adapter(std::move(first), std::move(last)), cb, allow_exceptions, ignore_comments).parse(true, result);
+        return result;
+    }
+
+    JSON_HEDLEY_WARN_UNUSED_RESULT
+    JSON_HEDLEY_DEPRECATED_FOR(3.8.0, parse(ptr, ptr + len))
+    static basic_json parse(detail::span_input_adapter&& i,
+                            const parser_callback_t cb = nullptr,
+                            const bool allow_exceptions = true,
+                            const bool ignore_comments = false)
+    {
+        basic_json result;
+        parser(i.get(), cb, allow_exceptions, ignore_comments).parse(true, result);
+        return result;
+    }
+
+    /// @brief check if the input is valid JSON
+    /// @sa https://json.nlohmann.me/api/basic_json/accept/
+    template<typename InputType>
+    static bool accept(InputType&& i,
+                       const bool ignore_comments = false)
+    {
+        return parser(detail::input_adapter(std::forward<InputType>(i)), nullptr, false, ignore_comments).accept(true);
+    }
+
+    /// @brief check if the input is valid JSON
+    /// @sa https://json.nlohmann.me/api/basic_json/accept/
+    template<typename IteratorType>
+    static bool accept(IteratorType first, IteratorType last,
+                       const bool ignore_comments = false)
+    {
+        return parser(detail::input_adapter(std::move(first), std::move(last)), nullptr, false, ignore_comments).accept(true);
+    }
+
+    JSON_HEDLEY_WARN_UNUSED_RESULT
+    JSON_HEDLEY_DEPRECATED_FOR(3.8.0, accept(ptr, ptr + len))
+    static bool accept(detail::span_input_adapter&& i,
+                       const bool ignore_comments = false)
+    {
+        return parser(i.get(), nullptr, false, ignore_comments).accept(true);
+    }
+
+    /// @brief generate SAX events
+    /// @sa https://json.nlohmann.me/api/basic_json/sax_parse/
+    template <typename InputType, typename SAX>
+    JSON_HEDLEY_NON_NULL(2)
+    static bool sax_parse(InputType&& i, SAX* sax,
+                          input_format_t format = input_format_t::json,
+                          const bool strict = true,
+                          const bool ignore_comments = false)
+    {
+        auto ia = detail::input_adapter(std::forward<InputType>(i));
+        return format == input_format_t::json
+               ? parser(std::move(ia), nullptr, true, ignore_comments).sax_parse(sax, strict)
+               : detail::binary_reader<basic_json, decltype(ia), SAX>(std::move(ia), format).sax_parse(format, sax, strict);
+    }
+
+    /// @brief generate SAX events
+    /// @sa https://json.nlohmann.me/api/basic_json/sax_parse/
+    template<class IteratorType, class SAX>
+    JSON_HEDLEY_NON_NULL(3)
+    static bool sax_parse(IteratorType first, IteratorType last, SAX* sax,
+                          input_format_t format = input_format_t::json,
+                          const bool strict = true,
+                          const bool ignore_comments = false)
+    {
+        auto ia = detail::input_adapter(std::move(first), std::move(last));
+        return format == input_format_t::json
+               ? parser(std::move(ia), nullptr, true, ignore_comments).sax_parse(sax, strict)
+               : detail::binary_reader<basic_json, decltype(ia), SAX>(std::move(ia), format).sax_parse(format, sax, strict);
+    }
+
+    /// @brief generate SAX events
+    /// @sa https://json.nlohmann.me/api/basic_json/sax_parse/
+    /// @deprecated This function is deprecated since 3.8.0 and will be removed in
+    ///             version 4.0.0 of the library. Please use
+    ///             sax_parse(ptr, ptr + len) instead.
+    template <typename SAX>
+    JSON_HEDLEY_DEPRECATED_FOR(3.8.0, sax_parse(ptr, ptr + len, ...))
+    JSON_HEDLEY_NON_NULL(2)
+    static bool sax_parse(detail::span_input_adapter&& i, SAX* sax,
+                          input_format_t format = input_format_t::json,
+                          const bool strict = true,
+                          const bool ignore_comments = false)
+    {
+        auto ia = i.get();
+        return format == input_format_t::json
+               // NOLINTNEXTLINE(hicpp-move-const-arg,performance-move-const-arg)
+               ? parser(std::move(ia), nullptr, true, ignore_comments).sax_parse(sax, strict)
+               // NOLINTNEXTLINE(hicpp-move-const-arg,performance-move-const-arg)
+               : detail::binary_reader<basic_json, decltype(ia), SAX>(std::move(ia), format).sax_parse(format, sax, strict);
+    }
+#ifndef JSON_NO_IO
+    /// @brief deserialize from stream
+    /// @sa https://json.nlohmann.me/api/basic_json/operator_gtgt/
+    /// @deprecated This stream operator is deprecated since 3.0.0 and will be removed in
+    ///             version 4.0.0 of the library. Please use
+    ///             operator>>(std::istream&, basic_json&) instead; that is,
+    ///             replace calls like `j << i;` with `i >> j;`.
+    JSON_HEDLEY_DEPRECATED_FOR(3.0.0, operator>>(std::istream&, basic_json&))
+    friend std::istream& operator<<(basic_json& j, std::istream& i)
+    {
+        return operator>>(i, j);
+    }
+
+    /// @brief deserialize from stream
+    /// @sa https://json.nlohmann.me/api/basic_json/operator_gtgt/
+    friend std::istream& operator>>(std::istream& i, basic_json& j)
+    {
+        parser(detail::input_adapter(i)).parse(false, j);
+        return i;
+    }
+#endif  // JSON_NO_IO
+    /// @}
+
+    ///////////////////////////
+    // convenience functions //
+    ///////////////////////////
+
+    /// @brief return the type as string
+    /// @sa https://json.nlohmann.me/api/basic_json/type_name/
+    JSON_HEDLEY_RETURNS_NON_NULL
+    const char* type_name() const noexcept
+    {
+        switch (m_type)
+        {
+            case value_t::null:
+                return "null";
+            case value_t::object:
+                return "object";
+            case value_t::array:
+                return "array";
+            case value_t::string:
+                return "string";
+            case value_t::boolean:
+                return "boolean";
+            case value_t::binary:
+                return "binary";
+            case value_t::discarded:
+                return "discarded";
+            case value_t::number_integer:
+            case value_t::number_unsigned:
+            case value_t::number_float:
+            default:
+                return "number";
+        }
+    }
+
+
+  JSON_PRIVATE_UNLESS_TESTED:
+    //////////////////////
+    // member variables //
+    //////////////////////
+
+    /// the type of the current element
+    value_t m_type = value_t::null;
+
+    /// the value of the current element
+    json_value m_value = {};
+
+#if JSON_DIAGNOSTICS
+    /// a pointer to a parent value (for debugging purposes)
+    basic_json* m_parent = nullptr;
+#endif
+
+    //////////////////////////////////////////
+    // binary serialization/deserialization //
+    //////////////////////////////////////////
+
+    /// @name binary serialization/deserialization support
+    /// @{
+
+  public:
+    /// @brief create a CBOR serialization of a given JSON value
+    /// @sa https://json.nlohmann.me/api/basic_json/to_cbor/
+    static std::vector<std::uint8_t> to_cbor(const basic_json& j)
+    {
+        std::vector<std::uint8_t> result;
+        to_cbor(j, result);
+        return result;
+    }
+
+    /// @brief create a CBOR serialization of a given JSON value
+    /// @sa https://json.nlohmann.me/api/basic_json/to_cbor/
+    static void to_cbor(const basic_json& j, detail::output_adapter<std::uint8_t> o)
+    {
+        binary_writer<std::uint8_t>(o).write_cbor(j);
+    }
+
+    /// @brief create a CBOR serialization of a given JSON value
+    /// @sa https://json.nlohmann.me/api/basic_json/to_cbor/
+    static void to_cbor(const basic_json& j, detail::output_adapter<char> o)
+    {
+        binary_writer<char>(o).write_cbor(j);
+    }
+
+    /// @brief create a MessagePack serialization of a given JSON value
+    /// @sa https://json.nlohmann.me/api/basic_json/to_msgpack/
+    static std::vector<std::uint8_t> to_msgpack(const basic_json& j)
+    {
+        std::vector<std::uint8_t> result;
+        to_msgpack(j, result);
+        return result;
+    }
+
+    /// @brief create a MessagePack serialization of a given JSON value
+    /// @sa https://json.nlohmann.me/api/basic_json/to_msgpack/
+    static void to_msgpack(const basic_json& j, detail::output_adapter<std::uint8_t> o)
+    {
+        binary_writer<std::uint8_t>(o).write_msgpack(j);
+    }
+
+    /// @brief create a MessagePack serialization of a given JSON value
+    /// @sa https://json.nlohmann.me/api/basic_json/to_msgpack/
+    static void to_msgpack(const basic_json& j, detail::output_adapter<char> o)
+    {
+        binary_writer<char>(o).write_msgpack(j);
+    }
+
+    /// @brief create a UBJSON serialization of a given JSON value
+    /// @sa https://json.nlohmann.me/api/basic_json/to_ubjson/
+    static std::vector<std::uint8_t> to_ubjson(const basic_json& j,
+            const bool use_size = false,
+            const bool use_type = false)
+    {
+        std::vector<std::uint8_t> result;
+        to_ubjson(j, result, use_size, use_type);
+        return result;
+    }
+
+    /// @brief create a UBJSON serialization of a given JSON value
+    /// @sa https://json.nlohmann.me/api/basic_json/to_ubjson/
+    static void to_ubjson(const basic_json& j, detail::output_adapter<std::uint8_t> o,
+                          const bool use_size = false, const bool use_type = false)
+    {
+        binary_writer<std::uint8_t>(o).write_ubjson(j, use_size, use_type);
+    }
+
+    /// @brief create a UBJSON serialization of a given JSON value
+    /// @sa https://json.nlohmann.me/api/basic_json/to_ubjson/
+    static void to_ubjson(const basic_json& j, detail::output_adapter<char> o,
+                          const bool use_size = false, const bool use_type = false)
+    {
+        binary_writer<char>(o).write_ubjson(j, use_size, use_type);
+    }
+
+    /// @brief create a BJData serialization of a given JSON value
+    /// @sa https://json.nlohmann.me/api/basic_json/to_bjdata/
+    static std::vector<std::uint8_t> to_bjdata(const basic_json& j,
+            const bool use_size = false,
+            const bool use_type = false)
+    {
+        std::vector<std::uint8_t> result;
+        to_bjdata(j, result, use_size, use_type);
+        return result;
+    }
+
+    /// @brief create a BJData serialization of a given JSON value
+    /// @sa https://json.nlohmann.me/api/basic_json/to_bjdata/
+    static void to_bjdata(const basic_json& j, detail::output_adapter<std::uint8_t> o,
+                          const bool use_size = false, const bool use_type = false)
+    {
+        binary_writer<std::uint8_t>(o).write_ubjson(j, use_size, use_type, true, true);
+    }
+
+    /// @brief create a BJData serialization of a given JSON value
+    /// @sa https://json.nlohmann.me/api/basic_json/to_bjdata/
+    static void to_bjdata(const basic_json& j, detail::output_adapter<char> o,
+                          const bool use_size = false, const bool use_type = false)
+    {
+        binary_writer<char>(o).write_ubjson(j, use_size, use_type, true, true);
+    }
+
+    /// @brief create a BSON serialization of a given JSON value
+    /// @sa https://json.nlohmann.me/api/basic_json/to_bson/
+    static std::vector<std::uint8_t> to_bson(const basic_json& j)
+    {
+        std::vector<std::uint8_t> result;
+        to_bson(j, result);
+        return result;
+    }
+
+    /// @brief create a BSON serialization of a given JSON value
+    /// @sa https://json.nlohmann.me/api/basic_json/to_bson/
+    static void to_bson(const basic_json& j, detail::output_adapter<std::uint8_t> o)
+    {
+        binary_writer<std::uint8_t>(o).write_bson(j);
+    }
+
+    /// @brief create a BSON serialization of a given JSON value
+    /// @sa https://json.nlohmann.me/api/basic_json/to_bson/
+    static void to_bson(const basic_json& j, detail::output_adapter<char> o)
+    {
+        binary_writer<char>(o).write_bson(j);
+    }
+
+    /// @brief create a JSON value from an input in CBOR format
+    /// @sa https://json.nlohmann.me/api/basic_json/from_cbor/
+    template<typename InputType>
+    JSON_HEDLEY_WARN_UNUSED_RESULT
+    static basic_json from_cbor(InputType&& i,
+                                const bool strict = true,
+                                const bool allow_exceptions = true,
+                                const cbor_tag_handler_t tag_handler = cbor_tag_handler_t::error)
+    {
+        basic_json result;
+        detail::json_sax_dom_parser<basic_json> sdp(result, allow_exceptions);
+        auto ia = detail::input_adapter(std::forward<InputType>(i));
+        const bool res = binary_reader<decltype(ia)>(std::move(ia), input_format_t::cbor).sax_parse(input_format_t::cbor, &sdp, strict, tag_handler);
+        return res ? result : basic_json(value_t::discarded);
+    }
+
+    /// @brief create a JSON value from an input in CBOR format
+    /// @sa https://json.nlohmann.me/api/basic_json/from_cbor/
+    template<typename IteratorType>
+    JSON_HEDLEY_WARN_UNUSED_RESULT
+    static basic_json from_cbor(IteratorType first, IteratorType last,
+                                const bool strict = true,
+                                const bool allow_exceptions = true,
+                                const cbor_tag_handler_t tag_handler = cbor_tag_handler_t::error)
+    {
+        basic_json result;
+        detail::json_sax_dom_parser<basic_json> sdp(result, allow_exceptions);
+        auto ia = detail::input_adapter(std::move(first), std::move(last));
+        const bool res = binary_reader<decltype(ia)>(std::move(ia), input_format_t::cbor).sax_parse(input_format_t::cbor, &sdp, strict, tag_handler);
+        return res ? result : basic_json(value_t::discarded);
+    }
+
+    template<typename T>
+    JSON_HEDLEY_WARN_UNUSED_RESULT
+    JSON_HEDLEY_DEPRECATED_FOR(3.8.0, from_cbor(ptr, ptr + len))
+    static basic_json from_cbor(const T* ptr, std::size_t len,
+                                const bool strict = true,
+                                const bool allow_exceptions = true,
+                                const cbor_tag_handler_t tag_handler = cbor_tag_handler_t::error)
+    {
+        return from_cbor(ptr, ptr + len, strict, allow_exceptions, tag_handler);
+    }
+
+
+    JSON_HEDLEY_WARN_UNUSED_RESULT
+    JSON_HEDLEY_DEPRECATED_FOR(3.8.0, from_cbor(ptr, ptr + len))
+    static basic_json from_cbor(detail::span_input_adapter&& i,
+                                const bool strict = true,
+                                const bool allow_exceptions = true,
+                                const cbor_tag_handler_t tag_handler = cbor_tag_handler_t::error)
+    {
+        basic_json result;
+        detail::json_sax_dom_parser<basic_json> sdp(result, allow_exceptions);
+        auto ia = i.get();
+        // NOLINTNEXTLINE(hicpp-move-const-arg,performance-move-const-arg)
+        const bool res = binary_reader<decltype(ia)>(std::move(ia), input_format_t::cbor).sax_parse(input_format_t::cbor, &sdp, strict, tag_handler);
+        return res ? result : basic_json(value_t::discarded);
+    }
+
+    /// @brief create a JSON value from an input in MessagePack format
+    /// @sa https://json.nlohmann.me/api/basic_json/from_msgpack/
+    template<typename InputType>
+    JSON_HEDLEY_WARN_UNUSED_RESULT
+    static basic_json from_msgpack(InputType&& i,
+                                   const bool strict = true,
+                                   const bool allow_exceptions = true)
+    {
+        basic_json result;
+        detail::json_sax_dom_parser<basic_json> sdp(result, allow_exceptions);
+        auto ia = detail::input_adapter(std::forward<InputType>(i));
+        const bool res = binary_reader<decltype(ia)>(std::move(ia), input_format_t::msgpack).sax_parse(input_format_t::msgpack, &sdp, strict);
+        return res ? result : basic_json(value_t::discarded);
+    }
+
+    /// @brief create a JSON value from an input in MessagePack format
+    /// @sa https://json.nlohmann.me/api/basic_json/from_msgpack/
+    template<typename IteratorType>
+    JSON_HEDLEY_WARN_UNUSED_RESULT
+    static basic_json from_msgpack(IteratorType first, IteratorType last,
+                                   const bool strict = true,
+                                   const bool allow_exceptions = true)
+    {
+        basic_json result;
+        detail::json_sax_dom_parser<basic_json> sdp(result, allow_exceptions);
+        auto ia = detail::input_adapter(std::move(first), std::move(last));
+        const bool res = binary_reader<decltype(ia)>(std::move(ia), input_format_t::msgpack).sax_parse(input_format_t::msgpack, &sdp, strict);
+        return res ? result : basic_json(value_t::discarded);
+    }
+
+    template<typename T>
+    JSON_HEDLEY_WARN_UNUSED_RESULT
+    JSON_HEDLEY_DEPRECATED_FOR(3.8.0, from_msgpack(ptr, ptr + len))
+    static basic_json from_msgpack(const T* ptr, std::size_t len,
+                                   const bool strict = true,
+                                   const bool allow_exceptions = true)
+    {
+        return from_msgpack(ptr, ptr + len, strict, allow_exceptions);
+    }
+
+    JSON_HEDLEY_WARN_UNUSED_RESULT
+    JSON_HEDLEY_DEPRECATED_FOR(3.8.0, from_msgpack(ptr, ptr + len))
+    static basic_json from_msgpack(detail::span_input_adapter&& i,
+                                   const bool strict = true,
+                                   const bool allow_exceptions = true)
+    {
+        basic_json result;
+        detail::json_sax_dom_parser<basic_json> sdp(result, allow_exceptions);
+        auto ia = i.get();
+        // NOLINTNEXTLINE(hicpp-move-const-arg,performance-move-const-arg)
+        const bool res = binary_reader<decltype(ia)>(std::move(ia), input_format_t::msgpack).sax_parse(input_format_t::msgpack, &sdp, strict);
+        return res ? result : basic_json(value_t::discarded);
+    }
+
+    /// @brief create a JSON value from an input in UBJSON format
+    /// @sa https://json.nlohmann.me/api/basic_json/from_ubjson/
+    template<typename InputType>
+    JSON_HEDLEY_WARN_UNUSED_RESULT
+    static basic_json from_ubjson(InputType&& i,
+                                  const bool strict = true,
+                                  const bool allow_exceptions = true)
+    {
+        basic_json result;
+        detail::json_sax_dom_parser<basic_json> sdp(result, allow_exceptions);
+        auto ia = detail::input_adapter(std::forward<InputType>(i));
+        const bool res = binary_reader<decltype(ia)>(std::move(ia), input_format_t::ubjson).sax_parse(input_format_t::ubjson, &sdp, strict);
+        return res ? result : basic_json(value_t::discarded);
+    }
+
+    /// @brief create a JSON value from an input in UBJSON format
+    /// @sa https://json.nlohmann.me/api/basic_json/from_ubjson/
+    template<typename IteratorType>
+    JSON_HEDLEY_WARN_UNUSED_RESULT
+    static basic_json from_ubjson(IteratorType first, IteratorType last,
+                                  const bool strict = true,
+                                  const bool allow_exceptions = true)
+    {
+        basic_json result;
+        detail::json_sax_dom_parser<basic_json> sdp(result, allow_exceptions);
+        auto ia = detail::input_adapter(std::move(first), std::move(last));
+        const bool res = binary_reader<decltype(ia)>(std::move(ia), input_format_t::ubjson).sax_parse(input_format_t::ubjson, &sdp, strict);
+        return res ? result : basic_json(value_t::discarded);
+    }
+
+    template<typename T>
+    JSON_HEDLEY_WARN_UNUSED_RESULT
+    JSON_HEDLEY_DEPRECATED_FOR(3.8.0, from_ubjson(ptr, ptr + len))
+    static basic_json from_ubjson(const T* ptr, std::size_t len,
+                                  const bool strict = true,
+                                  const bool allow_exceptions = true)
+    {
+        return from_ubjson(ptr, ptr + len, strict, allow_exceptions);
+    }
+
+    JSON_HEDLEY_WARN_UNUSED_RESULT
+    JSON_HEDLEY_DEPRECATED_FOR(3.8.0, from_ubjson(ptr, ptr + len))
+    static basic_json from_ubjson(detail::span_input_adapter&& i,
+                                  const bool strict = true,
+                                  const bool allow_exceptions = true)
+    {
+        basic_json result;
+        detail::json_sax_dom_parser<basic_json> sdp(result, allow_exceptions);
+        auto ia = i.get();
+        // NOLINTNEXTLINE(hicpp-move-const-arg,performance-move-const-arg)
+        const bool res = binary_reader<decltype(ia)>(std::move(ia), input_format_t::ubjson).sax_parse(input_format_t::ubjson, &sdp, strict);
+        return res ? result : basic_json(value_t::discarded);
+    }
+
+
+    /// @brief create a JSON value from an input in BJData format
+    /// @sa https://json.nlohmann.me/api/basic_json/from_bjdata/
+    template<typename InputType>
+    JSON_HEDLEY_WARN_UNUSED_RESULT
+    static basic_json from_bjdata(InputType&& i,
+                                  const bool strict = true,
+                                  const bool allow_exceptions = true)
+    {
+        basic_json result;
+        detail::json_sax_dom_parser<basic_json> sdp(result, allow_exceptions);
+        auto ia = detail::input_adapter(std::forward<InputType>(i));
+        const bool res = binary_reader<decltype(ia)>(std::move(ia), input_format_t::bjdata).sax_parse(input_format_t::bjdata, &sdp, strict);
+        return res ? result : basic_json(value_t::discarded);
+    }
+
+    /// @brief create a JSON value from an input in BJData format
+    /// @sa https://json.nlohmann.me/api/basic_json/from_bjdata/
+    template<typename IteratorType>
+    JSON_HEDLEY_WARN_UNUSED_RESULT
+    static basic_json from_bjdata(IteratorType first, IteratorType last,
+                                  const bool strict = true,
+                                  const bool allow_exceptions = true)
+    {
+        basic_json result;
+        detail::json_sax_dom_parser<basic_json> sdp(result, allow_exceptions);
+        auto ia = detail::input_adapter(std::move(first), std::move(last));
+        const bool res = binary_reader<decltype(ia)>(std::move(ia), input_format_t::bjdata).sax_parse(input_format_t::bjdata, &sdp, strict);
+        return res ? result : basic_json(value_t::discarded);
+    }
+
+    /// @brief create a JSON value from an input in BSON format
+    /// @sa https://json.nlohmann.me/api/basic_json/from_bson/
+    template<typename InputType>
+    JSON_HEDLEY_WARN_UNUSED_RESULT
+    static basic_json from_bson(InputType&& i,
+                                const bool strict = true,
+                                const bool allow_exceptions = true)
+    {
+        basic_json result;
+        detail::json_sax_dom_parser<basic_json> sdp(result, allow_exceptions);
+        auto ia = detail::input_adapter(std::forward<InputType>(i));
+        const bool res = binary_reader<decltype(ia)>(std::move(ia), input_format_t::bson).sax_parse(input_format_t::bson, &sdp, strict);
+        return res ? result : basic_json(value_t::discarded);
+    }
+
+    /// @brief create a JSON value from an input in BSON format
+    /// @sa https://json.nlohmann.me/api/basic_json/from_bson/
+    template<typename IteratorType>
+    JSON_HEDLEY_WARN_UNUSED_RESULT
+    static basic_json from_bson(IteratorType first, IteratorType last,
+                                const bool strict = true,
+                                const bool allow_exceptions = true)
+    {
+        basic_json result;
+        detail::json_sax_dom_parser<basic_json> sdp(result, allow_exceptions);
+        auto ia = detail::input_adapter(std::move(first), std::move(last));
+        const bool res = binary_reader<decltype(ia)>(std::move(ia), input_format_t::bson).sax_parse(input_format_t::bson, &sdp, strict);
+        return res ? result : basic_json(value_t::discarded);
+    }
+
+    template<typename T>
+    JSON_HEDLEY_WARN_UNUSED_RESULT
+    JSON_HEDLEY_DEPRECATED_FOR(3.8.0, from_bson(ptr, ptr + len))
+    static basic_json from_bson(const T* ptr, std::size_t len,
+                                const bool strict = true,
+                                const bool allow_exceptions = true)
+    {
+        return from_bson(ptr, ptr + len, strict, allow_exceptions);
+    }
+
+    JSON_HEDLEY_WARN_UNUSED_RESULT
+    JSON_HEDLEY_DEPRECATED_FOR(3.8.0, from_bson(ptr, ptr + len))
+    static basic_json from_bson(detail::span_input_adapter&& i,
+                                const bool strict = true,
+                                const bool allow_exceptions = true)
+    {
+        basic_json result;
+        detail::json_sax_dom_parser<basic_json> sdp(result, allow_exceptions);
+        auto ia = i.get();
+        // NOLINTNEXTLINE(hicpp-move-const-arg,performance-move-const-arg)
+        const bool res = binary_reader<decltype(ia)>(std::move(ia), input_format_t::bson).sax_parse(input_format_t::bson, &sdp, strict);
+        return res ? result : basic_json(value_t::discarded);
+    }
+    /// @}
+
+    //////////////////////////
+    // JSON Pointer support //
+    //////////////////////////
+
+    /// @name JSON Pointer functions
+    /// @{
+
+    /// @brief access specified element via JSON Pointer
+    /// @sa https://json.nlohmann.me/api/basic_json/operator%5B%5D/
+    reference operator[](const json_pointer& ptr)
+    {
+        return ptr.get_unchecked(this);
+    }
+
+    template<typename BasicJsonType, detail::enable_if_t<detail::is_basic_json<BasicJsonType>::value, int> = 0>
+    JSON_HEDLEY_DEPRECATED_FOR(3.11.0, basic_json::json_pointer or nlohmann::json_pointer<basic_json::string_t>) // NOLINT(readability/alt_tokens)
+    reference operator[](const ::nlohmann::json_pointer<BasicJsonType>& ptr)
+    {
+        return ptr.get_unchecked(this);
+    }
+
+    /// @brief access specified element via JSON Pointer
+    /// @sa https://json.nlohmann.me/api/basic_json/operator%5B%5D/
+    const_reference operator[](const json_pointer& ptr) const
+    {
+        return ptr.get_unchecked(this);
+    }
+
+    template<typename BasicJsonType, detail::enable_if_t<detail::is_basic_json<BasicJsonType>::value, int> = 0>
+    JSON_HEDLEY_DEPRECATED_FOR(3.11.0, basic_json::json_pointer or nlohmann::json_pointer<basic_json::string_t>) // NOLINT(readability/alt_tokens)
+    const_reference operator[](const ::nlohmann::json_pointer<BasicJsonType>& ptr) const
+    {
+        return ptr.get_unchecked(this);
+    }
+
+    /// @brief access specified element via JSON Pointer
+    /// @sa https://json.nlohmann.me/api/basic_json/at/
+    reference at(const json_pointer& ptr)
+    {
+        return ptr.get_checked(this);
+    }
+
+    template<typename BasicJsonType, detail::enable_if_t<detail::is_basic_json<BasicJsonType>::value, int> = 0>
+    JSON_HEDLEY_DEPRECATED_FOR(3.11.0, basic_json::json_pointer or nlohmann::json_pointer<basic_json::string_t>) // NOLINT(readability/alt_tokens)
+    reference at(const ::nlohmann::json_pointer<BasicJsonType>& ptr)
+    {
+        return ptr.get_checked(this);
+    }
+
+    /// @brief access specified element via JSON Pointer
+    /// @sa https://json.nlohmann.me/api/basic_json/at/
+    const_reference at(const json_pointer& ptr) const
+    {
+        return ptr.get_checked(this);
+    }
+
+    template<typename BasicJsonType, detail::enable_if_t<detail::is_basic_json<BasicJsonType>::value, int> = 0>
+    JSON_HEDLEY_DEPRECATED_FOR(3.11.0, basic_json::json_pointer or nlohmann::json_pointer<basic_json::string_t>) // NOLINT(readability/alt_tokens)
+    const_reference at(const ::nlohmann::json_pointer<BasicJsonType>& ptr) const
+    {
+        return ptr.get_checked(this);
+    }
+
+    /// @brief return flattened JSON value
+    /// @sa https://json.nlohmann.me/api/basic_json/flatten/
+    basic_json flatten() const
+    {
+        basic_json result(value_t::object);
+        json_pointer::flatten("", *this, result);
+        return result;
+    }
+
+    /// @brief unflatten a previously flattened JSON value
+    /// @sa https://json.nlohmann.me/api/basic_json/unflatten/
+    basic_json unflatten() const
+    {
+        return json_pointer::unflatten(*this);
+    }
+
+    /// @}
+
+    //////////////////////////
+    // JSON Patch functions //
+    //////////////////////////
+
+    /// @name JSON Patch functions
+    /// @{
+
+    /// @brief applies a JSON patch in-place without copying the object
+    /// @sa https://json.nlohmann.me/api/basic_json/patch/
+    void patch_inplace(const basic_json& json_patch)
+    {
+        basic_json& result = *this;
+        // the valid JSON Patch operations
+        enum class patch_operations {add, remove, replace, move, copy, test, invalid};
+
+        const auto get_op = [](const std::string & op)
+        {
+            if (op == "add")
+            {
+                return patch_operations::add;
+            }
+            if (op == "remove")
+            {
+                return patch_operations::remove;
+            }
+            if (op == "replace")
+            {
+                return patch_operations::replace;
+            }
+            if (op == "move")
+            {
+                return patch_operations::move;
+            }
+            if (op == "copy")
+            {
+                return patch_operations::copy;
+            }
+            if (op == "test")
+            {
+                return patch_operations::test;
+            }
+
+            return patch_operations::invalid;
+        };
+
+        // wrapper for "add" operation; add value at ptr
+        const auto operation_add = [&result](json_pointer & ptr, basic_json val)
+        {
+            // adding to the root of the target document means replacing it
+            if (ptr.empty())
+            {
+                result = val;
+                return;
+            }
+
+            // make sure the top element of the pointer exists
+            json_pointer top_pointer = ptr.top();
+            if (top_pointer != ptr)
+            {
+                result.at(top_pointer);
+            }
+
+            // get reference to parent of JSON pointer ptr
+            const auto last_path = ptr.back();
+            ptr.pop_back();
+            // parent must exist when performing patch add per RFC6902 specs
+            basic_json& parent = result.at(ptr);
+
+            switch (parent.m_type)
+            {
+                case value_t::null:
+                case value_t::object:
+                {
+                    // use operator[] to add value
+                    parent[last_path] = val;
+                    break;
+                }
+
+                case value_t::array:
+                {
+                    if (last_path == "-")
+                    {
+                        // special case: append to back
+                        parent.push_back(val);
+                    }
+                    else
+                    {
+                        const auto idx = json_pointer::template array_index<basic_json_t>(last_path);
+                        if (JSON_HEDLEY_UNLIKELY(idx > parent.size()))
+                        {
+                            // avoid undefined behavior
+                            JSON_THROW(out_of_range::create(401, detail::concat("array index ", std::to_string(idx), " is out of range"), &parent));
+                        }
+
+                        // default case: insert add offset
+                        parent.insert(parent.begin() + static_cast<difference_type>(idx), val);
+                    }
+                    break;
+                }
+
+                // if there exists a parent it cannot be primitive
+                case value_t::string: // LCOV_EXCL_LINE
+                case value_t::boolean: // LCOV_EXCL_LINE
+                case value_t::number_integer: // LCOV_EXCL_LINE
+                case value_t::number_unsigned: // LCOV_EXCL_LINE
+                case value_t::number_float: // LCOV_EXCL_LINE
+                case value_t::binary: // LCOV_EXCL_LINE
+                case value_t::discarded: // LCOV_EXCL_LINE
+                default:            // LCOV_EXCL_LINE
+                    JSON_ASSERT(false); // NOLINT(cert-dcl03-c,hicpp-static-assert,misc-static-assert) LCOV_EXCL_LINE
+            }
+        };
+
+        // wrapper for "remove" operation; remove value at ptr
+        const auto operation_remove = [this, &result](json_pointer & ptr)
+        {
+            // get reference to parent of JSON pointer ptr
+            const auto last_path = ptr.back();
+            ptr.pop_back();
+            basic_json& parent = result.at(ptr);
+
+            // remove child
+            if (parent.is_object())
+            {
+                // perform range check
+                auto it = parent.find(last_path);
+                if (JSON_HEDLEY_LIKELY(it != parent.end()))
+                {
+                    parent.erase(it);
+                }
+                else
+                {
+                    JSON_THROW(out_of_range::create(403, detail::concat("key '", last_path, "' not found"), this));
+                }
+            }
+            else if (parent.is_array())
+            {
+                // note erase performs range check
+                parent.erase(json_pointer::template array_index<basic_json_t>(last_path));
+            }
+        };
+
+        // type check: top level value must be an array
+        if (JSON_HEDLEY_UNLIKELY(!json_patch.is_array()))
+        {
+            JSON_THROW(parse_error::create(104, 0, "JSON patch must be an array of objects", &json_patch));
+        }
+
+        // iterate and apply the operations
+        for (const auto& val : json_patch)
+        {
+            // wrapper to get a value for an operation
+            const auto get_value = [&val](const std::string & op,
+                                          const std::string & member,
+                                          bool string_type) -> basic_json &
+            {
+                // find value
+                auto it = val.m_value.object->find(member);
+
+                // context-sensitive error message
+                const auto error_msg = (op == "op") ? "operation" : detail::concat("operation '", op, '\'');
+
+                // check if desired value is present
+                if (JSON_HEDLEY_UNLIKELY(it == val.m_value.object->end()))
+                {
+                    // NOLINTNEXTLINE(performance-inefficient-string-concatenation)
+                    JSON_THROW(parse_error::create(105, 0, detail::concat(error_msg, " must have member '", member, "'"), &val));
+                }
+
+                // check if result is of type string
+                if (JSON_HEDLEY_UNLIKELY(string_type && !it->second.is_string()))
+                {
+                    // NOLINTNEXTLINE(performance-inefficient-string-concatenation)
+                    JSON_THROW(parse_error::create(105, 0, detail::concat(error_msg, " must have string member '", member, "'"), &val));
+                }
+
+                // no error: return value
+                return it->second;
+            };
+
+            // type check: every element of the array must be an object
+            if (JSON_HEDLEY_UNLIKELY(!val.is_object()))
+            {
+                JSON_THROW(parse_error::create(104, 0, "JSON patch must be an array of objects", &val));
+            }
+
+            // collect mandatory members
+            const auto op = get_value("op", "op", true).template get<std::string>();
+            const auto path = get_value(op, "path", true).template get<std::string>();
+            json_pointer ptr(path);
+
+            switch (get_op(op))
+            {
+                case patch_operations::add:
+                {
+                    operation_add(ptr, get_value("add", "value", false));
+                    break;
+                }
+
+                case patch_operations::remove:
+                {
+                    operation_remove(ptr);
+                    break;
+                }
+
+                case patch_operations::replace:
+                {
+                    // the "path" location must exist - use at()
+                    result.at(ptr) = get_value("replace", "value", false);
+                    break;
+                }
+
+                case patch_operations::move:
+                {
+                    const auto from_path = get_value("move", "from", true).template get<std::string>();
+                    json_pointer from_ptr(from_path);
+
+                    // the "from" location must exist - use at()
+                    basic_json v = result.at(from_ptr);
+
+                    // The move operation is functionally identical to a
+                    // "remove" operation on the "from" location, followed
+                    // immediately by an "add" operation at the target
+                    // location with the value that was just removed.
+                    operation_remove(from_ptr);
+                    operation_add(ptr, v);
+                    break;
+                }
+
+                case patch_operations::copy:
+                {
+                    const auto from_path = get_value("copy", "from", true).template get<std::string>();
+                    const json_pointer from_ptr(from_path);
+
+                    // the "from" location must exist - use at()
+                    basic_json v = result.at(from_ptr);
+
+                    // The copy is functionally identical to an "add"
+                    // operation at the target location using the value
+                    // specified in the "from" member.
+                    operation_add(ptr, v);
+                    break;
+                }
+
+                case patch_operations::test:
+                {
+                    bool success = false;
+                    JSON_TRY
+                    {
+                        // check if "value" matches the one at "path"
+                        // the "path" location must exist - use at()
+                        success = (result.at(ptr) == get_value("test", "value", false));
+                    }
+                    JSON_INTERNAL_CATCH (out_of_range&)
+                    {
+                        // ignore out of range errors: success remains false
+                    }
+
+                    // throw an exception if test fails
+                    if (JSON_HEDLEY_UNLIKELY(!success))
+                    {
+                        JSON_THROW(other_error::create(501, detail::concat("unsuccessful: ", val.dump()), &val));
+                    }
+
+                    break;
+                }
+
+                case patch_operations::invalid:
+                default:
+                {
+                    // op must be "add", "remove", "replace", "move", "copy", or
+                    // "test"
+                    JSON_THROW(parse_error::create(105, 0, detail::concat("operation value '", op, "' is invalid"), &val));
+                }
+            }
+        }
+    }
+
+    /// @brief applies a JSON patch to a copy of the current object
+    /// @sa https://json.nlohmann.me/api/basic_json/patch/
+    basic_json patch(const basic_json& json_patch) const
+    {
+        basic_json result = *this;
+        result.patch_inplace(json_patch);
+        return result;
+    }
+
+    /// @brief creates a diff as a JSON patch
+    /// @sa https://json.nlohmann.me/api/basic_json/diff/
+    JSON_HEDLEY_WARN_UNUSED_RESULT
+    static basic_json diff(const basic_json& source, const basic_json& target,
+                           const std::string& path = "")
+    {
+        // the patch
+        basic_json result(value_t::array);
+
+        // if the values are the same, return empty patch
+        if (source == target)
+        {
+            return result;
+        }
+
+        if (source.type() != target.type())
+        {
+            // different types: replace value
+            result.push_back(
+            {
+                {"op", "replace"}, {"path", path}, {"value", target}
+            });
+            return result;
+        }
+
+        switch (source.type())
+        {
+            case value_t::array:
+            {
+                // first pass: traverse common elements
+                std::size_t i = 0;
+                while (i < source.size() && i < target.size())
+                {
+                    // recursive call to compare array values at index i
+                    auto temp_diff = diff(source[i], target[i], detail::concat(path, '/', std::to_string(i)));
+                    result.insert(result.end(), temp_diff.begin(), temp_diff.end());
+                    ++i;
+                }
+
+                // We now reached the end of at least one array
+                // in a second pass, traverse the remaining elements
+
+                // remove my remaining elements
+                const auto end_index = static_cast<difference_type>(result.size());
+                while (i < source.size())
+                {
+                    // add operations in reverse order to avoid invalid
+                    // indices
+                    result.insert(result.begin() + end_index, object(
+                    {
+                        {"op", "remove"},
+                        {"path", detail::concat(path, '/', std::to_string(i))}
+                    }));
+                    ++i;
+                }
+
+                // add other remaining elements
+                while (i < target.size())
+                {
+                    result.push_back(
+                    {
+                        {"op", "add"},
+                        {"path", detail::concat(path, "/-")},
+                        {"value", target[i]}
+                    });
+                    ++i;
+                }
+
+                break;
+            }
+
+            case value_t::object:
+            {
+                // first pass: traverse this object's elements
+                for (auto it = source.cbegin(); it != source.cend(); ++it)
+                {
+                    // escape the key name to be used in a JSON patch
+                    const auto path_key = detail::concat(path, '/', detail::escape(it.key()));
+
+                    if (target.find(it.key()) != target.end())
+                    {
+                        // recursive call to compare object values at key it
+                        auto temp_diff = diff(it.value(), target[it.key()], path_key);
+                        result.insert(result.end(), temp_diff.begin(), temp_diff.end());
+                    }
+                    else
+                    {
+                        // found a key that is not in o -> remove it
+                        result.push_back(object(
+                        {
+                            {"op", "remove"}, {"path", path_key}
+                        }));
+                    }
+                }
+
+                // second pass: traverse other object's elements
+                for (auto it = target.cbegin(); it != target.cend(); ++it)
+                {
+                    if (source.find(it.key()) == source.end())
+                    {
+                        // found a key that is not in this -> add it
+                        const auto path_key = detail::concat(path, '/', detail::escape(it.key()));
+                        result.push_back(
+                        {
+                            {"op", "add"}, {"path", path_key},
+                            {"value", it.value()}
+                        });
+                    }
+                }
+
+                break;
+            }
+
+            case value_t::null:
+            case value_t::string:
+            case value_t::boolean:
+            case value_t::number_integer:
+            case value_t::number_unsigned:
+            case value_t::number_float:
+            case value_t::binary:
+            case value_t::discarded:
+            default:
+            {
+                // both primitive type: replace value
+                result.push_back(
+                {
+                    {"op", "replace"}, {"path", path}, {"value", target}
+                });
+                break;
+            }
+        }
+
+        return result;
+    }
+    /// @}
+
+    ////////////////////////////////
+    // JSON Merge Patch functions //
+    ////////////////////////////////
+
+    /// @name JSON Merge Patch functions
+    /// @{
+
+    /// @brief applies a JSON Merge Patch
+    /// @sa https://json.nlohmann.me/api/basic_json/merge_patch/
+    void merge_patch(const basic_json& apply_patch)
+    {
+        if (apply_patch.is_object())
+        {
+            if (!is_object())
+            {
+                *this = object();
+            }
+            for (auto it = apply_patch.begin(); it != apply_patch.end(); ++it)
+            {
+                if (it.value().is_null())
+                {
+                    erase(it.key());
+                }
+                else
+                {
+                    operator[](it.key()).merge_patch(it.value());
+                }
+            }
+        }
+        else
+        {
+            *this = apply_patch;
+        }
+    }
+
+    /// @}
+};
+
+/// @brief user-defined to_string function for JSON values
+/// @sa https://json.nlohmann.me/api/basic_json/to_string/
+NLOHMANN_BASIC_JSON_TPL_DECLARATION
+std::string to_string(const NLOHMANN_BASIC_JSON_TPL& j)
+{
+    return j.dump();
+}
+
+inline namespace literals
+{
+inline namespace json_literals
+{
+
+/// @brief user-defined string literal for JSON values
+/// @sa https://json.nlohmann.me/api/basic_json/operator_literal_json/
+JSON_HEDLEY_NON_NULL(1)
+inline nlohmann::json operator "" _json(const char* s, std::size_t n)
+{
+    return nlohmann::json::parse(s, s + n);
+}
+
+/// @brief user-defined string literal for JSON pointer
+/// @sa https://json.nlohmann.me/api/basic_json/operator_literal_json_pointer/
+JSON_HEDLEY_NON_NULL(1)
+inline nlohmann::json::json_pointer operator "" _json_pointer(const char* s, std::size_t n)
+{
+    return nlohmann::json::json_pointer(std::string(s, n));
+}
+
+}  // namespace json_literals
+}  // namespace literals
+NLOHMANN_JSON_NAMESPACE_END
+
+///////////////////////
+// nonmember support //
+///////////////////////
+
+namespace std // NOLINT(cert-dcl58-cpp)
+{
+
+/// @brief hash value for JSON objects
+/// @sa https://json.nlohmann.me/api/basic_json/std_hash/
+NLOHMANN_BASIC_JSON_TPL_DECLARATION
+struct hash<nlohmann::NLOHMANN_BASIC_JSON_TPL>
+{
+    std::size_t operator()(const nlohmann::NLOHMANN_BASIC_JSON_TPL& j) const
+    {
+        return nlohmann::detail::hash(j);
+    }
+};
+
+// specialization for std::less<value_t>
+template<>
+struct less< ::nlohmann::detail::value_t> // do not remove the space after '<', see https://github.com/nlohmann/json/pull/679
+{
+    /*!
+    @brief compare two value_t enum values
+    @since version 3.0.0
+    */
+    bool operator()(::nlohmann::detail::value_t lhs,
+                    ::nlohmann::detail::value_t rhs) const noexcept
+    {
+#if JSON_HAS_THREE_WAY_COMPARISON
+        return std::is_lt(lhs <=> rhs); // *NOPAD*
+#else
+        return ::nlohmann::detail::operator<(lhs, rhs);
+#endif
+    }
+};
+
+// C++20 prohibit function specialization in the std namespace.
+#ifndef JSON_HAS_CPP_20
+
+/// @brief exchanges the values of two JSON objects
+/// @sa https://json.nlohmann.me/api/basic_json/std_swap/
+NLOHMANN_BASIC_JSON_TPL_DECLARATION
+inline void swap(nlohmann::NLOHMANN_BASIC_JSON_TPL& j1, nlohmann::NLOHMANN_BASIC_JSON_TPL& j2) noexcept(  // NOLINT(readability-inconsistent-declaration-parameter-name)
+    is_nothrow_move_constructible<nlohmann::NLOHMANN_BASIC_JSON_TPL>::value&&                          // NOLINT(misc-redundant-expression)
+    is_nothrow_move_assignable<nlohmann::NLOHMANN_BASIC_JSON_TPL>::value)
+{
+    j1.swap(j2);
+}
+
+#endif
+
+}  // namespace std
+
+#if JSON_USE_GLOBAL_UDLS
+    using nlohmann::literals::json_literals::operator "" _json; // NOLINT(misc-unused-using-decls,google-global-names-in-headers)
+    using nlohmann::literals::json_literals::operator "" _json_pointer; //NOLINT(misc-unused-using-decls,google-global-names-in-headers)
+#endif
+
+// #include <nlohmann/detail/macro_unscope.hpp>
+//     __ _____ _____ _____
+//  __|  |   __|     |   | |  JSON for Modern C++
+// |  |  |__   |  |  | | | |  version 3.11.2
+// |_____|_____|_____|_|___|  https://github.com/nlohmann/json
+//
+// SPDX-FileCopyrightText: 2013-2022 Niels Lohmann <https://nlohmann.me>
+// SPDX-License-Identifier: MIT
+
+
+
+// restore clang diagnostic settings
+#if defined(__clang__)
+    #pragma clang diagnostic pop
+#endif
+
+// clean up
+#undef JSON_ASSERT
+#undef JSON_INTERNAL_CATCH
+#undef JSON_THROW
+#undef JSON_PRIVATE_UNLESS_TESTED
+#undef NLOHMANN_BASIC_JSON_TPL_DECLARATION
+#undef NLOHMANN_BASIC_JSON_TPL
+#undef JSON_EXPLICIT
+#undef NLOHMANN_CAN_CALL_STD_FUNC_IMPL
+#undef JSON_INLINE_VARIABLE
+#undef JSON_NO_UNIQUE_ADDRESS
+#undef JSON_DISABLE_ENUM_SERIALIZATION
+#undef JSON_USE_GLOBAL_UDLS
+
+#ifndef JSON_TEST_KEEP_MACROS
+    #undef JSON_CATCH
+    #undef JSON_TRY
+    #undef JSON_HAS_CPP_11
+    #undef JSON_HAS_CPP_14
+    #undef JSON_HAS_CPP_17
+    #undef JSON_HAS_CPP_20
+    #undef JSON_HAS_FILESYSTEM
+    #undef JSON_HAS_EXPERIMENTAL_FILESYSTEM
+    #undef JSON_HAS_THREE_WAY_COMPARISON
+    #undef JSON_HAS_RANGES
+    #undef JSON_USE_LEGACY_DISCARDED_VALUE_COMPARISON
+#endif
+
+// #include <nlohmann/thirdparty/hedley/hedley_undef.hpp>
+//     __ _____ _____ _____
+//  __|  |   __|     |   | |  JSON for Modern C++
+// |  |  |__   |  |  | | | |  version 3.11.2
+// |_____|_____|_____|_|___|  https://github.com/nlohmann/json
+//
+// SPDX-FileCopyrightText: 2013-2022 Niels Lohmann <https://nlohmann.me>
+// SPDX-License-Identifier: MIT
+
+
+
+#undef JSON_HEDLEY_ALWAYS_INLINE
+#undef JSON_HEDLEY_ARM_VERSION
+#undef JSON_HEDLEY_ARM_VERSION_CHECK
+#undef JSON_HEDLEY_ARRAY_PARAM
+#undef JSON_HEDLEY_ASSUME
+#undef JSON_HEDLEY_BEGIN_C_DECLS
+#undef JSON_HEDLEY_CLANG_HAS_ATTRIBUTE
+#undef JSON_HEDLEY_CLANG_HAS_BUILTIN
+#undef JSON_HEDLEY_CLANG_HAS_CPP_ATTRIBUTE
+#undef JSON_HEDLEY_CLANG_HAS_DECLSPEC_DECLSPEC_ATTRIBUTE
+#undef JSON_HEDLEY_CLANG_HAS_EXTENSION
+#undef JSON_HEDLEY_CLANG_HAS_FEATURE
+#undef JSON_HEDLEY_CLANG_HAS_WARNING
+#undef JSON_HEDLEY_COMPCERT_VERSION
+#undef JSON_HEDLEY_COMPCERT_VERSION_CHECK
+#undef JSON_HEDLEY_CONCAT
+#undef JSON_HEDLEY_CONCAT3
+#undef JSON_HEDLEY_CONCAT3_EX
+#undef JSON_HEDLEY_CONCAT_EX
+#undef JSON_HEDLEY_CONST
+#undef JSON_HEDLEY_CONSTEXPR
+#undef JSON_HEDLEY_CONST_CAST
+#undef JSON_HEDLEY_CPP_CAST
+#undef JSON_HEDLEY_CRAY_VERSION
+#undef JSON_HEDLEY_CRAY_VERSION_CHECK
+#undef JSON_HEDLEY_C_DECL
+#undef JSON_HEDLEY_DEPRECATED
+#undef JSON_HEDLEY_DEPRECATED_FOR
+#undef JSON_HEDLEY_DIAGNOSTIC_DISABLE_CAST_QUAL
+#undef JSON_HEDLEY_DIAGNOSTIC_DISABLE_CPP98_COMPAT_WRAP_
+#undef JSON_HEDLEY_DIAGNOSTIC_DISABLE_DEPRECATED
+#undef JSON_HEDLEY_DIAGNOSTIC_DISABLE_UNKNOWN_CPP_ATTRIBUTES
+#undef JSON_HEDLEY_DIAGNOSTIC_DISABLE_UNKNOWN_PRAGMAS
+#undef JSON_HEDLEY_DIAGNOSTIC_DISABLE_UNUSED_FUNCTION
+#undef JSON_HEDLEY_DIAGNOSTIC_POP
+#undef JSON_HEDLEY_DIAGNOSTIC_PUSH
+#undef JSON_HEDLEY_DMC_VERSION
+#undef JSON_HEDLEY_DMC_VERSION_CHECK
+#undef JSON_HEDLEY_EMPTY_BASES
+#undef JSON_HEDLEY_EMSCRIPTEN_VERSION
+#undef JSON_HEDLEY_EMSCRIPTEN_VERSION_CHECK
+#undef JSON_HEDLEY_END_C_DECLS
+#undef JSON_HEDLEY_FLAGS
+#undef JSON_HEDLEY_FLAGS_CAST
+#undef JSON_HEDLEY_GCC_HAS_ATTRIBUTE
+#undef JSON_HEDLEY_GCC_HAS_BUILTIN
+#undef JSON_HEDLEY_GCC_HAS_CPP_ATTRIBUTE
+#undef JSON_HEDLEY_GCC_HAS_DECLSPEC_ATTRIBUTE
+#undef JSON_HEDLEY_GCC_HAS_EXTENSION
+#undef JSON_HEDLEY_GCC_HAS_FEATURE
+#undef JSON_HEDLEY_GCC_HAS_WARNING
+#undef JSON_HEDLEY_GCC_NOT_CLANG_VERSION_CHECK
+#undef JSON_HEDLEY_GCC_VERSION
+#undef JSON_HEDLEY_GCC_VERSION_CHECK
+#undef JSON_HEDLEY_GNUC_HAS_ATTRIBUTE
+#undef JSON_HEDLEY_GNUC_HAS_BUILTIN
+#undef JSON_HEDLEY_GNUC_HAS_CPP_ATTRIBUTE
+#undef JSON_HEDLEY_GNUC_HAS_DECLSPEC_ATTRIBUTE
+#undef JSON_HEDLEY_GNUC_HAS_EXTENSION
+#undef JSON_HEDLEY_GNUC_HAS_FEATURE
+#undef JSON_HEDLEY_GNUC_HAS_WARNING
+#undef JSON_HEDLEY_GNUC_VERSION
+#undef JSON_HEDLEY_GNUC_VERSION_CHECK
+#undef JSON_HEDLEY_HAS_ATTRIBUTE
+#undef JSON_HEDLEY_HAS_BUILTIN
+#undef JSON_HEDLEY_HAS_CPP_ATTRIBUTE
+#undef JSON_HEDLEY_HAS_CPP_ATTRIBUTE_NS
+#undef JSON_HEDLEY_HAS_DECLSPEC_ATTRIBUTE
+#undef JSON_HEDLEY_HAS_EXTENSION
+#undef JSON_HEDLEY_HAS_FEATURE
+#undef JSON_HEDLEY_HAS_WARNING
+#undef JSON_HEDLEY_IAR_VERSION
+#undef JSON_HEDLEY_IAR_VERSION_CHECK
+#undef JSON_HEDLEY_IBM_VERSION
+#undef JSON_HEDLEY_IBM_VERSION_CHECK
+#undef JSON_HEDLEY_IMPORT
+#undef JSON_HEDLEY_INLINE
+#undef JSON_HEDLEY_INTEL_CL_VERSION
+#undef JSON_HEDLEY_INTEL_CL_VERSION_CHECK
+#undef JSON_HEDLEY_INTEL_VERSION
+#undef JSON_HEDLEY_INTEL_VERSION_CHECK
+#undef JSON_HEDLEY_IS_CONSTANT
+#undef JSON_HEDLEY_IS_CONSTEXPR_
+#undef JSON_HEDLEY_LIKELY
+#undef JSON_HEDLEY_MALLOC
+#undef JSON_HEDLEY_MCST_LCC_VERSION
+#undef JSON_HEDLEY_MCST_LCC_VERSION_CHECK
+#undef JSON_HEDLEY_MESSAGE
+#undef JSON_HEDLEY_MSVC_VERSION
+#undef JSON_HEDLEY_MSVC_VERSION_CHECK
+#undef JSON_HEDLEY_NEVER_INLINE
+#undef JSON_HEDLEY_NON_NULL
+#undef JSON_HEDLEY_NO_ESCAPE
+#undef JSON_HEDLEY_NO_RETURN
+#undef JSON_HEDLEY_NO_THROW
+#undef JSON_HEDLEY_NULL
+#undef JSON_HEDLEY_PELLES_VERSION
+#undef JSON_HEDLEY_PELLES_VERSION_CHECK
+#undef JSON_HEDLEY_PGI_VERSION
+#undef JSON_HEDLEY_PGI_VERSION_CHECK
+#undef JSON_HEDLEY_PREDICT
+#undef JSON_HEDLEY_PRINTF_FORMAT
+#undef JSON_HEDLEY_PRIVATE
+#undef JSON_HEDLEY_PUBLIC
+#undef JSON_HEDLEY_PURE
+#undef JSON_HEDLEY_REINTERPRET_CAST
+#undef JSON_HEDLEY_REQUIRE
+#undef JSON_HEDLEY_REQUIRE_CONSTEXPR
+#undef JSON_HEDLEY_REQUIRE_MSG
+#undef JSON_HEDLEY_RESTRICT
+#undef JSON_HEDLEY_RETURNS_NON_NULL
+#undef JSON_HEDLEY_SENTINEL
+#undef JSON_HEDLEY_STATIC_ASSERT
+#undef JSON_HEDLEY_STATIC_CAST
+#undef JSON_HEDLEY_STRINGIFY
+#undef JSON_HEDLEY_STRINGIFY_EX
+#undef JSON_HEDLEY_SUNPRO_VERSION
+#undef JSON_HEDLEY_SUNPRO_VERSION_CHECK
+#undef JSON_HEDLEY_TINYC_VERSION
+#undef JSON_HEDLEY_TINYC_VERSION_CHECK
+#undef JSON_HEDLEY_TI_ARMCL_VERSION
+#undef JSON_HEDLEY_TI_ARMCL_VERSION_CHECK
+#undef JSON_HEDLEY_TI_CL2000_VERSION
+#undef JSON_HEDLEY_TI_CL2000_VERSION_CHECK
+#undef JSON_HEDLEY_TI_CL430_VERSION
+#undef JSON_HEDLEY_TI_CL430_VERSION_CHECK
+#undef JSON_HEDLEY_TI_CL6X_VERSION
+#undef JSON_HEDLEY_TI_CL6X_VERSION_CHECK
+#undef JSON_HEDLEY_TI_CL7X_VERSION
+#undef JSON_HEDLEY_TI_CL7X_VERSION_CHECK
+#undef JSON_HEDLEY_TI_CLPRU_VERSION
+#undef JSON_HEDLEY_TI_CLPRU_VERSION_CHECK
+#undef JSON_HEDLEY_TI_VERSION
+#undef JSON_HEDLEY_TI_VERSION_CHECK
+#undef JSON_HEDLEY_UNAVAILABLE
+#undef JSON_HEDLEY_UNLIKELY
+#undef JSON_HEDLEY_UNPREDICTABLE
+#undef JSON_HEDLEY_UNREACHABLE
+#undef JSON_HEDLEY_UNREACHABLE_RETURN
+#undef JSON_HEDLEY_VERSION
+#undef JSON_HEDLEY_VERSION_DECODE_MAJOR
+#undef JSON_HEDLEY_VERSION_DECODE_MINOR
+#undef JSON_HEDLEY_VERSION_DECODE_REVISION
+#undef JSON_HEDLEY_VERSION_ENCODE
+#undef JSON_HEDLEY_WARNING
+#undef JSON_HEDLEY_WARN_UNUSED_RESULT
+#undef JSON_HEDLEY_WARN_UNUSED_RESULT_MSG
+#undef JSON_HEDLEY_FALL_THROUGH
+
+
+
+#endif  // INCLUDE_NLOHMANN_JSON_HPP_
diff --git a/src/whisper_cpp/examples/main/CMakeLists.txt b/src/whisper_cpp/examples/main/CMakeLists.txt
new file mode 100644
index 00000000..1e66e4b5
--- /dev/null
+++ b/src/whisper_cpp/examples/main/CMakeLists.txt
@@ -0,0 +1,6 @@
+set(TARGET main)
+add_executable(${TARGET} main.cpp)
+
+include(DefaultTargetOptions)
+
+target_link_libraries(${TARGET} PRIVATE common whisper ${FFMPEG_LIBRARIES} ${CMAKE_THREAD_LIBS_INIT})
diff --git a/src/whisper_cpp/examples/main/README.md b/src/whisper_cpp/examples/main/README.md
new file mode 100644
index 00000000..aca2a9fd
--- /dev/null
+++ b/src/whisper_cpp/examples/main/README.md
@@ -0,0 +1,53 @@
+# main
+
+This is the main example demonstrating most of the functionality of the Whisper model.
+It can be used as a reference for using the `whisper.cpp` library in other projects.
+
+```
+./main -h
+
+usage: ./main [options] file0.wav file1.wav ...
+
+options:
+  -h,        --help              [default] show this help message and exit
+  -t N,      --threads N         [4      ] number of threads to use during computation
+  -p N,      --processors N      [1      ] number of processors to use during computation
+  -ot N,     --offset-t N        [0      ] time offset in milliseconds
+  -on N,     --offset-n N        [0      ] segment index offset
+  -d  N,     --duration N        [0      ] duration of audio to process in milliseconds
+  -mc N,     --max-context N     [-1     ] maximum number of text context tokens to store
+  -ml N,     --max-len N         [0      ] maximum segment length in characters
+  -sow,      --split-on-word     [false  ] split on word rather than on token
+  -bo N,     --best-of N         [5      ] number of best candidates to keep
+  -bs N,     --beam-size N       [5      ] beam size for beam search
+  -wt N,     --word-thold N      [0.01   ] word timestamp probability threshold
+  -et N,     --entropy-thold N   [2.40   ] entropy threshold for decoder fail
+  -lpt N,    --logprob-thold N   [-1.00  ] log probability threshold for decoder fail
+  -debug,    --debug-mode        [false  ] enable debug mode (eg. dump log_mel)
+  -tr,       --translate         [false  ] translate from source language to english
+  -di,       --diarize           [false  ] stereo audio diarization
+  -tdrz,     --tinydiarize       [false  ] enable tinydiarize (requires a tdrz model)
+  -nf,       --no-fallback       [false  ] do not use temperature fallback while decoding
+  -otxt,     --output-txt        [false  ] output result in a text file
+  -ovtt,     --output-vtt        [false  ] output result in a vtt file
+  -osrt,     --output-srt        [false  ] output result in a srt file
+  -olrc,     --output-lrc        [false  ] output result in a lrc file
+  -owts,     --output-words      [false  ] output script for generating karaoke video
+  -fp,       --font-path         [/System/Library/Fonts/Supplemental/Courier New Bold.ttf] path to a monospace font for karaoke video
+  -ocsv,     --output-csv        [false  ] output result in a CSV file
+  -oj,       --output-json       [false  ] output result in a JSON file
+  -ojf,      --output-json-full  [false  ] include more information in the JSON file
+  -of FNAME, --output-file FNAME [       ] output file path (without file extension)
+  -ps,       --print-special     [false  ] print special tokens
+  -pc,       --print-colors      [false  ] print colors
+  -pp,       --print-progress    [false  ] print progress
+  -nt,       --no-timestamps     [false  ] do not print timestamps
+  -l LANG,   --language LANG     [en     ] spoken language ('auto' for auto-detect)
+  -dl,       --detect-language   [false  ] exit after automatically detecting language
+             --prompt PROMPT     [       ] initial prompt
+  -m FNAME,  --model FNAME       [models/ggml-base.en.bin] model path
+  -f FNAME,  --file FNAME        [       ] input WAV file path
+  -oved D,   --ov-e-device DNAME [CPU    ] the OpenVINO device used for encode inference
+  -ls,       --log-score         [false  ] log best decoder scores of tokens
+  -ng,       --no-gpu            [false  ] disable GPU
+```
diff --git a/src/whisper_cpp/examples/main/main.cpp b/src/whisper_cpp/examples/main/main.cpp
new file mode 100644
index 00000000..8847e9d3
--- /dev/null
+++ b/src/whisper_cpp/examples/main/main.cpp
@@ -0,0 +1,1244 @@
+#include "common.h"
+
+#include "whisper.h"
+#include "grammar-parser.h"
+
+#include <cmath>
+#include <fstream>
+#include <cstdio>
+#include <regex>
+#include <string>
+#include <thread>
+#include <vector>
+#include <cstring>
+
+#if defined(_MSC_VER)
+#pragma warning(disable: 4244 4267) // possible loss of data
+#endif
+
+// helper function to replace substrings
+static void replace_all(std::string & s, const std::string & search, const std::string & replace) {
+    for (size_t pos = 0; ; pos += replace.length()) {
+        pos = s.find(search, pos);
+        if (pos == std::string::npos) break;
+        s.erase(pos, search.length());
+        s.insert(pos, replace);
+    }
+}
+
+// command-line parameters
+struct whisper_params {
+    int32_t n_threads     = std::min(4, (int32_t) std::thread::hardware_concurrency());
+    int32_t n_processors  = 1;
+    int32_t offset_t_ms   = 0;
+    int32_t offset_n      = 0;
+    int32_t duration_ms   = 0;
+    int32_t progress_step = 5;
+    int32_t max_context   = -1;
+    int32_t max_len       = 0;
+    int32_t best_of       = whisper_full_default_params(WHISPER_SAMPLING_GREEDY).greedy.best_of;
+    int32_t beam_size     = whisper_full_default_params(WHISPER_SAMPLING_BEAM_SEARCH).beam_search.beam_size;
+    int32_t audio_ctx     = 0;
+
+    float word_thold      =  0.01f;
+    float entropy_thold   =  2.40f;
+    float logprob_thold   = -1.00f;
+    float grammar_penalty = 100.0f;
+    float temperature     = 0.0f;
+    float temperature_inc = 0.2f;
+
+    bool debug_mode      = false;
+    bool translate       = false;
+    bool detect_language = false;
+    bool diarize         = false;
+    bool tinydiarize     = false;
+    bool split_on_word   = false;
+    bool no_fallback     = false;
+    bool output_txt      = false;
+    bool output_vtt      = false;
+    bool output_srt      = false;
+    bool output_wts      = false;
+    bool output_csv      = false;
+    bool output_jsn      = false;
+    bool output_jsn_full = false;
+    bool output_lrc      = false;
+    bool no_prints       = false;
+    bool print_special   = false;
+    bool print_colors    = false;
+    bool print_progress  = false;
+    bool no_timestamps   = false;
+    bool log_score       = false;
+    bool use_gpu         = true;
+    bool flash_attn      = false;
+
+    std::string language  = "en";
+    std::string prompt;
+    std::string font_path = "/System/Library/Fonts/Supplemental/Courier New Bold.ttf";
+    std::string model     = "models/ggml-base.en.bin";
+    std::string grammar;
+    std::string grammar_rule;
+
+    // [TDRZ] speaker turn string
+    std::string tdrz_speaker_turn = " [SPEAKER_TURN]"; // TODO: set from command line
+
+    // A regular expression that matches tokens to suppress
+    std::string suppress_regex;
+
+    std::string openvino_encode_device = "CPU";
+
+    std::string dtw = "";
+
+    std::vector<std::string> fname_inp = {};
+    std::vector<std::string> fname_out = {};
+
+    grammar_parser::parse_state grammar_parsed;
+};
+
+static void whisper_print_usage(int argc, char ** argv, const whisper_params & params);
+
+static char * whisper_param_turn_lowercase(char * in){
+    int string_len = strlen(in);
+    for (int i = 0; i < string_len; i++){
+        *(in+i) = tolower((unsigned char)*(in+i));
+    }
+    return in;
+}
+
+static bool whisper_params_parse(int argc, char ** argv, whisper_params & params) {
+    for (int i = 1; i < argc; i++) {
+        std::string arg = argv[i];
+
+        if (arg == "-"){
+            params.fname_inp.push_back(arg);
+            continue;
+        }
+
+        if (arg[0] != '-') {
+            params.fname_inp.push_back(arg);
+            continue;
+        }
+
+        if (arg == "-h" || arg == "--help") {
+            whisper_print_usage(argc, argv, params);
+            exit(0);
+        }
+        else if (arg == "-t"    || arg == "--threads")         { params.n_threads       = std::stoi(argv[++i]); }
+        else if (arg == "-p"    || arg == "--processors")      { params.n_processors    = std::stoi(argv[++i]); }
+        else if (arg == "-ot"   || arg == "--offset-t")        { params.offset_t_ms     = std::stoi(argv[++i]); }
+        else if (arg == "-on"   || arg == "--offset-n")        { params.offset_n        = std::stoi(argv[++i]); }
+        else if (arg == "-d"    || arg == "--duration")        { params.duration_ms     = std::stoi(argv[++i]); }
+        else if (arg == "-mc"   || arg == "--max-context")     { params.max_context     = std::stoi(argv[++i]); }
+        else if (arg == "-ml"   || arg == "--max-len")         { params.max_len         = std::stoi(argv[++i]); }
+        else if (arg == "-bo"   || arg == "--best-of")         { params.best_of         = std::stoi(argv[++i]); }
+        else if (arg == "-bs"   || arg == "--beam-size")       { params.beam_size       = std::stoi(argv[++i]); }
+        else if (arg == "-ac"   || arg == "--audio-ctx")       { params.audio_ctx       = std::stoi(argv[++i]); }
+        else if (arg == "-wt"   || arg == "--word-thold")      { params.word_thold      = std::stof(argv[++i]); }
+        else if (arg == "-et"   || arg == "--entropy-thold")   { params.entropy_thold   = std::stof(argv[++i]); }
+        else if (arg == "-lpt"  || arg == "--logprob-thold")   { params.logprob_thold   = std::stof(argv[++i]); }
+        else if (arg == "-tp"   || arg == "--temperature")     { params.temperature     = std::stof(argv[++i]); }
+        else if (arg == "-tpi"  || arg == "--temperature-inc") { params.temperature_inc = std::stof(argv[++i]); }
+        else if (arg == "-debug"|| arg == "--debug-mode")      { params.debug_mode      = true; }
+        else if (arg == "-tr"   || arg == "--translate")       { params.translate       = true; }
+        else if (arg == "-di"   || arg == "--diarize")         { params.diarize         = true; }
+        else if (arg == "-tdrz" || arg == "--tinydiarize")     { params.tinydiarize     = true; }
+        else if (arg == "-sow"  || arg == "--split-on-word")   { params.split_on_word   = true; }
+        else if (arg == "-nf"   || arg == "--no-fallback")     { params.no_fallback     = true; }
+        else if (arg == "-otxt" || arg == "--output-txt")      { params.output_txt      = true; }
+        else if (arg == "-ovtt" || arg == "--output-vtt")      { params.output_vtt      = true; }
+        else if (arg == "-osrt" || arg == "--output-srt")      { params.output_srt      = true; }
+        else if (arg == "-owts" || arg == "--output-words")    { params.output_wts      = true; }
+        else if (arg == "-olrc" || arg == "--output-lrc")      { params.output_lrc      = true; }
+        else if (arg == "-fp"   || arg == "--font-path")       { params.font_path       = argv[++i]; }
+        else if (arg == "-ocsv" || arg == "--output-csv")      { params.output_csv      = true; }
+        else if (arg == "-oj"   || arg == "--output-json")     { params.output_jsn      = true; }
+        else if (arg == "-ojf"  || arg == "--output-json-full"){ params.output_jsn_full = params.output_jsn = true; }
+        else if (arg == "-of"   || arg == "--output-file")     { params.fname_out.emplace_back(argv[++i]); }
+        else if (arg == "-np"   || arg == "--no-prints")       { params.no_prints       = true; }
+        else if (arg == "-ps"   || arg == "--print-special")   { params.print_special   = true; }
+        else if (arg == "-pc"   || arg == "--print-colors")    { params.print_colors    = true; }
+        else if (arg == "-pp"   || arg == "--print-progress")  { params.print_progress  = true; }
+        else if (arg == "-nt"   || arg == "--no-timestamps")   { params.no_timestamps   = true; }
+        else if (arg == "-l"    || arg == "--language")        { params.language        = whisper_param_turn_lowercase(argv[++i]); }
+        else if (arg == "-dl"   || arg == "--detect-language") { params.detect_language = true; }
+        else if (                  arg == "--prompt")          { params.prompt          = argv[++i]; }
+        else if (arg == "-m"    || arg == "--model")           { params.model           = argv[++i]; }
+        else if (arg == "-f"    || arg == "--file")            { params.fname_inp.emplace_back(argv[++i]); }
+        else if (arg == "-oved" || arg == "--ov-e-device")     { params.openvino_encode_device = argv[++i]; }
+        else if (arg == "-dtw"  || arg == "--dtw")             { params.dtw             = argv[++i]; }
+        else if (arg == "-ls"   || arg == "--log-score")       { params.log_score       = true; }
+        else if (arg == "-ng"   || arg == "--no-gpu")          { params.use_gpu         = false; }
+        else if (arg == "-fa"   || arg == "--flash-attn")      { params.flash_attn      = true; }
+        else if (                  arg == "--suppress-regex")  { params.suppress_regex  = argv[++i]; }
+        else if (                  arg == "--grammar")         { params.grammar         = argv[++i]; }
+        else if (                  arg == "--grammar-rule")    { params.grammar_rule    = argv[++i]; }
+        else if (                  arg == "--grammar-penalty") { params.grammar_penalty = std::stof(argv[++i]); }
+        else {
+            fprintf(stderr, "error: unknown argument: %s\n", arg.c_str());
+            whisper_print_usage(argc, argv, params);
+            exit(0);
+        }
+    }
+
+    return true;
+}
+
+static void whisper_print_usage(int /*argc*/, char ** argv, const whisper_params & params) {
+    fprintf(stderr, "\n");
+    fprintf(stderr, "usage: %s [options] file0.wav file1.wav ...\n", argv[0]);
+    fprintf(stderr, "\n");
+    fprintf(stderr, "options:\n");
+    fprintf(stderr, "  -h,        --help              [default] show this help message and exit\n");
+    fprintf(stderr, "  -t N,      --threads N         [%-7d] number of threads to use during computation\n",    params.n_threads);
+    fprintf(stderr, "  -p N,      --processors N      [%-7d] number of processors to use during computation\n", params.n_processors);
+    fprintf(stderr, "  -ot N,     --offset-t N        [%-7d] time offset in milliseconds\n",                    params.offset_t_ms);
+    fprintf(stderr, "  -on N,     --offset-n N        [%-7d] segment index offset\n",                           params.offset_n);
+    fprintf(stderr, "  -d  N,     --duration N        [%-7d] duration of audio to process in milliseconds\n",   params.duration_ms);
+    fprintf(stderr, "  -mc N,     --max-context N     [%-7d] maximum number of text context tokens to store\n", params.max_context);
+    fprintf(stderr, "  -ml N,     --max-len N         [%-7d] maximum segment length in characters\n",           params.max_len);
+    fprintf(stderr, "  -sow,      --split-on-word     [%-7s] split on word rather than on token\n",             params.split_on_word ? "true" : "false");
+    fprintf(stderr, "  -bo N,     --best-of N         [%-7d] number of best candidates to keep\n",              params.best_of);
+    fprintf(stderr, "  -bs N,     --beam-size N       [%-7d] beam size for beam search\n",                      params.beam_size);
+    fprintf(stderr, "  -ac N,     --audio-ctx N       [%-7d] audio context size (0 - all)\n",                   params.audio_ctx);
+    fprintf(stderr, "  -wt N,     --word-thold N      [%-7.2f] word timestamp probability threshold\n",         params.word_thold);
+    fprintf(stderr, "  -et N,     --entropy-thold N   [%-7.2f] entropy threshold for decoder fail\n",           params.entropy_thold);
+    fprintf(stderr, "  -lpt N,    --logprob-thold N   [%-7.2f] log probability threshold for decoder fail\n",   params.logprob_thold);
+    fprintf(stderr, "  -tp,       --temperature N     [%-7.2f] The sampling temperature, between 0 and 1\n",    params.temperature);
+    fprintf(stderr, "  -tpi,      --temperature-inc N [%-7.2f] The increment of temperature, between 0 and 1\n",params.temperature_inc);
+    fprintf(stderr, "  -debug,    --debug-mode        [%-7s] enable debug mode (eg. dump log_mel)\n",           params.debug_mode ? "true" : "false");
+    fprintf(stderr, "  -tr,       --translate         [%-7s] translate from source language to english\n",      params.translate ? "true" : "false");
+    fprintf(stderr, "  -di,       --diarize           [%-7s] stereo audio diarization\n",                       params.diarize ? "true" : "false");
+    fprintf(stderr, "  -tdrz,     --tinydiarize       [%-7s] enable tinydiarize (requires a tdrz model)\n",     params.tinydiarize ? "true" : "false");
+    fprintf(stderr, "  -nf,       --no-fallback       [%-7s] do not use temperature fallback while decoding\n", params.no_fallback ? "true" : "false");
+    fprintf(stderr, "  -otxt,     --output-txt        [%-7s] output result in a text file\n",                   params.output_txt ? "true" : "false");
+    fprintf(stderr, "  -ovtt,     --output-vtt        [%-7s] output result in a vtt file\n",                    params.output_vtt ? "true" : "false");
+    fprintf(stderr, "  -osrt,     --output-srt        [%-7s] output result in a srt file\n",                    params.output_srt ? "true" : "false");
+    fprintf(stderr, "  -olrc,     --output-lrc        [%-7s] output result in a lrc file\n",                    params.output_lrc ? "true" : "false");
+    fprintf(stderr, "  -owts,     --output-words      [%-7s] output script for generating karaoke video\n",     params.output_wts ? "true" : "false");
+    fprintf(stderr, "  -fp,       --font-path         [%-7s] path to a monospace font for karaoke video\n",     params.font_path.c_str());
+    fprintf(stderr, "  -ocsv,     --output-csv        [%-7s] output result in a CSV file\n",                    params.output_csv ? "true" : "false");
+    fprintf(stderr, "  -oj,       --output-json       [%-7s] output result in a JSON file\n",                   params.output_jsn ? "true" : "false");
+    fprintf(stderr, "  -ojf,      --output-json-full  [%-7s] include more information in the JSON file\n",      params.output_jsn_full ? "true" : "false");
+    fprintf(stderr, "  -of FNAME, --output-file FNAME [%-7s] output file path (without file extension)\n",      "");
+    fprintf(stderr, "  -np,       --no-prints         [%-7s] do not print anything other than the results\n",   params.no_prints ? "true" : "false");
+    fprintf(stderr, "  -ps,       --print-special     [%-7s] print special tokens\n",                           params.print_special ? "true" : "false");
+    fprintf(stderr, "  -pc,       --print-colors      [%-7s] print colors\n",                                   params.print_colors ? "true" : "false");
+    fprintf(stderr, "  -pp,       --print-progress    [%-7s] print progress\n",                                 params.print_progress ? "true" : "false");
+    fprintf(stderr, "  -nt,       --no-timestamps     [%-7s] do not print timestamps\n",                        params.no_timestamps ? "true" : "false");
+    fprintf(stderr, "  -l LANG,   --language LANG     [%-7s] spoken language ('auto' for auto-detect)\n",       params.language.c_str());
+    fprintf(stderr, "  -dl,       --detect-language   [%-7s] exit after automatically detecting language\n",    params.detect_language ? "true" : "false");
+    fprintf(stderr, "             --prompt PROMPT     [%-7s] initial prompt (max n_text_ctx/2 tokens)\n",       params.prompt.c_str());
+    fprintf(stderr, "  -m FNAME,  --model FNAME       [%-7s] model path\n",                                     params.model.c_str());
+    fprintf(stderr, "  -f FNAME,  --file FNAME        [%-7s] input WAV file path\n",                            "");
+    fprintf(stderr, "  -oved D,   --ov-e-device DNAME [%-7s] the OpenVINO device used for encode inference\n",  params.openvino_encode_device.c_str());
+    fprintf(stderr, "  -dtw MODEL --dtw MODEL         [%-7s] compute token-level timestamps\n",                 params.dtw.c_str());
+    fprintf(stderr, "  -ls,       --log-score         [%-7s] log best decoder scores of tokens\n",              params.log_score?"true":"false");
+    fprintf(stderr, "  -ng,       --no-gpu            [%-7s] disable GPU\n",                                    params.use_gpu ? "false" : "true");
+    fprintf(stderr, "  -fa,       --flash-attn        [%-7s] flash attention\n",                                params.flash_attn ? "true" : "false");
+    fprintf(stderr, "  --suppress-regex REGEX         [%-7s] regular expression matching tokens to suppress\n", params.suppress_regex.c_str());
+    fprintf(stderr, "  --grammar GRAMMAR              [%-7s] GBNF grammar to guide decoding\n",                 params.grammar.c_str());
+    fprintf(stderr, "  --grammar-rule RULE            [%-7s] top-level GBNF grammar rule name\n",               params.grammar_rule.c_str());
+    fprintf(stderr, "  --grammar-penalty N            [%-7.1f] scales down logits of nongrammar tokens\n",      params.grammar_penalty);
+    fprintf(stderr, "\n");
+}
+
+struct whisper_print_user_data {
+    const whisper_params * params;
+
+    const std::vector<std::vector<float>> * pcmf32s;
+    int progress_prev;
+};
+
+static std::string estimate_diarization_speaker(std::vector<std::vector<float>> pcmf32s, int64_t t0, int64_t t1, bool id_only = false) {
+    std::string speaker = "";
+    const int64_t n_samples = pcmf32s[0].size();
+
+    const int64_t is0 = timestamp_to_sample(t0, n_samples, WHISPER_SAMPLE_RATE);
+    const int64_t is1 = timestamp_to_sample(t1, n_samples, WHISPER_SAMPLE_RATE);
+
+    double energy0 = 0.0f;
+    double energy1 = 0.0f;
+
+    for (int64_t j = is0; j < is1; j++) {
+        energy0 += fabs(pcmf32s[0][j]);
+        energy1 += fabs(pcmf32s[1][j]);
+    }
+
+    if (energy0 > 1.1*energy1) {
+        speaker = "0";
+    } else if (energy1 > 1.1*energy0) {
+        speaker = "1";
+    } else {
+        speaker = "?";
+    }
+
+    //printf("is0 = %lld, is1 = %lld, energy0 = %f, energy1 = %f, speaker = %s\n", is0, is1, energy0, energy1, speaker.c_str());
+
+    if (!id_only) {
+        speaker.insert(0, "(speaker ");
+        speaker.append(")");
+    }
+
+    return speaker;
+}
+
+static void whisper_print_progress_callback(struct whisper_context * /*ctx*/, struct whisper_state * /*state*/, int progress, void * user_data) {
+    int progress_step = ((whisper_print_user_data *) user_data)->params->progress_step;
+    int * progress_prev  = &(((whisper_print_user_data *) user_data)->progress_prev);
+    if (progress >= *progress_prev + progress_step) {
+        *progress_prev += progress_step;
+        fprintf(stderr, "%s: progress = %3d%%\n", __func__, progress);
+    }
+}
+
+static void whisper_print_segment_callback(struct whisper_context * ctx, struct whisper_state * /*state*/, int n_new, void * user_data) {
+    const auto & params  = *((whisper_print_user_data *) user_data)->params;
+    const auto & pcmf32s = *((whisper_print_user_data *) user_data)->pcmf32s;
+
+    const int n_segments = whisper_full_n_segments(ctx);
+
+    std::string speaker = "";
+
+    int64_t t0 = 0;
+    int64_t t1 = 0;
+
+    // print the last n_new segments
+    const int s0 = n_segments - n_new;
+
+    if (s0 == 0) {
+        printf("\n");
+    }
+
+    for (int i = s0; i < n_segments; i++) {
+        if (!params.no_timestamps || params.diarize) {
+            t0 = whisper_full_get_segment_t0(ctx, i);
+            t1 = whisper_full_get_segment_t1(ctx, i);
+        }
+
+        if (!params.no_timestamps) {
+            printf("[%s --> %s]  ", to_timestamp(t0).c_str(), to_timestamp(t1).c_str());
+        }
+
+        if (params.diarize && pcmf32s.size() == 2) {
+            speaker = estimate_diarization_speaker(pcmf32s, t0, t1);
+        }
+
+        if (params.print_colors) {
+            for (int j = 0; j < whisper_full_n_tokens(ctx, i); ++j) {
+                if (params.print_special == false) {
+                    const whisper_token id = whisper_full_get_token_id(ctx, i, j);
+                    if (id >= whisper_token_eot(ctx)) {
+                        continue;
+                    }
+                }
+
+                const char * text = whisper_full_get_token_text(ctx, i, j);
+                const float  p    = whisper_full_get_token_p   (ctx, i, j);
+
+                const int col = std::max(0, std::min((int) k_colors.size() - 1, (int) (std::pow(p, 3)*float(k_colors.size()))));
+
+                printf("%s%s%s%s", speaker.c_str(), k_colors[col].c_str(), text, "\033[0m");
+            }
+        } else {
+            const char * text = whisper_full_get_segment_text(ctx, i);
+
+            printf("%s%s", speaker.c_str(), text);
+        }
+
+        if (params.tinydiarize) {
+            if (whisper_full_get_segment_speaker_turn_next(ctx, i)) {
+                printf("%s", params.tdrz_speaker_turn.c_str());
+            }
+        }
+
+        // with timestamps or speakers: each segment on new line
+        if (!params.no_timestamps || params.diarize) {
+            printf("\n");
+        }
+
+        fflush(stdout);
+    }
+}
+
+static bool output_txt(struct whisper_context * ctx, const char * fname, const whisper_params & params, std::vector<std::vector<float>> pcmf32s) {
+    std::ofstream fout(fname);
+    if (!fout.is_open()) {
+        fprintf(stderr, "%s: failed to open '%s' for writing\n", __func__, fname);
+        return false;
+    }
+
+    fprintf(stderr, "%s: saving output to '%s'\n", __func__, fname);
+
+    const int n_segments = whisper_full_n_segments(ctx);
+    for (int i = 0; i < n_segments; ++i) {
+        const char * text = whisper_full_get_segment_text(ctx, i);
+        std::string speaker = "";
+
+        if (params.diarize && pcmf32s.size() == 2)
+        {
+            const int64_t t0 = whisper_full_get_segment_t0(ctx, i);
+            const int64_t t1 = whisper_full_get_segment_t1(ctx, i);
+            speaker = estimate_diarization_speaker(pcmf32s, t0, t1);
+        }
+
+        fout << speaker << text << "\n";
+    }
+
+    return true;
+}
+
+static bool output_vtt(struct whisper_context * ctx, const char * fname, const whisper_params & params, std::vector<std::vector<float>> pcmf32s) {
+    std::ofstream fout(fname);
+    if (!fout.is_open()) {
+        fprintf(stderr, "%s: failed to open '%s' for writing\n", __func__, fname);
+        return false;
+    }
+
+    fprintf(stderr, "%s: saving output to '%s'\n", __func__, fname);
+
+    fout << "WEBVTT\n\n";
+
+    const int n_segments = whisper_full_n_segments(ctx);
+    for (int i = 0; i < n_segments; ++i) {
+        const char * text = whisper_full_get_segment_text(ctx, i);
+        const int64_t t0 = whisper_full_get_segment_t0(ctx, i);
+        const int64_t t1 = whisper_full_get_segment_t1(ctx, i);
+        std::string speaker = "";
+
+        if (params.diarize && pcmf32s.size() == 2)
+        {
+            speaker = estimate_diarization_speaker(pcmf32s, t0, t1, true);
+            speaker.insert(0, "<v Speaker");
+            speaker.append(">");
+        }
+
+        fout << to_timestamp(t0) << " --> " << to_timestamp(t1) << "\n";
+        fout << speaker << text << "\n\n";
+    }
+
+    return true;
+}
+
+static bool output_srt(struct whisper_context * ctx, const char * fname, const whisper_params & params, std::vector<std::vector<float>> pcmf32s) {
+    std::ofstream fout(fname);
+    if (!fout.is_open()) {
+        fprintf(stderr, "%s: failed to open '%s' for writing\n", __func__, fname);
+        return false;
+    }
+
+    fprintf(stderr, "%s: saving output to '%s'\n", __func__, fname);
+
+    const int n_segments = whisper_full_n_segments(ctx);
+    for (int i = 0; i < n_segments; ++i) {
+        const char * text = whisper_full_get_segment_text(ctx, i);
+        const int64_t t0 = whisper_full_get_segment_t0(ctx, i);
+        const int64_t t1 = whisper_full_get_segment_t1(ctx, i);
+        std::string speaker = "";
+
+        if (params.diarize && pcmf32s.size() == 2)
+        {
+            speaker = estimate_diarization_speaker(pcmf32s, t0, t1);
+        }
+
+        fout << i + 1 + params.offset_n << "\n";
+        fout << to_timestamp(t0, true) << " --> " << to_timestamp(t1, true) << "\n";
+        fout << speaker << text << "\n\n";
+    }
+
+    return true;
+}
+
+static char * escape_double_quotes_and_backslashes(const char * str) {
+    if (str == NULL) {
+        return NULL;
+    }
+
+    size_t escaped_length = strlen(str) + 1;
+
+    for (size_t i = 0; str[i] != '\0'; i++) {
+        if (str[i] == '"' || str[i] == '\\') {
+            escaped_length++;
+        }
+    }
+
+    char * escaped = (char *)calloc(escaped_length, 1); // pre-zeroed
+    if (escaped == NULL) {
+        return NULL;
+    }
+
+    size_t pos = 0;
+    for (size_t i = 0; str[i] != '\0'; i++) {
+        if (str[i] == '"' || str[i] == '\\') {
+            escaped[pos++] = '\\';
+        }
+        escaped[pos++] = str[i];
+    }
+
+    // no need to set zero due to calloc() being used prior
+
+    return escaped;
+}
+
+// double quote should be escaped by another double quote. (rfc4180)
+static char * escape_double_quotes_in_csv(const char * str) {
+    if (str == NULL) {
+        return NULL;
+    }
+
+    size_t escaped_length = strlen(str) + 1;
+
+    for (size_t i = 0; str[i] != '\0'; i++) {
+        if (str[i] == '"') {
+            escaped_length++;
+        }
+    }
+
+    char *escaped = (char *)calloc(escaped_length, 1); // pre-zeroed
+    if (escaped == NULL) {
+        return NULL;
+    }
+
+    size_t pos = 0;
+    for (size_t i = 0; str[i] != '\0'; i++) {
+        if (str[i] == '"') {
+            escaped[pos++] = '"';
+        }
+        escaped[pos++] = str[i];
+    }
+
+    // no need to set zero due to calloc() being used prior
+
+    return escaped;
+}
+
+static bool output_csv(struct whisper_context * ctx, const char * fname, const whisper_params & params, std::vector<std::vector<float>> pcmf32s) {
+    std::ofstream fout(fname);
+    if (!fout.is_open()) {
+        fprintf(stderr, "%s: failed to open '%s' for writing\n", __func__, fname);
+        return false;
+    }
+
+    fprintf(stderr, "%s: saving output to '%s'\n", __func__, fname);
+
+    const int n_segments = whisper_full_n_segments(ctx);
+    fout << "start,end,";
+    if (params.diarize && pcmf32s.size() == 2)
+    {
+        fout << "speaker,";
+    }
+    fout << "text\n";
+
+    for (int i = 0; i < n_segments; ++i) {
+        const char * text = whisper_full_get_segment_text(ctx, i);
+        const int64_t t0 = whisper_full_get_segment_t0(ctx, i);
+        const int64_t t1 = whisper_full_get_segment_t1(ctx, i);
+        char * text_escaped = escape_double_quotes_in_csv(text);
+
+        //need to multiply times returned from whisper_full_get_segment_t{0,1}() by 10 to get milliseconds.
+        fout << 10 * t0 << "," << 10 * t1 << ",";
+        if (params.diarize && pcmf32s.size() == 2)
+        {
+            fout << estimate_diarization_speaker(pcmf32s, t0, t1, true) << ",";
+        }
+        fout << "\"" << text_escaped << "\"\n";
+    }
+
+    return true;
+}
+
+static bool output_score(struct whisper_context * ctx, const char * fname, const whisper_params & /*params*/, std::vector<std::vector<float>> /*pcmf32s*/) {
+    std::ofstream fout(fname);
+    fprintf(stderr, "%s: saving output to '%s'\n", __func__, fname);
+
+    const int n_segments = whisper_full_n_segments(ctx);
+    // fprintf(stderr,"segments: %d\n",n_segments);
+    for (int i = 0; i < n_segments; ++i) {
+        const int n_tokens = whisper_full_n_tokens(ctx, i);
+        // fprintf(stderr,"tokens: %d\n",n_tokens);
+        for (int j = 0; j < n_tokens; j++) {
+            auto token = whisper_full_get_token_text(ctx, i, j);
+            auto probability = whisper_full_get_token_p(ctx, i, j);
+            fout << token << '\t' << probability << std::endl;
+            // fprintf(stderr,"token: %s %f\n",token,probability);
+	    }
+    }
+    return true;
+}
+
+static bool output_json(
+             struct whisper_context * ctx,
+                         const char * fname,
+               const whisper_params & params,
+    std::vector<std::vector<float>>   pcmf32s,
+                               bool   full) {
+    std::ofstream fout(fname);
+    int indent = 0;
+
+    auto doindent = [&]() {
+        for (int i = 0; i < indent; i++) fout << "\t";
+    };
+
+    auto start_arr = [&](const char *name) {
+        doindent();
+        fout << "\"" << name << "\": [\n";
+        indent++;
+    };
+
+    auto end_arr = [&](bool end) {
+        indent--;
+        doindent();
+        fout << (end ? "]\n" : "],\n");
+    };
+
+    auto start_obj = [&](const char *name) {
+        doindent();
+        if (name) {
+            fout << "\"" << name << "\": {\n";
+        } else {
+            fout << "{\n";
+        }
+        indent++;
+    };
+
+    auto end_obj = [&](bool end) {
+        indent--;
+        doindent();
+        fout << (end ? "}\n" : "},\n");
+    };
+
+    auto start_value = [&](const char *name) {
+        doindent();
+        fout << "\"" << name << "\": ";
+    };
+
+    auto value_s = [&](const char *name, const char *val, bool end) {
+        start_value(name);
+        char * val_escaped = escape_double_quotes_and_backslashes(val);
+        fout << "\"" << val_escaped << (end ? "\"\n" : "\",\n");
+        free(val_escaped);
+    };
+
+    auto end_value = [&](bool end) {
+        fout << (end ? "\n" : ",\n");
+    };
+
+    auto value_i = [&](const char *name, const int64_t val, bool end) {
+        start_value(name);
+        fout << val;
+        end_value(end);
+    };
+
+    auto value_f = [&](const char *name, const float val, bool end) {
+        start_value(name);
+        fout << val;
+        end_value(end);
+    };
+
+    auto value_b = [&](const char *name, const bool val, bool end) {
+        start_value(name);
+        fout << (val ? "true" : "false");
+        end_value(end);
+    };
+
+    auto times_o = [&](int64_t t0, int64_t t1, bool end) {
+        start_obj("timestamps");
+        value_s("from", to_timestamp(t0, true).c_str(), false);
+        value_s("to", to_timestamp(t1, true).c_str(), true);
+        end_obj(false);
+        start_obj("offsets");
+        value_i("from", t0 * 10, false);
+        value_i("to", t1 * 10, true);
+        end_obj(end);
+    };
+
+    if (!fout.is_open()) {
+        fprintf(stderr, "%s: failed to open '%s' for writing\n", __func__, fname);
+        return false;
+    }
+
+    fprintf(stderr, "%s: saving output to '%s'\n", __func__, fname);
+    start_obj(nullptr);
+        value_s("systeminfo", whisper_print_system_info(), false);
+        start_obj("model");
+            value_s("type", whisper_model_type_readable(ctx), false);
+            value_b("multilingual", whisper_is_multilingual(ctx), false);
+            value_i("vocab", whisper_model_n_vocab(ctx), false);
+            start_obj("audio");
+                value_i("ctx", whisper_model_n_audio_ctx(ctx), false);
+                value_i("state", whisper_model_n_audio_state(ctx), false);
+                value_i("head", whisper_model_n_audio_head(ctx), false);
+                value_i("layer", whisper_model_n_audio_layer(ctx), true);
+            end_obj(false);
+            start_obj("text");
+                value_i("ctx", whisper_model_n_text_ctx(ctx), false);
+                value_i("state", whisper_model_n_text_state(ctx), false);
+                value_i("head", whisper_model_n_text_head(ctx), false);
+                value_i("layer", whisper_model_n_text_layer(ctx), true);
+            end_obj(false);
+            value_i("mels", whisper_model_n_mels(ctx), false);
+            value_i("ftype", whisper_model_ftype(ctx), true);
+        end_obj(false);
+        start_obj("params");
+            value_s("model", params.model.c_str(), false);
+            value_s("language", params.language.c_str(), false);
+            value_b("translate", params.translate, true);
+        end_obj(false);
+        start_obj("result");
+            value_s("language", whisper_lang_str(whisper_full_lang_id(ctx)), true);
+        end_obj(false);
+        start_arr("transcription");
+
+            const int n_segments = whisper_full_n_segments(ctx);
+            for (int i = 0; i < n_segments; ++i) {
+                const char * text = whisper_full_get_segment_text(ctx, i);
+
+                const int64_t t0 = whisper_full_get_segment_t0(ctx, i);
+                const int64_t t1 = whisper_full_get_segment_t1(ctx, i);
+
+                start_obj(nullptr);
+                    times_o(t0, t1, false);
+                    value_s("text", text, !params.diarize && !params.tinydiarize && !full);
+
+                    if (full) {
+                        start_arr("tokens");
+                        const int n = whisper_full_n_tokens(ctx, i);
+                        for (int j = 0; j < n; ++j) {
+                            auto token = whisper_full_get_token_data(ctx, i, j);
+                            start_obj(nullptr);
+                                value_s("text", whisper_token_to_str(ctx, token.id), false);
+                                if(token.t0 > -1 && token.t1 > -1) {
+                                    // If we have per-token timestamps, write them out
+                                    times_o(token.t0, token.t1, false);
+                                }
+                                value_i("id", token.id, false);
+                                value_f("p", token.p, false);
+                                value_f("t_dtw", token.t_dtw, true);
+                            end_obj(j == (n - 1));
+                        }
+                        end_arr(!params.diarize && !params.tinydiarize);
+                    }
+
+                    if (params.diarize && pcmf32s.size() == 2) {
+                        value_s("speaker", estimate_diarization_speaker(pcmf32s, t0, t1, true).c_str(), true);
+                    }
+
+                    if (params.tinydiarize) {
+                        value_b("speaker_turn_next", whisper_full_get_segment_speaker_turn_next(ctx, i), true);
+                    }
+                end_obj(i == (n_segments - 1));
+            }
+
+        end_arr(true);
+    end_obj(true);
+    return true;
+}
+
+// karaoke video generation
+// outputs a bash script that uses ffmpeg to generate a video with the subtitles
+// TODO: font parameter adjustments
+static bool output_wts(struct whisper_context * ctx, const char * fname, const char * fname_inp, const whisper_params & params, float t_sec, std::vector<std::vector<float>> pcmf32s) {
+    std::ofstream fout(fname);
+
+    fprintf(stderr, "%s: saving output to '%s'\n", __func__, fname);
+
+    static const char * font = params.font_path.c_str();
+
+    std::ifstream fin(font);
+    if (!fin.is_open()) {
+        fprintf(stderr, "%s: font not found at '%s', please specify a monospace font with -fp\n", __func__, font);
+        return false;
+    }
+
+    fout << "#!/bin/bash" << "\n";
+    fout << "\n";
+
+    fout << "ffmpeg -i " << fname_inp << " -f lavfi -i color=size=1200x120:duration=" << t_sec << ":rate=25:color=black -vf \"";
+
+    for (int i = 0; i < whisper_full_n_segments(ctx); i++) {
+        const int64_t t0 = whisper_full_get_segment_t0(ctx, i);
+        const int64_t t1 = whisper_full_get_segment_t1(ctx, i);
+
+        const int n = whisper_full_n_tokens(ctx, i);
+
+        std::vector<whisper_token_data> tokens(n);
+        for (int j = 0; j < n; ++j) {
+            tokens[j] = whisper_full_get_token_data(ctx, i, j);
+        }
+
+        if (i > 0) {
+            fout << ",";
+        }
+
+        // background text
+        fout << "drawtext=fontfile='" << font << "':fontsize=24:fontcolor=gray:x=(w-text_w)/2:y=h/2:text='':enable='between(t," << t0/100.0 << "," << t0/100.0 << ")'";
+
+        bool is_first = true;
+        std::string speaker = "";
+
+        if (params.diarize && pcmf32s.size() == 2) {
+            speaker = estimate_diarization_speaker(pcmf32s, t0, t1);
+        }
+
+        for (int j = 0; j < n; ++j) {
+            const auto & token = tokens[j];
+
+            if (tokens[j].id >= whisper_token_eot(ctx)) {
+                continue;
+            }
+
+            std::string txt_bg = "";
+            std::string txt_fg = ""; // highlight token
+            std::string txt_ul = ""; // underline
+
+            if (params.diarize && pcmf32s.size() == 2) {
+                txt_bg = speaker;
+                txt_fg = speaker;
+                txt_ul = "\\ \\ \\ \\ \\ \\ \\ \\ \\ \\ \\ ";
+            }
+
+            txt_bg.append("> ");
+            txt_fg.append("> ");
+            txt_ul.append("\\ \\ ");
+
+            {
+                for (int k = 0; k < n; ++k) {
+                    const auto & token2 = tokens[k];
+
+                    if (tokens[k].id >= whisper_token_eot(ctx)) {
+                        continue;
+                    }
+
+                    const std::string txt = whisper_token_to_str(ctx, token2.id);
+
+                    txt_bg += txt;
+
+                    if (k == j) {
+                        for (int l = 0; l < (int) txt.size(); ++l) {
+                            txt_fg += txt[l];
+                            txt_ul += "_";
+                        }
+                        txt_fg += "|";
+                    } else {
+                        for (int l = 0; l < (int) txt.size(); ++l) {
+                            txt_fg += "\\ ";
+                            txt_ul += "\\ ";
+                        }
+                    }
+                }
+
+                ::replace_all(txt_bg, "'", "\u2019");
+                ::replace_all(txt_bg, "\"", "\\\"");
+                ::replace_all(txt_fg, "'", "\u2019");
+                ::replace_all(txt_fg, "\"", "\\\"");
+            }
+
+            if (is_first) {
+                // background text
+                fout << ",drawtext=fontfile='" << font << "':fontsize=24:fontcolor=gray:x=(w-text_w)/2:y=h/2:text='" << txt_bg << "':enable='between(t," << t0/100.0 << "," << t1/100.0 << ")'";
+                is_first = false;
+            }
+
+            // foreground text
+            fout << ",drawtext=fontfile='" << font << "':fontsize=24:fontcolor=lightgreen:x=(w-text_w)/2+8:y=h/2:text='" << txt_fg << "':enable='between(t," << token.t0/100.0 << "," << token.t1/100.0 << ")'";
+
+            // underline
+            fout << ",drawtext=fontfile='" << font << "':fontsize=24:fontcolor=lightgreen:x=(w-text_w)/2+8:y=h/2+16:text='" << txt_ul << "':enable='between(t," << token.t0/100.0 << "," << token.t1/100.0 << ")'";
+        }
+    }
+
+    fout << "\" -c:v libx264 -pix_fmt yuv420p -y " << fname_inp << ".mp4" << "\n";
+
+    fout << "\n\n";
+    fout << "echo \"Your video has been saved to " << fname_inp << ".mp4\"" << "\n";
+    fout << "\n";
+    fout << "echo \"  ffplay " << fname_inp << ".mp4\"\n";
+    fout << "\n";
+
+    fout.close();
+
+    fprintf(stderr, "%s: run 'source %s' to generate karaoke video\n", __func__, fname);
+
+    return true;
+}
+
+static bool output_lrc(struct whisper_context * ctx, const char * fname, const whisper_params & params, std::vector<std::vector<float>> pcmf32s) {
+    std::ofstream fout(fname);
+    if (!fout.is_open()) {
+        fprintf(stderr, "%s: failed to open '%s' for writing\n", __func__, fname);
+        return false;
+    }
+
+    fprintf(stderr, "%s: saving output to '%s'\n", __func__, fname);
+
+    fout << "[by:whisper.cpp]\n";
+
+    const int n_segments = whisper_full_n_segments(ctx);
+    for (int i = 0; i < n_segments; ++i) {
+        const char * text = whisper_full_get_segment_text(ctx, i);
+        const int64_t t = whisper_full_get_segment_t0(ctx, i);
+
+        int64_t msec = t * 10;
+        int64_t min = msec / (1000 * 60);
+        msec = msec - min * (1000 * 60);
+        int64_t sec = msec / 1000;
+        msec = msec - sec * 1000;
+
+        char buf[16];
+        snprintf(buf, sizeof(buf), "%02d:%02d.%02d", (int) min, (int) sec, (int) ( msec / 10));
+        std::string timestamp_lrc = std::string(buf);
+        std::string speaker = "";
+
+        if (params.diarize && pcmf32s.size() == 2)
+        {
+            const int64_t t0 = whisper_full_get_segment_t0(ctx, i);
+            const int64_t t1 = whisper_full_get_segment_t1(ctx, i);
+            speaker = estimate_diarization_speaker(pcmf32s, t0, t1);
+        }
+
+        fout <<  '[' << timestamp_lrc << ']' << speaker << text << "\n";
+    }
+
+    return true;
+}
+
+
+static void cb_log_disable(enum ggml_log_level , const char * , void * ) { }
+
+int main(int argc, char ** argv) {
+    whisper_params params;
+
+    // If the only argument starts with "@", read arguments line-by-line
+    // from the given file.
+    std::vector<std::string> vec_args;
+    if (argc == 2 && argv != nullptr && argv[1] != nullptr && argv[1][0] == '@') {
+        // Save the name of the executable.
+        vec_args.push_back(argv[0]);
+
+        // Open the response file.
+        char const * rspfile = argv[1] + sizeof(char);
+        std::ifstream fin(rspfile);
+        if (fin.is_open() == false) {
+            fprintf(stderr, "error: response file '%s' not found\n", rspfile);
+            return 1;
+        }
+
+        // Read the entire response file.
+        std::string line;
+        while (std::getline(fin, line)) {
+            vec_args.push_back(line);
+        }
+
+        // Use the contents of the response file as the command-line arguments.
+        argc = static_cast<int>(vec_args.size());
+        argv = static_cast<char **>(alloca(argc * sizeof (char *)));
+        for (int i = 0; i < argc; ++i) {
+            argv[i] = const_cast<char *>(vec_args[i].c_str());
+        }
+    }
+
+    if (whisper_params_parse(argc, argv, params) == false) {
+        whisper_print_usage(argc, argv, params);
+        return 1;
+    }
+
+    // remove non-existent files
+    for (auto it = params.fname_inp.begin(); it != params.fname_inp.end();) {
+        const auto fname_inp = it->c_str();
+
+        if (*it != "-" && !is_file_exist(fname_inp)) {
+            fprintf(stderr, "error: input file not found '%s'\n", fname_inp);
+            it = params.fname_inp.erase(it);
+            continue;
+        }
+
+        it++;
+    }
+
+    if (params.fname_inp.empty()) {
+        fprintf(stderr, "error: no input files specified\n");
+        whisper_print_usage(argc, argv, params);
+        return 2;
+    }
+
+    if (params.language != "auto" && whisper_lang_id(params.language.c_str()) == -1) {
+        fprintf(stderr, "error: unknown language '%s'\n", params.language.c_str());
+        whisper_print_usage(argc, argv, params);
+        exit(0);
+    }
+
+    if (params.diarize && params.tinydiarize) {
+        fprintf(stderr, "error: cannot use both --diarize and --tinydiarize\n");
+        whisper_print_usage(argc, argv, params);
+        exit(0);
+    }
+
+    if (params.no_prints) {
+        whisper_log_set(cb_log_disable, NULL);
+    }
+
+    // whisper init
+
+    struct whisper_context_params cparams = whisper_context_default_params();
+
+    cparams.use_gpu    = params.use_gpu;
+    cparams.flash_attn = params.flash_attn;
+
+    if (!params.dtw.empty()) {
+        cparams.dtw_token_timestamps = true;
+        cparams.dtw_aheads_preset = WHISPER_AHEADS_NONE;
+
+        if (params.dtw == "tiny")      cparams.dtw_aheads_preset = WHISPER_AHEADS_TINY;
+        if (params.dtw == "tiny.en")   cparams.dtw_aheads_preset = WHISPER_AHEADS_TINY_EN;
+        if (params.dtw == "base")      cparams.dtw_aheads_preset = WHISPER_AHEADS_BASE;
+        if (params.dtw == "base.en")   cparams.dtw_aheads_preset = WHISPER_AHEADS_BASE_EN;
+        if (params.dtw == "small")     cparams.dtw_aheads_preset = WHISPER_AHEADS_SMALL;
+        if (params.dtw == "small.en")  cparams.dtw_aheads_preset = WHISPER_AHEADS_SMALL_EN;
+        if (params.dtw == "medium")    cparams.dtw_aheads_preset = WHISPER_AHEADS_MEDIUM;
+        if (params.dtw == "medium.en") cparams.dtw_aheads_preset = WHISPER_AHEADS_MEDIUM_EN;
+        if (params.dtw == "large.v1")  cparams.dtw_aheads_preset = WHISPER_AHEADS_LARGE_V1;
+        if (params.dtw == "large.v2")  cparams.dtw_aheads_preset = WHISPER_AHEADS_LARGE_V2;
+        if (params.dtw == "large.v3")  cparams.dtw_aheads_preset = WHISPER_AHEADS_LARGE_V3;
+
+        if (cparams.dtw_aheads_preset == WHISPER_AHEADS_NONE) {
+            fprintf(stderr, "error: unknown DTW preset '%s'\n", params.dtw.c_str());
+            return 3;
+        }
+    }
+
+    struct whisper_context * ctx = whisper_init_from_file_with_params(params.model.c_str(), cparams);
+
+    if (ctx == nullptr) {
+        fprintf(stderr, "error: failed to initialize whisper context\n");
+        return 3;
+    }
+
+    // initialize openvino encoder. this has no effect on whisper.cpp builds that don't have OpenVINO configured
+    whisper_ctx_init_openvino_encoder(ctx, nullptr, params.openvino_encode_device.c_str(), nullptr);
+
+    if (!params.grammar.empty()) {
+        auto & grammar = params.grammar_parsed;
+        if (is_file_exist(params.grammar.c_str())) {
+            // read grammar from file
+            std::ifstream ifs(params.grammar.c_str());
+            const std::string txt = std::string((std::istreambuf_iterator<char>(ifs)), std::istreambuf_iterator<char>());
+            grammar = grammar_parser::parse(txt.c_str());
+        } else {
+            // read grammar from string
+            grammar = grammar_parser::parse(params.grammar.c_str());
+        }
+
+        // will be empty (default) if there are parse errors
+        if (grammar.rules.empty()) {
+            fprintf(stderr, "error: failed to parse grammar \"%s\"\n", params.grammar.c_str());
+            return 4;
+        } else {
+            fprintf(stderr, "%s: grammar:\n", __func__);
+            grammar_parser::print_grammar(stderr, grammar);
+            fprintf(stderr, "\n");
+        }
+    }
+
+    for (int f = 0; f < (int) params.fname_inp.size(); ++f) {
+        const auto fname_inp = params.fname_inp[f];
+		const auto fname_out = f < (int) params.fname_out.size() && !params.fname_out[f].empty() ? params.fname_out[f] : params.fname_inp[f];
+
+        std::vector<float> pcmf32;               // mono-channel F32 PCM
+        std::vector<std::vector<float>> pcmf32s; // stereo-channel F32 PCM
+
+        if (!::read_wav(fname_inp, pcmf32, pcmf32s, params.diarize)) {
+            fprintf(stderr, "error: failed to read WAV file '%s'\n", fname_inp.c_str());
+            continue;
+        }
+
+        if (!whisper_is_multilingual(ctx)) {
+            if (params.language != "en" || params.translate) {
+                params.language = "en";
+                params.translate = false;
+                fprintf(stderr, "%s: WARNING: model is not multilingual, ignoring language and translation options\n", __func__);
+            }
+        }
+        if (params.detect_language) {
+            params.language = "auto";
+        }
+
+        if (!params.no_prints) {
+            // print system information
+            fprintf(stderr, "\n");
+            fprintf(stderr, "system_info: n_threads = %d / %d | %s\n",
+                    params.n_threads*params.n_processors, std::thread::hardware_concurrency(), whisper_print_system_info());
+
+            // print some info about the processing
+            fprintf(stderr, "\n");
+            fprintf(stderr, "%s: processing '%s' (%d samples, %.1f sec), %d threads, %d processors, %d beams + best of %d, lang = %s, task = %s, %stimestamps = %d ...\n",
+                    __func__, fname_inp.c_str(), int(pcmf32.size()), float(pcmf32.size())/WHISPER_SAMPLE_RATE,
+                    params.n_threads, params.n_processors, params.beam_size, params.best_of,
+                    params.language.c_str(),
+                    params.translate ? "translate" : "transcribe",
+                    params.tinydiarize ? "tdrz = 1, " : "",
+                    params.no_timestamps ? 0 : 1);
+
+            fprintf(stderr, "\n");
+        }
+
+        // run the inference
+        {
+            whisper_full_params wparams = whisper_full_default_params(WHISPER_SAMPLING_GREEDY);
+
+            const bool use_grammar = (!params.grammar_parsed.rules.empty() && !params.grammar_rule.empty());
+            wparams.strategy = (params.beam_size > 1 || use_grammar) ? WHISPER_SAMPLING_BEAM_SEARCH : WHISPER_SAMPLING_GREEDY;
+
+            wparams.print_realtime   = false;
+            wparams.print_progress   = params.print_progress;
+            wparams.print_timestamps = !params.no_timestamps;
+            wparams.print_special    = params.print_special;
+            wparams.translate        = params.translate;
+            wparams.language         = params.language.c_str();
+            wparams.detect_language  = params.detect_language;
+            wparams.n_threads        = params.n_threads;
+            wparams.n_max_text_ctx   = params.max_context >= 0 ? params.max_context : wparams.n_max_text_ctx;
+            wparams.offset_ms        = params.offset_t_ms;
+            wparams.duration_ms      = params.duration_ms;
+
+            wparams.token_timestamps = params.output_wts || params.output_jsn_full || params.max_len > 0;
+            wparams.thold_pt         = params.word_thold;
+            wparams.max_len          = params.output_wts && params.max_len == 0 ? 60 : params.max_len;
+            wparams.split_on_word    = params.split_on_word;
+            wparams.audio_ctx        = params.audio_ctx;
+
+            wparams.debug_mode       = params.debug_mode;
+
+            wparams.tdrz_enable      = params.tinydiarize; // [TDRZ]
+
+            wparams.suppress_regex   = params.suppress_regex.empty() ? nullptr : params.suppress_regex.c_str();
+
+            wparams.initial_prompt   = params.prompt.c_str();
+
+            wparams.greedy.best_of        = params.best_of;
+            wparams.beam_search.beam_size = params.beam_size;
+
+            wparams.temperature_inc  = params.no_fallback ? 0.0f : params.temperature_inc;
+            wparams.temperature      = params.temperature;
+
+            wparams.entropy_thold    = params.entropy_thold;
+            wparams.logprob_thold    = params.logprob_thold;
+
+            wparams.no_timestamps    = params.no_timestamps;
+
+            whisper_print_user_data user_data = { &params, &pcmf32s, 0 };
+
+            const auto & grammar_parsed = params.grammar_parsed;
+            auto grammar_rules = grammar_parsed.c_rules();
+
+            if (use_grammar) {
+                if (grammar_parsed.symbol_ids.find(params.grammar_rule) == grammar_parsed.symbol_ids.end()) {
+                    fprintf(stderr, "%s: warning: grammar rule '%s' not found - skipping grammar sampling\n", __func__, params.grammar_rule.c_str());
+                } else {
+                    wparams.grammar_rules = grammar_rules.data();
+                    wparams.n_grammar_rules = grammar_rules.size();
+                    wparams.i_start_rule = grammar_parsed.symbol_ids.at(params.grammar_rule);
+                    wparams.grammar_penalty = params.grammar_penalty;
+                }
+            }
+
+            // this callback is called on each new segment
+            if (!wparams.print_realtime) {
+                wparams.new_segment_callback           = whisper_print_segment_callback;
+                wparams.new_segment_callback_user_data = &user_data;
+            }
+
+            if (wparams.print_progress) {
+                wparams.progress_callback           = whisper_print_progress_callback;
+                wparams.progress_callback_user_data = &user_data;
+            }
+
+            // examples for abort mechanism
+            // in examples below, we do not abort the processing, but we could if the flag is set to true
+
+            // the callback is called before every encoder run - if it returns false, the processing is aborted
+            {
+                static bool is_aborted = false; // NOTE: this should be atomic to avoid data race
+
+                wparams.encoder_begin_callback = [](struct whisper_context * /*ctx*/, struct whisper_state * /*state*/, void * user_data) {
+                    bool is_aborted = *(bool*)user_data;
+                    return !is_aborted;
+                };
+                wparams.encoder_begin_callback_user_data = &is_aborted;
+            }
+
+            // the callback is called before every computation - if it returns true, the computation is aborted
+            {
+                static bool is_aborted = false; // NOTE: this should be atomic to avoid data race
+
+                wparams.abort_callback = [](void * user_data) {
+                    bool is_aborted = *(bool*)user_data;
+                    return is_aborted;
+                };
+                wparams.abort_callback_user_data = &is_aborted;
+            }
+
+            if (whisper_full_parallel(ctx, wparams, pcmf32.data(), pcmf32.size(), params.n_processors) != 0) {
+                fprintf(stderr, "%s: failed to process audio\n", argv[0]);
+                return 10;
+            }
+        }
+
+        // output stuff
+        {
+            printf("\n");
+
+            // output to text file
+            if (params.output_txt) {
+                const auto fname_txt = fname_out + ".txt";
+                output_txt(ctx, fname_txt.c_str(), params, pcmf32s);
+            }
+
+            // output to VTT file
+            if (params.output_vtt) {
+                const auto fname_vtt = fname_out + ".vtt";
+                output_vtt(ctx, fname_vtt.c_str(), params, pcmf32s);
+            }
+
+            // output to SRT file
+            if (params.output_srt) {
+                const auto fname_srt = fname_out + ".srt";
+                output_srt(ctx, fname_srt.c_str(), params, pcmf32s);
+            }
+
+            // output to WTS file
+            if (params.output_wts) {
+                const auto fname_wts = fname_out + ".wts";
+                output_wts(ctx, fname_wts.c_str(), fname_inp.c_str(), params, float(pcmf32.size() + 1000)/WHISPER_SAMPLE_RATE, pcmf32s);
+            }
+
+            // output to CSV file
+            if (params.output_csv) {
+                const auto fname_csv = fname_out + ".csv";
+                output_csv(ctx, fname_csv.c_str(), params, pcmf32s);
+            }
+
+            // output to JSON file
+            if (params.output_jsn) {
+                const auto fname_jsn = fname_out + ".json";
+                output_json(ctx, fname_jsn.c_str(), params, pcmf32s, params.output_jsn_full);
+            }
+
+            // output to LRC file
+            if (params.output_lrc) {
+                const auto fname_lrc = fname_out + ".lrc";
+                output_lrc(ctx, fname_lrc.c_str(), params, pcmf32s);
+            }
+
+            // output to score file
+            if (params.log_score) {
+                const auto fname_score = fname_out + ".score.txt";
+                output_score(ctx, fname_score.c_str(), params, pcmf32s);
+            }
+        }
+    }
+
+    if (!params.no_prints) {
+        whisper_print_timings(ctx);
+    }
+    whisper_free(ctx);
+
+    return 0;
+}
diff --git a/src/whisper_cpp/examples/server/CMakeLists.txt b/src/whisper_cpp/examples/server/CMakeLists.txt
new file mode 100644
index 00000000..96dd97f7
--- /dev/null
+++ b/src/whisper_cpp/examples/server/CMakeLists.txt
@@ -0,0 +1,10 @@
+set(TARGET server)
+add_executable(${TARGET} server.cpp httplib.h)
+
+include(DefaultTargetOptions)
+
+target_link_libraries(${TARGET} PRIVATE common json_cpp whisper ${CMAKE_THREAD_LIBS_INIT})
+
+if (WIN32)
+    target_link_libraries(${TARGET} PRIVATE ws2_32)
+endif()
diff --git a/src/whisper_cpp/examples/server/README.md b/src/whisper_cpp/examples/server/README.md
new file mode 100644
index 00000000..596fd769
--- /dev/null
+++ b/src/whisper_cpp/examples/server/README.md
@@ -0,0 +1,69 @@
+# whisper.cpp http server
+
+Simple http server. WAV Files are passed to the inference model via http requests.
+
+https://github.com/ggerganov/whisper.cpp/assets/1991296/e983ee53-8741-4eb5-9048-afe5e4594b8f
+
+## Usage
+
+```
+./server -h
+
+usage: ./bin/server [options]
+
+options:
+  -h,        --help              [default] show this help message and exit
+  -t N,      --threads N         [4      ] number of threads to use during computation
+  -p N,      --processors N      [1      ] number of processors to use during computation
+  -ot N,     --offset-t N        [0      ] time offset in milliseconds
+  -on N,     --offset-n N        [0      ] segment index offset
+  -d  N,     --duration N        [0      ] duration of audio to process in milliseconds
+  -mc N,     --max-context N     [-1     ] maximum number of text context tokens to store
+  -ml N,     --max-len N         [0      ] maximum segment length in characters
+  -sow,      --split-on-word     [false  ] split on word rather than on token
+  -bo N,     --best-of N         [2      ] number of best candidates to keep
+  -bs N,     --beam-size N       [-1     ] beam size for beam search
+  -wt N,     --word-thold N      [0.01   ] word timestamp probability threshold
+  -et N,     --entropy-thold N   [2.40   ] entropy threshold for decoder fail
+  -lpt N,    --logprob-thold N   [-1.00  ] log probability threshold for decoder fail
+  -debug,    --debug-mode        [false  ] enable debug mode (eg. dump log_mel)
+  -tr,       --translate         [false  ] translate from source language to english
+  -di,       --diarize           [false  ] stereo audio diarization
+  -tdrz,     --tinydiarize       [false  ] enable tinydiarize (requires a tdrz model)
+  -nf,       --no-fallback       [false  ] do not use temperature fallback while decoding
+  -ps,       --print-special     [false  ] print special tokens
+  -pc,       --print-colors      [false  ] print colors
+  -pr,       --print-realtime    [false  ] print output in realtime
+  -pp,       --print-progress    [false  ] print progress
+  -nt,       --no-timestamps     [false  ] do not print timestamps
+  -l LANG,   --language LANG     [en     ] spoken language ('auto' for auto-detect)
+  -dl,       --detect-language   [false  ] exit after automatically detecting language
+             --prompt PROMPT     [       ] initial prompt
+  -m FNAME,  --model FNAME       [models/ggml-base.en.bin] model path
+  -oved D,   --ov-e-device DNAME [CPU    ] the OpenVINO device used for encode inference
+  --host HOST,                   [127.0.0.1] Hostname/ip-adress for the server
+  --port PORT,                   [8080   ] Port number for the server
+  --convert,                     [false  ] Convert audio to WAV, requires ffmpeg on the server
+```
+
+> [!WARNING]
+> **Do not run the server example with administrative privileges and ensure it's operated in a sandbox environment, especially since it involves risky operations like accepting user file uploads and using ffmpeg for format conversions. Always validate and sanitize inputs to guard against potential security threats.**
+
+## request examples
+
+**/inference**
+```
+curl 127.0.0.1:8080/inference \
+-H "Content-Type: multipart/form-data" \
+-F file="@<file-path>" \
+-F temperature="0.0" \
+-F temperature_inc="0.2" \
+-F response_format="json"
+```
+
+**/load**
+```
+curl 127.0.0.1:8080/load \
+-H "Content-Type: multipart/form-data" \
+-F model="<path-to-model-file>"
+```
diff --git a/src/whisper_cpp/examples/server/httplib.h b/src/whisper_cpp/examples/server/httplib.h
new file mode 100644
index 00000000..feeb1308
--- /dev/null
+++ b/src/whisper_cpp/examples/server/httplib.h
@@ -0,0 +1,9262 @@
+//
+//  httplib.h
+//
+//  Copyright (c) 2023 Yuji Hirose. All rights reserved.
+//  MIT License
+//
+
+#ifndef CPPHTTPLIB_HTTPLIB_H
+#define CPPHTTPLIB_HTTPLIB_H
+
+#define CPPHTTPLIB_VERSION "0.14.1"
+
+/*
+ * Configuration
+ */
+
+#ifndef CPPHTTPLIB_KEEPALIVE_TIMEOUT_SECOND
+#define CPPHTTPLIB_KEEPALIVE_TIMEOUT_SECOND 5
+#endif
+
+#ifndef CPPHTTPLIB_KEEPALIVE_MAX_COUNT
+#define CPPHTTPLIB_KEEPALIVE_MAX_COUNT 5
+#endif
+
+#ifndef CPPHTTPLIB_CONNECTION_TIMEOUT_SECOND
+#define CPPHTTPLIB_CONNECTION_TIMEOUT_SECOND 300
+#endif
+
+#ifndef CPPHTTPLIB_CONNECTION_TIMEOUT_USECOND
+#define CPPHTTPLIB_CONNECTION_TIMEOUT_USECOND 0
+#endif
+
+#ifndef CPPHTTPLIB_READ_TIMEOUT_SECOND
+#define CPPHTTPLIB_READ_TIMEOUT_SECOND 5
+#endif
+
+#ifndef CPPHTTPLIB_READ_TIMEOUT_USECOND
+#define CPPHTTPLIB_READ_TIMEOUT_USECOND 0
+#endif
+
+#ifndef CPPHTTPLIB_WRITE_TIMEOUT_SECOND
+#define CPPHTTPLIB_WRITE_TIMEOUT_SECOND 5
+#endif
+
+#ifndef CPPHTTPLIB_WRITE_TIMEOUT_USECOND
+#define CPPHTTPLIB_WRITE_TIMEOUT_USECOND 0
+#endif
+
+#ifndef CPPHTTPLIB_IDLE_INTERVAL_SECOND
+#define CPPHTTPLIB_IDLE_INTERVAL_SECOND 0
+#endif
+
+#ifndef CPPHTTPLIB_IDLE_INTERVAL_USECOND
+#ifdef _WIN32
+#define CPPHTTPLIB_IDLE_INTERVAL_USECOND 10000
+#else
+#define CPPHTTPLIB_IDLE_INTERVAL_USECOND 0
+#endif
+#endif
+
+#ifndef CPPHTTPLIB_REQUEST_URI_MAX_LENGTH
+#define CPPHTTPLIB_REQUEST_URI_MAX_LENGTH 8192
+#endif
+
+#ifndef CPPHTTPLIB_HEADER_MAX_LENGTH
+#define CPPHTTPLIB_HEADER_MAX_LENGTH 8192
+#endif
+
+#ifndef CPPHTTPLIB_REDIRECT_MAX_COUNT
+#define CPPHTTPLIB_REDIRECT_MAX_COUNT 20
+#endif
+
+#ifndef CPPHTTPLIB_MULTIPART_FORM_DATA_FILE_MAX_COUNT
+#define CPPHTTPLIB_MULTIPART_FORM_DATA_FILE_MAX_COUNT 1024
+#endif
+
+#ifndef CPPHTTPLIB_PAYLOAD_MAX_LENGTH
+#define CPPHTTPLIB_PAYLOAD_MAX_LENGTH ((std::numeric_limits<size_t>::max)())
+#endif
+
+#ifndef CPPHTTPLIB_FORM_URL_ENCODED_PAYLOAD_MAX_LENGTH
+#define CPPHTTPLIB_FORM_URL_ENCODED_PAYLOAD_MAX_LENGTH 8192
+#endif
+
+#ifndef CPPHTTPLIB_TCP_NODELAY
+#define CPPHTTPLIB_TCP_NODELAY false
+#endif
+
+#ifndef CPPHTTPLIB_RECV_BUFSIZ
+#define CPPHTTPLIB_RECV_BUFSIZ size_t(4096u)
+#endif
+
+#ifndef CPPHTTPLIB_COMPRESSION_BUFSIZ
+#define CPPHTTPLIB_COMPRESSION_BUFSIZ size_t(16384u)
+#endif
+
+#ifndef CPPHTTPLIB_THREAD_POOL_COUNT
+#define CPPHTTPLIB_THREAD_POOL_COUNT                                           \
+  ((std::max)(8u, std::thread::hardware_concurrency() > 0                      \
+                      ? std::thread::hardware_concurrency() - 1                \
+                      : 0))
+#endif
+
+#ifndef CPPHTTPLIB_RECV_FLAGS
+#define CPPHTTPLIB_RECV_FLAGS 0
+#endif
+
+#ifndef CPPHTTPLIB_SEND_FLAGS
+#define CPPHTTPLIB_SEND_FLAGS 0
+#endif
+
+#ifndef CPPHTTPLIB_LISTEN_BACKLOG
+#define CPPHTTPLIB_LISTEN_BACKLOG 5
+#endif
+
+/*
+ * Headers
+ */
+
+#ifdef _WIN32
+#ifndef _CRT_SECURE_NO_WARNINGS
+#define _CRT_SECURE_NO_WARNINGS
+#endif //_CRT_SECURE_NO_WARNINGS
+
+#ifndef _CRT_NONSTDC_NO_DEPRECATE
+#define _CRT_NONSTDC_NO_DEPRECATE
+#endif //_CRT_NONSTDC_NO_DEPRECATE
+
+#if defined(_MSC_VER)
+#if _MSC_VER < 1900
+#error Sorry, Visual Studio versions prior to 2015 are not supported
+#endif
+
+#pragma comment(lib, "ws2_32.lib")
+
+#ifdef _WIN64
+using ssize_t = __int64;
+#else
+using ssize_t = long;
+#endif
+#endif // _MSC_VER
+
+#ifndef S_ISREG
+#define S_ISREG(m) (((m)&S_IFREG) == S_IFREG)
+#endif // S_ISREG
+
+#ifndef S_ISDIR
+#define S_ISDIR(m) (((m)&S_IFDIR) == S_IFDIR)
+#endif // S_ISDIR
+
+#ifndef NOMINMAX
+#define NOMINMAX
+#endif // NOMINMAX
+
+#include <io.h>
+#include <winsock2.h>
+#include <ws2tcpip.h>
+
+#ifndef WSA_FLAG_NO_HANDLE_INHERIT
+#define WSA_FLAG_NO_HANDLE_INHERIT 0x80
+#endif
+
+#ifndef strcasecmp
+#define strcasecmp _stricmp
+#endif // strcasecmp
+
+using socket_t = SOCKET;
+#ifdef CPPHTTPLIB_USE_POLL
+#define poll(fds, nfds, timeout) WSAPoll(fds, nfds, timeout)
+#endif
+
+#else // not _WIN32
+
+#include <arpa/inet.h>
+#if !defined(_AIX) && !defined(__MVS__)
+#include <ifaddrs.h>
+#endif
+#ifdef __MVS__
+#include <strings.h>
+#ifndef NI_MAXHOST
+#define NI_MAXHOST 1025
+#endif
+#endif
+#include <net/if.h>
+#include <netdb.h>
+#include <netinet/in.h>
+#ifdef __linux__
+#include <resolv.h>
+#endif
+#include <netinet/tcp.h>
+#ifdef CPPHTTPLIB_USE_POLL
+#include <poll.h>
+#endif
+#include <csignal>
+#include <pthread.h>
+#include <sys/mman.h>
+#include <sys/select.h>
+#include <sys/socket.h>
+#include <sys/un.h>
+#include <unistd.h>
+
+using socket_t = int;
+#ifndef INVALID_SOCKET
+#define INVALID_SOCKET (-1)
+#endif
+#endif //_WIN32
+
+#include <algorithm>
+#include <array>
+#include <atomic>
+#include <cassert>
+#include <cctype>
+#include <climits>
+#include <condition_variable>
+#include <cstring>
+#include <errno.h>
+#include <fcntl.h>
+#include <fstream>
+#include <functional>
+#include <iomanip>
+#include <iostream>
+#include <list>
+#include <map>
+#include <memory>
+#include <mutex>
+#include <random>
+#include <regex>
+#include <set>
+#include <sstream>
+#include <string>
+#include <sys/stat.h>
+#include <thread>
+#include <unordered_map>
+#include <unordered_set>
+#include <utility>
+
+#ifdef CPPHTTPLIB_OPENSSL_SUPPORT
+#ifdef _WIN32
+#include <wincrypt.h>
+
+// these are defined in wincrypt.h and it breaks compilation if BoringSSL is
+// used
+#undef X509_NAME
+#undef X509_CERT_PAIR
+#undef X509_EXTENSIONS
+#undef PKCS7_SIGNER_INFO
+
+#ifdef _MSC_VER
+#pragma comment(lib, "crypt32.lib")
+#pragma comment(lib, "cryptui.lib")
+#endif
+#elif defined(CPPHTTPLIB_USE_CERTS_FROM_MACOSX_KEYCHAIN) && defined(__APPLE__)
+#include <TargetConditionals.h>
+#if TARGET_OS_OSX
+#include <CoreFoundation/CoreFoundation.h>
+#include <Security/Security.h>
+#endif // TARGET_OS_OSX
+#endif // _WIN32
+
+#include <openssl/err.h>
+#include <openssl/evp.h>
+#include <openssl/ssl.h>
+#include <openssl/x509v3.h>
+
+#if defined(_WIN32) && defined(OPENSSL_USE_APPLINK)
+#include <openssl/applink.c>
+#endif
+
+#include <iostream>
+#include <sstream>
+
+#if OPENSSL_VERSION_NUMBER < 0x1010100fL
+#error Sorry, OpenSSL versions prior to 1.1.1 are not supported
+#elif OPENSSL_VERSION_NUMBER < 0x30000000L
+#define SSL_get1_peer_certificate SSL_get_peer_certificate
+#endif
+
+#endif
+
+#ifdef CPPHTTPLIB_ZLIB_SUPPORT
+#include <zlib.h>
+#endif
+
+#ifdef CPPHTTPLIB_BROTLI_SUPPORT
+#include <brotli/decode.h>
+#include <brotli/encode.h>
+#endif
+
+/*
+ * Declaration
+ */
+namespace httplib {
+
+namespace detail {
+
+/*
+ * Backport std::make_unique from C++14.
+ *
+ * NOTE: This code came up with the following stackoverflow post:
+ * https://stackoverflow.com/questions/10149840/c-arrays-and-make-unique
+ *
+ */
+
+template <class T, class... Args>
+typename std::enable_if<!std::is_array<T>::value, std::unique_ptr<T>>::type
+make_unique(Args &&...args) {
+  return std::unique_ptr<T>(new T(std::forward<Args>(args)...));
+}
+
+template <class T>
+typename std::enable_if<std::is_array<T>::value, std::unique_ptr<T>>::type
+make_unique(std::size_t n) {
+  typedef typename std::remove_extent<T>::type RT;
+  return std::unique_ptr<T>(new RT[n]);
+}
+
+struct ci {
+  bool operator()(const std::string &s1, const std::string &s2) const {
+    return std::lexicographical_compare(s1.begin(), s1.end(), s2.begin(),
+                                        s2.end(),
+                                        [](unsigned char c1, unsigned char c2) {
+                                          return ::tolower(c1) < ::tolower(c2);
+                                        });
+  }
+};
+
+// This is based on
+// "http://www.open-std.org/jtc1/sc22/wg21/docs/papers/2014/n4189".
+
+struct scope_exit {
+  explicit scope_exit(std::function<void(void)> &&f)
+      : exit_function(std::move(f)), execute_on_destruction{true} {}
+
+  scope_exit(scope_exit &&rhs)
+      : exit_function(std::move(rhs.exit_function)),
+        execute_on_destruction{rhs.execute_on_destruction} {
+    rhs.release();
+  }
+
+  ~scope_exit() {
+    if (execute_on_destruction) { this->exit_function(); }
+  }
+
+  void release() { this->execute_on_destruction = false; }
+
+private:
+  scope_exit(const scope_exit &) = delete;
+  void operator=(const scope_exit &) = delete;
+  scope_exit &operator=(scope_exit &&) = delete;
+
+  std::function<void(void)> exit_function;
+  bool execute_on_destruction;
+};
+
+} // namespace detail
+
+using Headers = std::multimap<std::string, std::string, detail::ci>;
+
+using Params = std::multimap<std::string, std::string>;
+using Match = std::smatch;
+
+using Progress = std::function<bool(uint64_t current, uint64_t total)>;
+
+struct Response;
+using ResponseHandler = std::function<bool(const Response &response)>;
+
+struct MultipartFormData {
+  std::string name;
+  std::string content;
+  std::string filename;
+  std::string content_type;
+};
+using MultipartFormDataItems = std::vector<MultipartFormData>;
+using MultipartFormDataMap = std::multimap<std::string, MultipartFormData>;
+
+class DataSink {
+public:
+  DataSink() : os(&sb_), sb_(*this) {}
+
+  DataSink(const DataSink &) = delete;
+  DataSink &operator=(const DataSink &) = delete;
+  DataSink(DataSink &&) = delete;
+  DataSink &operator=(DataSink &&) = delete;
+
+  std::function<bool(const char *data, size_t data_len)> write;
+  std::function<void()> done;
+  std::function<void(const Headers &trailer)> done_with_trailer;
+  std::ostream os;
+
+private:
+  class data_sink_streambuf : public std::streambuf {
+  public:
+    explicit data_sink_streambuf(DataSink &sink) : sink_(sink) {}
+
+  protected:
+    std::streamsize xsputn(const char *s, std::streamsize n) {
+      sink_.write(s, static_cast<size_t>(n));
+      return n;
+    }
+
+  private:
+    DataSink &sink_;
+  };
+
+  data_sink_streambuf sb_;
+};
+
+using ContentProvider =
+    std::function<bool(size_t offset, size_t length, DataSink &sink)>;
+
+using ContentProviderWithoutLength =
+    std::function<bool(size_t offset, DataSink &sink)>;
+
+using ContentProviderResourceReleaser = std::function<void(bool success)>;
+
+struct MultipartFormDataProvider {
+  std::string name;
+  ContentProviderWithoutLength provider;
+  std::string filename;
+  std::string content_type;
+};
+using MultipartFormDataProviderItems = std::vector<MultipartFormDataProvider>;
+
+using ContentReceiverWithProgress =
+    std::function<bool(const char *data, size_t data_length, uint64_t offset,
+                       uint64_t total_length)>;
+
+using ContentReceiver =
+    std::function<bool(const char *data, size_t data_length)>;
+
+using MultipartContentHeader =
+    std::function<bool(const MultipartFormData &file)>;
+
+class ContentReader {
+public:
+  using Reader = std::function<bool(ContentReceiver receiver)>;
+  using MultipartReader = std::function<bool(MultipartContentHeader header,
+                                             ContentReceiver receiver)>;
+
+  ContentReader(Reader reader, MultipartReader multipart_reader)
+      : reader_(std::move(reader)),
+        multipart_reader_(std::move(multipart_reader)) {}
+
+  bool operator()(MultipartContentHeader header,
+                  ContentReceiver receiver) const {
+    return multipart_reader_(std::move(header), std::move(receiver));
+  }
+
+  bool operator()(ContentReceiver receiver) const {
+    return reader_(std::move(receiver));
+  }
+
+  Reader reader_;
+  MultipartReader multipart_reader_;
+};
+
+using Range = std::pair<ssize_t, ssize_t>;
+using Ranges = std::vector<Range>;
+
+struct Request {
+  std::string method;
+  std::string path;
+  Headers headers;
+  std::string body;
+
+  std::string remote_addr;
+  int remote_port = -1;
+  std::string local_addr;
+  int local_port = -1;
+
+  // for server
+  std::string version;
+  std::string target;
+  Params params;
+  MultipartFormDataMap files;
+  Ranges ranges;
+  Match matches;
+  std::unordered_map<std::string, std::string> path_params;
+
+  // for client
+  ResponseHandler response_handler;
+  ContentReceiverWithProgress content_receiver;
+  Progress progress;
+#ifdef CPPHTTPLIB_OPENSSL_SUPPORT
+  const SSL *ssl = nullptr;
+#endif
+
+  bool has_header(const std::string &key) const;
+  std::string get_header_value(const std::string &key, size_t id = 0) const;
+  uint64_t get_header_value_u64(const std::string &key, size_t id = 0) const;
+  size_t get_header_value_count(const std::string &key) const;
+  void set_header(const std::string &key, const std::string &val);
+
+  bool has_param(const std::string &key) const;
+  std::string get_param_value(const std::string &key, size_t id = 0) const;
+  size_t get_param_value_count(const std::string &key) const;
+
+  bool is_multipart_form_data() const;
+
+  bool has_file(const std::string &key) const;
+  MultipartFormData get_file_value(const std::string &key) const;
+  std::vector<MultipartFormData> get_file_values(const std::string &key) const;
+
+  // private members...
+  size_t redirect_count_ = CPPHTTPLIB_REDIRECT_MAX_COUNT;
+  size_t content_length_ = 0;
+  ContentProvider content_provider_;
+  bool is_chunked_content_provider_ = false;
+  size_t authorization_count_ = 0;
+};
+
+struct Response {
+  std::string version;
+  int status = -1;
+  std::string reason;
+  Headers headers;
+  std::string body;
+  std::string location; // Redirect location
+
+  bool has_header(const std::string &key) const;
+  std::string get_header_value(const std::string &key, size_t id = 0) const;
+  uint64_t get_header_value_u64(const std::string &key, size_t id = 0) const;
+  size_t get_header_value_count(const std::string &key) const;
+  void set_header(const std::string &key, const std::string &val);
+
+  void set_redirect(const std::string &url, int status = 302);
+  void set_content(const char *s, size_t n, const std::string &content_type);
+  void set_content(const std::string &s, const std::string &content_type);
+
+  void set_content_provider(
+      size_t length, const std::string &content_type, ContentProvider provider,
+      ContentProviderResourceReleaser resource_releaser = nullptr);
+
+  void set_content_provider(
+      const std::string &content_type, ContentProviderWithoutLength provider,
+      ContentProviderResourceReleaser resource_releaser = nullptr);
+
+  void set_chunked_content_provider(
+      const std::string &content_type, ContentProviderWithoutLength provider,
+      ContentProviderResourceReleaser resource_releaser = nullptr);
+
+  Response() = default;
+  Response(const Response &) = default;
+  Response &operator=(const Response &) = default;
+  Response(Response &&) = default;
+  Response &operator=(Response &&) = default;
+  ~Response() {
+    if (content_provider_resource_releaser_) {
+      content_provider_resource_releaser_(content_provider_success_);
+    }
+  }
+
+  // private members...
+  size_t content_length_ = 0;
+  ContentProvider content_provider_;
+  ContentProviderResourceReleaser content_provider_resource_releaser_;
+  bool is_chunked_content_provider_ = false;
+  bool content_provider_success_ = false;
+};
+
+class Stream {
+public:
+  virtual ~Stream() = default;
+
+  virtual bool is_readable() const = 0;
+  virtual bool is_writable() const = 0;
+
+  virtual ssize_t read(char *ptr, size_t size) = 0;
+  virtual ssize_t write(const char *ptr, size_t size) = 0;
+  virtual void get_remote_ip_and_port(std::string &ip, int &port) const = 0;
+  virtual void get_local_ip_and_port(std::string &ip, int &port) const = 0;
+  virtual socket_t socket() const = 0;
+
+  template <typename... Args>
+  ssize_t write_format(const char *fmt, const Args &...args);
+  ssize_t write(const char *ptr);
+  ssize_t write(const std::string &s);
+};
+
+class TaskQueue {
+public:
+  TaskQueue() = default;
+  virtual ~TaskQueue() = default;
+
+  virtual void enqueue(std::function<void()> fn) = 0;
+  virtual void shutdown() = 0;
+
+  virtual void on_idle() {}
+};
+
+class ThreadPool : public TaskQueue {
+public:
+  explicit ThreadPool(size_t n) : shutdown_(false) {
+    while (n) {
+      threads_.emplace_back(worker(*this));
+      n--;
+    }
+  }
+
+  ThreadPool(const ThreadPool &) = delete;
+  ~ThreadPool() override = default;
+
+  void enqueue(std::function<void()> fn) override {
+    {
+      std::unique_lock<std::mutex> lock(mutex_);
+      jobs_.push_back(std::move(fn));
+    }
+
+    cond_.notify_one();
+  }
+
+  void shutdown() override {
+    // Stop all worker threads...
+    {
+      std::unique_lock<std::mutex> lock(mutex_);
+      shutdown_ = true;
+    }
+
+    cond_.notify_all();
+
+    // Join...
+    for (auto &t : threads_) {
+      t.join();
+    }
+  }
+
+private:
+  struct worker {
+    explicit worker(ThreadPool &pool) : pool_(pool) {}
+
+    void operator()() {
+      for (;;) {
+        std::function<void()> fn;
+        {
+          std::unique_lock<std::mutex> lock(pool_.mutex_);
+
+          pool_.cond_.wait(
+              lock, [&] { return !pool_.jobs_.empty() || pool_.shutdown_; });
+
+          if (pool_.shutdown_ && pool_.jobs_.empty()) { break; }
+
+          fn = std::move(pool_.jobs_.front());
+          pool_.jobs_.pop_front();
+        }
+
+        assert(true == static_cast<bool>(fn));
+        fn();
+      }
+    }
+
+    ThreadPool &pool_;
+  };
+  friend struct worker;
+
+  std::vector<std::thread> threads_;
+  std::list<std::function<void()>> jobs_;
+
+  bool shutdown_;
+
+  std::condition_variable cond_;
+  std::mutex mutex_;
+};
+
+using Logger = std::function<void(const Request &, const Response &)>;
+
+using SocketOptions = std::function<void(socket_t sock)>;
+
+void default_socket_options(socket_t sock);
+
+const char *status_message(int status);
+
+namespace detail {
+
+class MatcherBase {
+public:
+  virtual ~MatcherBase() = default;
+
+  // Match request path and populate its matches and
+  virtual bool match(Request &request) const = 0;
+};
+
+/**
+ * Captures parameters in request path and stores them in Request::path_params
+ *
+ * Capture name is a substring of a pattern from : to /.
+ * The rest of the pattern is matched agains the request path directly
+ * Parameters are captured starting from the next character after
+ * the end of the last matched static pattern fragment until the next /.
+ *
+ * Example pattern:
+ * "/path/fragments/:capture/more/fragments/:second_capture"
+ * Static fragments:
+ * "/path/fragments/", "more/fragments/"
+ *
+ * Given the following request path:
+ * "/path/fragments/:1/more/fragments/:2"
+ * the resulting capture will be
+ * {{"capture", "1"}, {"second_capture", "2"}}
+ */
+class PathParamsMatcher : public MatcherBase {
+public:
+  PathParamsMatcher(const std::string &pattern);
+
+  bool match(Request &request) const override;
+
+private:
+  static constexpr char marker = ':';
+  // Treat segment separators as the end of path parameter capture
+  // Does not need to handle query parameters as they are parsed before path
+  // matching
+  static constexpr char separator = '/';
+
+  // Contains static path fragments to match against, excluding the '/' after
+  // path params
+  // Fragments are separated by path params
+  std::vector<std::string> static_fragments_;
+  // Stores the names of the path parameters to be used as keys in the
+  // Request::path_params map
+  std::vector<std::string> param_names_;
+};
+
+/**
+ * Performs std::regex_match on request path
+ * and stores the result in Request::matches
+ *
+ * Note that regex match is performed directly on the whole request.
+ * This means that wildcard patterns may match multiple path segments with /:
+ * "/begin/(.*)/end" will match both "/begin/middle/end" and "/begin/1/2/end".
+ */
+class RegexMatcher : public MatcherBase {
+public:
+  RegexMatcher(const std::string &pattern) : regex_(pattern) {}
+
+  bool match(Request &request) const override;
+
+private:
+  std::regex regex_;
+};
+
+ssize_t write_headers(Stream &strm, const Headers &headers);
+
+} // namespace detail
+
+class Server {
+public:
+  using Handler = std::function<void(const Request &, Response &)>;
+
+  using ExceptionHandler =
+      std::function<void(const Request &, Response &, std::exception_ptr ep)>;
+
+  enum class HandlerResponse {
+    Handled,
+    Unhandled,
+  };
+  using HandlerWithResponse =
+      std::function<HandlerResponse(const Request &, Response &)>;
+
+  using HandlerWithContentReader = std::function<void(
+      const Request &, Response &, const ContentReader &content_reader)>;
+
+  using Expect100ContinueHandler =
+      std::function<int(const Request &, Response &)>;
+
+  Server();
+
+  virtual ~Server();
+
+  virtual bool is_valid() const;
+
+  Server &Get(const std::string &pattern, Handler handler);
+  Server &Post(const std::string &pattern, Handler handler);
+  Server &Post(const std::string &pattern, HandlerWithContentReader handler);
+  Server &Put(const std::string &pattern, Handler handler);
+  Server &Put(const std::string &pattern, HandlerWithContentReader handler);
+  Server &Patch(const std::string &pattern, Handler handler);
+  Server &Patch(const std::string &pattern, HandlerWithContentReader handler);
+  Server &Delete(const std::string &pattern, Handler handler);
+  Server &Delete(const std::string &pattern, HandlerWithContentReader handler);
+  Server &Options(const std::string &pattern, Handler handler);
+
+  bool set_base_dir(const std::string &dir,
+                    const std::string &mount_point = std::string());
+  bool set_mount_point(const std::string &mount_point, const std::string &dir,
+                       Headers headers = Headers());
+  bool remove_mount_point(const std::string &mount_point);
+  Server &set_file_extension_and_mimetype_mapping(const std::string &ext,
+                                                  const std::string &mime);
+  Server &set_default_file_mimetype(const std::string &mime);
+  Server &set_file_request_handler(Handler handler);
+
+  Server &set_error_handler(HandlerWithResponse handler);
+  Server &set_error_handler(Handler handler);
+  Server &set_exception_handler(ExceptionHandler handler);
+  Server &set_pre_routing_handler(HandlerWithResponse handler);
+  Server &set_post_routing_handler(Handler handler);
+
+  Server &set_expect_100_continue_handler(Expect100ContinueHandler handler);
+  Server &set_logger(Logger logger);
+
+  Server &set_address_family(int family);
+  Server &set_tcp_nodelay(bool on);
+  Server &set_socket_options(SocketOptions socket_options);
+
+  Server &set_default_headers(Headers headers);
+  Server &
+  set_header_writer(std::function<ssize_t(Stream &, Headers &)> const &writer);
+
+  Server &set_keep_alive_max_count(size_t count);
+  Server &set_keep_alive_timeout(time_t sec);
+
+  Server &set_read_timeout(time_t sec, time_t usec = 0);
+  template <class Rep, class Period>
+  Server &set_read_timeout(const std::chrono::duration<Rep, Period> &duration);
+
+  Server &set_write_timeout(time_t sec, time_t usec = 0);
+  template <class Rep, class Period>
+  Server &set_write_timeout(const std::chrono::duration<Rep, Period> &duration);
+
+  Server &set_idle_interval(time_t sec, time_t usec = 0);
+  template <class Rep, class Period>
+  Server &set_idle_interval(const std::chrono::duration<Rep, Period> &duration);
+
+  Server &set_payload_max_length(size_t length);
+
+  bool bind_to_port(const std::string &host, int port, int socket_flags = 0);
+  int bind_to_any_port(const std::string &host, int socket_flags = 0);
+  bool listen_after_bind();
+
+  bool listen(const std::string &host, int port, int socket_flags = 0);
+
+  bool is_running() const;
+  void wait_until_ready() const;
+  void stop();
+
+  std::function<TaskQueue *(void)> new_task_queue;
+
+protected:
+  bool process_request(Stream &strm, bool close_connection,
+                       bool &connection_closed,
+                       const std::function<void(Request &)> &setup_request);
+
+  std::atomic<socket_t> svr_sock_{INVALID_SOCKET};
+  size_t keep_alive_max_count_ = CPPHTTPLIB_KEEPALIVE_MAX_COUNT;
+  time_t keep_alive_timeout_sec_ = CPPHTTPLIB_KEEPALIVE_TIMEOUT_SECOND;
+  time_t read_timeout_sec_ = CPPHTTPLIB_READ_TIMEOUT_SECOND;
+  time_t read_timeout_usec_ = CPPHTTPLIB_READ_TIMEOUT_USECOND;
+  time_t write_timeout_sec_ = CPPHTTPLIB_WRITE_TIMEOUT_SECOND;
+  time_t write_timeout_usec_ = CPPHTTPLIB_WRITE_TIMEOUT_USECOND;
+  time_t idle_interval_sec_ = CPPHTTPLIB_IDLE_INTERVAL_SECOND;
+  time_t idle_interval_usec_ = CPPHTTPLIB_IDLE_INTERVAL_USECOND;
+  size_t payload_max_length_ = CPPHTTPLIB_PAYLOAD_MAX_LENGTH;
+
+private:
+  using Handlers =
+      std::vector<std::pair<std::unique_ptr<detail::MatcherBase>, Handler>>;
+  using HandlersForContentReader =
+      std::vector<std::pair<std::unique_ptr<detail::MatcherBase>,
+                            HandlerWithContentReader>>;
+
+  static std::unique_ptr<detail::MatcherBase>
+  make_matcher(const std::string &pattern);
+
+  socket_t create_server_socket(const std::string &host, int port,
+                                int socket_flags,
+                                SocketOptions socket_options) const;
+  int bind_internal(const std::string &host, int port, int socket_flags);
+  bool listen_internal();
+
+  bool routing(Request &req, Response &res, Stream &strm);
+  bool handle_file_request(const Request &req, Response &res,
+                           bool head = false);
+  bool dispatch_request(Request &req, Response &res, const Handlers &handlers);
+  bool
+  dispatch_request_for_content_reader(Request &req, Response &res,
+                                      ContentReader content_reader,
+                                      const HandlersForContentReader &handlers);
+
+  bool parse_request_line(const char *s, Request &req);
+  void apply_ranges(const Request &req, Response &res,
+                    std::string &content_type, std::string &boundary);
+  bool write_response(Stream &strm, bool close_connection, const Request &req,
+                      Response &res);
+  bool write_response_with_content(Stream &strm, bool close_connection,
+                                   const Request &req, Response &res);
+  bool write_response_core(Stream &strm, bool close_connection,
+                           const Request &req, Response &res,
+                           bool need_apply_ranges);
+  bool write_content_with_provider(Stream &strm, const Request &req,
+                                   Response &res, const std::string &boundary,
+                                   const std::string &content_type);
+  bool read_content(Stream &strm, Request &req, Response &res);
+  bool
+  read_content_with_content_receiver(Stream &strm, Request &req, Response &res,
+                                     ContentReceiver receiver,
+                                     MultipartContentHeader multipart_header,
+                                     ContentReceiver multipart_receiver);
+  bool read_content_core(Stream &strm, Request &req, Response &res,
+                         ContentReceiver receiver,
+                         MultipartContentHeader multipart_header,
+                         ContentReceiver multipart_receiver);
+
+  virtual bool process_and_close_socket(socket_t sock);
+
+  std::atomic<bool> is_running_{false};
+  std::atomic<bool> done_{false};
+
+  struct MountPointEntry {
+    std::string mount_point;
+    std::string base_dir;
+    Headers headers;
+  };
+  std::vector<MountPointEntry> base_dirs_;
+  std::map<std::string, std::string> file_extension_and_mimetype_map_;
+  std::string default_file_mimetype_ = "application/octet-stream";
+  Handler file_request_handler_;
+
+  Handlers get_handlers_;
+  Handlers post_handlers_;
+  HandlersForContentReader post_handlers_for_content_reader_;
+  Handlers put_handlers_;
+  HandlersForContentReader put_handlers_for_content_reader_;
+  Handlers patch_handlers_;
+  HandlersForContentReader patch_handlers_for_content_reader_;
+  Handlers delete_handlers_;
+  HandlersForContentReader delete_handlers_for_content_reader_;
+  Handlers options_handlers_;
+
+  HandlerWithResponse error_handler_;
+  ExceptionHandler exception_handler_;
+  HandlerWithResponse pre_routing_handler_;
+  Handler post_routing_handler_;
+  Expect100ContinueHandler expect_100_continue_handler_;
+
+  Logger logger_;
+
+  int address_family_ = AF_UNSPEC;
+  bool tcp_nodelay_ = CPPHTTPLIB_TCP_NODELAY;
+  SocketOptions socket_options_ = default_socket_options;
+
+  Headers default_headers_;
+  std::function<ssize_t(Stream &, Headers &)> header_writer_ =
+      detail::write_headers;
+};
+
+enum class Error {
+  Success = 0,
+  Unknown,
+  Connection,
+  BindIPAddress,
+  Read,
+  Write,
+  ExceedRedirectCount,
+  Canceled,
+  SSLConnection,
+  SSLLoadingCerts,
+  SSLServerVerification,
+  UnsupportedMultipartBoundaryChars,
+  Compression,
+  ConnectionTimeout,
+  ProxyConnection,
+
+  // For internal use only
+  SSLPeerCouldBeClosed_,
+};
+
+std::string to_string(const Error error);
+
+std::ostream &operator<<(std::ostream &os, const Error &obj);
+
+class Result {
+public:
+  Result() = default;
+  Result(std::unique_ptr<Response> &&res, Error err,
+         Headers &&request_headers = Headers{})
+      : res_(std::move(res)), err_(err),
+        request_headers_(std::move(request_headers)) {}
+  // Response
+  operator bool() const { return res_ != nullptr; }
+  bool operator==(std::nullptr_t) const { return res_ == nullptr; }
+  bool operator!=(std::nullptr_t) const { return res_ != nullptr; }
+  const Response &value() const { return *res_; }
+  Response &value() { return *res_; }
+  const Response &operator*() const { return *res_; }
+  Response &operator*() { return *res_; }
+  const Response *operator->() const { return res_.get(); }
+  Response *operator->() { return res_.get(); }
+
+  // Error
+  Error error() const { return err_; }
+
+  // Request Headers
+  bool has_request_header(const std::string &key) const;
+  std::string get_request_header_value(const std::string &key,
+                                       size_t id = 0) const;
+  uint64_t get_request_header_value_u64(const std::string &key,
+                                        size_t id = 0) const;
+  size_t get_request_header_value_count(const std::string &key) const;
+
+private:
+  std::unique_ptr<Response> res_;
+  Error err_ = Error::Unknown;
+  Headers request_headers_;
+};
+
+class ClientImpl {
+public:
+  explicit ClientImpl(const std::string &host);
+
+  explicit ClientImpl(const std::string &host, int port);
+
+  explicit ClientImpl(const std::string &host, int port,
+                      const std::string &client_cert_path,
+                      const std::string &client_key_path);
+
+  virtual ~ClientImpl();
+
+  virtual bool is_valid() const;
+
+  Result Get(const std::string &path);
+  Result Get(const std::string &path, const Headers &headers);
+  Result Get(const std::string &path, Progress progress);
+  Result Get(const std::string &path, const Headers &headers,
+             Progress progress);
+  Result Get(const std::string &path, ContentReceiver content_receiver);
+  Result Get(const std::string &path, const Headers &headers,
+             ContentReceiver content_receiver);
+  Result Get(const std::string &path, ContentReceiver content_receiver,
+             Progress progress);
+  Result Get(const std::string &path, const Headers &headers,
+             ContentReceiver content_receiver, Progress progress);
+  Result Get(const std::string &path, ResponseHandler response_handler,
+             ContentReceiver content_receiver);
+  Result Get(const std::string &path, const Headers &headers,
+             ResponseHandler response_handler,
+             ContentReceiver content_receiver);
+  Result Get(const std::string &path, ResponseHandler response_handler,
+             ContentReceiver content_receiver, Progress progress);
+  Result Get(const std::string &path, const Headers &headers,
+             ResponseHandler response_handler, ContentReceiver content_receiver,
+             Progress progress);
+
+  Result Get(const std::string &path, const Params &params,
+             const Headers &headers, Progress progress = nullptr);
+  Result Get(const std::string &path, const Params &params,
+             const Headers &headers, ContentReceiver content_receiver,
+             Progress progress = nullptr);
+  Result Get(const std::string &path, const Params &params,
+             const Headers &headers, ResponseHandler response_handler,
+             ContentReceiver content_receiver, Progress progress = nullptr);
+
+  Result Head(const std::string &path);
+  Result Head(const std::string &path, const Headers &headers);
+
+  Result Post(const std::string &path);
+  Result Post(const std::string &path, const Headers &headers);
+  Result Post(const std::string &path, const char *body, size_t content_length,
+              const std::string &content_type);
+  Result Post(const std::string &path, const Headers &headers, const char *body,
+              size_t content_length, const std::string &content_type);
+  Result Post(const std::string &path, const std::string &body,
+              const std::string &content_type);
+  Result Post(const std::string &path, const Headers &headers,
+              const std::string &body, const std::string &content_type);
+  Result Post(const std::string &path, size_t content_length,
+              ContentProvider content_provider,
+              const std::string &content_type);
+  Result Post(const std::string &path,
+              ContentProviderWithoutLength content_provider,
+              const std::string &content_type);
+  Result Post(const std::string &path, const Headers &headers,
+              size_t content_length, ContentProvider content_provider,
+              const std::string &content_type);
+  Result Post(const std::string &path, const Headers &headers,
+              ContentProviderWithoutLength content_provider,
+              const std::string &content_type);
+  Result Post(const std::string &path, const Params &params);
+  Result Post(const std::string &path, const Headers &headers,
+              const Params &params);
+  Result Post(const std::string &path, const MultipartFormDataItems &items);
+  Result Post(const std::string &path, const Headers &headers,
+              const MultipartFormDataItems &items);
+  Result Post(const std::string &path, const Headers &headers,
+              const MultipartFormDataItems &items, const std::string &boundary);
+  Result Post(const std::string &path, const Headers &headers,
+              const MultipartFormDataItems &items,
+              const MultipartFormDataProviderItems &provider_items);
+
+  Result Put(const std::string &path);
+  Result Put(const std::string &path, const char *body, size_t content_length,
+             const std::string &content_type);
+  Result Put(const std::string &path, const Headers &headers, const char *body,
+             size_t content_length, const std::string &content_type);
+  Result Put(const std::string &path, const std::string &body,
+             const std::string &content_type);
+  Result Put(const std::string &path, const Headers &headers,
+             const std::string &body, const std::string &content_type);
+  Result Put(const std::string &path, size_t content_length,
+             ContentProvider content_provider, const std::string &content_type);
+  Result Put(const std::string &path,
+             ContentProviderWithoutLength content_provider,
+             const std::string &content_type);
+  Result Put(const std::string &path, const Headers &headers,
+             size_t content_length, ContentProvider content_provider,
+             const std::string &content_type);
+  Result Put(const std::string &path, const Headers &headers,
+             ContentProviderWithoutLength content_provider,
+             const std::string &content_type);
+  Result Put(const std::string &path, const Params &params);
+  Result Put(const std::string &path, const Headers &headers,
+             const Params &params);
+  Result Put(const std::string &path, const MultipartFormDataItems &items);
+  Result Put(const std::string &path, const Headers &headers,
+             const MultipartFormDataItems &items);
+  Result Put(const std::string &path, const Headers &headers,
+             const MultipartFormDataItems &items, const std::string &boundary);
+  Result Put(const std::string &path, const Headers &headers,
+             const MultipartFormDataItems &items,
+             const MultipartFormDataProviderItems &provider_items);
+
+  Result Patch(const std::string &path);
+  Result Patch(const std::string &path, const char *body, size_t content_length,
+               const std::string &content_type);
+  Result Patch(const std::string &path, const Headers &headers,
+               const char *body, size_t content_length,
+               const std::string &content_type);
+  Result Patch(const std::string &path, const std::string &body,
+               const std::string &content_type);
+  Result Patch(const std::string &path, const Headers &headers,
+               const std::string &body, const std::string &content_type);
+  Result Patch(const std::string &path, size_t content_length,
+               ContentProvider content_provider,
+               const std::string &content_type);
+  Result Patch(const std::string &path,
+               ContentProviderWithoutLength content_provider,
+               const std::string &content_type);
+  Result Patch(const std::string &path, const Headers &headers,
+               size_t content_length, ContentProvider content_provider,
+               const std::string &content_type);
+  Result Patch(const std::string &path, const Headers &headers,
+               ContentProviderWithoutLength content_provider,
+               const std::string &content_type);
+
+  Result Delete(const std::string &path);
+  Result Delete(const std::string &path, const Headers &headers);
+  Result Delete(const std::string &path, const char *body,
+                size_t content_length, const std::string &content_type);
+  Result Delete(const std::string &path, const Headers &headers,
+                const char *body, size_t content_length,
+                const std::string &content_type);
+  Result Delete(const std::string &path, const std::string &body,
+                const std::string &content_type);
+  Result Delete(const std::string &path, const Headers &headers,
+                const std::string &body, const std::string &content_type);
+
+  Result Options(const std::string &path);
+  Result Options(const std::string &path, const Headers &headers);
+
+  bool send(Request &req, Response &res, Error &error);
+  Result send(const Request &req);
+
+  void stop();
+
+  std::string host() const;
+  int port() const;
+
+  size_t is_socket_open() const;
+  socket_t socket() const;
+
+  void set_hostname_addr_map(std::map<std::string, std::string> addr_map);
+
+  void set_default_headers(Headers headers);
+
+  void
+  set_header_writer(std::function<ssize_t(Stream &, Headers &)> const &writer);
+
+  void set_address_family(int family);
+  void set_tcp_nodelay(bool on);
+  void set_socket_options(SocketOptions socket_options);
+
+  void set_connection_timeout(time_t sec, time_t usec = 0);
+  template <class Rep, class Period>
+  void
+  set_connection_timeout(const std::chrono::duration<Rep, Period> &duration);
+
+  void set_read_timeout(time_t sec, time_t usec = 0);
+  template <class Rep, class Period>
+  void set_read_timeout(const std::chrono::duration<Rep, Period> &duration);
+
+  void set_write_timeout(time_t sec, time_t usec = 0);
+  template <class Rep, class Period>
+  void set_write_timeout(const std::chrono::duration<Rep, Period> &duration);
+
+  void set_basic_auth(const std::string &username, const std::string &password);
+  void set_bearer_token_auth(const std::string &token);
+#ifdef CPPHTTPLIB_OPENSSL_SUPPORT
+  void set_digest_auth(const std::string &username,
+                       const std::string &password);
+#endif
+
+  void set_keep_alive(bool on);
+  void set_follow_location(bool on);
+
+  void set_url_encode(bool on);
+
+  void set_compress(bool on);
+
+  void set_decompress(bool on);
+
+  void set_interface(const std::string &intf);
+
+  void set_proxy(const std::string &host, int port);
+  void set_proxy_basic_auth(const std::string &username,
+                            const std::string &password);
+  void set_proxy_bearer_token_auth(const std::string &token);
+#ifdef CPPHTTPLIB_OPENSSL_SUPPORT
+  void set_proxy_digest_auth(const std::string &username,
+                             const std::string &password);
+#endif
+
+#ifdef CPPHTTPLIB_OPENSSL_SUPPORT
+  void set_ca_cert_path(const std::string &ca_cert_file_path,
+                        const std::string &ca_cert_dir_path = std::string());
+  void set_ca_cert_store(X509_STORE *ca_cert_store);
+  X509_STORE *create_ca_cert_store(const char *ca_cert, std::size_t size);
+#endif
+
+#ifdef CPPHTTPLIB_OPENSSL_SUPPORT
+  void enable_server_certificate_verification(bool enabled);
+#endif
+
+  void set_logger(Logger logger);
+
+protected:
+  struct Socket {
+    socket_t sock = INVALID_SOCKET;
+#ifdef CPPHTTPLIB_OPENSSL_SUPPORT
+    SSL *ssl = nullptr;
+#endif
+
+    bool is_open() const { return sock != INVALID_SOCKET; }
+  };
+
+  virtual bool create_and_connect_socket(Socket &socket, Error &error);
+
+  // All of:
+  //   shutdown_ssl
+  //   shutdown_socket
+  //   close_socket
+  // should ONLY be called when socket_mutex_ is locked.
+  // Also, shutdown_ssl and close_socket should also NOT be called concurrently
+  // with a DIFFERENT thread sending requests using that socket.
+  virtual void shutdown_ssl(Socket &socket, bool shutdown_gracefully);
+  void shutdown_socket(Socket &socket);
+  void close_socket(Socket &socket);
+
+  bool process_request(Stream &strm, Request &req, Response &res,
+                       bool close_connection, Error &error);
+
+  bool write_content_with_provider(Stream &strm, const Request &req,
+                                   Error &error);
+
+  void copy_settings(const ClientImpl &rhs);
+
+  // Socket endpoint information
+  const std::string host_;
+  const int port_;
+  const std::string host_and_port_;
+
+  // Current open socket
+  Socket socket_;
+  mutable std::mutex socket_mutex_;
+  std::recursive_mutex request_mutex_;
+
+  // These are all protected under socket_mutex
+  size_t socket_requests_in_flight_ = 0;
+  std::thread::id socket_requests_are_from_thread_ = std::thread::id();
+  bool socket_should_be_closed_when_request_is_done_ = false;
+
+  // Hostname-IP map
+  std::map<std::string, std::string> addr_map_;
+
+  // Default headers
+  Headers default_headers_;
+
+  // Header writer
+  std::function<ssize_t(Stream &, Headers &)> header_writer_ =
+      detail::write_headers;
+
+  // Settings
+  std::string client_cert_path_;
+  std::string client_key_path_;
+
+  time_t connection_timeout_sec_ = CPPHTTPLIB_CONNECTION_TIMEOUT_SECOND;
+  time_t connection_timeout_usec_ = CPPHTTPLIB_CONNECTION_TIMEOUT_USECOND;
+  time_t read_timeout_sec_ = CPPHTTPLIB_READ_TIMEOUT_SECOND;
+  time_t read_timeout_usec_ = CPPHTTPLIB_READ_TIMEOUT_USECOND;
+  time_t write_timeout_sec_ = CPPHTTPLIB_WRITE_TIMEOUT_SECOND;
+  time_t write_timeout_usec_ = CPPHTTPLIB_WRITE_TIMEOUT_USECOND;
+
+  std::string basic_auth_username_;
+  std::string basic_auth_password_;
+  std::string bearer_token_auth_token_;
+#ifdef CPPHTTPLIB_OPENSSL_SUPPORT
+  std::string digest_auth_username_;
+  std::string digest_auth_password_;
+#endif
+
+  bool keep_alive_ = false;
+  bool follow_location_ = false;
+
+  bool url_encode_ = true;
+
+  int address_family_ = AF_UNSPEC;
+  bool tcp_nodelay_ = CPPHTTPLIB_TCP_NODELAY;
+  SocketOptions socket_options_ = nullptr;
+
+  bool compress_ = false;
+  bool decompress_ = true;
+
+  std::string interface_;
+
+  std::string proxy_host_;
+  int proxy_port_ = -1;
+
+  std::string proxy_basic_auth_username_;
+  std::string proxy_basic_auth_password_;
+  std::string proxy_bearer_token_auth_token_;
+#ifdef CPPHTTPLIB_OPENSSL_SUPPORT
+  std::string proxy_digest_auth_username_;
+  std::string proxy_digest_auth_password_;
+#endif
+
+#ifdef CPPHTTPLIB_OPENSSL_SUPPORT
+  std::string ca_cert_file_path_;
+  std::string ca_cert_dir_path_;
+
+  X509_STORE *ca_cert_store_ = nullptr;
+#endif
+
+#ifdef CPPHTTPLIB_OPENSSL_SUPPORT
+  bool server_certificate_verification_ = true;
+#endif
+
+  Logger logger_;
+
+private:
+  bool send_(Request &req, Response &res, Error &error);
+  Result send_(Request &&req);
+
+  socket_t create_client_socket(Error &error) const;
+  bool read_response_line(Stream &strm, const Request &req, Response &res);
+  bool write_request(Stream &strm, Request &req, bool close_connection,
+                     Error &error);
+  bool redirect(Request &req, Response &res, Error &error);
+  bool handle_request(Stream &strm, Request &req, Response &res,
+                      bool close_connection, Error &error);
+  std::unique_ptr<Response> send_with_content_provider(
+      Request &req, const char *body, size_t content_length,
+      ContentProvider content_provider,
+      ContentProviderWithoutLength content_provider_without_length,
+      const std::string &content_type, Error &error);
+  Result send_with_content_provider(
+      const std::string &method, const std::string &path,
+      const Headers &headers, const char *body, size_t content_length,
+      ContentProvider content_provider,
+      ContentProviderWithoutLength content_provider_without_length,
+      const std::string &content_type);
+  ContentProviderWithoutLength get_multipart_content_provider(
+      const std::string &boundary, const MultipartFormDataItems &items,
+      const MultipartFormDataProviderItems &provider_items);
+
+  std::string adjust_host_string(const std::string &host) const;
+
+  virtual bool process_socket(const Socket &socket,
+                              std::function<bool(Stream &strm)> callback);
+  virtual bool is_ssl() const;
+};
+
+class Client {
+public:
+  // Universal interface
+  explicit Client(const std::string &scheme_host_port);
+
+  explicit Client(const std::string &scheme_host_port,
+                  const std::string &client_cert_path,
+                  const std::string &client_key_path);
+
+  // HTTP only interface
+  explicit Client(const std::string &host, int port);
+
+  explicit Client(const std::string &host, int port,
+                  const std::string &client_cert_path,
+                  const std::string &client_key_path);
+
+  Client(Client &&) = default;
+
+  ~Client();
+
+  bool is_valid() const;
+
+  Result Get(const std::string &path);
+  Result Get(const std::string &path, const Headers &headers);
+  Result Get(const std::string &path, Progress progress);
+  Result Get(const std::string &path, const Headers &headers,
+             Progress progress);
+  Result Get(const std::string &path, ContentReceiver content_receiver);
+  Result Get(const std::string &path, const Headers &headers,
+             ContentReceiver content_receiver);
+  Result Get(const std::string &path, ContentReceiver content_receiver,
+             Progress progress);
+  Result Get(const std::string &path, const Headers &headers,
+             ContentReceiver content_receiver, Progress progress);
+  Result Get(const std::string &path, ResponseHandler response_handler,
+             ContentReceiver content_receiver);
+  Result Get(const std::string &path, const Headers &headers,
+             ResponseHandler response_handler,
+             ContentReceiver content_receiver);
+  Result Get(const std::string &path, const Headers &headers,
+             ResponseHandler response_handler, ContentReceiver content_receiver,
+             Progress progress);
+  Result Get(const std::string &path, ResponseHandler response_handler,
+             ContentReceiver content_receiver, Progress progress);
+
+  Result Get(const std::string &path, const Params &params,
+             const Headers &headers, Progress progress = nullptr);
+  Result Get(const std::string &path, const Params &params,
+             const Headers &headers, ContentReceiver content_receiver,
+             Progress progress = nullptr);
+  Result Get(const std::string &path, const Params &params,
+             const Headers &headers, ResponseHandler response_handler,
+             ContentReceiver content_receiver, Progress progress = nullptr);
+
+  Result Head(const std::string &path);
+  Result Head(const std::string &path, const Headers &headers);
+
+  Result Post(const std::string &path);
+  Result Post(const std::string &path, const Headers &headers);
+  Result Post(const std::string &path, const char *body, size_t content_length,
+              const std::string &content_type);
+  Result Post(const std::string &path, const Headers &headers, const char *body,
+              size_t content_length, const std::string &content_type);
+  Result Post(const std::string &path, const std::string &body,
+              const std::string &content_type);
+  Result Post(const std::string &path, const Headers &headers,
+              const std::string &body, const std::string &content_type);
+  Result Post(const std::string &path, size_t content_length,
+              ContentProvider content_provider,
+              const std::string &content_type);
+  Result Post(const std::string &path,
+              ContentProviderWithoutLength content_provider,
+              const std::string &content_type);
+  Result Post(const std::string &path, const Headers &headers,
+              size_t content_length, ContentProvider content_provider,
+              const std::string &content_type);
+  Result Post(const std::string &path, const Headers &headers,
+              ContentProviderWithoutLength content_provider,
+              const std::string &content_type);
+  Result Post(const std::string &path, const Params &params);
+  Result Post(const std::string &path, const Headers &headers,
+              const Params &params);
+  Result Post(const std::string &path, const MultipartFormDataItems &items);
+  Result Post(const std::string &path, const Headers &headers,
+              const MultipartFormDataItems &items);
+  Result Post(const std::string &path, const Headers &headers,
+              const MultipartFormDataItems &items, const std::string &boundary);
+  Result Post(const std::string &path, const Headers &headers,
+              const MultipartFormDataItems &items,
+              const MultipartFormDataProviderItems &provider_items);
+
+  Result Put(const std::string &path);
+  Result Put(const std::string &path, const char *body, size_t content_length,
+             const std::string &content_type);
+  Result Put(const std::string &path, const Headers &headers, const char *body,
+             size_t content_length, const std::string &content_type);
+  Result Put(const std::string &path, const std::string &body,
+             const std::string &content_type);
+  Result Put(const std::string &path, const Headers &headers,
+             const std::string &body, const std::string &content_type);
+  Result Put(const std::string &path, size_t content_length,
+             ContentProvider content_provider, const std::string &content_type);
+  Result Put(const std::string &path,
+             ContentProviderWithoutLength content_provider,
+             const std::string &content_type);
+  Result Put(const std::string &path, const Headers &headers,
+             size_t content_length, ContentProvider content_provider,
+             const std::string &content_type);
+  Result Put(const std::string &path, const Headers &headers,
+             ContentProviderWithoutLength content_provider,
+             const std::string &content_type);
+  Result Put(const std::string &path, const Params &params);
+  Result Put(const std::string &path, const Headers &headers,
+             const Params &params);
+  Result Put(const std::string &path, const MultipartFormDataItems &items);
+  Result Put(const std::string &path, const Headers &headers,
+             const MultipartFormDataItems &items);
+  Result Put(const std::string &path, const Headers &headers,
+             const MultipartFormDataItems &items, const std::string &boundary);
+  Result Put(const std::string &path, const Headers &headers,
+             const MultipartFormDataItems &items,
+             const MultipartFormDataProviderItems &provider_items);
+
+  Result Patch(const std::string &path);
+  Result Patch(const std::string &path, const char *body, size_t content_length,
+               const std::string &content_type);
+  Result Patch(const std::string &path, const Headers &headers,
+               const char *body, size_t content_length,
+               const std::string &content_type);
+  Result Patch(const std::string &path, const std::string &body,
+               const std::string &content_type);
+  Result Patch(const std::string &path, const Headers &headers,
+               const std::string &body, const std::string &content_type);
+  Result Patch(const std::string &path, size_t content_length,
+               ContentProvider content_provider,
+               const std::string &content_type);
+  Result Patch(const std::string &path,
+               ContentProviderWithoutLength content_provider,
+               const std::string &content_type);
+  Result Patch(const std::string &path, const Headers &headers,
+               size_t content_length, ContentProvider content_provider,
+               const std::string &content_type);
+  Result Patch(const std::string &path, const Headers &headers,
+               ContentProviderWithoutLength content_provider,
+               const std::string &content_type);
+
+  Result Delete(const std::string &path);
+  Result Delete(const std::string &path, const Headers &headers);
+  Result Delete(const std::string &path, const char *body,
+                size_t content_length, const std::string &content_type);
+  Result Delete(const std::string &path, const Headers &headers,
+                const char *body, size_t content_length,
+                const std::string &content_type);
+  Result Delete(const std::string &path, const std::string &body,
+                const std::string &content_type);
+  Result Delete(const std::string &path, const Headers &headers,
+                const std::string &body, const std::string &content_type);
+
+  Result Options(const std::string &path);
+  Result Options(const std::string &path, const Headers &headers);
+
+  bool send(Request &req, Response &res, Error &error);
+  Result send(const Request &req);
+
+  void stop();
+
+  std::string host() const;
+  int port() const;
+
+  size_t is_socket_open() const;
+  socket_t socket() const;
+
+  void set_hostname_addr_map(std::map<std::string, std::string> addr_map);
+
+  void set_default_headers(Headers headers);
+
+  void
+  set_header_writer(std::function<ssize_t(Stream &, Headers &)> const &writer);
+
+  void set_address_family(int family);
+  void set_tcp_nodelay(bool on);
+  void set_socket_options(SocketOptions socket_options);
+
+  void set_connection_timeout(time_t sec, time_t usec = 0);
+  template <class Rep, class Period>
+  void
+  set_connection_timeout(const std::chrono::duration<Rep, Period> &duration);
+
+  void set_read_timeout(time_t sec, time_t usec = 0);
+  template <class Rep, class Period>
+  void set_read_timeout(const std::chrono::duration<Rep, Period> &duration);
+
+  void set_write_timeout(time_t sec, time_t usec = 0);
+  template <class Rep, class Period>
+  void set_write_timeout(const std::chrono::duration<Rep, Period> &duration);
+
+  void set_basic_auth(const std::string &username, const std::string &password);
+  void set_bearer_token_auth(const std::string &token);
+#ifdef CPPHTTPLIB_OPENSSL_SUPPORT
+  void set_digest_auth(const std::string &username,
+                       const std::string &password);
+#endif
+
+  void set_keep_alive(bool on);
+  void set_follow_location(bool on);
+
+  void set_url_encode(bool on);
+
+  void set_compress(bool on);
+
+  void set_decompress(bool on);
+
+  void set_interface(const std::string &intf);
+
+  void set_proxy(const std::string &host, int port);
+  void set_proxy_basic_auth(const std::string &username,
+                            const std::string &password);
+  void set_proxy_bearer_token_auth(const std::string &token);
+#ifdef CPPHTTPLIB_OPENSSL_SUPPORT
+  void set_proxy_digest_auth(const std::string &username,
+                             const std::string &password);
+#endif
+
+#ifdef CPPHTTPLIB_OPENSSL_SUPPORT
+  void enable_server_certificate_verification(bool enabled);
+#endif
+
+  void set_logger(Logger logger);
+
+  // SSL
+#ifdef CPPHTTPLIB_OPENSSL_SUPPORT
+  void set_ca_cert_path(const std::string &ca_cert_file_path,
+                        const std::string &ca_cert_dir_path = std::string());
+
+  void set_ca_cert_store(X509_STORE *ca_cert_store);
+  void load_ca_cert_store(const char *ca_cert, std::size_t size);
+
+  long get_openssl_verify_result() const;
+
+  SSL_CTX *ssl_context() const;
+#endif
+
+private:
+  std::unique_ptr<ClientImpl> cli_;
+
+#ifdef CPPHTTPLIB_OPENSSL_SUPPORT
+  bool is_ssl_ = false;
+#endif
+};
+
+#ifdef CPPHTTPLIB_OPENSSL_SUPPORT
+class SSLServer : public Server {
+public:
+  SSLServer(const char *cert_path, const char *private_key_path,
+            const char *client_ca_cert_file_path = nullptr,
+            const char *client_ca_cert_dir_path = nullptr,
+            const char *private_key_password = nullptr);
+
+  SSLServer(X509 *cert, EVP_PKEY *private_key,
+            X509_STORE *client_ca_cert_store = nullptr);
+
+  SSLServer(
+      const std::function<bool(SSL_CTX &ssl_ctx)> &setup_ssl_ctx_callback);
+
+  ~SSLServer() override;
+
+  bool is_valid() const override;
+
+  SSL_CTX *ssl_context() const;
+
+private:
+  bool process_and_close_socket(socket_t sock) override;
+
+  SSL_CTX *ctx_;
+  std::mutex ctx_mutex_;
+};
+
+class SSLClient : public ClientImpl {
+public:
+  explicit SSLClient(const std::string &host);
+
+  explicit SSLClient(const std::string &host, int port);
+
+  explicit SSLClient(const std::string &host, int port,
+                     const std::string &client_cert_path,
+                     const std::string &client_key_path);
+
+  explicit SSLClient(const std::string &host, int port, X509 *client_cert,
+                     EVP_PKEY *client_key);
+
+  ~SSLClient() override;
+
+  bool is_valid() const override;
+
+  void set_ca_cert_store(X509_STORE *ca_cert_store);
+  void load_ca_cert_store(const char *ca_cert, std::size_t size);
+
+  long get_openssl_verify_result() const;
+
+  SSL_CTX *ssl_context() const;
+
+private:
+  bool create_and_connect_socket(Socket &socket, Error &error) override;
+  void shutdown_ssl(Socket &socket, bool shutdown_gracefully) override;
+  void shutdown_ssl_impl(Socket &socket, bool shutdown_socket);
+
+  bool process_socket(const Socket &socket,
+                      std::function<bool(Stream &strm)> callback) override;
+  bool is_ssl() const override;
+
+  bool connect_with_proxy(Socket &sock, Response &res, bool &success,
+                          Error &error);
+  bool initialize_ssl(Socket &socket, Error &error);
+
+  bool load_certs();
+
+  bool verify_host(X509 *server_cert) const;
+  bool verify_host_with_subject_alt_name(X509 *server_cert) const;
+  bool verify_host_with_common_name(X509 *server_cert) const;
+  bool check_host_name(const char *pattern, size_t pattern_len) const;
+
+  SSL_CTX *ctx_;
+  std::mutex ctx_mutex_;
+  std::once_flag initialize_cert_;
+
+  std::vector<std::string> host_components_;
+
+  long verify_result_ = 0;
+
+  friend class ClientImpl;
+};
+#endif
+
+/*
+ * Implementation of template methods.
+ */
+
+namespace detail {
+
+template <typename T, typename U>
+inline void duration_to_sec_and_usec(const T &duration, U callback) {
+  auto sec = std::chrono::duration_cast<std::chrono::seconds>(duration).count();
+  auto usec = std::chrono::duration_cast<std::chrono::microseconds>(
+                  duration - std::chrono::seconds(sec))
+                  .count();
+  callback(static_cast<time_t>(sec), static_cast<time_t>(usec));
+}
+
+inline uint64_t get_header_value_u64(const Headers &headers,
+                                     const std::string &key, size_t id,
+                                     uint64_t def) {
+  auto rng = headers.equal_range(key);
+  auto it = rng.first;
+  std::advance(it, static_cast<ssize_t>(id));
+  if (it != rng.second) {
+    return std::strtoull(it->second.data(), nullptr, 10);
+  }
+  return def;
+}
+
+} // namespace detail
+
+inline uint64_t Request::get_header_value_u64(const std::string &key,
+                                              size_t id) const {
+  return detail::get_header_value_u64(headers, key, id, 0);
+}
+
+inline uint64_t Response::get_header_value_u64(const std::string &key,
+                                               size_t id) const {
+  return detail::get_header_value_u64(headers, key, id, 0);
+}
+
+template <typename... Args>
+inline ssize_t Stream::write_format(const char *fmt, const Args &...args) {
+  const auto bufsiz = 2048;
+  std::array<char, bufsiz> buf{};
+
+  auto sn = snprintf(buf.data(), buf.size() - 1, fmt, args...);
+  if (sn <= 0) { return sn; }
+
+  auto n = static_cast<size_t>(sn);
+
+  if (n >= buf.size() - 1) {
+    std::vector<char> glowable_buf(buf.size());
+
+    while (n >= glowable_buf.size() - 1) {
+      glowable_buf.resize(glowable_buf.size() * 2);
+      n = static_cast<size_t>(
+          snprintf(&glowable_buf[0], glowable_buf.size() - 1, fmt, args...));
+    }
+    return write(&glowable_buf[0], n);
+  } else {
+    return write(buf.data(), n);
+  }
+}
+
+inline void default_socket_options(socket_t sock) {
+  int yes = 1;
+#ifdef _WIN32
+  setsockopt(sock, SOL_SOCKET, SO_REUSEADDR,
+             reinterpret_cast<const char *>(&yes), sizeof(yes));
+  setsockopt(sock, SOL_SOCKET, SO_EXCLUSIVEADDRUSE,
+             reinterpret_cast<const char *>(&yes), sizeof(yes));
+#else
+#ifdef SO_REUSEPORT
+  setsockopt(sock, SOL_SOCKET, SO_REUSEPORT,
+             reinterpret_cast<const void *>(&yes), sizeof(yes));
+#else
+  setsockopt(sock, SOL_SOCKET, SO_REUSEADDR,
+             reinterpret_cast<const void *>(&yes), sizeof(yes));
+#endif
+#endif
+}
+
+inline const char *status_message(int status) {
+  switch (status) {
+  case 100: return "Continue";
+  case 101: return "Switching Protocol";
+  case 102: return "Processing";
+  case 103: return "Early Hints";
+  case 200: return "OK";
+  case 201: return "Created";
+  case 202: return "Accepted";
+  case 203: return "Non-Authoritative Information";
+  case 204: return "No Content";
+  case 205: return "Reset Content";
+  case 206: return "Partial Content";
+  case 207: return "Multi-Status";
+  case 208: return "Already Reported";
+  case 226: return "IM Used";
+  case 300: return "Multiple Choice";
+  case 301: return "Moved Permanently";
+  case 302: return "Found";
+  case 303: return "See Other";
+  case 304: return "Not Modified";
+  case 305: return "Use Proxy";
+  case 306: return "unused";
+  case 307: return "Temporary Redirect";
+  case 308: return "Permanent Redirect";
+  case 400: return "Bad Request";
+  case 401: return "Unauthorized";
+  case 402: return "Payment Required";
+  case 403: return "Forbidden";
+  case 404: return "Not Found";
+  case 405: return "Method Not Allowed";
+  case 406: return "Not Acceptable";
+  case 407: return "Proxy Authentication Required";
+  case 408: return "Request Timeout";
+  case 409: return "Conflict";
+  case 410: return "Gone";
+  case 411: return "Length Required";
+  case 412: return "Precondition Failed";
+  case 413: return "Payload Too Large";
+  case 414: return "URI Too Long";
+  case 415: return "Unsupported Media Type";
+  case 416: return "Range Not Satisfiable";
+  case 417: return "Expectation Failed";
+  case 418: return "I'm a teapot";
+  case 421: return "Misdirected Request";
+  case 422: return "Unprocessable Entity";
+  case 423: return "Locked";
+  case 424: return "Failed Dependency";
+  case 425: return "Too Early";
+  case 426: return "Upgrade Required";
+  case 428: return "Precondition Required";
+  case 429: return "Too Many Requests";
+  case 431: return "Request Header Fields Too Large";
+  case 451: return "Unavailable For Legal Reasons";
+  case 501: return "Not Implemented";
+  case 502: return "Bad Gateway";
+  case 503: return "Service Unavailable";
+  case 504: return "Gateway Timeout";
+  case 505: return "HTTP Version Not Supported";
+  case 506: return "Variant Also Negotiates";
+  case 507: return "Insufficient Storage";
+  case 508: return "Loop Detected";
+  case 510: return "Not Extended";
+  case 511: return "Network Authentication Required";
+
+  default:
+  case 500: return "Internal Server Error";
+  }
+}
+
+template <class Rep, class Period>
+inline Server &
+Server::set_read_timeout(const std::chrono::duration<Rep, Period> &duration) {
+  detail::duration_to_sec_and_usec(
+      duration, [&](time_t sec, time_t usec) { set_read_timeout(sec, usec); });
+  return *this;
+}
+
+template <class Rep, class Period>
+inline Server &
+Server::set_write_timeout(const std::chrono::duration<Rep, Period> &duration) {
+  detail::duration_to_sec_and_usec(
+      duration, [&](time_t sec, time_t usec) { set_write_timeout(sec, usec); });
+  return *this;
+}
+
+template <class Rep, class Period>
+inline Server &
+Server::set_idle_interval(const std::chrono::duration<Rep, Period> &duration) {
+  detail::duration_to_sec_and_usec(
+      duration, [&](time_t sec, time_t usec) { set_idle_interval(sec, usec); });
+  return *this;
+}
+
+inline std::string to_string(const Error error) {
+  switch (error) {
+  case Error::Success: return "Success (no error)";
+  case Error::Connection: return "Could not establish connection";
+  case Error::BindIPAddress: return "Failed to bind IP address";
+  case Error::Read: return "Failed to read connection";
+  case Error::Write: return "Failed to write connection";
+  case Error::ExceedRedirectCount: return "Maximum redirect count exceeded";
+  case Error::Canceled: return "Connection handling canceled";
+  case Error::SSLConnection: return "SSL connection failed";
+  case Error::SSLLoadingCerts: return "SSL certificate loading failed";
+  case Error::SSLServerVerification: return "SSL server verification failed";
+  case Error::UnsupportedMultipartBoundaryChars:
+    return "Unsupported HTTP multipart boundary characters";
+  case Error::Compression: return "Compression failed";
+  case Error::ConnectionTimeout: return "Connection timed out";
+  case Error::ProxyConnection: return "Proxy connection failed";
+  case Error::Unknown: return "Unknown";
+  default: break;
+  }
+
+  return "Invalid";
+}
+
+inline std::ostream &operator<<(std::ostream &os, const Error &obj) {
+  os << to_string(obj);
+  os << " (" << static_cast<std::underlying_type<Error>::type>(obj) << ')';
+  return os;
+}
+
+inline uint64_t Result::get_request_header_value_u64(const std::string &key,
+                                                     size_t id) const {
+  return detail::get_header_value_u64(request_headers_, key, id, 0);
+}
+
+template <class Rep, class Period>
+inline void ClientImpl::set_connection_timeout(
+    const std::chrono::duration<Rep, Period> &duration) {
+  detail::duration_to_sec_and_usec(duration, [&](time_t sec, time_t usec) {
+    set_connection_timeout(sec, usec);
+  });
+}
+
+template <class Rep, class Period>
+inline void ClientImpl::set_read_timeout(
+    const std::chrono::duration<Rep, Period> &duration) {
+  detail::duration_to_sec_and_usec(
+      duration, [&](time_t sec, time_t usec) { set_read_timeout(sec, usec); });
+}
+
+template <class Rep, class Period>
+inline void ClientImpl::set_write_timeout(
+    const std::chrono::duration<Rep, Period> &duration) {
+  detail::duration_to_sec_and_usec(
+      duration, [&](time_t sec, time_t usec) { set_write_timeout(sec, usec); });
+}
+
+template <class Rep, class Period>
+inline void Client::set_connection_timeout(
+    const std::chrono::duration<Rep, Period> &duration) {
+  cli_->set_connection_timeout(duration);
+}
+
+template <class Rep, class Period>
+inline void
+Client::set_read_timeout(const std::chrono::duration<Rep, Period> &duration) {
+  cli_->set_read_timeout(duration);
+}
+
+template <class Rep, class Period>
+inline void
+Client::set_write_timeout(const std::chrono::duration<Rep, Period> &duration) {
+  cli_->set_write_timeout(duration);
+}
+
+/*
+ * Forward declarations and types that will be part of the .h file if split into
+ * .h + .cc.
+ */
+
+std::string hosted_at(const std::string &hostname);
+
+void hosted_at(const std::string &hostname, std::vector<std::string> &addrs);
+
+std::string append_query_params(const std::string &path, const Params &params);
+
+std::pair<std::string, std::string> make_range_header(Ranges ranges);
+
+std::pair<std::string, std::string>
+make_basic_authentication_header(const std::string &username,
+                                 const std::string &password,
+                                 bool is_proxy = false);
+
+namespace detail {
+
+std::string encode_query_param(const std::string &value);
+
+std::string decode_url(const std::string &s, bool convert_plus_to_space);
+
+void read_file(const std::string &path, std::string &out);
+
+std::string trim_copy(const std::string &s);
+
+void split(const char *b, const char *e, char d,
+           std::function<void(const char *, const char *)> fn);
+
+bool process_client_socket(socket_t sock, time_t read_timeout_sec,
+                           time_t read_timeout_usec, time_t write_timeout_sec,
+                           time_t write_timeout_usec,
+                           std::function<bool(Stream &)> callback);
+
+socket_t create_client_socket(
+    const std::string &host, const std::string &ip, int port,
+    int address_family, bool tcp_nodelay, SocketOptions socket_options,
+    time_t connection_timeout_sec, time_t connection_timeout_usec,
+    time_t read_timeout_sec, time_t read_timeout_usec, time_t write_timeout_sec,
+    time_t write_timeout_usec, const std::string &intf, Error &error);
+
+const char *get_header_value(const Headers &headers, const std::string &key,
+                             size_t id = 0, const char *def = nullptr);
+
+std::string params_to_query_str(const Params &params);
+
+void parse_query_text(const std::string &s, Params &params);
+
+bool parse_multipart_boundary(const std::string &content_type,
+                              std::string &boundary);
+
+bool parse_range_header(const std::string &s, Ranges &ranges);
+
+int close_socket(socket_t sock);
+
+ssize_t send_socket(socket_t sock, const void *ptr, size_t size, int flags);
+
+ssize_t read_socket(socket_t sock, void *ptr, size_t size, int flags);
+
+enum class EncodingType { None = 0, Gzip, Brotli };
+
+EncodingType encoding_type(const Request &req, const Response &res);
+
+class BufferStream : public Stream {
+public:
+  BufferStream() = default;
+  ~BufferStream() override = default;
+
+  bool is_readable() const override;
+  bool is_writable() const override;
+  ssize_t read(char *ptr, size_t size) override;
+  ssize_t write(const char *ptr, size_t size) override;
+  void get_remote_ip_and_port(std::string &ip, int &port) const override;
+  void get_local_ip_and_port(std::string &ip, int &port) const override;
+  socket_t socket() const override;
+
+  const std::string &get_buffer() const;
+
+private:
+  std::string buffer;
+  size_t position = 0;
+};
+
+class compressor {
+public:
+  virtual ~compressor() = default;
+
+  typedef std::function<bool(const char *data, size_t data_len)> Callback;
+  virtual bool compress(const char *data, size_t data_length, bool last,
+                        Callback callback) = 0;
+};
+
+class decompressor {
+public:
+  virtual ~decompressor() = default;
+
+  virtual bool is_valid() const = 0;
+
+  typedef std::function<bool(const char *data, size_t data_len)> Callback;
+  virtual bool decompress(const char *data, size_t data_length,
+                          Callback callback) = 0;
+};
+
+class nocompressor : public compressor {
+public:
+  virtual ~nocompressor() = default;
+
+  bool compress(const char *data, size_t data_length, bool /*last*/,
+                Callback callback) override;
+};
+
+#ifdef CPPHTTPLIB_ZLIB_SUPPORT
+class gzip_compressor : public compressor {
+public:
+  gzip_compressor();
+  ~gzip_compressor();
+
+  bool compress(const char *data, size_t data_length, bool last,
+                Callback callback) override;
+
+private:
+  bool is_valid_ = false;
+  z_stream strm_;
+};
+
+class gzip_decompressor : public decompressor {
+public:
+  gzip_decompressor();
+  ~gzip_decompressor();
+
+  bool is_valid() const override;
+
+  bool decompress(const char *data, size_t data_length,
+                  Callback callback) override;
+
+private:
+  bool is_valid_ = false;
+  z_stream strm_;
+};
+#endif
+
+#ifdef CPPHTTPLIB_BROTLI_SUPPORT
+class brotli_compressor : public compressor {
+public:
+  brotli_compressor();
+  ~brotli_compressor();
+
+  bool compress(const char *data, size_t data_length, bool last,
+                Callback callback) override;
+
+private:
+  BrotliEncoderState *state_ = nullptr;
+};
+
+class brotli_decompressor : public decompressor {
+public:
+  brotli_decompressor();
+  ~brotli_decompressor();
+
+  bool is_valid() const override;
+
+  bool decompress(const char *data, size_t data_length,
+                  Callback callback) override;
+
+private:
+  BrotliDecoderResult decoder_r;
+  BrotliDecoderState *decoder_s = nullptr;
+};
+#endif
+
+// NOTE: until the read size reaches `fixed_buffer_size`, use `fixed_buffer`
+// to store data. The call can set memory on stack for performance.
+class stream_line_reader {
+public:
+  stream_line_reader(Stream &strm, char *fixed_buffer,
+                     size_t fixed_buffer_size);
+  const char *ptr() const;
+  size_t size() const;
+  bool end_with_crlf() const;
+  bool getline();
+
+private:
+  void append(char c);
+
+  Stream &strm_;
+  char *fixed_buffer_;
+  const size_t fixed_buffer_size_;
+  size_t fixed_buffer_used_size_ = 0;
+  std::string glowable_buffer_;
+};
+
+class mmap {
+public:
+  mmap(const char *path);
+  ~mmap();
+
+  bool open(const char *path);
+  void close();
+
+  bool is_open() const;
+  size_t size() const;
+  const char *data() const;
+
+private:
+#if defined(_WIN32)
+  HANDLE hFile_;
+  HANDLE hMapping_;
+#else
+  int fd_;
+#endif
+  size_t size_;
+  void *addr_;
+};
+
+} // namespace detail
+
+// ----------------------------------------------------------------------------
+
+/*
+ * Implementation that will be part of the .cc file if split into .h + .cc.
+ */
+
+namespace detail {
+
+inline bool is_hex(char c, int &v) {
+  if (0x20 <= c && isdigit(c)) {
+    v = c - '0';
+    return true;
+  } else if ('A' <= c && c <= 'F') {
+    v = c - 'A' + 10;
+    return true;
+  } else if ('a' <= c && c <= 'f') {
+    v = c - 'a' + 10;
+    return true;
+  }
+  return false;
+}
+
+inline bool from_hex_to_i(const std::string &s, size_t i, size_t cnt,
+                          int &val) {
+  if (i >= s.size()) { return false; }
+
+  val = 0;
+  for (; cnt; i++, cnt--) {
+    if (!s[i]) { return false; }
+    auto v = 0;
+    if (is_hex(s[i], v)) {
+      val = val * 16 + v;
+    } else {
+      return false;
+    }
+  }
+  return true;
+}
+
+inline std::string from_i_to_hex(size_t n) {
+  static const auto charset = "0123456789abcdef";
+  std::string ret;
+  do {
+    ret = charset[n & 15] + ret;
+    n >>= 4;
+  } while (n > 0);
+  return ret;
+}
+
+inline size_t to_utf8(int code, char *buff) {
+  if (code < 0x0080) {
+    buff[0] = (code & 0x7F);
+    return 1;
+  } else if (code < 0x0800) {
+    buff[0] = static_cast<char>(0xC0 | ((code >> 6) & 0x1F));
+    buff[1] = static_cast<char>(0x80 | (code & 0x3F));
+    return 2;
+  } else if (code < 0xD800) {
+    buff[0] = static_cast<char>(0xE0 | ((code >> 12) & 0xF));
+    buff[1] = static_cast<char>(0x80 | ((code >> 6) & 0x3F));
+    buff[2] = static_cast<char>(0x80 | (code & 0x3F));
+    return 3;
+  } else if (code < 0xE000) { // D800 - DFFF is invalid...
+    return 0;
+  } else if (code < 0x10000) {
+    buff[0] = static_cast<char>(0xE0 | ((code >> 12) & 0xF));
+    buff[1] = static_cast<char>(0x80 | ((code >> 6) & 0x3F));
+    buff[2] = static_cast<char>(0x80 | (code & 0x3F));
+    return 3;
+  } else if (code < 0x110000) {
+    buff[0] = static_cast<char>(0xF0 | ((code >> 18) & 0x7));
+    buff[1] = static_cast<char>(0x80 | ((code >> 12) & 0x3F));
+    buff[2] = static_cast<char>(0x80 | ((code >> 6) & 0x3F));
+    buff[3] = static_cast<char>(0x80 | (code & 0x3F));
+    return 4;
+  }
+
+  // NOTREACHED
+  return 0;
+}
+
+// NOTE: This code came up with the following stackoverflow post:
+// https://stackoverflow.com/questions/180947/base64-decode-snippet-in-c
+inline std::string base64_encode(const std::string &in) {
+  static const auto lookup =
+      "ABCDEFGHIJKLMNOPQRSTUVWXYZabcdefghijklmnopqrstuvwxyz0123456789+/";
+
+  std::string out;
+  out.reserve(in.size());
+
+  auto val = 0;
+  auto valb = -6;
+
+  for (auto c : in) {
+    val = (val << 8) + static_cast<uint8_t>(c);
+    valb += 8;
+    while (valb >= 0) {
+      out.push_back(lookup[(val >> valb) & 0x3F]);
+      valb -= 6;
+    }
+  }
+
+  if (valb > -6) { out.push_back(lookup[((val << 8) >> (valb + 8)) & 0x3F]); }
+
+  while (out.size() % 4) {
+    out.push_back('=');
+  }
+
+  return out;
+}
+
+inline bool is_file(const std::string &path) {
+#ifdef _WIN32
+  return _access_s(path.c_str(), 0) == 0;
+#else
+  struct stat st;
+  return stat(path.c_str(), &st) >= 0 && S_ISREG(st.st_mode);
+#endif
+}
+
+inline bool is_dir(const std::string &path) {
+  struct stat st;
+  return stat(path.c_str(), &st) >= 0 && S_ISDIR(st.st_mode);
+}
+
+inline bool is_valid_path(const std::string &path) {
+  size_t level = 0;
+  size_t i = 0;
+
+  // Skip slash
+  while (i < path.size() && path[i] == '/') {
+    i++;
+  }
+
+  while (i < path.size()) {
+    // Read component
+    auto beg = i;
+    while (i < path.size() && path[i] != '/') {
+      i++;
+    }
+
+    auto len = i - beg;
+    assert(len > 0);
+
+    if (!path.compare(beg, len, ".")) {
+      ;
+    } else if (!path.compare(beg, len, "..")) {
+      if (level == 0) { return false; }
+      level--;
+    } else {
+      level++;
+    }
+
+    // Skip slash
+    while (i < path.size() && path[i] == '/') {
+      i++;
+    }
+  }
+
+  return true;
+}
+
+inline std::string encode_query_param(const std::string &value) {
+  std::ostringstream escaped;
+  escaped.fill('0');
+  escaped << std::hex;
+
+  for (auto c : value) {
+    if (std::isalnum(static_cast<uint8_t>(c)) || c == '-' || c == '_' ||
+        c == '.' || c == '!' || c == '~' || c == '*' || c == '\'' || c == '(' ||
+        c == ')') {
+      escaped << c;
+    } else {
+      escaped << std::uppercase;
+      escaped << '%' << std::setw(2)
+              << static_cast<int>(static_cast<unsigned char>(c));
+      escaped << std::nouppercase;
+    }
+  }
+
+  return escaped.str();
+}
+
+inline std::string encode_url(const std::string &s) {
+  std::string result;
+  result.reserve(s.size());
+
+  for (size_t i = 0; s[i]; i++) {
+    switch (s[i]) {
+    case ' ': result += "%20"; break;
+    case '+': result += "%2B"; break;
+    case '\r': result += "%0D"; break;
+    case '\n': result += "%0A"; break;
+    case '\'': result += "%27"; break;
+    case ',': result += "%2C"; break;
+    // case ':': result += "%3A"; break; // ok? probably...
+    case ';': result += "%3B"; break;
+    default:
+      auto c = static_cast<uint8_t>(s[i]);
+      if (c >= 0x80) {
+        result += '%';
+        char hex[4];
+        auto len = snprintf(hex, sizeof(hex) - 1, "%02X", c);
+        assert(len == 2);
+        result.append(hex, static_cast<size_t>(len));
+      } else {
+        result += s[i];
+      }
+      break;
+    }
+  }
+
+  return result;
+}
+
+inline std::string decode_url(const std::string &s,
+                              bool convert_plus_to_space) {
+  std::string result;
+
+  for (size_t i = 0; i < s.size(); i++) {
+    if (s[i] == '%' && i + 1 < s.size()) {
+      if (s[i + 1] == 'u') {
+        auto val = 0;
+        if (from_hex_to_i(s, i + 2, 4, val)) {
+          // 4 digits Unicode codes
+          char buff[4];
+          size_t len = to_utf8(val, buff);
+          if (len > 0) { result.append(buff, len); }
+          i += 5; // 'u0000'
+        } else {
+          result += s[i];
+        }
+      } else {
+        auto val = 0;
+        if (from_hex_to_i(s, i + 1, 2, val)) {
+          // 2 digits hex codes
+          result += static_cast<char>(val);
+          i += 2; // '00'
+        } else {
+          result += s[i];
+        }
+      }
+    } else if (convert_plus_to_space && s[i] == '+') {
+      result += ' ';
+    } else {
+      result += s[i];
+    }
+  }
+
+  return result;
+}
+
+inline void read_file(const std::string &path, std::string &out) {
+  std::ifstream fs(path, std::ios_base::binary);
+  fs.seekg(0, std::ios_base::end);
+  auto size = fs.tellg();
+  fs.seekg(0);
+  out.resize(static_cast<size_t>(size));
+  fs.read(&out[0], static_cast<std::streamsize>(size));
+}
+
+inline std::string file_extension(const std::string &path) {
+  std::smatch m;
+  static auto re = std::regex("\\.([a-zA-Z0-9]+)$");
+  if (std::regex_search(path, m, re)) { return m[1].str(); }
+  return std::string();
+}
+
+inline bool is_space_or_tab(char c) { return c == ' ' || c == '\t'; }
+
+inline std::pair<size_t, size_t> trim(const char *b, const char *e, size_t left,
+                                      size_t right) {
+  while (b + left < e && is_space_or_tab(b[left])) {
+    left++;
+  }
+  while (right > 0 && is_space_or_tab(b[right - 1])) {
+    right--;
+  }
+  return std::make_pair(left, right);
+}
+
+inline std::string trim_copy(const std::string &s) {
+  auto r = trim(s.data(), s.data() + s.size(), 0, s.size());
+  return s.substr(r.first, r.second - r.first);
+}
+
+inline std::string trim_double_quotes_copy(const std::string &s) {
+  if (s.length() >= 2 && s.front() == '"' && s.back() == '"') {
+    return s.substr(1, s.size() - 2);
+  }
+  return s;
+}
+
+inline void split(const char *b, const char *e, char d,
+                  std::function<void(const char *, const char *)> fn) {
+  size_t i = 0;
+  size_t beg = 0;
+
+  while (e ? (b + i < e) : (b[i] != '\0')) {
+    if (b[i] == d) {
+      auto r = trim(b, e, beg, i);
+      if (r.first < r.second) { fn(&b[r.first], &b[r.second]); }
+      beg = i + 1;
+    }
+    i++;
+  }
+
+  if (i) {
+    auto r = trim(b, e, beg, i);
+    if (r.first < r.second) { fn(&b[r.first], &b[r.second]); }
+  }
+}
+
+inline stream_line_reader::stream_line_reader(Stream &strm, char *fixed_buffer,
+                                              size_t fixed_buffer_size)
+    : strm_(strm), fixed_buffer_(fixed_buffer),
+      fixed_buffer_size_(fixed_buffer_size) {}
+
+inline const char *stream_line_reader::ptr() const {
+  if (glowable_buffer_.empty()) {
+    return fixed_buffer_;
+  } else {
+    return glowable_buffer_.data();
+  }
+}
+
+inline size_t stream_line_reader::size() const {
+  if (glowable_buffer_.empty()) {
+    return fixed_buffer_used_size_;
+  } else {
+    return glowable_buffer_.size();
+  }
+}
+
+inline bool stream_line_reader::end_with_crlf() const {
+  auto end = ptr() + size();
+  return size() >= 2 && end[-2] == '\r' && end[-1] == '\n';
+}
+
+inline bool stream_line_reader::getline() {
+  fixed_buffer_used_size_ = 0;
+  glowable_buffer_.clear();
+
+  for (size_t i = 0;; i++) {
+    char byte;
+    auto n = strm_.read(&byte, 1);
+
+    if (n < 0) {
+      return false;
+    } else if (n == 0) {
+      if (i == 0) {
+        return false;
+      } else {
+        break;
+      }
+    }
+
+    append(byte);
+
+    if (byte == '\n') { break; }
+  }
+
+  return true;
+}
+
+inline void stream_line_reader::append(char c) {
+  if (fixed_buffer_used_size_ < fixed_buffer_size_ - 1) {
+    fixed_buffer_[fixed_buffer_used_size_++] = c;
+    fixed_buffer_[fixed_buffer_used_size_] = '\0';
+  } else {
+    if (glowable_buffer_.empty()) {
+      assert(fixed_buffer_[fixed_buffer_used_size_] == '\0');
+      glowable_buffer_.assign(fixed_buffer_, fixed_buffer_used_size_);
+    }
+    glowable_buffer_ += c;
+  }
+}
+
+inline mmap::mmap(const char *path)
+#if defined(_WIN32)
+    : hFile_(NULL), hMapping_(NULL)
+#else
+    : fd_(-1)
+#endif
+      ,
+      size_(0), addr_(nullptr) {
+  if (!open(path)) { std::runtime_error(""); }
+}
+
+inline mmap::~mmap() { close(); }
+
+inline bool mmap::open(const char *path) {
+  close();
+
+#if defined(_WIN32)
+  hFile_ = ::CreateFileA(path, GENERIC_READ, FILE_SHARE_READ, NULL,
+                         OPEN_EXISTING, FILE_ATTRIBUTE_NORMAL, NULL);
+
+  if (hFile_ == INVALID_HANDLE_VALUE) { return false; }
+
+  size_ = ::GetFileSize(hFile_, NULL);
+
+  hMapping_ = ::CreateFileMapping(hFile_, NULL, PAGE_READONLY, 0, 0, NULL);
+
+  if (hMapping_ == NULL) {
+    close();
+    return false;
+  }
+
+  addr_ = ::MapViewOfFile(hMapping_, FILE_MAP_READ, 0, 0, 0);
+#else
+  fd_ = ::open(path, O_RDONLY);
+  if (fd_ == -1) { return false; }
+
+  struct stat sb;
+  if (fstat(fd_, &sb) == -1) {
+    close();
+    return false;
+  }
+  size_ = static_cast<size_t>(sb.st_size);
+
+  addr_ = ::mmap(NULL, size_, PROT_READ, MAP_PRIVATE, fd_, 0);
+#endif
+
+  if (addr_ == nullptr) {
+    close();
+    return false;
+  }
+
+  return true;
+}
+
+inline bool mmap::is_open() const { return addr_ != nullptr; }
+
+inline size_t mmap::size() const { return size_; }
+
+inline const char *mmap::data() const { return (const char *)addr_; }
+
+inline void mmap::close() {
+#if defined(_WIN32)
+  if (addr_) {
+    ::UnmapViewOfFile(addr_);
+    addr_ = nullptr;
+  }
+
+  if (hMapping_) {
+    ::CloseHandle(hMapping_);
+    hMapping_ = NULL;
+  }
+
+  if (hFile_ != INVALID_HANDLE_VALUE) {
+    ::CloseHandle(hFile_);
+    hFile_ = INVALID_HANDLE_VALUE;
+  }
+#else
+  if (addr_ != nullptr) {
+    munmap(addr_, size_);
+    addr_ = nullptr;
+  }
+
+  if (fd_ != -1) {
+    ::close(fd_);
+    fd_ = -1;
+  }
+#endif
+  size_ = 0;
+}
+inline int close_socket(socket_t sock) {
+#ifdef _WIN32
+  return closesocket(sock);
+#else
+  return close(sock);
+#endif
+}
+
+template <typename T> inline ssize_t handle_EINTR(T fn) {
+  ssize_t res = 0;
+  while (true) {
+    res = fn();
+    if (res < 0 && errno == EINTR) { continue; }
+    break;
+  }
+  return res;
+}
+
+inline ssize_t read_socket(socket_t sock, void *ptr, size_t size, int flags) {
+  return handle_EINTR([&]() {
+    return recv(sock,
+#ifdef _WIN32
+                static_cast<char *>(ptr), static_cast<int>(size),
+#else
+                ptr, size,
+#endif
+                flags);
+  });
+}
+
+inline ssize_t send_socket(socket_t sock, const void *ptr, size_t size,
+                           int flags) {
+  return handle_EINTR([&]() {
+    return send(sock,
+#ifdef _WIN32
+                static_cast<const char *>(ptr), static_cast<int>(size),
+#else
+                ptr, size,
+#endif
+                flags);
+  });
+}
+
+inline ssize_t select_read(socket_t sock, time_t sec, time_t usec) {
+#ifdef CPPHTTPLIB_USE_POLL
+  struct pollfd pfd_read;
+  pfd_read.fd = sock;
+  pfd_read.events = POLLIN;
+
+  auto timeout = static_cast<int>(sec * 1000 + usec / 1000);
+
+  return handle_EINTR([&]() { return poll(&pfd_read, 1, timeout); });
+#else
+#ifndef _WIN32
+  if (sock >= FD_SETSIZE) { return 1; }
+#endif
+
+  fd_set fds;
+  FD_ZERO(&fds);
+  FD_SET(sock, &fds);
+
+  timeval tv;
+  tv.tv_sec = static_cast<long>(sec);
+  tv.tv_usec = static_cast<decltype(tv.tv_usec)>(usec);
+
+  return handle_EINTR([&]() {
+    return select(static_cast<int>(sock + 1), &fds, nullptr, nullptr, &tv);
+  });
+#endif
+}
+
+inline ssize_t select_write(socket_t sock, time_t sec, time_t usec) {
+#ifdef CPPHTTPLIB_USE_POLL
+  struct pollfd pfd_read;
+  pfd_read.fd = sock;
+  pfd_read.events = POLLOUT;
+
+  auto timeout = static_cast<int>(sec * 1000 + usec / 1000);
+
+  return handle_EINTR([&]() { return poll(&pfd_read, 1, timeout); });
+#else
+#ifndef _WIN32
+  if (sock >= FD_SETSIZE) { return 1; }
+#endif
+
+  fd_set fds;
+  FD_ZERO(&fds);
+  FD_SET(sock, &fds);
+
+  timeval tv;
+  tv.tv_sec = static_cast<long>(sec);
+  tv.tv_usec = static_cast<decltype(tv.tv_usec)>(usec);
+
+  return handle_EINTR([&]() {
+    return select(static_cast<int>(sock + 1), nullptr, &fds, nullptr, &tv);
+  });
+#endif
+}
+
+inline Error wait_until_socket_is_ready(socket_t sock, time_t sec,
+                                        time_t usec) {
+#ifdef CPPHTTPLIB_USE_POLL
+  struct pollfd pfd_read;
+  pfd_read.fd = sock;
+  pfd_read.events = POLLIN | POLLOUT;
+
+  auto timeout = static_cast<int>(sec * 1000 + usec / 1000);
+
+  auto poll_res = handle_EINTR([&]() { return poll(&pfd_read, 1, timeout); });
+
+  if (poll_res == 0) { return Error::ConnectionTimeout; }
+
+  if (poll_res > 0 && pfd_read.revents & (POLLIN | POLLOUT)) {
+    auto error = 0;
+    socklen_t len = sizeof(error);
+    auto res = getsockopt(sock, SOL_SOCKET, SO_ERROR,
+                          reinterpret_cast<char *>(&error), &len);
+    auto successful = res >= 0 && !error;
+    return successful ? Error::Success : Error::Connection;
+  }
+
+  return Error::Connection;
+#else
+#ifndef _WIN32
+  if (sock >= FD_SETSIZE) { return Error::Connection; }
+#endif
+
+  fd_set fdsr;
+  FD_ZERO(&fdsr);
+  FD_SET(sock, &fdsr);
+
+  auto fdsw = fdsr;
+  auto fdse = fdsr;
+
+  timeval tv;
+  tv.tv_sec = static_cast<long>(sec);
+  tv.tv_usec = static_cast<decltype(tv.tv_usec)>(usec);
+
+  auto ret = handle_EINTR([&]() {
+    return select(static_cast<int>(sock + 1), &fdsr, &fdsw, &fdse, &tv);
+  });
+
+  if (ret == 0) { return Error::ConnectionTimeout; }
+
+  if (ret > 0 && (FD_ISSET(sock, &fdsr) || FD_ISSET(sock, &fdsw))) {
+    auto error = 0;
+    socklen_t len = sizeof(error);
+    auto res = getsockopt(sock, SOL_SOCKET, SO_ERROR,
+                          reinterpret_cast<char *>(&error), &len);
+    auto successful = res >= 0 && !error;
+    return successful ? Error::Success : Error::Connection;
+  }
+  return Error::Connection;
+#endif
+}
+
+inline bool is_socket_alive(socket_t sock) {
+  const auto val = detail::select_read(sock, 0, 0);
+  if (val == 0) {
+    return true;
+  } else if (val < 0 && errno == EBADF) {
+    return false;
+  }
+  char buf[1];
+  return detail::read_socket(sock, &buf[0], sizeof(buf), MSG_PEEK) > 0;
+}
+
+class SocketStream : public Stream {
+public:
+  SocketStream(socket_t sock, time_t read_timeout_sec, time_t read_timeout_usec,
+               time_t write_timeout_sec, time_t write_timeout_usec);
+  ~SocketStream() override;
+
+  bool is_readable() const override;
+  bool is_writable() const override;
+  ssize_t read(char *ptr, size_t size) override;
+  ssize_t write(const char *ptr, size_t size) override;
+  void get_remote_ip_and_port(std::string &ip, int &port) const override;
+  void get_local_ip_and_port(std::string &ip, int &port) const override;
+  socket_t socket() const override;
+
+private:
+  socket_t sock_;
+  time_t read_timeout_sec_;
+  time_t read_timeout_usec_;
+  time_t write_timeout_sec_;
+  time_t write_timeout_usec_;
+
+  std::vector<char> read_buff_;
+  size_t read_buff_off_ = 0;
+  size_t read_buff_content_size_ = 0;
+
+  static const size_t read_buff_size_ = 1024 * 4;
+};
+
+#ifdef CPPHTTPLIB_OPENSSL_SUPPORT
+class SSLSocketStream : public Stream {
+public:
+  SSLSocketStream(socket_t sock, SSL *ssl, time_t read_timeout_sec,
+                  time_t read_timeout_usec, time_t write_timeout_sec,
+                  time_t write_timeout_usec);
+  ~SSLSocketStream() override;
+
+  bool is_readable() const override;
+  bool is_writable() const override;
+  ssize_t read(char *ptr, size_t size) override;
+  ssize_t write(const char *ptr, size_t size) override;
+  void get_remote_ip_and_port(std::string &ip, int &port) const override;
+  void get_local_ip_and_port(std::string &ip, int &port) const override;
+  socket_t socket() const override;
+
+private:
+  socket_t sock_;
+  SSL *ssl_;
+  time_t read_timeout_sec_;
+  time_t read_timeout_usec_;
+  time_t write_timeout_sec_;
+  time_t write_timeout_usec_;
+};
+#endif
+
+inline bool keep_alive(socket_t sock, time_t keep_alive_timeout_sec) {
+  using namespace std::chrono;
+  auto start = steady_clock::now();
+  while (true) {
+    auto val = select_read(sock, 0, 10000);
+    if (val < 0) {
+      return false;
+    } else if (val == 0) {
+      auto current = steady_clock::now();
+      auto duration = duration_cast<milliseconds>(current - start);
+      auto timeout = keep_alive_timeout_sec * 1000;
+      if (duration.count() > timeout) { return false; }
+      std::this_thread::sleep_for(std::chrono::milliseconds(1));
+    } else {
+      return true;
+    }
+  }
+}
+
+template <typename T>
+inline bool
+process_server_socket_core(const std::atomic<socket_t> &svr_sock, socket_t sock,
+                           size_t keep_alive_max_count,
+                           time_t keep_alive_timeout_sec, T callback) {
+  assert(keep_alive_max_count > 0);
+  auto ret = false;
+  auto count = keep_alive_max_count;
+  while (svr_sock != INVALID_SOCKET && count > 0 &&
+         keep_alive(sock, keep_alive_timeout_sec)) {
+    auto close_connection = count == 1;
+    auto connection_closed = false;
+    ret = callback(close_connection, connection_closed);
+    if (!ret || connection_closed) { break; }
+    count--;
+  }
+  return ret;
+}
+
+template <typename T>
+inline bool
+process_server_socket(const std::atomic<socket_t> &svr_sock, socket_t sock,
+                      size_t keep_alive_max_count,
+                      time_t keep_alive_timeout_sec, time_t read_timeout_sec,
+                      time_t read_timeout_usec, time_t write_timeout_sec,
+                      time_t write_timeout_usec, T callback) {
+  return process_server_socket_core(
+      svr_sock, sock, keep_alive_max_count, keep_alive_timeout_sec,
+      [&](bool close_connection, bool &connection_closed) {
+        SocketStream strm(sock, read_timeout_sec, read_timeout_usec,
+                          write_timeout_sec, write_timeout_usec);
+        return callback(strm, close_connection, connection_closed);
+      });
+}
+
+inline bool process_client_socket(socket_t sock, time_t read_timeout_sec,
+                                  time_t read_timeout_usec,
+                                  time_t write_timeout_sec,
+                                  time_t write_timeout_usec,
+                                  std::function<bool(Stream &)> callback) {
+  SocketStream strm(sock, read_timeout_sec, read_timeout_usec,
+                    write_timeout_sec, write_timeout_usec);
+  return callback(strm);
+}
+
+inline int shutdown_socket(socket_t sock) {
+#ifdef _WIN32
+  return shutdown(sock, SD_BOTH);
+#else
+  return shutdown(sock, SHUT_RDWR);
+#endif
+}
+
+template <typename BindOrConnect>
+socket_t create_socket(const std::string &host, const std::string &ip, int port,
+                       int address_family, int socket_flags, bool tcp_nodelay,
+                       SocketOptions socket_options,
+                       BindOrConnect bind_or_connect) {
+  // Get address info
+  const char *node = nullptr;
+  struct addrinfo hints;
+  struct addrinfo *result;
+
+  memset(&hints, 0, sizeof(struct addrinfo));
+  hints.ai_socktype = SOCK_STREAM;
+  hints.ai_protocol = 0;
+
+  if (!ip.empty()) {
+    node = ip.c_str();
+    // Ask getaddrinfo to convert IP in c-string to address
+    hints.ai_family = AF_UNSPEC;
+    hints.ai_flags = AI_NUMERICHOST;
+  } else {
+    if (!host.empty()) { node = host.c_str(); }
+    hints.ai_family = address_family;
+    hints.ai_flags = socket_flags;
+  }
+
+#ifndef _WIN32
+  if (hints.ai_family == AF_UNIX) {
+    const auto addrlen = host.length();
+    if (addrlen > sizeof(sockaddr_un::sun_path)) return INVALID_SOCKET;
+
+    auto sock = socket(hints.ai_family, hints.ai_socktype, hints.ai_protocol);
+    if (sock != INVALID_SOCKET) {
+      sockaddr_un addr{};
+      addr.sun_family = AF_UNIX;
+      std::copy(host.begin(), host.end(), addr.sun_path);
+
+      hints.ai_addr = reinterpret_cast<sockaddr *>(&addr);
+      hints.ai_addrlen = static_cast<socklen_t>(
+          sizeof(addr) - sizeof(addr.sun_path) + addrlen);
+
+      fcntl(sock, F_SETFD, FD_CLOEXEC);
+      if (socket_options) { socket_options(sock); }
+
+      if (!bind_or_connect(sock, hints)) {
+        close_socket(sock);
+        sock = INVALID_SOCKET;
+      }
+    }
+    return sock;
+  }
+#endif
+
+  auto service = std::to_string(port);
+
+  if (getaddrinfo(node, service.c_str(), &hints, &result)) {
+#if defined __linux__ && !defined __ANDROID__
+    res_init();
+#endif
+    return INVALID_SOCKET;
+  }
+
+  for (auto rp = result; rp; rp = rp->ai_next) {
+    // Create a socket
+#ifdef _WIN32
+    auto sock =
+        WSASocketW(rp->ai_family, rp->ai_socktype, rp->ai_protocol, nullptr, 0,
+                   WSA_FLAG_NO_HANDLE_INHERIT | WSA_FLAG_OVERLAPPED);
+    /**
+     * Since the WSA_FLAG_NO_HANDLE_INHERIT is only supported on Windows 7 SP1
+     * and above the socket creation fails on older Windows Systems.
+     *
+     * Let's try to create a socket the old way in this case.
+     *
+     * Reference:
+     * https://docs.microsoft.com/en-us/windows/win32/api/winsock2/nf-winsock2-wsasocketa
+     *
+     * WSA_FLAG_NO_HANDLE_INHERIT:
+     * This flag is supported on Windows 7 with SP1, Windows Server 2008 R2 with
+     * SP1, and later
+     *
+     */
+    if (sock == INVALID_SOCKET) {
+      sock = socket(rp->ai_family, rp->ai_socktype, rp->ai_protocol);
+    }
+#else
+    auto sock = socket(rp->ai_family, rp->ai_socktype, rp->ai_protocol);
+#endif
+    if (sock == INVALID_SOCKET) { continue; }
+
+#ifndef _WIN32
+    if (fcntl(sock, F_SETFD, FD_CLOEXEC) == -1) {
+      close_socket(sock);
+      continue;
+    }
+#endif
+
+    if (tcp_nodelay) {
+      auto yes = 1;
+#ifdef _WIN32
+      setsockopt(sock, IPPROTO_TCP, TCP_NODELAY,
+                 reinterpret_cast<const char *>(&yes), sizeof(yes));
+#else
+      setsockopt(sock, IPPROTO_TCP, TCP_NODELAY,
+                 reinterpret_cast<const void *>(&yes), sizeof(yes));
+#endif
+    }
+
+    if (socket_options) { socket_options(sock); }
+
+    if (rp->ai_family == AF_INET6) {
+      auto no = 0;
+#ifdef _WIN32
+      setsockopt(sock, IPPROTO_IPV6, IPV6_V6ONLY,
+                 reinterpret_cast<const char *>(&no), sizeof(no));
+#else
+      setsockopt(sock, IPPROTO_IPV6, IPV6_V6ONLY,
+                 reinterpret_cast<const void *>(&no), sizeof(no));
+#endif
+    }
+
+    // bind or connect
+    if (bind_or_connect(sock, *rp)) {
+      freeaddrinfo(result);
+      return sock;
+    }
+
+    close_socket(sock);
+  }
+
+  freeaddrinfo(result);
+  return INVALID_SOCKET;
+}
+
+inline void set_nonblocking(socket_t sock, bool nonblocking) {
+#ifdef _WIN32
+  auto flags = nonblocking ? 1UL : 0UL;
+  ioctlsocket(sock, FIONBIO, &flags);
+#else
+  auto flags = fcntl(sock, F_GETFL, 0);
+  fcntl(sock, F_SETFL,
+        nonblocking ? (flags | O_NONBLOCK) : (flags & (~O_NONBLOCK)));
+#endif
+}
+
+inline bool is_connection_error() {
+#ifdef _WIN32
+  return WSAGetLastError() != WSAEWOULDBLOCK;
+#else
+  return errno != EINPROGRESS;
+#endif
+}
+
+inline bool bind_ip_address(socket_t sock, const std::string &host) {
+  struct addrinfo hints;
+  struct addrinfo *result;
+
+  memset(&hints, 0, sizeof(struct addrinfo));
+  hints.ai_family = AF_UNSPEC;
+  hints.ai_socktype = SOCK_STREAM;
+  hints.ai_protocol = 0;
+
+  if (getaddrinfo(host.c_str(), "0", &hints, &result)) { return false; }
+
+  auto ret = false;
+  for (auto rp = result; rp; rp = rp->ai_next) {
+    const auto &ai = *rp;
+    if (!::bind(sock, ai.ai_addr, static_cast<socklen_t>(ai.ai_addrlen))) {
+      ret = true;
+      break;
+    }
+  }
+
+  freeaddrinfo(result);
+  return ret;
+}
+
+#if !defined _WIN32 && !defined ANDROID && !defined _AIX && !defined __MVS__
+#define USE_IF2IP
+#endif
+
+#ifdef USE_IF2IP
+inline std::string if2ip(int address_family, const std::string &ifn) {
+  struct ifaddrs *ifap;
+  getifaddrs(&ifap);
+  std::string addr_candidate;
+  for (auto ifa = ifap; ifa; ifa = ifa->ifa_next) {
+    if (ifa->ifa_addr && ifn == ifa->ifa_name &&
+        (AF_UNSPEC == address_family ||
+         ifa->ifa_addr->sa_family == address_family)) {
+      if (ifa->ifa_addr->sa_family == AF_INET) {
+        auto sa = reinterpret_cast<struct sockaddr_in *>(ifa->ifa_addr);
+        char buf[INET_ADDRSTRLEN];
+        if (inet_ntop(AF_INET, &sa->sin_addr, buf, INET_ADDRSTRLEN)) {
+          freeifaddrs(ifap);
+          return std::string(buf, INET_ADDRSTRLEN);
+        }
+      } else if (ifa->ifa_addr->sa_family == AF_INET6) {
+        auto sa = reinterpret_cast<struct sockaddr_in6 *>(ifa->ifa_addr);
+        if (!IN6_IS_ADDR_LINKLOCAL(&sa->sin6_addr)) {
+          char buf[INET6_ADDRSTRLEN] = {};
+          if (inet_ntop(AF_INET6, &sa->sin6_addr, buf, INET6_ADDRSTRLEN)) {
+            // equivalent to mac's IN6_IS_ADDR_UNIQUE_LOCAL
+            auto s6_addr_head = sa->sin6_addr.s6_addr[0];
+            if (s6_addr_head == 0xfc || s6_addr_head == 0xfd) {
+              addr_candidate = std::string(buf, INET6_ADDRSTRLEN);
+            } else {
+              freeifaddrs(ifap);
+              return std::string(buf, INET6_ADDRSTRLEN);
+            }
+          }
+        }
+      }
+    }
+  }
+  freeifaddrs(ifap);
+  return addr_candidate;
+}
+#endif
+
+inline socket_t create_client_socket(
+    const std::string &host, const std::string &ip, int port,
+    int address_family, bool tcp_nodelay, SocketOptions socket_options,
+    time_t connection_timeout_sec, time_t connection_timeout_usec,
+    time_t read_timeout_sec, time_t read_timeout_usec, time_t write_timeout_sec,
+    time_t write_timeout_usec, const std::string &intf, Error &error) {
+  auto sock = create_socket(
+      host, ip, port, address_family, 0, tcp_nodelay, std::move(socket_options),
+      [&](socket_t sock2, struct addrinfo &ai) -> bool {
+        if (!intf.empty()) {
+#ifdef USE_IF2IP
+          auto ip_from_if = if2ip(address_family, intf);
+          if (ip_from_if.empty()) { ip_from_if = intf; }
+          if (!bind_ip_address(sock2, ip_from_if.c_str())) {
+            error = Error::BindIPAddress;
+            return false;
+          }
+#endif
+        }
+
+        set_nonblocking(sock2, true);
+
+        auto ret =
+            ::connect(sock2, ai.ai_addr, static_cast<socklen_t>(ai.ai_addrlen));
+
+        if (ret < 0) {
+          if (is_connection_error()) {
+            error = Error::Connection;
+            return false;
+          }
+          error = wait_until_socket_is_ready(sock2, connection_timeout_sec,
+                                             connection_timeout_usec);
+          if (error != Error::Success) { return false; }
+        }
+
+        set_nonblocking(sock2, false);
+
+        {
+#ifdef _WIN32
+          auto timeout = static_cast<uint32_t>(read_timeout_sec * 1000 +
+                                               read_timeout_usec / 1000);
+          setsockopt(sock2, SOL_SOCKET, SO_RCVTIMEO,
+                     reinterpret_cast<const char *>(&timeout), sizeof(timeout));
+#else
+          timeval tv;
+          tv.tv_sec = static_cast<long>(read_timeout_sec);
+          tv.tv_usec = static_cast<decltype(tv.tv_usec)>(read_timeout_usec);
+          setsockopt(sock2, SOL_SOCKET, SO_RCVTIMEO,
+                     reinterpret_cast<const void *>(&tv), sizeof(tv));
+#endif
+        }
+        {
+
+#ifdef _WIN32
+          auto timeout = static_cast<uint32_t>(write_timeout_sec * 1000 +
+                                               write_timeout_usec / 1000);
+          setsockopt(sock2, SOL_SOCKET, SO_SNDTIMEO,
+                     reinterpret_cast<const char *>(&timeout), sizeof(timeout));
+#else
+          timeval tv;
+          tv.tv_sec = static_cast<long>(write_timeout_sec);
+          tv.tv_usec = static_cast<decltype(tv.tv_usec)>(write_timeout_usec);
+          setsockopt(sock2, SOL_SOCKET, SO_SNDTIMEO,
+                     reinterpret_cast<const void *>(&tv), sizeof(tv));
+#endif
+        }
+
+        error = Error::Success;
+        return true;
+      });
+
+  if (sock != INVALID_SOCKET) {
+    error = Error::Success;
+  } else {
+    if (error == Error::Success) { error = Error::Connection; }
+  }
+
+  return sock;
+}
+
+inline bool get_ip_and_port(const struct sockaddr_storage &addr,
+                            socklen_t addr_len, std::string &ip, int &port) {
+  if (addr.ss_family == AF_INET) {
+    port = ntohs(reinterpret_cast<const struct sockaddr_in *>(&addr)->sin_port);
+  } else if (addr.ss_family == AF_INET6) {
+    port =
+        ntohs(reinterpret_cast<const struct sockaddr_in6 *>(&addr)->sin6_port);
+  } else {
+    return false;
+  }
+
+  std::array<char, NI_MAXHOST> ipstr{};
+  if (getnameinfo(reinterpret_cast<const struct sockaddr *>(&addr), addr_len,
+                  ipstr.data(), static_cast<socklen_t>(ipstr.size()), nullptr,
+                  0, NI_NUMERICHOST)) {
+    return false;
+  }
+
+  ip = ipstr.data();
+  return true;
+}
+
+inline void get_local_ip_and_port(socket_t sock, std::string &ip, int &port) {
+  struct sockaddr_storage addr;
+  socklen_t addr_len = sizeof(addr);
+  if (!getsockname(sock, reinterpret_cast<struct sockaddr *>(&addr),
+                   &addr_len)) {
+    get_ip_and_port(addr, addr_len, ip, port);
+  }
+}
+
+inline void get_remote_ip_and_port(socket_t sock, std::string &ip, int &port) {
+  struct sockaddr_storage addr;
+  socklen_t addr_len = sizeof(addr);
+
+  if (!getpeername(sock, reinterpret_cast<struct sockaddr *>(&addr),
+                   &addr_len)) {
+#ifndef _WIN32
+    if (addr.ss_family == AF_UNIX) {
+#if defined(__linux__)
+      struct ucred ucred;
+      socklen_t len = sizeof(ucred);
+      if (getsockopt(sock, SOL_SOCKET, SO_PEERCRED, &ucred, &len) == 0) {
+        port = ucred.pid;
+      }
+#elif defined(SOL_LOCAL) && defined(SO_PEERPID) // __APPLE__
+      pid_t pid;
+      socklen_t len = sizeof(pid);
+      if (getsockopt(sock, SOL_LOCAL, SO_PEERPID, &pid, &len) == 0) {
+        port = pid;
+      }
+#endif
+      return;
+    }
+#endif
+    get_ip_and_port(addr, addr_len, ip, port);
+  }
+}
+
+inline constexpr unsigned int str2tag_core(const char *s, size_t l,
+                                           unsigned int h) {
+  return (l == 0)
+             ? h
+             : str2tag_core(
+                   s + 1, l - 1,
+                   // Unsets the 6 high bits of h, therefore no overflow happens
+                   (((std::numeric_limits<unsigned int>::max)() >> 6) &
+                    h * 33) ^
+                       static_cast<unsigned char>(*s));
+}
+
+inline unsigned int str2tag(const std::string &s) {
+  return str2tag_core(s.data(), s.size(), 0);
+}
+
+namespace udl {
+
+inline constexpr unsigned int operator"" _t(const char *s, size_t l) {
+  return str2tag_core(s, l, 0);
+}
+
+} // namespace udl
+
+inline std::string
+find_content_type(const std::string &path,
+                  const std::map<std::string, std::string> &user_data,
+                  const std::string &default_content_type) {
+  auto ext = file_extension(path);
+
+  auto it = user_data.find(ext);
+  if (it != user_data.end()) { return it->second.c_str(); }
+
+  using udl::operator""_t;
+
+  switch (str2tag(ext)) {
+  default: return default_content_type;
+
+  case "css"_t: return "text/css";
+  case "csv"_t: return "text/csv";
+  case "htm"_t:
+  case "html"_t: return "text/html";
+  case "js"_t:
+  case "mjs"_t: return "text/javascript";
+  case "txt"_t: return "text/plain";
+  case "vtt"_t: return "text/vtt";
+
+  case "apng"_t: return "image/apng";
+  case "avif"_t: return "image/avif";
+  case "bmp"_t: return "image/bmp";
+  case "gif"_t: return "image/gif";
+  case "png"_t: return "image/png";
+  case "svg"_t: return "image/svg+xml";
+  case "webp"_t: return "image/webp";
+  case "ico"_t: return "image/x-icon";
+  case "tif"_t: return "image/tiff";
+  case "tiff"_t: return "image/tiff";
+  case "jpg"_t:
+  case "jpeg"_t: return "image/jpeg";
+
+  case "mp4"_t: return "video/mp4";
+  case "mpeg"_t: return "video/mpeg";
+  case "webm"_t: return "video/webm";
+
+  case "mp3"_t: return "audio/mp3";
+  case "mpga"_t: return "audio/mpeg";
+  case "weba"_t: return "audio/webm";
+  case "wav"_t: return "audio/wave";
+
+  case "otf"_t: return "font/otf";
+  case "ttf"_t: return "font/ttf";
+  case "woff"_t: return "font/woff";
+  case "woff2"_t: return "font/woff2";
+
+  case "7z"_t: return "application/x-7z-compressed";
+  case "atom"_t: return "application/atom+xml";
+  case "pdf"_t: return "application/pdf";
+  case "json"_t: return "application/json";
+  case "rss"_t: return "application/rss+xml";
+  case "tar"_t: return "application/x-tar";
+  case "xht"_t:
+  case "xhtml"_t: return "application/xhtml+xml";
+  case "xslt"_t: return "application/xslt+xml";
+  case "xml"_t: return "application/xml";
+  case "gz"_t: return "application/gzip";
+  case "zip"_t: return "application/zip";
+  case "wasm"_t: return "application/wasm";
+  }
+}
+
+inline bool can_compress_content_type(const std::string &content_type) {
+  using udl::operator""_t;
+
+  auto tag = str2tag(content_type);
+
+  switch (tag) {
+  case "image/svg+xml"_t:
+  case "application/javascript"_t:
+  case "application/json"_t:
+  case "application/xml"_t:
+  case "application/protobuf"_t:
+  case "application/xhtml+xml"_t: return true;
+
+  default:
+    return !content_type.rfind("text/", 0) && tag != "text/event-stream"_t;
+  }
+}
+
+inline EncodingType encoding_type(const Request &req, const Response &res) {
+  auto ret =
+      detail::can_compress_content_type(res.get_header_value("Content-Type"));
+  if (!ret) { return EncodingType::None; }
+
+  const auto &s = req.get_header_value("Accept-Encoding");
+  (void)(s);
+
+#ifdef CPPHTTPLIB_BROTLI_SUPPORT
+  // TODO: 'Accept-Encoding' has br, not br;q=0
+  ret = s.find("br") != std::string::npos;
+  if (ret) { return EncodingType::Brotli; }
+#endif
+
+#ifdef CPPHTTPLIB_ZLIB_SUPPORT
+  // TODO: 'Accept-Encoding' has gzip, not gzip;q=0
+  ret = s.find("gzip") != std::string::npos;
+  if (ret) { return EncodingType::Gzip; }
+#endif
+
+  return EncodingType::None;
+}
+
+inline bool nocompressor::compress(const char *data, size_t data_length,
+                                   bool /*last*/, Callback callback) {
+  if (!data_length) { return true; }
+  return callback(data, data_length);
+}
+
+#ifdef CPPHTTPLIB_ZLIB_SUPPORT
+inline gzip_compressor::gzip_compressor() {
+  std::memset(&strm_, 0, sizeof(strm_));
+  strm_.zalloc = Z_NULL;
+  strm_.zfree = Z_NULL;
+  strm_.opaque = Z_NULL;
+
+  is_valid_ = deflateInit2(&strm_, Z_DEFAULT_COMPRESSION, Z_DEFLATED, 31, 8,
+                           Z_DEFAULT_STRATEGY) == Z_OK;
+}
+
+inline gzip_compressor::~gzip_compressor() { deflateEnd(&strm_); }
+
+inline bool gzip_compressor::compress(const char *data, size_t data_length,
+                                      bool last, Callback callback) {
+  assert(is_valid_);
+
+  do {
+    constexpr size_t max_avail_in =
+        (std::numeric_limits<decltype(strm_.avail_in)>::max)();
+
+    strm_.avail_in = static_cast<decltype(strm_.avail_in)>(
+        (std::min)(data_length, max_avail_in));
+    strm_.next_in = const_cast<Bytef *>(reinterpret_cast<const Bytef *>(data));
+
+    data_length -= strm_.avail_in;
+    data += strm_.avail_in;
+
+    auto flush = (last && data_length == 0) ? Z_FINISH : Z_NO_FLUSH;
+    auto ret = Z_OK;
+
+    std::array<char, CPPHTTPLIB_COMPRESSION_BUFSIZ> buff{};
+    do {
+      strm_.avail_out = static_cast<uInt>(buff.size());
+      strm_.next_out = reinterpret_cast<Bytef *>(buff.data());
+
+      ret = deflate(&strm_, flush);
+      if (ret == Z_STREAM_ERROR) { return false; }
+
+      if (!callback(buff.data(), buff.size() - strm_.avail_out)) {
+        return false;
+      }
+    } while (strm_.avail_out == 0);
+
+    assert((flush == Z_FINISH && ret == Z_STREAM_END) ||
+           (flush == Z_NO_FLUSH && ret == Z_OK));
+    assert(strm_.avail_in == 0);
+  } while (data_length > 0);
+
+  return true;
+}
+
+inline gzip_decompressor::gzip_decompressor() {
+  std::memset(&strm_, 0, sizeof(strm_));
+  strm_.zalloc = Z_NULL;
+  strm_.zfree = Z_NULL;
+  strm_.opaque = Z_NULL;
+
+  // 15 is the value of wbits, which should be at the maximum possible value
+  // to ensure that any gzip stream can be decoded. The offset of 32 specifies
+  // that the stream type should be automatically detected either gzip or
+  // deflate.
+  is_valid_ = inflateInit2(&strm_, 32 + 15) == Z_OK;
+}
+
+inline gzip_decompressor::~gzip_decompressor() { inflateEnd(&strm_); }
+
+inline bool gzip_decompressor::is_valid() const { return is_valid_; }
+
+inline bool gzip_decompressor::decompress(const char *data, size_t data_length,
+                                          Callback callback) {
+  assert(is_valid_);
+
+  auto ret = Z_OK;
+
+  do {
+    constexpr size_t max_avail_in =
+        (std::numeric_limits<decltype(strm_.avail_in)>::max)();
+
+    strm_.avail_in = static_cast<decltype(strm_.avail_in)>(
+        (std::min)(data_length, max_avail_in));
+    strm_.next_in = const_cast<Bytef *>(reinterpret_cast<const Bytef *>(data));
+
+    data_length -= strm_.avail_in;
+    data += strm_.avail_in;
+
+    std::array<char, CPPHTTPLIB_COMPRESSION_BUFSIZ> buff{};
+    while (strm_.avail_in > 0 && ret == Z_OK) {
+      strm_.avail_out = static_cast<uInt>(buff.size());
+      strm_.next_out = reinterpret_cast<Bytef *>(buff.data());
+
+      ret = inflate(&strm_, Z_NO_FLUSH);
+
+      assert(ret != Z_STREAM_ERROR);
+      switch (ret) {
+      case Z_NEED_DICT:
+      case Z_DATA_ERROR:
+      case Z_MEM_ERROR: inflateEnd(&strm_); return false;
+      }
+
+      if (!callback(buff.data(), buff.size() - strm_.avail_out)) {
+        return false;
+      }
+    }
+
+    if (ret != Z_OK && ret != Z_STREAM_END) return false;
+
+  } while (data_length > 0);
+
+  return true;
+}
+#endif
+
+#ifdef CPPHTTPLIB_BROTLI_SUPPORT
+inline brotli_compressor::brotli_compressor() {
+  state_ = BrotliEncoderCreateInstance(nullptr, nullptr, nullptr);
+}
+
+inline brotli_compressor::~brotli_compressor() {
+  BrotliEncoderDestroyInstance(state_);
+}
+
+inline bool brotli_compressor::compress(const char *data, size_t data_length,
+                                        bool last, Callback callback) {
+  std::array<uint8_t, CPPHTTPLIB_COMPRESSION_BUFSIZ> buff{};
+
+  auto operation = last ? BROTLI_OPERATION_FINISH : BROTLI_OPERATION_PROCESS;
+  auto available_in = data_length;
+  auto next_in = reinterpret_cast<const uint8_t *>(data);
+
+  for (;;) {
+    if (last) {
+      if (BrotliEncoderIsFinished(state_)) { break; }
+    } else {
+      if (!available_in) { break; }
+    }
+
+    auto available_out = buff.size();
+    auto next_out = buff.data();
+
+    if (!BrotliEncoderCompressStream(state_, operation, &available_in, &next_in,
+                                     &available_out, &next_out, nullptr)) {
+      return false;
+    }
+
+    auto output_bytes = buff.size() - available_out;
+    if (output_bytes) {
+      callback(reinterpret_cast<const char *>(buff.data()), output_bytes);
+    }
+  }
+
+  return true;
+}
+
+inline brotli_decompressor::brotli_decompressor() {
+  decoder_s = BrotliDecoderCreateInstance(0, 0, 0);
+  decoder_r = decoder_s ? BROTLI_DECODER_RESULT_NEEDS_MORE_INPUT
+                        : BROTLI_DECODER_RESULT_ERROR;
+}
+
+inline brotli_decompressor::~brotli_decompressor() {
+  if (decoder_s) { BrotliDecoderDestroyInstance(decoder_s); }
+}
+
+inline bool brotli_decompressor::is_valid() const { return decoder_s; }
+
+inline bool brotli_decompressor::decompress(const char *data,
+                                            size_t data_length,
+                                            Callback callback) {
+  if (decoder_r == BROTLI_DECODER_RESULT_SUCCESS ||
+      decoder_r == BROTLI_DECODER_RESULT_ERROR) {
+    return 0;
+  }
+
+  auto next_in = reinterpret_cast<const uint8_t *>(data);
+  size_t avail_in = data_length;
+  size_t total_out;
+
+  decoder_r = BROTLI_DECODER_RESULT_NEEDS_MORE_OUTPUT;
+
+  std::array<char, CPPHTTPLIB_COMPRESSION_BUFSIZ> buff{};
+  while (decoder_r == BROTLI_DECODER_RESULT_NEEDS_MORE_OUTPUT) {
+    char *next_out = buff.data();
+    size_t avail_out = buff.size();
+
+    decoder_r = BrotliDecoderDecompressStream(
+        decoder_s, &avail_in, &next_in, &avail_out,
+        reinterpret_cast<uint8_t **>(&next_out), &total_out);
+
+    if (decoder_r == BROTLI_DECODER_RESULT_ERROR) { return false; }
+
+    if (!callback(buff.data(), buff.size() - avail_out)) { return false; }
+  }
+
+  return decoder_r == BROTLI_DECODER_RESULT_SUCCESS ||
+         decoder_r == BROTLI_DECODER_RESULT_NEEDS_MORE_INPUT;
+}
+#endif
+
+inline bool has_header(const Headers &headers, const std::string &key) {
+  return headers.find(key) != headers.end();
+}
+
+inline const char *get_header_value(const Headers &headers,
+                                    const std::string &key, size_t id,
+                                    const char *def) {
+  auto rng = headers.equal_range(key);
+  auto it = rng.first;
+  std::advance(it, static_cast<ssize_t>(id));
+  if (it != rng.second) { return it->second.c_str(); }
+  return def;
+}
+
+inline bool compare_case_ignore(const std::string &a, const std::string &b) {
+  if (a.size() != b.size()) { return false; }
+  for (size_t i = 0; i < b.size(); i++) {
+    if (::tolower(a[i]) != ::tolower(b[i])) { return false; }
+  }
+  return true;
+}
+
+template <typename T>
+inline bool parse_header(const char *beg, const char *end, T fn) {
+  // Skip trailing spaces and tabs.
+  while (beg < end && is_space_or_tab(end[-1])) {
+    end--;
+  }
+
+  auto p = beg;
+  while (p < end && *p != ':') {
+    p++;
+  }
+
+  if (p == end) { return false; }
+
+  auto key_end = p;
+
+  if (*p++ != ':') { return false; }
+
+  while (p < end && is_space_or_tab(*p)) {
+    p++;
+  }
+
+  if (p < end) {
+    auto key = std::string(beg, key_end);
+    auto val = compare_case_ignore(key, "Location")
+                   ? std::string(p, end)
+                   : decode_url(std::string(p, end), false);
+    fn(std::move(key), std::move(val));
+    return true;
+  }
+
+  return false;
+}
+
+inline bool read_headers(Stream &strm, Headers &headers) {
+  const auto bufsiz = 2048;
+  char buf[bufsiz];
+  stream_line_reader line_reader(strm, buf, bufsiz);
+
+  for (;;) {
+    if (!line_reader.getline()) { return false; }
+
+    // Check if the line ends with CRLF.
+    auto line_terminator_len = 2;
+    if (line_reader.end_with_crlf()) {
+      // Blank line indicates end of headers.
+      if (line_reader.size() == 2) { break; }
+#ifdef CPPHTTPLIB_ALLOW_LF_AS_LINE_TERMINATOR
+    } else {
+      // Blank line indicates end of headers.
+      if (line_reader.size() == 1) { break; }
+      line_terminator_len = 1;
+    }
+#else
+    } else {
+      continue; // Skip invalid line.
+    }
+#endif
+
+    if (line_reader.size() > CPPHTTPLIB_HEADER_MAX_LENGTH) { return false; }
+
+    // Exclude line terminator
+    auto end = line_reader.ptr() + line_reader.size() - line_terminator_len;
+
+    parse_header(line_reader.ptr(), end,
+                 [&](std::string &&key, std::string &&val) {
+                   headers.emplace(std::move(key), std::move(val));
+                 });
+  }
+
+  return true;
+}
+
+inline bool read_content_with_length(Stream &strm, uint64_t len,
+                                     Progress progress,
+                                     ContentReceiverWithProgress out) {
+  char buf[CPPHTTPLIB_RECV_BUFSIZ];
+
+  uint64_t r = 0;
+  while (r < len) {
+    auto read_len = static_cast<size_t>(len - r);
+    auto n = strm.read(buf, (std::min)(read_len, CPPHTTPLIB_RECV_BUFSIZ));
+    if (n <= 0) { return false; }
+
+    if (!out(buf, static_cast<size_t>(n), r, len)) { return false; }
+    r += static_cast<uint64_t>(n);
+
+    if (progress) {
+      if (!progress(r, len)) { return false; }
+    }
+  }
+
+  return true;
+}
+
+inline void skip_content_with_length(Stream &strm, uint64_t len) {
+  char buf[CPPHTTPLIB_RECV_BUFSIZ];
+  uint64_t r = 0;
+  while (r < len) {
+    auto read_len = static_cast<size_t>(len - r);
+    auto n = strm.read(buf, (std::min)(read_len, CPPHTTPLIB_RECV_BUFSIZ));
+    if (n <= 0) { return; }
+    r += static_cast<uint64_t>(n);
+  }
+}
+
+inline bool read_content_without_length(Stream &strm,
+                                        ContentReceiverWithProgress out) {
+  char buf[CPPHTTPLIB_RECV_BUFSIZ];
+  uint64_t r = 0;
+  for (;;) {
+    auto n = strm.read(buf, CPPHTTPLIB_RECV_BUFSIZ);
+    if (n < 0) {
+      return false;
+    } else if (n == 0) {
+      return true;
+    }
+
+    if (!out(buf, static_cast<size_t>(n), r, 0)) { return false; }
+    r += static_cast<uint64_t>(n);
+  }
+
+  return true;
+}
+
+template <typename T>
+inline bool read_content_chunked(Stream &strm, T &x,
+                                 ContentReceiverWithProgress out) {
+  const auto bufsiz = 16;
+  char buf[bufsiz];
+
+  stream_line_reader line_reader(strm, buf, bufsiz);
+
+  if (!line_reader.getline()) { return false; }
+
+  unsigned long chunk_len;
+  while (true) {
+    char *end_ptr;
+
+    chunk_len = std::strtoul(line_reader.ptr(), &end_ptr, 16);
+
+    if (end_ptr == line_reader.ptr()) { return false; }
+    if (chunk_len == ULONG_MAX) { return false; }
+
+    if (chunk_len == 0) { break; }
+
+    if (!read_content_with_length(strm, chunk_len, nullptr, out)) {
+      return false;
+    }
+
+    if (!line_reader.getline()) { return false; }
+
+    if (strcmp(line_reader.ptr(), "\r\n")) { return false; }
+
+    if (!line_reader.getline()) { return false; }
+  }
+
+  assert(chunk_len == 0);
+
+  // Trailer
+  if (!line_reader.getline()) { return false; }
+
+  while (strcmp(line_reader.ptr(), "\r\n")) {
+    if (line_reader.size() > CPPHTTPLIB_HEADER_MAX_LENGTH) { return false; }
+
+    // Exclude line terminator
+    constexpr auto line_terminator_len = 2;
+    auto end = line_reader.ptr() + line_reader.size() - line_terminator_len;
+
+    parse_header(line_reader.ptr(), end,
+                 [&](std::string &&key, std::string &&val) {
+                   x.headers.emplace(std::move(key), std::move(val));
+                 });
+
+    if (!line_reader.getline()) { return false; }
+  }
+
+  return true;
+}
+
+inline bool is_chunked_transfer_encoding(const Headers &headers) {
+  return !strcasecmp(get_header_value(headers, "Transfer-Encoding", 0, ""),
+                     "chunked");
+}
+
+template <typename T, typename U>
+bool prepare_content_receiver(T &x, int &status,
+                              ContentReceiverWithProgress receiver,
+                              bool decompress, U callback) {
+  if (decompress) {
+    std::string encoding = x.get_header_value("Content-Encoding");
+    std::unique_ptr<decompressor> decompressor;
+
+    if (encoding == "gzip" || encoding == "deflate") {
+#ifdef CPPHTTPLIB_ZLIB_SUPPORT
+      decompressor = detail::make_unique<gzip_decompressor>();
+#else
+      status = 415;
+      return false;
+#endif
+    } else if (encoding.find("br") != std::string::npos) {
+#ifdef CPPHTTPLIB_BROTLI_SUPPORT
+      decompressor = detail::make_unique<brotli_decompressor>();
+#else
+      status = 415;
+      return false;
+#endif
+    }
+
+    if (decompressor) {
+      if (decompressor->is_valid()) {
+        ContentReceiverWithProgress out = [&](const char *buf, size_t n,
+                                              uint64_t off, uint64_t len) {
+          return decompressor->decompress(buf, n,
+                                          [&](const char *buf2, size_t n2) {
+                                            return receiver(buf2, n2, off, len);
+                                          });
+        };
+        return callback(std::move(out));
+      } else {
+        status = 500;
+        return false;
+      }
+    }
+  }
+
+  ContentReceiverWithProgress out = [&](const char *buf, size_t n, uint64_t off,
+                                        uint64_t len) {
+    return receiver(buf, n, off, len);
+  };
+  return callback(std::move(out));
+}
+
+template <typename T>
+bool read_content(Stream &strm, T &x, size_t payload_max_length, int &status,
+                  Progress progress, ContentReceiverWithProgress receiver,
+                  bool decompress) {
+  return prepare_content_receiver(
+      x, status, std::move(receiver), decompress,
+      [&](const ContentReceiverWithProgress &out) {
+        auto ret = true;
+        auto exceed_payload_max_length = false;
+
+        if (is_chunked_transfer_encoding(x.headers)) {
+          ret = read_content_chunked(strm, x, out);
+        } else if (!has_header(x.headers, "Content-Length")) {
+          ret = read_content_without_length(strm, out);
+        } else {
+          auto len = get_header_value_u64(x.headers, "Content-Length", 0, 0);
+          if (len > payload_max_length) {
+            exceed_payload_max_length = true;
+            skip_content_with_length(strm, len);
+            ret = false;
+          } else if (len > 0) {
+            ret = read_content_with_length(strm, len, std::move(progress), out);
+          }
+        }
+
+        if (!ret) { status = exceed_payload_max_length ? 413 : 400; }
+        return ret;
+      });
+} // namespace detail
+
+inline ssize_t write_headers(Stream &strm, const Headers &headers) {
+  ssize_t write_len = 0;
+  for (const auto &x : headers) {
+    auto len =
+        strm.write_format("%s: %s\r\n", x.first.c_str(), x.second.c_str());
+    if (len < 0) { return len; }
+    write_len += len;
+  }
+  auto len = strm.write("\r\n");
+  if (len < 0) { return len; }
+  write_len += len;
+  return write_len;
+}
+
+inline bool write_data(Stream &strm, const char *d, size_t l) {
+  size_t offset = 0;
+  while (offset < l) {
+    auto length = strm.write(d + offset, l - offset);
+    if (length < 0) { return false; }
+    offset += static_cast<size_t>(length);
+  }
+  return true;
+}
+
+template <typename T>
+inline bool write_content(Stream &strm, const ContentProvider &content_provider,
+                          size_t offset, size_t length, T is_shutting_down,
+                          Error &error) {
+  size_t end_offset = offset + length;
+  auto ok = true;
+  DataSink data_sink;
+
+  data_sink.write = [&](const char *d, size_t l) -> bool {
+    if (ok) {
+      if (strm.is_writable() && write_data(strm, d, l)) {
+        offset += l;
+      } else {
+        ok = false;
+      }
+    }
+    return ok;
+  };
+
+  while (offset < end_offset && !is_shutting_down()) {
+    if (!strm.is_writable()) {
+      error = Error::Write;
+      return false;
+    } else if (!content_provider(offset, end_offset - offset, data_sink)) {
+      error = Error::Canceled;
+      return false;
+    } else if (!ok) {
+      error = Error::Write;
+      return false;
+    }
+  }
+
+  error = Error::Success;
+  return true;
+}
+
+template <typename T>
+inline bool write_content(Stream &strm, const ContentProvider &content_provider,
+                          size_t offset, size_t length,
+                          const T &is_shutting_down) {
+  auto error = Error::Success;
+  return write_content(strm, content_provider, offset, length, is_shutting_down,
+                       error);
+}
+
+template <typename T>
+inline bool
+write_content_without_length(Stream &strm,
+                             const ContentProvider &content_provider,
+                             const T &is_shutting_down) {
+  size_t offset = 0;
+  auto data_available = true;
+  auto ok = true;
+  DataSink data_sink;
+
+  data_sink.write = [&](const char *d, size_t l) -> bool {
+    if (ok) {
+      offset += l;
+      if (!strm.is_writable() || !write_data(strm, d, l)) { ok = false; }
+    }
+    return ok;
+  };
+
+  data_sink.done = [&](void) { data_available = false; };
+
+  while (data_available && !is_shutting_down()) {
+    if (!strm.is_writable()) {
+      return false;
+    } else if (!content_provider(offset, 0, data_sink)) {
+      return false;
+    } else if (!ok) {
+      return false;
+    }
+  }
+  return true;
+}
+
+template <typename T, typename U>
+inline bool
+write_content_chunked(Stream &strm, const ContentProvider &content_provider,
+                      const T &is_shutting_down, U &compressor, Error &error) {
+  size_t offset = 0;
+  auto data_available = true;
+  auto ok = true;
+  DataSink data_sink;
+
+  data_sink.write = [&](const char *d, size_t l) -> bool {
+    if (ok) {
+      data_available = l > 0;
+      offset += l;
+
+      std::string payload;
+      if (compressor.compress(d, l, false,
+                              [&](const char *data, size_t data_len) {
+                                payload.append(data, data_len);
+                                return true;
+                              })) {
+        if (!payload.empty()) {
+          // Emit chunked response header and footer for each chunk
+          auto chunk =
+              from_i_to_hex(payload.size()) + "\r\n" + payload + "\r\n";
+          if (!strm.is_writable() ||
+              !write_data(strm, chunk.data(), chunk.size())) {
+            ok = false;
+          }
+        }
+      } else {
+        ok = false;
+      }
+    }
+    return ok;
+  };
+
+  auto done_with_trailer = [&](const Headers *trailer) {
+    if (!ok) { return; }
+
+    data_available = false;
+
+    std::string payload;
+    if (!compressor.compress(nullptr, 0, true,
+                             [&](const char *data, size_t data_len) {
+                               payload.append(data, data_len);
+                               return true;
+                             })) {
+      ok = false;
+      return;
+    }
+
+    if (!payload.empty()) {
+      // Emit chunked response header and footer for each chunk
+      auto chunk = from_i_to_hex(payload.size()) + "\r\n" + payload + "\r\n";
+      if (!strm.is_writable() ||
+          !write_data(strm, chunk.data(), chunk.size())) {
+        ok = false;
+        return;
+      }
+    }
+
+    static const std::string done_marker("0\r\n");
+    if (!write_data(strm, done_marker.data(), done_marker.size())) {
+      ok = false;
+    }
+
+    // Trailer
+    if (trailer) {
+      for (const auto &kv : *trailer) {
+        std::string field_line = kv.first + ": " + kv.second + "\r\n";
+        if (!write_data(strm, field_line.data(), field_line.size())) {
+          ok = false;
+        }
+      }
+    }
+
+    static const std::string crlf("\r\n");
+    if (!write_data(strm, crlf.data(), crlf.size())) { ok = false; }
+  };
+
+  data_sink.done = [&](void) { done_with_trailer(nullptr); };
+
+  data_sink.done_with_trailer = [&](const Headers &trailer) {
+    done_with_trailer(&trailer);
+  };
+
+  while (data_available && !is_shutting_down()) {
+    if (!strm.is_writable()) {
+      error = Error::Write;
+      return false;
+    } else if (!content_provider(offset, 0, data_sink)) {
+      error = Error::Canceled;
+      return false;
+    } else if (!ok) {
+      error = Error::Write;
+      return false;
+    }
+  }
+
+  error = Error::Success;
+  return true;
+}
+
+template <typename T, typename U>
+inline bool write_content_chunked(Stream &strm,
+                                  const ContentProvider &content_provider,
+                                  const T &is_shutting_down, U &compressor) {
+  auto error = Error::Success;
+  return write_content_chunked(strm, content_provider, is_shutting_down,
+                               compressor, error);
+}
+
+template <typename T>
+inline bool redirect(T &cli, Request &req, Response &res,
+                     const std::string &path, const std::string &location,
+                     Error &error) {
+  Request new_req = req;
+  new_req.path = path;
+  new_req.redirect_count_ -= 1;
+
+  if (res.status == 303 && (req.method != "GET" && req.method != "HEAD")) {
+    new_req.method = "GET";
+    new_req.body.clear();
+    new_req.headers.clear();
+  }
+
+  Response new_res;
+
+  auto ret = cli.send(new_req, new_res, error);
+  if (ret) {
+    req = new_req;
+    res = new_res;
+
+    if (res.location.empty()) res.location = location;
+  }
+  return ret;
+}
+
+inline std::string params_to_query_str(const Params &params) {
+  std::string query;
+
+  for (auto it = params.begin(); it != params.end(); ++it) {
+    if (it != params.begin()) { query += "&"; }
+    query += it->first;
+    query += "=";
+    query += encode_query_param(it->second);
+  }
+  return query;
+}
+
+inline void parse_query_text(const std::string &s, Params &params) {
+  std::set<std::string> cache;
+  split(s.data(), s.data() + s.size(), '&', [&](const char *b, const char *e) {
+    std::string kv(b, e);
+    if (cache.find(kv) != cache.end()) { return; }
+    cache.insert(kv);
+
+    std::string key;
+    std::string val;
+    split(b, e, '=', [&](const char *b2, const char *e2) {
+      if (key.empty()) {
+        key.assign(b2, e2);
+      } else {
+        val.assign(b2, e2);
+      }
+    });
+
+    if (!key.empty()) {
+      params.emplace(decode_url(key, true), decode_url(val, true));
+    }
+  });
+}
+
+inline bool parse_multipart_boundary(const std::string &content_type,
+                                     std::string &boundary) {
+  auto boundary_keyword = "boundary=";
+  auto pos = content_type.find(boundary_keyword);
+  if (pos == std::string::npos) { return false; }
+  auto end = content_type.find(';', pos);
+  auto beg = pos + strlen(boundary_keyword);
+  boundary = trim_double_quotes_copy(content_type.substr(beg, end - beg));
+  return !boundary.empty();
+}
+
+inline void parse_disposition_params(const std::string &s, Params &params) {
+  std::set<std::string> cache;
+  split(s.data(), s.data() + s.size(), ';', [&](const char *b, const char *e) {
+    std::string kv(b, e);
+    if (cache.find(kv) != cache.end()) { return; }
+    cache.insert(kv);
+
+    std::string key;
+    std::string val;
+    split(b, e, '=', [&](const char *b2, const char *e2) {
+      if (key.empty()) {
+        key.assign(b2, e2);
+      } else {
+        val.assign(b2, e2);
+      }
+    });
+
+    if (!key.empty()) {
+      params.emplace(trim_double_quotes_copy((key)),
+                     trim_double_quotes_copy((val)));
+    }
+  });
+}
+
+#ifdef CPPHTTPLIB_NO_EXCEPTIONS
+inline bool parse_range_header(const std::string &s, Ranges &ranges) {
+#else
+inline bool parse_range_header(const std::string &s, Ranges &ranges) try {
+#endif
+  static auto re_first_range = std::regex(R"(bytes=(\d*-\d*(?:,\s*\d*-\d*)*))");
+  std::smatch m;
+  if (std::regex_match(s, m, re_first_range)) {
+    auto pos = static_cast<size_t>(m.position(1));
+    auto len = static_cast<size_t>(m.length(1));
+    auto all_valid_ranges = true;
+    split(&s[pos], &s[pos + len], ',', [&](const char *b, const char *e) {
+      if (!all_valid_ranges) return;
+      static auto re_another_range = std::regex(R"(\s*(\d*)-(\d*))");
+      std::cmatch cm;
+      if (std::regex_match(b, e, cm, re_another_range)) {
+        ssize_t first = -1;
+        if (!cm.str(1).empty()) {
+          first = static_cast<ssize_t>(std::stoll(cm.str(1)));
+        }
+
+        ssize_t last = -1;
+        if (!cm.str(2).empty()) {
+          last = static_cast<ssize_t>(std::stoll(cm.str(2)));
+        }
+
+        if (first != -1 && last != -1 && first > last) {
+          all_valid_ranges = false;
+          return;
+        }
+        ranges.emplace_back(std::make_pair(first, last));
+      }
+    });
+    return all_valid_ranges;
+  }
+  return false;
+#ifdef CPPHTTPLIB_NO_EXCEPTIONS
+}
+#else
+} catch (...) { return false; }
+#endif
+
+class MultipartFormDataParser {
+public:
+  MultipartFormDataParser() = default;
+
+  void set_boundary(std::string &&boundary) {
+    boundary_ = boundary;
+    dash_boundary_crlf_ = dash_ + boundary_ + crlf_;
+    crlf_dash_boundary_ = crlf_ + dash_ + boundary_;
+  }
+
+  bool is_valid() const { return is_valid_; }
+
+  bool parse(const char *buf, size_t n, const ContentReceiver &content_callback,
+             const MultipartContentHeader &header_callback) {
+
+    buf_append(buf, n);
+
+    while (buf_size() > 0) {
+      switch (state_) {
+      case 0: { // Initial boundary
+        buf_erase(buf_find(dash_boundary_crlf_));
+        if (dash_boundary_crlf_.size() > buf_size()) { return true; }
+        if (!buf_start_with(dash_boundary_crlf_)) { return false; }
+        buf_erase(dash_boundary_crlf_.size());
+        state_ = 1;
+        break;
+      }
+      case 1: { // New entry
+        clear_file_info();
+        state_ = 2;
+        break;
+      }
+      case 2: { // Headers
+        auto pos = buf_find(crlf_);
+        if (pos > CPPHTTPLIB_HEADER_MAX_LENGTH) { return false; }
+        while (pos < buf_size()) {
+          // Empty line
+          if (pos == 0) {
+            if (!header_callback(file_)) {
+              is_valid_ = false;
+              return false;
+            }
+            buf_erase(crlf_.size());
+            state_ = 3;
+            break;
+          }
+
+          static const std::string header_name = "content-type:";
+          const auto header = buf_head(pos);
+          if (start_with_case_ignore(header, header_name)) {
+            file_.content_type = trim_copy(header.substr(header_name.size()));
+          } else {
+            static const std::regex re_content_disposition(
+                R"~(^Content-Disposition:\s*form-data;\s*(.*)$)~",
+                std::regex_constants::icase);
+
+            std::smatch m;
+            if (std::regex_match(header, m, re_content_disposition)) {
+              Params params;
+              parse_disposition_params(m[1], params);
+
+              auto it = params.find("name");
+              if (it != params.end()) {
+                file_.name = it->second;
+              } else {
+                is_valid_ = false;
+                return false;
+              }
+
+              it = params.find("filename");
+              if (it != params.end()) { file_.filename = it->second; }
+
+              it = params.find("filename*");
+              if (it != params.end()) {
+                // Only allow UTF-8 enconnding...
+                static const std::regex re_rfc5987_encoding(
+                    R"~(^UTF-8''(.+?)$)~", std::regex_constants::icase);
+
+                std::smatch m2;
+                if (std::regex_match(it->second, m2, re_rfc5987_encoding)) {
+                  file_.filename = decode_url(m2[1], false); // override...
+                } else {
+                  is_valid_ = false;
+                  return false;
+                }
+              }
+            } else {
+              is_valid_ = false;
+              return false;
+            }
+          }
+          buf_erase(pos + crlf_.size());
+          pos = buf_find(crlf_);
+        }
+        if (state_ != 3) { return true; }
+        break;
+      }
+      case 3: { // Body
+        if (crlf_dash_boundary_.size() > buf_size()) { return true; }
+        auto pos = buf_find(crlf_dash_boundary_);
+        if (pos < buf_size()) {
+          if (!content_callback(buf_data(), pos)) {
+            is_valid_ = false;
+            return false;
+          }
+          buf_erase(pos + crlf_dash_boundary_.size());
+          state_ = 4;
+        } else {
+          auto len = buf_size() - crlf_dash_boundary_.size();
+          if (len > 0) {
+            if (!content_callback(buf_data(), len)) {
+              is_valid_ = false;
+              return false;
+            }
+            buf_erase(len);
+          }
+          return true;
+        }
+        break;
+      }
+      case 4: { // Boundary
+        if (crlf_.size() > buf_size()) { return true; }
+        if (buf_start_with(crlf_)) {
+          buf_erase(crlf_.size());
+          state_ = 1;
+        } else {
+          if (dash_.size() > buf_size()) { return true; }
+          if (buf_start_with(dash_)) {
+            buf_erase(dash_.size());
+            is_valid_ = true;
+            buf_erase(buf_size()); // Remove epilogue
+          } else {
+            return true;
+          }
+        }
+        break;
+      }
+      }
+    }
+
+    return true;
+  }
+
+private:
+  void clear_file_info() {
+    file_.name.clear();
+    file_.filename.clear();
+    file_.content_type.clear();
+  }
+
+  bool start_with_case_ignore(const std::string &a,
+                              const std::string &b) const {
+    if (a.size() < b.size()) { return false; }
+    for (size_t i = 0; i < b.size(); i++) {
+      if (::tolower(a[i]) != ::tolower(b[i])) { return false; }
+    }
+    return true;
+  }
+
+  const std::string dash_ = "--";
+  const std::string crlf_ = "\r\n";
+  std::string boundary_;
+  std::string dash_boundary_crlf_;
+  std::string crlf_dash_boundary_;
+
+  size_t state_ = 0;
+  bool is_valid_ = false;
+  MultipartFormData file_;
+
+  // Buffer
+  bool start_with(const std::string &a, size_t spos, size_t epos,
+                  const std::string &b) const {
+    if (epos - spos < b.size()) { return false; }
+    for (size_t i = 0; i < b.size(); i++) {
+      if (a[i + spos] != b[i]) { return false; }
+    }
+    return true;
+  }
+
+  size_t buf_size() const { return buf_epos_ - buf_spos_; }
+
+  const char *buf_data() const { return &buf_[buf_spos_]; }
+
+  std::string buf_head(size_t l) const { return buf_.substr(buf_spos_, l); }
+
+  bool buf_start_with(const std::string &s) const {
+    return start_with(buf_, buf_spos_, buf_epos_, s);
+  }
+
+  size_t buf_find(const std::string &s) const {
+    auto c = s.front();
+
+    size_t off = buf_spos_;
+    while (off < buf_epos_) {
+      auto pos = off;
+      while (true) {
+        if (pos == buf_epos_) { return buf_size(); }
+        if (buf_[pos] == c) { break; }
+        pos++;
+      }
+
+      auto remaining_size = buf_epos_ - pos;
+      if (s.size() > remaining_size) { return buf_size(); }
+
+      if (start_with(buf_, pos, buf_epos_, s)) { return pos - buf_spos_; }
+
+      off = pos + 1;
+    }
+
+    return buf_size();
+  }
+
+  void buf_append(const char *data, size_t n) {
+    auto remaining_size = buf_size();
+    if (remaining_size > 0 && buf_spos_ > 0) {
+      for (size_t i = 0; i < remaining_size; i++) {
+        buf_[i] = buf_[buf_spos_ + i];
+      }
+    }
+    buf_spos_ = 0;
+    buf_epos_ = remaining_size;
+
+    if (remaining_size + n > buf_.size()) { buf_.resize(remaining_size + n); }
+
+    for (size_t i = 0; i < n; i++) {
+      buf_[buf_epos_ + i] = data[i];
+    }
+    buf_epos_ += n;
+  }
+
+  void buf_erase(size_t size) { buf_spos_ += size; }
+
+  std::string buf_;
+  size_t buf_spos_ = 0;
+  size_t buf_epos_ = 0;
+};
+
+inline std::string to_lower(const char *beg, const char *end) {
+  std::string out;
+  auto it = beg;
+  while (it != end) {
+    out += static_cast<char>(::tolower(*it));
+    it++;
+  }
+  return out;
+}
+
+inline std::string make_multipart_data_boundary() {
+  static const char data[] =
+      "0123456789ABCDEFGHIJKLMNOPQRSTUVWXYZabcdefghijklmnopqrstuvwxyz";
+
+  // std::random_device might actually be deterministic on some
+  // platforms, but due to lack of support in the c++ standard library,
+  // doing better requires either some ugly hacks or breaking portability.
+  std::random_device seed_gen;
+
+  // Request 128 bits of entropy for initialization
+  std::seed_seq seed_sequence{seed_gen(), seed_gen(), seed_gen(), seed_gen()};
+  std::mt19937 engine(seed_sequence);
+
+  std::string result = "--cpp-httplib-multipart-data-";
+
+  for (auto i = 0; i < 16; i++) {
+    result += data[engine() % (sizeof(data) - 1)];
+  }
+
+  return result;
+}
+
+inline bool is_multipart_boundary_chars_valid(const std::string &boundary) {
+  auto valid = true;
+  for (size_t i = 0; i < boundary.size(); i++) {
+    auto c = boundary[i];
+    if (!std::isalnum(c) && c != '-' && c != '_') {
+      valid = false;
+      break;
+    }
+  }
+  return valid;
+}
+
+template <typename T>
+inline std::string
+serialize_multipart_formdata_item_begin(const T &item,
+                                        const std::string &boundary) {
+  std::string body = "--" + boundary + "\r\n";
+  body += "Content-Disposition: form-data; name=\"" + item.name + "\"";
+  if (!item.filename.empty()) {
+    body += "; filename=\"" + item.filename + "\"";
+  }
+  body += "\r\n";
+  if (!item.content_type.empty()) {
+    body += "Content-Type: " + item.content_type + "\r\n";
+  }
+  body += "\r\n";
+
+  return body;
+}
+
+inline std::string serialize_multipart_formdata_item_end() { return "\r\n"; }
+
+inline std::string
+serialize_multipart_formdata_finish(const std::string &boundary) {
+  return "--" + boundary + "--\r\n";
+}
+
+inline std::string
+serialize_multipart_formdata_get_content_type(const std::string &boundary) {
+  return "multipart/form-data; boundary=" + boundary;
+}
+
+inline std::string
+serialize_multipart_formdata(const MultipartFormDataItems &items,
+                             const std::string &boundary, bool finish = true) {
+  std::string body;
+
+  for (const auto &item : items) {
+    body += serialize_multipart_formdata_item_begin(item, boundary);
+    body += item.content + serialize_multipart_formdata_item_end();
+  }
+
+  if (finish) body += serialize_multipart_formdata_finish(boundary);
+
+  return body;
+}
+
+inline std::pair<size_t, size_t>
+get_range_offset_and_length(const Request &req, size_t content_length,
+                            size_t index) {
+  auto r = req.ranges[index];
+
+  if (r.first == -1 && r.second == -1) {
+    return std::make_pair(0, content_length);
+  }
+
+  auto slen = static_cast<ssize_t>(content_length);
+
+  if (r.first == -1) {
+    r.first = (std::max)(static_cast<ssize_t>(0), slen - r.second);
+    r.second = slen - 1;
+  }
+
+  if (r.second == -1) { r.second = slen - 1; }
+  return std::make_pair(r.first, static_cast<size_t>(r.second - r.first) + 1);
+}
+
+inline std::string
+make_content_range_header_field(const std::pair<ssize_t, ssize_t> &range,
+                                size_t content_length) {
+  std::string field = "bytes ";
+  if (range.first != -1) { field += std::to_string(range.first); }
+  field += "-";
+  if (range.second != -1) { field += std::to_string(range.second); }
+  field += "/";
+  field += std::to_string(content_length);
+  return field;
+}
+
+template <typename SToken, typename CToken, typename Content>
+bool process_multipart_ranges_data(const Request &req, Response &res,
+                                   const std::string &boundary,
+                                   const std::string &content_type,
+                                   SToken stoken, CToken ctoken,
+                                   Content content) {
+  for (size_t i = 0; i < req.ranges.size(); i++) {
+    ctoken("--");
+    stoken(boundary);
+    ctoken("\r\n");
+    if (!content_type.empty()) {
+      ctoken("Content-Type: ");
+      stoken(content_type);
+      ctoken("\r\n");
+    }
+
+    ctoken("Content-Range: ");
+    const auto &range = req.ranges[i];
+    stoken(make_content_range_header_field(range, res.content_length_));
+    ctoken("\r\n");
+    ctoken("\r\n");
+
+    auto offsets = get_range_offset_and_length(req, res.content_length_, i);
+    auto offset = offsets.first;
+    auto length = offsets.second;
+    if (!content(offset, length)) { return false; }
+    ctoken("\r\n");
+  }
+
+  ctoken("--");
+  stoken(boundary);
+  ctoken("--");
+
+  return true;
+}
+
+inline bool make_multipart_ranges_data(const Request &req, Response &res,
+                                       const std::string &boundary,
+                                       const std::string &content_type,
+                                       std::string &data) {
+  return process_multipart_ranges_data(
+      req, res, boundary, content_type,
+      [&](const std::string &token) { data += token; },
+      [&](const std::string &token) { data += token; },
+      [&](size_t offset, size_t length) {
+        if (offset < res.body.size()) {
+          data += res.body.substr(offset, length);
+          return true;
+        }
+        return false;
+      });
+}
+
+inline size_t
+get_multipart_ranges_data_length(const Request &req, Response &res,
+                                 const std::string &boundary,
+                                 const std::string &content_type) {
+  size_t data_length = 0;
+
+  process_multipart_ranges_data(
+      req, res, boundary, content_type,
+      [&](const std::string &token) { data_length += token.size(); },
+      [&](const std::string &token) { data_length += token.size(); },
+      [&](size_t /*offset*/, size_t length) {
+        data_length += length;
+        return true;
+      });
+
+  return data_length;
+}
+
+template <typename T>
+inline bool write_multipart_ranges_data(Stream &strm, const Request &req,
+                                        Response &res,
+                                        const std::string &boundary,
+                                        const std::string &content_type,
+                                        const T &is_shutting_down) {
+  return process_multipart_ranges_data(
+      req, res, boundary, content_type,
+      [&](const std::string &token) { strm.write(token); },
+      [&](const std::string &token) { strm.write(token); },
+      [&](size_t offset, size_t length) {
+        return write_content(strm, res.content_provider_, offset, length,
+                             is_shutting_down);
+      });
+}
+
+inline std::pair<size_t, size_t>
+get_range_offset_and_length(const Request &req, const Response &res,
+                            size_t index) {
+  auto r = req.ranges[index];
+
+  if (r.second == -1) {
+    r.second = static_cast<ssize_t>(res.content_length_) - 1;
+  }
+
+  return std::make_pair(r.first, r.second - r.first + 1);
+}
+
+inline bool expect_content(const Request &req) {
+  if (req.method == "POST" || req.method == "PUT" || req.method == "PATCH" ||
+      req.method == "PRI" || req.method == "DELETE") {
+    return true;
+  }
+  // TODO: check if Content-Length is set
+  return false;
+}
+
+inline bool has_crlf(const std::string &s) {
+  auto p = s.c_str();
+  while (*p) {
+    if (*p == '\r' || *p == '\n') { return true; }
+    p++;
+  }
+  return false;
+}
+
+#ifdef CPPHTTPLIB_OPENSSL_SUPPORT
+inline std::string message_digest(const std::string &s, const EVP_MD *algo) {
+  auto context = std::unique_ptr<EVP_MD_CTX, decltype(&EVP_MD_CTX_free)>(
+      EVP_MD_CTX_new(), EVP_MD_CTX_free);
+
+  unsigned int hash_length = 0;
+  unsigned char hash[EVP_MAX_MD_SIZE];
+
+  EVP_DigestInit_ex(context.get(), algo, nullptr);
+  EVP_DigestUpdate(context.get(), s.c_str(), s.size());
+  EVP_DigestFinal_ex(context.get(), hash, &hash_length);
+
+  std::stringstream ss;
+  for (auto i = 0u; i < hash_length; ++i) {
+    ss << std::hex << std::setw(2) << std::setfill('0')
+       << static_cast<unsigned int>(hash[i]);
+  }
+
+  return ss.str();
+}
+
+inline std::string MD5(const std::string &s) {
+  return message_digest(s, EVP_md5());
+}
+
+inline std::string SHA_256(const std::string &s) {
+  return message_digest(s, EVP_sha256());
+}
+
+inline std::string SHA_512(const std::string &s) {
+  return message_digest(s, EVP_sha512());
+}
+#endif
+
+#ifdef CPPHTTPLIB_OPENSSL_SUPPORT
+#ifdef _WIN32
+// NOTE: This code came up with the following stackoverflow post:
+// https://stackoverflow.com/questions/9507184/can-openssl-on-windows-use-the-system-certificate-store
+inline bool load_system_certs_on_windows(X509_STORE *store) {
+  auto hStore = CertOpenSystemStoreW((HCRYPTPROV_LEGACY)NULL, L"ROOT");
+  if (!hStore) { return false; }
+
+  auto result = false;
+  PCCERT_CONTEXT pContext = NULL;
+  while ((pContext = CertEnumCertificatesInStore(hStore, pContext)) !=
+         nullptr) {
+    auto encoded_cert =
+        static_cast<const unsigned char *>(pContext->pbCertEncoded);
+
+    auto x509 = d2i_X509(NULL, &encoded_cert, pContext->cbCertEncoded);
+    if (x509) {
+      X509_STORE_add_cert(store, x509);
+      X509_free(x509);
+      result = true;
+    }
+  }
+
+  CertFreeCertificateContext(pContext);
+  CertCloseStore(hStore, 0);
+
+  return result;
+}
+#elif defined(CPPHTTPLIB_USE_CERTS_FROM_MACOSX_KEYCHAIN) && defined(__APPLE__)
+#if TARGET_OS_OSX
+template <typename T>
+using CFObjectPtr =
+    std::unique_ptr<typename std::remove_pointer<T>::type, void (*)(CFTypeRef)>;
+
+inline void cf_object_ptr_deleter(CFTypeRef obj) {
+  if (obj) { CFRelease(obj); }
+}
+
+inline bool retrieve_certs_from_keychain(CFObjectPtr<CFArrayRef> &certs) {
+  CFStringRef keys[] = {kSecClass, kSecMatchLimit, kSecReturnRef};
+  CFTypeRef values[] = {kSecClassCertificate, kSecMatchLimitAll,
+                        kCFBooleanTrue};
+
+  CFObjectPtr<CFDictionaryRef> query(
+      CFDictionaryCreate(nullptr, reinterpret_cast<const void **>(keys), values,
+                         sizeof(keys) / sizeof(keys[0]),
+                         &kCFTypeDictionaryKeyCallBacks,
+                         &kCFTypeDictionaryValueCallBacks),
+      cf_object_ptr_deleter);
+
+  if (!query) { return false; }
+
+  CFTypeRef security_items = nullptr;
+  if (SecItemCopyMatching(query.get(), &security_items) != errSecSuccess ||
+      CFArrayGetTypeID() != CFGetTypeID(security_items)) {
+    return false;
+  }
+
+  certs.reset(reinterpret_cast<CFArrayRef>(security_items));
+  return true;
+}
+
+inline bool retrieve_root_certs_from_keychain(CFObjectPtr<CFArrayRef> &certs) {
+  CFArrayRef root_security_items = nullptr;
+  if (SecTrustCopyAnchorCertificates(&root_security_items) != errSecSuccess) {
+    return false;
+  }
+
+  certs.reset(root_security_items);
+  return true;
+}
+
+inline bool add_certs_to_x509_store(CFArrayRef certs, X509_STORE *store) {
+  auto result = false;
+  for (auto i = 0; i < CFArrayGetCount(certs); ++i) {
+    const auto cert = reinterpret_cast<const __SecCertificate *>(
+        CFArrayGetValueAtIndex(certs, i));
+
+    if (SecCertificateGetTypeID() != CFGetTypeID(cert)) { continue; }
+
+    CFDataRef cert_data = nullptr;
+    if (SecItemExport(cert, kSecFormatX509Cert, 0, nullptr, &cert_data) !=
+        errSecSuccess) {
+      continue;
+    }
+
+    CFObjectPtr<CFDataRef> cert_data_ptr(cert_data, cf_object_ptr_deleter);
+
+    auto encoded_cert = static_cast<const unsigned char *>(
+        CFDataGetBytePtr(cert_data_ptr.get()));
+
+    auto x509 =
+        d2i_X509(NULL, &encoded_cert, CFDataGetLength(cert_data_ptr.get()));
+
+    if (x509) {
+      X509_STORE_add_cert(store, x509);
+      X509_free(x509);
+      result = true;
+    }
+  }
+
+  return result;
+}
+
+inline bool load_system_certs_on_macos(X509_STORE *store) {
+  auto result = false;
+  CFObjectPtr<CFArrayRef> certs(nullptr, cf_object_ptr_deleter);
+  if (retrieve_certs_from_keychain(certs) && certs) {
+    result = add_certs_to_x509_store(certs.get(), store);
+  }
+
+  if (retrieve_root_certs_from_keychain(certs) && certs) {
+    result = add_certs_to_x509_store(certs.get(), store) || result;
+  }
+
+  return result;
+}
+#endif // TARGET_OS_OSX
+#endif // _WIN32
+#endif // CPPHTTPLIB_OPENSSL_SUPPORT
+
+#ifdef _WIN32
+class WSInit {
+public:
+  WSInit() {
+    WSADATA wsaData;
+    if (WSAStartup(0x0002, &wsaData) == 0) is_valid_ = true;
+  }
+
+  ~WSInit() {
+    if (is_valid_) WSACleanup();
+  }
+
+  bool is_valid_ = false;
+};
+
+static WSInit wsinit_;
+#endif
+
+#ifdef CPPHTTPLIB_OPENSSL_SUPPORT
+inline std::pair<std::string, std::string> make_digest_authentication_header(
+    const Request &req, const std::map<std::string, std::string> &auth,
+    size_t cnonce_count, const std::string &cnonce, const std::string &username,
+    const std::string &password, bool is_proxy = false) {
+  std::string nc;
+  {
+    std::stringstream ss;
+    ss << std::setfill('0') << std::setw(8) << std::hex << cnonce_count;
+    nc = ss.str();
+  }
+
+  std::string qop;
+  if (auth.find("qop") != auth.end()) {
+    qop = auth.at("qop");
+    if (qop.find("auth-int") != std::string::npos) {
+      qop = "auth-int";
+    } else if (qop.find("auth") != std::string::npos) {
+      qop = "auth";
+    } else {
+      qop.clear();
+    }
+  }
+
+  std::string algo = "MD5";
+  if (auth.find("algorithm") != auth.end()) { algo = auth.at("algorithm"); }
+
+  std::string response;
+  {
+    auto H = algo == "SHA-256"   ? detail::SHA_256
+             : algo == "SHA-512" ? detail::SHA_512
+                                 : detail::MD5;
+
+    auto A1 = username + ":" + auth.at("realm") + ":" + password;
+
+    auto A2 = req.method + ":" + req.path;
+    if (qop == "auth-int") { A2 += ":" + H(req.body); }
+
+    if (qop.empty()) {
+      response = H(H(A1) + ":" + auth.at("nonce") + ":" + H(A2));
+    } else {
+      response = H(H(A1) + ":" + auth.at("nonce") + ":" + nc + ":" + cnonce +
+                   ":" + qop + ":" + H(A2));
+    }
+  }
+
+  auto opaque = (auth.find("opaque") != auth.end()) ? auth.at("opaque") : "";
+
+  auto field = "Digest username=\"" + username + "\", realm=\"" +
+               auth.at("realm") + "\", nonce=\"" + auth.at("nonce") +
+               "\", uri=\"" + req.path + "\", algorithm=" + algo +
+               (qop.empty() ? ", response=\""
+                            : ", qop=" + qop + ", nc=" + nc + ", cnonce=\"" +
+                                  cnonce + "\", response=\"") +
+               response + "\"" +
+               (opaque.empty() ? "" : ", opaque=\"" + opaque + "\"");
+
+  auto key = is_proxy ? "Proxy-Authorization" : "Authorization";
+  return std::make_pair(key, field);
+}
+#endif
+
+inline bool parse_www_authenticate(const Response &res,
+                                   std::map<std::string, std::string> &auth,
+                                   bool is_proxy) {
+  auto auth_key = is_proxy ? "Proxy-Authenticate" : "WWW-Authenticate";
+  if (res.has_header(auth_key)) {
+    static auto re = std::regex(R"~((?:(?:,\s*)?(.+?)=(?:"(.*?)"|([^,]*))))~");
+    auto s = res.get_header_value(auth_key);
+    auto pos = s.find(' ');
+    if (pos != std::string::npos) {
+      auto type = s.substr(0, pos);
+      if (type == "Basic") {
+        return false;
+      } else if (type == "Digest") {
+        s = s.substr(pos + 1);
+        auto beg = std::sregex_iterator(s.begin(), s.end(), re);
+        for (auto i = beg; i != std::sregex_iterator(); ++i) {
+          const auto &m = *i;
+          auto key = s.substr(static_cast<size_t>(m.position(1)),
+                              static_cast<size_t>(m.length(1)));
+          auto val = m.length(2) > 0
+                         ? s.substr(static_cast<size_t>(m.position(2)),
+                                    static_cast<size_t>(m.length(2)))
+                         : s.substr(static_cast<size_t>(m.position(3)),
+                                    static_cast<size_t>(m.length(3)));
+          auth[key] = val;
+        }
+        return true;
+      }
+    }
+  }
+  return false;
+}
+
+// https://stackoverflow.com/questions/440133/how-do-i-create-a-random-alpha-numeric-string-in-c/440240#answer-440240
+inline std::string random_string(size_t length) {
+  auto randchar = []() -> char {
+    const char charset[] = "0123456789"
+                           "ABCDEFGHIJKLMNOPQRSTUVWXYZ"
+                           "abcdefghijklmnopqrstuvwxyz";
+    const size_t max_index = (sizeof(charset) - 1);
+    return charset[static_cast<size_t>(std::rand()) % max_index];
+  };
+  std::string str(length, 0);
+  std::generate_n(str.begin(), length, randchar);
+  return str;
+}
+
+class ContentProviderAdapter {
+public:
+  explicit ContentProviderAdapter(
+      ContentProviderWithoutLength &&content_provider)
+      : content_provider_(content_provider) {}
+
+  bool operator()(size_t offset, size_t, DataSink &sink) {
+    return content_provider_(offset, sink);
+  }
+
+private:
+  ContentProviderWithoutLength content_provider_;
+};
+
+} // namespace detail
+
+inline std::string hosted_at(const std::string &hostname) {
+  std::vector<std::string> addrs;
+  hosted_at(hostname, addrs);
+  if (addrs.empty()) { return std::string(); }
+  return addrs[0];
+}
+
+inline void hosted_at(const std::string &hostname,
+                      std::vector<std::string> &addrs) {
+  struct addrinfo hints;
+  struct addrinfo *result;
+
+  memset(&hints, 0, sizeof(struct addrinfo));
+  hints.ai_family = AF_UNSPEC;
+  hints.ai_socktype = SOCK_STREAM;
+  hints.ai_protocol = 0;
+
+  if (getaddrinfo(hostname.c_str(), nullptr, &hints, &result)) {
+#if defined __linux__ && !defined __ANDROID__
+    res_init();
+#endif
+    return;
+  }
+
+  for (auto rp = result; rp; rp = rp->ai_next) {
+    const auto &addr =
+        *reinterpret_cast<struct sockaddr_storage *>(rp->ai_addr);
+    std::string ip;
+    auto dummy = -1;
+    if (detail::get_ip_and_port(addr, sizeof(struct sockaddr_storage), ip,
+                                dummy)) {
+      addrs.push_back(ip);
+    }
+  }
+
+  freeaddrinfo(result);
+}
+
+inline std::string append_query_params(const std::string &path,
+                                       const Params &params) {
+  std::string path_with_query = path;
+  const static std::regex re("[^?]+\\?.*");
+  auto delm = std::regex_match(path, re) ? '&' : '?';
+  path_with_query += delm + detail::params_to_query_str(params);
+  return path_with_query;
+}
+
+// Header utilities
+inline std::pair<std::string, std::string> make_range_header(Ranges ranges) {
+  std::string field = "bytes=";
+  auto i = 0;
+  for (auto r : ranges) {
+    if (i != 0) { field += ", "; }
+    if (r.first != -1) { field += std::to_string(r.first); }
+    field += '-';
+    if (r.second != -1) { field += std::to_string(r.second); }
+    i++;
+  }
+  return std::make_pair("Range", std::move(field));
+}
+
+inline std::pair<std::string, std::string>
+make_basic_authentication_header(const std::string &username,
+                                 const std::string &password, bool is_proxy) {
+  auto field = "Basic " + detail::base64_encode(username + ":" + password);
+  auto key = is_proxy ? "Proxy-Authorization" : "Authorization";
+  return std::make_pair(key, std::move(field));
+}
+
+inline std::pair<std::string, std::string>
+make_bearer_token_authentication_header(const std::string &token,
+                                        bool is_proxy = false) {
+  auto field = "Bearer " + token;
+  auto key = is_proxy ? "Proxy-Authorization" : "Authorization";
+  return std::make_pair(key, std::move(field));
+}
+
+// Request implementation
+inline bool Request::has_header(const std::string &key) const {
+  return detail::has_header(headers, key);
+}
+
+inline std::string Request::get_header_value(const std::string &key,
+                                             size_t id) const {
+  return detail::get_header_value(headers, key, id, "");
+}
+
+inline size_t Request::get_header_value_count(const std::string &key) const {
+  auto r = headers.equal_range(key);
+  return static_cast<size_t>(std::distance(r.first, r.second));
+}
+
+inline void Request::set_header(const std::string &key,
+                                const std::string &val) {
+  if (!detail::has_crlf(key) && !detail::has_crlf(val)) {
+    headers.emplace(key, val);
+  }
+}
+
+inline bool Request::has_param(const std::string &key) const {
+  return params.find(key) != params.end();
+}
+
+inline std::string Request::get_param_value(const std::string &key,
+                                            size_t id) const {
+  auto rng = params.equal_range(key);
+  auto it = rng.first;
+  std::advance(it, static_cast<ssize_t>(id));
+  if (it != rng.second) { return it->second; }
+  return std::string();
+}
+
+inline size_t Request::get_param_value_count(const std::string &key) const {
+  auto r = params.equal_range(key);
+  return static_cast<size_t>(std::distance(r.first, r.second));
+}
+
+inline bool Request::is_multipart_form_data() const {
+  const auto &content_type = get_header_value("Content-Type");
+  return !content_type.rfind("multipart/form-data", 0);
+}
+
+inline bool Request::has_file(const std::string &key) const {
+  return files.find(key) != files.end();
+}
+
+inline MultipartFormData Request::get_file_value(const std::string &key) const {
+  auto it = files.find(key);
+  if (it != files.end()) { return it->second; }
+  return MultipartFormData();
+}
+
+inline std::vector<MultipartFormData>
+Request::get_file_values(const std::string &key) const {
+  std::vector<MultipartFormData> values;
+  auto rng = files.equal_range(key);
+  for (auto it = rng.first; it != rng.second; it++) {
+    values.push_back(it->second);
+  }
+  return values;
+}
+
+// Response implementation
+inline bool Response::has_header(const std::string &key) const {
+  return headers.find(key) != headers.end();
+}
+
+inline std::string Response::get_header_value(const std::string &key,
+                                              size_t id) const {
+  return detail::get_header_value(headers, key, id, "");
+}
+
+inline size_t Response::get_header_value_count(const std::string &key) const {
+  auto r = headers.equal_range(key);
+  return static_cast<size_t>(std::distance(r.first, r.second));
+}
+
+inline void Response::set_header(const std::string &key,
+                                 const std::string &val) {
+  if (!detail::has_crlf(key) && !detail::has_crlf(val)) {
+    headers.emplace(key, val);
+  }
+}
+
+inline void Response::set_redirect(const std::string &url, int stat) {
+  if (!detail::has_crlf(url)) {
+    set_header("Location", url);
+    if (300 <= stat && stat < 400) {
+      this->status = stat;
+    } else {
+      this->status = 302;
+    }
+  }
+}
+
+inline void Response::set_content(const char *s, size_t n,
+                                  const std::string &content_type) {
+  body.assign(s, n);
+
+  auto rng = headers.equal_range("Content-Type");
+  headers.erase(rng.first, rng.second);
+  set_header("Content-Type", content_type);
+}
+
+inline void Response::set_content(const std::string &s,
+                                  const std::string &content_type) {
+  set_content(s.data(), s.size(), content_type);
+}
+
+inline void Response::set_content_provider(
+    size_t in_length, const std::string &content_type, ContentProvider provider,
+    ContentProviderResourceReleaser resource_releaser) {
+  set_header("Content-Type", content_type);
+  content_length_ = in_length;
+  if (in_length > 0) { content_provider_ = std::move(provider); }
+  content_provider_resource_releaser_ = resource_releaser;
+  is_chunked_content_provider_ = false;
+}
+
+inline void Response::set_content_provider(
+    const std::string &content_type, ContentProviderWithoutLength provider,
+    ContentProviderResourceReleaser resource_releaser) {
+  set_header("Content-Type", content_type);
+  content_length_ = 0;
+  content_provider_ = detail::ContentProviderAdapter(std::move(provider));
+  content_provider_resource_releaser_ = resource_releaser;
+  is_chunked_content_provider_ = false;
+}
+
+inline void Response::set_chunked_content_provider(
+    const std::string &content_type, ContentProviderWithoutLength provider,
+    ContentProviderResourceReleaser resource_releaser) {
+  set_header("Content-Type", content_type);
+  content_length_ = 0;
+  content_provider_ = detail::ContentProviderAdapter(std::move(provider));
+  content_provider_resource_releaser_ = resource_releaser;
+  is_chunked_content_provider_ = true;
+}
+
+// Result implementation
+inline bool Result::has_request_header(const std::string &key) const {
+  return request_headers_.find(key) != request_headers_.end();
+}
+
+inline std::string Result::get_request_header_value(const std::string &key,
+                                                    size_t id) const {
+  return detail::get_header_value(request_headers_, key, id, "");
+}
+
+inline size_t
+Result::get_request_header_value_count(const std::string &key) const {
+  auto r = request_headers_.equal_range(key);
+  return static_cast<size_t>(std::distance(r.first, r.second));
+}
+
+// Stream implementation
+inline ssize_t Stream::write(const char *ptr) {
+  return write(ptr, strlen(ptr));
+}
+
+inline ssize_t Stream::write(const std::string &s) {
+  return write(s.data(), s.size());
+}
+
+namespace detail {
+
+// Socket stream implementation
+inline SocketStream::SocketStream(socket_t sock, time_t read_timeout_sec,
+                                  time_t read_timeout_usec,
+                                  time_t write_timeout_sec,
+                                  time_t write_timeout_usec)
+    : sock_(sock), read_timeout_sec_(read_timeout_sec),
+      read_timeout_usec_(read_timeout_usec),
+      write_timeout_sec_(write_timeout_sec),
+      write_timeout_usec_(write_timeout_usec), read_buff_(read_buff_size_, 0) {}
+
+inline SocketStream::~SocketStream() {}
+
+inline bool SocketStream::is_readable() const {
+  return select_read(sock_, read_timeout_sec_, read_timeout_usec_) > 0;
+}
+
+inline bool SocketStream::is_writable() const {
+  return select_write(sock_, write_timeout_sec_, write_timeout_usec_) > 0 &&
+         is_socket_alive(sock_);
+}
+
+inline ssize_t SocketStream::read(char *ptr, size_t size) {
+#ifdef _WIN32
+  size =
+      (std::min)(size, static_cast<size_t>((std::numeric_limits<int>::max)()));
+#else
+  size = (std::min)(size,
+                    static_cast<size_t>((std::numeric_limits<ssize_t>::max)()));
+#endif
+
+  if (read_buff_off_ < read_buff_content_size_) {
+    auto remaining_size = read_buff_content_size_ - read_buff_off_;
+    if (size <= remaining_size) {
+      memcpy(ptr, read_buff_.data() + read_buff_off_, size);
+      read_buff_off_ += size;
+      return static_cast<ssize_t>(size);
+    } else {
+      memcpy(ptr, read_buff_.data() + read_buff_off_, remaining_size);
+      read_buff_off_ += remaining_size;
+      return static_cast<ssize_t>(remaining_size);
+    }
+  }
+
+  if (!is_readable()) { return -1; }
+
+  read_buff_off_ = 0;
+  read_buff_content_size_ = 0;
+
+  if (size < read_buff_size_) {
+    auto n = read_socket(sock_, read_buff_.data(), read_buff_size_,
+                         CPPHTTPLIB_RECV_FLAGS);
+    if (n <= 0) {
+      return n;
+    } else if (n <= static_cast<ssize_t>(size)) {
+      memcpy(ptr, read_buff_.data(), static_cast<size_t>(n));
+      return n;
+    } else {
+      memcpy(ptr, read_buff_.data(), size);
+      read_buff_off_ = size;
+      read_buff_content_size_ = static_cast<size_t>(n);
+      return static_cast<ssize_t>(size);
+    }
+  } else {
+    return read_socket(sock_, ptr, size, CPPHTTPLIB_RECV_FLAGS);
+  }
+}
+
+inline ssize_t SocketStream::write(const char *ptr, size_t size) {
+  if (!is_writable()) { return -1; }
+
+#if defined(_WIN32) && !defined(_WIN64)
+  size =
+      (std::min)(size, static_cast<size_t>((std::numeric_limits<int>::max)()));
+#endif
+
+  return send_socket(sock_, ptr, size, CPPHTTPLIB_SEND_FLAGS);
+}
+
+inline void SocketStream::get_remote_ip_and_port(std::string &ip,
+                                                 int &port) const {
+  return detail::get_remote_ip_and_port(sock_, ip, port);
+}
+
+inline void SocketStream::get_local_ip_and_port(std::string &ip,
+                                                int &port) const {
+  return detail::get_local_ip_and_port(sock_, ip, port);
+}
+
+inline socket_t SocketStream::socket() const { return sock_; }
+
+// Buffer stream implementation
+inline bool BufferStream::is_readable() const { return true; }
+
+inline bool BufferStream::is_writable() const { return true; }
+
+inline ssize_t BufferStream::read(char *ptr, size_t size) {
+#if defined(_MSC_VER) && _MSC_VER < 1910
+  auto len_read = buffer._Copy_s(ptr, size, size, position);
+#else
+  auto len_read = buffer.copy(ptr, size, position);
+#endif
+  position += static_cast<size_t>(len_read);
+  return static_cast<ssize_t>(len_read);
+}
+
+inline ssize_t BufferStream::write(const char *ptr, size_t size) {
+  buffer.append(ptr, size);
+  return static_cast<ssize_t>(size);
+}
+
+inline void BufferStream::get_remote_ip_and_port(std::string & /*ip*/,
+                                                 int & /*port*/) const {}
+
+inline void BufferStream::get_local_ip_and_port(std::string & /*ip*/,
+                                                int & /*port*/) const {}
+
+inline socket_t BufferStream::socket() const { return 0; }
+
+inline const std::string &BufferStream::get_buffer() const { return buffer; }
+
+inline PathParamsMatcher::PathParamsMatcher(const std::string &pattern) {
+  // One past the last ending position of a path param substring
+  std::size_t last_param_end = 0;
+
+#ifndef CPPHTTPLIB_NO_EXCEPTIONS
+  // Needed to ensure that parameter names are unique during matcher
+  // construction
+  // If exceptions are disabled, only last duplicate path
+  // parameter will be set
+  std::unordered_set<std::string> param_name_set;
+#endif
+
+  while (true) {
+    const auto marker_pos = pattern.find(marker, last_param_end);
+    if (marker_pos == std::string::npos) { break; }
+
+    static_fragments_.push_back(
+        pattern.substr(last_param_end, marker_pos - last_param_end));
+
+    const auto param_name_start = marker_pos + 1;
+
+    auto sep_pos = pattern.find(separator, param_name_start);
+    if (sep_pos == std::string::npos) { sep_pos = pattern.length(); }
+
+    auto param_name =
+        pattern.substr(param_name_start, sep_pos - param_name_start);
+
+#ifndef CPPHTTPLIB_NO_EXCEPTIONS
+    if (param_name_set.find(param_name) != param_name_set.cend()) {
+      std::string msg = "Encountered path parameter '" + param_name +
+                        "' multiple times in route pattern '" + pattern + "'.";
+      throw std::invalid_argument(msg);
+    }
+#endif
+
+    param_names_.push_back(std::move(param_name));
+
+    last_param_end = sep_pos + 1;
+  }
+
+  if (last_param_end < pattern.length()) {
+    static_fragments_.push_back(pattern.substr(last_param_end));
+  }
+}
+
+inline bool PathParamsMatcher::match(Request &request) const {
+  request.matches = std::smatch();
+  request.path_params.clear();
+  request.path_params.reserve(param_names_.size());
+
+  // One past the position at which the path matched the pattern last time
+  std::size_t starting_pos = 0;
+  for (size_t i = 0; i < static_fragments_.size(); ++i) {
+    const auto &fragment = static_fragments_[i];
+
+    if (starting_pos + fragment.length() > request.path.length()) {
+      return false;
+    }
+
+    // Avoid unnecessary allocation by using strncmp instead of substr +
+    // comparison
+    if (std::strncmp(request.path.c_str() + starting_pos, fragment.c_str(),
+                     fragment.length()) != 0) {
+      return false;
+    }
+
+    starting_pos += fragment.length();
+
+    // Should only happen when we have a static fragment after a param
+    // Example: '/users/:id/subscriptions'
+    // The 'subscriptions' fragment here does not have a corresponding param
+    if (i >= param_names_.size()) { continue; }
+
+    auto sep_pos = request.path.find(separator, starting_pos);
+    if (sep_pos == std::string::npos) { sep_pos = request.path.length(); }
+
+    const auto &param_name = param_names_[i];
+
+    request.path_params.emplace(
+        param_name, request.path.substr(starting_pos, sep_pos - starting_pos));
+
+    // Mark everythin up to '/' as matched
+    starting_pos = sep_pos + 1;
+  }
+  // Returns false if the path is longer than the pattern
+  return starting_pos >= request.path.length();
+}
+
+inline bool RegexMatcher::match(Request &request) const {
+  request.path_params.clear();
+  return std::regex_match(request.path, request.matches, regex_);
+}
+
+} // namespace detail
+
+// HTTP server implementation
+inline Server::Server()
+    : new_task_queue(
+          [] { return new ThreadPool(CPPHTTPLIB_THREAD_POOL_COUNT); }) {
+#ifndef _WIN32
+  signal(SIGPIPE, SIG_IGN);
+#endif
+}
+
+inline Server::~Server() {}
+
+inline std::unique_ptr<detail::MatcherBase>
+Server::make_matcher(const std::string &pattern) {
+  if (pattern.find("/:") != std::string::npos) {
+    return detail::make_unique<detail::PathParamsMatcher>(pattern);
+  } else {
+    return detail::make_unique<detail::RegexMatcher>(pattern);
+  }
+}
+
+inline Server &Server::Get(const std::string &pattern, Handler handler) {
+  get_handlers_.push_back(
+      std::make_pair(make_matcher(pattern), std::move(handler)));
+  return *this;
+}
+
+inline Server &Server::Post(const std::string &pattern, Handler handler) {
+  post_handlers_.push_back(
+      std::make_pair(make_matcher(pattern), std::move(handler)));
+  return *this;
+}
+
+inline Server &Server::Post(const std::string &pattern,
+                            HandlerWithContentReader handler) {
+  post_handlers_for_content_reader_.push_back(
+      std::make_pair(make_matcher(pattern), std::move(handler)));
+  return *this;
+}
+
+inline Server &Server::Put(const std::string &pattern, Handler handler) {
+  put_handlers_.push_back(
+      std::make_pair(make_matcher(pattern), std::move(handler)));
+  return *this;
+}
+
+inline Server &Server::Put(const std::string &pattern,
+                           HandlerWithContentReader handler) {
+  put_handlers_for_content_reader_.push_back(
+      std::make_pair(make_matcher(pattern), std::move(handler)));
+  return *this;
+}
+
+inline Server &Server::Patch(const std::string &pattern, Handler handler) {
+  patch_handlers_.push_back(
+      std::make_pair(make_matcher(pattern), std::move(handler)));
+  return *this;
+}
+
+inline Server &Server::Patch(const std::string &pattern,
+                             HandlerWithContentReader handler) {
+  patch_handlers_for_content_reader_.push_back(
+      std::make_pair(make_matcher(pattern), std::move(handler)));
+  return *this;
+}
+
+inline Server &Server::Delete(const std::string &pattern, Handler handler) {
+  delete_handlers_.push_back(
+      std::make_pair(make_matcher(pattern), std::move(handler)));
+  return *this;
+}
+
+inline Server &Server::Delete(const std::string &pattern,
+                              HandlerWithContentReader handler) {
+  delete_handlers_for_content_reader_.push_back(
+      std::make_pair(make_matcher(pattern), std::move(handler)));
+  return *this;
+}
+
+inline Server &Server::Options(const std::string &pattern, Handler handler) {
+  options_handlers_.push_back(
+      std::make_pair(make_matcher(pattern), std::move(handler)));
+  return *this;
+}
+
+inline bool Server::set_base_dir(const std::string &dir,
+                                 const std::string &mount_point) {
+  return set_mount_point(mount_point, dir);
+}
+
+inline bool Server::set_mount_point(const std::string &mount_point,
+                                    const std::string &dir, Headers headers) {
+  if (detail::is_dir(dir)) {
+    std::string mnt = !mount_point.empty() ? mount_point : "/";
+    if (!mnt.empty() && mnt[0] == '/') {
+      base_dirs_.push_back({mnt, dir, std::move(headers)});
+      return true;
+    }
+  }
+  return false;
+}
+
+inline bool Server::remove_mount_point(const std::string &mount_point) {
+  for (auto it = base_dirs_.begin(); it != base_dirs_.end(); ++it) {
+    if (it->mount_point == mount_point) {
+      base_dirs_.erase(it);
+      return true;
+    }
+  }
+  return false;
+}
+
+inline Server &
+Server::set_file_extension_and_mimetype_mapping(const std::string &ext,
+                                                const std::string &mime) {
+  file_extension_and_mimetype_map_[ext] = mime;
+  return *this;
+}
+
+inline Server &Server::set_default_file_mimetype(const std::string &mime) {
+  default_file_mimetype_ = mime;
+  return *this;
+}
+
+inline Server &Server::set_file_request_handler(Handler handler) {
+  file_request_handler_ = std::move(handler);
+  return *this;
+}
+
+inline Server &Server::set_error_handler(HandlerWithResponse handler) {
+  error_handler_ = std::move(handler);
+  return *this;
+}
+
+inline Server &Server::set_error_handler(Handler handler) {
+  error_handler_ = [handler](const Request &req, Response &res) {
+    handler(req, res);
+    return HandlerResponse::Handled;
+  };
+  return *this;
+}
+
+inline Server &Server::set_exception_handler(ExceptionHandler handler) {
+  exception_handler_ = std::move(handler);
+  return *this;
+}
+
+inline Server &Server::set_pre_routing_handler(HandlerWithResponse handler) {
+  pre_routing_handler_ = std::move(handler);
+  return *this;
+}
+
+inline Server &Server::set_post_routing_handler(Handler handler) {
+  post_routing_handler_ = std::move(handler);
+  return *this;
+}
+
+inline Server &Server::set_logger(Logger logger) {
+  logger_ = std::move(logger);
+  return *this;
+}
+
+inline Server &
+Server::set_expect_100_continue_handler(Expect100ContinueHandler handler) {
+  expect_100_continue_handler_ = std::move(handler);
+  return *this;
+}
+
+inline Server &Server::set_address_family(int family) {
+  address_family_ = family;
+  return *this;
+}
+
+inline Server &Server::set_tcp_nodelay(bool on) {
+  tcp_nodelay_ = on;
+  return *this;
+}
+
+inline Server &Server::set_socket_options(SocketOptions socket_options) {
+  socket_options_ = std::move(socket_options);
+  return *this;
+}
+
+inline Server &Server::set_default_headers(Headers headers) {
+  default_headers_ = std::move(headers);
+  return *this;
+}
+
+inline Server &Server::set_header_writer(
+    std::function<ssize_t(Stream &, Headers &)> const &writer) {
+  header_writer_ = writer;
+  return *this;
+}
+
+inline Server &Server::set_keep_alive_max_count(size_t count) {
+  keep_alive_max_count_ = count;
+  return *this;
+}
+
+inline Server &Server::set_keep_alive_timeout(time_t sec) {
+  keep_alive_timeout_sec_ = sec;
+  return *this;
+}
+
+inline Server &Server::set_read_timeout(time_t sec, time_t usec) {
+  read_timeout_sec_ = sec;
+  read_timeout_usec_ = usec;
+  return *this;
+}
+
+inline Server &Server::set_write_timeout(time_t sec, time_t usec) {
+  write_timeout_sec_ = sec;
+  write_timeout_usec_ = usec;
+  return *this;
+}
+
+inline Server &Server::set_idle_interval(time_t sec, time_t usec) {
+  idle_interval_sec_ = sec;
+  idle_interval_usec_ = usec;
+  return *this;
+}
+
+inline Server &Server::set_payload_max_length(size_t length) {
+  payload_max_length_ = length;
+  return *this;
+}
+
+inline bool Server::bind_to_port(const std::string &host, int port,
+                                 int socket_flags) {
+  if (bind_internal(host, port, socket_flags) < 0) return false;
+  return true;
+}
+inline int Server::bind_to_any_port(const std::string &host, int socket_flags) {
+  return bind_internal(host, 0, socket_flags);
+}
+
+inline bool Server::listen_after_bind() {
+  auto se = detail::scope_exit([&]() { done_ = true; });
+  return listen_internal();
+}
+
+inline bool Server::listen(const std::string &host, int port,
+                           int socket_flags) {
+  auto se = detail::scope_exit([&]() { done_ = true; });
+  return bind_to_port(host, port, socket_flags) && listen_internal();
+}
+
+inline bool Server::is_running() const { return is_running_; }
+
+inline void Server::wait_until_ready() const {
+  while (!is_running() && !done_) {
+    std::this_thread::sleep_for(std::chrono::milliseconds{1});
+  }
+}
+
+inline void Server::stop() {
+  if (is_running_) {
+    assert(svr_sock_ != INVALID_SOCKET);
+    std::atomic<socket_t> sock(svr_sock_.exchange(INVALID_SOCKET));
+    detail::shutdown_socket(sock);
+    detail::close_socket(sock);
+  }
+}
+
+inline bool Server::parse_request_line(const char *s, Request &req) {
+  auto len = strlen(s);
+  if (len < 2 || s[len - 2] != '\r' || s[len - 1] != '\n') { return false; }
+  len -= 2;
+
+  {
+    size_t count = 0;
+
+    detail::split(s, s + len, ' ', [&](const char *b, const char *e) {
+      switch (count) {
+      case 0: req.method = std::string(b, e); break;
+      case 1: req.target = std::string(b, e); break;
+      case 2: req.version = std::string(b, e); break;
+      default: break;
+      }
+      count++;
+    });
+
+    if (count != 3) { return false; }
+  }
+
+  static const std::set<std::string> methods{
+      "GET",     "HEAD",    "POST",  "PUT",   "DELETE",
+      "CONNECT", "OPTIONS", "TRACE", "PATCH", "PRI"};
+
+  if (methods.find(req.method) == methods.end()) { return false; }
+
+  if (req.version != "HTTP/1.1" && req.version != "HTTP/1.0") { return false; }
+
+  {
+    // Skip URL fragment
+    for (size_t i = 0; i < req.target.size(); i++) {
+      if (req.target[i] == '#') {
+        req.target.erase(i);
+        break;
+      }
+    }
+
+    size_t count = 0;
+
+    detail::split(req.target.data(), req.target.data() + req.target.size(), '?',
+                  [&](const char *b, const char *e) {
+                    switch (count) {
+                    case 0:
+                      req.path = detail::decode_url(std::string(b, e), false);
+                      break;
+                    case 1: {
+                      if (e - b > 0) {
+                        detail::parse_query_text(std::string(b, e), req.params);
+                      }
+                      break;
+                    }
+                    default: break;
+                    }
+                    count++;
+                  });
+
+    if (count > 2) { return false; }
+  }
+
+  return true;
+}
+
+inline bool Server::write_response(Stream &strm, bool close_connection,
+                                   const Request &req, Response &res) {
+  return write_response_core(strm, close_connection, req, res, false);
+}
+
+inline bool Server::write_response_with_content(Stream &strm,
+                                                bool close_connection,
+                                                const Request &req,
+                                                Response &res) {
+  return write_response_core(strm, close_connection, req, res, true);
+}
+
+inline bool Server::write_response_core(Stream &strm, bool close_connection,
+                                        const Request &req, Response &res,
+                                        bool need_apply_ranges) {
+  assert(res.status != -1);
+
+  if (400 <= res.status && error_handler_ &&
+      error_handler_(req, res) == HandlerResponse::Handled) {
+    need_apply_ranges = true;
+  }
+
+  std::string content_type;
+  std::string boundary;
+  if (need_apply_ranges) { apply_ranges(req, res, content_type, boundary); }
+
+  // Prepare additional headers
+  if (close_connection || req.get_header_value("Connection") == "close") {
+    res.set_header("Connection", "close");
+  } else {
+    std::stringstream ss;
+    ss << "timeout=" << keep_alive_timeout_sec_
+       << ", max=" << keep_alive_max_count_;
+    res.set_header("Keep-Alive", ss.str());
+  }
+
+  if (!res.has_header("Content-Type") &&
+      (!res.body.empty() || res.content_length_ > 0 || res.content_provider_)) {
+    res.set_header("Content-Type", "text/plain");
+  }
+
+  if (!res.has_header("Content-Length") && res.body.empty() &&
+      !res.content_length_ && !res.content_provider_) {
+    res.set_header("Content-Length", "0");
+  }
+
+  if (!res.has_header("Accept-Ranges") && req.method == "HEAD") {
+    res.set_header("Accept-Ranges", "bytes");
+  }
+
+  if (post_routing_handler_) { post_routing_handler_(req, res); }
+
+  // Response line and headers
+  {
+    detail::BufferStream bstrm;
+
+    if (!bstrm.write_format("HTTP/1.1 %d %s\r\n", res.status,
+                            status_message(res.status))) {
+      return false;
+    }
+
+    if (!header_writer_(bstrm, res.headers)) { return false; }
+
+    // Flush buffer
+    auto &data = bstrm.get_buffer();
+    detail::write_data(strm, data.data(), data.size());
+  }
+
+  // Body
+  auto ret = true;
+  if (req.method != "HEAD") {
+    if (!res.body.empty()) {
+      if (!detail::write_data(strm, res.body.data(), res.body.size())) {
+        ret = false;
+      }
+    } else if (res.content_provider_) {
+      if (write_content_with_provider(strm, req, res, boundary, content_type)) {
+        res.content_provider_success_ = true;
+      } else {
+        res.content_provider_success_ = false;
+        ret = false;
+      }
+    }
+  }
+
+  // Log
+  if (logger_) { logger_(req, res); }
+
+  return ret;
+}
+
+inline bool
+Server::write_content_with_provider(Stream &strm, const Request &req,
+                                    Response &res, const std::string &boundary,
+                                    const std::string &content_type) {
+  auto is_shutting_down = [this]() {
+    return this->svr_sock_ == INVALID_SOCKET;
+  };
+
+  if (res.content_length_ > 0) {
+    if (req.ranges.empty()) {
+      return detail::write_content(strm, res.content_provider_, 0,
+                                   res.content_length_, is_shutting_down);
+    } else if (req.ranges.size() == 1) {
+      auto offsets =
+          detail::get_range_offset_and_length(req, res.content_length_, 0);
+      auto offset = offsets.first;
+      auto length = offsets.second;
+      return detail::write_content(strm, res.content_provider_, offset, length,
+                                   is_shutting_down);
+    } else {
+      return detail::write_multipart_ranges_data(
+          strm, req, res, boundary, content_type, is_shutting_down);
+    }
+  } else {
+    if (res.is_chunked_content_provider_) {
+      auto type = detail::encoding_type(req, res);
+
+      std::unique_ptr<detail::compressor> compressor;
+      if (type == detail::EncodingType::Gzip) {
+#ifdef CPPHTTPLIB_ZLIB_SUPPORT
+        compressor = detail::make_unique<detail::gzip_compressor>();
+#endif
+      } else if (type == detail::EncodingType::Brotli) {
+#ifdef CPPHTTPLIB_BROTLI_SUPPORT
+        compressor = detail::make_unique<detail::brotli_compressor>();
+#endif
+      } else {
+        compressor = detail::make_unique<detail::nocompressor>();
+      }
+      assert(compressor != nullptr);
+
+      return detail::write_content_chunked(strm, res.content_provider_,
+                                           is_shutting_down, *compressor);
+    } else {
+      return detail::write_content_without_length(strm, res.content_provider_,
+                                                  is_shutting_down);
+    }
+  }
+}
+
+inline bool Server::read_content(Stream &strm, Request &req, Response &res) {
+  MultipartFormDataMap::iterator cur;
+  auto file_count = 0;
+  if (read_content_core(
+          strm, req, res,
+          // Regular
+          [&](const char *buf, size_t n) {
+            if (req.body.size() + n > req.body.max_size()) { return false; }
+            req.body.append(buf, n);
+            return true;
+          },
+          // Multipart
+          [&](const MultipartFormData &file) {
+            if (file_count++ == CPPHTTPLIB_MULTIPART_FORM_DATA_FILE_MAX_COUNT) {
+              return false;
+            }
+            cur = req.files.emplace(file.name, file);
+            return true;
+          },
+          [&](const char *buf, size_t n) {
+            auto &content = cur->second.content;
+            if (content.size() + n > content.max_size()) { return false; }
+            content.append(buf, n);
+            return true;
+          })) {
+    const auto &content_type = req.get_header_value("Content-Type");
+    if (!content_type.find("application/x-www-form-urlencoded")) {
+      if (req.body.size() > CPPHTTPLIB_FORM_URL_ENCODED_PAYLOAD_MAX_LENGTH) {
+        res.status = 413; // NOTE: should be 414?
+        return false;
+      }
+      detail::parse_query_text(req.body, req.params);
+    }
+    return true;
+  }
+  return false;
+}
+
+inline bool Server::read_content_with_content_receiver(
+    Stream &strm, Request &req, Response &res, ContentReceiver receiver,
+    MultipartContentHeader multipart_header,
+    ContentReceiver multipart_receiver) {
+  return read_content_core(strm, req, res, std::move(receiver),
+                           std::move(multipart_header),
+                           std::move(multipart_receiver));
+}
+
+inline bool Server::read_content_core(Stream &strm, Request &req, Response &res,
+                                      ContentReceiver receiver,
+                                      MultipartContentHeader multipart_header,
+                                      ContentReceiver multipart_receiver) {
+  detail::MultipartFormDataParser multipart_form_data_parser;
+  ContentReceiverWithProgress out;
+
+  if (req.is_multipart_form_data()) {
+    const auto &content_type = req.get_header_value("Content-Type");
+    std::string boundary;
+    if (!detail::parse_multipart_boundary(content_type, boundary)) {
+      res.status = 400;
+      return false;
+    }
+
+    multipart_form_data_parser.set_boundary(std::move(boundary));
+    out = [&](const char *buf, size_t n, uint64_t /*off*/, uint64_t /*len*/) {
+      /* For debug
+      size_t pos = 0;
+      while (pos < n) {
+        auto read_size = (std::min)<size_t>(1, n - pos);
+        auto ret = multipart_form_data_parser.parse(
+            buf + pos, read_size, multipart_receiver, multipart_header);
+        if (!ret) { return false; }
+        pos += read_size;
+      }
+      return true;
+      */
+      return multipart_form_data_parser.parse(buf, n, multipart_receiver,
+                                              multipart_header);
+    };
+  } else {
+    out = [receiver](const char *buf, size_t n, uint64_t /*off*/,
+                     uint64_t /*len*/) { return receiver(buf, n); };
+  }
+
+  if (req.method == "DELETE" && !req.has_header("Content-Length")) {
+    return true;
+  }
+
+  if (!detail::read_content(strm, req, payload_max_length_, res.status, nullptr,
+                            out, true)) {
+    return false;
+  }
+
+  if (req.is_multipart_form_data()) {
+    if (!multipart_form_data_parser.is_valid()) {
+      res.status = 400;
+      return false;
+    }
+  }
+
+  return true;
+}
+
+inline bool Server::handle_file_request(const Request &req, Response &res,
+                                        bool head) {
+  for (const auto &entry : base_dirs_) {
+    // Prefix match
+    if (!req.path.compare(0, entry.mount_point.size(), entry.mount_point)) {
+      std::string sub_path = "/" + req.path.substr(entry.mount_point.size());
+      if (detail::is_valid_path(sub_path)) {
+        auto path = entry.base_dir + sub_path;
+        if (path.back() == '/') { path += "index.html"; }
+
+        if (detail::is_file(path)) {
+          for (const auto &kv : entry.headers) {
+            res.set_header(kv.first.c_str(), kv.second);
+          }
+
+          auto mm = std::make_shared<detail::mmap>(path.c_str());
+          if (!mm->is_open()) { return false; }
+
+          res.set_content_provider(
+              mm->size(),
+              detail::find_content_type(path, file_extension_and_mimetype_map_,
+                                        default_file_mimetype_),
+              [mm](size_t offset, size_t length, DataSink &sink) -> bool {
+                sink.write(mm->data() + offset, length);
+                return true;
+              });
+
+          if (!head && file_request_handler_) {
+            file_request_handler_(req, res);
+          }
+
+          return true;
+        }
+      }
+    }
+  }
+  return false;
+}
+
+inline socket_t
+Server::create_server_socket(const std::string &host, int port,
+                             int socket_flags,
+                             SocketOptions socket_options) const {
+  return detail::create_socket(
+      host, std::string(), port, address_family_, socket_flags, tcp_nodelay_,
+      std::move(socket_options),
+      [](socket_t sock, struct addrinfo &ai) -> bool {
+        if (::bind(sock, ai.ai_addr, static_cast<socklen_t>(ai.ai_addrlen))) {
+          return false;
+        }
+        if (::listen(sock, CPPHTTPLIB_LISTEN_BACKLOG)) { return false; }
+        return true;
+      });
+}
+
+inline int Server::bind_internal(const std::string &host, int port,
+                                 int socket_flags) {
+  if (!is_valid()) { return -1; }
+
+  svr_sock_ = create_server_socket(host, port, socket_flags, socket_options_);
+  if (svr_sock_ == INVALID_SOCKET) { return -1; }
+
+  if (port == 0) {
+    struct sockaddr_storage addr;
+    socklen_t addr_len = sizeof(addr);
+    if (getsockname(svr_sock_, reinterpret_cast<struct sockaddr *>(&addr),
+                    &addr_len) == -1) {
+      return -1;
+    }
+    if (addr.ss_family == AF_INET) {
+      return ntohs(reinterpret_cast<struct sockaddr_in *>(&addr)->sin_port);
+    } else if (addr.ss_family == AF_INET6) {
+      return ntohs(reinterpret_cast<struct sockaddr_in6 *>(&addr)->sin6_port);
+    } else {
+      return -1;
+    }
+  } else {
+    return port;
+  }
+}
+
+inline bool Server::listen_internal() {
+  auto ret = true;
+  is_running_ = true;
+  auto se = detail::scope_exit([&]() { is_running_ = false; });
+
+  {
+    std::unique_ptr<TaskQueue> task_queue(new_task_queue());
+
+    while (svr_sock_ != INVALID_SOCKET) {
+#ifndef _WIN32
+      if (idle_interval_sec_ > 0 || idle_interval_usec_ > 0) {
+#endif
+        auto val = detail::select_read(svr_sock_, idle_interval_sec_,
+                                       idle_interval_usec_);
+        if (val == 0) { // Timeout
+          task_queue->on_idle();
+          continue;
+        }
+#ifndef _WIN32
+      }
+#endif
+      socket_t sock = accept(svr_sock_, nullptr, nullptr);
+
+      if (sock == INVALID_SOCKET) {
+        if (errno == EMFILE) {
+          // The per-process limit of open file descriptors has been reached.
+          // Try to accept new connections after a short sleep.
+          std::this_thread::sleep_for(std::chrono::milliseconds(1));
+          continue;
+        } else if (errno == EINTR || errno == EAGAIN) {
+          continue;
+        }
+        if (svr_sock_ != INVALID_SOCKET) {
+          detail::close_socket(svr_sock_);
+          ret = false;
+        } else {
+          ; // The server socket was closed by user.
+        }
+        break;
+      }
+
+      {
+#ifdef _WIN32
+        auto timeout = static_cast<uint32_t>(read_timeout_sec_ * 1000 +
+                                             read_timeout_usec_ / 1000);
+        setsockopt(sock, SOL_SOCKET, SO_RCVTIMEO,
+                   reinterpret_cast<const char *>(&timeout), sizeof(timeout));
+#else
+        timeval tv;
+        tv.tv_sec = static_cast<long>(read_timeout_sec_);
+        tv.tv_usec = static_cast<decltype(tv.tv_usec)>(read_timeout_usec_);
+        setsockopt(sock, SOL_SOCKET, SO_RCVTIMEO,
+                   reinterpret_cast<const void *>(&tv), sizeof(tv));
+#endif
+      }
+      {
+
+#ifdef _WIN32
+        auto timeout = static_cast<uint32_t>(write_timeout_sec_ * 1000 +
+                                             write_timeout_usec_ / 1000);
+        setsockopt(sock, SOL_SOCKET, SO_SNDTIMEO,
+                   reinterpret_cast<const char *>(&timeout), sizeof(timeout));
+#else
+        timeval tv;
+        tv.tv_sec = static_cast<long>(write_timeout_sec_);
+        tv.tv_usec = static_cast<decltype(tv.tv_usec)>(write_timeout_usec_);
+        setsockopt(sock, SOL_SOCKET, SO_SNDTIMEO,
+                   reinterpret_cast<const void *>(&tv), sizeof(tv));
+#endif
+      }
+
+      task_queue->enqueue([this, sock]() { process_and_close_socket(sock); });
+    }
+
+    task_queue->shutdown();
+  }
+
+  return ret;
+}
+
+inline bool Server::routing(Request &req, Response &res, Stream &strm) {
+  if (pre_routing_handler_ &&
+      pre_routing_handler_(req, res) == HandlerResponse::Handled) {
+    return true;
+  }
+
+  // File handler
+  auto is_head_request = req.method == "HEAD";
+  if ((req.method == "GET" || is_head_request) &&
+      handle_file_request(req, res, is_head_request)) {
+    return true;
+  }
+
+  if (detail::expect_content(req)) {
+    // Content reader handler
+    {
+      ContentReader reader(
+          [&](ContentReceiver receiver) {
+            return read_content_with_content_receiver(
+                strm, req, res, std::move(receiver), nullptr, nullptr);
+          },
+          [&](MultipartContentHeader header, ContentReceiver receiver) {
+            return read_content_with_content_receiver(strm, req, res, nullptr,
+                                                      std::move(header),
+                                                      std::move(receiver));
+          });
+
+      if (req.method == "POST") {
+        if (dispatch_request_for_content_reader(
+                req, res, std::move(reader),
+                post_handlers_for_content_reader_)) {
+          return true;
+        }
+      } else if (req.method == "PUT") {
+        if (dispatch_request_for_content_reader(
+                req, res, std::move(reader),
+                put_handlers_for_content_reader_)) {
+          return true;
+        }
+      } else if (req.method == "PATCH") {
+        if (dispatch_request_for_content_reader(
+                req, res, std::move(reader),
+                patch_handlers_for_content_reader_)) {
+          return true;
+        }
+      } else if (req.method == "DELETE") {
+        if (dispatch_request_for_content_reader(
+                req, res, std::move(reader),
+                delete_handlers_for_content_reader_)) {
+          return true;
+        }
+      }
+    }
+
+    // Read content into `req.body`
+    if (!read_content(strm, req, res)) { return false; }
+  }
+
+  // Regular handler
+  if (req.method == "GET" || req.method == "HEAD") {
+    return dispatch_request(req, res, get_handlers_);
+  } else if (req.method == "POST") {
+    return dispatch_request(req, res, post_handlers_);
+  } else if (req.method == "PUT") {
+    return dispatch_request(req, res, put_handlers_);
+  } else if (req.method == "DELETE") {
+    return dispatch_request(req, res, delete_handlers_);
+  } else if (req.method == "OPTIONS") {
+    return dispatch_request(req, res, options_handlers_);
+  } else if (req.method == "PATCH") {
+    return dispatch_request(req, res, patch_handlers_);
+  }
+
+  res.status = 400;
+  return false;
+}
+
+inline bool Server::dispatch_request(Request &req, Response &res,
+                                     const Handlers &handlers) {
+  for (const auto &x : handlers) {
+    const auto &matcher = x.first;
+    const auto &handler = x.second;
+
+    if (matcher->match(req)) {
+      handler(req, res);
+      return true;
+    }
+  }
+  return false;
+}
+
+inline void Server::apply_ranges(const Request &req, Response &res,
+                                 std::string &content_type,
+                                 std::string &boundary) {
+  if (req.ranges.size() > 1) {
+    boundary = detail::make_multipart_data_boundary();
+
+    auto it = res.headers.find("Content-Type");
+    if (it != res.headers.end()) {
+      content_type = it->second;
+      res.headers.erase(it);
+    }
+
+    res.set_header("Content-Type",
+                   "multipart/byteranges; boundary=" + boundary);
+  }
+
+  auto type = detail::encoding_type(req, res);
+
+  if (res.body.empty()) {
+    if (res.content_length_ > 0) {
+      size_t length = 0;
+      if (req.ranges.empty()) {
+        length = res.content_length_;
+      } else if (req.ranges.size() == 1) {
+        auto offsets =
+            detail::get_range_offset_and_length(req, res.content_length_, 0);
+        length = offsets.second;
+
+        auto content_range = detail::make_content_range_header_field(
+            req.ranges[0], res.content_length_);
+        res.set_header("Content-Range", content_range);
+      } else {
+        length = detail::get_multipart_ranges_data_length(req, res, boundary,
+                                                          content_type);
+      }
+      res.set_header("Content-Length", std::to_string(length));
+    } else {
+      if (res.content_provider_) {
+        if (res.is_chunked_content_provider_) {
+          res.set_header("Transfer-Encoding", "chunked");
+          if (type == detail::EncodingType::Gzip) {
+            res.set_header("Content-Encoding", "gzip");
+          } else if (type == detail::EncodingType::Brotli) {
+            res.set_header("Content-Encoding", "br");
+          }
+        }
+      }
+    }
+  } else {
+    if (req.ranges.empty()) {
+      ;
+    } else if (req.ranges.size() == 1) {
+      auto content_range = detail::make_content_range_header_field(
+          req.ranges[0], res.body.size());
+      res.set_header("Content-Range", content_range);
+
+      auto offsets =
+          detail::get_range_offset_and_length(req, res.body.size(), 0);
+      auto offset = offsets.first;
+      auto length = offsets.second;
+
+      if (offset < res.body.size()) {
+        res.body = res.body.substr(offset, length);
+      } else {
+        res.body.clear();
+        res.status = 416;
+      }
+    } else {
+      std::string data;
+      if (detail::make_multipart_ranges_data(req, res, boundary, content_type,
+                                             data)) {
+        res.body.swap(data);
+      } else {
+        res.body.clear();
+        res.status = 416;
+      }
+    }
+
+    if (type != detail::EncodingType::None) {
+      std::unique_ptr<detail::compressor> compressor;
+      std::string content_encoding;
+
+      if (type == detail::EncodingType::Gzip) {
+#ifdef CPPHTTPLIB_ZLIB_SUPPORT
+        compressor = detail::make_unique<detail::gzip_compressor>();
+        content_encoding = "gzip";
+#endif
+      } else if (type == detail::EncodingType::Brotli) {
+#ifdef CPPHTTPLIB_BROTLI_SUPPORT
+        compressor = detail::make_unique<detail::brotli_compressor>();
+        content_encoding = "br";
+#endif
+      }
+
+      if (compressor) {
+        std::string compressed;
+        if (compressor->compress(res.body.data(), res.body.size(), true,
+                                 [&](const char *data, size_t data_len) {
+                                   compressed.append(data, data_len);
+                                   return true;
+                                 })) {
+          res.body.swap(compressed);
+          res.set_header("Content-Encoding", content_encoding);
+        }
+      }
+    }
+
+    auto length = std::to_string(res.body.size());
+    res.set_header("Content-Length", length);
+  }
+}
+
+inline bool Server::dispatch_request_for_content_reader(
+    Request &req, Response &res, ContentReader content_reader,
+    const HandlersForContentReader &handlers) {
+  for (const auto &x : handlers) {
+    const auto &matcher = x.first;
+    const auto &handler = x.second;
+
+    if (matcher->match(req)) {
+      handler(req, res, content_reader);
+      return true;
+    }
+  }
+  return false;
+}
+
+inline bool
+Server::process_request(Stream &strm, bool close_connection,
+                        bool &connection_closed,
+                        const std::function<void(Request &)> &setup_request) {
+  std::array<char, 2048> buf{};
+
+  detail::stream_line_reader line_reader(strm, buf.data(), buf.size());
+
+  // Connection has been closed on client
+  if (!line_reader.getline()) { return false; }
+
+  Request req;
+
+  Response res;
+  res.version = "HTTP/1.1";
+  res.headers = default_headers_;
+
+#ifdef _WIN32
+  // TODO: Increase FD_SETSIZE statically (libzmq), dynamically (MySQL).
+#else
+#ifndef CPPHTTPLIB_USE_POLL
+  // Socket file descriptor exceeded FD_SETSIZE...
+  if (strm.socket() >= FD_SETSIZE) {
+    Headers dummy;
+    detail::read_headers(strm, dummy);
+    res.status = 500;
+    return write_response(strm, close_connection, req, res);
+  }
+#endif
+#endif
+
+  // Check if the request URI doesn't exceed the limit
+  if (line_reader.size() > CPPHTTPLIB_REQUEST_URI_MAX_LENGTH) {
+    Headers dummy;
+    detail::read_headers(strm, dummy);
+    res.status = 414;
+    return write_response(strm, close_connection, req, res);
+  }
+
+  // Request line and headers
+  if (!parse_request_line(line_reader.ptr(), req) ||
+      !detail::read_headers(strm, req.headers)) {
+    res.status = 400;
+    return write_response(strm, close_connection, req, res);
+  }
+
+  if (req.get_header_value("Connection") == "close") {
+    connection_closed = true;
+  }
+
+  if (req.version == "HTTP/1.0" &&
+      req.get_header_value("Connection") != "Keep-Alive") {
+    connection_closed = true;
+  }
+
+  strm.get_remote_ip_and_port(req.remote_addr, req.remote_port);
+  req.set_header("REMOTE_ADDR", req.remote_addr);
+  req.set_header("REMOTE_PORT", std::to_string(req.remote_port));
+
+  strm.get_local_ip_and_port(req.local_addr, req.local_port);
+  req.set_header("LOCAL_ADDR", req.local_addr);
+  req.set_header("LOCAL_PORT", std::to_string(req.local_port));
+
+  if (req.has_header("Range")) {
+    const auto &range_header_value = req.get_header_value("Range");
+    if (!detail::parse_range_header(range_header_value, req.ranges)) {
+      res.status = 416;
+      return write_response(strm, close_connection, req, res);
+    }
+  }
+
+  if (setup_request) { setup_request(req); }
+
+  if (req.get_header_value("Expect") == "100-continue") {
+    auto status = 100;
+    if (expect_100_continue_handler_) {
+      status = expect_100_continue_handler_(req, res);
+    }
+    switch (status) {
+    case 100:
+    case 417:
+      strm.write_format("HTTP/1.1 %d %s\r\n\r\n", status,
+                        status_message(status));
+      break;
+    default: return write_response(strm, close_connection, req, res);
+    }
+  }
+
+  // Rounting
+  auto routed = false;
+#ifdef CPPHTTPLIB_NO_EXCEPTIONS
+  routed = routing(req, res, strm);
+#else
+  try {
+    routed = routing(req, res, strm);
+  } catch (std::exception &e) {
+    if (exception_handler_) {
+      auto ep = std::current_exception();
+      exception_handler_(req, res, ep);
+      routed = true;
+    } else {
+      res.status = 500;
+      std::string val;
+      auto s = e.what();
+      for (size_t i = 0; s[i]; i++) {
+        switch (s[i]) {
+        case '\r': val += "\\r"; break;
+        case '\n': val += "\\n"; break;
+        default: val += s[i]; break;
+        }
+      }
+      res.set_header("EXCEPTION_WHAT", val);
+    }
+  } catch (...) {
+    if (exception_handler_) {
+      auto ep = std::current_exception();
+      exception_handler_(req, res, ep);
+      routed = true;
+    } else {
+      res.status = 500;
+      res.set_header("EXCEPTION_WHAT", "UNKNOWN");
+    }
+  }
+#endif
+
+  if (routed) {
+    if (res.status == -1) { res.status = req.ranges.empty() ? 200 : 206; }
+    return write_response_with_content(strm, close_connection, req, res);
+  } else {
+    if (res.status == -1) { res.status = 404; }
+    return write_response(strm, close_connection, req, res);
+  }
+}
+
+inline bool Server::is_valid() const { return true; }
+
+inline bool Server::process_and_close_socket(socket_t sock) {
+  auto ret = detail::process_server_socket(
+      svr_sock_, sock, keep_alive_max_count_, keep_alive_timeout_sec_,
+      read_timeout_sec_, read_timeout_usec_, write_timeout_sec_,
+      write_timeout_usec_,
+      [this](Stream &strm, bool close_connection, bool &connection_closed) {
+        return process_request(strm, close_connection, connection_closed,
+                               nullptr);
+      });
+
+  detail::shutdown_socket(sock);
+  detail::close_socket(sock);
+  return ret;
+}
+
+// HTTP client implementation
+inline ClientImpl::ClientImpl(const std::string &host)
+    : ClientImpl(host, 80, std::string(), std::string()) {}
+
+inline ClientImpl::ClientImpl(const std::string &host, int port)
+    : ClientImpl(host, port, std::string(), std::string()) {}
+
+inline ClientImpl::ClientImpl(const std::string &host, int port,
+                              const std::string &client_cert_path,
+                              const std::string &client_key_path)
+    : host_(host), port_(port),
+      host_and_port_(adjust_host_string(host) + ":" + std::to_string(port)),
+      client_cert_path_(client_cert_path), client_key_path_(client_key_path) {}
+
+inline ClientImpl::~ClientImpl() {
+  std::lock_guard<std::mutex> guard(socket_mutex_);
+  shutdown_socket(socket_);
+  close_socket(socket_);
+}
+
+inline bool ClientImpl::is_valid() const { return true; }
+
+inline void ClientImpl::copy_settings(const ClientImpl &rhs) {
+  client_cert_path_ = rhs.client_cert_path_;
+  client_key_path_ = rhs.client_key_path_;
+  connection_timeout_sec_ = rhs.connection_timeout_sec_;
+  read_timeout_sec_ = rhs.read_timeout_sec_;
+  read_timeout_usec_ = rhs.read_timeout_usec_;
+  write_timeout_sec_ = rhs.write_timeout_sec_;
+  write_timeout_usec_ = rhs.write_timeout_usec_;
+  basic_auth_username_ = rhs.basic_auth_username_;
+  basic_auth_password_ = rhs.basic_auth_password_;
+  bearer_token_auth_token_ = rhs.bearer_token_auth_token_;
+#ifdef CPPHTTPLIB_OPENSSL_SUPPORT
+  digest_auth_username_ = rhs.digest_auth_username_;
+  digest_auth_password_ = rhs.digest_auth_password_;
+#endif
+  keep_alive_ = rhs.keep_alive_;
+  follow_location_ = rhs.follow_location_;
+  url_encode_ = rhs.url_encode_;
+  address_family_ = rhs.address_family_;
+  tcp_nodelay_ = rhs.tcp_nodelay_;
+  socket_options_ = rhs.socket_options_;
+  compress_ = rhs.compress_;
+  decompress_ = rhs.decompress_;
+  interface_ = rhs.interface_;
+  proxy_host_ = rhs.proxy_host_;
+  proxy_port_ = rhs.proxy_port_;
+  proxy_basic_auth_username_ = rhs.proxy_basic_auth_username_;
+  proxy_basic_auth_password_ = rhs.proxy_basic_auth_password_;
+  proxy_bearer_token_auth_token_ = rhs.proxy_bearer_token_auth_token_;
+#ifdef CPPHTTPLIB_OPENSSL_SUPPORT
+  proxy_digest_auth_username_ = rhs.proxy_digest_auth_username_;
+  proxy_digest_auth_password_ = rhs.proxy_digest_auth_password_;
+#endif
+#ifdef CPPHTTPLIB_OPENSSL_SUPPORT
+  ca_cert_file_path_ = rhs.ca_cert_file_path_;
+  ca_cert_dir_path_ = rhs.ca_cert_dir_path_;
+  ca_cert_store_ = rhs.ca_cert_store_;
+#endif
+#ifdef CPPHTTPLIB_OPENSSL_SUPPORT
+  server_certificate_verification_ = rhs.server_certificate_verification_;
+#endif
+  logger_ = rhs.logger_;
+}
+
+inline socket_t ClientImpl::create_client_socket(Error &error) const {
+  if (!proxy_host_.empty() && proxy_port_ != -1) {
+    return detail::create_client_socket(
+        proxy_host_, std::string(), proxy_port_, address_family_, tcp_nodelay_,
+        socket_options_, connection_timeout_sec_, connection_timeout_usec_,
+        read_timeout_sec_, read_timeout_usec_, write_timeout_sec_,
+        write_timeout_usec_, interface_, error);
+  }
+
+  // Check is custom IP specified for host_
+  std::string ip;
+  auto it = addr_map_.find(host_);
+  if (it != addr_map_.end()) ip = it->second;
+
+  return detail::create_client_socket(
+      host_, ip, port_, address_family_, tcp_nodelay_, socket_options_,
+      connection_timeout_sec_, connection_timeout_usec_, read_timeout_sec_,
+      read_timeout_usec_, write_timeout_sec_, write_timeout_usec_, interface_,
+      error);
+}
+
+inline bool ClientImpl::create_and_connect_socket(Socket &socket,
+                                                  Error &error) {
+  auto sock = create_client_socket(error);
+  if (sock == INVALID_SOCKET) { return false; }
+  socket.sock = sock;
+  return true;
+}
+
+inline void ClientImpl::shutdown_ssl(Socket & /*socket*/,
+                                     bool /*shutdown_gracefully*/) {
+  // If there are any requests in flight from threads other than us, then it's
+  // a thread-unsafe race because individual ssl* objects are not thread-safe.
+  assert(socket_requests_in_flight_ == 0 ||
+         socket_requests_are_from_thread_ == std::this_thread::get_id());
+}
+
+inline void ClientImpl::shutdown_socket(Socket &socket) {
+  if (socket.sock == INVALID_SOCKET) { return; }
+  detail::shutdown_socket(socket.sock);
+}
+
+inline void ClientImpl::close_socket(Socket &socket) {
+  // If there are requests in flight in another thread, usually closing
+  // the socket will be fine and they will simply receive an error when
+  // using the closed socket, but it is still a bug since rarely the OS
+  // may reassign the socket id to be used for a new socket, and then
+  // suddenly they will be operating on a live socket that is different
+  // than the one they intended!
+  assert(socket_requests_in_flight_ == 0 ||
+         socket_requests_are_from_thread_ == std::this_thread::get_id());
+
+  // It is also a bug if this happens while SSL is still active
+#ifdef CPPHTTPLIB_OPENSSL_SUPPORT
+  assert(socket.ssl == nullptr);
+#endif
+  if (socket.sock == INVALID_SOCKET) { return; }
+  detail::close_socket(socket.sock);
+  socket.sock = INVALID_SOCKET;
+}
+
+inline bool ClientImpl::read_response_line(Stream &strm, const Request &req,
+                                           Response &res) {
+  std::array<char, 2048> buf{};
+
+  detail::stream_line_reader line_reader(strm, buf.data(), buf.size());
+
+  if (!line_reader.getline()) { return false; }
+
+#ifdef CPPHTTPLIB_ALLOW_LF_AS_LINE_TERMINATOR
+  const static std::regex re("(HTTP/1\\.[01]) (\\d{3})(?: (.*?))?\r?\n");
+#else
+  const static std::regex re("(HTTP/1\\.[01]) (\\d{3})(?: (.*?))?\r\n");
+#endif
+
+  std::cmatch m;
+  if (!std::regex_match(line_reader.ptr(), m, re)) {
+    return req.method == "CONNECT";
+  }
+  res.version = std::string(m[1]);
+  res.status = std::stoi(std::string(m[2]));
+  res.reason = std::string(m[3]);
+
+  // Ignore '100 Continue'
+  while (res.status == 100) {
+    if (!line_reader.getline()) { return false; } // CRLF
+    if (!line_reader.getline()) { return false; } // next response line
+
+    if (!std::regex_match(line_reader.ptr(), m, re)) { return false; }
+    res.version = std::string(m[1]);
+    res.status = std::stoi(std::string(m[2]));
+    res.reason = std::string(m[3]);
+  }
+
+  return true;
+}
+
+inline bool ClientImpl::send(Request &req, Response &res, Error &error) {
+  std::lock_guard<std::recursive_mutex> request_mutex_guard(request_mutex_);
+  auto ret = send_(req, res, error);
+  if (error == Error::SSLPeerCouldBeClosed_) {
+    assert(!ret);
+    ret = send_(req, res, error);
+  }
+  return ret;
+}
+
+inline bool ClientImpl::send_(Request &req, Response &res, Error &error) {
+  {
+    std::lock_guard<std::mutex> guard(socket_mutex_);
+
+    // Set this to false immediately - if it ever gets set to true by the end of
+    // the request, we know another thread instructed us to close the socket.
+    socket_should_be_closed_when_request_is_done_ = false;
+
+    auto is_alive = false;
+    if (socket_.is_open()) {
+      is_alive = detail::is_socket_alive(socket_.sock);
+      if (!is_alive) {
+        // Attempt to avoid sigpipe by shutting down nongracefully if it seems
+        // like the other side has already closed the connection Also, there
+        // cannot be any requests in flight from other threads since we locked
+        // request_mutex_, so safe to close everything immediately
+        const bool shutdown_gracefully = false;
+        shutdown_ssl(socket_, shutdown_gracefully);
+        shutdown_socket(socket_);
+        close_socket(socket_);
+      }
+    }
+
+    if (!is_alive) {
+      if (!create_and_connect_socket(socket_, error)) { return false; }
+
+#ifdef CPPHTTPLIB_OPENSSL_SUPPORT
+      // TODO: refactoring
+      if (is_ssl()) {
+        auto &scli = static_cast<SSLClient &>(*this);
+        if (!proxy_host_.empty() && proxy_port_ != -1) {
+          auto success = false;
+          if (!scli.connect_with_proxy(socket_, res, success, error)) {
+            return success;
+          }
+        }
+
+        if (!scli.initialize_ssl(socket_, error)) { return false; }
+      }
+#endif
+    }
+
+    // Mark the current socket as being in use so that it cannot be closed by
+    // anyone else while this request is ongoing, even though we will be
+    // releasing the mutex.
+    if (socket_requests_in_flight_ > 1) {
+      assert(socket_requests_are_from_thread_ == std::this_thread::get_id());
+    }
+    socket_requests_in_flight_ += 1;
+    socket_requests_are_from_thread_ = std::this_thread::get_id();
+  }
+
+  for (const auto &header : default_headers_) {
+    if (req.headers.find(header.first) == req.headers.end()) {
+      req.headers.insert(header);
+    }
+  }
+
+  auto ret = false;
+  auto close_connection = !keep_alive_;
+
+  auto se = detail::scope_exit([&]() {
+    // Briefly lock mutex in order to mark that a request is no longer ongoing
+    std::lock_guard<std::mutex> guard(socket_mutex_);
+    socket_requests_in_flight_ -= 1;
+    if (socket_requests_in_flight_ <= 0) {
+      assert(socket_requests_in_flight_ == 0);
+      socket_requests_are_from_thread_ = std::thread::id();
+    }
+
+    if (socket_should_be_closed_when_request_is_done_ || close_connection ||
+        !ret) {
+      shutdown_ssl(socket_, true);
+      shutdown_socket(socket_);
+      close_socket(socket_);
+    }
+  });
+
+  ret = process_socket(socket_, [&](Stream &strm) {
+    return handle_request(strm, req, res, close_connection, error);
+  });
+
+  if (!ret) {
+    if (error == Error::Success) { error = Error::Unknown; }
+  }
+
+  return ret;
+}
+
+inline Result ClientImpl::send(const Request &req) {
+  auto req2 = req;
+  return send_(std::move(req2));
+}
+
+inline Result ClientImpl::send_(Request &&req) {
+  auto res = detail::make_unique<Response>();
+  auto error = Error::Success;
+  auto ret = send(req, *res, error);
+  return Result{ret ? std::move(res) : nullptr, error, std::move(req.headers)};
+}
+
+inline bool ClientImpl::handle_request(Stream &strm, Request &req,
+                                       Response &res, bool close_connection,
+                                       Error &error) {
+  if (req.path.empty()) {
+    error = Error::Connection;
+    return false;
+  }
+
+  auto req_save = req;
+
+  bool ret;
+
+  if (!is_ssl() && !proxy_host_.empty() && proxy_port_ != -1) {
+    auto req2 = req;
+    req2.path = "http://" + host_and_port_ + req.path;
+    ret = process_request(strm, req2, res, close_connection, error);
+    req = req2;
+    req.path = req_save.path;
+  } else {
+    ret = process_request(strm, req, res, close_connection, error);
+  }
+
+  if (!ret) { return false; }
+
+  if (res.get_header_value("Connection") == "close" ||
+      (res.version == "HTTP/1.0" && res.reason != "Connection established")) {
+    // TODO this requires a not-entirely-obvious chain of calls to be correct
+    // for this to be safe.
+
+    // This is safe to call because handle_request is only called by send_
+    // which locks the request mutex during the process. It would be a bug
+    // to call it from a different thread since it's a thread-safety issue
+    // to do these things to the socket if another thread is using the socket.
+    std::lock_guard<std::mutex> guard(socket_mutex_);
+    shutdown_ssl(socket_, true);
+    shutdown_socket(socket_);
+    close_socket(socket_);
+  }
+
+  if (300 < res.status && res.status < 400 && follow_location_) {
+    req = req_save;
+    ret = redirect(req, res, error);
+  }
+
+#ifdef CPPHTTPLIB_OPENSSL_SUPPORT
+  if ((res.status == 401 || res.status == 407) &&
+      req.authorization_count_ < 5) {
+    auto is_proxy = res.status == 407;
+    const auto &username =
+        is_proxy ? proxy_digest_auth_username_ : digest_auth_username_;
+    const auto &password =
+        is_proxy ? proxy_digest_auth_password_ : digest_auth_password_;
+
+    if (!username.empty() && !password.empty()) {
+      std::map<std::string, std::string> auth;
+      if (detail::parse_www_authenticate(res, auth, is_proxy)) {
+        Request new_req = req;
+        new_req.authorization_count_ += 1;
+        new_req.headers.erase(is_proxy ? "Proxy-Authorization"
+                                       : "Authorization");
+        new_req.headers.insert(detail::make_digest_authentication_header(
+            req, auth, new_req.authorization_count_, detail::random_string(10),
+            username, password, is_proxy));
+
+        Response new_res;
+
+        ret = send(new_req, new_res, error);
+        if (ret) { res = new_res; }
+      }
+    }
+  }
+#endif
+
+  return ret;
+}
+
+inline bool ClientImpl::redirect(Request &req, Response &res, Error &error) {
+  if (req.redirect_count_ == 0) {
+    error = Error::ExceedRedirectCount;
+    return false;
+  }
+
+  auto location = res.get_header_value("location");
+  if (location.empty()) { return false; }
+
+  const static std::regex re(
+      R"((?:(https?):)?(?://(?:\[([\d:]+)\]|([^:/?#]+))(?::(\d+))?)?([^?#]*)(\?[^#]*)?(?:#.*)?)");
+
+  std::smatch m;
+  if (!std::regex_match(location, m, re)) { return false; }
+
+  auto scheme = is_ssl() ? "https" : "http";
+
+  auto next_scheme = m[1].str();
+  auto next_host = m[2].str();
+  if (next_host.empty()) { next_host = m[3].str(); }
+  auto port_str = m[4].str();
+  auto next_path = m[5].str();
+  auto next_query = m[6].str();
+
+  auto next_port = port_;
+  if (!port_str.empty()) {
+    next_port = std::stoi(port_str);
+  } else if (!next_scheme.empty()) {
+    next_port = next_scheme == "https" ? 443 : 80;
+  }
+
+  if (next_scheme.empty()) { next_scheme = scheme; }
+  if (next_host.empty()) { next_host = host_; }
+  if (next_path.empty()) { next_path = "/"; }
+
+  auto path = detail::decode_url(next_path, true) + next_query;
+
+  if (next_scheme == scheme && next_host == host_ && next_port == port_) {
+    return detail::redirect(*this, req, res, path, location, error);
+  } else {
+    if (next_scheme == "https") {
+#ifdef CPPHTTPLIB_OPENSSL_SUPPORT
+      SSLClient cli(next_host.c_str(), next_port);
+      cli.copy_settings(*this);
+      if (ca_cert_store_) { cli.set_ca_cert_store(ca_cert_store_); }
+      return detail::redirect(cli, req, res, path, location, error);
+#else
+      return false;
+#endif
+    } else {
+      ClientImpl cli(next_host.c_str(), next_port);
+      cli.copy_settings(*this);
+      return detail::redirect(cli, req, res, path, location, error);
+    }
+  }
+}
+
+inline bool ClientImpl::write_content_with_provider(Stream &strm,
+                                                    const Request &req,
+                                                    Error &error) {
+  auto is_shutting_down = []() { return false; };
+
+  if (req.is_chunked_content_provider_) {
+    // TODO: Brotli support
+    std::unique_ptr<detail::compressor> compressor;
+#ifdef CPPHTTPLIB_ZLIB_SUPPORT
+    if (compress_) {
+      compressor = detail::make_unique<detail::gzip_compressor>();
+    } else
+#endif
+    {
+      compressor = detail::make_unique<detail::nocompressor>();
+    }
+
+    return detail::write_content_chunked(strm, req.content_provider_,
+                                         is_shutting_down, *compressor, error);
+  } else {
+    return detail::write_content(strm, req.content_provider_, 0,
+                                 req.content_length_, is_shutting_down, error);
+  }
+}
+
+inline bool ClientImpl::write_request(Stream &strm, Request &req,
+                                      bool close_connection, Error &error) {
+  // Prepare additional headers
+  if (close_connection) {
+    if (!req.has_header("Connection")) {
+      req.set_header("Connection", "close");
+    }
+  }
+
+  if (!req.has_header("Host")) {
+    if (is_ssl()) {
+      if (port_ == 443) {
+        req.set_header("Host", host_);
+      } else {
+        req.set_header("Host", host_and_port_);
+      }
+    } else {
+      if (port_ == 80) {
+        req.set_header("Host", host_);
+      } else {
+        req.set_header("Host", host_and_port_);
+      }
+    }
+  }
+
+  if (!req.has_header("Accept")) { req.set_header("Accept", "*/*"); }
+
+#ifndef CPPHTTPLIB_NO_DEFAULT_USER_AGENT
+  if (!req.has_header("User-Agent")) {
+    auto agent = std::string("cpp-httplib/") + CPPHTTPLIB_VERSION;
+    req.set_header("User-Agent", agent);
+  }
+#endif
+
+  if (req.body.empty()) {
+    if (req.content_provider_) {
+      if (!req.is_chunked_content_provider_) {
+        if (!req.has_header("Content-Length")) {
+          auto length = std::to_string(req.content_length_);
+          req.set_header("Content-Length", length);
+        }
+      }
+    } else {
+      if (req.method == "POST" || req.method == "PUT" ||
+          req.method == "PATCH") {
+        req.set_header("Content-Length", "0");
+      }
+    }
+  } else {
+    if (!req.has_header("Content-Type")) {
+      req.set_header("Content-Type", "text/plain");
+    }
+
+    if (!req.has_header("Content-Length")) {
+      auto length = std::to_string(req.body.size());
+      req.set_header("Content-Length", length);
+    }
+  }
+
+  if (!basic_auth_password_.empty() || !basic_auth_username_.empty()) {
+    if (!req.has_header("Authorization")) {
+      req.headers.insert(make_basic_authentication_header(
+          basic_auth_username_, basic_auth_password_, false));
+    }
+  }
+
+  if (!proxy_basic_auth_username_.empty() &&
+      !proxy_basic_auth_password_.empty()) {
+    if (!req.has_header("Proxy-Authorization")) {
+      req.headers.insert(make_basic_authentication_header(
+          proxy_basic_auth_username_, proxy_basic_auth_password_, true));
+    }
+  }
+
+  if (!bearer_token_auth_token_.empty()) {
+    if (!req.has_header("Authorization")) {
+      req.headers.insert(make_bearer_token_authentication_header(
+          bearer_token_auth_token_, false));
+    }
+  }
+
+  if (!proxy_bearer_token_auth_token_.empty()) {
+    if (!req.has_header("Proxy-Authorization")) {
+      req.headers.insert(make_bearer_token_authentication_header(
+          proxy_bearer_token_auth_token_, true));
+    }
+  }
+
+  // Request line and headers
+  {
+    detail::BufferStream bstrm;
+
+    const auto &path = url_encode_ ? detail::encode_url(req.path) : req.path;
+    bstrm.write_format("%s %s HTTP/1.1\r\n", req.method.c_str(), path.c_str());
+
+    header_writer_(bstrm, req.headers);
+
+    // Flush buffer
+    auto &data = bstrm.get_buffer();
+    if (!detail::write_data(strm, data.data(), data.size())) {
+      error = Error::Write;
+      return false;
+    }
+  }
+
+  // Body
+  if (req.body.empty()) {
+    return write_content_with_provider(strm, req, error);
+  }
+
+  if (!detail::write_data(strm, req.body.data(), req.body.size())) {
+    error = Error::Write;
+    return false;
+  }
+
+  return true;
+}
+
+inline std::unique_ptr<Response> ClientImpl::send_with_content_provider(
+    Request &req, const char *body, size_t content_length,
+    ContentProvider content_provider,
+    ContentProviderWithoutLength content_provider_without_length,
+    const std::string &content_type, Error &error) {
+  if (!content_type.empty()) { req.set_header("Content-Type", content_type); }
+
+#ifdef CPPHTTPLIB_ZLIB_SUPPORT
+  if (compress_) { req.set_header("Content-Encoding", "gzip"); }
+#endif
+
+#ifdef CPPHTTPLIB_ZLIB_SUPPORT
+  if (compress_ && !content_provider_without_length) {
+    // TODO: Brotli support
+    detail::gzip_compressor compressor;
+
+    if (content_provider) {
+      auto ok = true;
+      size_t offset = 0;
+      DataSink data_sink;
+
+      data_sink.write = [&](const char *data, size_t data_len) -> bool {
+        if (ok) {
+          auto last = offset + data_len == content_length;
+
+          auto ret = compressor.compress(
+              data, data_len, last,
+              [&](const char *compressed_data, size_t compressed_data_len) {
+                req.body.append(compressed_data, compressed_data_len);
+                return true;
+              });
+
+          if (ret) {
+            offset += data_len;
+          } else {
+            ok = false;
+          }
+        }
+        return ok;
+      };
+
+      while (ok && offset < content_length) {
+        if (!content_provider(offset, content_length - offset, data_sink)) {
+          error = Error::Canceled;
+          return nullptr;
+        }
+      }
+    } else {
+      if (!compressor.compress(body, content_length, true,
+                               [&](const char *data, size_t data_len) {
+                                 req.body.append(data, data_len);
+                                 return true;
+                               })) {
+        error = Error::Compression;
+        return nullptr;
+      }
+    }
+  } else
+#endif
+  {
+    if (content_provider) {
+      req.content_length_ = content_length;
+      req.content_provider_ = std::move(content_provider);
+      req.is_chunked_content_provider_ = false;
+    } else if (content_provider_without_length) {
+      req.content_length_ = 0;
+      req.content_provider_ = detail::ContentProviderAdapter(
+          std::move(content_provider_without_length));
+      req.is_chunked_content_provider_ = true;
+      req.set_header("Transfer-Encoding", "chunked");
+    } else {
+      req.body.assign(body, content_length);
+    }
+  }
+
+  auto res = detail::make_unique<Response>();
+  return send(req, *res, error) ? std::move(res) : nullptr;
+}
+
+inline Result ClientImpl::send_with_content_provider(
+    const std::string &method, const std::string &path, const Headers &headers,
+    const char *body, size_t content_length, ContentProvider content_provider,
+    ContentProviderWithoutLength content_provider_without_length,
+    const std::string &content_type) {
+  Request req;
+  req.method = method;
+  req.headers = headers;
+  req.path = path;
+
+  auto error = Error::Success;
+
+  auto res = send_with_content_provider(
+      req, body, content_length, std::move(content_provider),
+      std::move(content_provider_without_length), content_type, error);
+
+  return Result{std::move(res), error, std::move(req.headers)};
+}
+
+inline std::string
+ClientImpl::adjust_host_string(const std::string &host) const {
+  if (host.find(':') != std::string::npos) { return "[" + host + "]"; }
+  return host;
+}
+
+inline bool ClientImpl::process_request(Stream &strm, Request &req,
+                                        Response &res, bool close_connection,
+                                        Error &error) {
+  // Send request
+  if (!write_request(strm, req, close_connection, error)) { return false; }
+
+#ifdef CPPHTTPLIB_OPENSSL_SUPPORT
+  if (is_ssl()) {
+    auto is_proxy_enabled = !proxy_host_.empty() && proxy_port_ != -1;
+    if (!is_proxy_enabled) {
+      char buf[1];
+      if (SSL_peek(socket_.ssl, buf, 1) == 0 &&
+          SSL_get_error(socket_.ssl, 0) == SSL_ERROR_ZERO_RETURN) {
+        error = Error::SSLPeerCouldBeClosed_;
+        return false;
+      }
+    }
+  }
+#endif
+
+  // Receive response and headers
+  if (!read_response_line(strm, req, res) ||
+      !detail::read_headers(strm, res.headers)) {
+    error = Error::Read;
+    return false;
+  }
+
+  // Body
+  if ((res.status != 204) && req.method != "HEAD" && req.method != "CONNECT") {
+    auto redirect = 300 < res.status && res.status < 400 && follow_location_;
+
+    if (req.response_handler && !redirect) {
+      if (!req.response_handler(res)) {
+        error = Error::Canceled;
+        return false;
+      }
+    }
+
+    auto out =
+        req.content_receiver
+            ? static_cast<ContentReceiverWithProgress>(
+                  [&](const char *buf, size_t n, uint64_t off, uint64_t len) {
+                    if (redirect) { return true; }
+                    auto ret = req.content_receiver(buf, n, off, len);
+                    if (!ret) { error = Error::Canceled; }
+                    return ret;
+                  })
+            : static_cast<ContentReceiverWithProgress>(
+                  [&](const char *buf, size_t n, uint64_t /*off*/,
+                      uint64_t /*len*/) {
+                    if (res.body.size() + n > res.body.max_size()) {
+                      return false;
+                    }
+                    res.body.append(buf, n);
+                    return true;
+                  });
+
+    auto progress = [&](uint64_t current, uint64_t total) {
+      if (!req.progress || redirect) { return true; }
+      auto ret = req.progress(current, total);
+      if (!ret) { error = Error::Canceled; }
+      return ret;
+    };
+
+    int dummy_status;
+    if (!detail::read_content(strm, res, (std::numeric_limits<size_t>::max)(),
+                              dummy_status, std::move(progress), std::move(out),
+                              decompress_)) {
+      if (error != Error::Canceled) { error = Error::Read; }
+      return false;
+    }
+  }
+
+  // Log
+  if (logger_) { logger_(req, res); }
+
+  return true;
+}
+
+inline ContentProviderWithoutLength ClientImpl::get_multipart_content_provider(
+    const std::string &boundary, const MultipartFormDataItems &items,
+    const MultipartFormDataProviderItems &provider_items) {
+  size_t cur_item = 0, cur_start = 0;
+  // cur_item and cur_start are copied to within the std::function and maintain
+  // state between successive calls
+  return [&, cur_item, cur_start](size_t offset,
+                                  DataSink &sink) mutable -> bool {
+    if (!offset && items.size()) {
+      sink.os << detail::serialize_multipart_formdata(items, boundary, false);
+      return true;
+    } else if (cur_item < provider_items.size()) {
+      if (!cur_start) {
+        const auto &begin = detail::serialize_multipart_formdata_item_begin(
+            provider_items[cur_item], boundary);
+        offset += begin.size();
+        cur_start = offset;
+        sink.os << begin;
+      }
+
+      DataSink cur_sink;
+      auto has_data = true;
+      cur_sink.write = sink.write;
+      cur_sink.done = [&]() { has_data = false; };
+
+      if (!provider_items[cur_item].provider(offset - cur_start, cur_sink))
+        return false;
+
+      if (!has_data) {
+        sink.os << detail::serialize_multipart_formdata_item_end();
+        cur_item++;
+        cur_start = 0;
+      }
+      return true;
+    } else {
+      sink.os << detail::serialize_multipart_formdata_finish(boundary);
+      sink.done();
+      return true;
+    }
+  };
+}
+
+inline bool
+ClientImpl::process_socket(const Socket &socket,
+                           std::function<bool(Stream &strm)> callback) {
+  return detail::process_client_socket(
+      socket.sock, read_timeout_sec_, read_timeout_usec_, write_timeout_sec_,
+      write_timeout_usec_, std::move(callback));
+}
+
+inline bool ClientImpl::is_ssl() const { return false; }
+
+inline Result ClientImpl::Get(const std::string &path) {
+  return Get(path, Headers(), Progress());
+}
+
+inline Result ClientImpl::Get(const std::string &path, Progress progress) {
+  return Get(path, Headers(), std::move(progress));
+}
+
+inline Result ClientImpl::Get(const std::string &path, const Headers &headers) {
+  return Get(path, headers, Progress());
+}
+
+inline Result ClientImpl::Get(const std::string &path, const Headers &headers,
+                              Progress progress) {
+  Request req;
+  req.method = "GET";
+  req.path = path;
+  req.headers = headers;
+  req.progress = std::move(progress);
+
+  return send_(std::move(req));
+}
+
+inline Result ClientImpl::Get(const std::string &path,
+                              ContentReceiver content_receiver) {
+  return Get(path, Headers(), nullptr, std::move(content_receiver), nullptr);
+}
+
+inline Result ClientImpl::Get(const std::string &path,
+                              ContentReceiver content_receiver,
+                              Progress progress) {
+  return Get(path, Headers(), nullptr, std::move(content_receiver),
+             std::move(progress));
+}
+
+inline Result ClientImpl::Get(const std::string &path, const Headers &headers,
+                              ContentReceiver content_receiver) {
+  return Get(path, headers, nullptr, std::move(content_receiver), nullptr);
+}
+
+inline Result ClientImpl::Get(const std::string &path, const Headers &headers,
+                              ContentReceiver content_receiver,
+                              Progress progress) {
+  return Get(path, headers, nullptr, std::move(content_receiver),
+             std::move(progress));
+}
+
+inline Result ClientImpl::Get(const std::string &path,
+                              ResponseHandler response_handler,
+                              ContentReceiver content_receiver) {
+  return Get(path, Headers(), std::move(response_handler),
+             std::move(content_receiver), nullptr);
+}
+
+inline Result ClientImpl::Get(const std::string &path, const Headers &headers,
+                              ResponseHandler response_handler,
+                              ContentReceiver content_receiver) {
+  return Get(path, headers, std::move(response_handler),
+             std::move(content_receiver), nullptr);
+}
+
+inline Result ClientImpl::Get(const std::string &path,
+                              ResponseHandler response_handler,
+                              ContentReceiver content_receiver,
+                              Progress progress) {
+  return Get(path, Headers(), std::move(response_handler),
+             std::move(content_receiver), std::move(progress));
+}
+
+inline Result ClientImpl::Get(const std::string &path, const Headers &headers,
+                              ResponseHandler response_handler,
+                              ContentReceiver content_receiver,
+                              Progress progress) {
+  Request req;
+  req.method = "GET";
+  req.path = path;
+  req.headers = headers;
+  req.response_handler = std::move(response_handler);
+  req.content_receiver =
+      [content_receiver](const char *data, size_t data_length,
+                         uint64_t /*offset*/, uint64_t /*total_length*/) {
+        return content_receiver(data, data_length);
+      };
+  req.progress = std::move(progress);
+
+  return send_(std::move(req));
+}
+
+inline Result ClientImpl::Get(const std::string &path, const Params &params,
+                              const Headers &headers, Progress progress) {
+  if (params.empty()) { return Get(path, headers); }
+
+  std::string path_with_query = append_query_params(path, params);
+  return Get(path_with_query.c_str(), headers, progress);
+}
+
+inline Result ClientImpl::Get(const std::string &path, const Params &params,
+                              const Headers &headers,
+                              ContentReceiver content_receiver,
+                              Progress progress) {
+  return Get(path, params, headers, nullptr, content_receiver, progress);
+}
+
+inline Result ClientImpl::Get(const std::string &path, const Params &params,
+                              const Headers &headers,
+                              ResponseHandler response_handler,
+                              ContentReceiver content_receiver,
+                              Progress progress) {
+  if (params.empty()) {
+    return Get(path, headers, response_handler, content_receiver, progress);
+  }
+
+  std::string path_with_query = append_query_params(path, params);
+  return Get(path_with_query.c_str(), headers, response_handler,
+             content_receiver, progress);
+}
+
+inline Result ClientImpl::Head(const std::string &path) {
+  return Head(path, Headers());
+}
+
+inline Result ClientImpl::Head(const std::string &path,
+                               const Headers &headers) {
+  Request req;
+  req.method = "HEAD";
+  req.headers = headers;
+  req.path = path;
+
+  return send_(std::move(req));
+}
+
+inline Result ClientImpl::Post(const std::string &path) {
+  return Post(path, std::string(), std::string());
+}
+
+inline Result ClientImpl::Post(const std::string &path,
+                               const Headers &headers) {
+  return Post(path, headers, nullptr, 0, std::string());
+}
+
+inline Result ClientImpl::Post(const std::string &path, const char *body,
+                               size_t content_length,
+                               const std::string &content_type) {
+  return Post(path, Headers(), body, content_length, content_type);
+}
+
+inline Result ClientImpl::Post(const std::string &path, const Headers &headers,
+                               const char *body, size_t content_length,
+                               const std::string &content_type) {
+  return send_with_content_provider("POST", path, headers, body, content_length,
+                                    nullptr, nullptr, content_type);
+}
+
+inline Result ClientImpl::Post(const std::string &path, const std::string &body,
+                               const std::string &content_type) {
+  return Post(path, Headers(), body, content_type);
+}
+
+inline Result ClientImpl::Post(const std::string &path, const Headers &headers,
+                               const std::string &body,
+                               const std::string &content_type) {
+  return send_with_content_provider("POST", path, headers, body.data(),
+                                    body.size(), nullptr, nullptr,
+                                    content_type);
+}
+
+inline Result ClientImpl::Post(const std::string &path, const Params &params) {
+  return Post(path, Headers(), params);
+}
+
+inline Result ClientImpl::Post(const std::string &path, size_t content_length,
+                               ContentProvider content_provider,
+                               const std::string &content_type) {
+  return Post(path, Headers(), content_length, std::move(content_provider),
+              content_type);
+}
+
+inline Result ClientImpl::Post(const std::string &path,
+                               ContentProviderWithoutLength content_provider,
+                               const std::string &content_type) {
+  return Post(path, Headers(), std::move(content_provider), content_type);
+}
+
+inline Result ClientImpl::Post(const std::string &path, const Headers &headers,
+                               size_t content_length,
+                               ContentProvider content_provider,
+                               const std::string &content_type) {
+  return send_with_content_provider("POST", path, headers, nullptr,
+                                    content_length, std::move(content_provider),
+                                    nullptr, content_type);
+}
+
+inline Result ClientImpl::Post(const std::string &path, const Headers &headers,
+                               ContentProviderWithoutLength content_provider,
+                               const std::string &content_type) {
+  return send_with_content_provider("POST", path, headers, nullptr, 0, nullptr,
+                                    std::move(content_provider), content_type);
+}
+
+inline Result ClientImpl::Post(const std::string &path, const Headers &headers,
+                               const Params &params) {
+  auto query = detail::params_to_query_str(params);
+  return Post(path, headers, query, "application/x-www-form-urlencoded");
+}
+
+inline Result ClientImpl::Post(const std::string &path,
+                               const MultipartFormDataItems &items) {
+  return Post(path, Headers(), items);
+}
+
+inline Result ClientImpl::Post(const std::string &path, const Headers &headers,
+                               const MultipartFormDataItems &items) {
+  const auto &boundary = detail::make_multipart_data_boundary();
+  const auto &content_type =
+      detail::serialize_multipart_formdata_get_content_type(boundary);
+  const auto &body = detail::serialize_multipart_formdata(items, boundary);
+  return Post(path, headers, body, content_type.c_str());
+}
+
+inline Result ClientImpl::Post(const std::string &path, const Headers &headers,
+                               const MultipartFormDataItems &items,
+                               const std::string &boundary) {
+  if (!detail::is_multipart_boundary_chars_valid(boundary)) {
+    return Result{nullptr, Error::UnsupportedMultipartBoundaryChars};
+  }
+
+  const auto &content_type =
+      detail::serialize_multipart_formdata_get_content_type(boundary);
+  const auto &body = detail::serialize_multipart_formdata(items, boundary);
+  return Post(path, headers, body, content_type.c_str());
+}
+
+inline Result
+ClientImpl::Post(const std::string &path, const Headers &headers,
+                 const MultipartFormDataItems &items,
+                 const MultipartFormDataProviderItems &provider_items) {
+  const auto &boundary = detail::make_multipart_data_boundary();
+  const auto &content_type =
+      detail::serialize_multipart_formdata_get_content_type(boundary);
+  return send_with_content_provider(
+      "POST", path, headers, nullptr, 0, nullptr,
+      get_multipart_content_provider(boundary, items, provider_items),
+      content_type);
+}
+
+inline Result ClientImpl::Put(const std::string &path) {
+  return Put(path, std::string(), std::string());
+}
+
+inline Result ClientImpl::Put(const std::string &path, const char *body,
+                              size_t content_length,
+                              const std::string &content_type) {
+  return Put(path, Headers(), body, content_length, content_type);
+}
+
+inline Result ClientImpl::Put(const std::string &path, const Headers &headers,
+                              const char *body, size_t content_length,
+                              const std::string &content_type) {
+  return send_with_content_provider("PUT", path, headers, body, content_length,
+                                    nullptr, nullptr, content_type);
+}
+
+inline Result ClientImpl::Put(const std::string &path, const std::string &body,
+                              const std::string &content_type) {
+  return Put(path, Headers(), body, content_type);
+}
+
+inline Result ClientImpl::Put(const std::string &path, const Headers &headers,
+                              const std::string &body,
+                              const std::string &content_type) {
+  return send_with_content_provider("PUT", path, headers, body.data(),
+                                    body.size(), nullptr, nullptr,
+                                    content_type);
+}
+
+inline Result ClientImpl::Put(const std::string &path, size_t content_length,
+                              ContentProvider content_provider,
+                              const std::string &content_type) {
+  return Put(path, Headers(), content_length, std::move(content_provider),
+             content_type);
+}
+
+inline Result ClientImpl::Put(const std::string &path,
+                              ContentProviderWithoutLength content_provider,
+                              const std::string &content_type) {
+  return Put(path, Headers(), std::move(content_provider), content_type);
+}
+
+inline Result ClientImpl::Put(const std::string &path, const Headers &headers,
+                              size_t content_length,
+                              ContentProvider content_provider,
+                              const std::string &content_type) {
+  return send_with_content_provider("PUT", path, headers, nullptr,
+                                    content_length, std::move(content_provider),
+                                    nullptr, content_type);
+}
+
+inline Result ClientImpl::Put(const std::string &path, const Headers &headers,
+                              ContentProviderWithoutLength content_provider,
+                              const std::string &content_type) {
+  return send_with_content_provider("PUT", path, headers, nullptr, 0, nullptr,
+                                    std::move(content_provider), content_type);
+}
+
+inline Result ClientImpl::Put(const std::string &path, const Params &params) {
+  return Put(path, Headers(), params);
+}
+
+inline Result ClientImpl::Put(const std::string &path, const Headers &headers,
+                              const Params &params) {
+  auto query = detail::params_to_query_str(params);
+  return Put(path, headers, query, "application/x-www-form-urlencoded");
+}
+
+inline Result ClientImpl::Put(const std::string &path,
+                              const MultipartFormDataItems &items) {
+  return Put(path, Headers(), items);
+}
+
+inline Result ClientImpl::Put(const std::string &path, const Headers &headers,
+                              const MultipartFormDataItems &items) {
+  const auto &boundary = detail::make_multipart_data_boundary();
+  const auto &content_type =
+      detail::serialize_multipart_formdata_get_content_type(boundary);
+  const auto &body = detail::serialize_multipart_formdata(items, boundary);
+  return Put(path, headers, body, content_type);
+}
+
+inline Result ClientImpl::Put(const std::string &path, const Headers &headers,
+                              const MultipartFormDataItems &items,
+                              const std::string &boundary) {
+  if (!detail::is_multipart_boundary_chars_valid(boundary)) {
+    return Result{nullptr, Error::UnsupportedMultipartBoundaryChars};
+  }
+
+  const auto &content_type =
+      detail::serialize_multipart_formdata_get_content_type(boundary);
+  const auto &body = detail::serialize_multipart_formdata(items, boundary);
+  return Put(path, headers, body, content_type);
+}
+
+inline Result
+ClientImpl::Put(const std::string &path, const Headers &headers,
+                const MultipartFormDataItems &items,
+                const MultipartFormDataProviderItems &provider_items) {
+  const auto &boundary = detail::make_multipart_data_boundary();
+  const auto &content_type =
+      detail::serialize_multipart_formdata_get_content_type(boundary);
+  return send_with_content_provider(
+      "PUT", path, headers, nullptr, 0, nullptr,
+      get_multipart_content_provider(boundary, items, provider_items),
+      content_type);
+}
+inline Result ClientImpl::Patch(const std::string &path) {
+  return Patch(path, std::string(), std::string());
+}
+
+inline Result ClientImpl::Patch(const std::string &path, const char *body,
+                                size_t content_length,
+                                const std::string &content_type) {
+  return Patch(path, Headers(), body, content_length, content_type);
+}
+
+inline Result ClientImpl::Patch(const std::string &path, const Headers &headers,
+                                const char *body, size_t content_length,
+                                const std::string &content_type) {
+  return send_with_content_provider("PATCH", path, headers, body,
+                                    content_length, nullptr, nullptr,
+                                    content_type);
+}
+
+inline Result ClientImpl::Patch(const std::string &path,
+                                const std::string &body,
+                                const std::string &content_type) {
+  return Patch(path, Headers(), body, content_type);
+}
+
+inline Result ClientImpl::Patch(const std::string &path, const Headers &headers,
+                                const std::string &body,
+                                const std::string &content_type) {
+  return send_with_content_provider("PATCH", path, headers, body.data(),
+                                    body.size(), nullptr, nullptr,
+                                    content_type);
+}
+
+inline Result ClientImpl::Patch(const std::string &path, size_t content_length,
+                                ContentProvider content_provider,
+                                const std::string &content_type) {
+  return Patch(path, Headers(), content_length, std::move(content_provider),
+               content_type);
+}
+
+inline Result ClientImpl::Patch(const std::string &path,
+                                ContentProviderWithoutLength content_provider,
+                                const std::string &content_type) {
+  return Patch(path, Headers(), std::move(content_provider), content_type);
+}
+
+inline Result ClientImpl::Patch(const std::string &path, const Headers &headers,
+                                size_t content_length,
+                                ContentProvider content_provider,
+                                const std::string &content_type) {
+  return send_with_content_provider("PATCH", path, headers, nullptr,
+                                    content_length, std::move(content_provider),
+                                    nullptr, content_type);
+}
+
+inline Result ClientImpl::Patch(const std::string &path, const Headers &headers,
+                                ContentProviderWithoutLength content_provider,
+                                const std::string &content_type) {
+  return send_with_content_provider("PATCH", path, headers, nullptr, 0, nullptr,
+                                    std::move(content_provider), content_type);
+}
+
+inline Result ClientImpl::Delete(const std::string &path) {
+  return Delete(path, Headers(), std::string(), std::string());
+}
+
+inline Result ClientImpl::Delete(const std::string &path,
+                                 const Headers &headers) {
+  return Delete(path, headers, std::string(), std::string());
+}
+
+inline Result ClientImpl::Delete(const std::string &path, const char *body,
+                                 size_t content_length,
+                                 const std::string &content_type) {
+  return Delete(path, Headers(), body, content_length, content_type);
+}
+
+inline Result ClientImpl::Delete(const std::string &path,
+                                 const Headers &headers, const char *body,
+                                 size_t content_length,
+                                 const std::string &content_type) {
+  Request req;
+  req.method = "DELETE";
+  req.headers = headers;
+  req.path = path;
+
+  if (!content_type.empty()) { req.set_header("Content-Type", content_type); }
+  req.body.assign(body, content_length);
+
+  return send_(std::move(req));
+}
+
+inline Result ClientImpl::Delete(const std::string &path,
+                                 const std::string &body,
+                                 const std::string &content_type) {
+  return Delete(path, Headers(), body.data(), body.size(), content_type);
+}
+
+inline Result ClientImpl::Delete(const std::string &path,
+                                 const Headers &headers,
+                                 const std::string &body,
+                                 const std::string &content_type) {
+  return Delete(path, headers, body.data(), body.size(), content_type);
+}
+
+inline Result ClientImpl::Options(const std::string &path) {
+  return Options(path, Headers());
+}
+
+inline Result ClientImpl::Options(const std::string &path,
+                                  const Headers &headers) {
+  Request req;
+  req.method = "OPTIONS";
+  req.headers = headers;
+  req.path = path;
+
+  return send_(std::move(req));
+}
+
+inline void ClientImpl::stop() {
+  std::lock_guard<std::mutex> guard(socket_mutex_);
+
+  // If there is anything ongoing right now, the ONLY thread-safe thing we can
+  // do is to shutdown_socket, so that threads using this socket suddenly
+  // discover they can't read/write any more and error out. Everything else
+  // (closing the socket, shutting ssl down) is unsafe because these actions are
+  // not thread-safe.
+  if (socket_requests_in_flight_ > 0) {
+    shutdown_socket(socket_);
+
+    // Aside from that, we set a flag for the socket to be closed when we're
+    // done.
+    socket_should_be_closed_when_request_is_done_ = true;
+    return;
+  }
+
+  // Otherwise, still holding the mutex, we can shut everything down ourselves
+  shutdown_ssl(socket_, true);
+  shutdown_socket(socket_);
+  close_socket(socket_);
+}
+
+inline std::string ClientImpl::host() const { return host_; }
+
+inline int ClientImpl::port() const { return port_; }
+
+inline size_t ClientImpl::is_socket_open() const {
+  std::lock_guard<std::mutex> guard(socket_mutex_);
+  return socket_.is_open();
+}
+
+inline socket_t ClientImpl::socket() const { return socket_.sock; }
+
+inline void ClientImpl::set_connection_timeout(time_t sec, time_t usec) {
+  connection_timeout_sec_ = sec;
+  connection_timeout_usec_ = usec;
+}
+
+inline void ClientImpl::set_read_timeout(time_t sec, time_t usec) {
+  read_timeout_sec_ = sec;
+  read_timeout_usec_ = usec;
+}
+
+inline void ClientImpl::set_write_timeout(time_t sec, time_t usec) {
+  write_timeout_sec_ = sec;
+  write_timeout_usec_ = usec;
+}
+
+inline void ClientImpl::set_basic_auth(const std::string &username,
+                                       const std::string &password) {
+  basic_auth_username_ = username;
+  basic_auth_password_ = password;
+}
+
+inline void ClientImpl::set_bearer_token_auth(const std::string &token) {
+  bearer_token_auth_token_ = token;
+}
+
+#ifdef CPPHTTPLIB_OPENSSL_SUPPORT
+inline void ClientImpl::set_digest_auth(const std::string &username,
+                                        const std::string &password) {
+  digest_auth_username_ = username;
+  digest_auth_password_ = password;
+}
+#endif
+
+inline void ClientImpl::set_keep_alive(bool on) { keep_alive_ = on; }
+
+inline void ClientImpl::set_follow_location(bool on) { follow_location_ = on; }
+
+inline void ClientImpl::set_url_encode(bool on) { url_encode_ = on; }
+
+inline void
+ClientImpl::set_hostname_addr_map(std::map<std::string, std::string> addr_map) {
+  addr_map_ = std::move(addr_map);
+}
+
+inline void ClientImpl::set_default_headers(Headers headers) {
+  default_headers_ = std::move(headers);
+}
+
+inline void ClientImpl::set_header_writer(
+    std::function<ssize_t(Stream &, Headers &)> const &writer) {
+  header_writer_ = writer;
+}
+
+inline void ClientImpl::set_address_family(int family) {
+  address_family_ = family;
+}
+
+inline void ClientImpl::set_tcp_nodelay(bool on) { tcp_nodelay_ = on; }
+
+inline void ClientImpl::set_socket_options(SocketOptions socket_options) {
+  socket_options_ = std::move(socket_options);
+}
+
+inline void ClientImpl::set_compress(bool on) { compress_ = on; }
+
+inline void ClientImpl::set_decompress(bool on) { decompress_ = on; }
+
+inline void ClientImpl::set_interface(const std::string &intf) {
+  interface_ = intf;
+}
+
+inline void ClientImpl::set_proxy(const std::string &host, int port) {
+  proxy_host_ = host;
+  proxy_port_ = port;
+}
+
+inline void ClientImpl::set_proxy_basic_auth(const std::string &username,
+                                             const std::string &password) {
+  proxy_basic_auth_username_ = username;
+  proxy_basic_auth_password_ = password;
+}
+
+inline void ClientImpl::set_proxy_bearer_token_auth(const std::string &token) {
+  proxy_bearer_token_auth_token_ = token;
+}
+
+#ifdef CPPHTTPLIB_OPENSSL_SUPPORT
+inline void ClientImpl::set_proxy_digest_auth(const std::string &username,
+                                              const std::string &password) {
+  proxy_digest_auth_username_ = username;
+  proxy_digest_auth_password_ = password;
+}
+
+inline void ClientImpl::set_ca_cert_path(const std::string &ca_cert_file_path,
+                                         const std::string &ca_cert_dir_path) {
+  ca_cert_file_path_ = ca_cert_file_path;
+  ca_cert_dir_path_ = ca_cert_dir_path;
+}
+
+inline void ClientImpl::set_ca_cert_store(X509_STORE *ca_cert_store) {
+  if (ca_cert_store && ca_cert_store != ca_cert_store_) {
+    ca_cert_store_ = ca_cert_store;
+  }
+}
+
+inline X509_STORE *ClientImpl::create_ca_cert_store(const char *ca_cert,
+                                                    std::size_t size) {
+  auto mem = BIO_new_mem_buf(ca_cert, static_cast<int>(size));
+  if (!mem) return nullptr;
+
+  auto inf = PEM_X509_INFO_read_bio(mem, nullptr, nullptr, nullptr);
+  if (!inf) {
+    BIO_free_all(mem);
+    return nullptr;
+  }
+
+  auto cts = X509_STORE_new();
+  if (cts) {
+    for (auto i = 0; i < static_cast<int>(sk_X509_INFO_num(inf)); i++) {
+      auto itmp = sk_X509_INFO_value(inf, i);
+      if (!itmp) { continue; }
+
+      if (itmp->x509) { X509_STORE_add_cert(cts, itmp->x509); }
+      if (itmp->crl) { X509_STORE_add_crl(cts, itmp->crl); }
+    }
+  }
+
+  sk_X509_INFO_pop_free(inf, X509_INFO_free);
+  BIO_free_all(mem);
+  return cts;
+}
+
+inline void ClientImpl::enable_server_certificate_verification(bool enabled) {
+  server_certificate_verification_ = enabled;
+}
+#endif
+
+inline void ClientImpl::set_logger(Logger logger) {
+  logger_ = std::move(logger);
+}
+
+/*
+ * SSL Implementation
+ */
+#ifdef CPPHTTPLIB_OPENSSL_SUPPORT
+namespace detail {
+
+template <typename U, typename V>
+inline SSL *ssl_new(socket_t sock, SSL_CTX *ctx, std::mutex &ctx_mutex,
+                    U SSL_connect_or_accept, V setup) {
+  SSL *ssl = nullptr;
+  {
+    std::lock_guard<std::mutex> guard(ctx_mutex);
+    ssl = SSL_new(ctx);
+  }
+
+  if (ssl) {
+    set_nonblocking(sock, true);
+    auto bio = BIO_new_socket(static_cast<int>(sock), BIO_NOCLOSE);
+    BIO_set_nbio(bio, 1);
+    SSL_set_bio(ssl, bio, bio);
+
+    if (!setup(ssl) || SSL_connect_or_accept(ssl) != 1) {
+      SSL_shutdown(ssl);
+      {
+        std::lock_guard<std::mutex> guard(ctx_mutex);
+        SSL_free(ssl);
+      }
+      set_nonblocking(sock, false);
+      return nullptr;
+    }
+    BIO_set_nbio(bio, 0);
+    set_nonblocking(sock, false);
+  }
+
+  return ssl;
+}
+
+inline void ssl_delete(std::mutex &ctx_mutex, SSL *ssl,
+                       bool shutdown_gracefully) {
+  // sometimes we may want to skip this to try to avoid SIGPIPE if we know
+  // the remote has closed the network connection
+  // Note that it is not always possible to avoid SIGPIPE, this is merely a
+  // best-efforts.
+  if (shutdown_gracefully) { SSL_shutdown(ssl); }
+
+  std::lock_guard<std::mutex> guard(ctx_mutex);
+  SSL_free(ssl);
+}
+
+template <typename U>
+bool ssl_connect_or_accept_nonblocking(socket_t sock, SSL *ssl,
+                                       U ssl_connect_or_accept,
+                                       time_t timeout_sec,
+                                       time_t timeout_usec) {
+  auto res = 0;
+  while ((res = ssl_connect_or_accept(ssl)) != 1) {
+    auto err = SSL_get_error(ssl, res);
+    switch (err) {
+    case SSL_ERROR_WANT_READ:
+      if (select_read(sock, timeout_sec, timeout_usec) > 0) { continue; }
+      break;
+    case SSL_ERROR_WANT_WRITE:
+      if (select_write(sock, timeout_sec, timeout_usec) > 0) { continue; }
+      break;
+    default: break;
+    }
+    return false;
+  }
+  return true;
+}
+
+template <typename T>
+inline bool process_server_socket_ssl(
+    const std::atomic<socket_t> &svr_sock, SSL *ssl, socket_t sock,
+    size_t keep_alive_max_count, time_t keep_alive_timeout_sec,
+    time_t read_timeout_sec, time_t read_timeout_usec, time_t write_timeout_sec,
+    time_t write_timeout_usec, T callback) {
+  return process_server_socket_core(
+      svr_sock, sock, keep_alive_max_count, keep_alive_timeout_sec,
+      [&](bool close_connection, bool &connection_closed) {
+        SSLSocketStream strm(sock, ssl, read_timeout_sec, read_timeout_usec,
+                             write_timeout_sec, write_timeout_usec);
+        return callback(strm, close_connection, connection_closed);
+      });
+}
+
+template <typename T>
+inline bool
+process_client_socket_ssl(SSL *ssl, socket_t sock, time_t read_timeout_sec,
+                          time_t read_timeout_usec, time_t write_timeout_sec,
+                          time_t write_timeout_usec, T callback) {
+  SSLSocketStream strm(sock, ssl, read_timeout_sec, read_timeout_usec,
+                       write_timeout_sec, write_timeout_usec);
+  return callback(strm);
+}
+
+class SSLInit {
+public:
+  SSLInit() {
+    OPENSSL_init_ssl(
+        OPENSSL_INIT_LOAD_SSL_STRINGS | OPENSSL_INIT_LOAD_CRYPTO_STRINGS, NULL);
+  }
+};
+
+// SSL socket stream implementation
+inline SSLSocketStream::SSLSocketStream(socket_t sock, SSL *ssl,
+                                        time_t read_timeout_sec,
+                                        time_t read_timeout_usec,
+                                        time_t write_timeout_sec,
+                                        time_t write_timeout_usec)
+    : sock_(sock), ssl_(ssl), read_timeout_sec_(read_timeout_sec),
+      read_timeout_usec_(read_timeout_usec),
+      write_timeout_sec_(write_timeout_sec),
+      write_timeout_usec_(write_timeout_usec) {
+  SSL_clear_mode(ssl, SSL_MODE_AUTO_RETRY);
+}
+
+inline SSLSocketStream::~SSLSocketStream() {}
+
+inline bool SSLSocketStream::is_readable() const {
+  return detail::select_read(sock_, read_timeout_sec_, read_timeout_usec_) > 0;
+}
+
+inline bool SSLSocketStream::is_writable() const {
+  return select_write(sock_, write_timeout_sec_, write_timeout_usec_) > 0 &&
+         is_socket_alive(sock_);
+}
+
+inline ssize_t SSLSocketStream::read(char *ptr, size_t size) {
+  if (SSL_pending(ssl_) > 0) {
+    return SSL_read(ssl_, ptr, static_cast<int>(size));
+  } else if (is_readable()) {
+    auto ret = SSL_read(ssl_, ptr, static_cast<int>(size));
+    if (ret < 0) {
+      auto err = SSL_get_error(ssl_, ret);
+      auto n = 1000;
+#ifdef _WIN32
+      while (--n >= 0 && (err == SSL_ERROR_WANT_READ ||
+                          (err == SSL_ERROR_SYSCALL &&
+                           WSAGetLastError() == WSAETIMEDOUT))) {
+#else
+      while (--n >= 0 && err == SSL_ERROR_WANT_READ) {
+#endif
+        if (SSL_pending(ssl_) > 0) {
+          return SSL_read(ssl_, ptr, static_cast<int>(size));
+        } else if (is_readable()) {
+          std::this_thread::sleep_for(std::chrono::milliseconds(1));
+          ret = SSL_read(ssl_, ptr, static_cast<int>(size));
+          if (ret >= 0) { return ret; }
+          err = SSL_get_error(ssl_, ret);
+        } else {
+          return -1;
+        }
+      }
+    }
+    return ret;
+  }
+  return -1;
+}
+
+inline ssize_t SSLSocketStream::write(const char *ptr, size_t size) {
+  if (is_writable()) {
+    auto handle_size = static_cast<int>(
+        std::min<size_t>(size, (std::numeric_limits<int>::max)()));
+
+    auto ret = SSL_write(ssl_, ptr, static_cast<int>(handle_size));
+    if (ret < 0) {
+      auto err = SSL_get_error(ssl_, ret);
+      auto n = 1000;
+#ifdef _WIN32
+      while (--n >= 0 && (err == SSL_ERROR_WANT_WRITE ||
+                          (err == SSL_ERROR_SYSCALL &&
+                           WSAGetLastError() == WSAETIMEDOUT))) {
+#else
+      while (--n >= 0 && err == SSL_ERROR_WANT_WRITE) {
+#endif
+        if (is_writable()) {
+          std::this_thread::sleep_for(std::chrono::milliseconds(1));
+          ret = SSL_write(ssl_, ptr, static_cast<int>(handle_size));
+          if (ret >= 0) { return ret; }
+          err = SSL_get_error(ssl_, ret);
+        } else {
+          return -1;
+        }
+      }
+    }
+    return ret;
+  }
+  return -1;
+}
+
+inline void SSLSocketStream::get_remote_ip_and_port(std::string &ip,
+                                                    int &port) const {
+  detail::get_remote_ip_and_port(sock_, ip, port);
+}
+
+inline void SSLSocketStream::get_local_ip_and_port(std::string &ip,
+                                                   int &port) const {
+  detail::get_local_ip_and_port(sock_, ip, port);
+}
+
+inline socket_t SSLSocketStream::socket() const { return sock_; }
+
+static SSLInit sslinit_;
+
+} // namespace detail
+
+// SSL HTTP server implementation
+inline SSLServer::SSLServer(const char *cert_path, const char *private_key_path,
+                            const char *client_ca_cert_file_path,
+                            const char *client_ca_cert_dir_path,
+                            const char *private_key_password) {
+  ctx_ = SSL_CTX_new(TLS_server_method());
+
+  if (ctx_) {
+    SSL_CTX_set_options(ctx_,
+                        SSL_OP_NO_COMPRESSION |
+                            SSL_OP_NO_SESSION_RESUMPTION_ON_RENEGOTIATION);
+
+    SSL_CTX_set_min_proto_version(ctx_, TLS1_1_VERSION);
+
+    // add default password callback before opening encrypted private key
+    if (private_key_password != nullptr && (private_key_password[0] != '\0')) {
+      SSL_CTX_set_default_passwd_cb_userdata(
+          ctx_,
+          reinterpret_cast<void *>(const_cast<char *>(private_key_password)));
+    }
+
+    if (SSL_CTX_use_certificate_chain_file(ctx_, cert_path) != 1 ||
+        SSL_CTX_use_PrivateKey_file(ctx_, private_key_path, SSL_FILETYPE_PEM) !=
+            1) {
+      SSL_CTX_free(ctx_);
+      ctx_ = nullptr;
+    } else if (client_ca_cert_file_path || client_ca_cert_dir_path) {
+      SSL_CTX_load_verify_locations(ctx_, client_ca_cert_file_path,
+                                    client_ca_cert_dir_path);
+
+      SSL_CTX_set_verify(
+          ctx_, SSL_VERIFY_PEER | SSL_VERIFY_FAIL_IF_NO_PEER_CERT, nullptr);
+    }
+  }
+}
+
+inline SSLServer::SSLServer(X509 *cert, EVP_PKEY *private_key,
+                            X509_STORE *client_ca_cert_store) {
+  ctx_ = SSL_CTX_new(TLS_server_method());
+
+  if (ctx_) {
+    SSL_CTX_set_options(ctx_,
+                        SSL_OP_NO_COMPRESSION |
+                            SSL_OP_NO_SESSION_RESUMPTION_ON_RENEGOTIATION);
+
+    SSL_CTX_set_min_proto_version(ctx_, TLS1_1_VERSION);
+
+    if (SSL_CTX_use_certificate(ctx_, cert) != 1 ||
+        SSL_CTX_use_PrivateKey(ctx_, private_key) != 1) {
+      SSL_CTX_free(ctx_);
+      ctx_ = nullptr;
+    } else if (client_ca_cert_store) {
+      SSL_CTX_set_cert_store(ctx_, client_ca_cert_store);
+
+      SSL_CTX_set_verify(
+          ctx_, SSL_VERIFY_PEER | SSL_VERIFY_FAIL_IF_NO_PEER_CERT, nullptr);
+    }
+  }
+}
+
+inline SSLServer::SSLServer(
+    const std::function<bool(SSL_CTX &ssl_ctx)> &setup_ssl_ctx_callback) {
+  ctx_ = SSL_CTX_new(TLS_method());
+  if (ctx_) {
+    if (!setup_ssl_ctx_callback(*ctx_)) {
+      SSL_CTX_free(ctx_);
+      ctx_ = nullptr;
+    }
+  }
+}
+
+inline SSLServer::~SSLServer() {
+  if (ctx_) { SSL_CTX_free(ctx_); }
+}
+
+inline bool SSLServer::is_valid() const { return ctx_; }
+
+inline SSL_CTX *SSLServer::ssl_context() const { return ctx_; }
+
+inline bool SSLServer::process_and_close_socket(socket_t sock) {
+  auto ssl = detail::ssl_new(
+      sock, ctx_, ctx_mutex_,
+      [&](SSL *ssl2) {
+        return detail::ssl_connect_or_accept_nonblocking(
+            sock, ssl2, SSL_accept, read_timeout_sec_, read_timeout_usec_);
+      },
+      [](SSL * /*ssl2*/) { return true; });
+
+  auto ret = false;
+  if (ssl) {
+    ret = detail::process_server_socket_ssl(
+        svr_sock_, ssl, sock, keep_alive_max_count_, keep_alive_timeout_sec_,
+        read_timeout_sec_, read_timeout_usec_, write_timeout_sec_,
+        write_timeout_usec_,
+        [this, ssl](Stream &strm, bool close_connection,
+                    bool &connection_closed) {
+          return process_request(strm, close_connection, connection_closed,
+                                 [&](Request &req) { req.ssl = ssl; });
+        });
+
+    // Shutdown gracefully if the result seemed successful, non-gracefully if
+    // the connection appeared to be closed.
+    const bool shutdown_gracefully = ret;
+    detail::ssl_delete(ctx_mutex_, ssl, shutdown_gracefully);
+  }
+
+  detail::shutdown_socket(sock);
+  detail::close_socket(sock);
+  return ret;
+}
+
+// SSL HTTP client implementation
+inline SSLClient::SSLClient(const std::string &host)
+    : SSLClient(host, 443, std::string(), std::string()) {}
+
+inline SSLClient::SSLClient(const std::string &host, int port)
+    : SSLClient(host, port, std::string(), std::string()) {}
+
+inline SSLClient::SSLClient(const std::string &host, int port,
+                            const std::string &client_cert_path,
+                            const std::string &client_key_path)
+    : ClientImpl(host, port, client_cert_path, client_key_path) {
+  ctx_ = SSL_CTX_new(TLS_client_method());
+
+  detail::split(&host_[0], &host_[host_.size()], '.',
+                [&](const char *b, const char *e) {
+                  host_components_.emplace_back(std::string(b, e));
+                });
+
+  if (!client_cert_path.empty() && !client_key_path.empty()) {
+    if (SSL_CTX_use_certificate_file(ctx_, client_cert_path.c_str(),
+                                     SSL_FILETYPE_PEM) != 1 ||
+        SSL_CTX_use_PrivateKey_file(ctx_, client_key_path.c_str(),
+                                    SSL_FILETYPE_PEM) != 1) {
+      SSL_CTX_free(ctx_);
+      ctx_ = nullptr;
+    }
+  }
+}
+
+inline SSLClient::SSLClient(const std::string &host, int port,
+                            X509 *client_cert, EVP_PKEY *client_key)
+    : ClientImpl(host, port) {
+  ctx_ = SSL_CTX_new(TLS_client_method());
+
+  detail::split(&host_[0], &host_[host_.size()], '.',
+                [&](const char *b, const char *e) {
+                  host_components_.emplace_back(std::string(b, e));
+                });
+
+  if (client_cert != nullptr && client_key != nullptr) {
+    if (SSL_CTX_use_certificate(ctx_, client_cert) != 1 ||
+        SSL_CTX_use_PrivateKey(ctx_, client_key) != 1) {
+      SSL_CTX_free(ctx_);
+      ctx_ = nullptr;
+    }
+  }
+}
+
+inline SSLClient::~SSLClient() {
+  if (ctx_) { SSL_CTX_free(ctx_); }
+  // Make sure to shut down SSL since shutdown_ssl will resolve to the
+  // base function rather than the derived function once we get to the
+  // base class destructor, and won't free the SSL (causing a leak).
+  shutdown_ssl_impl(socket_, true);
+}
+
+inline bool SSLClient::is_valid() const { return ctx_; }
+
+inline void SSLClient::set_ca_cert_store(X509_STORE *ca_cert_store) {
+  if (ca_cert_store) {
+    if (ctx_) {
+      if (SSL_CTX_get_cert_store(ctx_) != ca_cert_store) {
+        // Free memory allocated for old cert and use new store `ca_cert_store`
+        SSL_CTX_set_cert_store(ctx_, ca_cert_store);
+      }
+    } else {
+      X509_STORE_free(ca_cert_store);
+    }
+  }
+}
+
+inline void SSLClient::load_ca_cert_store(const char *ca_cert,
+                                          std::size_t size) {
+  set_ca_cert_store(ClientImpl::create_ca_cert_store(ca_cert, size));
+}
+
+inline long SSLClient::get_openssl_verify_result() const {
+  return verify_result_;
+}
+
+inline SSL_CTX *SSLClient::ssl_context() const { return ctx_; }
+
+inline bool SSLClient::create_and_connect_socket(Socket &socket, Error &error) {
+  return is_valid() && ClientImpl::create_and_connect_socket(socket, error);
+}
+
+// Assumes that socket_mutex_ is locked and that there are no requests in flight
+inline bool SSLClient::connect_with_proxy(Socket &socket, Response &res,
+                                          bool &success, Error &error) {
+  success = true;
+  Response proxy_res;
+  if (!detail::process_client_socket(
+          socket.sock, read_timeout_sec_, read_timeout_usec_,
+          write_timeout_sec_, write_timeout_usec_, [&](Stream &strm) {
+            Request req2;
+            req2.method = "CONNECT";
+            req2.path = host_and_port_;
+            return process_request(strm, req2, proxy_res, false, error);
+          })) {
+    // Thread-safe to close everything because we are assuming there are no
+    // requests in flight
+    shutdown_ssl(socket, true);
+    shutdown_socket(socket);
+    close_socket(socket);
+    success = false;
+    return false;
+  }
+
+  if (proxy_res.status == 407) {
+    if (!proxy_digest_auth_username_.empty() &&
+        !proxy_digest_auth_password_.empty()) {
+      std::map<std::string, std::string> auth;
+      if (detail::parse_www_authenticate(proxy_res, auth, true)) {
+        proxy_res = Response();
+        if (!detail::process_client_socket(
+                socket.sock, read_timeout_sec_, read_timeout_usec_,
+                write_timeout_sec_, write_timeout_usec_, [&](Stream &strm) {
+                  Request req3;
+                  req3.method = "CONNECT";
+                  req3.path = host_and_port_;
+                  req3.headers.insert(detail::make_digest_authentication_header(
+                      req3, auth, 1, detail::random_string(10),
+                      proxy_digest_auth_username_, proxy_digest_auth_password_,
+                      true));
+                  return process_request(strm, req3, proxy_res, false, error);
+                })) {
+          // Thread-safe to close everything because we are assuming there are
+          // no requests in flight
+          shutdown_ssl(socket, true);
+          shutdown_socket(socket);
+          close_socket(socket);
+          success = false;
+          return false;
+        }
+      }
+    }
+  }
+
+  // If status code is not 200, proxy request is failed.
+  // Set error to ProxyConnection and return proxy response
+  // as the response of the request
+  if (proxy_res.status != 200) {
+    error = Error::ProxyConnection;
+    res = std::move(proxy_res);
+    // Thread-safe to close everything because we are assuming there are
+    // no requests in flight
+    shutdown_ssl(socket, true);
+    shutdown_socket(socket);
+    close_socket(socket);
+    return false;
+  }
+
+  return true;
+}
+
+inline bool SSLClient::load_certs() {
+  auto ret = true;
+
+  std::call_once(initialize_cert_, [&]() {
+    std::lock_guard<std::mutex> guard(ctx_mutex_);
+    if (!ca_cert_file_path_.empty()) {
+      if (!SSL_CTX_load_verify_locations(ctx_, ca_cert_file_path_.c_str(),
+                                         nullptr)) {
+        ret = false;
+      }
+    } else if (!ca_cert_dir_path_.empty()) {
+      if (!SSL_CTX_load_verify_locations(ctx_, nullptr,
+                                         ca_cert_dir_path_.c_str())) {
+        ret = false;
+      }
+    } else {
+      auto loaded = false;
+#ifdef _WIN32
+      loaded =
+          detail::load_system_certs_on_windows(SSL_CTX_get_cert_store(ctx_));
+#elif defined(CPPHTTPLIB_USE_CERTS_FROM_MACOSX_KEYCHAIN) && defined(__APPLE__)
+#if TARGET_OS_OSX
+      loaded = detail::load_system_certs_on_macos(SSL_CTX_get_cert_store(ctx_));
+#endif // TARGET_OS_OSX
+#endif // _WIN32
+      if (!loaded) { SSL_CTX_set_default_verify_paths(ctx_); }
+    }
+  });
+
+  return ret;
+}
+
+inline bool SSLClient::initialize_ssl(Socket &socket, Error &error) {
+  auto ssl = detail::ssl_new(
+      socket.sock, ctx_, ctx_mutex_,
+      [&](SSL *ssl2) {
+        if (server_certificate_verification_) {
+          if (!load_certs()) {
+            error = Error::SSLLoadingCerts;
+            return false;
+          }
+          SSL_set_verify(ssl2, SSL_VERIFY_NONE, nullptr);
+        }
+
+        if (!detail::ssl_connect_or_accept_nonblocking(
+                socket.sock, ssl2, SSL_connect, connection_timeout_sec_,
+                connection_timeout_usec_)) {
+          error = Error::SSLConnection;
+          return false;
+        }
+
+        if (server_certificate_verification_) {
+          verify_result_ = SSL_get_verify_result(ssl2);
+
+          if (verify_result_ != X509_V_OK) {
+            error = Error::SSLServerVerification;
+            return false;
+          }
+
+          auto server_cert = SSL_get1_peer_certificate(ssl2);
+
+          if (server_cert == nullptr) {
+            error = Error::SSLServerVerification;
+            return false;
+          }
+
+          if (!verify_host(server_cert)) {
+            X509_free(server_cert);
+            error = Error::SSLServerVerification;
+            return false;
+          }
+          X509_free(server_cert);
+        }
+
+        return true;
+      },
+      [&](SSL *ssl2) {
+        // NOTE: With -Wold-style-cast, this can produce a warning, since
+        //  SSL_set_tlsext_host_name is a macro (in OpenSSL), which contains
+        //  an old style cast. Short of doing compiler specific pragma's
+        //  here, we can't get rid of this warning. :'(
+        SSL_set_tlsext_host_name(ssl2, host_.c_str());
+        return true;
+      });
+
+  if (ssl) {
+    socket.ssl = ssl;
+    return true;
+  }
+
+  shutdown_socket(socket);
+  close_socket(socket);
+  return false;
+}
+
+inline void SSLClient::shutdown_ssl(Socket &socket, bool shutdown_gracefully) {
+  shutdown_ssl_impl(socket, shutdown_gracefully);
+}
+
+inline void SSLClient::shutdown_ssl_impl(Socket &socket,
+                                         bool shutdown_gracefully) {
+  if (socket.sock == INVALID_SOCKET) {
+    assert(socket.ssl == nullptr);
+    return;
+  }
+  if (socket.ssl) {
+    detail::ssl_delete(ctx_mutex_, socket.ssl, shutdown_gracefully);
+    socket.ssl = nullptr;
+  }
+  assert(socket.ssl == nullptr);
+}
+
+inline bool
+SSLClient::process_socket(const Socket &socket,
+                          std::function<bool(Stream &strm)> callback) {
+  assert(socket.ssl);
+  return detail::process_client_socket_ssl(
+      socket.ssl, socket.sock, read_timeout_sec_, read_timeout_usec_,
+      write_timeout_sec_, write_timeout_usec_, std::move(callback));
+}
+
+inline bool SSLClient::is_ssl() const { return true; }
+
+inline bool SSLClient::verify_host(X509 *server_cert) const {
+  /* Quote from RFC2818 section 3.1 "Server Identity"
+
+     If a subjectAltName extension of type dNSName is present, that MUST
+     be used as the identity. Otherwise, the (most specific) Common Name
+     field in the Subject field of the certificate MUST be used. Although
+     the use of the Common Name is existing practice, it is deprecated and
+     Certification Authorities are encouraged to use the dNSName instead.
+
+     Matching is performed using the matching rules specified by
+     [RFC2459].  If more than one identity of a given type is present in
+     the certificate (e.g., more than one dNSName name, a match in any one
+     of the set is considered acceptable.) Names may contain the wildcard
+     character * which is considered to match any single domain name
+     component or component fragment. E.g., *.a.com matches foo.a.com but
+     not bar.foo.a.com. f*.com matches foo.com but not bar.com.
+
+     In some cases, the URI is specified as an IP address rather than a
+     hostname. In this case, the iPAddress subjectAltName must be present
+     in the certificate and must exactly match the IP in the URI.
+
+  */
+  return verify_host_with_subject_alt_name(server_cert) ||
+         verify_host_with_common_name(server_cert);
+}
+
+inline bool
+SSLClient::verify_host_with_subject_alt_name(X509 *server_cert) const {
+  auto ret = false;
+
+  auto type = GEN_DNS;
+
+  struct in6_addr addr6;
+  struct in_addr addr;
+  size_t addr_len = 0;
+
+#ifndef __MINGW32__
+  if (inet_pton(AF_INET6, host_.c_str(), &addr6)) {
+    type = GEN_IPADD;
+    addr_len = sizeof(struct in6_addr);
+  } else if (inet_pton(AF_INET, host_.c_str(), &addr)) {
+    type = GEN_IPADD;
+    addr_len = sizeof(struct in_addr);
+  }
+#endif
+
+  auto alt_names = static_cast<const struct stack_st_GENERAL_NAME *>(
+      X509_get_ext_d2i(server_cert, NID_subject_alt_name, nullptr, nullptr));
+
+  if (alt_names) {
+    auto dsn_matched = false;
+    auto ip_matched = false;
+
+    auto count = sk_GENERAL_NAME_num(alt_names);
+
+    for (decltype(count) i = 0; i < count && !dsn_matched; i++) {
+      auto val = sk_GENERAL_NAME_value(alt_names, i);
+      if (val->type == type) {
+        auto name =
+            reinterpret_cast<const char *>(ASN1_STRING_get0_data(val->d.ia5));
+        auto name_len = static_cast<size_t>(ASN1_STRING_length(val->d.ia5));
+
+        switch (type) {
+        case GEN_DNS: dsn_matched = check_host_name(name, name_len); break;
+
+        case GEN_IPADD:
+          if (!memcmp(&addr6, name, addr_len) ||
+              !memcmp(&addr, name, addr_len)) {
+            ip_matched = true;
+          }
+          break;
+        }
+      }
+    }
+
+    if (dsn_matched || ip_matched) { ret = true; }
+  }
+
+  GENERAL_NAMES_free(const_cast<STACK_OF(GENERAL_NAME) *>(
+      reinterpret_cast<const STACK_OF(GENERAL_NAME) *>(alt_names)));
+  return ret;
+}
+
+inline bool SSLClient::verify_host_with_common_name(X509 *server_cert) const {
+  const auto subject_name = X509_get_subject_name(server_cert);
+
+  if (subject_name != nullptr) {
+    char name[BUFSIZ];
+    auto name_len = X509_NAME_get_text_by_NID(subject_name, NID_commonName,
+                                              name, sizeof(name));
+
+    if (name_len != -1) {
+      return check_host_name(name, static_cast<size_t>(name_len));
+    }
+  }
+
+  return false;
+}
+
+inline bool SSLClient::check_host_name(const char *pattern,
+                                       size_t pattern_len) const {
+  if (host_.size() == pattern_len && host_ == pattern) { return true; }
+
+  // Wildcard match
+  // https://bugs.launchpad.net/ubuntu/+source/firefox-3.0/+bug/376484
+  std::vector<std::string> pattern_components;
+  detail::split(&pattern[0], &pattern[pattern_len], '.',
+                [&](const char *b, const char *e) {
+                  pattern_components.emplace_back(std::string(b, e));
+                });
+
+  if (host_components_.size() != pattern_components.size()) { return false; }
+
+  auto itr = pattern_components.begin();
+  for (const auto &h : host_components_) {
+    auto &p = *itr;
+    if (p != h && p != "*") {
+      auto partial_match = (p.size() > 0 && p[p.size() - 1] == '*' &&
+                            !p.compare(0, p.size() - 1, h));
+      if (!partial_match) { return false; }
+    }
+    ++itr;
+  }
+
+  return true;
+}
+#endif
+
+// Universal client implementation
+inline Client::Client(const std::string &scheme_host_port)
+    : Client(scheme_host_port, std::string(), std::string()) {}
+
+inline Client::Client(const std::string &scheme_host_port,
+                      const std::string &client_cert_path,
+                      const std::string &client_key_path) {
+  const static std::regex re(
+      R"((?:([a-z]+):\/\/)?(?:\[([\d:]+)\]|([^:/?#]+))(?::(\d+))?)");
+
+  std::smatch m;
+  if (std::regex_match(scheme_host_port, m, re)) {
+    auto scheme = m[1].str();
+
+#ifdef CPPHTTPLIB_OPENSSL_SUPPORT
+    if (!scheme.empty() && (scheme != "http" && scheme != "https")) {
+#else
+    if (!scheme.empty() && scheme != "http") {
+#endif
+#ifndef CPPHTTPLIB_NO_EXCEPTIONS
+      std::string msg = "'" + scheme + "' scheme is not supported.";
+      throw std::invalid_argument(msg);
+#endif
+      return;
+    }
+
+    auto is_ssl = scheme == "https";
+
+    auto host = m[2].str();
+    if (host.empty()) { host = m[3].str(); }
+
+    auto port_str = m[4].str();
+    auto port = !port_str.empty() ? std::stoi(port_str) : (is_ssl ? 443 : 80);
+
+    if (is_ssl) {
+#ifdef CPPHTTPLIB_OPENSSL_SUPPORT
+      cli_ = detail::make_unique<SSLClient>(host, port, client_cert_path,
+                                            client_key_path);
+      is_ssl_ = is_ssl;
+#endif
+    } else {
+      cli_ = detail::make_unique<ClientImpl>(host, port, client_cert_path,
+                                             client_key_path);
+    }
+  } else {
+    cli_ = detail::make_unique<ClientImpl>(scheme_host_port, 80,
+                                           client_cert_path, client_key_path);
+  }
+}
+
+inline Client::Client(const std::string &host, int port)
+    : cli_(detail::make_unique<ClientImpl>(host, port)) {}
+
+inline Client::Client(const std::string &host, int port,
+                      const std::string &client_cert_path,
+                      const std::string &client_key_path)
+    : cli_(detail::make_unique<ClientImpl>(host, port, client_cert_path,
+                                           client_key_path)) {}
+
+inline Client::~Client() {}
+
+inline bool Client::is_valid() const {
+  return cli_ != nullptr && cli_->is_valid();
+}
+
+inline Result Client::Get(const std::string &path) { return cli_->Get(path); }
+inline Result Client::Get(const std::string &path, const Headers &headers) {
+  return cli_->Get(path, headers);
+}
+inline Result Client::Get(const std::string &path, Progress progress) {
+  return cli_->Get(path, std::move(progress));
+}
+inline Result Client::Get(const std::string &path, const Headers &headers,
+                          Progress progress) {
+  return cli_->Get(path, headers, std::move(progress));
+}
+inline Result Client::Get(const std::string &path,
+                          ContentReceiver content_receiver) {
+  return cli_->Get(path, std::move(content_receiver));
+}
+inline Result Client::Get(const std::string &path, const Headers &headers,
+                          ContentReceiver content_receiver) {
+  return cli_->Get(path, headers, std::move(content_receiver));
+}
+inline Result Client::Get(const std::string &path,
+                          ContentReceiver content_receiver, Progress progress) {
+  return cli_->Get(path, std::move(content_receiver), std::move(progress));
+}
+inline Result Client::Get(const std::string &path, const Headers &headers,
+                          ContentReceiver content_receiver, Progress progress) {
+  return cli_->Get(path, headers, std::move(content_receiver),
+                   std::move(progress));
+}
+inline Result Client::Get(const std::string &path,
+                          ResponseHandler response_handler,
+                          ContentReceiver content_receiver) {
+  return cli_->Get(path, std::move(response_handler),
+                   std::move(content_receiver));
+}
+inline Result Client::Get(const std::string &path, const Headers &headers,
+                          ResponseHandler response_handler,
+                          ContentReceiver content_receiver) {
+  return cli_->Get(path, headers, std::move(response_handler),
+                   std::move(content_receiver));
+}
+inline Result Client::Get(const std::string &path,
+                          ResponseHandler response_handler,
+                          ContentReceiver content_receiver, Progress progress) {
+  return cli_->Get(path, std::move(response_handler),
+                   std::move(content_receiver), std::move(progress));
+}
+inline Result Client::Get(const std::string &path, const Headers &headers,
+                          ResponseHandler response_handler,
+                          ContentReceiver content_receiver, Progress progress) {
+  return cli_->Get(path, headers, std::move(response_handler),
+                   std::move(content_receiver), std::move(progress));
+}
+inline Result Client::Get(const std::string &path, const Params &params,
+                          const Headers &headers, Progress progress) {
+  return cli_->Get(path, params, headers, progress);
+}
+inline Result Client::Get(const std::string &path, const Params &params,
+                          const Headers &headers,
+                          ContentReceiver content_receiver, Progress progress) {
+  return cli_->Get(path, params, headers, content_receiver, progress);
+}
+inline Result Client::Get(const std::string &path, const Params &params,
+                          const Headers &headers,
+                          ResponseHandler response_handler,
+                          ContentReceiver content_receiver, Progress progress) {
+  return cli_->Get(path, params, headers, response_handler, content_receiver,
+                   progress);
+}
+
+inline Result Client::Head(const std::string &path) { return cli_->Head(path); }
+inline Result Client::Head(const std::string &path, const Headers &headers) {
+  return cli_->Head(path, headers);
+}
+
+inline Result Client::Post(const std::string &path) { return cli_->Post(path); }
+inline Result Client::Post(const std::string &path, const Headers &headers) {
+  return cli_->Post(path, headers);
+}
+inline Result Client::Post(const std::string &path, const char *body,
+                           size_t content_length,
+                           const std::string &content_type) {
+  return cli_->Post(path, body, content_length, content_type);
+}
+inline Result Client::Post(const std::string &path, const Headers &headers,
+                           const char *body, size_t content_length,
+                           const std::string &content_type) {
+  return cli_->Post(path, headers, body, content_length, content_type);
+}
+inline Result Client::Post(const std::string &path, const std::string &body,
+                           const std::string &content_type) {
+  return cli_->Post(path, body, content_type);
+}
+inline Result Client::Post(const std::string &path, const Headers &headers,
+                           const std::string &body,
+                           const std::string &content_type) {
+  return cli_->Post(path, headers, body, content_type);
+}
+inline Result Client::Post(const std::string &path, size_t content_length,
+                           ContentProvider content_provider,
+                           const std::string &content_type) {
+  return cli_->Post(path, content_length, std::move(content_provider),
+                    content_type);
+}
+inline Result Client::Post(const std::string &path,
+                           ContentProviderWithoutLength content_provider,
+                           const std::string &content_type) {
+  return cli_->Post(path, std::move(content_provider), content_type);
+}
+inline Result Client::Post(const std::string &path, const Headers &headers,
+                           size_t content_length,
+                           ContentProvider content_provider,
+                           const std::string &content_type) {
+  return cli_->Post(path, headers, content_length, std::move(content_provider),
+                    content_type);
+}
+inline Result Client::Post(const std::string &path, const Headers &headers,
+                           ContentProviderWithoutLength content_provider,
+                           const std::string &content_type) {
+  return cli_->Post(path, headers, std::move(content_provider), content_type);
+}
+inline Result Client::Post(const std::string &path, const Params &params) {
+  return cli_->Post(path, params);
+}
+inline Result Client::Post(const std::string &path, const Headers &headers,
+                           const Params &params) {
+  return cli_->Post(path, headers, params);
+}
+inline Result Client::Post(const std::string &path,
+                           const MultipartFormDataItems &items) {
+  return cli_->Post(path, items);
+}
+inline Result Client::Post(const std::string &path, const Headers &headers,
+                           const MultipartFormDataItems &items) {
+  return cli_->Post(path, headers, items);
+}
+inline Result Client::Post(const std::string &path, const Headers &headers,
+                           const MultipartFormDataItems &items,
+                           const std::string &boundary) {
+  return cli_->Post(path, headers, items, boundary);
+}
+inline Result
+Client::Post(const std::string &path, const Headers &headers,
+             const MultipartFormDataItems &items,
+             const MultipartFormDataProviderItems &provider_items) {
+  return cli_->Post(path, headers, items, provider_items);
+}
+inline Result Client::Put(const std::string &path) { return cli_->Put(path); }
+inline Result Client::Put(const std::string &path, const char *body,
+                          size_t content_length,
+                          const std::string &content_type) {
+  return cli_->Put(path, body, content_length, content_type);
+}
+inline Result Client::Put(const std::string &path, const Headers &headers,
+                          const char *body, size_t content_length,
+                          const std::string &content_type) {
+  return cli_->Put(path, headers, body, content_length, content_type);
+}
+inline Result Client::Put(const std::string &path, const std::string &body,
+                          const std::string &content_type) {
+  return cli_->Put(path, body, content_type);
+}
+inline Result Client::Put(const std::string &path, const Headers &headers,
+                          const std::string &body,
+                          const std::string &content_type) {
+  return cli_->Put(path, headers, body, content_type);
+}
+inline Result Client::Put(const std::string &path, size_t content_length,
+                          ContentProvider content_provider,
+                          const std::string &content_type) {
+  return cli_->Put(path, content_length, std::move(content_provider),
+                   content_type);
+}
+inline Result Client::Put(const std::string &path,
+                          ContentProviderWithoutLength content_provider,
+                          const std::string &content_type) {
+  return cli_->Put(path, std::move(content_provider), content_type);
+}
+inline Result Client::Put(const std::string &path, const Headers &headers,
+                          size_t content_length,
+                          ContentProvider content_provider,
+                          const std::string &content_type) {
+  return cli_->Put(path, headers, content_length, std::move(content_provider),
+                   content_type);
+}
+inline Result Client::Put(const std::string &path, const Headers &headers,
+                          ContentProviderWithoutLength content_provider,
+                          const std::string &content_type) {
+  return cli_->Put(path, headers, std::move(content_provider), content_type);
+}
+inline Result Client::Put(const std::string &path, const Params &params) {
+  return cli_->Put(path, params);
+}
+inline Result Client::Put(const std::string &path, const Headers &headers,
+                          const Params &params) {
+  return cli_->Put(path, headers, params);
+}
+inline Result Client::Put(const std::string &path,
+                          const MultipartFormDataItems &items) {
+  return cli_->Put(path, items);
+}
+inline Result Client::Put(const std::string &path, const Headers &headers,
+                          const MultipartFormDataItems &items) {
+  return cli_->Put(path, headers, items);
+}
+inline Result Client::Put(const std::string &path, const Headers &headers,
+                          const MultipartFormDataItems &items,
+                          const std::string &boundary) {
+  return cli_->Put(path, headers, items, boundary);
+}
+inline Result
+Client::Put(const std::string &path, const Headers &headers,
+            const MultipartFormDataItems &items,
+            const MultipartFormDataProviderItems &provider_items) {
+  return cli_->Put(path, headers, items, provider_items);
+}
+inline Result Client::Patch(const std::string &path) {
+  return cli_->Patch(path);
+}
+inline Result Client::Patch(const std::string &path, const char *body,
+                            size_t content_length,
+                            const std::string &content_type) {
+  return cli_->Patch(path, body, content_length, content_type);
+}
+inline Result Client::Patch(const std::string &path, const Headers &headers,
+                            const char *body, size_t content_length,
+                            const std::string &content_type) {
+  return cli_->Patch(path, headers, body, content_length, content_type);
+}
+inline Result Client::Patch(const std::string &path, const std::string &body,
+                            const std::string &content_type) {
+  return cli_->Patch(path, body, content_type);
+}
+inline Result Client::Patch(const std::string &path, const Headers &headers,
+                            const std::string &body,
+                            const std::string &content_type) {
+  return cli_->Patch(path, headers, body, content_type);
+}
+inline Result Client::Patch(const std::string &path, size_t content_length,
+                            ContentProvider content_provider,
+                            const std::string &content_type) {
+  return cli_->Patch(path, content_length, std::move(content_provider),
+                     content_type);
+}
+inline Result Client::Patch(const std::string &path,
+                            ContentProviderWithoutLength content_provider,
+                            const std::string &content_type) {
+  return cli_->Patch(path, std::move(content_provider), content_type);
+}
+inline Result Client::Patch(const std::string &path, const Headers &headers,
+                            size_t content_length,
+                            ContentProvider content_provider,
+                            const std::string &content_type) {
+  return cli_->Patch(path, headers, content_length, std::move(content_provider),
+                     content_type);
+}
+inline Result Client::Patch(const std::string &path, const Headers &headers,
+                            ContentProviderWithoutLength content_provider,
+                            const std::string &content_type) {
+  return cli_->Patch(path, headers, std::move(content_provider), content_type);
+}
+inline Result Client::Delete(const std::string &path) {
+  return cli_->Delete(path);
+}
+inline Result Client::Delete(const std::string &path, const Headers &headers) {
+  return cli_->Delete(path, headers);
+}
+inline Result Client::Delete(const std::string &path, const char *body,
+                             size_t content_length,
+                             const std::string &content_type) {
+  return cli_->Delete(path, body, content_length, content_type);
+}
+inline Result Client::Delete(const std::string &path, const Headers &headers,
+                             const char *body, size_t content_length,
+                             const std::string &content_type) {
+  return cli_->Delete(path, headers, body, content_length, content_type);
+}
+inline Result Client::Delete(const std::string &path, const std::string &body,
+                             const std::string &content_type) {
+  return cli_->Delete(path, body, content_type);
+}
+inline Result Client::Delete(const std::string &path, const Headers &headers,
+                             const std::string &body,
+                             const std::string &content_type) {
+  return cli_->Delete(path, headers, body, content_type);
+}
+inline Result Client::Options(const std::string &path) {
+  return cli_->Options(path);
+}
+inline Result Client::Options(const std::string &path, const Headers &headers) {
+  return cli_->Options(path, headers);
+}
+
+inline bool Client::send(Request &req, Response &res, Error &error) {
+  return cli_->send(req, res, error);
+}
+
+inline Result Client::send(const Request &req) { return cli_->send(req); }
+
+inline void Client::stop() { cli_->stop(); }
+
+inline std::string Client::host() const { return cli_->host(); }
+
+inline int Client::port() const { return cli_->port(); }
+
+inline size_t Client::is_socket_open() const { return cli_->is_socket_open(); }
+
+inline socket_t Client::socket() const { return cli_->socket(); }
+
+inline void
+Client::set_hostname_addr_map(std::map<std::string, std::string> addr_map) {
+  cli_->set_hostname_addr_map(std::move(addr_map));
+}
+
+inline void Client::set_default_headers(Headers headers) {
+  cli_->set_default_headers(std::move(headers));
+}
+
+inline void Client::set_header_writer(
+    std::function<ssize_t(Stream &, Headers &)> const &writer) {
+  cli_->set_header_writer(writer);
+}
+
+inline void Client::set_address_family(int family) {
+  cli_->set_address_family(family);
+}
+
+inline void Client::set_tcp_nodelay(bool on) { cli_->set_tcp_nodelay(on); }
+
+inline void Client::set_socket_options(SocketOptions socket_options) {
+  cli_->set_socket_options(std::move(socket_options));
+}
+
+inline void Client::set_connection_timeout(time_t sec, time_t usec) {
+  cli_->set_connection_timeout(sec, usec);
+}
+
+inline void Client::set_read_timeout(time_t sec, time_t usec) {
+  cli_->set_read_timeout(sec, usec);
+}
+
+inline void Client::set_write_timeout(time_t sec, time_t usec) {
+  cli_->set_write_timeout(sec, usec);
+}
+
+inline void Client::set_basic_auth(const std::string &username,
+                                   const std::string &password) {
+  cli_->set_basic_auth(username, password);
+}
+inline void Client::set_bearer_token_auth(const std::string &token) {
+  cli_->set_bearer_token_auth(token);
+}
+#ifdef CPPHTTPLIB_OPENSSL_SUPPORT
+inline void Client::set_digest_auth(const std::string &username,
+                                    const std::string &password) {
+  cli_->set_digest_auth(username, password);
+}
+#endif
+
+inline void Client::set_keep_alive(bool on) { cli_->set_keep_alive(on); }
+inline void Client::set_follow_location(bool on) {
+  cli_->set_follow_location(on);
+}
+
+inline void Client::set_url_encode(bool on) { cli_->set_url_encode(on); }
+
+inline void Client::set_compress(bool on) { cli_->set_compress(on); }
+
+inline void Client::set_decompress(bool on) { cli_->set_decompress(on); }
+
+inline void Client::set_interface(const std::string &intf) {
+  cli_->set_interface(intf);
+}
+
+inline void Client::set_proxy(const std::string &host, int port) {
+  cli_->set_proxy(host, port);
+}
+inline void Client::set_proxy_basic_auth(const std::string &username,
+                                         const std::string &password) {
+  cli_->set_proxy_basic_auth(username, password);
+}
+inline void Client::set_proxy_bearer_token_auth(const std::string &token) {
+  cli_->set_proxy_bearer_token_auth(token);
+}
+#ifdef CPPHTTPLIB_OPENSSL_SUPPORT
+inline void Client::set_proxy_digest_auth(const std::string &username,
+                                          const std::string &password) {
+  cli_->set_proxy_digest_auth(username, password);
+}
+#endif
+
+#ifdef CPPHTTPLIB_OPENSSL_SUPPORT
+inline void Client::enable_server_certificate_verification(bool enabled) {
+  cli_->enable_server_certificate_verification(enabled);
+}
+#endif
+
+inline void Client::set_logger(Logger logger) {
+  cli_->set_logger(std::move(logger));
+}
+
+#ifdef CPPHTTPLIB_OPENSSL_SUPPORT
+inline void Client::set_ca_cert_path(const std::string &ca_cert_file_path,
+                                     const std::string &ca_cert_dir_path) {
+  cli_->set_ca_cert_path(ca_cert_file_path, ca_cert_dir_path);
+}
+
+inline void Client::set_ca_cert_store(X509_STORE *ca_cert_store) {
+  if (is_ssl_) {
+    static_cast<SSLClient &>(*cli_).set_ca_cert_store(ca_cert_store);
+  } else {
+    cli_->set_ca_cert_store(ca_cert_store);
+  }
+}
+
+inline void Client::load_ca_cert_store(const char *ca_cert, std::size_t size) {
+  set_ca_cert_store(cli_->create_ca_cert_store(ca_cert, size));
+}
+
+inline long Client::get_openssl_verify_result() const {
+  if (is_ssl_) {
+    return static_cast<SSLClient &>(*cli_).get_openssl_verify_result();
+  }
+  return -1; // NOTE: -1 doesn't match any of X509_V_ERR_???
+}
+
+inline SSL_CTX *Client::ssl_context() const {
+  if (is_ssl_) { return static_cast<SSLClient &>(*cli_).ssl_context(); }
+  return nullptr;
+}
+#endif
+
+// ----------------------------------------------------------------------------
+
+} // namespace httplib
+
+#if defined(_WIN32) && defined(CPPHTTPLIB_USE_POLL)
+#undef poll
+#endif
+
+#endif // CPPHTTPLIB_HTTPLIB_H
diff --git a/src/whisper_cpp/examples/server/server.cpp b/src/whisper_cpp/examples/server/server.cpp
new file mode 100644
index 00000000..e7d7c8ae
--- /dev/null
+++ b/src/whisper_cpp/examples/server/server.cpp
@@ -0,0 +1,1041 @@
+#include "common.h"
+
+#include "whisper.h"
+#include "httplib.h"
+#include "json.hpp"
+
+#include <cmath>
+#include <fstream>
+#include <cstdio>
+#include <string>
+#include <thread>
+#include <vector>
+#include <cstring>
+#include <sstream>
+
+#if defined(_MSC_VER)
+#pragma warning(disable: 4244 4267) // possible loss of data
+#endif
+
+using namespace httplib;
+using json = nlohmann::ordered_json;
+
+namespace {
+
+// output formats
+const std::string json_format   = "json";
+const std::string text_format   = "text";
+const std::string srt_format    = "srt";
+const std::string vjson_format  = "verbose_json";
+const std::string vtt_format    = "vtt";
+
+struct server_params
+{
+    std::string hostname = "127.0.0.1";
+    std::string public_path = "examples/server/public";
+    std::string request_path = "";
+    std::string inference_path = "/inference";
+
+    int32_t port          = 8080;
+    int32_t read_timeout  = 600;
+    int32_t write_timeout = 600;
+
+    bool ffmpeg_converter = false;
+};
+
+struct whisper_params {
+    int32_t n_threads     = std::min(4, (int32_t) std::thread::hardware_concurrency());
+    int32_t n_processors  = 1;
+    int32_t offset_t_ms   = 0;
+    int32_t offset_n      = 0;
+    int32_t duration_ms   = 0;
+    int32_t progress_step = 5;
+    int32_t max_context   = -1;
+    int32_t max_len       = 0;
+    int32_t best_of       = 2;
+    int32_t beam_size     = -1;
+    int32_t audio_ctx     = 0;
+
+    float word_thold      =  0.01f;
+    float entropy_thold   =  2.40f;
+    float logprob_thold   = -1.00f;
+    float temperature     =  0.00f;
+    float temperature_inc =  0.20f;
+
+    bool debug_mode      = false;
+    bool translate       = false;
+    bool detect_language = false;
+    bool diarize         = false;
+    bool tinydiarize     = false;
+    bool split_on_word   = false;
+    bool no_fallback     = false;
+    bool print_special   = false;
+    bool print_colors    = false;
+    bool print_realtime  = false;
+    bool print_progress  = false;
+    bool no_timestamps   = false;
+    bool use_gpu         = true;
+    bool flash_attn      = false;
+
+    std::string language        = "en";
+    std::string prompt          = "";
+    std::string font_path       = "/System/Library/Fonts/Supplemental/Courier New Bold.ttf";
+    std::string model           = "models/ggml-base.en.bin";
+
+    std::string response_format     = json_format;
+
+    // [TDRZ] speaker turn string
+    std::string tdrz_speaker_turn = " [SPEAKER_TURN]"; // TODO: set from command line
+
+    std::string openvino_encode_device = "CPU";
+
+    std::string dtw = "";
+};
+
+void whisper_print_usage(int /*argc*/, char ** argv, const whisper_params & params, const server_params& sparams) {
+    fprintf(stderr, "\n");
+    fprintf(stderr, "usage: %s [options] \n", argv[0]);
+    fprintf(stderr, "\n");
+    fprintf(stderr, "options:\n");
+    fprintf(stderr, "  -h,        --help              [default] show this help message and exit\n");
+    fprintf(stderr, "  -t N,      --threads N         [%-7d] number of threads to use during computation\n",    params.n_threads);
+    fprintf(stderr, "  -p N,      --processors N      [%-7d] number of processors to use during computation\n", params.n_processors);
+    fprintf(stderr, "  -ot N,     --offset-t N        [%-7d] time offset in milliseconds\n",                    params.offset_t_ms);
+    fprintf(stderr, "  -on N,     --offset-n N        [%-7d] segment index offset\n",                           params.offset_n);
+    fprintf(stderr, "  -d  N,     --duration N        [%-7d] duration of audio to process in milliseconds\n",   params.duration_ms);
+    fprintf(stderr, "  -mc N,     --max-context N     [%-7d] maximum number of text context tokens to store\n", params.max_context);
+    fprintf(stderr, "  -ml N,     --max-len N         [%-7d] maximum segment length in characters\n",           params.max_len);
+    fprintf(stderr, "  -sow,      --split-on-word     [%-7s] split on word rather than on token\n",             params.split_on_word ? "true" : "false");
+    fprintf(stderr, "  -bo N,     --best-of N         [%-7d] number of best candidates to keep\n",              params.best_of);
+    fprintf(stderr, "  -bs N,     --beam-size N       [%-7d] beam size for beam search\n",                      params.beam_size);
+    fprintf(stderr, "  -ac N,     --audio-ctx N       [%-7d] audio context size (0 - all)\n",                   params.audio_ctx);
+    fprintf(stderr, "  -wt N,     --word-thold N      [%-7.2f] word timestamp probability threshold\n",         params.word_thold);
+    fprintf(stderr, "  -et N,     --entropy-thold N   [%-7.2f] entropy threshold for decoder fail\n",           params.entropy_thold);
+    fprintf(stderr, "  -lpt N,    --logprob-thold N   [%-7.2f] log probability threshold for decoder fail\n",   params.logprob_thold);
+    fprintf(stderr, "  -debug,    --debug-mode        [%-7s] enable debug mode (eg. dump log_mel)\n",           params.debug_mode ? "true" : "false");
+    fprintf(stderr, "  -tr,       --translate         [%-7s] translate from source language to english\n",      params.translate ? "true" : "false");
+    fprintf(stderr, "  -di,       --diarize           [%-7s] stereo audio diarization\n",                       params.diarize ? "true" : "false");
+    fprintf(stderr, "  -tdrz,     --tinydiarize       [%-7s] enable tinydiarize (requires a tdrz model)\n",     params.tinydiarize ? "true" : "false");
+    fprintf(stderr, "  -nf,       --no-fallback       [%-7s] do not use temperature fallback while decoding\n", params.no_fallback ? "true" : "false");
+    fprintf(stderr, "  -ps,       --print-special     [%-7s] print special tokens\n",                           params.print_special ? "true" : "false");
+    fprintf(stderr, "  -pc,       --print-colors      [%-7s] print colors\n",                                   params.print_colors ? "true" : "false");
+    fprintf(stderr, "  -pr,       --print-realtime    [%-7s] print output in realtime\n",                       params.print_realtime ? "true" : "false");
+    fprintf(stderr, "  -pp,       --print-progress    [%-7s] print progress\n",                                 params.print_progress ? "true" : "false");
+    fprintf(stderr, "  -nt,       --no-timestamps     [%-7s] do not print timestamps\n",                        params.no_timestamps ? "true" : "false");
+    fprintf(stderr, "  -l LANG,   --language LANG     [%-7s] spoken language ('auto' for auto-detect)\n",       params.language.c_str());
+    fprintf(stderr, "  -dl,       --detect-language   [%-7s] exit after automatically detecting language\n",    params.detect_language ? "true" : "false");
+    fprintf(stderr, "             --prompt PROMPT     [%-7s] initial prompt\n",                                 params.prompt.c_str());
+    fprintf(stderr, "  -m FNAME,  --model FNAME       [%-7s] model path\n",                                     params.model.c_str());
+    fprintf(stderr, "  -oved D,   --ov-e-device DNAME [%-7s] the OpenVINO device used for encode inference\n",  params.openvino_encode_device.c_str());
+    // server params
+    fprintf(stderr, "  -dtw MODEL --dtw MODEL         [%-7s] compute token-level timestamps\n", params.dtw.c_str());
+    fprintf(stderr, "  --host HOST,                   [%-7s] Hostname/ip-adress for the server\n", sparams.hostname.c_str());
+    fprintf(stderr, "  --port PORT,                   [%-7d] Port number for the server\n", sparams.port);
+    fprintf(stderr, "  --public PATH,                 [%-7s] Path to the public folder\n", sparams.public_path.c_str());
+    fprintf(stderr, "  --request-path PATH,           [%-7s] Request path for all requests\n", sparams.request_path.c_str());
+    fprintf(stderr, "  --inference-path PATH,         [%-7s] Inference path for all requests\n", sparams.inference_path.c_str());
+    fprintf(stderr, "  --convert,                     [%-7s] Convert audio to WAV, requires ffmpeg on the server", sparams.ffmpeg_converter ? "true" : "false");
+    fprintf(stderr, "\n");
+}
+
+bool whisper_params_parse(int argc, char ** argv, whisper_params & params, server_params & sparams) {
+    for (int i = 1; i < argc; i++) {
+        std::string arg = argv[i];
+
+        if (arg == "-h" || arg == "--help") {
+            whisper_print_usage(argc, argv, params, sparams);
+            exit(0);
+        }
+        else if (arg == "-t"    || arg == "--threads")         { params.n_threads       = std::stoi(argv[++i]); }
+        else if (arg == "-p"    || arg == "--processors")      { params.n_processors    = std::stoi(argv[++i]); }
+        else if (arg == "-ot"   || arg == "--offset-t")        { params.offset_t_ms     = std::stoi(argv[++i]); }
+        else if (arg == "-on"   || arg == "--offset-n")        { params.offset_n        = std::stoi(argv[++i]); }
+        else if (arg == "-d"    || arg == "--duration")        { params.duration_ms     = std::stoi(argv[++i]); }
+        else if (arg == "-mc"   || arg == "--max-context")     { params.max_context     = std::stoi(argv[++i]); }
+        else if (arg == "-ml"   || arg == "--max-len")         { params.max_len         = std::stoi(argv[++i]); }
+        else if (arg == "-bo"   || arg == "--best-of")         { params.best_of         = std::stoi(argv[++i]); }
+        else if (arg == "-bs"   || arg == "--beam-size")       { params.beam_size       = std::stoi(argv[++i]); }
+        else if (arg == "-ac"   || arg == "--audio-ctx")       { params.audio_ctx       = std::stoi(argv[++i]); }
+        else if (arg == "-wt"   || arg == "--word-thold")      { params.word_thold      = std::stof(argv[++i]); }
+        else if (arg == "-et"   || arg == "--entropy-thold")   { params.entropy_thold   = std::stof(argv[++i]); }
+        else if (arg == "-lpt"  || arg == "--logprob-thold")   { params.logprob_thold   = std::stof(argv[++i]); }
+        else if (arg == "-debug"|| arg == "--debug-mode")      { params.debug_mode      = true; }
+        else if (arg == "-tr"   || arg == "--translate")       { params.translate       = true; }
+        else if (arg == "-di"   || arg == "--diarize")         { params.diarize         = true; }
+        else if (arg == "-tdrz" || arg == "--tinydiarize")     { params.tinydiarize     = true; }
+        else if (arg == "-sow"  || arg == "--split-on-word")   { params.split_on_word   = true; }
+        else if (arg == "-nf"   || arg == "--no-fallback")     { params.no_fallback     = true; }
+        else if (arg == "-fp"   || arg == "--font-path")       { params.font_path       = argv[++i]; }
+        else if (arg == "-ps"   || arg == "--print-special")   { params.print_special   = true; }
+        else if (arg == "-pc"   || arg == "--print-colors")    { params.print_colors    = true; }
+        else if (arg == "-pr"   || arg == "--print-realtime")  { params.print_realtime  = true; }
+        else if (arg == "-pp"   || arg == "--print-progress")  { params.print_progress  = true; }
+        else if (arg == "-nt"   || arg == "--no-timestamps")   { params.no_timestamps   = true; }
+        else if (arg == "-l"    || arg == "--language")        { params.language        = argv[++i]; }
+        else if (arg == "-dl"   || arg == "--detect-language") { params.detect_language = true; }
+        else if (                  arg == "--prompt")          { params.prompt          = argv[++i]; }
+        else if (arg == "-m"    || arg == "--model")           { params.model           = argv[++i]; }
+        else if (arg == "-oved" || arg == "--ov-e-device")     { params.openvino_encode_device = argv[++i]; }
+        else if (arg == "-dtw"  || arg == "--dtw")             { params.dtw             = argv[++i]; }
+        else if (arg == "-ng"   || arg == "--no-gpu")          { params.use_gpu         = false; }
+        else if (arg == "-fa"   || arg == "--flash-attn")      { params.flash_attn      = true; }
+        // server params
+        else if (                  arg == "--port")            { sparams.port        = std::stoi(argv[++i]); }
+        else if (                  arg == "--host")            { sparams.hostname    = argv[++i]; }
+        else if (                  arg == "--public")          { sparams.public_path = argv[++i]; }
+        else if (                  arg == "--request-path")    { sparams.request_path = argv[++i]; }
+        else if (                  arg == "--inference-path")  { sparams.inference_path = argv[++i]; }
+        else if (                  arg == "--convert")         { sparams.ffmpeg_converter     = true; }
+        else {
+            fprintf(stderr, "error: unknown argument: %s\n", arg.c_str());
+            whisper_print_usage(argc, argv, params, sparams);
+            exit(0);
+        }
+    }
+
+    return true;
+}
+
+struct whisper_print_user_data {
+    const whisper_params * params;
+
+    const std::vector<std::vector<float>> * pcmf32s;
+    int progress_prev;
+};
+
+void check_ffmpeg_availibility() {
+    int result = system("ffmpeg -version");
+
+    if (result == 0) {
+        std::cout << "ffmpeg is available." << std::endl;
+    } else {
+        // ffmpeg is not available
+        std::cout << "ffmpeg is not found. Please ensure that ffmpeg is installed ";
+        std::cout << "and that its executable is included in your system's PATH. ";
+        exit(0);
+    }
+}
+
+bool convert_to_wav(const std::string & temp_filename, std::string & error_resp) {
+    std::ostringstream cmd_stream;
+    std::string converted_filename_temp = temp_filename + "_temp.wav";
+    cmd_stream << "ffmpeg -i \"" << temp_filename << "\" -y -ar 16000 -ac 1 -c:a pcm_s16le \"" << converted_filename_temp << "\" 2>&1";
+    std::string cmd = cmd_stream.str();
+
+    int status = std::system(cmd.c_str());
+    if (status != 0) {
+        error_resp = "{\"error\":\"FFmpeg conversion failed.\"}";
+        return false;
+    }
+
+    // Remove the original file
+    if (remove(temp_filename.c_str()) != 0) {
+        error_resp = "{\"error\":\"Failed to remove the original file.\"}";
+        return false;
+    }
+
+    // Rename the temporary file to match the original filename
+    if (rename(converted_filename_temp.c_str(), temp_filename.c_str()) != 0) {
+        error_resp = "{\"error\":\"Failed to rename the temporary file.\"}";
+        return false;
+    }
+    return true;
+}
+
+std::string estimate_diarization_speaker(std::vector<std::vector<float>> pcmf32s, int64_t t0, int64_t t1, bool id_only = false) {
+    std::string speaker = "";
+    const int64_t n_samples = pcmf32s[0].size();
+
+    const int64_t is0 = timestamp_to_sample(t0, n_samples, WHISPER_SAMPLE_RATE);
+    const int64_t is1 = timestamp_to_sample(t1, n_samples, WHISPER_SAMPLE_RATE);
+
+    double energy0 = 0.0f;
+    double energy1 = 0.0f;
+
+    for (int64_t j = is0; j < is1; j++) {
+        energy0 += fabs(pcmf32s[0][j]);
+        energy1 += fabs(pcmf32s[1][j]);
+    }
+
+    if (energy0 > 1.1*energy1) {
+        speaker = "0";
+    } else if (energy1 > 1.1*energy0) {
+        speaker = "1";
+    } else {
+        speaker = "?";
+    }
+
+    //printf("is0 = %lld, is1 = %lld, energy0 = %f, energy1 = %f, speaker = %s\n", is0, is1, energy0, energy1, speaker.c_str());
+
+    if (!id_only) {
+        speaker.insert(0, "(speaker ");
+        speaker.append(")");
+    }
+
+    return speaker;
+}
+
+void whisper_print_progress_callback(struct whisper_context * /*ctx*/, struct whisper_state * /*state*/, int progress, void * user_data) {
+    int progress_step = ((whisper_print_user_data *) user_data)->params->progress_step;
+    int * progress_prev  = &(((whisper_print_user_data *) user_data)->progress_prev);
+    if (progress >= *progress_prev + progress_step) {
+        *progress_prev += progress_step;
+        fprintf(stderr, "%s: progress = %3d%%\n", __func__, progress);
+    }
+}
+
+void whisper_print_segment_callback(struct whisper_context * ctx, struct whisper_state * /*state*/, int n_new, void * user_data) {
+    const auto & params  = *((whisper_print_user_data *) user_data)->params;
+    const auto & pcmf32s = *((whisper_print_user_data *) user_data)->pcmf32s;
+
+    const int n_segments = whisper_full_n_segments(ctx);
+
+    std::string speaker = "";
+
+    int64_t t0 = 0;
+    int64_t t1 = 0;
+
+    // print the last n_new segments
+    const int s0 = n_segments - n_new;
+
+    if (s0 == 0) {
+        printf("\n");
+    }
+
+    for (int i = s0; i < n_segments; i++) {
+        if (!params.no_timestamps || params.diarize) {
+            t0 = whisper_full_get_segment_t0(ctx, i);
+            t1 = whisper_full_get_segment_t1(ctx, i);
+        }
+
+        if (!params.no_timestamps) {
+            printf("[%s --> %s]  ", to_timestamp(t0).c_str(), to_timestamp(t1).c_str());
+        }
+
+        if (params.diarize && pcmf32s.size() == 2) {
+            speaker = estimate_diarization_speaker(pcmf32s, t0, t1);
+        }
+
+        if (params.print_colors) {
+            for (int j = 0; j < whisper_full_n_tokens(ctx, i); ++j) {
+                if (params.print_special == false) {
+                    const whisper_token id = whisper_full_get_token_id(ctx, i, j);
+                    if (id >= whisper_token_eot(ctx)) {
+                        continue;
+                    }
+                }
+
+                const char * text = whisper_full_get_token_text(ctx, i, j);
+                const float  p    = whisper_full_get_token_p   (ctx, i, j);
+
+                const int col = std::max(0, std::min((int) k_colors.size() - 1, (int) (std::pow(p, 3)*float(k_colors.size()))));
+
+                printf("%s%s%s%s", speaker.c_str(), k_colors[col].c_str(), text, "\033[0m");
+            }
+        } else {
+            const char * text = whisper_full_get_segment_text(ctx, i);
+
+            printf("%s%s", speaker.c_str(), text);
+        }
+
+        if (params.tinydiarize) {
+            if (whisper_full_get_segment_speaker_turn_next(ctx, i)) {
+                printf("%s", params.tdrz_speaker_turn.c_str());
+            }
+        }
+
+        // with timestamps or speakers: each segment on new line
+        if (!params.no_timestamps || params.diarize) {
+            printf("\n");
+        }
+        fflush(stdout);
+    }
+}
+
+std::string output_str(struct whisper_context * ctx, const whisper_params & params, std::vector<std::vector<float>> pcmf32s) {
+    std::stringstream result;
+    const int n_segments = whisper_full_n_segments(ctx);
+    for (int i = 0; i < n_segments; ++i) {
+        const char * text = whisper_full_get_segment_text(ctx, i);
+        std::string speaker = "";
+
+        if (params.diarize && pcmf32s.size() == 2)
+        {
+            const int64_t t0 = whisper_full_get_segment_t0(ctx, i);
+            const int64_t t1 = whisper_full_get_segment_t1(ctx, i);
+            speaker = estimate_diarization_speaker(pcmf32s, t0, t1);
+        }
+
+        result << speaker << text << "\n";
+    }
+    return result.str();
+}
+
+bool parse_str_to_bool(const std::string & s) {
+    if (s == "true" || s == "1" || s == "yes" || s == "y") {
+        return true;
+    }
+    return false;
+}
+
+void get_req_parameters(const Request & req, whisper_params & params)
+{
+    if (req.has_file("offset_t"))
+    {
+        params.offset_t_ms = std::stoi(req.get_file_value("offset_t").content);
+    }
+    if (req.has_file("offset_n"))
+    {
+        params.offset_n = std::stoi(req.get_file_value("offset_n").content);
+    }
+    if (req.has_file("duration"))
+    {
+        params.duration_ms = std::stoi(req.get_file_value("duration").content);
+    }
+    if (req.has_file("max_context"))
+    {
+        params.max_context = std::stoi(req.get_file_value("max_context").content);
+    }
+    if (req.has_file("max_len"))
+    {
+        params.max_len = std::stoi(req.get_file_value("max_len").content);
+    }
+    if (req.has_file("best_of"))
+    {
+        params.best_of = std::stoi(req.get_file_value("best_of").content);
+    }
+    if (req.has_file("beam_size"))
+    {
+        params.beam_size = std::stoi(req.get_file_value("beam_size").content);
+    }
+    if (req.has_file("audio_ctx"))
+    {
+        params.audio_ctx = std::stof(req.get_file_value("audio_ctx").content);
+    }
+    if (req.has_file("word_thold"))
+    {
+        params.word_thold = std::stof(req.get_file_value("word_thold").content);
+    }
+    if (req.has_file("entropy_thold"))
+    {
+        params.entropy_thold = std::stof(req.get_file_value("entropy_thold").content);
+    }
+    if (req.has_file("logprob_thold"))
+    {
+        params.logprob_thold = std::stof(req.get_file_value("logprob_thold").content);
+    }
+    if (req.has_file("debug_mode"))
+    {
+        params.debug_mode = parse_str_to_bool(req.get_file_value("debug_mode").content);
+    }
+    if (req.has_file("translate"))
+    {
+        params.translate = parse_str_to_bool(req.get_file_value("translate").content);
+    }
+    if (req.has_file("diarize"))
+    {
+        params.diarize = parse_str_to_bool(req.get_file_value("diarize").content);
+    }
+    if (req.has_file("tinydiarize"))
+    {
+        params.tinydiarize = parse_str_to_bool(req.get_file_value("tinydiarize").content);
+    }
+    if (req.has_file("split_on_word"))
+    {
+        params.split_on_word = parse_str_to_bool(req.get_file_value("split_on_word").content);
+    }
+    if (req.has_file("no_timestamps"))
+    {
+        params.no_timestamps = parse_str_to_bool(req.get_file_value("no_timestamps").content);
+    }
+    if (req.has_file("language"))
+    {
+        params.language = req.get_file_value("language").content;
+    }
+    if (req.has_file("detect_language"))
+    {
+        params.detect_language = parse_str_to_bool(req.get_file_value("detect_language").content);
+    }
+    if (req.has_file("prompt"))
+    {
+        params.prompt = req.get_file_value("prompt").content;
+    }
+    if (req.has_file("response_format"))
+    {
+        params.response_format = req.get_file_value("response_format").content;
+    }
+    if (req.has_file("temperature"))
+    {
+        params.temperature = std::stof(req.get_file_value("temperature").content);
+    }
+    if (req.has_file("temperature_inc"))
+    {
+        params.temperature_inc = std::stof(req.get_file_value("temperature_inc").content);
+    }
+}
+
+}  // namespace
+
+int main(int argc, char ** argv) {
+    whisper_params params;
+    server_params sparams;
+
+    std::mutex whisper_mutex;
+
+    if (whisper_params_parse(argc, argv, params, sparams) == false) {
+        whisper_print_usage(argc, argv, params, sparams);
+        return 1;
+    }
+
+    if (params.language != "auto" && whisper_lang_id(params.language.c_str()) == -1) {
+        fprintf(stderr, "error: unknown language '%s'\n", params.language.c_str());
+        whisper_print_usage(argc, argv, params, sparams);
+        exit(0);
+    }
+
+    if (params.diarize && params.tinydiarize) {
+        fprintf(stderr, "error: cannot use both --diarize and --tinydiarize\n");
+        whisper_print_usage(argc, argv, params, sparams);
+        exit(0);
+    }
+
+    if (sparams.ffmpeg_converter) {
+        check_ffmpeg_availibility();
+    }
+    // whisper init
+    struct whisper_context_params cparams = whisper_context_default_params();
+
+    cparams.use_gpu    = params.use_gpu;
+    cparams.flash_attn = params.flash_attn;
+
+    if (!params.dtw.empty()) {
+        cparams.dtw_token_timestamps = true;
+        cparams.dtw_aheads_preset = WHISPER_AHEADS_NONE;
+
+        if (params.dtw == "tiny") {
+            cparams.dtw_aheads_preset = WHISPER_AHEADS_TINY;
+        }
+        if (params.dtw == "tiny.en") {
+            cparams.dtw_aheads_preset = WHISPER_AHEADS_TINY_EN;
+        }
+        if (params.dtw == "base") {
+            cparams.dtw_aheads_preset = WHISPER_AHEADS_BASE;
+        }
+        if (params.dtw == "base.en") {
+            cparams.dtw_aheads_preset = WHISPER_AHEADS_BASE_EN;
+        }
+        if (params.dtw == "small") {
+            cparams.dtw_aheads_preset = WHISPER_AHEADS_SMALL;
+        }
+        if (params.dtw == "small.en") {
+            cparams.dtw_aheads_preset = WHISPER_AHEADS_SMALL_EN;
+        }
+        if (params.dtw == "medium") {
+            cparams.dtw_aheads_preset = WHISPER_AHEADS_MEDIUM;
+        }
+        if (params.dtw == "medium.en") {
+            cparams.dtw_aheads_preset = WHISPER_AHEADS_MEDIUM_EN;
+        }
+        if (params.dtw == "large.v1") {
+            cparams.dtw_aheads_preset = WHISPER_AHEADS_LARGE_V1;
+        }
+        if (params.dtw == "large.v2") {
+            cparams.dtw_aheads_preset = WHISPER_AHEADS_LARGE_V2;
+        }
+        if (params.dtw == "large.v3") {
+            cparams.dtw_aheads_preset = WHISPER_AHEADS_LARGE_V3;
+        }
+
+        if (cparams.dtw_aheads_preset == WHISPER_AHEADS_NONE) {
+            fprintf(stderr, "error: unknown DTW preset '%s'\n", params.dtw.c_str());
+            return 3;
+        }
+    }
+
+    struct whisper_context * ctx = whisper_init_from_file_with_params(params.model.c_str(), cparams);
+
+    if (ctx == nullptr) {
+        fprintf(stderr, "error: failed to initialize whisper context\n");
+        return 3;
+    }
+
+    // initialize openvino encoder. this has no effect on whisper.cpp builds that don't have OpenVINO configured
+    whisper_ctx_init_openvino_encoder(ctx, nullptr, params.openvino_encode_device.c_str(), nullptr);
+
+    Server svr;
+    svr.set_default_headers({{"Server", "whisper.cpp"},
+                             {"Access-Control-Allow-Origin", "*"},
+                             {"Access-Control-Allow-Headers", "content-type, authorization"}});
+
+    std::string const default_content = R"(
+    <html>
+    <head>
+        <title>Whisper.cpp Server</title>
+        <meta charset="utf-8">
+        <meta name="viewport" content="width=device-width">
+        <style>
+        body {
+            font-family: sans-serif;
+        }
+        form {
+            display: flex;
+            flex-direction: column;
+            align-items: flex-start;
+        }
+        label {
+            margin-bottom: 0.5rem;
+        }
+        input, select {
+            margin-bottom: 1rem;
+        }
+        button {
+            margin-top: 1rem;
+        }
+        </style>
+    </head>
+    <body>
+        <h1>Whisper.cpp Server</h1>
+
+        <h2>/inference</h2>
+        <pre>
+    curl 127.0.0.1:)" + std::to_string(sparams.port) + R"(/inference \
+    -H "Content-Type: multipart/form-data" \
+    -F file="@&lt;file-path&gt;" \
+    -F temperature="0.0" \
+    -F temperature_inc="0.2" \
+    -F response_format="json"
+        </pre>
+
+        <h2>/load</h2>
+        <pre>
+    curl 127.0.0.1:)" + std::to_string(sparams.port) + R"(/load \
+    -H "Content-Type: multipart/form-data" \
+    -F model="&lt;path-to-model-file&gt;"
+        </pre>
+
+        <div>
+            <h2>Try it out</h2>
+            <form action="/inference" method="POST" enctype="multipart/form-data">
+                <label for="file">Choose an audio file:</label>
+                <input type="file" id="file" name="file" accept="audio/*" required><br>
+
+                <label for="temperature">Temperature:</label>
+                <input type="number" id="temperature" name="temperature" value="0.0" step="0.01" placeholder="e.g., 0.0"><br>
+
+                <label for="response_format">Response Format:</label>
+                <select id="response_format" name="response_format">
+                    <option value="verbose_json">Verbose JSON</option>
+                    <option value="json">JSON</option>
+                    <option value="text">Text</option>
+                    <option value="srt">SRT</option>
+                    <option value="vtt">VTT</option>
+                </select><br>
+
+                <button type="submit">Submit</button>
+            </form>
+        </div>
+    </body>
+    </html>
+    )";
+
+    // store default params so we can reset after each inference request
+    whisper_params default_params = params;
+
+    // this is only called if no index.html is found in the public --path
+    svr.Get(sparams.request_path + "/", [&default_content](const Request &, Response &res){
+        res.set_content(default_content, "text/html");
+        return false;
+    });
+
+    svr.Options(sparams.request_path + sparams.inference_path, [&](const Request &, Response &){
+    });
+
+    svr.Post(sparams.request_path + sparams.inference_path, [&](const Request &req, Response &res){
+        // acquire whisper model mutex lock
+        std::lock_guard<std::mutex> lock(whisper_mutex);
+
+        // first check user requested fields of the request
+        if (!req.has_file("file"))
+        {
+            fprintf(stderr, "error: no 'file' field in the request\n");
+            const std::string error_resp = "{\"error\":\"no 'file' field in the request\"}";
+            res.set_content(error_resp, "application/json");
+            return;
+        }
+        auto audio_file = req.get_file_value("file");
+
+        // check non-required fields
+        get_req_parameters(req, params);
+
+        std::string filename{audio_file.filename};
+        printf("Received request: %s\n", filename.c_str());
+
+        // audio arrays
+        std::vector<float> pcmf32;               // mono-channel F32 PCM
+        std::vector<std::vector<float>> pcmf32s; // stereo-channel F32 PCM
+
+        if (sparams.ffmpeg_converter) {
+            // if file is not wav, convert to wav
+            // write to temporary file
+            const std::string temp_filename_base = std::tmpnam(nullptr);
+            const std::string temp_filename = temp_filename_base + ".wav";
+            std::ofstream temp_file{temp_filename, std::ios::binary};
+            temp_file << audio_file.content;
+            temp_file.close();
+
+            std::string error_resp = "{\"error\":\"Failed to execute ffmpeg command.\"}";
+            const bool is_converted = convert_to_wav(temp_filename, error_resp);
+            if (!is_converted) {
+                res.set_content(error_resp, "application/json");
+                return;
+            }
+
+            // read wav content into pcmf32
+            if (!::read_wav(temp_filename, pcmf32, pcmf32s, params.diarize))
+            {
+                fprintf(stderr, "error: failed to read WAV file '%s'\n", temp_filename.c_str());
+                const std::string error_resp = "{\"error\":\"failed to read WAV file\"}";
+                res.set_content(error_resp, "application/json");
+                std::remove(temp_filename.c_str());
+                return;
+            }
+            // remove temp file
+            std::remove(temp_filename.c_str());
+        } else {
+            if (!::read_wav(audio_file.content, pcmf32, pcmf32s, params.diarize))
+            {
+                fprintf(stderr, "error: failed to read WAV file\n");
+                const std::string error_resp = "{\"error\":\"failed to read WAV file\"}";
+                res.set_content(error_resp, "application/json");
+                return;
+            }
+        }
+
+
+        printf("Successfully loaded %s\n", filename.c_str());
+
+        // print system information
+        {
+            fprintf(stderr, "\n");
+            fprintf(stderr, "system_info: n_threads = %d / %d | %s\n",
+                    params.n_threads*params.n_processors, std::thread::hardware_concurrency(), whisper_print_system_info());
+        }
+
+        // print some info about the processing
+        {
+            fprintf(stderr, "\n");
+            if (!whisper_is_multilingual(ctx)) {
+                if (params.language != "en" || params.translate) {
+                    params.language = "en";
+                    params.translate = false;
+                    fprintf(stderr, "%s: WARNING: model is not multilingual, ignoring language and translation options\n", __func__);
+                }
+            }
+            if (params.detect_language) {
+                params.language = "auto";
+            }
+            fprintf(stderr, "%s: processing '%s' (%d samples, %.1f sec), %d threads, %d processors, lang = %s, task = %s, %stimestamps = %d ...\n",
+                    __func__, filename.c_str(), int(pcmf32.size()), float(pcmf32.size())/WHISPER_SAMPLE_RATE,
+                    params.n_threads, params.n_processors,
+                    params.language.c_str(),
+                    params.translate ? "translate" : "transcribe",
+                    params.tinydiarize ? "tdrz = 1, " : "",
+                    params.no_timestamps ? 0 : 1);
+
+            fprintf(stderr, "\n");
+        }
+
+        // run the inference
+        {
+            printf("Running whisper.cpp inference on %s\n", filename.c_str());
+            whisper_full_params wparams = whisper_full_default_params(WHISPER_SAMPLING_GREEDY);
+
+            wparams.strategy = params.beam_size > 1 ? WHISPER_SAMPLING_BEAM_SEARCH : WHISPER_SAMPLING_GREEDY;
+
+            wparams.print_realtime   = false;
+            wparams.print_progress   = params.print_progress;
+            wparams.print_timestamps = !params.no_timestamps;
+            wparams.print_special    = params.print_special;
+            wparams.translate        = params.translate;
+            wparams.language         = params.language.c_str();
+            wparams.detect_language  = params.detect_language;
+            wparams.n_threads        = params.n_threads;
+            wparams.n_max_text_ctx   = params.max_context >= 0 ? params.max_context : wparams.n_max_text_ctx;
+            wparams.offset_ms        = params.offset_t_ms;
+            wparams.duration_ms      = params.duration_ms;
+
+            wparams.thold_pt         = params.word_thold;
+            wparams.max_len          = params.max_len == 0 ? 60 : params.max_len;
+            wparams.split_on_word    = params.split_on_word;
+            wparams.audio_ctx        = params.audio_ctx;
+
+            wparams.debug_mode       = params.debug_mode;
+
+            wparams.tdrz_enable      = params.tinydiarize; // [TDRZ]
+
+            wparams.initial_prompt   = params.prompt.c_str();
+
+            wparams.greedy.best_of        = params.best_of;
+            wparams.beam_search.beam_size = params.beam_size;
+
+            wparams.temperature      = params.temperature;
+            wparams.temperature_inc  = params.temperature_inc;
+            wparams.entropy_thold    = params.entropy_thold;
+            wparams.logprob_thold    = params.logprob_thold;
+
+            wparams.no_timestamps    = params.no_timestamps;
+            wparams.token_timestamps = !params.no_timestamps && params.response_format == vjson_format;
+
+            whisper_print_user_data user_data = { &params, &pcmf32s, 0 };
+
+            // this callback is called on each new segment
+            if (params.print_realtime) {
+                wparams.new_segment_callback           = whisper_print_segment_callback;
+                wparams.new_segment_callback_user_data = &user_data;
+            }
+
+            if (wparams.print_progress) {
+                wparams.progress_callback           = whisper_print_progress_callback;
+                wparams.progress_callback_user_data = &user_data;
+            }
+
+            // examples for abort mechanism
+            // in examples below, we do not abort the processing, but we could if the flag is set to true
+
+            // the callback is called before every encoder run - if it returns false, the processing is aborted
+            {
+                static bool is_aborted = false; // NOTE: this should be atomic to avoid data race
+
+                wparams.encoder_begin_callback = [](struct whisper_context * /*ctx*/, struct whisper_state * /*state*/, void * user_data) {
+                    bool is_aborted = *(bool*)user_data;
+                    return !is_aborted;
+                };
+                wparams.encoder_begin_callback_user_data = &is_aborted;
+            }
+
+            // the callback is called before every computation - if it returns true, the computation is aborted
+            {
+                static bool is_aborted = false; // NOTE: this should be atomic to avoid data race
+
+                wparams.abort_callback = [](void * user_data) {
+                    bool is_aborted = *(bool*)user_data;
+                    return is_aborted;
+                };
+                wparams.abort_callback_user_data = &is_aborted;
+            }
+
+            if (whisper_full_parallel(ctx, wparams, pcmf32.data(), pcmf32.size(), params.n_processors) != 0) {
+                fprintf(stderr, "%s: failed to process audio\n", argv[0]);
+                const std::string error_resp = "{\"error\":\"failed to process audio\"}";
+                res.set_content(error_resp, "application/json");
+                return;
+            }
+        }
+
+        // return results to user
+        if (params.response_format == text_format)
+        {
+            std::string results = output_str(ctx, params, pcmf32s);
+            res.set_content(results.c_str(), "text/html; charset=utf-8");
+        }
+        else if (params.response_format == srt_format)
+        {
+            std::stringstream ss;
+            const int n_segments = whisper_full_n_segments(ctx);
+            for (int i = 0; i < n_segments; ++i) {
+                const char * text = whisper_full_get_segment_text(ctx, i);
+                const int64_t t0 = whisper_full_get_segment_t0(ctx, i);
+                const int64_t t1 = whisper_full_get_segment_t1(ctx, i);
+                std::string speaker = "";
+
+                if (params.diarize && pcmf32s.size() == 2)
+                {
+                    speaker = estimate_diarization_speaker(pcmf32s, t0, t1);
+                }
+
+                ss << i + 1 + params.offset_n << "\n";
+                ss << to_timestamp(t0, true) << " --> " << to_timestamp(t1, true) << "\n";
+                ss << speaker << text << "\n\n";
+            }
+            res.set_content(ss.str(), "application/x-subrip");
+        } else if (params.response_format == vtt_format) {
+            std::stringstream ss;
+
+            ss << "WEBVTT\n\n";
+
+            const int n_segments = whisper_full_n_segments(ctx);
+            for (int i = 0; i < n_segments; ++i) {
+                const char * text = whisper_full_get_segment_text(ctx, i);
+                const int64_t t0 = whisper_full_get_segment_t0(ctx, i);
+                const int64_t t1 = whisper_full_get_segment_t1(ctx, i);
+                std::string speaker = "";
+
+                if (params.diarize && pcmf32s.size() == 2)
+                {
+                    speaker = estimate_diarization_speaker(pcmf32s, t0, t1, true);
+                    speaker.insert(0, "<v Speaker");
+                    speaker.append(">");
+                }
+
+                ss << to_timestamp(t0) << " --> " << to_timestamp(t1) << "\n";
+                ss << speaker << text << "\n\n";
+            }
+            res.set_content(ss.str(), "text/vtt");
+        } else if (params.response_format == vjson_format) {
+            /* try to match openai/whisper's Python format */
+            std::string results = output_str(ctx, params, pcmf32s);
+            json jres = json{
+                {"task", params.translate ? "translate" : "transcribe"},
+                {"language", whisper_lang_str_full(whisper_full_lang_id(ctx))},
+                {"duration", float(pcmf32.size())/WHISPER_SAMPLE_RATE},
+                {"text", results},
+                {"segments", json::array()}
+            };
+            const int n_segments = whisper_full_n_segments(ctx);
+            for (int i = 0; i < n_segments; ++i)
+            {
+                json segment = json{
+                    {"id", i},
+                    {"text", whisper_full_get_segment_text(ctx, i)},
+                };
+
+                if (!params.no_timestamps) {
+                    segment["start"] = whisper_full_get_segment_t0(ctx, i) * 0.01;
+                    segment["end"] = whisper_full_get_segment_t1(ctx, i) * 0.01;
+                }
+
+                float total_logprob = 0;
+                const int n_tokens = whisper_full_n_tokens(ctx, i);
+                for (int j = 0; j < n_tokens; ++j) {
+                    whisper_token_data token = whisper_full_get_token_data(ctx, i, j);
+                    if (token.id >= whisper_token_eot(ctx)) {
+                        continue;
+                    }
+
+                    segment["tokens"].push_back(token.id);
+                    json word = json{{"word", whisper_full_get_token_text(ctx, i, j)}};
+                    if (!params.no_timestamps) {
+                        word["start"] = token.t0 * 0.01;
+                        word["end"] = token.t1 * 0.01;
+                        word["t_dtw"] = token.t_dtw;
+                    }
+                    word["probability"] = token.p;
+                    total_logprob += token.plog;
+                    segment["words"].push_back(word);
+                }
+
+                segment["temperature"] = params.temperature;
+                segment["avg_logprob"] = total_logprob / n_tokens;
+
+                // TODO compression_ratio and no_speech_prob are not implemented yet
+                // segment["compression_ratio"] = 0;
+                // segment["no_speech_prob"] = 0;
+
+                jres["segments"].push_back(segment);
+            }
+            res.set_content(jres.dump(-1, ' ', false, json::error_handler_t::replace),
+                            "application/json");
+        }
+        // TODO add more output formats
+        else
+        {
+            std::string results = output_str(ctx, params, pcmf32s);
+            json jres = json{
+                {"text", results}
+            };
+            res.set_content(jres.dump(-1, ' ', false, json::error_handler_t::replace),
+                            "application/json");
+        }
+
+        // reset params to their defaults
+        params = default_params;
+    });
+    svr.Post(sparams.request_path + "/load", [&](const Request &req, Response &res){
+        std::lock_guard<std::mutex> lock(whisper_mutex);
+        if (!req.has_file("model"))
+        {
+            fprintf(stderr, "error: no 'model' field in the request\n");
+            const std::string error_resp = "{\"error\":\"no 'model' field in the request\"}";
+            res.set_content(error_resp, "application/json");
+            return;
+        }
+        std::string model = req.get_file_value("model").content;
+        if (!is_file_exist(model.c_str()))
+        {
+            fprintf(stderr, "error: 'model': %s not found!\n", model.c_str());
+            const std::string error_resp = "{\"error\":\"model not found!\"}";
+            res.set_content(error_resp, "application/json");
+            return;
+        }
+
+        // clean up
+        whisper_free(ctx);
+
+        // whisper init
+        ctx = whisper_init_from_file_with_params(model.c_str(), cparams);
+
+        // TODO perhaps load prior model here instead of exit
+        if (ctx == nullptr) {
+            fprintf(stderr, "error: model init  failed, no model loaded must exit\n");
+            exit(1);
+        }
+
+        // initialize openvino encoder. this has no effect on whisper.cpp builds that don't have OpenVINO configured
+        whisper_ctx_init_openvino_encoder(ctx, nullptr, params.openvino_encode_device.c_str(), nullptr);
+
+        const std::string success = "Load was successful!";
+        res.set_content(success, "application/text");
+
+        // check if the model is in the file system
+    });
+
+    svr.set_exception_handler([](const Request &, Response &res, std::exception_ptr ep) {
+        const char fmt[] = "500 Internal Server Error\n%s";
+        char buf[BUFSIZ];
+        try {
+            std::rethrow_exception(std::move(ep));
+        } catch (std::exception &e) {
+            snprintf(buf, sizeof(buf), fmt, e.what());
+        } catch (...) {
+            snprintf(buf, sizeof(buf), fmt, "Unknown Exception");
+        }
+        res.set_content(buf, "text/plain");
+        res.status = 500;
+    });
+
+    svr.set_error_handler([](const Request &req, Response &res) {
+        if (res.status == 400) {
+            res.set_content("Invalid request", "text/plain");
+        } else if (res.status != 500) {
+            res.set_content("File Not Found (" + req.path + ")", "text/plain");
+            res.status = 404;
+        }
+    });
+
+    // set timeouts and change hostname and port
+    svr.set_read_timeout(sparams.read_timeout);
+    svr.set_write_timeout(sparams.write_timeout);
+
+    if (!svr.bind_to_port(sparams.hostname, sparams.port))
+    {
+        fprintf(stderr, "\ncouldn't bind to server socket: hostname=%s port=%d\n\n",
+                sparams.hostname.c_str(), sparams.port);
+        return 1;
+    }
+
+    // Set the base directory for serving static files
+    svr.set_base_dir(sparams.public_path);
+
+    // to make it ctrl+clickable:
+    printf("\nwhisper server listening at http://%s:%d\n\n", sparams.hostname.c_str(), sparams.port);
+
+    if (!svr.listen_after_bind())
+    {
+        return 1;
+    }
+
+    whisper_print_timings(ctx);
+    whisper_free(ctx);
+
+    return 0;
+}
diff --git a/src/whisper_cpp/ggml-alloc.c b/src/whisper_cpp/ggml-alloc.c
deleted file mode 100644
index f53954fb..00000000
--- a/src/whisper_cpp/ggml-alloc.c
+++ /dev/null
@@ -1,811 +0,0 @@
-#include "R.h"
-#include "ggml-alloc.h"
-#include "ggml-backend-impl.h"
-#include "ggml.h"
-#include "ggml-impl.h"
-#include <assert.h>
-#include <limits.h>
-#include <stdarg.h>
-#include <stdio.h>
-#include <stdlib.h>
-#include <string.h>
-
-#define MAX(a, b) ((a) > (b) ? (a) : (b))
-#define MAX_FREE_BLOCKS 256
-
-//#define GGML_ALLOCATOR_DEBUG
-
-//#define AT_PRINTF(...) RRprintf(__VA_ARGS__)
-#define AT_PRINTF(...)
-
-// TODO: GGML_PAD ?
-static size_t aligned_offset(const void * buffer, size_t offset, size_t alignment) {
-    assert(alignment && !(alignment & (alignment - 1))); // power of 2
-    size_t align = (alignment - (((uintptr_t)buffer + offset) % alignment)) % alignment;
-    return offset + align;
-}
-
-struct free_block {
-    void * addr;
-    size_t size;
-};
-
-struct ggml_tallocr {
-    struct ggml_backend_buffer * buffer;
-    bool buffer_owned;
-    void * base;
-    size_t alignment;
-
-    int n_free_blocks;
-    struct free_block free_blocks[MAX_FREE_BLOCKS];
-
-    size_t max_size;
-
-    bool measure;
-
-#ifdef GGML_ALLOCATOR_DEBUG
-    struct ggml_tensor * allocated_tensors[1024];
-#endif
-};
-
-#ifdef GGML_ALLOCATOR_DEBUG
-static void add_allocated_tensor(ggml_tallocr_t alloc, struct ggml_tensor * tensor) {
-    for (int i = 0; i < 1024; i++) {
-        if (alloc->allocated_tensors[i] == NULL) {
-            alloc->allocated_tensors[i] = tensor;
-            return;
-        }
-    }
-    GGML_ASSERT(!"out of allocated_tensors");
-}
-static void remove_allocated_tensor(ggml_tallocr_t alloc, struct ggml_tensor * tensor) {
-    for (int i = 0; i < 1024; i++) {
-        if (alloc->allocated_tensors[i] == tensor ||
-            (alloc->allocated_tensors[i] != NULL && alloc->allocated_tensors[i]->data == tensor->data)) {
-            alloc->allocated_tensors[i] = NULL;
-            return;
-        }
-    }
-    Rprintf("tried to free tensor %s not found\n", tensor->name);
-    GGML_ASSERT(!"tensor not found");
-}
-#endif
-
-// check if a tensor is allocated by this buffer
-static bool ggml_tallocr_is_own(ggml_tallocr_t alloc, const struct ggml_tensor * tensor) {
-    return tensor->buffer == alloc->buffer && (!tensor->view_src || tensor->view_src->buffer == alloc->buffer);
-}
-
-static bool ggml_is_view(struct ggml_tensor * t) {
-    return t->view_src != NULL;
-}
-
-void ggml_tallocr_alloc(ggml_tallocr_t alloc, struct ggml_tensor * tensor) {
-    GGML_ASSERT(!ggml_is_view(tensor)); // views generally get data pointer from one of their sources
-    GGML_ASSERT(tensor->data == NULL); // avoid allocating tensor which already has memory allocated
-
-    size_t size = ggml_backend_buffer_get_alloc_size(alloc->buffer, tensor);
-    size = aligned_offset(NULL, size, alloc->alignment);
-
-    AT_PRINTF("%s: allocating %s (%zu bytes) - ", __func__, tensor->name, size);
-
-    size_t max_avail = 0;
-
-    // find the best fitting free block besides the last block
-    int best_fit_block = -1;
-    size_t best_fit_size = SIZE_MAX;
-    for (int i = 0; i < alloc->n_free_blocks - 1; i++) {
-        struct free_block * block = &alloc->free_blocks[i];
-        max_avail = MAX(max_avail, block->size);
-        if (block->size >= size && block->size <= best_fit_size) {
-            best_fit_block = i;
-            best_fit_size = block->size;
-        }
-    }
-
-    AT_PRINTF("block %d\n", best_fit_block);
-
-    if (best_fit_block == -1) {
-        // the last block is our last resort
-        struct free_block * block = &alloc->free_blocks[alloc->n_free_blocks - 1];
-        max_avail = MAX(max_avail, block->size);
-        if (block->size >= size) {
-            best_fit_block = alloc->n_free_blocks - 1;
-        } else {
-            Rprintf("%s: not enough space in the buffer (needed %zu, largest block available %zu)\n",
-                    __func__, size, max_avail);
-            GGML_ASSERT(!"not enough space in the buffer");
-            return;
-        }
-    }
-    struct free_block * block = &alloc->free_blocks[best_fit_block];
-    void * addr = block->addr;
-    block->addr = (char*)block->addr + size;
-    block->size -= size;
-    if (block->size == 0) {
-        // remove block if empty
-        alloc->n_free_blocks--;
-        for (int j = best_fit_block; j < alloc->n_free_blocks; j++) {
-            alloc->free_blocks[j] = alloc->free_blocks[j+1];
-        }
-    }
-
-    tensor->data = addr;
-    tensor->buffer = alloc->buffer;
-    if (!alloc->measure) {
-        ggml_backend_buffer_init_tensor(alloc->buffer, tensor);
-    }
-
-#ifdef GGML_ALLOCATOR_DEBUG
-    add_allocated_tensor(alloc, tensor);
-    size_t cur_max = (char*)addr - (char*)alloc->base + size;
-    if (cur_max > alloc->max_size) {
-        Rprintf("max_size = %.2f MB: tensors: ", cur_max / 1024.0 / 1024.0);
-        for (int i = 0; i < 1024; i++) {
-            if (alloc->allocated_tensors[i]) {
-                Rprintf("%s (%.2f MB) ", alloc->allocated_tensors[i]->name, ggml_nbytes(alloc->allocated_tensors[i]) / 1024.0 / 1024.0);
-            }
-        }
-        Rprintf("\n");
-    }
-#endif
-
-    alloc->max_size = MAX(alloc->max_size, (char*)addr - (char*)alloc->base + size);
-}
-
-// this is a very naive implementation, but for our case the number of free blocks should be very small
-static void ggml_tallocr_free_tensor(ggml_tallocr_t alloc, struct ggml_tensor * tensor) {
-    if (ggml_tallocr_is_own(alloc, tensor) == false) {
-        // the tensor was not allocated in this buffer
-        // this can happen because the graph allocator will try to free weights and other tensors from different buffers
-        // the easiest way to deal with this is just to ignore it
-        // AT_PRINTF("ignoring %s (their buffer: %p, our buffer: %p)\n", tensor->name, (void *)tensor->buffer, (void *)alloc->buffer);
-        return;
-    }
-
-    void * ptr = tensor->data;
-
-    size_t size = ggml_backend_buffer_get_alloc_size(alloc->buffer, tensor);
-    size = aligned_offset(NULL, size, alloc->alignment);
-    AT_PRINTF("%s: freeing %s at %p (%zu bytes) - n_free_blocks = %d\n", __func__, tensor->name, ptr, size, alloc->n_free_blocks);
-
-#ifdef GGML_ALLOCATOR_DEBUG
-    remove_allocated_tensor(alloc, tensor);
-#endif
-
-    // see if we can merge with an existing block
-    for (int i = 0; i < alloc->n_free_blocks; i++) {
-        struct free_block * block = &alloc->free_blocks[i];
-        // check if ptr is at the end of the block
-        if ((char*)block->addr + block->size == ptr) {
-            block->size += size;
-            // check if we can merge with the next block
-            if (i < alloc->n_free_blocks - 1 && (char*)block->addr + block->size == alloc->free_blocks[i+1].addr) {
-                block->size += alloc->free_blocks[i+1].size;
-                alloc->n_free_blocks--;
-                for (int j = i+1; j < alloc->n_free_blocks; j++) {
-                    alloc->free_blocks[j] = alloc->free_blocks[j+1];
-                }
-            }
-            return;
-        }
-        // check if ptr is at the beginning of the block
-        if ((char*)ptr + size == block->addr) {
-            block->addr = ptr;
-            block->size += size;
-            // check if we can merge with the previous block
-            if (i > 0 && (char*)alloc->free_blocks[i-1].addr + alloc->free_blocks[i-1].size == block->addr) {
-                alloc->free_blocks[i-1].size += block->size;
-                alloc->n_free_blocks--;
-                for (int j = i; j < alloc->n_free_blocks; j++) {
-                    alloc->free_blocks[j] = alloc->free_blocks[j+1];
-                }
-            }
-            return;
-        }
-    }
-    // otherwise, add a new block
-    GGML_ASSERT(alloc->n_free_blocks < MAX_FREE_BLOCKS && "out of free blocks");
-    // insert the new block in the correct position to keep the array sorted by address (to make merging blocks faster)
-    int insert_pos = 0;
-    while (insert_pos < alloc->n_free_blocks && alloc->free_blocks[insert_pos].addr < ptr) {
-        insert_pos++;
-    }
-    // shift all blocks from insert_pos onward to make room for the new block
-    for (int i = alloc->n_free_blocks; i > insert_pos; i--) {
-        alloc->free_blocks[i] = alloc->free_blocks[i-1];
-    }
-    // insert the new block
-    alloc->free_blocks[insert_pos].addr = ptr;
-    alloc->free_blocks[insert_pos].size = size;
-    alloc->n_free_blocks++;
-}
-
-void ggml_tallocr_reset(ggml_tallocr_t alloc) {
-    alloc->n_free_blocks = 1;
-    size_t align_offset = aligned_offset(alloc->base, 0, alloc->alignment);
-    alloc->free_blocks[0].addr = (char *)alloc->base + align_offset;
-
-    if (alloc->measure) {
-        alloc->free_blocks[0].size = SIZE_MAX/2; // restrict maximum size of a measure allocator to half size_t max to avoid overflows
-    } else {
-        alloc->free_blocks[0].size = ggml_backend_buffer_get_size(alloc->buffer) - align_offset;
-    }
-}
-
-ggml_tallocr_t ggml_tallocr_new(void * data, size_t size, size_t alignment) {
-    struct ggml_backend_buffer * buffer = ggml_backend_cpu_buffer_from_ptr(data, size);
-
-    ggml_tallocr_t alloc = (ggml_tallocr_t)malloc(sizeof(struct ggml_tallocr));
-
-    *alloc = (struct ggml_tallocr) {
-        /*.buffer        = */ buffer,
-        /*.buffer_owned  = */ true,
-        /*.base          = */ ggml_backend_buffer_get_base(buffer),
-        /*.alignment     = */ alignment,
-        /*.n_free_blocks = */ 0,
-        /*.free_blocks   = */ {{0}},
-        /*.max_size      = */ 0,
-        /*.measure       = */ false,
-#ifdef GGML_ALLOCATOR_DEBUG
-        /*.allocated_tensors = */ {0},
-#endif
-    };
-
-    ggml_tallocr_reset(alloc);
-
-    return alloc;
-}
-
-ggml_tallocr_t ggml_tallocr_new_measure(size_t alignment) {
-    ggml_tallocr_t alloc = ggml_tallocr_new((void *)0x1000, SIZE_MAX/2, alignment);
-    alloc->measure = true;
-
-    return alloc;
-}
-
-ggml_tallocr_t ggml_tallocr_new_measure_from_backend(struct ggml_backend * backend) {
-    // create a backend buffer to get the correct tensor allocation sizes
-    ggml_backend_buffer_t buffer = ggml_backend_alloc_buffer(backend, 1);
-
-    // TODO: move alloc initialization to a common ggml_tallocr_new_impl function
-    ggml_tallocr_t alloc = ggml_tallocr_new_from_buffer(buffer);
-    alloc->buffer_owned = true;
-    alloc->measure = true;
-    ggml_tallocr_reset(alloc);
-    return alloc;
-}
-
-ggml_tallocr_t ggml_tallocr_new_from_backend(struct ggml_backend * backend, size_t size) {
-    ggml_backend_buffer_t buffer = ggml_backend_alloc_buffer(backend, size);
-    ggml_tallocr_t alloc = ggml_tallocr_new_from_buffer(buffer);
-    alloc->buffer_owned = true;
-    return alloc;
-}
-
-ggml_tallocr_t ggml_tallocr_new_from_buffer(struct ggml_backend_buffer * buffer) {
-    ggml_tallocr_t alloc = (ggml_tallocr_t)malloc(sizeof(struct ggml_tallocr));
-
-    *alloc = (struct ggml_tallocr) {
-        /*.buffer        = */ buffer,
-        /*.buffer_owned  = */ false,
-        /*.base          = */ ggml_backend_buffer_get_base(buffer),
-        /*.alignment     = */ ggml_backend_buffer_get_alignment(buffer),
-        /*.n_free_blocks = */ 0,
-        /*.free_blocks   = */ {{0}},
-        /*.max_size      = */ 0,
-        /*.measure       = */ false,
-#ifdef GGML_ALLOCATOR_DEBUG
-        /*.allocated_tensors = */ {0},
-#endif
-    };
-
-    ggml_tallocr_reset(alloc);
-
-    return alloc;
-}
-
-struct ggml_backend_buffer * ggml_tallocr_get_buffer(ggml_tallocr_t alloc) {
-    return alloc->buffer;
-}
-
-void ggml_tallocr_free(ggml_tallocr_t alloc) {
-    if (alloc == NULL) {
-        return;
-    }
-
-    if (alloc->buffer_owned) {
-        ggml_backend_buffer_free(alloc->buffer);
-    }
-    free(alloc);
-}
-
-bool ggml_tallocr_is_measure(ggml_tallocr_t alloc) {
-    return alloc->measure;
-}
-
-size_t ggml_tallocr_max_size(ggml_tallocr_t alloc) {
-    return alloc->max_size;
-}
-
-// graph allocator
-
-struct hash_node {
-    int n_children;
-    int n_views;
-};
-
-struct ggml_gallocr {
-    ggml_tallocr_t talloc;
-    struct ggml_hash_set hash_set;
-    struct hash_node * hash_values;
-    size_t hash_values_size;
-    ggml_tallocr_t * hash_allocs;
-    int * parse_seq;
-    int parse_seq_len;
-};
-
-ggml_gallocr_t ggml_gallocr_new(void) {
-    ggml_gallocr_t galloc = (ggml_gallocr_t)malloc(sizeof(struct ggml_gallocr));
-
-    *galloc = (struct ggml_gallocr) {
-        /*.talloc           = */ NULL,
-        /*.hash_set         = */ {0},
-        /*.hash_values      = */ NULL,
-        /*.hash_values_size = */ 0,
-        /*.hash_allocs      = */ NULL,
-        /*.parse_seq        = */ NULL,
-        /*.parse_seq_len    = */ 0,
-    };
-
-    return galloc;
-}
-
-void ggml_gallocr_free(ggml_gallocr_t galloc) {
-    if (galloc == NULL) {
-        return;
-    }
-
-    if (galloc->hash_set.keys != NULL) {
-        free(galloc->hash_set.keys);
-    }
-    if (galloc->hash_values != NULL) {
-        free(galloc->hash_values);
-    }
-    if (galloc->hash_allocs != NULL) {
-        free(galloc->hash_allocs);
-    }
-    if (galloc->parse_seq != NULL) {
-        free(galloc->parse_seq);
-    }
-    free(galloc);
-}
-
-void ggml_gallocr_set_parse_seq(ggml_gallocr_t galloc, const int * list, int n) {
-    free(galloc->parse_seq);
-    galloc->parse_seq = malloc(sizeof(int) * n);
-
-    for (int i = 0; i < n; i++) {
-        galloc->parse_seq[i] = list[i];
-    }
-    galloc->parse_seq_len = n;
-}
-
-static struct hash_node * hash_get(ggml_gallocr_t galloc, struct ggml_tensor * t) {
-    size_t i = ggml_hash_find_or_insert(galloc->hash_set, t);
-    return &galloc->hash_values[i];
-}
-
-static bool ggml_are_same_layout(const struct ggml_tensor * a, const struct ggml_tensor * b) {
-    if (a->type != b->type) {
-        return false;
-    }
-    for (int i = 0; i < GGML_MAX_DIMS; i++) {
-        if (a->ne[i] != b->ne[i]) {
-            return false;
-        }
-        if (a->nb[i] != b->nb[i]) {
-            return false;
-        }
-    }
-    return true;
-}
-
-static bool ggml_op_can_inplace(enum ggml_op op) {
-    switch (op) {
-        case GGML_OP_SCALE:
-        case GGML_OP_DIAG_MASK_ZERO:
-        case GGML_OP_DIAG_MASK_INF:
-        case GGML_OP_ADD:
-        case GGML_OP_ADD1:
-        case GGML_OP_SUB:
-        case GGML_OP_MUL:
-        case GGML_OP_DIV:
-        case GGML_OP_SQR:
-        case GGML_OP_SQRT:
-        case GGML_OP_LOG:
-        case GGML_OP_UNARY:
-        case GGML_OP_ROPE:
-        case GGML_OP_RMS_NORM:
-        case GGML_OP_SOFT_MAX:
-            return true;
-
-        default:
-            return false;
-    }
-}
-
-static ggml_tallocr_t node_tallocr(ggml_gallocr_t galloc, struct ggml_tensor * node) {
-    if (galloc->talloc != NULL) {
-        return galloc->talloc;
-    }
-
-    return galloc->hash_allocs[ggml_hash_find_or_insert(galloc->hash_set, node)];
-}
-
-static void init_view(ggml_gallocr_t galloc, struct ggml_tensor * view, bool update_backend) {
-    ggml_tallocr_t alloc = node_tallocr(galloc, view);
-
-    GGML_ASSERT(view->view_src != NULL && view->view_src->data != NULL);
-    if (update_backend) {
-        view->backend = view->view_src->backend;
-    }
-    // views are initialized in the alloc buffer rather than the view_src buffer
-    view->buffer  = alloc->buffer;
-    view->data    = (char *)view->view_src->data + view->view_offs;
-
-    assert(ggml_tallocr_is_measure(alloc) || !view->buffer || view->buffer->buft == alloc->buffer->buft);
-
-    if (!alloc->measure) {
-        ggml_backend_buffer_init_tensor(alloc->buffer, view);
-    }
-}
-
-static void allocate_node(ggml_gallocr_t galloc, struct ggml_tensor * node) {
-    ggml_tallocr_t alloc = node_tallocr(galloc, node);
-
-    if (node->data == NULL) {
-        if (ggml_is_view(node)) {
-            init_view(galloc, node, true);
-        } else {
-            // see if we can reuse a parent's buffer (inplace)
-            if (ggml_op_can_inplace(node->op)) {
-                for (int i = 0; i < GGML_MAX_SRC; i++) {
-                    struct ggml_tensor * parent = node->src[i];
-                    if (parent == NULL) {
-                        break;
-                    }
-
-                    // if the node's data is external, then we cannot re-use it
-                    if (ggml_tallocr_is_own(alloc, parent) == false) {
-                        AT_PRINTF("not reusing parent %s for %s as %p is external\n", parent->name, node->name, parent->data);
-                        continue;
-                    }
-
-                    struct hash_node * p_hn = hash_get(galloc, parent);
-                    if (parent->data != NULL && p_hn->n_children == 1 && p_hn->n_views == 0 && ggml_are_same_layout(node, parent)) {
-                        if (ggml_is_view(parent)) {
-                            struct ggml_tensor * view_src = parent->view_src;
-                            struct hash_node * view_src_hn = hash_get(galloc, view_src);
-                            if (view_src_hn->n_views == 1 && view_src_hn->n_children == 0 && view_src->data == parent->data) {
-                                // TODO: the offset of the view parent must be kept to ensure that the op doesn't overwrite
-                                // the parent's data that it will need later (same layout requirement). the problem is that then
-                                // we cannot free the tensor because the original address of the allocation is lost.
-                                // adding a view_src pointer to the tensor would solve this and simplify the code dealing with views
-                                // for now, we only reuse the parent's data if the offset is zero (view_src->data == parent->data)
-                                AT_PRINTF("reusing view parent %s (%s) for %s\n", parent->name, view_src->name, node->name);
-                                node->view_src = view_src;
-                                view_src_hn->n_views += 1;
-                                init_view(galloc, node, false);
-                                return;
-                            }
-                        } else {
-                            AT_PRINTF("reusing parent %s for %s\n", parent->name, node->name);
-                            node->view_src = parent;
-                            p_hn->n_views += 1;
-                            init_view(galloc, node, false);
-                            return;
-                        }
-                    }
-                }
-            }
-            ggml_tallocr_alloc(alloc, node);
-        }
-    }
-}
-
-static void free_node(ggml_gallocr_t galloc, struct ggml_tensor * node) {
-    ggml_tallocr_t alloc = node_tallocr(galloc, node);
-
-    ggml_tallocr_free_tensor(alloc, node);
-}
-
-static void ggml_tallocr_alloc_graph_impl(ggml_gallocr_t galloc, struct ggml_cgraph * gf) {
-    const int * parse_seq     = galloc->parse_seq;
-    int         parse_seq_len = galloc->parse_seq_len;
-
-    // count number of children and views
-    for (int i = 0; i < gf->n_nodes; i++) {
-        struct ggml_tensor * node = gf->nodes[i];
-
-        if (ggml_is_view(node)) {
-            struct ggml_tensor * view_src = node->view_src;
-            hash_get(galloc, view_src)->n_views += 1;
-            if (node->buffer == NULL && node->data != NULL) {
-                // view of a pre-allocated tensor, didn't call init_view() yet
-                init_view(galloc, node, true);
-            }
-        }
-
-        for (int j = 0; j < GGML_MAX_SRC; j++) {
-            struct ggml_tensor * parent = node->src[j];
-            if (parent == NULL) {
-                break;
-            }
-            hash_get(galloc, parent)->n_children += 1;
-            if (ggml_is_view(parent) && parent->buffer == NULL && parent->data != NULL) {
-                init_view(galloc, parent, true);
-            }
-        }
-   }
-
-    // allocate tensors
-    // if we have parse_seq then we allocate nodes following the list, and we only free nodes at barriers
-    int last_barrier_pos = 0;
-    int n_nodes = parse_seq_len ? parse_seq_len : gf->n_nodes;
-
-    for (int ind = 0; ind < n_nodes; ind++) {
-        // allocate a node if there is no parse_seq or this is not a barrier
-        if (parse_seq_len == 0 || parse_seq[ind] != -1) {
-            int i = parse_seq_len ? parse_seq[ind] : ind;
-            struct ggml_tensor * node = gf->nodes[i];
-
-            // allocate parents (leafs)
-            for (int j = 0; j < GGML_MAX_SRC; j++) {
-                struct ggml_tensor * parent = node->src[j];
-                if (parent == NULL) {
-                    break;
-                }
-                allocate_node(galloc, parent);
-            }
-
-            // allocate node
-            allocate_node(galloc, node);
-
-            AT_PRINTF("exec: %s (%s) <= ", ggml_op_name(node->op), node->name);
-            for (int j = 0; j < GGML_MAX_SRC; j++) {
-                struct ggml_tensor * parent = node->src[j];
-                if (parent == NULL) {
-                    break;
-                }
-                AT_PRINTF("%s", parent->name);
-                if (j < GGML_MAX_SRC - 1 && node->src[j + 1] != NULL) {
-                    AT_PRINTF(", ");
-                }
-            }
-            AT_PRINTF("\n");
-        }
-
-        // update parents
-        // update immediately if there is no parse_seq
-        // update only at barriers if there is parse_seq
-        if ((parse_seq_len == 0) || parse_seq[ind] == -1) {
-            int update_start = parse_seq_len ? last_barrier_pos : ind;
-            int update_end   = parse_seq_len ? ind              : ind + 1;
-            for (int i = update_start; i < update_end; i++) {
-                int node_i = parse_seq_len ? parse_seq[i] : i;
-                struct ggml_tensor * node = gf->nodes[node_i];
-
-                for (int j = 0; j < GGML_MAX_SRC; j++) {
-                    struct ggml_tensor * parent = node->src[j];
-                    if (parent == NULL) {
-                        break;
-                    }
-                    struct hash_node * p_hn = hash_get(galloc, parent);
-                    p_hn->n_children -= 1;
-
-                    //AT_PRINTF("parent %s: %d children, %d views\n", parent->name, parent->n_children, parent->n_views);
-
-                    if (p_hn->n_children == 0 && p_hn->n_views == 0) {
-                        if (ggml_is_view(parent)) {
-                            struct ggml_tensor * view_src = parent->view_src;
-                            struct hash_node * view_src_hn = hash_get(galloc, view_src);
-                            view_src_hn->n_views -= 1;
-                            AT_PRINTF("view_src %s: %d children, %d views\n", view_src->name, view_src_hn->n_children, view_src_hn->n_views);
-                            if (view_src_hn->n_views == 0 && view_src_hn->n_children == 0) {
-                                free_node(galloc, view_src);
-                            }
-                        }
-                        else {
-                            free_node(galloc, parent);
-                        }
-                    }
-                }
-            }
-            AT_PRINTF("\n");
-            if (parse_seq_len) {
-                last_barrier_pos = ind + 1;
-            }
-        }
-    }
-}
-
-size_t ggml_gallocr_alloc_graph(ggml_gallocr_t galloc, ggml_tallocr_t talloc, struct ggml_cgraph * graph) {
-    size_t hash_size = graph->visited_hash_table.size;
-
-    // check if the hash table is initialized and large enough
-    if (galloc->hash_set.size < hash_size) {
-        if (galloc->hash_set.keys != NULL) {
-            free(galloc->hash_set.keys);
-        }
-        if (galloc->hash_values != NULL) {
-            free(galloc->hash_values);
-        }
-        galloc->hash_set.keys = malloc(sizeof(struct ggml_tensor *) * hash_size);
-        galloc->hash_set.size = hash_size;
-        galloc->hash_values = malloc(sizeof(struct hash_node) * hash_size);
-    }
-
-    // reset hash table
-    memset(galloc->hash_set.keys, 0, sizeof(struct ggml_tensor *) * hash_size);
-    memset(galloc->hash_values,   0, sizeof(struct hash_node) * hash_size);
-
-    galloc->talloc = talloc;
-    ggml_tallocr_alloc_graph_impl(galloc, graph);
-    galloc->talloc = NULL;
-
-    size_t max_size = ggml_tallocr_max_size(talloc);
-
-    return max_size;
-}
-
-void ggml_gallocr_alloc_graph_n(ggml_gallocr_t galloc, struct ggml_cgraph * graph, struct ggml_hash_set hash_set, ggml_tallocr_t * hash_node_talloc) {
-    const size_t hash_size = hash_set.size;
-
-    GGML_ASSERT(hash_size >= (size_t)(graph->n_nodes + graph->n_leafs));
-
-    galloc->talloc = NULL;
-
-    // alloc hash_values if needed
-    if (galloc->hash_values == NULL || galloc->hash_values_size < hash_size) {
-        free(galloc->hash_values);
-        galloc->hash_values      = malloc(sizeof(struct hash_node) * hash_size);
-        galloc->hash_values_size = hash_size;
-    }
-
-    // free hash_set.keys if needed
-    if (galloc->hash_set.keys != NULL) {
-        free(galloc->hash_set.keys);
-    }
-    galloc->hash_set = hash_set;
-
-    // reset hash values
-    memset(galloc->hash_values, 0, sizeof(struct hash_node) * hash_size);
-
-    galloc->hash_allocs = hash_node_talloc;
-
-    ggml_tallocr_alloc_graph_impl(galloc, graph);
-
-    // remove unowned resources
-    galloc->hash_set.keys = NULL;
-    galloc->hash_allocs = NULL;
-}
-
-// legacy API wrapper
-
-struct ggml_allocr {
-    ggml_tallocr_t talloc;
-    ggml_gallocr_t galloc;
-};
-
-static ggml_allocr_t ggml_allocr_new_impl(ggml_tallocr_t talloc) {
-    ggml_allocr_t alloc = (ggml_allocr_t)malloc(sizeof(struct ggml_allocr));
-    *alloc = (struct ggml_allocr) {
-        /*.talloc = */ talloc,
-        /*.galloc = */ ggml_gallocr_new(),
-    };
-    return alloc;
-}
-
-ggml_allocr_t ggml_allocr_new(void * data, size_t size, size_t alignment) {
-    return ggml_allocr_new_impl(ggml_tallocr_new(data, size, alignment));
-}
-
-ggml_allocr_t ggml_allocr_new_measure(size_t alignment) {
-    return ggml_allocr_new_impl(ggml_tallocr_new_measure(alignment));
-}
-
-ggml_allocr_t ggml_allocr_new_from_buffer(struct ggml_backend_buffer * buffer) {
-    return ggml_allocr_new_impl(ggml_tallocr_new_from_buffer(buffer));
-}
-
-ggml_allocr_t ggml_allocr_new_from_backend(struct ggml_backend * backend, size_t size) {
-    return ggml_allocr_new_impl(ggml_tallocr_new_from_backend(backend, size));
-}
-
-ggml_allocr_t ggml_allocr_new_measure_from_backend(struct ggml_backend * backend) {
-    return ggml_allocr_new_impl(ggml_tallocr_new_measure_from_backend(backend));
-}
-
-struct ggml_backend_buffer * ggml_allocr_get_buffer(ggml_allocr_t alloc) {
-    return ggml_tallocr_get_buffer(alloc->talloc);
-}
-
-void ggml_allocr_set_parse_seq(ggml_allocr_t alloc, const int * list, int n) {
-    ggml_gallocr_set_parse_seq(alloc->galloc, list, n);
-}
-
-void ggml_allocr_free(ggml_allocr_t alloc) {
-    if (alloc == NULL) {
-        return;
-    }
-
-    ggml_gallocr_free(alloc->galloc);
-    ggml_tallocr_free(alloc->talloc);
-    free(alloc);
-}
-
-bool ggml_allocr_is_measure(ggml_allocr_t alloc) {
-    return ggml_tallocr_is_measure(alloc->talloc);
-}
-
-void ggml_allocr_reset(ggml_allocr_t alloc) {
-    ggml_tallocr_reset(alloc->talloc);
-}
-
-void ggml_allocr_alloc(ggml_allocr_t alloc, struct ggml_tensor * tensor) {
-    ggml_tallocr_alloc(alloc->talloc, tensor);
-}
-
-size_t ggml_allocr_max_size(ggml_allocr_t alloc) {
-    return ggml_tallocr_max_size(alloc->talloc);
-}
-
-size_t ggml_allocr_alloc_graph(ggml_allocr_t alloc, struct ggml_cgraph * graph) {
-    return ggml_gallocr_alloc_graph(alloc->galloc, alloc->talloc, graph);
-}
-
-// utils
-ggml_backend_buffer_t ggml_backend_alloc_ctx_tensors_from_buft(struct ggml_context * ctx, ggml_backend_buffer_type_t buft) {
-    GGML_ASSERT(ggml_get_no_alloc(ctx) == true);
-
-    size_t alignment = ggml_backend_buft_get_alignment(buft);
-
-    size_t nbytes = 0;
-    for (struct ggml_tensor * t = ggml_get_first_tensor(ctx); t != NULL; t = ggml_get_next_tensor(ctx, t)) {
-        if (t->data == NULL && t->view_src == NULL) {
-            nbytes += GGML_PAD(ggml_backend_buft_get_alloc_size(buft, t), alignment);
-        }
-    }
-
-    if (nbytes == 0) {
-        // all the tensors in the context are already allocated
-        return NULL;
-    }
-
-    ggml_backend_buffer_t buffer = ggml_backend_buft_alloc_buffer(buft, nbytes);
-    ggml_tallocr_t tallocr = ggml_tallocr_new_from_buffer(buffer);
-
-    for (struct ggml_tensor * t = ggml_get_first_tensor(ctx); t != NULL; t = ggml_get_next_tensor(ctx, t)) {
-        if (t->data == NULL) {
-            if (t->view_src == NULL) {
-                ggml_tallocr_alloc(tallocr, t);
-            } else {
-                ggml_backend_view_init(buffer, t);
-            }
-        } else {
-            if (t->view_src != NULL) {
-                // view of a pre-allocated tensor
-                ggml_backend_view_init(buffer, t);
-            }
-        }
-    }
-
-    ggml_tallocr_free(tallocr);
-
-    return buffer;
-}
-
-ggml_backend_buffer_t ggml_backend_alloc_ctx_tensors(struct ggml_context * ctx, ggml_backend_t backend) {
-    return ggml_backend_alloc_ctx_tensors_from_buft(ctx, ggml_backend_get_default_buffer_type(backend));
-}
diff --git a/src/whisper_cpp/ggml-alloc.h b/src/whisper_cpp/ggml-alloc.h
deleted file mode 100644
index 64a41246..00000000
--- a/src/whisper_cpp/ggml-alloc.h
+++ /dev/null
@@ -1,92 +0,0 @@
-#pragma once
-
-#include "ggml.h"
-
-#ifdef  __cplusplus
-extern "C" {
-#endif
-
-struct ggml_backend;
-struct ggml_backend_buffer;
-struct ggml_backend_buffer_type;
-
-//
-// Legacy API
-//
-
-typedef struct ggml_allocr * ggml_allocr_t;
-
-// initialize allocator for use with CPU backend only
-GGML_API ggml_allocr_t ggml_allocr_new(void * data, size_t size, size_t alignment);
-GGML_API ggml_allocr_t ggml_allocr_new_measure(size_t alignment);
-
-// initialize allocator for use with ggml-backend
-GGML_API ggml_allocr_t ggml_allocr_new_from_buffer(struct ggml_backend_buffer * buffer);
-GGML_API ggml_allocr_t ggml_allocr_new_from_backend(struct ggml_backend * backend, size_t size); // allocates an owned buffer
-GGML_API ggml_allocr_t ggml_allocr_new_measure_from_backend(struct ggml_backend * backend);
-
-GGML_API struct ggml_backend_buffer * ggml_allocr_get_buffer(ggml_allocr_t alloc);
-
-// tell the allocator to parse nodes following the order described in the list
-// you should call this if your graph are optimized to execute out-of-order
-GGML_API void   ggml_allocr_set_parse_seq(ggml_allocr_t alloc, const int * list, int n);
-
-GGML_API void   ggml_allocr_free       (ggml_allocr_t alloc);
-GGML_API bool   ggml_allocr_is_measure (ggml_allocr_t alloc);
-GGML_API void   ggml_allocr_reset      (ggml_allocr_t alloc);
-GGML_API void   ggml_allocr_alloc      (ggml_allocr_t alloc, struct ggml_tensor * tensor);
-GGML_API size_t ggml_allocr_max_size   (ggml_allocr_t alloc);
-
-GGML_API size_t ggml_allocr_alloc_graph(ggml_allocr_t alloc, struct ggml_cgraph * graph);
-
-//
-// ggml-backend v2 API
-//
-
-// Separate tensor and graph allocator objects
-// This is necessary for multi-backend allocation because the graph allocator needs to use multiple tensor allocators
-// The original API is kept as a wrapper around the new API
-
-// Tensor allocator
-typedef struct ggml_tallocr * ggml_tallocr_t;
-
-GGML_API ggml_tallocr_t ggml_tallocr_new(void * data, size_t size, size_t alignment);
-GGML_API ggml_tallocr_t ggml_tallocr_new_measure(size_t alignment);
-GGML_API ggml_tallocr_t ggml_tallocr_new_from_buffer(struct ggml_backend_buffer * buffer);
-GGML_API ggml_tallocr_t ggml_tallocr_new_from_backend(struct ggml_backend * backend, size_t size); // allocates an owned buffer
-GGML_API ggml_tallocr_t ggml_tallocr_new_measure_from_backend(struct ggml_backend * backend);
-
-GGML_API struct ggml_backend_buffer * ggml_tallocr_get_buffer(ggml_tallocr_t talloc);
-
-GGML_API void   ggml_tallocr_free       (ggml_tallocr_t talloc);
-GGML_API bool   ggml_tallocr_is_measure (ggml_tallocr_t talloc);
-GGML_API void   ggml_tallocr_reset      (ggml_tallocr_t talloc);
-GGML_API void   ggml_tallocr_alloc      (ggml_tallocr_t talloc, struct ggml_tensor * tensor);
-GGML_API size_t ggml_tallocr_max_size   (ggml_tallocr_t talloc);
-
-
-// Graph allocator
-typedef struct ggml_gallocr * ggml_gallocr_t;
-
-GGML_API ggml_gallocr_t ggml_gallocr_new(void);
-GGML_API void   ggml_gallocr_free(ggml_gallocr_t galloc);
-
-GGML_API void   ggml_gallocr_set_parse_seq(ggml_gallocr_t galloc, const int * list, int n);
-GGML_API size_t ggml_gallocr_alloc_graph(ggml_gallocr_t galloc, ggml_tallocr_t talloc, struct ggml_cgraph * graph);
-
-// Allocate tensors from the allocators given by the hash table
-GGML_API void   ggml_gallocr_alloc_graph_n(
-                    ggml_gallocr_t galloc,
-                    struct ggml_cgraph * graph,
-                    struct ggml_hash_set hash_set,
-                    ggml_tallocr_t * hash_node_talloc);
-
-
-// Utils
-// Create a buffer and allocate all the tensors in a ggml_context
-GGML_API struct ggml_backend_buffer * ggml_backend_alloc_ctx_tensors_from_buft(struct ggml_context * ctx, struct ggml_backend_buffer_type * buft);
-GGML_API struct ggml_backend_buffer * ggml_backend_alloc_ctx_tensors(struct ggml_context * ctx, struct ggml_backend * backend);
-
-#ifdef  __cplusplus
-}
-#endif
diff --git a/src/whisper_cpp/ggml-backend-impl.h b/src/whisper_cpp/ggml-backend-impl.h
deleted file mode 100644
index ca21b474..00000000
--- a/src/whisper_cpp/ggml-backend-impl.h
+++ /dev/null
@@ -1,116 +0,0 @@
-#pragma once
-
-// ggml-backend internal header
-
-#include "ggml-backend.h"
-
-#ifdef  __cplusplus
-extern "C" {
-#endif
-
-    //
-    // Backend buffer
-    //
-
-    // buffer type
-    typedef void * ggml_backend_buffer_type_context_t;
-
-    struct ggml_backend_buffer_type_i {
-        ggml_backend_buffer_t (*alloc_buffer)    (ggml_backend_buffer_type_t buft, size_t size);
-        size_t                (*get_alignment)   (ggml_backend_buffer_type_t buft); // tensor alignment
-        size_t                (*get_alloc_size)  (ggml_backend_buffer_type_t buft, struct ggml_tensor * tensor); // data size needed to allocate the tensor, including padding
-        bool                  (*supports_backend)(ggml_backend_buffer_type_t buft, ggml_backend_t backend); // check if the buffer type is usable by the backend
-        // check if tensor data is in host memory
-        // should be equivalent to supports_backend(buft, ggml_backend_cpu_init())
-        bool                  (*is_host)         (ggml_backend_buffer_type_t buft);
-    };
-
-    struct ggml_backend_buffer_type {
-        struct ggml_backend_buffer_type_i  iface;
-        ggml_backend_buffer_type_context_t context;
-    };
-
-    // buffer
-    typedef void * ggml_backend_buffer_context_t;
-
-    struct ggml_backend_buffer_i {
-        void   (*free_buffer)    (ggml_backend_buffer_t buffer);
-        //void     (*reset)      (ggml_backend_buffer_t buffer); // reset any internal state due to tensor initialization, such as tensor extras
-        void * (*get_base)       (ggml_backend_buffer_t buffer);
-        void   (*init_tensor)    (ggml_backend_buffer_t buffer, struct ggml_tensor * tensor);
-        void   (*set_tensor)     (ggml_backend_buffer_t buffer,       struct ggml_tensor * tensor, const void * data, size_t offset, size_t size);
-        void   (*get_tensor)     (ggml_backend_buffer_t buffer, const struct ggml_tensor * tensor,       void * data, size_t offset, size_t size);
-        // (optional) copy tensor between different buffer-type, allow for single-copy tranfers
-        void   (*cpy_tensor_from)(ggml_backend_buffer_t buffer, struct ggml_tensor * src, struct ggml_tensor * dst);
-        void   (*cpy_tensor_to)  (ggml_backend_buffer_t buffer, struct ggml_tensor * src, struct ggml_tensor * dst);
-        void   (*clear)          (ggml_backend_buffer_t buffer, uint8_t value);
-    };
-
-    struct ggml_backend_buffer {
-        struct ggml_backend_buffer_i  iface;
-        ggml_backend_buffer_type_t    buft;
-        ggml_backend_buffer_context_t context;
-        size_t size;
-    };
-
-    ggml_backend_buffer_t ggml_backend_buffer_init(
-                   ggml_backend_buffer_type_t      buft,
-            struct ggml_backend_buffer_i           iface,
-                   ggml_backend_buffer_context_t   context,
-                   size_t                          size);
-
-
-    //
-    // Backend
-    //
-
-    typedef void * ggml_backend_context_t;
-
-    struct ggml_backend_i {
-        const char * (*get_name)(ggml_backend_t backend);
-
-        void (*free)(ggml_backend_t backend);
-
-        // buffer allocation
-        ggml_backend_buffer_type_t (*get_default_buffer_type)(ggml_backend_t backend);
-
-        // (optional) asynchroneous tensor data access
-        void (*set_tensor_async)(ggml_backend_t backend,       struct ggml_tensor * tensor, const void * data, size_t offset, size_t size);
-        void (*get_tensor_async)(ggml_backend_t backend, const struct ggml_tensor * tensor,       void * data, size_t offset, size_t size);
-
-        // (optional) asynchroneous tensor copy
-        void (*cpy_tensor_from_async)(ggml_backend_t backend, struct ggml_tensor * src, struct ggml_tensor * dst);
-        void (*cpy_tensor_to_async)  (ggml_backend_t backend, struct ggml_tensor * src, struct ggml_tensor * dst);
-
-        void (*synchronize)(ggml_backend_t backend);
-
-        // compute graph with a plan
-        ggml_backend_graph_plan_t (*graph_plan_create) (ggml_backend_t backend, struct ggml_cgraph * cgraph);
-        void                      (*graph_plan_free)   (ggml_backend_t backend, ggml_backend_graph_plan_t plan);
-        void                      (*graph_plan_compute)(ggml_backend_t backend, ggml_backend_graph_plan_t plan);
-
-        // compute graph without a plan
-        bool (*graph_compute)(ggml_backend_t backend, struct ggml_cgraph * cgraph);
-
-        // check if the backend supports an operation
-        bool (*supports_op)(ggml_backend_t backend, const struct ggml_tensor * op);
-    };
-
-    struct ggml_backend {
-        struct ggml_backend_i iface;
-
-        ggml_backend_context_t context;
-    };
-
-
-    //
-    // Backend registry
-    //
-
-    typedef ggml_backend_t (*ggml_backend_init_fn)(const char * params, void * user_data);
-
-    void ggml_backend_register(const char * name, ggml_backend_init_fn init_fn, ggml_backend_buffer_type_t default_buffer_type, void * user_data);
-
-#ifdef  __cplusplus
-}
-#endif
diff --git a/src/whisper_cpp/ggml-backend.c b/src/whisper_cpp/ggml-backend.c
deleted file mode 100644
index 8a26515e..00000000
--- a/src/whisper_cpp/ggml-backend.c
+++ /dev/null
@@ -1,1432 +0,0 @@
-#include "R.h"
-#include "ggml-backend-impl.h"
-#include "ggml-alloc.h"
-#include "ggml-impl.h"
-
-#include <assert.h>
-#include <limits.h>
-#include <stdarg.h>
-#include <stdio.h>
-#include <stdlib.h>
-#include <string.h>
-
-
-#define MAX(a, b) ((a) > (b) ? (a) : (b))
-
-
-// backend buffer type
-
-ggml_backend_buffer_t ggml_backend_buft_alloc_buffer(ggml_backend_buffer_type_t buft, size_t size) {
-    return buft->iface.alloc_buffer(buft, size);
-}
-
-size_t ggml_backend_buft_get_alignment(ggml_backend_buffer_type_t buft) {
-    return buft->iface.get_alignment(buft);
-}
-
-size_t ggml_backend_buft_get_alloc_size(ggml_backend_buffer_type_t buft, struct ggml_tensor * tensor) {
-    // get_alloc_size is optional, defaults to ggml_nbytes
-    if (buft->iface.get_alloc_size) {
-        return buft->iface.get_alloc_size(buft, tensor);
-    }
-    return ggml_nbytes(tensor);
-}
-
-bool ggml_backend_buft_supports_backend(ggml_backend_buffer_type_t buft, ggml_backend_t backend) {
-    return buft->iface.supports_backend(buft, backend);
-}
-
-bool ggml_backend_buft_is_host(ggml_backend_buffer_type_t buft) {
-    if (buft->iface.is_host) {
-        return buft->iface.is_host(buft);
-    }
-    return false;
-}
-
-// backend buffer
-
-ggml_backend_buffer_t ggml_backend_buffer_init(
-               ggml_backend_buffer_type_t      buft,
-        struct ggml_backend_buffer_i           iface,
-               ggml_backend_buffer_context_t   context,
-               size_t                          size) {
-    ggml_backend_buffer_t buffer = malloc(sizeof(struct ggml_backend_buffer));
-
-    GGML_ASSERT(iface.get_base != NULL);
-
-    (*buffer) = (struct ggml_backend_buffer) {
-        /* .interface = */ iface,
-        /* .buft      = */ buft,
-        /* .context   = */ context,
-        /* .size      = */ size,
-    };
-
-    return buffer;
-}
-
-void ggml_backend_buffer_free(ggml_backend_buffer_t buffer) {
-    if (buffer == NULL) {
-        return;
-    }
-
-    if (buffer->iface.free_buffer != NULL) {
-        buffer->iface.free_buffer(buffer);
-    }
-    free(buffer);
-}
-
-size_t ggml_backend_buffer_get_size(ggml_backend_buffer_t buffer) {
-    return buffer->size;
-}
-
-void * ggml_backend_buffer_get_base(ggml_backend_buffer_t buffer) {
-    void * base = buffer->iface.get_base(buffer);
-
-    GGML_ASSERT(base != NULL && "backend buffer base cannot be NULL");
-
-    return base;
-}
-
-void ggml_backend_buffer_init_tensor(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor) {
-    // init_tensor is optional
-    if (buffer->iface.init_tensor) {
-        buffer->iface.init_tensor(buffer, tensor);
-    }
-}
-
-size_t ggml_backend_buffer_get_alignment (ggml_backend_buffer_t buffer) {
-    return ggml_backend_buft_get_alignment(ggml_backend_buffer_type(buffer));
-}
-
-size_t ggml_backend_buffer_get_alloc_size(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor) {
-    return ggml_backend_buft_get_alloc_size(ggml_backend_buffer_type(buffer), tensor);
-}
-
-void ggml_backend_buffer_clear(ggml_backend_buffer_t buffer, uint8_t value) {
-    buffer->iface.clear(buffer, value);
-}
-
-bool ggml_backend_buffer_is_host(ggml_backend_buffer_t buffer) {
-    return ggml_backend_buft_is_host(ggml_backend_buffer_type(buffer));
-}
-
-ggml_backend_buffer_type_t ggml_backend_buffer_type(ggml_backend_buffer_t buffer) {
-    return buffer->buft;
-}
-
-// backend
-
-const char * ggml_backend_name(ggml_backend_t backend) {
-    if (backend == NULL) {
-        return "NULL";
-    }
-    return backend->iface.get_name(backend);
-}
-
-void ggml_backend_free(ggml_backend_t backend) {
-    if (backend == NULL) {
-        return;
-    }
-
-    backend->iface.free(backend);
-}
-
-ggml_backend_buffer_type_t ggml_backend_get_default_buffer_type(ggml_backend_t backend) {
-    return backend->iface.get_default_buffer_type(backend);
-}
-
-ggml_backend_buffer_t ggml_backend_alloc_buffer(ggml_backend_t backend, size_t size) {
-    return ggml_backend_buft_alloc_buffer(ggml_backend_get_default_buffer_type(backend), size);
-}
-
-size_t ggml_backend_get_alignment(ggml_backend_t backend) {
-    return ggml_backend_buft_get_alignment(ggml_backend_get_default_buffer_type(backend));
-}
-
-void ggml_backend_tensor_set_async(ggml_backend_t backend, struct ggml_tensor * tensor, const void * data, size_t offset, size_t size) {
-    GGML_ASSERT(tensor->data != NULL && "tensor not allocated");
-    GGML_ASSERT(offset + size <= ggml_nbytes(tensor) && "tensor write out of bounds");
-
-    backend->iface.set_tensor_async(backend, tensor, data, offset, size);
-}
-
-void ggml_backend_tensor_get_async(ggml_backend_t backend, const struct ggml_tensor * tensor, void * data, size_t offset, size_t size) {
-    GGML_ASSERT(tensor->data != NULL && "tensor not allocated");
-    GGML_ASSERT(offset + size <= ggml_nbytes(tensor) && "tensor read out of bounds");
-
-    backend->iface.get_tensor_async(backend, tensor, data, offset, size);
-}
-
-void ggml_backend_tensor_set(struct ggml_tensor * tensor, const void * data, size_t offset, size_t size) {
-    GGML_ASSERT(tensor->data != NULL && "tensor not allocated");
-    GGML_ASSERT(tensor->buffer != NULL && "tensor buffer not set");
-    GGML_ASSERT(offset + size <= ggml_nbytes(tensor) && "tensor write out of bounds");
-
-    tensor->buffer->iface.set_tensor(tensor->buffer, tensor, data, offset, size);
-}
-
-void ggml_backend_tensor_get(const struct ggml_tensor * tensor, void * data, size_t offset, size_t size) {
-    GGML_ASSERT(tensor->data != NULL && "tensor not allocated");
-    GGML_ASSERT(tensor->buffer != NULL && "tensor buffer not set");
-    GGML_ASSERT(offset + size <= ggml_nbytes(tensor) && "tensor read out of bounds");
-
-    tensor->buffer->iface.get_tensor(tensor->buffer, tensor, data, offset, size);
-}
-
-void ggml_backend_synchronize(ggml_backend_t backend) {
-    if (backend->iface.synchronize == NULL) {
-        return;
-    }
-
-    backend->iface.synchronize(backend);
-}
-
-ggml_backend_graph_plan_t ggml_backend_graph_plan_create(ggml_backend_t backend, struct ggml_cgraph * cgraph) {
-    return backend->iface.graph_plan_create(backend, cgraph);
-}
-
-void ggml_backend_graph_plan_free(ggml_backend_t backend, ggml_backend_graph_plan_t plan) {
-    backend->iface.graph_plan_free(backend, plan);
-}
-
-void ggml_backend_graph_plan_compute(ggml_backend_t backend, ggml_backend_graph_plan_t plan) {
-    backend->iface.graph_plan_compute(backend, plan);
-
-    // TODO: optional sync
-    ggml_backend_synchronize(backend);
-}
-
-bool ggml_backend_graph_compute(ggml_backend_t backend, struct ggml_cgraph * cgraph) {
-    if (!backend->iface.graph_compute(backend, cgraph)) {
-        return false;
-    }
-
-    // TODO: optional sync
-    ggml_backend_synchronize(backend);
-    return true;
-}
-
-bool ggml_backend_supports_op(ggml_backend_t backend, const struct ggml_tensor * op) {
-    return backend->iface.supports_op(backend, op);
-}
-
-// backend copy
-
-static bool ggml_are_same_layout(const struct ggml_tensor * a, const struct ggml_tensor * b) {
-    if (a->type != b->type) {
-        return false;
-    }
-    for (int i = 0; i < GGML_MAX_DIMS; i++) {
-        if (a->ne[i] != b->ne[i]) {
-            return false;
-        }
-        if (a->nb[i] != b->nb[i]) {
-            return false;
-        }
-    }
-    return true;
-}
-
-void ggml_backend_tensor_copy(struct ggml_tensor * src, struct ggml_tensor * dst) {
-    //printf("src: %s ne: [%d %d %d %d] nb: [%d %d %d %d]\n", src->name, (int)src->ne[0], (int)src->ne[1], (int)src->ne[2], (int)src->ne[3], (int)src->nb[0], (int)src->nb[1], (int)src->nb[2], (int)src->nb[3]);
-    //printf("dst: %s ne: [%d %d %d %d] nb: [%d %d %d %d]\n", dst->name, (int)dst->ne[0], (int)dst->ne[1], (int)dst->ne[2], (int)dst->ne[3], (int)dst->nb[0], (int)dst->nb[1], (int)dst->nb[2], (int)dst->nb[3]);
-    GGML_ASSERT(ggml_are_same_layout(src, dst) && "cannot copy tensors with different layouts");
-
-    // Rprintf("cpy tensor %s from %s to %s (%lu bytes)\n", src->name, ggml_backend_name(src->backend), ggml_backend_name(dst->backend), ggml_nbytes(src));
-
-    if (src == dst) {
-        return;
-    }
-
-    // TODO: allow backends to support copy to/from same backend
-
-    if (dst->buffer->iface.cpy_tensor_from != NULL) {
-        dst->buffer->iface.cpy_tensor_from(dst->buffer, src, dst);
-    } else if (src->buffer->iface.cpy_tensor_to != NULL) {
-        src->buffer->iface.cpy_tensor_to(src->buffer, src, dst);
-    } else {
-        // shouldn't be hit when copying from/to CPU
-        #ifndef NDEBUG
-        Rprintf("ggml_backend_tensor_copy: neither cpy_tensor_from nor cpy_tensor_to "
-                        "are implemented for %s and %s, falling back to get/set\n", src->name, dst->name);
-        #endif
-        size_t nbytes = ggml_nbytes(src);
-        void * data = malloc(nbytes);
-        ggml_backend_tensor_get(src, data, 0, nbytes);
-        ggml_backend_tensor_set(dst, data, 0, nbytes);
-        free(data);
-    }
-}
-
-// backend registry
-
-#define GGML_MAX_BACKENDS_REG 16
-
-struct ggml_backend_reg {
-    char name[128];
-    ggml_backend_init_fn init_fn;
-    ggml_backend_buffer_type_t default_buffer_type;
-    void * user_data;
-};
-
-static struct ggml_backend_reg ggml_backend_registry[GGML_MAX_BACKENDS_REG];
-static size_t ggml_backend_registry_count = 0;
-
-static ggml_backend_t ggml_backend_reg_cpu_init(const char * params, void * user_data);
-
-static void ggml_backend_registry_init(void) {
-    static bool initialized = false;
-
-    if (initialized) {
-        return;
-    }
-
-    initialized = true;
-
-    ggml_backend_register("CPU", ggml_backend_reg_cpu_init, ggml_backend_cpu_buffer_type(), NULL);
-
-    // add forward decls here to avoid including the backend headers
-#ifdef GGML_USE_CUBLAS
-    extern void ggml_backend_cuda_reg_devices(void);
-    ggml_backend_cuda_reg_devices();
-#endif
-
-#ifdef GGML_USE_METAL
-    extern ggml_backend_t ggml_backend_reg_metal_init(const char * params, void * user_data);
-    extern ggml_backend_buffer_type_t ggml_backend_metal_buffer_type(void);
-    ggml_backend_register("Metal", ggml_backend_reg_metal_init, ggml_backend_metal_buffer_type(), NULL);
-#endif
-}
-
-void ggml_backend_register(const char * name, ggml_backend_init_fn init_fn, ggml_backend_buffer_type_t default_buffer_type, void * user_data) {
-    GGML_ASSERT(ggml_backend_registry_count < GGML_MAX_BACKENDS_REG);
-
-    size_t id = ggml_backend_registry_count;
-
-    ggml_backend_registry[id] = (struct ggml_backend_reg) {
-        /* .name                = */ {0},
-        /* .fn                  = */ init_fn,
-        /* .default_buffer_type = */ default_buffer_type,
-        /* .user_data           = */ user_data,
-    };
-
-    snprintf(ggml_backend_registry[id].name, sizeof(ggml_backend_registry[id].name), "%s", name);
-
-#ifndef NDEBUG
-    Rprintf("%s: registered backend %s\n", __func__, name);
-#endif
-
-    ggml_backend_registry_count++;
-}
-
-size_t ggml_backend_reg_get_count(void) {
-    ggml_backend_registry_init();
-
-    return ggml_backend_registry_count;
-}
-
-size_t ggml_backend_reg_find_by_name(const char * name) {
-    ggml_backend_registry_init();
-
-    for (size_t i = 0; i < ggml_backend_registry_count; i++) {
-        // TODO: case insensitive in a portable way
-        if (strcmp(ggml_backend_registry[i].name, name) == 0) {
-            return i;
-        }
-    }
-
-    // not found
-    return SIZE_MAX;
-}
-
-// init from backend:params string
-ggml_backend_t ggml_backend_reg_init_backend_from_str(const char * backend_str) {
-    ggml_backend_registry_init();
-
-    const char * params = strchr(backend_str, ':');
-    char backend_name[128];
-    if (params == NULL) {
-        snprintf(backend_name, sizeof(backend_name), "%s", backend_str);
-        params = "";
-    } else {
-        snprintf(backend_name, sizeof(backend_name), "%.*s", (int)(params - backend_str), backend_str);
-        params++;
-    }
-
-    size_t backend_i = ggml_backend_reg_find_by_name(backend_name);
-
-    if (backend_i == SIZE_MAX) {
-        Rprintf("%s: backend %s not found\n", __func__, backend_name);
-        return NULL;
-    }
-
-    return ggml_backend_reg_init_backend(backend_i, params);
-}
-
-const char * ggml_backend_reg_get_name(size_t i) {
-    ggml_backend_registry_init();
-
-    GGML_ASSERT(i < ggml_backend_registry_count);
-    return ggml_backend_registry[i].name;
-}
-
-ggml_backend_t ggml_backend_reg_init_backend(size_t i, const char * params) {
-    ggml_backend_registry_init();
-
-    GGML_ASSERT(i < ggml_backend_registry_count);
-    return ggml_backend_registry[i].init_fn(params, ggml_backend_registry[i].user_data);
-}
-
-ggml_backend_buffer_type_t ggml_backend_reg_get_default_buffer_type(size_t i) {
-    ggml_backend_registry_init();
-
-    GGML_ASSERT(i < ggml_backend_registry_count);
-    return ggml_backend_registry[i].default_buffer_type;
-}
-
-ggml_backend_buffer_t ggml_backend_reg_alloc_buffer(size_t i, size_t size) {
-    ggml_backend_registry_init();
-
-    GGML_ASSERT(i < ggml_backend_registry_count);
-    return ggml_backend_buft_alloc_buffer(ggml_backend_registry[i].default_buffer_type, size);
-}
-
-// backend CPU
-
-static void * ggml_backend_cpu_buffer_get_base(ggml_backend_buffer_t buffer) {
-    return (void *)buffer->context;
-}
-
-static void ggml_backend_cpu_buffer_free_buffer(ggml_backend_buffer_t buffer) {
-    free(buffer->context);
-}
-
-static void ggml_backend_cpu_buffer_set_tensor(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor, const void * data, size_t offset, size_t size) {
-    memcpy((char *)tensor->data + offset, data, size);
-
-    GGML_UNUSED(buffer);
-}
-
-static void ggml_backend_cpu_buffer_get_tensor(ggml_backend_buffer_t buffer, const struct ggml_tensor * tensor, void * data, size_t offset, size_t size) {
-    memcpy(data, (const char *)tensor->data + offset, size);
-
-    GGML_UNUSED(buffer);
-}
-
-static void ggml_backend_cpu_buffer_cpy_tensor_from(ggml_backend_buffer_t buffer, struct ggml_tensor * src, struct ggml_tensor * dst) {
-    ggml_backend_tensor_get(src, dst->data, 0, ggml_nbytes(src));
-
-    GGML_UNUSED(buffer);
-}
-
-static void ggml_backend_cpu_buffer_cpy_tensor_to(ggml_backend_buffer_t buffer, struct ggml_tensor * src, struct ggml_tensor * dst) {
-    ggml_backend_tensor_set(dst, src->data, 0, ggml_nbytes(src));
-
-    GGML_UNUSED(buffer);
-}
-
-static void ggml_backend_cpu_buffer_clear(ggml_backend_buffer_t buffer, uint8_t value) {
-    memset(buffer->context, value, buffer->size);
-}
-
-static struct ggml_backend_buffer_i cpu_backend_buffer_i = {
-    /* .free_buffer     = */ ggml_backend_cpu_buffer_free_buffer,
-    /* .get_base        = */ ggml_backend_cpu_buffer_get_base,
-    /* .init_tensor     = */ NULL, // no initialization required
-    /* .set_tensor      = */ ggml_backend_cpu_buffer_set_tensor,
-    /* .get_tensor      = */ ggml_backend_cpu_buffer_get_tensor,
-    /* .cpy_tensor_from = */ ggml_backend_cpu_buffer_cpy_tensor_from,
-    /* .cpy_tensor_to   = */ ggml_backend_cpu_buffer_cpy_tensor_to,
-    /* .clear           = */ ggml_backend_cpu_buffer_clear,
-};
-
-// for buffers from ptr, free is not called
-static struct ggml_backend_buffer_i cpu_backend_buffer_i_from_ptr = {
-    /* .free_buffer     = */ NULL, // ptr is not owned by the buffer, so it does not need to be freed
-    /* .get_base        = */ ggml_backend_cpu_buffer_get_base,
-    /* .init_tensor     = */ NULL, // no initialization required
-    /* .set_tensor      = */ ggml_backend_cpu_buffer_set_tensor,
-    /* .get_tensor      = */ ggml_backend_cpu_buffer_get_tensor,
-    /* .cpy_tensor_from = */ ggml_backend_cpu_buffer_cpy_tensor_from,
-    /* .cpy_tensor_to   = */ ggml_backend_cpu_buffer_cpy_tensor_to,
-    /* .clear           = */ ggml_backend_cpu_buffer_clear,
-};
-
-static const size_t TENSOR_ALIGNMENT = 64; // should be enough for AVX 512
-
-static ggml_backend_buffer_t ggml_backend_cpu_buffer_type_alloc_buffer(ggml_backend_buffer_type_t buft, size_t size) {
-    size += TENSOR_ALIGNMENT;   // malloc may return an address that is not aligned
-    void * data = malloc(size); // TODO: maybe use GGML_ALIGNED_MALLOC?
-
-    GGML_ASSERT(data != NULL && "failed to allocate buffer");
-
-    return ggml_backend_buffer_init(buft, cpu_backend_buffer_i, data, size);
-}
-
-static size_t ggml_backend_cpu_buffer_type_get_alignment(ggml_backend_buffer_type_t buft) {
-    return TENSOR_ALIGNMENT;
-
-    GGML_UNUSED(buft);
-}
-
-static bool ggml_backend_cpu_buffer_type_supports_backend(ggml_backend_buffer_type_t buft, ggml_backend_t backend) {
-    return ggml_backend_is_cpu(backend);
-
-    GGML_UNUSED(buft);
-}
-
-static bool ggml_backend_cpu_buffer_type_is_host(ggml_backend_buffer_type_t buft) {
-    return true;
-
-    GGML_UNUSED(buft);
-}
-
-ggml_backend_buffer_type_t ggml_backend_cpu_buffer_type(void) {
-    static struct ggml_backend_buffer_type ggml_backend_cpu_buffer_type = {
-        /* .iface = */ {
-            /* .alloc_buffer     = */ ggml_backend_cpu_buffer_type_alloc_buffer,
-            /* .get_alignment    = */ ggml_backend_cpu_buffer_type_get_alignment,
-            /* .get_alloc_size   = */ NULL, // defaults to ggml_nbytes
-            /* .supports_backend = */ ggml_backend_cpu_buffer_type_supports_backend,
-            /* .is_host          = */ ggml_backend_cpu_buffer_type_is_host,
-        },
-        /* .context = */ NULL,
-    };
-
-    return &ggml_backend_cpu_buffer_type;
-}
-
-#ifdef GGML_USE_CPU_HBM
-
-// buffer type HBM
-
-#include <hbwmalloc.h>
-
-static void ggml_backend_cpu_hbm_buffer_free_buffer(ggml_backend_buffer_t buffer) {
-    hbw_free(buffer->context);
-}
-
-static ggml_backend_buffer_t ggml_backend_cpu_hbm_buffer_type_alloc_buffer(ggml_backend_buffer_type_t buft, size_t size) {
-    //void * ptr = hbw_malloc(size);
-    void * ptr;
-    int result = hbw_posix_memalign(&ptr, ggml_backend_cpu_buffer_type_get_alignment(buft), size);
-    if (result != 0) {
-        Rprintf("failed to allocate HBM buffer of size %zu\n", size);
-        return NULL;
-    }
-
-    // FIXME: this is a hack to avoid having to implement a new buffer type
-    ggml_backend_buffer_t buffer = ggml_backend_cpu_buffer_from_ptr(ptr, size);
-    buffer->buft = buft;
-    buffer->iface.free_buffer = ggml_backend_cpu_hbm_buffer_free_buffer;
-
-    return buffer;
-}
-
-ggml_backend_buffer_type_t ggml_backend_cpu_hbm_buffer_type() {
-    static struct ggml_backend_buffer_type ggml_backend_cpu_buffer_type_hbm = {
-        /* .iface    = */ {
-            /* .alloc_buffer     = */ ggml_backend_cpu_hbm_buffer_type_alloc_buffer,
-            /* .get_alignment    = */ ggml_backend_cpu_buffer_type_get_alignment,
-            /* .get_alloc_size   = */ NULL, // defaults to ggml_nbytes
-            /* .supports_backend = */ ggml_backend_cpu_buffer_type_supports_backend,
-            /* .is_host          = */ ggml_backend_cpu_buffer_type_is_host,
-        },
-        /* .context  = */ NULL,
-    };
-
-    return &ggml_backend_cpu_buffer_type_hbm;
-}
-#endif
-
-struct ggml_backend_cpu_context {
-    int n_threads;
-    void * work_data;
-    size_t work_size;
-};
-
-static const char * ggml_backend_cpu_name(ggml_backend_t backend) {
-    return "CPU";
-
-    GGML_UNUSED(backend);
-}
-
-static void ggml_backend_cpu_free(ggml_backend_t backend) {
-    struct ggml_backend_cpu_context * cpu_ctx = (struct ggml_backend_cpu_context *)backend->context;
-    free(cpu_ctx->work_data);
-    free(cpu_ctx);
-    free(backend);
-}
-
-static ggml_backend_buffer_type_t ggml_backend_cpu_get_default_buffer_type(ggml_backend_t backend) {
-    return ggml_backend_cpu_buffer_type();
-
-    GGML_UNUSED(backend);
-}
-
-struct ggml_backend_plan_cpu {
-    struct ggml_cplan cplan;
-    struct ggml_cgraph cgraph;
-};
-
-static ggml_backend_graph_plan_t ggml_backend_cpu_graph_plan_create(ggml_backend_t backend, struct ggml_cgraph * cgraph) {
-    struct ggml_backend_cpu_context * cpu_ctx = (struct ggml_backend_cpu_context *)backend->context;
-
-    struct ggml_backend_plan_cpu * cpu_plan = malloc(sizeof(struct ggml_backend_plan_cpu));
-
-    cpu_plan->cplan = ggml_graph_plan(cgraph, cpu_ctx->n_threads);
-    cpu_plan->cgraph = *cgraph; // FIXME: deep copy
-
-    if (cpu_plan->cplan.work_size > 0) {
-        cpu_plan->cplan.work_data = malloc(cpu_plan->cplan.work_size);
-    }
-
-    return cpu_plan;
-}
-
-static void ggml_backend_cpu_graph_plan_free(ggml_backend_t backend, ggml_backend_graph_plan_t plan) {
-    struct ggml_backend_plan_cpu * cpu_plan = (struct ggml_backend_plan_cpu *)plan;
-
-    free(cpu_plan->cplan.work_data);
-    free(cpu_plan);
-
-    GGML_UNUSED(backend);
-}
-
-static void ggml_backend_cpu_graph_plan_compute(ggml_backend_t backend, ggml_backend_graph_plan_t plan) {
-    struct ggml_backend_plan_cpu * cpu_plan = (struct ggml_backend_plan_cpu *)plan;
-
-    ggml_graph_compute(&cpu_plan->cgraph, &cpu_plan->cplan);
-
-    GGML_UNUSED(backend);
-}
-
-static bool ggml_backend_cpu_graph_compute(ggml_backend_t backend, struct ggml_cgraph * cgraph) {
-    struct ggml_backend_cpu_context * cpu_ctx = (struct ggml_backend_cpu_context *)backend->context;
-
-    struct ggml_cplan cplan = ggml_graph_plan(cgraph, cpu_ctx->n_threads);
-
-    if (cpu_ctx->work_size < cplan.work_size) {
-        // TODO: may be faster to free and use malloc to avoid the copy
-        cpu_ctx->work_data = realloc(cpu_ctx->work_data, cplan.work_size);
-        cpu_ctx->work_size = cplan.work_size;
-    }
-
-    cplan.work_data = cpu_ctx->work_data;
-
-    ggml_graph_compute(cgraph, &cplan);
-    return true;
-}
-
-static bool ggml_backend_cpu_supports_op(ggml_backend_t backend, const struct ggml_tensor * op) {
-    switch (op->op) {
-        case GGML_OP_MUL_MAT:
-            return op->src[1]->type == GGML_TYPE_F32 || op->src[1]->type == ggml_internal_get_type_traits(op->src[0]->type).vec_dot_type;
-        default:
-            return true;
-    }
-
-    GGML_UNUSED(backend);
-}
-
-static struct ggml_backend_i cpu_backend_i = {
-    /* .get_name                = */ ggml_backend_cpu_name,
-    /* .free                    = */ ggml_backend_cpu_free,
-    /* .get_default_buffer_type = */ ggml_backend_cpu_get_default_buffer_type,
-    /* .set_tensor_async        = */ NULL,
-    /* .get_tensor_async        = */ NULL,
-    /* .cpy_tensor_from_async   = */ NULL,
-    /* .cpy_tensor_to_async     = */ NULL,
-    /* .synchronize             = */ NULL,
-    /* .graph_plan_create       = */ ggml_backend_cpu_graph_plan_create,
-    /* .graph_plan_free         = */ ggml_backend_cpu_graph_plan_free,
-    /* .graph_plan_compute      = */ ggml_backend_cpu_graph_plan_compute,
-    /* .graph_compute           = */ ggml_backend_cpu_graph_compute,
-    /* .supports_op             = */ ggml_backend_cpu_supports_op,
-};
-
-ggml_backend_t ggml_backend_cpu_init(void) {
-    struct ggml_backend_cpu_context * ctx = malloc(sizeof(struct ggml_backend_cpu_context));
-
-    ctx->n_threads = GGML_DEFAULT_N_THREADS;
-    ctx->work_data = NULL;
-    ctx->work_size = 0;
-
-    ggml_backend_t cpu_backend = malloc(sizeof(struct ggml_backend));
-
-    *cpu_backend = (struct ggml_backend) {
-        /* .interface = */ cpu_backend_i,
-        /* .context   = */ ctx
-    };
-    return cpu_backend;
-}
-
-bool ggml_backend_is_cpu(ggml_backend_t backend) {
-    return backend->iface.get_name == ggml_backend_cpu_name;
-}
-
-void ggml_backend_cpu_set_n_threads(ggml_backend_t backend_cpu, int n_threads) {
-    GGML_ASSERT(ggml_backend_is_cpu(backend_cpu));
-
-    struct ggml_backend_cpu_context * ctx = (struct ggml_backend_cpu_context *)backend_cpu->context;
-    ctx->n_threads = n_threads;
-}
-
-ggml_backend_buffer_t ggml_backend_cpu_buffer_from_ptr(void * ptr, size_t size) {
-    return ggml_backend_buffer_init(ggml_backend_cpu_buffer_type(), cpu_backend_buffer_i_from_ptr, ptr, size);
-}
-
-static ggml_backend_t ggml_backend_reg_cpu_init(const char * params, void * user_data) {
-    return ggml_backend_cpu_init();
-
-    GGML_UNUSED(params);
-    GGML_UNUSED(user_data);
-}
-
-
-// scheduler
-
-#define GGML_MAX_BACKENDS 4
-#define GGML_MAX_SPLITS 256
-#define GGML_MAX_SPLIT_INPUTS 16
-
-struct ggml_backend_sched_split {
-    ggml_tallocr_t tallocr;
-    int i_start;
-    int i_end;
-    struct ggml_tensor * inputs[GGML_MAX_SPLIT_INPUTS];
-    int n_inputs;
-    struct ggml_cgraph graph;
-};
-
-struct ggml_backend_sched {
-    int n_backends;
-    ggml_backend_t backends[GGML_MAX_BACKENDS];
-    ggml_tallocr_t  tallocs[GGML_MAX_BACKENDS];
-
-    ggml_gallocr_t galloc;
-
-    struct ggml_hash_set    hash_set;
-    ggml_tallocr_t *        node_talloc;                     // [hash_set.size]
-    struct ggml_tensor * (* node_copies)[GGML_MAX_BACKENDS]; // [hash_set.size][GGML_MAX_BACKENDS]
-
-    struct ggml_cgraph * graph;
-    struct ggml_backend_sched_split splits[GGML_MAX_SPLITS];
-    int n_splits;
-
-    struct ggml_context * ctx;
-
-    // align context_buffer to GGML_MEM_ALIGN
-    #ifdef _MSC_VER
-    __declspec(align(GGML_MEM_ALIGN))
-    #else
-    __attribute__((aligned(GGML_MEM_ALIGN)))
-    #endif
-    char context_buffer[GGML_MAX_SPLITS*GGML_MAX_SPLIT_INPUTS*sizeof(struct ggml_tensor) + sizeof(struct ggml_cgraph)];
-};
-
-#define hash_id(node) ggml_hash_find_or_insert(sched->hash_set, node)
-#define node_allocr(node) sched->node_talloc[hash_id(node)]
-
-static bool ggml_is_view_op(enum ggml_op op) {
-    return op == GGML_OP_VIEW || op == GGML_OP_RESHAPE || op == GGML_OP_PERMUTE || op == GGML_OP_TRANSPOSE;
-}
-
-// returns the priority of the backend, lower is better
-static int sched_backend_prio(ggml_backend_sched_t sched, ggml_backend_t backend) {
-    for (int i = 0; i < sched->n_backends; i++) {
-        if (sched->backends[i] == backend) {
-            return i;
-        }
-    }
-    return INT_MAX;
-}
-
-static int sched_allocr_prio(ggml_backend_sched_t sched, ggml_tallocr_t allocr) {
-    for (int i = 0; i < sched->n_backends; i++) {
-        if (sched->tallocs[i] == allocr) {
-            return i;
-        }
-    }
-    return INT_MAX;
-}
-
-static ggml_backend_t get_buffer_backend(ggml_backend_sched_t sched, ggml_backend_buffer_t buffer) {
-    if (buffer == NULL) {
-        return NULL;
-    }
-    // find highest prio backend that supports the buffer type
-    for (int i = 0; i < sched->n_backends; i++) {
-        if (ggml_backend_buft_supports_backend(buffer->buft, sched->backends[i])) {
-            return sched->backends[i];
-        }
-    }
-    GGML_ASSERT(false && "tensor buffer type not supported by any backend");
-}
-
-static ggml_backend_t get_allocr_backend(ggml_backend_sched_t sched, ggml_tallocr_t allocr) {
-    if (allocr == NULL) {
-        return NULL;
-    }
-    // find highest prio backend that supports the buffer type
-    for (int i = 0; i < sched->n_backends; i++) {
-        if (sched->tallocs[i] == allocr) {
-            return sched->backends[i];
-        }
-    }
-    GGML_UNREACHABLE();
-}
-
-#if 0
-static char causes[GGML_DEFAULT_GRAPH_SIZE*8 + GGML_MAX_SPLITS*GGML_MAX_SPLIT_INPUTS][128]; // debug, remove
-#define SET_CAUSE(node, ...) sprintf(causes[hash_id(node)], __VA_ARGS__)
-#define GET_CAUSE(node) causes[hash_id(node)]
-#else
-#define SET_CAUSE(node, ...)
-#define GET_CAUSE(node) ""
-#endif
-
-// returns the backend that should be used for the node based on the current locations
-static ggml_backend_t sched_backend_from_cur(ggml_backend_sched_t sched, struct ggml_tensor * node) {
-    // if the dst tensor is already allocated in a buffer, we must assume that it is critical to keep it there
-    // ie. kv cache updates
-    // note that this doesn't allow fallback to CPU. need to add output tensors to the splits to copy the data back to the original backend.
-    // dst
-    ggml_backend_t cur_backend = get_buffer_backend(sched, node->buffer);
-    if (cur_backend != NULL) {
-        SET_CAUSE(node, "1.dst");
-        return cur_backend;
-    }
-
-    // view_src
-    if (node->view_src != NULL && get_buffer_backend(sched, node->view_src->buffer) != NULL) {
-        SET_CAUSE(node, "1.vsrc");
-        return get_buffer_backend(sched, node->view_src->buffer);
-    }
-
-    // src
-    int cur_prio = INT_MAX;
-    size_t cur_size = 0;
-
-    for (int i = 0; i < GGML_MAX_SRC; i++) {
-        const struct ggml_tensor * src = node->src[i];
-        if (src == NULL) {
-            break;
-        }
-        ggml_backend_t src_backend = get_buffer_backend(sched, src->buffer);
-        if (src_backend != NULL) {
-            int src_prio = sched_backend_prio(sched, src_backend);
-            size_t src_size = ggml_nbytes(src);
-            if (src_prio < cur_prio && src_size >= cur_size) {
-                cur_prio = src_prio;
-                cur_size = src_size;
-                cur_backend = src_backend;
-                SET_CAUSE(node, "1.src%d", i);
-            }
-        }
-    }
-    return cur_backend;
-}
-
-static char * fmt_size(size_t size) {
-    static char buffer[128];
-    if (size >= 1024*1024) {
-        sprintf(buffer, "%zuM", size/1024/1024);
-    } else {
-        sprintf(buffer, "%zuK", size/1024);
-    }
-    return buffer;
-}
-
-static void sched_print_assignments(ggml_backend_sched_t sched, struct ggml_cgraph * graph) {
-    int cur_split = 0;
-    for (int i = 0; i < graph->n_nodes; i++) {
-        if (cur_split < sched->n_splits && i == sched->splits[cur_split].i_start) {
-            ggml_backend_t split_backend = get_allocr_backend(sched, sched->splits[cur_split].tallocr);
-            Rprintf("\n## SPLIT #%d: %s # %d inputs: ", cur_split, ggml_backend_name(split_backend),
-                sched->splits[cur_split].n_inputs);
-            for (int j = 0; j < sched->splits[cur_split].n_inputs; j++) {
-                Rprintf("[%s (%5.5s)] ", sched->splits[cur_split].inputs[j]->name,
-                    fmt_size(ggml_nbytes(sched->splits[cur_split].inputs[j])));
-            }
-            Rprintf("\n");
-            cur_split++;
-        }
-        struct ggml_tensor * node = graph->nodes[i];
-        if (ggml_is_view_op(node->op)) {
-            continue;
-        }
-        ggml_tallocr_t node_allocr = node_allocr(node);
-        ggml_backend_t node_backend = node_allocr ? get_allocr_backend(sched, node_allocr) : NULL; // FIXME:
-        Rprintf("node #%3d (%10.10s): %20.20s (%4.4s) [%4.4s %8.8s]:", i, ggml_op_name(node->op), node->name,
-            fmt_size(ggml_nbytes(node)), node_allocr ? ggml_backend_name(node_backend) : "NULL", GET_CAUSE(node));
-        for (int j = 0; j < GGML_MAX_SRC; j++) {
-            struct ggml_tensor * src = node->src[j];
-            if (src == NULL) {
-                break;
-            }
-            ggml_tallocr_t src_allocr = node_allocr(src);
-            ggml_backend_t src_backend = src_allocr ? get_allocr_backend(sched, src_allocr) : NULL;
-            Rprintf(" %20.20s (%4.4s) [%4.4s %8.8s]", src->name,
-                fmt_size(ggml_nbytes(src)), src_backend ? ggml_backend_name(src_backend) : "NULL", GET_CAUSE(src));
-        }
-        Rprintf("\n");
-    }
-}
-
-// creates a copy of the tensor with the same memory layout
-static struct ggml_tensor * ggml_dup_tensor_layout(struct ggml_context * ctx, const struct ggml_tensor * tensor) {
-    struct ggml_tensor * dup = ggml_dup_tensor(ctx, tensor);
-    for (int i = 0; i < GGML_MAX_DIMS; i++) {
-        dup->nb[i] = tensor->nb[i];
-    }
-    return dup;
-}
-
-// assigns backends to ops and splits the graph into subgraphs that can be computed on the same backend
-// TODO: merge passes
-static void sched_split_graph(ggml_backend_sched_t sched, struct ggml_cgraph * graph) {
-    // reset state
-    size_t hash_size = sched->hash_set.size;
-    memset(sched->hash_set.keys, 0, sizeof(sched->hash_set.keys[0]) * hash_size);
-    memset(sched->node_talloc,   0, sizeof(sched->node_talloc[0])   * hash_size);
-    memset(sched->node_copies,   0, sizeof(sched->node_copies[0])   * hash_size);
-    sched->n_splits = 0;
-
-    struct ggml_init_params params = {
-        /* .mem_size =   */ sizeof(sched->context_buffer),
-        /* .mem_buffer = */ sched->context_buffer,
-        /* .no_alloc =   */ true
-    };
-
-    if (sched->ctx != NULL) {
-        ggml_free(sched->ctx);
-    }
-
-    sched->ctx = ggml_init(params);
-
-    // pass 1: assign backends to ops with allocated inputs
-    for (int i = 0; i < graph->n_leafs; i++) {
-        struct ggml_tensor * leaf = graph->leafs[i];
-        if (node_allocr(leaf) != NULL) {
-            // do not overwrite user assignments
-            continue;
-        }
-        ggml_backend_t leaf_backend = get_buffer_backend(sched, leaf->buffer);
-        if (leaf_backend == NULL && leaf->view_src != NULL) {
-            leaf_backend = get_buffer_backend(sched, leaf->view_src->buffer);
-        }
-        if (leaf_backend != NULL) {
-            node_allocr(leaf) = ggml_backend_sched_get_tallocr(sched, leaf_backend);
-        }
-    }
-
-    for (int i = 0; i < graph->n_nodes; i++) {
-        struct ggml_tensor * node = graph->nodes[i];
-        if (node_allocr(node) != NULL) {
-            // do not overwrite user assignments
-            continue;
-        }
-        ggml_backend_t node_backend = sched_backend_from_cur(sched, node);
-        if (node_backend != NULL) {
-            node_allocr(node) = ggml_backend_sched_get_tallocr(sched, node_backend);
-        }
-    }
-    //printf("PASS 1 ASSIGNMENTS\n"); sched_print_assignments(sched, graph);
-
-    // pass 2: assign backends to ops from current assignments
-    // TODO:
-    //  - reuse sched_backend_from_cur
-    for (int i = 0; i < graph->n_nodes; i++) {
-        struct ggml_tensor * node = graph->nodes[i];
-        ggml_tallocr_t node_allocr = node_allocr(node);
-        if (node_allocr == NULL) {
-            int    cur_prio = INT_MAX;
-            size_t cur_size = 0;
-            for (int j = 0; j < GGML_MAX_SRC; j++) {
-                struct ggml_tensor * src = node->src[j];
-                if (src == NULL) {
-                    break;
-                }
-                ggml_tallocr_t src_allocr = node_allocr(src);
-                if (src_allocr != NULL) {
-                    int    src_prio = sched_allocr_prio(sched, src_allocr);
-                    size_t src_size = ggml_nbytes(src);
-                    if (src_prio < cur_prio && src_size >= cur_size) {
-                        cur_prio = src_prio;
-                        cur_size = src_size;
-                        node_allocr = src_allocr;
-                        SET_CAUSE(node, "2.src%d", j);
-                    }
-                }
-            }
-            if (node_allocr != NULL) {
-                node_allocr(node) = node_allocr;
-            }
-        }
-    }
-    //printf("PASS 2 ASSIGNMENTS\n"); sched_print_assignments(sched, graph);
-
-    // pass 3: assign backends to remaining src from dst (should only be leafs)
-    for (int i = 0; i < graph->n_nodes; i++) {
-        struct ggml_tensor * node = graph->nodes[i];
-        ggml_tallocr_t node_allocr = node_allocr(node);
-        for (int j = 0; j < GGML_MAX_SRC; j++) {
-            struct ggml_tensor * src = node->src[j];
-            if (src == NULL) {
-                break;
-            }
-            ggml_tallocr_t src_allocr = node_allocr(src);
-            if (src_allocr == NULL) {
-                node_allocr(src) = node_allocr;
-            }
-        }
-    }
-    //printf("PASS 3 ASSIGNMENTS\n"); sched_print_assignments(sched, graph);
-
-    // pass 4: split graph, find tensors that need to be copied
-    // TODO:
-    //  - when switching from a less preferred backend to a more preferred backend, check if it is possible to move the switch to an earlier point for the same cost
-    // find first backend
-    int cur_split = 0;
-    for (int i = 0; i < graph->n_nodes; i++) {
-        struct ggml_tensor * node = graph->nodes[i];
-        if (node->view_src == NULL) {
-            sched->splits[0].tallocr = node_allocr(node);
-            break;
-        }
-    }
-    sched->splits[0].i_start = 0;
-    sched->splits[0].n_inputs = 0;
-    memset(sched->splits[0].inputs, 0, sizeof(sched->splits[0].inputs)); //HACK
-    ggml_tallocr_t cur_allocr = sched->splits[0].tallocr;
-    size_t cur_backend_id = sched_allocr_prio(sched, cur_allocr);
-    for (int i = 0; i < graph->n_nodes; i++) {
-        struct ggml_tensor * node = graph->nodes[i];
-
-        if (ggml_is_view_op(node->op)) {
-            continue;
-        }
-
-        ggml_tallocr_t node_allocr = node_allocr(node);
-
-        if (node_allocr != cur_allocr) {
-            sched->splits[cur_split].i_end = i;
-            cur_split++;
-            GGML_ASSERT(cur_split < GGML_MAX_SPLITS);
-            sched->splits[cur_split].tallocr = node_allocr;
-            sched->splits[cur_split].i_start = i;
-            sched->splits[cur_split].n_inputs = 0;
-            memset(sched->splits[cur_split].inputs, 0, sizeof(sched->splits[cur_split].inputs)); //HACK
-            cur_allocr = node_allocr;
-            cur_backend_id = sched_allocr_prio(sched, cur_allocr);
-        }
-
-        // find inputs that are not on the same backend
-        for (int j = 0; j < GGML_MAX_SRC; j++) {
-            struct ggml_tensor * src = node->src[j];
-            if (src == NULL) {
-                break;
-            }
-            ggml_tallocr_t src_allocr = node_allocr(src);
-            if (src_allocr != node_allocr) {
-                int n_inputs = sched->splits[cur_split].n_inputs++;
-                GGML_ASSERT(n_inputs < GGML_MAX_SPLIT_INPUTS);
-                sched->splits[cur_split].inputs[n_inputs] = (struct ggml_tensor *)src;
-
-                // create copies
-                size_t id = hash_id(src);
-                if (sched->node_copies[id][cur_backend_id] == NULL) {
-                    struct ggml_tensor * tensor_copy = ggml_dup_tensor_layout(sched->ctx, src);
-                    sched->node_copies[id][cur_backend_id] = tensor_copy;
-                    node_allocr(tensor_copy) = cur_allocr;
-                    ggml_backend_t backend = get_allocr_backend(sched, cur_allocr);
-                    ggml_format_name(tensor_copy, "%s#%s", ggml_backend_name(backend), src->name);
-                }
-                node->src[j] = sched->node_copies[id][cur_backend_id];
-            }
-        }
-    }
-    sched->splits[cur_split].i_end = graph->n_nodes;
-    sched->n_splits = cur_split + 1;
-
-    //Rprintf("PASS 4 ASSIGNMENTS\n"); sched_print_assignments(sched, graph); R_CheckUserInterrupt();
-
-#if 1
-    // sanity check: all sources should have the same backend as the node
-    for (int i = 0; i < graph->n_nodes; i++) {
-        struct ggml_tensor * node = graph->nodes[i];
-        ggml_tallocr_t node_allocr = node_allocr(node);
-        if (node_allocr == NULL) {
-            Rprintf("!!!!!!! %s has no backend\n", node->name);
-        }
-        for (int j = 0; j < GGML_MAX_SRC; j++) {
-            struct ggml_tensor * src = node->src[j];
-            if (src == NULL) {
-                break;
-            }
-            ggml_tallocr_t src_allocr = node_allocr(src);
-            if (src_allocr != node_allocr /* && src_backend != NULL */) { // ignore nulls for now
-                Rprintf("!!!! %s has backend %s, src %d (%s) has backend %s\n",
-                    node->name, node_allocr ? ggml_backend_name(get_allocr_backend(sched, node_allocr)) : "NULL",
-                    j, src->name, src_allocr ? ggml_backend_name(get_allocr_backend(sched, src_allocr)) : "NULL");
-            }
-        }
-    }
-#endif
-
-    // create copies of the graph for each split
-    // FIXME: avoid this copy, pass split inputs to ggml_gallocr_alloc_graph_n in some other way
-    struct ggml_cgraph * graph_copy = ggml_new_graph_custom(sched->ctx, graph->n_nodes + sched->n_splits*GGML_MAX_SPLIT_INPUTS, false);
-    for (int i = 0; i < sched->n_splits; i++) {
-        struct ggml_backend_sched_split * split = &sched->splits[i];
-        split->graph = ggml_graph_view(graph, split->i_start, split->i_end);
-
-        // add inputs to the graph copy so that they are allocated by ggml-alloc at the start of the split
-        for (int j = 0; j < split->n_inputs; j++) {
-            struct ggml_tensor * input = split->inputs[j];
-            struct ggml_tensor * input_cpy = sched->node_copies[hash_id(input)][sched_allocr_prio(sched, split->tallocr)];
-            input_cpy->src[0] = input;
-            graph_copy->nodes[graph_copy->n_nodes++] = input_cpy;
-        }
-
-        for (int j = split->i_start; j < split->i_end; j++) {
-            graph_copy->nodes[graph_copy->n_nodes++] = graph->nodes[j];
-        }
-    }
-    sched->graph = graph_copy;
-}
-
-static void sched_alloc_splits(ggml_backend_sched_t sched) {
-    ggml_gallocr_alloc_graph_n(
-        sched->galloc,
-        sched->graph,
-        sched->hash_set,
-        sched->node_talloc);
-}
-
-static void sched_compute_splits(ggml_backend_sched_t sched) {
-    uint64_t copy_us[GGML_MAX_BACKENDS] = {0};
-    uint64_t compute_us[GGML_MAX_BACKENDS] = {0};
-
-    struct ggml_backend_sched_split * splits = sched->splits;
-
-    for (int i = 0; i < sched->n_splits; i++) {
-        struct ggml_backend_sched_split * split = &splits[i];
-        ggml_backend_t split_backend = get_allocr_backend(sched, split->tallocr);
-        int split_backend_id = sched_backend_prio(sched, split_backend);
-
-        // copy the input tensors to the split backend
-        uint64_t copy_start_us = ggml_time_us();
-        for (int j = 0; j < split->n_inputs; j++) {
-            struct ggml_tensor * input = split->inputs[j];
-            struct ggml_tensor * input_cpy = sched->node_copies[hash_id(input)][sched_backend_prio(sched, split_backend)];
-            if (input->buffer == NULL) {
-                if (input->view_src == NULL) {
-                    Rprintf("input %s has no buffer and no view_src\n", input->name);
-                    Rf_error("whispercpp error");
-                }
-                // FIXME: may need to use the sched buffer instead
-                ggml_backend_view_init(input->view_src->buffer, input);
-            }
-            if (input_cpy->buffer == NULL) {
-                Rprintf("input_cpy %s has no buffer\n", input_cpy->name);
-                Rf_error("whispercpp error");
-            }
-            //GGML_ASSERT(input->buffer->backend != input_cpy->buffer->backend);
-            //GGML_ASSERT(input_cpy->buffer->backend == split_backend);
-            ggml_backend_tensor_copy(input, input_cpy);
-        }
-        // ggml_backend_synchronize(split_backend);
-        int64_t copy_end_us = ggml_time_us();
-        copy_us[split_backend_id] += copy_end_us - copy_start_us;
-
-#if 0
-        char split_filename[GGML_MAX_NAME];
-        snprintf(split_filename, GGML_MAX_NAME, "split_%i_%s.dot", i, ggml_backend_name(split_backend));
-        ggml_graph_dump_dot(split->graph, NULL, split_filename);
-#endif
-
-        uint64_t compute_start_us = ggml_time_us();
-        ggml_backend_graph_compute(split_backend, &split->graph);
-        // ggml_backend_synchronize(split_backend);
-        uint64_t compute_end_us = ggml_time_us();
-        compute_us[split_backend_id] += compute_end_us - compute_start_us;
-    }
-
-#if 0
-    // per-backend timings
-    Rprintf("sched_compute_splits times (%d splits):\n", sched->n_splits);
-    for (int i = 0; i < sched->n_backends; i++) {
-        if (copy_us[i] > 0 || compute_us[i] > 0) {
-            Rprintf("\t%5.5s: %lu us copy, %lu us compute\n", ggml_backend_name(sched->backends[i]), copy_us[i], compute_us[i]);
-        }
-    }
-#endif
-}
-
-static void sched_reset(ggml_backend_sched_t sched) {
-    for (int i = 0; i < sched->n_backends; i++) {
-        ggml_tallocr_reset(sched->tallocs[i]);
-    }
-}
-
-ggml_backend_sched_t ggml_backend_sched_new(ggml_backend_t * backends, int n_backends) {
-    GGML_ASSERT(n_backends <= GGML_MAX_BACKENDS);
-
-    struct ggml_backend_sched * sched = malloc(sizeof(struct ggml_backend_sched));
-    memset(sched, 0, sizeof(struct ggml_backend_sched));
-
-    sched->n_backends = n_backends;
-    for (int i = 0; i < n_backends; i++) {
-        sched->backends[i] = backends[i];
-    }
-
-    sched->galloc = ggml_gallocr_new();
-
-    // init measure allocs for each backend
-    for (int i = 0; i < n_backends; i++) {
-        sched->tallocs[i] = ggml_tallocr_new_measure_from_backend(backends[i]);
-    }
-
-    return sched;
-}
-
-void ggml_backend_sched_free(ggml_backend_sched_t sched) {
-    if (sched == NULL) {
-        return;
-    }
-    for (int i = 0; i < sched->n_backends; i++) {
-        ggml_tallocr_free(sched->tallocs[i]);
-    }
-    ggml_gallocr_free(sched->galloc);
-    free(sched->hash_set.keys);
-    free(sched->node_talloc);
-    free(sched->node_copies);
-    free(sched);
-}
-
-void ggml_backend_sched_init_measure(ggml_backend_sched_t sched, struct ggml_cgraph * measure_graph) {
-    // initialize hash tables
-    size_t hash_size = measure_graph->visited_hash_table.size + GGML_MAX_SPLITS*GGML_MAX_SPLIT_INPUTS;
-    sched->hash_set.size = hash_size;
-    sched->hash_set.keys = malloc(sizeof(sched->hash_set.keys[0]) * hash_size);
-    sched->node_talloc   = malloc(sizeof(sched->node_talloc[0])   * hash_size);
-    sched->node_copies   = malloc(sizeof(sched->node_copies[0])   * hash_size);
-
-    sched_split_graph(sched, measure_graph);
-    sched_alloc_splits(sched);
-
-    // allocate buffers and reset allocators
-    for (int i = 0; i < sched->n_backends; i++) {
-        size_t size = ggml_tallocr_max_size(sched->tallocs[i]);
-        ggml_tallocr_free(sched->tallocs[i]);
-        sched->tallocs[i] = ggml_tallocr_new_from_backend(sched->backends[i], size);
-    }
-
-    sched_reset(sched);
-}
-
-void ggml_backend_sched_graph_compute(ggml_backend_sched_t sched, struct ggml_cgraph * graph) {
-    GGML_ASSERT(sched->hash_set.size >= graph->visited_hash_table.size + GGML_MAX_SPLITS*GGML_MAX_SPLIT_INPUTS);
-
-    sched_split_graph(sched, graph);
-    sched_alloc_splits(sched);
-    sched_compute_splits(sched);
-    sched_reset(sched);
-}
-
-ggml_tallocr_t ggml_backend_sched_get_tallocr(ggml_backend_sched_t sched, ggml_backend_t backend) {
-    int backend_index = sched_backend_prio(sched, backend);
-    return sched->tallocs[backend_index];
-}
-
-ggml_backend_buffer_t ggml_backend_sched_get_buffer(ggml_backend_sched_t sched, ggml_backend_t backend) {
-    int backend_index = sched_backend_prio(sched, backend);
-    return ggml_tallocr_get_buffer(sched->tallocs[backend_index]);
-}
-
-void ggml_backend_sched_set_node_backend(ggml_backend_sched_t sched, struct ggml_tensor * node, ggml_backend_t backend) {
-    int backend_index = sched_backend_prio(sched, backend);
-    GGML_ASSERT(backend_index >= 0 && backend_index < sched->n_backends);
-    node_allocr(node) = sched->tallocs[backend_index];
-}
-
-// utils
-void ggml_backend_view_init(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor) {
-    GGML_ASSERT(tensor->buffer == NULL);
-    //GGML_ASSERT(tensor->data == NULL); // views of pre-allocted tensors may have the data set, but still need to be initialized
-    GGML_ASSERT(tensor->view_src != NULL);
-    GGML_ASSERT(tensor->view_src->buffer != NULL);
-    GGML_ASSERT(tensor->view_src->data != NULL);
-
-    tensor->buffer = buffer;
-    tensor->data = (char *)tensor->view_src->data + tensor->view_offs;
-    tensor->backend = tensor->view_src->backend;
-    ggml_backend_buffer_init_tensor(buffer, tensor);
-}
-
-void ggml_backend_tensor_alloc(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor, void * addr) {
-    GGML_ASSERT(tensor->buffer == NULL);
-    GGML_ASSERT(tensor->data == NULL);
-    GGML_ASSERT(tensor->view_src == NULL);
-    GGML_ASSERT(addr >= ggml_backend_buffer_get_base(buffer));
-    GGML_ASSERT((char *)addr + ggml_backend_buffer_get_alloc_size(buffer, tensor) <=
-                (char *)ggml_backend_buffer_get_base(buffer) + ggml_backend_buffer_get_size(buffer));
-
-    tensor->buffer = buffer;
-    tensor->data = addr;
-    ggml_backend_buffer_init_tensor(buffer, tensor);
-}
-
-static struct ggml_tensor * graph_dup_tensor(struct ggml_hash_set hash_set, struct ggml_tensor ** node_copies,
-    struct ggml_context * ctx_allocated, struct ggml_context * ctx_unallocated, struct ggml_tensor * src) {
-
-    GGML_ASSERT(src != NULL);
-    GGML_ASSERT(src->data && "graph must be allocated");
-
-    size_t id = ggml_hash_insert(hash_set, src);
-    if (id == GGML_HASHTABLE_ALREADY_EXISTS) {
-        return node_copies[ggml_hash_find(hash_set, src)];
-    }
-
-    struct ggml_tensor * dst = ggml_dup_tensor_layout(src->data && !src->view_src ? ctx_allocated : ctx_unallocated, src);
-    if (src->view_src != NULL) {
-        dst->view_src = graph_dup_tensor(hash_set, node_copies, ctx_allocated, ctx_unallocated, src->view_src);
-        dst->view_offs = src->view_offs;
-    }
-    dst->op = src->op;
-    memcpy(dst->op_params, src->op_params, sizeof(dst->op_params));
-    ggml_set_name(dst, src->name);
-
-    // copy src
-    for (int i = 0; i < GGML_MAX_SRC; i++) {
-        struct ggml_tensor * s = src->src[i];
-        if (s == NULL) {
-            break;
-        }
-        dst->src[i] = graph_dup_tensor(hash_set, node_copies, ctx_allocated, ctx_unallocated, s);
-    }
-
-    node_copies[id] = dst;
-    return dst;
-}
-
-static void graph_init_tensor(struct ggml_hash_set hash_set, struct ggml_tensor ** node_copies, bool * node_init, struct ggml_tensor * src) {
-    size_t id = ggml_hash_find(hash_set, src);
-    if (node_init[id]) {
-        return;
-    }
-    node_init[id] = true;
-
-    struct ggml_tensor * dst = node_copies[id];
-    if (dst->view_src != NULL) {
-        ggml_backend_view_init(dst->view_src->buffer, dst);
-    }
-    else {
-        ggml_backend_tensor_copy(src, dst);
-    }
-
-    // init src
-    for (int i = 0; i < GGML_MAX_SRC; i++) {
-        struct ggml_tensor * s = src->src[i];
-        if (s == NULL) {
-            break;
-        }
-        graph_init_tensor(hash_set, node_copies, node_init, s);
-    }
-}
-
-struct ggml_backend_graph_copy ggml_backend_graph_copy(ggml_backend_t backend, struct ggml_cgraph * graph) {
-    struct ggml_hash_set hash_set = {
-        /* .size = */ graph->visited_hash_table.size,
-        /* .keys = */ calloc(sizeof(hash_set.keys[0]) * graph->visited_hash_table.size, 1)
-    };
-    struct ggml_tensor ** node_copies = calloc(sizeof(node_copies[0]) * hash_set.size, 1);
-    bool * node_init = calloc(sizeof(node_init[0]) * hash_set.size, 1);
-
-    struct ggml_init_params params = {
-        /* .mem_size   = */ ggml_tensor_overhead()*hash_set.size + ggml_graph_overhead_custom(graph->size, false),
-        /* .mem_buffer = */ NULL,
-        /* .no_alloc   = */ true
-    };
-
-    struct ggml_context * ctx_allocated = ggml_init(params);
-    struct ggml_context * ctx_unallocated = ggml_init(params);
-
-    // dup nodes
-    for (int i = 0; i < graph->n_nodes; i++) {
-        struct ggml_tensor * node = graph->nodes[i];
-        graph_dup_tensor(hash_set, node_copies, ctx_allocated, ctx_unallocated, node);
-    }
-
-    // allocate nodes
-    ggml_backend_buffer_t buffer = ggml_backend_alloc_ctx_tensors(ctx_allocated, backend);
-
-    //printf("copy buffer size: %zu MB\n", ggml_backend_buffer_get_size(buffer) / 1024 / 1024);
-
-    // copy data and init views
-    for (int i = 0; i < graph->n_nodes; i++) {
-        struct ggml_tensor * node = graph->nodes[i];
-        graph_init_tensor(hash_set, node_copies, node_init, node);
-    }
-
-    // build graph copy
-    struct ggml_cgraph * graph_copy = ggml_new_graph_custom(ctx_allocated, graph->size, false);
-    for (int i = 0; i < graph->n_nodes; i++) {
-        struct ggml_tensor * node = graph->nodes[i];
-        struct ggml_tensor * node_copy = node_copies[ggml_hash_find(hash_set, node)];
-        graph_copy->nodes[i] = node_copy;
-    }
-    graph_copy->n_nodes = graph->n_nodes;
-
-    free(hash_set.keys);
-    free(node_copies);
-    free(node_init);
-
-    return (struct ggml_backend_graph_copy) {
-        /* .buffer           = */ buffer,
-        /* .ctx_allocated    = */ ctx_allocated,
-        /* .ctx_unallocated  = */ ctx_unallocated,
-        /* .graph            = */ graph_copy,
-    };
-}
-
-void ggml_backend_graph_copy_free(struct ggml_backend_graph_copy copy) {
-    ggml_backend_buffer_free(copy.buffer);
-    ggml_free(copy.ctx_allocated);
-    ggml_free(copy.ctx_unallocated);
-}
-
-void ggml_backend_compare_graph_backend(ggml_backend_t backend1, ggml_backend_t backend2, struct ggml_cgraph * graph, ggml_backend_eval_callback callback, void * user_data) {
-    struct ggml_backend_graph_copy copy = ggml_backend_graph_copy(backend2, graph);
-    struct ggml_cgraph * g1 = graph;
-    struct ggml_cgraph * g2 = copy.graph;
-
-    assert(g1->n_nodes == g2->n_nodes);
-
-    for (int i = 0; i < g1->n_nodes; i++) {
-        //printf("eval %d/%d\n", i, g1->n_nodes);
-        struct ggml_tensor * t1 = g1->nodes[i];
-        struct ggml_tensor * t2 = g2->nodes[i];
-
-        assert(t1->op == t2->op && ggml_are_same_layout(t1, t2));
-
-        struct ggml_cgraph g1v = ggml_graph_view(g1, i, i + 1);
-        struct ggml_cgraph g2v = ggml_graph_view(g2, i, i + 1);
-
-        ggml_backend_graph_compute(backend1, &g1v);
-        ggml_backend_graph_compute(backend2, &g2v);
-
-        if (ggml_is_view_op(t1->op)) {
-            continue;
-        }
-
-        // compare results, calculate rms etc
-        if (!callback(i, t1, t2, user_data)) {
-            break;
-        }
-    }
-
-    ggml_backend_graph_copy_free(copy);
-}
diff --git a/src/whisper_cpp/ggml-backend.h b/src/whisper_cpp/ggml-backend.h
deleted file mode 100644
index 85ff67b0..00000000
--- a/src/whisper_cpp/ggml-backend.h
+++ /dev/null
@@ -1,188 +0,0 @@
-#pragma once
-
-#include "ggml.h"
-#include "ggml-alloc.h"
-
-#ifdef  __cplusplus
-extern "C" {
-#endif
-
-    typedef struct ggml_backend_buffer_type * ggml_backend_buffer_type_t;
-    typedef struct ggml_backend_buffer * ggml_backend_buffer_t;
-    typedef struct ggml_backend * ggml_backend_t;
-    typedef void * ggml_backend_graph_plan_t;
-
-    //
-    // Backend buffer
-    //
-
-    // buffer type
-    GGML_API ggml_backend_buffer_t ggml_backend_buft_alloc_buffer(ggml_backend_buffer_type_t buft, size_t size);
-    GGML_API size_t ggml_backend_buft_get_alignment (ggml_backend_buffer_type_t buft);
-    GGML_API size_t ggml_backend_buft_get_alloc_size(ggml_backend_buffer_type_t buft, struct ggml_tensor * tensor);
-    GGML_API bool ggml_backend_buft_supports_backend(ggml_backend_buffer_type_t buft, ggml_backend_t backend);
-    GGML_API bool ggml_backend_buft_is_host         (ggml_backend_buffer_type_t buft);
-
-    // buffer
-    GGML_API void   ggml_backend_buffer_free          (ggml_backend_buffer_t buffer);
-    GGML_API void * ggml_backend_buffer_get_base      (ggml_backend_buffer_t buffer);
-    GGML_API size_t ggml_backend_buffer_get_size      (ggml_backend_buffer_t buffer);
-    GGML_API void   ggml_backend_buffer_init_tensor   (ggml_backend_buffer_t buffer, struct ggml_tensor * tensor);
-    GGML_API size_t ggml_backend_buffer_get_alignment (ggml_backend_buffer_t buffer);
-    GGML_API size_t ggml_backend_buffer_get_alloc_size(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor);
-    GGML_API void   ggml_backend_buffer_clear         (ggml_backend_buffer_t buffer, uint8_t value);
-    GGML_API bool   ggml_backend_buffer_is_host       (ggml_backend_buffer_t buffer);
-    GGML_API ggml_backend_buffer_type_t ggml_backend_buffer_type(ggml_backend_buffer_t buffer);
-
-    //
-    // Backend
-    //
-
-
-    GGML_API const char * ggml_backend_name(ggml_backend_t backend);
-    GGML_API void         ggml_backend_free(ggml_backend_t backend);
-
-    GGML_API ggml_backend_buffer_type_t ggml_backend_get_default_buffer_type(ggml_backend_t backend);
-    GGML_API ggml_backend_buffer_t      ggml_backend_alloc_buffer(ggml_backend_t backend, size_t size);
-    GGML_API size_t                     ggml_backend_get_alignment(ggml_backend_t backend);
-
-    GGML_API void ggml_backend_tensor_set_async(ggml_backend_t backend,       struct ggml_tensor * tensor, const void * data, size_t offset, size_t size);
-    GGML_API void ggml_backend_tensor_get_async(ggml_backend_t backend, const struct ggml_tensor * tensor,       void * data, size_t offset, size_t size);
-
-    GGML_API void ggml_backend_tensor_set(      struct ggml_tensor * tensor, const void * data, size_t offset, size_t size);
-    GGML_API void ggml_backend_tensor_get(const struct ggml_tensor * tensor,       void * data, size_t offset, size_t size);
-
-    GGML_API void ggml_backend_synchronize(ggml_backend_t backend);
-
-    GGML_API ggml_backend_graph_plan_t ggml_backend_graph_plan_create (ggml_backend_t backend, struct ggml_cgraph * cgraph);
-
-    GGML_API void ggml_backend_graph_plan_free   (ggml_backend_t backend, ggml_backend_graph_plan_t plan);
-    GGML_API void ggml_backend_graph_plan_compute(ggml_backend_t backend, ggml_backend_graph_plan_t plan);
-    GGML_API bool ggml_backend_graph_compute     (ggml_backend_t backend, struct ggml_cgraph * cgraph);
-    GGML_API bool ggml_backend_supports_op       (ggml_backend_t backend, const struct ggml_tensor * op);
-
-    // tensor copy between different backends
-    GGML_API void ggml_backend_tensor_copy(struct ggml_tensor * src, struct ggml_tensor * dst);
-    GGML_API void ggml_backend_tensor_copy_async(ggml_backend_t backend, struct ggml_tensor * src, struct ggml_tensor * dst); // automatic fallback to sync copy
-
-    //
-    // CPU backend
-    //
-
-    GGML_API ggml_backend_t ggml_backend_cpu_init(void);
-
-    GGML_API bool ggml_backend_is_cpu(ggml_backend_t backend);
-    GGML_API void ggml_backend_cpu_set_n_threads(ggml_backend_t backend_cpu, int n_threads);
-
-    // Create a backend buffer from an existing pointer
-    GGML_API ggml_backend_buffer_t ggml_backend_cpu_buffer_from_ptr(void * ptr, size_t size);
-
-    GGML_API ggml_backend_buffer_type_t ggml_backend_cpu_buffer_type(void);
-
-#ifdef GGML_USE_CPU_HBM
-    GGML_API ggml_backend_buffer_type_t ggml_backend_cpu_hbm_buffer_type(void);
-#endif
-
-    //
-    // Backend registry
-    //
-
-    // The backend registry is a registry of all the available backends, and allows initializing backends in a generic way
-
-    GGML_API size_t                     ggml_backend_reg_get_count(void);
-    GGML_API size_t                     ggml_backend_reg_find_by_name(const char * name);
-    GGML_API ggml_backend_t             ggml_backend_reg_init_backend_from_str(const char * backend_str); // str is name[:params]
-    GGML_API const char *               ggml_backend_reg_get_name(size_t i);
-    GGML_API ggml_backend_t             ggml_backend_reg_init_backend(size_t i, const char * params); // params is backend-specific
-    GGML_API ggml_backend_buffer_type_t ggml_backend_reg_get_default_buffer_type(size_t i);
-    GGML_API ggml_backend_buffer_t      ggml_backend_reg_alloc_buffer(size_t i, size_t size);
-
-    //
-    // Backend scheduler
-    //
-
-    // The backend scheduler allows for multiple backends to be used together
-    // Handles compute buffer allocation, assignment of tensors to backends, and copying of tensors between backends
-    // The backends are selected based on:
-    // - the backend that supports the operation
-    // - the location of the pre-allocated tensors (e.g. the weights)
-    /*
-      Example usage:
-
-        sched = ggml_backend_sched_new({backend_gpu, backend_gpu2, backend_cpu}, num_backends);
-        // sched is initialized with measure allocators and cannot be used until allocated with a measure graph
-
-        // initialize buffers from a measure graph
-        measure_graph = build_graph(sched); // use the allocr to allocate inputs as needed
-
-        // in build_graph:
-        build_graph(...) {
-            // allocating tensors in a specific backend (optional, recommended: pre-allocate inputs in a different buffer)
-            alloc_cpu = ggml_backend_sched_get_allocr(sched, backend_cpu);
-            ggml_allocr_alloc(alloc_cpu, tensor);
-
-            // manually assigning nodes to a backend (optional, shouldn't be needed in most cases)
-            struct ggml_tensor * node = ggml_mul_mat(ctx, ...);
-            ggml_backend_sched_set_node_backend(sched, node, backend_gpu);
-        }
-
-        // allocate backend buffers from measure graph
-        ggml_backend_sched_init_measure(sched, measure_graph);
-
-        // the scheduler is now ready to compute graphs
-
-        // compute
-        graph = build_graph(sched);
-        ggml_backend_sched_graph_compute(sched, graph);
-    */
-
-    struct ggml_backend_sched;
-    typedef struct ggml_backend_sched * ggml_backend_sched_t;
-
-    // Initialize a backend scheduler
-    GGML_API ggml_backend_sched_t ggml_backend_sched_new(ggml_backend_t * backends, int n_backends);
-
-    GGML_API void ggml_backend_sched_free(ggml_backend_sched_t sched);
-
-    // Initialize backend buffers from a measure graph
-    GGML_API void ggml_backend_sched_init_measure(ggml_backend_sched_t sched, struct ggml_cgraph * measure_graph);
-
-    GGML_API ggml_tallocr_t        ggml_backend_sched_get_tallocr(ggml_backend_sched_t sched, ggml_backend_t backend);
-    GGML_API ggml_backend_buffer_t ggml_backend_sched_get_buffer (ggml_backend_sched_t sched, ggml_backend_t backend);
-
-    GGML_API void ggml_backend_sched_set_node_backend(ggml_backend_sched_t sched, struct ggml_tensor * node, ggml_backend_t backend);
-
-    // Allocate a graph on the backend scheduler
-    GGML_API void ggml_backend_sched_graph_compute(
-            ggml_backend_sched_t sched,
-            struct ggml_cgraph * graph);
-
-
-    //
-    // Utils
-    //
-
-    struct ggml_backend_graph_copy {
-        ggml_backend_buffer_t buffer;
-        struct ggml_context * ctx_allocated;
-        struct ggml_context * ctx_unallocated;
-        struct ggml_cgraph * graph;
-    };
-
-    // Copy a graph to a different backend
-    GGML_API struct ggml_backend_graph_copy ggml_backend_graph_copy(ggml_backend_t backend, struct ggml_cgraph * graph);
-    GGML_API void                           ggml_backend_graph_copy_free(struct ggml_backend_graph_copy copy);
-
-    typedef bool (*ggml_backend_eval_callback)(int node_index, struct ggml_tensor * t1, struct ggml_tensor * t2, void * user_data);
-
-    // Compare the output of two backends
-    GGML_API void ggml_backend_compare_graph_backend(ggml_backend_t backend1, ggml_backend_t backend2, struct ggml_cgraph * graph, ggml_backend_eval_callback callback, void * user_data);
-
-    // Tensor initialization
-    GGML_API void ggml_backend_tensor_alloc(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor, void * addr);
-    GGML_API void ggml_backend_view_init(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor);
-
-
-#ifdef  __cplusplus
-}
-#endif
diff --git a/src/whisper_cpp/ggml-cuda.cu b/src/whisper_cpp/ggml-cuda.cu
deleted file mode 100644
index 7578d21c..00000000
--- a/src/whisper_cpp/ggml-cuda.cu
+++ /dev/null
@@ -1,10146 +0,0 @@
-#include <algorithm>
-#include <assert.h>
-#include <atomic>
-#include <cinttypes>
-#include <cstddef>
-#include <cstdint>
-#include <float.h>
-#include <limits>
-#include <stdint.h>
-#include <stdio.h>
-#include <vector>
-
-
-#if defined(GGML_USE_HIPBLAS)
-#include <hip/hip_runtime.h>
-#include <hipblas/hipblas.h>
-#include <hip/hip_fp16.h>
-#ifdef __HIP_PLATFORM_AMD__
-// for rocblas_initialize()
-#include "rocblas/rocblas.h"
-#endif // __HIP_PLATFORM_AMD__
-#define CUBLAS_COMPUTE_16F HIPBLAS_R_16F
-#define CUBLAS_COMPUTE_32F HIPBLAS_R_32F
-#define CUBLAS_COMPUTE_32F_FAST_16F HIPBLAS_R_32F
-#define CUBLAS_GEMM_DEFAULT HIPBLAS_GEMM_DEFAULT
-#define CUBLAS_GEMM_DEFAULT_TENSOR_OP HIPBLAS_GEMM_DEFAULT
-#define CUBLAS_OP_N HIPBLAS_OP_N
-#define CUBLAS_OP_T HIPBLAS_OP_T
-#define CUBLAS_STATUS_SUCCESS HIPBLAS_STATUS_SUCCESS
-#define CUBLAS_TF32_TENSOR_OP_MATH 0
-#define CUDA_R_16F  HIPBLAS_R_16F
-#define CUDA_R_32F  HIPBLAS_R_32F
-#define __shfl_xor_sync(mask, var, laneMask, width) __shfl_xor(var, laneMask, width)
-#define cublasComputeType_t hipblasDatatype_t //deprecated, new hipblasComputeType_t not in 5.6
-#define cublasCreate hipblasCreate
-#define cublasGemmEx hipblasGemmEx
-#define cublasGemmBatchedEx hipblasGemmBatchedEx
-#define cublasGemmStridedBatchedEx hipblasGemmStridedBatchedEx
-#define cublasHandle_t hipblasHandle_t
-#define cublasSetMathMode(handle, mode) CUBLAS_STATUS_SUCCESS
-#define cublasSetStream hipblasSetStream
-#define cublasSgemm hipblasSgemm
-#define cublasStatus_t hipblasStatus_t
-#define cudaDataType_t hipblasDatatype_t //deprecated, new hipblasDatatype not in 5.6
-#define cudaDeviceCanAccessPeer hipDeviceCanAccessPeer
-#define cudaDeviceDisablePeerAccess hipDeviceDisablePeerAccess
-#define cudaDeviceEnablePeerAccess hipDeviceEnablePeerAccess
-#define cudaDeviceProp hipDeviceProp_t
-#define cudaDeviceSynchronize hipDeviceSynchronize
-#define cudaError_t hipError_t
-#define cudaEventCreateWithFlags hipEventCreateWithFlags
-#define cudaEventDisableTiming hipEventDisableTiming
-#define cudaEventRecord hipEventRecord
-#define cudaEvent_t hipEvent_t
-#define cudaEventDestroy hipEventDestroy
-#define cudaFree hipFree
-#define cudaFreeHost hipHostFree
-#define cudaGetDevice hipGetDevice
-#define cudaGetDeviceCount hipGetDeviceCount
-#define cudaGetDeviceProperties hipGetDeviceProperties
-#define cudaGetErrorString hipGetErrorString
-#define cudaGetLastError hipGetLastError
-#ifdef GGML_HIP_UMA
-#define cudaMalloc hipMallocManaged
-#define cudaMallocHost(ptr, size) hipHostMalloc(ptr, size)
-#else
-#define cudaMalloc hipMalloc
-#define cudaMallocHost(ptr, size) hipHostMalloc(ptr, size, hipHostMallocDefault)
-#endif
-#define cudaMemcpy hipMemcpy
-#define cudaMemcpyAsync hipMemcpyAsync
-#define cudaMemcpyPeerAsync hipMemcpyPeerAsync
-#define cudaMemcpy2DAsync hipMemcpy2DAsync
-#define cudaMemcpyDeviceToDevice hipMemcpyDeviceToDevice
-#define cudaMemcpyDeviceToHost hipMemcpyDeviceToHost
-#define cudaMemcpyHostToDevice hipMemcpyHostToDevice
-#define cudaMemcpyKind hipMemcpyKind
-#define cudaMemset hipMemset
-#define cudaMemsetAsync hipMemsetAsync
-#define cudaOccupancyMaxPotentialBlockSize hipOccupancyMaxPotentialBlockSize
-#define cudaSetDevice hipSetDevice
-#define cudaStreamCreateWithFlags hipStreamCreateWithFlags
-#define cudaStreamFireAndForget hipStreamFireAndForget
-#define cudaStreamNonBlocking hipStreamNonBlocking
-#define cudaStreamSynchronize hipStreamSynchronize
-#define cudaStreamWaitEvent(stream, event, flags) hipStreamWaitEvent(stream, event, flags)
-#define cudaStream_t hipStream_t
-#define cudaSuccess hipSuccess
-#define __trap abort
-#define CUBLAS_STATUS_SUCCESS HIPBLAS_STATUS_SUCCESS
-#define CUBLAS_STATUS_NOT_INITIALIZED HIPBLAS_STATUS_NOT_INITIALIZED
-#define CUBLAS_STATUS_ALLOC_FAILED HIPBLAS_STATUS_ALLOC_FAILED
-#define CUBLAS_STATUS_INVALID_VALUE HIPBLAS_STATUS_INVALID_VALUE
-#define CUBLAS_STATUS_ARCH_MISMATCH HIPBLAS_STATUS_ARCH_MISMATCH
-#define CUBLAS_STATUS_MAPPING_ERROR HIPBLAS_STATUS_MAPPING_ERROR
-#define CUBLAS_STATUS_EXECUTION_FAILED HIPBLAS_STATUS_EXECUTION_FAILED
-#define CUBLAS_STATUS_INTERNAL_ERROR HIPBLAS_STATUS_INTERNAL_ERROR
-#define CUBLAS_STATUS_NOT_SUPPORTED HIPBLAS_STATUS_NOT_SUPPORTED
-#else
-#include <cuda_runtime.h>
-#include <cuda.h>
-#include <cublas_v2.h>
-#include <cuda_fp16.h>
-
-#if CUDART_VERSION < 11020
-#define CU_DEVICE_ATTRIBUTE_VIRTUAL_MEMORY_MANAGEMENT_SUPPORTED CU_DEVICE_ATTRIBUTE_VIRTUAL_ADDRESS_MANAGEMENT_SUPPORTED
-#define CUBLAS_TF32_TENSOR_OP_MATH CUBLAS_TENSOR_OP_MATH
-#define CUBLAS_COMPUTE_16F CUDA_R_16F
-#define CUBLAS_COMPUTE_32F CUDA_R_32F
-#define cublasComputeType_t cudaDataType_t
-#endif // CUDART_VERSION < 11020
-
-#endif // defined(GGML_USE_HIPBLAS)
-
-#include "ggml-cuda.h"
-#include "ggml.h"
-#include "ggml-backend-impl.h"
-
-#define MIN_CC_DP4A   610 // minimum compute capability for __dp4a, an intrinsic for byte-wise dot products
-#define CC_VOLTA      700
-#define CC_OFFSET_AMD 1000000
-#define CC_RDNA1      (CC_OFFSET_AMD + 1010)
-#define CC_RDNA2      (CC_OFFSET_AMD + 1030)
-#define CC_RDNA3      (CC_OFFSET_AMD + 1100)
-
-#define GGML_CUDA_MAX_NODES 8192
-
-// define this if you want to always fallback to MMQ kernels and not use cuBLAS for matrix multiplication
-// on modern hardware, using cuBLAS is recommended as it utilizes F16 tensor cores which are very performant
-// for large computational tasks. the drawback is that this requires some extra amount of VRAM:
-// -  7B quantum model: +100-200 MB
-// - 13B quantum model: +200-400 MB
-//
-//#define GGML_CUDA_FORCE_MMQ
-
-// TODO: improve this to be correct for more hardware
-//       for example, currently fails for GeForce GTX 1660 which is TURING arch (> VOLTA) but does not have tensor cores
-#if !defined(GGML_CUDA_FORCE_MMQ)
-#define CUDA_USE_TENSOR_CORES
-#endif
-
-// max batch size to use MMQ kernels when tensor cores are available
-#define MMQ_MAX_BATCH_SIZE 32
-
-#if defined(GGML_USE_HIPBLAS)
-#define __CUDA_ARCH__ 1300
-
-#if defined(__gfx1100__) || defined(__gfx1101__) || defined(__gfx1102__) || defined(__gfx1103__) || \
-    defined(__gfx1150__) || defined(__gfx1151__)
-#define RDNA3
-#endif
-
-#if defined(__gfx1030__) || defined(__gfx1031__) || defined(__gfx1032__) || defined(__gfx1033__) || \
-    defined(__gfx1034__) || defined(__gfx1035__) || defined(__gfx1036__) || defined(__gfx1037__)
-#define RDNA2
-#endif
-
-#ifndef __has_builtin
-    #define __has_builtin(x) 0
-#endif
-
-typedef int8_t int8x4_t __attribute__((ext_vector_type(4)));
-static __device__ __forceinline__ int __vsubss4(const int a, const int b) {
-    const int8x4_t va = reinterpret_cast<const int8x4_t&>(a);
-    const int8x4_t vb = reinterpret_cast<const int8x4_t&>(b);
-#if __has_builtin(__builtin_elementwise_sub_sat)
-    const int8x4_t c = __builtin_elementwise_sub_sat(va, vb);
-    return reinterpret_cast<const int &>(c);
-#else
-    int8x4_t c;
-    int16_t tmp;
-#pragma unroll
-    for (int i = 0; i < 4; i++) {
-        tmp = va[i] - vb[i];
-        if(tmp > std::numeric_limits<int8_t>::max()) tmp = std::numeric_limits<int8_t>::max();
-        if(tmp < std::numeric_limits<int8_t>::min()) tmp = std::numeric_limits<int8_t>::min();
-        c[i] = tmp;
-    }
-    return reinterpret_cast<int &>(c);
-#endif // __has_builtin(__builtin_elementwise_sub_sat)
-}
-
-static __device__ __forceinline__ int __dp4a(const int a, const int b, int c) {
-#if defined(__gfx906__) || defined(__gfx908__) || defined(__gfx90a__) || defined(__gfx1030__)
-    c = __builtin_amdgcn_sdot4(a, b, c, false);
-#elif defined(__gfx1100__)
-    c = __builtin_amdgcn_sudot4( true, a, true, b, c, false);
-#elif defined(__gfx1010__) || defined(__gfx900__)
-    int tmp1;
-    int tmp2;
-    asm("\n \
-        v_mul_i32_i24 %1, sext(%3), sext(%4) dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:BYTE_0 \n \
-        v_mul_i32_i24 %2, sext(%3), sext(%4) dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_1 src1_sel:BYTE_1 \n \
-        v_add3_u32 %0, %1, %2, %0 \n \
-        v_mul_i32_i24 %1, sext(%3), sext(%4) dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_2 src1_sel:BYTE_2 \n \
-        v_mul_i32_i24 %2, sext(%3), sext(%4) dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_3 src1_sel:BYTE_3 \n \
-        v_add3_u32 %0, %1, %2, %0 \n \
-        "
-        : "+v"(c), "=&v"(tmp1), "=&v"(tmp2)
-        : "v"(a), "v"(b)
-    );
-#else
-    const int8x4_t va = reinterpret_cast<const int8x4_t&>(a);
-    const int8x4_t vb = reinterpret_cast<const int8x4_t&>(b);
-    c += va[0] * vb[0] + va[1] * vb[1] + va[2] * vb[2] + va[3] * vb[3];
-#endif
-    return c;
-}
-#endif // defined(GGML_USE_HIPBLAS)
-
-#if defined(_MSC_VER)
-#pragma warning(disable: 4244 4267) // possible loss of data
-#endif
-
-static_assert(sizeof(half) == sizeof(ggml_fp16_t), "wrong fp16 size");
-
-[[noreturn]]
-static void ggml_cuda_error(const char * stmt, const char * func, const char * file, const int line, const char * msg) {
-    int id = -1; // in case cudaGetDevice fails
-    cudaGetDevice(&id);
-
-    fprintf(stderr, "CUDA error: %s\n", msg);
-    fprintf(stderr, "  current device: %d, in function %s at %s:%d\n", id, func, file, line);
-    fprintf(stderr, "  %s\n", stmt);
-    // abort with GGML_ASSERT to get a stack trace
-    GGML_ASSERT(!"CUDA error");
-}
-
-#define CUDA_CHECK_GEN(err, success, error_fn)                                      \
-     do {                                                                           \
-        auto err_ = (err);                                                          \
-        if (err_ != (success)) {                                                    \
-            ggml_cuda_error(#err, __func__, __FILE__, __LINE__, error_fn(err_));    \
-        }                                                                           \
-    } while (0)
-
-#define CUDA_CHECK(err) CUDA_CHECK_GEN(err, cudaSuccess, cudaGetErrorString)
-
-#if CUDART_VERSION >= 12000
-    static const char * cublas_get_error_str(const cublasStatus_t err) {
-        return cublasGetStatusString(err);
-    }
-#else
-    static const char * cublas_get_error_str(const cublasStatus_t err) {
-        switch (err) {
-            case CUBLAS_STATUS_SUCCESS: return "CUBLAS_STATUS_SUCCESS";
-            case CUBLAS_STATUS_NOT_INITIALIZED: return "CUBLAS_STATUS_NOT_INITIALIZED";
-            case CUBLAS_STATUS_ALLOC_FAILED: return "CUBLAS_STATUS_ALLOC_FAILED";
-            case CUBLAS_STATUS_INVALID_VALUE: return "CUBLAS_STATUS_INVALID_VALUE";
-            case CUBLAS_STATUS_ARCH_MISMATCH: return "CUBLAS_STATUS_ARCH_MISMATCH";
-            case CUBLAS_STATUS_MAPPING_ERROR: return "CUBLAS_STATUS_MAPPING_ERROR";
-            case CUBLAS_STATUS_EXECUTION_FAILED: return "CUBLAS_STATUS_EXECUTION_FAILED";
-            case CUBLAS_STATUS_INTERNAL_ERROR: return "CUBLAS_STATUS_INTERNAL_ERROR";
-            case CUBLAS_STATUS_NOT_SUPPORTED: return "CUBLAS_STATUS_NOT_SUPPORTED";
-            default: return "unknown error";
-        }
-    }
-#endif // CUDART_VERSION >= 12000
-
-#define CUBLAS_CHECK(err) CUDA_CHECK_GEN(err, CUBLAS_STATUS_SUCCESS, cublas_get_error_str)
-
-#if !defined(GGML_USE_HIPBLAS)
-static const char * cu_get_error_str(CUresult err) {
-    const char * err_str;
-    cuGetErrorString(err, &err_str);
-    return err_str;
-}
-#define CU_CHECK(err) CUDA_CHECK_GEN(err, CUDA_SUCCESS, cu_get_error_str)
-#endif
-
-#if CUDART_VERSION >= 11100
-#define GGML_CUDA_ASSUME(x) __builtin_assume(x)
-#else
-#define GGML_CUDA_ASSUME(x)
-#endif // CUDART_VERSION >= 11100
-
-#ifdef GGML_CUDA_F16
-typedef half dfloat; // dequantize float
-typedef half2 dfloat2;
-#else
-typedef float dfloat; // dequantize float
-typedef float2 dfloat2;
-#endif //GGML_CUDA_F16
-
-static __device__ __forceinline__ int get_int_from_int8(const int8_t * x8, const int & i32) {
-    const uint16_t * x16 = (const uint16_t *) (x8 + sizeof(int) * i32); // assume at least 2 byte alignment
-
-    int x32 = 0;
-    x32 |= x16[0] <<  0;
-    x32 |= x16[1] << 16;
-
-    return x32;
-}
-
-static __device__ __forceinline__ int get_int_from_uint8(const uint8_t * x8, const int & i32) {
-    const uint16_t * x16 = (const uint16_t *) (x8 + sizeof(int) * i32); // assume at least 2 byte alignment
-
-    int x32 = 0;
-    x32 |= x16[0] <<  0;
-    x32 |= x16[1] << 16;
-
-    return x32;
-}
-
-static __device__ __forceinline__ int get_int_from_int8_aligned(const int8_t * x8, const int & i32) {
-    return *((const int *) (x8 + sizeof(int) * i32)); // assume at least 4 byte alignment
-}
-
-static __device__ __forceinline__ int get_int_from_uint8_aligned(const uint8_t * x8, const int & i32) {
-    return *((const int *) (x8 + sizeof(int) * i32)); // assume at least 4 byte alignment
-}
-
-template<typename T>
-using to_t_cuda_t = void (*)(const void * __restrict__ x, T * __restrict__ y, int k, cudaStream_t stream);
-typedef to_t_cuda_t<float> to_fp32_cuda_t;
-typedef to_t_cuda_t<half> to_fp16_cuda_t;
-
-typedef void (*dequantize_kernel_t)(const void * vx, const int ib, const int iqs, dfloat2 & v);
-typedef void (*dot_kernel_k_t)(const void * __restrict__ vx, const int ib, const int iqs, const float * __restrict__ y, float & v);
-typedef void (*cpy_kernel_t)(const char * cx, char * cdst);
-typedef void (*ggml_cuda_func_t)(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst);
-typedef void (*ggml_cuda_op_mul_mat_t)(
-    const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, const char * src0_dd_i, const float * src1_ddf_i,
-    const char * src1_ddq_i, float * dst_dd_i, const int64_t row_low, const int64_t row_high, const int64_t src1_ncols,
-    const int64_t src1_padded_row_size, cudaStream_t stream);
-typedef void (*ggml_cuda_op_flatten_t)(
-    const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst,
-    const float * src0_dd, const float * src1_dd, float * dst_dd, cudaStream_t main_stream);
-
-// QK = number of values after dequantization
-// QR = QK / number of values before dequantization
-// QI = number of 32 bit integers before dequantization
-
-#define QK4_0 32
-#define QR4_0 2
-#define QI4_0 (QK4_0 / (4 * QR4_0))
-typedef struct {
-    half    d;              // delta
-    uint8_t qs[QK4_0 / 2];  // nibbles / quants
-} block_q4_0;
-static_assert(sizeof(block_q4_0) == sizeof(ggml_fp16_t) + QK4_0 / 2, "wrong q4_0 block size/padding");
-
-#define QK4_1 32
-#define QR4_1 2
-#define QI4_1 (QK4_1 / (4 * QR4_1))
-typedef struct {
-    half2   dm;             // dm.x = delta, dm.y = min
-    uint8_t qs[QK4_1 / 2];  // nibbles / quants
-} block_q4_1;
-static_assert(sizeof(block_q4_1) == sizeof(ggml_fp16_t) * 2 + QK4_1 / 2, "wrong q4_1 block size/padding");
-
-#define QK5_0 32
-#define QR5_0 2
-#define QI5_0 (QK5_0 / (4 * QR5_0))
-typedef struct {
-    half d;                 // delta
-    uint8_t qh[4];          // 5-th bit of quants
-    uint8_t qs[QK5_0 / 2];  // nibbles / quants
-} block_q5_0;
-static_assert(sizeof(block_q5_0) == sizeof(ggml_fp16_t) + sizeof(uint32_t) + QK5_0 / 2, "wrong q5_0 block size/padding");
-
-#define QK5_1 32
-#define QR5_1 2
-#define QI5_1 (QK5_1 / (4 * QR5_1))
-typedef struct {
-    half2 dm;               // dm.x = delta, dm.y = min
-    uint8_t qh[4];          // 5-th bit of quants
-    uint8_t qs[QK5_1 / 2];  // nibbles / quants
-} block_q5_1;
-static_assert(sizeof(block_q5_1) == 2 * sizeof(ggml_fp16_t) + sizeof(uint32_t) + QK5_1 / 2, "wrong q5_1 block size/padding");
-
-#define QK8_0 32
-#define QR8_0 1
-#define QI8_0 (QK8_0 / (4 * QR8_0))
-typedef struct {
-    half    d;              // delta
-    int8_t  qs[QK8_0];      // quants
-} block_q8_0;
-static_assert(sizeof(block_q8_0) == sizeof(ggml_fp16_t) + QK8_0, "wrong q8_0 block size/padding");
-
-#define QK8_1 32
-#define QR8_1 1
-#define QI8_1 (QK8_1 / (4 * QR8_1))
-typedef struct {
-    half2   ds;             // ds.x = delta, ds.y = sum
-    int8_t  qs[QK8_0];      // quants
-} block_q8_1;
-static_assert(sizeof(block_q8_1) == 2*sizeof(ggml_fp16_t) + QK8_0, "wrong q8_1 block size/padding");
-
-typedef float (*vec_dot_q_cuda_t)(const void * __restrict__ vbq, const block_q8_1 * __restrict__ bq8_1, const int & iqs);
-typedef void (*allocate_tiles_cuda_t)(int ** x_ql, half2 ** x_dm, int ** x_qh, int ** x_sc);
-typedef void (*load_tiles_cuda_t)(
-    const void * __restrict__ vx, int * __restrict__ x_ql, half2 * __restrict__ x_dm, int * __restrict__ x_qh,
-    int * __restrict__ x_sc, const int & i_offset, const int & i_max, const int & k, const int & blocks_per_row);
-typedef float (*vec_dot_q_mul_mat_cuda_t)(
-    const int * __restrict__ x_ql, const half2 * __restrict__ x_dm, const int * __restrict__ x_qh, const int * __restrict__ x_sc,
-    const int * __restrict__ y_qs, const half2 * __restrict__ y_ms, const int & i, const int & j, const int & k);
-
-//================================= k-quants
-
-#ifdef GGML_QKK_64
-#define QK_K 64
-#define K_SCALE_SIZE 4
-#else
-#define QK_K 256
-#define K_SCALE_SIZE 12
-#endif
-
-#define QR2_K 4
-#define QI2_K (QK_K / (4*QR2_K))
-typedef struct {
-    uint8_t scales[QK_K/16]; // scales and mins, quantized with 4 bits
-    uint8_t qs[QK_K/4];      // quants
-    half2 dm;                // super-block scale for quantized scales/mins
-} block_q2_K;
-static_assert(sizeof(block_q2_K) == 2*sizeof(ggml_fp16_t) + QK_K/16 + QK_K/4, "wrong q2_K block size/padding");
-
-#define QR3_K 4
-#define QI3_K (QK_K / (4*QR3_K))
-typedef struct {
-    uint8_t hmask[QK_K/8];     // quants - high bit
-    uint8_t qs[QK_K/4];        // quants - low 2 bits
-#ifdef GGML_QKK_64
-    uint8_t scales[2]; // scales, quantized with 8 bits
-#else
-    uint8_t scales[K_SCALE_SIZE]; // scales, quantized with 6 bits
-#endif
-    half d;             // super-block scale
-} block_q3_K;
-//static_assert(sizeof(block_q3_K) == sizeof(ggml_fp16_t) + QK_K / 4 + QK_K / 8 + K_SCALE_SIZE, "wrong q3_K block size/padding");
-
-#define QR4_K 2
-#define QI4_K (QK_K / (4*QR4_K))
-#ifdef GGML_QKK_64
-typedef struct {
-    half    dm[2];             // super-block scales/mins
-    uint8_t scales[2];         // 4-bit block scales/mins
-    uint8_t qs[QK_K/2];        // 4--bit quants
-} block_q4_K;
-static_assert(sizeof(block_q4_K) == sizeof(half2) + QK_K/2 + 2, "wrong q4_K block size/padding");
-#else
-typedef struct {
-    half2 dm;                  // super-block scale for quantized scales/mins
-    uint8_t scales[3*QK_K/64]; // scales, quantized with 6 bits
-    uint8_t qs[QK_K/2];        // 4--bit quants
-} block_q4_K;
-static_assert(sizeof(block_q4_K) == 2*sizeof(ggml_fp16_t) + 3*QK_K/64 + QK_K/2, "wrong q4_K block size/padding");
-#endif
-
-#define QR5_K 2
-#define QI5_K (QK_K / (4*QR5_K))
-#ifdef GGML_QKK_64
-typedef struct {
-    half d;                  // super-block scale
-    int8_t scales[QK_K/16];  // block scales
-    uint8_t qh[QK_K/8];      // quants, high bit
-    uint8_t qs[QK_K/2];      // quants, low 4 bits
-} block_q5_K;
-static_assert(sizeof(block_q5_K) == sizeof(ggml_fp16_t) + QK_K/2 + QK_K/8 + QK_K/16, "wrong q5_K block size/padding");
-#else
-typedef struct {
-    half2 dm;                     // super-block scale for quantized scales/mins
-    uint8_t scales[K_SCALE_SIZE]; // scales and mins, quantized with 6 bits
-    uint8_t qh[QK_K/8];           // quants, high bit
-    uint8_t qs[QK_K/2];           // quants, low 4 bits
-} block_q5_K;
-static_assert(sizeof(block_q5_K) == 2*sizeof(ggml_fp16_t) + K_SCALE_SIZE + QK_K/2 + QK_K/8, "wrong q5_K block size/padding");
-#endif
-
-#define QR6_K 2
-#define QI6_K (QK_K / (4*QR6_K))
-typedef struct {
-    uint8_t ql[QK_K/2];   // quants, lower 4 bits
-    uint8_t qh[QK_K/4];   // quants, upper 2 bits
-    int8_t  scales[QK_K/16]; // scales
-    half    d;         // delta
-} block_q6_K;
-static_assert(sizeof(block_q6_K) == sizeof(ggml_fp16_t) + 13*QK_K/16, "wrong q6_K block size/padding");
-
-#define WARP_SIZE 32
-#define MATRIX_ROW_PADDING 512 // last row of quant. matrices is a multiple of this to avoid out-of-bounds memory accesses
-
-#define CUDA_GELU_BLOCK_SIZE 256
-#define CUDA_SILU_BLOCK_SIZE 256
-#define CUDA_TANH_BLOCK_SIZE 256
-#define CUDA_RELU_BLOCK_SIZE 256
-#define CUDA_SQR_BLOCK_SIZE 256
-#define CUDA_CPY_BLOCK_SIZE 32
-#define CUDA_SCALE_BLOCK_SIZE 256
-#define CUDA_CLAMP_BLOCK_SIZE 256
-#define CUDA_ROPE_BLOCK_SIZE 256
-#define CUDA_SOFT_MAX_BLOCK_SIZE 1024
-#define CUDA_ALIBI_BLOCK_SIZE 32
-#define CUDA_DIAG_MASK_INF_BLOCK_SIZE 32
-#define CUDA_QUANTIZE_BLOCK_SIZE 256
-#define CUDA_DEQUANTIZE_BLOCK_SIZE 256
-#define CUDA_GET_ROWS_BLOCK_SIZE 256
-#define CUDA_UPSCALE_BLOCK_SIZE 256
-#define CUDA_CONCAT_BLOCK_SIZE 256
-#define CUDA_PAD_BLOCK_SIZE 256
-#define CUDA_ACC_BLOCK_SIZE 256
-#define CUDA_IM2COL_BLOCK_SIZE 256
-
-// dmmv = dequantize_mul_mat_vec
-#ifndef GGML_CUDA_DMMV_X
-#define GGML_CUDA_DMMV_X 32
-#endif
-#ifndef GGML_CUDA_MMV_Y
-#define GGML_CUDA_MMV_Y 1
-#endif
-
-#ifndef K_QUANTS_PER_ITERATION
-#define K_QUANTS_PER_ITERATION 2
-#else
-static_assert(K_QUANTS_PER_ITERATION == 1 || K_QUANTS_PER_ITERATION == 2, "K_QUANTS_PER_ITERATION must be 1 or 2");
-#endif
-
-#ifndef GGML_CUDA_PEER_MAX_BATCH_SIZE
-#define GGML_CUDA_PEER_MAX_BATCH_SIZE 128
-#endif // GGML_CUDA_PEER_MAX_BATCH_SIZE
-
-#define MUL_MAT_SRC1_COL_STRIDE 128
-
-#define MAX_STREAMS 8
-static cudaStream_t g_cudaStreams[GGML_CUDA_MAX_DEVICES][MAX_STREAMS] = { { nullptr } };
-
-struct ggml_tensor_extra_gpu {
-    void * data_device[GGML_CUDA_MAX_DEVICES]; // 1 pointer for each device for split tensors
-    cudaEvent_t events[GGML_CUDA_MAX_DEVICES][MAX_STREAMS]; // events for synchronizing multiple GPUs
-};
-
-// this is faster on Windows
-// probably because the Windows CUDA libraries forget to make this check before invoking the drivers
-static void ggml_cuda_set_device(const int device) {
-    int current_device;
-    CUDA_CHECK(cudaGetDevice(&current_device));
-
-    if (device == current_device) {
-        return;
-    }
-
-    CUDA_CHECK(cudaSetDevice(device));
-}
-
-static int g_device_count = -1;
-static int g_main_device = 0;
-static float g_tensor_split[GGML_CUDA_MAX_DEVICES] = {0};
-
-struct cuda_device_capabilities {
-    int     cc;                 // compute capability
-    bool    vmm;                // virtual memory support
-    size_t  vmm_granularity;    // granularity of virtual memory
-};
-
-static cuda_device_capabilities g_device_caps[GGML_CUDA_MAX_DEVICES] = { {0, false, 0} };
-
-static void * g_scratch_buffer = nullptr;
-static size_t g_scratch_size = 0; // disabled by default
-static size_t g_scratch_offset = 0;
-
-static cublasHandle_t g_cublas_handles[GGML_CUDA_MAX_DEVICES] = {nullptr};
-
-[[noreturn]]
-static __device__ void bad_arch() {
-    printf("ERROR: ggml-cuda was compiled without support for the current GPU architecture.\n");
-    __trap();
-
-    (void) bad_arch; // suppress unused function warning
-}
-
-static __device__ __forceinline__ float warp_reduce_sum(float x) {
-#pragma unroll
-    for (int mask = 16; mask > 0; mask >>= 1) {
-        x += __shfl_xor_sync(0xffffffff, x, mask, 32);
-    }
-    return x;
-}
-
-static __device__ __forceinline__ float2 warp_reduce_sum(float2 a) {
-#pragma unroll
-    for (int mask = 16; mask > 0; mask >>= 1) {
-        a.x += __shfl_xor_sync(0xffffffff, a.x, mask, 32);
-        a.y += __shfl_xor_sync(0xffffffff, a.y, mask, 32);
-    }
-    return a;
-}
-
-static __device__ __forceinline__ float warp_reduce_max(float x) {
-#pragma unroll
-    for (int mask = 16; mask > 0; mask >>= 1) {
-        x = fmaxf(x, __shfl_xor_sync(0xffffffff, x, mask, 32));
-    }
-    return x;
-}
-
-static __device__ __forceinline__ float op_repeat(const float a, const float b) {
-    return b;
-    GGML_UNUSED(a);
-}
-
-static __device__ __forceinline__ float op_add(const float a, const float b) {
-    return a + b;
-}
-
-static __device__ __forceinline__ float op_mul(const float a, const float b) {
-    return a * b;
-}
-
-static __device__ __forceinline__ float op_div(const float a, const float b) {
-    return a / b;
-}
-
-template<float (*bin_op)(const float, const float), typename src0_t, typename src1_t, typename dst_t>
-static __global__ void k_bin_bcast(const src0_t * src0, const src1_t * src1, dst_t * dst,
-        int ne0, int ne1, int ne2, int ne3,
-        int ne10, int ne11, int ne12, int ne13,
-        /*int s0, */ int s1,  int s2,  int s3,
-        /*int s10,*/ int s11, int s12, int s13) {
-    const int i0s = blockDim.x*blockIdx.x + threadIdx.x;
-    const int i1 = (blockDim.y*blockIdx.y + threadIdx.y);
-    const int i2 = (blockDim.z*blockIdx.z + threadIdx.z) / ne3;
-    const int i3 = (blockDim.z*blockIdx.z + threadIdx.z) % ne3;
-
-    if (i0s >= ne0 || i1 >= ne1 || i2 >= ne2 || i3 >= ne3) {
-        return;
-    }
-
-    const int i11 = i1 % ne11;
-    const int i12 = i2 % ne12;
-    const int i13 = i3 % ne13;
-
-    const size_t i_src0 = i3*s3 + i2*s2 + i1*s1;
-    const size_t i_src1 = i13*s13 + i12*s12 + i11*s11;
-    const size_t i_dst  = i_src0;
-
-    const src0_t * src0_row = src0 + i_src0;
-    const src1_t * src1_row = src1 + i_src1;
-    dst_t * dst_row = dst + i_dst;
-
-    for (int i0 = i0s; i0 < ne0; i0 += blockDim.x*gridDim.x) {
-        const int i10 = i0 % ne10;
-        dst_row[i0] = (dst_t)bin_op(src0 ? (float)src0_row[i0] : 0.0f, (float)src1_row[i10]);
-    }
-}
-
-template<float (*bin_op)(const float, const float), typename src0_t, typename src1_t, typename dst_t>
-static __global__ void k_bin_bcast_unravel(const src0_t * src0, const src1_t * src1, dst_t * dst,
-        int ne0, int ne1, int ne2, int ne3,
-        int ne10, int ne11, int ne12, int ne13,
-        /*int s0, */ int s1,  int s2,  int s3,
-        /*int s10,*/ int s11, int s12, int s13) {
-
-    const int i = blockDim.x*blockIdx.x + threadIdx.x;
-
-    const int i3 = i/(ne2*ne1*ne0);
-    const int i2 = (i/(ne1*ne0)) % ne2;
-    const int i1 = (i/ne0) % ne1;
-    const int i0 = i % ne0;
-
-    if (i0 >= ne0 || i1 >= ne1 || i2 >= ne2 || i3 >= ne3) {
-        return;
-    }
-
-    const int i11 = i1 % ne11;
-    const int i12 = i2 % ne12;
-    const int i13 = i3 % ne13;
-
-    const size_t i_src0 = i3*s3 + i2*s2 + i1*s1;
-    const size_t i_src1 = i13*s13 + i12*s12 + i11*s11;
-    const size_t i_dst  = i_src0;
-
-    const src0_t * src0_row = src0 + i_src0;
-    const src1_t * src1_row = src1 + i_src1;
-    dst_t * dst_row = dst + i_dst;
-
-    const int i10 = i0 % ne10;
-    dst_row[i0] = (dst_t)bin_op(src0 ? (float)src0_row[i0] : 0.0f, (float)src1_row[i10]);
-}
-
-static __global__ void acc_f32(const float * x, const float * y, float * dst, const int ne,
-    const int ne10, const int ne11, const int ne12,
-    const int nb1, const int nb2, int offset) {
-    const int i = blockDim.x * blockIdx.x + threadIdx.x;
-    if (i >= ne) {
-        return;
-    }
-    int src1_idx = i - offset;
-    int oz = src1_idx / nb2;
-    int oy = (src1_idx - (oz * nb2)) / nb1;
-    int ox = src1_idx % nb1;
-    if (src1_idx >= 0 && ox < ne10 && oy < ne11 && oz < ne12) {
-        dst[i] = x[i] + y[ox + oy * ne10 + oz * ne10 * ne11];
-    } else {
-        dst[i] = x[i];
-    }
-}
-
-static __global__ void gelu_f32(const float * x, float * dst, const int k) {
-    const float GELU_COEF_A    = 0.044715f;
-    const float SQRT_2_OVER_PI = 0.79788456080286535587989211986876f;
-    const int i = blockDim.x*blockIdx.x + threadIdx.x;
-
-    if (i >= k) {
-        return;
-    }
-
-    float xi = x[i];
-    dst[i] = 0.5f*xi*(1.0f + tanhf(SQRT_2_OVER_PI*xi*(1.0f + GELU_COEF_A*xi*xi)));
-}
-
-static __global__ void silu_f32(const float * x, float * dst, const int k) {
-    const int i = blockDim.x*blockIdx.x + threadIdx.x;
-
-    if (i >= k) {
-        return;
-    }
-    dst[i] = x[i] / (1.0f + expf(-x[i]));
-}
-
-static __global__ void gelu_quick_f32(const float * x, float * dst, int k) {
-    const float GELU_QUICK_COEF = -1.702f;
-    const int i  = blockDim.x*blockIdx.x + threadIdx.x;
-    if (i >= k) {
-        return;
-    }
-    dst[i] = x[i] * (1.0f / (1.0f + expf(GELU_QUICK_COEF * x[i])));
-}
-
-static __global__ void tanh_f32(const float * x, float * dst, int k) {
-    const int i  = blockDim.x*blockIdx.x + threadIdx.x;
-    if (i >= k) {
-        return;
-    }
-    dst[i] = tanhf(x[i]);
-}
-
-static __global__ void relu_f32(const float * x, float * dst, const int k) {
-    const int i = blockDim.x*blockIdx.x + threadIdx.x;
-
-    if (i >= k) {
-        return;
-    }
-    dst[i] = fmaxf(x[i], 0);
-}
-
-static __global__ void leaky_relu_f32(const float * x, float * dst, const int k, const float negative_slope) {
-    const int i  = blockDim.x*blockIdx.x + threadIdx.x;
-    if (i >= k) {
-        return;
-    }
-    dst[i] = fmaxf(x[i], 0) + fminf(x[i], 0.0f) * negative_slope;
-}
-
-static __global__ void sqr_f32(const float * x, float * dst, const int k) {
-    const int i = blockDim.x*blockIdx.x + threadIdx.x;
-
-    if (i >= k) {
-        return;
-    }
-    dst[i] = x[i] * x[i];
-}
-
-template <int block_size>
-static __global__ void norm_f32(const float * x, float * dst, const int ncols, const float eps) {
-    const int row = blockIdx.x*blockDim.y + threadIdx.y;
-    const int tid = threadIdx.x;
-
-    float2 mean_var = make_float2(0.f, 0.f);
-
-    for (int col = tid; col < ncols; col += block_size) {
-        const float xi = x[row*ncols + col];
-        mean_var.x += xi;
-        mean_var.y += xi * xi;
-    }
-
-    // sum up partial sums
-    mean_var = warp_reduce_sum(mean_var);
-    if (block_size > WARP_SIZE) {
-        __shared__ float2 s_sum[32];
-        int warp_id = threadIdx.x / WARP_SIZE;
-        int lane_id = threadIdx.x % WARP_SIZE;
-        if (lane_id == 0) {
-            s_sum[warp_id] = mean_var;
-        }
-        __syncthreads();
-        mean_var = s_sum[lane_id];
-        mean_var = warp_reduce_sum(mean_var);
-    }
-
-    const float mean = mean_var.x / ncols;
-    const float var = mean_var.y / ncols - mean * mean;
-    const float inv_std = rsqrtf(var + eps);
-
-    for (int col = tid; col < ncols; col += block_size) {
-        dst[row*ncols + col] = (x[row*ncols + col] - mean) * inv_std;
-    }
-}
-
-static __global__ void concat_f32(const float * x,const float * y, float * dst, const int ne0, const int ne02) {
-    int nidx = threadIdx.x + blockIdx.x * blockDim.x;
-    if (nidx >= ne0) {
-        return;
-    }
-    // operation
-    int offset_dst =
-        nidx +
-        blockIdx.y * ne0 +
-        blockIdx.z * ne0 * gridDim.y;
-    if (blockIdx.z < ne02) { // src0
-        int offset_src =
-            nidx +
-            blockIdx.y * ne0 +
-            blockIdx.z * ne0 * gridDim.y;
-            dst[offset_dst] = x[offset_src];
-    } else {
-        int offset_src =
-            nidx +
-            blockIdx.y * ne0 +
-            (blockIdx.z - ne02) * ne0 *  gridDim.y;
-            dst[offset_dst] = y[offset_src];
-    }
-}
-
-static __global__ void upscale_f32(const float * x, float * dst, const int ne00, const int nb02, const int scale_factor) {
-    int ne0 = ne00 * scale_factor;
-    int nidx = threadIdx.x + blockIdx.x * blockDim.x;
-    if (nidx >= ne0) {
-        return;
-    }
-    // operation
-    int i00 = nidx / scale_factor;
-    int i01 = blockIdx.y / scale_factor;
-    int offset_src =
-        i00 +
-        i01 * ne00 +
-        blockIdx.z * nb02;
-    int offset_dst =
-        nidx +
-        blockIdx.y * ne0 +
-        blockIdx.z * ne0 * gridDim.y;
-    dst[offset_dst] = x[offset_src];
-}
-
-static __global__ void pad_f32(const float * x, float * dst, const int ne0, const int ne00, const int ne01, const int ne02) {
-    int nidx = threadIdx.x + blockIdx.x * blockDim.x;
-    if (nidx >= ne0) {
-        return;
-    }
-
-    // operation
-    int offset_dst =
-        nidx +
-        blockIdx.y * ne0 +
-        blockIdx.z * ne0 * gridDim.y;
-    if (nidx < ne00 && blockIdx.y < ne01 && blockIdx.z < ne02) {
-        int offset_src =
-            nidx +
-            blockIdx.y * ne00 +
-            blockIdx.z * ne00 * ne01;
-            dst[offset_dst] = x[offset_src];
-    } else {
-        dst[offset_dst] = 0.0f;
-    }
-}
-
-template <int block_size>
-static __global__ void group_norm_f32(const float * x, float * dst, const int group_size, const int ne_elements, const float eps) {
-    int start = blockIdx.x * group_size;
-    int end = start + group_size;
-
-    start += threadIdx.x;
-
-    if (end >= ne_elements) {
-        end = ne_elements;
-    }
-
-    float tmp = 0.0f; // partial sum for thread in warp
-
-    for (int j = start; j < end; j += block_size) {
-        tmp += x[j];
-    }
-
-    tmp = warp_reduce_sum(tmp);
-    if (block_size > WARP_SIZE) {
-        __shared__ float s_sum[32];
-        int warp_id = threadIdx.x / WARP_SIZE;
-        int lane_id = threadIdx.x % WARP_SIZE;
-        if (lane_id == 0) {
-            s_sum[warp_id] = tmp;
-        }
-        __syncthreads();
-        tmp = s_sum[lane_id];
-        tmp = warp_reduce_sum(tmp);
-    }
-
-    float mean = tmp / group_size;
-    tmp = 0.0f;
-
-    for (int j = start; j < end; j += block_size) {
-        float xi = x[j] - mean;
-        dst[j] = xi;
-        tmp += xi * xi;
-    }
-
-    tmp = warp_reduce_sum(tmp);
-    if (block_size > WARP_SIZE) {
-        __shared__ float s_sum[32];
-        int warp_id = threadIdx.x / WARP_SIZE;
-        int lane_id = threadIdx.x % WARP_SIZE;
-        if (lane_id == 0) {
-            s_sum[warp_id] = tmp;
-        }
-        __syncthreads();
-        tmp = s_sum[lane_id];
-        tmp = warp_reduce_sum(tmp);
-    }
-
-    float variance = tmp / group_size;
-    float scale = rsqrtf(variance + eps);
-    for (int j = start; j < end; j += block_size) {
-        dst[j] *= scale;
-    }
-}
-
-template <int block_size>
-static __global__ void rms_norm_f32(const float * x, float * dst, const int ncols, const float eps) {
-    const int row = blockIdx.x*blockDim.y + threadIdx.y;
-    const int tid = threadIdx.x;
-
-    float tmp = 0.0f; // partial sum for thread in warp
-
-    for (int col = tid; col < ncols; col += block_size) {
-        const float xi = x[row*ncols + col];
-        tmp += xi * xi;
-    }
-
-    // sum up partial sums
-    tmp = warp_reduce_sum(tmp);
-    if (block_size > WARP_SIZE) {
-        __shared__ float s_sum[32];
-        int warp_id = threadIdx.x / WARP_SIZE;
-        int lane_id = threadIdx.x % WARP_SIZE;
-        if (lane_id == 0) {
-            s_sum[warp_id] = tmp;
-        }
-        __syncthreads();
-        tmp = s_sum[lane_id];
-        tmp = warp_reduce_sum(tmp);
-    }
-
-    const float mean = tmp / ncols;
-    const float scale = rsqrtf(mean + eps);
-
-    for (int col = tid; col < ncols; col += block_size) {
-        dst[row*ncols + col] = scale * x[row*ncols + col];
-    }
-}
-
-static __device__ __forceinline__ void dequantize_q4_0(const void * vx, const int ib, const int iqs, dfloat2 & v){
-    const block_q4_0 * x = (const block_q4_0 *) vx;
-
-    const dfloat d = x[ib].d;
-
-    const int vui = x[ib].qs[iqs];
-
-    v.x = vui & 0xF;
-    v.y = vui >> 4;
-
-#ifdef GGML_CUDA_F16
-    v = __hsub2(v, {8.0f, 8.0f});
-    v = __hmul2(v, {d, d});
-#else
-    v.x = (v.x - 8.0f) * d;
-    v.y = (v.y - 8.0f) * d;
-#endif // GGML_CUDA_F16
-}
-
-static __device__ __forceinline__ void dequantize_q4_1(const void * vx, const int ib, const int iqs, dfloat2 & v){
-    const block_q4_1 * x = (const block_q4_1 *) vx;
-
-    const dfloat d = __low2half(x[ib].dm);
-    const dfloat m = __high2half(x[ib].dm);
-
-    const int vui = x[ib].qs[iqs];
-
-    v.x = vui & 0xF;
-    v.y = vui >> 4;
-
-#ifdef GGML_CUDA_F16
-    v = __hmul2(v, {d, d});
-    v = __hadd2(v, {m, m});
-#else
-    v.x = (v.x * d) + m;
-    v.y = (v.y * d) + m;
-#endif // GGML_CUDA_F16
-}
-
-static __device__ __forceinline__ void dequantize_q5_0(const void * vx, const int ib, const int iqs, dfloat2 & v){
-    const block_q5_0 * x = (const block_q5_0 *) vx;
-
-    const dfloat d = x[ib].d;
-
-    uint32_t qh;
-    memcpy(&qh, x[ib].qh, sizeof(qh));
-
-    const int xh_0 = ((qh >> (iqs +  0)) << 4) & 0x10;
-    const int xh_1 = ((qh >> (iqs + 12))     ) & 0x10;
-
-    v.x = ((x[ib].qs[iqs] & 0xf) | xh_0);
-    v.y = ((x[ib].qs[iqs] >>  4) | xh_1);
-
-#ifdef GGML_CUDA_F16
-    v = __hsub2(v, {16.0f, 16.0f});
-    v = __hmul2(v, {d, d});
-#else
-    v.x = (v.x - 16.0f) * d;
-    v.y = (v.y - 16.0f) * d;
-#endif // GGML_CUDA_F16
-}
-
-static __device__ __forceinline__ void dequantize_q5_1(const void * vx, const int ib, const int iqs, dfloat2 & v){
-    const block_q5_1 * x = (const block_q5_1 *) vx;
-
-    const dfloat d = __low2half(x[ib].dm);
-    const dfloat m = __high2half(x[ib].dm);
-
-    uint32_t qh;
-    memcpy(&qh, x[ib].qh, sizeof(qh));
-
-    const int xh_0 = ((qh >> (iqs +  0)) << 4) & 0x10;
-    const int xh_1 = ((qh >> (iqs + 12))     ) & 0x10;
-
-    v.x = ((x[ib].qs[iqs] & 0xf) | xh_0);
-    v.y = ((x[ib].qs[iqs] >>  4) | xh_1);
-
-#ifdef GGML_CUDA_F16
-    v = __hmul2(v, {d, d});
-    v = __hadd2(v, {m, m});
-#else
-    v.x = (v.x * d) + m;
-    v.y = (v.y * d) + m;
-#endif // GGML_CUDA_F16
-}
-
-static __device__ __forceinline__ void dequantize_q8_0(const void * vx, const int ib, const int iqs, dfloat2 & v){
-    const block_q8_0 * x = (const block_q8_0 *) vx;
-
-    const dfloat d = x[ib].d;
-
-    v.x = x[ib].qs[iqs + 0];
-    v.y = x[ib].qs[iqs + 1];
-
-#ifdef GGML_CUDA_F16
-    v = __hmul2(v, {d, d});
-#else
-    v.x *= d;
-    v.y *= d;
-#endif // GGML_CUDA_F16
-}
-
-//================================== k-quants
-
-template<typename dst_t>
-static __global__ void dequantize_block_q2_K(const void * __restrict__ vx, dst_t * __restrict__ yy) {
-
-    const int i   = blockIdx.x;
-    const block_q2_K * x = (const block_q2_K *) vx;
-
-    const int tid = threadIdx.x;
-#if QK_K == 256
-    const int n   = tid/32;
-    const int l   = tid - 32*n;
-    const int is  = 8*n + l/16;
-
-    const uint8_t q = x[i].qs[32*n + l];
-    dst_t * y = yy + i*QK_K + 128*n;
-
-    float dall = __low2half(x[i].dm);
-    float dmin = __high2half(x[i].dm);
-    y[l+ 0] = dall * (x[i].scales[is+0] & 0xF) * ((q >> 0) & 3) - dmin * (x[i].scales[is+0] >> 4);
-    y[l+32] = dall * (x[i].scales[is+2] & 0xF) * ((q >> 2) & 3) - dmin * (x[i].scales[is+2] >> 4);
-    y[l+64] = dall * (x[i].scales[is+4] & 0xF) * ((q >> 4) & 3) - dmin * (x[i].scales[is+4] >> 4);
-    y[l+96] = dall * (x[i].scales[is+6] & 0xF) * ((q >> 6) & 3) - dmin * (x[i].scales[is+6] >> 4);
-#else
-    const int is = tid/16;  // 0 or 1
-    const int il = tid%16;  // 0...15
-    const uint8_t q = x[i].qs[il] >> (2*is);
-    dst_t * y = yy + i*QK_K + 16*is + il;
-    float dall = __low2half(x[i].dm);
-    float dmin = __high2half(x[i].dm);
-    y[ 0] = dall * (x[i].scales[is+0] & 0xF) * ((q >> 0) & 3) - dmin * (x[i].scales[is+0] >> 4);
-    y[32] = dall * (x[i].scales[is+2] & 0xF) * ((q >> 4) & 3) - dmin * (x[i].scales[is+2] >> 4);
-#endif
-
-}
-
-template<typename dst_t>
-static __global__ void dequantize_block_q3_K(const void * __restrict__ vx, dst_t * __restrict__ yy) {
-
-    const int i = blockIdx.x;
-    const block_q3_K * x = (const block_q3_K *) vx;
-
-#if QK_K == 256
-    const int r = threadIdx.x/4;
-    const int tid = r/2;
-    const int is0 = r%2;
-    const int l0 = 16*is0 + 4*(threadIdx.x%4);
-    const int n = tid / 4;
-    const int j = tid - 4*n;
-
-    uint8_t m = 1 << (4*n + j);
-    int is = 8*n + 2*j + is0;
-    int shift = 2*j;
-
-    int8_t us = is <  4 ? (x[i].scales[is-0] & 0xF) | (((x[i].scales[is+8] >> 0) & 3) << 4) :
-                is <  8 ? (x[i].scales[is-0] & 0xF) | (((x[i].scales[is+4] >> 2) & 3) << 4) :
-                is < 12 ? (x[i].scales[is-8] >>  4) | (((x[i].scales[is+0] >> 4) & 3) << 4) :
-                          (x[i].scales[is-8] >>  4) | (((x[i].scales[is-4] >> 6) & 3) << 4);
-    float d_all = x[i].d;
-    float dl = d_all * (us - 32);
-
-    dst_t * y = yy + i*QK_K + 128*n + 32*j;
-    const uint8_t * q = x[i].qs + 32*n;
-    const uint8_t * hm = x[i].hmask;
-
-    for (int l = l0; l < l0+4; ++l) y[l] = dl * ((int8_t)((q[l] >> shift) & 3) - ((hm[l] & m) ? 0 : 4));
-#else
-    const int tid = threadIdx.x;
-    const int is  = tid/16;  // 0 or 1
-    const int il  = tid%16;  // 0...15
-    const int im  = il/8;    // 0...1
-    const int in  = il%8;    // 0...7
-
-    dst_t * y = yy + i*QK_K + 16*is + il;
-
-    const uint8_t q = x[i].qs[il] >> (2*is);
-    const uint8_t h = x[i].hmask[in] >> (2*is + im);
-    const float   d = (float)x[i].d;
-
-    if (is == 0) {
-        y[ 0] = d * ((x[i].scales[0] & 0xF) - 8) * ((int8_t)((q >> 0) & 3) - ((h >> 0) & 1 ? 0 : 4));
-        y[32] = d * ((x[i].scales[1] & 0xF) - 8) * ((int8_t)((q >> 4) & 3) - ((h >> 4) & 1 ? 0 : 4));
-    } else {
-        y[ 0] = d * ((x[i].scales[0] >>  4) - 8) * ((int8_t)((q >> 0) & 3) - ((h >> 0) & 1 ? 0 : 4));
-        y[32] = d * ((x[i].scales[1] >>  4) - 8) * ((int8_t)((q >> 4) & 3) - ((h >> 4) & 1 ? 0 : 4));
-    }
-#endif
-
-}
-
-#if QK_K == 256
-static inline __device__ void get_scale_min_k4(int j, const uint8_t * q, uint8_t & d, uint8_t & m) {
-    if (j < 4) {
-        d = q[j] & 63; m = q[j + 4] & 63;
-    } else {
-        d = (q[j+4] & 0xF) | ((q[j-4] >> 6) << 4);
-        m = (q[j+4] >>  4) | ((q[j-0] >> 6) << 4);
-    }
-}
-#endif
-
-template<typename dst_t>
-static __global__ void dequantize_block_q4_K(const void * __restrict__ vx, dst_t * __restrict__ yy) {
-    const block_q4_K * x = (const block_q4_K *) vx;
-
-    const int i = blockIdx.x;
-
-#if QK_K == 256
-    // assume 32 threads
-    const int tid = threadIdx.x;
-    const int il  = tid/8;
-    const int ir  = tid%8;
-    const int is  = 2*il;
-    const int n   = 4;
-
-    dst_t * y = yy + i*QK_K + 64*il + n*ir;
-
-    const float dall = __low2half(x[i].dm);
-    const float dmin = __high2half(x[i].dm);
-
-    const uint8_t * q = x[i].qs + 32*il + n*ir;
-
-    uint8_t sc, m;
-    get_scale_min_k4(is + 0, x[i].scales, sc, m);
-    const float d1 = dall * sc; const float m1 = dmin * m;
-    get_scale_min_k4(is + 1, x[i].scales, sc, m);
-    const float d2 = dall * sc; const float m2 = dmin * m;
-    for (int l = 0; l < n; ++l) {
-        y[l + 0] = d1 * (q[l] & 0xF) - m1;
-        y[l +32] = d2 * (q[l] >>  4) - m2;
-    }
-#else
-    const int tid = threadIdx.x;
-    const uint8_t * q = x[i].qs;
-    dst_t * y = yy + i*QK_K;
-    const float d = (float)x[i].dm[0];
-    const float m = (float)x[i].dm[1];
-    y[tid+ 0] = d * (x[i].scales[0] & 0xF) * (q[tid] & 0xF) - m * (x[i].scales[0] >> 4);
-    y[tid+32] = d * (x[i].scales[1] & 0xF) * (q[tid] >>  4) - m * (x[i].scales[1] >> 4);
-#endif
-}
-
-template<typename dst_t>
-static __global__ void dequantize_block_q5_K(const void * __restrict__ vx, dst_t * __restrict__ yy) {
-    const block_q5_K * x = (const block_q5_K *) vx;
-
-    const int i = blockIdx.x;
-
-#if QK_K == 256
-    // assume 64 threads - this is very slightly better than the one below
-    const int tid = threadIdx.x;
-    const int il  = tid/16;   // il is in 0...3
-    const int ir  = tid%16;   // ir is in 0...15
-    const int is  = 2*il;     // is is in 0...6
-
-    dst_t * y = yy + i*QK_K + 64*il + 2*ir;
-
-    const float dall = __low2half(x[i].dm);
-    const float dmin = __high2half(x[i].dm);
-
-    const uint8_t * ql = x[i].qs + 32*il + 2*ir;
-    const uint8_t * qh = x[i].qh + 2*ir;
-
-    uint8_t sc, m;
-    get_scale_min_k4(is + 0, x[i].scales, sc, m);
-    const float d1 = dall * sc; const float m1 = dmin * m;
-    get_scale_min_k4(is + 1, x[i].scales, sc, m);
-    const float d2 = dall * sc; const float m2 = dmin * m;
-
-    uint8_t   hm  = 1 << (2*il);
-    y[ 0] = d1 * ((ql[ 0] & 0xF) + (qh[ 0] & hm ? 16 : 0)) - m1;
-    y[ 1] = d1 * ((ql[ 1] & 0xF) + (qh[ 1] & hm ? 16 : 0)) - m1;
-    hm <<= 1;
-    y[32] = d2 * ((ql[ 0] >>  4) + (qh[ 0] & hm ? 16 : 0)) - m2;
-    y[33] = d2 * ((ql[ 1] >>  4) + (qh[ 1] & hm ? 16 : 0)) - m2;
-#else
-    const int tid = threadIdx.x;
-    const uint8_t q = x[i].qs[tid];
-    const int im = tid/8;  // 0...3
-    const int in = tid%8;  // 0...7
-    const int is = tid/16; // 0 or 1
-    const uint8_t h = x[i].qh[in] >> im;
-    const float d = x[i].d;
-    dst_t * y = yy + i*QK_K + tid;
-    y[ 0] = d * x[i].scales[is+0] * ((q & 0xF) - ((h >> 0) & 1 ? 0 : 16));
-    y[32] = d * x[i].scales[is+2] * ((q >>  4) - ((h >> 4) & 1 ? 0 : 16));
-#endif
-}
-
-template<typename dst_t>
-static __global__ void dequantize_block_q6_K(const void * __restrict__ vx, dst_t * __restrict__ yy) {
-    const block_q6_K * x = (const block_q6_K *) vx;
-
-    const int i = blockIdx.x;
-#if QK_K == 256
-
-    // assume 64 threads - this is very slightly better than the one below
-    const int tid = threadIdx.x;
-    const int ip  = tid/32;   // ip is 0 or 1
-    const int il  = tid - 32*ip; // 0...32
-    const int is  = 8*ip + il/16;
-
-    dst_t * y = yy + i*QK_K + 128*ip + il;
-
-    const float d = x[i].d;
-
-    const uint8_t * ql = x[i].ql + 64*ip + il;
-    const uint8_t   qh = x[i].qh[32*ip + il];
-    const int8_t  * sc = x[i].scales + is;
-
-    y[ 0] = d * sc[0] * ((int8_t)((ql[ 0] & 0xF) | (((qh >> 0) & 3) << 4)) - 32);
-    y[32] = d * sc[2] * ((int8_t)((ql[32] & 0xF) | (((qh >> 2) & 3) << 4)) - 32);
-    y[64] = d * sc[4] * ((int8_t)((ql[ 0]  >> 4) | (((qh >> 4) & 3) << 4)) - 32);
-    y[96] = d * sc[6] * ((int8_t)((ql[32]  >> 4) | (((qh >> 6) & 3) << 4)) - 32);
-#else
-
-    // assume 32 threads
-    const int tid = threadIdx.x;
-    const int ip  = tid/16;         // 0 or 1
-    const int il  = tid - 16*ip;    // 0...15
-
-    dst_t * y = yy + i*QK_K + 16*ip + il;
-
-    const float d = x[i].d;
-
-    const uint8_t   ql = x[i].ql[16*ip + il];
-    const uint8_t   qh = x[i].qh[il] >> (2*ip);
-    const int8_t  * sc = x[i].scales;
-
-    y[ 0] = d * sc[ip+0] * ((int8_t)((ql & 0xF) | (((qh >> 0) & 3) << 4)) - 32);
-    y[32] = d * sc[ip+2] * ((int8_t)((ql  >> 4) | (((qh >> 4) & 3) << 4)) - 32);
-#endif
-}
-
-static __global__ void dequantize_mul_mat_vec_q2_k(const void * __restrict__ vx, const float * __restrict__ yy, float * __restrict__ dst, const int ncols, int nrows) {
-
-    static_assert(16%K_QUANTS_PER_ITERATION == 0, "16 must be divisible by K_QUANTS_PER_ITERATION");
-
-    const int row = blockIdx.x*blockDim.y + threadIdx.y;
-    if (row > nrows) return;
-
-    const int num_blocks_per_row = ncols / QK_K;
-    const int ib0 = row*num_blocks_per_row;
-
-    const block_q2_K * x = (const block_q2_K *)vx + ib0;
-
-    float tmp = 0; // partial sum for thread in warp
-
-#if QK_K == 256
-    const int tid = threadIdx.x/K_QUANTS_PER_ITERATION;  // 0...31 or 0...15
-    const int ix  = threadIdx.x%K_QUANTS_PER_ITERATION;  // 0 or 0,1
-
-    const int step = 16/K_QUANTS_PER_ITERATION;
-
-    const int im = tid/step;                             // 0 or 1. 0 computes 0..., 1 computes 128...
-    const int in = tid - step*im;                        // 0...15 or 0...7
-
-    const int l0 = K_QUANTS_PER_ITERATION*in;            // 0...15 or 0...14 in steps of 2
-    const int q_offset = 32*im + l0;
-    const int s_offset = 8*im;
-    const int y_offset = 128*im + l0;
-
-    uint32_t aux[4];
-    const uint8_t * d = (const uint8_t *)aux;
-    const uint8_t * m = (const uint8_t *)(aux + 2);
-
-    for (int i = ix; i < num_blocks_per_row; i += K_QUANTS_PER_ITERATION) {
-
-        const float   * y = yy + i * QK_K + y_offset;
-        const uint8_t * q = x[i].qs + q_offset;
-
-        const float dall = __low2half(x[i].dm);
-        const float dmin = __high2half(x[i].dm);
-
-        const uint32_t * a = (const uint32_t *)(x[i].scales + s_offset);
-        aux[0] = a[0] & 0x0f0f0f0f;
-        aux[1] = a[1] & 0x0f0f0f0f;
-        aux[2] = (a[0] >> 4) & 0x0f0f0f0f;
-        aux[3] = (a[1] >> 4) & 0x0f0f0f0f;
-
-        float sum1 = 0, sum2 = 0;
-        for (int l = 0; l < K_QUANTS_PER_ITERATION; ++l) {
-            sum1 += y[l+ 0] * d[0] * ((q[l+ 0] >> 0) & 3)
-                  + y[l+32] * d[2] * ((q[l+ 0] >> 2) & 3)
-                  + y[l+64] * d[4] * ((q[l+ 0] >> 4) & 3)
-                  + y[l+96] * d[6] * ((q[l+ 0] >> 6) & 3)
-                  + y[l+16] * d[1] * ((q[l+16] >> 0) & 3)
-                  + y[l+48] * d[3] * ((q[l+16] >> 2) & 3)
-                  + y[l+80] * d[5] * ((q[l+16] >> 4) & 3)
-                  +y[l+112] * d[7] * ((q[l+16] >> 6) & 3);
-            sum2 += y[l+ 0] * m[0] + y[l+32] * m[2] + y[l+64] * m[4] + y[ l+96] * m[6]
-                  + y[l+16] * m[1] + y[l+48] * m[3] + y[l+80] * m[5] + y[l+112] * m[7];
-
-        }
-        tmp += dall * sum1 - dmin * sum2;
-
-    }
-#else
-    const int tid = threadIdx.x/(2*K_QUANTS_PER_ITERATION);  // 0...15 or 0...7
-    const int ix  = threadIdx.x%(2*K_QUANTS_PER_ITERATION);  // 0....1 or 0...3
-    const int offset = tid * K_QUANTS_PER_ITERATION;
-
-    uint32_t uaux[2];
-    const uint8_t * d = (const uint8_t *)uaux;
-
-    for (int i = ix; i < num_blocks_per_row; i += 2*K_QUANTS_PER_ITERATION) {
-
-        const float   * y = yy + i * QK_K + offset;
-        const uint8_t * q = x[i].qs + offset;
-        const uint32_t * s = (const uint32_t *)x[i].scales;
-
-        uaux[0] = s[0] & 0x0f0f0f0f;
-        uaux[1] = (s[0] >> 4) & 0x0f0f0f0f;
-
-        const float2 dall = __half22float2(x[i].dm);
-
-        float sum1 = 0, sum2 = 0;
-        for (int l = 0; l < K_QUANTS_PER_ITERATION; ++l) {
-            const uint8_t ql = q[l];
-            sum1 += y[l+ 0] * d[0] * ((ql >> 0) & 3)
-                  + y[l+16] * d[1] * ((ql >> 2) & 3)
-                  + y[l+32] * d[2] * ((ql >> 4) & 3)
-                  + y[l+48] * d[3] * ((ql >> 6) & 3);
-            sum2 += y[l+0] * d[4] + y[l+16] * d[5] + y[l+32] * d[6] + y[l+48] * d[7];
-        }
-        tmp += dall.x * sum1 - dall.y * sum2;
-    }
-#endif
-
-    // sum up partial sums and write back result
-#pragma unroll
-    for (int mask = 16; mask > 0; mask >>= 1) {
-        tmp += __shfl_xor_sync(0xffffffff, tmp, mask, 32);
-    }
-
-    if (threadIdx.x == 0) {
-        dst[row] = tmp;
-    }
-}
-
-static __global__ void dequantize_mul_mat_vec_q3_k(const void * __restrict__ vx, const float * __restrict__ yy, float * __restrict__ dst, const int ncols, int nrows) {
-
-    const int row = blockIdx.x*blockDim.y + threadIdx.y;
-    if (row > nrows) return;
-
-    const int num_blocks_per_row = ncols / QK_K;
-    const int ib0 = row*num_blocks_per_row;
-
-    const block_q3_K * x = (const block_q3_K *)vx + ib0;
-
-    float tmp = 0; // partial sum for thread in warp
-
-#if QK_K == 256
-
-    const uint16_t kmask1 = 0x0303;
-    const uint16_t kmask2 = 0x0f0f;
-
-    const int tid = threadIdx.x/K_QUANTS_PER_ITERATION;  // 0...31 or 0...16
-    const int ix  = threadIdx.x%K_QUANTS_PER_ITERATION;  // 0 or 0,1
-
-    const int n  = K_QUANTS_PER_ITERATION;               // iterations in the inner loop
-    const int step = 16/K_QUANTS_PER_ITERATION;
-    const int im = tid/step;                             // 0 or 1. 0 computes 0..., 1 computes 128...
-    const int in = tid - step*im;                        // 0....15 or 0...7
-
-    const uint8_t m = 1 << (4*im);
-
-    const int l0 = n*in;                                 // 0...15 or 0...14 in steps of 2
-    const int q_offset =  32*im + l0;
-    const int y_offset = 128*im + l0;
-
-    uint16_t utmp[4];
-    const int8_t * s = (const int8_t *)utmp;
-
-    const uint16_t s_shift = 4*im;
-
-    for (int i = ix; i < num_blocks_per_row; i += K_QUANTS_PER_ITERATION) {
-
-        const float   * y  = yy + i * QK_K + y_offset;
-        const uint8_t * q = x[i].qs + q_offset;
-        const uint8_t * h = x[i].hmask + l0;
-
-        const uint16_t * a = (const uint16_t *)x[i].scales;
-        utmp[0] = ((a[0] >> s_shift) & kmask2) | (((a[4] >> (s_shift + 0)) & kmask1) << 4);
-        utmp[1] = ((a[1] >> s_shift) & kmask2) | (((a[5] >> (s_shift + 0)) & kmask1) << 4);
-        utmp[2] = ((a[2] >> s_shift) & kmask2) | (((a[4] >> (s_shift + 2)) & kmask1) << 4);
-        utmp[3] = ((a[3] >> s_shift) & kmask2) | (((a[5] >> (s_shift + 2)) & kmask1) << 4);
-
-        const float d = x[i].d;
-
-        float sum = 0;
-        for (int l = 0; l < n; ++l) {
-            sum += y[l+ 0] * (s[0] - 32) * (((q[l] >> 0) & 3) - (h[l] & (m << 0) ? 0 : 4))
-                 + y[l+32] * (s[2] - 32) * (((q[l] >> 2) & 3) - (h[l] & (m << 1) ? 0 : 4))
-                 + y[l+64] * (s[4] - 32) * (((q[l] >> 4) & 3) - (h[l] & (m << 2) ? 0 : 4))
-                 + y[l+96] * (s[6] - 32) * (((q[l] >> 6) & 3) - (h[l] & (m << 3) ? 0 : 4));
-            sum += y[l+16] * (s[1] - 32) * (((q[l+16] >> 0) & 3) - (h[l+16] & (m << 0) ? 0 : 4))
-                 + y[l+48] * (s[3] - 32) * (((q[l+16] >> 2) & 3) - (h[l+16] & (m << 1) ? 0 : 4))
-                 + y[l+80] * (s[5] - 32) * (((q[l+16] >> 4) & 3) - (h[l+16] & (m << 2) ? 0 : 4))
-                + y[l+112] * (s[7] - 32) * (((q[l+16] >> 6) & 3) - (h[l+16] & (m << 3) ? 0 : 4));
-        }
-        tmp += d * sum;
-
-    }
-#else
-
-    const int tid = threadIdx.x/(2*K_QUANTS_PER_ITERATION);  // 0...15 or 0...7
-    const int ix  = threadIdx.x%(2*K_QUANTS_PER_ITERATION);  // 0....1 or 0...3
-    const int offset = tid * K_QUANTS_PER_ITERATION;         // 0...15 or 0...14
-    const int in = offset/8;                                 // 0 or 1
-    const int im = offset%8;                                 // 0...7
-
-    for (int i = ix; i < num_blocks_per_row; i += 2*K_QUANTS_PER_ITERATION) {
-
-        const float   * y = yy + i * QK_K + offset;
-        const uint8_t * q = x[i].qs + offset;
-        const uint8_t * s = x[i].scales;
-
-        const float dall = (float)x[i].d;
-
-        float sum = 0;
-        for (int l = 0; l < K_QUANTS_PER_ITERATION; ++l) {
-            const uint8_t hl = x[i].hmask[im+l] >> in;
-            const uint8_t ql = q[l];
-            sum += y[l+ 0] * dall * ((s[0] & 0xF) - 8) * ((int8_t)((ql >> 0) & 3) - ((hl >> 0) & 1 ? 0 : 4))
-                 + y[l+16] * dall * ((s[0] >>  4) - 8) * ((int8_t)((ql >> 2) & 3) - ((hl >> 2) & 1 ? 0 : 4))
-                 + y[l+32] * dall * ((s[1] & 0xF) - 8) * ((int8_t)((ql >> 4) & 3) - ((hl >> 4) & 1 ? 0 : 4))
-                 + y[l+48] * dall * ((s[1] >>  4) - 8) * ((int8_t)((ql >> 6) & 3) - ((hl >> 6) & 1 ? 0 : 4));
-        }
-        tmp += sum;
-    }
-#endif
-
-    // sum up partial sums and write back result
-#pragma unroll
-    for (int mask = 16; mask > 0; mask >>= 1) {
-        tmp += __shfl_xor_sync(0xffffffff, tmp, mask, 32);
-    }
-
-    if (threadIdx.x == 0) {
-        dst[row] = tmp;
-    }
-}
-
-static __global__ void dequantize_mul_mat_vec_q4_k(const void * __restrict__ vx, const float * __restrict__ yy, float * __restrict__ dst, const int ncols, int nrows) {
-
-    const int row = blockIdx.x*blockDim.y + threadIdx.y;
-    if (row > nrows) return;
-    const int num_blocks_per_row = ncols / QK_K;
-    const int ib0 = row*num_blocks_per_row;
-
-    const block_q4_K * x = (const block_q4_K *)vx + ib0;
-
-#if QK_K == 256
-    const uint16_t kmask1 = 0x3f3f;
-    const uint16_t kmask2 = 0x0f0f;
-    const uint16_t kmask3 = 0xc0c0;
-
-    const int tid = threadIdx.x/K_QUANTS_PER_ITERATION;  // 0...31 or 0...16
-    const int ix  = threadIdx.x%K_QUANTS_PER_ITERATION;  // 0 or 0,1
-
-    const int step = 8/K_QUANTS_PER_ITERATION;           // 8 or 4
-
-    const int il  = tid/step;                            // 0...3
-    const int ir  = tid - step*il;                       // 0...7 or 0...3
-    const int n   = 2 * K_QUANTS_PER_ITERATION;          // 2 or 4
-
-    const int im = il/2;  // 0 or 1. 0 computes 0,32 + 128,160, 1 computes 64,96 + 192,224
-    const int in = il%2;
-
-    const int l0 = n*(2*ir + in);
-    const int q_offset = 32*im + l0;
-    const int y_offset = 64*im + l0;
-
-    uint16_t aux[4];
-    const uint8_t * sc = (const uint8_t *)aux;
-
-#if K_QUANTS_PER_ITERATION == 2
-    uint32_t q32[4];
-    const uint8_t * q4 = (const uint8_t *)q32;
-#else
-    uint16_t q16[4];
-    const uint8_t * q4 = (const uint8_t *)q16;
-#endif
-
-    float tmp = 0; // partial sum for thread in warp
-
-    for (int i = ix; i < num_blocks_per_row; i += K_QUANTS_PER_ITERATION) {
-
-        const float   * y1 = yy + i*QK_K + y_offset;
-        const float   * y2 = y1 + 128;
-
-        const float dall = __low2half(x[i].dm);
-        const float dmin = __high2half(x[i].dm);
-
-        const uint16_t * a = (const uint16_t *)x[i].scales;
-        aux[0] = a[im+0] & kmask1;
-        aux[1] = a[im+2] & kmask1;
-        aux[2] = ((a[im+4] >> 0) & kmask2) | ((a[im+0] & kmask3) >> 2);
-        aux[3] = ((a[im+4] >> 4) & kmask2) | ((a[im+2] & kmask3) >> 2);
-
-#if K_QUANTS_PER_ITERATION == 2
-        const uint32_t * q1 = (const uint32_t *)(x[i].qs + q_offset);
-        const uint32_t * q2 = q1 + 16;
-
-        q32[0] = q1[0] & 0x0f0f0f0f;
-        q32[1] = q1[0] & 0xf0f0f0f0;
-        q32[2] = q2[0] & 0x0f0f0f0f;
-        q32[3] = q2[0] & 0xf0f0f0f0;
-
-        float4 s = {0.f, 0.f, 0.f, 0.f};
-        float smin = 0;
-        for (int l = 0; l < 4; ++l) {
-            s.x += y1[l] * q4[l+0]; s.y += y1[l+32] * q4[l+ 4];
-            s.z += y2[l] * q4[l+8]; s.w += y2[l+32] * q4[l+12];
-            smin += y1[l] * sc[2] + y1[l+32] * sc[3] + y2[l] * sc[6] + y2[l+32] * sc[7];
-        }
-        tmp += dall * (s.x * sc[0] + s.y * sc[1] * 1.f/16.f + s.z * sc[4] + s.w * sc[5] * 1.f/16.f) - dmin * smin;
-#else
-        const uint16_t * q1 = (const uint16_t *)(x[i].qs + q_offset);
-        const uint16_t * q2 = q1 + 32;
-
-        q16[0] = q1[0] & 0x0f0f;
-        q16[1] = q1[0] & 0xf0f0;
-        q16[2] = q2[0] & 0x0f0f;
-        q16[3] = q2[0] & 0xf0f0;
-
-        float4 s = {0.f, 0.f, 0.f, 0.f};
-        float smin = 0;
-        for (int l = 0; l < 2; ++l) {
-            s.x += y1[l] * q4[l+0]; s.y += y1[l+32] * q4[l+2];
-            s.z += y2[l] * q4[l+4]; s.w += y2[l+32] * q4[l+6];
-            smin += y1[l] * sc[2] + y1[l+32] * sc[3] + y2[l] * sc[6] + y2[l+32] * sc[7];
-        }
-        tmp += dall * (s.x * sc[0] + s.y * sc[1] * 1.f/16.f + s.z * sc[4] + s.w * sc[5] * 1.f/16.f) - dmin * smin;
-#endif
-
-    }
-#else
-    const int tid = threadIdx.x/(2*K_QUANTS_PER_ITERATION);  // 0...15
-    const int ix  = threadIdx.x%(2*K_QUANTS_PER_ITERATION);
-
-    const int step = tid * K_QUANTS_PER_ITERATION;
-
-    uint16_t aux16[2];
-    const uint8_t * s = (const uint8_t *)aux16;
-
-    float tmp = 0;
-
-    for (int i = ix; i < num_blocks_per_row; i += 2*K_QUANTS_PER_ITERATION) {
-        const uint8_t * q = x[i].qs + step;
-        const float   * y = yy + i*QK_K + step;
-        const uint16_t * a = (const uint16_t *)x[i].scales;
-        aux16[0] = a[0] & 0x0f0f;
-        aux16[1] = (a[0] >> 4) & 0x0f0f;
-        const float d = (float)x[i].dm[0];
-        const float m = (float)x[i].dm[1];
-        float sum = 0.f;
-        for (int j = 0; j < K_QUANTS_PER_ITERATION; ++j) {
-            sum += y[j+ 0] * (d * s[0] * (q[j+ 0] & 0xF) - m * s[2])
-                 + y[j+16] * (d * s[0] * (q[j+16] & 0xF) - m * s[2])
-                 + y[j+32] * (d * s[1] * (q[j+ 0] >>  4) - m * s[3])
-                 + y[j+48] * (d * s[1] * (q[j+16] >>  4) - m * s[3]);
-        }
-        tmp += sum;
-    }
-
-#endif
-
-    // sum up partial sums and write back result
-#pragma unroll
-    for (int mask = 16; mask > 0; mask >>= 1) {
-        tmp += __shfl_xor_sync(0xffffffff, tmp, mask, 32);
-    }
-
-    if (tid == 0) {
-        dst[row] = tmp;
-    }
-}
-
-static __global__ void dequantize_mul_mat_vec_q5_k(const void * __restrict__ vx, const float * __restrict__ yy, float * __restrict__ dst, const int ncols) {
-
-    const int row = blockIdx.x;
-    const int num_blocks_per_row = ncols / QK_K;
-    const int ib0 = row*num_blocks_per_row;
-
-    const block_q5_K * x = (const block_q5_K *)vx + ib0;
-
-    float tmp = 0; // partial sum for thread in warp
-
-#if QK_K == 256
-    const uint16_t kmask1 = 0x3f3f;
-    const uint16_t kmask2 = 0x0f0f;
-    const uint16_t kmask3 = 0xc0c0;
-
-    const int tid = threadIdx.x/2;  // 0...15
-    const int ix  = threadIdx.x%2;
-
-    const int il  = tid/4;     // 0...3
-    const int ir  = tid - 4*il;// 0...3
-    const int n   = 2;
-
-    const int im = il/2;  // 0 or 1. 0 computes 0,32 + 128,160, 1 computes 64,96 + 192,224
-    const int in = il%2;
-
-    const int l0 = n*(2*ir + in);
-    const int q_offset = 32*im + l0;
-    const int y_offset = 64*im + l0;
-
-    const uint8_t hm1  = 1 << (2*im);
-    const uint8_t hm2  = hm1 << 4;
-
-    uint16_t aux[4];
-    const uint8_t * sc = (const uint8_t *)aux;
-
-    uint16_t q16[8];
-    const uint8_t * q4 = (const uint8_t *)q16;
-
-    for (int i = ix; i < num_blocks_per_row; i += 2) {
-
-        const uint8_t * ql1 = x[i].qs + q_offset;
-        const uint8_t * qh  = x[i].qh + l0;
-        const float   * y1  = yy + i*QK_K + y_offset;
-        const float   * y2  = y1 + 128;
-
-        const float dall = __low2half(x[i].dm);
-        const float dmin = __high2half(x[i].dm);
-
-        const uint16_t * a = (const uint16_t *)x[i].scales;
-        aux[0] = a[im+0] & kmask1;
-        aux[1] = a[im+2] & kmask1;
-        aux[2] = ((a[im+4] >> 0) & kmask2) | ((a[im+0] & kmask3) >> 2);
-        aux[3] = ((a[im+4] >> 4) & kmask2) | ((a[im+2] & kmask3) >> 2);
-
-        float4 sum = {0.f, 0.f, 0.f, 0.f};
-        float smin = 0;
-        const uint16_t * q1 = (const uint16_t *)ql1;
-        const uint16_t * q2 = q1 + 32;
-        q16[0] = q1[0] & 0x0f0f;
-        q16[1] = q1[8] & 0x0f0f;
-        q16[2] = (q1[0] >> 4) & 0x0f0f;
-        q16[3] = (q1[8] >> 4) & 0x0f0f;
-        q16[4] = q2[0] & 0x0f0f;
-        q16[5] = q2[8] & 0x0f0f;
-        q16[6] = (q2[0] >> 4) & 0x0f0f;
-        q16[7] = (q2[8] >> 4) & 0x0f0f;
-        for (int l = 0; l < n; ++l) {
-            sum.x += y1[l+ 0] * (q4[l +0] + (qh[l+ 0] & (hm1 << 0) ? 16 : 0))
-                   + y1[l+16] * (q4[l +2] + (qh[l+16] & (hm1 << 0) ? 16 : 0));
-            sum.y += y1[l+32] * (q4[l +4] + (qh[l+ 0] & (hm1 << 1) ? 16 : 0))
-                   + y1[l+48] * (q4[l +6] + (qh[l+16] & (hm1 << 1) ? 16 : 0));
-            sum.z += y2[l+ 0] * (q4[l +8] + (qh[l+ 0] & (hm2 << 0) ? 16 : 0))
-                   + y2[l+16] * (q4[l+10] + (qh[l+16] & (hm2 << 0) ? 16 : 0));
-            sum.w += y2[l+32] * (q4[l+12] + (qh[l+ 0] & (hm2 << 1) ? 16 : 0))
-                   + y2[l+48] * (q4[l+14] + (qh[l+16] & (hm2 << 1) ? 16 : 0));
-            smin += (y1[l] + y1[l+16]) * sc[2] + (y1[l+32] + y1[l+48]) * sc[3]
-                  + (y2[l] + y2[l+16]) * sc[6] + (y2[l+32] + y2[l+48]) * sc[7];
-        }
-        tmp += dall * (sum.x * sc[0] + sum.y * sc[1] + sum.z * sc[4] + sum.w * sc[5]) - dmin * smin;
-    }
-
-#else
-    const int tid = threadIdx.x/(2*K_QUANTS_PER_ITERATION);  // 0...15
-    const int ix  = threadIdx.x%(2*K_QUANTS_PER_ITERATION);
-    const int step = tid * K_QUANTS_PER_ITERATION;
-    const int im = step/8;
-    const int in = step%8;
-
-    for (int i = ix; i < num_blocks_per_row; i += 2*K_QUANTS_PER_ITERATION) {
-        const uint8_t * q = x[i].qs + step;
-        const int8_t  * s = x[i].scales;
-        const float   * y = yy + i*QK_K + step;
-        const float     d = x[i].d;
-        float sum = 0.f;
-        for (int j = 0; j < K_QUANTS_PER_ITERATION; ++j) {
-            const uint8_t h = x[i].qh[in+j] >> im;
-            sum += y[j+ 0] * d * s[0] * ((q[j+ 0] & 0xF) - ((h >> 0) & 1 ? 0 : 16))
-                 + y[j+16] * d * s[1] * ((q[j+16] & 0xF) - ((h >> 2) & 1 ? 0 : 16))
-                 + y[j+32] * d * s[2] * ((q[j+ 0] >>  4) - ((h >> 4) & 1 ? 0 : 16))
-                 + y[j+48] * d * s[3] * ((q[j+16] >>  4) - ((h >> 6) & 1 ? 0 : 16));
-        }
-        tmp += sum;
-    }
-#endif
-
-    // sum up partial sums and write back result
-#pragma unroll
-    for (int mask = 16; mask > 0; mask >>= 1) {
-        tmp += __shfl_xor_sync(0xffffffff, tmp, mask, 32);
-    }
-
-    if (threadIdx.x == 0) {
-        dst[row] = tmp;
-    }
-}
-
-static __global__ void dequantize_mul_mat_vec_q6_k(const void * __restrict__ vx, const float * __restrict__ yy, float * __restrict__ dst, const int ncols, int nrows) {
-
-    static_assert(16%K_QUANTS_PER_ITERATION == 0, "16 must be divisible by K_QUANTS_PER_ITERATION");
-
-    const int row = blockIdx.x*blockDim.y + threadIdx.y;
-    if (row > nrows) return;
-
-    const int num_blocks_per_row = ncols / QK_K;
-    const int ib0 = row*num_blocks_per_row;
-
-    const block_q6_K * x = (const block_q6_K *)vx + ib0;
-
-#if QK_K == 256
-
-    const int tid = threadIdx.x/K_QUANTS_PER_ITERATION;  // 0...31 or 0...16
-    const int ix  = threadIdx.x%K_QUANTS_PER_ITERATION;  // 0 or 0, 1
-
-    const int step = 16/K_QUANTS_PER_ITERATION;          // 16 or 8
-
-    const int im = tid/step;                             // 0 or 1. 0 computes 0..., 1 computes 128...
-    const int in = tid - step*im;                        // 0...15 or 0...7
-
-#if K_QUANTS_PER_ITERATION == 1
-    const int l0 = K_QUANTS_PER_ITERATION*in;            // 0...15
-    const int is = 0;
-#else
-    const int l0 = 4 * in;                               // 0, 4, 8, ..., 28
-    const int is = in / 4;
-#endif
-    const int ql_offset = 64*im + l0;
-    const int qh_offset = 32*im + l0;
-    const int s_offset  =  8*im + is;
-    const int y_offset = 128*im + l0;
-
-    float tmp = 0; // partial sum for thread in warp
-
-    for (int i = ix; i < num_blocks_per_row; i += K_QUANTS_PER_ITERATION) {
-
-        const float   * y  = yy + i * QK_K + y_offset;
-        const uint8_t * ql = x[i].ql + ql_offset;
-        const uint8_t * qh = x[i].qh + qh_offset;
-        const int8_t  * s  = x[i].scales + s_offset;
-
-        const float d = x[i].d;
-
-#if K_QUANTS_PER_ITERATION == 1
-        float sum = y[ 0] * s[0] * d * ((int8_t)((ql[ 0] & 0xF) | ((qh[ 0] & 0x03) << 4)) - 32)
-                  + y[16] * s[1] * d * ((int8_t)((ql[16] & 0xF) | ((qh[16] & 0x03) << 4)) - 32)
-                  + y[32] * s[2] * d * ((int8_t)((ql[32] & 0xF) | ((qh[ 0] & 0x0c) << 2)) - 32)
-                  + y[48] * s[3] * d * ((int8_t)((ql[48] & 0xF) | ((qh[16] & 0x0c) << 2)) - 32)
-                  + y[64] * s[4] * d * ((int8_t)((ql[ 0]  >> 4) | ((qh[ 0] & 0x30) >> 0)) - 32)
-                  + y[80] * s[5] * d * ((int8_t)((ql[16]  >> 4) | ((qh[16] & 0x30) >> 0)) - 32)
-                  + y[96] * s[6] * d * ((int8_t)((ql[32]  >> 4) | ((qh[ 0] & 0xc0) >> 2)) - 32)
-                  +y[112] * s[7] * d * ((int8_t)((ql[48]  >> 4) | ((qh[16] & 0xc0) >> 2)) - 32);
-        tmp += sum;
-#else
-        float sum = 0;
-        for (int l = 0; l < 4; ++l) {
-            sum += y[l+ 0] * s[0] * d * ((int8_t)((ql[l+ 0] & 0xF) | (((qh[l] >> 0) & 3) << 4)) - 32)
-                 + y[l+32] * s[2] * d * ((int8_t)((ql[l+32] & 0xF) | (((qh[l] >> 2) & 3) << 4)) - 32)
-                 + y[l+64] * s[4] * d * ((int8_t)((ql[l+ 0]  >> 4) | (((qh[l] >> 4) & 3) << 4)) - 32)
-                 + y[l+96] * s[6] * d * ((int8_t)((ql[l+32]  >> 4) | (((qh[l] >> 6) & 3) << 4)) - 32);
-        }
-        tmp += sum;
-#endif
-
-    }
-
-#else
-
-    const int tid = threadIdx.x/(2*K_QUANTS_PER_ITERATION);  // 0...7
-    const int ix  = threadIdx.x%(2*K_QUANTS_PER_ITERATION);  // 0...3
-
-    const int step = tid * K_QUANTS_PER_ITERATION;
-
-    float tmp = 0; // partial sum for thread in warp
-
-    for (int i = ix; i < num_blocks_per_row; i += 2*K_QUANTS_PER_ITERATION) {
-
-        const float   * y  = yy + i * QK_K + step;
-        const uint8_t * ql = x[i].ql + step;
-        const uint8_t * qh = x[i].qh + step;
-        const int8_t  * s  = x[i].scales;
-
-        const float d = x[i+0].d;
-
-        float sum = 0;
-        for (int j = 0; j < K_QUANTS_PER_ITERATION; ++j) {
-            sum += y[j+ 0] * s[0] * d * ((int8_t)((ql[j+ 0] & 0xF) | ((qh[j] & 0x03) << 4)) - 32)
-                 + y[j+16] * s[1] * d * ((int8_t)((ql[j+16] & 0xF) | ((qh[j] & 0x0c) << 2)) - 32)
-                 + y[j+32] * s[2] * d * ((int8_t)((ql[j+ 0] >>  4) | ((qh[j] & 0x30) >> 0)) - 32)
-                 + y[j+48] * s[3] * d * ((int8_t)((ql[j+16] >>  4) | ((qh[j] & 0xc0) >> 2)) - 32);
-        }
-        tmp += sum;
-
-    }
-
-#endif
-
-    // sum up partial sums and write back result
-#pragma unroll
-    for (int mask = 16; mask > 0; mask >>= 1) {
-        tmp += __shfl_xor_sync(0xffffffff, tmp, mask, 32);
-    }
-
-    if (tid == 0) {
-        dst[row] = tmp;
-    }
-}
-
-static __device__ void convert_f16(const void * vx, const int ib, const int iqs, dfloat2 & v){
-    const half * x = (const half *) vx;
-
-    // automatic half -> float type cast if dfloat == float
-    v.x = x[ib + iqs + 0];
-    v.y = x[ib + iqs + 1];
-}
-
-static __device__ void convert_f32(const void * vx, const int ib, const int iqs, dfloat2 & v){
-    const float * x = (const float *) vx;
-
-    // automatic half -> float type cast if dfloat == float
-    v.x = x[ib + iqs + 0];
-    v.y = x[ib + iqs + 1];
-}
-
-static __global__ void quantize_q8_1(const float * __restrict__ x, void * __restrict__ vy, const int kx, const int kx_padded) {
-    const int ix = blockDim.x*blockIdx.x + threadIdx.x;
-
-    if (ix >= kx_padded) {
-        return;
-    }
-
-    const int iy = blockDim.y*blockIdx.y + threadIdx.y;
-
-    const int i_padded = iy*kx_padded + ix;
-
-    block_q8_1 * y = (block_q8_1 *) vy;
-
-    const int ib = i_padded / QK8_1; // block index
-    const int iqs = i_padded % QK8_1; // quant index
-
-    const float xi = ix < kx ? x[iy*kx + ix] : 0.0f;
-    float amax = fabsf(xi);
-    float sum = xi;
-
-#pragma unroll
-    for (int mask = 16; mask > 0; mask >>= 1) {
-        amax = fmaxf(amax, __shfl_xor_sync(0xffffffff, amax, mask, 32));
-        sum += __shfl_xor_sync(0xffffffff, sum, mask, 32);
-    }
-
-    const float d = amax / 127;
-    const int8_t q = amax == 0.0f ? 0 : roundf(xi / d);
-
-    y[ib].qs[iqs] = q;
-
-    if (iqs > 0) {
-        return;
-    }
-
-    reinterpret_cast<half&>(y[ib].ds.x) = d;
-    reinterpret_cast<half&>(y[ib].ds.y) = sum;
-}
-
-template<int qk, int qr, dequantize_kernel_t dequantize_kernel, typename dst_t>
-static __global__ void k_get_rows(
-            const void * src0, const int32_t * src1, dst_t * dst,
-            int64_t ne00, /*int64_t ne01, int64_t ne02, int64_t ne03,*/
-            /*int64_t ne10, int64_t ne11,*/ int64_t ne12, /*int64_t ne13,*/
-            /*size_t s0,*/ size_t s1, size_t s2, size_t s3,
-            /*size_t nb00,*/ size_t nb01, size_t nb02, size_t nb03,
-            size_t s10, size_t s11, size_t s12/*, size_t s13*/) {
-
-    const int i00 = (blockIdx.x*blockDim.x + threadIdx.x)*2;
-    const int i10 = blockDim.y*blockIdx.y + threadIdx.y;
-    const int i11 = (blockIdx.z*blockDim.z + threadIdx.z)/ne12;
-    const int i12 = (blockIdx.z*blockDim.z + threadIdx.z)%ne12;
-
-    if (i00 >= ne00) {
-        return;
-    }
-
-    const int i01 = src1[i10*s10 + i11*s11 + i12*s12];
-
-    dst_t * dst_row = dst + i10*s1 + i11*s2 + i12*s3;
-    const void * src0_row = (const char *)src0 + i01*nb01 + i11*nb02 + i12*nb03;
-
-    const int ib = i00/qk; // block index
-    const int iqs = (i00%qk)/qr; // quant index
-    const int iybs = i00 - i00%qk; // dst block start index
-    const int y_offset = qr == 1 ? 1 : qk/2;
-
-    // dequantize
-    dfloat2 v;
-    dequantize_kernel(src0_row, ib, iqs, v);
-
-    dst_row[iybs + iqs + 0]        = v.x;
-    dst_row[iybs + iqs + y_offset] = v.y;
-}
-
-template<typename src0_t, typename dst_t>
-static __global__ void k_get_rows_float(
-            const src0_t * src0, const int32_t * src1, dst_t * dst,
-            int64_t ne00, /*int64_t ne01, int64_t ne02, int64_t ne03,*/
-            /*int64_t ne10, int64_t ne11,*/ int64_t ne12, /*int64_t ne13,*/
-            /*size_t s0,*/ size_t s1, size_t s2, size_t s3,
-            /*size_t nb00,*/ size_t nb01, size_t nb02, size_t nb03,
-            size_t s10, size_t s11, size_t s12/*, size_t s13*/) {
-
-    const int i00 = blockIdx.x*blockDim.x + threadIdx.x;
-    const int i10 = blockDim.y*blockIdx.y + threadIdx.y;
-    const int i11 = (blockIdx.z*blockDim.z + threadIdx.z)/ne12;
-    const int i12 = (blockIdx.z*blockDim.z + threadIdx.z)%ne12;
-
-    if (i00 >= ne00) {
-        return;
-    }
-
-    const int i01 = src1[i10*s10 + i11*s11 + i12*s12];
-
-    dst_t * dst_row = dst + i10*s1 + i11*s2 + i12*s3;
-    const src0_t * src0_row = (const src0_t *)((const char *)src0 + i01*nb01 + i11*nb02 + i12*nb03);
-
-    dst_row[i00] = src0_row[i00];
-}
-
-template <int qk, int qr, dequantize_kernel_t dequantize_kernel, typename dst_t>
-static __global__ void dequantize_block(const void * __restrict__ vx, dst_t * __restrict__ y, const int k) {
-    const int i = blockDim.x*blockIdx.x + 2*threadIdx.x;
-
-    if (i >= k) {
-        return;
-    }
-
-    const int ib = i/qk; // block index
-    const int iqs = (i%qk)/qr; // quant index
-    const int iybs = i - i%qk; // y block start index
-    const int y_offset = qr == 1 ? 1 : qk/2;
-
-    // dequantize
-    dfloat2 v;
-    dequantize_kernel(vx, ib, iqs, v);
-
-    y[iybs + iqs + 0]        = v.x;
-    y[iybs + iqs + y_offset] = v.y;
-}
-
-// VDR = vec dot ratio, how many contiguous integers each thread processes when the vec dot kernel is called
-// MMVQ = mul_mat_vec_q, MMQ = mul_mat_q
-
-#define VDR_Q4_0_Q8_1_MMVQ 2
-#define VDR_Q4_0_Q8_1_MMQ  4
-
-template <int vdr> static __device__ __forceinline__ float vec_dot_q4_0_q8_1_impl(
-    const int * v, const int * u, const float & d4, const half2 & ds8) {
-
-#if __CUDA_ARCH__ >= MIN_CC_DP4A // lowest compute capability for integer intrinsics
-    int sumi = 0;
-
-#pragma unroll
-    for (int i = 0; i < vdr; ++i) {
-        const int vi0 = (v[i] >> 0) & 0x0F0F0F0F;
-        const int vi1 = (v[i] >> 4) & 0x0F0F0F0F;
-
-        // SIMD dot product of quantized values
-        sumi = __dp4a(vi0, u[2*i+0], sumi);
-        sumi = __dp4a(vi1, u[2*i+1], sumi);
-    }
-
-    const float2 ds8f = __half22float2(ds8);
-
-    // second part effectively subtracts 8 from each quant value
-    return d4 * (sumi * ds8f.x - (8*vdr/QI4_0) * ds8f.y);
-#else
-    bad_arch();
-#endif // __CUDA_ARCH__ >= MIN_CC_DP4A
-}
-
-#define VDR_Q4_1_Q8_1_MMVQ 2
-#define VDR_Q4_1_Q8_1_MMQ  4
-
-template <int vdr> static __device__ __forceinline__ float vec_dot_q4_1_q8_1_impl(
-    const int * v, const int * u, const half2 & dm4, const half2 & ds8) {
-
-#if __CUDA_ARCH__ >= MIN_CC_DP4A // lowest compute capability for integer intrinsics
-    int sumi = 0;
-
-#pragma unroll
-    for (int i = 0; i < vdr; ++i) {
-        const int vi0 = (v[i] >> 0) & 0x0F0F0F0F;
-        const int vi1 = (v[i] >> 4) & 0x0F0F0F0F;
-
-        // SIMD dot product of quantized values
-        sumi = __dp4a(vi0, u[2*i+0], sumi);
-        sumi = __dp4a(vi1, u[2*i+1], sumi);
-    }
-
-#ifdef GGML_CUDA_F16
-    const float2 tmp = __half22float2(__hmul2(dm4, ds8));
-    const float d4d8 = tmp.x;
-    const float m4s8 = tmp.y;
-#else
-    const float2 dm4f = __half22float2(dm4);
-    const float2 ds8f = __half22float2(ds8);
-    const float d4d8 = dm4f.x * ds8f.x;
-    const float m4s8 = dm4f.y * ds8f.y;
-#endif // GGML_CUDA_F16
-
-    // scale second part of sum by QI8_1/(vdr * QR4_1) to compensate for multiple threads adding it
-    return sumi * d4d8 + m4s8 / (QI8_1 / (vdr * QR4_1));
-#else
-    bad_arch();
-#endif // __CUDA_ARCH__ >= MIN_CC_DP4A
-}
-
-#define VDR_Q5_0_Q8_1_MMVQ 2
-#define VDR_Q5_0_Q8_1_MMQ  4
-
-template <int vdr> static __device__ __forceinline__ float vec_dot_q5_0_q8_1_impl(
-    const int * vl, const int * vh, const int * u, const float & d5, const half2 & ds8) {
-
-#if __CUDA_ARCH__ >= MIN_CC_DP4A // lowest compute capability for integer intrinsics
-    int sumi = 0;
-
-#pragma unroll
-    for (int i = 0; i < vdr; ++i) {
-        int vi0 = (vl[i] >>  0) & 0x0F0F0F0F; // lower 4 qs bits, still need qh as 5th bits
-        vi0    |= (vh[i] <<  4) & 0x00000010; // 0 ->  4
-        vi0    |= (vh[i] << 11) & 0x00001000; // 1 -> 12
-        vi0    |= (vh[i] << 18) & 0x00100000; // 2 -> 20
-        vi0    |= (vh[i] << 25) & 0x10000000; // 3 -> 28
-        sumi = __dp4a(vi0, u[2*i+0], sumi); // SIMD dot product of quantized values
-
-        int vi1 = (vl[i] >>  4) & 0x0F0F0F0F; // upper 4 qs bits, still need qh as 5th bits
-        vi1    |= (vh[i] >> 12) & 0x00000010; // 16 ->  4
-        vi1    |= (vh[i] >>  5) & 0x00001000; // 17 -> 12
-        vi1    |= (vh[i] <<  2) & 0x00100000; // 18 -> 20
-        vi1    |= (vh[i] <<  9) & 0x10000000; // 19 -> 28
-        sumi = __dp4a(vi1, u[2*i+1], sumi); // SIMD dot product of quantized values
-    }
-
-    const float2 ds8f = __half22float2(ds8);
-
-    // second part effectively subtracts 16 from each quant value
-    return d5 * (sumi * ds8f.x - (16*vdr/QI5_0) * ds8f.y);
-#else
-    bad_arch();
-#endif // __CUDA_ARCH__ >= MIN_CC_DP4A
-}
-
-#define VDR_Q5_1_Q8_1_MMVQ 2
-#define VDR_Q5_1_Q8_1_MMQ  4
-
-template <int vdr> static __device__ __forceinline__ float vec_dot_q5_1_q8_1_impl(
-    const int * vl, const int * vh, const int * u, const half2 & dm5, const half2 & ds8) {
-
-#if __CUDA_ARCH__ >= MIN_CC_DP4A // lowest compute capability for integer intrinsics
-    int sumi = 0;
-
-#pragma unroll
-    for (int i = 0; i < vdr; ++i) {
-        int vi0 = (vl[i] >>  0) & 0x0F0F0F0F; // lower 4 qs bits, still need qh as 5th bits
-        vi0    |= (vh[i] <<  4) & 0x00000010; // 0 ->  4
-        vi0    |= (vh[i] << 11) & 0x00001000; // 1 -> 12
-        vi0    |= (vh[i] << 18) & 0x00100000; // 2 -> 20
-        vi0    |= (vh[i] << 25) & 0x10000000; // 3 -> 28
-        sumi = __dp4a(vi0, u[2*i+0], sumi); // SIMD dot product of quantized values
-
-        int vi1 = (vl[i] >>  4) & 0x0F0F0F0F; // upper 4 qs bits, still need qh as 5th bits
-        vi1    |= (vh[i] >> 12) & 0x00000010; // 16 ->  4
-        vi1    |= (vh[i] >>  5) & 0x00001000; // 17 -> 12
-        vi1    |= (vh[i] <<  2) & 0x00100000; // 18 -> 20
-        vi1    |= (vh[i] <<  9) & 0x10000000; // 19 -> 28
-        sumi = __dp4a(vi1, u[2*i+1], sumi); // SIMD dot product of quantized values
-    }
-
-#ifdef GGML_CUDA_F16
-    const float2 tmp = __half22float2(__hmul2(dm5, ds8));
-    const float d5d8 = tmp.x;
-    const float m5s8 = tmp.y;
-#else
-    const float2 dm5f = __half22float2(dm5);
-    const float2 ds8f = __half22float2(ds8);
-    const float d5d8 = dm5f.x * ds8f.x;
-    const float m5s8 = dm5f.y * ds8f.y;
-#endif // GGML_CUDA_F16
-
-    // scale second part of sum by QI5_1 / vdr to compensate for multiple threads adding it
-    return sumi*d5d8 + m5s8 / (QI5_1 / vdr);
-
-#else
-    bad_arch();
-#endif // __CUDA_ARCH__ >= MIN_CC_DP4A
-}
-
-#define VDR_Q8_0_Q8_1_MMVQ 2
-#define VDR_Q8_0_Q8_1_MMQ 8
-
-template <int vdr> static __device__ __forceinline__ float vec_dot_q8_0_q8_1_impl(
-    const int * v, const int * u, const float & d8_0, const float & d8_1) {
-
-#if __CUDA_ARCH__ >= MIN_CC_DP4A // lowest compute capability for integer intrinsics
-    int sumi = 0;
-
-#pragma unroll
-    for (int i = 0; i < vdr; ++i) {
-        // SIMD dot product of quantized values
-        sumi = __dp4a(v[i], u[i], sumi);
-    }
-
-    return d8_0*d8_1 * sumi;
-#else
-    bad_arch();
-#endif // __CUDA_ARCH__ >= MIN_CC_DP4A
-}
-
-template <int vdr> static __device__ __forceinline__ float vec_dot_q8_1_q8_1_impl(
-    const int * v, const int * u, const half2 & dm8, const half2 & ds8) {
-
-#if __CUDA_ARCH__ >= MIN_CC_DP4A // lowest compute capability for integer intrinsics
-    int sumi = 0;
-
-#pragma unroll
-    for (int i = 0; i < vdr; ++i) {
-        // SIMD dot product of quantized values
-        sumi = __dp4a(v[i], u[i], sumi);
-    }
-
-#ifdef GGML_CUDA_F16
-    const float2 tmp = __half22float2(__hmul2(dm8, ds8));
-    const float d8d8 = tmp.x;
-    const float m8s8 = tmp.y;
-#else
-    const float2 dm8f = __half22float2(dm8);
-    const float2 ds8f = __half22float2(ds8);
-    const float d8d8 = dm8f.x * ds8f.x;
-    const float m8s8 = dm8f.y * ds8f.y;
-#endif // GGML_CUDA_F16
-
-    // scale second part of sum by QI8_1/ vdr to compensate for multiple threads adding it
-    return sumi*d8d8 + m8s8 / (QI8_1 / vdr);
-#else
-    bad_arch();
-#endif // __CUDA_ARCH__ >= MIN_CC_DP4A
-}
-
-#define VDR_Q2_K_Q8_1_MMVQ 1
-#define VDR_Q2_K_Q8_1_MMQ  2
-
-// contiguous v/x values
-static __device__ __forceinline__ float vec_dot_q2_K_q8_1_impl_mmvq(
-    const int & v, const int * __restrict__ u, const uint8_t * __restrict__ scales,
-    const half2 & dm2, const float * __restrict__ d8) {
-
-#if __CUDA_ARCH__ >= MIN_CC_DP4A // lowest compute capability for integer intrinsics
-    float sumf_d = 0.0f;
-    float sumf_m = 0.0f;
-
-#pragma unroll
-    for (int i = 0; i < QR2_K; ++i) {
-        const int sc = scales[2*i];
-
-        const int vi = (v >> (2*i)) & 0x03030303;
-
-        sumf_d += d8[i] * (__dp4a(vi, u[i], 0) * (sc & 0xF)); // SIMD dot product
-
-        // fill int with 4x m
-        int m = sc >> 4;
-        m |= m <<  8;
-        m |= m << 16;
-        sumf_m += d8[i] * __dp4a(m, u[i], 0); // multiply constant q2_K part with sum of q8_1 values
-    }
-
-    const float2 dm2f = __half22float2(dm2);
-
-    return dm2f.x*sumf_d - dm2f.y*sumf_m;
-#else
-    bad_arch();
-#endif // __CUDA_ARCH__ >= MIN_CC_DP4A
-}
-
-// contiguous u/y values
-static __device__ __forceinline__ float vec_dot_q2_K_q8_1_impl_mmq(
-    const int * __restrict__ v, const int * __restrict__ u, const uint8_t * __restrict__ scales,
-    const half2 & dm2, const float & d8) {
-
-#if __CUDA_ARCH__ >= MIN_CC_DP4A // lowest compute capability for integer intrinsics
-    int sumi_d = 0;
-    int sumi_m = 0;
-
-#pragma unroll
-    for (int i0 = 0; i0 < QI8_1; i0 += QI8_1/2) {
-        int sumi_d_sc = 0;
-
-        const int sc = scales[i0 / (QI8_1/2)];
-
-        // fill int with 4x m
-        int m = sc >> 4;
-        m |= m <<  8;
-        m |= m << 16;
-
-#pragma unroll
-        for (int i = i0; i < i0 + QI8_1/2; ++i) {
-            sumi_d_sc = __dp4a(v[i], u[i], sumi_d_sc); // SIMD dot product
-            sumi_m    = __dp4a(m,    u[i], sumi_m); // multiply sum of q8_1 values with m
-        }
-
-        sumi_d += sumi_d_sc * (sc & 0xF);
-    }
-
-    const float2 dm2f = __half22float2(dm2);
-
-    return d8 * (dm2f.x*sumi_d - dm2f.y*sumi_m);
-#else
-    bad_arch();
-#endif // __CUDA_ARCH__ >= MIN_CC_DP4A
-}
-
-#define VDR_Q3_K_Q8_1_MMVQ 1
-#define VDR_Q3_K_Q8_1_MMQ  2
-
-// contiguous v/x values
-static __device__ __forceinline__ float vec_dot_q3_K_q8_1_impl_mmvq(
-    const int & vl, const int & vh, const int * __restrict__ u, const uint8_t * __restrict__ scales,
-    const int & scale_offset, const float & d3, const float * __restrict__ d8) {
-
-#if __CUDA_ARCH__ >= MIN_CC_DP4A // lowest compute capability for integer intrinsics
-    float sumf = 0.0f;
-
-#pragma unroll
-    for (int i = 0; i < QR3_K; ++i) {
-        const int isc = scale_offset + 2*i;
-
-        const int isc_low = isc % (QK_K/32);
-        const int sc_shift_low = 4 * (isc / (QK_K/32));
-        const int sc_low  = (scales[isc_low] >> sc_shift_low) & 0xF;
-
-        const int isc_high = isc % (QK_K/64);
-        const int sc_shift_high = 2 * (isc / (QK_K/64));
-        const int sc_high = ((scales[(QK_K/32) + isc_high] >> sc_shift_high) & 3) << 4;
-
-        const int sc = (sc_low | sc_high) - 32;
-
-        const int vil = (vl >> (2*i)) & 0x03030303;
-
-        const int vih = ((vh >> i) << 2) & 0x04040404;
-
-        const int vi = __vsubss4(vil, vih);
-
-        sumf += d8[i] * (__dp4a(vi, u[i], 0) * sc); // SIMD dot product
-    }
-
-    return d3 * sumf;
-#else
-    bad_arch();
-#endif // __CUDA_ARCH__ >= MIN_CC_DP4A
-}
-
-// contiguous u/y values
-static __device__ __forceinline__ float vec_dot_q3_K_q8_1_impl_mmq(
-    const int * __restrict__ v, const int * __restrict__ u, const int8_t * __restrict__ scales,
-    const float & d3, const float & d8) {
-
-#if __CUDA_ARCH__ >= MIN_CC_DP4A // lowest compute capability for integer intrinsics
-    int sumi = 0;
-
-#pragma unroll
-    for (int i0 = 0; i0 < QR3_K*VDR_Q3_K_Q8_1_MMQ; i0 += QI8_1/2) {
-        int sumi_sc = 0;
-
-        for (int i = i0; i < i0 + QI8_1/2; ++i) {
-            sumi_sc = __dp4a(v[i], u[i], sumi_sc); // SIMD dot product
-        }
-
-        sumi += sumi_sc * scales[i0 / (QI8_1/2)];
-    }
-
-    return d3*d8 * sumi;
-#else
-    bad_arch();
-#endif // __CUDA_ARCH__ >= MIN_CC_DP4A
-}
-
-#define VDR_Q4_K_Q8_1_MMVQ 2
-#define VDR_Q4_K_Q8_1_MMQ  8
-
-// contiguous v/x values
-static __device__ __forceinline__ float vec_dot_q4_K_q8_1_impl_vmmq(
-    const int * __restrict__ v, const int * __restrict__ u, const uint8_t * __restrict__ sc,
-    const uint8_t * __restrict__ m, const half2 & dm4, const float * __restrict__ d8) {
-
-#if __CUDA_ARCH__ >= MIN_CC_DP4A // lowest compute capability for integer intrinsics
-    float sumf_d = 0.0f;
-    float sumf_m = 0.0f;
-
-#pragma unroll
-    for (int i = 0; i < QR4_K; ++i) {
-        const int v0i = (v[0] >> (4*i)) & 0x0F0F0F0F;
-        const int v1i = (v[1] >> (4*i)) & 0x0F0F0F0F;
-
-        const int dot1 = __dp4a(v1i, u[2*i+1], __dp4a(v0i, u[2*i+0], 0)); // SIMD dot product
-        const int dot2 = __dp4a(0x01010101, u[2*i+1], __dp4a(0x01010101, u[2*i+0], 0)); // sum of u
-
-        sumf_d += d8[i] * (dot1 * sc[i]);
-        sumf_m += d8[i] * (dot2 * m[i]);  // multiply constant part of q4_K with sum of q8_1 values
-    }
-
-    const float2 dm4f = __half22float2(dm4);
-
-    return dm4f.x*sumf_d - dm4f.y*sumf_m;
-
-#else
-    bad_arch();
-#endif // __CUDA_ARCH__ >= MIN_CC_DP4A
-}
-
-// contiguous u/y values
-static __device__ __forceinline__ float vec_dot_q4_K_q8_1_impl_mmq(
-    const int * __restrict__ v, const int * __restrict__ u, const uint8_t * __restrict__ sc,
-    const uint8_t * __restrict__ m, const half2 & dm4, const half2 * __restrict__ ds8) {
-
-#if __CUDA_ARCH__ >= MIN_CC_DP4A // lowest compute capability for integer intrinsics
-    float sumf_d = 0.0f;
-    float sumf_m = 0.0f;
-
-#pragma unroll
-    for (int i = 0; i < QR4_K*VDR_Q4_K_Q8_1_MMQ/QI8_1; ++i) {
-        int sumi_d = 0;
-
-#pragma unroll
-        for (int j = 0; j < QI8_1; ++j) {
-            sumi_d = __dp4a((v[j] >> (4*i)) & 0x0F0F0F0F, u[i*QI8_1 + j], sumi_d); // SIMD dot product
-        }
-
-        const float2 ds8f = __half22float2(ds8[i]);
-
-        sumf_d += ds8f.x * (sc[i] * sumi_d);
-        sumf_m += ds8f.y *   m[i]; // sum of q8_1 block * q4_K min val
-    }
-
-    const float2 dm4f = __half22float2(dm4);
-
-    return dm4f.x*sumf_d - dm4f.y*sumf_m;
-
-#else
-    bad_arch();
-#endif // __CUDA_ARCH__ >= MIN_CC_DP4A
-}
-
-#define VDR_Q5_K_Q8_1_MMVQ 2
-#define VDR_Q5_K_Q8_1_MMQ  8
-
-// contiguous v/x values
-static __device__ __forceinline__ float vec_dot_q5_K_q8_1_impl_vmmq(
-    const int * __restrict__ vl, const int * __restrict__ vh, const int * __restrict__ u, const uint8_t * __restrict__ sc,
-    const uint8_t * __restrict__ m, const half2 & dm5, const float * __restrict__ d8) {
-
-#if __CUDA_ARCH__ >= MIN_CC_DP4A // lowest compute capability for integer intrinsics
-    float sumf_d = 0.0f;
-    float sumf_m = 0.0f;
-
-#pragma unroll
-    for (int i = 0; i < QR5_K; ++i) {
-        const int vl0i = (vl[0] >> (4*i)) & 0x0F0F0F0F;
-        const int vl1i = (vl[1] >> (4*i)) & 0x0F0F0F0F;
-
-        const int vh0i = ((vh[0] >> i) << 4) & 0x10101010;
-        const int vh1i = ((vh[1] >> i) << 4) & 0x10101010;
-
-        const int v0i = vl0i | vh0i;
-        const int v1i = vl1i | vh1i;
-
-        const int dot1 = __dp4a(v0i, u[2*i+0], __dp4a(v1i, u[2*i+1], 0)); // SIMD dot product
-        const int dot2 = __dp4a(0x01010101, u[2*i+0], __dp4a(0x01010101, u[2*i+1], 0)); // sum of u
-
-        sumf_d += d8[i] * (dot1 * sc[i]);
-        sumf_m += d8[i] * (dot2 * m[i]);
-
-    }
-
-    const float2 dm5f = __half22float2(dm5);
-
-    return dm5f.x*sumf_d - dm5f.y*sumf_m;
-
-#else
-    bad_arch();
-#endif // __CUDA_ARCH__ >= MIN_CC_DP4A
-}
-
-// contiguous u/y values
-static __device__ __forceinline__ float vec_dot_q5_K_q8_1_impl_mmq(
-    const int * __restrict__ v, const int * __restrict__ u, const uint8_t * __restrict__ sc,
-    const uint8_t * __restrict__ m, const half2 & dm4, const half2 * __restrict__ ds8) {
-
-#if __CUDA_ARCH__ >= MIN_CC_DP4A // lowest compute capability for integer intrinsics
-    float sumf_d = 0.0f;
-    float sumf_m = 0.0f;
-
-#pragma unroll
-    for (int i = 0; i < QR5_K*VDR_Q5_K_Q8_1_MMQ/QI8_1; ++i) {
-        int sumi_d = 0;
-
-#pragma unroll
-        for (int j = 0; j < QI8_1; ++j) {
-            sumi_d = __dp4a(v[i*QI8_1 + j], u[i*QI8_1 + j], sumi_d); // SIMD dot product
-        }
-
-        const float2 ds8f = __half22float2(ds8[i]);
-
-        sumf_d += ds8f.x * (sc[i] * sumi_d);
-        sumf_m += ds8f.y *   m[i]; // sum of q8_1 block * q4_K min val
-    }
-
-    const float2 dm4f = __half22float2(dm4);
-
-    return dm4f.x*sumf_d - dm4f.y*sumf_m;
-
-#else
-    bad_arch();
-#endif // __CUDA_ARCH__ >= MIN_CC_DP4A
-}
-
-#define VDR_Q6_K_Q8_1_MMVQ 1
-#define VDR_Q6_K_Q8_1_MMQ  8
-
-// contiguous v/x values
-static __device__ __forceinline__ float vec_dot_q6_K_q8_1_impl_mmvq(
-    const int & vl, const int & vh, const int * __restrict__ u, const int8_t * __restrict__ scales,
-    const float & d, const float * __restrict__ d8) {
-
-#if __CUDA_ARCH__ >= MIN_CC_DP4A // lowest compute capability for integer intrinsics
-    float sumf = 0.0f;
-
-#pragma unroll
-    for (int i = 0; i < QR6_K; ++i) {
-        const int sc = scales[4*i];
-
-        const int vil = (vl >> (4*i)) & 0x0F0F0F0F;
-
-        const int vih = ((vh >> (4*i)) << 4) & 0x30303030;
-
-        const int vi = __vsubss4((vil | vih), 0x20202020); // vi = (vil | vih) - 32
-
-        sumf += d8[i] * (__dp4a(vi, u[i], 0) * sc); // SIMD dot product
-    }
-
-    return d*sumf;
-#else
-    bad_arch();
-#endif // __CUDA_ARCH__ >= MIN_CC_DP4A
-}
-
-// contiguous u/y values
-static __device__ __forceinline__ float vec_dot_q6_K_q8_1_impl_mmq(
-    const int * __restrict__ v, const int * __restrict__ u, const int8_t * __restrict__ sc,
-    const float & d6, const float * __restrict__ d8) {
-
-#if __CUDA_ARCH__ >= MIN_CC_DP4A // lowest compute capability for integer intrinsics
-    float sumf_d = 0.0f;
-
-#pragma unroll
-    for (int i0 = 0; i0 < VDR_Q6_K_Q8_1_MMQ; i0 += 4) {
-        int2 sumi_d = {0, 0}; // 2 q6_K scales per q8_1 scale
-
-#pragma unroll
-        for (int i = i0; i < i0 + 2; ++i) {
-            sumi_d.x = __dp4a(v[2*i+0], u[2*i+0], sumi_d.x); // SIMD dot product
-            sumi_d.x = __dp4a(v[2*i+1], u[2*i+1], sumi_d.x); // SIMD dot product
-
-            sumi_d.y = __dp4a(v[2*i+4], u[2*i+4], sumi_d.y); // SIMD dot product
-            sumi_d.y = __dp4a(v[2*i+5], u[2*i+5], sumi_d.y); // SIMD dot product
-        }
-
-        sumf_d += d8[i0/4] * (sc[i0/2+0]*sumi_d.x + sc[i0/2+1]*sumi_d.y);
-    }
-
-    return d6 * sumf_d;
-
-#else
-    bad_arch();
-#endif // __CUDA_ARCH__ >= MIN_CC_DP4A
-}
-
-static __device__ __forceinline__ float vec_dot_q4_0_q8_1(
-    const void * __restrict__ vbq, const block_q8_1 * __restrict__ bq8_1, const int & iqs) {
-
-    const block_q4_0 * bq4_0 = (const block_q4_0 *) vbq;
-
-    int v[VDR_Q4_0_Q8_1_MMVQ];
-    int u[2*VDR_Q4_0_Q8_1_MMVQ];
-
-#pragma unroll
-    for (int i = 0; i < VDR_Q4_0_Q8_1_MMVQ; ++i) {
-        v[i]     = get_int_from_uint8(bq4_0->qs, iqs + i);
-        u[2*i+0] = get_int_from_int8_aligned(bq8_1->qs, iqs + i);
-        u[2*i+1] = get_int_from_int8_aligned(bq8_1->qs, iqs + i + QI4_0);
-    }
-
-    return vec_dot_q4_0_q8_1_impl<VDR_Q4_0_Q8_1_MMVQ>(v, u, bq4_0->d, bq8_1->ds);
-}
-
-template <int mmq_y> static __device__ __forceinline__ void allocate_tiles_q4_0(int ** x_ql, half2 ** x_dm, int ** x_qh, int ** x_sc) {
-    (void)x_qh; (void)x_sc;
-
-    __shared__ int  tile_x_qs[mmq_y * (WARP_SIZE)       + mmq_y];
-    __shared__ float tile_x_d[mmq_y * (WARP_SIZE/QI4_0) + mmq_y/QI4_0];
-
-    *x_ql = tile_x_qs;
-    *x_dm = (half2 *) tile_x_d;
-}
-
-template <int mmq_y, int nwarps, bool need_check> static __device__ __forceinline__ void load_tiles_q4_0(
-    const void * __restrict__ vx, int * __restrict__ x_ql, half2 * __restrict__ x_dm, int * __restrict__ x_qh,
-    int * __restrict__ x_sc, const int & i_offset, const int & i_max, const int & k, const int & blocks_per_row) {
-    (void)x_qh; (void)x_sc;
-    GGML_CUDA_ASSUME(i_offset >= 0);
-    GGML_CUDA_ASSUME(i_offset <  nwarps);
-    GGML_CUDA_ASSUME(k >= 0);
-    GGML_CUDA_ASSUME(k <  WARP_SIZE);
-
-    const int kbx  = k / QI4_0;
-    const int kqsx = k % QI4_0;
-
-    const block_q4_0 * bx0 = (const block_q4_0 *) vx;
-
-    float * x_dmf = (float *) x_dm;
-
-#pragma unroll
-    for (int i0 = 0; i0 < mmq_y; i0 += nwarps) {
-        int i = i0 + i_offset;
-
-        if (need_check) {
-            i = min(i, i_max);
-        }
-
-        const block_q4_0 * bxi = bx0 + i*blocks_per_row + kbx;
-
-        x_ql[i * (WARP_SIZE + 1) + k] = get_int_from_uint8(bxi->qs, kqsx);
-        // x_dmf[i * (WARP_SIZE/QI4_0) + i / QI4_0 + kbx] = bxi->d;
-    }
-
-    const int blocks_per_tile_x_row = WARP_SIZE / QI4_0;
-    const int kbxd = k % blocks_per_tile_x_row;
-
-#pragma unroll
-    for (int i0 = 0; i0 < mmq_y; i0 += nwarps * QI4_0) {
-        int i = i0 + i_offset * QI4_0 + k / blocks_per_tile_x_row;
-
-        if (need_check) {
-            i = min(i, i_max);
-        }
-
-        const block_q4_0 * bxi = bx0 + i*blocks_per_row + kbxd;
-
-        x_dmf[i * (WARP_SIZE/QI4_0) + i / QI4_0 + kbxd] = bxi->d;
-    }
-}
-
-static __device__ __forceinline__ float vec_dot_q4_0_q8_1_mul_mat(
-    const int * __restrict__ x_ql, const half2 * __restrict__ x_dm, const int * __restrict__ x_qh, const int * __restrict__ x_sc,
-    const int * __restrict__ y_qs, const half2 * __restrict__ y_ds, const int & i, const int & j, const int & k) {
-    (void)x_qh; (void)x_sc;
-
-    const int kyqs = k % (QI8_1/2) + QI8_1 * (k / (QI8_1/2));
-    const float * x_dmf = (const float *) x_dm;
-
-    int u[2*VDR_Q4_0_Q8_1_MMQ];
-
-#pragma unroll
-    for (int l = 0; l < VDR_Q4_0_Q8_1_MMQ; ++l) {
-        u[2*l+0] = y_qs[j * WARP_SIZE + (kyqs + l)         % WARP_SIZE];
-        u[2*l+1] = y_qs[j * WARP_SIZE + (kyqs + l + QI4_0) % WARP_SIZE];
-    }
-
-    return vec_dot_q4_0_q8_1_impl<VDR_Q4_0_Q8_1_MMQ>
-        (&x_ql[i * (WARP_SIZE + 1) + k], u, x_dmf[i * (WARP_SIZE/QI4_0) + i/QI4_0 + k/QI4_0],
-         y_ds[j * (WARP_SIZE/QI8_1) + (2*k/QI8_1) % (WARP_SIZE/QI8_1)]);
-}
-
-static __device__ __forceinline__ float vec_dot_q4_1_q8_1(
-    const void * __restrict__ vbq, const block_q8_1 * __restrict__ bq8_1, const int & iqs) {
-
-    const block_q4_1 * bq4_1 = (const block_q4_1 *) vbq;
-
-    int v[VDR_Q4_1_Q8_1_MMVQ];
-    int u[2*VDR_Q4_1_Q8_1_MMVQ];
-
-#pragma unroll
-    for (int i = 0; i < VDR_Q4_1_Q8_1_MMVQ; ++i) {
-        v[i]    = get_int_from_uint8_aligned(bq4_1->qs, iqs + i);
-        u[2*i+0] = get_int_from_int8_aligned(bq8_1->qs, iqs + i);
-        u[2*i+1] = get_int_from_int8_aligned(bq8_1->qs, iqs + i + QI4_1);
-    }
-
-    return vec_dot_q4_1_q8_1_impl<VDR_Q4_1_Q8_1_MMVQ>(v, u, bq4_1->dm, bq8_1->ds);
-}
-
-template <int mmq_y> static __device__ __forceinline__ void allocate_tiles_q4_1(int ** x_ql, half2 ** x_dm, int ** x_qh, int ** x_sc) {
-    (void)x_qh; (void)x_sc;
-
-    __shared__ int   tile_x_qs[mmq_y * (WARP_SIZE) +     + mmq_y];
-    __shared__ half2 tile_x_dm[mmq_y * (WARP_SIZE/QI4_1) + mmq_y/QI4_1];
-
-    *x_ql = tile_x_qs;
-    *x_dm = tile_x_dm;
-}
-
-template <int mmq_y, int nwarps, bool need_check> static __device__ __forceinline__ void load_tiles_q4_1(
-    const void * __restrict__ vx, int * __restrict__ x_ql, half2 * __restrict__ x_dm, int * __restrict__ x_qh,
-    int * __restrict__ x_sc, const int & i_offset, const int & i_max, const int & k, const int & blocks_per_row) {
-    (void)x_qh; (void)x_sc;
-
-    GGML_CUDA_ASSUME(i_offset >= 0);
-    GGML_CUDA_ASSUME(i_offset <  nwarps);
-    GGML_CUDA_ASSUME(k >= 0);
-    GGML_CUDA_ASSUME(k <  WARP_SIZE);
-
-    const int kbx  = k / QI4_1;
-    const int kqsx = k % QI4_1;
-
-    const block_q4_1 * bx0 = (const block_q4_1 *) vx;
-
-#pragma unroll
-    for (int i0 = 0; i0 < mmq_y; i0 += nwarps) {
-        int i = i0 + i_offset;
-
-        if (need_check) {
-            i = min(i, i_max);
-        }
-
-        const block_q4_1 * bxi = bx0 + i*blocks_per_row + kbx;
-
-        x_ql[i * (WARP_SIZE + 1) + k] = get_int_from_uint8_aligned(bxi->qs, kqsx);
-    }
-
-    const int blocks_per_tile_x_row = WARP_SIZE / QI4_1;
-    const int kbxd = k % blocks_per_tile_x_row;
-
-#pragma unroll
-    for (int i0 = 0; i0 < mmq_y; i0 += nwarps * QI4_1) {
-        int i = i0 + i_offset * QI4_1 + k / blocks_per_tile_x_row;
-
-        if (need_check) {
-            i = min(i, i_max);
-        }
-
-        const block_q4_1 * bxi = bx0 + i*blocks_per_row + kbxd;
-
-        x_dm[i * (WARP_SIZE/QI4_1) + i / QI4_1 + kbxd] = bxi->dm;
-    }
-}
-
-static __device__ __forceinline__ float vec_dot_q4_1_q8_1_mul_mat(
-    const int * __restrict__ x_ql, const half2 * __restrict__ x_dm, const int * __restrict__ x_qh, const int * __restrict__ x_sc,
-    const int * __restrict__ y_qs, const half2 * __restrict__ y_ds, const int & i, const int & j, const int & k) {
-    (void)x_qh; (void)x_sc;
-
-    const int kyqs = k % (QI8_1/2) + QI8_1 * (k / (QI8_1/2));
-
-    int u[2*VDR_Q4_1_Q8_1_MMQ];
-
-#pragma unroll
-    for (int l = 0; l < VDR_Q4_1_Q8_1_MMQ; ++l) {
-        u[2*l+0] = y_qs[j * WARP_SIZE + (kyqs + l)         % WARP_SIZE];
-        u[2*l+1] = y_qs[j * WARP_SIZE + (kyqs + l + QI4_1) % WARP_SIZE];
-    }
-
-    return vec_dot_q4_1_q8_1_impl<VDR_Q4_1_Q8_1_MMQ>
-        (&x_ql[i * (WARP_SIZE + 1) + k], u, x_dm[i * (WARP_SIZE/QI4_1) + i/QI4_1 + k/QI4_1],
-         y_ds[j * (WARP_SIZE/QI8_1) + (2*k/QI8_1) % (WARP_SIZE/QI8_1)]);
-}
-
-static __device__ __forceinline__ float vec_dot_q5_0_q8_1(
-    const void * __restrict__ vbq, const block_q8_1 * __restrict__ bq8_1, const int & iqs) {
-
-    const block_q5_0 * bq5_0 = (const block_q5_0 *) vbq;
-
-    int vl[VDR_Q5_0_Q8_1_MMVQ];
-    int vh[VDR_Q5_0_Q8_1_MMVQ];
-    int  u[2*VDR_Q5_0_Q8_1_MMVQ];
-
-#pragma unroll
-    for (int i = 0; i < VDR_Q5_0_Q8_1_MMVQ; ++i) {
-        vl[i]    = get_int_from_uint8(bq5_0->qs, iqs + i);
-        vh[i]    = get_int_from_uint8(bq5_0->qh, 0) >> (4 * (iqs + i));
-        u[2*i+0] = get_int_from_int8_aligned(bq8_1->qs, iqs + i);
-        u[2*i+1] = get_int_from_int8_aligned(bq8_1->qs, iqs + i + QI5_0);
-    }
-
-    return vec_dot_q5_0_q8_1_impl<VDR_Q5_0_Q8_1_MMVQ>(vl, vh, u, bq5_0->d, bq8_1->ds);
-}
-
-template <int mmq_y> static __device__ __forceinline__ void allocate_tiles_q5_0(int ** x_ql, half2 ** x_dm, int ** x_qh, int ** x_sc) {
-    (void)x_qh; (void)x_sc;
-
-    __shared__ int  tile_x_ql[mmq_y * (2*WARP_SIZE)     + mmq_y];
-    __shared__ float tile_x_d[mmq_y * (WARP_SIZE/QI5_0) + mmq_y/QI5_0];
-
-    *x_ql = tile_x_ql;
-    *x_dm = (half2 *) tile_x_d;
-}
-
-template <int mmq_y, int nwarps, bool need_check> static __device__ __forceinline__ void load_tiles_q5_0(
-    const void * __restrict__ vx, int * __restrict__ x_ql, half2 * __restrict__ x_dm, int * __restrict__ x_qh,
-    int * __restrict__ x_sc, const int & i_offset, const int & i_max, const int & k, const int & blocks_per_row) {
-    (void)x_qh; (void)x_sc;
-
-    GGML_CUDA_ASSUME(i_offset >= 0);
-    GGML_CUDA_ASSUME(i_offset <  nwarps);
-    GGML_CUDA_ASSUME(k >= 0);
-    GGML_CUDA_ASSUME(k <  WARP_SIZE);
-
-    const int kbx  = k / QI5_0;
-    const int kqsx = k % QI5_0;
-
-    const block_q5_0 * bx0 = (const block_q5_0 *) vx;
-
-#pragma unroll
-    for (int i0 = 0; i0 < mmq_y; i0 += nwarps) {
-        int i = i0 + i_offset;
-
-        if (need_check) {
-            i = min(i, i_max);
-        }
-
-        const block_q5_0 * bxi = bx0 + i*blocks_per_row + kbx;
-
-        const int ql = get_int_from_uint8(bxi->qs, kqsx);
-        const int qh = get_int_from_uint8(bxi->qh, 0) >> (4 * (k % QI5_0));
-
-        int qs0 = (ql >>  0)   & 0x0F0F0F0F;
-        qs0    |= (qh <<  4)   & 0x00000010;  // 0 ->  4
-        qs0    |= (qh << 11)   & 0x00001000;  // 1 -> 12
-        qs0    |= (qh << 18)   & 0x00100000;  // 2 -> 20
-        qs0    |= (qh << 25)   & 0x10000000;  // 3 -> 28
-        qs0     = __vsubss4(qs0, 0x10101010); // subtract 16
-
-        x_ql[i * (2*WARP_SIZE + 1) + 2*k+0] = qs0;
-
-        int qs1 = (ql >>  4)   & 0x0F0F0F0F;
-        qs1    |= (qh >> 12)   & 0x00000010;  // 16 ->  4
-        qs1    |= (qh >>  5)   & 0x00001000;  // 17 -> 12
-        qs1    |= (qh <<  2)   & 0x00100000;  // 18 -> 20
-        qs1    |= (qh <<  9)   & 0x10000000;  // 19 -> 28
-        qs1     = __vsubss4(qs1, 0x10101010); // subtract 16
-
-        x_ql[i * (2*WARP_SIZE + 1) + 2*k+1] = qs1;
-    }
-
-    const int blocks_per_tile_x_row = WARP_SIZE / QI5_0;
-    const int kbxd = k % blocks_per_tile_x_row;
-    float * x_dmf = (float *) x_dm;
-
-#pragma unroll
-    for (int i0 = 0; i0 < mmq_y; i0 += nwarps * QI5_0) {
-        int i = i0 + i_offset * QI5_0 + k / blocks_per_tile_x_row;
-
-        if (need_check) {
-            i = min(i, i_max);
-        }
-
-        const block_q5_0 * bxi = bx0 + i*blocks_per_row + kbxd;
-
-        x_dmf[i * (WARP_SIZE/QI5_0) + i / QI5_0 + kbxd] = bxi->d;
-    }
-}
-
-static __device__ __forceinline__ float vec_dot_q5_0_q8_1_mul_mat(
-    const int * __restrict__ x_ql, const half2 * __restrict__ x_dm, const int * __restrict__ x_qh, const int * __restrict__ x_sc,
-    const int * __restrict__ y_qs, const half2 * __restrict__ y_ds, const int & i, const int & j, const int & k) {
-    (void)x_qh; (void)x_sc;
-
-    const int kyqs = k % (QI8_1/2) + QI8_1 * (k / (QI8_1/2));
-    const int index_bx = i * (WARP_SIZE/QI5_0) + i/QI5_0 + k/QI5_0;
-    const float * x_dmf = (const float *) x_dm;
-    const float * y_df  = (const float *) y_ds;
-
-    int u[2*VDR_Q5_0_Q8_1_MMQ];
-
-#pragma unroll
-    for (int l = 0; l < VDR_Q5_0_Q8_1_MMQ; ++l) {
-        u[2*l+0] = y_qs[j * WARP_SIZE + (kyqs + l)         % WARP_SIZE];
-        u[2*l+1] = y_qs[j * WARP_SIZE + (kyqs + l + QI5_0) % WARP_SIZE];
-    }
-
-    return vec_dot_q8_0_q8_1_impl<QR5_0*VDR_Q5_0_Q8_1_MMQ>
-        (&x_ql[i * (2*WARP_SIZE + 1) + 2 * k], u, x_dmf[index_bx], y_df[j * (WARP_SIZE/QI8_1) + (2*k/QI8_1) % (WARP_SIZE/QI8_1)]);
-}
-
-static __device__ __forceinline__ float vec_dot_q5_1_q8_1(
-    const void * __restrict__ vbq, const block_q8_1 * __restrict__ bq8_1, const int & iqs) {
-
-    const block_q5_1 * bq5_1 = (const block_q5_1 *) vbq;
-
-    int vl[VDR_Q5_1_Q8_1_MMVQ];
-    int vh[VDR_Q5_1_Q8_1_MMVQ];
-    int  u[2*VDR_Q5_1_Q8_1_MMVQ];
-
-#pragma unroll
-    for (int i = 0; i < VDR_Q5_1_Q8_1_MMVQ; ++i) {
-        vl[i]   = get_int_from_uint8_aligned(bq5_1->qs, iqs + i);
-        vh[i]   = get_int_from_uint8_aligned(bq5_1->qh, 0) >> (4 * (iqs + i));
-        u[2*i+0] = get_int_from_int8_aligned(bq8_1->qs, iqs + i);
-        u[2*i+1] = get_int_from_int8_aligned(bq8_1->qs, iqs + i + QI5_1);
-    }
-
-    return vec_dot_q5_1_q8_1_impl<VDR_Q5_1_Q8_1_MMVQ>(vl, vh, u, bq5_1->dm, bq8_1->ds);
-}
-
-template <int mmq_y> static __device__ __forceinline__ void allocate_tiles_q5_1(int ** x_ql, half2 ** x_dm, int ** x_qh, int ** x_sc) {
-    (void)x_qh; (void)x_sc;
-
-    __shared__ int   tile_x_ql[mmq_y * (2*WARP_SIZE)     + mmq_y];
-    __shared__ half2 tile_x_dm[mmq_y * (WARP_SIZE/QI5_1) + mmq_y/QI5_1];
-
-    *x_ql = tile_x_ql;
-    *x_dm = tile_x_dm;
-}
-
-template <int mmq_y, int nwarps, bool need_check> static __device__ __forceinline__ void load_tiles_q5_1(
-    const void * __restrict__ vx, int * __restrict__ x_ql, half2 * __restrict__ x_dm, int * __restrict__ x_qh,
-    int * __restrict__ x_sc, const int & i_offset, const int & i_max, const int & k, const int & blocks_per_row) {
-    (void)x_qh; (void)x_sc;
-
-    GGML_CUDA_ASSUME(i_offset >= 0);
-    GGML_CUDA_ASSUME(i_offset < nwarps);
-    GGML_CUDA_ASSUME(k >= 0);
-    GGML_CUDA_ASSUME(k <  WARP_SIZE);
-
-    const int kbx  = k / QI5_1;
-    const int kqsx = k % QI5_1;
-
-    const block_q5_1 * bx0 = (const block_q5_1 *) vx;
-
-#pragma unroll
-    for (int i0 = 0; i0 < mmq_y; i0 += nwarps) {
-        int i = i0 + i_offset;
-
-        if (need_check) {
-            i = min(i, i_max);
-        }
-
-        const block_q5_1 * bxi = bx0 + i*blocks_per_row + kbx;
-
-        const int ql = get_int_from_uint8_aligned(bxi->qs, kqsx);
-        const int qh = get_int_from_uint8_aligned(bxi->qh, 0) >> (4 * (k % QI5_1));
-
-        int qs0 = (ql >>  0) & 0x0F0F0F0F;
-        qs0    |= (qh <<  4) & 0x00000010; // 0 ->  4
-        qs0    |= (qh << 11) & 0x00001000; // 1 -> 12
-        qs0    |= (qh << 18) & 0x00100000; // 2 -> 20
-        qs0    |= (qh << 25) & 0x10000000; // 3 -> 28
-
-        x_ql[i * (2*WARP_SIZE + 1) + 2*k+0] = qs0;
-
-        int qs1 = (ql >>  4) & 0x0F0F0F0F;
-        qs1    |= (qh >> 12) & 0x00000010; // 16 ->  4
-        qs1    |= (qh >>  5) & 0x00001000; // 17 -> 12
-        qs1    |= (qh <<  2) & 0x00100000; // 18 -> 20
-        qs1    |= (qh <<  9) & 0x10000000; // 19 -> 28
-
-        x_ql[i * (2*WARP_SIZE + 1) + 2*k+1] = qs1;
-    }
-
-    const int blocks_per_tile_x_row = WARP_SIZE / QI5_1;
-    const int kbxd = k % blocks_per_tile_x_row;
-
-#pragma unroll
-    for (int i0 = 0; i0 < mmq_y; i0 += nwarps * QI5_1) {
-        int i = i0 + i_offset * QI5_1 + k / blocks_per_tile_x_row;
-
-        if (need_check) {
-            i = min(i, i_max);
-        }
-
-        const block_q5_1 * bxi = bx0 + i*blocks_per_row + kbxd;
-
-        x_dm[i * (WARP_SIZE/QI5_1) + i / QI5_1 + kbxd] = bxi->dm;
-    }
-}
-
-static __device__ __forceinline__ float vec_dot_q5_1_q8_1_mul_mat(
-    const int * __restrict__ x_ql, const half2 * __restrict__ x_dm, const int * __restrict__ x_qh, const int * __restrict__ x_sc,
-    const int * __restrict__ y_qs, const half2 * __restrict__ y_ds, const int & i, const int & j, const int & k) {
-    (void)x_qh; (void)x_sc;
-
-    const int kyqs = k % (QI8_1/2) + QI8_1 * (k / (QI8_1/2));
-    const int index_bx = i * (WARP_SIZE/QI5_1) + + i/QI5_1 + k/QI5_1;
-
-    int u[2*VDR_Q5_1_Q8_1_MMQ];
-
-#pragma unroll
-    for (int l = 0; l < VDR_Q5_1_Q8_1_MMQ; ++l) {
-        u[2*l+0] = y_qs[j * WARP_SIZE + (kyqs + l)         % WARP_SIZE];
-        u[2*l+1] = y_qs[j * WARP_SIZE + (kyqs + l + QI5_1) % WARP_SIZE];
-    }
-
-    return vec_dot_q8_1_q8_1_impl<QR5_1*VDR_Q5_1_Q8_1_MMQ>
-        (&x_ql[i * (2*WARP_SIZE + 1) + 2 * k], u, x_dm[index_bx], y_ds[j * (WARP_SIZE/QI8_1) + (2*k/QI8_1) % (WARP_SIZE/QI8_1)]);
-}
-
-static __device__ __forceinline__ float vec_dot_q8_0_q8_1(
-    const void * __restrict__ vbq, const block_q8_1 * __restrict__ bq8_1, const int & iqs) {
-
-    const block_q8_0 * bq8_0 = (const block_q8_0 *) vbq;
-
-    int v[VDR_Q8_0_Q8_1_MMVQ];
-    int u[VDR_Q8_0_Q8_1_MMVQ];
-
-#pragma unroll
-    for (int i = 0; i < VDR_Q8_0_Q8_1_MMVQ; ++i) {
-        v[i] = get_int_from_int8(bq8_0->qs, iqs + i);
-        u[i] = get_int_from_int8_aligned(bq8_1->qs, iqs + i);
-    }
-
-    return vec_dot_q8_0_q8_1_impl<VDR_Q8_0_Q8_1_MMVQ>(v, u, bq8_0->d, __low2half(bq8_1->ds));
-}
-
-template <int mmq_y> static __device__ __forceinline__ void allocate_tiles_q8_0(int ** x_ql, half2 ** x_dm, int ** x_qh, int ** x_sc) {
-    (void)x_qh; (void)x_sc;
-
-    __shared__ int  tile_x_qs[mmq_y * (WARP_SIZE)       + mmq_y];
-    __shared__ float tile_x_d[mmq_y * (WARP_SIZE/QI8_0) + mmq_y/QI8_0];
-
-    *x_ql = tile_x_qs;
-    *x_dm = (half2 *) tile_x_d;
-}
-
-template <int mmq_y, int nwarps, bool need_check> static __device__ __forceinline__ void load_tiles_q8_0(
-    const void * __restrict__ vx, int * __restrict__ x_ql, half2 * __restrict__ x_dm, int * __restrict__ x_qh,
-    int * __restrict__ x_sc, const int & i_offset, const int & i_max, const int & k, const int & blocks_per_row) {
-    (void)x_qh; (void)x_sc;
-
-    GGML_CUDA_ASSUME(i_offset >= 0);
-    GGML_CUDA_ASSUME(i_offset <  nwarps);
-    GGML_CUDA_ASSUME(k >= 0);
-    GGML_CUDA_ASSUME(k <  WARP_SIZE);
-
-    const int kbx  = k / QI8_0;
-    const int kqsx = k % QI8_0;
-    float * x_dmf = (float *) x_dm;
-
-    const block_q8_0 * bx0 = (const block_q8_0 *) vx;
-
-#pragma unroll
-    for (int i0 = 0; i0 < mmq_y; i0 += nwarps) {
-        int i = i0 + i_offset;
-
-        if (need_check) {
-            i = min(i, i_max);
-        }
-
-        const block_q8_0 * bxi = bx0 + i*blocks_per_row + kbx;
-
-        x_ql[i * (WARP_SIZE + 1) + k] = get_int_from_int8(bxi->qs, kqsx);
-    }
-
-    const int blocks_per_tile_x_row = WARP_SIZE / QI8_0;
-    const int kbxd = k % blocks_per_tile_x_row;
-
-#pragma unroll
-    for (int i0 = 0; i0 < mmq_y; i0 += nwarps * QI8_0) {
-        int i = i0 + i_offset * QI8_0 + k / blocks_per_tile_x_row;
-
-        if (need_check) {
-            i = min(i, i_max);
-        }
-
-        const block_q8_0 * bxi = bx0 + i*blocks_per_row + kbxd;
-
-        x_dmf[i * (WARP_SIZE/QI8_0) + i / QI8_0 + kbxd] = bxi->d;
-    }
-}
-
-static __device__ __forceinline__ float vec_dot_q8_0_q8_1_mul_mat(
-    const int * __restrict__ x_ql, const half2 * __restrict__ x_dm, const int * __restrict__ x_qh, const int * __restrict__ x_sc,
-    const int * __restrict__ y_qs, const half2 * __restrict__ y_ds, const int & i, const int & j, const int & k) {
-    (void)x_qh; (void)x_sc;
-
-    const float * x_dmf = (const float *) x_dm;
-    const float * y_df  = (const float *) y_ds;
-
-    return vec_dot_q8_0_q8_1_impl<VDR_Q8_0_Q8_1_MMQ>
-        (&x_ql[i * (WARP_SIZE + 1) + k], &y_qs[j * WARP_SIZE + k], x_dmf[i * (WARP_SIZE/QI8_0) + i/QI8_0 + k/QI8_0],
-         y_df[j * (WARP_SIZE/QI8_1) + k/QI8_1]);
-}
-
-static __device__ __forceinline__ float vec_dot_q2_K_q8_1(
-    const void * __restrict__ vbq, const block_q8_1 * __restrict__ bq8_1, const int & iqs) {
-
-    const block_q2_K * bq2_K = (const block_q2_K *) vbq;
-
-    const int bq8_offset = QR2_K * (iqs / QI8_1);
-    const int scale_offset = iqs - iqs % QI8_1 + (iqs % QI8_1) / (QI8_1/2);
-
-    const uint8_t * scales = bq2_K->scales + scale_offset;
-
-    const int v = get_int_from_uint8_aligned(bq2_K->qs, iqs);
-    int    u[QR2_K];
-    float d8[QR2_K];
-
-#pragma unroll
-    for (int i = 0; i < QR2_K; ++ i) {
-        u[i]  = get_int_from_int8_aligned(bq8_1[bq8_offset + i].qs, iqs % QI8_1);
-        d8[i] = __low2half(bq8_1[bq8_offset + i].ds);
-    }
-
-    return vec_dot_q2_K_q8_1_impl_mmvq(v, u, scales, bq2_K->dm, d8);
-}
-
-template <int mmq_y> static __device__ __forceinline__ void allocate_tiles_q2_K(int ** x_ql, half2 ** x_dm, int ** x_qh, int ** x_sc) {
-    (void)x_qh;
-
-    __shared__ int   tile_x_ql[mmq_y * (WARP_SIZE)       + mmq_y];
-    __shared__ half2 tile_x_dm[mmq_y * (WARP_SIZE/QI2_K) + mmq_y/QI2_K];
-    __shared__ int   tile_x_sc[mmq_y * (WARP_SIZE/4)     + mmq_y/4];
-
-    *x_ql = tile_x_ql;
-    *x_dm = tile_x_dm;
-    *x_sc = tile_x_sc;
-}
-
-template <int mmq_y, int nwarps, bool need_check> static __device__ __forceinline__ void load_tiles_q2_K(
-    const void * __restrict__ vx, int * __restrict__ x_ql, half2 * __restrict__ x_dm, int * __restrict__ x_qh,
-    int * __restrict__ x_sc, const int & i_offset, const int & i_max, const int & k, const int & blocks_per_row) {
-    (void)x_qh;
-
-    GGML_CUDA_ASSUME(i_offset >= 0);
-    GGML_CUDA_ASSUME(i_offset <  nwarps);
-    GGML_CUDA_ASSUME(k >= 0);
-    GGML_CUDA_ASSUME(k <  WARP_SIZE);
-
-    const int kbx  = k / QI2_K;
-    const int kqsx = k % QI2_K;
-
-    const block_q2_K * bx0 = (const block_q2_K *) vx;
-
-#pragma unroll
-    for (int i0 = 0; i0 < mmq_y; i0 += nwarps) {
-        int i = i0 + i_offset;
-
-        if (need_check) {
-            i = min(i, i_max);
-        }
-
-        const block_q2_K * bxi = bx0 + i*blocks_per_row + kbx;
-
-        x_ql[i * (WARP_SIZE + 1) + k] = get_int_from_uint8_aligned(bxi->qs, kqsx);
-    }
-
-    const int blocks_per_tile_x_row = WARP_SIZE / QI2_K;
-    const int kbxd = k % blocks_per_tile_x_row;
-
-#pragma unroll
-    for (int i0 = 0; i0 < mmq_y; i0 += nwarps * QI2_K) {
-        int i = (i0 + i_offset * QI2_K + k / blocks_per_tile_x_row) % mmq_y;
-
-        if (need_check) {
-            i = min(i, i_max);
-        }
-
-        const block_q2_K * bxi = bx0 + i*blocks_per_row + kbxd;
-
-        x_dm[i * (WARP_SIZE/QI2_K) + i / QI2_K + kbxd] = bxi->dm;
-    }
-
-#pragma unroll
-    for (int i0 = 0; i0 < mmq_y; i0 += nwarps * 4) {
-        int i = i0 + i_offset * 4 + k / (WARP_SIZE/4);
-
-        if (need_check) {
-            i = min(i, i_max);
-        }
-
-        const block_q2_K * bxi = bx0 + i*blocks_per_row + (k % (WARP_SIZE/4)) / (QI2_K/4);
-
-        x_sc[i * (WARP_SIZE/4) + i / 4 + k % (WARP_SIZE/4)] = get_int_from_uint8_aligned(bxi->scales, k % (QI2_K/4));
-    }
-}
-
-static __device__ __forceinline__ float vec_dot_q2_K_q8_1_mul_mat(
-    const int * __restrict__ x_ql, const half2 * __restrict__ x_dm, const int * __restrict__ x_qh, const int * __restrict__ x_sc,
-    const int * __restrict__ y_qs, const half2 * __restrict__ y_ds, const int & i, const int & j, const int & k) {
-    (void)x_qh;
-
-    const int kbx = k / QI2_K;
-    const int ky  = (k % QI2_K) * QR2_K;
-    const float * y_df = (const float *) y_ds;
-
-    int v[QR2_K*VDR_Q2_K_Q8_1_MMQ];
-
-    const int kqsx = i * (WARP_SIZE + 1) + kbx*QI2_K + (QI2_K/2) * (ky/(2*QI2_K)) + ky % (QI2_K/2);
-    const int shift = 2 * ((ky % (2*QI2_K)) / (QI2_K/2));
-
-#pragma unroll
-    for (int l = 0; l < QR2_K*VDR_Q2_K_Q8_1_MMQ; ++l) {
-        v[l] = (x_ql[kqsx + l] >> shift) & 0x03030303;
-    }
-
-    const uint8_t * scales = ((const uint8_t *) &x_sc[i * (WARP_SIZE/4) + i/4 + kbx*4]) + ky/4;
-
-    const int index_y = j * WARP_SIZE + (QR2_K*k) % WARP_SIZE;
-    return vec_dot_q2_K_q8_1_impl_mmq(v, &y_qs[index_y], scales, x_dm[i * (WARP_SIZE/QI2_K) + i/QI2_K + kbx], y_df[index_y/QI8_1]);
-}
-
-static __device__ __forceinline__ float vec_dot_q3_K_q8_1(
-    const void * __restrict__ vbq, const block_q8_1 * __restrict__ bq8_1, const int & iqs) {
-
-    const block_q3_K * bq3_K = (const block_q3_K *) vbq;
-
-    const int bq8_offset = QR3_K * (iqs / (QI3_K/2));
-    const int scale_offset = iqs - iqs % QI8_1 + (iqs % QI8_1) / (QI8_1/2);
-
-    const float d = bq3_K->d;
-
-    const int vl = get_int_from_uint8(bq3_K->qs, iqs);
-
-    // invert the mask with ~ so that a 0/1 results in 4/0 being subtracted
-    const int vh = ~get_int_from_uint8(bq3_K->hmask, iqs % (QI3_K/2)) >> bq8_offset;
-
-    int    u[QR3_K];
-    float d8[QR3_K];
-
-#pragma unroll
-    for (int i = 0; i < QR3_K; ++i) {
-        u[i]  = get_int_from_int8_aligned(bq8_1[bq8_offset + i].qs, iqs % QI8_1);
-        d8[i] = __low2half(bq8_1[bq8_offset + i].ds);
-    }
-
-    return vec_dot_q3_K_q8_1_impl_mmvq(vl, vh, u, bq3_K->scales, scale_offset, d, d8);
-}
-
-template <int mmq_y> static __device__ __forceinline__ void allocate_tiles_q3_K(int ** x_ql, half2 ** x_dm, int ** x_qh, int ** x_sc) {
-
-    __shared__ int   tile_x_ql[mmq_y * (WARP_SIZE)       + mmq_y];
-    __shared__ half2 tile_x_dm[mmq_y * (WARP_SIZE/QI3_K) + mmq_y/QI3_K];
-    __shared__ int   tile_x_qh[mmq_y * (WARP_SIZE/2)     + mmq_y/2];
-    __shared__ int   tile_x_sc[mmq_y * (WARP_SIZE/4)     + mmq_y/4];
-
-    *x_ql = tile_x_ql;
-    *x_dm = tile_x_dm;
-    *x_qh = tile_x_qh;
-    *x_sc = tile_x_sc;
-}
-
-template <int mmq_y, int nwarps, bool need_check> static __device__ __forceinline__ void load_tiles_q3_K(
-    const void * __restrict__ vx, int * __restrict__ x_ql, half2 * __restrict__ x_dm, int * __restrict__ x_qh,
-    int * __restrict__ x_sc, const int & i_offset, const int & i_max, const int & k, const int & blocks_per_row) {
-
-    GGML_CUDA_ASSUME(i_offset >= 0);
-    GGML_CUDA_ASSUME(i_offset <  nwarps);
-    GGML_CUDA_ASSUME(k >= 0);
-    GGML_CUDA_ASSUME(k <  WARP_SIZE);
-
-    const int kbx  = k / QI3_K;
-    const int kqsx = k % QI3_K;
-
-    const block_q3_K * bx0 = (const block_q3_K *) vx;
-
-#pragma unroll
-    for (int i0 = 0; i0 < mmq_y; i0 += nwarps) {
-        int i = i0 + i_offset;
-
-        if (need_check) {
-            i = min(i, i_max);
-        }
-
-        const block_q3_K * bxi = bx0 + i*blocks_per_row + kbx;
-
-        x_ql[i * (WARP_SIZE + 1) + k] = get_int_from_uint8(bxi->qs, kqsx);
-    }
-
-    const int blocks_per_tile_x_row = WARP_SIZE / QI3_K;
-    const int kbxd = k % blocks_per_tile_x_row;
-    float * x_dmf = (float *) x_dm;
-
-#pragma unroll
-    for (int i0 = 0; i0 < mmq_y; i0 += nwarps * QI3_K) {
-        int i = (i0 + i_offset * QI3_K + k / blocks_per_tile_x_row) % mmq_y;
-
-        if (need_check) {
-            i = min(i, i_max);
-        }
-
-        const block_q3_K * bxi = bx0 + i*blocks_per_row + kbxd;
-
-        x_dmf[i * (WARP_SIZE/QI3_K) + i / QI3_K + kbxd] = bxi->d;
-    }
-
-#pragma unroll
-    for (int i0 = 0; i0 < mmq_y; i0 += nwarps * 2) {
-        int i = i0 + i_offset * 2 + k / (WARP_SIZE/2);
-
-        if (need_check) {
-            i = min(i, i_max);
-        }
-
-        const block_q3_K * bxi = bx0 + i*blocks_per_row + (k % (WARP_SIZE/2)) / (QI3_K/2);
-
-        // invert the mask with ~ so that a 0/1 results in 4/0 being subtracted
-        x_qh[i * (WARP_SIZE/2) + i / 2 + k % (WARP_SIZE/2)] = ~get_int_from_uint8(bxi->hmask, k % (QI3_K/2));
-    }
-
-#pragma unroll
-    for (int i0 = 0; i0 < mmq_y; i0 += nwarps * 4) {
-        int i = i0 + i_offset * 4 + k / (WARP_SIZE/4);
-
-        if (need_check) {
-            i = min(i, i_max);
-        }
-
-        const block_q3_K * bxi = bx0 + i*blocks_per_row + (k % (WARP_SIZE/4)) / (QI3_K/4);
-
-        const int ksc = k % (QI3_K/4);
-
-        const int ksc_low = ksc % (QI3_K/8);
-        const int shift_low = 4 * (ksc / (QI3_K/8));
-        const int sc_low = (get_int_from_uint8(bxi->scales, ksc_low) >> shift_low) & 0x0F0F0F0F;
-
-        const int ksc_high = QI3_K/8;
-        const int shift_high = 2 * ksc;
-        const int sc_high = ((get_int_from_uint8(bxi->scales, ksc_high) >> shift_high) << 4) & 0x30303030;
-
-        const int sc = __vsubss4(sc_low | sc_high, 0x20202020);
-
-        x_sc[i * (WARP_SIZE/4) + i / 4 + k % (WARP_SIZE/4)] = sc;
-    }
-}
-
-static __device__ __forceinline__ float vec_dot_q3_K_q8_1_mul_mat(
-    const int * __restrict__ x_ql, const half2 * __restrict__ x_dm, const int * __restrict__ x_qh, const int * __restrict__ x_sc,
-    const int * __restrict__ y_qs, const half2 * __restrict__ y_ds, const int & i, const int & j, const int & k) {
-
-    const int kbx  = k / QI3_K;
-    const int ky  = (k % QI3_K) * QR3_K;
-    const float * x_dmf = (const float *) x_dm;
-    const float * y_df  = (const float *) y_ds;
-
-    const int8_t * scales = ((const int8_t *) (x_sc + i * (WARP_SIZE/4) + i/4 + kbx*4)) + ky/4;
-
-    int v[QR3_K*VDR_Q3_K_Q8_1_MMQ];
-
-#pragma unroll
-    for (int l = 0; l < QR3_K*VDR_Q3_K_Q8_1_MMQ; ++l) {
-        const int kqsx = i * (WARP_SIZE + 1) + kbx*QI3_K + (QI3_K/2) * (ky/(2*QI3_K)) + ky % (QI3_K/2);
-        const int shift = 2 * ((ky % 32) / 8);
-        const int vll = (x_ql[kqsx + l] >> shift) & 0x03030303;
-
-        const int vh = x_qh[i * (WARP_SIZE/2) + i/2 + kbx * (QI3_K/2) + (ky+l)%8] >> ((ky+l) / 8);
-        const int vlh = (vh << 2) & 0x04040404;
-
-        v[l] = __vsubss4(vll, vlh);
-    }
-
-    const int index_y = j * WARP_SIZE + (k*QR3_K) % WARP_SIZE;
-    return vec_dot_q3_K_q8_1_impl_mmq(v, &y_qs[index_y], scales, x_dmf[i * (WARP_SIZE/QI3_K) + i/QI3_K + kbx], y_df[index_y/QI8_1]);
-}
-
-static __device__ __forceinline__ float vec_dot_q4_K_q8_1(
-    const void * __restrict__ vbq, const block_q8_1 * __restrict__ bq8_1, const int & iqs) {
-
-#ifndef GGML_QKK_64
-    const block_q4_K * bq4_K = (const block_q4_K *) vbq;
-
-    int    v[2];
-    int    u[2*QR4_K];
-    float d8[QR4_K];
-
-    // iqs is in 0,2..30. bq8_offset = iqs/4 -> bq8_offset = 0, 2, 4, 6
-    const int bq8_offset = QR4_K * ((iqs/2) / (QI8_1/2));
-
-    // iqs = 0....3 -> bq8_offset = 0, want q4_offset = 0, 4, 8, 12
-    // iqs = 4....7 -> bq8_offset = 2, want q4_offset = 32, 36, 40, 44
-    // iqs = 8...11 -> bq8_offset = 4, want q4_offset = 64, 68, 72, 76
-    // iqs = 12..15 -> bq8_offset = 6, want q4_offset = 96, 100, 104, 108
-
-    const int * q4 = (const int *)(bq4_K->qs + 16 * bq8_offset + 4 * ((iqs/2)%4));
-    v[0] = q4[0];
-    v[1] = q4[4];
-
-    const uint16_t * scales = (const uint16_t *)bq4_K->scales;
-    uint16_t aux[2];
-    const int j = bq8_offset/2;
-    if (j < 2) {
-        aux[0] = scales[j+0] & 0x3f3f;
-        aux[1] = scales[j+2] & 0x3f3f;
-    } else {
-        aux[0] = ((scales[j+2] >> 0) & 0x0f0f) | ((scales[j-2] & 0xc0c0) >> 2);
-        aux[1] = ((scales[j+2] >> 4) & 0x0f0f) | ((scales[j-0] & 0xc0c0) >> 2);
-    }
-    const uint8_t * sc = (const uint8_t *)aux;
-    const uint8_t * m  = sc + 2;
-
-    for (int i = 0; i < QR4_K; ++i) {
-        const block_q8_1 * bq8i = bq8_1 + bq8_offset + i;
-        d8[i] = __low2half(bq8i->ds);
-
-        const int * q8 = (const int *)bq8i->qs + ((iqs/2)%4);
-        u[2*i+0] = q8[0];
-        u[2*i+1] = q8[4];
-    }
-
-    return vec_dot_q4_K_q8_1_impl_vmmq(v, u, sc, m, bq4_K->dm, d8);
-
-#else
-
-#if __CUDA_ARCH__ >= MIN_CC_DP4A // lowest compute capability for integer intrinsics
-    const block_q4_K * bq4_K = (const block_q4_K *) vbq;
-
-    float sumf_d = 0.0f;
-    float sumf_m = 0.0f;
-
-    uint16_t aux16[2];
-    const uint8_t * s = (const uint8_t *)aux16;
-
-    const uint16_t * a = (const uint16_t *)bq4_K->scales;
-    aux16[0] = a[0] & 0x0f0f;
-    aux16[1] = (a[0] >> 4) & 0x0f0f;
-
-    const float dall = bq4_K->dm[0];
-    const float dmin = bq4_K->dm[1];
-
-    const float d8_1 = __low2float(bq8_1[0].ds);
-    const float d8_2 = __low2float(bq8_1[1].ds);
-
-    const int ui1 = *((const int *)bq8_1[0].qs + (iqs/2));
-    const int ui2 = *((const int *)bq8_1[0].qs + (iqs/2) + 4);
-    const int ui3 = *((const int *)bq8_1[1].qs + (iqs/2));
-    const int ui4 = *((const int *)bq8_1[1].qs + (iqs/2) + 4);
-
-    const int * q4 = (const int *)bq4_K->qs + (iqs/2);
-    const int v1 = q4[0];
-    const int v2 = q4[4];
-
-    const int dot1 = __dp4a(ui2, v2 & 0x0f0f0f0f, __dp4a(ui1, v1 & 0x0f0f0f0f, 0));
-    const int dot2 = __dp4a(ui4, (v2 >> 4) & 0x0f0f0f0f, __dp4a(ui3, (v1 >> 4) & 0x0f0f0f0f, 0));
-    const int dot3 = __dp4a(0x01010101, ui2, __dp4a(0x01010101, ui1, 0));
-    const int dot4 = __dp4a(0x01010101, ui4, __dp4a(0x01010101, ui3, 0));
-
-    sumf_d += d8_1 * (dot1 * s[0]) + d8_2 * (dot2 * s[1]);
-    sumf_m += d8_1 * (dot3 * s[2]) + d8_2 * (dot4 * s[3]);
-
-    return dall * sumf_d - dmin * sumf_m;
-
-#else
-    bad_arch();
-#endif // __CUDA_ARCH__ >= MIN_CC_DP4A
-
-#endif
-}
-
-template <int mmq_y> static __device__ __forceinline__ void allocate_tiles_q4_K(int ** x_ql, half2 ** x_dm, int ** x_qh, int ** x_sc) {
-    (void)x_qh;
-
-    __shared__ int   tile_x_ql[mmq_y * (WARP_SIZE)       + mmq_y];
-    __shared__ half2 tile_x_dm[mmq_y * (WARP_SIZE/QI4_K) + mmq_y/QI4_K];
-    __shared__ int   tile_x_sc[mmq_y * (WARP_SIZE/8)     + mmq_y/8];
-
-    *x_ql = tile_x_ql;
-    *x_dm = tile_x_dm;
-    *x_sc = tile_x_sc;
-}
-
-template <int mmq_y, int nwarps, bool need_check> static __device__ __forceinline__ void load_tiles_q4_K(
-    const void * __restrict__ vx, int * __restrict__ x_ql, half2 * __restrict__ x_dm, int * __restrict__ x_qh,
-    int * __restrict__ x_sc, const int & i_offset, const int & i_max, const int & k, const int & blocks_per_row) {
-    (void)x_qh;
-
-    GGML_CUDA_ASSUME(i_offset >= 0);
-    GGML_CUDA_ASSUME(i_offset <  nwarps);
-    GGML_CUDA_ASSUME(k >= 0);
-    GGML_CUDA_ASSUME(k <  WARP_SIZE);
-
-    const int kbx  = k / QI4_K; // == 0 if QK_K == 256
-    const int kqsx = k % QI4_K; // == k if QK_K == 256
-
-    const block_q4_K * bx0 = (const block_q4_K *) vx;
-
-#pragma unroll
-    for (int i0 = 0; i0 < mmq_y; i0 += nwarps) {
-        int i = i0 + i_offset;
-
-        if (need_check) {
-            i = min(i, i_max);
-        }
-
-        const block_q4_K * bxi = bx0 + i*blocks_per_row + kbx;
-
-        x_ql[i * (WARP_SIZE + 1) + k] = get_int_from_uint8_aligned(bxi->qs, kqsx);
-    }
-
-    const int blocks_per_tile_x_row = WARP_SIZE / QI4_K; // == 1 if QK_K == 256
-    const int kbxd = k % blocks_per_tile_x_row;          // == 0 if QK_K == 256
-
-#pragma unroll
-    for (int i0 = 0; i0 < mmq_y; i0 += nwarps * QI4_K) {
-        int i = (i0 + i_offset * QI4_K + k / blocks_per_tile_x_row) % mmq_y;
-
-        if (need_check) {
-            i = min(i, i_max);
-        }
-
-        const block_q4_K * bxi = bx0 + i*blocks_per_row + kbxd;
-
-#if QK_K == 256
-        x_dm[i * (WARP_SIZE/QI4_K) + i / QI4_K + kbxd] = bxi->dm;
-#else
-        x_dm[i * (WARP_SIZE/QI4_K) + i / QI4_K + kbxd] = {bxi->dm[0], bxi->dm[1]};
-#endif
-    }
-
-#pragma unroll
-    for (int i0 = 0; i0 < mmq_y; i0 += nwarps * 8) {
-        int i = (i0 + i_offset * 8 + k / (WARP_SIZE/8)) % mmq_y;
-
-        if (need_check) {
-            i = min(i, i_max);
-        }
-
-        const block_q4_K * bxi = bx0 + i*blocks_per_row + (k % (WARP_SIZE/8)) / (QI4_K/8);
-
-        const int * scales = (const int *) bxi->scales;
-
-        const int ksc = k % (WARP_SIZE/8);
-
-        // scale arrangement after the following two lines: sc0,...,sc3, sc4,...,sc7, m0,...,m3, m4,...,m8
-        int scales8 = (scales[(ksc%2) + (ksc!=0)] >> (4 * (ksc & (ksc/2)))) & 0x0F0F0F0F; // lower 4 bits
-        scales8    |= (scales[ksc/2]              >> (2 * (ksc % 2)))       & 0x30303030; // upper 2 bits
-
-        x_sc[i * (WARP_SIZE/8) + i / 8 + ksc] = scales8;
-    }
-}
-
-static __device__ __forceinline__ float vec_dot_q4_K_q8_1_mul_mat(
-    const int * __restrict__ x_ql, const half2 * __restrict__ x_dm, const int * __restrict__ x_qh, const int * __restrict__ x_sc,
-    const int * __restrict__ y_qs, const half2 * __restrict__ y_ds, const int & i, const int & j, const int & k) {
-    (void)x_qh;
-
-    const uint8_t * sc = ((const uint8_t *) &x_sc[i * (WARP_SIZE/8) + i/8 + k/16]) + 2*((k % 16) / 8);
-
-    const int index_y = j * WARP_SIZE + (QR4_K*k) % WARP_SIZE;
-    return vec_dot_q4_K_q8_1_impl_mmq(&x_ql[i * (WARP_SIZE + 1) + k], &y_qs[index_y], sc, sc+8,
-                                      x_dm[i * (WARP_SIZE/QI4_K) + i/QI4_K], &y_ds[index_y/QI8_1]);
-}
-
-static __device__ __forceinline__ float vec_dot_q5_K_q8_1(
-    const void * __restrict__ vbq, const block_q8_1 * __restrict__ bq8_1, const int & iqs) {
-
-#ifndef GGML_QKK_64
-    const block_q5_K * bq5_K = (const block_q5_K *) vbq;
-
-    int   vl[2];
-    int   vh[2];
-    int    u[2*QR5_K];
-    float d8[QR5_K];
-
-    const int bq8_offset = QR5_K * ((iqs/2) / (QI8_1/2));
-    const int * ql = (const int *)(bq5_K->qs + 16 * bq8_offset + 4 * ((iqs/2)%4));
-    const int * qh = (const int *)(bq5_K->qh + 4 * ((iqs/2)%4));
-
-    vl[0] = ql[0];
-    vl[1] = ql[4];
-
-    vh[0] = qh[0] >> bq8_offset;
-    vh[1] = qh[4] >> bq8_offset;
-
-    const uint16_t * scales = (const uint16_t *)bq5_K->scales;
-    uint16_t aux[2];
-    const int j = bq8_offset/2;
-    if (j < 2) {
-        aux[0] = scales[j+0] & 0x3f3f;
-        aux[1] = scales[j+2] & 0x3f3f;
-    } else {
-        aux[0] = ((scales[j+2] >> 0) & 0x0f0f) | ((scales[j-2] & 0xc0c0) >> 2);
-        aux[1] = ((scales[j+2] >> 4) & 0x0f0f) | ((scales[j-0] & 0xc0c0) >> 2);
-    }
-    const uint8_t * sc = (const uint8_t *)aux;
-    const uint8_t * m  = sc + 2;
-
-#pragma unroll
-    for (int i = 0; i < QR5_K; ++i) {
-        const block_q8_1 * bq8i = bq8_1 + bq8_offset + i;
-        d8[i] = __low2float(bq8i->ds);
-
-        const int * q8 = (const int *)bq8i->qs + ((iqs/2)%4);
-        u[2*i+0] = q8[0];
-        u[2*i+1] = q8[4];
-    }
-
-    return vec_dot_q5_K_q8_1_impl_vmmq(vl, vh, u, sc, m, bq5_K->dm, d8);
-
-#else
-
-#if __CUDA_ARCH__ >= MIN_CC_DP4A // lowest compute capability for integer intrinsics
-    const block_q5_K * bq5_K = (const block_q5_K *) vbq;
-
-    const int8_t * s = bq5_K->scales;
-
-    const float d = bq5_K->d;
-
-    const float d8_1 = __low2half(bq8_1[0].ds);
-    const float d8_2 = __low2half(bq8_1[1].ds);
-
-    const int ui1 = *((const int *)bq8_1[0].qs + (iqs/2));
-    const int ui2 = *((const int *)bq8_1[0].qs + (iqs/2) + 4);
-    const int ui3 = *((const int *)bq8_1[1].qs + (iqs/2));
-    const int ui4 = *((const int *)bq8_1[1].qs + (iqs/2) + 4);
-
-    const int * ql = (const int *)bq5_K->qs + (iqs/2);
-    const int vl1 = ql[0];
-    const int vl2 = ql[4];
-
-    const int step = 4 * (iqs/2); // 0, 4, 8, 12
-    const int im = step/8; // = 0 for iqs = 0, 2, = 1 for iqs = 4, 6
-    const int in = step%8; // 0, 4, 0, 4
-    const int vh = (*((const int *)(bq5_K->qh + in))) >> im;
-
-    const int v1 = (((vh << 4) & 0x10101010) ^ 0x10101010) | ((vl1 >> 0) & 0x0f0f0f0f);
-    const int v2 = (((vh << 2) & 0x10101010) ^ 0x10101010) | ((vl2 >> 0) & 0x0f0f0f0f);
-    const int v3 = (((vh >> 0) & 0x10101010) ^ 0x10101010) | ((vl1 >> 4) & 0x0f0f0f0f);
-    const int v4 = (((vh >> 2) & 0x10101010) ^ 0x10101010) | ((vl2 >> 4) & 0x0f0f0f0f);
-
-    const float sumf_d = d8_1 * (__dp4a(ui1, v1, 0) * s[0] + __dp4a(ui2, v2, 0) * s[1])
-                       + d8_2 * (__dp4a(ui3, v3, 0) * s[2] + __dp4a(ui4, v4, 0) * s[3]);
-
-    return d * sumf_d;
-
-#else
-    bad_arch();
-#endif // __CUDA_ARCH__ >= MIN_CC_DP4A
-
-#endif
-}
-
-template <int mmq_y> static __device__ __forceinline__ void allocate_tiles_q5_K(int ** x_ql, half2 ** x_dm, int ** x_qh, int ** x_sc) {
-    (void)x_qh;
-
-    __shared__ int   tile_x_ql[mmq_y * (2*WARP_SIZE)     + mmq_y];
-    __shared__ half2 tile_x_dm[mmq_y * (WARP_SIZE/QI5_K) + mmq_y/QI5_K];
-    __shared__ int   tile_x_sc[mmq_y * (WARP_SIZE/8)     + mmq_y/8];
-
-    *x_ql = tile_x_ql;
-    *x_dm = tile_x_dm;
-    *x_sc = tile_x_sc;
-}
-
-template <int mmq_y, int nwarps, bool need_check> static __device__ __forceinline__ void load_tiles_q5_K(
-    const void * __restrict__ vx, int * __restrict__ x_ql, half2 * __restrict__ x_dm, int * __restrict__ x_qh,
-    int * __restrict__ x_sc, const int & i_offset, const int & i_max, const int & k, const int & blocks_per_row) {
-    (void)x_qh;
-
-    GGML_CUDA_ASSUME(i_offset >= 0);
-    GGML_CUDA_ASSUME(i_offset <  nwarps);
-    GGML_CUDA_ASSUME(k >= 0);
-    GGML_CUDA_ASSUME(k <  WARP_SIZE);
-
-    const int kbx  = k / QI5_K; // == 0 if QK_K == 256
-    const int kqsx = k % QI5_K; // == k if QK_K == 256
-
-    const block_q5_K * bx0 = (const block_q5_K *) vx;
-
-#pragma unroll
-    for (int i0 = 0; i0 < mmq_y; i0 += nwarps) {
-        int i = i0 + i_offset;
-
-        if (need_check) {
-            i = min(i, i_max);
-        }
-
-        const block_q5_K * bxi = bx0 + i*blocks_per_row + kbx;
-        const int ky = QR5_K*kqsx;
-
-        const int ql = get_int_from_uint8_aligned(bxi->qs, kqsx);
-        const int ql0 = (ql >> 0) & 0x0F0F0F0F;
-        const int ql1 = (ql >> 4) & 0x0F0F0F0F;
-
-        const int qh = get_int_from_uint8_aligned(bxi->qh, kqsx % (QI5_K/4));
-        const int qh0 = ((qh >> (2 * (kqsx / (QI5_K/4)) + 0)) << 4) & 0x10101010;
-        const int qh1 = ((qh >> (2 * (kqsx / (QI5_K/4)) + 1)) << 4) & 0x10101010;
-
-        const int kq0 = ky - ky % (QI5_K/2) + k % (QI5_K/4) + 0;
-        const int kq1 = ky - ky % (QI5_K/2) + k % (QI5_K/4) + (QI5_K/4);
-
-        x_ql[i * (2*WARP_SIZE + 1) + kq0] = ql0 | qh0;
-        x_ql[i * (2*WARP_SIZE + 1) + kq1] = ql1 | qh1;
-    }
-
-    const int blocks_per_tile_x_row = WARP_SIZE / QI5_K; // == 1 if QK_K == 256
-    const int kbxd = k % blocks_per_tile_x_row;          // == 0 if QK_K == 256
-
-#pragma unroll
-    for (int i0 = 0; i0 < mmq_y; i0 += nwarps * QI5_K) {
-        int i = (i0 + i_offset * QI5_K + k / blocks_per_tile_x_row) % mmq_y;
-
-        if (need_check) {
-            i = min(i, i_max);
-        }
-
-        const block_q5_K * bxi = bx0 + i*blocks_per_row + kbxd;
-
-#if QK_K == 256
-        x_dm[i * (WARP_SIZE/QI5_K) + i / QI5_K + kbxd] = bxi->dm;
-#endif
-    }
-
-#pragma unroll
-    for (int i0 = 0; i0 < mmq_y; i0 += nwarps * 8) {
-        int i = (i0 + i_offset * 8 + k / (WARP_SIZE/8)) % mmq_y;
-
-        if (need_check) {
-            i = min(i, i_max);
-        }
-
-        const block_q5_K * bxi = bx0 + i*blocks_per_row + (k % (WARP_SIZE/8)) / (QI5_K/8);
-
-        const int * scales = (const int *) bxi->scales;
-
-        const int ksc = k % (WARP_SIZE/8);
-
-        // scale arrangement after the following two lines: sc0,...,sc3, sc4,...,sc7, m0,...,m3, m4,...,m8
-        int scales8 = (scales[(ksc%2) + (ksc!=0)] >> (4 * (ksc & (ksc/2)))) & 0x0F0F0F0F; // lower 4 bits
-        scales8    |= (scales[ksc/2]              >> (2 * (ksc % 2)))       & 0x30303030; // upper 2 bits
-
-        x_sc[i * (WARP_SIZE/8) + i / 8 + ksc] = scales8;
-    }
-}
-
-static __device__ __forceinline__ float vec_dot_q5_K_q8_1_mul_mat(
-    const int * __restrict__ x_ql, const half2 * __restrict__ x_dm, const int * __restrict__ x_qh, const int * __restrict__ x_sc,
-    const int * __restrict__ y_qs, const half2 * __restrict__ y_ds, const int & i, const int & j, const int & k) {
-    (void)x_qh;
-
-    const uint8_t * sc = ((const uint8_t *) &x_sc[i * (WARP_SIZE/8) + i/8 + k/16]) + 2 * ((k % 16) / 8);
-
-    const int index_x = i * (QR5_K*WARP_SIZE + 1) +  QR5_K*k;
-    const int index_y = j * WARP_SIZE             + (QR5_K*k) % WARP_SIZE;
-    return vec_dot_q5_K_q8_1_impl_mmq(&x_ql[index_x], &y_qs[index_y], sc, sc+8,
-                                      x_dm[i * (WARP_SIZE/QI5_K) + i/QI5_K], &y_ds[index_y/QI8_1]);
-}
-
-static __device__ __forceinline__ float vec_dot_q6_K_q8_1(
-    const void * __restrict__ vbq, const block_q8_1 * __restrict__ bq8_1, const int & iqs) {
-
-    const block_q6_K * bq6_K = (const block_q6_K *) vbq;
-
-    const int bq8_offset = 2 * QR6_K * (iqs / (QI6_K/2)) + (iqs % (QI6_K/2)) / (QI6_K/4);
-    const int scale_offset = (QI6_K/4) * (iqs / (QI6_K/2)) + (iqs % (QI6_K/2)) / (QI6_K/8);
-    const int vh_shift = 2 * ((iqs % (QI6_K/2)) / (QI6_K/4));
-
-    const int vl = get_int_from_uint8(bq6_K->ql, iqs);
-    const int vh = get_int_from_uint8(bq6_K->qh, (QI6_K/4) * (iqs / (QI6_K/2)) + iqs % (QI6_K/4)) >> vh_shift;
-
-    const int8_t * scales = bq6_K->scales + scale_offset;
-
-    int    u[QR6_K];
-    float d8[QR6_K];
-
-#pragma unroll
-    for (int i = 0; i < QR6_K; ++i) {
-        u[i]  = get_int_from_int8_aligned(bq8_1[bq8_offset + 2*i].qs, iqs % QI8_1);
-        d8[i] = __low2half(bq8_1[bq8_offset + 2*i].ds);
-    }
-
-    return vec_dot_q6_K_q8_1_impl_mmvq(vl, vh, u, scales, bq6_K->d, d8);
-}
-
-template <int mmq_y> static __device__ __forceinline__ void allocate_tiles_q6_K(int ** x_ql, half2 ** x_dm, int ** x_qh, int ** x_sc) {
-    (void)x_qh;
-
-    __shared__ int   tile_x_ql[mmq_y * (2*WARP_SIZE)     + mmq_y];
-    __shared__ half2 tile_x_dm[mmq_y * (WARP_SIZE/QI6_K) + mmq_y/QI6_K];
-    __shared__ int   tile_x_sc[mmq_y * (WARP_SIZE/8)     + mmq_y/8];
-
-    *x_ql = tile_x_ql;
-    *x_dm = tile_x_dm;
-    *x_sc = tile_x_sc;
-}
-
-template <int mmq_y, int nwarps, bool need_check> static __device__ __forceinline__ void load_tiles_q6_K(
-    const void * __restrict__ vx, int * __restrict__ x_ql, half2 * __restrict__ x_dm, int * __restrict__ x_qh,
-    int * __restrict__ x_sc, const int & i_offset, const int & i_max, const int & k, const int & blocks_per_row) {
-    (void)x_qh;
-
-    GGML_CUDA_ASSUME(i_offset >= 0);
-    GGML_CUDA_ASSUME(i_offset <  nwarps);
-    GGML_CUDA_ASSUME(k >= 0);
-    GGML_CUDA_ASSUME(k <  WARP_SIZE);
-
-    const int kbx  = k / QI6_K; // == 0 if QK_K == 256
-    const int kqsx = k % QI6_K; // == k if QK_K == 256
-
-    const block_q6_K * bx0 = (const block_q6_K *) vx;
-
-#pragma unroll
-    for (int i0 = 0; i0 < mmq_y; i0 += nwarps) {
-        int i = i0 + i_offset;
-
-        if (need_check) {
-            i = min(i, i_max);
-        }
-
-        const block_q6_K * bxi = bx0 + i*blocks_per_row + kbx;
-        const int ky = QR6_K*kqsx;
-
-        const int ql = get_int_from_uint8(bxi->ql, kqsx);
-        const int ql0 = (ql >> 0) & 0x0F0F0F0F;
-        const int ql1 = (ql >> 4) & 0x0F0F0F0F;
-
-        const int qh = get_int_from_uint8(bxi->qh, (QI6_K/4) * (kqsx / (QI6_K/2)) + kqsx % (QI6_K/4));
-        const int qh0 = ((qh >> (2 * ((kqsx % (QI6_K/2)) / (QI6_K/4)))) << 4) & 0x30303030;
-        const int qh1 =  (qh >> (2 * ((kqsx % (QI6_K/2)) / (QI6_K/4))))       & 0x30303030;
-
-        const int kq0 = ky - ky % QI6_K + k % (QI6_K/2) + 0;
-        const int kq1 = ky - ky % QI6_K + k % (QI6_K/2) + (QI6_K/2);
-
-        x_ql[i * (2*WARP_SIZE + 1) + kq0] = __vsubss4(ql0 | qh0, 0x20202020);
-        x_ql[i * (2*WARP_SIZE + 1) + kq1] = __vsubss4(ql1 | qh1, 0x20202020);
-    }
-
-    const int blocks_per_tile_x_row = WARP_SIZE / QI6_K; // == 1 if QK_K == 256
-    const int kbxd = k % blocks_per_tile_x_row;          // == 0 if QK_K == 256
-    float * x_dmf = (float *) x_dm;
-
-#pragma unroll
-    for (int i0 = 0; i0 < mmq_y; i0 += nwarps * QI6_K) {
-        int i = (i0 + i_offset * QI6_K + k / blocks_per_tile_x_row) % mmq_y;
-
-        if (need_check) {
-            i = min(i, i_max);
-        }
-
-        const block_q6_K * bxi = bx0 + i*blocks_per_row + kbxd;
-
-        x_dmf[i * (WARP_SIZE/QI6_K) + i / QI6_K + kbxd] = bxi->d;
-    }
-
-#pragma unroll
-    for (int i0 = 0; i0 < mmq_y; i0 += nwarps * 8) {
-        int i = (i0 + i_offset * 8 + k / (WARP_SIZE/8)) % mmq_y;
-
-        if (need_check) {
-            i = min(i, i_max);
-        }
-
-        const block_q6_K * bxi = bx0 + i*blocks_per_row + (k % (WARP_SIZE/8)) / 4;
-
-        x_sc[i * (WARP_SIZE/8) + i / 8 + k % (WARP_SIZE/8)] = get_int_from_int8(bxi->scales, k % (QI6_K/8));
-    }
-}
-
-static __device__ __forceinline__ float vec_dot_q6_K_q8_1_mul_mat(
-    const int * __restrict__ x_ql, const half2 * __restrict__ x_dm, const int * __restrict__ x_qh, const int * __restrict__ x_sc,
-    const int * __restrict__ y_qs, const half2 * __restrict__ y_ds, const int & i, const int & j, const int & k) {
-    (void)x_qh;
-
-    const float * x_dmf = (const float *) x_dm;
-    const float * y_df  = (const float *) y_ds;
-
-    const int8_t * sc = ((const int8_t *) &x_sc[i * (WARP_SIZE/8) + i/8 + k/8]);
-
-    const int index_x = i * (QR6_K*WARP_SIZE + 1) +  QR6_K*k;
-    const int index_y = j * WARP_SIZE             + (QR6_K*k) % WARP_SIZE;
-    return vec_dot_q6_K_q8_1_impl_mmq(&x_ql[index_x], &y_qs[index_y], sc, x_dmf[i * (WARP_SIZE/QI6_K) + i/QI6_K], &y_df[index_y/QI8_1]);
-}
-
-template <int qk, int qr, int qi, bool need_sum, typename block_q_t, int mmq_x, int mmq_y, int nwarps,
-              allocate_tiles_cuda_t allocate_tiles, load_tiles_cuda_t load_tiles, int vdr, vec_dot_q_mul_mat_cuda_t vec_dot>
-static __device__ __forceinline__ void mul_mat_q(
-    const void * __restrict__ vx, const void * __restrict__ vy, float * __restrict__ dst,
-    const int ncols_x, const int nrows_x, const int ncols_y, const int nrows_y, const int nrows_dst) {
-
-    const block_q_t  * x = (const block_q_t  *) vx;
-    const block_q8_1 * y = (const block_q8_1 *) vy;
-
-    const int blocks_per_row_x = ncols_x / qk;
-    const int blocks_per_col_y = nrows_y / QK8_1;
-    const int blocks_per_warp = WARP_SIZE / qi;
-
-    const int & ncols_dst = ncols_y;
-
-    const int row_dst_0 = blockIdx.x*mmq_y;
-    const int & row_x_0 = row_dst_0;
-
-    const int col_dst_0 = blockIdx.y*mmq_x;
-    const int & col_y_0 = col_dst_0;
-
-    int   * tile_x_ql = nullptr;
-    half2 * tile_x_dm = nullptr;
-    int   * tile_x_qh = nullptr;
-    int   * tile_x_sc = nullptr;
-
-    allocate_tiles(&tile_x_ql, &tile_x_dm, &tile_x_qh, &tile_x_sc);
-
-    __shared__ int    tile_y_qs[mmq_x * WARP_SIZE];
-    __shared__ half2  tile_y_ds[mmq_x * WARP_SIZE/QI8_1];
-
-    float sum[mmq_y/WARP_SIZE][mmq_x/nwarps] = {{0.0f}};
-
-    for (int ib0 = 0; ib0 < blocks_per_row_x; ib0 += blocks_per_warp) {
-
-        load_tiles(x + row_x_0*blocks_per_row_x + ib0, tile_x_ql, tile_x_dm, tile_x_qh, tile_x_sc,
-                   threadIdx.y, nrows_x-row_x_0-1, threadIdx.x, blocks_per_row_x);
-
-#pragma unroll
-        for (int ir = 0; ir < qr; ++ir) {
-            const int kqs = ir*WARP_SIZE + threadIdx.x;
-            const int kbxd = kqs / QI8_1;
-
-#pragma unroll
-            for (int i = 0; i < mmq_x; i += nwarps) {
-                const int col_y_eff = min(col_y_0 + threadIdx.y + i, ncols_y-1); // to prevent out-of-bounds memory accesses
-
-                const block_q8_1 * by0 = &y[col_y_eff*blocks_per_col_y + ib0 * (qk/QK8_1) + kbxd];
-
-                const int index_y = (threadIdx.y + i) * WARP_SIZE + kqs % WARP_SIZE;
-                tile_y_qs[index_y] = get_int_from_int8_aligned(by0->qs, threadIdx.x % QI8_1);
-            }
-
-#pragma unroll
-            for (int ids0 = 0; ids0 < mmq_x; ids0 += nwarps * QI8_1) {
-                const int ids = (ids0 + threadIdx.y * QI8_1 + threadIdx.x / (WARP_SIZE/QI8_1)) % mmq_x;
-                const int kby = threadIdx.x % (WARP_SIZE/QI8_1);
-                const int col_y_eff = min(col_y_0 + ids, ncols_y-1);
-
-                // if the sum is not needed it's faster to transform the scale to f32 ahead of time
-                const half2 * dsi_src = &y[col_y_eff*blocks_per_col_y + ib0 * (qk/QK8_1) + ir*(WARP_SIZE/QI8_1) + kby].ds;
-                half2       * dsi_dst = &tile_y_ds[ids * (WARP_SIZE/QI8_1) + kby];
-                if (need_sum) {
-                    *dsi_dst = *dsi_src;
-                } else {
-                    float * dfi_dst = (float *) dsi_dst;
-                    *dfi_dst = __low2half(*dsi_src);
-                }
-            }
-
-            __syncthreads();
-
-// #pragma unroll // unrolling this loop causes too much register pressure
-            for (int k = ir*WARP_SIZE/qr; k < (ir+1)*WARP_SIZE/qr; k += vdr) {
-#pragma unroll
-                for (int j = 0; j < mmq_x; j += nwarps) {
-#pragma unroll
-                    for (int i = 0; i < mmq_y; i += WARP_SIZE) {
-                        sum[i/WARP_SIZE][j/nwarps] += vec_dot(
-                            tile_x_ql, tile_x_dm, tile_x_qh, tile_x_sc, tile_y_qs, tile_y_ds,
-                            threadIdx.x + i, threadIdx.y + j, k);
-                    }
-                }
-            }
-
-            __syncthreads();
-        }
-    }
-
-#pragma unroll
-    for (int j = 0; j < mmq_x; j += nwarps) {
-        const int col_dst = col_dst_0 + j + threadIdx.y;
-
-        if (col_dst >= ncols_dst) {
-            return;
-        }
-
-#pragma unroll
-        for (int i = 0; i < mmq_y; i += WARP_SIZE) {
-            const int row_dst = row_dst_0 + threadIdx.x + i;
-
-            if (row_dst >= nrows_dst) {
-                continue;
-            }
-
-            dst[col_dst*nrows_dst + row_dst] = sum[i/WARP_SIZE][j/nwarps];
-        }
-    }
-}
-
-#define  MMQ_X_Q4_0_RDNA2  64
-#define  MMQ_Y_Q4_0_RDNA2  128
-#define NWARPS_Q4_0_RDNA2  8
-#define  MMQ_X_Q4_0_RDNA1  64
-#define  MMQ_Y_Q4_0_RDNA1  64
-#define NWARPS_Q4_0_RDNA1  8
-#if defined(CUDA_USE_TENSOR_CORES)
-#define  MMQ_X_Q4_0_AMPERE 4
-#define  MMQ_Y_Q4_0_AMPERE 32
-#define NWARPS_Q4_0_AMPERE 4
-#else
-#define  MMQ_X_Q4_0_AMPERE 64
-#define  MMQ_Y_Q4_0_AMPERE 128
-#define NWARPS_Q4_0_AMPERE 4
-#endif
-#define  MMQ_X_Q4_0_PASCAL 64
-#define  MMQ_Y_Q4_0_PASCAL 64
-#define NWARPS_Q4_0_PASCAL 8
-
-template <bool need_check> static __global__ void
-#if defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)
-#if defined(RDNA3) || defined(RDNA2)
-    __launch_bounds__(WARP_SIZE*NWARPS_Q4_0_RDNA2, 2)
-#endif // defined(RDNA3) || defined(RDNA2)
-#endif // defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)
-    mul_mat_q4_0(
-    const void * __restrict__ vx, const void * __restrict__ vy, float * __restrict__ dst,
-    const int ncols_x, const int nrows_x, const int ncols_y, const int nrows_y, const int nrows_dst) {
-
-#if defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)
-#if defined(RDNA3) || defined(RDNA2)
-    const int mmq_x  =  MMQ_X_Q4_0_RDNA2;
-    const int mmq_y  =  MMQ_Y_Q4_0_RDNA2;
-    const int nwarps = NWARPS_Q4_0_RDNA2;
-#else
-    const int mmq_x  =  MMQ_X_Q4_0_RDNA1;
-    const int mmq_y  =  MMQ_Y_Q4_0_RDNA1;
-    const int nwarps = NWARPS_Q4_0_RDNA1;
-#endif // defined(RDNA3) || defined(RDNA2)
-
-    mul_mat_q<QK4_0, QR4_0, QI4_0, true, block_q4_0, mmq_x, mmq_y, nwarps, allocate_tiles_q4_0<mmq_y>,
-        load_tiles_q4_0<mmq_y, nwarps, need_check>, VDR_Q4_0_Q8_1_MMQ, vec_dot_q4_0_q8_1_mul_mat>
-        (vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst);
-
-#elif __CUDA_ARCH__ >= CC_VOLTA
-    const int mmq_x  =  MMQ_X_Q4_0_AMPERE;
-    const int mmq_y  =  MMQ_Y_Q4_0_AMPERE;
-    const int nwarps = NWARPS_Q4_0_AMPERE;
-
-    mul_mat_q<QK4_0, QR4_0, QI4_0, true, block_q4_0, mmq_x, mmq_y, nwarps, allocate_tiles_q4_0<mmq_y>,
-        load_tiles_q4_0<mmq_y, nwarps, need_check>, VDR_Q4_0_Q8_1_MMQ, vec_dot_q4_0_q8_1_mul_mat>
-        (vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst);
-
-#elif __CUDA_ARCH__ >= MIN_CC_DP4A
-    const int mmq_x  =  MMQ_X_Q4_0_PASCAL;
-    const int mmq_y  =  MMQ_Y_Q4_0_PASCAL;
-    const int nwarps = NWARPS_Q4_0_PASCAL;
-
-    mul_mat_q<QK4_0, QR4_0, QI4_0, true, block_q4_0, mmq_x, mmq_y, nwarps, allocate_tiles_q4_0<mmq_y>,
-        load_tiles_q4_0<mmq_y, nwarps, need_check>, VDR_Q4_0_Q8_1_MMQ, vec_dot_q4_0_q8_1_mul_mat>
-        (vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst);
-#else
-    (void) vec_dot_q4_0_q8_1_mul_mat;
-    bad_arch();
-#endif // __CUDA_ARCH__ >= CC_VOLTA
-}
-
-#define  MMQ_X_Q4_1_RDNA2  64
-#define  MMQ_Y_Q4_1_RDNA2  128
-#define NWARPS_Q4_1_RDNA2  8
-#define  MMQ_X_Q4_1_RDNA1  64
-#define  MMQ_Y_Q4_1_RDNA1  64
-#define NWARPS_Q4_1_RDNA1  8
-#if defined(CUDA_USE_TENSOR_CORES)
-#define  MMQ_X_Q4_1_AMPERE 4
-#define  MMQ_Y_Q4_1_AMPERE 32
-#define NWARPS_Q4_1_AMPERE 4
-#else
-#define  MMQ_X_Q4_1_AMPERE 64
-#define  MMQ_Y_Q4_1_AMPERE 128
-#define NWARPS_Q4_1_AMPERE 4
-#endif
-#define  MMQ_X_Q4_1_PASCAL 64
-#define  MMQ_Y_Q4_1_PASCAL 64
-#define NWARPS_Q4_1_PASCAL 8
-
-template <bool need_check> static __global__ void
-#if defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)
-#if defined(RDNA3) || defined(RDNA2)
-    __launch_bounds__(WARP_SIZE*NWARPS_Q4_1_RDNA2, 2)
-#endif // defined(RDNA3) || defined(RDNA2)
-#elif __CUDA_ARCH__ < CC_VOLTA
-    __launch_bounds__(WARP_SIZE*NWARPS_Q4_1_PASCAL, 2)
-#endif // __CUDA_ARCH__ < CC_VOLTA
-    mul_mat_q4_1(
-    const void * __restrict__ vx, const void * __restrict__ vy, float * __restrict__ dst,
-    const int ncols_x, const int nrows_x, const int ncols_y, const int nrows_y, const int nrows_dst) {
-
-#if defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)
-#if defined(RDNA3) || defined(RDNA2)
-    const int mmq_x  =  MMQ_X_Q4_1_RDNA2;
-    const int mmq_y  =  MMQ_Y_Q4_1_RDNA2;
-    const int nwarps = NWARPS_Q4_1_RDNA2;
-#else
-    const int mmq_x  =  MMQ_X_Q4_1_RDNA1;
-    const int mmq_y  =  MMQ_Y_Q4_1_RDNA1;
-    const int nwarps = NWARPS_Q4_1_RDNA1;
-#endif // defined(RDNA3) || defined(RDNA2)
-
-    mul_mat_q<QK4_1, QR4_1, QI4_1, true, block_q4_1, mmq_x, mmq_y, nwarps, allocate_tiles_q4_1<mmq_y>,
-        load_tiles_q4_1<mmq_y, nwarps, need_check>, VDR_Q4_1_Q8_1_MMQ, vec_dot_q4_1_q8_1_mul_mat>
-        (vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst);
-
-#elif __CUDA_ARCH__ >= CC_VOLTA
-    const int mmq_x  =  MMQ_X_Q4_1_AMPERE;
-    const int mmq_y  =  MMQ_Y_Q4_1_AMPERE;
-    const int nwarps = NWARPS_Q4_1_AMPERE;
-
-    mul_mat_q<QK4_1, QR4_1, QI4_1, true, block_q4_1, mmq_x, mmq_y, nwarps, allocate_tiles_q4_1<mmq_y>,
-        load_tiles_q4_1<mmq_y, nwarps, need_check>, VDR_Q4_1_Q8_1_MMQ, vec_dot_q4_1_q8_1_mul_mat>
-        (vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst);
-
-#elif __CUDA_ARCH__ >= MIN_CC_DP4A
-    const int mmq_x  =  MMQ_X_Q4_1_PASCAL;
-    const int mmq_y  =  MMQ_Y_Q4_1_PASCAL;
-    const int nwarps = NWARPS_Q4_1_PASCAL;
-
-    mul_mat_q<QK4_1, QR4_1, QI4_1, true, block_q4_1, mmq_x, mmq_y, nwarps, allocate_tiles_q4_1<mmq_y>,
-        load_tiles_q4_1<mmq_y, nwarps, need_check>, VDR_Q4_1_Q8_1_MMQ, vec_dot_q4_1_q8_1_mul_mat>
-        (vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst);
-#else
-    (void) vec_dot_q4_1_q8_1_mul_mat;
-    bad_arch();
-#endif // __CUDA_ARCH__ >= CC_VOLTA
-}
-
-#define  MMQ_X_Q5_0_RDNA2  64
-#define  MMQ_Y_Q5_0_RDNA2  128
-#define NWARPS_Q5_0_RDNA2  8
-#define  MMQ_X_Q5_0_RDNA1  64
-#define  MMQ_Y_Q5_0_RDNA1  64
-#define NWARPS_Q5_0_RDNA1  8
-#if defined(CUDA_USE_TENSOR_CORES)
-#define  MMQ_X_Q5_0_AMPERE 4
-#define  MMQ_Y_Q5_0_AMPERE 32
-#define NWARPS_Q5_0_AMPERE 4
-#else
-#define  MMQ_X_Q5_0_AMPERE 128
-#define  MMQ_Y_Q5_0_AMPERE 64
-#define NWARPS_Q5_0_AMPERE 4
-#endif
-#define  MMQ_X_Q5_0_PASCAL 64
-#define  MMQ_Y_Q5_0_PASCAL 64
-#define NWARPS_Q5_0_PASCAL 8
-
-template <bool need_check> static __global__ void
-#if defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)
-#if defined(RDNA3) || defined(RDNA2)
-    __launch_bounds__(WARP_SIZE*NWARPS_Q5_0_RDNA2, 2)
-#endif // defined(RDNA3) || defined(RDNA2)
-#endif // defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)
-    mul_mat_q5_0(
-    const void * __restrict__ vx, const void * __restrict__ vy, float * __restrict__ dst,
-    const int ncols_x, const int nrows_x, const int ncols_y, const int nrows_y, const int nrows_dst) {
-
-#if defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)
-#if defined(RDNA3) || defined(RDNA2)
-    const int mmq_x  =  MMQ_X_Q5_0_RDNA2;
-    const int mmq_y  =  MMQ_Y_Q5_0_RDNA2;
-    const int nwarps = NWARPS_Q5_0_RDNA2;
-#else
-    const int mmq_x  =  MMQ_X_Q5_0_RDNA1;
-    const int mmq_y  =  MMQ_Y_Q5_0_RDNA1;
-    const int nwarps = NWARPS_Q5_0_RDNA1;
-#endif // defined(RDNA3) || defined(RDNA2)
-
-    mul_mat_q<QK5_0, QR5_0, QI5_0, false, block_q5_0, mmq_x, mmq_y, nwarps, allocate_tiles_q5_0<mmq_y>,
-        load_tiles_q5_0<mmq_y, nwarps, need_check>, VDR_Q5_0_Q8_1_MMQ, vec_dot_q5_0_q8_1_mul_mat>
-        (vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst);
-
-#elif __CUDA_ARCH__ >= CC_VOLTA
-    const int mmq_x  =  MMQ_X_Q5_0_AMPERE;
-    const int mmq_y  =  MMQ_Y_Q5_0_AMPERE;
-    const int nwarps = NWARPS_Q5_0_AMPERE;
-
-    mul_mat_q<QK5_0, QR5_0, QI5_0, false, block_q5_0, mmq_x, mmq_y, nwarps, allocate_tiles_q5_0<mmq_y>,
-        load_tiles_q5_0<mmq_y, nwarps, need_check>, VDR_Q5_0_Q8_1_MMQ, vec_dot_q5_0_q8_1_mul_mat>
-        (vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst);
-
-#elif __CUDA_ARCH__ >= MIN_CC_DP4A
-    const int mmq_x  =  MMQ_X_Q5_0_PASCAL;
-    const int mmq_y  =  MMQ_Y_Q5_0_PASCAL;
-    const int nwarps = NWARPS_Q5_0_PASCAL;
-
-    mul_mat_q<QK5_0, QR5_0, QI5_0, false, block_q5_0, mmq_x, mmq_y, nwarps, allocate_tiles_q5_0<mmq_y>,
-        load_tiles_q5_0<mmq_y, nwarps, need_check>, VDR_Q5_0_Q8_1_MMQ, vec_dot_q5_0_q8_1_mul_mat>
-        (vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst);
-#else
-    (void) vec_dot_q5_0_q8_1_mul_mat;
-    bad_arch();
-#endif // __CUDA_ARCH__ >= CC_VOLTA
-}
-
-#define  MMQ_X_Q5_1_RDNA2  64
-#define  MMQ_Y_Q5_1_RDNA2  128
-#define NWARPS_Q5_1_RDNA2  8
-#define  MMQ_X_Q5_1_RDNA1  64
-#define  MMQ_Y_Q5_1_RDNA1  64
-#define NWARPS_Q5_1_RDNA1  8
-#if defined(CUDA_USE_TENSOR_CORES)
-#define  MMQ_X_Q5_1_AMPERE 4
-#define  MMQ_Y_Q5_1_AMPERE 32
-#define NWARPS_Q5_1_AMPERE 4
-#else
-#define  MMQ_X_Q5_1_AMPERE 128
-#define  MMQ_Y_Q5_1_AMPERE 64
-#define NWARPS_Q5_1_AMPERE 4
-#endif
-#define  MMQ_X_Q5_1_PASCAL 64
-#define  MMQ_Y_Q5_1_PASCAL 64
-#define NWARPS_Q5_1_PASCAL 8
-
-template <bool need_check> static __global__ void
-#if defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)
-#if defined(RDNA3) || defined(RDNA2)
-    __launch_bounds__(WARP_SIZE*NWARPS_Q5_1_RDNA2, 2)
-#endif // defined(RDNA3) || defined(RDNA2)
-#endif // defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)
-mul_mat_q5_1(
-    const void * __restrict__ vx, const void * __restrict__ vy, float * __restrict__ dst,
-    const int ncols_x, const int nrows_x, const int ncols_y, const int nrows_y, const int nrows_dst) {
-
-#if defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)
-#if defined(RDNA3) || defined(RDNA2)
-    const int mmq_x  =  MMQ_X_Q5_1_RDNA2;
-    const int mmq_y  =  MMQ_Y_Q5_1_RDNA2;
-    const int nwarps = NWARPS_Q5_1_RDNA2;
-#else
-    const int mmq_x  =  MMQ_X_Q5_1_RDNA1;
-    const int mmq_y  =  MMQ_Y_Q5_1_RDNA1;
-    const int nwarps = NWARPS_Q5_1_RDNA1;
-#endif // defined(RDNA3) || defined(RDNA2)
-
-    mul_mat_q<QK5_1, QR5_1, QI5_1, true, block_q5_1, mmq_x, mmq_y, nwarps, allocate_tiles_q5_1<mmq_y>,
-        load_tiles_q5_1<mmq_y, nwarps, need_check>, VDR_Q5_1_Q8_1_MMQ, vec_dot_q5_1_q8_1_mul_mat>
-        (vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst);
-
-#elif __CUDA_ARCH__ >= CC_VOLTA
-    const int mmq_x  =  MMQ_X_Q5_1_AMPERE;
-    const int mmq_y  =  MMQ_Y_Q5_1_AMPERE;
-    const int nwarps = NWARPS_Q5_1_AMPERE;
-
-    mul_mat_q<QK5_1, QR5_1, QI5_1, true, block_q5_1, mmq_x, mmq_y, nwarps, allocate_tiles_q5_1<mmq_y>,
-        load_tiles_q5_1<mmq_y, nwarps, need_check>, VDR_Q5_1_Q8_1_MMQ, vec_dot_q5_1_q8_1_mul_mat>
-        (vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst);
-
-#elif __CUDA_ARCH__ >= MIN_CC_DP4A
-    const int mmq_x  =  MMQ_X_Q5_1_PASCAL;
-    const int mmq_y  =  MMQ_Y_Q5_1_PASCAL;
-    const int nwarps = NWARPS_Q5_1_PASCAL;
-
-    mul_mat_q<QK5_1, QR5_1, QI5_1, true, block_q5_1, mmq_x, mmq_y, nwarps, allocate_tiles_q5_1<mmq_y>,
-        load_tiles_q5_1<mmq_y, nwarps, need_check>, VDR_Q5_1_Q8_1_MMQ, vec_dot_q5_1_q8_1_mul_mat>
-        (vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst);
-#else
-    (void) vec_dot_q5_1_q8_1_mul_mat;
-    bad_arch();
-#endif // __CUDA_ARCH__ >= CC_VOLTA
-}
-
-#define  MMQ_X_Q8_0_RDNA2  64
-#define  MMQ_Y_Q8_0_RDNA2  128
-#define NWARPS_Q8_0_RDNA2  8
-#define  MMQ_X_Q8_0_RDNA1  64
-#define  MMQ_Y_Q8_0_RDNA1  64
-#define NWARPS_Q8_0_RDNA1  8
-#if defined(CUDA_USE_TENSOR_CORES)
-#define  MMQ_X_Q8_0_AMPERE 4
-#define  MMQ_Y_Q8_0_AMPERE 32
-#define NWARPS_Q8_0_AMPERE 4
-#else
-#define  MMQ_X_Q8_0_AMPERE 128
-#define  MMQ_Y_Q8_0_AMPERE 64
-#define NWARPS_Q8_0_AMPERE 4
-#endif
-#define  MMQ_X_Q8_0_PASCAL 64
-#define  MMQ_Y_Q8_0_PASCAL 64
-#define NWARPS_Q8_0_PASCAL 8
-
-template <bool need_check> static __global__ void
-#if defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)
-#if defined(RDNA3) || defined(RDNA2)
-    __launch_bounds__(WARP_SIZE*NWARPS_Q8_0_RDNA2, 2)
-#endif // defined(RDNA3) || defined(RDNA2)
-#endif // defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)
-    mul_mat_q8_0(
-    const void * __restrict__ vx, const void * __restrict__ vy, float * __restrict__ dst,
-    const int ncols_x, const int nrows_x, const int ncols_y, const int nrows_y, const int nrows_dst) {
-
-#if defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)
-#if defined(RDNA3) || defined(RDNA2)
-    const int mmq_x  =  MMQ_X_Q8_0_RDNA2;
-    const int mmq_y  =  MMQ_Y_Q8_0_RDNA2;
-    const int nwarps = NWARPS_Q8_0_RDNA2;
-#else
-    const int mmq_x  =  MMQ_X_Q8_0_RDNA1;
-    const int mmq_y  =  MMQ_Y_Q8_0_RDNA1;
-    const int nwarps = NWARPS_Q8_0_RDNA1;
-#endif // defined(RDNA3) || defined(RDNA2)
-
-    mul_mat_q<QK8_0, QR8_0, QI8_0, false, block_q8_0, mmq_x, mmq_y, nwarps, allocate_tiles_q8_0<mmq_y>,
-        load_tiles_q8_0<mmq_y, nwarps, need_check>, VDR_Q8_0_Q8_1_MMQ, vec_dot_q8_0_q8_1_mul_mat>
-        (vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst);
-
-#elif __CUDA_ARCH__ >= CC_VOLTA
-    const int mmq_x  =  MMQ_X_Q8_0_AMPERE;
-    const int mmq_y  =  MMQ_Y_Q8_0_AMPERE;
-    const int nwarps = NWARPS_Q8_0_AMPERE;
-
-    mul_mat_q<QK8_0, QR8_0, QI8_0, false, block_q8_0, mmq_x, mmq_y, nwarps, allocate_tiles_q8_0<mmq_y>,
-        load_tiles_q8_0<mmq_y, nwarps, need_check>, VDR_Q8_0_Q8_1_MMQ, vec_dot_q8_0_q8_1_mul_mat>
-        (vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst);
-
-#elif __CUDA_ARCH__ >= MIN_CC_DP4A
-    const int mmq_x  =  MMQ_X_Q8_0_PASCAL;
-    const int mmq_y  =  MMQ_Y_Q8_0_PASCAL;
-    const int nwarps = NWARPS_Q8_0_PASCAL;
-
-    mul_mat_q<QK8_0, QR8_0, QI8_0, false, block_q8_0, mmq_x, mmq_y, nwarps, allocate_tiles_q8_0<mmq_y>,
-        load_tiles_q8_0<mmq_y, nwarps, need_check>, VDR_Q8_0_Q8_1_MMQ, vec_dot_q8_0_q8_1_mul_mat>
-        (vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst);
-#else
-    (void) vec_dot_q8_0_q8_1_mul_mat;
-    bad_arch();
-#endif // __CUDA_ARCH__ >= CC_VOLTA
-}
-
-#define  MMQ_X_Q2_K_RDNA2  64
-#define  MMQ_Y_Q2_K_RDNA2  128
-#define NWARPS_Q2_K_RDNA2  8
-#define  MMQ_X_Q2_K_RDNA1  128
-#define  MMQ_Y_Q2_K_RDNA1  32
-#define NWARPS_Q2_K_RDNA1  8
-#if defined(CUDA_USE_TENSOR_CORES)
-#define  MMQ_X_Q2_K_AMPERE 4
-#define  MMQ_Y_Q2_K_AMPERE 32
-#define NWARPS_Q2_K_AMPERE 4
-#else
-#define  MMQ_X_Q2_K_AMPERE 64
-#define  MMQ_Y_Q2_K_AMPERE 128
-#define NWARPS_Q2_K_AMPERE 4
-#endif
-#define  MMQ_X_Q2_K_PASCAL 64
-#define  MMQ_Y_Q2_K_PASCAL 64
-#define NWARPS_Q2_K_PASCAL 8
-
-template <bool need_check> static __global__ void
-#if defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)
-#if defined(RDNA3) || defined(RDNA2)
-    __launch_bounds__(WARP_SIZE*NWARPS_Q2_K_RDNA2, 2)
-#endif // defined(RDNA3) || defined(RDNA2)
-#endif // defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)
-mul_mat_q2_K(
-    const void * __restrict__ vx, const void * __restrict__ vy, float * __restrict__ dst,
-    const int ncols_x, const int nrows_x, const int ncols_y, const int nrows_y, const int nrows_dst) {
-
-#if defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)
-#if defined(RDNA3) || defined(RDNA2)
-    const int mmq_x  =  MMQ_X_Q2_K_RDNA2;
-    const int mmq_y  =  MMQ_Y_Q2_K_RDNA2;
-    const int nwarps = NWARPS_Q2_K_RDNA2;
-#else
-    const int mmq_x  =  MMQ_X_Q2_K_RDNA1;
-    const int mmq_y  =  MMQ_Y_Q2_K_RDNA1;
-    const int nwarps = NWARPS_Q2_K_RDNA1;
-#endif // defined(RDNA3) || defined(RDNA2)
-
-    mul_mat_q<QK_K, QR2_K, QI2_K, false, block_q2_K, mmq_x, mmq_y, nwarps, allocate_tiles_q2_K<mmq_y>,
-        load_tiles_q2_K<mmq_y, nwarps, need_check>, VDR_Q2_K_Q8_1_MMQ, vec_dot_q2_K_q8_1_mul_mat>
-        (vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst);
-
-#elif __CUDA_ARCH__ >= CC_VOLTA
-    const int mmq_x  =  MMQ_X_Q2_K_AMPERE;
-    const int mmq_y  =  MMQ_Y_Q2_K_AMPERE;
-    const int nwarps = NWARPS_Q2_K_AMPERE;
-
-    mul_mat_q<QK_K, QR2_K, QI2_K, false, block_q2_K, mmq_x, mmq_y, nwarps, allocate_tiles_q2_K<mmq_y>,
-        load_tiles_q2_K<mmq_y, nwarps, need_check>, VDR_Q2_K_Q8_1_MMQ, vec_dot_q2_K_q8_1_mul_mat>
-        (vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst);
-
-#elif __CUDA_ARCH__ >= MIN_CC_DP4A
-    const int mmq_x  =  MMQ_X_Q2_K_PASCAL;
-    const int mmq_y  =  MMQ_Y_Q2_K_PASCAL;
-    const int nwarps = NWARPS_Q2_K_PASCAL;
-
-    mul_mat_q<QK_K, QR2_K, QI2_K, false, block_q2_K, mmq_x, mmq_y, nwarps, allocate_tiles_q2_K<mmq_y>,
-        load_tiles_q2_K<mmq_y, nwarps, need_check>, VDR_Q2_K_Q8_1_MMQ, vec_dot_q2_K_q8_1_mul_mat>
-        (vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst);
-#else
-    (void) vec_dot_q2_K_q8_1_mul_mat;
-    bad_arch();
-#endif // __CUDA_ARCH__ >= CC_VOLTA
-}
-
-#define  MMQ_X_Q3_K_RDNA2  128
-#define  MMQ_Y_Q3_K_RDNA2  64
-#define NWARPS_Q3_K_RDNA2  8
-#define  MMQ_X_Q3_K_RDNA1  32
-#define  MMQ_Y_Q3_K_RDNA1  128
-#define NWARPS_Q3_K_RDNA1  8
-#if defined(CUDA_USE_TENSOR_CORES)
-#define  MMQ_X_Q3_K_AMPERE 4
-#define  MMQ_Y_Q3_K_AMPERE 32
-#define NWARPS_Q3_K_AMPERE 4
-#else
-#define  MMQ_X_Q3_K_AMPERE 128
-#define  MMQ_Y_Q3_K_AMPERE 128
-#define NWARPS_Q3_K_AMPERE 4
-#endif
-#define  MMQ_X_Q3_K_PASCAL 64
-#define  MMQ_Y_Q3_K_PASCAL 64
-#define NWARPS_Q3_K_PASCAL 8
-
-template <bool need_check> static __global__ void
-#if defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)
-#if defined(RDNA3) || defined(RDNA2)
-    __launch_bounds__(WARP_SIZE*NWARPS_Q3_K_RDNA2, 2)
-#endif // defined(RDNA3) || defined(RDNA2)
-#elif __CUDA_ARCH__ < CC_VOLTA
-    __launch_bounds__(WARP_SIZE*NWARPS_Q3_K_PASCAL, 2)
-#endif // __CUDA_ARCH__ < CC_VOLTA
-    mul_mat_q3_K(
-    const void * __restrict__ vx, const void * __restrict__ vy, float * __restrict__ dst,
-    const int ncols_x, const int nrows_x, const int ncols_y, const int nrows_y, const int nrows_dst) {
-
-#if defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)
-#if defined(RDNA3) || defined(RDNA2)
-    const int mmq_x  =  MMQ_X_Q3_K_RDNA2;
-    const int mmq_y  =  MMQ_Y_Q3_K_RDNA2;
-    const int nwarps = NWARPS_Q3_K_RDNA2;
-#else
-    const int mmq_x  =  MMQ_X_Q3_K_RDNA1;
-    const int mmq_y  =  MMQ_Y_Q3_K_RDNA1;
-    const int nwarps = NWARPS_Q3_K_RDNA1;
-#endif // defined(RDNA3) || defined(RDNA2)
-
-    mul_mat_q<QK_K, QR3_K, QI3_K, false, block_q3_K, mmq_x, mmq_y, nwarps, allocate_tiles_q3_K<mmq_y>,
-        load_tiles_q3_K<mmq_y, nwarps, need_check>, VDR_Q3_K_Q8_1_MMQ, vec_dot_q3_K_q8_1_mul_mat>
-        (vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst);
-
-#elif __CUDA_ARCH__ >= CC_VOLTA
-    const int mmq_x  =  MMQ_X_Q3_K_AMPERE;
-    const int mmq_y  =  MMQ_Y_Q3_K_AMPERE;
-    const int nwarps = NWARPS_Q3_K_AMPERE;
-
-    mul_mat_q<QK_K, QR3_K, QI3_K, false, block_q3_K, mmq_x, mmq_y, nwarps, allocate_tiles_q3_K<mmq_y>,
-        load_tiles_q3_K<mmq_y, nwarps, need_check>, VDR_Q3_K_Q8_1_MMQ, vec_dot_q3_K_q8_1_mul_mat>
-        (vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst);
-
-#elif __CUDA_ARCH__ >= MIN_CC_DP4A
-    const int mmq_x  =  MMQ_X_Q3_K_PASCAL;
-    const int mmq_y  =  MMQ_Y_Q3_K_PASCAL;
-    const int nwarps = NWARPS_Q3_K_PASCAL;
-
-    mul_mat_q<QK_K, QR3_K, QI3_K, false, block_q3_K, mmq_x, mmq_y, nwarps, allocate_tiles_q3_K<mmq_y>,
-        load_tiles_q3_K<mmq_y, nwarps, need_check>, VDR_Q3_K_Q8_1_MMQ, vec_dot_q3_K_q8_1_mul_mat>
-        (vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst);
-#else
-    (void) vec_dot_q3_K_q8_1_mul_mat;
-    bad_arch();
-#endif // __CUDA_ARCH__ >= CC_VOLTA
-}
-
-#define  MMQ_X_Q4_K_RDNA2  64
-#define  MMQ_Y_Q4_K_RDNA2  128
-#define NWARPS_Q4_K_RDNA2  8
-#define  MMQ_X_Q4_K_RDNA1  32
-#define  MMQ_Y_Q4_K_RDNA1  64
-#define NWARPS_Q4_K_RDNA1  8
-#if defined(CUDA_USE_TENSOR_CORES)
-#define  MMQ_X_Q4_K_AMPERE 4
-#define  MMQ_Y_Q4_K_AMPERE 32
-#define NWARPS_Q4_K_AMPERE 4
-#else
-#define  MMQ_X_Q4_K_AMPERE 64
-#define  MMQ_Y_Q4_K_AMPERE 128
-#define NWARPS_Q4_K_AMPERE 4
-#endif
-#define  MMQ_X_Q4_K_PASCAL 64
-#define  MMQ_Y_Q4_K_PASCAL 64
-#define NWARPS_Q4_K_PASCAL 8
-
-template <bool need_check> static __global__ void
-#if defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)
-#if defined(RDNA3) || defined(RDNA2)
-    __launch_bounds__(WARP_SIZE*NWARPS_Q4_K_RDNA2, 2)
-#endif // defined(RDNA3) || defined(RDNA2)
-#elif __CUDA_ARCH__ < CC_VOLTA
-    __launch_bounds__(WARP_SIZE*NWARPS_Q4_K_PASCAL, 2)
-#endif // __CUDA_ARCH__ < CC_VOLTA
-    mul_mat_q4_K(
-    const void * __restrict__ vx, const void * __restrict__ vy, float * __restrict__ dst,
-    const int ncols_x, const int nrows_x, const int ncols_y, const int nrows_y, const int nrows_dst) {
-
-#if defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)
-#if defined(RDNA3) || defined(RDNA2)
-    const int mmq_x  =  MMQ_X_Q4_K_RDNA2;
-    const int mmq_y  =  MMQ_Y_Q4_K_RDNA2;
-    const int nwarps = NWARPS_Q4_K_RDNA2;
-#else
-    const int mmq_x  =  MMQ_X_Q4_K_RDNA1;
-    const int mmq_y  =  MMQ_Y_Q4_K_RDNA1;
-    const int nwarps = NWARPS_Q4_K_RDNA1;
-#endif // defined(RDNA3) || defined(RDNA2)
-
-    mul_mat_q<QK_K, QR4_K, QI4_K, true, block_q4_K, mmq_x, mmq_y, nwarps, allocate_tiles_q4_K<mmq_y>,
-        load_tiles_q4_K<mmq_y, nwarps, need_check>, VDR_Q4_K_Q8_1_MMQ, vec_dot_q4_K_q8_1_mul_mat>
-        (vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst);
-
-#elif __CUDA_ARCH__ >= CC_VOLTA
-    const int mmq_x  =  MMQ_X_Q4_K_AMPERE;
-    const int mmq_y  =  MMQ_Y_Q4_K_AMPERE;
-    const int nwarps = NWARPS_Q4_K_AMPERE;
-
-    mul_mat_q<QK_K, QR4_K, QI4_K, true, block_q4_K, mmq_x, mmq_y, nwarps, allocate_tiles_q4_K<mmq_y>,
-        load_tiles_q4_K<mmq_y, nwarps, need_check>, VDR_Q4_K_Q8_1_MMQ, vec_dot_q4_K_q8_1_mul_mat>
-        (vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst);
-
-#elif __CUDA_ARCH__ >= MIN_CC_DP4A
-    const int mmq_x  =  MMQ_X_Q4_K_PASCAL;
-    const int mmq_y  =  MMQ_Y_Q4_K_PASCAL;
-    const int nwarps = NWARPS_Q4_K_PASCAL;
-
-    mul_mat_q<QK_K, QR4_K, QI4_K, true, block_q4_K, mmq_x, mmq_y, nwarps, allocate_tiles_q4_K<mmq_y>,
-        load_tiles_q4_K<mmq_y, nwarps, need_check>, VDR_Q4_K_Q8_1_MMQ, vec_dot_q4_K_q8_1_mul_mat>
-        (vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst);
-#else
-    (void) vec_dot_q4_K_q8_1_mul_mat;
-    bad_arch();
-#endif // __CUDA_ARCH__ >= CC_VOLTA
-}
-
-#define  MMQ_X_Q5_K_RDNA2  64
-#define  MMQ_Y_Q5_K_RDNA2  128
-#define NWARPS_Q5_K_RDNA2  8
-#define  MMQ_X_Q5_K_RDNA1  32
-#define  MMQ_Y_Q5_K_RDNA1  64
-#define NWARPS_Q5_K_RDNA1  8
-#if defined(CUDA_USE_TENSOR_CORES)
-#define  MMQ_X_Q5_K_AMPERE 4
-#define  MMQ_Y_Q5_K_AMPERE 32
-#define NWARPS_Q5_K_AMPERE 4
-#else
-#define  MMQ_X_Q5_K_AMPERE 64
-#define  MMQ_Y_Q5_K_AMPERE 128
-#define NWARPS_Q5_K_AMPERE 4
-#endif
-#define  MMQ_X_Q5_K_PASCAL 64
-#define  MMQ_Y_Q5_K_PASCAL 64
-#define NWARPS_Q5_K_PASCAL 8
-
-template <bool need_check> static __global__ void
-#if defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)
-#if defined(RDNA3) || defined(RDNA2)
-    __launch_bounds__(WARP_SIZE*NWARPS_Q5_K_RDNA2, 2)
-#endif // defined(RDNA3) || defined(RDNA2)
-#endif // defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)
-mul_mat_q5_K(
-    const void * __restrict__ vx, const void * __restrict__ vy, float * __restrict__ dst,
-    const int ncols_x, const int nrows_x, const int ncols_y, const int nrows_y, const int nrows_dst) {
-
-#if defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)
-#if defined(RDNA3) || defined(RDNA2)
-    const int mmq_x  =  MMQ_X_Q5_K_RDNA2;
-    const int mmq_y  =  MMQ_Y_Q5_K_RDNA2;
-    const int nwarps = NWARPS_Q5_K_RDNA2;
-#else
-    const int mmq_x  =  MMQ_X_Q5_K_RDNA1;
-    const int mmq_y  =  MMQ_Y_Q5_K_RDNA1;
-    const int nwarps = NWARPS_Q5_K_RDNA1;
-#endif // defined(RDNA3) || defined(RDNA2)
-
-    mul_mat_q<QK_K, QR5_K, QI5_K, true, block_q5_K, mmq_x, mmq_y, nwarps, allocate_tiles_q5_K<mmq_y>,
-        load_tiles_q5_K<mmq_y, nwarps, need_check>, VDR_Q5_K_Q8_1_MMQ, vec_dot_q5_K_q8_1_mul_mat>
-        (vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst);
-
-#elif __CUDA_ARCH__ >= CC_VOLTA
-    const int mmq_x  =  MMQ_X_Q5_K_AMPERE;
-    const int mmq_y  =  MMQ_Y_Q5_K_AMPERE;
-    const int nwarps = NWARPS_Q5_K_AMPERE;
-
-    mul_mat_q<QK_K, QR5_K, QI5_K, true, block_q5_K, mmq_x, mmq_y, nwarps, allocate_tiles_q5_K<mmq_y>,
-        load_tiles_q5_K<mmq_y, nwarps, need_check>, VDR_Q5_K_Q8_1_MMQ, vec_dot_q5_K_q8_1_mul_mat>
-        (vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst);
-
-#elif __CUDA_ARCH__ >= MIN_CC_DP4A
-    const int mmq_x  =  MMQ_X_Q5_K_PASCAL;
-    const int mmq_y  =  MMQ_Y_Q5_K_PASCAL;
-    const int nwarps = NWARPS_Q5_K_PASCAL;
-
-    mul_mat_q<QK_K, QR5_K, QI5_K, true, block_q5_K, mmq_x, mmq_y, nwarps, allocate_tiles_q5_K<mmq_y>,
-        load_tiles_q5_K<mmq_y, nwarps, need_check>, VDR_Q5_K_Q8_1_MMQ, vec_dot_q5_K_q8_1_mul_mat>
-        (vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst);
-#else
-    (void) vec_dot_q5_K_q8_1_mul_mat;
-    bad_arch();
-#endif // __CUDA_ARCH__ >= CC_VOLTA
-}
-
-#define  MMQ_X_Q6_K_RDNA2  64
-#define  MMQ_Y_Q6_K_RDNA2  128
-#define NWARPS_Q6_K_RDNA2  8
-#define  MMQ_X_Q6_K_RDNA1  32
-#define  MMQ_Y_Q6_K_RDNA1  64
-#define NWARPS_Q6_K_RDNA1  8
-#if defined(CUDA_USE_TENSOR_CORES)
-#define  MMQ_X_Q6_K_AMPERE 4
-#define  MMQ_Y_Q6_K_AMPERE 32
-#define NWARPS_Q6_K_AMPERE 4
-#else
-#define  MMQ_X_Q6_K_AMPERE 64
-#define  MMQ_Y_Q6_K_AMPERE 64
-#define NWARPS_Q6_K_AMPERE 4
-#endif
-#define  MMQ_X_Q6_K_PASCAL 64
-#define  MMQ_Y_Q6_K_PASCAL 64
-#define NWARPS_Q6_K_PASCAL 8
-
-template <bool need_check> static __global__ void
-#if defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)
-#if defined(RDNA3) || defined(RDNA2)
-    __launch_bounds__(WARP_SIZE*NWARPS_Q6_K_RDNA2, 2)
-#endif // defined(RDNA3) || defined(RDNA2)
-#elif __CUDA_ARCH__ < CC_VOLTA
-    __launch_bounds__(WARP_SIZE*NWARPS_Q6_K_PASCAL, 2)
-#endif // __CUDA_ARCH__ < CC_VOLTA
-    mul_mat_q6_K(
-    const void * __restrict__ vx, const void * __restrict__ vy, float * __restrict__ dst,
-    const int ncols_x, const int nrows_x, const int ncols_y, const int nrows_y, const int nrows_dst) {
-
-#if defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)
-#if defined(RDNA3) || defined(RDNA2)
-    const int mmq_x  =  MMQ_X_Q6_K_RDNA2;
-    const int mmq_y  =  MMQ_Y_Q6_K_RDNA2;
-    const int nwarps = NWARPS_Q6_K_RDNA2;
-#else
-    const int mmq_x  =  MMQ_X_Q6_K_RDNA1;
-    const int mmq_y  =  MMQ_Y_Q6_K_RDNA1;
-    const int nwarps = NWARPS_Q6_K_RDNA1;
-#endif // defined(RDNA3) || defined(RDNA2)
-
-    mul_mat_q<QK_K, QR6_K, QI6_K, false, block_q6_K, mmq_x, mmq_y, nwarps, allocate_tiles_q6_K<mmq_y>,
-        load_tiles_q6_K<mmq_y, nwarps, need_check>, VDR_Q6_K_Q8_1_MMQ, vec_dot_q6_K_q8_1_mul_mat>
-        (vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst);
-
-#elif __CUDA_ARCH__ >= CC_VOLTA
-    const int mmq_x  =  MMQ_X_Q6_K_AMPERE;
-    const int mmq_y  =  MMQ_Y_Q6_K_AMPERE;
-    const int nwarps = NWARPS_Q6_K_AMPERE;
-
-    mul_mat_q<QK_K, QR6_K, QI6_K, false, block_q6_K, mmq_x, mmq_y, nwarps, allocate_tiles_q6_K<mmq_y>,
-        load_tiles_q6_K<mmq_y, nwarps, need_check>, VDR_Q6_K_Q8_1_MMQ, vec_dot_q6_K_q8_1_mul_mat>
-        (vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst);
-
-#elif __CUDA_ARCH__ >= MIN_CC_DP4A
-    const int mmq_x  =  MMQ_X_Q6_K_PASCAL;
-    const int mmq_y  =  MMQ_Y_Q6_K_PASCAL;
-    const int nwarps = NWARPS_Q6_K_PASCAL;
-
-    mul_mat_q<QK_K, QR6_K, QI6_K, false, block_q6_K, mmq_x, mmq_y, nwarps, allocate_tiles_q6_K<mmq_y>,
-        load_tiles_q6_K<mmq_y, nwarps, need_check>, VDR_Q6_K_Q8_1_MMQ, vec_dot_q6_K_q8_1_mul_mat>
-        (vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst);
-#else
-    (void) vec_dot_q6_K_q8_1_mul_mat;
-    bad_arch();
-#endif // __CUDA_ARCH__ >= CC_VOLTA
-}
-
-template <int qk, int qi, typename block_q_t, int vdr, vec_dot_q_cuda_t vec_dot_q_cuda>
-static __global__ void mul_mat_vec_q(const void * __restrict__ vx, const void * __restrict__ vy, float * __restrict__ dst, const int ncols, const int nrows) {
-    const int row = blockIdx.x*blockDim.y + threadIdx.y;
-
-    if (row >= nrows) {
-        return;
-    }
-
-    const int blocks_per_row = ncols / qk;
-    const int blocks_per_warp = vdr * WARP_SIZE / qi;
-
-// partial sum for each thread
-    float tmp = 0.0f;
-
-    const block_q_t  * x = (const block_q_t  *) vx;
-    const block_q8_1 * y = (const block_q8_1 *) vy;
-
-    for (int i = 0; i < blocks_per_row; i += blocks_per_warp) {
-        const int ibx = row*blocks_per_row + i + threadIdx.x / (qi/vdr); // x block index
-
-        const int iby = (i + threadIdx.x / (qi/vdr)) * (qk/QK8_1); // y block index that aligns with ibx
-
-        const int iqs  = vdr * (threadIdx.x % (qi/vdr)); // x block quant index when casting the quants to int
-
-        tmp += vec_dot_q_cuda(&x[ibx], &y[iby], iqs);
-    }
-
-    // sum up partial sums and write back result
-#pragma unroll
-    for (int mask = 16; mask > 0; mask >>= 1) {
-        tmp += __shfl_xor_sync(0xffffffff, tmp, mask, 32);
-    }
-
-    if (threadIdx.x == 0) {
-        dst[row] = tmp;
-    }
-}
-
-template <int qk, int qr, dequantize_kernel_t dequantize_kernel>
-static __global__ void dequantize_mul_mat_vec(const void * __restrict__ vx, const dfloat * __restrict__ y, float * __restrict__ dst, const int ncols, const int nrows) {
-    // qk = quantized weights per x block
-    // qr = number of quantized weights per data value in x block
-    const int row = blockIdx.x*blockDim.y + threadIdx.y;
-
-    if (row >= nrows) {
-        return;
-    }
-
-    const int tid = threadIdx.x;
-
-    const int iter_stride = 2*GGML_CUDA_DMMV_X;
-    const int vals_per_iter = iter_stride / WARP_SIZE; // num quantized vals per thread and i iter
-    const int y_offset = qr == 1 ? 1 : qk/2;
-
-// partial sum for each thread
-#ifdef GGML_CUDA_F16
-    half2 tmp = {0.0f, 0.0f}; // two sums for f16 to take advantage of half2 intrinsics
-#else
-    float tmp = 0.0f;
-#endif // GGML_CUDA_F16
-
-    for (int i = 0; i < ncols; i += iter_stride) {
-        const int col = i + vals_per_iter*tid;
-        const int ib = (row*ncols + col)/qk; // x block index
-        const int iqs = (col%qk)/qr; // x quant index
-        const int iybs = col - col%qk; // y block start index
-
-// processing >2 values per i iter is faster for fast GPUs
-#pragma unroll
-        for (int j = 0; j < vals_per_iter; j += 2) {
-            // process 2 vals per j iter
-
-            // dequantize
-            // for qr = 2 the iqs needs to increase by 1 per j iter because 2 weights per data val
-            dfloat2 v;
-            dequantize_kernel(vx, ib, iqs + j/qr, v);
-
-            // matrix multiplication
-            // for qr = 2 the y index needs to increase by 1 per j iter because of y_offset = qk/2
-#ifdef GGML_CUDA_F16
-            tmp += __hmul2(v, {
-                y[iybs + iqs + j/qr + 0],
-                y[iybs + iqs + j/qr + y_offset]
-            });
-#else
-            tmp += v.x * y[iybs + iqs + j/qr + 0];
-            tmp += v.y * y[iybs + iqs + j/qr + y_offset];
-#endif // GGML_CUDA_F16
-        }
-    }
-
-    // sum up partial sums and write back result
-#pragma unroll
-    for (int mask = 16; mask > 0; mask >>= 1) {
-        tmp += __shfl_xor_sync(0xffffffff, tmp, mask, 32);
-    }
-
-    if (tid == 0) {
-#ifdef GGML_CUDA_F16
-        dst[row] = tmp.x + tmp.y;
-#else
-        dst[row] = tmp;
-#endif // GGML_CUDA_F16
-    }
-}
-
-static __global__ void mul_mat_p021_f16_f32(
-    const void * __restrict__ vx, const float * __restrict__ y, float * __restrict__ dst,
-    const int ncols_x, const int nrows_x, const int nchannels_x, const int nchannels_y) {
-
-    const half * x = (const half *) vx;
-
-    const int row_x = blockDim.y*blockIdx.y + threadIdx.y;
-    const int channel = blockDim.z*blockIdx.z + threadIdx.z;
-    const int channel_x = channel / (nchannels_y / nchannels_x);
-
-    const int nrows_y = ncols_x;
-    const int nrows_dst = nrows_x;
-    const int row_dst = row_x;
-
-    float tmp = 0.0f;
-
-    for (int col_x0 = 0; col_x0 < ncols_x; col_x0 += blockDim.x) {
-        const int col_x = col_x0 + threadIdx.x;
-
-        if (col_x >= ncols_x) {
-            break;
-        }
-
-        // x is transposed and permuted
-        const int ix = row_x*nchannels_x*ncols_x + channel_x*ncols_x + col_x;
-        const float xi = __half2float(x[ix]);
-
-        const int row_y = col_x;
-
-        // y is not transposed but permuted
-        const int iy = channel*nrows_y + row_y;
-
-        tmp += xi * y[iy];
-    }
-
-    // dst is not transposed and not permuted
-    const int idst = channel*nrows_dst + row_dst;
-
-    // sum up partial sums and write back result
-#pragma unroll
-    for (int mask = 16; mask > 0; mask >>= 1) {
-        tmp += __shfl_xor_sync(0xffffffff, tmp, mask, 32);
-    }
-
-    if (threadIdx.x == 0) {
-        dst[idst] = tmp;
-    }
-}
-
-static __global__ void mul_mat_vec_nc_f16_f32( // nc == non-contiguous
-    const void * __restrict__ vx, const float * __restrict__ y, float * __restrict__ dst, const int ncols_x, const int nrows_x,
-    const int row_stride_x, const int channel_stride_x, const int channel_x_divisor) {
-
-    const half * x = (const half *) vx;
-
-    const int row_x     = blockDim.y*blockIdx.y + threadIdx.y;
-    const int channel   = blockDim.z*blockIdx.z + threadIdx.z;
-    const int channel_x = channel / channel_x_divisor;
-
-    const int nrows_y   = ncols_x;
-    const int nrows_dst = nrows_x;
-    const int row_dst   = row_x;
-
-    const int idst = channel*nrows_dst + row_dst;
-
-    float tmp = 0.0f;
-
-    for (int col_x0 = 0; col_x0 < ncols_x; col_x0 += blockDim.x) {
-        const int col_x = col_x0 + threadIdx.x;
-
-        if (col_x >= ncols_x) {
-            break;
-        }
-
-        const int row_y = col_x;
-
-        const int ix = channel_x*channel_stride_x + row_x*row_stride_x + col_x;
-        const int iy = channel*nrows_y + row_y;
-
-        const float xi = __half2float(x[ix]);
-
-        tmp += xi * y[iy];
-    }
-
-    // sum up partial sums and write back result
-#pragma unroll
-    for (int mask = 16; mask > 0; mask >>= 1) {
-        tmp += __shfl_xor_sync(0xffffffff, tmp, mask, 32);
-    }
-
-    if (threadIdx.x == 0) {
-        dst[idst] = tmp;
-    }
-}
-
-static __device__ void cpy_1_f32_f32(const char * cxi, char * cdsti) {
-    const float * xi = (const float *) cxi;
-    float * dsti = (float *) cdsti;
-
-    *dsti = *xi;
-}
-
-static __device__ void cpy_1_f32_f16(const char * cxi, char * cdsti) {
-    const float * xi = (const float *) cxi;
-    half * dsti = (half *) cdsti;
-
-    *dsti = __float2half(*xi);
-}
-
-static __device__ void cpy_1_f16_f16(const char * cxi, char * cdsti) {
-    const half * xi = (const half *) cxi;
-    half * dsti = (half *) cdsti;
-
-    *dsti = *xi;
-}
-
-template <cpy_kernel_t cpy_1>
-static __global__ void cpy_f32_f16(const char * cx, char * cdst, const int ne,
-                                   const int ne00, const int ne01, const int nb00, const int nb01, const int nb02,
-                                   const int ne10, const int ne11, const int nb10, const int nb11, const int nb12) {
-    const int i = blockDim.x*blockIdx.x + threadIdx.x;
-
-    if (i >= ne) {
-        return;
-    }
-
-    // determine indices i02/i12, i01/i11, i00/i10 as a function of index i of flattened tensor
-    // then combine those indices with the corresponding byte offsets to get the total offsets
-    const int i02 = i / (ne00*ne01);
-    const int i01 = (i - i02*ne01*ne00) / ne00;
-    const int i00 = i - i02*ne01*ne00 - i01*ne00;
-    const int x_offset = i00*nb00 + i01*nb01 + i02*nb02;
-
-    const int i12 = i / (ne10*ne11);
-    const int i11 = (i - i12*ne10*ne11) / ne10;
-    const int i10 = i - i12*ne10*ne11 - i11*ne10;
-    const int dst_offset = i10*nb10 + i11*nb11 + i12*nb12;
-
-    cpy_1(cx + x_offset, cdst + dst_offset);
-}
-
-static __device__ void cpy_blck_f32_q8_0(const char * cxi, char * cdsti) {
-    const float * xi = (const float *) cxi;
-    block_q8_0 * dsti = (block_q8_0 *) cdsti;
-
-    float amax = 0.0f; // absolute max
-
-    for (int j = 0; j < QK8_0; j++) {
-        const float v = xi[j];
-        amax = fmaxf(amax, fabsf(v));
-    }
-
-    const float d = amax / ((1 << 7) - 1);
-    const float id = d ? 1.0f/d : 0.0f;
-
-    dsti->d = d;
-
-    for (int j = 0; j < QK8_0; ++j) {
-        const float x0 = xi[j]*id;
-
-        dsti->qs[j] = roundf(x0);
-    }
-}
-
-static __device__ void cpy_blck_f32_q4_0(const char * cxi, char * cdsti) {
-    const float * xi = (const float *) cxi;
-    block_q4_0 * dsti = (block_q4_0 *) cdsti;
-
-    float amax = 0.0f;
-    float vmax = 0.0f;
-
-    for (int j = 0; j < QK4_0; ++j) {
-        const float v = xi[j];
-        if (amax < fabsf(v)) {
-            amax = fabsf(v);
-            vmax = v;
-        }
-    }
-
-    const float d  = vmax / -8;
-    const float id = d ? 1.0f/d : 0.0f;
-
-    dsti->d = d;
-
-    for (int j = 0; j < QK4_0/2; ++j) {
-        const float x0 = xi[0       + j]*id;
-        const float x1 = xi[QK4_0/2 + j]*id;
-
-        const uint8_t xi0 = min(15, (int8_t)(x0 + 8.5f));
-        const uint8_t xi1 = min(15, (int8_t)(x1 + 8.5f));
-
-        dsti->qs[j]  = xi0;
-        dsti->qs[j] |= xi1 << 4;
-    }
-}
-
-static __device__ void cpy_blck_f32_q4_1(const char * cxi, char * cdsti) {
-    const float * xi = (const float *) cxi;
-    block_q4_1 * dsti = (block_q4_1 *) cdsti;
-
-    float vmin = FLT_MAX;
-    float vmax = -FLT_MAX;
-
-    for (int j = 0; j < QK4_1; ++j) {
-        const float v = xi[j];
-
-        if (v < vmin) vmin = v;
-        if (v > vmax) vmax = v;
-    }
-
-    const float d  = (vmax - vmin) / ((1 << 4) - 1);
-    const float id = d ? 1.0f/d : 0.0f;
-
-    dsti->dm.x = d;
-    dsti->dm.y = vmin;
-
-    for (int j = 0; j < QK4_1/2; ++j) {
-        const float x0 = (xi[0       + j] - vmin)*id;
-        const float x1 = (xi[QK4_1/2 + j] - vmin)*id;
-
-        const uint8_t xi0 = min(15, (int8_t)(x0 + 0.5f));
-        const uint8_t xi1 = min(15, (int8_t)(x1 + 0.5f));
-
-        dsti->qs[j]  = xi0;
-        dsti->qs[j] |= xi1 << 4;
-    }
-}
-
-template <cpy_kernel_t cpy_blck, int qk>
-static __global__ void cpy_f32_q(const char * cx, char * cdst, const int ne,
-                                 const int ne00, const int ne01, const int nb00, const int nb01, const int nb02,
-                                 const int ne10, const int ne11, const int nb10, const int nb11, const int nb12) {
-    const int i = (blockDim.x*blockIdx.x + threadIdx.x)*qk;
-
-    if (i >= ne) {
-        return;
-    }
-
-    const int i02 = i / (ne00*ne01);
-    const int i01 = (i - i02*ne01*ne00) / ne00;
-    const int i00 = (i - i02*ne01*ne00 - i01*ne00);
-    const int x_offset = i00*nb00 + i01*nb01 + i02*nb02;
-
-    const int i12 = i / (ne10*ne11);
-    const int i11 = (i - i12*ne10*ne11) / ne10;
-    const int i10 = (i - i12*ne10*ne11 - i11*ne10)/qk;
-    const int dst_offset = i10*nb10 + i11*nb11 + i12*nb12;
-
-    cpy_blck(cx + x_offset, cdst + dst_offset);
-}
-
-static __device__ float rope_yarn_ramp(const float low, const float high, const int i0) {
-    const float y = (i0 / 2 - low) / max(0.001f, high - low);
-    return 1.0f - min(1.0f, max(0.0f, y));
-}
-
-struct rope_corr_dims {
-    float v[4];
-};
-
-// YaRN algorithm based on LlamaYaRNScaledRotaryEmbedding.py from https://github.com/jquesnelle/yarn
-// MIT licensed. Copyright (c) 2023 Jeffrey Quesnelle and Bowen Peng.
-static __device__ void rope_yarn(
-    float theta_extrap, float freq_scale, rope_corr_dims corr_dims, int64_t i0, float ext_factor, float mscale,
-    float * cos_theta, float * sin_theta
-) {
-    // Get n-d rotational scaling corrected for extrapolation
-    float theta_interp = freq_scale * theta_extrap;
-    float theta = theta_interp;
-    if (ext_factor != 0.0f) {
-        float ramp_mix = rope_yarn_ramp(corr_dims.v[0], corr_dims.v[1], i0) * ext_factor;
-        theta = theta_interp * (1 - ramp_mix) + theta_extrap * ramp_mix;
-
-        // Get n-d magnitude scaling corrected for interpolation
-        mscale *= 1.0f + 0.1f * logf(1.0f / freq_scale);
-    }
-    *cos_theta = cosf(theta) * mscale;
-    *sin_theta = sinf(theta) * mscale;
-}
-
-// rope == RoPE == rotary positional embedding
-template<typename T, bool has_pos>
-static __global__ void rope(
-    const T * x, T * dst, int ncols, const int32_t * pos, float freq_scale, int p_delta_rows, float freq_base,
-    float ext_factor, float attn_factor, rope_corr_dims corr_dims
-) {
-    const int col = 2*(blockDim.y*blockIdx.y + threadIdx.y);
-
-    if (col >= ncols) {
-        return;
-    }
-
-    const int row = blockDim.x*blockIdx.x + threadIdx.x;
-    const int i = row*ncols + col;
-    const int i2 = row/p_delta_rows;
-
-    const int p = has_pos ? pos[i2] : 0;
-    const float theta_base = p*powf(freq_base, -float(col)/ncols);
-
-    float cos_theta, sin_theta;
-    rope_yarn(theta_base, freq_scale, corr_dims, col, ext_factor, attn_factor, &cos_theta, &sin_theta);
-
-    const float x0 = x[i + 0];
-    const float x1 = x[i + 1];
-
-    dst[i + 0] = x0*cos_theta - x1*sin_theta;
-    dst[i + 1] = x0*sin_theta + x1*cos_theta;
-}
-
-template<typename T, bool has_pos>
-static __global__ void rope_neox(
-    const T * x, T * dst, int ncols, int n_dims, const int32_t * pos, float freq_scale, int p_delta_rows,
-    float ext_factor, float attn_factor, rope_corr_dims corr_dims, float theta_scale, float inv_ndims
-) {
-    const int col = 2*(blockDim.y*blockIdx.y + threadIdx.y);
-
-    if (col >= ncols) {
-        return;
-    }
-
-    const int row = blockDim.x*blockIdx.x + threadIdx.x;
-    const int ib = col / n_dims;
-    const int ic = col % n_dims;
-
-    if (ib > 0) {
-        const int i = row*ncols + ib*n_dims + ic;
-
-        dst[i + 0] = x[i + 0];
-        dst[i + 1] = x[i + 1];
-
-        return;
-    }
-
-    const int i  = row*ncols + ib*n_dims + ic/2;
-    const int i2 = row/p_delta_rows;
-
-    float cur_rot = inv_ndims * ic - ib;
-
-    const int p = has_pos ? pos[i2] : 0;
-    const float theta_base = p*freq_scale*powf(theta_scale, col/2.0f);
-
-    float cos_theta, sin_theta;
-    rope_yarn(theta_base, freq_scale, corr_dims, cur_rot, ext_factor, attn_factor, &cos_theta, &sin_theta);
-
-    const float x0 = x[i + 0];
-    const float x1 = x[i + n_dims/2];
-
-    dst[i + 0]        = x0*cos_theta - x1*sin_theta;
-    dst[i + n_dims/2] = x0*sin_theta + x1*cos_theta;
-}
-
-static __global__ void rope_glm_f32(
-    const float * x, float * dst, int ncols, const int32_t * pos, float freq_scale, int p_delta_rows, float freq_base,
-    int n_ctx
-) {
-    const int col = blockDim.x*blockIdx.x + threadIdx.x;
-    const int half_n_dims = ncols/4;
-
-    if (col >= half_n_dims) {
-        return;
-    }
-
-    const int row = blockDim.y*blockIdx.y + threadIdx.y;
-    const int i = row*ncols + col;
-    const int i2 = row/p_delta_rows;
-
-    const float col_theta_scale = powf(freq_base, -2.0f*col/ncols);
-     // FIXME: this is likely wrong
-    const int p = pos != nullptr ? pos[i2] : 0;
-
-    const float theta = min(p, n_ctx - 2)*freq_scale*col_theta_scale;
-    const float sin_theta = sinf(theta);
-    const float cos_theta = cosf(theta);
-
-    const float x0 = x[i + 0];
-    const float x1 = x[i + half_n_dims];
-
-    dst[i + 0]           = x0*cos_theta - x1*sin_theta;
-    dst[i + half_n_dims] = x0*sin_theta + x1*cos_theta;
-
-    const float block_theta = ((float)max(p - n_ctx - 2, 0))*col_theta_scale;
-    const float sin_block_theta = sinf(block_theta);
-    const float cos_block_theta = cosf(block_theta);
-
-    const float x2 = x[i + half_n_dims * 2];
-    const float x3 = x[i + half_n_dims * 3];
-
-    dst[i + half_n_dims * 2] = x2*cos_block_theta - x3*sin_block_theta;
-    dst[i + half_n_dims * 3] = x2*sin_block_theta + x3*cos_block_theta;
-}
-
-static __global__ void alibi_f32(const float * x, float * dst, const int ncols, const int k_rows,
-                                 const int n_heads_log2_floor, const float m0, const float m1) {
-    const int col = blockDim.x*blockIdx.x + threadIdx.x;
-
-    if (col >= ncols) {
-        return;
-    }
-
-    const int row = blockDim.y*blockIdx.y + threadIdx.y;
-    const int i = row*ncols + col;
-
-    const int k = row/k_rows;
-
-    float m_k;
-    if (k < n_heads_log2_floor) {
-        m_k = powf(m0, k + 1);
-    } else {
-        m_k = powf(m1, 2 * (k - n_heads_log2_floor) + 1);
-    }
-
-    dst[i] = col * m_k + x[i];
-}
-
-static __global__ void k_sum_rows_f32(const float * x, float * dst, const int ncols) {
-    const int row = blockIdx.y;
-    const int col = threadIdx.x;
-
-    float sum = 0.0f;
-    for (int i = col; i < ncols; i += blockDim.x) {
-        sum += x[row * ncols + i];
-    }
-
-    sum = warp_reduce_sum(sum);
-
-    if (col == 0) {
-        dst[row] = sum;
-    }
-}
-
-template<typename T>
-static inline __device__ void swap(T & a, T & b) {
-    T tmp = a;
-    a = b;
-    b = tmp;
-}
-
-template<ggml_sort_order order>
-static __global__ void k_argsort_f32_i32(const float * x, int * dst, const int ncols) {
-    // bitonic sort
-    int col = threadIdx.x;
-    int row = blockIdx.y;
-
-    if (col >= ncols) return;
-
-    const float * x_row = x + row * ncols;
-    int * dst_row = dst + row * ncols;
-
-    // initialize indices
-    if (col < ncols) {
-        dst_row[col] = col;
-    }
-    __syncthreads();
-
-    for (int k = 2; k <= ncols; k *= 2) {
-        for (int j = k / 2; j > 0; j /= 2) {
-            int ixj = col ^ j;
-            if (ixj > col) {
-                if ((col & k) == 0) {
-                    if (order == GGML_SORT_ASC ? x_row[dst_row[col]] > x_row[dst_row[ixj]] : x_row[dst_row[col]] < x_row[dst_row[ixj]]) {
-                        swap(dst_row[col], dst_row[ixj]);
-                    }
-                } else {
-                    if (order == GGML_SORT_ASC ? x_row[dst_row[col]] < x_row[dst_row[ixj]] : x_row[dst_row[col]] > x_row[dst_row[ixj]]) {
-                        swap(dst_row[col], dst_row[ixj]);
-                    }
-                }
-            }
-            __syncthreads();
-        }
-    }
-}
-
-static __global__ void diag_mask_inf_f32(const float * x, float * dst, const int ncols, const int rows_per_channel, const int n_past) {
-    const int col = blockDim.y*blockIdx.y + threadIdx.y;
-    const int row = blockDim.x*blockIdx.x + threadIdx.x;
-
-    if (col >= ncols) {
-        return;
-    }
-
-    const int i = row*ncols + col;
-    //dst[i] = col > (n_past + row % rows_per_channel) ? -INFINITY : x[i];
-    //dst[i] = x[i] - (col > n_past + row % rows_per_channel) * INT_MAX; // equivalent within rounding error but slightly faster on GPU
-    dst[i] = x[i] - (col > n_past + row % rows_per_channel) * FLT_MAX;
-}
-
-static __global__ void soft_max_f32(const float * x, const float * y, float * dst, const int ncols, const int nrows_y, const float scale) {
-    const int tid  = threadIdx.x;
-    const int rowx = blockIdx.x;
-    const int rowy = rowx % nrows_y; // broadcast the mask (y) in the row dimension
-
-    const int block_size = blockDim.x;
-
-    const int warp_id = threadIdx.x / WARP_SIZE;
-    const int lane_id = threadIdx.x % WARP_SIZE;
-
-    __shared__ float buf[CUDA_SOFT_MAX_BLOCK_SIZE/WARP_SIZE];
-
-    float max_val = -INFINITY;
-
-    for (int col = tid; col < ncols; col += block_size) {
-        const int ix = rowx*ncols + col;
-        const int iy = rowy*ncols + col;
-        max_val = max(max_val, x[ix]*scale + (y ? y[iy] : 0.0f));
-    }
-
-    // find the max value in the block
-    max_val = warp_reduce_max(max_val);
-    if (block_size > WARP_SIZE) {
-        if (warp_id == 0) {
-            buf[lane_id] = -INFINITY;
-        }
-        __syncthreads();
-
-        if (lane_id == 0) {
-            buf[warp_id] = max_val;
-        }
-        __syncthreads();
-
-        max_val = buf[lane_id];
-        max_val = warp_reduce_max(max_val);
-    }
-
-    float tmp = 0.f;
-
-    for (int col = tid; col < ncols; col += block_size) {
-        const int ix = rowx*ncols + col;
-        const int iy = rowy*ncols + col;
-        const float val = expf((x[ix]*scale + (y ? y[iy] : 0.0f)) - max_val);
-        tmp += val;
-        dst[ix] = val;
-    }
-
-    // find the sum of exps in the block
-    tmp = warp_reduce_sum(tmp);
-    if (block_size > WARP_SIZE) {
-        if (warp_id == 0) {
-            buf[lane_id] = 0.f;
-        }
-        __syncthreads();
-
-        if (lane_id == 0) {
-            buf[warp_id] = tmp;
-        }
-        __syncthreads();
-
-        tmp = buf[lane_id];
-        tmp = warp_reduce_sum(tmp);
-    }
-
-    const float inv_tmp = 1.f / tmp;
-
-    for (int col = tid; col < ncols; col += block_size) {
-        const int i = rowx*ncols + col;
-        dst[i] *= inv_tmp;
-    }
-}
-
-static __global__ void scale_f32(const float * x, float * dst, const float scale, const int k) {
-    const int i = blockDim.x*blockIdx.x + threadIdx.x;
-
-    if (i >= k) {
-        return;
-    }
-
-    dst[i] = scale * x[i];
-}
-
-static __global__ void clamp_f32(const float * x, float * dst, const float min, const float max, const int k) {
-    const int i = blockDim.x*blockIdx.x + threadIdx.x;
-
-    if (i >= k) {
-        return;
-    }
-
-    dst[i] = x[i] < min ? min : (x[i] > max ? max : x[i]);
-}
-
-static  __global__ void im2col_f32_f16(
-        const float * x, half * dst,
-        int offset_delta, int IW, int IH, int OW, int KW, int KH, int pelements, int CHW,
-        int s0, int s1, int p0, int p1, int d0, int d1) {
-    const int i = threadIdx.x + blockIdx.x * blockDim.x;
-    if (i >= pelements) {
-        return;
-    }
-
-    const int ksize = OW * (KH > 1 ? KW : 1);
-    const int kx = i / ksize;
-    const int kd = kx * ksize;
-    const int ky = (i - kd) / OW;
-    const int ix = i % OW;
-
-    const int64_t iiw = ix * s0 + kx * d0 - p0;
-    const int64_t iih = blockIdx.y * s1 + ky * d1 - p1;
-
-    const int64_t offset_dst =
-        (blockIdx.y * OW + ix) * CHW +
-        (blockIdx.z * (KW * KH) + ky * KW + kx);
-
-    if (iih < 0 || iih >= IH || iiw < 0 || iiw >= IW) {
-        dst[offset_dst] = __float2half(0.0f);
-    } else {
-        const int64_t offset_src = blockIdx.z * offset_delta;
-        dst[offset_dst] = __float2half(x[offset_src + iih * IW + iiw]);
-    }
-}
-
-template<int qk, int qr, dequantize_kernel_t dq>
-static void get_rows_cuda(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst,
-                            const void * src0_dd, const int32_t * src1_dd, float * dst_dd, cudaStream_t stream) {
-
-    GGML_TENSOR_BINARY_OP_LOCALS
-
-    const dim3 block_dims(CUDA_GET_ROWS_BLOCK_SIZE, 1, 1);
-    const int block_num_x = (ne00 + 2*CUDA_GET_ROWS_BLOCK_SIZE - 1) / (2*CUDA_GET_ROWS_BLOCK_SIZE);
-    const dim3 block_nums(block_num_x, ne10, ne11*ne12);
-
-    // strides in elements
-    //const size_t s0 = nb0 / ggml_element_size(dst);
-    const size_t s1 = nb1 / ggml_element_size(dst);
-    const size_t s2 = nb2 / ggml_element_size(dst);
-    const size_t s3 = nb3 / ggml_element_size(dst);
-
-    const size_t s10 = nb10 / ggml_element_size(src1);
-    const size_t s11 = nb11 / ggml_element_size(src1);
-    const size_t s12 = nb12 / ggml_element_size(src1);
-    //const size_t s13 = nb13 / ggml_element_size(src1);
-
-    GGML_ASSERT(ne00 % 2 == 0);
-
-    k_get_rows<qk, qr, dq><<<block_nums, block_dims, 0, stream>>>(
-            src0_dd, src1_dd, dst_dd,
-            ne00, /*ne01, ne02, ne03,*/
-            /*ne10, ne11,*/ ne12, /*ne13,*/
-            /* s0,*/ s1, s2, s3,
-            /* nb00,*/ nb01, nb02, nb03,
-            s10, s11, s12/*, s13*/);
-
-    (void) dst;
-}
-
-template<typename src0_t>
-static void get_rows_cuda_float(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst,
-                                const src0_t * src0_dd, const int32_t * src1_dd, float * dst_dd, cudaStream_t stream) {
-
-    GGML_TENSOR_BINARY_OP_LOCALS
-
-    const dim3 block_dims(CUDA_GET_ROWS_BLOCK_SIZE, 1, 1);
-    const int block_num_x = (ne00 + CUDA_GET_ROWS_BLOCK_SIZE - 1) / CUDA_GET_ROWS_BLOCK_SIZE;
-    const dim3 block_nums(block_num_x, ne10, ne11*ne12);
-
-    // strides in elements
-    //const size_t s0 = nb0 / ggml_element_size(dst);
-    const size_t s1 = nb1 / ggml_element_size(dst);
-    const size_t s2 = nb2 / ggml_element_size(dst);
-    const size_t s3 = nb3 / ggml_element_size(dst);
-
-    const size_t s10 = nb10 / ggml_element_size(src1);
-    const size_t s11 = nb11 / ggml_element_size(src1);
-    const size_t s12 = nb12 / ggml_element_size(src1);
-    //const size_t s13 = nb13 / ggml_element_size(src1);
-
-    k_get_rows_float<<<block_nums, block_dims, 0, stream>>>(
-            src0_dd, src1_dd, dst_dd,
-            ne00, /*ne01, ne02, ne03,*/
-            /*ne10, ne11,*/ ne12, /*ne13,*/
-            /* s0,*/ s1, s2, s3,
-            /* nb00,*/ nb01, nb02, nb03,
-            s10, s11, s12/*, s13*/);
-
-    (void) dst;
-}
-
-template<float (*bin_op)(const float, const float)>
-struct bin_bcast_cuda {
-    template<typename src0_t, typename src1_t, typename dst_t>
-    void operator()(const struct ggml_tensor * src0, const struct ggml_tensor * src1, struct ggml_tensor * dst,
-            const src0_t * src0_dd, const src1_t * src1_dd, dst_t * dst_dd,
-            cudaStream_t stream) {
-
-        GGML_TENSOR_BINARY_OP_LOCALS
-
-        int nr0 = ne10/ne0;
-        int nr1 = ne11/ne1;
-        int nr2 = ne12/ne2;
-        int nr3 = ne13/ne3;
-
-        int nr[4] = { nr0, nr1, nr2, nr3 };
-
-        // collapse dimensions until first broadcast dimension
-        int64_t cne0[] = {ne0, ne1, ne2, ne3};
-        int64_t cne1[] = {ne10, ne11, ne12, ne13};
-        size_t cnb0[] = {nb0, nb1, nb2, nb3};
-        size_t cnb1[] = {nb10, nb11, nb12, nb13};
-        auto collapse = [](int64_t cne[]) {
-            cne[0] *= cne[1];
-            cne[1] = cne[2];
-            cne[2] = cne[3];
-            cne[3] = 1;
-        };
-
-        auto collapse_nb = [](size_t cnb[], const int64_t cne[]) {
-            cnb[1] *= cne[1];
-            cnb[2] *= cne[2];
-            cnb[3] *= cne[3];
-        };
-
-        for (int i = 0; i < 4; i++) {
-            if (nr[i] != 1) {
-                break;
-            }
-            if (i > 0) {
-                collapse_nb(cnb0, cne0);
-                collapse_nb(cnb1, cne1);
-                collapse(cne0);
-                collapse(cne1);
-            }
-        }
-        {
-            int64_t ne0 = cne0[0];
-            int64_t ne1 = cne0[1];
-            int64_t ne2 = cne0[2];
-            int64_t ne3 = cne0[3];
-
-            int64_t ne10 = cne1[0];
-            int64_t ne11 = cne1[1];
-            int64_t ne12 = cne1[2];
-            int64_t ne13 = cne1[3];
-
-            size_t nb0 = cnb0[0];
-            size_t nb1 = cnb0[1];
-            size_t nb2 = cnb0[2];
-            size_t nb3 = cnb0[3];
-
-            size_t nb10 = cnb1[0];
-            size_t nb11 = cnb1[1];
-            size_t nb12 = cnb1[2];
-            size_t nb13 = cnb1[3];
-
-            size_t s0 = nb0 / sizeof(dst_t);
-            size_t s1 = nb1 / sizeof(dst_t);
-            size_t s2 = nb2 / sizeof(dst_t);
-            size_t s3 = nb3 / sizeof(dst_t);
-
-            size_t s10 = nb10 / sizeof(src1_t);
-            size_t s11 = nb11 / sizeof(src1_t);
-            size_t s12 = nb12 / sizeof(src1_t);
-            size_t s13 = nb13 / sizeof(src1_t);
-
-            GGML_ASSERT(s0 == 1);
-            GGML_ASSERT(s10 == 1);
-
-            const int block_size = 128;
-
-            int64_t hne0 = std::max(ne0/2LL, 1LL);
-
-            dim3 block_dims;
-            block_dims.x = std::min<unsigned int>(hne0, block_size);
-            block_dims.y = std::min<unsigned int>(ne1, block_size / block_dims.x);
-            block_dims.z = std::min(std::min<unsigned int>(ne2*ne3, block_size / block_dims.x / block_dims.y), 64U);
-
-            dim3 block_nums(
-                (hne0 + block_dims.x - 1) / block_dims.x,
-                (ne1 + block_dims.y - 1) / block_dims.y,
-                (ne2*ne3 + block_dims.z - 1) / block_dims.z
-            );
-
-            if (block_nums.z > 65535) {
-                // this is the maximum number of blocks in z direction, fallback to 1D grid kernel
-                int block_num = (ne0*ne1*ne2*ne3 + block_size - 1) / block_size;
-                k_bin_bcast_unravel<bin_op><<<block_num, block_size, 0, stream>>>(
-                    src0_dd, src1_dd, dst_dd,
-                    ne0, ne1, ne2, ne3,
-                    ne10, ne11, ne12, ne13,
-                    /* s0, */ s1, s2, s3,
-                    /* s10, */ s11, s12, s13);
-            } else {
-                k_bin_bcast<bin_op><<<block_nums, block_dims, 0, stream>>>(
-                    src0_dd, src1_dd, dst_dd,
-                    ne0, ne1, ne2, ne3,
-                    ne10, ne11, ne12, ne13,
-                    /* s0, */ s1, s2, s3,
-                    /* s10, */ s11, s12, s13);
-            }
-        }
-    }
-};
-
-static void acc_f32_cuda(const float * x, const float * y, float * dst, const int n_elements,
-    const int ne10, const int ne11, const int ne12,
-    const int nb1, const int nb2, const int offset, cudaStream_t stream) {
-    int num_blocks = (n_elements + CUDA_ACC_BLOCK_SIZE - 1) / CUDA_ACC_BLOCK_SIZE;
-    acc_f32<<<num_blocks, CUDA_ACC_BLOCK_SIZE, 0, stream>>>(x, y, dst, n_elements, ne10, ne11, ne12, nb1, nb2, offset);
-}
-
-static void gelu_f32_cuda(const float * x, float * dst, const int k, cudaStream_t stream) {
-    const int num_blocks = (k + CUDA_GELU_BLOCK_SIZE - 1) / CUDA_GELU_BLOCK_SIZE;
-    gelu_f32<<<num_blocks, CUDA_GELU_BLOCK_SIZE, 0, stream>>>(x, dst, k);
-}
-
-static void silu_f32_cuda(const float * x, float * dst, const int k, cudaStream_t stream) {
-    const int num_blocks = (k + CUDA_SILU_BLOCK_SIZE - 1) / CUDA_SILU_BLOCK_SIZE;
-    silu_f32<<<num_blocks, CUDA_SILU_BLOCK_SIZE, 0, stream>>>(x, dst, k);
-}
-
-static void gelu_quick_f32_cuda(const float * x, float * dst, const int k, cudaStream_t stream) {
-    const int num_blocks = (k + CUDA_GELU_BLOCK_SIZE - 1) / CUDA_GELU_BLOCK_SIZE;
-    gelu_quick_f32<<<num_blocks, CUDA_GELU_BLOCK_SIZE, 0, stream>>>(x, dst, k);
-}
-
-static void tanh_f32_cuda(const float * x, float * dst, const int k, cudaStream_t stream) {
-    const int num_blocks = (k + CUDA_TANH_BLOCK_SIZE - 1) / CUDA_TANH_BLOCK_SIZE;
-    tanh_f32<<<num_blocks, CUDA_TANH_BLOCK_SIZE, 0, stream>>>(x, dst, k);
-}
-
-static void relu_f32_cuda(const float * x, float * dst, const int k, cudaStream_t stream) {
-    const int num_blocks = (k + CUDA_RELU_BLOCK_SIZE - 1) / CUDA_RELU_BLOCK_SIZE;
-    relu_f32<<<num_blocks, CUDA_RELU_BLOCK_SIZE, 0, stream>>>(x, dst, k);
-}
-
-static void leaky_relu_f32_cuda(const float * x, float * dst, const int k, const float negative_slope, cudaStream_t stream) {
-    const int num_blocks = (k + CUDA_RELU_BLOCK_SIZE - 1) / CUDA_RELU_BLOCK_SIZE;
-    leaky_relu_f32<<<num_blocks, CUDA_RELU_BLOCK_SIZE, 0, stream>>>(x, dst, k, negative_slope);
-}
-
-static void sqr_f32_cuda(const float * x, float * dst, const int k, cudaStream_t stream) {
-    const int num_blocks = (k + CUDA_SQR_BLOCK_SIZE - 1) / CUDA_SQR_BLOCK_SIZE;
-    sqr_f32<<<num_blocks, CUDA_SQR_BLOCK_SIZE, 0, stream>>>(x, dst, k);
-}
-
-static void norm_f32_cuda(const float * x, float * dst, const int ncols, const int nrows, const float eps, cudaStream_t stream) {
-    GGML_ASSERT(ncols % WARP_SIZE == 0);
-    if (ncols < 1024) {
-        const dim3 block_dims(WARP_SIZE, 1, 1);
-        norm_f32<WARP_SIZE><<<nrows, block_dims, 0, stream>>>(x, dst, ncols, eps);
-    } else {
-        const dim3 block_dims(1024, 1, 1);
-        norm_f32<1024><<<nrows, block_dims, 0, stream>>>(x, dst, ncols, eps);
-    }
-}
-
-static void group_norm_f32_cuda(const float * x, float * dst, const int num_groups, const int group_size, const int ne_elements, cudaStream_t stream) {
-    static const float eps = 1e-6f;
-    if (group_size < 1024) {
-        const dim3 block_dims(WARP_SIZE, 1, 1);
-        group_norm_f32<WARP_SIZE><<<num_groups, block_dims, 0, stream>>>(x, dst, group_size, ne_elements, eps);
-    } else {
-        const dim3 block_dims(1024, 1, 1);
-        group_norm_f32<1024><<<num_groups, block_dims, 0, stream>>>(x, dst, group_size, ne_elements, eps);
-    }
-}
-
-static void concat_f32_cuda(const float * x, const float * y, float * dst, const int ne0, int ne1, int ne2, int ne02, cudaStream_t stream) {
-    int num_blocks = (ne0 + CUDA_CONCAT_BLOCK_SIZE - 1) / CUDA_CONCAT_BLOCK_SIZE;
-    dim3 gridDim(num_blocks, ne1, ne2);
-    concat_f32<<<gridDim, CUDA_CONCAT_BLOCK_SIZE, 0, stream>>>(x, y, dst, ne0, ne02);
-}
-
-static void upscale_f32_cuda(const float * x, float * dst, const int ne00, const int ne01, const int ne02, const int scale_factor, cudaStream_t stream) {
-    int ne0 = (ne00 * scale_factor);
-    int num_blocks = (ne0 + CUDA_UPSCALE_BLOCK_SIZE - 1) / CUDA_UPSCALE_BLOCK_SIZE;
-    dim3 gridDim(num_blocks, (ne01 * scale_factor), ne02);
-    upscale_f32<<<gridDim, CUDA_UPSCALE_BLOCK_SIZE, 0, stream>>>(x, dst, ne00, ne00 * ne01, scale_factor);
-}
-
-static void pad_f32_cuda(const float * x, float * dst,
-    const int ne00, const int ne01, const int ne02,
-    const int ne0, const int ne1, const int ne2, cudaStream_t stream) {
-    int num_blocks = (ne0 + CUDA_PAD_BLOCK_SIZE - 1) / CUDA_PAD_BLOCK_SIZE;
-    dim3 gridDim(num_blocks, ne1, ne2);
-    pad_f32<<<gridDim, CUDA_PAD_BLOCK_SIZE, 0, stream>>>(x, dst, ne0, ne00, ne01, ne02);
-}
-
-static void rms_norm_f32_cuda(const float * x, float * dst, const int ncols, const int nrows, const float eps, cudaStream_t stream) {
-    GGML_ASSERT(ncols % WARP_SIZE == 0);
-    if (ncols < 1024) {
-        const dim3 block_dims(WARP_SIZE, 1, 1);
-        rms_norm_f32<WARP_SIZE><<<nrows, block_dims, 0, stream>>>(x, dst, ncols, eps);
-    } else {
-        const dim3 block_dims(1024, 1, 1);
-        rms_norm_f32<1024><<<nrows, block_dims, 0, stream>>>(x, dst, ncols, eps);
-    }
-}
-
-static void quantize_row_q8_1_cuda(const float * x, void * vy, const int kx, const int ky, const int kx_padded, cudaStream_t stream) {
-    const int block_num_x = (kx_padded + CUDA_QUANTIZE_BLOCK_SIZE - 1) / CUDA_QUANTIZE_BLOCK_SIZE;
-    const dim3 num_blocks(block_num_x, ky, 1);
-    const dim3 block_size(CUDA_DEQUANTIZE_BLOCK_SIZE, 1, 1);
-    quantize_q8_1<<<num_blocks, block_size, 0, stream>>>(x, vy, kx, kx_padded);
-}
-
-template <int qk, int qr, dequantize_kernel_t dequantize_kernel, typename dst_t>
-static void dequantize_block_cuda(const void * __restrict__ vx, dst_t * __restrict__ y, const int k, cudaStream_t stream) {
-    const int num_blocks = (k + CUDA_DEQUANTIZE_BLOCK_SIZE - 1) / CUDA_DEQUANTIZE_BLOCK_SIZE;
-    dequantize_block<qk, qr, dequantize_kernel><<<num_blocks, CUDA_DEQUANTIZE_BLOCK_SIZE, 0, stream>>>(vx, y, k);
-}
-
-template<typename dst_t>
-static void dequantize_row_q2_K_cuda(const void * vx, dst_t * y, const int k, cudaStream_t stream) {
-    const int nb = k / QK_K;
-#if QK_K == 256
-    dequantize_block_q2_K<<<nb, 64, 0, stream>>>(vx, y);
-#else
-    dequantize_block_q2_K<<<nb, 32, 0, stream>>>(vx, y);
-#endif
-}
-
-template<typename dst_t>
-static void dequantize_row_q3_K_cuda(const void * vx, dst_t * y, const int k, cudaStream_t stream) {
-    const int nb = k / QK_K;
-#if QK_K == 256
-    dequantize_block_q3_K<<<nb, 64, 0, stream>>>(vx, y);
-#else
-    dequantize_block_q3_K<<<nb, 32, 0, stream>>>(vx, y);
-#endif
-}
-
-template<typename dst_t>
-static void dequantize_row_q4_K_cuda(const void * vx, dst_t * y, const int k, cudaStream_t stream) {
-    const int nb = k / QK_K;
-    dequantize_block_q4_K<<<nb, 32, 0, stream>>>(vx, y);
-}
-
-template<typename dst_t>
-static void dequantize_row_q5_K_cuda(const void * vx, dst_t * y, const int k, cudaStream_t stream) {
-    const int nb = k / QK_K;
-#if QK_K == 256
-    dequantize_block_q5_K<<<nb, 64, 0, stream>>>(vx, y);
-#else
-    dequantize_block_q5_K<<<nb, 32, 0, stream>>>(vx, y);
-#endif
-}
-
-template<typename dst_t>
-static void dequantize_row_q6_K_cuda(const void * vx, dst_t * y, const int k, cudaStream_t stream) {
-    const int nb = k / QK_K;
-#if QK_K == 256
-    dequantize_block_q6_K<<<nb, 64, 0, stream>>>(vx, y);
-#else
-    dequantize_block_q6_K<<<nb, 32, 0, stream>>>(vx, y);
-#endif
-}
-
-static to_fp16_cuda_t ggml_get_to_fp16_cuda(ggml_type type) {
-    switch (type) {
-        case GGML_TYPE_Q4_0:
-            return dequantize_block_cuda<QK4_0, QR4_0, dequantize_q4_0>;
-        case GGML_TYPE_Q4_1:
-            return dequantize_block_cuda<QK4_1, QR4_1, dequantize_q4_1>;
-        case GGML_TYPE_Q5_0:
-            return dequantize_block_cuda<QK5_0, QR5_0, dequantize_q5_0>;
-        case GGML_TYPE_Q5_1:
-            return dequantize_block_cuda<QK5_1, QR5_1, dequantize_q5_1>;
-        case GGML_TYPE_Q8_0:
-            return dequantize_block_cuda<QK8_0, QR8_0, dequantize_q8_0>;
-        case GGML_TYPE_Q2_K:
-            return dequantize_row_q2_K_cuda;
-        case GGML_TYPE_Q3_K:
-            return dequantize_row_q3_K_cuda;
-        case GGML_TYPE_Q4_K:
-            return dequantize_row_q4_K_cuda;
-        case GGML_TYPE_Q5_K:
-            return dequantize_row_q5_K_cuda;
-        case GGML_TYPE_Q6_K:
-            return dequantize_row_q6_K_cuda;
-        case GGML_TYPE_F32:
-            return dequantize_block_cuda<1, 1, convert_f32>;
-        default:
-            return nullptr;
-    }
-}
-
-static to_fp32_cuda_t ggml_get_to_fp32_cuda(ggml_type type) {
-    switch (type) {
-        case GGML_TYPE_Q4_0:
-            return dequantize_block_cuda<QK4_0, QR4_0, dequantize_q4_0>;
-        case GGML_TYPE_Q4_1:
-            return dequantize_block_cuda<QK4_1, QR4_1, dequantize_q4_1>;
-        case GGML_TYPE_Q5_0:
-            return dequantize_block_cuda<QK5_0, QR5_0, dequantize_q5_0>;
-        case GGML_TYPE_Q5_1:
-            return dequantize_block_cuda<QK5_1, QR5_1, dequantize_q5_1>;
-        case GGML_TYPE_Q8_0:
-            return dequantize_block_cuda<QK8_0, QR8_0, dequantize_q8_0>;
-        case GGML_TYPE_Q2_K:
-            return dequantize_row_q2_K_cuda;
-        case GGML_TYPE_Q3_K:
-            return dequantize_row_q3_K_cuda;
-        case GGML_TYPE_Q4_K:
-            return dequantize_row_q4_K_cuda;
-        case GGML_TYPE_Q5_K:
-            return dequantize_row_q5_K_cuda;
-        case GGML_TYPE_Q6_K:
-            return dequantize_row_q6_K_cuda;
-        case GGML_TYPE_F16:
-            return dequantize_block_cuda<1, 1, convert_f16>;
-        default:
-            return nullptr;
-    }
-}
-
-static void dequantize_mul_mat_vec_q4_0_cuda(const void * vx, const dfloat * y, float * dst, const int ncols, const int nrows, cudaStream_t stream) {
-    GGML_ASSERT(ncols % GGML_CUDA_DMMV_X == 0);
-    const int block_num_y = (nrows + GGML_CUDA_MMV_Y - 1) / GGML_CUDA_MMV_Y;
-    // the number of rows may exceed maximum grid size in the y or z dimensions, use the x dimension instead
-    const dim3 block_nums(block_num_y, 1, 1);
-    const dim3 block_dims(WARP_SIZE, GGML_CUDA_MMV_Y, 1);
-    dequantize_mul_mat_vec<QK4_0, QR4_0, dequantize_q4_0>
-        <<<block_nums, block_dims, 0, stream>>>(vx, y, dst, ncols, nrows);
-}
-
-static void dequantize_mul_mat_vec_q4_1_cuda(const void * vx, const dfloat * y, float * dst, const int ncols, const int nrows, cudaStream_t stream) {
-    GGML_ASSERT(ncols % GGML_CUDA_DMMV_X == 0);
-    const int block_num_y = (nrows + GGML_CUDA_MMV_Y - 1) / GGML_CUDA_MMV_Y;
-    const dim3 block_nums(block_num_y, 1, 1);
-    const dim3 block_dims(WARP_SIZE, GGML_CUDA_MMV_Y, 1);
-    dequantize_mul_mat_vec<QK4_1, QR4_1, dequantize_q4_1>
-        <<<block_nums, block_dims, 0, stream>>>(vx, y, dst, ncols, nrows);
-}
-
-static void dequantize_mul_mat_vec_q5_0_cuda(const void * vx, const dfloat * y, float * dst, const int ncols, const int nrows, cudaStream_t stream) {
-    GGML_ASSERT(ncols % GGML_CUDA_DMMV_X == 0);
-    const int block_num_y = (nrows + GGML_CUDA_MMV_Y - 1) / GGML_CUDA_MMV_Y;
-    const dim3 block_nums(block_num_y, 1, 1);
-    const dim3 block_dims(WARP_SIZE, GGML_CUDA_MMV_Y, 1);
-    dequantize_mul_mat_vec<QK5_0, QR5_0, dequantize_q5_0>
-        <<<block_nums, block_dims, 0, stream>>>(vx, y, dst, ncols, nrows);
-}
-
-static void dequantize_mul_mat_vec_q5_1_cuda(const void * vx, const dfloat * y, float * dst, const int ncols, const int nrows, cudaStream_t stream) {
-    GGML_ASSERT(ncols % GGML_CUDA_DMMV_X == 0);
-    const int block_num_y = (nrows + GGML_CUDA_MMV_Y - 1) / GGML_CUDA_MMV_Y;
-    const dim3 block_nums(block_num_y, 1, 1);
-    const dim3 block_dims(WARP_SIZE, GGML_CUDA_MMV_Y, 1);
-    dequantize_mul_mat_vec<QK5_1, QR5_1, dequantize_q5_1>
-        <<<block_nums, block_dims, 0, stream>>>(vx, y, dst, ncols, nrows);
-}
-
-static void dequantize_mul_mat_vec_q8_0_cuda(const void * vx, const dfloat * y, float * dst, const int ncols, const int nrows, cudaStream_t stream) {
-    GGML_ASSERT(ncols % GGML_CUDA_DMMV_X == 0);
-    const int block_num_y = (nrows + GGML_CUDA_MMV_Y - 1) / GGML_CUDA_MMV_Y;
-    const dim3 block_nums(block_num_y, 1, 1);
-    const dim3 block_dims(WARP_SIZE, GGML_CUDA_MMV_Y, 1);
-    dequantize_mul_mat_vec<QK8_0, QR8_0, dequantize_q8_0>
-        <<<block_nums, block_dims, 0, stream>>>(vx, y, dst, ncols, nrows);
-}
-
-static void dequantize_mul_mat_vec_q2_K_cuda(const void * vx, const float * y, float * dst, const int ncols, const int nrows, cudaStream_t stream) {
-    GGML_ASSERT(ncols % QK_K == 0);
-    const int ny = 2; // very slightly faster than 1 even when K_QUANTS_PER_ITERATION = 2
-    const int block_num_y = (nrows + ny - 1) / ny;
-    const dim3 block_nums(block_num_y, 1, 1);
-    const dim3 block_dims(32, ny, 1);
-    dequantize_mul_mat_vec_q2_k<<<block_nums, block_dims, 0, stream>>>(vx, y, dst, ncols, nrows);
-}
-
-static void dequantize_mul_mat_vec_q3_K_cuda(const void * vx, const float * y, float * dst, const int ncols, const int nrows, cudaStream_t stream) {
-    GGML_ASSERT(ncols % QK_K == 0);
-    const int ny = 2 / K_QUANTS_PER_ITERATION;
-    const int block_num_y = (nrows + ny - 1) / ny;
-    const dim3 block_nums(block_num_y, 1, 1);
-    const dim3 block_dims(32, ny, 1);
-    dequantize_mul_mat_vec_q3_k<<<block_nums, block_dims, 0, stream>>>(vx, y, dst, ncols, nrows);
-}
-
-static void dequantize_mul_mat_vec_q4_K_cuda(const void * vx, const float * y, float * dst, const int ncols, const int nrows, cudaStream_t stream) {
-    GGML_ASSERT(ncols % QK_K == 0);
-    const int ny = 2 / K_QUANTS_PER_ITERATION;
-    const int block_num_y = (nrows + ny - 1) / ny;
-    const dim3 block_nums(block_num_y, 1, 1);
-    const dim3 block_dims(32, ny, 1);
-    dequantize_mul_mat_vec_q4_k<<<block_nums, block_dims, 0, stream>>>(vx, y, dst, ncols, nrows);
-}
-
-static void dequantize_mul_mat_vec_q5_K_cuda(const void * vx, const float * y, float * dst, const int ncols, const int nrows, cudaStream_t stream) {
-    GGML_ASSERT(ncols % QK_K == 0);
-    const dim3 block_dims(32, 1, 1);
-    dequantize_mul_mat_vec_q5_k<<<nrows, block_dims, 0, stream>>>(vx, y, dst, ncols);
-}
-
-static void dequantize_mul_mat_vec_q6_K_cuda(const void * vx, const float * y, float * dst, const int ncols, const int nrows, cudaStream_t stream) {
-    GGML_ASSERT(ncols % QK_K == 0);
-    const int ny = 2 / K_QUANTS_PER_ITERATION;
-    const int block_num_y = (nrows + ny - 1) / ny;
-    const dim3 block_nums(block_num_y, 1, 1);
-    const dim3 block_dims(32, ny, 1);
-    dequantize_mul_mat_vec_q6_k<<<block_nums, block_dims, 0, stream>>>(vx, y, dst, ncols, nrows);
-}
-
-static void convert_mul_mat_vec_f16_cuda(const void * vx, const dfloat * y, float * dst, const int ncols, const int nrows, cudaStream_t stream) {
-    GGML_ASSERT(ncols % GGML_CUDA_DMMV_X == 0);
-    const int block_num_y = (nrows + GGML_CUDA_MMV_Y - 1) / GGML_CUDA_MMV_Y;
-    const dim3 block_nums(block_num_y, 1, 1);
-    const dim3 block_dims(WARP_SIZE, GGML_CUDA_MMV_Y, 1);
-    dequantize_mul_mat_vec<1, 1, convert_f16>
-        <<<block_nums, block_dims, 0, stream>>>(vx, y, dst, ncols, nrows);
-}
-
-static void mul_mat_vec_q4_0_q8_1_cuda(const void * vx, const void * vy, float * dst, const int ncols, const int nrows, cudaStream_t stream) {
-    GGML_ASSERT(ncols % QK4_0 == 0);
-    const int block_num_y = (nrows + GGML_CUDA_MMV_Y - 1) / GGML_CUDA_MMV_Y;
-    const dim3 block_nums(block_num_y, 1, 1);
-    const dim3 block_dims(WARP_SIZE, GGML_CUDA_MMV_Y, 1);
-    mul_mat_vec_q<QK4_0, QI4_0, block_q4_0, VDR_Q4_0_Q8_1_MMVQ, vec_dot_q4_0_q8_1>
-        <<<block_nums, block_dims, 0, stream>>>(vx, vy, dst, ncols, nrows);
-}
-
-static void mul_mat_vec_q4_1_q8_1_cuda(const void * vx, const void * vy, float * dst, const int ncols, const int nrows, cudaStream_t stream) {
-    GGML_ASSERT(ncols % QK4_1 == 0);
-    const int block_num_y = (nrows + GGML_CUDA_MMV_Y - 1) / GGML_CUDA_MMV_Y;
-    const dim3 block_nums(block_num_y, 1, 1);
-    const dim3 block_dims(WARP_SIZE, GGML_CUDA_MMV_Y, 1);
-    mul_mat_vec_q<QK4_0, QI4_1, block_q4_1, VDR_Q4_1_Q8_1_MMVQ, vec_dot_q4_1_q8_1>
-        <<<block_nums, block_dims, 0, stream>>>(vx, vy, dst, ncols, nrows);
-}
-
-static void mul_mat_vec_q5_0_q8_1_cuda(const void * vx, const void * vy, float * dst, const int ncols, const int nrows, cudaStream_t stream) {
-    GGML_ASSERT(ncols % QK5_0 == 0);
-    const int block_num_y = (nrows + GGML_CUDA_MMV_Y - 1) / GGML_CUDA_MMV_Y;
-    const dim3 block_nums(block_num_y, 1, 1);
-    const dim3 block_dims(WARP_SIZE, GGML_CUDA_MMV_Y, 1);
-    mul_mat_vec_q<QK5_0, QI5_0, block_q5_0, VDR_Q5_0_Q8_1_MMVQ, vec_dot_q5_0_q8_1>
-        <<<block_nums, block_dims, 0, stream>>>(vx, vy, dst, ncols, nrows);
-}
-
-static void mul_mat_vec_q5_1_q8_1_cuda(const void * vx, const void * vy, float * dst, const int ncols, const int nrows, cudaStream_t stream) {
-    GGML_ASSERT(ncols % QK5_1 == 0);
-    const int block_num_y = (nrows + GGML_CUDA_MMV_Y - 1) / GGML_CUDA_MMV_Y;
-    const dim3 block_nums(block_num_y, 1, 1);
-    const dim3 block_dims(WARP_SIZE, GGML_CUDA_MMV_Y, 1);
-    mul_mat_vec_q<QK5_1, QI5_1, block_q5_1, VDR_Q5_1_Q8_1_MMVQ, vec_dot_q5_1_q8_1>
-        <<<block_nums, block_dims, 0, stream>>>(vx, vy, dst, ncols, nrows);
-}
-
-static void mul_mat_vec_q8_0_q8_1_cuda(const void * vx, const void * vy, float * dst, const int ncols, const int nrows, cudaStream_t stream) {
-    GGML_ASSERT(ncols % QK8_0 == 0);
-    const int block_num_y = (nrows + GGML_CUDA_MMV_Y - 1) / GGML_CUDA_MMV_Y;
-    const dim3 block_nums(block_num_y, 1, 1);
-    const dim3 block_dims(WARP_SIZE, GGML_CUDA_MMV_Y, 1);
-    mul_mat_vec_q<QK8_0, QI8_0, block_q8_0, VDR_Q8_0_Q8_1_MMVQ, vec_dot_q8_0_q8_1>
-        <<<block_nums, block_dims, 0, stream>>>(vx, vy, dst, ncols, nrows);
-}
-
-static void mul_mat_vec_q2_K_q8_1_cuda(const void * vx, const void * vy, float * dst, const int ncols, const int nrows, cudaStream_t stream) {
-    GGML_ASSERT(ncols % QK_K == 0);
-    const int block_num_y = (nrows + GGML_CUDA_MMV_Y - 1) / GGML_CUDA_MMV_Y;
-    const dim3 block_nums(block_num_y, 1, 1);
-    const dim3 block_dims(WARP_SIZE, GGML_CUDA_MMV_Y, 1);
-    mul_mat_vec_q<QK_K, QI2_K, block_q2_K, VDR_Q2_K_Q8_1_MMVQ, vec_dot_q2_K_q8_1>
-        <<<block_nums, block_dims, 0, stream>>>(vx, vy, dst, ncols, nrows);
-}
-
-static void mul_mat_vec_q3_K_q8_1_cuda(const void * vx, const void * vy, float * dst, const int ncols, const int nrows, cudaStream_t stream) {
-    GGML_ASSERT(ncols % QK_K == 0);
-    const int block_num_y = (nrows + GGML_CUDA_MMV_Y - 1) / GGML_CUDA_MMV_Y;
-    const dim3 block_nums(block_num_y, 1, 1);
-    const dim3 block_dims(WARP_SIZE, GGML_CUDA_MMV_Y, 1);
-    mul_mat_vec_q<QK_K, QI3_K, block_q3_K, VDR_Q3_K_Q8_1_MMVQ, vec_dot_q3_K_q8_1>
-        <<<block_nums, block_dims, 0, stream>>>(vx, vy, dst, ncols, nrows);
-}
-
-static void mul_mat_vec_q4_K_q8_1_cuda(const void * vx, const void * vy, float * dst, const int ncols, const int nrows, cudaStream_t stream) {
-    GGML_ASSERT(ncols % QK_K == 0);
-    const int block_num_y = (nrows + GGML_CUDA_MMV_Y - 1) / GGML_CUDA_MMV_Y;
-    const dim3 block_nums(block_num_y, 1, 1);
-    const dim3 block_dims(WARP_SIZE, GGML_CUDA_MMV_Y, 1);
-    mul_mat_vec_q<QK_K, QI4_K, block_q4_K, VDR_Q4_K_Q8_1_MMVQ, vec_dot_q4_K_q8_1>
-        <<<block_nums, block_dims, 0, stream>>>(vx, vy, dst, ncols, nrows);
-}
-
-static void mul_mat_vec_q5_K_q8_1_cuda(const void * vx, const void * vy, float * dst, const int ncols, const int nrows, cudaStream_t stream) {
-    GGML_ASSERT(ncols % QK_K == 0);
-    const int block_num_y = (nrows + GGML_CUDA_MMV_Y - 1) / GGML_CUDA_MMV_Y;
-    const dim3 block_nums(block_num_y, 1, 1);
-    const dim3 block_dims(WARP_SIZE, GGML_CUDA_MMV_Y, 1);
-    mul_mat_vec_q<QK_K, QI5_K, block_q5_K, VDR_Q5_K_Q8_1_MMVQ, vec_dot_q5_K_q8_1>
-        <<<block_nums, block_dims, 0, stream>>>(vx, vy, dst, ncols, nrows);
-}
-
-static void mul_mat_vec_q6_K_q8_1_cuda(const void * vx, const void * vy, float * dst, const int ncols, const int nrows, cudaStream_t stream) {
-    GGML_ASSERT(ncols % QK_K == 0);
-    const int block_num_y = (nrows + GGML_CUDA_MMV_Y - 1) / GGML_CUDA_MMV_Y;
-    const dim3 block_nums(block_num_y, 1, 1);
-    const dim3 block_dims(WARP_SIZE, GGML_CUDA_MMV_Y, 1);
-    mul_mat_vec_q<QK_K, QI6_K, block_q6_K, VDR_Q6_K_Q8_1_MMVQ, vec_dot_q6_K_q8_1>
-        <<<block_nums, block_dims, 0, stream>>>(vx, vy, dst, ncols, nrows);
-}
-
-static void ggml_mul_mat_q4_0_q8_1_cuda(
-    const void * vx, const void * vy, float * dst, const int ncols_x, const int nrows_x,
-    const int ncols_y, const int nrows_y, const int nrows_dst, cudaStream_t stream) {
-
-    int id;
-    CUDA_CHECK(cudaGetDevice(&id));
-    const int compute_capability = g_device_caps[id].cc;
-
-    int mmq_x, mmq_y, nwarps;
-    if (compute_capability >= CC_RDNA2) {
-        mmq_x  =  MMQ_X_Q4_0_RDNA2;
-        mmq_y  =  MMQ_Y_Q4_0_RDNA2;
-        nwarps = NWARPS_Q4_0_RDNA2;
-    } else if (compute_capability >= CC_OFFSET_AMD) {
-        mmq_x  =  MMQ_X_Q4_0_RDNA1;
-        mmq_y  =  MMQ_Y_Q4_0_RDNA1;
-        nwarps = NWARPS_Q4_0_RDNA1;
-    } else if (compute_capability >= CC_VOLTA) {
-        mmq_x  =  MMQ_X_Q4_0_AMPERE;
-        mmq_y  =  MMQ_Y_Q4_0_AMPERE;
-        nwarps = NWARPS_Q4_0_AMPERE;
-    } else if (compute_capability >= MIN_CC_DP4A) {
-        mmq_x  =  MMQ_X_Q4_0_PASCAL;
-        mmq_y  =  MMQ_Y_Q4_0_PASCAL;
-        nwarps = NWARPS_Q4_0_PASCAL;
-    } else {
-        GGML_ASSERT(false);
-    }
-
-    const int block_num_x = (nrows_x + mmq_y - 1) / mmq_y;
-    const int block_num_y = (ncols_y + mmq_x - 1) / mmq_x;
-    const dim3 block_nums(block_num_x, block_num_y, 1);
-    const dim3 block_dims(WARP_SIZE, nwarps, 1);
-
-    if (nrows_x % mmq_y == 0) {
-        const bool need_check = false;
-        mul_mat_q4_0<need_check><<<block_nums, block_dims, 0, stream>>>
-            (vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst);
-    } else {
-        const bool need_check = true;
-        mul_mat_q4_0<need_check><<<block_nums, block_dims, 0, stream>>>
-            (vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst);
-    }
-}
-
-static void ggml_mul_mat_q4_1_q8_1_cuda(
-    const void * vx, const void * vy, float * dst, const int ncols_x, const int nrows_x,
-    const int ncols_y, const int nrows_y, const int nrows_dst, cudaStream_t stream) {
-
-    int id;
-    CUDA_CHECK(cudaGetDevice(&id));
-    const int compute_capability = g_device_caps[id].cc;
-
-    int mmq_x, mmq_y, nwarps;
-    if (compute_capability >= CC_RDNA2) {
-        mmq_x  =  MMQ_X_Q4_1_RDNA2;
-        mmq_y  =  MMQ_Y_Q4_1_RDNA2;
-        nwarps = NWARPS_Q4_1_RDNA2;
-    } else if (compute_capability >= CC_OFFSET_AMD) {
-        mmq_x  =  MMQ_X_Q4_1_RDNA1;
-        mmq_y  =  MMQ_Y_Q4_1_RDNA1;
-        nwarps = NWARPS_Q4_1_RDNA1;
-    } else if (compute_capability >= CC_VOLTA) {
-        mmq_x  =  MMQ_X_Q4_1_AMPERE;
-        mmq_y  =  MMQ_Y_Q4_1_AMPERE;
-        nwarps = NWARPS_Q4_1_AMPERE;
-    } else if (compute_capability >= MIN_CC_DP4A) {
-        mmq_x  =  MMQ_X_Q4_1_PASCAL;
-        mmq_y  =  MMQ_Y_Q4_1_PASCAL;
-        nwarps = NWARPS_Q4_1_PASCAL;
-    } else {
-        GGML_ASSERT(false);
-    }
-
-    const int block_num_x = (nrows_x + mmq_y - 1) / mmq_y;
-    const int block_num_y = (ncols_y + mmq_x - 1) / mmq_x;
-    const dim3 block_nums(block_num_x, block_num_y, 1);
-    const dim3 block_dims(WARP_SIZE, nwarps, 1);
-
-    if (nrows_x % mmq_y == 0) {
-        const bool need_check = false;
-        mul_mat_q4_1<need_check><<<block_nums, block_dims, 0, stream>>>
-            (vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst);
-    } else {
-        const bool need_check = true;
-        mul_mat_q4_1<need_check><<<block_nums, block_dims, 0, stream>>>
-            (vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst);
-    }
-}
-
-static void ggml_mul_mat_q5_0_q8_1_cuda(
-    const void * vx, const void * vy, float * dst, const int ncols_x, const int nrows_x,
-    const int ncols_y, const int nrows_y, const int nrows_dst, cudaStream_t stream) {
-
-    int id;
-    CUDA_CHECK(cudaGetDevice(&id));
-    const int compute_capability = g_device_caps[id].cc;
-
-    int mmq_x, mmq_y, nwarps;
-    if (compute_capability >= CC_RDNA2) {
-        mmq_x  =  MMQ_X_Q5_0_RDNA2;
-        mmq_y  =  MMQ_Y_Q5_0_RDNA2;
-        nwarps = NWARPS_Q5_0_RDNA2;
-    } else if (compute_capability >= CC_OFFSET_AMD) {
-        mmq_x  =  MMQ_X_Q5_0_RDNA1;
-        mmq_y  =  MMQ_Y_Q5_0_RDNA1;
-        nwarps = NWARPS_Q5_0_RDNA1;
-    } else if (compute_capability >= CC_VOLTA) {
-        mmq_x  =  MMQ_X_Q5_0_AMPERE;
-        mmq_y  =  MMQ_Y_Q5_0_AMPERE;
-        nwarps = NWARPS_Q5_0_AMPERE;
-    } else if (compute_capability >= MIN_CC_DP4A) {
-        mmq_x  =  MMQ_X_Q5_0_PASCAL;
-        mmq_y  =  MMQ_Y_Q5_0_PASCAL;
-        nwarps = NWARPS_Q5_0_PASCAL;
-    } else {
-        GGML_ASSERT(false);
-    }
-
-    const int block_num_x = (nrows_x + mmq_y - 1) / mmq_y;
-    const int block_num_y = (ncols_y + mmq_x - 1) / mmq_x;
-    const dim3 block_nums(block_num_x, block_num_y, 1);
-    const dim3 block_dims(WARP_SIZE, nwarps, 1);
-
-    if (nrows_x % mmq_y == 0) {
-        const bool need_check = false;
-        mul_mat_q5_0<need_check><<<block_nums, block_dims, 0, stream>>>
-            (vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst);
-    } else {
-        const bool need_check = true;
-        mul_mat_q5_0<need_check><<<block_nums, block_dims, 0, stream>>>
-            (vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst);
-    }
-}
-
-static void ggml_mul_mat_q5_1_q8_1_cuda(
-    const void * vx, const void * vy, float * dst, const int ncols_x, const int nrows_x,
-    const int ncols_y, const int nrows_y, const int nrows_dst, cudaStream_t stream) {
-
-    int id;
-    CUDA_CHECK(cudaGetDevice(&id));
-    const int compute_capability = g_device_caps[id].cc;
-
-    int mmq_x, mmq_y, nwarps;
-    if (compute_capability >= CC_RDNA2) {
-        mmq_x  =  MMQ_X_Q5_1_RDNA2;
-        mmq_y  =  MMQ_Y_Q5_1_RDNA2;
-        nwarps = NWARPS_Q5_1_RDNA2;
-    } else if (compute_capability >= CC_OFFSET_AMD) {
-        mmq_x  =  MMQ_X_Q5_1_RDNA1;
-        mmq_y  =  MMQ_Y_Q5_1_RDNA1;
-        nwarps = NWARPS_Q5_1_RDNA1;
-    } else if (compute_capability >= CC_VOLTA) {
-        mmq_x  =  MMQ_X_Q5_1_AMPERE;
-        mmq_y  =  MMQ_Y_Q5_1_AMPERE;
-        nwarps = NWARPS_Q5_1_AMPERE;
-    } else if (compute_capability >= MIN_CC_DP4A) {
-        mmq_x  =  MMQ_X_Q5_1_PASCAL;
-        mmq_y  =  MMQ_Y_Q5_1_PASCAL;
-        nwarps = NWARPS_Q5_1_PASCAL;
-    } else {
-        GGML_ASSERT(false);
-    }
-
-    const int block_num_x = (nrows_x + mmq_y - 1) / mmq_y;
-    const int block_num_y = (ncols_y + mmq_x - 1) / mmq_x;
-    const dim3 block_nums(block_num_x, block_num_y, 1);
-    const dim3 block_dims(WARP_SIZE, nwarps, 1);
-
-    if (nrows_x % mmq_y == 0) {
-        const bool need_check = false;
-        mul_mat_q5_1<need_check><<<block_nums, block_dims, 0, stream>>>
-            (vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst);
-    } else {
-        const bool need_check = true;
-        mul_mat_q5_1<need_check><<<block_nums, block_dims, 0, stream>>>
-            (vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst);
-    }
-}
-
-static void ggml_mul_mat_q8_0_q8_1_cuda(
-    const void * vx, const void * vy, float * dst, const int ncols_x, const int nrows_x,
-    const int ncols_y, const int nrows_y, const int nrows_dst, cudaStream_t stream) {
-
-    int id;
-    CUDA_CHECK(cudaGetDevice(&id));
-    const int compute_capability = g_device_caps[id].cc;
-
-    int mmq_x, mmq_y, nwarps;
-    if (compute_capability >= CC_RDNA2) {
-        mmq_x  =  MMQ_X_Q8_0_RDNA2;
-        mmq_y  =  MMQ_Y_Q8_0_RDNA2;
-        nwarps = NWARPS_Q8_0_RDNA2;
-    } else if (compute_capability >= CC_OFFSET_AMD) {
-        mmq_x  =  MMQ_X_Q8_0_RDNA1;
-        mmq_y  =  MMQ_Y_Q8_0_RDNA1;
-        nwarps = NWARPS_Q8_0_RDNA1;
-    } else if (compute_capability >= CC_VOLTA) {
-        mmq_x  =  MMQ_X_Q8_0_AMPERE;
-        mmq_y  =  MMQ_Y_Q8_0_AMPERE;
-        nwarps = NWARPS_Q8_0_AMPERE;
-    } else if (compute_capability >= MIN_CC_DP4A) {
-        mmq_x  =  MMQ_X_Q8_0_PASCAL;
-        mmq_y  =  MMQ_Y_Q8_0_PASCAL;
-        nwarps = NWARPS_Q8_0_PASCAL;
-    } else {
-        GGML_ASSERT(false);
-    }
-
-    const int block_num_x = (nrows_x + mmq_y - 1) / mmq_y;
-    const int block_num_y = (ncols_y + mmq_x - 1) / mmq_x;
-    const dim3 block_nums(block_num_x, block_num_y, 1);
-    const dim3 block_dims(WARP_SIZE, nwarps, 1);
-
-    if (nrows_x % mmq_y == 0) {
-        const bool need_check = false;
-        mul_mat_q8_0<need_check><<<block_nums, block_dims, 0, stream>>>
-            (vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst);
-    } else {
-        const bool need_check = true;
-        mul_mat_q8_0<need_check><<<block_nums, block_dims, 0, stream>>>
-            (vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst);
-    }
-}
-
-static void ggml_mul_mat_q2_K_q8_1_cuda(
-    const void * vx, const void * vy, float * dst, const int ncols_x, const int nrows_x,
-    const int ncols_y, const int nrows_y, const int nrows_dst, cudaStream_t stream) {
-
-    int id;
-    CUDA_CHECK(cudaGetDevice(&id));
-    const int compute_capability = g_device_caps[id].cc;
-
-    int mmq_x, mmq_y, nwarps;
-    if (compute_capability >= CC_RDNA2) {
-        mmq_x  =  MMQ_X_Q2_K_RDNA2;
-        mmq_y  =  MMQ_Y_Q2_K_RDNA2;
-        nwarps = NWARPS_Q2_K_RDNA2;
-    } else if (compute_capability >= CC_OFFSET_AMD) {
-        mmq_x  =  MMQ_X_Q2_K_RDNA1;
-        mmq_y  =  MMQ_Y_Q2_K_RDNA1;
-        nwarps = NWARPS_Q2_K_RDNA1;
-    } else if (compute_capability >= CC_VOLTA) {
-        mmq_x  =  MMQ_X_Q2_K_AMPERE;
-        mmq_y  =  MMQ_Y_Q2_K_AMPERE;
-        nwarps = NWARPS_Q2_K_AMPERE;
-    } else if (compute_capability >= MIN_CC_DP4A) {
-        mmq_x  =  MMQ_X_Q2_K_PASCAL;
-        mmq_y  =  MMQ_Y_Q2_K_PASCAL;
-        nwarps = NWARPS_Q2_K_PASCAL;
-    } else {
-        GGML_ASSERT(false);
-    }
-
-    const int block_num_x = (nrows_x + mmq_y - 1) / mmq_y;
-    const int block_num_y = (ncols_y + mmq_x - 1) / mmq_x;
-    const dim3 block_nums(block_num_x, block_num_y, 1);
-    const dim3 block_dims(WARP_SIZE, nwarps, 1);
-
-    if (nrows_x % mmq_y == 0) {
-        const bool need_check = false;
-        mul_mat_q2_K<need_check><<<block_nums, block_dims, 0, stream>>>
-            (vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst);
-    } else {
-        const bool need_check = true;
-        mul_mat_q2_K<need_check><<<block_nums, block_dims, 0, stream>>>
-            (vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst);
-    }
-}
-
-static void ggml_mul_mat_q3_K_q8_1_cuda(
-    const void * vx, const void * vy, float * dst, const int ncols_x, const int nrows_x,
-    const int ncols_y, const int nrows_y, const int nrows_dst, cudaStream_t stream) {
-
-#if QK_K == 256
-
-    int id;
-    CUDA_CHECK(cudaGetDevice(&id));
-    const int compute_capability = g_device_caps[id].cc;
-
-    int mmq_x, mmq_y, nwarps;
-    if (compute_capability >= CC_RDNA2) {
-        mmq_x  =  MMQ_X_Q3_K_RDNA2;
-        mmq_y  =  MMQ_Y_Q3_K_RDNA2;
-        nwarps = NWARPS_Q3_K_RDNA2;
-    } else if (compute_capability >= CC_OFFSET_AMD) {
-        mmq_x  =  MMQ_X_Q3_K_RDNA1;
-        mmq_y  =  MMQ_Y_Q3_K_RDNA1;
-        nwarps = NWARPS_Q3_K_RDNA1;
-    } else if (compute_capability >= CC_VOLTA) {
-        mmq_x  =  MMQ_X_Q3_K_AMPERE;
-        mmq_y  =  MMQ_Y_Q3_K_AMPERE;
-        nwarps = NWARPS_Q3_K_AMPERE;
-    } else if (compute_capability >= MIN_CC_DP4A) {
-        mmq_x  =  MMQ_X_Q3_K_PASCAL;
-        mmq_y  =  MMQ_Y_Q3_K_PASCAL;
-        nwarps = NWARPS_Q3_K_PASCAL;
-    } else {
-        GGML_ASSERT(false);
-    }
-
-    const int block_num_x = (nrows_x + mmq_y - 1) / mmq_y;
-    const int block_num_y = (ncols_y + mmq_x - 1) / mmq_x;
-    const dim3 block_nums(block_num_x, block_num_y, 1);
-    const dim3 block_dims(WARP_SIZE, nwarps, 1);
-
-    if (nrows_x % mmq_y == 0) {
-        const bool need_check = false;
-        mul_mat_q3_K<need_check><<<block_nums, block_dims, 0, stream>>>
-            (vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst);
-    } else {
-        const bool need_check = true;
-        mul_mat_q3_K<need_check><<<block_nums, block_dims, 0, stream>>>
-            (vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst);
-    }
-#endif
-}
-
-static void ggml_mul_mat_q4_K_q8_1_cuda(
-    const void * vx, const void * vy, float * dst, const int ncols_x, const int nrows_x,
-    const int ncols_y, const int nrows_y, const int nrows_dst, cudaStream_t stream) {
-
-    int id;
-    CUDA_CHECK(cudaGetDevice(&id));
-    const int compute_capability = g_device_caps[id].cc;
-
-    int mmq_x, mmq_y, nwarps;
-    if (compute_capability >= CC_RDNA2) {
-        mmq_x  =  MMQ_X_Q4_K_RDNA2;
-        mmq_y  =  MMQ_Y_Q4_K_RDNA2;
-        nwarps = NWARPS_Q4_K_RDNA2;
-    } else if (compute_capability >= CC_OFFSET_AMD) {
-        mmq_x  =  MMQ_X_Q4_K_RDNA1;
-        mmq_y  =  MMQ_Y_Q4_K_RDNA1;
-        nwarps = NWARPS_Q4_K_RDNA1;
-    } else if (compute_capability >= CC_VOLTA) {
-        mmq_x  =  MMQ_X_Q4_K_AMPERE;
-        mmq_y  =  MMQ_Y_Q4_K_AMPERE;
-        nwarps = NWARPS_Q4_K_AMPERE;
-    } else if (compute_capability >= MIN_CC_DP4A) {
-        mmq_x  =  MMQ_X_Q4_K_PASCAL;
-        mmq_y  =  MMQ_Y_Q4_K_PASCAL;
-        nwarps = NWARPS_Q4_K_PASCAL;
-    } else {
-        GGML_ASSERT(false);
-    }
-
-    const int block_num_x = (nrows_x + mmq_y - 1) / mmq_y;
-    const int block_num_y = (ncols_y + mmq_x - 1) / mmq_x;
-    const dim3 block_nums(block_num_x, block_num_y, 1);
-    const dim3 block_dims(WARP_SIZE, nwarps, 1);
-
-    if (nrows_x % mmq_y == 0) {
-        const bool need_check = false;
-        mul_mat_q4_K<need_check><<<block_nums, block_dims, 0, stream>>>
-            (vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst);
-    } else {
-        const bool need_check = true;
-        mul_mat_q4_K<need_check><<<block_nums, block_dims, 0, stream>>>
-            (vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst);
-    }
-}
-
-static void ggml_mul_mat_q5_K_q8_1_cuda(
-    const void * vx, const void * vy, float * dst, const int ncols_x, const int nrows_x,
-    const int ncols_y, const int nrows_y, const int nrows_dst, cudaStream_t stream) {
-
-    int id;
-    CUDA_CHECK(cudaGetDevice(&id));
-    const int compute_capability = g_device_caps[id].cc;
-
-    int mmq_x, mmq_y, nwarps;
-    if (compute_capability >= CC_RDNA2) {
-        mmq_x  =  MMQ_X_Q5_K_RDNA2;
-        mmq_y  =  MMQ_Y_Q5_K_RDNA2;
-        nwarps = NWARPS_Q5_K_RDNA2;
-    } else if (compute_capability >= CC_OFFSET_AMD) {
-        mmq_x  =  MMQ_X_Q5_K_RDNA1;
-        mmq_y  =  MMQ_Y_Q5_K_RDNA1;
-        nwarps = NWARPS_Q5_K_RDNA1;
-    } else if (compute_capability >= CC_VOLTA) {
-        mmq_x  =  MMQ_X_Q5_K_AMPERE;
-        mmq_y  =  MMQ_Y_Q5_K_AMPERE;
-        nwarps = NWARPS_Q5_K_AMPERE;
-    } else if (compute_capability >= MIN_CC_DP4A) {
-        mmq_x  =  MMQ_X_Q5_K_PASCAL;
-        mmq_y  =  MMQ_Y_Q5_K_PASCAL;
-        nwarps = NWARPS_Q5_K_PASCAL;
-    } else {
-        GGML_ASSERT(false);
-    }
-
-    const int block_num_x = (nrows_x + mmq_y - 1) / mmq_y;
-    const int block_num_y = (ncols_y + mmq_x - 1) / mmq_x;
-    const dim3 block_nums(block_num_x, block_num_y, 1);
-    const dim3 block_dims(WARP_SIZE, nwarps, 1);
-
-    if (nrows_x % mmq_y == 0) {
-        const bool need_check = false;
-        mul_mat_q5_K<need_check><<<block_nums, block_dims, 0, stream>>>
-            (vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst);
-    } else {
-        const bool need_check = true;
-        mul_mat_q5_K<need_check><<<block_nums, block_dims, 0, stream>>>
-            (vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst);
-    }
-}
-
-static void ggml_mul_mat_q6_K_q8_1_cuda(
-    const void * vx, const void * vy, float * dst, const int ncols_x, const int nrows_x,
-    const int ncols_y, const int nrows_y, const int nrows_dst, cudaStream_t stream) {
-
-    int id;
-    CUDA_CHECK(cudaGetDevice(&id));
-    const int compute_capability = g_device_caps[id].cc;
-
-    int mmq_x, mmq_y, nwarps;
-    if (compute_capability >= CC_RDNA2) {
-        mmq_x  =  MMQ_X_Q6_K_RDNA2;
-        mmq_y  =  MMQ_Y_Q6_K_RDNA2;
-        nwarps = NWARPS_Q6_K_RDNA2;
-    } else if (compute_capability >= CC_OFFSET_AMD) {
-        mmq_x  =  MMQ_X_Q6_K_RDNA1;
-        mmq_y  =  MMQ_Y_Q6_K_RDNA1;
-        nwarps = NWARPS_Q6_K_RDNA1;
-    } else if (compute_capability >= CC_VOLTA) {
-        mmq_x  =  MMQ_X_Q6_K_AMPERE;
-        mmq_y  =  MMQ_Y_Q6_K_AMPERE;
-        nwarps = NWARPS_Q6_K_AMPERE;
-    } else if (compute_capability >= MIN_CC_DP4A) {
-        mmq_x  =  MMQ_X_Q6_K_PASCAL;
-        mmq_y  =  MMQ_Y_Q6_K_PASCAL;
-        nwarps = NWARPS_Q6_K_PASCAL;
-    } else {
-        GGML_ASSERT(false);
-    }
-
-    const int block_num_x = (nrows_x + mmq_y - 1) / mmq_y;
-    const int block_num_y = (ncols_y + mmq_x - 1) / mmq_x;
-    const dim3 block_nums(block_num_x, block_num_y, 1);
-    const dim3 block_dims(WARP_SIZE, nwarps, 1);
-
-    if (nrows_x % mmq_y == 0) {
-        const bool need_check = false;
-        mul_mat_q6_K<need_check><<<block_nums, block_dims, 0, stream>>>
-            (vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst);
-    } else {
-        const bool need_check = true;
-        mul_mat_q6_K<need_check><<<block_nums, block_dims, 0, stream>>>
-            (vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst);
-    }
-}
-
-static void ggml_mul_mat_p021_f16_f32_cuda(
-    const void * vx, const float * y, float * dst, const int ncols_x, const int nrows_x,
-    const int nchannels_x, const int nchannels_y, cudaStream_t stream) {
-
-    const dim3 block_nums(1, nrows_x, nchannels_y);
-    const dim3 block_dims(WARP_SIZE, 1, 1);
-    mul_mat_p021_f16_f32<<<block_nums, block_dims, 0, stream>>>(vx, y, dst, ncols_x, nrows_x, nchannels_x, nchannels_y);
-}
-
-static void ggml_mul_mat_vec_nc_f16_f32_cuda(
-    const void * vx, const float * y, float * dst, const int ncols_x, const int nrows_x, const int row_stride_x,
-    const int nchannels_x, const int nchannels_y, const int channel_stride_x, cudaStream_t stream) {
-
-    const dim3 block_nums(1, nrows_x, nchannels_y);
-    const dim3 block_dims(WARP_SIZE, 1, 1);
-    mul_mat_vec_nc_f16_f32<<<block_nums, block_dims, 0, stream>>>
-        (vx, y, dst, ncols_x, nrows_x, row_stride_x, channel_stride_x, nchannels_y/nchannels_x);
-}
-
-static void ggml_cpy_f32_f32_cuda(
-    const char * cx, char * cdst, const int ne,
-    const int ne00, const int ne01, const int nb00, const int nb01, const int nb02,
-    const int ne10, const int ne11, const int nb10, const int nb11, const int nb12, cudaStream_t stream) {
-
-    const int num_blocks = (ne + CUDA_CPY_BLOCK_SIZE - 1) / CUDA_CPY_BLOCK_SIZE;
-    cpy_f32_f16<cpy_1_f32_f32><<<num_blocks, CUDA_CPY_BLOCK_SIZE, 0, stream>>>
-        (cx, cdst, ne, ne00, ne01, nb00, nb01, nb02, ne10, ne11, nb10, nb11, nb12);
-}
-
-static void ggml_cpy_f32_f16_cuda(
-    const char * cx, char * cdst, const int ne,
-    const int ne00, const int ne01, const int nb00, const int nb01, const int nb02,
-    const int ne10, const int ne11, const int nb10, const int nb11, const int nb12, cudaStream_t stream) {
-
-    const int num_blocks = (ne + CUDA_CPY_BLOCK_SIZE - 1) / CUDA_CPY_BLOCK_SIZE;
-    cpy_f32_f16<cpy_1_f32_f16><<<num_blocks, CUDA_CPY_BLOCK_SIZE, 0, stream>>>
-        (cx, cdst, ne, ne00, ne01, nb00, nb01, nb02, ne10, ne11, nb10, nb11, nb12);
-}
-
-static void ggml_cpy_f32_q8_0_cuda(
-    const char * cx, char * cdst, const int ne,
-    const int ne00, const int ne01, const int nb00, const int nb01, const int nb02,
-    const int ne10, const int ne11, const int nb10, const int nb11, const int nb12, cudaStream_t stream) {
-
-    GGML_ASSERT(ne % QK8_0 == 0);
-    const int num_blocks = ne / QK8_0;
-    cpy_f32_q<cpy_blck_f32_q8_0, QK8_0><<<num_blocks, 1, 0, stream>>>
-        (cx, cdst, ne, ne00, ne01, nb00, nb01, nb02, ne10, ne11, nb10, nb11, nb12);
-}
-
-static void ggml_cpy_f32_q4_0_cuda(
-    const char * cx, char * cdst, const int ne,
-    const int ne00, const int ne01, const int nb00, const int nb01, const int nb02,
-    const int ne10, const int ne11, const int nb10, const int nb11, const int nb12, cudaStream_t stream) {
-
-    GGML_ASSERT(ne % QK4_0 == 0);
-    const int num_blocks = ne / QK4_0;
-    cpy_f32_q<cpy_blck_f32_q4_0, QK4_0><<<num_blocks, 1, 0, stream>>>
-        (cx, cdst, ne, ne00, ne01, nb00, nb01, nb02, ne10, ne11, nb10, nb11, nb12);
-}
-
-static void ggml_cpy_f32_q4_1_cuda(
-    const char * cx, char * cdst, const int ne,
-    const int ne00, const int ne01, const int nb00, const int nb01, const int nb02,
-    const int ne10, const int ne11, const int nb10, const int nb11, const int nb12, cudaStream_t stream) {
-
-    GGML_ASSERT(ne % QK4_1 == 0);
-    const int num_blocks = ne / QK4_1;
-    cpy_f32_q<cpy_blck_f32_q4_1, QK4_1><<<num_blocks, 1, 0, stream>>>
-        (cx, cdst, ne, ne00, ne01, nb00, nb01, nb02, ne10, ne11, nb10, nb11, nb12);
-}
-
-static void ggml_cpy_f16_f16_cuda(
-    const char * cx, char * cdst, const int ne,
-    const int ne00, const int ne01, const int nb00, const int nb01, const int nb02,
-    const int ne10, const int ne11, const int nb10, const int nb11, const int nb12, cudaStream_t stream) {
-
-    const int num_blocks = (ne + CUDA_CPY_BLOCK_SIZE - 1) / CUDA_CPY_BLOCK_SIZE;
-    cpy_f32_f16<cpy_1_f16_f16><<<num_blocks, CUDA_CPY_BLOCK_SIZE, 0, stream>>>
-        (cx, cdst, ne, ne00, ne01, nb00, nb01, nb02, ne10, ne11, nb10, nb11, nb12);
-}
-
-static void scale_f32_cuda(const float * x, float * dst, const float scale, const int k, cudaStream_t stream) {
-    const int num_blocks = (k + CUDA_SCALE_BLOCK_SIZE - 1) / CUDA_SCALE_BLOCK_SIZE;
-    scale_f32<<<num_blocks, CUDA_SCALE_BLOCK_SIZE, 0, stream>>>(x, dst, scale, k);
-}
-
-static void clamp_f32_cuda(const float * x, float * dst, const float min, const float max, const int k, cudaStream_t stream) {
-    const int num_blocks = (k + CUDA_CLAMP_BLOCK_SIZE - 1) / CUDA_CLAMP_BLOCK_SIZE;
-    clamp_f32<<<num_blocks, CUDA_CLAMP_BLOCK_SIZE, 0, stream>>>(x, dst, min, max, k);
-}
-
-template<typename T>
-static void rope_cuda(
-    const T * x, T * dst, int ncols, int nrows, const int32_t * pos, float freq_scale, int p_delta_rows,
-    float freq_base, float ext_factor, float attn_factor, rope_corr_dims corr_dims, cudaStream_t stream
-) {
-    GGML_ASSERT(ncols % 2 == 0);
-    const dim3 block_dims(1, CUDA_ROPE_BLOCK_SIZE, 1);
-    const int num_blocks_x = (ncols + 2*CUDA_ROPE_BLOCK_SIZE - 1) / (2*CUDA_ROPE_BLOCK_SIZE);
-    const dim3 block_nums(nrows, num_blocks_x, 1);
-    if (pos == nullptr) {
-        rope<T, false><<<block_nums, block_dims, 0, stream>>>(
-            x, dst, ncols, pos, freq_scale, p_delta_rows, freq_base, ext_factor, attn_factor, corr_dims
-        );
-    } else {
-        rope<T, true><<<block_nums, block_dims, 0, stream>>>(
-            x, dst, ncols, pos, freq_scale, p_delta_rows, freq_base, ext_factor, attn_factor, corr_dims
-        );
-    }
-}
-
-template<typename T>
-static void rope_neox_cuda(
-    const T * x, T * dst, int ncols, int n_dims, int nrows, const int32_t * pos, float freq_scale, int p_delta_rows,
-    float freq_base, float ext_factor, float attn_factor, rope_corr_dims corr_dims, cudaStream_t stream
-) {
-    GGML_ASSERT(ncols % 2 == 0);
-    const dim3 block_dims(1, CUDA_ROPE_BLOCK_SIZE, 1);
-    const int num_blocks_x = (ncols + 2*CUDA_ROPE_BLOCK_SIZE - 1) / (2*CUDA_ROPE_BLOCK_SIZE);
-    const dim3 block_nums(nrows, num_blocks_x, 1);
-
-    const float theta_scale = powf(freq_base, -2.0f/n_dims);
-    const float inv_ndims = -1.0f / n_dims;
-
-    if (pos == nullptr) {
-        rope_neox<T, false><<<block_nums, block_dims, 0, stream>>>(
-            x, dst, ncols, n_dims, pos, freq_scale, p_delta_rows, ext_factor, attn_factor, corr_dims,
-            theta_scale, inv_ndims
-        );
-    } else {
-        rope_neox<T, true><<<block_nums, block_dims, 0, stream>>>(
-            x, dst, ncols, n_dims, pos, freq_scale, p_delta_rows, ext_factor, attn_factor, corr_dims,
-            theta_scale, inv_ndims
-        );
-    }
-}
-
-static void rope_glm_f32_cuda(
-    const float * x, float * dst, int ncols, int nrows, const int32_t * pos, float freq_scale, int p_delta_rows,
-    float freq_base, int n_ctx, cudaStream_t stream
-) {
-    GGML_ASSERT(ncols % 4 == 0);
-    const dim3 block_dims(CUDA_ROPE_BLOCK_SIZE/4, 1, 1);
-    const int num_blocks_x = (ncols + CUDA_ROPE_BLOCK_SIZE - 1) / CUDA_ROPE_BLOCK_SIZE;
-    const dim3 block_nums(num_blocks_x, nrows, 1);
-    rope_glm_f32<<<block_nums, block_dims, 0, stream>>>(x, dst, ncols, pos, freq_scale, p_delta_rows, freq_base, n_ctx);
-}
-
-static void alibi_f32_cuda(const float * x, float * dst, const int ncols, const int nrows,
-                           const int k_rows, const int n_heads_log2_floor, const float m0,
-                           const float m1, cudaStream_t stream) {
-    const dim3 block_dims(CUDA_ALIBI_BLOCK_SIZE, 1, 1);
-    const int num_blocks_x = (ncols + CUDA_ALIBI_BLOCK_SIZE - 1) / (CUDA_ALIBI_BLOCK_SIZE);
-    const dim3 block_nums(num_blocks_x, nrows, 1);
-    alibi_f32<<<block_nums, block_dims, 0, stream>>>(x, dst, ncols, k_rows, n_heads_log2_floor, m0, m1);
-}
-
-static void sum_rows_f32_cuda(const float * x, float * dst, const int ncols, const int nrows, cudaStream_t stream) {
-    const dim3 block_dims(WARP_SIZE, 1, 1);
-    const dim3 block_nums(1, nrows, 1);
-    k_sum_rows_f32<<<block_nums, block_dims, 0, stream>>>(x, dst, ncols);
-}
-
-static void argsort_f32_i32_cuda(const float * x, int * dst, const int ncols, const int nrows, ggml_sort_order order, cudaStream_t stream) {
-    // bitonic sort requires ncols to be power of 2
-    GGML_ASSERT((ncols & (ncols - 1)) == 0);
-
-    const dim3 block_dims(ncols, 1, 1);
-    const dim3 block_nums(1, nrows, 1);
-    if (order == GGML_SORT_ASC) {
-        k_argsort_f32_i32<GGML_SORT_ASC><<<block_nums, block_dims, 0, stream>>>(x, dst, ncols);
-    } else if (order == GGML_SORT_DESC) {
-        k_argsort_f32_i32<GGML_SORT_DESC><<<block_nums, block_dims, 0, stream>>>(x, dst, ncols);
-    } else {
-        GGML_ASSERT(false);
-    }
-}
-
-static void diag_mask_inf_f32_cuda(const float * x, float * dst, const int ncols_x, const int nrows_x, const int rows_per_channel, const int n_past, cudaStream_t stream) {
-    const dim3 block_dims(1, CUDA_DIAG_MASK_INF_BLOCK_SIZE, 1);
-    const int block_num_x = (ncols_x + CUDA_DIAG_MASK_INF_BLOCK_SIZE - 1) / CUDA_DIAG_MASK_INF_BLOCK_SIZE;
-    const dim3 block_nums(nrows_x, block_num_x, 1);
-    diag_mask_inf_f32<<<block_nums, block_dims, 0, stream>>>(x, dst, ncols_x, rows_per_channel, n_past);
-}
-
-static void soft_max_f32_cuda(const float * x, const float * y, float * dst, const int ncols_x, const int nrows_x, const int nrows_y, const float scale, cudaStream_t stream) {
-    int nth = WARP_SIZE;
-    while (nth < ncols_x && nth < CUDA_SOFT_MAX_BLOCK_SIZE) nth *= 2;
-    const dim3 block_dims(nth,     1, 1);
-    const dim3 block_nums(nrows_x, 1, 1);
-    soft_max_f32<<<block_nums, block_dims, 0, stream>>>(x, y, dst, ncols_x, nrows_y, scale);
-}
-
-static void im2col_f32_f16_cuda(const float* x, half* dst,
-    int IW, int IH, int OW, int OH, int KW, int KH, int IC,
-    int offset_delta,
-    int s0,int s1,int p0,int p1,int d0,int d1, cudaStream_t stream) {
-    const int parallel_elements = OW * KW * KH;
-    const int num_blocks = (parallel_elements + CUDA_IM2COL_BLOCK_SIZE - 1) / CUDA_IM2COL_BLOCK_SIZE;
-    dim3 block_nums(num_blocks, OH, IC);
-    im2col_f32_f16<<<block_nums, CUDA_IM2COL_BLOCK_SIZE, 0, stream>>>(x, dst, offset_delta, IW, IH, OW, KW, KH, parallel_elements, (IC * KH * KW), s0, s1, p0, p1, d0, d1);
-}
-
-// buffer pool for cuda
-#define MAX_CUDA_BUFFERS 256
-
-struct scoped_spin_lock {
-    std::atomic_flag& lock;
-    scoped_spin_lock(std::atomic_flag& lock) : lock(lock) {
-        while (lock.test_and_set(std::memory_order_acquire)) {
-            ; // spin
-        }
-    }
-    ~scoped_spin_lock() {
-        lock.clear(std::memory_order_release);
-    }
-    scoped_spin_lock(const scoped_spin_lock&) = delete;
-    scoped_spin_lock& operator=(const scoped_spin_lock&) = delete;
-};
-
-static std::atomic_flag g_cuda_pool_lock = ATOMIC_FLAG_INIT;
-
-// #define DEBUG_CUDA_MALLOC
-struct ggml_cuda_buffer {
-    void * ptr = nullptr;
-    size_t size = 0;
-};
-
-static ggml_cuda_buffer g_cuda_buffer_pool[GGML_CUDA_MAX_DEVICES][MAX_CUDA_BUFFERS];
-static size_t g_cuda_pool_size[GGML_CUDA_MAX_DEVICES] = {0};
-
-static void * ggml_cuda_pool_malloc_leg(int device, size_t size, size_t * actual_size) {
-    scoped_spin_lock lock(g_cuda_pool_lock);
-#ifdef DEBUG_CUDA_MALLOC
-    int nnz = 0;
-    size_t max_size = 0;
-#endif
-    size_t best_diff = 1ull << 36;
-    int ibest = -1;
-    for (int i = 0; i < MAX_CUDA_BUFFERS; ++i) {
-        ggml_cuda_buffer& b = g_cuda_buffer_pool[device][i];
-        if (b.ptr != nullptr) {
-#ifdef DEBUG_CUDA_MALLOC
-            ++nnz;
-            if (b.size > max_size) max_size = b.size;
-#endif
-            if (b.size >= size) {
-                size_t diff = b.size - size;
-                if (diff < best_diff) {
-                    best_diff = diff;
-                    ibest = i;
-                    if (!best_diff) {
-                        void * ptr = b.ptr;
-                        *actual_size = b.size;
-                        b.ptr = nullptr;
-                        b.size = 0;
-                        return ptr;
-                    }
-                }
-            }
-        }
-    }
-    if (ibest >= 0) {
-        ggml_cuda_buffer& b = g_cuda_buffer_pool[device][ibest];
-        void * ptr = b.ptr;
-        *actual_size = b.size;
-        b.ptr = nullptr;
-        b.size = 0;
-        return ptr;
-    }
-    void * ptr;
-    size_t look_ahead_size = (size_t) (1.05 * size);
-    look_ahead_size = 256 * ((look_ahead_size + 255)/256);
-    ggml_cuda_set_device(device);
-    CUDA_CHECK(cudaMalloc((void **) &ptr, look_ahead_size));
-    *actual_size = look_ahead_size;
-    g_cuda_pool_size[device] += look_ahead_size;
-#ifdef DEBUG_CUDA_MALLOC
-    fprintf(stderr, "%s[%d]: %d buffers, max_size = %u MB, pool_size = %u MB, requested %u MB\n", __func__, id, nnz,
-            (uint32_t)(max_size/1024/1024), (uint32_t)(g_cuda_pool_size[id]/1024/1024), (uint32_t)(size/1024/1024));
-#endif
-    return ptr;
-}
-
-static void ggml_cuda_pool_free_leg(int device, void * ptr, size_t size) {
-    scoped_spin_lock lock(g_cuda_pool_lock);
-
-    for (int i = 0; i < MAX_CUDA_BUFFERS; ++i) {
-        ggml_cuda_buffer& b = g_cuda_buffer_pool[device][i];
-        if (b.ptr == nullptr) {
-            b.ptr = ptr;
-            b.size = size;
-            return;
-        }
-    }
-    fprintf(stderr, "WARNING: cuda buffer pool full, increase MAX_CUDA_BUFFERS\n");
-    ggml_cuda_set_device(device);
-    CUDA_CHECK(cudaFree(ptr));
-    g_cuda_pool_size[device] -= size;
-}
-
-#if !defined(GGML_USE_HIPBLAS)
-// pool with virtual memory
-static CUdeviceptr g_cuda_pool_addr[GGML_CUDA_MAX_DEVICES] = {0};
-static size_t g_cuda_pool_used[GGML_CUDA_MAX_DEVICES] = {0};
-static const size_t CUDA_POOL_VMM_MAX_SIZE = 1ull << 35; // 32 GB
-
-static void * ggml_cuda_pool_malloc_vmm(int device, size_t size, size_t * actual_size) {
-    scoped_spin_lock lock(g_cuda_pool_lock);
-
-    // round up the allocation size to the alignment to ensure that all allocations are aligned for all data types
-    const size_t alignment = 128;
-    size = alignment * ((size + alignment - 1) / alignment);
-
-    size_t avail = g_cuda_pool_size[device] - g_cuda_pool_used[device];
-
-    if (size > avail) {
-        // round up to the next multiple of the granularity
-        size_t reserve_size = size - avail;
-        const size_t granularity = g_device_caps[device].vmm_granularity;
-        reserve_size = granularity * ((reserve_size + granularity - 1) / granularity);
-
-        GGML_ASSERT(g_cuda_pool_size[device] + reserve_size <= CUDA_POOL_VMM_MAX_SIZE);
-
-        // allocate more physical memory
-        CUmemAllocationProp prop = {};
-        prop.type = CU_MEM_ALLOCATION_TYPE_PINNED;
-        prop.location.type = CU_MEM_LOCATION_TYPE_DEVICE;
-        prop.location.id = device;
-        CUmemGenericAllocationHandle handle;
-        CU_CHECK(cuMemCreate(&handle, reserve_size, &prop, 0));
-
-        // reserve virtual address space (if not already reserved)
-        if (g_cuda_pool_addr[device] == 0) {
-            CU_CHECK(cuMemAddressReserve(&g_cuda_pool_addr[device], CUDA_POOL_VMM_MAX_SIZE, 0, 0, 0));
-        }
-
-        // map at the end of the pool
-        CU_CHECK(cuMemMap(g_cuda_pool_addr[device] + g_cuda_pool_size[device], reserve_size, 0, handle, 0));
-
-        // the memory allocation handle is no longer needed after mapping
-        CU_CHECK(cuMemRelease(handle));
-
-        // set access
-        CUmemAccessDesc access = {};
-        access.location.type = CU_MEM_LOCATION_TYPE_DEVICE;
-        access.location.id = device;
-        access.flags = CU_MEM_ACCESS_FLAGS_PROT_READWRITE;
-        CU_CHECK(cuMemSetAccess(g_cuda_pool_addr[device] + g_cuda_pool_size[device], reserve_size, &access, 1));
-
-        // add to the pool
-        g_cuda_pool_size[device] += reserve_size;
-
-        //printf("cuda pool[%d]: size increased to %llu MB (reserved %llu MB)\n",
-        //       id, (unsigned long long) (g_cuda_pool_size[id]/1024/1024),
-        //       (unsigned long long) (reserve_size/1024/1024));
-    }
-
-    GGML_ASSERT(g_cuda_pool_addr[device] != 0);
-
-    void * ptr = (void *) (g_cuda_pool_addr[device] + g_cuda_pool_used[device]);
-    *actual_size = size;
-    g_cuda_pool_used[device] += size;
-
-#ifdef DEBUG_CUDA_MALLOC
-    printf("cuda pool[%d]: allocated %llu bytes at %llx [%s]\n", id, (unsigned long long) size, ptr);
-#endif
-
-    return ptr;
-}
-
-static void ggml_cuda_pool_free_vmm(int device, void * ptr, size_t size) {
-    scoped_spin_lock lock(g_cuda_pool_lock);
-
-#ifdef DEBUG_CUDA_MALLOC
-    printf("cuda pool[%d]: freed %llu bytes at %llx\n", id, (unsigned long long) size, ptr);
-#endif
-
-    g_cuda_pool_used[device] -= size;
-
-    // all deallocations must be in reverse order of the allocations
-    GGML_ASSERT(ptr == (void *) (g_cuda_pool_addr[device] + g_cuda_pool_used[device]));
-}
-
-static void * ggml_cuda_pool_malloc(int device, size_t size, size_t * actual_size) {
-    if (g_device_caps[device].vmm) {
-        return ggml_cuda_pool_malloc_vmm(device, size, actual_size);
-    } else {
-        return ggml_cuda_pool_malloc_leg(device, size, actual_size);
-    }
-}
-
-static void ggml_cuda_pool_free(int device, void * ptr, size_t size) {
-    if (g_device_caps[device].vmm) {
-        ggml_cuda_pool_free_vmm(device, ptr, size);
-    } else {
-        ggml_cuda_pool_free_leg(device, ptr, size);
-    }
-}
-#else
-#define ggml_cuda_pool_malloc ggml_cuda_pool_malloc_leg
-#define ggml_cuda_pool_free ggml_cuda_pool_free_leg
-#endif // !defined(GGML_USE_HIPBLAS)
-
-template<typename T>
-struct cuda_pool_alloc {
-    int device = -1;
-    T * ptr = nullptr;
-    size_t actual_size = 0;
-
-    // size is in number of elements
-    T * alloc(size_t size) {
-        GGML_ASSERT(ptr == nullptr);
-        CUDA_CHECK(cudaGetDevice(&device));
-        ptr = (T *) ggml_cuda_pool_malloc(device, size * sizeof(T), &this->actual_size);
-        return ptr;
-    }
-
-    cuda_pool_alloc(size_t size) {
-        alloc(size);
-    }
-
-    ~cuda_pool_alloc() {
-        if (ptr != nullptr) {
-            ggml_cuda_pool_free(device, ptr, actual_size);
-        }
-    }
-
-    T * get() {
-        return ptr;
-    }
-
-    cuda_pool_alloc() = default;
-    cuda_pool_alloc(const cuda_pool_alloc &) = delete;
-    cuda_pool_alloc(cuda_pool_alloc &&) = delete;
-    cuda_pool_alloc& operator=(const cuda_pool_alloc &) = delete;
-    cuda_pool_alloc& operator=(cuda_pool_alloc &&) = delete;
-};
-
-static bool g_cublas_loaded = false;
-
-bool ggml_cublas_loaded(void) {
-    return g_cublas_loaded;
-}
-
-void ggml_init_cublas() {
-    static bool initialized = false;
-
-    if (!initialized) {
-
-#ifdef __HIP_PLATFORM_AMD__
-        // Workaround for a rocBLAS bug when using multiple graphics cards:
-        // https://github.com/ROCmSoftwarePlatform/rocBLAS/issues/1346
-        rocblas_initialize();
-        CUDA_CHECK(cudaDeviceSynchronize());
-#endif
-
-        if (cudaGetDeviceCount(&g_device_count) != cudaSuccess) {
-            initialized = true;
-            g_cublas_loaded = false;
-            return;
-        }
-
-        GGML_ASSERT(g_device_count <= GGML_CUDA_MAX_DEVICES);
-        int64_t total_vram = 0;
-#if defined(GGML_CUDA_FORCE_MMQ)
-        fprintf(stderr, "%s: GGML_CUDA_FORCE_MMQ:   yes\n", __func__);
-#else
-        fprintf(stderr, "%s: GGML_CUDA_FORCE_MMQ:   no\n", __func__);
-#endif
-#if defined(CUDA_USE_TENSOR_CORES)
-        fprintf(stderr, "%s: CUDA_USE_TENSOR_CORES: yes\n", __func__);
-#else
-        fprintf(stderr, "%s: CUDA_USE_TENSOR_CORES: no\n", __func__);
-#endif
-        fprintf(stderr, "%s: found %d " GGML_CUDA_NAME " devices:\n", __func__, g_device_count);
-        for (int id = 0; id < g_device_count; ++id) {
-            int device_vmm = 0;
-
-#if !defined(GGML_USE_HIPBLAS)
-            CUdevice device;
-            CU_CHECK(cuDeviceGet(&device, id));
-            CU_CHECK(cuDeviceGetAttribute(&device_vmm, CU_DEVICE_ATTRIBUTE_VIRTUAL_MEMORY_MANAGEMENT_SUPPORTED, device));
-
-            if (device_vmm) {
-                CUmemAllocationProp alloc_prop = {};
-                alloc_prop.type = CU_MEM_ALLOCATION_TYPE_PINNED;
-                alloc_prop.location.type = CU_MEM_LOCATION_TYPE_DEVICE;
-                alloc_prop.location.id = id;
-                CU_CHECK(cuMemGetAllocationGranularity(&g_device_caps[id].vmm_granularity, &alloc_prop, CU_MEM_ALLOC_GRANULARITY_RECOMMENDED));
-            }
-#endif // !defined(GGML_USE_HIPBLAS)
-            g_device_caps[id].vmm = !!device_vmm;
-
-            cudaDeviceProp prop;
-            CUDA_CHECK(cudaGetDeviceProperties(&prop, id));
-            fprintf(stderr, "  Device %d: %s, compute capability %d.%d, VMM: %s\n", id, prop.name, prop.major, prop.minor, device_vmm ? "yes" : "no");
-
-            g_tensor_split[id] = total_vram;
-            total_vram += prop.totalGlobalMem;
-#if defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)
-            g_device_caps[id].cc = 100*prop.major + 10*prop.minor + CC_OFFSET_AMD;
-#else
-            g_device_caps[id].cc = 100*prop.major + 10*prop.minor;
-#endif // defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)
-        }
-        for (int id = 0; id < g_device_count; ++id) {
-            g_tensor_split[id] /= total_vram;
-        }
-
-        for (int id = 0; id < g_device_count; ++id) {
-            ggml_cuda_set_device(id);
-
-            // create cuda streams
-            for (int is = 0; is < MAX_STREAMS; ++is) {
-                CUDA_CHECK(cudaStreamCreateWithFlags(&g_cudaStreams[id][is], cudaStreamNonBlocking));
-            }
-
-            // create cublas handle
-            CUBLAS_CHECK(cublasCreate(&g_cublas_handles[id]));
-            CUBLAS_CHECK(cublasSetMathMode(g_cublas_handles[id], CUBLAS_TF32_TENSOR_OP_MATH));
-        }
-
-        // configure logging to stdout
-        // CUBLAS_CHECK(cublasLoggerConfigure(1, 1, 0, nullptr));
-
-        initialized = true;
-        g_cublas_loaded = true;
-    }
-}
-
-void ggml_cuda_set_tensor_split(const float * tensor_split) {
-    if (tensor_split == nullptr) {
-        return;
-    }
-    bool all_zero = true;
-    for (int i = 0; i < g_device_count; ++i) {
-        if (tensor_split[i] != 0.0f) {
-            all_zero = false;
-            break;
-        }
-    }
-    if (all_zero) {
-        return;
-    }
-    float split_sum = 0.0f;
-    for (int i = 0; i < g_device_count; ++i) {
-        g_tensor_split[i] = split_sum;
-        split_sum += tensor_split[i];
-    }
-    for (int i = 0; i < g_device_count; ++i) {
-        g_tensor_split[i] /= split_sum;
-    }
-}
-
-void * ggml_cuda_host_malloc(size_t size) {
-    if (getenv("GGML_CUDA_NO_PINNED") != nullptr) {
-        return nullptr;
-    }
-
-    void * ptr = nullptr;
-    cudaError_t err = cudaMallocHost((void **) &ptr, size);
-    if (err != cudaSuccess) {
-        // clear the error
-        cudaGetLastError();
-        fprintf(stderr, "WARNING: failed to allocate %.2f MB of pinned memory: %s\n",
-            size/1024.0/1024.0, cudaGetErrorString(err));
-        return nullptr;
-    }
-
-    return ptr;
-}
-
-void ggml_cuda_host_free(void * ptr) {
-    CUDA_CHECK(cudaFreeHost(ptr));
-}
-
-static cudaError_t ggml_cuda_cpy_tensor_2d(
-    void * dst, const struct ggml_tensor * src, int64_t i3, int64_t i2, int64_t i1_low, int64_t i1_high, cudaStream_t stream) {
-
-    cudaMemcpyKind kind;
-    char * src_ptr;
-    if (src->backend == GGML_BACKEND_CPU) {
-        kind = cudaMemcpyHostToDevice;
-        src_ptr = (char *) src->data;
-    } else if (src->backend == GGML_BACKEND_GPU || src->backend == GGML_BACKEND_GPU_SPLIT) {
-        GGML_ASSERT(src->backend != GGML_BACKEND_GPU_SPLIT || (i1_low == 0 && i1_high == src->ne[1]));
-        kind = cudaMemcpyDeviceToDevice;
-        ggml_tensor_extra_gpu * extra = (ggml_tensor_extra_gpu *) src->extra;
-        int id;
-        CUDA_CHECK(cudaGetDevice(&id));
-        src_ptr = (char *) extra->data_device[id];
-    } else {
-        GGML_ASSERT(false);
-    }
-    char * dst_ptr = (char *) dst;
-
-    const int64_t ne0 = src->ne[0];
-    const int64_t nb0 = src->nb[0];
-    const int64_t nb1 = src->nb[1];
-    const int64_t nb2 = src->nb[2];
-    const int64_t nb3 = src->nb[3];
-    const enum ggml_type type = src->type;
-    const int64_t ts = ggml_type_size(type);
-    const int64_t bs = ggml_blck_size(type);
-    int64_t i1_diff = i1_high - i1_low;
-
-    const char * x = src_ptr + i1_low*nb1 + i2*nb2 + i3*nb3;
-    if (nb0 == ts && nb1 == ts*ne0/bs) {
-        return cudaMemcpyAsync(dst_ptr, x, i1_diff*nb1, kind, stream);
-    } else if (nb0 == ts) {
-        return cudaMemcpy2DAsync(dst_ptr, ts*ne0/bs, x, nb1, ts*ne0/bs, i1_diff, kind, stream);
-    } else {
-        for (int64_t i1 = 0; i1 < i1_diff; i1++) {
-            const void * rx = (const void *) ((const char *) x + i1*nb1);
-            void * rd = (void *) (dst_ptr + i1*ts*ne0/bs);
-            // pretend the row is a matrix with cols=1
-            cudaError_t r = cudaMemcpy2DAsync(rd, ts/bs, rx, nb0, ts/bs, ne0, kind, stream);
-            if (r != cudaSuccess) return r;
-        }
-        return cudaSuccess;
-    }
-}
-
-static void ggml_cuda_op_get_rows(
-    const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst,
-    const float * src0_d, const float * src1_d, float * dst_d, cudaStream_t stream) {
-
-    GGML_ASSERT(src1->type == GGML_TYPE_I32);
-    GGML_ASSERT(dst->type == GGML_TYPE_F32);
-
-    GGML_ASSERT(src0->nb[0] == ggml_type_size(src0->type));
-    GGML_ASSERT(src1->nb[0] == ggml_type_size(src1->type));
-    GGML_ASSERT(dst->nb[0] == ggml_type_size(dst->type));
-
-    const int32_t * src1_i32 = (const int32_t *) src1_d;
-
-    switch (src0->type) {
-        case GGML_TYPE_F16:
-            get_rows_cuda_float(src0, src1, dst, (const half *)src0_d, src1_i32, dst_d, stream);
-            break;
-        case GGML_TYPE_F32:
-            get_rows_cuda_float(src0, src1, dst, src0_d, src1_i32, dst_d, stream);
-            break;
-        case GGML_TYPE_Q4_0:
-            get_rows_cuda<QK4_0, QR4_0, dequantize_q4_0>(src0, src1, dst, src0_d, src1_i32, dst_d, stream);
-            break;
-        case GGML_TYPE_Q4_1:
-            get_rows_cuda<QK4_1, QR4_1, dequantize_q4_1>(src0, src1, dst, src0_d, src1_i32, dst_d, stream);
-            break;
-        case GGML_TYPE_Q5_0:
-            get_rows_cuda<QK5_0, QR5_0, dequantize_q5_0>(src0, src1, dst, src0_d, src1_i32, dst_d, stream);
-            break;
-        case GGML_TYPE_Q5_1:
-            get_rows_cuda<QK5_1, QR5_1, dequantize_q5_1>(src0, src1, dst, src0_d, src1_i32, dst_d, stream);
-            break;
-        case GGML_TYPE_Q8_0:
-            get_rows_cuda<QK8_0, QR8_0, dequantize_q8_0>(src0, src1, dst, src0_d, src1_i32, dst_d, stream);
-            break;
-        default:
-            // TODO: k-quants
-            fprintf(stderr, "%s: unsupported type: %s\n", __func__, ggml_type_name(src0->type));
-            GGML_ASSERT(false);
-            break;
-    }
-}
-
-template<class op>
-static void ggml_cuda_op_bin_bcast(
-    const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst,
-    const float * src0_dd, const float * src1_dd, float * dst_dd, cudaStream_t main_stream) {
-
-    GGML_ASSERT(src1->type == GGML_TYPE_F32);
-
-    if (src0->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32) {
-        op()(src0, src1, dst, src0_dd, src1_dd, dst_dd, main_stream);
-    } else if (src0->type == GGML_TYPE_F16 && dst->type == GGML_TYPE_F16) {
-        op()(src0, src1, dst, (const half *) src0_dd, src1_dd, (half *) dst_dd, main_stream);
-    } else if (src0->type == GGML_TYPE_F16 && dst->type == GGML_TYPE_F32) {
-        op()(src0, src1, dst, (const half *) src0_dd, src1_dd, dst_dd, main_stream);
-    } else {
-        fprintf(stderr, "%s: unsupported types: dst: %s, src0: %s, src1: %s\n", __func__,
-            ggml_type_name(dst->type), ggml_type_name(src0->type), ggml_type_name(src1->type));
-        GGML_ASSERT(false);
-    }
-}
-
-static void ggml_cuda_op_repeat(
-    const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst,
-    const float * src0_d, const float * src1_d, float * dst_d, cudaStream_t main_stream) {
-
-    ggml_cuda_op_bin_bcast<bin_bcast_cuda<op_repeat>>(dst, src0, dst, nullptr, src0_d, dst_d, main_stream);
-
-    (void) src1;
-    (void) src1_d;
-}
-
-static void ggml_cuda_op_add(
-    const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst,
-    const float * src0_dd, const float * src1_dd, float * dst_dd, cudaStream_t main_stream) {
-
-    ggml_cuda_op_bin_bcast<bin_bcast_cuda<op_add>>(src0, src1, dst, src0_dd, src1_dd, dst_dd, main_stream);
-}
-
-static void ggml_cuda_op_acc(
-    const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst,
-    const float * src0_dd, const float * src1_dd, float * dst_dd, cudaStream_t main_stream) {
-
-    GGML_ASSERT(src0->type == GGML_TYPE_F32);
-    GGML_ASSERT(src1->type == GGML_TYPE_F32);
-    GGML_ASSERT( dst->type == GGML_TYPE_F32);
-    GGML_ASSERT(dst->ne[3] == 1); // just 3D tensors supported
-
-    int nb1 = dst->op_params[0] / 4; // 4 bytes of float32
-    int nb2 = dst->op_params[1] / 4; // 4 bytes of float32
-    // int nb3 = dst->op_params[2] / 4; // 4 bytes of float32 - unused
-    int offset = dst->op_params[3] / 4; // offset in bytes
-
-    acc_f32_cuda(src0_dd, src1_dd, dst_dd, ggml_nelements(dst), src1->ne[0], src1->ne[1], src1->ne[2], nb1, nb2, offset, main_stream);
-
-    (void) dst;
-}
-
-static void ggml_cuda_op_mul(
-    const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst,
-    const float * src0_dd, const float * src1_dd, float * dst_dd, cudaStream_t main_stream) {
-
-    ggml_cuda_op_bin_bcast<bin_bcast_cuda<op_mul>>(src0, src1, dst, src0_dd, src1_dd, dst_dd, main_stream);
-}
-
-static void ggml_cuda_op_div(
-    const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst,
-    const float * src0_dd, const float * src1_dd, float * dst_dd, cudaStream_t main_stream) {
-
-    ggml_cuda_op_bin_bcast<bin_bcast_cuda<op_div>>(src0, src1, dst, src0_dd, src1_dd, dst_dd, main_stream);
-}
-
-static void ggml_cuda_op_gelu(
-    const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst,
-    const float * src0_dd, const float * src1_dd, float * dst_dd, cudaStream_t main_stream) {
-
-    GGML_ASSERT(src0->type == GGML_TYPE_F32);
-    GGML_ASSERT( dst->type == GGML_TYPE_F32);
-
-    gelu_f32_cuda(src0_dd, dst_dd, ggml_nelements(src0), main_stream);
-
-    (void) src1;
-    (void) dst;
-    (void) src1_dd;
-}
-
-static void ggml_cuda_op_silu(
-    const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst,
-    const float * src0_dd, const float * src1_dd, float * dst_dd, cudaStream_t main_stream) {
-
-    GGML_ASSERT(src0->type == GGML_TYPE_F32);
-    GGML_ASSERT( dst->type == GGML_TYPE_F32);
-
-    silu_f32_cuda(src0_dd, dst_dd, ggml_nelements(src0), main_stream);
-
-    (void) src1;
-    (void) dst;
-    (void) src1_dd;
-}
-
-static void ggml_cuda_op_gelu_quick(
-    const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst,
-    const float * src0_dd, const float * src1_dd, float * dst_dd, cudaStream_t main_stream) {
-
-    GGML_ASSERT(src0->type == GGML_TYPE_F32);
-    GGML_ASSERT( dst->type == GGML_TYPE_F32);
-
-    gelu_quick_f32_cuda(src0_dd, dst_dd, ggml_nelements(src0), main_stream);
-
-    (void) src1;
-    (void) dst;
-    (void) src1_dd;
-}
-
-static void ggml_cuda_op_tanh(
-    const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst,
-    const float * src0_dd, const float * src1_dd, float * dst_dd, cudaStream_t main_stream) {
-
-    GGML_ASSERT(src0->type == GGML_TYPE_F32);
-    GGML_ASSERT( dst->type == GGML_TYPE_F32);
-
-    tanh_f32_cuda(src0_dd, dst_dd, ggml_nelements(src0), main_stream);
-
-    (void) src1;
-    (void) dst;
-    (void) src1_dd;
-}
-
-static void ggml_cuda_op_relu(
-    const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst,
-    const float * src0_dd, const float * src1_dd, float * dst_dd, cudaStream_t main_stream) {
-
-    GGML_ASSERT(src0->type == GGML_TYPE_F32);
-    GGML_ASSERT( dst->type == GGML_TYPE_F32);
-
-    relu_f32_cuda(src0_dd, dst_dd, ggml_nelements(src0), main_stream);
-
-    (void) src1;
-    (void) dst;
-    (void) src1_dd;
-}
-
-static void ggml_cuda_op_leaky_relu(
-    const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst,
-    const float * src0_dd, const float * src1_dd, float * dst_dd, cudaStream_t main_stream) {
-
-    GGML_ASSERT(src0->type == GGML_TYPE_F32);
-    GGML_ASSERT( dst->type == GGML_TYPE_F32);
-
-    float negative_slope;
-    memcpy(&negative_slope, dst->op_params, sizeof(float));
-
-    leaky_relu_f32_cuda(src0_dd, dst_dd, ggml_nelements(src0), negative_slope, main_stream);
-
-    (void) src1;
-    (void) dst;
-    (void) src1_dd;
-}
-
-static void ggml_cuda_op_sqr(
-    const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst,
-    const float * src0_dd, const float * src1_dd, float * dst_dd, cudaStream_t main_stream) {
-
-    GGML_ASSERT(src0->type == GGML_TYPE_F32);
-    GGML_ASSERT( dst->type == GGML_TYPE_F32);
-
-    sqr_f32_cuda(src0_dd, dst_dd, ggml_nelements(src0), main_stream);
-
-    (void) src1;
-    (void) dst;
-    (void) src1_dd;
-}
-
-static void ggml_cuda_op_norm(
-    const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst,
-    const float * src0_dd, const float * src1_dd, float * dst_dd, cudaStream_t main_stream) {
-
-    GGML_ASSERT(src0->type == GGML_TYPE_F32);
-    GGML_ASSERT( dst->type == GGML_TYPE_F32);
-
-    const int64_t ne00 = src0->ne[0];
-    const int64_t nrows = ggml_nrows(src0);
-
-    float eps;
-    memcpy(&eps, dst->op_params, sizeof(float));
-
-    norm_f32_cuda(src0_dd, dst_dd, ne00, nrows, eps, main_stream);
-
-    (void) src1;
-    (void) dst;
-    (void) src1_dd;
-}
-
-static void ggml_cuda_op_group_norm(
-    const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst,
-    const float * src0_dd, const float * src1_dd, float * dst_dd, cudaStream_t main_stream) {
-
-    GGML_ASSERT(src0->type == GGML_TYPE_F32);
-    GGML_ASSERT( dst->type == GGML_TYPE_F32);
-
-    int num_groups = dst->op_params[0];
-    int group_size = src0->ne[0] * src0->ne[1] * ((src0->ne[2] + num_groups - 1) / num_groups);
-    group_norm_f32_cuda(src0_dd, dst_dd, num_groups, group_size, src0->ne[0] * src0->ne[1] * src0->ne[2], main_stream);
-
-    (void) src1;
-    (void) dst;
-    (void) src1_dd;
-}
-
-static void ggml_cuda_op_concat(
-    const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst,
-    const float * src0_dd, const float * src1_dd, float * dst_dd, cudaStream_t main_stream) {
-
-    GGML_ASSERT(src0->type == GGML_TYPE_F32);
-    GGML_ASSERT(src1->type == GGML_TYPE_F32);
-    GGML_ASSERT(dst->type == GGML_TYPE_F32);
-
-    for (int i3 = 0; i3 < dst->ne[3]; i3++) {
-        concat_f32_cuda(src0_dd + i3 * (src0->nb[3] / 4), src1_dd + i3 * (src1->nb[3] / 4), dst_dd + i3 * (dst->nb[3] / 4), dst->ne[0], dst->ne[1], dst->ne[2], src0->ne[2], main_stream);
-    }
-
-    (void) src1;
-    (void) dst;
-}
-
-static void ggml_cuda_op_upscale(
-    const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst,
-    const float * src0_dd, const float * src1_dd, float * dst_dd, cudaStream_t main_stream) {
-
-    GGML_ASSERT(src0->type == GGML_TYPE_F32);
-    GGML_ASSERT(dst->type == GGML_TYPE_F32);
-    GGML_ASSERT(src0->ne[3] == 1 && dst->ne[3] == 1); // just 3D tensors
-
-    const int scale_factor = dst->op_params[0];
-
-    upscale_f32_cuda(src0_dd, dst_dd, src0->ne[0], src0->ne[1], src0->ne[2], scale_factor, main_stream);
-
-    (void) src1;
-    (void) dst;
-    (void) src1_dd;
-}
-
-static void ggml_cuda_op_pad(
-    const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst,
-    const float * src0_dd, const float * src1_dd, float * dst_dd, cudaStream_t main_stream) {
-
-    GGML_ASSERT(src0->type == GGML_TYPE_F32);
-    GGML_ASSERT(dst->type == GGML_TYPE_F32);
-    GGML_ASSERT(src0->ne[3] == 1 && dst->ne[3] == 1); // just 3D tensors
-
-    pad_f32_cuda(src0_dd, dst_dd,
-        src0->ne[0], src0->ne[1], src0->ne[2],
-        dst->ne[0], dst->ne[1], dst->ne[2], main_stream);
-
-    (void) src1;
-    (void) dst;
-    (void) src1_dd;
-}
-
-static void ggml_cuda_op_rms_norm(
-    const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst,
-    const float * src0_dd, const float * src1_dd, float * dst_dd, cudaStream_t main_stream) {
-
-    GGML_ASSERT(src0->type == GGML_TYPE_F32);
-    GGML_ASSERT( dst->type == GGML_TYPE_F32);
-
-    const int64_t ne00 = src0->ne[0];
-    const int64_t nrows = ggml_nrows(src0);
-
-    float eps;
-    memcpy(&eps, dst->op_params, sizeof(float));
-
-    rms_norm_f32_cuda(src0_dd, dst_dd, ne00, nrows, eps, main_stream);
-
-    (void) src1;
-    (void) dst;
-    (void) src1_dd;
-}
-
-static void ggml_cuda_op_mul_mat_q(
-    const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, const char * src0_dd_i, const float * src1_ddf_i,
-    const char * src1_ddq_i, float * dst_dd_i, const int64_t row_low, const int64_t row_high, const int64_t src1_ncols,
-    const int64_t src1_padded_row_size, cudaStream_t stream) {
-
-    const int64_t ne00 = src0->ne[0];
-
-    const int64_t ne10 = src1->ne[0];
-    GGML_ASSERT(ne10 % QK8_1 == 0);
-
-    const int64_t ne0 = dst->ne[0];
-
-    const int64_t row_diff = row_high - row_low;
-
-    int id;
-    CUDA_CHECK(cudaGetDevice(&id));
-
-    // the main device has a larger memory buffer to hold the results from all GPUs
-    // nrows_dst == nrows of the matrix that the dequantize_mul_mat kernel writes into
-    const int64_t nrows_dst = dst->backend == GGML_BACKEND_GPU && id == g_main_device ? ne0 : row_diff;
-
-    switch (src0->type) {
-        case GGML_TYPE_Q4_0:
-            ggml_mul_mat_q4_0_q8_1_cuda(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, src1_ncols, src1_padded_row_size, nrows_dst, stream);
-            break;
-        case GGML_TYPE_Q4_1:
-            ggml_mul_mat_q4_1_q8_1_cuda(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, src1_ncols, src1_padded_row_size, nrows_dst, stream);
-            break;
-        case GGML_TYPE_Q5_0:
-            ggml_mul_mat_q5_0_q8_1_cuda(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, src1_ncols, src1_padded_row_size, nrows_dst, stream);
-            break;
-        case GGML_TYPE_Q5_1:
-            ggml_mul_mat_q5_1_q8_1_cuda(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, src1_ncols, src1_padded_row_size, nrows_dst, stream);
-            break;
-        case GGML_TYPE_Q8_0:
-            ggml_mul_mat_q8_0_q8_1_cuda(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, src1_ncols, src1_padded_row_size, nrows_dst, stream);
-            break;
-        case GGML_TYPE_Q2_K:
-            ggml_mul_mat_q2_K_q8_1_cuda(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, src1_ncols, src1_padded_row_size, nrows_dst, stream);
-            break;
-        case GGML_TYPE_Q3_K:
-            ggml_mul_mat_q3_K_q8_1_cuda(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, src1_ncols, src1_padded_row_size, nrows_dst, stream);
-            break;
-        case GGML_TYPE_Q4_K:
-            ggml_mul_mat_q4_K_q8_1_cuda(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, src1_ncols, src1_padded_row_size, nrows_dst, stream);
-            break;
-        case GGML_TYPE_Q5_K:
-            ggml_mul_mat_q5_K_q8_1_cuda(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, src1_ncols, src1_padded_row_size, nrows_dst, stream);
-            break;
-        case GGML_TYPE_Q6_K:
-            ggml_mul_mat_q6_K_q8_1_cuda(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, src1_ncols, src1_padded_row_size, nrows_dst, stream);
-            break;
-        default:
-            GGML_ASSERT(false);
-            break;
-    }
-
-    (void) src1;
-    (void) dst;
-    (void) src1_ddf_i;
-}
-
-static int64_t get_row_rounding(ggml_type type) {
-    int64_t min_compute_capability = INT_MAX;
-    int64_t max_compute_capability = INT_MIN;
-    for (int id = 0; id < g_device_count; ++id) {
-        if (g_tensor_split[id] < (id + 1 < g_device_count ? g_tensor_split[id + 1] : 1.0f)) {
-            if (min_compute_capability > g_device_caps[id].cc) {
-                min_compute_capability = g_device_caps[id].cc;
-            }
-            if (max_compute_capability < g_device_caps[id].cc) {
-                max_compute_capability = g_device_caps[id].cc;
-            }
-        }
-    }
-
-#if defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)
-    switch(type) {
-        case GGML_TYPE_Q4_0:
-        case GGML_TYPE_Q4_1:
-        case GGML_TYPE_Q5_0:
-        case GGML_TYPE_Q5_1:
-        case GGML_TYPE_Q8_0:
-            return max_compute_capability >= CC_RDNA2 ? 128 : 64;
-        case GGML_TYPE_F16:
-        case GGML_TYPE_F32:
-            return 1;
-        case GGML_TYPE_Q2_K:
-            return max_compute_capability >= CC_RDNA2 ? 128 : 32;
-        case GGML_TYPE_Q3_K:
-            return min_compute_capability < CC_RDNA2 ? 128 : 64;
-        case GGML_TYPE_Q4_K:
-        case GGML_TYPE_Q5_K:
-        case GGML_TYPE_Q6_K:
-            return max_compute_capability >= CC_RDNA2 ? 128 : 64;
-        default:
-            GGML_ASSERT(false);
-    }
-#else
-    switch(type) {
-        case GGML_TYPE_Q4_0:
-        case GGML_TYPE_Q4_1:
-            return max_compute_capability >= CC_VOLTA ? 128 : 64;
-        case GGML_TYPE_Q5_0:
-        case GGML_TYPE_Q5_1:
-        case GGML_TYPE_Q8_0:
-            return 64;
-        case GGML_TYPE_F16:
-        case GGML_TYPE_F32:
-            return 1;
-        case GGML_TYPE_Q2_K:
-        case GGML_TYPE_Q3_K:
-        case GGML_TYPE_Q4_K:
-        case GGML_TYPE_Q5_K:
-            return max_compute_capability >= CC_VOLTA ? 128 : 64;
-        case GGML_TYPE_Q6_K:
-            return 64;
-        default:
-            GGML_ASSERT(false);
-    }
-#endif // defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)
-}
-
-static void ggml_cuda_op_mul_mat_vec_q(
-    const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, const char * src0_dd_i, const float * src1_ddf_i,
-    const char * src1_ddq_i, float * dst_dd_i, const int64_t row_low, const int64_t row_high, const int64_t src1_ncols,
-    const int64_t src1_padded_row_size, cudaStream_t stream) {
-
-    GGML_ASSERT(ggml_nrows(src1) == 1);
-
-    const int64_t ne00 = src0->ne[0];
-    const int64_t row_diff = row_high - row_low;
-
-    switch (src0->type) {
-        case GGML_TYPE_Q4_0:
-            mul_mat_vec_q4_0_q8_1_cuda(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, stream);
-            break;
-        case GGML_TYPE_Q4_1:
-            mul_mat_vec_q4_1_q8_1_cuda(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, stream);
-            break;
-        case GGML_TYPE_Q5_0:
-            mul_mat_vec_q5_0_q8_1_cuda(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, stream);
-            break;
-        case GGML_TYPE_Q5_1:
-            mul_mat_vec_q5_1_q8_1_cuda(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, stream);
-            break;
-        case GGML_TYPE_Q8_0:
-            mul_mat_vec_q8_0_q8_1_cuda(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, stream);
-            break;
-        case GGML_TYPE_Q2_K:
-            mul_mat_vec_q2_K_q8_1_cuda(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, stream);
-            break;
-        case GGML_TYPE_Q3_K:
-            mul_mat_vec_q3_K_q8_1_cuda(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, stream);
-            break;
-        case GGML_TYPE_Q4_K:
-            mul_mat_vec_q4_K_q8_1_cuda(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, stream);
-            break;
-        case GGML_TYPE_Q5_K:
-            mul_mat_vec_q5_K_q8_1_cuda(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, stream);
-            break;
-        case GGML_TYPE_Q6_K:
-            mul_mat_vec_q6_K_q8_1_cuda(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, stream);
-            break;
-        default:
-            GGML_ASSERT(false);
-            break;
-    }
-
-    (void) src1;
-    (void) dst;
-    (void) src1_ddf_i;
-    (void) src1_ncols;
-    (void) src1_padded_row_size;
-}
-
-static void ggml_cuda_op_dequantize_mul_mat_vec(
-    const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, const char * src0_dd_i, const float * src1_ddf_i,
-    const char * src1_ddq_i, float * dst_dd_i, const int64_t row_low, const int64_t row_high, const int64_t src1_ncols,
-    const int64_t src1_padded_row_size, cudaStream_t stream) {
-
-    const int64_t ne00 = src0->ne[0];
-    const int64_t row_diff = row_high - row_low;
-
-    GGML_ASSERT(src1->type == GGML_TYPE_F32);
-
-    // on some GPUs it is faster to convert src1 to half and to use half precision intrinsics
-#ifdef GGML_CUDA_F16
-    cuda_pool_alloc<half> src1_dfloat_a;
-    half * src1_dfloat = nullptr; // dfloat == half
-
-    bool src1_convert_f16 =
-        src0->type == GGML_TYPE_Q4_0 || src0->type == GGML_TYPE_Q4_1 ||
-        src0->type == GGML_TYPE_Q5_0 || src0->type == GGML_TYPE_Q5_1 ||
-        src0->type == GGML_TYPE_Q8_0 || src0->type == GGML_TYPE_F16;
-
-    if (src1_convert_f16) {
-        src1_dfloat = src1_dfloat_a.alloc(ne00);
-        ggml_cpy_f32_f16_cuda((const char *) src1_ddf_i, (char *) src1_dfloat, ne00,
-                                ne00, 1, sizeof(float), 0, 0,
-                                ne00, 1, sizeof(half),  0, 0, stream);
-    }
-#else
-    const dfloat * src1_dfloat = (const dfloat *) src1_ddf_i; // dfloat == float, no conversion
-#endif // GGML_CUDA_F16
-
-    switch (src0->type) {
-        case GGML_TYPE_Q4_0:
-            dequantize_mul_mat_vec_q4_0_cuda(src0_dd_i, src1_dfloat, dst_dd_i, ne00, row_diff, stream);
-            break;
-        case GGML_TYPE_Q4_1:
-            dequantize_mul_mat_vec_q4_1_cuda(src0_dd_i, src1_dfloat, dst_dd_i, ne00, row_diff, stream);
-            break;
-        case GGML_TYPE_Q5_0:
-            dequantize_mul_mat_vec_q5_0_cuda(src0_dd_i, src1_dfloat, dst_dd_i, ne00, row_diff, stream);
-            break;
-        case GGML_TYPE_Q5_1:
-            dequantize_mul_mat_vec_q5_1_cuda(src0_dd_i, src1_dfloat, dst_dd_i, ne00, row_diff, stream);
-            break;
-        case GGML_TYPE_Q8_0:
-            dequantize_mul_mat_vec_q8_0_cuda(src0_dd_i, src1_dfloat, dst_dd_i, ne00, row_diff, stream);
-            break;
-        case GGML_TYPE_Q2_K:
-            dequantize_mul_mat_vec_q2_K_cuda(src0_dd_i, src1_ddf_i, dst_dd_i, ne00, row_diff, stream);
-            break;
-        case GGML_TYPE_Q3_K:
-            dequantize_mul_mat_vec_q3_K_cuda(src0_dd_i, src1_ddf_i, dst_dd_i, ne00, row_diff, stream);
-            break;
-        case GGML_TYPE_Q4_K:
-            dequantize_mul_mat_vec_q4_K_cuda(src0_dd_i, src1_ddf_i, dst_dd_i, ne00, row_diff, stream);
-            break;
-        case GGML_TYPE_Q5_K:
-            dequantize_mul_mat_vec_q5_K_cuda(src0_dd_i, src1_ddf_i, dst_dd_i, ne00, row_diff, stream);
-            break;
-        case GGML_TYPE_Q6_K:
-            dequantize_mul_mat_vec_q6_K_cuda(src0_dd_i, src1_ddf_i, dst_dd_i, ne00, row_diff, stream);
-            break;
-        case GGML_TYPE_F16:
-            convert_mul_mat_vec_f16_cuda(src0_dd_i, src1_dfloat, dst_dd_i, ne00, row_diff, stream);
-            break;
-        default:
-            GGML_ASSERT(false);
-            break;
-    }
-
-    (void) src1;
-    (void) dst;
-    (void) src1_ddq_i;
-    (void) src1_ncols;
-    (void) src1_padded_row_size;
-}
-
-static void ggml_cuda_op_mul_mat_cublas(
-    const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, const char * src0_dd_i, const float * src1_ddf_i,
-    const char * src1_ddq_i, float * dst_dd_i, const int64_t row_low, const int64_t row_high, const int64_t src1_ncols,
-    const int64_t src1_padded_row_size, cudaStream_t stream) {
-
-    GGML_ASSERT(src0_dd_i  != nullptr);
-    GGML_ASSERT(src1_ddf_i != nullptr);
-    GGML_ASSERT(dst_dd_i   != nullptr);
-
-    const int64_t ne00 = src0->ne[0];
-    const int64_t ne10 = src1->ne[0];
-
-    const int64_t ne0 = dst->ne[0];
-
-    const int64_t row_diff = row_high - row_low;
-
-    int id;
-    CUDA_CHECK(cudaGetDevice(&id));
-
-    // the main device has a larger memory buffer to hold the results from all GPUs
-    // ldc == nrows of the matrix that cuBLAS writes into
-    int ldc = dst->backend == GGML_BACKEND_GPU && id == g_main_device ? ne0 : row_diff;
-
-    const int compute_capability = g_device_caps[id].cc;
-
-    if (compute_capability >= CC_VOLTA && (src0->type == GGML_TYPE_F16 || ggml_is_quantized(src0->type)) && ggml_is_contiguous(src0) && row_diff == src0->ne[1] && dst->op_params[0] == GGML_PREC_DEFAULT) {
-        //printf("this branch\n");
-        // convert src0 and src1 to fp16, multiply as fp16, convert dst to fp32
-        cuda_pool_alloc<half> src0_as_f16;
-        if (src0->type != GGML_TYPE_F16) {
-            const to_fp16_cuda_t to_fp16_cuda = ggml_get_to_fp16_cuda(src0->type);
-            GGML_ASSERT(to_fp16_cuda != nullptr);
-            size_t ne = row_diff*ne00;
-            src0_as_f16.alloc(ne);
-            to_fp16_cuda(src0_dd_i, src0_as_f16.get(), ne, stream);
-        }
-        const half * src0_ptr = src0->type == GGML_TYPE_F16 ? (const half *) src0_dd_i : src0_as_f16.get();
-
-        cuda_pool_alloc<half> src1_as_f16;
-        if (src1->type != GGML_TYPE_F16) {
-            const to_fp16_cuda_t to_fp16_cuda = ggml_get_to_fp16_cuda(src1->type);
-            GGML_ASSERT(to_fp16_cuda != nullptr);
-            size_t ne = src1_ncols*ne10;
-            src1_as_f16.alloc(ne);
-            to_fp16_cuda(src1_ddf_i, src1_as_f16.get(), ne, stream);
-        }
-        const half * src1_ptr = src1->type == GGML_TYPE_F16 ? (const half *) src1_ddf_i : src1_as_f16.get();
-        cuda_pool_alloc<half> dst_f16(row_diff*src1_ncols);
-
-        const half alpha_f16 = 1.0f;
-        const half beta_f16 = 0.0f;
-
-        CUBLAS_CHECK(cublasSetStream(g_cublas_handles[id], stream));
-        CUBLAS_CHECK(
-            cublasGemmEx(g_cublas_handles[id], CUBLAS_OP_T, CUBLAS_OP_N,
-                    row_diff, src1_ncols, ne10,
-                    &alpha_f16, src0_ptr,       CUDA_R_16F, ne00,
-                                src1_ptr,       CUDA_R_16F, ne10,
-                    &beta_f16,   dst_f16.get(), CUDA_R_16F, ldc,
-                    CUBLAS_COMPUTE_16F,
-                    CUBLAS_GEMM_DEFAULT_TENSOR_OP));
-
-        const to_fp32_cuda_t to_fp32_cuda = ggml_get_to_fp32_cuda(GGML_TYPE_F16);
-        to_fp32_cuda(dst_f16.get(), dst_dd_i, row_diff*src1_ncols, stream);
-    } else {
-        cuda_pool_alloc<float> src0_ddq_as_f32;
-        cuda_pool_alloc<float> src1_ddq_as_f32;
-
-        if (src0->type != GGML_TYPE_F32) {
-            const to_fp32_cuda_t to_fp32_cuda = ggml_get_to_fp32_cuda(src0->type);
-            GGML_ASSERT(to_fp32_cuda != nullptr);
-            src0_ddq_as_f32.alloc(row_diff*ne00);
-            to_fp32_cuda(src0_dd_i, src0_ddq_as_f32.get(), row_diff*ne00, stream);
-        }
-        if (src1->type != GGML_TYPE_F32) {
-            const to_fp32_cuda_t to_fp32_cuda = ggml_get_to_fp32_cuda(src1->type);
-            GGML_ASSERT(to_fp32_cuda != nullptr);
-            src1_ddq_as_f32.alloc(src1_ncols*ne10);
-            to_fp32_cuda(src1_ddf_i, src1_ddq_as_f32.get(), src1_ncols*ne10, stream);
-        }
-
-        const float * src0_ddf_i = src0->type == GGML_TYPE_F32 ? (const float *) src0_dd_i : src0_ddq_as_f32.get();
-        const float * src1_ddf1_i = src1->type == GGML_TYPE_F32 ? (const float *) src1_ddf_i : src1_ddq_as_f32.get();
-
-        const float alpha = 1.0f;
-        const float beta = 0.0f;
-
-        CUBLAS_CHECK(cublasSetStream(g_cublas_handles[id], stream));
-        CUBLAS_CHECK(
-            cublasSgemm(g_cublas_handles[id], CUBLAS_OP_T, CUBLAS_OP_N,
-                    row_diff, src1_ncols, ne10,
-                    &alpha, src0_ddf_i,  ne00,
-                            src1_ddf1_i, ne10,
-                    &beta,  dst_dd_i,    ldc));
-    }
-
-    (void) dst;
-    (void) src1_ddq_i;
-    (void) src1_padded_row_size;
-}
-
-static void ggml_cuda_op_rope(
-    const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst,
-    const float * src0_dd, const float * src1_dd, float * dst_dd, cudaStream_t main_stream) {
-
-    GGML_ASSERT(src0->type == GGML_TYPE_F32 || src0->type == GGML_TYPE_F16);
-    GGML_ASSERT( dst->type == GGML_TYPE_F32 ||  dst->type == GGML_TYPE_F16);
-    GGML_ASSERT(src0->type == dst->type);
-
-    const int64_t ne00 = src0->ne[0];
-    const int64_t ne01 = src0->ne[1];
-    const int64_t ne2 = dst->ne[2];
-    const int64_t nrows = ggml_nrows(src0);
-
-    //const int n_past      = ((int32_t *) dst->op_params)[0];
-    const int n_dims      = ((int32_t *) dst->op_params)[1];
-    const int mode        = ((int32_t *) dst->op_params)[2];
-    const int n_ctx       = ((int32_t *) dst->op_params)[3];
-    const int n_orig_ctx  = ((int32_t *) dst->op_params)[4];
-
-    // RoPE alteration for extended context
-    float freq_base, freq_scale, ext_factor, attn_factor, beta_fast, beta_slow;
-    memcpy(&freq_base,   (int32_t *) dst->op_params +  5, sizeof(float));
-    memcpy(&freq_scale,  (int32_t *) dst->op_params +  6, sizeof(float));
-    memcpy(&ext_factor,  (int32_t *) dst->op_params +  7, sizeof(float));
-    memcpy(&attn_factor, (int32_t *) dst->op_params +  8, sizeof(float));
-    memcpy(&beta_fast,   (int32_t *) dst->op_params +  9, sizeof(float));
-    memcpy(&beta_slow,   (int32_t *) dst->op_params + 10, sizeof(float));
-
-    const int32_t * pos = nullptr;
-    if ((mode & 1) == 0) {
-        GGML_ASSERT(src1->type == GGML_TYPE_I32);
-        GGML_ASSERT(src1->ne[0] == ne2);
-        pos = (const int32_t *) src1_dd;
-    }
-
-    const bool is_neox = mode & 2;
-    const bool is_glm  = mode & 4;
-
-    rope_corr_dims corr_dims;
-    ggml_rope_yarn_corr_dims(n_dims, n_orig_ctx, freq_base, beta_fast, beta_slow, corr_dims.v);
-
-    // compute
-    if (is_glm) {
-        GGML_ASSERT(false);
-        rope_glm_f32_cuda(src0_dd, dst_dd, ne00, nrows, pos, freq_scale, ne01, freq_base, n_ctx, main_stream);
-    } else if (is_neox) {
-        if (src0->type == GGML_TYPE_F32) {
-            rope_neox_cuda(
-                (const float *)src0_dd, (float *)dst_dd, ne00, n_dims, nrows, pos, freq_scale, ne01, freq_base, ext_factor,
-                attn_factor, corr_dims, main_stream
-            );
-        } else if (src0->type == GGML_TYPE_F16) {
-            rope_neox_cuda(
-                (const half *)src0_dd, (half *)dst_dd, ne00, n_dims, nrows, pos, freq_scale, ne01, freq_base, ext_factor,
-                attn_factor, corr_dims, main_stream
-            );
-        } else {
-            GGML_ASSERT(false);
-        }
-    } else {
-        if (src0->type == GGML_TYPE_F32) {
-            rope_cuda(
-                (const float *)src0_dd, (float *)dst_dd, ne00, nrows, pos, freq_scale, ne01, freq_base, ext_factor,
-                attn_factor, corr_dims, main_stream
-            );
-        } else if (src0->type == GGML_TYPE_F16) {
-            rope_cuda(
-                (const half *)src0_dd, (half *)dst_dd, ne00, nrows, pos, freq_scale, ne01, freq_base, ext_factor,
-                attn_factor, corr_dims, main_stream
-            );
-        } else {
-            GGML_ASSERT(false);
-        }
-    }
-
-    (void) src1;
-    (void) dst;
-    (void) src1_dd;
-}
-
-static void ggml_cuda_op_alibi(
-    const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst,
-    const float * src0_dd, const float * src1_dd, float * dst_dd, cudaStream_t main_stream) {
-
-    GGML_ASSERT(src0->type == GGML_TYPE_F32);
-    GGML_ASSERT( dst->type == GGML_TYPE_F32);
-
-    const int64_t ne00 = src0->ne[0];
-    const int64_t ne01 = src0->ne[1];
-    const int64_t ne02 = src0->ne[2];
-    const int64_t nrows = ggml_nrows(src0);
-
-    //const int n_past = ((int32_t *) dst->op_params)[0];
-    const int n_head = ((int32_t *) dst->op_params)[1];
-    float max_bias;
-    memcpy(&max_bias, (int32_t *) dst->op_params + 2, sizeof(float));
-
-    //GGML_ASSERT(ne01 + n_past == ne00);
-    GGML_ASSERT(n_head == ne02);
-
-    const int n_heads_log2_floor = 1 << (int) floor(log2(n_head));
-
-    const float m0 = powf(2.0f, -(max_bias) / n_heads_log2_floor);
-    const float m1 = powf(2.0f, -(max_bias / 2.0f) / n_heads_log2_floor);
-
-    alibi_f32_cuda(src0_dd, dst_dd, ne00, nrows, ne01, n_heads_log2_floor, m0, m1, main_stream);
-
-    (void) src1;
-    (void) src1_dd;
-}
-
-static void ggml_cuda_op_im2col(
-    const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst,
-    const float * src0_dd, const float * src1_dd, float * dst_dd, cudaStream_t main_stream) {
-
-    GGML_ASSERT(src0->type == GGML_TYPE_F16);
-    GGML_ASSERT(src1->type == GGML_TYPE_F32);
-    GGML_ASSERT( dst->type == GGML_TYPE_F16);
-
-    const int32_t s0 = ((const int32_t*)(dst->op_params))[0];
-    const int32_t s1 = ((const int32_t*)(dst->op_params))[1];
-    const int32_t p0 = ((const int32_t*)(dst->op_params))[2];
-    const int32_t p1 = ((const int32_t*)(dst->op_params))[3];
-    const int32_t d0 = ((const int32_t*)(dst->op_params))[4];
-    const int32_t d1 = ((const int32_t*)(dst->op_params))[5];
-
-    const bool is_2D = ((const int32_t*)(dst->op_params))[6] == 1;
-
-    const int64_t IC = src1->ne[is_2D ? 2 : 1];
-    const int64_t IH = is_2D ? src1->ne[1] : 1;
-    const int64_t IW =         src1->ne[0];
-
-    const int64_t KH = is_2D ? src0->ne[1] : 1;
-    const int64_t KW =         src0->ne[0];
-
-    const int64_t OH = is_2D ? dst->ne[2] : 1;
-    const int64_t OW =         dst->ne[1];
-
-    const size_t delta_offset = src1->nb[is_2D ? 2 : 1] / 4; // nb is byte offset, src is type float32
-
-    im2col_f32_f16_cuda(src1_dd, (half*) dst_dd, IW, IH, OW, OH, KW, KH, IC, delta_offset, s0, s1, p0, p1, d0, d1, main_stream);
-
-    (void) src0;
-    (void) src0_dd;
-}
-
-static void ggml_cuda_op_sum_rows(
-    const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst,
-    const float * src0_dd, const float * src1_dd, float * dst_dd, cudaStream_t main_stream) {
-
-    GGML_ASSERT(src0->type == GGML_TYPE_F32);
-    GGML_ASSERT( dst->type == GGML_TYPE_F32);
-
-    const int64_t ncols = src0->ne[0];
-    const int64_t nrows = ggml_nrows(src0);
-
-    sum_rows_f32_cuda(src0_dd, dst_dd, ncols, nrows, main_stream);
-
-    (void) src1;
-    (void) dst;
-    (void) src1_dd;
-}
-
-static void ggml_cuda_op_argsort(
-    const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst,
-    const float * src0_dd, const float * src1_dd, float * dst_dd, cudaStream_t main_stream) {
-
-    GGML_ASSERT(src0->type == GGML_TYPE_F32);
-    GGML_ASSERT( dst->type == GGML_TYPE_I32);
-
-    const int64_t ncols = src0->ne[0];
-    const int64_t nrows = ggml_nrows(src0);
-
-    enum ggml_sort_order order = (enum ggml_sort_order) dst->op_params[0];
-
-    argsort_f32_i32_cuda(src0_dd, (int *)dst_dd, ncols, nrows, order, main_stream);
-
-    (void) src1;
-    (void) dst;
-    (void) src1_dd;
-}
-
-static void ggml_cuda_op_diag_mask_inf(
-    const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst,
-    const float * src0_dd, const float * src1_dd, float * dst_dd, cudaStream_t main_stream) {
-
-    GGML_ASSERT(src0->type == GGML_TYPE_F32);
-    GGML_ASSERT( dst->type == GGML_TYPE_F32);
-
-    const int64_t ne00 = src0->ne[0];
-    const int64_t ne01 = src0->ne[1];
-    const int nrows0 = ggml_nrows(src0);
-
-    const int n_past = ((int32_t *) dst->op_params)[0];
-
-    diag_mask_inf_f32_cuda(src0_dd, dst_dd, ne00, nrows0, ne01, n_past, main_stream);
-
-    (void) src1;
-    (void) dst;
-    (void) src1_dd;
-}
-
-static void ggml_cuda_op_soft_max(
-    const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst,
-    const float * src0_dd, const float * src1_dd, float * dst_dd, cudaStream_t main_stream) {
-
-    GGML_ASSERT(src0->type == GGML_TYPE_F32);
-    GGML_ASSERT( dst->type == GGML_TYPE_F32);
-
-    GGML_ASSERT(!src1 || src1->type == GGML_TYPE_F32); // src1 contains mask and it is optional
-
-    const int64_t ne00 = src0->ne[0];
-    const int64_t nrows_x = ggml_nrows(src0);
-    const int64_t nrows_y = src1 ? ggml_nrows(src1) : 1;
-
-    float scale = 1.0f;
-    memcpy(&scale, dst->op_params, sizeof(float));
-
-    soft_max_f32_cuda(src0_dd, src1 ? src1_dd : nullptr, dst_dd, ne00, nrows_x, nrows_y, scale, main_stream);
-
-    (void) dst;
-}
-
-static void ggml_cuda_op_scale(
-    const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst,
-    const float * src0_dd, const float * src1_dd, float * dst_dd, cudaStream_t main_stream) {
-
-    GGML_ASSERT(src0->type == GGML_TYPE_F32);
-    GGML_ASSERT( dst->type == GGML_TYPE_F32);
-
-    float scale;
-    memcpy(&scale, dst->op_params, sizeof(float));
-
-    scale_f32_cuda(src0_dd, dst_dd, scale, ggml_nelements(src0), main_stream);
-    CUDA_CHECK(cudaGetLastError());
-
-    (void) src1;
-    (void) dst;
-    (void) src1_dd;
-}
-
-static void ggml_cuda_op_clamp(
-    const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst,
-    const float * src0_dd, const float * src1_dd, float * dst_dd, cudaStream_t main_stream) {
-
-    GGML_ASSERT(src0->type == GGML_TYPE_F32);
-    GGML_ASSERT( dst->type == GGML_TYPE_F32);
-
-    float min;
-    float max;
-    memcpy(&min, dst->op_params, sizeof(float));
-    memcpy(&max, (float *) dst->op_params + 1, sizeof(float));
-
-    clamp_f32_cuda(src0_dd, dst_dd, min, max, ggml_nelements(src0), main_stream);
-    CUDA_CHECK(cudaGetLastError());
-
-    (void) src1;
-    (void) dst;
-    (void) src1_dd;
-}
-
-static void ggml_cuda_op_flatten(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, const ggml_cuda_op_flatten_t op) {
-    const int64_t nrows0 = ggml_nrows(src0);
-
-    const bool use_src1 = src1 != nullptr;
-    const int64_t nrows1 = use_src1 ? ggml_nrows(src1) : 1;
-
-    GGML_ASSERT(!use_src1 || src1->backend != GGML_BACKEND_GPU_SPLIT);
-    GGML_ASSERT(              dst->backend != GGML_BACKEND_GPU_SPLIT);
-
-    ggml_tensor_extra_gpu * src0_extra =            (ggml_tensor_extra_gpu *) src0->extra;
-    ggml_tensor_extra_gpu * src1_extra = use_src1 ? (ggml_tensor_extra_gpu *) src1->extra : nullptr;
-    ggml_tensor_extra_gpu * dst_extra  =            (ggml_tensor_extra_gpu *)  dst->extra;
-
-    const bool src0_on_device =             src0->backend == GGML_BACKEND_GPU || src0->backend == GGML_BACKEND_GPU_SPLIT;
-    const bool src1_on_device = use_src1 && src1->backend == GGML_BACKEND_GPU;
-    const bool  dst_on_device =              dst->backend == GGML_BACKEND_GPU;
-
-    // dd = data device
-    float * src0_ddf = nullptr;
-    float * src1_ddf = nullptr;
-    float *  dst_ddf = nullptr;
-
-    cuda_pool_alloc<float> src0_f;
-    cuda_pool_alloc<float> src1_f;
-    cuda_pool_alloc<float>  dst_f;
-
-    ggml_cuda_set_device(g_main_device);
-    cudaStream_t main_stream = g_cudaStreams[g_main_device][0];
-
-    if (src0_on_device) {
-        src0_ddf = (float *) src0_extra->data_device[g_main_device];
-    } else {
-        src0_ddf = src0_f.alloc(ggml_nelements(src0));
-        CUDA_CHECK(ggml_cuda_cpy_tensor_2d(src0_ddf, src0, 0, 0, 0, nrows0, main_stream));
-    }
-
-    if (use_src1) {
-        if (src1_on_device) {
-            src1_ddf = (float *) src1_extra->data_device[g_main_device];
-        } else {
-            src1_ddf = src1_f.alloc(ggml_nelements(src1));
-            CUDA_CHECK(ggml_cuda_cpy_tensor_2d(src1_ddf, src1, 0, 0, 0, nrows1, main_stream));
-        }
-    }
-    if (dst_on_device) {
-        dst_ddf = (float *) dst_extra->data_device[g_main_device];
-    } else {
-        dst_ddf = dst_f.alloc(ggml_nelements(dst));
-    }
-
-    // do the computation
-    op(src0, src1, dst, src0_ddf, src1_ddf, dst_ddf, main_stream);
-    CUDA_CHECK(cudaGetLastError());
-
-    // copy dst to host if necessary
-    if (!dst_on_device) {
-        CUDA_CHECK(cudaMemcpyAsync(dst->data, dst_ddf, ggml_nbytes(dst), cudaMemcpyDeviceToHost, main_stream));
-    }
-
-    if (dst->backend == GGML_BACKEND_CPU) {
-        CUDA_CHECK(cudaDeviceSynchronize());
-    }
-}
-
-static void ggml_cuda_set_peer_access(const int n_tokens) {
-    static bool peer_access_enabled = false;
-
-    const bool enable_peer_access = n_tokens <= GGML_CUDA_PEER_MAX_BATCH_SIZE;
-
-    if (peer_access_enabled == enable_peer_access) {
-        return;
-    }
-
-#ifdef NDEBUG
-    for (int id = 0; id < g_device_count; ++id) {
-        ggml_cuda_set_device(id);
-        CUDA_CHECK(cudaDeviceSynchronize());
-    }
-
-    for (int id = 0; id < g_device_count; ++id) {
-        ggml_cuda_set_device(id);
-
-        for (int id_other = 0; id_other < g_device_count; ++id_other) {
-            if (id == id_other) {
-                continue;
-            }
-            if (id != g_main_device && id_other != g_main_device) {
-                continue;
-            }
-
-            int can_access_peer;
-            CUDA_CHECK(cudaDeviceCanAccessPeer(&can_access_peer, id, id_other));
-            if (can_access_peer) {
-                if (enable_peer_access) {
-                    CUDA_CHECK(cudaDeviceEnablePeerAccess(id_other, 0));
-                } else {
-                    CUDA_CHECK(cudaDeviceDisablePeerAccess(id_other));
-                }
-            }
-        }
-    }
-#endif // NDEBUG
-
-    peer_access_enabled = enable_peer_access;
-}
-
-static void ggml_cuda_op_mul_mat(
-    const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, ggml_cuda_op_mul_mat_t op,
-    const bool convert_src1_to_q8_1) {
-
-    const int64_t ne00 = src0->ne[0];
-    const int64_t ne01 = src0->ne[1];
-    const int64_t ne02 = src0->ne[2];
-    const int64_t ne03 = src0->ne[3];
-
-    const int64_t ne10 = src1->ne[0];
-    const int64_t ne11 = src1->ne[1];
-    const int64_t ne12 = src1->ne[2];
-    const int64_t ne13 = src1->ne[3];
-    const int64_t nrows1 = ggml_nrows(src1);
-
-    GGML_ASSERT(ne03 == ne13);
-
-    const int64_t ne0 = dst->ne[0];
-    const int64_t ne1 = dst->ne[1];
-
-    const int nb2 = dst->nb[2];
-    const int nb3 = dst->nb[3];
-
-    GGML_ASSERT(dst->backend != GGML_BACKEND_GPU_SPLIT);
-    GGML_ASSERT(src1->backend != GGML_BACKEND_GPU_SPLIT);
-    GGML_ASSERT(src1->type == GGML_TYPE_F32 || (src1->ne[2] == 1 && src1->ne[3] == 1));
-
-    GGML_ASSERT(ne12 >= ne02 && ne12 % ne02 == 0);
-
-    const int64_t i02_divisor = ne12 / ne02;
-
-    const size_t src0_ts = ggml_type_size(src0->type);
-    const size_t src0_bs = ggml_blck_size(src0->type);
-    const size_t q8_1_ts = sizeof(block_q8_1);
-    const size_t q8_1_bs = QK8_1;
-
-    ggml_tensor_extra_gpu * src0_extra = (ggml_tensor_extra_gpu *) src0->extra;
-    ggml_tensor_extra_gpu * src1_extra = (ggml_tensor_extra_gpu *) src1->extra;
-    ggml_tensor_extra_gpu *  dst_extra = (ggml_tensor_extra_gpu *)  dst->extra;
-
-    const bool src0_on_device = src0->backend == GGML_BACKEND_GPU || src0->backend == GGML_BACKEND_GPU_SPLIT;
-    const bool src0_is_contiguous = ggml_is_contiguous(src0);
-    const bool src1_is_contiguous = ggml_is_contiguous(src1);
-
-    const int64_t src1_padded_col_size = GGML_PAD(ne10, MATRIX_ROW_PADDING);
-
-    const bool split = src0->backend == GGML_BACKEND_GPU_SPLIT;
-    GGML_ASSERT(!(split && ne02 > 1));
-    GGML_ASSERT(!(split && ne03 > 1));
-    GGML_ASSERT(!(split && ne02 < ne12));
-
-    struct dev_data {
-        cuda_pool_alloc<char>  src0_dd_alloc;
-        cuda_pool_alloc<float> src1_ddf_alloc;
-        cuda_pool_alloc<char>  src1_ddq_alloc;
-        cuda_pool_alloc<float>   dst_dd_alloc;
-
-        char  *  src0_dd = nullptr;
-        float * src1_ddf = nullptr; // float
-        char  * src1_ddq = nullptr; // q8_1
-        float *   dst_dd = nullptr;
-
-        int64_t  row_low;
-        int64_t row_high;
-    };
-
-    dev_data dev[GGML_CUDA_MAX_DEVICES];
-
-    int used_devices = 0;
-
-    for (int id = 0; id < g_device_count; ++id) {
-        // by default, use all rows
-        dev[id].row_low  = 0;
-        dev[id].row_high = ne01;
-
-        // for multi GPU, get the row boundaries from tensor split
-        // and round to mul_mat_q tile sizes
-        if (split) {
-            const int64_t rounding = get_row_rounding(src0->type);
-
-            if (id != 0) {
-                dev[id].row_low  = ne01*g_tensor_split[id];
-                if (dev[id].row_low < ne01) {
-                    dev[id].row_low -= dev[id].row_low % rounding;
-                }
-            }
-
-            if (id != g_device_count - 1) {
-                dev[id].row_high  = ne01*g_tensor_split[id + 1];
-                if (dev[id].row_high < ne01) {
-                    dev[id].row_high -= dev[id].row_high % rounding;
-                }
-            }
-        }
-    }
-
-    for (int id = 0; id < g_device_count; ++id) {
-        if ((!split && id != g_main_device) || dev[id].row_low == dev[id].row_high) {
-            continue;
-        }
-
-        used_devices++;
-
-        const bool src1_on_device = src1->backend == GGML_BACKEND_GPU && id == g_main_device;
-        const bool  dst_on_device =  dst->backend == GGML_BACKEND_GPU && id == g_main_device;
-
-        ggml_cuda_set_device(id);
-        cudaStream_t stream = g_cudaStreams[id][0];
-
-        if (src0_on_device && src0_is_contiguous) {
-            dev[id].src0_dd = (char *) src0_extra->data_device[id];
-        } else {
-            dev[id].src0_dd = dev[id].src0_dd_alloc.alloc(ggml_nbytes(src0));
-        }
-
-        if (src1_on_device && src1_is_contiguous) {
-            dev[id].src1_ddf = (float *) src1_extra->data_device[id];
-        } else {
-            dev[id].src1_ddf = dev[id].src1_ddf_alloc.alloc(ggml_nelements(src1));
-        }
-
-        if (convert_src1_to_q8_1) {
-            dev[id].src1_ddq = dev[id].src1_ddq_alloc.alloc(nrows1*src1_padded_col_size*q8_1_ts/q8_1_bs);
-
-            if (src1_on_device && src1_is_contiguous) {
-                quantize_row_q8_1_cuda(dev[id].src1_ddf, dev[id].src1_ddq, ne10, nrows1, src1_padded_col_size, stream);
-                CUDA_CHECK(cudaGetLastError());
-            }
-        }
-
-        if (dst_on_device) {
-            dev[id].dst_dd = (float *) dst_extra->data_device[id];
-        } else {
-            const size_t size_dst_ddf = split ? (dev[id].row_high - dev[id].row_low)*ne1 : ggml_nelements(dst);
-            dev[id].dst_dd = dev[id].dst_dd_alloc.alloc(size_dst_ddf);
-        }
-    }
-
-    // if multiple devices are used they need to wait for the main device
-    // here an event is recorded that signals that the main device has finished calculating the input data
-    if (split && used_devices > 1) {
-        ggml_cuda_set_device(g_main_device);
-        CUDA_CHECK(cudaEventRecord(src0_extra->events[g_main_device][0], g_cudaStreams[g_main_device][0]));
-    }
-
-    const int64_t src1_col_stride = split && used_devices > 1 ? MUL_MAT_SRC1_COL_STRIDE : ne11;
-    for (int64_t src1_col_0 = 0; src1_col_0 < ne11; src1_col_0 += src1_col_stride) {
-        const int64_t is = split ? (src1_col_0/src1_col_stride) % MAX_STREAMS : 0;
-        const int64_t src1_ncols = src1_col_0 + src1_col_stride > ne11 ? ne11 - src1_col_0 : src1_col_stride;
-
-        for (int id = 0; id < g_device_count; ++id) {
-            if ((!split && id != g_main_device) || dev[id].row_low == dev[id].row_high) {
-                continue;
-            }
-
-            const bool src1_on_device = src1->backend == GGML_BACKEND_GPU && id == g_main_device;
-            const bool  dst_on_device =  dst->backend == GGML_BACKEND_GPU && id == g_main_device;
-            const int64_t row_diff = dev[id].row_high - dev[id].row_low;
-
-            ggml_cuda_set_device(id);
-            cudaStream_t stream = g_cudaStreams[id][is];
-
-            // wait for main GPU data if necessary
-            if (split && (id != g_main_device || is != 0)) {
-                CUDA_CHECK(cudaStreamWaitEvent(stream, src0_extra->events[g_main_device][0], 0));
-            }
-
-            for (int64_t i0 = 0; i0 < ne13*ne12; ++i0) {
-                const int64_t i03 = i0 / ne12;
-                const int64_t i02 = i0 % ne12;
-
-                const size_t src1_ddq_i_offset = (i0*ne11 + src1_col_0) * src1_padded_col_size*q8_1_ts/q8_1_bs;
-
-                // for split tensors the data begins at i0 == i0_offset_low
-                char  *  src0_dd_i =  dev[id].src0_dd + (i0/i02_divisor) * (ne01*ne00*src0_ts)/src0_bs;
-                float * src1_ddf_i = dev[id].src1_ddf + (i0*ne11 + src1_col_0) * ne10;
-                char  * src1_ddq_i = dev[id].src1_ddq +  src1_ddq_i_offset;
-                float *   dst_dd_i =   dev[id].dst_dd + (i0*ne1  + src1_col_0) * (dst_on_device ? ne0 : row_diff);
-
-                // the main device memory buffer can be on VRAM scratch, with space for all partial results
-                // in that case an offset on dst_ddf_i is needed
-                if (dst->backend == GGML_BACKEND_GPU && id == g_main_device) {
-                    dst_dd_i += dev[id].row_low; // offset is 0 if no tensor split
-                }
-
-                // copy src0, src1 to device if necessary
-                if (src1->backend == GGML_BACKEND_GPU && src1_is_contiguous) {
-                    if (id != g_main_device) {
-                        if (convert_src1_to_q8_1) {
-                            char * src1_ddq_i_source = dev[g_main_device].src1_ddq + src1_ddq_i_offset;
-                            CUDA_CHECK(cudaMemcpyPeerAsync(src1_ddq_i, id, src1_ddq_i_source, g_main_device,
-                                                            src1_ncols*src1_padded_col_size*q8_1_ts/q8_1_bs, stream));
-                        } else {
-                            float * src1_ddf_i_source = (float *) src1_extra->data_device[g_main_device];
-                            src1_ddf_i_source += (i0*ne11 + src1_col_0) * ne10;
-                            CUDA_CHECK(cudaMemcpyPeerAsync(src1_ddf_i, id, src1_ddf_i_source, g_main_device,
-                                                            src1_ncols*ne10*sizeof(float), stream));
-                        }
-                    }
-                } else if (src1->backend == GGML_BACKEND_CPU || (src1_on_device && !src1_is_contiguous)) {
-                    CUDA_CHECK(ggml_cuda_cpy_tensor_2d(
-                                src1_ddf_i, src1, i03, i02, src1_col_0, src1_col_0+src1_ncols, stream));
-                } else {
-                    GGML_ASSERT(false);
-                }
-
-                if (convert_src1_to_q8_1 && (src1->backend == GGML_BACKEND_CPU || !src1_is_contiguous)) {
-                    quantize_row_q8_1_cuda(src1_ddf_i, src1_ddq_i, ne10, src1_ncols, src1_padded_col_size, stream);
-                    CUDA_CHECK(cudaGetLastError());
-                }
-
-                if (src1_col_0 == 0 && (!src0_on_device || !src0_is_contiguous) && i02 % i02_divisor == 0) {
-                    CUDA_CHECK(ggml_cuda_cpy_tensor_2d(src0_dd_i, src0, i03, i02/i02_divisor, dev[id].row_low, dev[id].row_high, stream));
-                }
-
-                // do the computation
-                op(src0, src1, dst, src0_dd_i, src1_ddf_i, src1_ddq_i, dst_dd_i,
-                    dev[id].row_low, dev[id].row_high, src1_ncols, src1_padded_col_size, stream);
-                CUDA_CHECK(cudaGetLastError());
-
-                // copy dst to host or other device if necessary
-                if (!dst_on_device) {
-                    void * dst_off_device;
-                    cudaMemcpyKind kind;
-                    if (dst->backend == GGML_BACKEND_CPU) {
-                        dst_off_device = dst->data;
-                        kind = cudaMemcpyDeviceToHost;
-                    } else if (dst->backend == GGML_BACKEND_GPU) {
-                        dst_off_device = dst_extra->data_device[g_main_device];
-                        kind = cudaMemcpyDeviceToDevice;
-                    } else {
-                        GGML_ASSERT(false);
-                    }
-                    if (split) {
-                        // src0 = weight matrix is saved as a transposed matrix for better memory layout.
-                        // dst is NOT transposed.
-                        // The outputs of matrix matrix multiplications can therefore NOT simply be concatenated for >1 GPU.
-                        // Instead they need to be copied to the correct slice in ne0 = dst row index.
-                        // If dst is a vector with ne0 == 1 then you don't have to do this but it still produces correct results.
-                        float * dhf_dst_i = (float *) ((char *) dst_off_device + i02*nb2 + i03*nb3);
-                        GGML_ASSERT(dst->nb[1] == ne0*sizeof(float));
-                        dhf_dst_i += src1_col_0*ne0 + dev[id].row_low;
-#if !defined(GGML_USE_HIPBLAS)
-                        if (kind == cudaMemcpyDeviceToDevice) {
-                            // cudaMemcpy2DAsync may fail with copies between vmm pools of different devices
-                            cudaMemcpy3DPeerParms p = {};
-                            p.dstDevice = g_main_device;
-                            p.dstPtr = make_cudaPitchedPtr(dhf_dst_i, ne0*sizeof(float), row_diff, src1_ncols);
-                            p.srcDevice = id;
-                            p.srcPtr = make_cudaPitchedPtr(dst_dd_i, row_diff*sizeof(float), row_diff, src1_ncols);
-                            p.extent = make_cudaExtent(row_diff*sizeof(float), src1_ncols, 1);
-                            CUDA_CHECK(cudaMemcpy3DPeerAsync(&p, stream));
-                        } else
-#endif
-                        {
-                            CUDA_CHECK(cudaMemcpy2DAsync(dhf_dst_i, ne0*sizeof(float),
-                                                            dst_dd_i, row_diff*sizeof(float),
-                                                            row_diff*sizeof(float), src1_ncols,
-                                                            kind, stream));
-                        }
-                    } else {
-                        float * dhf_dst_i = (float *) ((char *) dst_off_device + i02*nb2 + i03*nb3);
-                        GGML_ASSERT(dst->nb[1] == ne0*sizeof(float));
-                        dhf_dst_i += src1_col_0*ne0;
-                        CUDA_CHECK(cudaMemcpyAsync(dhf_dst_i, dst_dd_i, src1_ncols*ne0*sizeof(float), kind, stream));
-                    }
-                }
-
-                // add event for the main device to wait on until other device is done
-                if (split && (id != g_main_device || is != 0)) {
-                    CUDA_CHECK(cudaEventRecord(src0_extra->events[id][is], stream));
-                }
-            }
-        }
-    }
-
-    // main device waits for all other devices to be finished
-    if (split && g_device_count > 1) {
-        int64_t is_max = (ne11 + MUL_MAT_SRC1_COL_STRIDE - 1) / MUL_MAT_SRC1_COL_STRIDE;
-        is_max = is_max <= MAX_STREAMS ? is_max : MAX_STREAMS;
-
-        ggml_cuda_set_device(g_main_device);
-        for (int id = 0; id < g_device_count; ++id) {
-            if (dev[id].row_low == dev[id].row_high) {
-                continue;
-            }
-            for (int64_t is = 0; is < is_max; ++is) {
-                CUDA_CHECK(cudaStreamWaitEvent(g_cudaStreams[g_main_device][0], src0_extra->events[id][is], 0));
-            }
-        }
-    }
-
-    if (dst->backend == GGML_BACKEND_CPU) {
-        ggml_cuda_set_device(g_main_device);
-        CUDA_CHECK(cudaDeviceSynchronize());
-    }
-}
-
-static void ggml_cuda_repeat(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
-    ggml_cuda_op_flatten(src0, src1, dst, ggml_cuda_op_repeat);
-}
-
-static void ggml_cuda_get_rows(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
-    ggml_cuda_op_flatten(src0, src1, dst, ggml_cuda_op_get_rows);
-}
-
-static void ggml_cuda_add(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
-    ggml_cuda_op_flatten(src0, src1, dst, ggml_cuda_op_add);
-}
-
-static void ggml_cuda_acc(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
-    ggml_cuda_op_flatten(src0, src1, dst, ggml_cuda_op_acc);
-}
-
-static void ggml_cuda_mul(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
-    ggml_cuda_op_flatten(src0, src1, dst, ggml_cuda_op_mul);
-}
-
-static void ggml_cuda_div(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
-    ggml_cuda_op_flatten(src0, src1, dst, ggml_cuda_op_div);
-}
-
-static void ggml_cuda_gelu(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
-    ggml_cuda_op_flatten(src0, src1, dst, ggml_cuda_op_gelu);
-}
-
-static void ggml_cuda_silu(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
-    ggml_cuda_op_flatten(src0, src1, dst, ggml_cuda_op_silu);
-}
-
-static void ggml_cuda_gelu_quick(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
-    ggml_cuda_op_flatten(src0, src1, dst, ggml_cuda_op_gelu_quick);
-}
-
-static void ggml_cuda_tanh(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
-    ggml_cuda_op_flatten(src0, src1, dst, ggml_cuda_op_tanh);
-}
-
-static void ggml_cuda_relu(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
-    ggml_cuda_op_flatten(src0, src1, dst, ggml_cuda_op_relu);
-}
-
-static void ggml_cuda_leaky_relu(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
-    ggml_cuda_op_flatten(src0, src1, dst, ggml_cuda_op_leaky_relu);
-}
-
-static void ggml_cuda_sqr(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
-    ggml_cuda_op_flatten(src0, src1, dst, ggml_cuda_op_sqr);
-}
-
-static void ggml_cuda_norm(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
-    ggml_cuda_op_flatten(src0, src1, dst, ggml_cuda_op_norm);
-}
-
-static void ggml_cuda_group_norm(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
-    ggml_cuda_op_flatten(src0, src1, dst, ggml_cuda_op_group_norm);
-}
-
-static void ggml_cuda_concat(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
-    ggml_cuda_op_flatten(src0, src1, dst, ggml_cuda_op_concat);
-}
-
-static void ggml_cuda_upscale(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
-    ggml_cuda_op_flatten(src0, src1, dst, ggml_cuda_op_upscale);
-}
-
-static void ggml_cuda_pad(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
-    ggml_cuda_op_flatten(src0, src1, dst, ggml_cuda_op_pad);
-}
-
-static void ggml_cuda_rms_norm(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
-    ggml_cuda_op_flatten(src0, src1, dst, ggml_cuda_op_rms_norm);
-}
-
-bool ggml_cuda_can_mul_mat(const struct ggml_tensor * src0, const struct ggml_tensor * src1, struct ggml_tensor * dst) {
-    if (!g_cublas_loaded) return false;
-
-    const int64_t ne10 = src1->ne[0];
-
-    const int64_t ne0 = dst->ne[0];
-    const int64_t ne1 = dst->ne[1];
-
-    // TODO: find the optimal values for these
-    return (src0->type == GGML_TYPE_F32 || src0->type == GGML_TYPE_F16 || ggml_is_quantized(src0->type)) &&
-            src1->type == GGML_TYPE_F32 &&
-             dst->type == GGML_TYPE_F32 &&
-            (ne0 >= 32 && ne1 >= 32 && ne10 >= 32);
-}
-
-static void ggml_cuda_mul_mat_vec_p021(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst){
-    GGML_ASSERT(ggml_is_permuted(src0) && ggml_is_permuted(src1));
-    GGML_ASSERT(src0->backend != GGML_BACKEND_GPU_SPLIT);
-    GGML_ASSERT(src0->nb[0] <= src0->nb[1] && src0->nb[2] <= src0->nb[3]); // 0213 permutation
-    GGML_ASSERT(src1->nb[0] <= src1->nb[1] && src1->nb[2] <= src1->nb[3]); // 0213 permutation
-    GGML_ASSERT(src0->type == GGML_TYPE_F16);
-    GGML_ASSERT(src1->type == GGML_TYPE_F32);
-
-    const int64_t ne00 = src0->ne[0];
-    const int64_t ne01 = src0->ne[1];
-    const int64_t ne02 = src0->ne[2];
-
-    const int64_t ne12 = src1->ne[2];
-
-    ggml_cuda_set_device(g_main_device);
-    cudaStream_t main_stream = g_cudaStreams[g_main_device][0];
-
-    ggml_tensor_extra_gpu * src0_extra = (ggml_tensor_extra_gpu *) src0->extra;
-    void * src0_ddq = src0_extra->data_device[g_main_device];
-
-    ggml_tensor_extra_gpu * src1_extra = (ggml_tensor_extra_gpu *) src1->extra;
-    float * src1_ddf = (float *) src1_extra->data_device[g_main_device];
-
-    ggml_tensor_extra_gpu * dst_extra = (ggml_tensor_extra_gpu *) dst->extra;
-    float * dst_ddf = (float *) dst_extra->data_device[g_main_device];
-
-    ggml_mul_mat_p021_f16_f32_cuda(src0_ddq, src1_ddf, dst_ddf, ne00, ne01, ne02, ne12, main_stream);
-}
-
-static void ggml_cuda_mul_mat_vec_nc(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst){
-    GGML_ASSERT(!ggml_is_transposed(src0));
-    GGML_ASSERT(!ggml_is_transposed(src1));
-    GGML_ASSERT(!ggml_is_permuted(src0));
-    GGML_ASSERT(src0->backend != GGML_BACKEND_GPU_SPLIT);
-    GGML_ASSERT(src0->type == GGML_TYPE_F16);
-    GGML_ASSERT(src1->type == GGML_TYPE_F32);
-
-    const int64_t ne00 = src0->ne[0];
-    const int64_t ne01 = src0->ne[1];
-    const int64_t ne02 = src0->ne[2];
-
-    const int64_t nb01 = src0->nb[1];
-    const int64_t nb02 = src0->nb[2];
-
-    const int64_t ne12 = src1->ne[2];
-
-    ggml_cuda_set_device(g_main_device);
-    cudaStream_t main_stream = g_cudaStreams[g_main_device][0];
-
-    ggml_tensor_extra_gpu * src0_extra = (ggml_tensor_extra_gpu *) src0->extra;
-    void * src0_ddq = src0_extra->data_device[g_main_device];
-
-    ggml_tensor_extra_gpu * src1_extra = (ggml_tensor_extra_gpu *) src1->extra;
-    float * src1_ddf = (float *) src1_extra->data_device[g_main_device];
-
-    ggml_tensor_extra_gpu * dst_extra = (ggml_tensor_extra_gpu *) dst->extra;
-    float * dst_ddf = (float *) dst_extra->data_device[g_main_device];
-
-    const int64_t row_stride_x = nb01 / sizeof(half);
-    const int64_t channel_stride_x = nb02 / sizeof(half);
-
-    ggml_mul_mat_vec_nc_f16_f32_cuda(src0_ddq, src1_ddf, dst_ddf, ne00, ne01, row_stride_x, ne02, ne12, channel_stride_x, main_stream);
-}
-
-static __global__ void k_compute_batched_ptrs(
-        const half * src0_as_f16, const half * src1_as_f16, char * dst,
-        const void ** ptrs_src, void ** ptrs_dst,
-        int64_t ne12, int64_t ne13,
-        int64_t ne23,
-        size_t  nb02, size_t  nb03,
-        size_t  nb12, size_t  nb13,
-        size_t  nbd2, size_t  nbd3,
-        int64_t r2,   int64_t r3) {
-    int64_t i13 = blockIdx.x * blockDim.x + threadIdx.x;
-    int64_t i12 = blockIdx.y * blockDim.y + threadIdx.y;
-
-    if (i13 >= ne13 || i12 >= ne12) {
-        return;
-    }
-
-    int64_t i03 = i13 / r3;
-    int64_t i02 = i12 / r2;
-
-    ptrs_src[0*ne23 + i12 + i13*ne12] = (const char *) src0_as_f16 + i02*nb02 + i03*nb03;
-    ptrs_src[1*ne23 + i12 + i13*ne12] = (const char *) src1_as_f16 + i12*nb12 + i13*nb13;
-    ptrs_dst[0*ne23 + i12 + i13*ne12] = (      char *)         dst + i12*nbd2 + i13*nbd3;
-}
-
-static void ggml_cuda_mul_mat_mat_batched_cublas(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
-    GGML_ASSERT(!ggml_is_transposed(src0));
-    GGML_ASSERT(!ggml_is_transposed(src1));
-
-    GGML_ASSERT(src0->backend != GGML_BACKEND_GPU_SPLIT);
-    GGML_ASSERT(src0->type == GGML_TYPE_F16);
-
-    GGML_TENSOR_BINARY_OP_LOCALS
-
-    const int64_t ne_dst = ggml_nelements(dst);
-
-    ggml_cuda_set_device(g_main_device);
-    cudaStream_t main_stream = g_cudaStreams[g_main_device][0];
-
-    CUBLAS_CHECK(cublasSetStream(g_cublas_handles[g_main_device], main_stream));
-
-    ggml_tensor_extra_gpu * src0_extra = (ggml_tensor_extra_gpu *) src0->extra;
-    void * src0_ddq = src0_extra->data_device[g_main_device];
-    half * src0_f16 = (half *) src0_ddq;
-
-    ggml_tensor_extra_gpu * src1_extra = (ggml_tensor_extra_gpu *) src1->extra;
-    float * src1_ddf = (float *) src1_extra->data_device[g_main_device];
-
-    ggml_tensor_extra_gpu * dst_extra = (ggml_tensor_extra_gpu *) dst->extra;
-    float * dst_ddf = (float *) dst_extra->data_device[g_main_device];
-
-    // convert src1 to fp16
-    cuda_pool_alloc<half> src1_f16_alloc;
-    if (src1->type != GGML_TYPE_F16) {
-        const to_fp16_cuda_t to_fp16_cuda = ggml_get_to_fp16_cuda(src1->type);
-        const int64_t ne_src1 = ggml_nelements(src1);
-        src1_f16_alloc.alloc(ne_src1);
-        GGML_ASSERT(to_fp16_cuda != nullptr);
-        to_fp16_cuda(src1_ddf, src1_f16_alloc.get(), ne_src1, main_stream);
-    }
-    half * src1_f16 = src1->type == GGML_TYPE_F16 ? (half *) src1_ddf : src1_f16_alloc.get();
-
-    cuda_pool_alloc<half> dst_f16;
-    char * dst_t;
-
-    cublasComputeType_t cu_compute_type = CUBLAS_COMPUTE_16F;
-    cudaDataType_t      cu_data_type    = CUDA_R_16F;
-
-    // dst strides
-    size_t nbd2 = dst->nb[2];
-    size_t nbd3 = dst->nb[3];
-
-    const half  alpha_f16 = 1.0f;
-    const half  beta_f16  = 0.0f;
-
-    const float alpha_f32 = 1.0f;
-    const float beta_f32  = 0.0f;
-
-    const void * alpha = &alpha_f16;
-    const void * beta  = &beta_f16;
-
-    if (dst->op_params[0] == GGML_PREC_DEFAULT) {
-        dst_t = (char *) dst_f16.alloc(ne_dst);
-
-        nbd2 /= sizeof(float) / sizeof(half);
-        nbd3 /= sizeof(float) / sizeof(half);
-    } else {
-        dst_t = (char *) dst_ddf;
-
-        cu_compute_type = CUBLAS_COMPUTE_32F;
-        cu_data_type    = CUDA_R_32F;
-
-        alpha = &alpha_f32;
-        beta  = &beta_f32;
-    }
-
-    GGML_ASSERT(ne12 % ne02 == 0);
-    GGML_ASSERT(ne13 % ne03 == 0);
-
-    // broadcast factors
-    const int64_t r2 = ne12/ne02;
-    const int64_t r3 = ne13/ne03;
-
-#if 0
-    // use cublasGemmEx
-    {
-        for (int i13 = 0; i13 < ne13; ++i13) {
-            for (int i12 = 0; i12 < ne12; ++i12) {
-                int i03 = i13 / r3;
-                int i02 = i12 / r2;
-
-                CUBLAS_CHECK(
-                        cublasGemmEx(g_cublas_handles[g_main_device], CUBLAS_OP_T, CUBLAS_OP_N,
-                            ne01, ne11, ne10,
-                            alpha, (const char *) src0_as_f16 + i02*src0->nb[2]   + i03*src0->nb[3]  , CUDA_R_16F,   nb01/sizeof(half),
-                                   (const char *) src1_as_f16 + i12*src1->nb[2]/2 + i13*src1->nb[3]/2, CUDA_R_16F,   nb11/sizeof(float),
-                            beta,  (      char *)       dst_t + i12*nbd2          + i13*nbd3,          cu_data_type, ne01,
-                            cu_compute_type,
-                            CUBLAS_GEMM_DEFAULT_TENSOR_OP));
-            }
-        }
-    }
-#else
-    if (r2 == 1 && r3 == 1 && src0->nb[2]*src0->ne[2] == src0->nb[3] && src1->nb[2]*src1->ne[2] == src1->nb[3]) {
-        // there is no broadcast and src0, src1 are contiguous across dims 2, 3
-        // use cublasGemmStridedBatchedEx
-        CUBLAS_CHECK(
-        cublasGemmStridedBatchedEx(g_cublas_handles[g_main_device], CUBLAS_OP_T, CUBLAS_OP_N,
-                ne01, ne11, ne10,
-                alpha, (const char *) src0_f16, CUDA_R_16F,   nb01/nb00, nb02/nb00,  // strideA
-                       (const char *) src1_f16, CUDA_R_16F,   nb11/nb10, nb12/nb10,  // strideB
-                beta,  (      char *)    dst_t, cu_data_type, ne01,       nb2/nb0,   // strideC
-                ne12*ne13,
-                cu_compute_type,
-                CUBLAS_GEMM_DEFAULT_TENSOR_OP));
-    } else {
-        // use cublasGemmBatchedEx
-        const int ne23 = ne12*ne13;
-
-        cuda_pool_alloc<const void *> ptrs_src(2*ne23);
-        cuda_pool_alloc<      void *> ptrs_dst(1*ne23);
-
-        dim3 block_dims(ne13, ne12);
-        k_compute_batched_ptrs<<<1, block_dims, 0, main_stream>>>(
-                src0_f16, src1_f16, dst_t,
-                ptrs_src.get(), ptrs_dst.get(),
-                ne12, ne13,
-                ne23,
-                nb02, nb03,
-                src1->type == GGML_TYPE_F16 ? nb12 : nb12/2,
-                src1->type == GGML_TYPE_F16 ? nb13 : nb13/2,
-                nbd2, nbd3,
-                r2, r3);
-        CUDA_CHECK(cudaGetLastError());
-
-        CUBLAS_CHECK(
-        cublasGemmBatchedEx(g_cublas_handles[g_main_device], CUBLAS_OP_T, CUBLAS_OP_N,
-                ne01, ne11, ne10,
-                alpha, (const void **) (ptrs_src.get() + 0*ne23), CUDA_R_16F,   nb01/nb00,
-                       (const void **) (ptrs_src.get() + 1*ne23), CUDA_R_16F,   nb11/nb10,
-                beta,  (      void **) (ptrs_dst.get() + 0*ne23), cu_data_type, ne01,
-                ne23,
-                cu_compute_type,
-                CUBLAS_GEMM_DEFAULT_TENSOR_OP));
-    }
-#endif
-
-    if (dst->op_params[0] == GGML_PREC_DEFAULT) {
-        const to_fp32_cuda_t to_fp32_cuda = ggml_get_to_fp32_cuda(GGML_TYPE_F16);
-        to_fp32_cuda(dst_f16.get(), dst_ddf, ne_dst, main_stream);
-    }
-}
-
-static void ggml_cuda_mul_mat(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
-    const bool all_on_device =
-        (src0->backend == GGML_BACKEND_GPU || src0->backend == GGML_BACKEND_GPU_SPLIT) &&
-        (src1->backend == GGML_BACKEND_GPU) &&
-        ( dst->backend == GGML_BACKEND_GPU);
-
-    const bool split = src0->backend == GGML_BACKEND_GPU_SPLIT;
-
-    int64_t min_compute_capability = INT_MAX;
-    for (int id = 0; id < g_device_count; ++id) {
-        if (min_compute_capability > g_device_caps[id].cc && g_tensor_split[id] < (id + 1 < g_device_count ? g_tensor_split[id + 1] : 1.0f)) {
-            min_compute_capability = g_device_caps[id].cc;
-        }
-    }
-
-#if defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)
-
-    const bool fp16_performance_good = min_compute_capability >= CC_RDNA1;
-    bool               use_mul_mat_q = ggml_is_quantized(src0->type);
-#ifdef CUDA_USE_TENSOR_CORES
-    use_mul_mat_q = use_mul_mat_q && min_compute_capability < CC_RDNA3;
-#endif // CUDA_USE_TENSOR_CORES
-
-#else
-
-    const bool fp16_performance_good = min_compute_capability >= CC_VOLTA;
-    bool               use_mul_mat_q = min_compute_capability >= MIN_CC_DP4A && ggml_is_quantized(src0->type);
-#ifdef CUDA_USE_TENSOR_CORES
-    // when tensor cores are available, use them for large batch size
-    // ref: https://github.com/ggerganov/llama.cpp/pull/3776
-    use_mul_mat_q = use_mul_mat_q && !(fp16_performance_good && src1->ne[1] > MMQ_MAX_BATCH_SIZE);
-#endif // CUDA_USE_TENSOR_CORES
-
-#endif // defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)
-
-    // debug helpers
-    //printf("src0: %8d %8d %8d %8d\n", src0->ne[0], src0->ne[1], src0->ne[2], src0->ne[3]);
-    //printf("      %8d %8d %8d %8d\n", src0->nb[0], src0->nb[1], src0->nb[2], src0->nb[3]);
-    //printf("src1: %8d %8d %8d %8d\n", src1->ne[0], src1->ne[1], src1->ne[2], src1->ne[3]);
-    //printf("      %8d %8d %8d %8d\n", src1->nb[0], src1->nb[1], src1->nb[2], src1->nb[3]);
-    //printf("src0 is contiguous %d, transposed %d, type = %s, name = %s\n", ggml_is_contiguous(src0), ggml_is_transposed(src0), ggml_type_name(src0->type), src0->name);
-    //printf("src1 is contiguous %d, transposed %d, type = %s, name = %s\n", ggml_is_contiguous(src1), ggml_is_transposed(src1), ggml_type_name(src1->type), src1->name);
-
-    if (!split && all_on_device && !fp16_performance_good && src0->type == GGML_TYPE_F16 && ggml_is_permuted(src0) && ggml_is_permuted(src1) && src1->ne[1] == 1) {
-        // KQ single-batch
-        ggml_cuda_mul_mat_vec_p021(src0, src1, dst);
-    } else if (!split && all_on_device && !fp16_performance_good && src0->type == GGML_TYPE_F16 && !ggml_is_contiguous(src0) && !ggml_is_transposed(src1) && src1->ne[1] == 1) {
-        // KQV single-batch
-        ggml_cuda_mul_mat_vec_nc(src0, src1, dst);
-    } else if (!split && all_on_device && fp16_performance_good && src0->type == GGML_TYPE_F16 && !ggml_is_transposed(src0) && !ggml_is_transposed(src1)) {
-        // KQ + KQV multi-batch
-        ggml_cuda_mul_mat_mat_batched_cublas(src0, src1, dst);
-    } else if (src0->type == GGML_TYPE_F32) {
-        ggml_cuda_op_mul_mat(src0, src1, dst, ggml_cuda_op_mul_mat_cublas, false);
-    } else if (ggml_is_quantized(src0->type) || src0->type == GGML_TYPE_F16) {
-        if (src1->ne[1] == 1 && src0->ne[0] % GGML_CUDA_DMMV_X == 0 && src1->type == GGML_TYPE_F32) {
-#ifdef GGML_CUDA_FORCE_DMMV
-            const bool use_mul_mat_vec_q = false;
-#else
-            const bool use_mul_mat_vec_q = min_compute_capability >= MIN_CC_DP4A && ggml_is_quantized(src0->type) && ggml_nrows(src1) == 1;
-#endif // GGML_CUDA_FORCE_DMMV
-
-            if (use_mul_mat_vec_q) {
-                // NOTE: this kernel does not support ggml_nrows(src1) > 1
-                ggml_cuda_op_mul_mat(src0, src1, dst, ggml_cuda_op_mul_mat_vec_q, true);
-            } else {
-                ggml_cuda_op_mul_mat(src0, src1, dst, ggml_cuda_op_dequantize_mul_mat_vec, false);
-            }
-        } else {
-            if (use_mul_mat_q) {
-                ggml_cuda_op_mul_mat(src0, src1, dst, ggml_cuda_op_mul_mat_q, true);
-            } else {
-                ggml_cuda_op_mul_mat(src0, src1, dst, ggml_cuda_op_mul_mat_cublas, false);
-            }
-        }
-    } else {
-        GGML_ASSERT(false);
-    }
-}
-
-#if 0
-template<typename ... Srcs>
-static __global__ void k_compute_batched_ptrs_id(
-        const void ** ptrs_src, void ** ptrs_dst,
-        int ne12, int ne13,
-        int ne23,
-        int nb02, int nb03,
-        int nb12, int nb13,
-        int nb2, int nb3,
-        int r2, int r3,
-        ggml_type src0_type, half * src0_as_f16, int64_t src0_ne,
-        const half * src1_f16, half * dst_f16,
-        const int32_t * ids, const int id,
-        Srcs... src0s) {
-
-    int i = ids[id];
-
-    half * src0_f16;
-    const void * srcs_ar[] = { (const half *) src0s... };
-    if (src0_type == GGML_TYPE_F16) {
-        src0_f16 = (half *) srcs_ar[i];
-    } else {
-        src0_f16 = src0_as_f16;
-        if (threadIdx.x == 0 && threadIdx.y == 0) {
-            const to_fp16_cuda_t to_fp16 = ggml_get_to_fp16_cuda(src0_type);
-            to_fp16(srcs_ar[i], src0_f16, src0_ne, cudaStreamFireAndForget);
-        }
-    }
-
-    int i13 = blockIdx.x * blockDim.x + threadIdx.x;
-    int i12 = blockIdx.y * blockDim.y + threadIdx.y;
-
-    if (i13 >= ne13 || i12 >= ne12) {
-        return;
-    }
-
-    int i03 = i13 / r3;
-    int i02 = i12 / r2;
-
-    ptrs_src[0*ne23 + i12 + i13*ne12] = (const char *) src0_f16 + i02*nb02   + i03*nb03;
-    ptrs_src[1*ne23 + i12 + i13*ne12] = (const char *) src1_f16 + i12*nb12/2 + i13*nb13/2;
-    ptrs_dst[0*ne23 + i12 + i13*ne12] = (      char *)  dst_f16 + i12* nb2/2 + i13* nb3/2;
-}
-
-static void ggml_cuda_mul_mat_id_cublas(ggml_tensor * dst) {
-    const struct ggml_tensor * ids = dst->src[0];
-    const struct ggml_tensor * src1 = dst->src[1];
-    const struct ggml_tensor * src00 = dst->src[2];
-
-    const int id = dst->op_params[0];
-
-    GGML_ASSERT(!ggml_is_transposed(src00));
-    GGML_ASSERT(!ggml_is_transposed(src1));
-
-    GGML_ASSERT(src00->backend != GGML_BACKEND_GPU_SPLIT);
-    GGML_ASSERT(src1->type == GGML_TYPE_F32);
-
-    const int64_t ne00 = src00->ne[0]; GGML_UNUSED(ne00);
-    const int64_t ne01 = src00->ne[1];
-    const int64_t ne02 = src00->ne[2];
-    const int64_t ne03 = src00->ne[3];
-
-    //const int64_t nb01 = src00->nb[1];
-    const int64_t nb02 = src00->nb[2]; GGML_UNUSED(nb02);
-    const int64_t nb03 = src00->nb[3]; GGML_UNUSED(nb03);
-
-    const int64_t ne10 = src1->ne[0];
-    const int64_t ne11 = src1->ne[1];
-    const int64_t ne12 = src1->ne[2];
-    const int64_t ne13 = src1->ne[3];
-
-    //const int64_t nb11 = src1->nb[1];
-    const int64_t nb12 = src1->nb[2]; GGML_UNUSED(nb12);
-    const int64_t nb13 = src1->nb[3]; GGML_UNUSED(nb13);
-
-    const int64_t ne1 = ggml_nelements(src1);
-    const int64_t ne  = ggml_nelements(dst);
-
-    ggml_cuda_set_device(g_main_device);
-    cudaStream_t main_stream = g_cudaStreams[g_main_device][0];
-
-    CUBLAS_CHECK(cublasSetStream(g_cublas_handles[g_main_device], main_stream));
-
-    //ggml_tensor_extra_gpu * src0_extra = (ggml_tensor_extra_gpu *) src0->extra;
-    //void * src0_ddq = src0_extra->data_device[g_main_device];
-    //half * src0_as_f16 = (half *) src0_ddq;
-
-    ggml_tensor_extra_gpu * src1_extra = (ggml_tensor_extra_gpu *) src1->extra;
-    float * src1_ddf = (float *) src1_extra->data_device[g_main_device];
-
-    ggml_tensor_extra_gpu * dst_extra = (ggml_tensor_extra_gpu *) dst->extra;
-    float * dst_ddf = (float *) dst_extra->data_device[g_main_device];
-
-    // convert src1 to fp16
-    const to_fp16_cuda_t to_fp16_cuda = ggml_get_to_fp16_cuda(src1->type);
-    GGML_ASSERT(to_fp16_cuda != nullptr);
-
-    size_t src1_as = 0;
-    half * src1_as_f16 = (half *) ggml_cuda_pool_malloc(ne1 * sizeof(half), &src1_as);
-    to_fp16_cuda(src1_ddf, src1_as_f16, ne1, main_stream);
-
-    size_t dst_as = 0;
-    half * dst_f16 = (half *) ggml_cuda_pool_malloc(ne * sizeof(half), &dst_as);
-
-    GGML_ASSERT(ne12 % ne02 == 0);
-    GGML_ASSERT(ne13 % ne03 == 0);
-
-    // broadcast factors
-    const int64_t r2 = ne12/ne02;
-    const int64_t r3 = ne13/ne03;
-
-    const half alpha_f16 = 1.0f;
-    const half beta_f16  = 0.0f;
-
-    // use cublasGemmBatchedEx
-    const int ne23 = ne12*ne13;
-
-    const void ** ptrs_src = nullptr;
-          void ** ptrs_dst = nullptr;
-
-    size_t ptrs_src_s = 0;
-    size_t ptrs_dst_s = 0;
-
-    ptrs_src = (const void **) ggml_cuda_pool_malloc(2*ne23*sizeof(void *), &ptrs_src_s);
-    ptrs_dst = (      void **) ggml_cuda_pool_malloc(1*ne23*sizeof(void *), &ptrs_dst_s);
-
-    int64_t src0_ne = ggml_nelements(src00);
-    half * src0_as_f16 = nullptr;
-    size_t src0_as = 0;
-    if (src00->type != GGML_TYPE_F16) {
-        src0_as_f16 = (half *) ggml_cuda_pool_malloc(src0_ne * sizeof(half), &src0_as);
-    }
-
-    static_assert(GGML_MAX_SRC == 6, "GGML_MAX_SRC == 6");
-    dim3 block_dims(ne13, ne12);
-    k_compute_batched_ptrs_id<<<1, block_dims, 0, main_stream>>>(
-            ptrs_src, ptrs_dst,
-            ne12, ne13,
-            ne23,
-            ne00*ne01*sizeof(half), ne00*ne01*ne02*sizeof(half),
-            nb12, nb13,
-            dst->nb[2], dst->nb[3],
-            r2, r3,
-            src00->type, src0_as_f16, src0_ne,
-            src1_as_f16, dst_f16,
-            (const int *)((ggml_tensor_extra_gpu *)ids->extra)->data_device[g_main_device], id,
-            dst->src[2] ? (const half *)((ggml_tensor_extra_gpu *)dst->src[2]->extra)->data_device[g_main_device] : nullptr,
-            dst->src[3] ? (const half *)((ggml_tensor_extra_gpu *)dst->src[3]->extra)->data_device[g_main_device] : nullptr,
-            dst->src[4] ? (const half *)((ggml_tensor_extra_gpu *)dst->src[4]->extra)->data_device[g_main_device] : nullptr,
-            dst->src[5] ? (const half *)((ggml_tensor_extra_gpu *)dst->src[5]->extra)->data_device[g_main_device] : nullptr
-    );
-    CUDA_CHECK(cudaGetLastError());
-
-    CUBLAS_CHECK(
-    cublasGemmBatchedEx(g_cublas_handles[g_main_device], CUBLAS_OP_T, CUBLAS_OP_N,
-            ne01, ne11, ne10,
-            &alpha_f16, (const void **) (ptrs_src + 0*ne23), CUDA_R_16F, ne00,
-                        (const void **) (ptrs_src + 1*ne23), CUDA_R_16F, ne10,
-            &beta_f16,  (      void **) (ptrs_dst + 0*ne23), CUDA_R_16F, ne01,
-            ne23,
-            CUBLAS_COMPUTE_16F,
-            CUBLAS_GEMM_DEFAULT_TENSOR_OP));
-
-    if (src0_as != 0) {
-        ggml_cuda_pool_free(src0_as_f16, src0_as);
-    }
-    if (ptrs_src_s != 0) {
-        ggml_cuda_pool_free(ptrs_src, ptrs_src_s);
-    }
-    if (ptrs_dst_s != 0) {
-        ggml_cuda_pool_free(ptrs_dst, ptrs_dst_s);
-    }
-
-    const to_fp32_cuda_t to_fp32_cuda = ggml_get_to_fp32_cuda(GGML_TYPE_F16);
-    to_fp32_cuda(dst_f16, dst_ddf, ne, main_stream);
-
-    ggml_cuda_pool_free(src1_as_f16, src1_as);
-    ggml_cuda_pool_free(dst_f16, dst_as);
-}
-#endif
-
-static void ggml_cuda_mul_mat_id(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
-#if 0
-    ggml_cuda_mul_mat_id_cublas(dst);
-    // TODO: mmq/mmv support
-#endif
-
-    const int64_t nb11 = src1->nb[1];
-    const int64_t nb1  =  dst->nb[1];
-
-    const struct ggml_tensor * ids = src0;
-    const int32_t id = ((int32_t *) dst->op_params)[0];
-    const int32_t n_as = ((int32_t *) dst->op_params)[1];
-
-    std::vector<char> ids_host(ggml_nbytes(ids));
-
-    cudaStream_t stream = g_cudaStreams[g_main_device][0];
-
-    if (ids->backend == GGML_BACKEND_GPU) {
-        const char * ids_dev = (const char *)((const ggml_tensor_extra_gpu *)ids->extra)->data_device[g_main_device];
-        CUDA_CHECK(cudaMemcpyAsync(ids_host.data(), ids_dev, ggml_nbytes(ids), cudaMemcpyDeviceToHost, stream));
-        CUDA_CHECK(cudaStreamSynchronize(stream));
-    } else {
-        memcpy(ids_host.data(), ids->data, ggml_nbytes(ids));
-    }
-
-    const ggml_tensor_extra_gpu * src1_extra = (const ggml_tensor_extra_gpu *) src1->extra;
-    const ggml_tensor_extra_gpu * dst_extra = (const ggml_tensor_extra_gpu *) dst->extra;
-
-    ggml_tensor_extra_gpu src1_row_extra;
-    ggml_tensor_extra_gpu dst_row_extra;
-
-    ggml_tensor src1_row = *src1;
-    ggml_tensor dst_row = *dst;
-
-    src1_row.backend = GGML_BACKEND_GPU;
-    dst_row.backend  = GGML_BACKEND_GPU;
-
-    src1_row.extra = &src1_row_extra;
-    dst_row.extra = &dst_row_extra;
-
-    char * src1_original = src1->backend == GGML_BACKEND_CPU ?
-        (char *) src1->data : (char *) src1_extra->data_device[g_main_device];
-    char * dst_original  =  dst->backend == GGML_BACKEND_CPU ?
-        (char *)  dst->data : (char *)  dst_extra->data_device[g_main_device];
-
-    if (src1->ne[1] == 1) {
-        GGML_ASSERT(src1->backend == GGML_BACKEND_GPU);
-        GGML_ASSERT(dst->backend  == GGML_BACKEND_GPU);
-
-        for (int64_t i01 = 0; i01 < ids->ne[1]; i01++) {
-            //int32_t row_id;
-            //CUDA_CHECK(cudaMemcpyAsync(&row_id, ids_dev + i01*ids->nb[1] + id*ids->nb[0], sizeof(int32_t), cudaMemcpyDeviceToHost, g_cudaStreams[g_main_device][0]));
-            //CUDA_CHECK(cudaStreamSynchronize(g_cudaStreams[g_main_device][0]));
-
-            const int32_t row_id = *(const int32_t *) (ids_host.data() + i01*ids->nb[1] + id*ids->nb[0]);
-
-            GGML_ASSERT(row_id >= 0 && row_id < n_as);
-
-            const struct ggml_tensor * src0_row = dst->src[row_id + 2];
-
-            src1_row_extra.data_device[g_main_device] = src1_original + i01*src1->nb[1];
-            src1_row.data = (char *) src1->data + i01*src1->nb[1]; // TODO why is this set?
-
-            dst_row_extra.data_device[g_main_device] = dst_original + i01*dst->nb[1];
-            dst_row.data = (char *) dst->data + i01*dst->nb[1]; // TODO why is this set?
-
-            ggml_cuda_mul_mat(src0_row, &src1_row, &dst_row);
-        }
-    } else {
-        cuda_pool_alloc<char> src1_contiguous(sizeof(float)*ggml_nelements(src1));
-        cuda_pool_alloc<char>  dst_contiguous(sizeof(float)*ggml_nelements(dst));
-
-        src1_row_extra.data_device[g_main_device] = src1_contiguous.get();
-        dst_row_extra.data_device[g_main_device]  =  dst_contiguous.get();
-
-        const cudaMemcpyKind src1_kind = src1->backend == GGML_BACKEND_CPU ?
-            cudaMemcpyHostToDevice : cudaMemcpyDeviceToDevice;
-        const cudaMemcpyKind dst_kind  =  dst->backend == GGML_BACKEND_CPU ?
-            cudaMemcpyDeviceToHost : cudaMemcpyDeviceToDevice;
-
-        for (int32_t row_id = 0; row_id < n_as; ++row_id) {
-            const struct ggml_tensor * src0_row = dst->src[row_id + 2];
-
-            int64_t num_src1_rows = 0;
-            for (int64_t i01 = 0; i01 < ids->ne[1]; i01++) {
-                const int32_t row_id_i = *(const int32_t *) (ids_host.data() + i01*ids->nb[1] + id*ids->nb[0]);
-
-                if (row_id_i != row_id) {
-                    continue;
-                }
-
-                GGML_ASSERT(row_id >= 0 && row_id < n_as);
-
-                CUDA_CHECK(cudaMemcpyAsync(src1_contiguous.get() + num_src1_rows*nb11, src1_original + i01*nb11,
-                                        nb11, src1_kind, stream));
-                num_src1_rows++;
-            }
-
-            if (num_src1_rows == 0) {
-                continue;
-            }
-
-            src1_row.ne[1] = num_src1_rows;
-            dst_row.ne[1] = num_src1_rows;
-
-            src1_row.nb[1] = nb11;
-            src1_row.nb[2] = num_src1_rows*nb11;
-            src1_row.nb[3] = num_src1_rows*nb11;
-
-            dst_row.nb[1] = nb1;
-            dst_row.nb[2] = num_src1_rows*nb1;
-            dst_row.nb[3] = num_src1_rows*nb1;
-
-            ggml_cuda_mul_mat(src0_row, &src1_row, &dst_row);
-
-            num_src1_rows = 0;
-            for (int64_t i01 = 0; i01 < ids->ne[1]; i01++) {
-                const int32_t row_id_i = *(const int32_t *) (ids_host.data() + i01*ids->nb[1] + id*ids->nb[0]);
-
-                if (row_id_i != row_id) {
-                    continue;
-                }
-
-                GGML_ASSERT(row_id >= 0 && row_id < n_as);
-
-                CUDA_CHECK(cudaMemcpyAsync(dst_original + i01*nb1, dst_contiguous.get() + num_src1_rows*nb1,
-                                        nb1, dst_kind, stream));
-                num_src1_rows++;
-            }
-        }
-    }
-
-    if (dst->backend == GGML_BACKEND_CPU) {
-        CUDA_CHECK(cudaStreamSynchronize(stream));
-    }
-}
-
-static void ggml_cuda_scale(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
-    ggml_cuda_op_flatten(src0, src1, dst, ggml_cuda_op_scale);
-}
-
-static void ggml_cuda_clamp(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
-    ggml_cuda_op_flatten(src0, src1, dst, ggml_cuda_op_clamp);
-}
-
-static void ggml_cuda_cpy(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
-    const int64_t ne = ggml_nelements(src0);
-    GGML_ASSERT(ne == ggml_nelements(src1));
-
-    GGML_ASSERT(src0->backend == GGML_BACKEND_GPU);
-    GGML_ASSERT(src1->backend == GGML_BACKEND_GPU);
-
-    GGML_ASSERT(ggml_nbytes(src0) <= INT_MAX);
-    GGML_ASSERT(ggml_nbytes(src1) <= INT_MAX);
-
-    const int64_t ne00 = src0->ne[0];
-    const int64_t ne01 = src0->ne[1];
-    GGML_ASSERT(src0->ne[3] == 1);
-
-    const int64_t nb00 = src0->nb[0];
-    const int64_t nb01 = src0->nb[1];
-    const int64_t nb02 = src0->nb[2];
-
-    const int64_t ne10 = src1->ne[0];
-    const int64_t ne11 = src1->ne[1];
-    GGML_ASSERT(src1->ne[3] == 1);
-
-    const int64_t nb10 = src1->nb[0];
-    const int64_t nb11 = src1->nb[1];
-    const int64_t nb12 = src1->nb[2];
-
-    ggml_cuda_set_device(g_main_device);
-    cudaStream_t main_stream = g_cudaStreams[g_main_device][0];
-
-    const ggml_tensor_extra_gpu * src0_extra = (ggml_tensor_extra_gpu *) src0->extra;
-    const ggml_tensor_extra_gpu * src1_extra = (ggml_tensor_extra_gpu *) src1->extra;
-
-    char * src0_ddc = (char *) src0_extra->data_device[g_main_device];
-    char * src1_ddc = (char *) src1_extra->data_device[g_main_device];
-
-    if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_F32) {
-        ggml_cpy_f32_f32_cuda (src0_ddc, src1_ddc, ne, ne00, ne01, nb00, nb01, nb02, ne10, ne11, nb10, nb11, nb12, main_stream);
-    } else if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_F16) {
-        ggml_cpy_f32_f16_cuda (src0_ddc, src1_ddc, ne, ne00, ne01, nb00, nb01, nb02, ne10, ne11, nb10, nb11, nb12, main_stream);
-    } else if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_Q8_0) {
-        ggml_cpy_f32_q8_0_cuda(src0_ddc, src1_ddc, ne, ne00, ne01, nb00, nb01, nb02, ne10, ne11, nb10, nb11, nb12, main_stream);
-    } else if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_Q4_0) {
-        ggml_cpy_f32_q4_0_cuda(src0_ddc, src1_ddc, ne, ne00, ne01, nb00, nb01, nb02, ne10, ne11, nb10, nb11, nb12, main_stream);
-    } else if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_Q4_1) {
-        ggml_cpy_f32_q4_1_cuda(src0_ddc, src1_ddc, ne, ne00, ne01, nb00, nb01, nb02, ne10, ne11, nb10, nb11, nb12, main_stream);
-    } else if (src0->type == GGML_TYPE_F16 && src1->type == GGML_TYPE_F16) {
-        ggml_cpy_f16_f16_cuda (src0_ddc, src1_ddc, ne, ne00, ne01, nb00, nb01, nb02, ne10, ne11, nb10, nb11, nb12, main_stream);
-    } else {
-        fprintf(stderr, "%s: unsupported type combination (%s to %s)\n", __func__,
-                ggml_type_name(src0->type), ggml_type_name(src1->type));
-        GGML_ASSERT(false);
-    }
-
-    (void) dst;
-}
-
-static void ggml_cuda_dup(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
-    // TODO: why do we pass dst as src1 here?
-    ggml_cuda_cpy(src0, dst, nullptr);
-    (void) src1;
-}
-
-static void ggml_cuda_diag_mask_inf(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
-    ggml_cuda_op_flatten(src0, src1, dst, ggml_cuda_op_diag_mask_inf);
-}
-
-static void ggml_cuda_soft_max(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
-    ggml_cuda_op_flatten(src0, src1, dst, ggml_cuda_op_soft_max);
-}
-
-static void ggml_cuda_rope(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
-    GGML_ASSERT(ggml_is_contiguous(src0)); // TODO: this restriction is temporary until non-cont support is implemented
-    ggml_cuda_op_flatten(src0, src1, dst, ggml_cuda_op_rope);
-}
-
-static void ggml_cuda_alibi(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
-    ggml_cuda_op_flatten(src0, src1, dst, ggml_cuda_op_alibi);
-}
-
-static void ggml_cuda_im2col(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
-    ggml_cuda_op_flatten(src0, src1, dst, ggml_cuda_op_im2col);
-}
-
-static void ggml_cuda_sum_rows(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
-    GGML_ASSERT(ggml_is_contiguous(src0));
-    ggml_cuda_op_flatten(src0, src1, dst, ggml_cuda_op_sum_rows);
-}
-
-static void ggml_cuda_argsort(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
-    GGML_ASSERT(ggml_is_contiguous(src0));
-    ggml_cuda_op_flatten(src0, src1, dst, ggml_cuda_op_argsort);
-}
-
-static void ggml_cuda_nop(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
-    (void) src0;
-    (void) src1;
-    (void) dst;
-}
-
-static size_t ggml_nbytes_split(const struct ggml_tensor * tensor, int nrows_split) {
-    static_assert(GGML_MAX_DIMS == 4, "GGML_MAX_DIMS is not 4 - update this function");
-
-    return nrows_split*ggml_row_size(tensor->type, tensor->ne[0]);
-}
-
-void ggml_cuda_transform_tensor(void * data, struct ggml_tensor * tensor) {
-    const int64_t nrows = ggml_nrows(tensor);
-
-    const int64_t ne0 = tensor->ne[0];
-
-    const size_t nb1 = tensor->nb[1];
-
-    ggml_backend_type backend = tensor->backend;
-    ggml_tensor_extra_gpu * extra = new struct ggml_tensor_extra_gpu;
-    memset(extra, 0, sizeof(*extra));
-
-    for (int id = 0; id < g_device_count; ++id) {
-        if (backend == GGML_BACKEND_GPU && id != g_main_device) {
-            continue;
-        }
-
-        ggml_cuda_set_device(id);
-
-        int64_t row_low, row_high;
-        if (backend == GGML_BACKEND_GPU) {
-            row_low = 0;
-            row_high = nrows;
-        } else if (backend == GGML_BACKEND_GPU_SPLIT) {
-            const int64_t rounding = get_row_rounding(tensor->type);
-
-            row_low = id == 0 ? 0 : nrows*g_tensor_split[id];
-            row_low -= row_low % rounding;
-
-            if (id == g_device_count - 1) {
-                row_high = nrows;
-            } else {
-                row_high = nrows*g_tensor_split[id + 1];
-                row_high -= row_high % rounding;
-            }
-        } else {
-            GGML_ASSERT(false);
-        }
-        if (row_low == row_high) {
-            continue;
-        }
-
-        int64_t nrows_split = row_high - row_low;
-
-        const size_t offset_split = row_low*nb1;
-        size_t size = ggml_nbytes_split(tensor, nrows_split);
-        const size_t original_size = size;
-
-        // pad last row to a multiple of 512 elements to avoid out-of-bounds memory accesses
-        if (ne0 % MATRIX_ROW_PADDING != 0) {
-            size += ggml_row_size(tensor->type, MATRIX_ROW_PADDING - ne0 % MATRIX_ROW_PADDING);
-        }
-
-        char * buf;
-        CUDA_CHECK(cudaMalloc(&buf, size));
-        char * buf_host = (char *)data + offset_split;
-
-        // set padding to 0 to avoid possible NaN values
-        if (size > original_size) {
-            CUDA_CHECK(cudaMemset(buf + original_size, 0, size - original_size));
-        }
-
-        CUDA_CHECK(cudaMemcpy(buf, buf_host, original_size, cudaMemcpyHostToDevice));
-
-        extra->data_device[id] = buf;
-
-        if (backend == GGML_BACKEND_GPU_SPLIT) {
-            for (int64_t is = 0; is < MAX_STREAMS; ++is) {
-                CUDA_CHECK(cudaEventCreateWithFlags(&extra->events[id][is], cudaEventDisableTiming));
-            }
-        }
-    }
-
-    tensor->extra = extra;
-}
-
-void ggml_cuda_free_data(struct ggml_tensor * tensor) {
-    if (!tensor || !tensor->extra || (tensor->backend != GGML_BACKEND_GPU && tensor->backend != GGML_BACKEND_GPU_SPLIT) ) {
-        return;
-    }
-
-    ggml_tensor_extra_gpu * extra = (ggml_tensor_extra_gpu *) tensor->extra;
-
-    for (int id = 0; id < g_device_count; ++id) {
-        ggml_cuda_set_device(id);
-        if (extra->data_device[id] != nullptr) {
-            CUDA_CHECK(cudaFree(extra->data_device[id]));
-        }
-
-        for (int64_t is = 0; is < MAX_STREAMS; ++is) {
-            if (extra->events[id][is] != nullptr) {
-                CUDA_CHECK(cudaEventDestroy(extra->events[id][is]));
-            }
-        }
-    }
-
-    delete extra;
-}
-
-static ggml_tensor_extra_gpu * g_temp_tensor_extras = nullptr;
-static size_t g_temp_tensor_extra_index = 0;
-
-static ggml_tensor_extra_gpu * ggml_cuda_alloc_temp_tensor_extra() {
-    if (g_temp_tensor_extras == nullptr) {
-        g_temp_tensor_extras = new ggml_tensor_extra_gpu[GGML_CUDA_MAX_NODES];
-    }
-
-    size_t alloc_index = g_temp_tensor_extra_index;
-    g_temp_tensor_extra_index = (g_temp_tensor_extra_index + 1) % GGML_CUDA_MAX_NODES;
-    ggml_tensor_extra_gpu * extra = &g_temp_tensor_extras[alloc_index];
-    memset(extra, 0, sizeof(*extra));
-
-    return extra;
-}
-
-static void ggml_cuda_assign_buffers_impl(struct ggml_tensor * tensor, bool scratch, bool force_inplace, bool no_alloc) {
-    if (scratch && g_scratch_size == 0) {
-        return;
-    }
-
-    tensor->backend = GGML_BACKEND_GPU;
-
-    // recursively assign CUDA buffers until a compute tensor is found
-    if (tensor->src[0] != nullptr && tensor->src[0]->backend == GGML_BACKEND_CPU) {
-        const ggml_op src0_op = tensor->src[0]->op;
-        if (src0_op == GGML_OP_RESHAPE || src0_op == GGML_OP_TRANSPOSE || src0_op == GGML_OP_VIEW || src0_op == GGML_OP_PERMUTE) {
-            ggml_cuda_assign_buffers_impl(tensor->src[0], scratch, force_inplace, no_alloc);
-        }
-    }
-    if (tensor->op == GGML_OP_CPY && tensor->src[1]->backend == GGML_BACKEND_CPU) {
-        ggml_cuda_assign_buffers_impl(tensor->src[1], scratch, force_inplace, no_alloc);
-    }
-
-    if (scratch && no_alloc) {
-        return;
-    }
-
-    ggml_tensor_extra_gpu * extra;
-
-    const bool inplace = (tensor->src[0] != nullptr && tensor->src[0]->data == tensor->data) ||
-        tensor->op == GGML_OP_VIEW ||
-        force_inplace;
-    const size_t size = ggml_nbytes(tensor);
-
-    ggml_cuda_set_device(g_main_device);
-    if (inplace && (tensor->src[0]->backend == GGML_BACKEND_GPU || tensor->src[0]->backend == GGML_BACKEND_GPU_SPLIT)) {
-        ggml_tensor_extra_gpu * src0_extra = (ggml_tensor_extra_gpu * ) tensor->src[0]->extra;
-        char * src0_ddc = (char *) src0_extra->data_device[g_main_device];
-        size_t offset = 0;
-        if (tensor->op == GGML_OP_VIEW) {
-            memcpy(&offset, tensor->op_params, sizeof(size_t));
-        }
-        extra = ggml_cuda_alloc_temp_tensor_extra();
-        extra->data_device[g_main_device] = src0_ddc + offset;
-    } else if (tensor->op == GGML_OP_CPY) {
-        ggml_tensor_extra_gpu * src1_extra = (ggml_tensor_extra_gpu * ) tensor->src[1]->extra;
-        void * src1_ddv = src1_extra->data_device[g_main_device];
-        extra = ggml_cuda_alloc_temp_tensor_extra();
-        extra->data_device[g_main_device] = src1_ddv;
-    } else if (scratch) {
-        GGML_ASSERT(size <= g_scratch_size);
-        if (g_scratch_offset + size > g_scratch_size) {
-            g_scratch_offset = 0;
-        }
-
-        char * data = (char *) g_scratch_buffer;
-        if (data == nullptr) {
-            CUDA_CHECK(cudaMalloc(&data, g_scratch_size));
-            g_scratch_buffer = data;
-        }
-        extra = ggml_cuda_alloc_temp_tensor_extra();
-        extra->data_device[g_main_device] = data + g_scratch_offset;
-
-        g_scratch_offset += size;
-
-        GGML_ASSERT(g_scratch_offset <= g_scratch_size);
-    } else { // allocate new buffers outside of scratch
-        void * data;
-        CUDA_CHECK(cudaMalloc(&data, size));
-        CUDA_CHECK(cudaMemset(data, 0, size));
-        extra = new ggml_tensor_extra_gpu;
-        memset(extra, 0, sizeof(*extra));
-        extra->data_device[g_main_device] = data;
-    }
-
-    tensor->extra = extra;
-}
-
-void ggml_cuda_assign_scratch_offset(struct ggml_tensor * tensor, size_t offset) {
-    if (g_scratch_size == 0) {
-        return;
-    }
-    if (g_scratch_buffer == nullptr) {
-        ggml_cuda_set_device(g_main_device);
-        CUDA_CHECK(cudaMalloc(&g_scratch_buffer, g_scratch_size));
-    }
-
-    ggml_tensor_extra_gpu * extra = ggml_cuda_alloc_temp_tensor_extra();
-
-    const bool inplace = tensor->view_src != nullptr;
-
-    if (inplace && (tensor->view_src->backend == GGML_BACKEND_GPU || tensor->view_src->backend == GGML_BACKEND_GPU_SPLIT)) {
-        ggml_tensor_extra_gpu * src0_extra = (ggml_tensor_extra_gpu * ) tensor->view_src->extra;
-        char * src0_ddc = (char *) src0_extra->data_device[g_main_device];
-        size_t view_offset = 0;
-        if (tensor->op == GGML_OP_VIEW) {
-            memcpy(&view_offset, tensor->op_params, sizeof(size_t));
-        }
-        extra->data_device[g_main_device] = src0_ddc + view_offset;
-    } else {
-        extra->data_device[g_main_device] = (char *) g_scratch_buffer + offset;
-    }
-
-    tensor->extra = extra;
-}
-
-void ggml_cuda_copy_to_device(struct ggml_tensor * tensor) {
-    GGML_ASSERT(tensor->backend == GGML_BACKEND_GPU);
-    GGML_ASSERT(ggml_is_contiguous(tensor));
-
-    ggml_tensor_extra_gpu * extra = (ggml_tensor_extra_gpu *) tensor->extra;
-    ggml_cuda_set_device(g_main_device);
-    CUDA_CHECK(cudaMemcpy(extra->data_device[g_main_device], tensor->data, ggml_nbytes(tensor), cudaMemcpyHostToDevice));
-}
-
-void ggml_cuda_assign_buffers(struct ggml_tensor * tensor) {
-    ggml_cuda_assign_buffers_impl(tensor, true, false, false);
-}
-
-void ggml_cuda_assign_buffers_no_alloc(struct ggml_tensor * tensor) {
-    ggml_cuda_assign_buffers_impl(tensor, true, false, true);
-}
-
-void ggml_cuda_assign_buffers_no_scratch(struct ggml_tensor * tensor) {
-    ggml_cuda_assign_buffers_impl(tensor, false, false, false);
-}
-
-void ggml_cuda_assign_buffers_force_inplace(struct ggml_tensor * tensor) {
-    ggml_cuda_assign_buffers_impl(tensor, false, true, false);
-}
-
-void ggml_cuda_set_main_device(const int main_device) {
-    if (main_device >= g_device_count) {
-        fprintf(stderr, "warning: cannot set main_device=%d because there are only %d devices. Using device %d instead.\n",
-                main_device, g_device_count, g_main_device);
-        return;
-    }
-
-    if (g_main_device != main_device && g_device_count > 1) {
-        g_main_device = main_device;
-        cudaDeviceProp prop;
-        CUDA_CHECK(cudaGetDeviceProperties(&prop, g_main_device));
-        fprintf(stderr, "%s: using device %d (%s) as main device\n", __func__, g_main_device, prop.name);
-    }
-}
-
-void ggml_cuda_set_scratch_size(const size_t scratch_size) {
-    // this is a hack to not completely break llama.cpp when using multiple models or contexts simultaneously
-    // it still won't always work as expected, but it's better than nothing
-    if (scratch_size > g_scratch_size) {
-        ggml_cuda_free_scratch();
-    }
-    g_scratch_size = std::max(g_scratch_size, scratch_size);
-}
-
-void ggml_cuda_free_scratch() {
-    if (g_scratch_buffer == nullptr) {
-        return;
-    }
-
-    CUDA_CHECK(cudaFree(g_scratch_buffer));
-    g_scratch_buffer = nullptr;
-}
-
-bool ggml_cuda_compute_forward(struct ggml_compute_params * params, struct ggml_tensor * tensor) {
-    if (!g_cublas_loaded) return false;
-
-    ggml_cuda_func_t func;
-    const bool any_on_device = tensor->backend == GGML_BACKEND_GPU
-        || (tensor->src[0] != nullptr && (tensor->src[0]->backend == GGML_BACKEND_GPU || tensor->src[0]->backend == GGML_BACKEND_GPU_SPLIT))
-        || (tensor->src[1] != nullptr && tensor->src[1]->backend == GGML_BACKEND_GPU);
-
-    if (!any_on_device && tensor->op != GGML_OP_MUL_MAT && tensor->op != GGML_OP_MUL_MAT_ID) {
-        return false;
-    }
-
-    if (tensor->op == GGML_OP_MUL_MAT) {
-        if (tensor->src[0]->ne[3] != tensor->src[1]->ne[3]) {
-#ifndef NDEBUG
-            fprintf(stderr, "%s: cannot compute %s: src0->ne[3] = %" PRId64 ", src1->ne[3] = %" PRId64 " - fallback to CPU\n", __func__, tensor->name, tensor->src[0]->ne[3], tensor->src[1]->ne[3]);
-#endif
-            return false;
-        }
-    }
-
-    switch (tensor->op) {
-        case GGML_OP_REPEAT:
-            func = ggml_cuda_repeat;
-            break;
-        case GGML_OP_GET_ROWS:
-            func = ggml_cuda_get_rows;
-            break;
-        case GGML_OP_DUP:
-            func = ggml_cuda_dup;
-            break;
-        case GGML_OP_ADD:
-            func = ggml_cuda_add;
-            break;
-        case GGML_OP_ACC:
-            func = ggml_cuda_acc;
-            break;
-        case GGML_OP_MUL:
-            func = ggml_cuda_mul;
-            break;
-        case GGML_OP_DIV:
-            func = ggml_cuda_div;
-            break;
-        case GGML_OP_UNARY:
-            switch (ggml_get_unary_op(tensor)) {
-                case GGML_UNARY_OP_GELU:
-                    func = ggml_cuda_gelu;
-                    break;
-                case GGML_UNARY_OP_SILU:
-                    func = ggml_cuda_silu;
-                    break;
-                case GGML_UNARY_OP_GELU_QUICK:
-                    func = ggml_cuda_gelu_quick;
-                    break;
-                case GGML_UNARY_OP_TANH:
-                    func = ggml_cuda_tanh;
-                    break;
-                case GGML_UNARY_OP_RELU:
-                    func = ggml_cuda_relu;
-                    break;
-                default:
-                    return false;
-            }
-            break;
-        case GGML_OP_NORM:
-            func = ggml_cuda_norm;
-            break;
-        case GGML_OP_GROUP_NORM:
-            func = ggml_cuda_group_norm;
-            break;
-        case GGML_OP_CONCAT:
-            func = ggml_cuda_concat;
-            break;
-        case GGML_OP_UPSCALE:
-            func = ggml_cuda_upscale;
-            break;
-        case GGML_OP_PAD:
-            func = ggml_cuda_pad;
-            break;
-        case GGML_OP_LEAKY_RELU:
-            func = ggml_cuda_leaky_relu;
-            break;
-        case GGML_OP_RMS_NORM:
-            func = ggml_cuda_rms_norm;
-            break;
-        case GGML_OP_MUL_MAT:
-            if (!any_on_device && !ggml_cuda_can_mul_mat(tensor->src[0], tensor->src[1], tensor)) {
-                return false;
-            }
-            func = ggml_cuda_mul_mat;
-            break;
-        case GGML_OP_MUL_MAT_ID:
-            if (!any_on_device && !ggml_cuda_can_mul_mat(tensor->src[2], tensor->src[1], tensor)) {
-                return false;
-            }
-            func = ggml_cuda_mul_mat_id;
-            break;
-        case GGML_OP_SCALE:
-            func = ggml_cuda_scale;
-            break;
-        case GGML_OP_SQR:
-            func = ggml_cuda_sqr;
-            break;
-        case GGML_OP_CLAMP:
-            func = ggml_cuda_clamp;
-            break;
-        case GGML_OP_CPY:
-            func = ggml_cuda_cpy;
-            break;
-        case GGML_OP_CONT:
-            func = ggml_cuda_dup;
-            break;
-        case GGML_OP_NONE:
-        case GGML_OP_RESHAPE:
-        case GGML_OP_VIEW:
-        case GGML_OP_PERMUTE:
-        case GGML_OP_TRANSPOSE:
-            func = ggml_cuda_nop;
-            break;
-        case GGML_OP_DIAG_MASK_INF:
-            func = ggml_cuda_diag_mask_inf;
-            break;
-        case GGML_OP_SOFT_MAX:
-            func = ggml_cuda_soft_max;
-            break;
-        case GGML_OP_ROPE:
-            func = ggml_cuda_rope;
-            break;
-        case GGML_OP_ALIBI:
-            func = ggml_cuda_alibi;
-            break;
-        case GGML_OP_IM2COL:
-            func = ggml_cuda_im2col;
-            break;
-        case GGML_OP_SUM_ROWS:
-            func = ggml_cuda_sum_rows;
-            break;
-        case GGML_OP_ARGSORT:
-            func = ggml_cuda_argsort;
-            break;
-        default:
-            return false;
-    }
-
-    if (tensor->src[0] != nullptr && tensor->src[0]->backend == GGML_BACKEND_GPU_SPLIT) {
-        ggml_cuda_set_peer_access(tensor->src[1]->ne[1]);
-    }
-
-    if (params->ith != 0) {
-        return true;
-    }
-    if (params->type == GGML_TASK_INIT || params->type == GGML_TASK_FINALIZE) {
-        return true;
-    }
-    func(tensor->src[0], tensor->src[1], tensor);
-    return true;
-}
-
-int ggml_cuda_get_device_count() {
-    int device_count;
-    if (cudaGetDeviceCount(&device_count) != cudaSuccess) {
-        return 0;
-    }
-    return device_count;
-}
-
-void ggml_cuda_get_device_description(int device, char * description, size_t description_size) {
-    cudaDeviceProp prop;
-    CUDA_CHECK(cudaGetDeviceProperties(&prop, device));
-    snprintf(description, description_size, "%s", prop.name);
-}
-
-////////////////////////////////////////////////////////////////////////////////
-
-// backend interface
-
-#define UNUSED GGML_UNUSED
-
-// cuda buffer
-
-struct ggml_backend_buffer_context_cuda {
-    int device;
-    void * dev_ptr = nullptr;
-    ggml_tensor_extra_gpu * temp_tensor_extras = nullptr;
-    size_t temp_tensor_extra_index = 0;
-
-    ggml_backend_buffer_context_cuda(int device, void * dev_ptr) : device(device), dev_ptr(dev_ptr) {}
-
-    ~ggml_backend_buffer_context_cuda() {
-        delete[] temp_tensor_extras;
-    }
-
-    ggml_tensor_extra_gpu * ggml_cuda_alloc_temp_tensor_extra() {
-        if (temp_tensor_extras == nullptr) {
-            temp_tensor_extras = new ggml_tensor_extra_gpu[GGML_CUDA_MAX_NODES];
-        }
-
-        size_t alloc_index = temp_tensor_extra_index;
-        temp_tensor_extra_index = (temp_tensor_extra_index + 1) % GGML_CUDA_MAX_NODES;
-        ggml_tensor_extra_gpu * extra = &temp_tensor_extras[alloc_index];
-        memset(extra, 0, sizeof(*extra));
-
-        return extra;
-    }
-};
-
-static void ggml_backend_cuda_buffer_free_buffer(ggml_backend_buffer_t buffer) {
-    ggml_backend_buffer_context_cuda * ctx = (ggml_backend_buffer_context_cuda *)buffer->context;
-    CUDA_CHECK(cudaFree(ctx->dev_ptr));
-    delete ctx;
-}
-
-static void * ggml_backend_cuda_buffer_get_base(ggml_backend_buffer_t buffer) {
-    ggml_backend_buffer_context_cuda * ctx = (ggml_backend_buffer_context_cuda *)buffer->context;
-    return ctx->dev_ptr;
-}
-
-static void ggml_backend_cuda_buffer_init_tensor(ggml_backend_buffer_t buffer, ggml_tensor * tensor) {
-    ggml_backend_buffer_context_cuda * ctx = (ggml_backend_buffer_context_cuda *)buffer->context;
-
-    if (tensor->view_src != NULL && tensor->view_offs == 0) {
-        assert(tensor->view_src->buffer->buft == buffer->buft);
-        tensor->backend = tensor->view_src->backend;
-        tensor->extra = tensor->view_src->extra;
-        return;
-    }
-
-    ggml_tensor_extra_gpu * extra = ctx->ggml_cuda_alloc_temp_tensor_extra();
-
-    extra->data_device[ctx->device] = tensor->data;
-
-    tensor->backend = GGML_BACKEND_GPU;
-    tensor->extra = extra;
-
-    if (ggml_is_quantized(tensor->type)) {
-        // initialize padding to 0 to avoid possible NaN values
-        int64_t row_low = 0;
-        int64_t row_high = ggml_nrows(tensor);
-        int64_t nrows_split = row_high - row_low;
-
-        size_t original_size = ggml_nbytes_split(tensor, nrows_split);
-        size_t padded_size = ggml_backend_buft_get_alloc_size(buffer->buft, tensor);
-
-        if (padded_size > original_size && tensor->view_src == nullptr) {
-            CUDA_CHECK(cudaMemsetAsync((char *)tensor->data + original_size, 0, padded_size - original_size, g_cudaStreams[ctx->device][0]));
-        }
-    }
-
-    UNUSED(buffer);
-}
-
-static void ggml_backend_cuda_buffer_set_tensor(ggml_backend_buffer_t buffer, ggml_tensor * tensor, const void * data, size_t offset, size_t size) {
-    GGML_ASSERT(tensor->backend == GGML_BACKEND_GPU);
-
-    ggml_backend_buffer_context_cuda * ctx = (ggml_backend_buffer_context_cuda *)buffer->context;
-
-    ggml_cuda_set_device(ctx->device);
-    CUDA_CHECK(cudaDeviceSynchronize());
-    CUDA_CHECK(cudaMemcpy((char *)tensor->data + offset, data, size, cudaMemcpyHostToDevice));
-    CUDA_CHECK(cudaDeviceSynchronize());
-}
-
-static void ggml_backend_cuda_buffer_get_tensor(ggml_backend_buffer_t buffer, const ggml_tensor * tensor, void * data, size_t offset, size_t size) {
-    GGML_ASSERT(tensor->backend == GGML_BACKEND_GPU);
-
-    ggml_backend_buffer_context_cuda * ctx = (ggml_backend_buffer_context_cuda *)buffer->context;
-
-    ggml_cuda_set_device(ctx->device);
-    CUDA_CHECK(cudaDeviceSynchronize());
-
-    CUDA_CHECK(cudaMemcpy(data, (const char *)tensor->data + offset, size, cudaMemcpyDeviceToHost));
-}
-
-static void ggml_backend_cuda_buffer_clear(ggml_backend_buffer_t buffer, uint8_t value) {
-    ggml_backend_buffer_context_cuda * ctx = (ggml_backend_buffer_context_cuda *)buffer->context;
-
-    ggml_cuda_set_device(ctx->device);
-    CUDA_CHECK(cudaDeviceSynchronize());
-
-    CUDA_CHECK(cudaMemset(ctx->dev_ptr, value, buffer->size));
-}
-
-static struct ggml_backend_buffer_i cuda_backend_buffer_interface = {
-    /* .free_buffer     = */ ggml_backend_cuda_buffer_free_buffer,
-    /* .get_base        = */ ggml_backend_cuda_buffer_get_base,
-    /* .init_tensor     = */ ggml_backend_cuda_buffer_init_tensor,
-    /* .set_tensor      = */ ggml_backend_cuda_buffer_set_tensor,
-    /* .get_tensor      = */ ggml_backend_cuda_buffer_get_tensor,
-    /* .cpy_tensor_from = */ NULL,
-    /* .cpy_tensor_to   = */ NULL,
-    /* .clear           = */ ggml_backend_cuda_buffer_clear,
-};
-
-// cuda buffer type
-
-static ggml_backend_buffer_t ggml_backend_cuda_buffer_type_alloc_buffer(ggml_backend_buffer_type_t buft, size_t size) {
-    int device = (int) (intptr_t) buft->context;
-
-    ggml_cuda_set_device(device);
-
-    size = std::max(size, (size_t)1); // cudaMalloc returns null for size 0
-
-    void * dev_ptr;
-    CUDA_CHECK(cudaMalloc(&dev_ptr, size));
-
-    ggml_backend_buffer_context_cuda * ctx = new ggml_backend_buffer_context_cuda(device, dev_ptr);
-
-    return ggml_backend_buffer_init(buft, cuda_backend_buffer_interface, ctx, size);
-}
-
-static size_t ggml_backend_cuda_buffer_type_get_alignment(ggml_backend_buffer_type_t buft) {
-    return 128;
-
-    UNUSED(buft);
-}
-
-static size_t ggml_backend_cuda_buffer_type_get_alloc_size(ggml_backend_buffer_type_t buft, ggml_tensor * tensor) {
-    int64_t row_low = 0;
-    int64_t row_high = ggml_nrows(tensor);
-    int64_t nrows_split = row_high - row_low;
-
-    size_t size = ggml_nbytes_split(tensor, nrows_split);
-
-    int64_t ne0 = tensor->ne[0];
-
-    if (ggml_is_quantized(tensor->type)) {
-        if (ne0 % MATRIX_ROW_PADDING != 0) {
-            size += ggml_row_size(tensor->type, MATRIX_ROW_PADDING - ne0 % MATRIX_ROW_PADDING);
-        }
-    }
-
-    return size;
-
-    UNUSED(buft);
-}
-
-static bool ggml_backend_cuda_buffer_type_supports_backend(ggml_backend_buffer_type_t buft, ggml_backend_t backend) {
-    return ggml_backend_is_cuda(backend);
-
-    UNUSED(buft);
-}
-
-static ggml_backend_buffer_type_i ggml_backend_cuda_buffer_type_interface = {
-    /* .alloc_buffer     = */ ggml_backend_cuda_buffer_type_alloc_buffer,
-    /* .get_alignment    = */ ggml_backend_cuda_buffer_type_get_alignment,
-    /* .get_alloc_size   = */ ggml_backend_cuda_buffer_type_get_alloc_size,
-    /* .supports_backend = */ ggml_backend_cuda_buffer_type_supports_backend,
-    /* .is_host          = */ nullptr,
-};
-
-ggml_backend_buffer_type_t ggml_backend_cuda_buffer_type(int device) {
-    static struct ggml_backend_buffer_type ggml_backend_cuda_buffer_types[GGML_CUDA_MAX_DEVICES];
-
-    static bool ggml_backend_cuda_buffer_type_initialized = false;
-
-    if (!ggml_backend_cuda_buffer_type_initialized) {
-        for (int i = 0; i < GGML_CUDA_MAX_DEVICES; i++) {
-            ggml_backend_cuda_buffer_types[i] = {
-                /* .iface    = */ ggml_backend_cuda_buffer_type_interface,
-                /* .context  = */ (ggml_backend_buffer_type_context_t) (intptr_t) i,
-            };
-        }
-        ggml_backend_cuda_buffer_type_initialized = true;
-    }
-
-    return &ggml_backend_cuda_buffer_types[device];
-}
-
-// host buffer type
-
-static void ggml_backend_cuda_host_buffer_free_buffer(ggml_backend_buffer_t buffer) {
-    ggml_cuda_host_free(buffer->context);
-}
-
-static ggml_backend_buffer_t ggml_backend_cuda_host_buffer_type_alloc_buffer(ggml_backend_buffer_type_t buft, size_t size) {
-    void * ptr = ggml_cuda_host_malloc(size);
-
-    if (ptr == nullptr) {
-        // fallback to cpu buffer
-        return ggml_backend_buft_alloc_buffer(ggml_backend_cpu_buffer_type(), size);
-    }
-
-    // FIXME: this is a hack to avoid having to implement a new buffer type
-    ggml_backend_buffer_t buffer = ggml_backend_cpu_buffer_from_ptr(ptr, size);
-    buffer->buft = buft;
-    buffer->iface.free_buffer = ggml_backend_cuda_host_buffer_free_buffer;
-
-    return buffer;
-}
-
-ggml_backend_buffer_type_t ggml_backend_cuda_host_buffer_type() {
-    static struct ggml_backend_buffer_type ggml_backend_cuda_buffer_type_host = {
-        /* .iface    = */ {
-            /* .alloc_buffer     = */ ggml_backend_cuda_host_buffer_type_alloc_buffer,
-            /* .get_alignment    = */ ggml_backend_cpu_buffer_type()->iface.get_alignment,
-            /* .get_alloc_size   = */ ggml_backend_cpu_buffer_type()->iface.get_alloc_size,
-            /* .supports_backend = */ ggml_backend_cpu_buffer_type()->iface.supports_backend,
-            /* .is_host          = */ ggml_backend_cpu_buffer_type()->iface.is_host,
-        },
-        /* .context  = */ nullptr,
-    };
-
-    return &ggml_backend_cuda_buffer_type_host;
-}
-
-// backend
-
-struct ggml_backend_context_cuda {
-    int device;
-};
-
-static const char * ggml_backend_cuda_name(ggml_backend_t backend) {
-    return GGML_CUDA_NAME;
-
-    UNUSED(backend);
-}
-
-static void ggml_backend_cuda_free(ggml_backend_t backend) {
-    ggml_backend_context_cuda * cuda_ctx = (ggml_backend_context_cuda *)backend->context;
-
-    delete cuda_ctx;
-    delete backend;
-}
-
-static ggml_backend_buffer_type_t ggml_backend_cuda_get_default_buffer_type(ggml_backend_t backend) {
-    ggml_backend_context_cuda * cuda_ctx = (ggml_backend_context_cuda *)backend->context;
-
-    return ggml_backend_cuda_buffer_type(cuda_ctx->device);
-}
-
-static void ggml_backend_cuda_set_tensor_async(ggml_backend_t backend, ggml_tensor * tensor, const void * data, size_t offset, size_t size) {
-    ggml_backend_context_cuda * cuda_ctx = (ggml_backend_context_cuda *)backend->context;
-
-    GGML_ASSERT(tensor->buffer->buft == ggml_backend_cuda_buffer_type(cuda_ctx->device) && "unsupported buffer type");
-    GGML_ASSERT(tensor->backend == GGML_BACKEND_GPU);
-
-    CUDA_CHECK(cudaMemcpyAsync((char *)tensor->data + offset, data, size, cudaMemcpyHostToDevice, g_cudaStreams[cuda_ctx->device][0]));
-}
-
-static void ggml_backend_cuda_get_tensor_async(ggml_backend_t backend, const ggml_tensor * tensor, void * data, size_t offset, size_t size) {
-    ggml_backend_context_cuda * cuda_ctx = (ggml_backend_context_cuda *)backend->context;
-
-    GGML_ASSERT(tensor->buffer->buft == ggml_backend_cuda_buffer_type(cuda_ctx->device) && "unsupported buffer type");
-    GGML_ASSERT(tensor->backend == GGML_BACKEND_GPU);
-
-    CUDA_CHECK(cudaMemcpyAsync(data, (const char *)tensor->data + offset, size, cudaMemcpyDeviceToHost, g_cudaStreams[cuda_ctx->device][0]));
-}
-
-static void ggml_backend_cuda_synchronize(ggml_backend_t backend) {
-    ggml_backend_context_cuda * cuda_ctx = (ggml_backend_context_cuda *)backend->context;
-
-    CUDA_CHECK(cudaStreamSynchronize(g_cudaStreams[cuda_ctx->device][0]));
-
-    UNUSED(backend);
-}
-
-static ggml_backend_graph_plan_t ggml_backend_cuda_graph_plan_create(ggml_backend_t backend, ggml_cgraph * cgraph) {
-    GGML_ASSERT(!"not implemented");
-
-    return nullptr;
-
-    UNUSED(backend);
-    UNUSED(cgraph);
-}
-
-static void ggml_backend_cuda_graph_plan_free(ggml_backend_t backend, ggml_backend_graph_plan_t plan) {
-    GGML_ASSERT(!"not implemented");
-
-    UNUSED(backend);
-    UNUSED(plan);
-}
-
-static void ggml_backend_cuda_graph_plan_compute(ggml_backend_t backend, ggml_backend_graph_plan_t plan) {
-    GGML_ASSERT(!"not implemented");
-
-    UNUSED(backend);
-    UNUSED(plan);
-}
-
-static bool ggml_backend_cuda_graph_compute(ggml_backend_t backend, ggml_cgraph * cgraph) {
-    ggml_backend_context_cuda * cuda_ctx = (ggml_backend_context_cuda *)backend->context;
-
-    ggml_cuda_set_main_device(cuda_ctx->device);
-
-    ggml_compute_params params = {};
-    params.type = GGML_TASK_COMPUTE;
-    params.ith = 0;
-    for (int i = 0; i < cgraph->n_nodes; i++) {
-        ggml_tensor * node = cgraph->nodes[i];
-
-        if (node->op == GGML_OP_RESHAPE || node->op == GGML_OP_TRANSPOSE || node->op == GGML_OP_VIEW || node->op == GGML_OP_PERMUTE)
-            continue;
-
-        assert(node->backend == GGML_BACKEND_GPU);
-        assert(node->buffer->buft == ggml_backend_cuda_buffer_type(cuda_ctx->device));
-        assert(node->extra != nullptr);
-
-        for (int j = 0; j < GGML_MAX_SRC; j++) {
-            if (node->src[j] != nullptr) {
-                assert(node->src[j]->backend == GGML_BACKEND_GPU);
-                assert(node->src[j]->buffer->buft == ggml_backend_cuda_buffer_type(cuda_ctx->device));
-                assert(node->src[j]->extra != nullptr);
-            }
-        }
-
-        bool ok = ggml_cuda_compute_forward(&params, node);
-        if (!ok) {
-            fprintf(stderr, "%s: error: op not supported %s (%s)\n", __func__, node->name, ggml_op_name(node->op));
-        }
-        GGML_ASSERT(ok);
-
-#if 0
-        if (node->type == GGML_TYPE_F32) {
-            cudaDeviceSynchronize();
-            std::vector<float> tmp(ggml_nelements(node), 0.0f);
-            cudaMemcpy(tmp.data(), node->data, ggml_nelements(node)*sizeof(float), cudaMemcpyDeviceToHost);
-            printf("\n%s (%s) (%s %s) (%s %s): ", node->name, ggml_op_name(node->op),
-                ggml_type_name(node->src[0]->type),
-                node->src[1] ? ggml_type_name(node->src[1]->type) : "none",
-                node->src[0]->name,
-                node->src[1] ? node->src[1]->name : "none");
-            double sum = 0.0;
-            double sq_sum = 0.0;
-            for (int i = 0; i < ggml_nelements(node); i++) {
-                printf("%f ", tmp[i]);
-                sum += tmp[i];
-                sq_sum += tmp[i]*tmp[i];
-            }
-            printf("\n");
-            printf("sum: %f, ", sum);
-            printf("sq_sum: %f\n", sq_sum);
-        }
-#endif
-    }
-
-    UNUSED(backend);
-
-    return true;
-}
-
-static bool ggml_backend_cuda_supports_op(ggml_backend_t backend, const ggml_tensor * op) {
-    switch (op->op) {
-        case GGML_OP_UNARY:
-            switch (ggml_get_unary_op(op)) {
-                case GGML_UNARY_OP_GELU:
-                case GGML_UNARY_OP_SILU:
-                case GGML_UNARY_OP_RELU:
-                case GGML_UNARY_OP_GELU_QUICK:
-                case GGML_UNARY_OP_TANH:
-                    return true;
-                default:
-                    return false;
-            }
-            break;
-        case GGML_OP_MUL_MAT:
-        case GGML_OP_MUL_MAT_ID:
-            {
-                struct ggml_tensor * a;
-                struct ggml_tensor * b;
-                if (op->op == GGML_OP_MUL_MAT) {
-                    a = op->src[0];
-                    b = op->src[1];
-                } else {
-                    a = op->src[2];
-                    b = op->src[1];
-                }
-                if (a->ne[3] != b->ne[3]) {
-                    return false;
-                }
-                return true;
-            } break;
-        case GGML_OP_GET_ROWS:
-            {
-                switch (op->src[0]->type) {
-                    case GGML_TYPE_F16:
-                    case GGML_TYPE_F32:
-                    case GGML_TYPE_Q4_0:
-                    case GGML_TYPE_Q4_1:
-                    case GGML_TYPE_Q5_0:
-                    case GGML_TYPE_Q5_1:
-                    case GGML_TYPE_Q8_0:
-                        return true;
-                    default:
-                        return false;
-                }
-            } break;
-        case GGML_OP_CPY:
-            {
-                ggml_type src0_type = op->src[0]->type;
-                ggml_type src1_type = op->src[1]->type;
-                if (src0_type == GGML_TYPE_F32 && src1_type == GGML_TYPE_F32) {
-                    return true;
-                }
-                if (src0_type == GGML_TYPE_F32 && src1_type == GGML_TYPE_F16) {
-                    return true;
-                }
-                if (src0_type == GGML_TYPE_F32 && src1_type == GGML_TYPE_Q8_0) {
-                    return true;
-                }
-                if (src0_type == GGML_TYPE_F32 && src1_type == GGML_TYPE_Q4_0) {
-                    return true;
-                }
-                if (src0_type == GGML_TYPE_F32 && src1_type == GGML_TYPE_Q4_1) {
-                    return true;
-                }
-                if (src0_type == GGML_TYPE_F16 && src1_type == GGML_TYPE_F16) {
-                    return true;
-                }
-                return false;
-            } break;
-        case GGML_OP_DUP:
-        case GGML_OP_REPEAT:
-        case GGML_OP_CONCAT:
-            {
-                ggml_type src0_type = op->src[0]->type;
-                return src0_type != GGML_TYPE_I32 && src0_type != GGML_TYPE_I16;
-            } break;
-        case GGML_OP_NONE:
-        case GGML_OP_RESHAPE:
-        case GGML_OP_VIEW:
-        case GGML_OP_PERMUTE:
-        case GGML_OP_TRANSPOSE:
-        case GGML_OP_NORM:
-        case GGML_OP_ADD:
-        case GGML_OP_MUL:
-        case GGML_OP_DIV:
-        case GGML_OP_RMS_NORM:
-        case GGML_OP_SCALE:
-        case GGML_OP_SQR:
-        case GGML_OP_CLAMP:
-        case GGML_OP_CONT:
-        case GGML_OP_DIAG_MASK_INF:
-        case GGML_OP_SOFT_MAX:
-        case GGML_OP_ROPE:
-        case GGML_OP_ALIBI:
-        case GGML_OP_IM2COL:
-        case GGML_OP_SUM_ROWS:
-        case GGML_OP_ARGSORT:
-        case GGML_OP_ACC:
-        case GGML_OP_GROUP_NORM:
-        case GGML_OP_UPSCALE:
-        case GGML_OP_PAD:
-        case GGML_OP_LEAKY_RELU:
-            return true;
-        default:
-            return false;
-    }
-
-    UNUSED(backend);
-}
-
-static ggml_backend_i cuda_backend_i = {
-    /* .get_name                = */ ggml_backend_cuda_name,
-    /* .free                    = */ ggml_backend_cuda_free,
-    /* .get_default_buffer_type = */ ggml_backend_cuda_get_default_buffer_type,
-    /* .set_tensor_async        = */ ggml_backend_cuda_set_tensor_async,
-    /* .get_tensor_async        = */ ggml_backend_cuda_get_tensor_async,
-    /* .cpy_tensor_from_async   = */ NULL,
-    /* .cpy_tensor_to_async     = */ NULL,
-    /* .synchronize             = */ ggml_backend_cuda_synchronize,
-    /* .graph_plan_create       = */ ggml_backend_cuda_graph_plan_create,
-    /* .graph_plan_free         = */ ggml_backend_cuda_graph_plan_free,
-    /* .graph_plan_compute      = */ ggml_backend_cuda_graph_plan_compute,
-    /* .graph_compute           = */ ggml_backend_cuda_graph_compute,
-    /* .supports_op             = */ ggml_backend_cuda_supports_op,
-};
-
-ggml_backend_t ggml_backend_cuda_init(int device) {
-    ggml_init_cublas(); // TODO: remove from ggml.c
-
-    if (device < 0 || device >= ggml_cuda_get_device_count()) {
-        fprintf(stderr, "%s: error: invalid device %d\n", __func__, device);
-        return nullptr;
-    }
-
-    // not strictly necessary, but it may reduce the overhead of the first graph_compute
-    ggml_cuda_set_main_device(device);
-
-    ggml_backend_context_cuda * ctx = new ggml_backend_context_cuda {
-        /* .device = */ device
-    };
-
-    ggml_backend_t cuda_backend = new ggml_backend {
-        /* .interface = */ cuda_backend_i,
-        /* .context   = */ ctx
-    };
-
-    return cuda_backend;
-}
-
-bool ggml_backend_is_cuda(ggml_backend_t backend) {
-    return backend->iface.get_name == ggml_backend_cuda_name;
-}
-
-static ggml_backend_t ggml_backend_reg_cuda_init(const char * params, void * user_data) {
-    ggml_backend_t cuda_backend = ggml_backend_cuda_init((int) (intptr_t) user_data);
-    return cuda_backend;
-
-    UNUSED(params);
-}
-
-extern "C" int ggml_backend_cuda_reg_devices();
-
-int ggml_backend_cuda_reg_devices() {
-    int device_count = ggml_cuda_get_device_count();
-    //int device_count = 1; // DEBUG: some tools require delaying CUDA initialization
-    for (int i = 0; i < device_count; i++) {
-        char name[128];
-        snprintf(name, sizeof(name), "%s%d", GGML_CUDA_NAME, i);
-        ggml_backend_register(name, ggml_backend_reg_cuda_init, ggml_backend_cuda_buffer_type(i), (void *) (intptr_t) i);
-    }
-    return device_count;
-}
diff --git a/src/whisper_cpp/ggml-cuda.h b/src/whisper_cpp/ggml-cuda.h
deleted file mode 100644
index cdb0c0c4..00000000
--- a/src/whisper_cpp/ggml-cuda.h
+++ /dev/null
@@ -1,64 +0,0 @@
-#pragma once
-
-#include "ggml.h"
-#include "ggml-backend.h"
-
-#ifdef GGML_USE_HIPBLAS
-#define GGML_CUDA_NAME "ROCm"
-#define GGML_CUBLAS_NAME "hipBLAS"
-#else
-#define GGML_CUDA_NAME "CUDA"
-#define GGML_CUBLAS_NAME "cuBLAS"
-#endif
-
-#ifdef  __cplusplus
-extern "C" {
-#endif
-
-#define GGML_CUDA_MAX_DEVICES       16
-
-// Always success. To check if CUDA is actually loaded, use `ggml_cublas_loaded`.
-GGML_API void   ggml_init_cublas(void);
-
-// Returns `true` if there are available CUDA devices and cublas loads successfully; otherwise, it returns `false`.
-GGML_API bool   ggml_cublas_loaded(void);
-
-GGML_API void * ggml_cuda_host_malloc(size_t size);
-GGML_API void   ggml_cuda_host_free(void * ptr);
-
-GGML_API bool   ggml_cuda_can_mul_mat(const struct ggml_tensor * src0, const struct ggml_tensor * src1, struct ggml_tensor * dst);
-GGML_API void   ggml_cuda_set_tensor_split(const float * tensor_split);
-GGML_API void   ggml_cuda_transform_tensor(void * data, struct ggml_tensor * tensor);
-GGML_API void   ggml_cuda_free_data(struct ggml_tensor * tensor);
-
-GGML_API void   ggml_cuda_assign_buffers(struct ggml_tensor * tensor);
-GGML_API void   ggml_cuda_assign_buffers_no_scratch(struct ggml_tensor * tensor);
-GGML_API void   ggml_cuda_assign_buffers_force_inplace(struct ggml_tensor * tensor);
-
-GGML_API void   ggml_cuda_assign_buffers_no_alloc(struct ggml_tensor * tensor);
-GGML_API void   ggml_cuda_assign_scratch_offset(struct ggml_tensor * tensor, size_t offset);
-GGML_API void   ggml_cuda_copy_to_device(struct ggml_tensor * tensor);
-
-GGML_API void   ggml_cuda_set_main_device(int main_device);
-GGML_API void   ggml_cuda_set_mul_mat_q(bool mul_mat_q);
-GGML_API void   ggml_cuda_set_scratch_size(size_t scratch_size);
-GGML_API void   ggml_cuda_free_scratch(void);
-GGML_API bool   ggml_cuda_compute_forward(struct ggml_compute_params * params, struct ggml_tensor * tensor);
-
-GGML_API int    ggml_cuda_get_device_count(void);
-GGML_API void   ggml_cuda_get_device_description(int device, char * description, size_t description_size);
-
-// backend API
-GGML_API ggml_backend_t ggml_backend_cuda_init(int device);
-
-GGML_API bool ggml_backend_is_cuda(ggml_backend_t backend);
-GGML_API int  ggml_backend_cuda_get_device(ggml_backend_t backend);
-
-GGML_API ggml_backend_buffer_type_t ggml_backend_cuda_buffer_type(int device);
-
-// pinned host buffer for use with CPU backend for faster copies between CPU and GPU
-GGML_API ggml_backend_buffer_type_t ggml_backend_cuda_host_buffer_type(void);
-
-#ifdef  __cplusplus
-}
-#endif
diff --git a/src/whisper_cpp/ggml-impl.h b/src/whisper_cpp/ggml-impl.h
deleted file mode 100644
index 1f5610a8..00000000
--- a/src/whisper_cpp/ggml-impl.h
+++ /dev/null
@@ -1,243 +0,0 @@
-#pragma once
-
-#include "ggml.h"
-
-// GGML internal header
-
-#include <assert.h>
-#include <stddef.h>
-#include <stdbool.h>
-#include <string.h> // memcpy
-#include <math.h>   // fabsf
-
-#ifdef __cplusplus
-extern "C" {
-#endif
-
-// static_assert should be a #define, but if it's not,
-// fall back to the _Static_assert C11 keyword.
-// if C99 - static_assert is noop
-// ref: https://stackoverflow.com/a/53923785/4039976
-#ifndef static_assert
-#if defined(__STDC_VERSION__) && (__STDC_VERSION__ >= 201100L)
-#define static_assert(cond, msg) _Static_assert(cond, msg)
-#else
-#define static_assert(cond, msg) struct global_scope_noop_trick
-#endif
-#endif
-
-// __FMA__ and __F16C__ are not defined in MSVC, however they are implied with AVX2/AVX512
-#if defined(_MSC_VER) && (defined(__AVX2__) || defined(__AVX512F__))
-#ifndef __FMA__
-#define __FMA__
-#endif
-#ifndef __F16C__
-#define __F16C__
-#endif
-#ifndef __SSE3__
-#define __SSE3__
-#endif
-#endif
-
-// 16-bit float
-// on Arm, we use __fp16
-// on x86, we use uint16_t
-#if defined(__ARM_NEON) && !defined(_MSC_VER)
-
-// if YCM cannot find <arm_neon.h>, make a symbolic link to it, for example:
-//
-//   $ ln -sfn /Library/Developer/CommandLineTools/usr/lib/clang/13.1.6/include/arm_neon.h ./src/
-//
-#include <arm_neon.h>
-
-#define GGML_COMPUTE_FP16_TO_FP32(x) ((float) (x))
-#define GGML_COMPUTE_FP32_TO_FP16(x) (x)
-
-#define GGML_FP16_TO_FP32(x) ((float) (x))
-#define GGML_FP32_TO_FP16(x) (x)
-
-#else
-
-#ifdef __wasm_simd128__
-#include <wasm_simd128.h>
-#else
-#ifdef __POWER9_VECTOR__
-#include <altivec.h>
-#undef bool
-#define bool _Bool
-#else
-#if defined(_MSC_VER) || defined(__MINGW32__)
-#include <intrin.h>
-#else
-#if defined(__AVX__) || defined(__AVX2__) || defined(__AVX512F__) || defined(__SSSE3__) || defined(__SSE3__)
-#if !defined(__riscv)
-#include <immintrin.h>
-#endif
-#endif
-#endif
-#endif
-#endif
-
-#ifdef __riscv_v_intrinsic
-#include <riscv_vector.h>
-#endif
-
-#ifdef __F16C__
-
-#ifdef _MSC_VER
-#define GGML_COMPUTE_FP16_TO_FP32(x) _mm_cvtss_f32(_mm_cvtph_ps(_mm_cvtsi32_si128(x)))
-#define GGML_COMPUTE_FP32_TO_FP16(x) _mm_extract_epi16(_mm_cvtps_ph(_mm_set_ss(x), 0), 0)
-#else
-#define GGML_COMPUTE_FP16_TO_FP32(x) _cvtsh_ss(x)
-#define GGML_COMPUTE_FP32_TO_FP16(x) _cvtss_sh(x, 0)
-#endif
-
-#elif defined(__POWER9_VECTOR__)
-
-#define GGML_COMPUTE_FP16_TO_FP32(x) ggml_compute_fp16_to_fp32(x)
-#define GGML_COMPUTE_FP32_TO_FP16(x) ggml_compute_fp32_to_fp16(x)
-/* the inline asm below is about 12% faster than the lookup method */
-#define GGML_FP16_TO_FP32(x) GGML_COMPUTE_FP16_TO_FP32(x)
-#define GGML_FP32_TO_FP16(x) GGML_COMPUTE_FP32_TO_FP16(x)
-
-static inline float ggml_compute_fp16_to_fp32(ggml_fp16_t h) {
-    register float f;
-    register double d;
-    __asm__(
-        "mtfprd %0,%2\n"
-        "xscvhpdp %0,%0\n"
-        "frsp %1,%0\n" :
-        /* temp */ "=d"(d),
-        /* out */  "=f"(f):
-        /* in */   "r"(h));
-    return f;
-}
-
-static inline ggml_fp16_t ggml_compute_fp32_to_fp16(float f) {
-    register double d;
-    register ggml_fp16_t r;
-    __asm__( /* xscvdphp can work on double or single precision */
-        "xscvdphp %0,%2\n"
-        "mffprd %1,%0\n" :
-        /* temp */ "=d"(d),
-        /* out */  "=r"(r):
-        /* in */   "f"(f));
-    return r;
-}
-
-#else
-
-// FP16 <-> FP32
-// ref: https://github.com/Maratyszcza/FP16
-
-static inline float fp32_from_bits(uint32_t w) {
-    union {
-        uint32_t as_bits;
-        float as_value;
-    } fp32;
-    fp32.as_bits = w;
-    return fp32.as_value;
-}
-
-static inline uint32_t fp32_to_bits(float f) {
-    union {
-        float as_value;
-        uint32_t as_bits;
-    } fp32;
-    fp32.as_value = f;
-    return fp32.as_bits;
-}
-
-static inline float ggml_compute_fp16_to_fp32(ggml_fp16_t h) {
-    const uint32_t w = (uint32_t) h << 16;
-    const uint32_t sign = w & UINT32_C(0x80000000);
-    const uint32_t two_w = w + w;
-
-    const uint32_t exp_offset = UINT32_C(0xE0) << 23;
-#if defined(__STDC_VERSION__) && (__STDC_VERSION__ >= 199901L) || defined(__GNUC__) && !defined(__STRICT_ANSI__)
-    const float exp_scale = 0x1.0p-112f;
-#else
-    const float exp_scale = fp32_from_bits(UINT32_C(0x7800000));
-#endif
-    const float normalized_value = fp32_from_bits((two_w >> 4) + exp_offset) * exp_scale;
-
-    const uint32_t magic_mask = UINT32_C(126) << 23;
-    const float magic_bias = 0.5f;
-    const float denormalized_value = fp32_from_bits((two_w >> 17) | magic_mask) - magic_bias;
-
-    const uint32_t denormalized_cutoff = UINT32_C(1) << 27;
-    const uint32_t result = sign |
-        (two_w < denormalized_cutoff ? fp32_to_bits(denormalized_value) : fp32_to_bits(normalized_value));
-    return fp32_from_bits(result);
-}
-
-static inline ggml_fp16_t ggml_compute_fp32_to_fp16(float f) {
-#if defined(__STDC_VERSION__) && (__STDC_VERSION__ >= 199901L) || defined(__GNUC__) && !defined(__STRICT_ANSI__)
-    const float scale_to_inf = 0x1.0p+112f;
-    const float scale_to_zero = 0x1.0p-110f;
-#else
-    const float scale_to_inf = fp32_from_bits(UINT32_C(0x77800000));
-    const float scale_to_zero = fp32_from_bits(UINT32_C(0x08800000));
-#endif
-    float base = (fabsf(f) * scale_to_inf) * scale_to_zero;
-
-    const uint32_t w = fp32_to_bits(f);
-    const uint32_t shl1_w = w + w;
-    const uint32_t sign = w & UINT32_C(0x80000000);
-    uint32_t bias = shl1_w & UINT32_C(0xFF000000);
-    if (bias < UINT32_C(0x71000000)) {
-        bias = UINT32_C(0x71000000);
-    }
-
-    base = fp32_from_bits((bias >> 1) + UINT32_C(0x07800000)) + base;
-    const uint32_t bits = fp32_to_bits(base);
-    const uint32_t exp_bits = (bits >> 13) & UINT32_C(0x00007C00);
-    const uint32_t mantissa_bits = bits & UINT32_C(0x00000FFF);
-    const uint32_t nonsign = exp_bits + mantissa_bits;
-    return (sign >> 16) | (shl1_w > UINT32_C(0xFF000000) ? UINT16_C(0x7E00) : nonsign);
-}
-
-#define GGML_COMPUTE_FP16_TO_FP32(x) ggml_compute_fp16_to_fp32(x)
-#define GGML_COMPUTE_FP32_TO_FP16(x) ggml_compute_fp32_to_fp16(x)
-
-#endif // __F16C__
-
-#endif // __ARM_NEON
-
-// precomputed f32 table for f16 (256 KB)
-// defined in ggml.c, initialized in ggml_init()
-extern float ggml_table_f32_f16[1 << 16];
-
-// On ARM NEON, it's quicker to directly convert x -> x instead of calling into ggml_lookup_fp16_to_fp32,
-// so we define GGML_FP16_TO_FP32 and GGML_FP32_TO_FP16 elsewhere for NEON.
-// This is also true for POWER9.
-#if !defined(GGML_FP16_TO_FP32) || !defined(GGML_FP32_TO_FP16)
-
-inline static float ggml_lookup_fp16_to_fp32(ggml_fp16_t f) {
-    uint16_t s;
-    memcpy(&s, &f, sizeof(uint16_t));
-    return ggml_table_f32_f16[s];
-}
-
-#define GGML_FP16_TO_FP32(x) ggml_lookup_fp16_to_fp32(x)
-#define GGML_FP32_TO_FP16(x) GGML_COMPUTE_FP32_TO_FP16(x)
-
-#endif
-
-#define GGML_HASHTABLE_FULL ((size_t)-1)
-#define GGML_HASHTABLE_ALREADY_EXISTS ((size_t)-2)
-
-bool   ggml_hash_contains      (const struct ggml_hash_set hash_set, struct ggml_tensor * key);
-
-// returns GGML_HASHTABLE_FULL if table is full, otherwise the current index of the key or where it should be inserted
-size_t ggml_hash_find          (const struct ggml_hash_set hash_set, struct ggml_tensor * key);
-
-// returns GGML_HASHTABLE_ALREADY_EXISTS if key already exists, index otherwise, asserts if table is full
-size_t ggml_hash_insert        (      struct ggml_hash_set hash_set, struct ggml_tensor * key);
-
-// return index, asserts if table is full
-size_t ggml_hash_find_or_insert(      struct ggml_hash_set hash_set, struct ggml_tensor * key);
-
-#ifdef __cplusplus
-}
-#endif
diff --git a/src/whisper_cpp/ggml-metal.h b/src/whisper_cpp/ggml-metal.h
deleted file mode 100644
index c4b7325d..00000000
--- a/src/whisper_cpp/ggml-metal.h
+++ /dev/null
@@ -1,115 +0,0 @@
-// An interface allowing to compute ggml_cgraph with Metal
-//
-// This is a fully functional interface that extends ggml with GPU support for Apple devices.
-// A similar interface can be created for other GPU backends (e.g. Vulkan, CUDA, OpenCL, etc.)
-//
-// How it works?
-//
-// As long as your program can create and evaluate a ggml_cgraph on the CPU, you can use this
-// interface to evaluate the same graph on the GPU. Instead of using ggml_graph_compute(), you
-// use ggml_metal_graph_compute() (or ggml_vulkan_graph_compute(), etc.)
-//
-// You only need to make sure that all memory buffers that you used during the graph creation
-// are mapped to the device memory with the ggml_metal_add_buffer() function. This mapping is
-// used during the graph evaluation to determine the arguments of the compute kernels.
-//
-// Synchronization between device and host memory (for example for input and output tensors)
-// is done with the ggml_metal_set_tensor() and ggml_metal_get_tensor() functions.
-//
-
-#pragma once
-
-#include "ggml.h"
-#include "ggml-backend.h"
-
-#include <stddef.h>
-#include <stdbool.h>
-
-// max memory buffers that can be mapped to the device
-#define GGML_METAL_MAX_BUFFERS 64
-#define GGML_METAL_MAX_COMMAND_BUFFERS 32
-
-struct ggml_tensor;
-struct ggml_cgraph;
-
-#ifdef __cplusplus
-extern "C" {
-#endif
-
-//
-// internal API
-// temporary exposed to user-code
-//
-
-struct ggml_metal_context;
-
-void ggml_metal_log_set_callback(ggml_log_callback log_callback, void * user_data);
-
-// number of command buffers to use
-struct ggml_metal_context * ggml_metal_init(int n_cb);
-void ggml_metal_free(struct ggml_metal_context * ctx);
-
-void * ggml_metal_host_malloc(size_t n);
-void   ggml_metal_host_free  (void * data);
-
-// set the number of command buffers to use
-void ggml_metal_set_n_cb(struct ggml_metal_context * ctx, int n_cb);
-
-// creates a mapping between a host memory buffer and a device memory buffer
-// - make sure to map all buffers used in the graph before calling ggml_metal_graph_compute
-// - the mapping is used during computation to determine the arguments of the compute kernels
-// - you don't need to keep the host memory buffer allocated as it is never accessed by Metal
-// - max_size specifies the maximum size of a tensor and is used to create shared views such
-//   that it is guaranteed that the tensor will fit in at least one of the views
-//
-bool ggml_metal_add_buffer(
-        struct ggml_metal_context * ctx,
-                       const char * name,
-                             void * data,
-                           size_t   size,
-                           size_t   max_size);
-
-// set data from host memory into the device
-void ggml_metal_set_tensor(struct ggml_metal_context * ctx, struct ggml_tensor * t);
-
-// get data from the device into host memory
-void ggml_metal_get_tensor(struct ggml_metal_context * ctx, struct ggml_tensor * t);
-
-// try to find operations that can be run concurrently in the graph
-// you should run it again if the topology of your graph changes
-void ggml_metal_graph_find_concurrency(struct ggml_metal_context * ctx, struct ggml_cgraph * gf, bool check_mem);
-
-// if the graph has been optimized for concurrently dispatch, return length of the concur_list if optimized
-int ggml_metal_if_optimized(struct ggml_metal_context * ctx);
-
-// output the concur_list for ggml_alloc
-int * ggml_metal_get_concur_list(struct ggml_metal_context * ctx);
-
-// same as ggml_graph_compute but uses Metal
-// creates gf->n_threads command buffers in parallel
-bool ggml_metal_graph_compute(struct ggml_metal_context * ctx, struct ggml_cgraph * gf);
-
-//
-// backend API
-// user-code should use only these functions
-//
-
-GGML_API ggml_backend_t ggml_backend_metal_init(void);
-
-GGML_API bool ggml_backend_is_metal(ggml_backend_t backend);
-
-GGML_API ggml_backend_buffer_t ggml_backend_metal_buffer_from_ptr(void * data, size_t size, size_t max_size);
-
-GGML_API void ggml_backend_metal_set_n_cb(ggml_backend_t backend, int n_cb);
-
-GGML_API ggml_backend_buffer_type_t ggml_backend_metal_buffer_type(void);
-
-// helper to check if the device supports a specific family
-// ideally, the user code should be doing these checks
-// ref: https://developer.apple.com/metal/Metal-Feature-Set-Tables.pdf
-GGML_API bool ggml_backend_metal_supports_family(ggml_backend_t backend, int family);
-
-#ifdef __cplusplus
-}
-#endif
-
diff --git a/src/whisper_cpp/ggml-metal.m b/src/whisper_cpp/ggml-metal.m
deleted file mode 100644
index 005a94ef..00000000
--- a/src/whisper_cpp/ggml-metal.m
+++ /dev/null
@@ -1,2768 +0,0 @@
-#ifndef R_NO_REMAP
-#define R_NO_REMAP 1
-#endif
-#import "R.h"
-#import "ggml-metal.h"
-
-#import "ggml-backend-impl.h"
-#import "ggml.h"
-
-#import <Foundation/Foundation.h>
-
-#import <Metal/Metal.h>
-
-#undef MIN
-#undef MAX
-#define MIN(a, b) ((a) < (b) ? (a) : (b))
-#define MAX(a, b) ((a) > (b) ? (a) : (b))
-
-#ifdef GGML_METAL_NDEBUG
-#define GGML_METAL_LOG_INFO(...)
-#define GGML_METAL_LOG_WARN(...)
-#define GGML_METAL_LOG_ERROR(...)
-#else
-#define GGML_METAL_LOG_INFO(...)  ggml_metal_log(GGML_LOG_LEVEL_INFO, __VA_ARGS__)
-#define GGML_METAL_LOG_WARN(...)  ggml_metal_log(GGML_LOG_LEVEL_WARN, __VA_ARGS__)
-#define GGML_METAL_LOG_ERROR(...) ggml_metal_log(GGML_LOG_LEVEL_ERROR, __VA_ARGS__)
-#endif
-
-#define UNUSED(x) (void)(x)
-
-#define GGML_MAX_CONCUR (2*GGML_DEFAULT_GRAPH_SIZE)
-
-struct ggml_metal_buffer {
-    const char * name;
-
-    void   * data;
-    size_t   size;
-
-    id<MTLBuffer> metal;
-};
-
-struct ggml_metal_context {
-    int n_cb;
-
-    id<MTLDevice>       device;
-    id<MTLCommandQueue> queue;
-    id<MTLLibrary>      library;
-
-    id<MTLCommandBuffer>         command_buffers [GGML_METAL_MAX_COMMAND_BUFFERS];
-    id<MTLComputeCommandEncoder> command_encoders[GGML_METAL_MAX_COMMAND_BUFFERS];
-
-    dispatch_queue_t d_queue;
-
-    int n_buffers;
-    struct ggml_metal_buffer buffers[GGML_METAL_MAX_BUFFERS];
-
-    int concur_list[GGML_MAX_CONCUR];
-    int concur_list_len;
-
-    // custom kernels
-#define GGML_METAL_DECL_KERNEL(name) \
-    id<MTLFunction>             function_##name; \
-    id<MTLComputePipelineState> pipeline_##name
-
-    GGML_METAL_DECL_KERNEL(add);
-    GGML_METAL_DECL_KERNEL(add_row); // TODO: avoid this extra kernel, instead extend the "add" kernel to support broadcast
-    GGML_METAL_DECL_KERNEL(mul);
-    GGML_METAL_DECL_KERNEL(mul_row); // TODO: avoid this extra kernel, instead extend the "mul" kernel to support broadcast
-    GGML_METAL_DECL_KERNEL(div);
-    GGML_METAL_DECL_KERNEL(div_row);
-    GGML_METAL_DECL_KERNEL(scale);
-    GGML_METAL_DECL_KERNEL(scale_4);
-    GGML_METAL_DECL_KERNEL(tanh);
-    GGML_METAL_DECL_KERNEL(relu);
-    GGML_METAL_DECL_KERNEL(gelu);
-    GGML_METAL_DECL_KERNEL(gelu_quick);
-    GGML_METAL_DECL_KERNEL(silu);
-    GGML_METAL_DECL_KERNEL(soft_max);
-    GGML_METAL_DECL_KERNEL(soft_max_4);
-    GGML_METAL_DECL_KERNEL(diag_mask_inf);
-    GGML_METAL_DECL_KERNEL(diag_mask_inf_8);
-    GGML_METAL_DECL_KERNEL(get_rows_f32);
-    GGML_METAL_DECL_KERNEL(get_rows_f16);
-    GGML_METAL_DECL_KERNEL(get_rows_q4_0);
-    GGML_METAL_DECL_KERNEL(get_rows_q4_1);
-    GGML_METAL_DECL_KERNEL(get_rows_q5_0);
-    GGML_METAL_DECL_KERNEL(get_rows_q5_1);
-    GGML_METAL_DECL_KERNEL(get_rows_q8_0);
-    GGML_METAL_DECL_KERNEL(get_rows_q2_K);
-    GGML_METAL_DECL_KERNEL(get_rows_q3_K);
-    GGML_METAL_DECL_KERNEL(get_rows_q4_K);
-    GGML_METAL_DECL_KERNEL(get_rows_q5_K);
-    GGML_METAL_DECL_KERNEL(get_rows_q6_K);
-    GGML_METAL_DECL_KERNEL(get_rows_i32);
-    GGML_METAL_DECL_KERNEL(rms_norm);
-    GGML_METAL_DECL_KERNEL(group_norm);
-    GGML_METAL_DECL_KERNEL(norm);
-    GGML_METAL_DECL_KERNEL(mul_mv_f32_f32);
-    GGML_METAL_DECL_KERNEL(mul_mv_f16_f16);
-    GGML_METAL_DECL_KERNEL(mul_mv_f16_f32);
-    GGML_METAL_DECL_KERNEL(mul_mv_f16_f32_1row);
-    GGML_METAL_DECL_KERNEL(mul_mv_f16_f32_l4);
-    GGML_METAL_DECL_KERNEL(mul_mv_q4_0_f32);
-    GGML_METAL_DECL_KERNEL(mul_mv_q4_1_f32);
-    GGML_METAL_DECL_KERNEL(mul_mv_q5_0_f32);
-    GGML_METAL_DECL_KERNEL(mul_mv_q5_1_f32);
-    GGML_METAL_DECL_KERNEL(mul_mv_q8_0_f32);
-    GGML_METAL_DECL_KERNEL(mul_mv_q2_K_f32);
-    GGML_METAL_DECL_KERNEL(mul_mv_q3_K_f32);
-    GGML_METAL_DECL_KERNEL(mul_mv_q4_K_f32);
-    GGML_METAL_DECL_KERNEL(mul_mv_q5_K_f32);
-    GGML_METAL_DECL_KERNEL(mul_mv_q6_K_f32);
-    GGML_METAL_DECL_KERNEL(mul_mv_id_f32_f32);
-    //GGML_METAL_DECL_KERNEL(mul_mv_id_f16_f16);
-    GGML_METAL_DECL_KERNEL(mul_mv_id_f16_f32);
-    //GGML_METAL_DECL_KERNEL(mul_mv_id_f16_f32_1row);
-    //GGML_METAL_DECL_KERNEL(mul_mv_id_f16_f32_l4);
-    GGML_METAL_DECL_KERNEL(mul_mv_id_q4_0_f32);
-    GGML_METAL_DECL_KERNEL(mul_mv_id_q4_1_f32);
-    GGML_METAL_DECL_KERNEL(mul_mv_id_q5_0_f32);
-    GGML_METAL_DECL_KERNEL(mul_mv_id_q5_1_f32);
-    GGML_METAL_DECL_KERNEL(mul_mv_id_q8_0_f32);
-    GGML_METAL_DECL_KERNEL(mul_mv_id_q2_K_f32);
-    GGML_METAL_DECL_KERNEL(mul_mv_id_q3_K_f32);
-    GGML_METAL_DECL_KERNEL(mul_mv_id_q4_K_f32);
-    GGML_METAL_DECL_KERNEL(mul_mv_id_q5_K_f32);
-    GGML_METAL_DECL_KERNEL(mul_mv_id_q6_K_f32);
-    GGML_METAL_DECL_KERNEL(mul_mm_f32_f32);
-    GGML_METAL_DECL_KERNEL(mul_mm_f16_f32);
-    GGML_METAL_DECL_KERNEL(mul_mm_q4_0_f32);
-    GGML_METAL_DECL_KERNEL(mul_mm_q4_1_f32);
-    GGML_METAL_DECL_KERNEL(mul_mm_q5_0_f32);
-    GGML_METAL_DECL_KERNEL(mul_mm_q5_1_f32);
-    GGML_METAL_DECL_KERNEL(mul_mm_q8_0_f32);
-    GGML_METAL_DECL_KERNEL(mul_mm_q2_K_f32);
-    GGML_METAL_DECL_KERNEL(mul_mm_q3_K_f32);
-    GGML_METAL_DECL_KERNEL(mul_mm_q4_K_f32);
-    GGML_METAL_DECL_KERNEL(mul_mm_q5_K_f32);
-    GGML_METAL_DECL_KERNEL(mul_mm_q6_K_f32);
-    GGML_METAL_DECL_KERNEL(mul_mm_id_f32_f32);
-    GGML_METAL_DECL_KERNEL(mul_mm_id_f16_f32);
-    GGML_METAL_DECL_KERNEL(mul_mm_id_q4_0_f32);
-    GGML_METAL_DECL_KERNEL(mul_mm_id_q4_1_f32);
-    GGML_METAL_DECL_KERNEL(mul_mm_id_q5_0_f32);
-    GGML_METAL_DECL_KERNEL(mul_mm_id_q5_1_f32);
-    GGML_METAL_DECL_KERNEL(mul_mm_id_q8_0_f32);
-    GGML_METAL_DECL_KERNEL(mul_mm_id_q2_K_f32);
-    GGML_METAL_DECL_KERNEL(mul_mm_id_q3_K_f32);
-    GGML_METAL_DECL_KERNEL(mul_mm_id_q4_K_f32);
-    GGML_METAL_DECL_KERNEL(mul_mm_id_q5_K_f32);
-    GGML_METAL_DECL_KERNEL(mul_mm_id_q6_K_f32);
-    GGML_METAL_DECL_KERNEL(rope_f32);
-    GGML_METAL_DECL_KERNEL(rope_f16);
-    GGML_METAL_DECL_KERNEL(alibi_f32);
-    GGML_METAL_DECL_KERNEL(im2col_f16);
-    GGML_METAL_DECL_KERNEL(upscale_f32);
-    GGML_METAL_DECL_KERNEL(pad_f32);
-    GGML_METAL_DECL_KERNEL(argsort_f32_i32_asc);
-    GGML_METAL_DECL_KERNEL(argsort_f32_i32_desc);
-    GGML_METAL_DECL_KERNEL(leaky_relu_f32);
-    GGML_METAL_DECL_KERNEL(cpy_f32_f16);
-    GGML_METAL_DECL_KERNEL(cpy_f32_f32);
-    GGML_METAL_DECL_KERNEL(cpy_f32_q8_0);
-    GGML_METAL_DECL_KERNEL(cpy_f32_q4_0);
-    GGML_METAL_DECL_KERNEL(cpy_f32_q4_1);
-    //GGML_METAL_DECL_KERNEL(cpy_f32_q5_0);
-    //GGML_METAL_DECL_KERNEL(cpy_f32_q5_1);
-    GGML_METAL_DECL_KERNEL(cpy_f16_f16);
-    GGML_METAL_DECL_KERNEL(cpy_f16_f32);
-    GGML_METAL_DECL_KERNEL(concat);
-    GGML_METAL_DECL_KERNEL(sqr);
-    GGML_METAL_DECL_KERNEL(sum_rows);
-
-#undef GGML_METAL_DECL_KERNEL
-};
-
-// MSL code
-// TODO: move the contents here when ready
-//       for now it is easier to work in a separate file
-//static NSString * const msl_library_source = @"see metal.metal";
-
-// Here to assist with NSBundle Path Hack
-@interface GGMLMetalClass : NSObject
-@end
-@implementation GGMLMetalClass
-@end
-
-
-static void ggml_metal_default_log_callback(enum ggml_log_level level, const char * msg, void * user_data) {
-    Rprintf("%s", msg);
-
-    UNUSED(level);
-    UNUSED(user_data);
-}
-
-ggml_log_callback ggml_metal_log_callback = ggml_metal_default_log_callback;
-void * ggml_metal_log_user_data = NULL;
-
-void ggml_metal_log_set_callback(ggml_log_callback log_callback, void * user_data) {
-    ggml_metal_log_callback  = log_callback;
-    ggml_metal_log_user_data = user_data;
-}
-
-GGML_ATTRIBUTE_FORMAT(2, 3)
-static void ggml_metal_log(enum ggml_log_level level, const char * format, ...){
-    if (ggml_metal_log_callback != NULL) {
-        va_list args;
-        va_start(args, format);
-        char buffer[128];
-        int len = vsnprintf(buffer, 128, format, args);
-        if (len < 128) {
-            ggml_metal_log_callback(level, buffer, ggml_metal_log_user_data);
-        } else {
-            char* buffer2 = malloc(len+1);
-            va_end(args);
-            va_start(args, format);
-            vsnprintf(buffer2, len+1, format, args);
-            buffer2[len] = 0;
-            ggml_metal_log_callback(level, buffer2, ggml_metal_log_user_data);
-            free(buffer2);
-        }
-        va_end(args);
-    }
-}
-
-struct ggml_metal_context * ggml_metal_init(int n_cb) {
-    GGML_METAL_LOG_INFO("%s: allocating\n", __func__);
-
-    id<MTLDevice> device;
-    NSString * s;
-
-#if TARGET_OS_OSX
-    // Show all the Metal device instances in the system
-    NSArray * devices = MTLCopyAllDevices();
-    for (device in devices) {
-        s = [device name];
-        GGML_METAL_LOG_INFO("%s: found device: %s\n", __func__, [s UTF8String]);
-    }
-#endif
-
-    // Pick and show default Metal device
-    device = MTLCreateSystemDefaultDevice();
-    s = [device name];
-    GGML_METAL_LOG_INFO("%s: picking default device: %s\n", __func__, [s UTF8String]);
-
-    // Configure context
-    struct ggml_metal_context * ctx = malloc(sizeof(struct ggml_metal_context));
-    ctx->device = device;
-    ctx->n_cb   = MIN(n_cb, GGML_METAL_MAX_BUFFERS);
-    ctx->queue  = [ctx->device newCommandQueue];
-    ctx->n_buffers = 0;
-    ctx->concur_list_len = 0;
-
-    ctx->d_queue = dispatch_queue_create("ggml-metal", DISPATCH_QUEUE_CONCURRENT);
-
-    // load library
-    {
-        NSBundle * bundle = nil;
-#ifdef SWIFT_PACKAGE
-        bundle = SWIFTPM_MODULE_BUNDLE;
-#else
-        bundle = [NSBundle bundleForClass:[GGMLMetalClass class]];
-#endif
-        NSError * error = nil;
-        NSString * libPath = [bundle pathForResource:@"default" ofType:@"metallib"];
-        if (libPath != nil) {
-            // pre-compiled library found
-            NSURL * libURL = [NSURL fileURLWithPath:libPath];
-            GGML_METAL_LOG_INFO("%s: loading '%s'\n", __func__, [libPath UTF8String]);
-            ctx->library = [ctx->device newLibraryWithURL:libURL error:&error];
-        } else {
-            GGML_METAL_LOG_INFO("%s: default.metallib not found, loading from source\n", __func__);
-
-            NSString * sourcePath;
-            NSString * ggmlMetalPathResources = [[NSProcessInfo processInfo].environment objectForKey:@"GGML_METAL_PATH_RESOURCES"];
-
-            GGML_METAL_LOG_INFO("%s: GGML_METAL_PATH_RESOURCES = %s\n", __func__, ggmlMetalPathResources ? [ggmlMetalPathResources UTF8String] : "nil");
-
-            if (ggmlMetalPathResources) {
-                sourcePath = [ggmlMetalPathResources stringByAppendingPathComponent:@"ggml-metal.metal"];
-            } else {
-                sourcePath = [bundle pathForResource:@"ggml-metal" ofType:@"metal"];
-            }
-            if (sourcePath == nil) {
-                GGML_METAL_LOG_WARN("%s: error: could not use bundle path to find ggml-metal.metal, falling back to trying cwd\n", __func__);
-                sourcePath = @"ggml-metal.metal";
-            }
-            GGML_METAL_LOG_INFO("%s: loading '%s'\n", __func__, [sourcePath UTF8String]);
-            NSString * src = [NSString stringWithContentsOfFile:sourcePath encoding:NSUTF8StringEncoding error:&error];
-            if (error) {
-                GGML_METAL_LOG_ERROR("%s: error: %s\n", __func__, [[error description] UTF8String]);
-                return NULL;
-            }
-
-            MTLCompileOptions* options = nil;
-#ifdef GGML_QKK_64
-            options = [MTLCompileOptions new];
-            options.preprocessorMacros = @{ @"QK_K" : @(64) };
-#endif
-            // try to disable fast-math
-            // NOTE: this seems to have no effect whatsoever
-            //       instead, in order to disable fast-math, we have to build default.metallib from the command line
-            //       using xcrun -sdk macosx metal -fno-fast-math -c ggml-metal.metal -o ggml-metal.air
-            //       and go through the "pre-compiled library found" path above
-            //[options setFastMathEnabled:false];
-
-            ctx->library = [ctx->device newLibraryWithSource:src options:options error:&error];
-        }
-
-        if (error) {
-            GGML_METAL_LOG_ERROR("%s: error: %s\n", __func__, [[error description] UTF8String]);
-            return NULL;
-        }
-    }
-
-#if TARGET_OS_OSX
-    // print MTL GPU family:
-    GGML_METAL_LOG_INFO("%s: GPU name:   %s\n", __func__, [[ctx->device name] UTF8String]);
-
-    // determine max supported GPU family
-    // https://developer.apple.com/metal/Metal-Shading-Language-Specification.pdf
-    // https://developer.apple.com/metal/Metal-Feature-Set-Tables.pdf
-    for (int i = MTLGPUFamilyApple1 + 20; i >= MTLGPUFamilyApple1; --i) {
-        if ([ctx->device supportsFamily:i]) {
-            GGML_METAL_LOG_INFO("%s: GPU family: MTLGPUFamilyApple%d (%d)\n", __func__, i - (int) MTLGPUFamilyApple1 + 1, i);
-            break;
-        }
-    }
-
-    GGML_METAL_LOG_INFO("%s: hasUnifiedMemory              = %s\n",       __func__, ctx->device.hasUnifiedMemory ? "true" : "false");
-    GGML_METAL_LOG_INFO("%s: recommendedMaxWorkingSetSize  = %8.2f MB\n", __func__, ctx->device.recommendedMaxWorkingSetSize / 1e6);
-    if (ctx->device.maxTransferRate != 0) {
-        GGML_METAL_LOG_INFO("%s: maxTransferRate               = %8.2f MB/s\n", __func__, ctx->device.maxTransferRate / 1e6);
-    } else {
-        GGML_METAL_LOG_INFO("%s: maxTransferRate               = built-in GPU\n", __func__);
-    }
-#endif
-
-    // load kernels
-    {
-        NSError * error = nil;
-
-        /*
-        GGML_METAL_LOG_INFO("%s: loaded %-32s %16p | th_max = %4d | th_width = %4d\n", __func__, "kernel_"#name, (void *) ctx->pipeline_##name, \
-                (int) ctx->pipeline_##name.maxTotalThreadsPerThreadgroup, \
-                (int) ctx->pipeline_##name.threadExecutionWidth); \
-        */
-#define GGML_METAL_ADD_KERNEL(name) \
-        ctx->function_##name = [ctx->library newFunctionWithName:@"kernel_"#name]; \
-        ctx->pipeline_##name = [ctx->device newComputePipelineStateWithFunction:ctx->function_##name error:&error]; \
-        if (error) { \
-            GGML_METAL_LOG_ERROR("%s: error: load pipeline error: %s\n", __func__, [[error description] UTF8String]); \
-            return NULL; \
-        }
-
-        GGML_METAL_ADD_KERNEL(add);
-        GGML_METAL_ADD_KERNEL(add_row);
-        GGML_METAL_ADD_KERNEL(mul);
-        GGML_METAL_ADD_KERNEL(mul_row);
-        GGML_METAL_ADD_KERNEL(div);
-        GGML_METAL_ADD_KERNEL(div_row);
-        GGML_METAL_ADD_KERNEL(scale);
-        GGML_METAL_ADD_KERNEL(scale_4);
-        GGML_METAL_ADD_KERNEL(tanh);
-        GGML_METAL_ADD_KERNEL(relu);
-        GGML_METAL_ADD_KERNEL(gelu);
-        GGML_METAL_ADD_KERNEL(gelu_quick);
-        GGML_METAL_ADD_KERNEL(silu);
-        GGML_METAL_ADD_KERNEL(soft_max);
-        GGML_METAL_ADD_KERNEL(soft_max_4);
-        GGML_METAL_ADD_KERNEL(diag_mask_inf);
-        GGML_METAL_ADD_KERNEL(diag_mask_inf_8);
-        GGML_METAL_ADD_KERNEL(get_rows_f32);
-        GGML_METAL_ADD_KERNEL(get_rows_f16);
-        GGML_METAL_ADD_KERNEL(get_rows_q4_0);
-        GGML_METAL_ADD_KERNEL(get_rows_q4_1);
-        GGML_METAL_ADD_KERNEL(get_rows_q5_0);
-        GGML_METAL_ADD_KERNEL(get_rows_q5_1);
-        GGML_METAL_ADD_KERNEL(get_rows_q8_0);
-        GGML_METAL_ADD_KERNEL(get_rows_q2_K);
-        GGML_METAL_ADD_KERNEL(get_rows_q3_K);
-        GGML_METAL_ADD_KERNEL(get_rows_q4_K);
-        GGML_METAL_ADD_KERNEL(get_rows_q5_K);
-        GGML_METAL_ADD_KERNEL(get_rows_q6_K);
-        GGML_METAL_ADD_KERNEL(get_rows_i32);
-        GGML_METAL_ADD_KERNEL(rms_norm);
-        GGML_METAL_ADD_KERNEL(group_norm);
-        GGML_METAL_ADD_KERNEL(norm);
-        GGML_METAL_ADD_KERNEL(mul_mv_f32_f32);
-        GGML_METAL_ADD_KERNEL(mul_mv_f16_f16);
-        GGML_METAL_ADD_KERNEL(mul_mv_f16_f32);
-        GGML_METAL_ADD_KERNEL(mul_mv_f16_f32_1row);
-        GGML_METAL_ADD_KERNEL(mul_mv_f16_f32_l4);
-        GGML_METAL_ADD_KERNEL(mul_mv_q4_0_f32);
-        GGML_METAL_ADD_KERNEL(mul_mv_q4_1_f32);
-        GGML_METAL_ADD_KERNEL(mul_mv_q5_0_f32);
-        GGML_METAL_ADD_KERNEL(mul_mv_q5_1_f32);
-        GGML_METAL_ADD_KERNEL(mul_mv_q8_0_f32);
-        GGML_METAL_ADD_KERNEL(mul_mv_q2_K_f32);
-        GGML_METAL_ADD_KERNEL(mul_mv_q3_K_f32);
-        GGML_METAL_ADD_KERNEL(mul_mv_q4_K_f32);
-        GGML_METAL_ADD_KERNEL(mul_mv_q5_K_f32);
-        GGML_METAL_ADD_KERNEL(mul_mv_q6_K_f32);
-        GGML_METAL_ADD_KERNEL(mul_mv_id_f32_f32);
-        //GGML_METAL_ADD_KERNEL(mul_mv_id_f16_f16);
-        GGML_METAL_ADD_KERNEL(mul_mv_id_f16_f32);
-        //GGML_METAL_ADD_KERNEL(mul_mv_id_f16_f32_1row);
-        //GGML_METAL_ADD_KERNEL(mul_mv_id_f16_f32_l4);
-        GGML_METAL_ADD_KERNEL(mul_mv_id_q4_0_f32);
-        GGML_METAL_ADD_KERNEL(mul_mv_id_q4_1_f32);
-        GGML_METAL_ADD_KERNEL(mul_mv_id_q5_0_f32);
-        GGML_METAL_ADD_KERNEL(mul_mv_id_q5_1_f32);
-        GGML_METAL_ADD_KERNEL(mul_mv_id_q8_0_f32);
-        GGML_METAL_ADD_KERNEL(mul_mv_id_q2_K_f32);
-        GGML_METAL_ADD_KERNEL(mul_mv_id_q3_K_f32);
-        GGML_METAL_ADD_KERNEL(mul_mv_id_q4_K_f32);
-        GGML_METAL_ADD_KERNEL(mul_mv_id_q5_K_f32);
-        GGML_METAL_ADD_KERNEL(mul_mv_id_q6_K_f32);
-        if ([ctx->device supportsFamily:MTLGPUFamilyApple7]) {
-            GGML_METAL_ADD_KERNEL(mul_mm_f32_f32);
-            GGML_METAL_ADD_KERNEL(mul_mm_f16_f32);
-            GGML_METAL_ADD_KERNEL(mul_mm_q4_0_f32);
-            GGML_METAL_ADD_KERNEL(mul_mm_q4_1_f32);
-            GGML_METAL_ADD_KERNEL(mul_mm_q5_0_f32);
-            GGML_METAL_ADD_KERNEL(mul_mm_q5_1_f32);
-            GGML_METAL_ADD_KERNEL(mul_mm_q8_0_f32);
-            GGML_METAL_ADD_KERNEL(mul_mm_q2_K_f32);
-            GGML_METAL_ADD_KERNEL(mul_mm_q3_K_f32);
-            GGML_METAL_ADD_KERNEL(mul_mm_q4_K_f32);
-            GGML_METAL_ADD_KERNEL(mul_mm_q5_K_f32);
-            GGML_METAL_ADD_KERNEL(mul_mm_q6_K_f32);
-            GGML_METAL_ADD_KERNEL(mul_mm_id_f32_f32);
-            GGML_METAL_ADD_KERNEL(mul_mm_id_f16_f32);
-            GGML_METAL_ADD_KERNEL(mul_mm_id_q4_0_f32);
-            GGML_METAL_ADD_KERNEL(mul_mm_id_q4_1_f32);
-            GGML_METAL_ADD_KERNEL(mul_mm_id_q5_0_f32);
-            GGML_METAL_ADD_KERNEL(mul_mm_id_q5_1_f32);
-            GGML_METAL_ADD_KERNEL(mul_mm_id_q8_0_f32);
-            GGML_METAL_ADD_KERNEL(mul_mm_id_q2_K_f32);
-            GGML_METAL_ADD_KERNEL(mul_mm_id_q3_K_f32);
-            GGML_METAL_ADD_KERNEL(mul_mm_id_q4_K_f32);
-            GGML_METAL_ADD_KERNEL(mul_mm_id_q5_K_f32);
-            GGML_METAL_ADD_KERNEL(mul_mm_id_q6_K_f32);
-        }
-        GGML_METAL_ADD_KERNEL(rope_f32);
-        GGML_METAL_ADD_KERNEL(rope_f16);
-        GGML_METAL_ADD_KERNEL(alibi_f32);
-        GGML_METAL_ADD_KERNEL(im2col_f16);
-        GGML_METAL_ADD_KERNEL(upscale_f32);
-        GGML_METAL_ADD_KERNEL(pad_f32);
-        GGML_METAL_ADD_KERNEL(argsort_f32_i32_asc);
-        GGML_METAL_ADD_KERNEL(argsort_f32_i32_desc);
-        GGML_METAL_ADD_KERNEL(leaky_relu_f32);
-        GGML_METAL_ADD_KERNEL(cpy_f32_f16);
-        GGML_METAL_ADD_KERNEL(cpy_f32_f32);
-        GGML_METAL_ADD_KERNEL(cpy_f32_q8_0);
-        GGML_METAL_ADD_KERNEL(cpy_f32_q4_0);
-        GGML_METAL_ADD_KERNEL(cpy_f32_q4_1);
-        //GGML_METAL_ADD_KERNEL(cpy_f32_q5_0);
-        //GGML_METAL_ADD_KERNEL(cpy_f32_q5_1);
-        GGML_METAL_ADD_KERNEL(cpy_f16_f16);
-        GGML_METAL_ADD_KERNEL(cpy_f16_f32);
-        GGML_METAL_ADD_KERNEL(concat);
-        GGML_METAL_ADD_KERNEL(sqr);
-        GGML_METAL_ADD_KERNEL(sum_rows);
-
-#undef GGML_METAL_ADD_KERNEL
-    }
-
-    return ctx;
-}
-
-void ggml_metal_free(struct ggml_metal_context * ctx) {
-    GGML_METAL_LOG_INFO("%s: deallocating\n", __func__);
-#define GGML_METAL_DEL_KERNEL(name) \
-    [ctx->function_##name release]; \
-    [ctx->pipeline_##name release];
-
-    GGML_METAL_DEL_KERNEL(add);
-    GGML_METAL_DEL_KERNEL(add_row);
-    GGML_METAL_DEL_KERNEL(mul);
-    GGML_METAL_DEL_KERNEL(mul_row);
-    GGML_METAL_DEL_KERNEL(div);
-    GGML_METAL_DEL_KERNEL(div_row);
-    GGML_METAL_DEL_KERNEL(scale);
-    GGML_METAL_DEL_KERNEL(scale_4);
-    GGML_METAL_DEL_KERNEL(tanh);
-    GGML_METAL_DEL_KERNEL(relu);
-    GGML_METAL_DEL_KERNEL(gelu);
-    GGML_METAL_DEL_KERNEL(gelu_quick);
-    GGML_METAL_DEL_KERNEL(silu);
-    GGML_METAL_DEL_KERNEL(soft_max);
-    GGML_METAL_DEL_KERNEL(soft_max_4);
-    GGML_METAL_DEL_KERNEL(diag_mask_inf);
-    GGML_METAL_DEL_KERNEL(diag_mask_inf_8);
-    GGML_METAL_DEL_KERNEL(get_rows_f32);
-    GGML_METAL_DEL_KERNEL(get_rows_f16);
-    GGML_METAL_DEL_KERNEL(get_rows_q4_0);
-    GGML_METAL_DEL_KERNEL(get_rows_q4_1);
-    GGML_METAL_DEL_KERNEL(get_rows_q5_0);
-    GGML_METAL_DEL_KERNEL(get_rows_q5_1);
-    GGML_METAL_DEL_KERNEL(get_rows_q8_0);
-    GGML_METAL_DEL_KERNEL(get_rows_q2_K);
-    GGML_METAL_DEL_KERNEL(get_rows_q3_K);
-    GGML_METAL_DEL_KERNEL(get_rows_q4_K);
-    GGML_METAL_DEL_KERNEL(get_rows_q5_K);
-    GGML_METAL_DEL_KERNEL(get_rows_q6_K);
-    GGML_METAL_DEL_KERNEL(get_rows_i32);
-    GGML_METAL_DEL_KERNEL(rms_norm);
-    GGML_METAL_DEL_KERNEL(group_norm);
-    GGML_METAL_DEL_KERNEL(norm);
-    GGML_METAL_DEL_KERNEL(mul_mv_f32_f32);
-    GGML_METAL_DEL_KERNEL(mul_mv_f16_f16);
-    GGML_METAL_DEL_KERNEL(mul_mv_f16_f32);
-    GGML_METAL_DEL_KERNEL(mul_mv_f16_f32_1row);
-    GGML_METAL_DEL_KERNEL(mul_mv_f16_f32_l4);
-    GGML_METAL_DEL_KERNEL(mul_mv_q4_0_f32);
-    GGML_METAL_DEL_KERNEL(mul_mv_q4_1_f32);
-    GGML_METAL_DEL_KERNEL(mul_mv_q5_0_f32);
-    GGML_METAL_DEL_KERNEL(mul_mv_q5_1_f32);
-    GGML_METAL_DEL_KERNEL(mul_mv_q8_0_f32);
-    GGML_METAL_DEL_KERNEL(mul_mv_q2_K_f32);
-    GGML_METAL_DEL_KERNEL(mul_mv_q3_K_f32);
-    GGML_METAL_DEL_KERNEL(mul_mv_q4_K_f32);
-    GGML_METAL_DEL_KERNEL(mul_mv_q5_K_f32);
-    GGML_METAL_DEL_KERNEL(mul_mv_q6_K_f32);
-    GGML_METAL_DEL_KERNEL(mul_mv_id_f32_f32);
-    //GGML_METAL_DEL_KERNEL(mul_mv_id_f16_f16);
-    GGML_METAL_DEL_KERNEL(mul_mv_id_f16_f32);
-    //GGML_METAL_DEL_KERNEL(mul_mv_id_f16_f32_1row);
-    //GGML_METAL_DEL_KERNEL(mul_mv_id_f16_f32_l4);
-    GGML_METAL_DEL_KERNEL(mul_mv_id_q4_0_f32);
-    GGML_METAL_DEL_KERNEL(mul_mv_id_q4_1_f32);
-    GGML_METAL_DEL_KERNEL(mul_mv_id_q5_0_f32);
-    GGML_METAL_DEL_KERNEL(mul_mv_id_q5_1_f32);
-    GGML_METAL_DEL_KERNEL(mul_mv_id_q8_0_f32);
-    GGML_METAL_DEL_KERNEL(mul_mv_id_q2_K_f32);
-    GGML_METAL_DEL_KERNEL(mul_mv_id_q3_K_f32);
-    GGML_METAL_DEL_KERNEL(mul_mv_id_q4_K_f32);
-    GGML_METAL_DEL_KERNEL(mul_mv_id_q5_K_f32);
-    GGML_METAL_DEL_KERNEL(mul_mv_id_q6_K_f32);
-    if ([ctx->device supportsFamily:MTLGPUFamilyApple7]) {
-        GGML_METAL_DEL_KERNEL(mul_mm_f32_f32);
-        GGML_METAL_DEL_KERNEL(mul_mm_f16_f32);
-        GGML_METAL_DEL_KERNEL(mul_mm_q4_0_f32);
-        GGML_METAL_DEL_KERNEL(mul_mm_q4_1_f32);
-        GGML_METAL_DEL_KERNEL(mul_mm_q5_0_f32);
-        GGML_METAL_DEL_KERNEL(mul_mm_q5_1_f32);
-        GGML_METAL_DEL_KERNEL(mul_mm_q8_0_f32);
-        GGML_METAL_DEL_KERNEL(mul_mm_q2_K_f32);
-        GGML_METAL_DEL_KERNEL(mul_mm_q3_K_f32);
-        GGML_METAL_DEL_KERNEL(mul_mm_q4_K_f32);
-        GGML_METAL_DEL_KERNEL(mul_mm_q5_K_f32);
-        GGML_METAL_DEL_KERNEL(mul_mm_q6_K_f32);
-        GGML_METAL_DEL_KERNEL(mul_mm_id_f32_f32);
-        GGML_METAL_DEL_KERNEL(mul_mm_id_f16_f32);
-        GGML_METAL_DEL_KERNEL(mul_mm_id_q4_0_f32);
-        GGML_METAL_DEL_KERNEL(mul_mm_id_q4_1_f32);
-        GGML_METAL_DEL_KERNEL(mul_mm_id_q5_0_f32);
-        GGML_METAL_DEL_KERNEL(mul_mm_id_q5_1_f32);
-        GGML_METAL_DEL_KERNEL(mul_mm_id_q8_0_f32);
-        GGML_METAL_DEL_KERNEL(mul_mm_id_q2_K_f32);
-        GGML_METAL_DEL_KERNEL(mul_mm_id_q3_K_f32);
-        GGML_METAL_DEL_KERNEL(mul_mm_id_q4_K_f32);
-        GGML_METAL_DEL_KERNEL(mul_mm_id_q5_K_f32);
-        GGML_METAL_DEL_KERNEL(mul_mm_id_q6_K_f32);
-    }
-    GGML_METAL_DEL_KERNEL(rope_f32);
-    GGML_METAL_DEL_KERNEL(rope_f16);
-    GGML_METAL_DEL_KERNEL(alibi_f32);
-    GGML_METAL_DEL_KERNEL(im2col_f16);
-    GGML_METAL_DEL_KERNEL(upscale_f32);
-    GGML_METAL_DEL_KERNEL(pad_f32);
-    GGML_METAL_DEL_KERNEL(argsort_f32_i32_asc);
-    GGML_METAL_DEL_KERNEL(argsort_f32_i32_desc);
-    GGML_METAL_DEL_KERNEL(leaky_relu_f32);
-    GGML_METAL_DEL_KERNEL(cpy_f32_f16);
-    GGML_METAL_DEL_KERNEL(cpy_f32_f32);
-    GGML_METAL_DEL_KERNEL(cpy_f32_q8_0);
-    GGML_METAL_DEL_KERNEL(cpy_f32_q4_0);
-    GGML_METAL_DEL_KERNEL(cpy_f32_q4_1);
-    //GGML_METAL_DEL_KERNEL(cpy_f32_q5_0);
-    //GGML_METAL_DEL_KERNEL(cpy_f32_q5_1);
-    GGML_METAL_DEL_KERNEL(cpy_f16_f16);
-    GGML_METAL_DEL_KERNEL(cpy_f16_f32);
-    GGML_METAL_DEL_KERNEL(concat);
-    GGML_METAL_DEL_KERNEL(sqr);
-    GGML_METAL_DEL_KERNEL(sum_rows);
-
-#undef GGML_METAL_DEL_KERNEL
-
-    for (int i = 0; i < ctx->n_buffers; ++i) {
-        [ctx->buffers[i].metal release];
-    }
-
-    [ctx->library release];
-    [ctx->queue release];
-    [ctx->device release];
-
-    dispatch_release(ctx->d_queue);
-
-    free(ctx);
-}
-
-void * ggml_metal_host_malloc(size_t n) {
-    void * data = NULL;
-    const int result = posix_memalign((void **) &data, sysconf(_SC_PAGESIZE), n);
-    if (result != 0) {
-        GGML_METAL_LOG_ERROR("%s: error: posix_memalign failed\n", __func__);
-        return NULL;
-    }
-
-    return data;
-}
-
-void ggml_metal_host_free(void * data) {
-    free(data);
-}
-
-void ggml_metal_set_n_cb(struct ggml_metal_context * ctx, int n_cb) {
-    ctx->n_cb = MIN(n_cb, GGML_METAL_MAX_BUFFERS);
-}
-
-int ggml_metal_if_optimized(struct ggml_metal_context * ctx) {
-    return ctx->concur_list_len;
-}
-
-int * ggml_metal_get_concur_list(struct ggml_metal_context * ctx) {
-    return ctx->concur_list;
-}
-
-// temporarily defined here for compatibility between ggml-backend and the old API
-
-struct ggml_backend_metal_buffer {
-    void   * data;
-    size_t   size;
-
-    id<MTLBuffer> metal;
-};
-
-struct ggml_backend_metal_buffer_context {
-    void * all_data;
-    size_t all_size;
-    bool owned;
-
-    // multiple buffers are used only to avoid the maximum buffer size limitation when using mmap
-    int n_buffers;
-    struct ggml_backend_metal_buffer buffers[GGML_METAL_MAX_BUFFERS];
-};
-
-// finds the Metal buffer that contains the tensor data on the GPU device
-// the assumption is that there is 1-to-1 mapping between the host and device memory buffers, so we can find the
-// Metal buffer based on the host memory pointer
-//
-static id<MTLBuffer> ggml_metal_get_buffer(struct ggml_metal_context * ctx, struct ggml_tensor * t, size_t * offs) {
-    //GGML_METAL_LOG_INFO("%s: data tensor '%16s', offs_data = %8ld, offs_eval = %8ld, offs_cach = %8ld\n", __func__, t->name, offs_data, offs_eval, offs_cach);
-
-    const int64_t tsize = ggml_nbytes(t);
-
-    ggml_backend_buffer_t buffer = t->view_src ? t->view_src->buffer : t->buffer;
-
-    // compatibility with ggml-backend
-    if (buffer && buffer->buft == ggml_backend_metal_buffer_type()) {
-        struct ggml_backend_metal_buffer_context * buf_ctx = (struct ggml_backend_metal_buffer_context *) buffer->context;
-
-        // find the view that contains the tensor fully
-        for (int i = 0; i < buf_ctx->n_buffers; ++i) {
-            const int64_t ioffs = (int64_t) t->data - (int64_t) buf_ctx->buffers[i].data;
-
-            //GGML_METAL_LOG_INFO("ioffs = %10ld, tsize = %10ld, sum = %10ld, buf_ctx->buffers[%d].size = %10ld\n", ioffs, tsize, ioffs + tsize, i, buf_ctx->buffers[i].size);
-            if (ioffs >= 0 && ioffs + tsize <= (int64_t) buf_ctx->buffers[i].size) {
-                *offs = (size_t) ioffs;
-
-                //GGML_METAL_LOG_INFO("%s: tensor '%16s', offs = %8ld\n", __func__, t->name, *offs);
-
-                return buf_ctx->buffers[i].metal;
-            }
-        }
-
-        GGML_METAL_LOG_ERROR("%s: error: tensor '%s' buffer is nil\n", __func__, t->name);
-
-        return nil;
-    }
-
-    // find the view that contains the tensor fully
-    for (int i = 0; i < ctx->n_buffers; ++i) {
-        const int64_t ioffs = (int64_t) t->data - (int64_t) ctx->buffers[i].data;
-
-        //GGML_METAL_LOG_INFO("ioffs = %10ld, tsize = %10ld, sum = %10ld, ctx->buffers[%d].size = %10ld, name = %s\n", ioffs, tsize, ioffs + tsize, i, ctx->buffers[i].size, ctx->buffers[i].name);
-        if (ioffs >= 0 && ioffs + tsize <= (int64_t) ctx->buffers[i].size) {
-            *offs = (size_t) ioffs;
-
-            //GGML_METAL_LOG_INFO("%s: '%s' tensor '%16s', offs = %8ld\n", __func__, ctx->buffers[i].name, t->name, *offs);
-
-            return ctx->buffers[i].metal;
-        }
-    }
-
-    GGML_METAL_LOG_ERROR("%s: error: buffer is nil\n", __func__);
-
-    return nil;
-}
-
-bool ggml_metal_add_buffer(
-        struct ggml_metal_context * ctx,
-                     const char * name,
-                           void * data,
-                         size_t   size,
-                         size_t   max_size) {
-    if (ctx->n_buffers >= GGML_METAL_MAX_BUFFERS) {
-        GGML_METAL_LOG_ERROR("%s: error: too many buffers\n", __func__);
-        return false;
-    }
-
-    if (data) {
-        // verify that the buffer does not overlap with any of the existing buffers
-        for (int i = 0; i < ctx->n_buffers; ++i) {
-            const int64_t ioffs = (int64_t) data - (int64_t) ctx->buffers[i].data;
-
-            if (ioffs >= 0 && ioffs < (int64_t) ctx->buffers[i].size) {
-                GGML_METAL_LOG_ERROR("%s: error: buffer '%s' overlaps with '%s'\n", __func__, name, ctx->buffers[i].name);
-                return false;
-            }
-        }
-
-        const size_t size_page = sysconf(_SC_PAGESIZE);
-
-        size_t size_aligned = size;
-        if ((size_aligned % size_page) != 0) {
-            size_aligned += (size_page - (size_aligned % size_page));
-        }
-
-        // the buffer fits into the max buffer size allowed by the device
-        if (size_aligned <= ctx->device.maxBufferLength) {
-            ctx->buffers[ctx->n_buffers].name = name;
-            ctx->buffers[ctx->n_buffers].data = data;
-            ctx->buffers[ctx->n_buffers].size = size;
-
-            ctx->buffers[ctx->n_buffers].metal = [ctx->device newBufferWithBytesNoCopy:data length:size_aligned options:MTLResourceStorageModeShared deallocator:nil];
-
-            if (ctx->buffers[ctx->n_buffers].metal == nil) {
-                GGML_METAL_LOG_ERROR("%s: error: failed to allocate '%-16s' buffer, size = %8.2f MiB\n", __func__, name, size_aligned / 1024.0 / 1024.0);
-                return false;
-            }
-
-            GGML_METAL_LOG_INFO("%s: allocated '%-16s' buffer, size = %8.2f MiB", __func__, name, size_aligned / 1024.0 / 1024.0);
-
-            ++ctx->n_buffers;
-        } else {
-            // this overlap between the views will guarantee that the tensor with the maximum size will fully fit into
-            // one of the views
-            const size_t size_ovlp = ((max_size + size_page - 1) / size_page + 1) * size_page; // round-up 2 pages just in case
-            const size_t size_step = ctx->device.maxBufferLength - size_ovlp;
-            const size_t size_view = ctx->device.maxBufferLength;
-
-            for (size_t i = 0; i < size; i += size_step) {
-                const size_t size_step_aligned = (i + size_view <= size) ? size_view : (size_aligned - i);
-
-                ctx->buffers[ctx->n_buffers].name = name;
-                ctx->buffers[ctx->n_buffers].data = (void *) ((uint8_t *) data + i);
-                ctx->buffers[ctx->n_buffers].size = size_step_aligned;
-
-                ctx->buffers[ctx->n_buffers].metal = [ctx->device newBufferWithBytesNoCopy:(void *) ((uint8_t *) data + i) length:size_step_aligned options:MTLResourceStorageModeShared deallocator:nil];
-
-                if (ctx->buffers[ctx->n_buffers].metal == nil) {
-                    GGML_METAL_LOG_ERROR("%s: error: failed to allocate '%-16s' buffer, size = %8.2f MiB\n", __func__, name, size_step_aligned / 1024.0 / 1024.0);
-                    return false;
-                }
-
-                GGML_METAL_LOG_INFO("%s: allocated '%-16s' buffer, size = %8.2f MiB, offs = %12ld", __func__, name, size_step_aligned / 1024.0 / 1024.0, i);
-                if (i + size_step < size) {
-                    GGML_METAL_LOG_INFO("\n");
-                }
-
-                ++ctx->n_buffers;
-            }
-        }
-
-#if TARGET_OS_OSX
-        GGML_METAL_LOG_INFO(", (%8.2f / %8.2f)",
-                ctx->device.currentAllocatedSize / 1024.0 / 1024.0,
-                ctx->device.recommendedMaxWorkingSetSize / 1024.0 / 1024.0);
-
-        if (ctx->device.currentAllocatedSize > ctx->device.recommendedMaxWorkingSetSize) {
-            GGML_METAL_LOG_WARN("%s: warning: current allocated size is greater than the recommended max working set size\n", __func__);
-        } else {
-            GGML_METAL_LOG_INFO("\n");
-        }
-#else
-        GGML_METAL_LOG_INFO(", (%8.2f)\n", ctx->device.currentAllocatedSize / 1024.0 / 1024.0);
-#endif
-    }
-
-    return true;
-}
-
-void ggml_metal_set_tensor(
-        struct ggml_metal_context * ctx,
-        struct ggml_tensor * t) {
-    size_t offs;
-    id<MTLBuffer> id_dst = ggml_metal_get_buffer(ctx, t, &offs);
-
-    memcpy((void *) ((uint8_t *) id_dst.contents + offs), t->data, ggml_nbytes(t));
-}
-
-void ggml_metal_get_tensor(
-        struct ggml_metal_context * ctx,
-        struct ggml_tensor * t) {
-    size_t offs;
-    id<MTLBuffer> id_src = ggml_metal_get_buffer(ctx, t, &offs);
-
-    memcpy(t->data, (void *) ((uint8_t *) id_src.contents + offs), ggml_nbytes(t));
-}
-
-void ggml_metal_graph_find_concurrency(
-        struct ggml_metal_context * ctx,
-        struct ggml_cgraph * gf, bool check_mem) {
-    int search_depth = gf->n_nodes; //we only find concurrency in this range to avoid wasting too much time
-    int nodes_unused[GGML_MAX_CONCUR];
-
-    for (int i = 0; i < GGML_MAX_CONCUR; i++) { ctx->concur_list[i] = 0; }
-    for (int i = 0; i < gf->n_nodes;     i++) { nodes_unused[i]     = 1; }
-    ctx->concur_list_len = 0;
-
-    int n_left    = gf->n_nodes;
-    int n_start   = 0; // all nodes before n_start at nodes_unused array have been sorted and store back to ctx->concur_list
-    int level_pos = 0; // at ctx->concur_list, the last layer (level) ends at level_pos
-
-    while (n_left > 0) {
-        // number of nodes at a layer (that can be issued concurrently)
-        int concurrency = 0;
-        for (int i = n_start; i < ((n_start + search_depth > gf->n_nodes) ? gf->n_nodes : n_start + search_depth); i++) {
-            if (nodes_unused[i]) {
-                // if the requirements for gf->nodes[i] are satisfied
-                int exe_flag = 1;
-
-                // scan all srcs
-                for (int src_ind = 0; src_ind < GGML_MAX_SRC; src_ind++) {
-                    struct ggml_tensor * src_cur = gf->nodes[i]->src[src_ind];
-                    if (src_cur) {
-                        // if is leaf nodes it's satisfied.
-                        // TODO: ggml_is_leaf()
-                        if (src_cur->op == GGML_OP_NONE && src_cur->grad == NULL) {
-                            continue;
-                        }
-
-                        // otherwise this src should be the output from previous nodes.
-                        int is_found = 0;
-
-                        // scan 2*search_depth back because we inserted barrier.
-                        //for (int j = ((level_pos - 2*search_depth) < 0 ? 0 : (level_pos - 2*search_depth)); j < level_pos; j++) {
-                        for (int j = MAX(0, level_pos - 2*search_depth); j < level_pos; j++) {
-                            if (ctx->concur_list[j] >= 0 && gf->nodes[ctx->concur_list[j]] == src_cur) {
-                                is_found = 1;
-                                break;
-                            }
-                        }
-                        if (is_found == 0) {
-                            exe_flag = 0;
-                            break;
-                        }
-                    }
-                }
-                if (exe_flag && check_mem) {
-                    // check if nodes[i]'s data will be overwritten by a node before nodes[i].
-                    // if node[5] and node[3] write to the same memory region, then we can't issue node[5] before node[3]
-                    int64_t data_start = (int64_t) gf->nodes[i]->data;
-                    int64_t length     = (int64_t) ggml_nbytes(gf->nodes[i]);
-                    for (int j = n_start; j < i; j++) {
-                        if (nodes_unused[j] && gf->nodes[j]->op != GGML_OP_RESHAPE \
-                                            && gf->nodes[j]->op != GGML_OP_VIEW \
-                                            && gf->nodes[j]->op != GGML_OP_TRANSPOSE \
-                                            && gf->nodes[j]->op != GGML_OP_PERMUTE) {
-                            if (((int64_t)gf->nodes[j]->data) >= data_start + length || \
-                                ((int64_t)gf->nodes[j]->data) + (int64_t) ggml_nbytes(gf->nodes[j]) <= data_start) {
-                                continue;
-                            }
-
-                            exe_flag = 0;
-                        }
-                    }
-                }
-                if (exe_flag) {
-                    ctx->concur_list[level_pos + concurrency] = i;
-                    nodes_unused[i] = 0;
-                    concurrency++;
-                    ctx->concur_list_len++;
-                }
-            }
-        }
-        n_left -= concurrency;
-        // adding a barrier different layer
-        ctx->concur_list[level_pos + concurrency] = -1;
-        ctx->concur_list_len++;
-        // jump all sorted nodes at nodes_bak
-        while (!nodes_unused[n_start]) {
-            n_start++;
-        }
-        level_pos += concurrency + 1;
-    }
-
-    if (ctx->concur_list_len > GGML_MAX_CONCUR) {
-        GGML_METAL_LOG_WARN("%s: too many elements for metal ctx->concur_list!\n", __func__);
-    }
-}
-
-static bool ggml_metal_supports_op(const struct ggml_tensor * op) {
-    switch (op->op) {
-        case GGML_OP_UNARY:
-            switch (ggml_get_unary_op(op)) {
-                case GGML_UNARY_OP_TANH:
-                case GGML_UNARY_OP_RELU:
-                case GGML_UNARY_OP_GELU:
-                case GGML_UNARY_OP_GELU_QUICK:
-                case GGML_UNARY_OP_SILU:
-                    return true;
-                default:
-                    return false;
-            }
-        case GGML_OP_NONE:
-        case GGML_OP_RESHAPE:
-        case GGML_OP_VIEW:
-        case GGML_OP_TRANSPOSE:
-        case GGML_OP_PERMUTE:
-        case GGML_OP_CONCAT:
-        case GGML_OP_ADD:
-        case GGML_OP_ACC:
-        case GGML_OP_MUL:
-        case GGML_OP_DIV:
-        case GGML_OP_SCALE:
-        case GGML_OP_SQR:
-        case GGML_OP_SUM_ROWS:
-        case GGML_OP_SOFT_MAX:
-        case GGML_OP_RMS_NORM:
-        case GGML_OP_GROUP_NORM:
-        case GGML_OP_NORM:
-        case GGML_OP_ALIBI:
-        case GGML_OP_ROPE:
-        case GGML_OP_IM2COL:
-        case GGML_OP_UPSCALE:
-        case GGML_OP_PAD:
-        case GGML_OP_ARGSORT:
-        case GGML_OP_LEAKY_RELU:
-        case GGML_OP_MUL_MAT:
-        case GGML_OP_MUL_MAT_ID:
-            return true;
-        case GGML_OP_CPY:
-        case GGML_OP_DUP:
-        case GGML_OP_CONT:
-            {
-                switch (op->src[0]->type) {
-                    case GGML_TYPE_F32:
-                        switch (op->type) {
-                           case GGML_TYPE_F16:
-                           case GGML_TYPE_F32:
-                           case GGML_TYPE_Q8_0:
-                           case GGML_TYPE_Q4_0:
-                           case GGML_TYPE_Q4_1:
-                                return true;
-                           default:
-                                return false;
-                        }
-                    case GGML_TYPE_F16:
-                        switch (op->type) {
-                           case GGML_TYPE_F16:
-                           case GGML_TYPE_F32:
-                                return true;
-                           default:
-                                return false;
-                        }
-                    default:
-                        return false;
-                };
-            }
-        case GGML_OP_DIAG_MASK_INF:
-        case GGML_OP_GET_ROWS:
-            {
-                return op->ne[3] == 1;
-            }
-        default:
-            return false;
-    }
-}
-bool ggml_metal_graph_compute(
-        struct ggml_metal_context * ctx,
-               struct ggml_cgraph * gf) {
-    @autoreleasepool {
-
-    // if there is ctx->concur_list, dispatch concurrently
-    // else fallback to serial dispatch
-    MTLComputePassDescriptor * edesc = MTLComputePassDescriptor.computePassDescriptor;
-
-    const bool has_concur = ctx->concur_list_len && ctx->concur_list_len <= GGML_MAX_CONCUR;
-
-    const int n_nodes  = has_concur ? ctx->concur_list_len      : gf->n_nodes;
-    edesc.dispatchType = has_concur ? MTLDispatchTypeConcurrent : MTLDispatchTypeSerial;
-
-    // create multiple command buffers and enqueue them
-    // then, we encode the graph into the command buffers in parallel
-
-    const int n_cb = ctx->n_cb;
-
-    for (int i = 0; i < n_cb; ++i) {
-        ctx->command_buffers[i] = [ctx->queue commandBuffer];
-
-        // enqueue the command buffers in order to specify their execution order
-        [ctx->command_buffers[i] enqueue];
-
-        ctx->command_encoders[i] = [ctx->command_buffers[i] computeCommandEncoderWithDescriptor: edesc];
-    }
-
-    for (int cb_idx = 0; cb_idx < n_cb; ++cb_idx) {
-        const int n_nodes_per_cb = (n_nodes + n_cb - 1) / n_cb;
-
-        dispatch_async(ctx->d_queue, ^{
-            size_t offs_src0 = 0;
-            size_t offs_src1 = 0;
-            size_t offs_dst  = 0;
-
-            id<MTLCommandBuffer> command_buffer  = ctx->command_buffers[cb_idx];
-            id<MTLComputeCommandEncoder> encoder = ctx->command_encoders[cb_idx];
-
-            const int node_start =                                      (cb_idx + 0) * n_nodes_per_cb;
-            const int node_end   = MIN((cb_idx == n_cb - 1) ? n_nodes : (cb_idx + 1) * n_nodes_per_cb, n_nodes);
-
-            for (int ind = node_start; ind < node_end; ++ind) {
-                const int i = has_concur ? ctx->concur_list[ind] : ind;
-
-                if (i == -1) {
-                    [encoder memoryBarrierWithScope:MTLBarrierScopeBuffers];
-                    continue;
-                }
-
-                //GGML_METAL_LOG_INFO("%s: encoding node %3d, op = %8s\n", __func__, i, ggml_op_name(gf->nodes[i]->op));
-
-                struct ggml_tensor * src0 = gf->nodes[i]->src[0];
-                struct ggml_tensor * src1 = gf->nodes[i]->src[1];
-                struct ggml_tensor * dst  = gf->nodes[i];
-
-                switch (dst->op) {
-                    case GGML_OP_NONE:
-                    case GGML_OP_RESHAPE:
-                    case GGML_OP_VIEW:
-                    case GGML_OP_TRANSPOSE:
-                    case GGML_OP_PERMUTE:
-                        {
-                            // noop -> next node
-                        } continue;
-                    default:
-                        {
-                        } break;
-                }
-
-                if (!ggml_metal_supports_op(dst)) {
-                    GGML_METAL_LOG_ERROR("%s: error: unsupported op '%s'\n", __func__, ggml_op_desc(dst));
-                    GGML_ASSERT(!"unsupported op");
-                }
-
-                const int64_t  ne00 = src0 ? src0->ne[0] : 0;
-                const int64_t  ne01 = src0 ? src0->ne[1] : 0;
-                const int64_t  ne02 = src0 ? src0->ne[2] : 0;
-                const int64_t  ne03 = src0 ? src0->ne[3] : 0;
-
-                const uint64_t nb00 = src0 ? src0->nb[0] : 0;
-                const uint64_t nb01 = src0 ? src0->nb[1] : 0;
-                const uint64_t nb02 = src0 ? src0->nb[2] : 0;
-                const uint64_t nb03 = src0 ? src0->nb[3] : 0;
-
-                const int64_t  ne10 = src1 ? src1->ne[0] : 0;
-                const int64_t  ne11 = src1 ? src1->ne[1] : 0;
-                const int64_t  ne12 = src1 ? src1->ne[2] : 0;
-                const int64_t  ne13 = src1 ? src1->ne[3] : 0; UNUSED(ne13);
-
-                const uint64_t nb10 = src1 ? src1->nb[0] : 0;
-                const uint64_t nb11 = src1 ? src1->nb[1] : 0;
-                const uint64_t nb12 = src1 ? src1->nb[2] : 0;
-                const uint64_t nb13 = src1 ? src1->nb[3] : 0; UNUSED(nb13);
-
-                const int64_t  ne0  = dst ? dst->ne[0] : 0;
-                const int64_t  ne1  = dst ? dst->ne[1] : 0;
-                const int64_t  ne2  = dst ? dst->ne[2] : 0;
-                const int64_t  ne3  = dst ? dst->ne[3] : 0;
-
-                const uint64_t nb0  = dst ? dst->nb[0] : 0;
-                const uint64_t nb1  = dst ? dst->nb[1] : 0;
-                const uint64_t nb2  = dst ? dst->nb[2] : 0;
-                const uint64_t nb3  = dst ? dst->nb[3] : 0;
-
-                const enum ggml_type src0t = src0 ? src0->type : GGML_TYPE_COUNT;
-                const enum ggml_type src1t = src1 ? src1->type : GGML_TYPE_COUNT;
-                const enum ggml_type dstt  = dst  ? dst->type  : GGML_TYPE_COUNT;
-
-                id<MTLBuffer> id_src0 = src0 ? ggml_metal_get_buffer(ctx, src0, &offs_src0) : nil;
-                id<MTLBuffer> id_src1 = src1 ? ggml_metal_get_buffer(ctx, src1, &offs_src1) : nil;
-                id<MTLBuffer> id_dst  = dst  ? ggml_metal_get_buffer(ctx, dst,  &offs_dst)  : nil;
-
-                //GGML_METAL_LOG_INFO("%s: op - %s\n", __func__, ggml_op_name(dst->op));
-                //if (src0) {
-                //    GGML_METAL_LOG_INFO("%s: src0 - %4s [%5lld, %5lld, %5lld], %d, %s\n", __func__, ggml_type_name(src0t), ne00, ne01, ne02,
-                //            ggml_is_contiguous(src0), src0->name);
-                //}
-                //if (src1) {
-                //    GGML_METAL_LOG_INFO("%s: src1 - %4s [%5lld, %5lld, %5lld], %d, %s\n", __func__, ggml_type_name(src1t), ne10, ne11, ne12,
-                //            ggml_is_contiguous(src1), src1->name);
-                //}
-                //if (dst) {
-                //    GGML_METAL_LOG_INFO("%s: dst  - %4s [%5lld, %5lld, %5lld], 1, %s\n",  __func__, ggml_type_name(dstt),  ne0,  ne1,  ne2,
-                //            dst->name);
-                //}
-
-                switch (dst->op) {
-                    case GGML_OP_CONCAT:
-                        {
-                            const int64_t nb = ne00;
-
-                            [encoder setComputePipelineState:ctx->pipeline_concat];
-                            [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
-                            [encoder setBuffer:id_src1 offset:offs_src1 atIndex:1];
-                            [encoder setBuffer:id_dst  offset:offs_dst  atIndex:2];
-                            [encoder setBytes:&ne00 length:sizeof(ne00) atIndex:3];
-                            [encoder setBytes:&ne01 length:sizeof(ne01) atIndex:4];
-                            [encoder setBytes:&ne02 length:sizeof(ne02) atIndex:5];
-                            [encoder setBytes:&ne03 length:sizeof(ne03) atIndex:6];
-                            [encoder setBytes:&nb00 length:sizeof(nb00) atIndex:7];
-                            [encoder setBytes:&nb01 length:sizeof(nb01) atIndex:8];
-                            [encoder setBytes:&nb02 length:sizeof(nb02) atIndex:9];
-                            [encoder setBytes:&nb03 length:sizeof(nb03) atIndex:10];
-                            [encoder setBytes:&ne10 length:sizeof(ne10) atIndex:11];
-                            [encoder setBytes:&ne11 length:sizeof(ne11) atIndex:12];
-                            [encoder setBytes:&ne12 length:sizeof(ne12) atIndex:13];
-                            [encoder setBytes:&ne13 length:sizeof(ne13) atIndex:14];
-                            [encoder setBytes:&nb10 length:sizeof(nb10) atIndex:15];
-                            [encoder setBytes:&nb11 length:sizeof(nb11) atIndex:16];
-                            [encoder setBytes:&nb12 length:sizeof(nb12) atIndex:17];
-                            [encoder setBytes:&nb13 length:sizeof(nb13) atIndex:18];
-                            [encoder setBytes:&ne0  length:sizeof(ne0)  atIndex:19];
-                            [encoder setBytes:&ne1  length:sizeof(ne1)  atIndex:20];
-                            [encoder setBytes:&ne2  length:sizeof(ne2)  atIndex:21];
-                            [encoder setBytes:&ne3  length:sizeof(ne3)  atIndex:22];
-                            [encoder setBytes:&nb0  length:sizeof(nb0)  atIndex:23];
-                            [encoder setBytes:&nb1  length:sizeof(nb1)  atIndex:24];
-                            [encoder setBytes:&nb2  length:sizeof(nb2)  atIndex:25];
-                            [encoder setBytes:&nb3  length:sizeof(nb3)  atIndex:26];
-                            [encoder setBytes:&nb   length:sizeof(nb)   atIndex:27];
-
-                            const int nth = MIN(1024, ne0);
-
-                            [encoder dispatchThreadgroups:MTLSizeMake(ne1, ne2, ne3) threadsPerThreadgroup:MTLSizeMake(nth, 1, 1)];
-                        } break;
-                    case GGML_OP_ADD:
-                    case GGML_OP_MUL:
-                    case GGML_OP_DIV:
-                        {
-                            const size_t offs = 0;
-
-                            bool bcast_row = false;
-
-                            int64_t nb = ne00;
-
-                            id<MTLComputePipelineState> pipeline = nil;
-
-                            if (ggml_nelements(src1) == ne10 && ggml_is_contiguous(src1) && ne00 % 4 == 0 && ne10 % 4 == 0) {
-                                GGML_ASSERT(ggml_is_contiguous(src0));
-
-                                // src1 is a row
-                                GGML_ASSERT(ne11 == 1);
-
-                                nb = ne00 / 4;
-                                switch (dst->op) {
-                                    case GGML_OP_ADD: pipeline = ctx->pipeline_add_row; break;
-                                    case GGML_OP_MUL: pipeline = ctx->pipeline_mul_row; break;
-                                    case GGML_OP_DIV: pipeline = ctx->pipeline_div_row; break;
-                                    default: GGML_ASSERT(false);
-                                }
-
-                                bcast_row = true;
-                            } else {
-                                switch (dst->op) {
-                                    case GGML_OP_ADD: pipeline = ctx->pipeline_add; break;
-                                    case GGML_OP_MUL: pipeline = ctx->pipeline_mul; break;
-                                    case GGML_OP_DIV: pipeline = ctx->pipeline_div; break;
-                                    default: GGML_ASSERT(false);
-                                }
-                            }
-
-                            [encoder setComputePipelineState:pipeline];
-                            [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
-                            [encoder setBuffer:id_src1 offset:offs_src1 atIndex:1];
-                            [encoder setBuffer:id_dst  offset:offs_dst  atIndex:2];
-                            [encoder setBytes:&ne00 length:sizeof(ne00) atIndex:3];
-                            [encoder setBytes:&ne01 length:sizeof(ne01) atIndex:4];
-                            [encoder setBytes:&ne02 length:sizeof(ne02) atIndex:5];
-                            [encoder setBytes:&ne03 length:sizeof(ne03) atIndex:6];
-                            [encoder setBytes:&nb00 length:sizeof(nb00) atIndex:7];
-                            [encoder setBytes:&nb01 length:sizeof(nb01) atIndex:8];
-                            [encoder setBytes:&nb02 length:sizeof(nb02) atIndex:9];
-                            [encoder setBytes:&nb03 length:sizeof(nb03) atIndex:10];
-                            [encoder setBytes:&ne10 length:sizeof(ne10) atIndex:11];
-                            [encoder setBytes:&ne11 length:sizeof(ne11) atIndex:12];
-                            [encoder setBytes:&ne12 length:sizeof(ne12) atIndex:13];
-                            [encoder setBytes:&ne13 length:sizeof(ne13) atIndex:14];
-                            [encoder setBytes:&nb10 length:sizeof(nb10) atIndex:15];
-                            [encoder setBytes:&nb11 length:sizeof(nb11) atIndex:16];
-                            [encoder setBytes:&nb12 length:sizeof(nb12) atIndex:17];
-                            [encoder setBytes:&nb13 length:sizeof(nb13) atIndex:18];
-                            [encoder setBytes:&ne0  length:sizeof(ne0)  atIndex:19];
-                            [encoder setBytes:&ne1  length:sizeof(ne1)  atIndex:20];
-                            [encoder setBytes:&ne2  length:sizeof(ne2)  atIndex:21];
-                            [encoder setBytes:&ne3  length:sizeof(ne3)  atIndex:22];
-                            [encoder setBytes:&nb0  length:sizeof(nb0)  atIndex:23];
-                            [encoder setBytes:&nb1  length:sizeof(nb1)  atIndex:24];
-                            [encoder setBytes:&nb2  length:sizeof(nb2)  atIndex:25];
-                            [encoder setBytes:&nb3  length:sizeof(nb3)  atIndex:26];
-                            [encoder setBytes:&offs length:sizeof(offs) atIndex:27];
-                            [encoder setBytes:&nb   length:sizeof(nb)   atIndex:28];
-
-                            if (bcast_row) {
-                                const int64_t n = ggml_nelements(dst)/4;
-
-                                [encoder dispatchThreadgroups:MTLSizeMake(n, 1, 1) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)];
-                            } else {
-                                const int nth = MIN((int) pipeline.maxTotalThreadsPerThreadgroup, ne0);
-
-                                [encoder dispatchThreadgroups:MTLSizeMake(ne01, ne02, ne03) threadsPerThreadgroup:MTLSizeMake(nth, 1, 1)];
-                            }
-                        } break;
-                    case GGML_OP_ACC:
-                        {
-                            GGML_ASSERT(src0t == GGML_TYPE_F32);
-                            GGML_ASSERT(src1t == GGML_TYPE_F32);
-                            GGML_ASSERT(dstt  == GGML_TYPE_F32);
-
-                            GGML_ASSERT(ggml_is_contiguous(src0));
-                            GGML_ASSERT(ggml_is_contiguous(src1));
-
-                            const size_t pnb1 = ((int32_t *) dst->op_params)[0];
-                            const size_t pnb2 = ((int32_t *) dst->op_params)[1];
-                            const size_t pnb3 = ((int32_t *) dst->op_params)[2];
-                            const size_t offs = ((int32_t *) dst->op_params)[3];
-
-                            const bool inplace = (bool) ((int32_t *) dst->op_params)[4];
-
-                            if (!inplace) {
-                                // run a separete kernel to cpy src->dst
-                                // not sure how to avoid this
-                                // TODO: make a simpler cpy_bytes kernel
-
-                                const int nth = MIN((int) ctx->pipeline_cpy_f32_f32.maxTotalThreadsPerThreadgroup, ne00);
-
-                                [encoder setComputePipelineState:ctx->pipeline_cpy_f32_f32];
-                                [encoder setBuffer:id_src0 offset:offs_src0        atIndex:0];
-                                [encoder setBuffer:id_dst  offset:offs_dst         atIndex:1];
-                                [encoder setBytes:&ne00    length:sizeof( int64_t) atIndex:2];
-                                [encoder setBytes:&ne01    length:sizeof( int64_t) atIndex:3];
-                                [encoder setBytes:&ne02    length:sizeof( int64_t) atIndex:4];
-                                [encoder setBytes:&ne03    length:sizeof( int64_t) atIndex:5];
-                                [encoder setBytes:&nb00    length:sizeof(uint64_t) atIndex:6];
-                                [encoder setBytes:&nb01    length:sizeof(uint64_t) atIndex:7];
-                                [encoder setBytes:&nb02    length:sizeof(uint64_t) atIndex:8];
-                                [encoder setBytes:&nb03    length:sizeof(uint64_t) atIndex:9];
-                                [encoder setBytes:&ne0     length:sizeof( int64_t) atIndex:10];
-                                [encoder setBytes:&ne1     length:sizeof( int64_t) atIndex:11];
-                                [encoder setBytes:&ne2     length:sizeof( int64_t) atIndex:12];
-                                [encoder setBytes:&ne3     length:sizeof( int64_t) atIndex:13];
-                                [encoder setBytes:&nb0     length:sizeof(uint64_t) atIndex:14];
-                                [encoder setBytes:&nb1     length:sizeof(uint64_t) atIndex:15];
-                                [encoder setBytes:&nb2     length:sizeof(uint64_t) atIndex:16];
-                                [encoder setBytes:&nb3     length:sizeof(uint64_t) atIndex:17];
-
-                                [encoder dispatchThreadgroups:MTLSizeMake(ne01, ne02, ne03) threadsPerThreadgroup:MTLSizeMake(nth, 1, 1)];
-                            }
-
-                            [encoder setComputePipelineState:ctx->pipeline_add];
-                            [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
-                            [encoder setBuffer:id_src1 offset:offs_src1 atIndex:1];
-                            [encoder setBuffer:id_dst  offset:offs_dst  atIndex:2];
-                            [encoder setBytes:&ne00 length:sizeof(ne00) atIndex:3];
-                            [encoder setBytes:&ne01 length:sizeof(ne01) atIndex:4];
-                            [encoder setBytes:&ne02 length:sizeof(ne02) atIndex:5];
-                            [encoder setBytes:&ne03 length:sizeof(ne03) atIndex:6];
-                            [encoder setBytes:&nb00 length:sizeof(nb00) atIndex:7];
-                            [encoder setBytes:&pnb1 length:sizeof(pnb1) atIndex:8];
-                            [encoder setBytes:&pnb2 length:sizeof(pnb2) atIndex:9];
-                            [encoder setBytes:&pnb3 length:sizeof(pnb3) atIndex:10];
-                            [encoder setBytes:&ne10 length:sizeof(ne10) atIndex:11];
-                            [encoder setBytes:&ne11 length:sizeof(ne11) atIndex:12];
-                            [encoder setBytes:&ne12 length:sizeof(ne12) atIndex:13];
-                            [encoder setBytes:&ne13 length:sizeof(ne13) atIndex:14];
-                            [encoder setBytes:&nb10 length:sizeof(nb10) atIndex:15];
-                            [encoder setBytes:&nb11 length:sizeof(nb11) atIndex:16];
-                            [encoder setBytes:&nb12 length:sizeof(nb12) atIndex:17];
-                            [encoder setBytes:&nb13 length:sizeof(nb13) atIndex:18];
-                            [encoder setBytes:&ne0  length:sizeof(ne0)  atIndex:19];
-                            [encoder setBytes:&ne1  length:sizeof(ne1)  atIndex:20];
-                            [encoder setBytes:&ne2  length:sizeof(ne2)  atIndex:21];
-                            [encoder setBytes:&ne3  length:sizeof(ne3)  atIndex:22];
-                            [encoder setBytes:&nb0  length:sizeof(nb0)  atIndex:23];
-                            [encoder setBytes:&pnb1 length:sizeof(pnb1) atIndex:24];
-                            [encoder setBytes:&pnb2 length:sizeof(pnb2) atIndex:25];
-                            [encoder setBytes:&pnb3 length:sizeof(pnb3) atIndex:26];
-                            [encoder setBytes:&offs length:sizeof(offs) atIndex:27];
-
-                            const int nth = MIN((int) ctx->pipeline_add.maxTotalThreadsPerThreadgroup, ne00);
-
-                            [encoder dispatchThreadgroups:MTLSizeMake(ne11, ne12, ne13) threadsPerThreadgroup:MTLSizeMake(nth, 1, 1)];
-                        } break;
-                    case GGML_OP_SCALE:
-                        {
-                            GGML_ASSERT(ggml_is_contiguous(src0));
-
-                            const float scale = *(const float *) dst->op_params;
-
-                            int64_t n = ggml_nelements(dst);
-
-                            if (n % 4 == 0) {
-                                n /= 4;
-                                [encoder setComputePipelineState:ctx->pipeline_scale_4];
-                            } else {
-                                [encoder setComputePipelineState:ctx->pipeline_scale];
-                            }
-
-                            [encoder setBuffer:id_src0   offset:offs_src0 atIndex:0];
-                            [encoder setBuffer:id_dst    offset:offs_dst  atIndex:1];
-                            [encoder setBytes:&scale length:sizeof(scale) atIndex:2];
-
-                            [encoder dispatchThreadgroups:MTLSizeMake(n, 1, 1) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)];
-                        } break;
-                    case GGML_OP_UNARY:
-                        switch (ggml_get_unary_op(gf->nodes[i])) {
-                            case GGML_UNARY_OP_TANH:
-                                {
-                                    [encoder setComputePipelineState:ctx->pipeline_tanh];
-                                    [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
-                                    [encoder setBuffer:id_dst  offset:offs_dst  atIndex:1];
-
-                                    const int64_t n = ggml_nelements(dst);
-
-                                    [encoder dispatchThreadgroups:MTLSizeMake(n, 1, 1) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)];
-                                } break;
-                            case GGML_UNARY_OP_RELU:
-                                {
-                                    [encoder setComputePipelineState:ctx->pipeline_relu];
-                                    [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
-                                    [encoder setBuffer:id_dst  offset:offs_dst  atIndex:1];
-
-                                    const int64_t n = ggml_nelements(dst);
-
-                                    [encoder dispatchThreadgroups:MTLSizeMake(n, 1, 1) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)];
-                                } break;
-                            case GGML_UNARY_OP_GELU:
-                                {
-                                    [encoder setComputePipelineState:ctx->pipeline_gelu];
-                                    [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
-                                    [encoder setBuffer:id_dst  offset:offs_dst  atIndex:1];
-
-                                    const int64_t n = ggml_nelements(dst);
-                                    GGML_ASSERT(n % 4 == 0);
-
-                                    [encoder dispatchThreadgroups:MTLSizeMake(n/4, 1, 1) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)];
-                                } break;
-                            case GGML_UNARY_OP_GELU_QUICK:
-                                {
-                                    [encoder setComputePipelineState:ctx->pipeline_gelu_quick];
-                                    [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
-                                    [encoder setBuffer:id_dst  offset:offs_dst  atIndex:1];
-
-                                    const int64_t n = ggml_nelements(dst);
-                                    GGML_ASSERT(n % 4 == 0);
-
-                                    [encoder dispatchThreadgroups:MTLSizeMake(n/4, 1, 1) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)];
-                                } break;
-                            case GGML_UNARY_OP_SILU:
-                                {
-                                    [encoder setComputePipelineState:ctx->pipeline_silu];
-                                    [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
-                                    [encoder setBuffer:id_dst  offset:offs_dst  atIndex:1];
-
-                                    const int64_t n = ggml_nelements(dst);
-                                    GGML_ASSERT(n % 4 == 0);
-
-                                    [encoder dispatchThreadgroups:MTLSizeMake(n/4, 1, 1) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)];
-                                } break;
-                            default:
-                                {
-                                    GGML_METAL_LOG_WARN("%s: node %3d, op = %8s not implemented\n", __func__, i, ggml_op_name(dst->op));
-                                    GGML_ASSERT(false);
-                                }
-                        } break;
-                    case GGML_OP_SQR:
-                        {
-                            GGML_ASSERT(ggml_is_contiguous(src0));
-
-                            [encoder setComputePipelineState:ctx->pipeline_sqr];
-                            [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
-                            [encoder setBuffer:id_dst  offset:offs_dst atIndex:1];
-
-                            const int64_t n = ggml_nelements(dst);
-                            [encoder dispatchThreadgroups:MTLSizeMake(n, 1, 1) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)];
-                        } break;
-                    case GGML_OP_SUM_ROWS:
-                        {
-                            GGML_ASSERT(src0->nb[0] == ggml_type_size(src0->type));
-
-                            [encoder setComputePipelineState:ctx->pipeline_sum_rows];
-                            [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
-                            [encoder setBuffer:id_dst  offset:offs_dst  atIndex:1];
-                            [encoder setBytes:&ne00 length:sizeof(ne00) atIndex:2];
-                            [encoder setBytes:&ne01 length:sizeof(ne01) atIndex:3];
-                            [encoder setBytes:&ne02 length:sizeof(ne02) atIndex:4];
-                            [encoder setBytes:&ne03 length:sizeof(ne03) atIndex:5];
-                            [encoder setBytes:&nb00 length:sizeof(nb00) atIndex:6];
-                            [encoder setBytes:&nb01 length:sizeof(nb01) atIndex:7];
-                            [encoder setBytes:&nb02 length:sizeof(nb02) atIndex:8];
-                            [encoder setBytes:&nb03 length:sizeof(nb03) atIndex:9];
-                            [encoder setBytes:&ne10 length:sizeof(ne10) atIndex:10];
-                            [encoder setBytes:&ne11 length:sizeof(ne11) atIndex:11];
-                            [encoder setBytes:&ne12 length:sizeof(ne12) atIndex:12];
-                            [encoder setBytes:&ne13 length:sizeof(ne13) atIndex:13];
-                            [encoder setBytes:&nb10 length:sizeof(nb10) atIndex:14];
-                            [encoder setBytes:&nb11 length:sizeof(nb11) atIndex:15];
-                            [encoder setBytes:&nb12 length:sizeof(nb12) atIndex:16];
-                            [encoder setBytes:&nb13 length:sizeof(nb13) atIndex:17];
-                            [encoder setBytes:&ne0  length:sizeof(ne0)  atIndex:18];
-                            [encoder setBytes:&ne1  length:sizeof(ne1)  atIndex:19];
-                            [encoder setBytes:&ne2  length:sizeof(ne2)  atIndex:20];
-                            [encoder setBytes:&ne3  length:sizeof(ne3)  atIndex:21];
-                            [encoder setBytes:&nb0  length:sizeof(nb0)  atIndex:22];
-                            [encoder setBytes:&nb1  length:sizeof(nb1)  atIndex:23];
-                            [encoder setBytes:&nb2  length:sizeof(nb2)  atIndex:24];
-                            [encoder setBytes:&nb3  length:sizeof(nb3)  atIndex:25];
-
-                            [encoder dispatchThreadgroups:MTLSizeMake(ne01, ne02, ne03) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)];
-                        } break;
-                    case GGML_OP_SOFT_MAX:
-                        {
-                            int nth = 32; // SIMD width
-
-                            if (ne00%4 == 0) {
-                                while (nth < ne00/4 && nth < 256) {
-                                    nth *= 2;
-                                }
-                                [encoder setComputePipelineState:ctx->pipeline_soft_max_4];
-                            } else {
-                                while (nth < ne00 && nth < 1024) {
-                                    nth *= 2;
-                                }
-                                [encoder setComputePipelineState:ctx->pipeline_soft_max];
-                            }
-
-                            const float scale = ((float *) dst->op_params)[0];
-
-                            [encoder setBuffer:id_src0 offset:offs_src0   atIndex:0];
-                            if (id_src1) {
-                                [encoder setBuffer:id_src1 offset:offs_src1   atIndex:1];
-                            } else {
-                                [encoder setBuffer:id_src0 offset:offs_src0   atIndex:1];
-                            }
-                            [encoder setBuffer:id_dst  offset:offs_dst    atIndex:2];
-                            [encoder setBytes:&ne00  length:sizeof(ne00)  atIndex:3];
-                            [encoder setBytes:&ne01  length:sizeof(ne01)  atIndex:4];
-                            [encoder setBytes:&ne02  length:sizeof(ne02)  atIndex:5];
-                            [encoder setBytes:&scale length:sizeof(scale) atIndex:6];
-                            [encoder setThreadgroupMemoryLength:32*sizeof(float) atIndex:0];
-
-                            [encoder dispatchThreadgroups:MTLSizeMake(ne01*ne02*ne03, 1, 1) threadsPerThreadgroup:MTLSizeMake(nth, 1, 1)];
-                        } break;
-                    case GGML_OP_DIAG_MASK_INF:
-                        {
-                            const int n_past = ((int32_t *)(dst->op_params))[0];
-
-                            if (ne00%8 == 0) {
-                                [encoder setComputePipelineState:ctx->pipeline_diag_mask_inf_8];
-                            } else {
-                                [encoder setComputePipelineState:ctx->pipeline_diag_mask_inf];
-                            }
-                            [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
-                            [encoder setBuffer:id_dst  offset:offs_dst  atIndex:1];
-                            [encoder setBytes:&ne00   length:sizeof(ne00) atIndex:2];
-                            [encoder setBytes:&ne01   length:sizeof(ne01) atIndex:3];
-                            [encoder setBytes:&n_past length:sizeof(int)  atIndex:4];
-
-                            if (ne00%8 == 0) {
-                                [encoder dispatchThreadgroups:MTLSizeMake(ne00*ne01*ne02/8, 1, 1) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)];
-                            }
-                            else {
-                                [encoder dispatchThreadgroups:MTLSizeMake(ne00, ne01, ne02) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)];
-                            }
-                        } break;
-                    case GGML_OP_MUL_MAT:
-                        {
-                            GGML_ASSERT(ne00 == ne10);
-
-                            // TODO: assert that dim2 and dim3 are contiguous
-                            GGML_ASSERT(ne12 % ne02 == 0);
-                            GGML_ASSERT(ne13 % ne03 == 0);
-
-                            const uint r2 = ne12/ne02;
-                            const uint r3 = ne13/ne03;
-
-                            // find the break-even point where the matrix-matrix kernel becomes more efficient compared
-                            // to the matrix-vector kernel
-                            int ne11_mm_min = 1;
-
-#if 0
-                            // the numbers below are measured on M2 Ultra for 7B and 13B models
-                            // these numbers do not translate to other devices or model sizes
-                            // TODO: need to find a better approach
-                            if ([ctx->device.name isEqualToString:@"Apple M2 Ultra"]) {
-                                switch (src0t) {
-                                    case GGML_TYPE_F16:  ne11_mm_min = 2;  break;
-                                    case GGML_TYPE_Q8_0: ne11_mm_min = 7;  break;
-                                    case GGML_TYPE_Q2_K: ne11_mm_min = 15; break;
-                                    case GGML_TYPE_Q3_K: ne11_mm_min = 7;  break;
-                                    case GGML_TYPE_Q4_0:
-                                    case GGML_TYPE_Q4_1: ne11_mm_min = 15; break;
-                                    case GGML_TYPE_Q4_K: ne11_mm_min = 11; break;
-                                    case GGML_TYPE_Q5_0:                          // not tested yet
-                                    case GGML_TYPE_Q5_1: ne11_mm_min = 13; break; // not tested yet
-                                    case GGML_TYPE_Q5_K: ne11_mm_min = 7;  break;
-                                    case GGML_TYPE_Q6_K: ne11_mm_min = 7;  break;
-                                    default:             ne11_mm_min = 1;  break;
-                                }
-                            }
-#endif
-
-                            // for now the matrix-matrix multiplication kernel only works on A14+/M1+ SoCs
-                            // AMD GPU and older A-chips will reuse matrix-vector multiplication kernel
-                            if ([ctx->device supportsFamily:MTLGPUFamilyApple7] &&
-                                !ggml_is_transposed(src0) &&
-                                !ggml_is_transposed(src1) &&
-                                src1t == GGML_TYPE_F32 &&
-                                ne00 % 32 == 0 && ne00 >= 64 &&
-                                (ne11 > ne11_mm_min || (ggml_is_quantized(src0t) && ne12 > 1))) {
-                                //printf("matrix: ne00 = %6d, ne01 = %6d, ne02 = %6d, ne11 = %6d, ne12 = %6d\n", ne00, ne01, ne02, ne11, ne12);
-                                switch (src0->type) {
-                                    case GGML_TYPE_F32:  [encoder setComputePipelineState:ctx->pipeline_mul_mm_f32_f32];  break;
-                                    case GGML_TYPE_F16:  [encoder setComputePipelineState:ctx->pipeline_mul_mm_f16_f32];  break;
-                                    case GGML_TYPE_Q4_0: [encoder setComputePipelineState:ctx->pipeline_mul_mm_q4_0_f32]; break;
-                                    case GGML_TYPE_Q4_1: [encoder setComputePipelineState:ctx->pipeline_mul_mm_q4_1_f32]; break;
-                                    case GGML_TYPE_Q5_0: [encoder setComputePipelineState:ctx->pipeline_mul_mm_q5_0_f32]; break;
-                                    case GGML_TYPE_Q5_1: [encoder setComputePipelineState:ctx->pipeline_mul_mm_q5_1_f32]; break;
-                                    case GGML_TYPE_Q8_0: [encoder setComputePipelineState:ctx->pipeline_mul_mm_q8_0_f32]; break;
-                                    case GGML_TYPE_Q2_K: [encoder setComputePipelineState:ctx->pipeline_mul_mm_q2_K_f32]; break;
-                                    case GGML_TYPE_Q3_K: [encoder setComputePipelineState:ctx->pipeline_mul_mm_q3_K_f32]; break;
-                                    case GGML_TYPE_Q4_K: [encoder setComputePipelineState:ctx->pipeline_mul_mm_q4_K_f32]; break;
-                                    case GGML_TYPE_Q5_K: [encoder setComputePipelineState:ctx->pipeline_mul_mm_q5_K_f32]; break;
-                                    case GGML_TYPE_Q6_K: [encoder setComputePipelineState:ctx->pipeline_mul_mm_q6_K_f32]; break;
-                                    default: GGML_ASSERT(false && "MUL MAT-MAT not implemented");
-                                }
-                                [encoder setBuffer:id_src0 offset:offs_src0    atIndex:0];
-                                [encoder setBuffer:id_src1 offset:offs_src1    atIndex:1];
-                                [encoder setBuffer:id_dst  offset:offs_dst     atIndex:2];
-                                [encoder setBytes:&ne00    length:sizeof(ne00) atIndex:3];
-                                [encoder setBytes:&ne02    length:sizeof(ne02) atIndex:4];
-                                [encoder setBytes:&nb01    length:sizeof(nb01) atIndex:5];
-                                [encoder setBytes:&nb02    length:sizeof(nb02) atIndex:6];
-                                [encoder setBytes:&ne12    length:sizeof(ne12) atIndex:7];
-                                [encoder setBytes:&nb10    length:sizeof(nb10) atIndex:8];
-                                [encoder setBytes:&nb11    length:sizeof(nb11) atIndex:9];
-                                [encoder setBytes:&nb12    length:sizeof(nb12) atIndex:10];
-                                [encoder setBytes:&ne0     length:sizeof(ne0)  atIndex:11];
-                                [encoder setBytes:&ne1     length:sizeof(ne1)  atIndex:12];
-                                [encoder setBytes:&r2      length:sizeof(r2)   atIndex:13];
-                                [encoder setBytes:&r3      length:sizeof(r3)   atIndex:14];
-                                [encoder setThreadgroupMemoryLength:8192 atIndex:0];
-                                [encoder dispatchThreadgroups:MTLSizeMake( (ne11 + 31)/32, (ne01 + 63)/64, ne12*ne13) threadsPerThreadgroup:MTLSizeMake(128, 1, 1)];
-                            } else {
-                                int nth0 = 32;
-                                int nth1 = 1;
-                                int nrows = 1;
-                                //printf("vector: ne00 = %6d, ne01 = %6d, ne02 = %6d, ne11 = %6d, ne12 = %6d\n", ne00, ne01, ne02, ne11, ne12);
-
-                                // use custom matrix x vector kernel
-                                switch (src0t) {
-                                    case GGML_TYPE_F32:
-                                        {
-                                            GGML_ASSERT(src1t == GGML_TYPE_F32);
-                                            [encoder setComputePipelineState:ctx->pipeline_mul_mv_f32_f32];
-                                            nrows = 4;
-                                        } break;
-                                    case GGML_TYPE_F16:
-                                        {
-                                            nth0 = 32;
-                                            nth1 = 1;
-                                            if (src1t == GGML_TYPE_F32) {
-                                                if (ne11 * ne12 < 4) {
-                                                    [encoder setComputePipelineState:ctx->pipeline_mul_mv_f16_f32_1row];
-                                                } else if (ne00 >= 128 && ne01 >= 8 && ne00%4 == 0) {
-                                                    [encoder setComputePipelineState:ctx->pipeline_mul_mv_f16_f32_l4];
-                                                    nrows = ne11;
-                                                } else {
-                                                    [encoder setComputePipelineState:ctx->pipeline_mul_mv_f16_f32];
-                                                    nrows = 4;
-                                                }
-                                            } else {
-                                                [encoder setComputePipelineState:ctx->pipeline_mul_mv_f16_f16];
-                                                nrows = 4;
-                                            }
-                                        } break;
-                                    case GGML_TYPE_Q4_0:
-                                        {
-                                            nth0 = 8;
-                                            nth1 = 8;
-                                            [encoder setComputePipelineState:ctx->pipeline_mul_mv_q4_0_f32];
-                                        } break;
-                                    case GGML_TYPE_Q4_1:
-                                        {
-                                            nth0 = 8;
-                                            nth1 = 8;
-                                            [encoder setComputePipelineState:ctx->pipeline_mul_mv_q4_1_f32];
-                                        } break;
-                                    case GGML_TYPE_Q5_0:
-                                        {
-                                            nth0 = 8;
-                                            nth1 = 8;
-                                            [encoder setComputePipelineState:ctx->pipeline_mul_mv_q5_0_f32];
-                                        } break;
-                                    case GGML_TYPE_Q5_1:
-                                        {
-                                            nth0 = 8;
-                                            nth1 = 8;
-                                            [encoder setComputePipelineState:ctx->pipeline_mul_mv_q5_1_f32];
-                                        } break;
-                                    case GGML_TYPE_Q8_0:
-                                        {
-                                            nth0 = 8;
-                                            nth1 = 8;
-                                            [encoder setComputePipelineState:ctx->pipeline_mul_mv_q8_0_f32];
-                                        } break;
-                                    case GGML_TYPE_Q2_K:
-                                        {
-                                            nth0 = 2;
-                                            nth1 = 32;
-                                            [encoder setComputePipelineState:ctx->pipeline_mul_mv_q2_K_f32];
-                                        } break;
-                                    case GGML_TYPE_Q3_K:
-                                        {
-                                            nth0 = 2;
-                                            nth1 = 32;
-                                            [encoder setComputePipelineState:ctx->pipeline_mul_mv_q3_K_f32];
-                                        } break;
-                                    case GGML_TYPE_Q4_K:
-                                        {
-                                            nth0 = 4; //1;
-                                            nth1 = 8; //32;
-                                            [encoder setComputePipelineState:ctx->pipeline_mul_mv_q4_K_f32];
-                                        } break;
-                                    case GGML_TYPE_Q5_K:
-                                        {
-                                            nth0 = 2;
-                                            nth1 = 32;
-                                            [encoder setComputePipelineState:ctx->pipeline_mul_mv_q5_K_f32];
-                                        } break;
-                                    case GGML_TYPE_Q6_K:
-                                        {
-                                            nth0 = 2;
-                                            nth1 = 32;
-                                            [encoder setComputePipelineState:ctx->pipeline_mul_mv_q6_K_f32];
-                                        } break;
-                                    default:
-                                        {
-                                            GGML_METAL_LOG_ERROR("Asserting on type %d\n", (int)src0t);
-                                            GGML_ASSERT(false && "not implemented");
-                                        }
-                                };
-
-                                if (ggml_is_quantized(src0t)) {
-                                    GGML_ASSERT(ne00 >= nth0*nth1);
-                                }
-
-                                [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
-                                [encoder setBuffer:id_src1 offset:offs_src1 atIndex:1];
-                                [encoder setBuffer:id_dst  offset:offs_dst  atIndex:2];
-                                [encoder setBytes:&ne00 length:sizeof(ne00) atIndex:3];
-                                [encoder setBytes:&ne01 length:sizeof(ne01) atIndex:4];
-                                [encoder setBytes:&ne02 length:sizeof(ne02) atIndex:5];
-                                [encoder setBytes:&nb00 length:sizeof(nb00) atIndex:6];
-                                [encoder setBytes:&nb01 length:sizeof(nb01) atIndex:7];
-                                [encoder setBytes:&nb02 length:sizeof(nb02) atIndex:8];
-                                [encoder setBytes:&ne10 length:sizeof(ne10) atIndex:9];
-                                [encoder setBytes:&ne11 length:sizeof(ne11) atIndex:10];
-                                [encoder setBytes:&ne12 length:sizeof(ne12) atIndex:11];
-                                [encoder setBytes:&nb10 length:sizeof(nb10) atIndex:12];
-                                [encoder setBytes:&nb11 length:sizeof(nb11) atIndex:13];
-                                [encoder setBytes:&nb12 length:sizeof(nb12) atIndex:14];
-                                [encoder setBytes:&ne0  length:sizeof(ne0)  atIndex:15];
-                                [encoder setBytes:&ne1  length:sizeof(ne1)  atIndex:16];
-                                [encoder setBytes:&r2   length:sizeof(r2)   atIndex:17];
-                                [encoder setBytes:&r3   length:sizeof(r3)   atIndex:18];
-
-                                if (src0t == GGML_TYPE_Q4_0 || src0t == GGML_TYPE_Q4_1 ||
-                                    src0t == GGML_TYPE_Q5_0 || src0t == GGML_TYPE_Q5_1 || src0t == GGML_TYPE_Q8_0 ||
-                                    src0t == GGML_TYPE_Q2_K) { // || src0t == GGML_TYPE_Q4_K) {
-                                    [encoder dispatchThreadgroups:MTLSizeMake((ne01 + 7)/8, ne11, ne12*ne13) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
-                                }
-                                else if (src0t == GGML_TYPE_Q4_K) {
-                                    [encoder dispatchThreadgroups:MTLSizeMake((ne01 + 3)/4, ne11, ne12*ne13) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
-                                }
-                                else if (src0t == GGML_TYPE_Q3_K) {
-#ifdef GGML_QKK_64
-                                    [encoder dispatchThreadgroups:MTLSizeMake((ne01 + 1)/2, ne11, ne12*ne13) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
-#else
-                                    [encoder dispatchThreadgroups:MTLSizeMake((ne01 + 3)/4, ne11, ne12*ne13) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
-#endif
-                                }
-                                else if (src0t == GGML_TYPE_Q5_K) {
-                                    [encoder dispatchThreadgroups:MTLSizeMake((ne01 + 3)/4, ne11, ne12*ne13) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
-                                }
-                                else if (src0t == GGML_TYPE_Q6_K) {
-                                    [encoder dispatchThreadgroups:MTLSizeMake((ne01 + 1)/2, ne11, ne12*ne13) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
-                                } else {
-                                    const int64_t ny = (ne11 + nrows - 1)/nrows;
-                                    [encoder dispatchThreadgroups:MTLSizeMake(ne01, ny, ne12*ne13) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
-                                }
-                            }
-                        } break;
-                    case GGML_OP_MUL_MAT_ID:
-                        {
-                            //GGML_ASSERT(ne00 == ne10);
-                            //GGML_ASSERT(ne03 == ne13);
-
-                            GGML_ASSERT(src0t == GGML_TYPE_I32);
-
-                            const int n_as = ((int32_t *) dst->op_params)[1];
-
-                            // TODO: make this more general
-                            GGML_ASSERT(n_as <= 8);
-
-                            // max size of the src1ids array in the kernel stack
-                            GGML_ASSERT(ne11 <= 512);
-
-                            struct ggml_tensor * src2 = gf->nodes[i]->src[2];
-
-                            const int64_t  ne20 = src2 ? src2->ne[0] : 0;
-                            const int64_t  ne21 = src2 ? src2->ne[1] : 0;
-                            const int64_t  ne22 = src2 ? src2->ne[2] : 0;
-                            const int64_t  ne23 = src2 ? src2->ne[3] : 0; GGML_UNUSED(ne23);
-
-                            const uint64_t nb20 = src2 ? src2->nb[0] : 0; GGML_UNUSED(nb20);
-                            const uint64_t nb21 = src2 ? src2->nb[1] : 0;
-                            const uint64_t nb22 = src2 ? src2->nb[2] : 0;
-                            const uint64_t nb23 = src2 ? src2->nb[3] : 0; GGML_UNUSED(nb23);
-
-                            const enum ggml_type src2t = src2 ? src2->type : GGML_TYPE_COUNT; GGML_UNUSED(src2t);
-
-                            GGML_ASSERT(!ggml_is_transposed(src2));
-                            GGML_ASSERT(!ggml_is_transposed(src1));
-
-                            GGML_ASSERT(src1t == GGML_TYPE_F32);
-
-                            const uint r2 = ne12/ne22;
-                            const uint r3 = ne13/ne23;
-
-                            // find the break-even point where the matrix-matrix kernel becomes more efficient compared
-                            // to the matrix-vector kernel
-                            int ne11_mm_min = n_as;
-
-                            const int idx = ((int32_t *) dst->op_params)[0];
-
-                            // batch size
-                            GGML_ASSERT(ne01 == ne11);
-
-                            // for now the matrix-matrix multiplication kernel only works on A14+/M1+ SoCs
-                            // AMD GPU and older A-chips will reuse matrix-vector multiplication kernel
-                            // !!!
-                            // TODO: for now, always use mat-vec kernels until we figure out how to improve the
-                            //       indirect matrix multiplication
-                            // !!!
-                            if ([ctx->device supportsFamily:MTLGPUFamilyApple7] &&
-                                ne20 % 32 == 0 && ne20 >= 64 &&
-                                ne11 > ne11_mm_min) {
-                                switch (src2->type) {
-                                    case GGML_TYPE_F32:  [encoder setComputePipelineState:ctx->pipeline_mul_mm_id_f32_f32];  break;
-                                    case GGML_TYPE_F16:  [encoder setComputePipelineState:ctx->pipeline_mul_mm_id_f16_f32];  break;
-                                    case GGML_TYPE_Q4_0: [encoder setComputePipelineState:ctx->pipeline_mul_mm_id_q4_0_f32]; break;
-                                    case GGML_TYPE_Q4_1: [encoder setComputePipelineState:ctx->pipeline_mul_mm_id_q4_1_f32]; break;
-                                    case GGML_TYPE_Q5_0: [encoder setComputePipelineState:ctx->pipeline_mul_mm_id_q5_0_f32]; break;
-                                    case GGML_TYPE_Q5_1: [encoder setComputePipelineState:ctx->pipeline_mul_mm_id_q5_1_f32]; break;
-                                    case GGML_TYPE_Q8_0: [encoder setComputePipelineState:ctx->pipeline_mul_mm_id_q8_0_f32]; break;
-                                    case GGML_TYPE_Q2_K: [encoder setComputePipelineState:ctx->pipeline_mul_mm_id_q2_K_f32]; break;
-                                    case GGML_TYPE_Q3_K: [encoder setComputePipelineState:ctx->pipeline_mul_mm_id_q3_K_f32]; break;
-                                    case GGML_TYPE_Q4_K: [encoder setComputePipelineState:ctx->pipeline_mul_mm_id_q4_K_f32]; break;
-                                    case GGML_TYPE_Q5_K: [encoder setComputePipelineState:ctx->pipeline_mul_mm_id_q5_K_f32]; break;
-                                    case GGML_TYPE_Q6_K: [encoder setComputePipelineState:ctx->pipeline_mul_mm_id_q6_K_f32]; break;
-                                    default: GGML_ASSERT(false && "MUL_MAT_ID not implemented");
-                                }
-                                [encoder setBuffer:id_src0 offset:offs_src0    atIndex:0];
-                                [encoder setBuffer:id_src1 offset:offs_src1    atIndex:1];
-                                [encoder setBuffer:id_dst  offset:offs_dst     atIndex:2];
-                                [encoder setBytes:&nb01    length:sizeof(nb01) atIndex:3];
-                                [encoder setBytes:&ne20    length:sizeof(ne20) atIndex:4];
-                                [encoder setBytes:&ne22    length:sizeof(ne22) atIndex:5];
-                                [encoder setBytes:&nb21    length:sizeof(nb21) atIndex:6];
-                                [encoder setBytes:&nb22    length:sizeof(nb22) atIndex:7];
-                                [encoder setBytes:&ne12    length:sizeof(ne12) atIndex:8];
-                                [encoder setBytes:&ne13    length:sizeof(ne13) atIndex:9];
-                                [encoder setBytes:&nb10    length:sizeof(nb10) atIndex:10];
-                                [encoder setBytes:&nb11    length:sizeof(nb11) atIndex:11];
-                                [encoder setBytes:&nb12    length:sizeof(nb12) atIndex:12];
-                                [encoder setBytes:&ne0     length:sizeof(ne0)  atIndex:13];
-                                [encoder setBytes:&ne1     length:sizeof(ne1)  atIndex:14];
-                                [encoder setBytes:&nb1     length:sizeof(nb1)  atIndex:15];
-                                [encoder setBytes:&r2      length:sizeof(r2)   atIndex:16];
-                                [encoder setBytes:&r3      length:sizeof(r3)   atIndex:17];
-                                [encoder setBytes:&idx     length:sizeof(idx)  atIndex:18];
-                                // TODO: how to make this an array? read Metal docs
-                                for (int j = 0; j < 8; ++j) {
-                                    // NOTE: this is done like this to avoid uninitialized kernel arguments when n_as < 8
-                                    struct ggml_tensor * src_cur = dst->src[2 + (j % n_as)];
-
-                                    size_t offs_src_cur = 0;
-                                    id<MTLBuffer> id_src_cur = ggml_metal_get_buffer(ctx, src_cur, &offs_src_cur);
-
-                                    [encoder setBuffer:id_src_cur offset:offs_src_cur atIndex:19 + j];
-                                }
-
-                                [encoder setThreadgroupMemoryLength:8192 atIndex:0];
-
-                                [encoder dispatchThreadgroups:MTLSizeMake((ne11 + 31)/32, (ne21 + 63)/64, n_as*ne12*ne13) threadsPerThreadgroup:MTLSizeMake(128, 1, 1)];
-                            } else {
-                                int nth0 = 32;
-                                int nth1 = 1;
-                                int nrows = 1;
-                                //printf("vector: ne00 = %6d, ne01 = %6d, ne02 = %6d, ne11 = %6d, ne12 = %6d\n", ne00, ne01, ne02, ne11, ne12);
-
-                                // use custom matrix x vector kernel
-                                switch (src2t) {
-                                    case GGML_TYPE_F32:
-                                        {
-                                            GGML_ASSERT(src1t == GGML_TYPE_F32);
-                                            [encoder setComputePipelineState:ctx->pipeline_mul_mv_id_f32_f32];
-                                        } break;
-                                    case GGML_TYPE_F16:
-                                        {
-                                            GGML_ASSERT(src1t == GGML_TYPE_F32);
-                                            nth0 = 32;
-                                            nth1 = 1;
-                                            [encoder setComputePipelineState:ctx->pipeline_mul_mv_id_f16_f32];
-                                        } break;
-                                    case GGML_TYPE_Q4_0:
-                                        {
-                                            nth0 = 8;
-                                            nth1 = 8;
-                                            [encoder setComputePipelineState:ctx->pipeline_mul_mv_id_q4_0_f32];
-                                        } break;
-                                    case GGML_TYPE_Q4_1:
-                                        {
-                                            nth0 = 8;
-                                            nth1 = 8;
-                                            [encoder setComputePipelineState:ctx->pipeline_mul_mv_id_q4_1_f32];
-                                        } break;
-                                    case GGML_TYPE_Q5_0:
-                                        {
-                                            nth0 = 8;
-                                            nth1 = 8;
-                                            [encoder setComputePipelineState:ctx->pipeline_mul_mv_id_q5_0_f32];
-                                        } break;
-                                    case GGML_TYPE_Q5_1:
-                                        {
-                                            nth0 = 8;
-                                            nth1 = 8;
-                                            [encoder setComputePipelineState:ctx->pipeline_mul_mv_id_q5_1_f32];
-                                        } break;
-                                    case GGML_TYPE_Q8_0:
-                                        {
-                                            nth0 = 8;
-                                            nth1 = 8;
-                                            [encoder setComputePipelineState:ctx->pipeline_mul_mv_id_q8_0_f32];
-                                        } break;
-                                    case GGML_TYPE_Q2_K:
-                                        {
-                                            nth0 = 2;
-                                            nth1 = 32;
-                                            [encoder setComputePipelineState:ctx->pipeline_mul_mv_id_q2_K_f32];
-                                        } break;
-                                    case GGML_TYPE_Q3_K:
-                                        {
-                                            nth0 = 2;
-                                            nth1 = 32;
-                                            [encoder setComputePipelineState:ctx->pipeline_mul_mv_id_q3_K_f32];
-                                        } break;
-                                    case GGML_TYPE_Q4_K:
-                                        {
-                                            nth0 = 4; //1;
-                                            nth1 = 8; //32;
-                                            [encoder setComputePipelineState:ctx->pipeline_mul_mv_id_q4_K_f32];
-                                        } break;
-                                    case GGML_TYPE_Q5_K:
-                                        {
-                                            nth0 = 2;
-                                            nth1 = 32;
-                                            [encoder setComputePipelineState:ctx->pipeline_mul_mv_id_q5_K_f32];
-                                        } break;
-                                    case GGML_TYPE_Q6_K:
-                                        {
-                                            nth0 = 2;
-                                            nth1 = 32;
-                                            [encoder setComputePipelineState:ctx->pipeline_mul_mv_id_q6_K_f32];
-                                        } break;
-                                    default:
-                                        {
-                                            GGML_METAL_LOG_ERROR("Asserting on type %d\n", (int)src2t);
-                                            GGML_ASSERT(false && "not implemented");
-                                        }
-                                };
-
-                                if (ggml_is_quantized(src2t)) {
-                                    GGML_ASSERT(ne20 >= nth0*nth1);
-                                }
-
-                                const int64_t _ne1 = 1; // kernels needs a reference in constant memory
-
-                                [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
-                                [encoder setBuffer:id_src1 offset:offs_src1 atIndex:1];
-                                [encoder setBuffer:id_dst  offset:offs_dst  atIndex:2];
-                                [encoder setBytes:&nb01 length:sizeof(nb01) atIndex:3];
-                                [encoder setBytes:&ne20 length:sizeof(ne20) atIndex:4];
-                                [encoder setBytes:&ne21 length:sizeof(ne21) atIndex:5];
-                                [encoder setBytes:&ne22 length:sizeof(ne22) atIndex:6];
-                                [encoder setBytes:&nb20 length:sizeof(nb20) atIndex:7];
-                                [encoder setBytes:&nb21 length:sizeof(nb21) atIndex:8];
-                                [encoder setBytes:&nb22 length:sizeof(nb22) atIndex:9];
-                                [encoder setBytes:&ne10 length:sizeof(ne10) atIndex:10];
-                                [encoder setBytes:&_ne1 length:sizeof(_ne1) atIndex:11];
-                                [encoder setBytes:&ne12 length:sizeof(ne12) atIndex:12];
-                                [encoder setBytes:&ne13 length:sizeof(ne13) atIndex:13];
-                                [encoder setBytes:&nb10 length:sizeof(nb10) atIndex:14];
-                                [encoder setBytes:&nb11 length:sizeof(nb11) atIndex:15];
-                                [encoder setBytes:&nb12 length:sizeof(nb12) atIndex:16];
-                                [encoder setBytes:&ne0  length:sizeof(ne0)  atIndex:17];
-                                [encoder setBytes:&_ne1 length:sizeof(_ne1) atIndex:18];
-                                [encoder setBytes:&nb1  length:sizeof(nb1)  atIndex:19];
-                                [encoder setBytes:&r2   length:sizeof(r2)   atIndex:20];
-                                [encoder setBytes:&r3   length:sizeof(r3)   atIndex:21];
-                                [encoder setBytes:&idx  length:sizeof(idx)  atIndex:22];
-                                // TODO: how to make this an array? read Metal docs
-                                for (int j = 0; j < 8; ++j) {
-                                    // NOTE: this is done like this to avoid uninitialized kernel arguments when n_as < 8
-                                    struct ggml_tensor * src_cur = dst->src[2 + (j % n_as)];
-
-                                    size_t offs_src_cur = 0;
-                                    id<MTLBuffer> id_src_cur = ggml_metal_get_buffer(ctx, src_cur, &offs_src_cur);
-
-                                    [encoder setBuffer:id_src_cur offset:offs_src_cur atIndex:23 + j];
-                                }
-
-                                if (src2t == GGML_TYPE_Q4_0 || src2t == GGML_TYPE_Q4_1 ||
-                                    src2t == GGML_TYPE_Q5_0 || src2t == GGML_TYPE_Q5_1 || src2t == GGML_TYPE_Q8_0 ||
-                                    src2t == GGML_TYPE_Q2_K) { // || src2t == GGML_TYPE_Q4_K) {
-                                    [encoder dispatchThreadgroups:MTLSizeMake((ne21 + 7)/8, _ne1, ne01*ne12*ne13) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
-                                }
-                                else if (src2t == GGML_TYPE_Q4_K) {
-                                    [encoder dispatchThreadgroups:MTLSizeMake((ne21 + 3)/4, _ne1, ne01*ne12*ne13) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
-                                }
-                                else if (src2t == GGML_TYPE_Q3_K) {
-#ifdef GGML_QKK_64
-                                    [encoder dispatchThreadgroups:MTLSizeMake((ne21 + 1)/2, _ne1, ne01*ne12*ne13) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
-#else
-                                    [encoder dispatchThreadgroups:MTLSizeMake((ne21 + 3)/4, _ne1, ne01*ne12*ne13) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
-#endif
-                                }
-                                else if (src2t == GGML_TYPE_Q5_K) {
-                                    [encoder dispatchThreadgroups:MTLSizeMake((ne21 + 3)/4, _ne1, ne01*ne12*ne13) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
-                                }
-                                else if (src2t == GGML_TYPE_Q6_K) {
-                                    [encoder dispatchThreadgroups:MTLSizeMake((ne21 + 1)/2, _ne1, ne01*ne12*ne13) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
-                                } else {
-                                    const int64_t ny = (_ne1 + nrows - 1)/nrows;
-                                    [encoder dispatchThreadgroups:MTLSizeMake(ne21, ny, ne01*ne12*ne13) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
-                                }
-                            }
-                        } break;
-                    case GGML_OP_GET_ROWS:
-                        {
-                            switch (src0->type) {
-                                case GGML_TYPE_F32:  [encoder setComputePipelineState:ctx->pipeline_get_rows_f32];  break;
-                                case GGML_TYPE_F16:  [encoder setComputePipelineState:ctx->pipeline_get_rows_f16];  break;
-                                case GGML_TYPE_Q4_0: [encoder setComputePipelineState:ctx->pipeline_get_rows_q4_0]; break;
-                                case GGML_TYPE_Q4_1: [encoder setComputePipelineState:ctx->pipeline_get_rows_q4_1]; break;
-                                case GGML_TYPE_Q5_0: [encoder setComputePipelineState:ctx->pipeline_get_rows_q5_0]; break;
-                                case GGML_TYPE_Q5_1: [encoder setComputePipelineState:ctx->pipeline_get_rows_q5_1]; break;
-                                case GGML_TYPE_Q8_0: [encoder setComputePipelineState:ctx->pipeline_get_rows_q8_0]; break;
-                                case GGML_TYPE_Q2_K: [encoder setComputePipelineState:ctx->pipeline_get_rows_q2_K]; break;
-                                case GGML_TYPE_Q3_K: [encoder setComputePipelineState:ctx->pipeline_get_rows_q3_K]; break;
-                                case GGML_TYPE_Q4_K: [encoder setComputePipelineState:ctx->pipeline_get_rows_q4_K]; break;
-                                case GGML_TYPE_Q5_K: [encoder setComputePipelineState:ctx->pipeline_get_rows_q5_K]; break;
-                                case GGML_TYPE_Q6_K: [encoder setComputePipelineState:ctx->pipeline_get_rows_q6_K]; break;
-                                case GGML_TYPE_I32:  [encoder setComputePipelineState:ctx->pipeline_get_rows_i32];  break;
-                                default: GGML_ASSERT(false && "not implemented");
-                            }
-
-                            [encoder setBuffer:id_src0     offset:offs_src0 atIndex:0];
-                            [encoder setBuffer:id_src1     offset:offs_src1 atIndex:1];
-                            [encoder setBuffer:id_dst      offset:offs_dst  atIndex:2];
-                            [encoder setBytes:&ne00 length:sizeof( int64_t) atIndex:3];
-                            [encoder setBytes:&nb01 length:sizeof(uint64_t) atIndex:4];
-                            [encoder setBytes:&nb02 length:sizeof(uint64_t) atIndex:5];
-                            [encoder setBytes:&ne10 length:sizeof( int64_t) atIndex:6];
-                            [encoder setBytes:&nb10 length:sizeof( int64_t) atIndex:7];
-                            [encoder setBytes:&nb11 length:sizeof( int64_t) atIndex:8];
-                            [encoder setBytes:&nb1  length:sizeof(uint64_t) atIndex:9];
-                            [encoder setBytes:&nb2  length:sizeof(uint64_t) atIndex:10];
-
-                            [encoder dispatchThreadgroups:MTLSizeMake(ne10, ne11, 1) threadsPerThreadgroup:MTLSizeMake(32, 1, 1)];
-                        } break;
-                    case GGML_OP_RMS_NORM:
-                        {
-                            GGML_ASSERT(ne00 % 4 == 0);
-
-                            float eps;
-                            memcpy(&eps, dst->op_params, sizeof(float));
-
-                            int nth = 32; // SIMD width
-
-                            while (nth < ne00/4 && nth < 1024) {
-                                nth *= 2;
-                            }
-
-                            [encoder setComputePipelineState:ctx->pipeline_rms_norm];
-                            [encoder setBuffer:id_src0 offset:offs_src0        atIndex:0];
-                            [encoder setBuffer:id_dst  offset:offs_dst         atIndex:1];
-                            [encoder setBytes:&ne00    length:sizeof( int64_t) atIndex:2];
-                            [encoder setBytes:&nb01    length:sizeof(uint64_t) atIndex:3];
-                            [encoder setBytes:&eps     length:sizeof(   float) atIndex:4];
-                            [encoder setThreadgroupMemoryLength:32*sizeof(float) atIndex:0];
-
-                            const int64_t nrows = ggml_nrows(src0);
-
-                            [encoder dispatchThreadgroups:MTLSizeMake(nrows, 1, 1) threadsPerThreadgroup:MTLSizeMake(nth, 1, 1)];
-                        } break;
-                    case GGML_OP_GROUP_NORM:
-                        {
-                            GGML_ASSERT(ne00 % 4 == 0);
-
-                            //float eps;
-                            //memcpy(&eps, dst->op_params, sizeof(float));
-
-                            const float eps = 1e-6f; // TODO: temporarily hardcoded
-
-                            const int32_t n_groups = ((int32_t *) dst->op_params)[0];
-
-                            int nth = 32; // SIMD width
-
-                            //while (nth < ne00/4 && nth < 1024) {
-                            //    nth *= 2;
-                            //}
-
-                            [encoder setComputePipelineState:ctx->pipeline_group_norm];
-                            [encoder setBuffer:id_src0  offset:offs_src0        atIndex:0];
-                            [encoder setBuffer:id_dst   offset:offs_dst         atIndex:1];
-                            [encoder setBytes:&ne00     length:sizeof( int64_t) atIndex:2];
-                            [encoder setBytes:&ne01     length:sizeof( int64_t) atIndex:3];
-                            [encoder setBytes:&ne02     length:sizeof( int64_t) atIndex:4];
-                            [encoder setBytes:&nb00     length:sizeof(uint64_t) atIndex:5];
-                            [encoder setBytes:&nb01     length:sizeof(uint64_t) atIndex:6];
-                            [encoder setBytes:&nb02     length:sizeof(uint64_t) atIndex:7];
-                            [encoder setBytes:&n_groups length:sizeof( int32_t) atIndex:8];
-                            [encoder setBytes:&eps      length:sizeof(   float) atIndex:9];
-                            [encoder setThreadgroupMemoryLength:32*sizeof(float) atIndex:0];
-
-                            [encoder dispatchThreadgroups:MTLSizeMake(n_groups, 1, 1) threadsPerThreadgroup:MTLSizeMake(nth, 1, 1)];
-                        } break;
-                    case GGML_OP_NORM:
-                        {
-                            float eps;
-                            memcpy(&eps, dst->op_params, sizeof(float));
-
-                            const int nth = MIN(256, ne00);
-
-                            [encoder setComputePipelineState:ctx->pipeline_norm];
-                            [encoder setBuffer:id_src0 offset:offs_src0        atIndex:0];
-                            [encoder setBuffer:id_dst  offset:offs_dst         atIndex:1];
-                            [encoder setBytes:&ne00    length:sizeof( int64_t) atIndex:2];
-                            [encoder setBytes:&nb01    length:sizeof(uint64_t) atIndex:3];
-                            [encoder setBytes:&eps     length:sizeof(   float) atIndex:4];
-                            [encoder setThreadgroupMemoryLength:GGML_PAD(nth*sizeof(float), 16) atIndex:0];
-
-                            const int64_t nrows = ggml_nrows(src0);
-
-                            [encoder dispatchThreadgroups:MTLSizeMake(nrows, 1, 1) threadsPerThreadgroup:MTLSizeMake(nth, 1, 1)];
-                        } break;
-                    case GGML_OP_ALIBI:
-                        {
-                            GGML_ASSERT((src0t == GGML_TYPE_F32));
-
-                            const int nth = MIN(1024, ne00);
-
-                            //const int n_past = ((int32_t *) dst->op_params)[0];
-                            const int n_head = ((int32_t *) dst->op_params)[1];
-                            float max_bias;
-                            memcpy(&max_bias, (int32_t *) dst->op_params + 2, sizeof(float));
-
-                            const int n_heads_log2_floor = 1 << (int) floor(log2(n_head));
-                            const float m0 = powf(2.0f, -(max_bias) / n_heads_log2_floor);
-                            const float m1 = powf(2.0f, -(max_bias / 2.0f) / n_heads_log2_floor);
-
-                            [encoder setComputePipelineState:ctx->pipeline_alibi_f32];
-                            [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
-                            [encoder setBuffer:id_dst  offset:offs_dst  atIndex:1];
-                            [encoder setBytes:&ne00 length:sizeof( int64_t) atIndex:2];
-                            [encoder setBytes:&ne01 length:sizeof( int64_t) atIndex:3];
-                            [encoder setBytes:&ne02 length:sizeof( int64_t) atIndex:4];
-                            [encoder setBytes:&ne03 length:sizeof( int64_t) atIndex:5];
-                            [encoder setBytes:&nb00 length:sizeof(uint64_t) atIndex:6];
-                            [encoder setBytes:&nb01 length:sizeof(uint64_t) atIndex:7];
-                            [encoder setBytes:&nb02 length:sizeof(uint64_t) atIndex:8];
-                            [encoder setBytes:&nb03 length:sizeof(uint64_t) atIndex:9];
-                            [encoder setBytes:&ne0  length:sizeof( int64_t) atIndex:10];
-                            [encoder setBytes:&ne1  length:sizeof( int64_t) atIndex:11];
-                            [encoder setBytes:&ne2  length:sizeof( int64_t) atIndex:12];
-                            [encoder setBytes:&ne3  length:sizeof( int64_t) atIndex:13];
-                            [encoder setBytes:&nb0  length:sizeof(uint64_t) atIndex:14];
-                            [encoder setBytes:&nb1  length:sizeof(uint64_t) atIndex:15];
-                            [encoder setBytes:&nb2  length:sizeof(uint64_t) atIndex:16];
-                            [encoder setBytes:&nb3  length:sizeof(uint64_t) atIndex:17];
-                            [encoder setBytes:&m0   length:sizeof(   float) atIndex:18];
-                            [encoder setBytes:&m1   length:sizeof(   float) atIndex:19];
-                            [encoder setBytes:&n_heads_log2_floor   length:sizeof(int) atIndex:20];
-
-                            [encoder dispatchThreadgroups:MTLSizeMake(ne01, ne02, ne03) threadsPerThreadgroup:MTLSizeMake(nth, 1, 1)];
-                        } break;
-                    case GGML_OP_ROPE:
-                        {
-                            GGML_ASSERT(ne10 == ne02);
-
-                            const int nth = MIN(1024, ne00);
-
-                            const int n_past     = ((int32_t *) dst->op_params)[0];
-                            const int n_dims     = ((int32_t *) dst->op_params)[1];
-                            const int mode       = ((int32_t *) dst->op_params)[2];
-                            // skip 3, n_ctx, used in GLM RoPE, unimplemented in metal
-                            const int n_orig_ctx = ((int32_t *) dst->op_params)[4];
-
-                            float freq_base, freq_scale, ext_factor, attn_factor, beta_fast, beta_slow;
-                            memcpy(&freq_base,   (int32_t *) dst->op_params +  5, sizeof(float));
-                            memcpy(&freq_scale,  (int32_t *) dst->op_params +  6, sizeof(float));
-                            memcpy(&ext_factor,  (int32_t *) dst->op_params +  7, sizeof(float));
-                            memcpy(&attn_factor, (int32_t *) dst->op_params +  8, sizeof(float));
-                            memcpy(&beta_fast,   (int32_t *) dst->op_params +  9, sizeof(float));
-                            memcpy(&beta_slow,   (int32_t *) dst->op_params + 10, sizeof(float));
-
-                            switch (src0->type) {
-                                case GGML_TYPE_F32: [encoder setComputePipelineState:ctx->pipeline_rope_f32]; break;
-                                case GGML_TYPE_F16: [encoder setComputePipelineState:ctx->pipeline_rope_f16]; break;
-                                default: GGML_ASSERT(false);
-                            };
-
-                            [encoder setBuffer:id_src0     offset:offs_src0        atIndex:0];
-                            [encoder setBuffer:id_src1     offset:offs_src1        atIndex:1];
-                            [encoder setBuffer:id_dst      offset:offs_dst         atIndex:2];
-                            [encoder setBytes:&ne00        length:sizeof( int64_t) atIndex:3];
-                            [encoder setBytes:&ne01        length:sizeof( int64_t) atIndex:4];
-                            [encoder setBytes:&ne02        length:sizeof( int64_t) atIndex:5];
-                            [encoder setBytes:&ne03        length:sizeof( int64_t) atIndex:6];
-                            [encoder setBytes:&nb00        length:sizeof(uint64_t) atIndex:7];
-                            [encoder setBytes:&nb01        length:sizeof(uint64_t) atIndex:8];
-                            [encoder setBytes:&nb02        length:sizeof(uint64_t) atIndex:9];
-                            [encoder setBytes:&nb03        length:sizeof(uint64_t) atIndex:10];
-                            [encoder setBytes:&ne0         length:sizeof( int64_t) atIndex:11];
-                            [encoder setBytes:&ne1         length:sizeof( int64_t) atIndex:12];
-                            [encoder setBytes:&ne2         length:sizeof( int64_t) atIndex:13];
-                            [encoder setBytes:&ne3         length:sizeof( int64_t) atIndex:14];
-                            [encoder setBytes:&nb0         length:sizeof(uint64_t) atIndex:15];
-                            [encoder setBytes:&nb1         length:sizeof(uint64_t) atIndex:16];
-                            [encoder setBytes:&nb2         length:sizeof(uint64_t) atIndex:17];
-                            [encoder setBytes:&nb3         length:sizeof(uint64_t) atIndex:18];
-                            [encoder setBytes:&n_past      length:sizeof(     int) atIndex:19];
-                            [encoder setBytes:&n_dims      length:sizeof(     int) atIndex:20];
-                            [encoder setBytes:&mode        length:sizeof(     int) atIndex:21];
-                            [encoder setBytes:&n_orig_ctx  length:sizeof(     int) atIndex:22];
-                            [encoder setBytes:&freq_base   length:sizeof(   float) atIndex:23];
-                            [encoder setBytes:&freq_scale  length:sizeof(   float) atIndex:24];
-                            [encoder setBytes:&ext_factor  length:sizeof(   float) atIndex:25];
-                            [encoder setBytes:&attn_factor length:sizeof(   float) atIndex:26];
-                            [encoder setBytes:&beta_fast   length:sizeof(   float) atIndex:27];
-                            [encoder setBytes:&beta_slow   length:sizeof(   float) atIndex:28];
-
-                            [encoder dispatchThreadgroups:MTLSizeMake(ne01, ne02, ne03) threadsPerThreadgroup:MTLSizeMake(nth, 1, 1)];
-                        } break;
-                    case GGML_OP_IM2COL:
-                        {
-                            GGML_ASSERT(src0->type == GGML_TYPE_F16);
-                            GGML_ASSERT(src1->type == GGML_TYPE_F32);
-                            GGML_ASSERT( dst->type == GGML_TYPE_F16);
-
-                            const int32_t s0 = ((const int32_t *)(dst->op_params))[0];
-                            const int32_t s1 = ((const int32_t *)(dst->op_params))[1];
-                            const int32_t p0 = ((const int32_t *)(dst->op_params))[2];
-                            const int32_t p1 = ((const int32_t *)(dst->op_params))[3];
-                            const int32_t d0 = ((const int32_t *)(dst->op_params))[4];
-                            const int32_t d1 = ((const int32_t *)(dst->op_params))[5];
-                            const bool is_2D = ((const int32_t *)(dst->op_params))[6] == 1;
-
-                            const int32_t N  = src1->ne[is_2D ? 3 : 2];
-                            const int32_t IC = src1->ne[is_2D ? 2 : 1];
-                            const int32_t IH = is_2D ? src1->ne[1] : 1;
-                            const int32_t IW =         src1->ne[0];
-
-                            const int32_t KH = is_2D ? src0->ne[1] : 1;
-                            const int32_t KW =         src0->ne[0];
-
-                            const int32_t OH = is_2D ? dst->ne[2] : 1;
-                            const int32_t OW =         dst->ne[1];
-
-                            const int32_t CHW = IC * KH * KW;
-
-                            const int32_t ofs0 = src1->nb[is_2D ? 3 : 2] / 4;
-                            const int32_t ofs1 = src1->nb[is_2D ? 2 : 1] / 4;
-
-                            switch (src0->type) {
-                                case GGML_TYPE_F32: GGML_ASSERT(false && "not implemented"); break;
-                                case GGML_TYPE_F16: [encoder setComputePipelineState:ctx->pipeline_im2col_f16]; break;
-                                default: GGML_ASSERT(false);
-                            };
-
-                            [encoder setBuffer:id_src1 offset:offs_src1        atIndex:0];
-                            [encoder setBuffer:id_dst  offset:offs_dst         atIndex:1];
-                            [encoder setBytes:&ofs0    length:sizeof( int32_t) atIndex:2];
-                            [encoder setBytes:&ofs1    length:sizeof( int32_t) atIndex:3];
-                            [encoder setBytes:&IW      length:sizeof( int32_t) atIndex:4];
-                            [encoder setBytes:&IH      length:sizeof( int32_t) atIndex:5];
-                            [encoder setBytes:&CHW     length:sizeof( int32_t) atIndex:6];
-                            [encoder setBytes:&s0      length:sizeof( int32_t) atIndex:7];
-                            [encoder setBytes:&s1      length:sizeof( int32_t) atIndex:8];
-                            [encoder setBytes:&p0      length:sizeof( int32_t) atIndex:9];
-                            [encoder setBytes:&p1      length:sizeof( int32_t) atIndex:10];
-                            [encoder setBytes:&d0      length:sizeof( int32_t) atIndex:11];
-                            [encoder setBytes:&d1      length:sizeof( int32_t) atIndex:12];
-
-                            [encoder dispatchThreadgroups:MTLSizeMake(IC, OH, OW) threadsPerThreadgroup:MTLSizeMake(N, KH, KW)];
-                        } break;
-                    case GGML_OP_UPSCALE:
-                        {
-                            GGML_ASSERT(src0->type == GGML_TYPE_F32);
-
-                            const int sf = dst->op_params[0];
-
-                            [encoder setComputePipelineState:ctx->pipeline_upscale_f32];
-                            [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
-                            [encoder setBuffer:id_dst  offset:offs_dst  atIndex:1];
-                            [encoder setBytes:&ne00 length:sizeof(ne00) atIndex:2];
-                            [encoder setBytes:&ne01 length:sizeof(ne01) atIndex:3];
-                            [encoder setBytes:&ne02 length:sizeof(ne02) atIndex:4];
-                            [encoder setBytes:&ne03 length:sizeof(ne03) atIndex:5];
-                            [encoder setBytes:&nb00 length:sizeof(nb00) atIndex:6];
-                            [encoder setBytes:&nb01 length:sizeof(nb01) atIndex:7];
-                            [encoder setBytes:&nb02 length:sizeof(nb02) atIndex:8];
-                            [encoder setBytes:&nb03 length:sizeof(nb03) atIndex:9];
-                            [encoder setBytes:&ne0  length:sizeof(ne0)  atIndex:10];
-                            [encoder setBytes:&ne1  length:sizeof(ne1)  atIndex:11];
-                            [encoder setBytes:&ne2  length:sizeof(ne2)  atIndex:12];
-                            [encoder setBytes:&ne3  length:sizeof(ne3)  atIndex:13];
-                            [encoder setBytes:&nb0  length:sizeof(nb0)  atIndex:14];
-                            [encoder setBytes:&nb1  length:sizeof(nb1)  atIndex:15];
-                            [encoder setBytes:&nb2  length:sizeof(nb2)  atIndex:16];
-                            [encoder setBytes:&nb3  length:sizeof(nb3)  atIndex:17];
-                            [encoder setBytes:&sf   length:sizeof(sf)   atIndex:18];
-
-                            const int nth = MIN((int) ctx->pipeline_upscale_f32.maxTotalThreadsPerThreadgroup, ne0);
-
-                            [encoder dispatchThreadgroups:MTLSizeMake(ne1, ne2, ne3) threadsPerThreadgroup:MTLSizeMake(nth, 1, 1)];
-                        } break;
-                    case GGML_OP_PAD:
-                        {
-                            GGML_ASSERT(src0->type == GGML_TYPE_F32);
-
-                            [encoder setComputePipelineState:ctx->pipeline_pad_f32];
-                            [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
-                            [encoder setBuffer:id_dst  offset:offs_dst  atIndex:1];
-                            [encoder setBytes:&ne00 length:sizeof(ne00) atIndex:2];
-                            [encoder setBytes:&ne01 length:sizeof(ne01) atIndex:3];
-                            [encoder setBytes:&ne02 length:sizeof(ne02) atIndex:4];
-                            [encoder setBytes:&ne03 length:sizeof(ne03) atIndex:5];
-                            [encoder setBytes:&nb00 length:sizeof(nb00) atIndex:6];
-                            [encoder setBytes:&nb01 length:sizeof(nb01) atIndex:7];
-                            [encoder setBytes:&nb02 length:sizeof(nb02) atIndex:8];
-                            [encoder setBytes:&nb03 length:sizeof(nb03) atIndex:9];
-                            [encoder setBytes:&ne0  length:sizeof(ne0)  atIndex:10];
-                            [encoder setBytes:&ne1  length:sizeof(ne1)  atIndex:11];
-                            [encoder setBytes:&ne2  length:sizeof(ne2)  atIndex:12];
-                            [encoder setBytes:&ne3  length:sizeof(ne3)  atIndex:13];
-                            [encoder setBytes:&nb0  length:sizeof(nb0)  atIndex:14];
-                            [encoder setBytes:&nb1  length:sizeof(nb1)  atIndex:15];
-                            [encoder setBytes:&nb2  length:sizeof(nb2)  atIndex:16];
-                            [encoder setBytes:&nb3  length:sizeof(nb3)  atIndex:17];
-
-                            const int nth = MIN(1024, ne0);
-
-                            [encoder dispatchThreadgroups:MTLSizeMake(ne1, ne2, ne3) threadsPerThreadgroup:MTLSizeMake(nth, 1, 1)];
-                        } break;
-                    case GGML_OP_ARGSORT:
-                        {
-                            GGML_ASSERT(src0->type == GGML_TYPE_F32);
-                            GGML_ASSERT( dst->type == GGML_TYPE_I32);
-
-                            const int nrows = ggml_nrows(src0);
-
-                            enum ggml_sort_order order = (enum ggml_sort_order) dst->op_params[0];
-
-                            switch (order) {
-                                case GGML_SORT_ASC:  [encoder setComputePipelineState:ctx->pipeline_argsort_f32_i32_asc];  break;
-                                case GGML_SORT_DESC: [encoder setComputePipelineState:ctx->pipeline_argsort_f32_i32_desc]; break;
-                                default: GGML_ASSERT(false);
-                            };
-
-                            [encoder setBuffer:id_src0 offset:offs_src0        atIndex:0];
-                            [encoder setBuffer:id_dst  offset:offs_dst         atIndex:1];
-                            [encoder setBytes:&ne00    length:sizeof( int64_t) atIndex:2];
-
-                            [encoder dispatchThreadgroups:MTLSizeMake(1, nrows, 1) threadsPerThreadgroup:MTLSizeMake(ne00, 1, 1)];
-                        } break;
-                    case GGML_OP_LEAKY_RELU:
-                        {
-                            GGML_ASSERT(src0->type == GGML_TYPE_F32);
-
-                            float slope;
-                            memcpy(&slope, dst->op_params, sizeof(float));
-
-                            [encoder setComputePipelineState:ctx->pipeline_leaky_relu_f32];
-                            [encoder setBuffer:id_src0 offset:offs_src0   atIndex:0];
-                            [encoder setBuffer:id_dst  offset:offs_dst    atIndex:1];
-                            [encoder setBytes:&slope length:sizeof(slope) atIndex:2];
-
-                            const int64_t n = ggml_nelements(dst);
-
-                            [encoder dispatchThreadgroups:MTLSizeMake(n, 1, 1) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)];
-                        } break;
-                    case GGML_OP_DUP:
-                    case GGML_OP_CPY:
-                    case GGML_OP_CONT:
-                        {
-                            GGML_ASSERT(ne00 % ggml_blck_size(src0->type) == 0);
-
-                            int nth = MIN(1024, ne00/ggml_blck_size(src0->type));
-
-                            switch (src0t) {
-                                case GGML_TYPE_F32:
-                                    {
-                                        GGML_ASSERT(ne0 % ggml_blck_size(dst->type) == 0);
-
-                                        switch (dstt) {
-                                            case GGML_TYPE_F16:  [encoder setComputePipelineState:ctx->pipeline_cpy_f32_f16];  break;
-                                            case GGML_TYPE_F32:  [encoder setComputePipelineState:ctx->pipeline_cpy_f32_f32];  break;
-                                            case GGML_TYPE_Q8_0: [encoder setComputePipelineState:ctx->pipeline_cpy_f32_q8_0]; break;
-                                            case GGML_TYPE_Q4_0: [encoder setComputePipelineState:ctx->pipeline_cpy_f32_q4_0]; break;
-                                            case GGML_TYPE_Q4_1: [encoder setComputePipelineState:ctx->pipeline_cpy_f32_q4_1]; break;
-                                            //case GGML_TYPE_Q5_0: [encoder setComputePipelineState:ctx->pipeline_cpy_f32_q5_0]; break;
-                                            //case GGML_TYPE_Q5_1: [encoder setComputePipelineState:ctx->pipeline_cpy_f32_q5_1]; break;
-                                            default: GGML_ASSERT(false && "not implemented");
-                                        };
-                                    } break;
-                                case GGML_TYPE_F16:
-                                    {
-                                        switch (dstt) {
-                                            case GGML_TYPE_F16: [encoder setComputePipelineState:ctx->pipeline_cpy_f16_f16]; break;
-                                            case GGML_TYPE_F32: [encoder setComputePipelineState:ctx->pipeline_cpy_f16_f32]; break;
-                                            default: GGML_ASSERT(false && "not implemented");
-                                        };
-                                    } break;
-                                default: GGML_ASSERT(false && "not implemented");
-                            }
-
-                            [encoder setBuffer:id_src0 offset:offs_src0        atIndex:0];
-                            [encoder setBuffer:id_dst  offset:offs_dst         atIndex:1];
-                            [encoder setBytes:&ne00    length:sizeof( int64_t) atIndex:2];
-                            [encoder setBytes:&ne01    length:sizeof( int64_t) atIndex:3];
-                            [encoder setBytes:&ne02    length:sizeof( int64_t) atIndex:4];
-                            [encoder setBytes:&ne03    length:sizeof( int64_t) atIndex:5];
-                            [encoder setBytes:&nb00    length:sizeof(uint64_t) atIndex:6];
-                            [encoder setBytes:&nb01    length:sizeof(uint64_t) atIndex:7];
-                            [encoder setBytes:&nb02    length:sizeof(uint64_t) atIndex:8];
-                            [encoder setBytes:&nb03    length:sizeof(uint64_t) atIndex:9];
-                            [encoder setBytes:&ne0     length:sizeof( int64_t) atIndex:10];
-                            [encoder setBytes:&ne1     length:sizeof( int64_t) atIndex:11];
-                            [encoder setBytes:&ne2     length:sizeof( int64_t) atIndex:12];
-                            [encoder setBytes:&ne3     length:sizeof( int64_t) atIndex:13];
-                            [encoder setBytes:&nb0     length:sizeof(uint64_t) atIndex:14];
-                            [encoder setBytes:&nb1     length:sizeof(uint64_t) atIndex:15];
-                            [encoder setBytes:&nb2     length:sizeof(uint64_t) atIndex:16];
-                            [encoder setBytes:&nb3     length:sizeof(uint64_t) atIndex:17];
-
-                            [encoder dispatchThreadgroups:MTLSizeMake(ne01, ne02, ne03) threadsPerThreadgroup:MTLSizeMake(nth, 1, 1)];
-                        } break;
-                    default:
-                        {
-                            GGML_METAL_LOG_ERROR("%s: error: node %3d, op = %8s not implemented\n", __func__, i, ggml_op_name(dst->op));
-                            GGML_ASSERT(false);
-                        }
-                }
-            }
-
-            if (encoder != nil) {
-                [encoder endEncoding];
-                encoder = nil;
-            }
-
-            [command_buffer commit];
-        });
-    }
-
-    // wait for all threads to finish
-    dispatch_barrier_sync(ctx->d_queue, ^{});
-
-    // check status of command buffers
-    // needed to detect if the device ran out-of-memory for example (#1881)
-    for (int i = 0; i < n_cb; i++) {
-        [ctx->command_buffers[i] waitUntilCompleted];
-
-        MTLCommandBufferStatus status = (MTLCommandBufferStatus) [ctx->command_buffers[i] status];
-        if (status != MTLCommandBufferStatusCompleted) {
-            GGML_METAL_LOG_INFO("%s: command buffer %d failed with status %lu\n", __func__, i, status);
-            return false;
-        }
-    }
-
-    return true;
-    }
-}
-
-////////////////////////////////////////////////////////////////////////////////
-
-// backend interface
-
-// default buffer
-static id<MTLDevice> g_backend_device = nil;
-static int g_backend_device_ref_count = 0;
-
-static id<MTLDevice> ggml_backend_metal_get_device(void) {
-    if (g_backend_device == nil) {
-        g_backend_device = MTLCreateSystemDefaultDevice();
-    }
-
-    g_backend_device_ref_count++;
-
-    return g_backend_device;
-}
-
-static void ggml_backend_metal_free_device(void) {
-    assert(g_backend_device_ref_count > 0);
-
-    g_backend_device_ref_count--;
-
-    if (g_backend_device_ref_count == 0) {
-        [g_backend_device release];
-        g_backend_device = nil;
-    }
-}
-
-static void * ggml_backend_metal_buffer_get_base(ggml_backend_buffer_t buffer) {
-    struct ggml_backend_metal_buffer_context * ctx = (struct ggml_backend_metal_buffer_context *)buffer->context;
-
-    return ctx->all_data;
-}
-
-static void ggml_backend_metal_buffer_free_buffer(ggml_backend_buffer_t buffer) {
-    struct ggml_backend_metal_buffer_context * ctx = (struct ggml_backend_metal_buffer_context *)buffer->context;
-
-    for (int i = 0; i < ctx->n_buffers; i++) {
-        [ctx->buffers[i].metal release];
-    }
-    ggml_backend_metal_free_device();
-
-    if (ctx->owned) {
-        free(ctx->all_data);
-    }
-
-    free(ctx);
-}
-
-static void ggml_backend_metal_buffer_set_tensor(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor, const void * data, size_t offset, size_t size) {
-    memcpy((char *)tensor->data + offset, data, size);
-
-    UNUSED(buffer);
-}
-
-static void ggml_backend_metal_buffer_get_tensor(ggml_backend_buffer_t buffer, const struct ggml_tensor * tensor, void * data, size_t offset, size_t size) {
-    memcpy(data, (const char *)tensor->data + offset, size);
-
-    UNUSED(buffer);
-}
-
-static void ggml_backend_metal_buffer_cpy_tensor_from(ggml_backend_buffer_t buffer, struct ggml_tensor * src, struct ggml_tensor * dst) {
-    ggml_backend_tensor_get(src, dst->data, 0, ggml_nbytes(src));
-
-    UNUSED(buffer);
-}
-
-static void ggml_backend_metal_buffer_cpy_tensor_to(ggml_backend_buffer_t buffer, struct ggml_tensor * src, struct ggml_tensor * dst) {
-    ggml_backend_tensor_set(dst, src->data, 0, ggml_nbytes(src));
-
-    UNUSED(buffer);
-}
-
-static void ggml_backend_metal_buffer_clear(ggml_backend_buffer_t buffer, uint8_t value) {
-    struct ggml_backend_metal_buffer_context * ctx = (struct ggml_backend_metal_buffer_context *)buffer->context;
-
-    memset(ctx->all_data, value, ctx->all_size);
-}
-
-static struct ggml_backend_buffer_i ggml_backend_metal_buffer_i = {
-    /* .free_buffer     = */ ggml_backend_metal_buffer_free_buffer,
-    /* .get_base        = */ ggml_backend_metal_buffer_get_base,
-    /* .init_tensor     = */ NULL,
-    /* .set_tensor      = */ ggml_backend_metal_buffer_set_tensor,
-    /* .get_tensor      = */ ggml_backend_metal_buffer_get_tensor,
-    /* .cpy_tensor_from = */ ggml_backend_metal_buffer_cpy_tensor_from,
-    /* .cpy_tensor_to   = */ ggml_backend_metal_buffer_cpy_tensor_to,
-    /* .clear           = */ ggml_backend_metal_buffer_clear,
-};
-
-// default buffer type
-
-static ggml_backend_buffer_t ggml_backend_metal_buffer_type_alloc_buffer(ggml_backend_buffer_type_t buft, size_t size) {
-    struct ggml_backend_metal_buffer_context * ctx = malloc(sizeof(struct ggml_backend_metal_buffer_context));
-
-    const size_t size_page = sysconf(_SC_PAGESIZE);
-
-    size_t size_aligned = size;
-    if ((size_aligned % size_page) != 0) {
-        size_aligned += (size_page - (size_aligned % size_page));
-    }
-
-    id<MTLDevice> device = ggml_backend_metal_get_device();
-
-    ctx->all_data = ggml_metal_host_malloc(size_aligned);
-    ctx->all_size = size_aligned;
-    ctx->owned = true;
-    ctx->n_buffers = 1;
-
-    ctx->buffers[0].data = ctx->all_data;
-    ctx->buffers[0].size = size;
-    ctx->buffers[0].metal = [device newBufferWithBytesNoCopy:ctx->all_data
-                    length:size_aligned
-                    options:MTLResourceStorageModeShared
-                    deallocator:nil];
-
-    if (ctx->buffers[0].metal == nil) {
-        GGML_METAL_LOG_ERROR("%s: error: failed to allocate buffer, size = %8.2f MiB\n", __func__, size_aligned / 1024.0 / 1024.0);
-        free(ctx);
-        ggml_backend_metal_free_device();
-        return NULL;
-    }
-
-    GGML_METAL_LOG_INFO("%s: allocated buffer, size = %8.2f MiB", __func__, size_aligned / 1024.0 / 1024.0);
-
-
-#if TARGET_OS_OSX
-    GGML_METAL_LOG_INFO(", (%8.2f / %8.2f)",
-            device.currentAllocatedSize / 1024.0 / 1024.0,
-            device.recommendedMaxWorkingSetSize / 1024.0 / 1024.0);
-
-    if (device.currentAllocatedSize > device.recommendedMaxWorkingSetSize) {
-        GGML_METAL_LOG_WARN("%s: warning: current allocated size is greater than the recommended max working set size\n", __func__);
-    } else {
-        GGML_METAL_LOG_INFO("\n");
-    }
-#else
-    GGML_METAL_LOG_INFO(", (%8.2f)\n", device.currentAllocatedSize / 1024.0 / 1024.0);
-#endif
-
-
-    return ggml_backend_buffer_init(buft, ggml_backend_metal_buffer_i, ctx, size);
-}
-
-static size_t ggml_backend_metal_buffer_type_get_alignment(ggml_backend_buffer_type_t buft) {
-    return 32;
-    UNUSED(buft);
-}
-
-static bool ggml_backend_metal_buffer_type_supports_backend(ggml_backend_buffer_type_t buft, ggml_backend_t backend) {
-    return ggml_backend_is_metal(backend) || ggml_backend_is_cpu(backend);
-
-    UNUSED(buft);
-}
-
-static bool ggml_backend_metal_buffer_type_is_host(ggml_backend_buffer_type_t buft) {
-    return true;
-
-    UNUSED(buft);
-}
-
-ggml_backend_buffer_type_t ggml_backend_metal_buffer_type(void) {
-    static struct ggml_backend_buffer_type ggml_backend_buffer_type_metal = {
-        /* .iface = */ {
-            /* .alloc_buffer     = */ ggml_backend_metal_buffer_type_alloc_buffer,
-            /* .get_alignment    = */ ggml_backend_metal_buffer_type_get_alignment,
-            /* .get_alloc_size   = */ NULL, // defaults to ggml_nbytes
-            /* .supports_backend = */ ggml_backend_metal_buffer_type_supports_backend,
-            /* .is_host          = */ ggml_backend_metal_buffer_type_is_host,
-        },
-        /* .context = */ NULL,
-    };
-
-    return &ggml_backend_buffer_type_metal;
-}
-
-// buffer from ptr
-
-ggml_backend_buffer_t ggml_backend_metal_buffer_from_ptr(void * data, size_t size, size_t max_size) {
-    struct ggml_backend_metal_buffer_context * ctx = malloc(sizeof(struct ggml_backend_metal_buffer_context));
-
-    ctx->all_data = data;
-    ctx->all_size = size;
-    ctx->owned = false;
-    ctx->n_buffers = 0;
-
-    const size_t size_page = sysconf(_SC_PAGESIZE);
-    size_t size_aligned = size;
-    if ((size_aligned % size_page) != 0) {
-        size_aligned += (size_page - (size_aligned % size_page));
-    }
-
-    id<MTLDevice> device = ggml_backend_metal_get_device();
-
-    // the buffer fits into the max buffer size allowed by the device
-    if (size_aligned <= device.maxBufferLength) {
-        ctx->buffers[ctx->n_buffers].data = data;
-        ctx->buffers[ctx->n_buffers].size = size;
-
-        ctx->buffers[ctx->n_buffers].metal = [device newBufferWithBytesNoCopy:data length:size_aligned options:MTLResourceStorageModeShared deallocator:nil];
-
-        if (ctx->buffers[ctx->n_buffers].metal == nil) {
-            GGML_METAL_LOG_ERROR("%s: error: failed to allocate buffer, size = %8.2f MiB\n", __func__, size_aligned / 1024.0 / 1024.0);
-            return false;
-        }
-
-        GGML_METAL_LOG_INFO("%s: allocated buffer, size = %8.2f MiB", __func__, size_aligned / 1024.0 / 1024.0);
-
-        ++ctx->n_buffers;
-    } else {
-        // this overlap between the views will guarantee that the tensor with the maximum size will fully fit into
-        // one of the views
-        const size_t size_ovlp = ((max_size + size_page - 1) / size_page + 1) * size_page; // round-up 2 pages just in case
-        const size_t size_step = device.maxBufferLength - size_ovlp;
-        const size_t size_view = device.maxBufferLength;
-
-        for (size_t i = 0; i < size; i += size_step) {
-            const size_t size_step_aligned = (i + size_view <= size) ? size_view : (size_aligned - i);
-
-            ctx->buffers[ctx->n_buffers].data = (void *) ((uint8_t *) data + i);
-            ctx->buffers[ctx->n_buffers].size = size_step_aligned;
-
-            ctx->buffers[ctx->n_buffers].metal = [device newBufferWithBytesNoCopy:(void *) ((uint8_t *) data + i) length:size_step_aligned options:MTLResourceStorageModeShared deallocator:nil];
-
-            if (ctx->buffers[ctx->n_buffers].metal == nil) {
-                GGML_METAL_LOG_ERROR("%s: error: failed to allocate buffer, size = %8.2f MiB\n", __func__, size_step_aligned / 1024.0 / 1024.0);
-                return false;
-            }
-
-            GGML_METAL_LOG_INFO("%s: allocated buffer, size = %8.2f MiB, offs = %12ld", __func__, size_step_aligned / 1024.0 / 1024.0, i);
-            if (i + size_step < size) {
-                GGML_METAL_LOG_INFO("\n");
-            }
-
-            ++ctx->n_buffers;
-        }
-    }
-
-#if TARGET_OS_OSX
-    GGML_METAL_LOG_INFO(", (%8.2f / %8.2f)",
-            device.currentAllocatedSize / 1024.0 / 1024.0,
-            device.recommendedMaxWorkingSetSize / 1024.0 / 1024.0);
-
-    if (device.currentAllocatedSize > device.recommendedMaxWorkingSetSize) {
-        GGML_METAL_LOG_WARN("%s: warning: current allocated size is greater than the recommended max working set size\n", __func__);
-    } else {
-        GGML_METAL_LOG_INFO("\n");
-    }
-#else
-    GGML_METAL_LOG_INFO(", (%8.2f)\n", device.currentAllocatedSize / 1024.0 / 1024.0);
-#endif
-
-    return ggml_backend_buffer_init(ggml_backend_metal_buffer_type(), ggml_backend_metal_buffer_i, ctx, size);
-}
-
-// backend
-
-static const char * ggml_backend_metal_name(ggml_backend_t backend) {
-    return "Metal";
-
-    UNUSED(backend);
-}
-
-static void ggml_backend_metal_free(ggml_backend_t backend) {
-    struct ggml_metal_context * ctx = (struct ggml_metal_context *)backend->context;
-    ggml_metal_free(ctx);
-    free(backend);
-}
-
-static ggml_backend_buffer_type_t ggml_backend_metal_get_default_buffer_type(ggml_backend_t backend) {
-    return ggml_backend_metal_buffer_type();
-
-    UNUSED(backend);
-}
-
-static bool ggml_backend_metal_graph_compute(ggml_backend_t backend, struct ggml_cgraph * cgraph) {
-    struct ggml_metal_context * metal_ctx = (struct ggml_metal_context *)backend->context;
-
-    return ggml_metal_graph_compute(metal_ctx, cgraph);
-}
-
-static bool ggml_backend_metal_supports_op(ggml_backend_t backend, const struct ggml_tensor * op) {
-    return ggml_metal_supports_op(op);
-
-    UNUSED(backend);
-}
-
-static struct ggml_backend_i metal_backend_i = {
-    /* .get_name                = */ ggml_backend_metal_name,
-    /* .free                    = */ ggml_backend_metal_free,
-    /* .get_default_buffer_type = */ ggml_backend_metal_get_default_buffer_type,
-    /* .set_tensor_async        = */ NULL,
-    /* .get_tensor_async        = */ NULL,
-    /* .cpy_tensor_from_async   = */ NULL,
-    /* .cpy_tensor_to_async     = */ NULL,
-    /* .synchronize             = */ NULL,
-    /* .graph_plan_create       = */ NULL,
-    /* .graph_plan_free         = */ NULL,
-    /* .graph_plan_compute      = */ NULL,
-    /* .graph_compute           = */ ggml_backend_metal_graph_compute,
-    /* .supports_op             = */ ggml_backend_metal_supports_op,
-};
-
-ggml_backend_t ggml_backend_metal_init(void) {
-    struct ggml_metal_context * ctx = ggml_metal_init(GGML_DEFAULT_N_THREADS);
-
-    if (ctx == NULL) {
-        return NULL;
-    }
-
-    ggml_backend_t metal_backend = malloc(sizeof(struct ggml_backend));
-
-    *metal_backend = (struct ggml_backend) {
-        /* .interface = */ metal_backend_i,
-        /* .context   = */ ctx,
-    };
-
-    return metal_backend;
-}
-
-bool ggml_backend_is_metal(ggml_backend_t backend) {
-    return backend->iface.get_name == ggml_backend_metal_name;
-}
-
-void ggml_backend_metal_set_n_cb(ggml_backend_t backend, int n_cb) {
-    GGML_ASSERT(ggml_backend_is_metal(backend));
-
-    struct ggml_metal_context * ctx = (struct ggml_metal_context *)backend->context;
-
-    ggml_metal_set_n_cb(ctx, n_cb);
-}
-
-bool ggml_backend_metal_supports_family(ggml_backend_t backend, int family) {
-    GGML_ASSERT(ggml_backend_is_metal(backend));
-
-    struct ggml_metal_context * ctx = (struct ggml_metal_context *)backend->context;
-
-    return [ctx->device supportsFamily:(MTLGPUFamilyApple1 + family - 1)];
-}
-
-ggml_backend_t ggml_backend_reg_metal_init(const char * params, void * user_data); // silence warning
-
-ggml_backend_t ggml_backend_reg_metal_init(const char * params, void * user_data) {
-    return ggml_backend_metal_init();
-
-    GGML_UNUSED(params);
-    GGML_UNUSED(user_data);
-}
diff --git a/src/whisper_cpp/ggml-opencl.cpp b/src/whisper_cpp/ggml-opencl.cpp
deleted file mode 100644
index 496f9cdc..00000000
--- a/src/whisper_cpp/ggml-opencl.cpp
+++ /dev/null
@@ -1,1897 +0,0 @@
-#include "ggml.h"
-#include "ggml-opencl.h"
-
-#include <array>
-#include <atomic>
-#include <cstdio>
-#include <cstdlib>
-#include <cstring>
-#include <limits>
-#include <sstream>
-#include <vector>
-
-#define CL_TARGET_OPENCL_VERSION 110
-#include <clblast.h>
-
-#if defined(_MSC_VER)
-#pragma warning(disable: 4244 4267) // possible loss of data
-#endif
-
-#define CL_DMMV_LOCAL_SIZE 32
-
-#ifndef K_QUANTS_PER_ITERATION
-#define K_QUANTS_PER_ITERATION 1
-#else
-static_assert(K_QUANTS_PER_ITERATION == 1 || K_QUANTS_PER_ITERATION == 2, "K_QUANTS_PER_ITERATION must be 1 or 2");
-#endif
-
-#define MULTILINE_QUOTE(...) #__VA_ARGS__
-static std::string program_source = MULTILINE_QUOTE(
-
-typedef char int8_t;
-typedef uchar uint8_t;
-typedef short int16_t;
-typedef ushort uint16_t;
-typedef int int32_t;
-typedef uint uint32_t;
-
-struct __attribute__ ((packed)) block_q4_0
-{
-    half d;
-    uint8_t qs[QK4_0 / 2];
-};
-
-struct __attribute__ ((packed)) block_q4_1
-{
-    half d;
-    half m;
-    uint8_t qs[QK4_1 / 2];
-};
-
-struct __attribute__ ((packed)) block_q5_0
-{
-    half d;
-    uint32_t qh;
-    uint8_t qs[QK5_0 / 2];
-};
-
-struct __attribute__ ((packed)) block_q5_1
-{
-    half d;
-    half m;
-    uint32_t qh;
-    uint8_t qs[QK5_1 / 2];
-};
-
-struct __attribute__ ((packed)) block_q8_0
-{
-    half d;
-    int8_t qs[QK8_0];
-};
-
-struct __attribute__((packed)) block_q2_K
-{
-    uint8_t scales[16];
-    uint8_t qs[64];
-    half d;
-    half dmin;
-};
-
-struct __attribute__((packed)) block_q3_K
-{
-    uint8_t hmask[32];
-    uint8_t qs[64];
-    uint8_t scales[12];
-    half d;
-};
-
-struct __attribute__((packed)) block_q4_K
-{
-    half d;
-    half dmin;
-    uint8_t scales[12];
-    uint8_t qs[128];
-};
-
-struct __attribute__((packed)) block_q5_K
-{
-    half d;
-    half dmin;
-    uint8_t scales[12];
-    uint8_t qh[32];
-    uint8_t qs[128];
-};
-
-struct __attribute__((packed)) block_q6_K
-{
-    uint8_t ql[128];
-    uint8_t qh[64];
-    int8_t scales[16];
-    half d;
-};
-
-__kernel void convert_fp16_to_fp32(__global half* x, __global float* y) {
-    const uint i = get_global_id(0);
-
-    y[i] = vload_half(0, &x[i]);
-}
-
-void dequantize_q4_0(__global const struct block_q4_0* x, const int ib, const int iqs, float* v0, float* v1) {
-    const float d = vload_half(0, &x[ib].d);
-
-    const uint8_t vui = x[ib].qs[iqs];
-
-    const int8_t vi0 = vui & 0xF;
-    const int8_t vi1 = vui >> 4;
-
-    *v0 = (vi0 - 8)*d;
-    *v1 = (vi1 - 8)*d;
-}
-void dequantize_q4_1(__global const struct block_q4_1* x, const int ib, const int iqs, float* v0, float* v1) {
-    const float d = vload_half(0, &x[ib].d);
-    const float m = vload_half(0, &x[ib].m);
-
-    const uint8_t vui = x[ib].qs[iqs];
-
-    const int8_t vi0 = vui & 0xF;
-    const int8_t vi1 = vui >> 4;
-
-    *v0 = vi0*d + m;
-    *v1 = vi1*d + m;
-}
-void dequantize_q5_0(__global const struct block_q5_0* x, const int ib, const int iqs, float* v0, float* v1) {
-    const float d = vload_half(0, &x[ib].d);
-
-    uint32_t qh = x[ib].qh;
-
-    const uint8_t xh_0 = ((qh >> (iqs +  0)) << 4) & 0x10;
-    const uint8_t xh_1 = ((qh >> (iqs + 12))     ) & 0x10;
-
-    const int32_t x0 = ((x[ib].qs[iqs] & 0xf) | xh_0) - 16;
-    const int32_t x1 = ((x[ib].qs[iqs] >>  4) | xh_1) - 16;
-
-    *v0 = x0*d;
-    *v1 = x1*d;
-}
-void dequantize_q5_1(__global const struct block_q5_1* x, const int ib, const int iqs, float* v0, float* v1) {
-    const float d = vload_half(0, &x[ib].d);
-    const float m = vload_half(0, &x[ib].m);
-
-    uint32_t qh = x[ib].qh;
-
-    const uint8_t xh_0 = ((qh >> (iqs +  0)) << 4) & 0x10;
-    const uint8_t xh_1 = ((qh >> (iqs + 12))     ) & 0x10;
-
-    const int32_t x0 = ((x[ib].qs[iqs] & 0xf) | xh_0);
-    const int32_t x1 = ((x[ib].qs[iqs] >>  4) | xh_1);
-
-    *v0 = x0*d + m;
-    *v1 = x1*d + m;
-}
-void dequantize_q8_0(__global const struct block_q8_0* x, const int ib, const int iqs, float* v0, float* v1) {
-    const float d = vload_half(0, &x[ib].d);
-
-    const int8_t vi0 = x[ib].qs[iqs + 0];
-    const int8_t vi1 = x[ib].qs[iqs + 1];
-
-    *v0 = vi0*d;
-    *v1 = vi1*d;
-}
-void convert_f16(__global half* x, const int ib, const int iqs, float* v0, float* v1){
-    *v0 = vload_half(0, &x[ib + 0]);
-    *v1 = vload_half(0, &x[ib + 1]);
-}
-);
-
-static std::string k_quants_source = MULTILINE_QUOTE(
-inline void get_scale_min_k4(int j, const __global uint8_t *q, uint8_t *d, uint8_t *m)
-{
-    if (j < 4)
-    {
-        *d = q[j] & 63;
-        *m = q[j + 4] & 63;
-    }
-    else
-    {
-        *d = (q[j + 4] & 0xF) | ((q[j - 4] >> 6) << 4);
-        *m = (q[j + 4] >> 4) | ((q[j - 0] >> 6) << 4);
-    }
-}
-
-__kernel void dequantize_block_q2_K(__global const struct block_q2_K *x, __global float *yy)
-{
-    const int i = get_group_id(0) + get_global_offset(0);
-    const int tid = get_local_id(0);
-    const int n = tid / 32;
-    const int l = tid - 32 * n;
-    const int is = 8 * n + l / 16;
-
-    const uint8_t q = x[i].qs[32 * n + l];
-    __global float *y = yy + get_group_id(0) * QK_K + 128 * n;
-
-    const float dall = vload_half(0, &x[i].d);
-    const float dmin = vload_half(0, &x[i].dmin);
-
-    y[l + 0] = dall * (x[i].scales[is + 0] & 0xF) * ((q >> 0) & 3) - dmin * (x[i].scales[is + 0] >> 4);
-    y[l + 32] = dall * (x[i].scales[is + 2] & 0xF) * ((q >> 2) & 3) - dmin * (x[i].scales[is + 2] >> 4);
-    y[l + 64] = dall * (x[i].scales[is + 4] & 0xF) * ((q >> 4) & 3) - dmin * (x[i].scales[is + 4] >> 4);
-    y[l + 96] = dall * (x[i].scales[is + 6] & 0xF) * ((q >> 6) & 3) - dmin * (x[i].scales[is + 6] >> 4);
-}
-
-__kernel void dequantize_block_q3_K(__global const struct block_q3_K *x, __global float *yy)
-{
-    int r = get_local_id(0) / 4;
-    int i = get_group_id(0) + get_global_offset(0);
-    int tid = r / 2;
-    int is0 = r % 2;
-    int l0 = 16 * is0 + 4 * (get_local_id(0) % 4);
-    int n = tid / 4;
-    int j = tid - 4 * n;
-
-    uint8_t m = 1 << (4 * n + j);
-    int is = 8 * n + 2 * j + is0;
-    int shift = 2 * j;
-
-    int8_t us = is < 4 ? (x[i].scales[is - 0] & 0xF) | (((x[i].scales[is + 8] >> 0) & 3) << 4)
-              : is < 8 ? (x[i].scales[is - 0] & 0xF) | (((x[i].scales[is + 4] >> 2) & 3) << 4)
-              : is < 12  ? (x[i].scales[is - 8] >> 4) | (((x[i].scales[is + 0] >> 4) & 3) << 4)
-              : (x[i].scales[is - 8] >> 4) | (((x[i].scales[is - 4] >> 6) & 3) << 4);
-    float d_all = vload_half(0, &x[i].d);
-    float dl = d_all * (us - 32);
-
-    __global float *y = yy + get_group_id(0) * QK_K + 128 * n + 32 * j;
-    const __global uint8_t *q = x[i].qs + 32 * n;
-    const __global uint8_t *hm = x[i].hmask;
-
-    for (int l = l0; l < l0 + 4; ++l)
-        y[l] = dl * ((int8_t)((q[l] >> shift) & 3) - ((hm[l] & m) ? 0 : 4));
-}
-
-__kernel void dequantize_block_q4_K(__global const struct block_q4_K *x, __global float *yy)
-{
-    const int i = get_group_id(0) + get_global_offset(0);
-    const int tid = get_local_id(0);
-    const int il = tid / 8;
-    const int ir = tid % 8;
-    const int is = 2 * il;
-    const int n = 4;
-
-    __global float *y = yy + get_group_id(0) * QK_K + 64 * il + n * ir;
-
-    const float dall = vload_half(0, &x[i].d);
-    const float dmin = vload_half(0, &x[i].dmin);
-
-    __global const uint8_t *q = x[i].qs + 32 * il + n * ir;
-
-    uint8_t sc, m;
-    get_scale_min_k4(is + 0, x[i].scales, &sc, &m);
-    float d1 = dall * sc;
-    float m1 = dmin * m;
-    get_scale_min_k4(is + 1, x[i].scales, &sc, &m);
-    float d2 = dall * sc;
-    float m2 = dmin * m;
-    for (int l = 0; l < n; ++l)
-    {
-        y[l + 0] = d1 * (q[l] & 0xF) - m1;
-        y[l + 32] = d2 * (q[l] >> 4) - m2;
-    }
-}
-
-__kernel void dequantize_block_q5_K(__global const struct block_q5_K *x, __global float *yy)
-{
-    const int i = get_group_id(0) + get_global_offset(0);
-    const int tid = get_local_id(0);
-    const int il = tid / 16;
-    const int ir = tid % 16;
-    const int is = 2 * il;
-
-    __global float *y = yy + get_group_id(0) * QK_K + 64 * il + 2 * ir;
-
-    const float dall = vload_half(0, &x[i].d);
-    const float dmin = vload_half(0, &x[i].dmin);
-
-    __global const uint8_t *ql = x[i].qs + 32 * il + 2 * ir;
-    __global const uint8_t *qh = x[i].qh + 2 * ir;
-
-    uint8_t sc, m;
-    get_scale_min_k4(is + 0, x[i].scales, &sc, &m);
-    const float d1 = dall * sc;
-    const float m1 = dmin * m;
-    get_scale_min_k4(is + 1, x[i].scales, &sc, &m);
-    const float d2 = dall * sc;
-    const float m2 = dmin * m;
-
-    uint8_t hm = 1 << (2 * il);
-    y[0] = d1 * ((ql[0] & 0xF) + (qh[0] & hm ? 16 : 0)) - m1;
-    y[1] = d1 * ((ql[1] & 0xF) + (qh[1] & hm ? 16 : 0)) - m1;
-    hm <<= 1;
-    y[32] = d2 * ((ql[0] >> 4) + (qh[0] & hm ? 16 : 0)) - m2;
-    y[33] = d2 * ((ql[1] >> 4) + (qh[1] & hm ? 16 : 0)) - m2;
-}
-
-__kernel void dequantize_block_q6_K(__global const struct block_q6_K *x, __global float *yy)
-{
-    const int i = get_group_id(0) + get_global_offset(0);
-    const int tid = get_local_id(0);
-    const int ip = tid / 32;
-    const int il = tid - 32 * ip;
-    const int is = 8 * ip + il / 16;
-
-    __global float *y = yy + get_group_id(0) * QK_K + 128 * ip + il;
-
-    const float d = vload_half(0, &x[i].d);
-
-    __global const uint8_t *ql = x[i].ql + 64 * ip + il;
-    const uint8_t qh = x[i].qh[32 * ip + il];
-    __global const int8_t *sc = x[i].scales + is;
-
-    y[0] = d * sc[0] * ((int8_t)((ql[0] & 0xF) | (((qh >> 0) & 3) << 4)) - 32);
-    y[32] = d * sc[2] * ((int8_t)((ql[32] & 0xF) | (((qh >> 2) & 3) << 4)) - 32);
-    y[64] = d * sc[4] * ((int8_t)((ql[0] >> 4) | (((qh >> 4) & 3) << 4)) - 32);
-    y[96] = d * sc[6] * ((int8_t)((ql[32] >> 4) | (((qh >> 6) & 3) << 4)) - 32);
-}
-
-__kernel void dequantize_mul_mat_vec_q2_K(__global const struct block_q2_K * xx, __local float* tmp, __global float* yy, __global float* dst, const int ncols) {
-
-    const int row = get_group_id(0);
-
-    const int num_blocks_per_row = ncols / QK_K;
-    const int ib0 = row*num_blocks_per_row + get_global_offset(0);
-
-    __global const struct block_q2_K * x = xx + ib0;
-
-    const int tid = get_local_id(0)/K_QUANTS_PER_ITERATION;  // 0...31 or 0...15
-    const int ix  = get_local_id(0)%K_QUANTS_PER_ITERATION;  // 0 or 0,1
-
-    const int step = 16/K_QUANTS_PER_ITERATION;
-
-    const int im = tid/step;                             // 0 or 1. 0 computes 0..., 1 computes 128...
-    const int in = tid - step*im;                        // 0...15 or 0...7
-
-    const int l0 = K_QUANTS_PER_ITERATION*in;            // 0...15 or 0...14 in steps of 2
-    const int q_offset = 32*im + l0;
-    const int s_offset = 8*im;
-    const int y_offset = 128*im + l0;
-
-    tmp[16 * ix + tid] = 0;
-
-    uint32_t aux[4];
-    const uint8_t * d = (const uint8_t *)aux;
-    const uint8_t * m = (const uint8_t *)(aux + 2);
-
-    for (int i = ix; i < num_blocks_per_row; i += K_QUANTS_PER_ITERATION) {
-
-        __global const float   * y = yy + i * QK_K + y_offset;
-        __global const uint8_t * q = x[i].qs + q_offset;
-
-        const float dall = vload_half(0, &x[i].d);
-        const float dmin = vload_half(0, &x[i].dmin);
-
-        __global const uint32_t * a = (__global const uint32_t *)(x[i].scales + s_offset);
-        aux[0] = a[0] & 0x0f0f0f0f;
-        aux[1] = a[1] & 0x0f0f0f0f;
-        aux[2] = (a[0] >> 4) & 0x0f0f0f0f;
-        aux[3] = (a[1] >> 4) & 0x0f0f0f0f;
-
-        float sum1 = 0, sum2 = 0;
-        for (int l = 0; l < K_QUANTS_PER_ITERATION; ++l) {
-            sum1 += y[l+ 0] * d[0] * ((q[l+ 0] >> 0) & 3)
-                  + y[l+32] * d[2] * ((q[l+ 0] >> 2) & 3)
-                  + y[l+64] * d[4] * ((q[l+ 0] >> 4) & 3)
-                  + y[l+96] * d[6] * ((q[l+ 0] >> 6) & 3)
-                  + y[l+16] * d[1] * ((q[l+16] >> 0) & 3)
-                  + y[l+48] * d[3] * ((q[l+16] >> 2) & 3)
-                  + y[l+80] * d[5] * ((q[l+16] >> 4) & 3)
-                  +y[l+112] * d[7] * ((q[l+16] >> 6) & 3);
-            sum2 += y[l+ 0] * m[0] + y[l+32] * m[2] + y[l+64] * m[4] + y[ l+96] * m[6]
-                  + y[l+16] * m[1] + y[l+48] * m[3] + y[l+80] * m[5] + y[l+112] * m[7];
-
-        }
-        tmp[16 * ix + tid] += dall * sum1 - dmin * sum2;
-
-    }
-
-    // sum up partial sums and write back result
-    barrier(CLK_LOCAL_MEM_FENCE);
-    for (int s=16; s>0; s>>=1) {
-        if (tid < s) {
-            tmp[tid] += tmp[tid + s];
-        }
-        barrier(CLK_LOCAL_MEM_FENCE);
-    }
-    if (tid == 0) {
-        dst[row] = tmp[0];
-    }
-}
-
-__kernel void dequantize_mul_mat_vec_q3_K(__global const struct block_q3_K * xx, __local float* tmp, __global float* yy, __global float* dst, const int ncols) {
-    const uint16_t kmask1 = 0x0303;
-    const uint16_t kmask2 = 0x0f0f;
-
-    const int row = get_group_id(0);
-
-    const int num_blocks_per_row = ncols / QK_K;
-    const int ib0 = row*num_blocks_per_row + get_global_offset(0);
-
-    __global const struct block_q3_K * x = xx + ib0;
-
-    const int tid = get_local_id(0)/K_QUANTS_PER_ITERATION;  // 0...31 or 0...16
-    const int ix  = get_local_id(0)%K_QUANTS_PER_ITERATION;  // 0 or 0,1
-
-    const int n  = K_QUANTS_PER_ITERATION;               // iterations in the inner loop
-    const int step = 16/K_QUANTS_PER_ITERATION;
-    const int im = tid/step;                             // 0 or 1. 0 computes 0..., 1 computes 128...
-    const int in = tid - step*im;                        // 0....15 or 0...7
-
-    const uint8_t m = 1 << (4*im);
-
-    const int l0 = n*in;                                 // 0...15 or 0...14 in steps of 2
-    const int q_offset =  32*im + l0;
-    const int y_offset = 128*im + l0;
-
-    uint16_t utmp[4];
-    const int8_t * s = (const int8_t *)utmp;
-
-    const uint16_t s_shift = 4*im;
-
-    tmp[16 * ix + tid] = 0;
-
-    for (int i = ix; i < num_blocks_per_row; i += K_QUANTS_PER_ITERATION) {
-
-        __global const float   * y  = yy + i * QK_K + y_offset;
-        __global const uint8_t * q = x[i].qs + q_offset;
-        __global const uint8_t * h = x[i].hmask + l0;
-
-        __global const uint16_t * a = (__global const uint16_t *)x[i].scales;
-        utmp[0] = ((a[0] >> s_shift) & kmask2) | (((a[4] >> (s_shift + 0)) & kmask1) << 4);
-        utmp[1] = ((a[1] >> s_shift) & kmask2) | (((a[5] >> (s_shift + 0)) & kmask1) << 4);
-        utmp[2] = ((a[2] >> s_shift) & kmask2) | (((a[4] >> (s_shift + 2)) & kmask1) << 4);
-        utmp[3] = ((a[3] >> s_shift) & kmask2) | (((a[5] >> (s_shift + 2)) & kmask1) << 4);
-
-        const float d = vload_half(0, &x[i].d);
-
-        float sum = 0;
-        for (int l = 0; l < n; ++l) {
-            sum += y[l+ 0] * (s[0] - 32) * (((q[l] >> 0) & 3) - (h[l] & (m << 0) ? 0 : 4))
-                 + y[l+32] * (s[2] - 32) * (((q[l] >> 2) & 3) - (h[l] & (m << 1) ? 0 : 4))
-                 + y[l+64] * (s[4] - 32) * (((q[l] >> 4) & 3) - (h[l] & (m << 2) ? 0 : 4))
-                 + y[l+96] * (s[6] - 32) * (((q[l] >> 6) & 3) - (h[l] & (m << 3) ? 0 : 4));
-            sum += y[l+16] * (s[1] - 32) * (((q[l+16] >> 0) & 3) - (h[l+16] & (m << 0) ? 0 : 4))
-                 + y[l+48] * (s[3] - 32) * (((q[l+16] >> 2) & 3) - (h[l+16] & (m << 1) ? 0 : 4))
-                 + y[l+80] * (s[5] - 32) * (((q[l+16] >> 4) & 3) - (h[l+16] & (m << 2) ? 0 : 4))
-                + y[l+112] * (s[7] - 32) * (((q[l+16] >> 6) & 3) - (h[l+16] & (m << 3) ? 0 : 4));
-        }
-        tmp[16 * ix + tid] += d * sum;
-
-    }
-
-    // sum up partial sums and write back result
-    barrier(CLK_LOCAL_MEM_FENCE);
-    for (int s=16; s>0; s>>=1) {
-        if (tid < s) {
-            tmp[tid] += tmp[tid + s];
-        }
-        barrier(CLK_LOCAL_MEM_FENCE);
-    }
-    if (tid == 0) {
-        dst[row] = tmp[0];
-    }
-}
-
-__kernel void dequantize_mul_mat_vec_q4_K(__global const struct block_q4_K * xx, __local float* tmp, __global float* yy, __global float* dst, const int ncols) {
-
-    //to rename it later, just to test now
-    const uint16_t kmask1 = 0x3f3f;
-    const uint16_t kmask2 = 0x0f0f;
-    const uint16_t kmask3 = 0xc0c0;
-
-    const int row = get_group_id(0);
-    const int num_blocks_per_row = ncols / QK_K;
-    const int ib0 = row*num_blocks_per_row + get_global_offset(0);
-
-    const int tid = get_local_id(0)/K_QUANTS_PER_ITERATION;  // 0...15
-    const int ix  = get_local_id(0)%K_QUANTS_PER_ITERATION;
-
-    const int step = 8/K_QUANTS_PER_ITERATION;
-
-    const int il  = tid/step;     // 0...3
-    const int ir  = tid - step*il;// 0...3
-    const int n   = 2*K_QUANTS_PER_ITERATION;
-
-    const int im = il/2;  // 0 or 1. 0 computes 0,32 + 128,160, 1 computes 64,96 + 192,224
-    const int in = il%2;
-
-    const int l0 = n*(2*ir + in);
-    const int q_offset = 32*im + l0;
-    const int y_offset = 64*im + l0;
-
-    uint16_t aux[4];
-    const uint8_t * sc = (const uint8_t *)aux;
-
-    __global const struct block_q4_K * x = xx + ib0;
-
-    tmp[16 * ix + tid] = 0;
-
-    for (int i = ix; i < num_blocks_per_row; i += K_QUANTS_PER_ITERATION) {
-
-        __global const uint8_t * q1 = x[i].qs + q_offset;
-        __global const uint8_t * q2 = q1 + 64;
-        __global const float   * y1 = yy + i*QK_K + y_offset;
-        __global const float   * y2 = y1 + 128;
-
-        const float dall = vload_half(0, &x[i].d);
-        const float dmin = vload_half(0, &x[i].dmin);
-
-        __global const uint16_t * a = (__global const uint16_t *)x[i].scales;
-        aux[0] = a[im+0] & kmask1;
-        aux[1] = a[im+2] & kmask1;
-        aux[2] = ((a[im+4] >> 0) & kmask2) | ((a[im+0] & kmask3) >> 2);
-        aux[3] = ((a[im+4] >> 4) & kmask2) | ((a[im+2] & kmask3) >> 2);
-
-        float4 s = (float4)(0.f);
-        float smin = 0;
-        for (int l = 0; l < n; ++l) {
-            s.x += y1[l] * (q1[l] & 0xF); s.y += y1[l+32] * (q1[l] >> 4);
-            s.z += y2[l] * (q2[l] & 0xF); s.w += y2[l+32] * (q2[l] >> 4);
-            smin += y1[l] * sc[2] + y1[l+32] * sc[3] + y2[l] * sc[6] + y2[l+32] * sc[7];
-        }
-        tmp[16 * ix + tid] += dall * (s.x * sc[0] + s.y * sc[1] + s.z * sc[4] + s.w * sc[5]) - dmin * smin;
-
-    }
-
-    // sum up partial sums and write back result
-    barrier(CLK_LOCAL_MEM_FENCE);
-    for (int s=16; s>0; s>>=1) {
-        if (tid < s) {
-            tmp[tid] += tmp[tid + s];
-        }
-        barrier(CLK_LOCAL_MEM_FENCE);
-    }
-    if (tid == 0) {
-        dst[row] = tmp[0];
-    }
-}
-
-__kernel void dequantize_mul_mat_vec_q5_K(__global const struct block_q5_K * xx, __local float* tmp, __global float* yy, __global float* dst, const int ncols) {
-
-    const uint16_t kmask1 = 0x3f3f;
-    const uint16_t kmask2 = 0x0f0f;
-    const uint16_t kmask3 = 0xc0c0;
-
-    const int row = get_group_id(0);
-    const int num_blocks_per_row = ncols / QK_K;
-    const int ib0 = row*num_blocks_per_row + get_global_offset(0);
-
-    const int tid = get_local_id(0)/2;  // 0...15
-    const int ix  = get_local_id(0)%2;
-
-    const int il  = tid/4;     // 0...3
-    const int ir  = tid - 4*il;// 0...3
-    const int n   = 2;
-
-    const int im = il/2;  // 0 or 1. 0 computes 0,32 + 128,160, 1 computes 64,96 + 192,224
-    const int in = il%2;
-
-    const int l0 = n*(2*ir + in);
-    const int q_offset = 32*im + l0;
-    const int y_offset = 64*im + l0;
-
-    const uint8_t hm1  = 1 << (2*im);
-    const uint8_t hm2  = hm1 << 4;
-
-    uint16_t aux[4];
-    const uint8_t * sc = (const uint8_t *)aux;
-
-    __global const struct block_q5_K * x = xx + ib0;
-
-    tmp[16 * ix + tid] = 0;
-
-    for (int i = ix; i < num_blocks_per_row; i += 2) {
-
-        __global const uint8_t * ql1 = x[i].qs + q_offset;
-        __global const uint8_t * ql2 = ql1 + 64;
-        __global const uint8_t * qh  = x[i].qh + l0;
-        __global const float   * y1  = yy + i*QK_K + y_offset;
-        __global const float   * y2  = y1 + 128;
-
-        const float dall = vload_half(0, &x[i].d);
-        const float dmin = vload_half(0, &x[i].dmin);
-
-        __global const uint16_t * a = (__global const uint16_t *)x[i].scales;
-        aux[0] = a[im+0] & kmask1;
-        aux[1] = a[im+2] & kmask1;
-        aux[2] = ((a[im+4] >> 0) & kmask2) | ((a[im+0] & kmask3) >> 2);
-        aux[3] = ((a[im+4] >> 4) & kmask2) | ((a[im+2] & kmask3) >> 2);
-
-        float4 sum = (float4)(0.f);
-        float smin = 0;
-        for (int l = 0; l < n; ++l) {
-            sum.x += y1[l+ 0] * ((ql1[l+ 0] & 0xF) + (qh[l+ 0] & (hm1 << 0) ? 16 : 0))
-                   + y1[l+16] * ((ql1[l+16] & 0xF) + (qh[l+16] & (hm1 << 0) ? 16 : 0));
-            sum.y += y1[l+32] * ((ql1[l+ 0] >>  4) + (qh[l+ 0] & (hm1 << 1) ? 16 : 0))
-                   + y1[l+48] * ((ql1[l+16] >>  4) + (qh[l+16] & (hm1 << 1) ? 16 : 0));
-            sum.z += y2[l+ 0] * ((ql2[l+ 0] & 0xF) + (qh[l+ 0] & (hm2 << 0) ? 16 : 0))
-                   + y2[l+16] * ((ql2[l+16] & 0xF) + (qh[l+16] & (hm2 << 0) ? 16 : 0));
-            sum.w += y2[l+32] * ((ql2[l+ 0] >>  4) + (qh[l+ 0] & (hm2 << 1) ? 16 : 0))
-                   + y2[l+48] * ((ql2[l+16] >>  4) + (qh[l+16] & (hm2 << 1) ? 16 : 0));
-            smin += (y1[l] + y1[l+16]) * sc[2] + (y1[l+32] + y1[l+48]) * sc[3]
-                  + (y2[l] + y2[l+16]) * sc[6] + (y2[l+32] + y2[l+48]) * sc[7];
-        }
-        tmp[16 * ix + tid] += dall * (sum.x * sc[0] + sum.y * sc[1] + sum.z * sc[4] + sum.w * sc[5]) - dmin * smin;
-
-    }
-
-    // sum up partial sums and write back result
-    barrier(CLK_LOCAL_MEM_FENCE);
-    for (int s=16; s>0; s>>=1) {
-        if (tid < s) {
-            tmp[tid] += tmp[tid + s];
-        }
-        barrier(CLK_LOCAL_MEM_FENCE);
-    }
-    if (tid == 0) {
-        dst[row] = tmp[0];
-    }
-}
-
-__kernel void dequantize_mul_mat_vec_q6_K(__global const struct block_q6_K * xx, __local float* tmp, __global const float * yy, __global float * dst, const int ncols) {
-
-    const int row = get_group_id(0);
-
-    const int num_blocks_per_row = ncols / QK_K;
-    const int ib0 = row*num_blocks_per_row + get_global_offset(0);
-
-    __global const struct block_q6_K * x = xx + ib0;
-
-    const int tid = get_local_id(0)/K_QUANTS_PER_ITERATION;  // 0...31 or 0...16
-    const int ix  = get_local_id(0)%K_QUANTS_PER_ITERATION;  // 0 or 0, 1
-
-    const int step = 16/K_QUANTS_PER_ITERATION;          // 16 or 8
-
-    const int im = tid/step;                             // 0 or 1. 0 computes 0..., 1 computes 128...
-    const int in = tid - step*im;                        // 0...15 or 0...7
-
-\n#if K_QUANTS_PER_ITERATION == 1\n
-    const int l0 = K_QUANTS_PER_ITERATION*in;            // 0...15
-    const int is = 0;
-
-\n#else\n
-
-    const int l0 = 4 * in;                               // 0, 4, 8, ..., 28
-    const int is = in / 4;
-
-\n#endif\n
-
-    const int ql_offset = 64*im + l0;
-    const int qh_offset = 32*im + l0;
-    const int s_offset  =  8*im + is;
-    const int y_offset = 128*im + l0;
-
-    tmp[16 * ix + tid] = 0; // partial sum for thread in warp
-
-    for (int i = ix; i < num_blocks_per_row; i += K_QUANTS_PER_ITERATION) {
-
-        __global const float   * y  = yy + i * QK_K + y_offset;
-        __global const uint8_t * ql = x[i].ql + ql_offset;
-        __global const uint8_t * qh = x[i].qh + qh_offset;
-        __global const int8_t  * s  = x[i].scales + s_offset;
-
-        const float d = vload_half(0, &x[i].d);
-
-\n#if K_QUANTS_PER_ITERATION == 1\n
-        float sum = y[ 0] * s[0] * d * ((int8_t)((ql[ 0] & 0xF) | ((qh[ 0] & 0x03) << 4)) - 32)
-                  + y[16] * s[1] * d * ((int8_t)((ql[16] & 0xF) | ((qh[16] & 0x03) << 4)) - 32)
-                  + y[32] * s[2] * d * ((int8_t)((ql[32] & 0xF) | ((qh[ 0] & 0x0c) << 2)) - 32)
-                  + y[48] * s[3] * d * ((int8_t)((ql[48] & 0xF) | ((qh[16] & 0x0c) << 2)) - 32)
-                  + y[64] * s[4] * d * ((int8_t)((ql[ 0]  >> 4) | ((qh[ 0] & 0x30) >> 0)) - 32)
-                  + y[80] * s[5] * d * ((int8_t)((ql[16]  >> 4) | ((qh[16] & 0x30) >> 0)) - 32)
-                  + y[96] * s[6] * d * ((int8_t)((ql[32]  >> 4) | ((qh[ 0] & 0xc0) >> 2)) - 32)
-                  +y[112] * s[7] * d * ((int8_t)((ql[48]  >> 4) | ((qh[16] & 0xc0) >> 2)) - 32);
-        tmp[16 * ix + tid] += sum;
-\n#else\n
-        float sum = 0;
-        for (int l = 0; l < 4; ++l) {
-            sum += y[l+ 0] * s[0] * d * ((int8_t)((ql[l+ 0] & 0xF) | (((qh[l] >> 0) & 3) << 4)) - 32)
-                 + y[l+32] * s[2] * d * ((int8_t)((ql[l+32] & 0xF) | (((qh[l] >> 2) & 3) << 4)) - 32)
-                 + y[l+64] * s[4] * d * ((int8_t)((ql[l+ 0]  >> 4) | (((qh[l] >> 4) & 3) << 4)) - 32)
-                 + y[l+96] * s[6] * d * ((int8_t)((ql[l+32]  >> 4) | (((qh[l] >> 6) & 3) << 4)) - 32);
-        }
-        tmp[16 * ix + tid] += sum;
-\n#endif\n
-
-    }
-
-    // sum up partial sums and write back result
-    barrier(CLK_LOCAL_MEM_FENCE);
-    for (int s=16; s>0; s>>=1) {
-        if (tid < s) {
-            tmp[tid] += tmp[tid + s];
-        }
-        barrier(CLK_LOCAL_MEM_FENCE);
-    }
-    if (tid == 0) {
-        dst[row] = tmp[0];
-    }
-}
-
-);
-
-
-std::string dequant_template = MULTILINE_QUOTE(
-__kernel void KERNEL_NAME(__global X_TYPE* x, __global float* y) {
-    const int i = get_group_id(0)*get_local_size(0) + get_local_id(0)*2;
-
-    if (i >= get_global_size(0)) {
-        return;
-    }
-
-    const uint qk = QUANT_K;
-    const uint qr = QUANT_R;
-
-    const int ib = i/qk + get_global_offset(0); // block index
-    const int iqs = (i%qk)/qr; // quant index
-    const int iybs = i - i%qk; // y block start index
-    const int y_offset = qr == 1 ? 1 : qk/2;
-
-    // dequantize
-    float v0, v1;
-    DEQUANT_FUNC(x, ib, iqs, &v0, &v1);
-    y[iybs + iqs + 0] = v0;
-    y[iybs + iqs + y_offset] = v1;
-}
-);
-
-std::string dequant_mul_mat_vec_template = MULTILINE_QUOTE(
-__kernel void KERNEL_NAME(__global X_TYPE* x, __local float* tmp, __global float* y, __global float* dst, const int ncols) {
-    const int local_size = get_local_size(0);
-    const int row = get_group_id(0);
-    const int tid = get_local_id(0);
-
-    const uint qk = QUANT_K;
-    const uint qr = QUANT_R;
-
-    const int col_step = local_size * 2;
-    const int y_offset = qr == 1 ? 1 : qk/2;
-
-    x += get_global_offset(0);
-
-    tmp[tid] = 0;
-
-    for (int col = tid*2; col < ncols; col += col_step) {
-        const int ib = (row*ncols + col)/qk; // block index
-        const int iqs = (col%qk)/qr; // quant index
-        const int iybs = col - col%qk; // y block start index
-
-        // dequantize
-        float v0, v1;
-        DEQUANT_FUNC(x, ib, iqs, &v0, &v1);
-
-        // matrix multiplication
-        tmp[tid] += v0 * y[iybs + iqs + 0];
-        tmp[tid] += v1 * y[iybs + iqs + y_offset];
-    }
-
-    // sum up partial sums and write back result
-    barrier(CLK_LOCAL_MEM_FENCE);
-    for (int s=local_size/2; s>0; s>>=1) {
-        if (tid < s) {
-            tmp[tid] += tmp[tid + s];
-        }
-        barrier(CLK_LOCAL_MEM_FENCE);
-    }
-    if (tid == 0) {
-        dst[row] = tmp[0];
-    }
-}
-);
-
-
-std::string mul_template = MULTILINE_QUOTE(
-__kernel void KERNEL_NAME(__global TYPE* x, const int x_offset, __global TYPE* y, const int y_offset, __global TYPE* dst, const int dst_offset, const int ky) {
-    const int i = get_group_id(0)*get_local_size(0) + get_local_id(0);
-
-    if (i >= get_global_size(0)) {
-        return;
-    }
-
-    dst[dst_offset + i] = x[x_offset + i] * y[y_offset + i%ky];
-}
-);
-
-#define CL_CHECK(err)                                               \
-    do {                                                            \
-        cl_int err_ = (err);                                        \
-        if (err_ != CL_SUCCESS) {                                   \
-            fprintf(stderr, "ggml_opencl: %s error %d at %s:%d\n",  \
-                #err, err_, __FILE__, __LINE__);                    \
-            exit(1);                                                \
-        }                                                           \
-    } while (0)
-
-#define CLBLAST_CHECK(err)                                          \
-    do {                                                            \
-        CLBlastStatusCode err_ = (err);                             \
-        if (err_ != CLBlastSuccess) {                               \
-            fprintf(stderr, "ggml_opencl: %s error %d at %s:%d\n",  \
-                #err, err_, __FILE__, __LINE__);                    \
-            exit(1);                                                \
-        }                                                           \
-    } while (0)
-
-std::array<std::string, 5> dequant_str_keys = {
-    "KERNEL_NAME", "X_TYPE", "QUANT_K", "QUANT_R", "DEQUANT_FUNC"
-};
-
-std::array<std::string, 30> dequant_str_values = {
-    "dequantize_row_q4_0", "struct block_q4_0", "QK4_0", "QR4_0", "dequantize_q4_0",
-    "dequantize_row_q4_1", "struct block_q4_1", "QK4_1", "QR4_1", "dequantize_q4_1",
-    "dequantize_row_q5_0", "struct block_q5_0", "QK5_0", "QR5_0", "dequantize_q5_0",
-    "dequantize_row_q5_1", "struct block_q5_1", "QK5_1", "QR5_1", "dequantize_q5_1",
-    "dequantize_row_q8_0", "struct block_q8_0", "QK8_0", "QR8_0", "dequantize_q8_0",
-    "convert_row_f16", "half", "1", "1", "convert_f16"
-};
-
-std::array<std::string, 30> dequant_mul_mat_vec_str_values = {
-    "dequantize_mul_mat_vec_q4_0", "struct block_q4_0", "QK4_0", "QR4_0", "dequantize_q4_0",
-    "dequantize_mul_mat_vec_q4_1", "struct block_q4_1", "QK4_1", "QR4_1", "dequantize_q4_1",
-    "dequantize_mul_mat_vec_q5_0", "struct block_q5_0", "QK5_0", "QR5_0", "dequantize_q5_0",
-    "dequantize_mul_mat_vec_q5_1", "struct block_q5_1", "QK5_1", "QR5_1", "dequantize_q5_1",
-    "dequantize_mul_mat_vec_q8_0", "struct block_q8_0", "QK8_0", "QR8_0", "dequantize_q8_0",
-    "convert_mul_mat_vec_f16", "half", "1", "1", "convert_f16"
-};
-
-std::array<std::string, 2> mul_str_keys = {
-    "KERNEL_NAME", "TYPE"
-};
-std::array<std::string, 2> mul_str_values = {
-    "mul_f32", "float"
-};
-
-static std::string& replace(std::string& s, const std::string& from, const std::string& to) {
-    size_t pos = 0;
-    while ((pos = s.find(from, pos)) != std::string::npos) {
-         s.replace(pos, from.length(), to);
-         pos += to.length();
-    }
-    return s;
-}
-
-static std::string generate_kernels() {
-    std::stringstream src;
-    src << program_source << '\n';
-    src << k_quants_source << '\n';
-    for (size_t i = 0; i < dequant_str_values.size(); i += dequant_str_keys.size()) {
-        std::string dequant_kernel = dequant_template;
-        std::string dmmv_kernel = dequant_mul_mat_vec_template;
-        for (size_t j = 0; j < dequant_str_keys.size(); j++) {
-            replace(dequant_kernel, dequant_str_keys[j], dequant_str_values[i + j]);
-            replace(dmmv_kernel, dequant_str_keys[j], dequant_mul_mat_vec_str_values[i + j]);
-        }
-        src << dequant_kernel << '\n';
-        src << dmmv_kernel << '\n';
-    }
-    for (size_t i = 0; i < mul_str_values.size(); i += mul_str_keys.size()) {
-        std::string mul_kernel = mul_template;
-        for (size_t j = 0; j < mul_str_keys.size(); j++) {
-            replace(mul_kernel, mul_str_keys[j], mul_str_values[i + j]);
-        }
-        src << mul_kernel << '\n';
-    }
-
-    return src.str();
-}
-
-static cl_platform_id platform;
-static cl_device_id device;
-static cl_context context;
-static cl_command_queue queue;
-static cl_program program;
-static cl_kernel convert_row_f16_cl;
-static cl_kernel dequantize_row_q4_0_cl, dequantize_row_q4_1_cl, dequantize_row_q5_0_cl, dequantize_row_q5_1_cl, dequantize_row_q8_0_cl;
-static cl_kernel dequantize_mul_mat_vec_q4_0_cl, dequantize_mul_mat_vec_q4_1_cl, dequantize_mul_mat_vec_q5_0_cl, dequantize_mul_mat_vec_q5_1_cl, dequantize_mul_mat_vec_q8_0_cl, convert_mul_mat_vec_f16_cl;
-static cl_kernel dequantize_block_q2_k_cl, dequantize_block_q3_k_cl, dequantize_block_q4_k_cl, dequantize_block_q5_k_cl, dequantize_block_q6_k_cl;
-static cl_kernel dequantize_mul_mat_vec_q2_K_cl, dequantize_mul_mat_vec_q3_K_cl, dequantize_mul_mat_vec_q4_K_cl, dequantize_mul_mat_vec_q5_K_cl, dequantize_mul_mat_vec_q6_K_cl;
-static cl_kernel mul_f32_cl;
-static bool fp16_support;
-
-static cl_program build_program_from_source(cl_context ctx, cl_device_id dev, const char* program_buffer) {
-    cl_program p;
-    char *program_log;
-    size_t program_size;
-    size_t log_size;
-    int err;
-
-    program_size = strlen(program_buffer);
-
-    p = clCreateProgramWithSource(ctx, 1, (const char**)&program_buffer, &program_size, &err);
-    if(err < 0) {
-        fprintf(stderr, "OpenCL error creating program");
-        exit(1);
-    }
-
-    std::string compile_opts = "-cl-mad-enable -cl-unsafe-math-optimizations -cl-finite-math-only -cl-fast-relaxed-math "
-                               "-DQK4_0=32 -DQR4_0=2 -DQK4_1=32 -DQR4_1=2 -DQK5_0=32 -DQR5_0=2 -DQK5_1=32 -DQR5_1=2 -DQK8_0=32 -DQR8_0=1 "
-                               "-DQK_K=256 -DK_QUANTS_PER_ITERATION=" + std::to_string(K_QUANTS_PER_ITERATION);
-
-    err = clBuildProgram(p, 0, NULL, compile_opts.c_str(), NULL, NULL);
-    if(err < 0) {
-
-        clGetProgramBuildInfo(p, dev, CL_PROGRAM_BUILD_LOG, 0, NULL, &log_size);
-        program_log = (char*) malloc(log_size + 1);
-        program_log[log_size] = '\0';
-        clGetProgramBuildInfo(p, dev, CL_PROGRAM_BUILD_LOG, log_size + 1, program_log, NULL);
-        fprintf(stderr, "ggml_opencl: kernel compile error:\n\n%s\n", program_log);
-        free(program_log);
-        exit(1);
-    }
-
-    return p;
-}
-
-void ggml_cl_init(void) {
-    cl_int err;
-
-    struct cl_device;
-    struct cl_platform {
-        cl_platform_id id;
-        unsigned number;
-        char name[128];
-        char vendor[128];
-        struct cl_device * devices;
-        unsigned n_devices;
-        struct cl_device * default_device;
-    };
-
-    struct cl_device {
-        struct cl_platform * platform;
-        cl_device_id id;
-        unsigned number;
-        cl_device_type type;
-        char name[128];
-    };
-
-    enum { NPLAT = 16, NDEV = 16 };
-
-    struct cl_platform platforms[NPLAT];
-    unsigned n_platforms = 0;
-    struct cl_device devices[NDEV];
-    unsigned n_devices = 0;
-    struct cl_device * default_device = NULL;
-
-    platform = NULL;
-    device = NULL;
-
-    cl_platform_id platform_ids[NPLAT];
-    CL_CHECK(clGetPlatformIDs(NPLAT, platform_ids, &n_platforms));
-
-    for (unsigned i = 0; i < n_platforms; i++) {
-        struct cl_platform * p = &platforms[i];
-        p->number = i;
-        p->id = platform_ids[i];
-        CL_CHECK(clGetPlatformInfo(p->id, CL_PLATFORM_NAME, sizeof(p->name), &p->name, NULL));
-        CL_CHECK(clGetPlatformInfo(p->id, CL_PLATFORM_VENDOR, sizeof(p->vendor), &p->vendor, NULL));
-
-        cl_device_id device_ids[NDEV];
-        cl_int clGetDeviceIDsError = clGetDeviceIDs(p->id, CL_DEVICE_TYPE_ALL, NDEV, device_ids, &p->n_devices);
-        if (clGetDeviceIDsError == CL_DEVICE_NOT_FOUND) {
-            p->n_devices = 0;
-        } else {
-            CL_CHECK(clGetDeviceIDsError);
-        }
-        p->devices = p->n_devices > 0 ? &devices[n_devices] : NULL;
-        p->default_device = NULL;
-
-        for (unsigned j = 0; j < p->n_devices; j++) {
-            struct cl_device * d = &devices[n_devices];
-            d->number = n_devices++;
-            d->id = device_ids[j];
-            d->platform = p;
-            CL_CHECK(clGetDeviceInfo(d->id, CL_DEVICE_NAME, sizeof(d->name), &d->name, NULL));
-            CL_CHECK(clGetDeviceInfo(d->id, CL_DEVICE_TYPE, sizeof(d->type), &d->type, NULL));
-
-            if (p->default_device == NULL && d->type == CL_DEVICE_TYPE_GPU) {
-                p->default_device = d;
-            }
-        }
-
-        if (default_device == NULL && p->default_device != NULL) {
-            default_device = p->default_device;
-        }
-    }
-
-    if (n_devices == 0) {
-        fprintf(stderr, "ggml_opencl: could find any OpenCL devices.\n");
-        exit(1);
-    }
-
-    char * user_platform_string = getenv("GGML_OPENCL_PLATFORM");
-    char * user_device_string = getenv("GGML_OPENCL_DEVICE");
-    int user_platform_number = -1;
-    int user_device_number = -1;
-
-    unsigned n;
-    if (user_platform_string != NULL && sscanf(user_platform_string, " %u", &n) == 1 && n < n_platforms) {
-        user_platform_number = (int)n;
-    }
-    if (user_device_string != NULL && sscanf(user_device_string, " %u", &n) == 1 && n < n_devices) {
-        user_device_number = (int)n;
-    }
-    if (user_platform_number != -1 && user_device_number != -1) {
-        cl_platform* platform = &platforms[user_platform_number];
-        if ((unsigned)user_device_number >= platform->n_devices) {
-            fprintf(stderr, "ggml_opencl: invalid device number %d\n", user_device_number);
-            exit(1);
-        }
-        default_device = &platform->devices[user_device_number];
-    } else {
-
-        struct cl_device * selected_devices = devices;
-        unsigned n_selected_devices = n_devices;
-
-        if (user_platform_number == -1 && user_platform_string != NULL && user_platform_string[0] != 0) {
-            for (unsigned i = 0; i < n_platforms; i++) {
-                struct cl_platform * p = &platforms[i];
-                if (strstr(p->name, user_platform_string) != NULL ||
-                    strstr(p->vendor, user_platform_string) != NULL) {
-                    user_platform_number = (int)i;
-                    break;
-                }
-            }
-            if (user_platform_number == -1) {
-                fprintf(stderr, "ggml_opencl: no platform matching '%s' was found.\n", user_platform_string);
-                exit(1);
-            }
-        }
-        if (user_platform_number != -1) {
-            struct cl_platform * p = &platforms[user_platform_number];
-            selected_devices = p->devices;
-            n_selected_devices = p->n_devices;
-            default_device = p->default_device;
-            if (n_selected_devices == 0) {
-                fprintf(stderr, "ggml_opencl: selected platform '%s' does not have any devices.\n", p->name);
-                exit(1);
-            }
-        }
-
-        if (user_device_number == -1 && user_device_string != NULL && user_device_string[0] != 0) {
-            for (unsigned i = 0; i < n_selected_devices; i++) {
-                struct cl_device * d = &selected_devices[i];
-                if (strstr(d->name, user_device_string) != NULL) {
-                    user_device_number = d->number;
-                    break;
-                }
-            }
-            if (user_device_number == -1) {
-                fprintf(stderr, "ggml_opencl: no device matching '%s' was found.\n", user_device_string);
-                exit(1);
-            }
-        }
-        if (user_device_number != -1) {
-            selected_devices = &devices[user_device_number];
-            n_selected_devices = 1;
-            default_device = &selected_devices[0];
-        }
-
-        GGML_ASSERT(n_selected_devices > 0);
-
-        if (default_device == NULL) {
-            default_device = &selected_devices[0];
-        }
-    }
-
-    fprintf(stderr, "ggml_opencl: selecting platform: '%s'\n", default_device->platform->name);
-    fprintf(stderr, "ggml_opencl: selecting device: '%s'\n", default_device->name);
-    if (default_device->type != CL_DEVICE_TYPE_GPU) {
-        fprintf(stderr, "ggml_opencl: warning, not a GPU: '%s'.\n", default_device->name);
-    }
-
-    platform = default_device->platform->id;
-    device = default_device->id;
-
-    size_t ext_str_size;
-    clGetDeviceInfo(device, CL_DEVICE_EXTENSIONS, 0, NULL, &ext_str_size);
-    char *ext_buffer = (char *)alloca(ext_str_size + 1);
-    clGetDeviceInfo(device, CL_DEVICE_EXTENSIONS, ext_str_size, ext_buffer, NULL);
-    ext_buffer[ext_str_size] = '\0'; // ensure it is null terminated
-    // Check if ext_buffer contains cl_khr_fp16
-    fp16_support = strstr(ext_buffer, "cl_khr_fp16") != NULL;
-    fprintf(stderr, "ggml_opencl: device FP16 support: %s\n", fp16_support ? "true" : "false");
-
-    cl_context_properties properties[] = {
-        (intptr_t)CL_CONTEXT_PLATFORM, (intptr_t)platform, 0
-    };
-
-    CL_CHECK((context = clCreateContext(properties, 1, &device, NULL, NULL, &err), err));
-
-    CL_CHECK((queue = clCreateCommandQueue(context, device, CL_QUEUE_OUT_OF_ORDER_EXEC_MODE_ENABLE, &err),
-        (err != CL_INVALID_QUEUE_PROPERTIES && err != CL_INVALID_VALUE ? err :
-        (queue = clCreateCommandQueue(context, device, 0, &err), err)
-    )));
-
-    const std::string kernel_src = generate_kernels();
-
-    program = build_program_from_source(context, device, kernel_src.c_str());
-
-    // FP16 to FP32 kernel
-    CL_CHECK((convert_row_f16_cl = clCreateKernel(program, "convert_row_f16", &err), err));
-
-    // Dequantize kernels
-    CL_CHECK((dequantize_row_q4_0_cl = clCreateKernel(program, "dequantize_row_q4_0", &err), err));
-    CL_CHECK((dequantize_row_q4_1_cl = clCreateKernel(program, "dequantize_row_q4_1", &err), err));
-    CL_CHECK((dequantize_row_q5_0_cl = clCreateKernel(program, "dequantize_row_q5_0", &err), err));
-    CL_CHECK((dequantize_row_q5_1_cl = clCreateKernel(program, "dequantize_row_q5_1", &err), err));
-    CL_CHECK((dequantize_row_q8_0_cl = clCreateKernel(program, "dequantize_row_q8_0", &err), err));
-    CL_CHECK((dequantize_row_q8_0_cl = clCreateKernel(program, "dequantize_row_q8_0", &err), err));
-    CL_CHECK((dequantize_block_q2_k_cl = clCreateKernel(program, "dequantize_block_q2_K", &err), err));
-    CL_CHECK((dequantize_block_q3_k_cl = clCreateKernel(program, "dequantize_block_q3_K", &err), err));
-    CL_CHECK((dequantize_block_q4_k_cl = clCreateKernel(program, "dequantize_block_q4_K", &err), err));
-    CL_CHECK((dequantize_block_q5_k_cl = clCreateKernel(program, "dequantize_block_q5_K", &err), err));
-    CL_CHECK((dequantize_block_q6_k_cl = clCreateKernel(program, "dequantize_block_q6_K", &err), err));
-
-    // dequant mul mat kernel
-    CL_CHECK((dequantize_mul_mat_vec_q4_0_cl = clCreateKernel(program, "dequantize_mul_mat_vec_q4_0", &err), err));
-    CL_CHECK((dequantize_mul_mat_vec_q4_1_cl = clCreateKernel(program, "dequantize_mul_mat_vec_q4_1", &err), err));
-    CL_CHECK((dequantize_mul_mat_vec_q5_0_cl = clCreateKernel(program, "dequantize_mul_mat_vec_q5_0", &err), err));
-    CL_CHECK((dequantize_mul_mat_vec_q5_1_cl = clCreateKernel(program, "dequantize_mul_mat_vec_q5_1", &err), err));
-    CL_CHECK((dequantize_mul_mat_vec_q8_0_cl = clCreateKernel(program, "dequantize_mul_mat_vec_q8_0", &err), err));
-    CL_CHECK((convert_mul_mat_vec_f16_cl = clCreateKernel(program, "convert_mul_mat_vec_f16", &err), err));
-    CL_CHECK((dequantize_mul_mat_vec_q2_K_cl = clCreateKernel(program, "dequantize_mul_mat_vec_q2_K", &err), err));
-    CL_CHECK((dequantize_mul_mat_vec_q3_K_cl = clCreateKernel(program, "dequantize_mul_mat_vec_q3_K", &err), err));
-    CL_CHECK((dequantize_mul_mat_vec_q4_K_cl = clCreateKernel(program, "dequantize_mul_mat_vec_q4_K", &err), err));
-    CL_CHECK((dequantize_mul_mat_vec_q5_K_cl = clCreateKernel(program, "dequantize_mul_mat_vec_q5_K", &err), err));
-    CL_CHECK((dequantize_mul_mat_vec_q6_K_cl = clCreateKernel(program, "dequantize_mul_mat_vec_q6_K", &err), err));
-
-    // mul kernel
-    CL_CHECK((mul_f32_cl = clCreateKernel(program, "mul_f32", &err), err));
-}
-
-static cl_kernel* ggml_get_to_fp32_cl(ggml_type type) {
-    switch (type) {
-        case GGML_TYPE_Q4_0:
-            return &dequantize_row_q4_0_cl;
-        case GGML_TYPE_Q4_1:
-            return &dequantize_row_q4_1_cl;
-        case GGML_TYPE_Q5_0:
-            return &dequantize_row_q5_0_cl;
-        case GGML_TYPE_Q5_1:
-            return &dequantize_row_q5_1_cl;
-        case GGML_TYPE_Q8_0:
-            return &dequantize_row_q8_0_cl;
-        case GGML_TYPE_Q2_K:
-            return &dequantize_block_q2_k_cl;
-        case GGML_TYPE_Q3_K:
-            return &dequantize_block_q3_k_cl;
-        case GGML_TYPE_Q4_K:
-            return &dequantize_block_q4_k_cl;
-        case GGML_TYPE_Q5_K:
-            return &dequantize_block_q5_k_cl;
-        case GGML_TYPE_Q6_K:
-            return &dequantize_block_q6_k_cl;
-        case GGML_TYPE_F16:
-            return &convert_row_f16_cl;
-        default:
-            return nullptr;
-    }
-}
-
-static size_t ggml_cl_global_denom(ggml_type type) {
-    switch (type) {
-        case GGML_TYPE_Q4_0:
-        case GGML_TYPE_Q4_1:
-        case GGML_TYPE_Q5_0:
-        case GGML_TYPE_Q5_1:
-        case GGML_TYPE_Q8_0:
-            return 1;
-        case GGML_TYPE_Q2_K:
-        case GGML_TYPE_Q3_K:
-            return 4;
-        case GGML_TYPE_Q4_K:
-            return 8;
-        case GGML_TYPE_Q5_K:
-        case GGML_TYPE_Q6_K:
-            return 4;
-        case GGML_TYPE_F16:
-        default:
-            return 1;
-    }
-}
-
-static size_t ggml_cl_local_size(ggml_type type) {
-    switch (type) {
-        case GGML_TYPE_Q4_0:
-        case GGML_TYPE_Q4_1:
-        case GGML_TYPE_Q5_0:
-        case GGML_TYPE_Q5_1:
-        case GGML_TYPE_Q8_0:
-            return 0;
-        case GGML_TYPE_Q2_K:
-        case GGML_TYPE_Q3_K:
-            return 64;
-        case GGML_TYPE_Q4_K:
-            return 32;
-        case GGML_TYPE_Q5_K:
-        case GGML_TYPE_Q6_K:
-            return 64;
-        case GGML_TYPE_F16:
-        default:
-            return 0;
-    }
-}
-
-static cl_kernel* ggml_get_dequantize_mul_mat_vec_cl(ggml_type type) {
-    switch (type) {
-        case GGML_TYPE_Q4_0:
-            return &dequantize_mul_mat_vec_q4_0_cl;
-        case GGML_TYPE_Q4_1:
-            return &dequantize_mul_mat_vec_q4_1_cl;
-        case GGML_TYPE_Q5_0:
-            return &dequantize_mul_mat_vec_q5_0_cl;
-        case GGML_TYPE_Q5_1:
-            return &dequantize_mul_mat_vec_q5_1_cl;
-        case GGML_TYPE_Q8_0:
-            return &dequantize_mul_mat_vec_q8_0_cl;
-        case GGML_TYPE_F16:
-            return &convert_mul_mat_vec_f16_cl;
-        case GGML_TYPE_Q2_K:
-            return &dequantize_mul_mat_vec_q2_K_cl;
-        case GGML_TYPE_Q3_K:
-            return &dequantize_mul_mat_vec_q3_K_cl;
-        case GGML_TYPE_Q4_K:
-            return &dequantize_mul_mat_vec_q4_K_cl;
-        case GGML_TYPE_Q5_K:
-            return &dequantize_mul_mat_vec_q5_K_cl;
-        case GGML_TYPE_Q6_K:
-            return &dequantize_mul_mat_vec_q6_K_cl;
-        default:
-            return nullptr;
-    }
-}
-
-// buffer pool for cl
-#define MAX_CL_BUFFERS 256
-
-struct scoped_spin_lock {
-    std::atomic_flag& lock;
-    scoped_spin_lock(std::atomic_flag& lock) : lock(lock) {
-        while (lock.test_and_set(std::memory_order_acquire)) {
-            ; // spin
-        }
-    }
-    ~scoped_spin_lock() {
-        lock.clear(std::memory_order_release);
-    }
-    scoped_spin_lock(const scoped_spin_lock&) = delete;
-    scoped_spin_lock& operator=(const scoped_spin_lock&) = delete;
-};
-
-struct cl_buffer {
-    cl_mem mem;
-    size_t size = 0;
-};
-
-static cl_buffer g_cl_buffer_pool[MAX_CL_BUFFERS];
-static std::atomic_flag g_cl_pool_lock = ATOMIC_FLAG_INIT;
-
-static cl_mem ggml_cl_pool_malloc(size_t size, size_t * actual_size) {
-    scoped_spin_lock lock(g_cl_pool_lock);
-    cl_int err;
-
-    int best_i = -1;
-    size_t best_size = std::numeric_limits<size_t>::max(); //smallest unused buffer that fits our needs
-    int worst_i = -1;
-    size_t worst_size = 0; //largest unused buffer seen so far
-    for (int i = 0; i < MAX_CL_BUFFERS; ++i) {
-        cl_buffer &b = g_cl_buffer_pool[i];
-        if (b.size > 0 && b.size >= size && b.size < best_size)
-        {
-            best_i = i;
-            best_size = b.size;
-        }
-        if (b.size > 0 && b.size > worst_size)
-        {
-            worst_i = i;
-            worst_size = b.size;
-        }
-    }
-    if(best_i!=-1) //found the smallest buffer that fits our needs
-    {
-        cl_buffer& b = g_cl_buffer_pool[best_i];
-        cl_mem mem = b.mem;
-        *actual_size = b.size;
-        b.size = 0;
-        return mem;
-    }
-    if(worst_i!=-1) //no buffer that fits our needs, resize largest one to save memory
-    {
-         cl_buffer& b = g_cl_buffer_pool[worst_i];
-         cl_mem mem = b.mem;
-         b.size = 0;
-         clReleaseMemObject(mem);
-    }
-    cl_mem mem;
-    CL_CHECK((mem = clCreateBuffer(context, CL_MEM_READ_WRITE, size, NULL, &err), err));
-    *actual_size = size;
-    return mem;
-}
-
-static void ggml_cl_pool_free(cl_mem mem, size_t size) {
-    scoped_spin_lock lock(g_cl_pool_lock);
-
-    for (int i = 0; i < MAX_CL_BUFFERS; ++i) {
-        cl_buffer& b = g_cl_buffer_pool[i];
-        if (b.size == 0) {
-            b.mem = mem;
-            b.size = size;
-            return;
-        }
-    }
-    fprintf(stderr, "WARNING: cl buffer pool full, increase MAX_CL_BUFFERS\n");
-    clReleaseMemObject(mem);
-}
-
-void ggml_cl_free_data(const struct ggml_tensor* tensor) {
-    if (tensor->backend != GGML_BACKEND_GPU) {
-        return;
-    }
-
-    cl_mem mem = (cl_mem)tensor->extra;
-    clReleaseMemObject(mem);
-}
-
-static cl_int ggml_cl_h2d_tensor_2d(cl_command_queue queue, cl_mem dst, size_t offset, const struct ggml_tensor * src, uint64_t i3, uint64_t i2, cl_event* ev) {
-    cl_int err;
-    const uint64_t ne0 = src->ne[0];
-    const uint64_t ne1 = src->ne[1];
-    const uint64_t nb0 = src->nb[0];
-    const uint64_t nb1 = src->nb[1];
-    const uint64_t nb2 = src->nb[2];
-    const uint64_t nb3 = src->nb[3];
-    const enum ggml_type type = src->type;
-    const size_t ts = ggml_type_size(type);
-    const size_t bs = ggml_blck_size(type);
-    const uint64_t row_size = ts*ne0/bs;
-
-    const char * x = (const char *) src->data + i2*nb2 + i3*nb3;
-    if (nb0 == ts && nb1 == row_size) {
-        return clEnqueueWriteBuffer(queue, dst, CL_FALSE, offset, ne1*row_size, x, 0, NULL, ev);
-    }
-    if (nb0 == ts) {
-        const size_t buffer_origin[3] = { offset, 0, 0 };
-        const size_t host_origin[3] = { 0, 0, 0 };
-        const size_t region[3] = { row_size, ne1, 1 };
-        return clEnqueueWriteBufferRect(queue, dst, CL_FALSE, buffer_origin, host_origin, region, row_size, 0, nb1, 0, x, 0, NULL, ev);
-    }
-    std::vector<cl_event> events;
-    if (ev && ne1>1) events.reserve(ne1-1);
-    for (uint64_t i1 = 0; i1 < ne1; i1++) {
-        // pretend the row is a matrix with cols=1
-        const size_t buffer_origin[3] = { offset + i1*row_size, 0, 0 };
-        const size_t host_origin[3] = { 0, 0, 0 };
-        const size_t region[3] = { ts, ne0/bs, 1 };
-        // if an event is requested, make the last write wait for all previous writes to complete
-        if (ev && i1) {
-            events.push_back(*ev);
-        }
-        cl_uint nevents = i1 == ne1-1 ? events.size() : 0U;
-        err = clEnqueueWriteBufferRect(queue, dst, CL_FALSE, buffer_origin, host_origin, region, ts, 0, nb0, 0, x + i1*nb1, nevents, nevents ? events.data() : nullptr, ev);
-        if (err != CL_SUCCESS) {
-            for (auto event : events) {
-                clReleaseEvent(event);
-            }
-            return err;
-        }
-    }
-    for (auto event : events) {
-        CL_CHECK(clReleaseEvent(event));
-    }
-    return CL_SUCCESS;
-}
-
-static void ggml_cl_mul_f32(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
-    GGML_ASSERT(src1->backend == GGML_BACKEND_GPU);
-    const int64_t ne00 = src0->ne[0];
-    const int64_t ne01 = src0->ne[1];
-    const int64_t ne02 = src0->ne[2];
-    const int64_t ne03 = src0->ne[3];
-    const int64_t ne10 = src1->ne[0];
-    const int64_t ne11 = src1->ne[1];
-    const int64_t ne12 = src1->ne[2];
-    const int64_t ne13 = src1->ne[3];
-    const int nb2  = dst->nb[2];
-    const int nb3  = dst->nb[3];
-    size_t x_size;
-    size_t d_size;
-
-    cl_mem d_X = ggml_cl_pool_malloc(ne00 * ne01 * sizeof(float), &x_size); // src0
-    cl_mem d_Y = (cl_mem) src1->extra; // src1 is already on device, broadcasted.
-    cl_mem d_D = ggml_cl_pool_malloc(ne00 * ne01 * sizeof(float), &d_size); // dst
-
-
-    for (int64_t i03 = 0; i03 < ne03; i03++) {
-        for (int64_t i02 = 0; i02 < ne02; i02++) {
-            cl_event ev;
-
-            // copy src0 to device
-            CL_CHECK(ggml_cl_h2d_tensor_2d(queue, d_X, 0, src0, i03, i02, &ev));
-
-            const int64_t i13 = i03%ne13;
-            const int64_t i12 = i02%ne12;
-            const int i1 = i13*ne12*ne11 + i12*ne11;
-
-            cl_int x_offset = 0;
-            cl_int y_offset = i1*ne10;
-            cl_int d_offset = 0;
-
-            size_t global = ne00 * ne01;
-            cl_int ky = ne10 * ne11;
-
-            CL_CHECK(clSetKernelArg(mul_f32_cl, 0, sizeof(cl_mem), &d_X));
-            CL_CHECK(clSetKernelArg(mul_f32_cl, 1, sizeof(cl_int), &x_offset));
-            CL_CHECK(clSetKernelArg(mul_f32_cl, 2, sizeof(cl_mem), &d_Y));
-            CL_CHECK(clSetKernelArg(mul_f32_cl, 3, sizeof(cl_int), &y_offset));
-            CL_CHECK(clSetKernelArg(mul_f32_cl, 4, sizeof(cl_mem), &d_D));
-            CL_CHECK(clSetKernelArg(mul_f32_cl, 5, sizeof(cl_int), &d_offset));
-            CL_CHECK(clSetKernelArg(mul_f32_cl, 6, sizeof(cl_int), &ky));
-            CL_CHECK(clEnqueueNDRangeKernel(queue, mul_f32_cl, 1, NULL, &global, NULL, 1, &ev, NULL));
-
-            CL_CHECK(clReleaseEvent(ev));
-            CL_CHECK(clFinish(queue));
-
-            // copy dst to host
-            float * d = (float *) ((char *) dst->data + i02*nb2 + i03*nb3);
-            CL_CHECK(clEnqueueReadBuffer(queue, d_D, true, 0, sizeof(float) * ne00*ne01, d, 0, NULL, NULL));
-        }
-    }
-    ggml_cl_pool_free(d_X, x_size);
-    ggml_cl_pool_free(d_D, d_size);
-}
-
-void ggml_cl_mul(const struct ggml_tensor * src0, const struct ggml_tensor * src1, struct ggml_tensor * dst) {
-    GGML_ASSERT(src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32);
-    ggml_cl_mul_f32(src0, src1, dst);
-}
-
-static void ggml_cl_mul_mat_f32(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
-    const int64_t ne00 = src0->ne[0];
-    const int64_t ne01 = src0->ne[1];
-    const int64_t ne02 = src0->ne[2];
-    const int64_t ne03 = src0->ne[3];
-
-    const int64_t ne10 = src1->ne[0];
-    const int64_t ne11 = src1->ne[1];
-    const int64_t ne12 = src1->ne[2];
-    const int64_t ne13 = src1->ne[3];
-
-    const int nb2  = dst->nb[2];
-    const int nb3  = dst->nb[3];
-
-    const int64_t r2 = ne12 / ne02;
-    const int64_t r3 = ne13 / ne03;
-
-    const float alpha = 1.0f;
-    const float beta = 0.0f;
-    const int x_ne = ne01 * ne00;
-    const int y_ne = ne11 * ne10;
-    const int d_ne = ne11 * ne01;
-
-    size_t x_size;
-    size_t y_size;
-    size_t d_size;
-    cl_mem d_X;
-    if (src0->backend == GGML_BACKEND_GPU) { // NOLINT
-        d_X = (cl_mem) src0->extra;
-    } else {
-        d_X = ggml_cl_pool_malloc(sizeof(float) * x_ne, &x_size);
-    }
-    cl_mem d_Y = ggml_cl_pool_malloc(sizeof(float) * y_ne, &y_size);
-    cl_mem d_D = ggml_cl_pool_malloc(sizeof(float) * d_ne, &d_size);
-
-    size_t x_offset = 0;
-
-    for (int64_t i03 = 0; i03 < ne03; i03++) {
-        // TODO: copy src0 here when r3>1
-        for (int64_t i13 = i03 * r3, e13 = i13 + r3; i13 < e13; i13++) {
-            for (int64_t i02 = 0; i02 < ne02; i02++) {
-                if (src0->backend == GGML_BACKEND_GPU) {
-                    x_offset = (i03 * ne02 + i02) * x_ne;
-                } else {
-                    // copy src0 to device
-                    CL_CHECK(ggml_cl_h2d_tensor_2d(queue, d_X, 0, src0, i03, i02, NULL));
-                }
-
-                for (int64_t i12 = i02 * r2, e12 = i12 + r2; i12 < e12; i12++) {
-                    // copy src1 to device
-                    CL_CHECK(ggml_cl_h2d_tensor_2d(queue, d_Y, 0, src1, i13, i12, NULL));
-
-                    CL_CHECK(clFinish(queue));
-
-                    // compute
-                    cl_event ev_sgemm;
-                    clblast::StatusCode status = clblast::Gemm<cl_float>(clblast::Layout::kColMajor,
-                                                               clblast::Transpose::kYes, clblast::Transpose::kNo,
-                                                               ne01, ne11, ne10,
-                                                               alpha,
-                                                               d_X, x_offset, ne00,
-                                                               d_Y, 0, ne10,
-                                                               beta,
-                                                               d_D, 0, ne01,
-                                                               &queue, &ev_sgemm);
-
-                    if (status != clblast::StatusCode::kSuccess) {
-                        GGML_ASSERT(false);
-                    }
-
-                    // copy dst to host
-                    float * d = (float *) ((char *) dst->data + i12*nb2 + i13*nb3);
-                    CL_CHECK(clEnqueueReadBuffer(queue, d_D, true, 0, sizeof(float) * d_ne, d, 1, &ev_sgemm, NULL));
-                }
-            }
-        }
-    }
-
-    if (src0->backend != GGML_BACKEND_GPU) {
-        ggml_cl_pool_free(d_X, x_size);
-    }
-    ggml_cl_pool_free(d_Y, y_size);
-    ggml_cl_pool_free(d_D, d_size);
-}
-
-static void ggml_cl_mul_mat_f16(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, void * wdata, size_t wsize) {
-    GGML_ASSERT(fp16_support);
-
-    const int64_t ne00 = src0->ne[0];
-    const int64_t ne01 = src0->ne[1];
-    const int64_t ne02 = src0->ne[2];
-    const int64_t ne03 = src0->ne[3];
-
-    const int64_t ne10 = src1->ne[0];
-    const int64_t ne11 = src1->ne[1];
-    const int64_t ne12 = src1->ne[2];
-    const int64_t ne13 = src1->ne[3];
-
-    const int nb10 = src1->nb[0];
-    const int nb11 = src1->nb[1];
-    const int nb12 = src1->nb[2];
-    const int nb13 = src1->nb[3];
-
-    const int nb2  = dst->nb[2];
-    const int nb3  = dst->nb[3];
-
-    const int64_t r2 = ne12 / ne02;
-    const int64_t r3 = ne13 / ne03;
-
-    const ggml_fp16_t alpha = ggml_fp32_to_fp16(1.0f);
-    const ggml_fp16_t beta = ggml_fp32_to_fp16(0.0f);
-    const int x_ne = ne01 * ne00;
-    const int y_ne = ne11 * ne10;
-    const int d_ne = ne11 * ne01;
-
-    GGML_ASSERT(wsize >= sizeof(ggml_fp16_t) * y_ne);
-    GGML_ASSERT(wsize >= sizeof(ggml_fp16_t) * d_ne);
-    ggml_fp16_t * const tmp = (ggml_fp16_t *) wdata;
-
-    size_t x_size;
-    size_t y_size;
-    size_t d_size;
-    cl_mem d_X;
-    if (src0->backend == GGML_BACKEND_GPU) { // NOLINT
-        d_X = (cl_mem) src0->extra;
-    } else {
-        d_X = ggml_cl_pool_malloc(sizeof(ggml_fp16_t) * x_ne, &x_size);
-    }
-    cl_mem d_Y = ggml_cl_pool_malloc(sizeof(ggml_fp16_t) * y_ne, &y_size);
-    cl_mem d_D = ggml_cl_pool_malloc(sizeof(ggml_fp16_t) * d_ne, &d_size);
-
-    bool src1_cont_rows = nb10 == sizeof(float);
-    bool src1_cont_cols = (size_t)nb11 == ne11*sizeof(float);
-
-    size_t x_offset = 0;
-
-    for (int64_t i03 = 0; i03 < ne03; i03++) {
-        // TODO: copy src0 here when r3>1
-        for (int64_t i13 = i03 * r3, e13 = i13 + r3; i13 < e13; i13++) {
-            for (int64_t i02 = 0; i02 < ne02; i02++) {
-                if (src0->backend == GGML_BACKEND_GPU) {
-                    x_offset = (i03 * ne02 + i02) * x_ne;
-                } else {
-                    // copy src0 to device
-                    CL_CHECK(ggml_cl_h2d_tensor_2d(queue, d_X, 0, src0, i03, i02, NULL));
-                }
-
-                for (int64_t i12 = i02 * r2, e12 = i12 + r2; i12 < e12; i12++) {
-                    // convert src1 to fp16
-                    // TODO: use multiple threads
-                    char * src1i = (char *) src1->data + i13*nb13 + i12*nb12;
-                    if (src1_cont_rows) {
-                        if (src1_cont_cols) {
-                            ggml_fp32_to_fp16_row((float *) src1i, tmp, ne10*ne11);
-                        }
-                        else {
-                            for (int64_t i11 = 0; i11 < ne11; i11++) {
-                                ggml_fp32_to_fp16_row((float *) (src1i + i11*nb11), tmp + i11*ne10, ne10);
-                            }
-                        }
-                    }
-                    else {
-                        for (int64_t i11 = 0; i11 < ne11; i11++) {
-                            for (int64_t i10 = 0; i10 < ne10; i10++) {
-                                // very slow due to no inlining
-                                tmp[i11*ne10 + i10] = ggml_fp32_to_fp16(*(float *) (src1i + i11*nb11 + i10*nb10));
-                            }
-                        }
-                    }
-
-                    // copy src1 to device
-                    CL_CHECK(clEnqueueWriteBuffer(queue, d_Y, false, 0, sizeof(ggml_fp16_t) * y_ne, tmp, 0, NULL, NULL));
-
-                    CL_CHECK(clFinish(queue));
-
-                    // compute
-                    cl_event ev_sgemm;
-                    clblast::StatusCode status = clblast::Gemm<cl_half>(clblast::Layout::kColMajor,
-                                                               clblast::Transpose::kYes, clblast::Transpose::kNo,
-                                                               ne01, ne11, ne10,
-                                                               alpha,
-                                                               d_X, x_offset, ne00,
-                                                               d_Y, 0, ne10,
-                                                               beta,
-                                                               d_D, 0, ne01,
-                                                               &queue, &ev_sgemm);
-
-                    if (status != clblast::StatusCode::kSuccess) {
-                        GGML_ASSERT(false);
-                    }
-
-                    // copy dst to host, then convert to float
-                    CL_CHECK(clEnqueueReadBuffer(queue, d_D, true, 0, sizeof(ggml_fp16_t) * d_ne, tmp, 1, &ev_sgemm, NULL));
-
-                    float * d = (float *) ((char *) dst->data + i12*nb2 + i13*nb3);
-
-                    ggml_fp16_to_fp32_row(tmp, d, d_ne);
-                }
-            }
-        }
-    }
-
-    if (src0->backend != GGML_BACKEND_GPU) {
-        ggml_cl_pool_free(d_X, x_size);
-    }
-    ggml_cl_pool_free(d_Y, y_size);
-    ggml_cl_pool_free(d_D, d_size);
-}
-
-static void ggml_cl_mul_mat_q_f32(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
-    const int64_t ne00 = src0->ne[0];
-    const int64_t ne01 = src0->ne[1];
-    const int64_t ne02 = src0->ne[2];
-    const int64_t ne03 = src0->ne[3];
-
-    const int64_t ne10 = src1->ne[0];
-    const int64_t ne11 = src1->ne[1];
-    const int64_t ne12 = src1->ne[2];
-    const int64_t ne13 = src1->ne[3];
-
-    const int nb2  = dst->nb[2];
-    const int nb3  = dst->nb[3];
-    const ggml_type type = src0->type;
-    const bool mul_mat_vec = ne11 == 1 && ne00%2 == 0;
-
-    const int64_t r2 = ne12 / ne02;
-    const int64_t r3 = ne13 / ne03;
-
-    const float alpha = 1.0f;
-    const float beta = 0.0f;
-    const int x_ne = ne01 * ne00;
-    const int y_ne = ne11 * ne10;
-    const int d_ne = ne11 * ne01;
-    const int x_bps = x_ne / ggml_blck_size(type); // blocks per 2D slice
-    const size_t q_sz = ggml_type_size(type) * x_bps;
-
-    size_t x_size;
-    size_t y_size;
-    size_t d_size;
-    size_t q_size;
-    cl_mem d_X;
-    if (!mul_mat_vec) {
-        d_X = ggml_cl_pool_malloc(sizeof(float) * x_ne, &x_size);
-    }
-    cl_mem d_Y = ggml_cl_pool_malloc(sizeof(float) * y_ne, &y_size);
-    cl_mem d_D = ggml_cl_pool_malloc(sizeof(float) * d_ne, &d_size);
-    cl_mem d_Q;
-    if (src0->backend == GGML_BACKEND_CPU) {
-        d_Q = ggml_cl_pool_malloc(q_sz, &q_size);
-    }
-
-    cl_kernel* to_fp32_cl = ggml_get_to_fp32_cl(type);
-    cl_kernel* dmmv = ggml_get_dequantize_mul_mat_vec_cl(type);
-    GGML_ASSERT(to_fp32_cl != nullptr);
-
-    const size_t global_denom = ggml_cl_global_denom(type);
-    const size_t local = mul_mat_vec ? CL_DMMV_LOCAL_SIZE : ggml_cl_local_size(type);
-
-    size_t ev_idx = 0;
-    std::vector<cl_event> events;
-
-    for (int64_t i03 = 0; i03 < ne03; i03++) {
-        // TODO: copy and dequantize src0 here when r3>1
-        for (int64_t i13 = i03 * r3, e13 = i13 + r3; i13 < e13; i13++) {
-            for (int64_t i02 = 0; i02 < ne02; i02++) {
-                // copy src0 to device if necessary
-                if (src0->backend == GGML_BACKEND_CPU) {
-                    events.emplace_back();
-                    CL_CHECK(ggml_cl_h2d_tensor_2d(queue, d_Q, 0, src0, i03, i02, events.data() + ev_idx++));
-                } else if (src0->backend == GGML_BACKEND_GPU) {
-                    d_Q = (cl_mem) src0->extra;
-                } else {
-                    GGML_ASSERT(false);
-                }
-
-                if (!mul_mat_vec) {
-                    // convert src0 to fp32 on device
-                    const size_t global = x_ne / global_denom;
-                    const size_t offset = src0->backend == GGML_BACKEND_GPU ? (i03 * ne02 + i02) * x_bps : 0;
-                    CL_CHECK(clSetKernelArg(*to_fp32_cl, 0, sizeof(cl_mem), &d_Q));
-                    CL_CHECK(clSetKernelArg(*to_fp32_cl, 1, sizeof(cl_mem), &d_X));
-                    CL_CHECK(clEnqueueNDRangeKernel(queue, *to_fp32_cl, 1, &offset, &global, local > 0 ? &local : NULL, events.size(), !events.empty() ? events.data() : NULL, NULL));
-                }
-
-                for (int64_t i12 = i02 * r2, e12 = i12 + r2; i12 < e12; i12++) {
-                    if (mul_mat_vec) { // specialized dequantize_mul_mat_vec kernel
-                        // copy src1 to device
-                        events.emplace_back();
-                        CL_CHECK(ggml_cl_h2d_tensor_2d(queue, d_Y, 0, src1, i13, i12, events.data() + ev_idx++));
-
-                        // compute
-                        const size_t global = ne01 * local;
-                        const size_t offset = src0->backend == GGML_BACKEND_GPU ? (i03 * ne02 + i02) * x_bps : 0;
-                        const cl_int ncols = ne00;
-                        events.emplace_back();
-                        CL_CHECK(clSetKernelArg(*dmmv, 0, sizeof(cl_mem), &d_Q));
-                        CL_CHECK(clSetKernelArg(*dmmv, 1, sizeof(float) * local, NULL));
-                        CL_CHECK(clSetKernelArg(*dmmv, 2, sizeof(cl_mem), &d_Y));
-                        CL_CHECK(clSetKernelArg(*dmmv, 3, sizeof(cl_mem), &d_D));
-                        CL_CHECK(clSetKernelArg(*dmmv, 4, sizeof(cl_int), &ncols));
-                        CL_CHECK(clEnqueueNDRangeKernel(queue, *dmmv, 1, &offset, &global, &local, events.size() - 1, events.data(), events.data() + ev_idx++));
-                    } else { // CLBlast matrix matrix multiplication
-                        // copy src1 to device
-                        CL_CHECK(ggml_cl_h2d_tensor_2d(queue, d_Y, 0, src1, i13, i12, NULL));
-
-                        // wait for conversion
-                        CL_CHECK(clFinish(queue));
-
-                        // compute
-                        events.emplace_back();
-                        clblast::StatusCode status = clblast::Gemm<cl_float>(clblast::Layout::kColMajor,
-                                                                   clblast::Transpose::kYes, clblast::Transpose::kNo,
-                                                                   ne01, ne11, ne10,
-                                                                   alpha,
-                                                                   d_X, 0, ne00,
-                                                                   d_Y, 0, ne10,
-                                                                   beta,
-                                                                   d_D, 0, ne01,
-                                                                   &queue, events.data() + ev_idx++);
-
-                        if (status != clblast::StatusCode::kSuccess) {
-                            GGML_ASSERT(false);
-                        }
-                    }
-
-                    // copy dst to host
-                    float * d = (float *) ((char *) dst->data + i12*nb2 + i13*nb3);
-                    CL_CHECK(clEnqueueReadBuffer(queue, d_D, true, 0, sizeof(float) * d_ne, d, 1, &events[events.size() - 1], NULL));
-                    for (auto *event : events) {
-                        clReleaseEvent(event);
-                    }
-
-                    ev_idx = 0;
-                    events.clear();
-                }
-            }
-        }
-    }
-
-    if (!mul_mat_vec) {
-        ggml_cl_pool_free(d_X, x_size);
-    }
-    ggml_cl_pool_free(d_Y, y_size);
-    ggml_cl_pool_free(d_D, d_size);
-    if (src0->backend == GGML_BACKEND_CPU) {
-        ggml_cl_pool_free(d_Q, q_size);
-    }
-}
-
-
-bool ggml_cl_can_mul_mat(const struct ggml_tensor * src0, const struct ggml_tensor * src1, struct ggml_tensor * dst) {
-    const int64_t ne10 = src1->ne[0];
-
-    const int64_t ne0 = dst->ne[0];
-    const int64_t ne1 = dst->ne[1];
-
-    // TODO: find the optimal values for these
-    if ((src0->type == GGML_TYPE_F32 || src0->type == GGML_TYPE_F16 || ggml_is_quantized(src0->type)) &&
-        src1->type == GGML_TYPE_F32 &&
-        dst->type == GGML_TYPE_F32 &&
-        ((ne0 >= 32 && ne1 >= 32 && ne10 >= 32) || src0->backend == GGML_BACKEND_GPU)) {
-        return true;
-    }
-
-    return false;
-}
-
-static bool ggml_cl_mul_mat_use_f16(const struct ggml_tensor * src0, const struct ggml_tensor * src1, struct ggml_tensor * /* dst */) {
-    // If device doesn't support FP16
-    if (!fp16_support) {
-        return false;
-    }
-
-    size_t src0_sz = ggml_nbytes(src0);
-    size_t src1_sz = ggml_nbytes(src1);
-
-    // mul_mat_q: src0 is converted to fp32 on device
-    size_t mul_mat_q_transfer = src0_sz + src1_sz;
-
-    // mul_mat_f16: src1 is converted to fp16 on cpu
-    size_t mul_mat_f16_transfer = src0_sz + sizeof(ggml_fp16_t) * ggml_nelements(src1);
-
-    // choose the smaller one to transfer to the device
-    // TODO: this is not always the best choice due to the overhead of converting to fp16
-    return mul_mat_f16_transfer < mul_mat_q_transfer;
-}
-
-void ggml_cl_mul_mat(const struct ggml_tensor * src0, const struct ggml_tensor * src1, struct ggml_tensor * dst, void * wdata, size_t wsize) {
-    GGML_ASSERT(ggml_cl_can_mul_mat(src0, src1, dst));
-
-    if (src0->type == GGML_TYPE_F32) {
-        ggml_cl_mul_mat_f32(src0, src1, dst);
-    }
-    else if (src0->type == GGML_TYPE_F16) {
-        if (ggml_cl_mul_mat_use_f16(src0, src1, dst)) {
-            ggml_cl_mul_mat_f16(src0, src1, dst, wdata, wsize);
-        }
-        else {
-            ggml_cl_mul_mat_q_f32(src0, src1, dst);
-        }
-    }
-    else if (ggml_is_quantized(src0->type)) {
-        ggml_cl_mul_mat_q_f32(src0, src1, dst);
-    }
-    else {
-        GGML_ASSERT(false);
-    }
-}
-
-size_t ggml_cl_mul_mat_get_wsize(const struct ggml_tensor * src0, const struct ggml_tensor * src1, struct ggml_tensor * dst) {
-    if (src0->type == GGML_TYPE_F16 && ggml_cl_mul_mat_use_f16(src0, src1, dst)) {
-        return sizeof(ggml_fp16_t) * std::max(src1->ne[0] * src1->ne[1], dst->ne[0] * dst->ne[1]);
-    }
-    return 0;
-}
-
-void ggml_cl_transform_tensor(void * data, ggml_tensor * tensor) {
-    const int64_t ne0 = tensor->ne[0];
-    const int64_t ne1 = tensor->ne[1];
-    const int64_t ne2 = tensor->ne[2];
-    const int64_t ne3 = tensor->ne[3];
-
-    const ggml_type type = tensor->type;
-    const size_t s_sz = ggml_type_size(type) * (size_t) (ne0 * ne1 / ggml_blck_size(type));
-    const size_t q_sz = s_sz * (size_t) (ne2 * ne3);
-
-    size_t q_size;
-    cl_mem dst = ggml_cl_pool_malloc(q_sz, &q_size);
-
-    tensor->data = data;
-    // copy tensor to device
-    size_t offset = 0;
-    for (int64_t i3 = 0; i3 < ne3; i3++) {
-        for (int64_t i2 = 0; i2 < ne2; i2++) {
-            CL_CHECK(ggml_cl_h2d_tensor_2d(queue, dst, offset, tensor, i3, i2, NULL));
-            offset += s_sz;
-        }
-    }
-
-    CL_CHECK(clFinish(queue));
-
-    tensor->extra = dst;
-    GGML_ASSERT(tensor->backend == GGML_BACKEND_GPU);
-}
diff --git a/src/whisper_cpp/ggml-opencl.h b/src/whisper_cpp/ggml-opencl.h
deleted file mode 100644
index 44d05bd6..00000000
--- a/src/whisper_cpp/ggml-opencl.h
+++ /dev/null
@@ -1,25 +0,0 @@
-#pragma once
-
-#include "ggml.h"
-
-#ifdef  __cplusplus
-extern "C" {
-#endif
-
-GGML_API void ggml_cl_init(void);
-
-GGML_API void   ggml_cl_mul(const struct ggml_tensor * src0, const struct ggml_tensor * src1, struct ggml_tensor * dst);
-GGML_API bool   ggml_cl_can_mul_mat(const struct ggml_tensor * src0, const struct ggml_tensor * src1, struct ggml_tensor * dst);
-GGML_API size_t ggml_cl_mul_mat_get_wsize(const struct ggml_tensor * src0, const struct ggml_tensor * src1, struct ggml_tensor * dst);
-GGML_API void   ggml_cl_mul_mat(const struct ggml_tensor * src0, const struct ggml_tensor * src1, struct ggml_tensor * dst, void * wdata, size_t wsize);
-
-GGML_API void * ggml_cl_host_malloc(size_t size);
-GGML_API void   ggml_cl_host_free(void * ptr);
-
-GGML_API void ggml_cl_free_data(const struct ggml_tensor* tensor);
-
-GGML_API void ggml_cl_transform_tensor(void * data, struct ggml_tensor * tensor);
-
-#ifdef  __cplusplus
-}
-#endif
diff --git a/src/whisper_cpp/ggml-quants.c b/src/whisper_cpp/ggml-quants.c
deleted file mode 100644
index 55a9496d..00000000
--- a/src/whisper_cpp/ggml-quants.c
+++ /dev/null
@@ -1,7067 +0,0 @@
-#include "ggml-quants.h"
-#include "ggml-impl.h"
-
-#include <math.h>
-#include <string.h>
-#include <assert.h>
-#include <float.h>
-
-#ifdef __ARM_NEON
-
-// if YCM cannot find <arm_neon.h>, make a symbolic link to it, for example:
-//
-//   $ ln -sfn /Library/Developer/CommandLineTools/usr/lib/clang/13.1.6/include/arm_neon.h ./src/
-//
-#include <arm_neon.h>
-
-#else
-
-#ifdef __wasm_simd128__
-#include <wasm_simd128.h>
-#else
-#if defined(__POWER9_VECTOR__) || defined(__powerpc64__)
-#include <altivec.h>
-#undef bool
-#define bool _Bool
-#else
-#if defined(_MSC_VER) || defined(__MINGW32__)
-#include <intrin.h>
-#else
-#if defined(__AVX__) || defined(__AVX2__) || defined(__AVX512F__) || defined(__SSSE3__) || defined(__SSE3__)
-#if !defined(__riscv)
-#include <immintrin.h>
-#endif
-#endif
-#endif
-#endif
-#endif
-#endif
-
-#ifdef __riscv_v_intrinsic
-#include <riscv_vector.h>
-#endif
-
-#undef MIN
-#undef MAX
-
-#define MIN(a, b) ((a) < (b) ? (a) : (b))
-#define MAX(a, b) ((a) > (b) ? (a) : (b))
-
-#define MM256_SET_M128I(a, b) _mm256_insertf128_si256(_mm256_castsi128_si256(b), (a), 1)
-
-#if defined(__AVX__) || defined(__AVX2__) || defined(__AVX512F__) || defined(__SSSE3__)
-// multiply int8_t, add results pairwise twice
-static inline __m128i mul_sum_i8_pairs(const __m128i x, const __m128i y) {
-    // Get absolute values of x vectors
-    const __m128i ax = _mm_sign_epi8(x, x);
-    // Sign the values of the y vectors
-    const __m128i sy = _mm_sign_epi8(y, x);
-    // Perform multiplication and create 16-bit values
-    const __m128i dot = _mm_maddubs_epi16(ax, sy);
-    const __m128i ones = _mm_set1_epi16(1);
-    return _mm_madd_epi16(ones, dot);
-}
-
-#if __AVX__ || __AVX2__ || __AVX512F__
-// horizontally add 8 floats
-static inline float hsum_float_8(const __m256 x) {
-    __m128 res = _mm256_extractf128_ps(x, 1);
-    res = _mm_add_ps(res, _mm256_castps256_ps128(x));
-    res = _mm_add_ps(res, _mm_movehl_ps(res, res));
-    res = _mm_add_ss(res, _mm_movehdup_ps(res));
-    return _mm_cvtss_f32(res);
-}
-
-// horizontally add 8 int32_t
-static inline int hsum_i32_8(const __m256i a) {
-    const __m128i sum128 = _mm_add_epi32(_mm256_castsi256_si128(a), _mm256_extractf128_si256(a, 1));
-    const __m128i hi64 = _mm_unpackhi_epi64(sum128, sum128);
-    const __m128i sum64 = _mm_add_epi32(hi64, sum128);
-    const __m128i hi32  = _mm_shuffle_epi32(sum64, _MM_SHUFFLE(2, 3, 0, 1));
-    return _mm_cvtsi128_si32(_mm_add_epi32(sum64, hi32));
-}
-
-// horizontally add 4 int32_t
-static inline int hsum_i32_4(const __m128i a) {
-    const __m128i hi64 = _mm_unpackhi_epi64(a, a);
-    const __m128i sum64 = _mm_add_epi32(hi64, a);
-    const __m128i hi32  = _mm_shuffle_epi32(sum64, _MM_SHUFFLE(2, 3, 0, 1));
-    return _mm_cvtsi128_si32(_mm_add_epi32(sum64, hi32));
-}
-
-#if defined(__AVX2__) || defined(__AVX512F__)
-// spread 32 bits to 32 bytes { 0x00, 0xFF }
-static inline __m256i bytes_from_bits_32(const uint8_t * x) {
-    uint32_t x32;
-    memcpy(&x32, x, sizeof(uint32_t));
-    const __m256i shuf_mask = _mm256_set_epi64x(
-            0x0303030303030303, 0x0202020202020202,
-            0x0101010101010101, 0x0000000000000000);
-    __m256i bytes = _mm256_shuffle_epi8(_mm256_set1_epi32(x32), shuf_mask);
-    const __m256i bit_mask = _mm256_set1_epi64x(0x7fbfdfeff7fbfdfe);
-    bytes = _mm256_or_si256(bytes, bit_mask);
-    return _mm256_cmpeq_epi8(bytes, _mm256_set1_epi64x(-1));
-}
-
-// Unpack 32 4-bit fields into 32 bytes
-// The output vector contains 32 bytes, each one in [ 0 .. 15 ] interval
-static inline __m256i bytes_from_nibbles_32(const uint8_t * rsi)
-{
-    const __m128i tmp = _mm_loadu_si128((const __m128i *)rsi);
-    const __m256i bytes = MM256_SET_M128I(_mm_srli_epi16(tmp, 4), tmp);
-    const __m256i lowMask = _mm256_set1_epi8( 0xF );
-    return _mm256_and_si256(lowMask, bytes);
-}
-
-// add int16_t pairwise and return as float vector
-static inline __m256 sum_i16_pairs_float(const __m256i x) {
-    const __m256i ones = _mm256_set1_epi16(1);
-    const __m256i summed_pairs = _mm256_madd_epi16(ones, x);
-    return _mm256_cvtepi32_ps(summed_pairs);
-}
-
-static inline __m256 mul_sum_us8_pairs_float(const __m256i ax, const __m256i sy) {
-#if __AVXVNNI__
-    const __m256i zero = _mm256_setzero_si256();
-    const __m256i summed_pairs = _mm256_dpbusd_epi32(zero, ax, sy);
-    return _mm256_cvtepi32_ps(summed_pairs);
-#else
-    // Perform multiplication and create 16-bit values
-    const __m256i dot = _mm256_maddubs_epi16(ax, sy);
-    return sum_i16_pairs_float(dot);
-#endif
-}
-
-// multiply int8_t, add results pairwise twice and return as float vector
-static inline __m256 mul_sum_i8_pairs_float(const __m256i x, const __m256i y) {
-#if __AVXVNNIINT8__
-    const __m256i zero = _mm256_setzero_si256();
-    const __m256i summed_pairs = _mm256_dpbssd_epi32(zero, x, y);
-    return _mm256_cvtepi32_ps(summed_pairs);
-#else
-    // Get absolute values of x vectors
-    const __m256i ax = _mm256_sign_epi8(x, x);
-    // Sign the values of the y vectors
-    const __m256i sy = _mm256_sign_epi8(y, x);
-    return mul_sum_us8_pairs_float(ax, sy);
-#endif
-}
-
-static inline __m128i packNibbles( __m256i bytes )
-{
-    // Move bits within 16-bit lanes from 0000_abcd_0000_efgh into 0000_0000_abcd_efgh
-#if __AVX512F__
-    const __m256i bytes_srli_4 = _mm256_srli_epi16(bytes, 4);   // 0000_0000_abcd_0000
-    bytes = _mm256_or_si256(bytes, bytes_srli_4);               // 0000_abcd_abcd_efgh
-    return _mm256_cvtepi16_epi8(bytes);                         // abcd_efgh
-#else
-    const __m256i lowByte = _mm256_set1_epi16( 0xFF );
-    __m256i high = _mm256_andnot_si256( lowByte, bytes );
-    __m256i low = _mm256_and_si256( lowByte, bytes );
-    high = _mm256_srli_epi16( high, 4 );
-    bytes = _mm256_or_si256( low, high );
-
-    // Compress uint16_t lanes into bytes
-    __m128i r0 = _mm256_castsi256_si128( bytes );
-    __m128i r1 = _mm256_extracti128_si256( bytes, 1 );
-    return _mm_packus_epi16( r0, r1 );
-#endif
-}
-#elif defined(__AVX__)
-// spread 32 bits to 32 bytes { 0x00, 0xFF }
-static inline __m256i bytes_from_bits_32(const uint8_t * x) {
-    uint32_t x32;
-    memcpy(&x32, x, sizeof(uint32_t));
-    const __m128i shuf_maskl = _mm_set_epi64x(0x0101010101010101, 0x0000000000000000);
-    const __m128i shuf_maskh = _mm_set_epi64x(0x0303030303030303, 0x0202020202020202);
-    __m128i bytesl = _mm_shuffle_epi8(_mm_set1_epi32(x32), shuf_maskl);
-    __m128i bytesh = _mm_shuffle_epi8(_mm_set1_epi32(x32), shuf_maskh);
-    const __m128i bit_mask = _mm_set1_epi64x(0x7fbfdfeff7fbfdfe);
-    bytesl = _mm_or_si128(bytesl, bit_mask);
-    bytesh = _mm_or_si128(bytesh, bit_mask);
-    bytesl = _mm_cmpeq_epi8(bytesl, _mm_set1_epi64x(-1));
-    bytesh = _mm_cmpeq_epi8(bytesh, _mm_set1_epi64x(-1));
-    return MM256_SET_M128I(bytesh, bytesl);
-}
-
-// Unpack 32 4-bit fields into 32 bytes
-// The output vector contains 32 bytes, each one in [ 0 .. 15 ] interval
-static inline __m256i bytes_from_nibbles_32(const uint8_t * rsi)
-{
-    // Load 16 bytes from memory
-    __m128i tmpl = _mm_loadu_si128((const __m128i *)rsi);
-    __m128i tmph = _mm_srli_epi16(tmpl, 4);
-    const __m128i lowMask = _mm_set1_epi8(0xF);
-    tmpl = _mm_and_si128(lowMask, tmpl);
-    tmph = _mm_and_si128(lowMask, tmph);
-    return MM256_SET_M128I(tmph, tmpl);
-}
-
-// add int16_t pairwise and return as float vector
-static inline __m256 sum_i16_pairs_float(const __m128i xh, const __m128i xl) {
-    const __m128i ones = _mm_set1_epi16(1);
-    const __m128i summed_pairsl = _mm_madd_epi16(ones, xl);
-    const __m128i summed_pairsh = _mm_madd_epi16(ones, xh);
-    const __m256i summed_pairs = MM256_SET_M128I(summed_pairsh, summed_pairsl);
-    return _mm256_cvtepi32_ps(summed_pairs);
-}
-
-static inline __m256 mul_sum_us8_pairs_float(const __m256i ax, const __m256i sy) {
-    const __m128i axl = _mm256_castsi256_si128(ax);
-    const __m128i axh = _mm256_extractf128_si256(ax, 1);
-    const __m128i syl = _mm256_castsi256_si128(sy);
-    const __m128i syh = _mm256_extractf128_si256(sy, 1);
-    // Perform multiplication and create 16-bit values
-    const __m128i dotl = _mm_maddubs_epi16(axl, syl);
-    const __m128i doth = _mm_maddubs_epi16(axh, syh);
-    return sum_i16_pairs_float(doth, dotl);
-}
-
-// multiply int8_t, add results pairwise twice and return as float vector
-static inline __m256 mul_sum_i8_pairs_float(const __m256i x, const __m256i y) {
-    const __m128i xl = _mm256_castsi256_si128(x);
-    const __m128i xh = _mm256_extractf128_si256(x, 1);
-    const __m128i yl = _mm256_castsi256_si128(y);
-    const __m128i yh = _mm256_extractf128_si256(y, 1);
-    // Get absolute values of x vectors
-    const __m128i axl = _mm_sign_epi8(xl, xl);
-    const __m128i axh = _mm_sign_epi8(xh, xh);
-    // Sign the values of the y vectors
-    const __m128i syl = _mm_sign_epi8(yl, xl);
-    const __m128i syh = _mm_sign_epi8(yh, xh);
-    // Perform multiplication and create 16-bit values
-    const __m128i dotl = _mm_maddubs_epi16(axl, syl);
-    const __m128i doth = _mm_maddubs_epi16(axh, syh);
-    return sum_i16_pairs_float(doth, dotl);
-}
-
-static inline __m128i packNibbles( __m128i bytes1, __m128i bytes2 )
-{
-    // Move bits within 16-bit lanes from 0000_abcd_0000_efgh into 0000_0000_abcd_efgh
-    const __m128i lowByte = _mm_set1_epi16( 0xFF );
-    __m128i high = _mm_andnot_si128( lowByte, bytes1 );
-    __m128i low = _mm_and_si128( lowByte, bytes1 );
-    high = _mm_srli_epi16( high, 4 );
-    bytes1 = _mm_or_si128( low, high );
-    high = _mm_andnot_si128( lowByte, bytes2 );
-    low = _mm_and_si128( lowByte, bytes2 );
-    high = _mm_srli_epi16( high, 4 );
-    bytes2 = _mm_or_si128( low, high );
-
-    return _mm_packus_epi16( bytes1, bytes2);
-}
-#endif
-#elif defined(__SSSE3__)
-// horizontally add 4x4 floats
-static inline float hsum_float_4x4(const __m128 a, const __m128 b, const __m128 c, const __m128 d) {
-    __m128 res_0 =_mm_hadd_ps(a, b);
-    __m128 res_1 =_mm_hadd_ps(c, d);
-    __m128 res =_mm_hadd_ps(res_0, res_1);
-    res =_mm_hadd_ps(res, res);
-    res =_mm_hadd_ps(res, res);
-
-    return _mm_cvtss_f32(res);
-}
-#endif // __AVX__ || __AVX2__ || __AVX512F__
-#endif // defined(__AVX__) || defined(__AVX2__) || defined(__AVX512F__) || defined(__SSSE3__)
-
-#if defined(__ARM_NEON)
-#if !defined(__aarch64__)
-
-// 64-bit compatibility
-
-// vaddvq_s16
-// vpaddq_s16
-// vaddvq_s32
-// vaddvq_f32
-// vmaxvq_f32
-// vcvtnq_s32_f32
-
-inline static int32_t vaddvq_s16(int16x8_t v) {
-    return
-        (int32_t)vgetq_lane_s16(v, 0) + (int32_t)vgetq_lane_s16(v, 1) +
-        (int32_t)vgetq_lane_s16(v, 2) + (int32_t)vgetq_lane_s16(v, 3) +
-        (int32_t)vgetq_lane_s16(v, 4) + (int32_t)vgetq_lane_s16(v, 5) +
-        (int32_t)vgetq_lane_s16(v, 6) + (int32_t)vgetq_lane_s16(v, 7);
-}
-
-inline static int16x8_t vpaddq_s16(int16x8_t a, int16x8_t b) {
-    int16x4_t a0 = vpadd_s16(vget_low_s16(a), vget_high_s16(a));
-    int16x4_t b0 = vpadd_s16(vget_low_s16(b), vget_high_s16(b));
-    return vcombine_s16(a0, b0);
-}
-
-inline static int32_t vaddvq_s32(int32x4_t v) {
-    return vgetq_lane_s32(v, 0) + vgetq_lane_s32(v, 1) + vgetq_lane_s32(v, 2) + vgetq_lane_s32(v, 3);
-}
-
-inline static float vaddvq_f32(float32x4_t v) {
-    return vgetq_lane_f32(v, 0) + vgetq_lane_f32(v, 1) + vgetq_lane_f32(v, 2) + vgetq_lane_f32(v, 3);
-}
-
-inline static float vmaxvq_f32(float32x4_t v) {
-    return
-        MAX(MAX(vgetq_lane_f32(v, 0), vgetq_lane_f32(v, 1)),
-            MAX(vgetq_lane_f32(v, 2), vgetq_lane_f32(v, 3)));
-}
-
-inline static int32x4_t vcvtnq_s32_f32(float32x4_t v) {
-    int32x4_t res;
-
-    res[0] = roundf(vgetq_lane_f32(v, 0));
-    res[1] = roundf(vgetq_lane_f32(v, 1));
-    res[2] = roundf(vgetq_lane_f32(v, 2));
-    res[3] = roundf(vgetq_lane_f32(v, 3));
-
-    return res;
-}
-
-// vld1q_s16_x2
-// vld1q_u8_x2
-// vld1q_u8_x4
-// vld1q_s8_x2
-// vld1q_s8_x4
-// TODO: double-check these work correctly
-
-typedef struct ggml_int16x8x2_t {
-    int16x8_t val[2];
-} ggml_int16x8x2_t;
-
-inline static ggml_int16x8x2_t ggml_vld1q_s16_x2(const int16_t * ptr) {
-    ggml_int16x8x2_t res;
-
-    res.val[0] = vld1q_s16(ptr + 0);
-    res.val[1] = vld1q_s16(ptr + 8);
-
-    return res;
-}
-
-typedef struct ggml_uint8x16x2_t {
-    uint8x16_t val[2];
-} ggml_uint8x16x2_t;
-
-inline static ggml_uint8x16x2_t ggml_vld1q_u8_x2(const uint8_t * ptr) {
-    ggml_uint8x16x2_t res;
-
-    res.val[0] = vld1q_u8(ptr + 0);
-    res.val[1] = vld1q_u8(ptr + 16);
-
-    return res;
-}
-
-typedef struct ggml_uint8x16x4_t {
-    uint8x16_t val[4];
-} ggml_uint8x16x4_t;
-
-inline static ggml_uint8x16x4_t ggml_vld1q_u8_x4(const uint8_t * ptr) {
-    ggml_uint8x16x4_t res;
-
-    res.val[0] = vld1q_u8(ptr + 0);
-    res.val[1] = vld1q_u8(ptr + 16);
-    res.val[2] = vld1q_u8(ptr + 32);
-    res.val[3] = vld1q_u8(ptr + 48);
-
-    return res;
-}
-
-typedef struct ggml_int8x16x2_t {
-    int8x16_t val[2];
-} ggml_int8x16x2_t;
-
-inline static ggml_int8x16x2_t ggml_vld1q_s8_x2(const int8_t * ptr) {
-    ggml_int8x16x2_t res;
-
-    res.val[0] = vld1q_s8(ptr + 0);
-    res.val[1] = vld1q_s8(ptr + 16);
-
-    return res;
-}
-
-typedef struct ggml_int8x16x4_t {
-    int8x16_t val[4];
-} ggml_int8x16x4_t;
-
-inline static ggml_int8x16x4_t ggml_vld1q_s8_x4(const int8_t * ptr) {
-    ggml_int8x16x4_t res;
-
-    res.val[0] = vld1q_s8(ptr + 0);
-    res.val[1] = vld1q_s8(ptr + 16);
-    res.val[2] = vld1q_s8(ptr + 32);
-    res.val[3] = vld1q_s8(ptr + 48);
-
-    return res;
-}
-
-#else
-
-#define ggml_int16x8x2_t  int16x8x2_t
-#define ggml_uint8x16x2_t uint8x16x2_t
-#define ggml_uint8x16x4_t uint8x16x4_t
-#define ggml_int8x16x2_t  int8x16x2_t
-#define ggml_int8x16x4_t  int8x16x4_t
-
-#define ggml_vld1q_s16_x2 vld1q_s16_x2
-#define ggml_vld1q_u8_x2  vld1q_u8_x2
-#define ggml_vld1q_u8_x4  vld1q_u8_x4
-#define ggml_vld1q_s8_x2  vld1q_s8_x2
-#define ggml_vld1q_s8_x4  vld1q_s8_x4
-
-#endif
-
-#if !defined(__ARM_FEATURE_DOTPROD)
-
-inline static int32x4_t ggml_vdotq_s32(int32x4_t acc, int8x16_t a, int8x16_t b) {
-    const int16x8_t p0 = vmull_s8(vget_low_s8 (a), vget_low_s8 (b));
-    const int16x8_t p1 = vmull_s8(vget_high_s8(a), vget_high_s8(b));
-
-    return vaddq_s32(acc, vaddq_s32(vpaddlq_s16(p0), vpaddlq_s16(p1)));
-}
-
-#else
-
-#define ggml_vdotq_s32(a, b, c) vdotq_s32(a, b, c)
-
-#endif
-
-#endif
-
-#if defined(__ARM_NEON) || defined(__wasm_simd128__)
-#define B1(c,s,n)  0x ## n ## c ,  0x ## n ## s
-#define B2(c,s,n) B1(c,s,n ## c), B1(c,s,n ## s)
-#define B3(c,s,n) B2(c,s,n ## c), B2(c,s,n ## s)
-#define B4(c,s,n) B3(c,s,n ## c), B3(c,s,n ## s)
-#define B5(c,s,n) B4(c,s,n ## c), B4(c,s,n ## s)
-#define B6(c,s,n) B5(c,s,n ## c), B5(c,s,n ## s)
-#define B7(c,s,n) B6(c,s,n ## c), B6(c,s,n ## s)
-#define B8(c,s  ) B7(c,s,     c), B7(c,s,     s)
-
-// precomputed tables for expanding 8bits to 8 bytes:
-static const uint64_t table_b2b_0[1 << 8] = { B8(00, 10) }; // ( b) << 4
-static const uint64_t table_b2b_1[1 << 8] = { B8(10, 00) }; // (!b) << 4
-#endif
-
-// reference implementation for deterministic creation of model files
-void quantize_row_q4_0_reference(const float * restrict x, block_q4_0 * restrict y, int k) {
-    static const int qk = QK4_0;
-
-    assert(k % qk == 0);
-
-    const int nb = k / qk;
-
-    for (int i = 0; i < nb; i++) {
-        float amax = 0.0f; // absolute max
-        float max  = 0.0f;
-
-        for (int j = 0; j < qk; j++) {
-            const float v = x[i*qk + j];
-            if (amax < fabsf(v)) {
-                amax = fabsf(v);
-                max  = v;
-            }
-        }
-
-        const float d  = max / -8;
-        const float id = d ? 1.0f/d : 0.0f;
-
-        y[i].d = GGML_FP32_TO_FP16(d);
-
-        for (int j = 0; j < qk/2; ++j) {
-            const float x0 = x[i*qk + 0    + j]*id;
-            const float x1 = x[i*qk + qk/2 + j]*id;
-
-            const uint8_t xi0 = MIN(15, (int8_t)(x0 + 8.5f));
-            const uint8_t xi1 = MIN(15, (int8_t)(x1 + 8.5f));
-
-            y[i].qs[j]  = xi0;
-            y[i].qs[j] |= xi1 << 4;
-        }
-    }
-}
-
-void quantize_row_q4_0(const float * restrict x, void * restrict y, int k) {
-    quantize_row_q4_0_reference(x, y, k);
-}
-
-void quantize_row_q4_1_reference(const float * restrict x, block_q4_1 * restrict y, int k) {
-    const int qk = QK4_1;
-
-    assert(k % qk == 0);
-
-    const int nb = k / qk;
-
-    for (int i = 0; i < nb; i++) {
-        float min = FLT_MAX;
-        float max = -FLT_MAX;
-
-        for (int j = 0; j < qk; j++) {
-            const float v = x[i*qk + j];
-
-            if (v < min) min = v;
-            if (v > max) max = v;
-        }
-
-        const float d  = (max - min) / ((1 << 4) - 1);
-        const float id = d ? 1.0f/d : 0.0f;
-
-        y[i].d = GGML_FP32_TO_FP16(d);
-        y[i].m = GGML_FP32_TO_FP16(min);
-
-        for (int j = 0; j < qk/2; ++j) {
-            const float x0 = (x[i*qk + 0    + j] - min)*id;
-            const float x1 = (x[i*qk + qk/2 + j] - min)*id;
-
-            const uint8_t xi0 = MIN(15, (int8_t)(x0 + 0.5f));
-            const uint8_t xi1 = MIN(15, (int8_t)(x1 + 0.5f));
-
-            y[i].qs[j]  = xi0;
-            y[i].qs[j] |= xi1 << 4;
-        }
-    }
-}
-
-void quantize_row_q4_1(const float * restrict x, void * restrict y, int k) {
-    quantize_row_q4_1_reference(x, y, k);
-}
-
-void quantize_row_q5_0_reference(const float * restrict x, block_q5_0 * restrict y, int k) {
-    static const int qk = QK5_0;
-
-    assert(k % qk == 0);
-
-    const int nb = k / qk;
-
-    for (int i = 0; i < nb; i++) {
-        float amax = 0.0f; // absolute max
-        float max  = 0.0f;
-
-        for (int j = 0; j < qk; j++) {
-            const float v = x[i*qk + j];
-            if (amax < fabsf(v)) {
-                amax = fabsf(v);
-                max  = v;
-            }
-        }
-
-        const float d  = max / -16;
-        const float id = d ? 1.0f/d : 0.0f;
-
-        y[i].d = GGML_FP32_TO_FP16(d);
-
-        uint32_t qh = 0;
-
-        for (int j = 0; j < qk/2; ++j) {
-            const float x0 = x[i*qk + 0    + j]*id;
-            const float x1 = x[i*qk + qk/2 + j]*id;
-
-            const uint8_t xi0 = MIN(31, (int8_t)(x0 + 16.5f));
-            const uint8_t xi1 = MIN(31, (int8_t)(x1 + 16.5f));
-
-            y[i].qs[j] = (xi0 & 0x0F) | ((xi1 & 0x0F) << 4);
-
-            // get the 5-th bit and store it in qh at the right position
-            qh |= ((xi0 & 0x10u) >> 4) << (j + 0);
-            qh |= ((xi1 & 0x10u) >> 4) << (j + qk/2);
-        }
-
-        memcpy(&y[i].qh, &qh, sizeof(qh));
-    }
-}
-
-void quantize_row_q5_0(const float * restrict x, void * restrict y, int k) {
-    quantize_row_q5_0_reference(x, y, k);
-}
-
-void quantize_row_q5_1_reference(const float * restrict x, block_q5_1 * restrict y, int k) {
-    const int qk = QK5_1;
-
-    assert(k % qk == 0);
-
-    const int nb = k / qk;
-
-    for (int i = 0; i < nb; i++) {
-        float min = FLT_MAX;
-        float max = -FLT_MAX;
-
-        for (int j = 0; j < qk; j++) {
-            const float v = x[i*qk + j];
-
-            if (v < min) min = v;
-            if (v > max) max = v;
-        }
-
-        const float d  = (max - min) / ((1 << 5) - 1);
-        const float id = d ? 1.0f/d : 0.0f;
-
-        y[i].d = GGML_FP32_TO_FP16(d);
-        y[i].m = GGML_FP32_TO_FP16(min);
-
-        uint32_t qh = 0;
-
-        for (int j = 0; j < qk/2; ++j) {
-            const float x0 = (x[i*qk + 0    + j] - min)*id;
-            const float x1 = (x[i*qk + qk/2 + j] - min)*id;
-
-            const uint8_t xi0 = (uint8_t)(x0 + 0.5f);
-            const uint8_t xi1 = (uint8_t)(x1 + 0.5f);
-
-            y[i].qs[j] = (xi0 & 0x0F) | ((xi1 & 0x0F) << 4);
-
-            // get the 5-th bit and store it in qh at the right position
-            qh |= ((xi0 & 0x10u) >> 4) << (j + 0);
-            qh |= ((xi1 & 0x10u) >> 4) << (j + qk/2);
-        }
-
-        memcpy(&y[i].qh, &qh, sizeof(y[i].qh));
-    }
-}
-
-void quantize_row_q5_1(const float * restrict x, void * restrict y, int k) {
-    quantize_row_q5_1_reference(x, y, k);
-}
-
-// reference implementation for deterministic creation of model files
-void quantize_row_q8_0_reference(const float * restrict x, block_q8_0 * restrict y, int k) {
-    assert(k % QK8_0 == 0);
-    const int nb = k / QK8_0;
-
-    for (int i = 0; i < nb; i++) {
-        float amax = 0.0f; // absolute max
-
-        for (int j = 0; j < QK8_0; j++) {
-            const float v = x[i*QK8_0 + j];
-            amax = MAX(amax, fabsf(v));
-        }
-
-        const float d = amax / ((1 << 7) - 1);
-        const float id = d ? 1.0f/d : 0.0f;
-
-        y[i].d = GGML_FP32_TO_FP16(d);
-
-        for (int j = 0; j < QK8_0; ++j) {
-            const float x0 = x[i*QK8_0 + j]*id;
-
-            y[i].qs[j] = roundf(x0);
-        }
-    }
-}
-
-void quantize_row_q8_0(const float * restrict x, void * restrict vy, int k) {
-    assert(QK8_0 == 32);
-    assert(k % QK8_0 == 0);
-    const int nb = k / QK8_0;
-
-    block_q8_0 * restrict y = vy;
-
-#if defined(__ARM_NEON)
-    for (int i = 0; i < nb; i++) {
-        float32x4_t srcv [8];
-        float32x4_t asrcv[8];
-        float32x4_t amaxv[8];
-
-        for (int j = 0; j < 8; j++) srcv[j]  = vld1q_f32(x + i*32 + 4*j);
-        for (int j = 0; j < 8; j++) asrcv[j] = vabsq_f32(srcv[j]);
-
-        for (int j = 0; j < 4; j++) amaxv[2*j] = vmaxq_f32(asrcv[2*j], asrcv[2*j+1]);
-        for (int j = 0; j < 2; j++) amaxv[4*j] = vmaxq_f32(amaxv[4*j], amaxv[4*j+2]);
-        for (int j = 0; j < 1; j++) amaxv[8*j] = vmaxq_f32(amaxv[8*j], amaxv[8*j+4]);
-
-        const float amax = vmaxvq_f32(amaxv[0]);
-
-        const float d = amax / ((1 << 7) - 1);
-        const float id = d ? 1.0f/d : 0.0f;
-
-        y[i].d = GGML_FP32_TO_FP16(d);
-
-        for (int j = 0; j < 8; j++) {
-            const float32x4_t v  = vmulq_n_f32(srcv[j], id);
-            const int32x4_t   vi = vcvtnq_s32_f32(v);
-
-            y[i].qs[4*j + 0] = vgetq_lane_s32(vi, 0);
-            y[i].qs[4*j + 1] = vgetq_lane_s32(vi, 1);
-            y[i].qs[4*j + 2] = vgetq_lane_s32(vi, 2);
-            y[i].qs[4*j + 3] = vgetq_lane_s32(vi, 3);
-        }
-    }
-#elif defined(__wasm_simd128__)
-    for (int i = 0; i < nb; i++) {
-        v128_t srcv [8];
-        v128_t asrcv[8];
-        v128_t amaxv[8];
-
-        for (int j = 0; j < 8; j++) srcv[j]  = wasm_v128_load(x + i*32 + 4*j);
-        for (int j = 0; j < 8; j++) asrcv[j] = wasm_f32x4_abs(srcv[j]);
-
-        for (int j = 0; j < 4; j++) amaxv[2*j] = wasm_f32x4_max(asrcv[2*j], asrcv[2*j+1]);
-        for (int j = 0; j < 2; j++) amaxv[4*j] = wasm_f32x4_max(amaxv[4*j], amaxv[4*j+2]);
-        for (int j = 0; j < 1; j++) amaxv[8*j] = wasm_f32x4_max(amaxv[8*j], amaxv[8*j+4]);
-
-        const float amax = MAX(MAX(wasm_f32x4_extract_lane(amaxv[0], 0),
-                                   wasm_f32x4_extract_lane(amaxv[0], 1)),
-                               MAX(wasm_f32x4_extract_lane(amaxv[0], 2),
-                                   wasm_f32x4_extract_lane(amaxv[0], 3)));
-
-        const float d = amax / ((1 << 7) - 1);
-        const float id = d ? 1.0f/d : 0.0f;
-
-        y[i].d = GGML_FP32_TO_FP16(d);
-
-        for (int j = 0; j < 8; j++) {
-            const v128_t v  = wasm_f32x4_mul(srcv[j], wasm_f32x4_splat(id));
-            const v128_t vi = wasm_i32x4_trunc_sat_f32x4(v);
-
-            y[i].qs[4*j + 0] = wasm_i32x4_extract_lane(vi, 0);
-            y[i].qs[4*j + 1] = wasm_i32x4_extract_lane(vi, 1);
-            y[i].qs[4*j + 2] = wasm_i32x4_extract_lane(vi, 2);
-            y[i].qs[4*j + 3] = wasm_i32x4_extract_lane(vi, 3);
-        }
-    }
-#elif defined(__AVX2__) || defined(__AVX__)
-    for (int i = 0; i < nb; i++) {
-        // Load elements into 4 AVX vectors
-        __m256 v0 = _mm256_loadu_ps( x );
-        __m256 v1 = _mm256_loadu_ps( x + 8 );
-        __m256 v2 = _mm256_loadu_ps( x + 16 );
-        __m256 v3 = _mm256_loadu_ps( x + 24 );
-        x += 32;
-
-        // Compute max(abs(e)) for the block
-        const __m256 signBit = _mm256_set1_ps( -0.0f );
-        __m256 maxAbs = _mm256_andnot_ps( signBit, v0 );
-        maxAbs = _mm256_max_ps( maxAbs, _mm256_andnot_ps( signBit, v1 ) );
-        maxAbs = _mm256_max_ps( maxAbs, _mm256_andnot_ps( signBit, v2 ) );
-        maxAbs = _mm256_max_ps( maxAbs, _mm256_andnot_ps( signBit, v3 ) );
-
-        __m128 max4 = _mm_max_ps( _mm256_extractf128_ps( maxAbs, 1 ), _mm256_castps256_ps128( maxAbs ) );
-        max4 = _mm_max_ps( max4, _mm_movehl_ps( max4, max4 ) );
-        max4 = _mm_max_ss( max4, _mm_movehdup_ps( max4 ) );
-        const float maxScalar = _mm_cvtss_f32( max4 );
-
-        // Quantize these floats
-        const float d = maxScalar / 127.f;
-        y[i].d = GGML_FP32_TO_FP16(d);
-        const float id = ( maxScalar != 0.0f ) ? 127.f / maxScalar : 0.0f;
-        const __m256 mul = _mm256_set1_ps( id );
-
-        // Apply the multiplier
-        v0 = _mm256_mul_ps( v0, mul );
-        v1 = _mm256_mul_ps( v1, mul );
-        v2 = _mm256_mul_ps( v2, mul );
-        v3 = _mm256_mul_ps( v3, mul );
-
-        // Round to nearest integer
-        v0 = _mm256_round_ps( v0, _MM_ROUND_NEAREST );
-        v1 = _mm256_round_ps( v1, _MM_ROUND_NEAREST );
-        v2 = _mm256_round_ps( v2, _MM_ROUND_NEAREST );
-        v3 = _mm256_round_ps( v3, _MM_ROUND_NEAREST );
-
-        // Convert floats to integers
-        __m256i i0 = _mm256_cvtps_epi32( v0 );
-        __m256i i1 = _mm256_cvtps_epi32( v1 );
-        __m256i i2 = _mm256_cvtps_epi32( v2 );
-        __m256i i3 = _mm256_cvtps_epi32( v3 );
-
-#if defined(__AVX2__)
-        // Convert int32 to int16
-        i0 = _mm256_packs_epi32( i0, i1 );	// 0, 1, 2, 3,  8, 9, 10, 11,  4, 5, 6, 7, 12, 13, 14, 15
-        i2 = _mm256_packs_epi32( i2, i3 );	// 16, 17, 18, 19,  24, 25, 26, 27,  20, 21, 22, 23, 28, 29, 30, 31
-                                            // Convert int16 to int8
-        i0 = _mm256_packs_epi16( i0, i2 );	// 0, 1, 2, 3,  8, 9, 10, 11,  16, 17, 18, 19,  24, 25, 26, 27,  4, 5, 6, 7, 12, 13, 14, 15, 20, 21, 22, 23, 28, 29, 30, 31
-
-        // We got our precious signed bytes, but the order is now wrong
-        // These AVX2 pack instructions process 16-byte pieces independently
-        // The following instruction is fixing the order
-        const __m256i perm = _mm256_setr_epi32( 0, 4, 1, 5, 2, 6, 3, 7 );
-        i0 = _mm256_permutevar8x32_epi32( i0, perm );
-
-        _mm256_storeu_si256((__m256i *)y[i].qs, i0);
-#else
-        // Since we don't have in AVX some necessary functions,
-        // we split the registers in half and call AVX2 analogs from SSE
-        __m128i ni0 = _mm256_castsi256_si128( i0 );
-        __m128i ni1 = _mm256_extractf128_si256( i0, 1);
-        __m128i ni2 = _mm256_castsi256_si128( i1 );
-        __m128i ni3 = _mm256_extractf128_si256( i1, 1);
-        __m128i ni4 = _mm256_castsi256_si128( i2 );
-        __m128i ni5 = _mm256_extractf128_si256( i2, 1);
-        __m128i ni6 = _mm256_castsi256_si128( i3 );
-        __m128i ni7 = _mm256_extractf128_si256( i3, 1);
-
-        // Convert int32 to int16
-        ni0 = _mm_packs_epi32( ni0, ni1 );
-        ni2 = _mm_packs_epi32( ni2, ni3 );
-        ni4 = _mm_packs_epi32( ni4, ni5 );
-        ni6 = _mm_packs_epi32( ni6, ni7 );
-        // Convert int16 to int8
-        ni0 = _mm_packs_epi16( ni0, ni2 );
-        ni4 = _mm_packs_epi16( ni4, ni6 );
-
-        _mm_storeu_si128((__m128i *)(y[i].qs +  0), ni0);
-        _mm_storeu_si128((__m128i *)(y[i].qs + 16), ni4);
-#endif
-    }
-#elif defined(__riscv_v_intrinsic)
-
-    size_t vl = __riscv_vsetvl_e32m4(QK8_0);
-
-    for (int i = 0; i < nb; i++) {
-        // load elements
-        vfloat32m4_t v_x   = __riscv_vle32_v_f32m4(x+i*QK8_0, vl);
-
-        vfloat32m4_t vfabs = __riscv_vfabs_v_f32m4(v_x, vl);
-        vfloat32m1_t tmp   = __riscv_vfmv_v_f_f32m1(0.0f, vl);
-        vfloat32m1_t vmax  = __riscv_vfredmax_vs_f32m4_f32m1(vfabs, tmp, vl);
-        float amax = __riscv_vfmv_f_s_f32m1_f32(vmax);
-
-        const float d = amax / ((1 << 7) - 1);
-        const float id = d ? 1.0f/d : 0.0f;
-
-        y[i].d = GGML_FP32_TO_FP16(d);
-
-        vfloat32m4_t x0 = __riscv_vfmul_vf_f32m4(v_x, id, vl);
-
-        // convert to integer
-        vint16m2_t   vi = __riscv_vfncvt_x_f_w_i16m2(x0, vl);
-        vint8m1_t    vs = __riscv_vncvt_x_x_w_i8m1(vi, vl);
-
-        // store result
-        __riscv_vse8_v_i8m1(y[i].qs , vs, vl);
-    }
-#else
-    GGML_UNUSED(nb);
-    // scalar
-    quantize_row_q8_0_reference(x, y, k);
-#endif
-}
-
-// reference implementation for deterministic creation of model files
-void quantize_row_q8_1_reference(const float * restrict x, block_q8_1 * restrict y, int k) {
-    assert(QK8_1 == 32);
-    assert(k % QK8_1 == 0);
-    const int nb = k / QK8_1;
-
-    for (int i = 0; i < nb; i++) {
-        float amax = 0.0f; // absolute max
-
-        for (int j = 0; j < QK8_1; j++) {
-            const float v = x[i*QK8_1 + j];
-            amax = MAX(amax, fabsf(v));
-        }
-
-        const float d = amax / ((1 << 7) - 1);
-        const float id = d ? 1.0f/d : 0.0f;
-
-        y[i].d = d;
-
-        int sum = 0;
-
-        for (int j = 0; j < QK8_1/2; ++j) {
-            const float v0 = x[i*QK8_1           + j]*id;
-            const float v1 = x[i*QK8_1 + QK8_1/2 + j]*id;
-
-            y[i].qs[          j] = roundf(v0);
-            y[i].qs[QK8_1/2 + j] = roundf(v1);
-
-            sum += y[i].qs[          j];
-            sum += y[i].qs[QK8_1/2 + j];
-        }
-
-        y[i].s = sum*d;
-    }
-}
-
-void quantize_row_q8_1(const float * restrict x, void * restrict vy, int k) {
-    assert(k % QK8_1 == 0);
-    const int nb = k / QK8_1;
-
-    block_q8_1 * restrict y = vy;
-
-#if defined(__ARM_NEON)
-    for (int i = 0; i < nb; i++) {
-        float32x4_t srcv [8];
-        float32x4_t asrcv[8];
-        float32x4_t amaxv[8];
-
-        for (int j = 0; j < 8; j++) srcv[j]  = vld1q_f32(x + i*32 + 4*j);
-        for (int j = 0; j < 8; j++) asrcv[j] = vabsq_f32(srcv[j]);
-
-        for (int j = 0; j < 4; j++) amaxv[2*j] = vmaxq_f32(asrcv[2*j], asrcv[2*j+1]);
-        for (int j = 0; j < 2; j++) amaxv[4*j] = vmaxq_f32(amaxv[4*j], amaxv[4*j+2]);
-        for (int j = 0; j < 1; j++) amaxv[8*j] = vmaxq_f32(amaxv[8*j], amaxv[8*j+4]);
-
-        const float amax = vmaxvq_f32(amaxv[0]);
-
-        const float d = amax / ((1 << 7) - 1);
-        const float id = d ? 1.0f/d : 0.0f;
-
-        y[i].d = d;
-
-        int32x4_t accv = vdupq_n_s32(0);
-
-        for (int j = 0; j < 8; j++) {
-            const float32x4_t v  = vmulq_n_f32(srcv[j], id);
-            const int32x4_t   vi = vcvtnq_s32_f32(v);
-
-            y[i].qs[4*j + 0] = vgetq_lane_s32(vi, 0);
-            y[i].qs[4*j + 1] = vgetq_lane_s32(vi, 1);
-            y[i].qs[4*j + 2] = vgetq_lane_s32(vi, 2);
-            y[i].qs[4*j + 3] = vgetq_lane_s32(vi, 3);
-
-            accv = vaddq_s32(accv, vi);
-        }
-
-        y[i].s = d * vaddvq_s32(accv);
-    }
-#elif defined(__wasm_simd128__)
-    for (int i = 0; i < nb; i++) {
-        v128_t srcv [8];
-        v128_t asrcv[8];
-        v128_t amaxv[8];
-
-        for (int j = 0; j < 8; j++) srcv[j]  = wasm_v128_load(x + i*32 + 4*j);
-        for (int j = 0; j < 8; j++) asrcv[j] = wasm_f32x4_abs(srcv[j]);
-
-        for (int j = 0; j < 4; j++) amaxv[2*j] = wasm_f32x4_max(asrcv[2*j], asrcv[2*j+1]);
-        for (int j = 0; j < 2; j++) amaxv[4*j] = wasm_f32x4_max(amaxv[4*j], amaxv[4*j+2]);
-        for (int j = 0; j < 1; j++) amaxv[8*j] = wasm_f32x4_max(amaxv[8*j], amaxv[8*j+4]);
-
-        const float amax = MAX(MAX(wasm_f32x4_extract_lane(amaxv[0], 0),
-                                   wasm_f32x4_extract_lane(amaxv[0], 1)),
-                               MAX(wasm_f32x4_extract_lane(amaxv[0], 2),
-                                   wasm_f32x4_extract_lane(amaxv[0], 3)));
-
-        const float d = amax / ((1 << 7) - 1);
-        const float id = d ? 1.0f/d : 0.0f;
-
-        y[i].d = d;
-
-        v128_t accv = wasm_i32x4_splat(0);
-
-        for (int j = 0; j < 8; j++) {
-            const v128_t v  = wasm_f32x4_mul(srcv[j], wasm_f32x4_splat(id));
-            const v128_t vi = wasm_i32x4_trunc_sat_f32x4(v);
-
-            y[i].qs[4*j + 0] = wasm_i32x4_extract_lane(vi, 0);
-            y[i].qs[4*j + 1] = wasm_i32x4_extract_lane(vi, 1);
-            y[i].qs[4*j + 2] = wasm_i32x4_extract_lane(vi, 2);
-            y[i].qs[4*j + 3] = wasm_i32x4_extract_lane(vi, 3);
-
-            accv = wasm_i32x4_add(accv, vi);
-        }
-
-        y[i].s = d * (wasm_i32x4_extract_lane(accv, 0) +
-                      wasm_i32x4_extract_lane(accv, 1) +
-                      wasm_i32x4_extract_lane(accv, 2) +
-                      wasm_i32x4_extract_lane(accv, 3));
-    }
-#elif defined(__AVX2__) || defined(__AVX__)
-    for (int i = 0; i < nb; i++) {
-        // Load elements into 4 AVX vectors
-        __m256 v0 = _mm256_loadu_ps( x );
-        __m256 v1 = _mm256_loadu_ps( x + 8 );
-        __m256 v2 = _mm256_loadu_ps( x + 16 );
-        __m256 v3 = _mm256_loadu_ps( x + 24 );
-        x += 32;
-
-        // Compute max(abs(e)) for the block
-        const __m256 signBit = _mm256_set1_ps( -0.0f );
-        __m256 maxAbs = _mm256_andnot_ps( signBit, v0 );
-        maxAbs = _mm256_max_ps( maxAbs, _mm256_andnot_ps( signBit, v1 ) );
-        maxAbs = _mm256_max_ps( maxAbs, _mm256_andnot_ps( signBit, v2 ) );
-        maxAbs = _mm256_max_ps( maxAbs, _mm256_andnot_ps( signBit, v3 ) );
-
-        __m128 max4 = _mm_max_ps( _mm256_extractf128_ps( maxAbs, 1 ), _mm256_castps256_ps128( maxAbs ) );
-        max4 = _mm_max_ps( max4, _mm_movehl_ps( max4, max4 ) );
-        max4 = _mm_max_ss( max4, _mm_movehdup_ps( max4 ) );
-        const float maxScalar = _mm_cvtss_f32( max4 );
-
-        // Quantize these floats
-        const float d = maxScalar / 127.f;
-        y[i].d = d;
-        const float id = ( maxScalar != 0.0f ) ? 127.f / maxScalar : 0.0f;
-        const __m256 mul = _mm256_set1_ps( id );
-
-        // Apply the multiplier
-        v0 = _mm256_mul_ps( v0, mul );
-        v1 = _mm256_mul_ps( v1, mul );
-        v2 = _mm256_mul_ps( v2, mul );
-        v3 = _mm256_mul_ps( v3, mul );
-
-        // Round to nearest integer
-        v0 = _mm256_round_ps( v0, _MM_ROUND_NEAREST );
-        v1 = _mm256_round_ps( v1, _MM_ROUND_NEAREST );
-        v2 = _mm256_round_ps( v2, _MM_ROUND_NEAREST );
-        v3 = _mm256_round_ps( v3, _MM_ROUND_NEAREST );
-
-        // Convert floats to integers
-        __m256i i0 = _mm256_cvtps_epi32( v0 );
-        __m256i i1 = _mm256_cvtps_epi32( v1 );
-        __m256i i2 = _mm256_cvtps_epi32( v2 );
-        __m256i i3 = _mm256_cvtps_epi32( v3 );
-
-#if defined(__AVX2__)
-        // Compute the sum of the quants and set y[i].s
-        y[i].s = d * hsum_i32_8(_mm256_add_epi32(_mm256_add_epi32(i0, i1), _mm256_add_epi32(i2, i3)));
-
-        // Convert int32 to int16
-        i0 = _mm256_packs_epi32( i0, i1 );	// 0, 1, 2, 3,  8, 9, 10, 11,  4, 5, 6, 7, 12, 13, 14, 15
-        i2 = _mm256_packs_epi32( i2, i3 );	// 16, 17, 18, 19,  24, 25, 26, 27,  20, 21, 22, 23, 28, 29, 30, 31
-                                            // Convert int16 to int8
-        i0 = _mm256_packs_epi16( i0, i2 );	// 0, 1, 2, 3,  8, 9, 10, 11,  16, 17, 18, 19,  24, 25, 26, 27,  4, 5, 6, 7, 12, 13, 14, 15, 20, 21, 22, 23, 28, 29, 30, 31
-
-        // We got our precious signed bytes, but the order is now wrong
-        // These AVX2 pack instructions process 16-byte pieces independently
-        // The following instruction is fixing the order
-        const __m256i perm = _mm256_setr_epi32( 0, 4, 1, 5, 2, 6, 3, 7 );
-        i0 = _mm256_permutevar8x32_epi32( i0, perm );
-
-        _mm256_storeu_si256((__m256i *)y[i].qs, i0);
-#else
-        // Since we don't have in AVX some necessary functions,
-        // we split the registers in half and call AVX2 analogs from SSE
-        __m128i ni0 = _mm256_castsi256_si128( i0 );
-        __m128i ni1 = _mm256_extractf128_si256( i0, 1);
-        __m128i ni2 = _mm256_castsi256_si128( i1 );
-        __m128i ni3 = _mm256_extractf128_si256( i1, 1);
-        __m128i ni4 = _mm256_castsi256_si128( i2 );
-        __m128i ni5 = _mm256_extractf128_si256( i2, 1);
-        __m128i ni6 = _mm256_castsi256_si128( i3 );
-        __m128i ni7 = _mm256_extractf128_si256( i3, 1);
-
-        // Compute the sum of the quants and set y[i].s
-        const __m128i s0 = _mm_add_epi32(_mm_add_epi32(ni0, ni1), _mm_add_epi32(ni2, ni3));
-        const __m128i s1 = _mm_add_epi32(_mm_add_epi32(ni4, ni5), _mm_add_epi32(ni6, ni7));
-        y[i].s = d * hsum_i32_4(_mm_add_epi32(s0, s1));
-
-        // Convert int32 to int16
-        ni0 = _mm_packs_epi32( ni0, ni1 );
-        ni2 = _mm_packs_epi32( ni2, ni3 );
-        ni4 = _mm_packs_epi32( ni4, ni5 );
-        ni6 = _mm_packs_epi32( ni6, ni7 );
-        // Convert int16 to int8
-        ni0 = _mm_packs_epi16( ni0, ni2 );
-        ni4 = _mm_packs_epi16( ni4, ni6 );
-
-        _mm_storeu_si128((__m128i *)(y[i].qs +  0), ni0);
-        _mm_storeu_si128((__m128i *)(y[i].qs + 16), ni4);
-#endif
-    }
-#elif defined(__riscv_v_intrinsic)
-
-    size_t vl = __riscv_vsetvl_e32m4(QK8_1);
-
-    for (int i = 0; i < nb; i++) {
-        // load elements
-        vfloat32m4_t v_x   = __riscv_vle32_v_f32m4(x+i*QK8_1, vl);
-
-        vfloat32m4_t vfabs = __riscv_vfabs_v_f32m4(v_x, vl);
-        vfloat32m1_t tmp   = __riscv_vfmv_v_f_f32m1(0.0, vl);
-        vfloat32m1_t vmax  = __riscv_vfredmax_vs_f32m4_f32m1(vfabs, tmp, vl);
-        float amax = __riscv_vfmv_f_s_f32m1_f32(vmax);
-
-        const float d  = amax / ((1 << 7) - 1);
-        const float id = d ? 1.0f/d : 0.0f;
-
-        y[i].d = d;
-
-        vfloat32m4_t x0 = __riscv_vfmul_vf_f32m4(v_x, id, vl);
-
-        // convert to integer
-        vint16m2_t   vi = __riscv_vfncvt_x_f_w_i16m2(x0, vl);
-        vint8m1_t    vs = __riscv_vncvt_x_x_w_i8m1(vi, vl);
-
-        // store result
-        __riscv_vse8_v_i8m1(y[i].qs , vs, vl);
-
-        // compute sum for y[i].s
-        vint16m1_t tmp2 = __riscv_vmv_v_x_i16m1(0, vl);
-        vint16m1_t vwrs = __riscv_vwredsum_vs_i8m1_i16m1(vs, tmp2, vl);
-
-        // set y[i].s
-        int sum = __riscv_vmv_x_s_i16m1_i16(vwrs);
-        y[i].s = sum*d;
-    }
-#else
-    GGML_UNUSED(nb);
-    // scalar
-    quantize_row_q8_1_reference(x, y, k);
-#endif
-}
-
-void dequantize_row_q4_0(const block_q4_0 * restrict x, float * restrict y, int k) {
-    static const int qk = QK4_0;
-
-    assert(k % qk == 0);
-
-    const int nb = k / qk;
-
-    for (int i = 0; i < nb; i++) {
-        const float d = GGML_FP16_TO_FP32(x[i].d);
-
-        for (int j = 0; j < qk/2; ++j) {
-            const int x0 = (x[i].qs[j] & 0x0F) - 8;
-            const int x1 = (x[i].qs[j] >>   4) - 8;
-
-            y[i*qk + j + 0   ] = x0*d;
-            y[i*qk + j + qk/2] = x1*d;
-        }
-    }
-}
-
-void dequantize_row_q4_1(const block_q4_1 * restrict x, float * restrict y, int k) {
-    static const int qk = QK4_1;
-
-    assert(k % qk == 0);
-
-    const int nb = k / qk;
-
-    for (int i = 0; i < nb; i++) {
-        const float d = GGML_FP16_TO_FP32(x[i].d);
-        const float m = GGML_FP16_TO_FP32(x[i].m);
-
-        for (int j = 0; j < qk/2; ++j) {
-            const int x0 = (x[i].qs[j] & 0x0F);
-            const int x1 = (x[i].qs[j] >>   4);
-
-            y[i*qk + j + 0   ] = x0*d + m;
-            y[i*qk + j + qk/2] = x1*d + m;
-        }
-    }
-}
-
-void dequantize_row_q5_0(const block_q5_0 * restrict x, float * restrict y, int k) {
-    static const int qk = QK5_0;
-
-    assert(k % qk == 0);
-
-    const int nb = k / qk;
-
-    for (int i = 0; i < nb; i++) {
-        const float d = GGML_FP16_TO_FP32(x[i].d);
-
-        uint32_t qh;
-        memcpy(&qh, x[i].qh, sizeof(qh));
-
-        for (int j = 0; j < qk/2; ++j) {
-            const uint8_t xh_0 = ((qh >> (j +  0)) << 4) & 0x10;
-            const uint8_t xh_1 = ((qh >> (j + 12))     ) & 0x10;
-
-            const int32_t x0 = ((x[i].qs[j] & 0x0F) | xh_0) - 16;
-            const int32_t x1 = ((x[i].qs[j] >>   4) | xh_1) - 16;
-
-            y[i*qk + j + 0   ] = x0*d;
-            y[i*qk + j + qk/2] = x1*d;
-        }
-    }
-}
-
-void dequantize_row_q5_1(const block_q5_1 * restrict x, float * restrict y, int k) {
-    static const int qk = QK5_1;
-
-    assert(k % qk == 0);
-
-    const int nb = k / qk;
-
-    for (int i = 0; i < nb; i++) {
-        const float d = GGML_FP16_TO_FP32(x[i].d);
-        const float m = GGML_FP16_TO_FP32(x[i].m);
-
-        uint32_t qh;
-        memcpy(&qh, x[i].qh, sizeof(qh));
-
-        for (int j = 0; j < qk/2; ++j) {
-            const uint8_t xh_0 = ((qh >> (j +  0)) << 4) & 0x10;
-            const uint8_t xh_1 = ((qh >> (j + 12))     ) & 0x10;
-
-            const int x0 = (x[i].qs[j] & 0x0F) | xh_0;
-            const int x1 = (x[i].qs[j] >>   4) | xh_1;
-
-            y[i*qk + j + 0   ] = x0*d + m;
-            y[i*qk + j + qk/2] = x1*d + m;
-        }
-    }
-}
-
-void dequantize_row_q8_0(const block_q8_0 * restrict x, float * restrict y, int k) {
-    static const int qk = QK8_0;
-
-    assert(k % qk == 0);
-
-    const int nb = k / qk;
-
-    for (int i = 0; i < nb; i++) {
-        const float d = GGML_FP16_TO_FP32(x[i].d);
-
-        for (int j = 0; j < qk; ++j) {
-            y[i*qk + j] = x[i].qs[j]*d;
-        }
-    }
-}
-
-//
-// 2-6 bit quantization in super-blocks
-//
-
-//
-// ===================== Helper functions
-//
-static inline int nearest_int(float fval) {
-    assert(fval <= 4194303.f);
-    float val = fval + 12582912.f;
-    int i; memcpy(&i, &val, sizeof(int));
-    return (i & 0x007fffff) - 0x00400000;
-}
-
-static float make_qx_quants(int n, int nmax, const float * restrict x, int8_t * restrict L, int rmse_type) {
-    float max = 0;
-    float amax = 0;
-    for (int i = 0; i < n; ++i) {
-        float ax = fabsf(x[i]);
-        if (ax > amax) { amax = ax; max = x[i]; }
-    }
-    if (amax < 1e-30f) { // all zero
-        for (int i = 0; i < n; ++i) {
-            L[i] = 0;
-        }
-        return 0.f;
-    }
-    float iscale = -nmax / max;
-    if (rmse_type == 0) {
-        for (int i = 0; i < n; ++i) {
-            int l = nearest_int(iscale * x[i]);
-            L[i] = nmax + MAX(-nmax, MIN(nmax-1, l));
-        }
-        return 1/iscale;
-    }
-    bool return_early = false;
-    if (rmse_type < 0) {
-        rmse_type = -rmse_type;
-        return_early = true;
-    }
-    int weight_type = rmse_type%2;
-    float sumlx = 0;
-    float suml2 = 0;
-    for (int i = 0; i < n; ++i) {
-        int l = nearest_int(iscale * x[i]);
-        l = MAX(-nmax, MIN(nmax-1, l));
-        L[i] = l + nmax;
-        float w = weight_type == 1 ? x[i] * x[i] : 1;
-        sumlx += w*x[i]*l;
-        suml2 += w*l*l;
-    }
-    float scale = sumlx/suml2;
-    if (return_early) return suml2 > 0 ? 0.5f*(scale + 1/iscale) : 1/iscale;
-    float best = scale * sumlx;
-    for (int is = -9; is <= 9; ++is) {
-        if (is == 0) {
-            continue;
-        }
-        iscale = -(nmax + 0.1f*is) / max;
-        sumlx = suml2 = 0;
-        for (int i = 0; i < n; ++i) {
-            int l = nearest_int(iscale * x[i]);
-            l = MAX(-nmax, MIN(nmax-1, l));
-            float w = weight_type == 1 ? x[i] * x[i] : 1;
-            sumlx += w*x[i]*l;
-            suml2 += w*l*l;
-        }
-        if (suml2 > 0 && sumlx*sumlx > best*suml2) {
-            for (int i = 0; i < n; ++i) {
-                int l = nearest_int(iscale * x[i]);
-                L[i] = nmax + MAX(-nmax, MIN(nmax-1, l));
-            }
-            scale = sumlx/suml2; best = scale*sumlx;
-        }
-    }
-    return scale;
-}
-
-static float make_q3_quants(int n, int nmax, const float * restrict x, int8_t * restrict L, bool do_rmse) {
-    float max = 0;
-    float amax = 0;
-    for (int i = 0; i < n; ++i) {
-        float ax = fabsf(x[i]);
-        if (ax > amax) { amax = ax; max = x[i]; }
-    }
-    if (!amax) { // all zero
-        for (int i = 0; i < n; ++i) { L[i] = 0; }
-        return 0.f;
-    }
-    float iscale = -nmax / max;
-    if (do_rmse) {
-        float sumlx = 0;
-        float suml2 = 0;
-        for (int i = 0; i < n; ++i) {
-            int l = nearest_int(iscale * x[i]);
-            l = MAX(-nmax, MIN(nmax-1, l));
-            L[i] = l;
-            float w = x[i]*x[i];
-            sumlx += w*x[i]*l;
-            suml2 += w*l*l;
-        }
-        for (int itry = 0; itry < 5; ++itry) {
-            int n_changed = 0;
-            for (int i = 0; i < n; ++i) {
-                float w = x[i]*x[i];
-                float slx = sumlx - w*x[i]*L[i];
-                if (slx > 0) {
-                    float sl2 = suml2 - w*L[i]*L[i];
-                    int new_l = nearest_int(x[i] * sl2 / slx);
-                    new_l = MAX(-nmax, MIN(nmax-1, new_l));
-                    if (new_l != L[i]) {
-                        slx += w*x[i]*new_l;
-                        sl2 += w*new_l*new_l;
-                        if (sl2 > 0 && slx*slx*suml2 > sumlx*sumlx*sl2) {
-                            L[i] = new_l; sumlx = slx; suml2 = sl2;
-                            ++n_changed;
-                        }
-                    }
-                }
-            }
-            if (!n_changed) {
-                break;
-            }
-        }
-        for (int i = 0; i < n; ++i) {
-            L[i] += nmax;
-        }
-        return sumlx / suml2;
-    }
-    for (int i = 0; i < n; ++i) {
-        int l = nearest_int(iscale * x[i]);
-        l = MAX(-nmax, MIN(nmax-1, l));
-        L[i] = l + nmax;
-    }
-    return 1/iscale;
-}
-
-static float make_qkx1_quants(int n, int nmax, const float * restrict x, uint8_t * restrict L, float * restrict the_min,
-        int ntry, float alpha) {
-    float min = x[0];
-    float max = x[0];
-    for (int i = 1; i < n; ++i) {
-        if (x[i] < min) min = x[i];
-        if (x[i] > max) max = x[i];
-    }
-    if (max == min) {
-        for (int i = 0; i < n; ++i) L[i] = 0;
-        *the_min = 0;
-        return 0.f;
-    }
-    if (min > 0) min = 0;
-    float iscale = nmax/(max - min);
-    float scale = 1/iscale;
-    for (int itry = 0; itry < ntry; ++itry) {
-        float sumlx = 0; int suml2 = 0;
-        bool did_change = false;
-        for (int i = 0; i < n; ++i) {
-            int l = nearest_int(iscale*(x[i] - min));
-            l = MAX(0, MIN(nmax, l));
-            if (l != L[i]) {
-                L[i] = l;
-                did_change = true;
-            }
-            sumlx += (x[i] - min)*l;
-            suml2 += l*l;
-        }
-        scale = sumlx/suml2;
-        float sum = 0;
-        for (int i = 0; i < n; ++i) {
-            sum += x[i] - scale*L[i];
-        }
-        min = alpha*min + (1 - alpha)*sum/n;
-        if (min > 0) min = 0;
-        iscale = 1/scale;
-        if (!did_change) break;
-    }
-    *the_min = -min;
-    return scale;
-}
-
-static float make_qkx2_quants(int n, int nmax, const float * restrict x, const float * restrict weights,
-        uint8_t * restrict L, float * restrict the_min, uint8_t * restrict Laux,
-        float rmin, float rdelta, int nstep, bool use_mad) {
-    float min = x[0];
-    float max = x[0];
-    float sum_w = weights[0];
-    float sum_x = sum_w * x[0];
-#ifdef HAVE_BUGGY_APPLE_LINKER
-    // use 'volatile' to prevent unroll and work around a bug in Apple ld64 1015.7
-    for (volatile int i = 1; i < n; ++i) {
-#else
-    for (int i = 1; i < n; ++i) {
-#endif
-        if (x[i] < min) min = x[i];
-        if (x[i] > max) max = x[i];
-        float w = weights[i];
-        sum_w += w;
-        sum_x += w * x[i];
-    }
-    if (min > 0) min = 0;
-    if (max == min) {
-        for (int i = 0; i < n; ++i) L[i] = 0;
-        *the_min = -min;
-        return 0.f;
-    }
-    float iscale = nmax/(max - min);
-    float scale = 1/iscale;
-    float best_mad = 0;
-    for (int i = 0; i < n; ++i) {
-        int l = nearest_int(iscale*(x[i] - min));
-        L[i] = MAX(0, MIN(nmax, l));
-        float diff = scale * L[i] + min - x[i];
-        diff = use_mad ? fabsf(diff) : diff * diff;
-        float w = weights[i];
-        best_mad += w * diff;
-    }
-    if (nstep < 1) {
-        *the_min = -min;
-        return scale;
-    }
-    for (int is = 0; is <= nstep; ++is) {
-        iscale = (rmin + rdelta*is + nmax)/(max - min);
-        float sum_l = 0, sum_l2 = 0, sum_xl = 0;
-        for (int i = 0; i < n; ++i) {
-            int l = nearest_int(iscale*(x[i] - min));
-            l = MAX(0, MIN(nmax, l));
-            Laux[i] = l;
-            float w = weights[i];
-            sum_l += w*l;
-            sum_l2 += w*l*l;
-            sum_xl += w*l*x[i];
-        }
-        float D = sum_w * sum_l2 - sum_l * sum_l;
-        if (D > 0) {
-            float this_scale = (sum_w * sum_xl - sum_x * sum_l)/D;
-            float this_min   = (sum_l2 * sum_x - sum_l * sum_xl)/D;
-            if (this_min > 0) {
-                this_min = 0;
-                this_scale = sum_xl / sum_l2;
-            }
-            float mad = 0;
-            for (int i = 0; i < n; ++i) {
-                float diff = this_scale * Laux[i] + this_min - x[i];
-                diff = use_mad ? fabsf(diff) : diff * diff;
-                float w = weights[i];
-                mad += w * diff;
-            }
-            if (mad < best_mad) {
-                for (int i = 0; i < n; ++i) {
-                    L[i] = Laux[i];
-                }
-                best_mad = mad;
-                scale = this_scale;
-                min = this_min;
-            }
-        }
-    }
-    *the_min = -min;
-    return scale;
-}
-
-#if QK_K == 256
-static inline void get_scale_min_k4(int j, const uint8_t * restrict q, uint8_t * restrict d, uint8_t * restrict m) {
-    if (j < 4) {
-        *d = q[j] & 63; *m = q[j + 4] & 63;
-    } else {
-        *d = (q[j+4] & 0xF) | ((q[j-4] >> 6) << 4);
-        *m = (q[j+4] >>  4) | ((q[j-0] >> 6) << 4);
-    }
-}
-#endif
-
-//========================- 2-bit (de)-quantization
-
-void quantize_row_q2_K_reference(const float * restrict x, block_q2_K * restrict y, int k) {
-    assert(k % QK_K == 0);
-    const int nb = k / QK_K;
-
-    uint8_t L[QK_K];
-    uint8_t Laux[16];
-    float   weights[16];
-    float mins[QK_K/16];
-    float scales[QK_K/16];
-
-    const float q4scale = 15.f;
-
-    for (int i = 0; i < nb; i++) {
-        float max_scale = 0; // as we are deducting the min, scales are always positive
-        float max_min = 0;
-        for (int j = 0; j < QK_K/16; ++j) {
-            for (int l = 0; l < 16; ++l) weights[l] = fabsf(x[16*j + l]);
-            scales[j] = make_qkx2_quants(16, 3, x + 16*j, weights, L + 16*j, &mins[j], Laux, -0.5f, 0.1f, 15, true);
-            float scale = scales[j];
-            if (scale > max_scale) {
-                max_scale = scale;
-            }
-            float min = mins[j];
-            if (min > max_min) {
-                max_min = min;
-            }
-        }
-
-        if (max_scale > 0) {
-            float iscale = q4scale/max_scale;
-            for (int j = 0; j < QK_K/16; ++j) {
-                int l = nearest_int(iscale*scales[j]);
-                y[i].scales[j] = l;
-            }
-            y[i].d = GGML_FP32_TO_FP16(max_scale/q4scale);
-        } else {
-            for (int j = 0; j < QK_K/16; ++j) y[i].scales[j] = 0;
-            y[i].d = GGML_FP32_TO_FP16(0.f);
-        }
-        if (max_min > 0) {
-            float iscale = q4scale/max_min;
-            for (int j = 0; j < QK_K/16; ++j) {
-                int l = nearest_int(iscale*mins[j]);
-                y[i].scales[j] |= (l << 4);
-            }
-            y[i].dmin = GGML_FP32_TO_FP16(max_min/q4scale);
-        } else {
-            y[i].dmin = GGML_FP32_TO_FP16(0.f);
-        }
-        for (int j = 0; j < QK_K/16; ++j) {
-            const float d = GGML_FP16_TO_FP32(y[i].d) * (y[i].scales[j] & 0xF);
-            if (!d) continue;
-            const float dm = GGML_FP16_TO_FP32(y[i].dmin) * (y[i].scales[j] >> 4);
-            for (int ii = 0; ii < 16; ++ii) {
-                int l = nearest_int((x[16*j + ii] + dm)/d);
-                l = MAX(0, MIN(3, l));
-                L[16*j + ii] = l;
-            }
-        }
-
-#if QK_K == 256
-        for (int j = 0; j < QK_K; j += 128) {
-            for (int l = 0; l < 32; ++l) {
-                y[i].qs[j/4 + l] = L[j + l] | (L[j + l + 32] << 2) | (L[j + l + 64] << 4) | (L[j + l + 96] << 6);
-            }
-        }
-#else
-        for (int l = 0; l < 16; ++l) {
-            y[i].qs[l] = L[l] | (L[l + 16] << 2) | (L[l + 32] << 4) | (L[l + 48] << 6);
-        }
-#endif
-
-        x += QK_K;
-
-    }
-}
-
-void dequantize_row_q2_K(const block_q2_K * restrict x, float * restrict y, int k) {
-    assert(k % QK_K == 0);
-    const int nb = k / QK_K;
-
-    for (int i = 0; i < nb; i++) {
-
-        const float d = GGML_FP16_TO_FP32(x[i].d);
-        const float min = GGML_FP16_TO_FP32(x[i].dmin);
-
-        const uint8_t * q = x[i].qs;
-
-#if QK_K == 256
-        int is = 0;
-        float dl, ml;
-        for (int n = 0; n < QK_K; n += 128) {
-            int shift = 0;
-            for (int j = 0; j < 4; ++j) {
-
-                uint8_t sc = x[i].scales[is++];
-                dl = d * (sc & 0xF); ml = min * (sc >> 4);
-                for (int l = 0; l < 16; ++l) *y++ = dl * ((int8_t)((q[l] >> shift) & 3)) - ml;
-
-                sc = x[i].scales[is++];
-                dl = d * (sc & 0xF); ml = min * (sc >> 4);
-                for (int l = 0; l < 16; ++l) *y++ = dl * ((int8_t)((q[l+16] >> shift) & 3)) - ml;
-
-                shift += 2;
-            }
-            q += 32;
-        }
-#else
-        float dl1 = d * (x[i].scales[0] & 0xF), ml1 = min * (x[i].scales[0] >> 4);
-        float dl2 = d * (x[i].scales[1] & 0xF), ml2 = min * (x[i].scales[1] >> 4);
-        float dl3 = d * (x[i].scales[2] & 0xF), ml3 = min * (x[i].scales[2] >> 4);
-        float dl4 = d * (x[i].scales[3] & 0xF), ml4 = min * (x[i].scales[3] >> 4);
-        for (int l = 0; l < 16; ++l) {
-            y[l+ 0] = dl1 * ((int8_t)((q[l] >> 0) & 3)) - ml1;
-            y[l+16] = dl2 * ((int8_t)((q[l] >> 2) & 3)) - ml2;
-            y[l+32] = dl3 * ((int8_t)((q[l] >> 4) & 3)) - ml3;
-            y[l+48] = dl4 * ((int8_t)((q[l] >> 6) & 3)) - ml4;
-        }
-        y += QK_K;
-#endif
-    }
-}
-
-void quantize_row_q2_K(const float * restrict x, void * restrict vy, int k) {
-    quantize_row_q2_K_reference(x, vy, k);
-}
-
-size_t ggml_quantize_q2_K(const float * restrict src, void * restrict dst, int n, int k, int64_t * restrict hist) {
-    (void)hist; // TODO: collect histograms
-
-    for (int j = 0; j < n; j += k) {
-        block_q2_K * restrict y = (block_q2_K *)dst + j/QK_K;
-        quantize_row_q2_K_reference(src + j, y, k);
-    }
-    return (n/QK_K*sizeof(block_q2_K));
-}
-
-//========================= 3-bit (de)-quantization
-
-void quantize_row_q3_K_reference(const float * restrict x, block_q3_K * restrict y, int k) {
-    assert(k % QK_K == 0);
-    const int nb = k / QK_K;
-
-    int8_t L[QK_K];
-    float scales[QK_K / 16];
-
-    for (int i = 0; i < nb; i++) {
-
-        float max_scale = 0;
-        float amax = 0;
-        for (int j = 0; j < QK_K/16; ++j) {
-            scales[j] = make_q3_quants(16, 4, x + 16*j, L + 16*j, true);
-            float scale = fabsf(scales[j]);
-            if (scale > amax) {
-                amax = scale; max_scale = scales[j];
-            }
-        }
-
-#if QK_K == 256
-        memset(y[i].scales, 0, 12);
-        if (max_scale) {
-            float iscale = -32.f/max_scale;
-            for (int j = 0; j < QK_K/16; ++j) {
-                int8_t l = nearest_int(iscale*scales[j]);
-                l = MAX(-32, MIN(31, l)) + 32;
-                if (j < 8) {
-                    y[i].scales[j] = l & 0xF;
-                } else {
-                    y[i].scales[j-8] |= ((l & 0xF) << 4);
-                }
-                l >>= 4;
-                y[i].scales[j%4 + 8] |= (l << (2*(j/4)));
-            }
-            y[i].d = GGML_FP32_TO_FP16(1/iscale);
-        } else {
-            y[i].d = GGML_FP32_TO_FP16(0.f);
-        }
-
-        int8_t sc;
-        for (int j = 0; j < QK_K/16; ++j) {
-            sc = j < 8 ? y[i].scales[j] & 0xF : y[i].scales[j-8] >> 4;
-            sc = (sc | (((y[i].scales[8 + j%4] >> (2*(j/4))) & 3) << 4)) - 32;
-            float d = GGML_FP16_TO_FP32(y[i].d) * sc;
-            if (!d) {
-                continue;
-            }
-            for (int ii = 0; ii < 16; ++ii) {
-                int l = nearest_int(x[16*j + ii]/d);
-                l = MAX(-4, MIN(3, l));
-                L[16*j + ii] = l + 4;
-            }
-        }
-#else
-        if (max_scale) {
-            float iscale = -8.f/max_scale;
-            for (int j = 0; j < QK_K/16; j+=2) {
-                int l1 = nearest_int(iscale*scales[j]);
-                l1 = 8 + MAX(-8, MIN(7, l1));
-                int l2 = nearest_int(iscale*scales[j+1]);
-                l2 = 8 + MAX(-8, MIN(7, l2));
-                y[i].scales[j/2] = l1 | (l2 << 4);
-            }
-            y[i].d = GGML_FP32_TO_FP16(1/iscale);
-        } else {
-            for (int j = 0; j < QK_K/16; j+=2) {
-                y[i].scales[j/2] = 0;
-            }
-            y[i].d = GGML_FP32_TO_FP16(0.f);
-        }
-        for (int j = 0; j < QK_K/16; ++j) {
-            int s = j%2 == 0 ? y[i].scales[j/2] & 0xF : y[i].scales[j/2] >> 4;
-            float d = GGML_FP16_TO_FP32(y[i].d) * (s - 8);
-            if (!d) {
-                continue;
-            }
-            for (int ii = 0; ii < 16; ++ii) {
-                int l = nearest_int(x[16*j + ii]/d);
-                l = MAX(-4, MIN(3, l));
-                L[16*j + ii] = l + 4;
-            }
-        }
-#endif
-
-        memset(y[i].hmask, 0, QK_K/8);
-        // We put the high-bit for the 1st 8 quants into bit 0, the next 8 into bit 1, etc.
-        int m = 0;
-        uint8_t hm = 1;
-        for (int j = 0; j < QK_K; ++j) {
-            if (L[j] > 3) {
-                y[i].hmask[m] |= hm;
-                L[j] -= 4;
-            }
-            if (++m == QK_K/8) {
-                m = 0; hm <<= 1;
-            }
-        }
-#if QK_K == 256
-        for (int j = 0; j < QK_K; j += 128) {
-            for (int l = 0; l < 32; ++l) {
-                y[i].qs[j/4 + l] = L[j + l] | (L[j + l + 32] << 2) | (L[j + l + 64] << 4) | (L[j + l + 96] << 6);
-            }
-        }
-#else
-        for (int l = 0; l < 16; ++l) {
-            y[i].qs[l] = L[l] | (L[l + 16] << 2) | (L[l + 32] << 4) | (L[l + 48] << 6);
-        }
-#endif
-
-        x += QK_K;
-    }
-}
-
-#if QK_K == 256
-void dequantize_row_q3_K(const block_q3_K * restrict x, float * restrict y, int k) {
-    assert(k % QK_K == 0);
-    const int nb = k / QK_K;
-
-    const uint32_t kmask1 = 0x03030303;
-    const uint32_t kmask2 = 0x0f0f0f0f;
-
-    uint32_t aux[4];
-    const int8_t * scales = (const int8_t*)aux;
-
-    for (int i = 0; i < nb; i++) {
-
-        const float d_all = GGML_FP16_TO_FP32(x[i].d);
-
-        const uint8_t * restrict q = x[i].qs;
-        const uint8_t * restrict hm = x[i].hmask;
-        uint8_t m = 1;
-
-        memcpy(aux, x[i].scales, 12);
-        uint32_t tmp = aux[2];
-        aux[2] = ((aux[0] >> 4) & kmask2) | (((tmp >> 4) & kmask1) << 4);
-        aux[3] = ((aux[1] >> 4) & kmask2) | (((tmp >> 6) & kmask1) << 4);
-        aux[0] = (aux[0] & kmask2) | (((tmp >> 0) & kmask1) << 4);
-        aux[1] = (aux[1] & kmask2) | (((tmp >> 2) & kmask1) << 4);
-
-        int is = 0;
-        float dl;
-        for (int n = 0; n < QK_K; n += 128) {
-            int shift = 0;
-            for (int j = 0; j < 4; ++j) {
-
-                dl = d_all * (scales[is++] - 32);
-                for (int l = 0; l < 16; ++l) {
-                    *y++ = dl * ((int8_t)((q[l+ 0] >> shift) & 3) - ((hm[l+ 0] & m) ? 0 : 4));
-                }
-
-                dl = d_all * (scales[is++] - 32);
-                for (int l = 0; l < 16; ++l) {
-                    *y++ = dl * ((int8_t)((q[l+16] >> shift) & 3) - ((hm[l+16] & m) ? 0 : 4));
-                }
-
-                shift += 2;
-                m <<= 1;
-            }
-            q += 32;
-        }
-
-    }
-}
-#else
-void dequantize_row_q3_K(const block_q3_K * restrict x, float * restrict y, int k) {
-    assert(k % QK_K == 0);
-    assert(QK_K == 64);
-    const int nb = k / QK_K;
-
-    for (int i = 0; i < nb; i++) {
-
-        const float d_all = GGML_FP16_TO_FP32(x[i].d);
-
-        const uint8_t * restrict q = x[i].qs;
-        const uint8_t * restrict hm = x[i].hmask;
-
-        const float d1 = d_all * ((x[i].scales[0] & 0xF) - 8);
-        const float d2 = d_all * ((x[i].scales[0] >>  4) - 8);
-        const float d3 = d_all * ((x[i].scales[1] & 0xF) - 8);
-        const float d4 = d_all * ((x[i].scales[1] >>  4) - 8);
-
-        for (int l=0; l<8; ++l) {
-            uint8_t h = hm[l];
-            y[l+ 0] = d1 * ((int8_t)((q[l+0] >> 0) & 3) - ((h & 0x01) ? 0 : 4));
-            y[l+ 8] = d1 * ((int8_t)((q[l+8] >> 0) & 3) - ((h & 0x02) ? 0 : 4));
-            y[l+16] = d2 * ((int8_t)((q[l+0] >> 2) & 3) - ((h & 0x04) ? 0 : 4));
-            y[l+24] = d2 * ((int8_t)((q[l+8] >> 2) & 3) - ((h & 0x08) ? 0 : 4));
-            y[l+32] = d3 * ((int8_t)((q[l+0] >> 4) & 3) - ((h & 0x10) ? 0 : 4));
-            y[l+40] = d3 * ((int8_t)((q[l+8] >> 4) & 3) - ((h & 0x20) ? 0 : 4));
-            y[l+48] = d4 * ((int8_t)((q[l+0] >> 6) & 3) - ((h & 0x40) ? 0 : 4));
-            y[l+56] = d4 * ((int8_t)((q[l+8] >> 6) & 3) - ((h & 0x80) ? 0 : 4));
-        }
-        y += QK_K;
-    }
-}
-#endif
-
-void quantize_row_q3_K(const float * restrict x, void * restrict vy, int k) {
-    quantize_row_q3_K_reference(x, vy, k);
-}
-
-size_t ggml_quantize_q3_K(const float * restrict src, void * restrict dst, int n, int k, int64_t * restrict hist) {
-    (void)hist; // TODO: collect histograms
-
-    for (int j = 0; j < n; j += k) {
-        block_q3_K * restrict y = (block_q3_K *)dst + j/QK_K;
-        quantize_row_q3_K_reference(src + j, y, k);
-    }
-    return (n/QK_K*sizeof(block_q3_K));
-}
-
-// ====================== 4-bit (de)-quantization
-
-void quantize_row_q4_K_reference(const float * restrict x, block_q4_K * restrict y, int k) {
-    assert(k % QK_K == 0);
-    const int nb = k / QK_K;
-
-    uint8_t L[QK_K];
-    uint8_t Laux[32];
-    float   weights[32];
-    float mins[QK_K/32];
-    float scales[QK_K/32];
-
-    for (int i = 0; i < nb; i++) {
-
-        float max_scale = 0; // as we are deducting the min, scales are always positive
-        float max_min = 0;
-        for (int j = 0; j < QK_K/32; ++j) {
-            //scales[j] = make_qkx1_quants(32, 15, x + 32*j, L + 32*j, &mins[j], 9, 0.5f);
-            float sum_x2 = 0;
-            for (int l = 0; l < 32; ++l) sum_x2 += x[32*j + l] * x[32*j + l];
-            float av_x = sqrtf(sum_x2/32);
-            for (int l = 0; l < 32; ++l) weights[l] = av_x + fabsf(x[32*j + l]);
-            scales[j] = make_qkx2_quants(32, 15, x + 32*j, weights, L + 32*j, &mins[j], Laux, -1.f, 0.1f, 20, false);
-            float scale = scales[j];
-            if (scale > max_scale) {
-                max_scale = scale;
-            }
-            float min = mins[j];
-            if (min > max_min) {
-                max_min = min;
-            }
-        }
-
-#if QK_K == 256
-        float inv_scale = max_scale > 0 ? 63.f/max_scale : 0.f;
-        float inv_min   = max_min   > 0 ? 63.f/max_min   : 0.f;
-        for (int j = 0; j < QK_K/32; ++j) {
-            uint8_t ls = nearest_int(inv_scale*scales[j]);
-            uint8_t lm = nearest_int(inv_min*mins[j]);
-            ls = MIN(63, ls);
-            lm = MIN(63, lm);
-            if (j < 4) {
-                y[i].scales[j] = ls;
-                y[i].scales[j+4] = lm;
-            } else {
-                y[i].scales[j+4] = (ls & 0xF) | ((lm & 0xF) << 4);
-                y[i].scales[j-4] |= ((ls >> 4) << 6);
-                y[i].scales[j-0] |= ((lm >> 4) << 6);
-            }
-        }
-        y[i].d = GGML_FP32_TO_FP16(max_scale/63.f);
-        y[i].dmin = GGML_FP32_TO_FP16(max_min/63.f);
-
-        uint8_t sc, m;
-        for (int j = 0; j < QK_K/32; ++j) {
-            get_scale_min_k4(j, y[i].scales, &sc, &m);
-            const float d = GGML_FP16_TO_FP32(y[i].d) * sc;
-            if (!d) continue;
-            const float dm = GGML_FP16_TO_FP32(y[i].dmin) * m;
-            for (int ii = 0; ii < 32; ++ii) {
-                int l = nearest_int((x[32*j + ii] + dm)/d);
-                l = MAX(0, MIN(15, l));
-                L[32*j + ii] = l;
-            }
-        }
-#else
-        const float s_factor = 15.f;
-        float inv_scale = max_scale > 0 ? s_factor/max_scale : 0.f;
-        float inv_min   = max_min   > 0 ? s_factor/max_min   : 0.f;
-        int d1 = nearest_int(inv_scale*scales[0]);
-        int m1 = nearest_int(inv_min*mins[0]);
-        int d2 = nearest_int(inv_scale*scales[1]);
-        int m2 = nearest_int(inv_min*mins[1]);
-        y[i].scales[0] = d1 | (m1 << 4);
-        y[i].scales[1] = d2 | (m2 << 4);
-        y[i].d[0] = GGML_FP32_TO_FP16(max_scale/s_factor);
-        y[i].d[1] = GGML_FP32_TO_FP16(max_min/s_factor);
-
-        float sumlx = 0;
-        int   suml2 = 0;
-        for (int j = 0; j < QK_K/32; ++j) {
-            const uint8_t sd = y[i].scales[j] & 0xF;
-            const uint8_t sm = y[i].scales[j] >>  4;
-            const float d = GGML_FP16_TO_FP32(y[i].d[0]) * sd;
-            if (!d) continue;
-            const float m = GGML_FP16_TO_FP32(y[i].d[1]) * sm;
-            for (int ii = 0; ii < 32; ++ii) {
-                int l = nearest_int((x[32*j + ii] + m)/d);
-                l = MAX(0, MIN(15, l));
-                L[32*j + ii] = l;
-                sumlx += (x[32*j + ii] + m)*l*sd;
-                suml2 += l*l*sd*sd;
-            }
-        }
-        if (suml2) {
-            y[i].d[0] = GGML_FP32_TO_FP16(sumlx/suml2);
-        }
-#endif
-        uint8_t * q = y[i].qs;
-        for (int j = 0; j < QK_K; j += 64) {
-            for (int l = 0; l < 32; ++l) q[l] = L[j + l] | (L[j + l + 32] << 4);
-            q += 32;
-        }
-
-        x += QK_K;
-
-    }
-}
-
-void dequantize_row_q4_K(const block_q4_K * restrict x, float * restrict y, int k) {
-    assert(k % QK_K == 0);
-    const int nb = k / QK_K;
-
-    for (int i = 0; i < nb; i++) {
-
-        const uint8_t * q = x[i].qs;
-
-#if QK_K == 256
-
-        const float d   = GGML_FP16_TO_FP32(x[i].d);
-        const float min = GGML_FP16_TO_FP32(x[i].dmin);
-
-        int is = 0;
-        uint8_t sc, m;
-        for (int j = 0; j < QK_K; j += 64) {
-            get_scale_min_k4(is + 0, x[i].scales, &sc, &m);
-            const float d1 = d * sc; const float m1 = min * m;
-            get_scale_min_k4(is + 1, x[i].scales, &sc, &m);
-            const float d2 = d * sc; const float m2 = min * m;
-            for (int l = 0; l < 32; ++l) *y++ = d1 * (q[l] & 0xF) - m1;
-            for (int l = 0; l < 32; ++l) *y++ = d2 * (q[l]  >> 4) - m2;
-            q += 32; is += 2;
-        }
-#else
-        const float dall = GGML_FP16_TO_FP32(x[i].d[0]);
-        const float mall = GGML_FP16_TO_FP32(x[i].d[1]);
-        const float d1 = dall * (x[i].scales[0] & 0xF), m1 = mall * (x[i].scales[0] >> 4);
-        const float d2 = dall * (x[i].scales[1] & 0xF), m2 = mall * (x[i].scales[1] >> 4);
-        for (int l = 0; l < 32; ++l) {
-            y[l+ 0] = d1 * (q[l] & 0xF) - m1;
-            y[l+32] = d2 * (q[l] >>  4) - m2;
-        }
-        y += QK_K;
-#endif
-
-    }
-}
-
-void quantize_row_q4_K(const float * restrict x, void * restrict vy, int k) {
-    assert(k % QK_K == 0);
-    block_q4_K * restrict y = vy;
-    quantize_row_q4_K_reference(x, y, k);
-}
-
-size_t ggml_quantize_q4_K(const float * restrict src, void * restrict dst, int n, int k, int64_t * restrict hist) {
-    assert(k % QK_K == 0);
-    (void)hist; // TODO: collect histograms
-
-    for (int j = 0; j < n; j += k) {
-        block_q4_K * restrict y = (block_q4_K *)dst + j/QK_K;
-        quantize_row_q4_K_reference(src + j, y, k);
-    }
-    return (n/QK_K*sizeof(block_q4_K));
-}
-
-// ====================== 5-bit (de)-quantization
-
-void quantize_row_q5_K_reference(const float * restrict x, block_q5_K * restrict y, int k) {
-    assert(k % QK_K == 0);
-    const int nb = k / QK_K;
-
-#if QK_K == 256
-    uint8_t L[QK_K];
-    float mins[QK_K/32];
-    float scales[QK_K/32];
-    float weights[32];
-    uint8_t Laux[32];
-#else
-    int8_t L[QK_K];
-    float scales[QK_K/16];
-#endif
-
-    for (int i = 0; i < nb; i++) {
-
-#if QK_K == 256
-
-        float max_scale = 0; // as we are deducting the min, scales are always positive
-        float max_min = 0;
-        for (int j = 0; j < QK_K/32; ++j) {
-            //scales[j] = make_qkx1_quants(32, 31, x + 32*j, L + 32*j, &mins[j], 9, 0.5f);
-            float sum_x2 = 0;
-            for (int l = 0; l < 32; ++l) sum_x2 += x[32*j + l] * x[32*j + l];
-            float av_x = sqrtf(sum_x2/32);
-            for (int l = 0; l < 32; ++l) weights[l] = av_x + fabsf(x[32*j + l]);
-            scales[j] = make_qkx2_quants(32, 31, x + 32*j, weights, L + 32*j, &mins[j], Laux, -0.5f, 0.1f, 15, false);
-            float scale = scales[j];
-            if (scale > max_scale) {
-                max_scale = scale;
-            }
-            float min = mins[j];
-            if (min > max_min) {
-                max_min = min;
-            }
-        }
-
-        float inv_scale = max_scale > 0 ? 63.f/max_scale : 0.f;
-        float inv_min   = max_min   > 0 ? 63.f/max_min   : 0.f;
-        for (int j = 0; j < QK_K/32; ++j) {
-            uint8_t ls = nearest_int(inv_scale*scales[j]);
-            uint8_t lm = nearest_int(inv_min*mins[j]);
-            ls = MIN(63, ls);
-            lm = MIN(63, lm);
-            if (j < 4) {
-                y[i].scales[j] = ls;
-                y[i].scales[j+4] = lm;
-            } else {
-                y[i].scales[j+4] = (ls & 0xF) | ((lm & 0xF) << 4);
-                y[i].scales[j-4] |= ((ls >> 4) << 6);
-                y[i].scales[j-0] |= ((lm >> 4) << 6);
-            }
-        }
-        y[i].d = GGML_FP32_TO_FP16(max_scale/63.f);
-        y[i].dmin = GGML_FP32_TO_FP16(max_min/63.f);
-
-        uint8_t sc, m;
-        for (int j = 0; j < QK_K/32; ++j) {
-            get_scale_min_k4(j, y[i].scales, &sc, &m);
-            const float d = GGML_FP16_TO_FP32(y[i].d) * sc;
-            if (!d) continue;
-            const float dm = GGML_FP16_TO_FP32(y[i].dmin) * m;
-            for (int ii = 0; ii < 32; ++ii) {
-                int l = nearest_int((x[32*j + ii] + dm)/d);
-                l = MAX(0, MIN(31, l));
-                L[32*j + ii] = l;
-            }
-        }
-
-        uint8_t * restrict qh = y[i].qh;
-        uint8_t * restrict ql = y[i].qs;
-        memset(qh, 0, QK_K/8);
-
-        uint8_t m1 = 1, m2 = 2;
-        for (int n = 0; n < QK_K; n += 64) {
-            for (int j = 0; j < 32; ++j) {
-                int l1 = L[n + j];
-                if (l1 > 15) {
-                    l1 -= 16; qh[j] |= m1;
-                }
-                int l2 = L[n + j + 32];
-                if (l2 > 15) {
-                    l2 -= 16; qh[j] |= m2;
-                }
-                ql[j] = l1 | (l2 << 4);
-            }
-            m1 <<= 2; m2 <<= 2;
-            ql += 32;
-        }
-#else
-        float max_scale = 0, amax = 0;
-        for (int j = 0; j < QK_K/16; ++j) {
-            scales[j] = make_qx_quants(16, 16, x + 16*j, L + 16*j, 1);
-            float abs_scale = fabsf(scales[j]);
-            if (abs_scale > amax) {
-                amax = abs_scale;
-                max_scale = scales[j];
-            }
-        }
-
-        float iscale = -128.f/max_scale;
-        for (int j = 0; j < QK_K/16; ++j) {
-            int l = nearest_int(iscale*scales[j]);
-            y[i].scales[j] = MAX(-128, MIN(127, l));
-        }
-        y[i].d = GGML_FP32_TO_FP16(1/iscale);
-
-        for (int j = 0; j < QK_K/16; ++j) {
-            const float d = GGML_FP16_TO_FP32(y[i].d) * y[i].scales[j];
-            if (!d) continue;
-            for (int ii = 0; ii < 16; ++ii) {
-                int l = nearest_int(x[16*j + ii]/d);
-                l = MAX(-16, MIN(15, l));
-                L[16*j + ii] = l + 16;
-            }
-        }
-
-        uint8_t * restrict qh = y[i].qh;
-        uint8_t * restrict ql = y[i].qs;
-        memset(qh, 0, QK_K/8);
-
-        for (int j = 0; j < 32; ++j) {
-            int jm = j%8;
-            int is = j/8;
-            int l1 = L[j];
-            if (l1 > 15) {
-                l1 -= 16; qh[jm] |= (1 << is);
-            }
-            int l2 = L[j + 32];
-            if (l2 > 15) {
-                l2 -= 16; qh[jm] |= (1 << (4 + is));
-            }
-            ql[j] = l1 | (l2 << 4);
-        }
-#endif
-
-        x += QK_K;
-
-    }
-}
-
-void dequantize_row_q5_K(const block_q5_K * restrict x, float * restrict y, int k) {
-    assert(k % QK_K == 0);
-    const int nb = k / QK_K;
-
-    for (int i = 0; i < nb; i++) {
-
-        const uint8_t * ql = x[i].qs;
-        const uint8_t * qh = x[i].qh;
-
-#if QK_K == 256
-
-        const float d = GGML_FP16_TO_FP32(x[i].d);
-        const float min = GGML_FP16_TO_FP32(x[i].dmin);
-
-        int is = 0;
-        uint8_t sc, m;
-        uint8_t u1 = 1, u2 = 2;
-        for (int j = 0; j < QK_K; j += 64) {
-            get_scale_min_k4(is + 0, x[i].scales, &sc, &m);
-            const float d1 = d * sc; const float m1 = min * m;
-            get_scale_min_k4(is + 1, x[i].scales, &sc, &m);
-            const float d2 = d * sc; const float m2 = min * m;
-            for (int l = 0; l < 32; ++l) *y++ = d1 * ((ql[l] & 0xF) + (qh[l] & u1 ? 16 : 0)) - m1;
-            for (int l = 0; l < 32; ++l) *y++ = d2 * ((ql[l]  >> 4) + (qh[l] & u2 ? 16 : 0)) - m2;
-            ql += 32; is += 2;
-            u1 <<= 2; u2 <<= 2;
-        }
-#else
-        float d = GGML_FP16_TO_FP32(x[i].d);
-        const int8_t * restrict s = x[i].scales;
-        for (int l = 0; l < 8; ++l) {
-            y[l+ 0] = d * s[0] * ((ql[l+ 0] & 0xF) - (qh[l] & 0x01 ? 0 : 16));
-            y[l+ 8] = d * s[0] * ((ql[l+ 8] & 0xF) - (qh[l] & 0x02 ? 0 : 16));
-            y[l+16] = d * s[1] * ((ql[l+16] & 0xF) - (qh[l] & 0x04 ? 0 : 16));
-            y[l+24] = d * s[1] * ((ql[l+24] & 0xF) - (qh[l] & 0x08 ? 0 : 16));
-            y[l+32] = d * s[2] * ((ql[l+ 0] >>  4) - (qh[l] & 0x10 ? 0 : 16));
-            y[l+40] = d * s[2] * ((ql[l+ 8] >>  4) - (qh[l] & 0x20 ? 0 : 16));
-            y[l+48] = d * s[3] * ((ql[l+16] >>  4) - (qh[l] & 0x40 ? 0 : 16));
-            y[l+56] = d * s[3] * ((ql[l+24] >>  4) - (qh[l] & 0x80 ? 0 : 16));
-        }
-        y += QK_K;
-#endif
-    }
-}
-
-void quantize_row_q5_K(const float * restrict x, void * restrict vy, int k) {
-    assert(k % QK_K == 0);
-    block_q5_K * restrict y = vy;
-    quantize_row_q5_K_reference(x, y, k);
-}
-
-size_t ggml_quantize_q5_K(const float * restrict src, void * restrict dst, int n, int k, int64_t * restrict hist) {
-    assert(k % QK_K == 0);
-    (void)hist; // TODO: collect histograms
-
-    for (int j = 0; j < n; j += k) {
-        block_q5_K * restrict y = (block_q5_K *)dst + j/QK_K;
-        quantize_row_q5_K_reference(src + j, y, k);
-    }
-    return (n/QK_K*sizeof(block_q5_K));
-}
-
-// ====================== 6-bit (de)-quantization
-
-void quantize_row_q6_K_reference(const float * restrict x, block_q6_K * restrict y, int k) {
-    assert(k % QK_K == 0);
-    const int nb = k / QK_K;
-
-    int8_t L[QK_K];
-    float   scales[QK_K/16];
-
-    for (int i = 0; i < nb; i++) {
-
-        float max_scale = 0;
-        float max_abs_scale = 0;
-
-        for (int ib = 0; ib < QK_K/16; ++ib) {
-
-            const float scale = make_qx_quants(16, 32, x + 16*ib, L + 16*ib, 1);
-            scales[ib] = scale;
-
-            const float abs_scale = fabsf(scale);
-            if (abs_scale > max_abs_scale) {
-                max_abs_scale = abs_scale;
-                max_scale = scale;
-            }
-
-        }
-
-        if (!max_abs_scale) {
-            memset(&y[i], 0, sizeof(block_q6_K));
-            y[i].d = GGML_FP32_TO_FP16(0.f);
-            x += QK_K;
-            continue;
-        }
-
-        float iscale = -128.f/max_scale;
-        y[i].d = GGML_FP32_TO_FP16(1/iscale);
-        for (int ib = 0; ib < QK_K/16; ++ib) {
-            y[i].scales[ib] = MIN(127, nearest_int(iscale*scales[ib]));
-        }
-
-        for (int j = 0; j < QK_K/16; ++j) {
-            float d = GGML_FP16_TO_FP32(y[i].d) * y[i].scales[j];
-            if (!d) {
-                continue;
-            }
-            for (int ii = 0; ii < 16; ++ii) {
-                int l = nearest_int(x[16*j + ii]/d);
-                l = MAX(-32, MIN(31, l));
-                L[16*j + ii] = l + 32;
-            }
-        }
-
-        uint8_t * restrict ql = y[i].ql;
-        uint8_t * restrict qh = y[i].qh;
-#if QK_K == 256
-        for (int j = 0; j < QK_K; j += 128) {
-            for (int l = 0; l < 32; ++l) {
-                const uint8_t q1 = L[j + l +  0] & 0xF;
-                const uint8_t q2 = L[j + l + 32] & 0xF;
-                const uint8_t q3 = L[j + l + 64] & 0xF;
-                const uint8_t q4 = L[j + l + 96] & 0xF;
-                ql[l+ 0] = q1 | (q3 << 4);
-                ql[l+32] = q2 | (q4 << 4);
-                qh[l] = (L[j + l] >> 4) | ((L[j + l + 32] >> 4) << 2) | ((L[j + l + 64] >> 4) << 4) | ((L[j + l + 96] >> 4) << 6);
-            }
-            ql += 64;
-            qh += 32;
-        }
-#else
-        for (int l = 0; l < 32; ++l) {
-            const uint8_t q1 = L[l +  0] & 0xF;
-            const uint8_t q2 = L[l + 32] & 0xF;
-            ql[l] = q1 | (q2 << 4);
-        }
-        for (int l = 0; l < 16; ++l) {
-            qh[l] = (L[l] >> 4) | ((L[l + 16] >> 4) << 2) | ((L[l + 32] >> 4) << 4) | ((L[l + 48] >> 4) << 6);
-        }
-#endif
-
-        x += QK_K;
-
-    }
-}
-
-void dequantize_row_q6_K(const block_q6_K * restrict x, float * restrict y, int k) {
-    assert(k % QK_K == 0);
-    const int nb = k / QK_K;
-
-    for (int i = 0; i < nb; i++) {
-
-        const float d = GGML_FP16_TO_FP32(x[i].d);
-
-        const uint8_t * restrict ql = x[i].ql;
-        const uint8_t * restrict qh = x[i].qh;
-        const int8_t  * restrict sc = x[i].scales;
-
-#if QK_K == 256
-        for (int n = 0; n < QK_K; n += 128) {
-            for (int l = 0; l < 32; ++l) {
-                int is = l/16;
-                const int8_t q1 = (int8_t)((ql[l +  0] & 0xF) | (((qh[l] >> 0) & 3) << 4)) - 32;
-                const int8_t q2 = (int8_t)((ql[l + 32] & 0xF) | (((qh[l] >> 2) & 3) << 4)) - 32;
-                const int8_t q3 = (int8_t)((ql[l +  0]  >> 4) | (((qh[l] >> 4) & 3) << 4)) - 32;
-                const int8_t q4 = (int8_t)((ql[l + 32]  >> 4) | (((qh[l] >> 6) & 3) << 4)) - 32;
-                y[l +  0] = d * sc[is + 0] * q1;
-                y[l + 32] = d * sc[is + 2] * q2;
-                y[l + 64] = d * sc[is + 4] * q3;
-                y[l + 96] = d * sc[is + 6] * q4;
-            }
-            y  += 128;
-            ql += 64;
-            qh += 32;
-            sc += 8;
-        }
-#else
-        for (int l = 0; l < 16; ++l) {
-            const int8_t q1 = (int8_t)((ql[l+ 0] & 0xF) | (((qh[l] >> 0) & 3) << 4)) - 32;
-            const int8_t q2 = (int8_t)((ql[l+16] & 0xF) | (((qh[l] >> 2) & 3) << 4)) - 32;
-            const int8_t q3 = (int8_t)((ql[l+ 0]  >> 4) | (((qh[l] >> 4) & 3) << 4)) - 32;
-            const int8_t q4 = (int8_t)((ql[l+16]  >> 4) | (((qh[l] >> 6) & 3) << 4)) - 32;
-            y[l+ 0] = d * sc[0] * q1;
-            y[l+16] = d * sc[1] * q2;
-            y[l+32] = d * sc[2] * q3;
-            y[l+48] = d * sc[3] * q4;
-        }
-        y  += 64;
-#endif
-
-    }
-}
-
-void quantize_row_q6_K(const float * restrict x, void * restrict vy, int k) {
-    assert(k % QK_K == 0);
-    block_q6_K * restrict y = vy;
-    quantize_row_q6_K_reference(x, y, k);
-}
-
-size_t ggml_quantize_q6_K(const float * src, void * dst, int n, int k, int64_t * hist) {
-    assert(k % QK_K == 0);
-    (void)hist; // TODO: collect histograms
-
-    for (int j = 0; j < n; j += k) {
-        block_q6_K * restrict y = (block_q6_K *)dst + j/QK_K;
-        quantize_row_q6_K_reference(src + j, y, k);
-    }
-    return (n/QK_K*sizeof(block_q6_K));
-}
-
-//===================================== Q8_K ==============================================
-
-void quantize_row_q8_K_reference(const float * restrict x, block_q8_K * restrict y, int k) {
-    assert(k % QK_K == 0);
-    const int nb = k / QK_K;
-
-    for (int i = 0; i < nb; i++) {
-
-        float max = 0;
-        float amax = 0;
-        for (int j = 0; j < QK_K; ++j) {
-            float ax = fabsf(x[j]);
-            if (ax > amax) {
-                amax = ax; max = x[j];
-            }
-        }
-        if (!amax) {
-            y[i].d = 0;
-            memset(y[i].qs, 0, QK_K);
-            x += QK_K;
-            continue;
-        }
-        const float iscale = -128.f/max;
-        for (int j = 0; j < QK_K; ++j) {
-            int v = nearest_int(iscale*x[j]);
-            y[i].qs[j] = MIN(127, v);
-        }
-        for (int j = 0; j < QK_K/16; ++j) {
-            int sum = 0;
-            for (int ii = 0; ii < 16; ++ii) {
-                sum += y[i].qs[j*16 + ii];
-            }
-            y[i].bsums[j] = sum;
-        }
-        y[i].d = 1/iscale;
-        x += QK_K;
-    }
-}
-
-void dequantize_row_q8_K(const block_q8_K * restrict x, float * restrict y, int k) {
-    assert(k % QK_K == 0);
-    const int nb = k / QK_K;
-
-    for (int i = 0; i < nb; i++) {
-        for (int j = 0; j < QK_K; ++j) {
-            *y++ = x[i].d * x[i].qs[j];
-        }
-    }
-}
-
-void quantize_row_q8_K(const float * restrict x, void * restrict y, int k) {
-    quantize_row_q8_K_reference(x, y, k);
-}
-
-//===================================== Dot ptoducts =================================
-
-//
-// Helper functions
-//
-#if __AVX__ || __AVX2__ || __AVX512F__
-
-// shuffles to pick the required scales in dot products
-static inline __m256i get_scale_shuffle_q3k(int i) {
-    static const uint8_t k_shuffle[128] = {
-         0, 1, 0, 1, 0, 1, 0, 1, 0, 1, 0, 1, 0, 1, 0, 1,     2, 3, 2, 3, 2, 3, 2, 3, 2, 3, 2, 3, 2, 3, 2, 3,
-         4, 5, 4, 5, 4, 5, 4, 5, 4, 5, 4, 5, 4, 5, 4, 5,     6, 7, 6, 7, 6, 7, 6, 7, 6, 7, 6, 7, 6, 7, 6, 7,
-         8, 9, 8, 9, 8, 9, 8, 9, 8, 9, 8, 9, 8, 9, 8, 9,    10,11,10,11,10,11,10,11,10,11,10,11,10,11,10,11,
-        12,13,12,13,12,13,12,13,12,13,12,13,12,13,12,13,    14,15,14,15,14,15,14,15,14,15,14,15,14,15,14,15,
-    };
-    return _mm256_loadu_si256((const __m256i*)k_shuffle + i);
-}
-static inline __m256i get_scale_shuffle_k4(int i) {
-    static const uint8_t k_shuffle[256] = {
-         0, 1, 0, 1, 0, 1, 0, 1, 0, 1, 0, 1, 0, 1, 0, 1, 0, 1, 0, 1, 0, 1, 0, 1, 0, 1, 0, 1, 0, 1, 0, 1,
-         2, 3, 2, 3, 2, 3, 2, 3, 2, 3, 2, 3, 2, 3, 2, 3, 2, 3, 2, 3, 2, 3, 2, 3, 2, 3, 2, 3, 2, 3, 2, 3,
-         4, 5, 4, 5, 4, 5, 4, 5, 4, 5, 4, 5, 4, 5, 4, 5, 4, 5, 4, 5, 4, 5, 4, 5, 4, 5, 4, 5, 4, 5, 4, 5,
-         6, 7, 6, 7, 6, 7, 6, 7, 6, 7, 6, 7, 6, 7, 6, 7, 6, 7, 6, 7, 6, 7, 6, 7, 6, 7, 6, 7, 6, 7, 6, 7,
-         8, 9, 8, 9, 8, 9, 8, 9, 8, 9, 8, 9, 8, 9, 8, 9, 8, 9, 8, 9, 8, 9, 8, 9, 8, 9, 8, 9, 8, 9, 8, 9,
-        10,11,10,11,10,11,10,11,10,11,10,11,10,11,10,11,10,11,10,11,10,11,10,11,10,11,10,11,10,11,10,11,
-        12,13,12,13,12,13,12,13,12,13,12,13,12,13,12,13,12,13,12,13,12,13,12,13,12,13,12,13,12,13,12,13,
-        14,15,14,15,14,15,14,15,14,15,14,15,14,15,14,15,14,15,14,15,14,15,14,15,14,15,14,15,14,15,14,15
-    };
-    return _mm256_loadu_si256((const __m256i*)k_shuffle + i);
-}
-static inline __m128i get_scale_shuffle(int i) {
-    static const uint8_t k_shuffle[128] = {
-         0, 0, 0, 0, 0, 0, 0, 0, 1, 1, 1, 1, 1, 1, 1, 1,
-         2, 2, 2, 2, 2, 2, 2, 2, 3, 3, 3, 3, 3, 3, 3, 3,
-         4, 4, 4, 4, 4, 4, 4, 4, 5, 5, 5, 5, 5, 5, 5, 5,
-         6, 6, 6, 6, 6, 6, 6, 6, 7, 7, 7, 7, 7, 7, 7, 7,
-         8, 8, 8, 8, 8, 8, 8, 8, 9, 9, 9, 9, 9, 9, 9, 9,
-        10,10,10,10,10,10,10,10, 11,11,11,11,11,11,11,11,
-        12,12,12,12,12,12,12,12, 13,13,13,13,13,13,13,13,
-        14,14,14,14,14,14,14,14, 15,15,15,15,15,15,15,15
-    };
-    return _mm_loadu_si128((const __m128i*)k_shuffle + i);
-}
-#endif
-
-void ggml_vec_dot_q4_0_q8_0(int n, float * restrict s, const void * restrict vx, const void * restrict vy) {
-    const int qk = QK8_0;
-    const int nb = n / qk;
-
-    assert(n % qk == 0);
-
-    const block_q4_0 * restrict x = vx;
-    const block_q8_0 * restrict y = vy;
-
-#if defined(__ARM_NEON)
-    float32x4_t sumv0 = vdupq_n_f32(0.0f);
-    float32x4_t sumv1 = vdupq_n_f32(0.0f);
-
-    assert(nb % 2 == 0); // TODO: handle odd nb
-
-    for (int i = 0; i < nb; i += 2) {
-        const block_q4_0 * restrict x0 = &x[i + 0];
-        const block_q4_0 * restrict x1 = &x[i + 1];
-        const block_q8_0 * restrict y0 = &y[i + 0];
-        const block_q8_0 * restrict y1 = &y[i + 1];
-
-        const uint8x16_t m4b = vdupq_n_u8(0x0F);
-        const int8x16_t  s8b = vdupq_n_s8(0x8);
-
-        const uint8x16_t v0_0 = vld1q_u8(x0->qs);
-        const uint8x16_t v0_1 = vld1q_u8(x1->qs);
-
-        // 4-bit -> 8-bit
-        const int8x16_t v0_0l = vreinterpretq_s8_u8(vandq_u8  (v0_0, m4b));
-        const int8x16_t v0_0h = vreinterpretq_s8_u8(vshrq_n_u8(v0_0, 4));
-        const int8x16_t v0_1l = vreinterpretq_s8_u8(vandq_u8  (v0_1, m4b));
-        const int8x16_t v0_1h = vreinterpretq_s8_u8(vshrq_n_u8(v0_1, 4));
-
-        // sub 8
-        const int8x16_t v0_0ls = vsubq_s8(v0_0l, s8b);
-        const int8x16_t v0_0hs = vsubq_s8(v0_0h, s8b);
-        const int8x16_t v0_1ls = vsubq_s8(v0_1l, s8b);
-        const int8x16_t v0_1hs = vsubq_s8(v0_1h, s8b);
-
-        // load y
-        const int8x16_t v1_0l = vld1q_s8(y0->qs);
-        const int8x16_t v1_0h = vld1q_s8(y0->qs + 16);
-        const int8x16_t v1_1l = vld1q_s8(y1->qs);
-        const int8x16_t v1_1h = vld1q_s8(y1->qs + 16);
-
-        // dot product into int32x4_t
-        const int32x4_t p_0 = ggml_vdotq_s32(ggml_vdotq_s32(vdupq_n_s32(0), v0_0ls, v1_0l), v0_0hs, v1_0h);
-        const int32x4_t p_1 = ggml_vdotq_s32(ggml_vdotq_s32(vdupq_n_s32(0), v0_1ls, v1_1l), v0_1hs, v1_1h);
-
-        sumv0 = vmlaq_n_f32(sumv0, vcvtq_f32_s32(p_0), GGML_FP16_TO_FP32(x0->d)*GGML_FP16_TO_FP32(y0->d));
-        sumv1 = vmlaq_n_f32(sumv1, vcvtq_f32_s32(p_1), GGML_FP16_TO_FP32(x1->d)*GGML_FP16_TO_FP32(y1->d));
-    }
-
-    *s = vaddvq_f32(sumv0) + vaddvq_f32(sumv1);
-#elif defined(__AVX2__)
-    // Initialize accumulator with zeros
-    __m256 acc = _mm256_setzero_ps();
-
-    // Main loop
-    for (int i = 0; i < nb; ++i) {
-        /* Compute combined scale for the block */
-        const __m256 d = _mm256_set1_ps( GGML_FP16_TO_FP32(x[i].d) * GGML_FP16_TO_FP32(y[i].d) );
-
-        __m256i bx = bytes_from_nibbles_32(x[i].qs);
-
-        // Now we have a vector with bytes in [ 0 .. 15 ] interval. Offset them into [ -8 .. +7 ] interval.
-        const __m256i off = _mm256_set1_epi8( 8 );
-        bx = _mm256_sub_epi8( bx, off );
-
-        __m256i by = _mm256_loadu_si256((const __m256i *)y[i].qs);
-
-        const __m256 q = mul_sum_i8_pairs_float(bx, by);
-
-        /* Multiply q with scale and accumulate */
-        acc = _mm256_fmadd_ps( d, q, acc );
-    }
-
-    *s = hsum_float_8(acc);
-#elif defined(__AVX__)
-    // Initialize accumulator with zeros
-    __m256 acc = _mm256_setzero_ps();
-
-    // Main loop
-    for (int i = 0; i < nb; ++i) {
-        // Compute combined scale for the block
-        const __m256 d = _mm256_set1_ps( GGML_FP16_TO_FP32(x[i].d) * GGML_FP16_TO_FP32(y[i].d) );
-
-        const __m128i lowMask = _mm_set1_epi8(0xF);
-        const __m128i off = _mm_set1_epi8(8);
-
-        const __m128i tmp = _mm_loadu_si128((const __m128i *)x[i].qs);
-
-        __m128i bx = _mm_and_si128(lowMask, tmp);
-        __m128i by = _mm_loadu_si128((const __m128i *)y[i].qs);
-        bx = _mm_sub_epi8(bx, off);
-        const __m128i i32_0 = mul_sum_i8_pairs(bx, by);
-
-        bx = _mm_and_si128(lowMask, _mm_srli_epi64(tmp, 4));
-        by = _mm_loadu_si128((const __m128i *)(y[i].qs + 16));
-        bx = _mm_sub_epi8(bx, off);
-        const __m128i i32_1 = mul_sum_i8_pairs(bx, by);
-
-        // Convert int32_t to float
-        __m256 p = _mm256_cvtepi32_ps(MM256_SET_M128I(i32_0, i32_1));
-
-        // Apply the scale, and accumulate
-        acc = _mm256_add_ps(_mm256_mul_ps( d, p ), acc);
-    }
-
-    *s = hsum_float_8(acc);
-#elif defined(__SSSE3__)
-    // set constants
-    const __m128i lowMask = _mm_set1_epi8(0xF);
-    const __m128i off = _mm_set1_epi8(8);
-
-    // Initialize accumulator with zeros
-    __m128 acc_0 = _mm_setzero_ps();
-    __m128 acc_1 = _mm_setzero_ps();
-    __m128 acc_2 = _mm_setzero_ps();
-    __m128 acc_3 = _mm_setzero_ps();
-
-    // First round without accumulation
-    {
-        _mm_prefetch(&x[0] + sizeof(block_q4_0), _MM_HINT_T0);
-        _mm_prefetch(&y[0] + sizeof(block_q8_0), _MM_HINT_T0);
-
-        // Compute combined scale for the block 0 and 1
-        const __m128 d_0_1 = _mm_set1_ps( GGML_FP16_TO_FP32(x[0].d) * GGML_FP16_TO_FP32(y[0].d) );
-
-        const __m128i tmp_0_1 = _mm_loadu_si128((const __m128i *)x[0].qs);
-
-        __m128i bx_0 = _mm_and_si128(lowMask, tmp_0_1);
-        __m128i by_0 = _mm_loadu_si128((const __m128i *)y[0].qs);
-        bx_0 = _mm_sub_epi8(bx_0, off);
-        const __m128i i32_0 = mul_sum_i8_pairs(bx_0, by_0);
-
-        __m128i bx_1 = _mm_and_si128(lowMask, _mm_srli_epi64(tmp_0_1, 4));
-        __m128i by_1 = _mm_loadu_si128((const __m128i *)(y[0].qs + 16));
-        bx_1 = _mm_sub_epi8(bx_1, off);
-        const __m128i i32_1 = mul_sum_i8_pairs(bx_1, by_1);
-
-        _mm_prefetch(&x[1] + sizeof(block_q4_0), _MM_HINT_T0);
-        _mm_prefetch(&y[1] + sizeof(block_q8_0), _MM_HINT_T0);
-
-        // Compute combined scale for the block 2 and 3
-        const __m128 d_2_3 = _mm_set1_ps( GGML_FP16_TO_FP32(x[1].d) * GGML_FP16_TO_FP32(y[1].d) );
-
-        const __m128i tmp_2_3 = _mm_loadu_si128((const __m128i *)x[1].qs);
-
-        __m128i bx_2 = _mm_and_si128(lowMask, tmp_2_3);
-        __m128i by_2 = _mm_loadu_si128((const __m128i *)y[1].qs);
-        bx_2 = _mm_sub_epi8(bx_2, off);
-        const __m128i i32_2 = mul_sum_i8_pairs(bx_2, by_2);
-
-        __m128i bx_3 = _mm_and_si128(lowMask, _mm_srli_epi64(tmp_2_3, 4));
-        __m128i by_3 = _mm_loadu_si128((const __m128i *)(y[1].qs + 16));
-        bx_3 = _mm_sub_epi8(bx_3, off);
-        const __m128i i32_3 = mul_sum_i8_pairs(bx_3, by_3);
-
-        // Convert int32_t to float
-        __m128 p0 = _mm_cvtepi32_ps(i32_0);
-        __m128 p1 = _mm_cvtepi32_ps(i32_1);
-        __m128 p2 = _mm_cvtepi32_ps(i32_2);
-        __m128 p3 = _mm_cvtepi32_ps(i32_3);
-
-        // Apply the scale
-        acc_0 = _mm_mul_ps( d_0_1, p0 );
-        acc_1 = _mm_mul_ps( d_0_1, p1 );
-        acc_2 = _mm_mul_ps( d_2_3, p2 );
-        acc_3 = _mm_mul_ps( d_2_3, p3 );
-    }
-
-    assert(nb % 2 == 0); // TODO: handle odd nb
-
-    // Main loop
-    for (int i = 2; i < nb; i+=2) {
-        _mm_prefetch(&x[i] + sizeof(block_q4_0), _MM_HINT_T0);
-        _mm_prefetch(&y[i] + sizeof(block_q8_0), _MM_HINT_T0);
-
-        // Compute combined scale for the block 0 and 1
-        const __m128 d_0_1 = _mm_set1_ps( GGML_FP16_TO_FP32(x[i].d) * GGML_FP16_TO_FP32(y[i].d) );
-
-        const __m128i tmp_0_1 = _mm_loadu_si128((const __m128i *)x[i].qs);
-
-        __m128i bx_0 = _mm_and_si128(lowMask, tmp_0_1);
-        __m128i by_0 = _mm_loadu_si128((const __m128i *)y[i].qs);
-        bx_0 = _mm_sub_epi8(bx_0, off);
-        const __m128i i32_0 = mul_sum_i8_pairs(bx_0, by_0);
-
-        __m128i bx_1 = _mm_and_si128(lowMask, _mm_srli_epi64(tmp_0_1, 4));
-        __m128i by_1 = _mm_loadu_si128((const __m128i *)(y[i].qs + 16));
-        bx_1 = _mm_sub_epi8(bx_1, off);
-        const __m128i i32_1 = mul_sum_i8_pairs(bx_1, by_1);
-
-        _mm_prefetch(&x[i] + 2 * sizeof(block_q4_0), _MM_HINT_T0);
-        _mm_prefetch(&y[i] + 2 * sizeof(block_q8_0), _MM_HINT_T0);
-
-        // Compute combined scale for the block 2 and 3
-        const __m128 d_2_3 = _mm_set1_ps( GGML_FP16_TO_FP32(x[i + 1].d) * GGML_FP16_TO_FP32(y[i + 1].d) );
-
-        const __m128i tmp_2_3 = _mm_loadu_si128((const __m128i *)x[i + 1].qs);
-
-        __m128i bx_2 = _mm_and_si128(lowMask, tmp_2_3);
-        __m128i by_2 = _mm_loadu_si128((const __m128i *)y[i + 1].qs);
-        bx_2 = _mm_sub_epi8(bx_2, off);
-        const __m128i i32_2 = mul_sum_i8_pairs(bx_2, by_2);
-
-        __m128i bx_3 = _mm_and_si128(lowMask, _mm_srli_epi64(tmp_2_3, 4));
-        __m128i by_3 = _mm_loadu_si128((const __m128i *)(y[i + 1].qs + 16));
-        bx_3 = _mm_sub_epi8(bx_3, off);
-        const __m128i i32_3 = mul_sum_i8_pairs(bx_3, by_3);
-
-        // Convert int32_t to float
-        __m128 p0 = _mm_cvtepi32_ps(i32_0);
-        __m128 p1 = _mm_cvtepi32_ps(i32_1);
-        __m128 p2 = _mm_cvtepi32_ps(i32_2);
-        __m128 p3 = _mm_cvtepi32_ps(i32_3);
-
-        // Apply the scale
-        __m128 p0_d = _mm_mul_ps( d_0_1, p0 );
-        __m128 p1_d = _mm_mul_ps( d_0_1, p1 );
-        __m128 p2_d = _mm_mul_ps( d_2_3, p2 );
-        __m128 p3_d = _mm_mul_ps( d_2_3, p3 );
-
-        // Acummulate
-        acc_0 = _mm_add_ps(p0_d, acc_0);
-        acc_1 = _mm_add_ps(p1_d, acc_1);
-        acc_2 = _mm_add_ps(p2_d, acc_2);
-        acc_3 = _mm_add_ps(p3_d, acc_3);
-    }
-
-    *s = hsum_float_4x4(acc_0, acc_1, acc_2, acc_3);
-#elif defined(__riscv_v_intrinsic)
-    float sumf = 0.0;
-
-    size_t vl = __riscv_vsetvl_e8m1(qk/2);
-
-    for (int i = 0; i < nb; i++) {
-        // load elements
-        vuint8mf2_t tx = __riscv_vle8_v_u8mf2(x[i].qs, vl);
-
-        vint8mf2_t y0 = __riscv_vle8_v_i8mf2(y[i].qs, vl);
-        vint8mf2_t y1 = __riscv_vle8_v_i8mf2(y[i].qs+16, vl);
-
-        // mask and store lower part of x, and then upper part
-        vuint8mf2_t x_a = __riscv_vand_vx_u8mf2(tx, 0x0F, vl);
-        vuint8mf2_t x_l = __riscv_vsrl_vx_u8mf2(tx, 0x04, vl);
-
-        vint8mf2_t x_ai = __riscv_vreinterpret_v_u8mf2_i8mf2(x_a);
-        vint8mf2_t x_li = __riscv_vreinterpret_v_u8mf2_i8mf2(x_l);
-
-        // subtract offset
-        vint8mf2_t v0 = __riscv_vsub_vx_i8mf2(x_ai, 8, vl);
-        vint8mf2_t v1 = __riscv_vsub_vx_i8mf2(x_li, 8, vl);
-
-        vint16m1_t vec_mul1 = __riscv_vwmul_vv_i16m1(v0, y0, vl);
-        vint16m1_t vec_mul2 = __riscv_vwmul_vv_i16m1(v1, y1, vl);
-
-        vint32m1_t vec_zero = __riscv_vmv_v_x_i32m1(0, vl);
-
-        vint32m1_t vs1 = __riscv_vwredsum_vs_i16m1_i32m1(vec_mul1, vec_zero, vl);
-        vint32m1_t vs2 = __riscv_vwredsum_vs_i16m1_i32m1(vec_mul2, vs1, vl);
-
-        int sumi = __riscv_vmv_x_s_i32m1_i32(vs2);
-
-        sumf += sumi*GGML_FP16_TO_FP32(x[i].d)*GGML_FP16_TO_FP32(y[i].d);
-    }
-
-    *s = sumf;
-#else
-    // scalar
-    float sumf = 0.0;
-
-    for (int i = 0; i < nb; i++) {
-        int sumi = 0;
-
-        for (int j = 0; j < qk/2; ++j) {
-            const int v0 = (x[i].qs[j] & 0x0F) - 8;
-            const int v1 = (x[i].qs[j] >>   4) - 8;
-
-            sumi += (v0 * y[i].qs[j]) + (v1 * y[i].qs[j + qk/2]);
-        }
-
-        sumf += sumi*GGML_FP16_TO_FP32(x[i].d)*GGML_FP16_TO_FP32(y[i].d);
-    }
-
-    *s = sumf;
-#endif
-}
-
-void ggml_vec_dot_q4_1_q8_1(const int n, float * restrict s, const void * restrict vx, const void * restrict vy) {
-    const int qk = QK8_1;
-    const int nb = n / qk;
-
-    assert(n % qk == 0);
-
-    const block_q4_1 * restrict x = vx;
-    const block_q8_1 * restrict y = vy;
-
-    // TODO: add WASM SIMD
-#if defined(__ARM_NEON)
-    float32x4_t sumv0 = vdupq_n_f32(0.0f);
-    float32x4_t sumv1 = vdupq_n_f32(0.0f);
-
-    float summs = 0;
-
-    assert(nb % 2 == 0); // TODO: handle odd nb
-
-    for (int i = 0; i < nb; i += 2) {
-        const block_q4_1 * restrict x0 = &x[i + 0];
-        const block_q4_1 * restrict x1 = &x[i + 1];
-        const block_q8_1 * restrict y0 = &y[i + 0];
-        const block_q8_1 * restrict y1 = &y[i + 1];
-
-        summs += GGML_FP16_TO_FP32(x0->m) * y0->s + GGML_FP16_TO_FP32(x1->m) * y1->s;
-
-        const uint8x16_t m4b = vdupq_n_u8(0x0F);
-
-        const uint8x16_t v0_0 = vld1q_u8(x0->qs);
-        const uint8x16_t v0_1 = vld1q_u8(x1->qs);
-
-        // 4-bit -> 8-bit
-        const int8x16_t v0_0l = vreinterpretq_s8_u8(vandq_u8  (v0_0, m4b));
-        const int8x16_t v0_0h = vreinterpretq_s8_u8(vshrq_n_u8(v0_0, 4));
-        const int8x16_t v0_1l = vreinterpretq_s8_u8(vandq_u8  (v0_1, m4b));
-        const int8x16_t v0_1h = vreinterpretq_s8_u8(vshrq_n_u8(v0_1, 4));
-
-        // load y
-        const int8x16_t v1_0l = vld1q_s8(y0->qs);
-        const int8x16_t v1_0h = vld1q_s8(y0->qs + 16);
-        const int8x16_t v1_1l = vld1q_s8(y1->qs);
-        const int8x16_t v1_1h = vld1q_s8(y1->qs + 16);
-
-        // dot product into int32x4_t
-        const int32x4_t p_0 = ggml_vdotq_s32(ggml_vdotq_s32(vdupq_n_s32(0), v0_0l, v1_0l), v0_0h, v1_0h);
-        const int32x4_t p_1 = ggml_vdotq_s32(ggml_vdotq_s32(vdupq_n_s32(0), v0_1l, v1_1l), v0_1h, v1_1h);
-
-        sumv0 = vmlaq_n_f32(sumv0, vcvtq_f32_s32(p_0), GGML_FP16_TO_FP32(x0->d)*y0->d);
-        sumv1 = vmlaq_n_f32(sumv1, vcvtq_f32_s32(p_1), GGML_FP16_TO_FP32(x1->d)*y1->d);
-    }
-
-    *s = vaddvq_f32(sumv0) + vaddvq_f32(sumv1) + summs;
-#elif defined(__AVX2__) || defined(__AVX__)
-    // Initialize accumulator with zeros
-    __m256 acc = _mm256_setzero_ps();
-
-    float summs = 0;
-
-    // Main loop
-    for (int i = 0; i < nb; ++i) {
-        const float d0 = GGML_FP16_TO_FP32(x[i].d);
-        const float d1 = y[i].d;
-
-        summs += GGML_FP16_TO_FP32(x[i].m) * y[i].s;
-
-        const __m256 d0v = _mm256_set1_ps( d0 );
-        const __m256 d1v = _mm256_set1_ps( d1 );
-
-        // Compute combined scales
-        const __m256 d0d1 = _mm256_mul_ps( d0v, d1v );
-
-        // Load 16 bytes, and unpack 4 bit fields into bytes, making 32 bytes
-        const __m256i bx = bytes_from_nibbles_32(x[i].qs);
-        const __m256i by = _mm256_loadu_si256( (const __m256i *)y[i].qs );
-
-        const __m256 xy = mul_sum_us8_pairs_float(bx, by);
-
-        // Accumulate d0*d1*x*y
-#if defined(__AVX2__)
-        acc = _mm256_fmadd_ps( d0d1, xy, acc );
-#else
-        acc = _mm256_add_ps( _mm256_mul_ps( d0d1, xy ), acc );
-#endif
-    }
-
-    *s = hsum_float_8(acc) + summs;
-#elif defined(__riscv_v_intrinsic)
-    float sumf = 0.0;
-
-    size_t vl = __riscv_vsetvl_e8m1(qk/2);
-
-    for (int i = 0; i < nb; i++) {
-        // load elements
-        vuint8mf2_t tx = __riscv_vle8_v_u8mf2(x[i].qs, vl);
-
-        vint8mf2_t y0 = __riscv_vle8_v_i8mf2(y[i].qs, vl);
-        vint8mf2_t y1 = __riscv_vle8_v_i8mf2(y[i].qs+16, vl);
-
-        // mask and store lower part of x, and then upper part
-        vuint8mf2_t x_a = __riscv_vand_vx_u8mf2(tx, 0x0F, vl);
-        vuint8mf2_t x_l = __riscv_vsrl_vx_u8mf2(tx, 0x04, vl);
-
-        vint8mf2_t v0 = __riscv_vreinterpret_v_u8mf2_i8mf2(x_a);
-        vint8mf2_t v1 = __riscv_vreinterpret_v_u8mf2_i8mf2(x_l);
-
-        vint16m1_t vec_mul1 = __riscv_vwmul_vv_i16m1(v0, y0, vl);
-        vint16m1_t vec_mul2 = __riscv_vwmul_vv_i16m1(v1, y1, vl);
-
-        vint32m1_t vec_zero = __riscv_vmv_v_x_i32m1(0, vl);
-
-        vint32m1_t vs1 = __riscv_vwredsum_vs_i16m1_i32m1(vec_mul1, vec_zero, vl);
-        vint32m1_t vs2 = __riscv_vwredsum_vs_i16m1_i32m1(vec_mul2, vs1, vl);
-
-        int sumi = __riscv_vmv_x_s_i32m1_i32(vs2);
-
-        sumf += (GGML_FP16_TO_FP32(x[i].d)*y[i].d)*sumi + GGML_FP16_TO_FP32(x[i].m)*y[i].s;
-    }
-
-    *s = sumf;
-#else
-    // scalar
-    float sumf = 0.0;
-
-    for (int i = 0; i < nb; i++) {
-        int sumi = 0;
-
-        for (int j = 0; j < qk/2; ++j) {
-            const int v0 = (x[i].qs[j] & 0x0F);
-            const int v1 = (x[i].qs[j] >>   4);
-
-            sumi += (v0 * y[i].qs[j]) + (v1 * y[i].qs[j + qk/2]);
-        }
-
-        sumf += (GGML_FP16_TO_FP32(x[i].d)*y[i].d)*sumi + GGML_FP16_TO_FP32(x[i].m)*y[i].s;
-    }
-
-    *s = sumf;
-#endif
-}
-
-void ggml_vec_dot_q5_0_q8_0(const int n, float * restrict s, const void * restrict vx, const void * restrict vy) {
-    const int qk = QK8_0;
-    const int nb = n / qk;
-
-    assert(n % qk == 0);
-    assert(qk == QK5_0);
-
-    const block_q5_0 * restrict x = vx;
-    const block_q8_0 * restrict y = vy;
-
-#if defined(__ARM_NEON)
-    float32x4_t sumv0 = vdupq_n_f32(0.0f);
-    float32x4_t sumv1 = vdupq_n_f32(0.0f);
-
-    uint32_t qh0;
-    uint32_t qh1;
-
-    uint64_t tmp0[4];
-    uint64_t tmp1[4];
-
-    assert(nb % 2 == 0); // TODO: handle odd nb
-
-    for (int i = 0; i < nb; i += 2) {
-        const block_q5_0 * restrict x0 = &x[i];
-        const block_q5_0 * restrict x1 = &x[i + 1];
-        const block_q8_0 * restrict y0 = &y[i];
-        const block_q8_0 * restrict y1 = &y[i + 1];
-
-        const uint8x16_t m4b = vdupq_n_u8(0x0F);
-
-        // extract the 5th bit via lookup table ((!b) << 4)
-        memcpy(&qh0, x0->qh, sizeof(qh0));
-        memcpy(&qh1, x1->qh, sizeof(qh1));
-
-        tmp0[0] = table_b2b_1[(qh0 >>  0) & 0xFF];
-        tmp0[1] = table_b2b_1[(qh0 >>  8) & 0xFF];
-        tmp0[2] = table_b2b_1[(qh0 >> 16) & 0xFF];
-        tmp0[3] = table_b2b_1[(qh0 >> 24)       ];
-
-        tmp1[0] = table_b2b_1[(qh1 >>  0) & 0xFF];
-        tmp1[1] = table_b2b_1[(qh1 >>  8) & 0xFF];
-        tmp1[2] = table_b2b_1[(qh1 >> 16) & 0xFF];
-        tmp1[3] = table_b2b_1[(qh1 >> 24)       ];
-
-        const int8x16_t qhl0 = vld1q_s8((const int8_t *)(tmp0 + 0));
-        const int8x16_t qhh0 = vld1q_s8((const int8_t *)(tmp0 + 2));
-        const int8x16_t qhl1 = vld1q_s8((const int8_t *)(tmp1 + 0));
-        const int8x16_t qhh1 = vld1q_s8((const int8_t *)(tmp1 + 2));
-
-        const uint8x16_t v0_0 = vld1q_u8(x0->qs);
-        const uint8x16_t v0_1 = vld1q_u8(x1->qs);
-
-        // 4-bit -> 8-bit
-        int8x16_t v0_0l = vreinterpretq_s8_u8(vandq_u8  (v0_0, m4b));
-        int8x16_t v0_0h = vreinterpretq_s8_u8(vshrq_n_u8(v0_0, 4));
-        int8x16_t v0_1l = vreinterpretq_s8_u8(vandq_u8  (v0_1, m4b));
-        int8x16_t v0_1h = vreinterpretq_s8_u8(vshrq_n_u8(v0_1, 4));
-
-        // add high bit and sub 16 (equivalent to sub 0x10 when bit is zero)
-        const int8x16_t v0_0lf = vsubq_s8(v0_0l, qhl0);
-        const int8x16_t v0_0hf = vsubq_s8(v0_0h, qhh0);
-        const int8x16_t v0_1lf = vsubq_s8(v0_1l, qhl1);
-        const int8x16_t v0_1hf = vsubq_s8(v0_1h, qhh1);
-
-        // load y
-        const int8x16_t v1_0l = vld1q_s8(y0->qs);
-        const int8x16_t v1_0h = vld1q_s8(y0->qs + 16);
-        const int8x16_t v1_1l = vld1q_s8(y1->qs);
-        const int8x16_t v1_1h = vld1q_s8(y1->qs + 16);
-
-        sumv0 = vmlaq_n_f32(sumv0, vcvtq_f32_s32(vaddq_s32(
-                        ggml_vdotq_s32(vdupq_n_s32(0), v0_0lf, v1_0l),
-                        ggml_vdotq_s32(vdupq_n_s32(0), v0_0hf, v1_0h))), GGML_FP16_TO_FP32(x0->d)*GGML_FP16_TO_FP32(y0->d));
-        sumv1 = vmlaq_n_f32(sumv1, vcvtq_f32_s32(vaddq_s32(
-                        ggml_vdotq_s32(vdupq_n_s32(0), v0_1lf, v1_1l),
-                        ggml_vdotq_s32(vdupq_n_s32(0), v0_1hf, v1_1h))), GGML_FP16_TO_FP32(x1->d)*GGML_FP16_TO_FP32(y1->d));
-    }
-
-    *s = vaddvq_f32(sumv0) + vaddvq_f32(sumv1);
-#elif defined(__wasm_simd128__)
-    v128_t sumv = wasm_f32x4_splat(0.0f);
-
-    uint32_t qh;
-    uint64_t tmp[4];
-
-    // TODO: check if unrolling this is better
-    for (int i = 0; i < nb; ++i) {
-        const block_q5_0 * restrict x0 = &x[i];
-        const block_q8_0 * restrict y0 = &y[i];
-
-        const v128_t m4b  = wasm_i8x16_splat(0x0F);
-
-        // extract the 5th bit
-        memcpy(&qh, x0->qh, sizeof(qh));
-
-        tmp[0] = table_b2b_1[(qh >>  0) & 0xFF];
-        tmp[1] = table_b2b_1[(qh >>  8) & 0xFF];
-        tmp[2] = table_b2b_1[(qh >> 16) & 0xFF];
-        tmp[3] = table_b2b_1[(qh >> 24)       ];
-
-        const v128_t qhl = wasm_v128_load(tmp + 0);
-        const v128_t qhh = wasm_v128_load(tmp + 2);
-
-        const v128_t v0 = wasm_v128_load(x0->qs);
-
-        // 4-bit -> 8-bit
-        const v128_t v0l = wasm_v128_and (v0, m4b);
-        const v128_t v0h = wasm_u8x16_shr(v0, 4);
-
-        // add high bit and sub 16 (equivalent to sub 0x10 when bit is zero)
-        const v128_t v0lf = wasm_i8x16_sub(v0l, qhl);
-        const v128_t v0hf = wasm_i8x16_sub(v0h, qhh);
-
-        // load y
-        const v128_t v1l = wasm_v128_load(y0->qs);
-        const v128_t v1h = wasm_v128_load(y0->qs + 16);
-
-        // int8x16 -> int16x8
-        const v128_t v0lfl = wasm_i16x8_extend_low_i8x16 (v0lf);
-        const v128_t v0lfh = wasm_i16x8_extend_high_i8x16(v0lf);
-        const v128_t v0hfl = wasm_i16x8_extend_low_i8x16 (v0hf);
-        const v128_t v0hfh = wasm_i16x8_extend_high_i8x16(v0hf);
-
-        const v128_t v1ll = wasm_i16x8_extend_low_i8x16 (v1l);
-        const v128_t v1lh = wasm_i16x8_extend_high_i8x16(v1l);
-        const v128_t v1hl = wasm_i16x8_extend_low_i8x16 (v1h);
-        const v128_t v1hh = wasm_i16x8_extend_high_i8x16(v1h);
-
-        // dot product
-        sumv = wasm_f32x4_add(sumv, wasm_f32x4_mul(wasm_f32x4_convert_i32x4(
-                        wasm_i32x4_add(
-                            wasm_i32x4_add(wasm_i32x4_dot_i16x8(v0lfl, v1ll),
-                                           wasm_i32x4_dot_i16x8(v0lfh, v1lh)),
-                            wasm_i32x4_add(wasm_i32x4_dot_i16x8(v0hfl, v1hl),
-                                           wasm_i32x4_dot_i16x8(v0hfh, v1hh)))),
-                    wasm_f32x4_splat(GGML_FP16_TO_FP32(x0->d) * GGML_FP16_TO_FP32(y0->d))));
-    }
-
-    *s = wasm_f32x4_extract_lane(sumv, 0) + wasm_f32x4_extract_lane(sumv, 1) +
-         wasm_f32x4_extract_lane(sumv, 2) + wasm_f32x4_extract_lane(sumv, 3);
-#elif defined(__AVX2__)
-    // Initialize accumulator with zeros
-    __m256 acc = _mm256_setzero_ps();
-
-    // Main loop
-    for (int i = 0; i < nb; i++) {
-        /* Compute combined scale for the block */
-        const __m256 d = _mm256_set1_ps(GGML_FP16_TO_FP32(x[i].d) * GGML_FP16_TO_FP32(y[i].d));
-
-        __m256i bx = bytes_from_nibbles_32(x[i].qs);
-        __m256i bxhi = bytes_from_bits_32(x[i].qh);
-        bxhi = _mm256_andnot_si256(bxhi, _mm256_set1_epi8((char)0xF0));
-        bx = _mm256_or_si256(bx, bxhi);
-
-        __m256i by = _mm256_loadu_si256((const __m256i *)y[i].qs);
-
-        const __m256 q = mul_sum_i8_pairs_float(bx, by);
-
-        /* Multiply q with scale and accumulate */
-        acc = _mm256_fmadd_ps(d, q, acc);
-    }
-
-    *s = hsum_float_8(acc);
-#elif defined(__AVX__)
-    // Initialize accumulator with zeros
-    __m256 acc = _mm256_setzero_ps();
-    __m128i mask = _mm_set1_epi8((char)0xF0);
-
-    // Main loop
-    for (int i = 0; i < nb; i++) {
-        /* Compute combined scale for the block */
-        const __m256 d = _mm256_set1_ps(GGML_FP16_TO_FP32(x[i].d) * GGML_FP16_TO_FP32(y[i].d));
-
-        __m256i bx = bytes_from_nibbles_32(x[i].qs);
-        const __m256i bxhi = bytes_from_bits_32(x[i].qh);
-        __m128i bxhil = _mm256_castsi256_si128(bxhi);
-        __m128i bxhih = _mm256_extractf128_si256(bxhi, 1);
-        bxhil = _mm_andnot_si128(bxhil, mask);
-        bxhih = _mm_andnot_si128(bxhih, mask);
-        __m128i bxl = _mm256_castsi256_si128(bx);
-        __m128i bxh = _mm256_extractf128_si256(bx, 1);
-        bxl = _mm_or_si128(bxl, bxhil);
-        bxh = _mm_or_si128(bxh, bxhih);
-        bx = MM256_SET_M128I(bxh, bxl);
-
-        const __m256i by = _mm256_loadu_si256((const __m256i *)y[i].qs);
-
-        const __m256 q = mul_sum_i8_pairs_float(bx, by);
-
-        /* Multiply q with scale and accumulate */
-        acc = _mm256_add_ps(_mm256_mul_ps(d, q), acc);
-    }
-
-    *s = hsum_float_8(acc);
-#elif defined(__riscv_v_intrinsic)
-    float sumf = 0.0;
-
-    uint32_t qh;
-
-    size_t vl = __riscv_vsetvl_e8m1(qk/2);
-
-    // These temporary registers are for masking and shift operations
-    vuint32m2_t vt_1 = __riscv_vid_v_u32m2(vl);
-    vuint32m2_t vt_2 = __riscv_vsll_vv_u32m2(__riscv_vmv_v_x_u32m2(1, vl), vt_1, vl);
-
-    vuint32m2_t vt_3 = __riscv_vsll_vx_u32m2(vt_2, 16, vl);
-    vuint32m2_t vt_4 = __riscv_vadd_vx_u32m2(vt_1, 12, vl);
-
-    for (int i = 0; i < nb; i++) {
-        memcpy(&qh, x[i].qh, sizeof(uint32_t));
-
-        // ((qh & (1u << (j + 0 ))) >> (j + 0 )) << 4;
-        vuint32m2_t xha_0 = __riscv_vand_vx_u32m2(vt_2, qh, vl);
-        vuint32m2_t xhr_0 = __riscv_vsrl_vv_u32m2(xha_0, vt_1, vl);
-        vuint32m2_t xhl_0 = __riscv_vsll_vx_u32m2(xhr_0, 4, vl);
-
-        // ((qh & (1u << (j + 16))) >> (j + 12));
-        vuint32m2_t xha_1 = __riscv_vand_vx_u32m2(vt_3, qh, vl);
-        vuint32m2_t xhl_1 = __riscv_vsrl_vv_u32m2(xha_1, vt_4, vl);
-
-        // narrowing
-        vuint16m1_t xhc_0 = __riscv_vncvt_x_x_w_u16m1(xhl_0, vl);
-        vuint8mf2_t xh_0 = __riscv_vncvt_x_x_w_u8mf2(xhc_0, vl);
-
-        vuint16m1_t xhc_1 = __riscv_vncvt_x_x_w_u16m1(xhl_1, vl);
-        vuint8mf2_t xh_1 = __riscv_vncvt_x_x_w_u8mf2(xhc_1, vl);
-
-        // load
-        vuint8mf2_t tx = __riscv_vle8_v_u8mf2(x[i].qs, vl);
-
-        vint8mf2_t y0 = __riscv_vle8_v_i8mf2(y[i].qs, vl);
-        vint8mf2_t y1 = __riscv_vle8_v_i8mf2(y[i].qs+16, vl);
-
-        vuint8mf2_t x_at = __riscv_vand_vx_u8mf2(tx, 0x0F, vl);
-        vuint8mf2_t x_lt = __riscv_vsrl_vx_u8mf2(tx, 0x04, vl);
-
-        vuint8mf2_t x_a = __riscv_vor_vv_u8mf2(x_at, xh_0, vl);
-        vuint8mf2_t x_l = __riscv_vor_vv_u8mf2(x_lt, xh_1, vl);
-
-        vint8mf2_t x_ai = __riscv_vreinterpret_v_u8mf2_i8mf2(x_a);
-        vint8mf2_t x_li = __riscv_vreinterpret_v_u8mf2_i8mf2(x_l);
-
-        vint8mf2_t v0 = __riscv_vsub_vx_i8mf2(x_ai, 16, vl);
-        vint8mf2_t v1 = __riscv_vsub_vx_i8mf2(x_li, 16, vl);
-
-        vint16m1_t vec_mul1 = __riscv_vwmul_vv_i16m1(v0, y0, vl);
-        vint16m1_t vec_mul2 = __riscv_vwmul_vv_i16m1(v1, y1, vl);
-
-        vint32m1_t vec_zero = __riscv_vmv_v_x_i32m1(0, vl);
-
-        vint32m1_t vs1 = __riscv_vwredsum_vs_i16m1_i32m1(vec_mul1, vec_zero, vl);
-        vint32m1_t vs2 = __riscv_vwredsum_vs_i16m1_i32m1(vec_mul2, vs1, vl);
-
-        int sumi = __riscv_vmv_x_s_i32m1_i32(vs2);
-
-        sumf += (GGML_FP16_TO_FP32(x[i].d)*GGML_FP16_TO_FP32(y[i].d)) * sumi;
-    }
-
-    *s = sumf;
-#else
-    // scalar
-    float sumf = 0.0;
-
-    for (int i = 0; i < nb; i++) {
-        uint32_t qh;
-        memcpy(&qh, x[i].qh, sizeof(qh));
-
-        int sumi = 0;
-
-        for (int j = 0; j < qk/2; ++j) {
-            const uint8_t xh_0 = ((qh & (1u << (j + 0 ))) >> (j + 0 )) << 4;
-            const uint8_t xh_1 = ((qh & (1u << (j + 16))) >> (j + 12));
-
-            const int32_t x0 = ((x[i].qs[j] & 0x0F) | xh_0) - 16;
-            const int32_t x1 = ((x[i].qs[j] >>   4) | xh_1) - 16;
-
-            sumi += (x0 * y[i].qs[j]) + (x1 * y[i].qs[j + qk/2]);
-        }
-
-        sumf += (GGML_FP16_TO_FP32(x[i].d)*GGML_FP16_TO_FP32(y[i].d)) * sumi;
-    }
-
-    *s = sumf;
-#endif
-}
-
-void ggml_vec_dot_q5_1_q8_1(const int n, float * restrict s, const void * restrict vx, const void * restrict vy) {
-    const int qk = QK8_1;
-    const int nb = n / qk;
-
-    assert(n % qk == 0);
-    assert(qk == QK5_1);
-
-    const block_q5_1 * restrict x = vx;
-    const block_q8_1 * restrict y = vy;
-
-#if defined(__ARM_NEON)
-    float32x4_t sumv0 = vdupq_n_f32(0.0f);
-    float32x4_t sumv1 = vdupq_n_f32(0.0f);
-
-    float summs0 = 0.0f;
-    float summs1 = 0.0f;
-
-    uint32_t qh0;
-    uint32_t qh1;
-
-    uint64_t tmp0[4];
-    uint64_t tmp1[4];
-
-    assert(nb % 2 == 0); // TODO: handle odd nb
-
-    for (int i = 0; i < nb; i += 2) {
-        const block_q5_1 * restrict x0 = &x[i];
-        const block_q5_1 * restrict x1 = &x[i + 1];
-        const block_q8_1 * restrict y0 = &y[i];
-        const block_q8_1 * restrict y1 = &y[i + 1];
-
-        const uint8x16_t m4b = vdupq_n_u8(0x0F);
-
-        summs0 += GGML_FP16_TO_FP32(x0->m) * y0->s;
-        summs1 += GGML_FP16_TO_FP32(x1->m) * y1->s;
-
-        // extract the 5th bit via lookup table ((b) << 4)
-        memcpy(&qh0, x0->qh, sizeof(qh0));
-        memcpy(&qh1, x1->qh, sizeof(qh1));
-
-        tmp0[0] = table_b2b_0[(qh0 >>  0) & 0xFF];
-        tmp0[1] = table_b2b_0[(qh0 >>  8) & 0xFF];
-        tmp0[2] = table_b2b_0[(qh0 >> 16) & 0xFF];
-        tmp0[3] = table_b2b_0[(qh0 >> 24)       ];
-
-        tmp1[0] = table_b2b_0[(qh1 >>  0) & 0xFF];
-        tmp1[1] = table_b2b_0[(qh1 >>  8) & 0xFF];
-        tmp1[2] = table_b2b_0[(qh1 >> 16) & 0xFF];
-        tmp1[3] = table_b2b_0[(qh1 >> 24)       ];
-
-        const int8x16_t qhl0 = vld1q_s8((const int8_t *)(tmp0 + 0));
-        const int8x16_t qhh0 = vld1q_s8((const int8_t *)(tmp0 + 2));
-        const int8x16_t qhl1 = vld1q_s8((const int8_t *)(tmp1 + 0));
-        const int8x16_t qhh1 = vld1q_s8((const int8_t *)(tmp1 + 2));
-
-        const uint8x16_t v0_0 = vld1q_u8(x0->qs);
-        const uint8x16_t v0_1 = vld1q_u8(x1->qs);
-
-        // 4-bit -> 8-bit
-        const int8x16_t v0_0l = vreinterpretq_s8_u8(vandq_u8  (v0_0, m4b));
-        const int8x16_t v0_0h = vreinterpretq_s8_u8(vshrq_n_u8(v0_0, 4));
-        const int8x16_t v0_1l = vreinterpretq_s8_u8(vandq_u8  (v0_1, m4b));
-        const int8x16_t v0_1h = vreinterpretq_s8_u8(vshrq_n_u8(v0_1, 4));
-
-        // add high bit
-        const int8x16_t v0_0lf = vorrq_s8(v0_0l, qhl0);
-        const int8x16_t v0_0hf = vorrq_s8(v0_0h, qhh0);
-        const int8x16_t v0_1lf = vorrq_s8(v0_1l, qhl1);
-        const int8x16_t v0_1hf = vorrq_s8(v0_1h, qhh1);
-
-        // load y
-        const int8x16_t v1_0l = vld1q_s8(y0->qs);
-        const int8x16_t v1_0h = vld1q_s8(y0->qs + 16);
-        const int8x16_t v1_1l = vld1q_s8(y1->qs);
-        const int8x16_t v1_1h = vld1q_s8(y1->qs + 16);
-
-        sumv0 = vmlaq_n_f32(sumv0, vcvtq_f32_s32(vaddq_s32(
-                        ggml_vdotq_s32(vdupq_n_s32(0), v0_0lf, v1_0l),
-                        ggml_vdotq_s32(vdupq_n_s32(0), v0_0hf, v1_0h))), GGML_FP16_TO_FP32(x0->d)*y0->d);
-        sumv1 = vmlaq_n_f32(sumv1, vcvtq_f32_s32(vaddq_s32(
-                        ggml_vdotq_s32(vdupq_n_s32(0), v0_1lf, v1_1l),
-                        ggml_vdotq_s32(vdupq_n_s32(0), v0_1hf, v1_1h))), GGML_FP16_TO_FP32(x1->d)*y1->d);
-    }
-
-    *s = vaddvq_f32(sumv0) + vaddvq_f32(sumv1) + summs0 + summs1;
-#elif defined(__wasm_simd128__)
-    v128_t sumv = wasm_f32x4_splat(0.0f);
-
-    float summs = 0.0f;
-
-    uint32_t qh;
-    uint64_t tmp[4];
-
-    // TODO: check if unrolling this is better
-    for (int i = 0; i < nb; ++i) {
-        const block_q5_1 * restrict x0 = &x[i];
-        const block_q8_1 * restrict y0 = &y[i];
-
-        summs += GGML_FP16_TO_FP32(x0->m) * y0->s;
-
-        const v128_t m4b = wasm_i8x16_splat(0x0F);
-
-        // extract the 5th bit
-        memcpy(&qh, x0->qh, sizeof(qh));
-
-        tmp[0] = table_b2b_0[(qh >>  0) & 0xFF];
-        tmp[1] = table_b2b_0[(qh >>  8) & 0xFF];
-        tmp[2] = table_b2b_0[(qh >> 16) & 0xFF];
-        tmp[3] = table_b2b_0[(qh >> 24)       ];
-
-        const v128_t qhl = wasm_v128_load(tmp + 0);
-        const v128_t qhh = wasm_v128_load(tmp + 2);
-
-        const v128_t v0 = wasm_v128_load(x0->qs);
-
-        // 4-bit -> 8-bit
-        const v128_t v0l = wasm_v128_and (v0, m4b);
-        const v128_t v0h = wasm_u8x16_shr(v0, 4);
-
-        // add high bit
-        const v128_t v0lf = wasm_v128_or(v0l, qhl);
-        const v128_t v0hf = wasm_v128_or(v0h, qhh);
-
-        // load y
-        const v128_t v1l = wasm_v128_load(y0->qs);
-        const v128_t v1h = wasm_v128_load(y0->qs + 16);
-
-        // int8x16 -> int16x8
-        const v128_t v0lfl = wasm_i16x8_extend_low_i8x16 (v0lf);
-        const v128_t v0lfh = wasm_i16x8_extend_high_i8x16(v0lf);
-        const v128_t v0hfl = wasm_i16x8_extend_low_i8x16 (v0hf);
-        const v128_t v0hfh = wasm_i16x8_extend_high_i8x16(v0hf);
-
-        const v128_t v1ll = wasm_i16x8_extend_low_i8x16 (v1l);
-        const v128_t v1lh = wasm_i16x8_extend_high_i8x16(v1l);
-        const v128_t v1hl = wasm_i16x8_extend_low_i8x16 (v1h);
-        const v128_t v1hh = wasm_i16x8_extend_high_i8x16(v1h);
-
-        // dot product
-        sumv = wasm_f32x4_add(sumv,
-                wasm_f32x4_mul(wasm_f32x4_convert_i32x4(wasm_i32x4_add(
-                            wasm_i32x4_add(wasm_i32x4_dot_i16x8(v0lfl, v1ll),
-                                           wasm_i32x4_dot_i16x8(v0lfh, v1lh)),
-                            wasm_i32x4_add(wasm_i32x4_dot_i16x8(v0hfl, v1hl),
-                                           wasm_i32x4_dot_i16x8(v0hfh, v1hh)))),
-                    wasm_f32x4_splat(GGML_FP16_TO_FP32(x0->d) * y0->d)));
-    }
-
-    *s = wasm_f32x4_extract_lane(sumv, 0) + wasm_f32x4_extract_lane(sumv, 1) +
-         wasm_f32x4_extract_lane(sumv, 2) + wasm_f32x4_extract_lane(sumv, 3) + summs;
-#elif defined(__AVX2__)
-    // Initialize accumulator with zeros
-    __m256 acc = _mm256_setzero_ps();
-
-    float summs = 0.0f;
-
-    // Main loop
-    for (int i = 0; i < nb; i++) {
-        const __m256 dx = _mm256_set1_ps(GGML_FP16_TO_FP32(x[i].d));
-
-        summs += GGML_FP16_TO_FP32(x[i].m) * y[i].s;
-
-        __m256i bx = bytes_from_nibbles_32(x[i].qs);
-        __m256i bxhi = bytes_from_bits_32(x[i].qh);
-        bxhi = _mm256_and_si256(bxhi, _mm256_set1_epi8(0x10));
-        bx = _mm256_or_si256(bx, bxhi);
-
-        const __m256 dy = _mm256_set1_ps(y[i].d);
-        const __m256i by = _mm256_loadu_si256((const __m256i *)y[i].qs);
-
-        const __m256 q = mul_sum_us8_pairs_float(bx, by);
-
-        acc = _mm256_fmadd_ps(q, _mm256_mul_ps(dx, dy), acc);
-    }
-
-    *s = hsum_float_8(acc) + summs;
-#elif defined(__AVX__)
-    // Initialize accumulator with zeros
-    __m256 acc = _mm256_setzero_ps();
-    __m128i mask = _mm_set1_epi8(0x10);
-
-    float summs = 0.0f;
-
-    // Main loop
-    for (int i = 0; i < nb; i++) {
-        const __m256 dx = _mm256_set1_ps(GGML_FP16_TO_FP32(x[i].d));
-
-        summs += GGML_FP16_TO_FP32(x[i].m) * y[i].s;
-
-        __m256i bx = bytes_from_nibbles_32(x[i].qs);
-        const __m256i bxhi = bytes_from_bits_32(x[i].qh);
-        __m128i bxhil = _mm256_castsi256_si128(bxhi);
-        __m128i bxhih = _mm256_extractf128_si256(bxhi, 1);
-        bxhil = _mm_and_si128(bxhil, mask);
-        bxhih = _mm_and_si128(bxhih, mask);
-        __m128i bxl = _mm256_castsi256_si128(bx);
-        __m128i bxh = _mm256_extractf128_si256(bx, 1);
-        bxl = _mm_or_si128(bxl, bxhil);
-        bxh = _mm_or_si128(bxh, bxhih);
-        bx = MM256_SET_M128I(bxh, bxl);
-
-        const __m256 dy = _mm256_set1_ps(y[i].d);
-        const __m256i by = _mm256_loadu_si256((const __m256i *)y[i].qs);
-
-        const __m256 q = mul_sum_us8_pairs_float(bx, by);
-
-        acc = _mm256_add_ps(_mm256_mul_ps(q, _mm256_mul_ps(dx, dy)), acc);
-    }
-
-    *s = hsum_float_8(acc) + summs;
-#elif defined(__riscv_v_intrinsic)
-    float sumf = 0.0;
-
-    uint32_t qh;
-
-    size_t vl = __riscv_vsetvl_e8m1(qk/2);
-
-    // temporary registers for shift operations
-    vuint32m2_t vt_1 = __riscv_vid_v_u32m2(vl);
-    vuint32m2_t vt_2 = __riscv_vadd_vx_u32m2(vt_1, 12, vl);
-
-    for (int i = 0; i < nb; i++) {
-        memcpy(&qh, x[i].qh, sizeof(uint32_t));
-
-        // load qh
-        vuint32m2_t vqh = __riscv_vmv_v_x_u32m2(qh, vl);
-
-        // ((qh >> (j +  0)) << 4) & 0x10;
-        vuint32m2_t xhr_0 = __riscv_vsrl_vv_u32m2(vqh, vt_1, vl);
-        vuint32m2_t xhl_0 = __riscv_vsll_vx_u32m2(xhr_0, 4, vl);
-        vuint32m2_t xha_0 = __riscv_vand_vx_u32m2(xhl_0, 0x10, vl);
-
-        // ((qh >> (j + 12))     ) & 0x10;
-        vuint32m2_t xhr_1 = __riscv_vsrl_vv_u32m2(vqh, vt_2, vl);
-        vuint32m2_t xha_1 = __riscv_vand_vx_u32m2(xhr_1, 0x10, vl);
-
-        // narrowing
-        vuint16m1_t xhc_0 = __riscv_vncvt_x_x_w_u16m1(xha_0, vl);
-        vuint8mf2_t xh_0 = __riscv_vncvt_x_x_w_u8mf2(xhc_0, vl);
-
-        vuint16m1_t xhc_1 = __riscv_vncvt_x_x_w_u16m1(xha_1, vl);
-        vuint8mf2_t xh_1 = __riscv_vncvt_x_x_w_u8mf2(xhc_1, vl);
-
-        // load
-        vuint8mf2_t tx = __riscv_vle8_v_u8mf2(x[i].qs, vl);
-
-        vint8mf2_t y0 = __riscv_vle8_v_i8mf2(y[i].qs, vl);
-        vint8mf2_t y1 = __riscv_vle8_v_i8mf2(y[i].qs+16, vl);
-
-        vuint8mf2_t x_at = __riscv_vand_vx_u8mf2(tx, 0x0F, vl);
-        vuint8mf2_t x_lt = __riscv_vsrl_vx_u8mf2(tx, 0x04, vl);
-
-        vuint8mf2_t x_a = __riscv_vor_vv_u8mf2(x_at, xh_0, vl);
-        vuint8mf2_t x_l = __riscv_vor_vv_u8mf2(x_lt, xh_1, vl);
-
-        vint8mf2_t v0 = __riscv_vreinterpret_v_u8mf2_i8mf2(x_a);
-        vint8mf2_t v1 = __riscv_vreinterpret_v_u8mf2_i8mf2(x_l);
-
-        vint16m1_t vec_mul1 = __riscv_vwmul_vv_i16m1(v0, y0, vl);
-        vint16m1_t vec_mul2 = __riscv_vwmul_vv_i16m1(v1, y1, vl);
-
-        vint32m1_t vec_zero = __riscv_vmv_v_x_i32m1(0, vl);
-
-        vint32m1_t vs1 = __riscv_vwredsum_vs_i16m1_i32m1(vec_mul1, vec_zero, vl);
-        vint32m1_t vs2 = __riscv_vwredsum_vs_i16m1_i32m1(vec_mul2, vs1, vl);
-
-        int sumi = __riscv_vmv_x_s_i32m1_i32(vs2);
-
-        sumf += (GGML_FP16_TO_FP32(x[i].d)*y[i].d)*sumi + GGML_FP16_TO_FP32(x[i].m)*y[i].s;
-    }
-
-    *s = sumf;
-#else
-    // scalar
-    float sumf = 0.0;
-
-    for (int i = 0; i < nb; i++) {
-        uint32_t qh;
-        memcpy(&qh, x[i].qh, sizeof(qh));
-
-        int sumi = 0;
-
-        for (int j = 0; j < qk/2; ++j) {
-            const uint8_t xh_0 = ((qh >> (j +  0)) << 4) & 0x10;
-            const uint8_t xh_1 = ((qh >> (j + 12))     ) & 0x10;
-
-            const int32_t x0 = (x[i].qs[j] & 0xF) | xh_0;
-            const int32_t x1 = (x[i].qs[j] >>  4) | xh_1;
-
-            sumi += (x0 * y[i].qs[j]) + (x1 * y[i].qs[j + qk/2]);
-        }
-
-        sumf += (GGML_FP16_TO_FP32(x[i].d)*y[i].d)*sumi + GGML_FP16_TO_FP32(x[i].m)*y[i].s;
-    }
-
-    *s = sumf;
-#endif
-}
-
-void ggml_vec_dot_q8_0_q8_0(const int n, float * restrict s, const void * restrict vx, const void * restrict vy) {
-    const int qk = QK8_0;
-    const int nb = n / qk;
-
-    assert(n % qk == 0);
-
-    const block_q8_0 * restrict x = vx;
-    const block_q8_0 * restrict y = vy;
-
-#if defined(__ARM_NEON)
-    float32x4_t sumv0 = vdupq_n_f32(0.0f);
-    float32x4_t sumv1 = vdupq_n_f32(0.0f);
-
-    assert(nb % 2 == 0); // TODO: handle odd nb
-
-    for (int i = 0; i < nb; i += 2) {
-        const block_q8_0 * restrict x0 = &x[i + 0];
-        const block_q8_0 * restrict x1 = &x[i + 1];
-        const block_q8_0 * restrict y0 = &y[i + 0];
-        const block_q8_0 * restrict y1 = &y[i + 1];
-
-        const int8x16_t x0_0 = vld1q_s8(x0->qs);
-        const int8x16_t x0_1 = vld1q_s8(x0->qs + 16);
-        const int8x16_t x1_0 = vld1q_s8(x1->qs);
-        const int8x16_t x1_1 = vld1q_s8(x1->qs + 16);
-
-        // load y
-        const int8x16_t y0_0 = vld1q_s8(y0->qs);
-        const int8x16_t y0_1 = vld1q_s8(y0->qs + 16);
-        const int8x16_t y1_0 = vld1q_s8(y1->qs);
-        const int8x16_t y1_1 = vld1q_s8(y1->qs + 16);
-
-        sumv0 = vmlaq_n_f32(sumv0, vcvtq_f32_s32(vaddq_s32(
-                        ggml_vdotq_s32(vdupq_n_s32(0), x0_0, y0_0),
-                        ggml_vdotq_s32(vdupq_n_s32(0), x0_1, y0_1))), GGML_FP16_TO_FP32(x0->d)*GGML_FP16_TO_FP32(y0->d));
-
-        sumv1 = vmlaq_n_f32(sumv1, vcvtq_f32_s32(vaddq_s32(
-                        ggml_vdotq_s32(vdupq_n_s32(0), x1_0, y1_0),
-                        ggml_vdotq_s32(vdupq_n_s32(0), x1_1, y1_1))), GGML_FP16_TO_FP32(x1->d)*GGML_FP16_TO_FP32(y1->d));
-    }
-
-    *s = vaddvq_f32(sumv0) + vaddvq_f32(sumv1);
-#elif defined(__AVX2__) || defined(__AVX__)
-    // Initialize accumulator with zeros
-    __m256 acc = _mm256_setzero_ps();
-
-    // Main loop
-    for (int i = 0; i < nb; ++i) {
-        // Compute combined scale for the block
-        const __m256 d = _mm256_set1_ps(GGML_FP16_TO_FP32(x[i].d) * GGML_FP16_TO_FP32(y[i].d));
-        __m256i bx = _mm256_loadu_si256((const __m256i *)x[i].qs);
-        __m256i by = _mm256_loadu_si256((const __m256i *)y[i].qs);
-
-        const __m256 q = mul_sum_i8_pairs_float(bx, by);
-
-        // Multiply q with scale and accumulate
-#if defined(__AVX2__)
-        acc = _mm256_fmadd_ps( d, q, acc );
-#else
-        acc = _mm256_add_ps( _mm256_mul_ps( d, q ), acc );
-#endif
-    }
-
-    *s = hsum_float_8(acc);
-#elif defined(__riscv_v_intrinsic)
-    float sumf = 0.0;
-    size_t vl = __riscv_vsetvl_e8m1(qk);
-
-    for (int i = 0; i < nb; i++) {
-        // load elements
-        vint8m1_t bx = __riscv_vle8_v_i8m1(x[i].qs, vl);
-        vint8m1_t by = __riscv_vle8_v_i8m1(y[i].qs, vl);
-
-        vint16m2_t vw_mul = __riscv_vwmul_vv_i16m2(bx, by, vl);
-
-        vint32m1_t v_zero = __riscv_vmv_v_x_i32m1(0, vl);
-        vint32m1_t v_sum = __riscv_vwredsum_vs_i16m2_i32m1(vw_mul, v_zero, vl);
-
-        int sumi = __riscv_vmv_x_s_i32m1_i32(v_sum);
-
-        sumf += sumi*(GGML_FP16_TO_FP32(x[i].d)*GGML_FP16_TO_FP32(y[i].d));
-    }
-
-    *s = sumf;
-#else
-    // scalar
-    float sumf = 0.0;
-
-    for (int i = 0; i < nb; i++) {
-        int sumi = 0;
-
-        for (int j = 0; j < qk; j++) {
-            sumi += x[i].qs[j]*y[i].qs[j];
-        }
-
-        sumf += sumi*(GGML_FP16_TO_FP32(x[i].d)*GGML_FP16_TO_FP32(y[i].d));
-    }
-
-    *s = sumf;
-#endif
-}
-
-#if QK_K == 256
-void ggml_vec_dot_q2_K_q8_K(const int n, float * restrict s, const void * restrict vx, const void * restrict vy) {
-
-    const block_q2_K * restrict x = vx;
-    const block_q8_K * restrict y = vy;
-
-    const int nb = n / QK_K;
-
-#ifdef __ARM_NEON
-    const uint8x16_t m3 = vdupq_n_u8(0x3);
-    const uint8x16_t m4 = vdupq_n_u8(0xF);
-
-    const int32x4_t vzero = vdupq_n_s32(0);
-
-    ggml_int8x16x2_t q2bytes;
-    uint8_t aux[16];
-
-    float sum = 0;
-
-    for (int i = 0; i < nb; ++i) {
-        const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d);
-        const float dmin = -y[i].d * GGML_FP16_TO_FP32(x[i].dmin);
-
-        const uint8_t * restrict q2 = x[i].qs;
-        const int8_t  * restrict q8 = y[i].qs;
-        const uint8_t * restrict sc = x[i].scales;
-
-        const uint8x16_t mins_and_scales = vld1q_u8(sc);
-        const uint8x16_t scales = vandq_u8(mins_and_scales, m4);
-        vst1q_u8(aux, scales);
-
-        const uint8x16_t mins = vshrq_n_u8(mins_and_scales, 4);
-        const ggml_int16x8x2_t q8sums = ggml_vld1q_s16_x2(y[i].bsums);
-        const ggml_int16x8x2_t mins16 = {{vreinterpretq_s16_u16(vmovl_u8(vget_low_u8(mins))), vreinterpretq_s16_u16(vmovl_u8(vget_high_u8(mins)))}};
-        const int32x4_t s0 = vaddq_s32(vmull_s16(vget_low_s16 (mins16.val[0]), vget_low_s16 (q8sums.val[0])),
-                                       vmull_s16(vget_high_s16(mins16.val[0]), vget_high_s16(q8sums.val[0])));
-        const int32x4_t s1 = vaddq_s32(vmull_s16(vget_low_s16 (mins16.val[1]), vget_low_s16 (q8sums.val[1])),
-                                       vmull_s16(vget_high_s16(mins16.val[1]), vget_high_s16(q8sums.val[1])));
-        sum += dmin * vaddvq_s32(vaddq_s32(s0, s1));
-
-        int isum = 0;
-        int is = 0;
-
-// We use this macro instead of a function call because for some reason
-// the code runs 2-3% slower, even if the function is declared inline
-#define MULTIPLY_ACCUM_WITH_SCALE(index)\
-        isum += vaddvq_s32(ggml_vdotq_s32(vzero, q2bytes.val[0], q8bytes.val[0])) * aux[is+(index)];\
-        isum += vaddvq_s32(ggml_vdotq_s32(vzero, q2bytes.val[1], q8bytes.val[1])) * aux[is+1+(index)];
-
-#define SHIFT_MULTIPLY_ACCUM_WITH_SCALE(shift, index)\
-        q8bytes = ggml_vld1q_s8_x2(q8); q8 += 32;\
-        q2bytes.val[0] = vreinterpretq_s8_u8(vandq_u8(vshrq_n_u8(q2bits.val[0], (shift)), m3));\
-        q2bytes.val[1] = vreinterpretq_s8_u8(vandq_u8(vshrq_n_u8(q2bits.val[1], (shift)), m3));\
-        MULTIPLY_ACCUM_WITH_SCALE((index));
-
-        for (int j = 0; j < QK_K/128; ++j) {
-            const ggml_uint8x16x2_t q2bits = ggml_vld1q_u8_x2(q2); q2 += 32;
-
-            ggml_int8x16x2_t q8bytes = ggml_vld1q_s8_x2(q8); q8 += 32;
-            q2bytes.val[0] = vreinterpretq_s8_u8(vandq_u8(q2bits.val[0], m3));
-            q2bytes.val[1] = vreinterpretq_s8_u8(vandq_u8(q2bits.val[1], m3));
-
-            MULTIPLY_ACCUM_WITH_SCALE(0);
-
-            SHIFT_MULTIPLY_ACCUM_WITH_SCALE(2, 2);
-            SHIFT_MULTIPLY_ACCUM_WITH_SCALE(4, 4);
-            SHIFT_MULTIPLY_ACCUM_WITH_SCALE(6, 6);
-
-            is += 8;
-        }
-
-        sum += d * isum;
-    }
-
-    *s = sum;
-
-#elif defined __AVX2__
-
-    const __m256i m3 = _mm256_set1_epi8(3);
-    const __m128i m4 = _mm_set1_epi8(0xF);
-
-    __m256 acc = _mm256_setzero_ps();
-
-    for (int i = 0; i < nb; ++i) {
-
-        const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d);
-        const float dmin = -y[i].d * GGML_FP16_TO_FP32(x[i].dmin);
-
-        const uint8_t * restrict q2 = x[i].qs;
-        const int8_t  * restrict q8 = y[i].qs;
-
-        const __m128i mins_and_scales = _mm_loadu_si128((const __m128i*)x[i].scales);
-        const __m128i scales8 = _mm_and_si128(mins_and_scales, m4);
-        const __m128i mins8 = _mm_and_si128(_mm_srli_epi16(mins_and_scales, 4), m4);
-        const __m256i mins = _mm256_cvtepi8_epi16(mins8);
-        const __m256i prod = _mm256_madd_epi16(mins, _mm256_loadu_si256((const __m256i*)y[i].bsums));
-
-        acc = _mm256_fmadd_ps(_mm256_broadcast_ss(&dmin), _mm256_cvtepi32_ps(prod), acc);
-
-        const __m256i all_scales = _mm256_cvtepi8_epi16(scales8);
-        const __m128i l_scales = _mm256_extracti128_si256(all_scales, 0);
-        const __m128i h_scales = _mm256_extracti128_si256(all_scales, 1);
-        const __m256i scales[2] = {MM256_SET_M128I(l_scales, l_scales), MM256_SET_M128I(h_scales, h_scales)};
-
-        __m256i sumi = _mm256_setzero_si256();
-
-        for (int j = 0; j < QK_K/128; ++j) {
-
-            const __m256i q2bits = _mm256_loadu_si256((const __m256i*)q2); q2 += 32;
-
-            const __m256i q8_0 = _mm256_loadu_si256((const __m256i*)q8); q8 += 32;
-            const __m256i q8_1 = _mm256_loadu_si256((const __m256i*)q8); q8 += 32;
-            const __m256i q8_2 = _mm256_loadu_si256((const __m256i*)q8); q8 += 32;
-            const __m256i q8_3 = _mm256_loadu_si256((const __m256i*)q8); q8 += 32;
-
-            const __m256i q2_0 = _mm256_and_si256(q2bits, m3);
-            const __m256i q2_1 = _mm256_and_si256(_mm256_srli_epi16(q2bits, 2), m3);
-            const __m256i q2_2 = _mm256_and_si256(_mm256_srli_epi16(q2bits, 4), m3);
-            const __m256i q2_3 = _mm256_and_si256(_mm256_srli_epi16(q2bits, 6), m3);
-
-            __m256i p0 = _mm256_maddubs_epi16(q2_0, q8_0);
-            __m256i p1 = _mm256_maddubs_epi16(q2_1, q8_1);
-            __m256i p2 = _mm256_maddubs_epi16(q2_2, q8_2);
-            __m256i p3 = _mm256_maddubs_epi16(q2_3, q8_3);
-
-            p0 = _mm256_madd_epi16(_mm256_shuffle_epi8(scales[j], get_scale_shuffle_q3k(0)), p0);
-            p1 = _mm256_madd_epi16(_mm256_shuffle_epi8(scales[j], get_scale_shuffle_q3k(1)), p1);
-            p2 = _mm256_madd_epi16(_mm256_shuffle_epi8(scales[j], get_scale_shuffle_q3k(2)), p2);
-            p3 = _mm256_madd_epi16(_mm256_shuffle_epi8(scales[j], get_scale_shuffle_q3k(3)), p3);
-
-            p0 = _mm256_add_epi32(p0, p1);
-            p2 = _mm256_add_epi32(p2, p3);
-
-            sumi = _mm256_add_epi32(sumi, _mm256_add_epi32(p0, p2));
-        }
-
-        acc = _mm256_fmadd_ps(_mm256_broadcast_ss(&d), _mm256_cvtepi32_ps(sumi), acc);
-
-    }
-
-    *s = hsum_float_8(acc);
-
-#elif defined __AVX__
-
-    const __m128i m3 = _mm_set1_epi8(0x3);
-    const __m128i m4 = _mm_set1_epi8(0xF);
-    const __m128i m2 = _mm_set1_epi8(0x2);
-
-    __m256 acc = _mm256_setzero_ps();
-
-    for (int i = 0; i < nb; ++i) {
-
-        const float dall = y[i].d * GGML_FP16_TO_FP32(x[i].d);
-        const float dmin = -y[i].d * GGML_FP16_TO_FP32(x[i].dmin);
-
-        const uint8_t * restrict q2 = x[i].qs;
-        const int8_t  * restrict q8 = y[i].qs;
-
-        // load mins and scales from block_q2_K.scales[QK_K/16]
-        const __m128i mins_and_scales = _mm_loadu_si128((const __m128i*)x[i].scales);
-        const __m128i scales16 = _mm_and_si128(mins_and_scales, m4);
-        const __m128i mins16 = _mm_and_si128(_mm_srli_epi16(mins_and_scales, 4), m4);
-        const __m128i mins_0 = _mm_cvtepi8_epi16(mins16);
-        const __m128i mins_1 = _mm_cvtepi8_epi16(_mm_unpackhi_epi64(mins16, mins16));
-
-        // summs = y[i].bsums * (x[i].scales >> 4) in 16bits*8*2 to 32bits*4*2
-        const __m128i summs_0 = _mm_madd_epi16(mins_0, _mm_loadu_si128((const __m128i*)&y[i].bsums[0]));
-        const __m128i summs_1 = _mm_madd_epi16(mins_1, _mm_loadu_si128((const __m128i*)&y[i].bsums[8]));
-
-        // sumf += -dmin * summs in 32bits*8
-        acc = _mm256_add_ps(_mm256_mul_ps(_mm256_broadcast_ss(&dmin), _mm256_cvtepi32_ps(MM256_SET_M128I(summs_1, summs_0))), acc);
-
-        const __m128i scales_0 = _mm_cvtepi8_epi16(scales16);
-        const __m128i scales_1 = _mm_cvtepi8_epi16(_mm_unpackhi_epi64(scales16, scales16));
-        const __m128i scales[2] = { scales_0, scales_1 };
-
-        __m128i sumi_0 = _mm_setzero_si128();
-        __m128i sumi_1 = _mm_setzero_si128();
-
-        for (int j = 0; j < QK_K/128; ++j) {
-
-            // load Q8 quants int8*16*8 from block_q8_K.qs[QK_K]
-            const __m128i q8_0 = _mm_loadu_si128((const __m128i*)q8); q8 += 16;
-            const __m128i q8_1 = _mm_loadu_si128((const __m128i*)q8); q8 += 16;
-            const __m128i q8_2 = _mm_loadu_si128((const __m128i*)q8); q8 += 16;
-            const __m128i q8_3 = _mm_loadu_si128((const __m128i*)q8); q8 += 16;
-            const __m128i q8_4 = _mm_loadu_si128((const __m128i*)q8); q8 += 16;
-            const __m128i q8_5 = _mm_loadu_si128((const __m128i*)q8); q8 += 16;
-            const __m128i q8_6 = _mm_loadu_si128((const __m128i*)q8); q8 += 16;
-            const __m128i q8_7 = _mm_loadu_si128((const __m128i*)q8); q8 += 16;
-
-            // load 2bits*16*8 from block_q2_K.qs[QK_K/4]
-            __m128i q2bits = _mm_loadu_si128((const __m128i*)q2); q2 += 16;
-            const __m128i q2_0 = _mm_and_si128(q2bits, m3);
-            const __m128i q2_2 = _mm_and_si128(_mm_srli_epi16(q2bits, 2), m3);
-            const __m128i q2_4 = _mm_and_si128(_mm_srli_epi16(q2bits, 4), m3);
-            const __m128i q2_6 = _mm_and_si128(_mm_srli_epi16(q2bits, 6), m3);
-            q2bits = _mm_loadu_si128((const __m128i*)q2); q2 += 16;
-            const __m128i q2_1 = _mm_and_si128(q2bits, m3);
-            const __m128i q2_3 = _mm_and_si128(_mm_srli_epi16(q2bits, 2), m3);
-            const __m128i q2_5 = _mm_and_si128(_mm_srli_epi16(q2bits, 4), m3);
-            const __m128i q2_7 = _mm_and_si128(_mm_srli_epi16(q2bits, 6), m3);
-
-            // isuml = q8[l] * ((q2[l] >> shift) & 3) in 8bits*16*8 to 16bits*8*8
-            __m128i p0 = _mm_maddubs_epi16(q2_0, q8_0);
-            __m128i p1 = _mm_maddubs_epi16(q2_1, q8_1);
-            __m128i p2 = _mm_maddubs_epi16(q2_2, q8_2);
-            __m128i p3 = _mm_maddubs_epi16(q2_3, q8_3);
-            __m128i p4 = _mm_maddubs_epi16(q2_4, q8_4);
-            __m128i p5 = _mm_maddubs_epi16(q2_5, q8_5);
-            __m128i p6 = _mm_maddubs_epi16(q2_6, q8_6);
-            __m128i p7 = _mm_maddubs_epi16(q2_7, q8_7);
-
-            // isum += (x[i].scales[is++] & 0xF) * isuml in 16bits*8*8 to 32bits*4*8
-            __m128i shuffle = _mm_set1_epi16(0x0100);
-            p0 = _mm_madd_epi16(_mm_shuffle_epi8(scales[j], shuffle), p0);
-            shuffle = _mm_add_epi16(shuffle, m2);
-            p1 = _mm_madd_epi16(_mm_shuffle_epi8(scales[j], shuffle), p1);
-            shuffle = _mm_add_epi16(shuffle, m2);
-            p2 = _mm_madd_epi16(_mm_shuffle_epi8(scales[j], shuffle), p2);
-            shuffle = _mm_add_epi16(shuffle, m2);
-            p3 = _mm_madd_epi16(_mm_shuffle_epi8(scales[j], shuffle), p3);
-            shuffle = _mm_add_epi16(shuffle, m2);
-            p4 = _mm_madd_epi16(_mm_shuffle_epi8(scales[j], shuffle), p4);
-            shuffle = _mm_add_epi16(shuffle, m2);
-            p5 = _mm_madd_epi16(_mm_shuffle_epi8(scales[j], shuffle), p5);
-            shuffle = _mm_add_epi16(shuffle, m2);
-            p6 = _mm_madd_epi16(_mm_shuffle_epi8(scales[j], shuffle), p6);
-            shuffle = _mm_add_epi16(shuffle, m2);
-            p7 = _mm_madd_epi16(_mm_shuffle_epi8(scales[j], shuffle), p7);
-
-            p0 = _mm_add_epi32(p0, p1);
-            p2 = _mm_add_epi32(p2, p3);
-            p4 = _mm_add_epi32(p4, p5);
-            p6 = _mm_add_epi32(p6, p7);
-
-            // isum in 32bits*4*2
-            sumi_0 = _mm_add_epi32(sumi_0, _mm_add_epi32(p0, p2));
-            sumi_1 = _mm_add_epi32(sumi_1, _mm_add_epi32(p4, p6));
-        }
-
-        // sumf += dall * isum - dmin * summs in 32bits
-        __m256i sumi = MM256_SET_M128I(sumi_1, sumi_0);
-        acc = _mm256_add_ps(_mm256_mul_ps(_mm256_broadcast_ss(&dall), _mm256_cvtepi32_ps(sumi)), acc);
-    }
-
-    *s = hsum_float_8(acc);
-
-#elif defined __riscv_v_intrinsic
-
-    float sumf = 0;
-    uint8_t temp_01[32] = {0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
-                            1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1};
-
-    for (int i = 0; i < nb; ++i) {
-
-        const uint8_t * q2 = x[i].qs;
-        const  int8_t * q8 = y[i].qs;
-        const uint8_t * sc = x[i].scales;
-
-        const float dall = y[i].d * GGML_FP16_TO_FP32(x[i].d);
-        const float dmin = -y[i].d * GGML_FP16_TO_FP32(x[i].dmin);
-
-        size_t vl = 16;
-
-        vuint8m1_t scales = __riscv_vle8_v_u8m1(sc, vl);
-        vuint8m1_t aux = __riscv_vand_vx_u8m1(scales, 0x0F, vl);
-
-        vint16m1_t q8sums = __riscv_vle16_v_i16m1(y[i].bsums, vl);
-
-        vuint8mf2_t scales_2 = __riscv_vle8_v_u8mf2(sc, vl);
-        vuint8mf2_t mins8 = __riscv_vsrl_vx_u8mf2(scales_2, 0x4, vl);
-        vint16m1_t mins = __riscv_vreinterpret_v_u16m1_i16m1(__riscv_vzext_vf2_u16m1(mins8, vl));
-        vint32m2_t prod = __riscv_vwmul_vv_i32m2(q8sums, mins, vl);
-        vint32m1_t vsums = __riscv_vredsum_vs_i32m2_i32m1(prod, __riscv_vmv_v_x_i32m1(0, 1), vl);
-
-        sumf  += dmin * __riscv_vmv_x_s_i32m1_i32(vsums);
-
-        vl = 32;
-
-        vint32m1_t vzero = __riscv_vmv_v_x_i32m1(0, 1);
-        vuint8m1_t v_b = __riscv_vle8_v_u8m1(temp_01, vl);
-
-        uint8_t is=0;
-        int isum=0;
-
-        for (int j = 0; j < QK_K/128; ++j) {
-            // load Q2
-            vuint8m1_t q2_x = __riscv_vle8_v_u8m1(q2, vl);
-
-            vuint8m1_t q2_0 = __riscv_vand_vx_u8m1(q2_x, 0x03, vl);
-            vuint8m1_t q2_1 = __riscv_vand_vx_u8m1(__riscv_vsrl_vx_u8m1(q2_x, 0x2, vl), 0x03 , vl);
-            vuint8m1_t q2_2 = __riscv_vand_vx_u8m1(__riscv_vsrl_vx_u8m1(q2_x, 0x4, vl), 0x03 , vl);
-            vuint8m1_t q2_3 = __riscv_vand_vx_u8m1(__riscv_vsrl_vx_u8m1(q2_x, 0x6, vl), 0x03 , vl);
-
-            // duplicate scale elements for product
-            vuint8m1_t sc0 = __riscv_vrgather_vv_u8m1(aux, __riscv_vadd_vx_u8m1(v_b, 0+is, vl), vl);
-            vuint8m1_t sc1 = __riscv_vrgather_vv_u8m1(aux, __riscv_vadd_vx_u8m1(v_b, 2+is, vl), vl);
-            vuint8m1_t sc2 = __riscv_vrgather_vv_u8m1(aux, __riscv_vadd_vx_u8m1(v_b, 4+is, vl), vl);
-            vuint8m1_t sc3 = __riscv_vrgather_vv_u8m1(aux, __riscv_vadd_vx_u8m1(v_b, 6+is, vl), vl);
-
-            vint16m2_t p0 = __riscv_vreinterpret_v_u16m2_i16m2(__riscv_vwmulu_vv_u16m2(q2_0, sc0, vl));
-            vint16m2_t p1 = __riscv_vreinterpret_v_u16m2_i16m2(__riscv_vwmulu_vv_u16m2(q2_1, sc1, vl));
-            vint16m2_t p2 = __riscv_vreinterpret_v_u16m2_i16m2(__riscv_vwmulu_vv_u16m2(q2_2, sc2, vl));
-            vint16m2_t p3 = __riscv_vreinterpret_v_u16m2_i16m2(__riscv_vwmulu_vv_u16m2(q2_3, sc3, vl));
-
-            // load Q8
-            vint8m1_t q8_0 = __riscv_vle8_v_i8m1(q8, vl);
-            vint8m1_t q8_1 = __riscv_vle8_v_i8m1(q8+32, vl);
-            vint8m1_t q8_2 = __riscv_vle8_v_i8m1(q8+64, vl);
-            vint8m1_t q8_3 = __riscv_vle8_v_i8m1(q8+96, vl);
-
-            vint32m4_t s0 = __riscv_vwmul_vv_i32m4(p0, __riscv_vwcvt_x_x_v_i16m2(q8_0, vl), vl);
-            vint32m4_t s1 = __riscv_vwmul_vv_i32m4(p1, __riscv_vwcvt_x_x_v_i16m2(q8_1, vl), vl);
-            vint32m4_t s2 = __riscv_vwmul_vv_i32m4(p2, __riscv_vwcvt_x_x_v_i16m2(q8_2, vl), vl);
-            vint32m4_t s3 = __riscv_vwmul_vv_i32m4(p3, __riscv_vwcvt_x_x_v_i16m2(q8_3, vl), vl);
-
-            vint32m1_t isum0 = __riscv_vredsum_vs_i32m4_i32m1(__riscv_vadd_vv_i32m4(s0, s1, vl), vzero, vl);
-            vint32m1_t isum1 = __riscv_vredsum_vs_i32m4_i32m1(__riscv_vadd_vv_i32m4(s2, s3, vl), isum0, vl);
-
-            isum += __riscv_vmv_x_s_i32m1_i32(isum1);
-
-            q2+=32;  q8+=128;  is=8;
-
-        }
-
-        sumf += dall * isum;
-
-    }
-
-    *s = sumf;
-
-#else
-
-    float sumf = 0;
-
-    for (int i = 0; i < nb; ++i) {
-
-        const uint8_t * q2 = x[i].qs;
-        const  int8_t * q8 = y[i].qs;
-        const uint8_t * sc = x[i].scales;
-
-        int summs = 0;
-        for (int j = 0; j < 16; ++j) {
-            summs += y[i].bsums[j] * (sc[j] >> 4);
-        }
-
-        const float dall = y[i].d * GGML_FP16_TO_FP32(x[i].d);
-        const float dmin = y[i].d * GGML_FP16_TO_FP32(x[i].dmin);
-
-        int isum = 0;
-        int is = 0;
-        int d;
-        for (int k = 0; k < QK_K/128; ++k) {
-            int shift = 0;
-            for (int j = 0; j < 4; ++j) {
-                d = sc[is++] & 0xF;
-                int isuml = 0;
-                for (int l =  0; l < 16; ++l) isuml += q8[l] * ((q2[l] >> shift) & 3);
-                isum += d * isuml;
-                d = sc[is++] & 0xF;
-                isuml = 0;
-                for (int l = 16; l < 32; ++l) isuml += q8[l] * ((q2[l] >> shift) & 3);
-                isum += d * isuml;
-                shift += 2;
-                q8 += 32;
-            }
-            q2 += 32;
-        }
-        sumf += dall * isum - dmin * summs;
-    }
-    *s = sumf;
-#endif
-}
-
-#else
-
-void ggml_vec_dot_q2_K_q8_K(const int n, float * restrict s, const void * restrict vx, const void * restrict vy) {
-
-    const block_q2_K * restrict x = vx;
-    const block_q8_K * restrict y = vy;
-
-    const int nb = n / QK_K;
-
-#ifdef __ARM_NEON
-    const uint8x16_t m3 = vdupq_n_u8(0x3);
-
-    const int32x4_t vzero = vdupq_n_s32(0);
-
-    ggml_int8x16x4_t q2bytes;
-
-    uint32_t aux32[2];
-    const uint8_t * scales = (const uint8_t *)aux32;
-
-    float sum = 0;
-
-    for (int i = 0; i < nb; ++i) {
-
-        const float d = y[i].d * (float)x[i].d;
-        const float dmin = -y[i].d * (float)x[i].dmin;
-
-        const uint8_t * restrict q2 = x[i].qs;
-        const int8_t  * restrict q8 = y[i].qs;
-        const uint32_t * restrict sc = (const uint32_t *)x[i].scales;
-
-        aux32[0] = sc[0] & 0x0f0f0f0f;
-        aux32[1] = (sc[0] >> 4) & 0x0f0f0f0f;
-
-        sum += dmin * (scales[4] * y[i].bsums[0] + scales[5] * y[i].bsums[1] + scales[6] * y[i].bsums[2] + scales[7] * y[i].bsums[3]);
-
-        int isum1 = 0, isum2 = 0;
-
-        const uint8x16_t q2bits = vld1q_u8(q2);
-
-        const ggml_int8x16x4_t q8bytes = ggml_vld1q_s8_x4(q8);
-
-        q2bytes.val[0] = vreinterpretq_s8_u8(vandq_u8(q2bits, m3));
-        q2bytes.val[1] = vreinterpretq_s8_u8(vandq_u8(vshrq_n_u8(q2bits, 2), m3));
-        q2bytes.val[2] = vreinterpretq_s8_u8(vandq_u8(vshrq_n_u8(q2bits, 4), m3));
-        q2bytes.val[3] = vreinterpretq_s8_u8(vandq_u8(vshrq_n_u8(q2bits, 6), m3));
-
-        isum1 += vaddvq_s32(ggml_vdotq_s32(vzero, q2bytes.val[0], q8bytes.val[0])) * scales[0];
-        isum2 += vaddvq_s32(ggml_vdotq_s32(vzero, q2bytes.val[1], q8bytes.val[1])) * scales[1];
-        isum1 += vaddvq_s32(ggml_vdotq_s32(vzero, q2bytes.val[2], q8bytes.val[2])) * scales[2];
-        isum2 += vaddvq_s32(ggml_vdotq_s32(vzero, q2bytes.val[3], q8bytes.val[3])) * scales[3];
-
-        sum += d * (isum1 + isum2);
-    }
-
-    *s = sum;
-
-#elif defined __AVX2__
-
-    const __m256i m3 = _mm256_set1_epi8(3);
-
-    __m256 acc = _mm256_setzero_ps();
-
-    uint32_t ud, um;
-    const uint8_t * restrict db = (const uint8_t *)&ud;
-    const uint8_t * restrict mb = (const uint8_t *)&um;
-
-    float summs = 0;
-
-    // TODO: optimize this
-
-    for (int i = 0; i < nb; ++i) {
-
-        const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d);
-        const float dmin = -y[i].d * GGML_FP16_TO_FP32(x[i].dmin);
-
-        const uint8_t * restrict q2 = x[i].qs;
-        const int8_t  * restrict q8 = y[i].qs;
-
-        const uint32_t * restrict sc = (const uint32_t *)x[i].scales;
-        ud = (sc[0] >> 0) & 0x0f0f0f0f;
-        um = (sc[0] >> 4) & 0x0f0f0f0f;
-
-        int32_t smin = mb[0] * y[i].bsums[0] + mb[1] * y[i].bsums[1] + mb[2] * y[i].bsums[2] + mb[3] * y[i].bsums[3];
-        summs += dmin * smin;
-
-        const __m128i q2bits = _mm_loadu_si128((const __m128i*)q2);
-        const __m256i q2_0 = _mm256_and_si256(MM256_SET_M128I(_mm_srli_epi16(q2bits, 2), q2bits), m3);
-        const __m256i q2_1 = _mm256_and_si256(MM256_SET_M128I(_mm_srli_epi16(q2bits, 6), _mm_srli_epi16(q2bits, 4)), m3);
-
-        const __m256i q8_0 = _mm256_loadu_si256((const __m256i*)(q8+ 0));
-        const __m256i q8_1 = _mm256_loadu_si256((const __m256i*)(q8+32));
-
-        const __m256i p0 = _mm256_maddubs_epi16(q2_0, q8_0);
-        const __m256i p1 = _mm256_maddubs_epi16(q2_1, q8_1);
-
-        const __m256i p_0 = _mm256_cvtepi16_epi32(_mm256_extracti128_si256(p0, 0));
-        const __m256i p_1 = _mm256_cvtepi16_epi32(_mm256_extracti128_si256(p0, 1));
-        const __m256i p_2 = _mm256_cvtepi16_epi32(_mm256_extracti128_si256(p1, 0));
-        const __m256i p_3 = _mm256_cvtepi16_epi32(_mm256_extracti128_si256(p1, 1));
-
-        acc = _mm256_fmadd_ps(_mm256_set1_ps(d * db[0]), _mm256_cvtepi32_ps(p_0), acc);
-        acc = _mm256_fmadd_ps(_mm256_set1_ps(d * db[1]), _mm256_cvtepi32_ps(p_1), acc);
-        acc = _mm256_fmadd_ps(_mm256_set1_ps(d * db[2]), _mm256_cvtepi32_ps(p_2), acc);
-        acc = _mm256_fmadd_ps(_mm256_set1_ps(d * db[3]), _mm256_cvtepi32_ps(p_3), acc);
-    }
-
-    *s = hsum_float_8(acc) + summs;
-
-#elif defined __AVX__
-
-    const __m128i m3 = _mm_set1_epi8(3);
-
-    __m256 acc = _mm256_setzero_ps();
-
-    uint32_t ud, um;
-    const uint8_t * restrict db = (const uint8_t *)&ud;
-    const uint8_t * restrict mb = (const uint8_t *)&um;
-
-    float summs = 0;
-
-    // TODO: optimize this
-
-    for (int i = 0; i < nb; ++i) {
-
-        const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d);
-        const float dmin = -y[i].d * GGML_FP16_TO_FP32(x[i].dmin);
-
-        const uint8_t * restrict q2 = x[i].qs;
-        const int8_t  * restrict q8 = y[i].qs;
-
-        const uint32_t * restrict sc = (const uint32_t *)x[i].scales;
-        ud = (sc[0] >> 0) & 0x0f0f0f0f;
-        um = (sc[0] >> 4) & 0x0f0f0f0f;
-
-        int32_t smin = mb[0] * y[i].bsums[0] + mb[1] * y[i].bsums[1] + mb[2] * y[i].bsums[2] + mb[3] * y[i].bsums[3];
-        summs += dmin * smin;
-
-        const __m128i q2bits = _mm_loadu_si128((const __m128i*)q2);
-        const __m128i q2_0 = _mm_and_si128(q2bits, m3);
-        const __m128i q2_1 = _mm_and_si128(_mm_srli_epi16(q2bits, 2), m3);
-        const __m128i q2_2 = _mm_and_si128(_mm_srli_epi16(q2bits, 4), m3);
-        const __m128i q2_3 = _mm_and_si128(_mm_srli_epi16(q2bits, 6), m3);
-
-        const __m256i q8_0 = _mm256_loadu_si256((const __m256i*)(q8+ 0));
-        const __m256i q8_1 = _mm256_loadu_si256((const __m256i*)(q8+32));
-
-        const __m128i p0 = _mm_maddubs_epi16(q2_0, _mm256_extractf128_si256(q8_0, 0));
-        const __m128i p1 = _mm_maddubs_epi16(q2_1, _mm256_extractf128_si256(q8_0, 1));
-        const __m128i p2 = _mm_maddubs_epi16(q2_2, _mm256_extractf128_si256(q8_1, 0));
-        const __m128i p3 = _mm_maddubs_epi16(q2_3, _mm256_extractf128_si256(q8_1, 1));
-
-        const __m256i p_0 = MM256_SET_M128I(_mm_cvtepi16_epi32(_mm_unpackhi_epi64(p0, p0)), _mm_cvtepi16_epi32(p0));
-        const __m256i p_1 = MM256_SET_M128I(_mm_cvtepi16_epi32(_mm_unpackhi_epi64(p1, p1)), _mm_cvtepi16_epi32(p1));
-        const __m256i p_2 = MM256_SET_M128I(_mm_cvtepi16_epi32(_mm_unpackhi_epi64(p2, p2)), _mm_cvtepi16_epi32(p2));
-        const __m256i p_3 = MM256_SET_M128I(_mm_cvtepi16_epi32(_mm_unpackhi_epi64(p3, p3)), _mm_cvtepi16_epi32(p3));
-
-        acc = _mm256_add_ps(_mm256_mul_ps(_mm256_set1_ps(d * db[0]), _mm256_cvtepi32_ps(p_0)), acc);
-        acc = _mm256_add_ps(_mm256_mul_ps(_mm256_set1_ps(d * db[1]), _mm256_cvtepi32_ps(p_1)), acc);
-        acc = _mm256_add_ps(_mm256_mul_ps(_mm256_set1_ps(d * db[2]), _mm256_cvtepi32_ps(p_2)), acc);
-        acc = _mm256_add_ps(_mm256_mul_ps(_mm256_set1_ps(d * db[3]), _mm256_cvtepi32_ps(p_3)), acc);
-    }
-
-    *s = hsum_float_8(acc) + summs;
-
-#elif defined __riscv_v_intrinsic
-
-    uint32_t aux32[2];
-    const uint8_t * scales = (const uint8_t *)aux32;
-
-    float sumf = 0;
-
-    for (int i = 0; i < nb; ++i) {
-
-        const float d = y[i].d * (float)x[i].d;
-        const float dmin = -y[i].d * (float)x[i].dmin;
-
-        const uint8_t * restrict q2 = x[i].qs;
-        const int8_t  * restrict q8 = y[i].qs;
-        const uint32_t * restrict sc = (const uint32_t *)x[i].scales;
-
-        aux32[0] = sc[0] & 0x0f0f0f0f;
-        aux32[1] = (sc[0] >> 4) & 0x0f0f0f0f;
-
-        sumf += dmin * (scales[4] * y[i].bsums[0] + scales[5] * y[i].bsums[1] + scales[6] * y[i].bsums[2] + scales[7] * y[i].bsums[3]);
-
-        int isum1 = 0;
-        int isum2 = 0;
-
-        size_t vl = 16;
-
-        vint16m1_t vzero = __riscv_vmv_v_x_i16m1(0, 1);
-
-        // load Q2
-        vuint8mf2_t q2_x = __riscv_vle8_v_u8mf2(q2, vl);
-
-        vint8mf2_t q2_0 = __riscv_vreinterpret_v_u8mf2_i8mf2(__riscv_vand_vx_u8mf2(q2_x, 0x03, vl));
-        vint8mf2_t q2_1 = __riscv_vreinterpret_v_u8mf2_i8mf2(__riscv_vand_vx_u8mf2(__riscv_vsrl_vx_u8mf2(q2_x, 0x2, vl), 0x03 , vl));
-        vint8mf2_t q2_2 = __riscv_vreinterpret_v_u8mf2_i8mf2(__riscv_vand_vx_u8mf2(__riscv_vsrl_vx_u8mf2(q2_x, 0x4, vl), 0x03 , vl));
-        vint8mf2_t q2_3 = __riscv_vreinterpret_v_u8mf2_i8mf2(__riscv_vand_vx_u8mf2(__riscv_vsrl_vx_u8mf2(q2_x, 0x6, vl), 0x03 , vl));
-
-        // load Q8, and take product with Q2
-        vint16m1_t p0 = __riscv_vwmul_vv_i16m1(q2_0, __riscv_vle8_v_i8mf2(q8, vl), vl);
-        vint16m1_t p1 = __riscv_vwmul_vv_i16m1(q2_1, __riscv_vle8_v_i8mf2(q8+16, vl), vl);
-        vint16m1_t p2 = __riscv_vwmul_vv_i16m1(q2_2, __riscv_vle8_v_i8mf2(q8+32, vl), vl);
-        vint16m1_t p3 = __riscv_vwmul_vv_i16m1(q2_3, __riscv_vle8_v_i8mf2(q8+48, vl), vl);
-
-        vint16m1_t vs_0 = __riscv_vredsum_vs_i16m1_i16m1(p0, vzero, vl);
-        vint16m1_t vs_1 = __riscv_vredsum_vs_i16m1_i16m1(p1, vzero, vl);
-        vint16m1_t vs_2 = __riscv_vredsum_vs_i16m1_i16m1(p2, vzero, vl);
-        vint16m1_t vs_3 = __riscv_vredsum_vs_i16m1_i16m1(p3, vzero, vl);
-
-        isum1 += __riscv_vmv_x_s_i16m1_i16(vs_0) * scales[0];
-        isum2 += __riscv_vmv_x_s_i16m1_i16(vs_1) * scales[1];
-        isum1 += __riscv_vmv_x_s_i16m1_i16(vs_2) * scales[2];
-        isum2 += __riscv_vmv_x_s_i16m1_i16(vs_3) * scales[3];
-
-        sumf += d * (isum1 + isum2);
-
-    }
-
-    *s = sumf;
-
-#else
-
-    float sumf = 0;
-
-    int isum[4];
-
-    for (int i = 0; i < nb; ++i) {
-
-        const uint8_t * q2 = x[i].qs;
-        const  int8_t * q8 = y[i].qs;
-        const uint8_t * sc = x[i].scales;
-
-        int summs = 0;
-        for (int j = 0; j < QK_K/16; ++j) {
-            summs += y[i].bsums[j] * (sc[j] >> 4);
-        }
-
-        const float dall = y[i].d * GGML_FP16_TO_FP32(x[i].d);
-        const float dmin = y[i].d * GGML_FP16_TO_FP32(x[i].dmin);
-
-        isum[0] = isum[1] = isum[2] = isum[3] = 0;
-        for (int l =  0; l < 16; ++l) {
-            isum[0] += q8[l+ 0] * ((q2[l] >> 0) & 3);
-            isum[1] += q8[l+16] * ((q2[l] >> 2) & 3);
-            isum[2] += q8[l+32] * ((q2[l] >> 4) & 3);
-            isum[3] += q8[l+48] * ((q2[l] >> 6) & 3);
-        }
-        for (int l = 0; l < 4; ++l) {
-            isum[l] *= (sc[l] & 0xF);
-        }
-        sumf += dall * (isum[0] + isum[1] + isum[2] + isum[3]) - dmin * summs;
-    }
-    *s = sumf;
-#endif
-}
-#endif
-
-#if QK_K == 256
-void ggml_vec_dot_q3_K_q8_K(const int n, float * restrict s, const void * restrict vx, const void * restrict vy) {
-    assert(n % QK_K == 0);
-
-    const uint32_t kmask1 = 0x03030303;
-    const uint32_t kmask2 = 0x0f0f0f0f;
-
-    const block_q3_K * restrict x = vx;
-    const block_q8_K * restrict y = vy;
-
-    const int nb = n / QK_K;
-
-#ifdef __ARM_NEON
-
-    uint32_t aux[3];
-    uint32_t utmp[4];
-
-    const uint8x16_t m3b = vdupq_n_u8(0x3);
-    const int32x4_t  vzero = vdupq_n_s32(0);
-
-    const uint8x16_t m0 = vdupq_n_u8(1);
-    const uint8x16_t m1 = vshlq_n_u8(m0, 1);
-    const uint8x16_t m2 = vshlq_n_u8(m0, 2);
-    const uint8x16_t m3 = vshlq_n_u8(m0, 3);
-    const int8_t m32 = 32;
-
-    ggml_int8x16x4_t q3bytes;
-
-    float sum = 0;
-
-    for (int i = 0; i < nb; ++i) {
-
-        const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d);
-
-        const uint8_t * restrict q3 = x[i].qs;
-        const uint8_t * restrict qh = x[i].hmask;
-        const int8_t  * restrict q8 = y[i].qs;
-
-        ggml_uint8x16x2_t qhbits = ggml_vld1q_u8_x2(qh);
-
-        ggml_uint8x16x4_t q3h;
-
-        int32_t isum = 0;
-
-        // Set up scales
-        memcpy(aux, x[i].scales, 12);
-        utmp[3] = ((aux[1] >> 4) & kmask2) | (((aux[2] >> 6) & kmask1) << 4);
-        utmp[2] = ((aux[0] >> 4) & kmask2) | (((aux[2] >> 4) & kmask1) << 4);
-        utmp[1] = (aux[1] & kmask2) | (((aux[2] >> 2) & kmask1) << 4);
-        utmp[0] = (aux[0] & kmask2) | (((aux[2] >> 0) & kmask1) << 4);
-
-        int8_t * scale = (int8_t *)utmp;
-        for (int j = 0; j < 16; ++j) scale[j] -= m32;
-
-        for (int j = 0; j < QK_K/128; ++j) {
-
-            const ggml_uint8x16x2_t q3bits = ggml_vld1q_u8_x2(q3); q3 += 32;
-            const ggml_int8x16x4_t q8bytes_1 = ggml_vld1q_s8_x4(q8); q8 += 64;
-            const ggml_int8x16x4_t q8bytes_2 = ggml_vld1q_s8_x4(q8); q8 += 64;
-
-            q3h.val[0] = vshlq_n_u8(vbicq_u8(m0, qhbits.val[0]), 2);
-            q3h.val[1] = vshlq_n_u8(vbicq_u8(m0, qhbits.val[1]), 2);
-            q3h.val[2] = vshlq_n_u8(vbicq_u8(m1, qhbits.val[0]), 1);
-            q3h.val[3] = vshlq_n_u8(vbicq_u8(m1, qhbits.val[1]), 1);
-
-            q3bytes.val[0] = vsubq_s8(vreinterpretq_s8_u8(vandq_u8(q3bits.val[0], m3b)), vreinterpretq_s8_u8(q3h.val[0]));
-            q3bytes.val[1] = vsubq_s8(vreinterpretq_s8_u8(vandq_u8(q3bits.val[1], m3b)), vreinterpretq_s8_u8(q3h.val[1]));
-            q3bytes.val[2] = vsubq_s8(vreinterpretq_s8_u8(vandq_u8(vshrq_n_u8(q3bits.val[0], 2), m3b)), vreinterpretq_s8_u8(q3h.val[2]));
-            q3bytes.val[3] = vsubq_s8(vreinterpretq_s8_u8(vandq_u8(vshrq_n_u8(q3bits.val[1], 2), m3b)), vreinterpretq_s8_u8(q3h.val[3]));
-
-            isum += vaddvq_s32(ggml_vdotq_s32(vzero, q3bytes.val[0], q8bytes_1.val[0])) * scale[0];
-            isum += vaddvq_s32(ggml_vdotq_s32(vzero, q3bytes.val[1], q8bytes_1.val[1])) * scale[1];
-            isum += vaddvq_s32(ggml_vdotq_s32(vzero, q3bytes.val[2], q8bytes_1.val[2])) * scale[2];
-            isum += vaddvq_s32(ggml_vdotq_s32(vzero, q3bytes.val[3], q8bytes_1.val[3])) * scale[3];
-
-            scale += 4;
-
-            q3h.val[0] = vbicq_u8(m2, qhbits.val[0]);
-            q3h.val[1] = vbicq_u8(m2, qhbits.val[1]);
-            q3h.val[2] = vshrq_n_u8(vbicq_u8(m3, qhbits.val[0]), 1);
-            q3h.val[3] = vshrq_n_u8(vbicq_u8(m3, qhbits.val[1]), 1);
-
-            q3bytes.val[0] = vsubq_s8(vreinterpretq_s8_u8(vandq_u8(vshrq_n_u8(q3bits.val[0], 4), m3b)), vreinterpretq_s8_u8(q3h.val[0]));
-            q3bytes.val[1] = vsubq_s8(vreinterpretq_s8_u8(vandq_u8(vshrq_n_u8(q3bits.val[1], 4), m3b)), vreinterpretq_s8_u8(q3h.val[1]));
-            q3bytes.val[2] = vsubq_s8(vreinterpretq_s8_u8(vandq_u8(vshrq_n_u8(q3bits.val[0], 6), m3b)), vreinterpretq_s8_u8(q3h.val[2]));
-            q3bytes.val[3] = vsubq_s8(vreinterpretq_s8_u8(vandq_u8(vshrq_n_u8(q3bits.val[1], 6), m3b)), vreinterpretq_s8_u8(q3h.val[3]));
-
-            isum += vaddvq_s32(ggml_vdotq_s32(vzero, q3bytes.val[0], q8bytes_2.val[0])) * scale[0];
-            isum += vaddvq_s32(ggml_vdotq_s32(vzero, q3bytes.val[1], q8bytes_2.val[1])) * scale[1];
-            isum += vaddvq_s32(ggml_vdotq_s32(vzero, q3bytes.val[2], q8bytes_2.val[2])) * scale[2];
-            isum += vaddvq_s32(ggml_vdotq_s32(vzero, q3bytes.val[3], q8bytes_2.val[3])) * scale[3];
-
-            scale += 4;
-
-            if (j == 0) {
-                qhbits.val[0] = vshrq_n_u8(qhbits.val[0], 4);
-                qhbits.val[1] = vshrq_n_u8(qhbits.val[1], 4);
-            }
-
-        }
-        sum += d * isum;
-
-    }
-
-    *s = sum;
-
-#elif defined __AVX2__
-
-    const __m256i m3 = _mm256_set1_epi8(3);
-    const __m256i mone = _mm256_set1_epi8(1);
-    const __m128i m32 = _mm_set1_epi8(32);
-
-    __m256 acc = _mm256_setzero_ps();
-
-    uint32_t aux[3];
-
-    for (int i = 0; i < nb; ++i) {
-
-        const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d);
-
-        const uint8_t * restrict q3 = x[i].qs;
-        const int8_t  * restrict q8 = y[i].qs;
-
-        // Set up scales
-        memcpy(aux, x[i].scales, 12);
-        __m128i scales128 = _mm_set_epi32(
-                ((aux[1] >> 4) & kmask2) | (((aux[2] >> 6) & kmask1) << 4),
-                ((aux[0] >> 4) & kmask2) | (((aux[2] >> 4) & kmask1) << 4),
-                (aux[1] & kmask2) | (((aux[2] >> 2) & kmask1) << 4),
-                (aux[0] & kmask2) | (((aux[2] >> 0) & kmask1) << 4));
-        scales128 = _mm_sub_epi8(scales128, m32);
-        const __m256i all_scales = _mm256_cvtepi8_epi16(scales128);
-        const __m128i l_scales = _mm256_extracti128_si256(all_scales, 0);
-        const __m128i h_scales = _mm256_extracti128_si256(all_scales, 1);
-        const __m256i scales[2] = {MM256_SET_M128I(l_scales, l_scales), MM256_SET_M128I(h_scales, h_scales)};
-
-        // high bit
-        const __m256i hbits = _mm256_loadu_si256((const __m256i*)x[i].hmask);
-
-        // integer accumulator
-        __m256i sumi = _mm256_setzero_si256();
-
-        int bit = 0;
-        int is  = 0;
-
-        for (int j = 0; j < QK_K/128; ++j) {
-            // load low 2 bits
-            const __m256i q3bits = _mm256_loadu_si256((const __m256i*)q3); q3 += 32;
-
-            // prepare low and high bits
-            const __m256i q3l_0 = _mm256_and_si256(q3bits, m3);
-            const __m256i q3h_0 = _mm256_slli_epi16(_mm256_srli_epi16(_mm256_andnot_si256(hbits, _mm256_slli_epi16(mone, bit)), bit), 2);
-            ++bit;
-
-            const __m256i q3l_1 = _mm256_and_si256(_mm256_srli_epi16(q3bits, 2), m3);
-            const __m256i q3h_1 = _mm256_slli_epi16(_mm256_srli_epi16(_mm256_andnot_si256(hbits, _mm256_slli_epi16(mone, bit)), bit), 2);
-            ++bit;
-
-            const __m256i q3l_2 = _mm256_and_si256(_mm256_srli_epi16(q3bits, 4), m3);
-            const __m256i q3h_2 = _mm256_slli_epi16(_mm256_srli_epi16(_mm256_andnot_si256(hbits, _mm256_slli_epi16(mone, bit)), bit), 2);
-            ++bit;
-
-            const __m256i q3l_3 = _mm256_and_si256(_mm256_srli_epi16(q3bits, 6), m3);
-            const __m256i q3h_3 = _mm256_slli_epi16(_mm256_srli_epi16(_mm256_andnot_si256(hbits, _mm256_slli_epi16(mone, bit)), bit), 2);
-            ++bit;
-
-            // load Q8 quants
-            const __m256i q8_0 = _mm256_loadu_si256((const __m256i*)q8); q8 += 32;
-            const __m256i q8_1 = _mm256_loadu_si256((const __m256i*)q8); q8 += 32;
-            const __m256i q8_2 = _mm256_loadu_si256((const __m256i*)q8); q8 += 32;
-            const __m256i q8_3 = _mm256_loadu_si256((const __m256i*)q8); q8 += 32;
-
-            // Dot product: we multiply the 2 low bits and 1 high bit part separately, so we can use _mm256_maddubs_epi16,
-            // and then subtract. The high bit part has the 2 already subtracted (and so, it is zero if the high bit was not set,
-            // and 2 if the high bit was set)
-            __m256i q8s_0 = _mm256_maddubs_epi16(q3h_0, q8_0);
-            __m256i q8s_1 = _mm256_maddubs_epi16(q3h_1, q8_1);
-            __m256i q8s_2 = _mm256_maddubs_epi16(q3h_2, q8_2);
-            __m256i q8s_3 = _mm256_maddubs_epi16(q3h_3, q8_3);
-
-            __m256i p16_0 = _mm256_maddubs_epi16(q3l_0, q8_0);
-            __m256i p16_1 = _mm256_maddubs_epi16(q3l_1, q8_1);
-            __m256i p16_2 = _mm256_maddubs_epi16(q3l_2, q8_2);
-            __m256i p16_3 = _mm256_maddubs_epi16(q3l_3, q8_3);
-
-            p16_0 = _mm256_sub_epi16(p16_0, q8s_0);
-            p16_1 = _mm256_sub_epi16(p16_1, q8s_1);
-            p16_2 = _mm256_sub_epi16(p16_2, q8s_2);
-            p16_3 = _mm256_sub_epi16(p16_3, q8s_3);
-
-            // multiply with scales
-            p16_0 = _mm256_madd_epi16(_mm256_shuffle_epi8(scales[j], get_scale_shuffle_q3k(is + 0)), p16_0);
-            p16_1 = _mm256_madd_epi16(_mm256_shuffle_epi8(scales[j], get_scale_shuffle_q3k(is + 1)), p16_1);
-            p16_2 = _mm256_madd_epi16(_mm256_shuffle_epi8(scales[j], get_scale_shuffle_q3k(is + 2)), p16_2);
-            p16_3 = _mm256_madd_epi16(_mm256_shuffle_epi8(scales[j], get_scale_shuffle_q3k(is + 3)), p16_3);
-
-            // accumulate
-            p16_0 = _mm256_add_epi32(p16_0, p16_1);
-            p16_2 = _mm256_add_epi32(p16_2, p16_3);
-            sumi  = _mm256_add_epi32(sumi, _mm256_add_epi32(p16_0, p16_2));
-
-        }
-
-        // multiply with block scale and accumulate
-        acc = _mm256_fmadd_ps(_mm256_broadcast_ss(&d), _mm256_cvtepi32_ps(sumi), acc);
-
-    }
-
-    *s = hsum_float_8(acc);
-
-#elif defined __AVX__
-
-    const __m128i m3 = _mm_set1_epi8(3);
-    const __m128i mone = _mm_set1_epi8(1);
-    const __m128i m32 = _mm_set1_epi8(32);
-    const __m128i m2 = _mm_set1_epi8(2);
-
-    __m256 acc = _mm256_setzero_ps();
-
-    const uint32_t *aux;
-
-    for (int i = 0; i < nb; ++i) {
-
-        const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d);
-
-        const uint8_t * restrict q3 = x[i].qs;
-        const int8_t  * restrict q8 = y[i].qs;
-
-        // Set up scales
-        aux = (const uint32_t *)x[i].scales;
-        __m128i scales128 = _mm_set_epi32(
-                ((aux[1] >> 4) & kmask2) | (((aux[2] >> 6) & kmask1) << 4),
-                ((aux[0] >> 4) & kmask2) | (((aux[2] >> 4) & kmask1) << 4),
-                (aux[1] & kmask2) | (((aux[2] >> 2) & kmask1) << 4),
-                (aux[0] & kmask2) | (((aux[2] >> 0) & kmask1) << 4));
-        scales128 = _mm_sub_epi8(scales128, m32);
-        const __m128i scales_0 = _mm_cvtepi8_epi16(scales128);
-        const __m128i scales_1 = _mm_cvtepi8_epi16(_mm_unpackhi_epi64(scales128, scales128));
-        const __m128i scales[2] = { scales_0, scales_1 };
-
-        // high bit *128*2 from block_q3_K.hmask[QK_K/8]
-        const __m128i hbits_0 = _mm_loadu_si128((const __m128i*)&x[i].hmask[0]);
-        const __m128i hbits_1 = _mm_loadu_si128((const __m128i*)&x[i].hmask[16]);
-
-        // integer accumulator
-        __m128i sumi_0 = _mm_setzero_si128();
-        __m128i sumi_1 = _mm_setzero_si128();
-
-        for (int j = 0; j < QK_K/128; ++j) {
-            // load low 2 bits *64*2 from block_q3_K.qs[QK_K/4]
-            const __m128i q3bits_0 = _mm_loadu_si128((const __m128i*)q3); q3 += 16;
-            const __m128i q3bits_1 = _mm_loadu_si128((const __m128i*)q3); q3 += 16;
-
-            // prepare low and high bits
-            const int bit = j << 2;
-
-            const __m128i q3l_0 = _mm_and_si128(q3bits_0, m3);
-            const __m128i q3l_1 = _mm_and_si128(q3bits_1, m3);
-            const __m128i q3h_0 = _mm_slli_epi16(_mm_srli_epi16(_mm_andnot_si128(hbits_0, _mm_slli_epi16(mone, bit)), bit), 2);
-            const __m128i q3h_1 = _mm_slli_epi16(_mm_srli_epi16(_mm_andnot_si128(hbits_1, _mm_slli_epi16(mone, bit)), bit), 2);
-
-            const __m128i q3l_2 = _mm_and_si128(_mm_srli_epi16(q3bits_0, 2), m3);
-            const __m128i q3l_3 = _mm_and_si128(_mm_srli_epi16(q3bits_1, 2), m3);
-            const __m128i q3h_2 = _mm_slli_epi16(_mm_srli_epi16(_mm_andnot_si128(hbits_0, _mm_slli_epi16(mone, bit+1)), bit+1), 2);
-            const __m128i q3h_3 = _mm_slli_epi16(_mm_srli_epi16(_mm_andnot_si128(hbits_1, _mm_slli_epi16(mone, bit+1)), bit+1), 2);
-
-            const __m128i q3l_4 = _mm_and_si128(_mm_srli_epi16(q3bits_0, 4), m3);
-            const __m128i q3l_5 = _mm_and_si128(_mm_srli_epi16(q3bits_1, 4), m3);
-            const __m128i q3h_4 = _mm_slli_epi16(_mm_srli_epi16(_mm_andnot_si128(hbits_0, _mm_slli_epi16(mone, bit+2)), bit+2), 2);
-            const __m128i q3h_5 = _mm_slli_epi16(_mm_srli_epi16(_mm_andnot_si128(hbits_1, _mm_slli_epi16(mone, bit+2)), bit+2), 2);
-
-            const __m128i q3l_6 = _mm_and_si128(_mm_srli_epi16(q3bits_0, 6), m3);
-            const __m128i q3l_7 = _mm_and_si128(_mm_srli_epi16(q3bits_1, 6), m3);
-            const __m128i q3h_6 = _mm_slli_epi16(_mm_srli_epi16(_mm_andnot_si128(hbits_0, _mm_slli_epi16(mone, bit+3)), bit+3), 2);
-            const __m128i q3h_7 = _mm_slli_epi16(_mm_srli_epi16(_mm_andnot_si128(hbits_1, _mm_slli_epi16(mone, bit+3)), bit+3), 2);
-
-            // load Q8 quants from block_q8_K.qs[QK_K]
-            const __m128i q8_0 = _mm_loadu_si128((const __m128i*)q8); q8 += 16;
-            const __m128i q8_1 = _mm_loadu_si128((const __m128i*)q8); q8 += 16;
-            const __m128i q8_2 = _mm_loadu_si128((const __m128i*)q8); q8 += 16;
-            const __m128i q8_3 = _mm_loadu_si128((const __m128i*)q8); q8 += 16;
-            const __m128i q8_4 = _mm_loadu_si128((const __m128i*)q8); q8 += 16;
-            const __m128i q8_5 = _mm_loadu_si128((const __m128i*)q8); q8 += 16;
-            const __m128i q8_6 = _mm_loadu_si128((const __m128i*)q8); q8 += 16;
-            const __m128i q8_7 = _mm_loadu_si128((const __m128i*)q8); q8 += 16;
-
-            // Dot product: we multiply the 2 low bits and 1 high bit part separately, so we can use _mm256_maddubs_epi16,
-            // and then subtract. The high bit part has the 2 already subtracted (and so, it is zero if the high bit was not set,
-            // and 2 if the high bit was set)
-            __m128i q8s_0 = _mm_maddubs_epi16(q3h_0, q8_0);
-            __m128i q8s_1 = _mm_maddubs_epi16(q3h_1, q8_1);
-            __m128i q8s_2 = _mm_maddubs_epi16(q3h_2, q8_2);
-            __m128i q8s_3 = _mm_maddubs_epi16(q3h_3, q8_3);
-            __m128i q8s_4 = _mm_maddubs_epi16(q3h_4, q8_4);
-            __m128i q8s_5 = _mm_maddubs_epi16(q3h_5, q8_5);
-            __m128i q8s_6 = _mm_maddubs_epi16(q3h_6, q8_6);
-            __m128i q8s_7 = _mm_maddubs_epi16(q3h_7, q8_7);
-
-            __m128i p16_0 = _mm_maddubs_epi16(q3l_0, q8_0);
-            __m128i p16_1 = _mm_maddubs_epi16(q3l_1, q8_1);
-            __m128i p16_2 = _mm_maddubs_epi16(q3l_2, q8_2);
-            __m128i p16_3 = _mm_maddubs_epi16(q3l_3, q8_3);
-            __m128i p16_4 = _mm_maddubs_epi16(q3l_4, q8_4);
-            __m128i p16_5 = _mm_maddubs_epi16(q3l_5, q8_5);
-            __m128i p16_6 = _mm_maddubs_epi16(q3l_6, q8_6);
-            __m128i p16_7 = _mm_maddubs_epi16(q3l_7, q8_7);
-
-            p16_0 = _mm_sub_epi16(p16_0, q8s_0);
-            p16_1 = _mm_sub_epi16(p16_1, q8s_1);
-            p16_2 = _mm_sub_epi16(p16_2, q8s_2);
-            p16_3 = _mm_sub_epi16(p16_3, q8s_3);
-            p16_4 = _mm_sub_epi16(p16_4, q8s_4);
-            p16_5 = _mm_sub_epi16(p16_5, q8s_5);
-            p16_6 = _mm_sub_epi16(p16_6, q8s_6);
-            p16_7 = _mm_sub_epi16(p16_7, q8s_7);
-
-            // multiply with scales
-            __m128i shuffle = _mm_set1_epi16(0x0100);
-            p16_0 = _mm_madd_epi16(_mm_shuffle_epi8(scales[j], shuffle), p16_0);
-            shuffle = _mm_add_epi16(shuffle, m2);
-            p16_1 = _mm_madd_epi16(_mm_shuffle_epi8(scales[j], shuffle), p16_1);
-            shuffle = _mm_add_epi16(shuffle, m2);
-            p16_2 = _mm_madd_epi16(_mm_shuffle_epi8(scales[j], shuffle), p16_2);
-            shuffle = _mm_add_epi16(shuffle, m2);
-            p16_3 = _mm_madd_epi16(_mm_shuffle_epi8(scales[j], shuffle), p16_3);
-            shuffle = _mm_add_epi16(shuffle, m2);
-            p16_4 = _mm_madd_epi16(_mm_shuffle_epi8(scales[j], shuffle), p16_4);
-            shuffle = _mm_add_epi16(shuffle, m2);
-            p16_5 = _mm_madd_epi16(_mm_shuffle_epi8(scales[j], shuffle), p16_5);
-            shuffle = _mm_add_epi16(shuffle, m2);
-            p16_6 = _mm_madd_epi16(_mm_shuffle_epi8(scales[j], shuffle), p16_6);
-            shuffle = _mm_add_epi16(shuffle, m2);
-            p16_7 = _mm_madd_epi16(_mm_shuffle_epi8(scales[j], shuffle), p16_7);
-
-            // accumulate
-            p16_0 = _mm_add_epi32(p16_0, p16_1);
-            p16_2 = _mm_add_epi32(p16_2, p16_3);
-            p16_4 = _mm_add_epi32(p16_4, p16_5);
-            p16_6 = _mm_add_epi32(p16_6, p16_7);
-            sumi_0 = _mm_add_epi32(sumi_0, _mm_add_epi32(p16_0, p16_2));
-            sumi_1 = _mm_add_epi32(sumi_1, _mm_add_epi32(p16_4, p16_6));
-
-        }
-
-        // multiply with block scale and accumulate
-        __m256i sumi = MM256_SET_M128I(sumi_1, sumi_0);
-        acc = _mm256_add_ps(_mm256_mul_ps(_mm256_broadcast_ss(&d), _mm256_cvtepi32_ps(sumi)), acc);
-
-    }
-
-    *s = hsum_float_8(acc);
-
-#elif defined __riscv_v_intrinsic
-
-    uint32_t aux[3];
-    uint32_t utmp[4];
-
-    float sumf = 0;
-    for (int i = 0; i < nb; ++i) {
-
-        const uint8_t * restrict q3 = x[i].qs;
-        const uint8_t * restrict qh = x[i].hmask;
-        const  int8_t * restrict q8 = y[i].qs;
-
-        memcpy(aux, x[i].scales, 12);
-        utmp[3] = ((aux[1] >> 4) & kmask2) | (((aux[2] >> 6) & kmask1) << 4);
-        utmp[2] = ((aux[0] >> 4) & kmask2) | (((aux[2] >> 4) & kmask1) << 4);
-        utmp[1] = (aux[1] & kmask2) | (((aux[2] >> 2) & kmask1) << 4);
-        utmp[0] = (aux[0] & kmask2) | (((aux[2] >> 0) & kmask1) << 4);
-
-        int8_t * scale = (int8_t *)utmp;
-        for (int j = 0; j < 16; ++j) scale[j] -= 32;
-
-
-        size_t vl = 32;
-        uint8_t m =  1;
-
-        vint32m1_t vzero = __riscv_vmv_v_x_i32m1(0, 1);
-        vuint8m1_t vqh = __riscv_vle8_v_u8m1(qh, vl);
-
-        int sum_t = 0;
-
-        for (int j = 0; j < QK_K; j += 128) {
-
-            vl = 32;
-
-            // load Q3
-            vuint8m1_t q3_x = __riscv_vle8_v_u8m1(q3, vl);
-
-            vint8m1_t q3_0 = __riscv_vreinterpret_v_u8m1_i8m1(__riscv_vand_vx_u8m1(q3_x, 0x03, vl));
-            vint8m1_t q3_1 = __riscv_vreinterpret_v_u8m1_i8m1(__riscv_vand_vx_u8m1(__riscv_vsrl_vx_u8m1(q3_x, 0x2, vl), 0x03 , vl));
-            vint8m1_t q3_2 = __riscv_vreinterpret_v_u8m1_i8m1(__riscv_vand_vx_u8m1(__riscv_vsrl_vx_u8m1(q3_x, 0x4, vl), 0x03 , vl));
-            vint8m1_t q3_3 = __riscv_vreinterpret_v_u8m1_i8m1(__riscv_vand_vx_u8m1(__riscv_vsrl_vx_u8m1(q3_x, 0x6, vl), 0x03 , vl));
-
-            // compute mask for subtraction
-            vuint8m1_t qh_m0 = __riscv_vand_vx_u8m1(vqh, m, vl);
-            vbool8_t vmask_0 = __riscv_vmseq_vx_u8m1_b8(qh_m0, 0, vl);
-            vint8m1_t q3_m0 = __riscv_vsub_vx_i8m1_m(vmask_0, q3_0, 0x4, vl);
-            m <<= 1;
-
-            vuint8m1_t qh_m1 = __riscv_vand_vx_u8m1(vqh, m, vl);
-            vbool8_t vmask_1 = __riscv_vmseq_vx_u8m1_b8(qh_m1, 0, vl);
-            vint8m1_t q3_m1 = __riscv_vsub_vx_i8m1_m(vmask_1, q3_1, 0x4, vl);
-            m <<= 1;
-
-            vuint8m1_t qh_m2 = __riscv_vand_vx_u8m1(vqh, m, vl);
-            vbool8_t vmask_2 = __riscv_vmseq_vx_u8m1_b8(qh_m2, 0, vl);
-            vint8m1_t q3_m2 = __riscv_vsub_vx_i8m1_m(vmask_2, q3_2, 0x4, vl);
-            m <<= 1;
-
-            vuint8m1_t qh_m3 = __riscv_vand_vx_u8m1(vqh, m, vl);
-            vbool8_t vmask_3 = __riscv_vmseq_vx_u8m1_b8(qh_m3, 0, vl);
-            vint8m1_t q3_m3 = __riscv_vsub_vx_i8m1_m(vmask_3, q3_3, 0x4, vl);
-            m <<= 1;
-
-            // load Q8 and take product with Q3
-            vint16m2_t a0 = __riscv_vwmul_vv_i16m2(q3_m0, __riscv_vle8_v_i8m1(q8, vl), vl);
-            vint16m2_t a1 = __riscv_vwmul_vv_i16m2(q3_m1, __riscv_vle8_v_i8m1(q8+32, vl), vl);
-            vint16m2_t a2 = __riscv_vwmul_vv_i16m2(q3_m2, __riscv_vle8_v_i8m1(q8+64, vl), vl);
-            vint16m2_t a3 = __riscv_vwmul_vv_i16m2(q3_m3, __riscv_vle8_v_i8m1(q8+96, vl), vl);
-
-            vl = 16;
-
-            // retrieve lane to multiply with scale
-            vint32m2_t aux0_0 = __riscv_vwmul_vx_i32m2(__riscv_vget_v_i16m2_i16m1(a0, 0), (scale[0]), vl);
-            vint32m2_t aux0_1 = __riscv_vwmul_vx_i32m2(__riscv_vget_v_i16m2_i16m1(a0, 1), (scale[1]), vl);
-            vint32m2_t aux1_0 = __riscv_vwmul_vx_i32m2(__riscv_vget_v_i16m2_i16m1(a1, 0), (scale[2]), vl);
-            vint32m2_t aux1_1 = __riscv_vwmul_vx_i32m2(__riscv_vget_v_i16m2_i16m1(a1, 1), (scale[3]), vl);
-            vint32m2_t aux2_0 = __riscv_vwmul_vx_i32m2(__riscv_vget_v_i16m2_i16m1(a2, 0), (scale[4]), vl);
-            vint32m2_t aux2_1 = __riscv_vwmul_vx_i32m2(__riscv_vget_v_i16m2_i16m1(a2, 1), (scale[5]), vl);
-            vint32m2_t aux3_0 = __riscv_vwmul_vx_i32m2(__riscv_vget_v_i16m2_i16m1(a3, 0), (scale[6]), vl);
-            vint32m2_t aux3_1 = __riscv_vwmul_vx_i32m2(__riscv_vget_v_i16m2_i16m1(a3, 1), (scale[7]), vl);
-
-            vint32m1_t isum0 = __riscv_vredsum_vs_i32m2_i32m1(__riscv_vadd_vv_i32m2(aux0_0, aux0_1, vl), vzero, vl);
-            vint32m1_t isum1 = __riscv_vredsum_vs_i32m2_i32m1(__riscv_vadd_vv_i32m2(aux1_0, aux1_1, vl), isum0, vl);
-            vint32m1_t isum2 = __riscv_vredsum_vs_i32m2_i32m1(__riscv_vadd_vv_i32m2(aux2_0, aux2_1, vl), isum1, vl);
-            vint32m1_t isum3 = __riscv_vredsum_vs_i32m2_i32m1(__riscv_vadd_vv_i32m2(aux3_0, aux3_1, vl), isum2, vl);
-
-            sum_t +=  __riscv_vmv_x_s_i32m1_i32(isum3);
-
-            q3 += 32;    q8 += 128;   scale += 8;
-
-        }
-
-        const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
-
-        sumf += d*sum_t;
-
-    }
-
-    *s = sumf;
-
-#else
-    // scalar version
-    // This function is written like this so the compiler can manage to vectorize most of it
-    // Using -Ofast, GCC and clang manage to produce code that is within a factor of 2 or so from the
-    // manually vectorized version above. Every other version I tried would run at least 4 times slower.
-    // The ideal situation would be if we could just write the code once, and the compiler would
-    // automatically produce the best possible set of machine instructions, instead of us having to manually
-    // write vectorized versions for AVX, ARM_NEON, etc.
-
-    int8_t  aux8[QK_K];
-    int16_t aux16[8];
-    float   sums [8];
-    int32_t aux32[8];
-    memset(sums, 0, 8*sizeof(float));
-
-    uint32_t auxs[4];
-    const int8_t * scales = (const int8_t*)auxs;
-
-    float sumf = 0;
-    for (int i = 0; i < nb; ++i) {
-        const uint8_t * restrict q3 = x[i].qs;
-        const uint8_t * restrict hm = x[i].hmask;
-        const  int8_t * restrict q8 = y[i].qs;
-        memset(aux32, 0, 8*sizeof(int32_t));
-        int8_t * restrict a = aux8;
-        uint8_t m = 1;
-        for (int j = 0; j < QK_K; j += 128) {
-            for (int l = 0; l < 32; ++l) a[l] = q3[l] & 3;
-            for (int l = 0; l < 32; ++l) a[l] -= (hm[l] & m ? 0 : 4);
-            a += 32; m <<= 1;
-            for (int l = 0; l < 32; ++l) a[l] = (q3[l] >> 2) & 3;
-            for (int l = 0; l < 32; ++l) a[l] -= (hm[l] & m ? 0 : 4);
-            a += 32; m <<= 1;
-            for (int l = 0; l < 32; ++l) a[l] = (q3[l] >> 4) & 3;
-            for (int l = 0; l < 32; ++l) a[l] -= (hm[l] & m ? 0 : 4);
-            a += 32; m <<= 1;
-            for (int l = 0; l < 32; ++l) a[l] = (q3[l] >> 6) & 3;
-            for (int l = 0; l < 32; ++l) a[l] -= (hm[l] & m ? 0 : 4);
-            a += 32; m <<= 1;
-            q3 += 32;
-        }
-        a = aux8;
-
-        memcpy(auxs, x[i].scales, 12);
-        uint32_t tmp = auxs[2];
-        auxs[2] = ((auxs[0] >> 4) & kmask2) | (((tmp >> 4) & kmask1) << 4);
-        auxs[3] = ((auxs[1] >> 4) & kmask2) | (((tmp >> 6) & kmask1) << 4);
-        auxs[0] = (auxs[0] & kmask2) | (((tmp >> 0) & kmask1) << 4);
-        auxs[1] = (auxs[1] & kmask2) | (((tmp >> 2) & kmask1) << 4);
-        for (int j = 0; j < QK_K/16; ++j) {
-            for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
-            for (int l = 0; l < 8; ++l) aux32[l] += (scales[j] - 32) * aux16[l];
-            q8 += 8; a += 8;
-            for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
-            for (int l = 0; l < 8; ++l) aux32[l] += (scales[j] - 32) * aux16[l];
-            q8 += 8; a += 8;
-        }
-        const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
-        for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l];
-    }
-    for (int l = 0; l < 8; ++l) sumf += sums[l];
-    *s = sumf;
-
-#endif
-
-}
-
-#else
-
-void ggml_vec_dot_q3_K_q8_K(const int n, float * restrict s, const void * restrict vx, const void * restrict vy) {
-    assert(n % QK_K == 0);
-
-    const block_q3_K * restrict x = vx;
-    const block_q8_K * restrict y = vy;
-
-    const int nb = n / QK_K;
-
-#ifdef __ARM_NEON
-    const int32x4_t vzero = vdupq_n_s32(0);
-
-    const uint8x16_t m3b = vdupq_n_u8(0x3);
-    const uint8x16_t mh  = vdupq_n_u8(4);
-
-    ggml_int8x16x4_t q3bytes;
-
-    uint16_t aux16[2];
-    int8_t * scales = (int8_t *)aux16;
-
-    float sum = 0;
-
-    for (int i = 0; i < nb; ++i) {
-
-        ggml_uint8x16x4_t q3h;
-
-        const uint8x8_t  hbits    = vld1_u8(x[i].hmask);
-        const uint8x16_t q3bits   = vld1q_u8(x[i].qs);
-        const ggml_int8x16x4_t q8bytes = ggml_vld1q_s8_x4(y[i].qs);
-
-        const uint16_t a = *(const uint16_t *)x[i].scales;
-        aux16[0] = a & 0x0f0f;
-        aux16[1] = (a >> 4) & 0x0f0f;
-
-        for (int j = 0; j < 4; ++j) scales[j] -= 8;
-
-        int32_t isum = -4*(scales[0] * y[i].bsums[0] + scales[2] * y[i].bsums[1] + scales[1] * y[i].bsums[2] + scales[3] * y[i].bsums[3]);
-
-        const float d = y[i].d * (float)x[i].d;
-
-        const uint8x16_t htmp = vcombine_u8(hbits, vshr_n_u8(hbits, 1));
-        q3h.val[0] = vandq_u8(mh, vshlq_n_u8(htmp, 2));
-        q3h.val[1] = vandq_u8(mh, htmp);
-        q3h.val[2] = vandq_u8(mh, vshrq_n_u8(htmp, 2));
-        q3h.val[3] = vandq_u8(mh, vshrq_n_u8(htmp, 4));
-
-        q3bytes.val[0] = vreinterpretq_s8_u8(vorrq_u8(vandq_u8(q3bits, m3b),                q3h.val[0]));
-        q3bytes.val[1] = vreinterpretq_s8_u8(vorrq_u8(vandq_u8(vshrq_n_u8(q3bits, 2), m3b), q3h.val[1]));
-        q3bytes.val[2] = vreinterpretq_s8_u8(vorrq_u8(vandq_u8(vshrq_n_u8(q3bits, 4), m3b), q3h.val[2]));
-        q3bytes.val[3] = vreinterpretq_s8_u8(vorrq_u8(vshrq_n_u8(q3bits, 6),                q3h.val[3]));
-
-        isum += vaddvq_s32(ggml_vdotq_s32(vzero, q3bytes.val[0], q8bytes.val[0])) * scales[0];
-        isum += vaddvq_s32(ggml_vdotq_s32(vzero, q3bytes.val[1], q8bytes.val[1])) * scales[2];
-        isum += vaddvq_s32(ggml_vdotq_s32(vzero, q3bytes.val[2], q8bytes.val[2])) * scales[1];
-        isum += vaddvq_s32(ggml_vdotq_s32(vzero, q3bytes.val[3], q8bytes.val[3])) * scales[3];
-
-        sum += d * isum;
-
-    }
-
-    *s = sum;
-
-#elif defined __AVX2__
-
-    const __m256i m3 = _mm256_set1_epi8(3);
-    const __m256i m1 = _mm256_set1_epi8(1);
-
-    __m256 acc = _mm256_setzero_ps();
-
-    uint64_t aux64;
-
-    uint16_t aux16[2];
-    const int8_t * aux8 = (const int8_t *)aux16;
-
-    for (int i = 0; i < nb; ++i) {
-
-        const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d);
-
-        const uint8_t * restrict q3 = x[i].qs;
-        const int8_t  * restrict q8 = y[i].qs;
-
-        const uint16_t a = *(const uint16_t *)x[i].scales;
-        aux16[0] = a & 0x0f0f;
-        aux16[1] = (a >> 4) & 0x0f0f;
-
-        const __m256i scale_0 = MM256_SET_M128I(_mm_set1_epi16(aux8[2] - 8), _mm_set1_epi16(aux8[0] - 8));
-        const __m256i scale_1 = MM256_SET_M128I(_mm_set1_epi16(aux8[3] - 8), _mm_set1_epi16(aux8[1] - 8));
-
-        memcpy(&aux64, x[i].hmask, 8);
-
-        const __m128i haux = _mm_set_epi64x(aux64 >> 1, aux64 >> 0);
-        __m256i q3h_0 = MM256_SET_M128I(_mm_srli_epi16(haux, 2), haux);
-        __m256i q3h_1 = _mm256_srli_epi16(q3h_0, 4);
-        q3h_0 = _mm256_slli_epi16(_mm256_andnot_si256(q3h_0, m1), 2);
-        q3h_1 = _mm256_slli_epi16(_mm256_andnot_si256(q3h_1, m1), 2);
-
-        // load low 2 bits
-        const __m128i q3bits = _mm_loadu_si128((const __m128i*)q3);
-
-        // prepare low and high bits
-        const __m256i q3aux  = MM256_SET_M128I(_mm_srli_epi16(q3bits, 2), q3bits);
-        const __m256i q3l_0 = _mm256_and_si256(q3aux, m3);
-        const __m256i q3l_1 = _mm256_and_si256(_mm256_srli_epi16(q3aux, 4), m3);
-
-        // load Q8 quants
-        const __m256i q8_0 = _mm256_loadu_si256((const __m256i*)(q8+ 0));
-        const __m256i q8_1 = _mm256_loadu_si256((const __m256i*)(q8+32));
-
-        // Dot product: we multiply the 2 low bits and 1 high bit part separately, so we can use _mm256_maddubs_epi16,
-        // and then subtract. The high bit part has the 2 already subtracted (and so, it is zero if the high bit was not set,
-        // and 2 if the high bit was set)
-        const __m256i q8s_0 = _mm256_maddubs_epi16(q3h_0, q8_0);
-        const __m256i q8s_1 = _mm256_maddubs_epi16(q3h_1, q8_1);
-
-        __m256i p16_0 = _mm256_maddubs_epi16(q3l_0, q8_0);
-        __m256i p16_1 = _mm256_maddubs_epi16(q3l_1, q8_1);
-
-        p16_0 = _mm256_sub_epi16(p16_0, q8s_0);
-        p16_1 = _mm256_sub_epi16(p16_1, q8s_1);
-
-        // multiply with scales
-        p16_0 = _mm256_madd_epi16(scale_0, p16_0);
-        p16_1 = _mm256_madd_epi16(scale_1, p16_1);
-
-        p16_0 = _mm256_add_epi32(p16_0, p16_1);
-
-        // multiply with block scale and accumulate
-        acc = _mm256_fmadd_ps(_mm256_broadcast_ss(&d), _mm256_cvtepi32_ps(p16_0), acc);
-
-    }
-
-    *s = hsum_float_8(acc);
-
-#elif defined __AVX__
-
-    const __m128i m3 = _mm_set1_epi8(3);
-    const __m128i m1 = _mm_set1_epi8(1);
-
-    __m256 acc = _mm256_setzero_ps();
-
-    uint64_t aux64;
-
-    uint16_t aux16[2];
-    const int8_t * aux8 = (const int8_t *)aux16;
-
-    for (int i = 0; i < nb; ++i) {
-
-        const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d);
-
-        const uint8_t * restrict q3 = x[i].qs;
-        const int8_t  * restrict q8 = y[i].qs;
-
-        const uint16_t a = *(const uint16_t *)x[i].scales;
-        aux16[0] = a & 0x0f0f;
-        aux16[1] = (a >> 4) & 0x0f0f;
-
-        const __m128i scale_0 = _mm_set1_epi16(aux8[0] - 8);
-        const __m128i scale_1 = _mm_set1_epi16(aux8[2] - 8);
-        const __m128i scale_2 = _mm_set1_epi16(aux8[1] - 8);
-        const __m128i scale_3 = _mm_set1_epi16(aux8[3] - 8);
-
-        memcpy(&aux64, x[i].hmask, 8);
-
-        __m128i q3h_0 = _mm_set_epi64x(aux64 >> 1, aux64 >> 0);
-        __m128i q3h_1 = _mm_srli_epi16(q3h_0, 2);
-        __m128i q3h_2 = _mm_srli_epi16(q3h_0, 4);
-        __m128i q3h_3 = _mm_srli_epi16(q3h_0, 6);
-        q3h_0 = _mm_slli_epi16(_mm_andnot_si128(q3h_0, m1), 2);
-        q3h_1 = _mm_slli_epi16(_mm_andnot_si128(q3h_1, m1), 2);
-        q3h_2 = _mm_slli_epi16(_mm_andnot_si128(q3h_2, m1), 2);
-        q3h_3 = _mm_slli_epi16(_mm_andnot_si128(q3h_3, m1), 2);
-
-        // load low 2 bits
-        const __m128i q3bits = _mm_loadu_si128((const __m128i*)q3);
-
-        // prepare low and high bits
-        const __m128i q3l_0 = _mm_and_si128(q3bits, m3);
-        const __m128i q3l_1 = _mm_and_si128(_mm_srli_epi16(q3bits, 2), m3);
-        const __m128i q3l_2 = _mm_and_si128(_mm_srli_epi16(q3bits, 4), m3);
-        const __m128i q3l_3 = _mm_and_si128(_mm_srli_epi16(q3bits, 6), m3);
-
-        // load Q8 quants
-        const __m256i q8_0 = _mm256_loadu_si256((const __m256i*)(q8+ 0));
-        const __m256i q8_1 = _mm256_loadu_si256((const __m256i*)(q8+32));
-
-        // Dot product: we multiply the 2 low bits and 1 high bit part separately, so we can use _mm_maddubs_epi16,
-        // and then subtract. The high bit part has the 2 already subtracted (and so, it is zero if the high bit was not set,
-        // and 2 if the high bit was set)
-        const __m128i q8s_0 = _mm_maddubs_epi16(q3h_0, _mm256_extractf128_si256(q8_0, 0));
-        const __m128i q8s_1 = _mm_maddubs_epi16(q3h_1, _mm256_extractf128_si256(q8_0, 1));
-        const __m128i q8s_2 = _mm_maddubs_epi16(q3h_2, _mm256_extractf128_si256(q8_1, 0));
-        const __m128i q8s_3 = _mm_maddubs_epi16(q3h_3, _mm256_extractf128_si256(q8_1, 1));
-
-        __m128i p16_0 = _mm_maddubs_epi16(q3l_0, _mm256_extractf128_si256(q8_0, 0));
-        __m128i p16_1 = _mm_maddubs_epi16(q3l_1, _mm256_extractf128_si256(q8_0, 1));
-        __m128i p16_2 = _mm_maddubs_epi16(q3l_2, _mm256_extractf128_si256(q8_1, 0));
-        __m128i p16_3 = _mm_maddubs_epi16(q3l_3, _mm256_extractf128_si256(q8_1, 1));
-
-        p16_0 = _mm_sub_epi16(p16_0, q8s_0);
-        p16_1 = _mm_sub_epi16(p16_1, q8s_1);
-        p16_2 = _mm_sub_epi16(p16_2, q8s_2);
-        p16_3 = _mm_sub_epi16(p16_3, q8s_3);
-
-        // multiply with scales
-        p16_0 = _mm_madd_epi16(scale_0, p16_0);
-        p16_1 = _mm_madd_epi16(scale_1, p16_1);
-        p16_2 = _mm_madd_epi16(scale_2, p16_2);
-        p16_3 = _mm_madd_epi16(scale_3, p16_3);
-
-        p16_0 = _mm_add_epi32(p16_0, p16_2);
-        p16_1 = _mm_add_epi32(p16_1, p16_3);
-        __m256i p16 = MM256_SET_M128I(p16_1, p16_0);
-
-        // multiply with block scale and accumulate
-        acc = _mm256_add_ps(_mm256_mul_ps(_mm256_broadcast_ss(&d), _mm256_cvtepi32_ps(p16)), acc);
-
-    }
-
-    *s = hsum_float_8(acc);
-
-#elif defined __riscv_v_intrinsic
-
-    uint16_t aux16[2];
-    int8_t * scales = (int8_t *)aux16;
-
-    float sumf = 0;
-
-    for (int i = 0; i < nb; ++i) {
-
-        const uint8_t * restrict q3 = x[i].qs;
-        const int8_t  * restrict q8 = y[i].qs;
-
-        const uint16_t a = *(const uint16_t *)x[i].scales;
-        aux16[0] = a & 0x0f0f;
-        aux16[1] = (a >> 4) & 0x0f0f;
-
-        for (int j = 0; j < 4; ++j) scales[j] -= 8;
-
-        int32_t isum = -4*(scales[0] * y[i].bsums[0] + scales[2] * y[i].bsums[1] + scales[1] * y[i].bsums[2] + scales[3] * y[i].bsums[3]);
-
-        const float d = y[i].d * (float)x[i].d;
-
-        vint32m1_t vzero = __riscv_vmv_v_x_i32m1(0, 1);
-
-        // load qh
-        vuint8mf4_t qh_x1   = __riscv_vle8_v_u8mf4(x[i].hmask, 8);
-        vuint8mf2_t qh_x2   = __riscv_vlmul_ext_v_u8mf4_u8mf2(__riscv_vsrl_vx_u8mf4(qh_x1, 1, 8));
-
-        size_t vl = 16;
-
-        // extend and combine both qh_x1 and qh_x2
-        vuint8mf2_t qh_x = __riscv_vslideup_vx_u8mf2(__riscv_vlmul_ext_v_u8mf4_u8mf2(qh_x1), qh_x2, vl/2, vl);
-
-        vuint8mf2_t qh_0 = __riscv_vand_vx_u8mf2(__riscv_vsll_vx_u8mf2(qh_x, 0x2, vl), 0x4, vl);
-        vuint8mf2_t qh_1 = __riscv_vand_vx_u8mf2(qh_x, 0x4, vl);
-        vuint8mf2_t qh_2 = __riscv_vand_vx_u8mf2(__riscv_vsrl_vx_u8mf2(qh_x, 0x2, vl), 0x4, vl);
-        vuint8mf2_t qh_3 = __riscv_vand_vx_u8mf2(__riscv_vsrl_vx_u8mf2(qh_x, 0x4, vl), 0x4, vl);
-
-        // load Q3
-        vuint8mf2_t q3_x  = __riscv_vle8_v_u8mf2(q3, vl);
-
-        vuint8mf2_t q3h_0 = __riscv_vor_vv_u8mf2(__riscv_vand_vx_u8mf2(q3_x, 0x3, vl), qh_0, vl);
-        vuint8mf2_t q3h_1 = __riscv_vor_vv_u8mf2(__riscv_vand_vx_u8mf2(__riscv_vsrl_vx_u8mf2(q3_x, 2, vl), 0x3, vl), qh_1, vl);
-        vuint8mf2_t q3h_2 = __riscv_vor_vv_u8mf2(__riscv_vand_vx_u8mf2(__riscv_vsrl_vx_u8mf2(q3_x, 4, vl), 0x3, vl), qh_2, vl);
-        vuint8mf2_t q3h_3 = __riscv_vor_vv_u8mf2(__riscv_vsrl_vx_u8mf2(q3_x, 0x6, vl), qh_3, vl);
-
-        vint8mf2_t q3_0 = __riscv_vreinterpret_v_u8mf2_i8mf2(q3h_0);
-        vint8mf2_t q3_1 = __riscv_vreinterpret_v_u8mf2_i8mf2(q3h_1);
-        vint8mf2_t q3_2 = __riscv_vreinterpret_v_u8mf2_i8mf2(q3h_2);
-        vint8mf2_t q3_3 = __riscv_vreinterpret_v_u8mf2_i8mf2(q3h_3);
-
-        // load Q8 and take product with Q3
-        vint16m1_t p0 = __riscv_vwmul_vv_i16m1(q3_0, __riscv_vle8_v_i8mf2(q8, vl), vl);
-        vint16m1_t p1 = __riscv_vwmul_vv_i16m1(q3_1, __riscv_vle8_v_i8mf2(q8+16, vl), vl);
-        vint16m1_t p2 = __riscv_vwmul_vv_i16m1(q3_2, __riscv_vle8_v_i8mf2(q8+32, vl), vl);
-        vint16m1_t p3 = __riscv_vwmul_vv_i16m1(q3_3, __riscv_vle8_v_i8mf2(q8+48, vl), vl);
-
-        vint32m1_t vs_0 = __riscv_vwredsum_vs_i16m1_i32m1(p0, vzero, vl);
-        vint32m1_t vs_1 = __riscv_vwredsum_vs_i16m1_i32m1(p1, vzero, vl);
-        vint32m1_t vs_2 = __riscv_vwredsum_vs_i16m1_i32m1(p2, vzero, vl);
-        vint32m1_t vs_3 = __riscv_vwredsum_vs_i16m1_i32m1(p3, vzero, vl);
-
-        isum += __riscv_vmv_x_s_i32m1_i32(vs_0) * scales[0];
-        isum += __riscv_vmv_x_s_i32m1_i32(vs_1) * scales[2];
-        isum += __riscv_vmv_x_s_i32m1_i32(vs_2) * scales[1];
-        isum += __riscv_vmv_x_s_i32m1_i32(vs_3) * scales[3];
-
-        sumf += d * isum;
-
-    }
-
-    *s = sumf;
-
-#else
-
-    int8_t  aux8[QK_K];
-    int16_t aux16[8];
-    float   sums [8];
-    int32_t aux32[8];
-    int32_t scales[4];
-    memset(sums, 0, 8*sizeof(float));
-
-    float sumf = 0;
-    for (int i = 0; i < nb; ++i) {
-        const uint8_t * restrict q3 = x[i].qs;
-        const uint8_t * restrict hm = x[i].hmask;
-        const  int8_t * restrict q8 = y[i].qs;
-        int8_t * restrict a = aux8;
-        for (int l = 0; l < 8; ++l) {
-            a[l+ 0] = (int8_t)((q3[l+0] >> 0) & 3) - (hm[l] & 0x01 ? 0 : 4);
-            a[l+ 8] = (int8_t)((q3[l+8] >> 0) & 3) - (hm[l] & 0x02 ? 0 : 4);
-            a[l+16] = (int8_t)((q3[l+0] >> 2) & 3) - (hm[l] & 0x04 ? 0 : 4);
-            a[l+24] = (int8_t)((q3[l+8] >> 2) & 3) - (hm[l] & 0x08 ? 0 : 4);
-            a[l+32] = (int8_t)((q3[l+0] >> 4) & 3) - (hm[l] & 0x10 ? 0 : 4);
-            a[l+40] = (int8_t)((q3[l+8] >> 4) & 3) - (hm[l] & 0x20 ? 0 : 4);
-            a[l+48] = (int8_t)((q3[l+0] >> 6) & 3) - (hm[l] & 0x40 ? 0 : 4);
-            a[l+56] = (int8_t)((q3[l+8] >> 6) & 3) - (hm[l] & 0x80 ? 0 : 4);
-        }
-
-        scales[0] = (x[i].scales[0] & 0xF) - 8;
-        scales[1] = (x[i].scales[0] >>  4) - 8;
-        scales[2] = (x[i].scales[1] & 0xF) - 8;
-        scales[3] = (x[i].scales[1] >>  4) - 8;
-
-        memset(aux32, 0, 8*sizeof(int32_t));
-        for (int j = 0; j < QK_K/16; ++j) {
-            for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
-            q8 += 8; a += 8;
-            for (int l = 0; l < 8; ++l) aux16[l] += q8[l] * a[l];
-            q8 += 8; a += 8;
-            for (int l = 0; l < 8; ++l) aux32[l] += scales[j] * aux16[l];
-        }
-        const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
-        for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l];
-    }
-    for (int l = 0; l < 8; ++l) sumf += sums[l];
-    *s = sumf;
-
-#endif
-
-}
-#endif
-
-#if QK_K == 256
-void ggml_vec_dot_q4_K_q8_K(const int n, float * restrict s, const void * restrict vx, const void * restrict vy) {
-    assert(n % QK_K == 0);
-
-    const block_q4_K * restrict x = vx;
-    const block_q8_K * restrict y = vy;
-
-    const int nb = n / QK_K;
-
-    static const uint32_t kmask1 = 0x3f3f3f3f;
-    static const uint32_t kmask2 = 0x0f0f0f0f;
-    static const uint32_t kmask3 = 0x03030303;
-
-    uint32_t utmp[4];
-
-#ifdef __ARM_NEON
-    const uint8x16_t m4b = vdupq_n_u8(0xf);
-    const int32x4_t mzero = vdupq_n_s32(0);
-
-    ggml_int8x16x2_t q4bytes;
-    ggml_int8x16x2_t q8bytes;
-
-    float sumf = 0;
-
-    for (int i = 0; i < nb; ++i) {
-
-        const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d);
-        const float dmin = y[i].d * GGML_FP16_TO_FP32(x[i].dmin);
-
-        const int16x8_t q8sums = vpaddq_s16(vld1q_s16(y[i].bsums), vld1q_s16(y[i].bsums + 8));
-
-        memcpy(utmp, x[i].scales, 12);
-
-        uint32x2_t mins8 = { 0 };
-        mins8 = vset_lane_u32(utmp[1] & kmask1, mins8, 0);
-        mins8 = vset_lane_u32(((utmp[2] >> 4) & kmask2) | (((utmp[1] >> 6) & kmask3) << 4), mins8, 1);
-
-        utmp[1] = (utmp[2] & kmask2) | (((utmp[0] >> 6) & kmask3) << 4);
-        utmp[0] &= kmask1;
-
-        const int16x8_t mins = vreinterpretq_s16_u16(vmovl_u8(vreinterpret_u8_u32(mins8)));
-        const int32x4_t prod = vaddq_s32(vmull_s16(vget_low_s16 (q8sums), vget_low_s16 (mins)),
-                                         vmull_s16(vget_high_s16(q8sums), vget_high_s16(mins)));
-        sumf -= dmin * vaddvq_s32(prod);
-
-        const uint8_t * scales = (const uint8_t *)utmp;
-
-        const uint8_t * restrict q4 = x[i].qs;
-        const int8_t  * restrict q8 = y[i].qs;
-
-        int32_t sumi1 = 0;
-        int32_t sumi2 = 0;
-
-        for (int j = 0; j < QK_K/64; ++j) {
-            const ggml_uint8x16x2_t q4bits = ggml_vld1q_u8_x2(q4); q4 += 32;
-
-            q8bytes = ggml_vld1q_s8_x2(q8); q8 += 32;
-            q4bytes.val[0] = vreinterpretq_s8_u8(vandq_u8  (q4bits.val[0], m4b));
-            q4bytes.val[1] = vreinterpretq_s8_u8(vandq_u8  (q4bits.val[1], m4b));
-
-            const int32x4_t p1 = ggml_vdotq_s32(ggml_vdotq_s32(mzero, q4bytes.val[0], q8bytes.val[0]), q4bytes.val[1], q8bytes.val[1]);
-            sumi1 += vaddvq_s32(p1) * scales[2*j+0];
-
-            q8bytes = ggml_vld1q_s8_x2(q8); q8 += 32;
-            q4bytes.val[0] = vreinterpretq_s8_u8(vshrq_n_u8(q4bits.val[0], 4));
-            q4bytes.val[1] = vreinterpretq_s8_u8(vshrq_n_u8(q4bits.val[1], 4));
-
-            const int32x4_t p2 = ggml_vdotq_s32(ggml_vdotq_s32(mzero, q4bytes.val[0], q8bytes.val[0]), q4bytes.val[1], q8bytes.val[1]);
-
-            sumi2 += vaddvq_s32(p2) * scales[2*j+1];
-        }
-
-        sumf += d * (sumi1 + sumi2);
-
-    }
-
-    *s = sumf;
-
-#elif defined __AVX2__
-
-    const __m256i m4 = _mm256_set1_epi8(0xF);
-
-    __m256 acc = _mm256_setzero_ps();
-    __m128 acc_m = _mm_setzero_ps();
-
-   for (int i = 0; i < nb; ++i) {
-
-        const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d);
-        const float dmin = -y[i].d * GGML_FP16_TO_FP32(x[i].dmin);
-
-        memcpy(utmp, x[i].scales, 12);
-        utmp[3] = ((utmp[2] >> 4) & kmask2) | (((utmp[1] >> 6) & kmask3) << 4);
-        const uint32_t uaux = utmp[1] & kmask1;
-        utmp[1] = (utmp[2] & kmask2) | (((utmp[0] >> 6) & kmask3) << 4);
-        utmp[2] = uaux;
-        utmp[0] &= kmask1;
-
-        const uint8_t * restrict q4 = x[i].qs;
-        const int8_t  * restrict q8 = y[i].qs;
-
-        const __m256i mins_and_scales = _mm256_cvtepu8_epi16(_mm_set_epi32(utmp[3], utmp[2], utmp[1], utmp[0]));
-
-        const __m256i q8sums = _mm256_loadu_si256((const __m256i*)y[i].bsums);
-        const __m128i q8s = _mm_hadd_epi16(_mm256_extracti128_si256(q8sums, 0), _mm256_extracti128_si256(q8sums, 1));
-        const __m128i prod = _mm_madd_epi16(_mm256_extracti128_si256(mins_and_scales, 1), q8s);
-        acc_m = _mm_fmadd_ps(_mm_set1_ps(dmin), _mm_cvtepi32_ps(prod), acc_m);
-
-        const __m128i sc128  = _mm256_extracti128_si256(mins_and_scales, 0);
-        const __m256i scales = MM256_SET_M128I(sc128, sc128);
-
-        __m256i sumi = _mm256_setzero_si256();
-
-        for (int j = 0; j < QK_K/64; ++j) {
-
-            const __m256i scale_l = _mm256_shuffle_epi8(scales, get_scale_shuffle_k4(2*j+0));
-            const __m256i scale_h = _mm256_shuffle_epi8(scales, get_scale_shuffle_k4(2*j+1));
-
-            const __m256i q4bits = _mm256_loadu_si256((const __m256i*)q4); q4 += 32;
-            const __m256i q4l = _mm256_and_si256(q4bits, m4);
-            const __m256i q4h = _mm256_and_si256(_mm256_srli_epi16(q4bits, 4), m4);
-
-            const __m256i q8l = _mm256_loadu_si256((const __m256i*)q8); q8 += 32;
-            __m256i p16l = _mm256_maddubs_epi16(q4l, q8l);
-            p16l = _mm256_madd_epi16(scale_l, p16l);
-
-            const __m256i q8h = _mm256_loadu_si256((const __m256i*)q8); q8 += 32;
-            __m256i p16h = _mm256_maddubs_epi16(q4h, q8h);
-            p16h = _mm256_madd_epi16(scale_h, p16h);
-            const __m256i sumj = _mm256_add_epi32(p16l, p16h);
-
-            sumi = _mm256_add_epi32(sumi, sumj);
-        }
-
-        __m256 vd = _mm256_set1_ps(d);
-        acc = _mm256_fmadd_ps(vd, _mm256_cvtepi32_ps(sumi), acc);
-
-    }
-
-    acc_m = _mm_add_ps(acc_m, _mm_movehl_ps(acc_m, acc_m));
-    acc_m = _mm_add_ss(acc_m, _mm_movehdup_ps(acc_m));
-
-    *s = hsum_float_8(acc) + _mm_cvtss_f32(acc_m);
-
-#elif defined __AVX__
-
-    const __m128i m4 = _mm_set1_epi8(0xF);
-    const __m128i m2 = _mm_set1_epi8(0x2);
-
-    __m256 acc = _mm256_setzero_ps();
-    __m128 acc_m = _mm_setzero_ps();
-
-   for (int i = 0; i < nb; ++i) {
-
-        const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d);
-        const float dmin = -y[i].d * GGML_FP16_TO_FP32(x[i].dmin);
-
-        const uint8_t * restrict q4 = x[i].qs;
-        const int8_t  * restrict q8 = y[i].qs;
-
-        memcpy(utmp, x[i].scales, 12);
-        utmp[3] = ((utmp[2] >> 4) & kmask2) | (((utmp[1] >> 6) & kmask3) << 4);
-        const uint32_t uaux = utmp[1] & kmask1;
-        utmp[1] = (utmp[2] & kmask2) | (((utmp[0] >> 6) & kmask3) << 4);
-        utmp[2] = uaux;
-        utmp[0] &= kmask1;
-
-        const __m128i utmps = _mm_set_epi32(utmp[3], utmp[2], utmp[1], utmp[0]);
-        const __m128i scales = _mm_cvtepu8_epi16(utmps);
-        const __m128i mins = _mm_cvtepu8_epi16(_mm_unpackhi_epi64(utmps, utmps));
-
-        const __m128i q8sums_0 = _mm_loadu_si128((const __m128i*)&y[i].bsums[0]);
-        const __m128i q8sums_1 = _mm_loadu_si128((const __m128i*)&y[i].bsums[8]);
-        const __m128i q8s = _mm_hadd_epi16(q8sums_0, q8sums_1);
-        const __m128i prod = _mm_madd_epi16(mins, q8s);
-        acc_m = _mm_add_ps(_mm_mul_ps(_mm_set1_ps(dmin), _mm_cvtepi32_ps(prod)), acc_m);
-
-        __m128i sumi_0 = _mm_setzero_si128();
-        __m128i sumi_1 = _mm_setzero_si128();
-
-        __m128i shuffle = _mm_set1_epi16(0x0100);
-        for (int j = 0; j < QK_K/64; ++j) {
-
-            const __m128i scale_l = _mm_shuffle_epi8(scales, shuffle);
-            shuffle = _mm_add_epi16(shuffle, m2);
-            const __m128i scale_h = _mm_shuffle_epi8(scales, shuffle);
-            shuffle = _mm_add_epi16(shuffle, m2);
-
-            __m128i q4bits = _mm_loadu_si128((const __m128i*)q4); q4 += 16;
-            const __m128i q4l_0 = _mm_and_si128(q4bits, m4);
-            const __m128i q4h_0 = _mm_and_si128(_mm_srli_epi16(q4bits, 4), m4);
-            q4bits = _mm_loadu_si128((const __m128i*)q4); q4 += 16;
-            const __m128i q4l_1 = _mm_and_si128(q4bits, m4);
-            const __m128i q4h_1 = _mm_and_si128(_mm_srli_epi16(q4bits, 4), m4);
-
-            const __m128i q8l_0 = _mm_loadu_si128((const __m128i*)q8); q8 += 16;
-            __m128i p16l = _mm_maddubs_epi16(q4l_0, q8l_0);
-            p16l = _mm_madd_epi16(scale_l, p16l);
-            sumi_0 = _mm_add_epi32(sumi_0, p16l);
-            const __m128i q8l_1 = _mm_loadu_si128((const __m128i*)q8); q8 += 16;
-            p16l = _mm_maddubs_epi16(q4l_1, q8l_1);
-            p16l = _mm_madd_epi16(scale_l, p16l);
-            sumi_1 = _mm_add_epi32(sumi_1, p16l);
-
-            const __m128i q8h_0 = _mm_loadu_si128((const __m128i*)q8); q8 += 16;
-            __m128i p16h = _mm_maddubs_epi16(q4h_0, q8h_0);
-            p16h = _mm_madd_epi16(scale_h, p16h);
-            sumi_0 = _mm_add_epi32(sumi_0, p16h);
-            const __m128i q8h_1 = _mm_loadu_si128((const __m128i*)q8); q8 += 16;
-            p16h = _mm_maddubs_epi16(q4h_1, q8h_1);
-            p16h = _mm_madd_epi16(scale_h, p16h);
-            sumi_1 = _mm_add_epi32(sumi_1, p16h);
-
-        }
-
-        __m256 vd = _mm256_set1_ps(d);
-        __m256i sumi = MM256_SET_M128I(sumi_1, sumi_0);
-        acc = _mm256_add_ps(_mm256_mul_ps(vd, _mm256_cvtepi32_ps(sumi)), acc);
-
-    }
-
-    acc_m = _mm_add_ps(acc_m, _mm_movehl_ps(acc_m, acc_m));
-    acc_m = _mm_add_ss(acc_m, _mm_movehdup_ps(acc_m));
-
-    *s = hsum_float_8(acc) + _mm_cvtss_f32(acc_m);
-
-#elif defined __riscv_v_intrinsic
-
-    const uint8_t * scales = (const uint8_t*)&utmp[0];
-    const uint8_t * mins   = (const uint8_t*)&utmp[2];
-
-    float sumf = 0;
-
-    for (int i = 0; i < nb; ++i) {
-
-        size_t vl = 8;
-
-        const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d);
-        const float dmin = y[i].d * GGML_FP16_TO_FP32(x[i].dmin);
-
-        vint16mf2_t q8sums_0 = __riscv_vlse16_v_i16mf2(y[i].bsums, 4, vl);
-        vint16mf2_t q8sums_1 = __riscv_vlse16_v_i16mf2(y[i].bsums+1, 4, vl);
-        vint16mf2_t q8sums   = __riscv_vadd_vv_i16mf2(q8sums_0, q8sums_1, vl);
-
-        memcpy(utmp, x[i].scales, 12);
-        utmp[3] = ((utmp[2] >> 4) & kmask2) | (((utmp[1] >> 6) & kmask3) << 4);
-        const uint32_t uaux = utmp[1] & kmask1;
-        utmp[1] = (utmp[2] & kmask2) | (((utmp[0] >> 6) & kmask3) << 4);
-        utmp[2] = uaux;
-        utmp[0] &= kmask1;
-
-        vuint8mf4_t mins8  = __riscv_vle8_v_u8mf4(mins, vl);
-        vint16mf2_t v_mins = __riscv_vreinterpret_v_u16mf2_i16mf2(__riscv_vzext_vf2_u16mf2(mins8, vl));
-        vint32m1_t  prod   = __riscv_vwmul_vv_i32m1(q8sums, v_mins, vl);
-
-        vint32m1_t sumi = __riscv_vredsum_vs_i32m1_i32m1(prod, __riscv_vmv_v_x_i32m1(0, 1), vl);
-        sumf -= dmin * __riscv_vmv_x_s_i32m1_i32(sumi);
-
-        const uint8_t * restrict q4 = x[i].qs;
-        const int8_t  * restrict q8 = y[i].qs;
-
-        vl = 32;
-
-        int32_t sum_1 = 0;
-        int32_t sum_2 = 0;
-
-        vint16m1_t vzero = __riscv_vmv_v_x_i16m1(0, 1);
-
-        for (int j = 0; j < QK_K/64; ++j) {
-            // load Q4
-            vuint8m1_t q4_x = __riscv_vle8_v_u8m1(q4, vl);
-
-            // load Q8 and multiply it with lower Q4 nibble
-            vint8m1_t  q8_0 = __riscv_vle8_v_i8m1(q8, vl);
-            vint8m1_t  q4_0 = __riscv_vreinterpret_v_u8m1_i8m1(__riscv_vand_vx_u8m1(q4_x, 0x0F, vl));
-            vint16m2_t qv_0 = __riscv_vwmul_vv_i16m2(q4_0, q8_0, vl);
-            vint16m1_t vs_0 = __riscv_vredsum_vs_i16m2_i16m1(qv_0, vzero, vl);
-
-            sum_1 += __riscv_vmv_x_s_i16m1_i16(vs_0) * scales[2*j+0];
-
-            // load Q8 and multiply it with upper Q4 nibble
-            vint8m1_t  q8_1 = __riscv_vle8_v_i8m1(q8+32, vl);
-            vint8m1_t  q4_1 = __riscv_vreinterpret_v_u8m1_i8m1(__riscv_vsrl_vx_u8m1(q4_x, 0x04, vl));
-            vint16m2_t qv_1 = __riscv_vwmul_vv_i16m2(q4_1, q8_1, vl);
-            vint16m1_t vs_1 = __riscv_vredsum_vs_i16m2_i16m1(qv_1, vzero, vl);
-
-            sum_2 += __riscv_vmv_x_s_i16m1_i16(vs_1) * scales[2*j+1];
-
-            q4 += 32;    q8 += 64;
-
-        }
-
-        sumf += d*(sum_1 + sum_2);
-
-    }
-
-    *s = sumf;
-
-#else
-
-
-    const uint8_t * scales = (const uint8_t*)&utmp[0];
-    const uint8_t * mins   = (const uint8_t*)&utmp[2];
-
-    int8_t  aux8[QK_K];
-    int16_t aux16[8];
-    float   sums [8];
-    int32_t aux32[8];
-    memset(sums, 0, 8*sizeof(float));
-
-    float sumf = 0;
-    for (int i = 0; i < nb; ++i) {
-        const uint8_t * restrict q4 = x[i].qs;
-        const  int8_t * restrict q8 = y[i].qs;
-        memset(aux32, 0, 8*sizeof(int32_t));
-        int8_t * restrict a = aux8;
-        for (int j = 0; j < QK_K/64; ++j) {
-            for (int l = 0; l < 32; ++l) a[l] = (int8_t)(q4[l] & 0xF);
-            a += 32;
-            for (int l = 0; l < 32; ++l) a[l] = (int8_t)(q4[l]  >> 4);
-            a += 32; q4 += 32;
-        }
-        memcpy(utmp, x[i].scales, 12);
-        utmp[3] = ((utmp[2] >> 4) & kmask2) | (((utmp[1] >> 6) & kmask3) << 4);
-        const uint32_t uaux = utmp[1] & kmask1;
-        utmp[1] = (utmp[2] & kmask2) | (((utmp[0] >> 6) & kmask3) << 4);
-        utmp[2] = uaux;
-        utmp[0] &= kmask1;
-
-        int sumi = 0;
-        for (int j = 0; j < QK_K/16; ++j) sumi += y[i].bsums[j] * mins[j/2];
-        a = aux8;
-        int is = 0;
-        for (int j = 0; j < QK_K/32; ++j) {
-            int32_t scale = scales[is++];
-            for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
-            for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
-            q8 += 8; a += 8;
-            for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
-            for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
-            q8 += 8; a += 8;
-            for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
-            for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
-            q8 += 8; a += 8;
-            for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
-            for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
-            q8 += 8; a += 8;
-        }
-        const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
-        for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l];
-        const float dmin = GGML_FP16_TO_FP32(x[i].dmin) * y[i].d;
-        sumf -= dmin * sumi;
-    }
-    for (int l = 0; l < 8; ++l) sumf += sums[l];
-    *s = sumf;
-#endif
-}
-#else
-void ggml_vec_dot_q4_K_q8_K(const int n, float * restrict s, const void * restrict vx, const void * restrict vy) {
-    assert(n % QK_K == 0);
-
-    const block_q4_K * restrict x = vx;
-    const block_q8_K * restrict y = vy;
-
-    const int nb = n / QK_K;
-
-#ifdef __ARM_NEON
-    const uint8x16_t m4b = vdupq_n_u8(0xf);
-
-    const int32x4_t mzero = vdupq_n_s32(0);
-
-    float sumf = 0;
-
-    ggml_int8x16x2_t q4bytes;
-    ggml_int8x16x4_t q8bytes;
-
-    float sum_mins = 0.f;
-
-    uint16_t aux16[2];
-    const uint8_t * restrict scales = (const uint8_t *)aux16;
-
-    for (int i = 0; i < nb; ++i) {
-
-        const uint8_t * restrict q4 = x[i].qs;
-        const int8_t  * restrict q8 = y[i].qs;
-
-        const uint16_t * restrict a = (const uint16_t *)x[i].scales;
-        aux16[0] = a[0] & 0x0f0f;
-        aux16[1] = (a[0] >> 4) & 0x0f0f;
-
-        const int32_t summi = scales[2] * (y[i].bsums[0] + y[i].bsums[1]) + scales[3] * (y[i].bsums[2] + y[i].bsums[3]);
-        sum_mins += y[i].d * (float)x[i].d[1] * summi;
-
-        const float d = y[i].d * (float)x[i].d[0];
-
-        const ggml_uint8x16x2_t q4bits = ggml_vld1q_u8_x2(q4);
-
-        q8bytes = ggml_vld1q_s8_x4(q8);
-        q4bytes.val[0] = vreinterpretq_s8_u8(vandq_u8  (q4bits.val[0], m4b));
-        q4bytes.val[1] = vreinterpretq_s8_u8(vandq_u8  (q4bits.val[1], m4b));
-
-        const int32x4_t p1 = ggml_vdotq_s32(ggml_vdotq_s32(mzero, q4bytes.val[0], q8bytes.val[0]), q4bytes.val[1], q8bytes.val[1]);
-        const int32_t sumi1 = vaddvq_s32(p1) * scales[0];
-
-        q4bytes.val[0] = vreinterpretq_s8_u8(vshrq_n_u8(q4bits.val[0], 4));
-        q4bytes.val[1] = vreinterpretq_s8_u8(vshrq_n_u8(q4bits.val[1], 4));
-
-        const int32x4_t p2 = ggml_vdotq_s32(ggml_vdotq_s32(mzero, q4bytes.val[0], q8bytes.val[2]), q4bytes.val[1], q8bytes.val[3]);
-        const int32_t sumi2 = vaddvq_s32(p2) * scales[1];
-
-        sumf += d * (sumi1 + sumi2);
-    }
-
-    *s = sumf - sum_mins;
-
-#elif defined __AVX2__
-
-    const __m256i m4 = _mm256_set1_epi8(0xF);
-
-    __m256 acc = _mm256_setzero_ps();
-
-    float summs = 0;
-
-    uint16_t aux16[2];
-    const uint8_t * scales = (const uint8_t *)aux16;
-
-    for (int i = 0; i < nb; ++i) {
-
-        const float d = GGML_FP16_TO_FP32(x[i].d[0]) * y[i].d;
-        const float m = GGML_FP16_TO_FP32(x[i].d[1]) * y[i].d;
-        const __m256 vd = _mm256_set1_ps(d);
-
-        const uint16_t * a = (const uint16_t *)x[i].scales;
-        aux16[0] = a[0] & 0x0f0f;
-        aux16[1] = (a[0] >> 4) & 0x0f0f;
-
-        summs += m * (scales[2] * (y[i].bsums[0] + y[i].bsums[1]) + scales[3] * (y[i].bsums[2] + y[i].bsums[3]));
-
-        const uint8_t * restrict q4 = x[i].qs;
-        const int8_t  * restrict q8 = y[i].qs;
-
-        const __m256i q4bits = _mm256_loadu_si256((const __m256i*)q4);
-        const __m256i q4l = _mm256_and_si256(q4bits, m4);
-        const __m256i q4h = _mm256_and_si256(_mm256_srli_epi16(q4bits, 4), m4);
-
-        const __m256i q8l = _mm256_loadu_si256((const __m256i*)(q8+ 0));
-        const __m256i q8h = _mm256_loadu_si256((const __m256i*)(q8+32));
-
-        const __m256i p16l = _mm256_maddubs_epi16(q4l, q8l);
-        const __m256i p16h = _mm256_maddubs_epi16(q4h, q8h);
-
-        const __m256i p32l = _mm256_madd_epi16(_mm256_set1_epi16(scales[0]), p16l);
-        acc = _mm256_fmadd_ps(vd, _mm256_cvtepi32_ps(p32l), acc);
-
-        const __m256i p32h = _mm256_madd_epi16(_mm256_set1_epi16(scales[1]), p16h);
-        acc = _mm256_fmadd_ps(vd, _mm256_cvtepi32_ps(p32h), acc);
-
-    }
-
-    *s = hsum_float_8(acc) - summs;
-
-#elif defined __AVX__
-
-    const __m128i m4 = _mm_set1_epi8(0xF);
-
-    __m256 acc = _mm256_setzero_ps();
-
-    float summs = 0;
-
-    uint16_t aux16[2];
-    const uint8_t * scales = (const uint8_t *)aux16;
-
-    for (int i = 0; i < nb; ++i) {
-
-        const float d = GGML_FP16_TO_FP32(x[i].d[0]) * y[i].d;
-        const float m = GGML_FP16_TO_FP32(x[i].d[1]) * y[i].d;
-        const __m256 vd = _mm256_set1_ps(d);
-
-        const uint16_t * a = (const uint16_t *)x[i].scales;
-        aux16[0] = a[0] & 0x0f0f;
-        aux16[1] = (a[0] >> 4) & 0x0f0f;
-
-        summs += m * (scales[2] * (y[i].bsums[0] + y[i].bsums[1]) + scales[3] * (y[i].bsums[2] + y[i].bsums[3]));
-
-        const uint8_t * restrict q4 = x[i].qs;
-        const int8_t  * restrict q8 = y[i].qs;
-
-        const __m256i q4bits = _mm256_loadu_si256((const __m256i*)q4);
-        const __m128i q4bits_0 = _mm256_extractf128_si256(q4bits, 0);
-        const __m128i q4bits_1 = _mm256_extractf128_si256(q4bits, 1);
-        const __m128i q4_0 = _mm_and_si128(q4bits_0, m4);
-        const __m128i q4_1 = _mm_and_si128(q4bits_1, m4);
-        const __m128i q4_2 = _mm_and_si128(_mm_srli_epi16(q4bits_0, 4), m4);
-        const __m128i q4_3 = _mm_and_si128(_mm_srli_epi16(q4bits_1, 4), m4);
-
-        const __m256i q8_0 = _mm256_loadu_si256((const __m256i*)(q8+ 0));
-        const __m256i q8_1 = _mm256_loadu_si256((const __m256i*)(q8+32));
-
-        const __m128i p16_0 = _mm_maddubs_epi16(q4_0, _mm256_extractf128_si256(q8_0, 0));
-        const __m128i p16_1 = _mm_maddubs_epi16(q4_1, _mm256_extractf128_si256(q8_0, 1));
-        const __m128i p16_2 = _mm_maddubs_epi16(q4_2, _mm256_extractf128_si256(q8_1, 0));
-        const __m128i p16_3 = _mm_maddubs_epi16(q4_3, _mm256_extractf128_si256(q8_1, 1));
-
-        const __m128i p32_0 = _mm_madd_epi16(_mm_set1_epi16(scales[0]), p16_0);
-        const __m128i p32_1 = _mm_madd_epi16(_mm_set1_epi16(scales[0]), p16_1);
-        acc = _mm256_add_ps(_mm256_mul_ps(vd, _mm256_cvtepi32_ps(MM256_SET_M128I(p32_1, p32_0))), acc);
-
-        const __m128i p32_2 = _mm_madd_epi16(_mm_set1_epi16(scales[1]), p16_2);
-        const __m128i p32_3 = _mm_madd_epi16(_mm_set1_epi16(scales[1]), p16_3);
-        acc = _mm256_add_ps(_mm256_mul_ps(vd, _mm256_cvtepi32_ps(MM256_SET_M128I(p32_3, p32_2))), acc);
-
-    }
-
-    *s = hsum_float_8(acc) - summs;
-
-#elif defined __riscv_v_intrinsic
-
-    uint16_t s16[2];
-    const uint8_t * restrict scales = (const uint8_t *)s16;
-
-    float sumf = 0;
-
-    for (int i = 0; i < nb; ++i) {
-
-        const uint8_t * restrict q4 = x[i].qs;
-        const  int8_t * restrict q8 = y[i].qs;
-
-        const uint16_t * restrict b = (const uint16_t *)x[i].scales;
-        s16[0] = b[0] & 0x0f0f;
-        s16[1] = (b[0] >> 4) & 0x0f0f;
-
-        sumf -= y[i].d * GGML_FP16_TO_FP32(x[i].d[1]) * (scales[2] * (y[i].bsums[0] + y[i].bsums[1]) + scales[3] * (y[i].bsums[2] + y[i].bsums[3]));
-        const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d[0]);
-
-        size_t vl = 32;
-
-        vint16m1_t vzero = __riscv_vmv_v_x_i16m1(0, 1);
-
-        // load Q4
-        vuint8m1_t q4_x = __riscv_vle8_v_u8m1(q4, vl);
-
-        // load Q8 and multiply it with lower Q4 nibble
-        vint8m1_t  q4_a = __riscv_vreinterpret_v_u8m1_i8m1(__riscv_vand_vx_u8m1(q4_x, 0x0F, vl));
-        vint16m2_t va_0 = __riscv_vwmul_vv_i16m2(q4_a, __riscv_vle8_v_i8m1(q8, vl), vl);
-        vint16m1_t aux1 = __riscv_vredsum_vs_i16m2_i16m1(va_0, vzero, vl);
-
-        sumf += d*scales[0]*__riscv_vmv_x_s_i16m1_i16(aux1);
-
-        // load Q8 and multiply it with upper Q4 nibble
-        vint8m1_t  q4_s = __riscv_vreinterpret_v_u8m1_i8m1(__riscv_vsrl_vx_u8m1(q4_x, 0x04, vl));
-        vint16m2_t va_1 = __riscv_vwmul_vv_i16m2(q4_s, __riscv_vle8_v_i8m1(q8+32, vl), vl);
-        vint16m1_t aux2 = __riscv_vredsum_vs_i16m2_i16m1(va_1, vzero, vl);
-
-        sumf += d*scales[1]*__riscv_vmv_x_s_i16m1_i16(aux2);
-
-    }
-
-    *s = sumf;
-
-#else
-
-    uint8_t aux8[QK_K];
-    int16_t aux16[16];
-    float   sums [8];
-    memset(sums, 0, 8*sizeof(float));
-
-    uint16_t s16[2];
-    const uint8_t * restrict scales = (const uint8_t *)s16;
-
-    float sumf = 0;
-    for (int i = 0; i < nb; ++i) {
-        const uint8_t * restrict q4 = x[i].qs;
-        const  int8_t * restrict q8 = y[i].qs;
-        uint8_t * restrict a = aux8;
-        for (int l = 0; l < 32; ++l) a[l+ 0] = q4[l] & 0xF;
-        for (int l = 0; l < 32; ++l) a[l+32] = q4[l]  >> 4;
-
-        const uint16_t * restrict b = (const uint16_t *)x[i].scales;
-        s16[0] = b[0] & 0x0f0f;
-        s16[1] = (b[0] >> 4) & 0x0f0f;
-
-        sumf -= y[i].d * GGML_FP16_TO_FP32(x[i].d[1]) * (scales[2] * (y[i].bsums[0] + y[i].bsums[1]) + scales[3] * (y[i].bsums[2] + y[i].bsums[3]));
-
-        const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d[0]);
-
-        for (int j = 0; j < QK_K/32; ++j) {
-            for (int l = 0; l < 16; ++l) aux16[l] = q8[l] * a[l];
-            q8 += 16; a += 16;
-            for (int l = 0; l < 16; ++l) aux16[l] += q8[l] * a[l];
-            q8 += 16; a += 16;
-            const float dl = d * scales[j];
-            for (int l = 0; l < 8; ++l) sums[l] += dl * (aux16[l] + aux16[l+8]);
-        }
-    }
-    for (int l = 0; l < 8; ++l) sumf += sums[l];
-    *s = sumf;
-#endif
-}
-#endif
-
-#if QK_K == 256
-void ggml_vec_dot_q5_K_q8_K(const int n, float * restrict s, const void * restrict vx, const void * restrict vy) {
-    assert(n % QK_K == 0);
-
-    const block_q5_K * restrict x = vx;
-    const block_q8_K * restrict y = vy;
-
-    const int nb = n / QK_K;
-
-    static const uint32_t kmask1 = 0x3f3f3f3f;
-    static const uint32_t kmask2 = 0x0f0f0f0f;
-    static const uint32_t kmask3 = 0x03030303;
-
-    uint32_t utmp[4];
-
-#ifdef __ARM_NEON
-    const uint8x16_t m4b = vdupq_n_u8(0xf);
-    const uint8x16_t mone = vdupq_n_u8(1);
-    const uint8x16_t mtwo = vdupq_n_u8(2);
-    const int32x4_t mzero = vdupq_n_s32(0);
-
-    ggml_int8x16x4_t q5bytes;
-
-    float sumf = 0;
-
-    for (int i = 0; i < nb; ++i) {
-
-        const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d);
-        const float dmin = y[i].d * GGML_FP16_TO_FP32(x[i].dmin);
-
-        const int16x8_t q8sums = vpaddq_s16(vld1q_s16(y[i].bsums), vld1q_s16(y[i].bsums + 8));
-
-        memcpy(utmp, x[i].scales, 12);
-        utmp[3] = ((utmp[2] >> 4) & kmask2) | (((utmp[1] >> 6) & kmask3) << 4);
-        const uint32_t uaux = utmp[1] & kmask1;
-        utmp[1] = (utmp[2] & kmask2) | (((utmp[0] >> 6) & kmask3) << 4);
-        utmp[2] = uaux;
-        utmp[0] &= kmask1;
-
-        const uint8x8_t mins8 = vld1_u8((const uint8_t*)utmp + 8);
-        const int16x8_t mins = vreinterpretq_s16_u16(vmovl_u8(mins8));
-        const int32x4_t prod = vaddq_s32(vmull_s16(vget_low_s16 (q8sums), vget_low_s16 (mins)),
-                                         vmull_s16(vget_high_s16(q8sums), vget_high_s16(mins)));
-        int32_t sumi_mins = vaddvq_s32(prod);
-
-        const uint8_t * scales = (const uint8_t *)utmp;
-
-        const uint8_t * restrict q5 = x[i].qs;
-        const uint8_t * restrict qh = x[i].qh;
-        const int8_t  * restrict q8 = y[i].qs;
-
-        ggml_uint8x16x2_t qhbits = ggml_vld1q_u8_x2(qh);
-
-        ggml_uint8x16x4_t q5h;
-
-        int32_t sumi = 0;
-
-        for (int j = 0; j < QK_K/64; ++j) {
-
-            const ggml_uint8x16x2_t q5bits = ggml_vld1q_u8_x2(q5); q5 += 32;
-            const ggml_int8x16x4_t q8bytes = ggml_vld1q_s8_x4(q8); q8 += 64;
-
-            q5h.val[0] = vshlq_n_u8(vandq_u8(mone, qhbits.val[0]), 4);
-            q5h.val[1] = vshlq_n_u8(vandq_u8(mone, qhbits.val[1]), 4);
-            q5h.val[2] = vshlq_n_u8(vandq_u8(mtwo, qhbits.val[0]), 3);
-            q5h.val[3] = vshlq_n_u8(vandq_u8(mtwo, qhbits.val[1]), 3);
-            qhbits.val[0] = vshrq_n_u8(qhbits.val[0], 2);
-            qhbits.val[1] = vshrq_n_u8(qhbits.val[1], 2);
-
-            q5bytes.val[0] = vreinterpretq_s8_u8(vorrq_u8(vandq_u8(q5bits.val[0], m4b), q5h.val[0]));
-            q5bytes.val[1] = vreinterpretq_s8_u8(vorrq_u8(vandq_u8(q5bits.val[1], m4b), q5h.val[1]));
-            q5bytes.val[2] = vreinterpretq_s8_u8(vorrq_u8(vshrq_n_u8(q5bits.val[0], 4), q5h.val[2]));
-            q5bytes.val[3] = vreinterpretq_s8_u8(vorrq_u8(vshrq_n_u8(q5bits.val[1], 4), q5h.val[3]));
-
-            sumi += vaddvq_s32(ggml_vdotq_s32(ggml_vdotq_s32(mzero, q5bytes.val[0], q8bytes.val[0]), q5bytes.val[1], q8bytes.val[1])) * *scales++;
-            sumi += vaddvq_s32(ggml_vdotq_s32(ggml_vdotq_s32(mzero, q5bytes.val[2], q8bytes.val[2]), q5bytes.val[3], q8bytes.val[3])) * *scales++;
-        }
-
-        sumf += d * sumi - dmin * sumi_mins;
-    }
-
-    *s = sumf;
-
-#elif defined __AVX2__
-
-    const __m256i m4 = _mm256_set1_epi8(0xF);
-    const __m128i mzero = _mm_setzero_si128();
-    const __m256i mone  = _mm256_set1_epi8(1);
-
-    __m256 acc = _mm256_setzero_ps();
-
-    float summs = 0.f;
-
-   for (int i = 0; i < nb; ++i) {
-
-        const uint8_t * restrict q5 = x[i].qs;
-        const int8_t  * restrict q8 = y[i].qs;
-
-#if QK_K == 256
-        const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d);
-        const float dmin = -y[i].d * GGML_FP16_TO_FP32(x[i].dmin);
-
-        memcpy(utmp, x[i].scales, 12);
-        utmp[3] = ((utmp[2] >> 4) & kmask2) | (((utmp[1] >> 6) & kmask3) << 4);
-        const uint32_t uaux = utmp[1] & kmask1;
-        utmp[1] = (utmp[2] & kmask2) | (((utmp[0] >> 6) & kmask3) << 4);
-        utmp[2] = uaux;
-        utmp[0] &= kmask1;
-#else
-        // TODO
-        const float d = 0, dmin = 0;
-#endif
-
-        const __m256i mins_and_scales = _mm256_cvtepu8_epi16(_mm_set_epi32(utmp[3], utmp[2], utmp[1], utmp[0]));
-
-        const __m256i q8sums = _mm256_loadu_si256((const __m256i*)y[i].bsums);
-        const __m128i q8s = _mm_hadd_epi16(_mm256_extracti128_si256(q8sums, 0), _mm256_extracti128_si256(q8sums, 1));
-        const __m128i prod = _mm_madd_epi16(_mm256_extracti128_si256(mins_and_scales, 1), q8s);
-        const __m128i hsum = _mm_hadd_epi32(_mm_hadd_epi32(prod, mzero), mzero);
-        summs += dmin * _mm_extract_epi32(hsum, 0);
-
-        const __m128i sc128  = _mm256_extracti128_si256(mins_and_scales, 0);
-        const __m256i scales = MM256_SET_M128I(sc128, sc128);
-
-        const __m256i hbits = _mm256_loadu_si256((const __m256i*)x[i].qh);
-        __m256i hmask = mone;
-
-        __m256i sumi = _mm256_setzero_si256();
-
-        int bit = 0;
-
-        for (int j = 0; j < QK_K/64; ++j) {
-
-            const __m256i scale_0 = _mm256_shuffle_epi8(scales, get_scale_shuffle_k4(2*j+0));
-            const __m256i scale_1 = _mm256_shuffle_epi8(scales, get_scale_shuffle_k4(2*j+1));
-
-            const __m256i q5bits = _mm256_loadu_si256((const __m256i*)q5); q5 += 32;
-
-            const __m256i q5l_0 = _mm256_and_si256(q5bits, m4);
-            const __m256i q5h_0 = _mm256_slli_epi16(_mm256_srli_epi16(_mm256_and_si256(hbits, hmask), bit++), 4);
-            const __m256i q5_0  = _mm256_add_epi8(q5l_0, q5h_0);
-            hmask = _mm256_slli_epi16(hmask, 1);
-
-            const __m256i q5l_1 = _mm256_and_si256(_mm256_srli_epi16(q5bits, 4), m4);
-            const __m256i q5h_1 = _mm256_slli_epi16(_mm256_srli_epi16(_mm256_and_si256(hbits, hmask), bit++), 4);
-            const __m256i q5_1  = _mm256_add_epi8(q5l_1, q5h_1);
-            hmask = _mm256_slli_epi16(hmask, 1);
-
-            const __m256i q8_0 = _mm256_loadu_si256((const __m256i*)q8); q8 += 32;
-            const __m256i q8_1 = _mm256_loadu_si256((const __m256i*)q8); q8 += 32;
-
-            __m256i p16_0 = _mm256_maddubs_epi16(q5_0, q8_0);
-            __m256i p16_1 = _mm256_maddubs_epi16(q5_1, q8_1);
-
-            p16_0 = _mm256_madd_epi16(scale_0, p16_0);
-            p16_1 = _mm256_madd_epi16(scale_1, p16_1);
-
-            sumi = _mm256_add_epi32(sumi, _mm256_add_epi32(p16_0, p16_1));
-
-        }
-
-        __m256 vd = _mm256_set1_ps(d);
-        acc = _mm256_fmadd_ps(vd, _mm256_cvtepi32_ps(sumi), acc);
-
-    }
-
-    *s = hsum_float_8(acc) + summs;
-
-#elif defined __AVX__
-
-    const __m128i m4 = _mm_set1_epi8(0xF);
-    const __m128i mzero = _mm_setzero_si128();
-    const __m128i mone  = _mm_set1_epi8(1);
-    const __m128i m2 = _mm_set1_epi8(2);
-
-    __m256 acc = _mm256_setzero_ps();
-
-    float summs = 0.f;
-
-    for (int i = 0; i < nb; ++i) {
-
-        const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d);
-        const float dmin = -y[i].d * GGML_FP16_TO_FP32(x[i].dmin);
-
-        const uint8_t * restrict q5 = x[i].qs;
-        const int8_t  * restrict q8 = y[i].qs;
-
-        memcpy(utmp, x[i].scales, 12);
-        utmp[3] = ((utmp[2] >> 4) & kmask2) | (((utmp[1] >> 6) & kmask3) << 4);
-        const uint32_t uaux = utmp[1] & kmask1;
-        utmp[1] = (utmp[2] & kmask2) | (((utmp[0] >> 6) & kmask3) << 4);
-        utmp[2] = uaux;
-        utmp[0] &= kmask1;
-
-        const __m128i utmps = _mm_set_epi32(utmp[3], utmp[2], utmp[1], utmp[0]);
-        const __m128i scales = _mm_cvtepu8_epi16(utmps);
-        const __m128i mins = _mm_cvtepu8_epi16(_mm_unpackhi_epi64(utmps, utmps));
-
-        const __m128i q8sums_0 = _mm_loadu_si128((const __m128i*)&y[i].bsums[0]);
-        const __m128i q8sums_1 = _mm_loadu_si128((const __m128i*)&y[i].bsums[8]);
-        const __m128i q8s = _mm_hadd_epi16(q8sums_0, q8sums_1);
-        const __m128i prod = _mm_madd_epi16(mins, q8s);
-        const __m128i hsum = _mm_hadd_epi32(_mm_hadd_epi32(prod, mzero), mzero);
-        summs += dmin * _mm_extract_epi32(hsum, 0);
-
-        const __m128i hbits_0 = _mm_loadu_si128((const __m128i*)&x[i].qh[0]);
-        const __m128i hbits_1 = _mm_loadu_si128((const __m128i*)&x[i].qh[16]);
-        __m128i hmask = mone;
-
-        __m128i sumi_0 = _mm_setzero_si128();
-        __m128i sumi_1 = _mm_setzero_si128();
-
-        int bit = 0;
-
-        __m128i shuffle = _mm_set1_epi16(0x0100);
-        for (int j = 0; j < QK_K/64; ++j) {
-
-            const __m128i scale_0 = _mm_shuffle_epi8(scales, shuffle);
-            shuffle = _mm_add_epi16(shuffle, m2);
-            const __m128i scale_1 = _mm_shuffle_epi8(scales, shuffle);
-            shuffle = _mm_add_epi16(shuffle, m2);
-
-            const __m128i q5bits_0 = _mm_loadu_si128((const __m128i*)q5); q5 += 16;
-            const __m128i q5bits_1 = _mm_loadu_si128((const __m128i*)q5); q5 += 16;
-
-            __m128i q5l_0 = _mm_and_si128(q5bits_0, m4);
-            __m128i q5l_1 = _mm_and_si128(q5bits_1, m4);
-            __m128i q5h_0 = _mm_slli_epi16(_mm_srli_epi16(_mm_and_si128(hbits_0, hmask), bit), 4);
-            __m128i q5h_1 = _mm_slli_epi16(_mm_srli_epi16(_mm_and_si128(hbits_1, hmask), bit++), 4);
-            __m128i q5_0  = _mm_add_epi8(q5l_0, q5h_0);
-            __m128i q5_1  = _mm_add_epi8(q5l_1, q5h_1);
-            hmask = _mm_slli_epi16(hmask, 1);
-
-            __m128i q8_0 = _mm_loadu_si128((const __m128i*)q8); q8 += 16;
-            __m128i q8_1 = _mm_loadu_si128((const __m128i*)q8); q8 += 16;
-            __m128i p16_0 = _mm_maddubs_epi16(q5_0, q8_0);
-            __m128i p16_1 = _mm_maddubs_epi16(q5_1, q8_1);
-            p16_0 = _mm_madd_epi16(scale_0, p16_0);
-            p16_1 = _mm_madd_epi16(scale_0, p16_1);
-
-            q5l_0 = _mm_and_si128(_mm_srli_epi16(q5bits_0, 4), m4);
-            q5l_1 = _mm_and_si128(_mm_srli_epi16(q5bits_1, 4), m4);
-            q5h_0 = _mm_slli_epi16(_mm_srli_epi16(_mm_and_si128(hbits_0, hmask), bit), 4);
-            q5h_1 = _mm_slli_epi16(_mm_srli_epi16(_mm_and_si128(hbits_1, hmask), bit++), 4);
-            q5_0  = _mm_add_epi8(q5l_0, q5h_0);
-            q5_1  = _mm_add_epi8(q5l_1, q5h_1);
-            hmask = _mm_slli_epi16(hmask, 1);
-
-            q8_0 = _mm_loadu_si128((const __m128i*)q8); q8 += 16;
-            q8_1 = _mm_loadu_si128((const __m128i*)q8); q8 += 16;
-            __m128i p16_2 = _mm_maddubs_epi16(q5_0, q8_0);
-            __m128i p16_3 = _mm_maddubs_epi16(q5_1, q8_1);
-            p16_2 = _mm_madd_epi16(scale_1, p16_2);
-            p16_3 = _mm_madd_epi16(scale_1, p16_3);
-
-            sumi_0 = _mm_add_epi32(sumi_0, _mm_add_epi32(p16_0, p16_2));
-            sumi_1 = _mm_add_epi32(sumi_1, _mm_add_epi32(p16_1, p16_3));
-
-        }
-
-        __m256 vd = _mm256_set1_ps(d);
-        __m256i sumi = MM256_SET_M128I(sumi_1, sumi_0);
-        acc = _mm256_add_ps(_mm256_mul_ps(vd, _mm256_cvtepi32_ps(sumi)), acc);
-
-    }
-
-    *s = hsum_float_8(acc) + summs;
-
-#elif defined __riscv_v_intrinsic
-
-    const uint8_t * scales = (const uint8_t*)&utmp[0];
-    const uint8_t * mins   = (const uint8_t*)&utmp[2];
-
-    float sumf = 0;
-    float sums = 0.0;
-
-    size_t vl;
-
-    for (int i = 0; i < nb; ++i) {
-
-        vl = 8;
-
-        const uint8_t * restrict q5 = x[i].qs;
-        const uint8_t * restrict hm = x[i].qh;
-        const  int8_t * restrict q8 = y[i].qs;
-
-        const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
-        const float dmin = GGML_FP16_TO_FP32(x[i].dmin) * y[i].d;
-
-        vint16mf2_t q8sums_0 = __riscv_vlse16_v_i16mf2(y[i].bsums, 4, vl);
-        vint16mf2_t q8sums_1 = __riscv_vlse16_v_i16mf2(y[i].bsums+1, 4, vl);
-        vint16mf2_t q8sums = __riscv_vadd_vv_i16mf2(q8sums_0, q8sums_1, vl);
-
-        memcpy(utmp, x[i].scales, 12);
-        utmp[3] = ((utmp[2] >> 4) & kmask2) | (((utmp[1] >> 6) & kmask3) << 4);
-        const uint32_t uaux = utmp[1] & kmask1;
-        utmp[1] = (utmp[2] & kmask2) | (((utmp[0] >> 6) & kmask3) << 4);
-        utmp[2] = uaux;
-        utmp[0] &= kmask1;
-
-        vuint8mf4_t mins8 = __riscv_vle8_v_u8mf4(mins, vl);
-        vint16mf2_t v_mins = __riscv_vreinterpret_v_u16mf2_i16mf2(__riscv_vzext_vf2_u16mf2(mins8, vl));
-        vint32m1_t prod = __riscv_vwmul_vv_i32m1(q8sums, v_mins, vl);
-
-        vint32m1_t sumi = __riscv_vredsum_vs_i32m1_i32m1(prod, __riscv_vmv_v_x_i32m1(0, 1), vl);
-        sumf -= dmin * __riscv_vmv_x_s_i32m1_i32(sumi);
-
-        vl = 32;
-        int32_t aux32 = 0;
-        int is = 0;
-
-        uint8_t m = 1;
-        vint32m1_t vzero = __riscv_vmv_v_x_i32m1(0, 1);
-        vuint8m1_t vqh = __riscv_vle8_v_u8m1(hm, vl);
-
-        for (int j = 0; j < QK_K/64; ++j) {
-            // load Q5 and Q8
-            vuint8m1_t q5_x = __riscv_vle8_v_u8m1(q5, vl);
-            vint8m1_t  q8_y1 = __riscv_vle8_v_i8m1(q8, vl);
-            vint8m1_t  q8_y2 = __riscv_vle8_v_i8m1(q8+32, vl);
-
-            // compute mask for addition
-            vint8m1_t q5_a = __riscv_vreinterpret_v_u8m1_i8m1(__riscv_vand_vx_u8m1(q5_x, 0x0F, vl));
-            vuint8m1_t qh_m1 = __riscv_vand_vx_u8m1(vqh, m, vl);
-            vbool8_t vmask_1 = __riscv_vmsne_vx_u8m1_b8(qh_m1, 0, vl);
-            vint8m1_t q5_m1 = __riscv_vadd_vx_i8m1_m(vmask_1, q5_a, 16, vl);
-            m <<= 1;
-
-            vint8m1_t q5_l = __riscv_vreinterpret_v_u8m1_i8m1(__riscv_vsrl_vx_u8m1(q5_x, 0x04, vl));
-            vuint8m1_t qh_m2 = __riscv_vand_vx_u8m1(vqh, m, vl);
-            vbool8_t vmask_2 = __riscv_vmsne_vx_u8m1_b8(qh_m2, 0, vl);
-            vint8m1_t q5_m2 = __riscv_vadd_vx_i8m1_m(vmask_2, q5_l, 16, vl);
-            m <<= 1;
-
-            vint16m2_t v0 = __riscv_vwmul_vv_i16m2(q5_m1, q8_y1, vl);
-            vint16m2_t v1 = __riscv_vwmul_vv_i16m2(q5_m2, q8_y2, vl);
-
-            vint32m4_t vs1 = __riscv_vwmul_vx_i32m4(v0, scales[is++], vl);
-            vint32m4_t vs2 = __riscv_vwmul_vx_i32m4(v1, scales[is++], vl);
-
-            vint32m1_t vacc1 = __riscv_vredsum_vs_i32m4_i32m1(vs1, vzero, vl);
-            vint32m1_t vacc2 = __riscv_vredsum_vs_i32m4_i32m1(vs2, vzero, vl);
-
-            aux32 += __riscv_vmv_x_s_i32m1_i32(vacc1) + __riscv_vmv_x_s_i32m1_i32(vacc2);
-            q5 += 32;    q8 += 64;
-
-        }
-
-        vfloat32m1_t vaux = __riscv_vfmul_vf_f32m1(__riscv_vfmv_v_f_f32m1(aux32, 1), d, 1);
-        sums += __riscv_vfmv_f_s_f32m1_f32(vaux);
-
-    }
-
-    *s = sumf+sums;
-
-#else
-
-    const uint8_t * scales = (const uint8_t*)&utmp[0];
-    const uint8_t * mins   = (const uint8_t*)&utmp[2];
-
-    int8_t  aux8[QK_K];
-    int16_t aux16[8];
-    float   sums [8];
-    int32_t aux32[8];
-    memset(sums, 0, 8*sizeof(float));
-
-    float sumf = 0;
-    for (int i = 0; i < nb; ++i) {
-        const uint8_t * restrict q4 = x[i].qs;
-        const uint8_t * restrict hm = x[i].qh;
-        const  int8_t * restrict q8 = y[i].qs;
-        memset(aux32, 0, 8*sizeof(int32_t));
-        int8_t * restrict a = aux8;
-        uint8_t m = 1;
-        for (int j = 0; j < QK_K/64; ++j) {
-            for (int l = 0; l < 32; ++l) a[l] = (int8_t)(q4[l] & 0xF);
-            for (int l = 0; l < 32; ++l) a[l] += (hm[l] & m ? 16 : 0);
-            a += 32; m <<= 1;
-            for (int l = 0; l < 32; ++l) a[l] = (int8_t)(q4[l]  >> 4);
-            for (int l = 0; l < 32; ++l) a[l] += (hm[l] & m ? 16 : 0);
-            a += 32; m <<= 1;
-            q4 += 32;
-        }
-        memcpy(utmp, x[i].scales, 12);
-        utmp[3] = ((utmp[2] >> 4) & kmask2) | (((utmp[1] >> 6) & kmask3) << 4);
-        const uint32_t uaux = utmp[1] & kmask1;
-        utmp[1] = (utmp[2] & kmask2) | (((utmp[0] >> 6) & kmask3) << 4);
-        utmp[2] = uaux;
-        utmp[0] &= kmask1;
-
-        int sumi = 0;
-        for (int j = 0; j < QK_K/16; ++j) sumi += y[i].bsums[j] * mins[j/2];
-        a = aux8;
-        int is = 0;
-        for (int j = 0; j < QK_K/32; ++j) {
-            int32_t scale = scales[is++];
-            for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
-            for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
-            q8 += 8; a += 8;
-            for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
-            for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
-            q8 += 8; a += 8;
-            for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
-            for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
-            q8 += 8; a += 8;
-            for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
-            for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
-            q8 += 8; a += 8;
-        }
-        const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
-        for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l];
-        const float dmin = GGML_FP16_TO_FP32(x[i].dmin) * y[i].d;
-        sumf -= dmin * sumi;
-    }
-    for (int l = 0; l < 8; ++l) sumf += sums[l];
-    *s = sumf;
-#endif
-}
-
-#else
-
-void ggml_vec_dot_q5_K_q8_K(const int n, float * restrict s, const void * restrict vx, const void * restrict vy) {
-    assert(n % QK_K == 0);
-
-    const block_q5_K * restrict x = vx;
-    const block_q8_K * restrict y = vy;
-
-    const int nb = n / QK_K;
-
-#ifdef __ARM_NEON
-    const uint8x16_t m4b = vdupq_n_u8(0xf);
-    const uint8x16_t mh = vdupq_n_u8(16);
-    const int32x4_t mzero = vdupq_n_s32(0);
-
-    ggml_int8x16x4_t q5bytes;
-    ggml_uint8x16x4_t q5h;
-
-    float sumf = 0;
-
-    for (int i = 0; i < nb; ++i) {
-
-        const float d = y[i].d * (float)x[i].d;
-        const int8_t * sc = x[i].scales;
-
-        const uint8_t * restrict q5 = x[i].qs;
-        const uint8_t * restrict qh = x[i].qh;
-        const int8_t  * restrict q8 = y[i].qs;
-
-        const uint8x8_t qhbits = vld1_u8(qh);
-
-        const ggml_uint8x16x2_t q5bits = ggml_vld1q_u8_x2(q5);
-        const ggml_int8x16x4_t q8bytes = ggml_vld1q_s8_x4(q8);
-
-        const uint8x16_t htmp = vcombine_u8(qhbits, vshr_n_u8(qhbits, 1));
-        q5h.val[0] = vbicq_u8(mh, vshlq_n_u8(htmp, 4));
-        q5h.val[1] = vbicq_u8(mh, vshlq_n_u8(htmp, 2));
-        q5h.val[2] = vbicq_u8(mh, htmp);
-        q5h.val[3] = vbicq_u8(mh, vshrq_n_u8(htmp, 2));
-
-        q5bytes.val[0] = vsubq_s8(vreinterpretq_s8_u8(vandq_u8(q5bits.val[0], m4b)), vreinterpretq_s8_u8(q5h.val[0]));
-        q5bytes.val[1] = vsubq_s8(vreinterpretq_s8_u8(vandq_u8(q5bits.val[1], m4b)), vreinterpretq_s8_u8(q5h.val[1]));
-        q5bytes.val[2] = vsubq_s8(vreinterpretq_s8_u8(vshrq_n_u8(q5bits.val[0], 4)), vreinterpretq_s8_u8(q5h.val[2]));
-        q5bytes.val[3] = vsubq_s8(vreinterpretq_s8_u8(vshrq_n_u8(q5bits.val[1], 4)), vreinterpretq_s8_u8(q5h.val[3]));
-
-        int32_t sumi1 = sc[0] * vaddvq_s32(ggml_vdotq_s32(mzero, q5bytes.val[0], q8bytes.val[0]));
-        int32_t sumi2 = sc[1] * vaddvq_s32(ggml_vdotq_s32(mzero, q5bytes.val[1], q8bytes.val[1]));
-        int32_t sumi3 = sc[2] * vaddvq_s32(ggml_vdotq_s32(mzero, q5bytes.val[2], q8bytes.val[2]));
-        int32_t sumi4 = sc[3] * vaddvq_s32(ggml_vdotq_s32(mzero, q5bytes.val[3], q8bytes.val[3]));
-
-        sumf += d * (sumi1 + sumi2 + sumi3 + sumi4);
-    }
-
-    *s = sumf;
-
-#elif defined __AVX2__
-
-    const __m256i m4 = _mm256_set1_epi8(0xF);
-    const __m256i mone  = _mm256_set1_epi8(1);
-
-    __m256 acc = _mm256_setzero_ps();
-
-    for (int i = 0; i < nb; ++i) {
-
-        const uint8_t * restrict q5 = x[i].qs;
-        const int8_t  * restrict q8 = y[i].qs;
-
-        const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d);
-
-        const __m256i q5bits = _mm256_loadu_si256((const __m256i*)q5);
-
-        const __m256i scale_l = MM256_SET_M128I(_mm_set1_epi16(x[i].scales[1]), _mm_set1_epi16(x[i].scales[0]));
-        const __m256i scale_h = MM256_SET_M128I(_mm_set1_epi16(x[i].scales[3]), _mm_set1_epi16(x[i].scales[2]));
-
-        int64_t aux64;
-        memcpy(&aux64, x[i].qh, 8);
-        const __m128i haux128 = _mm_set_epi64x(aux64 >> 1, aux64);
-        const __m256i haux256 = MM256_SET_M128I(_mm_srli_epi16(haux128, 2), haux128);
-
-        const __m256i q5h_0 = _mm256_slli_epi16(_mm256_andnot_si256(haux256, mone), 4);
-        const __m256i q5h_1 = _mm256_slli_epi16(_mm256_andnot_si256(_mm256_srli_epi16(haux256, 4), mone), 4);
-
-        const __m256i q5l_0 = _mm256_and_si256(q5bits, m4);
-        const __m256i q5l_1 = _mm256_and_si256(_mm256_srli_epi16(q5bits, 4), m4);
-
-        const __m256i q8_0 = _mm256_loadu_si256((const __m256i*)(q8+ 0));
-        const __m256i q8_1 = _mm256_loadu_si256((const __m256i*)(q8+32));
-
-        const __m256i p16_0 = _mm256_madd_epi16(scale_l, _mm256_maddubs_epi16(q5l_0, q8_0));
-        const __m256i p16_1 = _mm256_madd_epi16(scale_h, _mm256_maddubs_epi16(q5l_1, q8_1));
-        const __m256i s16_0 = _mm256_madd_epi16(scale_l, _mm256_maddubs_epi16(q5h_0, q8_0));
-        const __m256i s16_1 = _mm256_madd_epi16(scale_h, _mm256_maddubs_epi16(q5h_1, q8_1));
-
-        const __m256i dot = _mm256_sub_epi32(_mm256_add_epi32(p16_0, p16_1), _mm256_add_epi32(s16_0, s16_1));
-
-        acc = _mm256_fmadd_ps(_mm256_set1_ps(d), _mm256_cvtepi32_ps(dot), acc);
-
-    }
-
-    *s = hsum_float_8(acc);
-
-#elif defined __AVX__
-
-    const __m128i m4 = _mm_set1_epi8(0xF);
-    const __m128i mone  = _mm_set1_epi8(1);
-
-    __m256 acc = _mm256_setzero_ps();
-
-    for (int i = 0; i < nb; ++i) {
-
-        const uint8_t * restrict q5 = x[i].qs;
-        const int8_t  * restrict q8 = y[i].qs;
-
-        const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d);
-
-        const __m256i q5bits = _mm256_loadu_si256((const __m256i*)q5);
-
-        const __m128i scale_0 = _mm_set1_epi16(x[i].scales[0]);
-        const __m128i scale_1 = _mm_set1_epi16(x[i].scales[1]);
-        const __m128i scale_2 = _mm_set1_epi16(x[i].scales[2]);
-        const __m128i scale_3 = _mm_set1_epi16(x[i].scales[3]);
-
-        int64_t aux64;
-        memcpy(&aux64, x[i].qh, 8);
-        const __m128i haux128_0 = _mm_set_epi64x(aux64 >> 1, aux64);
-        const __m128i haux128_1 = _mm_srli_epi16(haux128_0, 2);
-
-        const __m128i q5h_0 = _mm_slli_epi16(_mm_andnot_si128(haux128_0, mone), 4);
-        const __m128i q5h_1 = _mm_slli_epi16(_mm_andnot_si128(haux128_1, mone), 4);
-        const __m128i q5h_2 = _mm_slli_epi16(_mm_andnot_si128(_mm_srli_epi16(haux128_0, 4), mone), 4);
-        const __m128i q5h_3 = _mm_slli_epi16(_mm_andnot_si128(_mm_srli_epi16(haux128_1, 4), mone), 4);
-
-        const __m128i q5l_0 = _mm_and_si128(_mm256_extractf128_si256(q5bits, 0), m4);
-        const __m128i q5l_1 = _mm_and_si128(_mm256_extractf128_si256(q5bits, 1), m4);
-        const __m128i q5l_2 = _mm_and_si128(_mm_srli_epi16(_mm256_extractf128_si256(q5bits, 0), 4), m4);
-        const __m128i q5l_3 = _mm_and_si128(_mm_srli_epi16(_mm256_extractf128_si256(q5bits, 1), 4), m4);
-
-        const __m256i q8_0 = _mm256_loadu_si256((const __m256i*)(q8+ 0));
-        const __m256i q8_1 = _mm256_loadu_si256((const __m256i*)(q8+32));
-
-        const __m128i p16_0 = _mm_madd_epi16(scale_0, _mm_maddubs_epi16(q5l_0, _mm256_extractf128_si256(q8_0, 0)));
-        const __m128i p16_1 = _mm_madd_epi16(scale_1, _mm_maddubs_epi16(q5l_1, _mm256_extractf128_si256(q8_0, 1)));
-        const __m128i p16_2 = _mm_madd_epi16(scale_2, _mm_maddubs_epi16(q5l_2, _mm256_extractf128_si256(q8_1, 0)));
-        const __m128i p16_3 = _mm_madd_epi16(scale_3, _mm_maddubs_epi16(q5l_3, _mm256_extractf128_si256(q8_1, 1)));
-        const __m128i s16_0 = _mm_madd_epi16(scale_0, _mm_maddubs_epi16(q5h_0, _mm256_extractf128_si256(q8_0, 0)));
-        const __m128i s16_1 = _mm_madd_epi16(scale_1, _mm_maddubs_epi16(q5h_1, _mm256_extractf128_si256(q8_0, 1)));
-        const __m128i s16_2 = _mm_madd_epi16(scale_2, _mm_maddubs_epi16(q5h_2, _mm256_extractf128_si256(q8_1, 0)));
-        const __m128i s16_3 = _mm_madd_epi16(scale_3, _mm_maddubs_epi16(q5h_3, _mm256_extractf128_si256(q8_1, 1)));
-
-        const __m128i dot_0 = _mm_sub_epi32(_mm_add_epi32(p16_0, p16_2), _mm_add_epi32(s16_0, s16_2));
-        const __m128i dot_1 = _mm_sub_epi32(_mm_add_epi32(p16_1, p16_3), _mm_add_epi32(s16_1, s16_3));
-
-        acc = _mm256_add_ps(_mm256_mul_ps(_mm256_set1_ps(d), _mm256_cvtepi32_ps(MM256_SET_M128I(dot_1, dot_0))), acc);
-
-    }
-
-    *s = hsum_float_8(acc);
-
-#elif defined __riscv_v_intrinsic
-
-    float sumf = 0;
-
-    for (int i = 0; i < nb; ++i) {
-
-        const float d = y[i].d * (float)x[i].d;
-        const int8_t * sc = x[i].scales;
-
-        const uint8_t * restrict q5 = x[i].qs;
-        const uint8_t * restrict qh = x[i].qh;
-        const int8_t  * restrict q8 = y[i].qs;
-
-        vint32m1_t vzero = __riscv_vmv_v_x_i32m1(0, 1);
-
-        // load qh
-        vuint8mf4_t qh_x1   = __riscv_vle8_v_u8mf4(qh, 8);
-        vuint8mf2_t qh_x2   = __riscv_vlmul_ext_v_u8mf4_u8mf2(__riscv_vsrl_vx_u8mf4(qh_x1, 1, 8));
-
-        size_t vl = 16;
-
-        // combine both qh_1 and qh_2
-        vuint8mf2_t qh_x = __riscv_vslideup_vx_u8mf2(__riscv_vlmul_ext_v_u8mf4_u8mf2(qh_x1), qh_x2, vl/2, vl);
-
-        vuint8mf2_t qh_h0 = __riscv_vand_vx_u8mf2(__riscv_vnot_v_u8mf2(__riscv_vsll_vx_u8mf2(qh_x, 0x4, vl), vl), 16, vl);
-        vuint8mf2_t qh_h1 = __riscv_vand_vx_u8mf2(__riscv_vnot_v_u8mf2(__riscv_vsll_vx_u8mf2(qh_x, 0x2, vl), vl), 16, vl);
-        vuint8mf2_t qh_h2 = __riscv_vand_vx_u8mf2(__riscv_vnot_v_u8mf2(qh_x, vl), 16, vl);
-        vuint8mf2_t qh_h3 = __riscv_vand_vx_u8mf2(__riscv_vnot_v_u8mf2(__riscv_vsrl_vx_u8mf2(qh_x, 0x4, vl), vl), 16, vl);
-
-        vint8mf2_t qh_0 = __riscv_vreinterpret_v_u8mf2_i8mf2(qh_h0);
-        vint8mf2_t qh_1 = __riscv_vreinterpret_v_u8mf2_i8mf2(qh_h1);
-        vint8mf2_t qh_2 = __riscv_vreinterpret_v_u8mf2_i8mf2(qh_h2);
-        vint8mf2_t qh_3 = __riscv_vreinterpret_v_u8mf2_i8mf2(qh_h3);
-
-        // load q5
-        vuint8mf2_t q5_x1  = __riscv_vle8_v_u8mf2(q5, vl);
-        vuint8mf2_t q5_x2  = __riscv_vle8_v_u8mf2(q5+16, vl);
-
-        vint8mf2_t q5s_0 = __riscv_vreinterpret_v_u8mf2_i8mf2(__riscv_vand_vx_u8mf2(q5_x1, 0xF, vl));
-        vint8mf2_t q5s_1 = __riscv_vreinterpret_v_u8mf2_i8mf2(__riscv_vand_vx_u8mf2(q5_x2, 0xF, vl));
-        vint8mf2_t q5s_2 = __riscv_vreinterpret_v_u8mf2_i8mf2(__riscv_vsrl_vx_u8mf2(q5_x1, 0x4, vl));
-        vint8mf2_t q5s_3 = __riscv_vreinterpret_v_u8mf2_i8mf2(__riscv_vsrl_vx_u8mf2(q5_x2, 0x4, vl));
-
-        vint8mf2_t q5_0 = __riscv_vsub_vv_i8mf2(q5s_0, qh_0, vl);
-        vint8mf2_t q5_1 = __riscv_vsub_vv_i8mf2(q5s_1, qh_1, vl);
-        vint8mf2_t q5_2 = __riscv_vsub_vv_i8mf2(q5s_2, qh_2, vl);
-        vint8mf2_t q5_3 = __riscv_vsub_vv_i8mf2(q5s_3, qh_3, vl);
-
-        // load Q8 and multiply it with Q5
-        vint16m1_t p0 = __riscv_vwmul_vv_i16m1(q5_0, __riscv_vle8_v_i8mf2(q8, vl), vl);
-        vint16m1_t p1 = __riscv_vwmul_vv_i16m1(q5_1, __riscv_vle8_v_i8mf2(q8+16, vl), vl);
-        vint16m1_t p2 = __riscv_vwmul_vv_i16m1(q5_2, __riscv_vle8_v_i8mf2(q8+32, vl), vl);
-        vint16m1_t p3 = __riscv_vwmul_vv_i16m1(q5_3, __riscv_vle8_v_i8mf2(q8+48, vl), vl);
-
-        vint32m1_t vs_0 = __riscv_vwredsum_vs_i16m1_i32m1(p0, vzero, vl);
-        vint32m1_t vs_1 = __riscv_vwredsum_vs_i16m1_i32m1(p1, vzero, vl);
-        vint32m1_t vs_2 = __riscv_vwredsum_vs_i16m1_i32m1(p2, vzero, vl);
-        vint32m1_t vs_3 = __riscv_vwredsum_vs_i16m1_i32m1(p3, vzero, vl);
-
-        int32_t sumi1 = sc[0] * __riscv_vmv_x_s_i32m1_i32(vs_0);
-        int32_t sumi2 = sc[1] * __riscv_vmv_x_s_i32m1_i32(vs_1);
-        int32_t sumi3 = sc[2] * __riscv_vmv_x_s_i32m1_i32(vs_2);
-        int32_t sumi4 = sc[3] * __riscv_vmv_x_s_i32m1_i32(vs_3);
-
-        sumf += d * (sumi1 + sumi2 + sumi3 + sumi4);
-
-    }
-
-    *s = sumf;
-
-#else
-
-    int8_t aux8[QK_K];
-    int16_t aux16[16];
-    float   sums [8];
-    memset(sums, 0, 8*sizeof(float));
-
-    float sumf = 0;
-    for (int i = 0; i < nb; ++i) {
-        const uint8_t * restrict q4 = x[i].qs;
-        const uint8_t * restrict hm = x[i].qh;
-        const  int8_t * restrict q8 = y[i].qs;
-        int8_t * restrict a = aux8;
-        for (int l = 0; l < 32; ++l) {
-            a[l+ 0] = q4[l] & 0xF;
-            a[l+32] = q4[l]  >> 4;
-        }
-        for (int is = 0; is < 8; ++is) {
-            uint8_t m = 1 << is;
-            for (int l = 0; l < 8; ++l) a[8*is + l] -= (hm[l] & m ? 0 : 16);
-        }
-
-        const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d);
-        const int8_t * restrict sc = x[i].scales;
-
-        for (int j = 0; j < QK_K/16; ++j) {
-            const float dl = d * sc[j];
-            for (int l = 0; l < 16; ++l) aux16[l] = q8[l] * a[l];
-            for (int l = 0; l <  8; ++l) sums[l] += dl * (aux16[l] + aux16[8+l]);
-            q8 += 16; a += 16;
-        }
-    }
-    for (int l = 0; l < 8; ++l) sumf += sums[l];
-    *s = sumf;
-#endif
-}
-#endif
-
-
-#if QK_K == 256
-void ggml_vec_dot_q6_K_q8_K(const int n, float * restrict s, const void * restrict vx, const void * restrict vy) {
-    assert(n % QK_K == 0);
-
-    const block_q6_K * restrict x = vx;
-    const block_q8_K * restrict y = vy;
-
-    const int nb = n / QK_K;
-
-#ifdef __ARM_NEON
-    float sum = 0;
-
-    const uint8x16_t m4b = vdupq_n_u8(0xF);
-    const int32x4_t  vzero = vdupq_n_s32(0);
-    //const int8x16_t  m32s = vdupq_n_s8(32);
-
-    const uint8x16_t mone = vdupq_n_u8(3);
-
-    ggml_int8x16x4_t q6bytes;
-    ggml_uint8x16x4_t q6h;
-
-    for (int i = 0; i < nb; ++i) {
-
-        const float d_all = GGML_FP16_TO_FP32(x[i].d);
-
-        const uint8_t * restrict q6 = x[i].ql;
-        const uint8_t * restrict qh = x[i].qh;
-        const int8_t  * restrict q8 = y[i].qs;
-
-        const int8_t * restrict scale = x[i].scales;
-
-        const ggml_int16x8x2_t q8sums = ggml_vld1q_s16_x2(y[i].bsums);
-        const int8x16_t scales = vld1q_s8(scale);
-        const ggml_int16x8x2_t q6scales = {{vmovl_s8(vget_low_s8(scales)), vmovl_s8(vget_high_s8(scales))}};
-
-        const int32x4_t prod = vaddq_s32(vaddq_s32(vmull_s16(vget_low_s16 (q8sums.val[0]), vget_low_s16 (q6scales.val[0])),
-                                                   vmull_s16(vget_high_s16(q8sums.val[0]), vget_high_s16(q6scales.val[0]))),
-                                         vaddq_s32(vmull_s16(vget_low_s16 (q8sums.val[1]), vget_low_s16 (q6scales.val[1])),
-                                                   vmull_s16(vget_high_s16(q8sums.val[1]), vget_high_s16(q6scales.val[1]))));
-        int32_t isum_mins = vaddvq_s32(prod);
-
-        int32_t isum = 0;
-
-        for (int j = 0; j < QK_K/128; ++j) {
-
-            ggml_uint8x16x2_t qhbits = ggml_vld1q_u8_x2(qh); qh += 32;
-            ggml_uint8x16x4_t q6bits = ggml_vld1q_u8_x4(q6); q6 += 64;
-            ggml_int8x16x4_t q8bytes = ggml_vld1q_s8_x4(q8); q8 += 64;
-
-            q6h.val[0] = vshlq_n_u8(vandq_u8(mone, qhbits.val[0]), 4);
-            q6h.val[1] = vshlq_n_u8(vandq_u8(mone, qhbits.val[1]), 4);
-            uint8x16_t shifted = vshrq_n_u8(qhbits.val[0], 2);
-            q6h.val[2] = vshlq_n_u8(vandq_u8(mone, shifted), 4);
-            shifted = vshrq_n_u8(qhbits.val[1], 2);
-            q6h.val[3] = vshlq_n_u8(vandq_u8(mone, shifted), 4);
-
-            //q6bytes.val[0] = vsubq_s8(vreinterpretq_s8_u8(vorrq_u8(vandq_u8(q6bits.val[0], m4b), q6h.val[0])), m32s);
-            //q6bytes.val[1] = vsubq_s8(vreinterpretq_s8_u8(vorrq_u8(vandq_u8(q6bits.val[1], m4b), q6h.val[1])), m32s);
-            //q6bytes.val[2] = vsubq_s8(vreinterpretq_s8_u8(vorrq_u8(vandq_u8(q6bits.val[2], m4b), q6h.val[2])), m32s);
-            //q6bytes.val[3] = vsubq_s8(vreinterpretq_s8_u8(vorrq_u8(vandq_u8(q6bits.val[3], m4b), q6h.val[3])), m32s);
-            q6bytes.val[0] = vreinterpretq_s8_u8(vorrq_u8(vandq_u8(q6bits.val[0], m4b), q6h.val[0]));
-            q6bytes.val[1] = vreinterpretq_s8_u8(vorrq_u8(vandq_u8(q6bits.val[1], m4b), q6h.val[1]));
-            q6bytes.val[2] = vreinterpretq_s8_u8(vorrq_u8(vandq_u8(q6bits.val[2], m4b), q6h.val[2]));
-            q6bytes.val[3] = vreinterpretq_s8_u8(vorrq_u8(vandq_u8(q6bits.val[3], m4b), q6h.val[3]));
-
-            isum += vaddvq_s32(ggml_vdotq_s32(vzero, q6bytes.val[0], q8bytes.val[0])) * scale[0] +
-                    vaddvq_s32(ggml_vdotq_s32(vzero, q6bytes.val[1], q8bytes.val[1])) * scale[1] +
-                    vaddvq_s32(ggml_vdotq_s32(vzero, q6bytes.val[2], q8bytes.val[2])) * scale[2] +
-                    vaddvq_s32(ggml_vdotq_s32(vzero, q6bytes.val[3], q8bytes.val[3])) * scale[3];
-
-            scale += 4;
-
-            q8bytes = ggml_vld1q_s8_x4(q8); q8 += 64;
-
-            shifted = vshrq_n_u8(qhbits.val[0], 4);
-            q6h.val[0] = vshlq_n_u8(vandq_u8(mone, shifted), 4);
-            shifted = vshrq_n_u8(qhbits.val[1], 4);
-            q6h.val[1] = vshlq_n_u8(vandq_u8(mone, shifted), 4);
-            shifted = vshrq_n_u8(qhbits.val[0], 6);
-            q6h.val[2] = vshlq_n_u8(vandq_u8(mone, shifted), 4);
-            shifted = vshrq_n_u8(qhbits.val[1], 6);
-            q6h.val[3] = vshlq_n_u8(vandq_u8(mone, shifted), 4);
-
-            //q6bytes.val[0] = vsubq_s8(vreinterpretq_s8_u8(vorrq_u8(vshrq_n_u8(q6bits.val[0], 4), q6h.val[0])), m32s);
-            //q6bytes.val[1] = vsubq_s8(vreinterpretq_s8_u8(vorrq_u8(vshrq_n_u8(q6bits.val[1], 4), q6h.val[1])), m32s);
-            //q6bytes.val[2] = vsubq_s8(vreinterpretq_s8_u8(vorrq_u8(vshrq_n_u8(q6bits.val[2], 4), q6h.val[2])), m32s);
-            //q6bytes.val[3] = vsubq_s8(vreinterpretq_s8_u8(vorrq_u8(vshrq_n_u8(q6bits.val[3], 4), q6h.val[3])), m32s);
-            q6bytes.val[0] = vreinterpretq_s8_u8(vorrq_u8(vshrq_n_u8(q6bits.val[0], 4), q6h.val[0]));
-            q6bytes.val[1] = vreinterpretq_s8_u8(vorrq_u8(vshrq_n_u8(q6bits.val[1], 4), q6h.val[1]));
-            q6bytes.val[2] = vreinterpretq_s8_u8(vorrq_u8(vshrq_n_u8(q6bits.val[2], 4), q6h.val[2]));
-            q6bytes.val[3] = vreinterpretq_s8_u8(vorrq_u8(vshrq_n_u8(q6bits.val[3], 4), q6h.val[3]));
-
-            isum += vaddvq_s32(ggml_vdotq_s32(vzero, q6bytes.val[0], q8bytes.val[0])) * scale[0] +
-                    vaddvq_s32(ggml_vdotq_s32(vzero, q6bytes.val[1], q8bytes.val[1])) * scale[1] +
-                    vaddvq_s32(ggml_vdotq_s32(vzero, q6bytes.val[2], q8bytes.val[2])) * scale[2] +
-                    vaddvq_s32(ggml_vdotq_s32(vzero, q6bytes.val[3], q8bytes.val[3])) * scale[3];
-            scale += 4;
-        }
-        //sum += isum * d_all * y[i].d;
-        sum += d_all * y[i].d * (isum - 32 * isum_mins);
-
-    }
-    *s = sum;
-
-#elif defined __AVX2__
-
-    const __m256i m4 = _mm256_set1_epi8(0xF);
-    const __m256i m2 = _mm256_set1_epi8(3);
-    const __m256i m32s = _mm256_set1_epi8(32);
-
-    __m256 acc = _mm256_setzero_ps();
-
-    for (int i = 0; i < nb; ++i) {
-
-        const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d);
-
-        const uint8_t * restrict q4 = x[i].ql;
-        const uint8_t * restrict qh = x[i].qh;
-        const int8_t  * restrict q8 = y[i].qs;
-
-        const __m128i scales = _mm_loadu_si128((const __m128i*)x[i].scales);
-
-        __m256i sumi = _mm256_setzero_si256();
-
-        int is = 0;
-
-        for (int j = 0; j < QK_K/128; ++j) {
-
-            const __m128i scale_0 = _mm_shuffle_epi8(scales, get_scale_shuffle(is + 0));
-            const __m128i scale_1 = _mm_shuffle_epi8(scales, get_scale_shuffle(is + 1));
-            const __m128i scale_2 = _mm_shuffle_epi8(scales, get_scale_shuffle(is + 2));
-            const __m128i scale_3 = _mm_shuffle_epi8(scales, get_scale_shuffle(is + 3));
-            is += 4;
-
-            const __m256i q4bits1 = _mm256_loadu_si256((const __m256i*)q4); q4 += 32;
-            const __m256i q4bits2 = _mm256_loadu_si256((const __m256i*)q4); q4 += 32;
-            const __m256i q4bitsH = _mm256_loadu_si256((const __m256i*)qh); qh += 32;
-
-            const __m256i q4h_0 = _mm256_slli_epi16(_mm256_and_si256(q4bitsH, m2), 4);
-            const __m256i q4h_1 = _mm256_slli_epi16(_mm256_and_si256(_mm256_srli_epi16(q4bitsH, 2), m2), 4);
-            const __m256i q4h_2 = _mm256_slli_epi16(_mm256_and_si256(_mm256_srli_epi16(q4bitsH, 4), m2), 4);
-            const __m256i q4h_3 = _mm256_slli_epi16(_mm256_and_si256(_mm256_srli_epi16(q4bitsH, 6), m2), 4);
-
-            const __m256i q4_0 = _mm256_or_si256(_mm256_and_si256(q4bits1, m4), q4h_0);
-            const __m256i q4_1 = _mm256_or_si256(_mm256_and_si256(q4bits2, m4), q4h_1);
-            const __m256i q4_2 = _mm256_or_si256(_mm256_and_si256(_mm256_srli_epi16(q4bits1, 4), m4), q4h_2);
-            const __m256i q4_3 = _mm256_or_si256(_mm256_and_si256(_mm256_srli_epi16(q4bits2, 4), m4), q4h_3);
-
-            const __m256i q8_0 = _mm256_loadu_si256((const __m256i*)q8); q8 += 32;
-            const __m256i q8_1 = _mm256_loadu_si256((const __m256i*)q8); q8 += 32;
-            const __m256i q8_2 = _mm256_loadu_si256((const __m256i*)q8); q8 += 32;
-            const __m256i q8_3 = _mm256_loadu_si256((const __m256i*)q8); q8 += 32;
-
-            __m256i q8s_0 = _mm256_maddubs_epi16(m32s, q8_0);
-            __m256i q8s_1 = _mm256_maddubs_epi16(m32s, q8_1);
-            __m256i q8s_2 = _mm256_maddubs_epi16(m32s, q8_2);
-            __m256i q8s_3 = _mm256_maddubs_epi16(m32s, q8_3);
-
-            __m256i p16_0 = _mm256_maddubs_epi16(q4_0, q8_0);
-            __m256i p16_1 = _mm256_maddubs_epi16(q4_1, q8_1);
-            __m256i p16_2 = _mm256_maddubs_epi16(q4_2, q8_2);
-            __m256i p16_3 = _mm256_maddubs_epi16(q4_3, q8_3);
-
-            p16_0 = _mm256_sub_epi16(p16_0, q8s_0);
-            p16_1 = _mm256_sub_epi16(p16_1, q8s_1);
-            p16_2 = _mm256_sub_epi16(p16_2, q8s_2);
-            p16_3 = _mm256_sub_epi16(p16_3, q8s_3);
-
-            p16_0 = _mm256_madd_epi16(_mm256_cvtepi8_epi16(scale_0), p16_0);
-            p16_1 = _mm256_madd_epi16(_mm256_cvtepi8_epi16(scale_1), p16_1);
-            p16_2 = _mm256_madd_epi16(_mm256_cvtepi8_epi16(scale_2), p16_2);
-            p16_3 = _mm256_madd_epi16(_mm256_cvtepi8_epi16(scale_3), p16_3);
-
-            sumi = _mm256_add_epi32(sumi, _mm256_add_epi32(p16_0, p16_1));
-            sumi = _mm256_add_epi32(sumi, _mm256_add_epi32(p16_2, p16_3));
-
-        }
-
-        acc = _mm256_fmadd_ps(_mm256_broadcast_ss(&d), _mm256_cvtepi32_ps(sumi), acc);
-    }
-
-    *s = hsum_float_8(acc);
-
-#elif defined __AVX__
-
-    const __m128i m4 = _mm_set1_epi8(0xF);
-    const __m128i m3 = _mm_set1_epi8(3);
-    const __m128i m32s = _mm_set1_epi8(32);
-    const __m128i m2 = _mm_set1_epi8(2);
-
-    __m256 acc = _mm256_setzero_ps();
-
-    for (int i = 0; i < nb; ++i) {
-
-        const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d);
-
-        const uint8_t * restrict q4 = x[i].ql;
-        const uint8_t * restrict qh = x[i].qh;
-        const int8_t  * restrict q8 = y[i].qs;
-
-        const __m128i scales = _mm_loadu_si128((const __m128i*)x[i].scales);
-
-        __m128i sumi_0 = _mm_setzero_si128();
-        __m128i sumi_1 = _mm_setzero_si128();
-
-        __m128i shuffle = _mm_set_epi64x(0x0101010101010101, 0x0000000000000000);
-        for (int j = 0; j < QK_K/128; ++j) {
-
-            const __m128i q4bitsH_0 = _mm_loadu_si128((const __m128i*)qh); qh += 16;
-            const __m128i q4bitsH_1 = _mm_loadu_si128((const __m128i*)qh); qh += 16;
-
-            const __m128i q4h_0 = _mm_slli_epi16(_mm_and_si128(q4bitsH_0, m3), 4);
-            const __m128i q4h_1 = _mm_slli_epi16(_mm_and_si128(q4bitsH_1, m3), 4);
-            const __m128i q4h_2 = _mm_slli_epi16(_mm_and_si128(_mm_srli_epi16(q4bitsH_0, 2), m3), 4);
-            const __m128i q4h_3 = _mm_slli_epi16(_mm_and_si128(_mm_srli_epi16(q4bitsH_1, 2), m3), 4);
-            const __m128i q4h_4 = _mm_slli_epi16(_mm_and_si128(_mm_srli_epi16(q4bitsH_0, 4), m3), 4);
-            const __m128i q4h_5 = _mm_slli_epi16(_mm_and_si128(_mm_srli_epi16(q4bitsH_1, 4), m3), 4);
-            const __m128i q4h_6 = _mm_slli_epi16(_mm_and_si128(_mm_srli_epi16(q4bitsH_0, 6), m3), 4);
-            const __m128i q4h_7 = _mm_slli_epi16(_mm_and_si128(_mm_srli_epi16(q4bitsH_1, 6), m3), 4);
-
-            const __m128i q4bits1_0 = _mm_loadu_si128((const __m128i*)q4); q4 += 16;
-            const __m128i q4bits1_1 = _mm_loadu_si128((const __m128i*)q4); q4 += 16;
-            const __m128i q4bits2_0 = _mm_loadu_si128((const __m128i*)q4); q4 += 16;
-            const __m128i q4bits2_1 = _mm_loadu_si128((const __m128i*)q4); q4 += 16;
-
-            const __m128i q4_0 = _mm_or_si128(_mm_and_si128(q4bits1_0, m4), q4h_0);
-            const __m128i q4_1 = _mm_or_si128(_mm_and_si128(q4bits1_1, m4), q4h_1);
-            const __m128i q4_2 = _mm_or_si128(_mm_and_si128(q4bits2_0, m4), q4h_2);
-            const __m128i q4_3 = _mm_or_si128(_mm_and_si128(q4bits2_1, m4), q4h_3);
-            const __m128i q4_4 = _mm_or_si128(_mm_and_si128(_mm_srli_epi16(q4bits1_0, 4), m4), q4h_4);
-            const __m128i q4_5 = _mm_or_si128(_mm_and_si128(_mm_srli_epi16(q4bits1_1, 4), m4), q4h_5);
-            const __m128i q4_6 = _mm_or_si128(_mm_and_si128(_mm_srli_epi16(q4bits2_0, 4), m4), q4h_6);
-            const __m128i q4_7 = _mm_or_si128(_mm_and_si128(_mm_srli_epi16(q4bits2_1, 4), m4), q4h_7);
-
-            const __m128i q8_0 = _mm_loadu_si128((const __m128i*)q8); q8 += 16;
-            const __m128i q8_1 = _mm_loadu_si128((const __m128i*)q8); q8 += 16;
-            const __m128i q8_2 = _mm_loadu_si128((const __m128i*)q8); q8 += 16;
-            const __m128i q8_3 = _mm_loadu_si128((const __m128i*)q8); q8 += 16;
-            const __m128i q8_4 = _mm_loadu_si128((const __m128i*)q8); q8 += 16;
-            const __m128i q8_5 = _mm_loadu_si128((const __m128i*)q8); q8 += 16;
-            const __m128i q8_6 = _mm_loadu_si128((const __m128i*)q8); q8 += 16;
-            const __m128i q8_7 = _mm_loadu_si128((const __m128i*)q8); q8 += 16;
-
-            __m128i q8s_0 = _mm_maddubs_epi16(m32s, q8_0);
-            __m128i q8s_1 = _mm_maddubs_epi16(m32s, q8_1);
-            __m128i q8s_2 = _mm_maddubs_epi16(m32s, q8_2);
-            __m128i q8s_3 = _mm_maddubs_epi16(m32s, q8_3);
-            __m128i q8s_4 = _mm_maddubs_epi16(m32s, q8_4);
-            __m128i q8s_5 = _mm_maddubs_epi16(m32s, q8_5);
-            __m128i q8s_6 = _mm_maddubs_epi16(m32s, q8_6);
-            __m128i q8s_7 = _mm_maddubs_epi16(m32s, q8_7);
-
-            __m128i p16_0 = _mm_maddubs_epi16(q4_0, q8_0);
-            __m128i p16_1 = _mm_maddubs_epi16(q4_1, q8_1);
-            __m128i p16_2 = _mm_maddubs_epi16(q4_2, q8_2);
-            __m128i p16_3 = _mm_maddubs_epi16(q4_3, q8_3);
-            __m128i p16_4 = _mm_maddubs_epi16(q4_4, q8_4);
-            __m128i p16_5 = _mm_maddubs_epi16(q4_5, q8_5);
-            __m128i p16_6 = _mm_maddubs_epi16(q4_6, q8_6);
-            __m128i p16_7 = _mm_maddubs_epi16(q4_7, q8_7);
-
-            p16_0 = _mm_sub_epi16(p16_0, q8s_0);
-            p16_1 = _mm_sub_epi16(p16_1, q8s_1);
-            p16_2 = _mm_sub_epi16(p16_2, q8s_2);
-            p16_3 = _mm_sub_epi16(p16_3, q8s_3);
-            p16_4 = _mm_sub_epi16(p16_4, q8s_4);
-            p16_5 = _mm_sub_epi16(p16_5, q8s_5);
-            p16_6 = _mm_sub_epi16(p16_6, q8s_6);
-            p16_7 = _mm_sub_epi16(p16_7, q8s_7);
-
-            const __m128i scale_0 = _mm_shuffle_epi8(scales, shuffle);
-            shuffle = _mm_add_epi8(shuffle, m2);
-            const __m128i scale_1 = _mm_shuffle_epi8(scales, shuffle);
-            shuffle = _mm_add_epi8(shuffle, m2);
-            const __m128i scale_2 = _mm_shuffle_epi8(scales, shuffle);
-            shuffle = _mm_add_epi8(shuffle, m2);
-            const __m128i scale_3 = _mm_shuffle_epi8(scales, shuffle);
-            shuffle = _mm_add_epi8(shuffle, m2);
-
-            p16_0 = _mm_madd_epi16(_mm_cvtepi8_epi16(scale_0), p16_0);
-            p16_1 = _mm_madd_epi16(_mm_cvtepi8_epi16(_mm_unpackhi_epi64(scale_0, scale_0)), p16_1);
-            p16_2 = _mm_madd_epi16(_mm_cvtepi8_epi16(scale_1), p16_2);
-            p16_3 = _mm_madd_epi16(_mm_cvtepi8_epi16(_mm_unpackhi_epi64(scale_1, scale_1)), p16_3);
-            p16_4 = _mm_madd_epi16(_mm_cvtepi8_epi16(scale_2), p16_4);
-            p16_5 = _mm_madd_epi16(_mm_cvtepi8_epi16(_mm_unpackhi_epi64(scale_2, scale_2)), p16_5);
-            p16_6 = _mm_madd_epi16(_mm_cvtepi8_epi16(scale_3), p16_6);
-            p16_7 = _mm_madd_epi16(_mm_cvtepi8_epi16(_mm_unpackhi_epi64(scale_3, scale_3)), p16_7);
-
-            sumi_0 = _mm_add_epi32(sumi_0, _mm_add_epi32(p16_0, p16_2));
-            sumi_1 = _mm_add_epi32(sumi_1, _mm_add_epi32(p16_1, p16_3));
-            sumi_0 = _mm_add_epi32(sumi_0, _mm_add_epi32(p16_4, p16_6));
-            sumi_1 = _mm_add_epi32(sumi_1, _mm_add_epi32(p16_5, p16_7));
-
-        }
-
-        __m256i sumi = MM256_SET_M128I(sumi_1, sumi_0);
-        acc = _mm256_add_ps(_mm256_mul_ps(_mm256_broadcast_ss(&d), _mm256_cvtepi32_ps(sumi)), acc);
-    }
-
-    *s = hsum_float_8(acc);
-
-#elif defined __riscv_v_intrinsic
-
-    float sumf = 0;
-    for (int i = 0; i < nb; ++i) {
-
-        const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
-
-        const uint8_t * restrict q6 = x[i].ql;
-        const uint8_t * restrict qh = x[i].qh;
-        const  int8_t * restrict q8 = y[i].qs;
-
-        const int8_t * restrict scale = x[i].scales;
-
-        size_t vl;
-
-        vint32m1_t vzero = __riscv_vmv_v_x_i32m1(0, 1);
-
-        int sum_t = 0;
-        int is = 0;
-
-        for (int j = 0; j < QK_K/128; ++j) {
-
-            vl = 32;
-
-            // load qh
-            vuint8m1_t qh_x = __riscv_vle8_v_u8m1(qh, vl);
-
-            // load Q6
-            vuint8m1_t q6_0 = __riscv_vle8_v_u8m1(q6, vl);
-            vuint8m1_t q6_1 = __riscv_vle8_v_u8m1(q6+32, vl);
-
-            vuint8m1_t q6a_0 = __riscv_vand_vx_u8m1(q6_0, 0x0F, vl);
-            vuint8m1_t q6a_1 = __riscv_vand_vx_u8m1(q6_1, 0x0F, vl);
-            vuint8m1_t q6s_0 = __riscv_vsrl_vx_u8m1(q6_0, 0x04, vl);
-            vuint8m1_t q6s_1 = __riscv_vsrl_vx_u8m1(q6_1, 0x04, vl);
-
-            vuint8m1_t qh_0 = __riscv_vand_vx_u8m1(qh_x, 0x03, vl);
-            vuint8m1_t qh_1 = __riscv_vand_vx_u8m1(__riscv_vsrl_vx_u8m1(qh_x, 0x2, vl), 0x03 , vl);
-            vuint8m1_t qh_2 = __riscv_vand_vx_u8m1(__riscv_vsrl_vx_u8m1(qh_x, 0x4, vl), 0x03 , vl);
-            vuint8m1_t qh_3 = __riscv_vand_vx_u8m1(__riscv_vsrl_vx_u8m1(qh_x, 0x6, vl), 0x03 , vl);
-
-            vuint8m1_t qhi_0 = __riscv_vor_vv_u8m1(q6a_0, __riscv_vsll_vx_u8m1(qh_0, 0x04, vl), vl);
-            vuint8m1_t qhi_1 = __riscv_vor_vv_u8m1(q6a_1, __riscv_vsll_vx_u8m1(qh_1, 0x04, vl), vl);
-            vuint8m1_t qhi_2 = __riscv_vor_vv_u8m1(q6s_0, __riscv_vsll_vx_u8m1(qh_2, 0x04, vl), vl);
-            vuint8m1_t qhi_3 = __riscv_vor_vv_u8m1(q6s_1, __riscv_vsll_vx_u8m1(qh_3, 0x04, vl), vl);
-
-            vint8m1_t a_0 = __riscv_vsub_vx_i8m1(__riscv_vreinterpret_v_u8m1_i8m1(qhi_0), 32, vl);
-            vint8m1_t a_1 = __riscv_vsub_vx_i8m1(__riscv_vreinterpret_v_u8m1_i8m1(qhi_1), 32, vl);
-            vint8m1_t a_2 = __riscv_vsub_vx_i8m1(__riscv_vreinterpret_v_u8m1_i8m1(qhi_2), 32, vl);
-            vint8m1_t a_3 = __riscv_vsub_vx_i8m1(__riscv_vreinterpret_v_u8m1_i8m1(qhi_3), 32, vl);
-
-            // load Q8 and take product
-            vint16m2_t va_q_0 = __riscv_vwmul_vv_i16m2(a_0, __riscv_vle8_v_i8m1(q8, vl), vl);
-            vint16m2_t va_q_1 = __riscv_vwmul_vv_i16m2(a_1, __riscv_vle8_v_i8m1(q8+32, vl), vl);
-            vint16m2_t va_q_2 = __riscv_vwmul_vv_i16m2(a_2, __riscv_vle8_v_i8m1(q8+64, vl), vl);
-            vint16m2_t va_q_3 = __riscv_vwmul_vv_i16m2(a_3, __riscv_vle8_v_i8m1(q8+96, vl), vl);
-
-            vl = 16;
-
-            vint32m2_t vaux_0 = __riscv_vwmul_vx_i32m2(__riscv_vget_v_i16m2_i16m1(va_q_0, 0), scale[is+0], vl);
-            vint32m2_t vaux_1 = __riscv_vwmul_vx_i32m2(__riscv_vget_v_i16m2_i16m1(va_q_0, 1), scale[is+1], vl);
-            vint32m2_t vaux_2 = __riscv_vwmul_vx_i32m2(__riscv_vget_v_i16m2_i16m1(va_q_1, 0), scale[is+2], vl);
-            vint32m2_t vaux_3 = __riscv_vwmul_vx_i32m2(__riscv_vget_v_i16m2_i16m1(va_q_1, 1), scale[is+3], vl);
-            vint32m2_t vaux_4 = __riscv_vwmul_vx_i32m2(__riscv_vget_v_i16m2_i16m1(va_q_2, 0), scale[is+4], vl);
-            vint32m2_t vaux_5 = __riscv_vwmul_vx_i32m2(__riscv_vget_v_i16m2_i16m1(va_q_2, 1), scale[is+5], vl);
-            vint32m2_t vaux_6 = __riscv_vwmul_vx_i32m2(__riscv_vget_v_i16m2_i16m1(va_q_3, 0), scale[is+6], vl);
-            vint32m2_t vaux_7 = __riscv_vwmul_vx_i32m2(__riscv_vget_v_i16m2_i16m1(va_q_3, 1), scale[is+7], vl);
-
-            vint32m1_t isum0 = __riscv_vredsum_vs_i32m2_i32m1(__riscv_vadd_vv_i32m2(vaux_0, vaux_1, vl), vzero, vl);
-            vint32m1_t isum1 = __riscv_vredsum_vs_i32m2_i32m1(__riscv_vadd_vv_i32m2(vaux_2, vaux_3, vl), isum0, vl);
-            vint32m1_t isum2 = __riscv_vredsum_vs_i32m2_i32m1(__riscv_vadd_vv_i32m2(vaux_4, vaux_5, vl), isum1, vl);
-            vint32m1_t isum3 = __riscv_vredsum_vs_i32m2_i32m1(__riscv_vadd_vv_i32m2(vaux_6, vaux_7, vl), isum2, vl);
-
-            sum_t += __riscv_vmv_x_s_i32m1_i32(isum3);
-
-            q6 += 64;   qh += 32;   q8 += 128;   is=8;
-
-        }
-
-        sumf += d * sum_t;
-
-    }
-
-    *s = sumf;
-
-#else
-
-    int8_t  aux8[QK_K];
-    int16_t aux16[8];
-    float   sums [8];
-    int32_t aux32[8];
-    memset(sums, 0, 8*sizeof(float));
-
-    float sumf = 0;
-    for (int i = 0; i < nb; ++i) {
-        const uint8_t * restrict q4 = x[i].ql;
-        const uint8_t * restrict qh = x[i].qh;
-        const  int8_t * restrict q8 = y[i].qs;
-        memset(aux32, 0, 8*sizeof(int32_t));
-        int8_t * restrict a = aux8;
-        for (int j = 0; j < QK_K; j += 128) {
-            for (int l = 0; l < 32; ++l) {
-                a[l +  0] = (int8_t)((q4[l +  0] & 0xF) | (((qh[l] >> 0) & 3) << 4)) - 32;
-                a[l + 32] = (int8_t)((q4[l + 32] & 0xF) | (((qh[l] >> 2) & 3) << 4)) - 32;
-                a[l + 64] = (int8_t)((q4[l +  0] >>  4) | (((qh[l] >> 4) & 3) << 4)) - 32;
-                a[l + 96] = (int8_t)((q4[l + 32] >>  4) | (((qh[l] >> 6) & 3) << 4)) - 32;
-            }
-            a  += 128;
-            q4 += 64;
-            qh += 32;
-        }
-        a = aux8;
-        int is = 0;
-        for (int j = 0; j < QK_K/16; ++j) {
-            int scale = x[i].scales[is++];
-            for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
-            for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
-            q8 += 8; a += 8;
-            for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
-            for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
-            q8 += 8; a += 8;
-        }
-        const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
-        for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l];
-    }
-    for (int l = 0; l < 8; ++l) sumf += sums[l];
-    *s = sumf;
-#endif
-}
-
-#else
-
-void ggml_vec_dot_q6_K_q8_K(const int n, float * restrict s, const void * restrict vx, const void * restrict vy) {
-    assert(n % QK_K == 0);
-
-    const block_q6_K * restrict x = vx;
-    const block_q8_K * restrict y = vy;
-
-    const int nb = n / QK_K;
-
-#ifdef __ARM_NEON
-    float sum = 0;
-
-    const uint8x16_t m4b = vdupq_n_u8(0xF);
-    const int8x16_t  m32s = vdupq_n_s8(32);
-    const int32x4_t  vzero = vdupq_n_s32(0);
-
-    const uint8x16_t mone = vdupq_n_u8(3);
-
-    ggml_int8x16x4_t q6bytes;
-    ggml_uint8x16x4_t q6h;
-
-    for (int i = 0; i < nb; ++i) {
-
-        const float d_all = (float)x[i].d;
-
-        const uint8_t * restrict q6 = x[i].ql;
-        const uint8_t * restrict qh = x[i].qh;
-        const int8_t  * restrict q8 = y[i].qs;
-
-        const int8_t * restrict scale = x[i].scales;
-
-        int32_t isum = 0;
-
-        uint8x16_t qhbits = vld1q_u8(qh);
-        ggml_uint8x16x2_t q6bits = ggml_vld1q_u8_x2(q6);
-        ggml_int8x16x4_t q8bytes = ggml_vld1q_s8_x4(q8);
-
-        q6h.val[0] = vshlq_n_u8(vandq_u8(mone, qhbits), 4);
-        uint8x16_t shifted = vshrq_n_u8(qhbits, 2);
-        q6h.val[1] = vshlq_n_u8(vandq_u8(mone, shifted), 4);
-        shifted = vshrq_n_u8(qhbits, 4);
-        q6h.val[2] = vshlq_n_u8(vandq_u8(mone, shifted), 4);
-        shifted = vshrq_n_u8(qhbits, 6);
-        q6h.val[3] = vshlq_n_u8(vandq_u8(mone, shifted), 4);
-
-        q6bytes.val[0] = vsubq_s8(vreinterpretq_s8_u8(vorrq_u8(vandq_u8(q6bits.val[0], m4b), q6h.val[0])), m32s);
-        q6bytes.val[1] = vsubq_s8(vreinterpretq_s8_u8(vorrq_u8(vandq_u8(q6bits.val[1], m4b), q6h.val[1])), m32s);
-        q6bytes.val[2] = vsubq_s8(vreinterpretq_s8_u8(vorrq_u8(vshrq_n_u8(q6bits.val[0], 4), q6h.val[2])), m32s);
-        q6bytes.val[3] = vsubq_s8(vreinterpretq_s8_u8(vorrq_u8(vshrq_n_u8(q6bits.val[1], 4), q6h.val[3])), m32s);
-
-        isum += vaddvq_s32(ggml_vdotq_s32(vzero, q6bytes.val[0], q8bytes.val[0])) * scale[0] +
-                vaddvq_s32(ggml_vdotq_s32(vzero, q6bytes.val[1], q8bytes.val[1])) * scale[1] +
-                vaddvq_s32(ggml_vdotq_s32(vzero, q6bytes.val[2], q8bytes.val[2])) * scale[2] +
-                vaddvq_s32(ggml_vdotq_s32(vzero, q6bytes.val[3], q8bytes.val[3])) * scale[3];
-
-        sum += isum * d_all * y[i].d;
-
-    }
-    *s = sum;
-
-#elif defined __AVX2__
-
-    const __m256i m4 = _mm256_set1_epi8(0xF);
-    const __m256i m2 = _mm256_set1_epi8(3);
-    const __m256i m32s = _mm256_set1_epi8(32);
-
-    __m256 acc = _mm256_setzero_ps();
-
-    for (int i = 0; i < nb; ++i) {
-
-        const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d);
-
-        const uint8_t * restrict q4 = x[i].ql;
-        const uint8_t * restrict qh = x[i].qh;
-        const int8_t  * restrict q8 = y[i].qs;
-
-        const __m64 scales_1 = _mm_set1_pi8(x[i].scales[0]);
-        const __m64 scales_2 = _mm_set1_pi8(x[i].scales[1]);
-        const __m64 scales_3 = _mm_set1_pi8(x[i].scales[2]);
-        const __m64 scales_4 = _mm_set1_pi8(x[i].scales[3]);
-
-        __m256i sumi = _mm256_setzero_si256();
-
-        const __m128i scale_0 = _mm_set_epi64(scales_2, scales_1);
-        const __m128i scale_1 = _mm_set_epi64(scales_4, scales_3);
-
-        const __m256i q4bits1 = _mm256_loadu_si256((const __m256i*)q4);
-        const __m128i q4bitsH = _mm_loadu_si128((const __m128i*)qh);
-
-        const __m256i q4h_0 = _mm256_slli_epi16(_mm256_and_si256(MM256_SET_M128I(_mm_srli_epi16(q4bitsH, 2), q4bitsH), m2), 4);
-        const __m256i q4h_1 = _mm256_slli_epi16(_mm256_and_si256(MM256_SET_M128I(_mm_srli_epi16(q4bitsH, 6), _mm_srli_epi16(q4bitsH, 4)), m2), 4);
-
-        const __m256i q4_0 = _mm256_or_si256(_mm256_and_si256(q4bits1, m4), q4h_0);
-        const __m256i q4_1 = _mm256_or_si256(_mm256_and_si256(_mm256_srli_epi16(q4bits1, 4), m4), q4h_1);
-
-        const __m256i q8_0 = _mm256_loadu_si256((const __m256i*)(q8+ 0));
-        const __m256i q8_1 = _mm256_loadu_si256((const __m256i*)(q8+32));
-
-        __m256i q8s_0 = _mm256_maddubs_epi16(m32s, q8_0);
-        __m256i q8s_1 = _mm256_maddubs_epi16(m32s, q8_1);
-
-        __m256i p16_0 = _mm256_maddubs_epi16(q4_0, q8_0);
-        __m256i p16_1 = _mm256_maddubs_epi16(q4_1, q8_1);
-
-        p16_0 = _mm256_sub_epi16(p16_0, q8s_0);
-        p16_1 = _mm256_sub_epi16(p16_1, q8s_1);
-
-        p16_0 = _mm256_madd_epi16(_mm256_cvtepi8_epi16(scale_0), p16_0);
-        p16_1 = _mm256_madd_epi16(_mm256_cvtepi8_epi16(scale_1), p16_1);
-
-        sumi = _mm256_add_epi32(sumi, _mm256_add_epi32(p16_0, p16_1));
-
-        acc = _mm256_fmadd_ps(_mm256_broadcast_ss(&d), _mm256_cvtepi32_ps(sumi), acc);
-    }
-
-    *s = hsum_float_8(acc);
-
-#elif defined __AVX__
-
-    const __m128i m4 = _mm_set1_epi8(0xF);
-    const __m128i m2 = _mm_set1_epi8(3);
-    const __m128i m32s = _mm_set1_epi8(32);
-
-    __m256 acc = _mm256_setzero_ps();
-
-    for (int i = 0; i < nb; ++i) {
-
-        const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d);
-
-        const uint8_t * restrict q4 = x[i].ql;
-        const uint8_t * restrict qh = x[i].qh;
-        const int8_t  * restrict q8 = y[i].qs;
-
-        const __m64 scales_1 = _mm_set1_pi8(x[i].scales[0]);
-        const __m64 scales_2 = _mm_set1_pi8(x[i].scales[1]);
-        const __m64 scales_3 = _mm_set1_pi8(x[i].scales[2]);
-        const __m64 scales_4 = _mm_set1_pi8(x[i].scales[3]);
-
-        __m128i sumi_0 = _mm_setzero_si128();
-        __m128i sumi_1 = _mm_setzero_si128();
-
-        const __m128i scale_0 = _mm_set_epi64(scales_2, scales_1);
-        const __m128i scale_1 = _mm_set_epi64(scales_4, scales_3);
-
-        const __m256i q4bits1 = _mm256_loadu_si256((const __m256i*)q4);
-        const __m128i q4bitsH = _mm_loadu_si128((const __m128i*)qh);
-
-        const __m128i q4h_0 = _mm_slli_epi16(_mm_and_si128(q4bitsH, m2), 4);
-        const __m128i q4h_1 = _mm_slli_epi16(_mm_and_si128(_mm_srli_epi16(q4bitsH, 2), m2), 4);
-        const __m128i q4h_2 = _mm_slli_epi16(_mm_and_si128(_mm_srli_epi16(q4bitsH, 4), m2), 4);
-        const __m128i q4h_3 = _mm_slli_epi16(_mm_and_si128(_mm_srli_epi16(q4bitsH, 6), m2), 4);
-
-        const __m128i q4_0 = _mm_or_si128(_mm_and_si128(_mm256_extractf128_si256(q4bits1, 0), m4), q4h_0);
-        const __m128i q4_1 = _mm_or_si128(_mm_and_si128(_mm256_extractf128_si256(q4bits1, 1), m4), q4h_1);
-        const __m128i q4_2 = _mm_or_si128(_mm_and_si128(_mm_srli_epi16(_mm256_extractf128_si256(q4bits1, 0), 4), m4), q4h_2);
-        const __m128i q4_3 = _mm_or_si128(_mm_and_si128(_mm_srli_epi16(_mm256_extractf128_si256(q4bits1, 1), 4), m4), q4h_3);
-
-        const __m256i q8_0 = _mm256_loadu_si256((const __m256i*)(q8+ 0));
-        const __m256i q8_1 = _mm256_loadu_si256((const __m256i*)(q8+32));
-
-        __m128i q8s_0 = _mm_maddubs_epi16(m32s, _mm256_extractf128_si256(q8_0, 0));
-        __m128i q8s_1 = _mm_maddubs_epi16(m32s, _mm256_extractf128_si256(q8_0, 1));
-        __m128i q8s_2 = _mm_maddubs_epi16(m32s, _mm256_extractf128_si256(q8_1, 0));
-        __m128i q8s_3 = _mm_maddubs_epi16(m32s, _mm256_extractf128_si256(q8_1, 1));
-
-        __m128i p16_0 = _mm_maddubs_epi16(q4_0, _mm256_extractf128_si256(q8_0, 0));
-        __m128i p16_1 = _mm_maddubs_epi16(q4_1, _mm256_extractf128_si256(q8_0, 1));
-        __m128i p16_2 = _mm_maddubs_epi16(q4_2, _mm256_extractf128_si256(q8_1, 0));
-        __m128i p16_3 = _mm_maddubs_epi16(q4_3, _mm256_extractf128_si256(q8_1, 1));
-
-        p16_0 = _mm_sub_epi16(p16_0, q8s_0);
-        p16_1 = _mm_sub_epi16(p16_1, q8s_1);
-        p16_2 = _mm_sub_epi16(p16_2, q8s_2);
-        p16_3 = _mm_sub_epi16(p16_3, q8s_3);
-
-        p16_0 = _mm_madd_epi16(_mm_cvtepi8_epi16(scale_0), p16_0);
-        p16_1 = _mm_madd_epi16(_mm_cvtepi8_epi16(_mm_unpackhi_epi64(scale_0, scale_0)), p16_1);
-        p16_2 = _mm_madd_epi16(_mm_cvtepi8_epi16(scale_1), p16_2);
-        p16_3 = _mm_madd_epi16(_mm_cvtepi8_epi16(_mm_unpackhi_epi64(scale_1, scale_1)), p16_3);
-
-        sumi_0 = _mm_add_epi32(sumi_0, _mm_add_epi32(p16_0, p16_2));
-        sumi_1 = _mm_add_epi32(sumi_1, _mm_add_epi32(p16_1, p16_3));
-
-        acc = _mm256_add_ps(_mm256_mul_ps(_mm256_broadcast_ss(&d), _mm256_cvtepi32_ps(MM256_SET_M128I(sumi_1, sumi_0))), acc);
-    }
-
-    *s = hsum_float_8(acc);
-
-#elif defined __riscv_v_intrinsic
-
-    float sumf = 0;
-
-    for (int i = 0; i < nb; ++i) {
-
-        const float d_all = (float)x[i].d;
-
-        const uint8_t * restrict q6 = x[i].ql;
-        const uint8_t * restrict qh = x[i].qh;
-        const int8_t  * restrict q8 = y[i].qs;
-
-        const int8_t * restrict scale = x[i].scales;
-
-        int32_t isum = 0;
-
-        size_t vl = 16;
-
-        vint32m1_t vzero = __riscv_vmv_v_x_i32m1(0, 1);
-
-        // load Q6
-        vuint8mf2_t q6_0 = __riscv_vle8_v_u8mf2(q6, vl);
-        vuint8mf2_t q6_1 = __riscv_vle8_v_u8mf2(q6+16, vl);
-
-        // load qh
-        vuint8mf2_t qh_x = __riscv_vle8_v_u8mf2(qh, vl);
-
-        vuint8mf2_t qh0 = __riscv_vsll_vx_u8mf2(__riscv_vand_vx_u8mf2(qh_x, 0x3, vl), 0x4, vl);
-        qh_x = __riscv_vsrl_vx_u8mf2(qh_x, 0x2, vl);
-        vuint8mf2_t qh1 = __riscv_vsll_vx_u8mf2(__riscv_vand_vx_u8mf2(qh_x, 0x3, vl), 0x4, vl);
-        qh_x = __riscv_vsrl_vx_u8mf2(qh_x, 0x2, vl);
-        vuint8mf2_t qh2 = __riscv_vsll_vx_u8mf2(__riscv_vand_vx_u8mf2(qh_x, 0x3, vl), 0x4, vl);
-        qh_x = __riscv_vsrl_vx_u8mf2(qh_x, 0x2, vl);
-        vuint8mf2_t qh3 = __riscv_vsll_vx_u8mf2(__riscv_vand_vx_u8mf2(qh_x, 0x3, vl), 0x4, vl);
-
-        vuint8mf2_t q6h_0 = __riscv_vor_vv_u8mf2(__riscv_vand_vx_u8mf2(q6_0, 0xF, vl), qh0, vl);
-        vuint8mf2_t q6h_1 = __riscv_vor_vv_u8mf2(__riscv_vand_vx_u8mf2(q6_1, 0xF, vl), qh1, vl);
-        vuint8mf2_t q6h_2 = __riscv_vor_vv_u8mf2(__riscv_vsrl_vx_u8mf2(q6_0, 0x4, vl), qh2, vl);
-        vuint8mf2_t q6h_3 = __riscv_vor_vv_u8mf2(__riscv_vsrl_vx_u8mf2(q6_1, 0x4, vl), qh3, vl);
-
-        vint8mf2_t q6v_0 = __riscv_vsub_vx_i8mf2(__riscv_vreinterpret_v_u8mf2_i8mf2(q6h_0), 32, vl);
-        vint8mf2_t q6v_1 = __riscv_vsub_vx_i8mf2(__riscv_vreinterpret_v_u8mf2_i8mf2(q6h_1), 32, vl);
-        vint8mf2_t q6v_2 = __riscv_vsub_vx_i8mf2(__riscv_vreinterpret_v_u8mf2_i8mf2(q6h_2), 32, vl);
-        vint8mf2_t q6v_3 = __riscv_vsub_vx_i8mf2(__riscv_vreinterpret_v_u8mf2_i8mf2(q6h_3), 32, vl);
-
-        // load Q8 and take product
-        vint16m1_t p0 = __riscv_vwmul_vv_i16m1(q6v_0, __riscv_vle8_v_i8mf2(q8, vl), vl);
-        vint16m1_t p1 = __riscv_vwmul_vv_i16m1(q6v_1, __riscv_vle8_v_i8mf2(q8+16, vl), vl);
-        vint16m1_t p2 = __riscv_vwmul_vv_i16m1(q6v_2, __riscv_vle8_v_i8mf2(q8+32, vl), vl);
-        vint16m1_t p3 = __riscv_vwmul_vv_i16m1(q6v_3, __riscv_vle8_v_i8mf2(q8+48, vl), vl);
-
-        vint32m1_t vs_0 = __riscv_vwredsum_vs_i16m1_i32m1(p0, vzero, vl);
-        vint32m1_t vs_1 = __riscv_vwredsum_vs_i16m1_i32m1(p1, vzero, vl);
-        vint32m1_t vs_2 = __riscv_vwredsum_vs_i16m1_i32m1(p2, vzero, vl);
-        vint32m1_t vs_3 = __riscv_vwredsum_vs_i16m1_i32m1(p3, vzero, vl);
-
-        isum += __riscv_vmv_x_s_i32m1_i32(vs_0) * scale[0];
-        isum += __riscv_vmv_x_s_i32m1_i32(vs_1) * scale[1];
-        isum += __riscv_vmv_x_s_i32m1_i32(vs_2) * scale[2];
-        isum += __riscv_vmv_x_s_i32m1_i32(vs_3) * scale[3];
-
-        sumf += isum * d_all * y[i].d;
-
-    }
-
-    *s = sumf;
-
-#else
-
-    int8_t  aux8[QK_K];
-    int16_t aux16[8];
-    float   sums [8];
-    int32_t aux32[8];
-    memset(sums, 0, 8*sizeof(float));
-
-    float sumf = 0;
-    for (int i = 0; i < nb; ++i) {
-        const uint8_t * restrict q4 = x[i].ql;
-        const uint8_t * restrict qh = x[i].qh;
-        const  int8_t * restrict q8 = y[i].qs;
-        memset(aux32, 0, 8*sizeof(int32_t));
-        int8_t * restrict a = aux8;
-        for (int l = 0; l < 16; ++l) {
-            a[l+ 0] = (int8_t)((q4[l+ 0] & 0xF) | (((qh[l] >> 0) & 3) << 4)) - 32;
-            a[l+16] = (int8_t)((q4[l+16] & 0xF) | (((qh[l] >> 2) & 3) << 4)) - 32;
-            a[l+32] = (int8_t)((q4[l+ 0] >>  4) | (((qh[l] >> 4) & 3) << 4)) - 32;
-            a[l+48] = (int8_t)((q4[l+16] >>  4) | (((qh[l] >> 6) & 3) << 4)) - 32;
-        }
-        int is = 0;
-        for (int j = 0; j < QK_K/16; ++j) {
-            int scale = x[i].scales[is++];
-            for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
-            for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
-            q8 += 8; a += 8;
-            for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
-            for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
-            q8 += 8; a += 8;
-        }
-        const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
-        for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l];
-    }
-    for (int l = 0; l < 8; ++l) sumf += sums[l];
-    *s = sumf;
-#endif
-}
-
-#endif
diff --git a/src/whisper_cpp/ggml-quants.h b/src/whisper_cpp/ggml-quants.h
deleted file mode 100644
index 62c1df6c..00000000
--- a/src/whisper_cpp/ggml-quants.h
+++ /dev/null
@@ -1,224 +0,0 @@
-#pragma once
-
-#include "ggml-impl.h"
-
-// GGML internal header
-
-#include <stdint.h>
-#include <stddef.h>
-
-#define QK4_0 32
-typedef struct {
-    ggml_fp16_t d;          // delta
-    uint8_t qs[QK4_0 / 2];  // nibbles / quants
-} block_q4_0;
-static_assert(sizeof(block_q4_0) == sizeof(ggml_fp16_t) + QK4_0 / 2, "wrong q4_0 block size/padding");
-
-#define QK4_1 32
-typedef struct {
-    ggml_fp16_t d;          // delta
-    ggml_fp16_t m;          // min
-    uint8_t qs[QK4_1 / 2];  // nibbles / quants
-} block_q4_1;
-static_assert(sizeof(block_q4_1) == 2 * sizeof(ggml_fp16_t) + QK4_1 / 2, "wrong q4_1 block size/padding");
-
-#define QK5_0 32
-typedef struct {
-    ggml_fp16_t d;         // delta
-    uint8_t qh[4];         // 5-th bit of quants
-    uint8_t qs[QK5_0 / 2]; // nibbles / quants
-} block_q5_0;
-static_assert(sizeof(block_q5_0) == sizeof(ggml_fp16_t) + sizeof(uint32_t) + QK5_0 / 2, "wrong q5_0 block size/padding");
-
-#define QK5_1 32
-typedef struct {
-    ggml_fp16_t d;         // delta
-    ggml_fp16_t m;         // min
-    uint8_t qh[4];         // 5-th bit of quants
-    uint8_t qs[QK5_1 / 2]; // nibbles / quants
-} block_q5_1;
-static_assert(sizeof(block_q5_1) == 2 * sizeof(ggml_fp16_t) + sizeof(uint32_t) + QK5_1 / 2, "wrong q5_1 block size/padding");
-
-#define QK8_0 32
-typedef struct {
-    ggml_fp16_t d;         // delta
-    int8_t  qs[QK8_0];     // quants
-} block_q8_0;
-static_assert(sizeof(block_q8_0) == sizeof(ggml_fp16_t) + QK8_0, "wrong q8_0 block size/padding");
-
-#define QK8_1 32
-typedef struct {
-    float d;               // delta
-    float s;               // d * sum(qs[i])
-    int8_t  qs[QK8_1];     // quants
-} block_q8_1;
-static_assert(sizeof(block_q8_1) == 2*sizeof(float) + QK8_1, "wrong q8_1 block size/padding");
-
-//
-// Super-block quantization structures
-//
-
-// Super-block size
-#ifdef GGML_QKK_64
-#define QK_K 64
-#define K_SCALE_SIZE 4
-#else
-#define QK_K 256
-#define K_SCALE_SIZE 12
-#endif
-
-// 2-bit quantization
-// weight is represented as x = a * q + b
-// 16 blocks of 16 elements each
-// Effectively 2.625 bits per weight
-typedef struct {
-    uint8_t scales[QK_K/16]; // scales and mins, quantized with 4 bits
-    uint8_t qs[QK_K/4];      // quants
-    ggml_fp16_t d;           // super-block scale for quantized scales
-    ggml_fp16_t dmin;        // super-block scale for quantized mins
-} block_q2_K;
-static_assert(sizeof(block_q2_K) == 2*sizeof(ggml_fp16_t) + QK_K/16 + QK_K/4, "wrong q2_K block size/padding");
-
-// 3-bit quantization
-// weight is represented as x = a * q
-// 16 blocks of 16 elements each
-// Effectively 3.4375 bits per weight
-#ifdef GGML_QKK_64
-typedef struct {
-    uint8_t hmask[QK_K/8];     // quants - high bit
-    uint8_t qs[QK_K/4];        // quants - low 2 bits
-    uint8_t scales[2];
-    ggml_fp16_t d;             // super-block scale
-} block_q3_K;
-static_assert(sizeof(block_q3_K) == sizeof(ggml_fp16_t) + QK_K / 4 + QK_K / 8 + 2, "wrong q3_K block size/padding");
-#else
-typedef struct {
-    uint8_t hmask[QK_K/8];     // quants - high bit
-    uint8_t qs[QK_K/4];        // quants - low 2 bits
-    uint8_t scales[12];        // scales, quantized with 6 bits
-    ggml_fp16_t d;             // super-block scale
-} block_q3_K;
-static_assert(sizeof(block_q3_K) == sizeof(ggml_fp16_t) + QK_K / 4 + QK_K / 8 + 12, "wrong q3_K block size/padding");
-#endif
-
-// 4-bit quantization
-// 8 blocks of 32 elements each
-// weight is represented as x = a * q + b
-// Effectively 4.5 bits per weight
-#ifdef GGML_QKK_64
-typedef struct {
-    ggml_fp16_t d[2];          // super-block scales/mins
-    uint8_t scales[2];         // 4-bit block scales/mins
-    uint8_t qs[QK_K/2];        // 4--bit quants
-} block_q4_K;
-static_assert(sizeof(block_q4_K) == 2*sizeof(ggml_fp16_t) + QK_K/2 + 2, "wrong q4_K block size/padding");
-#else
-typedef struct {
-    ggml_fp16_t d;             // super-block scale for quantized scales
-    ggml_fp16_t dmin;          // super-block scale for quantized mins
-    uint8_t scales[K_SCALE_SIZE]; // scales and mins, quantized with 6 bits
-    uint8_t qs[QK_K/2];        // 4--bit quants
-} block_q4_K;
-static_assert(sizeof(block_q4_K) == 2*sizeof(ggml_fp16_t) + K_SCALE_SIZE + QK_K/2, "wrong q4_K block size/padding");
-#endif
-
-// 5-bit quantization
-// 8 blocks of 32 elements each
-// weight is represented as x = a * q + b
-// Effectively 5.5 bits per weight
-#ifdef GGML_QKK_64
-typedef struct {
-    ggml_fp16_t d;               // super-block scale
-    int8_t  scales[QK_K/16];     // 8-bit block scales
-    uint8_t qh[QK_K/8];          // quants, high bit
-    uint8_t qs[QK_K/2];          // quants, low 4 bits
-} block_q5_K;
-static_assert(sizeof(block_q5_K) == sizeof(ggml_fp16_t) + QK_K/2 + QK_K/8 + QK_K/16, "wrong q5_K block size/padding");
-#else
-typedef struct {
-    ggml_fp16_t d;               // super-block scale for quantized scales
-    ggml_fp16_t dmin;            // super-block scale for quantized mins
-    uint8_t scales[K_SCALE_SIZE];   // scales and mins, quantized with 6 bits
-    uint8_t qh[QK_K/8];          // quants, high bit
-    uint8_t qs[QK_K/2];          // quants, low 4 bits
-} block_q5_K;
-static_assert(sizeof(block_q5_K) == 2*sizeof(ggml_fp16_t) + K_SCALE_SIZE + QK_K/2 + QK_K/8, "wrong q5_K block size/padding");
-#endif
-
-// 6-bit quantization
-// weight is represented as x = a * q
-// 16 blocks of 16 elements each
-// Effectively 6.5625 bits per weight
-typedef struct {
-    uint8_t ql[QK_K/2];      // quants, lower 4 bits
-    uint8_t qh[QK_K/4];      // quants, upper 2 bits
-    int8_t  scales[QK_K/16]; // scales, quantized with 8 bits
-    ggml_fp16_t d;           // super-block scale
-} block_q6_K;
-static_assert(sizeof(block_q6_K) == sizeof(ggml_fp16_t) + QK_K / 16 + 3*QK_K/4, "wrong q6_K block size/padding");
-
-// This is only used for intermediate quantization and dot products
-typedef struct {
-    float   d;              // delta
-    int8_t  qs[QK_K];       // quants
-    int16_t bsums[QK_K/16]; // sum of quants in groups of 16
-} block_q8_K;
-static_assert(sizeof(block_q8_K) == sizeof(float) + QK_K + QK_K/16*sizeof(int16_t), "wrong q8_K block size/padding");
-
-
-// Quantization
-void quantize_row_q4_0_reference(const float * restrict x, block_q4_0 * restrict y, int k);
-void quantize_row_q4_1_reference(const float * restrict x, block_q4_1 * restrict y, int k);
-void quantize_row_q5_0_reference(const float * restrict x, block_q5_0 * restrict y, int k);
-void quantize_row_q5_1_reference(const float * restrict x, block_q5_1 * restrict y, int k);
-void quantize_row_q8_0_reference(const float * restrict x, block_q8_0 * restrict y, int k);
-void quantize_row_q8_1_reference(const float * restrict x, block_q8_1 * restrict y, int k);
-
-void quantize_row_q2_K_reference(const float * restrict x, block_q2_K * restrict y, int k);
-void quantize_row_q3_K_reference(const float * restrict x, block_q3_K * restrict y, int k);
-void quantize_row_q4_K_reference(const float * restrict x, block_q4_K * restrict y, int k);
-void quantize_row_q5_K_reference(const float * restrict x, block_q5_K * restrict y, int k);
-void quantize_row_q6_K_reference(const float * restrict x, block_q6_K * restrict y, int k);
-void quantize_row_q8_K_reference(const float * restrict x, block_q8_K * restrict y, int k);
-
-void quantize_row_q4_0(const float * restrict x, void * restrict y, int k);
-void quantize_row_q4_1(const float * restrict x, void * restrict y, int k);
-void quantize_row_q5_0(const float * restrict x, void * restrict y, int k);
-void quantize_row_q5_1(const float * restrict x, void * restrict y, int k);
-void quantize_row_q8_0(const float * restrict x, void * restrict y, int k);
-void quantize_row_q8_1(const float * restrict x, void * restrict y, int k);
-
-void quantize_row_q2_K(const float * restrict x, void * restrict y, int k);
-void quantize_row_q3_K(const float * restrict x, void * restrict y, int k);
-void quantize_row_q4_K(const float * restrict x, void * restrict y, int k);
-void quantize_row_q5_K(const float * restrict x, void * restrict y, int k);
-void quantize_row_q6_K(const float * restrict x, void * restrict y, int k);
-void quantize_row_q8_K(const float * restrict x, void * restrict y, int k);
-
-// Dequantization
-void dequantize_row_q4_0(const block_q4_0 * restrict x, float * restrict y, int k);
-void dequantize_row_q4_1(const block_q4_1 * restrict x, float * restrict y, int k);
-void dequantize_row_q5_0(const block_q5_0 * restrict x, float * restrict y, int k);
-void dequantize_row_q5_1(const block_q5_1 * restrict x, float * restrict y, int k);
-void dequantize_row_q8_0(const block_q8_0 * restrict x, float * restrict y, int k);
-//void dequantize_row_q8_1(const block_q8_1 * restrict x, float * restrict y, int k);
-
-void dequantize_row_q2_K(const block_q2_K * restrict x, float * restrict y, int k);
-void dequantize_row_q3_K(const block_q3_K * restrict x, float * restrict y, int k);
-void dequantize_row_q4_K(const block_q4_K * restrict x, float * restrict y, int k);
-void dequantize_row_q5_K(const block_q5_K * restrict x, float * restrict y, int k);
-void dequantize_row_q6_K(const block_q6_K * restrict x, float * restrict y, int k);
-void dequantize_row_q8_K(const block_q8_K * restrict x, float * restrict y, int k);
-
-// Dot product
-void ggml_vec_dot_q4_0_q8_0(int n, float * restrict s, const void * restrict vx, const void * restrict vy);
-void ggml_vec_dot_q4_1_q8_1(int n, float * restrict s, const void * restrict vx, const void * restrict vy);
-void ggml_vec_dot_q5_0_q8_0(int n, float * restrict s, const void * restrict vx, const void * restrict vy);
-void ggml_vec_dot_q5_1_q8_1(int n, float * restrict s, const void * restrict vx, const void * restrict vy);
-void ggml_vec_dot_q8_0_q8_0(int n, float * restrict s, const void * restrict vx, const void * restrict vy);
-
-void ggml_vec_dot_q2_K_q8_K(int n, float * restrict s, const void * restrict vx, const void * restrict vy);
-void ggml_vec_dot_q3_K_q8_K(int n, float * restrict s, const void * restrict vx, const void * restrict vy);
-void ggml_vec_dot_q4_K_q8_K(int n, float * restrict s, const void * restrict vx, const void * restrict vy);
-void ggml_vec_dot_q5_K_q8_K(int n, float * restrict s, const void * restrict vx, const void * restrict vy);
-void ggml_vec_dot_q6_K_q8_K(int n, float * restrict s, const void * restrict vx, const void * restrict vy);
diff --git a/src/whisper_cpp/ggml/.gitignore b/src/whisper_cpp/ggml/.gitignore
new file mode 100644
index 00000000..9dd9ddea
--- /dev/null
+++ b/src/whisper_cpp/ggml/.gitignore
@@ -0,0 +1 @@
+src/ggml-metal-embed.metal
diff --git a/src/whisper_cpp/ggml/CMakeLists.txt b/src/whisper_cpp/ggml/CMakeLists.txt
new file mode 100644
index 00000000..89fdf9d1
--- /dev/null
+++ b/src/whisper_cpp/ggml/CMakeLists.txt
@@ -0,0 +1,264 @@
+cmake_minimum_required(VERSION 3.14) # for add_link_options and implicit target directories.
+project("ggml" C CXX)
+include(CheckIncludeFileCXX)
+
+set(CMAKE_EXPORT_COMPILE_COMMANDS ON)
+
+if (NOT XCODE AND NOT MSVC AND NOT CMAKE_BUILD_TYPE)
+    set(CMAKE_BUILD_TYPE Release CACHE STRING "Build type" FORCE)
+    set_property(CACHE CMAKE_BUILD_TYPE PROPERTY STRINGS "Debug" "Release" "MinSizeRel" "RelWithDebInfo")
+endif()
+
+if (CMAKE_SOURCE_DIR STREQUAL CMAKE_CURRENT_SOURCE_DIR)
+    set(GGML_STANDALONE ON)
+
+    set(CMAKE_RUNTIME_OUTPUT_DIRECTORY ${CMAKE_BINARY_DIR}/bin)
+
+    # configure project version
+    # TODO
+else()
+    set(GGML_STANDALONE OFF)
+endif()
+
+if (EMSCRIPTEN)
+    set(BUILD_SHARED_LIBS_DEFAULT OFF)
+
+    option(GGML_WASM_SINGLE_FILE "ggml: embed WASM inside the generated ggml.js" ON)
+else()
+    if (MINGW)
+        set(BUILD_SHARED_LIBS_DEFAULT OFF)
+    else()
+        set(BUILD_SHARED_LIBS_DEFAULT ON)
+    endif()
+endif()
+
+option(BUILD_SHARED_LIBS "ggml: build shared libraries" ${BUILD_SHARED_LIBS_DEFAULT})
+
+#
+# option list
+#
+
+# TODO: mark all options as advanced when not GGML_STANDALONE
+
+if (APPLE)
+    set(GGML_METAL_DEFAULT ON)
+    set(GGML_BLAS_DEFAULT ON)
+    set(GGML_BLAS_VENDOR_DEFAULT "Apple")
+else()
+    set(GGML_METAL_DEFAULT OFF)
+    set(GGML_BLAS_DEFAULT OFF)
+    set(GGML_BLAS_VENDOR_DEFAULT "Generic")
+endif()
+
+if (CMAKE_CROSSCOMPILING)
+    set(GGML_NATIVE_DEFAULT OFF)
+else()
+    set(GGML_NATIVE_DEFAULT ON)
+endif()
+
+# defaults
+if (NOT GGML_LLAMAFILE_DEFAULT)
+    set(GGML_LLAMAFILE_DEFAULT OFF)
+endif()
+
+if (NOT GGML_CUDA_GRAPHS_DEFAULT)
+    set(GGML_CUDA_GRAPHS_DEFAULT OFF)
+endif()
+
+# general
+option(GGML_STATIC "ggml: static link libraries"         OFF)
+option(GGML_NATIVE "ggml: enable -march=native flag"     ${GGML_NATIVE_DEFAULT})
+option(GGML_LTO    "ggml: enable link time optimization" OFF)
+option(GGML_CCACHE "ggml: use ccache if available"       ON)
+
+# debug
+option(GGML_ALL_WARNINGS           "ggml: enable all compiler warnings"                   ON)
+option(GGML_ALL_WARNINGS_3RD_PARTY "ggml: enable all compiler warnings in 3rd party libs" OFF)
+option(GGML_GPROF                  "ggml: enable gprof"                                   OFF)
+
+# build
+option(GGML_FATAL_WARNINGS    "ggml: enable -Werror flag"    OFF)
+
+# sanitizers
+option(GGML_SANITIZE_THREAD    "ggml: enable thread sanitizer"    OFF)
+option(GGML_SANITIZE_ADDRESS   "ggml: enable address sanitizer"   OFF)
+option(GGML_SANITIZE_UNDEFINED "ggml: enable undefined sanitizer" OFF)
+
+# instruction set specific
+if (GGML_NATIVE OR NOT GGML_NATIVE_DEFAULT)
+    set(INS_ENB OFF)
+else()
+    set(INS_ENB ON)
+endif()
+
+option(GGML_CPU_HBM     "ggml: use memkind for CPU HBM" OFF)
+
+option(GGML_AVX         "ggml: enable AVX"              ${INS_ENB})
+option(GGML_AVX2        "ggml: enable AVX2"             ${INS_ENB})
+option(GGML_AVX512      "ggml: enable AVX512"           OFF)
+option(GGML_AVX512_VBMI "ggml: enable AVX512-VBMI"      OFF)
+option(GGML_AVX512_VNNI "ggml: enable AVX512-VNNI"      OFF)
+option(GGML_AVX512_BF16 "ggml: enable AVX512-BF16"      OFF)
+option(GGML_FMA         "ggml: enable FMA"              ${INS_ENB})
+if (NOT MSVC)
+    option(GGML_F16C    "ggml: enable F16C"             ${INS_ENB}) # in MSVC F16C is implied with AVX2/AVX512
+endif()
+option(GGML_LASX        "ggml: enable lasx"             ON)
+option(GGML_LSX         "ggml: enable lsx"              ON)
+option(GGML_SVE         "ggml: enable SVE"              OFF)
+
+if (WIN32)
+    set(GGML_WIN_VER "0x602" CACHE STRING "ggml: Windows Version")
+endif()
+
+# ggml core
+set(GGML_SCHED_MAX_COPIES  "4" CACHE STRING "ggml: max input copies for pipeline parallelism")
+
+# 3rd party libs / backends
+option(GGML_ACCELERATE                      "ggml: enable Accelerate framework"               ON)
+option(GGML_BLAS                            "ggml: use BLAS"                                  ${GGML_BLAS_DEFAULT})
+set(GGML_BLAS_VENDOR ${GGML_BLAS_VENDOR_DEFAULT} CACHE STRING
+                                            "ggml: BLAS library vendor")
+option(GGML_LLAMAFILE                       "ggml: use LLAMAFILE"                             ${GGML_LLAMAFILE_DEFAULT})
+
+option(GGML_CUDA                            "ggml: use CUDA"                                  OFF)
+option(GGML_MUSA                            "ggml: use MUSA"                                  OFF)
+option(GGML_CUDA_FORCE_DMMV                 "ggml: use dmmv instead of mmvq CUDA kernels"     OFF)
+option(GGML_CUDA_FORCE_MMQ                  "ggml: use mmq kernels instead of cuBLAS"         OFF)
+option(GGML_CUDA_FORCE_CUBLAS               "ggml: always use cuBLAS instead of mmq kernels"  OFF)
+set   (GGML_CUDA_DMMV_X   "32" CACHE STRING "ggml: x stride for dmmv CUDA kernels")
+set   (GGML_CUDA_MMV_Y     "1" CACHE STRING "ggml: y block size for mmv CUDA kernels")
+option(GGML_CUDA_F16                        "ggml: use 16 bit floats for some calculations"   OFF)
+set   (GGML_CUDA_KQUANTS_ITER "2" CACHE STRING
+                                            "ggml: iters./thread per block for Q2_K/Q6_K")
+set   (GGML_CUDA_PEER_MAX_BATCH_SIZE "128" CACHE STRING
+                                            "ggml: max. batch size for using peer access")
+option(GGML_CUDA_NO_PEER_COPY               "ggml: do not use peer to peer copies"            OFF)
+option(GGML_CUDA_NO_VMM                     "ggml: do not try to use CUDA VMM"                OFF)
+option(GGML_CUDA_FA_ALL_QUANTS              "ggml: compile all quants for FlashAttention"     OFF)
+option(GGML_CUDA_GRAPHS                     "ggml: use CUDA graphs (llama.cpp only)"          ${GGML_CUDA_GRAPHS_DEFAULT})
+
+option(GGML_HIPBLAS                         "ggml: use hipBLAS"                               OFF)
+option(GGML_HIP_UMA                         "ggml: use HIP unified memory architecture"       OFF)
+option(GGML_VULKAN                          "ggml: use Vulkan"                                OFF)
+option(GGML_VULKAN_CHECK_RESULTS            "ggml: run Vulkan op checks"                      OFF)
+option(GGML_VULKAN_DEBUG                    "ggml: enable Vulkan debug output"                OFF)
+option(GGML_VULKAN_MEMORY_DEBUG             "ggml: enable Vulkan memory debug output"         OFF)
+option(GGML_VULKAN_SHADER_DEBUG_INFO        "ggml: enable Vulkan shader debug info"           OFF)
+option(GGML_VULKAN_PERF                     "ggml: enable Vulkan perf output"                 OFF)
+option(GGML_VULKAN_VALIDATE                 "ggml: enable Vulkan validation"                  OFF)
+option(GGML_VULKAN_RUN_TESTS                "ggml: run Vulkan tests"                          OFF)
+option(GGML_KOMPUTE                         "ggml: use Kompute"                               OFF)
+option(GGML_METAL                           "ggml: use Metal"                                 ${GGML_METAL_DEFAULT})
+option(GGML_METAL_NDEBUG                    "ggml: disable Metal debugging"                   OFF)
+option(GGML_METAL_SHADER_DEBUG              "ggml: compile Metal with -fno-fast-math"         OFF)
+option(GGML_METAL_EMBED_LIBRARY             "ggml: embed Metal library"                       ${GGML_METAL})
+set   (GGML_METAL_MACOSX_VERSION_MIN "" CACHE STRING
+                                            "ggml: metal minimum macOS version")
+set   (GGML_METAL_STD "" CACHE STRING       "ggml: metal standard version (-std flag)")
+option(GGML_OPENMP                          "ggml: use OpenMP"                                ON)
+option(GGML_RPC                             "ggml: use RPC"                                   OFF)
+option(GGML_SYCL                            "ggml: use SYCL"                                  OFF)
+option(GGML_SYCL_F16                        "ggml: use 16 bit floats for sycl calculations"   OFF)
+set   (GGML_SYCL_TARGET "INTEL" CACHE STRING
+                                            "ggml: sycl target device")
+
+# extra artifacts
+option(GGML_BUILD_TESTS    "ggml: build tests"    ${GGML_STANDALONE})
+option(GGML_BUILD_EXAMPLES "ggml: build examples" ${GGML_STANDALONE})
+
+#
+# dependencies
+#
+
+set(CMAKE_C_STANDARD 11)
+set(CMAKE_C_STANDARD_REQUIRED true)
+
+if (GGML_SYCL)
+    set(CMAKE_CXX_STANDARD 17)
+else()
+    set(CMAKE_CXX_STANDARD 11)
+endif()
+set(CMAKE_CXX_STANDARD_REQUIRED true)
+
+set(THREADS_PREFER_PTHREAD_FLAG ON)
+
+find_package(Threads REQUIRED)
+
+#
+# build the library
+#
+
+add_subdirectory(src)
+
+#
+# tests and examples
+#
+
+if (GGML_BUILD_TESTS)
+    enable_testing()
+    add_subdirectory(tests)
+endif ()
+
+if (GGML_BUILD_EXAMPLES)
+    add_subdirectory(examples)
+endif ()
+
+#
+# install
+#
+
+include(GNUInstallDirs)
+include(CMakePackageConfigHelpers)
+
+# all public headers
+set(GGML_PUBLIC_HEADERS
+    include/ggml.h
+    include/ggml-alloc.h
+    include/ggml-backend.h
+    include/ggml-blas.h
+    include/ggml-cann.h
+    include/ggml-cuda.h
+    include/ggml.h
+    include/ggml-kompute.h
+    include/ggml-metal.h
+    include/ggml-rpc.h
+    include/ggml-sycl.h
+    include/ggml-vulkan.h)
+
+set_target_properties(ggml PROPERTIES PUBLIC_HEADER "${GGML_PUBLIC_HEADERS}")
+#if (GGML_METAL)
+#    set_target_properties(ggml PROPERTIES RESOURCE "${CMAKE_CURRENT_SOURCE_DIR}/src/ggml-metal.metal")
+#endif()
+install(TARGETS ggml PUBLIC_HEADER)
+
+if (BUILD_SHARED_LIBS)
+    install(TARGETS ggml LIBRARY)
+endif()
+
+if (GGML_METAL)
+    install(
+        FILES src/ggml-metal.metal
+        PERMISSIONS
+            OWNER_READ
+            OWNER_WRITE
+            GROUP_READ
+            WORLD_READ
+        DESTINATION ${CMAKE_INSTALL_BINDIR})
+
+    if (NOT GGML_METAL_EMBED_LIBRARY)
+        install(
+            FILES ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/default.metallib
+            DESTINATION ${CMAKE_INSTALL_BINDIR}
+        )
+    endif()
+endif()
+
+if (GGML_STANDALONE)
+    configure_file(${CMAKE_CURRENT_SOURCE_DIR}/ggml.pc.in
+        ${CMAKE_CURRENT_BINARY_DIR}/ggml.pc
+        @ONLY)
+
+    install(FILES ${CMAKE_CURRENT_BINARY_DIR}/ggml.pc
+        DESTINATION share/pkgconfig)
+endif()
diff --git a/src/whisper_cpp/ggml/cmake/FindSIMD.cmake b/src/whisper_cpp/ggml/cmake/FindSIMD.cmake
new file mode 100644
index 00000000..5533668e
--- /dev/null
+++ b/src/whisper_cpp/ggml/cmake/FindSIMD.cmake
@@ -0,0 +1,100 @@
+include(CheckCSourceRuns)
+
+set(AVX_CODE "
+    #include <immintrin.h>
+    int main()
+    {
+        __m256 a;
+        a = _mm256_set1_ps(0);
+        return 0;
+    }
+")
+
+set(AVX512_CODE "
+    #include <immintrin.h>
+    int main()
+    {
+        __m512i a = _mm512_set_epi8(0, 0, 0, 0, 0, 0, 0, 0,
+                                    0, 0, 0, 0, 0, 0, 0, 0,
+                                    0, 0, 0, 0, 0, 0, 0, 0,
+                                    0, 0, 0, 0, 0, 0, 0, 0,
+                                    0, 0, 0, 0, 0, 0, 0, 0,
+                                    0, 0, 0, 0, 0, 0, 0, 0,
+                                    0, 0, 0, 0, 0, 0, 0, 0,
+                                    0, 0, 0, 0, 0, 0, 0, 0);
+        __m512i b = a;
+        __mmask64 equality_mask = _mm512_cmp_epi8_mask(a, b, _MM_CMPINT_EQ);
+        return 0;
+    }
+")
+
+set(AVX2_CODE "
+    #include <immintrin.h>
+    int main()
+    {
+        __m256i a = {0};
+        a = _mm256_abs_epi16(a);
+        __m256i x;
+        _mm256_extract_epi64(x, 0); // we rely on this in our AVX2 code
+        return 0;
+    }
+")
+
+set(FMA_CODE "
+    #include <immintrin.h>
+    int main()
+    {
+        __m256 acc = _mm256_setzero_ps();
+        const __m256 d = _mm256_setzero_ps();
+        const __m256 p = _mm256_setzero_ps();
+        acc = _mm256_fmadd_ps( d, p, acc );
+        return 0;
+    }
+")
+
+macro(check_sse type flags)
+    set(__FLAG_I 1)
+    set(CMAKE_REQUIRED_FLAGS_SAVE ${CMAKE_REQUIRED_FLAGS})
+    foreach (__FLAG ${flags})
+        if (NOT ${type}_FOUND)
+            set(CMAKE_REQUIRED_FLAGS ${__FLAG})
+            check_c_source_runs("${${type}_CODE}" HAS_${type}_${__FLAG_I})
+            if (HAS_${type}_${__FLAG_I})
+                set(${type}_FOUND TRUE CACHE BOOL "${type} support")
+                set(${type}_FLAGS "${__FLAG}" CACHE STRING "${type} flags")
+            endif()
+            math(EXPR __FLAG_I "${__FLAG_I}+1")
+        endif()
+    endforeach()
+    set(CMAKE_REQUIRED_FLAGS ${CMAKE_REQUIRED_FLAGS_SAVE})
+
+    if (NOT ${type}_FOUND)
+        set(${type}_FOUND FALSE CACHE BOOL "${type} support")
+        set(${type}_FLAGS "" CACHE STRING "${type} flags")
+    endif()
+
+    mark_as_advanced(${type}_FOUND ${type}_FLAGS)
+endmacro()
+
+# flags are for MSVC only!
+check_sse("AVX" " ;/arch:AVX")
+if (NOT ${AVX_FOUND})
+    set(GGML_AVX OFF)
+else()
+    set(GGML_AVX ON)
+endif()
+
+check_sse("AVX2" " ;/arch:AVX2")
+check_sse("FMA" " ;/arch:AVX2")
+if ((NOT ${AVX2_FOUND}) OR (NOT ${FMA_FOUND}))
+    set(GGML_AVX2 OFF)
+else()
+    set(GGML_AVX2 ON)
+endif()
+
+check_sse("AVX512" " ;/arch:AVX512")
+if (NOT ${AVX512_FOUND})
+    set(GGML_AVX512 OFF)
+else()
+    set(GGML_AVX512 ON)
+endif()
diff --git a/src/whisper_cpp/ggml/ggml_vk_generate_shaders.py b/src/whisper_cpp/ggml/ggml_vk_generate_shaders.py
new file mode 100644
index 00000000..38914eed
--- /dev/null
+++ b/src/whisper_cpp/ggml/ggml_vk_generate_shaders.py
@@ -0,0 +1,220 @@
+#!/usr/bin/env python
+
+import logging
+import argparse
+import asyncio
+import os
+from tempfile import gettempdir
+
+logger = logging.getLogger("ggml-vk-generate-shaders")
+
+GLSLC = "glslc"
+
+type_names = [
+    "f32",
+    "f16",
+    "q4_0",
+    "q4_1",
+    "q5_0",
+    "q5_1",
+    "q8_0",
+    "q2_k",
+    "q3_k",
+    "q4_k",
+    "q5_k",
+    "q6_k",
+]
+
+ASYNCIO_CONCURRENCY = 64
+
+input_dir = "vulkan-shaders"
+output_dir = gettempdir()
+
+lock = asyncio.Lock()
+shader_fnames = []
+
+
+async def string_to_spv(name, in_fname, defines, fp16=True):
+    name = f"{name}{'_fp32' if not fp16 else ''}"
+    out_fname = os.path.join(output_dir, f"{name}.spv")
+
+    in_path = os.path.join(input_dir, in_fname)
+
+    cmd = [GLSLC, "-fshader-stage=compute", "--target-env=vulkan1.2", "-O", in_path, "-o", out_fname]
+
+    cmd.extend([f"-D{key}={value}" for key, value in defines.items()])
+
+    proc = await asyncio.create_subprocess_exec(*cmd, stdout=asyncio.subprocess.PIPE, stderr=asyncio.subprocess.PIPE)
+
+    stdout, stderr = await proc.communicate()
+
+    stdout = stdout.decode()
+    error = stderr.decode()
+
+    if proc.returncode:
+        cmd = " ".join(cmd)
+        logger.error(f"cannot compile {name}\n\n{cmd}\n\n{error}")
+        return
+
+    async with lock:
+        shader_fnames.append((name, out_fname))
+
+
+def matmul_shaders(tasks, fp16, matmul_id):
+    if fp16:
+        load_vec = "8"
+        aligned_b_type_f32 = "mat2x4"
+        aligned_b_type_f16 = "f16mat2x4"
+    else:
+        load_vec = "4"
+        aligned_b_type_f32 = "vec4"
+        aligned_b_type_f16 = "f16vec4"
+
+    base_dict = {"FLOAT_TYPE": "float" if not fp16 else "float16_t"}
+    shader_name = "matmul"
+
+    if matmul_id:
+        base_dict["MUL_MAT_ID"] = "1"
+        shader_name = "matmul_id"
+
+    if fp16:
+        base_dict["FLOAT16"] = "1"
+
+    # Shaders with f16 B_TYPE
+    tasks.append(string_to_spv(f"{shader_name}_f32_f16", "mul_mm.comp", base_dict | {"DATA_A_F32": "1", "B_TYPE": "float16_t", "D_TYPE": "float"}, fp16))
+    tasks.append(string_to_spv(f"{shader_name}_f32_f16_aligned", "mul_mm.comp", base_dict | {"DATA_A_F32": "1", "LOAD_VEC_A": load_vec, "LOAD_VEC_B": load_vec, "B_TYPE": aligned_b_type_f16, "D_TYPE": "float"}, fp16))
+
+    tasks.append(string_to_spv(f"{shader_name}_f16", "mul_mm.comp", base_dict | {"DATA_A_F16": "1", "B_TYPE": "float16_t", "D_TYPE": "float"}, fp16))
+    tasks.append(string_to_spv(f"{shader_name}_f16_aligned", "mul_mm.comp", base_dict | {"DATA_A_F16": "1", "LOAD_VEC_A": load_vec, "LOAD_VEC_B": load_vec, "B_TYPE": aligned_b_type_f16, "D_TYPE": "float"}, fp16))
+
+    for tname in type_names:
+        data_a_key = f"DATA_A_{tname.upper()}"
+        load_vec_a = load_vec if tname in ("f32", "f16") else "2"
+        tasks.append(string_to_spv(f"{shader_name}_{tname}_f32", "mul_mm.comp", base_dict | {data_a_key: "1", "B_TYPE": "float", "D_TYPE": "float"}, fp16))
+        tasks.append(string_to_spv(f"{shader_name}_{tname}_f32_aligned", "mul_mm.comp", base_dict | {data_a_key: "2", "LOAD_VEC_A": load_vec_a, "LOAD_VEC_B": load_vec, "B_TYPE": aligned_b_type_f32, "D_TYPE": "float"}, fp16))
+
+
+async def main():
+    logger.info("ggml_vulkan: Generating and compiling shaders to SPIR-V")
+
+    tasks = []
+
+    for fp16 in (False, True):
+        # MUL_MAT
+        matmul_shaders(tasks, fp16, False)
+        # MUL_MAT_ID
+        matmul_shaders(tasks, fp16, True)
+
+    for tname in type_names:
+        base_dict = {"FLOAT_TYPE": "float"}
+
+        # mul mat vec
+        data_a_key = f"DATA_A_{tname.upper()}"
+        shader = f"mul_mat_vec_{tname}.comp" if tname.endswith("_k") else "mul_mat_vec.comp"
+
+        tasks.append(string_to_spv(f"mul_mat_vec_{tname}_f32_f32", shader, base_dict | {data_a_key: "1", "B_TYPE": "float", "D_TYPE": "float"}))
+        tasks.append(string_to_spv(f"mul_mat_vec_{tname}_f16_f32", shader, base_dict | {data_a_key: "1", "B_TYPE": "float16_t", "D_TYPE": "float"}))
+
+        tasks.append(string_to_spv(f"mul_mat_vec_id_{tname}_f32", shader, base_dict | {"MUL_MAT_ID": "1", data_a_key: "1", "B_TYPE": "float", "D_TYPE": "float"}))
+
+        # Dequant shaders
+        if tname != "f16":
+            tasks.append(string_to_spv(f"dequant_{tname}", f"dequant_{tname}.comp", base_dict | {data_a_key: "1", "D_TYPE": "float16_t"}))
+
+        # get_rows
+        if not tname.endswith("_k"):
+            shader = "get_rows.comp" if tname in ("f32", "f16") else "get_rows_quant.comp"
+
+            if tname == "f16":
+                tasks.append(string_to_spv(f"get_rows_{tname}", shader, {data_a_key: "1", "B_TYPE": "int", "D_TYPE": "float16_t", "OPTIMIZATION_ERROR_WORKAROUND": "1"}))
+            else:
+                tasks.append(string_to_spv(f"get_rows_{tname}", shader, {data_a_key: "1", "B_TYPE": "int", "D_TYPE": "float16_t"}))
+            tasks.append(string_to_spv(f"get_rows_{tname}_f32", shader, {data_a_key: "1", "B_TYPE": "int", "D_TYPE": "float"}))
+
+    tasks.append(string_to_spv("mul_mat_vec_p021_f16_f32", "mul_mat_vec_p021.comp", {"A_TYPE": "float16_t", "B_TYPE": "float", "D_TYPE": "float"}))
+    tasks.append(string_to_spv("mul_mat_vec_nc_f16_f32", "mul_mat_vec_nc.comp", {"A_TYPE": "float16_t", "B_TYPE": "float", "D_TYPE": "float"}))
+
+    # Norms
+    tasks.append(string_to_spv("norm_f32", "norm.comp", base_dict | {"A_TYPE": "float", "D_TYPE": "float"}))
+    tasks.append(string_to_spv("rms_norm_f32", "rms_norm.comp", base_dict | {"A_TYPE": "float", "D_TYPE": "float"}))
+
+    tasks.append(string_to_spv("cpy_f32_f32", "copy.comp", {"A_TYPE": "float", "D_TYPE": "float"}))
+    tasks.append(string_to_spv("cpy_f32_f16", "copy.comp", {"A_TYPE": "float", "D_TYPE": "float16_t"}))
+    tasks.append(string_to_spv("cpy_f16_f16", "copy.comp", {"A_TYPE": "float16_t", "D_TYPE": "float16_t", "OPTIMIZATION_ERROR_WORKAROUND": "1"}))
+
+    tasks.append(string_to_spv("add_f32", "add.comp", {"A_TYPE": "float", "B_TYPE": "float", "D_TYPE": "float", "FLOAT_TYPE": "float"}))
+
+    tasks.append(string_to_spv("split_k_reduce", "mul_mat_split_k_reduce.comp", {}))
+
+    tasks.append(string_to_spv("mul_f32", "mul.comp", {"A_TYPE": "float", "B_TYPE": "float", "D_TYPE": "float", "FLOAT_TYPE": "float"}))
+
+    tasks.append(string_to_spv("div_f32", "div.comp", {"A_TYPE": "float", "B_TYPE": "float", "D_TYPE": "float", "FLOAT_TYPE": "float"}))
+
+    tasks.append(string_to_spv("scale_f32", "scale.comp", {"A_TYPE": "float", "D_TYPE": "float", "FLOAT_TYPE": "float"}))
+
+    tasks.append(string_to_spv("sqr_f32", "square.comp", {"A_TYPE": "float", "D_TYPE": "float", "FLOAT_TYPE": "float"}))
+
+    tasks.append(string_to_spv("clamp_f32", "clamp.comp", {"A_TYPE": "float", "D_TYPE": "float", "FLOAT_TYPE": "float"}))
+
+    tasks.append(string_to_spv("gelu_f32", "gelu.comp", {"A_TYPE": "float", "D_TYPE": "float"}))
+    tasks.append(string_to_spv("silu_f32", "silu.comp", {"A_TYPE": "float", "D_TYPE": "float"}))
+    tasks.append(string_to_spv("relu_f32", "relu.comp", {"A_TYPE": "float", "D_TYPE": "float"}))
+
+    tasks.append(string_to_spv("diag_mask_inf_f32", "diag_mask_inf.comp", {"A_TYPE": "float", "D_TYPE": "float"}))
+
+    tasks.append(string_to_spv("soft_max_f32", "soft_max.comp", base_dict | {"A_TYPE": "float", "B_TYPE": "float", "D_TYPE": "float"}))
+    tasks.append(string_to_spv("soft_max_f32_f16", "soft_max.comp", base_dict | {"A_TYPE": "float", "B_TYPE": "float16_t", "D_TYPE": "float"}))
+
+    tasks.append(string_to_spv("rope_norm_f32", "rope_norm.comp", {"A_TYPE": "float", "D_TYPE": "float"}))
+    tasks.append(string_to_spv("rope_norm_f16", "rope_norm.comp", {"A_TYPE": "float16_t", "D_TYPE": "float16_t"}))
+
+    tasks.append(string_to_spv("rope_neox_f32", "rope_neox.comp", {"A_TYPE": "float", "D_TYPE": "float"}))
+    tasks.append(string_to_spv("rope_neox_f16", "rope_neox.comp", {"A_TYPE": "float16_t", "D_TYPE": "float16_t"}))
+
+    tasks.append(string_to_spv("argsort_f32", "argsort.comp", {"A_TYPE": "float"}))
+
+    tasks.append(string_to_spv("sum_rows_f32", "sum_rows.comp", base_dict | {"A_TYPE": "float", "D_TYPE": "float"}))
+
+    # Helper to decorate tasks with semaphore acquisition.
+    async def withSemaphore(sem, task):
+        async with sem:
+            return await task
+
+    # Run tasks concurrently guarded by a concurrency limit.
+    sem = asyncio.Semaphore(ASYNCIO_CONCURRENCY)
+    await asyncio.gather(*(withSemaphore(sem, task) for task in tasks))
+
+    with open("ggml-vulkan-shaders.hpp", "w") as f:
+        f.write("#include <cstdint>\n\n")
+        for name, path in sorted(shader_fnames):
+
+            with open(path, "rb") as spv:
+                counter = 0
+                newline_counter = 0
+                f.write(f"unsigned char {name}_data[] = {{\n")
+                for val in spv.read():
+                    f.write(f"0x{val:02x},")
+                    newline_counter += 1
+                    counter += 1
+                    if newline_counter >= 12:
+                        newline_counter = 0
+                        f.write("\n")
+            f.write("\n};\n")
+            f.write(f"const uint64_t {name}_len = {counter};\n\n")
+            os.remove(path)
+
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser(description="GGML Vulkan Shader Generator")
+
+    parser.add_argument("--glslc", help="Path to glslc")
+    parser.add_argument("--verbose", action="store_true", help="increase output verbosity")
+
+    args = parser.parse_args()
+
+    logging.basicConfig(level=logging.DEBUG if args.verbose else logging.INFO)
+
+    if args.glslc:
+        GLSLC = args.glslc
+
+    asyncio.run(main())
diff --git a/src/whisper_cpp/ggml/include/ggml-alloc.h b/src/whisper_cpp/ggml/include/ggml-alloc.h
new file mode 100644
index 00000000..23600eea
--- /dev/null
+++ b/src/whisper_cpp/ggml/include/ggml-alloc.h
@@ -0,0 +1,76 @@
+#pragma once
+
+#include "ggml.h"
+
+#ifdef  __cplusplus
+extern "C" {
+#endif
+
+typedef struct ggml_backend_buffer_type * ggml_backend_buffer_type_t;
+typedef struct      ggml_backend_buffer * ggml_backend_buffer_t;
+typedef struct             ggml_backend * ggml_backend_t;
+
+// Tensor allocator
+struct ggml_tallocr {
+    ggml_backend_buffer_t buffer;
+    void * base;
+    size_t alignment;
+    size_t offset;
+};
+
+GGML_API struct ggml_tallocr ggml_tallocr_new(ggml_backend_buffer_t buffer);
+GGML_API void                ggml_tallocr_alloc(struct ggml_tallocr * talloc, struct ggml_tensor * tensor);
+
+// Graph allocator
+/*
+  Example usage:
+    ggml_gallocr_t galloc = ggml_gallocr_new(ggml_backend_cpu_buffer_type());
+
+    // optional: create a worst-case graph and reserve the buffers to avoid reallocations
+    ggml_gallocr_reserve(galloc, build_graph(max_batch));
+
+    // allocate the graph
+    struct ggml_cgraph * graph = build_graph(batch);
+    ggml_gallocr_alloc_graph(galloc, graph);
+
+    printf("compute buffer size: %zu bytes\n", ggml_gallocr_get_buffer_size(galloc, 0));
+
+    // evaluate the graph
+    ggml_backend_graph_compute(backend, graph);
+*/
+
+// special tensor flags for use with the graph allocator:
+//   ggml_set_input(): all input tensors are allocated at the beginning of the graph in non-overlapping addresses
+//   ggml_set_output(): output tensors are never freed and never overwritten
+
+typedef struct ggml_gallocr * ggml_gallocr_t;
+
+GGML_API ggml_gallocr_t ggml_gallocr_new(ggml_backend_buffer_type_t buft);
+GGML_API ggml_gallocr_t ggml_gallocr_new_n(ggml_backend_buffer_type_t * bufts, int n_bufs);
+GGML_API void           ggml_gallocr_free(ggml_gallocr_t galloc);
+
+// pre-allocate buffers from a measure graph - does not allocate or modify the graph
+// call with a worst-case graph to avoid buffer reallocations
+// not strictly required for single buffer usage: ggml_gallocr_alloc_graph will reallocate the buffers automatically if needed
+// returns false if the buffer allocation failed
+GGML_API bool ggml_gallocr_reserve(ggml_gallocr_t galloc, struct ggml_cgraph * graph);
+GGML_API bool ggml_gallocr_reserve_n(
+    ggml_gallocr_t galloc,
+    struct ggml_cgraph * graph,
+    const int * node_buffer_ids,
+    const int * leaf_buffer_ids);
+
+// automatic reallocation if the topology changes when using a single buffer
+// returns false if using multiple buffers and a re-allocation is needed (call ggml_gallocr_reserve_n first to set the node buffers)
+GGML_API bool ggml_gallocr_alloc_graph(ggml_gallocr_t galloc, struct ggml_cgraph * graph);
+
+GGML_API size_t ggml_gallocr_get_buffer_size(ggml_gallocr_t galloc, int buffer_id);
+
+// Utils
+// Create a buffer and allocate all the tensors in a ggml_context
+GGML_API struct ggml_backend_buffer * ggml_backend_alloc_ctx_tensors_from_buft(struct ggml_context * ctx, ggml_backend_buffer_type_t buft);
+GGML_API struct ggml_backend_buffer * ggml_backend_alloc_ctx_tensors(struct ggml_context * ctx, ggml_backend_t backend);
+
+#ifdef  __cplusplus
+}
+#endif
diff --git a/src/whisper_cpp/ggml/include/ggml-backend.h b/src/whisper_cpp/ggml/include/ggml-backend.h
new file mode 100644
index 00000000..4d7d2716
--- /dev/null
+++ b/src/whisper_cpp/ggml/include/ggml-backend.h
@@ -0,0 +1,323 @@
+#pragma once
+
+#include "ggml.h"
+#include "ggml-alloc.h"
+
+#ifdef  __cplusplus
+extern "C" {
+#endif
+
+    typedef struct ggml_backend_buffer_type * ggml_backend_buffer_type_t;
+    typedef struct ggml_backend_buffer * ggml_backend_buffer_t;
+    typedef struct ggml_backend_event * ggml_backend_event_t;
+    typedef struct ggml_backend * ggml_backend_t;
+    typedef void * ggml_backend_graph_plan_t;
+    typedef struct ggml_backend_reg * ggml_backend_reg_t;
+    typedef struct ggml_backend_device * ggml_backend_dev_t;
+
+
+    //
+    // Backend buffer type
+    //
+
+    GGML_API const char *          ggml_backend_buft_name          (ggml_backend_buffer_type_t buft);
+    GGML_API ggml_backend_buffer_t ggml_backend_buft_alloc_buffer  (ggml_backend_buffer_type_t buft, size_t size);
+    GGML_API size_t                ggml_backend_buft_get_alignment (ggml_backend_buffer_type_t buft);
+    GGML_API size_t                ggml_backend_buft_get_max_size  (ggml_backend_buffer_type_t buft);
+    GGML_API size_t                ggml_backend_buft_get_alloc_size(ggml_backend_buffer_type_t buft, struct ggml_tensor * tensor);
+    GGML_API bool                  ggml_backend_buft_is_host       (ggml_backend_buffer_type_t buft);
+    GGML_API ggml_backend_dev_t    ggml_backend_buft_get_device    (ggml_backend_buffer_type_t buft);
+
+    //
+    // Backend buffer
+    //
+
+    enum ggml_backend_buffer_usage {
+        GGML_BACKEND_BUFFER_USAGE_ANY = 0,
+        GGML_BACKEND_BUFFER_USAGE_WEIGHTS = 1,
+        GGML_BACKEND_BUFFER_USAGE_COMPUTE = 2,
+    };
+
+    GGML_API const char *                   ggml_backend_buffer_name          (ggml_backend_buffer_t buffer);
+    GGML_API void                           ggml_backend_buffer_free          (ggml_backend_buffer_t buffer);
+    GGML_API void *                         ggml_backend_buffer_get_base      (ggml_backend_buffer_t buffer);
+    GGML_API size_t                         ggml_backend_buffer_get_size      (ggml_backend_buffer_t buffer);
+    GGML_API void                           ggml_backend_buffer_init_tensor   (ggml_backend_buffer_t buffer, struct ggml_tensor * tensor);
+    GGML_API size_t                         ggml_backend_buffer_get_alignment (ggml_backend_buffer_t buffer);
+    GGML_API size_t                         ggml_backend_buffer_get_max_size  (ggml_backend_buffer_t buffer);
+    GGML_API size_t                         ggml_backend_buffer_get_alloc_size(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor);
+    GGML_API void                           ggml_backend_buffer_clear         (ggml_backend_buffer_t buffer, uint8_t value);
+    GGML_API bool                           ggml_backend_buffer_is_host       (ggml_backend_buffer_t buffer);
+    GGML_API void                           ggml_backend_buffer_set_usage     (ggml_backend_buffer_t buffer, enum ggml_backend_buffer_usage usage);
+    GGML_API enum ggml_backend_buffer_usage ggml_backend_buffer_get_usage     (ggml_backend_buffer_t buffer);
+    GGML_API ggml_backend_buffer_type_t     ggml_backend_buffer_get_type      (ggml_backend_buffer_t buffer);
+    GGML_API void                           ggml_backend_buffer_reset         (ggml_backend_buffer_t buffer);
+
+    // tensor copy between different backends
+    GGML_API void ggml_backend_tensor_copy(struct ggml_tensor * src, struct ggml_tensor * dst);
+
+    //
+    // Backend (stream)
+    //
+
+    GGML_API ggml_guid_t  ggml_backend_guid(ggml_backend_t backend);
+    GGML_API const char * ggml_backend_name(ggml_backend_t backend);
+    GGML_API void         ggml_backend_free(ggml_backend_t backend);
+
+    GGML_API ggml_backend_buffer_type_t ggml_backend_get_default_buffer_type(ggml_backend_t backend);
+    GGML_API ggml_backend_buffer_t      ggml_backend_alloc_buffer(ggml_backend_t backend, size_t size);
+    GGML_API size_t                     ggml_backend_get_alignment(ggml_backend_t backend);
+    GGML_API size_t                     ggml_backend_get_max_size(ggml_backend_t backend);
+
+    GGML_API void ggml_backend_tensor_set_async(ggml_backend_t backend,       struct ggml_tensor * tensor, const void * data, size_t offset, size_t size);
+    GGML_API void ggml_backend_tensor_get_async(ggml_backend_t backend, const struct ggml_tensor * tensor,       void * data, size_t offset, size_t size);
+
+    // "offset" refers to the offset of the tensor data for setting/getting data
+    GGML_API void ggml_backend_tensor_set(      struct ggml_tensor * tensor, const void * data, size_t offset, size_t size);
+    GGML_API void ggml_backend_tensor_get(const struct ggml_tensor * tensor,       void * data, size_t offset, size_t size);
+    GGML_API void ggml_backend_tensor_memset(   struct ggml_tensor * tensor,     uint8_t value, size_t offset, size_t size);
+
+    GGML_API void ggml_backend_synchronize(ggml_backend_t backend);
+
+    GGML_API ggml_backend_graph_plan_t ggml_backend_graph_plan_create(ggml_backend_t backend, struct ggml_cgraph * cgraph);
+    GGML_API void                      ggml_backend_graph_plan_free  (ggml_backend_t backend, ggml_backend_graph_plan_t plan);
+
+    GGML_API enum ggml_status ggml_backend_graph_plan_compute (ggml_backend_t backend, ggml_backend_graph_plan_t plan);
+    GGML_API enum ggml_status ggml_backend_graph_compute      (ggml_backend_t backend, struct ggml_cgraph * cgraph);
+    GGML_API enum ggml_status ggml_backend_graph_compute_async(ggml_backend_t backend, struct ggml_cgraph * cgraph);
+
+    // NOTE: will be removed, use device version instead
+    GGML_API bool ggml_backend_supports_op(ggml_backend_t backend, const struct ggml_tensor * op);
+    GGML_API bool ggml_backend_supports_buft(ggml_backend_t backend, ggml_backend_buffer_type_t buft);
+    GGML_API bool ggml_backend_offload_op(ggml_backend_t backend, const struct ggml_tensor * op);
+
+    // asynchronous copy
+    // the copy is performed after all the currently queued operations in backend_src
+    // backend_dst will wait for the copy to complete before performing other operations
+    // automatic fallback to sync copy if async is not supported
+    GGML_API void ggml_backend_tensor_copy_async(ggml_backend_t backend_src, ggml_backend_t backend_dst, struct ggml_tensor * src, struct ggml_tensor * dst);
+
+    GGML_API ggml_backend_dev_t ggml_backend_get_device(ggml_backend_t backend);
+
+    //
+    // Events
+    //
+
+    GGML_API ggml_backend_event_t ggml_backend_event_new(ggml_backend_dev_t device);
+    GGML_API void                 ggml_backend_event_free(ggml_backend_event_t event);
+    GGML_API void                 ggml_backend_event_record(ggml_backend_event_t event, ggml_backend_t backend);
+    GGML_API void                 ggml_backend_event_synchronize(ggml_backend_event_t event);
+    GGML_API void                 ggml_backend_event_wait(ggml_backend_t backend, ggml_backend_event_t event);
+
+    //
+    // Backend device
+    //
+
+    enum ggml_backend_dev_type {
+        GGML_BACKEND_DEVICE_TYPE_CPU,
+        GGML_BACKEND_DEVICE_TYPE_GPU,
+        // devices with full capabilities (excludes backends such as BLAS that only support matrix multiplication)
+        GGML_BACKEND_DEVICE_TYPE_CPU_FULL,
+        GGML_BACKEND_DEVICE_TYPE_GPU_FULL
+    };
+
+    // functionality supported by the device
+    struct ggml_backend_dev_caps {
+        // asynchronous operations
+        bool async;
+        // pinned host buffer
+        bool host_buffer;
+        // event synchronization
+        bool events;
+    };
+
+    // all the device properties
+    struct ggml_backend_dev_props {
+        const char * name;
+        const char * description;
+        size_t memory_free;
+        size_t memory_total;
+        enum ggml_backend_dev_type type;
+        struct ggml_backend_dev_caps caps;
+    };
+
+    GGML_API const char *                  ggml_backend_dev_name(ggml_backend_dev_t device);
+    GGML_API const char *                  ggml_backend_dev_description(ggml_backend_dev_t device);
+    GGML_API void                          ggml_backend_dev_memory(ggml_backend_dev_t device, size_t * free, size_t * total);
+    GGML_API enum ggml_backend_dev_type    ggml_backend_dev_type(ggml_backend_dev_t device);
+    GGML_API void                          ggml_backend_dev_get_props(ggml_backend_dev_t device, struct ggml_backend_dev_props * props);
+    GGML_API ggml_backend_reg_t            ggml_backend_dev_backend_reg(ggml_backend_dev_t device);
+    GGML_API ggml_backend_t                ggml_backend_dev_init(ggml_backend_dev_t device, const char * params);
+    GGML_API ggml_backend_buffer_type_t    ggml_backend_dev_buffer_type(ggml_backend_dev_t device);
+    GGML_API ggml_backend_buffer_type_t    ggml_backend_dev_host_buffer_type(ggml_backend_dev_t device);
+    GGML_API ggml_backend_buffer_t         ggml_backend_dev_buffer_from_host_ptr(ggml_backend_dev_t device, void * ptr, size_t size, size_t max_tensor_size);
+
+    GGML_API bool                          ggml_backend_dev_supports_op(ggml_backend_dev_t device, const struct ggml_tensor * op);
+    GGML_API bool                          ggml_backend_dev_supports_buft(ggml_backend_dev_t device, ggml_backend_buffer_type_t buft);
+    GGML_API bool                          ggml_backend_dev_offload_op(ggml_backend_dev_t device, const struct ggml_tensor * op);
+
+    //
+    // Backend (reg)
+    //
+
+    GGML_API const char *       ggml_backend_reg_name(ggml_backend_reg_t reg);
+    GGML_API size_t             ggml_backend_reg_dev_count(ggml_backend_reg_t reg);
+    GGML_API ggml_backend_dev_t ggml_backend_reg_dev_get(ggml_backend_reg_t reg, size_t index);
+    GGML_API void *             ggml_backend_reg_get_proc_address(ggml_backend_reg_t reg, const char * name);
+
+
+    // Functions that may be obtained using ggml_backend_reg_get_proc_address
+    typedef ggml_backend_buffer_type_t (*ggml_backend_split_buffer_type_t)(const float *);
+
+    //
+    // Backend registry
+    //
+
+    // Backend (reg) enumeration
+    GGML_API size_t             ggml_backend_reg_count(void);
+    GGML_API ggml_backend_reg_t ggml_backend_reg_get(size_t index);
+    GGML_API ggml_backend_reg_t ggml_backend_reg_by_name(const char * name);
+
+    // Device enumeration
+    GGML_API size_t             ggml_backend_dev_count(void);
+    GGML_API ggml_backend_dev_t ggml_backend_dev_get(size_t index);
+    GGML_API ggml_backend_dev_t ggml_backend_dev_by_name(const char * name);
+    GGML_API ggml_backend_dev_t ggml_backend_dev_by_type(enum ggml_backend_dev_type type);
+
+    // Direct backend (stream) initialization
+    // = ggml_backend_dev_init(ggml_backend_dev_by_name(name), params)
+    GGML_API ggml_backend_t ggml_backend_init_by_name(const char * name, const char * params);
+    // = ggml_backend_dev_init(ggml_backend_dev_by_type(type), params)
+    GGML_API ggml_backend_t ggml_backend_init_by_type(enum ggml_backend_dev_type type, const char * params);
+    // = ggml_backend_dev_init(ggml_backend_dev_by_type(GPU_FULL) OR ggml_backend_dev_by_type(CPU_FULL), NULL)
+    GGML_API ggml_backend_t ggml_backend_init_best(void);
+
+    //
+    // Backend scheduler
+    //
+
+    // The backend scheduler allows for multiple backend devices to be used together
+    // Handles compute buffer allocation, assignment of tensors to backends, and copying of tensors between backends
+    // The backends are selected based on:
+    // - the backend that supports the operation
+    // - the location of the pre-allocated tensors (e.g. the weights)
+    /*
+      Example usage:
+
+        // operations that use tensors allocated in a buffer with USAGE_WEIGHTS will be assigned
+        // preferrably to run on the same backend as the buffer
+        ggml_backend_buffer_set_usage(buf_weights, GGML_BACKEND_BUFFER_USAGE_WEIGHTS);
+
+        sched = ggml_backend_sched_new({backend_gpu, backend_gpu2, backend_cpu}, NULL, num_backends, GGML_DEFAULT_GRAPH_SIZE, false);
+
+        // initialize buffers from a max size graph (optional)
+        reserve_graph = build_graph(sched, max_batch_size);
+
+        // manually assign nodes to a backend (optional, should not be needed in most cases)
+        struct ggml_tensor * node = ggml_mul_mat(ctx, ...);
+        ggml_backend_sched_set_tensor_backend(sched, node, backend_gpu);
+
+        ggml_backend_sched_reserve(sched, reserve_graph);
+
+        // compute
+        graph = build_graph(sched);
+        ggml_backend_sched_graph_compute(sched, graph);
+
+        // if there are graph inputs:
+        ggml_backend_sched_reset(sched);
+        ggml_backend_sched_alloc_graph(sched, graph);
+        ggml_backend_tensor_set(input_tensor, ...);
+        ggml_backend_sched_graph_compute(sched, graph);
+    }
+    */
+
+    typedef struct ggml_backend_sched * ggml_backend_sched_t;
+
+    // Evaluation callback for each node in the graph (set with ggml_backend_sched_set_eval_callback)
+    // when ask == true, the scheduler wants to know if the user wants to observe this node
+    // this allows the scheduler to batch nodes together in order to evaluate them in a single call
+    //
+    // when ask == false, the scheduler is passing the node tensor to the user for observation
+    // if the user returns false, the scheduler will cancel the graph compute
+    //
+    typedef bool (*ggml_backend_sched_eval_callback)(struct ggml_tensor * t, bool ask, void * user_data);
+
+    // Initialize a backend scheduler
+    GGML_API ggml_backend_sched_t ggml_backend_sched_new(ggml_backend_t * backends, ggml_backend_buffer_type_t * bufts, int n_backends, size_t graph_size, bool parallel);
+    GGML_API void                 ggml_backend_sched_free(ggml_backend_sched_t sched);
+
+    // Initialize backend buffers from a measure graph
+    GGML_API bool                 ggml_backend_sched_reserve(ggml_backend_sched_t sched, struct ggml_cgraph * measure_graph); // returns success
+
+    GGML_API int                  ggml_backend_sched_get_n_backends(ggml_backend_sched_t sched);
+    GGML_API ggml_backend_t       ggml_backend_sched_get_backend(ggml_backend_sched_t sched, int i);
+
+    // Get the number of splits of the last graph
+    GGML_API int                  ggml_backend_sched_get_n_splits(ggml_backend_sched_t sched);
+    GGML_API int                  ggml_backend_sched_get_n_copies(ggml_backend_sched_t sched);
+
+    GGML_API size_t               ggml_backend_sched_get_buffer_size(ggml_backend_sched_t sched, ggml_backend_t backend);
+
+    GGML_API void                 ggml_backend_sched_set_tensor_backend(ggml_backend_sched_t sched, struct ggml_tensor * node, ggml_backend_t backend);
+    GGML_API ggml_backend_t       ggml_backend_sched_get_tensor_backend(ggml_backend_sched_t sched, struct ggml_tensor * node);
+
+    // Allocate and compute graph on the backend scheduler
+    GGML_API bool                 ggml_backend_sched_alloc_graph(ggml_backend_sched_t sched, struct ggml_cgraph * graph); // returns success
+    GGML_API enum ggml_status     ggml_backend_sched_graph_compute(ggml_backend_sched_t sched, struct ggml_cgraph * graph);
+    GGML_API enum ggml_status     ggml_backend_sched_graph_compute_async(ggml_backend_sched_t sched, struct ggml_cgraph * graph);
+    GGML_API void                 ggml_backend_sched_synchronize(ggml_backend_sched_t sched);
+
+    // Reset all assignments and allocators - must be called before changing the node backends
+    GGML_API void                 ggml_backend_sched_reset(ggml_backend_sched_t sched);
+
+    // Set a callback to be called for each resulting node during graph compute
+    GGML_API void                 ggml_backend_sched_set_eval_callback(ggml_backend_sched_t sched, ggml_backend_sched_eval_callback callback, void * user_data);
+
+    //
+    // Utils
+    //
+
+    struct ggml_backend_graph_copy {
+        ggml_backend_buffer_t buffer;
+        struct ggml_context * ctx_allocated;
+        struct ggml_context * ctx_unallocated;
+        struct ggml_cgraph * graph;
+    };
+
+    // Copy a graph to a different backend
+    GGML_API struct ggml_backend_graph_copy ggml_backend_graph_copy(ggml_backend_t backend, struct ggml_cgraph * graph);
+    GGML_API void                           ggml_backend_graph_copy_free(struct ggml_backend_graph_copy copy);
+
+    typedef bool (*ggml_backend_eval_callback)(int node_index, struct ggml_tensor * t1, struct ggml_tensor * t2, void * user_data);
+
+    // Compare the output of two backends
+    GGML_API bool ggml_backend_compare_graph_backend(ggml_backend_t backend1, ggml_backend_t backend2, struct ggml_cgraph * graph, ggml_backend_eval_callback callback, void * user_data);
+
+    // Tensor initialization
+    GGML_API void ggml_backend_tensor_alloc(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor, void * addr);
+    GGML_API void ggml_backend_view_init(struct ggml_tensor * tensor);
+
+    //
+    // CPU backend
+    //
+
+    GGML_API ggml_backend_t ggml_backend_cpu_init(void);
+
+    GGML_API bool ggml_backend_is_cpu                (ggml_backend_t backend);
+    GGML_API void ggml_backend_cpu_set_n_threads     (ggml_backend_t backend_cpu, int n_threads);
+    GGML_API void ggml_backend_cpu_set_threadpool    (ggml_backend_t backend_cpu, ggml_threadpool_t threadpool);
+    GGML_API void ggml_backend_cpu_set_abort_callback(ggml_backend_t backend_cpu, ggml_abort_callback abort_callback, void * abort_callback_data);
+
+    // Create a backend buffer from an existing pointer
+    GGML_API ggml_backend_buffer_t      ggml_backend_cpu_buffer_from_ptr(void * ptr, size_t size);
+    GGML_API ggml_backend_buffer_type_t ggml_backend_cpu_buffer_type(void);
+
+    GGML_API ggml_backend_reg_t ggml_backend_cpu_reg(void);
+
+#ifdef GGML_USE_CPU_HBM
+    GGML_API ggml_backend_buffer_type_t ggml_backend_cpu_hbm_buffer_type(void);
+#endif
+
+#ifdef  __cplusplus
+}
+#endif
diff --git a/src/whisper_cpp/ggml/include/ggml-blas.h b/src/whisper_cpp/ggml/include/ggml-blas.h
new file mode 100644
index 00000000..dd612860
--- /dev/null
+++ b/src/whisper_cpp/ggml/include/ggml-blas.h
@@ -0,0 +1,23 @@
+#pragma once
+
+#include "ggml.h"
+#include "ggml-backend.h"
+
+
+#ifdef  __cplusplus
+extern "C" {
+#endif
+
+// backend API
+GGML_API ggml_backend_t ggml_backend_blas_init(void);
+
+GGML_API bool ggml_backend_is_blas(ggml_backend_t backend);
+
+// number of threads used for conversion to float
+// for openblas and blis, this will also set the number of threads used for blas operations
+GGML_API void ggml_backend_blas_set_n_threads(ggml_backend_t backend_blas, int n_threads);
+
+
+#ifdef  __cplusplus
+}
+#endif
diff --git a/src/whisper_cpp/ggml/include/ggml-cann.h b/src/whisper_cpp/ggml/include/ggml-cann.h
new file mode 100644
index 00000000..95bdaf10
--- /dev/null
+++ b/src/whisper_cpp/ggml/include/ggml-cann.h
@@ -0,0 +1,121 @@
+/*
+ * Copyright (c) 2023-2024 The ggml authors
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a copy
+ * of this software and associated documentation files (the "Software"), to
+ * deal in the Software without restriction, including without limitation the
+ * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or
+ * sell copies of the Software, and to permit persons to whom the Software is
+ * furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice shall be included in
+ * all copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+ * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
+ * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
+ * IN THE SOFTWARE.
+ */
+
+#pragma once
+
+#include "ggml-backend.h"
+#include "ggml.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+/**
+ * @brief Maximum number of CANN devices supported.
+ */
+#define GGML_CANN_MAX_DEVICES 16
+
+/**
+ * @brief Initializes the CANN backend for a specified device.
+ *
+ * This function initializes the CANN backend for the given device.
+ * It verifies the device index, allocates a context, and creates a backend
+ * instance.
+ *
+ * @param device The index of the device to initialize.
+ * @return A pointer to the initialized backend instance, or nullptr on failure.
+ */
+GGML_API ggml_backend_t ggml_backend_cann_init(int32_t device);
+
+/**
+ * @brief Checks if a given backend is a CANN backend.
+ *
+ * This function verifies if the provided backend is a CANN backend by comparing
+ * its GUID with the CANN backend's GUID.
+ *
+ * @param backend The backend instance to check.
+ * @return True if the backend is a CANN backend, false otherwise.
+ */
+GGML_API bool ggml_backend_is_cann(ggml_backend_t backend);
+
+/**
+ * @brief Retrieves the CANN buffer type for a specified device.
+ *
+ * This function initializes and returns the buffer type interface associated
+ * with the given device. It ensures thread-safe access using a mutex.
+ *
+ * @param device The device index for which to retrieve the buffer type.
+ * @return A pointer to the buffer type interface for the specified device, or
+ * nullptr if the device index is out of range.
+ */
+GGML_API ggml_backend_buffer_type_t
+ggml_backend_cann_buffer_type(int32_t device);
+
+/**
+ * @brief Retrieves the number of CANN devices available.
+ *
+ * This function returns the number of CANN devices available based on
+ * information obtained from `ggml_cann_info()`.
+ *
+ * @return The number of CANN devices available.
+ */
+GGML_API int32_t ggml_backend_cann_get_device_count(void);
+
+/**
+ * @brief pinned host buffer for use with the CPU backend for faster copies between CPU and NPU.
+ *
+ * @return A pointer to the host buffer type interface.
+ */
+GGML_API ggml_backend_buffer_type_t ggml_backend_cann_host_buffer_type(void);
+
+/**
+ * @brief Retrieves the description of a specific CANN device.
+ *
+ * This function sets the specified device, retrieves the SoC name,
+ * and writes it into the provided description buffer.
+ *
+ * @param device The device index to retrieve the description for.
+ * @param description Pointer to a buffer where the description will be written.
+ * @param description_size Size of the description buffer.
+ */
+GGML_API void ggml_backend_cann_get_device_description(
+    int32_t device, char* description, size_t description_size);
+
+/**
+ * @brief Retrieves the memory information of a specific CANN device.
+ *
+ * This function sets the specified device, retrieves the free and total
+ * memory information of the specified type (ACL_HBM_MEM), and stores them
+ * in the provided pointers.
+ *
+ * @param device The device index to retrieve memory information for.
+ * @param free Pointer to a variable where the free memory size will be stored.
+ * @param total Pointer to a variable where the total memory size will be
+ * stored.
+ */
+GGML_API void ggml_backend_cann_get_device_memory(int32_t device,
+                                                  size_t* free,
+                                                  size_t* total);
+
+#ifdef __cplusplus
+}
+#endif
diff --git a/src/whisper_cpp/ggml/include/ggml-cuda.h b/src/whisper_cpp/ggml/include/ggml-cuda.h
new file mode 100644
index 00000000..f44d8f4e
--- /dev/null
+++ b/src/whisper_cpp/ggml/include/ggml-cuda.h
@@ -0,0 +1,47 @@
+#pragma once
+
+#include "ggml.h"
+#include "ggml-backend.h"
+
+#ifdef  __cplusplus
+extern "C" {
+#endif
+
+#ifdef GGML_USE_HIPBLAS
+#define GGML_CUDA_NAME "ROCm"
+#define GGML_CUBLAS_NAME "hipBLAS"
+#elif defined(GGML_USE_MUSA)
+#define GGML_CUDA_NAME "MUSA"
+#define GGML_CUBLAS_NAME "muBLAS"
+#else
+#define GGML_CUDA_NAME "CUDA"
+#define GGML_CUBLAS_NAME "cuBLAS"
+#endif
+#define GGML_CUDA_MAX_DEVICES       16
+
+// backend API
+GGML_API ggml_backend_t ggml_backend_cuda_init(int device);
+
+GGML_API bool ggml_backend_is_cuda(ggml_backend_t backend);
+
+// device buffer
+GGML_API ggml_backend_buffer_type_t ggml_backend_cuda_buffer_type(int device);
+
+// split tensor buffer that splits matrices by rows across multiple devices
+GGML_API ggml_backend_buffer_type_t ggml_backend_cuda_split_buffer_type(const float * tensor_split);
+
+// pinned host buffer for use with the CPU backend for faster copies between CPU and GPU
+GGML_API ggml_backend_buffer_type_t ggml_backend_cuda_host_buffer_type(void);
+
+GGML_API int  ggml_backend_cuda_get_device_count(void);
+GGML_API void ggml_backend_cuda_get_device_description(int device, char * description, size_t description_size);
+GGML_API void ggml_backend_cuda_get_device_memory(int device, size_t * free, size_t * total);
+
+GGML_API bool ggml_backend_cuda_register_host_buffer(void * buffer, size_t size);
+GGML_API void ggml_backend_cuda_unregister_host_buffer(void * buffer);
+
+GGML_API ggml_backend_reg_t ggml_backend_cuda_reg(void);
+
+#ifdef  __cplusplus
+}
+#endif
diff --git a/src/whisper_cpp/ggml/include/ggml-kompute.h b/src/whisper_cpp/ggml/include/ggml-kompute.h
new file mode 100644
index 00000000..17146545
--- /dev/null
+++ b/src/whisper_cpp/ggml/include/ggml-kompute.h
@@ -0,0 +1,46 @@
+#pragma once
+
+#include "ggml.h"
+#include "ggml-backend.h"
+
+#include <stdbool.h>
+#include <stddef.h>
+#include <stdint.h>
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+struct ggml_vk_device {
+    int index;
+    int type; // same as VkPhysicalDeviceType
+    size_t heapSize;
+    const char * name;
+    const char * vendor;
+    int subgroupSize;
+    uint64_t bufferAlignment;
+    uint64_t maxAlloc;
+};
+
+struct ggml_vk_device * ggml_vk_available_devices(size_t memoryRequired, size_t * count);
+bool ggml_vk_get_device(struct ggml_vk_device * device, size_t memoryRequired, const char * name);
+bool ggml_vk_has_vulkan(void);
+bool ggml_vk_has_device(void);
+struct ggml_vk_device ggml_vk_current_device(void);
+
+//
+// backend API
+//
+
+// forward declaration
+typedef struct ggml_backend * ggml_backend_t;
+
+GGML_API ggml_backend_t ggml_backend_kompute_init(int device);
+
+GGML_API bool ggml_backend_is_kompute(ggml_backend_t backend);
+
+GGML_API ggml_backend_buffer_type_t ggml_backend_kompute_buffer_type(int device);
+
+#ifdef __cplusplus
+}
+#endif
diff --git a/src/whisper_cpp/ggml/include/ggml-metal.h b/src/whisper_cpp/ggml/include/ggml-metal.h
new file mode 100644
index 00000000..c3ec572b
--- /dev/null
+++ b/src/whisper_cpp/ggml/include/ggml-metal.h
@@ -0,0 +1,62 @@
+// Note: this description is outdated
+//
+// An interface allowing to compute ggml_cgraph with Metal
+//
+// This is a fully functional interface that extends ggml with GPU support for Apple devices.
+// A similar interface can be created for other GPU backends (e.g. Vulkan, CUDA, etc.)
+//
+// How it works?
+//
+// As long as your program can create and evaluate a ggml_cgraph on the CPU, you can use this
+// interface to evaluate the same graph on the GPU. Instead of using ggml_graph_compute(), you
+// use ggml_metal_graph_compute() (or ggml_vulkan_graph_compute(), etc.)
+//
+// You only need to make sure that all memory buffers that you used during the graph creation
+// are mapped to the device memory with the ggml_metal_add_buffer() function. This mapping is
+// used during the graph evaluation to determine the arguments of the compute kernels.
+//
+// Synchronization between device and host memory (for example for input and output tensors)
+// is done with the ggml_metal_set_tensor() and ggml_metal_get_tensor() functions.
+//
+
+#pragma once
+
+#include "ggml.h"
+#include "ggml-backend.h"
+
+#include <stddef.h>
+#include <stdbool.h>
+
+struct ggml_tensor;
+struct ggml_cgraph;
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+//
+// backend API
+// user-code should use only these functions
+//
+
+GGML_API ggml_backend_t ggml_backend_metal_init(void);
+
+GGML_API bool ggml_backend_is_metal(ggml_backend_t backend);
+
+GGML_API ggml_backend_buffer_t ggml_backend_metal_buffer_from_ptr(void * data, size_t size, size_t max_size);
+
+GGML_API void ggml_backend_metal_set_abort_callback(ggml_backend_t backend, ggml_abort_callback abort_callback, void * user_data);
+
+GGML_API ggml_backend_buffer_type_t ggml_backend_metal_buffer_type(void);
+
+// helper to check if the device supports a specific family
+// ideally, the user code should be doing these checks
+// ref: https://developer.apple.com/metal/Metal-Feature-Set-Tables.pdf
+GGML_API bool ggml_backend_metal_supports_family(ggml_backend_t backend, int family);
+
+// capture all command buffers committed the next time `ggml_backend_graph_compute` is called
+GGML_API void ggml_backend_metal_capture_next_compute(ggml_backend_t backend);
+
+#ifdef __cplusplus
+}
+#endif
diff --git a/src/whisper_cpp/ggml/include/ggml-rpc.h b/src/whisper_cpp/ggml/include/ggml-rpc.h
new file mode 100644
index 00000000..64cde7f1
--- /dev/null
+++ b/src/whisper_cpp/ggml/include/ggml-rpc.h
@@ -0,0 +1,24 @@
+#pragma once
+
+#include "ggml.h"
+#include "ggml-backend.h"
+
+#ifdef  __cplusplus
+extern "C" {
+#endif
+
+#define GGML_RPC_MAX_SERVERS       16
+
+// backend API
+GGML_API ggml_backend_t ggml_backend_rpc_init(const char * endpoint);
+GGML_API bool ggml_backend_is_rpc(ggml_backend_t backend);
+
+GGML_API ggml_backend_buffer_type_t ggml_backend_rpc_buffer_type(const char * endpoint);
+
+GGML_API void ggml_backend_rpc_get_device_memory(const char * endpoint, size_t * free, size_t * total);
+
+GGML_API void start_rpc_server(ggml_backend_t backend, const char * endpoint, size_t free_mem, size_t total_mem);
+
+#ifdef  __cplusplus
+}
+#endif
diff --git a/src/whisper_cpp/ggml/include/ggml-sycl.h b/src/whisper_cpp/ggml/include/ggml-sycl.h
new file mode 100644
index 00000000..03b698e6
--- /dev/null
+++ b/src/whisper_cpp/ggml/include/ggml-sycl.h
@@ -0,0 +1,42 @@
+//
+//  MIT license
+//  Copyright (C) 2024 Intel Corporation
+//  SPDX-License-Identifier: MIT
+//
+
+#pragma once
+
+#include "ggml.h"
+#include "ggml-backend.h"
+
+#define GGML_SYCL_NAME "SYCL"
+#define GGML_SYCL_MAX_DEVICES 48
+
+#ifdef  __cplusplus
+extern "C" {
+#endif
+
+// backend API
+GGML_API ggml_backend_t ggml_backend_sycl_init(int device);
+
+// devide buffer
+GGML_API ggml_backend_buffer_type_t ggml_backend_sycl_buffer_type(int device);
+
+// split tensor buffer that splits matrices by rows across multiple devices
+GGML_API ggml_backend_buffer_type_t ggml_backend_sycl_split_buffer_type(const float * tensor_split);
+
+// pinned host buffer for use with the CPU backend for faster copies between CPU and GPU
+GGML_API ggml_backend_buffer_type_t ggml_backend_sycl_host_buffer_type(void);
+
+GGML_API void ggml_backend_sycl_print_sycl_devices(void);
+GGML_API void ggml_sycl_get_gpu_list(int *id_list, int max_len);
+GGML_API void ggml_sycl_get_device_description(int device, char *description, size_t description_size);
+GGML_API int  ggml_backend_sycl_get_device_count();
+GGML_API void ggml_backend_sycl_get_device_memory(int device, size_t *free, size_t *total);
+
+// SYCL doesn't support registering host memory, keep here for reference
+// GGML_API bool ggml_backend_sycl_register_host_buffer(void * buffer, size_t size);
+// GGML_API void ggml_backend_sycl_unregister_host_buffer(void * buffer);
+#ifdef  __cplusplus
+}
+#endif
diff --git a/src/whisper_cpp/ggml/include/ggml-vulkan.h b/src/whisper_cpp/ggml/include/ggml-vulkan.h
new file mode 100644
index 00000000..e074042e
--- /dev/null
+++ b/src/whisper_cpp/ggml/include/ggml-vulkan.h
@@ -0,0 +1,29 @@
+#pragma once
+
+#include "ggml.h"
+#include "ggml-backend.h"
+
+#ifdef  __cplusplus
+extern "C" {
+#endif
+
+#define GGML_VK_NAME "Vulkan"
+#define GGML_VK_MAX_DEVICES 16
+
+GGML_API void ggml_vk_instance_init(void);
+
+// backend API
+GGML_API ggml_backend_t ggml_backend_vk_init(size_t dev_num);
+
+GGML_API bool ggml_backend_is_vk(ggml_backend_t backend);
+GGML_API int  ggml_backend_vk_get_device_count(void);
+GGML_API void ggml_backend_vk_get_device_description(int device, char * description, size_t description_size);
+GGML_API void ggml_backend_vk_get_device_memory(int device, size_t * free, size_t * total);
+
+GGML_API ggml_backend_buffer_type_t ggml_backend_vk_buffer_type(size_t dev_num);
+// pinned host buffer for use with the CPU backend for faster copies between CPU and GPU
+GGML_API ggml_backend_buffer_type_t ggml_backend_vk_host_buffer_type(void);
+
+#ifdef  __cplusplus
+}
+#endif
diff --git a/src/whisper_cpp/ggml.h b/src/whisper_cpp/ggml/include/ggml.h
similarity index 71%
rename from src/whisper_cpp/ggml.h
rename to src/whisper_cpp/ggml/include/ggml.h
index e424db0b..e7678d07 100644
--- a/src/whisper_cpp/ggml.h
+++ b/src/whisper_cpp/ggml/include/ggml.h
@@ -1,6 +1,3 @@
-#include "R_ext/Print.h"
-#include "R_ext/Utils.h"
-#include "R_ext/Error.h"
 #pragma once
 
 //
@@ -71,7 +68,7 @@
 //
 //       ggml_graph_compute_with_ctx(ctx, &gf, n_threads);
 //
-//       Rprintf("f = %f\n", ggml_get_f32_1d(f, 0));
+//       printf("f = %f\n", ggml_get_f32_1d(f, 0));
 //
 //       ...
 //   }
@@ -207,12 +204,13 @@
 #    define GGML_ATTRIBUTE_FORMAT(...) __attribute__((format(printf, __VA_ARGS__)))
 #endif
 
-#include <stdint.h>
-#include <stddef.h>
 #include <stdbool.h>
+#include <stddef.h>
+#include <stdint.h>
+#include <stdio.h>
 
 #define GGML_FILE_MAGIC   0x67676d6c // "ggml"
-#define GGML_FILE_VERSION 1
+#define GGML_FILE_VERSION 2
 
 #define GGML_QNT_VERSION        2    // bump this on quantization format changes
 #define GGML_QNT_VERSION_FACTOR 1000 // do not change this
@@ -221,10 +219,16 @@
 #define GGML_MAX_PARAMS         2048
 #define GGML_MAX_CONTEXTS       64
 #define GGML_MAX_SRC            10
-#define GGML_MAX_NAME           64
+#define GGML_MAX_N_THREADS      512
 #define GGML_MAX_OP_PARAMS      64
+
+#ifndef GGML_MAX_NAME
+#   define GGML_MAX_NAME        64
+#endif
+
 #define GGML_DEFAULT_N_THREADS  4
 #define GGML_DEFAULT_GRAPH_SIZE 2048
+
 #if UINTPTR_MAX == 0xFFFFFFFF
     #define GGML_MEM_ALIGN 4
 #else
@@ -234,6 +238,8 @@
 #define GGML_EXIT_SUCCESS 0
 #define GGML_EXIT_ABORTED 1
 
+#define GGML_ROPE_TYPE_NEOX 2
+
 #define GGUF_MAGIC "GGUF"
 
 #define GGUF_VERSION 3
@@ -244,26 +250,27 @@
 
 #define GGML_PAD(x, n) (((x) + (n) - 1) & ~((n) - 1))
 
-#define GGML_ASSERT(x) \
-    do { \
-        if (!(x)) { \
-            R_CheckUserInterrupt(); \
-            Rprintf("GGML_ASSERT: %s:%d: %s\n", __FILE__, __LINE__, #x); \
-            ggml_print_backtrace(); \
-            Rf_error("whispercpp error"); \
-        } \
-    } while (0)
-
 #ifndef NDEBUG
-#define GGML_UNREACHABLE() GGML_ASSERT(!"statement should not be reached")
+#   define GGML_UNREACHABLE() do { fprintf(stderr, "statement should be unreachable\n"); abort(); } while(0)
 #elif defined(__GNUC__)
-#define GGML_UNREACHABLE() __builtin_unreachable()
+#   define GGML_UNREACHABLE() __builtin_unreachable()
+#elif defined(_MSC_VER)
+#   define GGML_UNREACHABLE() __assume(0)
+#else
+#   define GGML_UNREACHABLE() ((void) 0)
+#endif
+
+#ifdef __cplusplus
+#   define GGML_NORETURN [[noreturn]]
 #elif defined(_MSC_VER)
-#define GGML_UNREACHABLE() __assume(0)
+#   define GGML_NORETURN __declspec(noreturn)
 #else
-#define GGML_UNREACHABLE() ((void) 0)
+#   define GGML_NORETURN _Noreturn
 #endif
 
+#define GGML_ABORT(...) ggml_abort(__FILE__, __LINE__, __VA_ARGS__)
+#define GGML_ASSERT(x) if (!(x)) GGML_ABORT("GGML_ASSERT(%s) failed", #x)
+
 // used to copy the number of elements and stride in bytes of tensors into local variables.
 // main purpose is to reduce code duplication and improve readability.
 //
@@ -302,49 +309,87 @@
     GGML_TENSOR_LOCALS(int64_t, ne,  dst,  ne) \
     GGML_TENSOR_LOCALS(size_t,  nb,  dst,  nb)
 
+#define GGML_TENSOR_BINARY_OP_LOCALS01 \
+    GGML_TENSOR_LOCALS(int64_t, ne0, src0, ne) \
+    GGML_TENSOR_LOCALS(size_t,  nb0, src0, nb) \
+    GGML_TENSOR_LOCALS(int64_t, ne1, src1, ne) \
+    GGML_TENSOR_LOCALS(size_t,  nb1, src1, nb)
+
 #ifdef  __cplusplus
 extern "C" {
 #endif
 
-#if defined(__ARM_NEON) && defined(__CUDACC__)
-    typedef half ggml_fp16_t;
-#elif defined(__ARM_NEON) && !defined(_MSC_VER)
-    typedef __fp16 ggml_fp16_t;
-#else
-    typedef uint16_t ggml_fp16_t;
-#endif
+    GGML_NORETURN GGML_ATTRIBUTE_FORMAT(3, 4)
+    GGML_API void ggml_abort(const char * file, int line, const char * fmt, ...);
+
+    enum ggml_status {
+        GGML_STATUS_ALLOC_FAILED = -2,
+        GGML_STATUS_FAILED = -1,
+        GGML_STATUS_SUCCESS = 0,
+        GGML_STATUS_ABORTED = 1,
+    };
 
-    // convert FP16 <-> FP32
-    GGML_API float       ggml_fp16_to_fp32(ggml_fp16_t x);
-    GGML_API ggml_fp16_t ggml_fp32_to_fp16(float x);
+    // get ggml_status name string
+    GGML_API const char * ggml_status_to_string(enum ggml_status status);
 
-    GGML_API void ggml_fp16_to_fp32_row(const ggml_fp16_t * x, float * y, int n);
-    GGML_API void ggml_fp32_to_fp16_row(const float * x, ggml_fp16_t * y, int n);
+    // ieee 754-2008 half-precision float16
+    // todo: make this not an integral type
+    typedef uint16_t ggml_fp16_t;
+    GGML_API float       ggml_fp16_to_fp32(ggml_fp16_t);
+    GGML_API ggml_fp16_t ggml_fp32_to_fp16(float);
+    GGML_API void        ggml_fp16_to_fp32_row(const ggml_fp16_t *, float *, int64_t);
+    GGML_API void        ggml_fp32_to_fp16_row(const float *, ggml_fp16_t *, int64_t);
+
+    // google brain half-precision bfloat16
+    typedef struct { uint16_t bits; } ggml_bf16_t;
+    GGML_API ggml_bf16_t ggml_fp32_to_bf16(float);
+    GGML_API float       ggml_bf16_to_fp32(ggml_bf16_t);  // consider just doing << 16
+    GGML_API void        ggml_bf16_to_fp32_row(const ggml_bf16_t *, float *, int64_t);
+    GGML_API void        ggml_fp32_to_bf16_row_ref(const float *, ggml_bf16_t *, int64_t);
+    GGML_API void        ggml_fp32_to_bf16_row(const float *, ggml_bf16_t *, int64_t);
 
     struct ggml_object;
     struct ggml_context;
+    struct ggml_cgraph;
 
+    // NOTE: always add types at the end of the enum to keep backward compatibility
     enum ggml_type {
-        GGML_TYPE_F32  = 0,
-        GGML_TYPE_F16  = 1,
-        GGML_TYPE_Q4_0 = 2,
-        GGML_TYPE_Q4_1 = 3,
+        GGML_TYPE_F32     = 0,
+        GGML_TYPE_F16     = 1,
+        GGML_TYPE_Q4_0    = 2,
+        GGML_TYPE_Q4_1    = 3,
         // GGML_TYPE_Q4_2 = 4, support has been removed
-        // GGML_TYPE_Q4_3 (5) support has been removed
-        GGML_TYPE_Q5_0 = 6,
-        GGML_TYPE_Q5_1 = 7,
-        GGML_TYPE_Q8_0 = 8,
-        GGML_TYPE_Q8_1 = 9,
-        // k-quantizations
-        GGML_TYPE_Q2_K = 10,
-        GGML_TYPE_Q3_K = 11,
-        GGML_TYPE_Q4_K = 12,
-        GGML_TYPE_Q5_K = 13,
-        GGML_TYPE_Q6_K = 14,
-        GGML_TYPE_Q8_K = 15,
-        GGML_TYPE_I8,
-        GGML_TYPE_I16,
-        GGML_TYPE_I32,
+        // GGML_TYPE_Q4_3 = 5, support has been removed
+        GGML_TYPE_Q5_0    = 6,
+        GGML_TYPE_Q5_1    = 7,
+        GGML_TYPE_Q8_0    = 8,
+        GGML_TYPE_Q8_1    = 9,
+        GGML_TYPE_Q2_K    = 10,
+        GGML_TYPE_Q3_K    = 11,
+        GGML_TYPE_Q4_K    = 12,
+        GGML_TYPE_Q5_K    = 13,
+        GGML_TYPE_Q6_K    = 14,
+        GGML_TYPE_Q8_K    = 15,
+        GGML_TYPE_IQ2_XXS = 16,
+        GGML_TYPE_IQ2_XS  = 17,
+        GGML_TYPE_IQ3_XXS = 18,
+        GGML_TYPE_IQ1_S   = 19,
+        GGML_TYPE_IQ4_NL  = 20,
+        GGML_TYPE_IQ3_S   = 21,
+        GGML_TYPE_IQ2_S   = 22,
+        GGML_TYPE_IQ4_XS  = 23,
+        GGML_TYPE_I8      = 24,
+        GGML_TYPE_I16     = 25,
+        GGML_TYPE_I32     = 26,
+        GGML_TYPE_I64     = 27,
+        GGML_TYPE_F64     = 28,
+        GGML_TYPE_IQ1_M   = 29,
+        GGML_TYPE_BF16    = 30,
+        GGML_TYPE_Q4_0_4_4 = 31,
+        GGML_TYPE_Q4_0_4_8 = 32,
+        GGML_TYPE_Q4_0_8_8 = 33,
+        GGML_TYPE_TQ1_0   = 34,
+        GGML_TYPE_TQ2_0   = 35,
         GGML_TYPE_COUNT,
     };
 
@@ -355,27 +400,40 @@ extern "C" {
     };
 
     enum ggml_backend_type {
-        GGML_BACKEND_CPU = 0,
-        GGML_BACKEND_GPU = 10,
-        GGML_BACKEND_GPU_SPLIT = 20,
+        GGML_BACKEND_TYPE_CPU = 0,
+        GGML_BACKEND_TYPE_GPU = 10,
+        GGML_BACKEND_TYPE_GPU_SPLIT = 20,
     };
 
     // model file types
     enum ggml_ftype {
-        GGML_FTYPE_UNKNOWN     = -1,
-        GGML_FTYPE_ALL_F32     = 0,
-        GGML_FTYPE_MOSTLY_F16  = 1,  // except 1d tensors
-        GGML_FTYPE_MOSTLY_Q4_0 = 2,  // except 1d tensors
-        GGML_FTYPE_MOSTLY_Q4_1 = 3,  // except 1d tensors
+        GGML_FTYPE_UNKNOWN        = -1,
+        GGML_FTYPE_ALL_F32        = 0,
+        GGML_FTYPE_MOSTLY_F16     = 1,  // except 1d tensors
+        GGML_FTYPE_MOSTLY_Q4_0    = 2,  // except 1d tensors
+        GGML_FTYPE_MOSTLY_Q4_1    = 3,  // except 1d tensors
         GGML_FTYPE_MOSTLY_Q4_1_SOME_F16 = 4, // tok_embeddings.weight and output.weight are F16
-        GGML_FTYPE_MOSTLY_Q8_0 = 7,  // except 1d tensors
-        GGML_FTYPE_MOSTLY_Q5_0 = 8,  // except 1d tensors
-        GGML_FTYPE_MOSTLY_Q5_1 = 9,  // except 1d tensors
-        GGML_FTYPE_MOSTLY_Q2_K = 10, // except 1d tensors
-        GGML_FTYPE_MOSTLY_Q3_K = 11, // except 1d tensors
-        GGML_FTYPE_MOSTLY_Q4_K = 12, // except 1d tensors
-        GGML_FTYPE_MOSTLY_Q5_K = 13, // except 1d tensors
-        GGML_FTYPE_MOSTLY_Q6_K = 14, // except 1d tensors
+        GGML_FTYPE_MOSTLY_Q8_0    = 7,  // except 1d tensors
+        GGML_FTYPE_MOSTLY_Q5_0    = 8,  // except 1d tensors
+        GGML_FTYPE_MOSTLY_Q5_1    = 9,  // except 1d tensors
+        GGML_FTYPE_MOSTLY_Q2_K    = 10, // except 1d tensors
+        GGML_FTYPE_MOSTLY_Q3_K    = 11, // except 1d tensors
+        GGML_FTYPE_MOSTLY_Q4_K    = 12, // except 1d tensors
+        GGML_FTYPE_MOSTLY_Q5_K    = 13, // except 1d tensors
+        GGML_FTYPE_MOSTLY_Q6_K    = 14, // except 1d tensors
+        GGML_FTYPE_MOSTLY_IQ2_XXS = 15, // except 1d tensors
+        GGML_FTYPE_MOSTLY_IQ2_XS  = 16, // except 1d tensors
+        GGML_FTYPE_MOSTLY_IQ3_XXS = 17, // except 1d tensors
+        GGML_FTYPE_MOSTLY_IQ1_S   = 18, // except 1d tensors
+        GGML_FTYPE_MOSTLY_IQ4_NL  = 19, // except 1d tensors
+        GGML_FTYPE_MOSTLY_IQ3_S   = 20, // except 1d tensors
+        GGML_FTYPE_MOSTLY_IQ2_S   = 21, // except 1d tensors
+        GGML_FTYPE_MOSTLY_IQ4_XS  = 22, // except 1d tensors
+        GGML_FTYPE_MOSTLY_IQ1_M   = 23, // except 1d tensors
+        GGML_FTYPE_MOSTLY_BF16    = 24, // except 1d tensors
+        GGML_FTYPE_MOSTLY_Q4_0_4_4 = 25, // except 1d tensors
+        GGML_FTYPE_MOSTLY_Q4_0_4_8 = 26, // except 1d tensors
+        GGML_FTYPE_MOSTLY_Q4_0_8_8 = 27, // except 1d tensors
     };
 
     // available tensor operations:
@@ -392,10 +450,13 @@ extern "C" {
         GGML_OP_SQR,
         GGML_OP_SQRT,
         GGML_OP_LOG,
+        GGML_OP_SIN,
+        GGML_OP_COS,
         GGML_OP_SUM,
         GGML_OP_SUM_ROWS,
         GGML_OP_MEAN,
         GGML_OP_ARGMAX,
+        GGML_OP_COUNT_EQUAL,
         GGML_OP_REPEAT,
         GGML_OP_REPEAT_BACK,
         GGML_OP_CONCAT,
@@ -426,25 +487,30 @@ extern "C" {
         GGML_OP_SOFT_MAX_BACK,
         GGML_OP_ROPE,
         GGML_OP_ROPE_BACK,
-        GGML_OP_ALIBI,
         GGML_OP_CLAMP,
         GGML_OP_CONV_TRANSPOSE_1D,
         GGML_OP_IM2COL,
+        GGML_OP_IM2COL_BACK,
         GGML_OP_CONV_TRANSPOSE_2D,
         GGML_OP_POOL_1D,
         GGML_OP_POOL_2D,
+        GGML_OP_POOL_2D_BACK,
         GGML_OP_UPSCALE, // nearest interpolate
         GGML_OP_PAD,
+        GGML_OP_ARANGE,
+        GGML_OP_TIMESTEP_EMBEDDING,
         GGML_OP_ARGSORT,
         GGML_OP_LEAKY_RELU,
 
-        GGML_OP_FLASH_ATTN,
-        GGML_OP_FLASH_FF,
+        GGML_OP_FLASH_ATTN_EXT,
         GGML_OP_FLASH_ATTN_BACK,
+        GGML_OP_SSM_CONV,
+        GGML_OP_SSM_SCAN,
         GGML_OP_WIN_PART,
         GGML_OP_WIN_UNPART,
         GGML_OP_GET_REL_POS,
         GGML_OP_ADD_REL_POS,
+        GGML_OP_RWKV_WKV,
 
         GGML_OP_UNARY,
 
@@ -461,6 +527,7 @@ extern "C" {
 
         GGML_OP_CROSS_ENTROPY_LOSS,
         GGML_OP_CROSS_ENTROPY_LOSS_BACK,
+        GGML_OP_OPT_STEP_ADAMW,
 
         GGML_OP_COUNT,
     };
@@ -473,44 +540,45 @@ extern "C" {
         GGML_UNARY_OP_TANH,
         GGML_UNARY_OP_ELU,
         GGML_UNARY_OP_RELU,
+        GGML_UNARY_OP_SIGMOID,
         GGML_UNARY_OP_GELU,
         GGML_UNARY_OP_GELU_QUICK,
         GGML_UNARY_OP_SILU,
+        GGML_UNARY_OP_HARDSWISH,
+        GGML_UNARY_OP_HARDSIGMOID,
+        GGML_UNARY_OP_EXP,
 
         GGML_UNARY_OP_COUNT,
     };
 
     enum ggml_object_type {
-        GGML_OBJECT_TENSOR,
-        GGML_OBJECT_GRAPH,
-        GGML_OBJECT_WORK_BUFFER
+        GGML_OBJECT_TYPE_TENSOR,
+        GGML_OBJECT_TYPE_GRAPH,
+        GGML_OBJECT_TYPE_WORK_BUFFER
     };
 
     enum ggml_log_level {
-        GGML_LOG_LEVEL_ERROR = 2,
-        GGML_LOG_LEVEL_WARN = 3,
-        GGML_LOG_LEVEL_INFO = 4,
-        GGML_LOG_LEVEL_DEBUG = 5
+        GGML_LOG_LEVEL_NONE  = 0,
+        GGML_LOG_LEVEL_INFO  = 1,
+        GGML_LOG_LEVEL_WARN  = 2,
+        GGML_LOG_LEVEL_ERROR = 3,
+        GGML_LOG_LEVEL_DEBUG = 4,
+        GGML_LOG_LEVEL_CONT  = 5, // continue previous log
     };
 
-    // ggml object
-    struct ggml_object {
-        size_t offs;
-        size_t size;
-
-        struct ggml_object * next;
-
-        enum ggml_object_type type;
-
-        char padding[4];
+    // this tensor...
+    enum ggml_tensor_flag {
+        GGML_TENSOR_FLAG_INPUT  =  1, // ...is an input for the GGML compute graph
+        GGML_TENSOR_FLAG_OUTPUT =  2, // ...is an output for the GGML compute graph
+        GGML_TENSOR_FLAG_PARAM  =  4, // ...contains trainable parameters
+        GGML_TENSOR_FLAG_LOSS   =  8, // ...defines loss for numerical optimization (multiple loss tensors add up)
     };
 
-    static const size_t GGML_OBJECT_SIZE = sizeof(struct ggml_object);
-
     // n-dimensional tensor
     struct ggml_tensor {
-        enum ggml_type         type;
-        enum ggml_backend_type backend;
+        enum ggml_type type;
+
+        GGML_DEPRECATED(enum ggml_backend_type backend, "use the buffer type to find the storage location of the tensor");
 
         struct ggml_backend_buffer * buffer;
 
@@ -526,16 +594,12 @@ extern "C" {
         // op params - allocated as int32_t for alignment
         int32_t op_params[GGML_MAX_OP_PARAMS / sizeof(int32_t)];
 
-        bool is_param;
+        int32_t flags;
 
         struct ggml_tensor * grad;
         struct ggml_tensor * src[GGML_MAX_SRC];
 
-        // performance
-        int     perf_runs;
-        int64_t perf_cycles;
-        int64_t perf_time_us;
-
+        // source tensor and offset for views
         struct ggml_tensor * view_src;
         size_t               view_offs;
 
@@ -545,11 +609,39 @@ extern "C" {
 
         void * extra; // extra things e.g. for ggml-cuda.cu
 
-        char padding[8];
+        // char padding[4];
     };
 
     static const size_t GGML_TENSOR_SIZE = sizeof(struct ggml_tensor);
 
+    // Abort callback
+    // If not NULL, called before ggml computation
+    // If it returns true, the computation is aborted
+    typedef bool (*ggml_abort_callback)(void * data);
+
+    // Scheduling priorities
+    enum ggml_sched_priority {
+        GGML_SCHED_PRIO_NORMAL,
+        GGML_SCHED_PRIO_MEDIUM,
+        GGML_SCHED_PRIO_HIGH,
+        GGML_SCHED_PRIO_REALTIME
+    };
+
+    // Threadpool params
+    // Use ggml_threadpool_params_default() or ggml_threadpool_params_init() to populate the defaults
+    struct ggml_threadpool_params {
+        bool                cpumask[GGML_MAX_N_THREADS]; // mask of cpu cores (all-zeros means use default affinity settings)
+        int                 n_threads;                   // number of threads
+        enum ggml_sched_priority prio;                   // thread priority
+        uint32_t            poll;                        // polling level (0 - no polling, 100 - aggressive polling)
+        bool                strict_cpu;                  // strict cpu placement
+        bool                paused;                      // start in paused state
+    };
+
+    struct ggml_threadpool;     // forward declaration, see ggml.c
+
+    typedef struct ggml_threadpool * ggml_threadpool_t;
+
     // the compute plan that needs to be prepared for ggml_graph_compute()
     // since https://github.com/ggerganov/ggml/issues/287
     struct ggml_cplan {
@@ -557,41 +649,11 @@ extern "C" {
         uint8_t * work_data; // work buffer, to be allocated by caller before calling to `ggml_graph_compute()`
 
         int n_threads;
+        struct ggml_threadpool * threadpool;
 
         // abort ggml_graph_compute when true
-        bool (*abort_callback)(void * data);
-        void * abort_callback_data;
-    };
-
-    enum ggml_cgraph_eval_order {
-        GGML_CGRAPH_EVAL_ORDER_LEFT_TO_RIGHT = 0,
-        GGML_CGRAPH_EVAL_ORDER_RIGHT_TO_LEFT,
-        GGML_CGRAPH_EVAL_ORDER_COUNT
-    };
-
-    struct ggml_hash_set {
-        size_t size;
-        struct ggml_tensor ** keys;
-    };
-
-    // computation graph
-    struct ggml_cgraph {
-        int size;
-        int n_nodes;
-        int n_leafs;
-
-        struct ggml_tensor ** nodes;
-        struct ggml_tensor ** grads;
-        struct ggml_tensor ** leafs;
-
-        struct ggml_hash_set visited_hash_table;
-
-        enum ggml_cgraph_eval_order order;
-
-        // performance
-        int     perf_runs;
-        int64_t perf_cycles;
-        int64_t perf_time_us;
+        ggml_abort_callback abort_callback;
+        void *              abort_callback_data;
     };
 
     // scratch buffer
@@ -608,27 +670,25 @@ extern "C" {
         bool   no_alloc;   // don't allocate memory for the tensor data
     };
 
-
-    // compute types
-
-    // NOTE: the INIT or FINALIZE pass is not scheduled unless explicitly enabled.
-    // This behavior was changed since https://github.com/ggerganov/llama.cpp/pull/1995.
-    enum ggml_task_type {
-        GGML_TASK_INIT = 0,
-        GGML_TASK_COMPUTE,
-        GGML_TASK_FINALIZE,
+    // numa strategies
+    enum ggml_numa_strategy {
+        GGML_NUMA_STRATEGY_DISABLED   = 0,
+        GGML_NUMA_STRATEGY_DISTRIBUTE = 1,
+        GGML_NUMA_STRATEGY_ISOLATE    = 2,
+        GGML_NUMA_STRATEGY_NUMACTL    = 3,
+        GGML_NUMA_STRATEGY_MIRROR     = 4,
+        GGML_NUMA_STRATEGY_COUNT
     };
 
-    struct ggml_compute_params {
-        enum ggml_task_type type;
+    //
+    // GUID
+    //
 
-        // ith = thread index, nth = number of threads
-        int ith, nth;
+    // GUID types
+    typedef uint8_t ggml_guid[16];
+    typedef ggml_guid * ggml_guid_t;
 
-        // work buffer for all threads
-        size_t wsize;
-        void * wdata;
-    };
+    GGML_API bool ggml_guid_matches(ggml_guid_t guid_a, ggml_guid_t guid_b);
 
     // misc
 
@@ -638,22 +698,23 @@ extern "C" {
     GGML_API int64_t ggml_cycles(void);
     GGML_API int64_t ggml_cycles_per_ms(void);
 
-    GGML_API void    ggml_print_backtrace(void);
+    // accepts a UTF-8 path, even on Windows
+    GGML_API FILE *  ggml_fopen(const char * fname, const char * mode);
 
-    GGML_API void    ggml_numa_init(void); // call once for better performance on NUMA systems
+    GGML_API void    ggml_numa_init(enum ggml_numa_strategy numa); // call once for better performance on NUMA systems
     GGML_API bool    ggml_is_numa(void); // true if init detected that system has >1 NUMA node
 
     GGML_API void    ggml_print_object (const struct ggml_object * obj);
     GGML_API void    ggml_print_objects(const struct ggml_context * ctx);
 
-    GGML_API int64_t ggml_nelements   (const struct ggml_tensor * tensor);
-    GGML_API int64_t ggml_nrows       (const struct ggml_tensor * tensor);
-    GGML_API size_t  ggml_nbytes      (const struct ggml_tensor * tensor);
-    GGML_API size_t  ggml_nbytes_pad  (const struct ggml_tensor * tensor); // same as ggml_nbytes() but padded to GGML_MEM_ALIGN
+    GGML_API int64_t ggml_nelements (const struct ggml_tensor * tensor);
+    GGML_API int64_t ggml_nrows     (const struct ggml_tensor * tensor);
+    GGML_API size_t  ggml_nbytes    (const struct ggml_tensor * tensor);
+    GGML_API size_t  ggml_nbytes_pad(const struct ggml_tensor * tensor); // same as ggml_nbytes() but padded to GGML_MEM_ALIGN
 
-    GGML_API int    ggml_blck_size(enum ggml_type type);
-    GGML_API size_t ggml_type_size(enum ggml_type type);             // size in bytes for all elements in a block
-    GGML_API size_t ggml_row_size (enum ggml_type type, int64_t ne); // size in bytes for all elements in a row
+    GGML_API int64_t ggml_blck_size(enum ggml_type type);
+    GGML_API size_t  ggml_type_size(enum ggml_type type);             // size in bytes for all elements in a block
+    GGML_API size_t  ggml_row_size (enum ggml_type type, int64_t ne); // size in bytes for all elements in a row
 
     GGML_DEPRECATED(
     GGML_API double ggml_type_sizef(enum ggml_type type), // ggml_type_size()/ggml_blck_size() as float
@@ -674,19 +735,29 @@ extern "C" {
     GGML_API enum ggml_type ggml_ftype_to_ggml_type(enum ggml_ftype ftype);
 
     GGML_API bool ggml_is_transposed(const struct ggml_tensor * tensor);
-    GGML_API bool ggml_is_contiguous(const struct ggml_tensor * tensor);
     GGML_API bool ggml_is_permuted  (const struct ggml_tensor * tensor);
+    GGML_API bool ggml_is_empty     (const struct ggml_tensor * tensor);
     GGML_API bool ggml_is_scalar    (const struct ggml_tensor * tensor);
     GGML_API bool ggml_is_vector    (const struct ggml_tensor * tensor);
     GGML_API bool ggml_is_matrix    (const struct ggml_tensor * tensor);
     GGML_API bool ggml_is_3d        (const struct ggml_tensor * tensor);
     GGML_API int  ggml_n_dims       (const struct ggml_tensor * tensor); // returns 1 for scalars
 
-    GGML_API bool ggml_are_same_shape(const struct ggml_tensor * t0, const struct ggml_tensor * t1);
+    GGML_API bool ggml_is_contiguous  (const struct ggml_tensor * tensor);
+    GGML_API bool ggml_is_contiguous_0(const struct ggml_tensor * tensor); // same as ggml_is_contiguous()
+    GGML_API bool ggml_is_contiguous_1(const struct ggml_tensor * tensor); // contiguous for dims >= 1
+    GGML_API bool ggml_is_contiguous_2(const struct ggml_tensor * tensor); // contiguous for dims >= 2
+
+    GGML_API bool ggml_are_same_shape (const struct ggml_tensor * t0, const struct ggml_tensor * t1);
+    GGML_API bool ggml_are_same_stride(const struct ggml_tensor * t0, const struct ggml_tensor * t1);
+
+    GGML_API bool ggml_can_repeat(const struct ggml_tensor * t0, const struct ggml_tensor * t1);
 
     // use this to compute the memory overhead of a tensor
     GGML_API size_t ggml_tensor_overhead(void);
 
+    GGML_API bool ggml_validate_row_data(enum ggml_type type, const void * data, size_t nbytes);
+
     // main
 
     GGML_API struct ggml_context * ggml_init(struct ggml_init_params params);
@@ -888,6 +959,22 @@ extern "C" {
             struct ggml_context * ctx,
             struct ggml_tensor  * a);
 
+    GGML_API struct ggml_tensor * ggml_sin(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a);
+
+    GGML_API struct ggml_tensor * ggml_sin_inplace(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a);
+
+    GGML_API struct ggml_tensor * ggml_cos(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a);
+
+    GGML_API struct ggml_tensor * ggml_cos_inplace(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a);
+
     // return scalar
     GGML_API struct ggml_tensor * ggml_sum(
             struct ggml_context * ctx,
@@ -908,6 +995,12 @@ extern "C" {
             struct ggml_context * ctx,
             struct ggml_tensor  * a);
 
+    // count number of equal elements in a and b
+    GGML_API struct ggml_tensor * ggml_count_equal(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a,
+            struct ggml_tensor  * b);
+
     // if a is the same shape as b, and a is not parameter, return a
     // otherwise, return a new tensor: repeat(a) to fit in b
     GGML_API struct ggml_tensor * ggml_repeat(
@@ -921,12 +1014,13 @@ extern "C" {
             struct ggml_tensor  * a,
             struct ggml_tensor  * b);
 
-    // concat a and b on dim 2
+    // concat a and b along dim
     // used in stable-diffusion
     GGML_API struct ggml_tensor * ggml_concat(
             struct ggml_context * ctx,
             struct ggml_tensor  * a,
-            struct ggml_tensor  * b);
+            struct ggml_tensor  * b,
+            int                   dim);
 
     GGML_API struct ggml_tensor * ggml_abs(
             struct ggml_context * ctx,
@@ -988,6 +1082,14 @@ extern "C" {
             struct ggml_context * ctx,
             struct ggml_tensor  * a);
 
+    GGML_API struct ggml_tensor * ggml_sigmoid(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a);
+
+    GGML_API struct ggml_tensor * ggml_sigmoid_inplace(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a);
+
     GGML_API struct ggml_tensor * ggml_gelu(
             struct ggml_context * ctx,
             struct ggml_tensor  * a);
@@ -1019,6 +1121,24 @@ extern "C" {
             struct ggml_tensor  * a,
             struct ggml_tensor  * b);
 
+    // hardswish(x) = x * relu6(x + 3) / 6
+    GGML_API struct ggml_tensor * ggml_hardswish(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a);
+
+    // hardsigmoid(x) = relu6(x + 3) / 6
+    GGML_API struct ggml_tensor * ggml_hardsigmoid(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a);
+
+    GGML_API struct ggml_tensor * ggml_exp(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a);
+
+    GGML_API struct ggml_tensor * ggml_exp_inplace(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a);
+
     // normalize along rows
     GGML_API struct ggml_tensor * ggml_norm(
             struct ggml_context * ctx,
@@ -1042,16 +1162,17 @@ extern "C" {
 
     // group normalize along ne0*ne1*n_groups
     // used in stable-diffusion
-    // TODO: eps is hardcoded to 1e-6 for now
     GGML_API struct ggml_tensor * ggml_group_norm(
             struct ggml_context * ctx,
             struct ggml_tensor  * a,
-            int                   n_groups);
+            int                   n_groups,
+            float                 eps);
 
     GGML_API struct ggml_tensor * ggml_group_norm_inplace(
             struct ggml_context * ctx,
             struct ggml_tensor  * a,
-            int                   n_groups);
+            int                   n_groups,
+            float                 eps);
 
     // a - x
     // b - dy
@@ -1076,14 +1197,11 @@ extern "C" {
             enum ggml_prec       prec);
 
     // indirect matrix multiplication
-    //  ggml_mul_mat_id(ctx, as, ids, id, b) ~= ggml_mul_mat(as[ids[id]], b)
     GGML_API struct ggml_tensor * ggml_mul_mat_id(
             struct ggml_context * ctx,
-            struct ggml_tensor  * const as[],
-            int                   n_as,
-            struct ggml_tensor  * ids,
-            int                   id,
-            struct ggml_tensor  * b);
+            struct ggml_tensor  * as,
+            struct ggml_tensor  * b,
+            struct ggml_tensor  * ids);
 
     // A: m columns, n rows,
     // B: p columns, n rows,
@@ -1116,7 +1234,7 @@ extern "C" {
             size_t                nb1,
             size_t                nb2,
             size_t                nb3,
-            size_t                offset);
+            size_t                offset); // in bytes
 
     // b -> view(a,offset,nb1,nb2,3), return view(a)
     GGML_API struct ggml_tensor * ggml_set_inplace(
@@ -1126,19 +1244,19 @@ extern "C" {
             size_t                nb1,
             size_t                nb2,
             size_t                nb3,
-            size_t                offset);
+            size_t                offset); // in bytes
 
     GGML_API struct ggml_tensor * ggml_set_1d(
             struct ggml_context * ctx,
             struct ggml_tensor  * a,
             struct ggml_tensor  * b,
-            size_t                offset);
+            size_t                offset); // in bytes
 
     GGML_API struct ggml_tensor * ggml_set_1d_inplace(
             struct ggml_context * ctx,
             struct ggml_tensor  * a,
             struct ggml_tensor  * b,
-            size_t                offset);
+            size_t                offset); // in bytes
 
     // b -> view(a,offset,nb1,nb2,3), return modified a
     GGML_API struct ggml_tensor * ggml_set_2d(
@@ -1146,7 +1264,7 @@ extern "C" {
             struct ggml_tensor  * a,
             struct ggml_tensor  * b,
             size_t                nb1,
-            size_t                offset);
+            size_t                offset); // in bytes
 
     // b -> view(a,offset,nb1,nb2,3), return view(a)
     GGML_API struct ggml_tensor * ggml_set_2d_inplace(
@@ -1154,7 +1272,7 @@ extern "C" {
             struct ggml_tensor  * a,
             struct ggml_tensor  * b,
             size_t                nb1,
-            size_t                offset);
+            size_t                offset); // in bytes
 
     // a -> b, return view(b)
     GGML_API struct ggml_tensor * ggml_cpy(
@@ -1162,22 +1280,16 @@ extern "C" {
             struct ggml_tensor  * a,
             struct ggml_tensor  * b);
 
-    // a -> b, in-place, return view(b)
-    GGML_API struct ggml_tensor * ggml_cpy_inplace(
+    GGML_API struct ggml_tensor * ggml_cast(
             struct ggml_context * ctx,
             struct ggml_tensor  * a,
-            struct ggml_tensor  * b);
+            enum   ggml_type      type);
 
     // make contiguous
     GGML_API struct ggml_tensor * ggml_cont(
             struct ggml_context * ctx,
             struct ggml_tensor  * a);
 
-    // make contiguous, in-place
-    GGML_API struct ggml_tensor * ggml_cont_inplace(
-            struct ggml_context * ctx,
-            struct ggml_tensor  * a);
-
     // make contiguous, with new shape
     GGML_API struct ggml_tensor * ggml_cont_1d(
             struct ggml_context * ctx,
@@ -1295,14 +1407,14 @@ extern "C" {
     // supports 3D: a->ne[2] == b->ne[1]
     GGML_API struct ggml_tensor * ggml_get_rows(
             struct ggml_context * ctx,
-            struct ggml_tensor  * a,
-            struct ggml_tensor  * b);
+            struct ggml_tensor  * a,  // data
+            struct ggml_tensor  * b); // row indices
 
     GGML_API struct ggml_tensor * ggml_get_rows_back(
             struct ggml_context * ctx,
-            struct ggml_tensor  * a,
-            struct ggml_tensor  * b,
-            struct ggml_tensor  * c);
+            struct ggml_tensor  * a,  // gradients of ggml_get_rows result
+            struct ggml_tensor  * b,  // row indices
+            struct ggml_tensor  * c); // data for ggml_get_rows, only used for its shape
 
     GGML_API struct ggml_tensor * ggml_diag(
         struct ggml_context     * ctx,
@@ -1341,13 +1453,15 @@ extern "C" {
             struct ggml_context * ctx,
             struct ggml_tensor  * a);
 
-    // fused soft_max(a*scale + mask)
+    // fused soft_max(a*scale + mask*(ALiBi slope))
     // mask is optional
+    // max_bias = 0.0f for no ALiBi
     GGML_API struct ggml_tensor * ggml_soft_max_ext(
             struct ggml_context * ctx,
             struct ggml_tensor  * a,
             struct ggml_tensor  * mask,
-            float                 scale);
+            float                 scale,
+            float                 max_bias);
 
     GGML_API struct ggml_tensor * ggml_soft_max_back(
             struct ggml_context * ctx,
@@ -1361,9 +1475,8 @@ extern "C" {
             struct ggml_tensor  * b);
 
     // rotary position embedding
-    // if mode & 1 == 1, skip n_past elements (DEPRECATED)
-    // if mode & 2 == 1, GPT-NeoX style
-    // if mode & 4 == 1, ChatGLM style
+    // if (mode & 1) - skip n_past elements (NOT SUPPORTED)
+    // if (mode & GGML_ROPE_TYPE_NEOX) - GPT-NeoX style
     //
     // b is an int32 vector with size a->ne[2], it contains the positions
     GGML_API struct ggml_tensor * ggml_rope(
@@ -1371,8 +1484,7 @@ extern "C" {
             struct ggml_tensor  * a,
             struct ggml_tensor  * b,
             int                   n_dims,
-            int                   mode,
-            int                   n_ctx);
+            int                   mode);
 
     // in-place, returns view(a)
     GGML_API struct ggml_tensor * ggml_rope_inplace(
@@ -1380,18 +1492,18 @@ extern "C" {
             struct ggml_tensor  * a,
             struct ggml_tensor  * b,
             int                   n_dims,
-            int                   mode,
-            int                   n_ctx);
+            int                   mode);
 
     // custom RoPE
-    GGML_API struct ggml_tensor * ggml_rope_custom(
+    // c is freq factors (e.g. phi3-128k), (optional)
+    GGML_API struct ggml_tensor * ggml_rope_ext(
             struct ggml_context * ctx,
             struct ggml_tensor  * a,
             struct ggml_tensor  * b,
+            struct ggml_tensor  * c,
             int                   n_dims,
             int                   mode,
-            int                   n_ctx,
-            int                   n_orig_ctx,
+            int                   n_ctx_orig,
             float                 freq_base,
             float                 freq_scale,
             float                 ext_factor,
@@ -1400,14 +1512,14 @@ extern "C" {
             float                 beta_slow);
 
     // in-place, returns view(a)
-    GGML_API struct ggml_tensor * ggml_rope_custom_inplace(
+    GGML_API struct ggml_tensor * ggml_rope_ext_inplace(
             struct ggml_context * ctx,
             struct ggml_tensor  * a,
             struct ggml_tensor  * b,
+            struct ggml_tensor  * c,
             int                   n_dims,
             int                   mode,
-            int                   n_ctx,
-            int                   n_orig_ctx,
+            int                   n_ctx_orig,
             float                 freq_base,
             float                 freq_scale,
             float                 ext_factor,
@@ -1415,46 +1527,56 @@ extern "C" {
             float                 beta_fast,
             float                 beta_slow);
 
-    // compute correction dims for YaRN RoPE scaling
-    void ggml_rope_yarn_corr_dims(
-        int n_dims, int n_orig_ctx, float freq_base, float beta_fast, float beta_slow, float dims[2]);
-
-    // xPos RoPE, in-place, returns view(a)
-    GGML_API struct ggml_tensor * ggml_rope_xpos_inplace(
+    GGML_DEPRECATED(GGML_API struct ggml_tensor * ggml_rope_custom(
             struct ggml_context * ctx,
             struct ggml_tensor  * a,
             struct ggml_tensor  * b,
             int                   n_dims,
-            float                 base,
-            bool                  down);
+            int                   mode,
+            int                   n_ctx_orig,
+            float                 freq_base,
+            float                 freq_scale,
+            float                 ext_factor,
+            float                 attn_factor,
+            float                 beta_fast,
+            float                 beta_slow),
+        "use ggml_rope_ext instead");
 
-    // rotary position embedding backward, i.e compute dx from dy
-    // a - dy
-    GGML_API struct ggml_tensor * ggml_rope_back(
+    GGML_DEPRECATED(GGML_API struct ggml_tensor * ggml_rope_custom_inplace(
             struct ggml_context * ctx,
             struct ggml_tensor  * a,
             struct ggml_tensor  * b,
             int                   n_dims,
             int                   mode,
-            int                   n_ctx,
-            int                   n_orig_ctx,
+            int                   n_ctx_orig,
             float                 freq_base,
             float                 freq_scale,
             float                 ext_factor,
             float                 attn_factor,
             float                 beta_fast,
-            float                 beta_slow,
-            float                 xpos_base,
-            bool                  xpos_down);
+            float                 beta_slow),
+        "use ggml_rope_ext_inplace instead");
 
-    // alibi position embedding
-    // in-place, returns view(a)
-    GGML_API struct ggml_tensor * ggml_alibi(
+    // compute correction dims for YaRN RoPE scaling
+    void ggml_rope_yarn_corr_dims(
+        int n_dims, int n_ctx_orig, float freq_base, float beta_fast, float beta_slow, float dims[2]);
+
+    // rotary position embedding backward, i.e compute dx from dy
+    // a - dy
+    GGML_API struct ggml_tensor * ggml_rope_back(
             struct ggml_context * ctx,
-            struct ggml_tensor  * a,
-            int                   n_past,
-            int                   n_head,
-            float                 bias_max);
+            struct ggml_tensor  * a, // gradients of ggml_rope result
+            struct ggml_tensor  * b, // positions
+            struct ggml_tensor  * c, // freq factors
+            int                   n_dims,
+            int                   mode,
+            int                   n_ctx_orig,
+            float                 freq_base,
+            float                 freq_scale,
+            float                 ext_factor,
+            float                 attn_factor,
+            float                 beta_fast,
+            float                 beta_slow);
 
     // clamp
     // in-place, returns view(a)
@@ -1464,22 +1586,49 @@ extern "C" {
             float                 min,
             float                 max);
 
+    // im2col
+    // converts data into a format that effectively results in a convolution when combined with matrix multiplication
     GGML_API struct ggml_tensor * ggml_im2col(
             struct ggml_context * ctx,
-            struct ggml_tensor  * a,
-            struct ggml_tensor  * b,
-            int                  s0,
-            int                  s1,
-            int                  p0,
-            int                  p1,
-            int                  d0,
-            int                  d1,
-            bool                 is_2D);
+            struct ggml_tensor  * a,  // convolution kernel
+            struct ggml_tensor  * b,  // data
+            int                   s0, // stride dimension 0
+            int                   s1, // stride dimension 1
+            int                   p0, // padding dimension 0
+            int                   p1, // padding dimension 1
+            int                   d0, // dilation dimension 0
+            int                   d1, // dilation dimension 1
+            bool                  is_2D,
+            enum ggml_type        dst_type);
+
+    GGML_API struct ggml_tensor * ggml_im2col_back(
+        struct ggml_context * ctx,
+        struct ggml_tensor  * a,  // convolution kernel
+        struct ggml_tensor  * b,  // gradient of im2col output
+        int64_t             * ne, // shape of im2col input
+        int                   s0, // stride dimension 0
+        int                   s1, // stride dimension 1
+        int                   p0, // padding dimension 0
+        int                   p1, // padding dimension 1
+        int                   d0, // dilation dimension 0
+        int                   d1, // dilation dimension 1
+        bool                  is_2D);
+
+    GGML_API struct ggml_tensor * ggml_conv_depthwise_2d(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a,  // convolution kernel
+            struct ggml_tensor  * b,  // data
+            int                  s0,  // stride dimension 0
+            int                  s1,  // stride dimension 1
+            int                  p0,  // padding dimension 0
+            int                  p1,  // padding dimension 1
+            int                  d0,  // dilation dimension 0
+            int                  d1); // dilation dimension 1
 
     GGML_API struct ggml_tensor * ggml_conv_1d(
             struct ggml_context * ctx,
-            struct ggml_tensor  * a,
-            struct ggml_tensor  * b,
+            struct ggml_tensor  * a,   // convolution kernel
+            struct ggml_tensor  * b,   // data
             int                   s0,  // stride
             int                   p0,  // padding
             int                   d0); // dilation
@@ -1488,29 +1637,29 @@ extern "C" {
     // alias for ggml_conv_1d(a, b, s, a->ne[0]/2, d)
     GGML_API struct ggml_tensor* ggml_conv_1d_ph(
             struct ggml_context * ctx,
-            struct ggml_tensor  * a,
-            struct ggml_tensor  * b,
-            int                   s,
-            int                   d);
+            struct ggml_tensor  * a,  // convolution kernel
+            struct ggml_tensor  * b,  // data
+            int                   s,  // stride
+            int                   d); // dilation
 
     GGML_API struct ggml_tensor * ggml_conv_transpose_1d(
             struct ggml_context * ctx,
-            struct ggml_tensor  * a,
-            struct ggml_tensor  * b,
-            int                   s0,
-            int                   p0,
-            int                   d0);
+            struct ggml_tensor  * a,   // convolution kernel
+            struct ggml_tensor  * b,   // data
+            int                   s0,  // stride
+            int                   p0,  // padding
+            int                   d0); // dilation
 
     GGML_API struct ggml_tensor * ggml_conv_2d(
             struct ggml_context * ctx,
-            struct ggml_tensor  * a,
-            struct ggml_tensor  * b,
-            int                   s0,
-            int                   s1,
-            int                   p0,
-            int                   p1,
-            int                   d0,
-            int                   d1);
+            struct ggml_tensor  * a,   // convolution kernel
+            struct ggml_tensor  * b,   // data
+            int                   s0,  // stride dimension 0
+            int                   s1,  // stride dimension 1
+            int                   p0,  // padding dimension 0
+            int                   p1,  // padding dimension 1
+            int                   d0,  // dilation dimension 0
+            int                   d1); // dilation dimension 1
 
 
     // kernel size is a->ne[0] x a->ne[1]
@@ -1572,13 +1721,37 @@ extern "C" {
             float                 p0,
             float                 p1);
 
+    GGML_API struct ggml_tensor * ggml_pool_2d_back(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a,
+            struct ggml_tensor  * af, // "a"/input used in forward pass
+            enum ggml_op_pool     op,
+            int                   k0,
+            int                   k1,
+            int                   s0,
+            int                   s1,
+            float                 p0,
+            float                 p1);
+
     // nearest interpolate
+    // multiplies ne0 and ne1 by scale factor
     // used in stable-diffusion
     GGML_API struct ggml_tensor * ggml_upscale(
             struct ggml_context * ctx,
             struct ggml_tensor  * a,
             int                   scale_factor);
 
+    // nearest interpolate
+    // nearest interpolate to specified dimensions
+    // used in tortoise.cpp
+    GGML_API struct ggml_tensor * ggml_upscale_ext(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a,
+            int                   ne0,
+            int                   ne1,
+            int                   ne2,
+            int                   ne3);
+
     // pad each dimension with zeros: [x, ..., x] -> [x, ..., x, 0, ..., 0]
     GGML_API struct ggml_tensor * ggml_pad(
             struct ggml_context * ctx,
@@ -1588,10 +1761,19 @@ extern "C" {
             int                  p2,
             int                  p3);
 
+    // Ref: https://github.com/CompVis/stable-diffusion/blob/main/ldm/modules/diffusionmodules/util.py#L151
+    // timesteps: [N,]
+    // return: [N, dim]
+    GGML_API struct ggml_tensor * ggml_timestep_embedding(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * timesteps,
+            int                   dim,
+            int                   max_period);
+
     // sort rows
     enum ggml_sort_order {
-        GGML_SORT_ASC,
-        GGML_SORT_DESC,
+        GGML_SORT_ORDER_ASC,
+        GGML_SORT_ORDER_DESC,
     };
 
     GGML_API struct ggml_tensor * ggml_argsort(
@@ -1599,19 +1781,40 @@ extern "C" {
             struct ggml_tensor  * a,
             enum ggml_sort_order  order);
 
+    GGML_API struct ggml_tensor * ggml_arange(
+            struct ggml_context * ctx,
+            float                 start,
+            float                 stop,
+            float                 step);
+
     // top k elements per row
     GGML_API struct ggml_tensor * ggml_top_k(
             struct ggml_context * ctx,
             struct ggml_tensor  * a,
             int                   k);
 
-    GGML_API struct ggml_tensor * ggml_flash_attn(
+#define GGML_KQ_MASK_PAD 32
+
+    // q:    [n_embd, n_batch,     n_head,    1]
+    // k:    [n_embd, n_kv,        n_head_kv, 1]
+    // v:    [n_embd, n_kv,        n_head_kv, 1] !! not transposed !!
+    // mask: [n_kv,   n_batch_pad, 1,         1] !! n_batch_pad = GGML_PAD(n_batch, GGML_KQ_MASK_PAD) !!
+    // res:  [n_embd, n_head,      n_batch,   1] !! permuted !!
+    GGML_API struct ggml_tensor * ggml_flash_attn_ext(
             struct ggml_context * ctx,
             struct ggml_tensor  * q,
             struct ggml_tensor  * k,
             struct ggml_tensor  * v,
-            bool                  masked);
+            struct ggml_tensor  * mask,
+            float                 scale,
+            float                 max_bias,
+            float                 logit_softcap);
 
+    GGML_API void ggml_flash_attn_ext_set_prec(
+            struct ggml_tensor * a,
+            enum ggml_prec       prec);
+
+    // TODO: needs to be adapted to ggml_flash_attn_ext
     GGML_API struct ggml_tensor * ggml_flash_attn_back(
            struct ggml_context * ctx,
            struct ggml_tensor  * q,
@@ -1620,13 +1823,19 @@ extern "C" {
            struct ggml_tensor  * d,
            bool                  masked);
 
-    GGML_API struct ggml_tensor * ggml_flash_ff(
+    GGML_API struct ggml_tensor * ggml_ssm_conv(
             struct ggml_context * ctx,
-            struct ggml_tensor  * a,
-            struct ggml_tensor  * b0,
-            struct ggml_tensor  * b1,
-            struct ggml_tensor  * c0,
-            struct ggml_tensor  * c1);
+            struct ggml_tensor  * sx,
+            struct ggml_tensor  * c);
+
+    GGML_API struct ggml_tensor * ggml_ssm_scan(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * s,
+            struct ggml_tensor  * x,
+            struct ggml_tensor  * dt,
+            struct ggml_tensor  * A,
+            struct ggml_tensor  * B,
+            struct ggml_tensor  * C);
 
     // partition into non-overlapping windows with padding if needed
     // example:
@@ -1678,6 +1887,15 @@ extern "C" {
             struct ggml_tensor  * pw,
             struct ggml_tensor  * ph);
 
+    GGML_API struct ggml_tensor * ggml_rwkv_wkv(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * k,
+            struct ggml_tensor  * v,
+            struct ggml_tensor  * r,
+            struct ggml_tensor  * tf,
+            struct ggml_tensor  * td,
+            struct ggml_tensor  * state);
+
     // custom operators
 
     typedef void (*ggml_unary_op_f32_t) (const int, float *, const float *);
@@ -1761,7 +1979,8 @@ extern "C" {
     typedef void (*ggml_custom2_op_t)(struct ggml_tensor * dst , const struct ggml_tensor * a, const struct ggml_tensor * b, int ith, int nth, void * userdata);
     typedef void (*ggml_custom3_op_t)(struct ggml_tensor * dst , const struct ggml_tensor * a, const struct ggml_tensor * b, const struct ggml_tensor * c, int ith, int nth, void * userdata);
 
-    #define GGML_N_TASKS_MAX -1
+#define GGML_N_TASKS_MAX (-1)
+    // n_tasks == GGML_N_TASKS_MAX means to use max number of tasks
 
     GGML_API struct ggml_tensor * ggml_map_custom1(
             struct ggml_context   * ctx,
@@ -1814,48 +2033,87 @@ extern "C" {
     // loss function
 
     GGML_API struct ggml_tensor * ggml_cross_entropy_loss(
-            struct ggml_context         * ctx,
-            struct ggml_tensor          * a,
-            struct ggml_tensor          * b);
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a,  // logits
+            struct ggml_tensor  * b); // labels
 
     GGML_API struct ggml_tensor * ggml_cross_entropy_loss_back(
-            struct ggml_context         * ctx,
-            struct ggml_tensor          * a,
-            struct ggml_tensor          * b,
-            struct ggml_tensor          * c);
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a,  // logits
+            struct ggml_tensor  * b,  // labels
+            struct ggml_tensor  * c); // gradients of cross_entropy_loss result
+
+    // AdamW optimizer step
+    // Paper: https://arxiv.org/pdf/1711.05101v3.pdf
+    // PyTorch: https://pytorch.org/docs/stable/generated/torch.optim.AdamW.html
+    GGML_API struct ggml_tensor * ggml_opt_step_adamw(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a,
+            struct ggml_tensor  * grad,
+            float                 alpha,
+            float                 beta1,
+            float                 beta2,
+            float                 eps,
+            float                 wd); // weight decay
 
     //
     // automatic differentiation
     //
 
-    GGML_API void ggml_set_param(
-            struct ggml_context * ctx,
-            struct ggml_tensor  * tensor);
-
+    GGML_API void ggml_set_param(struct ggml_context * ctx, struct ggml_tensor * tensor);
+    GGML_API void ggml_set_loss(struct ggml_tensor * tensor);
 
     GGML_API void ggml_build_forward_expand (struct ggml_cgraph * cgraph, struct ggml_tensor * tensor);
-    GGML_API void ggml_build_backward_expand(struct ggml_context * ctx, struct ggml_cgraph * gf, struct ggml_cgraph * gb, bool keep);
+    GGML_API void ggml_build_backward_expand(struct ggml_context * ctx, struct ggml_cgraph * gf, struct ggml_cgraph * gb, bool accumulate);
+
+    GGML_API void ggml_build_opt_adamw(
+            struct ggml_context * ctx,
+            struct ggml_cgraph  * gf,
+            struct ggml_cgraph  * gb,
+            float                 alpha,
+            float                 beta1,
+            float                 beta2,
+            float                 eps,
+            float                 wd); // weight decay
 
     // graph allocation in a context
-    GGML_API struct ggml_cgraph * ggml_new_graph         (struct ggml_context * ctx); // size = GGML_DEFAULT_GRAPH_SIZE, grads = false
-    GGML_API struct ggml_cgraph * ggml_new_graph_custom  (struct ggml_context * ctx, size_t size, bool grads);
-    GGML_API struct ggml_cgraph * ggml_graph_dup         (struct ggml_context * ctx, struct ggml_cgraph * cgraph);
-    GGML_API struct ggml_cgraph   ggml_graph_view        (struct ggml_cgraph * cgraph, int i0, int i1);
-    GGML_API void                 ggml_graph_cpy         (struct ggml_cgraph * src, struct ggml_cgraph * dst);
-    GGML_API void                 ggml_graph_reset       (struct ggml_cgraph * cgraph);  // zero grads
-    GGML_API void                 ggml_graph_clear       (struct ggml_cgraph * cgraph);
+    GGML_API struct ggml_cgraph * ggml_new_graph       (struct ggml_context * ctx); // size = GGML_DEFAULT_GRAPH_SIZE, grads = false
+    GGML_API struct ggml_cgraph * ggml_new_graph_custom(struct ggml_context * ctx, size_t size, bool grads);
+    GGML_API struct ggml_cgraph * ggml_graph_dup       (struct ggml_context * ctx, struct ggml_cgraph * cgraph);
+    GGML_API void                 ggml_graph_cpy       (struct ggml_cgraph * src, struct ggml_cgraph * dst);
+    GGML_API void                 ggml_graph_reset     (struct ggml_cgraph * cgraph); // set regular grads + optimizer momenta to 0, set loss grad to 1
+    GGML_API void                 ggml_graph_clear     (struct ggml_cgraph * cgraph);
+
+    GGML_API int                   ggml_graph_size   (struct ggml_cgraph * cgraph);
+    GGML_API struct ggml_tensor *  ggml_graph_node   (struct ggml_cgraph * cgraph, int i); // if i < 0, returns nodes[n_nodes + i]
+    GGML_API struct ggml_tensor ** ggml_graph_nodes  (struct ggml_cgraph * cgraph);
+    GGML_API int                   ggml_graph_n_nodes(struct ggml_cgraph * cgraph);
+
+    GGML_API void   ggml_graph_add_node(struct ggml_cgraph * cgraph, struct ggml_tensor * tensor);
 
     GGML_API size_t ggml_graph_overhead(void);
     GGML_API size_t ggml_graph_overhead_custom(size_t size, bool grads);
 
+    GGML_API struct ggml_threadpool_params ggml_threadpool_params_default(int n_threads);
+    GGML_API void                          ggml_threadpool_params_init   (struct ggml_threadpool_params * p, int n_threads);
+    GGML_API bool                          ggml_threadpool_params_match  (const struct ggml_threadpool_params * p0, const struct ggml_threadpool_params * p1);
+    GGML_API struct ggml_threadpool *      ggml_threadpool_new          (struct ggml_threadpool_params  * params);
+    GGML_API void                          ggml_threadpool_free         (struct ggml_threadpool * threadpool);
+    GGML_API int                           ggml_threadpool_get_n_threads(struct ggml_threadpool * threadpool);
+    GGML_API void                          ggml_threadpool_pause        (struct ggml_threadpool * threadpool);
+    GGML_API void                          ggml_threadpool_resume       (struct ggml_threadpool * threadpool);
+
     // ggml_graph_plan() has to be called before ggml_graph_compute()
     // when plan.work_size > 0, caller must allocate memory for plan.work_data
-    GGML_API struct ggml_cplan ggml_graph_plan   (struct ggml_cgraph * cgraph, int n_threads /*= GGML_DEFAULT_N_THREADS*/);
-    GGML_API int               ggml_graph_compute(struct ggml_cgraph * cgraph, struct ggml_cplan * cplan);
+    GGML_API struct ggml_cplan ggml_graph_plan(
+                  const struct ggml_cgraph * cgraph,
+                                       int   n_threads, /* = GGML_DEFAULT_N_THREADS */
+                    struct ggml_threadpool * threadpool /* = NULL */ );
+    GGML_API enum ggml_status  ggml_graph_compute(struct ggml_cgraph * cgraph, struct ggml_cplan * cplan);
 
     // same as ggml_graph_compute() but the work data is allocated as a part of the context
     // note: the drawback of this API is that you must have ensured that the context has enough memory for the work data
-    GGML_API void ggml_graph_compute_with_ctx(struct ggml_context * ctx, struct ggml_cgraph * cgraph, int n_threads);
+    GGML_API enum ggml_status  ggml_graph_compute_with_ctx(struct ggml_context * ctx, struct ggml_cgraph * cgraph, int n_threads);
 
     GGML_API struct ggml_tensor * ggml_graph_get_tensor(struct ggml_cgraph * cgraph, const char * name);
 
@@ -1884,8 +2142,8 @@ extern "C" {
 
     // optimization methods
     enum ggml_opt_type {
-        GGML_OPT_ADAM,
-        GGML_OPT_LBFGS,
+        GGML_OPT_TYPE_ADAM,
+        GGML_OPT_TYPE_LBFGS,
     };
 
     // linesearch methods
@@ -1899,12 +2157,12 @@ extern "C" {
 
     // optimization return values
     enum ggml_opt_result {
-        GGML_OPT_OK = 0,
-        GGML_OPT_DID_NOT_CONVERGE,
-        GGML_OPT_NO_CONTEXT,
-        GGML_OPT_INVALID_WOLFE,
-        GGML_OPT_FAIL,
-        GGML_OPT_CANCEL,
+        GGML_OPT_RESULT_OK = 0,
+        GGML_OPT_RESULT_DID_NOT_CONVERGE,
+        GGML_OPT_RESULT_NO_CONTEXT,
+        GGML_OPT_RESULT_INVALID_WOLFE,
+        GGML_OPT_RESULT_FAIL,
+        GGML_OPT_RESULT_CANCEL,
 
         GGML_LINESEARCH_FAIL = -128,
         GGML_LINESEARCH_MINIMUM_STEP,
@@ -1916,6 +2174,10 @@ extern "C" {
     typedef void (*ggml_opt_callback)(void * data, int accum_step, float * sched, bool * cancel);
     typedef void (*ggml_log_callback)(enum ggml_log_level level, const char * text, void * user_data);
 
+    // Set callback for all future logging events.
+    // If this is not called, or NULL is supplied, everything is output on stderr.
+    GGML_API void ggml_log_set(ggml_log_callback log_callback, void * user_data);
+
     // optimization parameters
     //
     //   see ggml.c (ggml_opt_default_params) for default values
@@ -2055,23 +2317,39 @@ extern "C" {
             void * callback_data);
 
     //
-    // quantization
+    // tensor flags
     //
+    GGML_API void ggml_set_input(struct ggml_tensor * tensor);
+    GGML_API void ggml_set_output(struct ggml_tensor * tensor);
 
-    // TODO: these would probably get removed in favor of the more general ggml_quantize_chunk
-    GGML_API size_t ggml_quantize_q4_0(const float * src, void * dst, int n, int k, int64_t * hist);
-    GGML_API size_t ggml_quantize_q4_1(const float * src, void * dst, int n, int k, int64_t * hist);
-    GGML_API size_t ggml_quantize_q5_0(const float * src, void * dst, int n, int k, int64_t * hist);
-    GGML_API size_t ggml_quantize_q5_1(const float * src, void * dst, int n, int k, int64_t * hist);
-    GGML_API size_t ggml_quantize_q8_0(const float * src, void * dst, int n, int k, int64_t * hist);
-
-    GGML_API size_t ggml_quantize_q2_K(const float * src, void * dst, int n, int k, int64_t * hist);
-    GGML_API size_t ggml_quantize_q3_K(const float * src, void * dst, int n, int k, int64_t * hist);
-    GGML_API size_t ggml_quantize_q4_K(const float * src, void * dst, int n, int k, int64_t * hist);
-    GGML_API size_t ggml_quantize_q5_K(const float * src, void * dst, int n, int k, int64_t * hist);
-    GGML_API size_t ggml_quantize_q6_K(const float * src, void * dst, int n, int k, int64_t * hist);
+    //
+    // quantization
+    //
 
-    GGML_API size_t ggml_quantize_chunk(enum ggml_type type, const float * src, void * dst, int start, int n, int64_t * hist);
+    // - ggml_quantize_init can be called multiple times with the same type
+    //   it will only initialize the quantization tables for the first call or after ggml_quantize_free
+    //   automatically called by ggml_quantize_chunk for convenience
+    //
+    // - ggml_quantize_free will free any memory allocated by ggml_quantize_init
+    //   call this at the end of the program to avoid memory leaks
+    //
+    // note: these are thread-safe
+    //
+    GGML_API void ggml_quantize_init(enum ggml_type type);
+    GGML_API void ggml_quantize_free(void);
+
+    // some quantization type cannot be used without an importance matrix
+    GGML_API bool ggml_quantize_requires_imatrix(enum ggml_type type);
+
+    // calls ggml_quantize_init internally (i.e. can allocate memory)
+    GGML_API size_t ggml_quantize_chunk(
+            enum ggml_type   type,
+               const float * src,
+                      void * dst,
+                   int64_t   start,
+                   int64_t   nrows,
+                   int64_t   n_per_row,
+               const float * imatrix);
 
     //
     // gguf
@@ -2147,6 +2425,9 @@ extern "C" {
     GGML_API char *         gguf_get_tensor_name  (const struct gguf_context * ctx, int i);
     GGML_API enum ggml_type gguf_get_tensor_type  (const struct gguf_context * ctx, int i);
 
+    // removes key if it exists
+    GGML_API void gguf_remove_key(struct gguf_context * ctx, const char * key);
+
     // overrides existing values or adds a new one
     GGML_API void gguf_set_val_u8  (struct gguf_context * ctx, const char * key, uint8_t  val);
     GGML_API void gguf_set_val_i8  (struct gguf_context * ctx, const char * key, int8_t   val);
@@ -2206,20 +2487,32 @@ extern "C" {
     GGML_API int ggml_cpu_has_avx512     (void);
     GGML_API int ggml_cpu_has_avx512_vbmi(void);
     GGML_API int ggml_cpu_has_avx512_vnni(void);
+    GGML_API int ggml_cpu_has_avx512_bf16(void);
     GGML_API int ggml_cpu_has_fma        (void);
     GGML_API int ggml_cpu_has_neon       (void);
+    GGML_API int ggml_cpu_has_sve        (void);
     GGML_API int ggml_cpu_has_arm_fma    (void);
     GGML_API int ggml_cpu_has_metal      (void);
     GGML_API int ggml_cpu_has_f16c       (void);
     GGML_API int ggml_cpu_has_fp16_va    (void);
     GGML_API int ggml_cpu_has_wasm_simd  (void);
     GGML_API int ggml_cpu_has_blas       (void);
-    GGML_API int ggml_cpu_has_cublas     (void);
-    GGML_API int ggml_cpu_has_clblast    (void);
+    GGML_API int ggml_cpu_has_cuda       (void);
+    GGML_API int ggml_cpu_has_vulkan     (void);
+    GGML_API int ggml_cpu_has_kompute    (void);
     GGML_API int ggml_cpu_has_gpublas    (void);
     GGML_API int ggml_cpu_has_sse3       (void);
     GGML_API int ggml_cpu_has_ssse3      (void);
+    GGML_API int ggml_cpu_has_riscv_v    (void);
+    GGML_API int ggml_cpu_has_sycl       (void);
+    GGML_API int ggml_cpu_has_rpc        (void);
     GGML_API int ggml_cpu_has_vsx        (void);
+    GGML_API int ggml_cpu_has_matmul_int8(void);
+    GGML_API int ggml_cpu_has_cann       (void);
+    GGML_API int ggml_cpu_has_llamafile  (void);
+
+    // get the sve vector length in bytes
+    GGML_API int ggml_cpu_get_sve_cnt(void);
 
     //
     // Internal types and functions exposed for tests and benchmarks
@@ -2231,20 +2524,33 @@ extern "C" {
 #else
 #define GGML_RESTRICT restrict
 #endif
-    typedef void (*ggml_to_float_t)  (const void  * GGML_RESTRICT x, float * GGML_RESTRICT y, int k);
-    typedef void (*ggml_from_float_t)(const float * GGML_RESTRICT x, void  * GGML_RESTRICT y, int k);
-    typedef void (*ggml_vec_dot_t)   (const int n, float * GGML_RESTRICT s, const void * GGML_RESTRICT x, const void * GGML_RESTRICT y);
+    typedef void (*ggml_to_float_t)  (const void  * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
+    typedef void (*ggml_from_float_t)(const float * GGML_RESTRICT x, void  * GGML_RESTRICT y, int64_t k);
+    typedef void (*ggml_from_float_to_mat_t)
+                                     (const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t nr, int64_t k, int64_t bs);
+    typedef void (*ggml_vec_dot_t)  (int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT x, size_t bx,
+                                       const void * GGML_RESTRICT y, size_t by, int nrc);
+    typedef void (*ggml_gemv_t)     (int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT x,
+                                       const void * GGML_RESTRICT y, int nr, int nc);
+    typedef void (*ggml_gemm_t)     (int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT x,
+                                       const void * GGML_RESTRICT y, int nr, int nc);
 
     typedef struct {
-        const char      * type_name;
-        int               blck_size;
-        size_t            type_size;
-        bool              is_quantized;
-        ggml_to_float_t   to_float;
-        ggml_from_float_t from_float;
-        ggml_from_float_t from_float_reference;
-        ggml_vec_dot_t    vec_dot;
-        enum ggml_type    vec_dot_type;
+        const char             * type_name;
+        int64_t                  blck_size;
+        int64_t                  blck_size_interleave; // interleave elements in blocks
+        size_t                   type_size;
+        bool                     is_quantized;
+        ggml_to_float_t          to_float;
+        ggml_from_float_t        from_float;
+        ggml_from_float_t        from_float_ref;
+        ggml_from_float_to_mat_t from_float_to_mat;
+        ggml_vec_dot_t           vec_dot;
+        enum ggml_type           vec_dot_type;
+        int64_t                  nrows; // number of rows to process simultaneously
+        int64_t                  ncols; // number of columns to process simultaneously
+        ggml_gemv_t              gemv;
+        ggml_gemm_t              gemm;
     } ggml_type_traits_t;
 
     GGML_API ggml_type_traits_t ggml_internal_get_type_traits(enum ggml_type type);
diff --git a/src/whisper_cpp/ggml/src/CMakeLists.txt b/src/whisper_cpp/ggml/src/CMakeLists.txt
new file mode 100644
index 00000000..286bec25
--- /dev/null
+++ b/src/whisper_cpp/ggml/src/CMakeLists.txt
@@ -0,0 +1,1371 @@
+include(CheckCXXCompilerFlag)
+
+unset(GGML_CDEF_PUBLIC)
+
+add_compile_definitions(GGML_SCHED_MAX_COPIES=${GGML_SCHED_MAX_COPIES})
+
+# enable libstdc++ assertions for debug builds
+if (CMAKE_SYSTEM_NAME MATCHES "Linux")
+    add_compile_definitions($<$<CONFIG:Debug>:_GLIBCXX_ASSERTIONS>)
+endif()
+
+if (NOT MSVC)
+    if (GGML_SANITIZE_THREAD)
+        add_compile_options(-fsanitize=thread)
+        link_libraries     (-fsanitize=thread)
+    endif()
+
+    if (GGML_SANITIZE_ADDRESS)
+        add_compile_options(-fsanitize=address -fno-omit-frame-pointer)
+        link_libraries     (-fsanitize=address)
+    endif()
+
+    if (GGML_SANITIZE_UNDEFINED)
+        add_compile_options(-fsanitize=undefined)
+        link_libraries     (-fsanitize=undefined)
+    endif()
+endif()
+
+unset(GGML_EXTRA_LIBS_PRIVATE)
+unset(GGML_EXTRA_LIBS_PUBLIC)
+
+if (APPLE AND GGML_ACCELERATE)
+    find_library(ACCELERATE_FRAMEWORK Accelerate)
+    if (ACCELERATE_FRAMEWORK)
+        message(STATUS "Accelerate framework found")
+
+        add_compile_definitions(GGML_USE_ACCELERATE)
+        add_compile_definitions(ACCELERATE_NEW_LAPACK)
+        add_compile_definitions(ACCELERATE_LAPACK_ILP64)
+
+        list(APPEND GGML_EXTRA_LIBS_PRIVATE ${ACCELERATE_FRAMEWORK})
+    else()
+        message(WARNING "Accelerate framework not found")
+    endif()
+endif()
+
+if (GGML_METAL)
+    find_library(FOUNDATION_LIBRARY Foundation REQUIRED)
+    find_library(METAL_FRAMEWORK    Metal      REQUIRED)
+    find_library(METALKIT_FRAMEWORK MetalKit   REQUIRED)
+
+    message(STATUS "Metal framework found")
+    set(GGML_HEADERS_METAL ../include/ggml-metal.h)
+    set(GGML_SOURCES_METAL ggml-metal.m)
+
+    list(APPEND GGML_CDEF_PUBLIC GGML_USE_METAL)
+    if (GGML_METAL_NDEBUG)
+        add_compile_definitions(GGML_METAL_NDEBUG)
+    endif()
+
+    # copy ggml-common.h and ggml-metal.metal to bin directory
+    configure_file(ggml-common.h    ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/ggml-common.h    COPYONLY)
+    configure_file(ggml-metal.metal ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/ggml-metal.metal COPYONLY)
+
+    if (GGML_METAL_EMBED_LIBRARY)
+        enable_language(ASM)
+
+        add_compile_definitions(GGML_METAL_EMBED_LIBRARY)
+
+        set(METALLIB_COMMON "${CMAKE_CURRENT_SOURCE_DIR}/ggml-common.h")
+        set(METALLIB_SOURCE "${CMAKE_CURRENT_SOURCE_DIR}/ggml-metal.metal")
+
+        file(MAKE_DIRECTORY "${CMAKE_BINARY_DIR}/autogenerated")
+
+        # merge ggml-common.h and ggml-metal.metal into a single file
+        set(METALLIB_EMBED_ASM    "${CMAKE_BINARY_DIR}/autogenerated/ggml-metal-embed.s")
+        set(METALLIB_SOURCE_EMBED "${CMAKE_BINARY_DIR}/autogenerated/ggml-metal-embed.metal")
+
+        add_custom_command(
+            OUTPUT ${METALLIB_EMBED_ASM}
+            COMMAND echo "Embedding Metal library"
+            COMMAND sed -e '/\#include \"ggml-common.h\"/r ${METALLIB_COMMON}' -e '/\#include \"ggml-common.h\"/d' < ${METALLIB_SOURCE} > ${METALLIB_SOURCE_EMBED}
+            COMMAND echo ".section __DATA,__ggml_metallib"          >  ${METALLIB_EMBED_ASM}
+            COMMAND echo ".globl _ggml_metallib_start"              >> ${METALLIB_EMBED_ASM}
+            COMMAND echo "_ggml_metallib_start:"                    >> ${METALLIB_EMBED_ASM}
+            COMMAND echo ".incbin \\\"${METALLIB_SOURCE_EMBED}\\\"" >> ${METALLIB_EMBED_ASM}
+            COMMAND echo ".globl _ggml_metallib_end"                >> ${METALLIB_EMBED_ASM}
+            COMMAND echo "_ggml_metallib_end:"                      >> ${METALLIB_EMBED_ASM}
+            DEPENDS ggml-metal.metal ggml-common.h
+            COMMENT "Generate assembly for embedded Metal library"
+        )
+
+        list(APPEND GGML_SOURCES_METAL ${METALLIB_EMBED_ASM})
+    else()
+        if (GGML_METAL_SHADER_DEBUG)
+            # custom command to do the following:
+            #   xcrun -sdk macosx metal    -fno-fast-math -c ggml-metal.metal -o ggml-metal.air
+            #   xcrun -sdk macosx metallib                   ggml-metal.air   -o default.metallib
+            #
+            # note: this is the only way I found to disable fast-math in Metal. it's ugly, but at least it works
+            #       disabling fast math is needed in order to pass tests/test-backend-ops
+            # note: adding -fno-inline fixes the tests when using MTL_SHADER_VALIDATION=1
+            # note: unfortunately, we have to call it default.metallib instead of ggml.metallib
+            #       ref: https://github.com/ggerganov/whisper.cpp/issues/1720
+            set(XC_FLAGS -fno-fast-math -fno-inline -g)
+        else()
+            set(XC_FLAGS -O3)
+        endif()
+
+        # Append macOS metal versioning flags
+        if (GGML_METAL_MACOSX_VERSION_MIN)
+            message(STATUS "Adding  -mmacosx-version-min=${GGML_METAL_MACOSX_VERSION_MIN} flag to metal compilation")
+            list   (APPEND XC_FLAGS -mmacosx-version-min=${GGML_METAL_MACOSX_VERSION_MIN})
+        endif()
+
+        if (GGML_METAL_STD)
+            message(STATUS "Adding  -std=${GGML_METAL_STD} flag to metal compilation")
+            list   (APPEND XC_FLAGS -std=${GGML_METAL_STD})
+        endif()
+
+        add_custom_command(
+            OUTPUT ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/default.metallib
+            COMMAND xcrun -sdk macosx metal    ${XC_FLAGS} -c ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/ggml-metal.metal -o ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/ggml-metal.air
+            COMMAND xcrun -sdk macosx metallib                ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/ggml-metal.air   -o ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/default.metallib
+            COMMAND rm -f ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/ggml-metal.air
+            COMMAND rm -f ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/ggml-common.h
+            COMMAND rm -f ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/ggml-metal.metal
+            DEPENDS ggml-metal.metal ggml-common.h
+            COMMENT "Compiling Metal kernels"
+            )
+
+        add_custom_target(
+            ggml-metal ALL
+            DEPENDS ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/default.metallib
+            )
+    endif() # GGML_METAL_EMBED_LIBRARY
+
+    list(APPEND GGML_EXTRA_LIBS_PRIVATE
+        ${FOUNDATION_LIBRARY}
+        ${METAL_FRAMEWORK}
+        ${METALKIT_FRAMEWORK}
+        )
+endif()
+
+if (GGML_MUSA)
+    set(CMAKE_C_COMPILER clang)
+    set(CMAKE_C_EXTENSIONS OFF)
+    set(CMAKE_CXX_COMPILER clang++)
+    set(CMAKE_CXX_EXTENSIONS OFF)
+
+    set(GGML_CUDA ON)
+
+    list(APPEND GGML_CDEF_PUBLIC GGML_USE_MUSA)
+endif()
+
+if (GGML_OPENMP)
+    find_package(OpenMP)
+    if (OpenMP_FOUND)
+        message(STATUS "OpenMP found")
+
+        add_compile_definitions(GGML_USE_OPENMP)
+
+        list(APPEND GGML_EXTRA_LIBS_PRIVATE OpenMP::OpenMP_C OpenMP::OpenMP_CXX)
+
+        if (GGML_MUSA)
+            list(APPEND GGML_EXTRA_INCLUDES     "/usr/lib/llvm-10/include/openmp")
+            list(APPEND GGML_EXTRA_LIBS_PRIVATE "/usr/lib/llvm-10/lib/libomp.so")
+        endif()
+    else()
+        message(WARNING "OpenMP not found")
+    endif()
+endif()
+
+if (GGML_BLAS)
+    if (GGML_STATIC)
+        set(BLA_STATIC ON)
+    endif()
+    #if (CMAKE_VERSION VERSION_GREATER_EQUAL 3.22)
+    #    set(BLA_SIZEOF_INTEGER 8)
+    #endif()
+
+    set(BLA_VENDOR ${GGML_BLAS_VENDOR})
+    find_package(BLAS)
+
+    if (BLAS_FOUND)
+        message(STATUS "BLAS found, Libraries: ${BLAS_LIBRARIES}")
+
+        if (("${BLAS_INCLUDE_DIRS}" STREQUAL "") AND NOT (${GGML_BLAS_VENDOR} MATCHES "Apple"))
+            # BLAS_INCLUDE_DIRS is missing in FindBLAS.cmake.
+            # see https://gitlab.kitware.com/cmake/cmake/-/issues/20268
+            find_package(PkgConfig REQUIRED)
+            if (${GGML_BLAS_VENDOR} MATCHES "Generic")
+                pkg_check_modules(DepBLAS REQUIRED blas)
+            elseif (${GGML_BLAS_VENDOR} MATCHES "OpenBLAS")
+                # As of openblas v0.3.22, the 64-bit is named openblas64.pc
+                pkg_check_modules(DepBLAS openblas64)
+                if (NOT DepBLAS_FOUND)
+                    pkg_check_modules(DepBLAS REQUIRED openblas)
+                endif()
+            elseif (${GGML_BLAS_VENDOR} MATCHES "FLAME")
+                pkg_check_modules(DepBLAS REQUIRED blis)
+            elseif (${GGML_BLAS_VENDOR} MATCHES "ATLAS")
+                pkg_check_modules(DepBLAS REQUIRED blas-atlas)
+            elseif (${GGML_BLAS_VENDOR} MATCHES "FlexiBLAS")
+                pkg_check_modules(DepBLAS REQUIRED flexiblas_api)
+            elseif (${GGML_BLAS_VENDOR} MATCHES "Intel")
+                # all Intel* libraries share the same include path
+                pkg_check_modules(DepBLAS REQUIRED mkl-sdl)
+            elseif (${GGML_BLAS_VENDOR} MATCHES "NVHPC")
+                # this doesn't provide pkg-config
+                # suggest to assign BLAS_INCLUDE_DIRS on your own
+                if ("${NVHPC_VERSION}" STREQUAL "")
+                    message(WARNING "Better to set NVHPC_VERSION")
+                else()
+                    set(DepBLAS_FOUND ON)
+                    set(DepBLAS_INCLUDE_DIRS "/opt/nvidia/hpc_sdk/${CMAKE_SYSTEM_NAME}_${CMAKE_SYSTEM_PROCESSOR}/${NVHPC_VERSION}/math_libs/include")
+                endif()
+            endif()
+            if (DepBLAS_FOUND)
+                set(BLAS_INCLUDE_DIRS ${DepBLAS_INCLUDE_DIRS})
+            else()
+                message(WARNING "BLAS_INCLUDE_DIRS neither been provided nor been automatically"
+                " detected by pkgconfig, trying to find cblas.h from possible paths...")
+                find_path(BLAS_INCLUDE_DIRS
+                    NAMES cblas.h
+                    HINTS
+                        /usr/include
+                        /usr/local/include
+                        /usr/include/openblas
+                        /opt/homebrew/opt/openblas/include
+                        /usr/local/opt/openblas/include
+                        /usr/include/x86_64-linux-gnu/openblas/include
+                )
+            endif()
+        endif()
+
+        message(STATUS "BLAS found, Includes: ${BLAS_INCLUDE_DIRS}")
+
+        add_compile_options(${BLAS_LINKER_FLAGS})
+
+        list(APPEND GGML_CDEF_PUBLIC GGML_USE_BLAS)
+
+        if (${BLAS_INCLUDE_DIRS} MATCHES "mkl" AND (${GGML_BLAS_VENDOR} MATCHES "Generic" OR ${GGML_BLAS_VENDOR} MATCHES "Intel"))
+            add_compile_definitions(GGML_BLAS_USE_MKL)
+        endif()
+
+        set(GGML_HEADERS_BLAS ../include/ggml-blas.h)
+        set(GGML_SOURCES_BLAS ggml-blas.cpp)
+
+        list(APPEND GGML_EXTRA_LIBS_PRIVATE ${BLAS_LIBRARIES})
+        list(APPEND GGML_EXTRA_INCLUDES     ${BLAS_INCLUDE_DIRS})
+    else()
+        message(WARNING "BLAS not found, please refer to "
+        "https://cmake.org/cmake/help/latest/module/FindBLAS.html#blas-lapack-vendors"
+        " to set correct GGML_BLAS_VENDOR")
+    endif()
+endif()
+
+if (GGML_LLAMAFILE)
+    message(STATUS "Using llamafile")
+
+    add_compile_definitions(GGML_USE_LLAMAFILE)
+
+    set(GGML_HEADERS_LLAMAFILE llamafile/sgemm.h)
+    set(GGML_SOURCES_LLAMAFILE llamafile/sgemm.cpp)
+endif()
+
+if (GGML_CUDA)
+    cmake_minimum_required(VERSION 3.18)  # for CMAKE_CUDA_ARCHITECTURES
+
+    if (GGML_MUSA)
+        list(APPEND CMAKE_MODULE_PATH "/usr/local/musa/cmake/")
+        find_package(MUSAToolkit)
+        set(CUDAToolkit_FOUND ${MUSAToolkit_FOUND})
+    else()
+        find_package(CUDAToolkit)
+    endif()
+
+    if (CUDAToolkit_FOUND)
+        message(STATUS "CUDA found")
+
+        if (NOT DEFINED CMAKE_CUDA_ARCHITECTURES)
+            # 52 == lowest CUDA 12 standard
+            # 60 == FP16 CUDA intrinsics
+            # 61 == integer CUDA intrinsics
+            # 70 == compute capability at which unrolling a loop in mul_mat_q kernels is faster
+            if (GGML_CUDA_F16 OR GGML_CUDA_DMMV_F16)
+                set(CMAKE_CUDA_ARCHITECTURES "60;61;70;75")
+            else()
+                set(CMAKE_CUDA_ARCHITECTURES "52;61;70;75")
+                #set(CMAKE_CUDA_ARCHITECTURES "OFF") # use this to compile much faster, but only F16 models work
+            endif()
+        endif()
+        message(STATUS "Using CUDA architectures: ${CMAKE_CUDA_ARCHITECTURES}")
+
+        if (GGML_MUSA)
+            set(CMAKE_CUDA_COMPILER ${MUSAToolkit_MCC_EXECUTABLE})
+        else()
+            enable_language(CUDA)
+        endif()
+
+        file(GLOB   GGML_HEADERS_CUDA "ggml-cuda/*.cuh")
+        list(APPEND GGML_HEADERS_CUDA "../include/ggml-cuda.h")
+
+        file(GLOB   GGML_SOURCES_CUDA "ggml-cuda/*.cu")
+        list(APPEND GGML_SOURCES_CUDA "ggml-cuda.cu")
+        file(GLOB   SRCS "ggml-cuda/template-instances/fattn-wmma*.cu")
+        list(APPEND GGML_SOURCES_CUDA ${SRCS})
+        file(GLOB   SRCS "ggml-cuda/template-instances/mmq*.cu")
+        list(APPEND GGML_SOURCES_CUDA ${SRCS})
+
+        if (GGML_CUDA_FA_ALL_QUANTS)
+            file(GLOB   SRCS "ggml-cuda/template-instances/fattn-vec*.cu")
+            list(APPEND GGML_SOURCES_CUDA ${SRCS})
+            add_compile_definitions(GGML_CUDA_FA_ALL_QUANTS)
+        else()
+            file(GLOB   SRCS "ggml-cuda/template-instances/fattn-vec*q4_0-q4_0.cu")
+            list(APPEND GGML_SOURCES_CUDA ${SRCS})
+            file(GLOB   SRCS "ggml-cuda/template-instances/fattn-vec*q8_0-q8_0.cu")
+            list(APPEND GGML_SOURCES_CUDA ${SRCS})
+            file(GLOB   SRCS "ggml-cuda/template-instances/fattn-vec*f16-f16.cu")
+            list(APPEND GGML_SOURCES_CUDA ${SRCS})
+        endif()
+
+        list(APPEND GGML_CDEF_PUBLIC GGML_USE_CUDA)
+
+        add_compile_definitions(GGML_CUDA_DMMV_X=${GGML_CUDA_DMMV_X})
+        add_compile_definitions(GGML_CUDA_MMV_Y=${GGML_CUDA_MMV_Y})
+        add_compile_definitions(K_QUANTS_PER_ITERATION=${GGML_CUDA_KQUANTS_ITER})
+        add_compile_definitions(GGML_CUDA_PEER_MAX_BATCH_SIZE=${GGML_CUDA_PEER_MAX_BATCH_SIZE})
+
+        if (GGML_CUDA_GRAPHS)
+            add_compile_definitions(GGML_CUDA_USE_GRAPHS)
+        endif()
+
+        if (GGML_CUDA_FORCE_DMMV)
+            add_compile_definitions(GGML_CUDA_FORCE_DMMV)
+        endif()
+
+        if (GGML_CUDA_FORCE_MMQ)
+            add_compile_definitions(GGML_CUDA_FORCE_MMQ)
+        endif()
+
+        if (GGML_CUDA_FORCE_CUBLAS)
+            add_compile_definitions(GGML_CUDA_FORCE_CUBLAS)
+        endif()
+
+        if (GGML_CUDA_NO_VMM)
+            add_compile_definitions(GGML_CUDA_NO_VMM)
+        endif()
+
+        if (DEFINED GGML_CUDA_DMMV_Y)
+            add_compile_definitions(GGML_CUDA_MMV_Y=${GGML_CUDA_DMMV_Y}) # for backwards compatibility
+        endif()
+
+        if (GGML_CUDA_F16 OR GGML_CUDA_DMMV_F16)
+            add_compile_definitions(GGML_CUDA_F16)
+        endif()
+
+        if (GGML_CUDA_NO_PEER_COPY)
+            add_compile_definitions(GGML_CUDA_NO_PEER_COPY)
+        endif()
+
+        if (GGML_MUSA)
+            set_source_files_properties(${GGML_SOURCES_CUDA} PROPERTIES LANGUAGE CXX)
+            foreach(SOURCE ${GGML_SOURCES_CUDA})
+                set_property(SOURCE ${SOURCE} PROPERTY COMPILE_FLAGS "-x musa -mtgpu --cuda-gpu-arch=mp_21 --cuda-gpu-arch=mp_22")
+            endforeach()
+        endif()
+
+        if (GGML_STATIC)
+            if (WIN32)
+                # As of 12.3.1 CUDA Toolkit for Windows does not offer a static cublas library
+                list(APPEND GGML_EXTRA_LIBS_PRIVATE CUDA::cudart_static CUDA::cublas CUDA::cublasLt)
+            else ()
+                if (GGML_MUSA)
+                    list(APPEND GGML_EXTRA_LIBS_PRIVATE MUSA::musart_static MUSA::mublas_static)
+                else()
+                    list(APPEND GGML_EXTRA_LIBS_PRIVATE CUDA::cudart_static CUDA::cublas_static CUDA::cublasLt_static)
+                endif()
+            endif()
+        else()
+            if (GGML_MUSA)
+                list(APPEND GGML_EXTRA_LIBS_PRIVATE MUSA::musart MUSA::mublas)
+            else()
+                list(APPEND GGML_EXTRA_LIBS_PRIVATE CUDA::cudart CUDA::cublas CUDA::cublasLt)
+            endif()
+        endif()
+
+        if (GGML_CUDA_NO_VMM)
+            # No VMM requested, no need to link directly with the cuda driver lib (libcuda.so)
+        else()
+            if (GGML_MUSA)
+                list(APPEND GGML_EXTRA_LIBS_PRIVATE MUSA::musa_driver) # required by muDeviceGetAttribute(), muMemGetAllocationGranularity(...), ...
+            else()
+                list(APPEND GGML_EXTRA_LIBS_PRIVATE CUDA::cuda_driver) # required by cuDeviceGetAttribute(), cuMemGetAllocationGranularity(...), ...
+            endif()
+        endif()
+    else()
+        message(WARNING "CUDA not found")
+    endif()
+endif()
+
+if (GGML_HIPBLAS)
+    if (NOT EXISTS $ENV{ROCM_PATH})
+        if (NOT EXISTS /opt/rocm)
+            set(ROCM_PATH /usr)
+        else()
+            set(ROCM_PATH /opt/rocm)
+        endif()
+    else()
+        set(ROCM_PATH $ENV{ROCM_PATH})
+    endif()
+
+    list(APPEND CMAKE_PREFIX_PATH  ${ROCM_PATH})
+    list(APPEND CMAKE_PREFIX_PATH "${ROCM_PATH}/lib64/cmake")
+
+    # CMake on Windows doesn't support the HIP language yet
+    if (WIN32)
+        set(CXX_IS_HIPCC TRUE)
+    else()
+        string(REGEX MATCH "hipcc(\.bat)?$" CXX_IS_HIPCC "${CMAKE_CXX_COMPILER}")
+    endif()
+
+    if (CXX_IS_HIPCC)
+        if (LINUX)
+            if (NOT ${CMAKE_CXX_COMPILER_ID} MATCHES "Clang")
+                message(WARNING "Only LLVM is supported for HIP, hint: CXX=/opt/rocm/llvm/bin/clang++")
+            endif()
+
+            message(WARNING "Setting hipcc as the C++ compiler is legacy behavior."
+                    " Prefer setting the HIP compiler directly. See README for details.")
+        endif()
+    else()
+        # Forward AMDGPU_TARGETS to CMAKE_HIP_ARCHITECTURES.
+        if (AMDGPU_TARGETS AND NOT CMAKE_HIP_ARCHITECTURES)
+            set(CMAKE_HIP_ARCHITECTURES ${AMDGPU_TARGETS})
+        endif()
+        cmake_minimum_required(VERSION 3.21)
+        enable_language(HIP)
+    endif()
+
+    find_package(hip     REQUIRED)
+    find_package(hipblas REQUIRED)
+    find_package(rocblas REQUIRED)
+
+    message(STATUS "HIP and hipBLAS found")
+
+    file(GLOB   GGML_HEADERS_ROCM "ggml-cuda/*.cuh")
+    list(APPEND GGML_HEADERS_ROCM "../include/ggml-cuda.h")
+
+    file(GLOB   GGML_SOURCES_ROCM "ggml-cuda/*.cu")
+    list(APPEND GGML_SOURCES_ROCM "ggml-cuda.cu")
+    file(GLOB   SRCS "ggml-cuda/template-instances/fattn-wmma*.cu")
+    list(APPEND GGML_SOURCES_ROCM ${SRCS})
+    file(GLOB   SRCS "ggml-cuda/template-instances/mmq*.cu")
+    list(APPEND GGML_SOURCES_ROCM ${SRCS})
+
+    if (GGML_CUDA_FA_ALL_QUANTS)
+        file(GLOB   SRCS "ggml-cuda/template-instances/fattn-vec*.cu")
+        list(APPEND GGML_SOURCES_ROCM ${SRCS})
+        add_compile_definitions(GGML_CUDA_FA_ALL_QUANTS)
+    else()
+        file(GLOB   SRCS "ggml-cuda/template-instances/fattn-vec*q4_0-q4_0.cu")
+        list(APPEND GGML_SOURCES_ROCM ${SRCS})
+        file(GLOB   SRCS "ggml-cuda/template-instances/fattn-vec*q8_0-q8_0.cu")
+        list(APPEND GGML_SOURCES_ROCM ${SRCS})
+        file(GLOB   SRCS "ggml-cuda/template-instances/fattn-vec*f16-f16.cu")
+        list(APPEND GGML_SOURCES_ROCM ${SRCS})
+    endif()
+
+    list(APPEND GGML_CDEF_PUBLIC GGML_USE_CUDA)
+
+    add_compile_definitions(GGML_USE_HIPBLAS)
+    add_compile_definitions(GGML_CUDA_DMMV_X=${GGML_CUDA_DMMV_X})
+    add_compile_definitions(GGML_CUDA_MMV_Y=${GGML_CUDA_MMV_Y})
+    add_compile_definitions(K_QUANTS_PER_ITERATION=${GGML_CUDA_KQUANTS_ITER})
+
+    if (GGML_HIP_UMA)
+        add_compile_definitions(GGML_HIP_UMA)
+    endif()
+
+    if (GGML_CUDA_FORCE_DMMV)
+        add_compile_definitions(GGML_CUDA_FORCE_DMMV)
+    endif()
+
+    if (GGML_CUDA_FORCE_MMQ)
+        add_compile_definitions(GGML_CUDA_FORCE_MMQ)
+    endif()
+
+    if (GGML_CUDA_FORCE_CUBLAS)
+        add_compile_definitions(GGML_CUDA_FORCE_CUBLAS)
+    endif()
+
+    if (GGML_CUDA_NO_PEER_COPY)
+        add_compile_definitions(GGML_CUDA_NO_PEER_COPY)
+    endif()
+
+    if (CXX_IS_HIPCC)
+        set_source_files_properties(${GGML_SOURCES_ROCM} PROPERTIES LANGUAGE CXX)
+        list(APPEND GGML_EXTRA_LIBS_PRIVATE hip::device)
+    else()
+        set_source_files_properties(${GGML_SOURCES_ROCM} PROPERTIES LANGUAGE HIP)
+    endif()
+
+    if (GGML_STATIC)
+        message(FATAL_ERROR "Static linking not supported for HIP/ROCm")
+    endif()
+
+    list(APPEND GGML_EXTRA_LIBS_PUBLIC hip::host roc::rocblas roc::hipblas)
+endif()
+
+if (GGML_SYCL)
+    if (NOT GGML_SYCL_TARGET MATCHES "^(INTEL|NVIDIA|AMD)$")
+        message(FATAL_ERROR "Invalid backend chosen, supported options are INTEL, NVIDIA, or AMD")
+    endif()
+
+    check_cxx_compiler_flag("-fsycl" SUPPORTS_SYCL)
+
+    if (DEFINED ENV{ONEAPI_ROOT})
+        message(STATUS "Using oneAPI Release SYCL compiler (icpx).")
+    elseif(SUPPORTS_SYCL)
+        message(WARNING "Using open-source SYCL compiler (clang++). Didn't detect ENV {ONEAPI_ROOT}.
+         If you expected the oneAPI Release compiler, please install oneAPI & source it, like:
+         source /opt/intel/oneapi/setvars.sh")
+    else()
+        message(FATAL_ERROR, "C++ compiler lacks SYCL support.")
+    endif()
+    message(STATUS "SYCL found")
+    #todo: AOT
+
+    list(APPEND GGML_CDEF_PUBLIC GGML_USE_SYCL)
+
+    if (GGML_SYCL_F16)
+        if (GGML_SYCL_TARGET STREQUAL "AMD")
+            message(WARNING "AMD target does not entirely support FP16 in the SYCL backend.")
+        endif()
+        add_compile_definitions(GGML_SYCL_F16)
+    endif()
+
+    if (GGML_CUDA_FORCE_MMQ)
+        add_compile_definitions(GGML_SYCL_FORCE_MMQ)
+    endif()
+
+    set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -Wno-narrowing -fsycl")
+
+    if (GGML_SYCL_TARGET STREQUAL "NVIDIA")
+        add_compile_definitions(GGML_SYCL_WARP_SIZE=32)
+    elseif (GGML_SYCL_TARGET STREQUAL "AMD")
+        # INFO: Allowed Sub_group_sizes are not consistent through all
+        # hip targets. For example, 64 is used for certain models, but the backend
+        # does not support it.
+        # Target archs tested working: gfx1030, gfx1031, (Only tested sub_group_size = 32)
+        add_compile_definitions(GGML_SYCL_WARP_SIZE=32)
+    else()
+        add_compile_definitions(GGML_SYCL_WARP_SIZE=16)
+    endif()
+
+    file(GLOB   GGML_HEADERS_SYCL "ggml-sycl/*.hpp")
+    list(APPEND GGML_HEADERS_SYCL "../include/ggml-sycl.h")
+
+    file(GLOB   GGML_SOURCES_SYCL "ggml-sycl/*.cpp")
+    list(APPEND GGML_SOURCES_SYCL "ggml-sycl.cpp")
+
+    find_package(DNNL)
+    message("-- DNNL found:" ${DNNL_FOUND})
+
+    if (GGML_SYCL_TARGET STREQUAL "INTEL")
+        add_compile_definitions(GGML_SYCL_DNNL=${DNNL_FOUND})
+    else()
+        add_compile_definitions(GGML_SYCL_DNNL=0)
+    endif()
+
+    if (${DNNL_FOUND} AND GGML_SYCL_TARGET STREQUAL "INTEL")
+        list(APPEND GGML_EXTRA_LIBS_PRIVATE DNNL::dnnl)
+    endif()
+
+    if (WIN32)
+        find_package(IntelSYCL REQUIRED)
+        find_package(MKL REQUIRED)
+        list(APPEND GGML_EXTRA_LIBS_PRIVATE IntelSYCL::SYCL_CXX MKL::MKL MKL::MKL_SYCL)
+    else()
+        if (GGML_SYCL_TARGET STREQUAL "INTEL")
+            list(APPEND GGML_EXTRA_LIBS_PRIVATE sycl OpenCL mkl_core pthread m dl mkl_sycl_blas mkl_intel_ilp64 mkl_tbb_thread)
+        elseif (GGML_SYCL_TARGET STREQUAL "NVIDIA")
+            set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -fsycl-targets=nvptx64-nvidia-cuda")
+            list(APPEND GGML_EXTRA_LIBS_PRIVATE sycl pthread m dl onemkl)
+        elseif (GGML_SYCL_TARGET STREQUAL "AMD")
+            if (GGML_SYCL_HIP_TARGET STREQUAL "")
+                message(ERROR "Can't enable SYCL hip backend, GGML_SYCL_HIP_TARGET has not been set.")
+            endif()
+            set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -fsycl-targets=amdgcn-amd-amdhsa -Xsycl-target-backend --offload-arch=${GGML_SYCL_HIP_TARGET}")
+            list(APPEND GGML_EXTRA_LIBS_PRIVATE sycl pthread m dl onemkl)
+        endif()
+    endif()
+endif()
+
+if (GGML_RPC)
+    message(STATUS "RPC found")
+
+    list(APPEND GGML_CDEF_PUBLIC GGML_USE_RPC)
+
+    if (WIN32)
+        list(APPEND GGML_EXTRA_LIBS_PRIVATE ws2_32)
+    endif()
+
+    set(GGML_HEADERS_RPC ../include/ggml-rpc.h)
+    set(GGML_SOURCES_RPC ggml-rpc.cpp)
+endif()
+
+if (GGML_VULKAN)
+    find_package(Vulkan COMPONENTS glslc REQUIRED)
+
+    if (Vulkan_FOUND)
+        message(STATUS "Vulkan found")
+
+        list(APPEND GGML_CDEF_PUBLIC GGML_USE_VULKAN)
+
+        # Workaround to the "can't dereference invalidated vector iterator" bug in clang-cl debug build
+        # Posssibly relevant: https://stackoverflow.com/questions/74748276/visual-studio-no-displays-the-correct-length-of-stdvector
+        if (MSVC AND CMAKE_CXX_COMPILER_ID STREQUAL "Clang")
+            add_compile_definitions(_ITERATOR_DEBUG_LEVEL=0)
+        endif()
+
+        if (GGML_VULKAN_CHECK_RESULTS)
+            add_compile_definitions(GGML_VULKAN_CHECK_RESULTS)
+        endif()
+
+        if (GGML_VULKAN_DEBUG)
+            add_compile_definitions(GGML_VULKAN_DEBUG)
+        endif()
+
+        if (GGML_VULKAN_MEMORY_DEBUG)
+            add_compile_definitions(GGML_VULKAN_MEMORY_DEBUG)
+        endif()
+
+        if (GGML_VULKAN_SHADER_DEBUG_INFO)
+            add_compile_definitions(GGML_VULKAN_SHADER_DEBUG_INFO)
+        endif()
+
+        if (GGML_VULKAN_PERF)
+            add_compile_definitions(GGML_VULKAN_PERF)
+        endif()
+
+        if (GGML_VULKAN_VALIDATE)
+            add_compile_definitions(GGML_VULKAN_VALIDATE)
+        endif()
+
+        if (GGML_VULKAN_RUN_TESTS)
+            add_compile_definitions(GGML_VULKAN_RUN_TESTS)
+        endif()
+
+        add_subdirectory(vulkan-shaders)
+
+        set (_ggml_vk_genshaders_cmd vulkan-shaders-gen)
+        set (_ggml_vk_header     ${CMAKE_CURRENT_BINARY_DIR}/ggml-vulkan-shaders.hpp)
+        set (_ggml_vk_source     ${CMAKE_CURRENT_BINARY_DIR}/ggml-vulkan-shaders.cpp)
+        set (_ggml_vk_input_dir  ${CMAKE_CURRENT_SOURCE_DIR}/vulkan-shaders)
+        set (_ggml_vk_output_dir ${CMAKE_CURRENT_BINARY_DIR}/vulkan-shaders.spv)
+
+        file(GLOB _ggml_vk_shader_deps "${_ggml_vk_input_dir}/*.comp")
+
+        add_custom_command(
+            OUTPUT ${_ggml_vk_header}
+                   ${_ggml_vk_source}
+
+            COMMAND ${_ggml_vk_genshaders_cmd}
+                --glslc      ${Vulkan_GLSLC_EXECUTABLE}
+                --input-dir  ${_ggml_vk_input_dir}
+                --output-dir ${_ggml_vk_output_dir}
+                --target-hpp ${_ggml_vk_header}
+                --target-cpp ${_ggml_vk_source}
+                --no-clean
+
+            DEPENDS ${_ggml_vk_shader_deps}
+            COMMENT "Generate vulkan shaders"
+        )
+
+        set(GGML_HEADERS_VULKAN ${CMAKE_CURRENT_SOURCE_DIR}/../include/ggml-vulkan.h ${_ggml_vk_header})
+        set(GGML_SOURCES_VULKAN ggml-vulkan.cpp ${_ggml_vk_source})
+
+        list(APPEND GGML_EXTRA_LIBS_PRIVATE Vulkan::Vulkan)
+        list(APPEND GGML_EXTRA_INCLUDES     ${CMAKE_CURRENT_BINARY_DIR})
+    else()
+        message(WARNING "Vulkan not found")
+    endif()
+endif()
+
+if (GGML_KOMPUTE)
+    add_compile_definitions(VULKAN_HPP_DISPATCH_LOADER_DYNAMIC=1)
+
+    find_package(Vulkan COMPONENTS glslc REQUIRED)
+    find_program(glslc_executable NAMES glslc HINTS Vulkan::glslc)
+
+    if (NOT glslc_executable)
+        message(FATAL_ERROR "glslc not found")
+    endif()
+
+    function(compile_shader)
+        set(options)
+        set(oneValueArgs)
+        set(multiValueArgs SOURCES)
+        cmake_parse_arguments(compile_shader "${options}" "${oneValueArgs}" "${multiValueArgs}" ${ARGN})
+        foreach(source ${compile_shader_SOURCES})
+            get_filename_component(filename ${source} NAME)
+            set(spv_file ${filename}.spv)
+            add_custom_command(
+                OUTPUT ${spv_file}
+                DEPENDS ${CMAKE_CURRENT_SOURCE_DIR}/${source}
+                ${CMAKE_CURRENT_SOURCE_DIR}/kompute-shaders/common.comp
+                ${CMAKE_CURRENT_SOURCE_DIR}/kompute-shaders/op_getrows.comp
+                ${CMAKE_CURRENT_SOURCE_DIR}/kompute-shaders/op_mul_mv_q_n_pre.comp
+                ${CMAKE_CURRENT_SOURCE_DIR}/kompute-shaders/op_mul_mv_q_n.comp
+                COMMAND ${glslc_executable} --target-env=vulkan1.2 -o ${spv_file} ${CMAKE_CURRENT_SOURCE_DIR}/${source}
+                COMMENT "Compiling ${source} to ${spv_file}"
+                )
+
+            get_filename_component(RAW_FILE_NAME ${spv_file} NAME)
+            set(FILE_NAME "shader${RAW_FILE_NAME}")
+            string(REPLACE ".comp.spv" ".h" HEADER_FILE ${FILE_NAME})
+            string(TOUPPER ${HEADER_FILE} HEADER_FILE_DEFINE)
+            string(REPLACE "." "_" HEADER_FILE_DEFINE "${HEADER_FILE_DEFINE}")
+            set(OUTPUT_HEADER_FILE "${HEADER_FILE}")
+            message(STATUS "${HEADER_FILE} generating ${HEADER_FILE_DEFINE}")
+            if(CMAKE_GENERATOR MATCHES "Visual Studio")
+                add_custom_command(
+                    OUTPUT ${OUTPUT_HEADER_FILE}
+                    COMMAND ${CMAKE_COMMAND} -E echo "/*THIS FILE HAS BEEN AUTOMATICALLY GENERATED - DO NOT EDIT*/" > ${OUTPUT_HEADER_FILE}
+                    COMMAND ${CMAKE_COMMAND} -E echo \"\#ifndef ${HEADER_FILE_DEFINE}\" >> ${OUTPUT_HEADER_FILE}
+                    COMMAND ${CMAKE_COMMAND} -E echo \"\#define ${HEADER_FILE_DEFINE}\" >> ${OUTPUT_HEADER_FILE}
+                    COMMAND ${CMAKE_COMMAND} -E echo "namespace kp {" >> ${OUTPUT_HEADER_FILE}
+                    COMMAND ${CMAKE_COMMAND} -E echo "namespace shader_data {" >> ${OUTPUT_HEADER_FILE}
+                    COMMAND ${CMAKE_BINARY_DIR}/bin/$<CONFIG>/xxd -i ${RAW_FILE_NAME} >> ${OUTPUT_HEADER_FILE}
+                    COMMAND ${CMAKE_COMMAND} -E echo "}}" >> ${OUTPUT_HEADER_FILE}
+                    COMMAND ${CMAKE_COMMAND} -E echo \"\#endif // define ${HEADER_FILE_DEFINE}\" >> ${OUTPUT_HEADER_FILE}
+                    DEPENDS ${spv_file} xxd
+                    COMMENT "Converting to hpp: ${FILE_NAME} ${CMAKE_BINARY_DIR}/bin/$<CONFIG>/xxd"
+                    )
+            else()
+                add_custom_command(
+                    OUTPUT ${OUTPUT_HEADER_FILE}
+                    COMMAND ${CMAKE_COMMAND} -E echo "/*THIS FILE HAS BEEN AUTOMATICALLY GENERATED - DO NOT EDIT*/" > ${OUTPUT_HEADER_FILE}
+                    COMMAND ${CMAKE_COMMAND} -E echo \"\#ifndef ${HEADER_FILE_DEFINE}\" >> ${OUTPUT_HEADER_FILE}
+                    COMMAND ${CMAKE_COMMAND} -E echo \"\#define ${HEADER_FILE_DEFINE}\" >> ${OUTPUT_HEADER_FILE}
+                    COMMAND ${CMAKE_COMMAND} -E echo "namespace kp {" >> ${OUTPUT_HEADER_FILE}
+                    COMMAND ${CMAKE_COMMAND} -E echo "namespace shader_data {" >> ${OUTPUT_HEADER_FILE}
+                    COMMAND ${CMAKE_BINARY_DIR}/bin/xxd -i ${RAW_FILE_NAME} >> ${OUTPUT_HEADER_FILE}
+                    COMMAND ${CMAKE_COMMAND} -E echo "}}" >> ${OUTPUT_HEADER_FILE}
+                    COMMAND ${CMAKE_COMMAND} -E echo \"\#endif // define ${HEADER_FILE_DEFINE}\" >> ${OUTPUT_HEADER_FILE}
+                    DEPENDS ${spv_file} xxd
+                    COMMENT "Converting to hpp: ${FILE_NAME} ${CMAKE_BINARY_DIR}/bin/xxd"
+                    )
+            endif()
+        endforeach()
+    endfunction()
+
+    if (EXISTS "${CMAKE_CURRENT_SOURCE_DIR}/kompute/CMakeLists.txt")
+        message(STATUS "Kompute found")
+        set(KOMPUTE_OPT_LOG_LEVEL Error CACHE STRING "Kompute log level")
+        add_subdirectory(kompute)
+
+        # Compile our shaders
+        compile_shader(SOURCES
+            kompute-shaders/op_scale.comp
+            kompute-shaders/op_scale_8.comp
+            kompute-shaders/op_add.comp
+            kompute-shaders/op_addrow.comp
+            kompute-shaders/op_mul.comp
+            kompute-shaders/op_silu.comp
+            kompute-shaders/op_relu.comp
+            kompute-shaders/op_gelu.comp
+            kompute-shaders/op_softmax.comp
+            kompute-shaders/op_norm.comp
+            kompute-shaders/op_rmsnorm.comp
+            kompute-shaders/op_diagmask.comp
+            kompute-shaders/op_mul_mat_mat_f32.comp
+            kompute-shaders/op_mul_mat_f16.comp
+            kompute-shaders/op_mul_mat_q8_0.comp
+            kompute-shaders/op_mul_mat_q4_0.comp
+            kompute-shaders/op_mul_mat_q4_1.comp
+            kompute-shaders/op_mul_mat_q6_k.comp
+            kompute-shaders/op_getrows_f32.comp
+            kompute-shaders/op_getrows_f16.comp
+            kompute-shaders/op_getrows_q4_0.comp
+            kompute-shaders/op_getrows_q4_1.comp
+            kompute-shaders/op_getrows_q6_k.comp
+            kompute-shaders/op_rope_f16.comp
+            kompute-shaders/op_rope_f32.comp
+            kompute-shaders/op_cpy_f16_f16.comp
+            kompute-shaders/op_cpy_f16_f32.comp
+            kompute-shaders/op_cpy_f32_f16.comp
+            kompute-shaders/op_cpy_f32_f32.comp
+        )
+
+        # Create a custom target for our generated shaders
+        add_custom_target(generated_shaders DEPENDS
+            shaderop_scale.h
+            shaderop_scale_8.h
+            shaderop_add.h
+            shaderop_addrow.h
+            shaderop_mul.h
+            shaderop_silu.h
+            shaderop_relu.h
+            shaderop_gelu.h
+            shaderop_softmax.h
+            shaderop_norm.h
+            shaderop_rmsnorm.h
+            shaderop_diagmask.h
+            shaderop_mul_mat_mat_f32.h
+            shaderop_mul_mat_f16.h
+            shaderop_mul_mat_q8_0.h
+            shaderop_mul_mat_q4_0.h
+            shaderop_mul_mat_q4_1.h
+            shaderop_mul_mat_q6_k.h
+            shaderop_getrows_f32.h
+            shaderop_getrows_f16.h
+            shaderop_getrows_q4_0.h
+            shaderop_getrows_q4_1.h
+            shaderop_getrows_q6_k.h
+            shaderop_rope_f16.h
+            shaderop_rope_f32.h
+            shaderop_cpy_f16_f16.h
+            shaderop_cpy_f16_f32.h
+            shaderop_cpy_f32_f16.h
+            shaderop_cpy_f32_f32.h
+        )
+
+        # Create a custom command that depends on the generated_shaders
+        add_custom_command(
+            OUTPUT ${CMAKE_CURRENT_BINARY_DIR}/ggml-kompute.stamp
+            COMMAND ${CMAKE_COMMAND} -E touch ${CMAKE_CURRENT_BINARY_DIR}/ggml-kompute.stamp
+            DEPENDS generated_shaders
+            COMMENT "Ensuring shaders are generated before compiling ggml-kompute.cpp"
+        )
+
+        # Add the stamp to the main sources to ensure dependency tracking
+        set(GGML_SOURCES_KOMPUTE ggml-kompute.cpp           ${CMAKE_CURRENT_BINARY_DIR}/ggml-kompute.stamp)
+        set(GGML_HEADERS_KOMPUTE ../include/ggml-kompute.h  ${CMAKE_CURRENT_BINARY_DIR}/ggml-kompute.stamp)
+
+        list(APPEND GGML_CDEF_PUBLIC GGML_USE_KOMPUTE)
+
+        list(APPEND GGML_EXTRA_LIBS_PRIVATE kompute)
+        list(APPEND GGML_EXTRA_INCLUDES     ${CMAKE_CURRENT_BINARY_DIR})
+    else()
+        message(WARNING "Kompute not found")
+    endif()
+endif()
+
+if (GGML_CPU_HBM)
+    find_library(memkind memkind REQUIRED)
+
+    message(STATUS "Using memkind for CPU HBM")
+
+    add_compile_definitions(GGML_USE_CPU_HBM)
+
+    target_link_libraries(ggml PUBLIC memkind)
+endif()
+
+if (GGML_CANN)
+    if ("cann${CANN_INSTALL_DIR}" STREQUAL "cann" AND DEFINED ENV{ASCEND_TOOLKIT_HOME})
+        set(CANN_INSTALL_DIR $ENV{ASCEND_TOOLKIT_HOME})
+        message(STATUS "CANN: updated CANN_INSTALL_DIR from ASCEND_TOOLKIT_HOME=$ENV{ASCEND_TOOLKIT_HOME}")
+    endif()
+
+    if (CANN_INSTALL_DIR)
+        # Only Support Linux.
+        if (GGML_CANN)
+            if (NOT UNIX)
+                set(GGML_CANN OFF)
+                message(WARNING "CANN: CANN toolkit supports unix but not ${CMAKE_SYSTEM_NAME}. Turning off GGML_CANN")
+            endif()
+        endif()
+
+        # Supported platforms: x86-64, arm64
+        if (GGML_CANN)
+            if (CMAKE_SYSTEM_PROCESSOR STREQUAL "aarch64")
+            elseif (CMAKE_SYSTEM_PROCESSOR STREQUAL "x86_64" OR CMAKE_SYSTEM_PROCESSOR STREQUAL "amd64")
+            else()
+                set(GGML_CANN OFF)
+                message(WARNING "CANN: CANN toolkit supports x86-64 and arm64 but not ${CMAKE_SYSTEM_PROCESSOR}. Turning off GGML_CANN")
+            endif()
+        endif()
+
+        # Set header and libs
+        if(GGML_CANN)
+            set(CANN_INCLUDE_DIRS
+                ${CANN_INSTALL_DIR}/include
+                ${CANN_INSTALL_DIR}/include/aclnn
+                ${CANN_INSTALL_DIR}/acllib/include
+            )
+
+            add_subdirectory(ggml-cann/kernels)
+            list(APPEND CANN_LIBRARIES
+                ascendcl
+                nnopbase
+                opapi
+                acl_op_compiler
+                ascendc_kernels
+            )
+
+            set(GGML_HEADERS_CANN "../include/ggml-cann.h")
+            file(GLOB GGML_SOURCES_CANN "ggml-cann/*.cpp")
+            list(APPEND GGML_SOURCES_CANN "ggml-cann.cpp")
+
+            message(STATUS "CANN: CANN_INCLUDE_DIRS =  ${CANN_INCLUDE_DIRS}")
+            message(STATUS "CANN: CANN_LIBRARIES =  ${CANN_LIBRARIES}")
+
+            list(APPEND GGML_EXTRA_LIBS_PRIVATE ${CANN_LIBRARIES} )
+            list(APPEND GGML_EXTRA_INCLUDES     ${CANN_INCLUDE_DIRS})
+            list(APPEND GGML_EXTRA_LIBDIRS      ${CANN_INSTALL_DIR}/lib64)
+
+            list(APPEND GGML_CDEF_PUBLIC GGML_USE_CANN)
+        endif()
+    else()
+        set(GGML_CANN OFF)
+        message(WARNING "CANN: Can't find CANN_INSTALL_DIR, do you forget to source set_var.sh. Turning off GGML_CANN")
+    endif()
+
+    if(NOT GGML_CANN)
+        message(WARNING "CANN: GGML_CANN is turned OFF, see above for details.")
+    endif()
+endif()
+
+function(get_flags CCID CCVER)
+    set(C_FLAGS "")
+    set(CXX_FLAGS "")
+
+    if (CCID MATCHES "Clang")
+        set(C_FLAGS   -Wunreachable-code-break -Wunreachable-code-return)
+        set(CXX_FLAGS -Wunreachable-code-break -Wunreachable-code-return -Wmissing-prototypes -Wextra-semi)
+
+        if (
+            (CCID STREQUAL "Clang"      AND CCVER VERSION_GREATER_EQUAL 3.8.0) OR
+            (CCID STREQUAL "AppleClang" AND CCVER VERSION_GREATER_EQUAL 7.3.0)
+        )
+            list(APPEND C_FLAGS -Wdouble-promotion)
+        endif()
+    elseif (CCID STREQUAL "GNU")
+        set(C_FLAGS   -Wdouble-promotion)
+        set(CXX_FLAGS -Wno-array-bounds)
+
+        if (NOT GGML_MUSA)
+            if (CCVER VERSION_GREATER_EQUAL 7.1.0)
+                list(APPEND CXX_FLAGS -Wno-format-truncation)
+            endif()
+        endif()
+        if (CCVER VERSION_GREATER_EQUAL 8.1.0)
+            list(APPEND CXX_FLAGS -Wextra-semi)
+        endif()
+    endif()
+
+    set(GF_C_FLAGS   ${C_FLAGS}   PARENT_SCOPE)
+    set(GF_CXX_FLAGS ${CXX_FLAGS} PARENT_SCOPE)
+endfunction()
+
+if (GGML_FATAL_WARNINGS)
+    if (CMAKE_CXX_COMPILER_ID MATCHES "GNU" OR CMAKE_CXX_COMPILER_ID MATCHES "Clang")
+        list(APPEND C_FLAGS   -Werror)
+        list(APPEND CXX_FLAGS -Werror)
+    elseif (CMAKE_CXX_COMPILER_ID STREQUAL "MSVC")
+        add_compile_options(/WX)
+    endif()
+endif()
+
+if (GGML_ALL_WARNINGS)
+    if (NOT MSVC)
+        list(APPEND WARNING_FLAGS -Wall -Wextra -Wpedantic -Wcast-qual -Wno-unused-function)
+        list(APPEND C_FLAGS       -Wshadow -Wstrict-prototypes -Wpointer-arith -Wmissing-prototypes
+                                  -Werror=implicit-int -Werror=implicit-function-declaration)
+        list(APPEND CXX_FLAGS     -Wmissing-declarations -Wmissing-noreturn)
+
+        list(APPEND C_FLAGS   ${WARNING_FLAGS})
+        list(APPEND CXX_FLAGS ${WARNING_FLAGS})
+
+        get_flags(${CMAKE_CXX_COMPILER_ID} ${CMAKE_CXX_COMPILER_VERSION})
+
+        add_compile_options("$<$<COMPILE_LANGUAGE:C>:${C_FLAGS};${GF_C_FLAGS}>"
+                            "$<$<COMPILE_LANGUAGE:CXX>:${CXX_FLAGS};${GF_CXX_FLAGS}>")
+    else()
+        # todo : msvc
+        set(C_FLAGS   "")
+        set(CXX_FLAGS "")
+    endif()
+endif()
+
+set(CUDA_CXX_FLAGS "")
+
+if (GGML_CUDA)
+    set(CUDA_FLAGS -use_fast_math)
+
+    if (GGML_FATAL_WARNINGS)
+        list(APPEND CUDA_FLAGS -Werror all-warnings)
+    endif()
+
+    if (GGML_ALL_WARNINGS AND NOT MSVC)
+        set(NVCC_CMD ${CMAKE_CUDA_COMPILER} .c)
+        if (NOT CMAKE_CUDA_HOST_COMPILER STREQUAL "")
+            list(APPEND NVCC_CMD -ccbin ${CMAKE_CUDA_HOST_COMPILER})
+        endif()
+
+        execute_process(
+            COMMAND ${NVCC_CMD} -Xcompiler --version
+            OUTPUT_VARIABLE CUDA_CCFULLVER
+            ERROR_QUIET
+        )
+
+        if (NOT CUDA_CCFULLVER MATCHES clang)
+            set(CUDA_CCID "GNU")
+            execute_process(
+                COMMAND ${NVCC_CMD} -Xcompiler "-dumpfullversion -dumpversion"
+                OUTPUT_VARIABLE CUDA_CCVER
+                ERROR_QUIET
+            )
+        else()
+            if (CUDA_CCFULLVER MATCHES Apple)
+                set(CUDA_CCID "AppleClang")
+            else()
+                set(CUDA_CCID "Clang")
+            endif()
+            string(REGEX REPLACE "^.* version ([0-9.]*).*$" "\\1" CUDA_CCVER ${CUDA_CCFULLVER})
+        endif()
+
+        message("-- CUDA host compiler is ${CUDA_CCID} ${CUDA_CCVER}")
+
+        get_flags(${CUDA_CCID} ${CUDA_CCVER})
+        list(APPEND CUDA_CXX_FLAGS ${CXX_FLAGS} ${GF_CXX_FLAGS})  # This is passed to -Xcompiler later
+    endif()
+
+    if (NOT MSVC)
+        list(APPEND CUDA_CXX_FLAGS -Wno-pedantic)
+    endif()
+endif()
+
+if (GGML_LTO)
+    include(CheckIPOSupported)
+    check_ipo_supported(RESULT result OUTPUT output)
+    if (result)
+        set(CMAKE_INTERPROCEDURAL_OPTIMIZATION TRUE)
+    else()
+        message(WARNING "IPO is not supported: ${output}")
+    endif()
+endif()
+
+if (GGML_CCACHE)
+    find_program(GGML_CCACHE_FOUND ccache)
+
+    if (GGML_CCACHE_FOUND)
+        # TODO: should not be set globally
+        set_property(GLOBAL PROPERTY RULE_LAUNCH_COMPILE ccache)
+        set(ENV{CCACHE_SLOPPINESS} time_macros)
+        message(STATUS "ccache found, compilation results will be cached. Disable with GGML_CCACHE=OFF.")
+    else()
+        message(STATUS "Warning: ccache not found - consider installing it for faster compilation or disable this warning with GGML_CCACHE=OFF")
+    endif ()
+endif()
+
+# this version of Apple ld64 is buggy
+execute_process(
+    COMMAND ${CMAKE_C_COMPILER} ${CMAKE_EXE_LINKER_FLAGS} -Wl,-v
+    ERROR_VARIABLE output
+    OUTPUT_QUIET
+)
+
+if (output MATCHES "dyld-1015\.7")
+    add_compile_definitions(HAVE_BUGGY_APPLE_LINKER)
+endif()
+
+# architecture specific
+# TODO: probably these flags need to be tweaked on some architectures
+#       feel free to update the Makefile for your architecture and send a pull request or issue
+message(STATUS "CMAKE_SYSTEM_PROCESSOR: ${CMAKE_SYSTEM_PROCESSOR}")
+if (MSVC)
+    string(TOLOWER "${CMAKE_GENERATOR_PLATFORM}" CMAKE_GENERATOR_PLATFORM_LWR)
+    message(STATUS "CMAKE_GENERATOR_PLATFORM: ${CMAKE_GENERATOR_PLATFORM}")
+else ()
+    set(CMAKE_GENERATOR_PLATFORM_LWR "")
+endif ()
+
+if (NOT MSVC)
+    if (GGML_STATIC)
+        add_link_options(-static)
+        if (MINGW)
+            add_link_options(-static-libgcc -static-libstdc++)
+        endif()
+    endif()
+    if (GGML_GPROF)
+        add_compile_options(-pg)
+    endif()
+endif()
+
+set(ARCH_FLAGS "")
+
+if (CMAKE_OSX_ARCHITECTURES      STREQUAL "arm64" OR
+    CMAKE_GENERATOR_PLATFORM_LWR STREQUAL "arm64" OR
+    (NOT CMAKE_OSX_ARCHITECTURES      AND
+     NOT CMAKE_GENERATOR_PLATFORM_LWR AND
+         CMAKE_SYSTEM_PROCESSOR MATCHES "^(aarch64|arm.*|ARM64)$"))
+
+    message(STATUS "ARM detected")
+
+    if (MSVC)
+        add_compile_definitions(__aarch64__) # MSVC defines _M_ARM64 instead
+        add_compile_definitions(__ARM_NEON)
+        add_compile_definitions(__ARM_FEATURE_FMA)
+
+        set(CMAKE_REQUIRED_FLAGS_PREV ${CMAKE_REQUIRED_FLAGS})
+        string(JOIN " " CMAKE_REQUIRED_FLAGS ${CMAKE_REQUIRED_FLAGS} "/arch:armv8.2")
+
+        check_cxx_source_compiles("#include <arm_neon.h>\nint main() { int8x16_t _a, _b; int32x4_t _s = vdotq_s32(_s, _a, _b); return 0; }" GGML_COMPILER_SUPPORT_DOTPROD)
+        if (GGML_COMPILER_SUPPORT_DOTPROD)
+            add_compile_definitions(__ARM_FEATURE_DOTPROD)
+        endif ()
+
+        check_cxx_source_compiles("#include <arm_neon.h>\nint main() { int8x16_t _a, _b; int32x4_t _s = vmlaq_f32(_s, _a, _b); return 0; }" GGML_COMPILER_SUPPORT_MATMUL_INT8)
+
+        if (GGML_COMPILER_SUPPORT_MATMUL_INT8)
+            add_compile_definitions(__ARM_FEATURE_MATMUL_INT8)
+        endif ()
+
+        check_cxx_source_compiles("#include <arm_neon.h>\nint main() { float16_t _a; float16x8_t _s = vdupq_n_f16(_a); return 0; }" GGML_COMPILER_SUPPORT_FP16_VECTOR_ARITHMETIC)
+        if (GGML_COMPILER_SUPPORT_FP16_VECTOR_ARITHMETIC)
+            add_compile_definitions(__ARM_FEATURE_FP16_VECTOR_ARITHMETIC)
+        endif ()
+
+        set(CMAKE_REQUIRED_FLAGS ${CMAKE_REQUIRED_FLAGS_PREV})
+    else()
+        check_cxx_compiler_flag(-mfp16-format=ieee COMPILER_SUPPORTS_FP16_FORMAT_I3E)
+        if (NOT "${COMPILER_SUPPORTS_FP16_FORMAT_I3E}" STREQUAL "")
+            list(APPEND ARCH_FLAGS -mfp16-format=ieee)
+        endif()
+        if (${CMAKE_SYSTEM_PROCESSOR} MATCHES "armv6")
+            # Raspberry Pi 1, Zero
+            list(APPEND ARCH_FLAGS -mfpu=neon-fp-armv8 -mno-unaligned-access)
+        endif()
+        if (${CMAKE_SYSTEM_PROCESSOR} MATCHES "armv7")
+            if ("${CMAKE_SYSTEM_NAME}" STREQUAL "Android")
+                # Android armeabi-v7a
+                list(APPEND ARCH_FLAGS -mfpu=neon-vfpv4 -mno-unaligned-access -funsafe-math-optimizations)
+            else()
+                # Raspberry Pi 2
+                list(APPEND ARCH_FLAGS -mfpu=neon-fp-armv8 -mno-unaligned-access -funsafe-math-optimizations)
+            endif()
+        endif()
+        if (${CMAKE_SYSTEM_PROCESSOR} MATCHES "armv8")
+            # Android arm64-v8a
+            # Raspberry Pi 3, 4, Zero 2 (32-bit)
+            list(APPEND ARCH_FLAGS -mno-unaligned-access)
+        endif()
+        if (GGML_SVE)
+            list(APPEND ARCH_FLAGS -march=armv8.6-a+sve)
+        endif()
+    endif()
+elseif (CMAKE_OSX_ARCHITECTURES STREQUAL "x86_64" OR CMAKE_GENERATOR_PLATFORM_LWR MATCHES "^(x86_64|i686|amd64|x64|win32)$" OR
+        (NOT CMAKE_OSX_ARCHITECTURES AND NOT CMAKE_GENERATOR_PLATFORM_LWR AND
+         CMAKE_SYSTEM_PROCESSOR MATCHES "^(x86_64|i686|AMD64)$"))
+    message(STATUS "x86 detected")
+    if (MSVC)
+        # instruction set detection for MSVC only
+        if (GGML_NATIVE)
+            # TODO: improve, should not reference files from the parent folder
+            include(../cmake/FindSIMD.cmake)
+        endif ()
+        if (GGML_AVX512)
+            list(APPEND ARCH_FLAGS /arch:AVX512)
+            # MSVC has no compile-time flags enabling specific
+            # AVX512 extensions, neither it defines the
+            # macros corresponding to the extensions.
+            # Do it manually.
+            if (GGML_AVX512_VBMI)
+                add_compile_definitions($<$<COMPILE_LANGUAGE:C>:__AVX512VBMI__>)
+                add_compile_definitions($<$<COMPILE_LANGUAGE:CXX>:__AVX512VBMI__>)
+            endif()
+            if (GGML_AVX512_VNNI)
+                add_compile_definitions($<$<COMPILE_LANGUAGE:C>:__AVX512VNNI__>)
+                add_compile_definitions($<$<COMPILE_LANGUAGE:CXX>:__AVX512VNNI__>)
+            endif()
+            if (GGML_AVX512_BF16)
+                add_compile_definitions($<$<COMPILE_LANGUAGE:C>:__AVX512BF16__>)
+                add_compile_definitions($<$<COMPILE_LANGUAGE:CXX>:__AVX512BF16__>)
+            endif()
+        elseif (GGML_AVX2)
+            list(APPEND ARCH_FLAGS /arch:AVX2)
+        elseif (GGML_AVX)
+            list(APPEND ARCH_FLAGS /arch:AVX)
+        endif()
+    else()
+        if (GGML_NATIVE)
+            list(APPEND ARCH_FLAGS -march=native)
+        endif()
+        if (GGML_F16C)
+            list(APPEND ARCH_FLAGS -mf16c)
+        endif()
+        if (GGML_FMA)
+            list(APPEND ARCH_FLAGS -mfma)
+        endif()
+        if (GGML_AVX)
+            list(APPEND ARCH_FLAGS -mavx)
+        endif()
+        if (GGML_AVX2)
+            list(APPEND ARCH_FLAGS -mavx2)
+        endif()
+        if (GGML_AVX512)
+            list(APPEND ARCH_FLAGS -mavx512f)
+            list(APPEND ARCH_FLAGS -mavx512dq)
+            list(APPEND ARCH_FLAGS -mavx512bw)
+        endif()
+        if (GGML_AVX512_VBMI)
+            list(APPEND ARCH_FLAGS -mavx512vbmi)
+        endif()
+        if (GGML_AVX512_VNNI)
+            list(APPEND ARCH_FLAGS -mavx512vnni)
+        endif()
+        if (GGML_AVX512_BF16)
+            list(APPEND ARCH_FLAGS -mavx512bf16)
+        endif()
+    endif()
+elseif (${CMAKE_SYSTEM_PROCESSOR} MATCHES "ppc64")
+    message(STATUS "PowerPC detected")
+    if (${CMAKE_SYSTEM_PROCESSOR} MATCHES "ppc64le")
+        list(APPEND ARCH_FLAGS -mcpu=powerpc64le)
+    else()
+        list(APPEND ARCH_FLAGS -mcpu=native -mtune=native)
+        #TODO: Add  targets for Power8/Power9 (Altivec/VSX) and Power10(MMA) and query for big endian systems (ppc64/le/be)
+    endif()
+elseif (${CMAKE_SYSTEM_PROCESSOR} MATCHES "loongarch64")
+    message(STATUS "loongarch64 detected")
+
+    list(APPEND ARCH_FLAGS -march=loongarch64)
+    if (GGML_LASX)
+        list(APPEND ARCH_FLAGS -mlasx)
+    endif()
+    if (GGML_LSX)
+        list(APPEND ARCH_FLAGS -mlsx)
+    endif()
+else()
+    message(STATUS "Unknown architecture")
+endif()
+
+add_compile_options("$<$<COMPILE_LANGUAGE:CXX>:${ARCH_FLAGS}>")
+add_compile_options("$<$<COMPILE_LANGUAGE:C>:${ARCH_FLAGS}>")
+
+if (GGML_CUDA)
+    list(APPEND CUDA_CXX_FLAGS ${ARCH_FLAGS})
+    list(JOIN   CUDA_CXX_FLAGS " " CUDA_CXX_FLAGS_JOINED)  # pass host compiler flags as a single argument
+
+    if (NOT CUDA_CXX_FLAGS_JOINED STREQUAL "")
+        list(APPEND CUDA_FLAGS -Xcompiler ${CUDA_CXX_FLAGS_JOINED})
+    endif()
+
+    add_compile_options("$<$<COMPILE_LANGUAGE:CUDA>:${CUDA_FLAGS}>")
+endif()
+
+if (MINGW)
+    # Target Windows 8 for PrefetchVirtualMemory
+    add_compile_definitions(_WIN32_WINNT=${GGML_WIN_VER})
+endif()
+
+#
+# POSIX conformance
+#
+
+# clock_gettime came in POSIX.1b (1993)
+# CLOCK_MONOTONIC came in POSIX.1-2001 / SUSv3 as optional
+# posix_memalign came in POSIX.1-2001 / SUSv3
+# M_PI is an XSI extension since POSIX.1-2001 / SUSv3, came in XPG1 (1985)
+add_compile_definitions(_XOPEN_SOURCE=600)
+
+# Somehow in OpenBSD whenever POSIX conformance is specified
+# some string functions rely on locale_t availability,
+# which was introduced in POSIX.1-2008, forcing us to go higher
+if (CMAKE_SYSTEM_NAME MATCHES "OpenBSD")
+    remove_definitions(-D_XOPEN_SOURCE=600)
+    add_compile_definitions(_XOPEN_SOURCE=700)
+endif()
+
+# Data types, macros and functions related to controlling CPU affinity and
+# some memory allocation are available on Linux through GNU extensions in libc
+if (CMAKE_SYSTEM_NAME MATCHES "Linux" OR CMAKE_SYSTEM_NAME MATCHES "Android")
+    add_compile_definitions(_GNU_SOURCE)
+endif()
+
+# RLIMIT_MEMLOCK came in BSD, is not specified in POSIX.1,
+# and on macOS its availability depends on enabling Darwin extensions
+# similarly on DragonFly, enabling BSD extensions is necessary
+if (
+    CMAKE_SYSTEM_NAME MATCHES "Darwin" OR
+    CMAKE_SYSTEM_NAME MATCHES "iOS"    OR
+    CMAKE_SYSTEM_NAME MATCHES "tvOS"   OR
+    CMAKE_SYSTEM_NAME MATCHES "DragonFly"
+)
+    add_compile_definitions(_DARWIN_C_SOURCE)
+endif()
+
+# alloca is a non-standard interface that is not visible on BSDs when
+# POSIX conformance is specified, but not all of them provide a clean way
+# to enable it in such cases
+if (CMAKE_SYSTEM_NAME MATCHES "FreeBSD")
+    add_compile_definitions(__BSD_VISIBLE)
+endif()
+if (CMAKE_SYSTEM_NAME MATCHES "NetBSD")
+    add_compile_definitions(_NETBSD_SOURCE)
+endif()
+if (CMAKE_SYSTEM_NAME MATCHES "OpenBSD")
+    add_compile_definitions(_BSD_SOURCE)
+endif()
+
+if (WIN32)
+    add_compile_definitions(_CRT_SECURE_NO_WARNINGS)
+
+    if (BUILD_SHARED_LIBS)
+        # TODO: should not use this
+        set(CMAKE_WINDOWS_EXPORT_ALL_SYMBOLS ON)
+    endif()
+endif()
+
+#
+# libraries
+#
+
+# ggml
+
+add_library(ggml
+            ../include/ggml.h
+            ../include/ggml-alloc.h
+            ../include/ggml-backend.h
+            ggml.c
+            ggml-alloc.c
+            ggml-backend.cpp
+            ggml-quants.c
+            ggml-quants.h
+            ${GGML_SOURCES_CUDA}      ${GGML_HEADERS_CUDA}
+            ${GGML_SOURCES_METAL}     ${GGML_HEADERS_METAL}
+            ${GGML_SOURCES_RPC}       ${GGML_HEADERS_RPC}
+            ${GGML_SOURCES_EXTRA}     ${GGML_HEADERS_EXTRA}
+            ${GGML_SOURCES_SYCL}      ${GGML_HEADERS_SYCL}
+            ${GGML_SOURCES_KOMPUTE}   ${GGML_HEADERS_KOMPUTE}
+            ${GGML_SOURCES_VULKAN}    ${GGML_HEADERS_VULKAN}
+            ${GGML_SOURCES_ROCM}      ${GGML_HEADERS_ROCM}
+            ${GGML_SOURCES_BLAS}      ${GGML_HEADERS_BLAS}
+            ${GGML_SOURCES_LLAMAFILE} ${GGML_HEADERS_LLAMAFILE}
+            ${GGML_SOURCES_CANN}      ${GGML_HEADERS_CANN}
+            ggml-aarch64.c            ggml-aarch64.h
+            )
+
+if (EMSCRIPTEN)
+    set_target_properties(ggml PROPERTIES COMPILE_FLAGS "-msimd128")
+endif()
+
+target_compile_definitions(ggml PUBLIC    ${GGML_CDEF_PUBLIC})
+target_include_directories(ggml PUBLIC  ../include)
+target_include_directories(ggml PRIVATE . ${GGML_EXTRA_INCLUDES})
+target_link_directories   (ggml PRIVATE   ${GGML_EXTRA_LIBDIRS})
+target_compile_features   (ggml PRIVATE c_std_11) # don't bump
+
+list(APPEND GGML_EXTRA_LIBS_PRIVATE Threads::Threads)
+
+find_library(MATH_LIBRARY m)
+if (MATH_LIBRARY)
+    if (NOT WIN32 OR NOT GGML_SYCL)
+        list(APPEND GGML_EXTRA_LIBS_PRIVATE m)
+    endif()
+endif()
+
+list(REMOVE_DUPLICATES GGML_EXTRA_LIBS_PRIVATE)
+list(REMOVE_DUPLICATES GGML_EXTRA_LIBS_PUBLIC)
+target_link_libraries(ggml PRIVATE ${GGML_EXTRA_LIBS_PRIVATE} PUBLIC ${GGML_EXTRA_LIBS_PUBLIC})
+
+if (BUILD_SHARED_LIBS)
+    set_target_properties(ggml PROPERTIES POSITION_INDEPENDENT_CODE ON)
+    target_compile_definitions(ggml PRIVATE GGML_SHARED GGML_BUILD)
+endif()
diff --git a/src/whisper_cpp/ggml/src/ggml-aarch64.c b/src/whisper_cpp/ggml/src/ggml-aarch64.c
new file mode 100644
index 00000000..b27f4114
--- /dev/null
+++ b/src/whisper_cpp/ggml/src/ggml-aarch64.c
@@ -0,0 +1,3209 @@
+// SPDX-FileCopyrightText: Copyright 2024 Arm Limited and/or its affiliates <open-source-office@arm.com>
+// SPDX-License-Identifier: MIT
+//
+
+#define GGML_COMMON_IMPL_C
+#include "ggml-common.h"
+
+#include "ggml-quants.h"
+#include "ggml-impl.h"
+#include "ggml-cpu-impl.h"
+
+#include <math.h>
+#include <string.h>
+#include <assert.h>
+#include <float.h>
+#include <stdlib.h> // for qsort
+#include <stdio.h>  // for GGML_ASSERT
+
+#include "ggml-aarch64.h"
+
+#if defined(__GNUC__)
+#pragma GCC diagnostic ignored "-Woverlength-strings"
+#elif defined(_MSC_VER)
+#pragma warning(disable: 4244 4267) // possible loss of data
+#endif
+
+#define UNUSED GGML_UNUSED
+
+// Functions to create the interleaved data layout formats
+
+// interleave 4 block_q4_0s in blocks of blck_size_interleave
+// returns an interleaved block_q4_0x4
+// in the interleaved block_q4_0x4, place deltas for 4 block_q4_0 blocks
+// first, then interleave quants from 4 block_q4_0s in blocks of blck_size_interleave
+//
+// - in                  : an array of block_q4_0 pointers
+// - blck_size_interleave : the block_q4_0 quants bytes are interleaved in blocks of
+//                         blck_size_interleave bytes
+// - xor_mask            : the mask to convert the nibbles in block_q4_0 quants bytes
+//                         from bias offset form to pure sign form (this saves subtract
+//                         operations durin unpacking)
+//
+#if defined(__AVX__)
+#if defined(__F16C__)
+#if defined(__AVX512F__)
+#define GGML_F32Cx8x2_LOAD(x, y)     _mm512_cvtph_ps(_mm256_set_m128i(_mm_loadu_si128((const __m128i *)(y)), _mm_loadu_si128((const __m128i *)(x))))
+#define GGML_F32Cx16_REPEAT_LOAD(x)  _mm512_cvtph_ps(_mm256_set_m128i(x, x))
+#endif
+// the  _mm256_cvt intrinsics require F16C
+#define GGML_F32Cx8_LOAD(x)     _mm256_cvtph_ps(_mm_loadu_si128((const __m128i *)(x)))
+#define GGML_F32Cx8_REPEAT_LOAD(x, loadMask)     _mm256_cvtph_ps(_mm_shuffle_epi32(_mm_maskload_epi32((int const*)(x), loadMask), 68))
+#define GGML_F32Cx8_REARRANGE_LOAD(x, arrangeMask)     _mm256_cvtph_ps(_mm_shuffle_epi8(_mm_loadu_si128((const __m128i *) x), arrangeMask))
+#else
+#if defined(__AVX512F__)
+static inline __m512 __avx512_f32cx8x2_load(ggml_fp16_t *x, ggml_fp16_t *y) {
+    float tmp[16];
+
+    for (int i = 0; i < 8; i++) {
+        tmp[i] = GGML_FP16_TO_FP32(x[i]);
+    }
+
+    for (int i = 0; i < 8; i++) {
+        tmp[i + 8] = GGML_FP16_TO_FP32(y[i]);
+    }
+
+    return _mm512_loadu_ps(tmp);
+}
+static inline __m512 __avx512_repeat_f32cx16_load(__m128i x) {
+    float tmp[16];
+    uint16_t tmphalf[8];
+    _mm_storeu_si128((__m128i*)tmphalf, x);
+
+    for (int i = 0; i < 4; i++) {
+        tmp[i] = GGML_FP16_TO_FP32(tmphalf[i]);
+        tmp[i + 4] = GGML_FP16_TO_FP32(tmphalf[i]);
+        tmp[i + 8] = GGML_FP16_TO_FP32(tmphalf[i]);
+        tmp[i + 12] = GGML_FP16_TO_FP32(tmphalf[i]);
+    }
+
+    return _mm512_loadu_ps(tmp);
+}
+#endif
+static inline __m256 __avx_f32cx8_load(ggml_fp16_t *x) {
+    float tmp[8];
+
+    for (int i = 0; i < 8; i++) {
+        tmp[i] = GGML_FP16_TO_FP32(x[i]);
+    }
+
+    return _mm256_loadu_ps(tmp);
+}
+static inline __m256 __avx_repeat_f32cx8_load(ggml_fp16_t *x) {
+    float tmp[8];
+
+    for (int i = 0; i < 4; i++) {
+        tmp[i] = GGML_FP16_TO_FP32(x[i]);
+        tmp[i + 4] = GGML_FP16_TO_FP32(x[i]);
+    }
+
+    return _mm256_loadu_ps(tmp);
+}
+static inline __m256 __avx_rearranged_f32cx8_load(ggml_fp16_t *x, __m128i arrangeMask) {
+    uint16_t tmphalf[8];
+    float tmp[8];
+
+    _mm_storeu_si128((__m128i*)tmphalf, _mm_shuffle_epi8(_mm_loadu_si128((const __m128i *) x), arrangeMask));
+    for (int i = 0; i < 8; i++) {
+        tmp[i] = GGML_FP16_TO_FP32(tmphalf[i]);
+    }
+
+    return _mm256_loadu_ps(tmp);
+}
+
+#define GGML_F32Cx8_LOAD(x)     __avx_f32cx8_load(x)
+#define GGML_F32Cx8_REPEAT_LOAD(x, loadMask)     __avx_repeat_f32cx8_load(x)
+#define GGML_F32Cx8_REARRANGE_LOAD(x, arrangeMask)     __avx_rearranged_f32cx8_load(x, arrangeMask)
+#if defined(__AVX512F__)
+#define GGML_F32Cx8x2_LOAD(x, y)     __avx512_f32cx8x2_load(x, y)
+#define GGML_F32Cx16_REPEAT_LOAD(x)  __avx512_repeat_f32cx16_load(x)
+#endif
+#endif
+#endif
+
+
+#if defined(__AVX2__) || defined(__AVX512F__)
+#if defined(__AVX512F__)
+// add int16_t pairwise and return as 512 bit int vector
+static inline __m512i sum_i16_pairs_int_32x16(const __m512i x) {
+    const __m512i ones = _mm512_set1_epi16(1);
+    return _mm512_madd_epi16(ones, x);
+}
+
+static inline __m512i mul_sum_us8_pairs_int32x16(const __m512i ax, const __m512i sy) {
+#if defined(__AVXVNNI__) || (defined(__AVX512VNNI__) && defined(__AVX512VL__))
+    const __m512i zero = _mm512_setzero_si512();
+    return _mm512_dpbusd_epi32(zero, ax, sy);
+#else
+    // Perform multiplication and create 16-bit values
+    const __m512i dot = _mm512_maddubs_epi16(ax, sy);
+    return sum_i16_pairs_int_32x16(dot);
+#endif
+}
+
+// multiply int8_t, add results pairwise twice and return as 512 bit int vector
+static inline __m512i mul_sum_i8_pairs_int32x16(const __m512i x, const __m512i y) {
+    const __m512i zero = _mm512_setzero_si512();
+    // Get absolute values of x vectors
+    const __m512i ax = _mm512_abs_epi8(x);
+    // Sign the values of the y vectors
+    __mmask64 blt0 = _mm512_movepi8_mask(x);
+    const __m512i sy = _mm512_mask_sub_epi8(y, blt0, zero, y);
+    return mul_sum_us8_pairs_int32x16(ax, sy);
+}
+#endif
+
+// add int16_t pairwise and return as 256 bit int vector
+static inline __m256i sum_i16_pairs_int32x8(const __m256i x) {
+    const __m256i ones = _mm256_set1_epi16(1);
+    return _mm256_madd_epi16(ones, x);
+}
+
+static inline __m256i mul_sum_us8_pairs_int32x8(const __m256i ax, const __m256i sy) {
+#if defined(__AVXVNNI__) || (defined(__AVX512VNNI__) && defined(__AVX512VL__))
+    const __m256i zero = _mm256_setzero_si256();
+    return _mm256_dpbusd_epi32(zero, ax, sy);
+#else
+    // Perform multiplication and create 16-bit values
+    const __m256i dot = _mm256_maddubs_epi16(ax, sy);
+    return sum_i16_pairs_int32x8(dot);
+#endif
+}
+
+// Integer variant of the function defined in ggml-quants.c
+// multiply int8_t, add results pairwise twice and return as 256 bit int vector
+static inline __m256i mul_sum_i8_pairs_int32x8(const __m256i x, const __m256i y) {
+#if __AVXVNNIINT8__
+    const __m256i zero = _mm256_setzero_si256();
+    return _mm256_dpbssd_epi32(zero, x, y);
+#else
+    // Get absolute values of x vectors
+    const __m256i ax = _mm256_sign_epi8(x, x);
+    // Sign the values of the y vectors
+    const __m256i sy = _mm256_sign_epi8(y, x);
+    return mul_sum_us8_pairs_int32x8(ax, sy);
+#endif
+}
+#endif
+
+static block_q4_0x4 make_block_q4_0x4(block_q4_0 * in, unsigned int blck_size_interleave, unsigned int xor_mask) {
+    block_q4_0x4 out;
+
+    for (int i = 0; i < 4; i++) {
+        out.d[i] = in[i].d;
+    }
+
+    for (int i = 0; i < QK4_0 * 2; i++) {
+        int src_offset = (i / (4 * blck_size_interleave)) * blck_size_interleave;
+        int src_id = (i % (4 * blck_size_interleave)) / blck_size_interleave;
+        src_offset += (i % blck_size_interleave);
+
+        out.qs[i] = in[src_id].qs[src_offset] ^ xor_mask;
+    }
+
+    return out;
+}
+
+// interleave 8 block_q4_0s in blocks of blck_size_interleave
+// returns an interleaved block_q4_0x8
+// in the interleaved block_q4_0x8, place deltas for 8 block_q4_0 blocks
+// first, then interleave quants from 8 block_q4_0s in blocks of blck_size_interleave
+static block_q4_0x8 make_block_q4_0x8(block_q4_0 * in, unsigned int blck_size_interleave, unsigned int xor_mask) {
+    block_q4_0x8 out;
+
+    for (int i = 0; i < 8; i++) {
+        out.d[i] = in[i].d;
+    }
+
+    for (int i = 0; i < QK4_0 * 4; i++) {
+        int src_offset = (i / (8 * blck_size_interleave)) * blck_size_interleave;
+        int src_id = (i % (8 * blck_size_interleave)) / blck_size_interleave;
+        src_offset += (i % blck_size_interleave);
+
+        out.qs[i] = in[src_id].qs[src_offset] ^ xor_mask;
+    }
+
+    return out;
+}
+
+void quantize_q8_0_4x4(const float * restrict x, void * restrict vy, int64_t k) {
+    assert(QK8_0 == 32);
+    assert(k % QK8_0 == 0);
+    const int nb = k / QK8_0;
+
+    block_q8_0x4 * restrict y = (block_q8_0x4 *) vy;
+
+#if defined(__ARM_NEON)
+    float32x4_t srcv[4][8];
+    float id[4];
+
+    for (int i = 0; i < nb; i++) {
+        float32x4_t asrcv[8];
+        float32x4_t amaxv[8];
+
+        for (int row_iter = 0; row_iter < 4; row_iter++) {
+            for (int j = 0; j < 8; j++) srcv[row_iter][j] = vld1q_f32(x + row_iter * k + i * 32 + 4 * j);
+            for (int j = 0; j < 8; j++) asrcv[j] = vabsq_f32(srcv[row_iter][j]);
+
+            for (int j = 0; j < 4; j++) amaxv[2 * j] = vmaxq_f32(asrcv[2 * j], asrcv[2 * j + 1]);
+            for (int j = 0; j < 2; j++) amaxv[4 * j] = vmaxq_f32(amaxv[4 * j], amaxv[4 * j + 2]);
+            for (int j = 0; j < 1; j++) amaxv[8 * j] = vmaxq_f32(amaxv[8 * j], amaxv[8 * j + 4]);
+
+            const float amax = vmaxvq_f32(amaxv[0]);
+
+            const float d = amax / ((1 << 7) - 1);
+            id[row_iter] = d ? 1.0f / d : 0.0f;
+
+            y[i].d[row_iter] = GGML_FP32_TO_FP16(d);
+        }
+
+        for (int j = 0; j < 8; j++) {
+            float32x4_t v = vmulq_n_f32(srcv[0][j], id[0]);
+            int32x4_t vi = vcvtnq_s32_f32(v);
+            y[i].qs[16 * j + 0] = vgetq_lane_s32(vi, 0);
+            y[i].qs[16 * j + 1] = vgetq_lane_s32(vi, 1);
+            y[i].qs[16 * j + 2] = vgetq_lane_s32(vi, 2);
+            y[i].qs[16 * j + 3] = vgetq_lane_s32(vi, 3);
+
+            v = vmulq_n_f32(srcv[1][j], id[1]);
+            vi = vcvtnq_s32_f32(v);
+            y[i].qs[16 * j + 4] = vgetq_lane_s32(vi, 0);
+            y[i].qs[16 * j + 5] = vgetq_lane_s32(vi, 1);
+            y[i].qs[16 * j + 6] = vgetq_lane_s32(vi, 2);
+            y[i].qs[16 * j + 7] = vgetq_lane_s32(vi, 3);
+
+            v = vmulq_n_f32(srcv[2][j], id[2]);
+            vi = vcvtnq_s32_f32(v);
+            y[i].qs[16 * j + 8] = vgetq_lane_s32(vi, 0);
+            y[i].qs[16 * j + 9] = vgetq_lane_s32(vi, 1);
+            y[i].qs[16 * j + 10] = vgetq_lane_s32(vi, 2);
+            y[i].qs[16 * j + 11] = vgetq_lane_s32(vi, 3);
+
+            v = vmulq_n_f32(srcv[3][j], id[3]);
+            vi = vcvtnq_s32_f32(v);
+            y[i].qs[16 * j + 12] = vgetq_lane_s32(vi, 0);
+            y[i].qs[16 * j + 13] = vgetq_lane_s32(vi, 1);
+            y[i].qs[16 * j + 14] = vgetq_lane_s32(vi, 2);
+            y[i].qs[16 * j + 15] = vgetq_lane_s32(vi, 3);
+        }
+    }
+#else
+    // scalar
+    const int blck_size_interleave = 4;
+    float srcv[4][QK8_0];
+    float id[4];
+
+    for (int i = 0; i < nb; i++) {
+        for (int row_iter = 0; row_iter < 4; row_iter++) {
+            float amax = 0.0f; // absolute max
+
+            for (int j = 0; j < QK8_0; j++) {
+                srcv[row_iter][j] = x[row_iter * k + i * QK8_0 + j];
+                amax = MAX(amax, fabsf(srcv[row_iter][j]));
+            }
+
+            const float d = amax / ((1 << 7) - 1);
+            id[row_iter] = d ? 1.0f / d : 0.0f;
+
+            y[i].d[row_iter] = GGML_FP32_TO_FP16(d);
+        }
+
+        for (int j = 0; j < QK8_0 * 4; j++) {
+            int src_offset = (j / (4 * blck_size_interleave)) * blck_size_interleave;
+            int src_id = (j % (4 * blck_size_interleave)) / blck_size_interleave;
+            src_offset += (j % blck_size_interleave);
+
+            float x0 = srcv[src_id][src_offset] * id[src_id];
+            y[i].qs[j] = roundf(x0);
+        }
+    }
+#endif
+}
+
+void quantize_q8_0_4x8(const float * restrict x, void * restrict vy, int64_t k) {
+    assert(QK8_0 == 32);
+    assert(k % QK8_0 == 0);
+    const int nb = k / QK8_0;
+
+    block_q8_0x4 * restrict y = (block_q8_0x4 *) vy;
+
+#if defined(__ARM_NEON)
+    float32x4_t srcv[4][8];
+    float id[4];
+
+    for (int i = 0; i < nb; i++) {
+        float32x4_t asrcv[8];
+        float32x4_t amaxv[8];
+
+        for (int row_iter = 0; row_iter < 4; row_iter++) {
+            for (int j = 0; j < 8; j++) srcv[row_iter][j] = vld1q_f32(x + row_iter * k + i * 32 + 4 * j);
+            for (int j = 0; j < 8; j++) asrcv[j] = vabsq_f32(srcv[row_iter][j]);
+
+            for (int j = 0; j < 4; j++) amaxv[2 * j] = vmaxq_f32(asrcv[2 * j], asrcv[2 * j + 1]);
+            for (int j = 0; j < 2; j++) amaxv[4 * j] = vmaxq_f32(amaxv[4 * j], amaxv[4 * j + 2]);
+            for (int j = 0; j < 1; j++) amaxv[8 * j] = vmaxq_f32(amaxv[8 * j], amaxv[8 * j + 4]);
+
+            const float amax = vmaxvq_f32(amaxv[0]);
+
+            const float d = amax / ((1 << 7) - 1);
+            id[row_iter] = d ? 1.0f / d : 0.0f;
+
+            y[i].d[row_iter] = GGML_FP32_TO_FP16(d);
+        }
+
+        for (int j = 0; j < 4; j++) {
+            float32x4_t v = vmulq_n_f32(srcv[0][2 * j], id[0]);
+            int32x4_t vi = vcvtnq_s32_f32(v);
+            y[i].qs[32 * j + 0] = vgetq_lane_s32(vi, 0);
+            y[i].qs[32 * j + 1] = vgetq_lane_s32(vi, 1);
+            y[i].qs[32 * j + 2] = vgetq_lane_s32(vi, 2);
+            y[i].qs[32 * j + 3] = vgetq_lane_s32(vi, 3);
+            v = vmulq_n_f32(srcv[0][2 * j + 1], id[0]);
+            vi = vcvtnq_s32_f32(v);
+            y[i].qs[32 * j + 4] = vgetq_lane_s32(vi, 0);
+            y[i].qs[32 * j + 5] = vgetq_lane_s32(vi, 1);
+            y[i].qs[32 * j + 6] = vgetq_lane_s32(vi, 2);
+            y[i].qs[32 * j + 7] = vgetq_lane_s32(vi, 3);
+
+            v = vmulq_n_f32(srcv[1][2 * j], id[1]);
+            vi = vcvtnq_s32_f32(v);
+            y[i].qs[32 * j + 8] = vgetq_lane_s32(vi, 0);
+            y[i].qs[32 * j + 9] = vgetq_lane_s32(vi, 1);
+            y[i].qs[32 * j + 10] = vgetq_lane_s32(vi, 2);
+            y[i].qs[32 * j + 11] = vgetq_lane_s32(vi, 3);
+            v = vmulq_n_f32(srcv[1][2 * j + 1], id[1]);
+            vi = vcvtnq_s32_f32(v);
+            y[i].qs[32 * j + 12] = vgetq_lane_s32(vi, 0);
+            y[i].qs[32 * j + 13] = vgetq_lane_s32(vi, 1);
+            y[i].qs[32 * j + 14] = vgetq_lane_s32(vi, 2);
+            y[i].qs[32 * j + 15] = vgetq_lane_s32(vi, 3);
+
+            v = vmulq_n_f32(srcv[2][2 * j], id[2]);
+            vi = vcvtnq_s32_f32(v);
+            y[i].qs[32 * j + 16] = vgetq_lane_s32(vi, 0);
+            y[i].qs[32 * j + 17] = vgetq_lane_s32(vi, 1);
+            y[i].qs[32 * j + 18] = vgetq_lane_s32(vi, 2);
+            y[i].qs[32 * j + 19] = vgetq_lane_s32(vi, 3);
+            v = vmulq_n_f32(srcv[2][2 * j + 1], id[2]);
+            vi = vcvtnq_s32_f32(v);
+            y[i].qs[32 * j + 20] = vgetq_lane_s32(vi, 0);
+            y[i].qs[32 * j + 21] = vgetq_lane_s32(vi, 1);
+            y[i].qs[32 * j + 22] = vgetq_lane_s32(vi, 2);
+            y[i].qs[32 * j + 23] = vgetq_lane_s32(vi, 3);
+
+            v = vmulq_n_f32(srcv[3][2 * j], id[3]);
+            vi = vcvtnq_s32_f32(v);
+            y[i].qs[32 * j + 24] = vgetq_lane_s32(vi, 0);
+            y[i].qs[32 * j + 25] = vgetq_lane_s32(vi, 1);
+            y[i].qs[32 * j + 26] = vgetq_lane_s32(vi, 2);
+            y[i].qs[32 * j + 27] = vgetq_lane_s32(vi, 3);
+            v = vmulq_n_f32(srcv[3][2 * j + 1], id[3]);
+            vi = vcvtnq_s32_f32(v);
+            y[i].qs[32 * j + 28] = vgetq_lane_s32(vi, 0);
+            y[i].qs[32 * j + 29] = vgetq_lane_s32(vi, 1);
+            y[i].qs[32 * j + 30] = vgetq_lane_s32(vi, 2);
+            y[i].qs[32 * j + 31] = vgetq_lane_s32(vi, 3);
+        }
+    }
+#elif defined(__AVX2__) || defined(__AVX__)
+    float id[4];
+    __m256 srcv[4][4];
+    __m256 idvec[4];
+
+    for (int i = 0; i < nb; i++) {
+        for (int row_iter = 0; row_iter < 4; row_iter++) {
+            // Load elements into 4 AVX vectors
+            __m256 v0 = _mm256_loadu_ps( x + row_iter * k + i * 32 );
+            __m256 v1 = _mm256_loadu_ps( x + row_iter * k + i * 32 + 8 );
+            __m256 v2 = _mm256_loadu_ps( x + row_iter * k + i * 32 + 16 );
+            __m256 v3 = _mm256_loadu_ps( x + row_iter * k + i * 32 + 24 );
+
+            // Compute max(abs(e)) for the block
+            const __m256 signBit = _mm256_set1_ps( -0.0f );
+            __m256 maxAbs = _mm256_andnot_ps( signBit, v0 );
+            maxAbs = _mm256_max_ps( maxAbs, _mm256_andnot_ps( signBit, v1 ) );
+            maxAbs = _mm256_max_ps( maxAbs, _mm256_andnot_ps( signBit, v2 ) );
+            maxAbs = _mm256_max_ps( maxAbs, _mm256_andnot_ps( signBit, v3 ) );
+
+            __m128 max4 = _mm_max_ps( _mm256_extractf128_ps( maxAbs, 1 ), _mm256_castps256_ps128( maxAbs ) );
+            max4 = _mm_max_ps( max4, _mm_movehl_ps( max4, max4 ) );
+            max4 = _mm_max_ss( max4, _mm_movehdup_ps( max4 ) );
+            const float maxScalar = _mm_cvtss_f32( max4 );
+
+            // Divided by 127.f to mirror results in quantize_row_q8_0
+            const float d = maxScalar  / 127.f;
+            id[row_iter] = ( maxScalar != 0.0f ) ? 127.f / maxScalar : 0.0f; //d ? 1.0f / d : 0.0f;
+
+            // Store the scale for the individual block
+            y[i].d[row_iter] = GGML_FP32_TO_FP16(d);
+
+            // Store the values in blocks of eight values - Aim is to use these later for block interleaving
+            srcv[row_iter][0] = v0;
+            srcv[row_iter][1] = v1;
+            srcv[row_iter][2] = v2;
+            srcv[row_iter][3] = v3;
+            idvec[row_iter] = _mm256_set1_ps(id[row_iter]);
+        }
+
+        // The loop iterates four times - The aim is to get 4 corresponding chunks of eight bytes from the original weight blocks that are interleaved
+        for (int j = 0; j < 4; j++) {
+            // Apply the multiplier
+            __m256 v0 = _mm256_mul_ps(srcv[0][j], idvec[0]);
+            __m256 v1 = _mm256_mul_ps(srcv[1][j], idvec[1]);
+            __m256 v2 = _mm256_mul_ps(srcv[2][j], idvec[2]);
+            __m256 v3 = _mm256_mul_ps(srcv[3][j], idvec[3]);
+
+            // Round to nearest integer
+            v0 = _mm256_round_ps( v0, _MM_ROUND_NEAREST );
+            v1 = _mm256_round_ps( v1, _MM_ROUND_NEAREST );
+            v2 = _mm256_round_ps( v2, _MM_ROUND_NEAREST );
+            v3 = _mm256_round_ps( v3, _MM_ROUND_NEAREST );
+
+            // Convert floats to integers
+            __m256i i0 = _mm256_cvtps_epi32( v0 );
+            __m256i i1 = _mm256_cvtps_epi32( v1 );
+            __m256i i2 = _mm256_cvtps_epi32( v2 );
+            __m256i i3 = _mm256_cvtps_epi32( v3 );
+
+#if defined(__AVX2__)
+            // Convert int32 to int16
+            i0 = _mm256_packs_epi32( i0, i1 );
+            i2 = _mm256_packs_epi32( i2, i3 );
+            // Convert int16 to int8
+            i0 = _mm256_packs_epi16( i0, i2 );
+
+            //  Permute and store the quantized weights in the required order after the pack instruction
+            const __m256i perm = _mm256_setr_epi32( 0, 4, 1, 5, 2, 6, 3, 7 );
+            i0 = _mm256_permutevar8x32_epi32( i0, perm );
+
+            _mm256_storeu_si256((__m256i *)(y[i].qs + 32 * j), i0);
+#else
+            // Since we don't have in AVX some necessary functions,
+            // we split the registers in half and call AVX2 analogs from SSE
+            __m128i ni0 = _mm256_castsi256_si128( i0 );
+            __m128i ni1 = _mm256_extractf128_si256( i0, 1);
+            __m128i ni2 = _mm256_castsi256_si128( i1 );
+            __m128i ni3 = _mm256_extractf128_si256( i1, 1);
+            __m128i ni4 = _mm256_castsi256_si128( i2 );
+            __m128i ni5 = _mm256_extractf128_si256( i2, 1);
+            __m128i ni6 = _mm256_castsi256_si128( i3 );
+            __m128i ni7 = _mm256_extractf128_si256( i3, 1);
+
+            // Convert int32 to int16
+            ni0 = _mm_packs_epi32( ni0, ni1 );
+            ni2 = _mm_packs_epi32( ni2, ni3 );
+            ni4 = _mm_packs_epi32( ni4, ni5 );
+            ni6 = _mm_packs_epi32( ni6, ni7 );
+            // Convert int16 to int8
+            ni0 = _mm_packs_epi16( ni0, ni2 );
+            ni4 = _mm_packs_epi16( ni4, ni6 );
+            _mm_storeu_si128((__m128i *)(y[i].qs + 32 * j), ni0);
+            _mm_storeu_si128((__m128i *)(y[i].qs + 32 * j + 16), ni4);
+#endif
+        }
+    }
+#else
+    // scalar
+    const int blck_size_interleave = 8;
+    float srcv[4][QK8_0];
+    float id[4];
+
+    for (int i = 0; i < nb; i++) {
+        for (int row_iter = 0; row_iter < 4; row_iter++) {
+            float amax = 0.0f; // absolute max
+
+            for (int j = 0; j < QK8_0; j++) {
+                srcv[row_iter][j] = x[row_iter * k + i * QK8_0 + j];
+                amax = MAX(amax, fabsf(srcv[row_iter][j]));
+            }
+
+            const float d = amax / ((1 << 7) - 1);
+            id[row_iter] = d ? 1.0f / d : 0.0f;
+
+            y[i].d[row_iter] = GGML_FP32_TO_FP16(d);
+        }
+
+        for (int j = 0; j < QK8_0 * 4; j++) {
+            int src_offset = (j / (4 * blck_size_interleave)) * blck_size_interleave;
+            int src_id = (j % (4 * blck_size_interleave)) / blck_size_interleave;
+            src_offset += (j % blck_size_interleave);
+
+            float x0 = srcv[src_id][src_offset] * id[src_id];
+            y[i].qs[j] = roundf(x0);
+        }
+    }
+#endif
+}
+
+void quantize_mat_q8_0(const float * restrict x, void * restrict vy, int64_t nrow, int64_t n_per_row, int64_t blck_size_interleave) {
+    assert(nrow == 4);
+    UNUSED(nrow);
+    if (blck_size_interleave == 4) {
+        quantize_q8_0_4x4(x, vy, n_per_row);
+    } else if (blck_size_interleave == 8) {
+        quantize_q8_0_4x8(x, vy, n_per_row);
+    } else {
+        assert(false);
+    }
+}
+
+static size_t quantize_q4_0_nr_bl(const float * restrict src, void * restrict dst, int64_t nrow, int64_t n_per_row, int nrows_interleaved, int blck_size_interleave) {
+    assert(n_per_row % QK4_0 == 0);
+    const int nb = n_per_row / QK4_0;
+
+    void * out_ptr = NULL;
+    if (nrows_interleaved == 8) {
+        out_ptr = (block_q4_0x8 *) dst;
+    }
+    else if (nrows_interleaved == 4) {
+        out_ptr = (block_q4_0x4 *) dst;
+    }
+    assert(nrows_interleaved <= 8);
+    block_q4_0 dst_tmp[8];
+
+    for (int b = 0; b < (nrow * n_per_row); b += nrows_interleaved * n_per_row) {
+
+        for (int64_t x = 0; x < nb; x++) {
+
+            for (int i  = 0; i < nrows_interleaved; i++ ) {
+                quantize_row_q4_0_ref(src + b + i * n_per_row + x * QK4_0, (block_q4_0 *) dst_tmp + i, QK4_0);
+            }
+
+            if (nrows_interleaved == 8) {
+                *(block_q4_0x8 *) out_ptr = make_block_q4_0x8(dst_tmp, blck_size_interleave, 0x88);
+                out_ptr = (block_q4_0x8 *) out_ptr + 1;
+            }
+            else if (nrows_interleaved == 4) {
+                *(block_q4_0x4 *) out_ptr = make_block_q4_0x4(dst_tmp, blck_size_interleave, 0x88);
+                out_ptr = (block_q4_0x4 *) out_ptr + 1;
+            }
+        }
+    }
+
+    return ((nrow * n_per_row) / QK4_0 * sizeof(block_q4_0));
+}
+
+size_t quantize_q4_0_4x4(const float * restrict src, void * restrict dst, int64_t nrow, int64_t n_per_row, const float * quant_weights) {
+    UNUSED(quant_weights);
+    return quantize_q4_0_nr_bl(src, dst, nrow, n_per_row, 4, 4);
+}
+
+size_t quantize_q4_0_4x8(const float * restrict src, void * restrict dst, int64_t nrow, int64_t n_per_row, const float * quant_weights) {
+    UNUSED(quant_weights);
+    return quantize_q4_0_nr_bl(src, dst, nrow, n_per_row, 4, 8);
+}
+
+size_t quantize_q4_0_8x8(const float * restrict src, void * restrict dst, int64_t nrow, int64_t n_per_row, const float * quant_weights) {
+    UNUSED(quant_weights);
+    return quantize_q4_0_nr_bl(src, dst, nrow, n_per_row, 8, 8);
+}
+
+void ggml_gemv_q4_0_4x4_q8_0(int n, float * restrict s, size_t bs, const void * restrict vx, const void * restrict vy, int nr, int nc) {
+    const int qk = QK8_0;
+    const int nb = n / qk;
+    const int ncols_interleaved = 4;
+    const int blocklen = 4;
+
+    assert (n % qk == 0);
+    assert (nc % ncols_interleaved == 0);
+
+    UNUSED(s);
+    UNUSED(bs);
+    UNUSED(vx);
+    UNUSED(vy);
+    UNUSED(nr);
+    UNUSED(nc);
+    UNUSED(nb);
+    UNUSED(ncols_interleaved);
+    UNUSED(blocklen);
+
+#if ! ((defined(_MSC_VER)) && ! defined(__clang__)) && defined(__aarch64__) && defined(__ARM_NEON)
+    if (ggml_cpu_has_neon()) {
+        const void * b_ptr = vx;
+        const void * a_ptr = vy;
+        float * res_ptr = s;
+
+        __asm__ __volatile__(
+            "movi v31.16b, #0x4\n"
+            "movi v30.16b, #0xf0\n"
+            "add %x[b_ptr], %x[b_ptr], #0x8\n"
+            "1:"  // Column loop
+            "add x22, %x[a_ptr], #0x2\n"
+            "movi v29.16b, #0x0\n"
+            "mov x21, %x[nb]\n"
+            "2:"  // Block loop
+            "ldr q28, [%x[b_ptr], #0x0]\n"
+            "ldr q27, [x22, #0x0]\n"
+            "movi v26.4s, #0x0\n"
+            "sub x20, x22, #0x2\n"
+            "ldr q25, [x22, #0x10]\n"
+            "ldr q24, [%x[b_ptr], #0x10]\n"
+            "sub x21, x21, #0x1\n"
+            "add x22, x22, #0x22\n"
+            "ldr q23, [%x[b_ptr], #0x20]\n"
+            "ldr q22, [%x[b_ptr], #0x30]\n"
+            "ld1r { v21.8h }, [x20]\n"
+            "ldr q20, [%x[b_ptr], #-0x8]\n"
+            "sshl v16.16b, v28.16b, v31.16b\n"
+            "and v28.16b, v28.16b, v30.16b\n"
+            "sshl v19.16b, v24.16b, v31.16b\n"
+            "and v24.16b, v24.16b, v30.16b\n"
+            "add %x[b_ptr], %x[b_ptr], #0x48\n"
+            "sshl v18.16b, v23.16b, v31.16b\n"
+            "and v23.16b, v23.16b, v30.16b\n"
+            ".inst 0x4f9be21a  // sdot v26.4s, v16.16b, v27.4b[0]\n"
+            "sshl v17.16b, v22.16b, v31.16b\n"
+            "and v22.16b, v22.16b, v30.16b\n"
+            "fcvtl v21.4s, v21.4h\n"
+            "fcvtl v16.4s, v20.4h\n"
+            ".inst 0x4f99e39a  // sdot v26.4s, v28.16b, v25.4b[0]\n"
+            "fmul v16.4s, v16.4s, v21.4s\n"
+            ".inst 0x4fbbe27a  // sdot v26.4s, v19.16b, v27.4b[1]\n"
+            ".inst 0x4fb9e31a  // sdot v26.4s, v24.16b, v25.4b[1]\n"
+            ".inst 0x4f9bea5a  // sdot v26.4s, v18.16b, v27.4b[2]\n"
+            ".inst 0x4f99eafa  // sdot v26.4s, v23.16b, v25.4b[2]\n"
+            ".inst 0x4fbbea3a  // sdot v26.4s, v17.16b, v27.4b[3]\n"
+            ".inst 0x4fb9eada  // sdot v26.4s, v22.16b, v25.4b[3]\n"
+            "scvtf v26.4s, v26.4s, #0x4\n"
+            "fmla v29.4s, v26.4s, v16.4s\n"
+            "cbnz x21, 2b\n"
+            "sub %x[nc], %x[nc], #0x4\n"
+            "str q29, [%x[res_ptr], #0x0]\n"
+            "add %x[res_ptr], %x[res_ptr], #0x10\n"
+            "cbnz %x[nc], 1b\n"
+            : [b_ptr] "+&r" (b_ptr), [res_ptr] "+&r" (res_ptr), [nc] "+&r" (nc)
+            : [a_ptr] "r" (a_ptr), [nb] "r" (nb)
+            : "memory", "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30", "v31", "x20", "x21", "x22"
+            );
+        return;
+    }
+#endif // #if ! ((defined(_MSC_VER)) && ! defined(__clang__)) && defined(__aarch64__) && defined(__ARM_NEON)
+    float sumf[4];
+    int sumi;
+
+    const block_q8_0 * a_ptr = (const block_q8_0 *) vy;
+    for (int x = 0; x < nc / ncols_interleaved; x++) {
+        const block_q4_0x4 * b_ptr = (const block_q4_0x4 *) vx + (x * nb);
+
+        for (int j = 0; j < ncols_interleaved; j++) sumf[j] = 0.0;
+        for (int l = 0; l < nb; l++) {
+            for (int k = 0; k < (qk / (2 * blocklen)); k++) {
+                for (int j = 0; j < ncols_interleaved; j++) {
+                    sumi = 0;
+                    for (int i = 0; i < blocklen; ++i) {
+                        const int v0 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] << 4);
+                        const int v1 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] & 0xF0);
+                        sumi += ((v0 * a_ptr[l].qs[k * blocklen + i]) + (v1 * a_ptr[l].qs[k * blocklen + i + qk / 2])) >> 4;
+                    }
+                    sumf[j] += sumi * GGML_FP16_TO_FP32(b_ptr[l].d[j]) * GGML_FP16_TO_FP32(a_ptr[l].d);
+                }
+            }
+        }
+        for (int j = 0; j < ncols_interleaved; j++) s[x * ncols_interleaved + j] = sumf[j];
+    }
+}
+
+void ggml_gemv_q4_0_4x8_q8_0(int n, float * restrict s, size_t bs, const void * restrict vx, const void * restrict vy, int nr, int nc) {
+    const int qk = QK8_0;
+    const int nb = n / qk;
+    const int ncols_interleaved = 4;
+    const int blocklen = 8;
+
+    assert (n % qk == 0);
+    assert (nc % ncols_interleaved == 0);
+
+    UNUSED(s);
+    UNUSED(bs);
+    UNUSED(vx);
+    UNUSED(vy);
+    UNUSED(nr);
+    UNUSED(nc);
+    UNUSED(nb);
+    UNUSED(ncols_interleaved);
+    UNUSED(blocklen);
+
+#if ! ((defined(_MSC_VER)) && ! defined(__clang__)) && defined(__aarch64__) && defined(__ARM_NEON) && defined(__ARM_FEATURE_MATMUL_INT8)
+    if (ggml_cpu_has_neon() && ggml_cpu_has_matmul_int8()) {
+        const void * b_ptr = vx;
+        const void * a_ptr = vy;
+        float * res_ptr = s;
+
+        __asm__ __volatile__(
+            "movi v2.16b, #0x4\n"
+            "movi v1.16b, #0xf0\n"
+            "add %x[b_ptr], %x[b_ptr], #0x8\n"
+            "1:"  // Column loop
+            "add x23, %x[a_ptr], #0x2\n"
+            "movi v0.16b, #0x0\n"
+            "mov x22, %x[nb]\n"
+            "2:"  // Block loop
+            "ldr q31, [%x[b_ptr], #0x0]\n"
+            "ldr q30, [%x[b_ptr], #0x10]\n"
+            "mov x21, x23\n"
+            "movi v29.4s, #0x0\n"
+            "ldr q28, [%x[b_ptr], #0x20]\n"
+            "ldr q27, [%x[b_ptr], #0x30]\n"
+            "movi v26.4s, #0x0\n"
+            "sub x20, x23, #0x2\n"
+            "ld1r { v25.8h }, [x20]\n"
+            "ldr q24, [%x[b_ptr], #-0x8]\n"
+            "sub x22, x22, #0x1\n"
+            "add x23, x23, #0x22\n"
+            "ld1r { v23.2d }, [x21], #0x8\n"
+            "sshl v22.16b, v31.16b, v2.16b\n"
+            "sshl v16.16b, v30.16b, v2.16b\n"
+            "add %x[b_ptr], %x[b_ptr], #0x48\n"
+            "ld1r { v21.2d }, [x21], #0x8\n"
+            "sshl v20.16b, v28.16b, v2.16b\n"
+            "sshl v19.16b, v27.16b, v2.16b\n"
+            "ld1r { v18.2d }, [x21], #0x8\n"
+            "ld1r { v17.2d }, [x21], #0x8\n"
+            "and v31.16b, v31.16b, v1.16b\n"
+            "and v30.16b, v30.16b, v1.16b\n"
+            ".inst 0x4e9796dd  // sdot v29.4s, v22.16b, v23.16b\n"
+            ".inst 0x4e97961a  // sdot v26.4s, v16.16b, v23.16b\n"
+            "and v28.16b, v28.16b, v1.16b\n"
+            "and v27.16b, v27.16b, v1.16b\n"
+            "fcvtl v25.4s, v25.4h\n"
+            "fcvtl v16.4s, v24.4h\n"
+            ".inst 0x4e95969d  // sdot v29.4s, v20.16b, v21.16b\n"
+            ".inst 0x4e95967a  // sdot v26.4s, v19.16b, v21.16b\n"
+            "fmul v16.4s, v16.4s, v25.4s\n"
+            ".inst 0x4e9297fd  // sdot v29.4s, v31.16b, v18.16b\n"
+            ".inst 0x4e9297da  // sdot v26.4s, v30.16b, v18.16b\n"
+            ".inst 0x4e91979d  // sdot v29.4s, v28.16b, v17.16b\n"
+            ".inst 0x4e91977a  // sdot v26.4s, v27.16b, v17.16b\n"
+            "addp v29.4s, v29.4s, v26.4s\n"
+            "scvtf v29.4s, v29.4s, #0x4\n"
+            "fmla v0.4s, v29.4s, v16.4s\n"
+            "cbnz x22, 2b\n"
+            "sub %x[nc], %x[nc], #0x4\n"
+            "str q0, [%x[res_ptr], #0x0]\n"
+            "add %x[res_ptr], %x[res_ptr], #0x10\n"
+            "cbnz %x[nc], 1b\n"
+            : [b_ptr] "+&r" (b_ptr), [res_ptr] "+&r" (res_ptr), [nc] "+&r" (nc)
+            : [a_ptr] "r" (a_ptr), [nb] "r" (nb)
+            : "memory", "v0", "v1", "v2", "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30", "v31", "x20", "x21", "x22", "x23"
+        );
+        return;
+    }
+#endif // #if ! ((defined(_MSC_VER)) && ! defined(__clang__)) && defined(__aarch64__) && defined(__ARM_NEON) && defined(__ARM_FEATURE_MATMUL_INT8)
+    float sumf[4];
+    int sumi;
+
+    const block_q8_0 * a_ptr = (const block_q8_0 *) vy;
+    for (int x = 0; x < nc / ncols_interleaved; x++) {
+        const block_q4_0x4 * b_ptr = (const block_q4_0x4 *) vx + (x * nb);
+
+        for (int j = 0; j < ncols_interleaved; j++) sumf[j] = 0.0;
+        for (int l = 0; l < nb; l++) {
+            for (int k = 0; k < (qk / (2 * blocklen)); k++) {
+                for (int j = 0; j < ncols_interleaved; j++) {
+                    sumi = 0;
+                    for (int i = 0; i < blocklen; ++i) {
+                        const int v0 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] << 4);
+                        const int v1 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] & 0xF0);
+                        sumi += ((v0 * a_ptr[l].qs[k * blocklen + i]) + (v1 * a_ptr[l].qs[k * blocklen + i + qk / 2])) >> 4;
+                    }
+                    sumf[j] += sumi * GGML_FP16_TO_FP32(b_ptr[l].d[j]) * GGML_FP16_TO_FP32(a_ptr[l].d);
+                }
+            }
+        }
+        for (int j = 0; j < ncols_interleaved; j++) s[x * ncols_interleaved + j] = sumf[j];
+    }
+}
+
+void ggml_gemv_q4_0_8x8_q8_0(int n, float * restrict s, size_t bs, const void * restrict vx, const void * restrict vy, int nr, int nc) {
+    const int qk = QK8_0;
+    const int nb = n / qk;
+    const int ncols_interleaved = 8;
+    const int blocklen = 8;
+
+    assert (n % qk == 0);
+    assert (nc % ncols_interleaved == 0);
+
+    UNUSED(s);
+    UNUSED(bs);
+    UNUSED(vx);
+    UNUSED(vy);
+    UNUSED(nr);
+    UNUSED(nc);
+    UNUSED(nb);
+    UNUSED(ncols_interleaved);
+    UNUSED(blocklen);
+
+#if ! ((defined(_MSC_VER)) && ! defined(__clang__)) && defined(__aarch64__)
+#if defined(__ARM_FEATURE_SVE)
+    if (ggml_cpu_has_sve() && ggml_cpu_get_sve_cnt() == QK8_0) {
+        const void * b_ptr = vx;
+        const void * a_ptr = vy;
+        float * res_ptr = s;
+
+        __asm__ __volatile__(
+            "ptrue p0.b\n"
+            "add %x[b_ptr], %x[b_ptr], #0x10\n"
+            "1:"  // Column loop
+            "add x22, %x[a_ptr], #0x2\n"
+            "mov z31.b, #0x0\n"
+            "mov x21, %x[nb]\n"
+            "2:"  // Block loop
+            "ld1b { z30.b }, p0/Z, [%x[b_ptr]]\n"
+            "ld1b { z29.b }, p0/Z, [%x[b_ptr], #1, MUL VL]\n"
+            "mov z28.s, #0x0\n"
+            "mov z27.s, #0x0\n"
+            "ld1rd { z26.d }, p0/Z, [x22]\n"
+            "ld1b { z25.b }, p0/Z, [%x[b_ptr], #2, MUL VL]\n"
+            "sub x20, x22, #0x2\n"
+            "sub x21, x21, #0x1\n"
+            "ld1b { z24.b }, p0/Z, [%x[b_ptr], #3, MUL VL]\n"
+            "ld1rd { z23.d }, p0/Z, [x22, #8]\n"
+            "lsl z22.b, z30.b, #0x4\n"
+            "lsl z16.b, z29.b, #0x4\n"
+            "and z30.b, z30.b, #0xf0\n"
+            "and z29.b, z29.b, #0xf0\n"
+            "ld1rd { z21.d }, p0/Z, [x22, #16]\n"
+            "ld1rd { z20.d }, p0/Z, [x22, #24]\n"
+            "lsl z19.b, z25.b, #0x4\n"
+            "and z25.b, z25.b, #0xf0\n"
+            "ld1rh { z17.h }, p0/Z, [x20]\n"
+            "ld1h { z18.s }, p0/Z, [%x[b_ptr], #-1, MUL VL]\n"
+            "sdot z28.s, z22.b, z26.b\n"
+            "sdot z27.s, z16.b, z26.b\n"
+            "lsl z16.b, z24.b, #0x4\n"
+            "add x22, x22, #0x22\n"
+            "and z24.b, z24.b, #0xf0\n"
+            "add %x[b_ptr], %x[b_ptr], #0x90\n"
+            "fcvt z17.s, p0/m, z17.h\n"
+            "fcvt z18.s, p0/m, z18.h\n"
+            "sdot z28.s, z19.b, z23.b\n"
+            "sdot z27.s, z16.b, z23.b\n"
+            "fmul z18.s, z18.s, z17.s\n"
+            "sdot z28.s, z30.b, z21.b\n"
+            "sdot z27.s, z29.b, z21.b\n"
+            "sdot z28.s, z25.b, z20.b\n"
+            "sdot z27.s, z24.b, z20.b\n"
+            "uzp1 z17.s, z28.s, z27.s\n"
+            "uzp2 z16.s, z28.s, z27.s\n"
+            "add z17.s, z17.s, z16.s\n"
+            "asr z17.s, z17.s, #0x4\n"
+            "scvtf z17.s, p0/m, z17.s\n"
+            "fmla z31.s, p0/M, z17.s, z18.s\n"
+            "cbnz x21, 2b\n"
+            "sub %x[nc], %x[nc], #0x8\n"
+            "st1w { z31.s }, p0, [%x[res_ptr]]\n"
+            "add %x[res_ptr], %x[res_ptr], #0x20\n"
+            "cbnz %x[nc], 1b\n"
+            : [b_ptr] "+&r" (b_ptr), [res_ptr] "+&r" (res_ptr), [nc] "+&r" (nc)
+            : [a_ptr] "r" (a_ptr), [nb] "r" (nb)
+            : "memory", "p0", "x20", "x21", "x22", "z16", "z17", "z18", "z19", "z20", "z21", "z22", "z23", "z24", "z25", "z26", "z27", "z28", "z29", "z30", "z31"
+        );
+        return;
+    }
+#endif // #if defined(__ARM_FEATURE_SVE)
+#elif defined(__AVX2__)
+    // Lookup table to convert signed nibbles to signed bytes
+    __m256i signextendlut = _mm256_castsi128_si256(_mm_set_epi8(-1, -2, -3, -4, -5, -6, -7, -8, 7, 6, 5, 4, 3, 2, 1, 0));
+    signextendlut = _mm256_permute2f128_si256(signextendlut, signextendlut, 0);
+    __m128i changemask = _mm_set_epi8(15, 14, 7, 6, 13, 12, 5, 4, 11, 10, 3, 2, 9, 8, 1, 0);
+    __m256i finalpermutemask = _mm256_set_epi32(7, 5, 3, 1, 6, 4, 2, 0);
+
+    // Permute mask used for easier vector processing at later stages
+    const __m256i m4b = _mm256_set1_epi8(0x0F);
+
+    int64_t b_nb = n / QK4_0;
+
+    const block_q4_0x8 * b_ptr_start = (const block_q4_0x8 *)vx;
+    const block_q8_0 * a_ptr_start = (const block_q8_0 *)vy;
+
+    // Process Q8_0 blocks one by one
+    for (int64_t y = 0; y < nr; y++) {
+
+        // Pointers to LHS blocks of block_q8_0 format
+        const block_q8_0 * a_ptr = a_ptr_start + (y * nb);
+
+        // Take group of eight block_q4_0x8 structures at each pass of the loop and perform dot product operation
+        for (int64_t x = 0; x < nc / 8; x++) {
+
+            // Pointers to RHS blocks
+            const block_q4_0x8 * b_ptr = b_ptr_start + (x * b_nb);
+
+            // Master FP accumulator
+            __m256 acc_row = _mm256_setzero_ps();
+
+            for (int64_t b = 0; b < nb; b++) {
+                // Load 8 blocks of Q4_0 interleaved as 8 bytes (B0 - B7)
+                const __m256i rhs_raw_vec_0123_0 = _mm256_loadu_si256((const __m256i *)(b_ptr[b].qs));
+                const __m256i rhs_raw_vec_4567_0 = _mm256_loadu_si256((const __m256i *)(b_ptr[b].qs) + 1);
+                const __m256i rhs_raw_vec_0123_1 = _mm256_loadu_si256((const __m256i *)(b_ptr[b].qs) + 2);
+                const __m256i rhs_raw_vec_4567_1 = _mm256_loadu_si256((const __m256i *)(b_ptr[b].qs) + 3);
+
+                // 4-bit -> 8-bit - Sign is maintained
+                const __m256i rhs_vec_0123_0 = _mm256_shuffle_epi8(signextendlut, _mm256_and_si256(rhs_raw_vec_0123_0, m4b)); // B0(0-7) B1(0-7) B2(0-7) B3(0-7)
+                const __m256i rhs_vec_4567_0 = _mm256_shuffle_epi8(signextendlut, _mm256_and_si256(rhs_raw_vec_4567_0, m4b)); // B4(0-7) B5(0-7) B6(0-7) B7(0-7)
+                const __m256i rhs_vec_0123_1 = _mm256_shuffle_epi8(signextendlut, _mm256_and_si256(rhs_raw_vec_0123_1, m4b)); // B0(8-15) B1(8-15) B2(8-15) B3(8-15)
+                const __m256i rhs_vec_4567_1 = _mm256_shuffle_epi8(signextendlut, _mm256_and_si256(rhs_raw_vec_4567_1, m4b)); // B0(8-15) B1(8-15) B2(8-15) B3(8-15)
+
+                const __m256i rhs_vec_0123_2 = _mm256_shuffle_epi8(signextendlut, _mm256_and_si256(_mm256_srli_epi16(rhs_raw_vec_0123_0, 4), m4b)); // B0(16-23) B1(16-23) B2(16-23) B3(16-23)
+                const __m256i rhs_vec_4567_2 = _mm256_shuffle_epi8(signextendlut, _mm256_and_si256(_mm256_srli_epi16(rhs_raw_vec_4567_0, 4), m4b)); // B4(16-23) B5(16-23) B6(16-23) B7(16-23)
+                const __m256i rhs_vec_0123_3 = _mm256_shuffle_epi8(signextendlut, _mm256_and_si256(_mm256_srli_epi16(rhs_raw_vec_0123_1, 4), m4b)); // B0(24-31) B1(24-31) B2(24-31) B3(24-31)
+                const __m256i rhs_vec_4567_3 = _mm256_shuffle_epi8(signextendlut, _mm256_and_si256(_mm256_srli_epi16(rhs_raw_vec_4567_1, 4), m4b)); // B4(24-31) B5(24-31) B6(24-31) B7(24-31)
+
+                // Load the scale values for the 8 blocks interleaved in block_q4_0x8
+                const __m256 col_scale_f32 = GGML_F32Cx8_REARRANGE_LOAD(b_ptr[b].d, changemask);
+
+                // Load and convert to FP32 scale from block_q8_0
+                const __m256 row_scale_f32 = _mm256_set1_ps(GGML_FP16_TO_FP32(a_ptr[b].d));
+
+                // Load the block values in block_q8_0 in batches of 16 bytes and replicate the same across 256 bit vector
+                __m256i lhs_vec_0 = _mm256_castsi128_si256(_mm_loadu_si128((const __m128i *)a_ptr[b].qs));
+                __m256i lhs_vec_1 = _mm256_castsi128_si256(_mm_loadu_si128((const __m128i *)(a_ptr[b].qs + 16)));
+
+                lhs_vec_0 = _mm256_permute2f128_si256(lhs_vec_0, lhs_vec_0, 0); // A0 (0-15) A0(0-15)
+                lhs_vec_1 = _mm256_permute2f128_si256(lhs_vec_1, lhs_vec_1, 0); // A0 (16-31) A0(16-31))
+
+                __m256i iacc = _mm256_setzero_si256();
+
+                // Dot product done within 32 bit lanes and accumulated in the same vector
+                // B0(0-3) B4(0-3) B1(0-3) B5(0-3) B2(0-3) B6(0-3) B3(0-3) B7(0-3) with A0(0-3)
+                // B0(4-7) B4(4-7) B1(4-7) B5(4-7) B2(4-7) B6(4-7) B3(4-7) B7(4-7) with A0(4-7)
+                // ...........................................................................
+                // B0(28-31) B4(28-31) B1(28-31) B5(28-31) B2(28-31) B6(28-31) B3(28-31) B7(28-31) with A0(28-31)
+
+                iacc = _mm256_add_epi32(iacc, mul_sum_i8_pairs_int32x8(_mm256_blend_epi32(rhs_vec_0123_0 ,_mm256_shuffle_epi32(rhs_vec_4567_0, 177), 170), _mm256_shuffle_epi32(lhs_vec_0, 0)));
+                iacc = _mm256_add_epi32(iacc, mul_sum_i8_pairs_int32x8(_mm256_blend_epi32(_mm256_shuffle_epi32(rhs_vec_0123_0, 177) ,rhs_vec_4567_0, 170), _mm256_shuffle_epi32(lhs_vec_0, 85)));
+
+                iacc = _mm256_add_epi32(iacc, mul_sum_i8_pairs_int32x8(_mm256_blend_epi32(rhs_vec_0123_1 ,_mm256_shuffle_epi32(rhs_vec_4567_1, 177), 170), _mm256_shuffle_epi32(lhs_vec_0, 170)));
+                iacc = _mm256_add_epi32(iacc, mul_sum_i8_pairs_int32x8(_mm256_blend_epi32(_mm256_shuffle_epi32(rhs_vec_0123_1, 177) ,rhs_vec_4567_1, 170), _mm256_shuffle_epi32(lhs_vec_0, 255)));
+
+                iacc = _mm256_add_epi32(iacc, mul_sum_i8_pairs_int32x8(_mm256_blend_epi32(rhs_vec_0123_2 ,_mm256_shuffle_epi32(rhs_vec_4567_2, 177), 170), _mm256_shuffle_epi32(lhs_vec_1, 0)));
+                iacc = _mm256_add_epi32(iacc, mul_sum_i8_pairs_int32x8(_mm256_blend_epi32(_mm256_shuffle_epi32(rhs_vec_0123_2, 177) ,rhs_vec_4567_2, 170), _mm256_shuffle_epi32(lhs_vec_1, 85)));
+
+                iacc = _mm256_add_epi32(iacc, mul_sum_i8_pairs_int32x8(_mm256_blend_epi32(rhs_vec_0123_3 ,_mm256_shuffle_epi32(rhs_vec_4567_3, 177), 170), _mm256_shuffle_epi32(lhs_vec_1, 170)));
+                iacc = _mm256_add_epi32(iacc, mul_sum_i8_pairs_int32x8(_mm256_blend_epi32(_mm256_shuffle_epi32(rhs_vec_0123_3, 177) ,rhs_vec_4567_3, 170), _mm256_shuffle_epi32(lhs_vec_1, 255)));
+
+                // Accumulated values multipled with appropriate scales
+                acc_row = _mm256_fmadd_ps(_mm256_cvtepi32_ps(iacc), _mm256_mul_ps(col_scale_f32, row_scale_f32), acc_row);
+            }
+
+            // Accumulated output values permuted so as to be stored in appropriate order post accumulation
+            acc_row = _mm256_permutevar8x32_ps(acc_row, finalpermutemask);
+            _mm256_storeu_ps(s + (y * nr + x * 8), acc_row);
+        }
+    }
+    return;
+#endif // #if ! ((defined(_MSC_VER)) && ! defined(__clang__)) && defined(__aarch64__)
+    {
+        float sumf[8];
+        int sumi;
+
+        const block_q8_0 * a_ptr = (const block_q8_0 *) vy;
+        for (int x = 0; x < nc / ncols_interleaved; x++) {
+            const block_q4_0x8 * b_ptr = (const block_q4_0x8 *) vx + (x * nb);
+
+            for (int j = 0; j < ncols_interleaved; j++) sumf[j] = 0.0;
+            for (int l = 0; l < nb; l++) {
+                for (int k = 0; k < (qk / (2 * blocklen)); k++) {
+                    for (int j = 0; j < ncols_interleaved; j++) {
+                        sumi = 0;
+                        for (int i = 0; i < blocklen; ++i) {
+                            const int v0 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] << 4);
+                            const int v1 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] & 0xF0);
+                            sumi += ((v0 * a_ptr[l].qs[k * blocklen + i]) + (v1 * a_ptr[l].qs[k * blocklen + i + qk / 2])) >> 4;
+                        }
+                        sumf[j] += sumi * GGML_FP16_TO_FP32(b_ptr[l].d[j]) * GGML_FP16_TO_FP32(a_ptr[l].d);
+                    }
+                }
+            }
+            for (int j = 0; j < ncols_interleaved; j++) s[x * ncols_interleaved + j] = sumf[j];
+        }
+    }
+}
+
+void ggml_gemm_q4_0_4x4_q8_0(int n, float * restrict s, size_t bs, const void * restrict vx, const void * restrict vy, int nr, int nc) {
+    const int qk = QK8_0;
+    const int nb = n / qk;
+    const int ncols_interleaved = 4;
+    const int blocklen = 4;
+
+    assert (n % qk == 0);
+    assert (nr % 4 == 0);
+    assert (nc % ncols_interleaved == 0);
+
+    UNUSED(s);
+    UNUSED(bs);
+    UNUSED(vx);
+    UNUSED(vy);
+    UNUSED(nr);
+    UNUSED(nc);
+    UNUSED(nb);
+    UNUSED(ncols_interleaved);
+    UNUSED(blocklen);
+
+#if ! ((defined(_MSC_VER)) && ! defined(__clang__)) && defined(__aarch64__) && defined(__ARM_NEON)
+    if (ggml_cpu_has_neon()) {
+        const void * b_ptr = vx;
+        const void * a_ptr = vy;
+        float * res_ptr = s;
+        size_t res_stride = bs * sizeof(float);
+
+        __asm__ __volatile__(
+            "mov x10, %x[nr]\n"
+            "mov x9, #0x88\n"
+            "cmp x10, #0x10\n"
+            "mul x9, %x[nb], x9\n"
+            "blt 4f\n"
+            "1:"  // Row loop
+            "add x28, %x[b_ptr], #0x8\n"
+            "mov x27, %x[nc]\n"
+            "add x26, %x[res_ptr], %x[res_stride], LSL #4\n"
+            "2:"  // Column loop
+            "add x25, %x[a_ptr], #0x8\n"
+            "movi v15.16b, #0x0\n"
+            "movi v19.16b, #0x0\n"
+            "mov x24, %x[nb]\n"
+            "add x23, x25, x9\n"
+            "movi v18.16b, #0x0\n"
+            "movi v14.16b, #0x0\n"
+            "add x22, x23, x9\n"
+            "movi v11.16b, #0x0\n"
+            "movi v13.16b, #0x0\n"
+            "add x21, x22, x9\n"
+            "movi v23.16b, #0x0\n"
+            "movi v16.16b, #0x0\n"
+            "movi v25.16b, #0x0\n"
+            "movi v7.16b, #0x0\n"
+            "movi v0.16b, #0x0\n"
+            "movi v4.16b, #0x0\n"
+            "movi v5.16b, #0x0\n"
+            "movi v21.16b, #0x0\n"
+            "movi v8.16b, #0x0\n"
+            "movi v1.16b, #0x0\n"
+            "3:"  // Block loop
+            "ldr q3, [x28, #0x0]\n"
+            "ldr q31, [x25, #0x0]\n"
+            "movi v28.16b, #0x4\n"
+            "movi v10.4s, #0x0\n"
+            "ldr q22, [x28, #0x10]\n"
+            "ldr q6, [x25, #0x10]\n"
+            "movi v29.4s, #0x0\n"
+            "movi v9.4s, #0x0\n"
+            "ldr q27, [x28, #0x20]\n"
+            "ldr q30, [x28, #0x30]\n"
+            "movi v20.4s, #0x0\n"
+            "movi v24.16b, #0xf0\n"
+            "ldr d2, [x25, #-0x8]\n"
+            "ldr d26, [x23, #-0x8]\n"
+            "sshl v12.16b, v3.16b, v28.16b\n"
+            "sub x20, x28, #0x8\n"
+            "ldr d17, [x20, #0x0]\n"
+            "and v3.16b, v3.16b, v24.16b\n"
+            "subs x24, x24, #0x1\n"
+            "add x28, x28, #0x48\n"
+            ".inst 0x4f9fe18a  // sdot v10.4s, v12.16b, v31.4b[0]\n"
+            ".inst 0x4fbfe19d  // sdot v29.4s, v12.16b, v31.4b[1]\n"
+            ".inst 0x4f9fe989  // sdot v9.4s, v12.16b, v31.4b[2]\n"
+            ".inst 0x4fbfe994  // sdot v20.4s, v12.16b, v31.4b[3]\n"
+            "sshl v31.16b, v22.16b, v28.16b\n"
+            "and v22.16b, v22.16b, v24.16b\n"
+            "fcvtl v17.4s, v17.4h\n"
+            "fcvtl v2.4s, v2.4h\n"
+            "fcvtl v26.4s, v26.4h\n"
+            ".inst 0x4f86e3ea  // sdot v10.4s, v31.16b, v6.4b[0]\n"
+            ".inst 0x4fa6e3fd  // sdot v29.4s, v31.16b, v6.4b[1]\n"
+            ".inst 0x4f86ebe9  // sdot v9.4s, v31.16b, v6.4b[2]\n"
+            ".inst 0x4fa6ebf4  // sdot v20.4s, v31.16b, v6.4b[3]\n"
+            "sshl v6.16b, v27.16b, v28.16b\n"
+            "sshl v28.16b, v30.16b, v28.16b\n"
+            "and v27.16b, v27.16b, v24.16b\n"
+            "and v30.16b, v30.16b, v24.16b\n"
+            "ldr q24, [x25, #0x20]\n"
+            ".inst 0x4f98e0ca  // sdot v10.4s, v6.16b, v24.4b[0]\n"
+            ".inst 0x4fb8e0dd  // sdot v29.4s, v6.16b, v24.4b[1]\n"
+            ".inst 0x4f98e8c9  // sdot v9.4s, v6.16b, v24.4b[2]\n"
+            ".inst 0x4fb8e8d4  // sdot v20.4s, v6.16b, v24.4b[3]\n"
+            "ldr q24, [x25, #0x30]\n"
+            ".inst 0x4f98e38a  // sdot v10.4s, v28.16b, v24.4b[0]\n"
+            ".inst 0x4fb8e39d  // sdot v29.4s, v28.16b, v24.4b[1]\n"
+            ".inst 0x4f98eb89  // sdot v9.4s, v28.16b, v24.4b[2]\n"
+            ".inst 0x4fb8eb94  // sdot v20.4s, v28.16b, v24.4b[3]\n"
+            "ldr q24, [x25, #0x40]\n"
+            ".inst 0x4f98e06a  // sdot v10.4s, v3.16b, v24.4b[0]\n"
+            ".inst 0x4fb8e07d  // sdot v29.4s, v3.16b, v24.4b[1]\n"
+            ".inst 0x4f98e869  // sdot v9.4s, v3.16b, v24.4b[2]\n"
+            ".inst 0x4fb8e874  // sdot v20.4s, v3.16b, v24.4b[3]\n"
+            "ldr q24, [x25, #0x50]\n"
+            ".inst 0x4f98e2ca  // sdot v10.4s, v22.16b, v24.4b[0]\n"
+            ".inst 0x4fb8e2dd  // sdot v29.4s, v22.16b, v24.4b[1]\n"
+            ".inst 0x4f98eac9  // sdot v9.4s, v22.16b, v24.4b[2]\n"
+            ".inst 0x4fb8ead4  // sdot v20.4s, v22.16b, v24.4b[3]\n"
+            "ldr q24, [x25, #0x60]\n"
+            ".inst 0x4f98e36a  // sdot v10.4s, v27.16b, v24.4b[0]\n"
+            ".inst 0x4fb8e37d  // sdot v29.4s, v27.16b, v24.4b[1]\n"
+            ".inst 0x4f98eb69  // sdot v9.4s, v27.16b, v24.4b[2]\n"
+            ".inst 0x4fb8eb74  // sdot v20.4s, v27.16b, v24.4b[3]\n"
+            "ldr q24, [x25, #0x70]\n"
+            "add x25, x25, #0x88\n"
+            ".inst 0x4f98e3ca  // sdot v10.4s, v30.16b, v24.4b[0]\n"
+            ".inst 0x4fb8e3dd  // sdot v29.4s, v30.16b, v24.4b[1]\n"
+            ".inst 0x4f98ebc9  // sdot v9.4s, v30.16b, v24.4b[2]\n"
+            ".inst 0x4fb8ebd4  // sdot v20.4s, v30.16b, v24.4b[3]\n"
+            "fmul v24.4s, v17.4s, v2.s[0]\n"
+            "scvtf v10.4s, v10.4s, #0x4\n"
+            "scvtf v29.4s, v29.4s, #0x4\n"
+            "scvtf v9.4s, v9.4s, #0x4\n"
+            "scvtf v20.4s, v20.4s, #0x4\n"
+            "fmla v15.4s, v10.4s, v24.4s\n"
+            "ldr q24, [x23, #0x0]\n"
+            "fmul v10.4s, v17.4s, v2.s[1]\n"
+            "fmla v19.4s, v29.4s, v10.4s\n"
+            "ldr q10, [x23, #0x10]\n"
+            "fmul v29.4s, v17.4s, v2.s[2]\n"
+            "fmul v2.4s, v17.4s, v2.s[3]\n"
+            "fmla v18.4s, v9.4s, v29.4s\n"
+            "movi v9.4s, #0x0\n"
+            "movi v29.4s, #0x0\n"
+            ".inst 0x4f98e189  // sdot v9.4s, v12.16b, v24.4b[0]\n"
+            ".inst 0x4fb8e19d  // sdot v29.4s, v12.16b, v24.4b[1]\n"
+            "fmla v14.4s, v20.4s, v2.4s\n"
+            "movi v20.4s, #0x0\n"
+            "movi v2.4s, #0x0\n"
+            ".inst 0x4f98e994  // sdot v20.4s, v12.16b, v24.4b[2]\n"
+            ".inst 0x4fb8e982  // sdot v2.4s, v12.16b, v24.4b[3]\n"
+            "ldr q24, [x23, #0x20]\n"
+            ".inst 0x4f8ae3e9  // sdot v9.4s, v31.16b, v10.4b[0]\n"
+            ".inst 0x4faae3fd  // sdot v29.4s, v31.16b, v10.4b[1]\n"
+            ".inst 0x4f8aebf4  // sdot v20.4s, v31.16b, v10.4b[2]\n"
+            ".inst 0x4faaebe2  // sdot v2.4s, v31.16b, v10.4b[3]\n"
+            "ldr q10, [x23, #0x30]\n"
+            ".inst 0x4f98e0c9  // sdot v9.4s, v6.16b, v24.4b[0]\n"
+            ".inst 0x4fb8e0dd  // sdot v29.4s, v6.16b, v24.4b[1]\n"
+            ".inst 0x4f98e8d4  // sdot v20.4s, v6.16b, v24.4b[2]\n"
+            ".inst 0x4fb8e8c2  // sdot v2.4s, v6.16b, v24.4b[3]\n"
+            "ldr q24, [x23, #0x40]\n"
+            ".inst 0x4f8ae389  // sdot v9.4s, v28.16b, v10.4b[0]\n"
+            ".inst 0x4faae39d  // sdot v29.4s, v28.16b, v10.4b[1]\n"
+            ".inst 0x4f8aeb94  // sdot v20.4s, v28.16b, v10.4b[2]\n"
+            ".inst 0x4faaeb82  // sdot v2.4s, v28.16b, v10.4b[3]\n"
+            "ldr q10, [x23, #0x50]\n"
+            ".inst 0x4f98e069  // sdot v9.4s, v3.16b, v24.4b[0]\n"
+            ".inst 0x4fb8e07d  // sdot v29.4s, v3.16b, v24.4b[1]\n"
+            ".inst 0x4f98e874  // sdot v20.4s, v3.16b, v24.4b[2]\n"
+            ".inst 0x4fb8e862  // sdot v2.4s, v3.16b, v24.4b[3]\n"
+            "ldr q24, [x23, #0x60]\n"
+            ".inst 0x4f8ae2c9  // sdot v9.4s, v22.16b, v10.4b[0]\n"
+            ".inst 0x4faae2dd  // sdot v29.4s, v22.16b, v10.4b[1]\n"
+            ".inst 0x4f8aead4  // sdot v20.4s, v22.16b, v10.4b[2]\n"
+            ".inst 0x4faaeac2  // sdot v2.4s, v22.16b, v10.4b[3]\n"
+            "ldr q10, [x23, #0x70]\n"
+            "add x23, x23, #0x88\n"
+            ".inst 0x4f98e369  // sdot v9.4s, v27.16b, v24.4b[0]\n"
+            ".inst 0x4fb8e37d  // sdot v29.4s, v27.16b, v24.4b[1]\n"
+            ".inst 0x4f98eb74  // sdot v20.4s, v27.16b, v24.4b[2]\n"
+            ".inst 0x4fb8eb62  // sdot v2.4s, v27.16b, v24.4b[3]\n"
+            "ldr q24, [x22, #0x0]\n"
+            ".inst 0x4f8ae3c9  // sdot v9.4s, v30.16b, v10.4b[0]\n"
+            ".inst 0x4faae3dd  // sdot v29.4s, v30.16b, v10.4b[1]\n"
+            ".inst 0x4f8aebd4  // sdot v20.4s, v30.16b, v10.4b[2]\n"
+            ".inst 0x4faaebc2  // sdot v2.4s, v30.16b, v10.4b[3]\n"
+            "fmul v10.4s, v17.4s, v26.s[0]\n"
+            "scvtf v9.4s, v9.4s, #0x4\n"
+            "scvtf v29.4s, v29.4s, #0x4\n"
+            "scvtf v20.4s, v20.4s, #0x4\n"
+            "scvtf v2.4s, v2.4s, #0x4\n"
+            "fmla v11.4s, v9.4s, v10.4s\n"
+            "ldr q9, [x22, #0x10]\n"
+            "fmul v10.4s, v17.4s, v26.s[1]\n"
+            "fmla v13.4s, v29.4s, v10.4s\n"
+            "ldr d29, [x22, #-0x8]\n"
+            "fmul v10.4s, v17.4s, v26.s[2]\n"
+            "fmul v26.4s, v17.4s, v26.s[3]\n"
+            "fcvtl v29.4s, v29.4h\n"
+            "fmla v23.4s, v20.4s, v10.4s\n"
+            "movi v20.4s, #0x0\n"
+            "movi v10.4s, #0x0\n"
+            "fmla v16.4s, v2.4s, v26.4s\n"
+            "movi v26.4s, #0x0\n"
+            "movi v2.4s, #0x0\n"
+            ".inst 0x4f98e194  // sdot v20.4s, v12.16b, v24.4b[0]\n"
+            ".inst 0x4fb8e18a  // sdot v10.4s, v12.16b, v24.4b[1]\n"
+            ".inst 0x4f98e99a  // sdot v26.4s, v12.16b, v24.4b[2]\n"
+            ".inst 0x4fb8e982  // sdot v2.4s, v12.16b, v24.4b[3]\n"
+            "ldr q24, [x22, #0x20]\n"
+            ".inst 0x4f89e3f4  // sdot v20.4s, v31.16b, v9.4b[0]\n"
+            ".inst 0x4fa9e3ea  // sdot v10.4s, v31.16b, v9.4b[1]\n"
+            ".inst 0x4f89ebfa  // sdot v26.4s, v31.16b, v9.4b[2]\n"
+            ".inst 0x4fa9ebe2  // sdot v2.4s, v31.16b, v9.4b[3]\n"
+            "ldr q9, [x22, #0x30]\n"
+            ".inst 0x4f98e0d4  // sdot v20.4s, v6.16b, v24.4b[0]\n"
+            ".inst 0x4fb8e0ca  // sdot v10.4s, v6.16b, v24.4b[1]\n"
+            ".inst 0x4f98e8da  // sdot v26.4s, v6.16b, v24.4b[2]\n"
+            ".inst 0x4fb8e8c2  // sdot v2.4s, v6.16b, v24.4b[3]\n"
+            "ldr q24, [x22, #0x40]\n"
+            ".inst 0x4f89e394  // sdot v20.4s, v28.16b, v9.4b[0]\n"
+            ".inst 0x4fa9e38a  // sdot v10.4s, v28.16b, v9.4b[1]\n"
+            ".inst 0x4f89eb9a  // sdot v26.4s, v28.16b, v9.4b[2]\n"
+            ".inst 0x4fa9eb82  // sdot v2.4s, v28.16b, v9.4b[3]\n"
+            "ldr q9, [x22, #0x50]\n"
+            ".inst 0x4f98e074  // sdot v20.4s, v3.16b, v24.4b[0]\n"
+            ".inst 0x4fb8e06a  // sdot v10.4s, v3.16b, v24.4b[1]\n"
+            ".inst 0x4f98e87a  // sdot v26.4s, v3.16b, v24.4b[2]\n"
+            ".inst 0x4fb8e862  // sdot v2.4s, v3.16b, v24.4b[3]\n"
+            "ldr q24, [x22, #0x60]\n"
+            ".inst 0x4f89e2d4  // sdot v20.4s, v22.16b, v9.4b[0]\n"
+            ".inst 0x4fa9e2ca  // sdot v10.4s, v22.16b, v9.4b[1]\n"
+            ".inst 0x4f89eada  // sdot v26.4s, v22.16b, v9.4b[2]\n"
+            ".inst 0x4fa9eac2  // sdot v2.4s, v22.16b, v9.4b[3]\n"
+            "ldr q9, [x22, #0x70]\n"
+            "add x22, x22, #0x88\n"
+            ".inst 0x4f98e374  // sdot v20.4s, v27.16b, v24.4b[0]\n"
+            ".inst 0x4fb8e36a  // sdot v10.4s, v27.16b, v24.4b[1]\n"
+            ".inst 0x4f98eb7a  // sdot v26.4s, v27.16b, v24.4b[2]\n"
+            ".inst 0x4fb8eb62  // sdot v2.4s, v27.16b, v24.4b[3]\n"
+            "ldr q24, [x21, #0x0]\n"
+            ".inst 0x4f89e3d4  // sdot v20.4s, v30.16b, v9.4b[0]\n"
+            ".inst 0x4fa9e3ca  // sdot v10.4s, v30.16b, v9.4b[1]\n"
+            ".inst 0x4f89ebda  // sdot v26.4s, v30.16b, v9.4b[2]\n"
+            ".inst 0x4fa9ebc2  // sdot v2.4s, v30.16b, v9.4b[3]\n"
+            "fmul v9.4s, v17.4s, v29.s[0]\n"
+            "scvtf v20.4s, v20.4s, #0x4\n"
+            "scvtf v10.4s, v10.4s, #0x4\n"
+            "scvtf v26.4s, v26.4s, #0x4\n"
+            "scvtf v2.4s, v2.4s, #0x4\n"
+            "fmla v25.4s, v20.4s, v9.4s\n"
+            "ldr q9, [x21, #0x10]\n"
+            "fmul v20.4s, v17.4s, v29.s[1]\n"
+            "fmla v7.4s, v10.4s, v20.4s\n"
+            "ldr d20, [x21, #-0x8]\n"
+            "fmul v10.4s, v17.4s, v29.s[2]\n"
+            "fmul v29.4s, v17.4s, v29.s[3]\n"
+            "fcvtl v20.4s, v20.4h\n"
+            "fmla v0.4s, v26.4s, v10.4s\n"
+            "movi v26.4s, #0x0\n"
+            "movi v10.4s, #0x0\n"
+            "fmla v4.4s, v2.4s, v29.4s\n"
+            "movi v2.4s, #0x0\n"
+            "movi v29.4s, #0x0\n"
+            ".inst 0x4f98e19a  // sdot v26.4s, v12.16b, v24.4b[0]\n"
+            ".inst 0x4fb8e18a  // sdot v10.4s, v12.16b, v24.4b[1]\n"
+            ".inst 0x4f98e982  // sdot v2.4s, v12.16b, v24.4b[2]\n"
+            ".inst 0x4fb8e99d  // sdot v29.4s, v12.16b, v24.4b[3]\n"
+            "ldr q12, [x21, #0x20]\n"
+            "fmul v24.4s, v17.4s, v20.s[0]\n"
+            ".inst 0x4f89e3fa  // sdot v26.4s, v31.16b, v9.4b[0]\n"
+            ".inst 0x4fa9e3ea  // sdot v10.4s, v31.16b, v9.4b[1]\n"
+            ".inst 0x4f89ebe2  // sdot v2.4s, v31.16b, v9.4b[2]\n"
+            ".inst 0x4fa9ebfd  // sdot v29.4s, v31.16b, v9.4b[3]\n"
+            "ldr q9, [x21, #0x30]\n"
+            "fmul v31.4s, v17.4s, v20.s[1]\n"
+            ".inst 0x4f8ce0da  // sdot v26.4s, v6.16b, v12.4b[0]\n"
+            ".inst 0x4face0ca  // sdot v10.4s, v6.16b, v12.4b[1]\n"
+            ".inst 0x4f8ce8c2  // sdot v2.4s, v6.16b, v12.4b[2]\n"
+            ".inst 0x4face8dd  // sdot v29.4s, v6.16b, v12.4b[3]\n"
+            "ldr q12, [x21, #0x40]\n"
+            "fmul v6.4s, v17.4s, v20.s[2]\n"
+            "fmul v20.4s, v17.4s, v20.s[3]\n"
+            ".inst 0x4f89e39a  // sdot v26.4s, v28.16b, v9.4b[0]\n"
+            ".inst 0x4fa9e38a  // sdot v10.4s, v28.16b, v9.4b[1]\n"
+            ".inst 0x4f89eb82  // sdot v2.4s, v28.16b, v9.4b[2]\n"
+            ".inst 0x4fa9eb9d  // sdot v29.4s, v28.16b, v9.4b[3]\n"
+            "ldr q9, [x21, #0x50]\n"
+            ".inst 0x4f8ce07a  // sdot v26.4s, v3.16b, v12.4b[0]\n"
+            ".inst 0x4face06a  // sdot v10.4s, v3.16b, v12.4b[1]\n"
+            ".inst 0x4f8ce862  // sdot v2.4s, v3.16b, v12.4b[2]\n"
+            ".inst 0x4face87d  // sdot v29.4s, v3.16b, v12.4b[3]\n"
+            "ldr q12, [x21, #0x60]\n"
+            ".inst 0x4f89e2da  // sdot v26.4s, v22.16b, v9.4b[0]\n"
+            ".inst 0x4fa9e2ca  // sdot v10.4s, v22.16b, v9.4b[1]\n"
+            ".inst 0x4f89eac2  // sdot v2.4s, v22.16b, v9.4b[2]\n"
+            ".inst 0x4fa9eadd  // sdot v29.4s, v22.16b, v9.4b[3]\n"
+            "ldr q17, [x21, #0x70]\n"
+            "add x21, x21, #0x88\n"
+            ".inst 0x4f8ce37a  // sdot v26.4s, v27.16b, v12.4b[0]\n"
+            ".inst 0x4face36a  // sdot v10.4s, v27.16b, v12.4b[1]\n"
+            ".inst 0x4f8ceb62  // sdot v2.4s, v27.16b, v12.4b[2]\n"
+            ".inst 0x4faceb7d  // sdot v29.4s, v27.16b, v12.4b[3]\n"
+            ".inst 0x4f91e3da  // sdot v26.4s, v30.16b, v17.4b[0]\n"
+            ".inst 0x4fb1e3ca  // sdot v10.4s, v30.16b, v17.4b[1]\n"
+            ".inst 0x4f91ebc2  // sdot v2.4s, v30.16b, v17.4b[2]\n"
+            ".inst 0x4fb1ebdd  // sdot v29.4s, v30.16b, v17.4b[3]\n"
+            "scvtf v26.4s, v26.4s, #0x4\n"
+            "scvtf v10.4s, v10.4s, #0x4\n"
+            "fmla v5.4s, v26.4s, v24.4s\n"
+            "scvtf v2.4s, v2.4s, #0x4\n"
+            "scvtf v29.4s, v29.4s, #0x4\n"
+            "fmla v21.4s, v10.4s, v31.4s\n"
+            "fmla v8.4s, v2.4s, v6.4s\n"
+            "fmla v1.4s, v29.4s, v20.4s\n"
+            "bgt 3b\n"
+            "mov x20, %x[res_ptr]\n"
+            "subs x27, x27, #0x4\n"
+            "add %x[res_ptr], %x[res_ptr], #0x10\n"
+            "str q15, [x20, #0x0]\n"
+            "add x20, x20, %x[res_stride]\n"
+            "str q19, [x20, #0x0]\n"
+            "add x20, x20, %x[res_stride]\n"
+            "str q18, [x20, #0x0]\n"
+            "add x20, x20, %x[res_stride]\n"
+            "str q14, [x20, #0x0]\n"
+            "add x20, x20, %x[res_stride]\n"
+            "str q11, [x20, #0x0]\n"
+            "add x20, x20, %x[res_stride]\n"
+            "str q13, [x20, #0x0]\n"
+            "add x20, x20, %x[res_stride]\n"
+            "str q23, [x20, #0x0]\n"
+            "add x20, x20, %x[res_stride]\n"
+            "str q16, [x20, #0x0]\n"
+            "add x20, x20, %x[res_stride]\n"
+            "str q25, [x20, #0x0]\n"
+            "add x20, x20, %x[res_stride]\n"
+            "str q7, [x20, #0x0]\n"
+            "add x20, x20, %x[res_stride]\n"
+            "str q0, [x20, #0x0]\n"
+            "add x20, x20, %x[res_stride]\n"
+            "str q4, [x20, #0x0]\n"
+            "add x20, x20, %x[res_stride]\n"
+            "str q5, [x20, #0x0]\n"
+            "add x20, x20, %x[res_stride]\n"
+            "str q21, [x20, #0x0]\n"
+            "add x20, x20, %x[res_stride]\n"
+            "str q8, [x20, #0x0]\n"
+            "add x20, x20, %x[res_stride]\n"
+            "str q1, [x20, #0x0]\n"
+            "bne 2b\n"
+            "mov x20, #0x4\n"
+            "sub x10, x10, #0x10\n"
+            "cmp x10, #0x10\n"
+            "mov %x[res_ptr], x26\n"
+            "madd %x[a_ptr], x20, x9, %x[a_ptr]\n"
+            "bge 1b\n"
+            "4:"  // Row loop skip
+            "cbz x10, 9f\n"
+            "5:"  // Row tail: Row loop
+            "add x24, %x[b_ptr], #0x8\n"
+            "mov x23, %x[nc]\n"
+            "add x22, %x[res_ptr], %x[res_stride], LSL #2\n"
+            "6:"  // Row tail: Column loop
+            "movi v15.16b, #0x0\n"
+            "movi v19.16b, #0x0\n"
+            "add x25, %x[a_ptr], #0x8\n"
+            "mov x21, %x[nb]\n"
+            "movi v18.16b, #0x0\n"
+            "movi v14.16b, #0x0\n"
+            "7:"  // Row tail: Block loop
+            "ldr q7, [x24, #0x0]\n"
+            "ldr q5, [x25, #0x0]\n"
+            "movi v9.16b, #0x4\n"
+            "movi v4.4s, #0x0\n"
+            "ldr q3, [x24, #0x10]\n"
+            "ldr q2, [x25, #0x10]\n"
+            "movi v1.4s, #0x0\n"
+            "movi v0.4s, #0x0\n"
+            "ldr q13, [x24, #0x20]\n"
+            "ldr q31, [x25, #0x20]\n"
+            "movi v30.4s, #0x0\n"
+            "movi v29.16b, #0xf0\n"
+            "ldr q28, [x24, #0x30]\n"
+            "ldr q27, [x25, #0x30]\n"
+            "sshl v20.16b, v7.16b, v9.16b\n"
+            "sub x20, x24, #0x8\n"
+            "ldr q26, [x25, #0x40]\n"
+            "ldr q25, [x25, #0x50]\n"
+            "sshl v17.16b, v3.16b, v9.16b\n"
+            "and v7.16b, v7.16b, v29.16b\n"
+            "ldr q24, [x25, #0x60]\n"
+            "ldr q16, [x25, #0x70]\n"
+            "sshl v22.16b, v13.16b, v9.16b\n"
+            "and v3.16b, v3.16b, v29.16b\n"
+            "ldr d21, [x20, #0x0]\n"
+            "ldr d12, [x25, #-0x8]\n"
+            ".inst 0x4f85e284  // sdot v4.4s, v20.16b, v5.4b[0]\n"
+            ".inst 0x4fa5e281  // sdot v1.4s, v20.16b, v5.4b[1]\n"
+            ".inst 0x4f85ea80  // sdot v0.4s, v20.16b, v5.4b[2]\n"
+            ".inst 0x4fa5ea9e  // sdot v30.4s, v20.16b, v5.4b[3]\n"
+            "sshl v9.16b, v28.16b, v9.16b\n"
+            "subs x21, x21, #0x1\n"
+            "and v13.16b, v13.16b, v29.16b\n"
+            "and v28.16b, v28.16b, v29.16b\n"
+            "add x25, x25, #0x88\n"
+            "add x24, x24, #0x48\n"
+            "fcvtl v21.4s, v21.4h\n"
+            "fcvtl v12.4s, v12.4h\n"
+            ".inst 0x4f82e224  // sdot v4.4s, v17.16b, v2.4b[0]\n"
+            ".inst 0x4fa2e221  // sdot v1.4s, v17.16b, v2.4b[1]\n"
+            ".inst 0x4f82ea20  // sdot v0.4s, v17.16b, v2.4b[2]\n"
+            ".inst 0x4fa2ea3e  // sdot v30.4s, v17.16b, v2.4b[3]\n"
+            "fmul v11.4s, v21.4s, v12.s[0]\n"
+            "fmul v23.4s, v21.4s, v12.s[1]\n"
+            "fmul v17.4s, v21.4s, v12.s[2]\n"
+            ".inst 0x4f9fe2c4  // sdot v4.4s, v22.16b, v31.4b[0]\n"
+            "fmul v6.4s, v21.4s, v12.s[3]\n"
+            ".inst 0x4fbfe2c1  // sdot v1.4s, v22.16b, v31.4b[1]\n"
+            ".inst 0x4f9feac0  // sdot v0.4s, v22.16b, v31.4b[2]\n"
+            ".inst 0x4fbfeade  // sdot v30.4s, v22.16b, v31.4b[3]\n"
+            ".inst 0x4f9be124  // sdot v4.4s, v9.16b, v27.4b[0]\n"
+            ".inst 0x4fbbe121  // sdot v1.4s, v9.16b, v27.4b[1]\n"
+            ".inst 0x4f9be920  // sdot v0.4s, v9.16b, v27.4b[2]\n"
+            ".inst 0x4fbbe93e  // sdot v30.4s, v9.16b, v27.4b[3]\n"
+            ".inst 0x4f9ae0e4  // sdot v4.4s, v7.16b, v26.4b[0]\n"
+            ".inst 0x4fbae0e1  // sdot v1.4s, v7.16b, v26.4b[1]\n"
+            ".inst 0x4f9ae8e0  // sdot v0.4s, v7.16b, v26.4b[2]\n"
+            ".inst 0x4fbae8fe  // sdot v30.4s, v7.16b, v26.4b[3]\n"
+            ".inst 0x4f99e064  // sdot v4.4s, v3.16b, v25.4b[0]\n"
+            ".inst 0x4fb9e061  // sdot v1.4s, v3.16b, v25.4b[1]\n"
+            ".inst 0x4f99e860  // sdot v0.4s, v3.16b, v25.4b[2]\n"
+            ".inst 0x4fb9e87e  // sdot v30.4s, v3.16b, v25.4b[3]\n"
+            ".inst 0x4f98e1a4  // sdot v4.4s, v13.16b, v24.4b[0]\n"
+            ".inst 0x4fb8e1a1  // sdot v1.4s, v13.16b, v24.4b[1]\n"
+            ".inst 0x4f98e9a0  // sdot v0.4s, v13.16b, v24.4b[2]\n"
+            ".inst 0x4fb8e9be  // sdot v30.4s, v13.16b, v24.4b[3]\n"
+            ".inst 0x4f90e384  // sdot v4.4s, v28.16b, v16.4b[0]\n"
+            ".inst 0x4fb0e381  // sdot v1.4s, v28.16b, v16.4b[1]\n"
+            ".inst 0x4f90eb80  // sdot v0.4s, v28.16b, v16.4b[2]\n"
+            ".inst 0x4fb0eb9e  // sdot v30.4s, v28.16b, v16.4b[3]\n"
+            "scvtf v4.4s, v4.4s, #0x4\n"
+            "scvtf v1.4s, v1.4s, #0x4\n"
+            "scvtf v0.4s, v0.4s, #0x4\n"
+            "fmla v15.4s, v4.4s, v11.4s\n"
+            "scvtf v30.4s, v30.4s, #0x4\n"
+            "fmla v19.4s, v1.4s, v23.4s\n"
+            "fmla v18.4s, v0.4s, v17.4s\n"
+            "fmla v14.4s, v30.4s, v6.4s\n"
+            "bgt 7b\n"
+            "mov x20, %x[res_ptr]\n"
+            "cmp x10, #0x1\n"
+            "str q15, [x20, #0x0]\n"
+            "add x20, x20, %x[res_stride]\n"
+            "ble 8f\n"
+            "cmp x10, #0x2\n"
+            "str q19, [x20, #0x0]\n"
+            "add x20, x20, %x[res_stride]\n"
+            "ble 8f\n"
+            "cmp x10, #0x3\n"
+            "str q18, [x20, #0x0]\n"
+            "add x20, x20, %x[res_stride]\n"
+            "ble 8f\n"
+            "str q14, [x20, #0x0]\n"
+            "8:"  // Row tail: Accumulator store skip
+            "subs x23, x23, #0x4\n"
+            "add %x[res_ptr], %x[res_ptr], #0x10\n"
+            "bne 6b\n"
+            "subs x10, x10, #0x4\n"
+            "add %x[a_ptr], %x[a_ptr], x9\n"
+            "mov %x[res_ptr], x22\n"
+            "bgt 5b\n"
+            "9:"  // Row tail: Row loop skip
+            : [a_ptr] "+&r" (a_ptr), [res_ptr] "+&r" (res_ptr)
+            : [b_ptr] "r" (b_ptr), [nr] "r" (nr), [nb] "r" (nb), [res_stride] "r" (res_stride), [nc] "r" (nc)
+            : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v8", "v9", "v10", "v11", "v12", "v13", "v14", "v15", "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30", "v31", "x9", "x10", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "x27", "x28"
+        );
+        return;
+    }
+#endif // #if ! ((defined(_MSC_VER)) && ! defined(__clang__)) && defined(__aarch64__) && defined(__ARM_NEON)
+    {
+        float sumf[4][4];
+        int sumi;
+
+        for (int y = 0; y < nr / 4; y++) {
+            const block_q8_0x4 * a_ptr = (const block_q8_0x4 *) vy + (y * nb);
+            for (int x = 0; x < nc / ncols_interleaved; x++) {
+                const block_q4_0x4 * b_ptr = (const block_q4_0x4 *) vx + (x * nb);
+                for (int m = 0; m < 4; m++) {
+                    for (int j = 0; j < ncols_interleaved; j++) sumf[m][j] = 0.0;
+                }
+                for (int l = 0; l < nb; l++) {
+                    for (int k = 0; k < (qk / (2 * blocklen)); k++) {
+                        for (int m = 0; m < 4; m++) {
+                            for (int j = 0; j < ncols_interleaved; j++) {
+                                sumi = 0;
+                                for (int i = 0; i < blocklen; ++i) {
+                                    const int v0 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] << 4);
+                                    const int v1 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] & 0xF0);
+                                    sumi += ((v0 * a_ptr[l].qs[k * 4 * blocklen + m * blocklen + i]) +
+                                            (v1 * a_ptr[l].qs[k * 4 * blocklen + m * blocklen + i + qk / 2 * 4])) >> 4;
+                                }
+                                sumf[m][j] += sumi * GGML_FP16_TO_FP32(b_ptr[l].d[j]) * GGML_FP16_TO_FP32(a_ptr[l].d[m]);
+                            }
+                        }
+                    }
+                }
+                for (int m = 0; m < 4; m++) {
+                    for (int j = 0; j < ncols_interleaved; j++)
+                        s[(y * 4 + m) * bs + x * ncols_interleaved + j] = sumf[m][j];
+                }
+            }
+        }
+    }
+}
+
+void ggml_gemm_q4_0_4x8_q8_0(int n, float * restrict s, size_t bs, const void * restrict vx, const void * restrict vy, int nr, int nc) {
+    const int qk = QK8_0;
+    const int nb = n / qk;
+    const int ncols_interleaved = 4;
+    const int blocklen = 8;
+
+    assert (n % qk == 0);
+    assert (nr % 4 == 0);
+    assert (nc % ncols_interleaved == 0);
+
+    UNUSED(s);
+    UNUSED(bs);
+    UNUSED(vx);
+    UNUSED(vy);
+    UNUSED(nr);
+    UNUSED(nc);
+    UNUSED(nb);
+    UNUSED(ncols_interleaved);
+    UNUSED(blocklen);
+
+#if ! ((defined(_MSC_VER)) && ! defined(__clang__)) && defined(__aarch64__) && defined(__ARM_NEON) && defined(__ARM_FEATURE_MATMUL_INT8)
+    if (ggml_cpu_has_neon() && ggml_cpu_has_matmul_int8()) {
+        const void * b_ptr = vx;
+        const void * a_ptr = vy;
+        float * res_ptr = s;
+        size_t res_stride = bs * sizeof(float);
+
+        __asm__ __volatile__(
+            "mov x10, %x[nr]\n"
+            "mov x9, #0x88\n"
+            "cmp x10, #0x10\n"
+            "mul x9, %x[nb], x9\n"
+            "blt 4f\n"
+            "1:"  // Row loop
+            "add x28, %x[b_ptr], #0x8\n"
+            "mov x27, %x[nc]\n"
+            "add x26, %x[res_ptr], %x[res_stride], LSL #4\n"
+            "2:"  // Column loop
+            "add x25, %x[a_ptr], #0x8\n"
+            "movi v2.16b, #0x0\n"
+            "movi v10.16b, #0x0\n"
+            "mov x24, %x[nb]\n"
+            "add x23, x25, x9\n"
+            "movi v12.16b, #0x0\n"
+            "movi v28.16b, #0x0\n"
+            "add x22, x23, x9\n"
+            "movi v11.16b, #0x0\n"
+            "movi v13.16b, #0x0\n"
+            "add x21, x22, x9\n"
+            "movi v22.16b, #0x0\n"
+            "movi v23.16b, #0x0\n"
+            "movi v25.16b, #0x0\n"
+            "movi v5.16b, #0x0\n"
+            "movi v7.16b, #0x0\n"
+            "movi v4.16b, #0x0\n"
+            "movi v6.16b, #0x0\n"
+            "movi v30.16b, #0x0\n"
+            "movi v24.16b, #0x0\n"
+            "movi v14.16b, #0x0\n"
+            "3:"  // Block loop
+            "ldr q21, [x28, #0x0]\n"
+            "ldr q16, [x28, #0x10]\n"
+            "movi v1.16b, #0x4\n"
+            "movi v19.4s, #0x0\n"
+            "ldr q27, [x25, #0x0]\n"
+            "ldr q15, [x25, #0x10]\n"
+            "movi v26.4s, #0x0\n"
+            "movi v18.4s, #0x0\n"
+            "ldr q29, [x28, #0x20]\n"
+            "ldr q3, [x28, #0x30]\n"
+            "movi v17.4s, #0x0\n"
+            "movi v0.16b, #0xf0\n"
+            "ldr d20, [x25, #-0x8]\n"
+            "ldr d9, [x23, #-0x8]\n"
+            "sshl v8.16b, v21.16b, v1.16b\n"
+            "sshl v31.16b, v16.16b, v1.16b\n"
+            "and v21.16b, v21.16b, v0.16b\n"
+            "and v16.16b, v16.16b, v0.16b\n"
+            "sub x20, x28, #0x8\n"
+            "subs x24, x24, #0x1\n"
+            "add x28, x28, #0x48\n"
+            ".inst 0x4e88a773  // smmla v19.4s, v27.16b, v8.16b\n"
+            ".inst 0x4e9fa77a  // smmla v26.4s, v27.16b, v31.16b\n"
+            "ldr q27, [x25, #0x20]\n"
+            ".inst 0x4e88a5f2  // smmla v18.4s, v15.16b, v8.16b\n"
+            ".inst 0x4e9fa5f1  // smmla v17.4s, v15.16b, v31.16b\n"
+            "sshl v15.16b, v29.16b, v1.16b\n"
+            "sshl v1.16b, v3.16b, v1.16b\n"
+            "and v29.16b, v29.16b, v0.16b\n"
+            "and v3.16b, v3.16b, v0.16b\n"
+            "ldr q0, [x25, #0x30]\n"
+            "fcvtl v20.4s, v20.4h\n"
+            ".inst 0x4e8fa773  // smmla v19.4s, v27.16b, v15.16b\n"
+            "fcvtl v9.4s, v9.4h\n"
+            ".inst 0x4e81a77a  // smmla v26.4s, v27.16b, v1.16b\n"
+            "ldr q27, [x25, #0x40]\n"
+            ".inst 0x4e8fa412  // smmla v18.4s, v0.16b, v15.16b\n"
+            ".inst 0x4e81a411  // smmla v17.4s, v0.16b, v1.16b\n"
+            "ldr q0, [x25, #0x50]\n"
+            ".inst 0x4e95a773  // smmla v19.4s, v27.16b, v21.16b\n"
+            ".inst 0x4e90a77a  // smmla v26.4s, v27.16b, v16.16b\n"
+            "ldr q27, [x25, #0x60]\n"
+            ".inst 0x4e95a412  // smmla v18.4s, v0.16b, v21.16b\n"
+            ".inst 0x4e90a411  // smmla v17.4s, v0.16b, v16.16b\n"
+            "ldr q0, [x25, #0x70]\n"
+            "add x25, x25, #0x88\n"
+            ".inst 0x4e9da773  // smmla v19.4s, v27.16b, v29.16b\n"
+            ".inst 0x4e83a77a  // smmla v26.4s, v27.16b, v3.16b\n"
+            "ldr d27, [x20, #0x0]\n"
+            ".inst 0x4e9da412  // smmla v18.4s, v0.16b, v29.16b\n"
+            ".inst 0x4e83a411  // smmla v17.4s, v0.16b, v3.16b\n"
+            "fcvtl v27.4s, v27.4h\n"
+            "uzp1 v0.2d, v19.2d, v26.2d\n"
+            "uzp2 v26.2d, v19.2d, v26.2d\n"
+            "fmul v19.4s, v27.4s, v20.s[0]\n"
+            "scvtf v0.4s, v0.4s, #0x4\n"
+            "scvtf v26.4s, v26.4s, #0x4\n"
+            "fmla v2.4s, v0.4s, v19.4s\n"
+            "ldr q19, [x23, #0x0]\n"
+            "uzp1 v0.2d, v18.2d, v17.2d\n"
+            "uzp2 v18.2d, v18.2d, v17.2d\n"
+            "fmul v17.4s, v27.4s, v20.s[1]\n"
+            "scvtf v0.4s, v0.4s, #0x4\n"
+            "scvtf v18.4s, v18.4s, #0x4\n"
+            "fmla v10.4s, v26.4s, v17.4s\n"
+            "ldr q17, [x23, #0x10]\n"
+            "fmul v26.4s, v27.4s, v20.s[2]\n"
+            "fmul v20.4s, v27.4s, v20.s[3]\n"
+            "fmla v12.4s, v0.4s, v26.4s\n"
+            "ldr d0, [x22, #-0x8]\n"
+            "ldr d26, [x21, #-0x8]\n"
+            "fcvtl v0.4s, v0.4h\n"
+            "fmla v28.4s, v18.4s, v20.4s\n"
+            "movi v20.4s, #0x0\n"
+            "movi v18.4s, #0x0\n"
+            ".inst 0x4e88a674  // smmla v20.4s, v19.16b, v8.16b\n"
+            ".inst 0x4e9fa672  // smmla v18.4s, v19.16b, v31.16b\n"
+            "ldr q19, [x23, #0x20]\n"
+            "fcvtl v26.4s, v26.4h\n"
+            ".inst 0x4e8fa674  // smmla v20.4s, v19.16b, v15.16b\n"
+            ".inst 0x4e81a672  // smmla v18.4s, v19.16b, v1.16b\n"
+            "ldr q19, [x23, #0x40]\n"
+            ".inst 0x4e95a674  // smmla v20.4s, v19.16b, v21.16b\n"
+            ".inst 0x4e90a672  // smmla v18.4s, v19.16b, v16.16b\n"
+            "ldr q19, [x23, #0x60]\n"
+            ".inst 0x4e9da674  // smmla v20.4s, v19.16b, v29.16b\n"
+            ".inst 0x4e83a672  // smmla v18.4s, v19.16b, v3.16b\n"
+            "uzp1 v19.2d, v20.2d, v18.2d\n"
+            "scvtf v19.4s, v19.4s, #0x4\n"
+            "uzp2 v20.2d, v20.2d, v18.2d\n"
+            "fmul v18.4s, v27.4s, v9.s[0]\n"
+            "scvtf v20.4s, v20.4s, #0x4\n"
+            "fmla v11.4s, v19.4s, v18.4s\n"
+            "ldr q18, [x22, #0x0]\n"
+            "fmul v19.4s, v27.4s, v9.s[1]\n"
+            "fmla v13.4s, v20.4s, v19.4s\n"
+            "movi v19.4s, #0x0\n"
+            "movi v20.4s, #0x0\n"
+            ".inst 0x4e88a633  // smmla v19.4s, v17.16b, v8.16b\n"
+            ".inst 0x4e9fa634  // smmla v20.4s, v17.16b, v31.16b\n"
+            "ldr q17, [x23, #0x30]\n"
+            ".inst 0x4e8fa633  // smmla v19.4s, v17.16b, v15.16b\n"
+            ".inst 0x4e81a634  // smmla v20.4s, v17.16b, v1.16b\n"
+            "ldr q17, [x23, #0x50]\n"
+            ".inst 0x4e95a633  // smmla v19.4s, v17.16b, v21.16b\n"
+            ".inst 0x4e90a634  // smmla v20.4s, v17.16b, v16.16b\n"
+            "ldr q17, [x23, #0x70]\n"
+            "add x23, x23, #0x88\n"
+            ".inst 0x4e9da633  // smmla v19.4s, v17.16b, v29.16b\n"
+            ".inst 0x4e83a634  // smmla v20.4s, v17.16b, v3.16b\n"
+            "uzp1 v17.2d, v19.2d, v20.2d\n"
+            "scvtf v17.4s, v17.4s, #0x4\n"
+            "uzp2 v20.2d, v19.2d, v20.2d\n"
+            "fmul v19.4s, v27.4s, v9.s[2]\n"
+            "fmul v9.4s, v27.4s, v9.s[3]\n"
+            "scvtf v20.4s, v20.4s, #0x4\n"
+            "fmla v22.4s, v17.4s, v19.4s\n"
+            "ldr q17, [x22, #0x10]\n"
+            "movi v19.4s, #0x0\n"
+            ".inst 0x4e88a653  // smmla v19.4s, v18.16b, v8.16b\n"
+            "fmla v23.4s, v20.4s, v9.4s\n"
+            "movi v20.4s, #0x0\n"
+            "movi v9.4s, #0x0\n"
+            ".inst 0x4e9fa654  // smmla v20.4s, v18.16b, v31.16b\n"
+            "ldr q18, [x22, #0x20]\n"
+            ".inst 0x4e88a629  // smmla v9.4s, v17.16b, v8.16b\n"
+            ".inst 0x4e8fa653  // smmla v19.4s, v18.16b, v15.16b\n"
+            ".inst 0x4e81a654  // smmla v20.4s, v18.16b, v1.16b\n"
+            "ldr q18, [x22, #0x40]\n"
+            ".inst 0x4e95a653  // smmla v19.4s, v18.16b, v21.16b\n"
+            ".inst 0x4e90a654  // smmla v20.4s, v18.16b, v16.16b\n"
+            "ldr q18, [x22, #0x60]\n"
+            ".inst 0x4e9da653  // smmla v19.4s, v18.16b, v29.16b\n"
+            ".inst 0x4e83a654  // smmla v20.4s, v18.16b, v3.16b\n"
+            "movi v18.4s, #0x0\n"
+            ".inst 0x4e9fa632  // smmla v18.4s, v17.16b, v31.16b\n"
+            "ldr q17, [x22, #0x30]\n"
+            ".inst 0x4e8fa629  // smmla v9.4s, v17.16b, v15.16b\n"
+            ".inst 0x4e81a632  // smmla v18.4s, v17.16b, v1.16b\n"
+            "ldr q17, [x22, #0x50]\n"
+            ".inst 0x4e95a629  // smmla v9.4s, v17.16b, v21.16b\n"
+            ".inst 0x4e90a632  // smmla v18.4s, v17.16b, v16.16b\n"
+            "ldr q17, [x22, #0x70]\n"
+            "add x22, x22, #0x88\n"
+            ".inst 0x4e9da629  // smmla v9.4s, v17.16b, v29.16b\n"
+            ".inst 0x4e83a632  // smmla v18.4s, v17.16b, v3.16b\n"
+            "uzp1 v17.2d, v19.2d, v20.2d\n"
+            "uzp2 v20.2d, v19.2d, v20.2d\n"
+            "fmul v19.4s, v27.4s, v0.s[0]\n"
+            "scvtf v17.4s, v17.4s, #0x4\n"
+            "scvtf v20.4s, v20.4s, #0x4\n"
+            "fmla v25.4s, v17.4s, v19.4s\n"
+            "ldr q19, [x21, #0x0]\n"
+            "fmul v17.4s, v27.4s, v0.s[1]\n"
+            "fmla v5.4s, v20.4s, v17.4s\n"
+            "ldr q17, [x21, #0x10]\n"
+            "uzp1 v20.2d, v9.2d, v18.2d\n"
+            "uzp2 v9.2d, v9.2d, v18.2d\n"
+            "fmul v18.4s, v27.4s, v0.s[2]\n"
+            "fmul v0.4s, v27.4s, v0.s[3]\n"
+            "scvtf v20.4s, v20.4s, #0x4\n"
+            "scvtf v9.4s, v9.4s, #0x4\n"
+            "fmla v7.4s, v20.4s, v18.4s\n"
+            "movi v20.4s, #0x0\n"
+            "movi v18.4s, #0x0\n"
+            ".inst 0x4e88a674  // smmla v20.4s, v19.16b, v8.16b\n"
+            ".inst 0x4e9fa672  // smmla v18.4s, v19.16b, v31.16b\n"
+            "ldr q19, [x21, #0x20]\n"
+            "fmla v4.4s, v9.4s, v0.4s\n"
+            "movi v9.4s, #0x0\n"
+            "movi v0.4s, #0x0\n"
+            ".inst 0x4e88a629  // smmla v9.4s, v17.16b, v8.16b\n"
+            "fmul v8.4s, v27.4s, v26.s[0]\n"
+            ".inst 0x4e9fa620  // smmla v0.4s, v17.16b, v31.16b\n"
+            "ldr q17, [x21, #0x30]\n"
+            ".inst 0x4e8fa674  // smmla v20.4s, v19.16b, v15.16b\n"
+            "fmul v31.4s, v27.4s, v26.s[1]\n"
+            ".inst 0x4e81a672  // smmla v18.4s, v19.16b, v1.16b\n"
+            "ldr q19, [x21, #0x40]\n"
+            ".inst 0x4e8fa629  // smmla v9.4s, v17.16b, v15.16b\n"
+            "fmul v15.4s, v27.4s, v26.s[2]\n"
+            "fmul v27.4s, v27.4s, v26.s[3]\n"
+            ".inst 0x4e81a620  // smmla v0.4s, v17.16b, v1.16b\n"
+            "ldr q1, [x21, #0x50]\n"
+            ".inst 0x4e95a674  // smmla v20.4s, v19.16b, v21.16b\n"
+            ".inst 0x4e90a672  // smmla v18.4s, v19.16b, v16.16b\n"
+            "ldr q26, [x21, #0x60]\n"
+            ".inst 0x4e95a429  // smmla v9.4s, v1.16b, v21.16b\n"
+            ".inst 0x4e90a420  // smmla v0.4s, v1.16b, v16.16b\n"
+            "ldr q21, [x21, #0x70]\n"
+            "add x21, x21, #0x88\n"
+            ".inst 0x4e9da754  // smmla v20.4s, v26.16b, v29.16b\n"
+            ".inst 0x4e83a752  // smmla v18.4s, v26.16b, v3.16b\n"
+            ".inst 0x4e9da6a9  // smmla v9.4s, v21.16b, v29.16b\n"
+            ".inst 0x4e83a6a0  // smmla v0.4s, v21.16b, v3.16b\n"
+            "uzp1 v29.2d, v20.2d, v18.2d\n"
+            "uzp2 v21.2d, v20.2d, v18.2d\n"
+            "scvtf v29.4s, v29.4s, #0x4\n"
+            "uzp1 v18.2d, v9.2d, v0.2d\n"
+            "uzp2 v16.2d, v9.2d, v0.2d\n"
+            "scvtf v21.4s, v21.4s, #0x4\n"
+            "fmla v6.4s, v29.4s, v8.4s\n"
+            "scvtf v18.4s, v18.4s, #0x4\n"
+            "scvtf v16.4s, v16.4s, #0x4\n"
+            "fmla v30.4s, v21.4s, v31.4s\n"
+            "fmla v24.4s, v18.4s, v15.4s\n"
+            "fmla v14.4s, v16.4s, v27.4s\n"
+            "bgt 3b\n"
+            "mov x20, %x[res_ptr]\n"
+            "subs x27, x27, #0x4\n"
+            "add %x[res_ptr], %x[res_ptr], #0x10\n"
+            "str q2, [x20, #0x0]\n"
+            "add x20, x20, %x[res_stride]\n"
+            "str q10, [x20, #0x0]\n"
+            "add x20, x20, %x[res_stride]\n"
+            "str q12, [x20, #0x0]\n"
+            "add x20, x20, %x[res_stride]\n"
+            "str q28, [x20, #0x0]\n"
+            "add x20, x20, %x[res_stride]\n"
+            "str q11, [x20, #0x0]\n"
+            "add x20, x20, %x[res_stride]\n"
+            "str q13, [x20, #0x0]\n"
+            "add x20, x20, %x[res_stride]\n"
+            "str q22, [x20, #0x0]\n"
+            "add x20, x20, %x[res_stride]\n"
+            "str q23, [x20, #0x0]\n"
+            "add x20, x20, %x[res_stride]\n"
+            "str q25, [x20, #0x0]\n"
+            "add x20, x20, %x[res_stride]\n"
+            "str q5, [x20, #0x0]\n"
+            "add x20, x20, %x[res_stride]\n"
+            "str q7, [x20, #0x0]\n"
+            "add x20, x20, %x[res_stride]\n"
+            "str q4, [x20, #0x0]\n"
+            "add x20, x20, %x[res_stride]\n"
+            "str q6, [x20, #0x0]\n"
+            "add x20, x20, %x[res_stride]\n"
+            "str q30, [x20, #0x0]\n"
+            "add x20, x20, %x[res_stride]\n"
+            "str q24, [x20, #0x0]\n"
+            "add x20, x20, %x[res_stride]\n"
+            "str q14, [x20, #0x0]\n"
+            "bne 2b\n"
+            "mov x20, #0x4\n"
+            "sub x10, x10, #0x10\n"
+            "cmp x10, #0x10\n"
+            "mov %x[res_ptr], x26\n"
+            "madd %x[a_ptr], x20, x9, %x[a_ptr]\n"
+            "bge 1b\n"
+            "4:"  // Row loop skip
+            "cbz x10, 9f\n"
+            "5:"  // Row tail: Row loop
+            "add x24, %x[b_ptr], #0x8\n"
+            "mov x23, %x[nc]\n"
+            "add x22, %x[res_ptr], %x[res_stride], LSL #2\n"
+            "6:"  // Row tail: Column loop
+            "movi v2.16b, #0x0\n"
+            "movi v10.16b, #0x0\n"
+            "add x25, %x[a_ptr], #0x8\n"
+            "mov x21, %x[nb]\n"
+            "movi v12.16b, #0x0\n"
+            "movi v28.16b, #0x0\n"
+            "7:"  // Row tail: Block loop
+            "ldr q6, [x24, #0x0]\n"
+            "ldr q5, [x24, #0x10]\n"
+            "movi v17.16b, #0x4\n"
+            "movi v8.4s, #0x0\n"
+            "ldr q4, [x25, #0x0]\n"
+            "ldr q13, [x25, #0x10]\n"
+            "movi v27.4s, #0x0\n"
+            "movi v0.4s, #0x0\n"
+            "ldr q31, [x24, #0x20]\n"
+            "ldr q14, [x24, #0x30]\n"
+            "movi v29.4s, #0x0\n"
+            "movi v22.16b, #0xf0\n"
+            "ldr q11, [x25, #0x20]\n"
+            "ldr q23, [x25, #0x30]\n"
+            "sshl v21.16b, v6.16b, v17.16b\n"
+            "sshl v16.16b, v5.16b, v17.16b\n"
+            "ldr q20, [x25, #0x40]\n"
+            "ldr q26, [x25, #0x50]\n"
+            "and v6.16b, v6.16b, v22.16b\n"
+            "and v5.16b, v5.16b, v22.16b\n"
+            "ldr q25, [x25, #0x60]\n"
+            "ldr q3, [x25, #0x70]\n"
+            "sshl v19.16b, v31.16b, v17.16b\n"
+            "sshl v18.16b, v14.16b, v17.16b\n"
+            "ldr d17, [x25, #-0x8]\n"
+            ".inst 0x4e95a488  // smmla v8.4s, v4.16b, v21.16b\n"
+            ".inst 0x4e90a49b  // smmla v27.4s, v4.16b, v16.16b\n"
+            "and v31.16b, v31.16b, v22.16b\n"
+            ".inst 0x4e95a5a0  // smmla v0.4s, v13.16b, v21.16b\n"
+            ".inst 0x4e90a5bd  // smmla v29.4s, v13.16b, v16.16b\n"
+            "and v14.16b, v14.16b, v22.16b\n"
+            "sub x20, x24, #0x8\n"
+            "ldr d16, [x20, #0x0]\n"
+            "subs x21, x21, #0x1\n"
+            "add x25, x25, #0x88\n"
+            "fcvtl v17.4s, v17.4h\n"
+            "add x24, x24, #0x48\n"
+            ".inst 0x4e93a568  // smmla v8.4s, v11.16b, v19.16b\n"
+            ".inst 0x4e92a57b  // smmla v27.4s, v11.16b, v18.16b\n"
+            ".inst 0x4e93a6e0  // smmla v0.4s, v23.16b, v19.16b\n"
+            ".inst 0x4e92a6fd  // smmla v29.4s, v23.16b, v18.16b\n"
+            "fcvtl v16.4s, v16.4h\n"
+            ".inst 0x4e86a688  // smmla v8.4s, v20.16b, v6.16b\n"
+            ".inst 0x4e85a69b  // smmla v27.4s, v20.16b, v5.16b\n"
+            "fmul v23.4s, v16.4s, v17.s[0]\n"
+            "fmul v21.4s, v16.4s, v17.s[1]\n"
+            "fmul v1.4s, v16.4s, v17.s[2]\n"
+            "fmul v20.4s, v16.4s, v17.s[3]\n"
+            ".inst 0x4e86a740  // smmla v0.4s, v26.16b, v6.16b\n"
+            ".inst 0x4e85a75d  // smmla v29.4s, v26.16b, v5.16b\n"
+            ".inst 0x4e9fa728  // smmla v8.4s, v25.16b, v31.16b\n"
+            ".inst 0x4e8ea73b  // smmla v27.4s, v25.16b, v14.16b\n"
+            ".inst 0x4e9fa460  // smmla v0.4s, v3.16b, v31.16b\n"
+            ".inst 0x4e8ea47d  // smmla v29.4s, v3.16b, v14.16b\n"
+            "uzp1 v19.2d, v8.2d, v27.2d\n"
+            "uzp2 v18.2d, v8.2d, v27.2d\n"
+            "scvtf v19.4s, v19.4s, #0x4\n"
+            "uzp1 v17.2d, v0.2d, v29.2d\n"
+            "uzp2 v16.2d, v0.2d, v29.2d\n"
+            "scvtf v18.4s, v18.4s, #0x4\n"
+            "fmla v2.4s, v19.4s, v23.4s\n"
+            "scvtf v17.4s, v17.4s, #0x4\n"
+            "scvtf v16.4s, v16.4s, #0x4\n"
+            "fmla v10.4s, v18.4s, v21.4s\n"
+            "fmla v12.4s, v17.4s, v1.4s\n"
+            "fmla v28.4s, v16.4s, v20.4s\n"
+            "bgt 7b\n"
+            "mov x20, %x[res_ptr]\n"
+            "cmp x10, #0x1\n"
+            "str q2, [x20, #0x0]\n"
+            "add x20, x20, %x[res_stride]\n"
+            "ble 8f\n"
+            "cmp x10, #0x2\n"
+            "str q10, [x20, #0x0]\n"
+            "add x20, x20, %x[res_stride]\n"
+            "ble 8f\n"
+            "cmp x10, #0x3\n"
+            "str q12, [x20, #0x0]\n"
+            "add x20, x20, %x[res_stride]\n"
+            "ble 8f\n"
+            "str q28, [x20, #0x0]\n"
+            "8:"  // Row tail: Accumulator store skip
+            "subs x23, x23, #0x4\n"
+            "add %x[res_ptr], %x[res_ptr], #0x10\n"
+            "bne 6b\n"
+            "subs x10, x10, #0x4\n"
+            "add %x[a_ptr], %x[a_ptr], x9\n"
+            "mov %x[res_ptr], x22\n"
+            "bgt 5b\n"
+            "9:"  // Row tail: Row loop skip
+            : [a_ptr] "+&r" (a_ptr), [res_ptr] "+&r" (res_ptr)
+            : [b_ptr] "r" (b_ptr), [nr] "r" (nr), [nb] "r" (nb), [res_stride] "r" (res_stride), [nc] "r" (nc)
+            : "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v8", "v9", "v10", "v11", "v12", "v13", "v14", "v15", "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30", "v31", "x9", "x10", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "x27", "x28"
+        );
+        return;
+    }
+#endif // #if ! ((defined(_MSC_VER)) && ! defined(__clang__)) && defined(__aarch64__) && defined(__ARM_NEON) && defined(__ARM_FEATURE_MATMUL_INT8)
+    float sumf[4][4];
+    int sumi;
+
+    for (int y = 0; y < nr / 4; y++) {
+        const block_q8_0x4 * a_ptr = (const block_q8_0x4 *) vy + (y * nb);
+        for (int x = 0; x < nc / ncols_interleaved; x++) {
+            const block_q4_0x4 * b_ptr = (const block_q4_0x4 *) vx + (x * nb);
+            for (int m = 0; m < 4; m++) {
+                for (int j = 0; j < ncols_interleaved; j++) sumf[m][j] = 0.0;
+            }
+            for (int l = 0; l < nb; l++) {
+                for (int k = 0; k < (qk / (2 * blocklen)); k++) {
+                    for (int m = 0; m < 4; m++) {
+                        for (int j = 0; j < ncols_interleaved; j++) {
+                            sumi = 0;
+                            for (int i = 0; i < blocklen; ++i) {
+                                const int v0 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] << 4);
+                                const int v1 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] & 0xF0);
+                                sumi += ((v0 * a_ptr[l].qs[k * 4 * blocklen + m * blocklen + i]) +
+                                        (v1 * a_ptr[l].qs[k * 4 * blocklen + m * blocklen + i + qk / 2 * 4])) >> 4;
+                            }
+                            sumf[m][j] += sumi * GGML_FP16_TO_FP32(b_ptr[l].d[j]) * GGML_FP16_TO_FP32(a_ptr[l].d[m]);
+                        }
+                    }
+                }
+            }
+            for (int m = 0; m < 4; m++) {
+                for (int j = 0; j < ncols_interleaved; j++)
+                    s[(y * 4 + m) * bs + x * ncols_interleaved + j] = sumf[m][j];
+            }
+        }
+    }
+}
+
+void ggml_gemm_q4_0_8x8_q8_0(int n, float * restrict s, size_t bs, const void * restrict vx, const void * restrict vy, int nr, int nc) {
+    const int qk = QK8_0;
+    const int nb = n / qk;
+    const int ncols_interleaved = 8;
+    const int blocklen = 8;
+
+    assert (n % qk == 0);
+    assert (nr % 4 == 0);
+    assert (nc % ncols_interleaved == 0);
+
+    UNUSED(s);
+    UNUSED(bs);
+    UNUSED(vx);
+    UNUSED(vy);
+    UNUSED(nr);
+    UNUSED(nc);
+    UNUSED(nb);
+    UNUSED(ncols_interleaved);
+    UNUSED(blocklen);
+
+#if ! ((defined(_MSC_VER)) && ! defined(__clang__)) && defined(__aarch64__)
+#if defined(__ARM_FEATURE_SVE) && defined(__ARM_FEATURE_MATMUL_INT8)
+    if (ggml_cpu_has_sve() && ggml_cpu_has_matmul_int8() && ggml_cpu_get_sve_cnt() == QK8_0) {
+        const void * b_ptr = vx;
+        const void * a_ptr = vy;
+        float * res_ptr = s;
+        size_t res_stride = bs * sizeof(float);
+
+        __asm__ __volatile__(
+            "mov x20, #0x4\n"
+            "mov x13, %x[nr]\n"
+            "mov z28.s, #-0x4\n"
+            "mov x12, #0x88\n"
+            "ptrue p1.b\n"
+            "whilelt p0.s, XZR, x20\n"
+            "cmp x13, #0x10\n"
+            "mul x12, %x[nb], x12\n"
+            "blt 4f\n"
+            "1:"  // Row loop
+            "add x11, %x[b_ptr], #0x10\n"
+            "mov x10, %x[nc]\n"
+            "add x9, %x[res_ptr], %x[res_stride], LSL #4\n"
+            "2:"  // Column loop
+            "add x28, %x[a_ptr], #0x8\n"
+            "mov z24.b, #0x0\n"
+            "mov z15.b, #0x0\n"
+            "mov x27, %x[nb]\n"
+            "add x26, x28, x12\n"
+            "mov z12.b, #0x0\n"
+            "mov z0.b, #0x0\n"
+            "add x25, x26, x12\n"
+            "mov z13.b, #0x0\n"
+            "mov z1.b, #0x0\n"
+            "add x24, x25, x12\n"
+            "mov z20.b, #0x0\n"
+            "mov z25.b, #0x0\n"
+            "mov z11.b, #0x0\n"
+            "mov z16.b, #0x0\n"
+            "mov z19.b, #0x0\n"
+            "mov z26.b, #0x0\n"
+            "mov z8.b, #0x0\n"
+            "mov z29.b, #0x0\n"
+            "mov z27.b, #0x0\n"
+            "mov z10.b, #0x0\n"
+            "3:"  // Block loop
+            "ld1b { z30.b }, p1/Z, [x11]\n"
+            "ld1b { z21.b }, p1/Z, [x11, #1, MUL VL]\n"
+            "mov z18.s, #0x0\n"
+            "mov z7.s, #0x0\n"
+            "ld1rqb { z3.b }, p1/Z, [x28]\n"
+            "ld1rqb { z5.b }, p1/Z, [x28, #16]\n"
+            "mov z9.s, #0x0\n"
+            "mov z22.s, #0x0\n"
+            "ld1b { z4.b }, p1/Z, [x11, #2, MUL VL]\n"
+            "ld1b { z17.b }, p1/Z, [x11, #3, MUL VL]\n"
+            "sub x20, x11, #0x10\n"
+            "sub x23, x28, #0x8\n"
+            "lsl z31.b, z30.b, #0x4\n"
+            "lsl z6.b, z21.b, #0x4\n"
+            "ld1h { z23.s }, p1/Z, [x20]\n"
+            "sub x22, x26, #0x8\n"
+            "and z30.b, z30.b, #0xf0\n"
+            "and z21.b, z21.b, #0xf0\n"
+            "sub x21, x25, #0x8\n"
+            "sub x20, x24, #0x8\n"
+            "lsl z14.b, z4.b, #0x4\n"
+            "lsl z2.b, z17.b, #0x4\n"
+            "subs x27, x27, #0x1\n"
+            "add x11, x11, #0x90\n"
+            ".inst 0x451f9872  // smmla z18.s, z3.b, z31.b\n"
+            ".inst 0x45069867  // smmla z7.s, z3.b, z6.b\n"
+            "ld1rqb { z3.b }, p1/Z, [x28, #32]\n"
+            "and z4.b, z4.b, #0xf0\n"
+            ".inst 0x451f98a9  // smmla z9.s, z5.b, z31.b\n"
+            ".inst 0x450698b6  // smmla z22.s, z5.b, z6.b\n"
+            "ld1rqb { z5.b }, p1/Z, [x28, #48]\n"
+            "and z17.b, z17.b, #0xf0\n"
+            "fcvt z23.s, p1/m, z23.h\n"
+            ".inst 0x450e9872  // smmla z18.s, z3.b, z14.b\n"
+            ".inst 0x45029867  // smmla z7.s, z3.b, z2.b\n"
+            "ld1rqb { z3.b }, p1/Z, [x28, #64]\n"
+            ".inst 0x450e98a9  // smmla z9.s, z5.b, z14.b\n"
+            ".inst 0x450298b6  // smmla z22.s, z5.b, z2.b\n"
+            "ld1rqb { z5.b }, p1/Z, [x28, #80]\n"
+            "fscale z23.s, p1/m, z23.s, z28.s\n"
+            ".inst 0x451e9872  // smmla z18.s, z3.b, z30.b\n"
+            ".inst 0x45159867  // smmla z7.s, z3.b, z21.b\n"
+            "ld1rqb { z3.b }, p1/Z, [x28, #96]\n"
+            ".inst 0x451e98a9  // smmla z9.s, z5.b, z30.b\n"
+            ".inst 0x451598b6  // smmla z22.s, z5.b, z21.b\n"
+            "ld1rqb { z5.b }, p1/Z, [x28, #112]\n"
+            "add x28, x28, #0x88\n"
+            ".inst 0x45049872  // smmla z18.s, z3.b, z4.b\n"
+            ".inst 0x45119867  // smmla z7.s, z3.b, z17.b\n"
+            "ld1h { z3.s }, p0/Z, [x23]\n"
+            ".inst 0x450498a9  // smmla z9.s, z5.b, z4.b\n"
+            ".inst 0x451198b6  // smmla z22.s, z5.b, z17.b\n"
+            "fcvt z3.s, p1/m, z3.h\n"
+            "uzp1 z5.d, z18.d, z7.d\n"
+            "uzp2 z18.d, z18.d, z7.d\n"
+            "mov z3.q, z3.q[0]\n"
+            "uzp1 z7.d, z9.d, z22.d\n"
+            "uzp2 z22.d, z9.d, z22.d\n"
+            "fmul z9.s, z23.s, z3.s[0]\n"
+            "scvtf z5.s, p1/m, z5.s\n"
+            "scvtf z18.s, p1/m, z18.s\n"
+            "scvtf z7.s, p1/m, z7.s\n"
+            "scvtf z22.s, p1/m, z22.s\n"
+            "fmla z24.s, p1/M, z5.s, z9.s\n"
+            "ld1rqb { z5.b }, p1/Z, [x26]\n"
+            "fmul z9.s, z23.s, z3.s[1]\n"
+            "fmla z15.s, p1/M, z18.s, z9.s\n"
+            "ld1rqb { z18.b }, p1/Z, [x26, #16]\n"
+            "fmul z9.s, z23.s, z3.s[2]\n"
+            "fmul z3.s, z23.s, z3.s[3]\n"
+            "fmla z12.s, p1/M, z7.s, z9.s\n"
+            "mov z9.s, #0x0\n"
+            "ld1h { z7.s }, p0/Z, [x22]\n"
+            ".inst 0x451f98a9  // smmla z9.s, z5.b, z31.b\n"
+            "fmla z0.s, p1/M, z22.s, z3.s\n"
+            "mov z22.s, #0x0\n"
+            "ld1h { z3.s }, p0/Z, [x21]\n"
+            ".inst 0x450698b6  // smmla z22.s, z5.b, z6.b\n"
+            "ld1rqb { z5.b }, p1/Z, [x26, #32]\n"
+            "fcvt z7.s, p1/m, z7.h\n"
+            "fcvt z3.s, p1/m, z3.h\n"
+            ".inst 0x450e98a9  // smmla z9.s, z5.b, z14.b\n"
+            ".inst 0x450298b6  // smmla z22.s, z5.b, z2.b\n"
+            "ld1rqb { z5.b }, p1/Z, [x26, #64]\n"
+            "mov z7.q, z7.q[0]\n"
+            "mov z3.q, z3.q[0]\n"
+            ".inst 0x451e98a9  // smmla z9.s, z5.b, z30.b\n"
+            ".inst 0x451598b6  // smmla z22.s, z5.b, z21.b\n"
+            "ld1rqb { z5.b }, p1/Z, [x26, #96]\n"
+            ".inst 0x450498a9  // smmla z9.s, z5.b, z4.b\n"
+            ".inst 0x451198b6  // smmla z22.s, z5.b, z17.b\n"
+            "uzp1 z5.d, z9.d, z22.d\n"
+            "scvtf z5.s, p1/m, z5.s\n"
+            "uzp2 z22.d, z9.d, z22.d\n"
+            "fmul z9.s, z23.s, z7.s[0]\n"
+            "scvtf z22.s, p1/m, z22.s\n"
+            "fmla z13.s, p1/M, z5.s, z9.s\n"
+            "ld1rqb { z9.b }, p1/Z, [x25]\n"
+            "fmul z5.s, z23.s, z7.s[1]\n"
+            "fmla z1.s, p1/M, z22.s, z5.s\n"
+            "mov z5.s, #0x0\n"
+            "mov z22.s, #0x0\n"
+            ".inst 0x451f9a45  // smmla z5.s, z18.b, z31.b\n"
+            ".inst 0x45069a56  // smmla z22.s, z18.b, z6.b\n"
+            "ld1rqb { z18.b }, p1/Z, [x26, #48]\n"
+            ".inst 0x450e9a45  // smmla z5.s, z18.b, z14.b\n"
+            ".inst 0x45029a56  // smmla z22.s, z18.b, z2.b\n"
+            "ld1rqb { z18.b }, p1/Z, [x26, #80]\n"
+            ".inst 0x451e9a45  // smmla z5.s, z18.b, z30.b\n"
+            ".inst 0x45159a56  // smmla z22.s, z18.b, z21.b\n"
+            "ld1rqb { z18.b }, p1/Z, [x26, #112]\n"
+            "add x26, x26, #0x88\n"
+            ".inst 0x45049a45  // smmla z5.s, z18.b, z4.b\n"
+            ".inst 0x45119a56  // smmla z22.s, z18.b, z17.b\n"
+            "uzp1 z18.d, z5.d, z22.d\n"
+            "scvtf z18.s, p1/m, z18.s\n"
+            "uzp2 z22.d, z5.d, z22.d\n"
+            "fmul z5.s, z23.s, z7.s[2]\n"
+            "fmul z7.s, z23.s, z7.s[3]\n"
+            "scvtf z22.s, p1/m, z22.s\n"
+            "fmla z20.s, p1/M, z18.s, z5.s\n"
+            "ld1rqb { z18.b }, p1/Z, [x25, #16]\n"
+            "ld1h { z5.s }, p0/Z, [x20]\n"
+            "fcvt z5.s, p1/m, z5.h\n"
+            "fmla z25.s, p1/M, z22.s, z7.s\n"
+            "mov z22.s, #0x0\n"
+            "mov z7.s, #0x0\n"
+            ".inst 0x451f9936  // smmla z22.s, z9.b, z31.b\n"
+            ".inst 0x45069927  // smmla z7.s, z9.b, z6.b\n"
+            "ld1rqb { z9.b }, p1/Z, [x25, #32]\n"
+            "mov z5.q, z5.q[0]\n"
+            ".inst 0x450e9936  // smmla z22.s, z9.b, z14.b\n"
+            ".inst 0x45029927  // smmla z7.s, z9.b, z2.b\n"
+            "ld1rqb { z9.b }, p1/Z, [x25, #64]\n"
+            ".inst 0x451e9936  // smmla z22.s, z9.b, z30.b\n"
+            ".inst 0x45159927  // smmla z7.s, z9.b, z21.b\n"
+            "ld1rqb { z9.b }, p1/Z, [x25, #96]\n"
+            ".inst 0x45049936  // smmla z22.s, z9.b, z4.b\n"
+            ".inst 0x45119927  // smmla z7.s, z9.b, z17.b\n"
+            "uzp1 z9.d, z22.d, z7.d\n"
+            "scvtf z9.s, p1/m, z9.s\n"
+            "uzp2 z22.d, z22.d, z7.d\n"
+            "fmul z7.s, z23.s, z3.s[0]\n"
+            "scvtf z22.s, p1/m, z22.s\n"
+            "fmla z11.s, p1/M, z9.s, z7.s\n"
+            "ld1rqb { z9.b }, p1/Z, [x24]\n"
+            "fmul z7.s, z23.s, z3.s[1]\n"
+            "fmla z16.s, p1/M, z22.s, z7.s\n"
+            "mov z22.s, #0x0\n"
+            "mov z7.s, #0x0\n"
+            ".inst 0x451f9a56  // smmla z22.s, z18.b, z31.b\n"
+            ".inst 0x45069a47  // smmla z7.s, z18.b, z6.b\n"
+            "ld1rqb { z18.b }, p1/Z, [x25, #48]\n"
+            ".inst 0x450e9a56  // smmla z22.s, z18.b, z14.b\n"
+            ".inst 0x45029a47  // smmla z7.s, z18.b, z2.b\n"
+            "ld1rqb { z18.b }, p1/Z, [x25, #80]\n"
+            ".inst 0x451e9a56  // smmla z22.s, z18.b, z30.b\n"
+            ".inst 0x45159a47  // smmla z7.s, z18.b, z21.b\n"
+            "ld1rqb { z18.b }, p1/Z, [x25, #112]\n"
+            "add x25, x25, #0x88\n"
+            ".inst 0x45049a56  // smmla z22.s, z18.b, z4.b\n"
+            ".inst 0x45119a47  // smmla z7.s, z18.b, z17.b\n"
+            "uzp1 z18.d, z22.d, z7.d\n"
+            "scvtf z18.s, p1/m, z18.s\n"
+            "uzp2 z7.d, z22.d, z7.d\n"
+            "fmul z22.s, z23.s, z3.s[2]\n"
+            "fmul z3.s, z23.s, z3.s[3]\n"
+            "scvtf z7.s, p1/m, z7.s\n"
+            "fmla z19.s, p1/M, z18.s, z22.s\n"
+            "ld1rqb { z18.b }, p1/Z, [x24, #16]\n"
+            "fmul z22.s, z23.s, z5.s[0]\n"
+            "fmla z26.s, p1/M, z7.s, z3.s\n"
+            "mov z3.s, #0x0\n"
+            "mov z7.s, #0x0\n"
+            ".inst 0x451f9923  // smmla z3.s, z9.b, z31.b\n"
+            ".inst 0x45069927  // smmla z7.s, z9.b, z6.b\n"
+            "ld1rqb { z9.b }, p1/Z, [x24, #32]\n"
+            ".inst 0x450e9923  // smmla z3.s, z9.b, z14.b\n"
+            ".inst 0x45029927  // smmla z7.s, z9.b, z2.b\n"
+            "mov z9.s, #0x0\n"
+            ".inst 0x451f9a49  // smmla z9.s, z18.b, z31.b\n"
+            "mov z31.s, #0x0\n"
+            ".inst 0x45069a5f  // smmla z31.s, z18.b, z6.b\n"
+            "ld1rqb { z6.b }, p1/Z, [x24, #48]\n"
+            "ld1rqb { z18.b }, p1/Z, [x24, #64]\n"
+            ".inst 0x450e98c9  // smmla z9.s, z6.b, z14.b\n"
+            "fmul z14.s, z23.s, z5.s[1]\n"
+            ".inst 0x450298df  // smmla z31.s, z6.b, z2.b\n"
+            "ld1rqb { z6.b }, p1/Z, [x24, #80]\n"
+            "fmul z2.s, z23.s, z5.s[2]\n"
+            "fmul z23.s, z23.s, z5.s[3]\n"
+            ".inst 0x451e9a43  // smmla z3.s, z18.b, z30.b\n"
+            ".inst 0x45159a47  // smmla z7.s, z18.b, z21.b\n"
+            "ld1rqb { z5.b }, p1/Z, [x24, #96]\n"
+            ".inst 0x451e98c9  // smmla z9.s, z6.b, z30.b\n"
+            ".inst 0x451598df  // smmla z31.s, z6.b, z21.b\n"
+            "ld1rqb { z18.b }, p1/Z, [x24, #112]\n"
+            "add x24, x24, #0x88\n"
+            ".inst 0x450498a3  // smmla z3.s, z5.b, z4.b\n"
+            ".inst 0x451198a7  // smmla z7.s, z5.b, z17.b\n"
+            ".inst 0x45049a49  // smmla z9.s, z18.b, z4.b\n"
+            ".inst 0x45119a5f  // smmla z31.s, z18.b, z17.b\n"
+            "uzp1 z18.d, z3.d, z7.d\n"
+            "uzp2 z5.d, z3.d, z7.d\n"
+            "scvtf z18.s, p1/m, z18.s\n"
+            "uzp1 z6.d, z9.d, z31.d\n"
+            "uzp2 z9.d, z9.d, z31.d\n"
+            "scvtf z5.s, p1/m, z5.s\n"
+            "fmla z8.s, p1/M, z18.s, z22.s\n"
+            "scvtf z6.s, p1/m, z6.s\n"
+            "scvtf z9.s, p1/m, z9.s\n"
+            "fmla z29.s, p1/M, z5.s, z14.s\n"
+            "fmla z27.s, p1/M, z6.s, z2.s\n"
+            "fmla z10.s, p1/M, z9.s, z23.s\n"
+            "bgt 3b\n"
+            "mov x20, %x[res_ptr]\n"
+            "subs x10, x10, #0x8\n"
+            "add %x[res_ptr], %x[res_ptr], #0x20\n"
+            "st1w { z24.s }, p1, [x20]\n"
+            "add x20, x20, %x[res_stride]\n"
+            "st1w { z15.s }, p1, [x20]\n"
+            "add x20, x20, %x[res_stride]\n"
+            "st1w { z12.s }, p1, [x20]\n"
+            "add x20, x20, %x[res_stride]\n"
+            "st1w { z0.s }, p1, [x20]\n"
+            "add x20, x20, %x[res_stride]\n"
+            "st1w { z13.s }, p1, [x20]\n"
+            "add x20, x20, %x[res_stride]\n"
+            "st1w { z1.s }, p1, [x20]\n"
+            "add x20, x20, %x[res_stride]\n"
+            "st1w { z20.s }, p1, [x20]\n"
+            "add x20, x20, %x[res_stride]\n"
+            "st1w { z25.s }, p1, [x20]\n"
+            "add x20, x20, %x[res_stride]\n"
+            "st1w { z11.s }, p1, [x20]\n"
+            "add x20, x20, %x[res_stride]\n"
+            "st1w { z16.s }, p1, [x20]\n"
+            "add x20, x20, %x[res_stride]\n"
+            "st1w { z19.s }, p1, [x20]\n"
+            "add x20, x20, %x[res_stride]\n"
+            "st1w { z26.s }, p1, [x20]\n"
+            "add x20, x20, %x[res_stride]\n"
+            "st1w { z8.s }, p1, [x20]\n"
+            "add x20, x20, %x[res_stride]\n"
+            "st1w { z29.s }, p1, [x20]\n"
+            "add x20, x20, %x[res_stride]\n"
+            "st1w { z27.s }, p1, [x20]\n"
+            "add x20, x20, %x[res_stride]\n"
+            "st1w { z10.s }, p1, [x20]\n"
+            "bne 2b\n"
+            "mov x20, #0x4\n"
+            "sub x13, x13, #0x10\n"
+            "cmp x13, #0x10\n"
+            "mov %x[res_ptr], x9\n"
+            "madd %x[a_ptr], x20, x12, %x[a_ptr]\n"
+            "bge 1b\n"
+            "4:"  // Row loop skip
+            "cbz x13, 9f\n"
+            "5:"  // Row tail: Row loop
+            "add x25, %x[b_ptr], #0x10\n"
+            "mov x24, %x[nc]\n"
+            "add x23, %x[res_ptr], %x[res_stride], LSL #2\n"
+            "6:"  // Row tail: Column loop
+            "mov z24.b, #0x0\n"
+            "mov z15.b, #0x0\n"
+            "add x28, %x[a_ptr], #0x8\n"
+            "mov x22, %x[nb]\n"
+            "mov z12.b, #0x0\n"
+            "mov z0.b, #0x0\n"
+            "7:"  // Row tail: Block loop
+            "ld1b { z3.b }, p1/Z, [x25]\n"
+            "ld1b { z6.b }, p1/Z, [x25, #1, MUL VL]\n"
+            "mov z2.s, #0x0\n"
+            "mov z25.s, #0x0\n"
+            "ld1rqb { z26.b }, p1/Z, [x28]\n"
+            "ld1rqb { z21.b }, p1/Z, [x28, #16]\n"
+            "mov z27.s, #0x0\n"
+            "mov z19.s, #0x0\n"
+            "ld1b { z29.b }, p1/Z, [x25, #2, MUL VL]\n"
+            "ld1b { z16.b }, p1/Z, [x25, #3, MUL VL]\n"
+            "sub x21, x25, #0x10\n"
+            "sub x20, x28, #0x8\n"
+            "lsl z20.b, z3.b, #0x4\n"
+            "lsl z4.b, z6.b, #0x4\n"
+            "ld1rqb { z10.b }, p1/Z, [x28, #32]\n"
+            "ld1rqb { z23.b }, p1/Z, [x28, #48]\n"
+            "and z3.b, z3.b, #0xf0\n"
+            "and z6.b, z6.b, #0xf0\n"
+            "ld1rqb { z11.b }, p1/Z, [x28, #64]\n"
+            "ld1rqb { z7.b }, p1/Z, [x28, #80]\n"
+            "lsl z8.b, z29.b, #0x4\n"
+            "lsl z14.b, z16.b, #0x4\n"
+            "ld1rqb { z18.b }, p1/Z, [x28, #96]\n"
+            "ld1rqb { z30.b }, p1/Z, [x28, #112]\n"
+            ".inst 0x45149b42  // smmla z2.s, z26.b, z20.b\n"
+            ".inst 0x45049b59  // smmla z25.s, z26.b, z4.b\n"
+            "and z29.b, z29.b, #0xf0\n"
+            "ld1h { z17.s }, p1/Z, [x21]\n"
+            ".inst 0x45149abb  // smmla z27.s, z21.b, z20.b\n"
+            ".inst 0x45049ab3  // smmla z19.s, z21.b, z4.b\n"
+            "and z16.b, z16.b, #0xf0\n"
+            "ld1h { z4.s }, p0/Z, [x20]\n"
+            "subs x22, x22, #0x1\n"
+            "add x28, x28, #0x88\n"
+            "fcvt z17.s, p1/m, z17.h\n"
+            "add x25, x25, #0x90\n"
+            ".inst 0x45089942  // smmla z2.s, z10.b, z8.b\n"
+            ".inst 0x450e9959  // smmla z25.s, z10.b, z14.b\n"
+            "fcvt z4.s, p1/m, z4.h\n"
+            ".inst 0x45089afb  // smmla z27.s, z23.b, z8.b\n"
+            ".inst 0x450e9af3  // smmla z19.s, z23.b, z14.b\n"
+            "fscale z17.s, p1/m, z17.s, z28.s\n"
+            "mov z4.q, z4.q[0]\n"
+            ".inst 0x45039962  // smmla z2.s, z11.b, z3.b\n"
+            ".inst 0x45069979  // smmla z25.s, z11.b, z6.b\n"
+            "fmul z23.s, z17.s, z4.s[0]\n"
+            "fmul z9.s, z17.s, z4.s[1]\n"
+            "fmul z21.s, z17.s, z4.s[2]\n"
+            "fmul z4.s, z17.s, z4.s[3]\n"
+            ".inst 0x450398fb  // smmla z27.s, z7.b, z3.b\n"
+            ".inst 0x450698f3  // smmla z19.s, z7.b, z6.b\n"
+            ".inst 0x451d9a42  // smmla z2.s, z18.b, z29.b\n"
+            ".inst 0x45109a59  // smmla z25.s, z18.b, z16.b\n"
+            ".inst 0x451d9bdb  // smmla z27.s, z30.b, z29.b\n"
+            ".inst 0x45109bd3  // smmla z19.s, z30.b, z16.b\n"
+            "uzp1 z31.d, z2.d, z25.d\n"
+            "uzp2 z13.d, z2.d, z25.d\n"
+            "scvtf z31.s, p1/m, z31.s\n"
+            "uzp1 z17.d, z27.d, z19.d\n"
+            "uzp2 z18.d, z27.d, z19.d\n"
+            "scvtf z13.s, p1/m, z13.s\n"
+            "fmla z24.s, p1/M, z31.s, z23.s\n"
+            "scvtf z17.s, p1/m, z17.s\n"
+            "scvtf z18.s, p1/m, z18.s\n"
+            "fmla z15.s, p1/M, z13.s, z9.s\n"
+            "fmla z12.s, p1/M, z17.s, z21.s\n"
+            "fmla z0.s, p1/M, z18.s, z4.s\n"
+            "bgt 7b\n"
+            "mov x20, %x[res_ptr]\n"
+            "cmp x13, #0x1\n"
+            "st1w { z24.s }, p1, [x20]\n"
+            "add x20, x20, %x[res_stride]\n"
+            "ble 8f\n"
+            "cmp x13, #0x2\n"
+            "st1w { z15.s }, p1, [x20]\n"
+            "add x20, x20, %x[res_stride]\n"
+            "ble 8f\n"
+            "cmp x13, #0x3\n"
+            "st1w { z12.s }, p1, [x20]\n"
+            "add x20, x20, %x[res_stride]\n"
+            "ble 8f\n"
+            "st1w { z0.s }, p1, [x20]\n"
+            "8:"  // Row tail: Accumulator store skip
+            "subs x24, x24, #0x8\n"
+            "add %x[res_ptr], %x[res_ptr], #0x20\n"
+            "bne 6b\n"
+            "subs x13, x13, #0x4\n"
+            "add %x[a_ptr], %x[a_ptr], x12\n"
+            "mov %x[res_ptr], x23\n"
+            "bgt 5b\n"
+            "9:"  // Row tail: Row loop skip
+            : [a_ptr] "+&r" (a_ptr), [res_ptr] "+&r" (res_ptr)
+            : [b_ptr] "r" (b_ptr), [nr] "r" (nr), [nb] "r" (nb), [res_stride] "r" (res_stride), [nc] "r" (nc)
+            : "cc", "memory", "p0", "p1", "x9", "x10", "x11", "x12", "x13", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "x27", "x28", "z0", "z1", "z2", "z3", "z4", "z5", "z6", "z7", "z8", "z9", "z10", "z11", "z12", "z13", "z14", "z15", "z16", "z17", "z18", "z19", "z20", "z21", "z22", "z23", "z24", "z25", "z26", "z27", "z28", "z29", "z30", "z31"
+        );
+        return;
+    }
+#endif // #if defined(__ARM_FEATURE_SVE) && defined(__ARM_FEATURE_MATMUL_INT8)
+#elif defined(__AVX2__) || defined(__AVX512F__)
+    {
+        const block_q4_0x8 * b_ptr_start = (const block_q4_0x8 *)vx;
+        const block_q8_0x4 * a_ptr_start = (const block_q8_0x4 *)vy;
+        int64_t b_nb = n / QK4_0;
+        int64_t y = 0;
+        // Mask to mask out nibbles from packed bytes
+        const __m256i m4b = _mm256_set1_epi8(0x0F);
+        const __m128i loadMask = _mm_blend_epi32(_mm_setzero_si128(), _mm_set1_epi32(0xFFFFFFFF), 3);
+        // Lookup table to convert signed nibbles to signed bytes
+        __m256i signextendlut = _mm256_castsi128_si256(_mm_set_epi8(-1, -2, -3, -4, -5, -6, -7, -8, 7, 6, 5, 4, 3, 2, 1, 0));
+        signextendlut = _mm256_permute2f128_si256(signextendlut, signextendlut, 0);
+        // Permute mask used for easier vector processing at later stages
+        __m256i requiredOrder = _mm256_set_epi32(3, 2, 1, 0, 7, 6, 5, 4);
+        int64_t xstart = 0;
+        int anr = nr - nr%16; // Used to align nr with boundary of 16
+    #ifdef __AVX512F__
+        int anc = nc - nc%16; // Used to align nc with boundary of 16
+        // Mask to mask out nibbles from packed bytes expanded to 512 bit length
+        const __m512i m4bexpanded = _mm512_set1_epi8(0x0F);
+        // Lookup table to convert signed nibbles to signed bytes expanded to 512 bit length
+        __m512i signextendlutexpanded = _mm512_inserti32x8(_mm512_castsi256_si512(signextendlut), signextendlut, 1);
+
+        // Take group of four block_q8_0x4 structures at each pass of the loop and perform dot product operation
+        for (; y < anr / 4; y += 4) {
+
+            const block_q8_0x4 * a_ptrs[4];
+
+            a_ptrs[0] = a_ptr_start + (y * nb);
+            for (int i = 0; i < 3; ++i) {
+                a_ptrs[i + 1] = a_ptrs[i] + nb;
+            }
+
+            // Take group of two block_q4_0x8 structures at each pass of the loop and perform dot product operation
+            for (int64_t x = 0; x < anc / 8; x += 2) {
+
+                const block_q4_0x8 * b_ptr_0 = b_ptr_start + ((x)     * b_nb);
+                const block_q4_0x8 * b_ptr_1 = b_ptr_start + ((x + 1) * b_nb);
+
+                // Master FP accumulators
+                __m512 acc_rows[16];
+                for (int i = 0; i < 16; i++) {
+                    acc_rows[i] = _mm512_setzero_ps();
+                }
+
+                for (int64_t b = 0; b < nb; b++) {
+                    // Load the sixteen block_q4_0 quantized values interleaved with each other in chunks of eight - B0,B1 ....BE,BF
+                    const __m256i rhs_raw_mat_0123_0 = _mm256_loadu_si256((const __m256i *)(b_ptr_0[b].qs));
+                    const __m256i rhs_raw_mat_4567_0 = _mm256_loadu_si256((const __m256i *)(b_ptr_0[b].qs + 32));
+                    const __m256i rhs_raw_mat_0123_1 = _mm256_loadu_si256((const __m256i *)(b_ptr_0[b].qs + 64));
+                    const __m256i rhs_raw_mat_4567_1 = _mm256_loadu_si256((const __m256i *)(b_ptr_0[b].qs + 96));
+
+                    const __m256i rhs_raw_mat_89AB_0 = _mm256_loadu_si256((const __m256i *)(b_ptr_1[b].qs));
+                    const __m256i rhs_raw_mat_CDEF_0 = _mm256_loadu_si256((const __m256i *)(b_ptr_1[b].qs + 32));
+                    const __m256i rhs_raw_mat_89AB_1 = _mm256_loadu_si256((const __m256i *)(b_ptr_1[b].qs + 64));
+                    const __m256i rhs_raw_mat_CDEF_1 = _mm256_loadu_si256((const __m256i *)(b_ptr_1[b].qs + 96));
+
+                    // Save the values in the following vectors in the formats B0B1B4B5B8B9BCBD, B2B3B6B7BABBBEBF for further processing and storing of values
+                    const __m256i rhs_raw_mat_0145_0 = _mm256_blend_epi32(rhs_raw_mat_0123_0, _mm256_permutevar8x32_epi32(rhs_raw_mat_4567_0, requiredOrder), 240);
+                    const __m256i rhs_raw_mat_2367_0 = _mm256_blend_epi32(_mm256_permutevar8x32_epi32(rhs_raw_mat_0123_0, requiredOrder), rhs_raw_mat_4567_0, 240);
+                    const __m256i rhs_raw_mat_0145_1 = _mm256_blend_epi32(rhs_raw_mat_0123_1, _mm256_permutevar8x32_epi32(rhs_raw_mat_4567_1, requiredOrder), 240);
+                    const __m256i rhs_raw_mat_2367_1 = _mm256_blend_epi32(_mm256_permutevar8x32_epi32(rhs_raw_mat_0123_1, requiredOrder), rhs_raw_mat_4567_1, 240);
+
+                    const __m256i rhs_raw_mat_89CD_0 = _mm256_blend_epi32(rhs_raw_mat_89AB_0, _mm256_permutevar8x32_epi32(rhs_raw_mat_CDEF_0, requiredOrder), 240);
+                    const __m256i rhs_raw_mat_ABEF_0 = _mm256_blend_epi32(_mm256_permutevar8x32_epi32(rhs_raw_mat_89AB_0, requiredOrder), rhs_raw_mat_CDEF_0, 240);
+                    const __m256i rhs_raw_mat_89CD_1 = _mm256_blend_epi32(rhs_raw_mat_89AB_1, _mm256_permutevar8x32_epi32(rhs_raw_mat_CDEF_1, requiredOrder), 240);
+                    const __m256i rhs_raw_mat_ABEF_1 = _mm256_blend_epi32(_mm256_permutevar8x32_epi32(rhs_raw_mat_89AB_1, requiredOrder), rhs_raw_mat_CDEF_1, 240);
+
+                    const __m512i rhs_raw_mat_014589CD_0 = _mm512_inserti32x8(_mm512_castsi256_si512(rhs_raw_mat_0145_0), rhs_raw_mat_89CD_0, 1);
+                    const __m512i rhs_raw_mat_2367ABEF_0 = _mm512_inserti32x8(_mm512_castsi256_si512(rhs_raw_mat_2367_0), rhs_raw_mat_ABEF_0, 1);
+                    const __m512i rhs_raw_mat_014589CD_1 = _mm512_inserti32x8(_mm512_castsi256_si512(rhs_raw_mat_0145_1), rhs_raw_mat_89CD_1, 1);
+                    const __m512i rhs_raw_mat_2367ABEF_1 = _mm512_inserti32x8(_mm512_castsi256_si512(rhs_raw_mat_2367_1), rhs_raw_mat_ABEF_1, 1);
+
+                    // 4-bit -> 8-bit - Sign is maintained
+                    const __m512i rhs_mat_014589CD_0 = _mm512_shuffle_epi8(signextendlutexpanded, _mm512_and_si512(rhs_raw_mat_014589CD_0, m4bexpanded)); //B0(0-7) B1(0-7) B4(0-7) B5(0-7) B8(0-7) B9(0-7) BC(0-7) BD(0-7)
+                    const __m512i rhs_mat_2367ABEF_0 = _mm512_shuffle_epi8(signextendlutexpanded, _mm512_and_si512(rhs_raw_mat_2367ABEF_0, m4bexpanded)); //B2(0-7) B3(0-7) B6(0-7) B7(0-7) BA(0-7) BB(0-7) BE(0-7) BF(0-7)
+
+                    const __m512i rhs_mat_014589CD_1 = _mm512_shuffle_epi8(signextendlutexpanded, _mm512_and_si512(rhs_raw_mat_014589CD_1, m4bexpanded)); //B0(8-15) B1(8-15) B4(8-15) B5(8-15) B8(8-15) B9(8-15) BC(8-15) BD(8-15)
+                    const __m512i rhs_mat_2367ABEF_1 = _mm512_shuffle_epi8(signextendlutexpanded, _mm512_and_si512(rhs_raw_mat_2367ABEF_1, m4bexpanded)); //B2(8-15) B3(8-15) B6(8-15) B7(8-15) BA(8-15) BB(8-15) BE(8-15) BF(8-15)
+
+                    const __m512i rhs_mat_014589CD_2 = _mm512_shuffle_epi8(signextendlutexpanded, _mm512_and_si512(_mm512_srli_epi16(rhs_raw_mat_014589CD_0, 4), m4bexpanded)); //B0(16-23) B1(16-23) B4(16-23) B5(16-23) B8(16-23) B9(16-23) BC(16-23) BD(16-23)
+                    const __m512i rhs_mat_2367ABEF_2 = _mm512_shuffle_epi8(signextendlutexpanded, _mm512_and_si512(_mm512_srli_epi16(rhs_raw_mat_2367ABEF_0, 4), m4bexpanded)); //B2(16-23) B3(16-23) B6(16-23) B7(16-23) BA(16-23) BB(16-23) BE(16-23) BF(16-23)
+
+                    const __m512i rhs_mat_014589CD_3 = _mm512_shuffle_epi8(signextendlutexpanded, _mm512_and_si512(_mm512_srli_epi16(rhs_raw_mat_014589CD_1, 4), m4bexpanded)); //B0(24-31) B1(24-31) B4(24-31) B5(24-31) B8(24-31) B9(24-31) BC(24-31) BD(24-31)
+                    const __m512i rhs_mat_2367ABEF_3 = _mm512_shuffle_epi8(signextendlutexpanded, _mm512_and_si512(_mm512_srli_epi16(rhs_raw_mat_2367ABEF_1, 4), m4bexpanded)); //B2(24-31) B3(24-31) B6(24-31) B7(24-31) BA(24-31) BB(24-31) BE(24-31) BF(24-31)
+
+                    // Shuffle pattern one - right side input
+                    const __m512i rhs_mat_014589CD_0_sp1 = _mm512_shuffle_epi32(rhs_mat_014589CD_0, 136); //B0(0-3) B1(0-3) B0(0-3) B1(0-3) B4(0-3) B5(0-3) B4(0-3) B5(0-3) B8(0-3) B9(0-3) B8(0-3) B9(0-3) BC(0-3) BD(0-3) BC(0-3) BD(0-3)
+                    const __m512i rhs_mat_2367ABEF_0_sp1 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_0, 136); //B2(0-3) B3(0-3) B2(0-3) B3(0-3) B6(0-3) B7(0-3) B6(0-3) B7(0-3) BA(0-3) BB(0-3) BA(0-3) BB(0-3) BE(0-3) BF(0-3) BE(0-3) BF(0-3)
+
+                    const __m512i rhs_mat_014589CD_1_sp1 = _mm512_shuffle_epi32(rhs_mat_014589CD_1, 136); //B0(8-11) B1(8-11) B0(8-11) B1(8-11) B4(8-11) B5(8-11) B4(8-11) B5(8-11) B8(8-11) B9(8-11) B8(8-11) B9(8-11) BC(8-11) BD(8-11) BC(8-11) BD(8-11)
+                    const __m512i rhs_mat_2367ABEF_1_sp1 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_1, 136); //B2(8-11) B3(8-11) B2(8-11) B3(8-11) B6(8-11) B7(8-11) B6(8-11) B7(8-11) BA(8-11) BB(8-11) BA(8-11) BB(8-11) BE(8-11) BF(8-11) BE(8-11) BF(8-11)
+
+                    const __m512i rhs_mat_014589CD_2_sp1 = _mm512_shuffle_epi32(rhs_mat_014589CD_2, 136); //B0(16-19) B1(16-19) B0(16-19) B1(16-19) B4(16-19) B5(16-19) B4(16-19) B5(16-19) B8(16-19) B9(16-19) B8(16-19) B9(16-19) BC(16-19) BD(16-19) BC(16-19) BD(16-19)
+                    const __m512i rhs_mat_2367ABEF_2_sp1 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_2, 136); //B2(16-19) B3(16-19) B2(16-19) B3(16-19) B6(16-19) B7(16-19) B6(16-19) B7(16-19) BA(16-19) BB(16-19) BA(16-19) BB(16-19) BE(16-19) BF(16-19) BE(16-19) BF(16-19)
+
+                    const __m512i rhs_mat_014589CD_3_sp1 = _mm512_shuffle_epi32(rhs_mat_014589CD_3, 136); //B0(24-27) B1(24-27) B0(24-27) B1(24-27) B4(24-27) B5(24-27) B4(24-27) B5(24-27) B8(24-27) B9(24-27) B8(24-27) B9(24-27) BC(24-27) BD(24-27) BC(24-27) BD(24-27)
+                    const __m512i rhs_mat_2367ABEF_3_sp1 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_3, 136); //B2(24-27) B3(24-27) B2(24-27) B3(24-27) B6(24-27) B7(24-27) B6(24-27) B7(24-27) BA(24-27) BB(24-27) BA(24-27) BB(24-27) BE(24-27) BF(24-27) BE(24-27) BF(24-27)
+
+                    // Shuffle pattern two - right side input
+
+                    const __m512i rhs_mat_014589CD_0_sp2 = _mm512_shuffle_epi32(rhs_mat_014589CD_0, 221); //B0(4-7) B1(4-7) B0(4-7) B1(4-7) B4(4-7) B5(4-7) B4(4-7) B5(4-7) B8(4-7) B9(4-7) B8(4-7) B9(4-7) BC(4-7) BD(4-7) BC(4-7) BD(4-7)
+                    const __m512i rhs_mat_2367ABEF_0_sp2 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_0, 221); //B2(4-7) B3(4-7) B2(4-7) B3(4-7) B6(4-7) B7(4-7) B6(4-7) B7(4-7) BA(4-7) BB(4-7) BA(4-7) BB(4-7) BE(4-7) BF(4-7) BE(4-7) BF(4-7)
+
+                    const __m512i rhs_mat_014589CD_1_sp2 = _mm512_shuffle_epi32(rhs_mat_014589CD_1, 221); //B0(12-15) B1(12-15) B0(12-15) B1(12-15) B4(12-15) B5(12-15) B4(12-15) B5(12-15) B8(12-15) B9(12-15) B8(12-15) B9(12-15) BC(12-15) BD(12-15) BC(12-15) BD(12-15)
+                    const __m512i rhs_mat_2367ABEF_1_sp2 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_1, 221); //B2(12-15) B3(12-15) B2(12-15) B3(12-15) B6(12-15) B7(12-15) B6(12-15) B7(12-15) BA(12-15) BB(12-15) BA(12-15) BB(12-15) BE(12-15) BF(12-15) BE(12-15) BF(12-15)
+
+                    const __m512i rhs_mat_014589CD_2_sp2 = _mm512_shuffle_epi32(rhs_mat_014589CD_2, 221); //B0(20-23) B1(20-23) B0(20-23) B1(20-23) B4(20-23) B5(20-23) B4(20-23) B5(20-23) B8(20-23) B9(20-23) B8(20-23) B9(20-23) BC(20-23) BD(20-23) BC(20-23) BD(20-23)
+                    const __m512i rhs_mat_2367ABEF_2_sp2 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_2, 221); //B2(20-23) B3(20-23) B2(20-23) B3(20-23) B6(20-23) B7(20-23) B6(20-23) B7(20-23) BA(20-23) BB(20-23) BA(20-23) BB(20-23) BE(20-23) BF(20-23) BE(20-23) BF(20-23)
+
+                    const __m512i rhs_mat_014589CD_3_sp2 = _mm512_shuffle_epi32(rhs_mat_014589CD_3, 221); //B0(28-31) B1(28-31) B0(28-31) B1(28-31) B4(28-31) B5(28-31) B4(28-31) B5(28-31) B8(28-31) B9(28-31) B8(28-31) B9(28-31) BC(28-31) BD(28-31) BC(28-31) BD(28-31)
+                    const __m512i rhs_mat_2367ABEF_3_sp2 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_3, 221); //B2(28-31) B3(28-31) B2(28-31) B3(28-31) B6(28-31) B7(28-31) B6(28-31) B7(28-31) BA(28-31) BB(28-31) BA(28-31) BB(28-31) BE(28-31) BF(28-31) BE(28-31) BF(28-31)
+
+                    // Scale values - Load the weight scale values of two block_q4_0x8
+                    const __m512 col_scale_f32 = GGML_F32Cx8x2_LOAD(b_ptr_0[b].d, b_ptr_1[b].d);
+
+                    // Process LHS in pairs of rows
+                    for (int rp = 0; rp < 4; rp++) {
+
+                        // Load the four block_q4_0 quantized values interleaved with each other in chunks of eight - A0,A1,A2,A3
+                        // Loaded as set of 128 bit vectors and repeated and stored into a 256 bit vector before again repeating into 512 bit vector
+                        __m256i lhs_mat_ymm_0123_0 = _mm256_loadu_si256((const __m256i *)((a_ptrs[rp][b].qs)));
+                        __m256i lhs_mat_ymm_01_0 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_0, lhs_mat_ymm_0123_0, 0);
+                        __m256i lhs_mat_ymm_23_0 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_0, lhs_mat_ymm_0123_0, 17);
+                        __m256i lhs_mat_ymm_0123_1 = _mm256_loadu_si256((const __m256i *)((a_ptrs[rp][b].qs + 32)));
+                        __m256i lhs_mat_ymm_01_1 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_1, lhs_mat_ymm_0123_1, 0);
+                        __m256i lhs_mat_ymm_23_1 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_1, lhs_mat_ymm_0123_1, 17);
+                        __m256i lhs_mat_ymm_0123_2 = _mm256_loadu_si256((const __m256i *)((a_ptrs[rp][b].qs + 64)));
+                        __m256i lhs_mat_ymm_01_2 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_2, lhs_mat_ymm_0123_2, 0);
+                        __m256i lhs_mat_ymm_23_2 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_2, lhs_mat_ymm_0123_2, 17);
+                        __m256i lhs_mat_ymm_0123_3 = _mm256_loadu_si256((const __m256i *)((a_ptrs[rp][b].qs + 96)));
+                        __m256i lhs_mat_ymm_01_3 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_3, lhs_mat_ymm_0123_3, 0);
+                        __m256i lhs_mat_ymm_23_3 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_3, lhs_mat_ymm_0123_3, 17);
+
+                        __m512i lhs_mat_01_0 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_01_0), lhs_mat_ymm_01_0, 1);
+                        __m512i lhs_mat_23_0 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_23_0), lhs_mat_ymm_23_0, 1);
+                        __m512i lhs_mat_01_1 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_01_1), lhs_mat_ymm_01_1, 1);
+                        __m512i lhs_mat_23_1 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_23_1), lhs_mat_ymm_23_1, 1);
+                        __m512i lhs_mat_01_2 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_01_2), lhs_mat_ymm_01_2, 1);
+                        __m512i lhs_mat_23_2 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_23_2), lhs_mat_ymm_23_2, 1);
+                        __m512i lhs_mat_01_3 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_01_3), lhs_mat_ymm_01_3, 1);
+                        __m512i lhs_mat_23_3 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_23_3), lhs_mat_ymm_23_3, 1);
+
+                        // Shuffle pattern one - left side input
+
+                        const __m512i lhs_mat_01_0_sp1 = _mm512_shuffle_epi32(lhs_mat_01_0, 160);  //A0(0-3) A0(0-3) A1(0-3) A1(0-3) A0(0-3) A0(0-3) A1(0-3) A1(0-3) A0(0-3) A0(0-3) A1(0-3) A1(0-3) A0(0-3) A0(0-3) A1(0-3) A1(0-3)
+                        const __m512i lhs_mat_23_0_sp1 = _mm512_shuffle_epi32(lhs_mat_23_0, 160);  //A2(0-3) A2(0-3) A3(0-3) A3(0-3) A2(0-3) A2(0-3) A3(0-3) A3(0-3) A2(0-3) A2(0-3) A3(0-3) A3(0-3) A2(0-3) A2(0-3) A3(0-3) A3(0-3)
+
+                        const __m512i lhs_mat_01_1_sp1 = _mm512_shuffle_epi32(lhs_mat_01_1, 160);  //A0(8-11) A0(8-11) A1(8-11) A1(8-11) A0(8-11) A0(8-11) A1(8-11) A1(8-11) A0(8-11) A0(8-11) A1(8-11) A1(8-11) A0(8-11) A0(8-11) A1(8-11) A1(8-11)
+                        const __m512i lhs_mat_23_1_sp1 = _mm512_shuffle_epi32(lhs_mat_23_1, 160);  //A2(8-11) A2(8-11) A3(8-11) A3(8-11) A2(8-11) A2(8-11) A3(8-11) A3(8-11) A2(8-11) A2(8-11) A3(8-11) A3(8-11) A2(8-11) A2(8-11) A3(8-11) A3(8-11)
+
+                        const __m512i lhs_mat_01_2_sp1 = _mm512_shuffle_epi32(lhs_mat_01_2, 160);  //A0(16-19) A0(16-19) A1(16-19) A1(16-19) A0(16-19) A0(16-19) A1(16-19) A1(16-19) A0(16-19) A0(16-19) A1(16-19) A1(16-19) A0(16-19) A0(16-19) A1(16-19) A1(16-19)
+                        const __m512i lhs_mat_23_2_sp1 = _mm512_shuffle_epi32(lhs_mat_23_2, 160);  //A2(16-19) A2(16-19) A3(16-19) A3(16-19) A2(16-19) A2(16-19) A3(16-19) A3(16-19) A2(16-19) A2(16-19) A3(16-19) A3(16-19) A2(16-19) A2(16-19) A3(16-19) A3(16-19)
+
+                        const __m512i lhs_mat_01_3_sp1 = _mm512_shuffle_epi32(lhs_mat_01_3, 160);  //A0(24-27) A0(24-27) A1(24-27) A1(24-27) A0(24-27) A0(24-27) A1(24-27) A1(24-27) A0(24-27) A0(24-27) A1(24-27) A1(24-27) A0(24-27) A0(24-27) A1(24-27) A1(24-27)
+                        const __m512i lhs_mat_23_3_sp1 = _mm512_shuffle_epi32(lhs_mat_23_3, 160);  //A2(24-27) A2(24-27) A3(24-27) A3(24-27) A2(24-27) A2(24-27) A3(24-27) A3(24-27) A2(24-27) A2(24-27) A3(24-27) A3(24-27) A2(24-27) A2(24-27) A3(24-27) A3(24-27)
+
+                        // Shuffle pattern two - left side input
+
+                        const __m512i lhs_mat_01_0_sp2 = _mm512_shuffle_epi32(lhs_mat_01_0, 245);  //A0(4-7) A0(4-7) A1(4-7) A1(4-7) A0(4-7) A0(4-7) A1(4-7) A1(4-7) A0(4-7) A0(4-7) A1(4-7) A1(4-7) A0(4-7) A0(4-7) A1(4-7) A1(4-7)
+                        const __m512i lhs_mat_23_0_sp2 = _mm512_shuffle_epi32(lhs_mat_23_0, 245);  //A2(4-7) A2(4-7) A3(4-7) A3(4-7) A2(4-7) A2(4-7) A3(4-7) A3(4-7) A2(4-7) A2(4-7) A3(4-7) A3(4-7) A2(4-7) A2(4-7) A3(4-7) A3(4-7)
+
+                        const __m512i lhs_mat_01_1_sp2 = _mm512_shuffle_epi32(lhs_mat_01_1, 245);  //A0(12-15) A0(12-15) A1(12-15) A1(12-15) A0(12-15) A0(12-15) A1(12-15) A1(12-15) A0(12-15) A0(12-15) A1(12-15) A1(12-15) A0(12-15) A0(12-15) A1(12-15) A1(12-15)
+                        const __m512i lhs_mat_23_1_sp2 = _mm512_shuffle_epi32(lhs_mat_23_1, 245);  //A2(12-15) A2(12-15) A3(12-15) A3(12-15) A2(12-15) A2(12-15) A3(12-15) A3(12-15) A2(12-15) A2(12-15) A3(12-15) A3(12-15) A2(12-15) A2(12-15) A3(12-15) A3(12-15)
+
+                        const __m512i lhs_mat_01_2_sp2 = _mm512_shuffle_epi32(lhs_mat_01_2, 245);  //A0(20-23) A0(20-23) A1(20-23) A1(20-23) A0(20-23) A0(20-23) A1(20-23) A1(20-23) A0(20-23) A0(20-23) A1(20-23) A1(20-23) A0(20-23) A0(20-23) A1(20-23) A1(20-23)
+                        const __m512i lhs_mat_23_2_sp2 = _mm512_shuffle_epi32(lhs_mat_23_2, 245);  //A2(20-23) A2(20-23) A3(20-23) A3(20-23) A2(20-23) A2(20-23) A3(20-23) A3(20-23) A2(20-23) A2(20-23) A3(20-23) A3(20-23) A2(20-23) A2(20-23) A3(20-23) A3(20-23)
+
+                        const __m512i lhs_mat_01_3_sp2 = _mm512_shuffle_epi32(lhs_mat_01_3, 245);  //A0(28-31) A0(28-31) A1(28-31) A1(28-31) A0(28-31) A0(28-31) A1(28-31) A1(28-31) A0(28-31) A0(28-31) A1(28-31) A1(28-31) A0(28-31) A0(28-31) A1(28-31) A1(28-31)
+                        const __m512i lhs_mat_23_3_sp2 = _mm512_shuffle_epi32(lhs_mat_23_3, 245);  //A2(28-31) A2(28-31) A3(28-31) A3(28-31) A2(28-31) A2(28-31) A3(28-31) A3(28-31) A2(28-31) A2(28-31) A3(28-31) A3(28-31) A2(28-31) A2(28-31) A3(28-31) A3(28-31)
+
+                        // The values arranged in shuffle patterns are operated with dot product operation within 32 bit lane i.e corresponding bytes and multiplied and added into 32 bit integers within 32 bit lane
+                        // Resembles MMLAs into 2x2 matrices in ARM Version
+                        __m512i iacc_mat_00_sp1 =
+                            _mm512_add_epi32(_mm512_add_epi32(_mm512_add_epi32(mul_sum_i8_pairs_int32x16(lhs_mat_01_3_sp1, rhs_mat_014589CD_3_sp1), mul_sum_i8_pairs_int32x16(lhs_mat_01_2_sp1, rhs_mat_014589CD_2_sp1)), mul_sum_i8_pairs_int32x16(lhs_mat_01_1_sp1, rhs_mat_014589CD_1_sp1)), mul_sum_i8_pairs_int32x16(lhs_mat_01_0_sp1, rhs_mat_014589CD_0_sp1));
+                        __m512i iacc_mat_01_sp1 =
+                            _mm512_add_epi32(_mm512_add_epi32(_mm512_add_epi32(mul_sum_i8_pairs_int32x16(lhs_mat_01_3_sp1, rhs_mat_2367ABEF_3_sp1), mul_sum_i8_pairs_int32x16(lhs_mat_01_2_sp1, rhs_mat_2367ABEF_2_sp1)), mul_sum_i8_pairs_int32x16(lhs_mat_01_1_sp1, rhs_mat_2367ABEF_1_sp1)), mul_sum_i8_pairs_int32x16(lhs_mat_01_0_sp1, rhs_mat_2367ABEF_0_sp1));
+                        __m512i iacc_mat_10_sp1 =
+                            _mm512_add_epi32(_mm512_add_epi32(_mm512_add_epi32(mul_sum_i8_pairs_int32x16(lhs_mat_23_3_sp1, rhs_mat_014589CD_3_sp1), mul_sum_i8_pairs_int32x16(lhs_mat_23_2_sp1, rhs_mat_014589CD_2_sp1)), mul_sum_i8_pairs_int32x16(lhs_mat_23_1_sp1, rhs_mat_014589CD_1_sp1)), mul_sum_i8_pairs_int32x16(lhs_mat_23_0_sp1, rhs_mat_014589CD_0_sp1));
+                        __m512i iacc_mat_11_sp1 =
+                            _mm512_add_epi32(_mm512_add_epi32(_mm512_add_epi32(mul_sum_i8_pairs_int32x16(lhs_mat_23_3_sp1, rhs_mat_2367ABEF_3_sp1), mul_sum_i8_pairs_int32x16(lhs_mat_23_2_sp1, rhs_mat_2367ABEF_2_sp1)), mul_sum_i8_pairs_int32x16(lhs_mat_23_1_sp1, rhs_mat_2367ABEF_1_sp1)), mul_sum_i8_pairs_int32x16(lhs_mat_23_0_sp1, rhs_mat_2367ABEF_0_sp1));
+                        __m512i iacc_mat_00_sp2 =
+                            _mm512_add_epi32(_mm512_add_epi32(_mm512_add_epi32(mul_sum_i8_pairs_int32x16(lhs_mat_01_3_sp2, rhs_mat_014589CD_3_sp2), mul_sum_i8_pairs_int32x16(lhs_mat_01_2_sp2, rhs_mat_014589CD_2_sp2)), mul_sum_i8_pairs_int32x16(lhs_mat_01_1_sp2, rhs_mat_014589CD_1_sp2)), mul_sum_i8_pairs_int32x16(lhs_mat_01_0_sp2, rhs_mat_014589CD_0_sp2));
+                        __m512i iacc_mat_01_sp2 =
+                            _mm512_add_epi32(_mm512_add_epi32(_mm512_add_epi32(mul_sum_i8_pairs_int32x16(lhs_mat_01_3_sp2, rhs_mat_2367ABEF_3_sp2), mul_sum_i8_pairs_int32x16(lhs_mat_01_2_sp2, rhs_mat_2367ABEF_2_sp2)), mul_sum_i8_pairs_int32x16(lhs_mat_01_1_sp2, rhs_mat_2367ABEF_1_sp2)), mul_sum_i8_pairs_int32x16(lhs_mat_01_0_sp2, rhs_mat_2367ABEF_0_sp2));
+                        __m512i iacc_mat_10_sp2 =
+                            _mm512_add_epi32(_mm512_add_epi32(_mm512_add_epi32(mul_sum_i8_pairs_int32x16(lhs_mat_23_3_sp2, rhs_mat_014589CD_3_sp2), mul_sum_i8_pairs_int32x16(lhs_mat_23_2_sp2, rhs_mat_014589CD_2_sp2)), mul_sum_i8_pairs_int32x16(lhs_mat_23_1_sp2, rhs_mat_014589CD_1_sp2)), mul_sum_i8_pairs_int32x16(lhs_mat_23_0_sp2, rhs_mat_014589CD_0_sp2));
+                        __m512i iacc_mat_11_sp2 =
+                            _mm512_add_epi32(_mm512_add_epi32(_mm512_add_epi32(mul_sum_i8_pairs_int32x16(lhs_mat_23_3_sp2, rhs_mat_2367ABEF_3_sp2), mul_sum_i8_pairs_int32x16(lhs_mat_23_2_sp2, rhs_mat_2367ABEF_2_sp2)), mul_sum_i8_pairs_int32x16(lhs_mat_23_1_sp2, rhs_mat_2367ABEF_1_sp2)), mul_sum_i8_pairs_int32x16(lhs_mat_23_0_sp2, rhs_mat_2367ABEF_0_sp2));
+
+                        // Output of both shuffle patterns are added in order to sum dot product outputs of all 32 values in block
+                        __m512i iacc_mat_00 = _mm512_add_epi32(iacc_mat_00_sp1, iacc_mat_00_sp2);
+                        __m512i iacc_mat_01 = _mm512_add_epi32(iacc_mat_01_sp1, iacc_mat_01_sp2);
+                        __m512i iacc_mat_10 = _mm512_add_epi32(iacc_mat_10_sp1, iacc_mat_10_sp2);
+                        __m512i iacc_mat_11 = _mm512_add_epi32(iacc_mat_11_sp1, iacc_mat_11_sp2);
+
+
+                        // Straighten out to make 4 row vectors
+                        __m512i iacc_row_0 = _mm512_mask_blend_epi32(0xCCCC, iacc_mat_00, _mm512_shuffle_epi32(iacc_mat_01, 78));
+                        __m512i iacc_row_1 = _mm512_mask_blend_epi32(0xCCCC, _mm512_shuffle_epi32(iacc_mat_00, 78), iacc_mat_01);
+                        __m512i iacc_row_2 = _mm512_mask_blend_epi32(0xCCCC, iacc_mat_10, _mm512_shuffle_epi32(iacc_mat_11, 78));
+                        __m512i iacc_row_3 = _mm512_mask_blend_epi32(0xCCCC, _mm512_shuffle_epi32(iacc_mat_10, 78), iacc_mat_11);
+
+                        // Load the scale(d) values for all the 4 Q8_0 blocks and repeat it across lanes
+                        const __m128i row_scale_f16 = _mm_shuffle_epi32(_mm_maskload_epi32((int const*)(a_ptrs[rp][b].d), loadMask), 68);
+                        const __m512 row_scale_f32 = GGML_F32Cx16_REPEAT_LOAD(row_scale_f16);
+
+                        // Multiply with appropiate scales and accumulate
+                        acc_rows[rp * 4]     = _mm512_fmadd_ps(_mm512_cvtepi32_ps(iacc_row_0), _mm512_mul_ps(col_scale_f32, _mm512_shuffle_ps(row_scale_f32, row_scale_f32, 0)),   acc_rows[rp * 4]);
+                        acc_rows[rp * 4 + 1] = _mm512_fmadd_ps(_mm512_cvtepi32_ps(iacc_row_1), _mm512_mul_ps(col_scale_f32, _mm512_shuffle_ps(row_scale_f32, row_scale_f32, 85)),  acc_rows[rp * 4 + 1]);
+                        acc_rows[rp * 4 + 2] = _mm512_fmadd_ps(_mm512_cvtepi32_ps(iacc_row_2), _mm512_mul_ps(col_scale_f32, _mm512_shuffle_ps(row_scale_f32, row_scale_f32, 170)), acc_rows[rp * 4 + 2]);
+                        acc_rows[rp * 4 + 3] = _mm512_fmadd_ps(_mm512_cvtepi32_ps(iacc_row_3), _mm512_mul_ps(col_scale_f32, _mm512_shuffle_ps(row_scale_f32, row_scale_f32, 255)), acc_rows[rp * 4 + 3]);
+                    }
+                }
+
+                // Store the accumulated values
+                for (int i = 0; i < 16; i++) {
+                    _mm512_storeu_ps((float *)(s + ((y * 4 + i) * bs + x * 8)), acc_rows[i]);
+                }
+            }
+        }
+        // Take a block_q8_0x4 structures at each pass of the loop and perform dot product operation
+        for (; y < nr / 4; y ++) {
+
+            const block_q8_0x4 * a_ptr = a_ptr_start + (y * nb);
+
+            // Take group of two block_q4_0x8 structures at each pass of the loop and perform dot product operation
+            for (int64_t x = 0; x < anc / 8; x += 2) {
+
+                const block_q4_0x8 * b_ptr_0 = b_ptr_start + ((x)     * b_nb);
+                const block_q4_0x8 * b_ptr_1 = b_ptr_start + ((x + 1) * b_nb);
+
+                // Master FP accumulators
+                __m512 acc_rows[4];
+                for (int i = 0; i < 4; i++) {
+                    acc_rows[i] = _mm512_setzero_ps();
+                }
+
+                for (int64_t b = 0; b < nb; b++) {
+                    // Load the sixteen block_q4_0 quantized values interleaved with each other in chunks of eight - B0,B1 ....BE,BF
+                    const __m256i rhs_raw_mat_0123_0 = _mm256_loadu_si256((const __m256i *)(b_ptr_0[b].qs));
+                    const __m256i rhs_raw_mat_4567_0 = _mm256_loadu_si256((const __m256i *)(b_ptr_0[b].qs + 32));
+                    const __m256i rhs_raw_mat_0123_1 = _mm256_loadu_si256((const __m256i *)(b_ptr_0[b].qs + 64));
+                    const __m256i rhs_raw_mat_4567_1 = _mm256_loadu_si256((const __m256i *)(b_ptr_0[b].qs + 96));
+
+                    const __m256i rhs_raw_mat_89AB_0 = _mm256_loadu_si256((const __m256i *)(b_ptr_1[b].qs));
+                    const __m256i rhs_raw_mat_CDEF_0 = _mm256_loadu_si256((const __m256i *)(b_ptr_1[b].qs + 32));
+                    const __m256i rhs_raw_mat_89AB_1 = _mm256_loadu_si256((const __m256i *)(b_ptr_1[b].qs + 64));
+                    const __m256i rhs_raw_mat_CDEF_1 = _mm256_loadu_si256((const __m256i *)(b_ptr_1[b].qs + 96));
+
+                    // Save the values in the following vectors in the formats B0B1B4B5, B2B3B6B7 for further processing and storing of valuess
+                    const __m256i rhs_raw_mat_0145_0 = _mm256_blend_epi32(rhs_raw_mat_0123_0, _mm256_permutevar8x32_epi32(rhs_raw_mat_4567_0, requiredOrder), 240);
+                    const __m256i rhs_raw_mat_2367_0 = _mm256_blend_epi32(_mm256_permutevar8x32_epi32(rhs_raw_mat_0123_0, requiredOrder), rhs_raw_mat_4567_0, 240);
+                    const __m256i rhs_raw_mat_0145_1 = _mm256_blend_epi32(rhs_raw_mat_0123_1, _mm256_permutevar8x32_epi32(rhs_raw_mat_4567_1, requiredOrder), 240);
+                    const __m256i rhs_raw_mat_2367_1 = _mm256_blend_epi32(_mm256_permutevar8x32_epi32(rhs_raw_mat_0123_1, requiredOrder), rhs_raw_mat_4567_1, 240);
+
+                    const __m256i rhs_raw_mat_89CD_0 = _mm256_blend_epi32(rhs_raw_mat_89AB_0, _mm256_permutevar8x32_epi32(rhs_raw_mat_CDEF_0, requiredOrder), 240);
+                    const __m256i rhs_raw_mat_ABEF_0 = _mm256_blend_epi32(_mm256_permutevar8x32_epi32(rhs_raw_mat_89AB_0, requiredOrder), rhs_raw_mat_CDEF_0, 240);
+                    const __m256i rhs_raw_mat_89CD_1 = _mm256_blend_epi32(rhs_raw_mat_89AB_1, _mm256_permutevar8x32_epi32(rhs_raw_mat_CDEF_1, requiredOrder), 240);
+                    const __m256i rhs_raw_mat_ABEF_1 = _mm256_blend_epi32(_mm256_permutevar8x32_epi32(rhs_raw_mat_89AB_1, requiredOrder), rhs_raw_mat_CDEF_1, 240);
+
+                    const __m512i rhs_raw_mat_014589CD_0 = _mm512_inserti32x8(_mm512_castsi256_si512(rhs_raw_mat_0145_0), rhs_raw_mat_89CD_0, 1);
+                    const __m512i rhs_raw_mat_2367ABEF_0 = _mm512_inserti32x8(_mm512_castsi256_si512(rhs_raw_mat_2367_0), rhs_raw_mat_ABEF_0, 1);
+                    const __m512i rhs_raw_mat_014589CD_1 = _mm512_inserti32x8(_mm512_castsi256_si512(rhs_raw_mat_0145_1), rhs_raw_mat_89CD_1, 1);
+                    const __m512i rhs_raw_mat_2367ABEF_1 = _mm512_inserti32x8(_mm512_castsi256_si512(rhs_raw_mat_2367_1), rhs_raw_mat_ABEF_1, 1);
+
+                    // 4-bit -> 8-bit - Sign is maintained
+                    const __m512i rhs_mat_014589CD_0 = _mm512_shuffle_epi8(signextendlutexpanded, _mm512_and_si512(rhs_raw_mat_014589CD_0, m4bexpanded)); //B0(0-7) B1(0-7) B4(0-7) B5(0-7) B8(0-7) B9(0-7) BC(0-7) BD(0-7)
+                    const __m512i rhs_mat_2367ABEF_0 = _mm512_shuffle_epi8(signextendlutexpanded, _mm512_and_si512(rhs_raw_mat_2367ABEF_0, m4bexpanded)); //B2(0-7) B3(0-7) B6(0-7) B7(0-7) BA(0-7) BB(0-7) BE(0-7) BF(0-7)
+
+                    const __m512i rhs_mat_014589CD_1 = _mm512_shuffle_epi8(signextendlutexpanded, _mm512_and_si512(rhs_raw_mat_014589CD_1, m4bexpanded)); //B0(8-15) B1(8-15) B4(8-15) B5(8-15) B8(8-15) B9(8-15) BC(8-15) BD(8-15)
+                    const __m512i rhs_mat_2367ABEF_1 = _mm512_shuffle_epi8(signextendlutexpanded, _mm512_and_si512(rhs_raw_mat_2367ABEF_1, m4bexpanded)); //B2(8-15) B3(8-15) B6(8-15) B7(8-15) BA(8-15) BB(8-15) BE(8-15) BF(8-15)
+
+                    const __m512i rhs_mat_014589CD_2 = _mm512_shuffle_epi8(signextendlutexpanded, _mm512_and_si512(_mm512_srli_epi16(rhs_raw_mat_014589CD_0, 4), m4bexpanded)); //B0(16-23) B1(16-23) B4(16-23) B5(16-23) B8(16-23) B9(16-23) BC(16-23) BD(16-23)
+                    const __m512i rhs_mat_2367ABEF_2 = _mm512_shuffle_epi8(signextendlutexpanded, _mm512_and_si512(_mm512_srli_epi16(rhs_raw_mat_2367ABEF_0, 4), m4bexpanded)); //B2(16-23) B3(16-23) B6(16-23) B7(16-23) BA(16-23) BB(16-23) BE(16-23) BF(16-23)
+
+                    const __m512i rhs_mat_014589CD_3 = _mm512_shuffle_epi8(signextendlutexpanded, _mm512_and_si512(_mm512_srli_epi16(rhs_raw_mat_014589CD_1, 4), m4bexpanded)); //B0(24-31) B1(24-31) B4(24-31) B5(24-31) B8(24-31) B9(24-31) BC(24-31) BD(24-31)
+                    const __m512i rhs_mat_2367ABEF_3 = _mm512_shuffle_epi8(signextendlutexpanded, _mm512_and_si512(_mm512_srli_epi16(rhs_raw_mat_2367ABEF_1, 4), m4bexpanded)); //B2(24-31) B3(24-31) B6(24-31) B7(24-31) BA(24-31) BB(24-31) BE(24-31) BF(24-31)
+
+                    // Shuffle pattern one - right side input
+                    const __m512i rhs_mat_014589CD_0_sp1 = _mm512_shuffle_epi32(rhs_mat_014589CD_0, 136); //B0(0-3) B1(0-3) B0(0-3) B1(0-3) B4(0-3) B5(0-3) B4(0-3) B5(0-3) B8(0-3) B9(0-3) B8(0-3) B9(0-3) BC(0-3) BD(0-3) BC(0-3) BD(0-3)
+                    const __m512i rhs_mat_2367ABEF_0_sp1 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_0, 136); //B2(0-3) B3(0-3) B2(0-3) B3(0-3) B6(0-3) B7(0-3) B6(0-3) B7(0-3) BA(0-3) BB(0-3) BA(0-3) BB(0-3) BE(0-3) BF(0-3) BE(0-3) BF(0-3)
+
+                    const __m512i rhs_mat_014589CD_1_sp1 = _mm512_shuffle_epi32(rhs_mat_014589CD_1, 136); //B0(8-11) B1(8-11) B0(8-11) B1(8-11) B4(8-11) B5(8-11) B4(8-11) B5(8-11) B8(8-11) B9(8-11) B8(8-11) B9(8-11) BC(8-11) BD(8-11) BC(8-11) BD(8-11)
+                    const __m512i rhs_mat_2367ABEF_1_sp1 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_1, 136); //B2(8-11) B3(8-11) B2(8-11) B3(8-11) B6(8-11) B7(8-11) B6(8-11) B7(8-11) BA(8-11) BB(8-11) BA(8-11) BB(8-11) BE(8-11) BF(8-11) BE(8-11) BF(8-11)
+
+                    const __m512i rhs_mat_014589CD_2_sp1 = _mm512_shuffle_epi32(rhs_mat_014589CD_2, 136); //B0(16-19) B1(16-19) B0(16-19) B1(16-19) B4(16-19) B5(16-19) B4(16-19) B5(16-19) B8(16-19) B9(16-19) B8(16-19) B9(16-19) BC(16-19) BD(16-19) BC(16-19) BD(16-19)
+                    const __m512i rhs_mat_2367ABEF_2_sp1 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_2, 136); //B2(16-19) B3(16-19) B2(16-19) B3(16-19) B6(16-19) B7(16-19) B6(16-19) B7(16-19) BA(16-19) BB(16-19) BA(16-19) BB(16-19) BE(16-19) BF(16-19) BE(16-19) BF(16-19)
+
+                    const __m512i rhs_mat_014589CD_3_sp1 = _mm512_shuffle_epi32(rhs_mat_014589CD_3, 136); //B0(24-27) B1(24-27) B0(24-27) B1(24-27) B4(24-27) B5(24-27) B4(24-27) B5(24-27) B8(24-27) B9(24-27) B8(24-27) B9(24-27) BC(24-27) BD(24-27) BC(24-27) BD(24-27)
+                    const __m512i rhs_mat_2367ABEF_3_sp1 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_3, 136); //B2(24-27) B3(24-27) B2(24-27) B3(24-27) B6(24-27) B7(24-27) B6(24-27) B7(24-27) BA(24-27) BB(24-27) BA(24-27) BB(24-27) BE(24-27) BF(24-27) BE(24-27) BF(24-27)
+
+                    // Shuffle pattern two - right side input
+
+                    const __m512i rhs_mat_014589CD_0_sp2 = _mm512_shuffle_epi32(rhs_mat_014589CD_0, 221); //B0(4-7) B1(4-7) B0(4-7) B1(4-7) B4(4-7) B5(4-7) B4(4-7) B5(4-7) B8(4-7) B9(4-7) B8(4-7) B9(4-7) BC(4-7) BD(4-7) BC(4-7) BD(4-7)
+                    const __m512i rhs_mat_2367ABEF_0_sp2 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_0, 221); //B2(4-7) B3(4-7) B2(4-7) B3(4-7) B6(4-7) B7(4-7) B6(4-7) B7(4-7) BA(4-7) BB(4-7) BA(4-7) BB(4-7) BE(4-7) BF(4-7) BE(4-7) BF(4-7)
+
+                    const __m512i rhs_mat_014589CD_1_sp2 = _mm512_shuffle_epi32(rhs_mat_014589CD_1, 221); //B0(12-15) B1(12-15) B0(12-15) B1(12-15) B4(12-15) B5(12-15) B4(12-15) B5(12-15) B8(12-15) B9(12-15) B8(12-15) B9(12-15) BC(12-15) BD(12-15) BC(12-15) BD(12-15)
+                    const __m512i rhs_mat_2367ABEF_1_sp2 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_1, 221); //B2(12-15) B3(12-15) B2(12-15) B3(12-15) B6(12-15) B7(12-15) B6(12-15) B7(12-15) BA(12-15) BB(12-15) BA(12-15) BB(12-15) BE(12-15) BF(12-15) BE(12-15) BF(12-15)
+
+                    const __m512i rhs_mat_014589CD_2_sp2 = _mm512_shuffle_epi32(rhs_mat_014589CD_2, 221); //B0(20-23) B1(20-23) B0(20-23) B1(20-23) B4(20-23) B5(20-23) B4(20-23) B5(20-23) B8(20-23) B9(20-23) B8(20-23) B9(20-23) BC(20-23) BD(20-23) BC(20-23) BD(20-23)
+                    const __m512i rhs_mat_2367ABEF_2_sp2 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_2, 221); //B2(20-23) B3(20-23) B2(20-23) B3(20-23) B6(20-23) B7(20-23) B6(20-23) B7(20-23) BA(20-23) BB(20-23) BA(20-23) BB(20-23) BE(20-23) BF(20-23) BE(20-23) BF(20-23)
+
+                    const __m512i rhs_mat_014589CD_3_sp2 = _mm512_shuffle_epi32(rhs_mat_014589CD_3, 221); //B0(28-31) B1(28-31) B0(28-31) B1(28-31) B4(28-31) B5(28-31) B4(28-31) B5(28-31) B8(28-31) B9(28-31) B8(28-31) B9(28-31) BC(28-31) BD(28-31) BC(28-31) BD(28-31)
+                    const __m512i rhs_mat_2367ABEF_3_sp2 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_3, 221); //B2(28-31) B3(28-31) B2(28-31) B3(28-31) B6(28-31) B7(28-31) B6(28-31) B7(28-31) BA(28-31) BB(28-31) BA(28-31) BB(28-31) BE(28-31) BF(28-31) BE(28-31) BF(28-31)
+
+
+                    // Scale values - Load the weight scale values of two block_q4_0x8
+                    const __m512 col_scale_f32 = GGML_F32Cx8x2_LOAD(b_ptr_0[b].d, b_ptr_1[b].d);
+
+                    // Load the four block_q4_0 quantized values interleaved with each other in chunks of eight - A0,A1,A2,A3
+                    // Loaded as set of 128 bit vectors and repeated and stored into a 256 bit vector before again repeating into 512 bit vector
+                    __m256i lhs_mat_ymm_0123_0 = _mm256_loadu_si256((const __m256i *)((a_ptr[b].qs)));
+                    __m256i lhs_mat_ymm_01_0 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_0, lhs_mat_ymm_0123_0, 0);
+                    __m256i lhs_mat_ymm_23_0 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_0, lhs_mat_ymm_0123_0, 17);
+                    __m256i lhs_mat_ymm_0123_1 = _mm256_loadu_si256((const __m256i *)((a_ptr[b].qs + 32)));
+                    __m256i lhs_mat_ymm_01_1 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_1, lhs_mat_ymm_0123_1, 0);
+                    __m256i lhs_mat_ymm_23_1 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_1, lhs_mat_ymm_0123_1, 17);
+                    __m256i lhs_mat_ymm_0123_2 = _mm256_loadu_si256((const __m256i *)((a_ptr[b].qs + 64)));
+                    __m256i lhs_mat_ymm_01_2 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_2, lhs_mat_ymm_0123_2, 0);
+                    __m256i lhs_mat_ymm_23_2 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_2, lhs_mat_ymm_0123_2, 17);
+                    __m256i lhs_mat_ymm_0123_3 = _mm256_loadu_si256((const __m256i *)((a_ptr[b].qs + 96)));
+                    __m256i lhs_mat_ymm_01_3 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_3, lhs_mat_ymm_0123_3, 0);
+                    __m256i lhs_mat_ymm_23_3 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_3, lhs_mat_ymm_0123_3, 17);
+
+                    __m512i lhs_mat_01_0 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_01_0), lhs_mat_ymm_01_0, 1);
+                    __m512i lhs_mat_23_0 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_23_0), lhs_mat_ymm_23_0, 1);
+                    __m512i lhs_mat_01_1 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_01_1), lhs_mat_ymm_01_1, 1);
+                    __m512i lhs_mat_23_1 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_23_1), lhs_mat_ymm_23_1, 1);
+                    __m512i lhs_mat_01_2 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_01_2), lhs_mat_ymm_01_2, 1);
+                    __m512i lhs_mat_23_2 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_23_2), lhs_mat_ymm_23_2, 1);
+                    __m512i lhs_mat_01_3 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_01_3), lhs_mat_ymm_01_3, 1);
+                    __m512i lhs_mat_23_3 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_23_3), lhs_mat_ymm_23_3, 1);
+
+                    // Shuffle pattern one - left side input
+
+                    const __m512i lhs_mat_01_0_sp1 = _mm512_shuffle_epi32(lhs_mat_01_0, 160);  //A0(0-3) A0(0-3) A1(0-3) A1(0-3) A0(0-3) A0(0-3) A1(0-3) A1(0-3) A0(0-3) A0(0-3) A1(0-3) A1(0-3) A0(0-3) A0(0-3) A1(0-3) A1(0-3)
+                    const __m512i lhs_mat_23_0_sp1 = _mm512_shuffle_epi32(lhs_mat_23_0, 160);  //A2(0-3) A2(0-3) A3(0-3) A3(0-3) A2(0-3) A2(0-3) A3(0-3) A3(0-3) A2(0-3) A2(0-3) A3(0-3) A3(0-3) A2(0-3) A2(0-3) A3(0-3) A3(0-3)
+
+                    const __m512i lhs_mat_01_1_sp1 = _mm512_shuffle_epi32(lhs_mat_01_1, 160);  //A0(8-11) A0(8-11) A1(8-11) A1(8-11) A0(8-11) A0(8-11) A1(8-11) A1(8-11) A0(8-11) A0(8-11) A1(8-11) A1(8-11) A0(8-11) A0(8-11) A1(8-11) A1(8-11)
+                    const __m512i lhs_mat_23_1_sp1 = _mm512_shuffle_epi32(lhs_mat_23_1, 160);  //A2(8-11) A2(8-11) A3(8-11) A3(8-11) A2(8-11) A2(8-11) A3(8-11) A3(8-11) A2(8-11) A2(8-11) A3(8-11) A3(8-11) A2(8-11) A2(8-11) A3(8-11) A3(8-11)
+
+                    const __m512i lhs_mat_01_2_sp1 = _mm512_shuffle_epi32(lhs_mat_01_2, 160);  //A0(16-19) A0(16-19) A1(16-19) A1(16-19) A0(16-19) A0(16-19) A1(16-19) A1(16-19) A0(16-19) A0(16-19) A1(16-19) A1(16-19) A0(16-19) A0(16-19) A1(16-19) A1(16-19)
+                    const __m512i lhs_mat_23_2_sp1 = _mm512_shuffle_epi32(lhs_mat_23_2, 160);  //A2(16-19) A2(16-19) A3(16-19) A3(16-19) A2(16-19) A2(16-19) A3(16-19) A3(16-19) A2(16-19) A2(16-19) A3(16-19) A3(16-19) A2(16-19) A2(16-19) A3(16-19) A3(16-19)
+
+                    const __m512i lhs_mat_01_3_sp1 = _mm512_shuffle_epi32(lhs_mat_01_3, 160);  //A0(24-27) A0(24-27) A1(24-27) A1(24-27) A0(24-27) A0(24-27) A1(24-27) A1(24-27) A0(24-27) A0(24-27) A1(24-27) A1(24-27) A0(24-27) A0(24-27) A1(24-27) A1(24-27)
+                    const __m512i lhs_mat_23_3_sp1 = _mm512_shuffle_epi32(lhs_mat_23_3, 160);  //A2(24-27) A2(24-27) A3(24-27) A3(24-27) A2(24-27) A2(24-27) A3(24-27) A3(24-27) A2(24-27) A2(24-27) A3(24-27) A3(24-27) A2(24-27) A2(24-27) A3(24-27) A3(24-27)
+
+                    // Shuffle pattern two - left side input
+
+                    const __m512i lhs_mat_01_0_sp2 = _mm512_shuffle_epi32(lhs_mat_01_0, 245);  //A0(4-7) A0(4-7) A1(4-7) A1(4-7) A0(4-7) A0(4-7) A1(4-7) A1(4-7) A0(4-7) A0(4-7) A1(4-7) A1(4-7) A0(4-7) A0(4-7) A1(4-7) A1(4-7)
+                    const __m512i lhs_mat_23_0_sp2 = _mm512_shuffle_epi32(lhs_mat_23_0, 245);  //A2(4-7) A2(4-7) A3(4-7) A3(4-7) A2(4-7) A2(4-7) A3(4-7) A3(4-7) A2(4-7) A2(4-7) A3(4-7) A3(4-7) A2(4-7) A2(4-7) A3(4-7) A3(4-7)
+
+                    const __m512i lhs_mat_01_1_sp2 = _mm512_shuffle_epi32(lhs_mat_01_1, 245);  //A0(12-15) A0(12-15) A1(12-15) A1(12-15) A0(12-15) A0(12-15) A1(12-15) A1(12-15) A0(12-15) A0(12-15) A1(12-15) A1(12-15) A0(12-15) A0(12-15) A1(12-15) A1(12-15)
+                    const __m512i lhs_mat_23_1_sp2 = _mm512_shuffle_epi32(lhs_mat_23_1, 245);  //A2(12-15) A2(12-15) A3(12-15) A3(12-15) A2(12-15) A2(12-15) A3(12-15) A3(12-15) A2(12-15) A2(12-15) A3(12-15) A3(12-15) A2(12-15) A2(12-15) A3(12-15) A3(12-15)
+
+                    const __m512i lhs_mat_01_2_sp2 = _mm512_shuffle_epi32(lhs_mat_01_2, 245);  //A0(20-23) A0(20-23) A1(20-23) A1(20-23) A0(20-23) A0(20-23) A1(20-23) A1(20-23) A0(20-23) A0(20-23) A1(20-23) A1(20-23) A0(20-23) A0(20-23) A1(20-23) A1(20-23)
+                    const __m512i lhs_mat_23_2_sp2 = _mm512_shuffle_epi32(lhs_mat_23_2, 245);  //A2(20-23) A2(20-23) A3(20-23) A3(20-23) A2(20-23) A2(20-23) A3(20-23) A3(20-23) A2(20-23) A2(20-23) A3(20-23) A3(20-23) A2(20-23) A2(20-23) A3(20-23) A3(20-23)
+
+                    const __m512i lhs_mat_01_3_sp2 = _mm512_shuffle_epi32(lhs_mat_01_3, 245);  //A0(28-31) A0(28-31) A1(28-31) A1(28-31) A0(28-31) A0(28-31) A1(28-31) A1(28-31) A0(28-31) A0(28-31) A1(28-31) A1(28-31) A0(28-31) A0(28-31) A1(28-31) A1(28-31)
+                    const __m512i lhs_mat_23_3_sp2 = _mm512_shuffle_epi32(lhs_mat_23_3, 245);  //A2(28-31) A2(28-31) A3(28-31) A3(28-31) A2(28-31) A2(28-31) A3(28-31) A3(28-31) A2(28-31) A2(28-31) A3(28-31) A3(28-31) A2(28-31) A2(28-31) A3(28-31) A3(28-31)
+
+                    // The values arranged in shuffle patterns are operated with dot product operation within 32 bit lane i.e corresponding bytes and multiplied and added into 32 bit integers within 32 bit lane
+                    // Resembles MMLAs into 2x2 matrices in ARM Version
+                    __m512i iacc_mat_00_sp1 =
+                        _mm512_add_epi32(_mm512_add_epi32(_mm512_add_epi32(mul_sum_i8_pairs_int32x16(lhs_mat_01_3_sp1, rhs_mat_014589CD_3_sp1), mul_sum_i8_pairs_int32x16(lhs_mat_01_2_sp1, rhs_mat_014589CD_2_sp1)), mul_sum_i8_pairs_int32x16(lhs_mat_01_1_sp1, rhs_mat_014589CD_1_sp1)), mul_sum_i8_pairs_int32x16(lhs_mat_01_0_sp1, rhs_mat_014589CD_0_sp1));
+                    __m512i iacc_mat_01_sp1 =
+                        _mm512_add_epi32(_mm512_add_epi32(_mm512_add_epi32(mul_sum_i8_pairs_int32x16(lhs_mat_01_3_sp1, rhs_mat_2367ABEF_3_sp1), mul_sum_i8_pairs_int32x16(lhs_mat_01_2_sp1, rhs_mat_2367ABEF_2_sp1)), mul_sum_i8_pairs_int32x16(lhs_mat_01_1_sp1, rhs_mat_2367ABEF_1_sp1)), mul_sum_i8_pairs_int32x16(lhs_mat_01_0_sp1, rhs_mat_2367ABEF_0_sp1));
+                    __m512i iacc_mat_10_sp1 =
+                        _mm512_add_epi32(_mm512_add_epi32(_mm512_add_epi32(mul_sum_i8_pairs_int32x16(lhs_mat_23_3_sp1, rhs_mat_014589CD_3_sp1), mul_sum_i8_pairs_int32x16(lhs_mat_23_2_sp1, rhs_mat_014589CD_2_sp1)), mul_sum_i8_pairs_int32x16(lhs_mat_23_1_sp1, rhs_mat_014589CD_1_sp1)), mul_sum_i8_pairs_int32x16(lhs_mat_23_0_sp1, rhs_mat_014589CD_0_sp1));
+                    __m512i iacc_mat_11_sp1 =
+                        _mm512_add_epi32(_mm512_add_epi32(_mm512_add_epi32(mul_sum_i8_pairs_int32x16(lhs_mat_23_3_sp1, rhs_mat_2367ABEF_3_sp1), mul_sum_i8_pairs_int32x16(lhs_mat_23_2_sp1, rhs_mat_2367ABEF_2_sp1)), mul_sum_i8_pairs_int32x16(lhs_mat_23_1_sp1, rhs_mat_2367ABEF_1_sp1)), mul_sum_i8_pairs_int32x16(lhs_mat_23_0_sp1, rhs_mat_2367ABEF_0_sp1));
+                    __m512i iacc_mat_00_sp2 =
+                        _mm512_add_epi32(_mm512_add_epi32(_mm512_add_epi32(mul_sum_i8_pairs_int32x16(lhs_mat_01_3_sp2, rhs_mat_014589CD_3_sp2), mul_sum_i8_pairs_int32x16(lhs_mat_01_2_sp2, rhs_mat_014589CD_2_sp2)), mul_sum_i8_pairs_int32x16(lhs_mat_01_1_sp2, rhs_mat_014589CD_1_sp2)), mul_sum_i8_pairs_int32x16(lhs_mat_01_0_sp2, rhs_mat_014589CD_0_sp2));
+                    __m512i iacc_mat_01_sp2 =
+                        _mm512_add_epi32(_mm512_add_epi32(_mm512_add_epi32(mul_sum_i8_pairs_int32x16(lhs_mat_01_3_sp2, rhs_mat_2367ABEF_3_sp2), mul_sum_i8_pairs_int32x16(lhs_mat_01_2_sp2, rhs_mat_2367ABEF_2_sp2)), mul_sum_i8_pairs_int32x16(lhs_mat_01_1_sp2, rhs_mat_2367ABEF_1_sp2)), mul_sum_i8_pairs_int32x16(lhs_mat_01_0_sp2, rhs_mat_2367ABEF_0_sp2));
+                    __m512i iacc_mat_10_sp2 =
+                        _mm512_add_epi32(_mm512_add_epi32(_mm512_add_epi32(mul_sum_i8_pairs_int32x16(lhs_mat_23_3_sp2, rhs_mat_014589CD_3_sp2), mul_sum_i8_pairs_int32x16(lhs_mat_23_2_sp2, rhs_mat_014589CD_2_sp2)), mul_sum_i8_pairs_int32x16(lhs_mat_23_1_sp2, rhs_mat_014589CD_1_sp2)), mul_sum_i8_pairs_int32x16(lhs_mat_23_0_sp2, rhs_mat_014589CD_0_sp2));
+                    __m512i iacc_mat_11_sp2 =
+                        _mm512_add_epi32(_mm512_add_epi32(_mm512_add_epi32(mul_sum_i8_pairs_int32x16(lhs_mat_23_3_sp2, rhs_mat_2367ABEF_3_sp2), mul_sum_i8_pairs_int32x16(lhs_mat_23_2_sp2, rhs_mat_2367ABEF_2_sp2)), mul_sum_i8_pairs_int32x16(lhs_mat_23_1_sp2, rhs_mat_2367ABEF_1_sp2)), mul_sum_i8_pairs_int32x16(lhs_mat_23_0_sp2, rhs_mat_2367ABEF_0_sp2));
+
+                    // Output of both shuffle patterns are added in order to sum dot product outputs of all 32 values in block
+                    __m512i iacc_mat_00 = _mm512_add_epi32(iacc_mat_00_sp1, iacc_mat_00_sp2);
+                    __m512i iacc_mat_01 = _mm512_add_epi32(iacc_mat_01_sp1, iacc_mat_01_sp2);
+                    __m512i iacc_mat_10 = _mm512_add_epi32(iacc_mat_10_sp1, iacc_mat_10_sp2);
+                    __m512i iacc_mat_11 = _mm512_add_epi32(iacc_mat_11_sp1, iacc_mat_11_sp2);
+
+
+                    // Straighten out to make 4 row vectors
+                    __m512i iacc_row_0 = _mm512_mask_blend_epi32(0xCCCC, iacc_mat_00, _mm512_shuffle_epi32(iacc_mat_01, 78));
+                    __m512i iacc_row_1 = _mm512_mask_blend_epi32(0xCCCC, _mm512_shuffle_epi32(iacc_mat_00, 78), iacc_mat_01);
+                    __m512i iacc_row_2 = _mm512_mask_blend_epi32(0xCCCC, iacc_mat_10, _mm512_shuffle_epi32(iacc_mat_11, 78));
+                    __m512i iacc_row_3 = _mm512_mask_blend_epi32(0xCCCC, _mm512_shuffle_epi32(iacc_mat_10, 78), iacc_mat_11);
+
+                    // Load the scale(d) values for all the 4 Q8_0 blocks and repeat it across lanes
+                    const __m128i row_scale_f16 = _mm_shuffle_epi32(_mm_maskload_epi32((int const*)(a_ptr[b].d), loadMask), 68);
+                    const __m512 row_scale_f32 = GGML_F32Cx16_REPEAT_LOAD(row_scale_f16);
+
+                    // Multiply with appropiate scales and accumulate
+                    acc_rows[0] = _mm512_fmadd_ps(_mm512_cvtepi32_ps(iacc_row_0), _mm512_mul_ps(col_scale_f32, _mm512_shuffle_ps(row_scale_f32, row_scale_f32, 0)),   acc_rows[0]);
+                    acc_rows[1] = _mm512_fmadd_ps(_mm512_cvtepi32_ps(iacc_row_1), _mm512_mul_ps(col_scale_f32, _mm512_shuffle_ps(row_scale_f32, row_scale_f32, 85)),  acc_rows[1]);
+                    acc_rows[2] = _mm512_fmadd_ps(_mm512_cvtepi32_ps(iacc_row_2), _mm512_mul_ps(col_scale_f32, _mm512_shuffle_ps(row_scale_f32, row_scale_f32, 170)), acc_rows[2]);
+                    acc_rows[3] = _mm512_fmadd_ps(_mm512_cvtepi32_ps(iacc_row_3), _mm512_mul_ps(col_scale_f32, _mm512_shuffle_ps(row_scale_f32, row_scale_f32, 255)), acc_rows[3]);
+                }
+
+                // Store the accumulated values
+                for (int i = 0; i < 4; i++) {
+                    _mm512_storeu_ps((float *)(s + ((y * 4 + i) * bs + x * 8)), acc_rows[i]);
+                }
+            }
+        }
+        if (anc != nc) {
+            xstart = anc/8;
+            y = 0;
+        }
+    #endif // __AVX512F__
+
+        // Take group of four block_q8_0x4 structures at each pass of the loop and perform dot product operation
+
+        for (; y < anr / 4; y += 4) {
+            const block_q8_0x4 * a_ptrs[4];
+
+            a_ptrs[0] = a_ptr_start + (y * nb);
+            for (int i = 0; i < 3; ++i) {
+                a_ptrs[i + 1] = a_ptrs[i] + nb;
+            }
+
+            // Take group of eight block_q4_0x8 structures at each pass of the loop and perform dot product operation
+            for (int64_t x = xstart; x < nc / 8; x++) {
+
+                const block_q4_0x8 * b_ptr = b_ptr_start + (x * b_nb);
+
+                // Master FP accumulators
+                __m256 acc_rows[16];
+                for (int i = 0; i < 16; i++) {
+                    acc_rows[i] = _mm256_setzero_ps();
+                }
+
+                for (int64_t b = 0; b < nb; b++) {
+                    // Load the eight block_q4_0 quantized values interleaved with each other in chunks of eight - B0,B1 ....B6,B7
+                    const __m256i rhs_raw_mat_0123_0 = _mm256_loadu_si256((const __m256i *)(b_ptr[b].qs));
+                    const __m256i rhs_raw_mat_4567_0 = _mm256_loadu_si256((const __m256i *)(b_ptr[b].qs + 32));
+                    const __m256i rhs_raw_mat_0123_1 = _mm256_loadu_si256((const __m256i *)(b_ptr[b].qs + 64));
+                    const __m256i rhs_raw_mat_4567_1 = _mm256_loadu_si256((const __m256i *)(b_ptr[b].qs + 96));
+
+                    // Save the values in the following vectors in the formats B0B1B4B5, B2B3B6B7 for further processing and storing of values
+                    const __m256i rhs_raw_mat_0145_0 = _mm256_blend_epi32(rhs_raw_mat_0123_0, _mm256_permutevar8x32_epi32(rhs_raw_mat_4567_0, requiredOrder), 240);
+                    const __m256i rhs_raw_mat_2367_0 = _mm256_blend_epi32(_mm256_permutevar8x32_epi32(rhs_raw_mat_0123_0, requiredOrder), rhs_raw_mat_4567_0, 240);
+                    const __m256i rhs_raw_mat_0145_1 = _mm256_blend_epi32(rhs_raw_mat_0123_1, _mm256_permutevar8x32_epi32(rhs_raw_mat_4567_1, requiredOrder), 240);
+                    const __m256i rhs_raw_mat_2367_1 = _mm256_blend_epi32(_mm256_permutevar8x32_epi32(rhs_raw_mat_0123_1, requiredOrder), rhs_raw_mat_4567_1, 240);
+
+                    // 4-bit -> 8-bit - Sign is maintained
+                    const __m256i rhs_mat_0145_0 = _mm256_shuffle_epi8(signextendlut, _mm256_and_si256(rhs_raw_mat_0145_0, m4b)); //B0(0-7) B1(0-7) B4(0-7) B5(0-7)
+                    const __m256i rhs_mat_2367_0 = _mm256_shuffle_epi8(signextendlut, _mm256_and_si256(rhs_raw_mat_2367_0, m4b)); //B2(0-7) B3(0-7) B6(0-7) B7(0-7)
+
+                    const __m256i rhs_mat_0145_1 = _mm256_shuffle_epi8(signextendlut, _mm256_and_si256(rhs_raw_mat_0145_1, m4b)); //B0(8-15) B1(8-15) B4(8-15) B5(8-15)
+                    const __m256i rhs_mat_2367_1 = _mm256_shuffle_epi8(signextendlut, _mm256_and_si256(rhs_raw_mat_2367_1, m4b)); //B2(8-15) B3(8-15) B6(8-15) B7(8-15)
+
+                    const __m256i rhs_mat_0145_2 = _mm256_shuffle_epi8(signextendlut, _mm256_and_si256(_mm256_srli_epi16(rhs_raw_mat_0145_0, 4), m4b)); //B0(16-23) B1(16-23) B4(16-23) B5(16-23)
+                    const __m256i rhs_mat_2367_2 = _mm256_shuffle_epi8(signextendlut, _mm256_and_si256(_mm256_srli_epi16(rhs_raw_mat_2367_0, 4), m4b)); //B2(16-23) B3(16-23) B6(16-23) B7(16-23)
+
+                    const __m256i rhs_mat_0145_3 = _mm256_shuffle_epi8(signextendlut, _mm256_and_si256(_mm256_srli_epi16(rhs_raw_mat_0145_1, 4), m4b)); //B0(24-31) B1(24-31) B4(24-31) B5(24-31)
+                    const __m256i rhs_mat_2367_3 = _mm256_shuffle_epi8(signextendlut, _mm256_and_si256(_mm256_srli_epi16(rhs_raw_mat_2367_1, 4), m4b)); //B2(24-31) B3(24-31) B6(24-31) B7(24-31)
+
+                    // Shuffle pattern one - right side input
+                    const __m256i rhs_mat_0145_0_sp1 = _mm256_shuffle_epi32(rhs_mat_0145_0, 136);  //B0(0-3) B1(0-3) B0(0-3) B1(0-3) B4(0-3) B5(0-3) B4(0-3) B5(0-3)
+                    const __m256i rhs_mat_2367_0_sp1 = _mm256_shuffle_epi32(rhs_mat_2367_0, 136);  //B2(0-3) B3(0-3) B2(0-3) B3(0-3) B6(0-3) B7(0-3) B6(0-3) B7(0-3)
+
+                    const __m256i rhs_mat_0145_1_sp1 = _mm256_shuffle_epi32(rhs_mat_0145_1, 136);  //B0(8-11) B1(8-11) B0(8-11) B1(8-11) B4(8-11) B5(8-11) B4(8-11) B5(8-11)
+                    const __m256i rhs_mat_2367_1_sp1 = _mm256_shuffle_epi32(rhs_mat_2367_1, 136);  //B2(8-11) B3(8-11) B2(8-11) B3(8-11) B6(8-11) B7(8-11) B6(8-11) B7(8-11)
+
+                    const __m256i rhs_mat_0145_2_sp1 = _mm256_shuffle_epi32(rhs_mat_0145_2, 136);  //B0(16-19) B1(16-19) B0(16-19) B1(16-19) B4(16-19) B5(16-19) B4(16-19) B5(16-19)
+                    const __m256i rhs_mat_2367_2_sp1 = _mm256_shuffle_epi32(rhs_mat_2367_2, 136);  //B2(16-19) B3(16-19) B2(16-19) B3(16-19) B6(16-19) B7(16-19) B6(16-19) B7(16-19)
+
+                    const __m256i rhs_mat_0145_3_sp1 = _mm256_shuffle_epi32(rhs_mat_0145_3, 136);  //B0(24-27) B1(24-27) B0(24-27) B1(24-27) B4(24-27) B5(24-27) B4(24-27) B5(24-27)
+                    const __m256i rhs_mat_2367_3_sp1 = _mm256_shuffle_epi32(rhs_mat_2367_3, 136);  //B2(24-27) B3(24-27) B2(24-27) B3(24-27) B6(24-27) B7(24-27) B6(24-27) B7(24-27)
+
+                    // Shuffle pattern two - right side input
+
+                    const __m256i rhs_mat_0145_0_sp2 = _mm256_shuffle_epi32(rhs_mat_0145_0, 221);  //B0(4-7) B1(4-7) B0(4-7) B1(4-7) B4(4-7) B5(4-7) B4(4-7) B5(4-7)
+                    const __m256i rhs_mat_2367_0_sp2 = _mm256_shuffle_epi32(rhs_mat_2367_0, 221);  //B2(4-7) B3(4-7) B2(4-7) B3(4-7) B6(4-7) B7(4-7) B6(4-7) B7(4-7)
+
+                    const __m256i rhs_mat_0145_1_sp2 = _mm256_shuffle_epi32(rhs_mat_0145_1, 221);  //B0(12-15) B1(12-15) B0(12-15) B1(12-15) B4(12-15) B5(12-15) B4(12-15) B5(12-15)
+                    const __m256i rhs_mat_2367_1_sp2 = _mm256_shuffle_epi32(rhs_mat_2367_1, 221);  //B2(12-15) B3(12-15) B2(12-15) B3(12-15) B6(12-15) B7(12-15) B6(12-15) B7(12-15)
+
+                    const __m256i rhs_mat_0145_2_sp2 = _mm256_shuffle_epi32(rhs_mat_0145_2, 221);  //B0(20-23) B1(20-23) B0(20-23) B1(20-23) B4(20-23) B5(20-23) B4(20-23) B5(20-23)
+                    const __m256i rhs_mat_2367_2_sp2 = _mm256_shuffle_epi32(rhs_mat_2367_2, 221);  //B2(20-23) B3(20-23) B2(20-23) B3(20-23) B6(20-23) B7(20-23) B6(20-23) B7(20-23)
+
+                    const __m256i rhs_mat_0145_3_sp2 = _mm256_shuffle_epi32(rhs_mat_0145_3, 221);  //B0(28-31) B1(28-31) B0(28-31) B1(28-31) B4(28-31) B5(28-31) B4(28-31) B5(28-31)
+                    const __m256i rhs_mat_2367_3_sp2 = _mm256_shuffle_epi32(rhs_mat_2367_3, 221);  //B2(28-31) B3(28-31) B2(28-31) B3(28-31) B6(28-31) B7(28-31) B6(28-31) B7(28-31)
+
+                    // Scale values - Load the wight scale values of block_q4_0x8
+                    const __m256 col_scale_f32 = GGML_F32Cx8_LOAD(b_ptr[b].d);
+
+                    // Process LHS in groups of four
+                    for (int rp = 0; rp < 4; rp++) {
+                        // Load the four block_q4_0 quantized values interleaved with each other in chunks of eight - A0,A1,A2,A3
+                        // Loaded as set of 128 bit vectors and repeated into a 256 bit vector
+                        __m256i lhs_mat_0123_0 = _mm256_loadu_si256((const __m256i *)((a_ptrs[rp][b].qs)));
+                        __m256i lhs_mat_01_0 = _mm256_permute2f128_si256(lhs_mat_0123_0, lhs_mat_0123_0, 0);
+                        __m256i lhs_mat_23_0 = _mm256_permute2f128_si256(lhs_mat_0123_0, lhs_mat_0123_0, 17);
+                        __m256i lhs_mat_0123_1 = _mm256_loadu_si256((const __m256i *)((a_ptrs[rp][b].qs + 32)));
+                        __m256i lhs_mat_01_1 = _mm256_permute2f128_si256(lhs_mat_0123_1, lhs_mat_0123_1, 0);
+                        __m256i lhs_mat_23_1 = _mm256_permute2f128_si256(lhs_mat_0123_1, lhs_mat_0123_1, 17);
+                        __m256i lhs_mat_0123_2 = _mm256_loadu_si256((const __m256i *)((a_ptrs[rp][b].qs + 64)));
+                        __m256i lhs_mat_01_2 = _mm256_permute2f128_si256(lhs_mat_0123_2, lhs_mat_0123_2, 0);
+                        __m256i lhs_mat_23_2 = _mm256_permute2f128_si256(lhs_mat_0123_2, lhs_mat_0123_2, 17);
+                        __m256i lhs_mat_0123_3 = _mm256_loadu_si256((const __m256i *)((a_ptrs[rp][b].qs + 96)));
+                        __m256i lhs_mat_01_3 = _mm256_permute2f128_si256(lhs_mat_0123_3, lhs_mat_0123_3, 0);
+                        __m256i lhs_mat_23_3 = _mm256_permute2f128_si256(lhs_mat_0123_3, lhs_mat_0123_3, 17);
+
+                        // Shuffle pattern one - left side input
+                        const __m256i lhs_mat_01_0_sp1 = _mm256_shuffle_epi32(lhs_mat_01_0, 160);  //A0(0-3) A0(0-3) A1(0-3) A1(0-3) A0(0-3) A0(0-3) A1(0-3) A1(0-3)
+                        const __m256i lhs_mat_23_0_sp1 = _mm256_shuffle_epi32(lhs_mat_23_0, 160);  //A2(0-3) A2(0-3) A3(0-3) A3(0-3) A2(0-3) A2(0-3) A3(0-3) A3(0-3)
+
+                        const __m256i lhs_mat_01_1_sp1 = _mm256_shuffle_epi32(lhs_mat_01_1, 160);  //A0(8-11) A0(8-11) A1(8-11) A1(8-11) A0(8-11) A0(8-11) A1(8-11) A1(8-11)
+                        const __m256i lhs_mat_23_1_sp1 = _mm256_shuffle_epi32(lhs_mat_23_1, 160);  //A2(8-11) A2(8-11) A3(8-11) A3(8-11) A2(8-11) A2(8-11) A3(8-11) A3(8-11)
+
+                        const __m256i lhs_mat_01_2_sp1 = _mm256_shuffle_epi32(lhs_mat_01_2, 160);  //A0(16-19) A0(16-19) A1(16-19) A1(16-19) A0(16-19) A0(16-19) A1(16-19) A1(16-19)
+                        const __m256i lhs_mat_23_2_sp1 = _mm256_shuffle_epi32(lhs_mat_23_2, 160);  //A2(16-19) A2(16-19) A3(16-19) A3(16-19) A2(16-19) A2(16-19) A3(16-19) A3(16-19)
+
+                        const __m256i lhs_mat_01_3_sp1 = _mm256_shuffle_epi32(lhs_mat_01_3, 160);  //A0(24-27) A0(24-27) A1(24-27) A1(24-27) A0(24-27) A0(24-27) A1(24-27) A1(24-27)
+                        const __m256i lhs_mat_23_3_sp1 = _mm256_shuffle_epi32(lhs_mat_23_3, 160);  //A2(24-27) A2(24-27) A3(24-27) A3(24-27) A2(24-27) A2(24-27) A3(24-27) A3(24-27)
+
+                        // Shuffle pattern two - left side input
+                        const __m256i lhs_mat_01_0_sp2 = _mm256_shuffle_epi32(lhs_mat_01_0, 245);  //A0(4-7) A0(4-7) A1(4-7) A1(4-7) A0(4-7) A0(4-7) A1(4-7) A1(4-7)
+                        const __m256i lhs_mat_23_0_sp2 = _mm256_shuffle_epi32(lhs_mat_23_0, 245);  //A2(4-7) A2(4-7) A3(4-7) A3(4-7) A2(4-7) A2(4-7) A3(4-7) A3(4-7)
+
+                        const __m256i lhs_mat_01_1_sp2 = _mm256_shuffle_epi32(lhs_mat_01_1, 245);  //A0(12-15) A0(12-15) A1(12-15) A1(12-15) A0(12-15) A0(12-15) A1(12-15) A1(12-15)
+                        const __m256i lhs_mat_23_1_sp2 = _mm256_shuffle_epi32(lhs_mat_23_1, 245);  //A2(12-15) A2(12-15) A3(12-15) A3(12-15) A2(12-15) A2(12-15) A3(12-15) A3(12-15)
+
+                        const __m256i lhs_mat_01_2_sp2 = _mm256_shuffle_epi32(lhs_mat_01_2, 245);  //A0(20-23) A0(20-23) A1(20-23) A1(20-23) A0(20-23) A0(20-23) A1(20-23) A1(20-23)
+                        const __m256i lhs_mat_23_2_sp2 = _mm256_shuffle_epi32(lhs_mat_23_2, 245);  //A2(20-23) A2(20-23) A3(20-23) A3(20-23) A2(20-23) A2(20-23) A3(20-23) A3(20-23)
+
+                        const __m256i lhs_mat_01_3_sp2 = _mm256_shuffle_epi32(lhs_mat_01_3, 245);  //A0(28-31) A0(28-31) A1(28-31) A1(28-31) A0(28-31) A0(28-31) A1(28-31) A1(28-31)
+                        const __m256i lhs_mat_23_3_sp2 = _mm256_shuffle_epi32(lhs_mat_23_3, 245);  //A2(28-31) A2(28-31) A3(28-31) A3(28-31) A2(28-31) A2(28-31) A3(28-31) A3(28-31)
+
+                        // The values arranged in shuffle patterns are operated with dot product operation within 32 bit lane i.e corresponding bytes and multiplied and added into 32 bit integers within 32 bit lane
+                        // Resembles MMLAs into 2x2 matrices in ARM Version
+                        __m256i iacc_mat_00_sp1 =
+                            _mm256_add_epi32(_mm256_add_epi32(_mm256_add_epi32(mul_sum_i8_pairs_int32x8(lhs_mat_01_3_sp1, rhs_mat_0145_3_sp1), mul_sum_i8_pairs_int32x8(lhs_mat_01_2_sp1, rhs_mat_0145_2_sp1)), mul_sum_i8_pairs_int32x8(lhs_mat_01_1_sp1, rhs_mat_0145_1_sp1)), mul_sum_i8_pairs_int32x8(lhs_mat_01_0_sp1, rhs_mat_0145_0_sp1));
+                        __m256i iacc_mat_01_sp1 =
+                            _mm256_add_epi32(_mm256_add_epi32(_mm256_add_epi32(mul_sum_i8_pairs_int32x8(lhs_mat_01_3_sp1, rhs_mat_2367_3_sp1), mul_sum_i8_pairs_int32x8(lhs_mat_01_2_sp1, rhs_mat_2367_2_sp1)), mul_sum_i8_pairs_int32x8(lhs_mat_01_1_sp1, rhs_mat_2367_1_sp1)), mul_sum_i8_pairs_int32x8(lhs_mat_01_0_sp1, rhs_mat_2367_0_sp1));
+                        __m256i iacc_mat_10_sp1 =
+                            _mm256_add_epi32(_mm256_add_epi32(_mm256_add_epi32(mul_sum_i8_pairs_int32x8(lhs_mat_23_3_sp1, rhs_mat_0145_3_sp1), mul_sum_i8_pairs_int32x8(lhs_mat_23_2_sp1, rhs_mat_0145_2_sp1)), mul_sum_i8_pairs_int32x8(lhs_mat_23_1_sp1, rhs_mat_0145_1_sp1)), mul_sum_i8_pairs_int32x8(lhs_mat_23_0_sp1, rhs_mat_0145_0_sp1));
+                        __m256i iacc_mat_11_sp1 =
+                            _mm256_add_epi32(_mm256_add_epi32(_mm256_add_epi32(mul_sum_i8_pairs_int32x8(lhs_mat_23_3_sp1, rhs_mat_2367_3_sp1), mul_sum_i8_pairs_int32x8(lhs_mat_23_2_sp1, rhs_mat_2367_2_sp1)), mul_sum_i8_pairs_int32x8(lhs_mat_23_1_sp1, rhs_mat_2367_1_sp1)), mul_sum_i8_pairs_int32x8(lhs_mat_23_0_sp1, rhs_mat_2367_0_sp1));
+                        __m256i iacc_mat_00_sp2 =
+                            _mm256_add_epi32(_mm256_add_epi32(_mm256_add_epi32(mul_sum_i8_pairs_int32x8(lhs_mat_01_3_sp2, rhs_mat_0145_3_sp2), mul_sum_i8_pairs_int32x8(lhs_mat_01_2_sp2, rhs_mat_0145_2_sp2)), mul_sum_i8_pairs_int32x8(lhs_mat_01_1_sp2, rhs_mat_0145_1_sp2)), mul_sum_i8_pairs_int32x8(lhs_mat_01_0_sp2, rhs_mat_0145_0_sp2));
+                        __m256i iacc_mat_01_sp2 =
+                            _mm256_add_epi32(_mm256_add_epi32(_mm256_add_epi32(mul_sum_i8_pairs_int32x8(lhs_mat_01_3_sp2, rhs_mat_2367_3_sp2), mul_sum_i8_pairs_int32x8(lhs_mat_01_2_sp2, rhs_mat_2367_2_sp2)), mul_sum_i8_pairs_int32x8(lhs_mat_01_1_sp2, rhs_mat_2367_1_sp2)), mul_sum_i8_pairs_int32x8(lhs_mat_01_0_sp2, rhs_mat_2367_0_sp2));
+                        __m256i iacc_mat_10_sp2 =
+                            _mm256_add_epi32(_mm256_add_epi32(_mm256_add_epi32(mul_sum_i8_pairs_int32x8(lhs_mat_23_3_sp2, rhs_mat_0145_3_sp2), mul_sum_i8_pairs_int32x8(lhs_mat_23_2_sp2, rhs_mat_0145_2_sp2)), mul_sum_i8_pairs_int32x8(lhs_mat_23_1_sp2, rhs_mat_0145_1_sp2)), mul_sum_i8_pairs_int32x8(lhs_mat_23_0_sp2, rhs_mat_0145_0_sp2));
+                        __m256i iacc_mat_11_sp2 =
+                            _mm256_add_epi32(_mm256_add_epi32(_mm256_add_epi32(mul_sum_i8_pairs_int32x8(lhs_mat_23_3_sp2, rhs_mat_2367_3_sp2), mul_sum_i8_pairs_int32x8(lhs_mat_23_2_sp2, rhs_mat_2367_2_sp2)), mul_sum_i8_pairs_int32x8(lhs_mat_23_1_sp2, rhs_mat_2367_1_sp2)), mul_sum_i8_pairs_int32x8(lhs_mat_23_0_sp2, rhs_mat_2367_0_sp2));
+
+                        // Output of both shuffle patterns are added in order to sum dot product outputs of all 32 values in block
+                        __m256i iacc_mat_00 = _mm256_add_epi32(iacc_mat_00_sp1, iacc_mat_00_sp2);
+                        __m256i iacc_mat_01 = _mm256_add_epi32(iacc_mat_01_sp1, iacc_mat_01_sp2);
+                        __m256i iacc_mat_10 = _mm256_add_epi32(iacc_mat_10_sp1, iacc_mat_10_sp2);
+                        __m256i iacc_mat_11 = _mm256_add_epi32(iacc_mat_11_sp1, iacc_mat_11_sp2);
+
+                        // Straighten out to make 4 row vectors
+                        __m256i iacc_row_0 = _mm256_blend_epi32(iacc_mat_00, _mm256_shuffle_epi32(iacc_mat_01, 78), 204);
+                        __m256i iacc_row_1 = _mm256_blend_epi32(_mm256_shuffle_epi32(iacc_mat_00, 78), iacc_mat_01, 204);
+                        __m256i iacc_row_2 = _mm256_blend_epi32(iacc_mat_10, _mm256_shuffle_epi32(iacc_mat_11, 78), 204);
+                        __m256i iacc_row_3 = _mm256_blend_epi32(_mm256_shuffle_epi32(iacc_mat_10, 78), iacc_mat_11, 204);
+
+                        // Load the scale(d) values for all the 4 Q8_0 blocks and repeat it across lanes
+                        const __m256 row_scale_f32 = GGML_F32Cx8_REPEAT_LOAD(a_ptrs[rp][b].d, loadMask);
+
+                        // Multiply with appropiate scales and accumulate
+                        acc_rows[rp * 4] = _mm256_fmadd_ps(_mm256_cvtepi32_ps(iacc_row_0), _mm256_mul_ps(col_scale_f32, _mm256_shuffle_ps(row_scale_f32, row_scale_f32, 0)), acc_rows[rp * 4]);
+                        acc_rows[rp * 4 + 1] = _mm256_fmadd_ps(_mm256_cvtepi32_ps(iacc_row_1), _mm256_mul_ps(col_scale_f32, _mm256_shuffle_ps(row_scale_f32, row_scale_f32, 85)), acc_rows[rp * 4 + 1]);
+                        acc_rows[rp * 4 + 2] = _mm256_fmadd_ps(_mm256_cvtepi32_ps(iacc_row_2), _mm256_mul_ps(col_scale_f32, _mm256_shuffle_ps(row_scale_f32, row_scale_f32, 170)), acc_rows[rp * 4 + 2]);
+                        acc_rows[rp * 4 + 3] = _mm256_fmadd_ps(_mm256_cvtepi32_ps(iacc_row_3), _mm256_mul_ps(col_scale_f32, _mm256_shuffle_ps(row_scale_f32, row_scale_f32,  255)), acc_rows[rp * 4 + 3]);
+                    }
+                }
+
+                // Store the accumulated values
+                for (int i = 0; i < 16; i++) {
+                    _mm256_storeu_ps((float *)(s + ((y * 4 + i) * bs + x * 8)), acc_rows[i]);
+                }
+            }
+        }
+
+        // Take a block_q8_0x4 structures at each pass of the loop and perform dot product operation
+        for (; y < nr / 4; y ++) {
+
+            const block_q8_0x4 * a_ptr = a_ptr_start + (y * nb);
+
+            // Load the eight block_q4_0 quantized values interleaved with each other in chunks of eight - B0,B1 ....B6,B7
+            for (int64_t x = xstart; x < nc / 8; x++) {
+
+                const block_q4_0x8 * b_ptr = b_ptr_start + (x * b_nb);
+
+                // Master FP accumulators
+                __m256 acc_rows[4];
+                for (int i = 0; i < 4; i++) {
+                    acc_rows[i] = _mm256_setzero_ps();
+                }
+
+                for (int64_t b = 0; b < nb; b++) {
+                    // Load the eight block_q8_0 quantized values interleaved with each other in chunks of eight - B0,B1 ....B6,B7
+                    const __m256i rhs_raw_mat_0123_0 = _mm256_loadu_si256((const __m256i *)(b_ptr[b].qs));
+                    const __m256i rhs_raw_mat_4567_0 = _mm256_loadu_si256((const __m256i *)(b_ptr[b].qs + 32));
+                    const __m256i rhs_raw_mat_0123_1 = _mm256_loadu_si256((const __m256i *)(b_ptr[b].qs + 64));
+                    const __m256i rhs_raw_mat_4567_1 = _mm256_loadu_si256((const __m256i *)(b_ptr[b].qs + 96));
+
+                    // Save the values in the following vectors in the formats B0B1B4B5, B2B3B6B7 for further processing and storing of valuess
+                    const __m256i rhs_raw_mat_0145_0 = _mm256_blend_epi32(rhs_raw_mat_0123_0, _mm256_permutevar8x32_epi32(rhs_raw_mat_4567_0, requiredOrder), 240);
+                    const __m256i rhs_raw_mat_2367_0 = _mm256_blend_epi32(_mm256_permutevar8x32_epi32(rhs_raw_mat_0123_0, requiredOrder), rhs_raw_mat_4567_0, 240);
+                    const __m256i rhs_raw_mat_0145_1 = _mm256_blend_epi32(rhs_raw_mat_0123_1, _mm256_permutevar8x32_epi32(rhs_raw_mat_4567_1, requiredOrder), 240);
+                    const __m256i rhs_raw_mat_2367_1 = _mm256_blend_epi32(_mm256_permutevar8x32_epi32(rhs_raw_mat_0123_1, requiredOrder), rhs_raw_mat_4567_1, 240);
+
+                    // 4-bit -> 8-bit - Sign is maintained
+                    const __m256i rhs_mat_0145_0 = _mm256_shuffle_epi8(signextendlut, _mm256_and_si256(rhs_raw_mat_0145_0, m4b));  //B0(0-7) B1(0-7) B4(0-7) B5(0-7)
+                    const __m256i rhs_mat_2367_0 = _mm256_shuffle_epi8(signextendlut, _mm256_and_si256(rhs_raw_mat_2367_0, m4b));  //B2(0-7) B3(0-7) B6(0-7) B7(0-7)
+
+                    const __m256i rhs_mat_0145_1 = _mm256_shuffle_epi8(signextendlut, _mm256_and_si256(rhs_raw_mat_0145_1, m4b));  //B0(8-15) B1(8-15) B4(8-15) B5(8-15)
+                    const __m256i rhs_mat_2367_1 = _mm256_shuffle_epi8(signextendlut, _mm256_and_si256(rhs_raw_mat_2367_1, m4b));  //B2(8-15) B3(8-15) B6(8-15) B7(8-15)
+
+                    const __m256i rhs_mat_0145_2 = _mm256_shuffle_epi8(signextendlut, _mm256_and_si256(_mm256_srli_epi16(rhs_raw_mat_0145_0, 4), m4b));  //B0(16-23) B1(16-23) B4(16-23) B5(16-23)
+                    const __m256i rhs_mat_2367_2 = _mm256_shuffle_epi8(signextendlut, _mm256_and_si256(_mm256_srli_epi16(rhs_raw_mat_2367_0, 4), m4b));  //B2(16-23) B3(16-23) B6(16-23) B7(16-23)
+
+                    const __m256i rhs_mat_0145_3 = _mm256_shuffle_epi8(signextendlut, _mm256_and_si256(_mm256_srli_epi16(rhs_raw_mat_0145_1, 4), m4b));  //B0(24-31) B1(24-31) B4(24-31) B5(24-31)
+                    const __m256i rhs_mat_2367_3 = _mm256_shuffle_epi8(signextendlut, _mm256_and_si256(_mm256_srli_epi16(rhs_raw_mat_2367_1, 4), m4b));  //B2(24-31) B3(24-31) B6(24-31) B7(24-31)
+
+                    // Shuffle pattern one - right side input
+                    const __m256i rhs_mat_0145_0_sp1 = _mm256_shuffle_epi32(rhs_mat_0145_0, 136);  //B0(0-3) B1(0-3) B0(0-3) B1(0-3) B4(0-3) B5(0-3) B4(0-3) B5(0-3)
+                    const __m256i rhs_mat_2367_0_sp1 = _mm256_shuffle_epi32(rhs_mat_2367_0, 136);  //B2(0-3) B3(0-3) B2(0-3) B3(0-3) B6(0-3) B7(0-3) B6(0-3) B7(0-3)
+
+                    const __m256i rhs_mat_0145_1_sp1 = _mm256_shuffle_epi32(rhs_mat_0145_1, 136);  //B0(8-11) B1(8-11) B0(8-11) B1(8-11) B4(8-11) B5(8-11) B4(8-11) B5(8-11)
+                    const __m256i rhs_mat_2367_1_sp1 = _mm256_shuffle_epi32(rhs_mat_2367_1, 136);  //B2(8-11) B3(8-11) B2(8-11) B3(8-11) B6(8-11) B7(8-11) B6(8-11) B7(8-11)
+
+                    const __m256i rhs_mat_0145_2_sp1 = _mm256_shuffle_epi32(rhs_mat_0145_2, 136);  //B0(16-19) B1(16-19) B0(16-19) B1(16-19) B4(16-19) B5(16-19) B4(16-19) B5(16-19)
+                    const __m256i rhs_mat_2367_2_sp1 = _mm256_shuffle_epi32(rhs_mat_2367_2, 136);  //B2(16-19) B3(16-19) B2(16-19) B3(16-19) B6(16-19) B7(16-19) B6(16-19) B7(16-19)
+
+                    const __m256i rhs_mat_0145_3_sp1 = _mm256_shuffle_epi32(rhs_mat_0145_3, 136);  //B0(24-27) B1(24-27) B0(24-27) B1(24-27) B4(24-27) B5(24-27) B4(24-27) B5(24-27)
+                    const __m256i rhs_mat_2367_3_sp1 = _mm256_shuffle_epi32(rhs_mat_2367_3, 136);  //B2(24-27) B3(24-27) B2(24-27) B3(24-27) B6(24-27) B7(24-27) B6(24-27) B7(24-27)
+
+                    // Shuffle pattern two - right side input
+
+                    const __m256i rhs_mat_0145_0_sp2 = _mm256_shuffle_epi32(rhs_mat_0145_0, 221);  //B0(4-7) B1(4-7) B0(4-7) B1(4-7) B4(4-7) B5(4-7) B4(4-7) B5(4-7)
+                    const __m256i rhs_mat_2367_0_sp2 = _mm256_shuffle_epi32(rhs_mat_2367_0, 221);  //B2(4-7) B3(4-7) B2(4-7) B3(4-7) B6(4-7) B7(4-7) B6(4-7) B7(4-7)
+
+                    const __m256i rhs_mat_0145_1_sp2 = _mm256_shuffle_epi32(rhs_mat_0145_1, 221);  //B0(12-15) B1(12-15) B0(12-15) B1(12-15) B4(12-15) B5(12-15) B4(12-15) B5(12-15)
+                    const __m256i rhs_mat_2367_1_sp2 = _mm256_shuffle_epi32(rhs_mat_2367_1, 221);  //B2(12-15) B3(12-15) B2(12-15) B3(12-15) B6(12-15) B7(12-15) B6(12-15) B7(12-15)
+
+                    const __m256i rhs_mat_0145_2_sp2 = _mm256_shuffle_epi32(rhs_mat_0145_2, 221);  //B0(20-23) B1(20-23) B0(20-23) B1(20-23) B4(20-23) B5(20-23) B4(20-23) B5(20-23)
+                    const __m256i rhs_mat_2367_2_sp2 = _mm256_shuffle_epi32(rhs_mat_2367_2, 221);  //B2(20-23) B3(20-23) B2(20-23) B3(20-23) B6(20-23) B7(20-23) B6(20-23) B7(20-23)
+
+                    const __m256i rhs_mat_0145_3_sp2 = _mm256_shuffle_epi32(rhs_mat_0145_3, 221);  //B0(28-31) B1(28-31) B0(28-31) B1(28-31) B4(28-31) B5(28-31) B4(28-31) B5(28-31)
+                    const __m256i rhs_mat_2367_3_sp2 = _mm256_shuffle_epi32(rhs_mat_2367_3, 221);  //B2(28-31) B3(28-31) B2(28-31) B3(28-31) B6(28-31) B7(28-31) B6(28-31) B7(28-31)
+
+                    // Scale values - Load the wight scale values of block_q4_0x8
+                    const __m256 col_scale_f32 = GGML_F32Cx8_LOAD(b_ptr[b].d);
+
+                    // Load the four block_q4_0 quantized values interleaved with each other in chunks of eight - A0,A1,A2,A3
+                    // Loaded as set of 128 bit vectors and repeated into a 256 bit vector
+                    __m256i lhs_mat_0123_0 = _mm256_loadu_si256((const __m256i *)((a_ptr[b].qs)));
+                    __m256i lhs_mat_01_0 = _mm256_permute2f128_si256(lhs_mat_0123_0, lhs_mat_0123_0, 0);
+                    __m256i lhs_mat_23_0 = _mm256_permute2f128_si256(lhs_mat_0123_0, lhs_mat_0123_0, 17);
+                    __m256i lhs_mat_0123_1 = _mm256_loadu_si256((const __m256i *)((a_ptr[b].qs + 32)));
+                    __m256i lhs_mat_01_1 = _mm256_permute2f128_si256(lhs_mat_0123_1, lhs_mat_0123_1, 0);
+                    __m256i lhs_mat_23_1 = _mm256_permute2f128_si256(lhs_mat_0123_1, lhs_mat_0123_1, 17);
+                    __m256i lhs_mat_0123_2 = _mm256_loadu_si256((const __m256i *)((a_ptr[b].qs + 64)));
+                    __m256i lhs_mat_01_2 = _mm256_permute2f128_si256(lhs_mat_0123_2, lhs_mat_0123_2, 0);
+                    __m256i lhs_mat_23_2 = _mm256_permute2f128_si256(lhs_mat_0123_2, lhs_mat_0123_2, 17);
+                    __m256i lhs_mat_0123_3 = _mm256_loadu_si256((const __m256i *)((a_ptr[b].qs + 96)));
+                    __m256i lhs_mat_01_3 = _mm256_permute2f128_si256(lhs_mat_0123_3, lhs_mat_0123_3, 0);
+                    __m256i lhs_mat_23_3 = _mm256_permute2f128_si256(lhs_mat_0123_3, lhs_mat_0123_3, 17);
+
+                    // Shuffle pattern one - left side input
+
+                    const __m256i lhs_mat_01_0_sp1 = _mm256_shuffle_epi32(lhs_mat_01_0, 160);  //A0(0-3) A0(0-3) A1(0-3) A1(0-3) A0(0-3) A0(0-3) A1(0-3) A1(0-3)
+                    const __m256i lhs_mat_23_0_sp1 = _mm256_shuffle_epi32(lhs_mat_23_0, 160);  //A2(0-3) A2(0-3) A3(0-3) A3(0-3) A2(0-3) A2(0-3) A3(0-3) A3(0-3)
+
+                    const __m256i lhs_mat_01_1_sp1 = _mm256_shuffle_epi32(lhs_mat_01_1, 160);  //A0(8-11) A0(8-11) A1(8-11) A1(8-11) A0(8-11) A0(8-11) A1(8-11) A1(8-11)
+                    const __m256i lhs_mat_23_1_sp1 = _mm256_shuffle_epi32(lhs_mat_23_1, 160);  //A2(8-11) A2(8-11) A3(8-11) A3(8-11) A2(8-11) A2(8-11) A3(8-11) A3(8-11)
+
+                    const __m256i lhs_mat_01_2_sp1 = _mm256_shuffle_epi32(lhs_mat_01_2, 160);  //A0(16-19) A0(16-19) A1(16-19) A1(16-19) A0(16-19) A0(16-19) A1(16-19) A1(16-19)
+                    const __m256i lhs_mat_23_2_sp1 = _mm256_shuffle_epi32(lhs_mat_23_2, 160);  //A2(16-19) A2(16-19) A3(16-19) A3(16-19) A2(16-19) A2(16-19) A3(16-19) A3(16-19)
+
+                    const __m256i lhs_mat_01_3_sp1 = _mm256_shuffle_epi32(lhs_mat_01_3, 160);  //A0(24-27) A0(24-27) A1(24-27) A1(24-27) A0(24-27) A0(24-27) A1(24-27) A1(24-27)
+                    const __m256i lhs_mat_23_3_sp1 = _mm256_shuffle_epi32(lhs_mat_23_3, 160);  //A2(24-27) A2(24-27) A3(24-27) A3(24-27) A2(24-27) A2(24-27) A3(24-27) A3(24-27)
+
+                    // Shuffle pattern two - left side input
+
+                    const __m256i lhs_mat_01_0_sp2 = _mm256_shuffle_epi32(lhs_mat_01_0, 245);  //A0(4-7) A0(4-7) A1(4-7) A1(4-7) A0(4-7) A0(4-7) A1(4-7) A1(4-7)
+                    const __m256i lhs_mat_23_0_sp2 = _mm256_shuffle_epi32(lhs_mat_23_0, 245);  //A2(4-7) A2(4-7) A3(4-7) A3(4-7) A2(4-7) A2(4-7) A3(4-7) A3(4-7)
+
+                    const __m256i lhs_mat_01_1_sp2 = _mm256_shuffle_epi32(lhs_mat_01_1, 245);  //A0(12-15) A0(12-15) A1(12-15) A1(12-15) A0(12-15) A0(12-15) A1(12-15) A1(12-15)
+                    const __m256i lhs_mat_23_1_sp2 = _mm256_shuffle_epi32(lhs_mat_23_1, 245);  //A2(12-15) A2(12-15) A3(12-15) A3(12-15) A2(12-15) A2(12-15) A3(12-15) A3(12-15)
+
+                    const __m256i lhs_mat_01_2_sp2 = _mm256_shuffle_epi32(lhs_mat_01_2, 245);  //A0(20-23) A0(20-23) A1(20-23) A1(20-23) A0(20-23) A0(20-23) A1(20-23) A1(20-23)
+                    const __m256i lhs_mat_23_2_sp2 = _mm256_shuffle_epi32(lhs_mat_23_2, 245);  //A2(20-23) A2(20-23) A3(20-23) A3(20-23) A2(20-23) A2(20-23) A3(20-23) A3(20-23)
+
+                    const __m256i lhs_mat_01_3_sp2 = _mm256_shuffle_epi32(lhs_mat_01_3, 245);  //A0(28-31) A0(28-31) A1(28-31) A1(28-31) A0(28-31) A0(28-31) A1(28-31) A1(28-31)
+                    const __m256i lhs_mat_23_3_sp2 = _mm256_shuffle_epi32(lhs_mat_23_3, 245);  //A2(28-31) A2(28-31) A3(28-31) A3(28-31) A2(28-31) A2(28-31) A3(28-31) A3(28-31)
+
+                    // The values arranged in shuffle patterns are operated with dot product operation within 32 bit lane i.e corresponding bytes and multiplied and added into 32 bit integers within 32 bit lane
+                    // Resembles MMLAs into 2x2 matrices in ARM Version
+                    __m256i iacc_mat_00_sp1 =
+                        _mm256_add_epi32(_mm256_add_epi32(_mm256_add_epi32(mul_sum_i8_pairs_int32x8(lhs_mat_01_3_sp1, rhs_mat_0145_3_sp1), mul_sum_i8_pairs_int32x8(lhs_mat_01_2_sp1, rhs_mat_0145_2_sp1)), mul_sum_i8_pairs_int32x8(lhs_mat_01_1_sp1, rhs_mat_0145_1_sp1)), mul_sum_i8_pairs_int32x8(lhs_mat_01_0_sp1, rhs_mat_0145_0_sp1));
+                    __m256i iacc_mat_01_sp1 =
+                        _mm256_add_epi32(_mm256_add_epi32(_mm256_add_epi32(mul_sum_i8_pairs_int32x8(lhs_mat_01_3_sp1, rhs_mat_2367_3_sp1), mul_sum_i8_pairs_int32x8(lhs_mat_01_2_sp1, rhs_mat_2367_2_sp1)), mul_sum_i8_pairs_int32x8(lhs_mat_01_1_sp1, rhs_mat_2367_1_sp1)), mul_sum_i8_pairs_int32x8(lhs_mat_01_0_sp1, rhs_mat_2367_0_sp1));
+                    __m256i iacc_mat_10_sp1 =
+                        _mm256_add_epi32(_mm256_add_epi32(_mm256_add_epi32(mul_sum_i8_pairs_int32x8(lhs_mat_23_3_sp1, rhs_mat_0145_3_sp1), mul_sum_i8_pairs_int32x8(lhs_mat_23_2_sp1, rhs_mat_0145_2_sp1)), mul_sum_i8_pairs_int32x8(lhs_mat_23_1_sp1, rhs_mat_0145_1_sp1)), mul_sum_i8_pairs_int32x8(lhs_mat_23_0_sp1, rhs_mat_0145_0_sp1));
+                    __m256i iacc_mat_11_sp1 =
+                        _mm256_add_epi32(_mm256_add_epi32(_mm256_add_epi32(mul_sum_i8_pairs_int32x8(lhs_mat_23_3_sp1, rhs_mat_2367_3_sp1), mul_sum_i8_pairs_int32x8(lhs_mat_23_2_sp1, rhs_mat_2367_2_sp1)), mul_sum_i8_pairs_int32x8(lhs_mat_23_1_sp1, rhs_mat_2367_1_sp1)), mul_sum_i8_pairs_int32x8(lhs_mat_23_0_sp1, rhs_mat_2367_0_sp1));
+                    __m256i iacc_mat_00_sp2 =
+                        _mm256_add_epi32(_mm256_add_epi32(_mm256_add_epi32(mul_sum_i8_pairs_int32x8(lhs_mat_01_3_sp2, rhs_mat_0145_3_sp2), mul_sum_i8_pairs_int32x8(lhs_mat_01_2_sp2, rhs_mat_0145_2_sp2)), mul_sum_i8_pairs_int32x8(lhs_mat_01_1_sp2, rhs_mat_0145_1_sp2)), mul_sum_i8_pairs_int32x8(lhs_mat_01_0_sp2, rhs_mat_0145_0_sp2));
+                    __m256i iacc_mat_01_sp2 =
+                        _mm256_add_epi32(_mm256_add_epi32(_mm256_add_epi32(mul_sum_i8_pairs_int32x8(lhs_mat_01_3_sp2, rhs_mat_2367_3_sp2), mul_sum_i8_pairs_int32x8(lhs_mat_01_2_sp2, rhs_mat_2367_2_sp2)), mul_sum_i8_pairs_int32x8(lhs_mat_01_1_sp2, rhs_mat_2367_1_sp2)), mul_sum_i8_pairs_int32x8(lhs_mat_01_0_sp2, rhs_mat_2367_0_sp2));
+                    __m256i iacc_mat_10_sp2 =
+                        _mm256_add_epi32(_mm256_add_epi32(_mm256_add_epi32(mul_sum_i8_pairs_int32x8(lhs_mat_23_3_sp2, rhs_mat_0145_3_sp2), mul_sum_i8_pairs_int32x8(lhs_mat_23_2_sp2, rhs_mat_0145_2_sp2)), mul_sum_i8_pairs_int32x8(lhs_mat_23_1_sp2, rhs_mat_0145_1_sp2)), mul_sum_i8_pairs_int32x8(lhs_mat_23_0_sp2, rhs_mat_0145_0_sp2));
+                    __m256i iacc_mat_11_sp2 =
+                        _mm256_add_epi32(_mm256_add_epi32(_mm256_add_epi32(mul_sum_i8_pairs_int32x8(lhs_mat_23_3_sp2, rhs_mat_2367_3_sp2), mul_sum_i8_pairs_int32x8(lhs_mat_23_2_sp2, rhs_mat_2367_2_sp2)), mul_sum_i8_pairs_int32x8(lhs_mat_23_1_sp2, rhs_mat_2367_1_sp2)), mul_sum_i8_pairs_int32x8(lhs_mat_23_0_sp2, rhs_mat_2367_0_sp2));
+
+                    // Output of both shuffle patterns are added in order to sum dot product outputs of all 32 values in block
+                    __m256i iacc_mat_00 = _mm256_add_epi32(iacc_mat_00_sp1, iacc_mat_00_sp2);
+                    __m256i iacc_mat_01 = _mm256_add_epi32(iacc_mat_01_sp1, iacc_mat_01_sp2);
+                    __m256i iacc_mat_10 = _mm256_add_epi32(iacc_mat_10_sp1, iacc_mat_10_sp2);
+                    __m256i iacc_mat_11 = _mm256_add_epi32(iacc_mat_11_sp1, iacc_mat_11_sp2);
+
+
+                    // Straighten out to make 4 row vectors
+                    __m256i iacc_row_0 = _mm256_blend_epi32(iacc_mat_00, _mm256_shuffle_epi32(iacc_mat_01, 78), 204);
+                    __m256i iacc_row_1 = _mm256_blend_epi32(_mm256_shuffle_epi32(iacc_mat_00, 78), iacc_mat_01, 204);
+                    __m256i iacc_row_2 = _mm256_blend_epi32(iacc_mat_10, _mm256_shuffle_epi32(iacc_mat_11, 78), 204);
+                    __m256i iacc_row_3 = _mm256_blend_epi32(_mm256_shuffle_epi32(iacc_mat_10, 78), iacc_mat_11, 204);
+
+                    // Load the scale(d) values for all the 4 Q8_0 blocks and repeat it across lanes
+                    const __m256 row_scale_f32 = GGML_F32Cx8_REPEAT_LOAD(a_ptr[b].d, loadMask);
+
+                    // Multiply with appropiate scales and accumulate
+                    acc_rows[0] = _mm256_fmadd_ps(_mm256_cvtepi32_ps(iacc_row_0), _mm256_mul_ps(col_scale_f32, _mm256_shuffle_ps(row_scale_f32, row_scale_f32, 0)), acc_rows[0]);
+                    acc_rows[1] = _mm256_fmadd_ps(_mm256_cvtepi32_ps(iacc_row_1), _mm256_mul_ps(col_scale_f32, _mm256_shuffle_ps(row_scale_f32, row_scale_f32, 85)), acc_rows[1]);
+                    acc_rows[2] = _mm256_fmadd_ps(_mm256_cvtepi32_ps(iacc_row_2), _mm256_mul_ps(col_scale_f32, _mm256_shuffle_ps(row_scale_f32, row_scale_f32, 170)), acc_rows[2]);
+                    acc_rows[3] = _mm256_fmadd_ps(_mm256_cvtepi32_ps(iacc_row_3), _mm256_mul_ps(col_scale_f32, _mm256_shuffle_ps(row_scale_f32, row_scale_f32, 255)), acc_rows[3]);
+                }
+
+                // Store the accumulated values
+                for (int i = 0; i < 4; i++) {
+                    _mm256_storeu_ps((float *)(s + ((y * 4 + i) * bs + x * 8)), acc_rows[i]);
+                }
+            }
+        }
+        return;
+    }
+#endif // #if ! ((defined(_MSC_VER)) && ! defined(__clang__)) && defined(__aarch64__)
+    float sumf[4][8];
+    int sumi;
+
+    for (int y = 0; y < nr / 4; y++) {
+        const block_q8_0x4 * a_ptr = (const block_q8_0x4 *) vy + (y * nb);
+        for (int x = 0; x < nc / ncols_interleaved; x++) {
+            const block_q4_0x8 * b_ptr = (const block_q4_0x8 *) vx + (x * nb);
+            for (int m = 0; m < 4; m++) {
+                for (int j = 0; j < ncols_interleaved; j++) sumf[m][j] = 0.0;
+            }
+            for (int l = 0; l < nb; l++) {
+                for (int k = 0; k < (qk / (2 * blocklen)); k++) {
+                    for (int m = 0; m < 4; m++) {
+                        for (int j = 0; j < ncols_interleaved; j++) {
+                            sumi = 0;
+                            for (int i = 0; i < blocklen; ++i) {
+                                const int v0 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] << 4);
+                                const int v1 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] & 0xF0);
+                                sumi += ((v0 * a_ptr[l].qs[k * 4 * blocklen + m * blocklen + i]) +
+                                         (v1 * a_ptr[l].qs[k * 4 * blocklen + m * blocklen + i + qk / 2 * 4])) >> 4;
+                            }
+                            sumf[m][j] += sumi * GGML_FP16_TO_FP32(b_ptr[l].d[j]) * GGML_FP16_TO_FP32(a_ptr[l].d[m]);
+                        }
+                    }
+                }
+            }
+            for (int m = 0; m < 4; m++) {
+                for (int j = 0; j < ncols_interleaved; j++)
+                    s[(y * 4 + m) * bs + x * ncols_interleaved + j] = sumf[m][j];
+            }
+        }
+    }
+}
diff --git a/src/whisper_cpp/ggml/src/ggml-aarch64.h b/src/whisper_cpp/ggml/src/ggml-aarch64.h
new file mode 100644
index 00000000..517babaf
--- /dev/null
+++ b/src/whisper_cpp/ggml/src/ggml-aarch64.h
@@ -0,0 +1,39 @@
+// SPDX-FileCopyrightText: Copyright 2024 Arm Ltd.
+#pragma once
+
+#define GGML_COMMON_DECL_C
+#include "ggml-common.h"
+
+#include "ggml.h"
+
+// GGML internal header
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+// Quantization
+void quantize_q8_0_4x4(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
+void quantize_q8_0_4x8(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
+
+void quantize_mat_q8_0(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t nrows, int64_t n_per_row, int64_t blck_size_interleave);
+
+// Quantization utilizing an importance matrix (a.k.a. "Activation aWare Quantization")
+size_t quantize_q4_0_4x4(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
+size_t quantize_q4_0_4x8(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
+size_t quantize_q4_0_8x8(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
+
+// GEMV
+void ggml_gemv_q4_0_4x4_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
+void ggml_gemv_q4_0_4x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
+void ggml_gemv_q4_0_8x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
+
+// GEMM
+void ggml_gemm_q4_0_4x4_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
+void ggml_gemm_q4_0_4x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
+void ggml_gemm_q4_0_8x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
+
+#ifdef __cplusplus
+}
+#endif
+
diff --git a/src/whisper_cpp/ggml/src/ggml-alloc.c b/src/whisper_cpp/ggml/src/ggml-alloc.c
new file mode 100644
index 00000000..70187b9b
--- /dev/null
+++ b/src/whisper_cpp/ggml/src/ggml-alloc.c
@@ -0,0 +1,1042 @@
+#include "ggml-alloc.h"
+#include "ggml-backend-impl.h"
+#include "ggml.h"
+#include "ggml-impl.h"
+#include <assert.h>
+#include <limits.h>
+#include <stdarg.h>
+#include <stdio.h>
+#include <stdlib.h>
+#include <string.h>
+
+#define MAX(a, b) ((a) > (b) ? (a) : (b))
+#define MAX_FREE_BLOCKS 256
+
+//#define GGML_ALLOCATOR_DEBUG
+
+//#define AT_PRINTF(...) fprintf(stderr, __VA_ARGS__)
+#define AT_PRINTF(...)
+
+
+static bool ggml_is_view(const struct ggml_tensor * t) {
+    return t->view_src != NULL;
+}
+
+static bool ggml_are_same_layout(const struct ggml_tensor * a, const struct ggml_tensor * b) {
+    if (a->type != b->type) {
+        return false;
+    }
+    for (int i = 0; i < GGML_MAX_DIMS; i++) {
+        if (a->ne[i] != b->ne[i]) {
+            return false;
+        }
+        if (a->nb[i] != b->nb[i]) {
+            return false;
+        }
+    }
+    return true;
+}
+
+static bool ggml_op_can_inplace(enum ggml_op op) {
+    switch (op) {
+        case GGML_OP_SCALE:
+        case GGML_OP_DIAG_MASK_ZERO:
+        case GGML_OP_DIAG_MASK_INF:
+        case GGML_OP_ADD:
+        case GGML_OP_ADD1:
+        case GGML_OP_SUB:
+        case GGML_OP_MUL:
+        case GGML_OP_DIV:
+        case GGML_OP_SQR:
+        case GGML_OP_SQRT:
+        case GGML_OP_LOG:
+        case GGML_OP_UNARY:
+        case GGML_OP_ROPE:
+        case GGML_OP_RMS_NORM:
+        case GGML_OP_SOFT_MAX:
+            return true;
+
+        default:
+            return false;
+    }
+}
+
+static size_t aligned_offset(const void * buffer, size_t offset, size_t alignment) {
+    assert(alignment && !(alignment & (alignment - 1))); // power of 2
+    size_t align = (alignment - (((uintptr_t)buffer + offset) % alignment)) % alignment;
+    return offset + align;
+}
+
+// tallocr
+
+struct ggml_tallocr ggml_tallocr_new(ggml_backend_buffer_t buffer) {
+    void * base = ggml_backend_buffer_get_base(buffer);
+    size_t align = ggml_backend_buffer_get_alignment(buffer);
+
+    assert(align && !(align & (align - 1))); // power of 2
+
+    struct ggml_tallocr talloc = (struct ggml_tallocr) {
+        /*.buffer    = */ buffer,
+        /*.base      = */ base,
+        /*.alignment = */ align,
+        /*.offset    = */ aligned_offset(base, 0, align),
+    };
+    return talloc;
+}
+
+void ggml_tallocr_alloc(struct ggml_tallocr * talloc, struct ggml_tensor * tensor) {
+    size_t size = ggml_backend_buffer_get_alloc_size(talloc->buffer, tensor);
+    size = GGML_PAD(size, talloc->alignment);
+
+    if (talloc->offset + size > ggml_backend_buffer_get_size(talloc->buffer)) {
+        fprintf(stderr, "%s: not enough space in the buffer to allocate %s (needed %zu, available %zu)\n",
+                __func__, tensor->name, size, ggml_backend_buffer_get_size(talloc->buffer) - talloc->offset);
+        GGML_ABORT("not enough space in the buffer");
+    }
+
+    void * addr = (char *)ggml_backend_buffer_get_base(talloc->buffer) + talloc->offset;
+    talloc->offset += size;
+
+    assert(((uintptr_t)addr % talloc->alignment) == 0);
+
+    ggml_backend_tensor_alloc(talloc->buffer, tensor, addr);
+}
+
+// dynamic tensor allocator
+
+struct free_block {
+    size_t offset;
+    size_t size;
+};
+
+struct ggml_dyn_tallocr {
+    size_t alignment;
+    int n_free_blocks;
+    struct free_block free_blocks[MAX_FREE_BLOCKS];
+    size_t max_size;
+
+#ifdef GGML_ALLOCATOR_DEBUG
+    struct {
+        const struct ggml_tensor * tensor;
+        size_t offset;
+    } allocated_tensors[1024];
+#endif
+};
+
+#ifdef GGML_ALLOCATOR_DEBUG
+static void add_allocated_tensor(struct ggml_dyn_tallocr * alloc, size_t offset, const struct ggml_tensor * tensor) {
+    for (int i = 0; i < 1024; i++) {
+        if (alloc->allocated_tensors[i].tensor == NULL) {
+            alloc->allocated_tensors[i].tensor = tensor;
+            alloc->allocated_tensors[i].offset = offset;
+            return;
+        }
+    }
+    GGML_ABORT("out of allocated_tensors");
+}
+static void remove_allocated_tensor(struct ggml_dyn_tallocr * alloc, size_t offset, const struct ggml_tensor * tensor) {
+    for (int i = 0; i < 1024; i++) {
+        if (alloc->allocated_tensors[i].offset == offset) {
+            alloc->allocated_tensors[i].tensor = NULL;
+            return;
+        }
+    }
+    GGML_ABORT("tried to free tensor %s not found\n", tensor->name);
+}
+#endif
+
+static size_t ggml_dyn_tallocr_alloc(struct ggml_dyn_tallocr * alloc, size_t size, const struct ggml_tensor * tensor) {
+    size = aligned_offset(NULL, size, alloc->alignment);
+
+    AT_PRINTF("%s: allocating %s (%zu bytes) - ", __func__, tensor->name, size);
+
+    size_t max_avail = 0;
+
+    // find the best fitting free block besides the last block
+    int best_fit_block = -1;
+    size_t best_fit_size = SIZE_MAX;
+    for (int i = 0; i < alloc->n_free_blocks - 1; i++) {
+        struct free_block * block = &alloc->free_blocks[i];
+        max_avail = MAX(max_avail, block->size);
+        if (block->size >= size && block->size <= best_fit_size) {
+            best_fit_block = i;
+            best_fit_size = block->size;
+        }
+    }
+
+    if (best_fit_block == -1) {
+        // the last block is our last resort
+        struct free_block * block = &alloc->free_blocks[alloc->n_free_blocks - 1];
+        max_avail = MAX(max_avail, block->size);
+        if (block->size >= size) {
+            best_fit_block = alloc->n_free_blocks - 1;
+        } else {
+            // this should never happen
+            fprintf(stderr, "%s: not enough space in the buffer to allocate %zu bytes, largest block available %zu bytes\n",
+                    __func__, size, max_avail);
+            GGML_ABORT("not enough space in the buffer");
+        }
+    }
+
+    struct free_block * block = &alloc->free_blocks[best_fit_block];
+    size_t offset = block->offset;
+    block->offset = offset + size;
+    block->size -= size;
+    if (block->size == 0) {
+        // remove block if empty
+        alloc->n_free_blocks--;
+        for (int j = best_fit_block; j < alloc->n_free_blocks; j++) {
+            alloc->free_blocks[j] = alloc->free_blocks[j+1];
+        }
+    }
+
+    AT_PRINTF("block %d, offset %zu\n", best_fit_block, offset);
+
+#ifdef GGML_ALLOCATOR_DEBUG
+    add_allocated_tensor(alloc, offset, tensor);
+    size_t cur_max = offset + size;
+    if (cur_max > alloc->max_size) {
+        // sort allocated_tensors by offset
+        for (int i = 0; i < 1024; i++) {
+            for (int j = i + 1; j < 1024; j++) {
+                if (alloc->allocated_tensors[i].offset > alloc->allocated_tensors[j].offset) {
+                    const struct ggml_tensor * tmp_tensor = alloc->allocated_tensors[i].tensor;
+                    size_t tmp_offset = alloc->allocated_tensors[i].offset;
+                    alloc->allocated_tensors[i].tensor = alloc->allocated_tensors[j].tensor;
+                    alloc->allocated_tensors[i].offset = alloc->allocated_tensors[j].offset;
+                    alloc->allocated_tensors[j].tensor = tmp_tensor;
+                    alloc->allocated_tensors[j].offset = tmp_offset;
+                }
+            }
+        }
+        fprintf(stderr, "max_size = %.2f MB: tensors: ", cur_max / 1024.0 / 1024.0);
+        for (int i = 0; i < 1024; i++) {
+            if (alloc->allocated_tensors[i].tensor) {
+                fprintf(stderr, "%s [%zx-%zx] (%.2f MB) ", alloc->allocated_tensors[i].tensor->name,
+                    alloc->allocated_tensors[i].offset,
+                    alloc->allocated_tensors[i].offset + ggml_nbytes(alloc->allocated_tensors[i].tensor),
+                    ggml_nbytes(alloc->allocated_tensors[i].tensor) / 1024.0 / 1024.0);
+            }
+        }
+        fprintf(stderr, "\n");
+    }
+#endif
+
+    alloc->max_size = MAX(alloc->max_size, offset + size);
+
+    return offset;
+
+    GGML_UNUSED(tensor);
+}
+
+// this is a very naive implementation, but for our case the number of free blocks should be very small
+static void ggml_dyn_tallocr_free_tensor(struct ggml_dyn_tallocr * alloc, size_t offset, size_t size, const struct ggml_tensor * tensor) {
+    size = aligned_offset(NULL, size, alloc->alignment);
+
+    AT_PRINTF("%s: freeing %s at %zu (%zu bytes) - n_free_blocks = %d\n", __func__, tensor->name, offset, size, alloc->n_free_blocks);
+
+#ifdef GGML_ALLOCATOR_DEBUG
+    remove_allocated_tensor(alloc, offset, tensor);
+#endif
+
+    // see if we can merge with an existing block
+    for (int i = 0; i < alloc->n_free_blocks; i++) {
+        struct free_block * block = &alloc->free_blocks[i];
+        // check if ptr is at the end of the block
+        if (block->offset + block->size == offset) {
+            block->size += size;
+            // check if we can merge with the next block
+            if (i < alloc->n_free_blocks - 1 && block->offset + block->size == alloc->free_blocks[i+1].offset) {
+                block->size += alloc->free_blocks[i+1].size;
+                alloc->n_free_blocks--;
+                for (int j = i+1; j < alloc->n_free_blocks; j++) {
+                    alloc->free_blocks[j] = alloc->free_blocks[j+1];
+                }
+            }
+            return;
+        }
+        // check if ptr is at the beginning of the block
+        if (offset + size == block->offset) {
+            block->offset = offset;
+            block->size += size;
+            // check if we can merge with the previous block
+            if (i > 0 && alloc->free_blocks[i-1].offset + alloc->free_blocks[i-1].size == block->offset) {
+                alloc->free_blocks[i-1].size += block->size;
+                alloc->n_free_blocks--;
+                for (int j = i; j < alloc->n_free_blocks; j++) {
+                    alloc->free_blocks[j] = alloc->free_blocks[j+1];
+                }
+            }
+            return;
+        }
+    }
+    // otherwise, add a new block
+    GGML_ASSERT(alloc->n_free_blocks < MAX_FREE_BLOCKS && "out of free blocks");
+    // insert the new block in the correct position to keep the array sorted by address (to make merging blocks faster)
+    int insert_pos = 0;
+    while (insert_pos < alloc->n_free_blocks && alloc->free_blocks[insert_pos].offset < offset) {
+        insert_pos++;
+    }
+    // shift all blocks from insert_pos onward to make room for the new block
+    for (int i = alloc->n_free_blocks; i > insert_pos; i--) {
+        alloc->free_blocks[i] = alloc->free_blocks[i-1];
+    }
+    // insert the new block
+    alloc->free_blocks[insert_pos].offset = offset;
+    alloc->free_blocks[insert_pos].size = size;
+    alloc->n_free_blocks++;
+
+    GGML_UNUSED(tensor);
+}
+
+static void ggml_dyn_tallocr_reset(struct ggml_dyn_tallocr * alloc) {
+    alloc->n_free_blocks = 1;
+    alloc->free_blocks[0].offset = 0;
+    alloc->free_blocks[0].size = SIZE_MAX/2; // restrict maximum size of a measure allocator to half size_t max to avoid overflows
+    alloc->max_size = 0;
+
+#ifdef GGML_ALLOCATOR_DEBUG
+    for (int i = 0; i < 1024; i++) {
+        alloc->allocated_tensors[i].tensor = NULL;
+    }
+#endif
+}
+
+static struct ggml_dyn_tallocr * ggml_dyn_tallocr_new(size_t alignment) {
+    struct ggml_dyn_tallocr * alloc = (struct ggml_dyn_tallocr *)malloc(sizeof(struct ggml_dyn_tallocr));
+
+    *alloc = (struct ggml_dyn_tallocr) {
+        /*.alignment     = */ alignment,
+        /*.n_free_blocks = */ 0,
+        /*.free_blocks   = */ {{0}},
+        /*.max_size      = */ 0,
+#ifdef GGML_ALLOCATOR_DEBUG
+        /*.allocated_tensors = */ {{0}},
+#endif
+    };
+
+    ggml_dyn_tallocr_reset(alloc);
+
+    return alloc;
+}
+
+static void ggml_dyn_tallocr_free(struct ggml_dyn_tallocr * alloc) {
+    free(alloc);
+}
+
+static size_t ggml_dyn_tallocr_max_size(struct ggml_dyn_tallocr * alloc) {
+    return alloc->max_size;
+}
+
+
+/////////////////////////////////////
+
+// graph allocator
+
+struct hash_node {
+    int n_children;
+    int n_views;
+    int buffer_id;
+    size_t offset; // offset within the buffer
+    bool allocated;
+};
+
+struct tensor_alloc {
+    int buffer_id;
+    size_t offset;
+    size_t size_max; // 0 = pre-allocated, unused, or view
+};
+
+struct leaf_alloc {
+    int buffer_id;
+    struct tensor_alloc leaf;
+};
+
+struct node_alloc {
+    struct tensor_alloc dst;
+    struct tensor_alloc src[GGML_MAX_SRC];
+};
+
+struct ggml_gallocr {
+    ggml_backend_buffer_type_t * bufts; // [n_buffers]
+    ggml_backend_buffer_t * buffers; // [n_buffers]
+    struct ggml_dyn_tallocr ** buf_tallocs; // [n_buffers]
+    int n_buffers;
+
+    struct ggml_hash_set hash_set;
+    struct hash_node * hash_values; // [hash_set.size]
+
+    struct node_alloc * node_allocs; // [n_nodes]
+    int n_nodes;
+
+    struct leaf_alloc * leaf_allocs; // [n_leafs]
+    int n_leafs;
+};
+
+ggml_gallocr_t ggml_gallocr_new_n(ggml_backend_buffer_type_t * bufts, int n_bufs) {
+    ggml_gallocr_t galloc = (ggml_gallocr_t)calloc(1, sizeof(struct ggml_gallocr));
+    GGML_ASSERT(galloc != NULL);
+
+    galloc->bufts = calloc(n_bufs, sizeof(ggml_backend_buffer_type_t));
+    GGML_ASSERT(galloc->bufts != NULL);
+
+    galloc->buffers = calloc(n_bufs, sizeof(ggml_backend_buffer_t));
+    GGML_ASSERT(galloc->buffers != NULL);
+
+    galloc->buf_tallocs = calloc(n_bufs, sizeof(struct ggml_dyn_tallocr *));
+    GGML_ASSERT(galloc->buf_tallocs != NULL);
+
+    for (int i = 0; i < n_bufs; i++) {
+        galloc->bufts[i] = bufts[i];
+        galloc->buffers[i] = NULL;
+
+        // check if the same buffer type is used multiple times and reuse the same allocator
+        for (int j = 0; j < i; j++) {
+            if (bufts[i] == bufts[j]) {
+                galloc->buf_tallocs[i] = galloc->buf_tallocs[j];
+                break;
+            }
+        }
+
+        if (galloc->buf_tallocs[i] == NULL) {
+            size_t alignment = ggml_backend_buft_get_alignment(bufts[i]);
+            galloc->buf_tallocs[i] = ggml_dyn_tallocr_new(alignment);
+        }
+    }
+    galloc->n_buffers = n_bufs;
+
+    return galloc;
+}
+
+ggml_gallocr_t ggml_gallocr_new(ggml_backend_buffer_type_t buft) {
+    return ggml_gallocr_new_n(&buft, 1);
+}
+
+void ggml_gallocr_free(ggml_gallocr_t galloc) {
+    if (galloc == NULL) {
+        return;
+    }
+
+    for (int i = 0; i < galloc->n_buffers; i++) {
+        if (galloc->buffers != NULL) {
+            // skip if already freed
+            bool freed = false;
+            for (int j = 0; j < i; j++) {
+                if (galloc->buffers[j] == galloc->buffers[i]) {
+                    freed = true;
+                    break;
+                }
+            }
+            if (!freed) {
+                ggml_backend_buffer_free(galloc->buffers[i]);
+            }
+        }
+        if (galloc->buf_tallocs != NULL) {
+            // skip if already freed
+            bool freed = false;
+            for (int j = 0; j < i; j++) {
+                if (galloc->buf_tallocs[j] == galloc->buf_tallocs[i]) {
+                    freed = true;
+                    break;
+                }
+            }
+            if (!freed) {
+                ggml_dyn_tallocr_free(galloc->buf_tallocs[i]);
+            }
+        }
+    }
+
+    ggml_hash_set_free(&galloc->hash_set);
+    free(galloc->hash_values);
+    free(galloc->bufts);
+    free(galloc->buffers);
+    free(galloc->buf_tallocs);
+    free(galloc->node_allocs);
+    free(galloc->leaf_allocs);
+    free(galloc);
+}
+
+typedef struct ggml_gallocr * ggml_gallocr_t;
+
+static struct hash_node * ggml_gallocr_hash_get(ggml_gallocr_t galloc, struct ggml_tensor * t) {
+    size_t i = ggml_hash_find_or_insert(&galloc->hash_set, t);
+    return &galloc->hash_values[i];
+}
+
+static bool ggml_gallocr_is_own(ggml_gallocr_t galloc, struct ggml_tensor * t) {
+    return ggml_gallocr_hash_get(galloc, t)->allocated;
+}
+
+static void ggml_gallocr_set_node_offset(ggml_gallocr_t galloc, struct ggml_tensor * node, int buffer_id, size_t offset) {
+    struct hash_node * hn = ggml_gallocr_hash_get(galloc, node);
+    hn->buffer_id = buffer_id;
+    hn->offset = offset;
+    hn->allocated = true;
+}
+
+static bool ggml_gallocr_is_allocated(ggml_gallocr_t galloc, struct ggml_tensor * t) {
+    return t->data != NULL || ggml_gallocr_hash_get(galloc, t)->allocated;
+}
+
+static void ggml_gallocr_allocate_node(ggml_gallocr_t galloc, struct ggml_tensor * node, int buffer_id) {
+    struct hash_node * hn = ggml_gallocr_hash_get(galloc, node);
+
+    if (!ggml_gallocr_is_allocated(galloc, node) && !ggml_is_view(node)) {
+        hn->allocated = true;
+        assert(hn->offset == 0);
+
+        // try to reuse a parent's buffer (inplace)
+        if (ggml_op_can_inplace(node->op)) {
+            for (int i = 0; i < GGML_MAX_SRC; i++) {
+                struct ggml_tensor * parent = node->src[i];
+                if (parent == NULL) {
+                    continue;
+                }
+
+                // if the node's data is external, then we cannot re-use it
+                if (!ggml_gallocr_is_own(galloc, parent)) {
+                    AT_PRINTF("not reusing parent %s for %s as %p is external\n", parent->name, node->name, parent->data);
+                    continue;
+                }
+
+                // outputs cannot be reused
+                if (parent->flags & GGML_TENSOR_FLAG_OUTPUT || (parent->view_src != NULL && parent->view_src->flags & GGML_TENSOR_FLAG_OUTPUT)) {
+                    AT_PRINTF("not reusing parent %s for %s as it is an output\n", parent->name, node->name);
+                    continue;
+                }
+
+                if (!ggml_are_same_layout(node, parent)) {
+                    AT_PRINTF("not reusing parent %s for %s as layouts are different\n", parent->name, node->name);
+                    continue;
+                }
+
+                struct hash_node * p_hn = ggml_gallocr_hash_get(galloc, parent);
+                if (p_hn->n_children == 1 && p_hn->n_views == 0) {
+                    if (ggml_is_view(parent)) {
+                        struct ggml_tensor * view_src = parent->view_src;
+                        struct hash_node * view_src_hn = ggml_gallocr_hash_get(galloc, view_src);
+                        if (view_src_hn->n_views == 1 && view_src_hn->n_children == 0 && view_src->data == parent->data) {
+                            AT_PRINTF("reusing view parent %s (%s) for %s\n", parent->name, view_src->name, node->name);
+                            assert(view_src_hn->offset == p_hn->offset);
+                            hn->buffer_id = p_hn->buffer_id;
+                            hn->offset = p_hn->offset;
+                            p_hn->allocated = false; // avoid freeing the parent
+                            view_src_hn->allocated = false;
+                            return;
+                        }
+                    } else {
+                        AT_PRINTF("reusing parent %s for %s\n", parent->name, node->name);
+                        hn->buffer_id = p_hn->buffer_id;
+                        hn->offset = p_hn->offset;
+                        p_hn->allocated = false; // avoid freeing the parent
+                        return;
+                    }
+                }
+            }
+        }
+        // allocate tensor from the buffer
+        struct ggml_dyn_tallocr * alloc = galloc->buf_tallocs[buffer_id];
+        ggml_backend_buffer_type_t buft = galloc->bufts[buffer_id];
+        size_t size = ggml_backend_buft_get_alloc_size(buft, node);
+        size_t offset = ggml_dyn_tallocr_alloc(alloc, size, node);
+        hn->buffer_id = buffer_id;
+        hn->offset = offset;
+        return;
+    }
+}
+
+static void ggml_gallocr_free_node(ggml_gallocr_t galloc, struct ggml_tensor * node) {
+    // graph outputs are never freed
+    if (node->flags & GGML_TENSOR_FLAG_OUTPUT) {
+        AT_PRINTF("not freeing output %s\n", node->name);
+        return;
+    }
+
+    struct hash_node * hn = ggml_gallocr_hash_get(galloc, node);
+    size_t offset = hn->offset;
+    int buffer_id = hn->buffer_id;
+    struct ggml_dyn_tallocr * alloc = galloc->buf_tallocs[buffer_id];
+    ggml_backend_buffer_type_t buft = galloc->bufts[buffer_id];
+    size_t size = ggml_backend_buft_get_alloc_size(buft, node);
+    ggml_dyn_tallocr_free_tensor(alloc, offset, size, node);
+    hn->allocated = false;
+}
+
+static int get_node_buffer_id(const int * node_buffer_ids, int i) {
+    return node_buffer_ids ? node_buffer_ids[i] : 0;
+}
+
+static void ggml_gallocr_alloc_graph_impl(ggml_gallocr_t galloc, struct ggml_cgraph * graph, const int * node_buffer_ids, const int * leaf_buffer_ids) {
+    // clear hash tables
+    ggml_hash_set_reset(&galloc->hash_set);
+    memset(galloc->hash_values, 0, sizeof(struct hash_node) * galloc->hash_set.size);
+
+    // allocate leafs
+    // these may be tensors that the application is not using in the graph, but may still want to allocate for other purposes
+    for (int i = 0; i < graph->n_leafs; i++) {
+        struct ggml_tensor * leaf = graph->leafs[i];
+        ggml_gallocr_allocate_node(galloc, leaf, get_node_buffer_id(leaf_buffer_ids, i));
+    }
+
+    // count number of children and views
+    // allocate other graph inputs and leafs first to avoid overwriting them
+    for (int i = 0; i < graph->n_nodes; i++) {
+        struct ggml_tensor * node = graph->nodes[i];
+
+        // TODO: better way to add external dependencies
+        // GGML_OP_NONE does not appear normally in the graph nodes, but is used by ggml-backend to add dependencies to
+        // control when some tensors are allocated and freed. in this case, the dependencies are in `src`, but the node
+        // itself is never used and should not be considered a dependency
+        if (ggml_is_view(node) && node->op != GGML_OP_NONE) {
+            struct ggml_tensor * view_src = node->view_src;
+            ggml_gallocr_hash_get(galloc, view_src)->n_views += 1;
+        }
+
+        if (node->flags & GGML_TENSOR_FLAG_INPUT) {
+            ggml_gallocr_allocate_node(galloc, graph->nodes[i], get_node_buffer_id(node_buffer_ids, i));
+        }
+
+        for (int j = 0; j < GGML_MAX_SRC; j++) {
+            struct ggml_tensor * src = node->src[j];
+            if (src == NULL) {
+                continue;
+            }
+
+            ggml_gallocr_hash_get(galloc, src)->n_children += 1;
+
+            // allocate explicit inputs
+            if (src->flags & GGML_TENSOR_FLAG_INPUT) {
+                ggml_gallocr_allocate_node(galloc, src, get_node_buffer_id(node_buffer_ids, i));
+            }
+        }
+    }
+
+    // allocate tensors
+    for (int i = 0; i < graph->n_nodes; i++) {
+        struct ggml_tensor * node = graph->nodes[i];
+        int buffer_id = get_node_buffer_id(node_buffer_ids, i);
+
+        // allocate parents (only leafs need to be allocated at this point)
+        for (int j = 0; j < GGML_MAX_SRC; j++) {
+            struct ggml_tensor * parent = node->src[j];
+            if (parent == NULL) {
+                continue;
+            }
+            ggml_gallocr_allocate_node(galloc, parent, buffer_id);
+        }
+
+        // allocate node
+        ggml_gallocr_allocate_node(galloc, node, buffer_id);
+
+        AT_PRINTF("exec: %s (%s) <= ", ggml_op_desc(node), node->name);
+        for (int j = 0; j < GGML_MAX_SRC; j++) {
+            struct ggml_tensor * parent = node->src[j];
+            if (parent == NULL) {
+                continue;
+            }
+            AT_PRINTF("%s", parent->name);
+            if (j < GGML_MAX_SRC - 1 && node->src[j + 1] != NULL) {
+                AT_PRINTF(", ");
+            }
+        }
+        AT_PRINTF("\n");
+
+        // update parents
+        for (int j = 0; j < GGML_MAX_SRC; j++) {
+            struct ggml_tensor * parent = node->src[j];
+            if (parent == NULL) {
+                continue;
+            }
+            struct hash_node * p_hn = ggml_gallocr_hash_get(galloc, parent);
+            p_hn->n_children -= 1;
+
+            AT_PRINTF("parent %s: %d children, %d views, allocated: %d\n",
+                parent->name, p_hn->n_children, p_hn->n_views, p_hn->allocated);
+
+            if (p_hn->n_children == 0 && p_hn->n_views == 0) {
+                if (ggml_is_view(parent)) {
+                    struct ggml_tensor * view_src = parent->view_src;
+                    struct hash_node * view_src_hn = ggml_gallocr_hash_get(galloc, view_src);
+                    view_src_hn->n_views -= 1;
+                    AT_PRINTF("view_src %s: %d children, %d views\n",
+                        view_src->name, view_src_hn->n_children, view_src_hn->n_views);
+                    if (view_src_hn->n_views == 0 && view_src_hn->n_children == 0 && view_src_hn->allocated) {
+                        ggml_gallocr_free_node(galloc, view_src);
+                    }
+                }
+                else if (p_hn->allocated) {
+                    ggml_gallocr_free_node(galloc, parent);
+                }
+            }
+            AT_PRINTF("\n");
+        }
+    }
+}
+
+bool ggml_gallocr_reserve_n(ggml_gallocr_t galloc, struct ggml_cgraph * graph, const int * node_buffer_ids, const int * leaf_buffer_ids) {
+    size_t min_hash_size = graph->n_nodes + graph->n_leafs;
+    // add 25% margin to avoid hash collisions
+    min_hash_size += min_hash_size / 4;
+
+    // initialize hash table
+    if (galloc->hash_set.size < min_hash_size) {
+        ggml_hash_set_free(&galloc->hash_set);
+        galloc->hash_set = ggml_hash_set_new(min_hash_size);
+        GGML_ASSERT(galloc->hash_set.keys != NULL);
+
+        free(galloc->hash_values);
+        galloc->hash_values = malloc(sizeof(struct hash_node) * galloc->hash_set.size);
+        GGML_ASSERT(galloc->hash_values != NULL);
+    }
+
+    // reset allocators
+    for (int i = 0; i < galloc->n_buffers; i++) {
+        ggml_dyn_tallocr_reset(galloc->buf_tallocs[i]);
+    }
+
+    // allocate in hash table
+    ggml_gallocr_alloc_graph_impl(galloc, graph, node_buffer_ids, leaf_buffer_ids);
+
+    // set the node_allocs from the hash table
+    if (galloc->n_nodes < graph->n_nodes) {
+        free(galloc->node_allocs);
+        galloc->node_allocs = calloc(graph->n_nodes, sizeof(struct node_alloc));
+        GGML_ASSERT(galloc->node_allocs != NULL);
+    }
+    galloc->n_nodes = graph->n_nodes;
+    for (int i = 0; i < graph->n_nodes; i++) {
+        struct ggml_tensor * node = graph->nodes[i];
+        struct node_alloc * node_alloc = &galloc->node_allocs[i];
+        if (node->view_src || node->data) {
+            node_alloc->dst.buffer_id = -1;
+            node_alloc->dst.offset = SIZE_MAX;
+            node_alloc->dst.size_max = 0;
+        } else {
+            struct hash_node * hn = ggml_gallocr_hash_get(galloc, node);
+            node_alloc->dst.buffer_id = hn->buffer_id;
+            node_alloc->dst.offset    = hn->offset;
+            node_alloc->dst.size_max  = ggml_backend_buft_get_alloc_size(galloc->bufts[hn->buffer_id], node);
+        }
+        for (int j = 0; j < GGML_MAX_SRC; j++) {
+            struct ggml_tensor * src = node->src[j];
+            if (!src || src->view_src || src->data) {
+                node_alloc->src[j].buffer_id = -1;
+                node_alloc->src[j].offset = SIZE_MAX;
+                node_alloc->src[j].size_max = 0;
+            } else {
+                struct hash_node * hn = ggml_gallocr_hash_get(galloc, src);
+                node_alloc->src[j].buffer_id = hn->buffer_id;
+                node_alloc->src[j].offset   = hn->offset;
+                node_alloc->src[j].size_max = ggml_backend_buft_get_alloc_size(galloc->bufts[hn->buffer_id], src);
+            }
+        }
+    }
+    if (galloc->n_leafs < graph->n_leafs) {
+        free(galloc->leaf_allocs);
+        galloc->leaf_allocs = calloc(graph->n_leafs, sizeof(galloc->leaf_allocs[0]));
+        GGML_ASSERT(galloc->leaf_allocs != NULL);
+    }
+    galloc->n_leafs = graph->n_leafs;
+    for (int i = 0; i < graph->n_leafs; i++) {
+        struct ggml_tensor * leaf = graph->leafs[i];
+        struct hash_node * hn = ggml_gallocr_hash_get(galloc, leaf);
+        galloc->leaf_allocs[i].buffer_id = hn->buffer_id;
+        if (leaf->view_src || leaf->data) {
+            galloc->leaf_allocs[i].leaf.buffer_id = -1;
+            galloc->leaf_allocs[i].leaf.offset = SIZE_MAX;
+            galloc->leaf_allocs[i].leaf.size_max = 0;
+        } else {
+            galloc->leaf_allocs[i].leaf.buffer_id = hn->buffer_id;
+            galloc->leaf_allocs[i].leaf.offset = hn->offset;
+            galloc->leaf_allocs[i].leaf.size_max = ggml_backend_buft_get_alloc_size(galloc->bufts[hn->buffer_id], leaf);
+        }
+    }
+
+    // reallocate buffers if needed
+    for (int i = 0; i < galloc->n_buffers; i++) {
+        // if the buffer type is used multiple times, we reuse the same buffer
+        for (int j = 0; j < i; j++) {
+            if (galloc->buf_tallocs[j] == galloc->buf_tallocs[i]) {
+                galloc->buffers[i] = galloc->buffers[j];
+                break;
+            }
+        }
+
+        size_t cur_size = galloc->buffers[i] ? ggml_backend_buffer_get_size(galloc->buffers[i]) : 0;
+        size_t new_size = ggml_dyn_tallocr_max_size(galloc->buf_tallocs[i]);
+
+        // even if there are no tensors allocated in this buffer, we still need to allocate it to initialize views
+        if (new_size > cur_size || galloc->buffers[i] == NULL) {
+#ifndef NDEBUG
+            fprintf(stderr, "%s: reallocating %s buffer from size %.02f MiB to %.02f MiB\n", __func__, ggml_backend_buft_name(galloc->bufts[i]), cur_size / 1024.0 / 1024.0, new_size / 1024.0 / 1024.0);
+#endif
+
+            ggml_backend_buffer_free(galloc->buffers[i]);
+            galloc->buffers[i] = ggml_backend_buft_alloc_buffer(galloc->bufts[i], new_size);
+            if (galloc->buffers[i] == NULL) {
+                fprintf(stderr, "%s: failed to allocate %s buffer of size %zu\n", __func__, ggml_backend_buft_name(galloc->bufts[i]), new_size);
+                return false;
+            }
+            ggml_backend_buffer_set_usage(galloc->buffers[i], GGML_BACKEND_BUFFER_USAGE_COMPUTE);
+        }
+    }
+
+    return true;
+}
+
+bool ggml_gallocr_reserve(ggml_gallocr_t galloc, struct ggml_cgraph *graph) {
+    return ggml_gallocr_reserve_n(galloc, graph, NULL, NULL);
+}
+
+static void ggml_gallocr_init_tensor(ggml_gallocr_t galloc, struct ggml_tensor * tensor, struct tensor_alloc * tensor_alloc) {
+    int buffer_id = tensor_alloc->buffer_id;
+    assert(tensor->data || tensor->view_src || ggml_backend_buffer_get_alloc_size(galloc->buffers[buffer_id], tensor) <= tensor_alloc->size_max);
+
+    if (tensor->view_src != NULL) {
+        if (tensor->buffer == NULL) {
+            assert(tensor_alloc->offset == SIZE_MAX);
+            if (tensor->view_src->buffer == NULL) {
+                // this tensor was allocated without ggml-backend
+                return;
+            }
+            ggml_backend_view_init(tensor);
+        }
+    } else {
+        if (tensor->data == NULL) {
+            assert(tensor_alloc->offset != SIZE_MAX);
+            assert(ggml_backend_buffer_get_alloc_size(galloc->buffers[buffer_id], tensor) <= tensor_alloc->size_max);
+            void * base = ggml_backend_buffer_get_base(galloc->buffers[buffer_id]);
+            void * addr = (char *)base + tensor_alloc->offset;
+            ggml_backend_tensor_alloc(galloc->buffers[buffer_id], tensor, addr);
+        } else {
+            if (tensor->buffer == NULL) {
+                // this tensor was allocated without ggml-backend
+                return;
+            }
+        }
+    }
+}
+
+static bool ggml_gallocr_node_needs_realloc(ggml_gallocr_t galloc, struct ggml_tensor * node, struct tensor_alloc * talloc) {
+    size_t node_size = (node->data || node->view_src) ? 0 : ggml_backend_buft_get_alloc_size(galloc->bufts[talloc->buffer_id], node);
+    return talloc->size_max >= node_size;
+}
+
+static bool ggml_gallocr_needs_realloc(ggml_gallocr_t galloc, struct ggml_cgraph * graph) {
+    if (galloc->n_nodes != graph->n_nodes) {
+#ifndef NDEBUG
+        fprintf(stderr, "%s: graph has different number of nodes\n", __func__);
+#endif
+        return true;
+    }
+
+    if (galloc->n_leafs != graph->n_leafs) {
+#ifndef NDEBUG
+        fprintf(stderr, "%s: graph has different number of leafs\n", __func__);
+#endif
+        return true;
+    }
+
+    for (int i = 0; i < graph->n_nodes; i++) {
+        struct ggml_tensor * node = graph->nodes[i];
+        struct node_alloc * node_alloc = &galloc->node_allocs[i];
+
+        if (!ggml_gallocr_node_needs_realloc(galloc, node, &node_alloc->dst)) {
+#ifndef NDEBUG
+            fprintf(stderr, "%s: node %s is not valid\n", __func__, node->name);
+#endif
+            return true;
+        }
+
+        for (int j = 0; j < GGML_MAX_SRC; j++) {
+            struct ggml_tensor * src = node->src[j];
+            if (src == NULL) {
+                continue;
+            }
+            if (!ggml_gallocr_node_needs_realloc(galloc, src, &node_alloc->src[j])) {
+#ifndef NDEBUG
+                fprintf(stderr, "%s: src %d (%s) of node %s is not valid\n", __func__, j, src->name, node->name);
+#endif
+                return true;
+            }
+        }
+    }
+
+    return false;
+}
+
+bool ggml_gallocr_alloc_graph(ggml_gallocr_t galloc, struct ggml_cgraph * graph) {
+    if (ggml_gallocr_needs_realloc(galloc, graph)) {
+        if (galloc->n_buffers == 1) {
+#ifndef NDEBUG
+            fprintf(stderr, "%s: reallocating buffers automatically\n", __func__);
+#endif
+            if (!ggml_gallocr_reserve(galloc, graph)) {
+                return false;
+            }
+        } else {
+#ifndef NDEBUG
+            fprintf(stderr, "%s: cannot reallocate multi buffer graph automatically, call reserve\n", __func__);
+#endif
+            return false;
+        }
+    }
+
+    // reset buffers
+    for (int i = 0; i < galloc->n_buffers; i++) {
+        if (galloc->buffers[i] != NULL) {
+            ggml_backend_buffer_reset(galloc->buffers[i]);
+        }
+    }
+
+    // allocate the graph tensors from the previous assignments
+    // leafs
+    for (int i = 0; i < graph->n_leafs; i++) {
+        struct ggml_tensor * leaf = graph->leafs[i];
+        struct leaf_alloc * leaf_alloc = &galloc->leaf_allocs[i];
+        ggml_gallocr_init_tensor(galloc, leaf, &leaf_alloc->leaf);
+    }
+    // nodes
+    for (int i = 0; i < graph->n_nodes; i++) {
+        struct ggml_tensor * node = graph->nodes[i];
+        struct node_alloc * node_alloc = &galloc->node_allocs[i];
+        for (int j = 0; j < GGML_MAX_SRC; j++) {
+            struct ggml_tensor * src = node->src[j];
+            if (src == NULL) {
+                continue;
+            }
+            ggml_gallocr_init_tensor(galloc, src, &node_alloc->src[j]);
+        }
+        ggml_gallocr_init_tensor(galloc, node, &node_alloc->dst);
+    }
+
+    return true;
+}
+
+size_t ggml_gallocr_get_buffer_size(ggml_gallocr_t galloc, int buffer_id) {
+    GGML_ASSERT(buffer_id >= 0 && buffer_id < galloc->n_buffers);
+
+    if (galloc->buffers[buffer_id] == NULL) {
+        return 0;
+    }
+
+    for (int i = 0; i < buffer_id; i++) {
+        if (galloc->buffers[i] == galloc->buffers[buffer_id]) {
+            // this buffer is the same as a previous one due to the same buffer type being used multiple times
+            // only return the buffer size the first time it appears to avoid double counting
+            return 0;
+        }
+    }
+
+    return ggml_backend_buffer_get_size(galloc->buffers[buffer_id]);
+}
+
+// utils
+
+static bool alloc_tensor_range(struct ggml_context * ctx,
+        struct ggml_tensor * first, struct ggml_tensor * last,
+        ggml_backend_buffer_type_t buft, size_t size,
+        ggml_backend_buffer_t ** buffers, size_t * n_buffers) {
+    ggml_backend_buffer_t buffer = ggml_backend_buft_alloc_buffer(buft, size);
+    if (buffer == NULL) {
+#ifndef NDEBUG
+        fprintf(stderr, "%s: failed to allocate %s buffer of size %zu\n", __func__, ggml_backend_buft_name(buft), size);
+#endif
+        for (size_t i = 0; i < *n_buffers; i++) {
+            ggml_backend_buffer_free((*buffers)[i]);
+        }
+        free(*buffers);
+        return false;
+    }
+
+    struct ggml_tallocr tallocr = ggml_tallocr_new(buffer);
+
+    for (struct ggml_tensor * t = first; t != last; t = ggml_get_next_tensor(ctx, t)) {
+        if (t->data == NULL) {
+            if (t->view_src == NULL) {
+                ggml_tallocr_alloc(&tallocr, t);
+            } else if (t->buffer == NULL) {
+                ggml_backend_view_init(t);
+            }
+        } else {
+            if (t->view_src != NULL && t->buffer == NULL) {
+                // view of a pre-allocated tensor
+                ggml_backend_view_init(t);
+            }
+        }
+    }
+
+    *buffers = realloc(*buffers, sizeof(ggml_backend_buffer_t) * (*n_buffers + 1));
+    (*buffers)[(*n_buffers)++] = buffer;
+
+    return true;
+}
+
+ggml_backend_buffer_t ggml_backend_alloc_ctx_tensors_from_buft(struct ggml_context * ctx, ggml_backend_buffer_type_t buft) {
+    GGML_ASSERT(ggml_get_no_alloc(ctx) == true);
+
+    size_t alignment = ggml_backend_buft_get_alignment(buft);
+    size_t max_size = ggml_backend_buft_get_max_size(buft);
+
+    ggml_backend_buffer_t * buffers = NULL;
+    size_t n_buffers = 0;
+
+    size_t cur_buf_size = 0;
+    struct ggml_tensor * first = ggml_get_first_tensor(ctx);
+    for (struct ggml_tensor * t = first; t != NULL; t = ggml_get_next_tensor(ctx, t)) {
+        size_t this_size = 0;
+        if (t->data == NULL && t->view_src == NULL) {
+            this_size = GGML_PAD(ggml_backend_buft_get_alloc_size(buft, t), alignment);
+        }
+
+        if (this_size > max_size) {
+            fprintf(stderr, "%s: tensor %s is too large to fit in a %s buffer (tensor size: %zu, max buffer size: %zu)\n",
+                    __func__, t->name,
+                    ggml_backend_buft_name(buft),
+                    this_size, max_size);
+            for (size_t i = 0; i < n_buffers; i++) {
+                ggml_backend_buffer_free(buffers[i]);
+            }
+            free(buffers);
+            return NULL;
+        }
+
+        if ((cur_buf_size + this_size) > max_size) {
+            // allocate tensors in the current buffer
+            if (!alloc_tensor_range(ctx, first, t, buft, cur_buf_size, &buffers, &n_buffers)) {
+                return NULL;
+            }
+            first = t;
+            cur_buf_size = this_size;
+        } else {
+            cur_buf_size += this_size;
+        }
+    }
+
+    // allocate remaining tensors
+    if (cur_buf_size > 0) {
+        if (!alloc_tensor_range(ctx, first, NULL, buft, cur_buf_size, &buffers, &n_buffers)) {
+            return NULL;
+        }
+    }
+
+    if (n_buffers == 0) {
+#ifndef NDEBUG
+        fprintf(stderr, "%s: all tensors in the context are already allocated\n", __func__);
+#endif
+        return NULL;
+    }
+
+    ggml_backend_buffer_t buffer;
+    if (n_buffers == 1) {
+        buffer = buffers[0];
+    } else {
+        buffer = ggml_backend_multi_buffer_alloc_buffer(buffers, n_buffers);
+    }
+    free(buffers);
+    return buffer;
+}
+
+ggml_backend_buffer_t ggml_backend_alloc_ctx_tensors(struct ggml_context * ctx, ggml_backend_t backend) {
+    return ggml_backend_alloc_ctx_tensors_from_buft(ctx, ggml_backend_get_default_buffer_type(backend));
+}
diff --git a/src/whisper_cpp/ggml/src/ggml-backend-impl.h b/src/whisper_cpp/ggml/src/ggml-backend-impl.h
new file mode 100644
index 00000000..ba2e2699
--- /dev/null
+++ b/src/whisper_cpp/ggml/src/ggml-backend-impl.h
@@ -0,0 +1,235 @@
+#pragma once
+
+// ggml-backend internal header
+
+#include "ggml-backend.h"
+
+#ifdef  __cplusplus
+extern "C" {
+#endif
+
+    //
+    // Backend buffer type
+    //
+
+    struct ggml_backend_buffer_type_i {
+        const char *          (*get_name)      (ggml_backend_buffer_type_t buft);
+        // allocate a buffer of this type
+        ggml_backend_buffer_t (*alloc_buffer)  (ggml_backend_buffer_type_t buft, size_t size);
+        // tensor alignment
+        size_t                (*get_alignment) (ggml_backend_buffer_type_t buft);
+        // (optional) max buffer size that can be allocated (defaults to SIZE_MAX)
+        size_t                (*get_max_size)  (ggml_backend_buffer_type_t buft);
+        // (optional) data size needed to allocate the tensor, including padding (defaults to ggml_nbytes)
+        size_t                (*get_alloc_size)(ggml_backend_buffer_type_t buft, const struct ggml_tensor * tensor);
+        // (optional) check if tensor data is in host memory (defaults to false)
+        bool                  (*is_host)       (ggml_backend_buffer_type_t buft);
+    };
+
+    struct ggml_backend_buffer_type {
+        struct ggml_backend_buffer_type_i  iface;
+        ggml_backend_dev_t device;
+        void * context;
+    };
+
+    //
+    // Backend buffer
+    //
+
+    struct ggml_backend_buffer_i {
+        const char * (*get_name)     (ggml_backend_buffer_t buffer);
+        // (optional) free the buffer
+        void         (*free_buffer)  (ggml_backend_buffer_t buffer);
+        // base address of the buffer
+        void *       (*get_base)     (ggml_backend_buffer_t buffer);
+        // (optional) initialize a tensor in the buffer (eg. add tensor extras)
+        void         (*init_tensor)  (ggml_backend_buffer_t buffer, struct ggml_tensor * tensor);
+        // tensor data access
+        void         (*memset_tensor)(ggml_backend_buffer_t buffer,       struct ggml_tensor * tensor,     uint8_t value, size_t offset, size_t size);
+        void         (*set_tensor)   (ggml_backend_buffer_t buffer,       struct ggml_tensor * tensor, const void * data, size_t offset, size_t size);
+        void         (*get_tensor)   (ggml_backend_buffer_t buffer, const struct ggml_tensor * tensor,       void * data, size_t offset, size_t size);
+        // (optional) tensor copy: dst is in the buffer, src may be in any buffer, including buffers from a different backend (return false if not supported)
+        bool         (*cpy_tensor)   (ggml_backend_buffer_t buffer, const struct ggml_tensor * src, struct ggml_tensor * dst);
+        // clear the entire buffer
+        void         (*clear)        (ggml_backend_buffer_t buffer, uint8_t value);
+        // (optional) reset any internal state due to tensor initialization, such as tensor extras
+        void         (*reset)        (ggml_backend_buffer_t buffer);
+    };
+
+    struct ggml_backend_buffer {
+        struct ggml_backend_buffer_i  iface;
+        ggml_backend_buffer_type_t    buft;
+        void * context;
+        size_t size;
+        enum ggml_backend_buffer_usage usage;
+    };
+
+    ggml_backend_buffer_t ggml_backend_buffer_init(
+                   ggml_backend_buffer_type_t buft,
+            struct ggml_backend_buffer_i      iface,
+                   void *                     context,
+                   size_t                     size);
+
+    // do not use directly, use ggml_backend_tensor_copy instead
+    bool ggml_backend_buffer_copy_tensor(const struct ggml_tensor * src, struct ggml_tensor * dst);
+
+    // multi-buffer
+    // buffer that contains a collection of buffers
+    ggml_backend_buffer_t ggml_backend_multi_buffer_alloc_buffer(ggml_backend_buffer_t * buffers, size_t n_buffers);
+    bool                  ggml_backend_buffer_is_multi_buffer(ggml_backend_buffer_t buffer);
+    void                  ggml_backend_multi_buffer_set_usage(ggml_backend_buffer_t buffer, enum ggml_backend_buffer_usage usage);
+
+    //
+    // Backend (stream)
+    //
+
+    struct ggml_backend_i {
+        const char * (*get_name)(ggml_backend_t backend);
+
+        void (*free)(ggml_backend_t backend);
+
+        // buffer allocation
+        ggml_backend_buffer_type_t (*get_default_buffer_type)(ggml_backend_t backend);
+
+        // (optional) asynchronous tensor data access
+        void (*set_tensor_async)(ggml_backend_t backend,       struct ggml_tensor * tensor, const void * data, size_t offset, size_t size);
+        void (*get_tensor_async)(ggml_backend_t backend, const struct ggml_tensor * tensor,       void * data, size_t offset, size_t size);
+        bool (*cpy_tensor_async)(ggml_backend_t backend_src, ggml_backend_t backend_dst, const struct ggml_tensor * src, struct ggml_tensor * dst);
+
+        // (optional) complete all pending operations
+        void (*synchronize)(ggml_backend_t backend);
+
+        // (optional) compute graph with a plan (not used currently)
+        ggml_backend_graph_plan_t (*graph_plan_create) (ggml_backend_t backend, const struct ggml_cgraph * cgraph);
+        void                      (*graph_plan_free)   (ggml_backend_t backend, ggml_backend_graph_plan_t plan);
+        // update the plan with a new graph - this should be faster than creating a new plan when the graph has the same topology
+        void                      (*graph_plan_update) (ggml_backend_t backend, ggml_backend_graph_plan_t plan, const struct ggml_cgraph * cgraph);
+        // compute the graph with the plan
+        enum ggml_status          (*graph_plan_compute)(ggml_backend_t backend, ggml_backend_graph_plan_t plan);
+
+        // compute graph (always async if supported by the backend)
+        enum ggml_status          (*graph_compute)     (ggml_backend_t backend, struct ggml_cgraph * cgraph);
+
+        // IMPORTANT: these functions have been moved to the device interface and will be removed from the backend interface
+        //            new backends should implement the device interface instead
+
+        // These functions are being moved to the device interface
+        // check if the backend can compute an operation
+        bool (*supports_op)  (ggml_backend_t backend, const struct ggml_tensor * op);
+
+        // check if the backend can use tensors allocated in a buffer type
+        bool (*supports_buft)(ggml_backend_t backend, ggml_backend_buffer_type_t buft);
+
+        // check if the backend wants to run an operation, even if the weights are allocated in a CPU buffer
+        // these should be expensive operations with large batch sizes that may benefit from running on this backend
+        // even if the weight has to be copied from the CPU temporarily
+        bool (*offload_op)   (ggml_backend_t backend, const struct ggml_tensor * op);
+
+        // (optional) event synchronization
+        // record an event on this stream
+        void (*event_record)(ggml_backend_t backend, ggml_backend_event_t event);
+        // wait for an event on on a different stream
+        void (*event_wait)  (ggml_backend_t backend, ggml_backend_event_t event);
+    };
+
+    struct ggml_backend {
+        ggml_guid_t guid;
+        struct ggml_backend_i iface;
+        ggml_backend_dev_t device;
+        void * context;
+    };
+
+    struct ggml_backend_event {
+        struct ggml_backend_device * device;
+        void * context;
+    };
+
+    //
+    // Backend device
+    //
+
+    // Note: if additional properties are needed, we should add a struct with all of them
+    //       the current functions to obtain the properties can remain, since they are more convenient for often used properties
+    struct ggml_backend_device_i {
+        // device name: short identifier for this device, such as "CPU" or "CUDA0"
+        const char * (*get_name)(ggml_backend_dev_t dev);
+
+        // device description: short informative description of the device, could be the model name
+        const char * (*get_description)(ggml_backend_dev_t dev);
+
+        // device memory in bytes
+        void         (*get_memory)(ggml_backend_dev_t dev, size_t * free, size_t * total);
+
+        // device type
+        enum ggml_backend_dev_type (*get_type)(ggml_backend_dev_t dev);
+
+        // device properties
+        void (*get_props)(ggml_backend_dev_t dev, struct ggml_backend_dev_props * props);
+
+        // backend (stream) initialization
+        ggml_backend_t (*init_backend)(ggml_backend_dev_t dev, const char * params);
+
+        // preferred buffer type
+        ggml_backend_buffer_type_t (*get_buffer_type)(ggml_backend_dev_t dev);
+
+        // (optional) host buffer type (in system memory, typically this is a pinned memory buffer for faster transfers between host and device)
+        ggml_backend_buffer_type_t (*get_host_buffer_type)(ggml_backend_dev_t dev);
+
+        // (optional) buffer from pointer: create a buffer from a host pointer (useful for memory mapped models and importing data from other libraries)
+        ggml_backend_buffer_t (*buffer_from_host_ptr)(ggml_backend_dev_t dev, void * ptr, size_t size, size_t max_tensor_size);
+
+        // check if the backend can compute an operation
+        bool (*supports_op)(ggml_backend_dev_t dev, const struct ggml_tensor * op);
+
+        // check if the backend can use tensors allocated in a buffer type
+        bool (*supports_buft)(ggml_backend_dev_t dev, ggml_backend_buffer_type_t buft);
+
+        // check if the backend wants to run an operation, even if the weights are allocated in a CPU buffer
+        // these should be expensive operations with large batch sizes that may benefit from running on this backend
+        // even if the weight has to be copied from the CPU temporarily
+        bool (*offload_op)(ggml_backend_dev_t dev, const struct ggml_tensor * op);
+
+        // (optional) event synchronization
+        ggml_backend_event_t (*event_new)         (ggml_backend_dev_t dev);
+        void                 (*event_free)        (ggml_backend_dev_t dev, ggml_backend_event_t event);
+        void                 (*event_synchronize) (ggml_backend_dev_t dev, ggml_backend_event_t event);
+    };
+
+    struct ggml_backend_device {
+        struct ggml_backend_device_i iface;
+        ggml_backend_reg_t reg;
+        void * context;
+    };
+
+    //
+    // Backend (reg)
+    //
+
+    struct ggml_backend_reg_i {
+        const char * (*get_name)(ggml_backend_reg_t reg);
+
+        // enumerate available devices
+        size_t             (*get_device_count)(ggml_backend_reg_t reg);
+        ggml_backend_dev_t (*get_device)(ggml_backend_reg_t reg, size_t index);
+
+        // (optional) get a pointer to a function in the backend
+        // backends can add custom functions that are not part of the standard ggml-backend interface
+        void * (*get_proc_address)(ggml_backend_reg_t reg, const char * name);
+    };
+
+    struct ggml_backend_reg {
+        // int api_version; // TODO: for dynamic loading
+        struct ggml_backend_reg_i iface;
+        void * context;
+    };
+
+
+    // Internal backend registry API
+    void ggml_backend_register(ggml_backend_reg_t reg);
+    void ggml_backend_device_register(ggml_backend_dev_t device);
+    // TODO: backends can be loaded as a dynamic library, in which case it needs to export this function
+    // typedef ggml_backend_register_t * (*ggml_backend_init)(void);
+
+#ifdef  __cplusplus
+}
+#endif
diff --git a/src/whisper_cpp/ggml/src/ggml-backend.cpp b/src/whisper_cpp/ggml/src/ggml-backend.cpp
new file mode 100644
index 00000000..0551764f
--- /dev/null
+++ b/src/whisper_cpp/ggml/src/ggml-backend.cpp
@@ -0,0 +1,2560 @@
+// Note: porting this file to C++ is a work in progress
+
+#ifdef _WIN32
+#define WIN32_LEAN_AND_MEAN
+#ifndef NOMINMAX
+#   define NOMINMAX
+#endif
+#include <windows.h>
+#endif
+
+#include "ggml-backend-impl.h"
+#include "ggml-alloc.h"
+#include "ggml-impl.h"
+
+#include <assert.h>
+#include <limits.h>
+#include <stdarg.h>
+#include <stdio.h>
+#include <stdlib.h>
+#include <string.h>
+#include <string>
+#include <vector>
+
+#ifdef __APPLE__
+#include <sys/types.h>
+#include <sys/sysctl.h>
+#endif
+
+
+// backend buffer type
+
+const char * ggml_backend_buft_name(ggml_backend_buffer_type_t buft) {
+    return buft->iface.get_name(buft);
+}
+
+ggml_backend_buffer_t ggml_backend_buft_alloc_buffer(ggml_backend_buffer_type_t buft, size_t size) {
+    return buft->iface.alloc_buffer(buft, size);
+}
+
+size_t ggml_backend_buft_get_alignment(ggml_backend_buffer_type_t buft) {
+    return buft->iface.get_alignment(buft);
+}
+
+size_t ggml_backend_buft_get_max_size(ggml_backend_buffer_type_t buft) {
+    // get_max_size is optional, defaults to SIZE_MAX
+    if (buft->iface.get_max_size) {
+        return buft->iface.get_max_size(buft);
+    }
+    return SIZE_MAX;
+}
+
+size_t ggml_backend_buft_get_alloc_size(ggml_backend_buffer_type_t buft, struct ggml_tensor * tensor) {
+    // get_alloc_size is optional, defaults to ggml_nbytes
+    if (buft->iface.get_alloc_size) {
+        size_t size = buft->iface.get_alloc_size(buft, tensor);
+        assert(size >= ggml_nbytes(tensor));
+        return size;
+    }
+    return ggml_nbytes(tensor);
+}
+
+bool ggml_backend_buft_is_host(ggml_backend_buffer_type_t buft) {
+    if (buft->iface.is_host) {
+        return buft->iface.is_host(buft);
+    }
+    return false;
+}
+
+ggml_backend_dev_t ggml_backend_buft_get_device(ggml_backend_buffer_type_t buft) {
+    return buft->device;
+}
+
+// backend buffer
+
+ggml_backend_buffer_t ggml_backend_buffer_init(
+               ggml_backend_buffer_type_t buft,
+        struct ggml_backend_buffer_i      iface,
+               void *                     context,
+               size_t                     size) {
+    ggml_backend_buffer_t buffer = new ggml_backend_buffer {
+        /* .interface = */ iface,
+        /* .buft      = */ buft,
+        /* .context   = */ context,
+        /* .size      = */ size,
+        /* .usage     = */ GGML_BACKEND_BUFFER_USAGE_ANY
+    };
+
+    return buffer;
+}
+
+const char * ggml_backend_buffer_name(ggml_backend_buffer_t buffer) {
+    return buffer->iface.get_name(buffer);
+}
+
+void ggml_backend_buffer_free(ggml_backend_buffer_t buffer) {
+    if (buffer == NULL) {
+        return;
+    }
+
+    if (buffer->iface.free_buffer != NULL) {
+        buffer->iface.free_buffer(buffer);
+    }
+    delete buffer;
+}
+
+size_t ggml_backend_buffer_get_size(ggml_backend_buffer_t buffer) {
+    return buffer->size;
+}
+
+void * ggml_backend_buffer_get_base(ggml_backend_buffer_t buffer) {
+    void * base = buffer->iface.get_base(buffer);
+
+    GGML_ASSERT(base != NULL && "backend buffer base cannot be NULL");
+
+    return base;
+}
+
+void ggml_backend_buffer_init_tensor(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor) {
+    // init_tensor is optional
+    if (buffer->iface.init_tensor) {
+        buffer->iface.init_tensor(buffer, tensor);
+    }
+}
+
+size_t ggml_backend_buffer_get_alignment(ggml_backend_buffer_t buffer) {
+    return ggml_backend_buft_get_alignment(ggml_backend_buffer_get_type(buffer));
+}
+
+size_t ggml_backend_buffer_get_max_size(ggml_backend_buffer_t buffer) {
+    return ggml_backend_buft_get_max_size(ggml_backend_buffer_get_type(buffer));
+}
+
+size_t ggml_backend_buffer_get_alloc_size(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor) {
+    return ggml_backend_buft_get_alloc_size(ggml_backend_buffer_get_type(buffer), tensor);
+}
+
+void ggml_backend_buffer_clear(ggml_backend_buffer_t buffer, uint8_t value) {
+    buffer->iface.clear(buffer, value);
+}
+
+bool ggml_backend_buffer_is_host(ggml_backend_buffer_t buffer) {
+    return ggml_backend_buft_is_host(ggml_backend_buffer_get_type(buffer));
+}
+
+void ggml_backend_buffer_set_usage(ggml_backend_buffer_t buffer, enum ggml_backend_buffer_usage usage) {
+    buffer->usage = usage;
+
+    // FIXME: add a generic callback to the buffer interface
+    if (ggml_backend_buffer_is_multi_buffer(buffer)) {
+        ggml_backend_multi_buffer_set_usage(buffer, usage);
+    }
+}
+
+enum ggml_backend_buffer_usage ggml_backend_buffer_get_usage(ggml_backend_buffer_t buffer) {
+    return buffer->usage;
+}
+
+ggml_backend_buffer_type_t ggml_backend_buffer_get_type(ggml_backend_buffer_t buffer) {
+    return buffer->buft;
+}
+
+void ggml_backend_buffer_reset(ggml_backend_buffer_t buffer) {
+    if (buffer->iface.reset) {
+        buffer->iface.reset(buffer);
+    }
+}
+
+bool ggml_backend_buffer_copy_tensor(const struct ggml_tensor * src, struct ggml_tensor * dst) {
+    ggml_backend_buffer_t dst_buf = dst->view_src ? dst->view_src->buffer : dst->buffer;
+    if (dst_buf->iface.cpy_tensor) {
+        return dst_buf->iface.cpy_tensor(dst_buf, src, dst);
+    }
+    return false;
+}
+
+// backend
+
+ggml_guid_t ggml_backend_guid(ggml_backend_t backend) {
+    if (backend == NULL) {
+        return NULL;
+    }
+    return backend->guid;
+}
+
+const char * ggml_backend_name(ggml_backend_t backend) {
+    if (backend == NULL) {
+        return "NULL";
+    }
+    return backend->iface.get_name(backend);
+}
+
+void ggml_backend_free(ggml_backend_t backend) {
+    if (backend == NULL) {
+        return;
+    }
+
+    backend->iface.free(backend);
+}
+
+ggml_backend_buffer_type_t ggml_backend_get_default_buffer_type(ggml_backend_t backend) {
+    return backend->iface.get_default_buffer_type(backend);
+}
+
+ggml_backend_buffer_t ggml_backend_alloc_buffer(ggml_backend_t backend, size_t size) {
+    return ggml_backend_buft_alloc_buffer(ggml_backend_get_default_buffer_type(backend), size);
+}
+
+size_t ggml_backend_get_alignment(ggml_backend_t backend) {
+    return ggml_backend_buft_get_alignment(ggml_backend_get_default_buffer_type(backend));
+}
+
+size_t ggml_backend_get_max_size(ggml_backend_t backend) {
+    return ggml_backend_buft_get_max_size(ggml_backend_get_default_buffer_type(backend));
+}
+
+void ggml_backend_tensor_set_async(ggml_backend_t backend, struct ggml_tensor * tensor, const void * data, size_t offset, size_t size) {
+    GGML_ASSERT(tensor->data != NULL && "tensor not allocated");
+    GGML_ASSERT(offset + size <= ggml_nbytes(tensor) && "tensor write out of bounds");
+
+    if (backend->iface.set_tensor_async == NULL) {
+        ggml_backend_tensor_set(tensor, data, offset, size);
+    } else {
+        backend->iface.set_tensor_async(backend, tensor, data, offset, size);
+    }
+}
+
+void ggml_backend_tensor_get_async(ggml_backend_t backend, const struct ggml_tensor * tensor, void * data, size_t offset, size_t size) {
+    GGML_ASSERT(tensor->data != NULL && "tensor not allocated");
+    GGML_ASSERT(offset + size <= ggml_nbytes(tensor) && "tensor read out of bounds");
+
+    if (backend->iface.get_tensor_async == NULL) {
+        ggml_backend_tensor_get(tensor, data, offset, size);
+    } else {
+        backend->iface.get_tensor_async(backend, tensor, data, offset, size);
+    }
+}
+
+void ggml_backend_tensor_set(struct ggml_tensor * tensor, const void * data, size_t offset, size_t size) {
+    ggml_backend_buffer_t buf = tensor->view_src ? tensor->view_src->buffer : tensor->buffer;
+
+    GGML_ASSERT(buf != NULL && "tensor buffer not set");
+    GGML_ASSERT(tensor->data != NULL && "tensor not allocated");
+    GGML_ASSERT(offset + size <= ggml_nbytes(tensor) && "tensor write out of bounds");
+
+    if (!size) {
+        return;
+    }
+
+    buf->iface.set_tensor(buf, tensor, data, offset, size);
+}
+
+void ggml_backend_tensor_get(const struct ggml_tensor * tensor, void * data, size_t offset, size_t size) {
+    ggml_backend_buffer_t buf = tensor->view_src ? tensor->view_src->buffer : tensor->buffer;
+
+    GGML_ASSERT(buf != NULL && "tensor buffer not set");
+    GGML_ASSERT(tensor->data != NULL && "tensor not allocated");
+    GGML_ASSERT(offset + size <= ggml_nbytes(tensor) && "tensor read out of bounds");
+
+    if (!size) {
+        return;
+    }
+
+    buf->iface.get_tensor(buf, tensor, data, offset, size);
+}
+
+GGML_API void ggml_backend_tensor_memset(struct ggml_tensor * tensor, uint8_t value, size_t offset, size_t size) {
+    ggml_backend_buffer_t buf = tensor->view_src ? tensor->view_src->buffer : tensor->buffer;
+
+    GGML_ASSERT(buf != NULL && "tensor buffer not set");
+    GGML_ASSERT(tensor->data != NULL && "tensor not allocated");
+    GGML_ASSERT(offset + size <= ggml_nbytes(tensor) && "tensor write out of bounds");
+
+    if (!size) {
+        return;
+    }
+
+    GGML_ASSERT(buf->iface.memset_tensor != NULL && "memset not supported by backend buffer");
+
+    buf->iface.memset_tensor(buf, tensor, value, offset, size);
+}
+
+void ggml_backend_synchronize(ggml_backend_t backend) {
+    if (backend->iface.synchronize == NULL) {
+        return;
+    }
+
+    backend->iface.synchronize(backend);
+}
+
+ggml_backend_graph_plan_t ggml_backend_graph_plan_create(ggml_backend_t backend, struct ggml_cgraph * cgraph) {
+    GGML_ASSERT(backend->iface.graph_plan_create != NULL);
+
+    return backend->iface.graph_plan_create(backend, cgraph);
+}
+
+void ggml_backend_graph_plan_free(ggml_backend_t backend, ggml_backend_graph_plan_t plan) {
+    GGML_ASSERT(backend->iface.graph_plan_free != NULL);
+
+    backend->iface.graph_plan_free(backend, plan);
+}
+
+enum ggml_status ggml_backend_graph_plan_compute(ggml_backend_t backend, ggml_backend_graph_plan_t plan) {
+    GGML_ASSERT(backend->iface.graph_plan_compute != NULL);
+
+    return backend->iface.graph_plan_compute(backend, plan);
+}
+
+enum ggml_status ggml_backend_graph_compute(ggml_backend_t backend, struct ggml_cgraph * cgraph) {
+    enum ggml_status err = ggml_backend_graph_compute_async(backend, cgraph);
+    ggml_backend_synchronize(backend);
+    return err;
+}
+
+enum ggml_status ggml_backend_graph_compute_async(ggml_backend_t backend, struct ggml_cgraph * cgraph) {
+    return backend->iface.graph_compute(backend, cgraph);
+}
+
+bool ggml_backend_supports_op(ggml_backend_t backend, const struct ggml_tensor * op) {
+    // helper to ease transition to device interface
+    if (backend->device) {
+        return ggml_backend_dev_supports_op(backend->device, op);
+    }
+
+    return backend->iface.supports_op(backend, op);
+}
+
+bool ggml_backend_supports_buft(ggml_backend_t backend, ggml_backend_buffer_type_t buft) {
+    // helper to ease transition to device interface
+    if (backend->device) {
+        return ggml_backend_dev_supports_buft(backend->device, buft);
+    }
+
+    return backend->iface.supports_buft(backend, buft);
+}
+
+bool ggml_backend_offload_op(ggml_backend_t backend, const struct ggml_tensor * op) {
+    // helper to ease transition to device interface
+    if (backend->device) {
+        return ggml_backend_dev_offload_op(backend->device, op);
+    }
+
+    if (backend->iface.offload_op != NULL) {
+        return backend->iface.offload_op(backend, op);
+    }
+    return false;
+}
+
+ggml_backend_dev_t ggml_backend_get_device(ggml_backend_t backend) {
+    return backend->device;
+}
+
+// backend copy
+
+static bool ggml_are_same_layout(const struct ggml_tensor * a, const struct ggml_tensor * b) {
+    if (a->type != b->type) {
+        return false;
+    }
+    for (int i = 0; i < GGML_MAX_DIMS; i++) {
+        if (a->ne[i] != b->ne[i]) {
+            return false;
+        }
+        if (a->nb[i] != b->nb[i]) {
+            return false;
+        }
+    }
+    return true;
+}
+
+void ggml_backend_tensor_copy(struct ggml_tensor * src, struct ggml_tensor * dst) {
+    GGML_ASSERT(ggml_are_same_layout(src, dst) && "cannot copy tensors with different layouts");
+
+    if (src == dst) {
+        return;
+    }
+
+    if (ggml_backend_buffer_is_host(src->buffer)) {
+        ggml_backend_tensor_set(dst, src->data, 0, ggml_nbytes(src));
+    } else if (ggml_backend_buffer_is_host(dst->buffer)) {
+        ggml_backend_tensor_get(src, dst->data, 0, ggml_nbytes(src));
+    } else if (!ggml_backend_buffer_copy_tensor(src, dst)) {
+#ifndef NDEBUG
+        fprintf(stderr, "%s: warning: slow copy from %s to %s\n", __func__, ggml_backend_buffer_name(src->buffer), ggml_backend_buffer_name(dst->buffer));
+#endif
+        size_t nbytes = ggml_nbytes(src);
+        void * data = malloc(nbytes);
+        ggml_backend_tensor_get(src, data, 0, nbytes);
+        ggml_backend_tensor_set(dst, data, 0, nbytes);
+        free(data);
+    }
+}
+
+void ggml_backend_tensor_copy_async(ggml_backend_t backend_src, ggml_backend_t backend_dst, struct ggml_tensor * src, struct ggml_tensor * dst) {
+    GGML_ASSERT(ggml_are_same_layout(src, dst) && "cannot copy tensors with different layouts");
+
+    if (src == dst) {
+        return;
+    }
+
+    if (backend_dst->iface.cpy_tensor_async != NULL) {
+        if (backend_dst->iface.cpy_tensor_async(backend_src, backend_dst, src, dst)) {
+            return;
+        }
+    }
+
+    // an async copy would normally happen after all the queued operations on both backends are completed
+    // to simulate the same behavior, we need to synchronize both backends first, and do a blocking copy
+    ggml_backend_synchronize(backend_src);
+    ggml_backend_synchronize(backend_dst);
+    ggml_backend_tensor_copy(src, dst);
+}
+
+// events
+
+ggml_backend_event_t ggml_backend_event_new(ggml_backend_dev_t device) {
+    // null device is allowed for the transition period to the device interface
+    if (device == NULL || device->iface.event_new == NULL) {
+        return NULL;
+    }
+    return device->iface.event_new(device);
+}
+
+void ggml_backend_event_free(ggml_backend_event_t event) {
+    if (event == NULL) {
+        return;
+    }
+    event->device->iface.event_free(event->device, event);
+}
+
+void ggml_backend_event_record(ggml_backend_event_t event, ggml_backend_t backend) {
+    GGML_ASSERT(backend->iface.event_record != NULL);
+
+    backend->iface.event_record(backend, event);
+}
+
+void ggml_backend_event_synchronize(ggml_backend_event_t event) {
+    GGML_ASSERT(event->device->iface.event_synchronize);
+
+    event->device->iface.event_synchronize(event->device, event);
+}
+
+void ggml_backend_event_wait(ggml_backend_t backend, ggml_backend_event_t event) {
+    GGML_ASSERT(backend->iface.event_wait != NULL);
+
+    backend->iface.event_wait(backend, event);
+}
+
+// Backend device
+
+const char * ggml_backend_dev_name(ggml_backend_dev_t device) {
+    return device->iface.get_name(device);
+}
+
+const char * ggml_backend_dev_description(ggml_backend_dev_t device) {
+    return device->iface.get_description(device);
+}
+
+void ggml_backend_dev_memory(ggml_backend_dev_t device, size_t * free, size_t * total) {
+    device->iface.get_memory(device, free, total);
+}
+
+enum ggml_backend_dev_type ggml_backend_dev_type(ggml_backend_dev_t device) {
+    return device->iface.get_type(device);
+}
+
+void ggml_backend_dev_get_props(ggml_backend_dev_t device, struct ggml_backend_dev_props * props) {
+    device->iface.get_props(device, props);
+}
+
+ggml_backend_reg_t ggml_backend_dev_backend_reg(ggml_backend_dev_t device) {
+    return device->reg;
+}
+
+ggml_backend_t ggml_backend_dev_init(ggml_backend_dev_t device, const char * params) {
+    return device->iface.init_backend(device, params);
+}
+
+ggml_backend_buffer_type_t ggml_backend_dev_buffer_type(ggml_backend_dev_t device) {
+    return device->iface.get_buffer_type(device);
+}
+
+ggml_backend_buffer_type_t ggml_backend_dev_host_buffer_type(ggml_backend_dev_t device) {
+    return device->iface.get_host_buffer_type(device);
+}
+
+ggml_backend_buffer_t ggml_backend_dev_buffer_from_host_ptr(ggml_backend_dev_t device, void * ptr, size_t size, size_t max_tensor_size) {
+    return device->iface.buffer_from_host_ptr(device, ptr, size, max_tensor_size);
+}
+
+bool ggml_backend_dev_supports_op(ggml_backend_dev_t device, const struct ggml_tensor * op) {
+    return device->iface.supports_op(device, op);
+}
+
+bool ggml_backend_dev_supports_buft(ggml_backend_dev_t device, ggml_backend_buffer_type_t buft) {
+    return device->iface.supports_buft(device, buft);
+}
+
+bool ggml_backend_dev_offload_op(ggml_backend_dev_t device, const struct ggml_tensor * op) {
+    return device->iface.offload_op(device, op);
+}
+
+// Backend (reg)
+
+const char * ggml_backend_reg_name(ggml_backend_reg_t reg) {
+    return reg->iface.get_name(reg);
+}
+
+size_t ggml_backend_reg_dev_count(ggml_backend_reg_t reg) {
+    return reg->iface.get_device_count(reg);
+}
+
+ggml_backend_dev_t ggml_backend_reg_dev_get(ggml_backend_reg_t reg, size_t index) {
+    return reg->iface.get_device(reg, index);
+}
+
+void * ggml_backend_reg_get_proc_address(ggml_backend_reg_t reg, const char * name) {
+    if (!reg->iface.get_proc_address) {
+        return NULL;
+    }
+    return reg->iface.get_proc_address(reg, name);
+}
+
+// Backend registry
+
+#ifdef GGML_USE_CUDA
+#include "ggml-cuda.h"
+#endif
+
+struct ggml_backend_registry {
+    std::vector<ggml_backend_reg_t> backends;
+    std::vector<ggml_backend_dev_t> devices;
+
+    ggml_backend_registry() {
+#ifdef GGML_USE_CUDA
+        register_backend(ggml_backend_cuda_reg());
+#endif
+
+        register_backend(ggml_backend_cpu_reg());
+
+        // TODO: sycl, metal, vulkan, kompute, cann
+    }
+
+    void register_backend(ggml_backend_reg_t reg) {
+#ifndef NDEBUG
+        fprintf(stderr, "%s: registered backend %s (%zu devices)\n",
+            __func__, ggml_backend_reg_name(reg), ggml_backend_reg_dev_count(reg));
+#endif
+        backends.push_back(reg);
+        for (size_t i = 0; i < ggml_backend_reg_dev_count(reg); i++) {
+            register_device(ggml_backend_reg_dev_get(reg, i));
+        }
+    }
+
+    void register_device(ggml_backend_dev_t device) {
+#ifndef NDEBUG
+        fprintf(stderr, "%s: registered device %s (%s)\n", __func__, ggml_backend_dev_name(device), ggml_backend_dev_description(device));
+#endif
+        devices.push_back(device);
+    }
+};
+
+static ggml_backend_registry & get_reg() {
+    static ggml_backend_registry reg;
+    return reg;
+}
+
+// Internal API
+void ggml_backend_register(ggml_backend_reg_t reg) {
+    get_reg().register_backend(reg);
+}
+
+void ggml_backend_device_register(ggml_backend_dev_t device) {
+    get_reg().register_device(device);
+}
+
+// Backend (reg) enumeration
+size_t ggml_backend_reg_count() {
+    return get_reg().backends.size();
+}
+
+ggml_backend_reg_t ggml_backend_reg_get(size_t index) {
+    GGML_ASSERT(index < ggml_backend_reg_count());
+    return get_reg().backends[index];
+}
+
+ggml_backend_reg_t ggml_backend_reg_by_name(const char * name) {
+    for (size_t i = 0; i < ggml_backend_reg_count(); i++) {
+        ggml_backend_reg_t reg = ggml_backend_reg_get(i);
+        if (strcmp(ggml_backend_reg_name(reg), name) == 0) {
+            return reg;
+        }
+    }
+    return NULL;
+}
+
+// Device enumeration
+size_t ggml_backend_dev_count() {
+    return get_reg().devices.size();
+}
+
+ggml_backend_dev_t ggml_backend_dev_get(size_t index) {
+    GGML_ASSERT(index < ggml_backend_dev_count());
+    return get_reg().devices[index];
+}
+
+ggml_backend_dev_t ggml_backend_dev_by_name(const char * name) {
+    for (size_t i = 0; i < ggml_backend_dev_count(); i++) {
+        ggml_backend_dev_t dev = ggml_backend_dev_get(i);
+        if (strcmp(ggml_backend_dev_name(dev), name) == 0) {
+            return dev;
+        }
+    }
+    return NULL;
+}
+
+ggml_backend_dev_t ggml_backend_dev_by_type(enum ggml_backend_dev_type type) {
+    for (size_t i = 0; i < ggml_backend_dev_count(); i++) {
+        ggml_backend_dev_t dev = ggml_backend_dev_get(i);
+        if (ggml_backend_dev_type(dev) == type) {
+            return dev;
+        }
+    }
+    return NULL;
+}
+
+// Convenience functions
+ggml_backend_t ggml_backend_init_by_name(const char * name, const char * params) {
+    ggml_backend_dev_t dev = ggml_backend_dev_by_name(name);
+    if (!dev) {
+        return NULL;
+    }
+    return ggml_backend_dev_init(dev, params);
+}
+
+ggml_backend_t ggml_backend_init_by_type(enum ggml_backend_dev_type type, const char * params) {
+    ggml_backend_dev_t dev = ggml_backend_dev_by_type(type);
+    if (!dev) {
+        return NULL;
+    }
+    return ggml_backend_dev_init(dev, params);
+}
+
+ggml_backend_t ggml_backend_init_best(void) {
+    ggml_backend_dev_t dev = ggml_backend_dev_by_type(GGML_BACKEND_DEVICE_TYPE_GPU_FULL);
+    if (!dev) {
+        dev = ggml_backend_dev_by_type(GGML_BACKEND_DEVICE_TYPE_CPU_FULL);
+    }
+    if (!dev) {
+        return NULL;
+    }
+    return ggml_backend_dev_init(dev, NULL);
+}
+
+// backend CPU
+
+static const size_t TENSOR_ALIGNMENT = 32; // required for mmap as gguf only guarantees 32-byte alignment
+
+static const char * ggml_backend_cpu_buffer_get_name(ggml_backend_buffer_t buffer) {
+    return "CPU";
+
+    GGML_UNUSED(buffer);
+}
+
+static void * ggml_backend_cpu_buffer_get_base(ggml_backend_buffer_t buffer) {
+    uintptr_t data = (uintptr_t)buffer->context;
+
+    // align the buffer
+    if (data % TENSOR_ALIGNMENT != 0) {
+        data = GGML_PAD(data, TENSOR_ALIGNMENT);
+    }
+
+    return (void *)data;
+}
+
+static void ggml_backend_cpu_buffer_free_buffer(ggml_backend_buffer_t buffer) {
+    free(buffer->context);
+}
+
+static void ggml_backend_cpu_buffer_memset_tensor(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor, uint8_t value, size_t offset, size_t size) {
+    memset((char *)tensor->data + offset, value, size);
+
+    GGML_UNUSED(buffer);
+}
+
+static void ggml_backend_cpu_buffer_set_tensor(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor, const void * data, size_t offset, size_t size) {
+    memcpy((char *)tensor->data + offset, data, size);
+
+    GGML_UNUSED(buffer);
+}
+
+static void ggml_backend_cpu_buffer_get_tensor(ggml_backend_buffer_t buffer, const struct ggml_tensor * tensor, void * data, size_t offset, size_t size) {
+    memcpy(data, (const char *)tensor->data + offset, size);
+
+    GGML_UNUSED(buffer);
+}
+
+static bool ggml_backend_cpu_buffer_cpy_tensor(ggml_backend_buffer_t buffer, const struct ggml_tensor * src, struct ggml_tensor * dst) {
+    if (ggml_backend_buffer_is_host(src->buffer)) {
+        memcpy(dst->data, src->data, ggml_nbytes(src));
+        return true;
+    }
+    return false;
+
+    GGML_UNUSED(buffer);
+}
+
+static void ggml_backend_cpu_buffer_clear(ggml_backend_buffer_t buffer, uint8_t value) {
+    memset(buffer->context, value, buffer->size);
+}
+
+static const struct ggml_backend_buffer_i ggml_backend_cpu_buffer_i = {
+    /* .get_name        = */ ggml_backend_cpu_buffer_get_name,
+    /* .free_buffer     = */ ggml_backend_cpu_buffer_free_buffer,
+    /* .get_base        = */ ggml_backend_cpu_buffer_get_base,
+    /* .init_tensor     = */ NULL, // no initialization required
+    /* .memset_tensor   = */ ggml_backend_cpu_buffer_memset_tensor,
+    /* .set_tensor      = */ ggml_backend_cpu_buffer_set_tensor,
+    /* .get_tensor      = */ ggml_backend_cpu_buffer_get_tensor,
+    /* .cpy_tensor      = */ ggml_backend_cpu_buffer_cpy_tensor,
+    /* .clear           = */ ggml_backend_cpu_buffer_clear,
+    /* .reset           = */ NULL,
+};
+
+static const struct ggml_backend_buffer_i ggml_backend_cpu_buffer_from_ptr_i = {
+    /* .get_name        = */ ggml_backend_cpu_buffer_get_name,
+    /* .free_buffer     = */ NULL, // ptr is not owned by the buffer, so it does not need to be freed
+    /* .get_base        = */ ggml_backend_cpu_buffer_get_base,
+    /* .init_tensor     = */ NULL, // no initialization required
+    /* .memset_tensor   = */ ggml_backend_cpu_buffer_memset_tensor,
+    /* .set_tensor      = */ ggml_backend_cpu_buffer_set_tensor,
+    /* .get_tensor      = */ ggml_backend_cpu_buffer_get_tensor,
+    /* .cpy_tensor      = */ ggml_backend_cpu_buffer_cpy_tensor,
+    /* .clear           = */ ggml_backend_cpu_buffer_clear,
+    /* .reset           = */ NULL,
+};
+
+static const char * ggml_backend_cpu_buffer_type_get_name(ggml_backend_buffer_type_t buft) {
+    return "CPU";
+
+    GGML_UNUSED(buft);
+}
+
+static ggml_backend_buffer_t ggml_backend_cpu_buffer_type_alloc_buffer(ggml_backend_buffer_type_t buft, size_t size) {
+    size += TENSOR_ALIGNMENT;   // malloc may return an address that is not aligned
+    void * data = malloc(size); // TODO: use GGML_ALIGNED_MALLOC (move to ggml-impl.h)
+    if (data == NULL) {
+        fprintf(stderr, "%s: failed to allocate buffer of size %zu\n", __func__, size);
+        return NULL;
+    }
+
+    return ggml_backend_buffer_init(buft, ggml_backend_cpu_buffer_i, data, size);
+}
+
+static size_t ggml_backend_cpu_buffer_type_get_alignment(ggml_backend_buffer_type_t buft) {
+    return TENSOR_ALIGNMENT;
+
+    GGML_UNUSED(buft);
+}
+
+static bool ggml_backend_cpu_buffer_type_is_host(ggml_backend_buffer_type_t buft) {
+    return true;
+
+    GGML_UNUSED(buft);
+}
+
+ggml_backend_buffer_type_t ggml_backend_cpu_buffer_type(void) {
+    static struct ggml_backend_buffer_type ggml_backend_cpu_buffer_type = {
+        /* .iface   = */ {
+            /* .get_name         = */ ggml_backend_cpu_buffer_type_get_name,
+            /* .alloc_buffer     = */ ggml_backend_cpu_buffer_type_alloc_buffer,
+            /* .get_alignment    = */ ggml_backend_cpu_buffer_type_get_alignment,
+            /* .get_max_size     = */ NULL, // defaults to SIZE_MAX
+            /* .get_alloc_size   = */ NULL, // defaults to ggml_nbytes
+            /* .is_host          = */ ggml_backend_cpu_buffer_type_is_host,
+        },
+        /* .device  = */ ggml_backend_reg_dev_get(ggml_backend_cpu_reg(), 0),
+        /* .context = */ NULL,
+    };
+
+    return &ggml_backend_cpu_buffer_type;
+}
+
+#ifdef GGML_USE_CPU_HBM
+
+// buffer type HBM
+
+#include <hbwmalloc.h>
+
+static const char * ggml_backend_cpu_hbm_buffer_type_get_name(ggml_backend_buffer_type_t buft) {
+    return "CPU_HBM";
+
+    GGML_UNUSED(buft);
+}
+
+static const char * ggml_backend_cpu_hbm_buffer_get_name(ggml_backend_buffer_t buf) {
+    return "CPU_HBM";
+
+    GGML_UNUSED(buf);
+}
+
+static void ggml_backend_cpu_hbm_buffer_free_buffer(ggml_backend_buffer_t buffer) {
+    hbw_free(buffer->context);
+}
+
+static ggml_backend_buffer_t ggml_backend_cpu_hbm_buffer_type_alloc_buffer(ggml_backend_buffer_type_t buft, size_t size) {
+    //void * ptr = hbw_malloc(size);
+    void * ptr;
+    int result = hbw_posix_memalign(&ptr, ggml_backend_cpu_buffer_type_get_alignment(buft), size);
+    if (result != 0) {
+        fprintf(stderr, "failed to allocate HBM buffer of size %zu\n", size);
+        return NULL;
+    }
+
+    ggml_backend_buffer_t buffer = ggml_backend_cpu_buffer_from_ptr(ptr, size);
+    buffer->buft = buft;
+    buffer->iface.get_name = ggml_backend_cpu_hbm_buffer_get_name;
+    buffer->iface.free_buffer = ggml_backend_cpu_hbm_buffer_free_buffer;
+
+    return buffer;
+}
+
+ggml_backend_buffer_type_t ggml_backend_cpu_hbm_buffer_type(void) {
+    static struct ggml_backend_buffer_type ggml_backend_cpu_buffer_type_hbm = {
+        /* .iface    = */ {
+            /* .get_name         = */ ggml_backend_cpu_hbm_buffer_type_get_name,
+            /* .alloc_buffer     = */ ggml_backend_cpu_hbm_buffer_type_alloc_buffer,
+            /* .get_alignment    = */ ggml_backend_cpu_buffer_type_get_alignment,
+            /* .get_max_size     = */ NULL, // defaults to SIZE_MAX
+            /* .get_alloc_size   = */ NULL, // defaults to ggml_nbytes
+            /* .is_host          = */ ggml_backend_cpu_buffer_type_is_host,
+        },
+        /* .context  = */ NULL,
+    };
+
+    return &ggml_backend_cpu_buffer_type_hbm;
+}
+#endif
+
+struct ggml_backend_cpu_context {
+    int                 n_threads;
+    ggml_threadpool_t   threadpool;
+
+    uint8_t *           work_data;
+    size_t              work_size;
+
+    ggml_abort_callback abort_callback;
+    void *              abort_callback_data;
+};
+
+static const char * ggml_backend_cpu_get_name(ggml_backend_t backend) {
+    return "CPU";
+
+    GGML_UNUSED(backend);
+}
+
+static void ggml_backend_cpu_free(ggml_backend_t backend) {
+    struct ggml_backend_cpu_context * cpu_ctx = (struct ggml_backend_cpu_context *)backend->context;
+    delete[] cpu_ctx->work_data;
+    delete cpu_ctx;
+    delete backend;
+}
+
+static ggml_backend_buffer_type_t ggml_backend_cpu_get_default_buffer_type(ggml_backend_t backend) {
+    return ggml_backend_cpu_buffer_type();
+
+    GGML_UNUSED(backend);
+}
+
+struct ggml_backend_plan_cpu {
+    struct ggml_cplan cplan;
+    struct ggml_cgraph cgraph;
+};
+
+static ggml_backend_graph_plan_t ggml_backend_cpu_graph_plan_create(ggml_backend_t backend, const struct ggml_cgraph * cgraph) {
+    struct ggml_backend_cpu_context * cpu_ctx = (struct ggml_backend_cpu_context *)backend->context;
+
+    struct ggml_backend_plan_cpu * cpu_plan = new ggml_backend_plan_cpu;
+
+    cpu_plan->cplan = ggml_graph_plan(cgraph, cpu_ctx->n_threads, cpu_ctx->threadpool);
+    cpu_plan->cgraph = *cgraph; // FIXME: deep copy
+
+    if (cpu_plan->cplan.work_size > 0) {
+        cpu_plan->cplan.work_data = new uint8_t[cpu_plan->cplan.work_size];
+        if (cpu_plan->cplan.work_data == NULL) {
+            delete cpu_plan;
+            return NULL;
+        }
+    }
+
+    cpu_plan->cplan.abort_callback      = cpu_ctx->abort_callback;
+    cpu_plan->cplan.abort_callback_data = cpu_ctx->abort_callback_data;
+
+    return cpu_plan;
+}
+
+static void ggml_backend_cpu_graph_plan_free(ggml_backend_t backend, ggml_backend_graph_plan_t plan) {
+    struct ggml_backend_plan_cpu * cpu_plan = (struct ggml_backend_plan_cpu *)plan;
+
+    delete[] cpu_plan->cplan.work_data;
+    delete cpu_plan;
+
+    GGML_UNUSED(backend);
+}
+
+static enum ggml_status ggml_backend_cpu_graph_plan_compute(ggml_backend_t backend, ggml_backend_graph_plan_t plan) {
+    struct ggml_backend_plan_cpu * cpu_plan = (struct ggml_backend_plan_cpu *)plan;
+
+    return ggml_graph_compute(&cpu_plan->cgraph, &cpu_plan->cplan);
+
+    GGML_UNUSED(backend);
+}
+
+static enum ggml_status ggml_backend_cpu_graph_compute(ggml_backend_t backend, struct ggml_cgraph * cgraph) {
+    struct ggml_backend_cpu_context * cpu_ctx = (struct ggml_backend_cpu_context *)backend->context;
+
+    struct ggml_cplan cplan = ggml_graph_plan(cgraph, cpu_ctx->n_threads, cpu_ctx->threadpool);
+
+    if (cpu_ctx->work_size < cplan.work_size) {
+        delete[] cpu_ctx->work_data;
+        cpu_ctx->work_data = new uint8_t[cplan.work_size];
+        if (cpu_ctx->work_data == NULL) {
+            cpu_ctx->work_size = 0;
+            return GGML_STATUS_ALLOC_FAILED;
+        }
+        cpu_ctx->work_size = cplan.work_size;
+    }
+    cplan.work_data = (uint8_t *)cpu_ctx->work_data;
+
+    cplan.abort_callback      = cpu_ctx->abort_callback;
+    cplan.abort_callback_data = cpu_ctx->abort_callback_data;
+
+    return ggml_graph_compute(cgraph, &cplan);
+}
+
+static const struct ggml_backend_i ggml_backend_cpu_i = {
+    /* .get_name                = */ ggml_backend_cpu_get_name,
+    /* .free                    = */ ggml_backend_cpu_free,
+    /* .get_default_buffer_type = */ ggml_backend_cpu_get_default_buffer_type,
+    /* .set_tensor_async        = */ NULL,
+    /* .get_tensor_async        = */ NULL,
+    /* .cpy_tensor_async        = */ NULL,
+    /* .synchronize             = */ NULL,
+    /* .graph_plan_create       = */ ggml_backend_cpu_graph_plan_create,
+    /* .graph_plan_free         = */ ggml_backend_cpu_graph_plan_free,
+    /* .graph_plan_update       = */ NULL,
+    /* .graph_plan_compute      = */ ggml_backend_cpu_graph_plan_compute,
+    /* .graph_compute           = */ ggml_backend_cpu_graph_compute,
+    /* .supports_op             = */ NULL,
+    /* .supports_buft           = */ NULL,
+    /* .offload_op              = */ NULL,
+    /* .event_record            = */ NULL,
+    /* .event_wait              = */ NULL,
+};
+
+static ggml_guid_t ggml_backend_cpu_guid(void) {
+    static ggml_guid guid = { 0xaa, 0x67, 0xc7, 0x43, 0x96, 0xe6, 0xa3, 0x8a, 0xe3, 0xaf, 0xea, 0x92, 0x36, 0xbc, 0xfc, 0x89 };
+    return &guid;
+}
+
+ggml_backend_t ggml_backend_cpu_init(void) {
+    struct ggml_backend_cpu_context * ctx = new ggml_backend_cpu_context;
+    if (ctx == NULL) {
+        return NULL;
+    }
+
+    ctx->n_threads           = GGML_DEFAULT_N_THREADS;
+    ctx->threadpool          = NULL;
+    ctx->work_data           = NULL;
+    ctx->work_size           = 0;
+    ctx->abort_callback      = NULL;
+    ctx->abort_callback_data = NULL;
+
+    ggml_backend_t cpu_backend = new ggml_backend {
+        /* .guid      = */ ggml_backend_cpu_guid(),
+        /* .interface = */ ggml_backend_cpu_i,
+        /* .device    = */ ggml_backend_reg_dev_get(ggml_backend_cpu_reg(), 0),
+        /* .context   = */ ctx,
+    };
+
+    if (cpu_backend == NULL) {
+        delete ctx;
+        return NULL;
+    }
+
+    return cpu_backend;
+}
+
+bool ggml_backend_is_cpu(ggml_backend_t backend) {
+    return backend != NULL && ggml_guid_matches(backend->guid, ggml_backend_cpu_guid());
+}
+
+void ggml_backend_cpu_set_n_threads(ggml_backend_t backend_cpu, int n_threads) {
+    GGML_ASSERT(ggml_backend_is_cpu(backend_cpu));
+
+    struct ggml_backend_cpu_context * ctx = (struct ggml_backend_cpu_context *)backend_cpu->context;
+    ctx->n_threads = n_threads;
+}
+
+void ggml_backend_cpu_set_threadpool(ggml_backend_t backend_cpu, ggml_threadpool_t threadpool) {
+    GGML_ASSERT(ggml_backend_is_cpu(backend_cpu));
+
+    struct ggml_backend_cpu_context * ctx = (struct ggml_backend_cpu_context *)backend_cpu->context;
+
+    if (ctx->threadpool && ctx->threadpool != threadpool) {
+        // already had a different threadpool, pause/suspend it before switching
+        ggml_threadpool_pause(ctx->threadpool);
+    }
+    ctx->threadpool = threadpool;
+}
+
+void ggml_backend_cpu_set_abort_callback(ggml_backend_t backend_cpu, ggml_abort_callback abort_callback, void * abort_callback_data) {
+    GGML_ASSERT(ggml_backend_is_cpu(backend_cpu));
+
+    struct ggml_backend_cpu_context * ctx = (struct ggml_backend_cpu_context *)backend_cpu->context;
+    ctx->abort_callback = abort_callback;
+    ctx->abort_callback_data = abort_callback_data;
+}
+
+ggml_backend_buffer_t ggml_backend_cpu_buffer_from_ptr(void * ptr, size_t size) {
+    GGML_ASSERT((uintptr_t)ptr % TENSOR_ALIGNMENT == 0 && "buffer pointer must be aligned");
+    return ggml_backend_buffer_init(ggml_backend_cpu_buffer_type(), ggml_backend_cpu_buffer_from_ptr_i, ptr, size);
+}
+
+////////////////////////
+
+struct ggml_backend_cpu_device_context {
+    std::string description = "CPU";
+
+    ggml_backend_cpu_device_context() {
+#ifdef __APPLE__
+        size_t len = 0;
+        if (!sysctlbyname("machdep.cpu.brand_string", NULL, &len, NULL, 0)) {
+            description.resize(len);
+            sysctlbyname("machdep.cpu.brand_string", &description[0], &len, NULL, 0); // NOLINT
+        }
+#elif defined(__linux__)
+        FILE * f = fopen("/proc/cpuinfo", "r");
+        if (f) {
+            char buf[1024];
+            while (fgets(buf, sizeof(buf), f)) {
+                if (strncmp(buf, "model name", 10) == 0) {
+                    char * p = strchr(buf, ':');
+                    if (p) {
+                        p++;
+                        while (std::isspace(*p)) {
+                            p++;
+                        }
+                        while (std::isspace(p[strlen(p) - 1])) {
+                            p[strlen(p) - 1] = '\0';
+                        }
+                        description = p;
+                        break;
+                    }
+                }
+            }
+            fclose(f);
+        }
+#elif defined(_WIN32)
+        HKEY hKey;
+        if (RegOpenKeyEx(HKEY_LOCAL_MACHINE,
+                        TEXT("HARDWARE\\DESCRIPTION\\System\\CentralProcessor\\0"),
+                        0,
+                        KEY_READ,
+                        &hKey) == ERROR_SUCCESS) {
+            DWORD cpu_brand_size = 0;
+            if (RegQueryValueExA(hKey,
+                                TEXT("ProcessorNameString"),
+                                NULL,
+                                NULL,
+                                NULL,
+                                &cpu_brand_size) == ERROR_SUCCESS) {
+                description.resize(cpu_brand_size);
+                if (RegQueryValueExA(hKey,
+                                    TEXT("ProcessorNameString"),
+                                    NULL,
+                                    NULL,
+                                    (LPBYTE)&description[0], // NOLINT
+                                    &cpu_brand_size) == ERROR_SUCCESS) {
+                    if (description.find('\0') != std::string::npos) {
+                        description.resize(description.find('\0'));
+                    }
+                }
+            }
+            RegCloseKey(hKey);
+        }
+#endif
+    }
+};
+
+static const char * ggml_backend_cpu_device_get_name(ggml_backend_dev_t dev) {
+    return "CPU";
+
+    GGML_UNUSED(dev);
+}
+
+static const char * ggml_backend_cpu_device_get_description(ggml_backend_dev_t dev) {
+    struct ggml_backend_cpu_device_context * ctx = (struct ggml_backend_cpu_device_context *)dev->context;
+
+    return ctx->description.c_str();
+}
+
+static void ggml_backend_cpu_device_get_memory(ggml_backend_dev_t dev, size_t * free, size_t * total) {
+    // TODO
+    *free = 0;
+    *total = 0;
+
+    GGML_UNUSED(dev);
+}
+
+static enum ggml_backend_dev_type ggml_backend_cpu_device_get_type(ggml_backend_dev_t dev) {
+    return GGML_BACKEND_DEVICE_TYPE_CPU_FULL;
+
+    GGML_UNUSED(dev);
+}
+
+static void ggml_backend_cpu_device_get_props(ggml_backend_dev_t dev, struct ggml_backend_dev_props * props) {
+    props->name        = ggml_backend_cpu_device_get_name(dev);
+    props->description = ggml_backend_cpu_device_get_description(dev);
+    props->type        = ggml_backend_cpu_device_get_type(dev);
+    ggml_backend_cpu_device_get_memory(dev, &props->memory_free, &props->memory_total);
+    props->caps = {
+        /* async       */ false,
+        /* host_buffer */ false,
+        /* events      */ false,
+    };
+}
+
+static ggml_backend_t ggml_backend_cpu_device_init(ggml_backend_dev_t dev, const char * params) {
+    return ggml_backend_cpu_init();
+
+    GGML_UNUSED(dev);
+    GGML_UNUSED(params);
+}
+
+static ggml_backend_buffer_type_t ggml_backend_cpu_device_get_buffer_type(ggml_backend_dev_t dev) {
+    return ggml_backend_cpu_buffer_type();
+
+    GGML_UNUSED(dev);
+}
+
+static ggml_backend_buffer_t ggml_backend_cpu_device_buffer_from_ptr(ggml_backend_dev_t dev, void * ptr, size_t size, size_t max_tensor_size) {
+    return ggml_backend_cpu_buffer_from_ptr(ptr, size);
+
+    GGML_UNUSED(dev);
+    GGML_UNUSED(max_tensor_size);
+}
+
+static bool ggml_backend_cpu_device_supports_op(ggml_backend_dev_t dev, const struct ggml_tensor * op) {
+    switch (op->op) {
+        case GGML_OP_CPY:
+            return
+                op->type != GGML_TYPE_IQ2_XXS &&
+                op->type != GGML_TYPE_IQ2_XS  &&
+                op->type != GGML_TYPE_IQ1_S   &&
+                op->type != GGML_TYPE_IQ1_M; // missing type_traits.from_float
+        case GGML_OP_MUL_MAT:
+            return op->src[1]->type == GGML_TYPE_F32 || op->src[1]->type == ggml_internal_get_type_traits(op->src[0]->type).vec_dot_type;
+        case GGML_OP_ROPE_BACK:
+            return op->src[2] == NULL && (op->op_params[2] & 4) == 0;
+        case GGML_OP_IM2COL_BACK:
+            return op->src[0]->type == GGML_TYPE_F32 && op->src[1]->type == GGML_TYPE_F32;
+        case GGML_OP_OUT_PROD:
+            return (op->src[0]->type == GGML_TYPE_F32 || ggml_is_quantized(op->src[0]->type)) && op->src[1]->type == GGML_TYPE_F32;
+        default:
+            return true;
+    }
+
+    GGML_UNUSED(dev);
+}
+
+static bool ggml_backend_cpu_device_supports_buft(ggml_backend_dev_t dev, ggml_backend_buffer_type_t buft) {
+    return ggml_backend_buft_is_host(buft);
+
+    GGML_UNUSED(dev);
+}
+
+static const struct ggml_backend_device_i ggml_backend_cpu_device_i = {
+    /* .get_name             = */ ggml_backend_cpu_device_get_name,
+    /* .get_description      = */ ggml_backend_cpu_device_get_description,
+    /* .get_memory           = */ ggml_backend_cpu_device_get_memory,
+    /* .get_type             = */ ggml_backend_cpu_device_get_type,
+    /* .get_props            = */ ggml_backend_cpu_device_get_props,
+    /* .init_backend         = */ ggml_backend_cpu_device_init,
+    /* .get_buffer_type      = */ ggml_backend_cpu_device_get_buffer_type,
+    /* .get_host_buffer_type = */ NULL,
+    /* .buffer_from_host_ptr = */ ggml_backend_cpu_device_buffer_from_ptr,
+    /* .supports_op          = */ ggml_backend_cpu_device_supports_op,
+    /* .supports_buft        = */ ggml_backend_cpu_device_supports_buft,
+    /* .offload_op           = */ NULL,
+    /* .event_new            = */ NULL,
+    /* .event_free           = */ NULL,
+    /* .event_synchronize    = */ NULL,
+};
+
+////////////////////////
+
+static const char * ggml_backend_cpu_reg_get_name(ggml_backend_reg_t reg) {
+    return "CPU";
+
+    GGML_UNUSED(reg);
+}
+
+static size_t ggml_backend_cpu_reg_get_device_count(ggml_backend_reg_t reg) {
+    return 1;
+
+    GGML_UNUSED(reg);
+}
+
+static ggml_backend_dev_t ggml_backend_cpu_reg_get_device(ggml_backend_reg_t reg, size_t index) {
+    GGML_ASSERT(index == 0);
+
+    static ggml_backend_cpu_device_context ctx;
+    static ggml_backend_device ggml_backend_cpu_device = {
+        /* .iface   = */ ggml_backend_cpu_device_i,
+        /* .reg     = */ reg,
+        /* .context = */ &ctx,
+    };
+
+    return &ggml_backend_cpu_device;
+
+    GGML_UNUSED(reg);
+    GGML_UNUSED(index);
+}
+
+static const struct ggml_backend_reg_i ggml_backend_cpu_reg_i = {
+    /* .get_name         = */ ggml_backend_cpu_reg_get_name,
+    /* .get_device_count = */ ggml_backend_cpu_reg_get_device_count,
+    /* .get_device       = */ ggml_backend_cpu_reg_get_device,
+    /* .get_proc_address = */ NULL,
+};
+
+ggml_backend_reg_t ggml_backend_cpu_reg(void) {
+    static struct ggml_backend_reg ggml_backend_cpu_reg = {
+        /* .iface   = */ ggml_backend_cpu_reg_i,
+        /* .context = */ NULL,
+    };
+
+    return &ggml_backend_cpu_reg;
+}
+
+// multi-buffer buffer
+
+struct ggml_backend_multi_buffer_context {
+    ggml_backend_buffer_t * buffers;
+    size_t n_buffers;
+};
+
+static const char * ggml_backend_multi_buffer_get_name(ggml_backend_buffer_t buffer) {
+    ggml_backend_multi_buffer_context * ctx = (ggml_backend_multi_buffer_context *) buffer->context;
+
+    return ctx->buffers[0]->iface.get_name(ctx->buffers[0]);
+}
+
+static void ggml_backend_multi_buffer_free_buffer(ggml_backend_buffer_t buffer) {
+    ggml_backend_multi_buffer_context * ctx = (ggml_backend_multi_buffer_context *) buffer->context;
+    for (size_t i = 0; i < ctx->n_buffers; i++) {
+        ggml_backend_buffer_free(ctx->buffers[i]);
+    }
+
+    free(ctx->buffers);
+    free(ctx);
+}
+
+static void ggml_backend_multi_buffer_clear(ggml_backend_buffer_t buffer, uint8_t value) {
+    ggml_backend_multi_buffer_context * ctx = (ggml_backend_multi_buffer_context *) buffer->context;
+    for (size_t i = 0; i < ctx->n_buffers; i++) {
+        ggml_backend_buffer_clear(ctx->buffers[i], value);
+    }
+}
+
+static const struct ggml_backend_buffer_i ggml_backend_multi_buffer_i = {
+    /* .get_name        = */ ggml_backend_multi_buffer_get_name,
+    /* .free_buffer     = */ ggml_backend_multi_buffer_free_buffer,
+    /* .get_base        = */ NULL,
+    /* .init_tensor     = */ NULL,
+    /* .memset_tensor   = */ NULL,
+    /* .set_tensor      = */ NULL,
+    /* .get_tensor      = */ NULL,
+    /* .cpy_tensor      = */ NULL,
+    /* .clear           = */ ggml_backend_multi_buffer_clear,
+    /* .reset           = */ NULL,
+};
+
+ggml_backend_buffer_t ggml_backend_multi_buffer_alloc_buffer(ggml_backend_buffer_t * buffers, size_t n_buffers) {
+    ggml_backend_multi_buffer_context * ctx = (ggml_backend_multi_buffer_context *) malloc(sizeof(struct ggml_backend_multi_buffer_context));
+    ctx->n_buffers = n_buffers;
+    ctx->buffers = (ggml_backend_buffer_t *) malloc(n_buffers * sizeof(ggml_backend_buffer_t));
+
+    GGML_ASSERT(ctx->buffers != NULL);
+
+    size_t total_size = 0;
+    for (size_t i = 0; i < n_buffers; i++) {
+        ctx->buffers[i] = buffers[i];
+        total_size += ggml_backend_buffer_get_size(buffers[i]);
+    }
+
+    return ggml_backend_buffer_init(buffers[0]->buft, ggml_backend_multi_buffer_i, ctx, total_size);
+}
+
+bool ggml_backend_buffer_is_multi_buffer(ggml_backend_buffer_t buffer) {
+    return buffer->iface.get_name == ggml_backend_multi_buffer_get_name;
+}
+
+void ggml_backend_multi_buffer_set_usage(ggml_backend_buffer_t buffer, enum ggml_backend_buffer_usage usage) {
+    GGML_ASSERT(ggml_backend_buffer_is_multi_buffer(buffer));
+    ggml_backend_multi_buffer_context * ctx = (ggml_backend_multi_buffer_context *) buffer->context;
+    for (size_t i = 0; i < ctx->n_buffers; i++) {
+        ggml_backend_buffer_set_usage(ctx->buffers[i], usage);
+    }
+}
+
+// creates a copy of the tensor with the same memory layout
+static struct ggml_tensor * ggml_dup_tensor_layout(struct ggml_context * ctx, const struct ggml_tensor * tensor) {
+    struct ggml_tensor * dup = ggml_dup_tensor(ctx, tensor);
+    for (int i = 0; i < GGML_MAX_DIMS; i++) {
+        dup->nb[i] = tensor->nb[i];
+    }
+    return dup;
+}
+
+static bool ggml_is_view_op(enum ggml_op op) {
+    return op == GGML_OP_VIEW || op == GGML_OP_RESHAPE || op == GGML_OP_PERMUTE || op == GGML_OP_TRANSPOSE;
+}
+
+// scheduler
+
+#ifndef GGML_SCHED_MAX_BACKENDS
+#define GGML_SCHED_MAX_BACKENDS 16
+#endif
+
+#ifndef GGML_SCHED_MAX_SPLIT_INPUTS
+#define GGML_SCHED_MAX_SPLIT_INPUTS GGML_MAX_SRC
+#endif
+
+#ifndef GGML_SCHED_MAX_COPIES
+#define GGML_SCHED_MAX_COPIES 4
+#endif
+
+struct ggml_backend_sched_split {
+    int backend_id;
+    int i_start;
+    int i_end;
+    struct ggml_tensor * inputs[GGML_SCHED_MAX_SPLIT_INPUTS];
+    int n_inputs;
+    // graph view of this split
+    struct ggml_cgraph graph;
+};
+
+struct ggml_backend_sched {
+    bool is_reset; // true if the scheduler has been reset since the last graph split
+    bool is_alloc;
+
+    int n_backends;
+
+    ggml_backend_t backends[GGML_SCHED_MAX_BACKENDS];
+    ggml_backend_buffer_type_t bufts[GGML_SCHED_MAX_BACKENDS];
+    ggml_gallocr_t galloc;
+
+    // hash map of the nodes in the graph
+    struct ggml_hash_set  hash_set;
+    int                 * hv_tensor_backend_ids; // [hash_set.size]
+    struct ggml_tensor ** hv_tensor_copies;      // [hash_set.size][n_backends][n_copies]
+
+    int * node_backend_ids; // [graph_size]
+    int * leaf_backend_ids; // [graph_size]
+
+    int * prev_node_backend_ids; // [graph_size]
+    int * prev_leaf_backend_ids; // [graph_size]
+
+    // copy of the graph with modified inputs
+    struct ggml_cgraph graph;
+
+    // graph splits
+    struct ggml_backend_sched_split * splits;
+    int n_splits;
+    int splits_capacity;
+
+    // pipeline parallelism support
+    int n_copies;
+    int cur_copy;
+    ggml_backend_event_t events[GGML_SCHED_MAX_BACKENDS][GGML_SCHED_MAX_COPIES];
+    struct ggml_tensor * graph_inputs[GGML_SCHED_MAX_SPLIT_INPUTS];
+    int n_graph_inputs;
+
+    struct ggml_context * ctx;
+
+    ggml_backend_sched_eval_callback callback_eval;
+    void * callback_eval_user_data;
+
+    char * context_buffer;
+    size_t context_buffer_size;
+
+    bool debug;
+};
+
+#define hash_id(tensor) ggml_hash_find_or_insert(&sched->hash_set, tensor)
+#define tensor_backend_id(tensor) sched->hv_tensor_backend_ids[hash_id(tensor)]
+#define tensor_id_copy(id, backend_id, copy_id) sched->hv_tensor_copies[(id) * sched->n_backends * sched->n_copies + (backend_id) * sched->n_copies + (copy_id)]
+#define tensor_copy(tensor, backend_id, copy_id) tensor_id_copy(hash_id(tensor), backend_id, copy_id)
+
+// returns the priority of the backend, lower id is higher priority
+static int ggml_backend_sched_backend_id(ggml_backend_sched_t sched, ggml_backend_t backend) {
+    for (int i = 0; i < sched->n_backends; i++) {
+        if (sched->backends[i] == backend) {
+            return i;
+        }
+    }
+    return -1;
+}
+
+static int ggml_backend_sched_backend_from_buffer(ggml_backend_sched_t sched, const struct ggml_tensor * tensor, const struct ggml_tensor * op) {
+    ggml_backend_buffer_t buffer = tensor->buffer;
+    if (buffer == NULL) {
+        return -1;
+    }
+
+    // find highest prio backend that supports the buffer type and the op
+    for (int i = 0; i < sched->n_backends; i++) {
+        if (ggml_backend_supports_buft(sched->backends[i], buffer->buft) &&
+            ggml_backend_supports_op(sched->backends[i], op)) {
+            return i;
+        }
+    }
+
+#ifndef NDEBUG
+    fprintf(stderr, "%s: warning: no backend supports op %s with a weight with buffer type %s used in tensor %s, the weight will need to be copied\n",
+        __func__, ggml_op_desc(tensor), ggml_backend_buffer_name(buffer), tensor->name);
+#endif
+
+    return -1;
+}
+
+#if 0
+#define GGML_SCHED_MAX_SPLITS_DEBUG 4096
+static char causes[GGML_DEFAULT_GRAPH_SIZE*16 + GGML_SCHED_MAX_SPLITS_DEBUG*GGML_SCHED_MAX_SPLIT_INPUTS][128]; // debug only
+#define SET_CAUSE(node, ...) sprintf(causes[hash_id(node)], __VA_ARGS__)
+#define GET_CAUSE(node) causes[hash_id(node)]
+#else
+#define SET_CAUSE(node, ...)
+#define GET_CAUSE(node) ""
+#endif
+
+// returns the backend that should be used for the node based on the current locations
+static int ggml_backend_sched_backend_id_from_cur(ggml_backend_sched_t sched, struct ggml_tensor * tensor) {
+    // TODO: use supports_op to check if the backend supports the op
+
+    // assign pre-allocated nodes to their backend
+    int cur_backend_id = ggml_backend_sched_backend_from_buffer(sched, tensor, tensor);
+    if (cur_backend_id != -1) {
+        SET_CAUSE(tensor, "1.dst");
+        return cur_backend_id;
+    }
+
+    // view_src
+    if (tensor->view_src != NULL) {
+        cur_backend_id = ggml_backend_sched_backend_from_buffer(sched, tensor->view_src, tensor);
+        if (cur_backend_id != -1) {
+            SET_CAUSE(tensor, "1.vsrc");
+            return cur_backend_id;
+        }
+    }
+
+    if (tensor->buffer || (tensor->view_src && tensor->view_src->buffer)) {
+        // since the tensor is pre-allocated, it cannot be moved to another backend
+        GGML_ABORT("pre-allocated tensor in a backend that cannot run the operation");
+    }
+
+    // graph input
+    if (tensor->flags & GGML_TENSOR_FLAG_INPUT) {
+        cur_backend_id = sched->n_backends - 1; // last backend (assumed CPU)
+        SET_CAUSE(tensor, "1.inp");
+        return cur_backend_id;
+    }
+
+    // operations with weights are preferably run on the same backend as the weights
+    for (int i = 0; i < GGML_MAX_SRC; i++) {
+        const struct ggml_tensor * src = tensor->src[i];
+        if (src == NULL) {
+            continue;
+        }
+        if (src->buffer != NULL && src->buffer->usage == GGML_BACKEND_BUFFER_USAGE_WEIGHTS) {
+            int src_backend_id = ggml_backend_sched_backend_from_buffer(sched, src, tensor);
+            // check if a backend with higher prio wants to offload the op
+            if (src_backend_id == sched->n_backends - 1) {
+                for (int b = 0; b < src_backend_id; b++) {
+                    if (ggml_backend_supports_op(sched->backends[b], tensor) && ggml_backend_offload_op(sched->backends[b], tensor)) {
+                        SET_CAUSE(tensor, "1.off");
+                        return b;
+                    }
+                }
+            }
+            SET_CAUSE(tensor, "1.wgt%d", i);
+            return src_backend_id;
+        }
+    }
+
+    return -1;
+}
+
+static char * fmt_size(size_t size) {
+    static char buffer[128];
+    if (size >= 1024*1024) {
+        snprintf(buffer, sizeof(buffer), "%zuM", size/1024/1024);
+    } else {
+        snprintf(buffer, sizeof(buffer), "%zuK", size/1024);
+    }
+    return buffer;
+}
+
+static void ggml_backend_sched_print_assignments(ggml_backend_sched_t sched, struct ggml_cgraph * graph) {
+    int cur_split = 0;
+    for (int i = 0; i < graph->n_nodes; i++) {
+        if (cur_split < sched->n_splits && i == sched->splits[cur_split].i_start) {
+            ggml_backend_t split_backend = sched->backends[sched->splits[cur_split].backend_id];
+            fprintf(stderr, "\n## SPLIT #%d: %s # %d inputs: ", cur_split, ggml_backend_name(split_backend),
+                sched->splits[cur_split].n_inputs);
+            for (int j = 0; j < sched->splits[cur_split].n_inputs; j++) {
+                fprintf(stderr, "[%s (%5.5s)] ", sched->splits[cur_split].inputs[j]->name,
+                    fmt_size(ggml_nbytes(sched->splits[cur_split].inputs[j])));
+            }
+            fprintf(stderr, "\n");
+            cur_split++;
+        }
+        struct ggml_tensor * node = graph->nodes[i];
+        if (ggml_is_view_op(node->op)) {
+            continue;
+        }
+        ggml_backend_t tensor_backend = ggml_backend_sched_get_tensor_backend(sched, node);
+        fprintf(stderr, "node #%3d (%10.10s): %20.20s (%5.5s) [%5.5s %8.8s]:", i, ggml_op_name(node->op), node->name,
+            fmt_size(ggml_nbytes(node)), tensor_backend ? ggml_backend_name(tensor_backend) : "NULL", GET_CAUSE(node));
+        for (int j = 0; j < GGML_MAX_SRC; j++) {
+            struct ggml_tensor * src = node->src[j];
+            if (src == NULL) {
+                continue;
+            }
+            ggml_backend_t src_backend = ggml_backend_sched_get_tensor_backend(sched, src);
+            fprintf(stderr, " %20.20s (%5.5s) [%5.5s %8.8s]", src->name,
+                fmt_size(ggml_nbytes(src)), src_backend ? ggml_backend_name(src_backend) : "NULL", GET_CAUSE(src));
+        }
+        fprintf(stderr, "\n");
+    }
+}
+
+static bool ggml_backend_sched_buffer_supported(ggml_backend_sched_t sched, struct ggml_tensor * t, int backend_id) {
+    ggml_backend_buffer_t buf = t->view_src ? t->view_src->buffer : t->buffer;
+    ggml_backend_buffer_type_t buft = NULL;
+
+    if (buf) {
+        // the tensor is already allocated
+        buft = buf->buft;
+    } else {
+        // see if the tensor already has a backend assigned, and use the buffer type of that backend
+        int tensor_backend_id = tensor_backend_id(t);
+        if (tensor_backend_id == -1 && t->view_src) {
+            tensor_backend_id = tensor_backend_id(t->view_src);
+        }
+        if (tensor_backend_id != -1) {
+            buft = sched->bufts[tensor_backend_id];
+        }
+    }
+
+    return buft != NULL && ggml_backend_supports_buft(sched->backends[backend_id], buft);
+}
+
+static void ggml_backend_sched_set_if_supported(ggml_backend_sched_t sched, struct ggml_tensor * node, int cur_backend_id, int * node_backend_id) {
+    if (ggml_backend_supports_op(sched->backends[cur_backend_id], node)) {
+        *node_backend_id = cur_backend_id;
+        SET_CAUSE(node, "2.sup");
+    }
+}
+
+// assigns backends to ops and splits the graph into subgraphs that can be computed on the same backend
+static void ggml_backend_sched_split_graph(ggml_backend_sched_t sched, struct ggml_cgraph * graph) {
+    // reset splits
+    sched->n_splits = 0;
+    sched->n_graph_inputs = 0;
+    sched->is_reset = false;
+
+    struct ggml_init_params params = {
+        /* .mem_size =   */ sched->context_buffer_size,
+        /* .mem_buffer = */ sched->context_buffer,
+        /* .no_alloc =   */ true
+    };
+
+    ggml_free(sched->ctx);
+
+    sched->ctx = ggml_init(params);
+    if (sched->ctx == NULL) {
+        GGML_ABORT("%s: failed to initialize context\n", __func__);
+    }
+
+    // pass 1: assign backends to ops with pre-allocated inputs
+    for (int i = 0; i < graph->n_leafs; i++) {
+        struct ggml_tensor * leaf = graph->leafs[i];
+        int * leaf_backend_id = &tensor_backend_id(leaf);
+        // do not overwrite user assignments
+        if (*leaf_backend_id == -1) {
+            *leaf_backend_id = ggml_backend_sched_backend_id_from_cur(sched, leaf);
+        }
+    }
+
+    for (int i = 0; i < graph->n_nodes; i++) {
+        struct ggml_tensor * node = graph->nodes[i];
+        int * node_backend_id = &tensor_backend_id(node);
+        // do not overwrite user assignments
+        if (*node_backend_id == -1) {
+            *node_backend_id = ggml_backend_sched_backend_id_from_cur(sched, node);
+
+#if 0
+            // src
+            if (node->op == GGML_OP_NONE) {
+                continue;
+            }
+
+            for (int j = 0; j < GGML_MAX_SRC; j++) {
+                struct ggml_tensor * src = node->src[j];
+                if (src == NULL) {
+                    continue;
+                }
+                int * src_backend_id = &tensor_backend_id(src);
+                if (*src_backend_id == -1) {
+                    *src_backend_id = ggml_backend_sched_backend_id_from_cur(sched, src);
+                }
+            }
+#endif
+        }
+    }
+
+    // pass 2: expand current backend assignments
+    // assign the same backend to adjacent nodes
+    // expand gpu backends (i.e. non last prio) up and down, ignoring cpu (the lowest priority backend)
+    // thus, cpu will never be used unless weights are on cpu, or there are no gpu ops between cpu ops
+    // ops unsupported by the backend being expanded will be left unassigned so that they can be assigned later when the locations of its inputs are known
+    // expand gpu down
+    {
+        int cur_backend_id = -1;
+        for (int i = 0; i < graph->n_nodes; i++) {
+            struct ggml_tensor * node = graph->nodes[i];
+            if (ggml_is_view_op(node->op)) {
+                continue;
+            }
+            int * node_backend_id = &tensor_backend_id(node);
+            if (*node_backend_id != -1) {
+                if (*node_backend_id == sched->n_backends - 1) {
+                    // skip cpu (lowest prio backend)
+                    cur_backend_id = -1;
+                } else {
+                    cur_backend_id = *node_backend_id;
+                }
+            } else if (cur_backend_id != -1) {
+                ggml_backend_sched_set_if_supported(sched, node, cur_backend_id, node_backend_id);
+            }
+        }
+    }
+    // expand gpu up
+    {
+        int cur_backend_id = -1;
+        for (int i = graph->n_nodes - 1; i >= 0; i--) {
+            struct ggml_tensor * node = graph->nodes[i];
+            if (ggml_is_view_op(node->op)) {
+                continue;
+            }
+            int * node_backend_id = &tensor_backend_id(node);
+            if (*node_backend_id != -1) {
+                if (*node_backend_id == sched->n_backends - 1) {
+                    // skip cpu (lowest prio backend)
+                    cur_backend_id = -1;
+                } else {
+                    cur_backend_id = *node_backend_id;
+                }
+            } else if (cur_backend_id != -1) {
+                ggml_backend_sched_set_if_supported(sched, node, cur_backend_id, node_backend_id);
+            }
+        }
+    }
+    // expand rest down
+    {
+        int cur_backend_id = -1;
+        for (int i = 0; i < graph->n_nodes; i++) {
+            struct ggml_tensor * node = graph->nodes[i];
+            if (ggml_is_view_op(node->op)) {
+                continue;
+            }
+            int * node_backend_id = &tensor_backend_id(node);
+            if (*node_backend_id != -1) {
+                cur_backend_id = *node_backend_id;
+            } else if (cur_backend_id != -1) {
+                ggml_backend_sched_set_if_supported(sched, node, cur_backend_id, node_backend_id);
+            }
+        }
+    }
+    // expand rest up
+    {
+        int cur_backend_id = -1;
+        for (int i = graph->n_nodes - 1; i >= 0; i--) {
+            struct ggml_tensor * node = graph->nodes[i];
+            if (ggml_is_view_op(node->op)) {
+                continue;
+            }
+            int * node_backend_id = &tensor_backend_id(node);
+            if (*node_backend_id != -1) {
+                cur_backend_id = *node_backend_id;
+            } else if (cur_backend_id != -1) {
+                ggml_backend_sched_set_if_supported(sched, node, cur_backend_id, node_backend_id);
+            }
+        }
+    }
+
+    // pass 3: upgrade nodes to higher prio backends with compatible buffer types
+    // if the tensor is already in the same buffer type (*) as another higher priority backend, we should move it there
+    // however, we also need to verify that the sources are in compatible buffer types
+    // (*) the actual requirement is more relaxed, the buffer type of the backend should be supported by all the users of this tensor further down the graph
+    // however, this is slow to verify, so we have a more strict requirement that the buffer type is the same
+    // this is not uncommon since multiple backends can use host memory, with the same buffer type (eg. BLAS and CPU)
+    // additionally, set remaining unassigned nodes to the backend with the most supported inputs
+    // only nodes that could not be assigned during expansion due to the backend not supporting the op should be unassigned at this point
+    for (int i = 0; i < graph->n_nodes; i++) {
+        struct ggml_tensor * node = graph->nodes[i];
+        if (ggml_is_view_op(node->op)) {
+            continue;
+        }
+        int * node_backend_id = &tensor_backend_id(node);
+        if (*node_backend_id == -1) {
+            // unassigned node: find the backend with the most supported inputs
+            int n_supported_best = -1;
+            for (int b = 0; b < sched->n_backends; b++) {
+                if (ggml_backend_supports_op(sched->backends[b], node)) {
+                    int n_supported = 0;
+                    for (int j = 0; j < GGML_MAX_SRC; j++) {
+                        struct ggml_tensor * src = node->src[j];
+                        if (src == NULL) {
+                            continue;
+                        }
+                        if ((tensor_backend_id(src) != -1 || tensor_backend_id(src->view_src) != -1) && ggml_backend_sched_buffer_supported(sched, src, b)) {
+                            n_supported++;
+                        }
+                    }
+                    if (n_supported > n_supported_best) {
+                        n_supported_best = n_supported;
+                        *node_backend_id = b;
+                        SET_CAUSE(node, "3.best");
+                    }
+                }
+            }
+        } else {
+            // assigned node: upgrade to higher prio backend if possible
+            for (int b = 0; b < *node_backend_id; b++) {
+                if (sched->bufts[b] == sched->bufts[*node_backend_id] && ggml_backend_supports_op(sched->backends[b], node)) {
+                    bool supported = true;
+                    for (int j = 0; j < GGML_MAX_SRC; j++) {
+                        struct ggml_tensor * src = node->src[j];
+                        if (src == NULL) {
+                            continue;
+                        }
+                        if (!ggml_backend_sched_buffer_supported(sched, src, b)) {
+                            supported = false;
+                            break;
+                        }
+                    }
+                    if (supported) {
+                        *node_backend_id = b;
+                        SET_CAUSE(node, "3.upg");
+                        break;
+                    }
+                }
+            }
+        }
+    }
+
+    // pass 4: assign backends to remaining src from dst and view_src
+    for (int i = 0; i < graph->n_nodes; i++) {
+        struct ggml_tensor * node = graph->nodes[i];
+        int * cur_backend_id = &tensor_backend_id(node);
+        if (node->view_src != NULL && *cur_backend_id == -1) {
+            *cur_backend_id = tensor_backend_id(node->view_src);
+            SET_CAUSE(node, "4.vsrc");
+        }
+        for (int j = 0; j < GGML_MAX_SRC; j++) {
+            struct ggml_tensor * src = node->src[j];
+            if (src == NULL) {
+                continue;
+            }
+            int * src_backend_id = &tensor_backend_id(src);
+            if (*src_backend_id == -1) {
+                if (src->view_src != NULL) {
+                    // views are always on the same backend as the source
+                    *src_backend_id = tensor_backend_id(src->view_src);
+                    SET_CAUSE(src, "4.vsrc");
+                } else {
+                    *src_backend_id = *cur_backend_id;
+                    SET_CAUSE(src, "4.cur");
+                }
+            }
+        }
+    }
+
+    // pass 5: split graph, find tensors that need to be copied
+    {
+        int i_split = 0;
+        struct ggml_backend_sched_split * split = &sched->splits[0];
+        // find the backend of the first split, skipping view ops
+        int i = 0;
+        for (; i < graph->n_nodes; i++) {
+            struct ggml_tensor * node = graph->nodes[i];
+            if (!ggml_is_view_op(node->op)) {
+                split->backend_id = tensor_backend_id(node);
+                break;
+            }
+        }
+        split->i_start = 0;
+        split->n_inputs = 0;
+        int cur_backend_id = split->backend_id;
+        for (; i < graph->n_nodes; i++) {
+            struct ggml_tensor * node = graph->nodes[i];
+
+            if (ggml_is_view_op(node->op)) {
+                continue;
+            }
+
+            const int node_backend_id = tensor_backend_id(node);
+
+            assert(node_backend_id != -1); // all nodes should be assigned by now
+
+            // check if we should start a new split based on the sources of the current node
+            bool need_new_split = false;
+            if (node_backend_id == cur_backend_id && split->n_inputs > 0) {
+                for (int j = 0; j < GGML_MAX_SRC; j++) {
+                    struct ggml_tensor * src = node->src[j];
+                    if (src == NULL) {
+                        continue;
+                    }
+                    // check if a weight is on a different backend
+                    // by starting a new split, the memory of the previously offloaded weights can be reused
+                    if (src->buffer != NULL && src->buffer->usage == GGML_BACKEND_BUFFER_USAGE_WEIGHTS) {
+                        int src_backend_id = tensor_backend_id(src);
+                        if (src_backend_id != cur_backend_id) {
+                            need_new_split = true;
+                            break;
+                        }
+                    }
+                    // check if the split has too many inputs
+                    // FIXME: count the number of inputs instead of only checking when full
+                    if (split->n_inputs == GGML_SCHED_MAX_SPLIT_INPUTS) {
+                        const size_t id = hash_id(src);
+                        int src_backend_id = sched->hv_tensor_backend_ids[id];
+                        bool supported = ggml_backend_sched_buffer_supported(sched, src, cur_backend_id);
+                        if (src_backend_id != cur_backend_id && tensor_id_copy(id, cur_backend_id, 0) == NULL && !supported) {
+                            //printf("starting new split because of too many inputs: node %s, input %s\n", node->name, src->name);
+                            need_new_split = true;
+                            break;
+                        }
+                    }
+                }
+            }
+
+            if (node_backend_id != cur_backend_id || need_new_split) {
+                split->i_end = i;
+                i_split++;
+                if (i_split >= sched->splits_capacity) {
+                    sched->splits_capacity *= 2;
+                    sched->splits = (ggml_backend_sched_split *)
+                        realloc(sched->splits, sched->splits_capacity * sizeof(struct ggml_backend_sched_split));
+                    GGML_ASSERT(sched->splits != NULL);
+                }
+                split = &sched->splits[i_split];
+                split->backend_id = node_backend_id;
+                split->i_start = i;
+                split->n_inputs = 0;
+                cur_backend_id = node_backend_id;
+            }
+
+            // find inputs that are not on the same backend
+            for (int j = 0; j < GGML_MAX_SRC; j++) {
+                struct ggml_tensor * src = node->src[j];
+                if (src == NULL) {
+                    continue;
+                }
+
+                size_t src_id = hash_id(src);
+                const int src_backend_id = sched->hv_tensor_backend_ids[src_id];
+                assert(src_backend_id != -1); // all inputs should be assigned by now
+
+                if (src->flags & GGML_TENSOR_FLAG_INPUT && sched->n_copies > 1) {
+                    if (tensor_id_copy(src_id, src_backend_id, 0) == NULL) {
+                        ggml_backend_t backend = sched->backends[src_backend_id];
+                        for (int c = 0; c < sched->n_copies; c++) {
+                            struct ggml_tensor * tensor_copy;
+                            if (c == sched->cur_copy) {
+                                tensor_copy = src; // use the original tensor as the current copy
+                            } else {
+                                tensor_copy = ggml_dup_tensor_layout(sched->ctx, src);
+                                ggml_format_name(tensor_copy, "%s#%s#%d", ggml_backend_name(backend), src->name, c);
+                            }
+                            if (sched->n_copies > 1) {
+                                ggml_set_input(tensor_copy);
+                                ggml_set_output(tensor_copy); // prevent ggml-alloc from overwriting the tensor
+                            }
+                            tensor_id_copy(src_id, src_backend_id, c) = tensor_copy;
+                            SET_CAUSE(tensor_copy, "4.cpy");
+                        }
+                        int n_graph_inputs = sched->n_graph_inputs++;
+                        GGML_ASSERT(n_graph_inputs < GGML_SCHED_MAX_SPLIT_INPUTS);
+                        sched->graph_inputs[n_graph_inputs] = src;
+                    }
+                }
+
+                if (src_backend_id != cur_backend_id && !ggml_backend_sched_buffer_supported(sched, src, cur_backend_id)) {
+                    // create a copy of the input in the split's backend
+                    if (tensor_id_copy(src_id, cur_backend_id, 0) == NULL) {
+                        ggml_backend_t backend = sched->backends[cur_backend_id];
+                        for (int c = 0; c < sched->n_copies; c++) {
+                            struct ggml_tensor * tensor_copy = ggml_dup_tensor_layout(sched->ctx, src);
+                            ggml_format_name(tensor_copy, "%s#%s#%d", ggml_backend_name(backend), src->name, c);
+                            if (sched->n_copies > 1) {
+                                ggml_set_input(tensor_copy);
+                                ggml_set_output(tensor_copy); // prevent ggml-alloc from overwriting the tensor
+                            }
+                            tensor_id_copy(src_id, cur_backend_id, c) = tensor_copy;
+                            SET_CAUSE(tensor_copy, "4.cpy");
+                        }
+                        int n_inputs = split->n_inputs++;
+                        GGML_ASSERT(n_inputs < GGML_SCHED_MAX_SPLIT_INPUTS);
+                        split->inputs[n_inputs] = src;
+                    }
+                    node->src[j] = tensor_id_copy(src_id, cur_backend_id, sched->cur_copy);
+                }
+            }
+        }
+        split->i_end = graph->n_nodes;
+        sched->n_splits = i_split + 1;
+    }
+
+    if (sched->debug) {
+        ggml_backend_sched_print_assignments(sched, graph);
+    }
+
+    // swap node_backend_ids and leaf _backend_ids with prevs
+    {
+        int * tmp = sched->node_backend_ids;
+        sched->node_backend_ids = sched->prev_node_backend_ids;
+        sched->prev_node_backend_ids = tmp;
+
+        tmp = sched->leaf_backend_ids;
+        sched->leaf_backend_ids = sched->prev_leaf_backend_ids;
+        sched->prev_leaf_backend_ids = tmp;
+    }
+
+    int graph_size = std::max(graph->n_nodes, graph->n_leafs) + sched->n_splits*GGML_SCHED_MAX_SPLIT_INPUTS*2*sched->n_copies;
+    if (sched->graph.size < graph_size) {
+        sched->graph.size = graph_size;
+        sched->graph.nodes = (ggml_tensor **) realloc(sched->graph.nodes, graph_size * sizeof(struct ggml_tensor *));
+        sched->graph.leafs = (ggml_tensor **) realloc(sched->graph.leafs, graph_size * sizeof(struct ggml_tensor *));
+        GGML_ASSERT(sched->graph.nodes != NULL);
+        GGML_ASSERT(sched->graph.leafs != NULL);
+    }
+    sched->graph.n_nodes = 0;
+    sched->graph.n_leafs = 0;
+
+    struct ggml_cgraph * graph_copy = &sched->graph;
+
+    for (int i = 0; i < sched->n_splits; i++) {
+        struct ggml_backend_sched_split * split = &sched->splits[i];
+        split->graph = ggml_graph_view(graph, split->i_start, split->i_end);
+
+        // add inputs to the graph copy so that they are allocated by ggml-alloc at the start of the split
+        for (int j = 0; j < split->n_inputs; j++) {
+            assert(graph_copy->size > (graph_copy->n_nodes + 1));
+
+            struct ggml_tensor * input = split->inputs[j];
+            const size_t input_id = hash_id(input);
+            struct ggml_tensor * input_cpy = tensor_id_copy(input_id, split->backend_id, sched->cur_copy);
+
+            // add a dependency to the input source so that it is not freed before the copy is done
+            struct ggml_tensor * input_dep = ggml_view_tensor(sched->ctx, input);
+            input_dep->src[0] = input;
+            sched->node_backend_ids[graph_copy->n_nodes] = sched->hv_tensor_backend_ids[input_id];
+            graph_copy->nodes[graph_copy->n_nodes++] = input_dep;
+
+            // add a dependency to the input copy so that it is allocated at the start of the split
+            sched->node_backend_ids[graph_copy->n_nodes] = split->backend_id;
+            graph_copy->nodes[graph_copy->n_nodes++] = input_cpy;
+        }
+
+        for (int j = split->i_start; j < split->i_end; j++) {
+            assert(graph_copy->size > graph_copy->n_nodes);
+            sched->node_backend_ids[graph_copy->n_nodes] = tensor_backend_id(graph->nodes[j]);
+            graph_copy->nodes[graph_copy->n_nodes++] = graph->nodes[j];
+        }
+    }
+
+    if (sched->n_copies > 1) {
+        // add input copies as leafs so that they are allocated first
+        for (int i = 0; i < sched->n_graph_inputs; i++) {
+            struct ggml_tensor * input = sched->graph_inputs[i];
+            size_t id = hash_id(input);
+            int backend_id = tensor_backend_id(input);
+            for (int c = 0; c < sched->n_copies; c++) {
+                struct ggml_tensor * input_cpy = tensor_id_copy(id, backend_id, c);
+                sched->leaf_backend_ids[graph_copy->n_leafs] = backend_id;
+                assert(graph_copy->size > graph_copy->n_leafs);
+                graph_copy->leafs[graph_copy->n_leafs++] = input_cpy;
+            }
+        }
+
+        for (int i = 0; i < sched->n_splits; i++) {
+            struct ggml_backend_sched_split * split = &sched->splits[i];
+            int backend_id = split->backend_id;
+            for (int j = 0; j < split->n_inputs; j++) {
+                struct ggml_tensor * input = split->inputs[j];
+                size_t id = hash_id(input);
+                for (int c = 0; c < sched->n_copies; c++) {
+                    struct ggml_tensor * input_cpy = tensor_id_copy(id, backend_id, c);
+                    sched->leaf_backend_ids[graph_copy->n_leafs] = backend_id;
+                    assert(graph_copy->size > graph_copy->n_leafs);
+                    graph_copy->leafs[graph_copy->n_leafs++] = input_cpy;
+                }
+            }
+        }
+    }
+
+    // add leafs from the original graph
+    for (int i = 0; i < graph->n_leafs; i++) {
+        struct ggml_tensor * leaf = graph->leafs[i];
+        sched->leaf_backend_ids[graph_copy->n_leafs] = tensor_backend_id(leaf);
+        assert(graph_copy->size > graph_copy->n_leafs);
+        graph_copy->leafs[graph_copy->n_leafs++] = leaf;
+    }
+}
+
+static bool ggml_backend_sched_alloc_splits(ggml_backend_sched_t sched) {
+    bool backend_ids_changed = false;
+    for (int i = 0; i < sched->graph.n_nodes; i++) {
+        if (sched->node_backend_ids[i] != sched->prev_node_backend_ids[i] &&
+            sched->bufts[sched->node_backend_ids[i]] != sched->bufts[sched->prev_node_backend_ids[i]]) {
+            backend_ids_changed = true;
+            break;
+        }
+    }
+    if (!backend_ids_changed) {
+        for (int i = 0; i < sched->graph.n_leafs; i++) {
+            if (sched->leaf_backend_ids[i] != sched->prev_leaf_backend_ids[i] &&
+                sched->bufts[sched->leaf_backend_ids[i]] != sched->bufts[sched->prev_leaf_backend_ids[i]]) {
+                backend_ids_changed = true;
+                break;
+            }
+        }
+    }
+
+    // allocate graph
+    if (backend_ids_changed || !ggml_gallocr_alloc_graph(sched->galloc, &sched->graph)) {
+        // the re-allocation may cause the split inputs to be moved to a different address
+        ggml_backend_sched_synchronize(sched);
+#ifndef NDEBUG
+        fprintf(stderr, "%s: failed to allocate graph, reserving (backend_ids_changed = %d)\n", __func__, backend_ids_changed);
+#endif
+        ggml_gallocr_reserve_n(sched->galloc, &sched->graph, sched->node_backend_ids, sched->leaf_backend_ids);
+        if (!ggml_gallocr_alloc_graph(sched->galloc, &sched->graph)) {
+            fprintf(stderr, "%s: failed to allocate graph\n", __func__);
+            return false;
+        }
+    }
+
+    return true;
+}
+
+static enum ggml_status ggml_backend_sched_compute_splits(ggml_backend_sched_t sched) {
+    struct ggml_backend_sched_split * splits = sched->splits;
+
+    for (int i = 0; i < sched->n_splits; i++) {
+        struct ggml_backend_sched_split * split = &splits[i];
+        int split_backend_id = split->backend_id;
+        ggml_backend_t split_backend = sched->backends[split_backend_id];
+
+        // copy the input tensors to the split backend
+        for (int j = 0; j < split->n_inputs; j++) {
+            ggml_backend_t input_backend = ggml_backend_sched_get_tensor_backend(sched, split->inputs[j]);
+            struct ggml_tensor * input = split->inputs[j];
+            struct ggml_tensor * input_cpy = tensor_copy(input, split_backend_id, sched->cur_copy);
+
+            if (input->flags & GGML_TENSOR_FLAG_INPUT) {
+                // inputs from the user must be copied immediately to prevent the user overwriting the data before the copy is done
+                if (sched->events[split_backend_id][sched->cur_copy] != NULL) {
+                    ggml_backend_event_synchronize(sched->events[split_backend_id][sched->cur_copy]);
+                } else {
+                    ggml_backend_synchronize(split_backend);
+                }
+                ggml_backend_tensor_copy(input, input_cpy);
+            } else {
+                // wait for the split backend to finish using the input before overwriting it
+                if (sched->events[split_backend_id][sched->cur_copy] != NULL) {
+                    ggml_backend_event_wait(split_backend, sched->events[split_backend_id][sched->cur_copy]);
+                } else {
+                    ggml_backend_synchronize(split_backend);
+                }
+                // try async copy, but if not possible, we can still use a sync copy without synchronizing the dst backend, since we handle the synchronization here with multiple copies and events
+                // TODO: add public function to facilitate this, since applications do not have direct access to the backend interface
+                if (!split_backend->iface.cpy_tensor_async || !split_backend->iface.cpy_tensor_async(input_backend, split_backend, input, input_cpy)) {
+                    ggml_backend_synchronize(input_backend);
+                    if (sched->events[split_backend_id][sched->cur_copy] != NULL) {
+                        ggml_backend_event_synchronize(sched->events[split_backend_id][sched->cur_copy]);
+                    } else {
+                        ggml_backend_synchronize(split_backend);
+                    }
+                    ggml_backend_tensor_copy(input, input_cpy);
+                }
+            }
+        }
+
+        if (!sched->callback_eval) {
+            enum ggml_status ec = ggml_backend_graph_compute_async(split_backend, &split->graph);
+            if (ec != GGML_STATUS_SUCCESS) {
+                return ec;
+            }
+        } else {
+            // similar to ggml_backend_compare_graph_backend
+            for (int j0 = 0; j0 < split->graph.n_nodes; j0++) {
+                struct ggml_tensor * t = split->graph.nodes[j0];
+
+                // check if the user needs data from this node
+                bool need = sched->callback_eval(t, true, sched->callback_eval_user_data);
+
+                int j1 = j0;
+
+                // determine the range [j0, j1] of nodes that can be computed together
+                while (!need && j1 < split->graph.n_nodes - 1) {
+                    t = split->graph.nodes[++j1];
+                    need = sched->callback_eval(t, true, sched->callback_eval_user_data);
+                }
+
+                struct ggml_cgraph gv = ggml_graph_view(&split->graph, j0, j1 + 1);
+
+                enum ggml_status ec = ggml_backend_graph_compute_async(split_backend, &gv);
+                if (ec != GGML_STATUS_SUCCESS) {
+                    return ec;
+                }
+
+                // TODO: pass backend to the callback, then the user can decide if they want to synchronize
+                ggml_backend_synchronize(split_backend);
+
+                if (need && !sched->callback_eval(t, false, sched->callback_eval_user_data)) {
+                    break;
+                }
+
+                j0 = j1;
+            }
+        }
+
+        // record the event of this copy
+        if (split->n_inputs > 0) {
+            if (sched->events[split_backend_id][sched->cur_copy] != NULL) {
+                ggml_backend_event_record(sched->events[split_backend_id][sched->cur_copy], split_backend);
+            }
+        }
+    }
+
+    sched->cur_copy = (sched->cur_copy + 1) % sched->n_copies;
+
+    return GGML_STATUS_SUCCESS;
+}
+
+ggml_backend_sched_t ggml_backend_sched_new(
+        ggml_backend_t * backends,
+        ggml_backend_buffer_type_t * bufts,
+        int n_backends,
+        size_t graph_size,
+        bool parallel) {
+    GGML_ASSERT(n_backends > 0);
+    GGML_ASSERT(n_backends <= GGML_SCHED_MAX_BACKENDS);
+    GGML_ASSERT(ggml_backend_is_cpu(backends[n_backends - 1])); // last backend must be CPU
+
+    struct ggml_backend_sched * sched = (ggml_backend_sched *) calloc(1, sizeof(struct ggml_backend_sched));
+
+    sched->debug = getenv("GGML_SCHED_DEBUG") != NULL;
+    sched->n_backends = n_backends;
+    sched->n_copies = parallel ? GGML_SCHED_MAX_COPIES : 1;
+
+    // initialize hash table
+    // FIXME: needs to be size*2 to account for leafs (do it in graph_split instead)
+    sched->hash_set    = ggml_hash_set_new(graph_size);
+    sched->hv_tensor_backend_ids = (int *) malloc(sched->hash_set.size * sizeof(sched->hv_tensor_backend_ids[0]));
+    sched->hv_tensor_copies      = (ggml_tensor **) malloc(sched->hash_set.size * sched->n_backends * sched->n_copies * sizeof(struct ggml_tensor *));
+
+    const size_t ggml_sched_max_splits = graph_size; // at most there is one split for each node in the graph
+    const size_t nodes_size = graph_size + ggml_sched_max_splits*GGML_SCHED_MAX_SPLIT_INPUTS*2;
+    sched->node_backend_ids = (int *) calloc(nodes_size, sizeof(sched->node_backend_ids[0]));
+    sched->leaf_backend_ids = (int *) calloc(nodes_size, sizeof(sched->leaf_backend_ids[0]));
+    sched->prev_node_backend_ids = (int *) calloc(nodes_size, sizeof(sched->prev_node_backend_ids[0]));
+    sched->prev_leaf_backend_ids = (int *) calloc(nodes_size, sizeof(sched->prev_leaf_backend_ids[0]));
+
+    sched->context_buffer_size = ggml_sched_max_splits*GGML_SCHED_MAX_SPLIT_INPUTS*2*sizeof(struct ggml_tensor) + ggml_graph_overhead_custom(graph_size, false);
+    sched->context_buffer = (char *) malloc(sched->context_buffer_size);
+
+    const int initial_splits_capacity = 16;
+    sched->splits = (ggml_backend_sched_split *) calloc(initial_splits_capacity, sizeof(sched->splits[0]));
+    sched->splits_capacity = initial_splits_capacity;
+
+    for (int b = 0; b < n_backends; b++) {
+        sched->backends[b] = backends[b];
+        sched->bufts[b] = bufts ? bufts[b] : ggml_backend_get_default_buffer_type(backends[b]);
+        GGML_ASSERT(ggml_backend_supports_buft(backends[b], sched->bufts[b]));
+        if (sched->n_copies > 1) {
+            for (int c = 0; c < sched->n_copies; c++) {
+                sched->events[b][c] = ggml_backend_event_new(backends[b]->device);
+            }
+        }
+    }
+
+    sched->galloc = ggml_gallocr_new_n(sched->bufts, n_backends);
+
+    ggml_backend_sched_reset(sched);
+
+    return sched;
+}
+
+void ggml_backend_sched_free(ggml_backend_sched_t sched) {
+    if (sched == NULL) {
+        return;
+    }
+    for (int b = 0; b < sched->n_backends; b++) {
+        for (int c = 0; c < sched->n_copies; c++) {
+            ggml_backend_event_free(sched->events[b][c]);
+        }
+    }
+    ggml_gallocr_free(sched->galloc);
+    ggml_free(sched->ctx);
+    ggml_hash_set_free(&sched->hash_set);
+    free(sched->splits);
+    free(sched->hv_tensor_backend_ids);
+    free(sched->hv_tensor_copies);
+    free(sched->node_backend_ids);
+    free(sched->leaf_backend_ids);
+    free(sched->prev_node_backend_ids);
+    free(sched->prev_leaf_backend_ids);
+    free(sched->context_buffer);
+    free(sched->graph.nodes);
+    free(sched->graph.leafs);
+    free(sched);
+}
+
+void ggml_backend_sched_reset(ggml_backend_sched_t sched) {
+    // reset state for the next run
+    if (!sched->is_reset) {
+        ggml_hash_set_reset(&sched->hash_set);
+        memset(sched->hv_tensor_backend_ids, -1, sched->hash_set.size * sizeof(sched->hv_tensor_backend_ids[0]));
+        memset(sched->hv_tensor_copies,       0, sched->hash_set.size * sched->n_backends * sched->n_copies * sizeof(struct ggml_tensor *));
+        sched->is_reset = true;
+    }
+    sched->is_alloc = false;
+}
+
+bool ggml_backend_sched_reserve(ggml_backend_sched_t sched, struct ggml_cgraph * measure_graph) {
+    GGML_ASSERT((int)sched->hash_set.size >= measure_graph->n_nodes + measure_graph->n_leafs);
+
+    ggml_backend_sched_split_graph(sched, measure_graph);
+
+    if (!ggml_gallocr_reserve_n(sched->galloc, &sched->graph, sched->node_backend_ids, sched->leaf_backend_ids)) {
+        return false;
+    }
+
+    ggml_backend_sched_reset(sched);
+    ggml_backend_sched_synchronize(sched);
+
+    return true;
+}
+
+bool ggml_backend_sched_alloc_graph(ggml_backend_sched_t sched, struct ggml_cgraph * graph) {
+    GGML_ASSERT((int)sched->hash_set.size >= graph->n_nodes + graph->n_leafs);
+
+    ggml_backend_sched_split_graph(sched, graph);
+
+
+    if (!ggml_backend_sched_alloc_splits(sched)) {
+        return false;
+    }
+
+    sched->is_alloc = true;
+
+    return true;
+}
+
+enum ggml_status ggml_backend_sched_graph_compute(ggml_backend_sched_t sched, struct ggml_cgraph * graph) {
+    enum ggml_status err = ggml_backend_sched_graph_compute_async(sched, graph);
+    ggml_backend_sched_synchronize(sched);
+    return err;
+}
+
+enum ggml_status ggml_backend_sched_graph_compute_async(ggml_backend_sched_t sched, struct ggml_cgraph * graph) {
+    if (!sched->is_reset && !sched->is_alloc) {
+        ggml_backend_sched_reset(sched);
+    }
+
+    if (!sched->is_alloc) {
+        if (!ggml_backend_sched_alloc_graph(sched, graph)) {
+            return GGML_STATUS_ALLOC_FAILED;
+        }
+    }
+
+    return ggml_backend_sched_compute_splits(sched);
+}
+
+void ggml_backend_sched_synchronize(ggml_backend_sched_t sched) {
+    for (int i = 0; i < sched->n_backends; i++) {
+        ggml_backend_synchronize(sched->backends[i]);
+    }
+}
+
+void ggml_backend_sched_set_eval_callback(ggml_backend_sched_t sched, ggml_backend_sched_eval_callback callback, void * user_data) {
+    sched->callback_eval = callback;
+    sched->callback_eval_user_data = user_data;
+}
+
+int ggml_backend_sched_get_n_splits(ggml_backend_sched_t sched) {
+    return sched->n_splits;
+}
+
+int ggml_backend_sched_get_n_copies(ggml_backend_sched_t sched) {
+    return sched->n_copies;
+}
+
+int ggml_backend_sched_get_n_backends(ggml_backend_sched_t sched) {
+    return sched->n_backends;
+}
+
+ggml_backend_t ggml_backend_sched_get_backend(ggml_backend_sched_t sched, int i) {
+    GGML_ASSERT(i >= 0 && i < sched->n_backends);
+    return sched->backends[i];
+}
+
+size_t ggml_backend_sched_get_buffer_size(ggml_backend_sched_t sched, ggml_backend_t backend) {
+    int backend_index = ggml_backend_sched_backend_id(sched, backend);
+    GGML_ASSERT(backend_index >= 0 && backend_index < sched->n_backends);
+
+    return ggml_gallocr_get_buffer_size(sched->galloc, backend_index);
+}
+
+void ggml_backend_sched_set_tensor_backend(ggml_backend_sched_t sched, struct ggml_tensor * node, ggml_backend_t backend) {
+    int backend_index = ggml_backend_sched_backend_id(sched, backend);
+    GGML_ASSERT(backend_index >= 0 && backend_index < sched->n_backends);
+    tensor_backend_id(node) = backend_index;
+    SET_CAUSE(node, "usr");
+    sched->is_reset = false;
+}
+
+ggml_backend_t ggml_backend_sched_get_tensor_backend(ggml_backend_sched_t sched, struct ggml_tensor * node) {
+    int backend_index = tensor_backend_id(node);
+    if (backend_index == -1) {
+        return NULL;
+    }
+    return sched->backends[backend_index];
+}
+
+// utils
+
+void ggml_backend_view_init(struct ggml_tensor * tensor) {
+    GGML_ASSERT(tensor->buffer == NULL);
+    GGML_ASSERT(tensor->view_src != NULL);
+    GGML_ASSERT(tensor->view_src->buffer != NULL);
+    GGML_ASSERT(tensor->view_src->data != NULL);
+
+    tensor->buffer = tensor->view_src->buffer;
+    tensor->data = (char *)tensor->view_src->data + tensor->view_offs;
+    ggml_backend_buffer_init_tensor(tensor->buffer, tensor);
+}
+
+void ggml_backend_tensor_alloc(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor, void * addr) {
+    GGML_ASSERT(tensor->buffer == NULL);
+    GGML_ASSERT(tensor->data == NULL);
+    GGML_ASSERT(tensor->view_src == NULL);
+    GGML_ASSERT(addr >= ggml_backend_buffer_get_base(buffer));
+    GGML_ASSERT((char *)addr + ggml_backend_buffer_get_alloc_size(buffer, tensor) <=
+                (char *)ggml_backend_buffer_get_base(buffer) + ggml_backend_buffer_get_size(buffer));
+
+    tensor->buffer = buffer;
+    tensor->data = addr;
+    ggml_backend_buffer_init_tensor(buffer, tensor);
+}
+
+static struct ggml_tensor * graph_copy_dup_tensor(struct ggml_hash_set hash_set, struct ggml_tensor ** node_copies,
+    struct ggml_context * ctx_allocated, struct ggml_context * ctx_unallocated, struct ggml_tensor * src) {
+
+    GGML_ASSERT(src != NULL);
+    GGML_ASSERT(src->data && "graph must be allocated");
+
+    size_t id = ggml_hash_insert(&hash_set, src);
+    if (id == GGML_HASHSET_ALREADY_EXISTS) {
+        return node_copies[ggml_hash_find(&hash_set, src)];
+    }
+
+    struct ggml_tensor * dst = ggml_dup_tensor_layout(src->data && !src->view_src ? ctx_allocated : ctx_unallocated, src);
+    if (src->view_src != NULL) {
+        dst->view_src = graph_copy_dup_tensor(hash_set, node_copies, ctx_allocated, ctx_unallocated, src->view_src);
+        dst->view_offs = src->view_offs;
+    }
+    dst->op = src->op;
+    memcpy(dst->op_params, src->op_params, sizeof(dst->op_params));
+    ggml_set_name(dst, src->name);
+
+    // copy src
+    for (int i = 0; i < GGML_MAX_SRC; i++) {
+        struct ggml_tensor * s = src->src[i];
+        if (s == NULL) {
+            continue;
+        }
+        dst->src[i] = graph_copy_dup_tensor(hash_set, node_copies, ctx_allocated, ctx_unallocated, s);
+    }
+
+    node_copies[id] = dst;
+    return dst;
+}
+
+static void graph_copy_init_tensor(struct ggml_hash_set * hash_set, struct ggml_tensor ** node_copies, bool * node_init, struct ggml_tensor * src) {
+    size_t id = ggml_hash_find(hash_set, src);
+    if (node_init[id]) {
+        return;
+    }
+    node_init[id] = true;
+
+    struct ggml_tensor * dst = node_copies[id];
+    if (dst->view_src != NULL) {
+        graph_copy_init_tensor(hash_set, node_copies, node_init, src->view_src);
+        ggml_backend_view_init(dst);
+    }
+    else {
+        ggml_backend_tensor_copy(src, dst);
+    }
+
+    // init src
+    for (int i = 0; i < GGML_MAX_SRC; i++) {
+        struct ggml_tensor * s = src->src[i];
+        if (s == NULL) {
+            continue;
+        }
+        graph_copy_init_tensor(hash_set, node_copies, node_init, s);
+    }
+}
+
+struct ggml_backend_graph_copy ggml_backend_graph_copy(ggml_backend_t backend, struct ggml_cgraph * graph) {
+    struct ggml_hash_set hash_set = ggml_hash_set_new(graph->visited_hash_set.size);
+    struct ggml_tensor ** node_copies = (ggml_tensor **) calloc(hash_set.size, sizeof(node_copies[0])); // NOLINT
+    bool * node_init = (bool *) calloc(hash_set.size, sizeof(node_init[0]));
+
+    struct ggml_init_params params = {
+        /* .mem_size   = */ ggml_tensor_overhead()*hash_set.size + ggml_graph_overhead_custom(graph->size, false),
+        /* .mem_buffer = */ NULL,
+        /* .no_alloc   = */ true
+    };
+
+    struct ggml_context * ctx_allocated = ggml_init(params);
+    struct ggml_context * ctx_unallocated = ggml_init(params);
+
+    if (ctx_allocated == NULL || ctx_unallocated == NULL) {
+        fprintf(stderr, "failed to allocate context for graph copy\n");
+        ggml_hash_set_free(&hash_set);
+        free(node_copies);
+        free(node_init);
+        ggml_free(ctx_allocated);
+        ggml_free(ctx_unallocated);
+        return {
+            /* .buffer           = */ NULL,
+            /* .ctx_allocated    = */ NULL,
+            /* .ctx_unallocated  = */ NULL,
+            /* .graph            = */ NULL,
+        };
+    }
+
+    // dup nodes
+    for (int i = 0; i < graph->n_nodes; i++) {
+        struct ggml_tensor * node = graph->nodes[i];
+        graph_copy_dup_tensor(hash_set, node_copies, ctx_allocated, ctx_unallocated, node);
+    }
+
+    // allocate nodes
+    ggml_backend_buffer_t buffer = ggml_backend_alloc_ctx_tensors(ctx_allocated, backend);
+    if (buffer == NULL) {
+        fprintf(stderr, "failed to allocate buffer for graph copy\n");
+        ggml_hash_set_free(&hash_set);
+        free(node_copies);
+        free(node_init);
+        ggml_free(ctx_allocated);
+        ggml_free(ctx_unallocated);
+        return {
+            /* .buffer           = */ NULL,
+            /* .ctx_allocated    = */ NULL,
+            /* .ctx_unallocated  = */ NULL,
+            /* .graph            = */ NULL,
+        };
+    }
+
+    //printf("copy buffer size: %zu MB\n", ggml_backend_buffer_get_size(buffer) / 1024 / 1024);
+
+    // copy data and init views
+    for (int i = 0; i < graph->n_nodes; i++) {
+        struct ggml_tensor * node = graph->nodes[i];
+        graph_copy_init_tensor(&hash_set, node_copies, node_init, node);
+    }
+
+    // build graph copy
+    struct ggml_cgraph * graph_copy = ggml_new_graph_custom(ctx_allocated, graph->size, false);
+    for (int i = 0; i < graph->n_nodes; i++) {
+        struct ggml_tensor * node = graph->nodes[i];
+        struct ggml_tensor * node_copy = node_copies[ggml_hash_find(&hash_set, node)];
+        graph_copy->nodes[i] = node_copy;
+    }
+    graph_copy->n_nodes = graph->n_nodes;
+
+    ggml_hash_set_free(&hash_set);
+    free(node_copies);
+    free(node_init);
+
+    return {
+        /* .buffer           = */ buffer,
+        /* .ctx_allocated    = */ ctx_allocated,
+        /* .ctx_unallocated  = */ ctx_unallocated,
+        /* .graph            = */ graph_copy,
+    };
+}
+
+void ggml_backend_graph_copy_free(struct ggml_backend_graph_copy copy) {
+    ggml_backend_buffer_free(copy.buffer);
+    ggml_free(copy.ctx_allocated);
+    ggml_free(copy.ctx_unallocated);
+}
+
+bool ggml_backend_compare_graph_backend(ggml_backend_t backend1, ggml_backend_t backend2, struct ggml_cgraph * graph, ggml_backend_eval_callback callback, void * user_data) {
+    struct ggml_backend_graph_copy copy = ggml_backend_graph_copy(backend2, graph);
+    if (copy.buffer == NULL) {
+        return false;
+    }
+
+    struct ggml_cgraph * g1 = graph;
+    struct ggml_cgraph * g2 = copy.graph;
+
+    assert(g1->n_nodes == g2->n_nodes);
+
+    for (int i = 0; i < g1->n_nodes; i++) {
+        //printf("eval %d/%d\n", i, g1->n_nodes);
+        struct ggml_tensor * t1 = g1->nodes[i];
+        struct ggml_tensor * t2 = g2->nodes[i];
+
+        assert(t1->op == t2->op && ggml_are_same_layout(t1, t2));
+
+        struct ggml_cgraph g1v = ggml_graph_view(g1, i, i + 1);
+        struct ggml_cgraph g2v = ggml_graph_view(g2, i, i + 1);
+
+        ggml_backend_graph_compute(backend1, &g1v);
+        ggml_backend_graph_compute(backend2, &g2v);
+
+        if (ggml_is_view_op(t1->op)) {
+            continue;
+        }
+
+        // compare results, calculate rms etc
+        if (!callback(i, t1, t2, user_data)) {
+            break;
+        }
+    }
+
+    ggml_backend_graph_copy_free(copy);
+
+    return true;
+}
diff --git a/src/whisper_cpp/ggml/src/ggml-blas.cpp b/src/whisper_cpp/ggml/src/ggml-blas.cpp
new file mode 100644
index 00000000..b850e6a8
--- /dev/null
+++ b/src/whisper_cpp/ggml/src/ggml-blas.cpp
@@ -0,0 +1,366 @@
+#include "ggml-impl.h"
+#include "ggml-blas.h"
+#include "ggml-backend-impl.h"
+
+#include <future>
+#include <vector>
+
+#if defined(GGML_USE_ACCELERATE)
+#   include <Accelerate/Accelerate.h>
+#elif defined(GGML_BLAS_USE_MKL)
+#   include <mkl.h>
+#elif defined(GGML_BLAS_USE_BLIS)
+#   include <blis.h>
+#elif defined(GGML_BLAS_USE_NVPL)
+#   include <nvpl_blas.h>
+#else
+#   include <cblas.h>
+#endif
+
+struct ggml_backend_blas_context {
+    int n_threads = GGML_DEFAULT_N_THREADS;
+    std::unique_ptr<char[]> work_data;
+    size_t work_size = 0;
+#ifndef GGML_USE_OPENMP
+    std::vector<std::future<void>> tasks;
+#endif
+};
+
+// helper function to determine if it is better to use BLAS or not
+// for large matrices, BLAS is faster
+static bool ggml_backend_blas_use_blas(const struct ggml_tensor * dst) {
+    const struct ggml_tensor * src0 = dst->src[0];
+    const struct ggml_tensor * src1 = dst->src[1];
+
+    const int64_t ne10 = src1->ne[0];
+
+    const int64_t ne0 = dst->ne[0];
+    const int64_t ne1 = dst->ne[1];
+
+    // TODO: find the optimal values for these
+    if (ggml_is_contiguous(src0) &&
+        ggml_is_contiguous(src1) &&
+        src1->type == GGML_TYPE_F32 &&
+        (ne0 >= 32 && ne1 >= 32 && ne10 >= 32)) {
+
+        /*printf("BLAS: %d %d %d %d %d\n", ne0, ne1, ne10, ne00, ne01);*/
+        return true;
+    }
+
+    return false;
+}
+
+static void ggml_backend_blas_mul_mat(ggml_backend_blas_context * ctx, struct ggml_tensor * dst) {
+    const struct ggml_tensor * src0 = dst->src[0];
+    const struct ggml_tensor * src1 = dst->src[1];
+
+    GGML_TENSOR_BINARY_OP_LOCALS
+
+    const enum ggml_type type = src0->type;
+
+    GGML_ASSERT(ne0 == ne01);
+    GGML_ASSERT(ne1 == ne11);
+    GGML_ASSERT(ne2 == ne12);
+    GGML_ASSERT(ne3 == ne13);
+
+    // we don't support permuted src0 or src1
+    GGML_ASSERT(nb00 == ggml_type_size(type));
+    GGML_ASSERT(nb10 == ggml_type_size(src1->type));
+
+    // dst cannot be transposed or permuted
+    GGML_ASSERT(nb0 == sizeof(float));
+    GGML_ASSERT(nb0 <= nb1);
+    GGML_ASSERT(nb1 <= nb2);
+    GGML_ASSERT(nb2 <= nb3);
+
+    // broadcast factors
+    const int64_t r2 = ne12/ne02;
+    const int64_t r3 = ne13/ne03;
+
+    const int64_t ne_plane      = ne01*ne00;
+    const size_t  desired_wsize = type == GGML_TYPE_F32 ? 0 : ne03*ne02*ne_plane*sizeof(float);
+
+    if (ctx->work_size < desired_wsize) {
+        ctx->work_data.reset(new char[desired_wsize]);
+        ctx->work_size = desired_wsize;
+    }
+    void * wdata = ctx->work_data.get();
+
+    // convert src0 to float
+    if (type != GGML_TYPE_F32) {
+        ggml_type_traits_t type_traits = ggml_internal_get_type_traits(type);
+        ggml_to_float_t const to_float = type_traits.to_float;
+
+        for (int64_t i03 = 0; i03 < ne03; i03++) {
+            for (int64_t i02 = 0; i02 < ne02; i02++) {
+                const void  *       x      = (char *)  src0->data + i02*nb02          + i03*nb03;
+                      float * const wplane = (float *) wdata      + i02*ne_plane      + i03*ne02*ne_plane;
+
+                const int min_cols_per_thread = 4096;
+                const int min_rows_per_thread = std::max((int)(min_cols_per_thread/ne00), 1);
+                const int n_threads = std::max(std::min(ctx->n_threads, (int)(ne01/min_rows_per_thread)), 1);
+
+#ifdef GGML_USE_OPENMP
+                #pragma omp parallel for num_threads(n_threads)
+                for (int64_t i01 = 0; i01 < ne01; i01++) {
+                    to_float((const char *) x + i01*nb01, wplane + i01*ne00, ne00);
+                }
+#else
+                for (int i = 1; i < n_threads; i++) {
+                    const int64_t start =       i*ne01/n_threads;
+                    const int64_t end   = (i + 1)*ne01/n_threads;
+                    if (start < end) {
+                        ctx->tasks.push_back(std::async(std::launch::async, [=]() {
+                            for (int64_t i01 = start; i01 < end; i01++) {
+                                to_float((const char *) x + i01*nb01, wplane + i01*ne00, ne00);
+                            }
+                        }));
+                    }
+                }
+                {
+                    // reuse the current thread for the first task
+                    const int64_t start = 0;
+                    const int64_t end   = ne01/n_threads;
+                    for (int64_t i01 = start; i01 < end; i01++) {
+                        to_float((const char *) x + i01*nb01, wplane + i01*ne00, ne00);
+                    }
+                }
+#endif
+            }
+        }
+
+#ifndef GGML_USE_OPENMP
+        // wait for all tasks to finish
+        for (auto & task : ctx->tasks) {
+            task.get();
+        }
+        ctx->tasks.clear();
+#endif
+    }
+
+#if defined(OPENBLAS_VERSION)
+    openblas_set_num_threads(ctx->n_threads);
+#endif
+
+#if defined(GGML_BLAS_USE_BLIS)
+    bli_thread_set_num_threads(ctx->n_threads);
+#endif
+
+#if defined(GGML_BLAS_USE_NVPL)
+    nvpl_blas_set_num_threads(ctx->n_threads);
+#endif
+
+    for (int64_t i13 = 0; i13 < ne13; i13++) {
+        for (int64_t i12 = 0; i12 < ne12; i12++) {
+            const int64_t i03 = i13/r3;
+            const int64_t i02 = i12/r2;
+
+            const float * x = (float *) ((char *) src0->data + i02*nb02 + i03*nb03);
+            const float * y = (float *) ((char *) src1->data + i12*nb12 + i13*nb13);
+                  float * d = (float *) ((char *)  dst->data + i12*nb2  + i13*nb3);
+
+            if (type != GGML_TYPE_F32) {
+                x = (float *) wdata + i02*ne_plane + i03*ne02*ne_plane;
+            }
+
+            cblas_sgemm(CblasRowMajor, CblasNoTrans, CblasTrans,
+                        ne1, ne01, ne10,
+                        1.0f,   y, ne10,
+                                x, ne00,
+                        0.0f,   d, ne01);
+        }
+    }
+}
+
+static void ggml_backend_blas_out_prod(ggml_backend_blas_context * ctx, struct ggml_tensor * dst) {
+    const struct ggml_tensor * src0 = dst->src[0];
+    const struct ggml_tensor * src1 = dst->src[1];
+
+    GGML_TENSOR_BINARY_OP_LOCALS
+
+    GGML_ASSERT(ne0  == ne00);
+    GGML_ASSERT(ne1  == ne10);
+    GGML_ASSERT(ne2  == ne02);
+    GGML_ASSERT(ne02 == ne12);
+    GGML_ASSERT(ne3  == ne13);
+    GGML_ASSERT(ne03 == ne13);
+
+    // we don't support permuted src0 or src1
+    GGML_ASSERT(nb00 == sizeof(float));
+
+    // dst cannot be transposed or permuted
+    GGML_ASSERT(nb0 == sizeof(float));
+    // GGML_ASSERT(nb0 <= nb1);
+    // GGML_ASSERT(nb1 <= nb2);
+    // GGML_ASSERT(nb2 <= nb3);
+
+    // Arguments to ggml_compute_forward_out_prod (expressed as major,minor)
+    // src0: (k,n)
+    // src1: (k,m)
+    // dst:  (m,n)
+    //
+    // Arguments to sgemm (see https://github.com/Reference-LAPACK/lapack/blob/master/BLAS/SRC/sgemm.f)
+    // Also expressed as (major,minor)
+    // a: (m,k): so src1 transposed
+    // b: (k,n): so src0
+    // c: (m,n)
+    //
+    // However, if ggml_is_transposed(src1) is true, then
+    // src1->data already contains a transposed version, so sgemm mustn't
+    // transpose it further.
+
+    int n = src0->ne[0];
+    int k = src0->ne[1];
+    int m = src1->ne[0];
+
+    CBLAS_TRANSPOSE transposeA;
+    int lda;
+
+    if (!ggml_is_transposed(src1)) {
+        transposeA = CblasTrans;
+        lda = m;
+    } else {
+        transposeA = CblasNoTrans;
+        lda = k;
+    }
+
+    float * a = (float *) ((char *) src1->data);
+    float * b = (float *) ((char *) src0->data);
+    float * c = (float *) ((char *) dst->data);
+
+    cblas_sgemm(CblasRowMajor, transposeA, CblasNoTrans, m, n, k, 1.0, a, lda, b, n, 0.0, c, n);
+
+    GGML_UNUSED(ctx);
+}
+
+// backend interface
+
+static const char * ggml_backend_blas_name(ggml_backend_t backend) {
+    return "BLAS";
+
+    GGML_UNUSED(backend);
+}
+
+static void ggml_backend_blas_free(ggml_backend_t backend) {
+    ggml_backend_blas_context * ctx = (ggml_backend_blas_context *)backend->context;
+    delete ctx;
+    delete backend;
+}
+
+static ggml_backend_buffer_type_t ggml_backend_blas_get_default_buffer_type(ggml_backend_t backend) {
+    return ggml_backend_cpu_buffer_type();
+
+    GGML_UNUSED(backend);
+}
+
+static enum ggml_status ggml_backend_blas_graph_compute(ggml_backend_t backend, struct ggml_cgraph * cgraph) {
+    ggml_backend_blas_context * ctx = (ggml_backend_blas_context *)backend->context;
+
+    for (int i = 0; i < cgraph->n_nodes; i++) {
+        struct ggml_tensor * node = cgraph->nodes[i];
+
+        switch (node->op) {
+            case GGML_OP_MUL_MAT:
+                ggml_backend_blas_mul_mat(ctx, node);
+                break;
+
+            case GGML_OP_OUT_PROD:
+                ggml_backend_blas_out_prod(ctx, node);
+                break;
+
+            case GGML_OP_NONE:
+            case GGML_OP_RESHAPE:
+            case GGML_OP_VIEW:
+            case GGML_OP_PERMUTE:
+            case GGML_OP_TRANSPOSE:
+                break;
+
+            default:
+                GGML_ABORT("%s: unsupported op %s\n", __func__, ggml_op_desc(node));
+        }
+    }
+
+    return GGML_STATUS_SUCCESS;
+
+    GGML_UNUSED(backend);
+}
+
+static bool ggml_backend_blas_supports_op(ggml_backend_t backend, const struct ggml_tensor * op) {
+    const struct ggml_tensor * src0 = op->src[0];
+    const struct ggml_tensor * src1 = op->src[1];
+
+    return (op->op == GGML_OP_MUL_MAT  && ggml_backend_blas_use_blas(op)) ||
+           (op->op == GGML_OP_OUT_PROD && op->src[0]->type == GGML_TYPE_F32 &&
+                                          op->src[1]->type == GGML_TYPE_F32 &&
+                                          ggml_is_matrix(src0) &&
+                                          ggml_is_matrix(src1) &&
+                                          ggml_is_contiguous(src0) &&
+                                          (ggml_is_contiguous(src1) || ggml_is_transposed(src1)));
+
+    GGML_UNUSED(backend);
+}
+
+static bool ggml_backend_blas_supports_buft(ggml_backend_t backend, ggml_backend_buffer_type_t buft) {
+    return ggml_backend_buft_is_host(buft);
+
+    GGML_UNUSED(backend);
+}
+
+static struct ggml_backend_i blas_backend_i = {
+    /* .get_name                = */ ggml_backend_blas_name,
+    /* .free                    = */ ggml_backend_blas_free,
+    /* .get_default_buffer_type = */ ggml_backend_blas_get_default_buffer_type,
+    /* .set_tensor_async        = */ NULL,
+    /* .get_tensor_async        = */ NULL,
+    /* .cpy_tensor_async        = */ NULL,
+    /* .synchronize             = */ NULL,
+    /* .graph_plan_create       = */ NULL,
+    /* .graph_plan_free         = */ NULL,
+    /* .graph_plan_update       = */ NULL,
+    /* .graph_plan_compute      = */ NULL,
+    /* .graph_compute           = */ ggml_backend_blas_graph_compute,
+    /* .supports_op             = */ ggml_backend_blas_supports_op,
+    /* .supports_buft           = */ ggml_backend_blas_supports_buft,
+    /* .offload_op              = */ NULL,
+    /* .event_record            = */ NULL,
+    /* .event_wait              = */ NULL,
+};
+
+static ggml_guid_t ggml_backend_blas_guid(void) {
+    static ggml_guid guid = { 0x12, 0xa8, 0xae, 0xf4, 0xc0, 0x1e, 0x61, 0x97, 0x8f, 0xeb, 0x33, 0x04, 0xa1, 0x33, 0x51, 0x2d };
+    return &guid;
+}
+
+ggml_backend_t ggml_backend_blas_init(void) {
+    ggml_backend_blas_context * ctx = new ggml_backend_blas_context;
+
+    ggml_backend_t backend = new ggml_backend {
+        /* .guid      = */ ggml_backend_blas_guid(),
+        /* .interface = */ blas_backend_i,
+        /* .device    = */ nullptr,
+        /* .context   = */ ctx,
+    };
+
+#if !defined(NDEBUG) && defined(OPENBLAS_VERSION) && defined(GGML_USE_OPENMP)
+    if (openblas_get_parallel() != OPENBLAS_OPENMP) {
+        fprintf(stderr, "%s: warning: ggml is using OpenMP, but OpenBLAS was compiled without OpenMP support\n", __func__);
+    }
+#endif
+
+#if !defined(NDEBUG) && defined(BLIS_ENABLE_CBLAS) && defined(GGML_USE_OPENMP) && !defined(BLIS_ENABLE_OPENMP)
+    fprintf(stderr, "%s: warning: ggml is using OpenMP, but BLIS was compiled without OpenMP support\n", __func__);
+#endif
+
+    return backend;
+}
+
+bool ggml_backend_is_blas(ggml_backend_t backend) {
+    return backend != NULL && ggml_guid_matches(backend->guid, ggml_backend_blas_guid());
+}
+
+void ggml_backend_blas_set_n_threads(ggml_backend_t backend_blas, int n_threads) {
+    GGML_ASSERT(ggml_backend_is_blas(backend_blas));
+
+    ggml_backend_blas_context * ctx = (ggml_backend_blas_context *)backend_blas->context;
+    ctx->n_threads = n_threads;
+}
diff --git a/src/whisper_cpp/ggml/src/ggml-cann.cpp b/src/whisper_cpp/ggml/src/ggml-cann.cpp
new file mode 100644
index 00000000..db5f8f18
--- /dev/null
+++ b/src/whisper_cpp/ggml/src/ggml-cann.cpp
@@ -0,0 +1,2022 @@
+/*
+ * Copyright (c) 2023-2024 The ggml authors
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a copy
+ * of this software and associated documentation files (the "Software"), to
+ * deal in the Software without restriction, including without limitation the
+ * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or
+ * sell copies of the Software, and to permit persons to whom the Software is
+ * furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice shall be included in
+ * all copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+ * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
+ * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
+ * IN THE SOFTWARE.
+ */
+
+#include "ggml-cann.h"
+
+#include <acl/acl.h>
+#include <stdarg.h>
+
+#include <cmath>
+#include <cstdio>
+#include <cstring>
+#include <mutex>
+
+#include "ggml-impl.h"
+#include "ggml-backend-impl.h"
+#include "ggml-cann/aclnn_ops.h"
+#include "ggml-cann/common.h"
+
+#define GGML_COMMON_DECL_C
+
+#include "ggml-common.h"
+
+/**
+ * @brief Handles CANN errors by printing an error message and aborting.
+ *
+ * @param stmt The statement that caused the error.
+ * @param func The function in which the error occurred.
+ * @param file The file in which the error occurred.
+ * @param line The line number where the error occurred.
+ * @param msg The error message.
+ */
+[[noreturn]] void ggml_cann_error(const char* stmt, const char* func,
+                                  const char* file, int line, const char* msg) {
+    int32_t id = -1;
+    aclrtGetDevice(&id);
+
+    GGML_LOG_ERROR("CANN error: %s\n", msg);
+    GGML_LOG_ERROR("  current device: %d, in function %s at %s:%d\n", id, func,
+            file, line);
+    GGML_LOG_ERROR("  %s\n", stmt);
+    // abort with GGML_ASSERT to get a stack trace
+    GGML_ABORT("CANN error");
+}
+
+/**
+ * @brief Sets the device to be used by CANN.
+ *
+ * @param device The device ID to set.
+ */
+void ggml_cann_set_device(const int32_t device) {
+    // TODO: uncomment these lines after empty context has fixed.
+    // int current_device;
+    // ACL_CHECK(aclrtGetDevice(&current_device));
+
+    // if (device == current_device) {
+    //   return;
+    // }
+    ACL_CHECK(aclrtSetDevice(device));
+}
+
+/**
+ * @brief Retrieves the current device ID.
+ *
+ * @return The current device ID.
+ */
+int32_t ggml_cann_get_device() {
+    int32_t id;
+    ACL_CHECK(aclrtGetDevice(&id));
+    return id;
+}
+
+/**
+ * @brief Initialize the CANN device information.
+ *
+ * This function initializes the CANN device information by obtaining the
+ * device count and setting the memory allocation granularity for each device.
+ *
+ * @return A structure containing the device information.
+ */
+static ggml_cann_device_info ggml_cann_init() {
+    ggml_cann_device_info info = {};
+
+    aclError err = aclrtGetDeviceCount((uint32_t*)&info.device_count);
+
+    if (err != ACL_SUCCESS) {
+        GGML_LOG_ERROR("%s: failed to initialize CANN: %s\n",
+                __func__, aclGetRecentErrMsg());
+        return info;
+    }
+
+    GGML_ASSERT(info.device_count <= GGML_CANN_MAX_DEVICES);
+
+    for (int id = 0; id < info.device_count; ++id) {
+        aclrtPhysicalMemProp prop = {};
+        prop.handleType = ACL_MEM_HANDLE_TYPE_NONE;
+        prop.allocationType = ACL_MEM_ALLOCATION_TYPE_PINNED;
+        prop.memAttr = ACL_HBM_MEM_HUGE;
+        prop.location.type = ACL_MEM_LOCATION_TYPE_DEVICE;
+        prop.location.id = id;
+        prop.reserve = 0;
+        ACL_CHECK(aclrtMemGetAllocationGranularity(
+            &prop, ACL_RT_MEM_ALLOC_GRANULARITY_RECOMMENDED,
+            &info.devices[id].vmm_granularity));
+    }
+
+    // TODO: add more device info later.
+    return info;
+}
+
+/**
+ * @brief Retrieve the CANN device information.
+ *
+ * This function returns a reference to a structure containing the CANN device
+ * information. The device information is initialized once and reused on
+ * subsequent calls.
+ *
+ * @return A reference to the structure containing the device information.
+ */
+const ggml_cann_device_info& ggml_cann_info() {
+    static ggml_cann_device_info info = ggml_cann_init();
+    return info;
+}
+
+//#define DEBUG_CANN_MALLOC
+/**
+ * @brief A pool of CANN buffers(legacy).
+ *
+ * This class manages a pool of CANN buffers for a specific device.
+ */
+struct ggml_cann_pool_leg : public ggml_cann_pool {
+    /**
+     * @brief The maximum number of buffers in the pool.
+     */
+    static const int MAX_BUFFERS = 256;
+
+    /**
+     * @brief The device ID associated with this buffer pool.
+     */
+    int device;
+
+    /**
+     * @brief Structure representing a CANN buffer.
+     */
+    struct ggml_cann_buffer {
+        void* ptr = nullptr;  ///< Pointer to the buffer memory.
+        size_t size = 0;      ///< Size of the buffer.
+    };
+
+    /**
+     * @brief Array of CANN buffers in the pool.
+     */
+    ggml_cann_buffer buffer_pool[MAX_BUFFERS] = {};
+
+    /**
+     * @brief Total size of all buffers in the pool.
+     */
+    size_t pool_size = 0;
+
+    /**
+     * @brief Constructor to initialize the buffer pool for a specific device.
+     *
+     * @param device The device ID to associate with this buffer pool.
+     */
+    explicit ggml_cann_pool_leg(int device) : device(device) {}
+
+    /**
+     * @brief Destructor to free all buffers in the pool.
+     */
+    ~ggml_cann_pool_leg() {
+        ggml_cann_set_device(device);
+        for (int i = 0; i < MAX_BUFFERS; ++i) {
+            ggml_cann_buffer& b = buffer_pool[i];
+            if (b.ptr != nullptr) {
+                ACL_CHECK(aclrtFree(b.ptr));
+                pool_size -= b.size;
+            }
+        }
+        GGML_ASSERT(pool_size == 0);
+    }
+
+    /**
+     * @brief Allocate a buffer of the given size.
+     *
+     * @param size The size of the buffer to allocate.
+     * @param actual_size A pointer to a variable to receive the actual size of
+     * the allocated buffer.
+     * @return A pointer to the allocated buffer.
+     */
+    void* alloc(size_t size, size_t* actual_size) override {
+#ifdef DEBUG_CANN_MALLOC
+        int nnz = 0;
+        size_t max_size = 0;
+#endif
+        size_t best_diff = 1ull << 36;
+        int ibest = -1;
+        for (int i = 0; i < MAX_BUFFERS; ++i) {
+            ggml_cann_buffer& b = buffer_pool[i];
+            if (b.ptr != nullptr) {
+#ifdef DEBUG_CANN_MALLOC
+                ++nnz;
+                if (b.size > max_size) max_size = b.size;
+#endif
+                if (b.size >= size) {
+                    size_t diff = b.size - size;
+                    if (diff < best_diff) {
+                        best_diff = diff;
+                        ibest = i;
+                        if (!best_diff) {
+                            void* ptr = b.ptr;
+                            *actual_size = b.size;
+                            b.ptr = nullptr;
+                            b.size = 0;
+                            return ptr;
+                        }
+                    }
+                }
+            }
+        }
+        if (ibest >= 0) {
+            ggml_cann_buffer& b = buffer_pool[ibest];
+            void* ptr = b.ptr;
+            *actual_size = b.size;
+            b.ptr = nullptr;
+            b.size = 0;
+            return ptr;
+        }
+        void* ptr;
+        size_t look_ahead_size = (size_t)(1.05 * size);
+        look_ahead_size = 256 * ((look_ahead_size + 255) / 256);
+        ggml_cann_set_device(device);
+        ACL_CHECK(
+            aclrtMalloc(&ptr, look_ahead_size, ACL_MEM_MALLOC_HUGE_FIRST));
+        *actual_size = look_ahead_size;
+        pool_size += look_ahead_size;
+#ifdef DEBUG_CANN_MALLOC
+        GGML_LOG_INFO(
+            "%s[%d]: %d buffers, max_size = %u MB, pool_size = %u MB, "
+            "requested %u MB\n",
+            __func__, device, nnz, (uint32_t)(max_size / 1024 / 1024),
+            (uint32_t)(pool_size / 1024 / 1024),
+            (uint32_t)(size / 1024 / 1024));
+#endif
+        return ptr;
+    }
+
+    /**
+     * @brief Free a buffer and return it to the pool.
+     *
+     * @param ptr Pointer to the buffer to free.
+     * @param size Size of the buffer to free.
+     */
+    void free(void* ptr, size_t size) override {
+        for (int i = 0; i < MAX_BUFFERS; ++i) {
+            ggml_cann_buffer& b = buffer_pool[i];
+            if (b.ptr == nullptr) {
+                b.ptr = ptr;
+                b.size = size;
+                return;
+            }
+        }
+        // memory should always buffered. these memory may still needed by
+        // tasks in stream.
+        // TODO, fix me.
+        GGML_ABORT("Cann buffer pool full, increase MAX_CANN_BUFFERS\n");
+    }
+};
+
+/**
+ * @brief A pool of CANN buffers with virtual memory.
+ *
+ * This class manages a pool of CANN buffers with virtual memory for a specific
+ * device.
+ */
+struct ggml_cann_pool_vmm : public ggml_cann_pool {
+    /**
+     * @brief The maximum size of the virtual memory pool (32 GB).
+     */
+    static const size_t CANN_POOL_VMM_MAX_SIZE = 1ull << 35;  // 32 GB
+
+    /**
+     * @brief The device ID associated with this buffer pool.
+     */
+    int device;
+
+    /**
+     * @brief Pointer to the start of the virtual memory pool.
+     */
+    void* pool_addr = 0;
+
+    /**
+     * @brief Amount of virtual memory used in the pool.
+     */
+    size_t pool_used = 0;
+
+    /**
+     * @brief Total size of the virtual memory pool.
+     */
+    size_t pool_size = 0;
+
+    /**
+     * @brief Allocation granularity for the virtual memory pool.
+     */
+    size_t granularity;
+
+    /**
+     * @brief Handles for the physical memory allocated.
+     */
+    std::vector<aclrtDrvMemHandle> handles;
+
+    /**
+     * @brief Offsets for the mapped memory regions.
+     */
+    std::vector<void*> map_offsets;
+
+    /**
+     * @brief Constructor to initialize the buffer pool with virtual memory for
+     * a specific device.
+     *
+     * @param device The device ID to associate with this buffer pool.
+     */
+    explicit ggml_cann_pool_vmm(int device)
+        : device(device),
+          granularity(ggml_cann_info().devices[device].vmm_granularity) {}
+
+    /**
+     * @brief Destructor to free all buffers in the virtual memory pool.
+     */
+    ~ggml_cann_pool_vmm() {
+        if (pool_addr != 0) {
+            for (auto& offset : map_offsets) {
+                ACL_CHECK(aclrtUnmapMem(offset));
+            }
+            for (auto& handle : handles) {
+                ACL_CHECK(aclrtFreePhysical(handle));
+            }
+            ACL_CHECK(aclrtReleaseMemAddress(pool_addr));
+        }
+    }
+
+    /**
+     * @brief Allocate a buffer of the given size in the virtual memory pool.
+     *
+     * @param size The size of the buffer to allocate.
+     * @param actual_size A pointer to a variable to receive the actual size of
+     * the allocated buffer.
+     * @return A pointer to the allocated buffer.
+     */
+    void* alloc(size_t size, size_t* actual_size) override {
+        // round up the allocation size to the alignment to ensure that all
+        // allocations are aligned for all data types
+        const size_t alignment = 128;
+        size = alignment * ((size + alignment - 1) / alignment);
+
+        size_t avail = pool_size - pool_used;
+
+        if (size > avail) {
+            // round up to the next multiple of the granularity
+            size_t reserve_size = size - avail;
+            reserve_size =
+                granularity * ((reserve_size + granularity - 1) / granularity);
+
+            GGML_ASSERT(pool_size + reserve_size <= CANN_POOL_VMM_MAX_SIZE);
+
+            // allocate more physical memory
+            aclrtPhysicalMemProp prop = {};
+            prop.handleType = ACL_MEM_HANDLE_TYPE_NONE;
+            prop.allocationType = ACL_MEM_ALLOCATION_TYPE_PINNED;
+            prop.memAttr = ACL_HBM_MEM_HUGE;
+            prop.location.type = ACL_MEM_LOCATION_TYPE_DEVICE;
+            prop.location.id = device;
+            prop.reserve = 0;
+            aclrtDrvMemHandle handle;
+            ACL_CHECK(aclrtMallocPhysical(&handle, reserve_size, &prop, 0));
+
+            // reserve virtual address space (if not already reserved)
+            if (pool_addr == 0) {
+                ACL_CHECK(aclrtReserveMemAddress(
+                    &pool_addr, CANN_POOL_VMM_MAX_SIZE, 0, NULL, 1));
+            }
+
+            // map at the end of the pool
+            ACL_CHECK(aclrtMapMem((char*)pool_addr + pool_size, reserve_size, 0,
+                                  handle, 0));
+
+            handles.push_back(handle);
+            map_offsets.push_back((char*)pool_addr + pool_size);
+
+            // add to the pool
+            pool_size += reserve_size;
+
+            // GGML_LOG_INFO("cann pool[%d]: size increased to %llu MB (
+            // reserved %llu MB)\n",
+            //       device, (unsigned long long) (pool_size/1024/1024),
+            //       (unsigned long long) (reserve_size/1024/1024));
+        }
+
+        GGML_ASSERT(pool_addr != 0);
+
+        void* ptr = (void*)((char*)pool_addr + pool_used);
+        *actual_size = size;
+        pool_used += size;
+
+#ifdef DEBUG_CANN_MALLOC
+        GGML_LOG_INFO("cann pool[%d]: allocated %llu bytes at %llx\n", device,
+               (unsigned long long)size, (unsigned long long)ptr);
+#endif
+        return ptr;
+    }
+
+    /**
+     * @brief Free a buffer and return it to the virtual memory pool.
+     *
+     * @param ptr Pointer to the buffer to free.
+     * @param size Size of the buffer to free.
+     */
+    void free(void* ptr, size_t size) override {
+#ifdef DEBUG_CANN_MALLOC
+        GGML_LOG_INFO("cann pool[%d]: freed %llu bytes at %llx\n", device,
+               (unsigned long long)size, (unsigned long long)ptr);
+#endif
+
+        pool_used -= size;
+
+        // all deallocations must be in reverse order of the allocations
+        GGML_ASSERT(ptr == (void*)((char*)pool_addr + pool_used));
+    }
+};
+
+/**
+ * @brief Create a new CANN pool for a specific device.
+ *
+ * Factory method to create a new CANN pool object based on the device type.
+ *
+ * @param device The device ID for which to create the pool.
+ * @return A unique pointer to the created CANN pool.
+ */
+std::unique_ptr<ggml_cann_pool> ggml_backend_cann_context::new_pool_for_device(
+    int device) {
+    // return std::unique_ptr<ggml_cann_pool>(new ggml_cann_pool_leg(device));
+    return std::unique_ptr<ggml_cann_pool>(new ggml_cann_pool_vmm(device));
+}
+
+// cann buffer
+/**
+ * @brief Context for managing a CANN buffer associated with a specific device.
+ *
+ * This structure holds information about a CANN buffer, including the device
+ * ID, device pointer, and a name derived from GGML_CANN_NAME and the device ID.
+ */
+struct ggml_backend_cann_buffer_context {
+    int32_t device;  ///< The device ID associated with this buffer context.
+    void* dev_ptr =
+        nullptr;  ///< Pointer to the device memory allocated for the buffer.
+
+    /**
+     * @brief Constructor to initialize the CANN buffer context.
+     *
+     * @param device The device ID associated with this buffer context.
+     * @param dev_ptr Pointer to the device memory allocated for the buffer.
+     */
+    ggml_backend_cann_buffer_context(int32_t device, void* dev_ptr)
+        : device(device),
+          dev_ptr(dev_ptr) {}
+
+    /**
+     * @brief Destructor to free the device memory allocated for the buffer.
+     */
+    ~ggml_backend_cann_buffer_context() { ACL_CHECK(aclrtFree(dev_ptr)); }
+};
+
+/**
+ * @brief Retrieve the name associated with a CANN buffer.
+ *
+ * This function returns the name of a CANN buffer, which is stored in the
+ * context of the buffer.
+ *
+ * @param buffer The CANN buffer whose name is to be retrieved.
+ * @return A pointer to a C-string containing the name of the buffer.
+ */
+
+static const char* ggml_backend_cann_buffer_get_name(
+    ggml_backend_buffer_t buffer) {
+    return "CANN";
+
+    GGML_UNUSED(buffer);
+}
+
+/**
+ * @brief Check if a buffer is a CANN buffer.
+ *
+ * This function checks if a given buffer is a CANN buffer by comparing its
+ * `get_name` function pointer to `ggml_backend_cann_buffer_get_name`.
+ *
+ * @param buffer The buffer to check.
+ * @return true if the buffer is a CANN buffer, false otherwise.
+ */
+static bool ggml_backend_buffer_is_cann(
+    ggml_backend_buffer_t buffer) {
+    return buffer->iface.get_name == ggml_backend_cann_buffer_get_name;
+}
+
+/**
+ * @brief Free resources associated with a CANN buffer.
+ *
+ * This function frees the resources associated with a CANN buffer, including
+ * its context.
+ *
+ * @param buffer The CANN buffer to free.
+ */
+static void ggml_backend_cann_buffer_free_buffer(
+    ggml_backend_buffer_t buffer) {
+    ggml_backend_cann_buffer_context* ctx =
+        (ggml_backend_cann_buffer_context*)buffer->context;
+    delete ctx;
+}
+
+/**
+ * @brief Retrieve the base pointer of a CANN buffer.
+ *
+ * This function returns the base pointer of a CANN buffer, which points to the
+ * device memory allocated for the buffer.
+ *
+ * @param buffer The CANN buffer whose base pointer is to be retrieved.
+ * @return A pointer to the base of the device memory allocated for the buffer.
+ */
+static void* ggml_backend_cann_buffer_get_base(
+    ggml_backend_buffer_t buffer) {
+    ggml_backend_cann_buffer_context* ctx =
+        (ggml_backend_cann_buffer_context*)buffer->context;
+    return ctx->dev_ptr;
+}
+
+/**
+ * @brief Transform quantized Q4.0 tensor data into a format suitable for CANN
+ * processing.
+ *
+ * This function transforms quantized Q4.0 tensor data into a format suitable
+ * for CANN processing. It extracts quantization values and scales from the
+ * source data and prepares them in a format expected by CANN operations.
+ *
+ * @param tensor Pointer to the tensor information.
+ * @param src Pointer to the source data in Q4.0 format.
+ * @param dst Pointer to the destination buffer where transformed data will be
+ * stored.
+ */
+static void ggml_backend_cann_transform_q4_0(ggml_tensor* tensor,
+                                             const void* src,
+                                             void* dst) {
+
+    int64_t n_elems = ggml_nelements(tensor);
+    int64_t groups = n_elems / QK4_0;
+    size_t quant_bytes = n_elems * sizeof(uint8_t) / 2;
+
+    uint8_t* quant_offset = (uint8_t*)dst;
+    uint16_t* scale_offset = (uint16_t*)((char*)dst + quant_bytes);
+
+    for (int i = 0; i < groups; i++) {
+        const block_q4_0* group =
+            (const block_q4_0*)((const char*)src + i * sizeof(block_q4_0));
+        *scale_offset = group->d;
+        scale_offset++;
+
+        // 0-15
+        for (int j = 0; j < QK4_0 / 2; j += 2) {
+            (*quant_offset) = (group->qs[j] & 0x0F);
+            (*quant_offset) |= ((group->qs[j + 1] << 4));
+            quant_offset++;
+        }
+
+        // 16-31
+        for (int j = 0; j < QK4_0 / 2; j += 2) {
+            (*quant_offset) = (group->qs[j] >> 4);
+            (*quant_offset) |= (group->qs[j + 1] & 0xF0);
+            quant_offset++;
+        }
+    }
+
+    // put (uint4b_t -8) into int4b_t
+    for (quant_offset = (uint8_t*)dst;
+         quant_offset < (uint8_t*)dst + quant_bytes; quant_offset++) {
+        (*quant_offset) ^= 0x88;
+    }
+}
+
+/**
+ * @brief Transform CANN processed data back into quantized Q4.0 format.
+ *
+ * This function transforms CANN processed data back into quantized Q4.0 format.
+ * It reverses the transformation performed by
+ * ggml_backend_cann_transform_q4_0(), converting the data back into its
+ * original quantized form.
+ *
+ * @param tensor Pointer to the tensor information.
+ * @param src Pointer to the source buffer containing transformed data.
+ * @param dst Pointer to the destination buffer where the Q4.0 formatted data
+ * will be stored.
+ */
+static void ggml_backend_cann_transform_back_q4_0(
+    const ggml_tensor* tensor, void* src, void* dst) {
+
+    int64_t n_elems = ggml_nelements(tensor);
+    int64_t groups = n_elems / QK4_0;
+    size_t quant_bytes = n_elems * sizeof(uint8_t) / 2;
+
+    uint8_t* quant_offset = (uint8_t*)src;
+    uint16_t* scale_offset = (uint16_t*)((char*)src + quant_bytes);
+
+    for (; quant_offset < (uint8_t*)src + quant_bytes; quant_offset++) {
+        (*quant_offset) ^= 0x88;
+    }
+    quant_offset = (uint8_t*)src;
+
+    for (int i = 0; i < groups; i++) {
+        block_q4_0* group = (block_q4_0*)((char*)dst + i * sizeof(block_q4_0));
+        group->d = *scale_offset;
+        scale_offset++;
+
+        // 0-15
+        for (int j = 0; j < QK4_0 / 2; j += 2) {
+            group->qs[j] = ((*quant_offset) & 0x0F);
+            group->qs[j + 1] = ((*quant_offset) >> 4);
+            quant_offset++;
+        }
+
+        // 16-31
+        for (int j = 0; j < QK4_0 / 2; j += 2) {
+            group->qs[j] |= ((*quant_offset) << 4);
+            group->qs[j + 1] |= ((*quant_offset) & 0xF0);
+            quant_offset++;
+        }
+    }
+}
+
+/**
+ * @brief Transform quantized Q8.0 tensor data into a format suitable for CANN
+ * processing.
+ *
+ * This function transforms quantized Q8.0 tensor data into a format suitable
+ * for CANN processing. It extracts quantization values and scales from the
+ * source data and prepares them in a format expected by CANN operations.
+ *
+ * @param tensor Pointer to the tensor information.
+ * @param src Pointer to the source data in Q8.0 format.
+ * @param dst Pointer to the destination buffer where transformed data will be
+ * stored.
+ */
+static void ggml_backend_cann_transform_q8_0(ggml_tensor* tensor,
+                                             const void* src,
+                                             void* dst) {
+    int64_t n_elems = ggml_nelements(tensor);
+    int64_t groups = n_elems / QK8_0;
+    size_t quant_bytes = n_elems * sizeof(uint8_t);
+
+    uint8_t* quant_offset = (uint8_t*)dst;
+    uint16_t* scale_offset = (uint16_t*)((char*)dst + quant_bytes);
+
+    for (int i = 0; i < groups; i++) {
+        const block_q8_0* group =
+            (const block_q8_0*)((const char*)src + i * sizeof(block_q8_0));
+        *scale_offset = group->d;
+        scale_offset++;
+        size_t group_quant_size = QK8_0 * sizeof(uint8_t);
+        memcpy(quant_offset, group->qs, group_quant_size);
+        quant_offset += group_quant_size;
+    }
+}
+
+/**
+ * @brief Transform CANN processed data back into quantized Q8.0 format.
+ *
+ * This function transforms CANN processed data back into quantized Q8.0 format.
+ * It reverses the transformation performed by
+ * ggml_backend_cann_transform_q8_0(), converting the data back into its
+ * original quantized form.
+ *
+ * @param tensor Pointer to the tensor information.
+ * @param src Pointer to the source buffer containing transformed data.
+ * @param dst Pointer to the destination buffer where the Q8.0 formatted data
+ * will be stored.
+ */
+static void ggml_backend_cann_transform_back_q8_0(
+    const ggml_tensor* tensor, const void* src, void* dst) {
+    int64_t n_elems = ggml_nelements(tensor);
+    int64_t groups = n_elems / QK8_0;
+    size_t quant_bytes = n_elems * sizeof(uint8_t);
+
+    const uint8_t* quant_offset = (const uint8_t*)src;
+    const uint16_t* scale_offset =
+        (const uint16_t*)((const char*)src + quant_bytes);
+
+    for (int i = 0; i < groups; i++) {
+        block_q8_0* group = (block_q8_0*)((char*)dst + i * sizeof(block_q8_0));
+        group->d = *scale_offset;
+        scale_offset++;
+        size_t group_quant_size = QK8_0 * sizeof(uint8_t);
+        memcpy(group->qs, quant_offset, group_quant_size);
+        quant_offset += group_quant_size;
+    }
+}
+
+/**
+ * @brief Transform tensor data based on its type for CANN processing.
+ *
+ * This function transforms tensor data based on its quantization type for CANN
+ * processing. It dispatches the transformation based on the tensor's type to
+ * specialized functions handling Q4.0 and Q8.0 formats.
+ *
+ * @param tensor Pointer to the tensor information.
+ * @param src Pointer to the source data to be transformed.
+ * @param dst Pointer to the destination buffer where transformed data will be
+ * stored.
+ */
+static void ggml_backend_cann_transform(ggml_tensor* tensor,
+                                        const void* src, void* dst) {
+    switch (tensor->type) {
+        case GGML_TYPE_Q4_0:
+            ggml_backend_cann_transform_q4_0(tensor, src, dst);
+            break;
+        case GGML_TYPE_Q8_0:
+            ggml_backend_cann_transform_q8_0(tensor, src, dst);
+            break;
+        default:
+            break;
+    }
+}
+
+/**
+ * @brief Transform CANN processed data back into tensor data based on its type.
+ *
+ * This function transforms CANN processed data back into tensor data based on
+ * its quantization type for Q4.0 and Q8.0 formats. It dispatches the
+ * transformation based on the tensor's type to specialized functions.
+ *
+ * @param tensor Pointer to the tensor information.
+ * @param src Pointer to the source data containing CANN processed data.
+ * @param dst Pointer to the destination buffer where transformed tensor data
+ * will be stored.
+ */
+static void ggml_backend_cann_transform_back(
+    const ggml_tensor* tensor, void* src, void* dst) {
+    switch (tensor->type) {
+        case GGML_TYPE_Q4_0:
+            ggml_backend_cann_transform_back_q4_0(tensor, src, dst);
+            break;
+        case GGML_TYPE_Q8_0:
+            ggml_backend_cann_transform_back_q8_0(tensor, src, dst);
+            break;
+        default:
+            break;
+    }
+}
+
+/**
+ * @brief Check if transformation is needed for a given tensor type.
+ *
+ * This function checks if transformation is needed for a given tensor type
+ * to prepare data for CANN processing.
+ *
+ * @param type The tensor type to check.
+ * @return true if transformation is needed, false otherwise.
+ */
+static bool need_transform(ggml_type type) {
+    switch (type) {
+        case GGML_TYPE_Q4_0:
+        case GGML_TYPE_Q8_0:
+            return true;
+        default:
+            return false;
+    }
+}
+
+/**
+ * @brief Initialize a tensor using data from a CANN buffer.
+ *
+ * This function initializes a tensor using data from a CANN buffer.
+ * It handles special cases such as views and quantization.
+ *
+ * @param buffer The CANN buffer from which to initialize the tensor.
+ * @param tensor Pointer to the tensor to be initialized.
+ */
+static void ggml_backend_cann_buffer_init_tensor(
+    ggml_backend_buffer_t buffer, ggml_tensor* tensor) {
+    if (tensor->view_src != NULL && tensor->view_offs == 0) {
+        GGML_ASSERT(tensor->view_src->buffer->buft == buffer->buft);
+        return;
+    }
+
+    // TODO: can backend doesn't support quantized yet. Just leave the code
+    // here.
+    if (ggml_is_quantized(tensor->type)) {
+        // Initialize padding to 0 to avoid possible NaN values
+        size_t original_size = ggml_nbytes(tensor);
+        size_t padded_size =
+            ggml_backend_buft_get_alloc_size(buffer->buft, tensor);
+
+        if (padded_size > original_size && tensor->view_src == nullptr) {
+            size_t memset_size = padded_size - original_size;
+            ACL_CHECK(aclrtMemset((char*)tensor->data + original_size,
+                                  memset_size, 0, memset_size));
+        }
+    }
+}
+
+// TODO: need handle tensor which has paddings.
+/**
+ * @brief Set tensor data in a CANN buffer.
+ *
+ * This function sets tensor data in a CANN buffer, handling transformations
+ * if needed based on the tensor's type.
+ *
+ * @param buffer The CANN buffer where the tensor data will be set.
+ * @param tensor Pointer to the tensor whose data will be set.
+ * @param data Pointer to the source data to be copied into the tensor.
+ * @param offset Offset in the source data from where to start copying.
+ * @param size Size of the data to be copied, in bytes.
+ */
+static void ggml_backend_cann_buffer_set_tensor(
+    ggml_backend_buffer_t buffer, ggml_tensor *tensor, const void *data,
+    size_t offset, size_t size) {
+    ggml_backend_cann_buffer_context *ctx =
+        (ggml_backend_cann_buffer_context *)buffer->context;
+
+    ggml_cann_set_device(ctx->device);
+    // TODO: refer to cann(#6017), it use thread's default stream.
+    // For acl, synchronous functions use this default stream.
+    // Why aclrtSynchronizeDevice?
+
+    if (!need_transform(tensor->type)) {
+        ACL_CHECK(aclrtMemcpy((char *)tensor->data + offset, size, data, size,
+                              ACL_MEMCPY_HOST_TO_DEVICE));
+    } else {
+        void *transform_buffer = malloc(size);
+        ggml_backend_cann_transform(tensor, data, transform_buffer);
+
+#ifndef NDEBUG
+        void *check_buffer = malloc(size);
+        ggml_backend_cann_transform_back(tensor, transform_buffer,
+                                         check_buffer);
+        GGML_ASSERT(memcmp(data, check_buffer, size) == 0);
+        free(check_buffer);
+#endif
+        ACL_CHECK(aclrtMemcpy((char *)tensor->data + offset, size,
+                              transform_buffer, size,
+                              ACL_MEMCPY_HOST_TO_DEVICE));
+        free(transform_buffer);
+    }
+}
+
+/**
+ * @brief Get tensor data from a CANN buffer.
+ *
+ * This function retrieves tensor data from a CANN buffer, handling
+ * transformations if needed based on the tensor's type.
+ *
+ * @param buffer The CANN buffer from which to retrieve tensor data.
+ * @param tensor Pointer to the tensor whose data will be retrieved.
+ * @param data Pointer to the destination buffer where the tensor data will be
+ * copied.
+ * @param offset Offset in the destination buffer where to start copying.
+ * @param size Size of the data to be copied, in bytes.
+ */
+static void ggml_backend_cann_buffer_get_tensor(
+    ggml_backend_buffer_t buffer, const ggml_tensor* tensor, void* data,
+    size_t offset, size_t size) {
+    ggml_backend_cann_buffer_context* ctx =
+        (ggml_backend_cann_buffer_context*)buffer->context;
+
+    ggml_cann_set_device(ctx->device);
+
+    if (!need_transform(tensor->type)) {
+        ACL_CHECK(aclrtMemcpy(data, size, (char*)tensor->data + offset, size,
+                              ACL_MEMCPY_DEVICE_TO_HOST));
+    } else {
+        void* transform_buffer = malloc(size);
+        ACL_CHECK(aclrtMemcpy(transform_buffer, size,
+                              (char*)tensor->data + offset, size,
+                              ACL_MEMCPY_DEVICE_TO_HOST));
+        ggml_backend_cann_transform_back(tensor, transform_buffer, data);
+        free(transform_buffer);
+    }
+}
+
+/**
+ * @brief Copy tensor data between CANN buffers if possible.
+ *
+ * This function copies tensor data between CANN buffers if the source and
+ * destination buffers are CANN buffers and they meet the necessary conditions
+ * (same device or devices can access each other).
+ *
+ * @param buffer The destination CANN buffer where the tensor data will be
+ * copied.
+ * @param src Pointer to the source tensor whose data will be copied.
+ * @param dst Pointer to the destination tensor where the data will be copied.
+ * @return true if the copy operation succeeded, false otherwise.
+ */
+static bool ggml_backend_cann_buffer_cpy_tensor(
+    ggml_backend_buffer_t buffer, const ggml_tensor* src, ggml_tensor* dst) {
+    if (ggml_backend_buffer_is_cann(src->buffer)) {
+        ggml_backend_cann_buffer_context* src_ctx =
+            (ggml_backend_cann_buffer_context*)src->buffer->context;
+        ggml_backend_cann_buffer_context* dst_ctx =
+            (ggml_backend_cann_buffer_context*)buffer->context;
+
+        size_t memcpy_size = ggml_nbytes(src);
+        // Same device.
+        if (src_ctx->device == dst_ctx->device) {
+            ACL_CHECK(aclrtMemcpy((char*)dst->data, memcpy_size,
+                                  (const char*)src->data, memcpy_size,
+                                  ACL_MEMCPY_DEVICE_TO_DEVICE));
+            return true;
+        } else {
+            // Different device but can access by peer.
+            int32_t canAccessPeer = 0;
+            ACL_CHECK(aclrtDeviceCanAccessPeer(&canAccessPeer, src_ctx->device,
+                                               dst_ctx->device));
+            if (canAccessPeer) {
+                ggml_cann_set_device(src_ctx->device);
+                ACL_CHECK(aclrtDeviceEnablePeerAccess(dst_ctx->device, 0));
+                ACL_CHECK(aclrtMemcpy((char*)dst->data, memcpy_size,
+                                      (const char*)src->data, memcpy_size,
+                                      ACL_MEMCPY_DEVICE_TO_DEVICE));
+                return true;
+            }
+        }
+    }
+    return false;
+}
+
+/**
+ * @brief Clear a CANN buffer by setting all its memory to a specified value.
+ *
+ * This function clears a CANN buffer by setting all its memory to a specified
+ * value.
+ *
+ * @param buffer The CANN buffer to be cleared.
+ * @param value The value to which each byte in the buffer will be set.
+ */
+static void ggml_backend_cann_buffer_clear(
+    ggml_backend_buffer_t buffer, uint8_t value) {
+    ggml_backend_cann_buffer_context* ctx =
+        (ggml_backend_cann_buffer_context*)buffer->context;
+
+    ggml_cann_set_device(ctx->device);
+    ACL_CHECK(aclrtMemset(ctx->dev_ptr, buffer->size, value, buffer->size));
+}
+
+/**
+ * @brief Interface for a CANN buffer in the backend.
+ *
+ * This structure defines function pointers to operations that can be performed
+ * on a CANN buffer within the backend.
+ */
+static ggml_backend_buffer_i ggml_backend_cann_buffer_interface = {
+    /* .get_name        = */ ggml_backend_cann_buffer_get_name,
+    /* .free_buffer     = */ ggml_backend_cann_buffer_free_buffer,
+    /* .get_base        = */ ggml_backend_cann_buffer_get_base,
+    /* .init_tensor     = */ ggml_backend_cann_buffer_init_tensor,
+    /* .memset_tensor   = */ NULL,
+    /* .set_tensor      = */ ggml_backend_cann_buffer_set_tensor,
+    /* .get_tensor      = */ ggml_backend_cann_buffer_get_tensor,
+    /* .cpy_tensor      = */ ggml_backend_cann_buffer_cpy_tensor,
+    /* .clear           = */ ggml_backend_cann_buffer_clear,
+    /* .reset           = */ NULL,
+};
+
+// cann buffer type
+/**
+ * @brief Structure representing context information for a specific backend
+ * buffer type.
+ */
+struct ggml_backend_cann_buffer_type_context {
+    int32_t
+        device; /**< Device identifier associated with the buffer context. */
+    std::string name; /**< Name associated with the buffer context. */
+};
+
+/**
+ * @brief Retrieves the name associated with a CANN buffer type.
+ *
+ * This function returns the descriptive name associated with the specified
+ * CANN buffer type context.
+ *
+ * @param buft Pointer to the buffer type context.
+ * @return Const pointer to the C-style string containing the name.
+ */
+static const char* ggml_backend_cann_buffer_type_name(
+    ggml_backend_buffer_type_t buft) {
+    return "CANN";
+
+    GGML_UNUSED(buft);
+}
+
+/**
+ * @brief Allocates a new CANN buffer of the specified type and size.
+ *
+ * This function allocates a new CANN buffer on the specified device with the
+ * given size.
+ *
+ * @param buft Pointer to the buffer type context.
+ * @param size Size in bytes of the buffer to allocate.
+ * @return Pointer to the allocated buffer, or nullptr if allocation fails.
+ */
+static ggml_backend_buffer_t
+ggml_backend_cann_buffer_type_alloc_buffer(ggml_backend_buffer_type_t buft,
+                                           size_t size) {
+    ggml_backend_cann_buffer_type_context* buft_ctx =
+        (ggml_backend_cann_buffer_type_context*)buft->context;
+
+    ggml_cann_set_device(buft_ctx->device);
+
+    size = std::max(size, (size_t)1);
+
+    void* dev_ptr;
+    aclError err = aclrtMalloc(&dev_ptr, size, ACL_MEM_MALLOC_HUGE_FIRST);
+    if (err != ACL_SUCCESS) {
+        GGML_LOG_ERROR(
+            "%s: allocating %.2f MiB on device %d: aclrtMalloc failed: %s\n",
+            __func__, size / 1024.0 / 1024.0, buft_ctx->device,
+            aclGetRecentErrMsg());
+        return nullptr;
+    }
+
+    ggml_backend_cann_buffer_context* ctx =
+        new ggml_backend_cann_buffer_context(buft_ctx->device, dev_ptr);
+
+    return ggml_backend_buffer_init(buft, ggml_backend_cann_buffer_interface,
+                                    ctx, size);
+}
+
+/**
+ * @brief Retrieves the memory alignment requirement for CANN buffers of this
+ * type.
+ *
+ * This function returns the alignment requirement in bytes for memory allocated
+ * by the CANN buffer type.
+ *
+ * @param buft Pointer to the buffer type context (unused in this
+ * implementation).
+ * @return The alignment requirement in bytes (fixed at 128 bytes for CANN
+ * buffers).
+ */
+static size_t ggml_backend_cann_buffer_type_get_alignment(
+    ggml_backend_buffer_type_t buft) {
+    return 128;
+
+    GGML_UNUSED(buft);
+}
+
+/**
+ * @brief Calculates the allocation size required for a tensor in a CANN buffer.
+ *
+ * Computes the total allocation size needed for storing the tensor's data in a
+ * CANN buffer, considering any necessary padding or adjustments for quantized
+ * types.
+ *
+ * @param buft Pointer to the buffer type context (unused in this
+ * implementation).
+ * @param tensor Pointer to the tensor for which the allocation size is
+ * calculated.
+ * @return The total allocation size in bytes required for the tensor in the
+ * CANN buffer.
+ */
+static size_t ggml_backend_cann_buffer_type_get_alloc_size(
+    ggml_backend_buffer_type_t buft, const ggml_tensor* tensor) {
+    size_t size = ggml_nbytes(tensor);
+    int64_t ne0 = tensor->ne[0];
+
+    // last line must bigger than 32, because every single op deal at
+    // least 32 bytes.
+    // TODO: quantized type?
+    // int64_t line_size = ne0 * ggml_element_size(tensor);
+    // int64_t line_size_align_32 = (line_size + 31) & ~31;
+    // size += (line_size_align_32 - line_size);
+
+    // TODO: not support quantized yet.
+    // TODO: consider un-continue tensor.
+    if (ggml_is_quantized(tensor->type)) {
+        if (ne0 % MATRIX_ROW_PADDING != 0) {
+            size += ggml_row_size(
+                tensor->type, MATRIX_ROW_PADDING - ne0 % MATRIX_ROW_PADDING);
+        }
+    }
+
+    return size;
+
+    GGML_UNUSED(buft);
+}
+
+/**
+ * @brief Interface for managing CANN buffer types in the GGML backend.
+ *
+ * Provides function pointers for allocating, querying properties, and managing
+ * memory for CANN buffer types in the GGML backend.
+ */
+static ggml_backend_buffer_type_i ggml_backend_cann_buffer_type_interface = {
+    /* .get_name         = */ ggml_backend_cann_buffer_type_name,
+    /* .alloc_buffer     = */ ggml_backend_cann_buffer_type_alloc_buffer,
+    /* .get_alignment    = */ ggml_backend_cann_buffer_type_get_alignment,
+    /* .get_max_size     = */ NULL,  // defaults to SIZE_MAX
+    /* .get_alloc_size   = */ ggml_backend_cann_buffer_type_get_alloc_size,
+    /* .is_host          = */ NULL,
+};
+
+/**
+ * @brief Retrieves the CANN buffer type for a specified device.
+ *
+ * This function initializes and returns the buffer type interface associated
+ * with the given device. It ensures thread-safe access using a mutex.
+ *
+ * @param device The device index for which to retrieve the buffer type.
+ * @return A pointer to the buffer type interface for the specified device, or
+ * nullptr if the device index is out of range.
+ */
+ggml_backend_buffer_type_t
+ggml_backend_cann_buffer_type(int32_t device) {
+    static std::mutex mutex;
+    std::lock_guard<std::mutex> lock(mutex);
+
+    if (device >= ggml_backend_cann_get_device_count()) {
+        return nullptr;
+    }
+
+    static ggml_backend_buffer_type
+        ggml_backend_cann_buffer_types[GGML_CANN_MAX_DEVICES];
+
+    static bool ggml_backend_cann_buffer_type_initialized = false;
+
+    if (!ggml_backend_cann_buffer_type_initialized) {
+        for (int32_t i = 0; i < GGML_CANN_MAX_DEVICES; i++) {
+            ggml_backend_cann_buffer_types[i] = {
+                /* .iface    = */ ggml_backend_cann_buffer_type_interface,
+                /* .context  = */
+                 new ggml_backend_cann_buffer_type_context{
+                    i, "CANN" + std::to_string(i)},
+            };
+        }
+        ggml_backend_cann_buffer_type_initialized = true;
+    }
+
+    return &ggml_backend_cann_buffer_types[device];
+}
+
+/**
+ * @brief Retrieves the name associated with a CANN host buffer type.
+ *
+ * This function returns the descriptive name associated with the specified
+ * CANN host buffer type context.
+ *
+ * @param buft Pointer to the host buffer type context.
+ * @return Const pointer to the C-style string containing the name.
+ */
+static const char * ggml_backend_cann_host_buffer_type_name(ggml_backend_buffer_type_t buft) {
+    return "CANN_Host";
+
+    GGML_UNUSED(buft);
+}
+
+/**
+ * @brief Retrieves the name associated with a CANN host buffer.
+ *
+ * This function returns the descriptive name associated with the specified
+ * CANN host buffer context.
+ *
+ * @param buft Pointer to the host buffer context.
+ * @return Const pointer to the C-style string containing the name.
+ */
+static const char * ggml_backend_cann_host_buffer_name(ggml_backend_buffer_t buffer) {
+    return "CANN_Host";
+
+    GGML_UNUSED(buffer);
+}
+
+/**
+ * @brief Free resources associated with a CANN host buffer.
+ *
+ * This function frees the resources associated with a CANN host buffer, including
+ * its context.
+ *
+ * @param buffer The CANN host buffer to free.
+ */
+static void ggml_backend_cann_host_buffer_free(ggml_backend_buffer_t buffer) {
+    ACL_CHECK(aclrtFreeHost(buffer->context));
+}
+
+/**
+ * @brief Allocates a new CANN host buffer of the specified size.
+ *
+ * This function allocates a new CANN host buffer with the given size.
+ * @param size Size in bytes of the host buffer to allocate.
+ * @return Pointer to the allocated host buffer, or nullptr if allocation fails.
+ */
+static void * ggml_cann_host_malloc(size_t size) {
+    if (getenv("GGML_CANN_NO_PINNED") != nullptr) {
+        return nullptr;
+    }
+
+    void * hostPtr = nullptr;
+    aclError err = aclrtMallocHost((void **) &hostPtr, size);
+    if (err != ACL_SUCCESS) {
+
+        GGML_LOG_WARN("%s: failed to allocate %.2f MiB of pinned memory: %s\n", __func__,
+                           size / 1024.0 / 1024.0, aclGetRecentErrMsg());
+        return nullptr;
+    }
+    return hostPtr;
+}
+
+/**
+ * @brief Allocates a new CANN host buffer of the specified type and size.
+ *
+ * @param buft Pointer to the host buffer type context.
+ * @param size Size in bytes of the host buffer to allocate.
+ * @return Pointer to the allocated host buffer, or CPU buffer pointer if allocation fails.
+ */
+static ggml_backend_buffer_t ggml_backend_cann_host_buffer_type_alloc_buffer(ggml_backend_buffer_type_t buft, size_t size) {
+    void * hostPtr = ggml_cann_host_malloc(size);
+
+    if (hostPtr == nullptr) {
+        // fallback to cpu buffer
+        return ggml_backend_buft_alloc_buffer(ggml_backend_cpu_buffer_type(), size);
+    }
+
+    ggml_backend_buffer_t buffer = ggml_backend_cpu_buffer_from_ptr(hostPtr, size);
+    buffer->buft = buft;
+    buffer->iface.get_name = ggml_backend_cann_host_buffer_name;
+    buffer->iface.free_buffer = ggml_backend_cann_host_buffer_free;
+
+    return buffer;
+}
+
+/**
+ * @brief Interface for managing CANN host buffer types in the GGML backend.
+ *
+ * Provides function pointers for allocating, querying properties, and managing
+ * memory for CANN buffer types in the GGML backend.
+ */
+ggml_backend_buffer_type_t ggml_backend_cann_host_buffer_type() {
+    static struct ggml_backend_buffer_type ggml_backend_cann_buffer_type_host = {
+        /* .iface    = */ {
+            /* .get_name         = */ ggml_backend_cann_host_buffer_type_name,
+            /* .alloc_buffer     = */ ggml_backend_cann_host_buffer_type_alloc_buffer,
+            /* .get_alignment    = */ ggml_backend_cpu_buffer_type()->iface.get_alignment,
+            /* .get_max_size     = */ NULL, // defaults to SIZE_MAX
+            /* .get_alloc_size   = */ ggml_backend_cpu_buffer_type()->iface.get_alloc_size,
+            /* .is_host          = */ ggml_backend_cpu_buffer_type()->iface.is_host,
+        },
+        /* .device   = */ nullptr,
+        /* .context  = */ nullptr,
+    };
+
+    return &ggml_backend_cann_buffer_type_host;
+}
+
+/**
+ * @brief Computes the forward operation for a given tensor using CANN
+ * operations.
+ *
+ * This function selects the appropriate CANN operation based on the type of
+ * operation specified in the tensor and performs the computation.
+ *
+ * @param ctx The CANN context containing necessary resources and
+ * configurations.
+ * @param dst The destination tensor where the result of the computation will be
+ * stored.
+ * @return true if the computation was successful; false otherwise.
+ */
+static bool ggml_cann_compute_forward(ggml_backend_cann_context& ctx,
+                                      struct ggml_tensor* dst) {
+    switch (dst->op) {
+        case GGML_OP_REPEAT:
+            ggml_cann_repeat(ctx, dst);
+            break;
+        case GGML_OP_GET_ROWS:
+            ggml_cann_get_rows(ctx, dst);
+            break;
+        case GGML_OP_DUP:
+            ggml_cann_dup(ctx, dst);
+            break;
+        case GGML_OP_ADD:
+            ggml_cann_add(ctx, dst);
+            break;
+        case GGML_OP_ACC:
+            ggml_cann_acc(ctx, dst);
+            break;
+        case GGML_OP_MUL:
+            ggml_cann_mul_div<aclnnMulGetWorkspaceSize, aclnnMul>(ctx, dst);
+            break;
+        case GGML_OP_DIV:
+            ggml_cann_mul_div<aclnnDivGetWorkspaceSize, aclnnDiv>(ctx, dst);
+            break;
+        case GGML_OP_UNARY:
+            switch (ggml_get_unary_op(dst)) {
+                case GGML_UNARY_OP_GELU:
+                    ggml_cann_activation<aclnnGeluGetWorkspaceSize, aclnnGelu>(
+                        ctx, dst);
+                    break;
+                case GGML_UNARY_OP_SILU:
+                    ggml_cann_activation<aclnnSiluGetWorkspaceSize, aclnnSilu>(
+                        ctx, dst);
+                    break;
+                // TODO: Use faster gelu??
+                case GGML_UNARY_OP_GELU_QUICK:
+                    ggml_cann_activation<aclnnGeluGetWorkspaceSize, aclnnGelu>(
+                        ctx, dst);
+                    break;
+                case GGML_UNARY_OP_TANH:
+                    ggml_cann_activation<aclnnTanhGetWorkspaceSize, aclnnTanh>(
+                        ctx, dst);
+                    break;
+                case GGML_UNARY_OP_RELU:
+                    ggml_cann_activation<aclnnReluGetWorkspaceSize, aclnnRelu>(
+                        ctx, dst);
+                    break;
+                case GGML_UNARY_OP_HARDSIGMOID:
+                    ggml_cann_activation<aclnnHardsigmoidGetWorkspaceSize,
+                                         aclnnHardsigmoid>(ctx, dst);
+                    break;
+                case GGML_UNARY_OP_HARDSWISH:
+                    ggml_cann_activation<aclnnHardswishGetWorkspaceSize,
+                                         aclnnHardswish>(ctx, dst);
+                    break;
+                default:
+                    return false;
+            }
+            break;
+        case GGML_OP_NORM:
+            ggml_cann_norm(ctx, dst);
+            break;
+        case GGML_OP_GROUP_NORM:
+            ggml_cann_group_norm(ctx, dst);
+            break;
+        case GGML_OP_CONCAT:
+            ggml_cann_concat(ctx, dst);
+            break;
+        case GGML_OP_UPSCALE:
+            ggml_cann_upsample_nearest2d(ctx, dst);
+            break;
+        case GGML_OP_PAD:
+            ggml_cann_pad(ctx, dst);
+            break;
+        case GGML_OP_ARANGE:
+            ggml_cann_arange(ctx, dst);
+            break;
+        case GGML_OP_TIMESTEP_EMBEDDING:
+            ggml_cann_timestep_embedding(ctx, dst);
+            break;
+        case GGML_OP_LEAKY_RELU:
+            ggml_cann_leaky_relu(ctx, dst);
+            break;
+        case GGML_OP_RMS_NORM:
+            ggml_cann_rms_norm(ctx, dst);
+            break;
+        case GGML_OP_MUL_MAT:
+            ggml_cann_mul_mat(ctx, dst);
+            break;
+        case GGML_OP_MUL_MAT_ID:
+            return false;
+        case GGML_OP_SCALE:
+            ggml_cann_scale(ctx, dst);
+            break;
+        case GGML_OP_SQR:
+            ggml_cann_sqr(ctx, dst);
+            break;
+        case GGML_OP_CLAMP:
+            ggml_cann_clamp(ctx, dst);
+            break;
+        case GGML_OP_CPY:
+            ggml_cann_cpy(ctx, dst);
+            break;
+        case GGML_OP_CONT:
+            ggml_cann_dup(ctx, dst);
+            break;
+        case GGML_OP_NONE:
+        case GGML_OP_RESHAPE:
+        case GGML_OP_VIEW:
+        case GGML_OP_PERMUTE:
+        case GGML_OP_TRANSPOSE:
+            break;
+        case GGML_OP_DIAG_MASK_INF:
+            ggml_cann_diag_mask(ctx, dst, -INFINITY);
+            break;
+        case GGML_OP_SOFT_MAX:
+            ggml_cann_softmax(ctx, dst);
+            break;
+        case GGML_OP_ROPE:
+            ggml_cann_rope(ctx, dst);
+            break;
+        case GGML_OP_IM2COL:
+            ggml_cann_im2col(ctx, dst);
+            break;
+        case GGML_OP_POOL_2D:
+            ggml_cann_pool2d(ctx, dst);
+            break;
+        case GGML_OP_SUM_ROWS:
+            ggml_cann_sum_rows(ctx, dst);
+            break;
+        case GGML_OP_ARGSORT:
+            ggml_cann_argsort(ctx, dst);
+            break;
+        default:
+            return false;
+    }
+
+    return true;
+}
+
+// backend
+/**
+ * @brief Retrieves the name associated with the CANN backend.
+ *
+ * This function returns the name assigned to the CANN backend, which is stored
+ * in the context of the provided backend structure.
+ *
+ * @param backend Pointer to the CANN backend structure.
+ * @return A pointer to a constant string representing the backend name.
+ */
+static const char* ggml_backend_cann_name(ggml_backend_t backend) {
+    ggml_backend_cann_context* cann_ctx =
+        (ggml_backend_cann_context*)backend->context;
+
+    return cann_ctx->name.c_str();
+}
+
+/**
+ * @brief Frees resources associated with the CANN backend.
+ *
+ * This function releases resources associated with the CANN backend context
+ * and resets the device associated with the backend to its initial state.
+ *
+ * @param backend Pointer to the CANN backend structure to be freed.
+ */
+static void ggml_backend_cann_free(ggml_backend_t backend) {
+    ggml_backend_cann_context* cann_ctx =
+        (ggml_backend_cann_context*)backend->context;
+    ACL_CHECK(aclrtSynchronizeDevice());
+    ACL_CHECK(aclrtResetDevice(cann_ctx->device));
+
+    // finalize when last backend freed.
+    if (cann_ctx->device == ggml_backend_cann_get_device_count() - 1) {
+        ACL_CHECK(aclFinalize());
+    }
+
+    delete cann_ctx;
+    delete backend;
+}
+
+/**
+ * @brief Retrieves the default buffer type associated with the CANN backend.
+ *
+ * This function returns the buffer type specific to the device associated
+ * with the CANN backend. It is used to allocate buffers for computations
+ * performed by the backend.
+ *
+ * @param backend Pointer to the CANN backend structure.
+ * @return Pointer to the buffer type structure for the CANN backend.
+ */
+static ggml_backend_buffer_type_t
+ggml_backend_cann_get_default_buffer_type(ggml_backend_t backend) {
+    ggml_backend_cann_context* cann_ctx =
+        (ggml_backend_cann_context*)backend->context;
+
+    return ggml_backend_cann_buffer_type(cann_ctx->device);
+}
+
+/**
+ * @brief Sets tensor data asynchronously in the CANN backend.
+ *
+ * This function asynchronously sets tensor data in the CANN backend. Depending
+ * on the tensor type, it may perform data transformations before copying data
+ * to the device.
+ *
+ * @param backend Pointer to the CANN backend structure.
+ * @param tensor Pointer to the tensor structure to set data for.
+ * @param data Pointer to the host data to copy to the tensor.
+ * @param offset Offset in bytes within the host data.
+ * @param size Size of the data to copy in bytes.
+ */
+static void ggml_backend_cann_set_tensor_async(ggml_backend_t backend,
+                                               ggml_tensor *tensor,
+                                               const void *data,
+                                               size_t offset,
+                                               size_t size) {
+    ggml_backend_cann_context *cann_ctx =
+        (ggml_backend_cann_context *)backend->context;
+
+    if (!need_transform(tensor->type)) {
+        ACL_CHECK(aclrtMemcpyAsync((char *)tensor->data + offset, size, data,
+                                   size, ACL_MEMCPY_HOST_TO_DEVICE,
+                                   cann_ctx->stream()));
+    } else {
+        void *transform_buffer = malloc(size);
+        ggml_backend_cann_transform(tensor, data, transform_buffer);
+
+#ifndef NDEBUG
+        void *check_buffer = malloc(size);
+        ggml_backend_cann_transform_back(tensor, transform_buffer,
+                                         check_buffer);
+        GGML_ASSERT(memcmp(data, check_buffer, size));
+        free(check_buffer);
+#endif
+        ACL_CHECK(aclrtMemcpyAsync(
+            (char *)tensor->data + offset, size, transform_buffer, size,
+            ACL_MEMCPY_HOST_TO_DEVICE, cann_ctx->stream()));
+        ACL_CHECK(aclrtSynchronizeStream(cann_ctx->stream()));
+        free(transform_buffer);
+    }
+}
+
+static void ggml_backend_cann_get_tensor_async(
+    ggml_backend_t backend, const ggml_tensor *tensor, void *data,
+    size_t offset, size_t size) {
+    ggml_backend_cann_context *cann_ctx =
+        (ggml_backend_cann_context *)backend->context;
+    ggml_backend_buffer_t buf =
+        tensor->view_src ? tensor->view_src->buffer : tensor->buffer;
+
+    GGML_ASSERT(buf->buft == ggml_backend_cann_buffer_type(cann_ctx->device) &&
+                "unsupported buffer type");
+
+    if (!need_transform(tensor->type)) {
+        ACL_CHECK(aclrtMemcpyAsync(data, size, (char *)tensor->data + offset,
+                                   size, ACL_MEMCPY_DEVICE_TO_HOST,
+                                   cann_ctx->stream()));
+    } else {
+        void *transform_buffer = malloc(size);
+        ACL_CHECK(aclrtMemcpyAsync(
+            transform_buffer, size, (char *)tensor->data + offset, size,
+            ACL_MEMCPY_DEVICE_TO_HOST, cann_ctx->stream()));
+        ACL_CHECK(aclrtSynchronizeStream(cann_ctx->stream()));
+        ggml_backend_cann_transform_back(tensor, transform_buffer, data);
+        free(transform_buffer);
+    }
+}
+
+/**
+ * @brief Asynchronously copies tensor data between CANN backends.
+ *
+ * This function copies tensor data asynchronously between two CANN backends. It
+ * checks if both tensors reside in CANN buffers and whether the devices support
+ * peer-to-peer access for direct copying. If not, it returns false.
+ *
+ * @param backend_src Pointer to the source CANN backend structure.
+ * @param backend_dst Pointer to the destination CANN backend structure.
+ * @param src Pointer to the source tensor to copy data from.
+ * @param dst Pointer to the destination tensor to copy data to.
+ * @return true if the copy operation succeeds, false otherwise.
+ */
+static bool ggml_backend_cann_cpy_tensor_async(
+    ggml_backend_t backend_src, ggml_backend_t backend_dst,
+    const ggml_tensor* src, ggml_tensor* dst) {
+    GGML_ASSERT(ggml_backend_is_cann(backend_src) ||
+                ggml_backend_is_cann(backend_dst));
+
+    if (!ggml_backend_buffer_is_cann(src->buffer) ||
+        !ggml_backend_buffer_is_cann(dst->buffer)) {
+        return false;
+    }
+
+    ggml_backend_buffer_t buf_src =
+        src->view_src ? src->view_src->buffer : src->buffer;
+    ggml_backend_buffer_t buf_dst =
+        dst->view_src ? dst->view_src->buffer : dst->buffer;
+
+    ggml_backend_cann_context* cann_ctx_src =
+        (ggml_backend_cann_context*)backend_src->context;
+    ggml_backend_cann_context* cann_ctx_dst =
+        (ggml_backend_cann_context*)backend_dst->context;
+
+    size_t copy_size = ggml_nbytes(dst);
+    if (backend_src != backend_dst) {
+        ggml_backend_cann_buffer_context* buf_ctx_src =
+            (ggml_backend_cann_buffer_context*)buf_src->context;
+        ggml_backend_cann_buffer_context* buf_ctx_dst =
+            (ggml_backend_cann_buffer_context*)buf_dst->context;
+
+        GGML_ASSERT(cann_ctx_src->device == buf_ctx_src->device);
+        GGML_ASSERT(cann_ctx_dst->device == buf_ctx_dst->device);
+
+        int32_t canAccessPeer = 0;
+        ACL_CHECK(aclrtDeviceCanAccessPeer(&canAccessPeer, cann_ctx_src->device,
+                                           cann_ctx_dst->device));
+        if (!canAccessPeer) {
+            return false;
+        }
+
+        // need open both directions for memcpyasync between devices.
+        ggml_cann_set_device(cann_ctx_dst->device);
+        ACL_CHECK(aclrtDeviceEnablePeerAccess(cann_ctx_src->device, 0));
+        ggml_cann_set_device(cann_ctx_src->device);
+        ACL_CHECK(aclrtDeviceEnablePeerAccess(cann_ctx_dst->device, 0));
+
+        ACL_CHECK(aclrtMemcpyAsync(dst->data, copy_size, src->data, copy_size,
+                                   ACL_MEMCPY_DEVICE_TO_DEVICE,
+                                   cann_ctx_src->stream()));
+
+        //TODO: workaround for Event didn`t work here.
+        aclrtSynchronizeStream(cann_ctx_src->stream());
+    } else {
+        // src and dst are on the same backend
+        ACL_CHECK(aclrtMemcpyAsync(dst->data, copy_size, src->data, copy_size,
+                                   ACL_MEMCPY_DEVICE_TO_DEVICE,
+                                   cann_ctx_dst->stream()));
+    }
+
+    return true;
+}
+
+/**
+ * @brief Synchronizes a CANN backend.
+ *
+ * This function synchronizes the specified CANN backend by waiting for all
+ * operations in its associated stream to complete.
+ *
+ * @param backend Pointer to the CANN backend structure to synchronize.
+ */
+static void ggml_backend_cann_synchronize(ggml_backend_t backend) {
+    ggml_backend_cann_context* cann_ctx =
+        (ggml_backend_cann_context*)backend->context;
+
+    ggml_cann_set_device(cann_ctx->device);
+
+    ACL_CHECK(aclrtSynchronizeStream(cann_ctx->stream()));
+}
+
+/**
+ * @brief Computes a computational graph using a CANN backend.
+ *
+ * This function computes the operations defined in the computational graph
+ * using the specified CANN backend.
+ *
+ * @param backend Pointer to the CANN backend structure to use for computation.
+ * @param cgraph Pointer to the computational graph structure containing nodes
+ *               representing operations to be computed.
+ * @return enum ggml_status Returns GGML_STATUS_SUCCESS if computation
+ *         completes successfully, otherwise an appropriate error status.
+ */
+static enum ggml_status ggml_backend_cann_graph_compute(
+    ggml_backend_t backend, ggml_cgraph* cgraph) {
+    ggml_backend_cann_context* cann_ctx =
+        (ggml_backend_cann_context*)backend->context;
+
+    ggml_cann_set_device(cann_ctx->device);
+
+    for (int i = 0; i < cgraph->n_nodes; i++) {
+        ggml_tensor* node = cgraph->nodes[i];
+
+        if (ggml_is_empty(node) || node->op == GGML_OP_NONE) {
+            continue;
+        }
+
+        bool ok = ggml_cann_compute_forward(*cann_ctx, node);
+
+        if (!ok) {
+            GGML_LOG_ERROR("%s: error: op not supported %s (%s)\n", __func__,
+                    node->name, ggml_op_name(node->op));
+        }
+        GGML_ASSERT(ok);
+    }
+
+    return GGML_STATUS_SUCCESS;
+}
+
+/**
+ * @brief Checks if the CANN backend supports a specific operation.
+ *
+ * This function checks whether the specified operation is supported by the
+ * CANN backend.
+ *
+ * @param backend Pointer to the CANN backend structure to check support for
+ *                the operation.
+ * @param op Pointer to the tensor representing the operation to check.
+ * @return bool Returns true if the operation is supported by the backend,
+ *              otherwise false.
+ */
+static bool ggml_backend_cann_supports_op(ggml_backend_t backend,
+                                                    const ggml_tensor* op) {
+    switch (op->op) {
+        case GGML_OP_UNARY:
+            switch (ggml_get_unary_op(op)) {
+                case GGML_UNARY_OP_GELU:
+                case GGML_UNARY_OP_SILU:
+                case GGML_UNARY_OP_RELU:
+                case GGML_UNARY_OP_HARDSIGMOID:
+                case GGML_UNARY_OP_HARDSWISH:
+                case GGML_UNARY_OP_GELU_QUICK:
+                case GGML_UNARY_OP_TANH:
+                    return true;
+                default:
+                    return false;
+            }
+        case GGML_OP_MUL_MAT: {
+            switch (op->src[0]->type) {
+                case GGML_TYPE_F16:
+                case GGML_TYPE_F32:
+                case GGML_TYPE_Q8_0:
+                    // TODO: fix me
+                    // Current groupsize should not be greater than k-1 in
+                    // aclnnWeightQuantBatchMatmulV2GetWorkspaceSize().
+                case GGML_TYPE_Q4_0:
+                    return true;
+                default:
+                    return false;
+            }
+        }
+        case GGML_OP_MUL_MAT_ID:
+            return false;
+        // embedding
+        case GGML_OP_GET_ROWS: {
+            switch (op->src[0]->type) {
+                case GGML_TYPE_F32:
+                case GGML_TYPE_F16:
+                case GGML_TYPE_Q4_0:
+                case GGML_TYPE_Q8_0:
+                    return true;
+                default:
+                    return false;
+            }
+        } break;
+        case GGML_OP_CPY: {
+            switch (op->type) {
+                case GGML_TYPE_F32:
+                case GGML_TYPE_F16:
+                case GGML_TYPE_Q8_0:
+                case GGML_TYPE_Q4_0:
+                    return true;
+                default:
+                    return false;
+            }
+        }
+        case GGML_OP_DUP:
+        case GGML_OP_REPEAT:
+        case GGML_OP_CONCAT:
+        case GGML_OP_NONE:
+        case GGML_OP_RESHAPE:
+        case GGML_OP_VIEW:
+        case GGML_OP_PERMUTE:
+        case GGML_OP_TRANSPOSE:
+        case GGML_OP_NORM:
+        case GGML_OP_ADD:
+        case GGML_OP_MUL:
+        case GGML_OP_DIV:
+        case GGML_OP_RMS_NORM:
+        case GGML_OP_SCALE:
+        case GGML_OP_SQR:
+        case GGML_OP_CLAMP:
+        case GGML_OP_CONT:
+        case GGML_OP_DIAG_MASK_INF:
+        case GGML_OP_SOFT_MAX:
+        case GGML_OP_ROPE:
+        case GGML_OP_IM2COL:
+        case GGML_OP_POOL_2D:
+        case GGML_OP_SUM_ROWS:
+        case GGML_OP_ARGSORT:
+        case GGML_OP_ACC:
+        case GGML_OP_GROUP_NORM:
+        case GGML_OP_UPSCALE:
+        case GGML_OP_PAD:
+        case GGML_OP_ARANGE:
+        case GGML_OP_TIMESTEP_EMBEDDING:
+        case GGML_OP_LEAKY_RELU:
+            return true;
+        default:
+            return false;
+    }
+
+    GGML_UNUSED(backend);
+}
+
+/**
+ * @brief Checks if the backend buffer type is associated with the CANN backend.
+ *
+ * This function checks whether the provided backend buffer type is associated
+ * with the CANN backend based on the comparison of its name retrieval function
+ * pointer.
+ *
+ * @param buft Pointer to the backend buffer type to check.
+ * @return bool Returns true if the buffer type is associated with the CANN
+ * backend, otherwise false.
+ */
+static bool ggml_backend_buft_is_cann(ggml_backend_buffer_type_t buft) {
+    return buft->iface.get_name == ggml_backend_cann_buffer_type_name;
+}
+
+/**
+ * @brief Checks if the CANN backend supports a specific backend buffer type.
+ *
+ * This function determines whether the CANN backend supports the given backend
+ * buffer type by comparing the device context of the backend and buffer type.
+ * It returns true if the devices are same between the backend context and
+ * buffer type context.
+ *
+ * @param backend Pointer to the CANN backend.
+ * @param buft Pointer to the backend buffer type to check.
+ * @return bool Returns true if the CANN backend supports the buffer type,
+ *              otherwise false.
+ */
+static bool ggml_backend_cann_supports_buft(
+    ggml_backend_t backend, ggml_backend_buffer_type_t buft) {
+    if (ggml_backend_buft_is_cann(buft)) {
+        ggml_backend_cann_context * cann_ctx =
+                        (ggml_backend_cann_context *)backend->context;
+        ggml_backend_cann_buffer_type_context * buft_ctx =
+                        (ggml_backend_cann_buffer_type_context *)buft->context;
+        return buft_ctx->device == cann_ctx->device;
+    }
+    return false;
+}
+
+/**
+ * @brief Determines if a tensor operation should be offloaded to the CANN
+ * backend.
+ *
+ * This function checks if a given tensor operation should be offloaded to the
+ * CANN backend based on the operation type and the size of the tensor. It
+ * returns true if the second dimension (ne[1]) of the tensor is greater than or
+ * equal to the minimum batch size and the operation is not GGML_OP_GET_ROWS.
+ *
+ * @param backend Pointer to the CANN backend.
+ * @param op Pointer to the tensor operation to check.
+ * @return bool Returns true if the operation should be offloaded, otherwise
+ * false.
+ */
+static bool ggml_backend_cann_offload_op(ggml_backend_t backend,
+                                                   const ggml_tensor* op) {
+    const int min_batch_size = 32;
+    GGML_UNUSED(backend);
+
+    return op->ne[1] >= min_batch_size && op->op != GGML_OP_GET_ROWS;
+}
+
+/**
+ * @brief Creates a new event for the CANN backend.
+ *
+ * This function initializes a new event for the CANN backend by setting the
+ * device and creating an ACL runtime event. The created event is then wrapped
+ * in a ggml_backend_event structure and returned.
+ *
+ * @param backend Pointer to the CANN backend.
+ * @return ggml_backend_event_t Returns a pointer to the new event structure.
+ */
+static ggml_backend_event_t ggml_backend_cann_event_new(
+    ggml_backend_t backend) {
+    ggml_backend_cann_context* cann_ctx =
+        (ggml_backend_cann_context*)backend->context;
+
+    ggml_cann_set_device(cann_ctx->device);
+
+    aclrtEvent event;
+    ACL_CHECK(aclrtCreateEvent(&event));
+
+    return new ggml_backend_event{
+        /* .backend = */ backend,
+        /* .context = */ event,
+    };
+}
+
+/**
+ * @brief Frees a CANN backend event.
+ *
+ * This function destroys the ACL runtime event associated with the given CANN
+ * backend event and then deletes the event structure itself.
+ *
+ * @param event Pointer to the event structure to be freed.
+ */
+static void ggml_backend_cann_event_free(ggml_backend_event_t event) {
+    ACL_CHECK(aclrtDestroyEvent((aclrtEvent)event->context));
+
+    delete event;
+}
+
+/**
+ * @brief Records an event on the CANN backend stream.
+ *
+ * This function records the given event on the ACL runtime stream associated
+ * with the backend context.
+ *
+ * @param event Pointer to the event structure to be recorded.
+ */
+static void ggml_backend_cann_event_record(ggml_backend_event_t event) {
+    ggml_backend_cann_context* cann_ctx =
+        (ggml_backend_cann_context*)event->backend->context;
+
+    ACL_CHECK(aclrtRecordEvent((aclrtEvent)event->context, cann_ctx->stream()));
+}
+
+/**
+ * @brief Waits for a recorded event to complete on the CANN backend stream.
+ *
+ * This function makes the given backend wait for the event to complete on its
+ * ACL runtime stream.
+ *
+ * @param backend Pointer to the backend structure.
+ * @param event Pointer to the event structure that the backend needs to wait
+ * for.
+ */
+static void ggml_backend_cann_event_wait(ggml_backend_t backend,
+                                         ggml_backend_event_t event) {
+    ggml_backend_cann_context* cann_ctx =
+        (ggml_backend_cann_context*)backend->context;
+
+    if (ggml_backend_is_cann(event->backend)) {
+        ACL_CHECK(aclrtStreamWaitEvent(cann_ctx->stream(),
+                                       (aclrtEvent)event->context));
+    } else {
+        GGML_ABORT("fatal error");
+    }
+}
+
+/**
+ * @brief Synchronizes the given event on the CANN backend.
+ *
+ * This function waits for the specified event to complete on the ACL runtime.
+ *
+ * @param event Pointer to the event structure to be synchronized.
+ */
+static void ggml_backend_cann_event_synchronize(ggml_backend_event_t event) {
+    ACL_CHECK(aclrtSynchronizeEvent((aclrtEvent)event->context));
+}
+
+/**
+ * @brief Structure defining the interface for the CANN backend.
+ *
+ * This structure contains function pointers for various operations
+ * supported by the CANN backend, including name retrieval, memory
+ * management, tensor operations, synchronization, and event handling.
+ */
+static ggml_backend_i ggml_backend_cann_interface = {
+    /* .get_name                = */ ggml_backend_cann_name,
+    /* .free                    = */ ggml_backend_cann_free,
+    /* .get_default_buffer_type = */ ggml_backend_cann_get_default_buffer_type,
+    /* .set_tensor_async        = */ ggml_backend_cann_set_tensor_async,
+    /* .get_tensor_async        = */ ggml_backend_cann_get_tensor_async,
+    /* .cpy_tensor_async        = */ ggml_backend_cann_cpy_tensor_async,
+    /* .synchronize             = */ ggml_backend_cann_synchronize,
+    /* .graph_plan_create       = */ NULL,
+    /* .graph_plan_free         = */ NULL,
+    /* .graph_plan_update       = */ NULL,
+    /* .graph_plan_compute      = */ NULL,
+    /* .graph_compute           = */ ggml_backend_cann_graph_compute,
+    /* .supports_op             = */ ggml_backend_cann_supports_op,
+    /* .supports_buft           = */ ggml_backend_cann_supports_buft,
+    /* .offload_op              = */ ggml_backend_cann_offload_op,
+    /* .event_record            = */ ggml_backend_cann_event_record,
+    /* .event_wait              = */ ggml_backend_cann_event_wait,
+};
+
+/**
+ * @brief Return the hardcoded GUID for the CANN backend.
+ *
+ * This function returns a static GUID which uniquely identifies the CANN
+ * backend.
+ *
+ * @return A pointer to the static GUID.
+ */
+static ggml_guid_t ggml_backend_cann_guid() {
+    static ggml_guid guid = {0xa1, 0x94, 0xaf, 0xac, 0xbd, 0x4f, 0x47, 0x34,
+                             0xbe, 0x1a, 0x9e, 0x71, 0x1f, 0x9e, 0xed, 0x64};
+    return &guid;
+}
+
+ggml_backend_t ggml_backend_cann_init(int32_t device) {
+    aclInit(nullptr);
+    if (device < 0 || device >= ggml_backend_cann_get_device_count()) {
+        GGML_LOG_ERROR("%s: error: invalid device %d\n", __func__, device);
+        return nullptr;
+    }
+
+    ggml_backend_cann_context* ctx = new ggml_backend_cann_context(device);
+    if (ctx == nullptr) {
+        GGML_LOG_ERROR("%s: error: failed to allocate context\n", __func__);
+        return nullptr;
+    }
+    ggml_cann_set_device(ctx->device);
+    ggml_backend_t cann_backend =
+        new ggml_backend{/* .guid      = */ ggml_backend_cann_guid(),
+                         /* .interface = */ ggml_backend_cann_interface,
+                         /* .device    = */ nullptr,
+                         /* .context   = */ ctx};
+
+    return cann_backend;
+}
+
+bool ggml_backend_is_cann(ggml_backend_t backend) {
+    return backend != NULL &&
+           ggml_guid_matches(backend->guid, ggml_backend_cann_guid());
+}
+
+int32_t ggml_backend_cann_get_device_count() {
+    return ggml_cann_info().device_count;
+}
+
+void ggml_backend_cann_get_device_description(
+    int32_t device, char* description, size_t description_size) {
+    ggml_cann_set_device(device);
+    const char* soc_name = aclrtGetSocName();
+    snprintf(description, description_size, "%s", soc_name);
+}
+
+void ggml_backend_cann_get_device_memory(int32_t device, size_t* free,
+                                         size_t* total) {
+    ggml_cann_set_device(device);
+    ACL_CHECK(aclrtGetMemInfo(ACL_HBM_MEM, free, total));
+}
diff --git a/src/whisper_cpp/ggml/src/ggml-cann/Doxyfile b/src/whisper_cpp/ggml/src/ggml-cann/Doxyfile
new file mode 100644
index 00000000..3290a485
--- /dev/null
+++ b/src/whisper_cpp/ggml/src/ggml-cann/Doxyfile
@@ -0,0 +1,2579 @@
+# Doxyfile 1.8.17
+
+# This file describes the settings to be used by the documentation system
+# doxygen (www.doxygen.org) for a project.
+#
+# All text after a double hash (##) is considered a comment and is placed in
+# front of the TAG it is preceding.
+#
+# All text after a single hash (#) is considered a comment and will be ignored.
+# The format is:
+# TAG = value [value, ...]
+# For lists, items can also be appended using:
+# TAG += value [value, ...]
+# Values that contain spaces should be placed between quotes (\" \").
+
+#---------------------------------------------------------------------------
+# Project related configuration options
+#---------------------------------------------------------------------------
+
+# This tag specifies the encoding used for all characters in the configuration
+# file that follow. The default is UTF-8 which is also the encoding used for all
+# text before the first occurrence of this tag. Doxygen uses libiconv (or the
+# iconv built into libc) for the transcoding. See
+# https://www.gnu.org/software/libiconv/ for the list of possible encodings.
+# The default value is: UTF-8.
+
+DOXYFILE_ENCODING      = UTF-8
+
+# The PROJECT_NAME tag is a single word (or a sequence of words surrounded by
+# double-quotes, unless you are using Doxywizard) that should identify the
+# project for which the documentation is generated. This name is used in the
+# title of most generated pages and in a few other places.
+# The default value is: My Project.
+
+PROJECT_NAME           = "ggml"
+
+# The PROJECT_NUMBER tag can be used to enter a project or revision number. This
+# could be handy for archiving the generated documentation or if some version
+# control system is used.
+
+PROJECT_NUMBER         =
+
+# Using the PROJECT_BRIEF tag one can provide an optional one line description
+# for a project that appears at the top of each page and should give viewer a
+# quick idea about the purpose of the project. Keep the description short.
+
+PROJECT_BRIEF          = "Tensor library for machine learning"
+
+# With the PROJECT_LOGO tag one can specify a logo or an icon that is included
+# in the documentation. The maximum height of the logo should not exceed 55
+# pixels and the maximum width should not exceed 200 pixels. Doxygen will copy
+# the logo to the output directory.
+
+PROJECT_LOGO           =
+
+# The OUTPUT_DIRECTORY tag is used to specify the (relative or absolute) path
+# into which the generated documentation will be written. If a relative path is
+# entered, it will be relative to the location where doxygen was started. If
+# left blank the current directory will be used.
+
+OUTPUT_DIRECTORY       = docs
+
+# If the CREATE_SUBDIRS tag is set to YES then doxygen will create 4096 sub-
+# directories (in 2 levels) under the output directory of each output format and
+# will distribute the generated files over these directories. Enabling this
+# option can be useful when feeding doxygen a huge amount of source files, where
+# putting all generated files in the same directory would otherwise causes
+# performance problems for the file system.
+# The default value is: NO.
+
+CREATE_SUBDIRS         = NO
+
+# If the ALLOW_UNICODE_NAMES tag is set to YES, doxygen will allow non-ASCII
+# characters to appear in the names of generated files. If set to NO, non-ASCII
+# characters will be escaped, for example _xE3_x81_x84 will be used for Unicode
+# U+3044.
+# The default value is: NO.
+
+ALLOW_UNICODE_NAMES    = NO
+
+# The OUTPUT_LANGUAGE tag is used to specify the language in which all
+# documentation generated by doxygen is written. Doxygen will use this
+# information to generate all constant output in the proper language.
+# Possible values are: Afrikaans, Arabic, Armenian, Brazilian, Catalan, Chinese,
+# Chinese-Traditional, Croatian, Czech, Danish, Dutch, English (United States),
+# Esperanto, Farsi (Persian), Finnish, French, German, Greek, Hungarian,
+# Indonesian, Italian, Japanese, Japanese-en (Japanese with English messages),
+# Korean, Korean-en (Korean with English messages), Latvian, Lithuanian,
+# Macedonian, Norwegian, Persian (Farsi), Polish, Portuguese, Romanian, Russian,
+# Serbian, Serbian-Cyrillic, Slovak, Slovene, Spanish, Swedish, Turkish,
+# Ukrainian and Vietnamese.
+# The default value is: English.
+
+OUTPUT_LANGUAGE        = English
+
+# The OUTPUT_TEXT_DIRECTION tag is used to specify the direction in which all
+# documentation generated by doxygen is written. Doxygen will use this
+# information to generate all generated output in the proper direction.
+# Possible values are: None, LTR, RTL and Context.
+# The default value is: None.
+
+OUTPUT_TEXT_DIRECTION  = None
+
+# If the BRIEF_MEMBER_DESC tag is set to YES, doxygen will include brief member
+# descriptions after the members that are listed in the file and class
+# documentation (similar to Javadoc). Set to NO to disable this.
+# The default value is: YES.
+
+BRIEF_MEMBER_DESC      = YES
+
+# If the REPEAT_BRIEF tag is set to YES, doxygen will prepend the brief
+# description of a member or function before the detailed description
+#
+# Note: If both HIDE_UNDOC_MEMBERS and BRIEF_MEMBER_DESC are set to NO, the
+# brief descriptions will be completely suppressed.
+# The default value is: YES.
+
+REPEAT_BRIEF           = YES
+
+# This tag implements a quasi-intelligent brief description abbreviator that is
+# used to form the text in various listings. Each string in this list, if found
+# as the leading text of the brief description, will be stripped from the text
+# and the result, after processing the whole list, is used as the annotated
+# text. Otherwise, the brief description is used as-is. If left blank, the
+# following values are used ($name is automatically replaced with the name of
+# the entity):The $name class, The $name widget, The $name file, is, provides,
+# specifies, contains, represents, a, an and the.
+
+ABBREVIATE_BRIEF       = "The $name class" \
+                         "The $name widget" \
+                         "The $name file" \
+                         is \
+                         provides \
+                         specifies \
+                         contains \
+                         represents \
+                         a \
+                         an \
+                         the
+
+# If the ALWAYS_DETAILED_SEC and REPEAT_BRIEF tags are both set to YES then
+# doxygen will generate a detailed section even if there is only a brief
+# description.
+# The default value is: NO.
+
+ALWAYS_DETAILED_SEC    = NO
+
+# If the INLINE_INHERITED_MEMB tag is set to YES, doxygen will show all
+# inherited members of a class in the documentation of that class as if those
+# members were ordinary class members. Constructors, destructors and assignment
+# operators of the base classes will not be shown.
+# The default value is: NO.
+
+INLINE_INHERITED_MEMB  = NO
+
+# If the FULL_PATH_NAMES tag is set to YES, doxygen will prepend the full path
+# before files name in the file list and in the header files. If set to NO the
+# shortest path that makes the file name unique will be used
+# The default value is: YES.
+
+FULL_PATH_NAMES        = YES
+
+# The STRIP_FROM_PATH tag can be used to strip a user-defined part of the path.
+# Stripping is only done if one of the specified strings matches the left-hand
+# part of the path. The tag can be used to show relative paths in the file list.
+# If left blank the directory from which doxygen is run is used as the path to
+# strip.
+#
+# Note that you can specify absolute paths here, but also relative paths, which
+# will be relative from the directory where doxygen is started.
+# This tag requires that the tag FULL_PATH_NAMES is set to YES.
+
+STRIP_FROM_PATH        =
+
+# The STRIP_FROM_INC_PATH tag can be used to strip a user-defined part of the
+# path mentioned in the documentation of a class, which tells the reader which
+# header file to include in order to use a class. If left blank only the name of
+# the header file containing the class definition is used. Otherwise one should
+# specify the list of include paths that are normally passed to the compiler
+# using the -I flag.
+
+STRIP_FROM_INC_PATH    =
+
+# If the SHORT_NAMES tag is set to YES, doxygen will generate much shorter (but
+# less readable) file names. This can be useful is your file systems doesn't
+# support long names like on DOS, Mac, or CD-ROM.
+# The default value is: NO.
+
+SHORT_NAMES            = NO
+
+# If the JAVADOC_AUTOBRIEF tag is set to YES then doxygen will interpret the
+# first line (until the first dot) of a Javadoc-style comment as the brief
+# description. If set to NO, the Javadoc-style will behave just like regular Qt-
+# style comments (thus requiring an explicit @brief command for a brief
+# description.)
+# The default value is: NO.
+
+JAVADOC_AUTOBRIEF      = NO
+
+# If the JAVADOC_BANNER tag is set to YES then doxygen will interpret a line
+# such as
+# /***************
+# as being the beginning of a Javadoc-style comment "banner". If set to NO, the
+# Javadoc-style will behave just like regular comments and it will not be
+# interpreted by doxygen.
+# The default value is: NO.
+
+JAVADOC_BANNER         = NO
+
+# If the QT_AUTOBRIEF tag is set to YES then doxygen will interpret the first
+# line (until the first dot) of a Qt-style comment as the brief description. If
+# set to NO, the Qt-style will behave just like regular Qt-style comments (thus
+# requiring an explicit \brief command for a brief description.)
+# The default value is: NO.
+
+QT_AUTOBRIEF           = NO
+
+# The MULTILINE_CPP_IS_BRIEF tag can be set to YES to make doxygen treat a
+# multi-line C++ special comment block (i.e. a block of //! or /// comments) as
+# a brief description. This used to be the default behavior. The new default is
+# to treat a multi-line C++ comment block as a detailed description. Set this
+# tag to YES if you prefer the old behavior instead.
+#
+# Note that setting this tag to YES also means that rational rose comments are
+# not recognized any more.
+# The default value is: NO.
+
+MULTILINE_CPP_IS_BRIEF = NO
+
+# If the INHERIT_DOCS tag is set to YES then an undocumented member inherits the
+# documentation from any documented member that it re-implements.
+# The default value is: YES.
+
+INHERIT_DOCS           = YES
+
+# If the SEPARATE_MEMBER_PAGES tag is set to YES then doxygen will produce a new
+# page for each member. If set to NO, the documentation of a member will be part
+# of the file/class/namespace that contains it.
+# The default value is: NO.
+
+SEPARATE_MEMBER_PAGES  = NO
+
+# The TAB_SIZE tag can be used to set the number of spaces in a tab. Doxygen
+# uses this value to replace tabs by spaces in code fragments.
+# Minimum value: 1, maximum value: 16, default value: 4.
+
+TAB_SIZE               = 4
+
+# This tag can be used to specify a number of aliases that act as commands in
+# the documentation. An alias has the form:
+# name=value
+# For example adding
+# "sideeffect=@par Side Effects:\n"
+# will allow you to put the command \sideeffect (or @sideeffect) in the
+# documentation, which will result in a user-defined paragraph with heading
+# "Side Effects:". You can put \n's in the value part of an alias to insert
+# newlines (in the resulting output). You can put ^^ in the value part of an
+# alias to insert a newline as if a physical newline was in the original file.
+# When you need a literal { or } or , in the value part of an alias you have to
+# escape them by means of a backslash (\), this can lead to conflicts with the
+# commands \{ and \} for these it is advised to use the version @{ and @} or use
+# a double escape (\\{ and \\})
+
+ALIASES                =
+
+# This tag can be used to specify a number of word-keyword mappings (TCL only).
+# A mapping has the form "name=value". For example adding "class=itcl::class"
+# will allow you to use the command class in the itcl::class meaning.
+
+TCL_SUBST              =
+
+# Set the OPTIMIZE_OUTPUT_FOR_C tag to YES if your project consists of C sources
+# only. Doxygen will then generate output that is more tailored for C. For
+# instance, some of the names that are used will be different. The list of all
+# members will be omitted, etc.
+# The default value is: NO.
+
+OPTIMIZE_OUTPUT_FOR_C  = NO
+
+# Set the OPTIMIZE_OUTPUT_JAVA tag to YES if your project consists of Java or
+# Python sources only. Doxygen will then generate output that is more tailored
+# for that language. For instance, namespaces will be presented as packages,
+# qualified scopes will look different, etc.
+# The default value is: NO.
+
+OPTIMIZE_OUTPUT_JAVA   = NO
+
+# Set the OPTIMIZE_FOR_FORTRAN tag to YES if your project consists of Fortran
+# sources. Doxygen will then generate output that is tailored for Fortran.
+# The default value is: NO.
+
+OPTIMIZE_FOR_FORTRAN   = NO
+
+# Set the OPTIMIZE_OUTPUT_VHDL tag to YES if your project consists of VHDL
+# sources. Doxygen will then generate output that is tailored for VHDL.
+# The default value is: NO.
+
+OPTIMIZE_OUTPUT_VHDL   = NO
+
+# Set the OPTIMIZE_OUTPUT_SLICE tag to YES if your project consists of Slice
+# sources only. Doxygen will then generate output that is more tailored for that
+# language. For instance, namespaces will be presented as modules, types will be
+# separated into more groups, etc.
+# The default value is: NO.
+
+OPTIMIZE_OUTPUT_SLICE  = NO
+
+# Doxygen selects the parser to use depending on the extension of the files it
+# parses. With this tag you can assign which parser to use for a given
+# extension. Doxygen has a built-in mapping, but you can override or extend it
+# using this tag. The format is ext=language, where ext is a file extension, and
+# language is one of the parsers supported by doxygen: IDL, Java, JavaScript,
+# Csharp (C#), C, C++, D, PHP, md (Markdown), Objective-C, Python, Slice,
+# Fortran (fixed format Fortran: FortranFixed, free formatted Fortran:
+# FortranFree, unknown formatted Fortran: Fortran. In the later case the parser
+# tries to guess whether the code is fixed or free formatted code, this is the
+# default for Fortran type files), VHDL, tcl. For instance to make doxygen treat
+# .inc files as Fortran files (default is PHP), and .f files as C (default is
+# Fortran), use: inc=Fortran f=C.
+#
+# Note: For files without extension you can use no_extension as a placeholder.
+#
+# Note that for custom extensions you also need to set FILE_PATTERNS otherwise
+# the files are not read by doxygen.
+
+EXTENSION_MAPPING      =
+
+# If the MARKDOWN_SUPPORT tag is enabled then doxygen pre-processes all comments
+# according to the Markdown format, which allows for more readable
+# documentation. See https://daringfireball.net/projects/markdown/ for details.
+# The output of markdown processing is further processed by doxygen, so you can
+# mix doxygen, HTML, and XML commands with Markdown formatting. Disable only in
+# case of backward compatibilities issues.
+# The default value is: YES.
+
+MARKDOWN_SUPPORT       = YES
+
+# When the TOC_INCLUDE_HEADINGS tag is set to a non-zero value, all headings up
+# to that level are automatically included in the table of contents, even if
+# they do not have an id attribute.
+# Note: This feature currently applies only to Markdown headings.
+# Minimum value: 0, maximum value: 99, default value: 5.
+# This tag requires that the tag MARKDOWN_SUPPORT is set to YES.
+
+TOC_INCLUDE_HEADINGS   = 5
+
+# When enabled doxygen tries to link words that correspond to documented
+# classes, or namespaces to their corresponding documentation. Such a link can
+# be prevented in individual cases by putting a % sign in front of the word or
+# globally by setting AUTOLINK_SUPPORT to NO.
+# The default value is: YES.
+
+AUTOLINK_SUPPORT       = YES
+
+# If you use STL classes (i.e. std::string, std::vector, etc.) but do not want
+# to include (a tag file for) the STL sources as input, then you should set this
+# tag to YES in order to let doxygen match functions declarations and
+# definitions whose arguments contain STL classes (e.g. func(std::string);
+# versus func(std::string) {}). This also make the inheritance and collaboration
+# diagrams that involve STL classes more complete and accurate.
+# The default value is: NO.
+
+BUILTIN_STL_SUPPORT    = NO
+
+# If you use Microsoft's C++/CLI language, you should set this option to YES to
+# enable parsing support.
+# The default value is: NO.
+
+CPP_CLI_SUPPORT        = NO
+
+# Set the SIP_SUPPORT tag to YES if your project consists of sip (see:
+# https://www.riverbankcomputing.com/software/sip/intro) sources only. Doxygen
+# will parse them like normal C++ but will assume all classes use public instead
+# of private inheritance when no explicit protection keyword is present.
+# The default value is: NO.
+
+SIP_SUPPORT            = NO
+
+# For Microsoft's IDL there are propget and propput attributes to indicate
+# getter and setter methods for a property. Setting this option to YES will make
+# doxygen to replace the get and set methods by a property in the documentation.
+# This will only work if the methods are indeed getting or setting a simple
+# type. If this is not the case, or you want to show the methods anyway, you
+# should set this option to NO.
+# The default value is: YES.
+
+IDL_PROPERTY_SUPPORT   = YES
+
+# If member grouping is used in the documentation and the DISTRIBUTE_GROUP_DOC
+# tag is set to YES then doxygen will reuse the documentation of the first
+# member in the group (if any) for the other members of the group. By default
+# all members of a group must be documented explicitly.
+# The default value is: NO.
+
+DISTRIBUTE_GROUP_DOC   = NO
+
+# If one adds a struct or class to a group and this option is enabled, then also
+# any nested class or struct is added to the same group. By default this option
+# is disabled and one has to add nested compounds explicitly via \ingroup.
+# The default value is: NO.
+
+GROUP_NESTED_COMPOUNDS = NO
+
+# Set the SUBGROUPING tag to YES to allow class member groups of the same type
+# (for instance a group of public functions) to be put as a subgroup of that
+# type (e.g. under the Public Functions section). Set it to NO to prevent
+# subgrouping. Alternatively, this can be done per class using the
+# \nosubgrouping command.
+# The default value is: YES.
+
+SUBGROUPING            = YES
+
+# When the INLINE_GROUPED_CLASSES tag is set to YES, classes, structs and unions
+# are shown inside the group in which they are included (e.g. using \ingroup)
+# instead of on a separate page (for HTML and Man pages) or section (for LaTeX
+# and RTF).
+#
+# Note that this feature does not work in combination with
+# SEPARATE_MEMBER_PAGES.
+# The default value is: NO.
+
+INLINE_GROUPED_CLASSES = NO
+
+# When the INLINE_SIMPLE_STRUCTS tag is set to YES, structs, classes, and unions
+# with only public data fields or simple typedef fields will be shown inline in
+# the documentation of the scope in which they are defined (i.e. file,
+# namespace, or group documentation), provided this scope is documented. If set
+# to NO, structs, classes, and unions are shown on a separate page (for HTML and
+# Man pages) or section (for LaTeX and RTF).
+# The default value is: NO.
+
+INLINE_SIMPLE_STRUCTS  = NO
+
+# When TYPEDEF_HIDES_STRUCT tag is enabled, a typedef of a struct, union, or
+# enum is documented as struct, union, or enum with the name of the typedef. So
+# typedef struct TypeS {} TypeT, will appear in the documentation as a struct
+# with name TypeT. When disabled the typedef will appear as a member of a file,
+# namespace, or class. And the struct will be named TypeS. This can typically be
+# useful for C code in case the coding convention dictates that all compound
+# types are typedef'ed and only the typedef is referenced, never the tag name.
+# The default value is: NO.
+
+TYPEDEF_HIDES_STRUCT   = NO
+
+# The size of the symbol lookup cache can be set using LOOKUP_CACHE_SIZE. This
+# cache is used to resolve symbols given their name and scope. Since this can be
+# an expensive process and often the same symbol appears multiple times in the
+# code, doxygen keeps a cache of pre-resolved symbols. If the cache is too small
+# doxygen will become slower. If the cache is too large, memory is wasted. The
+# cache size is given by this formula: 2^(16+LOOKUP_CACHE_SIZE). The valid range
+# is 0..9, the default is 0, corresponding to a cache size of 2^16=65536
+# symbols. At the end of a run doxygen will report the cache usage and suggest
+# the optimal cache size from a speed point of view.
+# Minimum value: 0, maximum value: 9, default value: 0.
+
+LOOKUP_CACHE_SIZE      = 0
+
+#---------------------------------------------------------------------------
+# Build related configuration options
+#---------------------------------------------------------------------------
+
+# If the EXTRACT_ALL tag is set to YES, doxygen will assume all entities in
+# documentation are documented, even if no documentation was available. Private
+# class members and static file members will be hidden unless the
+# EXTRACT_PRIVATE respectively EXTRACT_STATIC tags are set to YES.
+# Note: This will also disable the warnings about undocumented members that are
+# normally produced when WARNINGS is set to YES.
+# The default value is: NO.
+
+EXTRACT_ALL            = YES
+
+# If the EXTRACT_PRIVATE tag is set to YES, all private members of a class will
+# be included in the documentation.
+# The default value is: NO.
+
+EXTRACT_PRIVATE        = YES
+
+# If the EXTRACT_PRIV_VIRTUAL tag is set to YES, documented private virtual
+# methods of a class will be included in the documentation.
+# The default value is: NO.
+
+EXTRACT_PRIV_VIRTUAL   = YES
+
+# If the EXTRACT_PACKAGE tag is set to YES, all members with package or internal
+# scope will be included in the documentation.
+# The default value is: NO.
+
+EXTRACT_PACKAGE        = YES
+
+# If the EXTRACT_STATIC tag is set to YES, all static members of a file will be
+# included in the documentation.
+# The default value is: NO.
+
+EXTRACT_STATIC         = YES
+
+# If the EXTRACT_LOCAL_CLASSES tag is set to YES, classes (and structs) defined
+# locally in source files will be included in the documentation. If set to NO,
+# only classes defined in header files are included. Does not have any effect
+# for Java sources.
+# The default value is: YES.
+
+EXTRACT_LOCAL_CLASSES  = YES
+
+# This flag is only useful for Objective-C code. If set to YES, local methods,
+# which are defined in the implementation section but not in the interface are
+# included in the documentation. If set to NO, only methods in the interface are
+# included.
+# The default value is: NO.
+
+EXTRACT_LOCAL_METHODS  = YES
+
+# If this flag is set to YES, the members of anonymous namespaces will be
+# extracted and appear in the documentation as a namespace called
+# 'anonymous_namespace{file}', where file will be replaced with the base name of
+# the file that contains the anonymous namespace. By default anonymous namespace
+# are hidden.
+# The default value is: NO.
+
+EXTRACT_ANON_NSPACES   = NO
+
+# If the HIDE_UNDOC_MEMBERS tag is set to YES, doxygen will hide all
+# undocumented members inside documented classes or files. If set to NO these
+# members will be included in the various overviews, but no documentation
+# section is generated. This option has no effect if EXTRACT_ALL is enabled.
+# The default value is: NO.
+
+HIDE_UNDOC_MEMBERS     = NO
+
+# If the HIDE_UNDOC_CLASSES tag is set to YES, doxygen will hide all
+# undocumented classes that are normally visible in the class hierarchy. If set
+# to NO, these classes will be included in the various overviews. This option
+# has no effect if EXTRACT_ALL is enabled.
+# The default value is: NO.
+
+HIDE_UNDOC_CLASSES     = NO
+
+# If the HIDE_FRIEND_COMPOUNDS tag is set to YES, doxygen will hide all friend
+# declarations. If set to NO, these declarations will be included in the
+# documentation.
+# The default value is: NO.
+
+HIDE_FRIEND_COMPOUNDS  = NO
+
+# If the HIDE_IN_BODY_DOCS tag is set to YES, doxygen will hide any
+# documentation blocks found inside the body of a function. If set to NO, these
+# blocks will be appended to the function's detailed documentation block.
+# The default value is: NO.
+
+HIDE_IN_BODY_DOCS      = NO
+
+# The INTERNAL_DOCS tag determines if documentation that is typed after a
+# \internal command is included. If the tag is set to NO then the documentation
+# will be excluded. Set it to YES to include the internal documentation.
+# The default value is: NO.
+
+INTERNAL_DOCS          = NO
+
+# If the CASE_SENSE_NAMES tag is set to NO then doxygen will only generate file
+# names in lower-case letters. If set to YES, upper-case letters are also
+# allowed. This is useful if you have classes or files whose names only differ
+# in case and if your file system supports case sensitive file names. Windows
+# (including Cygwin) ands Mac users are advised to set this option to NO.
+# The default value is: system dependent.
+
+CASE_SENSE_NAMES       = YES
+
+# If the HIDE_SCOPE_NAMES tag is set to NO then doxygen will show members with
+# their full class and namespace scopes in the documentation. If set to YES, the
+# scope will be hidden.
+# The default value is: NO.
+
+HIDE_SCOPE_NAMES       = NO
+
+# If the HIDE_COMPOUND_REFERENCE tag is set to NO (default) then doxygen will
+# append additional text to a page's title, such as Class Reference. If set to
+# YES the compound reference will be hidden.
+# The default value is: NO.
+
+HIDE_COMPOUND_REFERENCE= NO
+
+# If the SHOW_INCLUDE_FILES tag is set to YES then doxygen will put a list of
+# the files that are included by a file in the documentation of that file.
+# The default value is: YES.
+
+SHOW_INCLUDE_FILES     = YES
+
+# If the SHOW_GROUPED_MEMB_INC tag is set to YES then Doxygen will add for each
+# grouped member an include statement to the documentation, telling the reader
+# which file to include in order to use the member.
+# The default value is: NO.
+
+SHOW_GROUPED_MEMB_INC  = NO
+
+# If the FORCE_LOCAL_INCLUDES tag is set to YES then doxygen will list include
+# files with double quotes in the documentation rather than with sharp brackets.
+# The default value is: NO.
+
+FORCE_LOCAL_INCLUDES   = NO
+
+# If the INLINE_INFO tag is set to YES then a tag [inline] is inserted in the
+# documentation for inline members.
+# The default value is: YES.
+
+INLINE_INFO            = YES
+
+# If the SORT_MEMBER_DOCS tag is set to YES then doxygen will sort the
+# (detailed) documentation of file and class members alphabetically by member
+# name. If set to NO, the members will appear in declaration order.
+# The default value is: YES.
+
+SORT_MEMBER_DOCS       = YES
+
+# If the SORT_BRIEF_DOCS tag is set to YES then doxygen will sort the brief
+# descriptions of file, namespace and class members alphabetically by member
+# name. If set to NO, the members will appear in declaration order. Note that
+# this will also influence the order of the classes in the class list.
+# The default value is: NO.
+
+SORT_BRIEF_DOCS        = NO
+
+# If the SORT_MEMBERS_CTORS_1ST tag is set to YES then doxygen will sort the
+# (brief and detailed) documentation of class members so that constructors and
+# destructors are listed first. If set to NO the constructors will appear in the
+# respective orders defined by SORT_BRIEF_DOCS and SORT_MEMBER_DOCS.
+# Note: If SORT_BRIEF_DOCS is set to NO this option is ignored for sorting brief
+# member documentation.
+# Note: If SORT_MEMBER_DOCS is set to NO this option is ignored for sorting
+# detailed member documentation.
+# The default value is: NO.
+
+SORT_MEMBERS_CTORS_1ST = NO
+
+# If the SORT_GROUP_NAMES tag is set to YES then doxygen will sort the hierarchy
+# of group names into alphabetical order. If set to NO the group names will
+# appear in their defined order.
+# The default value is: NO.
+
+SORT_GROUP_NAMES       = NO
+
+# If the SORT_BY_SCOPE_NAME tag is set to YES, the class list will be sorted by
+# fully-qualified names, including namespaces. If set to NO, the class list will
+# be sorted only by class name, not including the namespace part.
+# Note: This option is not very useful if HIDE_SCOPE_NAMES is set to YES.
+# Note: This option applies only to the class list, not to the alphabetical
+# list.
+# The default value is: NO.
+
+SORT_BY_SCOPE_NAME     = NO
+
+# If the STRICT_PROTO_MATCHING option is enabled and doxygen fails to do proper
+# type resolution of all parameters of a function it will reject a match between
+# the prototype and the implementation of a member function even if there is
+# only one candidate or it is obvious which candidate to choose by doing a
+# simple string match. By disabling STRICT_PROTO_MATCHING doxygen will still
+# accept a match between prototype and implementation in such cases.
+# The default value is: NO.
+
+STRICT_PROTO_MATCHING  = NO
+
+# The GENERATE_TODOLIST tag can be used to enable (YES) or disable (NO) the todo
+# list. This list is created by putting \todo commands in the documentation.
+# The default value is: YES.
+
+GENERATE_TODOLIST      = YES
+
+# The GENERATE_TESTLIST tag can be used to enable (YES) or disable (NO) the test
+# list. This list is created by putting \test commands in the documentation.
+# The default value is: YES.
+
+GENERATE_TESTLIST      = YES
+
+# The GENERATE_BUGLIST tag can be used to enable (YES) or disable (NO) the bug
+# list. This list is created by putting \bug commands in the documentation.
+# The default value is: YES.
+
+GENERATE_BUGLIST       = YES
+
+# The GENERATE_DEPRECATEDLIST tag can be used to enable (YES) or disable (NO)
+# the deprecated list. This list is created by putting \deprecated commands in
+# the documentation.
+# The default value is: YES.
+
+GENERATE_DEPRECATEDLIST= YES
+
+# The ENABLED_SECTIONS tag can be used to enable conditional documentation
+# sections, marked by \if <section_label> ... \endif and \cond <section_label>
+# ... \endcond blocks.
+
+ENABLED_SECTIONS       =
+
+# The MAX_INITIALIZER_LINES tag determines the maximum number of lines that the
+# initial value of a variable or macro / define can have for it to appear in the
+# documentation. If the initializer consists of more lines than specified here
+# it will be hidden. Use a value of 0 to hide initializers completely. The
+# appearance of the value of individual variables and macros / defines can be
+# controlled using \showinitializer or \hideinitializer command in the
+# documentation regardless of this setting.
+# Minimum value: 0, maximum value: 10000, default value: 30.
+
+MAX_INITIALIZER_LINES  = 30
+
+# Set the SHOW_USED_FILES tag to NO to disable the list of files generated at
+# the bottom of the documentation of classes and structs. If set to YES, the
+# list will mention the files that were used to generate the documentation.
+# The default value is: YES.
+
+SHOW_USED_FILES        = YES
+
+# Set the SHOW_FILES tag to NO to disable the generation of the Files page. This
+# will remove the Files entry from the Quick Index and from the Folder Tree View
+# (if specified).
+# The default value is: YES.
+
+SHOW_FILES             = YES
+
+# Set the SHOW_NAMESPACES tag to NO to disable the generation of the Namespaces
+# page. This will remove the Namespaces entry from the Quick Index and from the
+# Folder Tree View (if specified).
+# The default value is: YES.
+
+SHOW_NAMESPACES        = YES
+
+# The FILE_VERSION_FILTER tag can be used to specify a program or script that
+# doxygen should invoke to get the current version for each file (typically from
+# the version control system). Doxygen will invoke the program by executing (via
+# popen()) the command command input-file, where command is the value of the
+# FILE_VERSION_FILTER tag, and input-file is the name of an input file provided
+# by doxygen. Whatever the program writes to standard output is used as the file
+# version. For an example see the documentation.
+
+FILE_VERSION_FILTER    =
+
+# The LAYOUT_FILE tag can be used to specify a layout file which will be parsed
+# by doxygen. The layout file controls the global structure of the generated
+# output files in an output format independent way. To create the layout file
+# that represents doxygen's defaults, run doxygen with the -l option. You can
+# optionally specify a file name after the option, if omitted DoxygenLayout.xml
+# will be used as the name of the layout file.
+#
+# Note that if you run doxygen from a directory containing a file called
+# DoxygenLayout.xml, doxygen will parse it automatically even if the LAYOUT_FILE
+# tag is left empty.
+
+LAYOUT_FILE            =
+
+# The CITE_BIB_FILES tag can be used to specify one or more bib files containing
+# the reference definitions. This must be a list of .bib files. The .bib
+# extension is automatically appended if omitted. This requires the bibtex tool
+# to be installed. See also https://en.wikipedia.org/wiki/BibTeX for more info.
+# For LaTeX the style of the bibliography can be controlled using
+# LATEX_BIB_STYLE. To use this feature you need bibtex and perl available in the
+# search path. See also \cite for info how to create references.
+
+CITE_BIB_FILES         =
+
+#---------------------------------------------------------------------------
+# Configuration options related to warning and progress messages
+#---------------------------------------------------------------------------
+
+# The QUIET tag can be used to turn on/off the messages that are generated to
+# standard output by doxygen. If QUIET is set to YES this implies that the
+# messages are off.
+# The default value is: NO.
+
+QUIET                  = NO
+
+# The WARNINGS tag can be used to turn on/off the warning messages that are
+# generated to standard error (stderr) by doxygen. If WARNINGS is set to YES
+# this implies that the warnings are on.
+#
+# Tip: Turn warnings on while writing the documentation.
+# The default value is: YES.
+
+WARNINGS               = YES
+
+# If the WARN_IF_UNDOCUMENTED tag is set to YES then doxygen will generate
+# warnings for undocumented members. If EXTRACT_ALL is set to YES then this flag
+# will automatically be disabled.
+# The default value is: YES.
+
+WARN_IF_UNDOCUMENTED   = YES
+
+# If the WARN_IF_DOC_ERROR tag is set to YES, doxygen will generate warnings for
+# potential errors in the documentation, such as not documenting some parameters
+# in a documented function, or documenting parameters that don't exist or using
+# markup commands wrongly.
+# The default value is: YES.
+
+WARN_IF_DOC_ERROR      = YES
+
+# This WARN_NO_PARAMDOC option can be enabled to get warnings for functions that
+# are documented, but have no documentation for their parameters or return
+# value. If set to NO, doxygen will only warn about wrong or incomplete
+# parameter documentation, but not about the absence of documentation. If
+# EXTRACT_ALL is set to YES then this flag will automatically be disabled.
+# The default value is: NO.
+
+WARN_NO_PARAMDOC       = NO
+
+# If the WARN_AS_ERROR tag is set to YES then doxygen will immediately stop when
+# a warning is encountered.
+# The default value is: NO.
+
+WARN_AS_ERROR          = NO
+
+# The WARN_FORMAT tag determines the format of the warning messages that doxygen
+# can produce. The string should contain the $file, $line, and $text tags, which
+# will be replaced by the file and line number from which the warning originated
+# and the warning text. Optionally the format may contain $version, which will
+# be replaced by the version of the file (if it could be obtained via
+# FILE_VERSION_FILTER)
+# The default value is: $file:$line: $text.
+
+WARN_FORMAT            = "$file:$line: $text"
+
+# The WARN_LOGFILE tag can be used to specify a file to which warning and error
+# messages should be written. If left blank the output is written to standard
+# error (stderr).
+
+WARN_LOGFILE           =
+
+#---------------------------------------------------------------------------
+# Configuration options related to the input files
+#---------------------------------------------------------------------------
+
+# The INPUT tag is used to specify the files and/or directories that contain
+# documented source files. You may enter file names like myfile.cpp or
+# directories like /usr/src/myproject. Separate the files or directories with
+# spaces. See also FILE_PATTERNS and EXTENSION_MAPPING
+# Note: If this tag is empty the current directory is searched.
+
+INPUT                  =
+
+# This tag can be used to specify the character encoding of the source files
+# that doxygen parses. Internally doxygen uses the UTF-8 encoding. Doxygen uses
+# libiconv (or the iconv built into libc) for the transcoding. See the libiconv
+# documentation (see: https://www.gnu.org/software/libiconv/) for the list of
+# possible encodings.
+# The default value is: UTF-8.
+
+INPUT_ENCODING         = UTF-8
+
+# If the value of the INPUT tag contains directories, you can use the
+# FILE_PATTERNS tag to specify one or more wildcard patterns (like *.cpp and
+# *.h) to filter out the source-files in the directories.
+#
+# Note that for custom extensions or not directly supported extensions you also
+# need to set EXTENSION_MAPPING for the extension otherwise the files are not
+# read by doxygen.
+#
+# If left blank the following patterns are tested:*.c, *.cc, *.cxx, *.cpp,
+# *.c++, *.java, *.ii, *.ixx, *.ipp, *.i++, *.inl, *.idl, *.ddl, *.odl, *.h,
+# *.hh, *.hxx, *.hpp, *.h++, *.cs, *.d, *.php, *.php4, *.php5, *.phtml, *.inc,
+# *.m, *.markdown, *.md, *.mm, *.dox (to be provided as doxygen C comment),
+# *.doc (to be provided as doxygen C comment), *.txt (to be provided as doxygen
+# C comment), *.py, *.pyw, *.f90, *.f95, *.f03, *.f08, *.f, *.for, *.tcl, *.vhd,
+# *.vhdl, *.ucf, *.qsf and *.ice.
+
+FILE_PATTERNS          = *.c \
+                         *.cc \
+                         *.cxx \
+                         *.cpp \
+                         *.c++ \
+                         *.java \
+                         *.ii \
+                         *.ixx \
+                         *.ipp \
+                         *.i++ \
+                         *.inl \
+                         *.idl \
+                         *.ddl \
+                         *.odl \
+                         *.h \
+                         *.hh \
+                         *.hxx \
+                         *.hpp \
+                         *.h++ \
+                         *.cs \
+                         *.d \
+                         *.php \
+                         *.php4 \
+                         *.php5 \
+                         *.phtml \
+                         *.inc \
+                         *.m \
+                         *.markdown \
+                         *.md \
+                         *.mm \
+                         *.dox \
+                         *.doc \
+                         *.txt \
+                         *.py \
+                         *.pyw \
+                         *.f90 \
+                         *.f95 \
+                         *.f03 \
+                         *.f08 \
+                         *.f \
+                         *.for \
+                         *.tcl \
+                         *.vhd \
+                         *.vhdl \
+                         *.ucf \
+                         *.qsf \
+                         *.ice
+
+# The RECURSIVE tag can be used to specify whether or not subdirectories should
+# be searched for input files as well.
+# The default value is: NO.
+
+RECURSIVE              = YES
+
+# The EXCLUDE tag can be used to specify files and/or directories that should be
+# excluded from the INPUT source files. This way you can easily exclude a
+# subdirectory from a directory tree whose root is specified with the INPUT tag.
+#
+# Note that relative paths are relative to the directory from which doxygen is
+# run.
+
+EXCLUDE                =
+
+# The EXCLUDE_SYMLINKS tag can be used to select whether or not files or
+# directories that are symbolic links (a Unix file system feature) are excluded
+# from the input.
+# The default value is: NO.
+
+EXCLUDE_SYMLINKS       = NO
+
+# If the value of the INPUT tag contains directories, you can use the
+# EXCLUDE_PATTERNS tag to specify one or more wildcard patterns to exclude
+# certain files from those directories.
+#
+# Note that the wildcards are matched against the file with absolute path, so to
+# exclude all test directories for example use the pattern */test/*
+
+EXCLUDE_PATTERNS       =
+
+# The EXCLUDE_SYMBOLS tag can be used to specify one or more symbol names
+# (namespaces, classes, functions, etc.) that should be excluded from the
+# output. The symbol name can be a fully qualified name, a word, or if the
+# wildcard * is used, a substring. Examples: ANamespace, AClass,
+# AClass::ANamespace, ANamespace::*Test
+#
+# Note that the wildcards are matched against the file with absolute path, so to
+# exclude all test directories use the pattern */test/*
+
+EXCLUDE_SYMBOLS        =
+
+# The EXAMPLE_PATH tag can be used to specify one or more files or directories
+# that contain example code fragments that are included (see the \include
+# command).
+
+EXAMPLE_PATH           =
+
+# If the value of the EXAMPLE_PATH tag contains directories, you can use the
+# EXAMPLE_PATTERNS tag to specify one or more wildcard pattern (like *.cpp and
+# *.h) to filter out the source-files in the directories. If left blank all
+# files are included.
+
+EXAMPLE_PATTERNS       = *
+
+# If the EXAMPLE_RECURSIVE tag is set to YES then subdirectories will be
+# searched for input files to be used with the \include or \dontinclude commands
+# irrespective of the value of the RECURSIVE tag.
+# The default value is: NO.
+
+EXAMPLE_RECURSIVE      = NO
+
+# The IMAGE_PATH tag can be used to specify one or more files or directories
+# that contain images that are to be included in the documentation (see the
+# \image command).
+
+IMAGE_PATH             =
+
+# The INPUT_FILTER tag can be used to specify a program that doxygen should
+# invoke to filter for each input file. Doxygen will invoke the filter program
+# by executing (via popen()) the command:
+#
+# <filter> <input-file>
+#
+# where <filter> is the value of the INPUT_FILTER tag, and <input-file> is the
+# name of an input file. Doxygen will then use the output that the filter
+# program writes to standard output. If FILTER_PATTERNS is specified, this tag
+# will be ignored.
+#
+# Note that the filter must not add or remove lines; it is applied before the
+# code is scanned, but not when the output code is generated. If lines are added
+# or removed, the anchors will not be placed correctly.
+#
+# Note that for custom extensions or not directly supported extensions you also
+# need to set EXTENSION_MAPPING for the extension otherwise the files are not
+# properly processed by doxygen.
+
+INPUT_FILTER           =
+
+# The FILTER_PATTERNS tag can be used to specify filters on a per file pattern
+# basis. Doxygen will compare the file name with each pattern and apply the
+# filter if there is a match. The filters are a list of the form: pattern=filter
+# (like *.cpp=my_cpp_filter). See INPUT_FILTER for further information on how
+# filters are used. If the FILTER_PATTERNS tag is empty or if none of the
+# patterns match the file name, INPUT_FILTER is applied.
+#
+# Note that for custom extensions or not directly supported extensions you also
+# need to set EXTENSION_MAPPING for the extension otherwise the files are not
+# properly processed by doxygen.
+
+FILTER_PATTERNS        =
+
+# If the FILTER_SOURCE_FILES tag is set to YES, the input filter (if set using
+# INPUT_FILTER) will also be used to filter the input files that are used for
+# producing the source files to browse (i.e. when SOURCE_BROWSER is set to YES).
+# The default value is: NO.
+
+FILTER_SOURCE_FILES    = NO
+
+# The FILTER_SOURCE_PATTERNS tag can be used to specify source filters per file
+# pattern. A pattern will override the setting for FILTER_PATTERN (if any) and
+# it is also possible to disable source filtering for a specific pattern using
+# *.ext= (so without naming a filter).
+# This tag requires that the tag FILTER_SOURCE_FILES is set to YES.
+
+FILTER_SOURCE_PATTERNS =
+
+# If the USE_MDFILE_AS_MAINPAGE tag refers to the name of a markdown file that
+# is part of the input, its contents will be placed on the main page
+# (index.html). This can be useful if you have a project on for instance GitHub
+# and want to reuse the introduction page also for the doxygen output.
+
+USE_MDFILE_AS_MAINPAGE =
+
+#---------------------------------------------------------------------------
+# Configuration options related to source browsing
+#---------------------------------------------------------------------------
+
+# If the SOURCE_BROWSER tag is set to YES then a list of source files will be
+# generated. Documented entities will be cross-referenced with these sources.
+#
+# Note: To get rid of all source code in the generated output, make sure that
+# also VERBATIM_HEADERS is set to NO.
+# The default value is: NO.
+
+SOURCE_BROWSER         = NO
+
+# Setting the INLINE_SOURCES tag to YES will include the body of functions,
+# classes and enums directly into the documentation.
+# The default value is: NO.
+
+INLINE_SOURCES         = NO
+
+# Setting the STRIP_CODE_COMMENTS tag to YES will instruct doxygen to hide any
+# special comment blocks from generated source code fragments. Normal C, C++ and
+# Fortran comments will always remain visible.
+# The default value is: YES.
+
+STRIP_CODE_COMMENTS    = YES
+
+# If the REFERENCED_BY_RELATION tag is set to YES then for each documented
+# entity all documented functions referencing it will be listed.
+# The default value is: NO.
+
+REFERENCED_BY_RELATION = NO
+
+# If the REFERENCES_RELATION tag is set to YES then for each documented function
+# all documented entities called/used by that function will be listed.
+# The default value is: NO.
+
+REFERENCES_RELATION    = NO
+
+# If the REFERENCES_LINK_SOURCE tag is set to YES and SOURCE_BROWSER tag is set
+# to YES then the hyperlinks from functions in REFERENCES_RELATION and
+# REFERENCED_BY_RELATION lists will link to the source code. Otherwise they will
+# link to the documentation.
+# The default value is: YES.
+
+REFERENCES_LINK_SOURCE = YES
+
+# If SOURCE_TOOLTIPS is enabled (the default) then hovering a hyperlink in the
+# source code will show a tooltip with additional information such as prototype,
+# brief description and links to the definition and documentation. Since this
+# will make the HTML file larger and loading of large files a bit slower, you
+# can opt to disable this feature.
+# The default value is: YES.
+# This tag requires that the tag SOURCE_BROWSER is set to YES.
+
+SOURCE_TOOLTIPS        = YES
+
+# If the USE_HTAGS tag is set to YES then the references to source code will
+# point to the HTML generated by the htags(1) tool instead of doxygen built-in
+# source browser. The htags tool is part of GNU's global source tagging system
+# (see https://www.gnu.org/software/global/global.html). You will need version
+# 4.8.6 or higher.
+#
+# To use it do the following:
+# - Install the latest version of global
+# - Enable SOURCE_BROWSER and USE_HTAGS in the configuration file
+# - Make sure the INPUT points to the root of the source tree
+# - Run doxygen as normal
+#
+# Doxygen will invoke htags (and that will in turn invoke gtags), so these
+# tools must be available from the command line (i.e. in the search path).
+#
+# The result: instead of the source browser generated by doxygen, the links to
+# source code will now point to the output of htags.
+# The default value is: NO.
+# This tag requires that the tag SOURCE_BROWSER is set to YES.
+
+USE_HTAGS              = NO
+
+# If the VERBATIM_HEADERS tag is set the YES then doxygen will generate a
+# verbatim copy of the header file for each class for which an include is
+# specified. Set to NO to disable this.
+# See also: Section \class.
+# The default value is: YES.
+
+VERBATIM_HEADERS       = YES
+
+# If the CLANG_ASSISTED_PARSING tag is set to YES then doxygen will use the
+# clang parser (see: http://clang.llvm.org/) for more accurate parsing at the
+# cost of reduced performance. This can be particularly helpful with template
+# rich C++ code for which doxygen's built-in parser lacks the necessary type
+# information.
+# Note: The availability of this option depends on whether or not doxygen was
+# generated with the -Duse_libclang=ON option for CMake.
+# The default value is: NO.
+
+CLANG_ASSISTED_PARSING = NO
+
+# If clang assisted parsing is enabled you can provide the compiler with command
+# line options that you would normally use when invoking the compiler. Note that
+# the include paths will already be set by doxygen for the files and directories
+# specified with INPUT and INCLUDE_PATH.
+# This tag requires that the tag CLANG_ASSISTED_PARSING is set to YES.
+
+CLANG_OPTIONS          =
+
+# If clang assisted parsing is enabled you can provide the clang parser with the
+# path to the compilation database (see:
+# http://clang.llvm.org/docs/HowToSetupToolingForLLVM.html) used when the files
+# were built. This is equivalent to specifying the "-p" option to a clang tool,
+# such as clang-check. These options will then be passed to the parser.
+# Note: The availability of this option depends on whether or not doxygen was
+# generated with the -Duse_libclang=ON option for CMake.
+
+CLANG_DATABASE_PATH    =
+
+#---------------------------------------------------------------------------
+# Configuration options related to the alphabetical class index
+#---------------------------------------------------------------------------
+
+# If the ALPHABETICAL_INDEX tag is set to YES, an alphabetical index of all
+# compounds will be generated. Enable this if the project contains a lot of
+# classes, structs, unions or interfaces.
+# The default value is: YES.
+
+ALPHABETICAL_INDEX     = YES
+
+# The COLS_IN_ALPHA_INDEX tag can be used to specify the number of columns in
+# which the alphabetical index list will be split.
+# Minimum value: 1, maximum value: 20, default value: 5.
+# This tag requires that the tag ALPHABETICAL_INDEX is set to YES.
+
+COLS_IN_ALPHA_INDEX    = 5
+
+# In case all classes in a project start with a common prefix, all classes will
+# be put under the same header in the alphabetical index. The IGNORE_PREFIX tag
+# can be used to specify a prefix (or a list of prefixes) that should be ignored
+# while generating the index headers.
+# This tag requires that the tag ALPHABETICAL_INDEX is set to YES.
+
+IGNORE_PREFIX          =
+
+#---------------------------------------------------------------------------
+# Configuration options related to the HTML output
+#---------------------------------------------------------------------------
+
+# If the GENERATE_HTML tag is set to YES, doxygen will generate HTML output
+# The default value is: YES.
+
+GENERATE_HTML          = YES
+
+# The HTML_OUTPUT tag is used to specify where the HTML docs will be put. If a
+# relative path is entered the value of OUTPUT_DIRECTORY will be put in front of
+# it.
+# The default directory is: html.
+# This tag requires that the tag GENERATE_HTML is set to YES.
+
+HTML_OUTPUT            = html
+
+# The HTML_FILE_EXTENSION tag can be used to specify the file extension for each
+# generated HTML page (for example: .htm, .php, .asp).
+# The default value is: .html.
+# This tag requires that the tag GENERATE_HTML is set to YES.
+
+HTML_FILE_EXTENSION    = .html
+
+# The HTML_HEADER tag can be used to specify a user-defined HTML header file for
+# each generated HTML page. If the tag is left blank doxygen will generate a
+# standard header.
+#
+# To get valid HTML the header file that includes any scripts and style sheets
+# that doxygen needs, which is dependent on the configuration options used (e.g.
+# the setting GENERATE_TREEVIEW). It is highly recommended to start with a
+# default header using
+# doxygen -w html new_header.html new_footer.html new_stylesheet.css
+# YourConfigFile
+# and then modify the file new_header.html. See also section "Doxygen usage"
+# for information on how to generate the default header that doxygen normally
+# uses.
+# Note: The header is subject to change so you typically have to regenerate the
+# default header when upgrading to a newer version of doxygen. For a description
+# of the possible markers and block names see the documentation.
+# This tag requires that the tag GENERATE_HTML is set to YES.
+
+HTML_HEADER            =
+
+# The HTML_FOOTER tag can be used to specify a user-defined HTML footer for each
+# generated HTML page. If the tag is left blank doxygen will generate a standard
+# footer. See HTML_HEADER for more information on how to generate a default
+# footer and what special commands can be used inside the footer. See also
+# section "Doxygen usage" for information on how to generate the default footer
+# that doxygen normally uses.
+# This tag requires that the tag GENERATE_HTML is set to YES.
+
+HTML_FOOTER            =
+
+# The HTML_STYLESHEET tag can be used to specify a user-defined cascading style
+# sheet that is used by each HTML page. It can be used to fine-tune the look of
+# the HTML output. If left blank doxygen will generate a default style sheet.
+# See also section "Doxygen usage" for information on how to generate the style
+# sheet that doxygen normally uses.
+# Note: It is recommended to use HTML_EXTRA_STYLESHEET instead of this tag, as
+# it is more robust and this tag (HTML_STYLESHEET) will in the future become
+# obsolete.
+# This tag requires that the tag GENERATE_HTML is set to YES.
+
+HTML_STYLESHEET        =
+
+# The HTML_EXTRA_STYLESHEET tag can be used to specify additional user-defined
+# cascading style sheets that are included after the standard style sheets
+# created by doxygen. Using this option one can overrule certain style aspects.
+# This is preferred over using HTML_STYLESHEET since it does not replace the
+# standard style sheet and is therefore more robust against future updates.
+# Doxygen will copy the style sheet files to the output directory.
+# Note: The order of the extra style sheet files is of importance (e.g. the last
+# style sheet in the list overrules the setting of the previous ones in the
+# list). For an example see the documentation.
+# This tag requires that the tag GENERATE_HTML is set to YES.
+
+HTML_EXTRA_STYLESHEET  =
+
+# The HTML_EXTRA_FILES tag can be used to specify one or more extra images or
+# other source files which should be copied to the HTML output directory. Note
+# that these files will be copied to the base HTML output directory. Use the
+# $relpath^ marker in the HTML_HEADER and/or HTML_FOOTER files to load these
+# files. In the HTML_STYLESHEET file, use the file name only. Also note that the
+# files will be copied as-is; there are no commands or markers available.
+# This tag requires that the tag GENERATE_HTML is set to YES.
+
+HTML_EXTRA_FILES       =
+
+# The HTML_COLORSTYLE_HUE tag controls the color of the HTML output. Doxygen
+# will adjust the colors in the style sheet and background images according to
+# this color. Hue is specified as an angle on a colorwheel, see
+# https://en.wikipedia.org/wiki/Hue for more information. For instance the value
+# 0 represents red, 60 is yellow, 120 is green, 180 is cyan, 240 is blue, 300
+# purple, and 360 is red again.
+# Minimum value: 0, maximum value: 359, default value: 220.
+# This tag requires that the tag GENERATE_HTML is set to YES.
+
+HTML_COLORSTYLE_HUE    = 220
+
+# The HTML_COLORSTYLE_SAT tag controls the purity (or saturation) of the colors
+# in the HTML output. For a value of 0 the output will use grayscales only. A
+# value of 255 will produce the most vivid colors.
+# Minimum value: 0, maximum value: 255, default value: 100.
+# This tag requires that the tag GENERATE_HTML is set to YES.
+
+HTML_COLORSTYLE_SAT    = 100
+
+# The HTML_COLORSTYLE_GAMMA tag controls the gamma correction applied to the
+# luminance component of the colors in the HTML output. Values below 100
+# gradually make the output lighter, whereas values above 100 make the output
+# darker. The value divided by 100 is the actual gamma applied, so 80 represents
+# a gamma of 0.8, The value 220 represents a gamma of 2.2, and 100 does not
+# change the gamma.
+# Minimum value: 40, maximum value: 240, default value: 80.
+# This tag requires that the tag GENERATE_HTML is set to YES.
+
+HTML_COLORSTYLE_GAMMA  = 80
+
+# If the HTML_TIMESTAMP tag is set to YES then the footer of each generated HTML
+# page will contain the date and time when the page was generated. Setting this
+# to YES can help to show when doxygen was last run and thus if the
+# documentation is up to date.
+# The default value is: NO.
+# This tag requires that the tag GENERATE_HTML is set to YES.
+
+HTML_TIMESTAMP         = NO
+
+# If the HTML_DYNAMIC_MENUS tag is set to YES then the generated HTML
+# documentation will contain a main index with vertical navigation menus that
+# are dynamically created via JavaScript. If disabled, the navigation index will
+# consists of multiple levels of tabs that are statically embedded in every HTML
+# page. Disable this option to support browsers that do not have JavaScript,
+# like the Qt help browser.
+# The default value is: YES.
+# This tag requires that the tag GENERATE_HTML is set to YES.
+
+HTML_DYNAMIC_MENUS     = YES
+
+# If the HTML_DYNAMIC_SECTIONS tag is set to YES then the generated HTML
+# documentation will contain sections that can be hidden and shown after the
+# page has loaded.
+# The default value is: NO.
+# This tag requires that the tag GENERATE_HTML is set to YES.
+
+HTML_DYNAMIC_SECTIONS  = NO
+
+# With HTML_INDEX_NUM_ENTRIES one can control the preferred number of entries
+# shown in the various tree structured indices initially; the user can expand
+# and collapse entries dynamically later on. Doxygen will expand the tree to
+# such a level that at most the specified number of entries are visible (unless
+# a fully collapsed tree already exceeds this amount). So setting the number of
+# entries 1 will produce a full collapsed tree by default. 0 is a special value
+# representing an infinite number of entries and will result in a full expanded
+# tree by default.
+# Minimum value: 0, maximum value: 9999, default value: 100.
+# This tag requires that the tag GENERATE_HTML is set to YES.
+
+HTML_INDEX_NUM_ENTRIES = 100
+
+# If the GENERATE_DOCSET tag is set to YES, additional index files will be
+# generated that can be used as input for Apple's Xcode 3 integrated development
+# environment (see: https://developer.apple.com/xcode/), introduced with OSX
+# 10.5 (Leopard). To create a documentation set, doxygen will generate a
+# Makefile in the HTML output directory. Running make will produce the docset in
+# that directory and running make install will install the docset in
+# ~/Library/Developer/Shared/Documentation/DocSets so that Xcode will find it at
+# startup. See https://developer.apple.com/library/archive/featuredarticles/Doxy
+# genXcode/_index.html for more information.
+# The default value is: NO.
+# This tag requires that the tag GENERATE_HTML is set to YES.
+
+GENERATE_DOCSET        = NO
+
+# This tag determines the name of the docset feed. A documentation feed provides
+# an umbrella under which multiple documentation sets from a single provider
+# (such as a company or product suite) can be grouped.
+# The default value is: Doxygen generated docs.
+# This tag requires that the tag GENERATE_DOCSET is set to YES.
+
+DOCSET_FEEDNAME        = "Doxygen generated docs"
+
+# This tag specifies a string that should uniquely identify the documentation
+# set bundle. This should be a reverse domain-name style string, e.g.
+# com.mycompany.MyDocSet. Doxygen will append .docset to the name.
+# The default value is: org.doxygen.Project.
+# This tag requires that the tag GENERATE_DOCSET is set to YES.
+
+DOCSET_BUNDLE_ID       = org.doxygen.Project
+
+# The DOCSET_PUBLISHER_ID tag specifies a string that should uniquely identify
+# the documentation publisher. This should be a reverse domain-name style
+# string, e.g. com.mycompany.MyDocSet.documentation.
+# The default value is: org.doxygen.Publisher.
+# This tag requires that the tag GENERATE_DOCSET is set to YES.
+
+DOCSET_PUBLISHER_ID    = org.doxygen.Publisher
+
+# The DOCSET_PUBLISHER_NAME tag identifies the documentation publisher.
+# The default value is: Publisher.
+# This tag requires that the tag GENERATE_DOCSET is set to YES.
+
+DOCSET_PUBLISHER_NAME  = Publisher
+
+# If the GENERATE_HTMLHELP tag is set to YES then doxygen generates three
+# additional HTML index files: index.hhp, index.hhc, and index.hhk. The
+# index.hhp is a project file that can be read by Microsoft's HTML Help Workshop
+# (see: https://www.microsoft.com/en-us/download/details.aspx?id=21138) on
+# Windows.
+#
+# The HTML Help Workshop contains a compiler that can convert all HTML output
+# generated by doxygen into a single compiled HTML file (.chm). Compiled HTML
+# files are now used as the Windows 98 help format, and will replace the old
+# Windows help format (.hlp) on all Windows platforms in the future. Compressed
+# HTML files also contain an index, a table of contents, and you can search for
+# words in the documentation. The HTML workshop also contains a viewer for
+# compressed HTML files.
+# The default value is: NO.
+# This tag requires that the tag GENERATE_HTML is set to YES.
+
+GENERATE_HTMLHELP      = NO
+
+# The CHM_FILE tag can be used to specify the file name of the resulting .chm
+# file. You can add a path in front of the file if the result should not be
+# written to the html output directory.
+# This tag requires that the tag GENERATE_HTMLHELP is set to YES.
+
+CHM_FILE               =
+
+# The HHC_LOCATION tag can be used to specify the location (absolute path
+# including file name) of the HTML help compiler (hhc.exe). If non-empty,
+# doxygen will try to run the HTML help compiler on the generated index.hhp.
+# The file has to be specified with full path.
+# This tag requires that the tag GENERATE_HTMLHELP is set to YES.
+
+HHC_LOCATION           =
+
+# The GENERATE_CHI flag controls if a separate .chi index file is generated
+# (YES) or that it should be included in the master .chm file (NO).
+# The default value is: NO.
+# This tag requires that the tag GENERATE_HTMLHELP is set to YES.
+
+GENERATE_CHI           = NO
+
+# The CHM_INDEX_ENCODING is used to encode HtmlHelp index (hhk), content (hhc)
+# and project file content.
+# This tag requires that the tag GENERATE_HTMLHELP is set to YES.
+
+CHM_INDEX_ENCODING     =
+
+# The BINARY_TOC flag controls whether a binary table of contents is generated
+# (YES) or a normal table of contents (NO) in the .chm file. Furthermore it
+# enables the Previous and Next buttons.
+# The default value is: NO.
+# This tag requires that the tag GENERATE_HTMLHELP is set to YES.
+
+BINARY_TOC             = NO
+
+# The TOC_EXPAND flag can be set to YES to add extra items for group members to
+# the table of contents of the HTML help documentation and to the tree view.
+# The default value is: NO.
+# This tag requires that the tag GENERATE_HTMLHELP is set to YES.
+
+TOC_EXPAND             = NO
+
+# If the GENERATE_QHP tag is set to YES and both QHP_NAMESPACE and
+# QHP_VIRTUAL_FOLDER are set, an additional index file will be generated that
+# can be used as input for Qt's qhelpgenerator to generate a Qt Compressed Help
+# (.qch) of the generated HTML documentation.
+# The default value is: NO.
+# This tag requires that the tag GENERATE_HTML is set to YES.
+
+GENERATE_QHP           = NO
+
+# If the QHG_LOCATION tag is specified, the QCH_FILE tag can be used to specify
+# the file name of the resulting .qch file. The path specified is relative to
+# the HTML output folder.
+# This tag requires that the tag GENERATE_QHP is set to YES.
+
+QCH_FILE               =
+
+# The QHP_NAMESPACE tag specifies the namespace to use when generating Qt Help
+# Project output. For more information please see Qt Help Project / Namespace
+# (see: https://doc.qt.io/archives/qt-4.8/qthelpproject.html#namespace).
+# The default value is: org.doxygen.Project.
+# This tag requires that the tag GENERATE_QHP is set to YES.
+
+QHP_NAMESPACE          = org.doxygen.Project
+
+# The QHP_VIRTUAL_FOLDER tag specifies the namespace to use when generating Qt
+# Help Project output. For more information please see Qt Help Project / Virtual
+# Folders (see: https://doc.qt.io/archives/qt-4.8/qthelpproject.html#virtual-
+# folders).
+# The default value is: doc.
+# This tag requires that the tag GENERATE_QHP is set to YES.
+
+QHP_VIRTUAL_FOLDER     = doc
+
+# If the QHP_CUST_FILTER_NAME tag is set, it specifies the name of a custom
+# filter to add. For more information please see Qt Help Project / Custom
+# Filters (see: https://doc.qt.io/archives/qt-4.8/qthelpproject.html#custom-
+# filters).
+# This tag requires that the tag GENERATE_QHP is set to YES.
+
+QHP_CUST_FILTER_NAME   =
+
+# The QHP_CUST_FILTER_ATTRS tag specifies the list of the attributes of the
+# custom filter to add. For more information please see Qt Help Project / Custom
+# Filters (see: https://doc.qt.io/archives/qt-4.8/qthelpproject.html#custom-
+# filters).
+# This tag requires that the tag GENERATE_QHP is set to YES.
+
+QHP_CUST_FILTER_ATTRS  =
+
+# The QHP_SECT_FILTER_ATTRS tag specifies the list of the attributes this
+# project's filter section matches. Qt Help Project / Filter Attributes (see:
+# https://doc.qt.io/archives/qt-4.8/qthelpproject.html#filter-attributes).
+# This tag requires that the tag GENERATE_QHP is set to YES.
+
+QHP_SECT_FILTER_ATTRS  =
+
+# The QHG_LOCATION tag can be used to specify the location of Qt's
+# qhelpgenerator. If non-empty doxygen will try to run qhelpgenerator on the
+# generated .qhp file.
+# This tag requires that the tag GENERATE_QHP is set to YES.
+
+QHG_LOCATION           =
+
+# If the GENERATE_ECLIPSEHELP tag is set to YES, additional index files will be
+# generated, together with the HTML files, they form an Eclipse help plugin. To
+# install this plugin and make it available under the help contents menu in
+# Eclipse, the contents of the directory containing the HTML and XML files needs
+# to be copied into the plugins directory of eclipse. The name of the directory
+# within the plugins directory should be the same as the ECLIPSE_DOC_ID value.
+# After copying Eclipse needs to be restarted before the help appears.
+# The default value is: NO.
+# This tag requires that the tag GENERATE_HTML is set to YES.
+
+GENERATE_ECLIPSEHELP   = NO
+
+# A unique identifier for the Eclipse help plugin. When installing the plugin
+# the directory name containing the HTML and XML files should also have this
+# name. Each documentation set should have its own identifier.
+# The default value is: org.doxygen.Project.
+# This tag requires that the tag GENERATE_ECLIPSEHELP is set to YES.
+
+ECLIPSE_DOC_ID         = org.doxygen.Project
+
+# If you want full control over the layout of the generated HTML pages it might
+# be necessary to disable the index and replace it with your own. The
+# DISABLE_INDEX tag can be used to turn on/off the condensed index (tabs) at top
+# of each HTML page. A value of NO enables the index and the value YES disables
+# it. Since the tabs in the index contain the same information as the navigation
+# tree, you can set this option to YES if you also set GENERATE_TREEVIEW to YES.
+# The default value is: NO.
+# This tag requires that the tag GENERATE_HTML is set to YES.
+
+DISABLE_INDEX          = NO
+
+# The GENERATE_TREEVIEW tag is used to specify whether a tree-like index
+# structure should be generated to display hierarchical information. If the tag
+# value is set to YES, a side panel will be generated containing a tree-like
+# index structure (just like the one that is generated for HTML Help). For this
+# to work a browser that supports JavaScript, DHTML, CSS and frames is required
+# (i.e. any modern browser). Windows users are probably better off using the
+# HTML help feature. Via custom style sheets (see HTML_EXTRA_STYLESHEET) one can
+# further fine-tune the look of the index. As an example, the default style
+# sheet generated by doxygen has an example that shows how to put an image at
+# the root of the tree instead of the PROJECT_NAME. Since the tree basically has
+# the same information as the tab index, you could consider setting
+# DISABLE_INDEX to YES when enabling this option.
+# The default value is: NO.
+# This tag requires that the tag GENERATE_HTML is set to YES.
+
+GENERATE_TREEVIEW      = NO
+
+# The ENUM_VALUES_PER_LINE tag can be used to set the number of enum values that
+# doxygen will group on one line in the generated HTML documentation.
+#
+# Note that a value of 0 will completely suppress the enum values from appearing
+# in the overview section.
+# Minimum value: 0, maximum value: 20, default value: 4.
+# This tag requires that the tag GENERATE_HTML is set to YES.
+
+ENUM_VALUES_PER_LINE   = 4
+
+# If the treeview is enabled (see GENERATE_TREEVIEW) then this tag can be used
+# to set the initial width (in pixels) of the frame in which the tree is shown.
+# Minimum value: 0, maximum value: 1500, default value: 250.
+# This tag requires that the tag GENERATE_HTML is set to YES.
+
+TREEVIEW_WIDTH         = 250
+
+# If the EXT_LINKS_IN_WINDOW option is set to YES, doxygen will open links to
+# external symbols imported via tag files in a separate window.
+# The default value is: NO.
+# This tag requires that the tag GENERATE_HTML is set to YES.
+
+EXT_LINKS_IN_WINDOW    = NO
+
+# Use this tag to change the font size of LaTeX formulas included as images in
+# the HTML documentation. When you change the font size after a successful
+# doxygen run you need to manually remove any form_*.png images from the HTML
+# output directory to force them to be regenerated.
+# Minimum value: 8, maximum value: 50, default value: 10.
+# This tag requires that the tag GENERATE_HTML is set to YES.
+
+FORMULA_FONTSIZE       = 10
+
+# Use the FORMULA_TRANSPARENT tag to determine whether or not the images
+# generated for formulas are transparent PNGs. Transparent PNGs are not
+# supported properly for IE 6.0, but are supported on all modern browsers.
+#
+# Note that when changing this option you need to delete any form_*.png files in
+# the HTML output directory before the changes have effect.
+# The default value is: YES.
+# This tag requires that the tag GENERATE_HTML is set to YES.
+
+FORMULA_TRANSPARENT    = YES
+
+# The FORMULA_MACROFILE can contain LaTeX \newcommand and \renewcommand commands
+# to create new LaTeX commands to be used in formulas as building blocks. See
+# the section "Including formulas" for details.
+
+FORMULA_MACROFILE      =
+
+# Enable the USE_MATHJAX option to render LaTeX formulas using MathJax (see
+# https://www.mathjax.org) which uses client side JavaScript for the rendering
+# instead of using pre-rendered bitmaps. Use this if you do not have LaTeX
+# installed or if you want to formulas look prettier in the HTML output. When
+# enabled you may also need to install MathJax separately and configure the path
+# to it using the MATHJAX_RELPATH option.
+# The default value is: NO.
+# This tag requires that the tag GENERATE_HTML is set to YES.
+
+USE_MATHJAX            = YES
+
+# When MathJax is enabled you can set the default output format to be used for
+# the MathJax output. See the MathJax site (see:
+# http://docs.mathjax.org/en/latest/output.html) for more details.
+# Possible values are: HTML-CSS (which is slower, but has the best
+# compatibility), NativeMML (i.e. MathML) and SVG.
+# The default value is: HTML-CSS.
+# This tag requires that the tag USE_MATHJAX is set to YES.
+
+MATHJAX_FORMAT         = HTML-CSS
+
+# When MathJax is enabled you need to specify the location relative to the HTML
+# output directory using the MATHJAX_RELPATH option. The destination directory
+# should contain the MathJax.js script. For instance, if the mathjax directory
+# is located at the same level as the HTML output directory, then
+# MATHJAX_RELPATH should be ../mathjax. The default value points to the MathJax
+# Content Delivery Network so you can quickly see the result without installing
+# MathJax. However, it is strongly recommended to install a local copy of
+# MathJax from https://www.mathjax.org before deployment.
+# The default value is: https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.5/.
+# This tag requires that the tag USE_MATHJAX is set to YES.
+
+MATHJAX_RELPATH        = https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.5/
+
+# The MATHJAX_EXTENSIONS tag can be used to specify one or more MathJax
+# extension names that should be enabled during MathJax rendering. For example
+# MATHJAX_EXTENSIONS = TeX/AMSmath TeX/AMSsymbols
+# This tag requires that the tag USE_MATHJAX is set to YES.
+
+MATHJAX_EXTENSIONS     =
+
+# The MATHJAX_CODEFILE tag can be used to specify a file with javascript pieces
+# of code that will be used on startup of the MathJax code. See the MathJax site
+# (see: http://docs.mathjax.org/en/latest/output.html) for more details. For an
+# example see the documentation.
+# This tag requires that the tag USE_MATHJAX is set to YES.
+
+MATHJAX_CODEFILE       =
+
+# When the SEARCHENGINE tag is enabled doxygen will generate a search box for
+# the HTML output. The underlying search engine uses javascript and DHTML and
+# should work on any modern browser. Note that when using HTML help
+# (GENERATE_HTMLHELP), Qt help (GENERATE_QHP), or docsets (GENERATE_DOCSET)
+# there is already a search function so this one should typically be disabled.
+# For large projects the javascript based search engine can be slow, then
+# enabling SERVER_BASED_SEARCH may provide a better solution. It is possible to
+# search using the keyboard; to jump to the search box use <access key> + S
+# (what the <access key> is depends on the OS and browser, but it is typically
+# <CTRL>, <ALT>/<option>, or both). Inside the search box use the <cursor down
+# key> to jump into the search results window, the results can be navigated
+# using the <cursor keys>. Press <Enter> to select an item or <escape> to cancel
+# the search. The filter options can be selected when the cursor is inside the
+# search box by pressing <Shift>+<cursor down>. Also here use the <cursor keys>
+# to select a filter and <Enter> or <escape> to activate or cancel the filter
+# option.
+# The default value is: YES.
+# This tag requires that the tag GENERATE_HTML is set to YES.
+
+SEARCHENGINE           = YES
+
+# When the SERVER_BASED_SEARCH tag is enabled the search engine will be
+# implemented using a web server instead of a web client using JavaScript. There
+# are two flavors of web server based searching depending on the EXTERNAL_SEARCH
+# setting. When disabled, doxygen will generate a PHP script for searching and
+# an index file used by the script. When EXTERNAL_SEARCH is enabled the indexing
+# and searching needs to be provided by external tools. See the section
+# "External Indexing and Searching" for details.
+# The default value is: NO.
+# This tag requires that the tag SEARCHENGINE is set to YES.
+
+SERVER_BASED_SEARCH    = NO
+
+# When EXTERNAL_SEARCH tag is enabled doxygen will no longer generate the PHP
+# script for searching. Instead the search results are written to an XML file
+# which needs to be processed by an external indexer. Doxygen will invoke an
+# external search engine pointed to by the SEARCHENGINE_URL option to obtain the
+# search results.
+#
+# Doxygen ships with an example indexer (doxyindexer) and search engine
+# (doxysearch.cgi) which are based on the open source search engine library
+# Xapian (see: https://xapian.org/).
+#
+# See the section "External Indexing and Searching" for details.
+# The default value is: NO.
+# This tag requires that the tag SEARCHENGINE is set to YES.
+
+EXTERNAL_SEARCH        = NO
+
+# The SEARCHENGINE_URL should point to a search engine hosted by a web server
+# which will return the search results when EXTERNAL_SEARCH is enabled.
+#
+# Doxygen ships with an example indexer (doxyindexer) and search engine
+# (doxysearch.cgi) which are based on the open source search engine library
+# Xapian (see: https://xapian.org/). See the section "External Indexing and
+# Searching" for details.
+# This tag requires that the tag SEARCHENGINE is set to YES.
+
+SEARCHENGINE_URL       =
+
+# When SERVER_BASED_SEARCH and EXTERNAL_SEARCH are both enabled the unindexed
+# search data is written to a file for indexing by an external tool. With the
+# SEARCHDATA_FILE tag the name of this file can be specified.
+# The default file is: searchdata.xml.
+# This tag requires that the tag SEARCHENGINE is set to YES.
+
+SEARCHDATA_FILE        = searchdata.xml
+
+# When SERVER_BASED_SEARCH and EXTERNAL_SEARCH are both enabled the
+# EXTERNAL_SEARCH_ID tag can be used as an identifier for the project. This is
+# useful in combination with EXTRA_SEARCH_MAPPINGS to search through multiple
+# projects and redirect the results back to the right project.
+# This tag requires that the tag SEARCHENGINE is set to YES.
+
+EXTERNAL_SEARCH_ID     =
+
+# The EXTRA_SEARCH_MAPPINGS tag can be used to enable searching through doxygen
+# projects other than the one defined by this configuration file, but that are
+# all added to the same external search index. Each project needs to have a
+# unique id set via EXTERNAL_SEARCH_ID. The search mapping then maps the id of
+# to a relative location where the documentation can be found. The format is:
+# EXTRA_SEARCH_MAPPINGS = tagname1=loc1 tagname2=loc2 ...
+# This tag requires that the tag SEARCHENGINE is set to YES.
+
+EXTRA_SEARCH_MAPPINGS  =
+
+#---------------------------------------------------------------------------
+# Configuration options related to the LaTeX output
+#---------------------------------------------------------------------------
+
+# If the GENERATE_LATEX tag is set to YES, doxygen will generate LaTeX output.
+# The default value is: YES.
+
+GENERATE_LATEX         = YES
+
+# The LATEX_OUTPUT tag is used to specify where the LaTeX docs will be put. If a
+# relative path is entered the value of OUTPUT_DIRECTORY will be put in front of
+# it.
+# The default directory is: latex.
+# This tag requires that the tag GENERATE_LATEX is set to YES.
+
+LATEX_OUTPUT           = latex
+
+# The LATEX_CMD_NAME tag can be used to specify the LaTeX command name to be
+# invoked.
+#
+# Note that when not enabling USE_PDFLATEX the default is latex when enabling
+# USE_PDFLATEX the default is pdflatex and when in the later case latex is
+# chosen this is overwritten by pdflatex. For specific output languages the
+# default can have been set differently, this depends on the implementation of
+# the output language.
+# This tag requires that the tag GENERATE_LATEX is set to YES.
+
+LATEX_CMD_NAME         =
+
+# The MAKEINDEX_CMD_NAME tag can be used to specify the command name to generate
+# index for LaTeX.
+# Note: This tag is used in the Makefile / make.bat.
+# See also: LATEX_MAKEINDEX_CMD for the part in the generated output file
+# (.tex).
+# The default file is: makeindex.
+# This tag requires that the tag GENERATE_LATEX is set to YES.
+
+MAKEINDEX_CMD_NAME     = makeindex
+
+# The LATEX_MAKEINDEX_CMD tag can be used to specify the command name to
+# generate index for LaTeX. In case there is no backslash (\) as first character
+# it will be automatically added in the LaTeX code.
+# Note: This tag is used in the generated output file (.tex).
+# See also: MAKEINDEX_CMD_NAME for the part in the Makefile / make.bat.
+# The default value is: makeindex.
+# This tag requires that the tag GENERATE_LATEX is set to YES.
+
+LATEX_MAKEINDEX_CMD    = makeindex
+
+# If the COMPACT_LATEX tag is set to YES, doxygen generates more compact LaTeX
+# documents. This may be useful for small projects and may help to save some
+# trees in general.
+# The default value is: NO.
+# This tag requires that the tag GENERATE_LATEX is set to YES.
+
+COMPACT_LATEX          = NO
+
+# The PAPER_TYPE tag can be used to set the paper type that is used by the
+# printer.
+# Possible values are: a4 (210 x 297 mm), letter (8.5 x 11 inches), legal (8.5 x
+# 14 inches) and executive (7.25 x 10.5 inches).
+# The default value is: a4.
+# This tag requires that the tag GENERATE_LATEX is set to YES.
+
+PAPER_TYPE             = a4
+
+# The EXTRA_PACKAGES tag can be used to specify one or more LaTeX package names
+# that should be included in the LaTeX output. The package can be specified just
+# by its name or with the correct syntax as to be used with the LaTeX
+# \usepackage command. To get the times font for instance you can specify :
+# EXTRA_PACKAGES=times or EXTRA_PACKAGES={times}
+# To use the option intlimits with the amsmath package you can specify:
+# EXTRA_PACKAGES=[intlimits]{amsmath}
+# If left blank no extra packages will be included.
+# This tag requires that the tag GENERATE_LATEX is set to YES.
+
+EXTRA_PACKAGES         =
+
+# The LATEX_HEADER tag can be used to specify a personal LaTeX header for the
+# generated LaTeX document. The header should contain everything until the first
+# chapter. If it is left blank doxygen will generate a standard header. See
+# section "Doxygen usage" for information on how to let doxygen write the
+# default header to a separate file.
+#
+# Note: Only use a user-defined header if you know what you are doing! The
+# following commands have a special meaning inside the header: $title,
+# $datetime, $date, $doxygenversion, $projectname, $projectnumber,
+# $projectbrief, $projectlogo. Doxygen will replace $title with the empty
+# string, for the replacement values of the other commands the user is referred
+# to HTML_HEADER.
+# This tag requires that the tag GENERATE_LATEX is set to YES.
+
+LATEX_HEADER           =
+
+# The LATEX_FOOTER tag can be used to specify a personal LaTeX footer for the
+# generated LaTeX document. The footer should contain everything after the last
+# chapter. If it is left blank doxygen will generate a standard footer. See
+# LATEX_HEADER for more information on how to generate a default footer and what
+# special commands can be used inside the footer.
+#
+# Note: Only use a user-defined footer if you know what you are doing!
+# This tag requires that the tag GENERATE_LATEX is set to YES.
+
+LATEX_FOOTER           =
+
+# The LATEX_EXTRA_STYLESHEET tag can be used to specify additional user-defined
+# LaTeX style sheets that are included after the standard style sheets created
+# by doxygen. Using this option one can overrule certain style aspects. Doxygen
+# will copy the style sheet files to the output directory.
+# Note: The order of the extra style sheet files is of importance (e.g. the last
+# style sheet in the list overrules the setting of the previous ones in the
+# list).
+# This tag requires that the tag GENERATE_LATEX is set to YES.
+
+LATEX_EXTRA_STYLESHEET =
+
+# The LATEX_EXTRA_FILES tag can be used to specify one or more extra images or
+# other source files which should be copied to the LATEX_OUTPUT output
+# directory. Note that the files will be copied as-is; there are no commands or
+# markers available.
+# This tag requires that the tag GENERATE_LATEX is set to YES.
+
+LATEX_EXTRA_FILES      =
+
+# If the PDF_HYPERLINKS tag is set to YES, the LaTeX that is generated is
+# prepared for conversion to PDF (using ps2pdf or pdflatex). The PDF file will
+# contain links (just like the HTML output) instead of page references. This
+# makes the output suitable for online browsing using a PDF viewer.
+# The default value is: YES.
+# This tag requires that the tag GENERATE_LATEX is set to YES.
+
+PDF_HYPERLINKS         = YES
+
+# If the USE_PDFLATEX tag is set to YES, doxygen will use pdflatex to generate
+# the PDF file directly from the LaTeX files. Set this option to YES, to get a
+# higher quality PDF documentation.
+# The default value is: YES.
+# This tag requires that the tag GENERATE_LATEX is set to YES.
+
+USE_PDFLATEX           = YES
+
+# If the LATEX_BATCHMODE tag is set to YES, doxygen will add the \batchmode
+# command to the generated LaTeX files. This will instruct LaTeX to keep running
+# if errors occur, instead of asking the user for help. This option is also used
+# when generating formulas in HTML.
+# The default value is: NO.
+# This tag requires that the tag GENERATE_LATEX is set to YES.
+
+LATEX_BATCHMODE        = NO
+
+# If the LATEX_HIDE_INDICES tag is set to YES then doxygen will not include the
+# index chapters (such as File Index, Compound Index, etc.) in the output.
+# The default value is: NO.
+# This tag requires that the tag GENERATE_LATEX is set to YES.
+
+LATEX_HIDE_INDICES     = NO
+
+# If the LATEX_SOURCE_CODE tag is set to YES then doxygen will include source
+# code with syntax highlighting in the LaTeX output.
+#
+# Note that which sources are shown also depends on other settings such as
+# SOURCE_BROWSER.
+# The default value is: NO.
+# This tag requires that the tag GENERATE_LATEX is set to YES.
+
+LATEX_SOURCE_CODE      = NO
+
+# The LATEX_BIB_STYLE tag can be used to specify the style to use for the
+# bibliography, e.g. plainnat, or ieeetr. See
+# https://en.wikipedia.org/wiki/BibTeX and \cite for more info.
+# The default value is: plain.
+# This tag requires that the tag GENERATE_LATEX is set to YES.
+
+LATEX_BIB_STYLE        = plain
+
+# If the LATEX_TIMESTAMP tag is set to YES then the footer of each generated
+# page will contain the date and time when the page was generated. Setting this
+# to NO can help when comparing the output of multiple runs.
+# The default value is: NO.
+# This tag requires that the tag GENERATE_LATEX is set to YES.
+
+LATEX_TIMESTAMP        = NO
+
+# The LATEX_EMOJI_DIRECTORY tag is used to specify the (relative or absolute)
+# path from which the emoji images will be read. If a relative path is entered,
+# it will be relative to the LATEX_OUTPUT directory. If left blank the
+# LATEX_OUTPUT directory will be used.
+# This tag requires that the tag GENERATE_LATEX is set to YES.
+
+LATEX_EMOJI_DIRECTORY  =
+
+#---------------------------------------------------------------------------
+# Configuration options related to the RTF output
+#---------------------------------------------------------------------------
+
+# If the GENERATE_RTF tag is set to YES, doxygen will generate RTF output. The
+# RTF output is optimized for Word 97 and may not look too pretty with other RTF
+# readers/editors.
+# The default value is: NO.
+
+GENERATE_RTF           = NO
+
+# The RTF_OUTPUT tag is used to specify where the RTF docs will be put. If a
+# relative path is entered the value of OUTPUT_DIRECTORY will be put in front of
+# it.
+# The default directory is: rtf.
+# This tag requires that the tag GENERATE_RTF is set to YES.
+
+RTF_OUTPUT             = rtf
+
+# If the COMPACT_RTF tag is set to YES, doxygen generates more compact RTF
+# documents. This may be useful for small projects and may help to save some
+# trees in general.
+# The default value is: NO.
+# This tag requires that the tag GENERATE_RTF is set to YES.
+
+COMPACT_RTF            = NO
+
+# If the RTF_HYPERLINKS tag is set to YES, the RTF that is generated will
+# contain hyperlink fields. The RTF file will contain links (just like the HTML
+# output) instead of page references. This makes the output suitable for online
+# browsing using Word or some other Word compatible readers that support those
+# fields.
+#
+# Note: WordPad (write) and others do not support links.
+# The default value is: NO.
+# This tag requires that the tag GENERATE_RTF is set to YES.
+
+RTF_HYPERLINKS         = NO
+
+# Load stylesheet definitions from file. Syntax is similar to doxygen's
+# configuration file, i.e. a series of assignments. You only have to provide
+# replacements, missing definitions are set to their default value.
+#
+# See also section "Doxygen usage" for information on how to generate the
+# default style sheet that doxygen normally uses.
+# This tag requires that the tag GENERATE_RTF is set to YES.
+
+RTF_STYLESHEET_FILE    =
+
+# Set optional variables used in the generation of an RTF document. Syntax is
+# similar to doxygen's configuration file. A template extensions file can be
+# generated using doxygen -e rtf extensionFile.
+# This tag requires that the tag GENERATE_RTF is set to YES.
+
+RTF_EXTENSIONS_FILE    =
+
+# If the RTF_SOURCE_CODE tag is set to YES then doxygen will include source code
+# with syntax highlighting in the RTF output.
+#
+# Note that which sources are shown also depends on other settings such as
+# SOURCE_BROWSER.
+# The default value is: NO.
+# This tag requires that the tag GENERATE_RTF is set to YES.
+
+RTF_SOURCE_CODE        = NO
+
+#---------------------------------------------------------------------------
+# Configuration options related to the man page output
+#---------------------------------------------------------------------------
+
+# If the GENERATE_MAN tag is set to YES, doxygen will generate man pages for
+# classes and files.
+# The default value is: NO.
+
+GENERATE_MAN           = NO
+
+# The MAN_OUTPUT tag is used to specify where the man pages will be put. If a
+# relative path is entered the value of OUTPUT_DIRECTORY will be put in front of
+# it. A directory man3 will be created inside the directory specified by
+# MAN_OUTPUT.
+# The default directory is: man.
+# This tag requires that the tag GENERATE_MAN is set to YES.
+
+MAN_OUTPUT             = man
+
+# The MAN_EXTENSION tag determines the extension that is added to the generated
+# man pages. In case the manual section does not start with a number, the number
+# 3 is prepended. The dot (.) at the beginning of the MAN_EXTENSION tag is
+# optional.
+# The default value is: .3.
+# This tag requires that the tag GENERATE_MAN is set to YES.
+
+MAN_EXTENSION          = .3
+
+# The MAN_SUBDIR tag determines the name of the directory created within
+# MAN_OUTPUT in which the man pages are placed. If defaults to man followed by
+# MAN_EXTENSION with the initial . removed.
+# This tag requires that the tag GENERATE_MAN is set to YES.
+
+MAN_SUBDIR             =
+
+# If the MAN_LINKS tag is set to YES and doxygen generates man output, then it
+# will generate one additional man file for each entity documented in the real
+# man page(s). These additional files only source the real man page, but without
+# them the man command would be unable to find the correct page.
+# The default value is: NO.
+# This tag requires that the tag GENERATE_MAN is set to YES.
+
+MAN_LINKS              = NO
+
+#---------------------------------------------------------------------------
+# Configuration options related to the XML output
+#---------------------------------------------------------------------------
+
+# If the GENERATE_XML tag is set to YES, doxygen will generate an XML file that
+# captures the structure of the code including all documentation.
+# The default value is: NO.
+
+GENERATE_XML           = NO
+
+# The XML_OUTPUT tag is used to specify where the XML pages will be put. If a
+# relative path is entered the value of OUTPUT_DIRECTORY will be put in front of
+# it.
+# The default directory is: xml.
+# This tag requires that the tag GENERATE_XML is set to YES.
+
+XML_OUTPUT             = xml
+
+# If the XML_PROGRAMLISTING tag is set to YES, doxygen will dump the program
+# listings (including syntax highlighting and cross-referencing information) to
+# the XML output. Note that enabling this will significantly increase the size
+# of the XML output.
+# The default value is: YES.
+# This tag requires that the tag GENERATE_XML is set to YES.
+
+XML_PROGRAMLISTING     = YES
+
+# If the XML_NS_MEMB_FILE_SCOPE tag is set to YES, doxygen will include
+# namespace members in file scope as well, matching the HTML output.
+# The default value is: NO.
+# This tag requires that the tag GENERATE_XML is set to YES.
+
+XML_NS_MEMB_FILE_SCOPE = NO
+
+#---------------------------------------------------------------------------
+# Configuration options related to the DOCBOOK output
+#---------------------------------------------------------------------------
+
+# If the GENERATE_DOCBOOK tag is set to YES, doxygen will generate Docbook files
+# that can be used to generate PDF.
+# The default value is: NO.
+
+GENERATE_DOCBOOK       = NO
+
+# The DOCBOOK_OUTPUT tag is used to specify where the Docbook pages will be put.
+# If a relative path is entered the value of OUTPUT_DIRECTORY will be put in
+# front of it.
+# The default directory is: docbook.
+# This tag requires that the tag GENERATE_DOCBOOK is set to YES.
+
+DOCBOOK_OUTPUT         = docbook
+
+# If the DOCBOOK_PROGRAMLISTING tag is set to YES, doxygen will include the
+# program listings (including syntax highlighting and cross-referencing
+# information) to the DOCBOOK output. Note that enabling this will significantly
+# increase the size of the DOCBOOK output.
+# The default value is: NO.
+# This tag requires that the tag GENERATE_DOCBOOK is set to YES.
+
+DOCBOOK_PROGRAMLISTING = NO
+
+#---------------------------------------------------------------------------
+# Configuration options for the AutoGen Definitions output
+#---------------------------------------------------------------------------
+
+# If the GENERATE_AUTOGEN_DEF tag is set to YES, doxygen will generate an
+# AutoGen Definitions (see http://autogen.sourceforge.net/) file that captures
+# the structure of the code including all documentation. Note that this feature
+# is still experimental and incomplete at the moment.
+# The default value is: NO.
+
+GENERATE_AUTOGEN_DEF   = NO
+
+#---------------------------------------------------------------------------
+# Configuration options related to the Perl module output
+#---------------------------------------------------------------------------
+
+# If the GENERATE_PERLMOD tag is set to YES, doxygen will generate a Perl module
+# file that captures the structure of the code including all documentation.
+#
+# Note that this feature is still experimental and incomplete at the moment.
+# The default value is: NO.
+
+GENERATE_PERLMOD       = NO
+
+# If the PERLMOD_LATEX tag is set to YES, doxygen will generate the necessary
+# Makefile rules, Perl scripts and LaTeX code to be able to generate PDF and DVI
+# output from the Perl module output.
+# The default value is: NO.
+# This tag requires that the tag GENERATE_PERLMOD is set to YES.
+
+PERLMOD_LATEX          = NO
+
+# If the PERLMOD_PRETTY tag is set to YES, the Perl module output will be nicely
+# formatted so it can be parsed by a human reader. This is useful if you want to
+# understand what is going on. On the other hand, if this tag is set to NO, the
+# size of the Perl module output will be much smaller and Perl will parse it
+# just the same.
+# The default value is: YES.
+# This tag requires that the tag GENERATE_PERLMOD is set to YES.
+
+PERLMOD_PRETTY         = YES
+
+# The names of the make variables in the generated doxyrules.make file are
+# prefixed with the string contained in PERLMOD_MAKEVAR_PREFIX. This is useful
+# so different doxyrules.make files included by the same Makefile don't
+# overwrite each other's variables.
+# This tag requires that the tag GENERATE_PERLMOD is set to YES.
+
+PERLMOD_MAKEVAR_PREFIX =
+
+#---------------------------------------------------------------------------
+# Configuration options related to the preprocessor
+#---------------------------------------------------------------------------
+
+# If the ENABLE_PREPROCESSING tag is set to YES, doxygen will evaluate all
+# C-preprocessor directives found in the sources and include files.
+# The default value is: YES.
+
+ENABLE_PREPROCESSING   = YES
+
+# If the MACRO_EXPANSION tag is set to YES, doxygen will expand all macro names
+# in the source code. If set to NO, only conditional compilation will be
+# performed. Macro expansion can be done in a controlled way by setting
+# EXPAND_ONLY_PREDEF to YES.
+# The default value is: NO.
+# This tag requires that the tag ENABLE_PREPROCESSING is set to YES.
+
+MACRO_EXPANSION        = NO
+
+# If the EXPAND_ONLY_PREDEF and MACRO_EXPANSION tags are both set to YES then
+# the macro expansion is limited to the macros specified with the PREDEFINED and
+# EXPAND_AS_DEFINED tags.
+# The default value is: NO.
+# This tag requires that the tag ENABLE_PREPROCESSING is set to YES.
+
+EXPAND_ONLY_PREDEF     = NO
+
+# If the SEARCH_INCLUDES tag is set to YES, the include files in the
+# INCLUDE_PATH will be searched if a #include is found.
+# The default value is: YES.
+# This tag requires that the tag ENABLE_PREPROCESSING is set to YES.
+
+SEARCH_INCLUDES        = YES
+
+# The INCLUDE_PATH tag can be used to specify one or more directories that
+# contain include files that are not input files but should be processed by the
+# preprocessor.
+# This tag requires that the tag SEARCH_INCLUDES is set to YES.
+
+INCLUDE_PATH           =
+
+# You can use the INCLUDE_FILE_PATTERNS tag to specify one or more wildcard
+# patterns (like *.h and *.hpp) to filter out the header-files in the
+# directories. If left blank, the patterns specified with FILE_PATTERNS will be
+# used.
+# This tag requires that the tag ENABLE_PREPROCESSING is set to YES.
+
+INCLUDE_FILE_PATTERNS  =
+
+# The PREDEFINED tag can be used to specify one or more macro names that are
+# defined before the preprocessor is started (similar to the -D option of e.g.
+# gcc). The argument of the tag is a list of macros of the form: name or
+# name=definition (no spaces). If the definition and the "=" are omitted, "=1"
+# is assumed. To prevent a macro definition from being undefined via #undef or
+# recursively expanded use the := operator instead of the = operator.
+# This tag requires that the tag ENABLE_PREPROCESSING is set to YES.
+
+PREDEFINED             =
+
+# If the MACRO_EXPANSION and EXPAND_ONLY_PREDEF tags are set to YES then this
+# tag can be used to specify a list of macro names that should be expanded. The
+# macro definition that is found in the sources will be used. Use the PREDEFINED
+# tag if you want to use a different macro definition that overrules the
+# definition found in the source code.
+# This tag requires that the tag ENABLE_PREPROCESSING is set to YES.
+
+EXPAND_AS_DEFINED      =
+
+# If the SKIP_FUNCTION_MACROS tag is set to YES then doxygen's preprocessor will
+# remove all references to function-like macros that are alone on a line, have
+# an all uppercase name, and do not end with a semicolon. Such function macros
+# are typically used for boiler-plate code, and will confuse the parser if not
+# removed.
+# The default value is: YES.
+# This tag requires that the tag ENABLE_PREPROCESSING is set to YES.
+
+SKIP_FUNCTION_MACROS   = YES
+
+#---------------------------------------------------------------------------
+# Configuration options related to external references
+#---------------------------------------------------------------------------
+
+# The TAGFILES tag can be used to specify one or more tag files. For each tag
+# file the location of the external documentation should be added. The format of
+# a tag file without this location is as follows:
+# TAGFILES = file1 file2 ...
+# Adding location for the tag files is done as follows:
+# TAGFILES = file1=loc1 "file2 = loc2" ...
+# where loc1 and loc2 can be relative or absolute paths or URLs. See the
+# section "Linking to external documentation" for more information about the use
+# of tag files.
+# Note: Each tag file must have a unique name (where the name does NOT include
+# the path). If a tag file is not located in the directory in which doxygen is
+# run, you must also specify the path to the tagfile here.
+
+TAGFILES               =
+
+# When a file name is specified after GENERATE_TAGFILE, doxygen will create a
+# tag file that is based on the input files it reads. See section "Linking to
+# external documentation" for more information about the usage of tag files.
+
+GENERATE_TAGFILE       =
+
+# If the ALLEXTERNALS tag is set to YES, all external class will be listed in
+# the class index. If set to NO, only the inherited external classes will be
+# listed.
+# The default value is: NO.
+
+ALLEXTERNALS           = NO
+
+# If the EXTERNAL_GROUPS tag is set to YES, all external groups will be listed
+# in the modules index. If set to NO, only the current project's groups will be
+# listed.
+# The default value is: YES.
+
+EXTERNAL_GROUPS        = YES
+
+# If the EXTERNAL_PAGES tag is set to YES, all external pages will be listed in
+# the related pages index. If set to NO, only the current project's pages will
+# be listed.
+# The default value is: YES.
+
+EXTERNAL_PAGES         = YES
+
+#---------------------------------------------------------------------------
+# Configuration options related to the dot tool
+#---------------------------------------------------------------------------
+
+# If the CLASS_DIAGRAMS tag is set to YES, doxygen will generate a class diagram
+# (in HTML and LaTeX) for classes with base or super classes. Setting the tag to
+# NO turns the diagrams off. Note that this option also works with HAVE_DOT
+# disabled, but it is recommended to install and use dot, since it yields more
+# powerful graphs.
+# The default value is: YES.
+
+CLASS_DIAGRAMS         = YES
+
+# You can include diagrams made with dia in doxygen documentation. Doxygen will
+# then run dia to produce the diagram and insert it in the documentation. The
+# DIA_PATH tag allows you to specify the directory where the dia binary resides.
+# If left empty dia is assumed to be found in the default search path.
+
+DIA_PATH               =
+
+# If set to YES the inheritance and collaboration graphs will hide inheritance
+# and usage relations if the target is undocumented or is not a class.
+# The default value is: YES.
+
+HIDE_UNDOC_RELATIONS   = YES
+
+# If you set the HAVE_DOT tag to YES then doxygen will assume the dot tool is
+# available from the path. This tool is part of Graphviz (see:
+# http://www.graphviz.org/), a graph visualization toolkit from AT&T and Lucent
+# Bell Labs. The other options in this section have no effect if this option is
+# set to NO
+# The default value is: YES.
+
+HAVE_DOT               = YES
+
+# The DOT_NUM_THREADS specifies the number of dot invocations doxygen is allowed
+# to run in parallel. When set to 0 doxygen will base this on the number of
+# processors available in the system. You can set it explicitly to a value
+# larger than 0 to get control over the balance between CPU load and processing
+# speed.
+# Minimum value: 0, maximum value: 32, default value: 0.
+# This tag requires that the tag HAVE_DOT is set to YES.
+
+DOT_NUM_THREADS        = 0
+
+# When you want a differently looking font in the dot files that doxygen
+# generates you can specify the font name using DOT_FONTNAME. You need to make
+# sure dot is able to find the font, which can be done by putting it in a
+# standard location or by setting the DOTFONTPATH environment variable or by
+# setting DOT_FONTPATH to the directory containing the font.
+# The default value is: Helvetica.
+# This tag requires that the tag HAVE_DOT is set to YES.
+
+DOT_FONTNAME           = Helvetica
+
+# The DOT_FONTSIZE tag can be used to set the size (in points) of the font of
+# dot graphs.
+# Minimum value: 4, maximum value: 24, default value: 10.
+# This tag requires that the tag HAVE_DOT is set to YES.
+
+DOT_FONTSIZE           = 10
+
+# By default doxygen will tell dot to use the default font as specified with
+# DOT_FONTNAME. If you specify a different font using DOT_FONTNAME you can set
+# the path where dot can find it using this tag.
+# This tag requires that the tag HAVE_DOT is set to YES.
+
+DOT_FONTPATH           =
+
+# If the CLASS_GRAPH tag is set to YES then doxygen will generate a graph for
+# each documented class showing the direct and indirect inheritance relations.
+# Setting this tag to YES will force the CLASS_DIAGRAMS tag to NO.
+# The default value is: YES.
+# This tag requires that the tag HAVE_DOT is set to YES.
+
+CLASS_GRAPH            = YES
+
+# If the COLLABORATION_GRAPH tag is set to YES then doxygen will generate a
+# graph for each documented class showing the direct and indirect implementation
+# dependencies (inheritance, containment, and class references variables) of the
+# class with other documented classes.
+# The default value is: YES.
+# This tag requires that the tag HAVE_DOT is set to YES.
+
+COLLABORATION_GRAPH    = YES
+
+# If the GROUP_GRAPHS tag is set to YES then doxygen will generate a graph for
+# groups, showing the direct groups dependencies.
+# The default value is: YES.
+# This tag requires that the tag HAVE_DOT is set to YES.
+
+GROUP_GRAPHS           = YES
+
+# If the UML_LOOK tag is set to YES, doxygen will generate inheritance and
+# collaboration diagrams in a style similar to the OMG's Unified Modeling
+# Language.
+# The default value is: NO.
+# This tag requires that the tag HAVE_DOT is set to YES.
+
+UML_LOOK               = NO
+
+# If the UML_LOOK tag is enabled, the fields and methods are shown inside the
+# class node. If there are many fields or methods and many nodes the graph may
+# become too big to be useful. The UML_LIMIT_NUM_FIELDS threshold limits the
+# number of items for each type to make the size more manageable. Set this to 0
+# for no limit. Note that the threshold may be exceeded by 50% before the limit
+# is enforced. So when you set the threshold to 10, up to 15 fields may appear,
+# but if the number exceeds 15, the total amount of fields shown is limited to
+# 10.
+# Minimum value: 0, maximum value: 100, default value: 10.
+# This tag requires that the tag HAVE_DOT is set to YES.
+
+UML_LIMIT_NUM_FIELDS   = 10
+
+# If the TEMPLATE_RELATIONS tag is set to YES then the inheritance and
+# collaboration graphs will show the relations between templates and their
+# instances.
+# The default value is: NO.
+# This tag requires that the tag HAVE_DOT is set to YES.
+
+TEMPLATE_RELATIONS     = NO
+
+# If the INCLUDE_GRAPH, ENABLE_PREPROCESSING and SEARCH_INCLUDES tags are set to
+# YES then doxygen will generate a graph for each documented file showing the
+# direct and indirect include dependencies of the file with other documented
+# files.
+# The default value is: YES.
+# This tag requires that the tag HAVE_DOT is set to YES.
+
+INCLUDE_GRAPH          = YES
+
+# If the INCLUDED_BY_GRAPH, ENABLE_PREPROCESSING and SEARCH_INCLUDES tags are
+# set to YES then doxygen will generate a graph for each documented file showing
+# the direct and indirect include dependencies of the file with other documented
+# files.
+# The default value is: YES.
+# This tag requires that the tag HAVE_DOT is set to YES.
+
+INCLUDED_BY_GRAPH      = YES
+
+# If the CALL_GRAPH tag is set to YES then doxygen will generate a call
+# dependency graph for every global function or class method.
+#
+# Note that enabling this option will significantly increase the time of a run.
+# So in most cases it will be better to enable call graphs for selected
+# functions only using the \callgraph command. Disabling a call graph can be
+# accomplished by means of the command \hidecallgraph.
+# The default value is: NO.
+# This tag requires that the tag HAVE_DOT is set to YES.
+
+CALL_GRAPH             = NO
+
+# If the CALLER_GRAPH tag is set to YES then doxygen will generate a caller
+# dependency graph for every global function or class method.
+#
+# Note that enabling this option will significantly increase the time of a run.
+# So in most cases it will be better to enable caller graphs for selected
+# functions only using the \callergraph command. Disabling a caller graph can be
+# accomplished by means of the command \hidecallergraph.
+# The default value is: NO.
+# This tag requires that the tag HAVE_DOT is set to YES.
+
+CALLER_GRAPH           = NO
+
+# If the GRAPHICAL_HIERARCHY tag is set to YES then doxygen will graphical
+# hierarchy of all classes instead of a textual one.
+# The default value is: YES.
+# This tag requires that the tag HAVE_DOT is set to YES.
+
+GRAPHICAL_HIERARCHY    = YES
+
+# If the DIRECTORY_GRAPH tag is set to YES then doxygen will show the
+# dependencies a directory has on other directories in a graphical way. The
+# dependency relations are determined by the #include relations between the
+# files in the directories.
+# The default value is: YES.
+# This tag requires that the tag HAVE_DOT is set to YES.
+
+DIRECTORY_GRAPH        = YES
+
+# The DOT_IMAGE_FORMAT tag can be used to set the image format of the images
+# generated by dot. For an explanation of the image formats see the section
+# output formats in the documentation of the dot tool (Graphviz (see:
+# http://www.graphviz.org/)).
+# Note: If you choose svg you need to set HTML_FILE_EXTENSION to xhtml in order
+# to make the SVG files visible in IE 9+ (other browsers do not have this
+# requirement).
+# Possible values are: png, png:cairo, png:cairo:cairo, png:cairo:gd, png:gd,
+# png:gd:gd, jpg, jpg:cairo, jpg:cairo:gd, jpg:gd, jpg:gd:gd, gif, gif:cairo,
+# gif:cairo:gd, gif:gd, gif:gd:gd, svg, png:gd, png:gd:gd, png:cairo,
+# png:cairo:gd, png:cairo:cairo, png:cairo:gdiplus, png:gdiplus and
+# png:gdiplus:gdiplus.
+# The default value is: png.
+# This tag requires that the tag HAVE_DOT is set to YES.
+
+DOT_IMAGE_FORMAT       = png
+
+# If DOT_IMAGE_FORMAT is set to svg, then this option can be set to YES to
+# enable generation of interactive SVG images that allow zooming and panning.
+#
+# Note that this requires a modern browser other than Internet Explorer. Tested
+# and working are Firefox, Chrome, Safari, and Opera.
+# Note: For IE 9+ you need to set HTML_FILE_EXTENSION to xhtml in order to make
+# the SVG files visible. Older versions of IE do not have SVG support.
+# The default value is: NO.
+# This tag requires that the tag HAVE_DOT is set to YES.
+
+INTERACTIVE_SVG        = NO
+
+# The DOT_PATH tag can be used to specify the path where the dot tool can be
+# found. If left blank, it is assumed the dot tool can be found in the path.
+# This tag requires that the tag HAVE_DOT is set to YES.
+
+DOT_PATH               =
+
+# The DOTFILE_DIRS tag can be used to specify one or more directories that
+# contain dot files that are included in the documentation (see the \dotfile
+# command).
+# This tag requires that the tag HAVE_DOT is set to YES.
+
+DOTFILE_DIRS           =
+
+# The MSCFILE_DIRS tag can be used to specify one or more directories that
+# contain msc files that are included in the documentation (see the \mscfile
+# command).
+
+MSCFILE_DIRS           =
+
+# The DIAFILE_DIRS tag can be used to specify one or more directories that
+# contain dia files that are included in the documentation (see the \diafile
+# command).
+
+DIAFILE_DIRS           =
+
+# When using plantuml, the PLANTUML_JAR_PATH tag should be used to specify the
+# path where java can find the plantuml.jar file. If left blank, it is assumed
+# PlantUML is not used or called during a preprocessing step. Doxygen will
+# generate a warning when it encounters a \startuml command in this case and
+# will not generate output for the diagram.
+
+PLANTUML_JAR_PATH      =
+
+# When using plantuml, the PLANTUML_CFG_FILE tag can be used to specify a
+# configuration file for plantuml.
+
+PLANTUML_CFG_FILE      =
+
+# When using plantuml, the specified paths are searched for files specified by
+# the !include statement in a plantuml block.
+
+PLANTUML_INCLUDE_PATH  =
+
+# The DOT_GRAPH_MAX_NODES tag can be used to set the maximum number of nodes
+# that will be shown in the graph. If the number of nodes in a graph becomes
+# larger than this value, doxygen will truncate the graph, which is visualized
+# by representing a node as a red box. Note that doxygen if the number of direct
+# children of the root node in a graph is already larger than
+# DOT_GRAPH_MAX_NODES then the graph will not be shown at all. Also note that
+# the size of a graph can be further restricted by MAX_DOT_GRAPH_DEPTH.
+# Minimum value: 0, maximum value: 10000, default value: 50.
+# This tag requires that the tag HAVE_DOT is set to YES.
+
+DOT_GRAPH_MAX_NODES    = 50
+
+# The MAX_DOT_GRAPH_DEPTH tag can be used to set the maximum depth of the graphs
+# generated by dot. A depth value of 3 means that only nodes reachable from the
+# root by following a path via at most 3 edges will be shown. Nodes that lay
+# further from the root node will be omitted. Note that setting this option to 1
+# or 2 may greatly reduce the computation time needed for large code bases. Also
+# note that the size of a graph can be further restricted by
+# DOT_GRAPH_MAX_NODES. Using a depth of 0 means no depth restriction.
+# Minimum value: 0, maximum value: 1000, default value: 0.
+# This tag requires that the tag HAVE_DOT is set to YES.
+
+MAX_DOT_GRAPH_DEPTH    = 0
+
+# Set the DOT_TRANSPARENT tag to YES to generate images with a transparent
+# background. This is disabled by default, because dot on Windows does not seem
+# to support this out of the box.
+#
+# Warning: Depending on the platform used, enabling this option may lead to
+# badly anti-aliased labels on the edges of a graph (i.e. they become hard to
+# read).
+# The default value is: NO.
+# This tag requires that the tag HAVE_DOT is set to YES.
+
+DOT_TRANSPARENT        = NO
+
+# Set the DOT_MULTI_TARGETS tag to YES to allow dot to generate multiple output
+# files in one run (i.e. multiple -o and -T options on the command line). This
+# makes dot run faster, but since only newer versions of dot (>1.8.10) support
+# this, this feature is disabled by default.
+# The default value is: NO.
+# This tag requires that the tag HAVE_DOT is set to YES.
+
+DOT_MULTI_TARGETS      = NO
+
+# If the GENERATE_LEGEND tag is set to YES doxygen will generate a legend page
+# explaining the meaning of the various boxes and arrows in the dot generated
+# graphs.
+# The default value is: YES.
+# This tag requires that the tag HAVE_DOT is set to YES.
+
+GENERATE_LEGEND        = YES
+
+# If the DOT_CLEANUP tag is set to YES, doxygen will remove the intermediate dot
+# files that are used to generate the various graphs.
+# The default value is: YES.
+# This tag requires that the tag HAVE_DOT is set to YES.
+
+DOT_CLEANUP            = YES
diff --git a/src/whisper_cpp/ggml/src/ggml-cann/acl_tensor.cpp b/src/whisper_cpp/ggml/src/ggml-cann/acl_tensor.cpp
new file mode 100644
index 00000000..d120ce6a
--- /dev/null
+++ b/src/whisper_cpp/ggml/src/ggml-cann/acl_tensor.cpp
@@ -0,0 +1,175 @@
+/*
+ * Copyright (c) 2023-2024 The ggml authors
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a copy
+ * of this software and associated documentation files (the "Software"), to
+ * deal in the Software without restriction, including without limitation the
+ * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or
+ * sell copies of the Software, and to permit persons to whom the Software is
+ * furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice shall be included in
+ * all copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+ * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
+ * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
+ * IN THE SOFTWARE.
+ */
+
+#include "acl_tensor.h"
+
+#include <algorithm>
+#include <cstring>
+
+aclDataType ggml_cann_type_mapping(ggml_type type) {
+    switch (type) {
+        case GGML_TYPE_F32:
+            return ACL_FLOAT;
+        case GGML_TYPE_F16:
+            return ACL_FLOAT16;
+        case GGML_TYPE_I8:
+            return ACL_INT8;
+        case GGML_TYPE_I16:
+            return ACL_INT16;
+        case GGML_TYPE_I32:
+            return ACL_INT32;
+        case GGML_TYPE_Q4_0:
+            return ACL_INT4;
+        case GGML_TYPE_Q8_0:
+            return ACL_INT8;
+        default:
+            return ACL_DT_UNDEFINED;
+    }
+    return ACL_DT_UNDEFINED;
+}
+
+aclTensor* ggml_cann_create_tensor(const ggml_tensor* tensor, int64_t* ne,
+                                   size_t* nb, int64_t dims, aclFormat format,
+                                   size_t offset) {
+    // If tensor is bcasted, Up to GGML_MAX_DIMS additional dimensions will be
+    // added.
+    int64_t acl_ne[GGML_MAX_DIMS * 2], acl_stride[GGML_MAX_DIMS * 2];
+
+    int64_t acl_storage_len = 0;
+    if (ne == nullptr) {
+        acl_storage_len = ggml_nbytes(tensor);
+        for (int i = 0; i < GGML_MAX_DIMS; i++) {
+            acl_ne[i] = tensor->ne[i];
+            // The step size of acl is in elements.
+            acl_stride[i] = tensor->nb[i] / ggml_element_size(tensor);
+        }
+    } else {
+        // With bcast
+        for (int i = 0; i < dims; i++) {
+            acl_storage_len += (ne[i] - 1) * nb[i];
+            acl_ne[i] = ne[i];
+            acl_stride[i] = nb[i] / ggml_element_size(tensor);
+        }
+    }
+
+    // Reverse ne and stride.
+    int64_t final_dims = (dims == 0 ? GGML_MAX_DIMS : dims);
+    std::reverse(acl_ne, acl_ne + final_dims);
+    std::reverse(acl_stride, acl_stride + final_dims);
+
+    aclTensor* acl_tensor = aclCreateTensor(
+        acl_ne, final_dims, ggml_cann_type_mapping(tensor->type), acl_stride,
+        offset / ggml_element_size(tensor), format, &acl_storage_len, 1,
+        tensor->data);
+
+    return acl_tensor;
+}
+
+bool ggml_cann_need_bcast(const ggml_tensor* t0, const ggml_tensor* t1) {
+    for (int i = 0; i < GGML_MAX_DIMS; i++) {
+        if (t1->ne[i] != t0->ne[i] && t1->ne[i] != 1) {
+            return true;
+        }
+    }
+    return false;
+}
+
+int64_t ggml_cann_get_bcast_shape(const ggml_tensor* src0,
+                                  const ggml_tensor* src1,
+                                  int64_t* bcast_src0_ne,
+                                  int64_t* bcast_src1_ne, size_t* bcast_src0_nb,
+                                  size_t* bcast_src1_nb) {
+    GGML_ASSERT(ggml_can_repeat(src1, src0));
+    int bcast_dim_cnt = 0;
+    for (int i = 0; i < GGML_MAX_DIMS; i++) {
+        int64_t nr = src0->ne[i] / src1->ne[i];
+        bcast_src0_ne[bcast_dim_cnt] = src0->ne[i] / nr;
+        bcast_src1_ne[bcast_dim_cnt] = src1->ne[i];
+        bcast_src0_nb[bcast_dim_cnt] = src0->nb[i];
+        bcast_src1_nb[bcast_dim_cnt] = src1->nb[i];
+        bcast_dim_cnt++;
+        if (nr != 1) {
+            // Need to add an extra dim.
+            bcast_src0_ne[bcast_dim_cnt] = nr;
+            bcast_src1_ne[bcast_dim_cnt] = 1;
+            bcast_src0_nb[bcast_dim_cnt] = bcast_src0_nb[bcast_dim_cnt - 1] *
+                                           bcast_src0_ne[bcast_dim_cnt - 1];
+            bcast_src1_nb[bcast_dim_cnt] = bcast_src1_nb[bcast_dim_cnt - 1] *
+                                           bcast_src1_ne[bcast_dim_cnt - 1];
+            bcast_dim_cnt++;
+        }
+    }
+    return bcast_dim_cnt;
+}
+
+int64_t ggml_cann_get_mulmat_bcast_shape(
+    const int64_t* input_ne, const int64_t* weight_ne, const int64_t* dst_ne,
+    const size_t* input_nb, const size_t* weight_nb, const size_t* dst_nb,
+    int64_t* bcast_input_ne, int64_t* bcast_weight_ne, int64_t* bcast_dst_ne,
+    size_t* bcast_input_nb, size_t* bcast_weight_nb, size_t* bcast_dst_nb) {
+    // input and dst shoule in same shape, except first two dims.
+    GGML_ASSERT(input_ne[2] == dst_ne[2]);
+    GGML_ASSERT(input_ne[3] == dst_ne[3]);
+
+    int bcast_dim_cnt = 0;
+
+    // For mul_mat, a dimension needs to be added before the dimension that
+    // weight needs to be expanded to satisfy the bcast rule of matrix
+    // multiplication.
+    for (int i = 0; i < GGML_MAX_DIMS; i++) {
+        int64_t nr = input_ne[i] / weight_ne[i];
+        // Do not use bcast in the first two dimensions because we only support
+        // the bcast batch dimension. Just copy them.
+        if (i < 2 || nr == 1) {
+            bcast_input_ne[bcast_dim_cnt] = input_ne[i];
+            bcast_weight_ne[bcast_dim_cnt] = weight_ne[i];
+            bcast_dst_ne[bcast_dim_cnt] = dst_ne[i];
+
+            bcast_input_nb[bcast_dim_cnt] = input_nb[i];
+            bcast_weight_nb[bcast_dim_cnt] = weight_nb[i];
+            bcast_dst_nb[bcast_dim_cnt] = dst_nb[i];
+            bcast_dim_cnt++;
+        } else {
+            // Need to add an extra dim.
+            bcast_input_ne[bcast_dim_cnt] = nr;
+            bcast_dst_ne[bcast_dim_cnt] = nr;
+            bcast_weight_ne[bcast_dim_cnt] = 1;
+            bcast_input_nb[bcast_dim_cnt] = input_nb[i];
+            bcast_dst_nb[bcast_dim_cnt] = dst_nb[i];
+            bcast_weight_nb[bcast_dim_cnt] = weight_nb[i];
+            bcast_dim_cnt++;
+
+            bcast_input_ne[bcast_dim_cnt] = input_ne[i] / nr;
+            bcast_dst_ne[bcast_dim_cnt] = dst_ne[i] / nr;
+            bcast_weight_ne[bcast_dim_cnt] = weight_ne[i];
+            bcast_input_nb[bcast_dim_cnt] = bcast_input_nb[bcast_dim_cnt - 1] *
+                                            bcast_input_ne[bcast_dim_cnt - 1];
+            bcast_dst_nb[bcast_dim_cnt] = bcast_dst_nb[bcast_dim_cnt - 1] *
+                                          bcast_dst_ne[bcast_dim_cnt - 1];
+            bcast_weight_nb[bcast_dim_cnt] =
+                bcast_weight_nb[bcast_dim_cnt - 1] *
+                bcast_weight_ne[bcast_dim_cnt - 1];
+            bcast_dim_cnt++;
+        }
+    }
+    return bcast_dim_cnt;
+}
diff --git a/src/whisper_cpp/ggml/src/ggml-cann/acl_tensor.h b/src/whisper_cpp/ggml/src/ggml-cann/acl_tensor.h
new file mode 100644
index 00000000..4734a9cb
--- /dev/null
+++ b/src/whisper_cpp/ggml/src/ggml-cann/acl_tensor.h
@@ -0,0 +1,258 @@
+/*
+ * Copyright (c) 2023-2024 The ggml authors
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a copy
+ * of this software and associated documentation files (the "Software"), to
+ * deal in the Software without restriction, including without limitation the
+ * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or
+ * sell copies of the Software, and to permit persons to whom the Software is
+ * furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice shall be included in
+ * all copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+ * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
+ * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
+ * IN THE SOFTWARE.
+ */
+
+#ifndef CANN_ACL_TENSOR_H
+#define CANN_ACL_TENSOR_H
+
+#include <algorithm>
+#include <cstring>
+
+#include <aclnn/aclnn_base.h>
+#include "common.h"
+
+/**
+ * @brief	Maps a ggml_type to its corresponding aclDataType.
+ *
+ * @details	This function takes a ggml_type as input and returns the corresponding
+ *			aclDataType. It supports mapping for various ggml_types. If the input type
+ *			does not match any of the predefined ggml_types, the function returns
+ *          ACL_DT_UNDEFINED.
+ *
+ * @param	type    The ggml_type to be mapped.
+ * @return	The corresponding aclDataType. If the input type is not recognized,
+ *			ACL_DT_UNDEFINED is returned.
+ */
+aclDataType ggml_cann_type_mapping(ggml_type type);
+
+/**
+ * @brief   Creates an ACL tensor from a ggml_tensor with optional shape.
+ *
+ * @details This function creates an ACL tensor based on the properties of the
+ *          provided ggml_tensor. It supports customer shape by adjusting dimensions
+ *          and strides accordingly. If customer shape is applied, additional
+ *          dimensions and strides are calculated based on the provided parameters.
+ *
+ * @param   tensor      Pointer to the ggml_tensor to be converted to ACL tensor.
+ * @param   ne          Pointer to an array containing dimensions. Defaults to nullptr
+ *                      if no customer shape is applied.
+ * @param   nb          Pointer to an array containing strides. Defaults to nullptr
+ *                      if no customer shape is applied.
+ * @param   dims        Number of dimensions in the tensor. Defaults to 0 if no customer
+ *                      shape is applied.
+ * @param   format      ACL tensor format. Defaults to ACL_FORMAT_ND.
+ * @param   offset      Offset in bytes for the ACL tensor data. Defaults to 0.
+ * @return  Pointer to the created ACL tensor.
+ */
+aclTensor* ggml_cann_create_tensor(const ggml_tensor* tensor, int64_t* ne = nullptr,
+                             size_t* nb = nullptr, int64_t dims = 0,
+                             aclFormat format = ACL_FORMAT_ND,
+                             size_t offset = 0);
+
+/**
+ * @brief   Template for creating an ACL tensor from provided parameters. typename TYPE
+ *          should be size_t or float.
+ *
+ * @details This function creates an ACL tensor using the provided data pointer,
+ *          data type, dimensions, strides, format, offset, and additional parameters.
+ *          It calculates necessary dimensions and strides based on the provided ne and nb
+ *          arrays, adjusting them for the ACL tensor creation. The ACL storage length
+ *          is also calculated based on the provided dimensions and strides.
+ *
+ * @param   data_ptr    Pointer to the data buffer for the ACL tensor.
+ * @param   dtype       ACL data type of the tensor.
+ * @param   type_size   Size of each element in the tensor data buffer.
+ * @param   ne          Pointer to an array containing tensor dimensions.
+ * @param   nb          Pointer to an array containing tensor strides.
+ * @param   dims        Number of dimensions of the tensor.
+ * @param   format      ACL tensor format. Defaults to ACL_FORMAT_ND.
+ * @param   offset      Offset in bytes for the ACL tensor data. Defaults to 0.
+ * @return  Pointer to the created ACL tensor.
+ */
+template<typename TYPE>
+aclTensor* ggml_cann_create_tensor(void* data_ptr, aclDataType dtype,
+                                   TYPE type_size, int64_t* ne, TYPE* nb,
+                                   int64_t dims,
+                                   aclFormat format = ACL_FORMAT_ND,
+                                   size_t offset = 0) {
+    int64_t tmp_ne[GGML_MAX_DIMS * 2];
+    int64_t tmp_stride[GGML_MAX_DIMS * 2];
+
+    memcpy(tmp_ne, ne, dims * sizeof(int64_t));
+    for (int i = 0; i < dims; i++) {
+        tmp_stride[i] = nb[i] / type_size;
+    }
+
+    std::reverse(tmp_ne, tmp_ne + dims);
+    std::reverse(tmp_stride, tmp_stride + dims);
+
+    int64_t acl_storage_len = 0;
+    for (int i = 0; i < dims; i++) {
+        acl_storage_len += (ne[i] - 1) * nb[i];
+    }
+
+    aclTensor* acl_tensor =
+        aclCreateTensor(tmp_ne, dims, dtype, tmp_stride, offset / type_size,
+                        format, &acl_storage_len, 1, data_ptr);
+
+    return acl_tensor;
+}
+
+/**
+ * @brief   Checks if tensors require broadcasting based on their shapes.
+ *
+ * @details This function determines if two ggml_tensors need to be broadcasted for
+ *          element-wise operations. Broadcasting is necessary if the shapes of the
+ *          tensors are not identical and no dimension in either tensor equals 1.
+ *
+ * @param   t0      Pointer to the first ggml_tensor.
+ * @param   t1      Pointer to the second ggml_tensor.
+ * @return  True if broadcasting is needed, False otherwise.
+ *
+ * @remarks This function iterates over the dimensions of t0 and t1. It checks if each
+ *          dimension in t1 differs from t0's corresponding dimension and is not equal
+ *          to 1. If such a dimension is found, broadcasting is required to align t1
+ *          with t0 for element-wise operations.
+ */
+bool ggml_cann_need_bcast(const ggml_tensor* t0, const ggml_tensor* t1);
+
+/**
+ * @brief   Computes broadcast shapes and strides for two ggml_tensors.
+ *
+ * @details This function calculates the broadcast shapes and strides for two ggml_tensors,
+ *          following the broadcasting rules similar to numpy. It adjusts dimensions and
+ *          strides to ensure compatibility for element-wise operations where one tensor
+ *          can be broadcasted to match the shape of another tensor.
+ *
+ * @param   src0                Pointer to the first ggml_tensor.
+ * @param   src1                Pointer to the second ggml_tensor.
+ * @param   bcast_ne_src0       Output array to store broadcasted dimensions for src0.
+ * @param   bcast_ne_src1       Output array to store broadcasted dimensions for src1.
+ * @param   bcast_nb_src0       Output array to store broadcasted strides for src0.
+ * @param   bcast_nb_src1       Output array to store broadcasted strides for src1.
+ * @return  Number of dimensions in the broadcasted shape.
+ *
+ * @pre     ggml_can_repeat(src1, src0) must return true, indicating src1 can be broadcasted
+ *          to match src0.
+ *
+ * @remarks This function iterates over the dimensions of src0 and src1, calculating the
+ *          necessary broadcast dimensions and strides. If a dimension requires broadcasting
+ *          (i.e., its size in src1 is smaller than in src0), an additional dimension is
+ *          added with size calculated to match src0's dimension. This adjustment ensures
+ *          that src1 can be element-wise broadcasted to src0's shape.
+ *
+ *  How it works:
+ *
+ *  if dim0 has padding.
+ *  a -> (2, 2) padding = 2
+ *   a: [[1, 2, *, *]
+ *       [2, 3, *, *]]
+ *  nb = (8, 4, 2)
+ *
+ *  if a should bcast with b -> (2, 4)
+ *  b' -> (2, 2, 2)
+ *  b : [[1, 2, 3, 4, *, *]
+ *       [5, 6, 7, 8, *, *]]
+ *  nb = (12, 6, 1)
+ *
+ *  after bcast:
+ *  a' -> (2, 1, 2)
+ *  a': [[[1, 2], *, *]
+ *       [[2, 3], *, *]]
+ *  nb = (8, 4, 2, 1)
+ *
+ *  b' : [[[1, 2], [3, 4], *, *]
+ *        [[5, 6], [7, 8], *, *]]
+ *  nb = (12, 6, 2, 1)
+ *  \endcode
+ *
+ *  dim1 in a inserted dim, should add nb for dim1,
+ *  and all other nb moves to next in order.
+ */
+int64_t ggml_cann_get_bcast_shape(const ggml_tensor* src0, const ggml_tensor* src1,
+                        int64_t* bcast_ne_src0, int64_t* bcast_ne_src1,
+                        size_t* bcast_nb_src0, size_t* bcast_nb_src1);
+
+// Bcast macro to avoid duplicate code.
+#define BCAST_SHAPE(src0, src1)                                              \
+    int64_t bcast_##src0##_ne[GGML_MAX_DIMS * 2];                            \
+    int64_t bcast_##src1##_ne[GGML_MAX_DIMS * 2];                            \
+    size_t bcast_##src0##_nb[GGML_MAX_DIMS * 2];                             \
+    size_t bcast_##src1##_nb[GGML_MAX_DIMS * 2];                             \
+    int64_t bcast_dims = ggml_cann_get_bcast_shape(                          \
+        src0, src1, bcast_##src0##_ne, bcast_##src1##_ne, bcast_##src0##_nb, \
+        bcast_##src1##_nb);
+
+#define BCAST_PARAM(tensor) bcast_##tensor##_ne, bcast_##tensor##_nb, bcast_dims
+
+/**
+ * @brief Calculates broadcast shapes for matrix multiplication.
+ *
+ * @details This function computes the broadcast shapes required for matrix multiplication
+ *          based on the input, weight, and destination tensor shapes. It ensures that the
+ *          dimensions of weight tensors are expanded appropriately to satisfy matrix
+ *          multiplication broadcast rules.
+ *
+ * @param input_ne      Array containing the dimensions of the input tensor.
+ * @param weight_ne     Array containing the dimensions of the weight tensor.
+ * @param dst_ne        Array containing the dimensions of the destination tensor.
+ * @param input_nb      Array containing the strides of the input tensor.
+ * @param weight_nb     Array containing the strides of the weight tensor.
+ * @param dst_nb        Array containing the strides of the destination tensor.
+ * @param bcast_input_ne    Output array for broadcasted input tensor dimensions.
+ * @param bcast_weight_ne   Output array for broadcasted weight tensor dimensions.
+ * @param bcast_dst_ne      Output array for broadcasted destination tensor dimensions.
+ * @param bcast_input_nb    Output array for broadcasted input tensor strides.
+ * @param bcast_weight_nb   Output array for broadcasted weight tensor strides.
+ * @param bcast_dst_nb      Output array for broadcasted destination tensor strides.
+ * @return The number of dimensions in the broadcasted tensors.
+ *
+ * @remarks This function iterates over the tensor dimensions and calculates the broadcast
+ *          shapes needed for matrix multiplication. It ensures that dimensions where
+ *          weight tensor requires expansion are appropriately handled to conform with
+ *          broadcasting rules.
+ * @note compare with ggml_cann_get_bcast_shape, mul_mat broadcast need add this new dim
+ *       before cast dim.
+ * @sa ggml_cann_get_bcast_shape
+ */
+int64_t ggml_cann_get_mulmat_bcast_shape(
+    const int64_t* input_ne, const int64_t* weight_ne, const int64_t* dst_ne,
+    const size_t* input_nb, const size_t* weight_nb, const size_t* dst_nb,
+    int64_t* bcast_input_ne, int64_t* bcast_weight_ne, int64_t* bcast_dst_ne,
+    size_t* bcast_input_nb, size_t* bcast_weight_nb, size_t* bcast_dst_nb);
+
+// Bcast macro to avoid duplicate code.
+#define BCAST_MUL_MAT_SHAPE(input, weight, dst)                         \
+    int64_t bcast_##input##_ne[GGML_MAX_DIMS * 2];                      \
+    int64_t bcast_##weight##_ne[GGML_MAX_DIMS * 2];                     \
+    int64_t bcast_##dst##_ne[GGML_MAX_DIMS * 2];                        \
+    size_t bcast_##input##_nb[GGML_MAX_DIMS * 2];                       \
+    size_t bcast_##weight##_nb[GGML_MAX_DIMS * 2];                      \
+    size_t bcast_##dst##_nb[GGML_MAX_DIMS * 2];                         \
+    int64_t bcast_dims = ggml_cann_get_mulmat_bcast_shape(              \
+        input->ne, weight->ne, dst->ne, input->nb, weight->nb, dst->nb, \
+        bcast_##input##_ne, bcast_##weight##_ne, bcast_##dst##_ne,      \
+        bcast_##input##_nb, bcast_##weight##_nb, bcast_##dst##_nb);
+
+#define BCAST_MUL_MAT_PARAM(tensor) \
+    bcast_##tensor##_ne, bcast_##tensor##_nb, bcast_dims
+
+#endif  // CANN_ACL_TENSOR_H
diff --git a/src/whisper_cpp/ggml/src/ggml-cann/aclnn_ops.cpp b/src/whisper_cpp/ggml/src/ggml-cann/aclnn_ops.cpp
new file mode 100644
index 00000000..a4ec8418
--- /dev/null
+++ b/src/whisper_cpp/ggml/src/ggml-cann/aclnn_ops.cpp
@@ -0,0 +1,3082 @@
+/*
+ * Copyright (c) 2023-2024 The ggml authors
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a copy
+ * of this software and associated documentation files (the "Software"), to
+ * deal in the Software without restriction, including without limitation the
+ * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or
+ * sell copies of the Software, and to permit persons to whom the Software is
+ * furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice shall be included in
+ * all copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+ * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
+ * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
+ * IN THE SOFTWARE.
+ */
+
+#include "aclnn_ops.h"
+
+#include <aclnnop/aclnn_avgpool2d.h>
+#include <aclnnop/aclnn_cast.h>
+#include <aclnnop/aclnn_constant_pad_nd.h>
+#include <aclnnop/aclnn_copy.h>
+#include <aclnnop/aclnn_cos.h>
+#include <aclnnop/aclnn_exp.h>
+#include <aclnnop/aclnn_fill_scalar.h>
+#include <aclnnop/aclnn_group_norm.h>
+#include <aclnnop/aclnn_index_fill_tensor.h>
+#include <aclnnop/aclnn_layer_norm.h>
+#include <aclnnop/aclnn_matmul.h>
+#include <aclnnop/aclnn_max_pool.h>
+#include <aclnnop/aclnn_permute.h>
+#include <aclnnop/aclnn_pow_tensor_tensor.h>
+#include <aclnnop/aclnn_reduce_sum.h>
+#include <aclnnop/aclnn_repeat.h>
+#include <aclnnop/aclnn_repeat_interleave.h>
+#include <aclnnop/aclnn_roll.h>
+#include <aclnnop/aclnn_sin.h>
+#include <aclnnop/aclnn_softmax.h>
+#include <aclnnop/aclnn_tril.h>
+#include <aclnnop/aclnn_triu.h>
+#include <aclnnop/aclnn_upsample_nearest_2d.h>
+#include <aclnnop/aclnn_weight_quant_batch_matmul_v2.h>
+#include <float.h>
+
+#include <cmath>
+#include <cstring>
+#include <exception>
+#include <vector>
+
+#include "kernels/ascendc_kernels.h"
+
+#define GGML_COMMON_DECL_C
+
+#include "../ggml-common.h"
+
+/**
+ * @brief Repeats elements of a tensor along each dimension according to the
+ * specified repeat array.
+ *
+ * @param ctx The context for the CANN backend operations.
+ * @param acl_src The source tensor to be repeated.
+ * @param acl_dst The destination tensor after repeating.
+ * @param repeat_array The array specifying the number of repetitions along each
+ * dimension.
+ */
+static void aclnn_repeat(ggml_backend_cann_context& ctx, aclTensor* acl_src,
+                         aclTensor* acl_dst, int64_t* repeat_array) {
+    // repeat tensor along each dim with repeat_array
+    aclIntArray* repeats = aclCreateIntArray(repeat_array, GGML_MAX_DIMS);
+
+    uint64_t workspaceSize = 0;
+    aclOpExecutor* executor;
+    void* workspaceAddr = nullptr;
+
+    ACL_CHECK(aclnnRepeatGetWorkspaceSize(acl_src, repeats, acl_dst,
+                                          &workspaceSize, &executor));
+
+    if (workspaceSize > 0) {
+        // Memory from allocator will "free" immediately, and this memory
+        // will be alloced to other pointers, but it won't access before
+        // this async task end because all tasks in same stream will execute
+        // in queue.
+        ggml_cann_pool_alloc workspace_allocator(ctx.pool(), workspaceSize);
+        workspaceAddr = workspace_allocator.get();
+    }
+    ACL_CHECK(
+        aclnnRepeat(workspaceAddr, workspaceSize, executor, ctx.stream()));
+    ACL_CHECK(aclDestroyIntArray(repeats));
+}
+
+void ggml_cann_repeat(ggml_backend_cann_context& ctx, ggml_tensor* dst) {
+    ggml_tensor* src = dst->src[0];
+    GGML_ASSERT(ggml_can_repeat(src, dst));
+
+    aclTensor* acl_src = ggml_cann_create_tensor(src);
+    aclTensor* acl_dst = ggml_cann_create_tensor(dst);
+
+    int64_t repeatsArray[] = {dst->ne[3] / src->ne[3], dst->ne[2] / src->ne[2],
+                              dst->ne[1] / src->ne[1], dst->ne[0] / src->ne[0]};
+
+    aclnn_repeat(ctx, acl_src, acl_dst, repeatsArray);
+    ACL_CHECK(aclDestroyTensor(acl_src));
+    ACL_CHECK(aclDestroyTensor(acl_dst));
+}
+
+/**
+ * @brief Adds two tensors element-wise and stores the result in a destination
+ * tensor.
+ *
+ * This function performs the operation:
+ * \f[
+ *    dst = acl\_src0 + alpha \times acl\_src1
+ * \f]
+ * where alpha is a scalar value and defaults to 1.0f.
+ *
+ * @param ctx The context for the CANN backend operations.
+ * @param acl_src0 The first source tensor.
+ * @param acl_src1 The second source tensor.
+ * @param acl_dst The destination tensor where the result will be stored.
+ */
+static void aclnn_add(ggml_backend_cann_context& ctx, aclTensor* acl_src0,
+                      aclTensor* acl_src1, aclTensor* acl_dst) {
+    aclScalar* alpha = nullptr;
+    float alphaValue = 1.0f;
+    alpha = aclCreateScalar(&alphaValue, aclDataType::ACL_FLOAT);
+
+    uint64_t workspaceSize = 0;
+    aclOpExecutor* executor;
+    void* workspaceAddr = nullptr;
+
+    ACL_CHECK(aclnnAddGetWorkspaceSize(acl_src0, acl_src1, alpha, acl_dst,
+                                       &workspaceSize, &executor));
+    if (workspaceSize > 0) {
+        ggml_cann_pool_alloc workspace_allocator(ctx.pool(), workspaceSize);
+        workspaceAddr = workspace_allocator.get();
+    }
+
+    ACL_CHECK(aclnnAdd(workspaceAddr, workspaceSize, executor, ctx.stream()));
+
+    ACL_CHECK(aclDestroyScalar(alpha));
+}
+
+void ggml_cann_add(ggml_backend_cann_context& ctx, ggml_tensor* dst) {
+    ggml_tensor* src0 = dst->src[0];
+    ggml_tensor* src1 = dst->src[1];
+    GGML_ASSERT(ggml_can_repeat(src1, src0) && ggml_are_same_shape(src0, dst));
+
+    aclTensor* acl_src0;
+    aclTensor* acl_src1;
+    aclTensor* acl_dst;
+
+    // Need bcast
+    if (!ggml_are_same_shape(src0, src1) && ggml_cann_need_bcast(src0, src1)) {
+        BCAST_SHAPE(src0, src1)
+        acl_src0 = ggml_cann_create_tensor(src0, BCAST_PARAM(src0));
+        acl_src1 = ggml_cann_create_tensor(src1, BCAST_PARAM(src1));
+        acl_dst = ggml_cann_create_tensor(dst, BCAST_PARAM(src0));
+    } else {
+        acl_src0 = ggml_cann_create_tensor(src0);
+        acl_src1 = ggml_cann_create_tensor(src1);
+        acl_dst = ggml_cann_create_tensor(dst);
+    }
+
+    aclnn_add(ctx, acl_src0, acl_src1, acl_dst);
+
+    ACL_CHECK(aclDestroyTensor(acl_src0));
+    ACL_CHECK(aclDestroyTensor(acl_src1));
+    ACL_CHECK(aclDestroyTensor(acl_dst));
+}
+
+void ggml_cann_leaky_relu(ggml_backend_cann_context& ctx, ggml_tensor* dst) {
+    ggml_tensor* src = dst->src[0];
+
+    GGML_ASSERT(src->type == GGML_TYPE_F32);
+    GGML_ASSERT(dst->type == GGML_TYPE_F32);
+
+    aclTensor* acl_src = ggml_cann_create_tensor(src);
+    aclTensor* acl_dst = ggml_cann_create_tensor(dst);
+
+    float negative_slope;
+    memcpy(&negative_slope, dst->op_params, sizeof(float));
+    aclScalar* acl_negative_slope =
+        aclCreateScalar(&negative_slope, aclDataType::ACL_FLOAT);
+
+    uint64_t workspaceSize = 0;
+    aclOpExecutor* executor;
+    void* workspaceAddr = nullptr;
+
+    ACL_CHECK(aclnnLeakyReluGetWorkspaceSize(
+        acl_src, acl_negative_slope, acl_dst, &workspaceSize, &executor));
+    if (workspaceSize > 0) {
+        ggml_cann_pool_alloc workspace_allocator(ctx.pool(), workspaceSize);
+        workspaceAddr = workspace_allocator.get();
+    }
+
+    ACL_CHECK(
+        aclnnLeakyRelu(workspaceAddr, workspaceSize, executor, ctx.stream()));
+
+    ACL_CHECK(aclDestroyScalar(acl_negative_slope));
+    ACL_CHECK(aclDestroyTensor(acl_src));
+    ACL_CHECK(aclDestroyTensor(acl_dst));
+}
+
+/**
+ * @brief Concatenates a list of tensors along a specified dimension and stores
+ * the result in a destination tensor.
+ *
+ * @param ctx The context for the CANN backend operations.
+ * @param tensorList The list of tensors to be concatenated.
+ * @param acl_dst The destination tensor where the concatenated result will be
+ * stored.
+ * @param concat_dim The dimension along which the tensors will be concatenated.
+ */
+static void aclnn_concat(ggml_backend_cann_context& ctx,
+                         aclTensorList* tensorList, aclTensor* acl_dst,
+                         int64_t concat_dim) {
+    uint64_t workspaceSize = 0;
+    aclOpExecutor* executor;
+    void* workspaceAddr = nullptr;
+
+    ACL_CHECK(aclnnCatGetWorkspaceSize(tensorList, concat_dim, acl_dst,
+                                       &workspaceSize, &executor));
+    if (workspaceSize > 0) {
+        ggml_cann_pool_alloc workspace_allocator(ctx.pool(), workspaceSize);
+        workspaceAddr = workspace_allocator.get();
+    }
+
+    ACL_CHECK(aclnnCat(workspaceAddr, workspaceSize, executor, ctx.stream()));
+}
+
+void ggml_cann_concat(ggml_backend_cann_context& ctx, ggml_tensor* dst) {
+    ggml_tensor* src0 = dst->src[0];
+    ggml_tensor* src1 = dst->src[1];
+    aclTensor* acl_src0 = ggml_cann_create_tensor(src0);
+    aclTensor* acl_src1 = ggml_cann_create_tensor(src1);
+    aclTensor* acl_dst = ggml_cann_create_tensor(dst);
+
+    int64_t concat_dim = 1;
+    aclTensor* tensors[] = {acl_src0, acl_src1};
+    aclTensorList* tensorList = aclCreateTensorList(tensors, 2);
+    aclnn_concat(ctx, tensorList, acl_dst, concat_dim);
+
+    ACL_CHECK(aclDestroyTensorList(tensorList));
+    ACL_CHECK(aclDestroyTensor(acl_dst));
+}
+
+/**
+ * @brief Creates a tensor with values starting from `start`, incremented by
+ * `step`, and ending before `stop`.
+ *
+ * This function performs the operation:
+ * \f[
+ *    \text {out }_{i+1}=\text {out }_i+\text {step}
+ * \f]
+ * the range is [start, stop).
+ *
+ * @param ctx The context for the CANN backend operations.
+ * @param acl_dst The destination tensor where the values will be stored.
+ * @param start The starting value of the range.
+ * @param stop The ending value of the range (exclusive).
+ * @param step The step size between consecutive values.
+ * @param n_elements The number of elements in the destination tensor.
+ */
+static void aclnn_arange(ggml_backend_cann_context& ctx, aclTensor* acl_dst,
+                         float start, float stop, float step,
+                         int64_t n_elements) {
+    int64_t steps = (int64_t)std::ceil((stop - start) / step);
+    GGML_ASSERT(n_elements == steps);
+
+    uint64_t workspaceSize = 0;
+    aclOpExecutor* executor;
+    void* workspaceAddr = nullptr;
+
+    aclScalar* acl_start = aclCreateScalar(&start, aclDataType::ACL_FLOAT);
+    aclScalar* acl_end = aclCreateScalar(&stop, aclDataType::ACL_FLOAT);
+    aclScalar* acl_step = aclCreateScalar(&step, aclDataType::ACL_FLOAT);
+
+    ACL_CHECK(aclnnArangeGetWorkspaceSize(acl_start, acl_end, acl_step, acl_dst,
+                                          &workspaceSize, &executor));
+    if (workspaceSize > 0) {
+        ggml_cann_pool_alloc workspace_allocator(ctx.pool(), workspaceSize);
+        workspaceAddr = workspace_allocator.get();
+    }
+
+    ACL_CHECK(
+        aclnnArange(workspaceAddr, workspaceSize, executor, ctx.stream()));
+
+    ACL_CHECK(aclDestroyScalar(acl_start));
+    ACL_CHECK(aclDestroyScalar(acl_end));
+    ACL_CHECK(aclDestroyScalar(acl_step));
+}
+
+void ggml_cann_arange(ggml_backend_cann_context& ctx, ggml_tensor* dst) {
+    GGML_ASSERT(dst->type == GGML_TYPE_F32);
+
+    aclTensor* acl_dst = ggml_cann_create_tensor(dst);
+
+    int64_t n_elements = ggml_nelements(dst);
+    float start;
+    float stop;
+    float step;
+    memcpy(&start, (float*)dst->op_params + 0, sizeof(float));
+    memcpy(&stop, (float*)dst->op_params + 1, sizeof(float));
+    memcpy(&step, (float*)dst->op_params + 2, sizeof(float));
+
+    aclnn_arange(ctx, acl_dst, start, stop, step, n_elements);
+    ACL_CHECK(aclDestroyTensor(acl_dst));
+}
+
+void ggml_cann_sqr(ggml_backend_cann_context& ctx, ggml_tensor* dst) {
+    dst->src[1] = dst->src[0];
+    ggml_cann_mul_div<aclnnMulGetWorkspaceSize, aclnnMul>(ctx, dst);
+}
+
+void ggml_cann_clamp(ggml_backend_cann_context& ctx, ggml_tensor* dst) {
+    ggml_tensor* src = dst->src[0];
+    GGML_ASSERT(src->type == GGML_TYPE_F32);
+    GGML_ASSERT(dst->type == GGML_TYPE_F32);
+
+    float min;
+    float max;
+    memcpy(&min, dst->op_params, sizeof(float));
+    memcpy(&max, (float*)dst->op_params + 1, sizeof(float));
+
+    aclTensor* acl_src = ggml_cann_create_tensor(src);
+    aclTensor* acl_dst = ggml_cann_create_tensor(dst);
+
+    aclScalar* acl_min = aclCreateScalar(&min, aclDataType::ACL_FLOAT);
+    aclScalar* acl_max = aclCreateScalar(&max, aclDataType::ACL_FLOAT);
+
+    uint64_t workspaceSize = 0;
+    aclOpExecutor* executor;
+    void* workspaceAddr = nullptr;
+
+    ACL_CHECK(aclnnClampGetWorkspaceSize(acl_src, acl_min, acl_max, acl_dst,
+                                         &workspaceSize, &executor));
+    if (workspaceSize > 0) {
+        ggml_cann_pool_alloc workspace_allocator(ctx.pool(), workspaceSize);
+        workspaceAddr = workspace_allocator.get();
+    }
+
+    ACL_CHECK(aclnnClamp(workspaceAddr, workspaceSize, executor, ctx.stream()));
+
+    ACL_CHECK(aclDestroyScalar(acl_min));
+    ACL_CHECK(aclDestroyScalar(acl_max));
+    ACL_CHECK(aclDestroyTensor(acl_src));
+    ACL_CHECK(aclDestroyTensor(acl_dst));
+}
+
+void ggml_cann_scale(ggml_backend_cann_context& ctx, ggml_tensor* dst) {
+    ggml_tensor* src = dst->src[0];
+
+    // scale factor
+    float v;
+    memcpy(&v, dst->op_params, sizeof(float));
+
+    aclScalar* scale = aclCreateScalar(&v, aclDataType::ACL_FLOAT);
+    aclTensor* acl_src = ggml_cann_create_tensor(src);
+    aclTensor* acl_dst = ggml_cann_create_tensor(dst);
+
+    uint64_t workspaceSize = 0;
+    aclOpExecutor* executor;
+    void* workspaceAddr = nullptr;
+
+    ACL_CHECK(aclnnMulsGetWorkspaceSize(acl_src, scale, acl_dst, &workspaceSize,
+                                        &executor));
+    if (workspaceSize > 0) {
+        ggml_cann_pool_alloc workspace_allocator(ctx.pool(), workspaceSize);
+        workspaceAddr = workspace_allocator.get();
+    }
+
+    ACL_CHECK(aclnnMuls(workspaceAddr, workspaceSize, executor, ctx.stream()));
+
+    ACL_CHECK(aclDestroyScalar(scale));
+    ACL_CHECK(aclDestroyTensor(acl_src));
+    ACL_CHECK(aclDestroyTensor(acl_dst));
+}
+
+void ggml_cann_argsort(ggml_backend_cann_context& ctx, ggml_tensor* dst) {
+    ggml_tensor* src = dst->src[0];
+    enum ggml_sort_order order = (enum ggml_sort_order)dst->op_params[0];
+
+    aclTensor* acl_src = ggml_cann_create_tensor(src);
+    aclTensor* acl_dst = ggml_cann_create_tensor(dst);
+    ggml_cann_pool_alloc temp_buffer_allocator(
+        ctx.pool(), ggml_nelements(dst) * sizeof(int64_t));
+    void* buffer = temp_buffer_allocator.get();
+    aclTensor* tmp_tensor =
+        ggml_cann_create_tensor(buffer, ACL_INT64, ggml_type_size(dst->type),
+                                dst->ne, dst->nb, GGML_MAX_DIMS);
+
+    uint64_t workspaceSize = 0;
+    aclOpExecutor* executor;
+    void* workspaceAddr = nullptr;
+
+    ACL_CHECK(aclnnArgsortGetWorkspaceSize(
+        acl_src, -1, (order == GGML_SORT_ORDER_DESC ? true : false), tmp_tensor,
+        &workspaceSize, &executor));
+    if (workspaceSize > 0) {
+        ggml_cann_pool_alloc workspace_allocator(ctx.pool(), workspaceSize);
+        workspaceAddr = workspace_allocator.get();
+    }
+
+    ACL_CHECK(
+        aclnnArgsort(workspaceAddr, workspaceSize, executor, ctx.stream()));
+
+    workspaceSize = 0;
+    ACL_CHECK(aclnnCastGetWorkspaceSize(tmp_tensor,
+                                        ggml_cann_type_mapping(dst->type),
+                                        acl_dst, &workspaceSize, &executor));
+    if (workspaceSize > 0) {
+        ggml_cann_pool_alloc workspace_allocator(ctx.pool(), workspaceSize);
+        workspaceAddr = workspace_allocator.get();
+    }
+
+    ACL_CHECK(aclnnCast(workspaceAddr, workspaceSize, executor, ctx.stream()));
+
+    ACL_CHECK(aclDestroyTensor(acl_src));
+    ACL_CHECK(aclDestroyTensor(tmp_tensor));
+    ACL_CHECK(aclDestroyTensor(acl_dst));
+}
+
+void ggml_cann_norm(ggml_backend_cann_context& ctx, ggml_tensor* dst) {
+    ggml_tensor* src = dst->src[0];
+
+    aclTensor* acl_src = ggml_cann_create_tensor(src);
+    aclTensor* acl_dst = ggml_cann_create_tensor(dst);
+
+    float eps;
+    memcpy(&eps, dst->op_params, sizeof(float));
+
+    uint64_t workspaceSize = 0;
+    aclOpExecutor* executor;
+    void* workspaceAddr = nullptr;
+
+    std::vector<int64_t> normData = {dst->ne[0]};
+    aclIntArray* norm = aclCreateIntArray(normData.data(), normData.size());
+    ACL_CHECK(aclnnLayerNormGetWorkspaceSize(acl_src, norm, nullptr, nullptr,
+                                             eps, acl_dst, nullptr, nullptr,
+                                             &workspaceSize, &executor));
+
+    if (workspaceSize > 0) {
+        ggml_cann_pool_alloc workspace_allocator(ctx.pool(), workspaceSize);
+        workspaceAddr = workspace_allocator.get();
+    }
+
+    ACL_CHECK(
+        aclnnLayerNorm(workspaceAddr, workspaceSize, executor, ctx.stream()));
+
+    ACL_CHECK(aclDestroyIntArray(norm));
+    ACL_CHECK(aclDestroyTensor(acl_src));
+    ACL_CHECK(aclDestroyTensor(acl_dst));
+}
+
+void ggml_cann_group_norm(ggml_backend_cann_context& ctx, ggml_tensor* dst) {
+    ggml_tensor* src = dst->src[0];
+
+    aclTensor* acl_src = ggml_cann_create_tensor(src);
+    aclTensor* acl_dst = ggml_cann_create_tensor(dst);
+
+    int n_groups = dst->op_params[0];
+
+    float eps;
+    memcpy(&eps, dst->op_params + 1, sizeof(float));
+
+    uint64_t workspaceSize = 0;
+    aclOpExecutor* executor;
+    void* workspaceAddr = nullptr;
+
+    int64_t N = src->ne[3];
+    int64_t C = src->ne[2];
+    int64_t HxW = src->ne[1] * src->ne[0];
+
+    size_t type_size = ggml_type_size(src->type);
+    int64_t ne[] = {n_groups, N};
+    size_t nb[] = {type_size, type_size * n_groups};
+    size_t n_bytes = N * n_groups;
+
+    ggml_cann_pool_alloc temp_buffer_allocator(ctx.pool(), n_bytes * 2);
+    void* buffer = temp_buffer_allocator.get();
+    aclTensor* acl_mean_out = ggml_cann_create_tensor(
+        buffer, ACL_FLOAT, type_size, ne, nb, ACL_FORMAT_ND);
+    aclTensor* acl_rstd_out = ggml_cann_create_tensor(
+        (char*)buffer + n_bytes, ACL_FLOAT, type_size, ne, nb, ACL_FORMAT_ND);
+
+    ACL_CHECK(aclnnGroupNormGetWorkspaceSize(
+        acl_src, nullptr, nullptr, N, C, HxW, n_groups, eps, acl_dst,
+        acl_mean_out, acl_rstd_out, &workspaceSize, &executor));
+
+    if (workspaceSize > 0) {
+        ggml_cann_pool_alloc workspace_allocator(ctx.pool(), workspaceSize);
+        workspaceAddr = workspace_allocator.get();
+    }
+
+    ACL_CHECK(
+        aclnnGroupNorm(workspaceAddr, workspaceSize, executor, ctx.stream()));
+
+    ACL_CHECK(aclDestroyTensor(acl_src));
+    ACL_CHECK(aclDestroyTensor(acl_dst));
+    ACL_CHECK(aclDestroyTensor(acl_mean_out));
+    ACL_CHECK(aclDestroyTensor(acl_rstd_out));
+}
+
+void ggml_cann_acc(ggml_backend_cann_context& ctx, ggml_tensor* dst) {
+    ggml_tensor* src0 = dst->src[0];
+    ggml_tensor* src1 = dst->src[1];
+
+    size_t nb1 = ((int32_t*)dst->op_params)[0];
+    size_t nb2 = ((int32_t*)dst->op_params)[1];
+    size_t nb3 = ((int32_t*)dst->op_params)[2];
+    size_t offset = ((int32_t*)dst->op_params)[3];
+    bool inplace = (bool)((int32_t*)dst->op_params)[4];
+
+    size_t param_nb[] = {ggml_element_size(src0), nb1, nb2, nb3};
+
+    aclTensor* acl_dst = ggml_cann_create_tensor(
+        dst, src1->ne, param_nb, GGML_MAX_DIMS, ACL_FORMAT_ND, offset);
+    aclTensor* acl_src1 = ggml_cann_create_tensor(src1);
+
+    aclScalar* alpha = nullptr;
+    float alphaValue = 1.0f;
+    alpha = aclCreateScalar(&alphaValue, aclDataType::ACL_FLOAT);
+
+    uint64_t workspaceSize = 0;
+    aclOpExecutor* executor;
+    void* workspaceAddr = nullptr;
+
+    if (!inplace) {
+        size_t cpy_size = ggml_nbytes(dst);
+        ACL_CHECK(aclrtMemcpyAsync(dst->data, cpy_size, src0->data, cpy_size,
+                                   ACL_MEMCPY_DEVICE_TO_DEVICE, ctx.stream()));
+        aclTensor* acl_src0 = ggml_cann_create_tensor(
+            src0, src1->ne, src0->nb, GGML_MAX_DIMS, ACL_FORMAT_ND, offset);
+        ACL_CHECK(aclnnAddGetWorkspaceSize(acl_src0, acl_src1, alpha, acl_dst,
+                                           &workspaceSize, &executor));
+        if (workspaceSize > 0) {
+            ggml_cann_pool_alloc workspace_allocator(ctx.pool(), workspaceSize);
+            workspaceAddr = workspace_allocator.get();
+        }
+        ACL_CHECK(
+            aclnnAdd(workspaceAddr, workspaceSize, executor, ctx.stream()));
+        ACL_CHECK(aclDestroyTensor(acl_src0));
+    } else {
+        ACL_CHECK(aclnnInplaceAddGetWorkspaceSize(acl_dst, acl_src1, alpha,
+                                                  &workspaceSize, &executor));
+        if (workspaceSize > 0) {
+            ggml_cann_pool_alloc workspace_allocator(ctx.pool(), workspaceSize);
+            workspaceAddr = workspace_allocator.get();
+        }
+        ACL_CHECK(aclnnInplaceAdd(workspaceAddr, workspaceSize, executor,
+                                  ctx.stream()));
+    }
+
+    ACL_CHECK(aclDestroyTensor(acl_src1));
+    ACL_CHECK(aclDestroyTensor(acl_dst));
+}
+
+void ggml_cann_sum_rows(ggml_backend_cann_context& ctx, ggml_tensor* dst) {
+    ggml_tensor* src = dst->src[0];
+
+    aclTensor* acl_src = ggml_cann_create_tensor(src);
+
+    GGML_ASSERT(dst->ne[0] == 1);
+    aclTensor* acl_dst = ggml_cann_create_tensor(dst);
+
+    int64_t reduce_dims_host[] = {3};
+    aclIntArray* reduce_dims = aclCreateIntArray(reduce_dims_host, 1);
+
+    uint64_t workspaceSize = 0;
+    aclOpExecutor* executor;
+    void* workspaceAddr = nullptr;
+
+    ACL_CHECK(aclnnReduceSumGetWorkspaceSize(
+        acl_src, reduce_dims, true, ggml_cann_type_mapping(src->type), acl_dst,
+        &workspaceSize, &executor));
+    if (workspaceSize > 0) {
+        ggml_cann_pool_alloc workspace_allocator(ctx.pool(), workspaceSize);
+        workspaceAddr = workspace_allocator.get();
+    }
+
+    ACL_CHECK(
+        aclnnReduceSum(workspaceAddr, workspaceSize, executor, ctx.stream()));
+
+    ACL_CHECK(aclDestroyTensor(acl_src));
+    ACL_CHECK(aclDestroyTensor(acl_dst));
+}
+
+void ggml_cann_upsample_nearest2d(ggml_backend_cann_context& ctx,
+                                  ggml_tensor* dst) {
+    ggml_tensor* src = dst->src[0];
+    aclTensor* acl_src =
+        ggml_cann_create_tensor(src, nullptr, nullptr, 0, ACL_FORMAT_NCHW);
+    aclTensor* acl_dst =
+        ggml_cann_create_tensor(dst, nullptr, nullptr, 0, ACL_FORMAT_NCHW);
+
+    std::vector<int64_t> output_size{dst->ne[1], dst->ne[0]};
+    auto output_size_array = aclCreateIntArray(output_size.data(), 2);
+
+    uint64_t workspaceSize = 0;
+    aclOpExecutor* executor;
+    void* workspaceAddr = nullptr;
+
+    ACL_CHECK(aclnnUpsampleNearest2dGetWorkspaceSize(
+        acl_src, output_size_array, acl_dst, &workspaceSize, &executor));
+    if (workspaceSize > 0) {
+        ggml_cann_pool_alloc workspace_allocator(ctx.pool(), workspaceSize);
+        workspaceAddr = workspace_allocator.get();
+    }
+
+    ACL_CHECK(aclnnUpsampleNearest2d(workspaceAddr, workspaceSize, executor,
+                                     ctx.stream()));
+
+    ACL_CHECK(aclDestroyIntArray(output_size_array));
+    ACL_CHECK(aclDestroyTensor(acl_src));
+    ACL_CHECK(aclDestroyTensor(acl_dst));
+}
+
+/**
+ * @brief Pads a tensor with a specified value along each dimension.
+ *
+ * This function performs padding of the source tensor `acl_src` and stores the
+ * result in the destination tensor `acl_dst`. The padding values for each
+ * dimension are specified in the `paddings` array.
+ *
+ * @param ctx The context for the CANN backend operations.
+ * @param acl_src The source tensor to be padded.
+ * @param acl_dst The destination tensor where the padded result will be stored.
+ * @param paddings An array specifying the padding values for each dimension.
+ * The size of the array should be twice the number of dimensions of the tensor.
+ * @param value The value to be used for padding. The default value is 0.0.
+ */
+static void aclnn_pad(ggml_backend_cann_context& ctx, aclTensor* acl_src,
+                      aclTensor* acl_dst, int64_t* paddings,
+                      float value = 0.0f) {
+    aclIntArray* acl_pad = aclCreateIntArray(paddings, GGML_MAX_DIMS * 2);
+    aclScalar* acl_value = aclCreateScalar(&value, aclDataType::ACL_FLOAT);
+
+    uint64_t workspaceSize = 0;
+    aclOpExecutor* executor;
+    void* workspaceAddr = nullptr;
+
+    ACL_CHECK(aclnnConstantPadNdGetWorkspaceSize(
+        acl_src, acl_pad, acl_value, acl_dst, &workspaceSize, &executor));
+
+    if (workspaceSize > 0) {
+        ggml_cann_pool_alloc workspace_allocator(ctx.pool(), workspaceSize);
+        workspaceAddr = workspace_allocator.get();
+    }
+
+    ACL_CHECK(aclnnConstantPadNd(workspaceAddr, workspaceSize, executor,
+                                 ctx.stream()));
+
+    ACL_CHECK(aclDestroyIntArray(acl_pad));
+    ACL_CHECK(aclDestroyScalar(acl_value));
+}
+
+void ggml_cann_pad(ggml_backend_cann_context& ctx, ggml_tensor* dst) {
+    ggml_tensor* src = dst->src[0];
+    aclTensor* acl_src = ggml_cann_create_tensor(src);
+    aclTensor* acl_dst = ggml_cann_create_tensor(dst);
+
+    // padding: value in the array means how much distance will be padding.
+    // the position of elements in the array means which dirction to padding,
+    // each position means: [dim0.front, dim0.behind, dim1.front, dim1.behind,
+    //                       dim2.front, dim2.behind, dim3.front, dim3.behind]
+    int64_t paddings[] = {
+        0, dst->ne[0] - src->ne[0], 0, dst->ne[1] - src->ne[1],
+        0, dst->ne[2] - src->ne[2], 0, dst->ne[3] - src->ne[3]};
+    aclnn_pad(ctx, acl_src, acl_dst, paddings);
+
+    ACL_CHECK(aclDestroyTensor(acl_dst));
+    ACL_CHECK(aclDestroyTensor(acl_src));
+}
+
+/**
+ * @brief Performs 2D average pooling on the input tensor and stores the result
+ * in the destination tensor.
+ *
+ * This function performs average pooling on the source tensor and stores the
+ * result in the destination tensor. The pooling parameters (kernel size,
+ * strides, padding) are specified in the `op_params` of the destination tensor.
+ *
+ * @param ctx The context for the CANN backend operations.
+ * @param dst The destination tensor where the result will be stored. The source
+ * tensor is referenced by `dst->src[0]`.
+ */
+static void ggml_cann_avg_pool2d(ggml_backend_cann_context& ctx,
+                                 ggml_tensor* dst) {
+    ggml_tensor* src = dst->src[0];
+    GGML_ASSERT(src->type == GGML_TYPE_F32);
+    GGML_ASSERT(dst->type == GGML_TYPE_F32);
+
+    aclTensor* acl_src =
+        ggml_cann_create_tensor(src, nullptr, nullptr, 0, ACL_FORMAT_NCHW);
+    aclTensor* acl_dst =
+        ggml_cann_create_tensor(dst, nullptr, nullptr, 0, ACL_FORMAT_NCHW);
+
+    const int32_t* opts = (const int32_t*)dst->op_params;
+    const int k0 = opts[1];
+    const int k1 = opts[2];
+    const int s0 = opts[3];
+    const int s1 = opts[4];
+    const int p0 = opts[5];
+    const int p1 = opts[6];
+
+    std::vector<int64_t> kernel_dims = {k1, k0};
+    std::vector<int64_t> stride_dims = {s1, s0};
+    std::vector<int64_t> padding_avg_dims = {p1, p0};  // (padH, padW)
+
+    auto* kernel_size = aclCreateIntArray(kernel_dims.data(), 2);
+    auto* strides = aclCreateIntArray(stride_dims.data(), 2);
+    auto* paddings_avg = aclCreateIntArray(padding_avg_dims.data(), 2);
+
+    bool ceil_mode = false;
+    bool count_include_pad = true;
+    int64_t divisor_override = 0;
+    int8_t cube_math_type = 0;
+
+    uint64_t workspaceSize = 0;
+    aclOpExecutor* executor;
+    void* workspaceAddr = nullptr;
+
+    ACL_CHECK(aclnnAvgPool2dGetWorkspaceSize(
+        acl_src, kernel_size, strides, paddings_avg, ceil_mode,
+        count_include_pad, divisor_override, cube_math_type, acl_dst,
+        &workspaceSize, &executor));
+
+    if (workspaceSize > 0) {
+        ggml_cann_pool_alloc workspace_allocator(ctx.pool(), workspaceSize);
+        workspaceAddr = workspace_allocator.get();
+    }
+    ACL_CHECK(
+        aclnnAvgPool2d(workspaceAddr, workspaceSize, executor, ctx.stream()));
+
+    ACL_CHECK(aclDestroyTensor(acl_src));
+    ACL_CHECK(aclDestroyTensor(acl_dst));
+    ACL_CHECK(aclDestroyIntArray(kernel_size));
+    ACL_CHECK(aclDestroyIntArray(strides));
+    ACL_CHECK(aclDestroyIntArray(paddings_avg));
+}
+
+/**
+ * @brief Performs 2D max pooling on the input tensor and stores the result in
+ * the destination tensor.
+ *
+ * This function performs max pooling on the source tensor and stores the result
+ * in the destination tensor. The pooling parameters (kernel size, strides,
+ * padding) are specified in the `op_params` of the destination tensor.
+ *
+ * @param ctx The context for the CANN backend operations.
+ * @param dst The destination tensor where the result will be stored. The source
+ * tensor is referenced by `dst->src[0]`.
+ */
+static void ggml_cann_max_pool2d(ggml_backend_cann_context& ctx,
+                                 ggml_tensor* dst) {
+    ggml_tensor* src = dst->src[0];
+    GGML_ASSERT(src->type == GGML_TYPE_F32);
+    GGML_ASSERT(dst->type == GGML_TYPE_F32);
+
+    aclTensor* acl_src =
+        ggml_cann_create_tensor(src, nullptr, nullptr, 0, ACL_FORMAT_NCHW);
+    aclTensor* acl_dst =
+        ggml_cann_create_tensor(dst, nullptr, nullptr, 0, ACL_FORMAT_NCHW);
+
+    const int32_t* opts = (const int32_t*)dst->op_params;
+    const int k0 = opts[1];
+    const int k1 = opts[2];
+    const int s0 = opts[3];
+    const int s1 = opts[4];
+    const int p0 = opts[5];
+    const int p1 = opts[6];
+
+    int64_t temp_ne[] = {src->ne[0] + p0 * 2, src->ne[1] + p1 * 2, src->ne[2],
+                         src->ne[3]};
+    size_t temp_nb[GGML_MAX_DIMS];
+
+    temp_nb[0] = ggml_element_size(src);
+    for (int i = 1; i < GGML_MAX_DIMS; i++) {
+        temp_nb[i] = temp_nb[i - 1] * temp_ne[i - 1];
+    }
+
+    ggml_cann_pool_alloc temp_buffer_allocator(
+        ctx.pool(), ggml_nbytes(src) + p0 * 2 + p1 * 2 * src->nb[1]);
+    void* buffer = temp_buffer_allocator.get();
+    aclTensor* tmp_tensor = ggml_cann_create_tensor(
+        buffer, ACL_FLOAT, ggml_element_size(src), temp_ne, temp_nb,
+        GGML_MAX_DIMS, ACL_FORMAT_NCHW);
+
+    // pad: see padding in ggml_cann_pad()
+    int64_t paddings[] = {p0, p0, p1, p1, 0, 0, 0, 0};
+    float value = -FLT_MAX;
+    aclnn_pad(ctx, acl_src, tmp_tensor, paddings, value);
+
+    // max_pool
+    std::vector<int64_t> kernel_dims = {k1, k0};
+    std::vector<int64_t> stride_dims = {s1, s0};
+    // padding_max_dims: [dim0_start, dim0_end, dim1_start, dim1_end]
+    std::vector<int64_t> padding_max_dims = {0, 0, 0, 0};
+    std::vector<int64_t> dilation_size = {1, 1};
+    auto* kernel_size = aclCreateIntArray(kernel_dims.data(), 2);
+    auto* strides = aclCreateIntArray(stride_dims.data(), 2);
+    auto* paddings_max = aclCreateIntArray(padding_max_dims.data(), 4);
+    auto* dilations = aclCreateIntArray(dilation_size.data(), 2);
+
+    bool ceil_mode = false;
+    int64_t auto_pads = 0;
+
+    uint64_t workspaceSize = 0;
+    aclOpExecutor* executor;
+    void* workspaceAddr = nullptr;
+
+    ACL_CHECK(aclnnMaxPoolGetWorkspaceSize(
+        tmp_tensor, kernel_size, strides, auto_pads, paddings_max, dilations,
+        ceil_mode, acl_dst, &workspaceSize, &executor));
+    if (workspaceSize > 0) {
+        ggml_cann_pool_alloc workspace_allocator(ctx.pool(), workspaceSize);
+        workspaceAddr = workspace_allocator.get();
+    }
+
+    ACL_CHECK(
+        aclnnMaxPool(workspaceAddr, workspaceSize, executor, ctx.stream()));
+
+    ACL_CHECK(aclDestroyTensor(acl_src));
+    ACL_CHECK(aclDestroyTensor(acl_dst));
+    ACL_CHECK(aclDestroyTensor(tmp_tensor));
+    ACL_CHECK(aclDestroyIntArray(kernel_size));
+    ACL_CHECK(aclDestroyIntArray(strides));
+    ACL_CHECK(aclDestroyIntArray(paddings_max));
+    ACL_CHECK(aclDestroyIntArray(dilations));
+}
+
+void ggml_cann_pool2d(ggml_backend_cann_context& ctx, ggml_tensor* dst) {
+    const int32_t* opts = (const int32_t*)dst->op_params;
+    enum ggml_op_pool op = static_cast<ggml_op_pool>(opts[0]);
+    switch (op) {
+        case GGML_OP_POOL_AVG:
+            ggml_cann_avg_pool2d(ctx, dst);
+            break;
+        case GGML_OP_POOL_MAX:
+            ggml_cann_max_pool2d(ctx, dst);
+            break;
+        case GGML_OP_POOL_COUNT:
+            GGML_ABORT("fatal error");
+            break;
+    }
+}
+
+/**
+ * @brief Copies data from the source tensor to the destination tensor.
+ *
+ * This function copies data from the source tensor `acl_src` to the destination
+ * tensor `acl_dst`.
+ *
+ * @param ctx The context for the CANN backend operations.
+ * @param acl_src The source tensor from which data will be copied.
+ * @param acl_dst The destination tensor where the data will be copied to.
+ */
+static void cann_copy(ggml_backend_cann_context& ctx, aclTensor* acl_src,
+                      aclTensor* acl_dst) {
+    uint64_t workspaceSize = 0;
+    aclOpExecutor* executor;
+    void* workspaceAddr = nullptr;
+
+    ACL_CHECK(aclnnInplaceCopyGetWorkspaceSize(acl_dst, acl_src, &workspaceSize,
+                                               &executor));
+
+    if (workspaceSize > 0) {
+        ggml_cann_pool_alloc workspace_allocator(ctx.pool(), workspaceSize);
+        workspaceAddr = workspace_allocator.get();
+    }
+
+    ACL_CHECK(
+        aclnnInplaceCopy(workspaceAddr, workspaceSize, executor, ctx.stream()));
+}
+
+void ggml_cann_dup(ggml_backend_cann_context& ctx, ggml_tensor* dst) {
+    ggml_tensor* src = dst->src[0];
+
+    aclTensor* acl_src = ggml_cann_create_tensor(src);
+    aclTensor* acl_dst = ggml_cann_create_tensor(dst);
+
+    ggml_cann_pool_alloc src_extra_allocator(ctx.pool(), sizeof(ggml_tensor));
+    ggml_cann_pool_alloc dst_extra_allocator(ctx.pool(), sizeof(ggml_tensor));
+    src->extra = src_extra_allocator.get();
+    dst->extra = dst_extra_allocator.get();
+    ACL_CHECK(aclrtMemcpyAsync(src->extra, sizeof(ggml_tensor), src,
+                               sizeof(ggml_tensor), ACL_MEMCPY_HOST_TO_DEVICE,
+                               ctx.stream()));
+    ACL_CHECK(aclrtMemcpyAsync(dst->extra, sizeof(ggml_tensor), dst,
+                               sizeof(ggml_tensor), ACL_MEMCPY_HOST_TO_DEVICE,
+                               ctx.stream()));
+
+    if ((dst->type == GGML_TYPE_F16 || dst->type == GGML_TYPE_F32) &&
+        ggml_are_same_shape(src, dst)) {
+        cann_copy(ctx, acl_src, acl_dst);
+        ACL_CHECK(aclDestroyTensor(acl_src));
+        ACL_CHECK(aclDestroyTensor(acl_dst));
+        return;
+    }
+    // TODO: simplify
+    if (src->type == GGML_TYPE_F16) {
+        if (dst->type == GGML_TYPE_Q8_0) {
+            aclrtlaunch_ascendc_quantize_f16_q8_0(
+                24, ctx.stream(), src->data, dst->data,
+                ((ggml_tensor*)src->extra)->ne, ((ggml_tensor*)src->extra)->nb,
+                ((ggml_tensor*)dst->extra)->ne);
+            return;
+        }
+        if (dst->type == GGML_TYPE_Q4_0) {
+            aclrtlaunch_ascendc_quantize_f16_to_q4_0(
+                24, ctx.stream(), src->data, dst->data,
+                ((ggml_tensor*)src->extra)->ne, ((ggml_tensor*)src->extra)->nb,
+                ((ggml_tensor*)dst->extra)->ne);
+            return;
+        }
+        if (dst->type == GGML_TYPE_F16) {
+            if (ggml_are_same_shape(src, dst)) {
+                cann_copy(ctx, acl_src, acl_dst);
+                ACL_CHECK(aclDestroyTensor(acl_src));
+                ACL_CHECK(aclDestroyTensor(acl_dst));
+                return;
+            }
+            if (ggml_is_contiguous(dst)) {
+                const size_t src_type_size = ggml_type_size(src->type);
+                if (src->nb[0] == src_type_size) {
+                    // src0 is contigous on first dimension, copy by rows
+                    int64_t rows_num = ggml_nrows(src);
+
+                    aclrtlaunch_ascendc_dup_by_rows_fp16(
+                        rows_num, ctx.stream(), src->data, dst->data,
+                        ((ggml_tensor*)src->extra)->ne,
+                        ((ggml_tensor*)src->extra)->nb,
+                        ((ggml_tensor*)dst->extra)->ne,
+                        ((ggml_tensor*)dst->extra)->nb);
+                    return;
+                }
+                GGML_ABORT("fatal error");
+            }
+            GGML_ABORT("fatal error");
+        }
+        if (dst->type == GGML_TYPE_F32) {
+            if (ggml_are_same_shape(src, dst)) {
+                cann_copy(ctx, acl_src, acl_dst);
+                ACL_CHECK(aclDestroyTensor(acl_src));
+                ACL_CHECK(aclDestroyTensor(acl_dst));
+                return;
+            }
+            if (ggml_is_contiguous(dst)) {
+                const size_t src_type_size = ggml_type_size(src->type);
+                if (src->nb[0] == src_type_size) {
+                    // src0 is contigous on first dimension, copy by rows
+                    int64_t rows_num = ggml_nrows(src);
+                    aclrtlaunch_ascendc_dup_by_rows_fp16_to_fp32(
+                        rows_num, ctx.stream(), src->data, dst->data,
+                        ((ggml_tensor*)src->extra)->ne,
+                        ((ggml_tensor*)src->extra)->nb,
+                        ((ggml_tensor*)dst->extra)->ne,
+                        ((ggml_tensor*)dst->extra)->nb);
+                    return;
+                }
+                GGML_ABORT("fatal error");
+            }
+            GGML_ABORT("fatal error");
+        }
+        // TODO
+        GGML_ABORT("fatal error");
+    } else if (src->type == GGML_TYPE_F32) {
+        // TODO: if (src0->type == dst->type && ne00 == ne0 && nb00 == type_size
+        //          && nb0 == type_size)
+        if (dst->type == GGML_TYPE_Q8_0) {
+            aclrtlaunch_ascendc_quantize_f32_q8_0(
+                24, ctx.stream(), src->data, dst->data,
+                ((ggml_tensor*)src->extra)->ne, ((ggml_tensor*)src->extra)->nb,
+                ((ggml_tensor*)dst->extra)->ne);
+            return;
+        }
+        if (dst->type == GGML_TYPE_Q4_0) {
+            aclrtlaunch_ascendc_quantize_f32_to_q4_0(
+                24, ctx.stream(), src->data, dst->data,
+                ((ggml_tensor*)src->extra)->ne, ((ggml_tensor*)src->extra)->nb,
+                ((ggml_tensor*)dst->extra)->ne);
+            return;
+        }
+        if (dst->type == GGML_TYPE_F32) {
+            if (ggml_are_same_shape(src, dst)) {
+                cann_copy(ctx, acl_src, acl_dst);
+                ACL_CHECK(aclDestroyTensor(acl_src));
+                ACL_CHECK(aclDestroyTensor(acl_dst));
+                return;
+            }
+            if (ggml_is_contiguous(dst)) {
+                const size_t src_type_size = ggml_type_size(src->type);
+                if (src->nb[0] == src_type_size) {
+                    // src0 is contigous on first dimension, copy by rows
+                    int64_t rows_num = ggml_nrows(src);
+                    aclrtlaunch_ascendc_dup_by_rows_fp32(
+                        rows_num, ctx.stream(), src->data, dst->data,
+                        ((ggml_tensor*)src->extra)->ne,
+                        ((ggml_tensor*)src->extra)->nb,
+                        ((ggml_tensor*)dst->extra)->ne,
+                        ((ggml_tensor*)dst->extra)->nb);
+                    return;
+                }
+                GGML_ABORT("fatal error");
+            } else {
+                // TODO: dst not contiguous
+                GGML_ABORT("fatal error");
+            }
+        }
+        if (dst->type == GGML_TYPE_F16) {
+            if (ggml_are_same_shape(src, dst)) {
+                cann_copy(ctx, acl_src, acl_dst);
+                ACL_CHECK(aclDestroyTensor(acl_src));
+                ACL_CHECK(aclDestroyTensor(acl_dst));
+                return;
+            }
+            if (ggml_is_contiguous(dst)) {
+                const size_t src_type_size = ggml_type_size(src->type);
+                if (src->nb[0] == src_type_size) {
+                    // src0 is contigous on first dimension, copy by rows
+                    int64_t rows_num = ggml_nrows(src);
+                    aclrtlaunch_ascendc_dup_by_rows_fp32_to_fp16(
+                        rows_num, ctx.stream(), src->data, dst->data,
+                        ((ggml_tensor*)src->extra)->ne,
+                        ((ggml_tensor*)src->extra)->nb,
+                        ((ggml_tensor*)dst->extra)->ne,
+                        ((ggml_tensor*)dst->extra)->nb);
+                    return;
+                }
+                GGML_ABORT("fatal error");
+            }
+        }
+        // TODO
+        GGML_ABORT("fatal error");
+    } else {
+        if (ggml_are_same_shape(src, dst)) {
+            cann_copy(ctx, acl_src, acl_dst);
+            ACL_CHECK(aclDestroyTensor(acl_src));
+            ACL_CHECK(aclDestroyTensor(acl_dst));
+            return;
+        }
+        GGML_ABORT("fatal error");
+    }
+}
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+aclnnStatus aclnnRmsNormGetWorkspaceSize(const aclTensor* x,
+                                         const aclTensor* gamma, double epsilon,
+                                         const aclTensor* yOut,
+                                         const aclTensor* rstdOout,
+                                         uint64_t* workspaceSize,
+                                         aclOpExecutor** executor);
+aclnnStatus aclnnRmsNorm(void* workspace, uint64_t workspaceSize,
+                         aclOpExecutor* executor, aclrtStream stream);
+#ifdef __cplusplus
+}
+#endif
+
+/**
+ * @brief Creates an ACL tensor initialized with zeros using a provided buffer.
+ *
+ * This function initializes a tensor with zeros using the specified buffer and
+ * tensor parameters.
+ *
+ * @param ctx The context for the CANN backend operations.
+ * @param buffer The buffer to be used for the tensor data.
+ * @param n_bytes The size of the buffer in bytes.
+ * @param ne An array specifying the extents (sizes) of each dimension of the
+ * tensor.
+ * @param dims The number of dimensions of the tensor.
+ * @param type The data type of the tensor.
+ * @param type_size The size of each element in the tensor data type.
+ * @return An ACL tensor initialized with zeros.
+ */
+static aclTensor* aclnn_zero(ggml_backend_cann_context& ctx, void* buffer,
+                             size_t n_bytes, int64_t* ne, int64_t dims,
+                             aclDataType type, size_t type_size) {
+    size_t nb[GGML_MAX_DIMS];
+    nb[0] = type_size;
+    for (int i = 1; i < dims; i++) {
+        nb[i] = nb[i - 1] * ne[i - 1];
+    }
+
+    ACL_CHECK(aclrtMemsetAsync(buffer, n_bytes, 0, n_bytes, ctx.stream()));
+    aclTensor* zero =
+        ggml_cann_create_tensor(buffer, type, type_size, ne, nb, dims);
+    return zero;
+}
+
+/**
+ * @brief Creates an ACL tensor initialized with ones using a provided buffer.
+ *
+ * This function initializes a tensor with ones using the specified buffer and
+ * tensor parameters.
+ *
+ * @param ctx The context for the CANN backend operations.
+ * @param buffer The buffer to be used for the tensor data.
+ * @param n_bytes The size of the buffer in bytes.
+ * @param ne An array specifying the extents (sizes) of each dimension of the
+ * tensor.
+ * @param dims The number of dimensions of the tensor.
+ * @param type The data type of the tensor.
+ * @param type_size The size of each element in the tensor data type.
+ * @param value The value to be used for initializing the tensor (default
+ * is 1.0).
+ * @return An ACL tensor initialized with ones.
+ */
+static aclTensor* aclnn_ones(ggml_backend_cann_context& ctx, void* buffer,
+                             size_t n_bytes, int64_t* ne, int64_t dims,
+                             aclDataType type, size_t type_size,
+                             float value = 1.0f) {
+    aclTensor* acl_tensor =
+        aclnn_zero(ctx, buffer, n_bytes, ne, dims, type, type_size);
+    float alpha_host = 1.0f;
+    aclScalar* alpha = aclCreateScalar(&alpha_host, aclDataType::ACL_FLOAT);
+    aclScalar* other = aclCreateScalar(&value, aclDataType::ACL_FLOAT);
+
+    uint64_t workspaceSize = 0;
+    aclOpExecutor* executor;
+    void* workspaceAddr = nullptr;
+
+    ACL_CHECK(aclnnInplaceAddsGetWorkspaceSize(acl_tensor, other, alpha,
+                                               &workspaceSize, &executor));
+
+    if (workspaceSize > 0) {
+        ggml_cann_pool_alloc workspace_allocator(ctx.pool(), workspaceSize);
+        workspaceAddr = workspace_allocator.get();
+    }
+    ACL_CHECK(
+        aclnnInplaceAdds(workspaceAddr, workspaceSize, executor, ctx.stream()));
+
+    return acl_tensor;
+}
+
+void ggml_cann_rms_norm(ggml_backend_cann_context& ctx, ggml_tensor* dst) {
+    ggml_tensor* src = dst->src[0];
+
+    aclTensor* acl_src = ggml_cann_create_tensor(src);
+    aclTensor* acl_dst = ggml_cann_create_tensor(dst);
+
+    float eps;
+    memcpy(&eps, dst->op_params, sizeof(float));
+
+    GGML_ASSERT(eps > 0.0f);
+
+    uint64_t workspaceSize = 0;
+    aclOpExecutor* executor;
+    void* workspaceAddr = nullptr;
+
+    size_t one_tensor_n_bytes = src->ne[0] * ggml_element_size(src);
+    ggml_cann_pool_alloc one_tensor_allocator(ctx.pool(), one_tensor_n_bytes);
+
+    aclTensor* acl_gamma = aclnn_ones(
+        ctx, one_tensor_allocator.get(), one_tensor_n_bytes, src->ne, 1,
+        ggml_cann_type_mapping(src->type), ggml_element_size(src));
+
+    size_t zero_tensor_n_bytes =
+        src->ne[1] * src->ne[2] * src->ne[3] * ggml_element_size(src);
+    ggml_cann_pool_alloc zero_tensor_allocator(ctx.pool(), zero_tensor_n_bytes);
+    aclTensor* acl_rstd =
+        aclnn_zero(ctx, zero_tensor_allocator.get(), zero_tensor_n_bytes,
+                   src->ne, GGML_MAX_DIMS, ggml_cann_type_mapping(src->type),
+                   ggml_element_size(src));
+
+    ACL_CHECK(aclnnRmsNormGetWorkspaceSize(
+        acl_src, acl_gamma, eps, acl_dst, acl_rstd, &workspaceSize, &executor));
+
+    if (workspaceSize > 0) {
+        ggml_cann_pool_alloc workspace_allocator(ctx.pool(), workspaceSize);
+        workspaceAddr = workspace_allocator.get();
+    }
+
+    ACL_CHECK(
+        aclnnRmsNorm(workspaceAddr, workspaceSize, executor, ctx.stream()));
+
+    ACL_CHECK(aclDestroyTensor(acl_src));
+    ACL_CHECK(aclDestroyTensor(acl_dst));
+    ACL_CHECK(aclDestroyTensor(acl_gamma));
+    ACL_CHECK(aclDestroyTensor(acl_rstd));
+}
+
+// TODO: performace is low.
+void ggml_cann_diag_mask(ggml_backend_cann_context& ctx, ggml_tensor* dst,
+                         float value) {
+    ggml_tensor* src = dst->src[0];
+
+    aclTensor* acl_src = ggml_cann_create_tensor(src);
+    aclTensor* acl_dst = ggml_cann_create_tensor(dst);
+
+    const int n_past = ((int32_t*)dst->op_params)[0];
+
+    size_t one_tensor_n_bytes = src->ne[0] * src->ne[1] * src->ne[2] *
+                                src->ne[3] * ggml_element_size(src);
+    ggml_cann_pool_alloc one_tensor_allocator(ctx.pool(), one_tensor_n_bytes);
+
+    aclTensor* mask_tensor =
+        aclnn_ones(ctx, one_tensor_allocator.get(), one_tensor_n_bytes, src->ne,
+                   GGML_MAX_DIMS, ggml_cann_type_mapping(src->type),
+                   ggml_element_size(src), value);
+
+    uint64_t workspaceSize = 0;
+    aclOpExecutor* executor;
+    void* workspaceAddr = nullptr;
+
+    ACL_CHECK(aclnnInplaceTriuGetWorkspaceSize(mask_tensor, n_past + 1,
+                                               &workspaceSize, &executor));
+    if (workspaceSize > 0) {
+        ggml_cann_pool_alloc workspace_allocator(ctx.pool(), workspaceSize);
+        workspaceAddr = workspace_allocator.get();
+    }
+
+    ACL_CHECK(
+        aclnnInplaceTriu(workspaceAddr, workspaceSize, executor, ctx.stream()));
+
+    ACL_CHECK(aclnnTrilGetWorkspaceSize(acl_src, n_past + 1, acl_dst,
+                                        &workspaceSize, &executor));
+    if (workspaceSize > 0) {
+        ggml_cann_pool_alloc workspace_allocator(ctx.pool(), workspaceSize);
+        workspaceAddr = workspace_allocator.get();
+    }
+
+    ACL_CHECK(aclnnTril(workspaceAddr, workspaceSize, executor, ctx.stream()));
+
+    aclScalar* alpha = nullptr;
+    float alphaValue = 1.0f;
+    alpha = aclCreateScalar(&alphaValue, aclDataType::ACL_FLOAT);
+
+    ACL_CHECK(aclnnInplaceAddGetWorkspaceSize(acl_dst, mask_tensor, alpha,
+                                              &workspaceSize, &executor));
+    if (workspaceSize > 0) {
+        ggml_cann_pool_alloc workspace_allocator(ctx.pool(), workspaceSize);
+        workspaceAddr = workspace_allocator.get();
+    }
+    ACL_CHECK(
+        aclnnInplaceAdd(workspaceAddr, workspaceSize, executor, ctx.stream()));
+
+    ACL_CHECK(aclDestroyScalar(alpha));
+    ACL_CHECK(aclDestroyTensor(mask_tensor));
+    ACL_CHECK(aclDestroyTensor(acl_src));
+    ACL_CHECK(aclDestroyTensor(acl_dst));
+}
+
+/**
+ * @brief Casts the data type of a source tensor to a destination tensor.
+ *
+ * This function casts the data type of the source tensor `acl_src` to the
+ * specified data type `cast_data_type` and stores the result in the destination
+ * tensor `acl_dst`.
+ *
+ * @param ctx The context for the CANN backend operations.
+ * @param acl_src The source tensor whose data type will be casted.
+ * @param acl_dst The destination tensor where the casted result will be stored.
+ * @param cast_data_type The target data type to which the source tensor will be
+ * casted.
+ */
+static void aclnn_cast(ggml_backend_cann_context& ctx, aclTensor* acl_src,
+                       aclTensor* acl_dst, aclDataType cast_data_type) {
+    uint64_t workspaceSize = 0;
+    aclOpExecutor* executor;
+    void* workspaceAddr = nullptr;
+
+    ACL_CHECK(aclnnCastGetWorkspaceSize(acl_src, cast_data_type, acl_dst,
+                                        &workspaceSize, &executor));
+    if (workspaceSize > 0) {
+        ggml_cann_pool_alloc workspace_allocator(ctx.pool(), workspaceSize);
+        workspaceAddr = workspace_allocator.get();
+    }
+
+    ACL_CHECK(aclnnCast(workspaceAddr, workspaceSize, executor, ctx.stream()));
+}
+
+/**
+ * @brief Permutes the dimensions of a tensor according to a specified order.
+ *
+ * This function permutes the dimensions of the source tensor `acl_src`
+ * according to the order specified in the `new_dim` array and stores the result
+ * in the destination tensor `acl_dst`.
+ *
+ * @param ctx The context for the CANN backend operations.
+ * @param acl_src The source tensor whose dimensions will be permuted.
+ * @param acl_dst The destination tensor where the permuted result will be
+ * stored.
+ * @param new_dim An array specifying the new order of dimensions for the
+ * tensor.
+ * @param dims The number of dimensions in the tensor.
+ */
+static void aclnn_permute(ggml_backend_cann_context& ctx, aclTensor* acl_src,
+                          aclTensor* acl_dst, int64_t* new_dim, uint64_t dims) {
+    aclIntArray* acl_dims = aclCreateIntArray(new_dim, dims);
+
+    uint64_t workspaceSize = 0;
+    aclOpExecutor* executor;
+    void* workspaceAddr = nullptr;
+
+    ACL_CHECK(aclnnPermuteGetWorkspaceSize(acl_src, acl_dims, acl_dst,
+                                           &workspaceSize, &executor));
+    if (workspaceSize > 0) {
+        ggml_cann_pool_alloc workspace_allocator(ctx.pool(), workspaceSize);
+        workspaceAddr = workspace_allocator.get();
+    }
+
+    ACL_CHECK(
+        aclnnPermute(workspaceAddr, workspaceSize, executor, ctx.stream()));
+
+    ACL_CHECK(aclDestroyIntArray(acl_dims));
+}
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+aclnnStatus aclnnIm2colGetWorkspaceSize(const aclTensor* self,
+                                        const aclIntArray* kernelSize,
+                                        const aclIntArray* dilation,
+                                        const aclIntArray* padding,
+                                        const aclIntArray* stride,
+                                        aclTensor* out, uint64_t* workspaceSize,
+                                        aclOpExecutor** executor);
+aclnnStatus aclnnIm2col(void* workspace, uint64_t workspaceSize,
+                        aclOpExecutor* executor, aclrtStream stream);
+#ifdef __cplusplus
+}
+#endif
+
+static void ggml_cann_im2col_2d_post_process(ggml_backend_cann_context& ctx,
+                                             ggml_tensor* dst,
+                                             ggml_tensor* src1,
+                                             aclTensor* tmp_cast_tensor,
+                                             aclTensor* tmp_im2col_tensor) {
+    // Permute: [N, IC * KH * KW, OW * OH] -> [N, OW * OH, IC * KH * KW]
+    int64_t dst_ne[] = {dst->ne[0], dst->ne[1] * dst->ne[2], dst->ne[3]};
+    size_t dst_nb[] = {dst->nb[0], dst->nb[1], dst->nb[3]};
+    aclTensor* acl_dst =
+        ggml_cann_create_tensor(dst, dst_ne, dst_nb, GGML_MAX_DIMS - 1);
+
+    int64_t permute_dim[] = {0, 2, 1};
+    if (src1->type != dst->type) {
+        aclnn_permute(ctx, tmp_cast_tensor, acl_dst, permute_dim, 3);
+    } else {
+        aclnn_permute(ctx, tmp_im2col_tensor, acl_dst, permute_dim, 3);
+    }
+
+    // release
+    ACL_CHECK(aclDestroyTensor(acl_dst));
+}
+
+static void ggml_cann_im2col_1d_post_process(
+    ggml_backend_cann_context& ctx, ggml_tensor* dst, ggml_tensor* src1,
+    aclTensor* tmp_cast_tensor, aclTensor* tmp_im2col_tensor,
+    const std::vector<int64_t>& im2col_op_params) {
+    // get params
+    const int64_t KH = im2col_op_params[0];
+    const int64_t KW = im2col_op_params[1];
+    const int64_t IW = im2col_op_params[2];
+    const int64_t IC = im2col_op_params[3];
+    const int64_t N = im2col_op_params[4];
+    const int64_t OH = im2col_op_params[5];
+    const int64_t OW = im2col_op_params[6];
+    const int64_t s0 = im2col_op_params[7];
+    const int64_t p0 = im2col_op_params[8];
+    const int64_t d0 = im2col_op_params[9];
+    const int64_t n_bytes_factor = im2col_op_params[10];
+
+    // Permute: [N, IC * KH * KW, OW * OH] ->
+    // [N, OW * OH * n_bytes_factor, IC * KH * KW]
+    aclTensor* tmp_permute_tensor = nullptr;
+    ggml_cann_pool_alloc tmp_permute_allocator(ctx.pool());
+    tmp_permute_allocator.alloc(ggml_nbytes(dst) * n_bytes_factor);
+    void* tmp_permute_buffer = tmp_permute_allocator.get();
+
+    int64_t tmp_permute_ne[] = {IC * KH * KW, OW * OH * n_bytes_factor, N};
+    size_t tmp_permute_nb[GGML_MAX_DIMS - 1];
+    tmp_permute_nb[0] = ggml_type_size(dst->type);
+    for (int i = 1; i < GGML_MAX_DIMS - 1; i++) {
+        tmp_permute_nb[i] = tmp_permute_nb[i - 1] * tmp_permute_ne[i - 1];
+    }
+
+    tmp_permute_tensor = ggml_cann_create_tensor(
+        tmp_permute_buffer, ggml_cann_type_mapping(dst->type),
+        ggml_type_size(dst->type), tmp_permute_ne, tmp_permute_nb,
+        GGML_MAX_DIMS - 1, ACL_FORMAT_ND);
+
+    int64_t permute_dim[] = {0, 2, 1};
+    if (src1->type != dst->type) {
+        aclnn_permute(ctx, tmp_cast_tensor, tmp_permute_tensor, permute_dim, 3);
+    } else {
+        aclnn_permute(ctx, tmp_im2col_tensor, tmp_permute_tensor, permute_dim,
+                      3);
+    }
+
+    // number of times the kernel moves in W dimension
+    const int n_step_w = (IW + 2 * p0 - d0 * (KW - 1) - 1) / s0 + 1;
+    size_t offset;
+    void *cur_dst_buffer = dst->data, *cur_permute_buffer = tmp_permute_buffer;
+
+    // memory copy with offset to restore 1D im2col from 2d
+    if (IC > 1) {
+        offset = IC * KH * KW * n_step_w * ggml_type_size(dst->type);
+        size_t size_cpy = KH * KW * ggml_type_size(dst->type);
+
+        for (int c = 0; c < IC; c++) {
+            cur_permute_buffer = (char*)tmp_permute_buffer + offset +
+                                 KH * KW * c * ggml_type_size(dst->type);
+            cur_dst_buffer = (char*)dst->data +
+                             c * KH * KW * n_step_w * ggml_type_size(dst->type);
+
+            for (int i = 0; i < n_step_w; i++) {
+                ACL_CHECK(aclrtMemcpyAsync(
+                    cur_dst_buffer, size_cpy, cur_permute_buffer, size_cpy,
+                    ACL_MEMCPY_DEVICE_TO_DEVICE, ctx.stream()));
+                cur_dst_buffer =
+                    (char*)cur_dst_buffer + KH * KW * ggml_type_size(dst->type);
+                cur_permute_buffer = (char*)cur_permute_buffer +
+                                     KH * KW * IC * ggml_type_size(dst->type);
+            }
+        }
+    } else {
+        offset = KH * KW * n_step_w *
+                 ggml_type_size(dst->type);  // equal to ggml_nbytes(dst)
+        ACL_CHECK(aclrtMemcpyAsync(dst->data, offset,
+                                   (char*)tmp_permute_buffer + offset, offset,
+                                   ACL_MEMCPY_DEVICE_TO_DEVICE, ctx.stream()));
+    }
+
+    // release
+    ACL_CHECK(aclDestroyTensor(tmp_permute_tensor));
+}
+
+void ggml_cann_im2col(ggml_backend_cann_context& ctx, ggml_tensor* dst) {
+    ggml_tensor* src0 = dst->src[0];  // kernel
+    ggml_tensor* src1 = dst->src[1];  // input
+
+    GGML_ASSERT(src0->type == GGML_TYPE_F16);
+    GGML_ASSERT(src1->type == GGML_TYPE_F32);
+    GGML_ASSERT(dst->type == GGML_TYPE_F16 || dst->type == GGML_TYPE_F32);
+
+    GGML_TENSOR_BINARY_OP_LOCALS;
+
+    // aclnnIm2col only works on 2D. set s1, p1, d1 to 1 to perform 2D
+    // im2col and do post-processing to restore it to 1D.
+    const bool is_2D = ((const int32_t*)(dst->op_params))[6] == 1;
+    const int32_t s0 = ((const int32_t*)(dst->op_params))[0];
+    const int32_t s1 = is_2D ? ((const int32_t*)(dst->op_params))[1] : 1;
+    const int32_t p0 = ((const int32_t*)(dst->op_params))[2];
+    const int32_t p1 = is_2D ? ((const int32_t*)(dst->op_params))[3] : 1;
+    const int32_t d0 = ((const int32_t*)(dst->op_params))[4];
+    const int32_t d1 = is_2D ? ((const int32_t*)(dst->op_params))[5] : 1;
+
+    const int64_t N = ne13;
+    const int64_t IC = ne12;
+    const int64_t KH = ne01;
+    const int64_t KW = ne00;
+    const int64_t IW = ne10;
+
+    const int64_t OH = is_2D ? ne2 : 1;
+    const int64_t OW = ne1;
+
+    GGML_ASSERT(nb00 == sizeof(ggml_fp16_t));
+    GGML_ASSERT(nb10 == sizeof(float));
+
+    // memory allocated increased to 3x when is_2D == false
+    const int64_t n_bytes_factor = is_2D ? 1 : 3;
+
+    // im2col: [N,C,H,W] -> [N, IC * KH * KW, OW * OH * n_bytes_factor]
+    aclTensor* acl_src1 = ggml_cann_create_tensor(src1);
+    int64_t tmp_im2col_ne[] = {OW * OH * n_bytes_factor, IC * KH * KW, N};
+    size_t tmp_im2col_nb[GGML_MAX_DIMS - 1];
+
+    tmp_im2col_nb[0] = ggml_type_size(src1->type);
+    for (int i = 1; i < GGML_MAX_DIMS - 1; i++) {
+        tmp_im2col_nb[i] = tmp_im2col_nb[i - 1] * tmp_im2col_ne[i - 1];
+    }
+
+    // Calculate im2col.
+    // If dst is f16, tmp_buffer is f32, we need alloc src.typesize *
+    // dst.elemcount.
+    ggml_cann_pool_alloc im2col_allocator(
+        ctx.pool(),
+        ggml_nelements(dst) * ggml_element_size(src1) * n_bytes_factor);
+    void* tmp_im2col_buffer = im2col_allocator.get();
+
+    aclTensor* tmp_im2col_tensor = ggml_cann_create_tensor(
+        tmp_im2col_buffer, ggml_cann_type_mapping(src1->type),
+        ggml_type_size(src1->type), tmp_im2col_ne, tmp_im2col_nb,
+        GGML_MAX_DIMS - 1, ACL_FORMAT_ND);
+
+    std::vector<int64_t> kernel_dims = {KH, KW};
+    std::vector<int64_t> dilation_size = {d1, d0};
+    std::vector<int64_t> padding_dims = {p1, p0};
+    std::vector<int64_t> stride_dims = {s1, s0};
+    auto* kernel_size = aclCreateIntArray(kernel_dims.data(), 2);
+    auto* dilations = aclCreateIntArray(dilation_size.data(), 2);
+    auto* paddings = aclCreateIntArray(padding_dims.data(), 2);
+    auto* strides = aclCreateIntArray(stride_dims.data(), 2);
+
+    uint64_t workspaceSize = 0;
+    aclOpExecutor* executor;
+    void* workspaceAddr = nullptr;
+
+    ACL_CHECK(aclnnIm2colGetWorkspaceSize(acl_src1, kernel_size, dilations,
+                                          paddings, strides, tmp_im2col_tensor,
+                                          &workspaceSize, &executor));
+
+    ggml_cann_pool_alloc workspace_allocator(ctx.pool());
+    if (workspaceSize > 0) {
+        workspace_allocator.alloc(workspaceSize);
+        workspaceAddr = workspace_allocator.get();
+    }
+
+    ACL_CHECK(
+        aclnnIm2col(workspaceAddr, workspaceSize, executor, ctx.stream()));
+
+    // Cast if dst is f16.
+    aclTensor* tmp_cast_tensor = nullptr;
+    ggml_cann_pool_alloc tmp_cast_allocator(ctx.pool());
+    void* tmp_cast_buffer = nullptr;
+    if (src1->type != dst->type) {
+        tmp_cast_allocator.alloc(ggml_nbytes(dst) * n_bytes_factor);
+        tmp_cast_buffer = tmp_cast_allocator.get();
+        size_t temp_cast_nb[GGML_MAX_DIMS - 1];
+        temp_cast_nb[0] = ggml_type_size(dst->type);
+        for (int i = 1; i < GGML_MAX_DIMS - 1; i++) {
+            temp_cast_nb[i] = temp_cast_nb[i - 1] * tmp_im2col_ne[i - 1];
+        }
+
+        tmp_cast_tensor = ggml_cann_create_tensor(
+            tmp_cast_buffer, ggml_cann_type_mapping(dst->type),
+            ggml_type_size(dst->type), tmp_im2col_ne, temp_cast_nb,
+            GGML_MAX_DIMS - 1, ACL_FORMAT_ND);
+        aclnn_cast(ctx, tmp_im2col_tensor, tmp_cast_tensor,
+                   ggml_cann_type_mapping(dst->type));
+    }
+
+    // post-processing
+    if (is_2D) {
+        ggml_cann_im2col_2d_post_process(ctx, dst, src1, tmp_cast_tensor,
+                                         tmp_im2col_tensor);
+    } else {
+        std::vector<int64_t> im2col_op_params = {
+            KH, KW, IW, IC, N, OH, OW, s0, p0, d0, n_bytes_factor};
+        ggml_cann_im2col_1d_post_process(ctx, dst, src1, tmp_cast_tensor,
+                                         tmp_im2col_tensor, im2col_op_params);
+    }
+
+    // release
+    ACL_CHECK(aclDestroyTensor(acl_src1));
+    ACL_CHECK(aclDestroyTensor(tmp_im2col_tensor));
+    ACL_CHECK(aclDestroyTensor(tmp_cast_tensor));
+    ACL_CHECK(aclDestroyIntArray(kernel_size));
+    ACL_CHECK(aclDestroyIntArray(dilations));
+    ACL_CHECK(aclDestroyIntArray(paddings));
+    ACL_CHECK(aclDestroyIntArray(strides));
+}
+
+/**
+ * @brief Applies element-wise exponential function to the elements of a tensor.
+ *
+ * This function computes the exponential of each element in the source tensor
+ * `acl_src` and stores the result back into the same tensor.
+ * The operation is defined as:
+ * \f[
+ *     \text {acl_src }_i=e^{acl\_src_i}
+ * \f]
+ *
+ * @param ctx The context for the CANN backend operations.
+ * @param acl_src The tensor on which the exponential function will be applied.
+ */
+static void aclnn_exp(ggml_backend_cann_context& ctx, aclTensor* acl_src) {
+    uint64_t workspaceSize = 0;
+    aclOpExecutor* executor;
+    void* workspaceAddr = nullptr;
+
+    ACL_CHECK(
+        aclnnInplaceExpGetWorkspaceSize(acl_src, &workspaceSize, &executor));
+    if (workspaceSize > 0) {
+        ggml_cann_pool_alloc workspace_allocator(ctx.pool(), workspaceSize);
+        workspaceAddr = workspace_allocator.get();
+    }
+
+    ACL_CHECK(
+        aclnnInplaceExp(workspaceAddr, workspaceSize, executor, ctx.stream()));
+}
+
+/**
+ * @brief Multiplies elements of a tensor by a scalar value, optionally
+ * in-place.
+ *
+ * This function multiplies each element of the source tensor `acl_src` by the
+ * scalar `scale` and stores the result in the destination tensor `acl_dst`. If
+ * `inplace` is true, `acl_dst` will not be used and the operation is performed
+ *  in-place on `acl_src`.
+ * The operation is defined as:
+ * \f[
+ *     \text {acl_dst }_i=\text {acl_src }_i \times \text {scale}
+ * \f]
+ *
+ * @param ctx The context for the CANN backend operations.
+ * @param acl_src The source tensor whose elements will be multiplied.
+ * @param scale The scalar value by which each element of `acl_src` will be
+ * multiplied.
+ * @param acl_dst The destination tensor where the result will be stored if
+ * `inplace` is false.
+ * @param inplace Flag indicating whether to perform the operation in-place on
+ * `acl_src`.
+ */
+static void aclnn_muls(ggml_backend_cann_context& ctx, aclTensor* acl_src,
+                       float scale, aclTensor* acl_dst, bool inplace) {
+    aclScalar* acl_scale = aclCreateScalar(&scale, aclDataType::ACL_FLOAT);
+
+    uint64_t workspaceSize = 0;
+    aclOpExecutor* executor;
+    void* workspaceAddr = nullptr;
+
+    if (inplace) {
+        ACL_CHECK(aclnnInplaceMulsGetWorkspaceSize(acl_src, acl_scale,
+                                                   &workspaceSize, &executor));
+        if (workspaceSize > 0) {
+            ggml_cann_pool_alloc workspace_allocator(ctx.pool(), workspaceSize);
+            workspaceAddr = workspace_allocator.get();
+        }
+
+        ACL_CHECK(aclnnInplaceMuls(workspaceAddr, workspaceSize, executor,
+                                   ctx.stream()));
+    } else {
+        ACL_CHECK(aclnnMulsGetWorkspaceSize(acl_src, acl_scale, acl_dst,
+                                            &workspaceSize, &executor));
+        if (workspaceSize > 0) {
+            ggml_cann_pool_alloc workspace_allocator(ctx.pool(), workspaceSize);
+            workspaceAddr = workspace_allocator.get();
+        }
+
+        ACL_CHECK(
+            aclnnMuls(workspaceAddr, workspaceSize, executor, ctx.stream()));
+    }
+
+    ACL_CHECK(aclDestroyScalar(acl_scale));
+}
+
+/**
+ * @brief Performs an in-place element-wise multiplication of two tensors.
+ *
+ * This function performs an element-wise multiplication of the tensors
+ * `acl_src` and `acl_other` and stores the result in `acl_src`.
+ * The operation is defined as:
+ * \f[
+ *     \text {acl_src }_i=\text {acl_src }_i \times \text {acl_other }_i
+ * \f]
+ *
+ * @param ctx The context for the CANN backend operations.
+ * @param acl_src The source tensor where the multiplication result will be
+ * stored.
+ * @param acl_other The tensor whose elements will be multiplied with `acl_src`.
+ */
+static void aclnn_inplace_mul(ggml_backend_cann_context& ctx,
+                              aclTensor* acl_src, aclTensor* acl_other) {
+    uint64_t workspaceSize = 0;
+    aclOpExecutor* executor;
+    void* workspaceAddr = nullptr;
+
+    ACL_CHECK(aclnnInplaceMulGetWorkspaceSize(acl_src, acl_other,
+                                              &workspaceSize, &executor));
+    if (workspaceSize > 0) {
+        ggml_cann_pool_alloc workspace_allocator(ctx.pool(), workspaceSize);
+        workspaceAddr = workspace_allocator.get();
+    }
+
+    ACL_CHECK(
+        aclnnInplaceMul(workspaceAddr, workspaceSize, executor, ctx.stream()));
+}
+
+/**
+ * @brief Performs element-wise multiplication of two tensors and stores the
+ * result in a destination tensor.
+ *
+ * This function performs element-wise multiplication of the tensors `acl_src`
+ * and `acl_other` and stores the result in the destination tensor `acl_dst`.
+ * The operation is defined as:
+ * \f[
+ *     \text {acl_dst }_i=\text {acl_src }_i \times \text {acl_other }_i
+ * \f]
+ *
+ * @param ctx The context for the CANN backend operations.
+ * @param acl_src The first tensor for element-wise multiplication.
+ * @param acl_other The second tensor for element-wise multiplication.
+ * @param acl_dst The destination tensor where the result will be stored.
+ */
+static void aclnn_mul(ggml_backend_cann_context& ctx, aclTensor* acl_src,
+                      aclTensor* acl_other, aclTensor* acl_dst) {
+    uint64_t workspaceSize = 0;
+    aclOpExecutor* executor;
+    void* workspaceAddr = nullptr;
+
+    ACL_CHECK(aclnnMulGetWorkspaceSize(acl_src, acl_other, acl_dst,
+                                       &workspaceSize, &executor));
+    if (workspaceSize > 0) {
+        ggml_cann_pool_alloc workspace_allocator(ctx.pool(), workspaceSize);
+        workspaceAddr = workspace_allocator.get();
+    }
+
+    ACL_CHECK(aclnnMul(workspaceAddr, workspaceSize, executor, ctx.stream()));
+}
+
+/**
+ * @brief Applies element-wise cosine function to the elements of a tensor.
+ *
+ * This function computes the cosine of each element in the source tensor
+ * `acl_src` and stores the result in the destination tensor `acl_dst`. The
+ * operation is defined as: \f[ \text {acl_dst }_i=\cos \left(\text {acl_src
+ * }_i\right) \f]
+ *
+ * @param ctx The context for the CANN backend operations.
+ * @param acl_src The source tensor on which the cosine function will be
+ * applied.
+ * @param acl_dst The destination tensor where the cosine results will be
+ * stored.
+ */
+static void aclnn_cos(ggml_backend_cann_context& ctx, aclTensor* acl_src,
+                      aclTensor* acl_dst) {
+    uint64_t workspaceSize = 0;
+    aclOpExecutor* executor;
+    void* workspaceAddr = nullptr;
+
+    ACL_CHECK(
+        aclnnCosGetWorkspaceSize(acl_src, acl_dst, &workspaceSize, &executor));
+    if (workspaceSize > 0) {
+        ggml_cann_pool_alloc workspace_allocator(ctx.pool(), workspaceSize);
+        workspaceAddr = workspace_allocator.get();
+    }
+
+    ACL_CHECK(aclnnCos(workspaceAddr, workspaceSize, executor, ctx.stream()));
+}
+
+/**
+ * @brief Applies element-wise sine function to the elements of a tensor.
+ *
+ * This function computes the sine of each element in the source tensor
+ `acl_src`
+ * and stores the result in the destination tensor `acl_dst`.
+ * The operation is defined as:
+ * \f[
+ *     \text {acl_dst }_i=\sin \left(\text {acl_src }_i\right)
+ * \f]
+
+ * @param ctx The context for the CANN backend operations.
+ * @param acl_src The source tensor on which the sine function will be applied.
+ * @param acl_dst The destination tensor where the sine results will be stored.
+ */
+static void aclnn_sin(ggml_backend_cann_context& ctx, aclTensor* acl_src,
+                      aclTensor* acl_dst) {
+    uint64_t workspaceSize = 0;
+    aclOpExecutor* executor;
+    void* workspaceAddr = nullptr;
+
+    ACL_CHECK(
+        aclnnSinGetWorkspaceSize(acl_src, acl_dst, &workspaceSize, &executor));
+    if (workspaceSize > 0) {
+        ggml_cann_pool_alloc workspace_allocator(ctx.pool(), workspaceSize);
+        workspaceAddr = workspace_allocator.get();
+    }
+
+    ACL_CHECK(aclnnSin(workspaceAddr, workspaceSize, executor, ctx.stream()));
+}
+
+void ggml_cann_timestep_embedding(ggml_backend_cann_context& ctx,
+                                  ggml_tensor* dst) {
+    const ggml_tensor* src = dst->src[0];
+
+    GGML_ASSERT(src->type == GGML_TYPE_F32);
+    GGML_ASSERT(dst->type == GGML_TYPE_F32);
+
+    const int dim = dst->op_params[0];
+    const int max_period = dst->op_params[1];
+    int half = dim / 2;
+
+    aclTensor* acl_src = ggml_cann_create_tensor(src);
+
+    // arange: [0, ..., half)
+    float start = 0;
+    float stop = half;
+    float step = 1;
+    int64_t n_elements_arange = half;
+    int64_t tmp_arange_ne[] = {half};
+    size_t tmp_arange_nb[] = {sizeof(dst->type)};
+
+    ggml_cann_pool_alloc arange_allocator(ctx.pool(), half * sizeof(dst->type));
+    void* tmp_arange_buffer = arange_allocator.get();
+    aclTensor* tmp_arange_tensor = ggml_cann_create_tensor(
+        tmp_arange_buffer, ggml_cann_type_mapping(dst->type),
+        ggml_type_size(dst->type), tmp_arange_ne, tmp_arange_nb,
+        GGML_MAX_DIMS - 3, ACL_FORMAT_ND);
+
+    aclnn_arange(ctx, tmp_arange_tensor, start, stop, step, n_elements_arange);
+
+    // freq
+    float freq_param = -logf(max_period) / half;
+    bool inplace = true;
+    aclnn_muls(ctx, tmp_arange_tensor, freq_param, nullptr, inplace);
+    aclnn_exp(ctx, tmp_arange_tensor);
+
+    // permute: src [0,1,2,3]->[0,1,3,2]
+    int64_t tmp_permute_ne[] = {src->ne[1], src->ne[0], src->ne[2], src->ne[3]};
+    size_t tmp_permute_nb[GGML_MAX_DIMS];
+    tmp_permute_nb[0] = ggml_type_size(src->type);
+    for (int i = 1; i < GGML_MAX_DIMS; i++) {
+        tmp_permute_nb[i] = tmp_permute_nb[i - 1] * tmp_permute_ne[i - 1];
+    }
+
+    ggml_cann_pool_alloc permute_allocator(ctx.pool(), ggml_nbytes(src));
+    void* tmp_permute_buffer = permute_allocator.get();
+    aclTensor* tmp_permute_tenosr = ggml_cann_create_tensor(
+        tmp_permute_buffer, ggml_cann_type_mapping(src->type),
+        ggml_type_size(src->type), tmp_permute_ne, tmp_permute_nb,
+        GGML_MAX_DIMS, ACL_FORMAT_ND);
+    int64_t permute_dim[] = {0, 1, 3, 2};
+    int64_t num_dims = 4;
+    aclnn_permute(ctx, acl_src, tmp_permute_tenosr, permute_dim, num_dims);
+
+    // timestep * freq
+    int64_t tmp_mul_ne[] = {src->ne[1] * half, src->ne[0], src->ne[2],
+                            src->ne[3]};
+    size_t tmp_mul_nb[GGML_MAX_DIMS];
+    tmp_mul_nb[0] = ggml_type_size(src->type);
+    for (int i = 1; i < GGML_MAX_DIMS; i++) {
+        tmp_mul_nb[i] = tmp_mul_nb[i - 1] * tmp_mul_ne[i - 1];
+    }
+
+    int mul_nelements =
+        src->ne[1] * half * src->ne[0] * src->ne[2] * src->ne[3];
+
+    ggml_cann_pool_alloc mul_allocator(
+        ctx.pool(), mul_nelements * ggml_type_size(src->type));
+    void* tmp_mul_buffer = mul_allocator.get();
+    aclTensor* tmp_mul_tensor = ggml_cann_create_tensor(
+        tmp_mul_buffer, ggml_cann_type_mapping(src->type),
+        ggml_type_size(src->type), tmp_mul_ne, tmp_mul_nb, GGML_MAX_DIMS,
+        ACL_FORMAT_ND);
+    aclnn_mul(ctx, tmp_permute_tenosr, tmp_arange_tensor, tmp_mul_tensor);
+
+    // cos
+    ggml_cann_pool_alloc cos_allocator(
+        ctx.pool(), mul_nelements * ggml_type_size(src->type));
+    void* tmp_cos_buffer = cos_allocator.get();
+    aclTensor* tmp_cos_tensor = ggml_cann_create_tensor(
+        tmp_cos_buffer, ggml_cann_type_mapping(dst->type),
+        ggml_type_size(dst->type), tmp_mul_ne, tmp_mul_nb, GGML_MAX_DIMS,
+        ACL_FORMAT_ND);
+
+    aclnn_cos(ctx, tmp_mul_tensor, tmp_cos_tensor);
+
+    // sin
+    ggml_cann_pool_alloc sin_allocator(
+        ctx.pool(), mul_nelements * ggml_type_size(src->type));
+    void* tmp_sin_buffer = sin_allocator.get();
+    aclTensor* tmp_sin_tensor = ggml_cann_create_tensor(
+        tmp_sin_buffer, ggml_cann_type_mapping(dst->type),
+        ggml_type_size(dst->type), tmp_mul_ne, tmp_mul_nb, GGML_MAX_DIMS,
+        ACL_FORMAT_ND);
+
+    aclnn_sin(ctx, tmp_mul_tensor, tmp_sin_tensor);
+
+    // concat
+    int64_t concat_dim = 3;
+    aclTensor* acl_dst = ggml_cann_create_tensor(dst);
+    aclTensor* tensors[] = {tmp_cos_tensor, tmp_sin_tensor};
+    aclTensorList* tensorList = aclCreateTensorList(tensors, 2);
+    aclnn_concat(ctx, tensorList, acl_dst, concat_dim);
+
+    // release
+    // segmentation fault when delete both tensorList and his elements.
+    ACL_CHECK(aclDestroyTensorList(tensorList));
+    ACL_CHECK(aclDestroyTensor(acl_src));
+    ACL_CHECK(aclDestroyTensor(tmp_arange_tensor));
+    ACL_CHECK(aclDestroyTensor(tmp_permute_tenosr));
+    ACL_CHECK(aclDestroyTensor(tmp_mul_tensor));
+    ACL_CHECK(aclDestroyTensor(acl_dst));
+}
+
+/**
+ * @brief Fills a tensor with a scalar value.
+ *
+ * This function fills the destination tensor `acl_dst` with the scalar value
+ * `scalar`.
+ *
+ * @param ctx The context for the CANN backend operations.
+ * @param scalar The scalar value used to fill the tensor.
+ * @param acl_dst The destination tensor to be filled with the scalar value.
+ */
+static void aclnn_fill_scalar(ggml_backend_cann_context& ctx, float scalar,
+                              aclTensor* acl_dst) {
+    auto acl_scalar = aclCreateScalar(&scalar, aclDataType::ACL_FLOAT);
+
+    uint64_t workspaceSize = 0;
+    aclOpExecutor* executor;
+    void* workspaceAddr = nullptr;
+
+    ACL_CHECK(aclnnInplaceFillScalarGetWorkspaceSize(
+        acl_dst, acl_scalar, &workspaceSize, &executor));
+    if (workspaceSize > 0) {
+        ggml_cann_pool_alloc workspace_allocator(ctx.pool(), workspaceSize);
+        workspaceAddr = workspace_allocator.get();
+    }
+
+    ACL_CHECK(aclnnInplaceFillScalar(workspaceAddr, workspaceSize, executor,
+                                     ctx.stream()));
+    ACL_CHECK(aclDestroyScalar(acl_scalar));
+}
+
+/**
+ * @brief Raises each element of a tensor to the power of the corresponding
+ * element in another tensor.
+ *
+ * This function computes the element-wise power of the destination tensor
+ * `acl_dst` raised to the power of the exponent tensor `acl_exp`.
+ * The operation is defined as:
+ * \f[
+ *     \text {acl_dst }_i=acl\_dst_i^{\text {acl_exp }_i}
+ * \f]
+ *
+ * @param ctx The context for the CANN backend operations.
+ * @param acl_dst The destination tensor, which also serves as the base tensor.
+ * @param acl_exp The exponent tensor, each element of which is used to raise
+ * the corresponding element in the destination tensor.
+ */
+static void aclnn_pow_tensor_tensor(ggml_backend_cann_context& ctx,
+                                    aclTensor* acl_dst, aclTensor* acl_exp) {
+    uint64_t workspaceSize = 0;
+    aclOpExecutor* executor;
+    void* workspaceAddr = nullptr;
+
+    ACL_CHECK(aclnnInplacePowTensorTensorGetWorkspaceSize(
+        acl_dst, acl_exp, &workspaceSize, &executor));
+    if (workspaceSize > 0) {
+        ggml_cann_pool_alloc workspace_allocator(ctx.pool(), workspaceSize);
+        workspaceAddr = workspace_allocator.get();
+    }
+
+    ACL_CHECK(aclnnInplacePowTensorTensor(workspaceAddr, workspaceSize,
+                                          executor, ctx.stream()));
+}
+
+/**
+ * @brief   Applies the Alibi (Attention with Linear Biases) mechanism to the
+ * @details This function implements the Alibi mechanism, which introduces
+ *          learnable biases into the attention scores to simulate relative
+ *          position encoding without the need for explicit positional
+ *          embeddings.
+ *
+ * @param ctx          The backend CANN context for executing operations.
+ * @param acl_src      The source tensor representing the query or key.
+ * @param acl_position The position tensor containing relative positions.
+ * @param acl_dst      The destination tensor where the result will be stored.
+ * @param n_head       The number of attention heads.
+ * @param src_ne       The dimensions of the source tensor.
+ * @param src_nb0      The byte size of the first dimension of the source
+ tensor.
+ * @param max_bias     The maximum bias value used in the Alibi mechanism.
+ * @param dst          The destination tensor object for additional metadata.
+ *
+ * The function performs the following steps:
+ * 1. Calculates the logarithm floor of the number of heads to determine the
+      base for bias calculation.
+ * 2. Initializes arrays with arithmetic sequences and fills them with bias
+      values.
+ * 3. Computes the bias tensor based on the calculated biases and arithmetic
+      sequences.
+ * 4. Reshapes the bias tensor to match the dimensions of the input tensors.
+ * 5. Multiplies the position tensor by the bias tensor.
+ * 6. Adds the result of the multiplication to the source tensor to produce the
+      final output.
+ */
+static void aclnn_alibi(ggml_backend_cann_context& ctx, aclTensor* acl_src,
+                        aclTensor* acl_position, aclTensor* acl_dst,
+                        const int n_head, int64_t* src_ne, const size_t src_nb0,
+                        float max_bias, ggml_tensor* dst) {
+    const int64_t ne2_ne3 = src_ne[2] * src_ne[3];
+    GGML_ASSERT(src_nb0 == sizeof(float));
+    GGML_ASSERT(n_head == src_ne[2]);
+
+    const int n_heads_log2_floor = 1u << (uint32_t)floor(log2(n_head));
+
+    float m0 = powf(2.0f, -(max_bias) / n_heads_log2_floor);
+    float m1 = powf(2.0f, -(max_bias / 2.0f) / n_heads_log2_floor);
+
+    // init arange
+    ggml_cann_pool_alloc arange_allocator(ctx.pool(),
+                                          ne2_ne3 * ggml_type_size(dst->type));
+    void* tmp_arange_buffer = arange_allocator.get();
+
+    // arange1: [1, ..., n_heads_log2_floor+1)
+    float start = 1;
+    float stop = n_heads_log2_floor + 1;
+    float step = 1;
+    int64_t n_elements_arange = n_heads_log2_floor;
+
+    int64_t tmp_arange1_ne[] = {n_heads_log2_floor};
+    size_t tmp_arange1_nb[] = {sizeof(dst->type)};
+    aclTensor* tmp_arange1_tensor = ggml_cann_create_tensor(
+        tmp_arange_buffer, ggml_cann_type_mapping(dst->type),
+        ggml_type_size(dst->type), tmp_arange1_ne, tmp_arange1_nb,
+        GGML_MAX_DIMS - 3, ACL_FORMAT_ND);
+
+    aclnn_arange(ctx, tmp_arange1_tensor, start, stop, step, n_elements_arange);
+
+    aclTensor* tmp_arange2_tensor = nullptr;
+    if (n_heads_log2_floor < ne2_ne3) {
+        // arange2: [1, ..., 2 * (k - n_heads_log2_floor) + 1)
+        start = 1;
+        stop = 2 * (ne2_ne3 - n_heads_log2_floor) + 1;
+        step = 2;
+        n_elements_arange = ne2_ne3 - n_heads_log2_floor;
+        int64_t tmp_arange2_ne[] = {ne2_ne3 - n_heads_log2_floor};
+        size_t tmp_arange2_nb[] = {sizeof(dst->type)};
+
+        aclTensor* tmp_arange2_tensor = ggml_cann_create_tensor(
+            (char*)tmp_arange_buffer +
+                n_heads_log2_floor * ggml_type_size(dst->type),
+            ggml_cann_type_mapping(dst->type), ggml_type_size(dst->type),
+            tmp_arange2_ne, tmp_arange2_nb, GGML_MAX_DIMS - 3, ACL_FORMAT_ND);
+        aclnn_arange(ctx, tmp_arange2_tensor, start, stop, step,
+                     n_elements_arange);
+    }
+
+    // init mk_base
+    ggml_cann_pool_alloc mk_base_allocator(ctx.pool(),
+                                           ne2_ne3 * ggml_type_size(dst->type));
+    void* tmp_mk_base_buffer = mk_base_allocator.get();
+    int64_t tmp_mk_base1_ne[] = {n_heads_log2_floor};
+    size_t tmp_mk_base1_nb[] = {sizeof(dst->type)};
+    aclTensor* tmp_mk_base1_tensor = ggml_cann_create_tensor(
+        tmp_mk_base_buffer, ggml_cann_type_mapping(dst->type),
+        ggml_type_size(dst->type), tmp_mk_base1_ne, tmp_mk_base1_nb,
+        GGML_MAX_DIMS - 3, ACL_FORMAT_ND);
+
+    aclnn_fill_scalar(ctx, m0, tmp_mk_base1_tensor);
+
+    aclTensor* tmp_mk_base2_tensor = nullptr;
+    if (n_heads_log2_floor < ne2_ne3) {
+        int64_t tmp_mk_base2_ne[] = {ne2_ne3 - n_heads_log2_floor};
+        size_t tmp_mk_base2_nb[] = {sizeof(dst->type)};
+        aclTensor* tmp_mk_base2_tensor = ggml_cann_create_tensor(
+            (char*)tmp_mk_base_buffer +
+                n_heads_log2_floor * ggml_type_size(dst->type),
+            ggml_cann_type_mapping(dst->type), ggml_type_size(dst->type),
+            tmp_mk_base2_ne, tmp_mk_base2_nb, GGML_MAX_DIMS - 3, ACL_FORMAT_ND);
+        aclnn_fill_scalar(ctx, m1, tmp_mk_base2_tensor);
+    }
+
+    // init mk
+    int64_t tmp_mk_base_ne[] = {ne2_ne3};
+    size_t tmp_mk_base_nb[] = {sizeof(dst->type)};
+    aclTensor* tmp_mk_base_tensor = ggml_cann_create_tensor(
+        tmp_mk_base_buffer, ggml_cann_type_mapping(dst->type),
+        ggml_type_size(dst->type), tmp_mk_base_ne, tmp_mk_base_nb,
+        GGML_MAX_DIMS - 3, ACL_FORMAT_ND);
+    aclTensor* tmp_arange_tensor = ggml_cann_create_tensor(
+        tmp_arange_buffer, ggml_cann_type_mapping(dst->type),
+        ggml_type_size(dst->type), tmp_mk_base_ne, tmp_mk_base_nb,
+        GGML_MAX_DIMS - 3, ACL_FORMAT_ND);
+    aclnn_pow_tensor_tensor(ctx, tmp_mk_base_tensor, tmp_arange_tensor);
+
+    // reshape mk
+    int64_t tmp_mk_ne[] = {1, 1, src_ne[2], src_ne[3]};
+    size_t tmp_mk_nb[GGML_MAX_DIMS];
+    tmp_mk_nb[0] = ggml_type_size(dst->type);
+    for (int i = 1; i < GGML_MAX_DIMS; i++) {
+        tmp_mk_nb[i] = tmp_mk_nb[i - 1] * tmp_mk_ne[i - 1];
+    }
+    aclTensor* tmp_mk_tensor = ggml_cann_create_tensor(
+        tmp_mk_base_buffer, ggml_cann_type_mapping(dst->type),
+        ggml_type_size(dst->type), tmp_mk_ne, tmp_mk_nb, GGML_MAX_DIMS,
+        ACL_FORMAT_ND);
+
+    // acl_position * mk
+    int64_t tmp_output_ne[] = {src_ne[0], src_ne[1], src_ne[2], src_ne[3]};
+    size_t tmp_output_nb[GGML_MAX_DIMS];
+    tmp_output_nb[0] = ggml_type_size(dst->type);
+    for (int i = 1; i < GGML_MAX_DIMS; i++) {
+        tmp_output_nb[i] = tmp_output_nb[i - 1] * tmp_output_ne[i - 1];
+    }
+    ggml_cann_pool_alloc output_allocator(ctx.pool(), ggml_nbytes(dst));
+    void* tmp_output_buffer = output_allocator.get();
+    aclTensor* tmp_output_tensor = ggml_cann_create_tensor(
+        tmp_output_buffer, ggml_cann_type_mapping(dst->type),
+        ggml_type_size(dst->type), tmp_output_ne, tmp_output_nb, GGML_MAX_DIMS,
+        ACL_FORMAT_ND);
+    aclnn_mul(ctx, acl_position, tmp_mk_tensor, tmp_output_tensor);
+
+    // add
+    aclnn_add(ctx, tmp_output_tensor, acl_src, acl_dst);
+
+    ACL_CHECK(aclDestroyTensor(tmp_arange1_tensor));
+    ACL_CHECK(aclDestroyTensor(tmp_arange2_tensor));
+    ACL_CHECK(aclDestroyTensor(tmp_mk_base1_tensor));
+    ACL_CHECK(aclDestroyTensor(tmp_mk_base2_tensor));
+    ACL_CHECK(aclDestroyTensor(tmp_mk_base_tensor));
+    ACL_CHECK(aclDestroyTensor(tmp_arange_tensor));
+    ACL_CHECK(aclDestroyTensor(tmp_mk_tensor));
+    ACL_CHECK(aclDestroyTensor(tmp_output_tensor));
+}
+
+void ggml_cann_cpy(ggml_backend_cann_context& ctx, ggml_tensor* dst) {
+    ggml_cann_dup(ctx, dst);
+}
+
+/**
+ * @brief Performs element-wise addition of two tensors in place.
+ *
+ * This function adds the source tensor `acl_src` to the destination tensor
+ * `acl_dst` element-wise and stores the result in the destination tensor
+ * `acl_dst`.
+ *
+ * @param ctx The context for the CANN backend operations.
+ * @param acl_src The source tensor to be added.
+ * @param acl_dst The destination tensor which will hold the result of the
+ * addition.
+ */
+static void aclnn_inplace_add(ggml_backend_cann_context& ctx,
+                              aclTensor* acl_src, aclTensor* acl_dst) {
+    aclScalar* alpha = nullptr;
+    float alphaValue = 1.0f;
+    alpha = aclCreateScalar(&alphaValue, aclDataType::ACL_FLOAT);
+
+    uint64_t workspaceSize = 0;
+    aclOpExecutor* executor;
+    void* workspaceAddr = nullptr;
+
+    ACL_CHECK(aclnnInplaceAddGetWorkspaceSize(acl_dst, acl_src, alpha,
+                                              &workspaceSize, &executor));
+    if (workspaceSize > 0) {
+        ggml_cann_pool_alloc workspace_allocator(ctx.pool(), workspaceSize);
+        workspaceAddr = workspace_allocator.get();
+    }
+
+    ACL_CHECK(
+        aclnnInplaceAdd(workspaceAddr, workspaceSize, executor, ctx.stream()));
+
+    ACL_CHECK(aclDestroyScalar(alpha));
+}
+
+/**
+ * @brief Applies the softmax function to a tensor along a specified dimension.
+ *
+ * This function computes the softmax of the source tensor `acl_src` along the
+ * specified dimension `dim` and stores the result in the destination tensor
+ * `acl_dst`.
+ *
+ * @param ctx The context for the CANN backend operations.
+ * @param acl_src The source tensor on which the softmax function will be
+ * applied.
+ * @param dim The dimension along which the softmax function will be computed.
+ * @param acl_dst The destination tensor where the softmax results will be
+ * stored.
+ */
+static void aclnn_softmax(ggml_backend_cann_context& ctx, aclTensor* acl_src,
+                          int64_t dim, aclTensor* acl_dst) {
+    uint64_t workspaceSize = 0;
+    aclOpExecutor* executor;
+    void* workspaceAddr = nullptr;
+
+    ACL_CHECK(aclnnSoftmaxGetWorkspaceSize(acl_src, dim, acl_dst,
+                                           &workspaceSize, &executor));
+
+    if (workspaceSize > 0) {
+        ggml_cann_pool_alloc workspace_allocator(ctx.pool(), workspaceSize);
+        workspaceAddr = workspace_allocator.get();
+    }
+
+    aclrtStream stream = ctx.stream();
+    ACL_CHECK(aclnnSoftmax(workspaceAddr, workspaceSize, executor, stream));
+}
+
+void ggml_cann_softmax(ggml_backend_cann_context& ctx, ggml_tensor* dst) {
+    ggml_tensor* src0 = dst->src[0];
+    ggml_tensor* src1 = dst->src[1];  // mask
+
+    aclTensor* acl_src0 = ggml_cann_create_tensor(src0);
+    aclTensor* acl_dst = ggml_cann_create_tensor(dst);
+
+    float scale = 1.0f;
+    float max_bias = 0.0f;
+
+    memcpy(&scale, (float*)dst->op_params + 0, sizeof(float));
+    memcpy(&max_bias, (float*)dst->op_params + 1, sizeof(float));
+
+    // input mul scale
+    aclScalar* acl_scale = aclCreateScalar(&scale, aclDataType::ACL_FLOAT);
+
+    size_t n_bytes = ggml_nbytes(src0);
+    ggml_cann_pool_alloc mul_scale_allocator(ctx.pool(), n_bytes);
+    void* input_mul_scale_buffer = mul_scale_allocator.get();
+    aclTensor* acl_input_mul_scale_tensor = ggml_cann_create_tensor(
+        input_mul_scale_buffer, ACL_FLOAT, ggml_type_size(src0->type), src0->ne,
+        src0->nb, GGML_MAX_DIMS);
+
+    bool inplace = false;
+    aclnn_muls(ctx, acl_src0, scale, acl_input_mul_scale_tensor, inplace);
+
+    // mask
+    aclTensor* acl_src1_fp32_tensor = nullptr;
+    aclTensor* tmp_mask_tensor = nullptr;
+    ggml_cann_pool_alloc src1_fp32_allocator(ctx.pool());
+    if (src1) {
+        const bool use_f16 = src1->type == GGML_TYPE_F16;
+        if (use_f16) {
+            // cast to fp32
+            size_t n_bytes = ggml_nelements(src1) * sizeof(float_t);
+            size_t src1_fp32_nb[GGML_MAX_DIMS];
+            src1_fp32_nb[0] = sizeof(float_t);
+            for (int i = 1; i < GGML_MAX_DIMS; i++) {
+                src1_fp32_nb[i] = src1_fp32_nb[i - 1] * src1->ne[i - 1];
+            }
+            src1_fp32_allocator.alloc(n_bytes);
+            void* src1_fp32_buffer = src1_fp32_allocator.get();
+            acl_src1_fp32_tensor = ggml_cann_create_tensor(
+                src1_fp32_buffer, ACL_FLOAT, sizeof(float), src1->ne,
+                src1_fp32_nb, GGML_MAX_DIMS);
+            aclTensor* acl_src1 = ggml_cann_create_tensor(src1);
+            aclnn_cast(ctx, acl_src1, acl_src1_fp32_tensor, ACL_FLOAT);
+
+            ACL_CHECK(aclDestroyTensor(acl_src1));
+        } else {
+            acl_src1_fp32_tensor = ggml_cann_create_tensor(src1);
+        }
+
+        // broadcast the mask across rows, only use ne11 of ne01 in mask
+        if (src1->ne[1] != src0->ne[1]) {
+            // mask shape: [1,1,ne11,ne10]
+            int64_t tmp_mask_ne[] = {src0->ne[0], src0->ne[1], 1, 1};
+            size_t tmp_mask_nb[GGML_MAX_DIMS];
+            tmp_mask_nb[0] = sizeof(float_t);
+            for (int i = 1; i < GGML_MAX_DIMS; i++) {
+                tmp_mask_nb[i] = tmp_mask_nb[i - 1] * tmp_mask_ne[i - 1];
+            }
+            tmp_mask_tensor = ggml_cann_create_tensor(
+                src1->data, ACL_FLOAT, sizeof(float), tmp_mask_ne, tmp_mask_nb,
+                GGML_MAX_DIMS, ACL_FORMAT_ND);
+        }
+
+        // alibi
+        const int n_head = src0->ne[2];
+        const size_t src_nb0 = src0->nb[0];
+
+        n_bytes = ggml_nbytes(dst);
+        ggml_cann_pool_alloc output_allocator(ctx.pool(), n_bytes);
+        void* output_buffer = output_allocator.get();
+        aclTensor* alibi_output_tensor = ggml_cann_create_tensor(
+            output_buffer, ACL_FLOAT, ggml_type_size(dst->type), dst->ne,
+            dst->nb, GGML_MAX_DIMS);
+        if (max_bias <= 0.0f) {
+            // slope = 1.0
+            if (tmp_mask_tensor) {
+                aclnn_add(ctx, tmp_mask_tensor, acl_input_mul_scale_tensor,
+                          alibi_output_tensor);
+            } else {
+                aclnn_add(ctx, acl_src1_fp32_tensor, acl_input_mul_scale_tensor,
+                          alibi_output_tensor);
+            }
+        } else {
+            // slope != 1.0
+            if (tmp_mask_tensor) {
+                aclnn_alibi(ctx, acl_input_mul_scale_tensor, tmp_mask_tensor,
+                            alibi_output_tensor, n_head, src0->ne, src_nb0,
+                            max_bias, dst);
+            } else {
+                aclnn_alibi(ctx, acl_input_mul_scale_tensor,
+                            acl_src1_fp32_tensor, alibi_output_tensor, n_head,
+                            src0->ne, src_nb0, max_bias, dst);
+            }
+        }
+
+        // softmax
+        aclnn_softmax(ctx, alibi_output_tensor, 3, acl_dst);
+        ACL_CHECK(aclDestroyTensor(alibi_output_tensor));
+    } else {
+        aclnn_softmax(ctx, acl_input_mul_scale_tensor, 3, acl_dst);
+    }
+
+    ACL_CHECK(aclDestroyTensor(acl_src0));
+    ACL_CHECK(aclDestroyTensor(acl_src1_fp32_tensor));
+    ACL_CHECK(aclDestroyTensor(acl_dst));
+    ACL_CHECK(aclDestroyScalar(acl_scale));
+    ACL_CHECK(aclDestroyTensor(acl_input_mul_scale_tensor));
+    ACL_CHECK(aclDestroyTensor(tmp_mask_tensor));
+}
+
+void ggml_cann_get_rows(ggml_backend_cann_context& ctx, ggml_tensor* dst) {
+    ggml_tensor* src0 = dst->src[0];
+    ggml_tensor* src1 = dst->src[1];
+
+    ggml_cann_pool_alloc src0_extra_allocator(ctx.pool(), sizeof(ggml_tensor));
+    ggml_cann_pool_alloc src1_extra_allocator(ctx.pool(), sizeof(ggml_tensor));
+    ggml_cann_pool_alloc dst_extra_allocator(ctx.pool(), sizeof(ggml_tensor));
+    src0->extra = src0_extra_allocator.get();
+    src1->extra = src1_extra_allocator.get();
+    dst->extra = dst_extra_allocator.get();
+    ACL_CHECK(aclrtMemcpyAsync(src0->extra, sizeof(ggml_tensor), src0,
+                               sizeof(ggml_tensor), ACL_MEMCPY_HOST_TO_DEVICE,
+                               ctx.stream()));
+    ACL_CHECK(aclrtMemcpyAsync(src1->extra, sizeof(ggml_tensor), src1,
+                               sizeof(ggml_tensor), ACL_MEMCPY_HOST_TO_DEVICE,
+                               ctx.stream()));
+    ACL_CHECK(aclrtMemcpyAsync(dst->extra, sizeof(ggml_tensor), dst,
+                               sizeof(ggml_tensor), ACL_MEMCPY_HOST_TO_DEVICE,
+                               ctx.stream()));
+
+    switch (src0->type) {
+        case GGML_TYPE_F32:
+            aclrtlaunch_ascendc_get_row_f32(
+                24, ctx.stream(), src0->data, src1->data, dst->data,
+                ((ggml_tensor*)src0->extra)->ne,
+                ((ggml_tensor*)src0->extra)->nb,
+                ((ggml_tensor*)src1->extra)->ne,
+                ((ggml_tensor*)src1->extra)->nb, ((ggml_tensor*)dst->extra)->ne,
+                ((ggml_tensor*)dst->extra)->nb);
+            break;
+        case GGML_TYPE_F16:
+            aclrtlaunch_ascendc_get_row_f16(
+                24, ctx.stream(), src0->data, src1->data, dst->data,
+                ((ggml_tensor*)src0->extra)->ne,
+                ((ggml_tensor*)src0->extra)->nb,
+                ((ggml_tensor*)src1->extra)->ne,
+                ((ggml_tensor*)src1->extra)->nb, ((ggml_tensor*)dst->extra)->ne,
+                ((ggml_tensor*)dst->extra)->nb);
+            break;
+        case GGML_TYPE_Q4_0:
+            aclrtlaunch_ascendc_get_row_q4_0(
+                24, ctx.stream(), src0->data, src1->data, dst->data,
+                ((ggml_tensor*)src0->extra)->ne,
+                ((ggml_tensor*)src1->extra)->ne,
+                ((ggml_tensor*)src1->extra)->nb, ((ggml_tensor*)dst->extra)->ne,
+                ((ggml_tensor*)dst->extra)->nb);
+            break;
+        case GGML_TYPE_Q8_0:
+            aclrtlaunch_ascendc_get_row_q8_0(
+                24, ctx.stream(), src0->data, src1->data, dst->data,
+                ((ggml_tensor*)src0->extra)->ne,
+                ((ggml_tensor*)src1->extra)->ne,
+                ((ggml_tensor*)src1->extra)->nb, ((ggml_tensor*)dst->extra)->ne,
+                ((ggml_tensor*)dst->extra)->nb);
+            break;
+        default:
+            GGML_ABORT("fatal error");
+            break;
+    }
+}
+
+/**
+ * @brief Repeats elements of a tensor along a specified dimension.
+ *
+ * This function repeats each element of the source tensor `acl_src` a specified
+ * number of times (`repeats`) along the specified dimension `dim` and stores
+ * the result in the destination tensor `acl_dst`.
+ *
+ * @param ctx The context for the CANN backend operations.
+ * @param acl_src The source tensor whose elements will be repeated.
+ * @param acl_dst The destination tensor where the repeated elements will be
+ * stored.
+ * @param dim The dimension along which the elements will be repeated.
+ * @param repeats The number of times each element will be repeated.
+ * @param output_size The size of the output tensor.
+ */
+static void aclnn_repeat_interleave(ggml_backend_cann_context& ctx,
+                                    aclTensor* acl_src, aclTensor* acl_dst,
+                                    int64_t dim, int64_t repeats,
+                                    int64_t output_size) {
+    uint64_t workspaceSize = 0;
+    aclOpExecutor* executor;
+    void* workspaceAddr = nullptr;
+
+    ACL_CHECK(aclnnRepeatInterleaveIntWithDimGetWorkspaceSize(
+        acl_src, repeats, dim, output_size, acl_dst, &workspaceSize,
+        &executor));
+    if (workspaceSize > 0) {
+        ggml_cann_pool_alloc workspace_allocator(ctx.pool(), workspaceSize);
+        workspaceAddr = workspace_allocator.get();
+    }
+
+    ACL_CHECK(aclnnRepeatInterleaveIntWithDim(workspaceAddr, workspaceSize,
+                                              executor, ctx.stream()));
+}
+
+/**
+ * @brief Performs matrix multiplication of two tensors.
+ *
+ * This function computes the matrix multiplication of the input tensor
+ * `acl_input` and the weight tensor `acl_weight`, and stores the result in the
+ * destination tensor `acl_dst`.
+ * The operation is defined as:
+ * \f[
+ *     \text {acl_dst}=\text {acl_input@acl_weight}
+ * \f]
+ *
+ * @param ctx The context for the CANN backend operations.
+ * @param acl_input The input tensor for the matrix multiplication.
+ * @param acl_weight The weight tensor for the matrix multiplication.
+ * @param acl_dst The destination tensor where the result of the matrix
+ * multiplication will be stored.
+ */
+static void aclnn_mat_mul(ggml_backend_cann_context& ctx, aclTensor* acl_input,
+                          aclTensor* acl_weight, aclTensor* acl_dst) {
+    int8_t cube_math_type = 1;  // ALLOW_FP32_DOWN_PRECISION, when input is
+                                // fp32, atlas a2 will transpose it to HFLOAT32.
+
+    uint64_t workspaceSize = 0;
+    aclOpExecutor* executor;
+    void* workspaceAddr = nullptr;
+
+    ACL_CHECK(aclnnMatmulGetWorkspaceSize(acl_input, acl_weight, acl_dst,
+                                          cube_math_type, &workspaceSize,
+                                          &executor));
+
+    if (workspaceSize > 0) {
+        ggml_cann_pool_alloc workspace_allocator(ctx.pool(), workspaceSize);
+        workspaceAddr = workspace_allocator.get();
+    }
+
+    ACL_CHECK(
+        aclnnMatmul(workspaceAddr, workspaceSize, executor, ctx.stream()));
+}
+
+/**
+ * @brief Performs matrix multiplication with floating-point precision on
+ * tensors using the CANN backend.
+ *
+ * This function performs matrix multiplication of the input tensor and the
+ * weight tensor, handling broadcasting and transposing as needed, and stores
+ * the result in the destination tensor `dst`.
+ *
+ * @param ctx The context for the CANN backend operations.
+ * @param dst The destination tensor where the result of the matrix
+ * multiplication will be stored.
+ */
+static void ggml_cann_mat_mul_fp(ggml_backend_cann_context& ctx,
+                                 ggml_tensor* dst) {
+    ggml_tensor* weight = dst->src[0];  // weight
+    ggml_tensor* input = dst->src[1];   // input
+
+    // when weight ne2 or ne3 is 1, aclnnMatmulGetWorkspaceSize will auto
+    // broadcast, when weight ne2 or ne3 is not 1, weight need repeat.
+    BCAST_MUL_MAT_SHAPE(input, weight, dst);
+
+    // transpose weight: [1,2,3,4] -> [1,2,4,3]
+    int64_t transpose_ne[] = {bcast_weight_ne[1], bcast_weight_ne[0],
+                              bcast_weight_ne[2], bcast_weight_ne[3],
+                              bcast_weight_ne[4], bcast_weight_ne[5]};
+    size_t transpose_nb[] = {bcast_weight_nb[1], bcast_weight_nb[0],
+                             bcast_weight_nb[2], bcast_weight_nb[3],
+                             bcast_weight_nb[4], bcast_weight_nb[5]};
+
+    aclTensor* acl_weight_tensor =
+        ggml_cann_create_tensor(weight, transpose_ne, transpose_nb, bcast_dims);
+    aclTensor* acl_input_tensor =
+        ggml_cann_create_tensor(input, BCAST_MUL_MAT_PARAM(input));
+    aclTensor* acl_dst = ggml_cann_create_tensor(dst, BCAST_MUL_MAT_PARAM(dst));
+    aclnn_mat_mul(ctx, acl_input_tensor, acl_weight_tensor, acl_dst);
+
+    ACL_CHECK(aclDestroyTensor(acl_weight_tensor));
+    ACL_CHECK(aclDestroyTensor(acl_input_tensor));
+    ACL_CHECK(aclDestroyTensor(acl_dst));
+}
+
+/**
+ * @brief Performs matrix multiplication with quantized weights and
+ * floating-point inputs using the CANN backend.
+ *
+ * This function performs matrix multiplication of the input tensor `src1` and
+ * the weight tensor `src0`, handling broadcasting, transposing, and
+ * quantization as needed, and stores the result in the destination tensor
+ * `dst`.
+ *
+ * @param ctx The context for the CANN backend operations.
+ * @param dst The destination tensor where the result of the matrix
+ * multiplication will be stored.
+ */
+static void ggml_cann_mul_mat_quant(ggml_backend_cann_context& ctx,
+                                   ggml_tensor* dst,
+                                   const enum ggml_type type) {
+    ggml_tensor* src0 = dst->src[0];  // weight
+    ggml_tensor* src1 = dst->src[1];  // input
+
+    // The shape of the weight is NCHW. Matrix multiplication uses HW dims. HC
+    // is regarded as batch. weight need transpose.
+    int64_t weight_ne[] = {src0->ne[1], src0->ne[0]};
+    float weight_elem_size;
+    if (type == GGML_TYPE_Q4_0) {
+        weight_elem_size = float(sizeof(uint8_t)) / 2;
+    }
+    else if (type == GGML_TYPE_Q8_0) {
+        weight_elem_size = float(sizeof(uint8_t));
+    }
+    else {
+        GGML_ABORT("Only support Q4_0 and Q8_0 MUL_MAT");
+    }
+    float weight_nb[] = {weight_elem_size * src0->ne[0], weight_elem_size};
+
+    // size of one matrix is element_size * height * width.
+    size_t weight_stride = weight_elem_size * src0->ne[0] * src0->ne[1];
+    size_t weight_size = weight_stride * src0->ne[2] * src0->ne[3];
+
+    // scale stored at the end of weight. Also need transpose.
+    GGML_ASSERT(QK4_0 == QK8_0);
+    int64_t scale_ne[] = {src0->ne[1], src0->ne[0] / QK8_0};
+    size_t scale_elem_size = sizeof(uint16_t);
+    size_t scale_nb[] = {src0->ne[0] / QK8_0 * scale_elem_size,
+                         scale_elem_size};
+    size_t scale_stride = scale_elem_size * src0->ne[0] * src0->ne[1] / QK8_0;
+    char* scale_offset = (char*)src0->data + weight_size;
+
+    // input
+    void* input_buffer;
+    size_t input_elem_size = sizeof(uint16_t);
+    int64_t input_ne[] = {src1->ne[0], src1->ne[1]};
+    size_t input_nb[] = {input_elem_size, input_elem_size * src1->ne[0]};
+    size_t input_stride = input_elem_size * src1->ne[0] * src1->ne[1];
+
+    ggml_cann_pool_alloc input_alloctor(ctx.pool());
+    if (src1->type != GGML_TYPE_F16) {
+        aclTensor* acl_src1_tensor = ggml_cann_create_tensor(src1);
+        input_alloctor.alloc(ggml_nelements(src1) * input_elem_size);
+        input_buffer = input_alloctor.get();
+
+        int64_t* input_cast_ne = src1->ne;
+        size_t input_cast_nb[GGML_MAX_DIMS];
+        input_cast_nb[0] = sizeof(uint16_t);
+        for (int i = 1; i < GGML_MAX_DIMS; i++) {
+            input_cast_nb[i] = input_cast_nb[i - 1] * input_cast_ne[i - 1];
+        }
+
+        aclTensor* acl_input_tensor = ggml_cann_create_tensor(
+            input_buffer, ACL_FLOAT16, input_elem_size, input_cast_ne,
+            input_cast_nb, GGML_MAX_DIMS);
+        aclnn_cast(ctx, acl_src1_tensor, acl_input_tensor, ACL_FLOAT16);
+        ACL_CHECK(aclDestroyTensor(acl_input_tensor));
+        ACL_CHECK(aclDestroyTensor(acl_src1_tensor));
+    } else {
+        input_buffer = src1->data;
+    }
+
+    // output
+    size_t output_elem_size = sizeof(uint16_t);
+    int64_t output_ne[] = {dst->ne[0], dst->ne[1]};
+    size_t output_nb[] = {output_elem_size, output_elem_size * dst->ne[0]};
+    ggml_cann_pool_alloc output_alloctor(
+        ctx.pool(), ggml_nelements(dst) * output_elem_size);
+    void* output_buffer = output_alloctor.get();
+    size_t output_stride = output_elem_size * dst->ne[0] * dst->ne[1];
+
+    // aclnn
+    uint64_t workspaceSize = 0;
+    aclOpExecutor* executor;
+    void* workspaceAddr = nullptr;
+
+    for (int64_t n1 = 0; n1 < src1->ne[3]; n1++) {
+        for (int64_t c1 = 0; c1 < src1->ne[2]; c1++) {
+            int64_t n0 = n1 / (src1->ne[3] / src0->ne[3]);
+            int64_t c0 = c1 / (src1->ne[2] / src0->ne[2]);
+
+            int64_t batch1 = n1 * src1->ne[2] + c1;
+            int64_t batch0 = n0 * src0->ne[2] + c0;
+
+            aclTensor* acl_input_tensor = ggml_cann_create_tensor(
+                (char*)input_buffer + batch1 * input_stride, ACL_FLOAT16,
+                input_elem_size, input_ne, input_nb, 2);
+            aclTensor* acl_weight_tensor = ggml_cann_create_tensor(
+                (char*)src0->data + batch0 * weight_stride,
+                ggml_cann_type_mapping(type), weight_elem_size, weight_ne,
+                weight_nb, 2);
+            aclTensor* acl_scale_tensor = ggml_cann_create_tensor(
+                scale_offset + batch0 * scale_stride, ACL_FLOAT16,
+                scale_elem_size, scale_ne, scale_nb, 2);
+            aclTensor* acl_output_tensor = ggml_cann_create_tensor(
+                (char*)output_buffer + batch1 * output_stride, ACL_FLOAT16,
+                output_elem_size, output_ne, output_nb, 2);
+
+            ACL_CHECK(aclnnWeightQuantBatchMatmulV2GetWorkspaceSize(
+                acl_input_tensor, acl_weight_tensor, acl_scale_tensor, nullptr,
+                nullptr, nullptr, nullptr, QK8_0, acl_output_tensor,
+                &workspaceSize, &executor));
+
+            if (workspaceSize > 0 && workspaceAddr == nullptr) {
+                ggml_cann_pool_alloc workspace_allocator(ctx.pool(),
+                                                         workspaceSize);
+                workspaceAddr = workspace_allocator.get();
+            }
+
+            ACL_CHECK(aclnnWeightQuantBatchMatmulV2(
+                workspaceAddr, workspaceSize, executor, ctx.stream()));
+
+            ACL_CHECK(aclDestroyTensor(acl_input_tensor));
+            ACL_CHECK(aclDestroyTensor(acl_weight_tensor));
+            ACL_CHECK(aclDestroyTensor(acl_scale_tensor));
+            ACL_CHECK(aclDestroyTensor(acl_output_tensor));
+        }
+    }
+
+    // cast out
+    int64_t* output_cast_ne = dst->ne;
+    size_t output_cast_nb[GGML_MAX_DIMS];
+    output_cast_nb[0] = sizeof(uint16_t);
+    for (int i = 1; i < GGML_MAX_DIMS; i++) {
+        output_cast_nb[i] = output_cast_nb[i - 1] * output_cast_ne[i - 1];
+    }
+
+    aclTensor* acl_output_tensor =
+        ggml_cann_create_tensor(output_buffer, ACL_FLOAT16, output_elem_size,
+                                output_cast_ne, output_cast_nb, GGML_MAX_DIMS);
+    aclTensor* acl_dst_tensor = ggml_cann_create_tensor(dst);
+    aclnn_cast(ctx, acl_output_tensor, acl_dst_tensor, ACL_FLOAT);
+
+    ACL_CHECK(aclDestroyTensor(acl_output_tensor));
+    ACL_CHECK(aclDestroyTensor(acl_dst_tensor));
+}
+
+void ggml_cann_mul_mat(ggml_backend_cann_context& ctx, ggml_tensor* dst) {
+    const enum ggml_type type = dst->src[0]->type;
+    switch (type) {
+        case GGML_TYPE_F32:
+        case GGML_TYPE_F16:
+            ggml_cann_mat_mul_fp(ctx, dst);
+            break;
+        case GGML_TYPE_Q4_0:
+        case GGML_TYPE_Q8_0:
+            ggml_cann_mul_mat_quant(ctx, dst, type);
+            break;
+        default:
+            GGML_ABORT("fatal error");
+            break;
+    }
+}
+
+/**
+ * @brief Rolls the elements of a tensor along a specified dimension.
+ *
+ * This function rolls the elements of the source tensor `acl_src` by the
+ * specified shifts `shifts` along the specified dimensions `dims`, and stores
+ * the result in the destination tensor `acl_dst`.
+ *
+ * @param ctx The context for the CANN backend operations.
+ * @param acl_src The source tensor whose elements will be rolled.
+ * @param acl_dst The destination tensor where the rolled elements will be
+ * stored.
+ * @param shifts An array specifying the number of positions by which elements
+ * are shifted.
+ * @param dims An array specifying the dimensions along which elements are
+ * shifted.
+ */
+static void aclnn_roll(ggml_backend_cann_context& ctx, aclTensor* acl_src,
+                       aclTensor* acl_dst, int64_t* shifts, int64_t* dims) {
+    aclIntArray* acl_shifts = aclCreateIntArray(shifts, 1);
+    aclIntArray* acl_dims = aclCreateIntArray(dims, 1);
+
+    uint64_t workspaceSize = 0;
+    aclOpExecutor* executor;
+    void* workspaceAddr = nullptr;
+
+    ACL_CHECK(aclnnRollGetWorkspaceSize(acl_src, acl_shifts, acl_dims, acl_dst,
+                                        &workspaceSize, &executor));
+    if (workspaceSize > 0) {
+        ggml_cann_pool_alloc workspace_allocator(ctx.pool(), workspaceSize);
+        workspaceAddr = workspace_allocator.get();
+    }
+
+    ACL_CHECK(aclnnRoll(workspaceAddr, workspaceSize, executor, ctx.stream()));
+
+    ACL_CHECK(aclDestroyIntArray(acl_shifts));
+    ACL_CHECK(aclDestroyIntArray(acl_dims));
+}
+
+/**
+ * @brief Fills specified positions of a tensor with a scalar value.
+ *
+ * This function fills the positions in the source tensor `acl_src` specified by
+ * `index` along the dimension `dim` with the scalar value `value`.
+ *
+ * @param ctx The context for the CANN backend operations.
+ * @param acl_src The source tensor where the positions will be filled.
+ * @param dim The dimension along which the positions are specified.
+ * @param index An array specifying the positions to be filled.
+ * @param index_num The number of positions specified in the index array.
+ * @param value The scalar value used to fill the specified positions.
+ */
+static void aclnn_index_fill_tensor(ggml_backend_cann_context& ctx,
+                                    aclTensor* acl_src, int64_t dim,
+                                    int64_t* index, int64_t index_num,
+                                    float value) {
+    aclIntArray* acl_index = aclCreateIntArray(index, index_num);
+    aclScalar* acl_value = aclCreateScalar(&value, aclDataType::ACL_FLOAT);
+
+    uint64_t workspaceSize = 0;
+    aclOpExecutor* executor;
+    void* workspaceAddr = nullptr;
+
+    ACL_CHECK(aclnnInplaceIndexFillTensorGetWorkspaceSize(
+        acl_src, dim, acl_index, acl_value, &workspaceSize, &executor));
+    if (workspaceSize > 0) {
+        ggml_cann_pool_alloc workspace_allocator(ctx.pool(), workspaceSize);
+        workspaceAddr = workspace_allocator.get();
+    }
+
+    ACL_CHECK(aclnnInplaceIndexFillTensor(workspaceAddr, workspaceSize,
+                                          executor, ctx.stream()));
+
+    ACL_CHECK(aclDestroyIntArray(acl_index));
+    ACL_CHECK(aclDestroyScalar(acl_value));
+}
+
+static void aclnn_cache_init(ggml_backend_cann_context& ctx, ggml_tensor* dst,
+                             aclTensor* acl_cos_repeat_tensor,
+                             aclTensor* acl_sin_repeat_tensor,
+                             float theta_scale, bool is_neox) {
+    // int sin/cos cache, cache has different repeat method depond on
+    // @param.is_neox
+
+    ggml_tensor* src0 = dst->src[0];  // input
+    ggml_tensor* src1 = dst->src[1];  // position
+
+    // arange, [0,1,...,ne0/2]
+    int64_t arange_length = src0->ne[0] / 2;
+    ggml_cann_pool_alloc arange_allocator(ctx.pool(),
+                                          arange_length * sizeof(float_t));
+    void* arange_buffer = arange_allocator.get();
+    int64_t arange_ne[] = {arange_length, 1, 1, 1};
+    size_t arange_nb[] = {sizeof(float_t), sizeof(float_t), sizeof(float_t),
+                          arange_length * sizeof(float_t)};
+
+    aclTensor* acl_arange_tensor =
+        ggml_cann_create_tensor(arange_buffer, ACL_FLOAT, sizeof(float_t),
+                                arange_ne, arange_nb, GGML_MAX_DIMS);
+    float start = 0;
+    float step = 1;
+    float stop = src0->ne[0] / 2;
+    float n_elements = src0->ne[0] / 2;
+    aclnn_arange(ctx, acl_arange_tensor, start, stop, step, n_elements);
+
+    // power
+    // aclnnPowScalarTensor(): @param self is tensor which should be scalar, so
+    // use aclnn_pow_tensor_tensor() until fixed. aclScalar* acl_theta_scale =
+    // aclCreateScalar(&theta_scale, aclDataType::ACL_FLOAT);
+    // aclnn_power_scalar_tensor(ctx, acl_theta_scale, acl_arange_tensor,
+    // acl_power_tensor);
+    ggml_cann_pool_alloc theta_scale_allocator(ctx.pool(),
+                                               arange_length * sizeof(float_t));
+    void* theta_scale_buffer = theta_scale_allocator.get();
+    aclTensor* acl_theta_scale_tensor = aclnn_ones(
+        ctx, theta_scale_buffer, arange_length * sizeof(float_t), arange_ne,
+        GGML_MAX_DIMS, ACL_FLOAT, sizeof(float_t), theta_scale);
+    aclnn_pow_tensor_tensor(ctx, acl_theta_scale_tensor, acl_arange_tensor);
+
+    // position
+    GGML_ASSERT(src1->type == GGML_TYPE_I32);
+    int64_t position_length = src1->ne[0];
+    int64_t position_ne[] = {1, position_length, 1, 1};
+    size_t position_nb[] = {sizeof(int32_t), sizeof(int32_t),
+                            sizeof(int32_t) * position_length,
+                            sizeof(int32_t) * position_length};
+    aclTensor* acl_position_tensor = ggml_cann_create_tensor(
+        src1->data, ggml_cann_type_mapping(src1->type),
+        ggml_type_size(src1->type), position_ne, position_nb, GGML_MAX_DIMS);
+
+    // power * position
+    int64_t theta_length = arange_length * position_length;
+    ggml_cann_pool_alloc theta_allocator(ctx.pool(),
+                                         theta_length * sizeof(float_t));
+    void* theta_buffer = theta_allocator.get();
+    int64_t theta_ne[] = {arange_length, position_length, 1, 1};
+    size_t theta_nb[GGML_MAX_DIMS];
+    theta_nb[0] = sizeof(float_t);
+    for (int i = 1; i < GGML_MAX_DIMS; i++) {
+        theta_nb[i] = theta_nb[i - 1] * theta_ne[i - 1];
+    }
+    aclTensor* acl_theta_tensor =
+        ggml_cann_create_tensor(theta_buffer, ACL_FLOAT, sizeof(float_t),
+                                theta_ne, theta_nb, GGML_MAX_DIMS);
+    aclnn_mul(ctx, acl_position_tensor, acl_theta_scale_tensor,
+              acl_theta_tensor);
+
+    // permute: [0,1,2,3]->[0,2,1,3]
+    int64_t permute_ne[] = {arange_length, 1, position_length, 1};
+    size_t permute_nb[GGML_MAX_DIMS];
+    permute_nb[0] = sizeof(float_t);
+    for (int i = 1; i < GGML_MAX_DIMS; i++) {
+        permute_nb[i] = permute_nb[i - 1] * permute_ne[i - 1];
+    }
+    ggml_cann_pool_alloc permute_allocator(ctx.pool(),
+                                           theta_length * sizeof(float_t));
+    void* permute_buffer = permute_allocator.get();
+    aclTensor* acl_permute_tensor = ggml_cann_create_tensor(
+        permute_buffer, ACL_FLOAT, sizeof(float_t), permute_ne, permute_nb,
+        GGML_MAX_DIMS, ACL_FORMAT_ND);
+    int64_t permute_dim[] = {0, 2, 1, 3};
+    int64_t num_dims = 4;
+    aclnn_permute(ctx, acl_theta_tensor, acl_permute_tensor, permute_dim,
+                  num_dims);
+
+    // sin/cos
+    ggml_cann_pool_alloc sin_allocator(ctx.pool(),
+                                       theta_length * sizeof(float_t));
+    void* sin_buffer = sin_allocator.get();
+    aclTensor* acl_sin_tensor = ggml_cann_create_tensor(
+        sin_buffer, ACL_FLOAT, sizeof(float_t), permute_ne, permute_nb,
+        GGML_MAX_DIMS, ACL_FORMAT_ND);
+    aclnn_sin(ctx, acl_permute_tensor, acl_sin_tensor);
+
+    ggml_cann_pool_alloc cos_allocator(ctx.pool(),
+                                       theta_length * sizeof(float_t));
+    void* cos_buffer = cos_allocator.get();
+    aclTensor* acl_cos_tensor = ggml_cann_create_tensor(
+        cos_buffer, ACL_FLOAT, sizeof(float_t), permute_ne, permute_nb,
+        GGML_MAX_DIMS, ACL_FORMAT_ND);
+    aclnn_cos(ctx, acl_permute_tensor, acl_cos_tensor);
+
+    // repeat
+    if (is_neox) {
+        int64_t repeatsArray[] = {1, 1, 1, 2};
+        aclnn_repeat(ctx, acl_sin_tensor, acl_sin_repeat_tensor, repeatsArray);
+        aclnn_repeat(ctx, acl_cos_tensor, acl_cos_repeat_tensor, repeatsArray);
+    } else {
+        int64_t num_repeats = 2;
+        int64_t dim = 3;
+        int64_t output_size = arange_length * num_repeats;
+        aclnn_repeat_interleave(ctx, acl_sin_tensor, acl_sin_repeat_tensor, dim,
+                                num_repeats, output_size);
+        aclnn_repeat_interleave(ctx, acl_cos_tensor, acl_cos_repeat_tensor, dim,
+                                num_repeats, output_size);
+    }
+
+    // release
+    ACL_CHECK(aclDestroyTensor(acl_arange_tensor));
+    ACL_CHECK(aclDestroyTensor(acl_theta_scale_tensor));
+    ACL_CHECK(aclDestroyTensor(acl_position_tensor));
+    ACL_CHECK(aclDestroyTensor(acl_theta_tensor));
+    ACL_CHECK(aclDestroyTensor(acl_permute_tensor));
+    ACL_CHECK(aclDestroyTensor(acl_sin_tensor));
+    ACL_CHECK(aclDestroyTensor(acl_cos_tensor));
+}
+
+void ggml_cann_rope(ggml_backend_cann_context& ctx, ggml_tensor* dst) {
+    // TODO: use ascendc
+    // Only test with LLAMA model.
+    ggml_tensor* src0 = dst->src[0];  // input
+    ggml_tensor* src2 = dst->src[2];  // freq_factors
+
+    // TODO: with freq_factors
+    GGML_ASSERT(src2 == NULL);
+
+    // param
+    float freq_base, freq_scale, ext_factor, attn_factor, beta_fast, beta_slow;
+    // const int n_past     = ((int32_t *) dst->op_params)[0];
+    const int n_dims = ((int32_t*)dst->op_params)[1];
+    const int mode = ((int32_t*)dst->op_params)[2];
+    // const int n_ctx      = ((int32_t *) dst->op_params)[3];
+    const int n_ctx_orig = ((int32_t*)dst->op_params)[4];
+
+    GGML_TENSOR_UNARY_OP_LOCALS
+
+    memcpy(&freq_base, (int32_t*)dst->op_params + 5, sizeof(float));
+    memcpy(&freq_scale, (int32_t*)dst->op_params + 6, sizeof(float));
+    memcpy(&ext_factor, (int32_t*)dst->op_params + 7, sizeof(float));
+    memcpy(&attn_factor, (int32_t*)dst->op_params + 8, sizeof(float));
+    memcpy(&beta_fast, (int32_t*)dst->op_params + 9, sizeof(float));
+    memcpy(&beta_slow, (int32_t*)dst->op_params + 10, sizeof(float));
+
+    GGML_ASSERT(n_dims <= ne0);
+    GGML_ASSERT(n_dims % 2 == 0);
+
+    // TODO: ext_factor != 0
+    GGML_ASSERT(ext_factor == 0);
+    // TODO: freq_scale != 1
+    GGML_ASSERT(freq_scale == 1);
+
+    const float theta_scale = powf(freq_base, -2.0f / n_dims);
+
+    float corr_dims[2];
+    ggml_rope_yarn_corr_dims(n_dims, n_ctx_orig, freq_base, beta_fast,
+                             beta_slow, corr_dims);
+
+    const bool is_neox = mode & GGML_ROPE_TYPE_NEOX;
+
+    // init cos/sin cache
+    ggml_cann_pool_alloc sin_allocator(
+        ctx.pool(), src0->ne[0] * src0->ne[2] * sizeof(float_t));
+    ggml_cann_pool_alloc cos_allocator(
+        ctx.pool(), src0->ne[0] * src0->ne[2] * sizeof(float_t));
+    void* sin_buffer = sin_allocator.get();
+    void* cos_buffer = cos_allocator.get();
+
+    int64_t sin_reshape_ne[4] = {src0->ne[0], 1, src0->ne[2], 1};
+    size_t sin_reshape_nb[GGML_MAX_DIMS];
+    sin_reshape_nb[0] = sizeof(float_t);
+    for (int i = 1; i < GGML_MAX_DIMS; i++) {
+        sin_reshape_nb[i] = sin_reshape_nb[i - 1] * sin_reshape_ne[i - 1];
+    }
+    aclTensor* acl_sin_reshape_tensor =
+        ggml_cann_create_tensor(sin_buffer, ACL_FLOAT, sizeof(float_t),
+                                sin_reshape_ne, sin_reshape_nb, GGML_MAX_DIMS);
+    aclTensor* acl_cos_reshape_tensor =
+        ggml_cann_create_tensor(cos_buffer, ACL_FLOAT, sizeof(float_t),
+                                sin_reshape_ne, sin_reshape_nb, GGML_MAX_DIMS);
+    aclnn_cache_init(ctx, dst, acl_cos_reshape_tensor, acl_sin_reshape_tensor,
+                     theta_scale, is_neox);
+
+    // roll input
+    void* input_roll_buffer;
+    aclTensor* acl_minus_one_tensor;
+    void* minus_one_scale_buffer = nullptr;
+    ggml_cann_pool_alloc roll_allocator(ctx.pool(), ggml_nbytes(src0));
+    ggml_cann_pool_alloc minus_one_scale_allocator(
+        ctx.pool(), sizeof(float_t) * src0->ne[0]);
+    if (!is_neox) {
+        // roll input: [q0,q1,q2,q3,...] -> [q1,q0,q3,q2,...]
+        input_roll_buffer = roll_allocator.get();
+        int64_t input_roll_ne[4] = {2, src0->ne[1] * (src0->ne[0] / 2),
+                                    src0->ne[2], src0->ne[3]};
+        size_t input_roll_nb[GGML_MAX_DIMS];
+        input_roll_nb[0] = ggml_type_size(src0->type);
+        for (int i = 1; i < GGML_MAX_DIMS; i++) {
+            input_roll_nb[i] = input_roll_nb[i - 1] * input_roll_ne[i - 1];
+        }
+        aclTensor* acl_input_roll_tensor = ggml_cann_create_tensor(
+            input_roll_buffer, ggml_cann_type_mapping(src0->type),
+            ggml_type_size(src0->type), input_roll_ne, input_roll_nb,
+            GGML_MAX_DIMS);
+        aclTensor* acl_input_tensor = ggml_cann_create_tensor(
+            src0->data, ggml_cann_type_mapping(src0->type),
+            ggml_type_size(src0->type), input_roll_ne, input_roll_nb,
+            GGML_MAX_DIMS);
+
+        int64_t shifts[] = {1};
+        int64_t dims[] = {3};
+        aclnn_roll(ctx, acl_input_tensor, acl_input_roll_tensor, shifts, dims);
+        ACL_CHECK(aclDestroyTensor(acl_input_roll_tensor));
+        ACL_CHECK(aclDestroyTensor(acl_input_tensor));
+
+        // init [-1, 1, -1, 1, ...]
+        minus_one_scale_buffer = minus_one_scale_allocator.get();
+
+        int64_t minus_one_ne[4] = {src0->ne[0], 1, 1, 1};
+        size_t minus_one_nb[GGML_MAX_DIMS];
+        minus_one_nb[0] = sizeof(float_t);
+        for (int i = 1; i < GGML_MAX_DIMS; i++) {
+            minus_one_nb[i] = minus_one_nb[i - 1] * minus_one_ne[i - 1];
+        }
+        acl_minus_one_tensor = aclnn_ones(
+            ctx, minus_one_scale_buffer, sizeof(float_t) * src0->ne[0],
+            minus_one_ne, GGML_MAX_DIMS, ACL_FLOAT, sizeof(float_t), 1);
+        int64_t dim = 3;
+        int64_t* index = new int64_t[src0->ne[0]];
+        for (int i = 0; i < src0->ne[0]; i++) {
+            index[i] = i / 2 * 2;
+        }
+        int64_t index_num = src0->ne[0];
+        float value = -1;
+        aclnn_index_fill_tensor(ctx, acl_minus_one_tensor, dim, index,
+                                index_num, value);
+    } else {
+        // roll input: [q0,q1,q2,...] ->
+        // [q_half,q_half+1,...,q_end,q0,q1,...q_half-1]
+        input_roll_buffer = roll_allocator.get();
+        aclTensor* acl_input_roll_tensor = ggml_cann_create_tensor(
+            input_roll_buffer, ggml_cann_type_mapping(src0->type),
+            ggml_type_size(src0->type), src0->ne, src0->nb, GGML_MAX_DIMS);
+        aclTensor* acl_input_tensor = ggml_cann_create_tensor(src0);
+
+        int64_t shifts[] = {src0->ne[0] / 2};
+        int64_t dims[] = {3};
+        aclnn_roll(ctx, acl_input_tensor, acl_input_roll_tensor, shifts, dims);
+
+        ACL_CHECK(aclDestroyTensor(acl_input_roll_tensor));
+        ACL_CHECK(aclDestroyTensor(acl_input_tensor));
+
+        // init [-1, -1, -1, 1, 1，1，...]
+        minus_one_scale_buffer = minus_one_scale_allocator.get();
+
+        int64_t minus_one_ne[4] = {src0->ne[0], 1, 1, 1};
+        size_t minus_one_nb[GGML_MAX_DIMS];
+        minus_one_nb[0] = sizeof(float_t);
+        for (int i = 1; i < GGML_MAX_DIMS; i++) {
+            minus_one_nb[i] = minus_one_nb[i - 1] * minus_one_ne[i - 1];
+        }
+        acl_minus_one_tensor = aclnn_ones(
+            ctx, minus_one_scale_buffer, sizeof(float_t) * src0->ne[0],
+            minus_one_ne, GGML_MAX_DIMS, ACL_FLOAT, sizeof(float_t), 1);
+        // -1 * first half
+        int64_t first_half_ne[4] = {src0->ne[0] / 2, 1, 1, 1};
+        size_t first_half_nb[GGML_MAX_DIMS];
+        first_half_nb[0] = sizeof(float_t);
+        for (int i = 1; i < GGML_MAX_DIMS; i++) {
+            first_half_nb[i] = first_half_nb[i - 1] * first_half_ne[i - 1];
+        }
+        aclTensor* acl_first_half_tensor = ggml_cann_create_tensor(
+            minus_one_scale_buffer, ACL_FLOAT, sizeof(float_t), first_half_ne,
+            first_half_nb, GGML_MAX_DIMS);
+        bool inplace = true;
+        float scale = -1;
+        aclnn_muls(ctx, acl_first_half_tensor, scale, nullptr, inplace);
+        ACL_CHECK(aclDestroyTensor(acl_first_half_tensor));
+    }
+
+    // TODO: n_dims < ne0
+    GGML_ASSERT(n_dims == src0->ne[0]);
+
+    // input * scale
+    ggml_cann_pool_alloc roll_mul_scale_allocator(ctx.pool(),
+                                                  ggml_nbytes(src0));
+    void* input_roll_mul_scale_buffer = roll_mul_scale_allocator.get();
+    size_t input_nb[GGML_MAX_DIMS];
+    input_nb[0] = ggml_type_size(src0->type);
+    for (int i = 1; i < GGML_MAX_DIMS; i++) {
+        input_nb[i] = input_nb[i - 1] * src0->ne[i - 1];
+    }
+    aclTensor* acl_input_roll_mul_scale_tensor = ggml_cann_create_tensor(
+        input_roll_mul_scale_buffer, ggml_cann_type_mapping(src0->type),
+        ggml_type_size(src0->type), src0->ne, input_nb, GGML_MAX_DIMS);
+    aclTensor* acl_input_roll_reshape_tensor = ggml_cann_create_tensor(
+        input_roll_buffer, ggml_cann_type_mapping(src0->type),
+        ggml_type_size(src0->type), src0->ne, input_nb, GGML_MAX_DIMS);
+
+    aclnn_mul(ctx, acl_input_roll_reshape_tensor, acl_minus_one_tensor,
+              acl_input_roll_mul_scale_tensor);
+
+    // output
+    aclTensor* acl_src0 = ggml_cann_create_tensor(src0);
+    aclTensor* acl_dst = ggml_cann_create_tensor(dst);
+    void* output_fp32_buffer;
+    if (src0->type == GGML_TYPE_F32) {
+        aclnn_inplace_mul(ctx, acl_src0, acl_cos_reshape_tensor);
+        aclnn_inplace_mul(ctx, acl_input_roll_mul_scale_tensor,
+                          acl_sin_reshape_tensor);
+        aclnn_add(ctx, acl_src0, acl_input_roll_mul_scale_tensor, acl_dst);
+        // TODO: ne0 != n_dims in mode2
+    } else if (src0->type == GGML_TYPE_F16) {
+        size_t input_fp32_nb[GGML_MAX_DIMS];
+        input_fp32_nb[0] = sizeof(float_t);
+        for (int i = 1; i < GGML_MAX_DIMS; i++) {
+            input_fp32_nb[i] = input_fp32_nb[i - 1] * dst->ne[i - 1];
+        }
+        ggml_cann_pool_alloc fp32_allocator1(
+            ctx.pool(), ggml_nelements(dst) * sizeof(float_t));
+        void* input_fp32_buffer1 = fp32_allocator1.get();
+        aclTensor* input_fp32_tensor1 = ggml_cann_create_tensor(
+            input_fp32_buffer1, ACL_FLOAT, sizeof(float_t), dst->ne,
+            input_fp32_nb, GGML_MAX_DIMS);
+        ggml_cann_pool_alloc fp32_allocator2(
+            ctx.pool(), ggml_nelements(dst) * sizeof(float_t));
+        void* input_fp32_buffer2 = fp32_allocator2.get();
+        aclTensor* input_fp32_tensor2 = ggml_cann_create_tensor(
+            input_fp32_buffer2, ACL_FLOAT, sizeof(float_t), dst->ne,
+            input_fp32_nb, GGML_MAX_DIMS);
+
+        ggml_cann_pool_alloc fp32_allocator(
+            ctx.pool(), ggml_nelements(dst) * sizeof(float_t));
+        output_fp32_buffer = fp32_allocator.get();
+        aclTensor* output_fp32_tensor = ggml_cann_create_tensor(
+            output_fp32_buffer, ACL_FLOAT, sizeof(float_t), dst->ne,
+            input_fp32_nb, GGML_MAX_DIMS);
+        aclnn_mul(ctx, acl_src0, acl_cos_reshape_tensor, input_fp32_tensor1);
+        aclnn_mul(ctx, acl_input_roll_mul_scale_tensor, acl_sin_reshape_tensor,
+                  input_fp32_tensor2);
+        aclnn_add(ctx, input_fp32_tensor1, input_fp32_tensor2,
+                  output_fp32_tensor);
+        aclnn_cast(ctx, output_fp32_tensor, acl_dst, ACL_FLOAT16);
+
+        ACL_CHECK(aclDestroyTensor(input_fp32_tensor1));
+        ACL_CHECK(aclDestroyTensor(input_fp32_tensor2));
+        ACL_CHECK(aclDestroyTensor(output_fp32_tensor));
+    }
+
+    ACL_CHECK(aclDestroyTensor(acl_sin_reshape_tensor));
+    ACL_CHECK(aclDestroyTensor(acl_cos_reshape_tensor));
+    ACL_CHECK(aclDestroyTensor(acl_minus_one_tensor));
+    ACL_CHECK(aclDestroyTensor(acl_input_roll_mul_scale_tensor));
+    ACL_CHECK(aclDestroyTensor(acl_input_roll_reshape_tensor));
+    ACL_CHECK(aclDestroyTensor(acl_src0));
+    ACL_CHECK(aclDestroyTensor(acl_dst));
+}
diff --git a/src/whisper_cpp/ggml/src/ggml-cann/aclnn_ops.h b/src/whisper_cpp/ggml/src/ggml-cann/aclnn_ops.h
new file mode 100644
index 00000000..680129c7
--- /dev/null
+++ b/src/whisper_cpp/ggml/src/ggml-cann/aclnn_ops.h
@@ -0,0 +1,592 @@
+#ifndef CANN_ACLNN_OPS
+#define CANN_ACLNN_OPS
+
+/**
+ * @file    acl_tensor
+ * @brief   This file contains related functions of ggml_tensor and acl_tensor.
+ *          Contains conversion from ggml_tensor to acl_tensor, broadcast and other
+ *          functions.
+ * @author  hipudding <huafengchun@gmail.com>
+ * @author  wangshuai09 <391746016@qq.com>
+ * @date    July 15, 2024
+ *
+ * Copyright (c) 2023-2024 The ggml authors
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a copy
+ * of this software and associated documentation files (the "Software"), to
+ * deal in the Software without restriction, including without limitation the
+ * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or
+ * sell copies of the Software, and to permit persons to whom the Software is
+ * furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice shall be included in
+ * all copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+ * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
+ * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
+ * IN THE SOFTWARE.
+ */
+
+#include <aclnnop/aclnn_add.h>
+#include <aclnnop/aclnn_arange.h>
+#include <aclnnop/aclnn_argsort.h>
+#include <aclnnop/aclnn_cat.h>
+#include <aclnnop/aclnn_clamp.h>
+#include <aclnnop/aclnn_div.h>
+#include <aclnnop/aclnn_gelu.h>
+#include <aclnnop/aclnn_hardsigmoid.h>
+#include <aclnnop/aclnn_hardswish.h>
+#include <aclnnop/aclnn_leaky_relu.h>
+#include <aclnnop/aclnn_mul.h>
+#include <aclnnop/aclnn_relu.h>
+#include <aclnnop/aclnn_silu.h>
+#include <aclnnop/aclnn_tanh.h>
+#include "acl_tensor.h"
+#include "common.h"
+
+/**
+ * @brief   Repeats a ggml tensor along each dimension to match the dimensions
+ *          of another tensor.
+ *
+ * @details This function repeats the elements of a source ggml tensor along
+ *          each dimension to create a destination tensor with the specified
+ *          dimensions. The operation is performed using the ACL backend and
+ *          executed asynchronously on the device.
+ *
+ * @param   ctx The CANN context used for operations.
+ * @param   dst The ggml tensor representing the destination, which op is
+ *              GGML_OP_REPEAT and specifies the desired dimensions.
+ */
+void ggml_cann_repeat(ggml_backend_cann_context& ctx, ggml_tensor* dst);
+
+/**
+ * @brief   Adds two ggml tensors using the CANN backend.
+ *
+ * @details This function performs an element-wise addition of two tensors. In
+ *          case the tensors do not have the same shape, one or both tensors
+ *          will be broadcasted to match the shape of the other before the
+ *          addition is performed.The formula for the operation is given by:
+ *          \f[
+ *              \text{dst} = \text{acl_src0} + \alpha \cdot \text{acl_src1}
+ *          \f]
+ *
+ * @param ctx The CANN context used for operations.
+ * @param dst The ggml tensor representing the destination, result of the
+ *            addition is stored at dst->data, and dst->op is `GGML_OP_ADD`
+ */
+void ggml_cann_add(ggml_backend_cann_context& ctx, ggml_tensor* dst);
+
+/**
+ * @brief   Applies the Leaky ReLU activation function to a tensor using the CANN
+ *          backend.
+ *
+ * @details This function computes the Leaky ReLU activation for each element of
+ *          the input tensor. The Leaky ReLU function allows a small gradient
+ *          when the unit is not active (i.e., when the input is negative). The
+ *          Leaky ReLU function is defined as:
+ *          \f[
+ *              \text{dst} = \max(0, src) + \text{negativeSlope} \cdot \min(0,
+ *               src)
+ *          \f]
+ *          `negativeSlope` is in dst->params.
+ *
+ * @param ctx The CANN context used for operations.
+ * @param dst The destination tensor where the result of the Leaky ReLU
+ *            activation is stored, which op is `GGML_OP_LEAKY_RELU`
+ */
+void ggml_cann_leaky_relu(ggml_backend_cann_context& ctx, ggml_tensor* dst);
+
+/**
+ * @brief    Concatenates multiple tensors along a specified dimension using the
+ *           CANN backend.
+ *
+ * @param ctx        The CANN context used for operations.
+ * @param tensorList A pointer to the list of tensors to be concatenated.
+ * @param dst        The destination tensor where the result of the
+ *                   concatenation is stored. dst->op is `GGML_OP_CONCAT`.
+ * @param concat_dim The dimension along which the tensors are concatenated.
+ *
+ * @attention tensorList length should be 2 and the dimension using for concat
+ *            default to 1.
+ */
+void ggml_cann_concat(ggml_backend_cann_context& ctx, ggml_tensor* dst);
+
+/**
+ * @brief   Generates a sequence of evenly spaced values within a specified
+ *          interval for a ggml tensor using the CANN backend.
+ *
+ * @details This function creates a sequence of numbers over a specified i
+ *          nterval, starting from `start`, ending before `stop`, and
+ *          incrementing by `step`. The sequence is stored in the destination
+ *          tensor `dst`.
+ *
+ * @param ctx The CANN context used for operations.
+ * @param dst The destination tensor where the generated sequence will be stored.
+ *            `start`, 'stop' and 'step' are in dst->op_params and dst->op is
+ *            `GGML_OP_ARANGE`.
+ */
+void ggml_cann_arange(ggml_backend_cann_context& ctx, ggml_tensor* dst);
+
+/**
+ * @brief   Computes the square of the elements of a ggml tensor using the CANN
+ *          backend.
+ * @details The function sets the second source tensor of the destination
+ *          tensor `dst` to be equal to the first source tensor. This is
+ *          effectively squaring the elements since the multiplication becomes
+ *          `element * element`.
+ * @param ctx The CANN context used for operations.
+ * @param dst The destination tensor where the squared values will be stored，
+ *            which dst->op is `GGML_OP_SQR`.
+ */
+void ggml_cann_sqr(ggml_backend_cann_context& ctx, ggml_tensor* dst);
+
+/**
+ * @brief   Applies a clamp operation to the elements of a ggml tensor using the
+ *          CANN backend.
+ *
+ * @details This function clamps the elements of the input tensor `src` to a
+ *          specified range defined by `min` and `max` values. The result is
+ *          stored in the destination tensor `dst`. The operation is defined as:
+ *          \f[
+ *              y = \max(\min(x, max\_value), min\_value)
+ *           \f]
+ *          where `x` is an element of the input tensor, and `y` is the
+ *          corresponding element in the output tensor.
+ * @param ctx The CANN context used for operations.
+ * @param dst The destination tensor where the clamped values will be stored.
+ *            dst->op is `GGML_OP_CLAMP`, `min` and `max` value is in dst->params.
+ */
+void ggml_cann_clamp(ggml_backend_cann_context& ctx, ggml_tensor* dst);
+
+/**
+ * @brief   Scales the elements of a ggml tensor by a constant factor using the
+ *          CANN backend.
+ *
+ * @details This function multiplies each element of the input tensor `src` by
+ *          a scaling factor `scale`, storing the result in the destination
+ *          tensor `dst`. The operation is defined as:
+ *          \f[
+ *             dst = src \times scale
+ *          \f]
+ *
+ * @param ctx The CANN context used for operations.
+ * @param dst The destination tensor where the scaled values will be stored.
+ *            dst->op is `GGML_OP_SCALE` and `scale` value is in dst->params.
+ */
+void ggml_cann_scale(ggml_backend_cann_context& ctx, ggml_tensor* dst);
+
+/**
+ * @brief   Sorts the elements of a ggml tensor and returns the indices that
+ *          would sort the tensor using the CANN backend.
+ *
+ * @details This function performs an argsort operation on the input tensor
+ *          `src`. It sorts the elements of `src` in either ascending or
+ *          descending order, depending on the `GGML_SORT_ORDER_DESC`,
+ *          and returns the indices that would sort the original tensor.
+ *
+ * @param ctx The CANN context used for operations.
+ * @param dst The destination tensor where the sorted indices will be stored.
+ *            dst->op is `GGML_OP_ARGSORT`.
+ */
+void ggml_cann_argsort(ggml_backend_cann_context& ctx, ggml_tensor* dst);
+
+/**
+ * @brief   Computes the Layer Normalization for a ggml tensor using the CANN
+ *          backend.
+ *
+ * @details This function applies the Layer Normalization operation on the
+ *          input tensor `src` and stores the result in the destination tensor
+ *          `dst`. Layer Normalization normalizes the features at each sample in
+ *          a mini-batch independently. It is commonly used in neural networks
+ *          to normalize the activations of a layer by adjusting and scaling
+ *          the outputs.
+ *          The operation is defined as:
+ *          \f[
+ *              \text { out }=\frac{x-\mathrm{E}[x]}{\sqrt{\text{Var}[x]+eps}}
+ *          \f]
+ *          `Var` defaults dst->ne[0]. `eps` is in dst->params.
+ *
+ * @param ctx The CANN context used for operations.
+ * @param dst The destination tensor where the normalized values will be stored.
+ * @attention `Var` defaults to dst->ne[0].
+ */
+void ggml_cann_norm(ggml_backend_cann_context& ctx, ggml_tensor* dst);
+
+/**
+ * @brief  Computes the Group Normalization for a ggml tensor using the CANN
+ *         backend.
+ *
+ * @brief  This function applies the Group Normalization operation on the input
+ *         tensor `src` and stores the result in the destination tensor `dst`.
+ *         Group Normalization divides the channels into groups and normalizes
+ *         the features within each group across spatial locations.
+ *         It is commonly used in convolutional neural networks to improve
+ *         training stability and performance.
+ *         The operation is defined as:
+ *         \f[
+ *             \text { out }=\frac{x-\mathrm{E}[x]}{\sqrt{\text{Var}[x]+eps}}
+ *         \f]
+ *
+ * @param ctx The CANN context used for operations.
+ * @param dst The destination tensor where the normalized values will be stored.
+ *            `n_groups` is in dst->params, which split C channel to `n_groups`.
+ *            dst->op is `GGML_OP_GROUP_NORM`.
+ *
+ * @attention eps defaults to 1e-6f.
+ */
+void ggml_cann_group_norm(ggml_backend_cann_context& ctx, ggml_tensor* dst);
+
+/**
+ * @brief   Computes the accumulation of tensors using the CANN backend.
+ *
+ * @details This function performs an accumulation operation on two tensors.
+ *          Depending on the `inplace` flag, it either updates the destination
+ *          tensor `dst` in place by adding `alpha * src1` to it, or it creates
+ *          a new tensor as the result of `src0 + alpha * src1` and stores it in
+ *          `dst`.
+ *          The operation is defined as:
+ *          \f[
+ *               dst = src0 + alpha \times src1
+ *          \f]
+ *          if `inplace` is `true`, `src0` is equal to 'dst'.
+ * @param ctx The CANN context used for operations.
+ * @param dst The destination tensor where the accumulated values will be stored.
+ *            `inplace` is in dst->params, and dst->op is `GGML_OP_ACC`.
+ */
+void ggml_cann_acc(ggml_backend_cann_context& ctx, ggml_tensor* dst);
+
+/**
+ * @brief   Computes the sum of elements along the last dimension of a ggml tensor
+ *          using the CANN backend.
+ *
+ * @details This function performs a reduction sum operation along the last
+ *          dimension of the input tensor `src`. The result of the sum is stored
+ *          in the destination tensor `dst`.
+ *
+ * @param ctx The CANN context used for operations.
+ * @param dst The destination tensor where the reduced values will be stored。
+ *            dst->op is `GGML_OP_SUM_ROWS`.
+ *
+ * @attention `reduce_dims` defaults to 3, which means the last dimension.
+ */
+void ggml_cann_sum_rows(ggml_backend_cann_context& ctx, ggml_tensor* dst);
+
+/**
+ * @brief   Upsamples a ggml tensor using nearest neighbor interpolation using
+ *          the CANN backend.
+ *
+ * @details This function performs upsampling of the input tensor `src` using
+ *          nearest neighbor interpolation. The upsampling is applied to the
+ *          height and width dimensions (last two dimensions) of the tensor. The
+ *          result is stored in the destination tensor `dst`, which must have
+ *          the appropriate dimensions for the upsampled output.
+ *
+ * @param ctx The CANN context used for operations.
+ * @param dst The destination tensor where the upsampled values will be stored.
+ *            dst->op is `GGML_OP_UPSCALE`.
+ */
+void ggml_cann_upsample_nearest2d(ggml_backend_cann_context& ctx,
+                                  ggml_tensor* dst);
+
+/**
+ * @brief   Pads a ggml tensor to match the dimensions of the destination tensor
+ *          using the CANN backend.
+ *
+ * @details This function pads the input tensor `src` so that it matches the
+ *          dimensions of the destination tensor `dst`. The amount of padding
+ *          is calculated based on the difference in sizes between `src` and
+ *          `dst` along each dimension. The padded tensor is stored in `dst`.
+ *
+ * @param ctx The CANN context used for operations.
+ * @param dst The destination tensor, which specifies the target dimensions for
+ *            padding. dst->op is `GGML_OP_PAD`.
+ */
+void ggml_cann_pad(ggml_backend_cann_context& ctx, ggml_tensor* dst);
+
+/**
+ * @brief   Executes a 2D pooling operation on a ggml tensor using the CANN
+ *          backend.
+ *
+ * @details This function dispatches the execution of a 2D pooling operation on
+ *          the input tensor `dst`. The type of pooling (average or max) is
+ *          determined by the `op` parameter, which is read from the operation
+ *          parameters of `dst`. The function supports average pooling
+ *          (`GGML_OP_POOL_AVG`) and max pooling (`GGML_OP_POOL_MAX`). If an
+ *          invalid operation is encountered, the function asserts a failure.
+ *
+ * @param ctx The CANN context used for operations.
+ * @param dst The destination tensor on which the pooling operation is to be
+ *            performed. dst->op is `GGML_OP_POOL_2D`.
+ */
+void ggml_cann_pool2d(ggml_backend_cann_context& ctx, ggml_tensor* dst);
+
+/**
+ * @brief   Duplicates a ggml tensor using the CANN backend.
+ *
+ * @details This function duplicates the contents of the source tensor `src` to
+ *          the destination tensor `dst`. The function supports various tensor
+ *          types and configurations, including handling of extra data, type
+ *          conversions, and special cases for contiguous and non-contiguous
+ *          tensors.
+ *
+ * @param ctx The CANN context used for operations.
+ * @param dst The destination tensor where the duplicated data will be stored.
+ *            dst->op is `GGML_OP_DUP`
+ *
+ * @attention Only support Fp16/FP32. Not support when src and dst have
+ *            different shape and dst is no-contiguous.
+ * @note:     This func need to simplify.
+ */
+void ggml_cann_dup(ggml_backend_cann_context& ctx, ggml_tensor* dst);
+
+/**
+ * @brief   Computes the Root Mean Square (RMS) normalization of a ggml tensor
+ *          using the CANN backend.
+ *
+ * @details This function applies RMS normalization to the input tensor `src`
+ *          and stores the result in the destination tensor `dst`. RMS
+ *          normalization involves computing the root mean square of the input
+ *          tensor along a specified dimension and then dividing each element of
+ *          the tensor by this value, adjusted by a small epsilon value to
+ *          prevent division by zero.
+ *          The operation is defined as:
+ *          \f[
+ *               \text{RmsNorm}\left(x_i\right)=\frac{x_i}{\text{Rms}(\mathbf{x})} g_i,
+ *               \quad \text { where } \text{Rms}(\mathbf{x})=\sqrt{\frac{1}{n} \sum_{i=1}^n x_i^2+e p s}
+ *          \f]
+ *          `eps` is in dst->op_params.
+ * @param ctx The CANN context used for operations.
+ * @param dst The destination tensor where the normalized values will be stored.
+ *            dst->op is `GGML_OP_RMS_NORM`.
+ */
+void ggml_cann_rms_norm(ggml_backend_cann_context& ctx, ggml_tensor* dst);
+
+/**
+ * @brief   Applies a diagonal mask to the tensor with a specified value.
+ *
+ * @details This function creates a mask tensor filled with ones, then applies
+ *          an upper triangular and lower triangular operation to it based on
+ *          the number of past elements specified. Afterward, it adds the masked
+ *          tensor to the destination tensor in-place.
+ *
+ * @param ctx The backend CANN context used for operations.
+ * @param dst The destination tensor where the result will be stored. dst->op is
+ *            `GGML_OP_DIAG_MASK`
+ * @param value The value to use for masking.
+ */
+void ggml_cann_diag_mask(ggml_backend_cann_context& ctx, ggml_tensor* dst, float value);
+
+/**
+ * @brief   Performs an image-to-column transformation on the input tensor.
+ *
+ * @details This function takes an input tensor and applies an image-to-column
+ *          operation, converting spatial dimensions into column-like
+ *          structures suitable for convolutional operations. It supports both
+ *          half-precision (F16) and single-precision (F32) floating-point data
+ *          types.
+ *
+ * @param ctx The backend CANN context for executing operations.
+ * @param dst The destination tensor that stores the result of the operation.
+ *            dst->op is `GGML_OP_IM2COL`.
+ */
+void ggml_cann_im2col(ggml_backend_cann_context& ctx, ggml_tensor* dst);
+
+/**
+ * @brief   Computes time step embeddings using sine and cosine functions.
+ *
+ * @details This function calculates time step embeddings by applying sine and
+ *          cosine transformations to a given input tensor, which is typically
+ *          used in temporal models like diffusion models or transformers to
+ *          encode time information effectively.
+ *
+ * @param ctx The backend CANN context for executing operations.
+ * @param dst The destination tensor where the result of the embedding operation
+ *            will be stored. dst->op is `GGML_OP_TIMESTEP_EMBEDDING`.
+ */
+void ggml_cann_timestep_embedding(ggml_backend_cann_context& ctx, ggml_tensor* dst);
+
+// @see ggml_cann_dup.
+void ggml_cann_cpy(ggml_backend_cann_context& ctx, ggml_tensor* dst);
+
+/**
+ * @brief   Computes the softmax activation with optional masking.
+ *
+ * @details This function computes the softmax activation over the input tensor,
+ *          optionally applying a mask and scaling factor. It supports both FP16
+ *          and FP32 data types and can handle masking by broadcasting the mask
+ *          across rows if necessary.
+ *          The function performs the following steps:
+ *          1. Multiplies the input tensor by a scale factor.
+ *          2. Optionally casts the mask tensor to FP32 if it is in FP16 format.
+ *          3. Broadcasts the mask tensor if its dimensions do not match the
+ *             input tensor's dimensions.
+ *          4. Adds the mask to the scaled input tensor.
+ *          5. Applies the softmax activation function along the specified
+ *             dimension.
+ *
+ * @param ctx The backend CANN context for executing operations.
+ * @param dst The destination tensor where the result will be stored. dst->op is
+ *            `GGML_OP_SOFTMAX`.
+ */
+void ggml_cann_softmax(ggml_backend_cann_context& ctx, ggml_tensor* dst);
+
+/**
+ * @brief   Extracts specific rows from a tensor based on indices.
+ *
+ * @details This function retrieves rows from a source tensor src0 according to
+ *          the indices provided in another tensor src1 and stores the result in
+ *          a destination tensor (\p dst). It supports different data types
+ *          including F32, F16, Q4_0, and Q8_0.
+ *
+ * @param ctx The backend CANN context for executing operations.
+ * @param dst The destination tensor where the extracted rows will be stored.
+ *            dst->op is `GGML_OP_GET_ROWS`.
+ */
+void ggml_cann_get_rows(ggml_backend_cann_context& ctx, ggml_tensor* dst);
+
+/**
+ * @brief   Executes matrix multiplication for the given tensor.
+ *
+ * @details This function performs matrix multiplication on the source tensors
+ *          associated with the destination tensor. It supports matrix
+ *          multiplication F32, F16, and Q8_0.
+ *
+ * @param ctx The backend CANN context for executing operations.
+ * @param dst The destination tensor for storing the result of the matrix
+ *            multiplication. dst->op is `GGML_OP_MUL_MAT`.
+ */
+void ggml_cann_mul_mat(ggml_backend_cann_context& ctx, ggml_tensor* dst);
+
+/**
+ * @brief Applies Rotary Positional Embedding (RoPE) to the input tensor.
+ *
+ * @details This function implements the RoPE mechanism, which is a method to
+ *          encode positional information into sequence data, particularly
+ *          useful in transformer models. It supports both F32 and F16 data
+ *          types.
+ *
+ * @param ctx The backend CANN context for executing operations.
+ * @param dst The destination tensor where the RoPE-transformed data will be
+ *            stored. dst->op is `GGML_OP_ROPE`.
+ *
+ * @note The function currently does not support cases where the n_dims is less
+ *       than the input tensor's first dimension.
+ * @note The function currently does not support cases where the freq_factors is
+ *       not NULL.
+ * @note The function currently does not support cases where the ext_factor is
+ *       not equal 0.
+ * @note The function currently does not support cases where the freq_scale is
+ *       not equal 1.
+ */
+void ggml_cann_rope(ggml_backend_cann_context& ctx, ggml_tensor* dst);
+
+template <aclnnStatus getWorkspaceSize(const aclTensor*, const aclTensor*,
+                                       aclTensor*, uint64_t*, aclOpExecutor**),
+          aclnnStatus execute(void*, uint64_t, aclOpExecutor*, aclrtStream)>
+void ggml_cann_mul_div(ggml_backend_cann_context& ctx, ggml_tensor* dst) {
+    ggml_tensor* src0 = dst->src[0];
+    ggml_tensor* src1 = dst->src[1];
+    GGML_ASSERT(ggml_can_repeat(src1, src0) && ggml_are_same_shape(src0, dst));
+
+    aclTensor* acl_src0;
+    aclTensor* acl_src1;
+    aclTensor* acl_dst;
+
+    // Need bcast
+    if (!ggml_are_same_shape(src0, src1) && ggml_cann_need_bcast(src0, src1)) {
+        BCAST_SHAPE(src0, src1)
+        acl_src0 = ggml_cann_create_tensor(src0, BCAST_PARAM(src0));
+        acl_src1 = ggml_cann_create_tensor(src1, BCAST_PARAM(src1));
+        acl_dst = ggml_cann_create_tensor(dst, BCAST_PARAM(src0));
+    } else {
+        acl_src0 = ggml_cann_create_tensor(src0);
+        acl_src1 = ggml_cann_create_tensor(src1);
+        acl_dst = ggml_cann_create_tensor(dst);
+    }
+
+    uint64_t workspaceSize = 0;
+    aclOpExecutor* executor;
+    void* workspaceAddr = nullptr;
+
+    ACL_CHECK(getWorkspaceSize(acl_src0, acl_src1, acl_dst, &workspaceSize,
+                               &executor));
+    if (workspaceSize > 0) {
+        ggml_cann_pool_alloc workspace_allocator(ctx.pool(), workspaceSize);
+        workspaceAddr = workspace_allocator.get();
+    }
+
+    aclrtStream main_stream = ctx.stream();
+    ACL_CHECK(execute(workspaceAddr, workspaceSize, executor, main_stream));
+
+    ACL_CHECK(aclDestroyTensor(acl_src0));
+    ACL_CHECK(aclDestroyTensor(acl_src1));
+    ACL_CHECK(aclDestroyTensor(acl_dst));
+}
+
+// Activation functions template.
+template <aclnnStatus getWorkspaceSize(const aclTensor*, aclTensor*, uint64_t*,
+                                       aclOpExecutor**),
+          aclnnStatus execute(void*, uint64_t, aclOpExecutor*,
+                              const aclrtStream)>
+void ggml_cann_activation(ggml_backend_cann_context& ctx, ggml_tensor* dst) {
+    ggml_tensor* src = dst->src[0];
+
+    GGML_ASSERT(src->type == GGML_TYPE_F32);
+    GGML_ASSERT(dst->type == GGML_TYPE_F32);
+
+    aclTensor* acl_src = ggml_cann_create_tensor(src);
+    aclTensor* acl_dst = ggml_cann_create_tensor(dst);
+
+    uint64_t workspaceSize = 0;
+    aclOpExecutor* executor;
+    void* workspaceAddr = nullptr;
+
+    ACL_CHECK(getWorkspaceSize(acl_src, acl_dst, &workspaceSize, &executor));
+    if (workspaceSize > 0) {
+        ggml_cann_pool_alloc workspace_allocator(ctx.pool(), workspaceSize);
+        workspaceAddr = workspace_allocator.get();
+    }
+
+    aclrtStream main_stream = ctx.stream();
+    ACL_CHECK(execute(workspaceAddr, workspaceSize, executor, main_stream));
+
+    ACL_CHECK(aclDestroyTensor(acl_src));
+    ACL_CHECK(aclDestroyTensor(acl_dst));
+}
+
+// Activation functions template for const aclTensors.
+template <aclnnStatus getWorkspaceSize(const aclTensor*, const aclTensor*,
+                                       uint64_t*, aclOpExecutor**),
+          aclnnStatus execute(void*, uint64_t, aclOpExecutor*,
+                              const aclrtStream)>
+void ggml_cann_activation(ggml_backend_cann_context& ctx, ggml_tensor* dst) {
+    ggml_tensor* src = dst->src[0];
+
+    GGML_ASSERT(src->type == GGML_TYPE_F32);
+    GGML_ASSERT(dst->type == GGML_TYPE_F32);
+
+    aclTensor* acl_src = ggml_cann_create_tensor(src);
+    aclTensor* acl_dst = ggml_cann_create_tensor(dst);
+
+    uint64_t workspaceSize = 0;
+    aclOpExecutor* executor;
+    void* workspaceAddr = nullptr;
+
+    ACL_CHECK(getWorkspaceSize(acl_src, acl_dst, &workspaceSize, &executor));
+    if (workspaceSize > 0) {
+        ggml_cann_pool_alloc workspace_allocator(ctx.pool(), workspaceSize);
+        workspaceAddr = workspace_allocator.get();
+    }
+
+    aclrtStream main_stream = ctx.stream();
+    ACL_CHECK(execute(workspaceAddr, workspaceSize, executor, main_stream));
+
+    ACL_CHECK(aclDestroyTensor(acl_src));
+    ACL_CHECK(aclDestroyTensor(acl_dst));
+}
+
+#endif  // CANN_ACLNN_OPS
diff --git a/src/whisper_cpp/ggml/src/ggml-cann/common.h b/src/whisper_cpp/ggml/src/ggml-cann/common.h
new file mode 100644
index 00000000..edfa4961
--- /dev/null
+++ b/src/whisper_cpp/ggml/src/ggml-cann/common.h
@@ -0,0 +1,283 @@
+/*
+ * Copyright (c) 2023-2024 The ggml authors
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a copy
+ * of this software and associated documentation files (the "Software"), to
+ * deal in the Software without restriction, including without limitation the
+ * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or
+ * sell copies of the Software, and to permit persons to whom the Software is
+ * furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice shall be included in
+ * all copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+ * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
+ * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
+ * IN THE SOFTWARE.
+ */
+
+#ifndef CANN_COMMON_H
+#define CANN_COMMON_H
+
+#include <acl/acl.h>
+
+#include <cstdio>
+#include <iostream>
+#include <map>
+#include <memory>
+#include <string>
+#include <vector>
+
+#include "../include/ggml-cann.h"
+#include "../include/ggml.h"
+
+#define MATRIX_ROW_PADDING 512
+#define GGML_CANN_MAX_STREAMS 8
+
+/**
+ * @brief Handles CANN-related errors by printing an error message and
+ *        terminating the program.
+ * @param stmt The statement that caused the error.
+ * @param func The function in which the error occurred.
+ * @param file The file in which the error occurred.
+ * @param line The line number at which the error occurred.
+ * @param msg The error message.
+ */
+[[noreturn]] void ggml_cann_error(const char* stmt, const char* func,
+                                  const char* file, int line, const char* msg);
+
+/**
+ * @brief Checks the result of a CANN function call and invokes the error
+ *        handler if the call fails.
+ * @param stmt The CANN function call to check.
+ * @param success The success code that indicates the call was successful.
+ * @param error_fn The function to call to retrieve the error message.
+ */
+#define ACL_CHECK_GEN(stmt, success, error_fn)                                \
+    do {                                                                      \
+        int err_code = (stmt);                                                \
+        if (err_code != (success)) {                                          \
+            ggml_cann_error(#stmt, __func__, __FILE__, __LINE__, error_fn()); \
+        }                                                                     \
+    } while (0);
+
+#define ACL_CHECK(stmt) ACL_CHECK_GEN(stmt, 0, aclGetRecentErrMsg)
+
+/**
+ * @brief Contains information about CANN devices.
+ */
+struct ggml_cann_device_info {
+    /**
+     * @brief Number of CANN devices available.
+     */
+    int32_t device_count;
+
+    /**
+     * @brief Information about a single CANN device.
+     */
+    struct cann_device_info {
+        int cc;                 /**< Compute capability.                   */
+        size_t smpb;            /**< Maximum shared memory per block.      */
+        bool vmm;               /**< Virtual memory support.               */
+        size_t vmm_granularity; /**< Granularity of virtual memory.        */
+        size_t total_vram;      /**< Total video RAM available on the device. */
+    };
+
+    cann_device_info devices[GGML_CANN_MAX_DEVICES] =
+        {}; /**< Array of CANN device information. */
+};
+
+const ggml_cann_device_info& ggml_cann_info();
+
+void ggml_cann_set_device(int32_t device);
+int32_t ggml_cann_get_device();
+
+/**
+ * @brief Abstract base class for memory pools used by CANN.
+ */
+struct ggml_cann_pool {
+    /**
+     * @brief Virtual destructor for the memory pool.
+     */
+    virtual ~ggml_cann_pool() = default;
+
+    /**
+     * @brief Allocates memory from the pool.
+     *
+     * @param size         The size of the memory block to allocate.
+     * @param actual_size  Pointer to a variable where the actual allocated size
+     *                     will be stored.
+     * @return             Pointer to the allocated memory block.
+     */
+    virtual void* alloc(size_t size, size_t* actual_size) = 0;
+
+    /**
+     * @brief Frees a previously allocated memory block.
+     *
+     * @param ptr   Pointer to the memory block to free.
+     * @param size  Size of the memory block to free.
+     * @note Note that all CANN opertors are running async. Make sure memory is
+     *       still avaiable before this operator finished.
+     */
+    virtual void free(void* ptr, size_t size) = 0;
+};
+
+/**
+ * @brief RAII wrapper for managing memory allocations from a CANN memory pool.
+ */
+struct ggml_cann_pool_alloc {
+    ggml_cann_pool* pool = nullptr; /**< Pointer to the memory pool. */
+    void* ptr = nullptr;    /**< Pointer to the allocated memory block. */
+    size_t actual_size = 0; /**< Actual size of the allocated memory block. */
+
+    /**
+     * @brief Default constructor.
+     */
+    ggml_cann_pool_alloc() = default;
+
+    /**
+     * @brief Constructor that initializes the memory pool.
+     * @param pool Reference to the memory pool.
+     */
+    explicit ggml_cann_pool_alloc(ggml_cann_pool& pool) : pool(&pool) {}
+
+    /**
+     * @brief Constructor that initializes the memory pool and allocates memory.
+     * @param pool Reference to the memory pool.
+     * @param size Size of the memory block to allocate.
+     */
+    ggml_cann_pool_alloc(ggml_cann_pool& pool, size_t size) : pool(&pool) {
+        alloc(size);
+    }
+
+    /**
+     * @brief Destructor that frees the allocated memory block.
+     */
+    ~ggml_cann_pool_alloc() {
+        if (ptr != nullptr) {
+            pool->free(ptr, actual_size);
+        }
+    }
+
+    /**
+     * @brief Allocates memory from the pool.
+     * @param size Size of the memory block to allocate.
+     * @return Pointer to the allocated memory block.
+     */
+    void* alloc(size_t size) {
+        GGML_ASSERT(pool != nullptr);
+        GGML_ASSERT(ptr == nullptr);
+        ptr = pool->alloc(size, &this->actual_size);
+        return ptr;
+    }
+
+    /**
+     * @brief Allocates memory from a specific memory pool.
+     * @param pool Reference to the memory pool.
+     * @param size Size of the memory block to allocate.
+     * @return Pointer to the allocated memory block.
+     */
+    void* alloc(ggml_cann_pool& pool, size_t size) {
+        this->pool = &pool;
+        return alloc(size);
+    }
+
+    /**
+     * @brief Gets the pointer to the allocated memory block.
+     * @return Pointer to the allocated memory block.
+     */
+    void* get() { return ptr; }
+
+    // Deleted copy constructor
+    ggml_cann_pool_alloc(const ggml_cann_pool_alloc&) = delete;
+
+    // Deleted move constructor
+    ggml_cann_pool_alloc(ggml_cann_pool_alloc&&) = delete;
+
+    // Deleted copy assignment operator
+    ggml_cann_pool_alloc& operator=(const ggml_cann_pool_alloc&) = delete;
+
+    // Deleted move assignment operator
+    ggml_cann_pool_alloc& operator=(ggml_cann_pool_alloc&&) = delete;
+};
+
+/**
+ * @brief Context for managing CANN backend operations.
+ */
+struct ggml_backend_cann_context {
+    int32_t device;                  /**< Device ID. */
+    std::string name;                /**< Name of the device. */
+    aclrtEvent copy_event = nullptr; /**< Event for managing copy operations. */
+
+    aclrtStream streams[GGML_CANN_MAX_STREAMS] = {
+        {nullptr}}; /**< Array of streams for the device. */
+
+    /**
+     * @brief Constructor for initializing the context with a given device.
+     * @param device Device ID.
+     */
+    explicit ggml_backend_cann_context(int device)
+        : device(device), name("CANN" + std::to_string(device)) {}
+
+    /**
+     * @brief Destructor for cleaning up resources.
+     */
+    ~ggml_backend_cann_context() {
+        ggml_cann_set_device(device);
+        if (copy_event != nullptr) {
+            ACL_CHECK(aclrtDestroyEvent(copy_event));
+        }
+        for (int i = 0; i < GGML_CANN_MAX_STREAMS; ++i) {
+            if (streams[i] != nullptr) {
+                ACL_CHECK(aclrtDestroyStream(streams[i]));
+            }
+        }
+    }
+
+    /**
+     * @brief Get or create a stream for a given index.
+     * @param stream Index of the stream.
+     * @return The stream corresponding to the given index.
+     */
+    aclrtStream stream(int stream) {
+        if (streams[stream] == nullptr) {
+            ggml_cann_set_device(device);
+            ACL_CHECK(aclrtCreateStream(&streams[stream]));
+        }
+        return streams[stream];
+    }
+
+    /**
+     * @brief Get or create the default stream (index 0).
+     * @return The default stream.
+     */
+    aclrtStream stream() { return stream(0); }
+
+    // TODO: each stream should have a memory pool.
+    std::unique_ptr<ggml_cann_pool>
+        mem_pool; /**< Memory pool for the device. */
+
+    /**
+     * @brief Create a new memory pool for a given device.
+     * @param device Device ID.
+     * @return A unique pointer to the new memory pool.
+     */
+    static std::unique_ptr<ggml_cann_pool> new_pool_for_device(int device);
+
+    /**
+     * @brief Get or create the memory pool for the context.
+     * @return Reference to the memory pool.
+     */
+    ggml_cann_pool& pool() {
+        if (mem_pool == nullptr) {
+            mem_pool = new_pool_for_device(device);
+        }
+        return *mem_pool;
+    }
+};
+
+#endif  // CANN_COMMON_H
diff --git a/src/whisper_cpp/ggml/src/ggml-cann/kernels/CMakeLists.txt b/src/whisper_cpp/ggml/src/ggml-cann/kernels/CMakeLists.txt
new file mode 100644
index 00000000..5b4fef91
--- /dev/null
+++ b/src/whisper_cpp/ggml/src/ggml-cann/kernels/CMakeLists.txt
@@ -0,0 +1,33 @@
+if (NOT SOC_TYPE)
+    set (SOC_TYPE "Ascend910B3")
+endif()
+
+file(GLOB SRC_FILES
+    get_row_f32.cpp
+    get_row_f16.cpp
+    get_row_q4_0.cpp
+    get_row_q8_0.cpp
+    quantize_f32_q8_0.cpp
+    quantize_f16_q8_0.cpp
+    quantize_float_to_q4_0.cpp
+    dup.cpp
+)
+
+string(TOLOWER ${SOC_TYPE} SOC_VERSION)
+set(ASCEND_CANN_PACKAGE_PATH ${CANN_INSTALL_DIR})
+set(RUN_MODE "npu" CACHE STRING "run mode: npu/sim")
+
+if(EXISTS ${ASCEND_CANN_PACKAGE_PATH}/compiler/tikcpp/ascendc_kernel_cmake)
+    set(ASCENDC_CMAKE_DIR ${ASCEND_CANN_PACKAGE_PATH}/compiler/tikcpp/ascendc_kernel_cmake)
+elseif(EXISTS ${ASCEND_CANN_PACKAGE_PATH}/ascendc_devkit/tikcpp/samples/cmake)
+    set(ASCENDC_CMAKE_DIR ${ASCEND_CANN_PACKAGE_PATH}/ascendc_devkit/tikcpp/samples/cmake)
+else()
+    message(FATAL_ERROR "ascendc_kernel_cmake does not exist, please check whether the compiler package is installed.")
+endif()
+include(${ASCENDC_CMAKE_DIR}/ascendc.cmake)
+
+ascendc_library(ascendc_kernels STATIC
+    ${SRC_FILES}
+)
+
+# ascendc_compile_definitions(ascendc_kernels PRIVATE -DASCENDC_DUMP)
diff --git a/src/whisper_cpp/ggml/src/ggml-cann/kernels/ascendc_kernels.h b/src/whisper_cpp/ggml/src/ggml-cann/kernels/ascendc_kernels.h
new file mode 100644
index 00000000..7e153208
--- /dev/null
+++ b/src/whisper_cpp/ggml/src/ggml-cann/kernels/ascendc_kernels.h
@@ -0,0 +1,19 @@
+#ifndef ASCENDC_KERNELS_H
+#define ASCENDC_KERNELS_H
+
+#include "aclrtlaunch_ascendc_get_row_f32.h"
+#include "aclrtlaunch_ascendc_get_row_f16.h"
+#include "aclrtlaunch_ascendc_get_row_q8_0.h"
+#include "aclrtlaunch_ascendc_get_row_q4_0.h"
+
+#include "aclrtlaunch_ascendc_quantize_f32_q8_0.h"
+#include "aclrtlaunch_ascendc_quantize_f16_q8_0.h"
+#include "aclrtlaunch_ascendc_quantize_f16_to_q4_0.h"
+#include "aclrtlaunch_ascendc_quantize_f32_to_q4_0.h"
+
+#include "aclrtlaunch_ascendc_dup_by_rows_fp16.h"
+#include "aclrtlaunch_ascendc_dup_by_rows_fp32.h"
+#include "aclrtlaunch_ascendc_dup_by_rows_fp32_to_fp16.h"
+#include "aclrtlaunch_ascendc_dup_by_rows_fp16_to_fp32.h"
+
+#endif  // ASCENDC_KERNELS_H
diff --git a/src/whisper_cpp/ggml/src/ggml-cann/kernels/dup.cpp b/src/whisper_cpp/ggml/src/ggml-cann/kernels/dup.cpp
new file mode 100644
index 00000000..e2c65115
--- /dev/null
+++ b/src/whisper_cpp/ggml/src/ggml-cann/kernels/dup.cpp
@@ -0,0 +1,223 @@
+#include "kernel_operator.h"
+
+#include <cmath>
+
+using namespace AscendC;
+
+#define BUFFER_NUM 2
+
+template <typename SRC_T, typename DST_T>
+class DupByRows {
+   public:
+    __aicore__ inline DupByRows() {}
+    __aicore__ inline void init(GM_ADDR src, GM_ADDR dst, int64_t *input_ne_ub,
+                                size_t *input_nb_ub) {
+        /* Dup by rows when src is contigous on first dimension and dst is
+        contiguous, each kernel process one row.
+        */
+
+        // Input has four dims.
+        int64_t op_block_num = GetBlockNum();
+        int64_t op_block_idx = GetBlockIdx();
+
+        // param
+        num_rows = input_ne_ub[1] * input_ne_ub[2] * input_ne_ub[3];
+        num_elem = input_ne_ub[0];
+
+        // index for (ne[1], ne[2], ne[3]): (idx_ne1, idx_ne2, idx_ne3)
+        idx_ne3 = op_block_idx / (input_ne_ub[1] * input_ne_ub[2]);
+        idx_ne2 = (op_block_idx - idx_ne3 * (input_ne_ub[1] * input_ne_ub[2]))
+                  / (input_ne_ub[1]);
+        idx_ne1 = op_block_idx - idx_ne3 * (input_ne_ub[1] * input_ne_ub[2])
+                - idx_ne2 * input_ne_ub[1];
+
+        // src may not contiguous in dim [1,2,3], so stride decited by ne&nb
+        src_stride = input_nb_ub[3] * idx_ne3 + input_nb_ub[2] * idx_ne2
+                     + input_nb_ub[1] * idx_ne1;
+
+        // dst is contiguous
+        dst_stride = op_block_idx * (input_ne_ub[0] * sizeof(DST_T));
+
+        src_gm.SetGlobalBuffer(reinterpret_cast<__gm__ SRC_T *>(src +
+                                                                src_stride));
+        dst_gm.SetGlobalBuffer(reinterpret_cast<__gm__ DST_T *>(dst +
+                                                                dst_stride));
+
+        pipe.InitBuffer(src_queue, BUFFER_NUM, (sizeof(SRC_T) * num_elem +
+                                                32 - 1) / 32 * 32);
+        pipe.InitBuffer(dst_queue, BUFFER_NUM, (sizeof(DST_T) * num_elem +
+                                                32 - 1) / 32 * 32);
+    }
+
+    __aicore__ inline void copy_in() {
+        LocalTensor<SRC_T> src_local = src_queue.AllocTensor<SRC_T>();
+
+        DataCopyExtParams dataCopyParams;
+        dataCopyParams.blockCount = 1;
+        dataCopyParams.blockLen = num_elem * sizeof(SRC_T);
+        DataCopyPadExtParams<SRC_T> padParams;
+        DataCopyPad(src_local, src_gm, dataCopyParams, padParams);
+
+        src_queue.EnQue(src_local);
+    }
+
+    __aicore__ inline void copy_out() {
+        LocalTensor<DST_T> dst_local = dst_queue.DeQue<DST_T>();
+
+        DataCopyExtParams dataCopyParams;
+        dataCopyParams.blockCount = 1;
+        dataCopyParams.blockLen = num_elem * sizeof(DST_T);
+        DataCopyPad(dst_gm, dst_local, dataCopyParams);
+
+        dst_queue.FreeTensor(dst_local);
+    }
+
+    __aicore__ inline void dup() {
+        // main process, copy one row data from src to dst.
+        copy_in();
+
+        LocalTensor<SRC_T> src_local = src_queue.DeQue<SRC_T>();
+        LocalTensor<DST_T> dst_local = dst_queue.AllocTensor<DST_T>();
+
+        int32_t BLOCK_NUM = 32 / sizeof(DST_T);
+        DataCopy(dst_local, src_local, (num_elem + BLOCK_NUM - 1)
+                                        / BLOCK_NUM * BLOCK_NUM);
+        dst_queue.EnQue<DST_T>(dst_local);
+
+        src_queue.FreeTensor(src_local);
+        copy_out();
+    }
+
+    __aicore__ inline void dup_with_cast() {
+        // main process, copy one row data from src to dst.
+        // cast dtype from src to dst.
+        copy_in();
+
+        LocalTensor<SRC_T> src_local = src_queue.DeQue<SRC_T>();
+        LocalTensor<DST_T> dst_local = dst_queue.AllocTensor<DST_T>();
+
+        Cast(dst_local, src_local, RoundMode::CAST_NONE, num_elem);
+        dst_queue.EnQue<DST_T>(dst_local);
+
+        src_queue.FreeTensor(src_local);
+        copy_out();
+    }
+
+   private:
+
+    TPipe pipe;
+    GlobalTensor<SRC_T> src_gm;
+    GlobalTensor<DST_T> dst_gm;
+
+    int64_t num_rows;
+    int64_t num_elem;
+    int64_t idx_ne3;
+    int64_t idx_ne2;
+    int64_t idx_ne1;
+    int64_t src_stride;
+    int64_t dst_stride;
+
+    TQue<QuePosition::VECIN, BUFFER_NUM> src_queue;
+    TQue<QuePosition::VECOUT, BUFFER_NUM> dst_queue;
+};
+
+template <typename T>
+__aicore__ inline void copy_to_ub(GM_ADDR gm, T *ub, size_t size) {
+    auto gm_ptr = (__gm__ uint8_t *)gm;
+    auto ub_ptr = (uint8_t *)(ub);
+    for (int32_t i = 0; i < size; ++i, ++ub_ptr, ++gm_ptr) {
+        *ub_ptr = *gm_ptr;
+    }
+}
+
+extern "C" __global__ __aicore__ void ascendc_dup_by_rows_fp16(
+                                                        GM_ADDR src_gm,
+                                                        GM_ADDR dst_gm,
+                                                        GM_ADDR input_ne_gm,
+                                                        GM_ADDR input_nb_gm,
+                                                        GM_ADDR output_ne_gm,
+                                                        GM_ADDR output_nb_gm) {
+
+    int64_t input_ne_ub[4];
+    size_t input_nb_ub[4];
+    int64_t output_ne_ub[4];
+    size_t output_nb_ub[4];
+
+    copy_to_ub(input_ne_gm, input_ne_ub, 32);
+    copy_to_ub(input_nb_gm, input_nb_ub, 32);
+    copy_to_ub(output_ne_gm, output_ne_ub, 32);
+    copy_to_ub(output_nb_gm, output_nb_ub, 32);
+
+    DupByRows<half, half> op;
+    op.init(src_gm, dst_gm, input_ne_ub, input_nb_ub);
+    op.dup();
+}
+
+extern "C" __global__ __aicore__ void ascendc_dup_by_rows_fp32(
+                                                        GM_ADDR src_gm,
+                                                        GM_ADDR dst_gm,
+                                                        GM_ADDR input_ne_gm,
+                                                        GM_ADDR input_nb_gm,
+                                                        GM_ADDR output_ne_gm,
+                                                        GM_ADDR output_nb_gm) {
+    int64_t input_ne_ub[4];
+    size_t input_nb_ub[4];
+    int64_t output_ne_ub[4];
+    size_t output_nb_ub[4];
+
+    copy_to_ub(input_ne_gm, input_ne_ub, 32);
+    copy_to_ub(input_nb_gm, input_nb_ub, 32);
+    copy_to_ub(output_ne_gm, output_ne_ub, 32);
+    copy_to_ub(output_nb_gm, output_nb_ub, 32);
+
+    DupByRows<float_t, float_t> op;
+    op.init(src_gm, dst_gm, input_ne_ub, input_nb_ub);
+    op.dup();
+}
+
+extern "C" __global__ __aicore__ void ascendc_dup_by_rows_fp32_to_fp16(
+                                                        GM_ADDR src_gm,
+                                                        GM_ADDR dst_gm,
+                                                        GM_ADDR input_ne_gm,
+                                                        GM_ADDR input_nb_gm,
+                                                        GM_ADDR output_ne_gm,
+                                                        GM_ADDR output_nb_gm) {
+
+    int64_t input_ne_ub[4];
+    size_t input_nb_ub[4];
+    int64_t output_ne_ub[4];
+    size_t output_nb_ub[4];
+
+    copy_to_ub(input_ne_gm, input_ne_ub, 32);
+    copy_to_ub(input_nb_gm, input_nb_ub, 32);
+    copy_to_ub(output_ne_gm, output_ne_ub, 32);
+    copy_to_ub(output_nb_gm, output_nb_ub, 32);
+
+    DupByRows<float_t, half> op;
+    op.init(src_gm, dst_gm, input_ne_ub, input_nb_ub);
+    op.dup_with_cast();
+}
+
+extern "C" __global__ __aicore__ void ascendc_dup_by_rows_fp16_to_fp32(
+                                                        GM_ADDR src_gm,
+                                                        GM_ADDR dst_gm,
+                                                        GM_ADDR input_ne_gm,
+                                                        GM_ADDR input_nb_gm,
+                                                        GM_ADDR output_ne_gm,
+                                                        GM_ADDR output_nb_gm) {
+
+    // copy params from gm to ub.
+    int64_t input_ne_ub[4];
+    size_t input_nb_ub[4];
+    int64_t output_ne_ub[4];
+    size_t output_nb_ub[4];
+
+    copy_to_ub(input_ne_gm, input_ne_ub, 32);
+    copy_to_ub(input_nb_gm, input_nb_ub, 32);
+    copy_to_ub(output_ne_gm, output_ne_ub, 32);
+    copy_to_ub(output_nb_gm, output_nb_ub, 32);
+
+    DupByRows<half, float_t> op;
+    op.init(src_gm, dst_gm, input_ne_ub, input_nb_ub);
+    op.dup_with_cast();
+}
diff --git a/src/whisper_cpp/ggml/src/ggml-cann/kernels/get_row_f16.cpp b/src/whisper_cpp/ggml/src/ggml-cann/kernels/get_row_f16.cpp
new file mode 100644
index 00000000..c704b5b2
--- /dev/null
+++ b/src/whisper_cpp/ggml/src/ggml-cann/kernels/get_row_f16.cpp
@@ -0,0 +1,186 @@
+#include "kernel_operator.h"
+
+// optimize me. Use template to avoid copy code.
+using namespace AscendC;
+
+#define BUFFER_NUM 2
+
+class GET_ROW_F16 {
+   public:
+    __aicore__ inline GET_ROW_F16() {}
+    __aicore__ inline void init(GM_ADDR input, GM_ADDR indices, GM_ADDR output,
+                                int64_t *input_ne_ub, size_t *input_nb_ub,
+                                int64_t *indices_ne_ub, size_t *indices_nb_ub,
+                                int64_t *output_ne_ub, size_t *output_nb_ub) {
+        // TODO, use template for F16/f32
+        int64_t op_block_num = GetBlockNum();
+        int64_t op_block_idx = GetBlockIdx();
+
+        for (int i = 0; i < 4; i++) {
+            input_ne[i] = input_ne_ub[i];
+            input_stride[i] = input_nb_ub[i] / input_nb_ub[0];
+
+            indices_ne[i] = indices_ne_ub[i];
+            indices_stride[i] = indices_nb_ub[i] / indices_nb_ub[0];
+
+            output_ne[i] = output_ne_ub[i];
+            output_stride[i] = output_nb_ub[i] / output_nb_ub[0];
+        }
+
+        // Indices has two dims. n_elements = all rows should get.
+        // dr, all rows should this thread get.
+        uint64_t n_elements =
+            indices_ne[0] * indices_ne[1] * indices_ne[2] * indices_ne[3];
+        dr = n_elements / op_block_num;
+
+        uint64_t tails = n_elements % op_block_num;
+        if (op_block_idx < tails) {
+            dr += 1;
+            ir = dr * op_block_idx;
+        } else {
+            ir = dr * op_block_idx + tails;
+        }
+
+        input_gm.SetGlobalBuffer((__gm__ half *)input);
+        indices_gm.SetGlobalBuffer((__gm__ int32_t *)indices);
+        output_gm.SetGlobalBuffer((__gm__ float *)output);
+
+        uint64_t input_local_buffer_size = ((input_ne[0] * sizeof(half) + 31)
+                                             & ~31);
+        uint64_t output_local_buffer_size = ((input_ne[0] * sizeof(float) + 31)
+                                              & ~31);
+
+        local_buffer_elems = input_local_buffer_size / sizeof(half);
+
+        // TODO, consider long row that can't put in UB.
+        // All data should asign to 32. It's ok because all data is align to 32.
+        pipe.InitBuffer(input_queue, BUFFER_NUM, input_local_buffer_size);
+        pipe.InitBuffer(output_queue, BUFFER_NUM, output_local_buffer_size);
+    }
+
+    __aicore__ inline void copy_in(uint32_t offset, size_t len) {
+        LocalTensor<half> input_local = input_queue.AllocTensor<half>();
+        size_t tail = len % 32;
+        len = len & ~31;
+        DataCopy(input_local, input_gm[offset], len);
+        if(tail != 0) {
+            DataCopyExtParams dataCopyParams;
+            dataCopyParams.blockCount = 1;
+            dataCopyParams.blockLen = tail * sizeof(half);
+            DataCopyPadExtParams<half> padParams;
+            DataCopyPad(input_local[len], input_gm[offset + len],
+                        dataCopyParams, padParams);
+        }
+        input_queue.EnQue(input_local);
+    }
+
+    __aicore__ inline void copy_out(uint32_t offset, size_t len) {
+        LocalTensor<float> output_local = output_queue.DeQue<float>();
+        size_t tail = len % 32;
+        len = len & ~31;
+        DataCopy(output_gm[offset], output_local, len);
+        if(tail != 0) {
+            DataCopyExtParams dataCopyParams;
+            dataCopyParams.blockCount = 1;
+            dataCopyParams.blockLen = tail * sizeof(float);
+            DataCopyPad(output_gm[offset + len], output_local[len],
+                        dataCopyParams);
+        }
+        output_queue.FreeTensor(output_local);
+    }
+
+    __aicore__ inline void calculate_row(int64_t idx) {
+        const int64_t indices_ne2_idx = idx / (indices_ne[0] * indices_ne[1]);
+        const int64_t indices_ne1_idx =
+            (idx - indices_ne2_idx * indices_ne[0] * indices_ne[1]) /
+            indices_ne[0];
+        const int64_t indices_ne0_idx =
+            (idx - indices_ne2_idx * indices_ne[0] * indices_ne[1] -
+             indices_ne1_idx * indices_ne[0]);
+
+        const int64_t indices_offset = indices_ne0_idx * indices_stride[0] +
+                                       indices_ne1_idx * indices_stride[1] +
+                                       indices_ne2_idx * indices_stride[2];
+        const int32_t selected_row_idx = indices_gm.GetValue(indices_offset);
+
+        const int64_t input_offset = selected_row_idx * input_stride[1] +
+                                     indices_ne1_idx * input_stride[2] +
+                                     indices_ne2_idx * input_stride[3];
+
+        const int64_t output_offset = indices_ne0_idx * output_stride[1] +
+                                      indices_ne1_idx * output_stride[2] +
+                                      indices_ne2_idx * output_stride[3];
+
+        copy_in(input_offset, input_ne[0]);
+        LocalTensor<half> input_local = input_queue.DeQue<half>();
+        LocalTensor<float> output_local = output_queue.AllocTensor<float>();
+
+        Cast(output_local, input_local, RoundMode::CAST_NONE,
+             local_buffer_elems);
+        output_queue.EnQue(output_local);
+        copy_out(output_offset, input_ne[0]);
+
+        input_queue.FreeTensor(input_local);
+    }
+
+    __aicore__ inline void calculate() {
+        for (int64_t i = ir; i < ir + dr; i++) {
+            calculate_row(i);
+        }
+    }
+
+   private:
+    int64_t input_ne[4];
+    size_t input_stride[4];
+
+    int64_t indices_ne[4];
+    size_t indices_stride[4];
+
+    int64_t output_ne[4];
+    size_t output_stride[4];
+
+    size_t local_buffer_elems;
+
+    int64_t ir;
+    int64_t dr;
+
+    TPipe pipe;
+    GlobalTensor<half> input_gm;
+    GlobalTensor<int32_t> indices_gm;
+    GlobalTensor<float> output_gm;
+    TQue<QuePosition::VECIN, BUFFER_NUM> input_queue;
+    TQue<QuePosition::VECOUT, BUFFER_NUM> output_queue;
+};
+
+template <typename T>
+__aicore__ inline void copy_to_ub(GM_ADDR gm, T *ub, size_t size) {
+    auto gm_ptr = (__gm__ uint8_t *)gm;
+    auto ub_ptr = (uint8_t *)(ub);
+    for (int32_t i = 0; i < size; ++i, ++ub_ptr, ++gm_ptr) {
+        *ub_ptr = *gm_ptr;
+    }
+}
+
+extern "C" __global__ __aicore__ void ascendc_get_row_f16(
+    GM_ADDR input_gm, GM_ADDR indices_gm, GM_ADDR output_gm,
+    GM_ADDR input_ne_gm, GM_ADDR input_nb_gm, GM_ADDR indices_ne_gm,
+    GM_ADDR indices_nb_gm, GM_ADDR output_ne_gm, GM_ADDR output_nb_gm) {
+    int64_t input_ne_ub[4];
+    size_t input_nb_ub[4];
+    int64_t indices_ne_ub[4];
+    size_t indices_nb_ub[4];
+    int64_t output_ne_ub[4];
+    size_t output_nb_ub[4];
+
+    copy_to_ub(input_ne_gm, input_ne_ub, 32);
+    copy_to_ub(input_nb_gm, input_nb_ub, 32);
+    copy_to_ub(indices_ne_gm, indices_ne_ub, 32);
+    copy_to_ub(indices_nb_gm, indices_nb_ub, 32);
+    copy_to_ub(output_ne_gm, output_ne_ub, 32);
+    copy_to_ub(output_nb_gm, output_nb_ub, 32);
+
+    GET_ROW_F16 op;
+    op.init(input_gm, indices_gm, output_gm, input_ne_ub, input_nb_ub,
+            indices_ne_ub, indices_nb_ub, output_ne_ub, output_nb_ub);
+    op.calculate();
+}
diff --git a/src/whisper_cpp/ggml/src/ggml-cann/kernels/get_row_f32.cpp b/src/whisper_cpp/ggml/src/ggml-cann/kernels/get_row_f32.cpp
new file mode 100644
index 00000000..9db080af
--- /dev/null
+++ b/src/whisper_cpp/ggml/src/ggml-cann/kernels/get_row_f32.cpp
@@ -0,0 +1,180 @@
+#include "kernel_operator.h"
+
+// optimize me. Use template to avoid copy code.
+using namespace AscendC;
+
+#define BUFFER_NUM 2
+
+class GET_ROW_F32 {
+   public:
+    __aicore__ inline GET_ROW_F32() {}
+    __aicore__ inline void init(GM_ADDR input, GM_ADDR indices, GM_ADDR output,
+                                int64_t *input_ne_ub, size_t *input_nb_ub,
+                                int64_t *indices_ne_ub, size_t *indices_nb_ub,
+                                int64_t *output_ne_ub, size_t *output_nb_ub) {
+        int64_t op_block_num = GetBlockNum();
+        int64_t op_block_idx = GetBlockIdx();
+
+        for (int i = 0; i < 4; i++) {
+            input_ne[i] = input_ne_ub[i];
+            input_stride[i] = input_nb_ub[i] / input_nb_ub[0];
+
+            indices_ne[i] = indices_ne_ub[i];
+            indices_stride[i] = indices_nb_ub[i] / indices_nb_ub[0];
+
+            output_ne[i] = output_ne_ub[i];
+            output_stride[i] = output_nb_ub[i] / output_nb_ub[0];
+        }
+
+        // Indices has two dims. n_elements = all rows should get.
+        // dr, all rows should this thread get.
+        uint64_t n_elements =
+            indices_ne[0] * indices_ne[1] * indices_ne[2] * indices_ne[3];
+        dr = n_elements / op_block_num;
+
+        uint64_t tails = n_elements % op_block_num;
+        if (op_block_idx < tails) {
+            dr += 1;
+            ir = dr * op_block_idx;
+        } else {
+            ir = dr * op_block_idx + tails;
+        }
+
+        input_gm.SetGlobalBuffer((__gm__ float *)input);
+        indices_gm.SetGlobalBuffer((__gm__ int32_t *)indices);
+        output_gm.SetGlobalBuffer((__gm__ float *)output);
+
+        uint64_t local_buffer_size = ((input_ne[0] * sizeof(float) + 31) & ~31);
+        local_buffer_elems = local_buffer_size / sizeof(float);
+
+        // TODO, consider long row that can't put in UB.
+        // All data should asign to 32. It's ok because all data is align to 32.
+        pipe.InitBuffer(input_queue, BUFFER_NUM, local_buffer_size);
+        pipe.InitBuffer(output_queue, BUFFER_NUM, local_buffer_size);
+    }
+
+    __aicore__ inline void copy_in(uint32_t offset, size_t len) {
+        LocalTensor<float> input_local = input_queue.AllocTensor<float>();
+        size_t tail = len % 32;
+        len = len & ~31;
+        DataCopy(input_local, input_gm[offset], len);
+        if(tail != 0) {
+            DataCopyExtParams dataCopyParams;
+            dataCopyParams.blockCount = 1;
+            dataCopyParams.blockLen = tail * sizeof(float);
+            DataCopyPadExtParams<float> padParams;
+            DataCopyPad(input_local[len], input_gm[offset + len],
+                        dataCopyParams, padParams);
+        }
+        input_queue.EnQue(input_local);
+    }
+
+    __aicore__ inline void copy_out(uint32_t offset, size_t len) {
+        LocalTensor<float> output_local = output_queue.DeQue<float>();
+        size_t tail = len % 32;
+        len = len & ~31;
+        DataCopy(output_gm[offset], output_local, len);
+        if(tail != 0) {
+            DataCopyExtParams dataCopyParams;
+            dataCopyParams.blockCount = 1;
+            dataCopyParams.blockLen = tail * sizeof(float);
+            DataCopyPad(output_gm[offset + len], output_local[len],
+                        dataCopyParams);
+        }
+        output_queue.FreeTensor(output_local);
+    }
+
+    __aicore__ inline void calculate_row(int64_t idx) {
+        const int64_t indices_ne2_idx = idx / (indices_ne[0] * indices_ne[1]);
+        const int64_t indices_ne1_idx =
+            (idx - indices_ne2_idx * indices_ne[0] * indices_ne[1]) /
+            indices_ne[0];
+        const int64_t indices_ne0_idx =
+            (idx - indices_ne2_idx * indices_ne[0] * indices_ne[1] -
+             indices_ne1_idx * indices_ne[0]);
+
+        const int64_t indices_offset = indices_ne0_idx * indices_stride[0] +
+                                       indices_ne1_idx * indices_stride[1] +
+                                       indices_ne2_idx * indices_stride[2];
+        const int32_t selected_row_idx = indices_gm.GetValue(indices_offset);
+
+        const int64_t input_offset = selected_row_idx * input_stride[1] +
+                                     indices_ne1_idx * input_stride[2] +
+                                     indices_ne2_idx * input_stride[3];
+
+        const int64_t output_offset = indices_ne0_idx * output_stride[1] +
+                                      indices_ne1_idx * output_stride[2] +
+                                      indices_ne2_idx * output_stride[3];
+
+        copy_in(input_offset, input_ne[0]);
+        LocalTensor<float> input_local = input_queue.DeQue<float>();
+        LocalTensor<float> output_local = output_queue.AllocTensor<float>();
+
+        DataCopy(output_local, input_local, local_buffer_elems);
+        output_queue.EnQue(output_local);
+        copy_out(output_offset, input_ne[0]);
+
+        input_queue.FreeTensor(input_local);
+    }
+
+    __aicore__ inline void calculate() {
+        for (int64_t i = ir; i < ir + dr; i++) {
+            calculate_row(i);
+        }
+    }
+
+   private:
+    int64_t input_ne[4];
+    size_t input_stride[4];
+
+    int64_t indices_ne[4];
+    size_t indices_stride[4];
+
+    int64_t output_ne[4];
+    size_t output_stride[4];
+
+    size_t local_buffer_elems;
+
+    int64_t ir;
+    int64_t dr;
+
+    TPipe pipe;
+    GlobalTensor<float> input_gm;
+    GlobalTensor<int32_t> indices_gm;
+    GlobalTensor<float> output_gm;
+    TQue<QuePosition::VECIN, BUFFER_NUM> input_queue;
+    TQue<QuePosition::VECOUT, BUFFER_NUM> output_queue;
+};
+
+template <typename T>
+__aicore__ inline void copy_to_ub(GM_ADDR gm, T *ub, size_t size) {
+    auto gm_ptr = (__gm__ uint8_t *)gm;
+    auto ub_ptr = (uint8_t *)(ub);
+    for (int32_t i = 0; i < size; ++i, ++ub_ptr, ++gm_ptr) {
+        *ub_ptr = *gm_ptr;
+    }
+}
+
+extern "C" __global__ __aicore__ void ascendc_get_row_f32(
+    GM_ADDR input_gm, GM_ADDR indices_gm, GM_ADDR output_gm,
+    GM_ADDR input_ne_gm, GM_ADDR input_nb_gm, GM_ADDR indices_ne_gm,
+    GM_ADDR indices_nb_gm, GM_ADDR output_ne_gm, GM_ADDR output_nb_gm) {
+    int64_t input_ne_ub[4];
+    size_t input_nb_ub[4];
+    int64_t indices_ne_ub[4];
+    size_t indices_nb_ub[4];
+    int64_t output_ne_ub[4];
+    size_t output_nb_ub[4];
+
+    copy_to_ub(input_ne_gm, input_ne_ub, 32);
+    copy_to_ub(input_nb_gm, input_nb_ub, 32);
+    copy_to_ub(indices_ne_gm, indices_ne_ub, 32);
+    copy_to_ub(indices_nb_gm, indices_nb_ub, 32);
+    copy_to_ub(output_ne_gm, output_ne_ub, 32);
+    copy_to_ub(output_nb_gm, output_nb_ub, 32);
+
+    GET_ROW_F32 op;
+    op.init(input_gm, indices_gm, output_gm, input_ne_ub, input_nb_ub,
+            indices_ne_ub, indices_nb_ub, output_ne_ub, output_nb_ub);
+    op.calculate();
+}
diff --git a/src/whisper_cpp/ggml/src/ggml-cann/kernels/get_row_q4_0.cpp b/src/whisper_cpp/ggml/src/ggml-cann/kernels/get_row_q4_0.cpp
new file mode 100644
index 00000000..a80bfeec
--- /dev/null
+++ b/src/whisper_cpp/ggml/src/ggml-cann/kernels/get_row_q4_0.cpp
@@ -0,0 +1,193 @@
+#include "kernel_operator.h"
+
+// optimize me. Use template to avoid copy code.
+using namespace AscendC;
+
+#define BUFFER_NUM 2
+
+#define QK4_0 32
+
+class GET_ROW_Q4_0 {
+   public:
+    __aicore__ inline GET_ROW_Q4_0() {}
+    __aicore__ inline void init(GM_ADDR input, GM_ADDR indices, GM_ADDR output,
+                                int64_t *input_ne_ub, int64_t *indices_ne_ub,
+                                size_t *indices_nb_ub, int64_t *output_ne_ub,
+                                size_t *output_nb_ub) {
+        int64_t op_block_num = GetBlockNum();
+        int64_t op_block_idx = GetBlockIdx();
+
+        for (int i = 0; i < 4; i++) {
+            input_ne[i] = input_ne_ub[i];
+            indices_ne[i] = indices_ne_ub[i];
+            indices_stride[i] = indices_nb_ub[i] / indices_nb_ub[0];
+            scale_ne[i] = input_ne_ub[i];
+            output_ne[i] = output_ne_ub[i];
+            output_stride[i] = output_nb_ub[i] / output_nb_ub[0];
+        }
+
+        // one scale for a group.
+        scale_ne[0] /= QK4_0;
+
+        input_stride[0] = 1;
+        scale_stride[0] = 1;
+        output_stride[0] = 1;
+        for (int i = 1; i < 4; i++) {
+            input_stride[i] = input_stride[i - 1] * input_ne[i - 1];
+            scale_stride[i] = scale_stride[i - 1] * scale_ne[i - 1];
+        }
+
+        group_size_in_row = input_ne[0] / QK4_0;
+        int64_t scale_offset = input_ne[0] * input_ne[1] * input_ne[2] *
+                               input_ne[3] / 2;
+
+        // Indices has two dims. n_elements = all rows should get.
+        // dr, all rows should this thread get.
+        uint64_t n_elements =
+            indices_ne[0] * indices_ne[1] * indices_ne[2] * indices_ne[3];
+        dr = n_elements / op_block_num;
+
+        uint64_t tails = n_elements % op_block_num;
+        if (op_block_idx < tails) {
+            dr += 1;
+            ir = dr * op_block_idx;
+        } else {
+            ir = dr * op_block_idx + tails;
+        }
+
+        input_gm.SetGlobalBuffer((__gm__ int4b_t *)input);
+        scale_gm.SetGlobalBuffer((__gm__ half *)(input + scale_offset));
+        indices_gm.SetGlobalBuffer((__gm__ int32_t *)indices);
+        output_gm.SetGlobalBuffer((__gm__ float *)output);
+
+        pipe.InitBuffer(input_queue, BUFFER_NUM, QK4_0 * sizeof(int4b_t));
+        pipe.InitBuffer(cast_queue, BUFFER_NUM, QK4_0 * sizeof(half));
+        pipe.InitBuffer(output_queue, BUFFER_NUM, QK4_0 * sizeof(float));
+    }
+
+    __aicore__ inline void copy_in(uint32_t offset) {
+        LocalTensor<int4b_t> input_local = input_queue.AllocTensor<int4b_t>();
+        // 32 * sizeof(int4b_t) = 16, which is not aligned to 32, why no error?
+        DataCopy(input_local, input_gm[offset], QK4_0);
+        input_queue.EnQue(input_local);
+    }
+
+    __aicore__ inline void copy_out(uint32_t offset) {
+        LocalTensor<float> output_local = output_queue.DeQue<float>();
+        DataCopy(output_gm[offset], output_local, QK4_0);
+        output_queue.FreeTensor(output_local);
+    }
+
+    __aicore__ inline void calculate_group(int64_t idx, int64_t group) {
+        const int64_t indices_ne2_idx = idx / (indices_ne[0] * indices_ne[1]);
+        const int64_t indices_ne1_idx =
+            (idx - indices_ne2_idx * indices_ne[0] * indices_ne[1]) /
+            indices_ne[0];
+        const int64_t indices_ne0_idx =
+            (idx - indices_ne2_idx * indices_ne[0] * indices_ne[1] -
+             indices_ne1_idx * indices_ne[0]);
+
+        const int64_t indices_offset = indices_ne0_idx * indices_stride[0] +
+                                       indices_ne1_idx * indices_stride[1] +
+                                       indices_ne2_idx * indices_stride[2];
+        const int32_t selected_row_idx = indices_gm.GetValue(indices_offset);
+
+        const int64_t input_offset = selected_row_idx * input_stride[1] +
+                                     indices_ne1_idx * input_stride[2] +
+                                     indices_ne2_idx * input_stride[3] +
+                                     group * QK4_0;
+        const int64_t scale_offset = selected_row_idx * scale_stride[1] +
+                                     indices_ne1_idx * scale_stride[2] +
+                                     indices_ne2_idx * scale_stride[3] + group;
+        const int64_t output_offset = indices_ne0_idx * output_stride[1] +
+                                      indices_ne1_idx * output_stride[2] +
+                                      indices_ne2_idx * output_stride[3] +
+                                      group * QK4_0;
+
+        copy_in(input_offset);
+        LocalTensor<int4b_t> input_local = input_queue.DeQue<int4b_t>();
+        LocalTensor<half> cast_local = cast_queue.AllocTensor<half>();
+        LocalTensor<float> output_local = output_queue.AllocTensor<float>();
+
+        // TODO: cast more data to speed up.
+        Cast(cast_local, input_local, RoundMode::CAST_NONE, QK4_0);
+        Cast(output_local, cast_local, RoundMode::CAST_NONE, QK4_0);
+
+        // Only mul need compile by group.
+        half scale = scale_gm.GetValue(scale_offset);
+
+        Muls(output_local, output_local, (float)scale, QK4_0);
+
+        input_queue.FreeTensor(input_local);
+        cast_queue.FreeTensor(cast_local);
+        output_queue.EnQue(output_local);
+
+        copy_out(output_offset);
+    }
+
+    __aicore__ inline void calculate() {
+        for (int64_t i = ir; i < ir + dr; i++) {
+            for (int64_t j = 0; j < group_size_in_row; j++) {
+                calculate_group(i, j);
+            }
+        }
+    }
+
+   private:
+    int64_t input_ne[4];
+    size_t input_stride[4];
+
+    int64_t scale_ne[4];
+    size_t scale_stride[4];
+
+    int64_t indices_ne[4];
+    size_t indices_stride[4];
+
+    int64_t output_ne[4];
+    size_t output_stride[4];
+
+    int64_t ir;
+    int64_t dr;
+
+    int64_t group_size_in_row;
+
+    TPipe pipe;
+    GlobalTensor<int4b_t> input_gm;
+    GlobalTensor<half> scale_gm;
+    GlobalTensor<int32_t> indices_gm;
+    GlobalTensor<float> output_gm;
+    TQue<QuePosition::VECIN, BUFFER_NUM> input_queue;
+    TQue<QuePosition::VECOUT, BUFFER_NUM> output_queue;
+    TQue<QuePosition::VECIN, BUFFER_NUM> cast_queue;
+};
+
+template <typename T>
+__aicore__ inline void copy_to_ub(GM_ADDR gm, T *ub, size_t size) {
+    auto gm_ptr = (__gm__ uint8_t *)gm;
+    auto ub_ptr = (uint8_t *)(ub);
+    for (int32_t i = 0; i < size; ++i, ++ub_ptr, ++gm_ptr) {
+        *ub_ptr = *gm_ptr;
+    }
+}
+
+extern "C" __global__ __aicore__ void ascendc_get_row_q4_0(
+    GM_ADDR input_gm, GM_ADDR indices_gm, GM_ADDR output_gm,
+    GM_ADDR input_ne_gm, GM_ADDR indices_ne_gm, GM_ADDR indices_nb_gm,
+    GM_ADDR output_ne_gm, GM_ADDR output_nb_gm) {
+    int64_t input_ne_ub[4];
+    int64_t indices_ne_ub[4];
+    size_t indices_nb_ub[4];
+    int64_t output_ne_ub[4];
+    size_t output_nb_ub[4];
+
+    copy_to_ub(input_ne_gm, input_ne_ub, 32);
+    copy_to_ub(indices_ne_gm, indices_ne_ub, 32);
+    copy_to_ub(indices_nb_gm, indices_nb_ub, 32);
+    copy_to_ub(output_ne_gm, output_ne_ub, 32);
+    copy_to_ub(output_nb_gm, output_nb_ub, 32);
+
+    GET_ROW_Q4_0 op;
+    op.init(input_gm, indices_gm, output_gm, input_ne_ub, indices_ne_ub,
+            indices_nb_ub, output_ne_ub, output_nb_ub);
+    op.calculate();
+}
diff --git a/src/whisper_cpp/ggml/src/ggml-cann/kernels/get_row_q8_0.cpp b/src/whisper_cpp/ggml/src/ggml-cann/kernels/get_row_q8_0.cpp
new file mode 100644
index 00000000..ba9ab3c0
--- /dev/null
+++ b/src/whisper_cpp/ggml/src/ggml-cann/kernels/get_row_q8_0.cpp
@@ -0,0 +1,191 @@
+#include "kernel_operator.h"
+
+// optimize me. Use template to avoid copy code.
+using namespace AscendC;
+
+#define BUFFER_NUM 2
+
+#define QK8_0 32
+
+class GET_ROW_Q8_0 {
+   public:
+    __aicore__ inline GET_ROW_Q8_0() {}
+    __aicore__ inline void init(GM_ADDR input, GM_ADDR indices, GM_ADDR output,
+                                int64_t *input_ne_ub, int64_t *indices_ne_ub,
+                                size_t *indices_nb_ub, int64_t *output_ne_ub,
+                                size_t *output_nb_ub) {
+        int64_t op_block_num = GetBlockNum();
+        int64_t op_block_idx = GetBlockIdx();
+
+        for (int i = 0; i < 4; i++) {
+            input_ne[i] = input_ne_ub[i];
+            indices_ne[i] = indices_ne_ub[i];
+            indices_stride[i] = indices_nb_ub[i] / indices_nb_ub[0];
+            scale_ne[i] = input_ne_ub[i];
+            output_ne[i] = output_ne_ub[i];
+            output_stride[i] = output_nb_ub[i] / output_nb_ub[0];
+        }
+
+        // one scale for a group.
+        scale_ne[0] /= QK8_0;
+
+        input_stride[0] = 1;
+        scale_stride[0] = 1;
+        output_stride[0] = 1;
+        for (int i = 1; i < 4; i++) {
+            input_stride[i] = input_stride[i - 1] * input_ne[i - 1];
+            scale_stride[i] = scale_stride[i - 1] * scale_ne[i - 1];
+        }
+
+        group_size_in_row = input_ne[0] / QK8_0;
+        int64_t scale_offset = input_ne[0] * input_ne[1] * input_ne[2] *
+                               input_ne[3] * sizeof(int8_t);
+
+        // Indices has two dims. n_elements = all rows should get.
+        // dr, all rows should this thread get.
+        uint64_t n_elements =
+            indices_ne[0] * indices_ne[1] * indices_ne[2] * indices_ne[3];
+        dr = n_elements / op_block_num;
+
+        uint64_t tails = n_elements % op_block_num;
+        if (op_block_idx < tails) {
+            dr += 1;
+            ir = dr * op_block_idx;
+        } else {
+            ir = dr * op_block_idx + tails;
+        }
+
+        input_gm.SetGlobalBuffer((__gm__ int8_t *)input);
+        scale_gm.SetGlobalBuffer((__gm__ half *)(input + scale_offset));
+        indices_gm.SetGlobalBuffer((__gm__ int32_t *)indices);
+        output_gm.SetGlobalBuffer((__gm__ float *)output);
+
+        pipe.InitBuffer(input_queue, BUFFER_NUM, QK8_0 * sizeof(int8_t));
+        pipe.InitBuffer(cast_queue, BUFFER_NUM, QK8_0 * sizeof(half));
+        pipe.InitBuffer(output_queue, BUFFER_NUM, QK8_0 * sizeof(float));
+    }
+
+    __aicore__ inline void copy_in(uint32_t offset) {
+        LocalTensor<int8_t> input_local = input_queue.AllocTensor<int8_t>();
+        DataCopy(input_local, input_gm[offset], QK8_0);
+        input_queue.EnQue(input_local);
+    }
+
+    __aicore__ inline void copy_out(uint32_t offset) {
+        LocalTensor<float> output_local = output_queue.DeQue<float>();
+        DataCopy(output_gm[offset], output_local, QK8_0);
+        output_queue.FreeTensor(output_local);
+    }
+
+    __aicore__ inline void calculate_group(int64_t idx, int64_t group) {
+        const int64_t indices_ne2_idx = idx / (indices_ne[0] * indices_ne[1]);
+        const int64_t indices_ne1_idx =
+            (idx - indices_ne2_idx * indices_ne[0] * indices_ne[1]) /
+            indices_ne[0];
+        const int64_t indices_ne0_idx =
+            (idx - indices_ne2_idx * indices_ne[0] * indices_ne[1] -
+             indices_ne1_idx * indices_ne[0]);
+
+        const int64_t indices_offset = indices_ne0_idx * indices_stride[0] +
+                                       indices_ne1_idx * indices_stride[1] +
+                                       indices_ne2_idx * indices_stride[2];
+        const int32_t selected_row_idx = indices_gm.GetValue(indices_offset);
+
+        const int64_t input_offset = selected_row_idx * input_stride[1] +
+                                     indices_ne1_idx * input_stride[2] +
+                                     indices_ne2_idx * input_stride[3] +
+                                     group * QK8_0;
+        const int64_t scale_offset = selected_row_idx * scale_stride[1] +
+                                     indices_ne1_idx * scale_stride[2] +
+                                     indices_ne2_idx * scale_stride[3] + group;
+        const int64_t output_offset = indices_ne0_idx * output_stride[1] +
+                                      indices_ne1_idx * output_stride[2] +
+                                      indices_ne2_idx * output_stride[3] +
+                                      group * QK8_0;
+
+        copy_in(input_offset);
+        LocalTensor<int8_t> input_local = input_queue.DeQue<int8_t>();
+        LocalTensor<half> cast_local = cast_queue.AllocTensor<half>();
+        LocalTensor<float> output_local = output_queue.AllocTensor<float>();
+
+        // TODO: cast more data to speed up.
+        Cast(cast_local, input_local, RoundMode::CAST_NONE, QK8_0);
+        Cast(output_local, cast_local, RoundMode::CAST_NONE, QK8_0);
+
+        // Only mul need compile by group.
+        half scale = scale_gm.GetValue(scale_offset);
+        Muls(output_local, output_local, (float)scale, QK8_0);
+
+        input_queue.FreeTensor(input_local);
+        cast_queue.FreeTensor(cast_local);
+        output_queue.EnQue(output_local);
+
+        copy_out(output_offset);
+    }
+
+    __aicore__ inline void calculate() {
+        for (int64_t i = ir; i < ir + dr; i++) {
+            for (int64_t j = 0; j < group_size_in_row; j++) {
+                calculate_group(i, j);
+            }
+        }
+    }
+
+   private:
+    int64_t input_ne[4];
+    size_t input_stride[4];
+
+    int64_t scale_ne[4];
+    size_t scale_stride[4];
+
+    int64_t indices_ne[4];
+    size_t indices_stride[4];
+
+    int64_t output_ne[4];
+    size_t output_stride[4];
+
+    int64_t ir;
+    int64_t dr;
+
+    int64_t group_size_in_row;
+
+    TPipe pipe;
+    GlobalTensor<int8_t> input_gm;
+    GlobalTensor<half> scale_gm;
+    GlobalTensor<int32_t> indices_gm;
+    GlobalTensor<float> output_gm;
+    TQue<QuePosition::VECIN, BUFFER_NUM> input_queue;
+    TQue<QuePosition::VECOUT, BUFFER_NUM> output_queue;
+    TQue<QuePosition::VECIN, BUFFER_NUM> cast_queue;
+};
+
+template <typename T>
+__aicore__ inline void copy_to_ub(GM_ADDR gm, T *ub, size_t size) {
+    auto gm_ptr = (__gm__ uint8_t *)gm;
+    auto ub_ptr = (uint8_t *)(ub);
+    for (int32_t i = 0; i < size; ++i, ++ub_ptr, ++gm_ptr) {
+        *ub_ptr = *gm_ptr;
+    }
+}
+
+extern "C" __global__ __aicore__ void ascendc_get_row_q8_0(
+    GM_ADDR input_gm, GM_ADDR indices_gm, GM_ADDR output_gm,
+    GM_ADDR input_ne_gm, GM_ADDR indices_ne_gm, GM_ADDR indices_nb_gm,
+    GM_ADDR output_ne_gm, GM_ADDR output_nb_gm) {
+    int64_t input_ne_ub[4];
+    int64_t indices_ne_ub[4];
+    size_t indices_nb_ub[4];
+    int64_t output_ne_ub[4];
+    size_t output_nb_ub[4];
+
+    copy_to_ub(input_ne_gm, input_ne_ub, 32);
+    copy_to_ub(indices_ne_gm, indices_ne_ub, 32);
+    copy_to_ub(indices_nb_gm, indices_nb_ub, 32);
+    copy_to_ub(output_ne_gm, output_ne_ub, 32);
+    copy_to_ub(output_nb_gm, output_nb_ub, 32);
+
+    GET_ROW_Q8_0 op;
+    op.init(input_gm, indices_gm, output_gm, input_ne_ub, indices_ne_ub,
+            indices_nb_ub, output_ne_ub, output_nb_ub);
+    op.calculate();
+}
diff --git a/src/whisper_cpp/ggml/src/ggml-cann/kernels/quantize_f16_q8_0.cpp b/src/whisper_cpp/ggml/src/ggml-cann/kernels/quantize_f16_q8_0.cpp
new file mode 100644
index 00000000..8423b3f0
--- /dev/null
+++ b/src/whisper_cpp/ggml/src/ggml-cann/kernels/quantize_f16_q8_0.cpp
@@ -0,0 +1,208 @@
+#include "kernel_operator.h"
+
+using namespace AscendC;
+
+#define BUFFER_NUM 2
+#define QK8_0 32
+
+class QUANTIZE_F16_Q8_0 {
+   public:
+    __aicore__ inline QUANTIZE_F16_Q8_0() {}
+    __aicore__ inline void init(GM_ADDR input, GM_ADDR output,
+                                int64_t *input_ne_ub, size_t *input_nb_ub,
+                                int64_t *output_ne_ub) {
+        int64_t op_block_num = GetBlockNum();
+        int64_t op_block_idx = GetBlockIdx();
+
+        for (int i = 0; i < 4; i++) {
+            input_ne[i] = input_ne_ub[i];
+            input_stride[i] = input_nb_ub[i] / input_nb_ub[0];
+
+            output_ne[i] = output_ne_ub[i];
+        }
+
+        output_stride[0] = 1;
+        for (int i = 1; i < 4; i++) {
+            output_stride[i] = output_stride[i - 1] * output_ne[i - 1];
+        }
+
+        scale_ne = input_ne;
+        scale_stride[0] = 1;
+        scale_stride[1] = input_ne[0] / QK8_0;
+        for (int i = 2; i < 4; i++) {
+            scale_stride[i] = scale_stride[i - 1] * scale_ne[i - 1];
+        }
+
+        // split input tensor by rows.
+        uint64_t nr = input_ne[1] * input_ne[2] * input_ne[3];
+        dr = nr / op_block_num;
+
+        uint64_t tails = nr % op_block_num;
+        if (op_block_idx < tails) {
+            dr += 1;
+            ir = dr * op_block_idx;
+        } else {
+            ir = dr * op_block_idx + tails;
+        }
+
+        group_size_in_row = scale_stride[1];
+        int64_t output_size = output_ne[0] * output_ne[1] * output_ne[2] *
+                              output_ne[3] * sizeof(uint8_t);
+
+        input_gm.SetGlobalBuffer((__gm__ half *)input);
+        output_gm.SetGlobalBuffer((__gm__ int8_t *)output);
+        scale_gm.SetGlobalBuffer((__gm__ half *)(output + output_size + ir *
+                                                 group_size_in_row *
+                                                 sizeof(half)));
+
+        pipe.InitBuffer(input_queue, BUFFER_NUM, QK8_0 * sizeof(half));
+        pipe.InitBuffer(output_queue, BUFFER_NUM, QK8_0 * sizeof(int8_t));
+        pipe.InitBuffer(work_queue, 1, 32);
+        pipe.InitBuffer(max_queue, 1, 32);
+        pipe.InitBuffer(abs_queue, 1, QK8_0 * sizeof(float));
+        pipe.InitBuffer(scale_queue, 1, 32);
+        pipe.InitBuffer(cast_queue ,1 ,QK8_0 * sizeof(float));
+    }
+
+    __aicore__ inline void copy_in(uint32_t offset) {
+        LocalTensor<half> input_local = input_queue.AllocTensor<half>();
+        DataCopy(input_local, input_gm[offset], QK8_0);
+        input_queue.EnQue(input_local);
+    }
+
+    __aicore__ inline void copy_out(uint32_t offset) {
+        LocalTensor<int8_t> output_local = output_queue.DeQue<int8_t>();
+        DataCopy(output_gm[offset], output_local, QK8_0);
+        output_queue.FreeTensor(output_local);
+    }
+
+    __aicore__ inline half calculate_group(int64_t row, int64_t group) {
+        const int64_t i3 = row / (input_ne[1] * input_ne[2]);
+        const int64_t i2 = (row - i3 * input_ne[1] * input_ne[2]) / input_ne[1];
+        const int64_t i1 =
+            row - i3 * input_ne[1] * input_ne[2] - i2 * input_ne[1];
+
+        const int64_t input_offset = i1 * input_stride[1] +
+                                     i2 * input_stride[2] +
+                                     i3 * input_stride[3] + QK8_0 * group;
+
+        const int64_t output_offset = i1 * output_stride[1] +
+                                      i2 * output_stride[2] +
+                                      i3 * output_stride[3] + QK8_0 * group;
+
+        copy_in(input_offset);
+        LocalTensor<half> input_local = input_queue.DeQue<half>();
+        LocalTensor<int8_t> output_local = output_queue.AllocTensor<int8_t>();
+        LocalTensor<float> work_local = work_queue.AllocTensor<float>();
+        LocalTensor<float> abs_local = abs_queue.AllocTensor<float>();
+        LocalTensor<float> max_local = max_queue.AllocTensor<float>();
+        LocalTensor<float> cast_local = cast_queue.AllocTensor<float>();
+
+        Cast(cast_local, input_local, RoundMode::CAST_NONE, QK8_0);
+        Abs(abs_local, cast_local, QK8_0);
+        ReduceMax(max_local, abs_local, work_local, QK8_0);
+
+        pipe_barrier(PIPE_ALL);
+        float d = max_local.GetValue(0);
+        d = d / ((1 << 7) - 1);
+        if (d != 0) {
+            Muls(cast_local, cast_local, 1.0f / d, QK8_0);
+        }
+
+        Cast(cast_local, cast_local, RoundMode::CAST_ROUND, QK8_0);
+        Cast(input_local, cast_local, RoundMode::CAST_ROUND, QK8_0);
+        Cast(output_local, input_local, RoundMode::CAST_ROUND, QK8_0);
+        output_queue.EnQue(output_local);
+        copy_out(output_offset);
+
+        input_queue.FreeTensor(input_local);
+        work_queue.FreeTensor(work_local);
+        abs_queue.FreeTensor(abs_local);
+        max_queue.FreeTensor(max_local);
+        cast_queue.FreeTensor(cast_local);
+        return (half)d;
+    }
+
+    __aicore__ inline void calculate() {
+        LocalTensor<half> scale_local = scale_queue.AllocTensor<half>();
+        uint32_t scale_local_offset = 0;
+        uint32_t scale_global_offset = 0;
+        for (int64_t i = ir; i < ir + dr; i++) {
+            for (int64_t j = 0; j < group_size_in_row; j++) {
+                half scale = calculate_group(i, j);
+                scale_local.SetValue(scale_local_offset++, scale);
+                if (scale_local_offset == 16) {
+                    scale_local_offset = 0;
+                    // TODO: OPTIMIZE ME
+                    pipe_barrier(PIPE_ALL);
+                    DataCopy(scale_gm[scale_global_offset], scale_local, 16);
+                    pipe_barrier(PIPE_ALL);
+                    scale_global_offset += 16;
+                }
+            }
+        }
+
+        if (scale_local_offset != 0) {
+            pipe_barrier(PIPE_ALL);
+            DataCopyExtParams dataCopyParams;
+            dataCopyParams.blockCount = 1;
+            dataCopyParams.blockLen = scale_local_offset * sizeof(half);
+            DataCopyPad(scale_gm[scale_global_offset], scale_local,
+                        dataCopyParams);
+            pipe_barrier(PIPE_ALL);
+        }
+    }
+
+   private:
+    int64_t input_ne[4];
+    size_t input_stride[4];
+
+    int64_t *scale_ne;
+    size_t scale_stride[4];
+
+    int64_t output_ne[4];
+    size_t output_stride[4];
+
+    int64_t group_size_in_row;
+
+    int64_t ir;
+    int64_t dr;
+
+    TPipe pipe;
+    GlobalTensor<half> input_gm;
+    GlobalTensor<half> scale_gm;
+    GlobalTensor<int8_t> output_gm;
+    TQue<QuePosition::VECIN, BUFFER_NUM> input_queue;
+    TQue<QuePosition::VECOUT, BUFFER_NUM> output_queue;
+    TQue<QuePosition::VECIN, 1> work_queue;
+    TQue<QuePosition::VECOUT, 1> max_queue;
+    TQue<QuePosition::VECIN, 1> abs_queue;
+    TQue<QuePosition::VECOUT, 1> scale_queue;
+    TQue<QuePosition::VECOUT, 1> cast_queue;
+
+};
+
+template <typename T>
+__aicore__ inline void copy_to_ub(GM_ADDR gm, T *ub, size_t size) {
+    auto gm_ptr = (__gm__ uint8_t *)gm;
+    auto ub_ptr = (uint8_t *)(ub);
+    for (int32_t i = 0; i < size; ++i, ++ub_ptr, ++gm_ptr) {
+        *ub_ptr = *gm_ptr;
+    }
+}
+
+extern "C" __global__ __aicore__ void ascendc_quantize_f16_q8_0(
+    GM_ADDR input_gm, GM_ADDR output_gm, GM_ADDR input_ne_gm,
+    GM_ADDR input_nb_gm, GM_ADDR output_ne_gm) {
+    int64_t input_ne_ub[4];
+    size_t input_nb_ub[4];
+    int64_t output_ne_ub[4];
+
+    copy_to_ub(input_ne_gm, input_ne_ub, 32);
+    copy_to_ub(input_nb_gm, input_nb_ub, 32);
+    copy_to_ub(output_ne_gm, output_ne_ub, 32);
+
+    QUANTIZE_F16_Q8_0 op;
+    op.init(input_gm, output_gm, input_ne_ub, input_nb_ub, output_ne_ub);
+    op.calculate();
+}
diff --git a/src/whisper_cpp/ggml/src/ggml-cann/kernels/quantize_f32_q8_0.cpp b/src/whisper_cpp/ggml/src/ggml-cann/kernels/quantize_f32_q8_0.cpp
new file mode 100644
index 00000000..b7c57509
--- /dev/null
+++ b/src/whisper_cpp/ggml/src/ggml-cann/kernels/quantize_f32_q8_0.cpp
@@ -0,0 +1,206 @@
+#include "kernel_operator.h"
+
+using namespace AscendC;
+
+#define BUFFER_NUM 2
+#define QK8_0 32
+
+class QUANTIZE_F32_Q8_0 {
+   public:
+    __aicore__ inline QUANTIZE_F32_Q8_0() {}
+    __aicore__ inline void init(GM_ADDR input, GM_ADDR output,
+                                int64_t *input_ne_ub, size_t *input_nb_ub,
+                                int64_t *output_ne_ub) {
+        int64_t op_block_num = GetBlockNum();
+        int64_t op_block_idx = GetBlockIdx();
+
+        for (int i = 0; i < 4; i++) {
+            input_ne[i] = input_ne_ub[i];
+            input_stride[i] = input_nb_ub[i] / input_nb_ub[0];
+
+            output_ne[i] = output_ne_ub[i];
+        }
+
+        output_stride[0] = 1;
+        for (int i = 1; i < 4; i++) {
+            output_stride[i] = output_stride[i - 1] * output_ne[i - 1];
+        }
+
+        scale_ne = input_ne;
+        scale_stride[0] = 1;
+        scale_stride[1] = input_ne[0] / QK8_0;
+        for (int i = 2; i < 4; i++) {
+            scale_stride[i] = scale_stride[i - 1] * scale_ne[i - 1];
+        }
+
+        // split input tensor by rows.
+        uint64_t nr = input_ne[1] * input_ne[2] * input_ne[3];
+        dr = nr / op_block_num;
+
+        uint64_t tails = nr % op_block_num;
+        if (op_block_idx < tails) {
+            dr += 1;
+            ir = dr * op_block_idx;
+        } else {
+            ir = dr * op_block_idx + tails;
+        }
+
+        group_size_in_row = scale_stride[1];
+        int64_t output_size = output_ne[0] * output_ne[1] * output_ne[2] *
+                              output_ne[3] * sizeof(uint8_t);
+
+        input_gm.SetGlobalBuffer((__gm__ float *)input);
+        output_gm.SetGlobalBuffer((__gm__ int8_t *)output);
+        scale_gm.SetGlobalBuffer((__gm__ half *)(output + output_size +
+                                                 ir * group_size_in_row *
+                                                 sizeof(half)));
+
+        pipe.InitBuffer(input_queue, BUFFER_NUM, QK8_0 * sizeof(float));
+        pipe.InitBuffer(output_queue, BUFFER_NUM, QK8_0 * sizeof(int8_t));
+        pipe.InitBuffer(work_queue, 1, 32);
+        pipe.InitBuffer(max_queue, 1, 32);
+        pipe.InitBuffer(abs_queue, 1, QK8_0 * sizeof(float));
+        pipe.InitBuffer(cast_queue, 1, QK8_0 * sizeof(half));
+        pipe.InitBuffer(scale_queue, 1, 32);
+    }
+
+    __aicore__ inline void copy_in(uint32_t offset) {
+        LocalTensor<float> input_local = input_queue.AllocTensor<float>();
+        DataCopy(input_local, input_gm[offset], QK8_0);
+        input_queue.EnQue(input_local);
+    }
+
+    __aicore__ inline void copy_out(uint32_t offset) {
+        LocalTensor<int8_t> output_local = output_queue.DeQue<int8_t>();
+        DataCopy(output_gm[offset], output_local, QK8_0);
+        output_queue.FreeTensor(output_local);
+    }
+
+    __aicore__ inline half calculate_group(int64_t row, int64_t group) {
+        const int64_t i3 = row / (input_ne[1] * input_ne[2]);
+        const int64_t i2 = (row - i3 * input_ne[1] * input_ne[2]) / input_ne[1];
+        const int64_t i1 =
+            row - i3 * input_ne[1] * input_ne[2] - i2 * input_ne[1];
+
+        const int64_t input_offset = i1 * input_stride[1] +
+                                     i2 * input_stride[2] +
+                                     i3 * input_stride[3] + QK8_0 * group;
+
+        const int64_t output_offset = i1 * output_stride[1] +
+                                      i2 * output_stride[2] +
+                                      i3 * output_stride[3] + QK8_0 * group;
+
+        copy_in(input_offset);
+        LocalTensor<float> input_local = input_queue.DeQue<float>();
+        LocalTensor<int8_t> output_local = output_queue.AllocTensor<int8_t>();
+        LocalTensor<float> work_local = work_queue.AllocTensor<float>();
+        LocalTensor<float> abs_local = abs_queue.AllocTensor<float>();
+        LocalTensor<float> max_local = max_queue.AllocTensor<float>();
+        LocalTensor<half> cast_local = cast_queue.AllocTensor<half>();
+
+        Abs(abs_local, input_local, QK8_0);
+        ReduceMax(max_local, abs_local, work_local, QK8_0);
+        pipe_barrier(PIPE_ALL);
+        float d = max_local.GetValue(0);
+        d = d / ((1 << 7) - 1);
+        if (d != 0) {
+            Muls(input_local, input_local, 1.0f / d, QK8_0);
+        }
+
+        Cast(input_local, input_local, RoundMode::CAST_ROUND, QK8_0);
+        Cast(cast_local, input_local, RoundMode::CAST_ROUND, QK8_0);
+        Cast(output_local, cast_local, RoundMode::CAST_ROUND, QK8_0);
+        output_queue.EnQue(output_local);
+        copy_out(output_offset);
+
+        input_queue.FreeTensor(input_local);
+        work_queue.FreeTensor(work_local);
+        abs_queue.FreeTensor(abs_local);
+        max_queue.FreeTensor(max_local);
+        cast_queue.FreeTensor(cast_local);
+
+        return (half)d;
+    }
+
+    __aicore__ inline void calculate() {
+        LocalTensor<half> scale_local = scale_queue.AllocTensor<half>();
+        uint32_t scale_local_offset = 0;
+        uint32_t scale_global_offset = 0;
+        for (int64_t i = ir; i < ir + dr; i++) {
+            for (int64_t j = 0; j < group_size_in_row; j++) {
+                half scale = calculate_group(i, j);
+                scale_local.SetValue(scale_local_offset++, scale);
+                if (scale_local_offset == 16) {
+                    scale_local_offset = 0;
+                    // TODO: OPTIMIZE ME
+                    pipe_barrier(PIPE_ALL);
+                    DataCopy(scale_gm[scale_global_offset], scale_local, 16);
+                    pipe_barrier(PIPE_ALL);
+                    scale_global_offset += 16;
+                }
+            }
+        }
+
+        if (scale_local_offset != 0) {
+            pipe_barrier(PIPE_ALL);
+            DataCopyExtParams dataCopyParams;
+            dataCopyParams.blockCount = 1;
+            dataCopyParams.blockLen = scale_local_offset * sizeof(half);
+            DataCopyPad(scale_gm[scale_global_offset], scale_local,
+                        dataCopyParams);
+            pipe_barrier(PIPE_ALL);
+        }
+    }
+
+   private:
+    int64_t input_ne[4];
+    size_t input_stride[4];
+
+    int64_t *scale_ne;
+    size_t scale_stride[4];
+
+    int64_t output_ne[4];
+    size_t output_stride[4];
+
+    int64_t group_size_in_row;
+
+    int64_t ir;
+    int64_t dr;
+
+    TPipe pipe;
+    GlobalTensor<float> input_gm;
+    GlobalTensor<half> scale_gm;
+    GlobalTensor<int8_t> output_gm;
+    TQue<QuePosition::VECIN, BUFFER_NUM> input_queue;
+    TQue<QuePosition::VECOUT, BUFFER_NUM> output_queue;
+    TQue<QuePosition::VECIN, 1> work_queue;
+    TQue<QuePosition::VECOUT, 1> max_queue;
+    TQue<QuePosition::VECIN, 1> abs_queue;
+    TQue<QuePosition::VECIN, 1> cast_queue;
+    TQue<QuePosition::VECOUT, 1> scale_queue;
+};
+
+template <typename T>
+__aicore__ inline void copy_to_ub(GM_ADDR gm, T *ub, size_t size) {
+    auto gm_ptr = (__gm__ uint8_t *)gm;
+    auto ub_ptr = (uint8_t *)(ub);
+    for (int32_t i = 0; i < size; ++i, ++ub_ptr, ++gm_ptr) {
+        *ub_ptr = *gm_ptr;
+    }
+}
+
+extern "C" __global__ __aicore__ void ascendc_quantize_f32_q8_0(
+    GM_ADDR input_gm, GM_ADDR output_gm, GM_ADDR input_ne_gm,
+    GM_ADDR input_nb_gm, GM_ADDR output_ne_gm) {
+    int64_t input_ne_ub[4];
+    size_t input_nb_ub[4];
+    int64_t output_ne_ub[4];
+
+    copy_to_ub(input_ne_gm, input_ne_ub, 32);
+    copy_to_ub(input_nb_gm, input_nb_ub, 32);
+    copy_to_ub(output_ne_gm, output_ne_ub, 32);
+
+    QUANTIZE_F32_Q8_0 op;
+    op.init(input_gm, output_gm, input_ne_ub, input_nb_ub, output_ne_ub);
+    op.calculate();
+}
diff --git a/src/whisper_cpp/ggml/src/ggml-cann/kernels/quantize_float_to_q4_0.cpp b/src/whisper_cpp/ggml/src/ggml-cann/kernels/quantize_float_to_q4_0.cpp
new file mode 100644
index 00000000..9c8c86b6
--- /dev/null
+++ b/src/whisper_cpp/ggml/src/ggml-cann/kernels/quantize_float_to_q4_0.cpp
@@ -0,0 +1,278 @@
+#include "kernel_operator.h"
+
+using namespace AscendC;
+
+#define BUFFER_NUM 2
+#define Group_Size 32
+
+template <typename SRC_T>
+class QUANTIZE_FLOAT_TO_Q4_0 {
+   public:
+    __aicore__ inline QUANTIZE_FLOAT_TO_Q4_0() {}
+    __aicore__ inline void init(GM_ADDR input, GM_ADDR output,
+                                int64_t *input_ne_ub, size_t *input_nb_ub,
+                                int64_t *output_ne_ub) {
+        // TODO: fix test_case CPY(type_src=f16,type_dst=q4_0,ne=[256,4,4,4],
+        //                         permute=[0,0,0,0]):
+        // [CPY] NMSE = 0.000008343 > 0.000001000 FAIL
+        int64_t op_block_num = GetBlockNum();
+        int64_t op_block_idx = GetBlockIdx();
+
+        // input stride of data elements
+        for (int i = 0; i < 4; i++) {
+            input_ne[i] = input_ne_ub[i];
+            input_stride[i] = input_nb_ub[i] / input_nb_ub[0];
+            output_ne[i] = output_ne_ub[i];
+        }
+
+        // output stride of data elements
+        output_stride[0] = 1;
+        for (int i = 1; i < 4; i++) {
+            output_stride[i] = output_stride[i - 1] * output_ne[i - 1];
+        }
+
+        // scale saved one by one after data:. [group1_scale, group2_scale, ...]
+        scale_ne = input_ne;
+        scale_stride[0] = 1;
+        scale_stride[1] = input_ne[0] / Group_Size;
+        for (int i = 2; i < 4; i++) {
+            scale_stride[i] = scale_stride[i - 1] * scale_ne[i - 1];
+        }
+
+        // split input tensor by rows.
+        uint64_t nr = input_ne[1] * input_ne[2] * input_ne[3];
+        dr = nr / op_block_num;
+
+        uint64_t tails = nr % op_block_num;
+        if (op_block_idx < tails) {
+            dr += 1;
+            ir = dr * op_block_idx;
+        } else {
+            ir = dr * op_block_idx + tails;
+        }
+
+        group_size_in_row = scale_stride[1];
+        int64_t scale_offset = output_ne[0] * output_ne[1] * output_ne[2] *
+                              output_ne[3] * sizeof(uint8_t) / 2;
+
+        input_gm.SetGlobalBuffer((__gm__ SRC_T *)input);
+        output_gm.SetGlobalBuffer((__gm__ int8_t *)output);
+        scale_gm.SetGlobalBuffer((__gm__ half *)(output + scale_offset + ir *
+                                                 group_size_in_row *
+                                                 sizeof(half)));
+
+        pipe.InitBuffer(input_queue, BUFFER_NUM, Group_Size * sizeof(SRC_T));
+        pipe.InitBuffer(output_queue, BUFFER_NUM,
+                            Group_Size * sizeof(int8_t) / 2);
+        pipe.InitBuffer(cast_queue , 1, Group_Size * sizeof(float));
+        pipe.InitBuffer(work_queue, 1, Group_Size * sizeof(float));
+        pipe.InitBuffer(max_queue, 1, Group_Size * sizeof(float));
+        pipe.InitBuffer(min_queue, 1, Group_Size * sizeof(float));
+        pipe.InitBuffer(scale_queue, 1, Group_Size / 2 * sizeof(half));
+        pipe.InitBuffer(int8_queue, 1, Group_Size * sizeof(int8_t));
+        pipe.InitBuffer(half_queue, 1, Group_Size * sizeof(half));
+    }
+
+    __aicore__ inline void copy_in(uint32_t offset) {
+        LocalTensor<SRC_T> input_local = input_queue.AllocTensor<SRC_T>();
+        DataCopy(input_local, input_gm[offset], Group_Size);
+        input_queue.EnQue(input_local);
+    }
+
+    __aicore__ inline void copy_out(uint32_t offset) {
+        // reinterpretcast Group_Size(32) * int4b_t to Group_Size / 2 * int8_t,
+        // and using DataCopyPad to avoid 32 bits align.
+        LocalTensor<int4b_t> output_local = output_queue.DeQue<int4b_t>();
+        LocalTensor<int8_t> output_int8_local =
+                                    output_local.ReinterpretCast<int8_t>();
+
+        DataCopyExtParams dataCopyParams;
+        dataCopyParams.blockCount = 1;
+        dataCopyParams.blockLen = Group_Size / 2  * sizeof(int8_t);
+        DataCopyPad(output_gm[offset], output_int8_local, dataCopyParams);
+
+        output_queue.FreeTensor(output_local);
+    }
+
+    __aicore__ inline void input_to_cast(LocalTensor<float> cast_local,
+                                         LocalTensor<float> input_local) {
+        DataCopy(cast_local, input_local, Group_Size);
+    }
+
+    __aicore__ inline void input_to_cast(LocalTensor<float> cast_local,
+                                         LocalTensor<half> input_local) {
+        Cast(cast_local, input_local, RoundMode::CAST_NONE, Group_Size);
+    }
+
+    __aicore__ inline half calculate_group(int64_t row, int64_t group) {
+        const int64_t i3 = row / (input_ne[1] * input_ne[2]);
+        const int64_t i2 = (row - i3 * input_ne[1] * input_ne[2]) / input_ne[1];
+        const int64_t i1 =
+            row - i3 * input_ne[1] * input_ne[2] - i2 * input_ne[1];
+
+        const int64_t input_offset = i1 * input_stride[1] +
+                                     i2 * input_stride[2] +
+                                     i3 * input_stride[3] + Group_Size * group;
+
+        // output_offset is stride for output_gm which datatype is int8_t and
+        // divided by 2 is needed for int4b_t.
+        const int64_t output_offset = (i1 * output_stride[1] +
+                                       i2 * output_stride[2] +
+                                       i3 * output_stride[3] +
+                                       Group_Size * group) / 2;
+        copy_in(input_offset);
+
+        LocalTensor<SRC_T> input_local = input_queue.DeQue<SRC_T>();
+        LocalTensor<int4b_t> output_local = output_queue.AllocTensor<int4b_t>();
+        LocalTensor<float> cast_local = cast_queue.AllocTensor<float>();
+        LocalTensor<float> work_local = work_queue.AllocTensor<float>();
+        LocalTensor<float> max_local = max_queue.AllocTensor<float>();
+        LocalTensor<float> min_local = min_queue.AllocTensor<float>();
+        LocalTensor<int8_t> int8_local = int8_queue.AllocTensor<int8_t>();
+        LocalTensor<half> half_local = half_queue.AllocTensor<half>();
+
+        input_to_cast(cast_local, input_local);
+
+        ReduceMax(max_local, cast_local, work_local, Group_Size);
+        ReduceMin(min_local, cast_local, work_local, Group_Size);
+        const float max_value = max_local.GetValue(0);
+        const float min_value = min_local.GetValue(0);
+        float d = max_value;
+        if (min_value < 0 && (-1 * min_value) > max_value) {
+            d = min_value;
+        }
+
+        d = d / (-8);
+        if (d != 0) {
+            Muls(cast_local, cast_local, 1.0f / d, Group_Size);
+        }
+
+        // range: [-8,8] -> [0.5,16.5] -> [0,16] -> [0,15] -> [-8,7]
+        float scalar = 8.5f;
+        Adds(cast_local, cast_local, scalar, Group_Size);
+        Cast(cast_local, cast_local, RoundMode::CAST_FLOOR, Group_Size);
+        scalar = 15.0f;
+        Mins(cast_local, cast_local, scalar, Group_Size);
+        scalar = -8.0f;
+        Adds(cast_local, cast_local, scalar, Group_Size);
+
+        // float->half->int4b
+        Cast(half_local, cast_local, RoundMode::CAST_NONE, Group_Size);
+        Cast(output_local, half_local, RoundMode::CAST_NONE, Group_Size);
+
+        output_queue.EnQue(output_local);
+        copy_out(output_offset);
+
+        input_queue.FreeTensor(input_local);
+        work_queue.FreeTensor(work_local);
+        max_queue.FreeTensor(max_local);
+        min_queue.FreeTensor(min_local);
+        int8_queue.FreeTensor(int8_local);
+        half_queue.FreeTensor(half_local);
+        cast_queue.FreeTensor(cast_local);
+        return (half)d;
+    }
+
+    __aicore__ inline void calculate() {
+        LocalTensor<half> scale_local = scale_queue.AllocTensor<half>();
+        uint32_t scale_local_offset = 0;
+        uint32_t scale_global_offset = 0;
+        for (int64_t i = ir; i < ir + dr; i++) {
+            for (int64_t j = 0; j < group_size_in_row; j++) {
+                half scale = calculate_group(i, j);
+                scale_local.SetValue(scale_local_offset++, scale);
+                // Copy Group_Size/2 length data each time.
+                if (scale_local_offset == Group_Size / 2) {
+                    scale_local_offset = 0;
+                    // TODO: OPTIMIZE ME
+                    pipe_barrier(PIPE_ALL);
+                    DataCopy(scale_gm[scale_global_offset], scale_local,
+                                      Group_Size / 2);
+                    pipe_barrier(PIPE_ALL);
+                    scale_global_offset += Group_Size / 2;
+                }
+            }
+        }
+
+        if (scale_local_offset != 0) {
+            pipe_barrier(PIPE_ALL);
+            DataCopyExtParams dataCopyParams;
+            dataCopyParams.blockCount = 1;
+            dataCopyParams.blockLen = scale_local_offset * sizeof(half);
+            DataCopyPad(scale_gm[scale_global_offset], scale_local,
+                        dataCopyParams);
+            pipe_barrier(PIPE_ALL);
+        }
+        scale_queue.FreeTensor(scale_local);
+    }
+
+   private:
+    int64_t input_ne[4];
+    size_t input_stride[4];
+
+    int64_t *scale_ne;
+    size_t scale_stride[4];
+
+    int64_t output_ne[4];
+    size_t output_stride[4];
+
+    int64_t group_size_in_row;
+
+    int64_t ir;
+    int64_t dr;
+
+    TPipe pipe;
+    GlobalTensor<SRC_T> input_gm;
+    GlobalTensor<half> scale_gm;
+    GlobalTensor<int8_t> output_gm;
+    TQue<QuePosition::VECIN, BUFFER_NUM> input_queue;
+    TQue<QuePosition::VECOUT, BUFFER_NUM> output_queue;
+    TQue<QuePosition::VECIN, BUFFER_NUM> work_queue;
+    TQue<QuePosition::VECOUT, BUFFER_NUM> max_queue;
+    TQue<QuePosition::VECOUT, BUFFER_NUM> min_queue;
+    TQue<QuePosition::VECOUT, BUFFER_NUM> scale_queue;
+    TQue<QuePosition::VECOUT, BUFFER_NUM> cast_queue;
+    TQue<QuePosition::VECOUT, BUFFER_NUM> int8_queue;
+    TQue<QuePosition::VECOUT, BUFFER_NUM> half_queue;
+};
+
+template <typename T>
+__aicore__ inline void copy_to_ub(GM_ADDR gm, T *ub, size_t size) {
+    auto gm_ptr = (__gm__ uint8_t *)gm;
+    auto ub_ptr = (uint8_t *)(ub);
+    for (int32_t i = 0; i < size; ++i, ++ub_ptr, ++gm_ptr) {
+        *ub_ptr = *gm_ptr;
+    }
+}
+
+extern "C" __global__ __aicore__ void ascendc_quantize_f16_to_q4_0(
+    GM_ADDR input_gm, GM_ADDR output_gm, GM_ADDR input_ne_gm,
+    GM_ADDR input_nb_gm, GM_ADDR output_ne_gm) {
+    int64_t input_ne_ub[4];
+    size_t input_nb_ub[4];
+    int64_t output_ne_ub[4];
+
+    copy_to_ub(input_ne_gm, input_ne_ub, 32);
+    copy_to_ub(input_nb_gm, input_nb_ub, 32);
+    copy_to_ub(output_ne_gm, output_ne_ub, 32);
+
+    QUANTIZE_FLOAT_TO_Q4_0<half> op;
+    op.init(input_gm, output_gm, input_ne_ub, input_nb_ub, output_ne_ub);
+    op.calculate();
+}
+
+extern "C" __global__ __aicore__ void ascendc_quantize_f32_to_q4_0(
+    GM_ADDR input_gm, GM_ADDR output_gm, GM_ADDR input_ne_gm,
+    GM_ADDR input_nb_gm, GM_ADDR output_ne_gm) {
+    int64_t input_ne_ub[4];
+    size_t input_nb_ub[4];
+    int64_t output_ne_ub[4];
+
+    copy_to_ub(input_ne_gm, input_ne_ub, 32);
+    copy_to_ub(input_nb_gm, input_nb_ub, 32);
+    copy_to_ub(output_ne_gm, output_ne_ub, 32);
+
+    QUANTIZE_FLOAT_TO_Q4_0<float> op;
+    op.init(input_gm, output_gm, input_ne_ub, input_nb_ub, output_ne_ub);
+    op.calculate();
+}
diff --git a/src/whisper_cpp/ggml/src/ggml-common.h b/src/whisper_cpp/ggml/src/ggml-common.h
new file mode 100644
index 00000000..05016139
--- /dev/null
+++ b/src/whisper_cpp/ggml/src/ggml-common.h
@@ -0,0 +1,1853 @@
+#ifndef GGML_COMMON_DECL
+
+#if defined(GGML_COMMON_DECL_C)
+#include <stdint.h>
+
+typedef uint16_t ggml_half;
+typedef uint32_t ggml_half2;
+
+#define GGML_COMMON_AGGR
+
+#define GGML_COMMON_DECL
+#elif defined(GGML_COMMON_DECL_METAL)
+#include <metal_stdlib>
+
+typedef half  ggml_half;
+typedef half2 ggml_half2;
+
+#define GGML_COMMON_AGGR
+
+#define GGML_COMMON_DECL
+#elif defined(GGML_COMMON_DECL_CUDA)
+#if defined(GGML_COMMON_DECL_MUSA)
+#include <musa_fp16.h>
+#else
+#include <cuda_fp16.h>
+#endif
+#include <cstdint>
+
+typedef half  ggml_half;
+typedef half2 ggml_half2;
+
+#define GGML_COMMON_AGGR data
+
+#define GGML_COMMON_DECL
+#elif defined(GGML_COMMON_DECL_HIP)
+#include <hip/hip_fp16.h>
+#include <cstdint>
+
+typedef half  ggml_half;
+typedef half2 ggml_half2;
+
+#define GGML_COMMON_AGGR data
+
+#define GGML_COMMON_DECL
+#elif defined(GGML_COMMON_DECL_SYCL)
+#include <sycl/half_type.hpp>
+#include <cstdint>
+
+typedef sycl::half  ggml_half;
+typedef sycl::half2 ggml_half2;
+
+#define GGML_COMMON_AGGR data
+
+#define GGML_COMMON_DECL
+#endif
+
+#if defined(GGML_COMMON_DECL)
+
+#ifndef __cplusplus
+#ifndef static_assert
+#if defined(__STDC_VERSION__) && (__STDC_VERSION__ >= 201100L)
+#define static_assert(cond, msg) _Static_assert(cond, msg)
+#else
+#define static_assert(cond, msg) struct global_scope_noop_trick
+#endif
+#endif
+#endif // __cplusplus
+
+// QK = number of values after dequantization
+// QK_K = super-block size
+
+#define QK_K 256
+#define K_SCALE_SIZE 12
+
+#if defined(GGML_COMMON_DECL_CUDA) || defined(GGML_COMMON_DECL_HIP) || defined(GGML_COMMON_DECL_SYCL)
+// QR = QK / number of values before dequantization
+// QI = number of 32 bit integers before dequantization
+
+#define QI4_0 (QK4_0 / (4 * QR4_0))
+#define QR4_0 2
+
+#define QI4_1 (QK4_1 / (4 * QR4_1))
+#define QR4_1 2
+
+#define QI5_0 (QK5_0 / (4 * QR5_0))
+#define QR5_0 2
+
+#define QI5_1 (QK5_1 / (4 * QR5_1))
+#define QR5_1 2
+
+#define QI8_0 (QK8_0 / (4 * QR8_0))
+#define QR8_0 1
+
+#define QI8_1 (QK8_1 / (4 * QR8_1))
+#define QR8_1 1
+
+#define QI2_K (QK_K / (4*QR2_K))
+#define QR2_K 4
+
+#define QI3_K (QK_K / (4*QR3_K))
+#define QR3_K 4
+
+#define QI4_K (QK_K / (4*QR4_K))
+#define QR4_K 2
+
+#define QI5_K (QK_K / (4*QR5_K))
+#define QR5_K 2
+
+#define QI6_K (QK_K / (4*QR6_K))
+#define QR6_K 2
+
+#define QI2_XXS (QK_K / (4*QR2_XXS))
+#define QR2_XXS 4
+
+#define QI2_XS (QK_K / (4*QR2_XS))
+#define QR2_XS 4
+
+#define QI2_S (QK_K / (4*QR2_S))
+#define QR2_S 4
+
+#define QI3_XXS (QK_K / (4*QR3_XXS))
+#define QR3_XXS 4
+
+#define QI3_XS (QK_K / (4*QR3_XS))
+#define QR3_XS 4
+
+#define QI1_S (QK_K / (4*QR1_S))
+#define QR1_S 8
+
+#define QI1_M (QK_K / (4*QR1_M))
+#define QR1_M 8
+
+#define QI4_NL (QK4_NL / (4*QR4_NL))
+#define QR4_NL 2
+
+#define QI4_XS (QK_K / (4*QR4_XS))
+#define QR4_XS 2
+
+#define QI3_S (QK_K / (4*QR3_S))
+#define QR3_S 4
+
+#endif // GGML_COMMON_DECL_CUDA || GGML_COMMON_DECL_HIP
+
+#define QK4_0 32
+typedef struct {
+    ggml_half d;           // delta
+    uint8_t qs[QK4_0 / 2]; // nibbles / quants
+} block_q4_0;
+static_assert(sizeof(block_q4_0) == sizeof(ggml_half) + QK4_0 / 2, "wrong q4_0 block size/padding");
+
+#define QK4_1 32
+typedef struct {
+    union {
+        struct {
+            ggml_half d; // delta
+            ggml_half m; // min
+        } GGML_COMMON_AGGR;
+        ggml_half2 dm;
+    };
+    uint8_t qs[QK4_1 / 2]; // nibbles / quants
+} block_q4_1;
+static_assert(sizeof(block_q4_1) == 2 * sizeof(ggml_half) + QK4_1 / 2, "wrong q4_1 block size/padding");
+
+#define QK5_0 32
+typedef struct {
+    ggml_half d;           // delta
+    uint8_t qh[4];         // 5-th bit of quants
+    uint8_t qs[QK5_0 / 2]; // nibbles / quants
+} block_q5_0;
+static_assert(sizeof(block_q5_0) == sizeof(ggml_half) + sizeof(uint32_t) + QK5_0 / 2, "wrong q5_0 block size/padding");
+
+#define QK5_1 32
+typedef struct {
+    union {
+        struct {
+            ggml_half d; // delta
+            ggml_half m; // min
+        } GGML_COMMON_AGGR;
+        ggml_half2 dm;
+    };
+    uint8_t qh[4];         // 5-th bit of quants
+    uint8_t qs[QK5_1 / 2]; // nibbles / quants
+} block_q5_1;
+static_assert(sizeof(block_q5_1) == 2 * sizeof(ggml_half) + sizeof(uint32_t) + QK5_1 / 2, "wrong q5_1 block size/padding");
+
+#define QK8_0 32
+typedef struct {
+    ggml_half d;       // delta
+    int8_t  qs[QK8_0]; // quants
+} block_q8_0;
+static_assert(sizeof(block_q8_0) == sizeof(ggml_half) + QK8_0, "wrong q8_0 block size/padding");
+
+#define QK8_1 32
+typedef struct {
+    union {
+        struct {
+            ggml_half d; // delta
+            ggml_half s; // d * sum(qs[i])
+        } GGML_COMMON_AGGR;
+        ggml_half2 ds;
+    };
+    int8_t qs[QK8_1]; // quants
+} block_q8_1;
+static_assert(sizeof(block_q8_1) == 2*sizeof(ggml_half) + QK8_1, "wrong q8_1 block size/padding");
+
+typedef struct {
+    ggml_half d[4];        // deltas for 4 q4_0 blocks
+    uint8_t qs[QK4_0 * 2]; // nibbles / quants for 4 q4_0 blocks
+} block_q4_0x4;
+static_assert(sizeof(block_q4_0x4) == 4 * sizeof(ggml_half) + QK4_0 * 2, "wrong q4_0x4 block size/padding");
+
+typedef struct {
+    ggml_half d[8];        // deltas for 8 q4_0 blocks
+    uint8_t qs[QK4_0 * 4]; // nibbles / quants for 8 q4_0 blocks
+} block_q4_0x8;
+static_assert(sizeof(block_q4_0x8) == 8 * sizeof(ggml_half) + QK4_0 * 4, "wrong q4_0x8 block size/padding");
+
+typedef struct {
+    ggml_half d[4];        // deltas for 4 q8_0 blocks
+    int8_t qs[QK8_0 * 4];  // quants for 4 q8_0 blocks
+} block_q8_0x4;
+static_assert(sizeof(block_q8_0x4) == 4 * sizeof(ggml_half) + QK8_0 * 4, "wrong q8_0x4 block size/padding");
+
+typedef struct {
+    ggml_half d[8];        // deltas for 8 q8_0 blocks
+    int8_t qs[QK8_0 * 8];  // quants for 8 q8_0 blocks
+} block_q8_0x8;
+static_assert(sizeof(block_q8_0x8) == 8 * sizeof(ggml_half) + QK8_0 * 8, "wrong q8_0x8 block size/padding");
+
+//
+// Ternary quantization
+//
+
+// 1.6875 bpw
+typedef struct {
+    uint8_t qs[(QK_K - 4 * QK_K / 64) / 5]; // 5 elements per byte (3^5 = 243 < 256)
+    uint8_t qh[QK_K/64]; // 4 elements per byte
+    ggml_half d;
+} block_tq1_0;
+static_assert(sizeof(block_tq1_0) == sizeof(ggml_half) + QK_K / 64 + (QK_K - 4 * QK_K / 64) / 5, "wrong tq1_0 block size/padding");
+
+// 2.0625 bpw
+typedef struct {
+    uint8_t qs[QK_K/4]; // 2 bits per element
+    ggml_half d;
+} block_tq2_0;
+static_assert(sizeof(block_tq2_0) == sizeof(ggml_half) + QK_K / 4, "wrong tq2_0 block size/padding");
+
+//
+// Super-block quantization structures
+//
+
+// 2-bit quantization
+// weight is represented as x = a * q + b
+// 16 blocks of 16 elements each
+// Effectively 2.625 bits per weight
+typedef struct {
+    uint8_t scales[QK_K/16]; // scales and mins, quantized with 4 bits
+    uint8_t qs[QK_K/4];      // quants
+    union {
+        struct {
+            ggml_half d;    // super-block scale for quantized scales
+            ggml_half dmin; // super-block scale for quantized mins
+        } GGML_COMMON_AGGR;
+        ggml_half2 dm;
+    };
+} block_q2_K;
+static_assert(sizeof(block_q2_K) == 2*sizeof(ggml_half) + QK_K/16 + QK_K/4, "wrong q2_K block size/padding");
+
+// 3-bit quantization
+// weight is represented as x = a * q
+// 16 blocks of 16 elements each
+// Effectively 3.4375 bits per weight
+typedef struct {
+    uint8_t hmask[QK_K/8]; // quants - high bit
+    uint8_t qs[QK_K/4];    // quants - low 2 bits
+    uint8_t scales[12];    // scales, quantized with 6 bits
+    ggml_half d;           // super-block scale
+} block_q3_K;
+static_assert(sizeof(block_q3_K) == sizeof(ggml_half) + QK_K / 4 + QK_K / 8 + 12, "wrong q3_K block size/padding");
+
+// 4-bit quantization
+// 8 blocks of 32 elements each
+// weight is represented as x = a * q + b
+// Effectively 4.5 bits per weight
+typedef struct {
+    union {
+        struct {
+            ggml_half d;    // super-block scale for quantized scales
+            ggml_half dmin; // super-block scale for quantized mins
+        } GGML_COMMON_AGGR;
+        ggml_half2 dm;
+    };
+    uint8_t scales[K_SCALE_SIZE]; // scales and mins, quantized with 6 bits
+    uint8_t qs[QK_K/2];           // 4--bit quants
+} block_q4_K;
+static_assert(sizeof(block_q4_K) == 2*sizeof(ggml_half) + K_SCALE_SIZE + QK_K/2, "wrong q4_K block size/padding");
+
+// 5-bit quantization
+// 8 blocks of 32 elements each
+// weight is represented as x = a * q + b
+// Effectively 5.5 bits per weight
+typedef struct {
+    union {
+        struct {
+            ggml_half d;    // super-block scale for quantized scales
+            ggml_half dmin; // super-block scale for quantized mins
+        } GGML_COMMON_AGGR;
+        ggml_half2 dm;
+    };
+    uint8_t scales[K_SCALE_SIZE]; // scales and mins, quantized with 6 bits
+    uint8_t qh[QK_K/8];           // quants, high bit
+    uint8_t qs[QK_K/2];           // quants, low 4 bits
+} block_q5_K;
+static_assert(sizeof(block_q5_K) == 2*sizeof(ggml_half) + K_SCALE_SIZE + QK_K/2 + QK_K/8, "wrong q5_K block size/padding");
+
+// 6-bit quantization
+// weight is represented as x = a * q
+// 16 blocks of 16 elements each
+// Effectively 6.5625 bits per weight
+typedef struct {
+    uint8_t ql[QK_K/2];      // quants, lower 4 bits
+    uint8_t qh[QK_K/4];      // quants, upper 2 bits
+    int8_t  scales[QK_K/16]; // scales, quantized with 8 bits
+    ggml_half d;             // super-block scale
+} block_q6_K;
+static_assert(sizeof(block_q6_K) == sizeof(ggml_half) + QK_K / 16 + 3*QK_K/4, "wrong q6_K block size/padding");
+
+// This is only used for intermediate quantization and dot products
+typedef struct {
+    float   d;              // delta
+    int8_t  qs[QK_K];       // quants
+    int16_t bsums[QK_K/16]; // sum of quants in groups of 16
+} block_q8_K;
+static_assert(sizeof(block_q8_K) == sizeof(float) + QK_K + QK_K/16*sizeof(int16_t), "wrong q8_K block size/padding");
+
+// (Almost) "true" 2-bit quantization.
+// Due to the need to use blocks as per ggml design, it ends up using
+// 2.0625 bpw because of the 16-bit scale for each block of 256.
+typedef struct {
+    ggml_half d;
+    uint16_t qs[QK_K/8];
+} block_iq2_xxs;
+static_assert(sizeof(block_iq2_xxs) == sizeof(ggml_half) + QK_K/8*sizeof(uint16_t), "wrong iq2_xxs block size/padding");
+
+// 2.3125 bpw quants
+typedef struct {
+    ggml_half d;
+    uint16_t qs[QK_K/8];
+    uint8_t  scales[QK_K/32];
+} block_iq2_xs;
+static_assert(sizeof(block_iq2_xs) == sizeof(ggml_half) + QK_K/8*sizeof(uint16_t) + QK_K/32, "wrong iq2_xs block size/padding");
+
+// 2.5625 bpw quants
+typedef struct {
+    ggml_half d;
+    uint8_t qs[QK_K/4];
+    uint8_t qh[QK_K/32];
+    uint8_t scales[QK_K/32];
+} block_iq2_s;
+static_assert(sizeof(block_iq2_s) == sizeof(ggml_half) + QK_K/4 + QK_K/16, "wrong iq2_s block size/padding");
+
+// (Almost) "true" 3-bit quantization.
+// Due to the need to use blocks as per ggml design, it ends up using
+// 3.0625 bpw because of the 16-bit scale for each block of 256.
+typedef struct {
+    ggml_half d;
+    uint8_t qs[3*QK_K/8];
+} block_iq3_xxs;
+static_assert(sizeof(block_iq3_xxs) == sizeof(ggml_half) + 3*(QK_K/8), "wrong iq3_xxs block size/padding");
+
+// 3.4375 bpw
+#define IQ3S_N_SCALE QK_K/64
+typedef struct {
+    ggml_half d;
+    uint8_t qs[QK_K/4];
+    uint8_t qh[QK_K/32];
+    uint8_t signs[QK_K/8];
+    uint8_t scales[IQ3S_N_SCALE];
+} block_iq3_s;
+static_assert(sizeof(block_iq3_s) == sizeof(ggml_half) + 13*(QK_K/32) + IQ3S_N_SCALE, "wrong iq3_s block size/padding");
+
+// 1.5625 bpw
+typedef struct {
+    ggml_half d;
+    uint8_t  qs[QK_K/8];
+    uint16_t qh[QK_K/32];
+} block_iq1_s;
+static_assert(sizeof(block_iq1_s) == sizeof(ggml_half) + QK_K/8 + QK_K/16, "wrong iq1_s block size/padding");
+
+// 1.75 bpw
+typedef struct {
+    uint8_t  qs[QK_K/8];      // grid index, low 8 bits
+    uint8_t  qh[QK_K/16];     // grid index, high 3 bits + grid shift bit (for two groups of 8)
+    uint8_t  scales[QK_K/32]; // 3-bit block scales (4-bit if QK_K == 64)
+} block_iq1_m;
+static_assert(sizeof(block_iq1_m) == QK_K/8 + QK_K/16 + QK_K/32, "wrong iq1_m block size/padding");
+
+// Used by IQ1_M quants
+typedef union {
+    ggml_half f16;
+    uint16_t  u16;
+} iq1m_scale_t;
+
+// Non-linear quants
+#define QK4_NL 32
+typedef struct {
+    ggml_half d;
+    uint8_t qs[QK4_NL/2];
+} block_iq4_nl;
+static_assert(sizeof(block_iq4_nl) == sizeof(ggml_half) + QK4_NL/2, "wrong iq4_nl block size/padding");
+
+typedef struct {
+    ggml_half d;
+    uint16_t scales_h;
+    uint8_t  scales_l[QK_K/64];
+    uint8_t  qs[QK_K/2];
+} block_iq4_xs;
+static_assert(sizeof(block_iq4_xs) == sizeof(ggml_half) + sizeof(uint16_t) + QK_K/64 + QK_K/2, "wrong iq4_xs block size/padding");
+
+#endif // GGML_COMMON_DECL
+#endif // GGML_COMMON_DECL
+
+////////////////////////////////////////////////////////////////////////////////
+
+#ifndef GGML_COMMON_IMPL
+
+#if defined(GGML_COMMON_IMPL_C)
+#include <stdint.h>
+
+#define GGML_TABLE_BEGIN(type, name, size) static const type name[size] = {
+#define GGML_TABLE_END() };
+
+#define GGML_COMMON_IMPL
+#elif defined(GGML_COMMON_IMPL_METAL)
+#include <metal_stdlib>
+
+#define GGML_TABLE_BEGIN(type, name, size) static const constant type name[size] = {
+#define GGML_TABLE_END() };
+
+#define GGML_COMMON_IMPL
+#elif defined(GGML_COMMON_IMPL_CUDA) || defined(GGML_COMMON_IMPL_HIP) || defined(GGML_COMMON_IMPL_MUSA)
+#include <cstdint>
+
+#define GGML_TABLE_BEGIN(type, name, size) static const __device__ type name[size] = {
+#define GGML_TABLE_END() };
+
+#define GGML_COMMON_IMPL
+#elif defined(GGML_COMMON_IMPL_SYCL)
+
+#include <cstdint>
+
+#define GGML_TABLE_BEGIN(type, name, size) static const type name[size] = {
+#define GGML_TABLE_END() };
+
+#define GGML_COMMON_IMPL
+#endif
+
+#if defined(GGML_COMMON_IMPL)
+
+GGML_TABLE_BEGIN(uint8_t, kmask_iq2xs, 8)
+    1, 2, 4, 8, 16, 32, 64, 128
+GGML_TABLE_END()
+
+GGML_TABLE_BEGIN(uint8_t, ksigns_iq2xs, 128)
+      0, 129, 130,   3, 132,   5,   6, 135, 136,   9,  10, 139,  12, 141, 142,  15,
+    144,  17,  18, 147,  20, 149, 150,  23,  24, 153, 154,  27, 156,  29,  30, 159,
+    160,  33,  34, 163,  36, 165, 166,  39,  40, 169, 170,  43, 172,  45,  46, 175,
+     48, 177, 178,  51, 180,  53,  54, 183, 184,  57,  58, 187,  60, 189, 190,  63,
+    192,  65,  66, 195,  68, 197, 198,  71,  72, 201, 202,  75, 204,  77,  78, 207,
+     80, 209, 210,  83, 212,  85,  86, 215, 216,  89,  90, 219,  92, 221, 222,  95,
+     96, 225, 226,  99, 228, 101, 102, 231, 232, 105, 106, 235, 108, 237, 238, 111,
+    240, 113, 114, 243, 116, 245, 246, 119, 120, 249, 250, 123, 252, 125, 126, 255,
+GGML_TABLE_END()
+
+//#if __CUDA_ARCH__ >= MIN_CC_DP4A // lowest compute capability for integer intrinsics
+GGML_TABLE_BEGIN(uint64_t, ksigns64, 128)
+    0x0000000000000000, 0xff000000000000ff, 0xff0000000000ff00, 0x000000000000ffff,
+    0xff00000000ff0000, 0x0000000000ff00ff, 0x0000000000ffff00, 0xff00000000ffffff,
+    0xff000000ff000000, 0x00000000ff0000ff, 0x00000000ff00ff00, 0xff000000ff00ffff,
+    0x00000000ffff0000, 0xff000000ffff00ff, 0xff000000ffffff00, 0x00000000ffffffff,
+    0xff0000ff00000000, 0x000000ff000000ff, 0x000000ff0000ff00, 0xff0000ff0000ffff,
+    0x000000ff00ff0000, 0xff0000ff00ff00ff, 0xff0000ff00ffff00, 0x000000ff00ffffff,
+    0x000000ffff000000, 0xff0000ffff0000ff, 0xff0000ffff00ff00, 0x000000ffff00ffff,
+    0xff0000ffffff0000, 0x000000ffffff00ff, 0x000000ffffffff00, 0xff0000ffffffffff,
+    0xff00ff0000000000, 0x0000ff00000000ff, 0x0000ff000000ff00, 0xff00ff000000ffff,
+    0x0000ff0000ff0000, 0xff00ff0000ff00ff, 0xff00ff0000ffff00, 0x0000ff0000ffffff,
+    0x0000ff00ff000000, 0xff00ff00ff0000ff, 0xff00ff00ff00ff00, 0x0000ff00ff00ffff,
+    0xff00ff00ffff0000, 0x0000ff00ffff00ff, 0x0000ff00ffffff00, 0xff00ff00ffffffff,
+    0x0000ffff00000000, 0xff00ffff000000ff, 0xff00ffff0000ff00, 0x0000ffff0000ffff,
+    0xff00ffff00ff0000, 0x0000ffff00ff00ff, 0x0000ffff00ffff00, 0xff00ffff00ffffff,
+    0xff00ffffff000000, 0x0000ffffff0000ff, 0x0000ffffff00ff00, 0xff00ffffff00ffff,
+    0x0000ffffffff0000, 0xff00ffffffff00ff, 0xff00ffffffffff00, 0x0000ffffffffffff,
+    0xffff000000000000, 0x00ff0000000000ff, 0x00ff00000000ff00, 0xffff00000000ffff,
+    0x00ff000000ff0000, 0xffff000000ff00ff, 0xffff000000ffff00, 0x00ff000000ffffff,
+    0x00ff0000ff000000, 0xffff0000ff0000ff, 0xffff0000ff00ff00, 0x00ff0000ff00ffff,
+    0xffff0000ffff0000, 0x00ff0000ffff00ff, 0x00ff0000ffffff00, 0xffff0000ffffffff,
+    0x00ff00ff00000000, 0xffff00ff000000ff, 0xffff00ff0000ff00, 0x00ff00ff0000ffff,
+    0xffff00ff00ff0000, 0x00ff00ff00ff00ff, 0x00ff00ff00ffff00, 0xffff00ff00ffffff,
+    0xffff00ffff000000, 0x00ff00ffff0000ff, 0x00ff00ffff00ff00, 0xffff00ffff00ffff,
+    0x00ff00ffffff0000, 0xffff00ffffff00ff, 0xffff00ffffffff00, 0x00ff00ffffffffff,
+    0x00ffff0000000000, 0xffffff00000000ff, 0xffffff000000ff00, 0x00ffff000000ffff,
+    0xffffff0000ff0000, 0x00ffff0000ff00ff, 0x00ffff0000ffff00, 0xffffff0000ffffff,
+    0xffffff00ff000000, 0x00ffff00ff0000ff, 0x00ffff00ff00ff00, 0xffffff00ff00ffff,
+    0x00ffff00ffff0000, 0xffffff00ffff00ff, 0xffffff00ffffff00, 0x00ffff00ffffffff,
+    0xffffffff00000000, 0x00ffffff000000ff, 0x00ffffff0000ff00, 0xffffffff0000ffff,
+    0x00ffffff00ff0000, 0xffffffff00ff00ff, 0xffffffff00ffff00, 0x00ffffff00ffffff,
+    0x00ffffffff000000, 0xffffffffff0000ff, 0xffffffffff00ff00, 0x00ffffffff00ffff,
+    0xffffffffffff0000, 0x00ffffffffff00ff, 0x00ffffffffffff00, 0xffffffffffffffff,
+GGML_TABLE_END()
+//#endif
+
+
+GGML_TABLE_BEGIN(uint64_t, iq2xxs_grid, 256)
+    0x0808080808080808, 0x080808080808082b, 0x0808080808081919, 0x0808080808082b08,
+    0x0808080808082b2b, 0x0808080808190819, 0x0808080808191908, 0x08080808082b0808,
+    0x08080808082b082b, 0x08080808082b2b08, 0x08080808082b2b2b, 0x0808080819080819,
+    0x0808080819081908, 0x0808080819190808, 0x0808080819192b08, 0x08080808192b0819,
+    0x08080808192b1908, 0x080808082b080808, 0x080808082b08082b, 0x080808082b082b2b,
+    0x080808082b2b082b, 0x0808081908080819, 0x0808081908081908, 0x0808081908190808,
+    0x0808081908191919, 0x0808081919080808, 0x080808192b081908, 0x080808192b192b08,
+    0x0808082b08080808, 0x0808082b0808082b, 0x0808082b082b082b, 0x0808082b2b08082b,
+    0x0808190808080819, 0x0808190808081908, 0x0808190808190808, 0x08081908082b0819,
+    0x08081908082b1908, 0x0808190819080808, 0x080819081908082b, 0x0808190819082b08,
+    0x08081908192b0808, 0x080819082b080819, 0x080819082b081908, 0x080819082b190808,
+    0x080819082b2b1908, 0x0808191908080808, 0x080819190808082b, 0x0808191908082b08,
+    0x08081919082b0808, 0x080819191908192b, 0x08081919192b2b19, 0x080819192b080808,
+    0x080819192b190819, 0x0808192b08082b19, 0x0808192b08190808, 0x0808192b19080808,
+    0x0808192b2b081908, 0x0808192b2b2b1908, 0x08082b0808080808, 0x08082b0808081919,
+    0x08082b0808082b08, 0x08082b0808191908, 0x08082b08082b2b08, 0x08082b0819080819,
+    0x08082b0819081908, 0x08082b0819190808, 0x08082b081919082b, 0x08082b082b082b08,
+    0x08082b1908081908, 0x08082b1919080808, 0x08082b2b0808082b, 0x08082b2b08191908,
+    0x0819080808080819, 0x0819080808081908, 0x0819080808190808, 0x08190808082b0819,
+    0x0819080819080808, 0x08190808192b0808, 0x081908082b081908, 0x081908082b190808,
+    0x081908082b191919, 0x0819081908080808, 0x0819081908082b08, 0x08190819082b0808,
+    0x0819081919190808, 0x0819081919192b2b, 0x081908192b080808, 0x0819082b082b1908,
+    0x0819082b19081919, 0x0819190808080808, 0x0819190808082b08, 0x08191908082b0808,
+    0x08191908082b1919, 0x0819190819082b19, 0x081919082b080808, 0x0819191908192b08,
+    0x08191919192b082b, 0x0819192b08080808, 0x0819192b0819192b, 0x08192b0808080819,
+    0x08192b0808081908, 0x08192b0808190808, 0x08192b0819080808, 0x08192b082b080819,
+    0x08192b1908080808, 0x08192b1908081919, 0x08192b192b2b0808, 0x08192b2b19190819,
+    0x082b080808080808, 0x082b08080808082b, 0x082b080808082b2b, 0x082b080819081908,
+    0x082b0808192b0819, 0x082b08082b080808, 0x082b08082b08082b, 0x082b0819082b2b19,
+    0x082b081919082b08, 0x082b082b08080808, 0x082b082b0808082b, 0x082b190808080819,
+    0x082b190808081908, 0x082b190808190808, 0x082b190819080808, 0x082b19081919192b,
+    0x082b191908080808, 0x082b191919080819, 0x082b1919192b1908, 0x082b192b2b190808,
+    0x082b2b0808082b08, 0x082b2b08082b0808, 0x082b2b082b191908, 0x082b2b2b19081908,
+    0x1908080808080819, 0x1908080808081908, 0x1908080808190808, 0x1908080808192b08,
+    0x19080808082b0819, 0x19080808082b1908, 0x1908080819080808, 0x1908080819082b08,
+    0x190808081919192b, 0x19080808192b0808, 0x190808082b080819, 0x190808082b081908,
+    0x190808082b190808, 0x1908081908080808, 0x19080819082b0808, 0x19080819192b0819,
+    0x190808192b080808, 0x190808192b081919, 0x1908082b08080819, 0x1908082b08190808,
+    0x1908082b19082b08, 0x1908082b1919192b, 0x1908082b192b2b08, 0x1908190808080808,
+    0x1908190808082b08, 0x19081908082b0808, 0x190819082b080808, 0x190819082b192b19,
+    0x190819190819082b, 0x19081919082b1908, 0x1908192b08080808, 0x19082b0808080819,
+    0x19082b0808081908, 0x19082b0808190808, 0x19082b0819080808, 0x19082b0819081919,
+    0x19082b1908080808, 0x19082b1919192b08, 0x19082b19192b0819, 0x19082b192b08082b,
+    0x19082b2b19081919, 0x19082b2b2b190808, 0x1919080808080808, 0x1919080808082b08,
+    0x1919080808190819, 0x1919080808192b19, 0x19190808082b0808, 0x191908082b080808,
+    0x191908082b082b08, 0x1919081908081908, 0x191908191908082b, 0x191908192b2b1908,
+    0x1919082b2b190819, 0x191919082b190808, 0x191919082b19082b, 0x1919191908082b2b,
+    0x1919192b08080819, 0x1919192b19191908, 0x19192b0808080808, 0x19192b0808190819,
+    0x19192b0808192b19, 0x19192b08192b1908, 0x19192b1919080808, 0x19192b2b08082b08,
+    0x192b080808081908, 0x192b080808190808, 0x192b080819080808, 0x192b0808192b2b08,
+    0x192b081908080808, 0x192b081919191919, 0x192b082b08192b08, 0x192b082b192b0808,
+    0x192b190808080808, 0x192b190808081919, 0x192b191908190808, 0x192b19190819082b,
+    0x192b19192b081908, 0x192b2b081908082b, 0x2b08080808080808, 0x2b0808080808082b,
+    0x2b08080808082b2b, 0x2b08080819080819, 0x2b0808082b08082b, 0x2b08081908081908,
+    0x2b08081908192b08, 0x2b08081919080808, 0x2b08082b08190819, 0x2b08190808080819,
+    0x2b08190808081908, 0x2b08190808190808, 0x2b08190808191919, 0x2b08190819080808,
+    0x2b081908192b0808, 0x2b08191908080808, 0x2b0819191908192b, 0x2b0819192b191908,
+    0x2b08192b08082b19, 0x2b08192b19080808, 0x2b08192b192b0808, 0x2b082b080808082b,
+    0x2b082b1908081908, 0x2b082b2b08190819, 0x2b19080808081908, 0x2b19080808190808,
+    0x2b190808082b1908, 0x2b19080819080808, 0x2b1908082b2b0819, 0x2b1908190819192b,
+    0x2b1908192b080808, 0x2b19082b19081919, 0x2b19190808080808, 0x2b191908082b082b,
+    0x2b19190819081908, 0x2b19191919190819, 0x2b192b082b080819, 0x2b192b19082b0808,
+    0x2b2b08080808082b, 0x2b2b080819190808, 0x2b2b08082b081919, 0x2b2b081908082b19,
+    0x2b2b082b08080808, 0x2b2b190808192b08, 0x2b2b2b0819190808, 0x2b2b2b1908081908,
+GGML_TABLE_END()
+
+GGML_TABLE_BEGIN(uint64_t, iq2xs_grid, 512)
+    0x0808080808080808, 0x080808080808082b, 0x0808080808081919, 0x0808080808082b08,
+    0x0808080808082b2b, 0x0808080808190819, 0x0808080808191908, 0x080808080819192b,
+    0x0808080808192b19, 0x08080808082b0808, 0x08080808082b082b, 0x08080808082b1919,
+    0x08080808082b2b08, 0x0808080819080819, 0x0808080819081908, 0x080808081908192b,
+    0x0808080819082b19, 0x0808080819190808, 0x080808081919082b, 0x0808080819191919,
+    0x0808080819192b08, 0x08080808192b0819, 0x08080808192b1908, 0x080808082b080808,
+    0x080808082b08082b, 0x080808082b081919, 0x080808082b082b08, 0x080808082b190819,
+    0x080808082b191908, 0x080808082b192b19, 0x080808082b2b0808, 0x0808081908080819,
+    0x0808081908081908, 0x080808190808192b, 0x0808081908082b19, 0x0808081908190808,
+    0x080808190819082b, 0x0808081908191919, 0x0808081908192b08, 0x0808081908192b2b,
+    0x08080819082b0819, 0x08080819082b1908, 0x0808081919080808, 0x080808191908082b,
+    0x0808081919081919, 0x0808081919082b08, 0x0808081919190819, 0x0808081919191908,
+    0x08080819192b0808, 0x08080819192b2b08, 0x080808192b080819, 0x080808192b081908,
+    0x080808192b190808, 0x0808082b08080808, 0x0808082b0808082b, 0x0808082b08081919,
+    0x0808082b08082b08, 0x0808082b08190819, 0x0808082b08191908, 0x0808082b082b0808,
+    0x0808082b19080819, 0x0808082b19081908, 0x0808082b19190808, 0x0808082b19191919,
+    0x0808082b2b080808, 0x0808082b2b082b2b, 0x0808190808080819, 0x0808190808081908,
+    0x080819080808192b, 0x0808190808082b19, 0x0808190808190808, 0x080819080819082b,
+    0x0808190808191919, 0x0808190808192b08, 0x08081908082b0819, 0x08081908082b1908,
+    0x0808190819080808, 0x080819081908082b, 0x0808190819081919, 0x0808190819082b08,
+    0x0808190819190819, 0x0808190819191908, 0x080819081919192b, 0x08081908192b0808,
+    0x080819082b080819, 0x080819082b081908, 0x080819082b190808, 0x0808191908080808,
+    0x080819190808082b, 0x0808191908081919, 0x0808191908082b08, 0x0808191908190819,
+    0x0808191908191908, 0x08081919082b0808, 0x0808191919080819, 0x0808191919081908,
+    0x0808191919190808, 0x08081919192b0819, 0x080819192b080808, 0x0808192b08080819,
+    0x0808192b08081908, 0x0808192b08190808, 0x0808192b082b192b, 0x0808192b19080808,
+    0x0808192b1908082b, 0x0808192b2b081908, 0x08082b0808080808, 0x08082b080808082b,
+    0x08082b0808081919, 0x08082b0808082b08, 0x08082b0808082b2b, 0x08082b0808190819,
+    0x08082b0808191908, 0x08082b08082b0808, 0x08082b08082b1919, 0x08082b0819080819,
+    0x08082b0819081908, 0x08082b0819190808, 0x08082b0819192b08, 0x08082b082b080808,
+    0x08082b082b2b0808, 0x08082b082b2b2b2b, 0x08082b1908080819, 0x08082b1908081908,
+    0x08082b1908190808, 0x08082b1919080808, 0x08082b192b080819, 0x08082b192b082b19,
+    0x08082b2b08080808, 0x08082b2b082b0808, 0x08082b2b082b2b08, 0x08082b2b2b19192b,
+    0x08082b2b2b2b0808, 0x0819080808080819, 0x0819080808081908, 0x081908080808192b,
+    0x0819080808082b19, 0x0819080808190808, 0x081908080819082b, 0x0819080808191919,
+    0x0819080808192b08, 0x08190808082b0819, 0x08190808082b1908, 0x0819080819080808,
+    0x081908081908082b, 0x0819080819081919, 0x0819080819082b08, 0x0819080819190819,
+    0x0819080819191908, 0x08190808192b0808, 0x08190808192b2b2b, 0x081908082b080819,
+    0x081908082b081908, 0x081908082b190808, 0x0819081908080808, 0x081908190808082b,
+    0x0819081908081919, 0x0819081908082b08, 0x0819081908190819, 0x0819081908191908,
+    0x08190819082b0808, 0x0819081919080819, 0x0819081919081908, 0x0819081919190808,
+    0x081908192b080808, 0x081908192b191908, 0x081908192b19192b, 0x0819082b08080819,
+    0x0819082b08081908, 0x0819082b0808192b, 0x0819082b08190808, 0x0819082b19080808,
+    0x0819082b192b0808, 0x0819190808080808, 0x081919080808082b, 0x0819190808081919,
+    0x0819190808082b08, 0x0819190808190819, 0x0819190808191908, 0x08191908082b0808,
+    0x0819190819080819, 0x0819190819081908, 0x0819190819082b19, 0x0819190819190808,
+    0x08191908192b1908, 0x081919082b080808, 0x0819191908080819, 0x0819191908081908,
+    0x0819191908190808, 0x0819191919080808, 0x0819192b08080808, 0x0819192b08191908,
+    0x0819192b19082b19, 0x08192b0808080819, 0x08192b0808081908, 0x08192b0808190808,
+    0x08192b080819082b, 0x08192b0819080808, 0x08192b0819191908, 0x08192b082b08192b,
+    0x08192b1908080808, 0x08192b1908081919, 0x08192b19192b192b, 0x08192b2b19190819,
+    0x08192b2b2b2b2b19, 0x082b080808080808, 0x082b08080808082b, 0x082b080808081919,
+    0x082b080808082b08, 0x082b080808082b2b, 0x082b080808190819, 0x082b080808191908,
+    0x082b0808082b0808, 0x082b080819080819, 0x082b080819081908, 0x082b080819190808,
+    0x082b08082b080808, 0x082b08082b2b0808, 0x082b081908080819, 0x082b081908081908,
+    0x082b081908190808, 0x082b081919080808, 0x082b081919082b08, 0x082b0819192b1919,
+    0x082b082b08080808, 0x082b082b082b082b, 0x082b082b2b080808, 0x082b082b2b2b2b08,
+    0x082b190808080819, 0x082b190808081908, 0x082b190808190808, 0x082b1908082b2b19,
+    0x082b190819080808, 0x082b191908080808, 0x082b191919080819, 0x082b19191919082b,
+    0x082b19192b192b19, 0x082b192b08080819, 0x082b192b08192b2b, 0x082b192b2b2b192b,
+    0x082b2b0808080808, 0x082b2b0808082b08, 0x082b2b0808082b2b, 0x082b2b08082b0808,
+    0x082b2b0819191919, 0x082b2b082b082b08, 0x082b2b082b2b082b, 0x082b2b19192b2b08,
+    0x082b2b192b190808, 0x082b2b2b08082b08, 0x082b2b2b082b0808, 0x082b2b2b2b08082b,
+    0x082b2b2b2b082b08, 0x082b2b2b2b082b2b, 0x1908080808080819, 0x1908080808081908,
+    0x190808080808192b, 0x1908080808082b19, 0x1908080808190808, 0x190808080819082b,
+    0x1908080808191919, 0x1908080808192b08, 0x19080808082b0819, 0x19080808082b1908,
+    0x1908080819080808, 0x190808081908082b, 0x1908080819081919, 0x1908080819082b08,
+    0x1908080819082b2b, 0x1908080819190819, 0x1908080819191908, 0x19080808192b0808,
+    0x19080808192b1919, 0x190808082b080819, 0x190808082b081908, 0x190808082b190808,
+    0x1908081908080808, 0x190808190808082b, 0x1908081908081919, 0x1908081908082b08,
+    0x1908081908190819, 0x1908081908191908, 0x19080819082b0808, 0x1908081919080819,
+    0x1908081919081908, 0x1908081919190808, 0x190808192b080808, 0x190808192b081919,
+    0x190808192b2b082b, 0x1908082b08080819, 0x1908082b08081908, 0x1908082b08190808,
+    0x1908082b0819082b, 0x1908082b082b2b19, 0x1908082b19080808, 0x1908190808080808,
+    0x190819080808082b, 0x1908190808081919, 0x1908190808082b08, 0x1908190808190819,
+    0x1908190808191908, 0x1908190808192b19, 0x19081908082b0808, 0x1908190819080819,
+    0x1908190819081908, 0x1908190819190808, 0x190819082b080808, 0x190819082b191908,
+    0x1908191908080819, 0x1908191908081908, 0x1908191908190808, 0x19081919082b1908,
+    0x1908191919080808, 0x190819192b192b2b, 0x1908192b08080808, 0x1908192b08082b2b,
+    0x1908192b19081908, 0x1908192b19190808, 0x19082b0808080819, 0x19082b0808081908,
+    0x19082b0808190808, 0x19082b0819080808, 0x19082b0819081919, 0x19082b0819191908,
+    0x19082b08192b082b, 0x19082b1908080808, 0x19082b1908190819, 0x19082b1919081908,
+    0x19082b1919190808, 0x19082b19192b2b19, 0x19082b2b08081908, 0x1919080808080808,
+    0x191908080808082b, 0x1919080808081919, 0x1919080808082b08, 0x1919080808190819,
+    0x1919080808191908, 0x19190808082b0808, 0x19190808082b2b08, 0x1919080819080819,
+    0x1919080819081908, 0x1919080819190808, 0x191908082b080808, 0x1919081908080819,
+    0x1919081908081908, 0x1919081908190808, 0x1919081908191919, 0x1919081919080808,
+    0x191908191908082b, 0x1919082b08080808, 0x1919082b19081908, 0x1919082b2b2b2b2b,
+    0x1919190808080819, 0x1919190808081908, 0x1919190808190808, 0x19191908082b0819,
+    0x1919190819080808, 0x19191908192b0808, 0x191919082b080819, 0x191919082b2b0819,
+    0x1919191908080808, 0x1919191908082b08, 0x191919192b080808, 0x191919192b082b08,
+    0x1919192b082b0819, 0x1919192b192b2b08, 0x1919192b2b2b0819, 0x19192b0808080808,
+    0x19192b0808191908, 0x19192b0819080819, 0x19192b0819190808, 0x19192b082b192b19,
+    0x19192b1908192b2b, 0x19192b1919080808, 0x19192b191908082b, 0x19192b2b2b081919,
+    0x192b080808080819, 0x192b080808081908, 0x192b080808190808, 0x192b080819080808,
+    0x192b080819191908, 0x192b0808192b082b, 0x192b08082b08192b, 0x192b08082b2b2b19,
+    0x192b081908080808, 0x192b082b082b1908, 0x192b082b19082b2b, 0x192b082b2b19082b,
+    0x192b190808080808, 0x192b19080819192b, 0x192b191908190808, 0x192b191919080808,
+    0x192b191919081919, 0x192b19192b2b1908, 0x192b2b0808080819, 0x192b2b08192b2b2b,
+    0x192b2b19082b1919, 0x192b2b2b0808192b, 0x192b2b2b19191908, 0x192b2b2b192b082b,
+    0x2b08080808080808, 0x2b0808080808082b, 0x2b08080808081919, 0x2b08080808082b08,
+    0x2b08080808190819, 0x2b08080808191908, 0x2b080808082b0808, 0x2b080808082b2b2b,
+    0x2b08080819080819, 0x2b08080819081908, 0x2b08080819190808, 0x2b0808082b080808,
+    0x2b0808082b08082b, 0x2b0808082b2b2b08, 0x2b0808082b2b2b2b, 0x2b08081908080819,
+    0x2b08081908081908, 0x2b0808190808192b, 0x2b08081908190808, 0x2b08081919080808,
+    0x2b08081919190819, 0x2b08081919192b19, 0x2b08082b08080808, 0x2b08082b082b0808,
+    0x2b08082b2b080808, 0x2b08082b2b08082b, 0x2b08082b2b2b0808, 0x2b08082b2b2b2b08,
+    0x2b08190808080819, 0x2b08190808081908, 0x2b08190808190808, 0x2b0819080819082b,
+    0x2b08190808191919, 0x2b08190819080808, 0x2b081908192b0808, 0x2b0819082b082b19,
+    0x2b08191908080808, 0x2b08191919081908, 0x2b0819192b2b1919, 0x2b08192b08192b08,
+    0x2b08192b192b2b2b, 0x2b082b0808080808, 0x2b082b0808082b08, 0x2b082b08082b1919,
+    0x2b082b0819192b2b, 0x2b082b082b080808, 0x2b082b082b08082b, 0x2b082b082b2b2b08,
+    0x2b082b190808192b, 0x2b082b2b082b082b, 0x2b082b2b2b080808, 0x2b082b2b2b082b08,
+    0x2b082b2b2b19192b, 0x2b082b2b2b2b2b08, 0x2b19080808080819, 0x2b19080808081908,
+    0x2b19080808190808, 0x2b19080819080808, 0x2b1908081919192b, 0x2b1908082b081908,
+    0x2b19081908080808, 0x2b190819082b082b, 0x2b190819192b1908, 0x2b19082b1919192b,
+    0x2b19082b2b082b19, 0x2b19190808080808, 0x2b19190808081919, 0x2b19190819081908,
+    0x2b19190819190808, 0x2b19190819192b08, 0x2b191919082b2b19, 0x2b1919192b190808,
+    0x2b1919192b19082b, 0x2b19192b19080819, 0x2b192b0819190819, 0x2b192b082b2b192b,
+    0x2b192b1919082b19, 0x2b192b2b08191919, 0x2b192b2b192b0808, 0x2b2b080808080808,
+    0x2b2b08080808082b, 0x2b2b080808082b08, 0x2b2b080808082b2b, 0x2b2b0808082b0808,
+    0x2b2b0808082b2b2b, 0x2b2b08082b2b0808, 0x2b2b081919190819, 0x2b2b081919192b19,
+    0x2b2b08192b2b192b, 0x2b2b082b08080808, 0x2b2b082b0808082b, 0x2b2b082b08082b08,
+    0x2b2b082b082b2b2b, 0x2b2b082b2b080808, 0x2b2b082b2b2b0808, 0x2b2b190819080808,
+    0x2b2b19082b191919, 0x2b2b192b192b1919, 0x2b2b192b2b192b08, 0x2b2b2b0808082b2b,
+    0x2b2b2b08082b0808, 0x2b2b2b08082b082b, 0x2b2b2b08082b2b08, 0x2b2b2b082b2b0808,
+    0x2b2b2b082b2b2b08, 0x2b2b2b1908081908, 0x2b2b2b192b081908, 0x2b2b2b192b08192b,
+    0x2b2b2b2b082b2b08, 0x2b2b2b2b082b2b2b, 0x2b2b2b2b2b190819, 0x2b2b2b2b2b2b2b2b,
+GGML_TABLE_END()
+
+GGML_TABLE_BEGIN(uint64_t, iq2s_grid, 1024)
+    0x0808080808080808, 0x080808080808082b, 0x0808080808081919, 0x0808080808082b08,
+    0x0808080808082b2b, 0x0808080808190819, 0x0808080808191908, 0x080808080819192b,
+    0x0808080808192b19, 0x08080808082b0808, 0x08080808082b082b, 0x08080808082b1919,
+    0x08080808082b2b08, 0x0808080819080819, 0x0808080819081908, 0x080808081908192b,
+    0x0808080819082b19, 0x0808080819190808, 0x080808081919082b, 0x0808080819191919,
+    0x0808080819192b08, 0x08080808192b0819, 0x08080808192b1908, 0x08080808192b192b,
+    0x08080808192b2b19, 0x080808082b080808, 0x080808082b08082b, 0x080808082b081919,
+    0x080808082b082b08, 0x080808082b190819, 0x080808082b191908, 0x080808082b2b0808,
+    0x080808082b2b1919, 0x080808082b2b2b2b, 0x0808081908080819, 0x0808081908081908,
+    0x080808190808192b, 0x0808081908082b19, 0x0808081908190808, 0x080808190819082b,
+    0x0808081908191919, 0x0808081908192b08, 0x08080819082b0819, 0x08080819082b1908,
+    0x0808081919080808, 0x080808191908082b, 0x0808081919081919, 0x0808081919082b08,
+    0x0808081919190819, 0x0808081919191908, 0x080808191919192b, 0x0808081919192b19,
+    0x08080819192b0808, 0x08080819192b1919, 0x08080819192b2b08, 0x080808192b080819,
+    0x080808192b081908, 0x080808192b190808, 0x080808192b19082b, 0x080808192b191919,
+    0x080808192b2b0819, 0x080808192b2b1908, 0x0808082b08080808, 0x0808082b0808082b,
+    0x0808082b08081919, 0x0808082b08082b08, 0x0808082b08190819, 0x0808082b08191908,
+    0x0808082b082b0808, 0x0808082b082b2b2b, 0x0808082b19080819, 0x0808082b19081908,
+    0x0808082b1908192b, 0x0808082b19082b19, 0x0808082b19190808, 0x0808082b19191919,
+    0x0808082b2b080808, 0x0808082b2b081919, 0x0808082b2b082b2b, 0x0808082b2b191908,
+    0x0808082b2b2b082b, 0x0808190808080819, 0x0808190808081908, 0x080819080808192b,
+    0x0808190808082b19, 0x0808190808190808, 0x080819080819082b, 0x0808190808191919,
+    0x0808190808192b08, 0x08081908082b0819, 0x08081908082b1908, 0x08081908082b192b,
+    0x08081908082b2b19, 0x0808190819080808, 0x080819081908082b, 0x0808190819081919,
+    0x0808190819082b08, 0x0808190819082b2b, 0x0808190819190819, 0x0808190819191908,
+    0x080819081919192b, 0x0808190819192b19, 0x08081908192b0808, 0x08081908192b082b,
+    0x08081908192b1919, 0x080819082b080819, 0x080819082b081908, 0x080819082b08192b,
+    0x080819082b082b19, 0x080819082b190808, 0x080819082b191919, 0x080819082b192b08,
+    0x080819082b2b0819, 0x080819082b2b1908, 0x0808191908080808, 0x080819190808082b,
+    0x0808191908081919, 0x0808191908082b08, 0x0808191908082b2b, 0x0808191908190819,
+    0x0808191908191908, 0x080819190819192b, 0x0808191908192b19, 0x08081919082b0808,
+    0x08081919082b1919, 0x08081919082b2b08, 0x0808191919080819, 0x0808191919081908,
+    0x080819191908192b, 0x0808191919082b19, 0x0808191919190808, 0x080819191919082b,
+    0x0808191919191919, 0x0808191919192b08, 0x08081919192b0819, 0x08081919192b1908,
+    0x080819192b080808, 0x080819192b08082b, 0x080819192b081919, 0x080819192b082b08,
+    0x080819192b190819, 0x080819192b191908, 0x080819192b2b0808, 0x0808192b08080819,
+    0x0808192b08081908, 0x0808192b0808192b, 0x0808192b08082b19, 0x0808192b08190808,
+    0x0808192b08191919, 0x0808192b19080808, 0x0808192b19081919, 0x0808192b19082b08,
+    0x0808192b19190819, 0x0808192b19191908, 0x0808192b192b0808, 0x0808192b2b080819,
+    0x0808192b2b081908, 0x0808192b2b190808, 0x08082b0808080808, 0x08082b080808082b,
+    0x08082b0808081919, 0x08082b0808082b08, 0x08082b0808190819, 0x08082b0808191908,
+    0x08082b080819192b, 0x08082b0808192b19, 0x08082b08082b0808, 0x08082b08082b1919,
+    0x08082b08082b2b2b, 0x08082b0819080819, 0x08082b0819081908, 0x08082b081908192b,
+    0x08082b0819082b19, 0x08082b0819190808, 0x08082b081919082b, 0x08082b0819191919,
+    0x08082b0819192b08, 0x08082b08192b0819, 0x08082b08192b1908, 0x08082b082b080808,
+    0x08082b082b081919, 0x08082b082b191908, 0x08082b082b2b2b2b, 0x08082b1908080819,
+    0x08082b1908081908, 0x08082b1908190808, 0x08082b190819082b, 0x08082b1908191919,
+    0x08082b1908192b08, 0x08082b19082b0819, 0x08082b1919080808, 0x08082b1919081919,
+    0x08082b1919082b08, 0x08082b1919190819, 0x08082b1919191908, 0x08082b19192b0808,
+    0x08082b192b080819, 0x08082b192b190808, 0x08082b2b08080808, 0x08082b2b08190819,
+    0x08082b2b08191908, 0x08082b2b082b082b, 0x08082b2b082b2b08, 0x08082b2b082b2b2b,
+    0x08082b2b19190808, 0x08082b2b2b192b19, 0x0819080808080819, 0x0819080808081908,
+    0x081908080808192b, 0x0819080808082b19, 0x0819080808190808, 0x081908080819082b,
+    0x0819080808191919, 0x0819080808192b08, 0x08190808082b0819, 0x08190808082b1908,
+    0x08190808082b192b, 0x0819080819080808, 0x081908081908082b, 0x0819080819081919,
+    0x0819080819082b08, 0x0819080819190819, 0x0819080819191908, 0x081908081919192b,
+    0x0819080819192b19, 0x08190808192b0808, 0x08190808192b082b, 0x08190808192b1919,
+    0x08190808192b2b08, 0x081908082b080819, 0x081908082b081908, 0x081908082b08192b,
+    0x081908082b190808, 0x081908082b191919, 0x081908082b192b08, 0x081908082b2b0819,
+    0x081908082b2b1908, 0x0819081908080808, 0x081908190808082b, 0x0819081908081919,
+    0x0819081908082b08, 0x0819081908082b2b, 0x0819081908190819, 0x0819081908191908,
+    0x081908190819192b, 0x0819081908192b19, 0x08190819082b0808, 0x08190819082b082b,
+    0x08190819082b1919, 0x08190819082b2b08, 0x0819081919080819, 0x0819081919081908,
+    0x081908191908192b, 0x0819081919082b19, 0x0819081919190808, 0x081908191919082b,
+    0x0819081919191919, 0x0819081919192b08, 0x08190819192b0819, 0x08190819192b1908,
+    0x081908192b080808, 0x081908192b08082b, 0x081908192b081919, 0x081908192b082b08,
+    0x081908192b190819, 0x081908192b191908, 0x0819082b08080819, 0x0819082b08081908,
+    0x0819082b08082b19, 0x0819082b08190808, 0x0819082b08191919, 0x0819082b082b0819,
+    0x0819082b082b1908, 0x0819082b19080808, 0x0819082b19081919, 0x0819082b19190819,
+    0x0819082b19191908, 0x0819082b2b080819, 0x0819082b2b081908, 0x0819082b2b190808,
+    0x0819190808080808, 0x081919080808082b, 0x0819190808081919, 0x0819190808082b08,
+    0x0819190808190819, 0x0819190808191908, 0x081919080819192b, 0x0819190808192b19,
+    0x08191908082b0808, 0x08191908082b1919, 0x08191908082b2b08, 0x0819190819080819,
+    0x0819190819081908, 0x081919081908192b, 0x0819190819082b19, 0x0819190819190808,
+    0x081919081919082b, 0x0819190819191919, 0x0819190819192b08, 0x08191908192b0819,
+    0x08191908192b1908, 0x081919082b080808, 0x081919082b08082b, 0x081919082b081919,
+    0x081919082b082b08, 0x081919082b190819, 0x081919082b191908, 0x081919082b2b0808,
+    0x0819191908080819, 0x0819191908081908, 0x081919190808192b, 0x0819191908082b19,
+    0x0819191908190808, 0x081919190819082b, 0x0819191908191919, 0x0819191908192b08,
+    0x08191919082b0819, 0x08191919082b1908, 0x0819191919080808, 0x081919191908082b,
+    0x0819191919081919, 0x0819191919082b08, 0x0819191919190819, 0x0819191919191908,
+    0x08191919192b0808, 0x081919192b080819, 0x081919192b081908, 0x081919192b190808,
+    0x0819192b08080808, 0x0819192b08081919, 0x0819192b08082b08, 0x0819192b08190819,
+    0x0819192b08191908, 0x0819192b082b0808, 0x0819192b19080819, 0x0819192b19081908,
+    0x0819192b19190808, 0x0819192b2b080808, 0x0819192b2b2b2b2b, 0x08192b0808080819,
+    0x08192b0808081908, 0x08192b080808192b, 0x08192b0808082b19, 0x08192b0808190808,
+    0x08192b0808191919, 0x08192b0808192b08, 0x08192b08082b0819, 0x08192b0819080808,
+    0x08192b081908082b, 0x08192b0819081919, 0x08192b0819082b08, 0x08192b0819190819,
+    0x08192b0819191908, 0x08192b08192b0808, 0x08192b082b080819, 0x08192b082b081908,
+    0x08192b1908080808, 0x08192b190808082b, 0x08192b1908081919, 0x08192b1908082b08,
+    0x08192b1908190819, 0x08192b1908191908, 0x08192b19082b0808, 0x08192b1919080819,
+    0x08192b1919081908, 0x08192b1919190808, 0x08192b19192b2b19, 0x08192b192b2b082b,
+    0x08192b2b08081908, 0x08192b2b08190808, 0x08192b2b19080808, 0x08192b2b1919192b,
+    0x082b080808080808, 0x082b08080808082b, 0x082b080808081919, 0x082b080808082b08,
+    0x082b080808190819, 0x082b080808191908, 0x082b08080819192b, 0x082b080808192b19,
+    0x082b0808082b0808, 0x082b0808082b1919, 0x082b0808082b2b2b, 0x082b080819080819,
+    0x082b080819081908, 0x082b080819190808, 0x082b08081919082b, 0x082b080819191919,
+    0x082b0808192b1908, 0x082b08082b080808, 0x082b08082b082b2b, 0x082b08082b191908,
+    0x082b08082b2b2b2b, 0x082b081908080819, 0x082b081908081908, 0x082b081908190808,
+    0x082b08190819082b, 0x082b081908191919, 0x082b0819082b0819, 0x082b081919080808,
+    0x082b08191908082b, 0x082b081919081919, 0x082b081919190819, 0x082b081919191908,
+    0x082b0819192b0808, 0x082b08192b080819, 0x082b08192b081908, 0x082b08192b190808,
+    0x082b082b08080808, 0x082b082b08082b2b, 0x082b082b082b082b, 0x082b082b082b2b08,
+    0x082b082b082b2b2b, 0x082b082b19081908, 0x082b082b19190808, 0x082b082b2b082b08,
+    0x082b082b2b082b2b, 0x082b082b2b2b2b08, 0x082b190808080819, 0x082b190808081908,
+    0x082b19080808192b, 0x082b190808082b19, 0x082b190808190808, 0x082b190808191919,
+    0x082b190808192b08, 0x082b1908082b0819, 0x082b1908082b1908, 0x082b190819080808,
+    0x082b19081908082b, 0x082b190819081919, 0x082b190819082b08, 0x082b190819190819,
+    0x082b190819191908, 0x082b1908192b0808, 0x082b19082b080819, 0x082b19082b081908,
+    0x082b19082b190808, 0x082b191908080808, 0x082b191908081919, 0x082b191908082b08,
+    0x082b191908190819, 0x082b191908191908, 0x082b1919082b0808, 0x082b191919080819,
+    0x082b191919081908, 0x082b191919190808, 0x082b1919192b192b, 0x082b19192b080808,
+    0x082b192b08080819, 0x082b192b08081908, 0x082b192b08190808, 0x082b192b19080808,
+    0x082b192b19192b19, 0x082b2b0808080808, 0x082b2b0808081919, 0x082b2b0808190819,
+    0x082b2b0808191908, 0x082b2b0819080819, 0x082b2b0819081908, 0x082b2b0819190808,
+    0x082b2b082b082b2b, 0x082b2b082b2b2b2b, 0x082b2b1908080819, 0x082b2b1908081908,
+    0x082b2b1908190808, 0x082b2b192b191919, 0x082b2b2b08082b2b, 0x082b2b2b082b082b,
+    0x082b2b2b192b1908, 0x082b2b2b2b082b08, 0x082b2b2b2b082b2b, 0x1908080808080819,
+    0x1908080808081908, 0x190808080808192b, 0x1908080808082b19, 0x1908080808190808,
+    0x190808080819082b, 0x1908080808191919, 0x1908080808192b08, 0x1908080808192b2b,
+    0x19080808082b0819, 0x19080808082b1908, 0x19080808082b192b, 0x1908080819080808,
+    0x190808081908082b, 0x1908080819081919, 0x1908080819082b08, 0x1908080819082b2b,
+    0x1908080819190819, 0x1908080819191908, 0x190808081919192b, 0x1908080819192b19,
+    0x19080808192b0808, 0x19080808192b082b, 0x19080808192b1919, 0x190808082b080819,
+    0x190808082b081908, 0x190808082b190808, 0x190808082b191919, 0x190808082b192b08,
+    0x190808082b2b0819, 0x190808082b2b1908, 0x1908081908080808, 0x190808190808082b,
+    0x1908081908081919, 0x1908081908082b08, 0x1908081908190819, 0x1908081908191908,
+    0x190808190819192b, 0x1908081908192b19, 0x19080819082b0808, 0x19080819082b082b,
+    0x19080819082b1919, 0x1908081919080819, 0x1908081919081908, 0x190808191908192b,
+    0x1908081919082b19, 0x1908081919190808, 0x190808191919082b, 0x1908081919191919,
+    0x1908081919192b08, 0x19080819192b0819, 0x19080819192b1908, 0x190808192b080808,
+    0x190808192b08082b, 0x190808192b081919, 0x190808192b082b08, 0x190808192b190819,
+    0x190808192b191908, 0x190808192b2b0808, 0x1908082b08080819, 0x1908082b08081908,
+    0x1908082b08190808, 0x1908082b0819082b, 0x1908082b08191919, 0x1908082b08192b08,
+    0x1908082b082b1908, 0x1908082b19080808, 0x1908082b19081919, 0x1908082b19082b08,
+    0x1908082b19190819, 0x1908082b19191908, 0x1908082b192b0808, 0x1908082b2b080819,
+    0x1908082b2b081908, 0x1908190808080808, 0x190819080808082b, 0x1908190808081919,
+    0x1908190808082b08, 0x1908190808082b2b, 0x1908190808190819, 0x1908190808191908,
+    0x190819080819192b, 0x1908190808192b19, 0x19081908082b0808, 0x19081908082b082b,
+    0x19081908082b1919, 0x19081908082b2b08, 0x1908190819080819, 0x1908190819081908,
+    0x190819081908192b, 0x1908190819082b19, 0x1908190819190808, 0x190819081919082b,
+    0x1908190819191919, 0x1908190819192b08, 0x19081908192b0819, 0x19081908192b1908,
+    0x190819082b080808, 0x190819082b08082b, 0x190819082b081919, 0x190819082b082b08,
+    0x190819082b190819, 0x190819082b191908, 0x190819082b2b0808, 0x1908191908080819,
+    0x1908191908081908, 0x190819190808192b, 0x1908191908082b19, 0x1908191908190808,
+    0x190819190819082b, 0x1908191908191919, 0x1908191908192b08, 0x19081919082b0819,
+    0x19081919082b1908, 0x1908191919080808, 0x190819191908082b, 0x1908191919081919,
+    0x1908191919082b08, 0x1908191919190819, 0x1908191919191908, 0x19081919192b0808,
+    0x19081919192b2b2b, 0x190819192b080819, 0x190819192b081908, 0x190819192b190808,
+    0x1908192b08080808, 0x1908192b0808082b, 0x1908192b08081919, 0x1908192b08082b08,
+    0x1908192b08190819, 0x1908192b08191908, 0x1908192b082b0808, 0x1908192b19080819,
+    0x1908192b19081908, 0x1908192b19190808, 0x1908192b2b080808, 0x1908192b2b2b1919,
+    0x19082b0808080819, 0x19082b0808081908, 0x19082b0808082b19, 0x19082b0808190808,
+    0x19082b080819082b, 0x19082b0808191919, 0x19082b0808192b08, 0x19082b08082b0819,
+    0x19082b08082b1908, 0x19082b0819080808, 0x19082b081908082b, 0x19082b0819081919,
+    0x19082b0819082b08, 0x19082b0819190819, 0x19082b0819191908, 0x19082b08192b0808,
+    0x19082b082b081908, 0x19082b082b190808, 0x19082b1908080808, 0x19082b190808082b,
+    0x19082b1908081919, 0x19082b1908082b08, 0x19082b1908190819, 0x19082b1908191908,
+    0x19082b19082b0808, 0x19082b1919080819, 0x19082b1919081908, 0x19082b1919190808,
+    0x19082b192b080808, 0x19082b192b19192b, 0x19082b2b08080819, 0x19082b2b08081908,
+    0x19082b2b08190808, 0x19082b2b19080808, 0x1919080808080808, 0x191908080808082b,
+    0x1919080808081919, 0x1919080808082b08, 0x1919080808190819, 0x1919080808191908,
+    0x191908080819192b, 0x1919080808192b19, 0x19190808082b0808, 0x19190808082b082b,
+    0x19190808082b1919, 0x19190808082b2b08, 0x1919080819080819, 0x1919080819081908,
+    0x191908081908192b, 0x1919080819082b19, 0x1919080819190808, 0x191908081919082b,
+    0x1919080819191919, 0x1919080819192b08, 0x19190808192b0819, 0x19190808192b1908,
+    0x191908082b080808, 0x191908082b08082b, 0x191908082b081919, 0x191908082b082b08,
+    0x191908082b190819, 0x191908082b191908, 0x1919081908080819, 0x1919081908081908,
+    0x191908190808192b, 0x1919081908082b19, 0x1919081908190808, 0x191908190819082b,
+    0x1919081908191919, 0x1919081908192b08, 0x19190819082b0819, 0x19190819082b1908,
+    0x1919081919080808, 0x191908191908082b, 0x1919081919081919, 0x1919081919082b08,
+    0x1919081919190819, 0x1919081919191908, 0x19190819192b0808, 0x191908192b080819,
+    0x191908192b081908, 0x191908192b190808, 0x1919082b08080808, 0x1919082b08081919,
+    0x1919082b08082b08, 0x1919082b08190819, 0x1919082b08191908, 0x1919082b082b0808,
+    0x1919082b19080819, 0x1919082b19081908, 0x1919082b19190808, 0x1919082b192b2b19,
+    0x1919082b2b080808, 0x1919190808080819, 0x1919190808081908, 0x191919080808192b,
+    0x1919190808082b19, 0x1919190808190808, 0x191919080819082b, 0x1919190808191919,
+    0x1919190808192b08, 0x19191908082b0819, 0x19191908082b1908, 0x1919190819080808,
+    0x191919081908082b, 0x1919190819081919, 0x1919190819082b08, 0x1919190819190819,
+    0x1919190819191908, 0x19191908192b0808, 0x191919082b080819, 0x191919082b081908,
+    0x191919082b190808, 0x1919191908080808, 0x191919190808082b, 0x1919191908081919,
+    0x1919191908082b08, 0x1919191908190819, 0x1919191908191908, 0x19191919082b0808,
+    0x1919191919080819, 0x1919191919081908, 0x1919191919190808, 0x191919192b080808,
+    0x1919192b08080819, 0x1919192b08081908, 0x1919192b08190808, 0x1919192b082b192b,
+    0x1919192b19080808, 0x19192b0808080808, 0x19192b080808082b, 0x19192b0808081919,
+    0x19192b0808082b08, 0x19192b0808190819, 0x19192b0808191908, 0x19192b08082b0808,
+    0x19192b0819080819, 0x19192b0819081908, 0x19192b0819190808, 0x19192b0819192b2b,
+    0x19192b082b080808, 0x19192b1908080819, 0x19192b1908081908, 0x19192b1908190808,
+    0x19192b1919080808, 0x19192b2b08080808, 0x19192b2b08192b19, 0x19192b2b2b081919,
+    0x19192b2b2b2b2b08, 0x192b080808080819, 0x192b080808081908, 0x192b08080808192b,
+    0x192b080808190808, 0x192b08080819082b, 0x192b080808191919, 0x192b080808192b08,
+    0x192b0808082b0819, 0x192b0808082b1908, 0x192b080819080808, 0x192b080819081919,
+    0x192b080819082b08, 0x192b080819190819, 0x192b080819191908, 0x192b0808192b0808,
+    0x192b08082b081908, 0x192b08082b190808, 0x192b081908080808, 0x192b08190808082b,
+    0x192b081908081919, 0x192b081908082b08, 0x192b081908190819, 0x192b081908191908,
+    0x192b0819082b0808, 0x192b081919080819, 0x192b081919081908, 0x192b081919190808,
+    0x192b08192b080808, 0x192b08192b192b19, 0x192b082b08081908, 0x192b082b08190808,
+    0x192b082b19080808, 0x192b082b1919192b, 0x192b082b2b2b0819, 0x192b190808080808,
+    0x192b190808081919, 0x192b190808082b08, 0x192b190808190819, 0x192b190808191908,
+    0x192b1908082b0808, 0x192b190819080819, 0x192b190819081908, 0x192b190819190808,
+    0x192b19082b080808, 0x192b191908080819, 0x192b191908081908, 0x192b191908190808,
+    0x192b191919080808, 0x192b191919082b2b, 0x192b1919192b2b08, 0x192b19192b19082b,
+    0x192b192b08080808, 0x192b192b2b191908, 0x192b2b0808080819, 0x192b2b0808081908,
+    0x192b2b0808190808, 0x192b2b08192b1919, 0x192b2b082b192b08, 0x192b2b1908080808,
+    0x192b2b19082b2b2b, 0x192b2b2b1908082b, 0x192b2b2b2b2b0819, 0x2b08080808080808,
+    0x2b0808080808082b, 0x2b08080808081919, 0x2b08080808082b08, 0x2b08080808190819,
+    0x2b08080808191908, 0x2b08080808192b19, 0x2b080808082b0808, 0x2b080808082b1919,
+    0x2b08080819080819, 0x2b08080819081908, 0x2b08080819190808, 0x2b0808081919082b,
+    0x2b08080819191919, 0x2b08080819192b08, 0x2b080808192b0819, 0x2b0808082b080808,
+    0x2b0808082b081919, 0x2b0808082b190819, 0x2b0808082b191908, 0x2b08081908080819,
+    0x2b08081908081908, 0x2b08081908082b19, 0x2b08081908190808, 0x2b0808190819082b,
+    0x2b08081908191919, 0x2b08081908192b08, 0x2b080819082b0819, 0x2b080819082b1908,
+    0x2b08081919080808, 0x2b0808191908082b, 0x2b08081919081919, 0x2b08081919082b08,
+    0x2b08081919190819, 0x2b08081919191908, 0x2b0808192b080819, 0x2b0808192b081908,
+    0x2b0808192b190808, 0x2b0808192b2b2b19, 0x2b08082b08080808, 0x2b08082b08081919,
+    0x2b08082b08082b2b, 0x2b08082b08190819, 0x2b08082b08191908, 0x2b08082b19080819,
+    0x2b08082b19081908, 0x2b08082b19190808, 0x2b08190808080819, 0x2b08190808081908,
+    0x2b0819080808192b, 0x2b08190808082b19, 0x2b08190808190808, 0x2b0819080819082b,
+    0x2b08190808191919, 0x2b08190808192b08, 0x2b081908082b0819, 0x2b08190819080808,
+    0x2b0819081908082b, 0x2b08190819081919, 0x2b08190819082b08, 0x2b08190819190819,
+    0x2b08190819191908, 0x2b081908192b0808, 0x2b0819082b080819, 0x2b0819082b081908,
+    0x2b0819082b190808, 0x2b08191908080808, 0x2b0819190808082b, 0x2b08191908081919,
+    0x2b08191908082b08, 0x2b08191908190819, 0x2b08191908191908, 0x2b081919082b0808,
+    0x2b08191919080819, 0x2b08191919081908, 0x2b08191919190808, 0x2b0819192b080808,
+    0x2b0819192b082b2b, 0x2b08192b08080819, 0x2b08192b08081908, 0x2b08192b08190808,
+    0x2b08192b082b2b19, 0x2b08192b19080808, 0x2b082b0808080808, 0x2b082b0808081919,
+    0x2b082b0808190819, 0x2b082b0808191908, 0x2b082b0819080819, 0x2b082b0819081908,
+    0x2b082b0819190808, 0x2b082b082b2b082b, 0x2b082b1908080819, 0x2b082b1908081908,
+    0x2b082b1919080808, 0x2b082b19192b1919, 0x2b082b2b082b082b, 0x2b082b2b19192b08,
+    0x2b082b2b19192b2b, 0x2b082b2b2b08082b, 0x2b082b2b2b2b082b, 0x2b19080808080819,
+    0x2b19080808081908, 0x2b19080808082b19, 0x2b19080808190808, 0x2b1908080819082b,
+    0x2b19080808191919, 0x2b19080808192b08, 0x2b190808082b1908, 0x2b19080819080808,
+    0x2b1908081908082b, 0x2b19080819081919, 0x2b19080819082b08, 0x2b19080819190819,
+    0x2b19080819191908, 0x2b190808192b0808, 0x2b1908082b080819, 0x2b1908082b081908,
+    0x2b1908082b190808, 0x2b19081908080808, 0x2b19081908081919, 0x2b19081908190819,
+    0x2b19081908191908, 0x2b19081919080819, 0x2b19081919081908, 0x2b19081919190808,
+    0x2b19081919192b2b, 0x2b19082b08080819, 0x2b19082b08081908, 0x2b19082b08190808,
+    0x2b19082b19080808, 0x2b19082b2b2b192b, 0x2b19190808080808, 0x2b1919080808082b,
+    0x2b19190808081919, 0x2b19190808082b08, 0x2b19190808190819, 0x2b19190808191908,
+    0x2b191908082b0808, 0x2b19190819080819, 0x2b19190819081908, 0x2b19190819190808,
+    0x2b1919082b080808, 0x2b1919082b19192b, 0x2b19191908080819, 0x2b19191908081908,
+    0x2b19191908190808, 0x2b19191919080808, 0x2b1919192b192b08, 0x2b1919192b2b0819,
+    0x2b19192b08080808, 0x2b19192b1908192b, 0x2b19192b192b1908, 0x2b192b0808080819,
+    0x2b192b0808081908, 0x2b192b0808190808, 0x2b192b08082b192b, 0x2b192b0819080808,
+    0x2b192b082b2b2b19, 0x2b192b1908080808, 0x2b192b1919082b19, 0x2b192b191919082b,
+    0x2b192b2b2b190808, 0x2b2b080808080808, 0x2b2b080808081919, 0x2b2b080808082b2b,
+    0x2b2b080808191908, 0x2b2b0808082b082b, 0x2b2b0808082b2b2b, 0x2b2b080819080819,
+    0x2b2b080819081908, 0x2b2b080819190808, 0x2b2b08082b2b082b, 0x2b2b08082b2b2b2b,
+    0x2b2b081919080808, 0x2b2b0819192b1919, 0x2b2b082b0808082b, 0x2b2b082b08082b2b,
+    0x2b2b082b082b082b, 0x2b2b082b082b2b08, 0x2b2b082b082b2b2b, 0x2b2b082b2b08082b,
+    0x2b2b082b2b082b08, 0x2b2b082b2b082b2b, 0x2b2b082b2b2b2b08, 0x2b2b190808080819,
+    0x2b2b190808081908, 0x2b2b190808190808, 0x2b2b190819080808, 0x2b2b19082b082b19,
+    0x2b2b19082b2b1908, 0x2b2b191908080808, 0x2b2b191908192b19, 0x2b2b192b19190819,
+    0x2b2b2b0808082b2b, 0x2b2b2b08082b2b08, 0x2b2b2b082b2b082b, 0x2b2b2b1919191908,
+    0x2b2b2b192b08192b, 0x2b2b2b2b08082b08, 0x2b2b2b2b08082b2b, 0x2b2b2b2b082b0808,
+    0x2b2b2b2b082b082b, 0x2b2b2b2b082b2b08, 0x2b2b2b2b2b082b08, 0x2b2b2b2b2b2b2b2b,
+GGML_TABLE_END()
+
+GGML_TABLE_BEGIN(uint32_t, iq3xxs_grid, 256)
+    0x04040404, 0x04040414, 0x04040424, 0x04040c0c, 0x04040c1c, 0x04040c3e, 0x04041404, 0x04041414,
+    0x04041c0c, 0x04042414, 0x04043e1c, 0x04043e2c, 0x040c040c, 0x040c041c, 0x040c0c04, 0x040c0c14,
+    0x040c140c, 0x040c142c, 0x040c1c04, 0x040c1c14, 0x040c240c, 0x040c2c24, 0x040c3e04, 0x04140404,
+    0x04140414, 0x04140424, 0x04140c0c, 0x04141404, 0x04141414, 0x04141c0c, 0x04141c1c, 0x04141c3e,
+    0x04142c0c, 0x04142c3e, 0x04143e2c, 0x041c040c, 0x041c043e, 0x041c0c04, 0x041c0c14, 0x041c142c,
+    0x041c3e04, 0x04240c1c, 0x04241c3e, 0x04242424, 0x04242c3e, 0x04243e1c, 0x04243e2c, 0x042c040c,
+    0x042c043e, 0x042c1c14, 0x042c2c14, 0x04341c2c, 0x04343424, 0x043e0c04, 0x043e0c24, 0x043e0c34,
+    0x043e241c, 0x043e340c, 0x0c04040c, 0x0c04041c, 0x0c040c04, 0x0c040c14, 0x0c04140c, 0x0c04141c,
+    0x0c041c04, 0x0c041c14, 0x0c041c24, 0x0c04243e, 0x0c042c04, 0x0c0c0404, 0x0c0c0414, 0x0c0c0c0c,
+    0x0c0c1404, 0x0c0c1414, 0x0c14040c, 0x0c14041c, 0x0c140c04, 0x0c140c14, 0x0c14140c, 0x0c141c04,
+    0x0c143e14, 0x0c1c0404, 0x0c1c0414, 0x0c1c1404, 0x0c1c1c0c, 0x0c1c2434, 0x0c1c3434, 0x0c24040c,
+    0x0c24042c, 0x0c242c04, 0x0c2c1404, 0x0c2c1424, 0x0c2c2434, 0x0c2c3e0c, 0x0c34042c, 0x0c3e1414,
+    0x0c3e2404, 0x14040404, 0x14040414, 0x14040c0c, 0x14040c1c, 0x14041404, 0x14041414, 0x14041434,
+    0x14041c0c, 0x14042414, 0x140c040c, 0x140c041c, 0x140c042c, 0x140c0c04, 0x140c0c14, 0x140c140c,
+    0x140c1c04, 0x140c341c, 0x140c343e, 0x140c3e04, 0x14140404, 0x14140414, 0x14140c0c, 0x14140c3e,
+    0x14141404, 0x14141414, 0x14141c3e, 0x14142404, 0x14142c2c, 0x141c040c, 0x141c0c04, 0x141c0c24,
+    0x141c3e04, 0x141c3e24, 0x14241c2c, 0x14242c1c, 0x142c041c, 0x142c143e, 0x142c240c, 0x142c3e24,
+    0x143e040c, 0x143e041c, 0x143e0c34, 0x143e242c, 0x1c04040c, 0x1c040c04, 0x1c040c14, 0x1c04140c,
+    0x1c04141c, 0x1c042c04, 0x1c04342c, 0x1c043e14, 0x1c0c0404, 0x1c0c0414, 0x1c0c1404, 0x1c0c1c0c,
+    0x1c0c2424, 0x1c0c2434, 0x1c14040c, 0x1c14041c, 0x1c140c04, 0x1c14142c, 0x1c142c14, 0x1c143e14,
+    0x1c1c0c0c, 0x1c1c1c1c, 0x1c241c04, 0x1c24243e, 0x1c243e14, 0x1c2c0404, 0x1c2c0434, 0x1c2c1414,
+    0x1c2c2c2c, 0x1c340c24, 0x1c341c34, 0x1c34341c, 0x1c3e1c1c, 0x1c3e3404, 0x24040424, 0x24040c3e,
+    0x24041c2c, 0x24041c3e, 0x24042c1c, 0x24042c3e, 0x240c3e24, 0x24141404, 0x24141c3e, 0x24142404,
+    0x24143404, 0x24143434, 0x241c043e, 0x241c242c, 0x24240424, 0x24242c0c, 0x24243424, 0x242c142c,
+    0x242c241c, 0x242c3e04, 0x243e042c, 0x243e0c04, 0x243e0c14, 0x243e1c04, 0x2c040c14, 0x2c04240c,
+    0x2c043e04, 0x2c0c0404, 0x2c0c0434, 0x2c0c1434, 0x2c0c2c2c, 0x2c140c24, 0x2c141c14, 0x2c143e14,
+    0x2c1c0414, 0x2c1c2c1c, 0x2c240c04, 0x2c24141c, 0x2c24143e, 0x2c243e14, 0x2c2c0414, 0x2c2c1c0c,
+    0x2c342c04, 0x2c3e1424, 0x2c3e2414, 0x34041424, 0x34042424, 0x34042434, 0x34043424, 0x340c140c,
+    0x340c340c, 0x34140c3e, 0x34143424, 0x341c1c04, 0x341c1c34, 0x34242424, 0x342c042c, 0x342c2c14,
+    0x34341c1c, 0x343e041c, 0x343e140c, 0x3e04041c, 0x3e04042c, 0x3e04043e, 0x3e040c04, 0x3e041c14,
+    0x3e042c14, 0x3e0c1434, 0x3e0c2404, 0x3e140c14, 0x3e14242c, 0x3e142c14, 0x3e1c0404, 0x3e1c0c2c,
+    0x3e1c1c1c, 0x3e1c3404, 0x3e24140c, 0x3e24240c, 0x3e2c0404, 0x3e2c0414, 0x3e2c1424, 0x3e341c04,
+GGML_TABLE_END()
+
+GGML_TABLE_BEGIN(uint32_t, iq3s_grid, 512)
+    0x01010101, 0x01010103, 0x01010105, 0x0101010b, 0x0101010f, 0x01010301, 0x01010303, 0x01010305,
+    0x01010309, 0x0101030d, 0x01010501, 0x01010503, 0x0101050b, 0x01010707, 0x01010901, 0x01010905,
+    0x0101090b, 0x0101090f, 0x01010b03, 0x01010b07, 0x01010d01, 0x01010d05, 0x01010f03, 0x01010f09,
+    0x01010f0f, 0x01030101, 0x01030103, 0x01030105, 0x01030109, 0x01030301, 0x01030303, 0x0103030b,
+    0x01030501, 0x01030507, 0x0103050f, 0x01030703, 0x0103070b, 0x01030909, 0x01030d03, 0x01030d0b,
+    0x01030f05, 0x01050101, 0x01050103, 0x0105010b, 0x0105010f, 0x01050301, 0x01050307, 0x0105030d,
+    0x01050503, 0x0105050b, 0x01050701, 0x01050709, 0x01050905, 0x0105090b, 0x0105090f, 0x01050b03,
+    0x01050b07, 0x01050f01, 0x01050f07, 0x01070107, 0x01070303, 0x0107030b, 0x01070501, 0x01070505,
+    0x01070703, 0x01070707, 0x0107070d, 0x01070909, 0x01070b01, 0x01070b05, 0x01070d0f, 0x01070f03,
+    0x01070f0b, 0x01090101, 0x01090307, 0x0109030f, 0x01090503, 0x01090509, 0x01090705, 0x01090901,
+    0x01090907, 0x01090b03, 0x01090f01, 0x010b0105, 0x010b0109, 0x010b0501, 0x010b0505, 0x010b050d,
+    0x010b0707, 0x010b0903, 0x010b090b, 0x010b090f, 0x010b0d0d, 0x010b0f07, 0x010d010d, 0x010d0303,
+    0x010d0307, 0x010d0703, 0x010d0b05, 0x010d0f03, 0x010f0101, 0x010f0105, 0x010f0109, 0x010f0501,
+    0x010f0505, 0x010f050d, 0x010f0707, 0x010f0b01, 0x010f0b09, 0x03010101, 0x03010103, 0x03010105,
+    0x03010109, 0x03010301, 0x03010303, 0x03010307, 0x0301030b, 0x0301030f, 0x03010501, 0x03010505,
+    0x03010703, 0x03010709, 0x0301070d, 0x03010b09, 0x03010b0d, 0x03010d03, 0x03010f05, 0x03030101,
+    0x03030103, 0x03030107, 0x0303010d, 0x03030301, 0x03030309, 0x03030503, 0x03030701, 0x03030707,
+    0x03030903, 0x03030b01, 0x03030b05, 0x03030f01, 0x03030f0d, 0x03050101, 0x03050305, 0x0305030b,
+    0x0305030f, 0x03050501, 0x03050509, 0x03050705, 0x03050901, 0x03050907, 0x03050b0b, 0x03050d01,
+    0x03050f05, 0x03070103, 0x03070109, 0x0307010f, 0x03070301, 0x03070307, 0x03070503, 0x0307050f,
+    0x03070701, 0x03070709, 0x03070903, 0x03070d05, 0x03070f01, 0x03090107, 0x0309010b, 0x03090305,
+    0x03090309, 0x03090703, 0x03090707, 0x03090905, 0x0309090d, 0x03090b01, 0x03090b09, 0x030b0103,
+    0x030b0301, 0x030b0307, 0x030b0503, 0x030b0701, 0x030b0705, 0x030b0b03, 0x030d0501, 0x030d0509,
+    0x030d050f, 0x030d0909, 0x030d090d, 0x030f0103, 0x030f0107, 0x030f0301, 0x030f0305, 0x030f0503,
+    0x030f070b, 0x030f0903, 0x030f0d05, 0x030f0f01, 0x05010101, 0x05010103, 0x05010107, 0x0501010b,
+    0x0501010f, 0x05010301, 0x05010305, 0x05010309, 0x0501030d, 0x05010503, 0x05010507, 0x0501050f,
+    0x05010701, 0x05010705, 0x05010903, 0x05010907, 0x0501090b, 0x05010b01, 0x05010b05, 0x05010d0f,
+    0x05010f01, 0x05010f07, 0x05010f0b, 0x05030101, 0x05030105, 0x05030301, 0x05030307, 0x0503030f,
+    0x05030505, 0x0503050b, 0x05030703, 0x05030709, 0x05030905, 0x05030b03, 0x05050103, 0x05050109,
+    0x0505010f, 0x05050503, 0x05050507, 0x05050701, 0x0505070f, 0x05050903, 0x05050b07, 0x05050b0f,
+    0x05050f03, 0x05050f09, 0x05070101, 0x05070105, 0x0507010b, 0x05070303, 0x05070505, 0x05070509,
+    0x05070703, 0x05070707, 0x05070905, 0x05070b01, 0x05070d0d, 0x05090103, 0x0509010f, 0x05090501,
+    0x05090507, 0x05090705, 0x0509070b, 0x05090903, 0x05090f05, 0x05090f0b, 0x050b0109, 0x050b0303,
+    0x050b0505, 0x050b070f, 0x050b0901, 0x050b0b07, 0x050b0f01, 0x050d0101, 0x050d0105, 0x050d010f,
+    0x050d0503, 0x050d0b0b, 0x050d0d03, 0x050f010b, 0x050f0303, 0x050f050d, 0x050f0701, 0x050f0907,
+    0x050f0b01, 0x07010105, 0x07010303, 0x07010307, 0x0701030b, 0x0701030f, 0x07010505, 0x07010703,
+    0x07010707, 0x0701070b, 0x07010905, 0x07010909, 0x0701090f, 0x07010b03, 0x07010d07, 0x07010f03,
+    0x07030103, 0x07030107, 0x0703010b, 0x07030309, 0x07030503, 0x07030507, 0x07030901, 0x07030d01,
+    0x07030f05, 0x07030f0d, 0x07050101, 0x07050305, 0x07050501, 0x07050705, 0x07050709, 0x07050b01,
+    0x07070103, 0x07070301, 0x07070309, 0x07070503, 0x07070507, 0x0707050f, 0x07070701, 0x07070903,
+    0x07070907, 0x0707090f, 0x07070b0b, 0x07070f07, 0x07090107, 0x07090303, 0x0709030d, 0x07090505,
+    0x07090703, 0x07090b05, 0x07090d01, 0x07090d09, 0x070b0103, 0x070b0301, 0x070b0305, 0x070b050b,
+    0x070b0705, 0x070b0909, 0x070b0b0d, 0x070b0f07, 0x070d030d, 0x070d0903, 0x070f0103, 0x070f0107,
+    0x070f0501, 0x070f0505, 0x070f070b, 0x09010101, 0x09010109, 0x09010305, 0x09010501, 0x09010509,
+    0x0901050f, 0x09010705, 0x09010903, 0x09010b01, 0x09010f01, 0x09030105, 0x0903010f, 0x09030303,
+    0x09030307, 0x09030505, 0x09030701, 0x0903070b, 0x09030907, 0x09030b03, 0x09030b0b, 0x09050103,
+    0x09050107, 0x09050301, 0x0905030b, 0x09050503, 0x09050707, 0x09050901, 0x09050b0f, 0x09050d05,
+    0x09050f01, 0x09070109, 0x09070303, 0x09070307, 0x09070501, 0x09070505, 0x09070703, 0x0907070b,
+    0x09090101, 0x09090105, 0x09090509, 0x0909070f, 0x09090901, 0x09090f03, 0x090b010b, 0x090b010f,
+    0x090b0503, 0x090b0d05, 0x090d0307, 0x090d0709, 0x090d0d01, 0x090f0301, 0x090f030b, 0x090f0701,
+    0x090f0907, 0x090f0b03, 0x0b010105, 0x0b010301, 0x0b010309, 0x0b010505, 0x0b010901, 0x0b010909,
+    0x0b01090f, 0x0b010b05, 0x0b010d0d, 0x0b010f09, 0x0b030103, 0x0b030107, 0x0b03010b, 0x0b030305,
+    0x0b030503, 0x0b030705, 0x0b030f05, 0x0b050101, 0x0b050303, 0x0b050507, 0x0b050701, 0x0b05070d,
+    0x0b050b07, 0x0b070105, 0x0b07010f, 0x0b070301, 0x0b07050f, 0x0b070909, 0x0b070b03, 0x0b070d0b,
+    0x0b070f07, 0x0b090103, 0x0b090109, 0x0b090501, 0x0b090705, 0x0b09090d, 0x0b0b0305, 0x0b0b050d,
+    0x0b0b0b03, 0x0b0b0b07, 0x0b0d0905, 0x0b0f0105, 0x0b0f0109, 0x0b0f0505, 0x0d010303, 0x0d010307,
+    0x0d01030b, 0x0d010703, 0x0d010707, 0x0d010d01, 0x0d030101, 0x0d030501, 0x0d03050f, 0x0d030d09,
+    0x0d050305, 0x0d050709, 0x0d050905, 0x0d050b0b, 0x0d050d05, 0x0d050f01, 0x0d070101, 0x0d070309,
+    0x0d070503, 0x0d070901, 0x0d09050b, 0x0d090907, 0x0d090d05, 0x0d0b0101, 0x0d0b0107, 0x0d0b0709,
+    0x0d0b0d01, 0x0d0d010b, 0x0d0d0901, 0x0d0f0303, 0x0d0f0307, 0x0f010101, 0x0f010109, 0x0f01010f,
+    0x0f010501, 0x0f010505, 0x0f01070d, 0x0f010901, 0x0f010b09, 0x0f010d05, 0x0f030105, 0x0f030303,
+    0x0f030509, 0x0f030907, 0x0f03090b, 0x0f050103, 0x0f050109, 0x0f050301, 0x0f05030d, 0x0f050503,
+    0x0f050701, 0x0f050b03, 0x0f070105, 0x0f070705, 0x0f07070b, 0x0f070b07, 0x0f090103, 0x0f09010b,
+    0x0f090307, 0x0f090501, 0x0f090b01, 0x0f0b0505, 0x0f0b0905, 0x0f0d0105, 0x0f0d0703, 0x0f0f0101,
+GGML_TABLE_END()
+
+#define NGRID_IQ1S 2048
+#define IQ1S_DELTA 0.125f
+#define IQ1M_DELTA 0.125f
+#if defined(GGML_COMMON_IMPL_C)
+GGML_TABLE_BEGIN(uint64_t, iq1s_grid, NGRID_IQ1S)
+    0xffffffffffffffff, 0xffffffffffffff01, 0xffffffffffff0000, 0xffffffffffff01ff,
+    0xffffffffffff0101, 0xffffffffff00ff00, 0xffffffffff000000, 0xffffffffff01ffff,
+    0xffffffffff01ff01, 0xffffffffff0101ff, 0xffffffffff010101, 0xffffffff00ff0000,
+    0xffffffff0000ff00, 0xffffffff000000ff, 0xffffffff00000001, 0xffffffff00010000,
+    0xffffffff01ffffff, 0xffffffff01ffff01, 0xffffffff01ff01ff, 0xffffffff01ff0101,
+    0xffffffff01000000, 0xffffffff0101ffff, 0xffffffff0101ff01, 0xffffffff010101ff,
+    0xffffffff01010101, 0xffffff00ffff00ff, 0xffffff00ffff0000, 0xffffff00ff00ff00,
+    0xffffff00ff0000ff, 0xffffff00ff000001, 0xffffff00ff000100, 0xffffff00ff000101,
+    0xffffff00ff010000, 0xffffff0000ffff00, 0xffffff0000ff0001, 0xffffff0000ff0100,
+    0xffffff000000ff01, 0xffffff0000000000, 0xffffff0000000101, 0xffffff000001ff00,
+    0xffffff00000100ff, 0xffffff0000010001, 0xffffff00000101ff, 0xffffff0001ff0000,
+    0xffffff000100ff00, 0xffffff00010000ff, 0xffffff0001000001, 0xffffff0001010000,
+    0xffffff01ffffffff, 0xffffff01ffffff01, 0xffffff01ffff01ff, 0xffffff01ffff0101,
+    0xffffff01ff000000, 0xffffff01ff01ffff, 0xffffff01ff01ff01, 0xffffff01ff0101ff,
+    0xffffff01ff010101, 0xffffff0100ff0000, 0xffffff010000ff00, 0xffffff0100000100,
+    0xffffff01000100ff, 0xffffff0100010100, 0xffffff0101ffffff, 0xffffff0101ffff01,
+    0xffffff0101ff01ff, 0xffffff0101ff0101, 0xffffff010100ff00, 0xffffff0101000000,
+    0xffffff0101000100, 0xffffff010101ffff, 0xffffff010101ff01, 0xffffff01010101ff,
+    0xffffff0101010101, 0xffff00ffff00ff00, 0xffff00ffff0000ff, 0xffff00ffff000001,
+    0xffff00ffff010000, 0xffff00ff00ffff00, 0xffff00ff00ff0100, 0xffff00ff00000000,
+    0xffff00ff00000101, 0xffff00ff000100ff, 0xffff00ff00010000, 0xffff00ff0100ff00,
+    0xffff00ff01000100, 0xffff00ff01010000, 0xffff0000ffffff00, 0xffff0000ffff00ff,
+    0xffff0000ffff0000, 0xffff0000ffff0001, 0xffff0000ff000000, 0xffff0000ff0001ff,
+    0xffff0000ff000101, 0xffff0000ff010100, 0xffff000000ffffff, 0xffff000000ff0000,
+    0xffff000000ff0101, 0xffff00000000ffff, 0xffff00000000ff00, 0xffff0000000000ff,
+    0xffff000000000000, 0xffff000000000001, 0xffff000000000100, 0xffff00000001ffff,
+    0xffff00000001ff01, 0xffff000000010000, 0xffff0000000101ff, 0xffff000000010101,
+    0xffff000001ffff00, 0xffff00000100ff00, 0xffff000001000000, 0xffff0000010001ff,
+    0xffff000001000101, 0xffff00000101ff00, 0xffff0000010100ff, 0xffff000001010000,
+    0xffff000001010001, 0xffff000001010100, 0xffff0001ff0000ff, 0xffff0001ff000100,
+    0xffff000100ffff00, 0xffff000100ff00ff, 0xffff00010000ffff, 0xffff00010000ff01,
+    0xffff000100000000, 0xffff0001000001ff, 0xffff00010001ffff, 0xffff00010001ff00,
+    0xffff000100010001, 0xffff000100010100, 0xffff000101ff0000, 0xffff00010100ff00,
+    0xffff0001010000ff, 0xffff000101000100, 0xffff01ffffffffff, 0xffff01ffffffff01,
+    0xffff01ffffff01ff, 0xffff01ffffff0101, 0xffff01ffff000000, 0xffff01ffff01ffff,
+    0xffff01ffff01ff01, 0xffff01ffff0101ff, 0xffff01ffff010101, 0xffff01ff00ff0000,
+    0xffff01ff0000ff00, 0xffff01ff00000001, 0xffff01ff00010000, 0xffff01ff01ffffff,
+    0xffff01ff01ffff01, 0xffff01ff01ff01ff, 0xffff01ff01ff0101, 0xffff01ff01000000,
+    0xffff01ff0101ffff, 0xffff01ff0101ff01, 0xffff01ff010101ff, 0xffff01ff01010101,
+    0xffff0100ffff0000, 0xffff0100ff00ff00, 0xffff0100ff0000ff, 0xffff0100ff000100,
+    0xffff0100ff0100ff, 0xffff0100ff010000, 0xffff010000ffff00, 0xffff01000000ffff,
+    0xffff01000000ff00, 0xffff010000000000, 0xffff01000001ff00, 0xffff0100000100ff,
+    0xffff010000010100, 0xffff01000100ff00, 0xffff0100010000ff, 0xffff010001000001,
+    0xffff010001000100, 0xffff010001010000, 0xffff0101ffffffff, 0xffff0101ffffff01,
+    0xffff0101ffff01ff, 0xffff0101ffff0101, 0xffff0101ff000000, 0xffff0101ff01ffff,
+    0xffff0101ff01ff01, 0xffff0101ff0101ff, 0xffff0101ff010101, 0xffff010100ff0000,
+    0xffff01010000ff00, 0xffff010100000100, 0xffff01010001ff00, 0xffff010100010000,
+    0xffff010101ffffff, 0xffff010101ffff01, 0xffff010101ff0000, 0xffff010101ff01ff,
+    0xffff010101ff0101, 0xffff010101000000, 0xffff01010101ffff, 0xffff01010101ff01,
+    0xffff0101010101ff, 0xffff010101010101, 0xff00ffffff00ffff, 0xff00ffffff00ff00,
+    0xff00ffffff0000ff, 0xff00ffffff000100, 0xff00ffffff0100ff, 0xff00ffffff010000,
+    0xff00ffff00ffff00, 0xff00ffff00ff00ff, 0xff00ffff0000ffff, 0xff00ffff00000000,
+    0xff00ffff000001ff, 0xff00ffff0001ff00, 0xff00ffff000100ff, 0xff00ffff00010000,
+    0xff00ffff00010100, 0xff00ffff0100ff00, 0xff00ffff010000ff, 0xff00ffff01000001,
+    0xff00ffff0101ff00, 0xff00ffff01010000, 0xff00ff00ffffff00, 0xff00ff00ffff00ff,
+    0xff00ff00ffff0001, 0xff00ff00ffff0100, 0xff00ff00ff00ffff, 0xff00ff00ff00ff01,
+    0xff00ff00ff000000, 0xff00ff00ff0001ff, 0xff00ff00ff01ff00, 0xff00ff00ff0100ff,
+    0xff00ff00ff010100, 0xff00ff0000ff0000, 0xff00ff0000ff0101, 0xff00ff000000ffff,
+    0xff00ff000000ff00, 0xff00ff000000ff01, 0xff00ff00000000ff, 0xff00ff0000000000,
+    0xff00ff0000000001, 0xff00ff0000000100, 0xff00ff000001ffff, 0xff00ff0000010000,
+    0xff00ff0001ff00ff, 0xff00ff000100ff01, 0xff00ff0001000000, 0xff00ff000101ff00,
+    0xff00ff00010100ff, 0xff00ff01ff00ff00, 0xff00ff01ff0000ff, 0xff00ff01ff000001,
+    0xff00ff01ff010000, 0xff00ff0100ffffff, 0xff00ff0100ff0001, 0xff00ff0100ff0100,
+    0xff00ff010000ff01, 0xff00ff0100000000, 0xff00ff01000001ff, 0xff00ff0100000101,
+    0xff00ff01000100ff, 0xff00ff0100010001, 0xff00ff0101ff0000, 0xff00ff010100ff00,
+    0xff00ff01010000ff, 0xff00ff0101000001, 0xff00ff0101010000, 0xff0000ffffffff00,
+    0xff0000ffffff0001, 0xff0000ffffff0100, 0xff0000ffff0000ff, 0xff0000ffff000000,
+    0xff0000ffff0001ff, 0xff0000ffff000100, 0xff0000ffff01ff00, 0xff0000ffff010001,
+    0xff0000ff00ffff00, 0xff0000ff00ff0000, 0xff0000ff00ff0001, 0xff0000ff00ff01ff,
+    0xff0000ff00ff0101, 0xff0000ff0000ff00, 0xff0000ff000000ff, 0xff0000ff00000000,
+    0xff0000ff00000001, 0xff0000ff00000100, 0xff0000ff0001ff01, 0xff0000ff00010000,
+    0xff0000ff000101ff, 0xff0000ff01ff00ff, 0xff0000ff01ff0100, 0xff0000ff0100ffff,
+    0xff0000ff010000ff, 0xff0000ff01000000, 0xff0000ff010001ff, 0xff0000ff01000100,
+    0xff0000ff01000101, 0xff0000ff0101ff00, 0xff0000ff010100ff, 0xff0000ff01010000,
+    0xff0000ff01010100, 0xff000000ffffff01, 0xff000000ffff0000, 0xff000000ffff0101,
+    0xff000000ff00ff00, 0xff000000ff0000ff, 0xff000000ff000000, 0xff000000ff000001,
+    0xff000000ff000100, 0xff000000ff01ffff, 0xff000000ff01ff01, 0xff000000ff010000,
+    0xff000000ff0101ff, 0xff000000ff010101, 0xff00000000ffff00, 0xff00000000ff00ff,
+    0xff00000000ff0000, 0xff00000000ff0001, 0xff0000000000ff00, 0xff0000000000ff01,
+    0xff000000000000ff, 0xff00000000000000, 0xff00000000000001, 0xff00000000000100,
+    0xff00000000000101, 0xff0000000001ff00, 0xff000000000100ff, 0xff00000000010000,
+    0xff00000000010001, 0xff00000000010100, 0xff00000001ffffff, 0xff00000001ffff01,
+    0xff00000001ff00ff, 0xff00000001ff0000, 0xff00000001ff01ff, 0xff00000001ff0101,
+    0xff0000000100ffff, 0xff0000000100ff00, 0xff000000010000ff, 0xff00000001000000,
+    0xff00000001000001, 0xff00000001000100, 0xff00000001000101, 0xff0000000101ffff,
+    0xff0000000101ff01, 0xff00000001010000, 0xff000001ffffff00, 0xff000001ffff00ff,
+    0xff000001ffff0000, 0xff000001ffff0001, 0xff000001ff000000, 0xff000001ff000001,
+    0xff000001ff0001ff, 0xff000001ff000101, 0xff000001ff01ff00, 0xff000001ff010001,
+    0xff00000100ffffff, 0xff00000100ffff01, 0xff00000100ff00ff, 0xff00000100ff0000,
+    0xff00000100ff01ff, 0xff00000100ff0101, 0xff0000010000ff00, 0xff00000100000000,
+    0xff00000100000001, 0xff000001000001ff, 0xff00000100000100, 0xff0000010001ff00,
+    0xff000001000100ff, 0xff00000100010000, 0xff000001000101ff, 0xff00000100010100,
+    0xff00000100010101, 0xff00000101ff0001, 0xff00000101ff0101, 0xff0000010100ff01,
+    0xff00000101000000, 0xff000001010100ff, 0xff00000101010100, 0xff0001ffff00ff00,
+    0xff0001ffff000001, 0xff0001ffff010000, 0xff0001ff00ffff00, 0xff0001ff00ff00ff,
+    0xff0001ff00ff0001, 0xff0001ff00ff0100, 0xff0001ff0000ffff, 0xff0001ff00000000,
+    0xff0001ff000001ff, 0xff0001ff00000101, 0xff0001ff0001ffff, 0xff0001ff0001ff00,
+    0xff0001ff000100ff, 0xff0001ff00010001, 0xff0001ff00010100, 0xff0001ff01ff0000,
+    0xff0001ff0100ff00, 0xff0001ff010000ff, 0xff0001ff01010000, 0xff000100ff00ffff,
+    0xff000100ff00ff01, 0xff000100ff000000, 0xff000100ff000101, 0xff000100ff01ff00,
+    0xff000100ff010000, 0xff00010000ffff01, 0xff00010000ff00ff, 0xff00010000ff0000,
+    0xff00010000ff01ff, 0xff0001000000ff00, 0xff000100000000ff, 0xff00010000000000,
+    0xff00010000000001, 0xff00010000000100, 0xff00010000000101, 0xff0001000001ffff,
+    0xff00010000010000, 0xff00010000010101, 0xff00010001ff0100, 0xff0001000100ff00,
+    0xff0001000100ff01, 0xff00010001000000, 0xff000100010001ff, 0xff0001000101ff00,
+    0xff00010001010001, 0xff00010001010100, 0xff000101ffff0100, 0xff000101ff000001,
+    0xff000101ff0100ff, 0xff000101ff010001, 0xff00010100ff00ff, 0xff00010100ff0001,
+    0xff00010100ff0100, 0xff0001010000ffff, 0xff0001010000ff01, 0xff00010100000000,
+    0xff000101000001ff, 0xff0001010001ff00, 0xff00010100010001, 0xff00010100010100,
+    0xff00010101ff0000, 0xff0001010100ff00, 0xff00010101000001, 0xff00010101000101,
+    0xff01ffffffffffff, 0xff01ffffffffff01, 0xff01ffffffff01ff, 0xff01ffffffff0101,
+    0xff01ffffff000000, 0xff01ffffff01ffff, 0xff01ffffff01ff01, 0xff01ffffff010000,
+    0xff01ffffff0101ff, 0xff01ffffff010101, 0xff01ffff00ff0000, 0xff01ffff0000ff00,
+    0xff01ffff00000100, 0xff01ffff0001ff00, 0xff01ffff00010000, 0xff01ffff01ffffff,
+    0xff01ffff01ffff01, 0xff01ffff01ff01ff, 0xff01ffff01ff0101, 0xff01ffff01000000,
+    0xff01ffff0101ffff, 0xff01ffff0101ff01, 0xff01ffff01010000, 0xff01ffff010101ff,
+    0xff01ffff01010101, 0xff01ff00ffff0000, 0xff01ff00ff00ff00, 0xff01ff00ff0000ff,
+    0xff01ff00ff000100, 0xff01ff00ff010000, 0xff01ff0000ffff01, 0xff01ff0000ff00ff,
+    0xff01ff0000ff0100, 0xff01ff0000000000, 0xff01ff00000001ff, 0xff01ff0000000101,
+    0xff01ff000001ff00, 0xff01ff00000100ff, 0xff01ff0000010000, 0xff01ff0000010001,
+    0xff01ff0001ff0000, 0xff01ff000100ffff, 0xff01ff0001000001, 0xff01ff0001000100,
+    0xff01ff0001010000, 0xff01ff01ffffff00, 0xff01ff01ffff01ff, 0xff01ff01ffff0101,
+    0xff01ff01ff00ff00, 0xff01ff01ff000000, 0xff01ff01ff01ffff, 0xff01ff01ff01ff01,
+    0xff01ff01ff0101ff, 0xff01ff01ff010101, 0xff01ff0100ff0000, 0xff01ff010000ff00,
+    0xff01ff0100000001, 0xff01ff0100000100, 0xff01ff0100010000, 0xff01ff0101ffff00,
+    0xff01ff0101ff01ff, 0xff01ff0101ff0101, 0xff01ff010100ff00, 0xff01ff0101000000,
+    0xff01ff010101ffff, 0xff01ff010101ff01, 0xff01ff01010101ff, 0xff01ff0101010101,
+    0xff0100ffffff0000, 0xff0100ffff0000ff, 0xff0100ffff000001, 0xff0100ffff000100,
+    0xff0100ffff010000, 0xff0100ff00ff00ff, 0xff0100ff00ff0000, 0xff0100ff00ff0001,
+    0xff0100ff00ff0100, 0xff0100ff0000ff01, 0xff0100ff00000000, 0xff0100ff000001ff,
+    0xff0100ff00000101, 0xff0100ff00010001, 0xff0100ff01ff0000, 0xff0100ff0100ff00,
+    0xff0100ff010000ff, 0xff0100ff01000100, 0xff0100ff0101ff00, 0xff0100ff01010000,
+    0xff010000ffff0100, 0xff010000ff000000, 0xff010000ff01ff00, 0xff010000ff010100,
+    0xff01000000ffffff, 0xff01000000ff0000, 0xff01000000ff01ff, 0xff0100000000ff00,
+    0xff010000000000ff, 0xff01000000000000, 0xff01000000000100, 0xff0100000001ff01,
+    0xff01000000010000, 0xff010000000101ff, 0xff01000001ff0100, 0xff0100000100ffff,
+    0xff010000010000ff, 0xff01000001000000, 0xff010000010001ff, 0xff01000001000101,
+    0xff0100000101ff00, 0xff010000010100ff, 0xff01000001010001, 0xff01000001010100,
+    0xff010001ffff0000, 0xff010001ff00ffff, 0xff010001ff00ff01, 0xff010001ff000100,
+    0xff010001ff010000, 0xff01000100ffff00, 0xff01000100ff0100, 0xff01000100000000,
+    0xff0100010001ffff, 0xff0100010001ff00, 0xff01000100010100, 0xff01000101ff00ff,
+    0xff01000101ff0001, 0xff0100010100ffff, 0xff01000101000101, 0xff0101ffffffffff,
+    0xff0101ffffffff01, 0xff0101ffffff01ff, 0xff0101ffffff0101, 0xff0101ffff000000,
+    0xff0101ffff01ffff, 0xff0101ffff01ff01, 0xff0101ffff0101ff, 0xff0101ffff010101,
+    0xff0101ff00ff0000, 0xff0101ff0000ff00, 0xff0101ff000000ff, 0xff0101ff00010000,
+    0xff0101ff01ffffff, 0xff0101ff01ffff01, 0xff0101ff01ff01ff, 0xff0101ff01ff0101,
+    0xff0101ff0101ffff, 0xff0101ff0101ff01, 0xff0101ff010101ff, 0xff0101ff01010101,
+    0xff010100ffff0100, 0xff010100ff00ff00, 0xff010100ff0000ff, 0xff010100ff000100,
+    0xff010100ff010000, 0xff01010000ff0001, 0xff01010000ff0100, 0xff0101000000ff01,
+    0xff01010000000000, 0xff0101000001ff00, 0xff010100000100ff, 0xff01010000010001,
+    0xff01010000010100, 0xff01010001ff0000, 0xff0101000100ffff, 0xff01010001000001,
+    0xff01010001000100, 0xff010100010100ff, 0xff01010001010000, 0xff010101ffffffff,
+    0xff010101ffffff01, 0xff010101ffff01ff, 0xff010101ffff0101, 0xff010101ff01ffff,
+    0xff010101ff01ff01, 0xff010101ff0101ff, 0xff010101ff010101, 0xff01010100ff0000,
+    0xff0101010000ff00, 0xff01010100000001, 0xff01010100000100, 0xff01010100010000,
+    0xff01010101ffffff, 0xff01010101ffff01, 0xff01010101ff01ff, 0xff01010101ff0101,
+    0xff01010101000000, 0xff0101010101ffff, 0xff0101010101ff01, 0xff010101010101ff,
+    0xff01010101010101, 0x00ffffffffff0000, 0x00ffffffff00ff00, 0x00ffffffff000001,
+    0x00ffffffff010000, 0x00ffffff00ff0100, 0x00ffffff0000ff01, 0x00ffffff00000000,
+    0x00ffffff000001ff, 0x00ffffff00000101, 0x00ffffff0001ff00, 0x00ffffff000100ff,
+    0x00ffffff00010001, 0x00ffffff010000ff, 0x00ffffff01000100, 0x00ffffff0101ff00,
+    0x00ffffff01010001, 0x00ffff00ffffffff, 0x00ffff00ffffff00, 0x00ffff00ffff00ff,
+    0x00ffff00ffff0001, 0x00ffff00ffff0100, 0x00ffff00ff00ff01, 0x00ffff00ff000000,
+    0x00ffff00ff000001, 0x00ffff00ff0001ff, 0x00ffff00ff000101, 0x00ffff00ff01ff00,
+    0x00ffff00ff010001, 0x00ffff00ff010100, 0x00ffff0000ff0000, 0x00ffff0000ff01ff,
+    0x00ffff0000ff0101, 0x00ffff000000ff00, 0x00ffff00000000ff, 0x00ffff0000000000,
+    0x00ffff0000000001, 0x00ffff0000000100, 0x00ffff0000000101, 0x00ffff0000010000,
+    0x00ffff00000101ff, 0x00ffff0000010101, 0x00ffff0001ffff00, 0x00ffff0001ff00ff,
+    0x00ffff0001ff0001, 0x00ffff000100ffff, 0x00ffff000100ff01, 0x00ffff0001000000,
+    0x00ffff000101ffff, 0x00ffff000101ff00, 0x00ffff000101ff01, 0x00ffff01ffff0000,
+    0x00ffff01ff00ff00, 0x00ffff01ff0000ff, 0x00ffff01ff000001, 0x00ffff01ff010000,
+    0x00ffff0100ffff00, 0x00ffff010000ff01, 0x00ffff0100000000, 0x00ffff0100000101,
+    0x00ffff01000100ff, 0x00ffff0100010100, 0x00ffff0101ff0100, 0x00ffff01010000ff,
+    0x00ffff0101010000, 0x00ff00ffffffff00, 0x00ff00ffff000000, 0x00ff00ffff000100,
+    0x00ff00ffff010100, 0x00ff00ff00ff0000, 0x00ff00ff00ff01ff, 0x00ff00ff00ff0101,
+    0x00ff00ff0000ff00, 0x00ff00ff000000ff, 0x00ff00ff00000000, 0x00ff00ff00000001,
+    0x00ff00ff0001ff00, 0x00ff00ff0001ff01, 0x00ff00ff00010000, 0x00ff00ff000101ff,
+    0x00ff00ff00010101, 0x00ff00ff01ffff00, 0x00ff00ff01ff0001, 0x00ff00ff01ff0100,
+    0x00ff00ff0100ffff, 0x00ff00ff0100ff01, 0x00ff00ff01000000, 0x00ff00ff0101ffff,
+    0x00ff00ff0101ff00, 0x00ff00ff01010100, 0x00ff0000ffffff00, 0x00ff0000ffffff01,
+    0x00ff0000ffff0000, 0x00ff0000ffff0101, 0x00ff0000ff00ff00, 0x00ff0000ff0000ff,
+    0x00ff0000ff000000, 0x00ff0000ff000001, 0x00ff0000ff000100, 0x00ff0000ff01ffff,
+    0x00ff0000ff010000, 0x00ff0000ff010101, 0x00ff000000ffff00, 0x00ff000000ff00ff,
+    0x00ff000000ff0000, 0x00ff000000ff0001, 0x00ff000000ff0100, 0x00ff00000000ffff,
+    0x00ff00000000ff00, 0x00ff0000000000ff, 0x00ff000000000000, 0x00ff000000000001,
+    0x00ff0000000001ff, 0x00ff000000000100, 0x00ff00000001ff00, 0x00ff0000000100ff,
+    0x00ff000000010000, 0x00ff000000010001, 0x00ff000000010100, 0x00ff000001ffff01,
+    0x00ff000001ff00ff, 0x00ff000001ff0000, 0x00ff000001ff01ff, 0x00ff00000100ff00,
+    0x00ff0000010000ff, 0x00ff000001000000, 0x00ff000001000001, 0x00ff000001000100,
+    0x00ff000001000101, 0x00ff000001010000, 0x00ff0000010101ff, 0x00ff000001010101,
+    0x00ff0001ffffff00, 0x00ff0001ffff0000, 0x00ff0001ffff0100, 0x00ff0001ff0000ff,
+    0x00ff0001ff000000, 0x00ff0001ff0001ff, 0x00ff0001ff000101, 0x00ff0001ff01ff00,
+    0x00ff0001ff0100ff, 0x00ff0001ff010100, 0x00ff000100ffffff, 0x00ff000100ffff01,
+    0x00ff000100ff0000, 0x00ff000100ff01ff, 0x00ff00010000ffff, 0x00ff00010000ff00,
+    0x00ff00010000ff01, 0x00ff000100000000, 0x00ff000100000001, 0x00ff000100000100,
+    0x00ff00010001ff01, 0x00ff000100010000, 0x00ff0001000101ff, 0x00ff000101ffff00,
+    0x00ff000101ff0000, 0x00ff000101ff0101, 0x00ff0001010000ff, 0x00ff000101000000,
+    0x00ff00010101ff00, 0x00ff0001010100ff, 0x00ff000101010001, 0x00ff01ffffff0000,
+    0x00ff01ffff00ff00, 0x00ff01ffff000000, 0x00ff01ffff000101, 0x00ff01ffff010000,
+    0x00ff01ff00ffff01, 0x00ff01ff00ff0100, 0x00ff01ff0000ffff, 0x00ff01ff00000000,
+    0x00ff01ff000001ff, 0x00ff01ff0001ff00, 0x00ff01ff000100ff, 0x00ff01ff00010001,
+    0x00ff01ff00010100, 0x00ff01ff01ff0000, 0x00ff01ff0100ff00, 0x00ff01ff010000ff,
+    0x00ff01ff01000001, 0x00ff01ff01000100, 0x00ff01ff01010000, 0x00ff0100ffffff00,
+    0x00ff0100ffff0000, 0x00ff0100ffff0001, 0x00ff0100ffff0101, 0x00ff0100ff00ffff,
+    0x00ff0100ff0000ff, 0x00ff0100ff000000, 0x00ff0100ff0001ff, 0x00ff0100ff01ff00,
+    0x00ff0100ff0100ff, 0x00ff0100ff010001, 0x00ff010000ffffff, 0x00ff010000ff0000,
+    0x00ff010000ff0101, 0x00ff01000000ff00, 0x00ff01000000ff01, 0x00ff0100000000ff,
+    0x00ff010000000000, 0x00ff010000000001, 0x00ff010000000100, 0x00ff01000001ffff,
+    0x00ff01000001ff01, 0x00ff010000010000, 0x00ff010000010001, 0x00ff010000010101,
+    0x00ff010001ff0001, 0x00ff010001ff0100, 0x00ff01000100ff01, 0x00ff010001000000,
+    0x00ff010001000001, 0x00ff0100010001ff, 0x00ff01000101ff00, 0x00ff0100010100ff,
+    0x00ff010001010001, 0x00ff010001010100, 0x00ff0101ff000001, 0x00ff010100ff00ff,
+    0x00ff010100ff0001, 0x00ff010100ff0100, 0x00ff010100000000, 0x00ff0101000001ff,
+    0x00ff010100000101, 0x00ff0101000100ff, 0x00ff010100010100, 0x00ff0101010000ff,
+    0x00ff010101010000, 0x0000ffffffffff00, 0x0000ffffffff00ff, 0x0000ffffffff0000,
+    0x0000ffffffff0001, 0x0000ffffffff0100, 0x0000ffffff00ff01, 0x0000ffffff000000,
+    0x0000ffffff000101, 0x0000ffffff01ff00, 0x0000ffffff0100ff, 0x0000ffffff010100,
+    0x0000ffff00ffffff, 0x0000ffff00ff0000, 0x0000ffff00ff01ff, 0x0000ffff0000ff00,
+    0x0000ffff000000ff, 0x0000ffff00000000, 0x0000ffff00000001, 0x0000ffff00000100,
+    0x0000ffff00010000, 0x0000ffff000101ff, 0x0000ffff01ff0001, 0x0000ffff01ff0100,
+    0x0000ffff01000000, 0x0000ffff010001ff, 0x0000ffff0101ffff, 0x0000ffff0101ff00,
+    0x0000ffff01010001, 0x0000ffff01010100, 0x0000ff00ffff0000, 0x0000ff00ffff01ff,
+    0x0000ff00ffff0100, 0x0000ff00ffff0101, 0x0000ff00ff00ff00, 0x0000ff00ff0000ff,
+    0x0000ff00ff000000, 0x0000ff00ff000001, 0x0000ff00ff0001ff, 0x0000ff00ff000100,
+    0x0000ff00ff01ffff, 0x0000ff00ff010000, 0x0000ff00ff010001, 0x0000ff00ff0101ff,
+    0x0000ff00ff010101, 0x0000ff0000ffff00, 0x0000ff0000ff00ff, 0x0000ff0000ff0000,
+    0x0000ff0000ff0001, 0x0000ff0000ff0100, 0x0000ff000000ffff, 0x0000ff000000ff00,
+    0x0000ff000000ff01, 0x0000ff00000000ff, 0x0000ff0000000000, 0x0000ff0000000001,
+    0x0000ff00000001ff, 0x0000ff0000000100, 0x0000ff0000000101, 0x0000ff000001ff00,
+    0x0000ff00000100ff, 0x0000ff0000010000, 0x0000ff0000010001, 0x0000ff0000010100,
+    0x0000ff0001ffff01, 0x0000ff0001ff0000, 0x0000ff000100ff00, 0x0000ff00010000ff,
+    0x0000ff0001000000, 0x0000ff0001000001, 0x0000ff0001000100, 0x0000ff000101ffff,
+    0x0000ff0001010000, 0x0000ff0001010101, 0x0000ff01ffffff00, 0x0000ff01ffff0001,
+    0x0000ff01ff00ff01, 0x0000ff01ff000000, 0x0000ff01ff000101, 0x0000ff01ff01ff00,
+    0x0000ff01ff0100ff, 0x0000ff0100ffff01, 0x0000ff0100ff0000, 0x0000ff0100ff0101,
+    0x0000ff010000ff00, 0x0000ff01000000ff, 0x0000ff0100000000, 0x0000ff0100000001,
+    0x0000ff0100000100, 0x0000ff010001ff01, 0x0000ff0100010000, 0x0000ff0101ff0000,
+    0x0000ff010100ffff, 0x0000ff010100ff01, 0x0000ff0101000000, 0x0000ff0101000100,
+    0x0000ff0101000101, 0x0000ff01010100ff, 0x000000ffffff00ff, 0x000000ffffff0000,
+    0x000000ffff00ff00, 0x000000ffff0000ff, 0x000000ffff000000, 0x000000ffff000001,
+    0x000000ffff0001ff, 0x000000ffff000100, 0x000000ffff01ff00, 0x000000ffff010000,
+    0x000000ffff0101ff, 0x000000ffff010101, 0x000000ff00ffff00, 0x000000ff00ff00ff,
+    0x000000ff00ff0000, 0x000000ff00ff0001, 0x000000ff00ff0100, 0x000000ff00ff0101,
+    0x000000ff0000ffff, 0x000000ff0000ff00, 0x000000ff000000ff, 0x000000ff00000000,
+    0x000000ff00000001, 0x000000ff000001ff, 0x000000ff00000100, 0x000000ff00000101,
+    0x000000ff0001ff00, 0x000000ff0001ff01, 0x000000ff000100ff, 0x000000ff00010000,
+    0x000000ff00010001, 0x000000ff00010100, 0x000000ff01ffffff, 0x000000ff01ff01ff,
+    0x000000ff01ff0101, 0x000000ff0100ff00, 0x000000ff010000ff, 0x000000ff01000000,
+    0x000000ff01000001, 0x000000ff01000100, 0x000000ff0101ff00, 0x000000ff010100ff,
+    0x000000ff01010000, 0x000000ff01010101, 0x00000000ffffff00, 0x00000000ffffff01,
+    0x00000000ffff00ff, 0x00000000ffff0000, 0x00000000ffff0001, 0x00000000ffff0100,
+    0x00000000ff00ffff, 0x00000000ff00ff00, 0x00000000ff00ff01, 0x00000000ff0000ff,
+    0x00000000ff000000, 0x00000000ff000001, 0x00000000ff000100, 0x00000000ff000101,
+    0x00000000ff01ff00, 0x00000000ff0100ff, 0x00000000ff010000, 0x00000000ff010001,
+    0x00000000ff010100, 0x0000000000ffffff, 0x0000000000ffff00, 0x0000000000ffff01,
+    0x0000000000ff00ff, 0x0000000000ff0000, 0x0000000000ff0001, 0x0000000000ff01ff,
+    0x0000000000ff0100, 0x000000000000ffff, 0x000000000000ff00, 0x000000000000ff01,
+    0x00000000000000ff, 0x0000000000000000, 0x0000000000000001, 0x00000000000001ff,
+    0x0000000000000100, 0x0000000000000101, 0x000000000001ffff, 0x000000000001ff00,
+    0x00000000000100ff, 0x0000000000010000, 0x0000000000010001, 0x00000000000101ff,
+    0x0000000000010100, 0x0000000000010101, 0x0000000001ffff00, 0x0000000001ff00ff,
+    0x0000000001ff0000, 0x0000000001ff0100, 0x0000000001ff0101, 0x000000000100ffff,
+    0x000000000100ff00, 0x00000000010000ff, 0x0000000001000000, 0x0000000001000001,
+    0x00000000010001ff, 0x0000000001000100, 0x000000000101ff00, 0x00000000010100ff,
+    0x0000000001010000, 0x0000000001010001, 0x0000000001010100, 0x00000001ffffffff,
+    0x00000001ffffff00, 0x00000001ffffff01, 0x00000001ffff00ff, 0x00000001ffff0001,
+    0x00000001ffff01ff, 0x00000001ffff0100, 0x00000001ff00ff00, 0x00000001ff0000ff,
+    0x00000001ff000000, 0x00000001ff0001ff, 0x00000001ff000100, 0x00000001ff01ffff,
+    0x00000001ff01ff00, 0x00000001ff01ff01, 0x00000001ff0100ff, 0x00000001ff010000,
+    0x00000001ff010001, 0x00000001ff0101ff, 0x00000001ff010100, 0x0000000100ffff00,
+    0x0000000100ff0000, 0x0000000100ff0001, 0x0000000100ff01ff, 0x0000000100ff0100,
+    0x0000000100ff0101, 0x000000010000ffff, 0x000000010000ff00, 0x000000010000ff01,
+    0x00000001000000ff, 0x0000000100000000, 0x0000000100000001, 0x00000001000001ff,
+    0x0000000100000100, 0x0000000100000101, 0x000000010001ff00, 0x00000001000100ff,
+    0x0000000100010000, 0x0000000100010100, 0x0000000101ffff01, 0x0000000101ff0000,
+    0x0000000101ff0001, 0x0000000101ff01ff, 0x0000000101ff0100, 0x0000000101ff0101,
+    0x000000010100ff00, 0x0000000101000000, 0x0000000101000101, 0x000000010101ff01,
+    0x0000000101010000, 0x0000000101010001, 0x00000001010101ff, 0x0000000101010100,
+    0x000001ffffff00ff, 0x000001ffffff0000, 0x000001ffffff0001, 0x000001ffffff0100,
+    0x000001ffff00ffff, 0x000001ffff000000, 0x000001ffff0001ff, 0x000001ffff01ff00,
+    0x000001ffff010101, 0x000001ff00ff0000, 0x000001ff00ff01ff, 0x000001ff00ff0101,
+    0x000001ff0000ff00, 0x000001ff000000ff, 0x000001ff00000000, 0x000001ff00000001,
+    0x000001ff000001ff, 0x000001ff00000100, 0x000001ff0001ffff, 0x000001ff0001ff01,
+    0x000001ff000100ff, 0x000001ff00010000, 0x000001ff01ffff01, 0x000001ff01ff0100,
+    0x000001ff0100ffff, 0x000001ff0100ff01, 0x000001ff01000000, 0x000001ff010001ff,
+    0x000001ff0101ff00, 0x000001ff01010100, 0x00000100ffffff00, 0x00000100ffffff01,
+    0x00000100ffff0000, 0x00000100ffff0101, 0x00000100ff00ff00, 0x00000100ff0000ff,
+    0x00000100ff000000, 0x00000100ff000001, 0x00000100ff000100, 0x00000100ff010000,
+    0x0000010000ffff00, 0x0000010000ff00ff, 0x0000010000ff0000, 0x0000010000ff0001,
+    0x0000010000ff0100, 0x000001000000ffff, 0x000001000000ff00, 0x000001000000ff01,
+    0x00000100000000ff, 0x0000010000000000, 0x0000010000000001, 0x00000100000001ff,
+    0x0000010000000100, 0x0000010000000101, 0x000001000001ff00, 0x00000100000100ff,
+    0x0000010000010000, 0x0000010000010001, 0x0000010000010100, 0x0000010001ffff00,
+    0x0000010001ff0000, 0x0000010001ff0100, 0x000001000100ff00, 0x00000100010000ff,
+    0x0000010001000000, 0x0000010001000001, 0x00000100010001ff, 0x0000010001000100,
+    0x0000010001010000, 0x00000101ffff00ff, 0x00000101ffff01ff, 0x00000101ff000000,
+    0x00000101ff000101, 0x00000101ff01ffff, 0x00000101ff010000, 0x00000101ff010001,
+    0x00000101ff010100, 0x0000010100ff0000, 0x0000010100ff01ff, 0x0000010100ff0100,
+    0x000001010000ff00, 0x0000010100000000, 0x0000010100000001, 0x00000101000001ff,
+    0x0000010100000100, 0x000001010001ff01, 0x0000010100010000, 0x00000101000101ff,
+    0x0000010100010101, 0x0000010101ffff00, 0x0000010101ff0101, 0x000001010100ff01,
+    0x0000010101000000, 0x0000010101000001, 0x00000101010001ff, 0x0000010101000101,
+    0x000001010101ff00, 0x0001ffffffff0000, 0x0001ffffff0000ff, 0x0001ffffff000001,
+    0x0001ffffff000100, 0x0001ffffff010000, 0x0001ffff00ff00ff, 0x0001ffff0000ffff,
+    0x0001ffff00000000, 0x0001ffff00000001, 0x0001ffff000001ff, 0x0001ffff00000101,
+    0x0001ffff0001ff00, 0x0001ffff000100ff, 0x0001ffff00010001, 0x0001ffff00010100,
+    0x0001ffff01ffff00, 0x0001ffff01000001, 0x0001ffff01010000, 0x0001ff00ffffff00,
+    0x0001ff00ffff00ff, 0x0001ff00ffff0001, 0x0001ff00ffff0100, 0x0001ff00ff00ff01,
+    0x0001ff00ff000000, 0x0001ff00ff01ff00, 0x0001ff00ff01ff01, 0x0001ff00ff010001,
+    0x0001ff00ff010100, 0x0001ff0000ff0000, 0x0001ff0000ff0100, 0x0001ff000000ff00,
+    0x0001ff0000000000, 0x0001ff0000000001, 0x0001ff0000000100, 0x0001ff0000010000,
+    0x0001ff0000010001, 0x0001ff0000010101, 0x0001ff0001ff00ff, 0x0001ff0001ff0101,
+    0x0001ff000100ff01, 0x0001ff0001000000, 0x0001ff000101ff00, 0x0001ff0001010001,
+    0x0001ff0001010100, 0x0001ff01ff00ff00, 0x0001ff01ff000001, 0x0001ff01ff000100,
+    0x0001ff0100ffffff, 0x0001ff0100ffff00, 0x0001ff0100ff0001, 0x0001ff0100000000,
+    0x0001ff0100000001, 0x0001ff01000001ff, 0x0001ff010001ffff, 0x0001ff0101ff0000,
+    0x0001ff010100ff00, 0x0001ff0101000001, 0x0001ff0101010000, 0x000100ffff00ff00,
+    0x000100ffff00ff01, 0x000100ffff000000, 0x000100ffff000001, 0x000100ffff000101,
+    0x000100ffff01ff00, 0x000100ffff010001, 0x000100ffff010100, 0x000100ff00ffffff,
+    0x000100ff00ffff01, 0x000100ff00ff0000, 0x000100ff00ff01ff, 0x000100ff00ff0101,
+    0x000100ff0000ff00, 0x000100ff000000ff, 0x000100ff00000000, 0x000100ff00000001,
+    0x000100ff00000100, 0x000100ff00000101, 0x000100ff0001ffff, 0x000100ff0001ff01,
+    0x000100ff00010000, 0x000100ff01ff00ff, 0x000100ff01ff0000, 0x000100ff01ff0100,
+    0x000100ff0100ffff, 0x000100ff0100ff01, 0x000100ff010000ff, 0x000100ff01000000,
+    0x000100ff01000001, 0x000100ff010001ff, 0x000100ff01000101, 0x000100ff0101ff00,
+    0x000100ff010100ff, 0x000100ff01010100, 0x00010000ffff0000, 0x00010000ffff01ff,
+    0x00010000ffff0101, 0x00010000ff00ff00, 0x00010000ff000000, 0x00010000ff000001,
+    0x00010000ff000100, 0x0001000000ff00ff, 0x0001000000ff0000, 0x0001000000ff0001,
+    0x0001000000ff0100, 0x000100000000ffff, 0x000100000000ff00, 0x00010000000000ff,
+    0x0001000000000000, 0x0001000000000001, 0x0001000000000100, 0x000100000001ff00,
+    0x00010000000100ff, 0x0001000000010000, 0x0001000000010001, 0x0001000000010100,
+    0x0001000001ff0001, 0x0001000001ff0100, 0x0001000001ff0101, 0x000100000100ff00,
+    0x0001000001000000, 0x0001000001000001, 0x0001000001000100, 0x0001000001000101,
+    0x000100000101ff01, 0x0001000001010000, 0x0001000001010001, 0x00010000010101ff,
+    0x00010001ffffff01, 0x00010001ffff0100, 0x00010001ff000000, 0x00010001ff01ffff,
+    0x00010001ff010001, 0x00010001ff0101ff, 0x00010001ff010100, 0x0001000100ffffff,
+    0x0001000100ff0000, 0x0001000100ff01ff, 0x0001000100ff0101, 0x000100010000ff00,
+    0x00010001000000ff, 0x0001000100000000, 0x0001000100000001, 0x00010001000001ff,
+    0x0001000100000101, 0x000100010001ffff, 0x0001000100010000, 0x00010001000101ff,
+    0x0001000101ffffff, 0x0001000101ffff01, 0x0001000101ff0000, 0x0001000101ff0101,
+    0x00010001010000ff, 0x0001000101000001, 0x00010001010001ff, 0x0001000101000100,
+    0x000100010101ffff, 0x00010001010100ff, 0x0001000101010001, 0x0001000101010101,
+    0x000101ffff000001, 0x000101ffff000100, 0x000101ffff010000, 0x000101ff00ffff00,
+    0x000101ff0000ff01, 0x000101ff00000000, 0x000101ff00000101, 0x000101ff0001ff00,
+    0x000101ff00010100, 0x000101ff01ff0000, 0x000101ff0100ff00, 0x000101ff010001ff,
+    0x000101ff01010001, 0x00010100ffffff00, 0x00010100ffff00ff, 0x00010100ff00ffff,
+    0x00010100ff000000, 0x00010100ff01ff00, 0x00010100ff0100ff, 0x00010100ff010001,
+    0x00010100ff010100, 0x0001010000ffffff, 0x0001010000ffff00, 0x0001010000ff0000,
+    0x0001010000ff0001, 0x0001010000ff01ff, 0x000101000000ff00, 0x00010100000000ff,
+    0x0001010000000000, 0x0001010000000001, 0x0001010000000100, 0x000101000001ffff,
+    0x0001010000010000, 0x0001010000010101, 0x0001010001ffff01, 0x0001010001ff00ff,
+    0x0001010001ff0101, 0x0001010001000000, 0x000101000101ff00, 0x00010100010100ff,
+    0x0001010001010000, 0x0001010001010100, 0x00010101ff00ff00, 0x00010101ff000001,
+    0x00010101ff0001ff, 0x0001010100ffff00, 0x0001010100ff00ff, 0x0001010100ff0100,
+    0x000101010000ffff, 0x0001010100000000, 0x00010101000001ff, 0x0001010100000101,
+    0x00010101000100ff, 0x0001010100010000, 0x0001010100010100, 0x0001010101ff0001,
+    0x00010101010000ff, 0x00010101010001ff, 0x0001010101000101, 0x0001010101010001,
+    0x01ffffffffffffff, 0x01ffffffffffff01, 0x01ffffffffff01ff, 0x01ffffffffff0101,
+    0x01ffffffff01ffff, 0x01ffffffff01ff01, 0x01ffffffff0101ff, 0x01ffffffff010101,
+    0x01ffffff00ff0000, 0x01ffffff0000ffff, 0x01ffffff0000ff00, 0x01ffffff000000ff,
+    0x01ffffff00000001, 0x01ffffff00000100, 0x01ffffff00010000, 0x01ffffff01ffffff,
+    0x01ffffff01ffff01, 0x01ffffff01ff01ff, 0x01ffffff01ff0101, 0x01ffffff01000000,
+    0x01ffffff0101ffff, 0x01ffffff0101ff01, 0x01ffffff010101ff, 0x01ffffff01010101,
+    0x01ffff00ffff0000, 0x01ffff00ff00ff00, 0x01ffff00ff0000ff, 0x01ffff00ff000001,
+    0x01ffff00ff000100, 0x01ffff00ff010000, 0x01ffff0000ffff00, 0x01ffff0000ff00ff,
+    0x01ffff0000ff0100, 0x01ffff000000ffff, 0x01ffff000000ff01, 0x01ffff0000000000,
+    0x01ffff0000000001, 0x01ffff00000001ff, 0x01ffff0000000100, 0x01ffff00000100ff,
+    0x01ffff0000010001, 0x01ffff0000010100, 0x01ffff0001ff0000, 0x01ffff0001ff0100,
+    0x01ffff00010000ff, 0x01ffff0001000001, 0x01ffff0001000100, 0x01ffff0001010000,
+    0x01ffff01ffffffff, 0x01ffff01ffffff01, 0x01ffff01ffff01ff, 0x01ffff01ffff0101,
+    0x01ffff01ff000000, 0x01ffff01ff01ffff, 0x01ffff01ff01ff01, 0x01ffff01ff0101ff,
+    0x01ffff01ff010101, 0x01ffff010000ff00, 0x01ffff01000000ff, 0x01ffff0100000100,
+    0x01ffff0100010000, 0x01ffff0101ffffff, 0x01ffff0101ffff01, 0x01ffff0101ff01ff,
+    0x01ffff0101ff0101, 0x01ffff0101000000, 0x01ffff010101ffff, 0x01ffff010101ff01,
+    0x01ffff01010101ff, 0x01ffff0101010101, 0x01ff00ffff0000ff, 0x01ff00ffff000100,
+    0x01ff00ff00ffff00, 0x01ff00ff00ff00ff, 0x01ff00ff0000ff00, 0x01ff00ff00000000,
+    0x01ff00ff00000101, 0x01ff00ff0001ff00, 0x01ff00ff000100ff, 0x01ff00ff00010100,
+    0x01ff00ff010000ff, 0x01ff00ff01000100, 0x01ff0000ffffff00, 0x01ff0000ffff0100,
+    0x01ff0000ff00ff01, 0x01ff0000ff000000, 0x01ff0000ff000101, 0x01ff0000ff010001,
+    0x01ff0000ff010100, 0x01ff000000ffffff, 0x01ff000000ffff00, 0x01ff000000ff0000,
+    0x01ff000000ff01ff, 0x01ff00000000ff00, 0x01ff0000000000ff, 0x01ff000000000000,
+    0x01ff000000000001, 0x01ff000000000100, 0x01ff000000000101, 0x01ff000000010000,
+    0x01ff000000010001, 0x01ff0000000101ff, 0x01ff000000010101, 0x01ff000001ffff00,
+    0x01ff000001ff00ff, 0x01ff000001ff0001, 0x01ff000001ff0100, 0x01ff00000100ffff,
+    0x01ff00000100ff01, 0x01ff000001000000, 0x01ff0000010001ff, 0x01ff000001010001,
+    0x01ff0001ff00ff00, 0x01ff0001ff000001, 0x01ff0001ff000100, 0x01ff0001ff010000,
+    0x01ff000100ffff00, 0x01ff000100ff00ff, 0x01ff000100ff0100, 0x01ff000100ff0101,
+    0x01ff00010000ffff, 0x01ff000100000000, 0x01ff000100000100, 0x01ff000100000101,
+    0x01ff00010001ff00, 0x01ff000100010001, 0x01ff000100010101, 0x01ff000101ff0000,
+    0x01ff00010100ff00, 0x01ff000101000101, 0x01ff0001010100ff, 0x01ff01ffffffffff,
+    0x01ff01ffffffff01, 0x01ff01ffffff01ff, 0x01ff01ffffff0101, 0x01ff01ffff000000,
+    0x01ff01ffff01ffff, 0x01ff01ffff01ff01, 0x01ff01ffff0101ff, 0x01ff01ffff010101,
+    0x01ff01ff00ffff00, 0x01ff01ff00ff0000, 0x01ff01ff0000ff00, 0x01ff01ff000000ff,
+    0x01ff01ff00000100, 0x01ff01ff00010000, 0x01ff01ff00010100, 0x01ff01ff01ffffff,
+    0x01ff01ff01ffff01, 0x01ff01ff01ff01ff, 0x01ff01ff01ff0101, 0x01ff01ff01000000,
+    0x01ff01ff0101ffff, 0x01ff01ff0101ff01, 0x01ff01ff010101ff, 0x01ff01ff01010101,
+    0x01ff0100ffff0000, 0x01ff0100ffff0001, 0x01ff0100ff00ff00, 0x01ff0100ff0000ff,
+    0x01ff0100ff000001, 0x01ff0100ff010000, 0x01ff010000ffff00, 0x01ff010000ff00ff,
+    0x01ff010000ff0001, 0x01ff010000ff0100, 0x01ff01000000ffff, 0x01ff01000000ff01,
+    0x01ff010000000000, 0x01ff010000000101, 0x01ff01000001ff00, 0x01ff0100000100ff,
+    0x01ff010001ff0000, 0x01ff010001000001, 0x01ff010001000100, 0x01ff010001010000,
+    0x01ff0101ffffffff, 0x01ff0101ffffff01, 0x01ff0101ffff01ff, 0x01ff0101ffff0101,
+    0x01ff0101ff000000, 0x01ff0101ff01ffff, 0x01ff0101ff01ff01, 0x01ff0101ff0101ff,
+    0x01ff0101ff010101, 0x01ff010100ff0000, 0x01ff01010000ff00, 0x01ff0101000000ff,
+    0x01ff010100000001, 0x01ff010101ffffff, 0x01ff010101ffff01, 0x01ff010101ff01ff,
+    0x01ff010101ff0101, 0x01ff010101000000, 0x01ff01010101ffff, 0x01ff01010101ff01,
+    0x01ff0101010101ff, 0x01ff010101010101, 0x0100ffffffff0000, 0x0100ffffff00ff00,
+    0x0100ffffff000001, 0x0100ffffff0001ff, 0x0100ffffff000100, 0x0100ffffff010000,
+    0x0100ffff00ffff00, 0x0100ffff00ff0001, 0x0100ffff00ff0100, 0x0100ffff00000000,
+    0x0100ffff000001ff, 0x0100ffff00000101, 0x0100ffff00010100, 0x0100ffff00010101,
+    0x0100ffff01ff0000, 0x0100ffff0100ff00, 0x0100ffff010000ff, 0x0100ffff01000001,
+    0x0100ffff01000100, 0x0100ffff01010000, 0x0100ff00ffffff00, 0x0100ff00ffff00ff,
+    0x0100ff00ffff0001, 0x0100ff00ffff0100, 0x0100ff00ff00ffff, 0x0100ff00ff000000,
+    0x0100ff00ff0001ff, 0x0100ff00ff000101, 0x0100ff00ff01ff00, 0x0100ff00ff0100ff,
+    0x0100ff00ff010001, 0x0100ff00ff010100, 0x0100ff0000ffffff, 0x0100ff0000ff0000,
+    0x0100ff000000ffff, 0x0100ff000000ff00, 0x0100ff00000000ff, 0x0100ff0000000000,
+    0x0100ff0000000001, 0x0100ff0000000100, 0x0100ff000001ff01, 0x0100ff0000010000,
+    0x0100ff0001ff00ff, 0x0100ff0001ff0001, 0x0100ff000100ff01, 0x0100ff0001000000,
+    0x0100ff00010001ff, 0x0100ff000101ff00, 0x0100ff00010100ff, 0x0100ff0001010001,
+    0x0100ff0001010100, 0x0100ff01ffff0000, 0x0100ff01ff00ff00, 0x0100ff01ff0000ff,
+    0x0100ff01ff000100, 0x0100ff01ff010000, 0x0100ff0100ff00ff, 0x0100ff0100ff0001,
+    0x0100ff0100ff0100, 0x0100ff010000ffff, 0x0100ff010000ff01, 0x0100ff0100000000,
+    0x0100ff01000001ff, 0x0100ff0100010001, 0x0100ff0100010100, 0x0100ff0101ff0000,
+    0x0100ff01010000ff, 0x0100ff0101000001, 0x0100ff0101010100, 0x010000ffffffff00,
+    0x010000ffffff00ff, 0x010000ffffff0001, 0x010000ffff00ffff, 0x010000ffff000000,
+    0x010000ffff0001ff, 0x010000ffff010001, 0x010000ff00ffffff, 0x010000ff00ff0101,
+    0x010000ff0000ff00, 0x010000ff000000ff, 0x010000ff00000000, 0x010000ff00000001,
+    0x010000ff000001ff, 0x010000ff00000100, 0x010000ff0001ffff, 0x010000ff0001ff00,
+    0x010000ff0001ff01, 0x010000ff00010000, 0x010000ff01ff00ff, 0x010000ff01ff0001,
+    0x010000ff0100ff01, 0x010000ff010000ff, 0x010000ff01000000, 0x010000ff010001ff,
+    0x010000ff0101ff00, 0x010000ff01010100, 0x01000000ffffffff, 0x01000000ffff0000,
+    0x01000000ffff01ff, 0x01000000ffff0101, 0x01000000ff00ffff, 0x01000000ff00ff00,
+    0x01000000ff0000ff, 0x01000000ff000000, 0x01000000ff000001, 0x01000000ff000100,
+    0x01000000ff01ff00, 0x01000000ff010000, 0x01000000ff010100, 0x01000000ff010101,
+    0x0100000000ffff00, 0x0100000000ff00ff, 0x0100000000ff0000, 0x0100000000ff0001,
+    0x0100000000ff0100, 0x010000000000ffff, 0x010000000000ff00, 0x010000000000ff01,
+    0x01000000000000ff, 0x0100000000000000, 0x0100000000000001, 0x01000000000001ff,
+    0x0100000000000100, 0x0100000000000101, 0x010000000001ff00, 0x01000000000100ff,
+    0x0100000000010000, 0x0100000000010001, 0x0100000000010100, 0x0100000001ffff00,
+    0x0100000001ff0000, 0x0100000001ff01ff, 0x010000000100ff00, 0x010000000100ff01,
+    0x01000000010000ff, 0x0100000001000000, 0x0100000001000001, 0x0100000001000100,
+    0x0100000001000101, 0x010000000101ffff, 0x010000000101ff01, 0x0100000001010000,
+    0x01000000010101ff, 0x0100000001010101, 0x01000001ffffff00, 0x01000001ffff00ff,
+    0x01000001ff00ffff, 0x01000001ff000000, 0x01000001ff000100, 0x01000001ff01ffff,
+    0x01000001ff010001, 0x01000001ff010100, 0x0100000100ff0000, 0x0100000100ff01ff,
+    0x0100000100ff0100, 0x010000010000ff00, 0x010000010000ff01, 0x0100000100000000,
+    0x0100000100000001, 0x0100000100000100, 0x0100000100010000, 0x01000001000101ff,
+    0x0100000101ffff01, 0x0100000101ff00ff, 0x0100000101ff0100, 0x0100000101ff0101,
+    0x010000010100ff01, 0x01000001010000ff, 0x0100000101000000, 0x01000001010100ff,
+    0x0100000101010001, 0x0100000101010100, 0x010001ffffff0000, 0x010001ffff000001,
+    0x010001ffff000100, 0x010001ffff010000, 0x010001ff00ffff00, 0x010001ff00ff0001,
+    0x010001ff0000ffff, 0x010001ff0000ff01, 0x010001ff00000000, 0x010001ff00000001,
+    0x010001ff00000101, 0x010001ff000100ff, 0x010001ff00010000, 0x010001ff01ff0000,
+    0x010001ff0100ff00, 0x010001ff01000001, 0x010001ff01000100, 0x010001ff01010000,
+    0x01000100ffff00ff, 0x01000100ffff0001, 0x01000100ffff0100, 0x01000100ff00ffff,
+    0x01000100ff00ff01, 0x01000100ff000000, 0x01000100ff0001ff, 0x01000100ff000101,
+    0x01000100ff01ffff, 0x01000100ff01ff00, 0x01000100ff0100ff, 0x01000100ff010001,
+    0x0100010000ffffff, 0x0100010000ffff01, 0x0100010000ff0000, 0x0100010000ff01ff,
+    0x0100010000ff0101, 0x010001000000ff00, 0x01000100000000ff, 0x0100010000000000,
+    0x0100010000000001, 0x0100010000000100, 0x010001000001ff01, 0x0100010000010000,
+    0x0100010000010001, 0x0100010000010101, 0x0100010001ffff00, 0x0100010001ff00ff,
+    0x010001000100ffff, 0x010001000100ff01, 0x0100010001000000, 0x0100010001000101,
+    0x010001000101ff00, 0x0100010001010001, 0x01000101ffff0000, 0x01000101ff000000,
+    0x01000101ff010000, 0x0100010100ff00ff, 0x0100010100ff0001, 0x0100010100ff0100,
+    0x010001010000ffff, 0x0100010100000000, 0x01000101000001ff, 0x010001010001ff00,
+    0x0100010101ff0000, 0x010001010100ff00, 0x01000101010000ff, 0x0100010101000000,
+    0x0100010101000001, 0x0101ffffffffffff, 0x0101ffffffffff01, 0x0101ffffffff01ff,
+    0x0101ffffffff0101, 0x0101ffffff000000, 0x0101ffffff01ffff, 0x0101ffffff01ff01,
+    0x0101ffffff0101ff, 0x0101ffffff010101, 0x0101ffff00ff0000, 0x0101ffff0000ff00,
+    0x0101ffff000000ff, 0x0101ffff00000001, 0x0101ffff00000100, 0x0101ffff01ffffff,
+    0x0101ffff01ffff01, 0x0101ffff01ff01ff, 0x0101ffff01ff0101, 0x0101ffff01000000,
+    0x0101ffff0101ffff, 0x0101ffff0101ff01, 0x0101ffff010101ff, 0x0101ffff01010101,
+    0x0101ff00ffff0000, 0x0101ff00ffff0100, 0x0101ff00ff00ff00, 0x0101ff00ff0000ff,
+    0x0101ff00ff000001, 0x0101ff00ff000100, 0x0101ff00ff000101, 0x0101ff0000ff0001,
+    0x0101ff0000ff0100, 0x0101ff000000ff00, 0x0101ff0000000000, 0x0101ff00000001ff,
+    0x0101ff0000000101, 0x0101ff000001ff00, 0x0101ff00000100ff, 0x0101ff0001ff0000,
+    0x0101ff000100ffff, 0x0101ff000100ff01, 0x0101ff0001000001, 0x0101ff0001000100,
+    0x0101ff01ffffff01, 0x0101ff01ffff01ff, 0x0101ff01ffff0101, 0x0101ff01ff00ffff,
+    0x0101ff01ff000100, 0x0101ff01ff01ff01, 0x0101ff01ff0101ff, 0x0101ff01ff010101,
+    0x0101ff0100ff0000, 0x0101ff010000ff00, 0x0101ff0100000001, 0x0101ff0100000100,
+    0x0101ff0100010000, 0x0101ff0101ffffff, 0x0101ff0101ffff01, 0x0101ff0101ff01ff,
+    0x0101ff0101ff0101, 0x0101ff0101000000, 0x0101ff010101ffff, 0x0101ff010101ff01,
+    0x0101ff01010101ff, 0x0101ff0101010101, 0x010100ffff000100, 0x010100ffff010000,
+    0x010100ff00ffff00, 0x010100ff00ff00ff, 0x010100ff0000ffff, 0x010100ff000000ff,
+    0x010100ff00000000, 0x010100ff000001ff, 0x010100ff00000101, 0x010100ff0001ff00,
+    0x010100ff00010000, 0x010100ff00010001, 0x010100ff000101ff, 0x010100ff00010100,
+    0x010100ff01ff0000, 0x01010000ffff0001, 0x01010000ffff0100, 0x01010000ff00ffff,
+    0x01010000ff00ff01, 0x01010000ff000000, 0x01010000ff0001ff, 0x01010000ff010001,
+    0x01010000ff010100, 0x0101000000ffff01, 0x0101000000ff0000, 0x010100000000ff00,
+    0x01010000000000ff, 0x0101000000000000, 0x0101000000000001, 0x0101000000000100,
+    0x0101000000010000, 0x0101000000010101, 0x0101000001ffff00, 0x0101000001ff00ff,
+    0x0101000001ff0000, 0x0101000001ff0001, 0x0101000001ff0100, 0x010100000100ff01,
+    0x0101000001000000, 0x01010000010001ff, 0x01010001ffff0000, 0x01010001ff00ff00,
+    0x01010001ff000001, 0x01010001ff000101, 0x01010001ff01ff00, 0x01010001ff010000,
+    0x0101000100ff00ff, 0x0101000100ff0001, 0x0101000100ff0101, 0x010100010000ff01,
+    0x0101000100000000, 0x0101000100000001, 0x01010001000001ff, 0x010100010001ffff,
+    0x010100010001ff01, 0x0101000101ff0001, 0x010100010100ffff, 0x0101000101000000,
+    0x0101000101000001, 0x0101000101000100, 0x010100010101ff00, 0x01010001010100ff,
+    0x0101000101010001, 0x010101ffffffffff, 0x010101ffffffff01, 0x010101ffffff01ff,
+    0x010101ffffff0101, 0x010101ffff01ffff, 0x010101ffff01ff01, 0x010101ffff0101ff,
+    0x010101ffff010101, 0x010101ff0000ff00, 0x010101ff000000ff, 0x010101ff00000001,
+    0x010101ff00000100, 0x010101ff01ffffff, 0x010101ff01ffff01, 0x010101ff01ff01ff,
+    0x010101ff01ff0101, 0x010101ff01000000, 0x010101ff0101ffff, 0x010101ff0101ff01,
+    0x010101ff010101ff, 0x010101ff01010101, 0x01010100ffff0000, 0x01010100ff0000ff,
+    0x01010100ff000100, 0x01010100ff01ff00, 0x01010100ff010000, 0x0101010000ffff00,
+    0x010101000000ffff, 0x0101010000000000, 0x0101010000000101, 0x010101000001ff00,
+    0x0101010000010001, 0x0101010000010100, 0x010101000100ffff, 0x0101010001000001,
+    0x01010101ffffffff, 0x01010101ffffff01, 0x01010101ffff01ff, 0x01010101ffff0101,
+    0x01010101ff01ffff, 0x01010101ff01ff01, 0x01010101ff0101ff, 0x01010101ff010101,
+    0x010101010000ff00, 0x01010101000000ff, 0x0101010100000001, 0x0101010101ffffff,
+    0x0101010101ffff01, 0x0101010101ff01ff, 0x0101010101ff0101, 0x0101010101000000,
+    0x010101010101ffff, 0x010101010101ff01, 0x01010101010101ff, 0x0101010101010101,
+GGML_TABLE_END()
+#else
+GGML_TABLE_BEGIN(uint32_t, iq1s_grid_gpu, NGRID_IQ1S)
+    0x00000000, 0x00000002, 0x00000101, 0x00000200, 0x00000202, 0x00010001, 0x00010101, 0x00020000,
+    0x00020002, 0x00020200, 0x00020202, 0x01000101, 0x01010001, 0x01010100, 0x01010102, 0x01020101,
+    0x02000000, 0x02000002, 0x02000200, 0x02000202, 0x02010101, 0x02020000, 0x02020002, 0x02020200,
+    0x02020202, 0x00000110, 0x00000111, 0x00010011, 0x00010110, 0x00010112, 0x00010211, 0x00010212,
+    0x00020111, 0x01000011, 0x01000112, 0x01000211, 0x01010012, 0x01010111, 0x01010212, 0x01020011,
+    0x01020110, 0x01020112, 0x01020210, 0x02000111, 0x02010011, 0x02010110, 0x02010112, 0x02020111,
+    0x00000020, 0x00000022, 0x00000220, 0x00000222, 0x00010121, 0x00020020, 0x00020022, 0x00020220,
+    0x00020222, 0x01000121, 0x01010021, 0x01010221, 0x01020120, 0x01020221, 0x02000020, 0x02000022,
+    0x02000220, 0x02000222, 0x02010021, 0x02010121, 0x02010221, 0x02020020, 0x02020022, 0x02020220,
+    0x02020222, 0x00011001, 0x00011100, 0x00011102, 0x00021101, 0x01001001, 0x01001201, 0x01011101,
+    0x01011202, 0x01021100, 0x01021101, 0x02011001, 0x02011201, 0x02021101, 0x00001011, 0x00001110,
+    0x00001111, 0x00001112, 0x00011111, 0x00011210, 0x00011212, 0x00021211, 0x01001010, 0x01001111,
+    0x01001212, 0x01011010, 0x01011011, 0x01011110, 0x01011111, 0x01011112, 0x01011211, 0x01021010,
+    0x01021012, 0x01021111, 0x01021210, 0x01021212, 0x02001011, 0x02011011, 0x02011111, 0x02011210,
+    0x02011212, 0x02021011, 0x02021110, 0x02021111, 0x02021112, 0x02021211, 0x00011120, 0x00011221,
+    0x01001021, 0x01001120, 0x01011020, 0x01011022, 0x01011121, 0x01011220, 0x01021020, 0x01021021,
+    0x01021122, 0x01021221, 0x02001121, 0x02011021, 0x02011120, 0x02011221, 0x00002000, 0x00002002,
+    0x00002200, 0x00002202, 0x00012101, 0x00022000, 0x00022002, 0x00022200, 0x00022202, 0x01002101,
+    0x01012001, 0x01012102, 0x01022101, 0x02002000, 0x02002002, 0x02002200, 0x02002202, 0x02012101,
+    0x02022000, 0x02022002, 0x02022200, 0x02022202, 0x00002111, 0x00012011, 0x00012110, 0x00012211,
+    0x00022110, 0x00022111, 0x01002011, 0x01012010, 0x01012011, 0x01012111, 0x01022011, 0x01022110,
+    0x01022211, 0x02012011, 0x02012110, 0x02012112, 0x02012211, 0x02022111, 0x00002020, 0x00002022,
+    0x00002220, 0x00002222, 0x00012121, 0x00022020, 0x00022022, 0x00022220, 0x00022222, 0x01002121,
+    0x01012021, 0x01012221, 0x01022021, 0x01022121, 0x02002020, 0x02002022, 0x02002121, 0x02002220,
+    0x02002222, 0x02012121, 0x02022020, 0x02022022, 0x02022220, 0x02022222, 0x00110000, 0x00110001,
+    0x00110100, 0x00110201, 0x00120100, 0x00120101, 0x01100001, 0x01100100, 0x01110000, 0x01110101,
+    0x01110200, 0x01120001, 0x01120100, 0x01120101, 0x01120201, 0x02110001, 0x02110100, 0x02110102,
+    0x02120001, 0x02120101, 0x00100011, 0x00100110, 0x00100112, 0x00100211, 0x00110010, 0x00110012,
+    0x00110111, 0x00110210, 0x00120011, 0x00120110, 0x00120211, 0x01100111, 0x01100212, 0x01110010,
+    0x01110011, 0x01110012, 0x01110110, 0x01110111, 0x01110112, 0x01110211, 0x01120010, 0x01120111,
+    0x02100110, 0x02110012, 0x02110111, 0x02120011, 0x02120110, 0x00110021, 0x00110120, 0x00110122,
+    0x00120121, 0x01100020, 0x01100122, 0x01100221, 0x01110022, 0x01110121, 0x01110220, 0x01110222,
+    0x01120120, 0x01120122, 0x02100121, 0x02110021, 0x02110120, 0x02110122, 0x02120121, 0x00101001,
+    0x00101102, 0x00101201, 0x00111100, 0x00111101, 0x00111200, 0x00111201, 0x00121001, 0x00121102,
+    0x01101001, 0x01101101, 0x01101102, 0x01101200, 0x01101202, 0x01111001, 0x01111100, 0x01111101,
+    0x01111102, 0x01111201, 0x01121002, 0x01121101, 0x01121200, 0x02101100, 0x02101201, 0x02111000,
+    0x02111100, 0x02111101, 0x02111200, 0x02111201, 0x02111202, 0x02121001, 0x02121100, 0x02121101,
+    0x02121201, 0x00101012, 0x00101111, 0x00101212, 0x00111011, 0x00111110, 0x00111111, 0x00111112,
+    0x00111211, 0x00121010, 0x00121012, 0x00121111, 0x00121210, 0x00121212, 0x01101011, 0x01101110,
+    0x01101111, 0x01101112, 0x01111011, 0x01111012, 0x01111110, 0x01111111, 0x01111112, 0x01111211,
+    0x01111212, 0x01121011, 0x01121110, 0x01121111, 0x01121112, 0x01121211, 0x02101010, 0x02101012,
+    0x02101110, 0x02101111, 0x02101210, 0x02101212, 0x02111010, 0x02111011, 0x02111110, 0x02111111,
+    0x02111112, 0x02111211, 0x02111212, 0x02121010, 0x02121012, 0x02121111, 0x00101021, 0x00101120,
+    0x00101121, 0x00101122, 0x00111121, 0x00111122, 0x00111220, 0x00111222, 0x00121021, 0x00121122,
+    0x01101020, 0x01101022, 0x01101120, 0x01101121, 0x01101220, 0x01101222, 0x01111021, 0x01111121,
+    0x01111122, 0x01111220, 0x01111221, 0x01121021, 0x01121120, 0x01121121, 0x01121220, 0x01121221,
+    0x01121222, 0x02101122, 0x02101222, 0x02111022, 0x02111121, 0x02121120, 0x02121221, 0x00112001,
+    0x00112102, 0x00122101, 0x01102001, 0x01102100, 0x01102102, 0x01102201, 0x01112000, 0x01112101,
+    0x01112200, 0x01112202, 0x01122000, 0x01122001, 0x01122100, 0x01122102, 0x01122201, 0x02102101,
+    0x02112001, 0x02112100, 0x02122101, 0x00112010, 0x00112012, 0x00112111, 0x00112212, 0x00122011,
+    0x00122111, 0x01102012, 0x01102110, 0x01102111, 0x01102210, 0x01112011, 0x01112110, 0x01112111,
+    0x01112112, 0x01112211, 0x01112212, 0x01122010, 0x01122111, 0x01122212, 0x02102211, 0x02112011,
+    0x02112012, 0x02112111, 0x02112210, 0x02122011, 0x02122112, 0x02122211, 0x00102221, 0x00112122,
+    0x00122120, 0x00122122, 0x01102120, 0x01102122, 0x01102221, 0x01112020, 0x01112022, 0x01112121,
+    0x01112220, 0x01122021, 0x01122122, 0x01122221, 0x02102121, 0x02112021, 0x02112122, 0x02112222,
+    0x00200000, 0x00200002, 0x00200200, 0x00200202, 0x00210101, 0x00220000, 0x00220002, 0x00220101,
+    0x00220200, 0x00220202, 0x01200101, 0x01210001, 0x01210201, 0x01220001, 0x01220101, 0x02200000,
+    0x02200002, 0x02200200, 0x02200202, 0x02210101, 0x02220000, 0x02220002, 0x02220101, 0x02220200,
+    0x02220202, 0x00200111, 0x00210011, 0x00210110, 0x00210211, 0x00220111, 0x01200012, 0x01200110,
+    0x01200211, 0x01210111, 0x01210210, 0x01210212, 0x01220011, 0x01220110, 0x01220111, 0x01220112,
+    0x02200111, 0x02210010, 0x02210112, 0x02210211, 0x02220111, 0x00200021, 0x00200220, 0x00200222,
+    0x00210021, 0x00210121, 0x00220020, 0x00220022, 0x00220220, 0x00220222, 0x01200121, 0x01210021,
+    0x01210122, 0x01210221, 0x01220121, 0x02200021, 0x02200220, 0x02200222, 0x02210021, 0x02210121,
+    0x02220020, 0x02220022, 0x02220220, 0x02220222, 0x00201101, 0x00211100, 0x00211102, 0x00211201,
+    0x00221101, 0x01201100, 0x01201101, 0x01201102, 0x01201201, 0x01211002, 0x01211101, 0x01211200,
+    0x01211202, 0x01221102, 0x02201101, 0x02211001, 0x02211100, 0x02211201, 0x02221001, 0x02221101,
+    0x00201211, 0x00211111, 0x00221011, 0x00221211, 0x01201010, 0x01201111, 0x01201210, 0x01211011,
+    0x01211110, 0x01211111, 0x01211211, 0x01221012, 0x01221111, 0x01221210, 0x02201211, 0x02211010,
+    0x02211110, 0x02211111, 0x02211210, 0x02211212, 0x02221011, 0x02221110, 0x02221112, 0x02221211,
+    0x00201121, 0x00211020, 0x00211022, 0x00211221, 0x00221121, 0x01201021, 0x01201221, 0x01211121,
+    0x01221020, 0x01221021, 0x01221221, 0x02201120, 0x02201122, 0x02211020, 0x02211222, 0x00202000,
+    0x00202002, 0x00202200, 0x00202202, 0x00212101, 0x00222000, 0x00222002, 0x00222200, 0x00222202,
+    0x01202101, 0x01212001, 0x01212100, 0x01222101, 0x02202000, 0x02202002, 0x02202200, 0x02202202,
+    0x02222000, 0x02222002, 0x02222200, 0x02222202, 0x00202211, 0x00212011, 0x00212110, 0x00212211,
+    0x00222111, 0x01202112, 0x01202211, 0x01212012, 0x01212111, 0x01222011, 0x01222110, 0x01222112,
+    0x01222211, 0x02202111, 0x02212010, 0x02212112, 0x02212211, 0x02222110, 0x02222111, 0x00202020,
+    0x00202022, 0x00202220, 0x00202222, 0x00222020, 0x00222022, 0x00222220, 0x00222222, 0x01202121,
+    0x01212021, 0x01212122, 0x01212221, 0x01222121, 0x02202020, 0x02202022, 0x02202220, 0x02202222,
+    0x02212121, 0x02222020, 0x02222022, 0x02222220, 0x02222222, 0x10000101, 0x10010001, 0x10010102,
+    0x10020101, 0x11000201, 0x11010002, 0x11010101, 0x11010200, 0x11010202, 0x11020001, 0x11020100,
+    0x11020102, 0x12010100, 0x12010201, 0x12020001, 0x12020102, 0x10000010, 0x10000011, 0x10000110,
+    0x10000112, 0x10000211, 0x10010012, 0x10010111, 0x10010112, 0x10010210, 0x10010212, 0x10020011,
+    0x10020112, 0x10020211, 0x11000111, 0x11000210, 0x11000212, 0x11010011, 0x11010110, 0x11010111,
+    0x11010112, 0x11010211, 0x11010212, 0x11020111, 0x11020210, 0x11020212, 0x12000011, 0x12000110,
+    0x12000112, 0x12010010, 0x12010012, 0x12010111, 0x12020010, 0x12020011, 0x12020012, 0x10000121,
+    0x10010021, 0x10010120, 0x10010122, 0x10020121, 0x11000021, 0x11010022, 0x11010121, 0x11010222,
+    0x11020120, 0x11020221, 0x12000221, 0x12010120, 0x12020121, 0x10001001, 0x10011101, 0x10011201,
+    0x10021201, 0x11001101, 0x11001200, 0x11001202, 0x11011001, 0x11011100, 0x11011101, 0x11011102,
+    0x11021001, 0x11021002, 0x11021101, 0x11021200, 0x11021202, 0x12001001, 0x12001102, 0x12001201,
+    0x12011000, 0x12011002, 0x12011101, 0x12021000, 0x12021001, 0x12021201, 0x10001011, 0x10001012,
+    0x10001111, 0x10001212, 0x10011011, 0x10011110, 0x10011111, 0x10011112, 0x10011211, 0x10021010,
+    0x10021111, 0x10021212, 0x11001011, 0x11001110, 0x11001111, 0x11001112, 0x11001211, 0x11011010,
+    0x11011011, 0x11011110, 0x11011111, 0x11011112, 0x11011210, 0x11011211, 0x11021011, 0x11021110,
+    0x11021111, 0x11021112, 0x11021211, 0x12001012, 0x12001110, 0x12001111, 0x12001210, 0x12011011,
+    0x12011110, 0x12011111, 0x12011112, 0x12011211, 0x12011212, 0x12021111, 0x12021210, 0x12021212,
+    0x10001021, 0x10001121, 0x10001221, 0x10011120, 0x10011121, 0x10011220, 0x10011222, 0x10021021,
+    0x10021120, 0x10021221, 0x11001020, 0x11001022, 0x11001121, 0x11001220, 0x11011020, 0x11011021,
+    0x11011022, 0x11011121, 0x11011122, 0x11011221, 0x11021022, 0x11021121, 0x11021220, 0x12001021,
+    0x12001121, 0x12001222, 0x12011120, 0x12011121, 0x12021021, 0x12021120, 0x12021122, 0x10002101,
+    0x10012001, 0x10012101, 0x10012202, 0x10022101, 0x11002002, 0x11002201, 0x11012000, 0x11012101,
+    0x11012200, 0x11022001, 0x11022100, 0x11022102, 0x11022201, 0x12002101, 0x12012001, 0x12012100,
+    0x12012102, 0x12012201, 0x12022101, 0x10002011, 0x10002111, 0x10002112, 0x10002212, 0x10012010,
+    0x10012110, 0x10012111, 0x10012210, 0x10022011, 0x10022110, 0x10022112, 0x11002010, 0x11002111,
+    0x11002212, 0x11012011, 0x11012012, 0x11012110, 0x11012111, 0x11012112, 0x11012211, 0x11022010,
+    0x11022012, 0x11022111, 0x11022112, 0x11022212, 0x12002112, 0x12002211, 0x12012012, 0x12012111,
+    0x12012112, 0x12012210, 0x12022011, 0x12022110, 0x12022112, 0x12022211, 0x10012122, 0x11002120,
+    0x11002122, 0x11002221, 0x11012121, 0x11012220, 0x11012222, 0x11022120, 0x11022221, 0x12012120,
+    0x12022121, 0x10100001, 0x10100100, 0x10100101, 0x10100102, 0x10100201, 0x10110002, 0x10110101,
+    0x10110202, 0x10120001, 0x10120100, 0x10120201, 0x11100000, 0x11100101, 0x11100200, 0x11110001,
+    0x11110100, 0x11110101, 0x11110102, 0x11110201, 0x11120101, 0x11120200, 0x12100102, 0x12100201,
+    0x12110101, 0x12110200, 0x12120000, 0x12120001, 0x12120102, 0x12120201, 0x10100111, 0x10100210,
+    0x10100211, 0x10100212, 0x10110011, 0x10110110, 0x10110111, 0x10110112, 0x10110210, 0x10110211,
+    0x10120010, 0x10120111, 0x10120112, 0x10120210, 0x10120212, 0x11100011, 0x11100110, 0x11100111,
+    0x11100112, 0x11100211, 0x11110010, 0x11110011, 0x11110012, 0x11110110, 0x11110111, 0x11110112,
+    0x11110210, 0x11110211, 0x11110212, 0x11120011, 0x11120110, 0x11120111, 0x11120112, 0x11120211,
+    0x12100012, 0x12100111, 0x12110011, 0x12110110, 0x12110111, 0x12110112, 0x12110211, 0x12120010,
+    0x12120111, 0x12120212, 0x10100021, 0x10100122, 0x10110022, 0x10110121, 0x10110222, 0x10120021,
+    0x10120120, 0x11100022, 0x11100121, 0x11100222, 0x11110021, 0x11110120, 0x11110121, 0x11110122,
+    0x11110221, 0x11120022, 0x11120121, 0x12100121, 0x12110020, 0x12110022, 0x12110121, 0x12110221,
+    0x12110222, 0x12120120, 0x10101100, 0x10101101, 0x10111001, 0x10111100, 0x10111101, 0x10111102,
+    0x10111200, 0x10111201, 0x10121001, 0x10121101, 0x10121200, 0x10121202, 0x11101001, 0x11101100,
+    0x11101101, 0x11101102, 0x11101201, 0x11101202, 0x11111000, 0x11111001, 0x11111100, 0x11111101,
+    0x11111102, 0x11111200, 0x11111201, 0x11111202, 0x11121001, 0x11121002, 0x11121100, 0x11121101,
+    0x11121102, 0x11121201, 0x12101000, 0x12101200, 0x12101202, 0x12111001, 0x12111100, 0x12111101,
+    0x12111102, 0x12111201, 0x12121001, 0x12121100, 0x12121101, 0x12121202, 0x10101011, 0x10101012,
+    0x10101110, 0x10101111, 0x10101112, 0x10101211, 0x10111010, 0x10111011, 0x10111012, 0x10111110,
+    0x10111111, 0x10111112, 0x10111211, 0x10111212, 0x10121011, 0x10121110, 0x10121111, 0x10121112,
+    0x10121211, 0x11101010, 0x11101011, 0x11101012, 0x11101110, 0x11101111, 0x11101112, 0x11101210,
+    0x11101211, 0x11111010, 0x11111011, 0x11111012, 0x11111110, 0x11111111, 0x11111112, 0x11111210,
+    0x11111211, 0x11111212, 0x11121010, 0x11121011, 0x11121110, 0x11121111, 0x11121112, 0x11121210,
+    0x11121211, 0x11121212, 0x12101011, 0x12101110, 0x12101111, 0x12101211, 0x12101212, 0x12111010,
+    0x12111011, 0x12111110, 0x12111111, 0x12111112, 0x12111210, 0x12111211, 0x12121011, 0x12121110,
+    0x12121111, 0x12121112, 0x12121211, 0x10101020, 0x10101021, 0x10101022, 0x10101120, 0x10101122,
+    0x10101220, 0x10101221, 0x10111021, 0x10111120, 0x10111121, 0x10111220, 0x10111221, 0x10121020,
+    0x10121021, 0x10121022, 0x10121120, 0x10121121, 0x10121122, 0x10121220, 0x10121221, 0x11101021,
+    0x11101121, 0x11101122, 0x11101220, 0x11101221, 0x11101222, 0x11111020, 0x11111021, 0x11111022,
+    0x11111120, 0x11111121, 0x11111122, 0x11111220, 0x11111221, 0x11111222, 0x11121021, 0x11121120,
+    0x11121121, 0x11121221, 0x12101022, 0x12101121, 0x12101122, 0x12101220, 0x12101221, 0x12101222,
+    0x12111021, 0x12111121, 0x12111222, 0x12121022, 0x12121121, 0x12121122, 0x12121220, 0x12121221,
+    0x10102100, 0x10102101, 0x10102102, 0x10102201, 0x10112000, 0x10112101, 0x10112200, 0x10122001,
+    0x10122202, 0x11102101, 0x11102200, 0x11102202, 0x11112001, 0x11112100, 0x11112101, 0x11112102,
+    0x11112200, 0x11112201, 0x11122000, 0x11122002, 0x11122100, 0x11122101, 0x12102002, 0x12102201,
+    0x12112000, 0x12112002, 0x12112101, 0x12112200, 0x12122001, 0x12122201, 0x10102011, 0x10102012,
+    0x10102111, 0x10102212, 0x10112011, 0x10112110, 0x10112111, 0x10112112, 0x10112211, 0x10122111,
+    0x11102011, 0x11102110, 0x11102111, 0x11102112, 0x11102211, 0x11112010, 0x11112011, 0x11112012,
+    0x11112110, 0x11112111, 0x11112112, 0x11112210, 0x11112211, 0x11112212, 0x11122011, 0x11122110,
+    0x11122111, 0x11122112, 0x11122211, 0x12102011, 0x12102111, 0x12102211, 0x12112011, 0x12112110,
+    0x12112111, 0x12112112, 0x12112210, 0x12112211, 0x12122111, 0x10102120, 0x10102220, 0x10112121,
+    0x10112222, 0x10122020, 0x10122121, 0x10122122, 0x10122221, 0x11102121, 0x11102220, 0x11102221,
+    0x11112021, 0x11112121, 0x11112122, 0x11112220, 0x11112221, 0x11122022, 0x11122121, 0x11122220,
+    0x11122222, 0x12102021, 0x12102222, 0x12112022, 0x12112121, 0x12112122, 0x12112220, 0x12112222,
+    0x12122021, 0x10200101, 0x10210100, 0x10210102, 0x10210201, 0x10220101, 0x11200100, 0x11210000,
+    0x11210101, 0x11210102, 0x11210200, 0x11210202, 0x11220001, 0x11220100, 0x11220102, 0x11220201,
+    0x12200001, 0x12210102, 0x12220101, 0x10200011, 0x10200110, 0x10200112, 0x10200211, 0x10210012,
+    0x10210111, 0x10220011, 0x10220012, 0x10220112, 0x10220211, 0x11200111, 0x11200211, 0x11210011,
+    0x11210111, 0x11210112, 0x11210211, 0x11220111, 0x11220112, 0x11220212, 0x12200110, 0x12200212,
+    0x12210012, 0x12210111, 0x12220011, 0x12220112, 0x12220211, 0x10210021, 0x10210122, 0x10210221,
+    0x11200020, 0x11200021, 0x11200122, 0x11210121, 0x11210122, 0x11210220, 0x11220020, 0x12200121,
+    0x12210021, 0x12210122, 0x12220121, 0x10211001, 0x10211002, 0x10211101, 0x10211102, 0x10211202,
+    0x10221001, 0x10221102, 0x10221201, 0x11201000, 0x11201002, 0x11201101, 0x11201200, 0x11201202,
+    0x11211001, 0x11211100, 0x11211101, 0x11211102, 0x11211201, 0x11211202, 0x11221000, 0x11221002,
+    0x11221101, 0x12201100, 0x12201101, 0x12201201, 0x12211000, 0x12211002, 0x12211100, 0x12211101,
+    0x12211102, 0x12211200, 0x12211202, 0x12221001, 0x12221100, 0x12221201, 0x10201111, 0x10201210,
+    0x10201212, 0x10211011, 0x10211111, 0x10211112, 0x10211211, 0x11201110, 0x11201111, 0x11201112,
+    0x11201211, 0x11211010, 0x11211011, 0x11211110, 0x11211111, 0x11211112, 0x11211211, 0x11221011,
+    0x11221110, 0x11221111, 0x11221112, 0x11221211, 0x12201112, 0x12201211, 0x12201212, 0x12211011,
+    0x12211111, 0x12211112, 0x12211211, 0x12211212, 0x12221012, 0x12221111, 0x12221112, 0x12221210,
+    0x10201022, 0x10201221, 0x10211121, 0x10221020, 0x10221122, 0x10221220, 0x10221221, 0x11201020,
+    0x11201121, 0x11201220, 0x11201222, 0x11211021, 0x11211120, 0x11211121, 0x11211122, 0x11211220,
+    0x11211222, 0x11221020, 0x11221121, 0x11221220, 0x12201020, 0x12201022, 0x12201121, 0x12201222,
+    0x12211120, 0x12211122, 0x12211220, 0x12211221, 0x12221020, 0x12221120, 0x12221122, 0x12221222,
+    0x10212102, 0x10212201, 0x10222101, 0x11202001, 0x11212002, 0x11212101, 0x11212202, 0x11222001,
+    0x11222201, 0x12202101, 0x12212001, 0x12212200, 0x12222102, 0x10202011, 0x10202110, 0x10212010,
+    0x10212111, 0x10222011, 0x10222110, 0x10222112, 0x10222211, 0x11202010, 0x11202011, 0x11202111,
+    0x11202112, 0x11202210, 0x11212011, 0x11212110, 0x11212111, 0x11212112, 0x11212211, 0x11222010,
+    0x11222111, 0x11222212, 0x12202012, 0x12202110, 0x12202212, 0x12212111, 0x12222011, 0x12222110,
+    0x12222111, 0x12222211, 0x10212021, 0x10212122, 0x10212220, 0x11202021, 0x11202120, 0x11202221,
+    0x11212020, 0x11212121, 0x11212220, 0x11212222, 0x11222120, 0x11222121, 0x11222221, 0x12202122,
+    0x12212120, 0x12212220, 0x12212222, 0x12222122, 0x20000000, 0x20000002, 0x20000200, 0x20000202,
+    0x20020000, 0x20020002, 0x20020200, 0x20020202, 0x21000101, 0x21010000, 0x21010001, 0x21010100,
+    0x21010102, 0x21010201, 0x21020101, 0x22000000, 0x22000002, 0x22000200, 0x22000202, 0x22010101,
+    0x22020000, 0x22020002, 0x22020200, 0x22020202, 0x20000111, 0x20010011, 0x20010110, 0x20010112,
+    0x20010211, 0x20020111, 0x21000011, 0x21000110, 0x21000211, 0x21010010, 0x21010012, 0x21010111,
+    0x21010112, 0x21010210, 0x21010211, 0x21020110, 0x21020112, 0x21020211, 0x22000111, 0x22000211,
+    0x22010110, 0x22010112, 0x22010211, 0x22020111, 0x20000020, 0x20000022, 0x20000220, 0x20000222,
+    0x20010121, 0x20020020, 0x20020022, 0x20020220, 0x20020222, 0x21010021, 0x21010120, 0x21010221,
+    0x21020121, 0x22000020, 0x22000022, 0x22000220, 0x22000222, 0x22010121, 0x22020020, 0x22020022,
+    0x22020220, 0x22020222, 0x20011100, 0x20011201, 0x21001001, 0x21001100, 0x21011001, 0x21011101,
+    0x21011202, 0x21021001, 0x21021100, 0x21021201, 0x22011100, 0x22011201, 0x20001011, 0x20001211,
+    0x20011012, 0x20011111, 0x20011212, 0x20021112, 0x20021211, 0x21001010, 0x21001011, 0x21001111,
+    0x21001210, 0x21011011, 0x21011110, 0x21011111, 0x21011112, 0x21011211, 0x21011212, 0x21021111,
+    0x21021112, 0x21021210, 0x21021212, 0x22001011, 0x22001110, 0x22001112, 0x22001211, 0x22011010,
+    0x22011012, 0x22011111, 0x22011210, 0x22021112, 0x20011021, 0x20011122, 0x20011221, 0x20021121,
+    0x21001021, 0x21001120, 0x21001221, 0x21001222, 0x21011020, 0x21011121, 0x21011221, 0x21011222,
+    0x21021021, 0x21021122, 0x21021222, 0x22001121, 0x22011021, 0x22011222, 0x22021120, 0x20002000,
+    0x20002002, 0x20002200, 0x20002202, 0x20012101, 0x20022000, 0x20022002, 0x20022200, 0x20022202,
+    0x21002001, 0x21002101, 0x21012001, 0x21012100, 0x21012201, 0x21022101, 0x21022201, 0x22002000,
+    0x22002002, 0x22002200, 0x22002202, 0x22012101, 0x22022000, 0x22022002, 0x22022200, 0x22022202,
+    0x20002111, 0x20002112, 0x20012011, 0x20012110, 0x20012112, 0x20022111, 0x21002011, 0x21002110,
+    0x21002112, 0x21002211, 0x21012010, 0x21012012, 0x21012111, 0x21012212, 0x21022011, 0x21022110,
+    0x22002111, 0x22012112, 0x22012211, 0x22022111, 0x20002020, 0x20002022, 0x20002220, 0x20002222,
+    0x20012121, 0x20022020, 0x20022022, 0x20022220, 0x20022222, 0x21002121, 0x21012021, 0x21012120,
+    0x21012122, 0x22002020, 0x22002022, 0x22002220, 0x22002222, 0x22012121, 0x22022020, 0x22022022,
+    0x22022220, 0x22022222, 0x20100101, 0x20110001, 0x20110102, 0x20110200, 0x20110201, 0x20120101,
+    0x21100001, 0x21100102, 0x21100201, 0x21110101, 0x21110200, 0x21110202, 0x21120201, 0x21120202,
+    0x22100101, 0x22110001, 0x22110100, 0x22110102, 0x22110201, 0x22120101, 0x20100011, 0x20100110,
+    0x20100112, 0x20100211, 0x20110010, 0x20110111, 0x20110210, 0x20110212, 0x20120011, 0x20120110,
+    0x20120112, 0x20120211, 0x21100010, 0x21100111, 0x21110010, 0x21110011, 0x21110110, 0x21110111,
+    0x21110112, 0x21110211, 0x21120012, 0x21120111, 0x22100110, 0x22100112, 0x22110012, 0x22110111,
+    0x22110210, 0x22120011, 0x22120110, 0x22120112, 0x22120211, 0x20100121, 0x20110021, 0x20110120,
+    0x20110221, 0x20120121, 0x21100120, 0x21100122, 0x21100221, 0x21110020, 0x21110022, 0x21110121,
+    0x21110220, 0x21120122, 0x21120221, 0x22100121, 0x22110120, 0x22110122, 0x22120221, 0x20101001,
+    0x20101100, 0x20101102, 0x20111000, 0x20111101, 0x20111200, 0x20121102, 0x21101000, 0x21101202,
+    0x21111001, 0x21111100, 0x21111101, 0x21111102, 0x21111200, 0x21111201, 0x21121000, 0x21121001,
+    0x21121002, 0x21121101, 0x22101100, 0x22101102, 0x22111002, 0x22111100, 0x22111101, 0x22111200,
+    0x22121001, 0x22121201, 0x20101010, 0x20101111, 0x20101210, 0x20101212, 0x20111010, 0x20111011,
+    0x20111110, 0x20111111, 0x20111112, 0x20111211, 0x20121011, 0x20121111, 0x20121211, 0x20121212,
+    0x21101011, 0x21101110, 0x21101111, 0x21101112, 0x21101211, 0x21111010, 0x21111011, 0x21111012,
+    0x21111110, 0x21111111, 0x21111112, 0x21111210, 0x21111211, 0x21111212, 0x21121011, 0x21121110,
+    0x21121111, 0x21121112, 0x21121211, 0x22101011, 0x22101111, 0x22101210, 0x22111011, 0x22111012,
+    0x22111110, 0x22111111, 0x22111112, 0x22111211, 0x22111212, 0x22121010, 0x22121012, 0x22121111,
+    0x22121210, 0x22121212, 0x20101021, 0x20101120, 0x20111020, 0x20111121, 0x20111221, 0x20121020,
+    0x20121122, 0x20121221, 0x21101121, 0x21101220, 0x21101221, 0x21111021, 0x21111022, 0x21111121,
+    0x21111122, 0x21111221, 0x21121121, 0x21121220, 0x22101022, 0x22101120, 0x22101221, 0x22101222,
+    0x22111022, 0x22111120, 0x22111121, 0x22121120, 0x22121122, 0x22121221, 0x20102101, 0x20112102,
+    0x20112201, 0x20122101, 0x21102001, 0x21102102, 0x21112000, 0x21112002, 0x21112101, 0x21112102,
+    0x21112202, 0x21122100, 0x21122101, 0x22102101, 0x22112001, 0x22112102, 0x22112201, 0x22122101,
+    0x20102110, 0x20102112, 0x20102211, 0x20112010, 0x20112012, 0x20112111, 0x20112210, 0x20112212,
+    0x20122010, 0x20122011, 0x20122110, 0x20122112, 0x21102010, 0x21102012, 0x21102111, 0x21102210,
+    0x21102212, 0x21112011, 0x21112110, 0x21112111, 0x21112112, 0x21112211, 0x21122012, 0x21122111,
+    0x21122112, 0x21122212, 0x22102011, 0x22102110, 0x22112010, 0x22112012, 0x22112111, 0x22112212,
+    0x22122011, 0x22122112, 0x20102121, 0x20112121, 0x20122121, 0x21102120, 0x21102122, 0x21102221,
+    0x21112020, 0x21112121, 0x21112220, 0x21122021, 0x22102121, 0x22112021, 0x22112120, 0x22112121,
+    0x22112122, 0x20200000, 0x20200002, 0x20200200, 0x20200202, 0x20210101, 0x20220000, 0x20220002,
+    0x20220200, 0x20220202, 0x21200101, 0x21210001, 0x21210100, 0x21210102, 0x21210201, 0x22200000,
+    0x22200002, 0x22200200, 0x22200202, 0x22210101, 0x22220000, 0x22220002, 0x22220200, 0x22220202,
+    0x20200111, 0x20200211, 0x20210011, 0x20210110, 0x20210112, 0x20210211, 0x20210212, 0x21200112,
+    0x21200211, 0x21210011, 0x21210111, 0x21210210, 0x21210212, 0x21220011, 0x21220110, 0x22200111,
+    0x22210010, 0x22210012, 0x22210112, 0x22210211, 0x20200022, 0x20200220, 0x20200222, 0x20210020,
+    0x20210221, 0x20220022, 0x20220220, 0x20220222, 0x21200121, 0x21210021, 0x21210122, 0x21210221,
+    0x21220121, 0x22200020, 0x22200022, 0x22200220, 0x22200222, 0x22210121, 0x22220020, 0x22220022,
+    0x22220220, 0x22220222, 0x20211201, 0x20221101, 0x21201001, 0x21201100, 0x21211000, 0x21211100,
+    0x21211101, 0x21211200, 0x21211202, 0x21221001, 0x21221101, 0x21221102, 0x21221200, 0x21221201,
+    0x22201101, 0x20201112, 0x20201211, 0x20211010, 0x20211012, 0x20211111, 0x20211210, 0x20221112,
+    0x20221211, 0x21201012, 0x21201111, 0x21211011, 0x21211110, 0x21211111, 0x21211112, 0x21211211,
+    0x21221111, 0x21221212, 0x22201011, 0x22201110, 0x22201111, 0x22201112, 0x22201211, 0x22211012,
+    0x22211111, 0x22211210, 0x20201121, 0x20211021, 0x20211122, 0x20211222, 0x20221021, 0x20221121,
+    0x21201120, 0x21201122, 0x21201222, 0x21211022, 0x21211121, 0x21211122, 0x21211220, 0x21221020,
+    0x21221022, 0x22201122, 0x22211020, 0x22211121, 0x22211122, 0x22211221, 0x22221021, 0x22221120,
+    0x22221122, 0x20202000, 0x20202002, 0x20202200, 0x20202202, 0x20222000, 0x20222002, 0x20222200,
+    0x20222202, 0x21212001, 0x21212100, 0x21212102, 0x21212201, 0x22202000, 0x22202002, 0x22202200,
+    0x22202202, 0x22212101, 0x22222000, 0x22222002, 0x22222200, 0x22222202, 0x20202111, 0x20212110,
+    0x20212211, 0x20222011, 0x20222111, 0x21202011, 0x21212010, 0x21212111, 0x21212212, 0x21222011,
+    0x21222112, 0x21222211, 0x22212010, 0x22212112, 0x20202020, 0x20202022, 0x20202220, 0x20202222,
+    0x20222020, 0x20222022, 0x20222220, 0x20222222, 0x21212021, 0x21212120, 0x21212122, 0x22202020,
+    0x22202022, 0x22202220, 0x22202222, 0x22212121, 0x22222020, 0x22222022, 0x22222220, 0x22222222,
+GGML_TABLE_END()
+#endif
+
+#endif // GGML_COMMON_IMPL
+#endif // GGML_COMMON_IMPL
diff --git a/src/whisper_cpp/ggml/src/ggml-cpu-impl.h b/src/whisper_cpp/ggml/src/ggml-cpu-impl.h
new file mode 100644
index 00000000..5b45155b
--- /dev/null
+++ b/src/whisper_cpp/ggml/src/ggml-cpu-impl.h
@@ -0,0 +1,614 @@
+#pragma once
+
+// GGML CPU internal header
+
+#include "ggml.h"
+#include "ggml-impl.h"
+#include <stdlib.h> // load `stdlib.h` before other headers to work around MinGW bug: https://sourceforge.net/p/mingw-w64/bugs/192/
+//#include <stddef.h>
+#include <stdbool.h>
+#include <string.h> // memcpy
+#include <math.h>   // fabsf
+
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+#if defined(_MSC_VER)
+
+#define m512bh(p) p
+#define m512i(p) p
+
+#else
+
+#define m512bh(p) (__m512bh)(p)
+#define m512i(p) (__m512i)(p)
+
+#endif
+
+/**
+ * Converts brain16 to float32.
+ *
+ * The bfloat16 floating point format has the following structure:
+ *
+ *       ┌sign
+ *       │
+ *       │   ┌exponent
+ *       │   │
+ *       │   │      ┌mantissa
+ *       │   │      │
+ *       │┌──┴───┐┌─┴───┐
+ *     0b0000000000000000 brain16
+ *
+ * Since bf16 has the same number of exponent bits as a 32bit float,
+ * encoding and decoding numbers becomes relatively straightforward.
+ *
+ *       ┌sign
+ *       │
+ *       │   ┌exponent
+ *       │   │
+ *       │   │      ┌mantissa
+ *       │   │      │
+ *       │┌──┴───┐┌─┴───────────────────┐
+ *     0b00000000000000000000000000000000 IEEE binary32
+ *
+ * For comparison, the standard fp16 format has fewer exponent bits.
+ *
+ *       ┌sign
+ *       │
+ *       │  ┌exponent
+ *       │  │
+ *       │  │    ┌mantissa
+ *       │  │    │
+ *       │┌─┴─┐┌─┴──────┐
+ *     0b0000000000000000 IEEE binary16
+ *
+ * @see IEEE 754-2008
+ */
+static inline float ggml_compute_bf16_to_fp32(ggml_bf16_t h) {
+    union {
+        float f;
+        uint32_t i;
+    } u;
+    u.i = (uint32_t)h.bits << 16;
+    return u.f;
+}
+
+/**
+ * Converts float32 to brain16.
+ *
+ * This is binary identical with Google Brain float conversion.
+ * Floats shall round to nearest even, and NANs shall be quiet.
+ * Subnormals aren't flushed to zero, except perhaps when used.
+ * This code should vectorize nicely if using modern compilers.
+ */
+static inline ggml_bf16_t ggml_compute_fp32_to_bf16(float s) {
+    ggml_bf16_t h;
+    union {
+        float f;
+        uint32_t i;
+    } u;
+    u.f = s;
+    if ((u.i & 0x7fffffff) > 0x7f800000) { /* nan */
+        h.bits = (u.i >> 16) | 64; /* force to quiet */
+        return h;
+    }
+    h.bits = (u.i + (0x7fff + ((u.i >> 16) & 1))) >> 16;
+    return h;
+}
+
+#define GGML_FP32_TO_BF16(x) ggml_compute_fp32_to_bf16(x)
+#define GGML_BF16_TO_FP32(x) ggml_compute_bf16_to_fp32(x)
+
+// __FMA__ and __F16C__ are not defined in MSVC, however they are implied with AVX2/AVX512
+#if defined(_MSC_VER) && (defined(__AVX2__) || defined(__AVX512F__))
+#ifndef __FMA__
+#define __FMA__
+#endif
+#ifndef __F16C__
+#define __F16C__
+#endif
+#endif
+
+// __SSE3__ and __SSSE3__ are not defined in MSVC, but SSE3/SSSE3 are present when AVX/AVX2/AVX512 are available
+#if defined(_MSC_VER) && (defined(__AVX__) || defined(__AVX2__) || defined(__AVX512F__))
+#ifndef __SSE3__
+#define __SSE3__
+#endif
+#ifndef __SSSE3__
+#define __SSSE3__
+#endif
+#endif
+
+#if defined(__ARM_FEATURE_SVE)
+#include <arm_sve.h>
+#include <sys/prctl.h>
+#endif
+
+// 16-bit float
+// on Arm, we use __fp16
+// on x86, we use uint16_t
+#if defined(__ARM_NEON)
+
+// if YCM cannot find <arm_neon.h>, make a symbolic link to it, for example:
+//
+//   $ ln -sfn /Library/Developer/CommandLineTools/usr/lib/clang/13.1.6/include/arm_neon.h ./src/
+//
+#include <arm_neon.h>
+
+#ifdef _MSC_VER
+
+typedef uint16_t ggml_fp16_internal_t;
+
+#define ggml_vld1q_u32(w,x,y,z) { ((w) + ((uint64_t)(x) << 32)), ((y) + ((uint64_t)(z) << 32)) }
+
+#else
+
+typedef __fp16 ggml_fp16_internal_t;
+
+#define ggml_vld1q_u32(w,x,y,z) { (w), (x), (y), (z) }
+
+#endif // _MSC_VER
+
+#if !defined(__aarch64__)
+
+// 32-bit ARM compatibility
+
+// vaddlvq_s16
+// vpaddq_s16
+// vpaddq_s32
+// vaddvq_s32
+// vaddvq_f32
+// vmaxvq_f32
+// vcvtnq_s32_f32
+// vzip1_u8
+// vzip2_u8
+
+inline static int32_t vaddlvq_s16(int16x8_t v) {
+    int32x4_t v0 = vreinterpretq_s32_s64(vpaddlq_s32(vpaddlq_s16(v)));
+    return vgetq_lane_s32(v0, 0) + vgetq_lane_s32(v0, 2);
+}
+
+inline static int16x8_t vpaddq_s16(int16x8_t a, int16x8_t b) {
+    int16x4_t a0 = vpadd_s16(vget_low_s16(a), vget_high_s16(a));
+    int16x4_t b0 = vpadd_s16(vget_low_s16(b), vget_high_s16(b));
+    return vcombine_s16(a0, b0);
+}
+
+inline static int32x4_t vpaddq_s32(int32x4_t a, int32x4_t b) {
+    int32x2_t a0 = vpadd_s32(vget_low_s32(a), vget_high_s32(a));
+    int32x2_t b0 = vpadd_s32(vget_low_s32(b), vget_high_s32(b));
+    return vcombine_s32(a0, b0);
+}
+
+inline static int32_t vaddvq_s32(int32x4_t v) {
+    return vgetq_lane_s32(v, 0) + vgetq_lane_s32(v, 1) + vgetq_lane_s32(v, 2) + vgetq_lane_s32(v, 3);
+}
+
+inline static float vaddvq_f32(float32x4_t v) {
+    return vgetq_lane_f32(v, 0) + vgetq_lane_f32(v, 1) + vgetq_lane_f32(v, 2) + vgetq_lane_f32(v, 3);
+}
+
+inline static float vmaxvq_f32(float32x4_t v) {
+    return
+        MAX(MAX(vgetq_lane_f32(v, 0), vgetq_lane_f32(v, 1)),
+            MAX(vgetq_lane_f32(v, 2), vgetq_lane_f32(v, 3)));
+}
+
+inline static int32x4_t vcvtnq_s32_f32(float32x4_t v) {
+    int32x4_t res;
+
+    res[0] = roundf(vgetq_lane_f32(v, 0));
+    res[1] = roundf(vgetq_lane_f32(v, 1));
+    res[2] = roundf(vgetq_lane_f32(v, 2));
+    res[3] = roundf(vgetq_lane_f32(v, 3));
+
+    return res;
+}
+
+inline static uint8x8_t vzip1_u8(uint8x8_t a, uint8x8_t b) {
+    uint8x8_t res;
+
+    res[0] = a[0]; res[1] = b[0];
+    res[2] = a[1]; res[3] = b[1];
+    res[4] = a[2]; res[5] = b[2];
+    res[6] = a[3]; res[7] = b[3];
+
+    return res;
+}
+
+inline static uint8x8_t vzip2_u8(uint8x8_t a, uint8x8_t b) {
+    uint8x8_t res;
+
+    res[0] = a[4]; res[1] = b[4];
+    res[2] = a[5]; res[3] = b[5];
+    res[4] = a[6]; res[5] = b[6];
+    res[6] = a[7]; res[7] = b[7];
+
+    return res;
+}
+
+// vld1q_s16_x2
+// vld1q_u8_x2
+// vld1q_u8_x4
+// vld1q_s8_x2
+// vld1q_s8_x4
+// TODO: double-check these work correctly
+
+typedef struct ggml_int16x8x2_t {
+    int16x8_t val[2];
+} ggml_int16x8x2_t;
+
+inline static ggml_int16x8x2_t ggml_vld1q_s16_x2(const int16_t * ptr) {
+    ggml_int16x8x2_t res;
+
+    res.val[0] = vld1q_s16(ptr + 0);
+    res.val[1] = vld1q_s16(ptr + 8);
+
+    return res;
+}
+
+typedef struct ggml_uint8x16x2_t {
+    uint8x16_t val[2];
+} ggml_uint8x16x2_t;
+
+inline static ggml_uint8x16x2_t ggml_vld1q_u8_x2(const uint8_t * ptr) {
+    ggml_uint8x16x2_t res;
+
+    res.val[0] = vld1q_u8(ptr + 0);
+    res.val[1] = vld1q_u8(ptr + 16);
+
+    return res;
+}
+
+typedef struct ggml_uint8x16x4_t {
+    uint8x16_t val[4];
+} ggml_uint8x16x4_t;
+
+inline static ggml_uint8x16x4_t ggml_vld1q_u8_x4(const uint8_t * ptr) {
+    ggml_uint8x16x4_t res;
+
+    res.val[0] = vld1q_u8(ptr + 0);
+    res.val[1] = vld1q_u8(ptr + 16);
+    res.val[2] = vld1q_u8(ptr + 32);
+    res.val[3] = vld1q_u8(ptr + 48);
+
+    return res;
+}
+
+typedef struct ggml_int8x16x2_t {
+    int8x16_t val[2];
+} ggml_int8x16x2_t;
+
+inline static ggml_int8x16x2_t ggml_vld1q_s8_x2(const int8_t * ptr) {
+    ggml_int8x16x2_t res;
+
+    res.val[0] = vld1q_s8(ptr + 0);
+    res.val[1] = vld1q_s8(ptr + 16);
+
+    return res;
+}
+
+typedef struct ggml_int8x16x4_t {
+    int8x16_t val[4];
+} ggml_int8x16x4_t;
+
+inline static ggml_int8x16x4_t ggml_vld1q_s8_x4(const int8_t * ptr) {
+    ggml_int8x16x4_t res;
+
+    res.val[0] = vld1q_s8(ptr + 0);
+    res.val[1] = vld1q_s8(ptr + 16);
+    res.val[2] = vld1q_s8(ptr + 32);
+    res.val[3] = vld1q_s8(ptr + 48);
+
+    return res;
+}
+
+// NOTE: not tested
+inline static int8x16_t ggml_vqtbl1q_s8(int8x16_t a, uint8x16_t b) {
+    int8x16_t res;
+
+    res[ 0] = a[b[ 0]];
+    res[ 1] = a[b[ 1]];
+    res[ 2] = a[b[ 2]];
+    res[ 3] = a[b[ 3]];
+    res[ 4] = a[b[ 4]];
+    res[ 5] = a[b[ 5]];
+    res[ 6] = a[b[ 6]];
+    res[ 7] = a[b[ 7]];
+    res[ 8] = a[b[ 8]];
+    res[ 9] = a[b[ 9]];
+    res[10] = a[b[10]];
+    res[11] = a[b[11]];
+    res[12] = a[b[12]];
+    res[13] = a[b[13]];
+    res[14] = a[b[14]];
+    res[15] = a[b[15]];
+
+    return res;
+}
+
+// NOTE: not tested
+inline static uint8x16_t ggml_vqtbl1q_u8(uint8x16_t a, uint8x16_t b) {
+    uint8x16_t res;
+
+    res[ 0] = a[b[ 0]];
+    res[ 1] = a[b[ 1]];
+    res[ 2] = a[b[ 2]];
+    res[ 3] = a[b[ 3]];
+    res[ 4] = a[b[ 4]];
+    res[ 5] = a[b[ 5]];
+    res[ 6] = a[b[ 6]];
+    res[ 7] = a[b[ 7]];
+    res[ 8] = a[b[ 8]];
+    res[ 9] = a[b[ 9]];
+    res[10] = a[b[10]];
+    res[11] = a[b[11]];
+    res[12] = a[b[12]];
+    res[13] = a[b[13]];
+    res[14] = a[b[14]];
+    res[15] = a[b[15]];
+
+    return res;
+}
+
+#else
+
+#define ggml_int16x8x2_t  int16x8x2_t
+#define ggml_uint8x16x2_t uint8x16x2_t
+#define ggml_uint8x16x4_t uint8x16x4_t
+#define ggml_int8x16x2_t  int8x16x2_t
+#define ggml_int8x16x4_t  int8x16x4_t
+
+#define ggml_vld1q_s16_x2 vld1q_s16_x2
+#define ggml_vld1q_u8_x2  vld1q_u8_x2
+#define ggml_vld1q_u8_x4  vld1q_u8_x4
+#define ggml_vld1q_s8_x2  vld1q_s8_x2
+#define ggml_vld1q_s8_x4  vld1q_s8_x4
+#define ggml_vqtbl1q_s8   vqtbl1q_s8
+#define ggml_vqtbl1q_u8   vqtbl1q_u8
+
+#endif // !defined(__aarch64__)
+
+#if !defined(__ARM_FEATURE_DOTPROD)
+
+inline static int32x4_t ggml_vdotq_s32(int32x4_t acc, int8x16_t a, int8x16_t b) {
+    const int16x8_t p0 = vmull_s8(vget_low_s8 (a), vget_low_s8 (b));
+    const int16x8_t p1 = vmull_s8(vget_high_s8(a), vget_high_s8(b));
+
+    return vaddq_s32(acc, vaddq_s32(vpaddlq_s16(p0), vpaddlq_s16(p1)));
+}
+
+#else
+
+#define ggml_vdotq_s32(a, b, c) vdotq_s32(a, b, c)
+
+#endif // !defined(__ARM_FEATURE_DOTPROD)
+
+#endif // defined(__ARM_NEON)
+
+#if defined(__ARM_NEON) && !defined(_MSC_VER)
+
+#define GGML_COMPUTE_FP16_TO_FP32(x) ggml_compute_fp16_to_fp32(x)
+#define GGML_COMPUTE_FP32_TO_FP16(x) ggml_compute_fp32_to_fp16(x)
+
+#define GGML_FP16_TO_FP32(x) ggml_compute_fp16_to_fp32(x)
+
+static inline float ggml_compute_fp16_to_fp32(ggml_fp16_t h) {
+    ggml_fp16_internal_t tmp;
+    memcpy(&tmp, &h, sizeof(ggml_fp16_t));
+    return (float)tmp;
+}
+
+static inline ggml_fp16_t ggml_compute_fp32_to_fp16(float f) {
+    ggml_fp16_t res;
+    ggml_fp16_internal_t tmp = f;
+    memcpy(&res, &tmp, sizeof(ggml_fp16_t));
+    return res;
+}
+
+#else
+
+#ifdef __wasm_simd128__
+#include <wasm_simd128.h>
+#else
+#ifdef __POWER9_VECTOR__
+#include <altivec.h>
+#undef bool
+#define bool _Bool
+#else
+#if defined(_MSC_VER) || defined(__MINGW32__)
+#include <intrin.h>
+#else
+#if defined(__AVX__) || defined(__AVX2__) || defined(__AVX512F__) || defined(__SSSE3__) || defined(__SSE3__) || defined(__SSE__)
+#if !defined(__riscv)
+#include <immintrin.h>
+#endif
+#endif
+#endif
+#endif
+#endif
+
+#ifdef __riscv_v_intrinsic
+#include <riscv_vector.h>
+#endif
+
+#if defined(__loongarch64)
+#if defined(__loongarch_asx)
+#include <lasxintrin.h>
+#endif
+#if defined(__loongarch_sx)
+#include <lsxintrin.h>
+#endif
+#endif
+
+#if defined(__loongarch_asx)
+
+typedef union {
+    int32_t i;
+    float f;
+} ft_union;
+
+/* float type data load instructions */
+static __m128 __lsx_vreplfr2vr_s(float val) {
+    ft_union fi_tmpval = {.f = val};
+    return (__m128)__lsx_vreplgr2vr_w(fi_tmpval.i);
+}
+
+static __m256 __lasx_xvreplfr2vr_s(float val) {
+    ft_union fi_tmpval = {.f = val};
+    return (__m256)__lasx_xvreplgr2vr_w(fi_tmpval.i);
+}
+#endif
+
+#ifdef __F16C__
+
+#ifdef _MSC_VER
+#define GGML_COMPUTE_FP16_TO_FP32(x) _mm_cvtss_f32(_mm_cvtph_ps(_mm_cvtsi32_si128(x)))
+#define GGML_COMPUTE_FP32_TO_FP16(x) _mm_extract_epi16(_mm_cvtps_ph(_mm_set_ss(x), 0), 0)
+#else
+#define GGML_COMPUTE_FP16_TO_FP32(x) _cvtsh_ss(x)
+#define GGML_COMPUTE_FP32_TO_FP16(x) _cvtss_sh(x, 0)
+#endif
+
+#elif defined(__POWER9_VECTOR__)
+
+#define GGML_COMPUTE_FP16_TO_FP32(x) ggml_compute_fp16_to_fp32(x)
+#define GGML_COMPUTE_FP32_TO_FP16(x) ggml_compute_fp32_to_fp16(x)
+/* the inline asm below is about 12% faster than the lookup method */
+#define GGML_FP16_TO_FP32(x) GGML_COMPUTE_FP16_TO_FP32(x)
+#define GGML_FP32_TO_FP16(x) GGML_COMPUTE_FP32_TO_FP16(x)
+
+static inline float ggml_compute_fp16_to_fp32(ggml_fp16_t h) {
+    register float f;
+    register double d;
+    __asm__(
+        "mtfprd %0,%2\n"
+        "xscvhpdp %0,%0\n"
+        "frsp %1,%0\n" :
+        /* temp */ "=d"(d),
+        /* out */  "=f"(f):
+        /* in */   "r"(h));
+    return f;
+}
+
+static inline ggml_fp16_t ggml_compute_fp32_to_fp16(float f) {
+    register double d;
+    register ggml_fp16_t r;
+    __asm__( /* xscvdphp can work on double or single precision */
+        "xscvdphp %0,%2\n"
+        "mffprd %1,%0\n" :
+        /* temp */ "=d"(d),
+        /* out */  "=r"(r):
+        /* in */   "f"(f));
+    return r;
+}
+
+#else
+
+// FP16 <-> FP32
+// ref: https://github.com/Maratyszcza/FP16
+
+static inline float fp32_from_bits(uint32_t w) {
+    union {
+        uint32_t as_bits;
+        float as_value;
+    } fp32;
+    fp32.as_bits = w;
+    return fp32.as_value;
+}
+
+static inline uint32_t fp32_to_bits(float f) {
+    union {
+        float as_value;
+        uint32_t as_bits;
+    } fp32;
+    fp32.as_value = f;
+    return fp32.as_bits;
+}
+
+static inline float ggml_compute_fp16_to_fp32(ggml_fp16_t h) {
+    const uint32_t w = (uint32_t) h << 16;
+    const uint32_t sign = w & UINT32_C(0x80000000);
+    const uint32_t two_w = w + w;
+
+    const uint32_t exp_offset = UINT32_C(0xE0) << 23;
+#if defined(__STDC_VERSION__) && (__STDC_VERSION__ >= 199901L) || defined(__GNUC__) && !defined(__STRICT_ANSI__)
+    const float exp_scale = 0x1.0p-112f;
+#else
+    const float exp_scale = fp32_from_bits(UINT32_C(0x7800000));
+#endif
+    const float normalized_value = fp32_from_bits((two_w >> 4) + exp_offset) * exp_scale;
+
+    const uint32_t magic_mask = UINT32_C(126) << 23;
+    const float magic_bias = 0.5f;
+    const float denormalized_value = fp32_from_bits((two_w >> 17) | magic_mask) - magic_bias;
+
+    const uint32_t denormalized_cutoff = UINT32_C(1) << 27;
+    const uint32_t result = sign |
+        (two_w < denormalized_cutoff ? fp32_to_bits(denormalized_value) : fp32_to_bits(normalized_value));
+    return fp32_from_bits(result);
+}
+
+static inline ggml_fp16_t ggml_compute_fp32_to_fp16(float f) {
+#if defined(__STDC_VERSION__) && (__STDC_VERSION__ >= 199901L) || defined(__GNUC__) && !defined(__STRICT_ANSI__)
+    const float scale_to_inf = 0x1.0p+112f;
+    const float scale_to_zero = 0x1.0p-110f;
+#else
+    const float scale_to_inf = fp32_from_bits(UINT32_C(0x77800000));
+    const float scale_to_zero = fp32_from_bits(UINT32_C(0x08800000));
+#endif
+    float base = (fabsf(f) * scale_to_inf) * scale_to_zero;
+
+    const uint32_t w = fp32_to_bits(f);
+    const uint32_t shl1_w = w + w;
+    const uint32_t sign = w & UINT32_C(0x80000000);
+    uint32_t bias = shl1_w & UINT32_C(0xFF000000);
+    if (bias < UINT32_C(0x71000000)) {
+        bias = UINT32_C(0x71000000);
+    }
+
+    base = fp32_from_bits((bias >> 1) + UINT32_C(0x07800000)) + base;
+    const uint32_t bits = fp32_to_bits(base);
+    const uint32_t exp_bits = (bits >> 13) & UINT32_C(0x00007C00);
+    const uint32_t mantissa_bits = bits & UINT32_C(0x00000FFF);
+    const uint32_t nonsign = exp_bits + mantissa_bits;
+    return (sign >> 16) | (shl1_w > UINT32_C(0xFF000000) ? UINT16_C(0x7E00) : nonsign);
+}
+
+#define GGML_COMPUTE_FP16_TO_FP32(x) ggml_compute_fp16_to_fp32(x)
+#define GGML_COMPUTE_FP32_TO_FP16(x) ggml_compute_fp32_to_fp16(x)
+
+#endif // __F16C__
+
+#endif // defined(__ARM_NEON) && (!defined(__MSC_VER)
+
+#ifdef __ARM_FEATURE_SVE
+#include <arm_sve.h>
+#endif // __ARM_FEATURE_SVE
+
+// precomputed f32 table for f16 (256 KB)
+// defined in ggml.c, initialized in ggml_init()
+extern float ggml_table_f32_f16[1 << 16];
+
+// On ARM NEON, it's quicker to directly convert x -> x instead of calling into ggml_lookup_fp16_to_fp32,
+// so we define GGML_FP16_TO_FP32 and GGML_FP32_TO_FP16 elsewhere for NEON.
+// This is also true for POWER9.
+#if !defined(GGML_FP16_TO_FP32)
+inline static float ggml_lookup_fp16_to_fp32(ggml_fp16_t f) {
+    uint16_t s;
+    memcpy(&s, &f, sizeof(uint16_t));
+    return ggml_table_f32_f16[s];
+}
+
+#define GGML_FP16_TO_FP32(x) ggml_lookup_fp16_to_fp32(x)
+#endif
+
+#if !defined(GGML_FP32_TO_FP16)
+#define GGML_FP32_TO_FP16(x) GGML_COMPUTE_FP32_TO_FP16(x)
+#endif
+
+#ifdef __cplusplus
+}
+#endif
diff --git a/src/whisper_cpp/ggml/src/ggml-cuda.cu b/src/whisper_cpp/ggml/src/ggml-cuda.cu
new file mode 100644
index 00000000..5b6f605b
--- /dev/null
+++ b/src/whisper_cpp/ggml/src/ggml-cuda.cu
@@ -0,0 +1,3357 @@
+#include "ggml-cuda.h"
+#include "ggml-impl.h"
+#include "ggml-backend-impl.h"
+
+#include "ggml-cuda/common.cuh"
+#include "ggml-cuda/acc.cuh"
+#include "ggml-cuda/arange.cuh"
+#include "ggml-cuda/argmax.cuh"
+#include "ggml-cuda/argsort.cuh"
+#include "ggml-cuda/binbcast.cuh"
+#include "ggml-cuda/clamp.cuh"
+#include "ggml-cuda/concat.cuh"
+#include "ggml-cuda/conv-transpose-1d.cuh"
+#include "ggml-cuda/convert.cuh"
+#include "ggml-cuda/count-equal.cuh"
+#include "ggml-cuda/cpy.cuh"
+#include "ggml-cuda/cross-entropy-loss.cuh"
+#include "ggml-cuda/diagmask.cuh"
+#include "ggml-cuda/dmmv.cuh"
+#include "ggml-cuda/fattn.cuh"
+#include "ggml-cuda/getrows.cuh"
+#include "ggml-cuda/im2col.cuh"
+#include "ggml-cuda/mmq.cuh"
+#include "ggml-cuda/mmvq.cuh"
+#include "ggml-cuda/norm.cuh"
+#include "ggml-cuda/opt-step-adamw.cuh"
+#include "ggml-cuda/out-prod.cuh"
+#include "ggml-cuda/pad.cuh"
+#include "ggml-cuda/pool2d.cuh"
+#include "ggml-cuda/quantize.cuh"
+#include "ggml-cuda/rope.cuh"
+#include "ggml-cuda/scale.cuh"
+#include "ggml-cuda/softmax.cuh"
+#include "ggml-cuda/sum.cuh"
+#include "ggml-cuda/sumrows.cuh"
+#include "ggml-cuda/tsembd.cuh"
+#include "ggml-cuda/unary.cuh"
+#include "ggml-cuda/upscale.cuh"
+#include "ggml-cuda/rwkv-wkv.cuh"
+
+#include <algorithm>
+#include <array>
+#include <atomic>
+#include <cinttypes>
+#include <cstddef>
+#include <cstdint>
+#include <float.h>
+#include <limits>
+#include <map>
+#include <memory>
+#include <mutex>
+#include <stdint.h>
+#include <stdio.h>
+#include <stdarg.h>
+#include <stdlib.h>
+#include <string>
+#include <vector>
+
+static_assert(sizeof(half) == sizeof(ggml_fp16_t), "wrong fp16 size");
+
+[[noreturn]]
+void ggml_cuda_error(const char * stmt, const char * func, const char * file, int line, const char * msg) {
+    int id = -1; // in case cudaGetDevice fails
+    cudaGetDevice(&id);
+
+    GGML_LOG_ERROR(GGML_CUDA_NAME " error: %s\n", msg);
+    GGML_LOG_ERROR("  current device: %d, in function %s at %s:%d\n", id, func, file, line);
+    GGML_LOG_ERROR("  %s\n", stmt);
+    // abort with GGML_ABORT to get a stack trace
+    GGML_ABORT(GGML_CUDA_NAME " error");
+}
+
+// this is faster on Windows
+// probably because the Windows CUDA libraries forget to make this check before invoking the drivers
+void ggml_cuda_set_device(int device) {
+    int current_device;
+    CUDA_CHECK(cudaGetDevice(&current_device));
+
+    if (device == current_device) {
+        return;
+    }
+
+    CUDA_CHECK(cudaSetDevice(device));
+}
+
+int ggml_cuda_get_device() {
+    int id;
+    CUDA_CHECK(cudaGetDevice(&id));
+    return id;
+}
+
+static cudaError_t ggml_cuda_device_malloc(void ** ptr, size_t size, int device) {
+    ggml_cuda_set_device(device);
+#if defined(GGML_USE_HIPBLAS) && defined(GGML_HIP_UMA)
+    auto res = hipMallocManaged(ptr, size);
+    if (res == hipSuccess) {
+        // if error we "need" to know why...
+        CUDA_CHECK(hipMemAdvise(*ptr, size, hipMemAdviseSetCoarseGrain, device));
+    }
+    return res;
+#else
+
+#if !defined(GGML_USE_HIPBLAS)
+    cudaError_t err;
+    if (getenv("GGML_CUDA_ENABLE_UNIFIED_MEMORY") != nullptr)
+    {
+        err = cudaMallocManaged(ptr, size);
+    }
+    else
+    {
+        err = cudaMalloc(ptr, size);
+    }
+    return err;
+#else
+    return cudaMalloc(ptr, size);
+#endif // !defined(GGML_USE_HIPBLAS)
+
+#endif
+}
+
+static ggml_cuda_device_info ggml_cuda_init() {
+#ifdef __HIP_PLATFORM_AMD__
+    // Workaround for a rocBLAS bug when using multiple graphics cards:
+    // https://github.com/ROCmSoftwarePlatform/rocBLAS/issues/1346
+    rocblas_initialize();
+    CUDA_CHECK(cudaDeviceSynchronize());
+#endif
+
+    ggml_cuda_device_info info = {};
+
+    cudaError_t err = cudaGetDeviceCount(&info.device_count);
+    if (err != cudaSuccess) {
+        GGML_LOG_ERROR("%s: failed to initialize " GGML_CUDA_NAME ": %s\n", __func__, cudaGetErrorString(err));
+        return info;
+    }
+
+    GGML_ASSERT(info.device_count <= GGML_CUDA_MAX_DEVICES);
+
+    int64_t total_vram = 0;
+#ifdef GGML_CUDA_FORCE_MMQ
+    GGML_LOG_INFO("%s: GGML_CUDA_FORCE_MMQ:    yes\n", __func__);
+#else
+    GGML_LOG_INFO("%s: GGML_CUDA_FORCE_MMQ:    no\n", __func__);
+#endif // GGML_CUDA_FORCE_MMQ
+#ifdef GGML_CUDA_FORCE_CUBLAS
+    GGML_LOG_INFO("%s: GGML_CUDA_FORCE_CUBLAS: yes\n", __func__);
+#else
+    GGML_LOG_INFO("%s: GGML_CUDA_FORCE_CUBLAS: no\n", __func__);
+#endif // GGML_CUDA_FORCE_CUBLAS
+    GGML_LOG_INFO("%s: found %d " GGML_CUDA_NAME " devices:\n", __func__, info.device_count);
+    for (int id = 0; id < info.device_count; ++id) {
+        int device_vmm = 0;
+
+#if !defined(GGML_USE_HIPBLAS) && !defined(GGML_CUDA_NO_VMM)
+        CUdevice device;
+        CU_CHECK(cuDeviceGet(&device, id));
+        CU_CHECK(cuDeviceGetAttribute(&device_vmm, CU_DEVICE_ATTRIBUTE_VIRTUAL_MEMORY_MANAGEMENT_SUPPORTED, device));
+
+        if (device_vmm) {
+            CUmemAllocationProp alloc_prop = {};
+            alloc_prop.type = CU_MEM_ALLOCATION_TYPE_PINNED;
+            alloc_prop.location.type = CU_MEM_LOCATION_TYPE_DEVICE;
+            alloc_prop.location.id = id;
+            CU_CHECK(cuMemGetAllocationGranularity(&info.devices[id].vmm_granularity, &alloc_prop, CU_MEM_ALLOC_GRANULARITY_RECOMMENDED));
+        }
+#endif // !defined(GGML_USE_HIPBLAS) && !defined(GGML_CUDA_NO_VMM)
+        info.devices[id].vmm = !!device_vmm;
+
+        cudaDeviceProp prop;
+        CUDA_CHECK(cudaGetDeviceProperties(&prop, id));
+        GGML_LOG_INFO("  Device %d: %s, compute capability %d.%d, VMM: %s\n", id, prop.name, prop.major, prop.minor, device_vmm ? "yes" : "no");
+
+        info.default_tensor_split[id] = total_vram;
+        total_vram += prop.totalGlobalMem;
+
+        info.devices[id].nsm   = prop.multiProcessorCount;
+        info.devices[id].smpb  = prop.sharedMemPerBlock;
+#if defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)
+        info.devices[id].smpbo = prop.sharedMemPerBlock;
+        info.devices[id].cc = 100*prop.major + 10*prop.minor + CC_OFFSET_AMD;
+#else
+        info.devices[id].smpbo = prop.sharedMemPerBlockOptin;
+        info.devices[id].cc = 100*prop.major + 10*prop.minor;
+#endif // defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)
+    }
+
+    for (int id = 0; id < info.device_count; ++id) {
+        info.default_tensor_split[id] /= total_vram;
+    }
+
+    // configure logging to stdout
+    // CUBLAS_CHECK(cublasLoggerConfigure(1, 1, 0, nullptr));
+
+    return info;
+}
+
+const ggml_cuda_device_info & ggml_cuda_info() {
+    static ggml_cuda_device_info info = ggml_cuda_init();
+    return info;
+}
+
+// #define DEBUG_CUDA_MALLOC
+
+// buffer pool for cuda (legacy)
+struct ggml_cuda_pool_leg : public ggml_cuda_pool {
+    static const int MAX_BUFFERS = 256;
+
+    int device;
+    struct ggml_cuda_buffer {
+        void * ptr = nullptr;
+        size_t size = 0;
+    };
+
+    ggml_cuda_buffer buffer_pool[MAX_BUFFERS] = {};
+    size_t pool_size = 0;
+
+    explicit ggml_cuda_pool_leg(int device) :
+        device(device) {
+    }
+
+    ~ggml_cuda_pool_leg() {
+        ggml_cuda_set_device(device);
+        for (int i = 0; i < MAX_BUFFERS; ++i) {
+            ggml_cuda_buffer & b = buffer_pool[i];
+            if (b.ptr != nullptr) {
+                CUDA_CHECK(cudaFree(b.ptr));
+                pool_size -= b.size;
+            }
+        }
+        GGML_ASSERT(pool_size == 0);
+    }
+
+    void * alloc(size_t size, size_t * actual_size) override {
+#ifdef DEBUG_CUDA_MALLOC
+        int nnz = 0;
+        size_t max_size = 0;
+#endif
+        size_t best_diff = 1ull << 36;
+        int ibest = -1;
+        for (int i = 0; i < MAX_BUFFERS; ++i) {
+            ggml_cuda_buffer& b = buffer_pool[i];
+            if (b.ptr != nullptr) {
+#ifdef DEBUG_CUDA_MALLOC
+                ++nnz;
+                if (b.size > max_size) max_size = b.size;
+#endif
+                if (b.size >= size) {
+                    size_t diff = b.size - size;
+                    if (diff < best_diff) {
+                        best_diff = diff;
+                        ibest = i;
+                        if (!best_diff) {
+                            void * ptr = b.ptr;
+                            *actual_size = b.size;
+                            b.ptr = nullptr;
+                            b.size = 0;
+                            return ptr;
+                        }
+                    }
+                }
+            }
+        }
+        if (ibest >= 0) {
+            ggml_cuda_buffer& b = buffer_pool[ibest];
+            void * ptr = b.ptr;
+            *actual_size = b.size;
+            b.ptr = nullptr;
+            b.size = 0;
+            return ptr;
+        }
+        void * ptr;
+        size_t look_ahead_size = (size_t) (1.05 * size);
+        look_ahead_size = 256 * ((look_ahead_size + 255)/256);
+        ggml_cuda_set_device(device);
+        CUDA_CHECK(ggml_cuda_device_malloc(&ptr, look_ahead_size, device));
+        *actual_size = look_ahead_size;
+        pool_size += look_ahead_size;
+#ifdef DEBUG_CUDA_MALLOC
+        GGML_LOG_INFO("%s[%d]: %d buffers, max_size = %u MB, pool_size = %u MB, requested %u MB\n", __func__, device, nnz,
+                           (uint32_t)(max_size / 1024 / 1024), (uint32_t)(pool_size / 1024 / 1024), (uint32_t)(size / 1024 / 1024));
+#endif
+        return ptr;
+    }
+
+    void free(void * ptr, size_t size) override {
+        for (int i = 0; i < MAX_BUFFERS; ++i) {
+            ggml_cuda_buffer& b = buffer_pool[i];
+            if (b.ptr == nullptr) {
+                b.ptr = ptr;
+                b.size = size;
+                return;
+            }
+        }
+        GGML_LOG_WARN(GGML_CUDA_NAME " buffer pool full, increase MAX_CUDA_BUFFERS\n");
+        ggml_cuda_set_device(device);
+        CUDA_CHECK(cudaFree(ptr));
+        pool_size -= size;
+    }
+};
+
+// pool with virtual memory
+#if !defined(GGML_USE_HIPBLAS) && !defined(GGML_CUDA_NO_VMM)
+struct ggml_cuda_pool_vmm : public ggml_cuda_pool {
+    static const size_t CUDA_POOL_VMM_MAX_SIZE = 1ull << 35; // 32 GB
+
+    int device;
+    CUdeviceptr pool_addr = 0;
+    size_t pool_used = 0;
+    size_t pool_size = 0;
+    size_t granularity;
+
+    explicit ggml_cuda_pool_vmm(int device) :
+        device(device),
+        granularity(ggml_cuda_info().devices[device].vmm_granularity) {
+    }
+
+    ~ggml_cuda_pool_vmm() {
+        if (pool_addr != 0) {
+            CU_CHECK(cuMemUnmap(pool_addr, pool_size));
+            CU_CHECK(cuMemAddressFree(pool_addr, CUDA_POOL_VMM_MAX_SIZE));
+        }
+    }
+
+    void * alloc(size_t size, size_t * actual_size) override {
+        // round up the allocation size to the alignment to ensure that all allocations are aligned for all data types
+        const size_t alignment = 128;
+        size = alignment * ((size + alignment - 1) / alignment);
+
+        size_t avail = pool_size - pool_used;
+
+        if (size > avail) {
+            // round up to the next multiple of the granularity
+            size_t reserve_size = size - avail;
+            reserve_size = granularity * ((reserve_size + granularity - 1) / granularity);
+
+            GGML_ASSERT(pool_size + reserve_size <= CUDA_POOL_VMM_MAX_SIZE);
+
+            // allocate more physical memory
+            CUmemAllocationProp prop = {};
+            prop.type = CU_MEM_ALLOCATION_TYPE_PINNED;
+            prop.location.type = CU_MEM_LOCATION_TYPE_DEVICE;
+            prop.location.id = device;
+            CUmemGenericAllocationHandle handle;
+            CU_CHECK(cuMemCreate(&handle, reserve_size, &prop, 0));
+
+            // reserve virtual address space (if not already reserved)
+            if (pool_addr == 0) {
+                CU_CHECK(cuMemAddressReserve(&pool_addr, CUDA_POOL_VMM_MAX_SIZE, 0, 0, 0));
+            }
+
+            // map at the end of the pool
+            CU_CHECK(cuMemMap(pool_addr + pool_size, reserve_size, 0, handle, 0));
+
+            // the memory allocation handle is no longer needed after mapping
+            CU_CHECK(cuMemRelease(handle));
+
+            // set access
+            CUmemAccessDesc access = {};
+            access.location.type = CU_MEM_LOCATION_TYPE_DEVICE;
+            access.location.id = device;
+            access.flags = CU_MEM_ACCESS_FLAGS_PROT_READWRITE;
+            CU_CHECK(cuMemSetAccess(pool_addr + pool_size, reserve_size, &access, 1));
+
+            // add to the pool
+            pool_size += reserve_size;
+
+            //printf("cuda pool[%d]: size increased to %llu MB (reserved %llu MB)\n",
+            //       device, (unsigned long long) (pool_size/1024/1024),
+            //       (unsigned long long) (reserve_size/1024/1024));
+        }
+
+        GGML_ASSERT(pool_addr != 0);
+
+        void * ptr = (void *) (pool_addr + pool_used);
+        *actual_size = size;
+        pool_used += size;
+
+#ifdef DEBUG_CUDA_MALLOC
+        printf("cuda pool[%d]: allocated %llu bytes at %llx\n", device, (unsigned long long) size, ptr);
+#endif
+
+        return ptr;
+    }
+
+    void free(void * ptr, size_t size) override {
+#ifdef DEBUG_CUDA_MALLOC
+        printf("cuda pool[%d]: freed %llu bytes at %llx\n", device, (unsigned long long) size, ptr);
+#endif
+
+        pool_used -= size;
+
+        // all deallocations must be in reverse order of the allocations
+        GGML_ASSERT(ptr == (void *) (pool_addr + pool_used));
+    }
+};
+#endif // !defined(GGML_USE_HIPBLAS) && !defined(GGML_CUDA_NO_VMM)
+
+std::unique_ptr<ggml_cuda_pool> ggml_backend_cuda_context::new_pool_for_device(int device) {
+#if !defined(GGML_USE_HIPBLAS) && !defined(GGML_CUDA_NO_VMM)
+    if (ggml_cuda_info().devices[device].vmm) {
+        return std::unique_ptr<ggml_cuda_pool>(new ggml_cuda_pool_vmm(device));
+    }
+#endif // !defined(GGML_USE_HIPBLAS) && !defined(GGML_CUDA_NO_VMM)
+    return std::unique_ptr<ggml_cuda_pool>(new ggml_cuda_pool_leg(device));
+}
+
+// cuda buffer
+
+struct ggml_backend_cuda_buffer_context {
+    int device;
+    void * dev_ptr = nullptr;
+    std::string name;
+
+    ggml_backend_cuda_buffer_context(int device, void * dev_ptr) :
+        device(device), dev_ptr(dev_ptr),
+        name(GGML_CUDA_NAME + std::to_string(device)) {
+    }
+
+    ~ggml_backend_cuda_buffer_context() {
+        CUDA_CHECK(cudaFree(dev_ptr));
+    }
+};
+
+static const char * ggml_backend_cuda_buffer_get_name(ggml_backend_buffer_t buffer) {
+    ggml_backend_cuda_buffer_context * ctx = (ggml_backend_cuda_buffer_context *)buffer->context;
+    return ctx->name.c_str();
+}
+
+static bool ggml_backend_buffer_is_cuda(ggml_backend_buffer_t buffer) {
+    return buffer->iface.get_name == ggml_backend_cuda_buffer_get_name;
+}
+
+static void ggml_backend_cuda_buffer_free_buffer(ggml_backend_buffer_t buffer) {
+    ggml_backend_cuda_buffer_context * ctx = (ggml_backend_cuda_buffer_context *)buffer->context;
+    delete ctx;
+}
+
+static void * ggml_backend_cuda_buffer_get_base(ggml_backend_buffer_t buffer) {
+    ggml_backend_cuda_buffer_context * ctx = (ggml_backend_cuda_buffer_context *)buffer->context;
+    return ctx->dev_ptr;
+}
+
+static void ggml_backend_cuda_buffer_init_tensor(ggml_backend_buffer_t buffer, ggml_tensor * tensor) {
+    ggml_backend_cuda_buffer_context * ctx = (ggml_backend_cuda_buffer_context *)buffer->context;
+
+    if (tensor->view_src != NULL) {
+        assert(tensor->view_src->buffer->buft == buffer->buft);
+        return;
+    }
+
+    if (ggml_is_quantized(tensor->type) && tensor->view_src == nullptr && ggml_backend_buffer_get_usage(buffer) != GGML_BACKEND_BUFFER_USAGE_COMPUTE) {
+        // initialize padding to 0 to avoid possible NaN values
+        size_t original_size = ggml_nbytes(tensor);
+        size_t padded_size = ggml_backend_buft_get_alloc_size(buffer->buft, tensor);
+
+        if (padded_size > original_size) {
+            ggml_cuda_set_device(ctx->device);
+            CUDA_CHECK(cudaMemset((char *)tensor->data + original_size, 0, padded_size - original_size));
+        }
+    }
+}
+
+static void ggml_backend_cuda_buffer_memset_tensor(ggml_backend_buffer_t buffer, ggml_tensor * tensor, uint8_t value, size_t offset, size_t size) {
+    ggml_backend_cuda_buffer_context * ctx = (ggml_backend_cuda_buffer_context *)buffer->context;
+
+    ggml_cuda_set_device(ctx->device);
+    CUDA_CHECK(cudaMemsetAsync((char *)tensor->data + offset, value, size, cudaStreamPerThread));
+    CUDA_CHECK(cudaStreamSynchronize(cudaStreamPerThread));
+}
+
+static void ggml_backend_cuda_buffer_set_tensor(ggml_backend_buffer_t buffer, ggml_tensor * tensor, const void * data, size_t offset, size_t size) {
+    ggml_backend_cuda_buffer_context * ctx = (ggml_backend_cuda_buffer_context *)buffer->context;
+
+    ggml_cuda_set_device(ctx->device);
+    CUDA_CHECK(cudaMemcpyAsync((char *)tensor->data + offset, data, size, cudaMemcpyHostToDevice, cudaStreamPerThread));
+    CUDA_CHECK(cudaStreamSynchronize(cudaStreamPerThread));
+}
+
+static void ggml_backend_cuda_buffer_get_tensor(ggml_backend_buffer_t buffer, const ggml_tensor * tensor, void * data, size_t offset, size_t size) {
+    ggml_backend_cuda_buffer_context * ctx = (ggml_backend_cuda_buffer_context *)buffer->context;
+
+    ggml_cuda_set_device(ctx->device);
+    CUDA_CHECK(cudaMemcpyAsync(data, (const char *)tensor->data + offset, size, cudaMemcpyDeviceToHost, cudaStreamPerThread));
+    CUDA_CHECK(cudaStreamSynchronize(cudaStreamPerThread));
+}
+
+static bool ggml_backend_cuda_buffer_cpy_tensor(ggml_backend_buffer_t buffer, const ggml_tensor * src, ggml_tensor * dst) {
+    if (ggml_backend_buffer_is_cuda(src->buffer)) {
+        ggml_backend_cuda_buffer_context * src_ctx = (ggml_backend_cuda_buffer_context *)src->buffer->context;
+        ggml_backend_cuda_buffer_context * dst_ctx = (ggml_backend_cuda_buffer_context *)dst->buffer->context;
+        if (src_ctx->device == dst_ctx->device) {
+            CUDA_CHECK(cudaMemcpyAsync(dst->data, src->data, ggml_nbytes(src), cudaMemcpyDeviceToDevice, cudaStreamPerThread));
+        } else {
+#ifdef GGML_CUDA_NO_PEER_COPY
+            return false;
+#else
+            CUDA_CHECK(cudaMemcpyPeerAsync(dst->data, dst_ctx->device, src->data, src_ctx->device, ggml_nbytes(src), cudaStreamPerThread));
+#endif
+        }
+        CUDA_CHECK(cudaStreamSynchronize(cudaStreamPerThread));
+        return true;
+    }
+    return false;
+
+    GGML_UNUSED(buffer);
+}
+
+static void ggml_backend_cuda_buffer_clear(ggml_backend_buffer_t buffer, uint8_t value) {
+    ggml_backend_cuda_buffer_context * ctx = (ggml_backend_cuda_buffer_context *)buffer->context;
+
+    ggml_cuda_set_device(ctx->device);
+    CUDA_CHECK(cudaDeviceSynchronize());
+    CUDA_CHECK(cudaMemset(ctx->dev_ptr, value, buffer->size));
+    CUDA_CHECK(cudaDeviceSynchronize());
+}
+
+static const ggml_backend_buffer_i ggml_backend_cuda_buffer_interface = {
+    /* .get_name        = */ ggml_backend_cuda_buffer_get_name,
+    /* .free_buffer     = */ ggml_backend_cuda_buffer_free_buffer,
+    /* .get_base        = */ ggml_backend_cuda_buffer_get_base,
+    /* .init_tensor     = */ ggml_backend_cuda_buffer_init_tensor,
+    /* .memset_tensor   = */ ggml_backend_cuda_buffer_memset_tensor,
+    /* .set_tensor      = */ ggml_backend_cuda_buffer_set_tensor,
+    /* .get_tensor      = */ ggml_backend_cuda_buffer_get_tensor,
+    /* .cpy_tensor      = */ ggml_backend_cuda_buffer_cpy_tensor,
+    /* .clear           = */ ggml_backend_cuda_buffer_clear,
+    /* .reset           = */ NULL,
+};
+
+// cuda buffer type
+struct ggml_backend_cuda_buffer_type_context {
+    int device;
+    std::string name;
+};
+
+static const char * ggml_backend_cuda_buffer_type_get_name(ggml_backend_buffer_type_t buft) {
+    ggml_backend_cuda_buffer_type_context * ctx = (ggml_backend_cuda_buffer_type_context *)buft->context;
+
+    return ctx->name.c_str();
+}
+
+static bool ggml_backend_buft_is_cuda(ggml_backend_buffer_type_t buft) {
+    return buft->iface.get_name == ggml_backend_cuda_buffer_type_get_name;
+}
+
+static ggml_backend_buffer_t ggml_backend_cuda_buffer_type_alloc_buffer(ggml_backend_buffer_type_t buft, size_t size) {
+    ggml_backend_cuda_buffer_type_context * buft_ctx = (ggml_backend_cuda_buffer_type_context *)buft->context;
+
+    ggml_cuda_set_device(buft_ctx->device);
+
+    size = std::max(size, (size_t)1); // cudaMalloc returns null for size 0
+
+    void * dev_ptr;
+    cudaError_t err = ggml_cuda_device_malloc(&dev_ptr, size, buft_ctx->device);
+    if (err != cudaSuccess) {
+        // clear the error
+        cudaGetLastError();
+        GGML_LOG_ERROR("%s: allocating %.2f MiB on device %d: cudaMalloc failed: %s\n", __func__, size / 1024.0 / 1024.0, buft_ctx->device, cudaGetErrorString(err));
+        return nullptr;
+    }
+
+    ggml_backend_cuda_buffer_context * ctx = new ggml_backend_cuda_buffer_context(buft_ctx->device, dev_ptr);
+
+    return ggml_backend_buffer_init(buft, ggml_backend_cuda_buffer_interface, ctx, size);
+}
+
+static size_t ggml_backend_cuda_buffer_type_get_alignment(ggml_backend_buffer_type_t buft) {
+    return 128;
+
+    GGML_UNUSED(buft);
+}
+
+static size_t ggml_backend_cuda_buffer_type_get_alloc_size(ggml_backend_buffer_type_t buft, const ggml_tensor * tensor) {
+    size_t size = ggml_nbytes(tensor);
+    int64_t ne0 = tensor->ne[0];
+
+    if (ggml_is_quantized(tensor->type)) {
+        if (ne0 % MATRIX_ROW_PADDING != 0) {
+            size += ggml_row_size(tensor->type, MATRIX_ROW_PADDING - ne0 % MATRIX_ROW_PADDING);
+        }
+    }
+
+    return size;
+
+    GGML_UNUSED(buft);
+}
+
+static const ggml_backend_buffer_type_i ggml_backend_cuda_buffer_type_interface = {
+    /* .get_name         = */ ggml_backend_cuda_buffer_type_get_name,
+    /* .alloc_buffer     = */ ggml_backend_cuda_buffer_type_alloc_buffer,
+    /* .get_alignment    = */ ggml_backend_cuda_buffer_type_get_alignment,
+    /* .get_max_size     = */ NULL, // defaults to SIZE_MAX
+    /* .get_alloc_size   = */ ggml_backend_cuda_buffer_type_get_alloc_size,
+    /* .is_host          = */ NULL,
+};
+
+ggml_backend_buffer_type_t ggml_backend_cuda_buffer_type(int device) {
+    static std::mutex mutex;
+    std::lock_guard<std::mutex> lock(mutex);
+
+    if (device >= ggml_backend_cuda_get_device_count()) {
+        return nullptr;
+    }
+
+    static ggml_backend_buffer_type ggml_backend_cuda_buffer_types[GGML_CUDA_MAX_DEVICES];
+
+    static bool ggml_backend_cuda_buffer_type_initialized = false;
+
+    if (!ggml_backend_cuda_buffer_type_initialized) {
+        for (int i = 0; i < ggml_backend_cuda_get_device_count(); i++) {
+            ggml_backend_cuda_buffer_types[i] = {
+                /* .iface    = */ ggml_backend_cuda_buffer_type_interface,
+                /* .device   = */ ggml_backend_reg_dev_get(ggml_backend_cuda_reg(), i),
+                /* .context  = */ new ggml_backend_cuda_buffer_type_context{i, GGML_CUDA_NAME + std::to_string(i)},
+            };
+        }
+        ggml_backend_cuda_buffer_type_initialized = true;
+    }
+
+    return &ggml_backend_cuda_buffer_types[device];
+}
+
+// cuda split buffer
+
+static int64_t get_row_rounding(const std::array<float, GGML_CUDA_MAX_DEVICES> & tensor_split) {
+    int64_t row_rounding = 0;
+    for (int id = 0; id < ggml_backend_cuda_get_device_count(); ++id) {
+        if (tensor_split[id] >= (id + 1 < ggml_backend_cuda_get_device_count() ? tensor_split[id + 1] : 1.0f)) {
+            continue;
+        }
+
+        const int cc = ggml_cuda_info().devices[id].cc;
+        row_rounding = std::max(row_rounding, (int64_t)get_mmq_y_host(cc));
+    }
+    return row_rounding;
+}
+
+static void get_row_split(int64_t * row_low, int64_t * row_high, const ggml_tensor * tensor, const std::array<float, GGML_CUDA_MAX_DEVICES> & tensor_split, int id) {
+    const int64_t nrows = ggml_nrows(tensor);
+    const int64_t rounding = get_row_rounding(tensor_split);
+
+    *row_low = id == 0 ? 0 : nrows*tensor_split[id];
+    *row_low -= *row_low % rounding;
+
+    if (id == ggml_backend_cuda_get_device_count() - 1) {
+        *row_high = nrows;
+    } else {
+        *row_high = nrows*tensor_split[id + 1];
+        *row_high -= *row_high % rounding;
+    }
+}
+
+static size_t ggml_nbytes_split(const struct ggml_tensor * tensor, int nrows_split) {
+    static_assert(GGML_MAX_DIMS == 4, "GGML_MAX_DIMS is not 4 - update this function");
+
+    return nrows_split*ggml_row_size(tensor->type, tensor->ne[0]);
+}
+
+struct ggml_backend_cuda_split_buffer_type_context {
+    std::array<float, GGML_CUDA_MAX_DEVICES> tensor_split;
+};
+
+struct ggml_backend_cuda_split_buffer_context {
+    ~ggml_backend_cuda_split_buffer_context() {
+        for (ggml_tensor_extra_gpu * extra : tensor_extras) {
+            for (int id = 0; id < GGML_CUDA_MAX_DEVICES; ++id) {
+                for (int64_t is = 0; is < GGML_CUDA_MAX_STREAMS; ++is) {
+                    if (extra->events[id][is] != nullptr) {
+                        CUDA_CHECK(cudaEventDestroy(extra->events[id][is]));
+                    }
+                }
+                if (extra->data_device[id] != nullptr) {
+                    CUDA_CHECK(cudaFree(extra->data_device[id]));
+                }
+            }
+            delete extra;
+        }
+    }
+
+    std::vector<ggml_tensor_extra_gpu *> tensor_extras;
+};
+
+static const char * ggml_backend_cuda_split_buffer_get_name(ggml_backend_buffer_t buffer) {
+    return GGML_CUDA_NAME "_Split";
+
+    GGML_UNUSED(buffer);
+}
+
+static bool ggml_backend_buffer_is_cuda_split(ggml_backend_buffer_t buffer) {
+    return buffer->iface.get_name == ggml_backend_cuda_split_buffer_get_name;
+    GGML_UNUSED(ggml_backend_buffer_is_cuda_split); // only used in debug builds currently, avoid unused function warning in release builds
+}
+
+static void ggml_backend_cuda_split_buffer_free_buffer(ggml_backend_buffer_t buffer) {
+    ggml_backend_cuda_split_buffer_context * ctx = (ggml_backend_cuda_split_buffer_context *)buffer->context;
+    delete ctx;
+}
+
+static void * ggml_backend_cuda_split_buffer_get_base(ggml_backend_buffer_t buffer) {
+    // the pointers are stored in the tensor extras, this is just a dummy address and never dereferenced
+    return (void *)0x1000;
+
+    GGML_UNUSED(buffer);
+}
+
+static void ggml_backend_cuda_split_buffer_init_tensor(ggml_backend_buffer_t buffer, ggml_tensor * tensor) {
+    GGML_ASSERT(tensor->view_src == nullptr); // views of split tensors are not supported
+
+    ggml_backend_cuda_split_buffer_context * ctx = (ggml_backend_cuda_split_buffer_context *)buffer->context;
+    ggml_backend_cuda_split_buffer_type_context * buft_ctx = (ggml_backend_cuda_split_buffer_type_context *)buffer->buft->context;
+
+    const int64_t ne0 = tensor->ne[0];
+
+    ggml_tensor_extra_gpu * extra = new ggml_tensor_extra_gpu{};
+    ctx->tensor_extras.push_back(extra);
+
+    for (int id = 0; id < ggml_backend_cuda_get_device_count(); ++id) {
+        int64_t row_low, row_high;
+        get_row_split(&row_low, &row_high, tensor, buft_ctx->tensor_split, id);
+
+        int64_t nrows_split = row_high - row_low;
+        if (nrows_split == 0) {
+            continue;
+        }
+
+        size_t size = ggml_nbytes_split(tensor, nrows_split);
+        const size_t original_size = size;
+
+        // pad last row to a multiple of 512 elements to avoid out-of-bounds memory accesses
+        if (ne0 % MATRIX_ROW_PADDING != 0) {
+            size += ggml_row_size(tensor->type, MATRIX_ROW_PADDING - ne0 % MATRIX_ROW_PADDING);
+        }
+
+        // FIXME: do not crash if cudaMalloc fails
+        // currently, init_tensor cannot fail, it needs to be fixed in ggml-backend first
+        ggml_cuda_set_device(id);
+        char * buf;
+        CUDA_CHECK(ggml_cuda_device_malloc((void**)&buf, size, id));
+
+        // set padding to 0 to avoid possible NaN values
+        if (size > original_size) {
+            CUDA_CHECK(cudaMemset(buf + original_size, 0, size - original_size));
+        }
+
+        extra->data_device[id] = buf;
+
+        for (int64_t is = 0; is < GGML_CUDA_MAX_STREAMS; ++is) {
+            CUDA_CHECK(cudaEventCreateWithFlags(&extra->events[id][is], cudaEventDisableTiming));
+        }
+    }
+    tensor->extra = extra;
+}
+
+static void ggml_backend_cuda_split_buffer_set_tensor(ggml_backend_buffer_t buffer, ggml_tensor * tensor, const void * data, size_t offset, size_t size) {
+    // split tensors must always be set in their entirety at once
+    GGML_ASSERT(offset == 0);
+    GGML_ASSERT(size == ggml_nbytes(tensor));
+
+    ggml_backend_cuda_split_buffer_type_context * buft_ctx = (ggml_backend_cuda_split_buffer_type_context *)buffer->buft->context;
+
+    const int64_t ne0 = tensor->ne[0];
+    const size_t nb1 = tensor->nb[1];
+    ggml_tensor_extra_gpu * extra = (ggml_tensor_extra_gpu *)tensor->extra;
+
+    for (int id = 0; id < ggml_backend_cuda_get_device_count(); ++id) {
+        int64_t row_low, row_high;
+        get_row_split(&row_low, &row_high, tensor, buft_ctx->tensor_split, id);
+
+        int64_t nrows_split = row_high - row_low;
+        if (nrows_split == 0) {
+            continue;
+        }
+
+        const size_t offset_split = row_low*nb1;
+        size_t size = ggml_nbytes_split(tensor, nrows_split);
+        const size_t original_size = size;
+
+        // pad last row to a multiple of 512 elements to avoid out-of-bounds memory accesses
+        if (ne0 % MATRIX_ROW_PADDING != 0) {
+            size += ggml_row_size(tensor->type, MATRIX_ROW_PADDING - ne0 % MATRIX_ROW_PADDING);
+        }
+
+        const char * buf_host = (const char *)data + offset_split;
+        CUDA_CHECK(cudaMemcpyAsync(extra->data_device[id], buf_host, original_size, cudaMemcpyHostToDevice, cudaStreamPerThread));
+    }
+
+    for (int id = 0; id < ggml_backend_cuda_get_device_count(); ++id) {
+        CUDA_CHECK(cudaStreamSynchronize(cudaStreamPerThread));
+    }
+}
+
+static void ggml_backend_cuda_split_buffer_get_tensor(ggml_backend_buffer_t buffer, const ggml_tensor * tensor, void * data, size_t offset, size_t size) {
+    // split tensors must always be set in their entirety at once
+    GGML_ASSERT(offset == 0);
+    GGML_ASSERT(size == ggml_nbytes(tensor));
+
+    ggml_backend_cuda_split_buffer_type_context * buft_ctx = (ggml_backend_cuda_split_buffer_type_context *)buffer->buft->context;
+
+    const int64_t ne0 = tensor->ne[0];
+    const size_t nb1 = tensor->nb[1];
+    ggml_tensor_extra_gpu * extra = (ggml_tensor_extra_gpu *)tensor->extra;
+
+    for (int id = 0; id < ggml_backend_cuda_get_device_count(); ++id) {
+        int64_t row_low, row_high;
+        get_row_split(&row_low, &row_high, tensor, buft_ctx->tensor_split, id);
+
+        int64_t nrows_split = row_high - row_low;
+        if (nrows_split == 0) {
+            continue;
+        }
+
+        const size_t offset_split = row_low*nb1;
+        size_t size = ggml_nbytes_split(tensor, nrows_split);
+        const size_t original_size = size;
+
+        // pad last row to a multiple of 512 elements to avoid out-of-bounds memory accesses
+        if (ne0 % MATRIX_ROW_PADDING != 0) {
+            size += ggml_row_size(tensor->type, MATRIX_ROW_PADDING - ne0 % MATRIX_ROW_PADDING);
+        }
+
+        char * buf_host = (char *)data + offset_split;
+        CUDA_CHECK(cudaMemcpyAsync(buf_host, extra->data_device[id], original_size, cudaMemcpyDeviceToHost, cudaStreamPerThread));
+    }
+
+    for (int id = 0; id < ggml_backend_cuda_get_device_count(); ++id) {
+        CUDA_CHECK(cudaStreamSynchronize(cudaStreamPerThread));
+    }
+}
+
+static void ggml_backend_cuda_split_buffer_clear(ggml_backend_buffer_t buffer, uint8_t value) {
+    GGML_UNUSED(buffer);
+    GGML_UNUSED(value);
+}
+
+static const ggml_backend_buffer_i ggml_backend_cuda_split_buffer_interface = {
+    /* .get_name        = */ ggml_backend_cuda_split_buffer_get_name,
+    /* .free_buffer     = */ ggml_backend_cuda_split_buffer_free_buffer,
+    /* .get_base        = */ ggml_backend_cuda_split_buffer_get_base,
+    /* .init_tensor     = */ ggml_backend_cuda_split_buffer_init_tensor,
+    /* .memset_tensor   = */ NULL,
+    /* .set_tensor      = */ ggml_backend_cuda_split_buffer_set_tensor,
+    /* .get_tensor      = */ ggml_backend_cuda_split_buffer_get_tensor,
+    /* .cpy_tensor      = */ NULL,
+    /* .clear           = */ ggml_backend_cuda_split_buffer_clear,
+    /* .reset           = */ NULL,
+};
+
+// cuda split buffer type
+
+static const char * ggml_backend_cuda_split_buffer_type_get_name(ggml_backend_buffer_type_t buft) {
+    return GGML_CUDA_NAME "_Split";
+
+    GGML_UNUSED(buft);
+}
+
+static bool ggml_backend_buft_is_cuda_split(ggml_backend_buffer_type_t buft) {
+    return buft->iface.get_name == ggml_backend_cuda_split_buffer_type_get_name;
+}
+
+static ggml_backend_buffer_t ggml_backend_cuda_split_buffer_type_alloc_buffer(ggml_backend_buffer_type_t buft, size_t size) {
+    // since we don't know the exact split after rounding, we cannot allocate the device buffers at this point
+    // instead, we allocate them for each tensor separately in init_tensor
+    // however, the size still represents the maximum cumulative size of all the device buffers after the tensors are allocated,
+    // as returned by get_alloc_size. this limit is enforced during tensor allocation by ggml-alloc, so it must be correct.
+    ggml_backend_cuda_split_buffer_context * ctx = new ggml_backend_cuda_split_buffer_context();
+
+    return ggml_backend_buffer_init(buft, ggml_backend_cuda_split_buffer_interface, ctx, size);
+}
+
+static size_t ggml_backend_cuda_split_buffer_type_get_alignment(ggml_backend_buffer_type_t buft) {
+    return 128;
+
+    GGML_UNUSED(buft);
+}
+
+static size_t ggml_backend_cuda_split_buffer_type_get_alloc_size(ggml_backend_buffer_type_t buft, const ggml_tensor * tensor) {
+    ggml_backend_cuda_split_buffer_type_context * ctx = (ggml_backend_cuda_split_buffer_type_context *)buft->context;
+
+    size_t total_size = 0;
+
+    const int64_t ne0 = tensor->ne[0];
+
+    for (int id = 0; id < ggml_backend_cuda_get_device_count(); ++id) {
+        int64_t row_low, row_high;
+        get_row_split(&row_low, &row_high, tensor, ctx->tensor_split, id);
+
+        int64_t nrows_split = row_high - row_low;
+        if (nrows_split == 0) {
+            continue;
+        }
+
+        total_size += ggml_nbytes_split(tensor, nrows_split);
+
+        // pad last row to a multiple of 512 elements to avoid out-of-bounds memory accesses
+        if (ne0 % MATRIX_ROW_PADDING != 0) {
+            total_size += ggml_row_size(tensor->type, MATRIX_ROW_PADDING - ne0 % MATRIX_ROW_PADDING);
+        }
+    }
+
+    return total_size;
+}
+
+static bool ggml_backend_cuda_split_buffer_type_is_host(ggml_backend_buffer_type_t buft) {
+    return false;
+
+    GGML_UNUSED(buft);
+}
+
+static const ggml_backend_buffer_type_i ggml_backend_cuda_split_buffer_type_interface = {
+    /* .get_name         = */ ggml_backend_cuda_split_buffer_type_get_name,
+    /* .alloc_buffer     = */ ggml_backend_cuda_split_buffer_type_alloc_buffer,
+    /* .get_alignment    = */ ggml_backend_cuda_split_buffer_type_get_alignment,
+    /* .get_max_size     = */ NULL, // defaults to SIZE_MAX
+    /* .get_alloc_size   = */ ggml_backend_cuda_split_buffer_type_get_alloc_size,
+    /* .is_host          = */ ggml_backend_cuda_split_buffer_type_is_host,
+};
+
+ggml_backend_buffer_type_t ggml_backend_cuda_split_buffer_type(const float * tensor_split) {
+    static std::mutex mutex;
+    std::lock_guard<std::mutex> lock(mutex);
+
+    static std::map<std::array<float, GGML_CUDA_MAX_DEVICES>, struct ggml_backend_buffer_type> buft_map;
+
+    std::array<float, GGML_CUDA_MAX_DEVICES> tensor_split_arr = {};
+
+    bool all_zero = tensor_split == nullptr || std::all_of(tensor_split, tensor_split + GGML_CUDA_MAX_DEVICES, [](float x) { return x == 0.0f; });
+    if (all_zero) {
+        tensor_split_arr = ggml_cuda_info().default_tensor_split;
+    } else {
+        float split_sum = 0.0f;
+        for (int i = 0; i < ggml_backend_cuda_get_device_count(); ++i) {
+            tensor_split_arr[i] = split_sum;
+            split_sum += tensor_split[i];
+        }
+        for (int i = 0; i < ggml_backend_cuda_get_device_count(); ++i) {
+            tensor_split_arr[i] /= split_sum;
+        }
+    }
+
+    auto it = buft_map.find(tensor_split_arr);
+    if (it != buft_map.end()) {
+        return &it->second;
+    }
+
+    struct ggml_backend_buffer_type buft {
+        /* .iface   = */ ggml_backend_cuda_split_buffer_type_interface,
+        /* .device  = */ ggml_backend_reg_dev_get(ggml_backend_cuda_reg(), 0),
+        /* .context = */ new ggml_backend_cuda_split_buffer_type_context{tensor_split_arr},
+    };
+
+    auto result = buft_map.emplace(tensor_split_arr, buft);
+    return &result.first->second;
+}
+
+// host buffer type
+
+static const char * ggml_backend_cuda_host_buffer_type_name(ggml_backend_buffer_type_t buft) {
+    return GGML_CUDA_NAME "_Host";
+
+    GGML_UNUSED(buft);
+}
+
+static const char * ggml_backend_cuda_host_buffer_name(ggml_backend_buffer_t buffer) {
+    return GGML_CUDA_NAME "_Host";
+
+    GGML_UNUSED(buffer);
+}
+
+static void ggml_backend_cuda_host_buffer_free_buffer(ggml_backend_buffer_t buffer) {
+    CUDA_CHECK(cudaFreeHost(buffer->context));
+}
+
+static void * ggml_cuda_host_malloc(size_t size) {
+    if (getenv("GGML_CUDA_NO_PINNED") != nullptr) {
+        return nullptr;
+    }
+
+    void * ptr = nullptr;
+    cudaError_t err = cudaMallocHost((void **) &ptr, size);
+    if (err != cudaSuccess) {
+        // clear the error
+        cudaGetLastError();
+        GGML_LOG_WARN("%s: failed to allocate %.2f MiB of pinned memory: %s\n", __func__,
+                           size / 1024.0 / 1024.0, cudaGetErrorString(err));
+        return nullptr;
+    }
+
+    return ptr;
+}
+
+static ggml_backend_buffer_t ggml_backend_cuda_host_buffer_type_alloc_buffer(ggml_backend_buffer_type_t buft, size_t size) {
+    void * ptr = ggml_cuda_host_malloc(size);
+
+    if (ptr == nullptr) {
+        // fallback to cpu buffer
+        return ggml_backend_buft_alloc_buffer(ggml_backend_cpu_buffer_type(), size);
+    }
+
+    ggml_backend_buffer_t buffer = ggml_backend_cpu_buffer_from_ptr(ptr, size);
+    buffer->buft = buft;
+    buffer->iface.get_name = ggml_backend_cuda_host_buffer_name;
+    buffer->iface.free_buffer = ggml_backend_cuda_host_buffer_free_buffer;
+
+    return buffer;
+}
+
+ggml_backend_buffer_type_t ggml_backend_cuda_host_buffer_type() {
+    static struct ggml_backend_buffer_type ggml_backend_cuda_buffer_type_host = {
+        /* .iface    = */ {
+            /* .get_name         = */ ggml_backend_cuda_host_buffer_type_name,
+            /* .alloc_buffer     = */ ggml_backend_cuda_host_buffer_type_alloc_buffer,
+            /* .get_alignment    = */ ggml_backend_cpu_buffer_type()->iface.get_alignment,
+            /* .get_max_size     = */ NULL, // defaults to SIZE_MAX
+            /* .get_alloc_size   = */ ggml_backend_cpu_buffer_type()->iface.get_alloc_size,
+            /* .is_host          = */ ggml_backend_cpu_buffer_type()->iface.is_host,
+        },
+        /* .device   = */ ggml_backend_reg_dev_get(ggml_backend_cuda_reg(), 0),
+        /* .context  = */ nullptr,
+    };
+
+    return &ggml_backend_cuda_buffer_type_host;
+}
+
+//static bool ggml_backend_buffer_is_cuda_host(ggml_backend_buffer_t buffer) {
+//    return buffer->buft->iface.get_name == ggml_backend_cuda_host_buffer_type_name;
+//}
+
+/// kernels
+
+typedef void (*ggml_cuda_op_mul_mat_t)(
+    ggml_backend_cuda_context & ctx,
+    const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, const char * src0_dd_i, const float * src1_ddf_i,
+    const char * src1_ddq_i, float * dst_dd_i, const int64_t row_low, const int64_t row_high, const int64_t src1_ncols,
+    const int64_t src1_padded_row_size, cudaStream_t stream);
+
+#ifndef GGML_CUDA_PEER_MAX_BATCH_SIZE
+#define GGML_CUDA_PEER_MAX_BATCH_SIZE 128
+#endif // GGML_CUDA_PEER_MAX_BATCH_SIZE
+
+#define MUL_MAT_SRC1_COL_STRIDE 128
+
+static __global__ void mul_mat_p021_f16_f32(
+    const void * __restrict__ vx, const float * __restrict__ y, float * __restrict__ dst,
+    const int ncols_x, const int nrows_x, const int nchannels_x, const int nchannels_y) {
+
+    const half * x = (const half *) vx;
+
+    const int row_x = blockDim.y*blockIdx.y + threadIdx.y;
+    const int channel = blockDim.z*blockIdx.z + threadIdx.z;
+    const int channel_x = channel / (nchannels_y / nchannels_x);
+
+    const int nrows_y = ncols_x;
+    const int nrows_dst = nrows_x;
+    const int row_dst = row_x;
+
+    float tmp = 0.0f;
+
+    for (int col_x0 = 0; col_x0 < ncols_x; col_x0 += blockDim.x) {
+        const int col_x = col_x0 + threadIdx.x;
+
+        if (col_x >= ncols_x) {
+            break;
+        }
+
+        // x is transposed and permuted
+        const int ix = row_x*nchannels_x*ncols_x + channel_x*ncols_x + col_x;
+        const float xi = __half2float(x[ix]);
+
+        const int row_y = col_x;
+
+        // y is not transposed but permuted
+        const int iy = channel*nrows_y + row_y;
+
+        tmp += xi * y[iy];
+    }
+
+    // dst is not transposed and not permuted
+    const int idst = channel*nrows_dst + row_dst;
+
+    // sum up partial sums and write back result
+    tmp = warp_reduce_sum(tmp);
+
+    if (threadIdx.x == 0) {
+        dst[idst] = tmp;
+    }
+}
+
+static __global__ void mul_mat_vec_nc_f16_f32( // nc == non-contiguous
+    const void * __restrict__ vx, const float * __restrict__ y, float * __restrict__ dst, const int ncols_x, const int nrows_x,
+    const int row_stride_x, const int channel_stride_x, const int channel_x_divisor) {
+
+    const half * x = (const half *) vx;
+
+    const int row_x     = blockDim.y*blockIdx.y + threadIdx.y;
+    const int channel   = blockDim.z*blockIdx.z + threadIdx.z;
+    const int channel_x = channel / channel_x_divisor;
+
+    const int nrows_y   = ncols_x;
+    const int nrows_dst = nrows_x;
+    const int row_dst   = row_x;
+
+    const int idst = channel*nrows_dst + row_dst;
+
+    float tmp = 0.0f;
+
+    for (int col_x0 = 0; col_x0 < ncols_x; col_x0 += blockDim.x) {
+        const int col_x = col_x0 + threadIdx.x;
+
+        if (col_x >= ncols_x) {
+            break;
+        }
+
+        const int row_y = col_x;
+
+        const int ix = channel_x*channel_stride_x + row_x*row_stride_x + col_x;
+        const int iy = channel*nrows_y + row_y;
+
+        const float xi = __half2float(x[ix]);
+
+        tmp += xi * y[iy];
+    }
+
+    // sum up partial sums and write back result
+    tmp = warp_reduce_sum(tmp);
+
+    if (threadIdx.x == 0) {
+        dst[idst] = tmp;
+    }
+}
+
+static void ggml_mul_mat_p021_f16_f32_cuda(
+    const void * vx, const float * y, float * dst, const int ncols_x, const int nrows_x,
+    const int nchannels_x, const int nchannels_y, cudaStream_t stream) {
+
+    const dim3 block_nums(1, nrows_x, nchannels_y);
+    const dim3 block_dims(WARP_SIZE, 1, 1);
+    mul_mat_p021_f16_f32<<<block_nums, block_dims, 0, stream>>>(vx, y, dst, ncols_x, nrows_x, nchannels_x, nchannels_y);
+}
+
+static void ggml_mul_mat_vec_nc_f16_f32_cuda(
+    const void * vx, const float * y, float * dst, const int ncols_x, const int nrows_x, const int row_stride_x,
+    const int nchannels_x, const int nchannels_y, const int channel_stride_x, cudaStream_t stream) {
+
+    const dim3 block_nums(1, nrows_x, nchannels_y);
+    const dim3 block_dims(WARP_SIZE, 1, 1);
+    mul_mat_vec_nc_f16_f32<<<block_nums, block_dims, 0, stream>>>
+        (vx, y, dst, ncols_x, nrows_x, row_stride_x, channel_stride_x, nchannels_y/nchannels_x);
+}
+
+static cudaError_t ggml_cuda_cpy_tensor_2d(
+    void * dst, const struct ggml_tensor * src, int64_t i3, int64_t i2, int64_t i1_low, int64_t i1_high, cudaStream_t stream) {
+
+    GGML_ASSERT(ggml_backend_buffer_is_cuda(src->buffer));
+    char * src_ptr = (char *) src->data;
+    char * dst_ptr = (char *) dst;
+
+    const int64_t ne0 = src->ne[0];
+    const int64_t nb0 = src->nb[0];
+    const int64_t nb1 = src->nb[1];
+    const int64_t nb2 = src->nb[2];
+    const int64_t nb3 = src->nb[3];
+    const enum ggml_type type = src->type;
+    const int64_t ts = ggml_type_size(type);
+    const int64_t bs = ggml_blck_size(type);
+    int64_t i1_diff = i1_high - i1_low;
+
+    const char * x = src_ptr + i1_low*nb1 + i2*nb2 + i3*nb3;
+    if (nb0 == ts && nb1 == ts*ne0/bs) {
+        return cudaMemcpyAsync(dst_ptr, x, i1_diff*nb1, cudaMemcpyDeviceToDevice, stream);
+    } else if (nb0 == ts) {
+        return cudaMemcpy2DAsync(dst_ptr, ts*ne0/bs, x, nb1, ts*ne0/bs, i1_diff, cudaMemcpyDeviceToDevice, stream);
+    } else {
+        for (int64_t i1 = 0; i1 < i1_diff; i1++) {
+            const void * rx = (const void *) ((const char *) x + i1*nb1);
+            void * rd = (void *) (dst_ptr + i1*ts*ne0/bs);
+            // pretend the row is a matrix with cols=1
+            cudaError_t r = cudaMemcpy2DAsync(rd, ts/bs, rx, nb0, ts/bs, ne0, cudaMemcpyDeviceToDevice, stream);
+            if (r != cudaSuccess) {
+                return r;
+            }
+        }
+        return cudaSuccess;
+    }
+}
+
+static void ggml_cuda_op_mul_mat_cublas(
+    ggml_backend_cuda_context & ctx,
+    const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, const char * src0_dd_i, const float * src1_ddf_i,
+    const char * src1_ddq_i, float * dst_dd_i, const int64_t row_low, const int64_t row_high, const int64_t src1_ncols,
+    const int64_t src1_padded_row_size, cudaStream_t stream) {
+
+    GGML_ASSERT(src0_dd_i  != nullptr);
+    GGML_ASSERT(src1_ddf_i != nullptr);
+    GGML_ASSERT(dst_dd_i   != nullptr);
+
+    const int64_t ne00 = src0->ne[0];
+    const int64_t ne10 = src1->ne[0];
+
+    const int64_t ne0 = dst->ne[0];
+
+    const int64_t row_diff = row_high - row_low;
+
+    int id = ggml_cuda_get_device();
+
+    // the main device has a larger memory buffer to hold the results from all GPUs
+    // ldc == nrows of the matrix that cuBLAS writes into
+    int64_t ldc = id == ctx.device ? ne0 : row_diff;
+
+    const int compute_capability = ggml_cuda_info().devices[id].cc;
+
+    if (compute_capability >= CC_VOLTA && (src0->type == GGML_TYPE_F16 || ggml_is_quantized(src0->type)) && ggml_is_contiguous(src0) && row_diff == src0->ne[1] && dst->op_params[0] == GGML_PREC_DEFAULT) {
+        // convert src0 and src1 to fp16, multiply as fp16, convert dst to fp32
+        ggml_cuda_pool_alloc<half> src0_as_f16(ctx.pool(id));
+        if (src0->type != GGML_TYPE_F16) {
+            const to_fp16_cuda_t to_fp16_cuda = ggml_get_to_fp16_cuda(src0->type);
+            GGML_ASSERT(to_fp16_cuda != nullptr);
+            size_t ne = row_diff*ne00;
+            src0_as_f16.alloc(ne);
+            to_fp16_cuda(src0_dd_i, src0_as_f16.get(), ne, stream);
+        }
+        const half * src0_ptr = src0->type == GGML_TYPE_F16 ? (const half *) src0_dd_i : src0_as_f16.get();
+
+        ggml_cuda_pool_alloc<half> src1_as_f16(ctx.pool(id));
+        if (src1->type != GGML_TYPE_F16) {
+            const to_fp16_cuda_t to_fp16_cuda = ggml_get_to_fp16_cuda(src1->type);
+            GGML_ASSERT(to_fp16_cuda != nullptr);
+            size_t ne = src1_ncols*ne10;
+            src1_as_f16.alloc(ne);
+            to_fp16_cuda(src1_ddf_i, src1_as_f16.get(), ne, stream);
+        }
+        const half * src1_ptr = src1->type == GGML_TYPE_F16 ? (const half *) src1_ddf_i : src1_as_f16.get();
+        ggml_cuda_pool_alloc<half> dst_f16(ctx.pool(id), row_diff*src1_ncols);
+
+        const half alpha_f16 = 1.0f;
+        const half beta_f16 = 0.0f;
+
+        CUBLAS_CHECK(cublasSetStream(ctx.cublas_handle(id), stream));
+        CUBLAS_CHECK(
+            cublasGemmEx(ctx.cublas_handle(id), CUBLAS_OP_T, CUBLAS_OP_N,
+                    row_diff, src1_ncols, ne10,
+                    &alpha_f16, src0_ptr,       CUDA_R_16F, ne00,
+                                src1_ptr,       CUDA_R_16F, ne10,
+                    &beta_f16,   dst_f16.get(), CUDA_R_16F, ldc,
+                    CUBLAS_COMPUTE_16F,
+                    CUBLAS_GEMM_DEFAULT_TENSOR_OP));
+
+        const to_fp32_cuda_t to_fp32_cuda = ggml_get_to_fp32_cuda(GGML_TYPE_F16);
+        to_fp32_cuda(dst_f16.get(), dst_dd_i, row_diff*src1_ncols, stream);
+    } else {
+        ggml_cuda_pool_alloc<float> src0_ddq_as_f32(ctx.pool(id));
+        ggml_cuda_pool_alloc<float> src1_ddq_as_f32(ctx.pool(id));
+
+        if (src0->type != GGML_TYPE_F32) {
+            const to_fp32_cuda_t to_fp32_cuda = ggml_get_to_fp32_cuda(src0->type);
+            GGML_ASSERT(to_fp32_cuda != nullptr);
+            src0_ddq_as_f32.alloc(row_diff*ne00);
+            to_fp32_cuda(src0_dd_i, src0_ddq_as_f32.get(), row_diff*ne00, stream);
+        }
+        if (src1->type != GGML_TYPE_F32) {
+            const to_fp32_cuda_t to_fp32_cuda = ggml_get_to_fp32_cuda(src1->type);
+            GGML_ASSERT(to_fp32_cuda != nullptr);
+            src1_ddq_as_f32.alloc(src1_ncols*ne10);
+            to_fp32_cuda(src1_ddf_i, src1_ddq_as_f32.get(), src1_ncols*ne10, stream);
+        }
+
+        const float * src0_ddf_i = src0->type == GGML_TYPE_F32 ? (const float *) src0_dd_i : src0_ddq_as_f32.get();
+        const float * src1_ddf1_i = src1->type == GGML_TYPE_F32 ? (const float *) src1_ddf_i : src1_ddq_as_f32.get();
+
+        const float alpha = 1.0f;
+        const float beta = 0.0f;
+
+        CUBLAS_CHECK(cublasSetStream(ctx.cublas_handle(id), stream));
+        CUBLAS_CHECK(
+            cublasSgemm(ctx.cublas_handle(id), CUBLAS_OP_T, CUBLAS_OP_N,
+                    row_diff, src1_ncols, ne10,
+                    &alpha, src0_ddf_i,  ne00,
+                            src1_ddf1_i, ne10,
+                    &beta,  dst_dd_i,    ldc));
+    }
+
+    GGML_UNUSED(dst);
+    GGML_UNUSED(src1_ddq_i);
+    GGML_UNUSED(src1_padded_row_size);
+}
+
+static void ggml_cuda_set_peer_access(const int n_tokens, int main_device) {
+    static bool peer_access_enabled = false;
+
+    const bool enable_peer_access = n_tokens <= GGML_CUDA_PEER_MAX_BATCH_SIZE;
+
+    if (peer_access_enabled == enable_peer_access) {
+        return;
+    }
+
+#ifdef NDEBUG
+    for (int id = 0; id < ggml_backend_cuda_get_device_count(); ++id) {
+        ggml_cuda_set_device(id);
+        CUDA_CHECK(cudaDeviceSynchronize());
+    }
+
+    for (int id = 0; id < ggml_backend_cuda_get_device_count(); ++id) {
+        ggml_cuda_set_device(id);
+
+        for (int id_other = 0; id_other < ggml_backend_cuda_get_device_count(); ++id_other) {
+            if (id == id_other) {
+                continue;
+            }
+            if (id != main_device && id_other != main_device) {
+                continue;
+            }
+
+            int can_access_peer;
+            CUDA_CHECK(cudaDeviceCanAccessPeer(&can_access_peer, id, id_other));
+            if (can_access_peer) {
+                if (enable_peer_access) {
+                    cudaError_t err = cudaDeviceEnablePeerAccess(id_other, 0);
+                    if (err != cudaErrorPeerAccessAlreadyEnabled) {
+                        CUDA_CHECK(err);
+                    }
+                } else {
+                    cudaError_t err = cudaDeviceDisablePeerAccess(id_other);
+                    if (err != cudaErrorPeerAccessNotEnabled) {
+                        CUDA_CHECK(err);
+                    }
+                }
+            }
+        }
+    }
+
+    ggml_cuda_set_device(main_device);
+#endif // NDEBUG
+
+    peer_access_enabled = enable_peer_access;
+
+    GGML_UNUSED(main_device);
+}
+
+static cudaError_t ggml_cuda_Memcpy2DPeerAsync(
+    void * dst, int dstDevice, size_t dpitch, void * src, int srcDevice, size_t spitch, size_t width, size_t height, cudaStream_t stream) {
+
+#if !defined(GGML_USE_HIPBLAS) && !defined(GGML_USE_MUSA)
+    // cudaMemcpy2DAsync may fail with copies between vmm pools of different devices
+    cudaMemcpy3DPeerParms p = {};
+    p.dstDevice = dstDevice;
+    p.dstPtr = make_cudaPitchedPtr(dst, dpitch, dpitch, height);
+    p.srcDevice = srcDevice;
+    p.srcPtr = make_cudaPitchedPtr(src, spitch, spitch, height);
+    p.extent = make_cudaExtent(width, height, 1);
+    return cudaMemcpy3DPeerAsync(&p, stream);
+#else
+    // HIP does not support cudaMemcpy3DPeerAsync or vmm pools
+    GGML_UNUSED(dstDevice);
+    GGML_UNUSED(srcDevice);
+    return cudaMemcpy2DAsync(dst, dpitch, src, spitch, width, height, cudaMemcpyDeviceToDevice, stream);
+#endif // !defined(GGML_USE_HIPBLAS) && !defined(GGML_USE_MUSA)
+}
+
+static void ggml_cuda_op_mul_mat(
+    ggml_backend_cuda_context & ctx,
+    const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, ggml_cuda_op_mul_mat_t op,
+    quantize_cuda_t quantize_src1) {
+
+    const int64_t ne00 = src0->ne[0];
+    const int64_t ne01 = src0->ne[1];
+    const int64_t ne02 = src0->ne[2];
+    const int64_t ne03 = src0->ne[3];
+
+    const int64_t ne10 = src1->ne[0];
+    const int64_t ne11 = src1->ne[1];
+    const int64_t ne12 = src1->ne[2];
+    const int64_t ne13 = src1->ne[3];
+    const int64_t nrows1 = ggml_nrows(src1);
+
+    GGML_ASSERT(ne03 == ne13);
+
+    const int64_t ne0 = dst->ne[0];
+    const int64_t ne1 = dst->ne[1];
+
+    const int64_t nb2 = dst->nb[2];
+    const int64_t nb3 = dst->nb[3];
+
+    GGML_ASSERT(ggml_backend_buffer_is_cuda(dst->buffer));
+    GGML_ASSERT(ggml_backend_buffer_is_cuda(src1->buffer));
+    ggml_backend_cuda_buffer_context * src1_ctx = (ggml_backend_cuda_buffer_context *) src1->buffer->context;
+    ggml_backend_cuda_buffer_context * dst_ctx  = (ggml_backend_cuda_buffer_context *) dst->buffer->context;
+
+    GGML_ASSERT(src1->type == GGML_TYPE_F32 || (src1->ne[2] == 1 && src1->ne[3] == 1));
+
+    GGML_ASSERT(ne12 >= ne02 && ne12 % ne02 == 0);
+
+    const int64_t i02_divisor = ne12 / ne02;
+
+    const size_t src0_ts = ggml_type_size(src0->type);
+    const size_t src0_bs = ggml_blck_size(src0->type);
+    const size_t q8_1_ts = sizeof(block_q8_1);
+    const size_t q8_1_bs = QK8_1;
+
+    const bool src0_is_contiguous = ggml_is_contiguous(src0);
+    const bool src1_is_contiguous = ggml_is_contiguous(src1);
+
+    const int64_t src1_padded_col_size = GGML_PAD(ne10, MATRIX_ROW_PADDING);
+
+    const bool split = ggml_backend_buffer_is_cuda_split(src0->buffer);
+    GGML_ASSERT(!(split && ne02 > 1));
+    GGML_ASSERT(!(split && ne03 > 1));
+    GGML_ASSERT(!(split && ne02 < ne12));
+
+    ggml_tensor_extra_gpu * src0_extra = split ? (ggml_tensor_extra_gpu *) src0->extra : nullptr;
+
+
+    std::array<float, GGML_CUDA_MAX_DEVICES> tensor_split;
+    if (split) {
+        ggml_backend_cuda_split_buffer_type_context * buft_ctx = (ggml_backend_cuda_split_buffer_type_context *) src0->buffer->buft->context;
+        tensor_split = buft_ctx->tensor_split;
+    }
+
+    struct dev_data {
+        int cc;
+
+        ggml_cuda_pool_alloc<char>   src0_dd_alloc;
+        ggml_cuda_pool_alloc<float> src1_ddf_alloc;
+        ggml_cuda_pool_alloc<char>  src1_ddq_alloc;
+        ggml_cuda_pool_alloc<float>   dst_dd_alloc;
+
+        char  *  src0_dd = nullptr;
+        float * src1_ddf = nullptr; // float
+        char  * src1_ddq = nullptr; // q8_1
+        float *   dst_dd = nullptr;
+
+        int64_t  row_low;
+        int64_t row_high;
+    };
+
+    dev_data dev[GGML_CUDA_MAX_DEVICES];
+
+    int used_devices = 0;
+
+    for (int id = 0; id < ggml_backend_cuda_get_device_count(); ++id) {
+        dev[id].cc = ggml_cuda_info().devices[id].cc;
+
+        // by default, use all rows
+        dev[id].row_low  = 0;
+        dev[id].row_high = ne01;
+
+        // for multi GPU, get the row boundaries from tensor split
+        // and round to mul_mat_q tile sizes
+        if (split) {
+            const int64_t rounding = get_row_rounding(tensor_split);
+
+            if (id != 0) {
+                dev[id].row_low  = ne01*tensor_split[id];
+                if (dev[id].row_low < ne01) {
+                    dev[id].row_low -= dev[id].row_low % rounding;
+                }
+            }
+
+            if (id != ggml_backend_cuda_get_device_count() - 1) {
+                dev[id].row_high  = ne01*tensor_split[id + 1];
+                if (dev[id].row_high < ne01) {
+                    dev[id].row_high -= dev[id].row_high % rounding;
+                }
+            }
+        }
+    }
+
+    for (int id = 0; id < ggml_backend_cuda_get_device_count(); ++id) {
+        if ((!split && id != ctx.device) || dev[id].row_low == dev[id].row_high) {
+            continue;
+        }
+
+        used_devices++;
+
+        const bool src1_on_device = id == src1_ctx->device;
+        const bool  dst_on_device = id == dst_ctx->device;
+
+        ggml_cuda_set_device(id);
+        cudaStream_t stream = ctx.stream(id, 0);
+
+        if (src0_is_contiguous) {
+            dev[id].src0_dd = split ? (char *) src0_extra->data_device[id] : (char *) src0->data;
+        } else {
+            dev[id].src0_dd = dev[id].src0_dd_alloc.alloc(ctx.pool(id), ggml_nbytes(src0));
+        }
+
+        // If src0 is on a temporary compute buffers (partial offloading) there may be some padding that needs to be cleared:
+        if (ne00 % MATRIX_ROW_PADDING != 0 && ggml_is_quantized(src0->type) && ggml_backend_buffer_get_usage(src0->buffer) == GGML_BACKEND_BUFFER_USAGE_COMPUTE && src0->view_src == nullptr) {
+            const int64_t nbytes_data    = ggml_row_size(src0->type, (dev[id].row_high - dev[id].row_low)*ne00);
+            const int64_t nbytes_padding = ggml_row_size(src0->type, MATRIX_ROW_PADDING - ne00 % MATRIX_ROW_PADDING);
+            CUDA_CHECK(cudaMemsetAsync(dev[id].src0_dd + nbytes_data , 0, nbytes_padding, stream));
+        }
+
+        if (src1_on_device && src1_is_contiguous) {
+            dev[id].src1_ddf = (float *) src1->data;
+        } else {
+            dev[id].src1_ddf = dev[id].src1_ddf_alloc.alloc(ctx.pool(id), ggml_nelements(src1));
+        }
+
+        if (quantize_src1) {
+            size_t src_1_ddq_size = nrows1*src1_padded_col_size*q8_1_ts/q8_1_bs;
+            if (quantize_src1 == quantize_mmq_q8_1_cuda) {
+                src_1_ddq_size += get_mmq_x_max_host(dev[id].cc)*sizeof(block_q8_1_mmq);
+            }
+            dev[id].src1_ddq = dev[id].src1_ddq_alloc.alloc(ctx.pool(id), src_1_ddq_size);
+
+            if (src1_on_device && src1_is_contiguous) {
+                quantize_src1(dev[id].src1_ddf, dev[id].src1_ddq, ne10, ne11, ne12*ne13, src1_padded_col_size, src0->type, stream);
+                CUDA_CHECK(cudaGetLastError());
+            }
+        }
+
+        if (dst_on_device) {
+            dev[id].dst_dd = (float *) dst->data;
+        } else {
+            const size_t size_dst_ddf = split ? (dev[id].row_high - dev[id].row_low)*ne1 : ggml_nelements(dst);
+            dev[id].dst_dd = dev[id].dst_dd_alloc.alloc(ctx.pool(id), size_dst_ddf);
+        }
+    }
+
+    // if multiple devices are used they need to wait for the main device
+    // here an event is recorded that signals that the main device has finished calculating the input data
+    if (split && used_devices > 1) {
+        ggml_cuda_set_device(ctx.device);
+        CUDA_CHECK(cudaEventRecord(src0_extra->events[ctx.device][0], ctx.stream()));
+    }
+
+    const int64_t src1_col_stride = split && used_devices > 1 ? MUL_MAT_SRC1_COL_STRIDE : ne11;
+    for (int64_t src1_col_0 = 0; src1_col_0 < ne11; src1_col_0 += src1_col_stride) {
+        const int64_t is = split ? (src1_col_0/src1_col_stride) % GGML_CUDA_MAX_STREAMS : 0;
+        const int64_t src1_ncols = src1_col_0 + src1_col_stride > ne11 ? ne11 - src1_col_0 : src1_col_stride;
+
+        for (int id = 0; id < ggml_backend_cuda_get_device_count(); ++id) {
+            if ((!split && id != ctx.device) || dev[id].row_low == dev[id].row_high) {
+                continue;
+            }
+
+            const bool src1_on_device = id == src1_ctx->device;
+            const bool  dst_on_device = id == dst_ctx->device;
+            const int64_t row_diff = dev[id].row_high - dev[id].row_low;
+
+            ggml_cuda_set_device(id);
+            cudaStream_t stream = ctx.stream(id, is);
+
+            // wait for main GPU data if necessary
+            if (split && (id != ctx.device || is != 0)) {
+                CUDA_CHECK(cudaStreamWaitEvent(stream, src0_extra->events[ctx.device][0], 0));
+            }
+
+            for (int64_t i0 = 0; i0 < ne13*ne12; ++i0) {
+                const int64_t i03 = i0 / ne12;
+                const int64_t i02 = i0 % ne12;
+
+                size_t src1_ddq_i_offset = i0*ne11 * src1_padded_col_size*q8_1_ts/q8_1_bs;
+                if (quantize_src1 == quantize_mmq_q8_1_cuda) {
+                    src1_ddq_i_offset += src1_col_0 * sizeof(block_q8_1_mmq);
+                } else {
+                    src1_ddq_i_offset += src1_col_0 * src1_padded_col_size*q8_1_ts/q8_1_bs;
+                }
+
+                // for split tensors the data begins at i0 == i0_offset_low
+                char  *  src0_dd_i =  dev[id].src0_dd + (i0/i02_divisor) * (ne01*ne00*src0_ts)/src0_bs;
+                float * src1_ddf_i = dev[id].src1_ddf + (i0*ne11 + src1_col_0) * ne10;
+                char  * src1_ddq_i = dev[id].src1_ddq +  src1_ddq_i_offset;
+                float *   dst_dd_i =   dev[id].dst_dd + (i0*ne1  + src1_col_0) * (dst_on_device ? ne0 : row_diff);
+
+                // the main device memory buffer can be on VRAM scratch, with space for all partial results
+                // in that case an offset on dst_ddf_i is needed
+                if (id == ctx.device) {
+                    dst_dd_i += dev[id].row_low; // offset is 0 if no tensor split
+                }
+
+                // copy src0, src1 to device if necessary
+                if (src1_is_contiguous) {
+                    if (id != ctx.device) {
+                        if (quantize_src1) {
+                            char * src1_ddq_i_source = dev[ctx.device].src1_ddq + src1_ddq_i_offset;
+                            if (quantize_src1 == quantize_mmq_q8_1_cuda) {
+                                const size_t pitch = ne11*sizeof(block_q8_1_mmq);
+                                const size_t width = src1_ncols*sizeof(block_q8_1_mmq);
+                                const size_t height = src1_padded_col_size/(4*QK8_1);
+                                CUDA_CHECK(ggml_cuda_Memcpy2DPeerAsync(src1_ddq_i, id, pitch, src1_ddq_i_source, ctx.device, pitch, width, height, stream));
+                            } else {
+                                CUDA_CHECK(cudaMemcpyPeerAsync(
+                                    src1_ddq_i, id, src1_ddq_i_source, ctx.device, src1_ncols*src1_padded_col_size*q8_1_ts/q8_1_bs, stream));
+                            }
+                        } else {
+                            float * src1_ddf_i_source = (float *) src1->data;
+                            src1_ddf_i_source += (i0*ne11 + src1_col_0) * ne10;
+                            CUDA_CHECK(cudaMemcpyPeerAsync(src1_ddf_i, id, src1_ddf_i_source, ctx.device,
+                                                            src1_ncols*ne10*sizeof(float), stream));
+                        }
+                    }
+                } else if (src1_on_device && !src1_is_contiguous) {
+                    CUDA_CHECK(ggml_cuda_cpy_tensor_2d(
+                                src1_ddf_i, src1, i03, i02, src1_col_0, src1_col_0+src1_ncols, stream));
+                } else {
+                    GGML_ABORT("fatal error");
+                }
+
+                if (quantize_src1 && !src1_is_contiguous) {
+                    quantize_src1(src1_ddf_i, src1_ddq_i, ne10, src1_ncols, 1, src1_padded_col_size, src0->type, stream);
+                    CUDA_CHECK(cudaGetLastError());
+                }
+
+                if (src1_col_0 == 0 && !src0_is_contiguous && i02 % i02_divisor == 0) {
+                    CUDA_CHECK(ggml_cuda_cpy_tensor_2d(src0_dd_i, src0, i03, i02/i02_divisor, dev[id].row_low, dev[id].row_high, stream));
+                }
+
+                // do the computation
+                op(ctx, src0, src1, dst, src0_dd_i, src1_ddf_i, src1_ddq_i, dst_dd_i,
+                    dev[id].row_low, dev[id].row_high, src1_ncols, src1_padded_col_size, stream);
+                CUDA_CHECK(cudaGetLastError());
+
+                // copy dst to host or other device if necessary
+                if (!dst_on_device) {
+                    void * dst_off_device = dst->data;
+                    if (split) {
+                        // src0 = weight matrix is saved as a transposed matrix for better memory layout.
+                        // dst is NOT transposed.
+                        // The outputs of matrix matrix multiplications can therefore NOT simply be concatenated for >1 GPU.
+                        // Instead they need to be copied to the correct slice in ne0 = dst row index.
+                        // If dst is a vector with ne0 == 1 then you don't have to do this but it still produces correct results.
+                        float * dhf_dst_i = (float *) ((char *) dst_off_device + i02*nb2 + i03*nb3);
+                        GGML_ASSERT(dst->nb[1] == ne0*sizeof(float));
+                        dhf_dst_i += src1_col_0*ne0 + dev[id].row_low;
+                        CUDA_CHECK(ggml_cuda_Memcpy2DPeerAsync(
+                            dhf_dst_i, ctx.device, ne0*sizeof(float), dst_dd_i, id, row_diff*sizeof(float), row_diff*sizeof(float), src1_ncols, stream));
+                    } else {
+                        float * dhf_dst_i = (float *) ((char *) dst_off_device + i02*nb2 + i03*nb3);
+                        GGML_ASSERT(dst->nb[1] == ne0*sizeof(float));
+                        dhf_dst_i += src1_col_0*ne0;
+                        CUDA_CHECK(cudaMemcpyAsync(dhf_dst_i, dst_dd_i, src1_ncols*ne0*sizeof(float), cudaMemcpyDeviceToDevice, stream));
+                    }
+                }
+
+                // add event for the main device to wait on until other device is done
+                if (split && (id != ctx.device || is != 0)) {
+                    CUDA_CHECK(cudaEventRecord(src0_extra->events[id][is], stream));
+                }
+            }
+        }
+    }
+
+    // main device waits for all other devices to be finished
+    if (split && ggml_backend_cuda_get_device_count() > 1) {
+        int64_t is_max = (ne11 + MUL_MAT_SRC1_COL_STRIDE - 1) / MUL_MAT_SRC1_COL_STRIDE;
+        is_max = is_max <= GGML_CUDA_MAX_STREAMS ? is_max : GGML_CUDA_MAX_STREAMS;
+
+        ggml_cuda_set_device(ctx.device);
+        for (int id = 0; id < ggml_backend_cuda_get_device_count(); ++id) {
+            if (dev[id].row_low == dev[id].row_high) {
+                continue;
+            }
+            for (int64_t is = 0; is < is_max; ++is) {
+                CUDA_CHECK(cudaStreamWaitEvent(ctx.stream(), src0_extra->events[id][is], 0));
+            }
+        }
+    }
+}
+
+static void ggml_cuda_mul_mat_vec_p021(ggml_backend_cuda_context & ctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
+    GGML_ASSERT(ggml_is_permuted(src0) && ggml_is_permuted(src1));
+    GGML_ASSERT(ggml_backend_buffer_is_cuda(src0->buffer));
+    GGML_ASSERT(src0->nb[0] <= src0->nb[1] && src0->nb[2] <= src0->nb[3]); // 0213 permutation
+    GGML_ASSERT(src1->nb[0] <= src1->nb[1] && src1->nb[2] <= src1->nb[3]); // 0213 permutation
+    GGML_ASSERT(src0->type == GGML_TYPE_F16);
+    GGML_ASSERT(src1->type == GGML_TYPE_F32);
+
+    const int64_t ne00 = src0->ne[0];
+    const int64_t ne01 = src0->ne[1];
+    const int64_t ne02 = src0->ne[2];
+
+    const int64_t ne12 = src1->ne[2];
+
+    cudaStream_t main_stream = ctx.stream();
+
+    void  * src0_ddq = src0->data;
+    float * src1_ddf = (float *) src1->data;
+    float * dst_ddf  = (float *) dst->data;
+
+    ggml_mul_mat_p021_f16_f32_cuda(src0_ddq, src1_ddf, dst_ddf, ne00, ne01, ne02, ne12, main_stream);
+}
+
+static void ggml_cuda_mul_mat_vec_nc(ggml_backend_cuda_context & ctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
+    GGML_ASSERT(!ggml_is_transposed(src0));
+    GGML_ASSERT(!ggml_is_transposed(src1));
+    GGML_ASSERT(!ggml_is_permuted(src0));
+    GGML_ASSERT(ggml_backend_buffer_is_cuda(src0->buffer));
+    GGML_ASSERT(src0->type == GGML_TYPE_F16);
+    GGML_ASSERT(src1->type == GGML_TYPE_F32);
+
+    const int64_t ne00 = src0->ne[0];
+    const int64_t ne01 = src0->ne[1];
+    const int64_t ne02 = src0->ne[2];
+
+    const int64_t nb01 = src0->nb[1];
+    const int64_t nb02 = src0->nb[2];
+
+    const int64_t ne12 = src1->ne[2];
+
+    cudaStream_t main_stream = ctx.stream();
+
+    void  * src0_ddq = src0->data;
+    float * src1_ddf = (float *) src1->data;
+    float * dst_ddf  = (float *) dst->data;
+
+    const int64_t row_stride_x = nb01 / sizeof(half);
+    const int64_t channel_stride_x = nb02 / sizeof(half);
+
+    ggml_mul_mat_vec_nc_f16_f32_cuda(src0_ddq, src1_ddf, dst_ddf, ne00, ne01, row_stride_x, ne02, ne12, channel_stride_x, main_stream);
+}
+
+static __global__ void k_compute_batched_ptrs(
+        const half * src0_as_f16, const half * src1_as_f16, char * dst,
+        const void ** ptrs_src, void ** ptrs_dst,
+        int64_t ne12, int64_t ne13,
+        int64_t ne23,
+        size_t  nb02, size_t  nb03,
+        size_t  nb12, size_t  nb13,
+        size_t  nbd2, size_t  nbd3,
+        int64_t r2,   int64_t r3) {
+    int64_t i13 = blockIdx.x * blockDim.x + threadIdx.x;
+    int64_t i12 = blockIdx.y * blockDim.y + threadIdx.y;
+
+    if (i13 >= ne13 || i12 >= ne12) {
+        return;
+    }
+
+    int64_t i03 = i13 / r3;
+    int64_t i02 = i12 / r2;
+
+    ptrs_src[0*ne23 + i12 + i13*ne12] = (const char *) src0_as_f16 + i02*nb02 + i03*nb03;
+    ptrs_src[1*ne23 + i12 + i13*ne12] = (const char *) src1_as_f16 + i12*nb12 + i13*nb13;
+    ptrs_dst[0*ne23 + i12 + i13*ne12] = (      char *)         dst + i12*nbd2 + i13*nbd3;
+}
+
+static void ggml_cuda_mul_mat_batched_cublas(ggml_backend_cuda_context & ctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
+    GGML_ASSERT(!ggml_is_transposed(src0));
+    GGML_ASSERT(!ggml_is_transposed(src1));
+
+    GGML_ASSERT(ggml_backend_buffer_is_cuda(src0->buffer));
+    GGML_ASSERT(src0->type == GGML_TYPE_F16);
+
+    GGML_TENSOR_BINARY_OP_LOCALS
+
+    const int64_t ne_dst = ggml_nelements(dst);
+
+    cudaStream_t main_stream = ctx.stream();
+
+    CUBLAS_CHECK(cublasSetStream(ctx.cublas_handle(), main_stream));
+
+    void * src0_ddq = src0->data;
+    half * src0_f16 = (half *) src0_ddq;
+    float * src1_ddf = (float *) src1->data;
+    float * dst_ddf  = (float *) dst->data;
+
+    // convert src1 to fp16
+    ggml_cuda_pool_alloc<half> src1_f16_alloc(ctx.pool());
+    if (src1->type != GGML_TYPE_F16) {
+        const to_fp16_cuda_t to_fp16_cuda = ggml_get_to_fp16_cuda(src1->type);
+        const int64_t ne_src1 = ggml_nelements(src1);
+        src1_f16_alloc.alloc(ne_src1);
+        GGML_ASSERT(to_fp16_cuda != nullptr);
+        to_fp16_cuda(src1_ddf, src1_f16_alloc.get(), ne_src1, main_stream);
+    }
+    half * src1_f16 = src1->type == GGML_TYPE_F16 ? (half *) src1_ddf : src1_f16_alloc.get();
+
+    ggml_cuda_pool_alloc<half> dst_f16(ctx.pool());
+    char * dst_t;
+
+    cublasComputeType_t cu_compute_type = CUBLAS_COMPUTE_16F;
+    cudaDataType_t      cu_data_type    = CUDA_R_16F;
+
+    // dst strides
+    size_t nbd2 = dst->nb[2];
+    size_t nbd3 = dst->nb[3];
+
+    const half  alpha_f16 = 1.0f;
+    const half  beta_f16  = 0.0f;
+
+    const float alpha_f32 = 1.0f;
+    const float beta_f32  = 0.0f;
+
+    const void * alpha = &alpha_f16;
+    const void * beta  = &beta_f16;
+
+    if (dst->op_params[0] == GGML_PREC_DEFAULT) {
+        dst_t = (char *) dst_f16.alloc(ne_dst);
+
+        nbd2 /= sizeof(float) / sizeof(half);
+        nbd3 /= sizeof(float) / sizeof(half);
+    } else {
+        dst_t = (char *) dst_ddf;
+
+        cu_compute_type = CUBLAS_COMPUTE_32F;
+        cu_data_type    = CUDA_R_32F;
+
+        alpha = &alpha_f32;
+        beta  = &beta_f32;
+    }
+
+    GGML_ASSERT(ne12 % ne02 == 0);
+    GGML_ASSERT(ne13 % ne03 == 0);
+
+    // broadcast factors
+    const int64_t r2 = ne12/ne02;
+    const int64_t r3 = ne13/ne03;
+
+#if 0
+    // use cublasGemmEx
+    {
+        for (int i13 = 0; i13 < ne13; ++i13) {
+            for (int i12 = 0; i12 < ne12; ++i12) {
+                int i03 = i13 / r3;
+                int i02 = i12 / r2;
+
+                CUBLAS_CHECK(
+                        cublasGemmEx(g_cublas_handles[g_main_device], CUBLAS_OP_T, CUBLAS_OP_N,
+                            ne01, ne11, ne10,
+                            alpha, (const char *) src0_as_f16 + i02*src0->nb[2]   + i03*src0->nb[3]  , CUDA_R_16F,   nb01/sizeof(half),
+                                   (const char *) src1_as_f16 + i12*src1->nb[2]/2 + i13*src1->nb[3]/2, CUDA_R_16F,   nb11/sizeof(float),
+                            beta,  (      char *)       dst_t + i12*nbd2          + i13*nbd3,          cu_data_type, ne01,
+                            cu_compute_type,
+                            CUBLAS_GEMM_DEFAULT_TENSOR_OP));
+            }
+        }
+    }
+#else
+#ifdef GGML_USE_MUSA
+    GGML_ASSERT(false);
+#else // !GGML_USE_MUSA
+    if (r2 == 1 && r3 == 1 && ggml_is_contiguous_2(src0) && ggml_is_contiguous_2(src1)) {
+        // there is no broadcast and src0, src1 are contiguous across dims 2, 3
+        // use cublasGemmStridedBatchedEx
+        CUBLAS_CHECK(
+        cublasGemmStridedBatchedEx(ctx.cublas_handle(), CUBLAS_OP_T, CUBLAS_OP_N,
+                ne01, ne11, ne10,
+                alpha, (const char *) src0_f16, CUDA_R_16F,   nb01/nb00, nb02/nb00,  // strideA
+                       (const char *) src1_f16, CUDA_R_16F,   nb11/nb10, nb12/nb10,  // strideB
+                beta,  (      char *)    dst_t, cu_data_type, ne01,       nb2/nb0,   // strideC
+                ne12*ne13,
+                cu_compute_type,
+                CUBLAS_GEMM_DEFAULT_TENSOR_OP));
+    } else {
+        // use cublasGemmBatchedEx
+        const int ne23 = ne12*ne13;
+
+        ggml_cuda_pool_alloc<const void *> ptrs_src(ctx.pool(), 2*ne23);
+        ggml_cuda_pool_alloc<      void *> ptrs_dst(ctx.pool(), 1*ne23);
+
+        dim3 block_dims(ne13, ne12);
+        k_compute_batched_ptrs<<<1, block_dims, 0, main_stream>>>(
+                src0_f16, src1_f16, dst_t,
+                ptrs_src.get(), ptrs_dst.get(),
+                ne12, ne13,
+                ne23,
+                nb02, nb03,
+                src1->type == GGML_TYPE_F16 ? nb12 : nb12/2,
+                src1->type == GGML_TYPE_F16 ? nb13 : nb13/2,
+                nbd2, nbd3,
+                r2, r3);
+        CUDA_CHECK(cudaGetLastError());
+
+        CUBLAS_CHECK(
+        cublasGemmBatchedEx(ctx.cublas_handle(), CUBLAS_OP_T, CUBLAS_OP_N,
+                ne01, ne11, ne10,
+                alpha, (const void **) (ptrs_src.get() + 0*ne23), CUDA_R_16F,   nb01/nb00,
+                       (const void **) (ptrs_src.get() + 1*ne23), CUDA_R_16F,   nb11/nb10,
+                beta,  (      void **) (ptrs_dst.get() + 0*ne23), cu_data_type, ne01,
+                ne23,
+                cu_compute_type,
+                CUBLAS_GEMM_DEFAULT_TENSOR_OP));
+    }
+#endif // GGML_USE_MUSA
+#endif
+
+    if (dst->op_params[0] == GGML_PREC_DEFAULT) {
+        const to_fp32_cuda_t to_fp32_cuda = ggml_get_to_fp32_cuda(GGML_TYPE_F16);
+        to_fp32_cuda(dst_f16.get(), dst_ddf, ne_dst, main_stream);
+    }
+}
+
+static void ggml_cuda_mul_mat(ggml_backend_cuda_context & ctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
+    const bool split = ggml_backend_buffer_is_cuda_split(src0->buffer);
+
+    bool use_dequantize_mul_mat_vec = ggml_cuda_dmmv_type_supported(src0->type)
+        && src1->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32
+        && src0->ne[0] % (GGML_CUDA_DMMV_X*2) == 0 && src1->ne[1] == 1;
+    bool          use_mul_mat_vec_q =  ggml_is_quantized(src0->type)
+        && src1->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32
+        && src1->ne[1] <= MMVQ_MAX_BATCH_SIZE;
+    bool              use_mul_mat_q =  ggml_is_quantized(src0->type)
+        && src1->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32;
+
+    // if mmvq is available it's a better choice than dmmv:
+#ifndef GGML_CUDA_FORCE_DMMV
+    use_dequantize_mul_mat_vec = use_dequantize_mul_mat_vec && !use_mul_mat_vec_q;
+#endif // GGML_CUDA_FORCE_DMMV
+
+    bool any_gpus_with_slow_fp16 = false;
+
+    if (split) {
+        ggml_backend_cuda_split_buffer_type_context * buft_ctx = (ggml_backend_cuda_split_buffer_type_context *) src0->buffer->buft->context;
+        auto & tensor_split = buft_ctx->tensor_split;
+        for (int id = 0; id < ggml_backend_cuda_get_device_count(); ++id) {
+            // skip devices that are not going to do any work:
+            if (tensor_split[id] >= (id + 1 < ggml_backend_cuda_get_device_count() ? tensor_split[id + 1] : 1.0f)) {
+                continue;
+            }
+
+            const int cc            = ggml_cuda_info().devices[id].cc;
+            use_mul_mat_q           = use_mul_mat_q           && ggml_cuda_should_use_mmq(src0->type, cc, src1->ne[1]);
+            any_gpus_with_slow_fp16 = any_gpus_with_slow_fp16 || !fast_fp16_available(cc);
+        }
+    } else {
+        const int cc            = ggml_cuda_info().devices[ctx.device].cc;
+        use_mul_mat_q           = use_mul_mat_q           && ggml_cuda_should_use_mmq(src0->type, cc, src1->ne[1]);
+        any_gpus_with_slow_fp16 = any_gpus_with_slow_fp16 || !fast_fp16_available(cc);
+    }
+
+    // debug helpers
+    //printf("src0: %8d %8d %8d %8d\n", src0->ne[0], src0->ne[1], src0->ne[2], src0->ne[3]);
+    //printf("      %8d %8d %8d %8d\n", src0->nb[0], src0->nb[1], src0->nb[2], src0->nb[3]);
+    //printf("src1: %8d %8d %8d %8d\n", src1->ne[0], src1->ne[1], src1->ne[2], src1->ne[3]);
+    //printf("      %8d %8d %8d %8d\n", src1->nb[0], src1->nb[1], src1->nb[2], src1->nb[3]);
+    //printf("src0 is contiguous %d, transposed %d, type = %s, name = %s\n", ggml_is_contiguous(src0), ggml_is_transposed(src0), ggml_type_name(src0->type), src0->name);
+    //printf("src1 is contiguous %d, transposed %d, type = %s, name = %s\n", ggml_is_contiguous(src1), ggml_is_transposed(src1), ggml_type_name(src1->type), src1->name);
+
+    if (!split && any_gpus_with_slow_fp16 && src0->type == GGML_TYPE_F16 && ggml_is_permuted(src0) && ggml_is_permuted(src1) && src1->ne[1] == 1) {
+        // FP32 precision KQ single-batch for batch size 1 without FlashAttention
+        ggml_cuda_mul_mat_vec_p021(ctx, src0, src1, dst);
+    } else if (!split && any_gpus_with_slow_fp16 && src0->type == GGML_TYPE_F16 && !ggml_is_contiguous(src0) && !ggml_is_transposed(src1) && src1->ne[1] == 1) {
+        // FP32 precision KQV single-batch for batch size 1 without FlashAttention
+        ggml_cuda_mul_mat_vec_nc(ctx, src0, src1, dst);
+    } else if (!split && src0->type == GGML_TYPE_F16 && (src1->type == GGML_TYPE_F16 || !any_gpus_with_slow_fp16)
+               && !ggml_is_transposed(src0) && !ggml_is_transposed(src1) && src1->ne[2]*src1->ne[3] > 1) {
+        // KQ + KQV multi-batch without FlashAttention
+        ggml_cuda_mul_mat_batched_cublas(ctx, src0, src1, dst);
+    } else if (use_dequantize_mul_mat_vec) {
+        ggml_cuda_op_mul_mat(ctx, src0, src1, dst, ggml_cuda_op_dequantize_mul_mat_vec, nullptr);
+    } else if (use_mul_mat_vec_q) {
+        ggml_cuda_op_mul_mat(ctx, src0, src1, dst, ggml_cuda_op_mul_mat_vec_q, quantize_row_q8_1_cuda);
+    } else if (use_mul_mat_q) {
+        ggml_cuda_op_mul_mat(ctx, src0, src1, dst, ggml_cuda_op_mul_mat_q, quantize_mmq_q8_1_cuda);
+    } else {
+        ggml_cuda_op_mul_mat(ctx, src0, src1, dst, ggml_cuda_op_mul_mat_cublas, nullptr);
+    }
+}
+
+struct mmid_row_mapping {
+    int32_t i1;
+    int32_t i2;
+};
+
+static __global__ void k_copy_src1_to_contiguous(const char * __restrict__ src1_original, char * __restrict__ src1_contiguous,
+                                                 int * __restrict__ cur_src1_row, mmid_row_mapping * __restrict__ row_mapping,
+                                                 const char * __restrict ids, int64_t i02, size_t ids_nb1, size_t ids_nb0,
+                                                 int64_t ne11, int64_t ne10,
+                                                 size_t nb11, size_t nb12) {
+    int32_t iid1 = blockIdx.x;
+    int32_t id = blockIdx.y;
+
+    const int32_t row_id_i = *(const int32_t *) (ids + iid1*ids_nb1 + id*ids_nb0);
+
+    if (row_id_i != i02) {
+        return;
+    }
+
+    const int64_t i11 = id % ne11;
+    const int64_t i12 = iid1;
+
+    __shared__ int src1_row;
+    if (threadIdx.x == 0) {
+        src1_row = atomicAdd(cur_src1_row, 1);
+        row_mapping[src1_row] = {id, iid1};
+    }
+    __syncthreads();
+
+    const float * src1_row_original = (const float *)(src1_original + i11*nb11 + i12*nb12);
+    float * src1_row_contiguous = (float *)(src1_contiguous + src1_row*nb11);
+
+    for (int i = threadIdx.x; i < ne10; i += blockDim.x) {
+        src1_row_contiguous[i] = src1_row_original[i];
+    }
+}
+
+static __global__ void k_copy_dst_from_contiguous(char * __restrict__ dst_original, const char * __restrict__ dst_contiguous,
+                                                  const mmid_row_mapping * __restrict__ row_mapping,
+                                                  int64_t ne0,
+                                                  size_t nb1, size_t nb2) {
+    int32_t i = blockIdx.x;
+
+    const int32_t i1 = row_mapping[i].i1;
+    const int32_t i2 = row_mapping[i].i2;
+
+    const float * dst_row_contiguous = (const float *)(dst_contiguous + i*nb1);
+    float * dst_row_original = (float *)(dst_original + i1*nb1 + i2*nb2);
+
+    for (int j = threadIdx.x; j < ne0; j += blockDim.x) {
+        dst_row_original[j] = dst_row_contiguous[j];
+    }
+}
+
+static void ggml_cuda_mul_mat_id(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
+    const ggml_tensor * src0 = dst->src[0];
+    const ggml_tensor * src1 = dst->src[1];
+    const ggml_tensor * ids  = dst->src[2];
+
+    GGML_TENSOR_BINARY_OP_LOCALS
+
+    GGML_ASSERT(!ggml_backend_buffer_is_cuda_split(src0->buffer) && "mul_mat_id does not support split buffers");
+
+    cudaStream_t stream = ctx.stream();
+
+    const int64_t n_as = ne02;
+    const int64_t n_ids = ids->ne[0];
+
+    std::vector<char> ids_host(ggml_nbytes(ids));
+    const char * ids_dev = (const char *) ids->data;
+    CUDA_CHECK(cudaMemcpyAsync(ids_host.data(), ids_dev, ggml_nbytes(ids), cudaMemcpyDeviceToHost, stream));
+    CUDA_CHECK(cudaStreamSynchronize(stream));
+
+    ggml_tensor src0_row = *src0;
+    ggml_tensor src1_row = *src1;
+    ggml_tensor dst_row  = *dst;
+
+    char * src0_original = (char *) src0->data;
+    char * src1_original = (char *) src1->data;
+    char * dst_original  = (char *)  dst->data;
+
+    src0_row.ne[2] = 1;
+    src0_row.ne[3] = 1;
+    src0_row.nb[3] = nb02;
+
+    src1_row.ne[1] = 1;
+    src1_row.ne[2] = 1;
+    src1_row.ne[3] = 1;
+    src1_row.nb[2] = nb11;
+    src1_row.nb[3] = nb11;
+
+    dst_row.ne[1] = 1;
+    dst_row.ne[2] = 1;
+    dst_row.ne[3] = 1;
+    dst_row.nb[2] = nb1;
+    dst_row.nb[3] = nb1;
+
+    if (ne12 == 1) {
+        for (int64_t iid1 = 0; iid1 < ids->ne[1]; iid1++) {
+            for (int64_t id = 0; id < n_ids; id++) {
+                const int32_t i02 = *(const int32_t *) (ids_host.data() + iid1*ids->nb[1] + id*ids->nb[0]);
+
+                GGML_ASSERT(i02 >= 0 && i02 < n_as);
+
+                const int64_t i11 = id % ne11;
+                const int64_t i12 = iid1;
+
+                const int64_t i1 = id;
+                const int64_t i2 = i12;
+
+                src0_row.data = src0_original + i02*nb02;
+                src1_row.data = src1_original + i11*nb11 + i12*nb12;
+                dst_row.data  =  dst_original + i1*nb1   + i2*nb2;
+
+                ggml_cuda_mul_mat(ctx, &src0_row, &src1_row, &dst_row);
+            }
+        }
+    } else {
+        ggml_cuda_pool_alloc<char> src1_contiguous(ctx.pool(), sizeof(float)*ggml_nelements(src1));
+        ggml_cuda_pool_alloc<char>  dst_contiguous(ctx.pool(), sizeof(float)*ggml_nelements(dst));
+
+        src1_row.data = src1_contiguous.get();
+        dst_row.data  =  dst_contiguous.get();
+
+        for (int64_t i02 = 0; i02 < n_as; i02++) {
+            int64_t num_src1_rows = 0;
+
+            for (int64_t iid1 = 0; iid1 < ids->ne[1]; iid1++) {
+                for (int64_t id = 0; id < n_ids; id++) {
+                    const int32_t row_id_i = *(const int32_t *) (ids_host.data() + iid1*ids->nb[1] + id*ids->nb[0]);
+
+                    GGML_ASSERT(row_id_i >= 0 && row_id_i < n_as);
+
+                    if (row_id_i != i02) {
+                        continue;
+                    }
+
+                    num_src1_rows++;
+                }
+            }
+
+            if (num_src1_rows == 0) {
+                continue;
+            }
+
+            ggml_cuda_pool_alloc<int> dev_cur_src1_row(ctx.pool(), 1);
+            ggml_cuda_pool_alloc<mmid_row_mapping> dev_row_mapping(ctx.pool(), num_src1_rows);
+            CUDA_CHECK(cudaMemsetAsync(dev_cur_src1_row.get(), 0, sizeof(int), stream));
+
+            {
+                dim3 block_dims(std::min((unsigned int)ne10, 768u));
+                dim3 grid_dims(ids->ne[1], n_ids);
+                k_copy_src1_to_contiguous<<<grid_dims, block_dims, 0, stream>>>(
+                        src1_original, src1_contiguous.get(),
+                        dev_cur_src1_row.get(), dev_row_mapping.get(),
+                        ids_dev, i02, ids->nb[1], ids->nb[0],
+                        ne11, ne10,
+                        nb11, nb12);
+                CUDA_CHECK(cudaGetLastError());
+            }
+
+            src0_row.data = src0_original + i02*nb02;
+
+            GGML_ASSERT(nb11 == sizeof(float)*ne10);
+            GGML_ASSERT(nb1 == sizeof(float)*ne0);
+
+            src1_row.ne[1] = num_src1_rows;
+            src1_row.nb[1] = nb11;
+            src1_row.nb[2] = num_src1_rows*nb11;
+            src1_row.nb[3] = num_src1_rows*nb11;
+
+            dst_row.ne[1] = num_src1_rows;
+            dst_row.nb[1] = nb1;
+            dst_row.nb[2] = num_src1_rows*nb1;
+            dst_row.nb[3] = num_src1_rows*nb1;
+
+            ggml_cuda_mul_mat(ctx, &src0_row, &src1_row, &dst_row);
+
+            {
+                dim3 block_dims(std::min((unsigned int)ne0, 768u));
+                dim3 grid_dims(num_src1_rows);
+                k_copy_dst_from_contiguous<<<grid_dims, block_dims, 0, stream>>>(
+                        dst_original, dst_contiguous.get(),
+                        dev_row_mapping.get(),
+                        ne0,
+                        nb1, nb2);
+                CUDA_CHECK(cudaGetLastError());
+            }
+        }
+    }
+}
+
+static bool ggml_cuda_compute_forward(ggml_backend_cuda_context & ctx, struct ggml_tensor * dst) {
+    // why is this here instead of mul_mat?
+    if (dst->src[0] != nullptr && ggml_backend_buffer_is_cuda_split(dst->src[0]->buffer)) {
+        ggml_cuda_set_peer_access(dst->src[1]->ne[1], ctx.device);
+    }
+
+    switch (dst->op) {
+        case GGML_OP_ARGMAX:
+            ggml_cuda_argmax(ctx, dst);
+            break;
+        case GGML_OP_COUNT_EQUAL:
+            ggml_cuda_count_equal(ctx, dst);
+            break;
+        case GGML_OP_REPEAT:
+            ggml_cuda_op_repeat(ctx, dst);
+            break;
+        case GGML_OP_REPEAT_BACK:
+            ggml_cuda_op_repeat_back(ctx, dst);
+            break;
+        case GGML_OP_GET_ROWS:
+            ggml_cuda_op_get_rows(ctx, dst);
+            break;
+        case GGML_OP_DUP:
+            ggml_cuda_dup(ctx, dst);
+            break;
+        case GGML_OP_CPY:
+            ggml_cuda_cpy(ctx, dst->src[0], dst->src[1]);
+            break;
+        case GGML_OP_CONT:
+            ggml_cuda_dup(ctx, dst);
+            break;
+        case GGML_OP_ADD:
+        case GGML_OP_ADD1: // TODO: more efficient implementation
+            ggml_cuda_op_add(ctx, dst);
+            break;
+        case GGML_OP_SUB:
+            ggml_cuda_op_sub(ctx, dst);
+            break;
+        case GGML_OP_ACC:
+            ggml_cuda_op_acc(ctx, dst);
+            break;
+        case GGML_OP_MUL:
+            ggml_cuda_op_mul(ctx, dst);
+            break;
+        case GGML_OP_DIV:
+            ggml_cuda_op_div(ctx, dst);
+            break;
+        case GGML_OP_UNARY:
+            switch (ggml_get_unary_op(dst)) {
+                case GGML_UNARY_OP_NEG:
+                    ggml_cuda_op_neg(ctx, dst);
+                    break;
+                case GGML_UNARY_OP_STEP:
+                    ggml_cuda_op_step(ctx, dst);
+                    break;
+                case GGML_UNARY_OP_GELU:
+                    ggml_cuda_op_gelu(ctx, dst);
+                    break;
+                case GGML_UNARY_OP_SILU:
+                    ggml_cuda_op_silu(ctx, dst);
+                    break;
+                case GGML_UNARY_OP_GELU_QUICK:
+                    ggml_cuda_op_gelu_quick(ctx, dst);
+                    break;
+                case GGML_UNARY_OP_TANH:
+                    ggml_cuda_op_tanh(ctx, dst);
+                    break;
+                case GGML_UNARY_OP_RELU:
+                    ggml_cuda_op_relu(ctx, dst);
+                    break;
+                case GGML_UNARY_OP_SIGMOID:
+                    ggml_cuda_op_sigmoid(ctx, dst);
+                    break;
+                case GGML_UNARY_OP_HARDSIGMOID:
+                    ggml_cuda_op_hardsigmoid(ctx, dst);
+                    break;
+                case GGML_UNARY_OP_HARDSWISH:
+                    ggml_cuda_op_hardswish(ctx, dst);
+                    break;
+                case GGML_UNARY_OP_EXP:
+                    ggml_cuda_op_exp(ctx, dst);
+                    break;
+                default:
+                    return false;
+            }
+            break;
+        case GGML_OP_NORM:
+            ggml_cuda_op_norm(ctx, dst);
+            break;
+        case GGML_OP_GROUP_NORM:
+            ggml_cuda_op_group_norm(ctx, dst);
+            break;
+        case GGML_OP_CONCAT:
+            ggml_cuda_op_concat(ctx, dst);
+            break;
+        case GGML_OP_UPSCALE:
+            ggml_cuda_op_upscale(ctx, dst);
+            break;
+        case GGML_OP_PAD:
+            ggml_cuda_op_pad(ctx, dst);
+            break;
+        case GGML_OP_ARANGE:
+            ggml_cuda_op_arange(ctx, dst);
+            break;
+        case GGML_OP_TIMESTEP_EMBEDDING:
+            ggml_cuda_op_timestep_embedding(ctx, dst);
+            break;
+        case GGML_OP_LEAKY_RELU:
+            ggml_cuda_op_leaky_relu(ctx, dst);
+            break;
+        case GGML_OP_RMS_NORM:
+            ggml_cuda_op_rms_norm(ctx, dst);
+            break;
+        case GGML_OP_MUL_MAT:
+            if (dst->src[0]->ne[3] != dst->src[1]->ne[3]) {
+                GGML_LOG_ERROR("%s: cannot compute %s: src0->ne[3] = %" PRId64 ", src1->ne[3] = %" PRId64 " - fallback to CPU\n", __func__, dst->name, dst->src[0]->ne[3], dst->src[1]->ne[3]);
+                return false;
+            } else {
+                ggml_cuda_mul_mat(ctx, dst->src[0], dst->src[1], dst);
+            }
+            break;
+        case GGML_OP_MUL_MAT_ID:
+            ggml_cuda_mul_mat_id(ctx, dst);
+            break;
+        case GGML_OP_OUT_PROD:
+            ggml_cuda_out_prod(ctx, dst);
+            break;
+        case GGML_OP_SCALE:
+            ggml_cuda_op_scale(ctx, dst);
+            break;
+        case GGML_OP_SQR:
+            ggml_cuda_op_sqr(ctx, dst);
+            break;
+        case GGML_OP_SQRT:
+            ggml_cuda_op_sqrt(ctx, dst);
+            break;
+        case GGML_OP_SIN:
+            ggml_cuda_op_sin(ctx, dst);
+            break;
+        case GGML_OP_COS:
+            ggml_cuda_op_cos(ctx, dst);
+            break;
+        case GGML_OP_CLAMP:
+            ggml_cuda_op_clamp(ctx, dst);
+            break;
+        case GGML_OP_NONE:
+        case GGML_OP_RESHAPE:
+        case GGML_OP_VIEW:
+        case GGML_OP_PERMUTE:
+        case GGML_OP_TRANSPOSE:
+                break;
+        case GGML_OP_DIAG_MASK_INF:
+            ggml_cuda_op_diag_mask_inf(ctx, dst);
+            break;
+        case GGML_OP_SOFT_MAX:
+            ggml_cuda_op_soft_max(ctx, dst);
+            break;
+        case GGML_OP_ROPE:
+            ggml_cuda_op_rope(ctx, dst);
+            break;
+        case GGML_OP_IM2COL:
+            ggml_cuda_op_im2col(ctx, dst);
+            break;
+        case GGML_OP_CONV_TRANSPOSE_1D:
+            ggml_cuda_op_conv_transpose_1d(ctx,dst);
+            break;
+        case GGML_OP_POOL_2D:
+            ggml_cuda_op_pool2d(ctx, dst);
+            break;
+        case GGML_OP_SUM:
+            ggml_cuda_op_sum(ctx, dst);
+            break;
+        case GGML_OP_SUM_ROWS:
+            ggml_cuda_op_sum_rows(ctx, dst);
+            break;
+        case GGML_OP_ARGSORT:
+            ggml_cuda_op_argsort(ctx, dst);
+            break;
+        case GGML_OP_FLASH_ATTN_EXT:
+            ggml_cuda_flash_attn_ext(ctx, dst);
+            break;
+        case GGML_OP_CROSS_ENTROPY_LOSS:
+            ggml_cuda_cross_entropy_loss(ctx, dst);
+            break;
+        case GGML_OP_RWKV_WKV:
+            ggml_cuda_op_rwkv_wkv(ctx, dst);
+            break;
+        case GGML_OP_CROSS_ENTROPY_LOSS_BACK:
+            ggml_cuda_cross_entropy_loss_back(ctx, dst);
+            break;
+        case GGML_OP_OPT_STEP_ADAMW:
+            ggml_cuda_opt_step_adamw(ctx, dst);
+            break;
+        default:
+            return false;
+    }
+
+    cudaError_t err = cudaGetLastError();
+    if (err != cudaSuccess) {
+        GGML_LOG_ERROR("%s: %s failed\n", __func__, ggml_op_desc(dst));
+        CUDA_CHECK(err);
+    }
+
+    return true;
+}
+
+////////////////////////////////////////////////////////////////////////////////
+
+// backend
+
+static const char * ggml_backend_cuda_get_name(ggml_backend_t backend) {
+    ggml_backend_cuda_context * cuda_ctx = (ggml_backend_cuda_context *)backend->context;
+
+    return cuda_ctx->name.c_str();
+}
+
+static void ggml_backend_cuda_free(ggml_backend_t backend) {
+    ggml_backend_cuda_context * cuda_ctx = (ggml_backend_cuda_context *)backend->context;
+
+    delete cuda_ctx;
+    delete backend;
+}
+
+static ggml_backend_buffer_type_t ggml_backend_cuda_get_default_buffer_type(ggml_backend_t backend) {
+    ggml_backend_cuda_context * cuda_ctx = (ggml_backend_cuda_context *)backend->context;
+
+    return ggml_backend_cuda_buffer_type(cuda_ctx->device);
+}
+
+static void ggml_backend_cuda_set_tensor_async(ggml_backend_t backend, ggml_tensor * tensor, const void * data, size_t offset, size_t size) {
+    ggml_backend_cuda_context * cuda_ctx = (ggml_backend_cuda_context *)backend->context;
+    ggml_backend_buffer_t buf = tensor->view_src ? tensor->view_src->buffer : tensor->buffer;
+
+    GGML_ASSERT(buf->buft == ggml_backend_cuda_buffer_type(cuda_ctx->device) && "unsupported buffer type");
+
+    CUDA_CHECK(cudaMemcpyAsync((char *)tensor->data + offset, data, size, cudaMemcpyHostToDevice, cuda_ctx->stream()));
+}
+
+static void ggml_backend_cuda_get_tensor_async(ggml_backend_t backend, const ggml_tensor * tensor, void * data, size_t offset, size_t size) {
+    ggml_backend_cuda_context * cuda_ctx = (ggml_backend_cuda_context *)backend->context;
+    ggml_backend_buffer_t buf = tensor->view_src ? tensor->view_src->buffer : tensor->buffer;
+
+    GGML_ASSERT(buf->buft == ggml_backend_cuda_buffer_type(cuda_ctx->device) && "unsupported buffer type");
+
+    CUDA_CHECK(cudaMemcpyAsync(data, (const char *)tensor->data + offset, size, cudaMemcpyDeviceToHost, cuda_ctx->stream()));
+}
+
+static bool ggml_backend_cuda_cpy_tensor_async(ggml_backend_t backend_src, ggml_backend_t backend_dst, const ggml_tensor * src, ggml_tensor * dst) {
+    ggml_backend_buffer_t buf_src = src->view_src ? src->view_src->buffer : src->buffer;
+    ggml_backend_buffer_t buf_dst = dst->view_src ? dst->view_src->buffer : dst->buffer;
+
+    if (!ggml_backend_is_cuda(backend_src) || !ggml_backend_is_cuda(backend_dst)) {
+        return false;
+    }
+
+    if (!ggml_backend_buffer_is_cuda(src->buffer) || !ggml_backend_buffer_is_cuda(dst->buffer)) {
+        return false;
+    }
+
+    // device -> device copy
+    ggml_backend_cuda_context * cuda_ctx_src = (ggml_backend_cuda_context *)backend_src->context;
+    ggml_backend_cuda_context * cuda_ctx_dst = (ggml_backend_cuda_context *)backend_dst->context;
+
+    ggml_backend_cuda_buffer_context * buf_ctx_src = (ggml_backend_cuda_buffer_context *)buf_src->context;
+    ggml_backend_cuda_buffer_context * buf_ctx_dst = (ggml_backend_cuda_buffer_context *)buf_dst->context;
+
+    if (cuda_ctx_src->device != buf_ctx_src->device || cuda_ctx_dst->device != buf_ctx_dst->device) {
+#ifndef NDEBUG
+        GGML_LOG_WARN("%s: backend and buffer devices do not match\n", __func__);
+#endif
+        return false;
+    }
+
+    if (backend_src != backend_dst) {
+        // copy on src stream
+        if (cuda_ctx_src->device == cuda_ctx_dst->device) {
+            CUDA_CHECK(cudaMemcpyAsync(dst->data, src->data, ggml_nbytes(dst), cudaMemcpyDeviceToDevice, cuda_ctx_src->stream()));
+        } else {
+#ifdef GGML_CUDA_NO_PEER_COPY
+            return false;
+#else
+            CUDA_CHECK(cudaMemcpyPeerAsync(dst->data, cuda_ctx_dst->device, src->data, cuda_ctx_src->device, ggml_nbytes(dst), cuda_ctx_src->stream()));
+#endif
+        }
+
+        // record event on src stream after the copy
+        if (!cuda_ctx_src->copy_event) {
+            ggml_cuda_set_device(cuda_ctx_src->device);
+            CUDA_CHECK(cudaEventCreateWithFlags(&cuda_ctx_src->copy_event, cudaEventDisableTiming));
+        }
+
+        CUDA_CHECK(cudaEventRecord(cuda_ctx_src->copy_event, cuda_ctx_src->stream()));
+
+        // wait on dst stream for the copy to complete
+        CUDA_CHECK(cudaStreamWaitEvent(cuda_ctx_dst->stream(), cuda_ctx_src->copy_event, 0));
+    } else {
+        // src and dst are on the same backend
+        CUDA_CHECK(cudaMemcpyAsync(dst->data, src->data, ggml_nbytes(dst), cudaMemcpyDeviceToDevice, cuda_ctx_src->stream()));
+    }
+    return true;
+}
+
+static void ggml_backend_cuda_synchronize(ggml_backend_t backend) {
+    ggml_backend_cuda_context * cuda_ctx = (ggml_backend_cuda_context *)backend->context;
+
+    CUDA_CHECK(cudaStreamSynchronize(cuda_ctx->stream()));
+
+    GGML_UNUSED(backend);
+}
+
+#ifdef USE_CUDA_GRAPH
+static void set_ggml_graph_node_properties(ggml_tensor * node, ggml_graph_node_properties * graph_node_properties) {
+    graph_node_properties->node_address = node->data;
+    graph_node_properties->node_op = node->op;
+    for (int i = 0; i < GGML_MAX_DIMS; i++) {
+        graph_node_properties->ne[i] = node->ne[i];
+        graph_node_properties->nb[i] = node->nb[i];
+    }
+    for (int i = 0; i < GGML_MAX_SRC; i++) {
+        graph_node_properties->src_address[i] = node->src[i] ? node->src[i]->data : nullptr;
+    }
+    memcpy(graph_node_properties->op_params, node->op_params, GGML_MAX_OP_PARAMS);
+}
+
+static bool ggml_graph_node_has_matching_properties(ggml_tensor * node, ggml_graph_node_properties * graph_node_properties) {
+    if (node->data != graph_node_properties->node_address &&
+          node->op != GGML_OP_CPY &&
+          node->op != GGML_OP_VIEW) {
+        return false;
+    }
+
+    if (node->op != graph_node_properties->node_op) {
+        return false;
+    }
+
+    for (int i = 0; i < GGML_MAX_DIMS; i++) {
+        if (node->ne[i] != graph_node_properties->ne[i]) {
+            return false;
+        }
+        if (node->nb[i] != graph_node_properties->nb[i]) {
+            return false;
+        }
+    }
+
+    for (int i = 0; i < GGML_MAX_SRC; i++) {
+        if (node->src[i] &&
+            node->src[i]->data != graph_node_properties->src_address[i] &&
+            node->op != GGML_OP_CPY &&
+            node->op != GGML_OP_VIEW
+        ) {
+            return false;
+        }
+    }
+
+    if (node->op == GGML_OP_SCALE &&
+        memcmp(graph_node_properties->op_params, node->op_params, GGML_MAX_OP_PARAMS) != 0) {
+        return false;
+    }
+
+    return true;
+}
+#endif
+
+static enum ggml_status ggml_backend_cuda_graph_compute(ggml_backend_t backend, ggml_cgraph * cgraph) {
+    ggml_backend_cuda_context * cuda_ctx = (ggml_backend_cuda_context *)backend->context;
+
+    ggml_cuda_set_device(cuda_ctx->device);
+
+#ifdef USE_CUDA_GRAPH
+    static const bool disable_cuda_graphs_due_to_env = (getenv("GGML_CUDA_DISABLE_GRAPHS") != nullptr);
+
+    // Objects required for CUDA Graph
+    if (cuda_ctx->cuda_graph == nullptr) {
+        cuda_ctx->cuda_graph.reset(new ggml_cuda_graph());
+    }
+
+    bool use_cuda_graph = true;
+    bool cuda_graph_update_required = false;
+    // vector of pointers to CUDA cpy kernels, which are required to identify
+    // kernel parameters which need updated in the graph for each token
+    std::vector<void *> ggml_cuda_cpy_fn_ptrs;
+
+    if (cuda_ctx->cuda_graph->graph == nullptr) {
+        if (ggml_cuda_info().devices[cuda_ctx->device].cc < CC_AMPERE) {
+            cuda_ctx->cuda_graph->disable_due_to_gpu_arch = true;
+#ifndef NDEBUG
+            GGML_LOG_WARN("%s: disabling CUDA graphs due to GPU architecture\n", __func__);
+#endif
+        }
+    }
+
+    // Disable CUDA graphs in presence of env var, old GPU, use-case which is changing too rapidly,
+    // or previous graph capture failure.
+    // Also disable for multi-gpu for now. TO DO investigate
+    if (disable_cuda_graphs_due_to_env
+        || cuda_ctx->cuda_graph->disable_due_to_gpu_arch
+        || cuda_ctx->cuda_graph->disable_due_to_too_many_updates
+        || cuda_ctx->cuda_graph->disable_due_to_failed_graph_capture) {
+        use_cuda_graph = false;
+    }
+
+    if (use_cuda_graph) {
+        if (cuda_ctx->cuda_graph->instance == nullptr) {
+            cuda_graph_update_required = true;
+        }
+
+        // Check if the graph size has changed
+        if (cuda_ctx->cuda_graph->ggml_graph_properties.size() != (size_t)cgraph->n_nodes) {
+            cuda_graph_update_required = true;
+            cuda_ctx->cuda_graph->ggml_graph_properties.resize(cgraph->n_nodes);
+        }
+
+        // Loop over nodes in GGML graph to determine if CUDA graph update is required
+        // and store properties to allow this comparison for the next token
+        for (int i = 0; i < cgraph->n_nodes; i++) {
+            bool has_matching_properties = true;
+            if (!cuda_graph_update_required) {
+                has_matching_properties = ggml_graph_node_has_matching_properties(cgraph->nodes[i], &cuda_ctx->cuda_graph->ggml_graph_properties[i]);
+            }
+            if (!has_matching_properties) {
+                cuda_graph_update_required = true;
+            }
+            set_ggml_graph_node_properties(cgraph->nodes[i], &cuda_ctx->cuda_graph->ggml_graph_properties[i]);
+        }
+
+        // Loop over nodes in GGML graph to obtain info needed for CUDA graph
+        cuda_ctx->cuda_graph->updated_kernel_arg.clear();
+        for (int i = 0; i < cgraph->n_nodes; i++) {
+            ggml_tensor * node = cgraph->nodes[i];
+
+            if (ggml_is_empty(node) || node->op == GGML_OP_RESHAPE || node->op == GGML_OP_TRANSPOSE || node->op == GGML_OP_VIEW || node->op == GGML_OP_PERMUTE || node->op == GGML_OP_NONE) {
+                continue;
+            }
+
+            if (node->src[0] && node->src[0]->buffer && ggml_backend_buffer_is_cuda_split(node->src[0]->buffer)) {
+                use_cuda_graph = false; // Split buffers are not supported by CUDA graph capture
+#ifndef NDEBUG
+                GGML_LOG_WARN("%s: disabling CUDA graphs due to split buffer\n", __func__);
+#endif
+            }
+
+            if (node->op == GGML_OP_MUL_MAT_ID) {
+                use_cuda_graph = false; // This node type is not supported by CUDA graph capture
+#ifndef NDEBUG
+                GGML_LOG_WARN("%s: disabling CUDA graphs due to mul_mat_id\n", __func__);
+#endif
+            }
+
+            if (node->op == GGML_OP_ADD && node->src[1] && node->src[1]->ne[1] > 1) {
+                // disable CUDA graphs for batch size > 1 for now.
+                // Changes in batch size or context size can cause changes to the grid size of some kernels.
+                use_cuda_graph = false;
+#ifndef NDEBUG
+                GGML_LOG_WARN("%s: disabling CUDA graphs due to batch size > 1 [%s] [%ld %ld %ld %ld]\n", __func__, node->name, node->ne[0], node->ne[1], node->ne[2], node->ne[3]);
+#endif
+            }
+
+            if (node->op == GGML_OP_CPY) {
+                // store the copy op parameter which changes with each token.
+                cuda_ctx->cuda_graph->updated_kernel_arg.push_back((char **) &(node->src[1]->data));
+                // store a pointer to each copy op CUDA kernel to identify it later
+                void * ptr = ggml_cuda_cpy_fn(node->src[0], node->src[1]);
+                if (!ptr) {
+                    use_cuda_graph = false;
+#ifndef NDEBUG
+                    GGML_LOG_WARN("%s: disabling CUDA graphs due to unsupported copy op\n", __func__);
+#endif
+                } else {
+                    if (std::find(ggml_cuda_cpy_fn_ptrs.begin(), ggml_cuda_cpy_fn_ptrs.end(), ptr) == ggml_cuda_cpy_fn_ptrs.end()) {
+                        ggml_cuda_cpy_fn_ptrs.push_back(ptr);
+                    }
+                }
+            }
+
+            if (!use_cuda_graph) {
+                break;
+            }
+        }
+
+        // Disable CUDA graphs (from the next token) if the use-case is demanding too many consecutive graph updates.
+        if (use_cuda_graph && cuda_graph_update_required) {
+            cuda_ctx->cuda_graph->number_consecutive_updates++;
+        } else {
+            cuda_ctx->cuda_graph->number_consecutive_updates = 0;
+        }
+
+        if (cuda_ctx->cuda_graph->number_consecutive_updates >= 4) {
+            cuda_ctx->cuda_graph->disable_due_to_too_many_updates = true;
+#ifndef NDEBUG
+            GGML_LOG_WARN("%s: disabling CUDA graphs due to too many consecutive updates\n", __func__);
+#endif
+        }
+    }
+
+    if (use_cuda_graph && cuda_graph_update_required) { // Start CUDA graph capture
+        CUDA_CHECK(cudaStreamBeginCapture(cuda_ctx->stream(), cudaStreamCaptureModeRelaxed));
+    }
+
+#else
+    bool use_cuda_graph = false;
+    bool cuda_graph_update_required = false;
+#endif // USE_CUDA_GRAPH
+
+    bool graph_evaluated_or_captured = false;
+
+    while (!graph_evaluated_or_captured) {
+        // Only perform the graph execution if CUDA graphs are not enabled, or we are capturing the graph.
+        // With the use of CUDA graphs, the execution will be performed by the graph launch.
+        if (!use_cuda_graph || cuda_graph_update_required) {
+            for (int i = 0; i < cgraph->n_nodes; i++) {
+                ggml_tensor * node = cgraph->nodes[i];
+
+                if (ggml_is_empty(node) || node->op == GGML_OP_RESHAPE || node->op == GGML_OP_TRANSPOSE || node->op == GGML_OP_VIEW || node->op == GGML_OP_PERMUTE || node->op == GGML_OP_NONE) {
+                    continue;
+                }
+
+#ifndef NDEBUG
+                assert(node->buffer->buft == ggml_backend_cuda_buffer_type(cuda_ctx->device));
+                for (int j = 0; j < GGML_MAX_SRC; j++) {
+                    if (node->src[j] != nullptr) {
+                        assert(node->src[j]->buffer);
+                        assert(node->src[j]->buffer->buft == ggml_backend_cuda_buffer_type(cuda_ctx->device) || ggml_backend_buffer_is_cuda_split(node->src[j]->buffer));
+                    }
+                }
+#endif
+
+                bool ok = ggml_cuda_compute_forward(*cuda_ctx, node);
+                if (!ok) {
+                    GGML_LOG_ERROR("%s: op not supported %s (%s)\n", __func__, node->name, ggml_op_name(node->op));
+                }
+                GGML_ASSERT(ok);
+            }
+        }
+
+#ifdef USE_CUDA_GRAPH
+        if (use_cuda_graph && cuda_graph_update_required) { // End CUDA graph capture
+            if (cuda_ctx->cuda_graph->graph != nullptr) {
+                CUDA_CHECK(cudaGraphDestroy(cuda_ctx->cuda_graph->graph));
+                cuda_ctx->cuda_graph->graph = nullptr;
+            }
+            CUDA_CHECK(cudaStreamEndCapture(cuda_ctx->stream(), &cuda_ctx->cuda_graph->graph));
+
+#if 0
+            if (disable_cuda_graphs_due_to_failed_capture) {
+                use_cuda_graph = false;
+                cuda_ctx->cuda_graph->disable_due_to_failed_graph_capture = true;
+#ifndef NDEBUG
+                GGML_LOG_WARN("%s: disabling CUDA graphs due to failed graph capture\n", __func__);
+#endif
+            } else {
+                graph_evaluated_or_captured = true; // CUDA graph has been captured
+            }
+#endif
+            graph_evaluated_or_captured = true; // CUDA graph has been captured
+        } else {
+            graph_evaluated_or_captured = true; // ggml graph has been directly evaluated
+        }
+    }
+
+    if (use_cuda_graph) {
+        if (cuda_ctx->cuda_graph->instance == nullptr) { // Create executable graph from captured graph.
+            CUDA_CHECK(cudaGraphInstantiate(&cuda_ctx->cuda_graph->instance, cuda_ctx->cuda_graph->graph, NULL, NULL, 0));
+        }
+
+        // Perform update to graph (if required for this token), and change copy parameter (required for every token)
+
+        if (cuda_graph_update_required) {
+            // Extract nodes from graph
+            // First call with null argument gets number of nodes in graph
+            CUDA_CHECK(cudaGraphGetNodes(cuda_ctx->cuda_graph->graph, nullptr, &cuda_ctx->cuda_graph->num_nodes));
+            // Subsequent call with non-null argument gets nodes
+            cuda_ctx->cuda_graph->nodes.clear();
+            cuda_ctx->cuda_graph->nodes.resize(cuda_ctx->cuda_graph->num_nodes);
+            cuda_ctx->cuda_graph->params.clear();
+            cuda_ctx->cuda_graph->params.resize(cuda_ctx->cuda_graph->num_nodes);
+            if (cuda_ctx->cuda_graph->num_nodes > 0) {
+                CUDA_CHECK(cudaGraphGetNodes(cuda_ctx->cuda_graph->graph, cuda_ctx->cuda_graph->nodes.data(), &cuda_ctx->cuda_graph->num_nodes));
+
+                // Loop over nodes, and extract kernel parameters from each node
+                for (size_t i = 0; i < cuda_ctx->cuda_graph->num_nodes; i++) {
+                    cudaGraphNodeType node_type;
+                    CUDA_CHECK(cudaGraphNodeGetType(cuda_ctx->cuda_graph->nodes[i], &node_type));
+                    if (node_type == cudaGraphNodeTypeKernel) {
+                        cudaError_t stat = cudaGraphKernelNodeGetParams(cuda_ctx->cuda_graph->nodes[i], &cuda_ctx->cuda_graph->params[i]); // Get params using runtime
+                        if (stat == cudaErrorInvalidDeviceFunction) {
+                            // Fails due to incorrect handling by CUDA runtime of CUDA BLAS node.
+                            // We don't need to update blas nodes, so clear error and move on.
+                            cudaGetLastError();
+                        } else {
+                            GGML_ASSERT(stat == cudaSuccess);
+                        }
+                    }
+                }
+            }
+        }
+
+        // One of the arguments to the copy kernel is updated for each token, hence we need to
+        // replace that argument with the updated value in the CUDA graph
+        if (!cuda_graph_update_required) { // on update steps, the live parameters will already be captured
+            int k = 0;
+            for (size_t i = 0; i < cuda_ctx->cuda_graph->num_nodes; i++) {
+                if(count(ggml_cuda_cpy_fn_ptrs.begin(), ggml_cuda_cpy_fn_ptrs.end(), cuda_ctx->cuda_graph->params[i].func) > 0) {
+                    char ** updated_kernel_arg_ptr = cuda_ctx->cuda_graph->updated_kernel_arg.at(k++);
+                    cuda_ctx->cuda_graph->params[i].kernelParams[1] = updated_kernel_arg_ptr;
+                    CUDA_CHECK(cudaGraphKernelNodeSetParams(cuda_ctx->cuda_graph->nodes[i], &cuda_ctx->cuda_graph->params[i]));
+                }
+            }
+        }
+
+        // Update graph executable
+        cudaGraphExecUpdateResultInfo result_info;
+        cudaError_t stat = cudaGraphExecUpdate(cuda_ctx->cuda_graph->instance, cuda_ctx->cuda_graph->graph, &result_info);
+        if (stat == cudaErrorGraphExecUpdateFailure) {
+#ifndef NDEBUG
+            GGML_LOG_ERROR("%s: CUDA graph update failed\n", __func__);
+#endif
+            // The pre-existing graph exec cannot be updated due to violated constraints
+            // so instead clear error and re-instantiate
+            cudaGetLastError();
+            CUDA_CHECK(cudaGraphExecDestroy(cuda_ctx->cuda_graph->instance));
+            cuda_ctx->cuda_graph->instance = nullptr;
+            CUDA_CHECK(cudaGraphInstantiate(&cuda_ctx->cuda_graph->instance, cuda_ctx->cuda_graph->graph, NULL, NULL, 0));
+        } else {
+            GGML_ASSERT(stat == cudaSuccess);
+        }
+        // Launch graph
+        CUDA_CHECK(cudaGraphLaunch(cuda_ctx->cuda_graph->instance, cuda_ctx->stream()));
+#else
+        graph_evaluated_or_captured = true;
+#endif // USE_CUDA_GRAPH
+    }
+
+    return GGML_STATUS_SUCCESS;
+}
+
+static void ggml_backend_cuda_event_record(ggml_backend_t backend, ggml_backend_event_t event) {
+    ggml_backend_cuda_context * cuda_ctx = (ggml_backend_cuda_context *)backend->context;
+
+    CUDA_CHECK(cudaEventRecord((cudaEvent_t)event->context, cuda_ctx->stream()));
+}
+
+static void ggml_backend_cuda_event_wait(ggml_backend_t backend, ggml_backend_event_t event) {
+    ggml_backend_cuda_context * cuda_ctx = (ggml_backend_cuda_context *)backend->context;
+
+    if (ggml_backend_is_cuda(backend)) {
+        CUDA_CHECK(cudaStreamWaitEvent(cuda_ctx->stream(), (cudaEvent_t)event->context, 0));
+    } else {
+#if 0
+        // untested
+        auto wait_fn = [](void * user_data) {
+            ggml_backend_event_t event = (ggml_backend_event_t)user_data;
+            ggml_backend_event_synchronize(event);
+        };
+
+        CUDA_CHECK(cudaLaunchHostFunc(cuda_ctx->stream(), wait_fn, event));
+#endif
+        GGML_ABORT("fatal error");
+    }
+}
+
+static const ggml_backend_i ggml_backend_cuda_interface = {
+    /* .get_name                = */ ggml_backend_cuda_get_name,
+    /* .free                    = */ ggml_backend_cuda_free,
+    /* .get_default_buffer_type = */ ggml_backend_cuda_get_default_buffer_type,
+    /* .set_tensor_async        = */ ggml_backend_cuda_set_tensor_async,
+    /* .get_tensor_async        = */ ggml_backend_cuda_get_tensor_async,
+    /* .cpy_tensor_async        = */ ggml_backend_cuda_cpy_tensor_async,
+    /* .synchronize             = */ ggml_backend_cuda_synchronize,
+    /* .graph_plan_create       = */ NULL,
+    /* .graph_plan_free         = */ NULL,
+    /* .graph_plan_update       = */ NULL,
+    /* .graph_plan_compute      = */ NULL,
+    /* .graph_compute           = */ ggml_backend_cuda_graph_compute,
+    /* .supports_op             = */ NULL, // moved to device
+    /* .supports_buft           = */ NULL, // moved to device
+    /* .offload_op              = */ NULL, // moved to device
+    /* .event_record            = */ ggml_backend_cuda_event_record,
+    /* .event_wait              = */ ggml_backend_cuda_event_wait,
+};
+
+static ggml_guid_t ggml_backend_cuda_guid() {
+    static ggml_guid guid = { 0x2c, 0xdd, 0xe8, 0x1c, 0x65, 0xb3, 0x65, 0x73, 0x6a, 0x12, 0x88, 0x61, 0x1c, 0xc9, 0xdc, 0x25 };
+    return &guid;
+}
+
+bool ggml_backend_is_cuda(ggml_backend_t backend) {
+    return backend != NULL && ggml_guid_matches(backend->guid, ggml_backend_cuda_guid());
+}
+
+int ggml_backend_cuda_get_device_count() {
+    return ggml_cuda_info().device_count;
+}
+
+void ggml_backend_cuda_get_device_description(int device, char * description, size_t description_size) {
+    cudaDeviceProp prop;
+    CUDA_CHECK(cudaGetDeviceProperties(&prop, device));
+    snprintf(description, description_size, "%s", prop.name);
+}
+
+void ggml_backend_cuda_get_device_memory(int device, size_t * free, size_t * total) {
+    ggml_cuda_set_device(device);
+
+    CUDA_CHECK(cudaMemGetInfo(free, total));
+}
+
+bool ggml_backend_cuda_register_host_buffer(void * buffer, size_t size) {
+    if (getenv("GGML_CUDA_REGISTER_HOST") == nullptr) {
+        return false;
+    }
+
+#if CUDART_VERSION >= 11100 || defined(GGML_USE_MUSA)
+    cudaError_t err = cudaHostRegister(buffer, size, cudaHostRegisterPortable | cudaHostRegisterReadOnly);
+    if (err != cudaSuccess) {
+        // clear the error
+        cudaGetLastError();
+
+        GGML_LOG_WARN("%s: failed to register %.2f MiB of pinned memory: %s\n", __func__,
+                           size / 1024.0 / 1024.0, cudaGetErrorString(err));
+        return false;
+    }
+    return true;
+#else
+    return false;
+#endif
+}
+
+void ggml_backend_cuda_unregister_host_buffer(void * buffer) {
+    if (getenv("GGML_CUDA_REGISTER_HOST") == nullptr) {
+        return;
+    }
+
+    cudaError_t err = cudaHostUnregister(buffer);
+    if (err != cudaSuccess) {
+        // clear the error
+        cudaGetLastError();
+    }
+}
+
+
+// backend device
+
+struct ggml_backend_cuda_device_context {
+    int device;
+    std::string name;
+    std::string description;
+};
+
+static const char * ggml_backend_cuda_device_get_name(ggml_backend_dev_t dev) {
+    ggml_backend_cuda_device_context * ctx = (ggml_backend_cuda_device_context *)dev->context;
+    return ctx->name.c_str();
+}
+
+static const char * ggml_backend_cuda_device_get_description(ggml_backend_dev_t dev) {
+    ggml_backend_cuda_device_context * ctx = (ggml_backend_cuda_device_context *)dev->context;
+    return ctx->description.c_str();
+}
+
+static void ggml_backend_cuda_device_get_memory(ggml_backend_dev_t dev, size_t * free, size_t * total) {
+    ggml_backend_cuda_device_context * ctx = (ggml_backend_cuda_device_context *)dev->context;
+    ggml_cuda_set_device(ctx->device);
+    CUDA_CHECK(cudaMemGetInfo(free, total));
+}
+
+static enum ggml_backend_dev_type ggml_backend_cuda_device_get_type(ggml_backend_dev_t dev) {
+    GGML_UNUSED(dev);
+    return GGML_BACKEND_DEVICE_TYPE_GPU_FULL;
+}
+
+static void ggml_backend_cuda_device_get_props(ggml_backend_dev_t dev, ggml_backend_dev_props * props) {
+    props->name        = ggml_backend_cuda_device_get_name(dev);
+    props->description = ggml_backend_cuda_device_get_description(dev);
+    props->type        = ggml_backend_cuda_device_get_type(dev);
+    ggml_backend_cuda_device_get_memory(dev, &props->memory_free, &props->memory_total);
+
+    bool host_buffer = getenv("GGML_CUDA_NO_PINNED") == nullptr;
+#ifdef GGML_CUDA_NO_PEER_COPY
+    bool events = false;
+#else
+    bool events = true;
+#endif
+
+    props->caps = {
+        /* async       */ true,
+        /* host_buffer */ host_buffer,
+        /* events      */ events,
+    };
+}
+
+static ggml_backend_t ggml_backend_cuda_device_init(ggml_backend_dev_t dev, const char * params) {
+    GGML_UNUSED(params);
+    ggml_backend_cuda_device_context * ctx = (ggml_backend_cuda_device_context *)dev->context;
+    return ggml_backend_cuda_init(ctx->device);
+}
+
+static ggml_backend_buffer_type_t ggml_backend_cuda_device_get_buffer_type(ggml_backend_dev_t dev) {
+    ggml_backend_cuda_device_context * ctx = (ggml_backend_cuda_device_context *)dev->context;
+    return ggml_backend_cuda_buffer_type(ctx->device);
+}
+
+static ggml_backend_buffer_type_t ggml_backend_cuda_device_get_host_buffer_type(ggml_backend_dev_t dev) {
+    GGML_UNUSED(dev);
+    return ggml_backend_cuda_host_buffer_type();
+}
+
+static ggml_backend_buffer_t ggml_backend_cuda_device_buffer_from_host_ptr(ggml_backend_dev_t dev, void * ptr, size_t size, size_t max_tensor_size) {
+    GGML_UNUSED(dev);
+    GGML_UNUSED(ptr);
+    GGML_UNUSED(size);
+    GGML_UNUSED(max_tensor_size);
+    return nullptr;
+}
+
+// TODO: move these functions here
+static bool ggml_backend_cuda_device_supports_op(ggml_backend_dev_t dev, const ggml_tensor * op) {
+    ggml_backend_cuda_device_context * dev_ctx = (ggml_backend_cuda_device_context *) dev->context;
+
+    switch (op->op) {
+        case GGML_OP_UNARY:
+            switch (ggml_get_unary_op(op)) {
+                case GGML_UNARY_OP_NEG:
+                case GGML_UNARY_OP_STEP:
+                case GGML_UNARY_OP_GELU:
+                case GGML_UNARY_OP_SILU:
+                case GGML_UNARY_OP_RELU:
+                case GGML_UNARY_OP_SIGMOID:
+                case GGML_UNARY_OP_HARDSIGMOID:
+                case GGML_UNARY_OP_HARDSWISH:
+                case GGML_UNARY_OP_GELU_QUICK:
+                case GGML_UNARY_OP_TANH:
+                case GGML_UNARY_OP_EXP:
+                    return ggml_is_contiguous(op->src[0]);
+                default:
+                    return false;
+            }
+            break;
+        case GGML_OP_MUL_MAT:
+        case GGML_OP_MUL_MAT_ID:
+            {
+                struct ggml_tensor * a = op->src[0];
+                struct ggml_tensor * b = op->src[1];
+                if (b->type == GGML_TYPE_F16 && a->type != GGML_TYPE_F16) {
+                    return false;
+                }
+                if (op->op == GGML_OP_MUL_MAT && a->ne[3] != b->ne[3]) {
+                    return false;
+                }
+#ifdef GGML_USE_MUSA
+                if (b->type == GGML_TYPE_F16 && b->ne[2]*b->ne[3] > 1 &&
+                    !ggml_is_transposed(a) && !ggml_is_transposed(b)) {
+                    return false;
+                }
+#endif // GGML_USE_MUSA
+                switch (a->type) {
+                    case GGML_TYPE_F32:
+                    case GGML_TYPE_F16:
+                    case GGML_TYPE_Q4_0:
+                    case GGML_TYPE_Q4_1:
+                    case GGML_TYPE_Q5_0:
+                    case GGML_TYPE_Q5_1:
+                    case GGML_TYPE_Q8_0:
+                    case GGML_TYPE_Q2_K:
+                    case GGML_TYPE_Q3_K:
+                    case GGML_TYPE_Q4_K:
+                    case GGML_TYPE_Q5_K:
+                    case GGML_TYPE_Q6_K:
+                    case GGML_TYPE_Q8_K:
+                    case GGML_TYPE_IQ1_M:
+                    case GGML_TYPE_IQ1_S:
+                    case GGML_TYPE_IQ2_S:
+                    case GGML_TYPE_IQ2_XS:
+                    case GGML_TYPE_IQ2_XXS:
+                    case GGML_TYPE_IQ3_S:
+                    case GGML_TYPE_IQ3_XXS:
+                    case GGML_TYPE_IQ4_NL:
+                    case GGML_TYPE_IQ4_XS:
+#ifdef GGML_USE_MUSA
+                        if (a->type == GGML_TYPE_Q3_K) {
+                            return false;
+                        }
+#endif // GGML_USE_MUSA
+                        return true;
+                    default:
+                        return false;
+                }
+            } break;
+        case GGML_OP_OUT_PROD:
+            return op->type == GGML_TYPE_F32 && op->src[0]->type == GGML_TYPE_F32 && op->src[1]->type == GGML_TYPE_F32 && op->ne[2] == 1 && op->ne[3] == 1;
+        case GGML_OP_GET_ROWS:
+            {
+                switch (op->src[0]->type) {
+                    case GGML_TYPE_F16:
+                    case GGML_TYPE_F32:
+                    case GGML_TYPE_Q4_0:
+                    case GGML_TYPE_Q4_1:
+                    case GGML_TYPE_Q5_0:
+                    case GGML_TYPE_Q5_1:
+                    case GGML_TYPE_Q8_0:
+                        return true;
+                    default:
+                        return false;
+                }
+            } break;
+        case GGML_OP_CPY:
+            {
+                ggml_type src0_type = op->src[0]->type;
+                ggml_type src1_type = op->src[1]->type;
+                if (src0_type == GGML_TYPE_F32 && src1_type == GGML_TYPE_F32) {
+                    return true;
+                }
+                if (src0_type == GGML_TYPE_F32 && src1_type == GGML_TYPE_F16) {
+                    return true;
+                }
+                if (src0_type == GGML_TYPE_F32 && src1_type == GGML_TYPE_Q8_0) {
+                    return true;
+                }
+                if (src0_type == GGML_TYPE_Q8_0 && src1_type == GGML_TYPE_F32) {
+                    return true;
+                }
+                if (src0_type == GGML_TYPE_F32 && src1_type == GGML_TYPE_Q4_0) {
+                    return true;
+                }
+                if (src0_type == GGML_TYPE_F32 && src1_type == GGML_TYPE_Q4_1) {
+                    return true;
+                }
+                if (src0_type == GGML_TYPE_F32 && src1_type == GGML_TYPE_Q5_0) {
+                    return true;
+                }
+                if (src0_type == GGML_TYPE_F32 && src1_type == GGML_TYPE_Q5_1) {
+                    return true;
+                }
+                if (src0_type == GGML_TYPE_F32 && src1_type == GGML_TYPE_IQ4_NL) {
+                    return true;
+                }
+                if (src0_type == GGML_TYPE_F16 && src1_type == GGML_TYPE_F16) {
+                    return true;
+                }
+                if (src0_type == GGML_TYPE_F16 && src1_type == GGML_TYPE_F32) {
+                    return true;
+                }
+                if (src0_type == src1_type && ggml_is_contiguous(op->src[0]) && ggml_is_contiguous(op->src[1])) {
+                    return true;
+                }
+                return false;
+            } break;
+        case GGML_OP_DUP:
+            {
+                ggml_type src0_type = op->src[0]->type;
+                return src0_type != GGML_TYPE_I32 && src0_type != GGML_TYPE_I16;
+            } break;
+        case GGML_OP_ARGMAX:
+        case GGML_OP_COUNT_EQUAL:
+            {
+                return true;
+            } break;
+        case GGML_OP_REPEAT:
+            {
+                ggml_type src0_type = op->src[0]->type;
+                return src0_type != GGML_TYPE_I32 && src0_type != GGML_TYPE_I16;
+            } break;
+        case GGML_OP_REPEAT_BACK:
+                return op->type == GGML_TYPE_F32 && op->src[0]->ne[3] == 1;
+        case GGML_OP_CONCAT:
+            {
+                ggml_type src0_type = op->src[0]->type;
+                return src0_type != GGML_TYPE_I32 && src0_type != GGML_TYPE_I16;
+            } break;
+        case GGML_OP_CONV_TRANSPOSE_1D:
+            {
+                ggml_type src0_type = op->src[0]->type;
+                ggml_type src1_type = op->src[1]->type;
+                if (src0_type == GGML_TYPE_F32 && src1_type == GGML_TYPE_F32) {
+                    return true;
+                }
+                return false;
+            } break;
+        case GGML_OP_NONE:
+        case GGML_OP_RESHAPE:
+        case GGML_OP_VIEW:
+        case GGML_OP_PERMUTE:
+        case GGML_OP_TRANSPOSE:
+        case GGML_OP_NORM:
+        case GGML_OP_ADD:
+        case GGML_OP_ADD1:
+        case GGML_OP_SUB:
+        case GGML_OP_MUL:
+        case GGML_OP_DIV:
+        case GGML_OP_RMS_NORM:
+        case GGML_OP_SCALE:
+        case GGML_OP_SQR:
+        case GGML_OP_SQRT:
+        case GGML_OP_SIN:
+        case GGML_OP_COS:
+        case GGML_OP_CLAMP:
+            return true;
+        case GGML_OP_CONT:
+            return op->src[0]->type != GGML_TYPE_BF16;
+        case GGML_OP_DIAG_MASK_INF:
+        case GGML_OP_SOFT_MAX:
+            return true;
+        case GGML_OP_ROPE:
+            return ggml_is_contiguous(op->src[0]);
+        case GGML_OP_IM2COL:
+            return op->src[0]->type == GGML_TYPE_F16;
+        case GGML_OP_POOL_2D:
+        case GGML_OP_SUM:
+        case GGML_OP_SUM_ROWS:
+        case GGML_OP_ARGSORT:
+        case GGML_OP_ACC:
+        case GGML_OP_GROUP_NORM:
+        case GGML_OP_UPSCALE:
+        case GGML_OP_PAD:
+        case GGML_OP_ARANGE:
+        case GGML_OP_TIMESTEP_EMBEDDING:
+        case GGML_OP_LEAKY_RELU:
+        case GGML_OP_RWKV_WKV:
+            return true;
+        case GGML_OP_FLASH_ATTN_EXT: {
+#ifndef FLASH_ATTN_AVAILABLE
+            return false;
+#endif
+            if (op->src[0]->ne[0] ==  64 && op->src[1]->type == GGML_TYPE_F16) {
+                return true;
+            }
+            if (op->src[0]->ne[0] == 128) {
+                return true;
+            }
+            if (op->src[0]->ne[0] == 256 && op->src[1]->type == GGML_TYPE_F16 && op->src[2]->type == GGML_TYPE_F16) {
+                return true;
+            }
+            const int cc = ggml_cuda_info().devices[dev_ctx->device].cc;
+            return cc >= CC_VOLTA && cc < CC_OFFSET_AMD && op->src[1]->type == GGML_TYPE_F16 && op->src[2]->type == GGML_TYPE_F16;
+        }
+        case GGML_OP_CROSS_ENTROPY_LOSS:
+        case GGML_OP_CROSS_ENTROPY_LOSS_BACK:
+        case GGML_OP_OPT_STEP_ADAMW:
+            return true;
+        default:
+            return false;
+    }
+}
+
+static bool ggml_backend_cuda_device_supports_buft(ggml_backend_dev_t dev, ggml_backend_buffer_type_t buft) {
+    if (ggml_backend_buft_is_cuda_split(buft)) {
+        return true;
+    }
+
+    if (ggml_backend_buft_is_cuda(buft)) {
+        ggml_backend_cuda_device_context * dev_ctx = (ggml_backend_cuda_device_context *)dev->context;
+        ggml_backend_cuda_buffer_type_context * buft_ctx = (ggml_backend_cuda_buffer_type_context *)buft->context;
+        return buft_ctx->device == dev_ctx->device;
+    }
+
+    return false;
+}
+
+static bool ggml_backend_cuda_device_offload_op(ggml_backend_dev_t dev, const ggml_tensor * op) {
+    const int min_batch_size = 32;
+
+    return (op->ne[1] >= min_batch_size && op->op != GGML_OP_GET_ROWS) ||
+           (op->ne[2] >= min_batch_size && op->op == GGML_OP_MUL_MAT_ID);
+
+    GGML_UNUSED(dev);
+}
+
+static ggml_backend_event_t ggml_backend_cuda_device_event_new(ggml_backend_dev_t dev) {
+#ifdef GGML_CUDA_NO_PEER_COPY
+    return nullptr;
+#else
+    ggml_backend_cuda_device_context * dev_ctx = (ggml_backend_cuda_device_context *)dev->context;
+
+    ggml_cuda_set_device(dev_ctx->device);
+
+    cudaEvent_t event;
+    CUDA_CHECK(cudaEventCreateWithFlags(&event, cudaEventDisableTiming));
+
+    return new ggml_backend_event {
+        /* .device  = */ dev,
+        /* .context = */ event,
+    };
+#endif
+}
+
+static void ggml_backend_cuda_device_event_free(ggml_backend_dev_t dev, ggml_backend_event_t event) {
+    GGML_UNUSED(dev);
+
+    CUDA_CHECK(cudaEventDestroy((cudaEvent_t)event->context));
+    delete event;
+}
+
+static void ggml_backend_cuda_device_event_synchronize(ggml_backend_dev_t dev, ggml_backend_event_t event) {
+    GGML_UNUSED(dev);
+    CUDA_CHECK(cudaEventSynchronize((cudaEvent_t)event->context));
+}
+
+static const ggml_backend_device_i ggml_backend_cuda_device_interface = {
+    /* .get_name                = */ ggml_backend_cuda_device_get_name,
+    /* .get_description         = */ ggml_backend_cuda_device_get_description,
+    /* .get_memory              = */ ggml_backend_cuda_device_get_memory,
+    /* .get_type                = */ ggml_backend_cuda_device_get_type,
+    /* .get_props               = */ ggml_backend_cuda_device_get_props,
+    /* .init_backend            = */ ggml_backend_cuda_device_init,
+    /* .get_buffer_type         = */ ggml_backend_cuda_device_get_buffer_type,
+    /* .get_host_buffer_type    = */ ggml_backend_cuda_device_get_host_buffer_type,
+    /* .buffer_from_host_ptr    = */ ggml_backend_cuda_device_buffer_from_host_ptr,
+    /* .supports_op             = */ ggml_backend_cuda_device_supports_op,
+    /* .supports_buft           = */ ggml_backend_cuda_device_supports_buft,
+    /* .offload_op              = */ ggml_backend_cuda_device_offload_op,
+    /* .event_new               = */ ggml_backend_cuda_device_event_new,
+    /* .event_free              = */ ggml_backend_cuda_device_event_free,
+    /* .event_synchronize       = */ ggml_backend_cuda_device_event_synchronize,
+};
+
+// backend reg
+
+struct ggml_backend_cuda_reg_context {
+    std::vector<ggml_backend_dev_t> devices;
+};
+
+static const char * ggml_backend_cuda_reg_get_name(ggml_backend_reg_t reg) {
+    GGML_UNUSED(reg);
+    return GGML_CUDA_NAME;
+}
+
+static size_t ggml_backend_cuda_reg_get_device_count(ggml_backend_reg_t reg) {
+    ggml_backend_cuda_reg_context * ctx = (ggml_backend_cuda_reg_context *)reg->context;
+    return ctx->devices.size();
+}
+
+static ggml_backend_dev_t ggml_backend_cuda_reg_get_device(ggml_backend_reg_t reg, size_t index) {
+    ggml_backend_cuda_reg_context * ctx = (ggml_backend_cuda_reg_context *)reg->context;
+    GGML_ASSERT(index < ctx->devices.size());
+    return ctx->devices[index];
+}
+
+static void * ggml_backend_cuda_reg_get_proc_address(ggml_backend_reg_t reg, const char * name) {
+    GGML_UNUSED(reg);
+    if (strcmp(name, "ggml_backend_split_buffer_type") == 0) {
+        return (void *)ggml_backend_cuda_split_buffer_type;
+    }
+    if (strcmp(name, "ggml_backend_register_host_buffer") == 0) {
+        return (void *)ggml_backend_cuda_register_host_buffer;
+    }
+    if (strcmp(name, "ggml_backend_unregister_host_buffer") == 0) {
+        return (void *)ggml_backend_cuda_unregister_host_buffer;
+    }
+    return nullptr;
+}
+
+static const ggml_backend_reg_i ggml_backend_cuda_reg_interface = {
+    /* .get_name          = */ ggml_backend_cuda_reg_get_name,
+    /* .get_device_count  = */ ggml_backend_cuda_reg_get_device_count,
+    /* .get_device_get    = */ ggml_backend_cuda_reg_get_device,
+    /* .get_proc_address  = */ ggml_backend_cuda_reg_get_proc_address,
+};
+
+// backend registry
+ggml_backend_reg_t ggml_backend_cuda_reg() {
+    static ggml_backend_reg reg;
+    static bool initialized = false;
+
+    {
+        static std::mutex mutex;
+        std::lock_guard<std::mutex> lock(mutex);
+        if (!initialized) {
+            ggml_backend_cuda_reg_context * ctx = new ggml_backend_cuda_reg_context;
+
+            for (int i = 0; i < ggml_cuda_info().device_count; i++) {
+                ggml_backend_cuda_device_context * dev_ctx = new ggml_backend_cuda_device_context;
+                dev_ctx->device = i;
+                dev_ctx->name = GGML_CUDA_NAME + std::to_string(i);
+
+                ggml_cuda_set_device(i);
+                cudaDeviceProp prop;
+                CUDA_CHECK(cudaGetDeviceProperties(&prop, i));
+                dev_ctx->description = prop.name;
+
+                ggml_backend_dev_t dev = new ggml_backend_device {
+                    /* .interface = */ ggml_backend_cuda_device_interface,
+                    /* .reg       = */ &reg,
+                    /* .context   = */ dev_ctx
+                };
+                ctx->devices.push_back(dev);
+            }
+
+            reg = ggml_backend_reg {
+                /* .interface = */ ggml_backend_cuda_reg_interface,
+                /* .context   = */ ctx
+            };
+        }
+
+        initialized = true;
+    }
+
+    return &reg;
+}
+
+ggml_backend_t ggml_backend_cuda_init(int device) {
+    if (device < 0 || device >= ggml_backend_cuda_get_device_count()) {
+        GGML_LOG_ERROR("%s: invalid device %d\n", __func__, device);
+        return nullptr;
+    }
+
+    ggml_backend_cuda_context * ctx = new ggml_backend_cuda_context(device);
+    if (ctx == nullptr) {
+        GGML_LOG_ERROR("%s: failed to allocate context\n", __func__);
+        return nullptr;
+    }
+
+    ggml_backend_t cuda_backend = new ggml_backend {
+        /* .guid      = */ ggml_backend_cuda_guid(),
+        /* .interface = */ ggml_backend_cuda_interface,
+        /* .device    = */ ggml_backend_reg_dev_get(ggml_backend_cuda_reg(), device),
+        /* .context   = */ ctx,
+    };
+
+    return cuda_backend;
+}
diff --git a/src/whisper_cpp/ggml/src/ggml-cuda/acc.cu b/src/whisper_cpp/ggml/src/ggml-cuda/acc.cu
new file mode 100644
index 00000000..96bfe1c9
--- /dev/null
+++ b/src/whisper_cpp/ggml/src/ggml-cuda/acc.cu
@@ -0,0 +1,47 @@
+#include "acc.cuh"
+
+static __global__ void acc_f32(const float * x, const float * y, float * dst, const int ne,
+    const int ne10, const int ne11, const int ne12,
+    const int nb1, const int nb2, int offset) {
+    const int i = blockDim.x * blockIdx.x + threadIdx.x;
+    if (i >= ne) {
+        return;
+    }
+    int src1_idx = i - offset;
+    int oz = src1_idx / nb2;
+    int oy = (src1_idx - (oz * nb2)) / nb1;
+    int ox = src1_idx % nb1;
+    if (src1_idx >= 0 && ox < ne10 && oy < ne11 && oz < ne12) {
+        dst[i] = x[i] + y[ox + oy * ne10 + oz * ne10 * ne11];
+    } else {
+        dst[i] = x[i];
+    }
+}
+
+static void acc_f32_cuda(const float * x, const float * y, float * dst, const int n_elements,
+    const int ne10, const int ne11, const int ne12,
+    const int nb1, const int nb2, const int offset, cudaStream_t stream) {
+    int num_blocks = (n_elements + CUDA_ACC_BLOCK_SIZE - 1) / CUDA_ACC_BLOCK_SIZE;
+    acc_f32<<<num_blocks, CUDA_ACC_BLOCK_SIZE, 0, stream>>>(x, y, dst, n_elements, ne10, ne11, ne12, nb1, nb2, offset);
+}
+
+void ggml_cuda_op_acc(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
+    const ggml_tensor * src0 = dst->src[0];
+    const ggml_tensor * src1 = dst->src[1];
+    const float * src0_d = (const float *)src0->data;
+    const float * src1_d = (const float *)src1->data;
+    float * dst_d = (float *)dst->data;
+    cudaStream_t stream = ctx.stream();
+
+    GGML_ASSERT(src0->type == GGML_TYPE_F32);
+    GGML_ASSERT(src1->type == GGML_TYPE_F32);
+    GGML_ASSERT( dst->type == GGML_TYPE_F32);
+    GGML_ASSERT(dst->ne[3] == 1); // just 3D tensors supported
+
+    int nb1 = dst->op_params[0] / 4; // 4 bytes of float32
+    int nb2 = dst->op_params[1] / 4; // 4 bytes of float32
+    // int nb3 = dst->op_params[2] / 4; // 4 bytes of float32 - unused
+    int offset = dst->op_params[3] / 4; // offset in bytes
+
+    acc_f32_cuda(src0_d, src1_d, dst_d, ggml_nelements(dst), src1->ne[0], src1->ne[1], src1->ne[2], nb1, nb2, offset, stream);
+}
diff --git a/src/whisper_cpp/ggml/src/ggml-cuda/acc.cuh b/src/whisper_cpp/ggml/src/ggml-cuda/acc.cuh
new file mode 100644
index 00000000..1168ea1b
--- /dev/null
+++ b/src/whisper_cpp/ggml/src/ggml-cuda/acc.cuh
@@ -0,0 +1,5 @@
+#include "common.cuh"
+
+#define CUDA_ACC_BLOCK_SIZE 256
+
+void ggml_cuda_op_acc(ggml_backend_cuda_context & ctx, ggml_tensor * dst);
diff --git a/src/whisper_cpp/ggml/src/ggml-cuda/arange.cu b/src/whisper_cpp/ggml/src/ggml-cuda/arange.cu
new file mode 100644
index 00000000..b5e495a2
--- /dev/null
+++ b/src/whisper_cpp/ggml/src/ggml-cuda/arange.cu
@@ -0,0 +1,34 @@
+#include "arange.cuh"
+
+static __global__ void arange_f32(float * dst, const int ne0, const float start, const float step) {
+    // blockIDx.x: idx of ne0 / BLOCK_SIZE
+    int nidx = threadIdx.x + blockIdx.x * blockDim.x;
+    if (nidx >= ne0) {
+        return;
+    }
+    dst[nidx] = start + step * nidx;
+}
+
+static void arange_f32_cuda(float * dst, const int ne0, const float start, const float step, cudaStream_t stream) {
+    int num_blocks = (ne0 + CUDA_ARANGE_BLOCK_SIZE - 1) / CUDA_ARANGE_BLOCK_SIZE;
+    arange_f32<<<num_blocks, CUDA_ARANGE_BLOCK_SIZE, 0, stream>>>(dst, ne0, start,  step);
+}
+
+void ggml_cuda_op_arange(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
+    float * dst_d = (float *)dst->data;
+    cudaStream_t stream = ctx.stream();
+
+    GGML_ASSERT(dst->type == GGML_TYPE_F32);
+
+    float start;
+    float stop;
+    float step;
+    memcpy(&start, (float *)dst->op_params + 0, sizeof(float));
+    memcpy(&stop,  (float *)dst->op_params + 1, sizeof(float));
+    memcpy(&step,  (float *)dst->op_params + 2, sizeof(float));
+
+    int64_t steps = (int64_t)ceil((stop - start) / step);
+    GGML_ASSERT(ggml_nelements(dst) == steps);
+
+    arange_f32_cuda(dst_d, dst->ne[0], start, step, stream);
+}
diff --git a/src/whisper_cpp/ggml/src/ggml-cuda/arange.cuh b/src/whisper_cpp/ggml/src/ggml-cuda/arange.cuh
new file mode 100644
index 00000000..41e74fdf
--- /dev/null
+++ b/src/whisper_cpp/ggml/src/ggml-cuda/arange.cuh
@@ -0,0 +1,5 @@
+#include "common.cuh"
+
+#define CUDA_ARANGE_BLOCK_SIZE 256
+
+void ggml_cuda_op_arange(ggml_backend_cuda_context & ctx, ggml_tensor * dst);
diff --git a/src/whisper_cpp/ggml/src/ggml-cuda/argmax.cu b/src/whisper_cpp/ggml/src/ggml-cuda/argmax.cu
new file mode 100644
index 00000000..aab04eca
--- /dev/null
+++ b/src/whisper_cpp/ggml/src/ggml-cuda/argmax.cu
@@ -0,0 +1,79 @@
+#include "common.cuh"
+#include "argmax.cuh"
+#include "sum.cuh"
+
+#include <cstdint>
+
+static __global__ void argmax_f32(
+    const float * x, int32_t * dst, const int64_t ncols, const int64_t nrows) {
+
+    int argmax_thread = 0;
+    const int64_t row0 = (int64_t)blockIdx.x*WARP_SIZE;
+
+#pragma unroll
+    for (int64_t row1 = 0; row1 < WARP_SIZE; ++row1) {
+        const int64_t row = row0 + row1;
+
+        if (row >= nrows) {
+            break;
+        }
+
+        float maxval = -FLT_MAX;
+        int   argmax = -1;
+
+        for (int32_t col = threadIdx.x; col < ncols; col += WARP_SIZE) {
+            const float val        = x[row*ncols + col];
+            const int   bigger     = val > maxval;
+            const int   not_bigger = bigger ^ 0x00000001;
+
+            maxval = maxval*not_bigger + val*bigger;
+            argmax = argmax*not_bigger + col*bigger;
+        }
+
+#pragma unroll
+        for (int mask = 16; mask > 0; mask >>= 1) {
+            const float val        = __shfl_xor_sync(0xFFFFFFFF, maxval, mask, WARP_SIZE);
+            const int   col        = __shfl_xor_sync(0xFFFFFFFF, argmax, mask, WARP_SIZE);
+            const int   bigger     = val > maxval;
+            const int   not_bigger = bigger ^ 0x00000001;
+
+            maxval = maxval*not_bigger + val*bigger;
+            argmax = argmax*not_bigger + col*bigger;
+        }
+
+        const int store = row1 == threadIdx.x;
+        argmax_thread += store*argmax;
+    }
+
+    const int row = row0 + threadIdx.x;
+
+    if (row >= nrows) {
+        return;
+    }
+
+    dst[row] = argmax_thread;
+}
+
+void ggml_cuda_argmax(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
+    const ggml_tensor * src0 = dst->src[0];
+
+    GGML_ASSERT(src0->type == GGML_TYPE_F32);
+    GGML_ASSERT( dst->type == GGML_TYPE_I32);
+
+    GGML_ASSERT(ggml_is_contiguous(src0));
+
+    const int64_t ne00  = src0->ne[0];
+    const int64_t nrows = ggml_nrows(src0);
+
+    const float * src0_d = (const float *) src0->data;
+    int32_t     * dst_d  = (int32_t     *) dst->data;
+
+    cudaStream_t stream = ctx.stream();
+
+    const int64_t num_blocks = (nrows + WARP_SIZE - 1) / WARP_SIZE;
+
+    const dim3 blocks_dim(WARP_SIZE, 1, 1);
+    const dim3 blocks_num(num_blocks, 1, 1);
+
+    argmax_f32<<<blocks_num, blocks_dim, 0, stream>>>(src0_d, dst_d, ne00, nrows);
+}
diff --git a/src/whisper_cpp/ggml/src/ggml-cuda/argmax.cuh b/src/whisper_cpp/ggml/src/ggml-cuda/argmax.cuh
new file mode 100644
index 00000000..5b7223ad
--- /dev/null
+++ b/src/whisper_cpp/ggml/src/ggml-cuda/argmax.cuh
@@ -0,0 +1,3 @@
+#include "common.cuh"
+
+void ggml_cuda_argmax(ggml_backend_cuda_context & ctx, ggml_tensor * dst);
diff --git a/src/whisper_cpp/ggml/src/ggml-cuda/argsort.cu b/src/whisper_cpp/ggml/src/ggml-cuda/argsort.cu
new file mode 100644
index 00000000..607ded85
--- /dev/null
+++ b/src/whisper_cpp/ggml/src/ggml-cuda/argsort.cu
@@ -0,0 +1,104 @@
+#include "argsort.cuh"
+
+template<typename T>
+static inline __device__ void ggml_cuda_swap(T & a, T & b) {
+    T tmp = a;
+    a = b;
+    b = tmp;
+}
+
+template<ggml_sort_order order>
+static __global__ void k_argsort_f32_i32(const float * x, int * dst, const int ncols, int ncols_pad) {
+    // bitonic sort
+    int col = threadIdx.x;
+    int row = blockIdx.y;
+
+    if (col >= ncols_pad) {
+        return;
+    }
+
+    const float * x_row = x + row * ncols;
+    extern __shared__ int dst_row[];
+
+    // initialize indices
+    dst_row[col] = col;
+
+    __syncthreads();
+
+    for (int k = 2; k <= ncols_pad; k *= 2) {
+        for (int j = k / 2; j > 0; j /= 2) {
+            int ixj = col ^ j;
+            if (ixj > col) {
+                if ((col & k) == 0) {
+                    if (dst_row[col] >= ncols ||
+                        (dst_row[ixj] < ncols && (order == GGML_SORT_ORDER_ASC ?
+                            x_row[dst_row[col]] > x_row[dst_row[ixj]] :
+                            x_row[dst_row[col]] < x_row[dst_row[ixj]]))
+                    ) {
+                        ggml_cuda_swap(dst_row[col], dst_row[ixj]);
+                    }
+                } else {
+                    if (dst_row[ixj] >= ncols ||
+                        (dst_row[col] < ncols && (order == GGML_SORT_ORDER_ASC ?
+                            x_row[dst_row[col]] < x_row[dst_row[ixj]] :
+                            x_row[dst_row[col]] > x_row[dst_row[ixj]]))
+                    ) {
+                        ggml_cuda_swap(dst_row[col], dst_row[ixj]);
+                    }
+                }
+            }
+            __syncthreads();
+        }
+    }
+
+    // copy the result to dst without the padding
+    if (col < ncols) {
+        dst[row * ncols + col] = dst_row[col];
+    }
+}
+
+static int next_power_of_2(int x) {
+    int n = 1;
+    while (n < x) {
+        n *= 2;
+    }
+    return n;
+}
+
+static void argsort_f32_i32_cuda(const float * x, int * dst, const int ncols, const int nrows, ggml_sort_order order, cudaStream_t stream) {
+    // bitonic sort requires ncols to be power of 2
+    const int ncols_pad = next_power_of_2(ncols);
+
+    const dim3 block_dims(ncols_pad, 1, 1);
+    const dim3 block_nums(1, nrows, 1);
+    const size_t shared_mem = ncols_pad * sizeof(int);
+
+    // FIXME: this limit could be raised by ~2-4x on Ampere or newer
+    GGML_ASSERT(shared_mem <= ggml_cuda_info().devices[ggml_cuda_get_device()].smpb);
+
+    if (order == GGML_SORT_ORDER_ASC) {
+        k_argsort_f32_i32<GGML_SORT_ORDER_ASC><<<block_nums, block_dims, shared_mem, stream>>>(x, dst, ncols, ncols_pad);
+    } else if (order == GGML_SORT_ORDER_DESC) {
+        k_argsort_f32_i32<GGML_SORT_ORDER_DESC><<<block_nums, block_dims, shared_mem, stream>>>(x, dst, ncols, ncols_pad);
+    } else {
+        GGML_ABORT("fatal error");
+    }
+}
+
+void ggml_cuda_op_argsort(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
+    const ggml_tensor * src0 = dst->src[0];
+    const float * src0_d = (const float *)src0->data;
+    float * dst_d = (float *)dst->data;
+    cudaStream_t stream = ctx.stream();
+
+    GGML_ASSERT(src0->type == GGML_TYPE_F32);
+    GGML_ASSERT( dst->type == GGML_TYPE_I32);
+    GGML_ASSERT(ggml_is_contiguous(src0));
+
+    const int64_t ncols = src0->ne[0];
+    const int64_t nrows = ggml_nrows(src0);
+
+    enum ggml_sort_order order = (enum ggml_sort_order) dst->op_params[0];
+
+    argsort_f32_i32_cuda(src0_d, (int *)dst_d, ncols, nrows, order, stream);
+}
diff --git a/src/whisper_cpp/ggml/src/ggml-cuda/argsort.cuh b/src/whisper_cpp/ggml/src/ggml-cuda/argsort.cuh
new file mode 100644
index 00000000..68a00154
--- /dev/null
+++ b/src/whisper_cpp/ggml/src/ggml-cuda/argsort.cuh
@@ -0,0 +1,3 @@
+#include "common.cuh"
+
+void ggml_cuda_op_argsort(ggml_backend_cuda_context & ctx, ggml_tensor * dst);
diff --git a/src/whisper_cpp/ggml/src/ggml-cuda/binbcast.cu b/src/whisper_cpp/ggml/src/ggml-cuda/binbcast.cu
new file mode 100644
index 00000000..c7b6be4e
--- /dev/null
+++ b/src/whisper_cpp/ggml/src/ggml-cuda/binbcast.cu
@@ -0,0 +1,355 @@
+#include "binbcast.cuh"
+#include <cstdint>
+
+static __device__ __forceinline__ float op_repeat(const float a, const float b) {
+    return b;
+    GGML_UNUSED(a);
+}
+
+static __device__ __forceinline__ float op_add(const float a, const float b) {
+    return a + b;
+}
+
+static __device__ __forceinline__ float op_sub(const float a, const float b) {
+    return a - b;
+}
+
+static __device__ __forceinline__ float op_mul(const float a, const float b) {
+    return a * b;
+}
+
+static __device__ __forceinline__ float op_div(const float a, const float b) {
+    return a / b;
+}
+
+template<float (*bin_op)(const float, const float), typename src0_t, typename src1_t, typename dst_t>
+static __global__ void k_bin_bcast(const src0_t * src0, const src1_t * src1, dst_t * dst,
+        int ne0, int ne1, int ne2, int ne3,
+        int ne10, int ne11, int ne12, int ne13,
+        /*int s0, */ int s1,  int s2,  int s3,
+        /*int s00,*/ int s01, int s02, int s03,
+        /*int s10,*/ int s11, int s12, int s13) {
+    const int i0s = blockDim.x*blockIdx.x + threadIdx.x;
+    const int i1 = (blockDim.y*blockIdx.y + threadIdx.y);
+    const int i2 = (blockDim.z*blockIdx.z + threadIdx.z) / ne3;
+    const int i3 = (blockDim.z*blockIdx.z + threadIdx.z) % ne3;
+
+    if (i0s >= ne0 || i1 >= ne1 || i2 >= ne2 || i3 >= ne3) {
+        return;
+    }
+
+    const int i11 = i1 % ne11;
+    const int i12 = i2 % ne12;
+    const int i13 = i3 % ne13;
+
+    const size_t i_src0 =  i3*s03 +  i2*s02 +  i1*s01;
+    const size_t i_src1 = i13*s13 + i12*s12 + i11*s11;
+    const size_t i_dst  =  i3*s3  +  i2*s2  +  i1*s1;
+
+    const src0_t * src0_row = src0 + i_src0;
+    const src1_t * src1_row = src1 + i_src1;
+    dst_t * dst_row = dst + i_dst;
+
+    for (int i0 = i0s; i0 < ne0; i0 += blockDim.x*gridDim.x) {
+        const int i10 = i0 % ne10;
+        dst_row[i0] = (dst_t)bin_op(src0 ? (float)src0_row[i0] : 0.0f, (float)src1_row[i10]);
+    }
+}
+
+template<float (*bin_op)(const float, const float), typename src0_t, typename src1_t, typename dst_t>
+static __global__ void k_bin_bcast_unravel(const src0_t * src0, const src1_t * src1, dst_t * dst,
+        int ne0, int ne1, int ne2, int ne3,
+        int ne10, int ne11, int ne12, int ne13,
+        /*int s0, */ int s1,  int s2,  int s3,
+        /*int s00,*/ int s01, int s02, int s03,
+        /*int s10,*/ int s11, int s12, int s13) {
+
+    const int i = blockDim.x*blockIdx.x + threadIdx.x;
+
+    const int i3 = i/(ne2*ne1*ne0);
+    const int i2 = (i/(ne1*ne0)) % ne2;
+    const int i1 = (i/ne0) % ne1;
+    const int i0 = i % ne0;
+
+    if (i0 >= ne0 || i1 >= ne1 || i2 >= ne2 || i3 >= ne3) {
+        return;
+    }
+
+    const int i11 = i1 % ne11;
+    const int i12 = i2 % ne12;
+    const int i13 = i3 % ne13;
+
+    const size_t i_src0 =  i3*s03 +  i2*s02 +  i1*s01;
+    const size_t i_src1 = i13*s13 + i12*s12 + i11*s11;
+    const size_t i_dst  =  i3*s3  +  i2*s2  +  i1*s1;
+
+    const src0_t * src0_row = src0 + i_src0;
+    const src1_t * src1_row = src1 + i_src1;
+    dst_t * dst_row = dst + i_dst;
+
+    const int i10 = i0 % ne10;
+    dst_row[i0] = (dst_t)bin_op(src0 ? (float)src0_row[i0] : 0.0f, (float)src1_row[i10]);
+}
+
+template <typename T>
+static __global__ void k_repeat_back(
+    const T * __restrict__ src, T * __restrict__ dst, const int64_t ne00, const int64_t ne01, const int64_t ne02,
+    const int64_t ne0, const int64_t ne1, const int64_t ne2) {
+
+    const int64_t tid0 = (int64_t) blockIdx.x*blockDim.x + threadIdx.x;
+    const int64_t tid1 = (int64_t) blockIdx.y*blockDim.y + threadIdx.y;
+    const int64_t tid2 = (int64_t) blockIdx.z*blockDim.z + threadIdx.z;
+
+    if (tid0 >= ne0) {
+        return;
+    }
+
+    T sum = 0;
+    for (int64_t i2 = tid2; i2 < ne02; i2 += ne2) {
+        for (int64_t i1 = tid1; i1 < ne01; i1 += ne1) {
+            for (int64_t i0 = tid0; i0 < ne00; i0 += ne0) {
+                sum += src[i2*ne01*ne00 + i1*ne00 + i0];
+            }
+        }
+    }
+    dst[tid2*ne1*ne0 + tid1*ne0 + tid0] = sum;
+}
+
+template<float (*bin_op)(const float, const float)>
+struct bin_bcast_cuda {
+    template<typename src0_t, typename src1_t, typename dst_t>
+    void operator()(const struct ggml_tensor * src0, const struct ggml_tensor * src1, struct ggml_tensor * dst,
+            const src0_t * src0_dd, const src1_t * src1_dd, dst_t * dst_dd,
+            cudaStream_t stream) {
+
+        GGML_TENSOR_BINARY_OP_LOCALS
+
+        int nr0 = ne10/ne0;
+        int nr1 = ne11/ne1;
+        int nr2 = ne12/ne2;
+        int nr3 = ne13/ne3;
+
+        int nr[4] = { nr0, nr1, nr2, nr3 };
+
+        // collapse dimensions until first broadcast dimension
+        int64_t cne[] = {ne0, ne1, ne2, ne3};
+        int64_t cne0[] = {ne00, ne01, ne02, ne03};
+        int64_t cne1[] = {ne10, ne11, ne12, ne13};
+
+        size_t cnb[] = {nb0, nb1, nb2, nb3};
+        size_t cnb0[] = {nb00, nb01, nb02, nb03};
+        size_t cnb1[] = {nb10, nb11, nb12, nb13};
+
+        auto collapse = [](int64_t cne[]) {
+            cne[0] *= cne[1];
+            cne[1] = cne[2];
+            cne[2] = cne[3];
+            cne[3] = 1;
+        };
+
+        auto collapse_nb = [](size_t cnb[], const int64_t cne[]) {
+            cnb[1] *= cne[1];
+            cnb[2] *= cne[2];
+            cnb[3] *= cne[3];
+        };
+
+        if (ggml_is_contiguous(src0) && ggml_is_contiguous(src1) && ggml_is_contiguous(dst)) {
+            for (int i = 0; i < 4; i++) {
+                if (nr[i] != 1) {
+                    break;
+                }
+                if (i > 0) {
+                    collapse_nb(cnb, cne);
+                    collapse_nb(cnb0, cne0);
+                    collapse_nb(cnb1, cne1);
+                    collapse(cne);
+                    collapse(cne0);
+                    collapse(cne1);
+                }
+            }
+        }
+
+        {
+            int64_t ne0 = cne[0];
+            int64_t ne1 = cne[1];
+            int64_t ne2 = cne[2];
+            int64_t ne3 = cne[3];
+
+            //int64_t ne00 = cne0[0]; GGML_UNUSED(ne00);
+            //int64_t ne01 = cne0[1]; GGML_UNUSED(ne01);
+            //int64_t ne02 = cne0[2]; GGML_UNUSED(ne02);
+            //int64_t ne03 = cne0[3]; GGML_UNUSED(ne03);
+
+            int64_t ne10 = cne1[0];
+            int64_t ne11 = cne1[1];
+            int64_t ne12 = cne1[2];
+            int64_t ne13 = cne1[3];
+
+            size_t nb0 = cnb[0];
+            size_t nb1 = cnb[1];
+            size_t nb2 = cnb[2];
+            size_t nb3 = cnb[3];
+
+            size_t nb00 = cnb0[0];
+            size_t nb01 = cnb0[1];
+            size_t nb02 = cnb0[2];
+            size_t nb03 = cnb0[3];
+
+            size_t nb10 = cnb1[0];
+            size_t nb11 = cnb1[1];
+            size_t nb12 = cnb1[2];
+            size_t nb13 = cnb1[3];
+
+            size_t s0 = nb0 / sizeof(dst_t);
+            size_t s1 = nb1 / sizeof(dst_t);
+            size_t s2 = nb2 / sizeof(dst_t);
+            size_t s3 = nb3 / sizeof(dst_t);
+
+            size_t s10 = nb10 / sizeof(src1_t);
+            size_t s11 = nb11 / sizeof(src1_t);
+            size_t s12 = nb12 / sizeof(src1_t);
+            size_t s13 = nb13 / sizeof(src1_t);
+
+            size_t s00 = nb00 / sizeof(src0_t);
+            size_t s01 = nb01 / sizeof(src0_t);
+            size_t s02 = nb02 / sizeof(src0_t);
+            size_t s03 = nb03 / sizeof(src0_t);
+
+            GGML_ASSERT(nb0 % sizeof(dst_t) == 0);
+            GGML_ASSERT(nb1 % sizeof(dst_t) == 0);
+            GGML_ASSERT(nb2 % sizeof(dst_t) == 0);
+            GGML_ASSERT(nb3 % sizeof(dst_t) == 0);
+
+            GGML_ASSERT(nb00 % sizeof(src0_t) == 0);
+            GGML_ASSERT(nb01 % sizeof(src0_t) == 0);
+            GGML_ASSERT(nb02 % sizeof(src0_t) == 0);
+            GGML_ASSERT(nb03 % sizeof(src0_t) == 0);
+
+            GGML_ASSERT(nb10 % sizeof(src1_t) == 0);
+            GGML_ASSERT(nb11 % sizeof(src1_t) == 0);
+            GGML_ASSERT(nb12 % sizeof(src1_t) == 0);
+            GGML_ASSERT(nb13 % sizeof(src1_t) == 0);
+
+            GGML_ASSERT(s0 == 1);
+            GGML_ASSERT(s00 == 1);
+            GGML_ASSERT(s10 == 1);
+
+            const int block_size = 128;
+
+            int64_t hne0 = std::max(ne0/2LL, 1LL);
+
+            dim3 block_dims;
+            block_dims.x = std::min<unsigned int>(hne0, block_size);
+            block_dims.y = std::min<unsigned int>(ne1, block_size / block_dims.x);
+            block_dims.z = std::min(std::min<unsigned int>(ne2*ne3, block_size / block_dims.x / block_dims.y), 64U);
+
+            dim3 block_nums(
+                (hne0 + block_dims.x - 1) / block_dims.x,
+                (ne1 + block_dims.y - 1) / block_dims.y,
+                (ne2*ne3 + block_dims.z - 1) / block_dims.z
+            );
+
+            if (block_nums.z > 65535) {
+                // this is the maximum number of blocks in z dimension, fallback to 1D grid kernel
+                int block_num = (ne0*ne1*ne2*ne3 + block_size - 1) / block_size;
+                k_bin_bcast_unravel<bin_op><<<block_num, block_size, 0, stream>>>(
+                    src0_dd, src1_dd, dst_dd,
+                    ne0, ne1, ne2, ne3,
+                    ne10, ne11, ne12, ne13,
+                    /* s0, */ s1, s2, s3,
+                    /* s00, */ s01, s02, s03,
+                    /* s10, */ s11, s12, s13);
+            } else {
+                k_bin_bcast<bin_op><<<block_nums, block_dims, 0, stream>>>(
+                    src0_dd, src1_dd, dst_dd,
+                    ne0, ne1, ne2, ne3,
+                    ne10, ne11, ne12, ne13,
+                    /* s0, */ s1, s2, s3,
+                    /* s00, */ s01, s02, s03,
+                    /* s10, */ s11, s12, s13);
+            }
+        }
+    }
+};
+
+template <typename T>
+static void repeat_back_cuda(
+    const T * src, T * dst, const int64_t ne00, const int64_t ne01, const int64_t ne02,
+    const int64_t ne0, const int64_t ne1, const int64_t ne2, cudaStream_t stream) {
+
+    const dim3 block_dims(WARP_SIZE, 1, 1);
+    const dim3 block_nums((ne0 + WARP_SIZE - 1) / WARP_SIZE, ne1, ne2);
+    k_repeat_back<T><<<block_nums, block_dims, 0, stream>>>(src, dst, ne00, ne01, ne02, ne0, ne1, ne2);
+}
+
+template<class op>
+static void ggml_cuda_op_bin_bcast(
+    const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst,
+    const void * src0_dd, const void * src1_dd, void * dst_dd, cudaStream_t stream) {
+
+    GGML_ASSERT(src1->type == GGML_TYPE_F32);
+
+    if (src0->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32) {
+        op()(src0, src1, dst, (const float *)src0_dd, (const float *)src1_dd, (float *)dst_dd, stream);
+    } else if (src0->type == GGML_TYPE_F16 && dst->type == GGML_TYPE_F16) {
+        op()(src0, src1, dst, (const half *) src0_dd, (const float *)src1_dd, (half *) dst_dd, stream);
+    } else if (src0->type == GGML_TYPE_F16 && dst->type == GGML_TYPE_F32) {
+        op()(src0, src1, dst, (const half *) src0_dd, (const float *)src1_dd, (float *)dst_dd, stream);
+    } else {
+        fprintf(stderr, "%s: unsupported types: dst: %s, src0: %s, src1: %s\n", __func__,
+            ggml_type_name(dst->type), ggml_type_name(src0->type), ggml_type_name(src1->type));
+        GGML_ABORT("fatal error");
+    }
+}
+
+void ggml_cuda_op_repeat(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
+    ggml_cuda_op_bin_bcast<bin_bcast_cuda<op_repeat>>(dst, dst->src[0], dst, nullptr, dst->src[0]->data, dst->data, ctx.stream());
+}
+
+void ggml_cuda_op_add(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
+    ggml_cuda_op_bin_bcast<bin_bcast_cuda<op_add>>(dst->src[0], dst->src[1], dst, dst->src[0]->data, dst->src[1]->data, dst->data, ctx.stream());
+}
+
+void ggml_cuda_op_sub(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
+    ggml_cuda_op_bin_bcast<bin_bcast_cuda<op_sub>>(dst->src[0], dst->src[1], dst, dst->src[0]->data, dst->src[1]->data, dst->data, ctx.stream());
+}
+
+void ggml_cuda_op_mul(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
+    ggml_cuda_op_bin_bcast<bin_bcast_cuda<op_mul>>(dst->src[0], dst->src[1], dst, dst->src[0]->data, dst->src[1]->data, dst->data, ctx.stream());
+}
+
+void ggml_cuda_op_div(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
+    ggml_cuda_op_bin_bcast<bin_bcast_cuda<op_div>>(dst->src[0], dst->src[1], dst, dst->src[0]->data, dst->src[1]->data, dst->data, ctx.stream());
+}
+
+void ggml_cuda_op_repeat_back(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
+    const ggml_tensor * src0 = dst->src[0];
+
+    GGML_ASSERT(src0->type == dst->type);
+    GGML_ASSERT(ggml_is_contiguous(src0));
+    GGML_ASSERT(ggml_is_contiguous(dst));
+    GGML_ASSERT(ggml_can_repeat(dst, src0));
+
+    cudaStream_t stream = ctx.stream();
+
+    const int64_t ne00 = src0->ne[0];
+    const int64_t ne01 = src0->ne[1];
+    const int64_t ne02 = src0->ne[2];
+    GGML_ASSERT(src0->ne[3] == 1);
+
+    const int64_t ne0 = dst->ne[0];
+    const int64_t ne1 = dst->ne[1];
+    const int64_t ne2 = dst->ne[2];
+    GGML_ASSERT(dst->ne[3] == 1);
+
+    switch (dst->type) {
+        case GGML_TYPE_F32: {
+            const float * src0_d = (const float *) src0->data;
+            float       * dst_d  = (float       *) dst->data;
+            repeat_back_cuda<float>(src0_d, dst_d, ne00, ne01, ne02, ne0, ne1, ne2, stream);
+        } break;
+        default: {
+            GGML_ASSERT(false);
+        } break;
+    }
+}
diff --git a/src/whisper_cpp/ggml/src/ggml-cuda/binbcast.cuh b/src/whisper_cpp/ggml/src/ggml-cuda/binbcast.cuh
new file mode 100644
index 00000000..3ac1c9b0
--- /dev/null
+++ b/src/whisper_cpp/ggml/src/ggml-cuda/binbcast.cuh
@@ -0,0 +1,9 @@
+#include "common.cuh"
+
+void ggml_cuda_op_repeat(ggml_backend_cuda_context & ctx, ggml_tensor * dst);
+void ggml_cuda_op_add(ggml_backend_cuda_context & ctx, ggml_tensor * dst);
+void ggml_cuda_op_sub(ggml_backend_cuda_context & ctx, ggml_tensor * dst);
+void ggml_cuda_op_mul(ggml_backend_cuda_context & ctx, ggml_tensor * dst);
+void ggml_cuda_op_div(ggml_backend_cuda_context & ctx, ggml_tensor * dst);
+
+void ggml_cuda_op_repeat_back(ggml_backend_cuda_context & ctx, ggml_tensor * dst);
diff --git a/src/whisper_cpp/ggml/src/ggml-cuda/clamp.cu b/src/whisper_cpp/ggml/src/ggml-cuda/clamp.cu
new file mode 100644
index 00000000..8009a3e3
--- /dev/null
+++ b/src/whisper_cpp/ggml/src/ggml-cuda/clamp.cu
@@ -0,0 +1,34 @@
+#include "clamp.cuh"
+
+static __global__ void clamp_f32(const float * x, float * dst, const float min, const float max, const int k) {
+    const int i = blockDim.x*blockIdx.x + threadIdx.x;
+
+    if (i >= k) {
+        return;
+    }
+
+    dst[i] = x[i] < min ? min : (x[i] > max ? max : x[i]);
+}
+
+static void clamp_f32_cuda(const float * x, float * dst, const float min, const float max, const int k, cudaStream_t stream) {
+    const int num_blocks = (k + CUDA_CLAMP_BLOCK_SIZE - 1) / CUDA_CLAMP_BLOCK_SIZE;
+    clamp_f32<<<num_blocks, CUDA_CLAMP_BLOCK_SIZE, 0, stream>>>(x, dst, min, max, k);
+}
+
+
+void ggml_cuda_op_clamp(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
+    const ggml_tensor * src0 = dst->src[0];
+    const float * src0_d = (const float *)src0->data;
+    float * dst_d = (float *)dst->data;
+    cudaStream_t stream = ctx.stream();
+
+    GGML_ASSERT(src0->type == GGML_TYPE_F32);
+    GGML_ASSERT( dst->type == GGML_TYPE_F32);
+
+    float min;
+    float max;
+    memcpy(&min, dst->op_params, sizeof(float));
+    memcpy(&max, (float *) dst->op_params + 1, sizeof(float));
+
+    clamp_f32_cuda(src0_d, dst_d, min, max, ggml_nelements(src0), stream);
+}
diff --git a/src/whisper_cpp/ggml/src/ggml-cuda/clamp.cuh b/src/whisper_cpp/ggml/src/ggml-cuda/clamp.cuh
new file mode 100644
index 00000000..7f9559dd
--- /dev/null
+++ b/src/whisper_cpp/ggml/src/ggml-cuda/clamp.cuh
@@ -0,0 +1,5 @@
+#include "common.cuh"
+
+#define CUDA_CLAMP_BLOCK_SIZE 256
+
+void ggml_cuda_op_clamp(ggml_backend_cuda_context & ctx, ggml_tensor * dst);
diff --git a/src/whisper_cpp/ggml/src/ggml-cuda/common.cuh b/src/whisper_cpp/ggml/src/ggml-cuda/common.cuh
new file mode 100644
index 00000000..dd203fcd
--- /dev/null
+++ b/src/whisper_cpp/ggml/src/ggml-cuda/common.cuh
@@ -0,0 +1,688 @@
+#pragma once
+
+#include "ggml.h"
+#include "ggml-cuda.h"
+
+#include <cstdint>
+#include <memory>
+
+#if defined(GGML_USE_HIPBLAS)
+#define GGML_COMMON_DECL_HIP
+#define GGML_COMMON_IMPL_HIP
+#else
+#define GGML_COMMON_DECL_CUDA
+#define GGML_COMMON_IMPL_CUDA
+#if defined(GGML_USE_MUSA)
+#define GGML_COMMON_DECL_MUSA
+#define GGML_COMMON_IMPL_MUSA
+#endif
+#endif
+#include "ggml-common.h"
+
+#include <cstdio>
+#include <array>
+#include <cassert>
+#include <cfloat>
+#include <string>
+#include <vector>
+
+#if defined(GGML_USE_HIPBLAS)
+#include "vendors/hip.h"
+#elif defined(GGML_USE_MUSA)
+#include "vendors/musa.h"
+#else
+#include "vendors/cuda.h"
+#endif // defined(GGML_USE_HIPBLAS)
+
+#define STRINGIZE_IMPL(...) #__VA_ARGS__
+#define STRINGIZE(...) STRINGIZE_IMPL(__VA_ARGS__)
+
+#define WARP_SIZE 32
+#define CUDART_HMAX   11070 // CUDA 11.7, min. ver. for which __hmax and __hmax2 are known to work (may be higher than needed)
+#define CUDART_HMASK  12000 // CUDA 12.0, min. ver. for half2 -> uint mask comparisons
+
+#define CC_PASCAL     600
+#define MIN_CC_DP4A   610 // minimum compute capability for __dp4a, an intrinsic for byte-wise dot products
+#define CC_VOLTA      700
+#define CC_TURING     750
+#define CC_AMPERE     800
+#define CC_OFFSET_AMD 1000000
+#define CC_RDNA1      (CC_OFFSET_AMD + 1010)
+#define CC_RDNA2      (CC_OFFSET_AMD + 1030)
+#define CC_RDNA3      (CC_OFFSET_AMD + 1100)
+#define CC_QY1        210
+#define CC_QY2        220
+
+#define MATRIX_ROW_PADDING 512 // last row of quant. matrices is a multiple of this to avoid out-of-bounds memory accesses
+
+#if defined(_MSC_VER)
+#pragma warning(disable: 4244 4267) // possible loss of data
+#endif
+
+#define GGML_CUDA_MAX_STREAMS 8
+
+[[noreturn]]
+void ggml_cuda_error(const char * stmt, const char * func, const char * file, int line, const char * msg);
+
+#define CUDA_CHECK_GEN(err, success, error_fn)                                      \
+     do {                                                                           \
+        auto err_ = (err);                                                          \
+        if (err_ != (success)) {                                                    \
+            ggml_cuda_error(#err, __func__, __FILE__, __LINE__, error_fn(err_));    \
+        }                                                                           \
+    } while (0)
+
+#define CUDA_CHECK(err) CUDA_CHECK_GEN(err, cudaSuccess, cudaGetErrorString)
+
+#if CUDART_VERSION >= 12000 || defined(GGML_USE_MUSA)
+    static const char * cublas_get_error_str(const cublasStatus_t err) {
+        return cublasGetStatusString(err);
+    }
+#else
+    static const char * cublas_get_error_str(const cublasStatus_t err) {
+        switch (err) {
+            case CUBLAS_STATUS_SUCCESS: return "CUBLAS_STATUS_SUCCESS";
+            case CUBLAS_STATUS_NOT_INITIALIZED: return "CUBLAS_STATUS_NOT_INITIALIZED";
+            case CUBLAS_STATUS_ALLOC_FAILED: return "CUBLAS_STATUS_ALLOC_FAILED";
+            case CUBLAS_STATUS_INVALID_VALUE: return "CUBLAS_STATUS_INVALID_VALUE";
+            case CUBLAS_STATUS_ARCH_MISMATCH: return "CUBLAS_STATUS_ARCH_MISMATCH";
+            case CUBLAS_STATUS_MAPPING_ERROR: return "CUBLAS_STATUS_MAPPING_ERROR";
+            case CUBLAS_STATUS_EXECUTION_FAILED: return "CUBLAS_STATUS_EXECUTION_FAILED";
+            case CUBLAS_STATUS_INTERNAL_ERROR: return "CUBLAS_STATUS_INTERNAL_ERROR";
+            case CUBLAS_STATUS_NOT_SUPPORTED: return "CUBLAS_STATUS_NOT_SUPPORTED";
+            default: return "unknown error";
+        }
+    }
+#endif // CUDART_VERSION >= 12000
+
+#define CUBLAS_CHECK(err) CUDA_CHECK_GEN(err, CUBLAS_STATUS_SUCCESS, cublas_get_error_str)
+
+#if !defined(GGML_USE_HIPBLAS)
+static const char * cu_get_error_str(CUresult err) {
+    const char * err_str;
+    cuGetErrorString(err, &err_str);
+    return err_str;
+}
+#define CU_CHECK(err) CUDA_CHECK_GEN(err, CUDA_SUCCESS, cu_get_error_str)
+#endif
+
+#if CUDART_VERSION >= 11100 || defined(GGML_USE_MUSA)
+#define GGML_CUDA_ASSUME(x) __builtin_assume(x)
+#else
+#define GGML_CUDA_ASSUME(x)
+#endif // CUDART_VERSION >= 11100
+
+#ifdef GGML_CUDA_F16
+typedef half dfloat; // dequantize float
+typedef half2 dfloat2;
+#else
+typedef float dfloat; // dequantize float
+typedef float2 dfloat2;
+#endif // GGML_CUDA_F16
+
+#if (defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)) || __CUDA_ARCH__ >= CC_PASCAL
+#define FP16_AVAILABLE
+#endif // (defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)) || __CUDA_ARCH__ >= CC_PASCAL
+
+#if defined(FP16_AVAILABLE) && __CUDA_ARCH__ != 610
+#define FAST_FP16_AVAILABLE
+#endif // defined(FP16_AVAILABLE) && __CUDA_ARCH__ != 610
+
+#if !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)) && __CUDA_ARCH__ >= CC_VOLTA
+#define FP16_MMA_AVAILABLE
+#endif // !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)) && __CUDA_ARCH__ >= CC_VOLTA
+
+#if !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)) && __CUDA_ARCH__ >= CC_TURING
+#define INT8_MMA_AVAILABLE
+#endif // !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)) && __CUDA_ARCH__ >= CC_TURING
+
+#if !(defined(GGML_USE_MUSA) && __MUSA_ARCH__ <= CC_QY1)
+#define FLASH_ATTN_AVAILABLE
+#endif // !(defined(GGML_USE_MUSA) && __MUSA_ARCH__ <= CC_QY1)
+
+static constexpr bool fast_fp16_available(const int cc) {
+    return cc >= CC_PASCAL && cc != 610;
+}
+
+static constexpr bool fp16_mma_available(const int cc) {
+    return cc < CC_OFFSET_AMD && cc >= CC_VOLTA;
+}
+
+static constexpr bool int8_mma_available(const int cc) {
+    return cc < CC_OFFSET_AMD && cc >= CC_TURING;
+}
+
+[[noreturn]]
+static __device__ void no_device_code(
+    const char * file_name, const int line, const char * function_name, const int arch, const char * arch_list) {
+
+#if defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)
+    printf("%s:%d: ERROR: HIP kernel %s has no device code compatible with HIP arch %d.\n",
+           file_name, line, function_name, arch);
+    GGML_UNUSED(arch_list);
+#else
+    printf("%s:%d: ERROR: CUDA kernel %s has no device code compatible with CUDA arch %d. ggml-cuda.cu was compiled for: %s\n",
+           file_name, line, function_name, arch, arch_list);
+#endif // defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)
+    __trap();
+
+    GGML_UNUSED(no_device_code); // suppress unused function warning
+}
+
+#ifdef __CUDA_ARCH__
+#define NO_DEVICE_CODE no_device_code(__FILE__, __LINE__, __FUNCTION__, __CUDA_ARCH__, STRINGIZE(__CUDA_ARCH_LIST__))
+#else
+#define NO_DEVICE_CODE //GGML_ABORT("NO_DEVICE_CODE not valid in host code.")
+#endif // __CUDA_ARCH__
+
+static __device__ __forceinline__ int warp_reduce_sum(int x) {
+#if !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)) && __CUDA_ARCH__ >= CC_AMPERE
+    return __reduce_add_sync(0xffffffff, x);
+#else
+#pragma unroll
+    for (int mask = 16; mask > 0; mask >>= 1) {
+        x += __shfl_xor_sync(0xffffffff, x, mask, 32);
+    }
+    return x;
+#endif // !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)) && __CUDA_ARCH__ >= CC_AMPERE
+}
+
+static __device__ __forceinline__ float warp_reduce_sum(float x) {
+#pragma unroll
+    for (int mask = 16; mask > 0; mask >>= 1) {
+        x += __shfl_xor_sync(0xffffffff, x, mask, 32);
+    }
+    return x;
+}
+
+static __device__ __forceinline__ float2 warp_reduce_sum(float2 a) {
+#pragma unroll
+    for (int mask = 16; mask > 0; mask >>= 1) {
+        a.x += __shfl_xor_sync(0xffffffff, a.x, mask, 32);
+        a.y += __shfl_xor_sync(0xffffffff, a.y, mask, 32);
+    }
+    return a;
+}
+
+static __device__ __forceinline__ half2 warp_reduce_sum(half2 a) {
+#ifdef FP16_AVAILABLE
+
+#if defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)
+#pragma unroll
+    for (int mask = 16; mask > 0; mask >>= 1) {
+        const half2 a_other = __shfl_xor_sync(0xffffffff, a, mask, 32);
+        reinterpret_cast<half&>(a.x) +=  __low2half(a_other);
+        reinterpret_cast<half&>(a.y) += __high2half(a_other);
+    }
+    return a;
+#else
+#pragma unroll
+    for (int mask = 16; mask > 0; mask >>= 1) {
+        a = __hadd2(a, __shfl_xor_sync(0xffffffff, a, mask, 32));
+    }
+    return a;
+#endif // defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)
+
+#else
+    NO_DEVICE_CODE;
+    return a;
+#endif // FP16_AVAILABLE
+}
+
+static __device__ __forceinline__ float warp_reduce_max(float x) {
+#pragma unroll
+    for (int mask = 16; mask > 0; mask >>= 1) {
+        x = fmaxf(x, __shfl_xor_sync(0xffffffff, x, mask, 32));
+    }
+    return x;
+}
+
+static __device__ __forceinline__ half ggml_cuda_hmax(const half a, const half b) {
+#ifdef FP16_AVAILABLE
+
+#if !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)) && CUDART_VERSION < CUDART_HMAX
+    return __float2half(fmaxf(__half2float(a), __half2float(b)));
+#else
+    return __hmax(a, b);
+#endif // !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)) && CUDART_VERSION < CUDART_HMAX
+
+#else
+   NO_DEVICE_CODE;
+   GGML_UNUSED(b);
+   return a;
+#endif // FP16_AVAILABLE
+}
+
+static __device__ __forceinline__ half2 ggml_cuda_hmax2(const half2 a, const half2 b) {
+#if !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__))
+
+#if CUDART_VERSION >= CUDART_HMAX
+    return __hmax2(a, b);
+#else
+    half2 ret;
+    reinterpret_cast<half&>(ret.x) = __float2half(fmaxf( __low2float(a),  __low2float(b)));
+    reinterpret_cast<half&>(ret.y) = __float2half(fmaxf(__high2float(a), __high2float(b)));
+    return ret;
+#endif // CUDART_VERSION >= CUDART_HMAX
+
+#else
+    GGML_UNUSED(a);
+    GGML_UNUSED(b);
+    NO_DEVICE_CODE;
+#endif // !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__))
+}
+
+static __device__ __forceinline__ half2 warp_reduce_max(half2 x) {
+#if !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)) && __CUDA_ARCH__ >= CC_PASCAL
+#pragma unroll
+   for (int mask = 16; mask > 0; mask >>= 1) {
+       x = ggml_cuda_hmax2(x, __shfl_xor_sync(0xffffffff, x, mask, 32));
+   }
+   return x;
+#else
+   GGML_UNUSED(x);
+   NO_DEVICE_CODE;
+#endif // !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)) && __CUDA_ARCH__ >= CC_PASCAL
+}
+
+#if CUDART_VERSION < CUDART_HMASK
+static __device__ __forceinline__ uint32_t __hgt2_mask(const half2 a, const half2 b) {
+    const uint32_t mask_low  = 0x0000FFFF * (float( __low2half(a)) > float( __low2half(b)));
+    const uint32_t mask_high = 0xFFFF0000 * (float(__high2half(a)) > float(__high2half(b)));
+    return mask_low | mask_high;
+}
+#endif // CUDART_VERSION < CUDART_HMASK
+
+static __device__ __forceinline__ int ggml_cuda_dp4a(const int a, const int b, int c) {
+#if defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)
+#if defined(__gfx906__) || defined(__gfx908__) || defined(__gfx90a__) || defined(RDNA2)
+    c = __builtin_amdgcn_sdot4(a, b, c, false);
+#elif defined(RDNA3)
+    c = __builtin_amdgcn_sudot4( true, a, true, b, c, false);
+#elif defined(__gfx1010__) || defined(__gfx900__)
+    int tmp1;
+    int tmp2;
+    asm("\n \
+        v_mul_i32_i24 %1, sext(%3), sext(%4) dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:BYTE_0 \n \
+        v_mul_i32_i24 %2, sext(%3), sext(%4) dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_1 src1_sel:BYTE_1 \n \
+        v_add3_u32 %0, %1, %2, %0 \n \
+        v_mul_i32_i24 %1, sext(%3), sext(%4) dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_2 src1_sel:BYTE_2 \n \
+        v_mul_i32_i24 %2, sext(%3), sext(%4) dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_3 src1_sel:BYTE_3 \n \
+        v_add3_u32 %0, %1, %2, %0 \n \
+        "
+        : "+v"(c), "=&v"(tmp1), "=&v"(tmp2)
+        : "v"(a), "v"(b)
+    );
+#else
+    const int8x4_t va = reinterpret_cast<const int8x4_t&>(a);
+    const int8x4_t vb = reinterpret_cast<const int8x4_t&>(b);
+    c += va[0] * vb[0] + va[1] * vb[1] + va[2] * vb[2] + va[3] * vb[3];
+#endif
+    return c;
+
+#else // defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)
+
+#if __CUDA_ARCH__ >= MIN_CC_DP4A
+    return __dp4a(a, b, c);
+#else // __CUDA_ARCH__ >= MIN_CC_DP4A
+    const int8_t * a8 = (const int8_t *) &a;
+    const int8_t * b8 = (const int8_t *) &b;
+    return c + a8[0]*b8[0] + a8[1]*b8[1] + a8[2]*b8[2] + a8[3]*b8[3];
+#endif // __CUDA_ARCH__ >= MIN_CC_DP4A
+
+#endif // defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)
+}
+
+// TODO: move to ggml-common.h
+static constexpr __device__ int8_t kvalues_iq4nl[16] = {-127, -104, -83, -65, -49, -35, -22, -10, 1, 13, 25, 38, 53, 69, 89, 113};
+
+typedef void (*dequantize_kernel_t)(const void * vx, const int64_t ib, const int iqs, dfloat2 & v);
+
+static __device__ __forceinline__ float get_alibi_slope(
+    const float max_bias, const uint32_t h, const uint32_t n_head_log2, const float m0, const float m1
+) {
+    if (max_bias <= 0.0f) {
+        return 1.0f;
+    }
+    const float base = h < n_head_log2 ? m0 : m1;
+    const int   exph = h < n_head_log2 ? h + 1 : 2*(h - n_head_log2) + 1;
+
+    return powf(base, exph);
+}
+
+template <ggml_type type>
+struct ggml_cuda_type_traits;
+
+template<>
+struct ggml_cuda_type_traits<GGML_TYPE_F16> {
+    static constexpr int qk = 1;
+    static constexpr int qr = 1;
+};
+
+template<>
+struct ggml_cuda_type_traits<GGML_TYPE_Q4_0> {
+    static constexpr int qk = QK4_0;
+    static constexpr int qr = QR4_0;
+    static constexpr int qi = QI4_0;
+};
+
+template<>
+struct ggml_cuda_type_traits<GGML_TYPE_Q4_1> {
+    static constexpr int qk = QK4_1;
+    static constexpr int qr = QR4_1;
+    static constexpr int qi = QI4_1;
+};
+
+template<>
+struct ggml_cuda_type_traits<GGML_TYPE_Q5_0> {
+    static constexpr int qk = QK5_0;
+    static constexpr int qr = QR5_0;
+    static constexpr int qi = QI5_0;
+};
+
+template<>
+struct ggml_cuda_type_traits<GGML_TYPE_Q5_1> {
+    static constexpr int qk = QK5_1;
+    static constexpr int qr = QR5_1;
+    static constexpr int qi = QI5_1;
+};
+
+template<>
+struct ggml_cuda_type_traits<GGML_TYPE_Q8_0> {
+    static constexpr int qk = QK8_0;
+    static constexpr int qr = QR8_0;
+    static constexpr int qi = QI8_0;
+};
+
+template<>
+struct ggml_cuda_type_traits<GGML_TYPE_Q2_K> {
+    static constexpr int qk = QK_K;
+    static constexpr int qr = QR2_K;
+    static constexpr int qi = QI2_K;
+};
+
+template<>
+struct ggml_cuda_type_traits<GGML_TYPE_Q3_K> {
+    static constexpr int qk = QK_K;
+    static constexpr int qr = QR3_K;
+    static constexpr int qi = QI3_K;
+};
+
+template<>
+struct ggml_cuda_type_traits<GGML_TYPE_Q4_K> {
+    static constexpr int qk = QK_K;
+    static constexpr int qr = QR4_K;
+    static constexpr int qi = QI4_K;
+};
+
+template<>
+struct ggml_cuda_type_traits<GGML_TYPE_Q5_K> {
+    static constexpr int qk = QK_K;
+    static constexpr int qr = QR5_K;
+    static constexpr int qi = QI5_K;
+};
+
+template<>
+struct ggml_cuda_type_traits<GGML_TYPE_Q6_K> {
+    static constexpr int qk = QK_K;
+    static constexpr int qr = QR6_K;
+    static constexpr int qi = QI6_K;
+};
+
+template<>
+struct ggml_cuda_type_traits<GGML_TYPE_IQ2_XXS> {
+    static constexpr int qk = QK_K;
+    static constexpr int qr = QR2_XXS;
+    static constexpr int qi = QI2_XXS;
+};
+
+template<>
+struct ggml_cuda_type_traits<GGML_TYPE_IQ2_XS> {
+    static constexpr int qk = QK_K;
+    static constexpr int qr = QR2_XS;
+    static constexpr int qi = QI2_XS;
+};
+
+template<>
+struct ggml_cuda_type_traits<GGML_TYPE_IQ2_S> {
+    static constexpr int qk = QK_K;
+    static constexpr int qr = QR2_S;
+    static constexpr int qi = QI2_S;
+};
+
+template<>
+struct ggml_cuda_type_traits<GGML_TYPE_IQ3_XXS> {
+    static constexpr int qk = QK_K;
+    static constexpr int qr = QR3_XXS;
+    static constexpr int qi = QI3_XXS;
+};
+
+template<>
+struct ggml_cuda_type_traits<GGML_TYPE_IQ1_S> {
+    static constexpr int qk = QK_K;
+    static constexpr int qr = QR1_S;
+    static constexpr int qi = QI1_S;
+};
+
+template<>
+struct ggml_cuda_type_traits<GGML_TYPE_IQ1_M> {
+    static constexpr int qk = QK_K;
+    static constexpr int qr = QR1_M;
+    static constexpr int qi = QI1_M;
+};
+
+template<>
+struct ggml_cuda_type_traits<GGML_TYPE_IQ4_NL> {
+    static constexpr int qk = QK4_NL;
+    static constexpr int qr = QR4_NL;
+    static constexpr int qi = QI4_NL;
+};
+
+template<>
+struct ggml_cuda_type_traits<GGML_TYPE_IQ4_XS> {
+    static constexpr int qk = QK_K;
+    static constexpr int qr = QR4_XS;
+    static constexpr int qi = QI4_XS;
+};
+
+template<>
+struct ggml_cuda_type_traits<GGML_TYPE_IQ3_S> {
+    static constexpr int qk = QK_K;
+    static constexpr int qr = QR3_S;
+    static constexpr int qi = QI3_S;
+};
+
+//////////////////////
+
+struct ggml_cuda_device_info {
+    int device_count;
+
+    struct cuda_device_info {
+        int     cc;                 // compute capability
+        int     nsm;                // number of streaming multiprocessors
+        size_t  smpb;               // max. shared memory per block
+        size_t  smpbo;              // max. shared memory per block (with opt-in)
+        bool    vmm;                // virtual memory support
+        size_t  vmm_granularity;    // granularity of virtual memory
+        size_t  total_vram;
+    };
+
+    cuda_device_info devices[GGML_CUDA_MAX_DEVICES] = {};
+
+    std::array<float, GGML_CUDA_MAX_DEVICES> default_tensor_split = {};
+};
+
+const ggml_cuda_device_info & ggml_cuda_info();
+
+void ggml_cuda_set_device(int device);
+int ggml_cuda_get_device();
+
+struct ggml_cuda_pool {
+    virtual ~ggml_cuda_pool() = default;
+
+    virtual void * alloc(size_t size, size_t * actual_size) = 0;
+    virtual void free(void * ptr, size_t size) = 0;
+};
+
+template<typename T>
+struct ggml_cuda_pool_alloc {
+    ggml_cuda_pool * pool = nullptr;
+    T * ptr = nullptr;
+    size_t actual_size = 0;
+
+    ggml_cuda_pool_alloc() = default;
+
+    explicit ggml_cuda_pool_alloc(ggml_cuda_pool & pool) : pool(&pool) {
+    }
+
+    ggml_cuda_pool_alloc(ggml_cuda_pool & pool, size_t size) : pool(&pool) {
+        alloc(size);
+    }
+
+    ~ggml_cuda_pool_alloc() {
+        if (ptr != nullptr) {
+            pool->free(ptr, actual_size);
+        }
+    }
+
+    // size is in number of elements
+    T * alloc(size_t size) {
+        GGML_ASSERT(pool != nullptr);
+        GGML_ASSERT(ptr == nullptr);
+        ptr = (T *) pool->alloc(size * sizeof(T), &this->actual_size);
+        return ptr;
+    }
+
+    T * alloc(ggml_cuda_pool & pool, size_t size) {
+        this->pool = &pool;
+        return alloc(size);
+    }
+
+    T * get() {
+        return ptr;
+    }
+
+    ggml_cuda_pool_alloc(const ggml_cuda_pool_alloc &) = delete;
+    ggml_cuda_pool_alloc(ggml_cuda_pool_alloc &&) = delete;
+    ggml_cuda_pool_alloc& operator=(const ggml_cuda_pool_alloc &) = delete;
+    ggml_cuda_pool_alloc& operator=(ggml_cuda_pool_alloc &&) = delete;
+};
+
+
+// backend interface
+
+struct ggml_tensor_extra_gpu {
+    void * data_device[GGML_CUDA_MAX_DEVICES]; // 1 pointer for each device for split tensors
+    cudaEvent_t events[GGML_CUDA_MAX_DEVICES][GGML_CUDA_MAX_STREAMS]; // events for synchronizing multiple GPUs
+};
+
+
+#if (CUDART_VERSION >= 12000) && defined(GGML_CUDA_USE_GRAPHS)
+#define USE_CUDA_GRAPH
+#endif
+
+struct ggml_graph_node_properties {
+    void * node_address;
+    ggml_op node_op;
+    int64_t ne[GGML_MAX_DIMS];
+    size_t nb[GGML_MAX_DIMS];
+    void * src_address[GGML_MAX_SRC];
+    int32_t op_params[GGML_MAX_OP_PARAMS / sizeof(int32_t)];
+};
+
+struct ggml_cuda_graph {
+#ifdef USE_CUDA_GRAPH
+    ~ggml_cuda_graph() {
+        if (instance != nullptr) {
+            CUDA_CHECK(cudaGraphExecDestroy(instance));
+        }
+        if (graph != nullptr) {
+            CUDA_CHECK(cudaGraphDestroy(graph));
+        }
+    }
+    cudaGraph_t graph = nullptr;
+    cudaGraphExec_t instance = nullptr;
+    size_t num_nodes = 0;
+    std::vector<cudaGraphNode_t> nodes;
+    std::vector<cudaKernelNodeParams> params;
+    bool disable_due_to_gpu_arch = false;
+    bool disable_due_to_too_many_updates = false;
+    bool disable_due_to_failed_graph_capture = false;
+    int number_consecutive_updates = 0;
+    std::vector<ggml_graph_node_properties> ggml_graph_properties;
+    std::vector<char **> updated_kernel_arg;
+#endif
+};
+
+struct ggml_backend_cuda_context {
+    int device;
+    std::string name;
+    cudaEvent_t copy_event = nullptr;
+
+    cudaStream_t streams[GGML_CUDA_MAX_DEVICES][GGML_CUDA_MAX_STREAMS] = { { nullptr } };
+    cublasHandle_t cublas_handles[GGML_CUDA_MAX_DEVICES] = {nullptr};
+
+    std::unique_ptr<ggml_cuda_graph> cuda_graph;
+
+    explicit ggml_backend_cuda_context(int device) :
+        device(device),
+        name(GGML_CUDA_NAME + std::to_string(device)) {
+    }
+
+    ~ggml_backend_cuda_context() {
+        if (copy_event != nullptr) {
+            CUDA_CHECK(cudaEventDestroy(copy_event));
+        }
+        for (int i = 0; i < GGML_CUDA_MAX_DEVICES; ++i) {
+            for (int j = 0; j < GGML_CUDA_MAX_STREAMS; ++j) {
+                if (streams[i][j] != nullptr) {
+                    CUDA_CHECK(cudaStreamDestroy(streams[i][j]));
+                }
+            }
+            if (cublas_handles[i] != nullptr) {
+                CUBLAS_CHECK(cublasDestroy(cublas_handles[i]));
+            }
+        }
+    }
+
+    cudaStream_t stream(int device, int stream) {
+        if (streams[device][stream] == nullptr) {
+            ggml_cuda_set_device(device);
+            CUDA_CHECK(cudaStreamCreateWithFlags(&streams[device][stream], cudaStreamNonBlocking));
+        }
+        return streams[device][stream];
+    }
+
+    cudaStream_t stream() {
+        return stream(device, 0);
+    }
+
+    cublasHandle_t cublas_handle(int device) {
+        if (cublas_handles[device] == nullptr) {
+            ggml_cuda_set_device(device);
+            CUBLAS_CHECK(cublasCreate(&cublas_handles[device]));
+            CUBLAS_CHECK(cublasSetMathMode(cublas_handles[device], CUBLAS_TF32_TENSOR_OP_MATH));
+        }
+        return cublas_handles[device];
+    }
+
+    cublasHandle_t cublas_handle() {
+        return cublas_handle(device);
+    }
+
+    // pool
+    std::unique_ptr<ggml_cuda_pool> pools[GGML_CUDA_MAX_DEVICES];
+
+    static std::unique_ptr<ggml_cuda_pool> new_pool_for_device(int device);
+
+    ggml_cuda_pool & pool(int device) {
+        if (pools[device] == nullptr) {
+            pools[device] = new_pool_for_device(device);
+        }
+        return *pools[device];
+    }
+
+    ggml_cuda_pool & pool() {
+        return pool(device);
+    }
+};
diff --git a/src/whisper_cpp/ggml/src/ggml-cuda/concat.cu b/src/whisper_cpp/ggml/src/ggml-cuda/concat.cu
new file mode 100644
index 00000000..dac10ec3
--- /dev/null
+++ b/src/whisper_cpp/ggml/src/ggml-cuda/concat.cu
@@ -0,0 +1,196 @@
+#include "concat.cuh"
+
+// contiguous kernels
+static __global__ void concat_f32_dim0(const float * x, const float * y, float * dst, const int ne0, const int ne00) {
+    int nidx = threadIdx.x + blockIdx.x * blockDim.x;
+    if (nidx >= ne0) {
+        return;
+    }
+
+    int offset_dst =
+        nidx +
+        blockIdx.y * ne0 +
+        blockIdx.z * ne0 * gridDim.y;
+
+    if (nidx < ne00) { // src0
+        int offset_src =
+            nidx +
+            blockIdx.y * ne00 +
+            blockIdx.z * ne00 * gridDim.y;
+        dst[offset_dst] = x[offset_src];
+    } else {
+        int offset_src =
+            (nidx - ne00) +
+            blockIdx.y * (ne0 - ne00) +
+            blockIdx.z * (ne0 - ne00) * gridDim.y;
+        dst[offset_dst] = y[offset_src];
+    }
+}
+
+static __global__ void concat_f32_dim1(const float * x, const float * y, float * dst, const int ne0, const int ne01) {
+    int nidx = threadIdx.x + blockIdx.x * blockDim.x;
+    if (nidx >= ne0) {
+        return;
+    }
+
+    int offset_dst =
+        nidx +
+        blockIdx.y * ne0 +
+        blockIdx.z * ne0 * gridDim.y;
+
+    if (blockIdx.y < ne01) { // src0
+        int offset_src =
+            nidx +
+            blockIdx.y * ne0 +
+            blockIdx.z * ne0 * ne01;
+        dst[offset_dst] = x[offset_src];
+    } else {
+        int offset_src =
+            nidx +
+            (blockIdx.y - ne01) * ne0 +
+            blockIdx.z * ne0 * (gridDim.y - ne01);
+        dst[offset_dst] = y[offset_src];
+    }
+}
+
+static __global__ void concat_f32_dim2(const float * x, const float * y, float * dst, const int ne0, const int ne02) {
+    int nidx = threadIdx.x + blockIdx.x * blockDim.x;
+    if (nidx >= ne0) {
+        return;
+    }
+
+    int offset_dst =
+        nidx +
+        blockIdx.y * ne0 +
+        blockIdx.z * ne0 * gridDim.y;
+
+    if (blockIdx.z < ne02) { // src0
+        int offset_src =
+            nidx +
+            blockIdx.y * ne0 +
+            blockIdx.z * ne0 * gridDim.y;
+        dst[offset_dst] = x[offset_src];
+    } else {
+        int offset_src =
+            nidx +
+            blockIdx.y * ne0 +
+            (blockIdx.z - ne02) * ne0 *  gridDim.y;
+        dst[offset_dst] = y[offset_src];
+    }
+}
+
+static void concat_f32_cuda(const float * x, const float * y, float * dst, int ne00, int ne01, int ne02, int ne0, int ne1, int ne2, int dim, cudaStream_t stream) {
+    int num_blocks = (ne0 + CUDA_CONCAT_BLOCK_SIZE - 1) / CUDA_CONCAT_BLOCK_SIZE;
+    dim3 gridDim(num_blocks, ne1, ne2);
+    if (dim == 0) {
+        concat_f32_dim0<<<gridDim, CUDA_CONCAT_BLOCK_SIZE, 0, stream>>>(x, y, dst, ne0, ne00);
+        return;
+    }
+    if (dim == 1) {
+        concat_f32_dim1<<<gridDim, CUDA_CONCAT_BLOCK_SIZE, 0, stream>>>(x, y, dst, ne0, ne01);
+        return;
+    }
+    concat_f32_dim2<<<gridDim, CUDA_CONCAT_BLOCK_SIZE, 0, stream>>>(x, y, dst, ne0, ne02);
+}
+
+// non-contiguous kernel (slow)
+static __global__ void concat_f32_non_cont(
+        const char * src0,
+        const char * src1,
+              char * dst,
+           int64_t   ne00,
+           int64_t   ne01,
+           int64_t   ne02,
+           int64_t   ne03,
+          uint64_t   nb00,
+          uint64_t   nb01,
+          uint64_t   nb02,
+          uint64_t   nb03,
+           int64_t /*ne10*/,
+           int64_t /*ne11*/,
+           int64_t /*ne12*/,
+           int64_t /*ne13*/,
+          uint64_t   nb10,
+          uint64_t   nb11,
+          uint64_t   nb12,
+          uint64_t   nb13,
+           int64_t   ne0,
+           int64_t /*ne1*/,
+           int64_t /*ne2*/,
+           int64_t /*ne3*/,
+          uint64_t   nb0,
+          uint64_t   nb1,
+          uint64_t   nb2,
+          uint64_t   nb3,
+          int32_t   dim) {
+    const int64_t i3 = blockIdx.z;
+    const int64_t i2 = blockIdx.y;
+    const int64_t i1 = blockIdx.x;
+
+    int64_t o[4] = {0, 0, 0, 0};
+    o[dim] = dim == 0 ? ne00 : (dim == 1 ? ne01 : (dim == 2 ? ne02 : ne03));
+
+    const float * x;
+
+    for (int i0 = threadIdx.x; i0 < ne0; i0 += blockDim.x) {
+        if (i0 < ne00 && i1 < ne01 && i2 < ne02 && i3 < ne03) {
+            x = (const float *)(src0 + (i3       )*nb03 + (i2       )*nb02 + (i1       )*nb01 + (i0       )*nb00);
+        } else {
+            x = (const float *)(src1 + (i3 - o[3])*nb13 + (i2 - o[2])*nb12 + (i1 - o[1])*nb11 + (i0 - o[0])*nb10);
+        }
+
+        float * y = (float *)(dst + i3*nb3 + i2*nb2 + i1*nb1 + i0*nb0);
+
+        *y = *x;
+    }
+}
+
+
+void ggml_cuda_op_concat(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
+    const ggml_tensor * src0 = dst->src[0];
+    const ggml_tensor * src1 = dst->src[1];
+
+    cudaStream_t stream = ctx.stream();
+
+    const int32_t dim = ((int32_t *) dst->op_params)[0];
+
+    GGML_ASSERT(src0->type == GGML_TYPE_F32);
+    GGML_ASSERT(src1->type == GGML_TYPE_F32);
+    GGML_ASSERT(dst->type  == GGML_TYPE_F32);
+
+    if (ggml_is_contiguous(src0) && ggml_is_contiguous(src1)) {
+        const float * src0_d = (const float *)src0->data;
+        const float * src1_d = (const float *)src1->data;
+
+        float * dst_d = (float *)dst->data;
+
+        if (dim != 3) {
+            for (int i3 = 0; i3 < dst->ne[3]; i3++) {
+                concat_f32_cuda(
+                        src0_d + i3 * (src0->nb[3] / 4),
+                        src1_d + i3 * (src1->nb[3] / 4),
+                        dst_d + i3 * ( dst->nb[3] / 4),
+                        src0->ne[0], src0->ne[1], src0->ne[2],
+                        dst->ne[0],  dst->ne[1],  dst->ne[2], dim, stream);
+            }
+        } else {
+            const size_t size0 = ggml_nbytes(src0);
+            const size_t size1 = ggml_nbytes(src1);
+
+            CUDA_CHECK(cudaMemcpyAsync(dst_d,           src0_d, size0, cudaMemcpyDeviceToDevice, stream));
+            CUDA_CHECK(cudaMemcpyAsync(dst_d + size0/4, src1_d, size1, cudaMemcpyDeviceToDevice, stream));
+        }
+    } else {
+        dim3 grid_dim(dst->ne[1], dst->ne[2], dst->ne[3]);
+        concat_f32_non_cont<<<grid_dim, CUDA_CONCAT_BLOCK_SIZE, 0, stream>>>(
+                (const char *)src0->data,
+                (const char *)src1->data,
+                (      char *)dst->data,
+                src0->ne[0], src0->ne[1], src0->ne[2], src0->ne[3],
+                src0->nb[0], src0->nb[1], src0->nb[2], src0->nb[3],
+                src1->ne[0], src1->ne[1], src1->ne[2], src1->ne[3],
+                src1->nb[0], src1->nb[1], src1->nb[2], src1->nb[3],
+                dst->ne[0],  dst->ne[1],  dst->ne[2],  dst->ne[3],
+                dst->nb[0],  dst->nb[1],  dst->nb[2],  dst->nb[3], dim);
+    }
+}
diff --git a/src/whisper_cpp/ggml/src/ggml-cuda/concat.cuh b/src/whisper_cpp/ggml/src/ggml-cuda/concat.cuh
new file mode 100644
index 00000000..aa506a05
--- /dev/null
+++ b/src/whisper_cpp/ggml/src/ggml-cuda/concat.cuh
@@ -0,0 +1,5 @@
+#include "common.cuh"
+
+#define CUDA_CONCAT_BLOCK_SIZE 256
+
+void ggml_cuda_op_concat(ggml_backend_cuda_context & ctx, ggml_tensor * dst);
diff --git a/src/whisper_cpp/ggml/src/ggml-cuda/conv-transpose-1d.cu b/src/whisper_cpp/ggml/src/ggml-cuda/conv-transpose-1d.cu
new file mode 100644
index 00000000..b1e94d6f
--- /dev/null
+++ b/src/whisper_cpp/ggml/src/ggml-cuda/conv-transpose-1d.cu
@@ -0,0 +1,87 @@
+#include "conv-transpose-1d.cuh"
+
+static  __global__ void conv_transpose_1d_kernel(
+        const int s0, const int p0, const int d0, const int output_size,
+        const int src0_ne0, const int src0_ne1, const int src0_ne2, const int src0_ne3,
+        const int src1_ne0, const int src1_ne1, const int src1_ne2, const int src1_ne3,
+        const int dst_ne0, const int dst_ne1, const int dst_ne2, const int dst_ne3,
+        const float * src0, const float * src1,  float * dst) {
+    int global_index = threadIdx.x + blockIdx.x * blockDim.x;
+    if (global_index >= output_size) {
+        return;
+    }
+
+    int out_index = global_index / dst_ne0;
+
+    float accumulator = 0;
+
+    for (int c = 0; c < src0_ne2; c++) {
+        int idx = global_index % dst_ne0;
+
+        int kernel_offset = (src0_ne0 * src0_ne1 * c) + (out_index * src0_ne0);
+        int input_offset = src1_ne0 * c;
+
+        for (int i = 0; i < src1_ne0; i++) {
+            if (!(idx >= i*s0 && idx < i*s0 + src0_ne0)) {
+                continue;
+            }
+            int weight_idx = idx - i*s0;
+
+            float kernel_weight = src0[kernel_offset + weight_idx];
+            float input_value =  src1[input_offset+i];
+
+            accumulator += kernel_weight * input_value;
+        }
+    }
+    dst[global_index] = accumulator;
+}
+
+static void conv_transpose_1d_f32_f32_cuda(
+        const int s0, const int p0, const int d0, const int output_size,
+        const int src0_ne0, const int src0_ne1, const int src0_ne2, const int src0_ne3,
+        const int src1_ne0, const int src1_ne1, const int src1_ne2, const int src1_ne3,
+        const int dst_ne0, const int dst_ne1, const int dst_ne2, const int dst_ne3,
+        const float * src0, const float * src1,  float * dst,
+        cudaStream_t stream) {
+
+    const int num_blocks = (output_size + CUDA_CONV_TRANPOSE_1D_BLOCK_SIZE - 1) / CUDA_CONV_TRANPOSE_1D_BLOCK_SIZE;
+    conv_transpose_1d_kernel<<<num_blocks,CUDA_CONV_TRANPOSE_1D_BLOCK_SIZE, 0, stream>>>(
+        s0,p0,d0,output_size,
+        src0_ne0, src0_ne1,  src0_ne2, src0_ne3,
+        src1_ne0, src1_ne1,  src1_ne2, src1_ne3,
+        dst_ne0,  dst_ne1,   dst_ne2,  dst_ne3,
+        src0,src1, dst);
+}
+
+void ggml_cuda_op_conv_transpose_1d(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
+    const ggml_tensor * src0 = dst->src[0];
+    const float * src0_d = (const float *)src0->data;
+
+    const ggml_tensor * src1 = dst->src[1];
+    const float * src1_d = (const float *)src1->data;
+
+    float * dst_d = (float *)dst->data;
+    cudaStream_t stream = ctx.stream();
+
+    GGML_ASSERT(src0->type == GGML_TYPE_F32);
+    GGML_ASSERT( dst->type == GGML_TYPE_F32);
+
+    GGML_ASSERT(ggml_is_contiguous(src0));
+    GGML_ASSERT(ggml_is_contiguous(src1));
+
+    const int32_t * opts = (const int32_t *)dst->op_params;
+
+    const int s0 = opts[0];
+    const int p0 = 0;//opts[3];
+    const int d0 = 1;//opts[4];
+
+    const int64_t kernel_size = ggml_nelements(src0);
+    const int64_t input_size = ggml_nelements(src1);
+    const int64_t output_size = ggml_nelements(dst);
+
+    conv_transpose_1d_f32_f32_cuda(s0, p0, d0, output_size,
+        src0->ne[0], src0->ne[1], src0->ne[2], src0->ne[3],
+        src1->ne[0], src1->ne[1], src1->ne[2], src1->ne[3],
+        dst->ne[0], dst->ne[1], dst->ne[2], dst->ne[3],
+        src0_d, src1_d, dst_d, stream);
+}
diff --git a/src/whisper_cpp/ggml/src/ggml-cuda/conv-transpose-1d.cuh b/src/whisper_cpp/ggml/src/ggml-cuda/conv-transpose-1d.cuh
new file mode 100644
index 00000000..6c2cf666
--- /dev/null
+++ b/src/whisper_cpp/ggml/src/ggml-cuda/conv-transpose-1d.cuh
@@ -0,0 +1,5 @@
+#include "common.cuh"
+
+#define CUDA_CONV_TRANPOSE_1D_BLOCK_SIZE 256
+
+void ggml_cuda_op_conv_transpose_1d(ggml_backend_cuda_context & ctx, ggml_tensor * dst);
diff --git a/src/whisper_cpp/ggml/src/ggml-cuda/convert.cu b/src/whisper_cpp/ggml/src/ggml-cuda/convert.cu
new file mode 100644
index 00000000..c0a44470
--- /dev/null
+++ b/src/whisper_cpp/ggml/src/ggml-cuda/convert.cu
@@ -0,0 +1,686 @@
+#include "convert.cuh"
+#include "dequantize.cuh"
+
+#define CUDA_Q8_0_NE_ALIGN 2048
+
+template <int qk, int qr, dequantize_kernel_t dequantize_kernel, typename dst_t>
+static __global__ void dequantize_block(const void * __restrict__ vx, dst_t * __restrict__ y, const int64_t k) {
+    const int64_t i = (int64_t)2*(blockDim.x*blockIdx.x + threadIdx.x);
+
+    if (i >= k) {
+        return;
+    }
+
+    const int64_t ib = i/qk; // block index
+    const int64_t iqs = (i%qk)/qr; // quant index
+    const int64_t iybs = i - i%qk; // y block start index
+    const int64_t y_offset = qr == 1 ? 1 : qk/2;
+
+    // dequantize
+    dfloat2 v;
+    dequantize_kernel(vx, ib, iqs, v);
+
+    y[iybs + iqs + 0]        = v.x;
+    y[iybs + iqs + y_offset] = v.y;
+}
+
+template <bool need_check>
+static __global__ void dequantize_block_q8_0_f16(const void * __restrict__ vx, half * __restrict__ y, const int64_t k) {
+#if __CUDA_ARCH__ >= CC_PASCAL
+    constexpr int nint = CUDA_Q8_0_NE_ALIGN/sizeof(int) + WARP_SIZE;
+
+    const int64_t   i0 = CUDA_Q8_0_NE_ALIGN*blockIdx.x;
+    const int * x0 = ((int *) vx) + blockIdx.x * nint;
+    half2 * y2 = (half2 *) (y + i0);
+
+    __shared__ int vals[nint];
+
+#pragma unroll
+    for (int ix0 = 0; ix0 < nint; ix0 += WARP_SIZE) {
+        if (need_check && i0*sizeof(block_q8_0)/QK8_0 + sizeof(int)*(ix0 + threadIdx.x) >= k*sizeof(block_q8_0)/QK8_0) {
+            break;
+        }
+
+        const int ix = ix0 + threadIdx.x;
+        vals[ix] = x0[ix];
+    }
+
+    __syncthreads();
+
+#pragma unroll
+    for (int iy = 0; iy < CUDA_Q8_0_NE_ALIGN; iy += 2*WARP_SIZE) {
+        if (need_check && i0 + iy + 2*threadIdx.x >= k) {
+            return;
+        }
+
+        const half * b0 = ((const half  *) vals) + (sizeof(block_q8_0)/sizeof(half)) * ((iy + 2*threadIdx.x)/QK8_0);
+        const half    d = *b0;
+        const char2  qs = ((const char2 *) (b0 + 1))[threadIdx.x % (QK8_0/2)];
+
+        y2[iy/2 + threadIdx.x] = __hmul2(make_half2(qs.x, qs.y), __half2half2(d));
+    }
+#else
+    GGML_UNUSED(vx);
+    GGML_UNUSED(y);
+    GGML_UNUSED(k);
+    NO_DEVICE_CODE;
+#endif // __CUDA_ARCH__ >= CC_PASCAL
+}
+
+template<typename dst_t>
+static __global__ void dequantize_block_q4_0(const void * __restrict__ vx, dst_t * __restrict__ yy, int nb32) {
+
+    const int64_t i = blockIdx.x;
+
+    // assume 32 threads
+    const int64_t tid = threadIdx.x;
+    const int64_t il  = tid/8;
+    const int64_t ir  = tid%8;
+    const int64_t ib = 8*i + ir;
+    if (ib >= nb32) {
+        return;
+    }
+
+    dst_t * y = yy + 256*i + 32*ir + 4*il;
+
+    const block_q4_0 * x = (const block_q4_0 *)vx + ib;
+    const float d = __half2float(x->d);
+    const float dm = -8*d;
+
+    const uint8_t * q = x->qs + 4*il;
+
+    for (int l = 0; l < 4; ++l) {
+        y[l+ 0] = d * (q[l] & 0xF) + dm;
+        y[l+16] = d * (q[l] >>  4) + dm;
+    }
+}
+
+template<typename dst_t>
+static __global__ void dequantize_block_q4_1(const void * __restrict__ vx, dst_t * __restrict__ yy, int nb32) {
+
+    const int64_t i = blockIdx.x;
+
+    // assume 32 threads
+    const int64_t tid = threadIdx.x;
+    const int64_t il  = tid/8;
+    const int64_t ir  = tid%8;
+    const int64_t ib = 8*i + ir;
+    if (ib >= nb32) {
+        return;
+    }
+
+    dst_t * y = yy + 256*i + 32*ir + 4*il;
+
+    const block_q4_1 * x = (const block_q4_1 *)vx + ib;
+    const float2 d = __half22float2(x->dm);
+
+    const uint8_t * q = x->qs + 4*il;
+
+    for (int l = 0; l < 4; ++l) {
+        y[l+ 0] = d.x * (q[l] & 0xF) + d.y;
+        y[l+16] = d.x * (q[l] >>  4) + d.y;
+    }
+}
+
+//================================== k-quants
+
+template<typename dst_t>
+static __global__ void dequantize_block_q2_K(const void * __restrict__ vx, dst_t * __restrict__ yy) {
+
+    const int64_t i   = blockIdx.x;
+    const block_q2_K * x = (const block_q2_K *) vx;
+
+    const int64_t tid = threadIdx.x;
+    const int64_t n   = tid/32;
+    const int64_t l   = tid - 32*n;
+    const int64_t is  = 8*n + l/16;
+
+    const uint8_t q = x[i].qs[32*n + l];
+    dst_t * y = yy + i*QK_K + 128*n;
+
+    float dall = __low2half(x[i].dm);
+    float dmin = __high2half(x[i].dm);
+    y[l+ 0] = dall * (x[i].scales[is+0] & 0xF) * ((q >> 0) & 3) - dmin * (x[i].scales[is+0] >> 4);
+    y[l+32] = dall * (x[i].scales[is+2] & 0xF) * ((q >> 2) & 3) - dmin * (x[i].scales[is+2] >> 4);
+    y[l+64] = dall * (x[i].scales[is+4] & 0xF) * ((q >> 4) & 3) - dmin * (x[i].scales[is+4] >> 4);
+    y[l+96] = dall * (x[i].scales[is+6] & 0xF) * ((q >> 6) & 3) - dmin * (x[i].scales[is+6] >> 4);
+}
+
+template<typename dst_t>
+static __global__ void dequantize_block_q3_K(const void * __restrict__ vx, dst_t * __restrict__ yy) {
+
+    const int64_t i = blockIdx.x;
+    const block_q3_K * x = (const block_q3_K *) vx;
+
+    const int64_t r = threadIdx.x/4;
+    const int64_t tid = r/2;
+    const int64_t is0 = r%2;
+    const int64_t l0 = 16*is0 + 4*(threadIdx.x%4);
+    const int64_t n = tid / 4;
+    const int64_t j = tid - 4*n;
+
+    uint8_t m = 1 << (4*n + j);
+    int64_t is = 8*n + 2*j + is0;
+    int shift = 2*j;
+
+    int8_t us = is <  4 ? (x[i].scales[is-0] & 0xF) | (((x[i].scales[is+8] >> 0) & 3) << 4) :
+                is <  8 ? (x[i].scales[is-0] & 0xF) | (((x[i].scales[is+4] >> 2) & 3) << 4) :
+                is < 12 ? (x[i].scales[is-8] >>  4) | (((x[i].scales[is+0] >> 4) & 3) << 4) :
+                          (x[i].scales[is-8] >>  4) | (((x[i].scales[is-4] >> 6) & 3) << 4);
+    float d_all = x[i].d;
+    float dl = d_all * (us - 32);
+
+    dst_t * y = yy + i*QK_K + 128*n + 32*j;
+    const uint8_t * q = x[i].qs + 32*n;
+    const uint8_t * hm = x[i].hmask;
+
+    for (int l = l0; l < l0+4; ++l) y[l] = dl * ((int8_t)((q[l] >> shift) & 3) - ((hm[l] & m) ? 0 : 4));
+}
+
+static inline __device__ void get_scale_min_k4(int j, const uint8_t * q, uint8_t & d, uint8_t & m) {
+    if (j < 4) {
+        d = q[j] & 63; m = q[j + 4] & 63;
+    } else {
+        d = (q[j+4] & 0xF) | ((q[j-4] >> 6) << 4);
+        m = (q[j+4] >>  4) | ((q[j-0] >> 6) << 4);
+    }
+}
+
+template<typename dst_t>
+static __global__ void dequantize_block_q4_K(const void * __restrict__ vx, dst_t * __restrict__ yy) {
+    const block_q4_K * x = (const block_q4_K *) vx;
+
+    const int64_t i = blockIdx.x;
+
+    // assume 32 threads
+    const int64_t tid = threadIdx.x;
+    const int64_t il  = tid/8;
+    const int64_t ir  = tid%8;
+    const int64_t is  = 2*il;
+    const int64_t n   = 4;
+
+    dst_t * y = yy + i*QK_K + 64*il + n*ir;
+
+    const float dall = __low2half(x[i].dm);
+    const float dmin = __high2half(x[i].dm);
+
+    const uint8_t * q = x[i].qs + 32*il + n*ir;
+
+    uint8_t sc, m;
+    get_scale_min_k4(is + 0, x[i].scales, sc, m);
+    const float d1 = dall * sc; const float m1 = dmin * m;
+    get_scale_min_k4(is + 1, x[i].scales, sc, m);
+    const float d2 = dall * sc; const float m2 = dmin * m;
+    for (int l = 0; l < n; ++l) {
+        y[l + 0] = d1 * (q[l] & 0xF) - m1;
+        y[l +32] = d2 * (q[l] >>  4) - m2;
+    }
+}
+
+template<typename dst_t>
+static __global__ void dequantize_block_q5_K(const void * __restrict__ vx, dst_t * __restrict__ yy) {
+    const block_q5_K * x = (const block_q5_K *) vx;
+
+    const int64_t i = blockIdx.x;
+
+    // assume 64 threads - this is very slightly better than the one below
+    const int64_t tid = threadIdx.x;
+    const int64_t il  = tid/16;   // il is in 0...3
+    const int64_t ir  = tid%16;   // ir is in 0...15
+    const int64_t is  = 2*il;     // is is in 0...6
+
+    dst_t * y = yy + i*QK_K + 64*il + 2*ir;
+
+    const float dall = __low2half(x[i].dm);
+    const float dmin = __high2half(x[i].dm);
+
+    const uint8_t * ql = x[i].qs + 32*il + 2*ir;
+    const uint8_t * qh = x[i].qh + 2*ir;
+
+    uint8_t sc, m;
+    get_scale_min_k4(is + 0, x[i].scales, sc, m);
+    const float d1 = dall * sc; const float m1 = dmin * m;
+    get_scale_min_k4(is + 1, x[i].scales, sc, m);
+    const float d2 = dall * sc; const float m2 = dmin * m;
+
+    uint8_t   hm  = 1 << (2*il);
+    y[ 0] = d1 * ((ql[ 0] & 0xF) + (qh[ 0] & hm ? 16 : 0)) - m1;
+    y[ 1] = d1 * ((ql[ 1] & 0xF) + (qh[ 1] & hm ? 16 : 0)) - m1;
+    hm <<= 1;
+    y[32] = d2 * ((ql[ 0] >>  4) + (qh[ 0] & hm ? 16 : 0)) - m2;
+    y[33] = d2 * ((ql[ 1] >>  4) + (qh[ 1] & hm ? 16 : 0)) - m2;
+}
+
+template<typename dst_t>
+static __global__ void dequantize_block_q6_K(const void * __restrict__ vx, dst_t * __restrict__ yy) {
+    const block_q6_K * x = (const block_q6_K *) vx;
+
+    const int64_t i = blockIdx.x;
+
+    // assume 64 threads - this is very slightly better than the one below
+    const int64_t tid = threadIdx.x;
+    const int64_t ip  = tid/32;   // ip is 0 or 1
+    const int64_t il  = tid - 32*ip; // 0...32
+    const int64_t is  = 8*ip + il/16;
+
+    dst_t * y = yy + i*QK_K + 128*ip + il;
+
+    const float d = x[i].d;
+
+    const uint8_t * ql = x[i].ql + 64*ip + il;
+    const uint8_t   qh = x[i].qh[32*ip + il];
+    const int8_t  * sc = x[i].scales + is;
+
+    y[ 0] = d * sc[0] * ((int8_t)((ql[ 0] & 0xF) | (((qh >> 0) & 3) << 4)) - 32);
+    y[32] = d * sc[2] * ((int8_t)((ql[32] & 0xF) | (((qh >> 2) & 3) << 4)) - 32);
+    y[64] = d * sc[4] * ((int8_t)((ql[ 0]  >> 4) | (((qh >> 4) & 3) << 4)) - 32);
+    y[96] = d * sc[6] * ((int8_t)((ql[32]  >> 4) | (((qh >> 6) & 3) << 4)) - 32);
+}
+
+template<typename dst_t>
+static __global__ void dequantize_block_iq2_xxs(const void * __restrict__ vx, dst_t * __restrict__ yy) {
+
+    const int64_t i   = blockIdx.x;
+    const block_iq2_xxs * x = (const block_iq2_xxs  *) vx;
+
+    const int64_t tid = threadIdx.x;
+    const int64_t il = tid/8; // 0...3
+    const int64_t ib = tid%8; // 0...7
+    dst_t * y = yy + i*QK_K + 32*ib + 8*il;
+    const uint16_t * q2 = x[i].qs + 4*ib;
+    const uint8_t  * aux8 = (const uint8_t *)q2;
+    const uint8_t  * grid = (const uint8_t *)(iq2xxs_grid + aux8[il]);
+    const uint32_t aux32 = q2[2] | (q2[3] << 16);
+    const float d = (float)x[i].d * (0.5f + (aux32 >> 28)) * 0.25f;
+    const uint8_t signs = ksigns_iq2xs[(aux32 >> 7*il) & 127];
+    for (int j = 0; j < 8; ++j) y[j] = d * grid[j] * (signs & kmask_iq2xs[j] ? -1.f : 1.f);
+}
+
+template<typename dst_t>
+static __global__ void dequantize_block_iq2_xs(const void * __restrict__ vx, dst_t * __restrict__ yy) {
+
+    const int64_t i   = blockIdx.x;
+    const block_iq2_xs * x = (const block_iq2_xs *) vx;
+
+    const int64_t tid = threadIdx.x;
+    const int64_t il = tid/8; // 0...3
+    const int64_t ib = tid%8; // 0...7
+    dst_t * y = yy + i*QK_K + 32*ib + 8*il;
+    const uint16_t * q2 = x[i].qs + 4*ib;
+    const uint8_t  * grid = (const uint8_t *)(iq2xs_grid + (q2[il] & 511));
+    const float d = (float)x[i].d * (0.5f + ((x[i].scales[ib] >> 4*(il/2)) & 0xf)) * 0.25f;
+    const uint8_t signs = ksigns_iq2xs[q2[il] >> 9];
+    for (int j = 0; j < 8; ++j) y[j] = d * grid[j] * (signs & kmask_iq2xs[j] ? -1.f : 1.f);
+}
+
+template<typename dst_t>
+static __global__ void dequantize_block_iq2_s(const void * __restrict__ vx, dst_t * __restrict__ yy) {
+
+    const int64_t i   = blockIdx.x;
+    const block_iq2_s * x = (const block_iq2_s *) vx;
+
+    const int64_t tid = threadIdx.x;
+    const int64_t il = tid/8; // 0...3
+    const int64_t ib = tid%8; // 0...7
+    dst_t * y = yy + i*QK_K + 32*ib + 8*il;
+    const uint8_t * grid = (const uint8_t *)(iq2s_grid + (x[i].qs[4*ib+il] | ((x[i].qh[ib] << (8-2*il)) & 0x300)));
+    const float d = (float)x[i].d * (0.5f + ((x[i].scales[ib] >> 4*(il/2)) & 0xf)) * 0.25f;
+    const uint8_t signs = x[i].qs[QK_K/8+4*ib+il];
+    for (int j = 0; j < 8; ++j) y[j] = d * grid[j] * (signs & kmask_iq2xs[j] ? -1.f : 1.f);
+}
+
+template<typename dst_t>
+static __global__ void dequantize_block_iq3_xxs(const void * __restrict__ vx, dst_t * __restrict__ yy) {
+
+    const int64_t i   = blockIdx.x;
+    const block_iq3_xxs * x = (const block_iq3_xxs  *) vx;
+
+    const int64_t tid = threadIdx.x;
+    const int64_t il = tid/8; // 0...3
+    const int64_t ib = tid%8; // 0...7
+    dst_t * y = yy + i*QK_K + 32*ib + 8*il;
+    const uint8_t  * q3 = x[i].qs + 8*ib;
+    const uint16_t * gas = (const uint16_t *)(x[i].qs + QK_K/4) + 2*ib;
+    const uint8_t  * grid1 = (const uint8_t *)(iq3xxs_grid + q3[2*il+0]);
+    const uint8_t  * grid2 = (const uint8_t *)(iq3xxs_grid + q3[2*il+1]);
+    const uint32_t aux32 = gas[0] | (gas[1] << 16);
+    const float d = (float)x[i].d * (0.5f + (aux32 >> 28)) * 0.5f;
+    const uint8_t signs = ksigns_iq2xs[(aux32 >> 7*il) & 127];
+    for (int j = 0; j < 4; ++j) {
+        y[j+0] = d * grid1[j] * (signs & kmask_iq2xs[j+0] ? -1.f : 1.f);
+        y[j+4] = d * grid2[j] * (signs & kmask_iq2xs[j+4] ? -1.f : 1.f);
+    }
+}
+
+template<typename dst_t>
+static __global__ void dequantize_block_iq3_s(const void * __restrict__ vx, dst_t * __restrict__ yy) {
+
+    const int64_t i   = blockIdx.x;
+    const block_iq3_s * x = (const block_iq3_s *) vx;
+
+    const int64_t tid = threadIdx.x;
+    const int64_t il = tid/8; // 0...3
+    const int64_t ib = tid%8; // 0...7
+    dst_t * y = yy + i*QK_K + 32*ib + 8*il;
+    const uint8_t * qs = x[i].qs + 8*ib;
+    const uint8_t * grid1 = (const uint8_t *)(iq3s_grid + (qs[2*il+0] | ((x[i].qh[ib] << (8-2*il)) & 256)));
+    const uint8_t * grid2 = (const uint8_t *)(iq3s_grid + (qs[2*il+1] | ((x[i].qh[ib] << (7-2*il)) & 256)));
+    const float d = (float)x[i].d * (1 + 2*((x[i].scales[ib/2] >> 4*(ib%2)) & 0xf));
+    const uint8_t signs = x[i].signs[4*ib + il];
+    for (int j = 0; j < 4; ++j) {
+        y[j+0] = d * grid1[j] * (signs & kmask_iq2xs[j+0] ? -1.f : 1.f);
+        y[j+4] = d * grid2[j] * (signs & kmask_iq2xs[j+4] ? -1.f : 1.f);
+    }
+}
+
+template<typename dst_t>
+static __global__ void dequantize_block_iq1_s(const void * __restrict__ vx, dst_t * __restrict__ yy) {
+
+    const int64_t i   = blockIdx.x;
+    const block_iq1_s * x = (const block_iq1_s  *) vx;
+
+    const int64_t tid = threadIdx.x;
+    const int64_t il = tid/8; // 0...3
+    const int64_t ib = tid%8; // 0...7
+    dst_t * y = yy + i*QK_K + 32*ib + 8*il;
+    const float delta = x[i].qh[ib] & 0x8000 ? -1 - IQ1S_DELTA : -1 + IQ1S_DELTA;
+    const float d = (float)x[i].d * (2*((x[i].qh[ib] >> 12) & 7) + 1);
+    uint32_t grid32[2]; const int8_t * q = (const int8_t *)grid32;
+    grid32[0] = iq1s_grid_gpu[x[i].qs[4*ib+il] | (((x[i].qh[ib] >> 3*il) & 7) << 8)];
+    grid32[1] = (grid32[0] >> 4) & 0x0f0f0f0f;
+    grid32[0] &= 0x0f0f0f0f;
+    for (int j = 0; j < 8; ++j) {
+        y[j] = d * (q[j] + delta);
+    }
+}
+
+template<typename dst_t>
+static __global__ void dequantize_block_iq1_m(const void * __restrict__ vx, dst_t * __restrict__ yy) {
+
+    const int64_t i   = blockIdx.x;
+    const block_iq1_m * x = (const block_iq1_m  *) vx;
+
+    const int64_t tid = threadIdx.x;
+    const int64_t il = tid/8; // 0...3
+    const int64_t ib = tid%8; // 0...7
+    dst_t * y = yy + i*QK_K + 32*ib + 8*il;
+    const uint16_t * sc = (const uint16_t *)x[i].scales;
+    iq1m_scale_t scale;
+    scale.u16 = (sc[0] >> 12) | ((sc[1] >> 8) & 0x00f0) | ((sc[2] >> 4) & 0x0f00) | (sc[3] & 0xf000);
+    const int64_t ib16 = 2*ib + il/2; // sc[ib16/4] >> 3*(ib16%4) -> sc[ib/2] >> 3*((2*ib+il/2)%4);
+    const float d = (float)scale.f16 * (2*((sc[ib16/4] >> 3*(ib16%4)) & 0x7) + 1);
+    const float delta = x[i].qh[2*ib+il/2] & (0x08 << 4*(il%2)) ? -1 - IQ1M_DELTA : -1 + IQ1M_DELTA;
+    uint32_t grid32[2]; const int8_t * q = (const int8_t *)grid32;
+    grid32[0] = iq1s_grid_gpu[x[i].qs[4*ib+il] | (((x[i].qh[2*ib+il/2] >> 4*(il%2)) & 7) << 8)];
+    grid32[1] = (grid32[0] >> 4) & 0x0f0f0f0f;
+    grid32[0] &= 0x0f0f0f0f;
+    for (int j = 0; j < 8; ++j) {
+        y[j] = d * (q[j] + delta);
+    }
+}
+
+template<typename dst_t>
+static __global__ void dequantize_block_iq4_nl(const void * __restrict__ vx, dst_t * __restrict__ yy) {
+
+    const int64_t i   = blockIdx.x;
+    const block_iq4_nl * x = (const block_iq4_nl *) vx + i*(QK_K/QK4_NL);
+
+    const int64_t tid = threadIdx.x;
+    const int64_t il = tid/8; // 0...3
+    const int64_t ib = tid%8; // 0...7
+    dst_t * y = yy + i*QK_K + 32*ib + 4*il;
+    const uint8_t  * q4 = x[ib].qs + 4*il;
+    const float d = (float)x[ib].d;
+    for (int j = 0; j < 4; ++j) {
+        y[j+ 0] = d * kvalues_iq4nl[q4[j] & 0xf];
+        y[j+16] = d * kvalues_iq4nl[q4[j] >>  4];
+    }
+}
+
+template<typename dst_t>
+static __global__ void dequantize_block_iq4_xs(const void * __restrict__ vx, dst_t * __restrict__ yy) {
+    const int64_t i   = blockIdx.x;
+    const block_iq4_xs * x = (const block_iq4_xs *)vx;
+
+    const int64_t tid = threadIdx.x;
+    const int64_t il = tid/8; // 0...3
+    const int64_t ib = tid%8; // 0...7
+    dst_t * y = yy + i*QK_K + 32*ib + 4*il;
+    const uint8_t  * q4 = x[i].qs + 16*ib + 4*il;
+    const float d = (float)x[i].d * ((((x[i].scales_l[ib/2] >> 4*(ib%2)) & 0xf) | (((x[i].scales_h >> 2*ib) & 3) << 4)) - 32);
+    for (int j = 0; j < 4; ++j) {
+        y[j+ 0] = d * kvalues_iq4nl[q4[j] & 0xf];
+        y[j+16] = d * kvalues_iq4nl[q4[j] >>  4];
+    }
+}
+
+template <int qk, int qr, dequantize_kernel_t dequantize_kernel, typename dst_t>
+static void dequantize_block_cuda(const void * __restrict__ vx, dst_t * __restrict__ y, const int64_t k, cudaStream_t stream) {
+    const int num_blocks = (k + 2*CUDA_DEQUANTIZE_BLOCK_SIZE - 1) / (2*CUDA_DEQUANTIZE_BLOCK_SIZE);
+    dequantize_block<qk, qr, dequantize_kernel><<<num_blocks, CUDA_DEQUANTIZE_BLOCK_SIZE, 0, stream>>>(vx, y, k);
+}
+
+static void dequantize_block_q8_0_f16_cuda(const void * __restrict__ vx, half * __restrict__ y, const int64_t k, cudaStream_t stream) {
+    const int num_blocks = (k + CUDA_Q8_0_NE_ALIGN - 1) / CUDA_Q8_0_NE_ALIGN;
+    if (k % CUDA_Q8_0_NE_ALIGN == 0) {
+        const bool need_check = false;
+        dequantize_block_q8_0_f16<need_check><<<num_blocks, WARP_SIZE, 0, stream>>>(vx, y, k);
+    } else {
+        const bool need_check = true;
+        dequantize_block_q8_0_f16<need_check><<<num_blocks, WARP_SIZE, 0, stream>>>(vx, y, k);
+    }
+}
+
+template<typename dst_t>
+static void dequantize_row_q2_K_cuda(const void * vx, dst_t * y, const int64_t k, cudaStream_t stream) {
+    const int nb = k / QK_K;
+    dequantize_block_q2_K<<<nb, 64, 0, stream>>>(vx, y);
+}
+
+template<typename dst_t>
+static void dequantize_row_q3_K_cuda(const void * vx, dst_t * y, const int64_t k, cudaStream_t stream) {
+    const int nb = k / QK_K;
+    dequantize_block_q3_K<<<nb, 64, 0, stream>>>(vx, y);
+}
+
+template<typename dst_t>
+static void dequantize_row_q4_0_cuda(const void * vx, dst_t * y, const int64_t k, cudaStream_t stream) {
+    const int nb32 = k / 32;
+    const int nb = (k + 255) / 256;
+    dequantize_block_q4_0<<<nb, 32, 0, stream>>>(vx, y, nb32);
+}
+
+template<typename dst_t>
+static void dequantize_row_q4_1_cuda(const void * vx, dst_t * y, const int64_t k, cudaStream_t stream) {
+    const int nb32 = k / 32;
+    const int nb = (k + 255) / 256;
+    dequantize_block_q4_1<<<nb, 32, 0, stream>>>(vx, y, nb32);
+}
+
+template<typename dst_t>
+static void dequantize_row_q4_K_cuda(const void * vx, dst_t * y, const int64_t k, cudaStream_t stream) {
+    const int nb = k / QK_K;
+    dequantize_block_q4_K<<<nb, 32, 0, stream>>>(vx, y);
+}
+
+template<typename dst_t>
+static void dequantize_row_q5_K_cuda(const void * vx, dst_t * y, const int64_t k, cudaStream_t stream) {
+    const int nb = k / QK_K;
+    dequantize_block_q5_K<<<nb, 64, 0, stream>>>(vx, y);
+}
+
+template<typename dst_t>
+static void dequantize_row_q6_K_cuda(const void * vx, dst_t * y, const int64_t k, cudaStream_t stream) {
+    const int nb = k / QK_K;
+    dequantize_block_q6_K<<<nb, 64, 0, stream>>>(vx, y);
+}
+
+template<typename dst_t>
+static void dequantize_row_iq2_xxs_cuda(const void * vx, dst_t * y, const int64_t k, cudaStream_t stream) {
+    const int nb = k / QK_K;
+    dequantize_block_iq2_xxs<<<nb, 32, 0, stream>>>(vx, y);
+}
+
+template<typename dst_t>
+static void dequantize_row_iq2_xs_cuda(const void * vx, dst_t * y, const int64_t k, cudaStream_t stream) {
+    const int nb = k / QK_K;
+    dequantize_block_iq2_xs<<<nb, 32, 0, stream>>>(vx, y);
+}
+
+template<typename dst_t>
+static void dequantize_row_iq2_s_cuda(const void * vx, dst_t * y, const int64_t k, cudaStream_t stream) {
+    const int nb = k / QK_K;
+    dequantize_block_iq2_s<<<nb, 32, 0, stream>>>(vx, y);
+}
+
+template<typename dst_t>
+static void dequantize_row_iq3_xxs_cuda(const void * vx, dst_t * y, const int64_t k, cudaStream_t stream) {
+    const int nb = k / QK_K;
+    dequantize_block_iq3_xxs<<<nb, 32, 0, stream>>>(vx, y);
+}
+
+template<typename dst_t>
+static void dequantize_row_iq3_s_cuda(const void * vx, dst_t * y, const int64_t k, cudaStream_t stream) {
+    const int nb = k / QK_K;
+    dequantize_block_iq3_s<<<nb, 32, 0, stream>>>(vx, y);
+}
+
+template<typename dst_t>
+static void dequantize_row_iq1_s_cuda(const void * vx, dst_t * y, const int64_t k, cudaStream_t stream) {
+    const int nb = k / QK_K;
+    dequantize_block_iq1_s<<<nb, 32, 0, stream>>>(vx, y);
+}
+
+template<typename dst_t>
+static void dequantize_row_iq4_nl_cuda(const void * vx, dst_t * y, const int64_t k, cudaStream_t stream) {
+    const int nb = (k + QK_K - 1) / QK_K;
+    dequantize_block_iq4_nl<<<nb, 32, 0, stream>>>(vx, y);
+}
+
+template<typename dst_t>
+static void dequantize_row_iq1_m_cuda(const void * vx, dst_t * y, const int64_t k, cudaStream_t stream) {
+    const int nb = k / QK_K;
+    dequantize_block_iq1_m<<<nb, 32, 0, stream>>>(vx, y);
+}
+
+template<typename dst_t>
+static void dequantize_row_iq4_xs_cuda(const void * vx, dst_t * y, const int64_t k, cudaStream_t stream) {
+    const int nb = (k + QK_K - 1) / QK_K;
+    dequantize_block_iq4_xs<<<nb, 32, 0, stream>>>(vx, y);
+}
+
+template <typename src_t, typename dst_t>
+static __global__ void convert_unary(const void * __restrict__ vx, dst_t * __restrict__ y, const int64_t k) {
+    const int64_t i = (int64_t)blockDim.x*blockIdx.x + threadIdx.x;
+
+    if (i >= k) {
+        return;
+    }
+
+    const src_t * x = (src_t *) vx;
+
+    y[i] = x[i];
+}
+
+template <typename src_t, typename dst_t>
+static void convert_unary_cuda(const void * __restrict__ vx, dst_t * __restrict__ y, const int64_t k, cudaStream_t stream) {
+    const int num_blocks = (k + CUDA_DEQUANTIZE_BLOCK_SIZE - 1) / CUDA_DEQUANTIZE_BLOCK_SIZE;
+    convert_unary<src_t><<<num_blocks, CUDA_DEQUANTIZE_BLOCK_SIZE, 0, stream>>>(vx, y, k);
+}
+
+to_fp16_cuda_t ggml_get_to_fp16_cuda(ggml_type type) {
+    switch (type) {
+        case GGML_TYPE_Q4_0:
+            return dequantize_row_q4_0_cuda;
+        case GGML_TYPE_Q4_1:
+            return dequantize_row_q4_1_cuda;
+        case GGML_TYPE_Q5_0:
+            return dequantize_block_cuda<QK5_0, QR5_0, dequantize_q5_0>;
+        case GGML_TYPE_Q5_1:
+            return dequantize_block_cuda<QK5_1, QR5_1, dequantize_q5_1>;
+        case GGML_TYPE_Q8_0:
+            if (ggml_cuda_info().devices[ggml_cuda_get_device()].cc >= CC_PASCAL) {
+                return dequantize_block_q8_0_f16_cuda;
+            }
+            return dequantize_block_cuda<QK8_0, QR8_0, dequantize_q8_0>;
+        case GGML_TYPE_Q2_K:
+            return dequantize_row_q2_K_cuda;
+        case GGML_TYPE_Q3_K:
+            return dequantize_row_q3_K_cuda;
+        case GGML_TYPE_Q4_K:
+            return dequantize_row_q4_K_cuda;
+        case GGML_TYPE_Q5_K:
+            return dequantize_row_q5_K_cuda;
+        case GGML_TYPE_Q6_K:
+            return dequantize_row_q6_K_cuda;
+        case GGML_TYPE_IQ2_XXS:
+            return dequantize_row_iq2_xxs_cuda;
+        case GGML_TYPE_IQ2_XS:
+            return dequantize_row_iq2_xs_cuda;
+        case GGML_TYPE_IQ2_S:
+            return dequantize_row_iq2_s_cuda;
+        case GGML_TYPE_IQ3_XXS:
+            return dequantize_row_iq3_xxs_cuda;
+        case GGML_TYPE_IQ1_S:
+            return dequantize_row_iq1_s_cuda;
+        case GGML_TYPE_IQ1_M:
+            return dequantize_row_iq1_m_cuda;
+        case GGML_TYPE_IQ4_NL:
+            return dequantize_row_iq4_nl_cuda;
+        case GGML_TYPE_IQ4_XS:
+            return dequantize_row_iq4_xs_cuda;
+        case GGML_TYPE_IQ3_S:
+            return dequantize_row_iq3_s_cuda;
+        case GGML_TYPE_F32:
+            return convert_unary_cuda<float>;
+        default:
+            return nullptr;
+    }
+}
+
+to_fp32_cuda_t ggml_get_to_fp32_cuda(ggml_type type) {
+    switch (type) {
+        case GGML_TYPE_Q4_0:
+            return dequantize_row_q4_0_cuda;
+        case GGML_TYPE_Q4_1:
+            return dequantize_row_q4_1_cuda;
+        case GGML_TYPE_Q5_0:
+            return dequantize_block_cuda<QK5_0, QR5_0, dequantize_q5_0>;
+        case GGML_TYPE_Q5_1:
+            return dequantize_block_cuda<QK5_1, QR5_1, dequantize_q5_1>;
+        case GGML_TYPE_Q8_0:
+            return dequantize_block_cuda<QK8_0, QR8_0, dequantize_q8_0>;
+        case GGML_TYPE_Q2_K:
+            return dequantize_row_q2_K_cuda;
+        case GGML_TYPE_Q3_K:
+            return dequantize_row_q3_K_cuda;
+        case GGML_TYPE_Q4_K:
+            return dequantize_row_q4_K_cuda;
+        case GGML_TYPE_Q5_K:
+            return dequantize_row_q5_K_cuda;
+        case GGML_TYPE_Q6_K:
+            return dequantize_row_q6_K_cuda;
+        case GGML_TYPE_IQ2_XXS:
+            return dequantize_row_iq2_xxs_cuda;
+        case GGML_TYPE_IQ2_XS:
+            return dequantize_row_iq2_xs_cuda;
+        case GGML_TYPE_IQ2_S:
+            return dequantize_row_iq2_s_cuda;
+        case GGML_TYPE_IQ3_XXS:
+            return dequantize_row_iq3_xxs_cuda;
+        case GGML_TYPE_IQ1_S:
+            return dequantize_row_iq1_s_cuda;
+        case GGML_TYPE_IQ1_M:
+            return dequantize_row_iq1_m_cuda;
+        case GGML_TYPE_IQ4_NL:
+            return dequantize_row_iq4_nl_cuda;
+        case GGML_TYPE_IQ4_XS:
+            return dequantize_row_iq4_xs_cuda;
+        case GGML_TYPE_IQ3_S:
+            return dequantize_row_iq3_s_cuda;
+        case GGML_TYPE_F16:
+            return convert_unary_cuda<half>;
+        default:
+            return nullptr;
+    }
+}
diff --git a/src/whisper_cpp/ggml/src/ggml-cuda/convert.cuh b/src/whisper_cpp/ggml/src/ggml-cuda/convert.cuh
new file mode 100644
index 00000000..5394be9f
--- /dev/null
+++ b/src/whisper_cpp/ggml/src/ggml-cuda/convert.cuh
@@ -0,0 +1,13 @@
+#include "common.cuh"
+
+#define CUDA_DEQUANTIZE_BLOCK_SIZE 256
+
+template<typename T>
+using to_t_cuda_t = void (*)(const void * __restrict__ x, T * __restrict__ y, int64_t k, cudaStream_t stream);
+
+typedef to_t_cuda_t<float> to_fp32_cuda_t;
+typedef to_t_cuda_t<half> to_fp16_cuda_t;
+
+to_fp16_cuda_t ggml_get_to_fp16_cuda(ggml_type type);
+
+to_fp32_cuda_t ggml_get_to_fp32_cuda(ggml_type type);
diff --git a/src/whisper_cpp/ggml/src/ggml-cuda/count-equal.cu b/src/whisper_cpp/ggml/src/ggml-cuda/count-equal.cu
new file mode 100644
index 00000000..ffb053b1
--- /dev/null
+++ b/src/whisper_cpp/ggml/src/ggml-cuda/count-equal.cu
@@ -0,0 +1,64 @@
+#include "common.cuh"
+#include "count-equal.cuh"
+
+#include <cstdint>
+
+template <typename T>
+static __global__ void count_equal(const T * __restrict__ x, const T * __restrict__ y, int64_t * __restrict__ dst, const int64_t dk, const int64_t k) {
+    const int64_t i0 = (int64_t) blockIdx.x*dk;
+    const int64_t i1 = min(i0 + dk, k);
+
+    int nequal = 0;
+
+    for (int64_t i = i0 + threadIdx.x; i < i1; i += WARP_SIZE) {
+        const T xi = x[i];
+        const T yi = y[i];
+        nequal += xi == yi;
+    }
+
+    nequal = warp_reduce_sum(nequal);
+
+    if (threadIdx.x != 0) {
+        return;
+    }
+
+    atomicAdd((int *) dst, nequal);
+}
+
+void ggml_cuda_count_equal(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
+    const ggml_tensor * src0 = dst->src[0];
+    const ggml_tensor * src1 = dst->src[1];
+
+    GGML_ASSERT(src0->type == src1->type);
+    GGML_ASSERT( dst->type == GGML_TYPE_I64);
+
+    GGML_ASSERT(ggml_are_same_shape(src0, src1));
+    GGML_ASSERT(ggml_is_contiguous(src0));
+    GGML_ASSERT(ggml_is_contiguous(src1));
+    GGML_ASSERT(ggml_is_contiguous(dst));
+
+    int64_t * dst_d  = (int64_t *) dst->data;
+
+    cudaStream_t stream = ctx.stream();
+    const int nsm = ggml_cuda_info().devices[ggml_cuda_get_device()].nsm;
+
+    const int64_t ne = ggml_nelements(src0);
+    GGML_ASSERT(ne < (1 << 30) && "atomicAdd implementation only supports int");
+    const int64_t dne = GGML_PAD(ne / (4*nsm), CUDA_COUNT_EQUAL_CHUNK_SIZE);
+
+    CUDA_CHECK(cudaMemsetAsync(dst_d, 0, ggml_nbytes(dst), stream));
+
+    const dim3 blocks_dim(WARP_SIZE, 1, 1);
+    const dim3 blocks_num(std::min((int64_t)4*nsm, (ne + CUDA_COUNT_EQUAL_CHUNK_SIZE - 1)/CUDA_COUNT_EQUAL_CHUNK_SIZE), 1, 1);
+
+    switch (src0->type) {
+        case GGML_TYPE_I32: {
+            const int * src0_d = (const int *) src0->data;
+            const int * src1_d = (const int *) src1->data;
+            count_equal<<<blocks_num, blocks_dim, 0, stream>>>(src0_d, src1_d, dst_d, dne, ne);
+        } break;
+        default:
+            GGML_ASSERT(false);
+            break;
+    }
+}
diff --git a/src/whisper_cpp/ggml/src/ggml-cuda/count-equal.cuh b/src/whisper_cpp/ggml/src/ggml-cuda/count-equal.cuh
new file mode 100644
index 00000000..8467da79
--- /dev/null
+++ b/src/whisper_cpp/ggml/src/ggml-cuda/count-equal.cuh
@@ -0,0 +1,5 @@
+#include "common.cuh"
+
+#define CUDA_COUNT_EQUAL_CHUNK_SIZE 128
+
+void ggml_cuda_count_equal(ggml_backend_cuda_context & ctx, ggml_tensor * dst);
diff --git a/src/whisper_cpp/ggml/src/ggml-cuda/cpy.cu b/src/whisper_cpp/ggml/src/ggml-cuda/cpy.cu
new file mode 100644
index 00000000..54c0f66d
--- /dev/null
+++ b/src/whisper_cpp/ggml/src/ggml-cuda/cpy.cu
@@ -0,0 +1,543 @@
+#include "cpy.cuh"
+
+typedef void (*cpy_kernel_t)(const char * cx, char * cdst);
+
+static __device__ void cpy_1_f32_f32(const char * cxi, char * cdsti) {
+    const float * xi = (const float *) cxi;
+    float * dsti = (float *) cdsti;
+
+    *dsti = *xi;
+}
+
+static __device__ void cpy_1_f32_f16(const char * cxi, char * cdsti) {
+    const float * xi = (const float *) cxi;
+    half * dsti = (half *) cdsti;
+
+    *dsti = __float2half(*xi);
+}
+
+static __device__ void cpy_1_f16_f16(const char * cxi, char * cdsti) {
+    const half * xi = (const half *) cxi;
+    half * dsti = (half *) cdsti;
+
+    *dsti = *xi;
+}
+
+static __device__ void cpy_1_f16_f32(const char * cxi, char * cdsti) {
+    const half * xi = (const half *) cxi;
+    float * dsti = (float *) cdsti;
+
+    *dsti = *xi;
+}
+
+template <cpy_kernel_t cpy_1>
+static __global__ void cpy_f32_f16(const char * cx, char * cdst, const int ne,
+                                   const int ne00, const int ne01, const int ne02, const int nb00, const int nb01, const int nb02,
+                                   const int nb03, const int ne10, const int ne11, const int ne12, const int nb10, const int nb11,
+                                   const int nb12, const int nb13) {
+    const int64_t i = blockDim.x*blockIdx.x + threadIdx.x;
+
+    if (i >= ne) {
+        return;
+    }
+
+    // determine indices i03/i13, i02/i12, i01/i11, i00/i10 as a function of index i of flattened tensor
+    // then combine those indices with the corresponding byte offsets to get the total offsets
+    const int64_t i03 = i/(ne00 * ne01 * ne02);
+    const int64_t i02 = (i - i03*ne00*ne01*ne02 )/ (ne00*ne01);
+    const int64_t i01 = (i - i03*ne00*ne01*ne02  -  i02*ne01*ne00) / ne00;
+    const int64_t i00 = i - i03*ne00*ne01*ne02 - i02*ne01*ne00 - i01*ne00;
+    const int64_t x_offset = i00*nb00 + i01*nb01 + i02*nb02 + i03 * nb03;
+
+    const int64_t i13 = i/(ne10 * ne11 * ne12);
+    const int64_t i12 = (i - i13*ne10*ne11*ne12) / (ne10*ne11);
+    const int64_t i11 = (i - i13*ne10*ne11*ne12 - i12*ne10*ne11) / ne10;
+    const int64_t i10 = i - i13*ne10*ne11*ne12 - i12*ne10*ne11 - i11*ne10;
+    const int64_t dst_offset = i10*nb10 + i11*nb11 + i12*nb12 + i13 * nb13;
+
+    cpy_1(cx + x_offset, cdst + dst_offset);
+}
+
+static __device__ void cpy_blck_f32_q8_0(const char * cxi, char * cdsti) {
+    const float * xi = (const float *) cxi;
+    block_q8_0 * dsti = (block_q8_0 *) cdsti;
+
+    float amax = 0.0f; // absolute max
+
+    for (int j = 0; j < QK8_0; j++) {
+        const float v = xi[j];
+        amax = fmaxf(amax, fabsf(v));
+    }
+
+    const float d = amax / ((1 << 7) - 1);
+    const float id = d ? 1.0f/d : 0.0f;
+
+    dsti->d = d;
+
+    for (int j = 0; j < QK8_0; ++j) {
+        const float x0 = xi[j]*id;
+
+        dsti->qs[j] = roundf(x0);
+    }
+}
+
+static __device__ void cpy_blck_q8_0_f32(const char * cxi, char * cdsti) {
+    const block_q8_0 * xi = (const block_q8_0 *) cxi;
+    float * dsti = (float *) cdsti;
+
+    const float d = (float)xi->d;
+
+    for (int j = 0; j < QK8_0; j++) {
+       dsti[j] = xi->qs[j] * d;
+    }
+}
+
+static __device__ void cpy_blck_f32_q4_0(const char * cxi, char * cdsti) {
+    const float * xi = (const float *) cxi;
+    block_q4_0 * dsti = (block_q4_0 *) cdsti;
+
+    float amax = 0.0f;
+    float vmax = 0.0f;
+
+    for (int j = 0; j < QK4_0; ++j) {
+        const float v = xi[j];
+        if (amax < fabsf(v)) {
+            amax = fabsf(v);
+            vmax = v;
+        }
+    }
+
+    const float d  = vmax / -8;
+    const float id = d ? 1.0f/d : 0.0f;
+
+    dsti->d = d;
+
+    for (int j = 0; j < QK4_0/2; ++j) {
+        const float x0 = xi[0       + j]*id;
+        const float x1 = xi[QK4_0/2 + j]*id;
+
+        const uint8_t xi0 = min(15, (int8_t)(x0 + 8.5f));
+        const uint8_t xi1 = min(15, (int8_t)(x1 + 8.5f));
+
+        dsti->qs[j]  = xi0;
+        dsti->qs[j] |= xi1 << 4;
+    }
+}
+
+static __device__ void cpy_blck_f32_q4_1(const char * cxi, char * cdsti) {
+    const float * xi = (const float *) cxi;
+    block_q4_1 * dsti = (block_q4_1 *) cdsti;
+
+    float vmin = FLT_MAX;
+    float vmax = -FLT_MAX;
+
+    for (int j = 0; j < QK4_1; ++j) {
+        const float v = xi[j];
+
+        if (v < vmin) vmin = v;
+        if (v > vmax) vmax = v;
+    }
+
+    const float d  = (vmax - vmin) / ((1 << 4) - 1);
+    const float id = d ? 1.0f/d : 0.0f;
+
+    dsti->dm.x = d;
+    dsti->dm.y = vmin;
+
+    for (int j = 0; j < QK4_1/2; ++j) {
+        const float x0 = (xi[0       + j] - vmin)*id;
+        const float x1 = (xi[QK4_1/2 + j] - vmin)*id;
+
+        const uint8_t xi0 = min(15, (int8_t)(x0 + 0.5f));
+        const uint8_t xi1 = min(15, (int8_t)(x1 + 0.5f));
+
+        dsti->qs[j]  = xi0;
+        dsti->qs[j] |= xi1 << 4;
+    }
+}
+
+static __device__ void cpy_blck_f32_q5_0(const char * cxi, char * cdsti) {
+    const float * xi = (const float *) cxi;
+    block_q5_0 * dsti = (block_q5_0 *) cdsti;
+
+    float amax = 0.0f;
+    float vmax = 0.0f;
+
+    for (int j = 0; j < QK5_0; ++j) {
+        const float v = xi[j];
+        if (amax < fabsf(v)) {
+            amax = fabsf(v);
+            vmax = v;
+        }
+    }
+
+    const float d  = vmax / -16;
+    const float id = d ? 1.0f/d : 0.0f;
+
+    dsti->d = d;
+
+    uint32_t qh = 0;
+    for (int j = 0; j < QK5_0/2; ++j) {
+        const float x0 = xi[0       + j]*id;
+        const float x1 = xi[QK5_0/2 + j]*id;
+
+        const uint8_t xi0 = min(31, (int8_t)(x0 + 16.5f));
+        const uint8_t xi1 = min(31, (int8_t)(x1 + 16.5f));
+
+        dsti->qs[j]  = (xi0 & 0xf) | ((xi1 & 0xf) << 4);
+        qh |= ((xi0 & 0x10u) >> 4) << (j + 0);
+        qh |= ((xi1 & 0x10u) >> 4) << (j + QK5_0/2);
+    }
+    memcpy(dsti->qh, &qh, sizeof(qh));
+}
+
+static __device__ void cpy_blck_f32_q5_1(const char * cxi, char * cdsti) {
+    const float * xi = (const float *) cxi;
+    block_q5_1 * dsti = (block_q5_1 *) cdsti;
+
+    float min = xi[0];
+    float max = xi[0];
+
+    for (int j = 1; j < QK5_1; ++j) {
+        const float v = xi[j];
+        min = v < min ? v : min;
+        max = v > max ? v : max;
+    }
+
+    const float d  = (max - min) / 31;
+    const float id = d ? 1.0f/d : 0.0f;
+
+    dsti->dm.x = d;
+    dsti->dm.y = min;
+
+    uint32_t qh = 0;
+    for (int j = 0; j < QK5_1/2; ++j) {
+        const float x0 = (xi[0       + j] - min)*id;
+        const float x1 = (xi[QK5_1/2 + j] - min)*id;
+
+        const uint8_t xi0 = (uint8_t)(x0 + 0.5f);
+        const uint8_t xi1 = (uint8_t)(x1 + 0.5f);
+
+        dsti->qs[j]  = (xi0 & 0xf) | ((xi1 & 0xf) << 4);
+        qh |= ((xi0 & 0x10u) >> 4) << (j + 0);
+        qh |= ((xi1 & 0x10u) >> 4) << (j + QK5_1/2);
+    }
+    memcpy(dsti->qh, &qh, sizeof(qh));
+}
+
+
+static __device__ __forceinline__ int best_index_int8(int n, const int8_t * val, float x) {
+    if (x <= val[0]) return 0;
+    if (x >= val[n-1]) return n-1;
+    int ml = 0, mu = n-1;
+    while (mu-ml > 1) {
+        int mav = (ml+mu)/2;
+        if (x < val[mav]) mu = mav; else ml = mav;
+    }
+    return x - val[mu-1] < val[mu] - x ? mu-1 : mu;
+}
+
+static __device__ void cpy_blck_f32_iq4_nl(const char * cxi, char * cdsti) {
+    const float * xi = (const float *) cxi;
+    block_iq4_nl * dsti = (block_iq4_nl *) cdsti;
+
+    float amax = 0.0f;
+    float vmax = 0.0f;
+
+    for (int j = 0; j < QK4_NL; ++j) {
+        const float v = xi[j];
+        if (amax < fabsf(v)) {
+            amax = fabsf(v);
+            vmax = v;
+        }
+    }
+
+    float d = vmax / kvalues_iq4nl[0];
+    const float id = d ? 1.0f/d : 0.0f;
+
+    float sumqx = 0, sumq2 = 0;
+    for (int j = 0; j < QK4_NL/2; ++j) {
+        const float x0 = xi[0        + j]*id;
+        const float x1 = xi[QK4_NL/2 + j]*id;
+        const uint8_t xi0 = best_index_int8(16, kvalues_iq4nl, x0);
+        const uint8_t xi1 = best_index_int8(16, kvalues_iq4nl, x1);
+        dsti->qs[j] = xi0 | (xi1 << 4);
+        const float v0 = kvalues_iq4nl[xi0];
+        const float v1 = kvalues_iq4nl[xi1];
+        const float w0 = xi[0        + j]*xi[0        + j];
+        const float w1 = xi[QK4_NL/2 + j]*xi[QK4_NL/2 + j];
+        sumqx += w0*v0*xi[j] + w1*v1*xi[QK4_NL/2 + j];
+        sumq2 += w0*v0*v0 + w1*v1*v1;
+    }
+
+    dsti->d = sumq2 > 0 ? sumqx/sumq2 : d;
+}
+
+template <cpy_kernel_t cpy_blck, int qk>
+static __global__ void cpy_f32_q(const char * cx, char * cdst, const int ne,
+                                 const int ne00, const int ne01, const int ne02, const int nb00, const int nb01, const int nb02,
+                                 const int nb03, const int ne10, const int ne11, const int ne12, const int nb10, const int nb11,
+                                 const int nb12, const int nb13) {
+    const int i = (blockDim.x*blockIdx.x + threadIdx.x)*qk;
+
+    if (i >= ne) {
+        return;
+    }
+
+    const int i03 = i/(ne00 * ne01 * ne02);
+    const int i02 = (i - i03*ne00*ne01*ne02 )/ (ne00*ne01);
+    const int i01 = (i - i03*ne00*ne01*ne02  -  i02*ne01*ne00) / ne00;
+    const int i00 = i - i03*ne00*ne01*ne02 - i02*ne01*ne00 - i01*ne00;
+    const int x_offset = i00*nb00 + i01*nb01 + i02*nb02 + i03 * nb03;
+
+    const int i13 = i/(ne10 * ne11 * ne12);
+    const int i12 = (i - i13*ne10*ne11*ne12) / (ne10*ne11);
+    const int i11 = (i - i13*ne10*ne11*ne12 - i12*ne10*ne11) / ne10;
+    const int i10 = i - i13*ne10*ne11*ne12 - i12*ne10*ne11 - i11*ne10;
+    const int dst_offset = (i10/qk)*nb10 + i11*nb11 + i12*nb12 + i13*nb13;
+
+    cpy_blck(cx + x_offset, cdst + dst_offset);
+}
+
+template <cpy_kernel_t cpy_blck, int qk>
+static __global__ void cpy_q_f32(const char * cx, char * cdst, const int ne,
+                                 const int ne00, const int ne01, const int ne02, const int nb00, const int nb01, const int nb02,
+                                 const int nb03, const int ne10, const int ne11, const int ne12, const int nb10, const int nb11,
+                                 const int nb12, const int nb13) {
+    const int i = (blockDim.x*blockIdx.x + threadIdx.x)*qk;
+
+    if (i >= ne) {
+        return;
+    }
+
+    const int i03 = i/(ne00 * ne01 * ne02);
+    const int i02 = (i - i03*ne00*ne01*ne02 )/ (ne00*ne01);
+    const int i01 = (i - i03*ne00*ne01*ne02  -  i02*ne01*ne00) / ne00;
+    const int i00 = i - i03*ne00*ne01*ne02 - i02*ne01*ne00 - i01*ne00;
+    const int x_offset = (i00/qk)*nb00 + i01*nb01 + i02*nb02 + i03 * nb03;
+
+    const int i13 = i/(ne10 * ne11 * ne12);
+    const int i12 = (i - i13*ne10*ne11*ne12) / (ne10*ne11);
+    const int i11 = (i - i13*ne10*ne11*ne12 - i12*ne10*ne11) / ne10;
+    const int i10 = i - i13*ne10*ne11*ne12 - i12*ne10*ne11 - i11*ne10;
+    const int dst_offset = i10*nb10 + i11*nb11 + i12*nb12 + i13*nb13;
+
+    cpy_blck(cx + x_offset, cdst + dst_offset);
+}
+
+static void ggml_cpy_f16_f32_cuda(
+    const char * cx, char * cdst, const int ne,
+    const int ne00, const int ne01, const int ne02, const int nb00, const int nb01, const int nb02,
+    const int nb03, const int ne10, const int ne11, const int ne12, const int nb10, const int nb11, const int nb12, const int nb13, cudaStream_t stream) {
+
+    const int num_blocks = (ne + CUDA_CPY_BLOCK_SIZE - 1) / CUDA_CPY_BLOCK_SIZE;
+    cpy_f32_f16<cpy_1_f16_f32><<<num_blocks, CUDA_CPY_BLOCK_SIZE, 0, stream>>>
+        (cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13);
+}
+
+static void ggml_cpy_f32_f32_cuda(
+    const char * cx, char * cdst, const int ne,
+    const int ne00, const int ne01, const int ne02, const int nb00, const int nb01, const int nb02,
+    const int nb03, const int ne10, const int ne11, const int ne12, const int nb10, const int nb11, const int nb12, const int nb13, cudaStream_t stream) {
+
+    const int num_blocks = (ne + CUDA_CPY_BLOCK_SIZE - 1) / CUDA_CPY_BLOCK_SIZE;
+    cpy_f32_f16<cpy_1_f32_f32><<<num_blocks, CUDA_CPY_BLOCK_SIZE, 0, stream>>>
+        (cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13);
+}
+
+static void ggml_cpy_f32_f16_cuda(
+    const char * cx, char * cdst, const int ne,
+    const int ne00, const int ne01, const int ne02, const int nb00, const int nb01, const int nb02,
+    const int nb03, const int ne10, const int ne11, const int ne12, const int nb10, const int nb11, const int nb12, const int nb13, cudaStream_t stream) {
+
+    const int num_blocks = (ne + CUDA_CPY_BLOCK_SIZE - 1) / CUDA_CPY_BLOCK_SIZE;
+    cpy_f32_f16<cpy_1_f32_f16><<<num_blocks, CUDA_CPY_BLOCK_SIZE, 0, stream>>>
+        (cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13);
+}
+
+static void ggml_cpy_f32_q8_0_cuda(
+    const char * cx, char * cdst, const int ne,
+    const int ne00, const int ne01, const int ne02, const int nb00, const int nb01, const int nb02,
+    const int nb03, const int ne10, const int ne11, const int ne12, const int nb10, const int nb11, const int nb12, const int nb13, cudaStream_t stream) {
+
+    GGML_ASSERT(ne % QK8_0 == 0);
+    const int num_blocks = ne / QK8_0;
+    cpy_f32_q<cpy_blck_f32_q8_0, QK8_0><<<num_blocks, 1, 0, stream>>>
+        (cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13);
+}
+
+static void ggml_cpy_q8_0_f32_cuda(
+    const char * cx, char * cdst, const int ne,
+    const int ne00, const int ne01, const int ne02, const int nb00, const int nb01, const int nb02,
+    const int nb03, const int ne10, const int ne11, const int ne12, const int nb10, const int nb11, const int nb12, const int nb13, cudaStream_t stream) {
+
+    const int num_blocks = ne;
+    cpy_q_f32<cpy_blck_q8_0_f32, QK8_0><<<num_blocks, 1, 0, stream>>>
+        (cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13);
+}
+
+static void ggml_cpy_f32_q4_0_cuda(
+    const char * cx, char * cdst, const int ne,
+    const int ne00, const int ne01, const int ne02, const int nb00, const int nb01, const int nb02,
+    const int nb03, const int ne10, const int ne11, const int ne12, const int nb10, const int nb11, const int nb12, const int nb13, cudaStream_t stream) {
+
+    GGML_ASSERT(ne % QK4_0 == 0);
+    const int num_blocks = ne / QK4_0;
+    cpy_f32_q<cpy_blck_f32_q4_0, QK4_0><<<num_blocks, 1, 0, stream>>>
+        (cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13);
+}
+
+static void ggml_cpy_f32_q4_1_cuda(
+    const char * cx, char * cdst, const int ne,
+    const int ne00, const int ne01, const int ne02, const int nb00, const int nb01, const int nb02,
+    const int nb03, const int ne10, const int ne11, const int ne12, const int nb10, const int nb11, const int nb12, const int nb13, cudaStream_t stream) {
+
+    GGML_ASSERT(ne % QK4_1 == 0);
+    const int num_blocks = ne / QK4_1;
+    cpy_f32_q<cpy_blck_f32_q4_1, QK4_1><<<num_blocks, 1, 0, stream>>>
+        (cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13);
+}
+
+static void ggml_cpy_f32_q5_0_cuda(
+    const char * cx, char * cdst, const int ne,
+    const int ne00, const int ne01, const int ne02, const int nb00, const int nb01, const int nb02,
+    const int nb03, const int ne10, const int ne11, const int ne12, const int nb10, const int nb11, const int nb12, const int nb13, cudaStream_t stream) {
+
+    GGML_ASSERT(ne % QK5_0 == 0);
+    const int num_blocks = ne / QK5_0;
+    cpy_f32_q<cpy_blck_f32_q5_0, QK5_0><<<num_blocks, 1, 0, stream>>>
+        (cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13);
+}
+
+static void ggml_cpy_f32_q5_1_cuda(
+    const char * cx, char * cdst, const int ne,
+    const int ne00, const int ne01, const int ne02, const int nb00, const int nb01, const int nb02,
+    const int nb03, const int ne10, const int ne11, const int ne12, const int nb10, const int nb11, const int nb12, const int nb13, cudaStream_t stream) {
+
+    GGML_ASSERT(ne % QK5_1 == 0);
+    const int num_blocks = ne / QK5_1;
+    cpy_f32_q<cpy_blck_f32_q5_1, QK5_1><<<num_blocks, 1, 0, stream>>>
+        (cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13);
+}
+
+static void ggml_cpy_f32_iq4_nl_cuda(
+    const char * cx, char * cdst, const int ne,
+    const int ne00, const int ne01, const int ne02, const int nb00, const int nb01, const int nb02,
+    const int nb03, const int ne10, const int ne11, const int ne12, const int nb10, const int nb11, const int nb12, const int nb13, cudaStream_t stream) {
+
+    GGML_ASSERT(ne % QK4_NL == 0);
+    const int num_blocks = ne / QK4_NL;
+    cpy_f32_q<cpy_blck_f32_iq4_nl, QK4_NL><<<num_blocks, 1, 0, stream>>>
+        (cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13);
+}
+
+static void ggml_cpy_f16_f16_cuda(
+    const char * cx, char * cdst, const int ne,
+    const int ne00, const int ne01, const int ne02, const int nb00, const int nb01, const int nb02,
+    const int nb03, const int ne10, const int ne11, const int ne12, const int nb10, const int nb11, const int nb12, const int nb13, cudaStream_t stream) {
+
+    const int num_blocks = (ne + CUDA_CPY_BLOCK_SIZE - 1) / CUDA_CPY_BLOCK_SIZE;
+    cpy_f32_f16<cpy_1_f16_f16><<<num_blocks, CUDA_CPY_BLOCK_SIZE, 0, stream>>>
+        (cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13);
+}
+
+void ggml_cuda_cpy(ggml_backend_cuda_context & ctx, const ggml_tensor * src0, ggml_tensor * src1) {
+    const int64_t ne = ggml_nelements(src0);
+    GGML_ASSERT(ne == ggml_nelements(src1));
+
+    GGML_ASSERT(ggml_nbytes(src0) <= INT_MAX);
+    GGML_ASSERT(ggml_nbytes(src1) <= INT_MAX);
+
+    const int64_t ne00 = src0->ne[0];
+    const int64_t ne01 = src0->ne[1];
+    const int64_t ne02 = src0->ne[2];
+
+    //GGML_ASSERT(src0->ne[3] == 1);
+
+    const int64_t nb00 = src0->nb[0];
+    const int64_t nb01 = src0->nb[1];
+    const int64_t nb02 = src0->nb[2];
+    const int64_t nb03 = src0->nb[3];
+
+    const int64_t ne10 = src1->ne[0];
+    const int64_t ne11 = src1->ne[1];
+    const int64_t ne12 = src1->ne[2];
+
+    //GGML_ASSERT(src1->ne[3] == 1);
+
+    const int64_t nb10 = src1->nb[0];
+    const int64_t nb11 = src1->nb[1];
+    const int64_t nb12 = src1->nb[2];
+    const int64_t nb13 = src1->nb[3];
+
+    cudaStream_t main_stream = ctx.stream();
+
+    char * src0_ddc = (char *) src0->data;
+    char * src1_ddc = (char *) src1->data;
+
+    if (src0->type == src1->type && ggml_is_contiguous(src0) && ggml_is_contiguous(src1)) {
+        GGML_ASSERT(ggml_nbytes(src0) == ggml_nbytes(src1));
+        CUDA_CHECK(cudaMemcpyAsync(src1_ddc, src0_ddc, ggml_nbytes(src0), cudaMemcpyDeviceToDevice, main_stream));
+    } else if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_F32) {
+        ggml_cpy_f32_f32_cuda (src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream);
+    } else if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_F16) {
+        ggml_cpy_f32_f16_cuda (src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream);
+    } else if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_Q8_0) {
+        ggml_cpy_f32_q8_0_cuda(src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream);
+    } else if (src0->type == GGML_TYPE_Q8_0 && src1->type == GGML_TYPE_F32) {
+        ggml_cpy_q8_0_f32_cuda(src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream);
+    } else if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_Q4_0) {
+        ggml_cpy_f32_q4_0_cuda(src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream);
+    } else if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_Q4_1) {
+        ggml_cpy_f32_q4_1_cuda(src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream);
+    } else if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_Q5_0) {
+        ggml_cpy_f32_q5_0_cuda(src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream);
+    } else if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_IQ4_NL) {
+        ggml_cpy_f32_iq4_nl_cuda(src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream);
+    } else if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_Q5_1) {
+        ggml_cpy_f32_q5_1_cuda(src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream);
+    } else if (src0->type == GGML_TYPE_F16 && src1->type == GGML_TYPE_F16) {
+        ggml_cpy_f16_f16_cuda (src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream);
+    } else if (src0->type == GGML_TYPE_F16 && src1->type == GGML_TYPE_F32) {
+        ggml_cpy_f16_f32_cuda (src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream);
+    } else {
+        GGML_ABORT("%s: unsupported type combination (%s to %s)\n", __func__,
+                ggml_type_name(src0->type), ggml_type_name(src1->type));
+    }
+}
+
+void ggml_cuda_dup(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
+    const ggml_tensor * src0 = dst->src[0];
+    ggml_cuda_cpy(ctx, src0, dst);
+}
+
+void* ggml_cuda_cpy_fn(const ggml_tensor * src0, ggml_tensor * src1) {
+    if (src0->type == src1->type && ggml_is_contiguous(src0) && ggml_is_contiguous(src1)) {
+        return nullptr;
+    } else if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_F32) {
+        return (void*) cpy_f32_f16<cpy_1_f32_f32>;
+    } else if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_F16) {
+        return (void*) cpy_f32_f16<cpy_1_f32_f16>;
+    } else if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_Q8_0) {
+        return (void*) cpy_f32_q<cpy_blck_f32_q8_0, QK8_0>;
+    } else if (src0->type == GGML_TYPE_Q8_0 && src1->type == GGML_TYPE_F32) {
+        return (void*) cpy_q_f32<cpy_blck_q8_0_f32, QK8_0>;
+    } else if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_Q4_0) {
+        return (void*) cpy_f32_q<cpy_blck_f32_q4_0, QK4_0>;
+    } else if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_Q4_1) {
+        return (void*) cpy_f32_q<cpy_blck_f32_q4_1, QK4_1>;
+    } else if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_Q5_0) {
+        return (void*) cpy_f32_q<cpy_blck_f32_q5_0, QK5_0>;
+    } else if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_IQ4_NL) {
+        return (void*) cpy_f32_q<cpy_blck_f32_iq4_nl, QK4_NL>;
+    } else if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_Q5_1) {
+        return (void*) cpy_f32_q<cpy_blck_f32_q5_1, QK5_1>;
+    } else if (src0->type == GGML_TYPE_F16 && src1->type == GGML_TYPE_F16) {
+        return (void*) cpy_f32_f16<cpy_1_f32_f16>;
+    } else if (src0->type == GGML_TYPE_F16 && src1->type == GGML_TYPE_F32) {
+        return (void*) cpy_f32_f16<cpy_1_f16_f32>;
+    } else {
+        GGML_ABORT("%s: unsupported type combination (%s to %s)\n", __func__,
+                ggml_type_name(src0->type), ggml_type_name(src1->type));
+    }
+}
diff --git a/src/whisper_cpp/ggml/src/ggml-cuda/cpy.cuh b/src/whisper_cpp/ggml/src/ggml-cuda/cpy.cuh
new file mode 100644
index 00000000..79616742
--- /dev/null
+++ b/src/whisper_cpp/ggml/src/ggml-cuda/cpy.cuh
@@ -0,0 +1,9 @@
+#include "common.cuh"
+
+#define CUDA_CPY_BLOCK_SIZE 32
+
+void ggml_cuda_cpy(ggml_backend_cuda_context & ctx, const ggml_tensor * src0, ggml_tensor * src1);
+
+void ggml_cuda_dup(ggml_backend_cuda_context & ctx, ggml_tensor * dst);
+
+void* ggml_cuda_cpy_fn(const ggml_tensor * src0, ggml_tensor * src1);
diff --git a/src/whisper_cpp/ggml/src/ggml-cuda/cross-entropy-loss.cu b/src/whisper_cpp/ggml/src/ggml-cuda/cross-entropy-loss.cu
new file mode 100644
index 00000000..ed09406a
--- /dev/null
+++ b/src/whisper_cpp/ggml/src/ggml-cuda/cross-entropy-loss.cu
@@ -0,0 +1,166 @@
+#include "common.cuh"
+#include "cross-entropy-loss.cuh"
+#include "sum.cuh"
+
+#include <cmath>
+#include <cstdint>
+
+static __global__ void cross_entropy_loss_f32(const float * logits, const float * labels, float * dst, const int nclasses, const int k) {
+    const int warp_id = threadIdx.x / WARP_SIZE;
+    const int lane_id = threadIdx.x % WARP_SIZE;
+    const int i0 = blockDim.x*blockIdx.x + warp_id*WARP_SIZE;
+
+    const int ne_tmp = WARP_SIZE*nclasses;
+
+    extern __shared__ float tmp_all[];
+    float * tmp_logits = tmp_all + (2*warp_id + 0)*ne_tmp;
+    float * tmp_labels = tmp_all + (2*warp_id + 1)*ne_tmp;
+
+    // Each warp first loads ne_tmp logits/labels into shared memory:
+    for (int i = lane_id; i < ne_tmp; i += WARP_SIZE) {
+        const int ig = i0*nclasses + i; // ig == i global
+
+        tmp_logits[i] = ig < k*nclasses ? logits[ig] : 0.0f;
+        tmp_labels[i] = ig < k*nclasses ? labels[ig] : 0.0f;
+    }
+
+    // Each thread in the warp then calculates the cross entropy loss for a single row.
+    // TODO: pad in order to avoid shared memory bank conflicts.
+
+    // Find maximum for softmax:
+    float max = -INFINITY;
+    for (int i = 0; i < nclasses; ++i) {
+        max = fmaxf(max, tmp_logits[lane_id*nclasses + i]);
+    }
+
+    // Calculate log(softmax(logits)) which is just logits - max:
+    float sum = 0.0f;
+    for (int i = 0; i < nclasses; ++i) {
+        float val = tmp_logits[lane_id*nclasses + i] - max;
+        sum += expf(val);
+        tmp_logits[lane_id*nclasses + i] = val;
+    }
+    sum = logf(sum);
+
+    // log(exp(logits - max) / sum) = (logits - max) - log(sum)
+    float loss = 0.0f;
+    for (int i = 0; i < nclasses; ++i) {
+        loss += (tmp_logits[lane_id*nclasses + i] - sum) * tmp_labels[lane_id*nclasses + i];
+    }
+    loss = -warp_reduce_sum(loss) / (float)k;
+
+    __syncthreads();
+
+    if (lane_id == 0) {
+        tmp_all[warp_id] = loss;
+    }
+
+    __syncthreads();
+
+    if (warp_id != 0) {
+        return;
+    }
+
+    loss = lane_id < CUDA_CROSS_ENTROPY_LOSS_BLOCK_SIZE/WARP_SIZE ? tmp_all[lane_id] : 0.0f;
+    loss = warp_reduce_sum(loss);
+
+    if (lane_id != 0) {
+        return;
+    }
+
+    dst[blockIdx.x] = loss;
+}
+
+static __global__ void cross_entropy_loss_back_f32(const float * logits, const float * labels, const float * loss, float * dst, const int nclasses) {
+    extern __shared__ float tmp[];
+
+    float maxval = -INFINITY;
+    for (int i = threadIdx.x; i < nclasses; i += WARP_SIZE) {
+        const float val = logits[blockIdx.x*nclasses + i];
+        maxval = fmaxf(maxval, val);
+        tmp[i] = val;
+    }
+    maxval = warp_reduce_max(maxval);
+
+    float sum = 0.0f;
+    for (int i = threadIdx.x; i < nclasses; i += WARP_SIZE) {
+        const float val = expf(tmp[i] - maxval);
+        sum += val;
+        tmp[i] = val;
+    }
+    sum = warp_reduce_sum(sum);
+    const float sm_scale = 1.0f/sum;
+
+    const float d_by_nrows = *loss/gridDim.x;
+    for (int i = threadIdx.x; i < nclasses; i += WARP_SIZE) {
+        dst[blockIdx.x*nclasses + i] = (tmp[i]*sm_scale - labels[blockIdx.x*nclasses + i])*d_by_nrows;
+    }
+}
+
+void ggml_cuda_cross_entropy_loss(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
+    const ggml_tensor * src0 = dst->src[0];
+    const ggml_tensor * src1 = dst->src[1];
+
+    GGML_ASSERT(src0->type == GGML_TYPE_F32);
+    GGML_ASSERT(src1->type == GGML_TYPE_F32);
+    GGML_ASSERT( dst->type == GGML_TYPE_F32);
+
+    GGML_ASSERT(ggml_is_contiguous(src0));
+    GGML_ASSERT(ggml_is_contiguous(src1));
+    GGML_ASSERT(ggml_is_contiguous(dst));
+
+    const int64_t ne00  = src0->ne[0];
+    const int64_t nrows = ggml_nrows(src0);
+
+    const float * src0_d = (const float *) src0->data;
+    const float * src1_d = (const float *) src1->data;
+    float       * dst_d  = (float       *) dst->data;
+
+    ggml_cuda_pool & pool = ctx.pool();
+    cudaStream_t stream = ctx.stream();
+
+    const dim3 blocks_dim(CUDA_CROSS_ENTROPY_LOSS_BLOCK_SIZE, 1, 1);
+    const dim3 blocks_num((nrows + CUDA_CROSS_ENTROPY_LOSS_BLOCK_SIZE - 1) / CUDA_CROSS_ENTROPY_LOSS_BLOCK_SIZE, 1, 1);
+    const int shmem = 2*CUDA_CROSS_ENTROPY_LOSS_BLOCK_SIZE*ne00*sizeof(float);
+
+    ggml_cuda_pool_alloc<float> dst_tmp(pool, blocks_num.x);
+
+    cross_entropy_loss_f32<<<blocks_num, blocks_dim, shmem, stream>>>(src0_d, src1_d, dst_tmp.ptr, ne00, nrows);
+
+    // Combine results from individual blocks:
+    sum_f32_cuda(pool, dst_tmp.ptr, dst_d, blocks_num.x, stream);
+}
+
+void ggml_cuda_cross_entropy_loss_back(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
+    const ggml_tensor * src0 = dst->src[0];
+    const ggml_tensor * src1 = dst->src[1];
+    const ggml_tensor * opt0 = dst->src[2];
+
+    GGML_ASSERT(src0->type == GGML_TYPE_F32);
+    GGML_ASSERT(src1->type == GGML_TYPE_F32);
+    GGML_ASSERT(opt0->type == GGML_TYPE_F32);
+    GGML_ASSERT( dst->type == GGML_TYPE_F32);
+
+    GGML_ASSERT(ggml_is_contiguous(src0));
+    GGML_ASSERT(ggml_is_contiguous(src1));
+    GGML_ASSERT(ggml_is_contiguous(opt0));
+    GGML_ASSERT(ggml_is_contiguous(dst));
+    GGML_ASSERT(ggml_are_same_shape(src0, src1));
+    GGML_ASSERT(ggml_are_same_shape(src0, dst));
+
+    const int64_t ne00  = src0->ne[0];
+    const int64_t nrows = ggml_nrows(src0);
+
+    const float * src0_d = (const float *) src0->data;
+    const float * src1_d = (const float *) src1->data;
+    const float * opt0_d = (const float *) opt0->data;
+    float       * dst_d  = (float       *) dst->data;
+
+    cudaStream_t stream = ctx.stream();
+
+    const dim3 blocks_dim(WARP_SIZE, 1, 1);
+    const dim3 blocks_num(nrows, 1, 1);
+    const int shmem = ne00*sizeof(float);
+
+    cross_entropy_loss_back_f32<<<blocks_num, blocks_dim, shmem, stream>>>(src0_d, src1_d, opt0_d, dst_d, ne00);
+}
diff --git a/src/whisper_cpp/ggml/src/ggml-cuda/cross-entropy-loss.cuh b/src/whisper_cpp/ggml/src/ggml-cuda/cross-entropy-loss.cuh
new file mode 100644
index 00000000..9ec7152f
--- /dev/null
+++ b/src/whisper_cpp/ggml/src/ggml-cuda/cross-entropy-loss.cuh
@@ -0,0 +1,7 @@
+#include "common.cuh"
+
+#define CUDA_CROSS_ENTROPY_LOSS_BLOCK_SIZE 256
+
+void ggml_cuda_cross_entropy_loss(ggml_backend_cuda_context & ctx, ggml_tensor * dst);
+
+void ggml_cuda_cross_entropy_loss_back(ggml_backend_cuda_context & ctx, ggml_tensor * dst);
diff --git a/src/whisper_cpp/ggml/src/ggml-cuda/dequantize.cuh b/src/whisper_cpp/ggml/src/ggml-cuda/dequantize.cuh
new file mode 100644
index 00000000..bd3c2d9d
--- /dev/null
+++ b/src/whisper_cpp/ggml/src/ggml-cuda/dequantize.cuh
@@ -0,0 +1,103 @@
+#include "common.cuh"
+
+static __device__ __forceinline__ void dequantize_q4_0(const void * vx, const int64_t ib, const int iqs, dfloat2 & v){
+    const block_q4_0 * x = (const block_q4_0 *) vx;
+
+    const dfloat d = x[ib].d;
+
+    const int vui = x[ib].qs[iqs];
+
+    v.x = vui & 0xF;
+    v.y = vui >> 4;
+
+#ifdef GGML_CUDA_F16
+    v = __hsub2(v, {8.0f, 8.0f});
+    v = __hmul2(v, {d, d});
+#else
+    v.x = (v.x - 8.0f) * d;
+    v.y = (v.y - 8.0f) * d;
+#endif // GGML_CUDA_F16
+}
+
+static __device__ __forceinline__ void dequantize_q4_1(const void * vx, const int64_t ib, const int iqs, dfloat2 & v){
+    const block_q4_1 * x = (const block_q4_1 *) vx;
+
+    const dfloat d = __low2half(x[ib].dm);
+    const dfloat m = __high2half(x[ib].dm);
+
+    const int vui = x[ib].qs[iqs];
+
+    v.x = vui & 0xF;
+    v.y = vui >> 4;
+
+#ifdef GGML_CUDA_F16
+    v = __hmul2(v, {d, d});
+    v = __hadd2(v, {m, m});
+#else
+    v.x = (v.x * d) + m;
+    v.y = (v.y * d) + m;
+#endif // GGML_CUDA_F16
+}
+
+static __device__ __forceinline__ void dequantize_q5_0(const void * vx, const int64_t ib, const int iqs, dfloat2 & v){
+    const block_q5_0 * x = (const block_q5_0 *) vx;
+
+    const dfloat d = x[ib].d;
+
+    uint32_t qh;
+    memcpy(&qh, x[ib].qh, sizeof(qh));
+
+    const int xh_0 = ((qh >> (iqs +  0)) << 4) & 0x10;
+    const int xh_1 = ((qh >> (iqs + 12))     ) & 0x10;
+
+    v.x = ((x[ib].qs[iqs] & 0xf) | xh_0);
+    v.y = ((x[ib].qs[iqs] >>  4) | xh_1);
+
+#ifdef GGML_CUDA_F16
+    v = __hsub2(v, {16.0f, 16.0f});
+    v = __hmul2(v, {d, d});
+#else
+    v.x = (v.x - 16.0f) * d;
+    v.y = (v.y - 16.0f) * d;
+#endif // GGML_CUDA_F16
+}
+
+static __device__ __forceinline__ void dequantize_q5_1(const void * vx, const int64_t ib, const int iqs, dfloat2 & v){
+    const block_q5_1 * x = (const block_q5_1 *) vx;
+
+    const dfloat d = __low2half(x[ib].dm);
+    const dfloat m = __high2half(x[ib].dm);
+
+    uint32_t qh;
+    memcpy(&qh, x[ib].qh, sizeof(qh));
+
+    const int xh_0 = ((qh >> (iqs +  0)) << 4) & 0x10;
+    const int xh_1 = ((qh >> (iqs + 12))     ) & 0x10;
+
+    v.x = ((x[ib].qs[iqs] & 0xf) | xh_0);
+    v.y = ((x[ib].qs[iqs] >>  4) | xh_1);
+
+#ifdef GGML_CUDA_F16
+    v = __hmul2(v, {d, d});
+    v = __hadd2(v, {m, m});
+#else
+    v.x = (v.x * d) + m;
+    v.y = (v.y * d) + m;
+#endif // GGML_CUDA_F16
+}
+
+static __device__ __forceinline__ void dequantize_q8_0(const void * vx, const int64_t ib, const int iqs, dfloat2 & v){
+    const block_q8_0 * x = (const block_q8_0 *) vx;
+
+    const dfloat d = x[ib].d;
+
+    v.x = x[ib].qs[iqs + 0];
+    v.y = x[ib].qs[iqs + 1];
+
+#ifdef GGML_CUDA_F16
+    v = __hmul2(v, {d, d});
+#else
+    v.x *= d;
+    v.y *= d;
+#endif // GGML_CUDA_F16
+}
diff --git a/src/whisper_cpp/ggml/src/ggml-cuda/diagmask.cu b/src/whisper_cpp/ggml/src/ggml-cuda/diagmask.cu
new file mode 100644
index 00000000..4b713ba2
--- /dev/null
+++ b/src/whisper_cpp/ggml/src/ggml-cuda/diagmask.cu
@@ -0,0 +1,40 @@
+#include "diagmask.cuh"
+
+static __global__ void diag_mask_inf_f32(const float * x, float * dst, const int ncols, const int rows_per_channel, const int n_past) {
+    const int col = blockDim.y*blockIdx.y + threadIdx.y;
+    const int row = blockDim.x*blockIdx.x + threadIdx.x;
+
+    if (col >= ncols) {
+        return;
+    }
+
+    const int i = row*ncols + col;
+    //dst[i] = col > (n_past + row % rows_per_channel) ? -INFINITY : x[i];
+    //dst[i] = x[i] - (col > n_past + row % rows_per_channel) * INT_MAX; // equivalent within rounding error but slightly faster on GPU
+    dst[i] = x[i] - (col > n_past + row % rows_per_channel) * FLT_MAX;
+}
+
+static void diag_mask_inf_f32_cuda(const float * x, float * dst, const int ncols_x, const int nrows_x, const int rows_per_channel, const int n_past, cudaStream_t stream) {
+    const dim3 block_dims(1, CUDA_DIAG_MASK_INF_BLOCK_SIZE, 1);
+    const int block_num_x = (ncols_x + CUDA_DIAG_MASK_INF_BLOCK_SIZE - 1) / CUDA_DIAG_MASK_INF_BLOCK_SIZE;
+    const dim3 block_nums(nrows_x, block_num_x, 1);
+    diag_mask_inf_f32<<<block_nums, block_dims, 0, stream>>>(x, dst, ncols_x, rows_per_channel, n_past);
+}
+
+void ggml_cuda_op_diag_mask_inf(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
+    const ggml_tensor * src0 = dst->src[0];
+    const float * src0_d = (const float *)src0->data;
+    float * dst_d = (float *)dst->data;
+    cudaStream_t stream = ctx.stream();
+
+    GGML_ASSERT(src0->type == GGML_TYPE_F32);
+    GGML_ASSERT( dst->type == GGML_TYPE_F32);
+
+    const int64_t ne00 = src0->ne[0];
+    const int64_t ne01 = src0->ne[1];
+    const int nrows0 = ggml_nrows(src0);
+
+    const int n_past = ((int32_t *) dst->op_params)[0];
+
+    diag_mask_inf_f32_cuda(src0_d, dst_d, ne00, nrows0, ne01, n_past, stream);
+}
diff --git a/src/whisper_cpp/ggml/src/ggml-cuda/diagmask.cuh b/src/whisper_cpp/ggml/src/ggml-cuda/diagmask.cuh
new file mode 100644
index 00000000..6cdbef17
--- /dev/null
+++ b/src/whisper_cpp/ggml/src/ggml-cuda/diagmask.cuh
@@ -0,0 +1,5 @@
+#include "common.cuh"
+
+#define CUDA_DIAG_MASK_INF_BLOCK_SIZE 32
+
+void ggml_cuda_op_diag_mask_inf(ggml_backend_cuda_context & ctx, ggml_tensor * dst);
diff --git a/src/whisper_cpp/ggml/src/ggml-cuda/dmmv.cu b/src/whisper_cpp/ggml/src/ggml-cuda/dmmv.cu
new file mode 100644
index 00000000..96a5adef
--- /dev/null
+++ b/src/whisper_cpp/ggml/src/ggml-cuda/dmmv.cu
@@ -0,0 +1,683 @@
+#include "dmmv.cuh"
+#include "dequantize.cuh"
+#include "convert.cuh"
+
+#ifndef K_QUANTS_PER_ITERATION
+#define K_QUANTS_PER_ITERATION 2
+#else
+static_assert(K_QUANTS_PER_ITERATION == 1 || K_QUANTS_PER_ITERATION == 2, "K_QUANTS_PER_ITERATION must be 1 or 2");
+#endif
+
+static __global__ void dequantize_mul_mat_vec_q2_k(const void * __restrict__ vx, const float * __restrict__ yy, float * __restrict__ dst, const int ncols, int nrows) {
+
+    static_assert(16%K_QUANTS_PER_ITERATION == 0, "16 must be divisible by K_QUANTS_PER_ITERATION");
+
+    const int row = blockIdx.x*blockDim.y + threadIdx.y;
+    if (row > nrows) return;
+
+    const int num_blocks_per_row = ncols / QK_K;
+    const int ib0 = row*num_blocks_per_row;
+
+    const block_q2_K * x = (const block_q2_K *)vx + ib0;
+
+    float tmp = 0; // partial sum for thread in warp
+
+    const int tid = threadIdx.x/K_QUANTS_PER_ITERATION;  // 0...31 or 0...15
+    const int ix  = threadIdx.x%K_QUANTS_PER_ITERATION;  // 0 or 0,1
+
+    const int step = 16/K_QUANTS_PER_ITERATION;
+
+    const int im = tid/step;                             // 0 or 1. 0 computes 0..., 1 computes 128...
+    const int in = tid - step*im;                        // 0...15 or 0...7
+
+    const int l0 = K_QUANTS_PER_ITERATION*in;            // 0...15 or 0...14 in steps of 2
+    const int q_offset = 32*im + l0;
+    const int s_offset = 8*im;
+    const int y_offset = 128*im + l0;
+
+    uint32_t aux[4];
+    const uint8_t * d = (const uint8_t *)aux;
+    const uint8_t * m = (const uint8_t *)(aux + 2);
+
+    for (int i = ix; i < num_blocks_per_row; i += K_QUANTS_PER_ITERATION) {
+
+        const float   * y = yy + i * QK_K + y_offset;
+        const uint8_t * q = x[i].qs + q_offset;
+
+        const float dall = __low2half(x[i].dm);
+        const float dmin = __high2half(x[i].dm);
+
+        const uint32_t * a = (const uint32_t *)(x[i].scales + s_offset);
+        aux[0] = a[0] & 0x0f0f0f0f;
+        aux[1] = a[1] & 0x0f0f0f0f;
+        aux[2] = (a[0] >> 4) & 0x0f0f0f0f;
+        aux[3] = (a[1] >> 4) & 0x0f0f0f0f;
+
+        float sum1 = 0, sum2 = 0;
+        for (int l = 0; l < K_QUANTS_PER_ITERATION; ++l) {
+            sum1 += y[l+ 0] * d[0] * ((q[l+ 0] >> 0) & 3)
+                  + y[l+32] * d[2] * ((q[l+ 0] >> 2) & 3)
+                  + y[l+64] * d[4] * ((q[l+ 0] >> 4) & 3)
+                  + y[l+96] * d[6] * ((q[l+ 0] >> 6) & 3)
+                  + y[l+16] * d[1] * ((q[l+16] >> 0) & 3)
+                  + y[l+48] * d[3] * ((q[l+16] >> 2) & 3)
+                  + y[l+80] * d[5] * ((q[l+16] >> 4) & 3)
+                  +y[l+112] * d[7] * ((q[l+16] >> 6) & 3);
+            sum2 += y[l+ 0] * m[0] + y[l+32] * m[2] + y[l+64] * m[4] + y[ l+96] * m[6]
+                  + y[l+16] * m[1] + y[l+48] * m[3] + y[l+80] * m[5] + y[l+112] * m[7];
+
+        }
+        tmp += dall * sum1 - dmin * sum2;
+
+    }
+
+    // sum up partial sums and write back result
+    tmp = warp_reduce_sum(tmp);
+
+    if (threadIdx.x == 0) {
+        dst[row] = tmp;
+    }
+}
+
+static __global__ void dequantize_mul_mat_vec_q3_k(const void * __restrict__ vx, const float * __restrict__ yy, float * __restrict__ dst, const int ncols, int nrows) {
+
+    const int row = blockIdx.x*blockDim.y + threadIdx.y;
+    if (row > nrows) return;
+
+    const int num_blocks_per_row = ncols / QK_K;
+    const int ib0 = row*num_blocks_per_row;
+
+    const block_q3_K * x = (const block_q3_K *)vx + ib0;
+
+    float tmp = 0; // partial sum for thread in warp
+
+    const uint16_t kmask1 = 0x0303;
+    const uint16_t kmask2 = 0x0f0f;
+
+    const int tid = threadIdx.x/K_QUANTS_PER_ITERATION;  // 0...31 or 0...16
+    const int ix  = threadIdx.x%K_QUANTS_PER_ITERATION;  // 0 or 0,1
+
+    const int n  = K_QUANTS_PER_ITERATION;               // iterations in the inner loop
+    const int step = 16/K_QUANTS_PER_ITERATION;
+    const int im = tid/step;                             // 0 or 1. 0 computes 0..., 1 computes 128...
+    const int in = tid - step*im;                        // 0....15 or 0...7
+
+    const uint8_t m = 1 << (4*im);
+
+    const int l0 = n*in;                                 // 0...15 or 0...14 in steps of 2
+    const int q_offset =  32*im + l0;
+    const int y_offset = 128*im + l0;
+
+    uint16_t utmp[4];
+    const int8_t * s = (const int8_t *)utmp;
+
+    const uint16_t s_shift = 4*im;
+
+    for (int i = ix; i < num_blocks_per_row; i += K_QUANTS_PER_ITERATION) {
+
+        const float   * y  = yy + i * QK_K + y_offset;
+        const uint8_t * q = x[i].qs + q_offset;
+        const uint8_t * h = x[i].hmask + l0;
+
+        const uint16_t * a = (const uint16_t *)x[i].scales;
+        utmp[0] = ((a[0] >> s_shift) & kmask2) | (((a[4] >> (s_shift + 0)) & kmask1) << 4);
+        utmp[1] = ((a[1] >> s_shift) & kmask2) | (((a[5] >> (s_shift + 0)) & kmask1) << 4);
+        utmp[2] = ((a[2] >> s_shift) & kmask2) | (((a[4] >> (s_shift + 2)) & kmask1) << 4);
+        utmp[3] = ((a[3] >> s_shift) & kmask2) | (((a[5] >> (s_shift + 2)) & kmask1) << 4);
+
+        const float d = x[i].d;
+
+        float sum = 0;
+        for (int l = 0; l < n; ++l) {
+            sum += y[l+ 0] * (s[0] - 32) * (((q[l] >> 0) & 3) - (h[l] & (m << 0) ? 0 : 4))
+                 + y[l+32] * (s[2] - 32) * (((q[l] >> 2) & 3) - (h[l] & (m << 1) ? 0 : 4))
+                 + y[l+64] * (s[4] - 32) * (((q[l] >> 4) & 3) - (h[l] & (m << 2) ? 0 : 4))
+                 + y[l+96] * (s[6] - 32) * (((q[l] >> 6) & 3) - (h[l] & (m << 3) ? 0 : 4));
+            sum += y[l+16] * (s[1] - 32) * (((q[l+16] >> 0) & 3) - (h[l+16] & (m << 0) ? 0 : 4))
+                 + y[l+48] * (s[3] - 32) * (((q[l+16] >> 2) & 3) - (h[l+16] & (m << 1) ? 0 : 4))
+                 + y[l+80] * (s[5] - 32) * (((q[l+16] >> 4) & 3) - (h[l+16] & (m << 2) ? 0 : 4))
+                + y[l+112] * (s[7] - 32) * (((q[l+16] >> 6) & 3) - (h[l+16] & (m << 3) ? 0 : 4));
+        }
+        tmp += d * sum;
+
+    }
+
+    // sum up partial sums and write back result
+    tmp = warp_reduce_sum(tmp);
+
+    if (threadIdx.x == 0) {
+        dst[row] = tmp;
+    }
+}
+
+static __global__ void dequantize_mul_mat_vec_q4_k(const void * __restrict__ vx, const float * __restrict__ yy, float * __restrict__ dst, const int ncols, int nrows) {
+
+    const int row = blockIdx.x*blockDim.y + threadIdx.y;
+    if (row > nrows) return;
+    const int num_blocks_per_row = ncols / QK_K;
+    const int ib0 = row*num_blocks_per_row;
+
+    const block_q4_K * x = (const block_q4_K *)vx + ib0;
+
+    const uint16_t kmask1 = 0x3f3f;
+    const uint16_t kmask2 = 0x0f0f;
+    const uint16_t kmask3 = 0xc0c0;
+
+    const int tid = threadIdx.x/K_QUANTS_PER_ITERATION;  // 0...31 or 0...16
+    const int ix  = threadIdx.x%K_QUANTS_PER_ITERATION;  // 0 or 0,1
+
+    const int step = 8/K_QUANTS_PER_ITERATION;           // 8 or 4
+
+    const int il  = tid/step;                            // 0...3
+    const int ir  = tid - step*il;                       // 0...7 or 0...3
+    const int n   = 2 * K_QUANTS_PER_ITERATION;          // 2 or 4
+
+    const int im = il/2;  // 0 or 1. 0 computes 0,32 + 128,160, 1 computes 64,96 + 192,224
+    const int in = il%2;
+
+    const int l0 = n*(2*ir + in);
+    const int q_offset = 32*im + l0;
+    const int y_offset = 64*im + l0;
+
+    uint16_t aux[4];
+    const uint8_t * sc = (const uint8_t *)aux;
+
+#if K_QUANTS_PER_ITERATION == 2
+    uint32_t q32[4];
+    const uint8_t * q4 = (const uint8_t *)q32;
+#else
+    uint16_t q16[4];
+    const uint8_t * q4 = (const uint8_t *)q16;
+#endif
+
+    float tmp = 0; // partial sum for thread in warp
+
+    for (int i = ix; i < num_blocks_per_row; i += K_QUANTS_PER_ITERATION) {
+
+        const float   * y1 = yy + i*QK_K + y_offset;
+        const float   * y2 = y1 + 128;
+
+        const float dall = __low2half(x[i].dm);
+        const float dmin = __high2half(x[i].dm);
+
+        const uint16_t * a = (const uint16_t *)x[i].scales;
+        aux[0] = a[im+0] & kmask1;
+        aux[1] = a[im+2] & kmask1;
+        aux[2] = ((a[im+4] >> 0) & kmask2) | ((a[im+0] & kmask3) >> 2);
+        aux[3] = ((a[im+4] >> 4) & kmask2) | ((a[im+2] & kmask3) >> 2);
+
+#if K_QUANTS_PER_ITERATION == 2
+        const uint32_t * q1 = (const uint32_t *)(x[i].qs + q_offset);
+        const uint32_t * q2 = q1 + 16;
+
+        q32[0] = q1[0] & 0x0f0f0f0f;
+        q32[1] = q1[0] & 0xf0f0f0f0;
+        q32[2] = q2[0] & 0x0f0f0f0f;
+        q32[3] = q2[0] & 0xf0f0f0f0;
+
+        float4 s = {0.f, 0.f, 0.f, 0.f};
+        float smin = 0;
+        for (int l = 0; l < 4; ++l) {
+            s.x += y1[l] * q4[l+0]; s.y += y1[l+32] * q4[l+ 4];
+            s.z += y2[l] * q4[l+8]; s.w += y2[l+32] * q4[l+12];
+            smin += y1[l] * sc[2] + y1[l+32] * sc[3] + y2[l] * sc[6] + y2[l+32] * sc[7];
+        }
+        tmp += dall * (s.x * sc[0] + s.y * sc[1] * 1.f/16.f + s.z * sc[4] + s.w * sc[5] * 1.f/16.f) - dmin * smin;
+#else
+        const uint16_t * q1 = (const uint16_t *)(x[i].qs + q_offset);
+        const uint16_t * q2 = q1 + 32;
+
+        q16[0] = q1[0] & 0x0f0f;
+        q16[1] = q1[0] & 0xf0f0;
+        q16[2] = q2[0] & 0x0f0f;
+        q16[3] = q2[0] & 0xf0f0;
+
+        float4 s = {0.f, 0.f, 0.f, 0.f};
+        float smin = 0;
+        for (int l = 0; l < 2; ++l) {
+            s.x += y1[l] * q4[l+0]; s.y += y1[l+32] * q4[l+2];
+            s.z += y2[l] * q4[l+4]; s.w += y2[l+32] * q4[l+6];
+            smin += y1[l] * sc[2] + y1[l+32] * sc[3] + y2[l] * sc[6] + y2[l+32] * sc[7];
+        }
+        tmp += dall * (s.x * sc[0] + s.y * sc[1] * 1.f/16.f + s.z * sc[4] + s.w * sc[5] * 1.f/16.f) - dmin * smin;
+#endif
+
+    }
+
+    // sum up partial sums and write back result
+    tmp = warp_reduce_sum(tmp);
+
+    if (tid == 0) {
+        dst[row] = tmp;
+    }
+}
+
+static __global__ void dequantize_mul_mat_vec_q5_k(const void * __restrict__ vx, const float * __restrict__ yy, float * __restrict__ dst, const int ncols) {
+
+    const int row = blockIdx.x;
+    const int num_blocks_per_row = ncols / QK_K;
+    const int ib0 = row*num_blocks_per_row;
+
+    const block_q5_K * x = (const block_q5_K *)vx + ib0;
+
+    float tmp = 0; // partial sum for thread in warp
+
+    const uint16_t kmask1 = 0x3f3f;
+    const uint16_t kmask2 = 0x0f0f;
+    const uint16_t kmask3 = 0xc0c0;
+
+    const int tid = threadIdx.x/2;  // 0...15
+    const int ix  = threadIdx.x%2;
+
+    const int il  = tid/4;     // 0...3
+    const int ir  = tid - 4*il;// 0...3
+    const int n   = 2;
+
+    const int im = il/2;  // 0 or 1. 0 computes 0,32 + 128,160, 1 computes 64,96 + 192,224
+    const int in = il%2;
+
+    const int l0 = n*(2*ir + in);
+    const int q_offset = 32*im + l0;
+    const int y_offset = 64*im + l0;
+
+    const uint8_t hm1  = 1 << (2*im);
+    const uint8_t hm2  = hm1 << 4;
+
+    uint16_t aux[4];
+    const uint8_t * sc = (const uint8_t *)aux;
+
+    uint16_t q16[8];
+    const uint8_t * q4 = (const uint8_t *)q16;
+
+    for (int i = ix; i < num_blocks_per_row; i += 2) {
+
+        const uint8_t * ql1 = x[i].qs + q_offset;
+        const uint8_t * qh  = x[i].qh + l0;
+        const float   * y1  = yy + i*QK_K + y_offset;
+        const float   * y2  = y1 + 128;
+
+        const float dall = __low2half(x[i].dm);
+        const float dmin = __high2half(x[i].dm);
+
+        const uint16_t * a = (const uint16_t *)x[i].scales;
+        aux[0] = a[im+0] & kmask1;
+        aux[1] = a[im+2] & kmask1;
+        aux[2] = ((a[im+4] >> 0) & kmask2) | ((a[im+0] & kmask3) >> 2);
+        aux[3] = ((a[im+4] >> 4) & kmask2) | ((a[im+2] & kmask3) >> 2);
+
+        float4 sum = {0.f, 0.f, 0.f, 0.f};
+        float smin = 0;
+        const uint16_t * q1 = (const uint16_t *)ql1;
+        const uint16_t * q2 = q1 + 32;
+        q16[0] = q1[0] & 0x0f0f;
+        q16[1] = q1[8] & 0x0f0f;
+        q16[2] = (q1[0] >> 4) & 0x0f0f;
+        q16[3] = (q1[8] >> 4) & 0x0f0f;
+        q16[4] = q2[0] & 0x0f0f;
+        q16[5] = q2[8] & 0x0f0f;
+        q16[6] = (q2[0] >> 4) & 0x0f0f;
+        q16[7] = (q2[8] >> 4) & 0x0f0f;
+        for (int l = 0; l < n; ++l) {
+            sum.x += y1[l+ 0] * (q4[l +0] + (qh[l+ 0] & (hm1 << 0) ? 16 : 0))
+                   + y1[l+16] * (q4[l +2] + (qh[l+16] & (hm1 << 0) ? 16 : 0));
+            sum.y += y1[l+32] * (q4[l +4] + (qh[l+ 0] & (hm1 << 1) ? 16 : 0))
+                   + y1[l+48] * (q4[l +6] + (qh[l+16] & (hm1 << 1) ? 16 : 0));
+            sum.z += y2[l+ 0] * (q4[l +8] + (qh[l+ 0] & (hm2 << 0) ? 16 : 0))
+                   + y2[l+16] * (q4[l+10] + (qh[l+16] & (hm2 << 0) ? 16 : 0));
+            sum.w += y2[l+32] * (q4[l+12] + (qh[l+ 0] & (hm2 << 1) ? 16 : 0))
+                   + y2[l+48] * (q4[l+14] + (qh[l+16] & (hm2 << 1) ? 16 : 0));
+            smin += (y1[l] + y1[l+16]) * sc[2] + (y1[l+32] + y1[l+48]) * sc[3]
+                  + (y2[l] + y2[l+16]) * sc[6] + (y2[l+32] + y2[l+48]) * sc[7];
+        }
+        tmp += dall * (sum.x * sc[0] + sum.y * sc[1] + sum.z * sc[4] + sum.w * sc[5]) - dmin * smin;
+    }
+
+    // sum up partial sums and write back result
+    tmp = warp_reduce_sum(tmp);
+
+    if (threadIdx.x == 0) {
+        dst[row] = tmp;
+    }
+}
+
+static __global__ void dequantize_mul_mat_vec_q6_k(const void * __restrict__ vx, const float * __restrict__ yy, float * __restrict__ dst, const int ncols, int nrows) {
+
+    static_assert(16%K_QUANTS_PER_ITERATION == 0, "16 must be divisible by K_QUANTS_PER_ITERATION");
+
+    const int row = blockIdx.x*blockDim.y + threadIdx.y;
+    if (row > nrows) return;
+
+    const int num_blocks_per_row = ncols / QK_K;
+    const int ib0 = row*num_blocks_per_row;
+
+    const block_q6_K * x = (const block_q6_K *)vx + ib0;
+
+    const int tid = threadIdx.x/K_QUANTS_PER_ITERATION;  // 0...31 or 0...16
+    const int ix  = threadIdx.x%K_QUANTS_PER_ITERATION;  // 0 or 0, 1
+
+    const int step = 16/K_QUANTS_PER_ITERATION;          // 16 or 8
+
+    const int im = tid/step;                             // 0 or 1. 0 computes 0..., 1 computes 128...
+    const int in = tid - step*im;                        // 0...15 or 0...7
+
+#if K_QUANTS_PER_ITERATION == 1
+    const int l0 = K_QUANTS_PER_ITERATION*in;            // 0...15
+    const int is = 0;
+#else
+    const int l0 = 4 * in;                               // 0, 4, 8, ..., 28
+    const int is = in / 4;
+#endif
+    const int ql_offset = 64*im + l0;
+    const int qh_offset = 32*im + l0;
+    const int s_offset  =  8*im + is;
+    const int y_offset = 128*im + l0;
+
+    float tmp = 0; // partial sum for thread in warp
+
+    for (int i = ix; i < num_blocks_per_row; i += K_QUANTS_PER_ITERATION) {
+
+        const float   * y  = yy + i * QK_K + y_offset;
+        const uint8_t * ql = x[i].ql + ql_offset;
+        const uint8_t * qh = x[i].qh + qh_offset;
+        const int8_t  * s  = x[i].scales + s_offset;
+
+        const float d = x[i].d;
+
+#if K_QUANTS_PER_ITERATION == 1
+        float sum = y[ 0] * s[0] * d * ((int8_t)((ql[ 0] & 0xF) | ((qh[ 0] & 0x03) << 4)) - 32)
+                  + y[16] * s[1] * d * ((int8_t)((ql[16] & 0xF) | ((qh[16] & 0x03) << 4)) - 32)
+                  + y[32] * s[2] * d * ((int8_t)((ql[32] & 0xF) | ((qh[ 0] & 0x0c) << 2)) - 32)
+                  + y[48] * s[3] * d * ((int8_t)((ql[48] & 0xF) | ((qh[16] & 0x0c) << 2)) - 32)
+                  + y[64] * s[4] * d * ((int8_t)((ql[ 0]  >> 4) | ((qh[ 0] & 0x30) >> 0)) - 32)
+                  + y[80] * s[5] * d * ((int8_t)((ql[16]  >> 4) | ((qh[16] & 0x30) >> 0)) - 32)
+                  + y[96] * s[6] * d * ((int8_t)((ql[32]  >> 4) | ((qh[ 0] & 0xc0) >> 2)) - 32)
+                  +y[112] * s[7] * d * ((int8_t)((ql[48]  >> 4) | ((qh[16] & 0xc0) >> 2)) - 32);
+        tmp += sum;
+#else
+        float sum = 0;
+        for (int l = 0; l < 4; ++l) {
+            sum += y[l+ 0] * s[0] * d * ((int8_t)((ql[l+ 0] & 0xF) | (((qh[l] >> 0) & 3) << 4)) - 32)
+                 + y[l+32] * s[2] * d * ((int8_t)((ql[l+32] & 0xF) | (((qh[l] >> 2) & 3) << 4)) - 32)
+                 + y[l+64] * s[4] * d * ((int8_t)((ql[l+ 0]  >> 4) | (((qh[l] >> 4) & 3) << 4)) - 32)
+                 + y[l+96] * s[6] * d * ((int8_t)((ql[l+32]  >> 4) | (((qh[l] >> 6) & 3) << 4)) - 32);
+        }
+        tmp += sum;
+#endif
+
+    }
+
+    // sum up partial sums and write back result
+    tmp = warp_reduce_sum(tmp);
+
+    if (tid == 0) {
+        dst[row] = tmp;
+    }
+}
+
+static __device__ void convert_f16(const void * vx, const int64_t ib, const int iqs, dfloat2 & v){
+    const half * x = (const half *) vx;
+
+    // automatic half -> float type cast if dfloat == float
+    v.x = x[ib + iqs + 0];
+    v.y = x[ib + iqs + 1];
+}
+
+static constexpr __device__ dequantize_kernel_t get_dequantize_kernel(ggml_type type) {
+    return type == GGML_TYPE_Q4_0 ? dequantize_q4_0 :
+        type == GGML_TYPE_Q4_1 ? dequantize_q4_1 :
+        type == GGML_TYPE_Q5_0 ? dequantize_q5_0 :
+        type == GGML_TYPE_Q5_1 ? dequantize_q5_1 :
+        type == GGML_TYPE_Q8_0 ? dequantize_q8_0 :
+        type == GGML_TYPE_F16 ? convert_f16 :
+        nullptr;
+}
+
+template <ggml_type type>
+static __global__ void dequantize_mul_mat_vec(const void * __restrict__ vx, const dfloat * __restrict__ y, float * __restrict__ dst, const int ncols, const int nrows) {
+    constexpr int qk = ggml_cuda_type_traits<type>::qk; // quantized weights per x block
+    constexpr int qr = ggml_cuda_type_traits<type>::qr; // number of quantized weights per data value in x block
+    constexpr dequantize_kernel_t dequantize_kernel = get_dequantize_kernel(type);
+
+    const int64_t row = (int64_t)blockIdx.x*blockDim.y + threadIdx.y;
+
+    if (row >= nrows) {
+        return;
+    }
+
+    const int tid = threadIdx.x;
+
+    const int iter_stride = 2*GGML_CUDA_DMMV_X;
+    const int vals_per_iter = iter_stride / WARP_SIZE; // num quantized vals per thread and i iter
+    const int y_offset = qr == 1 ? 1 : qk/2;
+
+// partial sum for each thread
+#ifdef GGML_CUDA_F16
+    half2 tmp = {0.0f, 0.0f}; // two sums for f16 to take advantage of half2 intrinsics
+#else
+    float tmp = 0.0f;
+#endif // GGML_CUDA_F16
+
+    for (int i = 0; i < ncols; i += iter_stride) {
+        const int col = i + vals_per_iter*tid;
+        const int64_t ib = ((int64_t)row*ncols + col)/qk; // x block index
+        const int iqs = (col%qk)/qr; // x quant index
+        const int iybs = col - col%qk; // y block start index
+
+// processing >2 values per i iter is faster for fast GPUs
+#pragma unroll
+        for (int j = 0; j < vals_per_iter; j += 2) {
+            // process 2 vals per j iter
+
+            // dequantize
+            // for qr = 2 the iqs needs to increase by 1 per j iter because 2 weights per data val
+            dfloat2 v;
+            dequantize_kernel(vx, ib, iqs + j/qr, v);
+
+            // matrix multiplication
+            // for qr = 2 the y index needs to increase by 1 per j iter because of y_offset = qk/2
+#ifdef GGML_CUDA_F16
+            tmp += __hmul2(v, {
+                y[iybs + iqs + j/qr + 0],
+                y[iybs + iqs + j/qr + y_offset]
+            });
+#else
+            tmp += v.x * y[iybs + iqs + j/qr + 0];
+            tmp += v.y * y[iybs + iqs + j/qr + y_offset];
+#endif // GGML_CUDA_F16
+        }
+    }
+
+    // sum up partial sums and write back result
+    tmp = warp_reduce_sum(tmp);
+
+    if (tid == 0) {
+#ifdef GGML_CUDA_F16
+        dst[row] = tmp.x + tmp.y;
+#else
+        dst[row] = tmp;
+#endif // GGML_CUDA_F16
+    }
+}
+
+static void dequantize_mul_mat_vec_q4_0_cuda(const void * vx, const dfloat * y, float * dst, const int ncols, const int nrows, cudaStream_t stream) {
+    GGML_ASSERT(ncols % (GGML_CUDA_DMMV_X*2) == 0);
+    const int block_num_y = (nrows + GGML_CUDA_MMV_Y - 1) / GGML_CUDA_MMV_Y;
+    // the number of rows may exceed maximum grid size in the y or z dimensions, use the x dimension instead
+    const dim3 block_nums(block_num_y, 1, 1);
+    const dim3 block_dims(WARP_SIZE, GGML_CUDA_MMV_Y, 1);
+    dequantize_mul_mat_vec<GGML_TYPE_Q4_0>
+        <<<block_nums, block_dims, 0, stream>>>(vx, y, dst, ncols, nrows);
+}
+
+static void dequantize_mul_mat_vec_q4_1_cuda(const void * vx, const dfloat * y, float * dst, const int ncols, const int nrows, cudaStream_t stream) {
+    GGML_ASSERT(ncols % (GGML_CUDA_DMMV_X*2) == 0);
+    const int block_num_y = (nrows + GGML_CUDA_MMV_Y - 1) / GGML_CUDA_MMV_Y;
+    const dim3 block_nums(block_num_y, 1, 1);
+    const dim3 block_dims(WARP_SIZE, GGML_CUDA_MMV_Y, 1);
+    dequantize_mul_mat_vec<GGML_TYPE_Q4_1>
+        <<<block_nums, block_dims, 0, stream>>>(vx, y, dst, ncols, nrows);
+}
+
+static void dequantize_mul_mat_vec_q5_0_cuda(const void * vx, const dfloat * y, float * dst, const int ncols, const int nrows, cudaStream_t stream) {
+    GGML_ASSERT(ncols % (GGML_CUDA_DMMV_X*2) == 0);
+    const int block_num_y = (nrows + GGML_CUDA_MMV_Y - 1) / GGML_CUDA_MMV_Y;
+    const dim3 block_nums(block_num_y, 1, 1);
+    const dim3 block_dims(WARP_SIZE, GGML_CUDA_MMV_Y, 1);
+    dequantize_mul_mat_vec<GGML_TYPE_Q5_0>
+        <<<block_nums, block_dims, 0, stream>>>(vx, y, dst, ncols, nrows);
+}
+
+static void dequantize_mul_mat_vec_q5_1_cuda(const void * vx, const dfloat * y, float * dst, const int ncols, const int nrows, cudaStream_t stream) {
+    GGML_ASSERT(ncols % (GGML_CUDA_DMMV_X*2) == 0);
+    const int block_num_y = (nrows + GGML_CUDA_MMV_Y - 1) / GGML_CUDA_MMV_Y;
+    const dim3 block_nums(block_num_y, 1, 1);
+    const dim3 block_dims(WARP_SIZE, GGML_CUDA_MMV_Y, 1);
+    dequantize_mul_mat_vec<GGML_TYPE_Q5_1>
+        <<<block_nums, block_dims, 0, stream>>>(vx, y, dst, ncols, nrows);
+}
+
+static void dequantize_mul_mat_vec_q8_0_cuda(const void * vx, const dfloat * y, float * dst, const int ncols, const int nrows, cudaStream_t stream) {
+    GGML_ASSERT(ncols % (GGML_CUDA_DMMV_X*2) == 0);
+    const int block_num_y = (nrows + GGML_CUDA_MMV_Y - 1) / GGML_CUDA_MMV_Y;
+    const dim3 block_nums(block_num_y, 1, 1);
+    const dim3 block_dims(WARP_SIZE, GGML_CUDA_MMV_Y, 1);
+    dequantize_mul_mat_vec<GGML_TYPE_Q8_0>
+        <<<block_nums, block_dims, 0, stream>>>(vx, y, dst, ncols, nrows);
+}
+
+static void dequantize_mul_mat_vec_q2_K_cuda(const void * vx, const float * y, float * dst, const int ncols, const int nrows, cudaStream_t stream) {
+    GGML_ASSERT(ncols % QK_K == 0);
+    const int ny = 2; // very slightly faster than 1 even when K_QUANTS_PER_ITERATION = 2
+    const int block_num_y = (nrows + ny - 1) / ny;
+    const dim3 block_nums(block_num_y, 1, 1);
+    const dim3 block_dims(32, ny, 1);
+    dequantize_mul_mat_vec_q2_k<<<block_nums, block_dims, 0, stream>>>(vx, y, dst, ncols, nrows);
+}
+
+static void dequantize_mul_mat_vec_q3_K_cuda(const void * vx, const float * y, float * dst, const int ncols, const int nrows, cudaStream_t stream) {
+    GGML_ASSERT(ncols % QK_K == 0);
+    const int ny = 2 / K_QUANTS_PER_ITERATION;
+    const int block_num_y = (nrows + ny - 1) / ny;
+    const dim3 block_nums(block_num_y, 1, 1);
+    const dim3 block_dims(32, ny, 1);
+    dequantize_mul_mat_vec_q3_k<<<block_nums, block_dims, 0, stream>>>(vx, y, dst, ncols, nrows);
+}
+
+static void dequantize_mul_mat_vec_q4_K_cuda(const void * vx, const float * y, float * dst, const int ncols, const int nrows, cudaStream_t stream) {
+    GGML_ASSERT(ncols % QK_K == 0);
+    const int ny = 2 / K_QUANTS_PER_ITERATION;
+    const int block_num_y = (nrows + ny - 1) / ny;
+    const dim3 block_nums(block_num_y, 1, 1);
+    const dim3 block_dims(32, ny, 1);
+    dequantize_mul_mat_vec_q4_k<<<block_nums, block_dims, 0, stream>>>(vx, y, dst, ncols, nrows);
+}
+
+static void dequantize_mul_mat_vec_q5_K_cuda(const void * vx, const float * y, float * dst, const int ncols, const int nrows, cudaStream_t stream) {
+    GGML_ASSERT(ncols % QK_K == 0);
+    const dim3 block_dims(32, 1, 1);
+    dequantize_mul_mat_vec_q5_k<<<nrows, block_dims, 0, stream>>>(vx, y, dst, ncols);
+}
+
+static void dequantize_mul_mat_vec_q6_K_cuda(const void * vx, const float * y, float * dst, const int ncols, const int nrows, cudaStream_t stream) {
+    GGML_ASSERT(ncols % QK_K == 0);
+    const int ny = 2 / K_QUANTS_PER_ITERATION;
+    const int block_num_y = (nrows + ny - 1) / ny;
+    const dim3 block_nums(block_num_y, 1, 1);
+    const dim3 block_dims(32, ny, 1);
+    dequantize_mul_mat_vec_q6_k<<<block_nums, block_dims, 0, stream>>>(vx, y, dst, ncols, nrows);
+}
+
+static void convert_mul_mat_vec_f16_cuda(const void * vx, const dfloat * y, float * dst, const int ncols, const int nrows, cudaStream_t stream) {
+    GGML_ASSERT(ncols % (GGML_CUDA_DMMV_X*2) == 0);
+    const int block_num_y = (nrows + GGML_CUDA_MMV_Y - 1) / GGML_CUDA_MMV_Y;
+    const dim3 block_nums(block_num_y, 1, 1);
+    const dim3 block_dims(WARP_SIZE, GGML_CUDA_MMV_Y, 1);
+    dequantize_mul_mat_vec<GGML_TYPE_F16>
+        <<<block_nums, block_dims, 0, stream>>>(vx, y, dst, ncols, nrows);
+}
+
+void ggml_cuda_op_dequantize_mul_mat_vec(
+    ggml_backend_cuda_context & ctx,
+    const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, const char * src0_dd_i, const float * src1_ddf_i,
+    const char * src1_ddq_i, float * dst_dd_i, const int64_t row_low, const int64_t row_high, const int64_t src1_ncols,
+    const int64_t src1_padded_row_size, cudaStream_t stream) {
+    GGML_UNUSED(ctx);
+    const int64_t ne00 = src0->ne[0];
+    const int64_t row_diff = row_high - row_low;
+
+    GGML_ASSERT(src1->type == GGML_TYPE_F32);
+
+    // on some GPUs it is faster to convert src1 to half and to use half precision intrinsics
+#ifdef GGML_CUDA_F16
+    ggml_cuda_pool_alloc<half> src1_dfloat_a(ctx.pool());
+    half * src1_dfloat = nullptr; // dfloat == half
+
+    bool src1_convert_f16 =
+        src0->type == GGML_TYPE_Q4_0 || src0->type == GGML_TYPE_Q4_1 ||
+        src0->type == GGML_TYPE_Q5_0 || src0->type == GGML_TYPE_Q5_1 ||
+        src0->type == GGML_TYPE_Q8_0 || src0->type == GGML_TYPE_F16;
+
+    if (src1_convert_f16) {
+        src1_dfloat = src1_dfloat_a.alloc(ne00);
+        const to_fp16_cuda_t to_fp16_cuda = ggml_get_to_fp16_cuda(src1->type);
+        GGML_ASSERT(to_fp16_cuda != nullptr);
+        to_fp16_cuda(src1_ddf_i, src1_dfloat, ne00, stream);
+    }
+#else
+    const dfloat * src1_dfloat = (const dfloat *) src1_ddf_i; // dfloat == float, no conversion
+#endif // GGML_CUDA_F16
+
+    switch (src0->type) {
+        case GGML_TYPE_Q4_0:
+            dequantize_mul_mat_vec_q4_0_cuda(src0_dd_i, src1_dfloat, dst_dd_i, ne00, row_diff, stream);
+            break;
+        case GGML_TYPE_Q4_1:
+            dequantize_mul_mat_vec_q4_1_cuda(src0_dd_i, src1_dfloat, dst_dd_i, ne00, row_diff, stream);
+            break;
+        case GGML_TYPE_Q5_0:
+            dequantize_mul_mat_vec_q5_0_cuda(src0_dd_i, src1_dfloat, dst_dd_i, ne00, row_diff, stream);
+            break;
+        case GGML_TYPE_Q5_1:
+            dequantize_mul_mat_vec_q5_1_cuda(src0_dd_i, src1_dfloat, dst_dd_i, ne00, row_diff, stream);
+            break;
+        case GGML_TYPE_Q8_0:
+            dequantize_mul_mat_vec_q8_0_cuda(src0_dd_i, src1_dfloat, dst_dd_i, ne00, row_diff, stream);
+            break;
+        case GGML_TYPE_Q2_K:
+            dequantize_mul_mat_vec_q2_K_cuda(src0_dd_i, src1_ddf_i, dst_dd_i, ne00, row_diff, stream);
+            break;
+        case GGML_TYPE_Q3_K:
+            dequantize_mul_mat_vec_q3_K_cuda(src0_dd_i, src1_ddf_i, dst_dd_i, ne00, row_diff, stream);
+            break;
+        case GGML_TYPE_Q4_K:
+            dequantize_mul_mat_vec_q4_K_cuda(src0_dd_i, src1_ddf_i, dst_dd_i, ne00, row_diff, stream);
+            break;
+        case GGML_TYPE_Q5_K:
+            dequantize_mul_mat_vec_q5_K_cuda(src0_dd_i, src1_ddf_i, dst_dd_i, ne00, row_diff, stream);
+            break;
+        case GGML_TYPE_Q6_K:
+            dequantize_mul_mat_vec_q6_K_cuda(src0_dd_i, src1_ddf_i, dst_dd_i, ne00, row_diff, stream);
+            break;
+        case GGML_TYPE_F16:
+            convert_mul_mat_vec_f16_cuda(src0_dd_i, src1_dfloat, dst_dd_i, ne00, row_diff, stream);
+            break;
+        default:
+            GGML_ABORT("fatal error");
+            break;
+    }
+
+    GGML_UNUSED(src1);
+    GGML_UNUSED(dst);
+    GGML_UNUSED(src1_ddq_i);
+    GGML_UNUSED(src1_ncols);
+    GGML_UNUSED(src1_padded_row_size);
+}
+
+bool ggml_cuda_dmmv_type_supported(ggml_type src0_type) {
+    return src0_type == GGML_TYPE_Q4_0 || src0_type == GGML_TYPE_Q4_1 ||
+        src0_type == GGML_TYPE_Q5_0 || src0_type == GGML_TYPE_Q5_1 ||
+        src0_type == GGML_TYPE_Q8_0 || src0_type == GGML_TYPE_Q2_K ||
+        src0_type == GGML_TYPE_Q3_K || src0_type == GGML_TYPE_Q4_K ||
+        src0_type == GGML_TYPE_Q5_K || src0_type == GGML_TYPE_Q6_K ||
+        src0_type == GGML_TYPE_F16;
+}
diff --git a/src/whisper_cpp/ggml/src/ggml-cuda/dmmv.cuh b/src/whisper_cpp/ggml/src/ggml-cuda/dmmv.cuh
new file mode 100644
index 00000000..e727eb97
--- /dev/null
+++ b/src/whisper_cpp/ggml/src/ggml-cuda/dmmv.cuh
@@ -0,0 +1,20 @@
+#include "common.cuh"
+
+// dmmv = dequantize_mul_mat_vec
+
+// TODO: remove this?
+#ifndef GGML_CUDA_DMMV_X
+#define GGML_CUDA_DMMV_X 32
+#endif
+
+#ifndef GGML_CUDA_MMV_Y
+#define GGML_CUDA_MMV_Y 1
+#endif
+
+void ggml_cuda_op_dequantize_mul_mat_vec(
+    ggml_backend_cuda_context & ctx,
+    const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, const char * src0_dd_i, const float * src1_ddf_i,
+    const char * src1_ddq_i, float * dst_dd_i, const int64_t row_low, const int64_t row_high, const int64_t src1_ncols,
+    const int64_t src1_padded_row_size, cudaStream_t stream);
+
+bool ggml_cuda_dmmv_type_supported(ggml_type src0_type);
diff --git a/src/whisper_cpp/ggml/src/ggml-cuda/fattn-common.cuh b/src/whisper_cpp/ggml/src/ggml-cuda/fattn-common.cuh
new file mode 100644
index 00000000..1fb5c09c
--- /dev/null
+++ b/src/whisper_cpp/ggml/src/ggml-cuda/fattn-common.cuh
@@ -0,0 +1,708 @@
+#pragma once
+
+#include "common.cuh"
+#include "convert.cuh"
+#include "vecdotq.cuh"
+
+#include <cstdint>
+
+#define FATTN_KQ_STRIDE       256
+#define HALF_MAX_HALF         __float2half(65504.0f/2) // Use neg. of this instead of -INFINITY to initialize KQ max vals to avoid NaN upon subtraction.
+#define SOFTMAX_FTZ_THRESHOLD -20.0f                   // Softmax exp. of values smaller than this are flushed to zero to avoid NaNs.
+
+typedef void (* fattn_kernel_t)(
+        const char * __restrict__ Q,
+        const char * __restrict__ K,
+        const char * __restrict__ V,
+        const char * __restrict__ mask,
+        float      * __restrict__ dst,
+        float2     * __restrict__ dst_meta,
+        const float scale,
+        const float max_bias,
+        const float m0,
+        const float m1,
+        const uint32_t n_head_log2,
+        const float logit_softcap,
+        const int ne00,
+        const int ne01,
+        const int ne02,
+        const int ne03,
+        const int ne10,
+        const int ne11,
+        const int ne12,
+        const int ne13,
+        const int ne31,
+        const int nb31,
+        const int nb01,
+        const int nb02,
+        const int nb03,
+        const int nb11,
+        const int nb12,
+        const int nb13,
+        const int nb21,
+        const int nb22,
+        const int nb23,
+        const int ne0,
+        const int ne1,
+        const int ne2,
+        const int ne3);
+
+typedef half (*vec_dot_KQ_f16_t)(
+    const char * __restrict__ K_c, const void * __restrict__ Q_v, const int * __restrict__ Q_q8 , const void * __restrict__ Q_ds);
+typedef float (*vec_dot_KQ_f32_t)(
+    const char * __restrict__ K_c, const void * __restrict__ Q_v, const int * __restrict__ Q_q8 , const void * __restrict__ Q_ds);
+
+template<typename T, int D>
+static __device__ __forceinline__ T vec_dot_fattn_vec_KQ_q4_0(
+    const char * __restrict__ K_c, const void * __restrict__ Q_v, const int * __restrict__ Q_q8, const void * __restrict__ Q_ds_v) {
+
+    const block_q4_0 * K_q4_0 = (const block_q4_0 *) K_c;
+    GGML_UNUSED(Q_v);
+
+    T sum = 0.0f;
+
+#pragma unroll
+    for (int k_KQ_0 = 0; k_KQ_0 < D/sizeof(int); k_KQ_0 += WARP_SIZE) {
+        const int k_KQ = k_KQ_0 + threadIdx.x;
+
+        const int ib    = k_KQ /  QI8_1;
+        const int iqs4  = k_KQ %  QI4_0;
+        const int shift = k_KQ & (QI8_1/2);
+
+        const int v = (get_int_b2(K_q4_0[ib].qs, iqs4) >> shift) & 0x0F0F0F0F;
+        const int u = Q_q8[k_KQ_0/WARP_SIZE];
+
+        const int sumi = ggml_cuda_dp4a(v, u, 0);
+
+#ifdef FP16_AVAILABLE
+        if (std::is_same<T, half>::value) {
+            const half2  * Q_ds = (const half2  *) Q_ds_v;
+
+            const half2 sum2 = __half2half2(K_q4_0[ib].d) * Q_ds[k_KQ_0/WARP_SIZE];
+            sum += (T) (((half) sumi)*__low2half(sum2) - __high2half(sum2) /* *8/QI8_1 == 1 */);
+        } else
+#endif // FP16_AVAILABLE
+        {
+            const float2 * Q_ds = (const float2 *) Q_ds_v;
+
+            sum += (T) (__half2float(K_q4_0[ib].d) * (sumi*Q_ds[k_KQ_0/WARP_SIZE].x - (8/QI8_1)*Q_ds[k_KQ_0/WARP_SIZE].y));
+        }
+    }
+
+    return sum;
+}
+
+template<typename T, int D>
+static __device__ __forceinline__ T vec_dot_fattn_vec_KQ_q4_1(
+    const char * __restrict__ K_c, const void * __restrict__ Q_v, const int * __restrict__ Q_q8, const void * __restrict__ Q_ds_v) {
+
+    const block_q4_1 * K_q4_1 = (const block_q4_1 *) K_c;
+    GGML_UNUSED(Q_v);
+
+    T sum = 0.0f;
+
+#pragma unroll
+    for (int k_KQ_0 = 0; k_KQ_0 < D/sizeof(int); k_KQ_0 += WARP_SIZE) {
+        const int k_KQ = k_KQ_0 + threadIdx.x;
+
+        const int ib    = k_KQ /  QI8_1;
+        const int iqs4  = k_KQ %  QI4_1;
+        const int shift = k_KQ & (QI8_1/2);
+
+        const int v = (get_int_b4(K_q4_1[ib].qs, iqs4) >> shift) & 0x0F0F0F0F;
+        const int u = Q_q8[k_KQ_0/WARP_SIZE];
+
+        const int sumi = ggml_cuda_dp4a(v, u, 0);
+
+#ifdef FP16_AVAILABLE
+        if (std::is_same<T, half>::value) {
+            const half2  * Q_ds = (const half2  *) Q_ds_v;
+
+            const half2 d4d8_m4s8 = K_q4_1[ib].dm * Q_ds[k_KQ_0/WARP_SIZE];
+            const half2 sumid4d8_m4s8scaled = d4d8_m4s8 * make_half2(sumi, 1.0f/QI8_1);
+            sum += (T) (__low2half(sumid4d8_m4s8scaled) + __high2half(sumid4d8_m4s8scaled));
+        } else
+#endif // FP16_AVAILABLE
+        {
+            const float2 * Q_ds = (const float2 *) Q_ds_v;
+
+            const float sumid4d8   =  __low2float(K_q4_1[ib].dm)*Q_ds[k_KQ_0/WARP_SIZE].x * sumi;
+            const float m4s8scaled = __high2float(K_q4_1[ib].dm)*Q_ds[k_KQ_0/WARP_SIZE].y / QI8_1;
+
+            sum += (T) (sumid4d8 + m4s8scaled);
+        }
+    }
+
+    return sum;
+}
+
+template<typename T, int D>
+static __device__ __forceinline__ T vec_dot_fattn_vec_KQ_q5_0(
+    const char * __restrict__ K_c, const void * __restrict__ Q_v, const int * __restrict__ Q_q8, const void * __restrict__ Q_ds_v) {
+
+    const block_q5_0 * K_q5_0 = (const block_q5_0 *) K_c;
+    GGML_UNUSED(Q_v);
+
+    T sum = 0.0f;
+
+#pragma unroll
+    for (int k_KQ_0 = 0; k_KQ_0 < D/sizeof(int); k_KQ_0 += WARP_SIZE) {
+        const int k_KQ = k_KQ_0 + threadIdx.x;
+
+        const int ib    = k_KQ /  QI8_1;
+        const int iqs4  = k_KQ %  QI5_0;
+        const int iqs8  = k_KQ %  QI8_1;
+        const int shift = k_KQ & (QI8_1/2);
+
+        int v = (get_int_b2(K_q5_0[ib].qs, iqs4) >> shift) & 0x0F0F0F0F;
+        const int vh = get_int_b2(K_q5_0[ib].qh, 0) >> (iqs8 * QI5_0);
+        v |= (vh <<  4) & 0x00000010; // 0 ->  4
+        v |= (vh << 11) & 0x00001000; // 1 -> 12
+        v |= (vh << 18) & 0x00100000; // 2 -> 20
+        v |= (vh << 25) & 0x10000000; // 3 -> 28
+
+        const int u = Q_q8[k_KQ_0/WARP_SIZE];
+
+        const int sumi = ggml_cuda_dp4a(v, u, 0);
+
+#ifdef FP16_AVAILABLE
+        if (std::is_same<T, half>::value) {
+            const half2  * Q_ds = (const half2  *) Q_ds_v;
+
+            const half2 sum2 = __half2half2(K_q5_0[ib].d) * Q_ds[k_KQ_0/WARP_SIZE];
+            sum += (T) (((half) sumi)*__low2half(sum2) - __high2half(sum2)*__float2half(2.0f)) /* *16/QI8_1 == 2 */;
+        } else
+#endif // FP16_AVAILABLE
+        {
+            const float2 * Q_ds = (const float2 *) Q_ds_v;
+
+            sum += (T) (__half2float(K_q5_0[ib].d) * (sumi*Q_ds[k_KQ_0/WARP_SIZE].x - (16/QI8_1)*Q_ds[k_KQ_0/WARP_SIZE].y));
+        }
+    }
+
+    return sum;
+}
+
+template<typename T, int D>
+static __device__ __forceinline__ T vec_dot_fattn_vec_KQ_q5_1(
+    const char * __restrict__ K_c, const void * __restrict__ Q_v, const int * __restrict__ Q_q8, const void * __restrict__ Q_ds_v) {
+
+    const block_q5_1 * K_q5_1 = (const block_q5_1 *) K_c;
+    GGML_UNUSED(Q_v);
+
+    T sum = 0.0f;
+
+#pragma unroll
+    for (int k_KQ_0 = 0; k_KQ_0 < D/sizeof(int); k_KQ_0 += WARP_SIZE) {
+        const int k_KQ = k_KQ_0 + threadIdx.x;
+
+        const int ib    = k_KQ /  QI8_1;
+        const int iqs4  = k_KQ %  QI5_1;
+        const int iqs8  = k_KQ %  QI8_1;
+        const int shift = k_KQ & (QI8_1/2);
+
+        int v = (get_int_b2(K_q5_1[ib].qs, iqs4) >> shift) & 0x0F0F0F0F;
+        const int vh = get_int_b2(K_q5_1[ib].qh, 0) >> (iqs8 * QI5_1);
+        v |= (vh <<  4) & 0x00000010; // 0 ->  4
+        v |= (vh << 11) & 0x00001000; // 1 -> 12
+        v |= (vh << 18) & 0x00100000; // 2 -> 20
+        v |= (vh << 25) & 0x10000000; // 3 -> 28
+
+        const int u = Q_q8[k_KQ_0/WARP_SIZE];
+
+        const int sumi = ggml_cuda_dp4a(v, u, 0);
+
+#ifdef FP16_AVAILABLE
+        if (std::is_same<T, half>::value) {
+            const half2  * Q_ds = (const half2  *) Q_ds_v;
+
+            const half2 d5d8_m5s8 = K_q5_1[ib].dm * Q_ds[k_KQ_0/WARP_SIZE];
+            const half2 sumid5d8_m5s8scaled = d5d8_m5s8 * make_half2(sumi, 1.0f/QI8_1);
+            sum += (T) (__low2half(sumid5d8_m5s8scaled) + __high2half(sumid5d8_m5s8scaled));
+        } else
+#endif // FP16_AVAILABLE
+        {
+            const float2 * Q_ds = (const float2 *) Q_ds_v;
+
+            const float sumid5d8   =  __low2float(K_q5_1[ib].dm)*Q_ds[k_KQ_0/WARP_SIZE].x * sumi;
+            const float m5s8scaled = __high2float(K_q5_1[ib].dm)*Q_ds[k_KQ_0/WARP_SIZE].y / QI8_1;
+
+            sum += (T) (sumid5d8 + m5s8scaled);
+        }
+    }
+
+    return sum;
+}
+
+template <typename T, int D>
+static __device__ __forceinline__ T vec_dot_fattn_vec_KQ_q8_0(
+    const char * __restrict__ K_c, const void * __restrict__ Q_v, const int * __restrict__ Q_q8, const void * __restrict__ Q_ds_v) {
+
+    const block_q8_0 * K_q8_0 = (const block_q8_0 *) K_c;
+    GGML_UNUSED(Q_v);
+
+    T sum = 0.0f;
+
+#pragma unroll
+    for (int k_KQ_0 = 0; k_KQ_0 < D/sizeof(int); k_KQ_0 += WARP_SIZE) {
+        const int k_KQ = k_KQ_0 + threadIdx.x;
+
+        const int ib  = k_KQ / QI8_0;
+        const int iqs = k_KQ % QI8_0;
+
+        const int v = get_int_b2(K_q8_0[ib].qs, iqs);
+
+        T Q_d;
+        if (std::is_same<T, half>::value) {
+            const half2  * Q_ds = (const half2  *) Q_ds_v;
+            Q_d = __low2half(Q_ds[k_KQ_0/WARP_SIZE]);
+        } else {
+            const float2 * Q_ds = (const float2 *) Q_ds_v;
+            Q_d = Q_ds[k_KQ_0/WARP_SIZE].x;
+        }
+
+        sum += vec_dot_q8_0_q8_1_impl<T, 1>(&v, &Q_q8[k_KQ_0/WARP_SIZE], K_q8_0[ib].d, Q_d);
+    }
+
+    return sum;
+}
+
+template <typename T, int D>
+static __device__ __forceinline__ T vec_dot_fattn_vec_KQ_f16(
+    const char * __restrict__ K_c, const void * __restrict__ Q_v, const int * __restrict__ Q_q8 , const void * __restrict__ Q_ds_v) {
+
+    const half2 * K_h2 = (const half2 *) K_c;
+    GGML_UNUSED(Q_q8);
+    GGML_UNUSED(Q_ds_v);
+
+#ifdef FP16_AVAILABLE
+    if (std::is_same<T, half>::value) {
+        const half2 * Q_h2 = (const half2 *) Q_v;
+
+        half2 sum2 = make_half2(0.0f, 0.0f);
+
+#pragma unroll
+        for (int k_KQ_0 = 0; k_KQ_0 < D/2; k_KQ_0 += WARP_SIZE) {
+            const int k_KQ = k_KQ_0 + threadIdx.x;
+
+            const half2 K_ik = K_h2[k_KQ];
+            sum2 += K_ik * Q_h2[k_KQ_0/WARP_SIZE];
+        }
+
+        return __low2half(sum2) + __high2half(sum2);
+    }
+#endif // FP16_AVAILABLE
+
+    const float2 * Q_f2 = (const float2 *) Q_v;
+
+    float sum = 0.0f;
+
+#pragma unroll
+    for (int k_KQ_0 = 0; k_KQ_0 < D/2; k_KQ_0 += WARP_SIZE) {
+        const int k_KQ = k_KQ_0 + threadIdx.x;
+
+        const half2 K_ik = K_h2[k_KQ];
+        sum +=  __low2float(K_ik) * Q_f2[k_KQ_0/WARP_SIZE].x;
+        sum += __high2float(K_ik) * Q_f2[k_KQ_0/WARP_SIZE].y;
+    }
+
+    return sum;
+}
+
+template <typename Tds>
+static __device__ __forceinline__ void quantize_q8_1_to_shared(
+    const float * __restrict__ x, const float scale, int * __restrict__ yq32, void * __restrict__ yds) {
+
+    float vals[sizeof(int)] = {0.0f};
+#pragma unroll
+    for (int l = 0; l < sizeof(int); ++l) {
+        vals[l] = scale * x[4*threadIdx.x + l];
+    }
+
+    float amax = fabsf(vals[0]);
+    float sum  = vals[0];
+#pragma unroll
+    for (int l = 1; l < sizeof(int); ++l) {
+        amax = fmaxf(amax, fabsf(vals[l]));
+        sum += vals[l];
+    }
+#pragma unroll
+    for (int mask = QI8_1/2; mask > 0; mask >>= 1) {
+        amax = fmaxf(amax, __shfl_xor_sync(0xFFFFFFFF, amax, mask, 32));
+        sum +=             __shfl_xor_sync(0xFFFFFFFF, sum,  mask, 32);
+    }
+
+    const float d = amax / 127;
+    int q32 = 0;
+    int8_t * q8 = (int8_t *) &q32;
+
+    if (d != 0.0f) {
+#pragma unroll
+        for (int l = 0; l < sizeof(int); ++l) {
+            q8[l] = roundf(vals[l] / d);
+        }
+    }
+
+    yq32[threadIdx.x] = q32;
+    if (threadIdx.x % QI8_1 == 0) {
+        if (std::is_same<Tds, half2>::value) {
+            ((half2  *) yds)[threadIdx.x/QI8_1] =  make_half2(d, sum);
+        } else {
+            ((float2 *) yds)[threadIdx.x/QI8_1] = make_float2(d, sum);
+        }
+    }
+}
+
+typedef half  (*dequantize_1_f16_t)(const void *, const int64_t);
+typedef float (*dequantize_1_f32_t)(const void *, const int64_t);
+
+template <typename T>
+static __device__ __forceinline__ T dequantize_1_q4_0(const void * __restrict__ vx, const int64_t i) {
+    const block_q4_0 * x = (const block_q4_0 *) vx;
+
+    const int64_t ib    =  i          /  QK4_0;
+    const int     iqs   =  i          % (QK4_0/2);
+    const int     shift = (i % QK4_0) / (QK4_0/2);
+
+    const T   d  = x[ib].d;
+    const int q0 = x[ib].qs[iqs];
+    const int q  = ((q0 >> (4*shift)) & 0x0F) - 8;
+
+#ifdef FP16_AVAILABLE
+    if (std::is_same<T, half>::value) {
+        return ((half) d)*((half) q);
+    }
+#endif // FP16_AVAILABLE
+
+    return ((float) d)*((float) q);
+}
+
+template <typename T>
+static __device__ __forceinline__ T dequantize_1_q4_1(const void * __restrict__ vx, const int64_t i) {
+    const block_q4_1 * x = (const block_q4_1 *) vx;
+
+    const int64_t ib    =  i          /  QK4_1;
+    const int     iqs   =  i          % (QK4_1/2);
+    const int     shift = (i % QK4_1) / (QK4_1/2);
+
+    const half2 dm = x[ib].dm;
+    const int   q0 = x[ib].qs[iqs];
+    const int   q  = ((q0 >> (4*shift)) & 0x0F);
+
+#ifdef FP16_AVAILABLE
+    if (std::is_same<T, half>::value) {
+        return __low2half(dm)*((half) q) + __high2half(dm);
+    }
+#endif // FP16_AVAILABLE
+
+    return __low2float(dm)*((float) q) + __high2float(dm);
+}
+
+template <typename T>
+static __device__ __forceinline__ T dequantize_1_q5_0(const void * __restrict__ vx, const int64_t i) {
+    const block_q5_0 * x = (const block_q5_0 *) vx;
+
+    const int64_t ib    =  i          /  QK5_0;
+    const int     idq   =  i          %  QK5_0;
+    const int     iqs   =  i          % (QK5_0/2);
+    const int     shift = (i % QK5_0) / (QK5_0/2);
+
+    const T   d   = x[ib].d;
+    const int ql0 = x[ib].qs[iqs];
+    const int qh0 = get_int_b2(x[ib].qh, 0);
+    const int ql  = ((ql0 >> (4*shift)) & 0x0F);
+    const int qh  = ((qh0 >> idq) << 4) & 0x10;
+    const int q   = (ql | qh) - 16;
+
+#ifdef FP16_AVAILABLE
+    if (std::is_same<T, half>::value) {
+        return ((half) d)*((half) q);
+    }
+#endif // FP16_AVAILABLE
+
+    return ((float) d)*((float) q);
+}
+
+template <typename T>
+static __device__ __forceinline__ T dequantize_1_q5_1(const void * __restrict__ vx, const int64_t i) {
+    const block_q5_1 * x = (const block_q5_1 *) vx;
+
+    const int64_t ib    =  i          /  QK5_1;
+    const int     idq   =  i          %  QK5_1;
+    const int     iqs   =  i          % (QK5_1/2);
+    const int     shift = (i % QK5_1) / (QK5_1/2);
+
+    const half2 dm  = x[ib].dm;
+    const int   ql0 = x[ib].qs[iqs];
+    const int   qh0 = get_int_b4(x[ib].qh, 0);
+    const int   ql  = ((ql0 >> (4*shift)) & 0x0F);
+    const int   qh  = ((qh0 >> idq) << 4) & 0x10;
+    const int   q   = (ql | qh);
+
+#ifdef FP16_AVAILABLE
+    if (std::is_same<T, half>::value) {
+        return __low2half(dm)*((half) q) + __high2half(dm);
+    }
+#endif // FP16_AVAILABLE
+
+    return __low2float(dm)*((float) q) + __high2float(dm);
+}
+
+template <typename T>
+static __device__ __forceinline__ T dequantize_1_q8_0(const void * __restrict__ vx, const int64_t i) {
+    const block_q8_0 * x = (const block_q8_0 *) vx;
+
+    const int64_t ib  = i / QK8_0;
+    const int     iqs = i % QK8_0;
+
+    const T   d = x[ib].d;
+    const int q = x[ib].qs[iqs];
+
+#ifdef FP16_AVAILABLE
+    if (std::is_same<T, half>::value) {
+        return ((half) d)*((half) q);
+    }
+#endif // FP16_AVAILABLE
+
+    return ((float) d)*((float) q);
+}
+
+template <typename T>
+static __device__ __forceinline__ T dequantize_1_f16(const void * __restrict__ vx, const int64_t i) {
+    const half * x = (const half *) vx;
+
+    return x[i];
+}
+
+template <int D>
+constexpr __device__ vec_dot_KQ_f16_t get_vec_dot_KQ_f16(ggml_type type_K) {
+    return type_K == GGML_TYPE_Q4_0 ? vec_dot_fattn_vec_KQ_q4_0<half, D> :
+        type_K == GGML_TYPE_Q4_1 ? vec_dot_fattn_vec_KQ_q4_1<half, D> :
+        type_K == GGML_TYPE_Q5_0 ? vec_dot_fattn_vec_KQ_q5_0<half, D> :
+        type_K == GGML_TYPE_Q5_1 ? vec_dot_fattn_vec_KQ_q5_1<half, D> :
+        type_K == GGML_TYPE_Q8_0 ? vec_dot_fattn_vec_KQ_q8_0<half, D> :
+        type_K == GGML_TYPE_F16 ? vec_dot_fattn_vec_KQ_f16<half, D> :
+        nullptr;
+}
+
+template <int D>
+constexpr __device__ vec_dot_KQ_f32_t get_vec_dot_KQ_f32(ggml_type type_K) {
+    return type_K == GGML_TYPE_Q4_0 ? vec_dot_fattn_vec_KQ_q4_0<float, D> :
+        type_K == GGML_TYPE_Q4_1 ? vec_dot_fattn_vec_KQ_q4_1<float, D> :
+        type_K == GGML_TYPE_Q5_0 ? vec_dot_fattn_vec_KQ_q5_0<float, D> :
+        type_K == GGML_TYPE_Q5_1 ? vec_dot_fattn_vec_KQ_q5_1<float, D> :
+        type_K == GGML_TYPE_Q8_0 ? vec_dot_fattn_vec_KQ_q8_0<float, D> :
+        type_K == GGML_TYPE_F16 ? vec_dot_fattn_vec_KQ_f16<float, D> :
+        nullptr;
+}
+
+constexpr __device__ dequantize_1_f16_t get_dequantize_1_f16(ggml_type type_V) {
+    return type_V == GGML_TYPE_Q4_0 ? dequantize_1_q4_0<half> :
+        type_V == GGML_TYPE_Q4_1 ? dequantize_1_q4_1<half> :
+        type_V == GGML_TYPE_Q5_0 ? dequantize_1_q5_0<half> :
+        type_V == GGML_TYPE_Q5_1 ? dequantize_1_q5_1<half> :
+        type_V == GGML_TYPE_Q8_0 ? dequantize_1_q8_0<half> :
+        type_V == GGML_TYPE_F16 ? dequantize_1_f16<half> :
+        nullptr;
+}
+
+constexpr __device__ dequantize_1_f32_t get_dequantize_1_f32(ggml_type type_V) {
+    return type_V == GGML_TYPE_Q4_0 ? dequantize_1_q4_0<float> :
+        type_V == GGML_TYPE_Q4_1 ? dequantize_1_q4_1<float> :
+        type_V == GGML_TYPE_Q5_0 ? dequantize_1_q5_0<float> :
+        type_V == GGML_TYPE_Q5_1 ? dequantize_1_q5_1<float> :
+        type_V == GGML_TYPE_Q8_0 ? dequantize_1_q8_0<float> :
+        type_V == GGML_TYPE_F16 ? dequantize_1_f16<float> :
+        nullptr;
+}
+
+template<int D, int parallel_blocks> // D == head size
+#if !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__))
+__launch_bounds__(D, 1)
+#endif // !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__))
+static __global__ void flash_attn_combine_results(
+        const float  * __restrict__ VKQ_parts,
+        const float2 * __restrict__ VKQ_meta,
+        float * __restrict__ dst) {
+    VKQ_parts += parallel_blocks*D * gridDim.y*blockIdx.x;
+    VKQ_meta  += parallel_blocks   * gridDim.y*blockIdx.x;
+    dst       +=                 D * gridDim.y*blockIdx.x;
+
+    const int tid = threadIdx.x;
+    __builtin_assume(tid < D);
+
+    __shared__ float2 meta[parallel_blocks];
+    if (tid < 2*parallel_blocks) {
+        ((float *) meta)[threadIdx.x] = ((const float *)VKQ_meta) [blockIdx.y*(2*parallel_blocks) + tid];
+    }
+
+    __syncthreads();
+
+    float kqmax = meta[0].x;
+#pragma unroll
+    for (int l = 1; l < parallel_blocks; ++l) {
+        kqmax = max(kqmax, meta[l].x);
+    }
+
+    float VKQ_numerator   = 0.0f;
+    float VKQ_denominator = 0.0f;
+#pragma unroll
+    for (int l = 0; l < parallel_blocks; ++l) {
+        const float diff = meta[l].x - kqmax;
+        const float KQ_max_scale = expf(diff);
+        const uint32_t ftz_mask = 0xFFFFFFFF * (diff > SOFTMAX_FTZ_THRESHOLD);
+        *((uint32_t *) &KQ_max_scale) &= ftz_mask;
+
+        VKQ_numerator   += KQ_max_scale * VKQ_parts[l*gridDim.y*D + blockIdx.y*D + tid];
+        VKQ_denominator += KQ_max_scale * meta[l].y;
+    }
+
+    dst[blockIdx.y*D + tid] = VKQ_numerator / VKQ_denominator;
+}
+
+static void on_no_fattn_vec_case(const int D) {
+    if (D == 64) {
+        fprintf(stderr, "Unsupported KV type combination for head_size 64.\n");
+        fprintf(stderr, "By default only f16 KV cache is supported.\n");
+        fprintf(stderr, "Compile with GGML_CUDA_FA_ALL_QUANTS for V cache quantization support.\n");
+        GGML_ABORT("fatal error");
+    } else if (D == 128) {
+        fprintf(stderr, "Unsupported KV type combination for head_size 128.\n");
+        fprintf(stderr, "Supported combinations:\n");
+        fprintf(stderr, "  - K == q4_0, V == q4_0,  4.50 BPV\n");
+        fprintf(stderr, "  - K == q8_0, V == q8_0,  8.50 BPV\n");
+        fprintf(stderr, "  - K == f16,  V == f16,  16.00 BPV\n");
+        fprintf(stderr, "Compile with GGML_CUDA_FA_ALL_QUANTS for all combinations of q4_0, q4_1, q5_0, q5_1, q8_0, and f16.\n");
+        GGML_ABORT("fatal error");
+    } else {
+        fprintf(stderr, "Unsupported KV type combination for head_size 256.\n");
+        fprintf(stderr, "Only f16 is supported.\n");
+        GGML_ABORT("fatal error");
+    }
+}
+
+template <int D, int parallel_blocks>
+void launch_fattn(
+    ggml_backend_cuda_context & ctx, ggml_tensor * dst, fattn_kernel_t fattn_kernel,
+    const int nwarps, const int cols_per_block, const bool need_f16_K, const bool need_f16_V
+) {
+    const ggml_tensor * Q = dst->src[0];
+    const ggml_tensor * K = dst->src[1];
+    const ggml_tensor * V = dst->src[2];
+
+    const ggml_tensor * mask = dst->src[3];
+
+    ggml_tensor * KQV = dst;
+
+    GGML_ASSERT(Q->type == GGML_TYPE_F32);
+    GGML_ASSERT(KQV->type == GGML_TYPE_F32);
+
+    GGML_ASSERT(!mask || mask->type == GGML_TYPE_F16);
+    GGML_ASSERT(!mask || mask->ne[1] >= GGML_PAD(Q->ne[1], 16) &&
+                                "the Flash-Attention CUDA kernel requires the mask to be padded to 16 and at least n_queries big");
+
+    GGML_ASSERT(K->ne[1] % FATTN_KQ_STRIDE == 0 && "Incorrect KV cache padding.");
+
+    ggml_cuda_pool & pool = ctx.pool();
+    cudaStream_t main_stream = ctx.stream();
+
+    ggml_cuda_pool_alloc<half>   K_f16(pool);
+    ggml_cuda_pool_alloc<half>   V_f16(pool);
+    ggml_cuda_pool_alloc<float>  dst_tmp(pool);
+    ggml_cuda_pool_alloc<float2> dst_tmp_meta(pool);
+
+    char * K_data = (char *) K->data;
+    size_t nb11 = K->nb[1];
+    size_t nb12 = K->nb[2];
+    size_t nb13 = K->nb[3];
+
+    char * V_data = (char *) V->data;
+    size_t nb21 = V->nb[1];
+    size_t nb22 = V->nb[2];
+    size_t nb23 = V->nb[3];
+
+    if (need_f16_K && K->type != GGML_TYPE_F16) {
+        K_f16.alloc(ggml_nelements(K));
+        to_fp16_cuda_t to_fp16 = ggml_get_to_fp16_cuda(K->type);
+        to_fp16(K_data, K_f16.ptr, ggml_nelements(K), main_stream);
+        K_data = (char *) K_f16.ptr;
+
+        const size_t bs = ggml_blck_size(K->type);
+        const size_t ts = ggml_type_size(K->type);
+
+        nb11 = nb11*bs*sizeof(half)/ts;
+        nb12 = nb12*bs*sizeof(half)/ts;
+        nb13 = nb13*bs*sizeof(half)/ts;
+    }
+
+    if (need_f16_V && V->type != GGML_TYPE_F16) {
+        V_f16.alloc(ggml_nelements(V));
+        to_fp16_cuda_t to_fp16 = ggml_get_to_fp16_cuda(V->type);
+        to_fp16(V_data, V_f16.ptr, ggml_nelements(V), main_stream);
+        V_data = (char *) V_f16.ptr;
+
+        const size_t bs = ggml_blck_size(V->type);
+        const size_t ts = ggml_type_size(V->type);
+
+        nb21 = nb21*bs*sizeof(half)/ts;
+        nb22 = nb22*bs*sizeof(half)/ts;
+        nb23 = nb23*bs*sizeof(half)/ts;
+    }
+
+    if (parallel_blocks > 1) {
+        dst_tmp.alloc(parallel_blocks*ggml_nelements(KQV));
+        dst_tmp_meta.alloc(parallel_blocks*ggml_nrows(KQV));
+    }
+
+    const dim3 block_dim(WARP_SIZE, nwarps, 1);
+    const dim3 blocks_num(parallel_blocks*((Q->ne[1] + cols_per_block - 1) / cols_per_block), Q->ne[2], Q->ne[3]);
+    const int  shmem = 0;
+
+    float scale         = 1.0f;
+    float max_bias      = 0.0f;
+    float logit_softcap = 0.0f;
+
+    memcpy(&scale,         (float *) KQV->op_params + 0, sizeof(float));
+    memcpy(&max_bias,      (float *) KQV->op_params + 1, sizeof(float));
+    memcpy(&logit_softcap, (float *) KQV->op_params + 2, sizeof(float));
+
+    if (logit_softcap != 0.0f) {
+        scale /= logit_softcap;
+    }
+
+    const uint32_t n_head      = Q->ne[2];
+    const uint32_t n_head_log2 = 1u << (uint32_t) floorf(log2f((float) n_head));
+
+    const float m0 = powf(2.0f, -(max_bias       ) / n_head_log2);
+    const float m1 = powf(2.0f, -(max_bias / 2.0f) / n_head_log2);
+
+    fattn_kernel<<<blocks_num, block_dim, shmem, main_stream>>>(
+        (const char *) Q->data,
+        K_data,
+        V_data,
+        mask ? ((const char *) mask->data) : nullptr,
+        (parallel_blocks) == 1 ? (float *) KQV->data : dst_tmp.ptr, dst_tmp_meta.ptr,
+        scale, max_bias, m0, m1, n_head_log2, logit_softcap,
+        Q->ne[0], Q->ne[1], Q->ne[2], Q->ne[3],
+        K->ne[0], K->ne[1], K->ne[2], K->ne[3],
+        mask ? mask->ne[1] : 0, mask ?  mask->nb[1] : 0,
+        Q->nb[1], Q->nb[2], Q->nb[3],
+        nb11, nb12, nb13,
+        nb21, nb22, nb23,
+        KQV->ne[0], KQV->ne[1], KQV->ne[2], KQV->ne[3]
+    );
+    CUDA_CHECK(cudaGetLastError());
+
+    if ((parallel_blocks) == 1) {
+        return;
+    }
+
+    const dim3 block_dim_combine(D, 1, 1);
+    const dim3 blocks_num_combine(Q->ne[1], blocks_num.y, blocks_num.z);
+    const int  shmem_combine = 0;
+
+    flash_attn_combine_results<D, parallel_blocks>
+        <<<blocks_num_combine, block_dim_combine, shmem_combine, main_stream>>>
+        (dst_tmp.ptr, dst_tmp_meta.ptr, (float *) KQV->data);
+    CUDA_CHECK(cudaGetLastError());
+}
diff --git a/src/whisper_cpp/ggml/src/ggml-cuda/fattn-tile-f16.cu b/src/whisper_cpp/ggml/src/ggml-cuda/fattn-tile-f16.cu
new file mode 100644
index 00000000..5af02c7e
--- /dev/null
+++ b/src/whisper_cpp/ggml/src/ggml-cuda/fattn-tile-f16.cu
@@ -0,0 +1,353 @@
+#include "common.cuh"
+#include "fattn-common.cuh"
+#include "fattn-tile-f16.cuh"
+
+#define FATTN_KQ_STRIDE_TILE_F16 64
+
+template<int D, int ncols, int nwarps, int parallel_blocks, bool use_logit_softcap> // D == head size
+#if !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__))
+__launch_bounds__(nwarps*WARP_SIZE, 1)
+#endif // !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__))
+static __global__ void flash_attn_tile_ext_f16(
+        const char * __restrict__ Q,
+        const char * __restrict__ K,
+        const char * __restrict__ V,
+        const char * __restrict__ mask,
+        float      * __restrict__ dst,
+        float2     * __restrict__ dst_meta,
+        const float scale,
+        const float max_bias,
+        const float m0,
+        const float m1,
+        const uint32_t n_head_log2,
+        const float logit_softcap,
+        const int ne00,
+        const int ne01,
+        const int ne02,
+        const int ne03,
+        const int ne10,
+        const int ne11,
+        const int ne12,
+        const int ne13,
+        const int ne31,
+        const int nb31,
+        const int nb01,
+        const int nb02,
+        const int nb03,
+        const int nb11,
+        const int nb12,
+        const int nb13,
+        const int nb21,
+        const int nb22,
+        const int nb23,
+        const int ne0,
+        const int ne1,
+        const int ne2,
+        const int ne3) {
+#ifdef FP16_AVAILABLE
+    // Skip unused kernel variants for faster compilation:
+    if (use_logit_softcap && !(D == 128 || D == 256)) {
+        NO_DEVICE_CODE;
+        return;
+    }
+
+    //In this kernel Q, K, V are matrices while i, j, k are matrix indices.
+
+    const int ic0 = (blockIdx.x / parallel_blocks) * ncols; // Index of the Q/QKV column to work on.
+    const int ip  =  blockIdx.x % parallel_blocks; // Index in group of blocks running for the same column in parallel.
+
+    const int gqa_ratio = ne02 / ne12; // With grouped query attention there are > 1 Q matrices per K, V matrix.
+    const float2 * Q_f2  = (const float2 *) (Q    + nb02* blockIdx.y              + nb01*ic0);
+    const half2  * K_h2  = (const half2  *) (K    + nb12*(blockIdx.y / gqa_ratio));
+    const half2  * V_h2  = (const half2  *) (V    + nb12*(blockIdx.y / gqa_ratio)); // K and V have same shape
+    const half   * maskh = (const half   *)  mask + ne11*ic0;
+
+    const int stride_KV2 = nb11 / sizeof(half2);
+
+    const float slopef = get_alibi_slope(max_bias, blockIdx.y, n_head_log2, m0, m1);
+    const half  slopeh = __float2half(slopef);
+
+    static_assert(D % (2*WARP_SIZE) == 0, "D not divisible by 2*WARP_SIZE == 64.");
+
+    __shared__ half KQ[ncols*FATTN_KQ_STRIDE_TILE_F16];
+    half2 * KQ2 = (half2 *) KQ;
+
+    __shared__ half2 KV_tmp[FATTN_KQ_STRIDE_TILE_F16][D/2 + 1]; // Pad D to avoid memory bank conflicts.
+
+    half kqmax[ncols/nwarps];
+#pragma unroll
+    for (int j0 = 0; j0 < ncols; j0 += nwarps) {
+        kqmax[j0/nwarps] = -HALF_MAX_HALF;
+    }
+    half2 kqsum[ncols/nwarps] = {{0.0f, 0.0f}};
+
+    half2 VKQ[ncols/nwarps][(D/2)/WARP_SIZE] = {{{0.0f, 0.0f}}};
+
+    // Convert Q to half2 and store in registers:
+    __shared__ half2 Q_h2[ncols][D/2];
+#pragma unroll
+    for (int j0 = 0; j0 < ncols; j0 += nwarps) {
+        const int j = j0 + threadIdx.y;
+
+#pragma unroll
+        for (int i0 = 0; i0 < D/2; i0 += WARP_SIZE) {
+            const int i = i0 + threadIdx.x;
+
+            const float2 tmp = ic0 + j < ne01 ? Q_f2[j*(nb01/sizeof(float2)) + i] : make_float2(0.0f, 0.0f);
+            Q_h2[j][i] = make_half2(scale, scale) * make_half2(tmp.x, tmp.y);
+        }
+    }
+
+    __syncthreads();
+
+    const int k_start = parallel_blocks == 1 ? 0 : ip*FATTN_KQ_STRIDE_TILE_F16;
+    for (int k_VKQ_0 = k_start; k_VKQ_0 < ne11; k_VKQ_0 += parallel_blocks*FATTN_KQ_STRIDE_TILE_F16) {
+        // Calculate KQ tile and keep track of new maximum KQ values:
+
+        half kqmax_new[ncols/nwarps];
+#pragma unroll
+        for (int j = 0; j < ncols/nwarps; ++j) {
+            kqmax_new[j] = kqmax[j];
+        }
+
+#pragma unroll
+        for (int i_KQ_0 = 0; i_KQ_0 < FATTN_KQ_STRIDE_TILE_F16; i_KQ_0 += nwarps) {
+            const int i_KQ = i_KQ_0 + threadIdx.y;
+
+#pragma unroll
+            for (int k_KQ_0 = 0; k_KQ_0 < D/2; k_KQ_0 += WARP_SIZE) {
+                const int k_KQ = k_KQ_0 + threadIdx.x;
+
+                KV_tmp[i_KQ][k_KQ] = K_h2[(k_VKQ_0 + i_KQ)*stride_KV2 + k_KQ];
+            }
+        }
+
+        __syncthreads();
+
+        half2 sum2[FATTN_KQ_STRIDE_TILE_F16/WARP_SIZE][ncols/nwarps] = {{{0.0f, 0.0f}}};
+
+#pragma unroll
+        for (int k_KQ = 0; k_KQ < D/2; ++k_KQ) {
+            half2 K_k[FATTN_KQ_STRIDE_TILE_F16/WARP_SIZE];
+            half2 Q_k[ncols/nwarps];
+
+#pragma unroll
+            for (int i_KQ_0 = 0; i_KQ_0 < FATTN_KQ_STRIDE_TILE_F16; i_KQ_0 += WARP_SIZE) {
+                const int i_KQ = i_KQ_0 + threadIdx.x;
+
+                K_k[i_KQ_0/WARP_SIZE] = KV_tmp[i_KQ][k_KQ];
+            }
+#pragma unroll
+            for (int j_KQ_0 = 0; j_KQ_0 < ncols; j_KQ_0 += nwarps) {
+                const int j_KQ = j_KQ_0 + threadIdx.y;
+
+                Q_k[j_KQ_0/nwarps] = Q_h2[j_KQ][k_KQ];
+            }
+
+#pragma unroll
+            for (int i_KQ_0 = 0; i_KQ_0 < FATTN_KQ_STRIDE_TILE_F16; i_KQ_0 += WARP_SIZE) {
+#pragma unroll
+                for (int j_KQ_0 = 0; j_KQ_0 < ncols; j_KQ_0 += nwarps) {
+                    sum2[i_KQ_0/WARP_SIZE][j_KQ_0/nwarps] += K_k[i_KQ_0/WARP_SIZE]*Q_k[j_KQ_0/nwarps];
+                }
+            }
+        }
+
+#pragma unroll
+        for (int i_KQ_0 = 0; i_KQ_0 < FATTN_KQ_STRIDE_TILE_F16; i_KQ_0 += WARP_SIZE) {
+            const int i_KQ = i_KQ_0 + threadIdx.x;
+
+#pragma unroll
+            for (int j_KQ_0 = 0; j_KQ_0 < ncols; j_KQ_0 += nwarps) {
+                const int j_KQ = j_KQ_0 + threadIdx.y;
+
+                half sum;
+                if (use_logit_softcap) {
+                    const float2 tmp = __half22float2(sum2[i_KQ_0/WARP_SIZE][j_KQ_0/nwarps]);
+                    sum = logit_softcap * tanhf(tmp.x + tmp.y);
+                } else {
+                    sum = __low2half(sum2[i_KQ_0/WARP_SIZE][j_KQ_0/nwarps]) + __high2half(sum2[i_KQ_0/WARP_SIZE][j_KQ_0/nwarps]);
+                }
+                sum += mask ? slopeh*maskh[j_KQ*ne11 + k_VKQ_0 + i_KQ] : __float2half(0.0f);
+
+                kqmax_new[j_KQ_0/nwarps] = ggml_cuda_hmax(kqmax_new[j_KQ_0/nwarps], sum);
+
+                KQ[j_KQ*FATTN_KQ_STRIDE_TILE_F16 + i_KQ] = sum;
+            }
+        }
+
+        __syncthreads();
+
+#pragma unroll
+        for (int j0 = 0; j0 < ncols; j0 += nwarps) {
+            const int j = j0 + threadIdx.y;
+
+            kqmax_new[j0/nwarps] = warp_reduce_max(kqmax_new[j0/nwarps]);
+            const half2 KQ_max_scale = __half2half2(hexp(kqmax[j0/nwarps] - kqmax_new[j0/nwarps]));
+            kqmax[j0/nwarps] = kqmax_new[j0/nwarps];
+
+#pragma unroll
+            for (int i0 = 0; i0 < FATTN_KQ_STRIDE_TILE_F16/2; i0 += WARP_SIZE) {
+                const int i = i0 + threadIdx.x;
+
+                const half2 diff = KQ2[j*(FATTN_KQ_STRIDE_TILE_F16/2) + i] - __half2half2(kqmax[j0/nwarps]);
+                const half2 val = h2exp(diff);
+                kqsum[j0/nwarps] = kqsum[j0/nwarps]*KQ_max_scale + val;
+                KQ2[j*(FATTN_KQ_STRIDE_TILE_F16/2) + i] = val;
+            }
+
+#pragma unroll
+            for (int i0 = 0; i0 < D/2; i0 += WARP_SIZE) {
+                VKQ[j0/nwarps][i0/WARP_SIZE] *= KQ_max_scale;
+            }
+        }
+
+        __syncthreads();
+
+#pragma unroll
+        for (int k0 = 0; k0 < FATTN_KQ_STRIDE_TILE_F16; k0 += nwarps) {
+            const int k = k0 + threadIdx.y;
+
+#pragma unroll
+            for (int i0 = 0; i0 < D/2; i0 += WARP_SIZE) {
+                const int i = i0 + threadIdx.x;
+
+                KV_tmp[k][i] = V_h2[(k_VKQ_0 + k)*stride_KV2 + i];
+            }
+        }
+
+        __syncthreads();
+
+#pragma unroll
+        for (int k0 = 0; k0 < FATTN_KQ_STRIDE_TILE_F16; k0 += 2) {
+            half2  V_k[(D/2)/WARP_SIZE][2];
+            half2 KQ_k[ncols/nwarps];
+
+#pragma unroll
+            for (int i0 = 0; i0 < D/2; i0 += WARP_SIZE) {
+                const int i = i0 + threadIdx.x;
+
+                V_k[i0/WARP_SIZE][0] = KV_tmp[k0 + 0][i];
+                V_k[i0/WARP_SIZE][1] = KV_tmp[k0 + 1][i];
+            }
+#pragma unroll
+            for (int j0 = 0; j0 < ncols; j0 += nwarps) {
+                const int j = j0 + threadIdx.y;
+
+                KQ_k[j0/nwarps] = KQ2[j*(FATTN_KQ_STRIDE_TILE_F16/2) + k0/2];
+            }
+
+#pragma unroll
+            for (int i0 = 0; i0 < D/2; i0 += WARP_SIZE) {
+#pragma unroll
+                for (int j0 = 0; j0 < ncols; j0 += nwarps) {
+                    VKQ[j0/nwarps][i0/WARP_SIZE] += V_k[i0/WARP_SIZE][0]* __low2half2(KQ_k[j0/nwarps]);
+                    VKQ[j0/nwarps][i0/WARP_SIZE] += V_k[i0/WARP_SIZE][1]*__high2half2(KQ_k[j0/nwarps]);
+                }
+            }
+        }
+
+        __syncthreads();
+    }
+
+#pragma unroll
+    for (int j_VKQ_0 = 0; j_VKQ_0 < ncols; j_VKQ_0 += nwarps) {
+        const int j_VKQ = j_VKQ_0 + threadIdx.y;
+
+        if (ic0 + j_VKQ >= ne01) {
+            return;
+        }
+
+        half kqsum_j = __low2half(kqsum[j_VKQ_0/nwarps]) + __high2half(kqsum[j_VKQ_0/nwarps]);
+        kqsum_j = warp_reduce_sum((float)kqsum_j);
+
+#pragma unroll
+        for (int i00 = 0; i00 < D; i00 += 2*WARP_SIZE) {
+            const int i0 = i00 + 2*threadIdx.x;
+
+            half2 dst_val = VKQ[j_VKQ_0/nwarps][i0/(2*WARP_SIZE)];
+            if (parallel_blocks == 1) {
+                dst_val /= __half2half2(kqsum_j);
+            }
+            const int j_dst = (ic0 + j_VKQ)*parallel_blocks + ip;
+            dst[j_dst*D*gridDim.y + D*blockIdx.y + i0 + 0] =  __low2float(dst_val);
+            dst[j_dst*D*gridDim.y + D*blockIdx.y + i0 + 1] = __high2float(dst_val);
+        }
+
+        if (parallel_blocks != 1 && threadIdx.x == 0) {
+            dst_meta[(ic0 + j_VKQ)*gridDim.y*parallel_blocks + blockIdx.y*parallel_blocks + ip] = make_float2(kqmax[j_VKQ_0/nwarps], kqsum_j);
+        }
+    }
+#else
+   NO_DEVICE_CODE;
+#endif // FP16_AVAILABLE
+}
+
+template <int cols_per_block, int parallel_blocks, bool use_logit_softcap>
+void launch_fattn_tile_f16_64_128(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
+    const ggml_tensor * Q = dst->src[0];
+    switch (Q->ne[0]) {
+        case  64: {
+            constexpr int      D = 64;
+            constexpr int nwarps = 8;
+            fattn_kernel_t fattn_kernel = flash_attn_tile_ext_f16<D, cols_per_block, nwarps, parallel_blocks, use_logit_softcap>;
+            launch_fattn<D, parallel_blocks>(ctx, dst, fattn_kernel, nwarps, cols_per_block, true, true);
+        } break;
+        case 128: {
+            constexpr int      D = 128;
+            constexpr int nwarps = 8;
+            fattn_kernel_t fattn_kernel = flash_attn_tile_ext_f16<D, cols_per_block, nwarps, parallel_blocks, use_logit_softcap>;
+            launch_fattn<D, parallel_blocks>(ctx, dst, fattn_kernel, nwarps, cols_per_block, true, true);
+        } break;
+        default: {
+            GGML_ABORT("FlashAttention without tensor cores only supports head sizes 64 and 128.");
+        } break;
+    }
+}
+
+void ggml_cuda_flash_attn_ext_tile_f16(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
+    const ggml_tensor * KQV = dst;
+    const ggml_tensor * Q   = dst->src[0];
+
+    const int32_t precision = KQV->op_params[3];
+    GGML_ASSERT(precision == GGML_PREC_DEFAULT);
+
+    float logit_softcap;
+    memcpy(&logit_softcap, (const float *) KQV->op_params + 2, sizeof(float));
+
+    if (Q->ne[1] <= 16) {
+        constexpr int cols_per_block = 16;
+        constexpr int parallel_blocks = 4;
+        if (logit_softcap == 0.0f) {
+            constexpr bool use_logit_softcap = false;
+            launch_fattn_tile_f16_64_128<cols_per_block, parallel_blocks, use_logit_softcap>(ctx, dst);
+        } else {
+            constexpr bool use_logit_softcap = true;
+            launch_fattn_tile_f16_64_128<cols_per_block, parallel_blocks, use_logit_softcap>(ctx, dst);
+        }
+        return;
+    }
+
+    if (Q->ne[1] <= 32) {
+        constexpr int cols_per_block = 32;
+        constexpr int parallel_blocks = 4;
+        if (logit_softcap == 0.0f) {
+            constexpr bool use_logit_softcap = false;
+            launch_fattn_tile_f16_64_128<cols_per_block, parallel_blocks, use_logit_softcap>(ctx, dst);
+        } else {
+            constexpr bool use_logit_softcap = true;
+            launch_fattn_tile_f16_64_128<cols_per_block, parallel_blocks, use_logit_softcap>(ctx, dst);
+        }
+        return;
+    }
+
+    constexpr int cols_per_block = 32;
+    constexpr int parallel_blocks = 1;
+    if (logit_softcap == 0.0f) {
+        constexpr bool use_logit_softcap = false;
+        launch_fattn_tile_f16_64_128<cols_per_block, parallel_blocks, use_logit_softcap>(ctx, dst);
+    } else {
+        constexpr bool use_logit_softcap = true;
+        launch_fattn_tile_f16_64_128<cols_per_block, parallel_blocks, use_logit_softcap>(ctx, dst);
+    }
+}
diff --git a/src/whisper_cpp/ggml/src/ggml-cuda/fattn-tile-f16.cuh b/src/whisper_cpp/ggml/src/ggml-cuda/fattn-tile-f16.cuh
new file mode 100644
index 00000000..ffc58784
--- /dev/null
+++ b/src/whisper_cpp/ggml/src/ggml-cuda/fattn-tile-f16.cuh
@@ -0,0 +1,3 @@
+#include "common.cuh"
+
+void ggml_cuda_flash_attn_ext_tile_f16(ggml_backend_cuda_context & ctx, ggml_tensor * dst);
diff --git a/src/whisper_cpp/ggml/src/ggml-cuda/fattn-tile-f32.cu b/src/whisper_cpp/ggml/src/ggml-cuda/fattn-tile-f32.cu
new file mode 100644
index 00000000..f402195c
--- /dev/null
+++ b/src/whisper_cpp/ggml/src/ggml-cuda/fattn-tile-f32.cu
@@ -0,0 +1,349 @@
+#include "common.cuh"
+#include "fattn-common.cuh"
+#include "fattn-tile-f32.cuh"
+
+#define FATTN_KQ_STRIDE_TILE_F32 32
+
+template<int D, int ncols, int nwarps, int parallel_blocks, bool use_logit_softcap> // D == head size
+#if !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__))
+__launch_bounds__(nwarps*WARP_SIZE, 1)
+#endif // !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__))
+static __global__ void flash_attn_tile_ext_f32(
+        const char * __restrict__ Q,
+        const char * __restrict__ K,
+        const char * __restrict__ V,
+        const char * __restrict__ mask,
+        float      * __restrict__ dst,
+        float2     * __restrict__ dst_meta,
+        const float scale,
+        const float max_bias,
+        const float m0,
+        const float m1,
+        const uint32_t n_head_log2,
+        const float logit_softcap,
+        const int ne00,
+        const int ne01,
+        const int ne02,
+        const int ne03,
+        const int ne10,
+        const int ne11,
+        const int ne12,
+        const int ne13,
+        const int ne31,
+        const int nb31,
+        const int nb01,
+        const int nb02,
+        const int nb03,
+        const int nb11,
+        const int nb12,
+        const int nb13,
+        const int nb21,
+        const int nb22,
+        const int nb23,
+        const int ne0,
+        const int ne1,
+        const int ne2,
+        const int ne3) {
+#ifndef FLASH_ATTN_AVAILABLE
+    NO_DEVICE_CODE;
+    return;
+#endif // FLASH_ATTN_AVAILABLE
+    // Skip unused kernel variants for faster compilation:
+    if (use_logit_softcap && !(D == 128 || D == 256)) {
+        NO_DEVICE_CODE;
+        return;
+    }
+
+    // In this kernel Q, K, V are matrices while i, j, k are matrix indices.
+
+    const int ic0 = (blockIdx.x / parallel_blocks) * ncols; // Index of the Q/QKV column to work on.
+    const int ip  =  blockIdx.x % parallel_blocks; // Index in group of blocks running for the same column in parallel.
+
+    const int gqa_ratio = ne02 / ne12; // With grouped query attention there are > 1 Q matrices per K, V matrix.
+    const float2 * Q_f2  = (const float2 *) (Q    + nb02* blockIdx.y              + nb01*ic0);
+    const half2  * K_h2  = (const half2  *) (K    + nb12*(blockIdx.y / gqa_ratio));
+    const half2  * V_h2  = (const half2  *) (V    + nb12*(blockIdx.y / gqa_ratio)); // K and V have same shape
+    const half   * maskh = (const half   *)  mask + ne11*ic0;
+
+    const int stride_KV2 = nb11 / sizeof(half2);
+
+    const float slope = get_alibi_slope(max_bias, blockIdx.y, n_head_log2, m0, m1);
+
+    static_assert(D % (2*WARP_SIZE) == 0, "D not divisible by 2*WARP_SIZE == 64.");
+
+    __shared__ float KQ[ncols*FATTN_KQ_STRIDE_TILE_F32];
+
+    __shared__ float KV_tmp[FATTN_KQ_STRIDE_TILE_F32][D + 1]; // Pad D to avoid memory bank conflicts.
+    float2 * KV_tmp2 = (float2 *) KV_tmp;
+
+    float kqmax[ncols/nwarps];
+#pragma unroll
+    for (int j0 = 0; j0 < ncols; j0 += nwarps) {
+        kqmax[j0/nwarps] = -FLT_MAX/2.0f;
+    }
+    float kqsum[ncols/nwarps] = {0.0f};
+
+    float2 VKQ[ncols/nwarps][(D/2)/WARP_SIZE] = {{{0.0f, 0.0f}}};
+
+    // Convert Q to half2 and store in registers:
+    __shared__ float Q_f[ncols][D];
+#pragma unroll
+    for (int j0 = 0; j0 < ncols; j0 += nwarps) {
+        const int j = j0 + threadIdx.y;
+
+#pragma unroll
+        for (int i0 = 0; i0 < D; i0 += 2*WARP_SIZE) {
+            float2 tmp = ic0 + j < ne01 ? Q_f2[j*(nb01/sizeof(float2)) + i0/2 + threadIdx.x] : make_float2(0.0f, 0.0f);
+            Q_f[j][i0 + 0*WARP_SIZE + threadIdx.x] = tmp.x * scale;
+            Q_f[j][i0 + 1*WARP_SIZE + threadIdx.x] = tmp.y * scale;
+        }
+    }
+
+    __syncthreads();
+
+    const int k_start = parallel_blocks == 1 ? 0 : ip*FATTN_KQ_STRIDE_TILE_F32;
+    for (int k_VKQ_0 = k_start; k_VKQ_0 < ne11; k_VKQ_0 += parallel_blocks*FATTN_KQ_STRIDE_TILE_F32) {
+        // Calculate KQ tile and keep track of new maximum KQ values:
+
+        float kqmax_new[ncols/nwarps];
+#pragma unroll
+        for (int j = 0; j < ncols/nwarps; ++j) {
+            kqmax_new[j] = kqmax[j];
+        }
+
+#pragma unroll
+        for (int i_KQ_0 = 0; i_KQ_0 < FATTN_KQ_STRIDE_TILE_F32; i_KQ_0 += nwarps) {
+            const int i_KQ = i_KQ_0 + threadIdx.y;
+
+#pragma unroll
+            for (int k_KQ_0 = 0; k_KQ_0 < D; k_KQ_0 += 2*WARP_SIZE) {
+                const half2 tmp = K_h2[(k_VKQ_0 + i_KQ)*stride_KV2 + k_KQ_0/2 + threadIdx.x];
+                KV_tmp[i_KQ][k_KQ_0 + 0*WARP_SIZE + threadIdx.x] =  __low2float(tmp);
+                KV_tmp[i_KQ][k_KQ_0 + 1*WARP_SIZE + threadIdx.x] = __high2float(tmp);
+            }
+        }
+
+        __syncthreads();
+
+        float sum[FATTN_KQ_STRIDE_TILE_F32/WARP_SIZE][ncols/nwarps] = {{0.0f}};
+
+#pragma unroll
+        for (int k_KQ = 0; k_KQ < D; ++k_KQ) {
+            float K_k[FATTN_KQ_STRIDE_TILE_F32/WARP_SIZE];
+            float Q_k[ncols/nwarps];
+
+#pragma unroll
+            for (int i_KQ_0 = 0; i_KQ_0 < FATTN_KQ_STRIDE_TILE_F32; i_KQ_0 += WARP_SIZE) {
+                const int i_KQ = i_KQ_0 + threadIdx.x;
+
+                K_k[i_KQ_0/WARP_SIZE] = KV_tmp[i_KQ][k_KQ];
+            }
+#pragma unroll
+            for (int j_KQ_0 = 0; j_KQ_0 < ncols; j_KQ_0 += nwarps) {
+                const int j_KQ = j_KQ_0 + threadIdx.y;
+
+                Q_k[j_KQ_0/nwarps] = Q_f[j_KQ][k_KQ];
+            }
+
+#pragma unroll
+            for (int i_KQ_0 = 0; i_KQ_0 < FATTN_KQ_STRIDE_TILE_F32; i_KQ_0 += WARP_SIZE) {
+#pragma unroll
+                for (int j_KQ_0 = 0; j_KQ_0 < ncols; j_KQ_0 += nwarps) {
+                    sum[i_KQ_0/WARP_SIZE][j_KQ_0/nwarps] += K_k[i_KQ_0/WARP_SIZE] * Q_k[j_KQ_0/nwarps];
+                }
+            }
+        }
+
+#pragma unroll
+        for (int i_KQ_0 = 0; i_KQ_0 < FATTN_KQ_STRIDE_TILE_F32; i_KQ_0 += WARP_SIZE) {
+            const int i_KQ = i_KQ_0 + threadIdx.x;
+
+#pragma unroll
+            for (int j_KQ_0 = 0; j_KQ_0 < ncols; j_KQ_0 += nwarps) {
+                const int j_KQ = j_KQ_0 + threadIdx.y;
+
+                if (use_logit_softcap) {
+                    sum[i_KQ_0/WARP_SIZE][j_KQ_0/nwarps] = logit_softcap * tanhf(sum[i_KQ_0/WARP_SIZE][j_KQ_0/nwarps]);
+                }
+
+                sum[i_KQ_0/WARP_SIZE][j_KQ_0/nwarps] += mask ? slope*__half2float(maskh[j_KQ*ne11 + k_VKQ_0 + i_KQ]) : 0.0f;
+
+                kqmax_new[j_KQ_0/nwarps] = fmaxf(kqmax_new[j_KQ_0/nwarps], sum[i_KQ_0/WARP_SIZE][j_KQ_0/nwarps]);
+
+                KQ[j_KQ*FATTN_KQ_STRIDE_TILE_F32 + i_KQ] = sum[i_KQ_0/WARP_SIZE][j_KQ_0/nwarps];
+            }
+        }
+
+        __syncthreads();
+
+#pragma unroll
+        for (int j0 = 0; j0 < ncols; j0 += nwarps) {
+            const int j = j0 + threadIdx.y;
+
+            kqmax_new[j0/nwarps] = warp_reduce_max(kqmax_new[j0/nwarps]);
+            const float KQ_max_scale = expf(kqmax[j0/nwarps] - kqmax_new[j0/nwarps]);
+            kqmax[j0/nwarps] = kqmax_new[j0/nwarps];
+
+            float kqsum_add = 0.0f;
+#pragma unroll
+            for (int i0 = 0; i0 < FATTN_KQ_STRIDE_TILE_F32; i0 += WARP_SIZE) {
+                const int i = i0 + threadIdx.x;
+
+                const float diff = KQ[j*FATTN_KQ_STRIDE_TILE_F32 + i] - kqmax[j0/nwarps];
+                const float val = expf(diff);
+                kqsum_add += val;
+                KQ[j*FATTN_KQ_STRIDE_TILE_F32 + i] = val;
+            }
+            kqsum[j0/nwarps] = kqsum[j0/nwarps]*KQ_max_scale + kqsum_add;
+
+#pragma unroll
+            for (int i0 = 0; i0 < D/2; i0 += WARP_SIZE) {
+                VKQ[j0/nwarps][i0/WARP_SIZE].x *= KQ_max_scale;
+                VKQ[j0/nwarps][i0/WARP_SIZE].y *= KQ_max_scale;
+            }
+        }
+
+        __syncthreads();
+
+#pragma unroll
+        for (int k0 = 0; k0 < FATTN_KQ_STRIDE_TILE_F32; k0 += nwarps) {
+            const int k = k0 + threadIdx.y;
+
+#pragma unroll
+            for (int i0 = 0; i0 < D/2; i0 += WARP_SIZE) {
+                const int i = i0 + threadIdx.x;
+
+                KV_tmp2[k*(D/2) + i].x =  __low2float(V_h2[(k_VKQ_0 + k)*stride_KV2 + i]);
+                KV_tmp2[k*(D/2) + i].y = __high2float(V_h2[(k_VKQ_0 + k)*stride_KV2 + i]);
+            }
+        }
+
+        __syncthreads();
+
+#pragma unroll
+        for (int k = 0; k < FATTN_KQ_STRIDE_TILE_F32; ++k) {
+            float2 V_k[(D/2)/WARP_SIZE];
+            float  KQ_k[ncols/nwarps];
+
+#pragma unroll
+            for (int i0 = 0; i0 < D/2; i0 += WARP_SIZE) {
+                const int i = i0 + threadIdx.x;
+
+                V_k[i0/WARP_SIZE] = KV_tmp2[k*(D/2) + i];
+            }
+#pragma unroll
+            for (int j0 = 0; j0 < ncols; j0 += nwarps) {
+                const int j = j0 + threadIdx.y;
+
+                KQ_k[j0/nwarps] = KQ[j*FATTN_KQ_STRIDE_TILE_F32 + k];
+            }
+
+#pragma unroll
+            for (int i0 = 0; i0 < D/2; i0 += WARP_SIZE) {
+#pragma unroll
+                for (int j0 = 0; j0 < ncols; j0 += nwarps) {
+                    VKQ[j0/nwarps][i0/WARP_SIZE].x += V_k[i0/WARP_SIZE].x*KQ_k[j0/nwarps];
+                    VKQ[j0/nwarps][i0/WARP_SIZE].y += V_k[i0/WARP_SIZE].y*KQ_k[j0/nwarps];
+                }
+            }
+        }
+
+        __syncthreads();
+    }
+
+#pragma unroll
+    for (int j_VKQ_0 = 0; j_VKQ_0 < ncols; j_VKQ_0 += nwarps) {
+        const int j_VKQ = j_VKQ_0 + threadIdx.y;
+
+        if (ic0 + j_VKQ >= ne01) {
+            return;
+        }
+
+        float kqsum_j = kqsum[j_VKQ_0/nwarps];
+        kqsum_j = warp_reduce_sum(kqsum_j);
+
+#pragma unroll
+        for (int i00 = 0; i00 < D; i00 += 2*WARP_SIZE) {
+            const int i0 = i00 + 2*threadIdx.x;
+
+            float2 dst_val = VKQ[j_VKQ_0/nwarps][i0/(2*WARP_SIZE)];
+            if (parallel_blocks == 1) {
+                dst_val.x /= kqsum_j;
+                dst_val.y /= kqsum_j;
+            }
+            const int j_dst = (ic0 + j_VKQ)*parallel_blocks + ip;
+            dst[j_dst*D*gridDim.y + D*blockIdx.y + i0 + 0] = dst_val.x;
+            dst[j_dst*D*gridDim.y + D*blockIdx.y + i0 + 1] = dst_val.y;
+        }
+
+        if (parallel_blocks != 1 && threadIdx.x == 0) {
+            dst_meta[(ic0 + j_VKQ)*gridDim.y*parallel_blocks + blockIdx.y*parallel_blocks + ip] = make_float2(kqmax[j_VKQ_0/nwarps], kqsum_j);
+        }
+    }
+}
+
+template <int cols_per_block, int parallel_blocks, bool use_logit_softcap>
+void launch_fattn_tile_f32_64_128(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
+    const ggml_tensor * Q = dst->src[0];
+    switch (Q->ne[0]) {
+        case  64: {
+            constexpr int      D = 64;
+            constexpr int nwarps = 8;
+            fattn_kernel_t fattn_kernel = flash_attn_tile_ext_f32<D, cols_per_block, nwarps, parallel_blocks, use_logit_softcap>;
+            launch_fattn<D, parallel_blocks>(ctx, dst, fattn_kernel, nwarps, cols_per_block, true, true);
+        } break;
+        case 128: {
+            constexpr int      D = 128;
+            constexpr int nwarps = 8;
+            fattn_kernel_t fattn_kernel = flash_attn_tile_ext_f32<D, cols_per_block, nwarps, parallel_blocks, use_logit_softcap>;
+            launch_fattn<D, parallel_blocks>(ctx, dst, fattn_kernel, nwarps, cols_per_block, true, true);
+        } break;
+        default: {
+            GGML_ABORT("FlashAttention without tensor cores only supports head sizes 64 and 128.");
+        } break;
+    }
+}
+
+void ggml_cuda_flash_attn_ext_tile_f32(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
+    const ggml_tensor * KQV = dst;
+    const ggml_tensor * Q = dst->src[0];
+
+    float logit_softcap;
+    memcpy(&logit_softcap, (const float *) KQV->op_params + 2, sizeof(float));
+
+    if (Q->ne[1] <= 16) {
+        constexpr int cols_per_block = 16;
+        constexpr int parallel_blocks = 4;
+        if (logit_softcap == 0.0f) {
+            constexpr bool use_logit_softcap = false;
+            launch_fattn_tile_f32_64_128<cols_per_block, parallel_blocks, use_logit_softcap>(ctx, dst);
+        } else {
+            constexpr bool use_logit_softcap = true;
+            launch_fattn_tile_f32_64_128<cols_per_block, parallel_blocks, use_logit_softcap>(ctx, dst);
+        }
+        return;
+    }
+
+    if (Q->ne[1] <= 32) {
+        constexpr int cols_per_block = 32;
+        constexpr int parallel_blocks = 4;
+        if (logit_softcap == 0.0f) {
+            constexpr bool use_logit_softcap = false;
+            launch_fattn_tile_f32_64_128<cols_per_block, parallel_blocks, use_logit_softcap>(ctx, dst);
+        } else {
+            constexpr bool use_logit_softcap = true;
+            launch_fattn_tile_f32_64_128<cols_per_block, parallel_blocks, use_logit_softcap>(ctx, dst);
+        }
+        return;
+    }
+
+    constexpr int cols_per_block = 32;
+    constexpr int parallel_blocks = 1;
+    if (logit_softcap == 0.0f) {
+        constexpr bool use_logit_softcap = false;
+        launch_fattn_tile_f32_64_128<cols_per_block, parallel_blocks, use_logit_softcap>(ctx, dst);
+    } else {
+        constexpr bool use_logit_softcap = true;
+        launch_fattn_tile_f32_64_128<cols_per_block, parallel_blocks, use_logit_softcap>(ctx, dst);
+    }
+}
diff --git a/src/whisper_cpp/ggml/src/ggml-cuda/fattn-tile-f32.cuh b/src/whisper_cpp/ggml/src/ggml-cuda/fattn-tile-f32.cuh
new file mode 100644
index 00000000..b1c546c8
--- /dev/null
+++ b/src/whisper_cpp/ggml/src/ggml-cuda/fattn-tile-f32.cuh
@@ -0,0 +1,3 @@
+#include "common.cuh"
+
+void ggml_cuda_flash_attn_ext_tile_f32(ggml_backend_cuda_context & ctx, ggml_tensor * dst);
diff --git a/src/whisper_cpp/ggml/src/ggml-cuda/fattn-vec-f16.cuh b/src/whisper_cpp/ggml/src/ggml-cuda/fattn-vec-f16.cuh
new file mode 100644
index 00000000..2ed6509a
--- /dev/null
+++ b/src/whisper_cpp/ggml/src/ggml-cuda/fattn-vec-f16.cuh
@@ -0,0 +1,442 @@
+#include "common.cuh"
+#include "fattn-common.cuh"
+
+template<int D, int ncols, int parallel_blocks, ggml_type type_K, ggml_type type_V, bool use_logit_softcap> // D == head size
+#if !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__))
+__launch_bounds__(D, 1)
+#endif // !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__))
+static __global__ void flash_attn_vec_ext_f16(
+        const char * __restrict__ Q,
+        const char * __restrict__ K,
+        const char * __restrict__ V,
+        const char * __restrict__ mask,
+        float      * __restrict__ dst,
+        float2     * __restrict__ dst_meta,
+        const float scale,
+        const float max_bias,
+        const float m0,
+        const float m1,
+        const uint32_t n_head_log2,
+        const float logit_softcap,
+        const int ne00,
+        const int ne01,
+        const int ne02,
+        const int ne03,
+        const int ne10,
+        const int ne11,
+        const int ne12,
+        const int ne13,
+        const int ne31,
+        const int nb31,
+        const int nb01,
+        const int nb02,
+        const int nb03,
+        const int nb11,
+        const int nb12,
+        const int nb13,
+        const int nb21,
+        const int nb22,
+        const int nb23,
+        const int ne0,
+        const int ne1,
+        const int ne2,
+        const int ne3) {
+#ifdef FP16_AVAILABLE
+    // Skip unused kernel variants for faster compilation:
+    if (use_logit_softcap && !(D == 128 || D == 256)) {
+        NO_DEVICE_CODE;
+        return;
+    }
+
+    //In this kernel Q, K, V are matrices while i, j, k are matrix indices.
+
+    constexpr vec_dot_KQ_f16_t vec_dot_KQ = get_vec_dot_KQ_f16<D>(type_K);
+    constexpr bool Q_q8_1 = type_K != GGML_TYPE_F16;
+    constexpr dequantize_1_f16_t dequantize_1_v = get_dequantize_1_f16(type_V);
+
+    const int ic0 = (blockIdx.x / parallel_blocks) * ncols; // Index of the Q/QKV column to work on.
+    const int ip  =  blockIdx.x % parallel_blocks; // Index in group of blocks running for the same column in parallel.
+
+    const int gqa_ratio = ne02 / ne12; // With grouped query attention there are > 1 Q matrices per K, V matrix.
+    Q += nb02* blockIdx.y              + nb01*ic0;
+    K += nb12*(blockIdx.y / gqa_ratio);
+    V += nb22*(blockIdx.y / gqa_ratio);
+
+    const half * maskh = (const half   *)  mask + ne11*ic0;
+
+    const float slopef = get_alibi_slope(max_bias, blockIdx.y, n_head_log2, m0, m1);
+    const half  slopeh = __float2half(slopef);
+
+    static_assert(D % (2*WARP_SIZE) == 0, "D not divisible by 2*WARP_SIZE == 64.");
+    constexpr int nwarps = D / WARP_SIZE;
+    const int tid = WARP_SIZE*threadIdx.y + threadIdx.x;
+    __builtin_assume(tid < D);
+
+    __shared__ half KQ[ncols*D];
+    half2 * KQ2 = (half2 *) KQ;
+
+    half kqmax[ncols];
+#pragma unroll
+    for (int j = 0; j < ncols; ++j) {
+        kqmax[j] = -HALF_MAX_HALF;
+    }
+    half kqsum[ncols] = {0.0f};
+
+    __shared__ half kqmax_shared[ncols][WARP_SIZE];
+    __shared__ half kqsum_shared[ncols][WARP_SIZE];
+#pragma unroll
+    for (int j = 0; j < ncols; ++j) {
+        if (threadIdx.y == 0) {
+            kqmax_shared[j][threadIdx.x] = -HALF_MAX_HALF;
+            kqsum_shared[j][threadIdx.x] = 0.0f;
+        }
+    }
+    __syncthreads();
+
+    // Convert Q to half2 (f16 K) or q8_1 (quantized K) and store in registers:
+    half2  Q_h2[ncols][D/(2*WARP_SIZE)];
+    int   Q_i32[ncols][D/(sizeof(int)*QK8_1) == 0 ? 1 : D/(sizeof(int)*QK8_1)];
+    half2  Q_ds[ncols][D/QK8_1 == 0 ? 1 : D/QK8_1];
+    if (Q_q8_1) {
+#pragma unroll
+        for (int j0 = 0; j0 < ncols; j0 += nwarps) {
+            const int j = j0 + threadIdx.y;
+
+            if (j0 + nwarps > ncols && j >= ncols) {
+                break;
+            }
+
+            // Reuse KQ as temporary storage for converting Q to q8_1:
+            int   * tmp_q_i32 = (int   *) &KQ[j*D];
+            half2 * tmp_q_ds  = (half2 *) (tmp_q_i32 + D/sizeof(int));
+
+            // Set memory to zero if out of bounds:
+            if (ncols > 2 && ic0 + j >= ne01) {
+#pragma unroll
+                for (int i0 = 0; i0 < D/sizeof(int); i0 += WARP_SIZE) {
+                    const int i = i0 + threadIdx.x;
+
+                    tmp_q_i32[i] = 0;
+                }
+                if (threadIdx.x < D/QK8_1) {
+                    tmp_q_ds[threadIdx.x] = make_half2(0.0f, 0.0f);
+                }
+                continue;
+            }
+
+            const float * Q_f = (const float *) (Q + j*nb01);
+#pragma unroll
+            for (int i0 = 0; i0 < D/sizeof(int); i0 += WARP_SIZE) {
+                quantize_q8_1_to_shared<half2>(Q_f + 4*i0, scale, tmp_q_i32, tmp_q_ds);
+            }
+        }
+
+        __syncthreads();
+
+#pragma unroll
+        for (int j = 0; j < ncols; ++j) {
+            int   * tmp_q_i32 = (int   *) &KQ[j*D];
+            half2 * tmp_q_ds  = (half2 *) (tmp_q_i32 + D/sizeof(int));
+
+#pragma unroll
+            for (int i0 = 0; i0 < D/sizeof(int); i0 += WARP_SIZE) {
+                const int i = i0 + threadIdx.x;
+
+                Q_i32[j][i0/WARP_SIZE] = tmp_q_i32[i];
+                Q_ds[j][i0/WARP_SIZE]  = tmp_q_ds[i/QI8_1];
+            }
+        }
+
+        __syncthreads();
+    } else {
+#pragma unroll
+        for (int j = 0; j < ncols; ++j) {
+            const float2 * Q_f2_j = (const float2 *) (Q + j*nb01);
+
+#pragma unroll
+            for (int i0 = 0; i0 < D/2; i0 += WARP_SIZE) {
+                const int i = i0 + threadIdx.x;
+
+                const float2 tmp = ncols <= 2 || ic0 + j < ne01 ? Q_f2_j[i] : make_float2(0.0f, 0.0f);
+                Q_h2[j][i0/WARP_SIZE] = make_half2(scale, scale) * make_half2(tmp.x, tmp.y);
+            }
+        }
+    }
+
+
+#pragma unroll
+    for (int j = 0; j < ncols; ++j) {
+        KQ[j*D + tid] = -HALF_MAX_HALF;
+    }
+
+    half2 VKQ[ncols] = {{0.0f, 0.0f}};
+
+    const int k_start = parallel_blocks == 1 ? 0 : ip*D;
+    for (int k_VKQ_0 = k_start; k_VKQ_0 < ne11; k_VKQ_0 += parallel_blocks*D) {
+        // Calculate KQ tile and keep track of new maximum KQ values:
+
+        // For unknown reasons using a half array of size 1 for kqmax_new causes a performance regression,
+        // see https://github.com/ggerganov/llama.cpp/pull/7061 .
+        // Therefore this variable is defined twice but only used once (so that the compiler can optimize out the unused variable).
+        half kqmax_new = kqmax[0];
+        half kqmax_new_arr[ncols];
+#pragma unroll
+        for (int j = 0; j < ncols; ++j) {
+            kqmax_new_arr[j] = kqmax[j];
+        }
+
+#pragma unroll
+        for (int i_KQ_0 = 0; i_KQ_0 < D; i_KQ_0 += nwarps) {
+            const int i_KQ = i_KQ_0 + threadIdx.y;
+
+            if ((i_KQ_0 + nwarps > D && i_KQ >= D) || (FATTN_KQ_STRIDE % D != 0 && k_VKQ_0 + i_KQ >= ne11)) {
+                break;
+            }
+
+#pragma unroll
+            for (int j = 0; j < ncols; ++j) {
+                half sum = vec_dot_KQ(K + (k_VKQ_0 + i_KQ)*nb11, Q_h2[j], Q_i32[j], Q_ds[j]);
+                sum = warp_reduce_sum((float)sum);
+
+                if (use_logit_softcap) {
+                    sum = logit_softcap*tanhf(sum);
+                }
+
+                sum += mask ? slopeh*maskh[j*ne11 + k_VKQ_0 + i_KQ] : __float2half(0.0f);
+
+                if (ncols == 1) {
+                    kqmax_new        = ggml_cuda_hmax(kqmax_new,        sum);
+                } else {
+                    kqmax_new_arr[j] = ggml_cuda_hmax(kqmax_new_arr[j], sum);
+                }
+
+                if (threadIdx.x == 0) {
+                    KQ[j*D + i_KQ] = sum;
+                }
+            }
+        }
+
+#pragma unroll
+        for (int j = 0; j < ncols; ++j) {
+            half kqmax_new_j = ncols == 1 ? kqmax_new : kqmax_new_arr[j];
+
+            kqmax_new_j = warp_reduce_max(kqmax_new_j);
+            if (threadIdx.x == 0) {
+                kqmax_shared[j][threadIdx.y] = kqmax_new_j;
+            }
+        }
+
+        __syncthreads();
+
+#pragma unroll
+        for (int j = 0; j < ncols; ++j) {
+            half kqmax_new_j = kqmax_shared[j][threadIdx.x];
+            kqmax_new_j = warp_reduce_max(kqmax_new_j);
+
+            const half KQ_max_scale = hexp(kqmax[j] - kqmax_new_j);
+            kqmax[j] = kqmax_new_j;
+
+            const half val = hexp(KQ[j*D + tid] - kqmax[j]);
+            kqsum[j] = kqsum[j]*KQ_max_scale + val;
+            KQ[j*D + tid] = val;
+
+            VKQ[j] *= __half2half2(KQ_max_scale);
+        }
+
+        __syncthreads();
+
+#pragma unroll
+        for (int k0 = 0; k0 < D; k0 += 2) {
+            if (FATTN_KQ_STRIDE % D != 0 && k_VKQ_0 + k0 >= ne11) {
+                break;
+            }
+
+            half2 V_k;
+            reinterpret_cast<half&>(V_k.x) = dequantize_1_v(V + (k_VKQ_0 + k0 + 0)*nb21, tid);
+            reinterpret_cast<half&>(V_k.y) = dequantize_1_v(V + (k_VKQ_0 + k0 + 1)*nb21, tid);
+#pragma unroll
+            for (int j = 0; j < ncols; ++j) {
+                VKQ[j] += V_k*KQ2[j*(D/2) + k0/2];
+            }
+        }
+
+        __syncthreads();
+    }
+
+#pragma unroll
+    for (int j = 0; j < ncols; ++j) {
+        kqsum[j] = warp_reduce_sum((float)kqsum[j]);
+        if (threadIdx.x == 0) {
+            kqsum_shared[j][threadIdx.y] = kqsum[j];
+        }
+    }
+
+    __syncthreads();
+
+#pragma unroll
+    for (int j_VKQ = 0; j_VKQ < ncols; ++j_VKQ) {
+        if (ncols > 2 && ic0 + j_VKQ >= ne01) {
+            break;
+        }
+
+        kqsum[j_VKQ] = kqsum_shared[j_VKQ][threadIdx.x];
+        kqsum[j_VKQ] = warp_reduce_sum((float)kqsum[j_VKQ]);
+
+        half dst_val = (__low2half(VKQ[j_VKQ]) + __high2half(VKQ[j_VKQ]));
+        if (parallel_blocks == 1) {
+            dst_val /= kqsum[j_VKQ];
+        }
+        const int j_dst = (ic0 + j_VKQ)*parallel_blocks + ip;
+        dst[j_dst*D*gridDim.y + D*blockIdx.y + tid] = dst_val;
+    }
+
+    if (parallel_blocks != 1 && tid < ncols && (ncols <= 2 || ic0 + tid < ne01)) {
+        dst_meta[(ic0 + tid)*gridDim.y*parallel_blocks + blockIdx.y*parallel_blocks + ip] = make_float2(kqmax[tid], kqsum[tid]);
+    }
+#else
+   NO_DEVICE_CODE;
+#endif // FP16_AVAILABLE
+}
+
+template <int D, int cols_per_block, int parallel_blocks, ggml_type type_K, ggml_type type_V, bool use_logit_softcap>
+void ggml_cuda_flash_attn_ext_vec_f16_case_impl(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
+    constexpr int nwarps = D/WARP_SIZE;
+    fattn_kernel_t fattn_kernel = flash_attn_vec_ext_f16<D, cols_per_block, parallel_blocks, type_K, type_V, use_logit_softcap>;
+    constexpr bool need_f16_K = D != 128;
+    constexpr bool need_f16_V = D != 128 && D != 64;
+    launch_fattn<D, parallel_blocks>(ctx, dst, fattn_kernel, nwarps, cols_per_block, need_f16_K, need_f16_V);
+}
+
+template <int D, ggml_type type_K, ggml_type type_V>
+void ggml_cuda_flash_attn_ext_vec_f16_case(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
+    const ggml_tensor * KQV = dst;
+    const ggml_tensor * Q   = dst->src[0];
+    const ggml_tensor * K   = dst->src[1];
+    const ggml_tensor * V   = dst->src[2];
+
+    const int32_t precision = KQV->op_params[3];
+    GGML_ASSERT(precision == GGML_PREC_DEFAULT);
+
+    GGML_ASSERT(K->type == type_K);
+    GGML_ASSERT(V->type == type_V);
+
+    float logit_softcap;
+    memcpy(&logit_softcap, (const float *) KQV->op_params + 2, sizeof(float));
+
+    if (Q->ne[1] == 1) {
+        constexpr int cols_per_block  = 1;
+        constexpr int parallel_blocks = 4;
+        if (logit_softcap == 0.0f) {
+            constexpr bool use_logit_softcap = false;
+            ggml_cuda_flash_attn_ext_vec_f16_case_impl<D, cols_per_block, parallel_blocks, type_K, type_V, use_logit_softcap>(ctx, dst);
+        } else {
+            constexpr bool use_logit_softcap = true;
+            ggml_cuda_flash_attn_ext_vec_f16_case_impl<D, cols_per_block, parallel_blocks, type_K, type_V, use_logit_softcap>(ctx, dst);
+        }
+        return;
+    }
+
+    if (Q->ne[1] == 2) {
+        constexpr int cols_per_block  = 2;
+        constexpr int parallel_blocks = 4;
+        if (logit_softcap == 0.0f) {
+            constexpr bool use_logit_softcap = false;
+            ggml_cuda_flash_attn_ext_vec_f16_case_impl<D, cols_per_block, parallel_blocks, type_K, type_V, use_logit_softcap>(ctx, dst);
+        } else {
+            constexpr bool use_logit_softcap = true;
+            ggml_cuda_flash_attn_ext_vec_f16_case_impl<D, cols_per_block, parallel_blocks, type_K, type_V, use_logit_softcap>(ctx, dst);
+        }
+        return;
+    }
+
+    if (Q->ne[1] <= 4) {
+        constexpr int cols_per_block  = 4;
+        constexpr int parallel_blocks = 4;
+        if (logit_softcap == 0.0f) {
+            constexpr bool use_logit_softcap = false;
+            ggml_cuda_flash_attn_ext_vec_f16_case_impl<D, cols_per_block, parallel_blocks, type_K, type_V, use_logit_softcap>(ctx, dst);
+        } else {
+            constexpr bool use_logit_softcap = true;
+            ggml_cuda_flash_attn_ext_vec_f16_case_impl<D, cols_per_block, parallel_blocks, type_K, type_V, use_logit_softcap>(ctx, dst);
+        }
+        return;
+    }
+
+    if (Q->ne[1] <= 8) {
+        constexpr int cols_per_block  = 8;
+        constexpr int parallel_blocks = 4;
+        if (logit_softcap == 0.0f) {
+            constexpr bool use_logit_softcap = false;
+            ggml_cuda_flash_attn_ext_vec_f16_case_impl<D, cols_per_block, parallel_blocks, type_K, type_V, use_logit_softcap>(ctx, dst);
+        } else {
+            constexpr bool use_logit_softcap = true;
+            ggml_cuda_flash_attn_ext_vec_f16_case_impl<D, cols_per_block, parallel_blocks, type_K, type_V, use_logit_softcap>(ctx, dst);
+        }
+        return;
+    }
+
+    constexpr int cols_per_block  = 8;
+    constexpr int parallel_blocks = 1;
+    if (logit_softcap == 0.0f) {
+        constexpr bool use_logit_softcap = false;
+        ggml_cuda_flash_attn_ext_vec_f16_case_impl<D, cols_per_block, parallel_blocks, type_K, type_V, use_logit_softcap>(ctx, dst);
+    } else {
+        constexpr bool use_logit_softcap = true;
+        ggml_cuda_flash_attn_ext_vec_f16_case_impl<D, cols_per_block, parallel_blocks, type_K, type_V, use_logit_softcap>(ctx, dst);
+    }
+}
+
+#define DECL_FATTN_VEC_F16_CASE(D, type_K, type_V)                          \
+    template void ggml_cuda_flash_attn_ext_vec_f16_case                     \
+    <D, type_K, type_V>(ggml_backend_cuda_context & ctx, ggml_tensor * dst) \
+
+extern DECL_FATTN_VEC_F16_CASE( 64, GGML_TYPE_F16, GGML_TYPE_Q4_0);
+extern DECL_FATTN_VEC_F16_CASE( 64, GGML_TYPE_F16, GGML_TYPE_Q4_1);
+extern DECL_FATTN_VEC_F16_CASE( 64, GGML_TYPE_F16, GGML_TYPE_Q5_0);
+extern DECL_FATTN_VEC_F16_CASE( 64, GGML_TYPE_F16, GGML_TYPE_Q5_1);
+extern DECL_FATTN_VEC_F16_CASE( 64, GGML_TYPE_F16, GGML_TYPE_Q8_0);
+extern DECL_FATTN_VEC_F16_CASE( 64, GGML_TYPE_F16, GGML_TYPE_F16);
+
+extern DECL_FATTN_VEC_F16_CASE(128, GGML_TYPE_Q4_0, GGML_TYPE_Q4_0);
+extern DECL_FATTN_VEC_F16_CASE(128, GGML_TYPE_Q4_1, GGML_TYPE_Q4_0);
+extern DECL_FATTN_VEC_F16_CASE(128, GGML_TYPE_Q5_0, GGML_TYPE_Q4_0);
+extern DECL_FATTN_VEC_F16_CASE(128, GGML_TYPE_Q5_1, GGML_TYPE_Q4_0);
+extern DECL_FATTN_VEC_F16_CASE(128, GGML_TYPE_Q8_0, GGML_TYPE_Q4_0);
+extern DECL_FATTN_VEC_F16_CASE(128, GGML_TYPE_F16,  GGML_TYPE_Q4_0);
+
+extern DECL_FATTN_VEC_F16_CASE(128, GGML_TYPE_Q4_0, GGML_TYPE_Q4_1);
+extern DECL_FATTN_VEC_F16_CASE(128, GGML_TYPE_Q4_1, GGML_TYPE_Q4_1);
+extern DECL_FATTN_VEC_F16_CASE(128, GGML_TYPE_Q5_0, GGML_TYPE_Q4_1);
+extern DECL_FATTN_VEC_F16_CASE(128, GGML_TYPE_Q5_1, GGML_TYPE_Q4_1);
+extern DECL_FATTN_VEC_F16_CASE(128, GGML_TYPE_Q8_0, GGML_TYPE_Q4_1);
+extern DECL_FATTN_VEC_F16_CASE(128, GGML_TYPE_F16,  GGML_TYPE_Q4_1);
+
+extern DECL_FATTN_VEC_F16_CASE(128, GGML_TYPE_Q4_0, GGML_TYPE_Q5_0);
+extern DECL_FATTN_VEC_F16_CASE(128, GGML_TYPE_Q4_1, GGML_TYPE_Q5_0);
+extern DECL_FATTN_VEC_F16_CASE(128, GGML_TYPE_Q5_0, GGML_TYPE_Q5_0);
+extern DECL_FATTN_VEC_F16_CASE(128, GGML_TYPE_Q5_1, GGML_TYPE_Q5_0);
+extern DECL_FATTN_VEC_F16_CASE(128, GGML_TYPE_Q8_0, GGML_TYPE_Q5_0);
+extern DECL_FATTN_VEC_F16_CASE(128, GGML_TYPE_F16,  GGML_TYPE_Q5_0);
+
+extern DECL_FATTN_VEC_F16_CASE(128, GGML_TYPE_Q4_0, GGML_TYPE_Q5_1);
+extern DECL_FATTN_VEC_F16_CASE(128, GGML_TYPE_Q4_1, GGML_TYPE_Q5_1);
+extern DECL_FATTN_VEC_F16_CASE(128, GGML_TYPE_Q5_0, GGML_TYPE_Q5_1);
+extern DECL_FATTN_VEC_F16_CASE(128, GGML_TYPE_Q5_1, GGML_TYPE_Q5_1);
+extern DECL_FATTN_VEC_F16_CASE(128, GGML_TYPE_Q8_0, GGML_TYPE_Q5_1);
+extern DECL_FATTN_VEC_F16_CASE(128, GGML_TYPE_F16,  GGML_TYPE_Q5_1);
+
+extern DECL_FATTN_VEC_F16_CASE(128, GGML_TYPE_Q4_0, GGML_TYPE_Q8_0);
+extern DECL_FATTN_VEC_F16_CASE(128, GGML_TYPE_Q4_1, GGML_TYPE_Q8_0);
+extern DECL_FATTN_VEC_F16_CASE(128, GGML_TYPE_Q5_0, GGML_TYPE_Q8_0);
+extern DECL_FATTN_VEC_F16_CASE(128, GGML_TYPE_Q5_1, GGML_TYPE_Q8_0);
+extern DECL_FATTN_VEC_F16_CASE(128, GGML_TYPE_Q8_0, GGML_TYPE_Q8_0);
+extern DECL_FATTN_VEC_F16_CASE(128, GGML_TYPE_F16,  GGML_TYPE_Q8_0);
+
+extern DECL_FATTN_VEC_F16_CASE(128, GGML_TYPE_Q4_0, GGML_TYPE_F16);
+extern DECL_FATTN_VEC_F16_CASE(128, GGML_TYPE_Q4_1, GGML_TYPE_F16);
+extern DECL_FATTN_VEC_F16_CASE(128, GGML_TYPE_Q5_0, GGML_TYPE_F16);
+extern DECL_FATTN_VEC_F16_CASE(128, GGML_TYPE_Q5_1, GGML_TYPE_F16);
+extern DECL_FATTN_VEC_F16_CASE(128, GGML_TYPE_Q8_0, GGML_TYPE_F16);
+extern DECL_FATTN_VEC_F16_CASE(128, GGML_TYPE_F16,  GGML_TYPE_F16);
+
+extern DECL_FATTN_VEC_F16_CASE(256, GGML_TYPE_F16, GGML_TYPE_F16);
diff --git a/src/whisper_cpp/ggml/src/ggml-cuda/fattn-vec-f32.cuh b/src/whisper_cpp/ggml/src/ggml-cuda/fattn-vec-f32.cuh
new file mode 100644
index 00000000..bf512590
--- /dev/null
+++ b/src/whisper_cpp/ggml/src/ggml-cuda/fattn-vec-f32.cuh
@@ -0,0 +1,420 @@
+#include "common.cuh"
+#include "fattn-common.cuh"
+
+template<int D, int ncols, int parallel_blocks, ggml_type type_K, ggml_type type_V, bool use_logit_softcap> // D == head size
+#if !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__))
+__launch_bounds__(D, 1)
+#endif // !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__))
+static __global__ void flash_attn_vec_ext_f32(
+        const char * __restrict__ Q,
+        const char * __restrict__ K,
+        const char * __restrict__ V,
+        const char * __restrict__ mask,
+        float      * __restrict__ dst,
+        float2     * __restrict__ dst_meta,
+        const float scale,
+        const float max_bias,
+        const float m0,
+        const float m1,
+        const uint32_t n_head_log2,
+        const float logit_softcap,
+        const int ne00,
+        const int ne01,
+        const int ne02,
+        const int ne03,
+        const int ne10,
+        const int ne11,
+        const int ne12,
+        const int ne13,
+        const int ne31,
+        const int nb31,
+        const int nb01,
+        const int nb02,
+        const int nb03,
+        const int nb11,
+        const int nb12,
+        const int nb13,
+        const int nb21,
+        const int nb22,
+        const int nb23,
+        const int ne0,
+        const int ne1,
+        const int ne2,
+        const int ne3) {
+    // Skip unused kernel variants for faster compilation:
+    if (use_logit_softcap && !(D == 128 || D == 256)) {
+        NO_DEVICE_CODE;
+        return;
+    }
+
+    //In this kernel Q, K, V are matrices while i, j, k are matrix indices.
+
+    constexpr vec_dot_KQ_f32_t vec_dot_KQ = get_vec_dot_KQ_f32<D>(type_K);
+    constexpr bool Q_q8_1 = type_K != GGML_TYPE_F16;
+    constexpr dequantize_1_f32_t dequantize_1_v = get_dequantize_1_f32(type_V);
+
+    const int ic0 = (blockIdx.x / parallel_blocks) * ncols; // Index of the Q/QKV column to work on.
+    const int ip  =  blockIdx.x % parallel_blocks; // Index in group of blocks running for the same column in parallel.
+
+    const int gqa_ratio = ne02 / ne12; // With grouped query attention there are > 1 Q matrices per K, V matrix.
+    Q += nb02* blockIdx.y              + nb01*ic0;
+    K += nb12*(blockIdx.y / gqa_ratio);
+    V += nb22*(blockIdx.y / gqa_ratio); // K and V have same shape
+    const half * maskh = (const half   *)  mask + ne11*ic0;
+
+    const float slope = get_alibi_slope(max_bias, blockIdx.y, n_head_log2, m0, m1);
+
+    static_assert(D % (2*WARP_SIZE) == 0, "D not divisible by 2*WARP_SIZE == 64.");
+    constexpr int nwarps = D / WARP_SIZE;
+    const int tid = WARP_SIZE*threadIdx.y + threadIdx.x;
+    __builtin_assume(tid < D);
+
+    __shared__ float KQ[ncols*D];
+#pragma unroll
+    for (int j = 0; j < ncols; ++j) {
+        KQ[j*D + tid] = -FLT_MAX/2.0f;
+    }
+
+    float kqmax[ncols];
+#pragma unroll
+    for (int j = 0; j < ncols; ++j) {
+        kqmax[j] = -FLT_MAX/2.0f;
+    }
+    float kqsum[ncols] = {0.0f};
+
+    __shared__ float kqmax_shared[ncols][WARP_SIZE];
+    __shared__ float kqsum_shared[ncols][WARP_SIZE];
+#pragma unroll
+    for (int j = 0; j < ncols; ++j) {
+        if (threadIdx.y == 0) {
+            kqmax_shared[j][threadIdx.x] = -FLT_MAX/2.0f;
+            kqsum_shared[j][threadIdx.x] = 0.0f;
+        }
+    }
+    __syncthreads();
+
+    // Convert Q to float2 (f16 K) or q8_1 (quantized K) and store in registers:
+    float2  Q_f2[ncols][D/(2*WARP_SIZE)];
+    int    Q_i32[ncols][D/(sizeof(int)*QK8_1) == 0 ? 1 : D >= D/(sizeof(int)*QK8_1)];
+    float2  Q_ds[ncols][D/QK8_1 == 0 ? 1 : D/QK8_1];
+    if (Q_q8_1) {
+#pragma unroll
+        for (int j0 = 0; j0 < ncols; j0 += nwarps) {
+            const int j = j0 + threadIdx.y;
+
+            if (j0 + nwarps > ncols && j >= ncols) {
+                break;
+            }
+
+            // Reuse KQ as temporary storage for converting Q to q8_1:
+            int    * tmp_q_i32 = (int    *) &KQ[j*D];
+            float2 * tmp_q_ds  = (float2 *) (tmp_q_i32 + D/sizeof(int));
+
+            // Set memory to zero if out of bounds:
+            if (ncols > 2 && ic0 + j >= ne01) {
+#pragma unroll
+                for (int i0 = 0; i0 < D/sizeof(int); i0 += WARP_SIZE) {
+                    const int i = i0 + threadIdx.x;
+
+                    tmp_q_i32[i] = 0;
+                }
+                if (threadIdx.x < D/QK8_1) {
+                    tmp_q_ds[threadIdx.x] = make_float2(0.0f, 0.0f);
+                }
+                continue;
+            }
+
+            const float * Q_f = (const float *) (Q + j*nb01);
+#pragma unroll
+            for (int i0 = 0; i0 < D/sizeof(int); i0 += WARP_SIZE) {
+                quantize_q8_1_to_shared<float2>(Q_f + 4*i0, scale, tmp_q_i32, tmp_q_ds);
+            }
+        }
+
+        __syncthreads();
+
+#pragma unroll
+        for (int j = 0; j < ncols; ++j) {
+            int    * tmp_q_i32 = (int    *) &KQ[j*D];
+            float2 * tmp_q_ds  = (float2 *) (tmp_q_i32 + D/sizeof(int));
+
+#pragma unroll
+            for (int i0 = 0; i0 < D/sizeof(int); i0 += WARP_SIZE) {
+                const int i = i0 + threadIdx.x;
+
+                Q_i32[j][i0/WARP_SIZE] = tmp_q_i32[i];
+                Q_ds[j][i0/WARP_SIZE]  = tmp_q_ds[i/QI8_1];
+            }
+        }
+
+        __syncthreads();
+    } else {
+#pragma unroll
+        for (int j = 0; j < ncols; ++j) {
+            const float2 * Q_f2_j = (const float2 *) (Q + j*nb01);
+#pragma unroll
+            for (int i0 = 0; i0 < D/2; i0 += WARP_SIZE) {
+                const int i = i0 + threadIdx.x;
+
+                Q_f2[j][i0/WARP_SIZE]    = ncols <= 2 || ic0 + j < ne01 ? Q_f2_j[i] : make_float2(0.0f, 0.0f);
+                Q_f2[j][i0/WARP_SIZE].x *= scale;
+                Q_f2[j][i0/WARP_SIZE].y *= scale;
+            }
+        }
+    }
+
+    float VKQ[ncols] = {0.0f};
+
+    const int k_start = parallel_blocks == 1 ? 0 : ip*D;
+    for (int k_VKQ_0 = k_start; k_VKQ_0 < ne11; k_VKQ_0 += parallel_blocks*D) {
+        // Calculate KQ tile and keep track of new maximum KQ values:
+
+        float kqmax_new_arr[ncols];
+#pragma unroll
+        for (int j = 0; j < ncols; ++j) {
+            kqmax_new_arr[j] = kqmax[j];
+        }
+
+#pragma unroll
+        for (int i_KQ_0 = 0; i_KQ_0 < D; i_KQ_0 += nwarps) {
+            const int i_KQ = i_KQ_0 + threadIdx.y;
+
+            if ((i_KQ_0 + nwarps > D && i_KQ >= D) || (FATTN_KQ_STRIDE % D != 0 && k_VKQ_0 + i_KQ >= ne11)) {
+                break;
+            }
+
+#pragma unroll
+            for (int j = 0; j < ncols; ++j) {
+                float sum = vec_dot_KQ(K + (k_VKQ_0 + i_KQ)*nb11, Q_f2[j], Q_i32[j], Q_ds[j]);
+                sum = warp_reduce_sum(sum);
+
+                if (use_logit_softcap) {
+                    sum = logit_softcap*tanhf(sum);
+                }
+
+                sum += mask ? slope*__half2float(maskh[j*ne11 + k_VKQ_0 + i_KQ]) : 0.0f;
+
+                kqmax_new_arr[j] = fmaxf(kqmax_new_arr[j], sum);
+
+                if (threadIdx.x == 0) {
+                    KQ[j*D + i_KQ] = sum;
+                }
+            }
+        }
+
+#pragma unroll
+        for (int j = 0; j < ncols; ++j) {
+            float kqmax_new_j = kqmax_new_arr[j];
+
+            kqmax_new_j = warp_reduce_max(kqmax_new_j);
+            if (threadIdx.x == 0) {
+                kqmax_shared[j][threadIdx.y] = kqmax_new_j;
+            }
+        }
+
+        __syncthreads();
+
+#pragma unroll
+        for (int j = 0; j < ncols; ++j) {
+            float kqmax_new_j = kqmax_shared[j][threadIdx.x];
+            kqmax_new_j = warp_reduce_max(kqmax_new_j);
+
+            const float KQ_max_scale = expf(kqmax[j] - kqmax_new_j);
+            kqmax[j] = kqmax_new_j;
+
+            const float val = expf(KQ[j*D + tid] - kqmax[j]);
+            kqsum[j] = kqsum[j]*KQ_max_scale + val;
+            KQ[j*D + tid] = val;
+
+            VKQ[j] *= KQ_max_scale;
+        }
+
+        __syncthreads();
+
+#pragma unroll
+        for (int k = 0; k < D; ++k) {
+            if (FATTN_KQ_STRIDE % D != 0 && k_VKQ_0 + k >= ne11) {
+                break;
+            }
+
+            const float V_ki = dequantize_1_v(V + (k_VKQ_0 + k)*nb21, tid);
+#pragma unroll
+            for (int j = 0; j < ncols; ++j) {
+                VKQ[j] += V_ki*KQ[j*D + k];
+            }
+        }
+
+        __syncthreads();
+    }
+
+#pragma unroll
+    for (int j = 0; j < ncols; ++j) {
+        kqsum[j] = warp_reduce_sum(kqsum[j]);
+        if (threadIdx.x == 0) {
+            kqsum_shared[j][threadIdx.y] = kqsum[j];
+        }
+    }
+
+    __syncthreads();
+
+#pragma unroll
+    for (int j_VKQ = 0; j_VKQ < ncols; ++j_VKQ) {
+        if (ncols > 2 && ic0 + j_VKQ >= ne01) {
+            break;
+        }
+
+        kqsum[j_VKQ] = kqsum_shared[j_VKQ][threadIdx.x];
+        kqsum[j_VKQ] = warp_reduce_sum(kqsum[j_VKQ]);
+
+        float dst_val = VKQ[j_VKQ];
+        if (parallel_blocks == 1) {
+            dst_val /= kqsum[j_VKQ];
+        }
+        const int j_dst = (ic0 + j_VKQ)*parallel_blocks + ip;
+        dst[j_dst*D*gridDim.y + D*blockIdx.y + tid] = dst_val;
+    }
+
+    if (parallel_blocks != 1 && tid < ncols && (ncols <= 2 || ic0 + tid < ne01)) {
+        dst_meta[(ic0 + tid)*gridDim.y*parallel_blocks + blockIdx.y*parallel_blocks + ip] = make_float2(kqmax[tid], kqsum[tid]);
+    }
+}
+
+template <int D, int cols_per_block, int parallel_blocks, ggml_type type_K, ggml_type type_V, bool use_logit_softcap>
+void ggml_cuda_flash_attn_ext_vec_f32_case_impl(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
+    constexpr int nwarps = D/WARP_SIZE;
+    fattn_kernel_t fattn_kernel = flash_attn_vec_ext_f32<D, cols_per_block, parallel_blocks, type_K, type_V, use_logit_softcap>;
+    constexpr bool need_f16_K = D != 128;
+    constexpr bool need_f16_V = D != 128 && D != 64;
+    launch_fattn<D, parallel_blocks>(ctx, dst, fattn_kernel, nwarps, cols_per_block, need_f16_K, need_f16_V);
+}
+
+template <int D, ggml_type type_K, ggml_type type_V>
+void ggml_cuda_flash_attn_ext_vec_f32_case(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
+    const ggml_tensor * KQV = dst;
+    const ggml_tensor * Q   = dst->src[0];
+    const ggml_tensor * K   = dst->src[1];
+    const ggml_tensor * V   = dst->src[2];
+
+    GGML_ASSERT(K->type == type_K);
+    GGML_ASSERT(V->type == type_V);
+
+    float logit_softcap;
+    memcpy(&logit_softcap, (const float *) KQV->op_params + 2, sizeof(float));
+
+    if (Q->ne[1] == 1) {
+        constexpr int cols_per_block  = 1;
+        constexpr int parallel_blocks = 4;
+        if (logit_softcap == 0.0f) {
+            constexpr bool use_logit_softcap = false;
+            ggml_cuda_flash_attn_ext_vec_f32_case_impl<D, cols_per_block, parallel_blocks, type_K, type_V, use_logit_softcap>(ctx, dst);
+        } else {
+            constexpr bool use_logit_softcap = true;
+            ggml_cuda_flash_attn_ext_vec_f32_case_impl<D, cols_per_block, parallel_blocks, type_K, type_V, use_logit_softcap>(ctx, dst);
+        }
+        return;
+    }
+
+    if (Q->ne[1] == 2) {
+        constexpr int cols_per_block  = 2;
+        constexpr int parallel_blocks = 4;
+        if (logit_softcap == 0.0f) {
+            constexpr bool use_logit_softcap = false;
+            ggml_cuda_flash_attn_ext_vec_f32_case_impl<D, cols_per_block, parallel_blocks, type_K, type_V, use_logit_softcap>(ctx, dst);
+        } else {
+            constexpr bool use_logit_softcap = true;
+            ggml_cuda_flash_attn_ext_vec_f32_case_impl<D, cols_per_block, parallel_blocks, type_K, type_V, use_logit_softcap>(ctx, dst);
+        }
+        return;
+    }
+
+    if (Q->ne[1] <= 4) {
+        constexpr int cols_per_block  = 4;
+        constexpr int parallel_blocks = 4;
+        if (logit_softcap == 0.0f) {
+            constexpr bool use_logit_softcap = false;
+            ggml_cuda_flash_attn_ext_vec_f32_case_impl<D, cols_per_block, parallel_blocks, type_K, type_V, use_logit_softcap>(ctx, dst);
+        } else {
+            constexpr bool use_logit_softcap = true;
+            ggml_cuda_flash_attn_ext_vec_f32_case_impl<D, cols_per_block, parallel_blocks, type_K, type_V, use_logit_softcap>(ctx, dst);
+        }
+        return;
+    }
+
+    if (Q->ne[1] <= 8) {
+        constexpr int cols_per_block  = 8;
+        constexpr int parallel_blocks = 4;
+        if (logit_softcap == 0.0f) {
+            constexpr bool use_logit_softcap = false;
+            ggml_cuda_flash_attn_ext_vec_f32_case_impl<D, cols_per_block, parallel_blocks, type_K, type_V, use_logit_softcap>(ctx, dst);
+        } else {
+            constexpr bool use_logit_softcap = true;
+            ggml_cuda_flash_attn_ext_vec_f32_case_impl<D, cols_per_block, parallel_blocks, type_K, type_V, use_logit_softcap>(ctx, dst);
+        }
+        return;
+    }
+
+    constexpr int cols_per_block  = 8;
+    constexpr int parallel_blocks = 1;
+    if (logit_softcap == 0.0f) {
+        constexpr bool use_logit_softcap = false;
+        ggml_cuda_flash_attn_ext_vec_f32_case_impl<D, cols_per_block, parallel_blocks, type_K, type_V, use_logit_softcap>(ctx, dst);
+    } else {
+        constexpr bool use_logit_softcap = true;
+        ggml_cuda_flash_attn_ext_vec_f32_case_impl<D, cols_per_block, parallel_blocks, type_K, type_V, use_logit_softcap>(ctx, dst);
+    }
+}
+
+#define DECL_FATTN_VEC_F32_CASE(D, type_K, type_V)                          \
+    template void ggml_cuda_flash_attn_ext_vec_f32_case                     \
+    <D, type_K, type_V>(ggml_backend_cuda_context & ctx, ggml_tensor * dst) \
+
+extern DECL_FATTN_VEC_F32_CASE( 64, GGML_TYPE_F16, GGML_TYPE_Q4_0);
+extern DECL_FATTN_VEC_F32_CASE( 64, GGML_TYPE_F16, GGML_TYPE_Q4_1);
+extern DECL_FATTN_VEC_F32_CASE( 64, GGML_TYPE_F16, GGML_TYPE_Q5_0);
+extern DECL_FATTN_VEC_F32_CASE( 64, GGML_TYPE_F16, GGML_TYPE_Q5_1);
+extern DECL_FATTN_VEC_F32_CASE( 64, GGML_TYPE_F16, GGML_TYPE_Q8_0);
+extern DECL_FATTN_VEC_F32_CASE( 64, GGML_TYPE_F16, GGML_TYPE_F16);
+
+extern DECL_FATTN_VEC_F32_CASE(128, GGML_TYPE_Q4_0, GGML_TYPE_Q4_0);
+extern DECL_FATTN_VEC_F32_CASE(128, GGML_TYPE_Q4_1, GGML_TYPE_Q4_0);
+extern DECL_FATTN_VEC_F32_CASE(128, GGML_TYPE_Q5_0, GGML_TYPE_Q4_0);
+extern DECL_FATTN_VEC_F32_CASE(128, GGML_TYPE_Q5_1, GGML_TYPE_Q4_0);
+extern DECL_FATTN_VEC_F32_CASE(128, GGML_TYPE_Q8_0, GGML_TYPE_Q4_0);
+extern DECL_FATTN_VEC_F32_CASE(128, GGML_TYPE_F16,  GGML_TYPE_Q4_0);
+
+extern DECL_FATTN_VEC_F32_CASE(128, GGML_TYPE_Q4_0, GGML_TYPE_Q4_1);
+extern DECL_FATTN_VEC_F32_CASE(128, GGML_TYPE_Q4_1, GGML_TYPE_Q4_1);
+extern DECL_FATTN_VEC_F32_CASE(128, GGML_TYPE_Q5_0, GGML_TYPE_Q4_1);
+extern DECL_FATTN_VEC_F32_CASE(128, GGML_TYPE_Q5_1, GGML_TYPE_Q4_1);
+extern DECL_FATTN_VEC_F32_CASE(128, GGML_TYPE_Q8_0, GGML_TYPE_Q4_1);
+extern DECL_FATTN_VEC_F32_CASE(128, GGML_TYPE_F16,  GGML_TYPE_Q4_1);
+
+extern DECL_FATTN_VEC_F32_CASE(128, GGML_TYPE_Q4_0, GGML_TYPE_Q5_0);
+extern DECL_FATTN_VEC_F32_CASE(128, GGML_TYPE_Q4_1, GGML_TYPE_Q5_0);
+extern DECL_FATTN_VEC_F32_CASE(128, GGML_TYPE_Q5_0, GGML_TYPE_Q5_0);
+extern DECL_FATTN_VEC_F32_CASE(128, GGML_TYPE_Q5_1, GGML_TYPE_Q5_0);
+extern DECL_FATTN_VEC_F32_CASE(128, GGML_TYPE_Q8_0, GGML_TYPE_Q5_0);
+extern DECL_FATTN_VEC_F32_CASE(128, GGML_TYPE_F16,  GGML_TYPE_Q5_0);
+
+extern DECL_FATTN_VEC_F32_CASE(128, GGML_TYPE_Q4_0, GGML_TYPE_Q5_1);
+extern DECL_FATTN_VEC_F32_CASE(128, GGML_TYPE_Q4_1, GGML_TYPE_Q5_1);
+extern DECL_FATTN_VEC_F32_CASE(128, GGML_TYPE_Q5_0, GGML_TYPE_Q5_1);
+extern DECL_FATTN_VEC_F32_CASE(128, GGML_TYPE_Q5_1, GGML_TYPE_Q5_1);
+extern DECL_FATTN_VEC_F32_CASE(128, GGML_TYPE_Q8_0, GGML_TYPE_Q5_1);
+extern DECL_FATTN_VEC_F32_CASE(128, GGML_TYPE_F16,  GGML_TYPE_Q5_1);
+
+extern DECL_FATTN_VEC_F32_CASE(128, GGML_TYPE_Q4_0, GGML_TYPE_Q8_0);
+extern DECL_FATTN_VEC_F32_CASE(128, GGML_TYPE_Q4_1, GGML_TYPE_Q8_0);
+extern DECL_FATTN_VEC_F32_CASE(128, GGML_TYPE_Q5_0, GGML_TYPE_Q8_0);
+extern DECL_FATTN_VEC_F32_CASE(128, GGML_TYPE_Q5_1, GGML_TYPE_Q8_0);
+extern DECL_FATTN_VEC_F32_CASE(128, GGML_TYPE_Q8_0, GGML_TYPE_Q8_0);
+extern DECL_FATTN_VEC_F32_CASE(128, GGML_TYPE_F16,  GGML_TYPE_Q8_0);
+
+extern DECL_FATTN_VEC_F32_CASE(128, GGML_TYPE_Q4_0, GGML_TYPE_F16);
+extern DECL_FATTN_VEC_F32_CASE(128, GGML_TYPE_Q4_1, GGML_TYPE_F16);
+extern DECL_FATTN_VEC_F32_CASE(128, GGML_TYPE_Q5_0, GGML_TYPE_F16);
+extern DECL_FATTN_VEC_F32_CASE(128, GGML_TYPE_Q5_1, GGML_TYPE_F16);
+extern DECL_FATTN_VEC_F32_CASE(128, GGML_TYPE_Q8_0, GGML_TYPE_F16);
+extern DECL_FATTN_VEC_F32_CASE(128, GGML_TYPE_F16,  GGML_TYPE_F16);
+
+extern DECL_FATTN_VEC_F32_CASE(256, GGML_TYPE_F16, GGML_TYPE_F16);
diff --git a/src/whisper_cpp/ggml/src/ggml-cuda/fattn-wmma-f16.cuh b/src/whisper_cpp/ggml/src/ggml-cuda/fattn-wmma-f16.cuh
new file mode 100644
index 00000000..b10d19d9
--- /dev/null
+++ b/src/whisper_cpp/ggml/src/ggml-cuda/fattn-wmma-f16.cuh
@@ -0,0 +1,543 @@
+#include "common.cuh"
+#include "fattn-common.cuh"
+
+#ifdef FP16_MMA_AVAILABLE
+#include <mma.h>
+#endif // FP16_MMA_AVAILABLE
+
+// D == head size, VKQ_stride == num VKQ rows calculated in parallel:
+template<int D, int ncols, int nwarps, int VKQ_stride, int parallel_blocks, typename KQ_acc_t, bool use_logit_softcap>
+#if !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__))
+__launch_bounds__(nwarps*WARP_SIZE, 1)
+#endif // !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__))
+static __global__ void flash_attn_ext_f16(
+        const char * __restrict__ Q,
+        const char * __restrict__ K,
+        const char * __restrict__ V,
+        const char * __restrict__ mask,
+        float      * __restrict__ dst,
+        float2     * __restrict__ dst_meta,
+        const float scale,
+        const float max_bias,
+        const float m0,
+        const float m1,
+        const uint32_t n_head_log2,
+        const float logit_softcap,
+        const int ne00,
+        const int ne01,
+        const int ne02,
+        const int ne03,
+        const int ne10,
+        const int ne11,
+        const int ne12,
+        const int ne13,
+        const int ne31,
+        const int nb31,
+        const int nb01,
+        const int nb02,
+        const int nb03,
+        const int nb11,
+        const int nb12,
+        const int nb13,
+        const int nb21,
+        const int nb22,
+        const int nb23,
+        const int ne0,
+        const int ne1,
+        const int ne2,
+        const int ne3) {
+#ifdef FP16_MMA_AVAILABLE
+    // Skip unused kernel variants for faster compilation:
+    if (use_logit_softcap && !(D == 128 || D == 256)) {
+        NO_DEVICE_CODE;
+        return;
+    }
+
+    //In this kernel Q, K, V are matrices while i, j, k are matrix indices.
+
+    const int ic0 = ncols*(blockIdx.x / parallel_blocks); // Index of the first Q/QKV column to work on.
+    const int ip  =        blockIdx.x % parallel_blocks;  // Index in group of blocks running for the same column in parallel.
+
+    static_assert(D <= FATTN_KQ_STRIDE, "D must be <= FATTN_KQ_STRIDE.");
+    static_assert(ncols == 8 || ncols % 16 == 0, "ncols must be 8 or a multiple of 16.");
+    constexpr int frag_m = ncols == 8 ? 32 : 16;
+    constexpr int frag_n = ncols == 8 ?  8 : 16;
+    static_assert(D % frag_m == 0, "If ncols == 8 then D % frag_m must be 0.");
+    typedef nvcuda::wmma::fragment<nvcuda::wmma::matrix_a,    frag_m, frag_n, 16, half, nvcuda::wmma::row_major> frag_a_K;
+    typedef nvcuda::wmma::fragment<nvcuda::wmma::matrix_a,    frag_m, frag_n, 16, half, nvcuda::wmma::col_major> frag_a_V;
+    typedef nvcuda::wmma::fragment<nvcuda::wmma::matrix_b,    frag_m, frag_n, 16, half, nvcuda::wmma::col_major> frag_b;
+    typedef nvcuda::wmma::fragment<nvcuda::wmma::accumulator, frag_m, frag_n, 16, KQ_acc_t>                      frag_c_KQ;
+    typedef nvcuda::wmma::fragment<nvcuda::wmma::accumulator, frag_m, frag_n, 16, half>                          frag_c_VKQ;
+
+    constexpr int KQ_stride_tc  = nwarps*frag_m; // Number of KQ rows calculated in parallel.
+    constexpr int VKQ_ratio = KQ_stride_tc/VKQ_stride; // Number of parallel VKQ accumulators needed to keep all warps busy.
+    static_assert(VKQ_ratio <= nwarps, "VKQ_ratio must be <= nwarps.");
+
+    // Pad internal representation of KQ, KQV to reduce shared memory bank conflicts:
+    constexpr int D_padded = D + 8;
+    constexpr int kqs_padded = FATTN_KQ_STRIDE + 8;
+    constexpr int kqar = sizeof(KQ_acc_t)/sizeof(half);
+
+    const int gqa_ratio = ne02 / ne12; // With grouped query attention there are > 1 Q matrices per K, V matrix.
+    const float * Q_f   = (const float *) (Q + nb02* blockIdx.y              + nb01*ic0);
+    const half  * K_h   = (const half  *) (K + nb12*(blockIdx.y / gqa_ratio));
+    const half  * V_h   = (const half  *) (V + nb12*(blockIdx.y / gqa_ratio)); // K and V have same shape
+    const half  * maskh = (const half  *)  mask + (nb31/sizeof(half))* ic0;
+    const half2 * mask2 = (const half2 *)  mask + (nb31/sizeof(half))*(ic0/2);
+
+    const int stride_Q  = nb01 / sizeof(float);
+    const int stride_KV = nb11 / sizeof(half);
+
+    const float slopef = get_alibi_slope(max_bias, blockIdx.y, n_head_log2, m0, m1);
+    const half  slopeh = __float2half(slopef);
+    const half2 slope2 = make_half2(slopef, slopef);
+
+    const half2 logit_softcap_2 = make_half2(logit_softcap, logit_softcap);
+
+    frag_b Q_b[D/16][ncols/frag_n];
+
+    // A single buffer for temporarily holding tiles of KQ and VKQ parts:
+    constexpr int mem_KQ = ncols*kqs_padded*kqar;
+    constexpr int mem_VKQ_parts = VKQ_ratio*ncols*D_padded;
+    __shared__ half KQ[mem_KQ >= mem_VKQ_parts ? mem_KQ : mem_VKQ_parts];
+    float * KQ_f = (float *) KQ;
+    half2 * KQ2 = (half2 *) KQ;
+
+    float    KQ_rowsum_f[ncols/nwarps] = {0.0f};
+    float       KQ_max_f[ncols/nwarps];
+    float KQ_max_scale_f[ncols/nwarps] = {0.0f};
+
+#pragma unroll
+    for (int j = 0; j < ncols/nwarps; ++j) {
+        KQ_max_f[j] = -FLT_MAX/2.0f;
+    }
+
+    half2    KQ_rowsum_h2[ncols/nwarps] = {{0.0f, 0.0f}};
+    half2       KQ_max_h2[ncols/nwarps];
+    half2 KQ_max_scale_h2[ncols/nwarps] = {{0.0f, 0.0f}};
+
+#pragma unroll
+    for (int j = 0; j < ncols/nwarps; ++j) {
+        KQ_max_h2[j] = make_half2(-HALF_MAX_HALF, -HALF_MAX_HALF);
+    }
+
+    __shared__ half VKQ[ncols*D_padded]; // Accumulator for final VKQ slice.
+    half2 * VKQ2 = (half2 *) VKQ;
+#pragma unroll
+    for (int j0 = 0; j0 < ncols; j0 += nwarps) {
+        const int j = j0 + threadIdx.y;
+#pragma unroll
+        for (int i0 = 0; i0 < D/2; i0 += WARP_SIZE) {
+            const int i = i0 + threadIdx.x;
+            if (i0 + WARP_SIZE > D/2 && i >= D/2) {
+                break;
+            }
+            VKQ2[j*(D_padded/2) + i] = make_half2(0.0f, 0.0f);
+        }
+    }
+
+    // Convert Q to half and apply scale, temporarily store in KQ:
+#pragma unroll
+    for (int j0 = 0; j0 < ncols; j0 += nwarps) {
+        const int j = j0 + threadIdx.y;
+#pragma unroll
+        for (int i0 = 0; i0 < D; i0 += WARP_SIZE) {
+            const int i = i0 + threadIdx.x;
+            if (i0 + WARP_SIZE > D && i >= D) {
+                break;
+            }
+            KQ[j*D_padded + i] = ic0 + j < ne01 ? Q_f[j*stride_Q + i] * scale : 0.0f;
+        }
+    }
+
+    __syncthreads();
+
+    // Load Q into tensor core fragments/registers since it will be used frequently:
+#pragma unroll
+    for (int i0 = 0; i0 < D; i0 += 16) {
+#pragma unroll
+        for (int j0 = 0; j0 < ncols; j0 += frag_n) {
+            nvcuda::wmma::load_matrix_sync(Q_b[i0/16][j0/frag_n], KQ + j0*D_padded + i0, D_padded);
+        }
+    }
+
+    __syncthreads();
+
+    // Iterate over ne11 == previous tokens:
+    for (int k_VKQ_0 = ip*FATTN_KQ_STRIDE; k_VKQ_0 < ne11; k_VKQ_0 += parallel_blocks*FATTN_KQ_STRIDE) {
+        // Calculate tile of KQ:
+#pragma unroll
+        for (int i_KQ_0 = 0; i_KQ_0 < FATTN_KQ_STRIDE; i_KQ_0 += KQ_stride_tc) {
+            frag_c_KQ KQ_c[ncols/frag_n];
+#pragma unroll
+            for (int j = 0; j < ncols/frag_n; ++j) {
+                nvcuda::wmma::fill_fragment(KQ_c[j], 0.0f);
+            }
+#pragma unroll
+            for (int k_KQ_0 = 0; k_KQ_0 < D; k_KQ_0 += 16) {
+                frag_a_K K_a;
+                nvcuda::wmma::load_matrix_sync(K_a, K_h + (k_VKQ_0 + i_KQ_0 + frag_m*threadIdx.y)*stride_KV + k_KQ_0, stride_KV);
+#pragma unroll
+                for (int j = 0; j < ncols/frag_n; ++j) {
+                    nvcuda::wmma::mma_sync(KQ_c[j], K_a, Q_b[k_KQ_0/16][j], KQ_c[j]);
+                }
+            }
+#pragma unroll
+            for (int j0 = 0; j0 < ncols; j0 += frag_n) {
+                nvcuda::wmma::store_matrix_sync((KQ_acc_t *) KQ + j0*kqs_padded + i_KQ_0 + frag_m*threadIdx.y, KQ_c[j0/frag_n], kqs_padded, nvcuda::wmma::mem_col_major);
+            }
+        }
+
+        __syncthreads();
+
+        // Calculate softmax for each KQ column using the current max. value.
+        // The divisor is stored in KQ_rowsum and will be applied at the end.
+#pragma unroll
+        for (int j0 = 0; j0 < ncols; j0 += nwarps) {
+            const int j = j0 + threadIdx.y;
+
+            if (std::is_same<KQ_acc_t, float>::value) {
+                float KQ_f_tmp[FATTN_KQ_STRIDE / WARP_SIZE];
+#pragma unroll
+                for (int k0 = 0; k0 < FATTN_KQ_STRIDE; k0 += WARP_SIZE) {
+                    const int k = k0 + threadIdx.x;
+
+                    KQ_f_tmp[k0/WARP_SIZE] = KQ_f[j*kqs_padded + k];
+
+                    if (use_logit_softcap) {
+                        KQ_f_tmp[k0/WARP_SIZE] = logit_softcap*tanhf(KQ_f_tmp[k0/WARP_SIZE]);
+                    }
+                }
+
+                float KQ_max_new = KQ_max_f[j0/nwarps];
+#pragma unroll
+                for (int k0 = 0; k0 < FATTN_KQ_STRIDE; k0 += WARP_SIZE) {
+                    const int k = k0 + threadIdx.x;
+
+                    KQ_f_tmp[k0/WARP_SIZE] += mask ? __half2float(slopeh*maskh[j*(nb31/sizeof(half)) + k_VKQ_0 + k]) : 0.0f;
+                    KQ_max_new = max(KQ_max_new, KQ_f_tmp[k0/WARP_SIZE]);
+                }
+                KQ_max_new = warp_reduce_max(KQ_max_new);
+
+                const float diff = KQ_max_f[j0/nwarps] - KQ_max_new;
+                KQ_max_scale_f[j0/nwarps] = expf(diff);
+                if (diff <= SOFTMAX_FTZ_THRESHOLD) {
+                    KQ_max_scale_f[j0/nwarps] = 0.0f;
+                }
+                KQ_max_f[j0/nwarps] = KQ_max_new;
+
+                float KQ_rowsum_add = 0.0f;
+#pragma unroll
+                for (int k0 = 0; k0 < FATTN_KQ_STRIDE; k0 += WARP_SIZE) {
+                    const int k = k0 + threadIdx.x;
+
+                    const float diff = KQ_f_tmp[k0/WARP_SIZE] - KQ_max_f[j0/nwarps];
+                    KQ_f_tmp[k0/WARP_SIZE] = expf(diff);
+                    if (diff <= SOFTMAX_FTZ_THRESHOLD) {
+                        KQ_f_tmp[k0/WARP_SIZE] = 0.0f;
+                    }
+                    KQ_rowsum_add += KQ_f_tmp[k0/WARP_SIZE];
+                    KQ[j*(kqar*kqs_padded) + k] = KQ_f_tmp[k0/WARP_SIZE];
+                }
+                KQ_rowsum_add = warp_reduce_sum(KQ_rowsum_add);
+
+                // Scale previous KQ_rowsum to account for a potential increase in KQ_max:
+                KQ_rowsum_f[j0/nwarps] = KQ_max_scale_f[j0/nwarps]*KQ_rowsum_f[j0/nwarps] + KQ_rowsum_add;
+            } else {
+                half2 KQ2_tmp[FATTN_KQ_STRIDE/(2*WARP_SIZE)];
+#pragma unroll
+                for (int k0 = 0; k0 < FATTN_KQ_STRIDE/2; k0 += WARP_SIZE) {
+                    const int k = k0 + threadIdx.x;
+
+                    KQ2_tmp[k0/WARP_SIZE] = KQ2[j*(kqs_padded/2) + k];
+
+                    if (use_logit_softcap) {
+                        // There is no dedicated tangens hyperbolicus function for half2.
+                        KQ2_tmp[k0/WARP_SIZE] = h2exp(KQ2_tmp[k0/WARP_SIZE]*make_half2(2.0f, 2.0f));
+                        KQ2_tmp[k0/WARP_SIZE] = (KQ2_tmp[k0/WARP_SIZE] - make_half2(1.0f, 1.0f))
+                                               /(KQ2_tmp[k0/WARP_SIZE] + make_half2(1.0f, 1.0f));
+
+                        KQ2_tmp[k0/WARP_SIZE] *= logit_softcap_2;
+                    }
+                }
+
+                half2 KQ_max_new = KQ_max_h2[j0/nwarps];
+#pragma unroll
+                for (int k0 = 0; k0 < FATTN_KQ_STRIDE/2; k0 += WARP_SIZE) {
+                    const int k = k0 + threadIdx.x;
+
+                    KQ2_tmp[k0/WARP_SIZE] += mask ? slope2*mask2[(j*ne11 + k_VKQ_0)/2 + k] : make_half2(0.0f, 0.0f);
+                    KQ_max_new = ggml_cuda_hmax2(KQ_max_new, KQ2_tmp[k0/WARP_SIZE]);
+                }
+                KQ_max_new = __half2half2(warp_reduce_max(ggml_cuda_hmax(__low2half(KQ_max_new), __high2half(KQ_max_new))));
+                const half2 diff = KQ_max_h2[j0/nwarps] - KQ_max_new;
+                KQ_max_scale_h2[j0/nwarps] = h2exp(diff);
+                const uint32_t ftz_mask = __hgt2_mask(diff, make_half2(SOFTMAX_FTZ_THRESHOLD, SOFTMAX_FTZ_THRESHOLD));
+                *((uint32_t *) &KQ_max_scale_h2[j0/nwarps]) &= ftz_mask;
+                KQ_max_h2[j0/nwarps] = KQ_max_new;
+
+                half2 KQ_rowsum_add = make_half2(0.0f, 0.0f);
+#pragma unroll
+                for (int k0 = 0; k0 < FATTN_KQ_STRIDE/2; k0 += WARP_SIZE) {
+                    const int k = k0 + threadIdx.x;
+
+                    const half2 diff = KQ2_tmp[k0/WARP_SIZE] - KQ_max_h2[j0/nwarps];
+                    KQ2_tmp[k0/WARP_SIZE] = h2exp(diff);
+                    const uint32_t ftz_mask = __hgt2_mask(diff, make_half2(SOFTMAX_FTZ_THRESHOLD, SOFTMAX_FTZ_THRESHOLD));
+                    *((uint32_t *) &KQ2_tmp[k0/WARP_SIZE]) &= ftz_mask;
+                    KQ_rowsum_add += KQ2_tmp[k0/WARP_SIZE];
+                    KQ2[j*(kqs_padded/2) + k] = KQ2_tmp[k0/WARP_SIZE];
+                }
+                KQ_rowsum_add = warp_reduce_sum(KQ_rowsum_add);
+
+                // Scale previous KQ_rowsum to account for a potential increase in KQ_max:
+                KQ_rowsum_h2[j0/nwarps] = KQ_max_scale_h2[j0/nwarps]*KQ_rowsum_h2[j0/nwarps] + KQ_rowsum_add;
+            }
+        }
+
+        __syncthreads();
+
+        frag_b KQ_b[FATTN_KQ_STRIDE/(VKQ_ratio*16)][ncols/frag_n];
+#pragma unroll
+        for (int j0 = 0; j0 < ncols; j0 += frag_n) {
+#pragma unroll
+            for (int k0 = 0; k0 < FATTN_KQ_STRIDE; k0 += VKQ_ratio*16) {
+                const int k = k0 + (threadIdx.y % VKQ_ratio)*16;
+                nvcuda::wmma::load_matrix_sync(
+                    KQ_b[k0/(VKQ_ratio*16)][j0/frag_n],
+                    KQ + j0*(kqar*kqs_padded) + k,
+                    kqar*kqs_padded);
+            }
+        }
+
+        frag_c_VKQ VKQ_c[D/VKQ_stride][ncols/frag_n];
+#pragma unroll
+        for (int i_VKQ_0 = 0; i_VKQ_0 < D; i_VKQ_0 += VKQ_stride) {
+#pragma unroll
+            for (int j = 0; j < ncols/frag_n; ++j) {
+                nvcuda::wmma::fill_fragment(VKQ_c[i_VKQ_0/VKQ_stride][j], 0.0f);
+            }
+
+#pragma unroll
+            for (int k0 = 0; k0 < FATTN_KQ_STRIDE; k0 += VKQ_ratio*16) {
+                const int k = k0 + (threadIdx.y % VKQ_ratio)*16;
+
+                frag_a_V v_a;
+                nvcuda::wmma::load_matrix_sync(v_a, V_h + (k_VKQ_0 + k)*stride_KV + i_VKQ_0 + frag_m*(threadIdx.y/VKQ_ratio), stride_KV);
+#pragma unroll
+                for (int j = 0; j < ncols/frag_n; ++j) {
+                    nvcuda::wmma::mma_sync(VKQ_c[i_VKQ_0/VKQ_stride][j], v_a, KQ_b[k0/(VKQ_ratio*16)][j], VKQ_c[i_VKQ_0/VKQ_stride][j]);
+                }
+            }
+        }
+
+        __syncthreads();
+
+        const int offset_k = (threadIdx.y % VKQ_ratio) * (ncols*D_padded);
+#pragma unroll
+        for (int i_KQ_0 = 0; i_KQ_0 < D; i_KQ_0 += VKQ_stride) {
+#pragma unroll
+            for (int j0 = 0; j0 < ncols; j0 += frag_n) {
+                nvcuda::wmma::store_matrix_sync(
+                    KQ + offset_k + j0*D_padded + i_KQ_0 + frag_m*(threadIdx.y/VKQ_ratio),
+                    VKQ_c[i_KQ_0/VKQ_stride][j0/frag_n],
+                    D_padded, nvcuda::wmma::mem_col_major);
+            }
+        }
+
+        __syncthreads();
+
+#pragma unroll
+        for (int j0 = 0; j0 < ncols; j0 += nwarps) {
+            const int j = j0 + threadIdx.y;
+
+            half2 VKQ_scale;
+            if (std::is_same<KQ_acc_t, float>::value) {
+                VKQ_scale = make_half2(KQ_max_scale_f[j0/nwarps], KQ_max_scale_f[j0/nwarps]);
+            } else {
+                VKQ_scale = KQ_max_scale_h2[j0/nwarps];
+            }
+
+#pragma unroll
+            for (int i0 = 0; i0 < D/2; i0 += WARP_SIZE) {
+                const int i = i0 + threadIdx.x;
+                if (i0 + WARP_SIZE > D/2 && i >= D/2) {
+                    break;
+                }
+
+                half2 VKQ_add = make_half2(0.0f, 0.0f);
+#pragma unroll
+                for (int l = 0; l < VKQ_ratio; ++l) {
+                    VKQ_add += KQ2[l*(ncols*D_padded/2) + j*(D_padded/2) + i];
+                }
+                VKQ2[j*(D_padded/2) + i] = VKQ_scale*VKQ2[j*(D_padded/2) + i] + VKQ_add;
+            }
+        }
+
+        __syncthreads();
+    }
+
+#pragma unroll
+    for (int j0 = 0; j0 < ncols; j0 += nwarps) {
+        const int j_VKQ = j0 + threadIdx.y;
+        if (ic0 + j_VKQ >= ne01) {
+            return;
+        }
+        const int j_dst = (ic0 + j_VKQ)*parallel_blocks + ip;
+
+        float KQ_rowsum_j;
+        if (std::is_same<KQ_acc_t, float>::value) {
+            KQ_rowsum_j = KQ_rowsum_f[j0/nwarps];
+        } else {
+            KQ_rowsum_j = __low2float(KQ_rowsum_h2[j0/nwarps]) + __high2float(KQ_rowsum_h2[j0/nwarps]);
+        }
+
+#pragma unroll
+        for (int i0 = 0; i0 < D; i0 += WARP_SIZE) {
+            const int i = i0 + threadIdx.x;
+            if (i0 + WARP_SIZE > D && i >= D) {
+                break;
+            }
+            float dst_val = VKQ[j_VKQ*D_padded + i];
+            if (parallel_blocks == 1) {
+                dst_val /= KQ_rowsum_j;
+            }
+            dst[j_dst*gridDim.y*D + blockIdx.y*D + i] = dst_val;
+        }
+
+        if (parallel_blocks == 1 || threadIdx.x != 0) {
+            continue;
+        }
+
+        float2 dst_meta_val;
+        if (std::is_same<KQ_acc_t, float>::value) {
+            dst_meta_val.x = KQ_max_f[j0/nwarps];
+        } else {
+            dst_meta_val.x = __low2float(KQ_max_h2[j0/nwarps]);
+        }
+        dst_meta_val.y = KQ_rowsum_j;
+        dst_meta[(ic0 + j_VKQ)*gridDim.y*parallel_blocks + blockIdx.y*parallel_blocks + ip] = dst_meta_val;
+    }
+#else
+   NO_DEVICE_CODE;
+#endif // FP16_MMA_AVAILABLE
+}
+
+constexpr int get_max_power_of_2(int x) {
+    return x % 2 == 0 ? 2*get_max_power_of_2(x/2) : 1;
+}
+
+static_assert(get_max_power_of_2(1) == 1, "Test failed.");
+static_assert(get_max_power_of_2(2) == 2, "Test failed.");
+static_assert(get_max_power_of_2(4) == 4, "Test failed.");
+static_assert(get_max_power_of_2(6) == 2, "Test failed.");
+
+// Number of VKQ rows calculated in parallel:
+constexpr int get_VKQ_stride(int D, int nwarps, int frag_m) {
+    return (get_max_power_of_2(D/frag_m) < nwarps ? get_max_power_of_2(D/frag_m) : nwarps)*frag_m;
+}
+
+static_assert(get_VKQ_stride(128, 1, 32) ==  32, "Test failed.");
+static_assert(get_VKQ_stride(128, 2, 32) ==  64, "Test failed.");
+static_assert(get_VKQ_stride(128, 4, 32) == 128, "Test failed.");
+static_assert(get_VKQ_stride( 64, 1, 32) ==  32, "Test failed.");
+static_assert(get_VKQ_stride( 64, 2, 32) ==  64, "Test failed.");
+static_assert(get_VKQ_stride( 64, 4, 32) ==  64, "Test failed.");
+static_assert(get_VKQ_stride( 80, 1, 16) ==  16, "Test failed.");
+static_assert(get_VKQ_stride( 80, 2, 16) ==  16, "Test failed.");
+static_assert(get_VKQ_stride( 80, 4, 16) ==  16, "Test failed.");
+
+template <int D, int cols_per_block, typename KQ_acc_t>
+void ggml_cuda_flash_attn_ext_wmma_f16_case(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
+    const ggml_tensor * KQV = dst;
+    const ggml_tensor * Q   = dst->src[0];
+
+    constexpr int nwarps = 4;
+
+    constexpr int frag_m = cols_per_block == 8 && D % 32 == 0 ? 32 : 16;
+    const int blocks_num_pb1 = ((Q->ne[1] + cols_per_block - 1) / cols_per_block)*Q->ne[2]*Q->ne[3];
+    const int nsm = ggml_cuda_info().devices[ggml_cuda_get_device()].nsm;
+
+    float logit_softcap;
+    memcpy(&logit_softcap, (const float *) KQV->op_params + 2, sizeof(float));
+
+    if (4*blocks_num_pb1 < 2*nsm) {
+        constexpr int parallel_blocks = 4;
+        fattn_kernel_t fattn_kernel;
+        if (logit_softcap == 0.0f) {
+            constexpr bool use_logit_softcap = false;
+            fattn_kernel = flash_attn_ext_f16<
+                D, cols_per_block, nwarps, get_VKQ_stride(D, nwarps, frag_m), parallel_blocks, KQ_acc_t, use_logit_softcap>;
+        } else {
+            constexpr bool use_logit_softcap = true;
+            fattn_kernel = flash_attn_ext_f16<
+                D, cols_per_block, nwarps, get_VKQ_stride(D, nwarps, frag_m), parallel_blocks, KQ_acc_t, use_logit_softcap>;
+        }
+        launch_fattn<D, parallel_blocks>(ctx, dst, fattn_kernel, nwarps, cols_per_block, true, true);
+        return;
+    }
+    if (2*blocks_num_pb1 < 2*nsm) {
+        constexpr int parallel_blocks = 2;
+        fattn_kernel_t fattn_kernel;
+        if (logit_softcap == 0.0f) {
+            constexpr bool use_logit_softcap = false;
+            fattn_kernel = flash_attn_ext_f16<
+                D, cols_per_block, nwarps, get_VKQ_stride(D, nwarps, frag_m), parallel_blocks, KQ_acc_t, use_logit_softcap>;
+        } else {
+            constexpr bool use_logit_softcap = true;
+            fattn_kernel = flash_attn_ext_f16<
+                D, cols_per_block, nwarps, get_VKQ_stride(D, nwarps, frag_m), parallel_blocks, KQ_acc_t, use_logit_softcap>;
+        }
+        launch_fattn<D, parallel_blocks>(ctx, dst, fattn_kernel, nwarps, cols_per_block, true, true);
+        return;
+    }
+    constexpr int parallel_blocks = 1;
+    fattn_kernel_t fattn_kernel;
+    if (logit_softcap == 0.0f) {
+        constexpr bool use_logit_softcap = false;
+        fattn_kernel = flash_attn_ext_f16<
+            D, cols_per_block, nwarps, get_VKQ_stride(D, nwarps, frag_m), parallel_blocks, KQ_acc_t, use_logit_softcap>;
+    } else {
+        constexpr bool use_logit_softcap = true;
+        fattn_kernel = flash_attn_ext_f16<
+            D, cols_per_block, nwarps, get_VKQ_stride(D, nwarps, frag_m), parallel_blocks, KQ_acc_t, use_logit_softcap>;
+    }
+    launch_fattn<D, parallel_blocks>(ctx, dst, fattn_kernel, nwarps, cols_per_block, true, true);
+}
+
+#define DECL_FATTN_WMMA_F16_CASE(D, cols_per_block, KQ_acc_t)                         \
+    template void ggml_cuda_flash_attn_ext_wmma_f16_case                              \
+    <D, cols_per_block, KQ_acc_t>(ggml_backend_cuda_context & ctx, ggml_tensor * dst) \
+
+extern DECL_FATTN_WMMA_F16_CASE( 64, 16, float);
+extern DECL_FATTN_WMMA_F16_CASE( 80, 16, float);
+extern DECL_FATTN_WMMA_F16_CASE( 96, 16, float);
+extern DECL_FATTN_WMMA_F16_CASE(112, 16, float);
+extern DECL_FATTN_WMMA_F16_CASE(128, 16, float);
+extern DECL_FATTN_WMMA_F16_CASE(256, 16, float);
+
+extern DECL_FATTN_WMMA_F16_CASE( 64, 32, float);
+extern DECL_FATTN_WMMA_F16_CASE( 80, 32, float);
+extern DECL_FATTN_WMMA_F16_CASE( 96, 32, float);
+extern DECL_FATTN_WMMA_F16_CASE(112, 32, float);
+extern DECL_FATTN_WMMA_F16_CASE(128, 32, float);
+// extern DECL_FATTN_WMMA_F16_CASE(256, 16, float);
+
+extern DECL_FATTN_WMMA_F16_CASE( 64,  8, half);
+extern DECL_FATTN_WMMA_F16_CASE( 96,  8, half);
+extern DECL_FATTN_WMMA_F16_CASE(128,  8, half);
+extern DECL_FATTN_WMMA_F16_CASE(256,  8, half);
+
+extern DECL_FATTN_WMMA_F16_CASE( 64, 16, half);
+extern DECL_FATTN_WMMA_F16_CASE( 80, 16, half);
+extern DECL_FATTN_WMMA_F16_CASE( 96, 16, half);
+extern DECL_FATTN_WMMA_F16_CASE(112, 16, half);
+extern DECL_FATTN_WMMA_F16_CASE(128, 16, half);
+extern DECL_FATTN_WMMA_F16_CASE(256, 16, half);
+
+extern DECL_FATTN_WMMA_F16_CASE( 64, 32, half);
+extern DECL_FATTN_WMMA_F16_CASE( 80, 32, half);
+extern DECL_FATTN_WMMA_F16_CASE( 96, 32, half);
+extern DECL_FATTN_WMMA_F16_CASE(112, 32, half);
+extern DECL_FATTN_WMMA_F16_CASE(128, 32, half);
+extern DECL_FATTN_WMMA_F16_CASE(256, 16, half);
diff --git a/src/whisper_cpp/ggml/src/ggml-cuda/fattn.cu b/src/whisper_cpp/ggml/src/ggml-cuda/fattn.cu
new file mode 100644
index 00000000..83e5589a
--- /dev/null
+++ b/src/whisper_cpp/ggml/src/ggml-cuda/fattn.cu
@@ -0,0 +1,345 @@
+#include "common.cuh"
+#include "fattn-common.cuh"
+#include "fattn-tile-f16.cuh"
+#include "fattn-tile-f32.cuh"
+#include "fattn-vec-f16.cuh"
+#include "fattn-vec-f32.cuh"
+#include "fattn-wmma-f16.cuh"
+#include "fattn.cuh"
+
+#include <cstdint>
+
+static void ggml_cuda_flash_attn_ext_wmma_f16(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
+    const ggml_tensor * KQV = dst;
+    const ggml_tensor * Q   = dst->src[0];
+
+    const int32_t precision = KQV->op_params[3];
+
+    if (precision != GGML_PREC_DEFAULT) {
+        if (Q->ne[1] <= 32 || Q->ne[0] > 128) {
+            constexpr int cols_per_block = 16;
+            switch (Q->ne[0]) {
+                case 64:
+                    ggml_cuda_flash_attn_ext_wmma_f16_case< 64, cols_per_block, float>(ctx, dst);
+                    break;
+                case 80:
+                    ggml_cuda_flash_attn_ext_wmma_f16_case< 80, cols_per_block, float>(ctx, dst);
+                    break;
+                case 96:
+                    ggml_cuda_flash_attn_ext_wmma_f16_case< 96, cols_per_block, float>(ctx, dst);
+                    break;
+                case 112:
+                    ggml_cuda_flash_attn_ext_wmma_f16_case<112, cols_per_block, float>(ctx, dst);
+                    break;
+                case 128:
+                    ggml_cuda_flash_attn_ext_wmma_f16_case<128, cols_per_block, float>(ctx, dst);
+                    break;
+                case 256:
+                    ggml_cuda_flash_attn_ext_wmma_f16_case<256, cols_per_block, float>(ctx, dst);
+                    break;
+                default:
+                    GGML_ABORT("fatal error");
+                    break;
+            }
+        } else {
+            constexpr int cols_per_block = 32;
+            switch (Q->ne[0]) {
+                case 64:
+                    ggml_cuda_flash_attn_ext_wmma_f16_case< 64, cols_per_block, float>(ctx, dst);
+                    break;
+                case 80:
+                    ggml_cuda_flash_attn_ext_wmma_f16_case< 80, cols_per_block, float>(ctx, dst);
+                    break;
+                case 96:
+                    ggml_cuda_flash_attn_ext_wmma_f16_case< 96, cols_per_block, float>(ctx, dst);
+                    break;
+                case 112:
+                    ggml_cuda_flash_attn_ext_wmma_f16_case<112, cols_per_block, float>(ctx, dst);
+                    break;
+                case 128:
+                    ggml_cuda_flash_attn_ext_wmma_f16_case<128, cols_per_block, float>(ctx, dst);
+                    break;
+                // case 256:
+                //     ggml_cuda_flash_attn_ext_wmma_f16_case<128, cols_per_block, float>(ctx, dst);
+                //     break;
+                default:
+                    GGML_ABORT("fatal error");
+                    break;
+            }
+        }
+        return;
+    }
+
+    if (Q->ne[1] <= 8 && Q->ne[0] % WARP_SIZE == 0) {
+        constexpr int cols_per_block = 8;
+        switch (Q->ne[0]) {
+            case 64:
+                ggml_cuda_flash_attn_ext_wmma_f16_case< 64, cols_per_block, half>(ctx, dst);
+                break;
+            case 96:
+                ggml_cuda_flash_attn_ext_wmma_f16_case< 96, cols_per_block, half>(ctx, dst);
+                break;
+            case 128:
+                ggml_cuda_flash_attn_ext_wmma_f16_case<128, cols_per_block, half>(ctx, dst);
+                break;
+            case 256:
+                ggml_cuda_flash_attn_ext_wmma_f16_case<256, cols_per_block, half>(ctx, dst);
+                break;
+            default:
+                GGML_ABORT("fatal error");
+                break;
+        }
+        return;
+    }
+
+    if (Q->ne[1] <= 32) {
+        constexpr int cols_per_block = 16;
+        switch (Q->ne[0]) {
+            case 64:
+                ggml_cuda_flash_attn_ext_wmma_f16_case< 64, cols_per_block, half>(ctx, dst);
+                break;
+            case 80:
+                ggml_cuda_flash_attn_ext_wmma_f16_case< 80, cols_per_block, half>(ctx, dst);
+                break;
+            case 96:
+                ggml_cuda_flash_attn_ext_wmma_f16_case< 96, cols_per_block, half>(ctx, dst);
+                break;
+            case 112:
+                ggml_cuda_flash_attn_ext_wmma_f16_case<112, cols_per_block, half>(ctx, dst);
+                break;
+            case 128:
+                ggml_cuda_flash_attn_ext_wmma_f16_case<128, cols_per_block, half>(ctx, dst);
+                break;
+            case 256:
+                ggml_cuda_flash_attn_ext_wmma_f16_case<256, cols_per_block, half>(ctx, dst);
+                break;
+            default:
+                GGML_ABORT("fatal error");
+                break;
+        }
+        return;
+    }
+
+    constexpr int cols_per_block = 32;
+    switch (Q->ne[0]) {
+        case 64:
+            ggml_cuda_flash_attn_ext_wmma_f16_case< 64, cols_per_block, half>(ctx, dst);
+            break;
+        case 80:
+            ggml_cuda_flash_attn_ext_wmma_f16_case< 80, cols_per_block, half>(ctx, dst);
+            break;
+        case 96:
+            ggml_cuda_flash_attn_ext_wmma_f16_case< 96, cols_per_block, half>(ctx, dst);
+            break;
+        case 112:
+            ggml_cuda_flash_attn_ext_wmma_f16_case<112, cols_per_block, half>(ctx, dst);
+            break;
+        case 128:
+            ggml_cuda_flash_attn_ext_wmma_f16_case<128, cols_per_block, half>(ctx, dst);
+            break;
+        case 256:
+            ggml_cuda_flash_attn_ext_wmma_f16_case<256, cols_per_block, half>(ctx, dst);
+            break;
+        default:
+            GGML_ABORT("fatal error");
+            break;
+    }
+}
+#define FATTN_VEC_F16_CASE(D, type_K, type_V)                               \
+    if (Q->ne[0] == (D) && K->type == (type_K) && V->type == (type_V)) {    \
+        ggml_cuda_flash_attn_ext_vec_f16_case<D, type_K, type_V>(ctx, dst); \
+        return;                                                             \
+    }                                                                       \
+
+static void ggml_cuda_flash_attn_ext_vec_f16(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
+    ggml_tensor * Q = dst->src[0];
+    ggml_tensor * K = dst->src[1];
+    ggml_tensor * V = dst->src[2];
+
+#ifdef GGML_CUDA_FA_ALL_QUANTS
+    FATTN_VEC_F16_CASE( 64, GGML_TYPE_F16, GGML_TYPE_Q4_0)
+    FATTN_VEC_F16_CASE( 64, GGML_TYPE_F16, GGML_TYPE_Q4_1)
+    FATTN_VEC_F16_CASE( 64, GGML_TYPE_F16, GGML_TYPE_Q5_0)
+    FATTN_VEC_F16_CASE( 64, GGML_TYPE_F16, GGML_TYPE_Q5_1)
+    FATTN_VEC_F16_CASE( 64, GGML_TYPE_F16, GGML_TYPE_Q8_0)
+    FATTN_VEC_F16_CASE( 64, GGML_TYPE_F16, GGML_TYPE_F16 )
+
+    FATTN_VEC_F16_CASE(128, GGML_TYPE_Q4_0, GGML_TYPE_Q4_0)
+    FATTN_VEC_F16_CASE(128, GGML_TYPE_Q4_1, GGML_TYPE_Q4_0)
+    FATTN_VEC_F16_CASE(128, GGML_TYPE_Q5_0, GGML_TYPE_Q4_0)
+    FATTN_VEC_F16_CASE(128, GGML_TYPE_Q5_1, GGML_TYPE_Q4_0)
+    FATTN_VEC_F16_CASE(128, GGML_TYPE_Q8_0, GGML_TYPE_Q4_0)
+    FATTN_VEC_F16_CASE(128, GGML_TYPE_F16,  GGML_TYPE_Q4_0)
+
+    FATTN_VEC_F16_CASE(128, GGML_TYPE_Q4_0, GGML_TYPE_Q4_1)
+    FATTN_VEC_F16_CASE(128, GGML_TYPE_Q4_1, GGML_TYPE_Q4_1)
+    FATTN_VEC_F16_CASE(128, GGML_TYPE_Q5_0, GGML_TYPE_Q4_1)
+    FATTN_VEC_F16_CASE(128, GGML_TYPE_Q5_1, GGML_TYPE_Q4_1)
+    FATTN_VEC_F16_CASE(128, GGML_TYPE_Q8_0, GGML_TYPE_Q4_1)
+    FATTN_VEC_F16_CASE(128, GGML_TYPE_F16,  GGML_TYPE_Q4_1)
+
+    FATTN_VEC_F16_CASE(128, GGML_TYPE_Q4_0, GGML_TYPE_Q5_0)
+    FATTN_VEC_F16_CASE(128, GGML_TYPE_Q4_1, GGML_TYPE_Q5_0)
+    FATTN_VEC_F16_CASE(128, GGML_TYPE_Q5_0, GGML_TYPE_Q5_0)
+    FATTN_VEC_F16_CASE(128, GGML_TYPE_Q5_1, GGML_TYPE_Q5_0)
+    FATTN_VEC_F16_CASE(128, GGML_TYPE_Q8_0, GGML_TYPE_Q5_0)
+    FATTN_VEC_F16_CASE(128, GGML_TYPE_F16,  GGML_TYPE_Q5_0)
+
+    FATTN_VEC_F16_CASE(128, GGML_TYPE_Q4_0, GGML_TYPE_Q5_1)
+    FATTN_VEC_F16_CASE(128, GGML_TYPE_Q4_1, GGML_TYPE_Q5_1)
+    FATTN_VEC_F16_CASE(128, GGML_TYPE_Q5_0, GGML_TYPE_Q5_1)
+    FATTN_VEC_F16_CASE(128, GGML_TYPE_Q5_1, GGML_TYPE_Q5_1)
+    FATTN_VEC_F16_CASE(128, GGML_TYPE_Q8_0, GGML_TYPE_Q5_1)
+    FATTN_VEC_F16_CASE(128, GGML_TYPE_F16,  GGML_TYPE_Q5_1)
+
+    FATTN_VEC_F16_CASE(128, GGML_TYPE_Q4_0, GGML_TYPE_Q8_0)
+    FATTN_VEC_F16_CASE(128, GGML_TYPE_Q4_1, GGML_TYPE_Q8_0)
+    FATTN_VEC_F16_CASE(128, GGML_TYPE_Q5_0, GGML_TYPE_Q8_0)
+    FATTN_VEC_F16_CASE(128, GGML_TYPE_Q5_1, GGML_TYPE_Q8_0)
+    FATTN_VEC_F16_CASE(128, GGML_TYPE_Q8_0, GGML_TYPE_Q8_0)
+    FATTN_VEC_F16_CASE(128, GGML_TYPE_F16,  GGML_TYPE_Q8_0)
+
+    FATTN_VEC_F16_CASE(128, GGML_TYPE_Q4_0, GGML_TYPE_F16)
+    FATTN_VEC_F16_CASE(128, GGML_TYPE_Q4_1, GGML_TYPE_F16)
+    FATTN_VEC_F16_CASE(128, GGML_TYPE_Q5_0, GGML_TYPE_F16)
+    FATTN_VEC_F16_CASE(128, GGML_TYPE_Q5_1, GGML_TYPE_F16)
+    FATTN_VEC_F16_CASE(128, GGML_TYPE_Q8_0, GGML_TYPE_F16)
+    FATTN_VEC_F16_CASE(128, GGML_TYPE_F16,  GGML_TYPE_F16)
+
+    FATTN_VEC_F16_CASE(256, GGML_TYPE_F16, GGML_TYPE_F16)
+#else
+    FATTN_VEC_F16_CASE(128, GGML_TYPE_Q4_0, GGML_TYPE_Q4_0)
+
+    FATTN_VEC_F16_CASE(128, GGML_TYPE_Q8_0, GGML_TYPE_Q8_0)
+
+    FATTN_VEC_F16_CASE( 64, GGML_TYPE_F16, GGML_TYPE_F16)
+    FATTN_VEC_F16_CASE(128, GGML_TYPE_F16, GGML_TYPE_F16)
+    FATTN_VEC_F16_CASE(256, GGML_TYPE_F16, GGML_TYPE_F16)
+#endif // GGML_CUDA_FA_ALL_QUANTS
+
+    on_no_fattn_vec_case(Q->ne[0]);
+}
+
+#define FATTN_VEC_F32_CASE(D, type_K, type_V)                               \
+    if (Q->ne[0] == (D) && K->type == (type_K) && V->type == (type_V)) {    \
+        ggml_cuda_flash_attn_ext_vec_f32_case<D, type_K, type_V>(ctx, dst); \
+        return;                                                             \
+    }                                                                       \
+
+static void ggml_cuda_flash_attn_ext_vec_f32(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
+    ggml_tensor * Q = dst->src[0];
+    ggml_tensor * K = dst->src[1];
+    ggml_tensor * V = dst->src[2];
+
+#ifdef GGML_CUDA_FA_ALL_QUANTS
+    FATTN_VEC_F32_CASE( 64, GGML_TYPE_F16, GGML_TYPE_Q4_0)
+    FATTN_VEC_F32_CASE( 64, GGML_TYPE_F16, GGML_TYPE_Q4_1)
+    FATTN_VEC_F32_CASE( 64, GGML_TYPE_F16, GGML_TYPE_Q5_0)
+    FATTN_VEC_F32_CASE( 64, GGML_TYPE_F16, GGML_TYPE_Q5_1)
+    FATTN_VEC_F32_CASE( 64, GGML_TYPE_F16, GGML_TYPE_Q8_0)
+    FATTN_VEC_F32_CASE( 64, GGML_TYPE_F16, GGML_TYPE_F16)
+
+    FATTN_VEC_F32_CASE(128, GGML_TYPE_Q4_0, GGML_TYPE_Q4_0)
+    FATTN_VEC_F32_CASE(128, GGML_TYPE_Q4_1, GGML_TYPE_Q4_0)
+    FATTN_VEC_F32_CASE(128, GGML_TYPE_Q5_0, GGML_TYPE_Q4_0)
+    FATTN_VEC_F32_CASE(128, GGML_TYPE_Q5_1, GGML_TYPE_Q4_0)
+    FATTN_VEC_F32_CASE(128, GGML_TYPE_Q8_0, GGML_TYPE_Q4_0)
+    FATTN_VEC_F32_CASE(128, GGML_TYPE_F16,  GGML_TYPE_Q4_0)
+
+    FATTN_VEC_F32_CASE(128, GGML_TYPE_Q4_0, GGML_TYPE_Q4_1)
+    FATTN_VEC_F32_CASE(128, GGML_TYPE_Q4_1, GGML_TYPE_Q4_1)
+    FATTN_VEC_F32_CASE(128, GGML_TYPE_Q5_0, GGML_TYPE_Q4_1)
+    FATTN_VEC_F32_CASE(128, GGML_TYPE_Q5_1, GGML_TYPE_Q4_1)
+    FATTN_VEC_F32_CASE(128, GGML_TYPE_Q8_0, GGML_TYPE_Q4_1)
+    FATTN_VEC_F32_CASE(128, GGML_TYPE_F16,  GGML_TYPE_Q4_1)
+
+    FATTN_VEC_F32_CASE(128, GGML_TYPE_Q4_0, GGML_TYPE_Q5_0)
+    FATTN_VEC_F32_CASE(128, GGML_TYPE_Q4_1, GGML_TYPE_Q5_0)
+    FATTN_VEC_F32_CASE(128, GGML_TYPE_Q5_0, GGML_TYPE_Q5_0)
+    FATTN_VEC_F32_CASE(128, GGML_TYPE_Q5_1, GGML_TYPE_Q5_0)
+    FATTN_VEC_F32_CASE(128, GGML_TYPE_Q8_0, GGML_TYPE_Q5_0)
+    FATTN_VEC_F32_CASE(128, GGML_TYPE_F16,  GGML_TYPE_Q5_0)
+
+    FATTN_VEC_F32_CASE(128, GGML_TYPE_Q4_0, GGML_TYPE_Q5_1)
+    FATTN_VEC_F32_CASE(128, GGML_TYPE_Q4_1, GGML_TYPE_Q5_1)
+    FATTN_VEC_F32_CASE(128, GGML_TYPE_Q5_0, GGML_TYPE_Q5_1)
+    FATTN_VEC_F32_CASE(128, GGML_TYPE_Q5_1, GGML_TYPE_Q5_1)
+    FATTN_VEC_F32_CASE(128, GGML_TYPE_Q8_0, GGML_TYPE_Q5_1)
+    FATTN_VEC_F32_CASE(128, GGML_TYPE_F16,  GGML_TYPE_Q5_1)
+
+    FATTN_VEC_F32_CASE(128, GGML_TYPE_Q4_0, GGML_TYPE_Q8_0)
+    FATTN_VEC_F32_CASE(128, GGML_TYPE_Q4_1, GGML_TYPE_Q8_0)
+    FATTN_VEC_F32_CASE(128, GGML_TYPE_Q5_0, GGML_TYPE_Q8_0)
+    FATTN_VEC_F32_CASE(128, GGML_TYPE_Q5_1, GGML_TYPE_Q8_0)
+    FATTN_VEC_F32_CASE(128, GGML_TYPE_Q8_0, GGML_TYPE_Q8_0)
+    FATTN_VEC_F32_CASE(128, GGML_TYPE_F16,  GGML_TYPE_Q8_0)
+
+    FATTN_VEC_F32_CASE(128, GGML_TYPE_Q4_0, GGML_TYPE_F16)
+    FATTN_VEC_F32_CASE(128, GGML_TYPE_Q4_1, GGML_TYPE_F16)
+    FATTN_VEC_F32_CASE(128, GGML_TYPE_Q5_0, GGML_TYPE_F16)
+    FATTN_VEC_F32_CASE(128, GGML_TYPE_Q5_1, GGML_TYPE_F16)
+    FATTN_VEC_F32_CASE(128, GGML_TYPE_Q8_0, GGML_TYPE_F16)
+    FATTN_VEC_F32_CASE(128, GGML_TYPE_F16,  GGML_TYPE_F16)
+
+    FATTN_VEC_F32_CASE(256, GGML_TYPE_F16, GGML_TYPE_F16)
+#else
+    FATTN_VEC_F32_CASE(128, GGML_TYPE_Q4_0, GGML_TYPE_Q4_0)
+
+    FATTN_VEC_F32_CASE(128, GGML_TYPE_Q8_0, GGML_TYPE_Q8_0)
+
+    FATTN_VEC_F32_CASE( 64, GGML_TYPE_F16, GGML_TYPE_F16)
+    FATTN_VEC_F32_CASE(128, GGML_TYPE_F16, GGML_TYPE_F16)
+    FATTN_VEC_F32_CASE(256, GGML_TYPE_F16, GGML_TYPE_F16)
+#endif // GGML_CUDA_FA_ALL_QUANTS
+
+    on_no_fattn_vec_case(Q->ne[0]);
+}
+
+void ggml_cuda_flash_attn_ext(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
+    const ggml_tensor * KQV = dst;
+    const ggml_tensor * Q   = dst->src[0];
+
+    ggml_cuda_set_device(ctx.device);
+    const int cc = ggml_cuda_info().devices[ggml_cuda_get_device()].cc;
+    const int32_t precision = KQV->op_params[3];
+
+    // On AMD the tile kernels perform poorly, use the vec kernel instead:
+    if (cc >= CC_OFFSET_AMD) {
+        if (precision == GGML_PREC_DEFAULT && fast_fp16_available(cc)) {
+            ggml_cuda_flash_attn_ext_vec_f16(ctx, dst);
+        } else {
+            ggml_cuda_flash_attn_ext_vec_f32(ctx, dst);
+        }
+        return;
+    }
+
+    if (!fast_fp16_available(cc)) {
+        if (Q->ne[1] <= 8 || Q->ne[0] == 256) {
+            ggml_cuda_flash_attn_ext_vec_f32(ctx, dst);
+        } else {
+            ggml_cuda_flash_attn_ext_tile_f32(ctx, dst);
+        }
+        return;
+    }
+
+    if (!fp16_mma_available(cc)) {
+        if (Q->ne[1] <= 8) {
+            ggml_cuda_flash_attn_ext_vec_f16(ctx, dst);
+        } else {
+            ggml_cuda_flash_attn_ext_tile_f16(ctx, dst);
+        }
+        return;
+    }
+
+    if (Q->ne[1] == 1 && Q->ne[0] % (2*WARP_SIZE) == 0) {
+        if (precision == GGML_PREC_DEFAULT) {
+            ggml_cuda_flash_attn_ext_vec_f16(ctx, dst);
+            return;
+        } else if(Q->ne[0] <= 128) {
+            ggml_cuda_flash_attn_ext_vec_f32(ctx, dst);
+            return;
+        }
+    }
+
+    ggml_cuda_flash_attn_ext_wmma_f16(ctx, dst);
+}
diff --git a/src/whisper_cpp/ggml/src/ggml-cuda/fattn.cuh b/src/whisper_cpp/ggml/src/ggml-cuda/fattn.cuh
new file mode 100644
index 00000000..ad3ca7a8
--- /dev/null
+++ b/src/whisper_cpp/ggml/src/ggml-cuda/fattn.cuh
@@ -0,0 +1,3 @@
+#include "common.cuh"
+
+void ggml_cuda_flash_attn_ext(ggml_backend_cuda_context & ctx, ggml_tensor * dst);
diff --git a/src/whisper_cpp/ggml/src/ggml-cuda/getrows.cu b/src/whisper_cpp/ggml/src/ggml-cuda/getrows.cu
new file mode 100644
index 00000000..4c370323
--- /dev/null
+++ b/src/whisper_cpp/ggml/src/ggml-cuda/getrows.cu
@@ -0,0 +1,177 @@
+#include "getrows.cuh"
+#include "dequantize.cuh"
+
+template<int qk, int qr, dequantize_kernel_t dequantize_kernel, typename dst_t>
+static __global__ void k_get_rows(
+            const void * src0, const int32_t * src1, dst_t * dst,
+            int64_t ne00, /*int64_t ne01, int64_t ne02, int64_t ne03,*/
+            /*int64_t ne10, int64_t ne11,*/ int64_t ne12, /*int64_t ne13,*/
+            /*size_t s0,*/ size_t s1, size_t s2, size_t s3,
+            /*size_t nb00,*/ size_t nb01, size_t nb02, size_t nb03,
+            size_t s10, size_t s11, size_t s12/*, size_t s13*/) {
+
+    const int i00 = (blockIdx.x*blockDim.x + threadIdx.x)*2;
+    const int i10 = blockDim.y*blockIdx.y + threadIdx.y;
+    const int i11 = (blockIdx.z*blockDim.z + threadIdx.z)/ne12;
+    const int i12 = (blockIdx.z*blockDim.z + threadIdx.z)%ne12;
+
+    if (i00 >= ne00) {
+        return;
+    }
+
+    const int i01 = src1[i10*s10 + i11*s11 + i12*s12];
+
+    dst_t * dst_row = dst + i10*s1 + i11*s2 + i12*s3;
+    const void * src0_row = (const char *)src0 + i01*nb01 + i11*nb02 + i12*nb03;
+
+    const int ib = i00/qk; // block index
+    const int iqs = (i00%qk)/qr; // quant index
+    const int iybs = i00 - i00%qk; // dst block start index
+    const int y_offset = qr == 1 ? 1 : qk/2;
+
+    // dequantize
+    dfloat2 v;
+    dequantize_kernel(src0_row, ib, iqs, v);
+
+    dst_row[iybs + iqs + 0]        = v.x;
+    dst_row[iybs + iqs + y_offset] = v.y;
+}
+
+template<typename src0_t, typename dst_t>
+static __global__ void k_get_rows_float(
+            const src0_t * src0, const int32_t * src1, dst_t * dst,
+            int64_t ne00, /*int64_t ne01, int64_t ne02, int64_t ne03,*/
+            /*int64_t ne10, int64_t ne11,*/ int64_t ne12, /*int64_t ne13,*/
+            /*size_t s0,*/ size_t s1, size_t s2, size_t s3,
+            /*size_t nb00,*/ size_t nb01, size_t nb02, size_t nb03,
+            size_t s10, size_t s11, size_t s12/*, size_t s13*/) {
+
+    const int i00 = blockIdx.x*blockDim.x + threadIdx.x;
+    const int i10 = blockDim.y*blockIdx.y + threadIdx.y;
+    const int i11 = (blockIdx.z*blockDim.z + threadIdx.z)/ne12;
+    const int i12 = (blockIdx.z*blockDim.z + threadIdx.z)%ne12;
+
+    if (i00 >= ne00) {
+        return;
+    }
+
+    const int i01 = src1[i10*s10 + i11*s11 + i12*s12];
+
+    dst_t * dst_row = dst + i10*s1 + i11*s2 + i12*s3;
+    const src0_t * src0_row = (const src0_t *)((const char *)src0 + i01*nb01 + i11*nb02 + i12*nb03);
+
+    dst_row[i00] = src0_row[i00];
+}
+
+template<int qk, int qr, dequantize_kernel_t dq>
+static void get_rows_cuda(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst,
+                            const void * src0_dd, const int32_t * src1_dd, float * dst_dd, cudaStream_t stream) {
+
+    GGML_TENSOR_BINARY_OP_LOCALS
+
+    const dim3 block_dims(CUDA_GET_ROWS_BLOCK_SIZE, 1, 1);
+    const int block_num_x = (ne00 + 2*CUDA_GET_ROWS_BLOCK_SIZE - 1) / (2*CUDA_GET_ROWS_BLOCK_SIZE);
+    const dim3 block_nums(block_num_x, ne10, ne11*ne12);
+
+    // strides in elements
+    //const size_t s0 = nb0 / ggml_element_size(dst);
+    const size_t s1 = nb1 / ggml_element_size(dst);
+    const size_t s2 = nb2 / ggml_element_size(dst);
+    const size_t s3 = nb3 / ggml_element_size(dst);
+
+    const size_t s10 = nb10 / ggml_element_size(src1);
+    const size_t s11 = nb11 / ggml_element_size(src1);
+    const size_t s12 = nb12 / ggml_element_size(src1);
+    //const size_t s13 = nb13 / ggml_element_size(src1);
+
+    GGML_ASSERT(ne00 % 2 == 0);
+
+    k_get_rows<qk, qr, dq><<<block_nums, block_dims, 0, stream>>>(
+            src0_dd, src1_dd, dst_dd,
+            ne00, /*ne01, ne02, ne03,*/
+            /*ne10, ne11,*/ ne12, /*ne13,*/
+            /* s0,*/ s1, s2, s3,
+            /* nb00,*/ nb01, nb02, nb03,
+            s10, s11, s12/*, s13*/);
+
+    GGML_UNUSED(dst);
+}
+
+template<typename src0_t>
+static void get_rows_cuda_float(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst,
+                                const src0_t * src0_dd, const int32_t * src1_dd, float * dst_dd, cudaStream_t stream) {
+
+    GGML_TENSOR_BINARY_OP_LOCALS
+
+    const dim3 block_dims(CUDA_GET_ROWS_BLOCK_SIZE, 1, 1);
+    const int block_num_x = (ne00 + CUDA_GET_ROWS_BLOCK_SIZE - 1) / CUDA_GET_ROWS_BLOCK_SIZE;
+    const dim3 block_nums(block_num_x, ne10, ne11*ne12);
+
+    // strides in elements
+    //const size_t s0 = nb0 / ggml_element_size(dst);
+    const size_t s1 = nb1 / ggml_element_size(dst);
+    const size_t s2 = nb2 / ggml_element_size(dst);
+    const size_t s3 = nb3 / ggml_element_size(dst);
+
+    const size_t s10 = nb10 / ggml_element_size(src1);
+    const size_t s11 = nb11 / ggml_element_size(src1);
+    const size_t s12 = nb12 / ggml_element_size(src1);
+    //const size_t s13 = nb13 / ggml_element_size(src1);
+
+    k_get_rows_float<<<block_nums, block_dims, 0, stream>>>(
+            src0_dd, src1_dd, dst_dd,
+            ne00, /*ne01, ne02, ne03,*/
+            /*ne10, ne11,*/ ne12, /*ne13,*/
+            /* s0,*/ s1, s2, s3,
+            /* nb00,*/ nb01, nb02, nb03,
+            s10, s11, s12/*, s13*/);
+
+    GGML_UNUSED(dst);
+}
+
+void ggml_cuda_op_get_rows(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
+    const ggml_tensor * src0 = dst->src[0];
+    const ggml_tensor * src1 = dst->src[1];
+    const float * src0_d = (const float *)src0->data;
+    const float * src1_d = (const float *)src1->data;
+    float * dst_d = (float *)dst->data;
+    cudaStream_t stream = ctx.stream();
+
+
+    GGML_ASSERT(src1->type == GGML_TYPE_I32);
+    GGML_ASSERT(dst->type == GGML_TYPE_F32);
+
+    GGML_ASSERT(src0->nb[0] == ggml_type_size(src0->type));
+    GGML_ASSERT(src1->nb[0] == ggml_type_size(src1->type));
+    GGML_ASSERT(dst->nb[0] == ggml_type_size(dst->type));
+
+    const int32_t * src1_i32 = (const int32_t *) src1_d;
+
+    switch (src0->type) {
+        case GGML_TYPE_F16:
+            get_rows_cuda_float(src0, src1, dst, (const half *)src0_d, src1_i32, dst_d, stream);
+            break;
+        case GGML_TYPE_F32:
+            get_rows_cuda_float(src0, src1, dst, src0_d, src1_i32, dst_d, stream);
+            break;
+        case GGML_TYPE_Q4_0:
+            get_rows_cuda<QK4_0, QR4_0, dequantize_q4_0>(src0, src1, dst, src0_d, src1_i32, dst_d, stream);
+            break;
+        case GGML_TYPE_Q4_1:
+            get_rows_cuda<QK4_1, QR4_1, dequantize_q4_1>(src0, src1, dst, src0_d, src1_i32, dst_d, stream);
+            break;
+        case GGML_TYPE_Q5_0:
+            get_rows_cuda<QK5_0, QR5_0, dequantize_q5_0>(src0, src1, dst, src0_d, src1_i32, dst_d, stream);
+            break;
+        case GGML_TYPE_Q5_1:
+            get_rows_cuda<QK5_1, QR5_1, dequantize_q5_1>(src0, src1, dst, src0_d, src1_i32, dst_d, stream);
+            break;
+        case GGML_TYPE_Q8_0:
+            get_rows_cuda<QK8_0, QR8_0, dequantize_q8_0>(src0, src1, dst, src0_d, src1_i32, dst_d, stream);
+            break;
+        default:
+            // TODO: k-quants
+            GGML_ABORT("%s: unsupported type: %s\n", __func__, ggml_type_name(src0->type));
+            break;
+    }
+}
diff --git a/src/whisper_cpp/ggml/src/ggml-cuda/getrows.cuh b/src/whisper_cpp/ggml/src/ggml-cuda/getrows.cuh
new file mode 100644
index 00000000..bbf13023
--- /dev/null
+++ b/src/whisper_cpp/ggml/src/ggml-cuda/getrows.cuh
@@ -0,0 +1,5 @@
+#include "common.cuh"
+
+#define CUDA_GET_ROWS_BLOCK_SIZE 256
+
+void ggml_cuda_op_get_rows(ggml_backend_cuda_context & ctx, ggml_tensor * dst);
diff --git a/src/whisper_cpp/ggml/src/ggml-cuda/im2col.cu b/src/whisper_cpp/ggml/src/ggml-cuda/im2col.cu
new file mode 100644
index 00000000..16463ab0
--- /dev/null
+++ b/src/whisper_cpp/ggml/src/ggml-cuda/im2col.cu
@@ -0,0 +1,103 @@
+#include "im2col.cuh"
+
+template <typename T>
+static  __global__ void im2col_kernel(
+        const float * x, T * dst, int64_t batch_offset,
+        int64_t offset_delta, int64_t IC, int64_t IW, int64_t IH, int64_t OH, int64_t OW, int64_t KW, int64_t KH, int64_t pelements, int64_t CHW,
+        int s0, int s1, int p0, int p1, int d0, int d1) {
+    const int64_t i = threadIdx.x + blockIdx.x * blockDim.x;
+    if (i >= pelements) {
+        return;
+    }
+
+    const int64_t  ksize = OW * (KH > 1 ? KW : 1);
+    const int64_t  kx = i / ksize;
+    const int64_t  kd = kx * ksize;
+    const int64_t  ky = (i - kd) / OW;
+    const int64_t  ix = i % OW;
+
+    const int64_t  oh = blockIdx.y;
+    const int64_t  batch = blockIdx.z / IC;
+    const int64_t  ic = blockIdx.z % IC;
+
+    const int64_t iiw = ix * s0 + kx * d0 - p0;
+    const int64_t iih = oh * s1 + ky * d1 - p1;
+
+    const int64_t offset_dst =
+        ((batch * OH + oh) * OW + ix) * CHW +
+        (ic * (KW * KH) + ky * KW + kx);
+
+    if (iih < 0 || iih >= IH || iiw < 0 || iiw >= IW) {
+        dst[offset_dst] = 0.0f;
+    } else {
+        const int64_t offset_src = ic * offset_delta + batch * batch_offset;
+        dst[offset_dst] = x[offset_src + iih * IW + iiw];
+    }
+}
+
+template <typename T>
+static void im2col_cuda(const float * x, T* dst,
+    int64_t IW, int64_t IH, int64_t OW, int64_t OH, int64_t KW, int64_t KH, int64_t IC,
+    int64_t batch, int64_t batch_offset, int64_t offset_delta,
+    int s0,int s1,int p0,int p1,int d0,int d1, cudaStream_t stream) {
+    const int parallel_elements = OW * KW * KH;
+    const int num_blocks = (parallel_elements + CUDA_IM2COL_BLOCK_SIZE - 1) / CUDA_IM2COL_BLOCK_SIZE;
+    dim3 block_nums(num_blocks, OH, batch * IC);
+    im2col_kernel<<<block_nums, CUDA_IM2COL_BLOCK_SIZE, 0, stream>>>(x, dst, batch_offset, offset_delta, IC, IW, IH, OH, OW, KW, KH, parallel_elements, (IC * KH * KW), s0, s1, p0, p1, d0, d1);
+}
+
+static void im2col_cuda_f16(const float * x, half * dst,
+    int64_t IW, int64_t IH, int64_t OW, int64_t OH, int64_t KW, int64_t KH, int64_t IC,
+    int64_t batch, int64_t batch_offset, int64_t offset_delta,
+    int s0,int s1,int p0,int p1,int d0,int d1, cudaStream_t stream) {
+
+    im2col_cuda<half>(x, dst, IW, IH, OW, OH, KW, KH, IC, batch, batch_offset, offset_delta, s0, s1, p0, p1, d0, d1, stream);
+}
+
+static void im2col_cuda_f32(const float * x, float * dst,
+    int64_t IW, int64_t IH, int64_t OW, int64_t OH, int64_t KW, int64_t KH, int64_t IC,
+    int64_t batch, int64_t batch_offset, int64_t offset_delta,
+    int s0,int s1,int p0,int p1,int d0,int d1, cudaStream_t stream) {
+
+    im2col_cuda<float>(x, dst, IW, IH, OW, OH, KW, KH, IC, batch, batch_offset, offset_delta, s0, s1, p0, p1, d0, d1, stream);
+}
+
+void ggml_cuda_op_im2col(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
+    const ggml_tensor * src0 = dst->src[0];
+    const ggml_tensor * src1 = dst->src[1];
+    const float * src1_d = (const float *)src1->data;
+    float * dst_d = (float *)dst->data;
+    cudaStream_t stream = ctx.stream();
+
+    GGML_ASSERT(src1->type == GGML_TYPE_F32);
+    GGML_ASSERT( dst->type == GGML_TYPE_F16 || dst->type == GGML_TYPE_F32);
+
+    const int32_t s0 = ((const int32_t*)(dst->op_params))[0];
+    const int32_t s1 = ((const int32_t*)(dst->op_params))[1];
+    const int32_t p0 = ((const int32_t*)(dst->op_params))[2];
+    const int32_t p1 = ((const int32_t*)(dst->op_params))[3];
+    const int32_t d0 = ((const int32_t*)(dst->op_params))[4];
+    const int32_t d1 = ((const int32_t*)(dst->op_params))[5];
+
+    const bool is_2D = ((const int32_t*)(dst->op_params))[6] == 1;
+
+    const int64_t IC = src1->ne[is_2D ? 2 : 1];
+    const int64_t IH = is_2D ? src1->ne[1] : 1;
+    const int64_t IW =         src1->ne[0];
+
+    const int64_t KH = is_2D ? src0->ne[1] : 1;
+    const int64_t KW =         src0->ne[0];
+
+    const int64_t OH = is_2D ? dst->ne[2] : 1;
+    const int64_t OW =         dst->ne[1];
+
+    const size_t delta_offset = src1->nb[is_2D ? 2 : 1] / 4; // nb is byte offset, src is type float32
+    const int64_t batch = src1->ne[3];
+    const size_t batch_offset = src1->nb[3] / 4; // nb is byte offset, src is type float32
+
+    if(dst->type == GGML_TYPE_F16) {
+        im2col_cuda_f16(src1_d, (half *) dst_d, IW, IH, OW, OH, KW, KH, IC, batch, batch_offset, delta_offset, s0, s1, p0, p1, d0, d1, stream);
+    } else {
+        im2col_cuda_f32(src1_d, (float *) dst_d, IW, IH, OW, OH, KW, KH, IC, batch, batch_offset, delta_offset, s0, s1, p0, p1, d0, d1, stream);
+    }
+}
diff --git a/src/whisper_cpp/ggml/src/ggml-cuda/im2col.cuh b/src/whisper_cpp/ggml/src/ggml-cuda/im2col.cuh
new file mode 100644
index 00000000..1ce8fae4
--- /dev/null
+++ b/src/whisper_cpp/ggml/src/ggml-cuda/im2col.cuh
@@ -0,0 +1,5 @@
+#include "common.cuh"
+
+#define CUDA_IM2COL_BLOCK_SIZE 256
+
+void ggml_cuda_op_im2col(ggml_backend_cuda_context & ctx, ggml_tensor * dst);
diff --git a/src/whisper_cpp/ggml/src/ggml-cuda/mma.cuh b/src/whisper_cpp/ggml/src/ggml-cuda/mma.cuh
new file mode 100644
index 00000000..a452a3cc
--- /dev/null
+++ b/src/whisper_cpp/ggml/src/ggml-cuda/mma.cuh
@@ -0,0 +1,221 @@
+#include "common.cuh"
+
+struct mma_int_A_I16K4 {
+    static constexpr int I  = 16;
+    static constexpr int K  = 4;
+    static constexpr int ne = 2;
+
+    int x[ne] = {0};
+
+    static __device__ __forceinline__ int get_i(const int l) {
+        const int ret = (l%2) * (I/2) + threadIdx.x / K;
+        GGML_CUDA_ASSUME(ret >= 0);
+        GGML_CUDA_ASSUME(ret <  I);
+        return ret;
+    }
+
+    static __device__ __forceinline__ int get_k(const int /* l */) {
+        const int ret = threadIdx.x % K;
+        GGML_CUDA_ASSUME(ret >= 0);
+        GGML_CUDA_ASSUME(ret <  K);
+        return ret;
+    }
+
+    __device__ __forceinline__ void load(const int * __restrict__ xs0, const int & stride) {
+#if defined(INT8_MMA_AVAILABLE)
+        const int * xs = xs0 + (threadIdx.x%I)*stride;
+        asm("ldmatrix.sync.aligned.m8n8.x2.b16 {%0, %1}, [%2];"
+            : "+r"(x[0]), "+r"(x[1])
+            : "l"(xs));
+#else
+#pragma unroll
+        for (int l = 0; l < ne; ++l) {
+            x[l] = xs0[get_i(l)*stride + get_k(l)];
+        }
+#endif // defined(INT8_MMA_AVAILABLE)
+    }
+};
+
+struct mma_int_A_I16K8 {
+    static constexpr int I  = 16;
+    static constexpr int K  = 8;
+    static constexpr int ne = 4;
+
+    int x[ne] = {0};
+
+    static __device__ __forceinline__ int get_i(const int l) {
+        const int ret = (l%2) * (I/2) + threadIdx.x / (K/2);
+        GGML_CUDA_ASSUME(ret >= 0);
+        GGML_CUDA_ASSUME(ret <  I);
+        return ret;
+    }
+
+    static __device__ __forceinline__ int get_k(const int l) {
+        const int ret = (l/2) * (K/2) + threadIdx.x % (K/2);
+        GGML_CUDA_ASSUME(ret >= 0);
+        GGML_CUDA_ASSUME(ret <  K);
+        return ret;
+    }
+
+    __device__ __forceinline__ void load(const int * __restrict__ xs0, const int & stride) {
+#if defined(INT8_MMA_AVAILABLE)
+        const int * xs = xs0 + (threadIdx.x%I)*stride + (threadIdx.x/I)*(K/2);
+        asm("ldmatrix.sync.aligned.m8n8.x4.b16 {%0, %1, %2, %3}, [%4];"
+            : "+r"(x[0]), "+r"(x[1]), "+r"(x[2]), "+r"(x[3])
+            : "l"(xs));
+#else
+#pragma unroll
+        for (int l = 0; l < ne; ++l) {
+            x[l] = xs0[get_i(l)*stride + get_k(l)];
+        }
+#endif // defined(INT8_MMA_AVAILABLE)
+    }
+
+    __device__ __forceinline__ void load_low(const int * __restrict__ xs0, const int & stride) {
+        ((mma_int_A_I16K4 *) x)[0].load(xs0, stride);
+    }
+};
+
+struct mma_int_B_J8K4 {
+    static constexpr int J  = 8;
+    static constexpr int K  = 4;
+    static constexpr int ne = 1;
+
+    int x[ne] = {0};
+
+    static __device__ __forceinline__ int get_j(const int /* l */) {
+        const int ret = threadIdx.x / K;
+        GGML_CUDA_ASSUME(ret >= 0);
+        GGML_CUDA_ASSUME(ret <  J);
+        return ret;
+    }
+
+    static __device__ __forceinline__ int get_k(const int /* l */) {
+        const int ret = threadIdx.x % K;
+        GGML_CUDA_ASSUME(ret >= 0);
+        GGML_CUDA_ASSUME(ret <  K);
+        return ret;
+    }
+
+    __device__ __forceinline__ void load(const int * __restrict__ xs0, const int & stride) {
+#if defined(INT8_MMA_AVAILABLE) && false // Loading as 4 byte values is faster
+        const int * xs = xs0 + (threadIdx.x%J)*stride;
+        asm("ldmatrix.sync.aligned.m8n8.x1.b16 {%0}, [%1];"
+            : "+r"(x[0])
+            : "l"(xs));
+#else
+#pragma unroll
+        for (int l = 0; l < ne; ++l) {
+            x[l] = xs0[get_j(l)*stride + get_k(l)];
+        }
+#endif // defined(INT8_MMA_AVAILABLE)
+    }
+};
+
+struct mma_int_B_J8K8 {
+    static constexpr int J  = 8;
+    static constexpr int K  = 8;
+    static constexpr int ne = 2;
+
+    int x[ne] = {0};
+
+    static __device__ __forceinline__ int get_j(const int /* l */) {
+        const int ret = threadIdx.x / (K/2);
+        GGML_CUDA_ASSUME(ret >= 0);
+        GGML_CUDA_ASSUME(ret <  J);
+        return ret;
+    }
+
+    static __device__ __forceinline__ int get_k(const int l) {
+        const int ret = l * (K/2) + threadIdx.x % (K/2);
+        GGML_CUDA_ASSUME(ret >= 0);
+        GGML_CUDA_ASSUME(ret <  K);
+        return ret;
+    }
+
+    __device__ __forceinline__ void load(const int * __restrict__ xs0, const int & stride) {
+#if defined(INT8_MMA_AVAILABLE) && false // Loading as 4 byte values is faster
+        const int * xs = xs0 + (threadIdx.x%J)*stride + ((threadIdx.x/J)*(K/2)) % K;
+        asm("ldmatrix.sync.aligned.m8n8.x2.b16 {%0, %1}, [%2];"
+            : "+r"(x[0]), "+r"(x[1])
+            : "l"(xs));
+#else
+#pragma unroll
+        for (int l = 0; l < ne; ++l) {
+            x[l] = xs0[get_j(l)*stride + get_k(l)];
+        }
+#endif // defined(INT8_MMA_AVAILABLE)
+    }
+};
+
+struct mma_int_C_I16J8 {
+    static constexpr int I  = 16;
+    static constexpr int J  = 8;
+    static constexpr int ne = 4;
+
+    int x[ne] = {0};
+
+    static __device__ __forceinline__ int get_i(const int l) {
+        const int ret = (l/2) * (I/2) + threadIdx.x / (J/2);
+        GGML_CUDA_ASSUME(ret >= 0);
+        GGML_CUDA_ASSUME(ret <  I);
+        return ret;
+    }
+
+    static __device__ __forceinline__ int get_j(const int l) {
+        const int ret = 2 * (threadIdx.x % (J/2)) + l%2;
+        GGML_CUDA_ASSUME(ret >= 0);
+        GGML_CUDA_ASSUME(ret <  J);
+        return ret;
+    }
+
+    __device__ __forceinline__ void mma_K4(const mma_int_A_I16K4 & mma_A, const mma_int_B_J8K4 & mma_B) {
+#ifdef INT8_MMA_AVAILABLE
+#if __CUDA_ARCH__ >= CC_AMPERE
+        asm("mma.sync.aligned.m16n8k16.row.col.s32.s8.s8.s32 {%0, %1, %2, %3}, {%4, %5}, {%6}, {%0, %1, %2, %3};"
+            : "+r"(x[0]), "+r"(x[1]), "+r"(x[2]), "+r"(x[3])
+            : "r"(mma_A.x[0]), "r"(mma_A.x[1]), "r"(mma_B.x[0]));
+#else
+        // On Turing m16n8k16 mma is not available, use 2x m8n8k16 mma instead:
+        asm("mma.sync.aligned.m8n8k16.row.col.s32.s8.s8.s32 {%0, %1}, {%2}, {%3}, {%0, %1};"
+            : "+r"(x[0]), "+r"(x[1])
+            : "r"(mma_A.x[0]), "r"(mma_B.x[0]));
+        asm("mma.sync.aligned.m8n8k16.row.col.s32.s8.s8.s32 {%0, %1}, {%2}, {%3}, {%0, %1};"
+            : "+r"(x[2]), "+r"(x[3])
+            : "r"(mma_A.x[1]), "r"(mma_B.x[0]));
+#endif // __CUDA_ARCH__ >= CC_AMPERE
+#else
+        GGML_UNUSED(mma_A);
+        GGML_UNUSED(mma_B);
+        NO_DEVICE_CODE;
+#endif // INT8_MMA_AVAILABLE
+    }
+
+    __device__ __forceinline__ void mma_K8(const mma_int_A_I16K8 & mma_A, const mma_int_B_J8K8 & mma_B) {
+#ifdef INT8_MMA_AVAILABLE
+#if __CUDA_ARCH__ >= CC_AMPERE
+        asm("mma.sync.aligned.m16n8k32.row.col.s32.s8.s8.s32 {%0, %1, %2, %3}, {%4, %5, %6, %7}, {%8, %9}, {%0, %1, %2, %3};"
+            : "+r"(x[0]), "+r"(x[1]), "+r"(x[2]), "+r"(x[3])
+            : "r"(mma_A.x[0]), "r"(mma_A.x[1]), "r"(mma_A.x[2]), "r"(mma_A.x[3]), "r"(mma_B.x[0]), "r"(mma_B.x[1]));
+#else
+        // On Turing m16n8k32 mma is not available, use 4x m8n8k16 mma instead:
+        asm("mma.sync.aligned.m8n8k16.row.col.s32.s8.s8.s32 {%0, %1}, {%2}, {%3}, {%0, %1};"
+            : "+r"(x[0]), "+r"(x[1])
+            : "r"(mma_A.x[0]), "r"(mma_B.x[0]));
+        asm("mma.sync.aligned.m8n8k16.row.col.s32.s8.s8.s32 {%0, %1}, {%2}, {%3}, {%0, %1};"
+            : "+r"(x[2]), "+r"(x[3])
+            : "r"(mma_A.x[1]), "r"(mma_B.x[0]));
+        asm("mma.sync.aligned.m8n8k16.row.col.s32.s8.s8.s32 {%0, %1}, {%2}, {%3}, {%0, %1};"
+            : "+r"(x[0]), "+r"(x[1])
+            : "r"(mma_A.x[2]), "r"(mma_B.x[1]));
+        asm("mma.sync.aligned.m8n8k16.row.col.s32.s8.s8.s32 {%0, %1}, {%2}, {%3}, {%0, %1};"
+            : "+r"(x[2]), "+r"(x[3])
+            : "r"(mma_A.x[3]), "r"(mma_B.x[1]));
+#endif // __CUDA_ARCH__ >= CC_AMPERE
+#else
+        GGML_UNUSED(mma_A);
+        GGML_UNUSED(mma_B);
+        NO_DEVICE_CODE;
+#endif // INT8_MMA_AVAILABLE
+    }
+};
diff --git a/src/whisper_cpp/ggml/src/ggml-cuda/mmq.cu b/src/whisper_cpp/ggml/src/ggml-cuda/mmq.cu
new file mode 100644
index 00000000..4935f881
--- /dev/null
+++ b/src/whisper_cpp/ggml/src/ggml-cuda/mmq.cu
@@ -0,0 +1,154 @@
+#include "mmq.cuh"
+
+void ggml_cuda_op_mul_mat_q(
+    ggml_backend_cuda_context & ctx,
+    const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, const char * src0_dd_i, const float * src1_ddf_i,
+    const char * src1_ddq_i, float * dst_dd_i, const int64_t row_low, const int64_t row_high, const int64_t src1_ncols,
+    const int64_t src1_padded_row_size, cudaStream_t stream) {
+
+    const int64_t ne00 = src0->ne[0];
+
+    const int64_t nb01 = src0->nb[1];
+
+    const int64_t ne10 = src1->ne[0];
+    const int64_t ne11 = src1->ne[1];
+    GGML_ASSERT(ne10 % QK8_1 == 0);
+
+    const int64_t ne0 = dst->ne[0];
+
+    const int64_t row_diff = row_high - row_low;
+    const int64_t stride00 = nb01 / ggml_type_size(src0->type);
+
+    int id = ggml_cuda_get_device();
+    const int compute_capability = ggml_cuda_info().devices[id].cc;
+
+    // the main device has a larger memory buffer to hold the results from all GPUs
+    // nrows_dst == nrows of the matrix that the kernel writes into
+    const int64_t nrows_dst = id == ctx.device ? ne0 : row_diff;
+
+    // The stream-k decomposition is only faster for recent NVIDIA GPUs.
+    // Also its fixup needs to allocate a temporary buffer in the memory pool.
+    // There are multiple parallel CUDA streams for src1_ncols != ne11 which would introduce a race condition for this buffer.
+    const bool use_stream_k = compute_capability >= CC_VOLTA && compute_capability < CC_OFFSET_AMD && src1_ncols == ne11;
+    const mmq_args args = {src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, stride00, src1_padded_row_size, src1_ncols, ne11, nrows_dst, use_stream_k};
+
+    switch (src0->type) {
+        case GGML_TYPE_Q4_0:
+            mul_mat_q_case<GGML_TYPE_Q4_0>(ctx, args, stream);
+            break;
+        case GGML_TYPE_Q4_1:
+            mul_mat_q_case<GGML_TYPE_Q4_1>(ctx, args, stream);
+            break;
+        case GGML_TYPE_Q5_0:
+            mul_mat_q_case<GGML_TYPE_Q5_0>(ctx, args, stream);
+            break;
+        case GGML_TYPE_Q5_1:
+            mul_mat_q_case<GGML_TYPE_Q5_1>(ctx, args, stream);
+            break;
+        case GGML_TYPE_Q8_0:
+            mul_mat_q_case<GGML_TYPE_Q8_0>(ctx, args, stream);
+            break;
+        case GGML_TYPE_Q2_K:
+            mul_mat_q_case<GGML_TYPE_Q2_K>(ctx, args, stream);
+            break;
+        case GGML_TYPE_Q3_K:
+            mul_mat_q_case<GGML_TYPE_Q3_K>(ctx, args, stream);
+            break;
+        case GGML_TYPE_Q4_K:
+            mul_mat_q_case<GGML_TYPE_Q4_K>(ctx, args, stream);
+            break;
+        case GGML_TYPE_Q5_K:
+            mul_mat_q_case<GGML_TYPE_Q5_K>(ctx, args, stream);
+            break;
+        case GGML_TYPE_Q6_K:
+            mul_mat_q_case<GGML_TYPE_Q6_K>(ctx, args, stream);
+            break;
+        case GGML_TYPE_IQ2_XXS:
+            mul_mat_q_case<GGML_TYPE_IQ2_XXS>(ctx, args, stream);
+            break;
+        case GGML_TYPE_IQ2_XS:
+            mul_mat_q_case<GGML_TYPE_IQ2_XS>(ctx, args, stream);
+            break;
+        case GGML_TYPE_IQ2_S:
+            mul_mat_q_case<GGML_TYPE_IQ2_S>(ctx, args, stream);
+            break;
+        case GGML_TYPE_IQ3_XXS:
+            mul_mat_q_case<GGML_TYPE_IQ3_XXS>(ctx, args, stream);
+            break;
+        case GGML_TYPE_IQ3_S:
+            mul_mat_q_case<GGML_TYPE_IQ3_S>(ctx, args, stream);
+            break;
+        case GGML_TYPE_IQ1_S:
+            mul_mat_q_case<GGML_TYPE_IQ1_S>(ctx, args, stream);
+            break;
+        case GGML_TYPE_IQ4_XS:
+            mul_mat_q_case<GGML_TYPE_IQ4_XS>(ctx, args, stream);
+            break;
+        case GGML_TYPE_IQ4_NL:
+            mul_mat_q_case<GGML_TYPE_IQ4_NL>(ctx, args, stream);
+            break;
+        default:
+            GGML_ABORT("fatal error");
+            break;
+    }
+
+    GGML_UNUSED(src1);
+    GGML_UNUSED(dst);
+    GGML_UNUSED(src1_ddf_i);
+}
+
+bool ggml_cuda_should_use_mmq(enum ggml_type type, int cc, int64_t ne11) {
+#ifdef GGML_CUDA_FORCE_CUBLAS
+    return false;
+#endif // GGML_CUDA_FORCE_CUBLAS
+
+    bool mmq_supported;
+
+    switch (type) {
+        case GGML_TYPE_Q4_0:
+        case GGML_TYPE_Q4_1:
+        case GGML_TYPE_Q5_0:
+        case GGML_TYPE_Q5_1:
+        case GGML_TYPE_Q8_0:
+        case GGML_TYPE_Q2_K:
+        case GGML_TYPE_Q3_K:
+        case GGML_TYPE_Q4_K:
+        case GGML_TYPE_Q5_K:
+        case GGML_TYPE_Q6_K:
+        case GGML_TYPE_IQ2_XXS:
+        case GGML_TYPE_IQ2_XS:
+        case GGML_TYPE_IQ2_S:
+        case GGML_TYPE_IQ3_XXS:
+        case GGML_TYPE_IQ3_S:
+        case GGML_TYPE_IQ1_S:
+        case GGML_TYPE_IQ4_XS:
+        case GGML_TYPE_IQ4_NL:
+            mmq_supported = true;
+            break;
+        default:
+            mmq_supported = false;
+            break;
+    }
+
+    if (!mmq_supported) {
+        return false;
+    }
+
+    if (int8_mma_available(cc)) {
+        return true;
+    }
+
+    if (cc < MIN_CC_DP4A) {
+        return false;
+    }
+
+#ifdef GGML_CUDA_FORCE_MMQ
+    return true;
+#endif //GGML_CUDA_FORCE_MMQ
+
+    if (cc < CC_OFFSET_AMD) {
+        return cc < CC_VOLTA || ne11 < MMQ_DP4A_MAX_BATCH_SIZE;
+    }
+
+    return cc < CC_RDNA3 || ne11 < MMQ_DP4A_MAX_BATCH_SIZE;
+}
diff --git a/src/whisper_cpp/ggml/src/ggml-cuda/mmq.cuh b/src/whisper_cpp/ggml/src/ggml-cuda/mmq.cuh
new file mode 100644
index 00000000..021a2568
--- /dev/null
+++ b/src/whisper_cpp/ggml/src/ggml-cuda/mmq.cuh
@@ -0,0 +1,2936 @@
+#pragma once
+
+#include "common.cuh"
+#include "vecdotq.cuh"
+#include "mma.cuh"
+
+#include <climits>
+#include <cstdint>
+
+#define MMQ_DP4A_MAX_BATCH_SIZE 64 // Max. batch size to use for dp4a MMQ kernels when FP16 tensor cores are available.
+#define MMQ_ITER_K 256
+#define MMQ_NWARPS 8
+
+typedef void (*load_tiles_mmq_t)(const char * __restrict__ x, int * x_tile, const int & kbx0, const int & i_max, const int & stride);
+typedef void (*vec_dot_mmq_t)(const int * __restrict__ x, const int * __restrict__ y, float * __restrict__ sum, const int & k00);
+typedef void (*mmq_write_back_t)(const float * __restrict__ sum, float * __restrict__ dst, const int & stride, const int & i_max, const int & j_max);
+
+enum mmq_q8_1_ds_layout {
+    MMQ_Q8_1_DS_LAYOUT_D4,
+    MMQ_Q8_1_DS_LAYOUT_DS4,
+    MMQ_Q8_1_DS_LAYOUT_D2S6,
+};
+
+struct block_q8_1_mmq {
+    // The y float data is converted to a data layout that can simply be copied to shared memory as a contiguous block.
+    // The y float data is first grouped as blocks of 128 values.
+    // These blocks are then treated as individual data values and transposed.
+    //
+    // To avoid shared memory bank conflicts each block is padded with 16 bytes.
+    // This padding is also used to store block scales/partial sums.
+    // The scales multiplied with the quantized data are equal to the unquantized values.
+    // The partial sums are obtained by summing up a subgroup of the contained values (prior to quantization)
+    //     and are only needed for performance reasons.
+    //
+    // The exact data stored depends on the x data type.
+    union {
+        float d4[4];    // 1 32 bit scale per 32 values, stored as d0,d1,d2,d3
+        half2 ds4[4];   // 1 16 bit scale + 1 16 bit partial sum per 32 values, stored as d0,s0,d1,s1,d2,s2,d3,s3
+        half  d2s6[8];  // 1 16 bit scale per 64 values + 1 16 bit partial sum per 16 values for the first 96 values,
+                        //     stored as d0,d1,s1,s2,s3,s4,s5
+    };
+    int8_t qs[4*QK8_1]; // 128 values quantized to 8 bit each
+};
+static_assert(sizeof(block_q8_1_mmq) == 4*QK8_1 + 4*sizeof(half2), "Unexpected block_q8_1_mmq size");
+static_assert(sizeof(block_q8_1_mmq) == 4*sizeof(block_q8_1),      "Unexpected block_q8_1_mmq size");
+
+static mmq_q8_1_ds_layout mmq_get_q8_1_ds_layout(const ggml_type type_x) {
+    switch (type_x) {
+        case GGML_TYPE_Q4_0:
+        case GGML_TYPE_Q4_1:
+            return MMQ_Q8_1_DS_LAYOUT_DS4;
+        case GGML_TYPE_Q5_0:
+            return MMQ_Q8_1_DS_LAYOUT_D4;
+        case GGML_TYPE_Q5_1:
+            return MMQ_Q8_1_DS_LAYOUT_DS4;
+        case GGML_TYPE_Q8_0:
+            return MMQ_Q8_1_DS_LAYOUT_D4;
+        case GGML_TYPE_Q2_K:
+            return MMQ_Q8_1_DS_LAYOUT_D2S6;
+        case GGML_TYPE_Q3_K:
+            return MMQ_Q8_1_DS_LAYOUT_D4;
+        case GGML_TYPE_Q4_K:
+        case GGML_TYPE_Q5_K:
+            return MMQ_Q8_1_DS_LAYOUT_DS4;
+        case GGML_TYPE_Q6_K:
+        case GGML_TYPE_IQ2_XXS:
+        case GGML_TYPE_IQ2_XS:
+        case GGML_TYPE_IQ2_S:
+        case GGML_TYPE_IQ3_XXS:
+        case GGML_TYPE_IQ3_S:
+            return MMQ_Q8_1_DS_LAYOUT_D4;
+        case GGML_TYPE_IQ1_S:
+            return MMQ_Q8_1_DS_LAYOUT_DS4;
+        case GGML_TYPE_IQ4_XS:
+        case GGML_TYPE_IQ4_NL:
+            return MMQ_Q8_1_DS_LAYOUT_D4;
+        default:
+            GGML_ABORT("fatal error");
+            break;
+    }
+}
+
+struct tile_x_sizes {
+    int qs;
+    int dm;
+    int sc;
+};
+
+static constexpr int get_mmq_x_max_host(const int cc) {
+    return int8_mma_available(cc) ? 128 :
+#ifdef GGML_CUDA_FORCE_MMQ
+        cc >= CC_VOLTA && cc < CC_OFFSET_AMD ? 128                     : 64;
+#else
+        cc >= CC_VOLTA && cc < CC_OFFSET_AMD ? MMQ_DP4A_MAX_BATCH_SIZE : 64;
+#endif // GGML_CUDA_FORCE_MMQ
+}
+
+static constexpr __device__ int get_mmq_x_max_device() {
+#ifdef INT8_MMA_AVAILABLE
+    return 128;
+#else // INT8_MMA_AVAILABLE
+
+#if defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)
+    return 128;
+#else // defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)
+
+#if __CUDA_ARCH__ >= CC_VOLTA
+#ifdef GGML_CUDA_FORCE_MMQ
+    return MMQ_DP4A_MAX_BATCH_SIZE;
+#else // GGML_CUDA_FORCE_MMQ
+    return 128;
+#endif // GGML_CUDA_FORCE_MMQ
+#else // __CUDA_ARCH__ >= CC_VOLTA
+
+    return 64;
+#endif // __CUDA_ARCH__ >= CC_VOLTA
+
+#endif // defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)
+#endif // INT8_MMA_AVAILABLE
+}
+
+static constexpr int get_mmq_y_host(const int cc) {
+    return cc >= CC_OFFSET_AMD ? (cc == CC_RDNA1 ? 64 : 128) : (cc >= CC_VOLTA ? 128 : 64);
+}
+
+static constexpr __device__ int get_mmq_y_device() {
+#if defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)
+#if defined(RDNA1)
+    return 64;
+#else
+    return 128;
+#endif // defined RDNA1
+#else
+#if __CUDA_ARCH__ >= CC_VOLTA
+    return 128;
+#else
+    return 64;
+#endif // __CUDA_ARCH__ >= CC_VOLTA
+#endif // defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)
+}
+
+#define MMQ_DP4A_TXS_Q4_0    tile_x_sizes{mmq_y*WARP_SIZE   + mmq_y, mmq_y*WARP_SIZE/QI4_0   + mmq_y/QI4_0,     0}
+#define MMQ_DP4A_TXS_Q4_1    tile_x_sizes{mmq_y*WARP_SIZE   + mmq_y, mmq_y*WARP_SIZE/QI4_1   + mmq_y/QI4_1,     0}
+#define MMQ_DP4A_TXS_Q8_0    tile_x_sizes{mmq_y*WARP_SIZE*2 + mmq_y, mmq_y*WARP_SIZE*2/QI8_0 + mmq_y/(QI8_0/2), 0}
+#define MMQ_DP4A_TXS_Q8_0_16 tile_x_sizes{mmq_y*WARP_SIZE*2 + mmq_y, mmq_y*WARP_SIZE*4/QI8_0 + mmq_y/(QI8_0/4), 0}
+#define MMQ_DP4A_TXS_Q8_1    tile_x_sizes{mmq_y*WARP_SIZE*2 + mmq_y, mmq_y*WARP_SIZE*2/QI8_1 + mmq_y/(QI8_1/2), 0}
+#define MMQ_DP4A_TXS_Q2_K    tile_x_sizes{mmq_y*WARP_SIZE*2 + mmq_y, mmq_y*WARP_SIZE         + mmq_y,           0}
+#define MMQ_DP4A_TXS_Q3_K    tile_x_sizes{mmq_y*WARP_SIZE*2 + mmq_y, mmq_y,                                     mmq_y*WARP_SIZE/8 + mmq_y/8}
+#define MMQ_DP4A_TXS_Q4_K    tile_x_sizes{mmq_y*WARP_SIZE   + mmq_y, mmq_y*WARP_SIZE/QI4_K,                     mmq_y*WARP_SIZE/8 + mmq_y/8}
+#define MMQ_DP4A_TXS_Q5_K    tile_x_sizes{mmq_y*WARP_SIZE*2 + mmq_y, mmq_y*WARP_SIZE/QI5_K   + mmq_y/QI5_K,     mmq_y*WARP_SIZE/8 + mmq_y/8}
+#define MMQ_DP4A_TXS_Q6_K    tile_x_sizes{mmq_y*WARP_SIZE*2 + mmq_y, mmq_y*WARP_SIZE/QI6_K   + mmq_y/QI6_K,     mmq_y*WARP_SIZE/8 + mmq_y/8}
+
+static constexpr __host__ __device__ tile_x_sizes mmq_get_dp4a_tile_x_sizes(ggml_type type, int mmq_y) {
+    return type == GGML_TYPE_Q4_0 ? MMQ_DP4A_TXS_Q4_0 :
+        type == GGML_TYPE_Q4_1    ? MMQ_DP4A_TXS_Q4_1 :
+        type == GGML_TYPE_Q5_0    ? MMQ_DP4A_TXS_Q8_0 :
+        type == GGML_TYPE_Q5_1    ? MMQ_DP4A_TXS_Q8_1 :
+        type == GGML_TYPE_Q8_0    ? MMQ_DP4A_TXS_Q8_0 :
+        type == GGML_TYPE_Q2_K    ? MMQ_DP4A_TXS_Q2_K :
+        type == GGML_TYPE_Q3_K    ? MMQ_DP4A_TXS_Q3_K :
+        type == GGML_TYPE_Q4_K    ? MMQ_DP4A_TXS_Q4_K :
+        type == GGML_TYPE_Q5_K    ? MMQ_DP4A_TXS_Q5_K :
+        type == GGML_TYPE_Q6_K    ? MMQ_DP4A_TXS_Q6_K :
+        type == GGML_TYPE_IQ2_XXS ? MMQ_DP4A_TXS_Q8_0 :
+        type == GGML_TYPE_IQ2_XS  ? MMQ_DP4A_TXS_Q8_0_16 :
+        type == GGML_TYPE_IQ2_S   ? MMQ_DP4A_TXS_Q8_0_16 :
+        type == GGML_TYPE_IQ3_XXS ? MMQ_DP4A_TXS_Q8_0 :
+        type == GGML_TYPE_IQ3_S   ? MMQ_DP4A_TXS_Q8_0 :
+        type == GGML_TYPE_IQ1_S   ? MMQ_DP4A_TXS_Q8_0 :
+        type == GGML_TYPE_IQ4_XS  ? MMQ_DP4A_TXS_Q8_0 :
+        type == GGML_TYPE_IQ4_NL  ? MMQ_DP4A_TXS_Q8_0 :
+        tile_x_sizes{0, 0, 0};
+}
+
+#define MMQ_MMA_TILE_X_K_Q8_0 (2*WARP_SIZE + 2*WARP_SIZE/QI8_0                 + 4)
+#define MMQ_MMA_TILE_X_K_Q8_1 (2*WARP_SIZE + 2*WARP_SIZE/QI8_0                 + 4)
+#define MMQ_MMA_TILE_X_K_Q2_K (2*WARP_SIZE + WARP_SIZE                         + 4)
+#define MMQ_MMA_TILE_X_K_Q3_K (2*WARP_SIZE + WARP_SIZE/2                       + 4)
+#define MMQ_MMA_TILE_X_K_Q6_K (2*WARP_SIZE + WARP_SIZE/QI6_K     + WARP_SIZE/8 + 7)
+
+static_assert(MMQ_MMA_TILE_X_K_Q8_0 % 8 == 4, "Wrong padding.");
+static_assert(MMQ_MMA_TILE_X_K_Q8_1 % 8 == 4, "Wrong padding.");
+static_assert(MMQ_MMA_TILE_X_K_Q2_K % 8 == 4, "Wrong padding.");
+static_assert(MMQ_MMA_TILE_X_K_Q3_K % 8 == 4, "Wrong padding.");
+static_assert(MMQ_MMA_TILE_X_K_Q6_K % 8 == 4, "Wrong padding.");
+
+static constexpr __host__ __device__ int mmq_get_mma_tile_x_k(ggml_type type) {
+    return type == GGML_TYPE_Q4_0 ? MMQ_MMA_TILE_X_K_Q8_0 :
+        type == GGML_TYPE_Q4_1    ? MMQ_MMA_TILE_X_K_Q8_1 :
+        type == GGML_TYPE_Q5_0    ? MMQ_MMA_TILE_X_K_Q8_0 :
+        type == GGML_TYPE_Q5_1    ? MMQ_MMA_TILE_X_K_Q8_1 :
+        type == GGML_TYPE_Q8_0    ? MMQ_MMA_TILE_X_K_Q8_0 :
+        type == GGML_TYPE_Q2_K    ? MMQ_MMA_TILE_X_K_Q2_K :
+        type == GGML_TYPE_Q3_K    ? MMQ_MMA_TILE_X_K_Q3_K :
+        type == GGML_TYPE_Q4_K    ? MMQ_MMA_TILE_X_K_Q8_1 :
+        type == GGML_TYPE_Q5_K    ? MMQ_MMA_TILE_X_K_Q8_1 :
+        type == GGML_TYPE_Q6_K    ? MMQ_MMA_TILE_X_K_Q6_K :
+        type == GGML_TYPE_IQ2_XXS ? MMQ_MMA_TILE_X_K_Q8_0 :
+        type == GGML_TYPE_IQ2_XS  ? MMQ_MMA_TILE_X_K_Q3_K :
+        type == GGML_TYPE_IQ2_S   ? MMQ_MMA_TILE_X_K_Q3_K :
+        type == GGML_TYPE_IQ3_XXS ? MMQ_MMA_TILE_X_K_Q8_0 :
+        type == GGML_TYPE_IQ3_S   ? MMQ_MMA_TILE_X_K_Q8_0 :
+        type == GGML_TYPE_IQ1_S   ? MMQ_MMA_TILE_X_K_Q8_0 :
+        type == GGML_TYPE_IQ4_XS  ? MMQ_MMA_TILE_X_K_Q8_0 :
+        type == GGML_TYPE_IQ4_NL  ? MMQ_MMA_TILE_X_K_Q8_0 :
+        0;
+}
+
+#define MMQ_TILE_Y_K (WARP_SIZE + WARP_SIZE/QI8_1)
+
+static int mmq_get_granularity_host(const int mmq_x, const int cc) {
+    return int8_mma_available(cc) && mmq_x >= 48 ? 16 : 8;
+}
+
+#ifdef INT8_MMA_AVAILABLE
+static constexpr __device__ int mmq_get_granularity_device(const int mmq_x) {
+    return mmq_x >= 48 ? 16 : 8;
+}
+#else
+static constexpr __device__ int mmq_get_granularity_device(const int /* mmq_x */) {
+    return 8;
+}
+#endif // INT8_MMA_AVAILABLE
+
+// ------------------------------------------------------------
+
+template <int mmq_y, int nwarps, bool need_check> static __device__ __forceinline__ void load_tiles_q4_0(
+    const char * __restrict__ x, int * __restrict__ x_tile, const int & kbx0, const int & i_max, const int & stride) {
+
+#ifdef INT8_MMA_AVAILABLE
+    int   * x_qs = (int   *)  x_tile;
+    float * x_df = (float *) (x_qs + 2*WARP_SIZE);
+#else
+    constexpr tile_x_sizes txs = mmq_get_dp4a_tile_x_sizes(GGML_TYPE_Q4_0, mmq_y);
+    int   * x_qs = (int   *)  x_tile;
+    float * x_df = (float *) (x_qs + txs.qs);
+#endif // INT8_MMA_AVAILABLE
+
+    const int kbx  = threadIdx.x / QI4_0;
+    const int kqsx = threadIdx.x % QI4_0;
+
+#pragma unroll
+    for (int i0 = 0; i0 < mmq_y; i0 += nwarps) {
+        int i = i0 + threadIdx.y;
+
+        if (need_check) {
+            i = min(i, i_max);
+        }
+
+        const block_q4_0 * bxi = (const block_q4_0 *) x + kbx0 + i*stride + kbx;
+        const int qs0 = get_int_b2(bxi->qs, kqsx);
+
+#ifdef INT8_MMA_AVAILABLE
+        x_qs[i*MMQ_MMA_TILE_X_K_Q8_0 + kbx*(2*QI4_0) + kqsx + 0]     = __vsubss4((qs0 >> 0) & 0x0F0F0F0F, 0x08080808);
+        x_qs[i*MMQ_MMA_TILE_X_K_Q8_0 + kbx*(2*QI4_0) + kqsx + QI4_0] = __vsubss4((qs0 >> 4) & 0x0F0F0F0F, 0x08080808);
+#else
+        x_qs[i*(WARP_SIZE + 1) + threadIdx.x] = qs0;
+#endif // INT8_MMA_AVAILABLE
+    }
+
+    const int blocks_per_tile_x_row = WARP_SIZE / QI4_0;
+    const int kbxd = threadIdx.x % blocks_per_tile_x_row;
+
+#pragma unroll
+    for (int i0 = 0; i0 < mmq_y; i0 += nwarps * QI4_0) {
+        int i = i0 + threadIdx.y * QI4_0 + threadIdx.x / blocks_per_tile_x_row;
+
+        if (need_check) {
+            i = min(i, i_max);
+        }
+
+        const block_q4_0 * bxi = (const block_q4_0 *) x + kbx0 + i*stride + kbxd;
+
+#ifdef INT8_MMA_AVAILABLE
+        x_df[i*MMQ_MMA_TILE_X_K_Q8_0       + kbxd] = bxi->d;
+#else
+        x_df[i*(WARP_SIZE/QI4_0) + i/QI4_0 + kbxd] = bxi->d;
+#endif // INT8_MMA_AVAILABLE
+    }
+}
+
+template <int mmq_x, int mmq_y, int nwarps>
+static __device__ __forceinline__ void vec_dot_q4_0_q8_1_dp4a(
+    const int * __restrict__ x, const int * __restrict__ y, float * __restrict__ sum, const int & k00) {
+
+    constexpr tile_x_sizes txs = mmq_get_dp4a_tile_x_sizes(GGML_TYPE_Q4_0, mmq_y);
+    const int   * x_qs = (const int   *) x;
+    const float * x_df = (const float *) x_qs + txs.qs;
+    const int   * y_qs = (const int   *) y + 4;
+    const half2 * y_ds = (const half2 *) y;
+
+// #pragma unroll
+    for (int k01 = 0; k01 < WARP_SIZE; k01 += QR4_0*VDR_Q4_0_Q8_1_MMQ) {
+        const int k0 = k00 + k01;
+
+#pragma unroll
+        for (int j0 = 0; j0 < mmq_x; j0 += nwarps) {
+            const int j = j0 + threadIdx.y;
+
+#pragma unroll
+            for (int i0 = 0; i0 < mmq_y; i0 += WARP_SIZE) {
+                const int i = i0 + threadIdx.x;
+
+                const int kyqs = QI8_1 * ((k01/2) / (QI8_1/2)) + (k01/2) % (QI8_1/2);
+
+                int u[2*VDR_Q4_0_Q8_1_MMQ];
+
+#pragma unroll
+                for (int l = 0; l < VDR_Q4_0_Q8_1_MMQ; ++l) {
+                    u[2*l+0] = y_qs[j*MMQ_TILE_Y_K + kyqs +  l];
+                    u[2*l+1] = y_qs[j*MMQ_TILE_Y_K + kyqs + (l + QI4_0)];
+                }
+
+                sum[j0/nwarps*mmq_y/WARP_SIZE + i0/WARP_SIZE] += vec_dot_q4_0_q8_1_impl<VDR_Q4_0_Q8_1_MMQ>
+                    (&x_qs[i*(WARP_SIZE + 1) + k0/QR4_0], u,
+                     x_df[i*(WARP_SIZE/QI4_0) + i/QI4_0 + k0/(QR4_0*QI4_0)], y_ds[j*MMQ_TILE_Y_K + k01/QI8_1]);
+            }
+        }
+    }
+}
+
+template <int mmq_y, int nwarps, bool need_check> static __device__ __forceinline__ void load_tiles_q4_1(
+    const char * __restrict__ x, int * __restrict__ x_tile, const int & kbx0, const int & i_max, const int & stride) {
+
+#ifdef INT8_MMA_AVAILABLE
+    int   * x_qs = (int   *)  x_tile;
+    half2 * x_dm = (half2 *) (x_qs + 2*WARP_SIZE);
+#else
+    constexpr tile_x_sizes txs = mmq_get_dp4a_tile_x_sizes(GGML_TYPE_Q4_1, mmq_y);
+    int   * x_qs = (int   *)  x_tile;
+    half2 * x_dm = (half2 *) (x_qs + txs.qs);
+#endif // INT8_MMA_AVAILABLE
+
+    const int kbx  = threadIdx.x / QI4_1;
+    const int kqsx = threadIdx.x % QI4_1;
+
+#pragma unroll
+    for (int i0 = 0; i0 < mmq_y; i0 += nwarps) {
+        int i = i0 + threadIdx.y;
+
+        if (need_check) {
+            i = min(i, i_max);
+        }
+
+        const block_q4_1 * bxi = (const block_q4_1 *) x + kbx0 + i*stride + kbx;
+        const int qs0 = get_int_b4(bxi->qs, kqsx);
+
+#ifdef INT8_MMA_AVAILABLE
+        x_qs[i*MMQ_MMA_TILE_X_K_Q8_1 + kbx*(2*QI4_1) + kqsx + 0]     = (qs0 >> 0) & 0x0F0F0F0F;
+        x_qs[i*MMQ_MMA_TILE_X_K_Q8_1 + kbx*(2*QI4_1) + kqsx + QI4_1] = (qs0 >> 4) & 0x0F0F0F0F;
+#else
+        x_qs[i*(WARP_SIZE + 1) + threadIdx.x] = qs0;
+#endif // INT8_MMA_AVAILABLE
+    }
+
+    const int blocks_per_tile_x_row = WARP_SIZE / QI4_1;
+    const int kbxd = threadIdx.x % blocks_per_tile_x_row;
+
+#pragma unroll
+    for (int i0 = 0; i0 < mmq_y; i0 += nwarps * QI4_1) {
+        int i = i0 + threadIdx.y * QI4_1 + threadIdx.x / blocks_per_tile_x_row;
+
+        if (need_check) {
+            i = min(i, i_max);
+        }
+
+        const block_q4_1 * bxi = (const block_q4_1 *) x + kbx0 + i*stride + kbxd;
+
+#ifdef INT8_MMA_AVAILABLE
+        x_dm[i*MMQ_MMA_TILE_X_K_Q8_1       + kbxd] = bxi->dm;
+#else
+        x_dm[i*(WARP_SIZE/QI4_1) + i/QI4_1 + kbxd] = bxi->dm;
+#endif // INT8_MMA_AVAILABLE
+    }
+}
+
+template <int mmq_x, int mmq_y, int nwarps>
+static __device__ __forceinline__ void vec_dot_q4_1_q8_1_dp4a(
+    const int * __restrict__ x, const int * __restrict__ y, float * __restrict__ sum, const int & k00) {
+
+    constexpr tile_x_sizes txs = mmq_get_dp4a_tile_x_sizes(GGML_TYPE_Q4_1, mmq_y);
+    const int   * x_qs = (const int   *) x;
+    const half2 * x_dm = (const half2 *) x_qs + txs.qs;
+    const int   * y_qs = (const int   *) y + 4;
+    const half2 * y_ds = (const half2 *) y;
+
+// #pragma unroll
+    for (int k01 = 0; k01 < WARP_SIZE; k01 += QR4_1*VDR_Q4_1_Q8_1_MMQ) {
+        const int k0 = k00 + k01;
+
+#pragma unroll
+        for (int j0 = 0; j0 < mmq_x; j0 += nwarps) {
+            const int j = j0 + threadIdx.y;
+
+#pragma unroll
+            for (int i0 = 0; i0 < mmq_y; i0 += WARP_SIZE) {
+                const int i = i0 + threadIdx.x;
+
+                const int kyqs = QI8_1 * ((k01/2) / (QI8_1/2)) + (k01/2) % (QI8_1/2);
+
+                int u[2*VDR_Q4_1_Q8_1_MMQ];
+
+#pragma unroll
+                for (int l = 0; l < VDR_Q4_1_Q8_1_MMQ; ++l) {
+                    u[2*l+0] = y_qs[j*MMQ_TILE_Y_K + kyqs +  l];
+                    u[2*l+1] = y_qs[j*MMQ_TILE_Y_K + kyqs + (l + QI4_1)];
+                }
+
+                sum[j0/nwarps*mmq_y/WARP_SIZE + i0/WARP_SIZE] += vec_dot_q4_1_q8_1_impl<VDR_Q4_1_Q8_1_MMQ>
+                    (&x_qs[i*(WARP_SIZE + 1) + k0/QR4_1], u,
+                     x_dm[i*(WARP_SIZE/QI4_1) + i/QI4_1 + k0/(QR4_1*QI4_1)], y_ds[j*MMQ_TILE_Y_K + k01/QI8_1]);
+            }
+        }
+    }
+}
+
+template <int mmq_y, int nwarps, bool need_check> static __device__ __forceinline__ void load_tiles_q5_0(
+    const char * __restrict__ x, int * __restrict__ x_tile, const int & kbx0, const int & i_max, const int & stride) {
+
+#ifdef INT8_MMA_AVAILABLE
+    int   * x_qs = (int   *)  x_tile;
+    float * x_df = (float *) (x_qs + WARP_SIZE*2);
+#else
+    constexpr tile_x_sizes txs = mmq_get_dp4a_tile_x_sizes(GGML_TYPE_Q5_0, mmq_y);
+    int   * x_qs = (int   *)  x_tile;
+    float * x_df = (float *) (x_qs + txs.qs);
+#endif // INT8_MMA_AVAILABLE
+
+    const int kbx  = threadIdx.x / QI5_0;
+    const int kqsx = threadIdx.x % QI5_0;
+
+#pragma unroll
+    for (int i0 = 0; i0 < mmq_y; i0 += nwarps) {
+        int i = i0 + threadIdx.y;
+
+        if (need_check) {
+            i = min(i, i_max);
+        }
+
+        const block_q5_0 * bxi = (const block_q5_0 *) x + kbx0 + i*stride + kbx;
+
+        const int ql = get_int_b2(bxi->qs, kqsx);
+        const int qh = get_int_b2(bxi->qh, 0) >> (4 * (threadIdx.x % QI5_0));
+
+        int qs0 = (ql >>  0)   & 0x0F0F0F0F;
+        qs0    |= (qh <<  4)   & 0x00000010;  // 0 ->  4
+        qs0    |= (qh << 11)   & 0x00001000;  // 1 -> 12
+        qs0    |= (qh << 18)   & 0x00100000;  // 2 -> 20
+        qs0    |= (qh << 25)   & 0x10000000;  // 3 -> 28
+        qs0     = __vsubss4(qs0, 0x10101010); // subtract 16
+
+        int qs1 = (ql >>  4)   & 0x0F0F0F0F;
+        qs1    |= (qh >> 12)   & 0x00000010;  // 16 ->  4
+        qs1    |= (qh >>  5)   & 0x00001000;  // 17 -> 12
+        qs1    |= (qh <<  2)   & 0x00100000;  // 18 -> 20
+        qs1    |= (qh <<  9)   & 0x10000000;  // 19 -> 28
+        qs1     = __vsubss4(qs1, 0x10101010); // subtract 16
+
+#ifdef INT8_MMA_AVAILABLE
+        x_qs[i*MMQ_MMA_TILE_X_K_Q8_0 + kbx*(2*QI5_0) + kqsx + 0]     = qs0;
+        x_qs[i*MMQ_MMA_TILE_X_K_Q8_0 + kbx*(2*QI5_0) + kqsx + QI5_0] = qs1;
+#else
+        x_qs[i*(2*WARP_SIZE + 1)     + kbx*(2*QI5_0) + kqsx + 0]     = qs0;
+        x_qs[i*(2*WARP_SIZE + 1)     + kbx*(2*QI5_0) + kqsx + QI5_0] = qs1;
+#endif // INT8_MMA_AVAILABLE
+    }
+
+    const int blocks_per_tile_x_row = WARP_SIZE / QI5_0;
+    const int kbxd = threadIdx.x % blocks_per_tile_x_row;
+
+#pragma unroll
+    for (int i0 = 0; i0 < mmq_y; i0 += nwarps * QI5_0) {
+        int i = i0 + threadIdx.y * QI5_0 + threadIdx.x / blocks_per_tile_x_row;
+
+        if (need_check) {
+            i = min(i, i_max);
+        }
+
+        const block_q5_0 * bxi = (const block_q5_0 *) x + kbx0 + i*stride + kbxd;
+
+#ifdef INT8_MMA_AVAILABLE
+        x_df[i*MMQ_MMA_TILE_X_K_Q8_0       + kbxd] = bxi->d;
+#else
+        x_df[i*(WARP_SIZE/QI5_0) + i/QI5_0 + kbxd] = bxi->d;
+#endif // INT8_MMA_AVAILABLE
+    }
+}
+
+template <int mmq_y, int nwarps, bool need_check> static __device__ __forceinline__ void load_tiles_q5_1(
+    const char * __restrict__ x, int * __restrict__ x_tile, const int & kbx0, const int & i_max, const int & stride) {
+
+#ifdef INT8_MMA_AVAILABLE
+    int   * x_qs = (int   *)  x_tile;
+    half2 * x_dm = (half2 *) (x_qs + 2*WARP_SIZE);
+#else
+    constexpr tile_x_sizes txs = mmq_get_dp4a_tile_x_sizes(GGML_TYPE_Q5_1, mmq_y);
+    int   * x_qs = (int   *)  x_tile;
+    half2 * x_dm = (half2 *) (x_qs + txs.qs);
+#endif // INT8_MMA_AVAILABLE
+
+    const int kbx  = threadIdx.x / QI5_1;
+    const int kqsx = threadIdx.x % QI5_1;
+
+#pragma unroll
+    for (int i0 = 0; i0 < mmq_y; i0 += nwarps) {
+        int i = i0 + threadIdx.y;
+
+        if (need_check) {
+            i = min(i, i_max);
+        }
+
+        const block_q5_1 * bxi = (const block_q5_1 *) x + kbx0 + i*stride + kbx;
+
+        const int ql = get_int_b4(bxi->qs, kqsx);
+        const int qh = get_int_b4(bxi->qh, 0) >> (4 * (threadIdx.x % QI5_1));
+
+        int qs0 = (ql >>  0) & 0x0F0F0F0F;
+        qs0    |= (qh <<  4) & 0x00000010; // 0 ->  4
+        qs0    |= (qh << 11) & 0x00001000; // 1 -> 12
+        qs0    |= (qh << 18) & 0x00100000; // 2 -> 20
+        qs0    |= (qh << 25) & 0x10000000; // 3 -> 28
+
+        int qs1 = (ql >>  4) & 0x0F0F0F0F;
+        qs1    |= (qh >> 12) & 0x00000010; // 16 ->  4
+        qs1    |= (qh >>  5) & 0x00001000; // 17 -> 12
+        qs1    |= (qh <<  2) & 0x00100000; // 18 -> 20
+        qs1    |= (qh <<  9) & 0x10000000; // 19 -> 28
+
+#ifdef INT8_MMA_AVAILABLE
+        x_qs[i*MMQ_MMA_TILE_X_K_Q8_1 + kbx*(2*QI5_1) + kqsx + 0]     = qs0;
+        x_qs[i*MMQ_MMA_TILE_X_K_Q8_1 + kbx*(2*QI5_1) + kqsx + QI5_1] = qs1;
+#else
+        x_qs[i*(2*WARP_SIZE + 1)     + kbx*(2*QI5_1) + kqsx + 0]     = qs0;
+        x_qs[i*(2*WARP_SIZE + 1)     + kbx*(2*QI5_1) + kqsx + QI5_1] = qs1;
+#endif // INT8_MMA_AVAILABLE
+    }
+
+    const int blocks_per_tile_x_row = WARP_SIZE / QI5_1;
+    const int kbxd = threadIdx.x % blocks_per_tile_x_row;
+
+#pragma unroll
+    for (int i0 = 0; i0 < mmq_y; i0 += nwarps * QI5_1) {
+        int i = i0 + threadIdx.y * QI5_1 + threadIdx.x / blocks_per_tile_x_row;
+
+        if (need_check) {
+            i = min(i, i_max);
+        }
+
+        const block_q5_1 * bxi = (const block_q5_1 *) x + kbx0 + i*stride + kbxd;
+
+#ifdef INT8_MMA_AVAILABLE
+        x_dm[i*MMQ_MMA_TILE_X_K_Q8_1       + kbxd] = bxi->dm;
+#else
+        x_dm[i*(WARP_SIZE/QI5_1) + i/QI5_1 + kbxd] = bxi->dm;
+#endif // INT8_MMA_AVAILABLE
+    }
+}
+
+template <int mmq_y, int nwarps, bool need_check> static __device__ __forceinline__ void load_tiles_q8_0(
+    const char * __restrict__ x, int * __restrict__ x_tile, const int & kbx0, const int & i_max, const int & stride) {
+
+#ifdef INT8_MMA_AVAILABLE
+    int   * x_qs = (int   *)  x_tile;
+    float * x_df = (float *) (x_tile + 2*WARP_SIZE);
+#else
+    constexpr tile_x_sizes txs = mmq_get_dp4a_tile_x_sizes(GGML_TYPE_Q8_0, mmq_y);
+    int   * x_qs = (int   *)  x_tile;
+    float * x_df = (float *) (x_qs + txs.qs);
+#endif // INT8_MMA_AVAILABLE
+
+    const int kbx  = threadIdx.x / QI8_0;
+    const int kqsx = threadIdx.x % QI8_0;
+
+#pragma unroll
+    for (int i0 = 0; i0 < mmq_y; i0 += nwarps) {
+        int i = i0 + threadIdx.y;
+
+        if (need_check) {
+            i = min(i, i_max);
+        }
+
+        const block_q8_0 * bxi = (const block_q8_0 *) x + kbx0 + i*stride + kbx;
+
+#ifdef INT8_MMA_AVAILABLE
+        x_qs[i*MMQ_MMA_TILE_X_K_Q8_0 + 0         + threadIdx.x] = get_int_b2(bxi[0].qs,               kqsx);
+        x_qs[i*MMQ_MMA_TILE_X_K_Q8_0 + WARP_SIZE + threadIdx.x] = get_int_b2(bxi[WARP_SIZE/QI8_0].qs, kqsx);
+#else
+        x_qs[i*(2*WARP_SIZE + 1)     + 0         + threadIdx.x] = get_int_b2(bxi[0].qs,               kqsx);
+        x_qs[i*(2*WARP_SIZE + 1)     + WARP_SIZE + threadIdx.x] = get_int_b2(bxi[WARP_SIZE/QI8_0].qs, kqsx);
+#endif // INT8_MMA_AVAILABLE
+    }
+
+    const int blocks_per_tile_x_row = 2*WARP_SIZE / QI8_0;
+    const int kbxd = threadIdx.x % blocks_per_tile_x_row;
+
+#pragma unroll
+    for (int i0 = 0; i0 < mmq_y; i0 += nwarps * QI8_0/2) {
+        int i = i0 + threadIdx.y * (QI8_0/2) + threadIdx.x / blocks_per_tile_x_row;
+
+        if (need_check) {
+            i = min(i, i_max);
+        }
+
+        const block_q8_0 * bxi = (const block_q8_0 *) x + kbx0 + i*stride + kbxd;
+
+#ifdef INT8_MMA_AVAILABLE
+        x_df[i*MMQ_MMA_TILE_X_K_Q8_0             + kbxd] = bxi->d;
+#else
+        x_df[i*(2*WARP_SIZE/QI8_0) + i/(QI8_0/2) + kbxd] = bxi->d;
+#endif // INT8_MMA_AVAILABLE
+    }
+}
+
+template <int mmq_x, int mmq_y, int nwarps>
+static __device__ __forceinline__ void vec_dot_q8_0_q8_1_dp4a(
+    const int * __restrict__ x, const int * __restrict__ y, float * __restrict__ sum, const int & k00) {
+
+    constexpr tile_x_sizes txs = mmq_get_dp4a_tile_x_sizes(GGML_TYPE_Q8_0, mmq_y);
+    const int   * x_qs = (const int   *) x;
+    const float * x_df = (const float *) x_qs + txs.qs;
+    const int   * y_qs = (const int   *) y + 4;
+    const float * y_df = (const float *) y;
+
+// #pragma unroll
+    for (int k01 = 0; k01 < WARP_SIZE; k01 += VDR_Q8_0_Q8_1_MMQ) {
+        const int k0 = k00 + k01;
+
+#pragma unroll
+        for (int j0 = 0; j0 < mmq_x; j0 += nwarps) {
+            const int j = j0 + threadIdx.y;
+
+#pragma unroll
+            for (int i0 = 0; i0 < mmq_y; i0 += WARP_SIZE) {
+                const int i = i0 + threadIdx.x;
+
+                sum[j0/nwarps*mmq_y/WARP_SIZE + i0/WARP_SIZE] += vec_dot_q8_0_q8_1_impl<float, VDR_Q8_0_Q8_1_MMQ>
+                    (&x_qs[i*(2*WARP_SIZE + 1) + k0], &y_qs[j*MMQ_TILE_Y_K + k0 % WARP_SIZE],
+                     x_df[i*(2*WARP_SIZE/QI8_0) + i/(QI8_0/2) + k0/QI8_0], y_df[j*MMQ_TILE_Y_K + (k0/QI8_1) % (WARP_SIZE/QI8_1)]);
+            }
+        }
+    }
+}
+
+template <int mmq_x, int mmq_y, int nwarps, mmq_q8_1_ds_layout ds_layout>
+static __device__ __forceinline__ void vec_dot_q8_0_q8_1_mma(
+    const int * __restrict__ x, const int * __restrict__ y, float * __restrict__ sum, const int & k00) {
+
+    typedef mma_int_A_I16K8 mma_A;
+    typedef mma_int_B_J8K8  mma_B;
+    typedef mma_int_C_I16J8 mma_C;
+
+    constexpr int granularity = mmq_get_granularity_device(mmq_x);
+    constexpr int rows_per_warp = 2 * granularity;
+    constexpr int ntx = rows_per_warp/mma_C::I; // Number of x minitiles per warp.
+
+    y += (threadIdx.y % ntx) * (mma_B::J*MMQ_TILE_Y_K);
+
+    const int   * x_qs = (const int   *) x;
+    const float * x_df = (const float *) x_qs + 2*WARP_SIZE;
+    const int   * y_qs = (const int   *) y + 4;
+    const float * y_df = (const float *) y;
+    const half2 * y_ds = (const half2 *) y;
+
+    mma_A A[ntx][WARP_SIZE/QI8_0];
+    float dA[ntx][mma_C::ne/2][WARP_SIZE/QI8_0];
+
+    const int i0 = (threadIdx.y/ntx)*rows_per_warp;
+
+#pragma unroll
+    for (int n = 0; n < ntx; ++n) {
+#pragma unroll
+        for (int k01 = 0; k01 < WARP_SIZE; k01 += QI8_0) {
+            const int k0 = k00 + k01;
+
+            A[n][k01/QI8_0].load(x_qs + (i0 + n*mma_A::I)*MMQ_MMA_TILE_X_K_Q8_0 + k0, MMQ_MMA_TILE_X_K_Q8_0);
+        }
+
+#pragma unroll
+        for (int l = 0; l < mma_C::ne/2; ++l) {
+            const int i = i0 + n*mma_A::I + mma_C::get_i(2*l);
+
+#pragma unroll
+            for (int k01 = 0; k01 < WARP_SIZE; k01 += QI8_0) {
+                const int k0 = k00 + k01;
+
+                dA[n][l][k01/QI8_0] = x_df[i*MMQ_MMA_TILE_X_K_Q8_0 + k0/QI8_0];
+            }
+        }
+    }
+
+#pragma unroll
+    for (int j0 = 0; j0 < mmq_x; j0 += ntx*mma_C::J) {
+#pragma unroll
+        for (int k01 = 0; k01 < WARP_SIZE; k01 += QI8_0) {
+            mma_B  B;
+            float dB[mma_C::ne/2];
+
+            B.load(y_qs + j0*MMQ_TILE_Y_K + k01, MMQ_TILE_Y_K);
+
+#pragma unroll
+            for (int l = 0; l < mma_C::ne/2; ++l) {
+                const int j = j0 + mma_C::get_j(l);
+
+                if (ds_layout == MMQ_Q8_1_DS_LAYOUT_D4) {
+                    dB[l] =             y_df[j*MMQ_TILE_Y_K + k01/QI8_1];
+                } else {
+                    dB[l] = __low2float(y_ds[j*MMQ_TILE_Y_K + k01/QI8_1]);
+                }
+            }
+
+#pragma unroll
+            for (int n = 0; n < ntx; ++n) {
+                mma_C C;
+                C.mma_K8(A[n][k01/QI8_0], B);
+
+#pragma unroll
+                for (int l = 0; l < mma_C::ne; ++l) {
+                    sum[(j0/mma_C::J + n)*mma_C::ne + l] += C.x[l]*dA[n][l/2][k01/QI8_0]*dB[l%2];
+                }
+            }
+        }
+    }
+}
+
+template <int mmq_x, int mmq_y, int nwarps>
+static __device__ __forceinline__ void vec_dot_q8_1_q8_1_dp4a(
+    const int * __restrict__ x, const int * __restrict__ y, float * __restrict__ sum, const int & k00) {
+
+    constexpr tile_x_sizes txs = mmq_get_dp4a_tile_x_sizes(GGML_TYPE_Q5_1, mmq_y);
+    const int   * x_qs = (const int   *) x;
+    const half2 * x_dm = (const half2 *) x_qs + txs.qs;
+    const int   * y_qs = (const int   *) y + 4;
+    const half2 * y_ds = (const half2 *) y;
+
+// #pragma unroll
+    for (int k01 = 0; k01 < WARP_SIZE; k01 += VDR_Q8_0_Q8_1_MMQ) {
+        const int k0 = k00 + k01;
+
+#pragma unroll
+        for (int j0 = 0; j0 < mmq_x; j0 += nwarps) {
+            const int j = j0 + threadIdx.y;
+
+#pragma unroll
+            for (int i0 = 0; i0 < mmq_y; i0 += WARP_SIZE) {
+                const int i = i0 + threadIdx.x;
+
+                sum[j0/nwarps*mmq_y/WARP_SIZE + i0/WARP_SIZE] += vec_dot_q8_1_q8_1_impl<QR5_1*VDR_Q5_1_Q8_1_MMQ>
+                    (&x_qs[i*(2*WARP_SIZE + 1) + k0], &y_qs[j*MMQ_TILE_Y_K + k01],
+                    x_dm[i*(WARP_SIZE/QI5_1) + i/QI5_1 + k0/QI8_1], y_ds[j*MMQ_TILE_Y_K + k01/QI8_1]);
+            }
+        }
+    }
+}
+
+template <int mmq_x, int mmq_y, int nwarps>
+static __device__ __forceinline__ void vec_dot_q8_1_q8_1_mma(
+    const int * __restrict__ x, const int * __restrict__ y, float * __restrict__ sum, const int & k00) {
+
+    typedef mma_int_A_I16K8 mma_A;
+    typedef mma_int_B_J8K8  mma_B;
+    typedef mma_int_C_I16J8 mma_C;
+
+    constexpr int granularity = mmq_get_granularity_device(mmq_x);
+    constexpr int rows_per_warp = 2 * granularity;
+    constexpr int ntx = rows_per_warp/mma_C::I; // Number of x minitiles per warp.
+
+    y += (threadIdx.y % ntx) * (mma_B::J*MMQ_TILE_Y_K);
+
+    const int   * x_qs = (const int   *) x;
+    const half2 * x_dm = (const half2 *) x_qs + 2*WARP_SIZE;
+    const int   * y_qs = (const int   *) y + 4;
+    const half2 * y_dm = (const half2 *) y;
+
+    mma_A    A[ntx][WARP_SIZE/QI8_1];
+    float2 dmA[ntx][mma_C::ne/2][WARP_SIZE/QI8_1];
+
+    const int i0 = (threadIdx.y/ntx)*rows_per_warp;
+
+#pragma unroll
+    for (int n = 0; n < ntx; ++n) {
+#pragma unroll
+        for (int k01 = 0; k01 < WARP_SIZE; k01 += QI8_1) {
+            const int k0 = k00 + k01;
+
+            A[n][k01/QI8_1].load(x_qs + (i0 + n*mma_A::I)*MMQ_MMA_TILE_X_K_Q8_1 + k0, MMQ_MMA_TILE_X_K_Q8_1);
+        }
+
+#pragma unroll
+        for (int l = 0; l < mma_C::ne/2; ++l) {
+            const int i = i0 + n*mma_A::I + mma_C::get_i(2*l);
+
+#pragma unroll
+            for (int k01 = 0; k01 < WARP_SIZE; k01 += QI8_1) {
+                const int k0 = k00 + k01;
+
+                dmA[n][l][k01/QI8_1] = __half22float2(x_dm[i*MMQ_MMA_TILE_X_K_Q8_1 + k0/QI8_1]);
+            }
+        }
+    }
+
+#pragma unroll
+    for (int j0 = 0; j0 < mmq_x; j0 += ntx*mma_C::J) {
+#pragma unroll
+        for (int k01 = 0; k01 < WARP_SIZE; k01 += QI8_1) {
+            mma_B    B;
+            float2 dsB[mma_C::ne/2];
+
+            B.load(y_qs + j0*MMQ_TILE_Y_K + k01, MMQ_TILE_Y_K);
+
+#pragma unroll
+            for (int l = 0; l < mma_C::ne/2; ++l) {
+                const int j = j0 + mma_C::get_j(l);
+
+                dsB[l] = __half22float2(y_dm[j*MMQ_TILE_Y_K + k01/QI8_1]);
+            }
+
+#pragma unroll
+            for (int n = 0; n < ntx; ++n) {
+                mma_C C;
+                C.mma_K8(A[n][k01/QI8_1], B);
+
+#pragma unroll
+                for (int l = 0; l < mma_C::ne; ++l) {
+                    sum[(j0/mma_C::J + n)*mma_C::ne + l] += dmA[n][l/2][k01/QI8_1].x*dsB[l%2].x*C.x[l];
+                    sum[(j0/mma_C::J + n)*mma_C::ne + l] += dmA[n][l/2][k01/QI8_1].y*dsB[l%2].y;
+                }
+            }
+        }
+    }
+}
+
+template <int mmq_x, int mmq_y, int nwarps>
+static __device__ __forceinline__ void vec_dot_q8_0_16_q8_1_dp4a(
+    const int * __restrict__ x, const int * __restrict__ y, float * __restrict__ sum, const int & k00) {
+
+    constexpr tile_x_sizes txs = MMQ_DP4A_TXS_Q8_0_16;
+    const int   * x_qs = (const int   *) x;
+    const float * x_df = (const float *) x_qs + txs.qs;
+    const int   * y_qs = (const int   *) y + 4;
+    const float * y_df = (const float *) y;
+
+// #pragma unroll
+    for (int k01 = 0; k01 < WARP_SIZE; k01 += QI8_0) {
+        const int k0 = k00 + k01;
+
+#pragma unroll
+        for (int j0 = 0; j0 < mmq_x; j0 += nwarps) {
+            const int j = j0 + threadIdx.y;
+
+#pragma unroll
+            for (int i0 = 0; i0 < mmq_y; i0 += WARP_SIZE) {
+                const int i = i0 + threadIdx.x;
+
+                sum[j0/nwarps*mmq_y/WARP_SIZE + i0/WARP_SIZE] += vec_dot_q8_0_16_q8_1_impl<QI8_0>(
+                    &x_qs[i*(2*WARP_SIZE + 1) + k0],
+                    &y_qs[j*MMQ_TILE_Y_K + k01],
+                    &x_df[i*(2*WARP_SIZE*2/QI8_0) + i/(QI8_0/4) + k0/(QI8_0/2)],
+                    y_df[j*MMQ_TILE_Y_K + k01/QI8_1]);
+            }
+        }
+    }
+}
+
+template <int mmq_x, int mmq_y, int nwarps>
+static __device__ __forceinline__ void vec_dot_q8_0_16_q8_1_mma(
+    const int * __restrict__ x, const int * __restrict__ y, float * __restrict__ sum, const int & k00) {
+#ifdef INT8_MMA_AVAILABLE
+
+    typedef mma_int_A_I16K4 mma_A;
+    typedef mma_int_A_I16K8 mma_A_K8;
+    typedef mma_int_B_J8K4  mma_B;
+    typedef mma_int_C_I16J8 mma_C;
+
+    constexpr int granularity = mmq_get_granularity_device(mmq_x);
+    constexpr int rows_per_warp = 2 * granularity;
+    constexpr int ntx = rows_per_warp/mma_C::I; // Number of x minitiles per warp.
+
+    y += (threadIdx.y % ntx) * (mma_B::J*MMQ_TILE_Y_K);
+
+    const int   * x_qs = (const int   *) x;
+    const float * x_df = (const float *) x_qs + WARP_SIZE*2;
+    const int   * y_qs = (const int   *) y + 4;
+    const float * y_df = (const float *) y;
+
+    const int i0 = (threadIdx.y / ntx) * (ntx*mma_A::I);
+
+    mma_A   A[ntx][8];
+    float  dA[ntx][mma_C::ne/2][8];
+
+#pragma unroll
+    for (int n = 0; n < ntx; ++n) {
+#pragma unroll
+        for (int k01 = 0; k01 < WARP_SIZE; k01 += 8) {
+            const int k0 = k00 + k01;
+
+            ((mma_A_K8 *) A[n])[k01/8].load(x_qs + (i0 + n*mma_A::I)*MMQ_MMA_TILE_X_K_Q3_K + k0, MMQ_MMA_TILE_X_K_Q3_K);
+        }
+
+#pragma unroll
+        for (int l = 0; l < mma_C::ne/2; ++l) {
+            const int i = i0 + n*mma_C::I + mma_C::get_i(2*l);
+
+#pragma unroll
+            for (int k01 = 0; k01 < WARP_SIZE; k01 += 4) {
+                const int k0 = k00 + k01;
+
+                dA[n][l][k01/4] = x_df[i*MMQ_MMA_TILE_X_K_Q3_K + k0/4];
+            }
+        }
+    }
+
+#pragma unroll
+    for (int j0 = 0; j0 < mmq_x; j0 += ntx*mma_C::J) {
+#pragma unroll
+        for (int k01 = 0; k01 < WARP_SIZE; k01 += QR3_K*VDR_Q3_K_Q8_1_MMQ) {
+            mma_B B[2];
+            float dB[mma_C::ne/2];
+
+            B[0].load(y_qs + j0*MMQ_TILE_Y_K + (k01 + 0),        MMQ_TILE_Y_K);
+            B[1].load(y_qs + j0*MMQ_TILE_Y_K + (k01 + mma_B::K), MMQ_TILE_Y_K);
+
+#pragma unroll
+            for (int l = 0; l < mma_C::ne/2; ++l) {
+                const int j = j0 + mma_C::get_j(l);
+
+                dB[l] = y_df[j*MMQ_TILE_Y_K + k01/QI8_1];
+            }
+
+#pragma unroll
+            for (int n = 0; n < ntx; ++n) {
+                mma_C C[2];
+                C[0].mma_K4(A[n][k01/4 + 0], B[0]);
+                C[1].mma_K4(A[n][k01/4 + 1], B[1]);
+
+#pragma unroll
+                for (int l = 0; l < mma_C::ne; ++l) {
+                    sum[(j0/mma_C::J + n)*mma_C::ne + l] += dB[l%2]*(C[0].x[l]*dA[n][l/2][k01/4 + 0] + C[1].x[l]*dA[n][l/2][k01/4 + 1]);
+                }
+            }
+        }
+    }
+#else
+    GGML_UNUSED(x); GGML_UNUSED(y); GGML_UNUSED(sum);
+    NO_DEVICE_CODE;
+#endif // INT8_MMA_AVAILABLE
+}
+
+template <int mmq_y, int nwarps, bool need_check> static __device__ __forceinline__ void load_tiles_q2_K(
+    const char * __restrict__ x, int * __restrict__ x_tile, const int & kbx0, const int & i_max, const int & stride) {
+
+#ifdef INT8_MMA_AVAILABLE
+    int   * x_qs = (int   *)  x_tile;
+    half2 * x_dm = (half2 *) (x_qs + 2*WARP_SIZE);
+#else
+    constexpr tile_x_sizes txs = mmq_get_dp4a_tile_x_sizes(GGML_TYPE_Q2_K, mmq_y);
+    int   * x_qs = (int   *)  x_tile;
+    half2 * x_dm = (half2 *) (x_qs + txs.qs);
+#endif // INT8_MMA_AVAILABLE
+
+    const int kqsx = threadIdx.x % QI2_K;
+
+#pragma unroll
+    for (int i0 = 0; i0 < mmq_y; i0 += nwarps * WARP_SIZE/QI2_K) {
+        int i = i0 + threadIdx.y*(WARP_SIZE/QI2_K) + threadIdx.x/QI2_K;
+
+        if (need_check) {
+            i = min(i, i_max);
+        }
+
+        const block_q2_K * bxi = (const block_q2_K *) x + kbx0 + i*stride;
+
+        const int x_ql_0 = get_int_b2(bxi->qs, kqsx);
+
+#pragma unroll
+        for (int l = 0; l < QR2_K; ++l) {
+            const int k = (kqsx/8)*32 + l*8 + kqsx % 8;
+
+            const int x_qs_k = (x_ql_0 >> (2*l)) & 0x03030303;
+
+#ifdef INT8_MMA_AVAILABLE
+            x_qs[i*MMQ_MMA_TILE_X_K_Q2_K + k] = x_qs_k;
+#else
+            x_qs[i*(2*WARP_SIZE + 1)     + k] = x_qs_k;
+#endif // INT8_MMA_AVAILABLE
+        }
+
+        const int sc_m = bxi->scales[kqsx];
+#ifdef FAST_FP16_AVAILABLE
+        const half2 x_dm_ik = __hmul2(bxi->dm, make_half2(sc_m & 0x0F, sc_m >> 4));
+#else
+        const float2 bxi_dmf = __half22float2(bxi->dm);
+        const half2 x_dm_ik = make_half2(bxi_dmf.x*(sc_m & 0x0F), bxi_dmf.y*(sc_m >> 4));
+#endif // FAST_FP16_AVAILABLE
+
+#ifdef INT8_MMA_AVAILABLE
+        x_dm[i*MMQ_MMA_TILE_X_K_Q2_K + kqsx] = x_dm_ik;
+#else
+        x_dm[i*(WARP_SIZE + 1)       + kqsx] = x_dm_ik;
+#endif // INT8_MMA_AVAILABLE
+    }
+}
+
+template <int mmq_x, int mmq_y, int nwarps>
+static __device__ __forceinline__ void vec_dot_q2_K_q8_1_dp4a(
+    const int * __restrict__ x, const int * __restrict__ y, float * __restrict__ sum, const int & k00) {
+
+    constexpr tile_x_sizes txs = mmq_get_dp4a_tile_x_sizes(GGML_TYPE_Q2_K, mmq_y);
+    const int   * x_qs = (const int   *) x;
+    const half2 * x_dm = (const half2 *) x_qs + txs.qs;
+    const int   * y_qs = (const int   *) y + 4;
+    const half2 * y_ds = (const half2 *) y;
+
+    float2 y_df[mmq_x/nwarps];
+#pragma unroll
+    for (int j0 = 0; j0 < mmq_x; j0 += nwarps) {
+        const int j = j0 + threadIdx.y;
+
+        y_df[j0/nwarps] = __half22float2(y_ds[j*MMQ_TILE_Y_K]);
+    }
+
+#pragma unroll
+    for (int k01 = 0; k01 < WARP_SIZE; k01 += QR2_K*VDR_Q2_K_Q8_1_MMQ) {
+        const int k0 = k00 + k01;
+
+#pragma unroll
+        for (int j0 = 0; j0 < mmq_x; j0 += nwarps) {
+            const int j = j0 + threadIdx.y;
+
+#pragma unroll
+            for (int i0 = 0; i0 < mmq_y; i0 += WARP_SIZE) {
+                const int i = i0 + threadIdx.x;
+
+                if (k01 < WARP_SIZE/2) {
+                    constexpr int ns = 2;
+                    sum[j0/nwarps*mmq_y/WARP_SIZE + i0/WARP_SIZE] += vec_dot_q2_K_q8_1_impl_mmq<ns>(
+                        &x_qs[i*(2*WARP_SIZE + 1) + k0], &y_qs[j*MMQ_TILE_Y_K + k01],
+                        &x_dm[i*(WARP_SIZE + 1) + k0/4], k01 < WARP_SIZE/2 ? y_df[j0/nwarps].x : y_df[j0/nwarps].y,
+                        &y_ds[j*MMQ_TILE_Y_K + (1 + k01/QI8_1)]);
+                } else {
+                    constexpr int ns = 1;
+                    sum[j0/nwarps*mmq_y/WARP_SIZE + i0/WARP_SIZE] += vec_dot_q2_K_q8_1_impl_mmq<ns>(
+                        &x_qs[i*(2*WARP_SIZE + 1) + k0], &y_qs[j*MMQ_TILE_Y_K + k01],
+                        &x_dm[i*(WARP_SIZE + 1) + k0/4], k01 < WARP_SIZE/2 ? y_df[j0/nwarps].x : y_df[j0/nwarps].y,
+                        &y_ds[j*MMQ_TILE_Y_K + (1 + k01/QI8_1)]);
+                }
+            }
+        }
+    }
+}
+
+template <int mmq_x, int mmq_y, int nwarps>
+static __device__ __forceinline__ void vec_dot_q2_K_q8_1_mma(
+    const int * __restrict__ x, const int * __restrict__ y, float * __restrict__ sum, const int & k00) {
+#ifdef INT8_MMA_AVAILABLE
+
+    typedef mma_int_A_I16K4 mma_A;
+    typedef mma_int_A_I16K8 mma_A_K8;
+    typedef mma_int_B_J8K4  mma_B;
+    typedef mma_int_C_I16J8 mma_C;
+
+    constexpr int granularity = mmq_get_granularity_device(mmq_x);
+    constexpr int rows_per_warp = 2 * granularity;
+    constexpr int ntx = rows_per_warp/mma_C::I; // Number of x minitiles per warp.
+
+    y += (threadIdx.y % ntx) * (mma_B::J*MMQ_TILE_Y_K);
+
+    const int   * x_qs = (const int   *) x;
+    const half2 * x_dm = (const half2 *) x_qs + WARP_SIZE*2;
+    const int   * y_qs = (const int   *) y + 4;
+    const half2 * y_ds = (const half2 *) y;
+
+    const int i0 = (threadIdx.y / ntx) * (ntx*mma_A::I);
+
+    mma_A   A[ntx][8];
+    float  dA[ntx][mma_C::ne/2][8];
+    float  mA[ntx][mma_C::ne/2][8];
+
+#pragma unroll
+    for (int n = 0; n < ntx; ++n) {
+#pragma unroll
+        for (int k01 = 0; k01 < WARP_SIZE; k01 += QI8_1) {
+            const int k0 = k00 + k01;
+
+            ((mma_A_K8 *) A[n])[k01/QI8_1].load(x_qs + (i0 + n*mma_A::I)*MMQ_MMA_TILE_X_K_Q2_K + k0, MMQ_MMA_TILE_X_K_Q2_K);
+        }
+    }
+
+#pragma unroll
+    for (int n = 0; n < ntx; ++n) {
+#pragma unroll
+        for (int l = 0; l < mma_C::ne/2; ++l) {
+            const int i = i0 + n*mma_C::I + mma_C::get_i(2*l);
+
+#pragma unroll
+            for (int k01 = 0; k01 < WARP_SIZE; k01 += QI8_1/2) {
+                const int k0 = k00 + k01;
+
+                const float2 dm = __half22float2(x_dm[i*MMQ_MMA_TILE_X_K_Q2_K + k0/(QI8_1/2)]);
+
+                dA[n][l][k01/(QI8_1/2)] = dm.x;
+                mA[n][l][k01/(QI8_1/2)] = dm.y;
+            }
+        }
+    }
+
+#pragma unroll
+    for (int j0 = 0; j0 < mmq_x; j0 += ntx*mma_C::J) {
+        float2 dB[mma_C::ne/2];
+
+#pragma unroll
+        for (int l = 0; l < mma_C::ne/2; ++l) {
+            const int j = j0 + mma_C::get_j(l);
+
+            dB[l] = __half22float2(y_ds[j*MMQ_TILE_Y_K]);
+        }
+
+#pragma unroll
+        for (int k01 = 0; k01 < WARP_SIZE; k01 += QI8_1) {
+            mma_B B[2];
+
+            B[0].load(y_qs + j0*MMQ_TILE_Y_K + (k01 + 0),        MMQ_TILE_Y_K);
+            B[1].load(y_qs + j0*MMQ_TILE_Y_K + (k01 + mma_B::K), MMQ_TILE_Y_K);
+
+            mma_C Cm[2];
+            if (k01 >= WARP_SIZE * 3/4) {
+                mma_A A1;
+                A1.x[0] = 0x01010101;
+                A1.x[1] = 0x01010101;
+                Cm[0].mma_K4(A1, B[0]);
+                Cm[1].mma_K4(A1, B[1]);
+            }
+
+#pragma unroll
+            for (int n = 0; n < ntx; ++n) {
+                mma_C Cd[2];
+
+                Cd[0].mma_K4(A[n][k01/4 + 0], B[0]);
+                Cd[1].mma_K4(A[n][k01/4 + 1], B[1]);
+
+#pragma unroll
+                for (int l = 0; l < mma_C::ne; ++l) {
+                    float tmp = Cd[0].x[l]*dA[n][l/2][k01/4 + 0] + Cd[1].x[l]*dA[n][l/2][k01/4 + 1];
+                    if (k01 >= WARP_SIZE * 3/4) {
+                        tmp -= Cm[0].x[l]*mA[n][l/2][k01/4 + 0] + Cm[1].x[l]*mA[n][l/2][k01/4 + 1];
+                    }
+                    sum[(j0/mma_C::J + n)*mma_C::ne + l] += tmp*(k01 < WARP_SIZE/2 ? dB[l%2].x : dB[l%2].y);
+                }
+            }
+        }
+
+#pragma unroll
+        for (int k01 = 0; k01 < WARP_SIZE * 3/4; k01 += QI8_1) {
+            float2 sB[mma_C::ne/2];
+
+#pragma unroll
+            for (int l = 0; l < mma_C::ne/2; ++l) {
+                const int j = j0 + mma_C::get_j(l);
+
+                sB[l] = __half22float2(y_ds[j*MMQ_TILE_Y_K + (1 + k01/QI8_1)]);
+            }
+
+#pragma unroll
+            for (int n = 0; n < ntx; ++n) {
+#pragma unroll
+                for (int l = 0; l < mma_C::ne; ++l) {
+                    sum[(j0/mma_C::J + n)*mma_C::ne + l] -= mA[n][l/2][k01/4 + 0]*sB[l%2].x;
+                    sum[(j0/mma_C::J + n)*mma_C::ne + l] -= mA[n][l/2][k01/4 + 1]*sB[l%2].y;
+                }
+            }
+        }
+    }
+#else
+    GGML_UNUSED(x); GGML_UNUSED(y); GGML_UNUSED(sum);
+    NO_DEVICE_CODE;
+#endif // INT8_MMA_AVAILABLE
+}
+
+template <int mmq_y, int nwarps, bool need_check> static __device__ __forceinline__ void load_tiles_q3_K(
+    const char * __restrict__ x, int * __restrict__ x_tile, const int & kbx0, const int & i_max, const int & stride) {
+
+#ifdef INT8_MMA_AVAILABLE
+    int   * x_qs = (int   *)  x_tile;
+    float * x_df = (float *) (x_qs + WARP_SIZE*2);
+#else
+    constexpr tile_x_sizes txs = mmq_get_dp4a_tile_x_sizes(GGML_TYPE_Q3_K, mmq_y);
+    int   * x_qs = (int   *)  x_tile;
+    float * x_df = (float *) (x_qs + txs.qs);
+    int   * x_sc = (int   *) (x_df + txs.dm);
+#endif // INT8_MMA_AVAILABLE
+
+    const int kqsx = threadIdx.x % QI3_K;
+
+#pragma unroll
+    for (int i0 = 0; i0 < mmq_y; i0 += nwarps * WARP_SIZE/QI3_K) {
+        int i = i0 + threadIdx.y * (WARP_SIZE/QI3_K) + threadIdx.x / QI3_K;
+
+        if (need_check) {
+            i = min(i, i_max);
+        }
+
+        const block_q3_K * bxi = (const block_q3_K *) x + kbx0 + i*stride;
+
+        const int x_ql_0 = get_int_b2(bxi->qs,    kqsx);
+        const int x_qh_0 = get_int_b2(bxi->hmask, kqsx % (QI3_K/2)) >> (4 * (kqsx / (QI3_K/2)));
+
+#pragma unroll
+        for (int l = 0; l < QR3_K; ++l) {
+            const int k = (kqsx/8)*32 + l*8 + kqsx % 8;
+
+            const int x_ql_k =  (x_ql_0 >> (2*l))       & 0x03030303;
+            const int x_qh_k = ((x_qh_0 >>    l)  << 2) & 0x04040404;
+
+            const int x_qs_k = __vsubss4(x_ql_k | x_qh_k, 0x04040404);
+
+#ifdef INT8_MMA_AVAILABLE
+            x_qs[i*MMQ_MMA_TILE_X_K_Q3_K + k] = x_qs_k;
+#else
+            x_qs[i*(2*WARP_SIZE + 1)     + k] = x_qs_k;
+#endif // INT8_MMA_AVAILABLE
+        }
+    }
+
+#pragma unroll
+    for (int i0 = 0; i0 < mmq_y; i0 += nwarps*8) {
+        int i = i0 + threadIdx.y*8 + threadIdx.x/(WARP_SIZE/8);
+
+        if (need_check) {
+            i = min(i, i_max);
+        }
+
+        const block_q3_K * bxi = (const block_q3_K *) x + kbx0 + i*stride;
+
+        const int ksc = threadIdx.x % (WARP_SIZE/8);
+
+        const int ksc_low = ksc % (QI3_K/8);
+        const int shift_low = 4 * (ksc / (QI3_K/8));
+        const int sc_low = (get_int_b2(bxi->scales, ksc_low) >> shift_low) & 0x0F0F0F0F;
+
+        const int ksc_high = QI3_K/8;
+        const int shift_high = 2 * ksc;
+        const int sc_high = ((get_int_b2(bxi->scales, ksc_high) >> shift_high) << 4) & 0x30303030;
+
+        const int sc = __vsubss4(sc_low | sc_high, 0x20202020);
+
+#ifdef INT8_MMA_AVAILABLE
+        const int8_t * sc8 = (const int8_t *) &sc;
+        const float d = bxi->d;
+
+#pragma unroll
+        for (int l = 0; l < sizeof(int); ++l) {
+            x_df[i*MMQ_MMA_TILE_X_K_Q3_K + sizeof(int)*(threadIdx.x % (WARP_SIZE/8)) + l] = d*sc8[l];
+        }
+#else
+        x_sc[i*(WARP_SIZE/8) + i/8 + threadIdx.x % (WARP_SIZE/8)] = sc;
+#endif // INT8_MMA_AVAILABLE
+    }
+
+#ifndef INT8_MMA_AVAILABLE
+#pragma unroll
+    for (int i0 = 0; i0 < mmq_y; i0 += nwarps*WARP_SIZE) {
+        int i = (i0 + threadIdx.y*WARP_SIZE + threadIdx.x) % mmq_y;
+
+        if (need_check) {
+            i = min(i, i_max);
+        }
+
+        const block_q3_K * bxi = (const block_q3_K *) x + kbx0 + i*stride;
+
+        x_df[i] = bxi->d;
+    }
+#endif // INT8_MMA_AVAILABLE
+}
+
+template <int mmq_x, int mmq_y, int nwarps>
+static __device__ __forceinline__ void vec_dot_q3_K_q8_1_dp4a(
+    const int * __restrict__ x, const int * __restrict__ y, float * __restrict__ sum, const int & k00) {
+
+    constexpr tile_x_sizes txs = mmq_get_dp4a_tile_x_sizes(GGML_TYPE_Q3_K, mmq_y);
+    const int   * x_qs = (const int   *) x;
+    const float * x_df = (const float *) x_qs + txs.qs;
+    const int   * x_sc = (const int   *) x_df + txs.dm;
+    const int   * y_qs = (const int   *) y + 4;
+    const float * y_df = (const float *) y;
+
+// #pragma unroll
+    for (int k01 = 0; k01 < WARP_SIZE; k01 += QR3_K*VDR_Q3_K_Q8_1_MMQ) {
+        const int k0 = k00 + k01;
+
+#pragma unroll
+        for (int j0 = 0; j0 < mmq_x; j0 += nwarps) {
+            const int j = j0 + threadIdx.y;
+
+#pragma unroll
+            for (int i0 = 0; i0 < mmq_y; i0 += WARP_SIZE) {
+                const int i = i0 + threadIdx.x;
+
+                const int8_t * scales = ((const int8_t *) (x_sc + i*(WARP_SIZE/8) + i/8)) + k0/4;
+
+                sum[j0/nwarps*mmq_y/WARP_SIZE + i0/WARP_SIZE] += vec_dot_q3_K_q8_1_impl_mmq(
+                    &x_qs[i*(2*WARP_SIZE + 1) + k0], &y_qs[j*MMQ_TILE_Y_K + k01], scales,
+                    x_df[i], y_df[j*MMQ_TILE_Y_K + k01/QI8_1]);
+            }
+        }
+    }
+}
+
+static __device__ __forceinline__ int unpack_scales_q45_K(const int * scales, const int ksc) {
+    // scale arrangement after the following two lines:
+    //   - ksc == 0: sc0, sc1, sc2, sc3
+    //   - ksc == 1: sc4, sc5, sc6, sc7
+    //   - ksc == 2:  m0,  m1,  m2,  m3
+    //   - ksc == 3:  m4,  m5,  m6,  m7
+    return ((scales[(ksc%2) + (ksc!=0)] >> (4 * (ksc & (ksc/2)))) & 0x0F0F0F0F) | // lower 4 bits
+           ((scales[ksc/2]              >> (2 * (ksc % 2)))       & 0x30303030);  // upper 2 bits
+}
+
+template <int mmq_y, int nwarps, bool need_check> static __device__ __forceinline__ void load_tiles_q4_K(
+    const char * __restrict__ x, int * __restrict__ x_tile, const int & kbx0, const int & i_max, const int & stride) {
+
+#ifdef INT8_MMA_AVAILABLE
+    int   * x_qs = (int   *)  x_tile;
+    half2 * x_dm = (half2 *) (x_qs + 2*WARP_SIZE);
+#else
+    constexpr tile_x_sizes txs = mmq_get_dp4a_tile_x_sizes(GGML_TYPE_Q4_K, mmq_y);
+    int   * x_qs = (int   *)  x_tile;
+    half2 * x_dm = (half2 *) (x_qs + txs.qs);
+    int   * x_sc = (int   *) (x_dm + txs.dm);
+#endif // INT8_MMA_AVAILABLE
+
+#pragma unroll
+    for (int i0 = 0; i0 < mmq_y; i0 += nwarps) {
+        int i = i0 + threadIdx.y;
+
+        if (need_check) {
+            i = min(i, i_max);
+        }
+
+        const block_q4_K * bxi = (const block_q4_K *) x + kbx0 + i*stride;
+        const int qs0 = get_int_b4(bxi->qs, threadIdx.x);
+
+#ifdef INT8_MMA_AVAILABLE
+        x_qs[i*MMQ_MMA_TILE_X_K_Q8_1 + 16*(threadIdx.x/8) + threadIdx.x % 8 + 0] = (qs0 >> 0) & 0x0F0F0F0F;
+        x_qs[i*MMQ_MMA_TILE_X_K_Q8_1 + 16*(threadIdx.x/8) + threadIdx.x % 8 + 8] = (qs0 >> 4) & 0x0F0F0F0F;
+#else
+        x_qs[i*(WARP_SIZE + 1) + threadIdx.x] = qs0;
+#endif // INT8_MMA_AVAILABLE
+    }
+
+#ifdef INT8_MMA_AVAILABLE
+
+#pragma unroll
+    for (int i0 = 0; i0 < mmq_y; i0 += nwarps*16) {
+        int i = (i0 + threadIdx.y*16 + threadIdx.x/(WARP_SIZE/16)) % mmq_y;
+
+        if (need_check) {
+            i = min(i, i_max);
+        }
+
+        const block_q4_K * bxi = (const block_q4_K *) x + kbx0 + i*stride;
+
+        const int * scales = (const int *) bxi->scales;
+        const int ksc = threadIdx.x % (WARP_SIZE/16);
+
+        const int sc32 = unpack_scales_q45_K(scales, ksc + 0);
+        const int  m32 = unpack_scales_q45_K(scales, ksc + 2);
+
+        const uint8_t * sc8 = (const uint8_t *) &sc32;
+        const uint8_t *  m8 = (const uint8_t *)  &m32;
+
+        const half2 dm = bxi->dm * make_half2(1.0f, -1.0f);
+
+#pragma unroll
+        for (int l = 0; l < sizeof(int); ++l) {
+            x_dm[i*MMQ_MMA_TILE_X_K_Q8_1 + sizeof(int)*ksc + l] = dm*make_half2(sc8[l], m8[l]);
+        }
+    }
+
+#else
+
+#pragma unroll
+    for (int i0 = 0; i0 < mmq_y; i0 += nwarps*QI4_K) {
+        int i = (i0 + threadIdx.y*QI4_K + threadIdx.x) % mmq_y;
+
+        if (need_check) {
+            i = min(i, i_max);
+        }
+
+        const block_q4_K * bxi = (const block_q4_K *) x + kbx0 + i*stride;
+
+        x_dm[i] = bxi->dm;
+    }
+
+#pragma unroll
+    for (int i0 = 0; i0 < mmq_y; i0 += nwarps * 8) {
+        int i = (i0 + threadIdx.y * 8 + threadIdx.x / (WARP_SIZE/8)) % mmq_y;
+
+        if (need_check) {
+            i = min(i, i_max);
+        }
+
+        const block_q4_K * bxi = (const block_q4_K *) x + kbx0 + i*stride + (threadIdx.x % (WARP_SIZE/8)) / (QI4_K/8);
+
+        const int * scales = (const int *) bxi->scales;
+
+        const int ksc = threadIdx.x % (WARP_SIZE/8);
+        const int scales8 = unpack_scales_q45_K(scales, ksc);
+
+        x_sc[i*(WARP_SIZE/8) + i/8 + ksc] = scales8;
+    }
+#endif // INT8_MMA_AVAILABLE
+}
+
+template <int mmq_x, int mmq_y, int nwarps>
+static __device__ __forceinline__ void vec_dot_q4_K_q8_1_dp4a(
+    const int * __restrict__ x, const int * __restrict__ y, float * __restrict__ sum, const int & k00) {
+
+    constexpr tile_x_sizes txs = mmq_get_dp4a_tile_x_sizes(GGML_TYPE_Q4_K, mmq_y);
+    const int   * x_qs = (const int   *) x;
+    const half2 * x_dm = (const half2 *) x_qs + txs.qs;
+    const int   * x_sc = (const int   *) x_dm + txs.dm;
+    const int   * y_qs = (const int   *) y + 4;
+    const half2 * y_ds = (const half2 *) y;
+
+// #pragma unroll
+    for (int k01 = 0; k01 < WARP_SIZE; k01 += QR4_K*VDR_Q4_K_Q8_1_MMQ) {
+        const int k0 = k00 + k01;
+
+#pragma unroll
+        for (int j0 = 0; j0 < mmq_x; j0 += nwarps) {
+            const int j = j0 + threadIdx.y;
+
+#pragma unroll
+            for (int i0 = 0; i0 < mmq_y; i0 += WARP_SIZE) {
+                const int i = i0 + threadIdx.x;
+
+                const uint8_t * sc = (const uint8_t *) &x_sc[i * (WARP_SIZE/8) + i/8 + k0/32] + 2*(k01/16);
+
+                sum[j0/nwarps*mmq_y/WARP_SIZE + i0/WARP_SIZE] += vec_dot_q4_K_q8_1_impl_mmq(
+                    &x_qs[i*(WARP_SIZE + 1) + k0/2], &y_qs[j*MMQ_TILE_Y_K + k01], sc, sc+8,
+                    x_dm[i], &y_ds[j*MMQ_TILE_Y_K + k01/QI8_1]);
+            }
+        }
+    }
+}
+
+template <int mmq_y, int nwarps, bool need_check> static __device__ __forceinline__ void load_tiles_q5_K(
+    const char * __restrict__ x, int * __restrict__ x_tile, const int & kbx0, const int & i_max, const int & stride) {
+
+#ifdef INT8_MMA_AVAILABLE
+    int   * x_qs = (int   *)  x_tile;
+    half2 * x_dm = (half2 *) (x_qs + WARP_SIZE*2);
+#else
+    constexpr tile_x_sizes txs = mmq_get_dp4a_tile_x_sizes(GGML_TYPE_Q5_K, mmq_y);
+    int   * x_qs = (int   *)  x_tile;
+    half2 * x_dm = (half2 *) (x_qs + txs.qs);
+    int   * x_sc = (int   *) (x_dm + txs.dm);
+#endif // INT8_MMA_AVAILABLE
+
+#pragma unroll
+    for (int i0 = 0; i0 < mmq_y; i0 += nwarps) {
+        int i = i0 + threadIdx.y;
+
+        if (need_check) {
+            i = min(i, i_max);
+        }
+
+        const block_q5_K * bxi = (const block_q5_K *) x + kbx0 + i*stride;
+        const int ky = QR5_K*threadIdx.x;
+
+        const int ql = get_int_b4(bxi->qs, threadIdx.x);
+        const int ql0 = (ql >> 0) & 0x0F0F0F0F;
+        const int ql1 = (ql >> 4) & 0x0F0F0F0F;
+
+        const int qh = get_int_b4(bxi->qh, threadIdx.x % (QI5_K/4));
+        const int qh0 = ((qh >> (2 * (threadIdx.x / (QI5_K/4)) + 0)) << 4) & 0x10101010;
+        const int qh1 = ((qh >> (2 * (threadIdx.x / (QI5_K/4)) + 1)) << 4) & 0x10101010;
+
+        const int kq0 = ky - ky % (QI5_K/2) + threadIdx.x % (QI5_K/4) + 0;
+        const int kq1 = ky - ky % (QI5_K/2) + threadIdx.x % (QI5_K/4) + QI5_K/4;
+
+#ifdef INT8_MMA_AVAILABLE
+        x_qs[i*MMQ_MMA_TILE_X_K_Q8_1 + kq0] = ql0 | qh0;
+        x_qs[i*MMQ_MMA_TILE_X_K_Q8_1 + kq1] = ql1 | qh1;
+#else
+        x_qs[i*(2*WARP_SIZE + 1)     + kq0] = ql0 | qh0;
+        x_qs[i*(2*WARP_SIZE + 1)     + kq1] = ql1 | qh1;
+#endif // INT8_MMA_AVAILABLE
+    }
+
+#ifdef INT8_MMA_AVAILABLE
+
+#pragma unroll
+    for (int i0 = 0; i0 < mmq_y; i0 += nwarps*16) {
+        int i = (i0 + threadIdx.y*16 + threadIdx.x/(WARP_SIZE/16)) % mmq_y;
+
+        if (need_check) {
+            i = min(i, i_max);
+        }
+
+        const block_q5_K * bxi = (const block_q5_K *) x + kbx0 + i*stride;
+
+        const int * scales = (const int *) bxi->scales;
+        const int ksc = threadIdx.x % (WARP_SIZE/16);
+
+        const int sc32 = unpack_scales_q45_K(scales, ksc + 0);
+        const int  m32 = unpack_scales_q45_K(scales, ksc + 2);
+
+        const uint8_t * sc8 = (const uint8_t *) &sc32;
+        const uint8_t *  m8 = (const uint8_t *)  &m32;
+
+        const half2 dm = bxi->dm * make_half2(1.0f, -1.0f);
+
+#pragma unroll
+        for (int l = 0; l < sizeof(int); ++l) {
+            x_dm[i*MMQ_MMA_TILE_X_K_Q8_1 + sizeof(int)*ksc + l] = dm*make_half2(sc8[l], m8[l]);
+        }
+    }
+
+#else
+
+#pragma unroll
+    for (int i0 = 0; i0 < mmq_y; i0 += nwarps*QI5_K) {
+        int i = (i0 + threadIdx.y*QI5_K + threadIdx.x) % mmq_y;
+
+        if (need_check) {
+            i = min(i, i_max);
+        }
+
+        const block_q5_K * bxi = (const block_q5_K *) x + kbx0 + i*stride;
+
+        x_dm[i] = bxi->dm;
+    }
+
+#pragma unroll
+    for (int i0 = 0; i0 < mmq_y; i0 += nwarps*8) {
+        int i = (i0 + threadIdx.y*8 + threadIdx.x/(WARP_SIZE/8)) % mmq_y;
+
+        if (need_check) {
+            i = min(i, i_max);
+        }
+
+        const block_q5_K * bxi = (const block_q5_K *) x + kbx0 + i*stride;
+
+        const int * scales = (const int *) bxi->scales;
+
+        const int ksc = threadIdx.x % (WARP_SIZE/8);
+        const int scales8 = unpack_scales_q45_K(scales, ksc);
+
+        x_sc[i*(WARP_SIZE/8) + i/8 + ksc] = scales8;
+    }
+#endif // INT8_MMA_AVAILABLE
+}
+
+template <int mmq_x, int mmq_y, int nwarps>
+static __device__ __forceinline__ void vec_dot_q5_K_q8_1_dp4a(
+    const int * __restrict__ x, const int * __restrict__ y, float * __restrict__ sum, const int & k00) {
+
+    constexpr tile_x_sizes txs = mmq_get_dp4a_tile_x_sizes(GGML_TYPE_Q5_K, mmq_y);
+    const int   * x_qs = (const int   *) x;
+    const half2 * x_dm = (const half2 *) x_qs + txs.qs;
+    const int   * x_sc = (const int   *) x_dm + txs.dm;
+    const int   * y_qs = (const int   *) y + 4;
+    const half2 * y_ds = (const half2 *) y;
+
+// #pragma unroll
+    for (int k01 = 0; k01 < WARP_SIZE; k01 += QR5_K*VDR_Q5_K_Q8_1_MMQ) {
+        const int k0 = k00 + k01;
+
+#pragma unroll
+        for (int j0 = 0; j0 < mmq_x; j0 += nwarps) {
+            const int j = j0 + threadIdx.y;
+
+#pragma unroll
+            for (int i0 = 0; i0 < mmq_y; i0 += WARP_SIZE) {
+                const int i = i0 + threadIdx.x;
+
+                const uint8_t * sc = ((const uint8_t *) &x_sc[i * (WARP_SIZE/8) + i/8 + k00/32]) + 2*(k01/16);
+
+                sum[j0/nwarps*mmq_y/WARP_SIZE + i0/WARP_SIZE] += vec_dot_q5_K_q8_1_impl_mmq(
+                    &x_qs[i*(QR5_K*WARP_SIZE + 1) + k0], &y_qs[j*MMQ_TILE_Y_K + k01], sc, sc+8,
+                    x_dm[i], &y_ds[j*MMQ_TILE_Y_K + k01/QI8_1]);
+            }
+        }
+    }
+}
+
+template <int mmq_y, int nwarps, bool need_check> static __device__ __forceinline__ void load_tiles_q6_K(
+    const char * __restrict__ x, int * __restrict__ x_tile, const int & kbx0, const int & i_max, const int & stride) {
+
+#ifdef INT8_MMA_AVAILABLE
+    int   * x_qs = (int   *)  x_tile;
+    float * x_df = (float *) (x_qs + WARP_SIZE*2);
+    int   * x_sc = (int   *) (x_df + WARP_SIZE/QI6_K);
+#else
+    constexpr tile_x_sizes txs = mmq_get_dp4a_tile_x_sizes(GGML_TYPE_Q6_K, mmq_y);
+    int   * x_qs = (int   *)  x_tile;
+    float * x_df = (float *) (x_qs + txs.qs);
+    int   * x_sc = (int   *) (x_df + txs.dm);
+#endif // INT8_MMA_AVAILABLE
+
+#pragma unroll
+    for (int i0 = 0; i0 < mmq_y; i0 += nwarps) {
+        int i = i0 + threadIdx.y;
+
+        if (need_check) {
+            i = min(i, i_max);
+        }
+
+        const block_q6_K * bxi = (const block_q6_K *) x + kbx0 + i*stride;
+
+        const int ql = get_int_b2(bxi->ql, threadIdx.x);
+        const int ql0 = (ql >> 0) & 0x0F0F0F0F;
+        const int ql1 = (ql >> 4) & 0x0F0F0F0F;
+
+        const int qh = get_int_b2(bxi->qh, (QI6_K/4) * (threadIdx.x / (QI6_K/2)) + threadIdx.x % (QI6_K/4));
+        const int qh0 = ((qh >> ((threadIdx.x & 0x08) >> 2)) << 4) & 0x30303030;
+        const int qh1 =  (qh >> ((threadIdx.x & 0x08) >> 2))       & 0x30303030;
+
+        const int kq0 = 2*threadIdx.x - threadIdx.x % (QI6_K/2) + 0;
+        const int kq1 = 2*threadIdx.x - threadIdx.x % (QI6_K/2) + QI6_K/2;
+
+#ifdef INT8_MMA_AVAILABLE
+        x_qs[i*MMQ_MMA_TILE_X_K_Q6_K + kq0] = __vsubss4(ql0 | qh0, 0x20202020);
+        x_qs[i*MMQ_MMA_TILE_X_K_Q6_K + kq1] = __vsubss4(ql1 | qh1, 0x20202020);
+#else
+        x_qs[i*(2*WARP_SIZE + 1)     + kq0] = __vsubss4(ql0 | qh0, 0x20202020);
+        x_qs[i*(2*WARP_SIZE + 1)     + kq1] = __vsubss4(ql1 | qh1, 0x20202020);
+#endif // INT8_MMA_AVAILABLE
+    }
+
+    const int blocks_per_tile_x_row = WARP_SIZE / QI6_K;  // == 1 if QK_K == 256
+    const int kbxd = threadIdx.x % blocks_per_tile_x_row; // == 0 if QK_K == 256
+
+#pragma unroll
+    for (int i0 = 0; i0 < mmq_y; i0 += nwarps * QI6_K) {
+        int i = (i0 + threadIdx.y * QI6_K + threadIdx.x / blocks_per_tile_x_row) % mmq_y;
+
+        if (need_check) {
+            i = min(i, i_max);
+        }
+
+        const block_q6_K * bxi = (const block_q6_K *) x + kbx0 + i*stride + kbxd;
+
+#ifdef INT8_MMA_AVAILABLE
+        x_df[i*MMQ_MMA_TILE_X_K_Q6_K       + kbxd] = bxi->d;
+#else
+        x_df[i*(WARP_SIZE/QI6_K) + i/QI6_K + kbxd] = bxi->d;
+#endif // INT8_MMA_AVAILABLE
+    }
+
+#pragma unroll
+    for (int i0 = 0; i0 < mmq_y; i0 += nwarps * 8) {
+        int i = (i0 + threadIdx.y * 8 + threadIdx.x / (WARP_SIZE/8)) % mmq_y;
+
+        if (need_check) {
+            i = min(i, i_max);
+        }
+
+        const block_q6_K * bxi = (const block_q6_K *) x + kbx0 + i*stride + (threadIdx.x % (WARP_SIZE/8)) / 4;
+
+#ifdef INT8_MMA_AVAILABLE
+        x_sc[i*MMQ_MMA_TILE_X_K_Q6_K + threadIdx.x % (WARP_SIZE/8)] = get_int_b2(bxi->scales, threadIdx.x % (QI6_K/8));
+#else
+        x_sc[i*(WARP_SIZE/8) + i/8   + threadIdx.x % (WARP_SIZE/8)] = get_int_b2(bxi->scales, threadIdx.x % (QI6_K/8));
+#endif // INT8_MMA_AVAILABLE
+    }
+}
+
+template <int mmq_x, int mmq_y, int nwarps>
+static __device__ __forceinline__ void vec_dot_q6_K_q8_1_dp4a(
+    const int * __restrict__ x, const int * __restrict__ y, float * __restrict__ sum, const int & k00) {
+
+    constexpr tile_x_sizes txs = mmq_get_dp4a_tile_x_sizes(GGML_TYPE_Q6_K, mmq_y);
+    const int   * x_qs = (const int   *) x;
+    const float * x_df = (const float *) x_qs + txs.qs;
+    const int   * x_sc = (const int   *) x_df + txs.dm;
+    const int   * y_qs = (const int   *) y + 4;
+    const float * y_df = (const float *) y;
+
+// #pragma unroll
+    for (int k01 = 0; k01 < WARP_SIZE; k01 += QR6_K*VDR_Q6_K_Q8_1_MMQ) {
+        const int k0 = k00 + k01;
+
+#pragma unroll
+        for (int j0 = 0; j0 < mmq_x; j0 += nwarps) {
+            const int j = j0 + threadIdx.y;
+
+#pragma unroll
+            for (int i0 = 0; i0 < mmq_y; i0 += WARP_SIZE) {
+                const int i = i0 + threadIdx.x;
+
+                const int8_t * sc = ((const int8_t *) &x_sc[i * (WARP_SIZE/8) + i/8 + k0/16]);
+
+                sum[j0/nwarps*mmq_y/WARP_SIZE + i0/WARP_SIZE] += vec_dot_q6_K_q8_1_impl_mmq(
+                    &x_qs[i*(QR6_K*WARP_SIZE + 1) + k0], &y_qs[j*MMQ_TILE_Y_K + k01], sc,
+                    x_df[i*(WARP_SIZE/QI6_K) + i/QI6_K], &y_df[j*MMQ_TILE_Y_K + k01/QI8_1]);
+            }
+        }
+    }
+}
+
+template <int mmq_x, int mmq_y, int nwarps>
+static __device__ __forceinline__ void vec_dot_q6_K_q8_1_mma(
+    const int * __restrict__ x, const int * __restrict__ y, float * __restrict__ sum, const int & k00) {
+#ifdef INT8_MMA_AVAILABLE
+
+    typedef mma_int_A_I16K4 mma_A;
+    typedef mma_int_B_J8K4  mma_B;
+    typedef mma_int_C_I16J8 mma_C;
+
+    constexpr int granularity = mmq_get_granularity_device(mmq_x);
+    constexpr int rows_per_warp = 2 * granularity;
+    constexpr int ntx = rows_per_warp/mma_C::I; // Number of x minitiles per warp.
+
+    y += (threadIdx.y % ntx) * (mma_B::J*MMQ_TILE_Y_K);
+
+    const int   * x_qs = (const int   *) x;
+    const float * x_df = (const float *) x_qs + WARP_SIZE*2;
+    const int   * x_sc = (const int   *) x_df + WARP_SIZE/QI6_K;
+    const int   * y_qs = (const int   *) y + 4;
+    const float * y_df = (const float *) y;
+
+    const int i0 = (threadIdx.y / ntx) * (ntx*mma_A::I);
+
+    mma_A   A[ntx][8];
+    int   scA[ntx][mma_C::ne/2][8];
+    float  dA[ntx][mma_C::ne/2];
+
+#pragma unroll
+    for (int n = 0; n < ntx; ++n) {
+#pragma unroll
+        for (int k01 = 0; k01 < WARP_SIZE; k01 += 8) {
+            const int k0 = k00 + k01;
+
+            A[n][k01/4 + 0].load(x_qs + (i0 + n*mma_A::I)*MMQ_MMA_TILE_X_K_Q6_K + (k0 + 0),        MMQ_MMA_TILE_X_K_Q6_K);
+            A[n][k01/4 + 1].load(x_qs + (i0 + n*mma_A::I)*MMQ_MMA_TILE_X_K_Q6_K + (k0 + mma_A::K), MMQ_MMA_TILE_X_K_Q6_K);
+        }
+
+#pragma unroll
+        for (int k01 = 0; k01 < WARP_SIZE; k01 += 16) {
+            const int k0 = k00 + k01;
+
+#pragma unroll
+            for (int l = 0; l < mma_C::ne/2; ++l) {
+                const int i = i0 + n*mma_C::I + mma_C::get_i(2*l);
+
+                const int      sc_packed = x_sc[i*MMQ_MMA_TILE_X_K_Q6_K + k0/16];
+                const int8_t * sc        = (const int8_t *) &sc_packed;
+
+#pragma unroll
+                for (int ksc = 0; ksc < sizeof(int); ++ksc) {
+                    scA[n][l][k01/4 + ksc] = sc[ksc];
+                }
+            }
+        }
+
+#pragma unroll
+        for (int l = 0; l < mma_C::ne/2; ++l) {
+            const int i = i0 + n*mma_C::I + mma_C::get_i(2*l);
+
+            dA[n][l] = x_df[i*MMQ_MMA_TILE_X_K_Q6_K];
+        }
+    }
+
+#pragma unroll
+    for (int j0 = 0; j0 < mmq_x; j0 += ntx*mma_C::J) {
+        float tmp[ntx][mma_C::ne] = {{0.0f}};
+
+#pragma unroll
+        for (int k01 = 0; k01 < WARP_SIZE; k01 += 8) {
+            mma_B B[2];
+            float dB[mma_C::ne/2];
+
+            B[0].load(y_qs + j0*MMQ_TILE_Y_K + 0        + k01, MMQ_TILE_Y_K);
+            B[1].load(y_qs + j0*MMQ_TILE_Y_K + mma_B::K + k01, MMQ_TILE_Y_K);
+
+#pragma unroll
+            for (int l = 0; l < mma_C::ne/2; ++l) {
+                const int j = j0 + mma_C::get_j(l);
+
+                dB[l] = y_df[j*MMQ_TILE_Y_K + k01/QI8_1];
+            }
+
+#pragma unroll
+            for (int n = 0; n < ntx; ++n) {
+                mma_C C[2];
+                C[0].mma_K4(A[n][k01/4 + 0], B[0]);
+                C[1].mma_K4(A[n][k01/4 + 1], B[1]);
+
+#pragma unroll
+                for (int l = 0; l < mma_C::ne; ++l) {
+                    tmp[n][l] += (C[0].x[l]*scA[n][l/2][k01/4 + 0] + C[1].x[l]*scA[n][l/2][k01/4 + 1])*dB[l%2];
+                }
+            }
+        }
+
+#pragma unroll
+        for (int n = 0; n < ntx; ++n) {
+#pragma unroll
+            for (int l = 0; l < mma_C::ne; ++l) {
+                sum[(j0/mma_C::J + n)*mma_C::ne + l] += tmp[n][l]*dA[n][l/2];
+            }
+        }
+    }
+#else
+    GGML_UNUSED(x); GGML_UNUSED(y); GGML_UNUSED(sum);
+    NO_DEVICE_CODE;
+#endif // INT8_MMA_AVAILABLE
+}
+
+template <int mmq_y, int nwarps, bool need_check> static __device__ __forceinline__ void load_tiles_iq4_nl(
+    const char * __restrict__ x, int * __restrict__ x_tile, const int & kbx0, const int & i_max, const int & stride) {
+
+#ifdef INT8_MMA_AVAILABLE
+    int   * x_qs = (int   *)  x_tile;
+    float * x_df = (float *) (x_qs + WARP_SIZE*2);
+#else
+    constexpr tile_x_sizes txs = mmq_get_dp4a_tile_x_sizes(GGML_TYPE_IQ4_NL, mmq_y);
+    int   * x_qs = (int   *)  x_tile;
+    float * x_df = (float *) (x_qs + txs.qs);
+#endif // INT8_MMA_AVAILABLE
+
+    const int kbx  = threadIdx.x / QI4_NL;
+    const int kqsx = threadIdx.x % QI4_NL;
+
+#pragma unroll
+    for (int i0 = 0; i0 < mmq_y; i0 += nwarps) {
+        int i = i0 + threadIdx.y;
+
+        if (need_check) {
+            i = min(i, i_max);
+        }
+
+        const block_iq4_nl * bxi = (const block_iq4_nl *) x + kbx0 + i*stride + kbx;
+
+        const int aux_q4 = get_int_b2(bxi->qs, kqsx);
+        const int2 v = get_int_from_table_16(aux_q4);
+        const int k0 = 8 * (threadIdx.x / 4) + threadIdx.x % 4;
+#ifdef INT8_MMA_AVAILABLE
+        x_qs[i*MMQ_MMA_TILE_X_K_Q8_0 + k0 + 0] = v.x;
+        x_qs[i*MMQ_MMA_TILE_X_K_Q8_0 + k0 + 4] = v.y;
+#else
+        x_qs[i*(2*WARP_SIZE + 1)     + k0 + 0] = v.x;
+        x_qs[i*(2*WARP_SIZE + 1)     + k0 + 4] = v.y;
+#endif // INT8_MMA_AVAILABLE
+    }
+
+    const int blocks_per_tile_x_row = WARP_SIZE / QI4_NL;
+    const int kbxd = threadIdx.x % blocks_per_tile_x_row;
+
+#pragma unroll
+    for (int i0 = 0; i0 < mmq_y; i0 += nwarps * QI4_NL) {
+        int i = i0 + threadIdx.y * QI4_NL + threadIdx.x / blocks_per_tile_x_row;
+
+        if (need_check) {
+            i = min(i, i_max);
+        }
+
+        const block_iq4_nl * bxi = (const block_iq4_nl *) x + kbx0 + i*stride + kbxd;
+
+#ifdef INT8_MMA_AVAILABLE
+        x_df[i*MMQ_MMA_TILE_X_K_Q8_0 + kbxd] = __half2float(bxi->d);
+#else
+        x_df[i*(WARP_SIZE/4) + i/4   + kbxd] = __half2float(bxi->d);
+#endif // INT8_MMA_AVAILABLE
+    }
+}
+
+template <int mmq_y, int nwarps, bool need_check> static __device__ __forceinline__ void load_tiles_iq2_xxs(
+    const char * __restrict__ x, int * __restrict__ x_tile, const int & kbx0, const int & i_max, const int & stride) {
+
+#ifdef INT8_MMA_AVAILABLE
+    int   * x_qs = (int   *)  x_tile;
+    float * x_df = (float *) (x_qs + WARP_SIZE*2);
+#else
+    constexpr tile_x_sizes txs = mmq_get_dp4a_tile_x_sizes(GGML_TYPE_IQ2_XXS, mmq_y);
+    int   * x_qs = (int   *)  x_tile;
+    float * x_df = (float *) (x_qs + txs.qs);
+#endif // INT8_MMA_AVAILABLE
+
+    const int kqsx = threadIdx.x % (QI2_XXS/2);
+
+#pragma unroll
+    for (int i0 = 0; i0 < mmq_y; i0 += nwarps * WARP_SIZE/(QI2_XXS/2)) {
+        int i = i0 + threadIdx.y*(2*WARP_SIZE/QI2_XXS) + threadIdx.x/(QI2_XXS/2);
+
+        if (need_check) {
+            i = min(i, i_max);
+        }
+
+        const block_iq2_xxs * bxi = (const block_iq2_xxs *) x + kbx0 + i*stride;
+
+        const int q2 = get_int_b2(bxi->qs, 2*kqsx+0);
+        const uint8_t * aux8 = (const uint8_t *) &q2;
+        const uint32_t aux32 = get_int_b2(bxi->qs, 2*kqsx+1);
+
+#pragma unroll
+        for (int l = 0; l < QR2_XXS; ++l) {
+            const int * grid_pos = (const int *) (iq2xxs_grid + aux8[l]);
+            const int signs_packed = ksigns_iq2xs[(aux32 >> (7*l)) & 0x7F];
+
+            const int signs0 = __vcmpne4(((signs_packed & 0x03) << 7) | ((signs_packed & 0x0C) << 21), 0x00000000);
+            const int grid0 = __vsub4(grid_pos[0] ^ signs0, signs0);
+
+            const int signs1 = __vcmpne4(((signs_packed & 0x30) << 3) | ((signs_packed & 0xC0) << 17), 0x00000000);
+            const int grid1 = __vsub4(grid_pos[1] ^ signs1, signs1);
+
+#ifdef INT8_MMA_AVAILABLE
+            x_qs[i*MMQ_MMA_TILE_X_K_Q8_0 + 8*kqsx + (2*l + 0)] = grid0;
+            x_qs[i*MMQ_MMA_TILE_X_K_Q8_0 + 8*kqsx + (2*l + 1)] = grid1;
+#else
+            x_qs[i*(2*WARP_SIZE + 1)     + 8*kqsx + (2*l + 0)] = grid0;
+            x_qs[i*(2*WARP_SIZE + 1)     + 8*kqsx + (2*l + 1)] = grid1;
+#endif // INT8_MMA_AVAILABLE
+        }
+
+        const int ls = aux32 >> 28;
+        const float d = bxi->d;
+#ifdef INT8_MMA_AVAILABLE
+        x_df[i*MMQ_MMA_TILE_X_K_Q8_0 + kqsx] = (ls*d + d/2)/4;
+#else
+        x_df[i*(WARP_SIZE/4) + i/4   + kqsx] = (ls*d + d/2)/4;
+#endif // INT8_MMA_AVAILABLE
+    }
+}
+
+template <int mmq_y, int nwarps, bool need_check> static __device__ __forceinline__ void load_tiles_iq2_xs(
+    const char * __restrict__ x, int * __restrict__ x_tile, const int & kbx0, const int & i_max, const int & stride) {
+
+#ifdef INT8_MMA_AVAILABLE
+    int   * x_qs = (int   *)  x_tile;
+    float * x_df = (float *) (x_qs + WARP_SIZE*2);
+#else
+    constexpr tile_x_sizes txs = MMQ_DP4A_TXS_Q8_0_16;
+    int   * x_qs = (int   *)  x_tile;
+    float * x_df = (float *) (x_qs + txs.qs);
+#endif // INT8_MMA_AVAILABLE
+
+    const int kqsx = threadIdx.x % (QI2_XS/2);
+
+#pragma unroll
+    for (int i0 = 0; i0 < mmq_y; i0 += nwarps * WARP_SIZE/(QI2_XS/2)) {
+        int i = i0 + threadIdx.y*(2*WARP_SIZE/QI2_XS) + threadIdx.x/(QI2_XS/2);
+
+        if (need_check) {
+            i = min(i, i_max);
+        }
+
+        const block_iq2_xs * bxi = (const block_iq2_xs *) x + kbx0 + i*stride;
+
+        const int2 q2_packed = make_int2(get_int_b2(bxi->qs, 2*kqsx+0), get_int_b2(bxi->qs, 2*kqsx+1));
+        const uint16_t * q2 = (const uint16_t *) &q2_packed;
+
+    #pragma unroll
+        for (int l = 0; l < QR2_XS; ++l) {
+            const uint32_t * grid_pos = (const uint32_t *)(iq2xs_grid + (q2[l] & 0x000001FF));
+            const uint32_t * signs    = (const uint32_t *)(ksigns64   + (q2[l] >> 9));
+
+            const int grid_l = __vsub4(grid_pos[0] ^ signs[0], signs[0]);
+            const int grid_h = __vsub4(grid_pos[1] ^ signs[1], signs[1]);
+
+#ifdef INT8_MMA_AVAILABLE
+            x_qs[i*MMQ_MMA_TILE_X_K_Q3_K + 8*kqsx + (2*l + 0)] = grid_l;
+            x_qs[i*MMQ_MMA_TILE_X_K_Q3_K + 8*kqsx + (2*l + 1)] = grid_h;
+#else
+            x_qs[i*(2*WARP_SIZE + 1)     + 8*kqsx + (2*l + 0)] = grid_l;
+            x_qs[i*(2*WARP_SIZE + 1)     + 8*kqsx + (2*l + 1)] = grid_h;
+#endif // INT8_MMA_AVAILABLE
+        }
+
+        const int ls = bxi->scales[kqsx];
+        const float d = bxi->d;
+#ifdef INT8_MMA_AVAILABLE
+        x_df[i*MMQ_MMA_TILE_X_K_Q3_K               + 2*kqsx+0] = ((ls &  0x0F)*d + d/2)/4;
+        x_df[i*MMQ_MMA_TILE_X_K_Q3_K               + 2*kqsx+1] = ((ls >>    4)*d + d/2)/4;
+#else
+        x_df[i*(2*WARP_SIZE*2/QI8_0) + i/(QI8_0/4) + 2*kqsx+0] = ((ls &  0x0F)*d + d/2)/4;
+        x_df[i*(2*WARP_SIZE*2/QI8_0) + i/(QI8_0/4) + 2*kqsx+1] = ((ls >>    4)*d + d/2)/4;
+#endif // INT8_MMA_AVAILABLE
+    }
+}
+
+template <int mmq_y, int nwarps, bool need_check> static __device__ __forceinline__ void load_tiles_iq2_s(
+    const char * __restrict__ x, int * __restrict__ x_tile, const int & kbx0, const int & i_max, const int & stride) {
+
+#ifdef INT8_MMA_AVAILABLE
+    int   * x_qs = (int   *)  x_tile;
+    float * x_df = (float *) (x_qs + WARP_SIZE*2);
+#else
+    constexpr tile_x_sizes txs = mmq_get_dp4a_tile_x_sizes(GGML_TYPE_IQ2_S, mmq_y);
+    int   * x_qs = (int   *)  x_tile;
+    float * x_df = (float *) (x_qs + txs.qs);
+#endif // INT8_MMA_AVAILABLE
+
+    const int kqsx = threadIdx.x % (QI2_S/2);
+
+#pragma unroll
+    for (int i0 = 0; i0 < mmq_y; i0 += nwarps * WARP_SIZE/(QI2_S/2)) {
+        int i = i0 + threadIdx.y*(2*WARP_SIZE/QI2_S) + threadIdx.x/(QI2_S/2);
+
+        if (need_check) {
+            i = min(i, i_max);
+        }
+
+        const block_iq2_s * bxi = (const block_iq2_s *) x + kbx0 + i*stride;
+
+        const int       qs_packed = get_int_b2(bxi->qs, kqsx);
+        const uint8_t * qs        = (const uint8_t *) &qs_packed;
+
+        const int qh = bxi->qh[kqsx];
+
+        const int       signs_packed_32 = get_int_b2(bxi->qs, QK_K/32 + kqsx);
+        const uint8_t * signs_packed_8  = (const uint8_t *) &signs_packed_32;
+
+#pragma unroll
+        for (int l = 0; l < QR2_S; ++l) {
+            const int * grid_pos = (const int *)(iq2s_grid + (qs[l] | ((qh << (8-2*l)) & 0x300)));
+
+            const int signs0 = __vcmpne4(((signs_packed_8[l] & 0x03) << 7) | ((signs_packed_8[l] & 0x0C) << 21), 0x00000000);
+            const int signs1 = __vcmpne4(((signs_packed_8[l] & 0x30) << 3) | ((signs_packed_8[l] & 0xC0) << 17), 0x00000000);
+
+            const int grid_l = __vsub4(grid_pos[0] ^ signs0, signs0);
+            const int grid_h = __vsub4(grid_pos[1] ^ signs1, signs1);
+
+#ifdef INT8_MMA_AVAILABLE
+            x_qs[i*MMQ_MMA_TILE_X_K_Q3_K + 8*kqsx + (2*l + 0)] = grid_l;
+            x_qs[i*MMQ_MMA_TILE_X_K_Q3_K + 8*kqsx + (2*l + 1)] = grid_h;
+#else
+            x_qs[i*(2*WARP_SIZE + 1)     + 8*kqsx + (2*l + 0)] = grid_l;
+            x_qs[i*(2*WARP_SIZE + 1)     + 8*kqsx + (2*l + 1)] = grid_h;
+#endif // INT8_MMA_AVAILABLE
+        }
+
+        const int ls = bxi->scales[kqsx];
+        const float d = bxi->d;
+#ifdef INT8_MMA_AVAILABLE
+        x_df[i*MMQ_MMA_TILE_X_K_Q3_K               + 2*kqsx+0] = ((ls &  0x0F)*d + d/2)/4;
+        x_df[i*MMQ_MMA_TILE_X_K_Q3_K               + 2*kqsx+1] = ((ls >>    4)*d + d/2)/4;
+#else
+        x_df[i*(2*WARP_SIZE*2/QI8_0) + i/(QI8_0/4) + 2*kqsx+0] = ((ls &  0x0F)*d + d/2)/4;
+        x_df[i*(2*WARP_SIZE*2/QI8_0) + i/(QI8_0/4) + 2*kqsx+1] = ((ls >>    4)*d + d/2)/4;
+#endif // INT8_MMA_AVAILABLE
+    }
+}
+
+template <int mmq_y, int nwarps, bool need_check> static __device__ __forceinline__ void load_tiles_iq3_xxs(
+    const char * __restrict__ x, int * __restrict__ x_tile, const int & kbx0, const int & i_max, const int & stride) {
+
+#ifdef INT8_MMA_AVAILABLE
+    int   * x_qs = (int   *)  x_tile;
+    float * x_df = (float *) (x_qs + WARP_SIZE*2);
+#else
+    constexpr tile_x_sizes txs = mmq_get_dp4a_tile_x_sizes(GGML_TYPE_IQ3_XXS, mmq_y);
+    int   * x_qs = (int   *)  x_tile;
+    float * x_df = (float *) (x_qs + txs.qs);
+#endif // INT8_MMA_AVAILABLE
+
+    const int kqsx = threadIdx.x % (QI3_XXS/2);
+
+#pragma unroll
+    for (int i0 = 0; i0 < mmq_y; i0 += nwarps * WARP_SIZE/(QI3_XXS/2)) {
+        int i = i0 + threadIdx.y*(2*WARP_SIZE/QI3_XXS) + threadIdx.x/(QI3_XXS/2);
+
+        if (need_check) {
+            i = min(i, i_max);
+        }
+
+        const block_iq3_xxs * bxi = (const block_iq3_xxs *) x + kbx0 + i*stride;
+
+        const int2 q3_packed = make_int2(get_int_b2(bxi->qs, 2*kqsx+0), get_int_b2(bxi->qs, 2*kqsx+1));
+        const uint8_t * q3 = (const uint8_t *) &q3_packed;
+        const uint32_t aux32 = get_int_b2(bxi->qs, QK_K/16 + kqsx);
+
+#pragma unroll
+        for (int l = 0; l < QR3_XXS; ++l) {
+            const int2 grid_pos = make_int2(iq3xxs_grid[q3[2*l+0]], iq3xxs_grid[q3[2*l+1]]);
+
+            const int * signs = (const int *)(ksigns64 + ((aux32 >> (7*l)) & 0x7F));
+
+            const int grid_l = __vsub4(grid_pos.x ^ signs[0], signs[0]);
+            const int grid_h = __vsub4(grid_pos.y ^ signs[1], signs[1]);
+
+#ifdef INT8_MMA_AVAILABLE
+            x_qs[i*MMQ_MMA_TILE_X_K_Q8_0 + 8*kqsx + (2*l + 0)] = grid_l;
+            x_qs[i*MMQ_MMA_TILE_X_K_Q8_0 + 8*kqsx + (2*l + 1)] = grid_h;
+#else
+            x_qs[i*(2*WARP_SIZE + 1)     + 8*kqsx + (2*l + 0)] = grid_l;
+            x_qs[i*(2*WARP_SIZE + 1)     + 8*kqsx + (2*l + 1)] = grid_h;
+#endif // INT8_MMA_AVAILABLE
+        }
+
+        const int ls = aux32 >> 28;
+        const float d = bxi->d;
+#ifdef INT8_MMA_AVAILABLE
+        x_df[i*MMQ_MMA_TILE_X_K_Q8_0 + kqsx] = (ls*d + d/2)/2;
+#else
+        x_df[i*(WARP_SIZE/4) + i/4   + kqsx] = (ls*d + d/2)/2;
+#endif // INT8_MMA_AVAILABLE
+    }
+}
+
+template <int mmq_y, int nwarps, bool need_check> static __device__ __forceinline__ void load_tiles_iq3_s(
+    const char * __restrict__ x, int * __restrict__ x_tile, const int & kbx0, const int & i_max, const int & stride) {
+
+#ifdef INT8_MMA_AVAILABLE
+    int   * x_qs = (int   *)  x_tile;
+    float * x_df = (float *) (x_qs + WARP_SIZE*2);
+#else
+    constexpr tile_x_sizes txs = mmq_get_dp4a_tile_x_sizes(GGML_TYPE_IQ3_S, mmq_y);
+    int   * x_qs = (int   *)  x_tile;
+    float * x_df = (float *) (x_qs + txs.qs);
+#endif // INT8_MMA_AVAILABLE
+
+    const int kqsx = threadIdx.x % (QI3_S/2);
+
+#pragma unroll
+    for (int i0 = 0; i0 < mmq_y; i0 += nwarps * WARP_SIZE/(QI3_S/2)) {
+        int i = i0 + threadIdx.y*(2*WARP_SIZE/QI3_S) + threadIdx.x/(QI3_S/2);
+
+        if (need_check) {
+            i = min(i, i_max);
+        }
+
+        const block_iq3_s * bxi = (const block_iq3_s *) x + kbx0 + i*stride;
+
+        const int2      qs_packed = make_int2(get_int_b2(bxi->qs, 2*kqsx+0), get_int_b2(bxi->qs, 2*kqsx+1));
+        const uint8_t * qs        = (const uint8_t *) &qs_packed;
+
+        const int qh = bxi->qh[kqsx];
+
+        const int       signs_packed_32 = get_int_b2(bxi->signs, kqsx);
+        const uint8_t * signs_packed_8  = (const uint8_t *) &signs_packed_32;
+
+#pragma unroll
+        for (int l = 0; l < QR3_S; ++l) {
+            const int2 grid_pos = make_int2(
+                iq3s_grid[qs[2*l+0] | ((qh << (8 - 2*l)) & 0x100)],
+                iq3s_grid[qs[2*l+1] | ((qh << (7 - 2*l)) & 0x100)]);
+
+            const int signs0 = __vcmpne4(((signs_packed_8[l] & 0x03) << 7) | ((signs_packed_8[l] & 0x0C) << 21), 0x00000000);
+            const int signs1 = __vcmpne4(((signs_packed_8[l] & 0x30) << 3) | ((signs_packed_8[l] & 0xC0) << 17), 0x00000000);
+
+            const int grid_l = __vsub4(grid_pos.x ^ signs0, signs0);
+            const int grid_h = __vsub4(grid_pos.y ^ signs1, signs1);
+
+#ifdef INT8_MMA_AVAILABLE
+            x_qs[i*MMQ_MMA_TILE_X_K_Q8_0 + 8*kqsx + (2*l+0)] = grid_l;
+            x_qs[i*MMQ_MMA_TILE_X_K_Q8_0 + 8*kqsx + (2*l+1)] = grid_h;
+#else
+            x_qs[i*(2*WARP_SIZE + 1)     + 8*kqsx + (2*l+0)] = grid_l;
+            x_qs[i*(2*WARP_SIZE + 1)     + 8*kqsx + (2*l+1)] = grid_h;
+#endif // INT8_MMA_AVAILABLE
+        }
+
+        const int ls = 1 + 2*((bxi->scales[kqsx/2] >> (((2*kqsx) << 1) & 0x04)) & 0x0F);
+        const float d = bxi->d;
+#ifdef INT8_MMA_AVAILABLE
+        x_df[i*MMQ_MMA_TILE_X_K_Q8_0 + kqsx] = ls*d;
+#else
+        x_df[i*(WARP_SIZE/4) + i/4   + kqsx] = ls*d;
+#endif // INT8_MMA_AVAILABLE
+    }
+}
+
+template <int mmq_y, int nwarps, bool need_check> static __device__ __forceinline__ void load_tiles_iq1_s(
+    const char * __restrict__ x, int * __restrict__ x_tile, const int & kbx0, const int & i_max, const int & stride) {
+
+#ifdef INT8_MMA_AVAILABLE
+    int   * x_qs = (int   *)  x_tile;
+    half2 * x_ds = (half2 *) (x_qs + WARP_SIZE*2);
+#else
+    constexpr tile_x_sizes txs = mmq_get_dp4a_tile_x_sizes(GGML_TYPE_IQ3_S, mmq_y);
+    int   * x_qs = (int   *)  x_tile;
+    half2 * x_ds = (half2 *) (x_qs + txs.qs);
+#endif // INT8_MMA_AVAILABLE
+
+    const int kqsx = threadIdx.x % QI1_S;
+
+#pragma unroll
+    for (int i0 = 0; i0 < mmq_y; i0 += nwarps * WARP_SIZE/QI1_S) {
+        int i = i0 + threadIdx.y*(WARP_SIZE/QI1_S) + threadIdx.x/QI1_S;
+
+        if (need_check) {
+            i = min(i, i_max);
+        }
+
+        const block_iq1_s * bxi = (const block_iq1_s *) x + kbx0 + i*stride;
+
+        const int       qs_packed = get_int_b2(bxi->qs, kqsx);
+        const uint8_t * qs        = (const uint8_t *) &qs_packed;
+
+        const int qh = bxi->qh[kqsx];
+
+    #pragma unroll
+        for (int l = 0; l < QR1_S/2; ++l) {
+            const int grid = iq1s_grid_gpu[qs[l] | (((qh >> (3*l)) & 0x07) << 8)];
+
+            const int grid0 = (grid >> 0) & 0x0F0F0F0F;
+            const int grid1 = (grid >> 4) & 0x0F0F0F0F;
+
+#ifdef INT8_MMA_AVAILABLE
+            x_qs[i*MMQ_MMA_TILE_X_K_Q8_1 + 8*kqsx + (2*l+0)] = grid0;
+            x_qs[i*MMQ_MMA_TILE_X_K_Q8_1 + 8*kqsx + (2*l+1)] = grid1;
+#else
+            x_qs[i*(2*WARP_SIZE + 1)     + 8*kqsx + (2*l+0)] = grid0;
+            x_qs[i*(2*WARP_SIZE + 1)     + 8*kqsx + (2*l+1)] = grid1;
+#endif // INT8_MMA_AVAILABLE
+        }
+
+        const float  d1q   = __half2float(bxi->d) * (((qh >> 11) & 0x0E) + 1);
+        const float  delta = -1.0f + IQ1S_DELTA - (qh & 0x8000) * (2.0f*IQ1S_DELTA/0x8000);
+
+#ifdef INT8_MMA_AVAILABLE
+        x_ds[i*MMQ_MMA_TILE_X_K_Q8_1 + kqsx] = make_half2(d1q, d1q*delta);
+#else
+        x_ds[i*(WARP_SIZE/4) + i/4   + kqsx] = make_half2(d1q, d1q*delta);
+#endif // INT8_MMA_AVAILABLE
+    }
+}
+
+template <int mmq_y, int nwarps, bool need_check> static __device__ __forceinline__ void load_tiles_iq4_xs(
+    const char * __restrict__ x, int * __restrict__ x_tile, const int & kbx0, const int & i_max, const int & stride) {
+
+#ifdef INT8_MMA_AVAILABLE
+    int   * x_qs = (int   *)  x_tile;
+    float * x_df = (float *) (x_qs + WARP_SIZE*2);
+#else
+    constexpr tile_x_sizes txs = mmq_get_dp4a_tile_x_sizes(GGML_TYPE_IQ4_XS, mmq_y);
+    int   * x_qs = (int   *)  x_tile;
+    float * x_df = (float *) (x_qs + txs.qs);
+#endif // INT8_MMA_AVAILABLE
+
+    const int kbx  = 0;           // threadIdx.x / QI4_XS
+    const int kqsx = threadIdx.x; // threadIdx.x % QI4_XS
+
+#pragma unroll
+    for (int i0 = 0; i0 < mmq_y; i0 += nwarps) {
+        int i = i0 + threadIdx.y;
+
+        if (need_check) {
+            i = min(i, i_max);
+        }
+
+        const block_iq4_xs * bxi = (const block_iq4_xs *) x + kbx0 + i*stride + kbx;
+
+        const int aux_q4 = get_int_b4(bxi->qs, kqsx);
+        const int2 v = get_int_from_table_16(aux_q4);
+        const int k0 = 8 * (threadIdx.x / 4) + threadIdx.x % 4;
+#ifdef INT8_MMA_AVAILABLE
+        x_qs[i*MMQ_MMA_TILE_X_K_Q8_0 + k0 + 0] = v.x;
+        x_qs[i*MMQ_MMA_TILE_X_K_Q8_0 + k0 + 4] = v.y;
+#else
+        x_qs[i*(2*WARP_SIZE + 1)     + k0 + 0] = v.x;
+        x_qs[i*(2*WARP_SIZE + 1)     + k0 + 4] = v.y;
+#endif // INT8_MMA_AVAILABLE
+    }
+
+#pragma unroll
+    for (int i0 = 0; i0 < mmq_y; i0 += nwarps * 4) {
+        int i = i0 + threadIdx.y * 4 + threadIdx.x / (WARP_SIZE/4);
+
+        if (need_check) {
+            i = min(i, i_max);
+        }
+
+        const block_iq4_xs * bxi = (const block_iq4_xs *) x + kbx0 + i*stride;
+
+        const float d = __half2float(bxi->d);
+
+        const int ls = ((bxi->scales_l[(threadIdx.x % 8)/2] >> (4*(threadIdx.x % 2))) & 0x0F)
+            | (((bxi->scales_h >> (2*(threadIdx.x % 8))) & 0x03) << 4);
+
+#ifdef INT8_MMA_AVAILABLE
+        x_df[i*MMQ_MMA_TILE_X_K_Q8_0 + threadIdx.x % 8] = d * (ls - 32);
+#else
+        x_df[i*(WARP_SIZE/4) + i/4   + threadIdx.x % 8] = d * (ls - 32);
+#endif // INT8_MMA_AVAILABLE
+    }
+}
+
+template<int mmq_x, int mmq_y, int nwarps, bool need_check>
+static __device__ __forceinline__ void mmq_write_back_dp4a(
+    const float * __restrict__ sum, float * __restrict__ dst, const int & stride, const int & i_max, const int & j_max) {
+
+#pragma unroll
+    for (int j0 = 0; j0 < mmq_x; j0 += nwarps) {
+        const int j = j0 + threadIdx.y;
+
+        if (j > j_max) {
+            return;
+        }
+
+#pragma unroll
+        for (int i0 = 0; i0 < mmq_y; i0 += WARP_SIZE) {
+            const int i = i0 + threadIdx.x;
+
+            if (need_check && i > i_max) {
+                continue;
+            }
+
+            dst[j*stride + i] = sum[(j0/nwarps) * (mmq_y/WARP_SIZE) + i0/WARP_SIZE];
+        }
+    }
+}
+
+template<int mmq_x, int mmq_y, int nwarps, bool need_check>
+static __device__ __forceinline__ void mmq_write_back_mma(
+    const float * __restrict__ sum, float * __restrict__ dst, const int & stride, const int & i_max, const int & j_max) {
+
+    typedef mma_int_C_I16J8 mma_C;
+
+    constexpr int granularity = mmq_get_granularity_device(mmq_x);
+    constexpr int rows_per_warp = 2 * granularity;
+    constexpr int ntx = rows_per_warp/mma_C::I; // Number of x minitiles per warp.
+
+    const int i0 = (threadIdx.y / ntx) * (ntx*mma_C::I);
+#ifdef INT8_MMA_AVAILABLE
+    static_assert(nwarps*mma_C::I == mmq_y, "nwarps*mma_C::I != mmq_y");
+#endif // INT8_MMA_AVAILABLE
+
+#pragma unroll
+    for (int j0 = 0; j0 < mmq_x; j0 += ntx*mma_C::J) {
+#pragma unroll
+        for (int n = 0; n < ntx; ++n) {
+#pragma unroll
+            for (int l = 0; l < mma_C::ne; ++l) {
+                const int j = j0 + (threadIdx.y % ntx) * mma_C::J + mma_C::get_j(l);
+
+                if (j > j_max) {
+                    continue;
+                }
+
+                const int i = i0 + n*mma_C::I + mma_C::get_i(l);
+
+                if (need_check && i > i_max) {
+                    continue;
+                }
+
+                dst[j*stride + i] = sum[(j0/mma_C::J + n)*mma_C::ne + l];
+            }
+        }
+    }
+}
+
+// -------------------------------------------------------------------------------------------------------------------------------------
+
+template <int mmq_x, int mmq_y, int nwarps, bool need_check, ggml_type type>
+struct mmq_type_traits;
+
+template <int mmq_x, int mmq_y, int nwarps, bool need_check>
+struct mmq_type_traits<mmq_x, mmq_y, nwarps, need_check, GGML_TYPE_Q4_0> {
+    static constexpr int              vdr          = VDR_Q4_0_Q8_1_MMQ;
+    static constexpr load_tiles_mmq_t load_tiles   = load_tiles_q4_0<mmq_y, nwarps, need_check>;
+    static constexpr vec_dot_mmq_t    vec_dot_mma  = vec_dot_q8_0_q8_1_mma<mmq_x, mmq_y, nwarps, MMQ_Q8_1_DS_LAYOUT_DS4>;
+    static constexpr vec_dot_mmq_t    vec_dot_dp4a = vec_dot_q4_0_q8_1_dp4a<mmq_x, mmq_y, nwarps>;
+};
+
+template <int mmq_x, int mmq_y, int nwarps, bool need_check>
+struct mmq_type_traits<mmq_x, mmq_y, nwarps, need_check, GGML_TYPE_Q4_1> {
+    static constexpr int              vdr          = VDR_Q4_1_Q8_1_MMQ;
+    static constexpr load_tiles_mmq_t load_tiles   = load_tiles_q4_1<mmq_y, nwarps, need_check>;
+    static constexpr vec_dot_mmq_t    vec_dot_mma  = vec_dot_q8_1_q8_1_mma<mmq_x, mmq_y, nwarps>;
+    static constexpr vec_dot_mmq_t    vec_dot_dp4a = vec_dot_q4_1_q8_1_dp4a<mmq_x, mmq_y, nwarps>;
+};
+
+template <int mmq_x, int mmq_y, int nwarps, bool need_check>
+struct mmq_type_traits<mmq_x, mmq_y, nwarps, need_check, GGML_TYPE_Q5_0> {
+    static constexpr int              vdr          = VDR_Q5_0_Q8_1_MMQ;
+    static constexpr load_tiles_mmq_t load_tiles   = load_tiles_q5_0<mmq_y, nwarps, need_check>;
+    static constexpr vec_dot_mmq_t    vec_dot_mma  = vec_dot_q8_0_q8_1_mma<mmq_x, mmq_y, nwarps, MMQ_Q8_1_DS_LAYOUT_D4>;
+    static constexpr vec_dot_mmq_t    vec_dot_dp4a = vec_dot_q8_0_q8_1_dp4a<mmq_x, mmq_y, nwarps>;
+};
+
+template <int mmq_x, int mmq_y, int nwarps, bool need_check>
+struct mmq_type_traits<mmq_x, mmq_y, nwarps, need_check, GGML_TYPE_Q5_1> {
+    static constexpr int              vdr          = VDR_Q5_1_Q8_1_MMQ;
+    static constexpr load_tiles_mmq_t load_tiles   = load_tiles_q5_1<mmq_y, nwarps, need_check>;
+    static constexpr vec_dot_mmq_t    vec_dot_mma  = vec_dot_q8_1_q8_1_mma<mmq_x, mmq_y, nwarps>;
+    static constexpr vec_dot_mmq_t    vec_dot_dp4a = vec_dot_q8_1_q8_1_dp4a<mmq_x, mmq_y, nwarps>;
+};
+
+template <int mmq_x, int mmq_y, int nwarps, bool need_check>
+struct mmq_type_traits<mmq_x, mmq_y, nwarps, need_check, GGML_TYPE_Q8_0> {
+    static constexpr int              vdr          = VDR_Q8_0_Q8_1_MMQ;
+    static constexpr load_tiles_mmq_t load_tiles   = load_tiles_q8_0<mmq_y, nwarps, need_check>;
+    static constexpr vec_dot_mmq_t    vec_dot_mma  = vec_dot_q8_0_q8_1_mma<mmq_x, mmq_y, nwarps, MMQ_Q8_1_DS_LAYOUT_D4>;
+    static constexpr vec_dot_mmq_t    vec_dot_dp4a = vec_dot_q8_0_q8_1_dp4a<mmq_x, mmq_y, nwarps>;
+};
+
+template <int mmq_x, int mmq_y, int nwarps, bool need_check>
+struct mmq_type_traits<mmq_x, mmq_y, nwarps, need_check, GGML_TYPE_Q2_K> {
+    static constexpr int              vdr          = VDR_Q2_K_Q8_1_MMQ;
+    static constexpr load_tiles_mmq_t load_tiles   = load_tiles_q2_K<mmq_y, nwarps, need_check>;
+    static constexpr vec_dot_mmq_t    vec_dot_mma  = vec_dot_q2_K_q8_1_mma<mmq_x, mmq_y, nwarps>;
+    static constexpr vec_dot_mmq_t    vec_dot_dp4a = vec_dot_q2_K_q8_1_dp4a<mmq_x, mmq_y, nwarps>;
+};
+
+template <int mmq_x, int mmq_y, int nwarps, bool need_check>
+struct mmq_type_traits<mmq_x, mmq_y, nwarps, need_check, GGML_TYPE_Q3_K> {
+    static constexpr int              vdr          = VDR_Q3_K_Q8_1_MMQ;
+    static constexpr load_tiles_mmq_t load_tiles   = load_tiles_q3_K<mmq_y, nwarps, need_check>;
+    static constexpr vec_dot_mmq_t    vec_dot_mma  = vec_dot_q8_0_16_q8_1_mma<mmq_x, mmq_y, nwarps>;
+    static constexpr vec_dot_mmq_t    vec_dot_dp4a = vec_dot_q3_K_q8_1_dp4a<mmq_x, mmq_y, nwarps>;
+};
+
+template <int mmq_x, int mmq_y, int nwarps, bool need_check>
+struct mmq_type_traits<mmq_x, mmq_y, nwarps, need_check, GGML_TYPE_Q4_K> {
+    static constexpr int              vdr          = VDR_Q4_K_Q8_1_MMQ;
+    static constexpr load_tiles_mmq_t load_tiles   = load_tiles_q4_K<mmq_y, nwarps, need_check>;
+    static constexpr vec_dot_mmq_t    vec_dot_mma  = vec_dot_q8_1_q8_1_mma<mmq_x, mmq_y, nwarps>;
+    static constexpr vec_dot_mmq_t    vec_dot_dp4a = vec_dot_q4_K_q8_1_dp4a<mmq_x, mmq_y, nwarps>;
+};
+
+template <int mmq_x, int mmq_y, int nwarps, bool need_check>
+struct mmq_type_traits<mmq_x, mmq_y, nwarps, need_check, GGML_TYPE_Q5_K> {
+    static constexpr int              vdr          = VDR_Q5_K_Q8_1_MMQ;
+    static constexpr load_tiles_mmq_t load_tiles   = load_tiles_q5_K<mmq_y, nwarps, need_check>;
+    static constexpr vec_dot_mmq_t    vec_dot_mma  = vec_dot_q8_1_q8_1_mma<mmq_x, mmq_y, nwarps>;
+    static constexpr vec_dot_mmq_t    vec_dot_dp4a = vec_dot_q5_K_q8_1_dp4a<mmq_x, mmq_y, nwarps>;
+};
+
+template <int mmq_x, int mmq_y, int nwarps, bool need_check>
+struct mmq_type_traits<mmq_x, mmq_y, nwarps, need_check, GGML_TYPE_Q6_K> {
+    static constexpr int              vdr          = VDR_Q6_K_Q8_1_MMQ;
+    static constexpr load_tiles_mmq_t load_tiles   = load_tiles_q6_K<mmq_y, nwarps, need_check>;
+    static constexpr vec_dot_mmq_t    vec_dot_mma  = vec_dot_q6_K_q8_1_mma<mmq_x, mmq_y, nwarps>;
+    static constexpr vec_dot_mmq_t    vec_dot_dp4a = vec_dot_q6_K_q8_1_dp4a<mmq_x, mmq_y, nwarps>;
+};
+
+template <int mmq_x, int mmq_y, int nwarps, bool need_check>
+struct mmq_type_traits<mmq_x, mmq_y, nwarps, need_check, GGML_TYPE_IQ2_XXS> {
+    static constexpr int              vdr          = VDR_IQ2_XXS_Q8_1_MMQ;
+    static constexpr load_tiles_mmq_t load_tiles   = load_tiles_iq2_xxs<mmq_y, nwarps, need_check>;
+    static constexpr vec_dot_mmq_t    vec_dot_mma  = vec_dot_q8_0_q8_1_mma<mmq_x, mmq_y, nwarps, MMQ_Q8_1_DS_LAYOUT_D4>;
+    static constexpr vec_dot_mmq_t    vec_dot_dp4a = vec_dot_q8_0_q8_1_dp4a<mmq_x, mmq_y, nwarps>;
+};
+
+template <int mmq_x, int mmq_y, int nwarps, bool need_check>
+struct mmq_type_traits<mmq_x, mmq_y, nwarps, need_check, GGML_TYPE_IQ2_XS> {
+    static constexpr int              vdr          = VDR_IQ2_XS_Q8_1_MMQ;
+    static constexpr load_tiles_mmq_t load_tiles   = load_tiles_iq2_xs<mmq_y, nwarps, need_check>;
+    static constexpr vec_dot_mmq_t    vec_dot_mma  = vec_dot_q8_0_16_q8_1_mma<mmq_x, mmq_y, nwarps>;
+    static constexpr vec_dot_mmq_t    vec_dot_dp4a = vec_dot_q8_0_16_q8_1_dp4a<mmq_x, mmq_y, nwarps>;
+};
+
+template <int mmq_x, int mmq_y, int nwarps, bool need_check>
+struct mmq_type_traits<mmq_x, mmq_y, nwarps, need_check, GGML_TYPE_IQ2_S> {
+    static constexpr int              vdr          = VDR_IQ2_S_Q8_1_MMQ;
+    static constexpr load_tiles_mmq_t load_tiles   = load_tiles_iq2_s<mmq_y, nwarps, need_check>;
+    static constexpr vec_dot_mmq_t    vec_dot_mma  = vec_dot_q8_0_16_q8_1_mma<mmq_x, mmq_y, nwarps>;
+    static constexpr vec_dot_mmq_t    vec_dot_dp4a = vec_dot_q8_0_16_q8_1_dp4a<mmq_x, mmq_y, nwarps>;
+};
+
+template <int mmq_x, int mmq_y, int nwarps, bool need_check>
+struct mmq_type_traits<mmq_x, mmq_y, nwarps, need_check, GGML_TYPE_IQ3_XXS> {
+    static constexpr int              vdr          = VDR_IQ3_XXS_Q8_1_MMQ;
+    static constexpr load_tiles_mmq_t load_tiles   = load_tiles_iq3_xxs<mmq_y, nwarps, need_check>;
+    static constexpr vec_dot_mmq_t    vec_dot_mma  = vec_dot_q8_0_q8_1_mma<mmq_x, mmq_y, nwarps, MMQ_Q8_1_DS_LAYOUT_D4>;
+    static constexpr vec_dot_mmq_t    vec_dot_dp4a = vec_dot_q8_0_q8_1_dp4a<mmq_x, mmq_y, nwarps>;
+};
+
+template <int mmq_x, int mmq_y, int nwarps, bool need_check>
+struct mmq_type_traits<mmq_x, mmq_y, nwarps, need_check, GGML_TYPE_IQ3_S> {
+    static constexpr int              vdr          = VDR_IQ3_S_Q8_1_MMQ;
+    static constexpr load_tiles_mmq_t load_tiles   = load_tiles_iq3_s<mmq_y, nwarps, need_check>;
+    static constexpr vec_dot_mmq_t    vec_dot_mma  = vec_dot_q8_0_q8_1_mma<mmq_x, mmq_y, nwarps, MMQ_Q8_1_DS_LAYOUT_D4>;
+    static constexpr vec_dot_mmq_t    vec_dot_dp4a = vec_dot_q8_0_q8_1_dp4a<mmq_x, mmq_y, nwarps>;
+};
+
+template <int mmq_x, int mmq_y, int nwarps, bool need_check>
+struct mmq_type_traits<mmq_x, mmq_y, nwarps, need_check, GGML_TYPE_IQ1_S> {
+    static constexpr int              vdr          = VDR_IQ1_S_Q8_1_MMQ;
+    static constexpr load_tiles_mmq_t load_tiles   = load_tiles_iq1_s<mmq_y, nwarps, need_check>;
+    static constexpr vec_dot_mmq_t    vec_dot_mma  = vec_dot_q8_1_q8_1_mma<mmq_x, mmq_y, nwarps>;
+    static constexpr vec_dot_mmq_t    vec_dot_dp4a = vec_dot_q8_1_q8_1_dp4a<mmq_x, mmq_y, nwarps>;
+};
+
+template <int mmq_x, int mmq_y, int nwarps, bool need_check>
+struct mmq_type_traits<mmq_x, mmq_y, nwarps, need_check, GGML_TYPE_IQ4_NL> {
+    static constexpr int              vdr          = VDR_IQ4_NL_Q8_1_MMQ;
+    static constexpr load_tiles_mmq_t load_tiles   = load_tiles_iq4_nl<mmq_y, nwarps, need_check>;
+    static constexpr vec_dot_mmq_t    vec_dot_mma  = vec_dot_q8_0_q8_1_mma<mmq_x, mmq_y, nwarps, MMQ_Q8_1_DS_LAYOUT_D4>;
+    static constexpr vec_dot_mmq_t    vec_dot_dp4a = vec_dot_q8_0_q8_1_dp4a<mmq_x, mmq_y, nwarps>;
+};
+
+template <int mmq_x, int mmq_y, int nwarps, bool need_check>
+struct mmq_type_traits<mmq_x, mmq_y, nwarps, need_check, GGML_TYPE_IQ4_XS> {
+    static constexpr int              vdr          = VDR_IQ4_XS_Q8_1_MMQ;
+    static constexpr load_tiles_mmq_t load_tiles   = load_tiles_iq4_xs<mmq_y, nwarps, need_check>;
+    static constexpr vec_dot_mmq_t    vec_dot_mma  = vec_dot_q8_0_q8_1_mma<mmq_x, mmq_y, nwarps, MMQ_Q8_1_DS_LAYOUT_D4>;
+    static constexpr vec_dot_mmq_t    vec_dot_dp4a = vec_dot_q8_0_q8_1_dp4a<mmq_x, mmq_y, nwarps>;
+};
+
+template <ggml_type type, int mmq_x, int nwarps, bool need_check, bool fixup>
+static __device__ void mul_mat_q_process_tile(
+    const char * __restrict__ x, const char * __restrict__ yc, float * __restrict__ dst, float * __restrict__ tmp_fixup,
+    const int & ne00, const int & ne01, const int & stride01, const int & ne10, const int & ne11, const int & stride11, const int & ne0,
+    const int & it, const int & jt, const int & kb0_start, const int & kb0_stop) {
+
+    constexpr int              qk         = ggml_cuda_type_traits<type>::qk;
+    constexpr int              mmq_y      = get_mmq_y_device();
+    constexpr load_tiles_mmq_t load_tiles = mmq_type_traits<mmq_x, mmq_y, nwarps, need_check, type>::load_tiles;
+
+    extern __shared__ char data_mul_mat_q[];
+    int * tile_y = (int *) data_mul_mat_q;
+    int * tile_x = tile_y + GGML_PAD(mmq_x*(WARP_SIZE + WARP_SIZE/QI8_1), nwarps*WARP_SIZE);
+
+#ifdef INT8_MMA_AVAILABLE
+    constexpr vec_dot_mmq_t    vec_dot    = mmq_type_traits<mmq_x, mmq_y, nwarps, need_check, type>::vec_dot_mma;
+    constexpr mmq_write_back_t write_back = mmq_write_back_mma<mmq_x, mmq_y, nwarps, need_check>;
+#else
+    constexpr vec_dot_mmq_t    vec_dot    = mmq_type_traits<mmq_x, mmq_y, nwarps, need_check, type>::vec_dot_dp4a;
+    constexpr mmq_write_back_t write_back = mmq_write_back_dp4a<mmq_x, mmq_y, nwarps, need_check>;
+#endif // INT8_MMA_AVAILABLE
+
+    constexpr int blocks_per_iter = MMQ_ITER_K / qk;
+
+    float sum[mmq_x*mmq_y / (nwarps*WARP_SIZE)] = {0.0f};
+
+    const int tile_x_max_i = ne01 - it*mmq_y - 1;
+    const int tile_y_max_j = ne11 - jt*mmq_x - 1;
+
+    const int * y = (const int *) yc + jt*(mmq_x*sizeof(block_q8_1_mmq)/sizeof(int));
+
+    for (int kb0 = kb0_start; kb0 < kb0_stop; kb0 += blocks_per_iter) {
+        load_tiles(x, tile_x, stride01*it*mmq_y + kb0, tile_x_max_i, stride01);
+
+        {
+            const int * by0 = y + stride11*(kb0*(qk*sizeof(block_q8_1_mmq) / (4*QK8_1*sizeof(int))) + 0*sizeof(block_q8_1_mmq)/sizeof(int));
+#pragma unroll
+            for (int l0 = 0; l0 < mmq_x*MMQ_TILE_Y_K; l0 += nwarps*WARP_SIZE) {
+                int l = l0 + threadIdx.y*WARP_SIZE + threadIdx.x;
+
+                tile_y[l] = by0[l];
+            }
+        }
+
+        __syncthreads();
+
+        vec_dot(tile_x, tile_y, sum, 0);
+
+        __syncthreads();
+
+        {
+            const int * by0 = y + stride11*(kb0*(qk*sizeof(block_q8_1_mmq) / (4*QK8_1*sizeof(int))) + 1*sizeof(block_q8_1_mmq)/sizeof(int));
+#pragma unroll
+            for (int l0 = 0; l0 < mmq_x*MMQ_TILE_Y_K; l0 += nwarps*WARP_SIZE) {
+                int l = l0 + threadIdx.y*WARP_SIZE + threadIdx.x;
+
+                tile_y[l] = by0[l];
+            }
+        }
+
+        __syncthreads();
+
+        vec_dot(tile_x, tile_y, sum, WARP_SIZE);
+
+        __syncthreads();
+    }
+
+    if (fixup) {
+        write_back(sum, tmp_fixup + blockIdx.x*(mmq_x*mmq_y), mmq_y, mmq_y, mmq_x);
+    } else {
+        write_back(sum, dst + jt*mmq_x*ne0 + it*mmq_y, ne0, tile_x_max_i, tile_y_max_j);
+    }
+}
+
+
+// The mul_mat_q kernel implements "stream-k" work partitioning as described in https://arxiv.org/abs/2301.03598
+
+template <ggml_type type, int mmq_x, int nwarps, bool need_check>
+#if defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)
+#if defined(RDNA3) || defined(RDNA2)
+    __launch_bounds__(WARP_SIZE*nwarps, 2)
+#endif // defined(RDNA3) || defined(RDNA2)
+#else
+#if __CUDA_ARCH__ >= CC_VOLTA
+    __launch_bounds__(WARP_SIZE*nwarps, 1)
+#else
+    __launch_bounds__(WARP_SIZE*nwarps, 2)
+#endif // __CUDA_ARCH__ >= CC_VOLTA
+#endif // defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)
+static __global__ void mul_mat_q(
+    const char * __restrict__ x, const char * __restrict__ yc, float * __restrict__ dst, float * __restrict__ tmp_fixup,
+    const int ne00, const int ne01, const int stride01, const int ne10, const int ne11, const int stride11, const int ne0) {
+
+    // Skip unused template specializations for faster compilation:
+    if (mmq_x > get_mmq_x_max_device() || mmq_x % mmq_get_granularity_device(mmq_x) != 0) {
+        NO_DEVICE_CODE;
+        return;
+    }
+
+    constexpr int qk    = ggml_cuda_type_traits<type>::qk;
+    constexpr int mmq_y = get_mmq_y_device();
+
+    // On AMD or old CUDA the performance with stream-k was worse, use conventional tiling instead:
+#if (defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)) || __CUDA_ARCH__ < CC_VOLTA
+    {
+        constexpr bool fixup = false;
+        mul_mat_q_process_tile<type, mmq_x, nwarps, need_check, fixup>
+            (x, yc, dst, tmp_fixup, ne00, ne01, stride01, ne10, ne11, stride11, ne0,
+                blockIdx.x, blockIdx.y, 0, ne00/qk);
+        return;
+    }
+#endif // (defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)) || __CUDA_ARCH__ < CC_VOLTA
+
+    const     int64_t blocks_per_ne00 = ne00 / qk;
+    constexpr int     blocks_per_iter = MMQ_ITER_K / qk;
+
+    const int ntx = (ne11 + mmq_x - 1) / mmq_x; // Number of tiles x
+    const int nty = (ne01 + mmq_y - 1) / mmq_y; // Number of tiles y
+
+    // kbc == k block continuous, current index in continuous ijk space.
+    int64_t kbc      = (int64_t) blockIdx.x     *blocks_per_ne00*ntx*nty / gridDim.x;
+    int64_t kbc_stop = (int64_t)(blockIdx.x + 1)*blocks_per_ne00*ntx*nty / gridDim.x;
+
+    kbc      -= (kbc      % blocks_per_ne00) % blocks_per_iter;
+    kbc_stop -= (kbc_stop % blocks_per_ne00) % blocks_per_iter;
+
+    // kb0 == k index when doing the matrix multiplication for an output tile.
+    int kb0_start = kbc % blocks_per_ne00;
+    int kb0_stop  = min(blocks_per_ne00, kb0_start + kbc_stop - kbc);
+    while (kbc < kbc_stop && kb0_stop == blocks_per_ne00) {
+        const int jt =  kbc /    (blocks_per_ne00*nty);                    // j index of current tile.
+        const int it = (kbc - jt*(blocks_per_ne00*nty)) / blocks_per_ne00; // i index of current tile.
+
+        constexpr bool fixup = false; // All but (potentially) the last iterations write their data to dst rather than the fixup buffer.
+        mul_mat_q_process_tile<type, mmq_x, nwarps, need_check, fixup>
+            (x, yc, dst, tmp_fixup, ne00, ne01, stride01, ne10, ne11, stride11, ne0,
+             it, jt, kb0_start, kb0_stop);
+
+        kbc += blocks_per_ne00;
+        kbc -= kbc % blocks_per_ne00;
+
+        kb0_start = 0;
+        kb0_stop  = min(blocks_per_ne00, kbc_stop - kbc);
+    }
+
+    if (kbc >= kbc_stop) {
+        return;
+    }
+
+    const int jt =  kbc /    (blocks_per_ne00*nty);
+    const int it = (kbc - jt*(blocks_per_ne00*nty)) / blocks_per_ne00;
+
+    constexpr bool fixup = true; // Last index writes it data to fixup buffer to avoid data races with other blocks.
+    mul_mat_q_process_tile<type, mmq_x, nwarps, need_check, fixup>
+        (x, yc, dst, tmp_fixup, ne00, ne01, stride01, ne10, ne11, stride11, ne0,
+            it, jt, kb0_start, kb0_stop);
+}
+
+
+template <ggml_type type, int mmq_x, int nwarps, bool need_check>
+static __global__ void mul_mat_q_stream_k_fixup(
+    float * __restrict__ dst, const float * __restrict__ tmp_last_tile, const int ne00, const int ne01, const int ne11, const int ne0, const int block_num_mmq) {
+
+    constexpr int     mmq_y           = get_mmq_y_device();
+    constexpr int     qk              = ggml_cuda_type_traits<type>::qk;
+    constexpr int     blocks_per_iter = MMQ_ITER_K / qk;
+    const     int64_t blocks_per_ne00 = ne00 / qk;
+
+    float sum[mmq_x*mmq_y / (nwarps*WARP_SIZE)] = {0.0f};
+
+    const int ntx = (ne11 + mmq_x - 1) / mmq_x;
+    const int nty = (ne01 + mmq_y - 1) / mmq_y;
+
+    bool any_fixup = false;
+
+    const int bidx_start = ((blockIdx.y*nty + blockIdx.x)     * block_num_mmq)                           / (gridDim.y*gridDim.x);
+    const int bidx_stop  = ((blockIdx.y*nty + blockIdx.x + 1) * block_num_mmq + gridDim.y*gridDim.x - 1) / (gridDim.y*gridDim.x);
+
+    int64_t kbc_0;
+    int64_t kbc_stop_0 = (int64_t) bidx_start*blocks_per_ne00*ntx*nty / block_num_mmq;
+
+    for (int bidx = bidx_start; bidx < bidx_stop; ++bidx) {
+        kbc_0 = kbc_stop_0;
+        kbc_stop_0 = (int64_t) (bidx + 1)*blocks_per_ne00*ntx*nty / block_num_mmq;
+
+        const int64_t kbc      = kbc_0      - (kbc_0      % blocks_per_ne00) % blocks_per_iter;
+        const int64_t kbc_stop = kbc_stop_0 - (kbc_stop_0 % blocks_per_ne00) % blocks_per_iter;
+
+        // Skip fixup tile if the MMQ CUDA block never wrote anything to it:
+        if (kbc == kbc_stop || kbc_stop % blocks_per_ne00 == 0) {
+            continue;
+        }
+
+        const int jt =  kbc_stop /    (blocks_per_ne00*nty);
+        const int it = (kbc_stop - jt*(blocks_per_ne00*nty)) / blocks_per_ne00;
+
+        // Skip fixup tile if it's unrelated to the output tile assigned to this CUDA block:
+        if (it != blockIdx.x || jt != blockIdx.y) {
+            continue;
+        }
+
+        any_fixup = true;
+
+#pragma unroll
+        for (int j0 = 0; j0 < mmq_x; j0 += nwarps) {
+            const int j = j0 + threadIdx.y;
+
+#pragma unroll
+            for (int i0 = 0; i0 < mmq_y; i0 += WARP_SIZE) {
+                const int i = i0 + threadIdx.x;
+
+                sum[(j0/nwarps) * (mmq_y/WARP_SIZE) + i0/WARP_SIZE] += tmp_last_tile[bidx*(mmq_x*mmq_y) + j*mmq_y + i];
+            }
+        }
+    }
+
+    if (!any_fixup) {
+        return;
+    }
+
+    dst += blockIdx.y*mmq_x*ne0 + blockIdx.x*mmq_y;
+
+    const int i_max = ne01 - blockIdx.x*mmq_y - 1;
+    const int j_max = ne11 - blockIdx.y*mmq_x - 1;
+
+#pragma unroll
+    for (int j0 = 0; j0 < mmq_x; j0 += nwarps) {
+        const int j = j0 + threadIdx.y;
+
+        if (j > j_max) {
+            return;
+        }
+
+#pragma unroll
+        for (int i0 = 0; i0 < mmq_y; i0 += WARP_SIZE) {
+            const int i = i0 + threadIdx.x;
+
+            if (need_check && i > i_max) {
+                continue;
+            }
+
+            dst[j*ne0 + i] += sum[(j0/nwarps) * (mmq_y/WARP_SIZE) + i0/WARP_SIZE];
+        }
+    }
+}
+
+struct mmq_args {
+    const char * x; const char * y; float * dst;
+    int64_t ne00; int64_t ne01; int64_t stride01;
+    int64_t ne10; int64_t ne11; int64_t stride11;
+    int64_t ne0;
+    bool use_stream_k;
+};
+
+template<ggml_type type>
+static int mmq_get_shmem(const int mmq_x, const int mmq_y, const int cc) {
+    const tile_x_sizes txs = mmq_get_dp4a_tile_x_sizes(type, mmq_y);
+    const int mmq_tile_x_k = mmq_get_mma_tile_x_k(type);
+    const int shmem_x = int8_mma_available(cc) ? mmq_y*mmq_tile_x_k*sizeof(int) : txs.qs*sizeof(int) + txs.dm*sizeof(half2) + txs.sc*sizeof(int);
+    const int shmem_y = mmq_x*sizeof(block_q8_1_mmq);
+    return shmem_x + GGML_PAD(shmem_y, MMQ_NWARPS*WARP_SIZE*sizeof(int));
+}
+
+template <ggml_type type, int mmq_x>
+static void launch_mul_mat_q(ggml_backend_cuda_context & ctx, const mmq_args & args, cudaStream_t stream) {
+    const int id = ggml_cuda_get_device();
+    const int cc = ggml_cuda_info().devices[id].cc;
+    const int nsm = ggml_cuda_info().devices[id].nsm;
+    const int mmq_y = get_mmq_y_host(cc);
+
+    const dim3 block_dims(WARP_SIZE, MMQ_NWARPS, 1);
+
+    const int shmem = mmq_get_shmem<type>(mmq_x, mmq_y, cc);
+
+#if !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__))
+    static bool shmem_limit_raised[GGML_CUDA_MAX_DEVICES] = {false};
+    if (!shmem_limit_raised[id]) {
+        CUDA_CHECK(cudaFuncSetAttribute(mul_mat_q<type, mmq_x, MMQ_NWARPS, false>, cudaFuncAttributeMaxDynamicSharedMemorySize, shmem));
+        CUDA_CHECK(cudaFuncSetAttribute(mul_mat_q<type, mmq_x, MMQ_NWARPS, true>,  cudaFuncAttributeMaxDynamicSharedMemorySize, shmem));
+        shmem_limit_raised[id] = true;
+    }
+#endif // !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__))
+
+    const int nty = (args.ne01 + mmq_y - 1) / mmq_y;
+    const int ntx = (args.ne11 + mmq_x - 1) / mmq_x;
+    const dim3 block_nums_xy_tiling(nty, ntx, 1);
+
+    if (!args.use_stream_k) {
+        if (args.ne01 % mmq_y == 0) {
+            constexpr bool need_check = false;
+            mul_mat_q<type, mmq_x, MMQ_NWARPS, need_check><<<block_nums_xy_tiling, block_dims, shmem, stream>>>
+                (args.x, args.y, args.dst, nullptr, args.ne00, args.ne01, args.stride01, args.ne10, args.ne11, args.stride11, args.ne0);
+        } else {
+            constexpr bool need_check = true;
+            mul_mat_q<type, mmq_x, MMQ_NWARPS, need_check><<<block_nums_xy_tiling, block_dims, shmem, stream>>>
+                (args.x, args.y, args.dst, nullptr, args.ne00, args.ne01, args.stride01, args.ne10, args.ne11, args.stride11, args.ne0);
+        }
+        return;
+    }
+
+    const dim3 block_nums_mmq(nsm, 1, 1);
+
+    ggml_cuda_pool & pool = ctx.pool(id);
+    ggml_cuda_pool_alloc<float> tmp_fixup(pool, block_nums_mmq.x * mmq_x*mmq_y);
+
+    if (args.ne01 % mmq_y == 0) {
+        constexpr bool need_check = false;
+
+        mul_mat_q<type, mmq_x, MMQ_NWARPS, need_check><<<block_nums_mmq, block_dims, shmem, stream>>>
+            (args.x, args.y, args.dst, tmp_fixup.ptr, args.ne00, args.ne01, args.stride01, args.ne10, args.ne11, args.stride11, args.ne0);
+
+        mul_mat_q_stream_k_fixup<type, mmq_x, MMQ_NWARPS, need_check><<<block_nums_xy_tiling, block_dims, 0, stream>>>
+            (args.dst, tmp_fixup.ptr, args.ne00, args.ne01, args.ne11, args.ne0, block_nums_mmq.x);
+    } else {
+        constexpr bool need_check = true;
+
+        mul_mat_q<type, mmq_x, MMQ_NWARPS, need_check><<<block_nums_mmq, block_dims, shmem, stream>>>
+            (args.x, args.y, args.dst, tmp_fixup.ptr, args.ne00, args.ne01, args.stride01, args.ne10, args.ne11, args.stride11, args.ne0);
+
+        mul_mat_q_stream_k_fixup<type, mmq_x, MMQ_NWARPS, need_check><<<block_nums_xy_tiling, block_dims, 0, stream>>>
+            (args.dst, tmp_fixup.ptr, args.ne00, args.ne01, args.ne11, args.ne0, block_nums_mmq.x);
+    }
+}
+
+template <ggml_type type>
+void mul_mat_q_case(ggml_backend_cuda_context & ctx, const mmq_args & args, cudaStream_t stream) {
+    const int id    = ggml_cuda_get_device();
+    const int nsm   = ggml_cuda_info().devices[id].nsm;
+    const int cc    = ggml_cuda_info().devices[id].cc;
+    const int smpbo = ggml_cuda_info().devices[id].smpbo;
+
+    const int mmq_x_max = get_mmq_x_max_host(cc);
+    const int mmq_y = get_mmq_y_host(cc);
+    const int block_num_y = (args.ne01 + mmq_y - 1) / mmq_y;
+    const bool use_stream_k = cc >= CC_VOLTA && cc < CC_OFFSET_AMD;
+
+    int mmq_x_best  = 0;
+    int nparts_best = INT_MAX;
+
+    for (int mmq_x = 8; mmq_x <= mmq_x_max && nparts_best > 1; mmq_x += 8) {
+        const int granularity = mmq_get_granularity_host(mmq_x, cc);
+
+        if (mmq_x % granularity != 0 || mmq_get_shmem<type>(mmq_x, mmq_y, cc) > smpbo) {
+            continue;
+        }
+
+        const int ntiles_x = (args.ne11 + mmq_x - 1) / mmq_x;
+        const int nwaves_xy_tiling = ntiles_x*block_num_y;
+        const int nparts = use_stream_k ? ntiles_x : nwaves_xy_tiling;
+
+        if (nparts < nparts_best) {
+            mmq_x_best  = mmq_x;
+            nparts_best = nparts;
+        }
+    }
+
+    switch (mmq_x_best) {
+        case   8:
+            launch_mul_mat_q<type,   8>(ctx, args, stream);
+            break;
+        case  16:
+            launch_mul_mat_q<type,  16>(ctx, args, stream);
+            break;
+        case  24:
+            launch_mul_mat_q<type,  24>(ctx, args, stream);
+            break;
+        case  32:
+            launch_mul_mat_q<type,  32>(ctx, args, stream);
+            break;
+        case  40:
+            launch_mul_mat_q<type,  40>(ctx, args, stream);
+            break;
+        case  48:
+            launch_mul_mat_q<type,  48>(ctx, args, stream);
+            break;
+        case  56:
+            launch_mul_mat_q<type,  56>(ctx, args, stream);
+            break;
+        case  64:
+            launch_mul_mat_q<type,  64>(ctx, args, stream);
+            break;
+        case  72:
+            launch_mul_mat_q<type,  72>(ctx, args, stream);
+            break;
+        case  80:
+            launch_mul_mat_q<type,  80>(ctx, args, stream);
+            break;
+        case  88:
+            launch_mul_mat_q<type,  88>(ctx, args, stream);
+            break;
+        case  96:
+            launch_mul_mat_q<type,  96>(ctx, args, stream);
+            break;
+        case 104:
+            launch_mul_mat_q<type, 104>(ctx, args, stream);
+            break;
+        case 112:
+            launch_mul_mat_q<type, 112>(ctx, args, stream);
+            break;
+        case 120:
+            launch_mul_mat_q<type, 120>(ctx, args, stream);
+            break;
+        case 128:
+            launch_mul_mat_q<type, 128>(ctx, args, stream);
+            break;
+        default:
+            fprintf(stderr, "mmq_x_best=%d\n", mmq_x_best);
+            GGML_ABORT("fatal error");
+            break;
+    }
+}
+
+#define DECL_MMQ_CASE(type)                                                        \
+    template void mul_mat_q_case<type>(ggml_backend_cuda_context & ctx, const mmq_args & args, cudaStream_t stream) \
+
+extern DECL_MMQ_CASE(GGML_TYPE_Q4_0);
+extern DECL_MMQ_CASE(GGML_TYPE_Q4_1);
+extern DECL_MMQ_CASE(GGML_TYPE_Q5_0);
+extern DECL_MMQ_CASE(GGML_TYPE_Q5_1);
+extern DECL_MMQ_CASE(GGML_TYPE_Q8_0);
+extern DECL_MMQ_CASE(GGML_TYPE_Q2_K);
+extern DECL_MMQ_CASE(GGML_TYPE_Q3_K);
+extern DECL_MMQ_CASE(GGML_TYPE_Q4_K);
+extern DECL_MMQ_CASE(GGML_TYPE_Q5_K);
+extern DECL_MMQ_CASE(GGML_TYPE_Q6_K);
+extern DECL_MMQ_CASE(GGML_TYPE_IQ2_XXS);
+extern DECL_MMQ_CASE(GGML_TYPE_IQ2_XS);
+extern DECL_MMQ_CASE(GGML_TYPE_IQ2_S);
+extern DECL_MMQ_CASE(GGML_TYPE_IQ3_XXS);
+extern DECL_MMQ_CASE(GGML_TYPE_IQ3_S);
+extern DECL_MMQ_CASE(GGML_TYPE_IQ1_S);
+extern DECL_MMQ_CASE(GGML_TYPE_IQ4_NL);
+extern DECL_MMQ_CASE(GGML_TYPE_IQ4_XS);
+
+// -------------------------------------------------------------------------------------------------------------------------
+
+void ggml_cuda_op_mul_mat_q(
+    ggml_backend_cuda_context & ctx,
+    const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, const char * src0_dd_i, const float * src1_ddf_i,
+    const char * src1_ddq_i, float * dst_dd_i, const int64_t row_low, const int64_t row_high, const int64_t src1_ncols,
+    const int64_t src1_padded_row_size, cudaStream_t stream);
+
+bool ggml_cuda_should_use_mmq(enum ggml_type type, int cc, int64_t ne11);
diff --git a/src/whisper_cpp/ggml/src/ggml-cuda/mmvq.cu b/src/whisper_cpp/ggml/src/ggml-cuda/mmvq.cu
new file mode 100644
index 00000000..7dbbc993
--- /dev/null
+++ b/src/whisper_cpp/ggml/src/ggml-cuda/mmvq.cu
@@ -0,0 +1,425 @@
+#include "mmvq.cuh"
+#include "vecdotq.cuh"
+
+typedef float (*vec_dot_q_cuda_t)(const void * __restrict__ vbq, const block_q8_1 * __restrict__ bq8_1, const int & kbx, const int & iqs);
+
+static constexpr __device__ vec_dot_q_cuda_t get_vec_dot_q_cuda(ggml_type type) {
+    return type == GGML_TYPE_Q4_0 ? vec_dot_q4_0_q8_1 :
+        type == GGML_TYPE_Q4_1 ? vec_dot_q4_1_q8_1 :
+        type == GGML_TYPE_Q5_0 ? vec_dot_q5_0_q8_1 :
+        type == GGML_TYPE_Q5_1 ? vec_dot_q5_1_q8_1 :
+        type == GGML_TYPE_Q8_0 ? vec_dot_q8_0_q8_1 :
+        type == GGML_TYPE_Q2_K ? vec_dot_q2_K_q8_1 :
+        type == GGML_TYPE_Q3_K ? vec_dot_q3_K_q8_1 :
+        type == GGML_TYPE_Q4_K ? vec_dot_q4_K_q8_1 :
+        type == GGML_TYPE_Q5_K ? vec_dot_q5_K_q8_1 :
+        type == GGML_TYPE_Q6_K ? vec_dot_q6_K_q8_1 :
+        type == GGML_TYPE_IQ2_XXS ? vec_dot_iq2_xxs_q8_1 :
+        type == GGML_TYPE_IQ2_XS ? vec_dot_iq2_xs_q8_1 :
+        type == GGML_TYPE_IQ2_S ? vec_dot_iq2_s_q8_1 :
+        type == GGML_TYPE_IQ3_XXS ? vec_dot_iq3_xxs_q8_1 :
+        type == GGML_TYPE_IQ1_S ? vec_dot_iq1_s_q8_1 :
+        type == GGML_TYPE_IQ1_M ? vec_dot_iq1_m_q8_1 :
+        type == GGML_TYPE_IQ4_NL ? vec_dot_iq4_nl_q8_1 :
+        type == GGML_TYPE_IQ4_XS ? vec_dot_iq4_xs_q8_1 :
+        type == GGML_TYPE_IQ3_S ? vec_dot_iq3_s_q8_1 :
+        nullptr;
+}
+
+static constexpr __device__ int get_vdr_mmvq(ggml_type type) {
+    return type == GGML_TYPE_Q4_0 ? VDR_Q4_0_Q8_1_MMVQ :
+        type == GGML_TYPE_Q4_1    ? VDR_Q4_1_Q8_1_MMVQ :
+        type == GGML_TYPE_Q5_0    ? VDR_Q5_0_Q8_1_MMVQ :
+        type == GGML_TYPE_Q5_1    ? VDR_Q5_1_Q8_1_MMVQ :
+        type == GGML_TYPE_Q8_0    ? VDR_Q8_0_Q8_1_MMVQ :
+        type == GGML_TYPE_Q2_K    ? VDR_Q2_K_Q8_1_MMVQ :
+        type == GGML_TYPE_Q3_K    ? VDR_Q3_K_Q8_1_MMVQ :
+        type == GGML_TYPE_Q4_K    ? VDR_Q4_K_Q8_1_MMVQ :
+        type == GGML_TYPE_Q5_K    ? VDR_Q5_K_Q8_1_MMVQ :
+        type == GGML_TYPE_Q6_K    ? VDR_Q6_K_Q8_1_MMVQ :
+        type == GGML_TYPE_IQ2_XXS ? VDR_IQ2_XXS_Q8_1_MMVQ :
+        type == GGML_TYPE_IQ2_XS  ? VDR_IQ2_XS_Q8_1_MMVQ :
+        type == GGML_TYPE_IQ2_S   ? VDR_IQ2_S_Q8_1_MMVQ :
+        type == GGML_TYPE_IQ3_XXS ? VDR_IQ3_XXS_Q8_1_MMVQ :
+        type == GGML_TYPE_IQ3_S   ? VDR_IQ3_S_Q8_1_MMVQ :
+        type == GGML_TYPE_IQ4_NL  ? VDR_IQ4_NL_Q8_1_MMVQ :
+        type == GGML_TYPE_IQ4_XS  ? VDR_IQ4_XS_Q8_1_MMVQ :
+        1;
+}
+
+template <ggml_type type, int ncols_y>
+#if !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__))
+// tell the compiler to use as many registers as it wants, see nwarps definition below
+__launch_bounds__((ncols_y <= 4 ? 4 : 2)*WARP_SIZE, 1)
+#endif // !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__))
+static __global__ void mul_mat_vec_q(
+    const void * __restrict__ vx, const void * __restrict__ vy, float * __restrict__ dst,
+    const int ncols_x, const int nrows_x, const int nrows_y, const int nrows_dst) {
+
+    constexpr int qk  = ggml_cuda_type_traits<type>::qk;
+    constexpr int qi  = ggml_cuda_type_traits<type>::qi;
+    constexpr int vdr = get_vdr_mmvq(type);
+
+    constexpr vec_dot_q_cuda_t vec_dot_q_cuda = get_vec_dot_q_cuda(type);
+
+#if defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__) && (defined(RDNA2) || defined(RDNA3))
+    constexpr int nwarps              = 1;
+    constexpr int rows_per_cuda_block = 1;
+#else
+    constexpr int nwarps              = ncols_y <= 4 ? 4 : 2;
+    constexpr int rows_per_cuda_block = ncols_y == 1 ? 1 : 2;
+#endif // defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__) && !defined(RDNA2) && !defined(RDNA3)
+
+    const     int tid = WARP_SIZE*threadIdx.y + threadIdx.x;
+    const     int row0 = rows_per_cuda_block*blockIdx.x;
+    const     int blocks_per_row_x = ncols_x / qk;
+    const     int blocks_per_col_y = nrows_y / QK8_1;
+    constexpr int blocks_per_iter = vdr * nwarps*WARP_SIZE / qi;
+
+// partial sum for each thread
+    float tmp[ncols_y][rows_per_cuda_block] = {0.0f};
+
+    const block_q8_1 * y = (const block_q8_1 *) vy;
+
+    for (int kbx = tid / (qi/vdr); kbx < blocks_per_row_x; kbx += blocks_per_iter) {
+        const int kby = kbx * (qk/QK8_1); // y block index that aligns with kbx
+
+        // x block quant index when casting the quants to int
+        const int kqs = vdr * (tid % (qi/vdr));
+
+#pragma unroll
+        for (int j = 0; j < ncols_y; ++j) {
+#pragma unroll
+            for (int i = 0; i < rows_per_cuda_block; ++i) {
+                tmp[j][i] += vec_dot_q_cuda(vx, &y[j*blocks_per_col_y + kby], (row0 + i)*blocks_per_row_x + kbx, kqs);
+            }
+        }
+    }
+
+    __shared__ float tmp_shared[nwarps-1 > 0 ? nwarps-1 : 1][ncols_y][rows_per_cuda_block][WARP_SIZE];
+    if (threadIdx.y > 0) {
+#pragma unroll
+        for (int j = 0; j < ncols_y; ++j) {
+#pragma unroll
+            for (int i = 0; i < rows_per_cuda_block; ++i) {
+                tmp_shared[threadIdx.y-1][j][i][threadIdx.x] = tmp[j][i];
+            }
+        }
+    }
+    __syncthreads();
+    if (threadIdx.y > 0) {
+        return;
+    }
+
+    // sum up partial sums and write back result
+#pragma unroll
+    for (int j = 0; j < ncols_y; ++j) {
+#pragma unroll
+        for (int i = 0; i < rows_per_cuda_block; ++i) {
+#pragma unroll
+            for (int l = 0; l < nwarps-1; ++l) {
+                tmp[j][i] += tmp_shared[l][j][i][threadIdx.x];
+            }
+            tmp[j][i] = warp_reduce_sum(tmp[j][i]);
+        }
+
+        if (threadIdx.x < rows_per_cuda_block && (rows_per_cuda_block == 1 || row0 + threadIdx.x < nrows_dst)) {
+            dst[j*nrows_dst + row0 + threadIdx.x] = tmp[j][threadIdx.x];
+        }
+    }
+}
+
+template <ggml_type type>
+static void mul_mat_vec_q_cuda(
+    const void * vx, const void * vy, float * dst,
+    const int ncols_x, const int nrows_x, const int nrows_y, const int ncols_y, const int nrows_dst, cudaStream_t stream) {
+
+    GGML_ASSERT(ncols_x % ggml_blck_size(type) == 0);
+    GGML_ASSERT(ncols_y <= MMVQ_MAX_BATCH_SIZE);
+
+    int id = ggml_cuda_get_device();
+
+    int64_t nwarps = 1;
+    int64_t rows_per_cuda_block = 1;
+
+    if (ggml_cuda_info().devices[id].cc < CC_RDNA2) { // NVIDIA and AMD older than RDNA2
+        switch(ncols_y) {
+            case 1:
+                nwarps = 4;
+                rows_per_cuda_block = 1;
+                break;
+            case 2:
+            case 3:
+            case 4:
+                nwarps = 4;
+                rows_per_cuda_block = 2;
+                break;
+            case 5:
+            case 6:
+            case 7:
+            case 8:
+                nwarps = 2;
+                rows_per_cuda_block = 2;
+                break;
+            default:
+                GGML_ABORT("fatal error");
+                break;
+        }
+    }
+    const int64_t nblocks = (nrows_x + rows_per_cuda_block - 1) / rows_per_cuda_block;
+    const dim3 block_nums(nblocks, 1, 1);
+    const dim3 block_dims(WARP_SIZE, nwarps, 1);
+
+    switch (ncols_y) {
+        case 1:
+            mul_mat_vec_q<type, 1><<<block_nums, block_dims, 0, stream>>>(vx, vy, dst, ncols_x, nrows_x, nrows_y, nrows_dst);
+            break;
+        case 2:
+            mul_mat_vec_q<type, 2><<<block_nums, block_dims, 0, stream>>>(vx, vy, dst, ncols_x, nrows_x, nrows_y, nrows_dst);
+            break;
+        case 3:
+            mul_mat_vec_q<type, 3><<<block_nums, block_dims, 0, stream>>>(vx, vy, dst, ncols_x, nrows_x, nrows_y, nrows_dst);
+            break;
+        case 4:
+            mul_mat_vec_q<type, 4><<<block_nums, block_dims, 0, stream>>>(vx, vy, dst, ncols_x, nrows_x, nrows_y, nrows_dst);
+            break;
+        case 5:
+            mul_mat_vec_q<type, 5><<<block_nums, block_dims, 0, stream>>>(vx, vy, dst, ncols_x, nrows_x, nrows_y, nrows_dst);
+            break;
+        case 6:
+            mul_mat_vec_q<type, 6><<<block_nums, block_dims, 0, stream>>>(vx, vy, dst, ncols_x, nrows_x, nrows_y, nrows_dst);
+            break;
+        case 7:
+            mul_mat_vec_q<type, 7><<<block_nums, block_dims, 0, stream>>>(vx, vy, dst, ncols_x, nrows_x, nrows_y, nrows_dst);
+            break;
+        case 8:
+            mul_mat_vec_q<type, 8><<<block_nums, block_dims, 0, stream>>>(vx, vy, dst, ncols_x, nrows_x, nrows_y, nrows_dst);
+            break;
+        default:
+            GGML_ABORT("fatal error");
+            break;
+    }
+}
+
+static void mul_mat_vec_q4_0_q8_1_cuda(
+    const void * vx, const void * vy, float * dst,
+    const int ncols_x, const int nrows_x, const int nrows_y, const int ncols_y, const int nrows_dst, cudaStream_t stream) {
+
+    mul_mat_vec_q_cuda<GGML_TYPE_Q4_0>(vx, vy, dst, ncols_x, nrows_x, nrows_y, ncols_y, nrows_dst, stream);
+}
+
+static void mul_mat_vec_q4_1_q8_1_cuda(
+    const void * vx, const void * vy, float * dst,
+    const int ncols_x, const int nrows_x, const int nrows_y, const int ncols_y, const int nrows_dst, cudaStream_t stream) {
+
+    mul_mat_vec_q_cuda<GGML_TYPE_Q4_1>(vx, vy, dst, ncols_x, nrows_x, nrows_y, ncols_y, nrows_dst, stream);
+}
+
+static void mul_mat_vec_q5_0_q8_1_cuda(
+    const void * vx, const void * vy, float * dst,
+    const int ncols_x, const int nrows_x, const int nrows_y, const int ncols_y, const int nrows_dst, cudaStream_t stream) {
+
+    mul_mat_vec_q_cuda<GGML_TYPE_Q5_0>(vx, vy, dst, ncols_x, nrows_x, nrows_y, ncols_y, nrows_dst, stream);
+}
+
+static void mul_mat_vec_q5_1_q8_1_cuda(
+    const void * vx, const void * vy, float * dst,
+    const int ncols_x, const int nrows_x, const int nrows_y, const int ncols_y, const int nrows_dst, cudaStream_t stream) {
+
+    mul_mat_vec_q_cuda<GGML_TYPE_Q5_1>(vx, vy, dst, ncols_x, nrows_x, nrows_y, ncols_y, nrows_dst, stream);
+}
+
+static void mul_mat_vec_q8_0_q8_1_cuda(
+    const void * vx, const void * vy, float * dst,
+    const int ncols_x, const int nrows_x, const int nrows_y, const int ncols_y, const int nrows_dst, cudaStream_t stream) {
+
+    mul_mat_vec_q_cuda<GGML_TYPE_Q8_0>(vx, vy, dst, ncols_x, nrows_x, nrows_y, ncols_y, nrows_dst, stream);
+}
+
+static void mul_mat_vec_q2_K_q8_1_cuda(
+    const void * vx, const void * vy, float * dst,
+    const int ncols_x, const int nrows_x, const int nrows_y, const int ncols_y, const int nrows_dst, cudaStream_t stream) {
+
+    mul_mat_vec_q_cuda<GGML_TYPE_Q2_K>(vx, vy, dst, ncols_x, nrows_x, nrows_y, ncols_y, nrows_dst, stream);
+}
+
+static void mul_mat_vec_q3_K_q8_1_cuda(
+    const void * vx, const void * vy, float * dst,
+    const int ncols_x, const int nrows_x, const int nrows_y, const int ncols_y, const int nrows_dst, cudaStream_t stream) {
+
+    mul_mat_vec_q_cuda<GGML_TYPE_Q3_K>(vx, vy, dst, ncols_x, nrows_x, nrows_y, ncols_y, nrows_dst, stream);
+}
+
+static void mul_mat_vec_q4_K_q8_1_cuda(
+    const void * vx, const void * vy, float * dst,
+    const int ncols_x, const int nrows_x, const int nrows_y, const int ncols_y, const int nrows_dst, cudaStream_t stream) {
+
+    mul_mat_vec_q_cuda<GGML_TYPE_Q4_K>(vx, vy, dst, ncols_x, nrows_x, nrows_y, ncols_y, nrows_dst, stream);
+}
+
+static void mul_mat_vec_q5_K_q8_1_cuda(
+    const void * vx, const void * vy, float * dst,
+    const int ncols_x, const int nrows_x, const int nrows_y, const int ncols_y, const int nrows_dst, cudaStream_t stream) {
+
+    mul_mat_vec_q_cuda<GGML_TYPE_Q5_K>(vx, vy, dst, ncols_x, nrows_x, nrows_y, ncols_y, nrows_dst, stream);
+}
+
+static void mul_mat_vec_q6_K_q8_1_cuda(
+    const void * vx, const void * vy, float * dst,
+    const int ncols_x, const int nrows_x, const int nrows_y, const int ncols_y, const int nrows_dst, cudaStream_t stream) {
+
+    mul_mat_vec_q_cuda<GGML_TYPE_Q6_K>(vx, vy, dst, ncols_x, nrows_x, nrows_y, ncols_y, nrows_dst, stream);
+}
+
+static void mul_mat_vec_iq2_xxs_q8_1_cuda(
+    const void * vx, const void * vy, float * dst,
+    const int ncols_x, const int nrows_x, const int nrows_y, const int ncols_y, const int nrows_dst, cudaStream_t stream) {
+
+    mul_mat_vec_q_cuda<GGML_TYPE_IQ2_XXS>(vx, vy, dst, ncols_x, nrows_x, nrows_y, ncols_y, nrows_dst, stream);
+}
+
+static void mul_mat_vec_iq2_xs_q8_1_cuda(
+    const void * vx, const void * vy, float * dst,
+    const int ncols_x, const int nrows_x, const int nrows_y, const int ncols_y, const int nrows_dst, cudaStream_t stream) {
+
+    mul_mat_vec_q_cuda<GGML_TYPE_IQ2_XS>(vx, vy, dst, ncols_x, nrows_x, nrows_y, ncols_y, nrows_dst, stream);
+}
+
+static void mul_mat_vec_iq2_s_q8_1_cuda(
+    const void * vx, const void * vy, float * dst,
+    const int ncols_x, const int nrows_x, const int nrows_y, const int ncols_y, const int nrows_dst, cudaStream_t stream) {
+
+    mul_mat_vec_q_cuda<GGML_TYPE_IQ2_S>(vx, vy, dst, ncols_x, nrows_x, nrows_y, ncols_y, nrows_dst, stream);
+}
+
+static void mul_mat_vec_iq3_xxs_q8_1_cuda(
+    const void * vx, const void * vy, float * dst,
+    const int ncols_x, const int nrows_x, const int nrows_y, const int ncols_y, const int nrows_dst, cudaStream_t stream) {
+
+    mul_mat_vec_q_cuda<GGML_TYPE_IQ3_XXS>(vx, vy, dst, ncols_x, nrows_x, nrows_y, ncols_y, nrows_dst, stream);
+}
+
+static void mul_mat_vec_iq1_s_q8_1_cuda(
+    const void * vx, const void * vy, float * dst,
+    const int ncols_x, const int nrows_x, const int nrows_y, const int ncols_y, const int nrows_dst, cudaStream_t stream) {
+
+    mul_mat_vec_q_cuda<GGML_TYPE_IQ1_S>(vx, vy, dst, ncols_x, nrows_x, nrows_y, ncols_y, nrows_dst, stream);
+}
+
+static void mul_mat_vec_iq1_m_q8_1_cuda(
+    const void * vx, const void * vy, float * dst,
+    const int ncols_x, const int nrows_x, const int nrows_y, const int ncols_y, const int nrows_dst, cudaStream_t stream) {
+
+    mul_mat_vec_q_cuda<GGML_TYPE_IQ1_M>(vx, vy, dst, ncols_x, nrows_x, nrows_y, ncols_y, nrows_dst, stream);
+}
+
+static void mul_mat_vec_iq4_nl_q8_1_cuda(
+    const void * vx, const void * vy, float * dst,
+    const int ncols_x, const int nrows_x, const int nrows_y, const int ncols_y, const int nrows_dst, cudaStream_t stream) {
+
+    mul_mat_vec_q_cuda<GGML_TYPE_IQ4_NL>(vx, vy, dst, ncols_x, nrows_x, nrows_y, ncols_y, nrows_dst, stream);
+}
+
+static void mul_mat_vec_iq4_xs_q8_1_cuda(
+    const void * vx, const void * vy, float * dst,
+    const int ncols_x, const int nrows_x, const int nrows_y, const int ncols_y, const int nrows_dst, cudaStream_t stream) {
+
+    mul_mat_vec_q_cuda<GGML_TYPE_IQ4_XS>(vx, vy, dst, ncols_x, nrows_x, nrows_y, ncols_y, nrows_dst, stream);
+}
+
+static void mul_mat_vec_iq3_s_q8_1_cuda(
+    const void * vx, const void * vy, float * dst,
+    const int ncols_x, const int nrows_x, const int nrows_y, const int ncols_y, const int nrows_dst, cudaStream_t stream) {
+
+    mul_mat_vec_q_cuda<GGML_TYPE_IQ3_S>(vx, vy, dst, ncols_x, nrows_x, nrows_y, ncols_y, nrows_dst, stream);
+}
+
+void ggml_cuda_op_mul_mat_vec_q(
+    ggml_backend_cuda_context & ctx,
+    const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, const char * src0_dd_i, const float * src1_ddf_i,
+    const char * src1_ddq_i, float * dst_dd_i, const int64_t row_low, const int64_t row_high, const int64_t src1_ncols,
+    const int64_t src1_padded_row_size, cudaStream_t stream) {
+
+    const int64_t ne00 = src0->ne[0];
+    const int64_t row_diff = row_high - row_low;
+
+    const int64_t ne10 = src1->ne[0];
+    GGML_ASSERT(ne10 % QK8_1 == 0);
+
+    const int64_t ne0 = dst->ne[0];
+
+    int id = ggml_cuda_get_device();
+
+    // the main device has a larger memory buffer to hold the results from all GPUs
+    // nrows_dst == nrows of the matrix that the kernel writes into
+    const int64_t nrows_dst = id == ctx.device ? ne0 : row_diff;
+
+    switch (src0->type) {
+        case GGML_TYPE_Q4_0:
+            mul_mat_vec_q4_0_q8_1_cuda(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, src1_padded_row_size, src1_ncols, nrows_dst, stream);
+            break;
+        case GGML_TYPE_Q4_1:
+            mul_mat_vec_q4_1_q8_1_cuda(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, src1_padded_row_size, src1_ncols, nrows_dst, stream);
+            break;
+        case GGML_TYPE_Q5_0:
+            mul_mat_vec_q5_0_q8_1_cuda(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, src1_padded_row_size, src1_ncols, nrows_dst, stream);
+            break;
+        case GGML_TYPE_Q5_1:
+            mul_mat_vec_q5_1_q8_1_cuda(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, src1_padded_row_size, src1_ncols, nrows_dst, stream);
+            break;
+        case GGML_TYPE_Q8_0:
+            mul_mat_vec_q8_0_q8_1_cuda(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, src1_padded_row_size, src1_ncols, nrows_dst, stream);
+            break;
+        case GGML_TYPE_Q2_K:
+            mul_mat_vec_q2_K_q8_1_cuda(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, src1_padded_row_size, src1_ncols, nrows_dst, stream);
+            break;
+        case GGML_TYPE_Q3_K:
+            mul_mat_vec_q3_K_q8_1_cuda(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, src1_padded_row_size, src1_ncols, nrows_dst, stream);
+            break;
+        case GGML_TYPE_Q4_K:
+            mul_mat_vec_q4_K_q8_1_cuda(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, src1_padded_row_size, src1_ncols, nrows_dst, stream);
+            break;
+        case GGML_TYPE_Q5_K:
+            mul_mat_vec_q5_K_q8_1_cuda(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, src1_padded_row_size, src1_ncols, nrows_dst, stream);
+            break;
+        case GGML_TYPE_Q6_K:
+            mul_mat_vec_q6_K_q8_1_cuda(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, src1_padded_row_size, src1_ncols, nrows_dst, stream);
+            break;
+        case GGML_TYPE_IQ2_XXS:
+            mul_mat_vec_iq2_xxs_q8_1_cuda(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, src1_padded_row_size, src1_ncols, nrows_dst, stream);
+            break;
+        case GGML_TYPE_IQ2_XS:
+            mul_mat_vec_iq2_xs_q8_1_cuda(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, src1_padded_row_size, src1_ncols, nrows_dst, stream);
+            break;
+        case GGML_TYPE_IQ2_S:
+            mul_mat_vec_iq2_s_q8_1_cuda(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, src1_padded_row_size, src1_ncols, nrows_dst, stream);
+            break;
+        case GGML_TYPE_IQ3_XXS:
+            mul_mat_vec_iq3_xxs_q8_1_cuda(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, src1_padded_row_size, src1_ncols, nrows_dst, stream);
+            break;
+        case GGML_TYPE_IQ1_S:
+            mul_mat_vec_iq1_s_q8_1_cuda(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, src1_padded_row_size, src1_ncols, nrows_dst, stream);
+            break;
+        case GGML_TYPE_IQ1_M:
+            mul_mat_vec_iq1_m_q8_1_cuda(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, src1_padded_row_size, src1_ncols, nrows_dst, stream);
+            break;
+        case GGML_TYPE_IQ4_NL:
+            mul_mat_vec_iq4_nl_q8_1_cuda(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, src1_padded_row_size, src1_ncols, nrows_dst, stream);
+            break;
+        case GGML_TYPE_IQ4_XS:
+            mul_mat_vec_iq4_xs_q8_1_cuda(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, src1_padded_row_size, src1_ncols, nrows_dst, stream);
+            break;
+        case GGML_TYPE_IQ3_S:
+            mul_mat_vec_iq3_s_q8_1_cuda(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, src1_padded_row_size, src1_ncols, nrows_dst, stream);
+            break;
+        default:
+            GGML_ABORT("fatal error");
+            break;
+    }
+
+    GGML_UNUSED(src1);
+    GGML_UNUSED(dst);
+    GGML_UNUSED(src1_ddf_i);
+    GGML_UNUSED(src1_ncols);
+    GGML_UNUSED(src1_padded_row_size);
+}
diff --git a/src/whisper_cpp/ggml/src/ggml-cuda/mmvq.cuh b/src/whisper_cpp/ggml/src/ggml-cuda/mmvq.cuh
new file mode 100644
index 00000000..d9e42fdd
--- /dev/null
+++ b/src/whisper_cpp/ggml/src/ggml-cuda/mmvq.cuh
@@ -0,0 +1,9 @@
+#include "common.cuh"
+
+#define MMVQ_MAX_BATCH_SIZE 8 // Max. batch size for which to use MMVQ kernels.
+
+void ggml_cuda_op_mul_mat_vec_q(
+    ggml_backend_cuda_context & ctx,
+    const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, const char * src0_dd_i, const float * src1_ddf_i,
+    const char * src1_ddq_i, float * dst_dd_i, const int64_t row_low, const int64_t row_high, const int64_t src1_ncols,
+    const int64_t src1_padded_row_size, cudaStream_t stream);
diff --git a/src/whisper_cpp/ggml/src/ggml-cuda/norm.cu b/src/whisper_cpp/ggml/src/ggml-cuda/norm.cu
new file mode 100644
index 00000000..133e219f
--- /dev/null
+++ b/src/whisper_cpp/ggml/src/ggml-cuda/norm.cu
@@ -0,0 +1,224 @@
+#include "norm.cuh"
+
+template <int block_size>
+static __global__ void norm_f32(const float * x, float * dst, const int ncols, const float eps) {
+    const int row = blockIdx.x*blockDim.y + threadIdx.y;
+    const int tid = threadIdx.x;
+
+    float2 mean_var = make_float2(0.f, 0.f);
+
+    for (int col = tid; col < ncols; col += block_size) {
+        const float xi = x[row*ncols + col];
+        mean_var.x += xi;
+        mean_var.y += xi * xi;
+    }
+
+    // sum up partial sums
+    mean_var = warp_reduce_sum(mean_var);
+    if (block_size > WARP_SIZE) {
+        __shared__ float2 s_sum[32];
+        int warp_id = threadIdx.x / WARP_SIZE;
+        int lane_id = threadIdx.x % WARP_SIZE;
+        if (lane_id == 0) {
+            s_sum[warp_id] = mean_var;
+        }
+        __syncthreads();
+        mean_var = s_sum[lane_id];
+        mean_var = warp_reduce_sum(mean_var);
+    }
+
+    const float mean = mean_var.x / ncols;
+    const float var = mean_var.y / ncols - mean * mean;
+    const float inv_std = rsqrtf(var + eps);
+
+    for (int col = tid; col < ncols; col += block_size) {
+        dst[row*ncols + col] = (x[row*ncols + col] - mean) * inv_std;
+    }
+}
+
+template <int block_size>
+static __global__ void group_norm_f32(const float * x, float * dst, const int group_size, const int ne_elements, const float eps) {
+    // blockIdx.x: num_groups idx
+    // threadIdx.x: block_size idx
+    int start = blockIdx.x * group_size;
+    int end = start + group_size;
+
+    start += threadIdx.x;
+
+    if (end >= ne_elements) {
+        end = ne_elements;
+    }
+
+    float tmp = 0.0f; // partial sum for thread in warp
+
+    for (int j = start; j < end; j += block_size) {
+        tmp += x[j];
+    }
+
+    tmp = warp_reduce_sum(tmp);
+    if (block_size > WARP_SIZE) {
+        __shared__ float s_sum[32];
+        int warp_id = threadIdx.x / WARP_SIZE;
+        int lane_id = threadIdx.x % WARP_SIZE;
+        if (lane_id == 0) {
+            s_sum[warp_id] = tmp;
+        }
+        __syncthreads();
+        tmp = s_sum[lane_id];
+        tmp = warp_reduce_sum(tmp);
+    }
+
+    float mean = tmp / group_size;
+    tmp = 0.0f;
+
+    for (int j = start; j < end; j += block_size) {
+        float xi = x[j] - mean;
+        dst[j] = xi;
+        tmp += xi * xi;
+    }
+
+    tmp = warp_reduce_sum(tmp);
+    if (block_size > WARP_SIZE) {
+        __shared__ float s_sum[32];
+        int warp_id = threadIdx.x / WARP_SIZE;
+        int lane_id = threadIdx.x % WARP_SIZE;
+        if (lane_id == 0) {
+            s_sum[warp_id] = tmp;
+        }
+        __syncthreads();
+        tmp = s_sum[lane_id];
+        tmp = warp_reduce_sum(tmp);
+    }
+
+    float variance = tmp / group_size;
+    float scale = rsqrtf(variance + eps);
+    for (int j = start; j < end; j += block_size) {
+        dst[j] *= scale;
+    }
+}
+
+template <int block_size>
+static __global__ void rms_norm_f32(const float * x, float * dst, const int ncols, const float eps) {
+    const int row = blockIdx.x*blockDim.y + threadIdx.y;
+    const int tid = threadIdx.x;
+
+    float tmp = 0.0f; // partial sum for thread in warp
+
+    for (int col = tid; col < ncols; col += block_size) {
+        const float xi = x[row*ncols + col];
+        tmp += xi * xi;
+    }
+
+    // sum up partial sums
+    tmp = warp_reduce_sum(tmp);
+    if (block_size > WARP_SIZE) {
+        __shared__ float s_sum[32];
+        int warp_id = threadIdx.x / WARP_SIZE;
+        int lane_id = threadIdx.x % WARP_SIZE;
+        if (lane_id == 0) {
+            s_sum[warp_id] = tmp;
+        }
+        __syncthreads();
+        tmp = s_sum[lane_id];
+        tmp = warp_reduce_sum(tmp);
+    }
+
+    const float mean = tmp / ncols;
+    const float scale = rsqrtf(mean + eps);
+
+    for (int col = tid; col < ncols; col += block_size) {
+        dst[row*ncols + col] = scale * x[row*ncols + col];
+    }
+}
+
+static void norm_f32_cuda(const float * x, float * dst, const int ncols, const int nrows, const float eps, cudaStream_t stream) {
+    GGML_ASSERT(ncols % WARP_SIZE == 0);
+    if (ncols < 1024) {
+        const dim3 block_dims(WARP_SIZE, 1, 1);
+        norm_f32<WARP_SIZE><<<nrows, block_dims, 0, stream>>>(x, dst, ncols, eps);
+    } else {
+        const dim3 block_dims(1024, 1, 1);
+        norm_f32<1024><<<nrows, block_dims, 0, stream>>>(x, dst, ncols, eps);
+    }
+}
+
+static void group_norm_f32_cuda(const float * x, float * dst, const int num_groups, const float eps, const int group_size, const int ne_elements, cudaStream_t stream) {
+    if (group_size < 1024) {
+        const dim3 block_dims(WARP_SIZE, 1, 1);
+        group_norm_f32<WARP_SIZE><<<num_groups, block_dims, 0, stream>>>(x, dst, group_size, ne_elements, eps);
+    } else {
+        const dim3 block_dims(1024, 1, 1);
+        group_norm_f32<1024><<<num_groups, block_dims, 0, stream>>>(x, dst, group_size, ne_elements, eps);
+    }
+}
+
+static void rms_norm_f32_cuda(const float * x, float * dst, const int ncols, const int nrows, const float eps, cudaStream_t stream) {
+    GGML_ASSERT(ncols % WARP_SIZE == 0);
+    if (ncols < 1024) {
+        const dim3 block_dims(WARP_SIZE, 1, 1);
+        rms_norm_f32<WARP_SIZE><<<nrows, block_dims, 0, stream>>>(x, dst, ncols, eps);
+    } else {
+        const dim3 block_dims(1024, 1, 1);
+        rms_norm_f32<1024><<<nrows, block_dims, 0, stream>>>(x, dst, ncols, eps);
+    }
+}
+
+void ggml_cuda_op_norm(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
+    const ggml_tensor * src0 = dst->src[0];
+    const float * src0_d = (const float *)src0->data;
+    float * dst_d = (float *)dst->data;
+    cudaStream_t stream = ctx.stream();
+
+    GGML_ASSERT(ggml_is_contiguous(src0));
+
+    GGML_ASSERT(src0->type == GGML_TYPE_F32);
+    GGML_ASSERT( dst->type == GGML_TYPE_F32);
+
+    const int64_t ne00 = src0->ne[0];
+    const int64_t nrows = ggml_nrows(src0);
+
+    float eps;
+    memcpy(&eps, dst->op_params, sizeof(float));
+
+    norm_f32_cuda(src0_d, dst_d, ne00, nrows, eps, stream);
+}
+
+void ggml_cuda_op_group_norm(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
+    const ggml_tensor * src0 = dst->src[0];
+    const float * src0_d = (const float *)src0->data;
+    float * dst_d = (float *)dst->data;
+    cudaStream_t stream = ctx.stream();
+
+    GGML_ASSERT(ggml_is_contiguous(src0));
+
+    GGML_ASSERT(src0->type == GGML_TYPE_F32);
+    GGML_ASSERT( dst->type == GGML_TYPE_F32);
+
+    int num_groups = dst->op_params[0];
+
+    float eps;
+    memcpy(&eps, dst->op_params + 1, sizeof(float));
+
+    int group_size = src0->ne[0] * src0->ne[1] * ((src0->ne[2] + num_groups - 1) / num_groups);
+    group_norm_f32_cuda(src0_d, dst_d, num_groups * src0->ne[3], eps, group_size, ggml_nelements(src0), stream);
+}
+
+void ggml_cuda_op_rms_norm(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
+    const ggml_tensor * src0 = dst->src[0];
+    const float * src0_d = (const float *)src0->data;
+    float * dst_d = (float *)dst->data;
+    cudaStream_t stream = ctx.stream();
+
+    GGML_ASSERT(ggml_is_contiguous(src0));
+
+    GGML_ASSERT(src0->type == GGML_TYPE_F32);
+    GGML_ASSERT( dst->type == GGML_TYPE_F32);
+
+    const int64_t ne00 = src0->ne[0];
+    const int64_t nrows = ggml_nrows(src0);
+
+    float eps;
+    memcpy(&eps, dst->op_params, sizeof(float));
+
+    rms_norm_f32_cuda(src0_d, dst_d, ne00, nrows, eps, stream);
+}
diff --git a/src/whisper_cpp/ggml/src/ggml-cuda/norm.cuh b/src/whisper_cpp/ggml/src/ggml-cuda/norm.cuh
new file mode 100644
index 00000000..431a8f74
--- /dev/null
+++ b/src/whisper_cpp/ggml/src/ggml-cuda/norm.cuh
@@ -0,0 +1,7 @@
+#include "common.cuh"
+
+void ggml_cuda_op_norm(ggml_backend_cuda_context & ctx, ggml_tensor * dst);
+
+void ggml_cuda_op_group_norm(ggml_backend_cuda_context & ctx, ggml_tensor * dst);
+
+void ggml_cuda_op_rms_norm(ggml_backend_cuda_context & ctx, ggml_tensor * dst);
diff --git a/src/whisper_cpp/ggml/src/ggml-cuda/opt-step-adamw.cu b/src/whisper_cpp/ggml/src/ggml-cuda/opt-step-adamw.cu
new file mode 100644
index 00000000..d6f13a9c
--- /dev/null
+++ b/src/whisper_cpp/ggml/src/ggml-cuda/opt-step-adamw.cu
@@ -0,0 +1,80 @@
+#include "opt-step-adamw.cuh"
+
+#include <cstdint>
+
+static __global__ void opt_step_adamw_f32(
+    float * __restrict__ x, const float * __restrict__ g, float * __restrict__ g_m, float * __restrict__ g_v, const int64_t k,
+    const float alpha, const float beta1, const float beta2, const float eps, const float wd,
+    const float beta1h, const float beta2h) {
+
+    const int64_t i = (int64_t) blockIdx.x*blockDim.x + threadIdx.x;
+
+    if (i >= k) {
+        return;
+    }
+
+    const float gi = g[i];
+    const float gmi = g_m[i]*beta1 +    gi*(1.0f - beta1);
+    const float gvi = g_v[i]*beta2 + gi*gi*(1.0f - beta2);
+
+    g_m[i] = gmi;
+    g_v[i] = gvi;
+
+    const float mh =       gmi*beta1h;
+    const float vh = sqrtf(gvi*beta2h) + eps;
+
+    x[i] = x[i]*(1.0f - alpha*wd) - mh/vh;
+}
+
+static void opt_step_adamw_f32_cuda(
+    float * x, const float * g, float * g_m, float * g_v, const int64_t k,
+    const float alpha, const float beta1, const float beta2, const float eps, const float wd,
+    const float beta1h, const float beta2h, cudaStream_t stream) {
+
+    const dim3 block_dims(CUDA_OPT_STEP_ADAMW_BLOCK_SIZE, 1, 1);
+    const dim3 block_nums((k + CUDA_OPT_STEP_ADAMW_BLOCK_SIZE - 1) / CUDA_OPT_STEP_ADAMW_BLOCK_SIZE, 1, 1);
+    opt_step_adamw_f32<<<block_nums, block_dims, 0, stream>>>(x, g, g_m, g_v, k, alpha, beta1, beta2, eps, wd, beta1h, beta2h);
+}
+
+void ggml_cuda_opt_step_adamw(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
+    const ggml_tensor * src0        = dst->src[0];
+    const ggml_tensor * src0_grad   = dst->src[1];
+    const ggml_tensor * src0_grad_m = dst->src[2];
+    const ggml_tensor * src0_grad_v = dst->src[3];
+
+    GGML_ASSERT(src0->type        == GGML_TYPE_F32);
+    GGML_ASSERT(src0_grad->type   == GGML_TYPE_F32);
+    GGML_ASSERT(src0_grad_m->type == GGML_TYPE_F32);
+    GGML_ASSERT(src0_grad_v->type == GGML_TYPE_F32);
+    GGML_ASSERT(ggml_is_contiguous(src0));
+    GGML_ASSERT(ggml_is_contiguous(src0_grad));
+    GGML_ASSERT(ggml_is_contiguous(src0_grad_m));
+    GGML_ASSERT(ggml_is_contiguous(src0_grad_v));
+    GGML_ASSERT(ggml_are_same_shape(src0, src0_grad));
+    GGML_ASSERT(ggml_are_same_shape(src0, src0_grad_m));
+    GGML_ASSERT(ggml_are_same_shape(src0, src0_grad_v));
+
+    float       * src0_d        = (float       *) src0->data;
+    const float * src0_grad_d   = (const float *) src0_grad->data;
+    float       * src0_grad_m_d = (float       *) src0_grad_m->data;
+    float       * src0_grad_v_d = (float       *) src0_grad_v->data;
+
+    cudaStream_t stream = ctx.stream();
+
+    const int64_t ne = ggml_nelements(src0);
+
+    int64_t iter;  memcpy(&iter,  &dst->op_params[0], sizeof(int64_t));
+    float   alpha; memcpy(&alpha, &dst->op_params[2], sizeof(float));
+    float   beta1; memcpy(&beta1, &dst->op_params[3], sizeof(float));
+    float   beta2; memcpy(&beta2, &dst->op_params[4], sizeof(float));
+    float   eps;   memcpy(&eps,   &dst->op_params[5], sizeof(float));
+    float   wd;    memcpy(&wd,    &dst->op_params[6], sizeof(float));
+
+    const float beta1h  = alpha/(1.0f - powf(beta1, iter));
+    const float beta2h  =  1.0f/(1.0f - powf(beta2, iter));
+
+    opt_step_adamw_f32_cuda(src0_d, src0_grad_d, src0_grad_m_d, src0_grad_v_d, ne, alpha, beta1, beta2, eps, wd, beta1h, beta2h, stream);
+
+    iter++;
+    memcpy(&dst->op_params[0], &iter, sizeof(int64_t));
+}
diff --git a/src/whisper_cpp/ggml/src/ggml-cuda/opt-step-adamw.cuh b/src/whisper_cpp/ggml/src/ggml-cuda/opt-step-adamw.cuh
new file mode 100644
index 00000000..58d6f6e5
--- /dev/null
+++ b/src/whisper_cpp/ggml/src/ggml-cuda/opt-step-adamw.cuh
@@ -0,0 +1,5 @@
+#include "common.cuh"
+
+#define CUDA_OPT_STEP_ADAMW_BLOCK_SIZE 256
+
+void ggml_cuda_opt_step_adamw(ggml_backend_cuda_context & ctx, ggml_tensor * dst);
diff --git a/src/whisper_cpp/ggml/src/ggml-cuda/out-prod.cu b/src/whisper_cpp/ggml/src/ggml-cuda/out-prod.cu
new file mode 100644
index 00000000..619cfdcb
--- /dev/null
+++ b/src/whisper_cpp/ggml/src/ggml-cuda/out-prod.cu
@@ -0,0 +1,51 @@
+#include "out-prod.cuh"
+
+#include <cstdint>
+
+void ggml_cuda_out_prod(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
+    const ggml_tensor * src0 = dst->src[0];
+    const ggml_tensor * src1 = dst->src[1];
+
+    GGML_TENSOR_BINARY_OP_LOCALS
+
+    GGML_ASSERT(src0->type == GGML_TYPE_F32);
+    GGML_ASSERT(src1->type == GGML_TYPE_F32);
+    GGML_ASSERT(dst->type  == GGML_TYPE_F32);
+    GGML_ASSERT(ggml_is_contiguous(src0));
+    GGML_ASSERT(ggml_is_contiguous(dst));
+
+    GGML_ASSERT(ne01 == ne11);
+    GGML_ASSERT(ne0 == ne00);
+    GGML_ASSERT(ne1 == ne10);
+
+    GGML_ASSERT(ne2 == src0->ne[2]);
+    GGML_ASSERT(ne2 == src1->ne[2]);
+    GGML_ASSERT(ne3 == src0->ne[3]);
+    GGML_ASSERT(ne3 == src1->ne[3]);
+
+    const float * src0_d = (const float *) src0->data;
+    const float * src1_d = (const float *) src1->data;
+    float       *  dst_d = (float       *)  dst->data;
+
+    cudaStream_t   stream = ctx.stream();
+    cublasHandle_t handle = ctx.cublas_handle();
+
+    const float alpha = 1.0f;
+    const float beta = 0.0f;
+
+    GGML_ASSERT(ne2 == 1);
+    GGML_ASSERT(ne3 == 1);
+    CUBLAS_CHECK(cublasSetStream(handle, stream));
+
+    const bool src1_T = ggml_is_transposed(src1);
+    const cublasOperation_t src1_cublas_op =  src1_T ? CUBLAS_OP_N : CUBLAS_OP_T;
+    const int64_t           ldb            = (src1_T ?        nb10 :        nb11) /  sizeof(float);
+    GGML_ASSERT(                             (src1_T ?        nb11 :        nb10) == sizeof(float));
+
+    CUBLAS_CHECK(
+        cublasSgemm(handle, CUBLAS_OP_N, src1_cublas_op,
+                ne0, ne1, ne01,
+                &alpha, src0_d, ne00,
+                        src1_d, ldb,
+                &beta,  dst_d,  ne0));
+}
diff --git a/src/whisper_cpp/ggml/src/ggml-cuda/out-prod.cuh b/src/whisper_cpp/ggml/src/ggml-cuda/out-prod.cuh
new file mode 100644
index 00000000..a0046f5f
--- /dev/null
+++ b/src/whisper_cpp/ggml/src/ggml-cuda/out-prod.cuh
@@ -0,0 +1,3 @@
+#include "common.cuh"
+
+void ggml_cuda_out_prod(ggml_backend_cuda_context & ctx, ggml_tensor * dst);
diff --git a/src/whisper_cpp/ggml/src/ggml-cuda/pad.cu b/src/whisper_cpp/ggml/src/ggml-cuda/pad.cu
new file mode 100644
index 00000000..aba539e8
--- /dev/null
+++ b/src/whisper_cpp/ggml/src/ggml-cuda/pad.cu
@@ -0,0 +1,49 @@
+#include "pad.cuh"
+
+static __global__ void pad_f32(const float * x, float * dst, const int ne0, const int ne00, const int ne01, const int ne02, const int ne03) {
+    // blockIdx.z: idx of ne2*ne3, aka ne02*ne03
+    // blockIdx.y: idx of ne1
+    // blockIDx.x: idx of ne0 / BLOCK_SIZE
+    int nidx = threadIdx.x + blockIdx.x * blockDim.x;
+    if (nidx >= ne0) {
+        return;
+    }
+
+    // operation
+    int offset_dst =
+        nidx +
+        blockIdx.y * ne0 +
+        blockIdx.z * ne0 * gridDim.y;
+    if (nidx < ne00 && blockIdx.y < ne01 && blockIdx.z < ne02*ne03) {
+        int offset_src =
+            nidx +
+            blockIdx.y * ne00 +
+            blockIdx.z * ne00 * ne01;
+        dst[offset_dst] = x[offset_src];
+    } else {
+        dst[offset_dst] = 0.0f;
+    }
+}
+
+static void pad_f32_cuda(const float * x, float * dst,
+    const int ne00, const int ne01, const int ne02, const int ne03,
+    const int ne0, const int ne1, const int ne2, const int ne3, cudaStream_t stream) {
+    int num_blocks = (ne0 + CUDA_PAD_BLOCK_SIZE - 1) / CUDA_PAD_BLOCK_SIZE;
+    dim3 gridDim(num_blocks, ne1, ne2*ne3);
+    pad_f32<<<gridDim, CUDA_PAD_BLOCK_SIZE, 0, stream>>>(x, dst, ne0, ne00, ne01, ne02, ne03);
+}
+
+void ggml_cuda_op_pad(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
+    const ggml_tensor * src0 = dst->src[0];
+    const float * src0_d = (const float *)src0->data;
+    float * dst_d = (float *)dst->data;
+    cudaStream_t stream = ctx.stream();
+
+    GGML_ASSERT(src0->type == GGML_TYPE_F32);
+    GGML_ASSERT(dst->type == GGML_TYPE_F32);
+    GGML_ASSERT(src0->ne[3] == 1 && dst->ne[3] == 1); // just 3D tensors
+
+    pad_f32_cuda(src0_d, dst_d,
+        src0->ne[0], src0->ne[1], src0->ne[2], src0->ne[3],
+        dst->ne[0], dst->ne[1], dst->ne[2], dst->ne[3], stream);
+}
diff --git a/src/whisper_cpp/ggml/src/ggml-cuda/pad.cuh b/src/whisper_cpp/ggml/src/ggml-cuda/pad.cuh
new file mode 100644
index 00000000..8fd386b0
--- /dev/null
+++ b/src/whisper_cpp/ggml/src/ggml-cuda/pad.cuh
@@ -0,0 +1,5 @@
+#include "common.cuh"
+
+#define CUDA_PAD_BLOCK_SIZE 256
+
+void ggml_cuda_op_pad(ggml_backend_cuda_context & ctx, ggml_tensor * dst);
diff --git a/src/whisper_cpp/ggml/src/ggml-cuda/pool2d.cu b/src/whisper_cpp/ggml/src/ggml-cuda/pool2d.cu
new file mode 100644
index 00000000..c6d51e4d
--- /dev/null
+++ b/src/whisper_cpp/ggml/src/ggml-cuda/pool2d.cu
@@ -0,0 +1,94 @@
+#include "pool2d.cuh"
+
+template <typename Ti, typename To>
+static  __global__ void pool2d_nchw_kernel(
+        const int ih, const int iw, const int oh, const int ow,
+        const int kh, const int kw, const int sh, const int sw,
+        const int ph, const int pw, const int parallel_elements,
+        const Ti* src, To* dst, const enum ggml_op_pool op) {
+    int idx = threadIdx.x + blockIdx.x * blockDim.x;
+    if (idx >= parallel_elements) {
+        return;
+    }
+
+    const int I_HW = ih * iw;
+    const int O_HW = oh * ow;
+    const int nc = idx / O_HW;
+    const int cur_oh = idx % O_HW / ow;
+    const int cur_ow = idx % O_HW % ow;
+    const Ti* i_ptr = src + nc * I_HW;
+    To* o_ptr = dst + nc * O_HW;
+    const int start_h = cur_oh * sh - ph;
+    const int bh = max(0, start_h);
+    const int eh = min(ih, start_h + kh);
+    const int start_w = cur_ow * sw - pw;
+    const int bw = max(0, start_w);
+    const int ew = min(iw, start_w + kw);
+    const To scale = 1. / (kh * kw);
+    To res = 0;
+
+    switch (op) {
+        case GGML_OP_POOL_AVG: res = 0; break;
+        case GGML_OP_POOL_MAX: res = -FLT_MAX; break;
+        default: assert(false);
+    }
+
+    for (int i = bh; i < eh; i += 1) {
+        for (int j = bw; j < ew; j += 1) {
+#if __CUDA_ARCH__ >= 350
+            Ti cur = __ldg(i_ptr + i * iw + j);
+#else
+            Ti cur = i_ptr[i * iw + j];
+#endif
+            switch (op) {
+                case GGML_OP_POOL_AVG: res += cur * scale; break;
+                case GGML_OP_POOL_MAX: res = max(res, (To)cur); break;
+                default: assert(false);
+            }
+        }
+    }
+    o_ptr[cur_oh * ow + cur_ow] = res;
+}
+
+static void pool2d_nchw_kernel_f32_f32_cuda(
+        const int ih, const int iw, const int oh, const int ow,
+        const int kh, const int kw, const int sh, const int sw,
+        const int ph, const int pw, const int parallel_elements,
+        const float * src, float * dst, const enum ggml_op_pool op,
+        cudaStream_t stream) {
+
+    const int num_blocks = (parallel_elements + CUDA_POOL2D_BLOCK_SIZE - 1) / CUDA_POOL2D_BLOCK_SIZE;
+    dim3 block_nums(num_blocks);
+    pool2d_nchw_kernel<<<block_nums, CUDA_POOL2D_BLOCK_SIZE, 0, stream>>>(ih, iw, oh, ow, kh, kw, sh, sw, ph, pw, parallel_elements, src, dst, op);
+}
+
+void ggml_cuda_op_pool2d(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
+    const ggml_tensor * src0 = dst->src[0];
+    const float * src0_d = (const float *)src0->data;
+    float * dst_d = (float *)dst->data;
+    cudaStream_t stream = ctx.stream();
+
+    GGML_ASSERT(src0->type == GGML_TYPE_F32);
+    GGML_ASSERT( dst->type == GGML_TYPE_F32);
+
+    const int32_t * opts = (const int32_t *)dst->op_params;
+    enum ggml_op_pool op = static_cast<ggml_op_pool>(opts[0]);
+    const int k0 = opts[1];
+    const int k1 = opts[2];
+    const int s0 = opts[3];
+    const int s1 = opts[4];
+    const int p0 = opts[5];
+    const int p1 = opts[6];
+
+    const int64_t IH = src0->ne[1];
+    const int64_t IW = src0->ne[0];
+
+    const int64_t N = dst->ne[3];
+    const int64_t OC = dst->ne[2];
+    const int64_t OH = dst->ne[1];
+    const int64_t OW = dst->ne[0];
+
+    const int parallel_elements = N * OC * OH * OW;
+
+    pool2d_nchw_kernel_f32_f32_cuda(IH, IW, OH, OW, k1, k0, s1, s0, p1, p0, parallel_elements, src0_d, dst_d, op, stream);
+}
diff --git a/src/whisper_cpp/ggml/src/ggml-cuda/pool2d.cuh b/src/whisper_cpp/ggml/src/ggml-cuda/pool2d.cuh
new file mode 100644
index 00000000..7841292b
--- /dev/null
+++ b/src/whisper_cpp/ggml/src/ggml-cuda/pool2d.cuh
@@ -0,0 +1,5 @@
+#include "common.cuh"
+
+#define CUDA_POOL2D_BLOCK_SIZE 256
+
+void ggml_cuda_op_pool2d(ggml_backend_cuda_context & ctx, ggml_tensor * dst);
diff --git a/src/whisper_cpp/ggml/src/ggml-cuda/quantize.cu b/src/whisper_cpp/ggml/src/ggml-cuda/quantize.cu
new file mode 100644
index 00000000..45408ce8
--- /dev/null
+++ b/src/whisper_cpp/ggml/src/ggml-cuda/quantize.cu
@@ -0,0 +1,169 @@
+#include "quantize.cuh"
+#include <cstdint>
+
+static __global__ void quantize_q8_1(const float * __restrict__ x, void * __restrict__ vy, const int64_t kx, const int64_t kx0_padded) {
+    const int64_t ix0 = (int64_t)blockDim.x*blockIdx.x + threadIdx.x;
+
+    if (ix0 >= kx0_padded) {
+        return;
+    }
+
+    const int64_t ix1 = blockIdx.y;
+
+    const int64_t i_padded = ix1*kx0_padded + ix0;
+
+    block_q8_1 * y = (block_q8_1 *) vy;
+
+    const int64_t ib = i_padded / QK8_1; // block index
+    const int64_t iqs = i_padded % QK8_1; // quant index
+
+    const float xi = ix0 < kx ? x[ix1*kx + ix0] : 0.0f;
+    float amax = fabsf(xi);
+    float sum = xi;
+
+    amax = warp_reduce_max(amax);
+    sum = warp_reduce_sum(sum);
+
+    const float d = amax / 127;
+    const int8_t q = amax == 0.0f ? 0 : roundf(xi / d);
+
+    y[ib].qs[iqs] = q;
+
+    if (iqs > 0) {
+        return;
+    }
+
+    reinterpret_cast<half&>(y[ib].ds.x) = d;
+    reinterpret_cast<half&>(y[ib].ds.y) = sum;
+}
+
+template <mmq_q8_1_ds_layout ds_layout>
+static __global__ void quantize_mmq_q8_1(
+    const float * __restrict__ x, void * __restrict__ vy, const int64_t kx0, const int64_t kx1, const int64_t kx0_padded) {
+
+    constexpr int vals_per_scale = ds_layout == MMQ_Q8_1_DS_LAYOUT_D2S6 ? 64 : 32;
+    constexpr int vals_per_sum   = ds_layout == MMQ_Q8_1_DS_LAYOUT_D2S6 ? 16 : 32;
+
+    const int64_t ix0 = ((int64_t)blockDim.x*blockIdx.x + threadIdx.x)*4;
+
+    if (ix0 >= kx0_padded) {
+        return;
+    }
+
+    const float4 * x4 = (const float4 *) x;
+
+    const int64_t ix1 = kx1*blockIdx.z + blockIdx.y;
+
+    block_q8_1_mmq * y = (block_q8_1_mmq *) vy;
+
+    const int64_t ib0 = blockIdx.z*((int64_t)gridDim.y*gridDim.x*blockDim.x/QK8_1); // first block of channel
+    const int64_t ib  = ib0 + (ix0 / (4*QK8_1))*kx1 + blockIdx.y;                   // block index in channel
+    const int64_t iqs = ix0 % (4*QK8_1);                                            // quant index in block
+
+    // Load 4 floats per thread and calculate max. abs. value between them:
+    const float4 xi = ix0 < kx0 ? x4[(ix1*kx0 + ix0)/4] : make_float4(0.0f, 0.0f, 0.0f, 0.0f);
+    float amax = fabsf(xi.x);
+    amax = fmaxf(amax, fabsf(xi.y));
+    amax = fmaxf(amax, fabsf(xi.z));
+    amax = fmaxf(amax, fabsf(xi.w));
+
+    // Exchange max. abs. value between vals_per_scale/4 threads.
+#pragma unroll
+    for (int mask = vals_per_scale/8; mask > 0; mask >>= 1) {
+        amax = fmaxf(amax, __shfl_xor_sync(0xFFFFFFFF, amax, mask, WARP_SIZE));
+    }
+
+    float sum;
+    if (ds_layout != MMQ_Q8_1_DS_LAYOUT_D4) {
+        sum = xi.x + xi.y + xi.z + xi.w;
+
+        // Exchange calculate sum across vals_per_sum/4 threads.
+#pragma unroll
+        for (int mask = vals_per_sum/8; mask > 0; mask >>= 1) {
+            sum += __shfl_xor_sync(0xFFFFFFFF, sum, mask, WARP_SIZE);
+        }
+    }
+
+    const float d_inv = 127.0f / amax;
+    char4 q;
+    q.x = roundf(xi.x*d_inv);
+    q.y = roundf(xi.y*d_inv);
+    q.z = roundf(xi.z*d_inv);
+    q.w = roundf(xi.w*d_inv);
+
+    // Write back 4 int8 values as a single 32 bit value for better memroy bandwidth:
+    char4 * yqs4 = (char4 *) y[ib].qs;
+    yqs4[iqs/4] = q;
+
+    if (ds_layout == MMQ_Q8_1_DS_LAYOUT_D2S6) {
+        if (iqs % 16 != 0 || iqs >= 96) {
+            return;
+        }
+
+        y[ib].d2s6[2 + iqs/16] = sum;
+
+        if (iqs % 64 != 0) {
+            return;
+        }
+
+        const float d = 1.0f / d_inv;
+
+        y[ib].d2s6[iqs/64] = d;
+
+        return;
+    }
+
+    if (iqs % 32 != 0) {
+        return;
+    }
+
+    const float d = 1.0f / d_inv;
+
+    if (ds_layout == MMQ_Q8_1_DS_LAYOUT_DS4) {
+        y[ib].ds4[iqs/32] = make_half2(d, sum);
+    } else {
+        y[ib].d4[iqs/32]  = d;
+    }
+}
+
+void quantize_row_q8_1_cuda(
+    const float * x, void * vy, const int64_t kx0, const int64_t kx1, const int64_t channels,
+    const int64_t kx0_padded, const ggml_type type_x, cudaStream_t stream) {
+
+    GGML_ASSERT(kx0_padded % QK8_1 == 0);
+
+    const int64_t block_num_x = (kx0_padded + CUDA_QUANTIZE_BLOCK_SIZE - 1) / CUDA_QUANTIZE_BLOCK_SIZE;
+    const dim3 num_blocks(block_num_x, kx1*channels, 1);
+    const dim3 block_size(CUDA_QUANTIZE_BLOCK_SIZE, 1, 1);
+    quantize_q8_1<<<num_blocks, block_size, 0, stream>>>(x, vy, kx0, kx0_padded);
+
+    GGML_UNUSED(type_x);
+}
+
+void quantize_mmq_q8_1_cuda(
+    const float * x, void * vy, const int64_t kx0, const int64_t kx1, const int64_t channels,
+    const int64_t kx0_padded, const ggml_type type_x, cudaStream_t stream) {
+
+    GGML_ASSERT(kx0_padded % (4*QK8_1) == 0);
+
+    const int64_t block_num_x = (kx0_padded + 4*CUDA_QUANTIZE_BLOCK_SIZE_MMQ - 1) / (4*CUDA_QUANTIZE_BLOCK_SIZE_MMQ);
+    const dim3 num_blocks(block_num_x, kx1, channels);
+    const dim3 block_size(CUDA_QUANTIZE_BLOCK_SIZE_MMQ, 1, 1);
+    switch (mmq_get_q8_1_ds_layout(type_x)) {
+        case MMQ_Q8_1_DS_LAYOUT_D4:
+            quantize_mmq_q8_1<MMQ_Q8_1_DS_LAYOUT_D4>
+                <<<num_blocks, block_size, 0, stream>>>(x, vy, kx0, kx1, kx0_padded);
+            break;
+        case MMQ_Q8_1_DS_LAYOUT_DS4:
+            quantize_mmq_q8_1<MMQ_Q8_1_DS_LAYOUT_DS4>
+                <<<num_blocks, block_size, 0, stream>>>(x, vy, kx0, kx1, kx0_padded);
+            break;
+        case MMQ_Q8_1_DS_LAYOUT_D2S6:
+            quantize_mmq_q8_1<MMQ_Q8_1_DS_LAYOUT_D2S6>
+                <<<num_blocks, block_size, 0, stream>>>(x, vy, kx0, kx1, kx0_padded);
+            break;
+        default:
+            GGML_ABORT("fatal error");
+            break;
+    }
+}
diff --git a/src/whisper_cpp/ggml/src/ggml-cuda/quantize.cuh b/src/whisper_cpp/ggml/src/ggml-cuda/quantize.cuh
new file mode 100644
index 00000000..03bf322b
--- /dev/null
+++ b/src/whisper_cpp/ggml/src/ggml-cuda/quantize.cuh
@@ -0,0 +1,24 @@
+#pragma once
+
+#include "common.cuh"
+#include "mmq.cuh"
+
+#include <cstdint>
+
+#define CUDA_QUANTIZE_BLOCK_SIZE     256
+#define CUDA_QUANTIZE_BLOCK_SIZE_MMQ 128
+
+static_assert(MATRIX_ROW_PADDING %    CUDA_QUANTIZE_BLOCK_SIZE      == 0, "Risk of out-of-bounds access.");
+static_assert(MATRIX_ROW_PADDING % (4*CUDA_QUANTIZE_BLOCK_SIZE_MMQ) == 0, "Risk of out-of-bounds access.");
+
+typedef void (*quantize_cuda_t)(
+    const float * x, void * vy, const int64_t kx0, const int64_t kx1, const int64_t channels, const int64_t kx0_padded,
+    const ggml_type type_x, cudaStream_t stream);
+
+void quantize_row_q8_1_cuda(
+    const float * x, void * vy, const int64_t kx0, const int64_t kx1, const int64_t channels, const int64_t kx0_padded,
+    const ggml_type type_x, cudaStream_t stream);
+
+void quantize_mmq_q8_1_cuda(
+    const float * x, void * vy, const int64_t kx0, const int64_t kx1, const int64_t channels, const int64_t kx0_padded,
+    const ggml_type type_x, cudaStream_t stream);
diff --git a/src/whisper_cpp/ggml/src/ggml-cuda/rope.cu b/src/whisper_cpp/ggml/src/ggml-cuda/rope.cu
new file mode 100644
index 00000000..88f586d6
--- /dev/null
+++ b/src/whisper_cpp/ggml/src/ggml-cuda/rope.cu
@@ -0,0 +1,271 @@
+#include "rope.cuh"
+
+struct rope_corr_dims {
+    float v[2];
+};
+
+static __device__ float rope_yarn_ramp(const float low, const float high, const int i0) {
+    const float y = (i0 / 2 - low) / max(0.001f, high - low);
+    return 1.0f - min(1.0f, max(0.0f, y));
+}
+
+// YaRN algorithm based on LlamaYaRNScaledRotaryEmbedding.py from https://github.com/jquesnelle/yarn
+// MIT licensed. Copyright (c) 2023 Jeffrey Quesnelle and Bowen Peng.
+static __device__ void rope_yarn(
+    float theta_extrap, float freq_scale, rope_corr_dims corr_dims, int64_t i0, float ext_factor, float mscale,
+    float * cos_theta, float * sin_theta) {
+    // Get n-d rotational scaling corrected for extrapolation
+    float theta_interp = freq_scale * theta_extrap;
+    float theta = theta_interp;
+    if (ext_factor != 0.0f) {
+        float ramp_mix = rope_yarn_ramp(corr_dims.v[0], corr_dims.v[1], i0) * ext_factor;
+        theta = theta_interp * (1 - ramp_mix) + theta_extrap * ramp_mix;
+
+        // Get n-d magnitude scaling corrected for interpolation
+        mscale *= 1.0f + 0.1f * logf(1.0f / freq_scale);
+    }
+    *cos_theta = cosf(theta) * mscale;
+    *sin_theta = sinf(theta) * mscale;
+}
+
+template<typename T, bool has_ff>
+static __global__ void rope_norm(
+    const T * x, T * dst, int ne0, int n_dims, const int32_t * pos, float freq_scale, int p_delta_rows,
+    float ext_factor, float attn_factor, rope_corr_dims corr_dims, float theta_scale, const float * freq_factors) {
+    const int i0 = 2*(blockDim.y*blockIdx.y + threadIdx.y);
+
+    if (i0 >= ne0) {
+        return;
+    }
+
+    const int row = blockDim.x*blockIdx.x + threadIdx.x;
+
+    if (i0 >= n_dims) {
+        const int i = row*ne0 + i0;
+
+        dst[i + 0] = x[i + 0];
+        dst[i + 1] = x[i + 1];
+
+        return;
+    }
+
+    const int i  = row*ne0 + i0;
+    const int i2 = row/p_delta_rows;
+
+    const float theta_base = pos[i2]*powf(theta_scale, i0/2.0f);
+
+    const float freq_factor = has_ff ? freq_factors[i0/2] : 1.0f;
+
+    float cos_theta;
+    float sin_theta;
+
+    rope_yarn(theta_base/freq_factor, freq_scale, corr_dims, i0, ext_factor, attn_factor, &cos_theta, &sin_theta);
+
+    const float x0 = x[i + 0];
+    const float x1 = x[i + 1];
+
+    dst[i + 0] = x0*cos_theta - x1*sin_theta;
+    dst[i + 1] = x0*sin_theta + x1*cos_theta;
+}
+
+template<typename T, bool has_ff>
+static __global__ void rope_neox(
+    const T * x, T * dst, int ne0, int n_dims, const int32_t * pos, float freq_scale, int p_delta_rows,
+    float ext_factor, float attn_factor, rope_corr_dims corr_dims, float theta_scale, const float * freq_factors) {
+    const int i0 = 2*(blockDim.y*blockIdx.y + threadIdx.y);
+
+    if (i0 >= ne0) {
+        return;
+    }
+
+    const int row = blockDim.x*blockIdx.x + threadIdx.x;
+
+    if (i0 >= n_dims) {
+        const int i = row*ne0 + i0;
+
+        dst[i + 0] = x[i + 0];
+        dst[i + 1] = x[i + 1];
+
+        return;
+    }
+
+    const int i  = row*ne0 + i0/2;
+    const int i2 = row/p_delta_rows;
+
+    const float theta_base = pos[i2]*powf(theta_scale, i0/2.0f);
+
+    const float freq_factor = has_ff ? freq_factors[i0/2] : 1.0f;
+
+    float cos_theta;
+    float sin_theta;
+
+    rope_yarn(theta_base/freq_factor, freq_scale, corr_dims, i0, ext_factor, attn_factor, &cos_theta, &sin_theta);
+
+    const float x0 = x[i + 0];
+    const float x1 = x[i + n_dims/2];
+
+    dst[i + 0]        = x0*cos_theta - x1*sin_theta;
+    dst[i + n_dims/2] = x0*sin_theta + x1*cos_theta;
+}
+
+template<typename T>
+static void rope_norm_cuda(
+    const T * x, T * dst, int ne0, int n_dims, int nr, const int32_t * pos, float freq_scale, int p_delta_rows,
+    float freq_base, float ext_factor, float attn_factor, rope_corr_dims corr_dims, const float * freq_factors, cudaStream_t stream) {
+    GGML_ASSERT(ne0 % 2 == 0);
+    const dim3 block_dims(1, CUDA_ROPE_BLOCK_SIZE, 1);
+    const int n_blocks_x = (ne0 + 2*CUDA_ROPE_BLOCK_SIZE - 1) / (2*CUDA_ROPE_BLOCK_SIZE);
+    const dim3 block_nums(nr, n_blocks_x, 1);
+
+    const float theta_scale = powf(freq_base, -2.0f/n_dims);
+
+    if (freq_factors == nullptr) {
+        rope_norm<T, false><<<block_nums, block_dims, 0, stream>>>(
+                x, dst, ne0, n_dims, pos, freq_scale, p_delta_rows, ext_factor, attn_factor, corr_dims,
+                theta_scale, freq_factors
+                );
+    } else {
+        rope_norm<T, true><<<block_nums, block_dims, 0, stream>>>(
+                x, dst, ne0, n_dims, pos, freq_scale, p_delta_rows, ext_factor, attn_factor, corr_dims,
+                theta_scale, freq_factors
+                );
+    }
+}
+
+template<typename T>
+static void rope_neox_cuda(
+    const T * x, T * dst, int ne0, int n_dims, int nr, const int32_t * pos, float freq_scale, int p_delta_rows,
+    float freq_base, float ext_factor, float attn_factor, rope_corr_dims corr_dims, const float * freq_factors, cudaStream_t stream) {
+    GGML_ASSERT(ne0 % 2 == 0);
+    const dim3 block_dims(1, CUDA_ROPE_BLOCK_SIZE, 1);
+    const int n_blocks_x = (ne0 + 2*CUDA_ROPE_BLOCK_SIZE - 1) / (2*CUDA_ROPE_BLOCK_SIZE);
+    const dim3 block_nums(nr, n_blocks_x, 1);
+
+    const float theta_scale = powf(freq_base, -2.0f/n_dims);
+
+    if (freq_factors == nullptr) {
+        rope_neox<T, false><<<block_nums, block_dims, 0, stream>>>(
+                x, dst, ne0, n_dims, pos, freq_scale, p_delta_rows, ext_factor, attn_factor, corr_dims,
+                theta_scale, freq_factors
+                );
+    } else {
+        rope_neox<T, true><<<block_nums, block_dims, 0, stream>>>(
+                x, dst, ne0, n_dims, pos, freq_scale, p_delta_rows, ext_factor, attn_factor, corr_dims,
+                theta_scale, freq_factors
+                );
+    }
+}
+
+static void rope_norm_cuda_f16(
+    const half * x, half * dst, int ne0, int n_dims, int nr, const int32_t * pos, float freq_scale, int p_delta_rows,
+    float freq_base, float ext_factor, float attn_factor, rope_corr_dims corr_dims, const float * freq_factors, cudaStream_t stream) {
+
+    rope_norm_cuda<half>(x, dst, ne0, n_dims, nr, pos, freq_scale, p_delta_rows, freq_base, ext_factor, attn_factor, corr_dims, freq_factors, stream);
+}
+
+static void rope_norm_cuda_f32(
+    const float * x, float * dst, int ne0, int n_dims, int nr, const int32_t * pos, float freq_scale, int p_delta_rows,
+    float freq_base, float ext_factor, float attn_factor, rope_corr_dims corr_dims, const float * freq_factors, cudaStream_t stream) {
+
+    rope_norm_cuda<float>(x, dst, ne0, n_dims, nr, pos, freq_scale, p_delta_rows, freq_base, ext_factor, attn_factor, corr_dims, freq_factors, stream);
+}
+
+static void rope_neox_cuda_f16(
+    const half * x, half * dst, int ne0, int n_dims, int nr, const int32_t * pos, float freq_scale, int p_delta_rows,
+    float freq_base, float ext_factor, float attn_factor, rope_corr_dims corr_dims, const float * freq_factors, cudaStream_t stream) {
+
+    rope_neox_cuda<half>(x, dst, ne0, n_dims, nr, pos, freq_scale, p_delta_rows, freq_base, ext_factor, attn_factor, corr_dims, freq_factors, stream);
+}
+
+static void rope_neox_cuda_f32(
+    const float * x, float * dst, int ne0, int n_dims, int nr, const int32_t * pos, float freq_scale, int p_delta_rows,
+    float freq_base, float ext_factor, float attn_factor, rope_corr_dims corr_dims, const float * freq_factors, cudaStream_t stream
+) {
+
+    rope_neox_cuda<float>(x, dst, ne0, n_dims, nr, pos, freq_scale, p_delta_rows, freq_base, ext_factor, attn_factor, corr_dims, freq_factors, stream);
+}
+
+void ggml_cuda_op_rope(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
+    const ggml_tensor * src0 = dst->src[0];
+    const ggml_tensor * src1 = dst->src[1];
+    const ggml_tensor * src2 = dst->src[2];
+
+    const float * src0_d = (const float *)src0->data;
+    const float * src1_d = (const float *)src1->data;
+
+    float * dst_d = (float *)dst->data;
+    cudaStream_t stream = ctx.stream();
+
+    GGML_ASSERT(ggml_is_contiguous(src0));
+    GGML_ASSERT(src0->type == GGML_TYPE_F32 || src0->type == GGML_TYPE_F16);
+    GGML_ASSERT( dst->type == GGML_TYPE_F32 ||  dst->type == GGML_TYPE_F16);
+    GGML_ASSERT(src0->type == dst->type);
+
+    const int64_t ne00 = src0->ne[0];
+    const int64_t ne01 = src0->ne[1];
+    const int64_t nr = ggml_nrows(src0);
+
+    //const int n_past     = ((int32_t *) dst->op_params)[0];
+    const int n_dims     = ((int32_t *) dst->op_params)[1];
+    const int mode       = ((int32_t *) dst->op_params)[2];
+    //const int n_ctx      = ((int32_t *) dst->op_params)[3];
+    const int n_ctx_orig = ((int32_t *) dst->op_params)[4];
+
+    // RoPE alteration for extended context
+    float freq_base;
+    float freq_scale;
+    float ext_factor;
+    float attn_factor;
+    float beta_fast;
+    float beta_slow;
+
+    memcpy(&freq_base,   (int32_t *) dst->op_params +  5, sizeof(float));
+    memcpy(&freq_scale,  (int32_t *) dst->op_params +  6, sizeof(float));
+    memcpy(&ext_factor,  (int32_t *) dst->op_params +  7, sizeof(float));
+    memcpy(&attn_factor, (int32_t *) dst->op_params +  8, sizeof(float));
+    memcpy(&beta_fast,   (int32_t *) dst->op_params +  9, sizeof(float));
+    memcpy(&beta_slow,   (int32_t *) dst->op_params + 10, sizeof(float));
+
+    const bool is_neox = mode & GGML_ROPE_TYPE_NEOX;
+
+    const int32_t * pos = (const int32_t *) src1_d;
+
+    const float * freq_factors = nullptr;
+    if (src2 != nullptr) {
+        freq_factors = (const float *) src2->data;
+    }
+
+    rope_corr_dims corr_dims;
+    ggml_rope_yarn_corr_dims(n_dims, n_ctx_orig, freq_base, beta_fast, beta_slow, corr_dims.v);
+
+    // compute
+    if (is_neox) {
+        if (src0->type == GGML_TYPE_F32) {
+            rope_neox_cuda_f32(
+                (const float *)src0_d, (float *)dst_d, ne00, n_dims, nr, pos, freq_scale, ne01, freq_base, ext_factor,
+                attn_factor, corr_dims, freq_factors, stream
+            );
+        } else if (src0->type == GGML_TYPE_F16) {
+            rope_neox_cuda_f16(
+                (const half *)src0_d, (half *)dst_d, ne00, n_dims, nr, pos, freq_scale, ne01, freq_base, ext_factor,
+                attn_factor, corr_dims, freq_factors, stream
+            );
+        } else {
+            GGML_ABORT("fatal error");
+        }
+    } else {
+        if (src0->type == GGML_TYPE_F32) {
+            rope_norm_cuda_f32(
+                (const float *)src0_d, (float *)dst_d, ne00, n_dims, nr, pos, freq_scale, ne01, freq_base, ext_factor,
+                attn_factor, corr_dims, freq_factors, stream
+            );
+        } else if (src0->type == GGML_TYPE_F16) {
+            rope_norm_cuda_f16(
+                (const half *)src0_d, (half *)dst_d, ne00, n_dims, nr, pos, freq_scale, ne01, freq_base, ext_factor,
+                attn_factor, corr_dims, freq_factors, stream
+            );
+        } else {
+            GGML_ABORT("fatal error");
+        }
+    }
+}
diff --git a/src/whisper_cpp/ggml/src/ggml-cuda/rope.cuh b/src/whisper_cpp/ggml/src/ggml-cuda/rope.cuh
new file mode 100644
index 00000000..0f787a0b
--- /dev/null
+++ b/src/whisper_cpp/ggml/src/ggml-cuda/rope.cuh
@@ -0,0 +1,5 @@
+#include "common.cuh"
+
+#define CUDA_ROPE_BLOCK_SIZE 256
+
+void ggml_cuda_op_rope(ggml_backend_cuda_context & ctx, ggml_tensor * dst);
diff --git a/src/whisper_cpp/ggml/src/ggml-cuda/rwkv-wkv.cu b/src/whisper_cpp/ggml/src/ggml-cuda/rwkv-wkv.cu
new file mode 100644
index 00000000..098e92d3
--- /dev/null
+++ b/src/whisper_cpp/ggml/src/ggml-cuda/rwkv-wkv.cu
@@ -0,0 +1,89 @@
+#include "common.cuh"
+#include "rwkv-wkv.cuh"
+
+static __global__ void rwkv_wkv_f32(const int B, const int T, const int C, const int H, const float * k, const float * v, const float * r, const float * tf, const float * td, const float * s, float * dst) {
+    const int tid = threadIdx.x;
+    const int bid = blockIdx.x;
+
+    const int head_size = CUDA_WKV_BLOCK_SIZE;
+    const int batch_i = bid / H;
+    const int head_i = bid % H;
+    const int state_size = C * head_size;
+    const int n_seq_tokens = T / B;
+
+    float state[head_size];
+    __shared__ float _k[head_size], _r[head_size], _tf[head_size], _td[head_size];
+
+    #pragma unroll
+    for (int i = 0; i < head_size; i++) {
+        state[i] = s[batch_i * state_size + head_i * head_size * head_size + i * head_size + tid];
+    }
+
+    __syncthreads();
+    _tf[tid] = tf[head_i * head_size + tid];
+    __syncthreads();
+
+    for (int t = batch_i * n_seq_tokens * C + head_i * head_size + tid; t < (batch_i + 1) * n_seq_tokens * C + head_i * head_size + tid; t += C) {
+        __syncthreads();
+        _k[tid] = k[t];
+        _r[tid] = r[t];
+        _td[tid] = td[t];
+        __syncthreads();
+
+        const float _v = v[t];
+        float y = 0;
+        for (int j = 0; j < head_size; j += 4) {
+            const float4& k = (float4&)(_k[j]);
+            const float4& r = (float4&)(_r[j]);
+            const float4& tf = (float4&)(_tf[j]);
+            const float4& td = (float4&)(_td[j]);
+            float4& s = (float4&)(state[j]);
+            float4 kv;
+
+            kv.x = k.x * _v;
+            kv.y = k.y * _v;
+            kv.z = k.z * _v;
+            kv.w = k.w * _v;
+
+            y += r.x * (tf.x * kv.x + s.x);
+            y += r.y * (tf.y * kv.y + s.y);
+            y += r.z * (tf.z * kv.z + s.z);
+            y += r.w * (tf.w * kv.w + s.w);
+
+            s.x = s.x * td.x + kv.x;
+            s.y = s.y * td.y + kv.y;
+            s.z = s.z * td.z + kv.z;
+            s.w = s.w * td.w + kv.w;
+        }
+        dst[t] = y;
+    }
+
+    #pragma unroll
+    for (int i = 0; i < head_size; i++) {
+        dst[T * C + batch_i * state_size + head_i * head_size * head_size + i * head_size + tid] = state[i];
+    }
+}
+
+void ggml_cuda_op_rwkv_wkv(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
+    const float * k_d  = (const float *)dst->src[0]->data;
+    const float * v_d  = (const float *)dst->src[1]->data;
+    const float * r_d  = (const float *)dst->src[2]->data;
+    const float * tf_d = (const float *)dst->src[3]->data;
+    const float * td_d = (const float *)dst->src[4]->data;
+    const float * s_d  = (const float *)dst->src[5]->data;
+
+    const int64_t B = dst->src[5]->ne[1];
+    const int64_t T = dst->src[0]->ne[3];
+    const int64_t C = dst->ne[0];
+    const int64_t H = dst->src[0]->ne[2];
+
+    float * dst_d = (float *)dst->data;
+
+    cudaStream_t stream = ctx.stream();
+
+    GGML_ASSERT(dst->src[5]->type == GGML_TYPE_F32);
+    GGML_ASSERT(C % H == 0);
+    GGML_ASSERT(C / H == CUDA_WKV_BLOCK_SIZE);
+
+    rwkv_wkv_f32<<<B * H, C / H, 0, stream>>>(B, T, C, H, k_d, v_d, r_d, tf_d, td_d, s_d, dst_d);
+}
diff --git a/src/whisper_cpp/ggml/src/ggml-cuda/rwkv-wkv.cuh b/src/whisper_cpp/ggml/src/ggml-cuda/rwkv-wkv.cuh
new file mode 100644
index 00000000..13795247
--- /dev/null
+++ b/src/whisper_cpp/ggml/src/ggml-cuda/rwkv-wkv.cuh
@@ -0,0 +1,5 @@
+#include "common.cuh"
+
+#define CUDA_WKV_BLOCK_SIZE 64
+
+void ggml_cuda_op_rwkv_wkv(ggml_backend_cuda_context & ctx, ggml_tensor * dst);
diff --git a/src/whisper_cpp/ggml/src/ggml-cuda/scale.cu b/src/whisper_cpp/ggml/src/ggml-cuda/scale.cu
new file mode 100644
index 00000000..1405e066
--- /dev/null
+++ b/src/whisper_cpp/ggml/src/ggml-cuda/scale.cu
@@ -0,0 +1,31 @@
+#include "scale.cuh"
+
+static __global__ void scale_f32(const float * x, float * dst, const float scale, const int k) {
+    const int i = blockDim.x*blockIdx.x + threadIdx.x;
+
+    if (i >= k) {
+        return;
+    }
+
+    dst[i] = scale * x[i];
+}
+
+static void scale_f32_cuda(const float * x, float * dst, const float scale, const int k, cudaStream_t stream) {
+    const int num_blocks = (k + CUDA_SCALE_BLOCK_SIZE - 1) / CUDA_SCALE_BLOCK_SIZE;
+    scale_f32<<<num_blocks, CUDA_SCALE_BLOCK_SIZE, 0, stream>>>(x, dst, scale, k);
+}
+
+void ggml_cuda_op_scale(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
+    const ggml_tensor * src0 = dst->src[0];
+    const float * src0_d = (const float *)src0->data;
+    float * dst_d = (float *)dst->data;
+    cudaStream_t stream = ctx.stream();
+
+    GGML_ASSERT(src0->type == GGML_TYPE_F32);
+    GGML_ASSERT( dst->type == GGML_TYPE_F32);
+
+    float scale;
+    memcpy(&scale, dst->op_params, sizeof(float));
+
+    scale_f32_cuda(src0_d, dst_d, scale, ggml_nelements(src0), stream);
+}
diff --git a/src/whisper_cpp/ggml/src/ggml-cuda/scale.cuh b/src/whisper_cpp/ggml/src/ggml-cuda/scale.cuh
new file mode 100644
index 00000000..8ff75c82
--- /dev/null
+++ b/src/whisper_cpp/ggml/src/ggml-cuda/scale.cuh
@@ -0,0 +1,5 @@
+#include "common.cuh"
+
+#define CUDA_SCALE_BLOCK_SIZE 256
+
+void ggml_cuda_op_scale(ggml_backend_cuda_context & ctx, ggml_tensor * dst);
diff --git a/src/whisper_cpp/ggml/src/ggml-cuda/softmax.cu b/src/whisper_cpp/ggml/src/ggml-cuda/softmax.cu
new file mode 100644
index 00000000..c24abae1
--- /dev/null
+++ b/src/whisper_cpp/ggml/src/ggml-cuda/softmax.cu
@@ -0,0 +1,206 @@
+#include "common.cuh"
+#include "softmax.cuh"
+
+template <typename T>
+static __device__ __forceinline__ float t2f32(T val) {
+    return (float) val;
+}
+
+template <>
+__device__ float __forceinline__ t2f32<half>(half val) {
+    return __half2float(val);
+}
+
+template <bool vals_smem, int ncols_template, int block_size_template, typename T>
+static __global__ void soft_max_f32(const float * x, const T * mask, float * dst, const int ncols_par, const int nrows_y, const float scale, const float max_bias, const float m0, const float m1, uint32_t n_head_log2) {
+    const int ncols = ncols_template == 0 ? ncols_par : ncols_template;
+
+    const int tid  = threadIdx.x;
+    const int rowx = blockIdx.x;
+    const int rowy = rowx % nrows_y; // broadcast the mask in the row dimension
+
+    const int block_size = block_size_template == 0 ? blockDim.x : block_size_template;
+
+    const int warp_id = threadIdx.x / WARP_SIZE;
+    const int lane_id = threadIdx.x % WARP_SIZE;
+
+    const float slope = get_alibi_slope(max_bias, rowx/nrows_y, n_head_log2, m0, m1);
+
+    extern __shared__ float data_soft_max_f32[];
+    float * buf_iw = data_soft_max_f32; // shared memory buffer for inter-warp communication
+    // shared memory buffer to cache values between iterations:
+    float * vals = vals_smem ? buf_iw + WARP_SIZE : dst + (int64_t)rowx*ncols;
+
+    float max_val = -INFINITY;
+
+#pragma unroll
+    for (int col0 = 0; col0 < ncols; col0 += block_size) {
+        const int col = col0 + tid;
+
+        if (ncols_template == 0 && col >= ncols) {
+            break;
+        }
+
+        const int64_t ix = (int64_t)rowx*ncols + col;
+        const int64_t iy = (int64_t)rowy*ncols + col;
+
+        const float val = x[ix]*scale + (mask ? slope*t2f32(mask[iy]) : 0.0f);
+
+        vals[col] = val;
+        max_val = max(max_val, val);
+    }
+
+    // find the max value in the block
+    max_val = warp_reduce_max(max_val);
+    if (block_size > WARP_SIZE) {
+        if (warp_id == 0) {
+            buf_iw[lane_id] = -INFINITY;
+        }
+        __syncthreads();
+
+        if (lane_id == 0) {
+            buf_iw[warp_id] = max_val;
+        }
+        __syncthreads();
+
+        max_val = buf_iw[lane_id];
+        max_val = warp_reduce_max(max_val);
+    }
+
+    float tmp = 0.0f; // partial sum
+
+#pragma unroll
+    for (int col0 = 0; col0 < ncols; col0 += block_size) {
+        const int col = col0 + tid;
+
+        if (ncols_template == 0 && col >= ncols) {
+            break;
+        }
+
+        const float val = expf(vals[col] - max_val);
+        tmp += val;
+        vals[col] = val;
+    }
+
+    // find the sum of exps in the block
+    tmp = warp_reduce_sum(tmp);
+    if (block_size > WARP_SIZE) {
+        __syncthreads();
+        if (warp_id == 0) {
+            buf_iw[lane_id] = 0.0f;
+        }
+        __syncthreads();
+
+        if (lane_id == 0) {
+            buf_iw[warp_id] = tmp;
+        }
+        __syncthreads();
+
+        tmp = buf_iw[lane_id];
+        tmp = warp_reduce_sum(tmp);
+    }
+
+    const float inv_sum = 1.0f / tmp;
+
+#pragma unroll
+    for (int col0 = 0; col0 < ncols; col0 += block_size) {
+        const int col = col0 + tid;
+
+        if (ncols_template == 0 && col >= ncols) {
+            return;
+        }
+
+        const int64_t idst = (int64_t)rowx*ncols + col;
+        dst[idst] = vals[col] * inv_sum;
+    }
+}
+
+template<typename T>
+static void soft_max_f32_cuda(const float * x, const T * mask, float * dst, const int ncols_x, const int nrows_x, const int nrows_y, const float scale, const float max_bias, cudaStream_t stream) {
+    int nth = WARP_SIZE;
+    while (nth < ncols_x && nth < CUDA_SOFT_MAX_BLOCK_SIZE) nth *= 2;
+    const dim3 block_dims(nth,     1, 1);
+    const dim3 block_nums(nrows_x, 1, 1);
+    const size_t shmem = (GGML_PAD(ncols_x, WARP_SIZE) + WARP_SIZE)*sizeof(float);
+    static_assert(CUDA_SOFT_MAX_BLOCK_SIZE == 1024, "These values need to be adjusted.");
+
+    const uint32_t n_head      = nrows_x/nrows_y;
+    const uint32_t n_head_log2 = 1u << (uint32_t) floorf(log2f((float) n_head));
+
+    const float m0 = powf(2.0f, -(max_bias       ) / n_head_log2);
+    const float m1 = powf(2.0f, -(max_bias / 2.0f) / n_head_log2);
+
+    // FIXME: this limit could be raised by ~2-4x on Ampere or newer
+    if (shmem < ggml_cuda_info().devices[ggml_cuda_get_device()].smpb) {
+        switch (ncols_x) {
+            case 32:
+                soft_max_f32<true, 32, 32><<<block_nums, block_dims, shmem, stream>>>(x, mask, dst, ncols_x, nrows_y, scale, max_bias, m0, m1, n_head_log2);
+                break;
+            case 64:
+                soft_max_f32<true, 64, 64><<<block_nums, block_dims, shmem, stream>>>(x, mask, dst, ncols_x, nrows_y, scale, max_bias, m0, m1, n_head_log2);
+                break;
+            case 128:
+                soft_max_f32<true, 128, 128><<<block_nums, block_dims, shmem, stream>>>(x, mask, dst, ncols_x, nrows_y, scale, max_bias, m0, m1, n_head_log2);
+                break;
+            case 256:
+                soft_max_f32<true, 256, 256><<<block_nums, block_dims, shmem, stream>>>(x, mask, dst, ncols_x, nrows_y, scale, max_bias, m0, m1, n_head_log2);
+                break;
+            case 512:
+                soft_max_f32<true, 512, 512><<<block_nums, block_dims, shmem, stream>>>(x, mask, dst, ncols_x, nrows_y, scale, max_bias, m0, m1, n_head_log2);
+                break;
+            case 1024:
+                soft_max_f32<true, 1024, 1024><<<block_nums, block_dims, shmem, stream>>>(x, mask, dst, ncols_x, nrows_y, scale, max_bias, m0, m1, n_head_log2);
+                break;
+            case 2048:
+                soft_max_f32<true, 2048, 1024><<<block_nums, block_dims, shmem, stream>>>(x, mask, dst, ncols_x, nrows_y, scale, max_bias, m0, m1, n_head_log2);
+                break;
+            case 4096:
+                soft_max_f32<true, 4096, 1024><<<block_nums, block_dims, shmem, stream>>>(x, mask, dst, ncols_x, nrows_y, scale, max_bias, m0, m1, n_head_log2);
+                break;
+            default:
+                soft_max_f32<true, 0, 0><<<block_nums, block_dims, shmem, stream>>>(x, mask, dst, ncols_x, nrows_y, scale, max_bias, m0, m1, n_head_log2);
+                break;
+        }
+    } else {
+        const size_t shmem_low = WARP_SIZE*sizeof(float);
+        soft_max_f32<false, 0, 0><<<block_nums, block_dims, shmem_low, stream>>>(x, mask, dst, ncols_x, nrows_y, scale, max_bias, m0, m1, n_head_log2);
+    }
+}
+
+void ggml_cuda_op_soft_max(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
+    const ggml_tensor * src0 = dst->src[0];
+    const ggml_tensor * src1 = dst->src[1];
+
+    const float * src0_d = (const float *)src0->data;
+    const void  * src1_d = src1 ? (const void *)src1->data : nullptr;
+
+    float * dst_d = (float *)dst->data;
+    cudaStream_t stream = ctx.stream();
+
+    GGML_ASSERT(src0->type == GGML_TYPE_F32);
+    GGML_ASSERT( dst->type == GGML_TYPE_F32);
+
+    GGML_ASSERT(!src1 || src1->type == GGML_TYPE_F16 || src1->type == GGML_TYPE_F32); // src1 contains mask and it is optional
+
+    const int64_t ne00    = src0->ne[0];
+    const int64_t nrows_x = ggml_nrows(src0);
+    const int64_t nrows_y = src0->ne[1];
+
+    float scale    = 1.0f;
+    float max_bias = 0.0f;
+
+    memcpy(&scale,    (float *) dst->op_params + 0, sizeof(float));
+    memcpy(&max_bias, (float *) dst->op_params + 1, sizeof(float));
+
+    const bool use_f16 = (src1 && src1->type == GGML_TYPE_F16);
+
+    if (use_f16) {
+        const half * src1_dd = (const half *)src1_d;
+
+        soft_max_f32_cuda(src0_d, src1_dd, dst_d, ne00, nrows_x, nrows_y, scale, max_bias, stream);
+    } else {
+        const float * src1_dd = (const float *)src1_d;
+
+        soft_max_f32_cuda(src0_d, src1_dd, dst_d, ne00, nrows_x, nrows_y, scale, max_bias, stream);
+    }
+}
diff --git a/src/whisper_cpp/ggml/src/ggml-cuda/softmax.cuh b/src/whisper_cpp/ggml/src/ggml-cuda/softmax.cuh
new file mode 100644
index 00000000..4ef4ff86
--- /dev/null
+++ b/src/whisper_cpp/ggml/src/ggml-cuda/softmax.cuh
@@ -0,0 +1,5 @@
+#include "common.cuh"
+
+#define CUDA_SOFT_MAX_BLOCK_SIZE 1024
+
+void ggml_cuda_op_soft_max(ggml_backend_cuda_context & ctx, ggml_tensor * dst);
diff --git a/src/whisper_cpp/ggml/src/ggml-cuda/sum.cu b/src/whisper_cpp/ggml/src/ggml-cuda/sum.cu
new file mode 100644
index 00000000..0583e4fe
--- /dev/null
+++ b/src/whisper_cpp/ggml/src/ggml-cuda/sum.cu
@@ -0,0 +1,47 @@
+#if !defined(GGML_USE_HIPBLAS) && !defined(GGML_USE_MUSA) && CUDART_VERSION >= 11700
+#define USE_CUB
+#endif // !defined(GGML_USE_HIPBLAS) && !defined(GGML_USE_MUSA) && CUDART_VERSION >= 11700
+
+#ifdef USE_CUB
+// On Windows CUB uses libraries with variables called CC_PASCAL which conflict with the define in common.cuh.
+// For this reason CUB must be included BEFORE anything else.
+#include <cub/cub.cuh>
+using namespace cub;
+#endif // USE_CUB
+
+#include "sumrows.cuh"
+#include "sum.cuh"
+
+#include <cstdint>
+
+void sum_f32_cuda(ggml_cuda_pool & pool, const float * x, float * dst, const int64_t ne, cudaStream_t stream) {
+#ifdef USE_CUB
+    size_t tmp_size = 0;
+    DeviceReduce::Sum(nullptr,       tmp_size, x, dst, ne, stream);
+    ggml_cuda_pool_alloc<uint8_t> tmp_alloc(pool, tmp_size);
+    DeviceReduce::Sum(tmp_alloc.ptr, tmp_size, x, dst, ne, stream);
+#else
+    // Use (inefficient) sum_rows implementation as a fallback.
+    // For AMD there is rocPRIM which could be used as a drop-in replacement via hipcub but this would require C++11 -> C++14.
+    sum_rows_f32_cuda(x, dst, ne, 1, stream);
+    GGML_UNUSED(pool);
+#endif // USE_CUB
+}
+
+void ggml_cuda_op_sum(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
+    const ggml_tensor * src0 = dst->src[0];
+
+    GGML_ASSERT(src0->type == GGML_TYPE_F32);
+    GGML_ASSERT( dst->type == GGML_TYPE_F32);
+    GGML_ASSERT(ggml_is_contiguous(src0));
+
+    const float * src0_d = (const float *) src0->data;
+    float * dst_d = (float *) dst->data;
+
+    const int64_t ne = ggml_nelements(src0);
+
+    ggml_cuda_pool & pool = ctx.pool();
+    cudaStream_t stream = ctx.stream();
+
+    sum_f32_cuda(pool, src0_d, dst_d, ne, stream);
+}
diff --git a/src/whisper_cpp/ggml/src/ggml-cuda/sum.cuh b/src/whisper_cpp/ggml/src/ggml-cuda/sum.cuh
new file mode 100644
index 00000000..8cadc373
--- /dev/null
+++ b/src/whisper_cpp/ggml/src/ggml-cuda/sum.cuh
@@ -0,0 +1,5 @@
+#include "common.cuh"
+
+void sum_f32_cuda(ggml_cuda_pool & pool, const float * x, float * dst, const int64_t ne, cudaStream_t stream);
+
+void ggml_cuda_op_sum(ggml_backend_cuda_context & ctx, ggml_tensor * dst);
diff --git a/src/whisper_cpp/ggml/src/ggml-cuda/sumrows.cu b/src/whisper_cpp/ggml/src/ggml-cuda/sumrows.cu
new file mode 100644
index 00000000..38dbf1b5
--- /dev/null
+++ b/src/whisper_cpp/ggml/src/ggml-cuda/sumrows.cu
@@ -0,0 +1,39 @@
+#include "sumrows.cuh"
+
+static __global__ void k_sum_rows_f32(const float * x, float * dst, const int ncols) {
+    const int row = blockIdx.x;
+    const int col = threadIdx.x;
+
+    float sum = 0.0f;
+    for (int i = col; i < ncols; i += blockDim.x) {
+        sum += x[row * ncols + i];
+    }
+
+    sum = warp_reduce_sum(sum);
+
+    if (col == 0) {
+        dst[row] = sum;
+    }
+}
+
+void sum_rows_f32_cuda(const float * x, float * dst, const int ncols, const int nrows, cudaStream_t stream) {
+    const dim3 block_dims(WARP_SIZE, 1, 1);
+    const dim3 block_nums(nrows, 1, 1);
+    k_sum_rows_f32<<<block_nums, block_dims, 0, stream>>>(x, dst, ncols);
+}
+
+void ggml_cuda_op_sum_rows(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
+    const ggml_tensor * src0 = dst->src[0];
+    const float * src0_d = (const float *)src0->data;
+    float * dst_d = (float *)dst->data;
+    cudaStream_t stream = ctx.stream();
+
+    GGML_ASSERT(src0->type == GGML_TYPE_F32);
+    GGML_ASSERT( dst->type == GGML_TYPE_F32);
+    GGML_ASSERT(ggml_is_contiguous(src0));
+
+    const int64_t ncols = src0->ne[0];
+    const int64_t nrows = ggml_nrows(src0);
+
+    sum_rows_f32_cuda(src0_d, dst_d, ncols, nrows, stream);
+}
diff --git a/src/whisper_cpp/ggml/src/ggml-cuda/sumrows.cuh b/src/whisper_cpp/ggml/src/ggml-cuda/sumrows.cuh
new file mode 100644
index 00000000..191db1c1
--- /dev/null
+++ b/src/whisper_cpp/ggml/src/ggml-cuda/sumrows.cuh
@@ -0,0 +1,5 @@
+#include "common.cuh"
+
+void sum_rows_f32_cuda(const float * x, float * dst, const int ncols, const int nrows, cudaStream_t stream);
+
+void ggml_cuda_op_sum_rows(ggml_backend_cuda_context & ctx, ggml_tensor * dst);
diff --git a/src/whisper_cpp/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-f16-f16.cu b/src/whisper_cpp/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-f16-f16.cu
new file mode 100644
index 00000000..6696a238
--- /dev/null
+++ b/src/whisper_cpp/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-f16-f16.cu
@@ -0,0 +1,5 @@
+// This file has been autogenerated by generate_cu_files.py, do not edit manually.
+
+#include "../fattn-vec-f16.cuh"
+
+DECL_FATTN_VEC_F16_CASE(128, GGML_TYPE_F16, GGML_TYPE_F16);
diff --git a/src/whisper_cpp/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-f16-q4_0.cu b/src/whisper_cpp/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-f16-q4_0.cu
new file mode 100644
index 00000000..dd070db2
--- /dev/null
+++ b/src/whisper_cpp/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-f16-q4_0.cu
@@ -0,0 +1,5 @@
+// This file has been autogenerated by generate_cu_files.py, do not edit manually.
+
+#include "../fattn-vec-f16.cuh"
+
+DECL_FATTN_VEC_F16_CASE(128, GGML_TYPE_F16, GGML_TYPE_Q4_0);
diff --git a/src/whisper_cpp/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-f16-q4_1.cu b/src/whisper_cpp/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-f16-q4_1.cu
new file mode 100644
index 00000000..54dcde6f
--- /dev/null
+++ b/src/whisper_cpp/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-f16-q4_1.cu
@@ -0,0 +1,5 @@
+// This file has been autogenerated by generate_cu_files.py, do not edit manually.
+
+#include "../fattn-vec-f16.cuh"
+
+DECL_FATTN_VEC_F16_CASE(128, GGML_TYPE_F16, GGML_TYPE_Q4_1);
diff --git a/src/whisper_cpp/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-f16-q5_0.cu b/src/whisper_cpp/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-f16-q5_0.cu
new file mode 100644
index 00000000..4ec22f79
--- /dev/null
+++ b/src/whisper_cpp/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-f16-q5_0.cu
@@ -0,0 +1,5 @@
+// This file has been autogenerated by generate_cu_files.py, do not edit manually.
+
+#include "../fattn-vec-f16.cuh"
+
+DECL_FATTN_VEC_F16_CASE(128, GGML_TYPE_F16, GGML_TYPE_Q5_0);
diff --git a/src/whisper_cpp/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-f16-q5_1.cu b/src/whisper_cpp/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-f16-q5_1.cu
new file mode 100644
index 00000000..3c15bf7f
--- /dev/null
+++ b/src/whisper_cpp/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-f16-q5_1.cu
@@ -0,0 +1,5 @@
+// This file has been autogenerated by generate_cu_files.py, do not edit manually.
+
+#include "../fattn-vec-f16.cuh"
+
+DECL_FATTN_VEC_F16_CASE(128, GGML_TYPE_F16, GGML_TYPE_Q5_1);
diff --git a/src/whisper_cpp/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-f16-q8_0.cu b/src/whisper_cpp/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-f16-q8_0.cu
new file mode 100644
index 00000000..7e61b5fd
--- /dev/null
+++ b/src/whisper_cpp/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-f16-q8_0.cu
@@ -0,0 +1,5 @@
+// This file has been autogenerated by generate_cu_files.py, do not edit manually.
+
+#include "../fattn-vec-f16.cuh"
+
+DECL_FATTN_VEC_F16_CASE(128, GGML_TYPE_F16, GGML_TYPE_Q8_0);
diff --git a/src/whisper_cpp/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q4_0-f16.cu b/src/whisper_cpp/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q4_0-f16.cu
new file mode 100644
index 00000000..fdb15b58
--- /dev/null
+++ b/src/whisper_cpp/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q4_0-f16.cu
@@ -0,0 +1,5 @@
+// This file has been autogenerated by generate_cu_files.py, do not edit manually.
+
+#include "../fattn-vec-f16.cuh"
+
+DECL_FATTN_VEC_F16_CASE(128, GGML_TYPE_Q4_0, GGML_TYPE_F16);
diff --git a/src/whisper_cpp/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q4_0-q4_0.cu b/src/whisper_cpp/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q4_0-q4_0.cu
new file mode 100644
index 00000000..0f7c417d
--- /dev/null
+++ b/src/whisper_cpp/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q4_0-q4_0.cu
@@ -0,0 +1,5 @@
+// This file has been autogenerated by generate_cu_files.py, do not edit manually.
+
+#include "../fattn-vec-f16.cuh"
+
+DECL_FATTN_VEC_F16_CASE(128, GGML_TYPE_Q4_0, GGML_TYPE_Q4_0);
diff --git a/src/whisper_cpp/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q4_0-q4_1.cu b/src/whisper_cpp/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q4_0-q4_1.cu
new file mode 100644
index 00000000..851f33c4
--- /dev/null
+++ b/src/whisper_cpp/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q4_0-q4_1.cu
@@ -0,0 +1,5 @@
+// This file has been autogenerated by generate_cu_files.py, do not edit manually.
+
+#include "../fattn-vec-f16.cuh"
+
+DECL_FATTN_VEC_F16_CASE(128, GGML_TYPE_Q4_0, GGML_TYPE_Q4_1);
diff --git a/src/whisper_cpp/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q4_0-q5_0.cu b/src/whisper_cpp/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q4_0-q5_0.cu
new file mode 100644
index 00000000..763809cb
--- /dev/null
+++ b/src/whisper_cpp/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q4_0-q5_0.cu
@@ -0,0 +1,5 @@
+// This file has been autogenerated by generate_cu_files.py, do not edit manually.
+
+#include "../fattn-vec-f16.cuh"
+
+DECL_FATTN_VEC_F16_CASE(128, GGML_TYPE_Q4_0, GGML_TYPE_Q5_0);
diff --git a/src/whisper_cpp/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q4_0-q5_1.cu b/src/whisper_cpp/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q4_0-q5_1.cu
new file mode 100644
index 00000000..f2a276e5
--- /dev/null
+++ b/src/whisper_cpp/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q4_0-q5_1.cu
@@ -0,0 +1,5 @@
+// This file has been autogenerated by generate_cu_files.py, do not edit manually.
+
+#include "../fattn-vec-f16.cuh"
+
+DECL_FATTN_VEC_F16_CASE(128, GGML_TYPE_Q4_0, GGML_TYPE_Q5_1);
diff --git a/src/whisper_cpp/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q4_0-q8_0.cu b/src/whisper_cpp/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q4_0-q8_0.cu
new file mode 100644
index 00000000..cb227f6f
--- /dev/null
+++ b/src/whisper_cpp/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q4_0-q8_0.cu
@@ -0,0 +1,5 @@
+// This file has been autogenerated by generate_cu_files.py, do not edit manually.
+
+#include "../fattn-vec-f16.cuh"
+
+DECL_FATTN_VEC_F16_CASE(128, GGML_TYPE_Q4_0, GGML_TYPE_Q8_0);
diff --git a/src/whisper_cpp/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q4_1-f16.cu b/src/whisper_cpp/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q4_1-f16.cu
new file mode 100644
index 00000000..97ac0520
--- /dev/null
+++ b/src/whisper_cpp/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q4_1-f16.cu
@@ -0,0 +1,5 @@
+// This file has been autogenerated by generate_cu_files.py, do not edit manually.
+
+#include "../fattn-vec-f16.cuh"
+
+DECL_FATTN_VEC_F16_CASE(128, GGML_TYPE_Q4_1, GGML_TYPE_F16);
diff --git a/src/whisper_cpp/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q4_1-q4_0.cu b/src/whisper_cpp/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q4_1-q4_0.cu
new file mode 100644
index 00000000..c772b426
--- /dev/null
+++ b/src/whisper_cpp/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q4_1-q4_0.cu
@@ -0,0 +1,5 @@
+// This file has been autogenerated by generate_cu_files.py, do not edit manually.
+
+#include "../fattn-vec-f16.cuh"
+
+DECL_FATTN_VEC_F16_CASE(128, GGML_TYPE_Q4_1, GGML_TYPE_Q4_0);
diff --git a/src/whisper_cpp/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q4_1-q4_1.cu b/src/whisper_cpp/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q4_1-q4_1.cu
new file mode 100644
index 00000000..5cb74308
--- /dev/null
+++ b/src/whisper_cpp/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q4_1-q4_1.cu
@@ -0,0 +1,5 @@
+// This file has been autogenerated by generate_cu_files.py, do not edit manually.
+
+#include "../fattn-vec-f16.cuh"
+
+DECL_FATTN_VEC_F16_CASE(128, GGML_TYPE_Q4_1, GGML_TYPE_Q4_1);
diff --git a/src/whisper_cpp/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q4_1-q5_0.cu b/src/whisper_cpp/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q4_1-q5_0.cu
new file mode 100644
index 00000000..98a709d1
--- /dev/null
+++ b/src/whisper_cpp/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q4_1-q5_0.cu
@@ -0,0 +1,5 @@
+// This file has been autogenerated by generate_cu_files.py, do not edit manually.
+
+#include "../fattn-vec-f16.cuh"
+
+DECL_FATTN_VEC_F16_CASE(128, GGML_TYPE_Q4_1, GGML_TYPE_Q5_0);
diff --git a/src/whisper_cpp/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q4_1-q5_1.cu b/src/whisper_cpp/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q4_1-q5_1.cu
new file mode 100644
index 00000000..4f2f947a
--- /dev/null
+++ b/src/whisper_cpp/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q4_1-q5_1.cu
@@ -0,0 +1,5 @@
+// This file has been autogenerated by generate_cu_files.py, do not edit manually.
+
+#include "../fattn-vec-f16.cuh"
+
+DECL_FATTN_VEC_F16_CASE(128, GGML_TYPE_Q4_1, GGML_TYPE_Q5_1);
diff --git a/src/whisper_cpp/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q4_1-q8_0.cu b/src/whisper_cpp/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q4_1-q8_0.cu
new file mode 100644
index 00000000..11f96b6f
--- /dev/null
+++ b/src/whisper_cpp/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q4_1-q8_0.cu
@@ -0,0 +1,5 @@
+// This file has been autogenerated by generate_cu_files.py, do not edit manually.
+
+#include "../fattn-vec-f16.cuh"
+
+DECL_FATTN_VEC_F16_CASE(128, GGML_TYPE_Q4_1, GGML_TYPE_Q8_0);
diff --git a/src/whisper_cpp/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q5_0-f16.cu b/src/whisper_cpp/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q5_0-f16.cu
new file mode 100644
index 00000000..b39bdc06
--- /dev/null
+++ b/src/whisper_cpp/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q5_0-f16.cu
@@ -0,0 +1,5 @@
+// This file has been autogenerated by generate_cu_files.py, do not edit manually.
+
+#include "../fattn-vec-f16.cuh"
+
+DECL_FATTN_VEC_F16_CASE(128, GGML_TYPE_Q5_0, GGML_TYPE_F16);
diff --git a/src/whisper_cpp/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q5_0-q4_0.cu b/src/whisper_cpp/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q5_0-q4_0.cu
new file mode 100644
index 00000000..bbd6a2c7
--- /dev/null
+++ b/src/whisper_cpp/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q5_0-q4_0.cu
@@ -0,0 +1,5 @@
+// This file has been autogenerated by generate_cu_files.py, do not edit manually.
+
+#include "../fattn-vec-f16.cuh"
+
+DECL_FATTN_VEC_F16_CASE(128, GGML_TYPE_Q5_0, GGML_TYPE_Q4_0);
diff --git a/src/whisper_cpp/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q5_0-q4_1.cu b/src/whisper_cpp/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q5_0-q4_1.cu
new file mode 100644
index 00000000..9d84ff2b
--- /dev/null
+++ b/src/whisper_cpp/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q5_0-q4_1.cu
@@ -0,0 +1,5 @@
+// This file has been autogenerated by generate_cu_files.py, do not edit manually.
+
+#include "../fattn-vec-f16.cuh"
+
+DECL_FATTN_VEC_F16_CASE(128, GGML_TYPE_Q5_0, GGML_TYPE_Q4_1);
diff --git a/src/whisper_cpp/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q5_0-q5_0.cu b/src/whisper_cpp/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q5_0-q5_0.cu
new file mode 100644
index 00000000..bc8a5bff
--- /dev/null
+++ b/src/whisper_cpp/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q5_0-q5_0.cu
@@ -0,0 +1,5 @@
+// This file has been autogenerated by generate_cu_files.py, do not edit manually.
+
+#include "../fattn-vec-f16.cuh"
+
+DECL_FATTN_VEC_F16_CASE(128, GGML_TYPE_Q5_0, GGML_TYPE_Q5_0);
diff --git a/src/whisper_cpp/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q5_0-q5_1.cu b/src/whisper_cpp/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q5_0-q5_1.cu
new file mode 100644
index 00000000..a679100c
--- /dev/null
+++ b/src/whisper_cpp/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q5_0-q5_1.cu
@@ -0,0 +1,5 @@
+// This file has been autogenerated by generate_cu_files.py, do not edit manually.
+
+#include "../fattn-vec-f16.cuh"
+
+DECL_FATTN_VEC_F16_CASE(128, GGML_TYPE_Q5_0, GGML_TYPE_Q5_1);
diff --git a/src/whisper_cpp/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q5_0-q8_0.cu b/src/whisper_cpp/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q5_0-q8_0.cu
new file mode 100644
index 00000000..8f21bccf
--- /dev/null
+++ b/src/whisper_cpp/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q5_0-q8_0.cu
@@ -0,0 +1,5 @@
+// This file has been autogenerated by generate_cu_files.py, do not edit manually.
+
+#include "../fattn-vec-f16.cuh"
+
+DECL_FATTN_VEC_F16_CASE(128, GGML_TYPE_Q5_0, GGML_TYPE_Q8_0);
diff --git a/src/whisper_cpp/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q5_1-f16.cu b/src/whisper_cpp/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q5_1-f16.cu
new file mode 100644
index 00000000..858b00fd
--- /dev/null
+++ b/src/whisper_cpp/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q5_1-f16.cu
@@ -0,0 +1,5 @@
+// This file has been autogenerated by generate_cu_files.py, do not edit manually.
+
+#include "../fattn-vec-f16.cuh"
+
+DECL_FATTN_VEC_F16_CASE(128, GGML_TYPE_Q5_1, GGML_TYPE_F16);
diff --git a/src/whisper_cpp/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q5_1-q4_0.cu b/src/whisper_cpp/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q5_1-q4_0.cu
new file mode 100644
index 00000000..0fc8011f
--- /dev/null
+++ b/src/whisper_cpp/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q5_1-q4_0.cu
@@ -0,0 +1,5 @@
+// This file has been autogenerated by generate_cu_files.py, do not edit manually.
+
+#include "../fattn-vec-f16.cuh"
+
+DECL_FATTN_VEC_F16_CASE(128, GGML_TYPE_Q5_1, GGML_TYPE_Q4_0);
diff --git a/src/whisper_cpp/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q5_1-q4_1.cu b/src/whisper_cpp/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q5_1-q4_1.cu
new file mode 100644
index 00000000..261fdf62
--- /dev/null
+++ b/src/whisper_cpp/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q5_1-q4_1.cu
@@ -0,0 +1,5 @@
+// This file has been autogenerated by generate_cu_files.py, do not edit manually.
+
+#include "../fattn-vec-f16.cuh"
+
+DECL_FATTN_VEC_F16_CASE(128, GGML_TYPE_Q5_1, GGML_TYPE_Q4_1);
diff --git a/src/whisper_cpp/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q5_1-q5_0.cu b/src/whisper_cpp/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q5_1-q5_0.cu
new file mode 100644
index 00000000..0fb82473
--- /dev/null
+++ b/src/whisper_cpp/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q5_1-q5_0.cu
@@ -0,0 +1,5 @@
+// This file has been autogenerated by generate_cu_files.py, do not edit manually.
+
+#include "../fattn-vec-f16.cuh"
+
+DECL_FATTN_VEC_F16_CASE(128, GGML_TYPE_Q5_1, GGML_TYPE_Q5_0);
diff --git a/src/whisper_cpp/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q5_1-q5_1.cu b/src/whisper_cpp/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q5_1-q5_1.cu
new file mode 100644
index 00000000..a9d9d089
--- /dev/null
+++ b/src/whisper_cpp/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q5_1-q5_1.cu
@@ -0,0 +1,5 @@
+// This file has been autogenerated by generate_cu_files.py, do not edit manually.
+
+#include "../fattn-vec-f16.cuh"
+
+DECL_FATTN_VEC_F16_CASE(128, GGML_TYPE_Q5_1, GGML_TYPE_Q5_1);
diff --git a/src/whisper_cpp/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q5_1-q8_0.cu b/src/whisper_cpp/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q5_1-q8_0.cu
new file mode 100644
index 00000000..7d7b2792
--- /dev/null
+++ b/src/whisper_cpp/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q5_1-q8_0.cu
@@ -0,0 +1,5 @@
+// This file has been autogenerated by generate_cu_files.py, do not edit manually.
+
+#include "../fattn-vec-f16.cuh"
+
+DECL_FATTN_VEC_F16_CASE(128, GGML_TYPE_Q5_1, GGML_TYPE_Q8_0);
diff --git a/src/whisper_cpp/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q8_0-f16.cu b/src/whisper_cpp/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q8_0-f16.cu
new file mode 100644
index 00000000..a092ee2d
--- /dev/null
+++ b/src/whisper_cpp/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q8_0-f16.cu
@@ -0,0 +1,5 @@
+// This file has been autogenerated by generate_cu_files.py, do not edit manually.
+
+#include "../fattn-vec-f16.cuh"
+
+DECL_FATTN_VEC_F16_CASE(128, GGML_TYPE_Q8_0, GGML_TYPE_F16);
diff --git a/src/whisper_cpp/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q8_0-q4_0.cu b/src/whisper_cpp/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q8_0-q4_0.cu
new file mode 100644
index 00000000..db55927a
--- /dev/null
+++ b/src/whisper_cpp/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q8_0-q4_0.cu
@@ -0,0 +1,5 @@
+// This file has been autogenerated by generate_cu_files.py, do not edit manually.
+
+#include "../fattn-vec-f16.cuh"
+
+DECL_FATTN_VEC_F16_CASE(128, GGML_TYPE_Q8_0, GGML_TYPE_Q4_0);
diff --git a/src/whisper_cpp/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q8_0-q4_1.cu b/src/whisper_cpp/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q8_0-q4_1.cu
new file mode 100644
index 00000000..c3c21cef
--- /dev/null
+++ b/src/whisper_cpp/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q8_0-q4_1.cu
@@ -0,0 +1,5 @@
+// This file has been autogenerated by generate_cu_files.py, do not edit manually.
+
+#include "../fattn-vec-f16.cuh"
+
+DECL_FATTN_VEC_F16_CASE(128, GGML_TYPE_Q8_0, GGML_TYPE_Q4_1);
diff --git a/src/whisper_cpp/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q8_0-q5_0.cu b/src/whisper_cpp/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q8_0-q5_0.cu
new file mode 100644
index 00000000..35dd9f52
--- /dev/null
+++ b/src/whisper_cpp/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q8_0-q5_0.cu
@@ -0,0 +1,5 @@
+// This file has been autogenerated by generate_cu_files.py, do not edit manually.
+
+#include "../fattn-vec-f16.cuh"
+
+DECL_FATTN_VEC_F16_CASE(128, GGML_TYPE_Q8_0, GGML_TYPE_Q5_0);
diff --git a/src/whisper_cpp/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q8_0-q5_1.cu b/src/whisper_cpp/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q8_0-q5_1.cu
new file mode 100644
index 00000000..050c22ac
--- /dev/null
+++ b/src/whisper_cpp/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q8_0-q5_1.cu
@@ -0,0 +1,5 @@
+// This file has been autogenerated by generate_cu_files.py, do not edit manually.
+
+#include "../fattn-vec-f16.cuh"
+
+DECL_FATTN_VEC_F16_CASE(128, GGML_TYPE_Q8_0, GGML_TYPE_Q5_1);
diff --git a/src/whisper_cpp/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q8_0-q8_0.cu b/src/whisper_cpp/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q8_0-q8_0.cu
new file mode 100644
index 00000000..de4866c5
--- /dev/null
+++ b/src/whisper_cpp/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs128-q8_0-q8_0.cu
@@ -0,0 +1,5 @@
+// This file has been autogenerated by generate_cu_files.py, do not edit manually.
+
+#include "../fattn-vec-f16.cuh"
+
+DECL_FATTN_VEC_F16_CASE(128, GGML_TYPE_Q8_0, GGML_TYPE_Q8_0);
diff --git a/src/whisper_cpp/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs256-f16-f16.cu b/src/whisper_cpp/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs256-f16-f16.cu
new file mode 100644
index 00000000..57a10bc4
--- /dev/null
+++ b/src/whisper_cpp/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs256-f16-f16.cu
@@ -0,0 +1,5 @@
+// This file has been autogenerated by generate_cu_files.py, do not edit manually.
+
+#include "../fattn-vec-f16.cuh"
+
+DECL_FATTN_VEC_F16_CASE(256, GGML_TYPE_F16, GGML_TYPE_F16);
diff --git a/src/whisper_cpp/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs64-f16-f16.cu b/src/whisper_cpp/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs64-f16-f16.cu
new file mode 100644
index 00000000..e0f08b46
--- /dev/null
+++ b/src/whisper_cpp/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs64-f16-f16.cu
@@ -0,0 +1,5 @@
+// This file has been autogenerated by generate_cu_files.py, do not edit manually.
+
+#include "../fattn-vec-f16.cuh"
+
+DECL_FATTN_VEC_F16_CASE(64, GGML_TYPE_F16, GGML_TYPE_F16);
diff --git a/src/whisper_cpp/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs64-f16-q4_0.cu b/src/whisper_cpp/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs64-f16-q4_0.cu
new file mode 100644
index 00000000..1c8e8a46
--- /dev/null
+++ b/src/whisper_cpp/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs64-f16-q4_0.cu
@@ -0,0 +1,5 @@
+// This file has been autogenerated by generate_cu_files.py, do not edit manually.
+
+#include "../fattn-vec-f16.cuh"
+
+DECL_FATTN_VEC_F16_CASE(64, GGML_TYPE_F16, GGML_TYPE_Q4_0);
diff --git a/src/whisper_cpp/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs64-f16-q4_1.cu b/src/whisper_cpp/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs64-f16-q4_1.cu
new file mode 100644
index 00000000..cefed83f
--- /dev/null
+++ b/src/whisper_cpp/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs64-f16-q4_1.cu
@@ -0,0 +1,5 @@
+// This file has been autogenerated by generate_cu_files.py, do not edit manually.
+
+#include "../fattn-vec-f16.cuh"
+
+DECL_FATTN_VEC_F16_CASE(64, GGML_TYPE_F16, GGML_TYPE_Q4_1);
diff --git a/src/whisper_cpp/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs64-f16-q5_0.cu b/src/whisper_cpp/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs64-f16-q5_0.cu
new file mode 100644
index 00000000..aede6e35
--- /dev/null
+++ b/src/whisper_cpp/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs64-f16-q5_0.cu
@@ -0,0 +1,5 @@
+// This file has been autogenerated by generate_cu_files.py, do not edit manually.
+
+#include "../fattn-vec-f16.cuh"
+
+DECL_FATTN_VEC_F16_CASE(64, GGML_TYPE_F16, GGML_TYPE_Q5_0);
diff --git a/src/whisper_cpp/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs64-f16-q5_1.cu b/src/whisper_cpp/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs64-f16-q5_1.cu
new file mode 100644
index 00000000..1a1a92c7
--- /dev/null
+++ b/src/whisper_cpp/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs64-f16-q5_1.cu
@@ -0,0 +1,5 @@
+// This file has been autogenerated by generate_cu_files.py, do not edit manually.
+
+#include "../fattn-vec-f16.cuh"
+
+DECL_FATTN_VEC_F16_CASE(64, GGML_TYPE_F16, GGML_TYPE_Q5_1);
diff --git a/src/whisper_cpp/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs64-f16-q8_0.cu b/src/whisper_cpp/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs64-f16-q8_0.cu
new file mode 100644
index 00000000..ad667473
--- /dev/null
+++ b/src/whisper_cpp/ggml/src/ggml-cuda/template-instances/fattn-vec-f16-instance-hs64-f16-q8_0.cu
@@ -0,0 +1,5 @@
+// This file has been autogenerated by generate_cu_files.py, do not edit manually.
+
+#include "../fattn-vec-f16.cuh"
+
+DECL_FATTN_VEC_F16_CASE(64, GGML_TYPE_F16, GGML_TYPE_Q8_0);
diff --git a/src/whisper_cpp/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-f16-f16.cu b/src/whisper_cpp/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-f16-f16.cu
new file mode 100644
index 00000000..c499f455
--- /dev/null
+++ b/src/whisper_cpp/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-f16-f16.cu
@@ -0,0 +1,5 @@
+// This file has been autogenerated by generate_cu_files.py, do not edit manually.
+
+#include "../fattn-vec-f32.cuh"
+
+DECL_FATTN_VEC_F32_CASE(128, GGML_TYPE_F16, GGML_TYPE_F16);
diff --git a/src/whisper_cpp/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-f16-q4_0.cu b/src/whisper_cpp/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-f16-q4_0.cu
new file mode 100644
index 00000000..8286ebf3
--- /dev/null
+++ b/src/whisper_cpp/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-f16-q4_0.cu
@@ -0,0 +1,5 @@
+// This file has been autogenerated by generate_cu_files.py, do not edit manually.
+
+#include "../fattn-vec-f32.cuh"
+
+DECL_FATTN_VEC_F32_CASE(128, GGML_TYPE_F16, GGML_TYPE_Q4_0);
diff --git a/src/whisper_cpp/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-f16-q4_1.cu b/src/whisper_cpp/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-f16-q4_1.cu
new file mode 100644
index 00000000..45878688
--- /dev/null
+++ b/src/whisper_cpp/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-f16-q4_1.cu
@@ -0,0 +1,5 @@
+// This file has been autogenerated by generate_cu_files.py, do not edit manually.
+
+#include "../fattn-vec-f32.cuh"
+
+DECL_FATTN_VEC_F32_CASE(128, GGML_TYPE_F16, GGML_TYPE_Q4_1);
diff --git a/src/whisper_cpp/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-f16-q5_0.cu b/src/whisper_cpp/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-f16-q5_0.cu
new file mode 100644
index 00000000..d89103ce
--- /dev/null
+++ b/src/whisper_cpp/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-f16-q5_0.cu
@@ -0,0 +1,5 @@
+// This file has been autogenerated by generate_cu_files.py, do not edit manually.
+
+#include "../fattn-vec-f32.cuh"
+
+DECL_FATTN_VEC_F32_CASE(128, GGML_TYPE_F16, GGML_TYPE_Q5_0);
diff --git a/src/whisper_cpp/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-f16-q5_1.cu b/src/whisper_cpp/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-f16-q5_1.cu
new file mode 100644
index 00000000..bb75fd42
--- /dev/null
+++ b/src/whisper_cpp/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-f16-q5_1.cu
@@ -0,0 +1,5 @@
+// This file has been autogenerated by generate_cu_files.py, do not edit manually.
+
+#include "../fattn-vec-f32.cuh"
+
+DECL_FATTN_VEC_F32_CASE(128, GGML_TYPE_F16, GGML_TYPE_Q5_1);
diff --git a/src/whisper_cpp/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-f16-q8_0.cu b/src/whisper_cpp/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-f16-q8_0.cu
new file mode 100644
index 00000000..b1629817
--- /dev/null
+++ b/src/whisper_cpp/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-f16-q8_0.cu
@@ -0,0 +1,5 @@
+// This file has been autogenerated by generate_cu_files.py, do not edit manually.
+
+#include "../fattn-vec-f32.cuh"
+
+DECL_FATTN_VEC_F32_CASE(128, GGML_TYPE_F16, GGML_TYPE_Q8_0);
diff --git a/src/whisper_cpp/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q4_0-f16.cu b/src/whisper_cpp/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q4_0-f16.cu
new file mode 100644
index 00000000..d8657604
--- /dev/null
+++ b/src/whisper_cpp/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q4_0-f16.cu
@@ -0,0 +1,5 @@
+// This file has been autogenerated by generate_cu_files.py, do not edit manually.
+
+#include "../fattn-vec-f32.cuh"
+
+DECL_FATTN_VEC_F32_CASE(128, GGML_TYPE_Q4_0, GGML_TYPE_F16);
diff --git a/src/whisper_cpp/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q4_0-q4_0.cu b/src/whisper_cpp/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q4_0-q4_0.cu
new file mode 100644
index 00000000..2e5bd2f1
--- /dev/null
+++ b/src/whisper_cpp/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q4_0-q4_0.cu
@@ -0,0 +1,5 @@
+// This file has been autogenerated by generate_cu_files.py, do not edit manually.
+
+#include "../fattn-vec-f32.cuh"
+
+DECL_FATTN_VEC_F32_CASE(128, GGML_TYPE_Q4_0, GGML_TYPE_Q4_0);
diff --git a/src/whisper_cpp/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q4_0-q4_1.cu b/src/whisper_cpp/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q4_0-q4_1.cu
new file mode 100644
index 00000000..be5f302d
--- /dev/null
+++ b/src/whisper_cpp/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q4_0-q4_1.cu
@@ -0,0 +1,5 @@
+// This file has been autogenerated by generate_cu_files.py, do not edit manually.
+
+#include "../fattn-vec-f32.cuh"
+
+DECL_FATTN_VEC_F32_CASE(128, GGML_TYPE_Q4_0, GGML_TYPE_Q4_1);
diff --git a/src/whisper_cpp/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q4_0-q5_0.cu b/src/whisper_cpp/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q4_0-q5_0.cu
new file mode 100644
index 00000000..8dd91cd7
--- /dev/null
+++ b/src/whisper_cpp/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q4_0-q5_0.cu
@@ -0,0 +1,5 @@
+// This file has been autogenerated by generate_cu_files.py, do not edit manually.
+
+#include "../fattn-vec-f32.cuh"
+
+DECL_FATTN_VEC_F32_CASE(128, GGML_TYPE_Q4_0, GGML_TYPE_Q5_0);
diff --git a/src/whisper_cpp/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q4_0-q5_1.cu b/src/whisper_cpp/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q4_0-q5_1.cu
new file mode 100644
index 00000000..4cb79150
--- /dev/null
+++ b/src/whisper_cpp/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q4_0-q5_1.cu
@@ -0,0 +1,5 @@
+// This file has been autogenerated by generate_cu_files.py, do not edit manually.
+
+#include "../fattn-vec-f32.cuh"
+
+DECL_FATTN_VEC_F32_CASE(128, GGML_TYPE_Q4_0, GGML_TYPE_Q5_1);
diff --git a/src/whisper_cpp/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q4_0-q8_0.cu b/src/whisper_cpp/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q4_0-q8_0.cu
new file mode 100644
index 00000000..09dea426
--- /dev/null
+++ b/src/whisper_cpp/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q4_0-q8_0.cu
@@ -0,0 +1,5 @@
+// This file has been autogenerated by generate_cu_files.py, do not edit manually.
+
+#include "../fattn-vec-f32.cuh"
+
+DECL_FATTN_VEC_F32_CASE(128, GGML_TYPE_Q4_0, GGML_TYPE_Q8_0);
diff --git a/src/whisper_cpp/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q4_1-f16.cu b/src/whisper_cpp/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q4_1-f16.cu
new file mode 100644
index 00000000..0fbb6076
--- /dev/null
+++ b/src/whisper_cpp/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q4_1-f16.cu
@@ -0,0 +1,5 @@
+// This file has been autogenerated by generate_cu_files.py, do not edit manually.
+
+#include "../fattn-vec-f32.cuh"
+
+DECL_FATTN_VEC_F32_CASE(128, GGML_TYPE_Q4_1, GGML_TYPE_F16);
diff --git a/src/whisper_cpp/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q4_1-q4_0.cu b/src/whisper_cpp/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q4_1-q4_0.cu
new file mode 100644
index 00000000..2aeab83b
--- /dev/null
+++ b/src/whisper_cpp/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q4_1-q4_0.cu
@@ -0,0 +1,5 @@
+// This file has been autogenerated by generate_cu_files.py, do not edit manually.
+
+#include "../fattn-vec-f32.cuh"
+
+DECL_FATTN_VEC_F32_CASE(128, GGML_TYPE_Q4_1, GGML_TYPE_Q4_0);
diff --git a/src/whisper_cpp/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q4_1-q4_1.cu b/src/whisper_cpp/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q4_1-q4_1.cu
new file mode 100644
index 00000000..599415b4
--- /dev/null
+++ b/src/whisper_cpp/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q4_1-q4_1.cu
@@ -0,0 +1,5 @@
+// This file has been autogenerated by generate_cu_files.py, do not edit manually.
+
+#include "../fattn-vec-f32.cuh"
+
+DECL_FATTN_VEC_F32_CASE(128, GGML_TYPE_Q4_1, GGML_TYPE_Q4_1);
diff --git a/src/whisper_cpp/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q4_1-q5_0.cu b/src/whisper_cpp/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q4_1-q5_0.cu
new file mode 100644
index 00000000..e4f8e308
--- /dev/null
+++ b/src/whisper_cpp/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q4_1-q5_0.cu
@@ -0,0 +1,5 @@
+// This file has been autogenerated by generate_cu_files.py, do not edit manually.
+
+#include "../fattn-vec-f32.cuh"
+
+DECL_FATTN_VEC_F32_CASE(128, GGML_TYPE_Q4_1, GGML_TYPE_Q5_0);
diff --git a/src/whisper_cpp/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q4_1-q5_1.cu b/src/whisper_cpp/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q4_1-q5_1.cu
new file mode 100644
index 00000000..34d16652
--- /dev/null
+++ b/src/whisper_cpp/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q4_1-q5_1.cu
@@ -0,0 +1,5 @@
+// This file has been autogenerated by generate_cu_files.py, do not edit manually.
+
+#include "../fattn-vec-f32.cuh"
+
+DECL_FATTN_VEC_F32_CASE(128, GGML_TYPE_Q4_1, GGML_TYPE_Q5_1);
diff --git a/src/whisper_cpp/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q4_1-q8_0.cu b/src/whisper_cpp/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q4_1-q8_0.cu
new file mode 100644
index 00000000..4bebef45
--- /dev/null
+++ b/src/whisper_cpp/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q4_1-q8_0.cu
@@ -0,0 +1,5 @@
+// This file has been autogenerated by generate_cu_files.py, do not edit manually.
+
+#include "../fattn-vec-f32.cuh"
+
+DECL_FATTN_VEC_F32_CASE(128, GGML_TYPE_Q4_1, GGML_TYPE_Q8_0);
diff --git a/src/whisper_cpp/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q5_0-f16.cu b/src/whisper_cpp/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q5_0-f16.cu
new file mode 100644
index 00000000..326468da
--- /dev/null
+++ b/src/whisper_cpp/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q5_0-f16.cu
@@ -0,0 +1,5 @@
+// This file has been autogenerated by generate_cu_files.py, do not edit manually.
+
+#include "../fattn-vec-f32.cuh"
+
+DECL_FATTN_VEC_F32_CASE(128, GGML_TYPE_Q5_0, GGML_TYPE_F16);
diff --git a/src/whisper_cpp/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q5_0-q4_0.cu b/src/whisper_cpp/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q5_0-q4_0.cu
new file mode 100644
index 00000000..511b58f4
--- /dev/null
+++ b/src/whisper_cpp/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q5_0-q4_0.cu
@@ -0,0 +1,5 @@
+// This file has been autogenerated by generate_cu_files.py, do not edit manually.
+
+#include "../fattn-vec-f32.cuh"
+
+DECL_FATTN_VEC_F32_CASE(128, GGML_TYPE_Q5_0, GGML_TYPE_Q4_0);
diff --git a/src/whisper_cpp/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q5_0-q4_1.cu b/src/whisper_cpp/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q5_0-q4_1.cu
new file mode 100644
index 00000000..d9906d14
--- /dev/null
+++ b/src/whisper_cpp/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q5_0-q4_1.cu
@@ -0,0 +1,5 @@
+// This file has been autogenerated by generate_cu_files.py, do not edit manually.
+
+#include "../fattn-vec-f32.cuh"
+
+DECL_FATTN_VEC_F32_CASE(128, GGML_TYPE_Q5_0, GGML_TYPE_Q4_1);
diff --git a/src/whisper_cpp/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q5_0-q5_0.cu b/src/whisper_cpp/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q5_0-q5_0.cu
new file mode 100644
index 00000000..f61c183a
--- /dev/null
+++ b/src/whisper_cpp/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q5_0-q5_0.cu
@@ -0,0 +1,5 @@
+// This file has been autogenerated by generate_cu_files.py, do not edit manually.
+
+#include "../fattn-vec-f32.cuh"
+
+DECL_FATTN_VEC_F32_CASE(128, GGML_TYPE_Q5_0, GGML_TYPE_Q5_0);
diff --git a/src/whisper_cpp/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q5_0-q5_1.cu b/src/whisper_cpp/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q5_0-q5_1.cu
new file mode 100644
index 00000000..c10450fd
--- /dev/null
+++ b/src/whisper_cpp/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q5_0-q5_1.cu
@@ -0,0 +1,5 @@
+// This file has been autogenerated by generate_cu_files.py, do not edit manually.
+
+#include "../fattn-vec-f32.cuh"
+
+DECL_FATTN_VEC_F32_CASE(128, GGML_TYPE_Q5_0, GGML_TYPE_Q5_1);
diff --git a/src/whisper_cpp/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q5_0-q8_0.cu b/src/whisper_cpp/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q5_0-q8_0.cu
new file mode 100644
index 00000000..2d5cb195
--- /dev/null
+++ b/src/whisper_cpp/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q5_0-q8_0.cu
@@ -0,0 +1,5 @@
+// This file has been autogenerated by generate_cu_files.py, do not edit manually.
+
+#include "../fattn-vec-f32.cuh"
+
+DECL_FATTN_VEC_F32_CASE(128, GGML_TYPE_Q5_0, GGML_TYPE_Q8_0);
diff --git a/src/whisper_cpp/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q5_1-f16.cu b/src/whisper_cpp/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q5_1-f16.cu
new file mode 100644
index 00000000..b384f34d
--- /dev/null
+++ b/src/whisper_cpp/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q5_1-f16.cu
@@ -0,0 +1,5 @@
+// This file has been autogenerated by generate_cu_files.py, do not edit manually.
+
+#include "../fattn-vec-f32.cuh"
+
+DECL_FATTN_VEC_F32_CASE(128, GGML_TYPE_Q5_1, GGML_TYPE_F16);
diff --git a/src/whisper_cpp/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q5_1-q4_0.cu b/src/whisper_cpp/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q5_1-q4_0.cu
new file mode 100644
index 00000000..446e293b
--- /dev/null
+++ b/src/whisper_cpp/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q5_1-q4_0.cu
@@ -0,0 +1,5 @@
+// This file has been autogenerated by generate_cu_files.py, do not edit manually.
+
+#include "../fattn-vec-f32.cuh"
+
+DECL_FATTN_VEC_F32_CASE(128, GGML_TYPE_Q5_1, GGML_TYPE_Q4_0);
diff --git a/src/whisper_cpp/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q5_1-q4_1.cu b/src/whisper_cpp/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q5_1-q4_1.cu
new file mode 100644
index 00000000..6f430298
--- /dev/null
+++ b/src/whisper_cpp/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q5_1-q4_1.cu
@@ -0,0 +1,5 @@
+// This file has been autogenerated by generate_cu_files.py, do not edit manually.
+
+#include "../fattn-vec-f32.cuh"
+
+DECL_FATTN_VEC_F32_CASE(128, GGML_TYPE_Q5_1, GGML_TYPE_Q4_1);
diff --git a/src/whisper_cpp/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q5_1-q5_0.cu b/src/whisper_cpp/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q5_1-q5_0.cu
new file mode 100644
index 00000000..1cd8ba88
--- /dev/null
+++ b/src/whisper_cpp/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q5_1-q5_0.cu
@@ -0,0 +1,5 @@
+// This file has been autogenerated by generate_cu_files.py, do not edit manually.
+
+#include "../fattn-vec-f32.cuh"
+
+DECL_FATTN_VEC_F32_CASE(128, GGML_TYPE_Q5_1, GGML_TYPE_Q5_0);
diff --git a/src/whisper_cpp/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q5_1-q5_1.cu b/src/whisper_cpp/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q5_1-q5_1.cu
new file mode 100644
index 00000000..1ee2eab6
--- /dev/null
+++ b/src/whisper_cpp/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q5_1-q5_1.cu
@@ -0,0 +1,5 @@
+// This file has been autogenerated by generate_cu_files.py, do not edit manually.
+
+#include "../fattn-vec-f32.cuh"
+
+DECL_FATTN_VEC_F32_CASE(128, GGML_TYPE_Q5_1, GGML_TYPE_Q5_1);
diff --git a/src/whisper_cpp/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q5_1-q8_0.cu b/src/whisper_cpp/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q5_1-q8_0.cu
new file mode 100644
index 00000000..2bc77816
--- /dev/null
+++ b/src/whisper_cpp/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q5_1-q8_0.cu
@@ -0,0 +1,5 @@
+// This file has been autogenerated by generate_cu_files.py, do not edit manually.
+
+#include "../fattn-vec-f32.cuh"
+
+DECL_FATTN_VEC_F32_CASE(128, GGML_TYPE_Q5_1, GGML_TYPE_Q8_0);
diff --git a/src/whisper_cpp/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q8_0-f16.cu b/src/whisper_cpp/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q8_0-f16.cu
new file mode 100644
index 00000000..d55ced08
--- /dev/null
+++ b/src/whisper_cpp/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q8_0-f16.cu
@@ -0,0 +1,5 @@
+// This file has been autogenerated by generate_cu_files.py, do not edit manually.
+
+#include "../fattn-vec-f32.cuh"
+
+DECL_FATTN_VEC_F32_CASE(128, GGML_TYPE_Q8_0, GGML_TYPE_F16);
diff --git a/src/whisper_cpp/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q8_0-q4_0.cu b/src/whisper_cpp/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q8_0-q4_0.cu
new file mode 100644
index 00000000..8361e99c
--- /dev/null
+++ b/src/whisper_cpp/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q8_0-q4_0.cu
@@ -0,0 +1,5 @@
+// This file has been autogenerated by generate_cu_files.py, do not edit manually.
+
+#include "../fattn-vec-f32.cuh"
+
+DECL_FATTN_VEC_F32_CASE(128, GGML_TYPE_Q8_0, GGML_TYPE_Q4_0);
diff --git a/src/whisper_cpp/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q8_0-q4_1.cu b/src/whisper_cpp/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q8_0-q4_1.cu
new file mode 100644
index 00000000..7507a67c
--- /dev/null
+++ b/src/whisper_cpp/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q8_0-q4_1.cu
@@ -0,0 +1,5 @@
+// This file has been autogenerated by generate_cu_files.py, do not edit manually.
+
+#include "../fattn-vec-f32.cuh"
+
+DECL_FATTN_VEC_F32_CASE(128, GGML_TYPE_Q8_0, GGML_TYPE_Q4_1);
diff --git a/src/whisper_cpp/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q8_0-q5_0.cu b/src/whisper_cpp/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q8_0-q5_0.cu
new file mode 100644
index 00000000..61f050b2
--- /dev/null
+++ b/src/whisper_cpp/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q8_0-q5_0.cu
@@ -0,0 +1,5 @@
+// This file has been autogenerated by generate_cu_files.py, do not edit manually.
+
+#include "../fattn-vec-f32.cuh"
+
+DECL_FATTN_VEC_F32_CASE(128, GGML_TYPE_Q8_0, GGML_TYPE_Q5_0);
diff --git a/src/whisper_cpp/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q8_0-q5_1.cu b/src/whisper_cpp/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q8_0-q5_1.cu
new file mode 100644
index 00000000..d4a49d9c
--- /dev/null
+++ b/src/whisper_cpp/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q8_0-q5_1.cu
@@ -0,0 +1,5 @@
+// This file has been autogenerated by generate_cu_files.py, do not edit manually.
+
+#include "../fattn-vec-f32.cuh"
+
+DECL_FATTN_VEC_F32_CASE(128, GGML_TYPE_Q8_0, GGML_TYPE_Q5_1);
diff --git a/src/whisper_cpp/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q8_0-q8_0.cu b/src/whisper_cpp/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q8_0-q8_0.cu
new file mode 100644
index 00000000..d1462789
--- /dev/null
+++ b/src/whisper_cpp/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs128-q8_0-q8_0.cu
@@ -0,0 +1,5 @@
+// This file has been autogenerated by generate_cu_files.py, do not edit manually.
+
+#include "../fattn-vec-f32.cuh"
+
+DECL_FATTN_VEC_F32_CASE(128, GGML_TYPE_Q8_0, GGML_TYPE_Q8_0);
diff --git a/src/whisper_cpp/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs256-f16-f16.cu b/src/whisper_cpp/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs256-f16-f16.cu
new file mode 100644
index 00000000..e73f917a
--- /dev/null
+++ b/src/whisper_cpp/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs256-f16-f16.cu
@@ -0,0 +1,5 @@
+// This file has been autogenerated by generate_cu_files.py, do not edit manually.
+
+#include "../fattn-vec-f32.cuh"
+
+DECL_FATTN_VEC_F32_CASE(256, GGML_TYPE_F16, GGML_TYPE_F16);
diff --git a/src/whisper_cpp/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs64-f16-f16.cu b/src/whisper_cpp/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs64-f16-f16.cu
new file mode 100644
index 00000000..d40825df
--- /dev/null
+++ b/src/whisper_cpp/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs64-f16-f16.cu
@@ -0,0 +1,5 @@
+// This file has been autogenerated by generate_cu_files.py, do not edit manually.
+
+#include "../fattn-vec-f32.cuh"
+
+DECL_FATTN_VEC_F32_CASE(64, GGML_TYPE_F16, GGML_TYPE_F16);
diff --git a/src/whisper_cpp/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs64-f16-q4_0.cu b/src/whisper_cpp/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs64-f16-q4_0.cu
new file mode 100644
index 00000000..b5c6869f
--- /dev/null
+++ b/src/whisper_cpp/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs64-f16-q4_0.cu
@@ -0,0 +1,5 @@
+// This file has been autogenerated by generate_cu_files.py, do not edit manually.
+
+#include "../fattn-vec-f32.cuh"
+
+DECL_FATTN_VEC_F32_CASE(64, GGML_TYPE_F16, GGML_TYPE_Q4_0);
diff --git a/src/whisper_cpp/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs64-f16-q4_1.cu b/src/whisper_cpp/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs64-f16-q4_1.cu
new file mode 100644
index 00000000..4e21b0cc
--- /dev/null
+++ b/src/whisper_cpp/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs64-f16-q4_1.cu
@@ -0,0 +1,5 @@
+// This file has been autogenerated by generate_cu_files.py, do not edit manually.
+
+#include "../fattn-vec-f32.cuh"
+
+DECL_FATTN_VEC_F32_CASE(64, GGML_TYPE_F16, GGML_TYPE_Q4_1);
diff --git a/src/whisper_cpp/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs64-f16-q5_0.cu b/src/whisper_cpp/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs64-f16-q5_0.cu
new file mode 100644
index 00000000..2eac321b
--- /dev/null
+++ b/src/whisper_cpp/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs64-f16-q5_0.cu
@@ -0,0 +1,5 @@
+// This file has been autogenerated by generate_cu_files.py, do not edit manually.
+
+#include "../fattn-vec-f32.cuh"
+
+DECL_FATTN_VEC_F32_CASE(64, GGML_TYPE_F16, GGML_TYPE_Q5_0);
diff --git a/src/whisper_cpp/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs64-f16-q5_1.cu b/src/whisper_cpp/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs64-f16-q5_1.cu
new file mode 100644
index 00000000..f7d2c3b4
--- /dev/null
+++ b/src/whisper_cpp/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs64-f16-q5_1.cu
@@ -0,0 +1,5 @@
+// This file has been autogenerated by generate_cu_files.py, do not edit manually.
+
+#include "../fattn-vec-f32.cuh"
+
+DECL_FATTN_VEC_F32_CASE(64, GGML_TYPE_F16, GGML_TYPE_Q5_1);
diff --git a/src/whisper_cpp/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs64-f16-q8_0.cu b/src/whisper_cpp/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs64-f16-q8_0.cu
new file mode 100644
index 00000000..a013f400
--- /dev/null
+++ b/src/whisper_cpp/ggml/src/ggml-cuda/template-instances/fattn-vec-f32-instance-hs64-f16-q8_0.cu
@@ -0,0 +1,5 @@
+// This file has been autogenerated by generate_cu_files.py, do not edit manually.
+
+#include "../fattn-vec-f32.cuh"
+
+DECL_FATTN_VEC_F32_CASE(64, GGML_TYPE_F16, GGML_TYPE_Q8_0);
diff --git a/src/whisper_cpp/ggml/src/ggml-cuda/template-instances/fattn-wmma-f16-instance-kqfloat-cpb16.cu b/src/whisper_cpp/ggml/src/ggml-cuda/template-instances/fattn-wmma-f16-instance-kqfloat-cpb16.cu
new file mode 100644
index 00000000..2d94e65c
--- /dev/null
+++ b/src/whisper_cpp/ggml/src/ggml-cuda/template-instances/fattn-wmma-f16-instance-kqfloat-cpb16.cu
@@ -0,0 +1,10 @@
+// This file has been autogenerated by generate_cu_files.py, do not edit manually.
+
+#include "../fattn-wmma-f16.cuh"
+
+DECL_FATTN_WMMA_F16_CASE(64, 16, float);
+DECL_FATTN_WMMA_F16_CASE(80, 16, float);
+DECL_FATTN_WMMA_F16_CASE(96, 16, float);
+DECL_FATTN_WMMA_F16_CASE(112, 16, float);
+DECL_FATTN_WMMA_F16_CASE(128, 16, float);
+DECL_FATTN_WMMA_F16_CASE(256, 16, float);
diff --git a/src/whisper_cpp/ggml/src/ggml-cuda/template-instances/fattn-wmma-f16-instance-kqfloat-cpb32.cu b/src/whisper_cpp/ggml/src/ggml-cuda/template-instances/fattn-wmma-f16-instance-kqfloat-cpb32.cu
new file mode 100644
index 00000000..c3d9df3c
--- /dev/null
+++ b/src/whisper_cpp/ggml/src/ggml-cuda/template-instances/fattn-wmma-f16-instance-kqfloat-cpb32.cu
@@ -0,0 +1,9 @@
+// This file has been autogenerated by generate_cu_files.py, do not edit manually.
+
+#include "../fattn-wmma-f16.cuh"
+
+DECL_FATTN_WMMA_F16_CASE(64, 32, float);
+DECL_FATTN_WMMA_F16_CASE(80, 32, float);
+DECL_FATTN_WMMA_F16_CASE(96, 32, float);
+DECL_FATTN_WMMA_F16_CASE(112, 32, float);
+DECL_FATTN_WMMA_F16_CASE(128, 32, float);
diff --git a/src/whisper_cpp/ggml/src/ggml-cuda/template-instances/fattn-wmma-f16-instance-kqhalf-cpb16.cu b/src/whisper_cpp/ggml/src/ggml-cuda/template-instances/fattn-wmma-f16-instance-kqhalf-cpb16.cu
new file mode 100644
index 00000000..bb680e40
--- /dev/null
+++ b/src/whisper_cpp/ggml/src/ggml-cuda/template-instances/fattn-wmma-f16-instance-kqhalf-cpb16.cu
@@ -0,0 +1,10 @@
+// This file has been autogenerated by generate_cu_files.py, do not edit manually.
+
+#include "../fattn-wmma-f16.cuh"
+
+DECL_FATTN_WMMA_F16_CASE(64, 16, half);
+DECL_FATTN_WMMA_F16_CASE(80, 16, half);
+DECL_FATTN_WMMA_F16_CASE(96, 16, half);
+DECL_FATTN_WMMA_F16_CASE(112, 16, half);
+DECL_FATTN_WMMA_F16_CASE(128, 16, half);
+DECL_FATTN_WMMA_F16_CASE(256, 16, half);
diff --git a/src/whisper_cpp/ggml/src/ggml-cuda/template-instances/fattn-wmma-f16-instance-kqhalf-cpb32.cu b/src/whisper_cpp/ggml/src/ggml-cuda/template-instances/fattn-wmma-f16-instance-kqhalf-cpb32.cu
new file mode 100644
index 00000000..073f71b1
--- /dev/null
+++ b/src/whisper_cpp/ggml/src/ggml-cuda/template-instances/fattn-wmma-f16-instance-kqhalf-cpb32.cu
@@ -0,0 +1,10 @@
+// This file has been autogenerated by generate_cu_files.py, do not edit manually.
+
+#include "../fattn-wmma-f16.cuh"
+
+DECL_FATTN_WMMA_F16_CASE(64, 32, half);
+DECL_FATTN_WMMA_F16_CASE(80, 32, half);
+DECL_FATTN_WMMA_F16_CASE(96, 32, half);
+DECL_FATTN_WMMA_F16_CASE(112, 32, half);
+DECL_FATTN_WMMA_F16_CASE(128, 32, half);
+DECL_FATTN_WMMA_F16_CASE(256, 32, half);
diff --git a/src/whisper_cpp/ggml/src/ggml-cuda/template-instances/fattn-wmma-f16-instance-kqhalf-cpb8.cu b/src/whisper_cpp/ggml/src/ggml-cuda/template-instances/fattn-wmma-f16-instance-kqhalf-cpb8.cu
new file mode 100644
index 00000000..d30710c5
--- /dev/null
+++ b/src/whisper_cpp/ggml/src/ggml-cuda/template-instances/fattn-wmma-f16-instance-kqhalf-cpb8.cu
@@ -0,0 +1,8 @@
+// This file has been autogenerated by generate_cu_files.py, do not edit manually.
+
+#include "../fattn-wmma-f16.cuh"
+
+DECL_FATTN_WMMA_F16_CASE(64, 8, half);
+DECL_FATTN_WMMA_F16_CASE(96, 8, half);
+DECL_FATTN_WMMA_F16_CASE(128, 8, half);
+DECL_FATTN_WMMA_F16_CASE(256, 8, half);
diff --git a/src/whisper_cpp/ggml/src/ggml-cuda/template-instances/generate_cu_files.py b/src/whisper_cpp/ggml/src/ggml-cuda/template-instances/generate_cu_files.py
new file mode 100644
index 00000000..d7874e6e
--- /dev/null
+++ b/src/whisper_cpp/ggml/src/ggml-cuda/template-instances/generate_cu_files.py
@@ -0,0 +1,77 @@
+#!/usr/bin/env python3
+
+from glob import glob
+import os
+
+TYPES_KV = ["GGML_TYPE_Q4_0", "GGML_TYPE_Q4_1", "GGML_TYPE_Q5_0", "GGML_TYPE_Q5_1", "GGML_TYPE_Q8_0", "GGML_TYPE_F16"]
+
+SOURCE_FATTN_VEC = """// This file has been autogenerated by generate_cu_files.py, do not edit manually.
+
+#include "../fattn-vec-f{vkq_size}.cuh"
+
+DECL_FATTN_VEC_F{vkq_size}_CASE({head_size}, {type_k}, {type_v});
+"""
+
+SOURCE_FATTN_WMMA_START = """// This file has been autogenerated by generate_cu_files.py, do not edit manually.
+
+#include "../fattn-wmma-f16.cuh"
+
+"""
+
+SOURCE_FATTN_WMMA_CASE = "DECL_FATTN_WMMA_F16_CASE({head_size}, {cols_per_block}, {kq_acc_t});\n"
+
+TYPES_MMQ = [
+    "GGML_TYPE_Q4_0", "GGML_TYPE_Q4_1", "GGML_TYPE_Q5_0", "GGML_TYPE_Q5_1", "GGML_TYPE_Q8_0",
+    "GGML_TYPE_Q2_K", "GGML_TYPE_Q3_K", "GGML_TYPE_Q4_K", "GGML_TYPE_Q5_K", "GGML_TYPE_Q6_K",
+    "GGML_TYPE_IQ2_XXS", "GGML_TYPE_IQ2_XS", "GGML_TYPE_IQ2_S", "GGML_TYPE_IQ3_XXS", "GGML_TYPE_IQ3_S",
+    "GGML_TYPE_IQ1_S", "GGML_TYPE_IQ4_NL", "GGML_TYPE_IQ4_XS"
+]
+
+SOURCE_MMQ = """// This file has been autogenerated by generate_cu_files.py, do not edit manually.
+
+#include "../mmq.cuh"
+
+DECL_MMQ_CASE({type});
+"""
+
+
+def get_short_name(long_quant_name):
+    return long_quant_name.replace("GGML_TYPE_", "").lower()
+
+
+def get_head_sizes(type_k, type_v):
+    if type_k == "GGML_TYPE_F16" and type_v == "GGML_TYPE_F16":
+        return [64, 128, 256]
+    if type_k == "GGML_TYPE_F16":
+        return [64, 128]
+    return [128]
+
+
+for filename in glob("*.cu"):
+    os.remove(filename)
+
+for vkq_size in [16, 32]:
+    for type_k in TYPES_KV:
+        for type_v in TYPES_KV:
+            for head_size in get_head_sizes(type_k, type_v):
+                with open(f"fattn-vec-f{vkq_size}-instance-hs{head_size}-{get_short_name(type_k)}-{get_short_name(type_v)}.cu", "w") as f:
+                    f.write(SOURCE_FATTN_VEC.format(vkq_size=vkq_size, head_size=head_size, type_k=type_k, type_v=type_v))
+
+for kq_acc_t in ["half", "float"]:
+    for cols_per_block in [8, 16, 32]:
+        if kq_acc_t == "float" and cols_per_block == 8:
+            continue
+
+        with open(f"fattn-wmma-f16-instance-kq{kq_acc_t}-cpb{cols_per_block}.cu", "w") as f:
+            f.write(SOURCE_FATTN_WMMA_START)
+
+            for head_size in [64, 80, 96, 112, 128, 256]:
+                if cols_per_block == 8 and head_size % 32 != 0: # wmma fragment is 8x32
+                    continue
+                if kq_acc_t == "float" and cols_per_block == 32 and head_size == 256: # register spilling, bad performance
+                    continue
+                f.write(SOURCE_FATTN_WMMA_CASE.format(kq_acc_t=kq_acc_t, cols_per_block=cols_per_block, head_size=head_size))
+
+for type in TYPES_MMQ:
+    with open(f"mmq-instance-{get_short_name(type)}.cu", "w") as f:
+        f.write(SOURCE_MMQ.format(type=type))
diff --git a/src/whisper_cpp/ggml/src/ggml-cuda/template-instances/mmq-instance-iq1_s.cu b/src/whisper_cpp/ggml/src/ggml-cuda/template-instances/mmq-instance-iq1_s.cu
new file mode 100644
index 00000000..84ec8502
--- /dev/null
+++ b/src/whisper_cpp/ggml/src/ggml-cuda/template-instances/mmq-instance-iq1_s.cu
@@ -0,0 +1,5 @@
+// This file has been autogenerated by generate_cu_files.py, do not edit manually.
+
+#include "../mmq.cuh"
+
+DECL_MMQ_CASE(GGML_TYPE_IQ1_S);
diff --git a/src/whisper_cpp/ggml/src/ggml-cuda/template-instances/mmq-instance-iq2_s.cu b/src/whisper_cpp/ggml/src/ggml-cuda/template-instances/mmq-instance-iq2_s.cu
new file mode 100644
index 00000000..583c4e5a
--- /dev/null
+++ b/src/whisper_cpp/ggml/src/ggml-cuda/template-instances/mmq-instance-iq2_s.cu
@@ -0,0 +1,5 @@
+// This file has been autogenerated by generate_cu_files.py, do not edit manually.
+
+#include "../mmq.cuh"
+
+DECL_MMQ_CASE(GGML_TYPE_IQ2_S);
diff --git a/src/whisper_cpp/ggml/src/ggml-cuda/template-instances/mmq-instance-iq2_xs.cu b/src/whisper_cpp/ggml/src/ggml-cuda/template-instances/mmq-instance-iq2_xs.cu
new file mode 100644
index 00000000..edaf1560
--- /dev/null
+++ b/src/whisper_cpp/ggml/src/ggml-cuda/template-instances/mmq-instance-iq2_xs.cu
@@ -0,0 +1,5 @@
+// This file has been autogenerated by generate_cu_files.py, do not edit manually.
+
+#include "../mmq.cuh"
+
+DECL_MMQ_CASE(GGML_TYPE_IQ2_XS);
diff --git a/src/whisper_cpp/ggml/src/ggml-cuda/template-instances/mmq-instance-iq2_xxs.cu b/src/whisper_cpp/ggml/src/ggml-cuda/template-instances/mmq-instance-iq2_xxs.cu
new file mode 100644
index 00000000..233d9342
--- /dev/null
+++ b/src/whisper_cpp/ggml/src/ggml-cuda/template-instances/mmq-instance-iq2_xxs.cu
@@ -0,0 +1,5 @@
+// This file has been autogenerated by generate_cu_files.py, do not edit manually.
+
+#include "../mmq.cuh"
+
+DECL_MMQ_CASE(GGML_TYPE_IQ2_XXS);
diff --git a/src/whisper_cpp/ggml/src/ggml-cuda/template-instances/mmq-instance-iq3_s.cu b/src/whisper_cpp/ggml/src/ggml-cuda/template-instances/mmq-instance-iq3_s.cu
new file mode 100644
index 00000000..6092dc71
--- /dev/null
+++ b/src/whisper_cpp/ggml/src/ggml-cuda/template-instances/mmq-instance-iq3_s.cu
@@ -0,0 +1,5 @@
+// This file has been autogenerated by generate_cu_files.py, do not edit manually.
+
+#include "../mmq.cuh"
+
+DECL_MMQ_CASE(GGML_TYPE_IQ3_S);
diff --git a/src/whisper_cpp/ggml/src/ggml-cuda/template-instances/mmq-instance-iq3_xxs.cu b/src/whisper_cpp/ggml/src/ggml-cuda/template-instances/mmq-instance-iq3_xxs.cu
new file mode 100644
index 00000000..1d5bd201
--- /dev/null
+++ b/src/whisper_cpp/ggml/src/ggml-cuda/template-instances/mmq-instance-iq3_xxs.cu
@@ -0,0 +1,5 @@
+// This file has been autogenerated by generate_cu_files.py, do not edit manually.
+
+#include "../mmq.cuh"
+
+DECL_MMQ_CASE(GGML_TYPE_IQ3_XXS);
diff --git a/src/whisper_cpp/ggml/src/ggml-cuda/template-instances/mmq-instance-iq4_nl.cu b/src/whisper_cpp/ggml/src/ggml-cuda/template-instances/mmq-instance-iq4_nl.cu
new file mode 100644
index 00000000..eb02fab0
--- /dev/null
+++ b/src/whisper_cpp/ggml/src/ggml-cuda/template-instances/mmq-instance-iq4_nl.cu
@@ -0,0 +1,5 @@
+// This file has been autogenerated by generate_cu_files.py, do not edit manually.
+
+#include "../mmq.cuh"
+
+DECL_MMQ_CASE(GGML_TYPE_IQ4_NL);
diff --git a/src/whisper_cpp/ggml/src/ggml-cuda/template-instances/mmq-instance-iq4_xs.cu b/src/whisper_cpp/ggml/src/ggml-cuda/template-instances/mmq-instance-iq4_xs.cu
new file mode 100644
index 00000000..1eb3b743
--- /dev/null
+++ b/src/whisper_cpp/ggml/src/ggml-cuda/template-instances/mmq-instance-iq4_xs.cu
@@ -0,0 +1,5 @@
+// This file has been autogenerated by generate_cu_files.py, do not edit manually.
+
+#include "../mmq.cuh"
+
+DECL_MMQ_CASE(GGML_TYPE_IQ4_XS);
diff --git a/src/whisper_cpp/ggml/src/ggml-cuda/template-instances/mmq-instance-q2_k.cu b/src/whisper_cpp/ggml/src/ggml-cuda/template-instances/mmq-instance-q2_k.cu
new file mode 100644
index 00000000..6415369d
--- /dev/null
+++ b/src/whisper_cpp/ggml/src/ggml-cuda/template-instances/mmq-instance-q2_k.cu
@@ -0,0 +1,5 @@
+// This file has been autogenerated by generate_cu_files.py, do not edit manually.
+
+#include "../mmq.cuh"
+
+DECL_MMQ_CASE(GGML_TYPE_Q2_K);
diff --git a/src/whisper_cpp/ggml/src/ggml-cuda/template-instances/mmq-instance-q3_k.cu b/src/whisper_cpp/ggml/src/ggml-cuda/template-instances/mmq-instance-q3_k.cu
new file mode 100644
index 00000000..ffb6213a
--- /dev/null
+++ b/src/whisper_cpp/ggml/src/ggml-cuda/template-instances/mmq-instance-q3_k.cu
@@ -0,0 +1,5 @@
+// This file has been autogenerated by generate_cu_files.py, do not edit manually.
+
+#include "../mmq.cuh"
+
+DECL_MMQ_CASE(GGML_TYPE_Q3_K);
diff --git a/src/whisper_cpp/ggml/src/ggml-cuda/template-instances/mmq-instance-q4_0.cu b/src/whisper_cpp/ggml/src/ggml-cuda/template-instances/mmq-instance-q4_0.cu
new file mode 100644
index 00000000..0c0b0c8a
--- /dev/null
+++ b/src/whisper_cpp/ggml/src/ggml-cuda/template-instances/mmq-instance-q4_0.cu
@@ -0,0 +1,5 @@
+// This file has been autogenerated by generate_cu_files.py, do not edit manually.
+
+#include "../mmq.cuh"
+
+DECL_MMQ_CASE(GGML_TYPE_Q4_0);
diff --git a/src/whisper_cpp/ggml/src/ggml-cuda/template-instances/mmq-instance-q4_1.cu b/src/whisper_cpp/ggml/src/ggml-cuda/template-instances/mmq-instance-q4_1.cu
new file mode 100644
index 00000000..ee67f694
--- /dev/null
+++ b/src/whisper_cpp/ggml/src/ggml-cuda/template-instances/mmq-instance-q4_1.cu
@@ -0,0 +1,5 @@
+// This file has been autogenerated by generate_cu_files.py, do not edit manually.
+
+#include "../mmq.cuh"
+
+DECL_MMQ_CASE(GGML_TYPE_Q4_1);
diff --git a/src/whisper_cpp/ggml/src/ggml-cuda/template-instances/mmq-instance-q4_k.cu b/src/whisper_cpp/ggml/src/ggml-cuda/template-instances/mmq-instance-q4_k.cu
new file mode 100644
index 00000000..9eeb3cd7
--- /dev/null
+++ b/src/whisper_cpp/ggml/src/ggml-cuda/template-instances/mmq-instance-q4_k.cu
@@ -0,0 +1,5 @@
+// This file has been autogenerated by generate_cu_files.py, do not edit manually.
+
+#include "../mmq.cuh"
+
+DECL_MMQ_CASE(GGML_TYPE_Q4_K);
diff --git a/src/whisper_cpp/ggml/src/ggml-cuda/template-instances/mmq-instance-q5_0.cu b/src/whisper_cpp/ggml/src/ggml-cuda/template-instances/mmq-instance-q5_0.cu
new file mode 100644
index 00000000..cc57fb97
--- /dev/null
+++ b/src/whisper_cpp/ggml/src/ggml-cuda/template-instances/mmq-instance-q5_0.cu
@@ -0,0 +1,5 @@
+// This file has been autogenerated by generate_cu_files.py, do not edit manually.
+
+#include "../mmq.cuh"
+
+DECL_MMQ_CASE(GGML_TYPE_Q5_0);
diff --git a/src/whisper_cpp/ggml/src/ggml-cuda/template-instances/mmq-instance-q5_1.cu b/src/whisper_cpp/ggml/src/ggml-cuda/template-instances/mmq-instance-q5_1.cu
new file mode 100644
index 00000000..721ac790
--- /dev/null
+++ b/src/whisper_cpp/ggml/src/ggml-cuda/template-instances/mmq-instance-q5_1.cu
@@ -0,0 +1,5 @@
+// This file has been autogenerated by generate_cu_files.py, do not edit manually.
+
+#include "../mmq.cuh"
+
+DECL_MMQ_CASE(GGML_TYPE_Q5_1);
diff --git a/src/whisper_cpp/ggml/src/ggml-cuda/template-instances/mmq-instance-q5_k.cu b/src/whisper_cpp/ggml/src/ggml-cuda/template-instances/mmq-instance-q5_k.cu
new file mode 100644
index 00000000..a2e90ffd
--- /dev/null
+++ b/src/whisper_cpp/ggml/src/ggml-cuda/template-instances/mmq-instance-q5_k.cu
@@ -0,0 +1,5 @@
+// This file has been autogenerated by generate_cu_files.py, do not edit manually.
+
+#include "../mmq.cuh"
+
+DECL_MMQ_CASE(GGML_TYPE_Q5_K);
diff --git a/src/whisper_cpp/ggml/src/ggml-cuda/template-instances/mmq-instance-q6_k.cu b/src/whisper_cpp/ggml/src/ggml-cuda/template-instances/mmq-instance-q6_k.cu
new file mode 100644
index 00000000..470938fe
--- /dev/null
+++ b/src/whisper_cpp/ggml/src/ggml-cuda/template-instances/mmq-instance-q6_k.cu
@@ -0,0 +1,5 @@
+// This file has been autogenerated by generate_cu_files.py, do not edit manually.
+
+#include "../mmq.cuh"
+
+DECL_MMQ_CASE(GGML_TYPE_Q6_K);
diff --git a/src/whisper_cpp/ggml/src/ggml-cuda/template-instances/mmq-instance-q8_0.cu b/src/whisper_cpp/ggml/src/ggml-cuda/template-instances/mmq-instance-q8_0.cu
new file mode 100644
index 00000000..974477bb
--- /dev/null
+++ b/src/whisper_cpp/ggml/src/ggml-cuda/template-instances/mmq-instance-q8_0.cu
@@ -0,0 +1,5 @@
+// This file has been autogenerated by generate_cu_files.py, do not edit manually.
+
+#include "../mmq.cuh"
+
+DECL_MMQ_CASE(GGML_TYPE_Q8_0);
diff --git a/src/whisper_cpp/ggml/src/ggml-cuda/tsembd.cu b/src/whisper_cpp/ggml/src/ggml-cuda/tsembd.cu
new file mode 100644
index 00000000..153ddbcd
--- /dev/null
+++ b/src/whisper_cpp/ggml/src/ggml-cuda/tsembd.cu
@@ -0,0 +1,47 @@
+#include "tsembd.cuh"
+
+static __global__ void timestep_embedding_f32(const float * timesteps, float * dst, const int nb1, const int dim, const int max_period) {
+    // blockIDx.y: idx of timesteps->ne[0]
+    // blockIDx.x: idx of ((dim + 1) / 2) / BLOCK_SIZE
+    int i = blockIdx.y;
+    int j = threadIdx.x + blockIdx.x * blockDim.x;
+    float * embed_data = (float *)((char *)dst +  i*nb1);
+
+    if (dim % 2 != 0 && j == ((dim + 1) / 2)) {
+        embed_data[dim] = 0.f;
+    }
+
+    int half = dim / 2;
+    if (j >= half) {
+        return;
+    }
+
+    float timestep = timesteps[i];
+    float freq = (float)expf(-logf(max_period) * j / half);
+    float arg = timestep * freq;
+    embed_data[j] = cosf(arg);
+    embed_data[j + half] = sinf(arg);
+}
+
+static void timestep_embedding_f32_cuda(const float * x, float * dst, const int ne00, const int nb1,
+                                        const int dim, const int max_period, cudaStream_t stream) {
+    int half_ceil = (dim + 1) / 2;
+    int num_blocks = (half_ceil + CUDA_TIMESTEP_EMBEDDING_BLOCK_SIZE - 1) / CUDA_TIMESTEP_EMBEDDING_BLOCK_SIZE;
+    dim3 gridDim(num_blocks, ne00, 1);
+    timestep_embedding_f32<<<gridDim, CUDA_TIMESTEP_EMBEDDING_BLOCK_SIZE, 0, stream>>>(x, dst, nb1, dim, max_period);
+}
+
+void ggml_cuda_op_timestep_embedding(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
+    const ggml_tensor * src0 = dst->src[0];
+    const float * src0_d = (const float *)src0->data;
+    float * dst_d = (float *)dst->data;
+    cudaStream_t stream = ctx.stream();
+
+    GGML_ASSERT(src0->type == GGML_TYPE_F32);
+    GGML_ASSERT(dst->type == GGML_TYPE_F32);
+
+    const int dim = dst->op_params[0];
+    const int max_period = dst->op_params[1];
+
+    timestep_embedding_f32_cuda(src0_d, dst_d, src0->ne[0], dst->nb[1], dim, max_period, stream);
+}
diff --git a/src/whisper_cpp/ggml/src/ggml-cuda/tsembd.cuh b/src/whisper_cpp/ggml/src/ggml-cuda/tsembd.cuh
new file mode 100644
index 00000000..84340e3d
--- /dev/null
+++ b/src/whisper_cpp/ggml/src/ggml-cuda/tsembd.cuh
@@ -0,0 +1,5 @@
+#include "common.cuh"
+
+#define CUDA_TIMESTEP_EMBEDDING_BLOCK_SIZE 256
+
+void ggml_cuda_op_timestep_embedding(ggml_backend_cuda_context & ctx, ggml_tensor * dst);
diff --git a/src/whisper_cpp/ggml/src/ggml-cuda/unary.cu b/src/whisper_cpp/ggml/src/ggml-cuda/unary.cu
new file mode 100644
index 00000000..81fc9220
--- /dev/null
+++ b/src/whisper_cpp/ggml/src/ggml-cuda/unary.cu
@@ -0,0 +1,456 @@
+#include "unary.cuh"
+
+static __global__ void neg_f32(const float * x, float * dst, const int k) {
+    const int i = blockDim.x*blockIdx.x + threadIdx.x;
+
+    if (i >= k) {
+        return;
+    }
+
+    dst[i] = -x[i];
+}
+
+static __global__ void step_f32(const float * x, float * dst, const int k) {
+    const int i = blockDim.x*blockIdx.x + threadIdx.x;
+
+    if (i >= k) {
+        return;
+    }
+
+    dst[i] = x[i] > 0.0f;
+}
+
+static __global__ void gelu_f32(const float * x, float * dst, const int k) {
+    const float GELU_COEF_A    = 0.044715f;
+    const float SQRT_2_OVER_PI = 0.79788456080286535587989211986876f;
+    const int i = blockDim.x*blockIdx.x + threadIdx.x;
+
+    if (i >= k) {
+        return;
+    }
+
+    float xi = x[i];
+    dst[i] = 0.5f*xi*(1.0f + tanhf(SQRT_2_OVER_PI*xi*(1.0f + GELU_COEF_A*xi*xi)));
+}
+
+static __global__ void gelu_quick_f32(const float * x, float * dst, int k) {
+    const float GELU_QUICK_COEF = -1.702f;
+    const int i  = blockDim.x*blockIdx.x + threadIdx.x;
+    if (i >= k) {
+        return;
+    }
+    dst[i] = x[i] * (1.0f / (1.0f + expf(GELU_QUICK_COEF * x[i])));
+}
+
+static __global__ void silu_f32(const float * x, float * dst, const int k) {
+    const int i = blockDim.x*blockIdx.x + threadIdx.x;
+
+    if (i >= k) {
+        return;
+    }
+    dst[i] = x[i] / (1.0f + expf(-x[i]));
+}
+
+static __global__ void tanh_f32(const float * x, float * dst, int k) {
+    const int i  = blockDim.x*blockIdx.x + threadIdx.x;
+    if (i >= k) {
+        return;
+    }
+    dst[i] = tanhf(x[i]);
+}
+
+static __global__ void relu_f32(const float * x, float * dst, const int k) {
+    const int i = blockDim.x*blockIdx.x + threadIdx.x;
+
+    if (i >= k) {
+        return;
+    }
+    dst[i] = fmaxf(x[i], 0);
+}
+
+static __global__ void sigmoid_f32(const float * x, float * dst, const int k) {
+    const int i = blockDim.x*blockIdx.x + threadIdx.x;
+
+    if (i >= k) {
+        return;
+    }
+    dst[i] = 1.0f / (1.0f + expf(-x[i]));
+}
+
+static __global__ void hardsigmoid_f32(const float * x, float * dst, const int k) {
+    const int i = blockDim.x*blockIdx.x + threadIdx.x;
+
+    if (i >= k) {
+        return;
+    }
+    dst[i] = fminf(1.0f, fmaxf(0.0f, (x[i] + 3.0f) / 6.0f));
+}
+
+static __global__ void hardswish_f32(const float * x, float * dst, const int k) {
+    const int i = blockDim.x*blockIdx.x + threadIdx.x;
+
+    if (i >= k) {
+        return;
+    }
+    dst[i] = x[i] * fminf(1.0f, fmaxf(0.0f, (x[i] + 3.0f) / 6.0f));
+}
+
+static __global__ void exp_f32(const float * x, float * dst, const int k) {
+    const int i = blockDim.x*blockIdx.x + threadIdx.x;
+
+    if (i >= k) {
+        return;
+    }
+    dst[i] = expf(x[i]);
+}
+
+static __global__ void leaky_relu_f32(const float * x, float * dst, const int k, const float negative_slope) {
+    const int i  = blockDim.x*blockIdx.x + threadIdx.x;
+    if (i >= k) {
+        return;
+    }
+    dst[i] = fmaxf(x[i], 0) + fminf(x[i], 0.0f) * negative_slope;
+}
+
+static __global__ void sqr_f32(const float * x, float * dst, const int k) {
+    const int i = blockDim.x*blockIdx.x + threadIdx.x;
+
+    if (i >= k) {
+        return;
+    }
+    dst[i] = x[i] * x[i];
+}
+
+static __global__ void sqrt_f32(const float * x, float * dst, const int k) {
+    const int i = blockDim.x*blockIdx.x + threadIdx.x;
+
+    if (i >= k) {
+        return;
+    }
+    dst[i] = sqrtf(x[i]);
+}
+
+static __global__ void sin_f32(const float * x, float * dst, const int k) {
+    const int i = blockDim.x*blockIdx.x + threadIdx.x;
+
+    if (i >= k) {
+        return;
+    }
+    dst[i] = sinf(x[i]);
+}
+
+static __global__ void cos_f32(const float * x, float * dst, const int k) {
+    const int i = blockDim.x*blockIdx.x + threadIdx.x;
+
+    if (i >= k) {
+        return;
+    }
+    dst[i] = cosf(x[i]);
+}
+
+static void neg_f32_cuda(const float * x, float * dst, const int k, cudaStream_t stream) {
+    const int num_blocks = (k + CUDA_NEG_BLOCK_SIZE - 1) / CUDA_NEG_BLOCK_SIZE;
+    neg_f32<<<num_blocks, CUDA_NEG_BLOCK_SIZE, 0, stream>>>(x, dst, k);
+}
+
+static void step_f32_cuda(const float * x, float * dst, const int k, cudaStream_t stream) {
+    const int num_blocks = (k + CUDA_STEP_BLOCK_SIZE - 1) / CUDA_STEP_BLOCK_SIZE;
+    step_f32<<<num_blocks, CUDA_STEP_BLOCK_SIZE, 0, stream>>>(x, dst, k);
+}
+
+static void gelu_f32_cuda(const float * x, float * dst, const int k, cudaStream_t stream) {
+    const int num_blocks = (k + CUDA_GELU_BLOCK_SIZE - 1) / CUDA_GELU_BLOCK_SIZE;
+    gelu_f32<<<num_blocks, CUDA_GELU_BLOCK_SIZE, 0, stream>>>(x, dst, k);
+}
+
+static void gelu_quick_f32_cuda(const float * x, float * dst, const int k, cudaStream_t stream) {
+    const int num_blocks = (k + CUDA_GELU_BLOCK_SIZE - 1) / CUDA_GELU_BLOCK_SIZE;
+    gelu_quick_f32<<<num_blocks, CUDA_GELU_BLOCK_SIZE, 0, stream>>>(x, dst, k);
+}
+
+static void silu_f32_cuda(const float * x, float * dst, const int k, cudaStream_t stream) {
+    const int num_blocks = (k + CUDA_SILU_BLOCK_SIZE - 1) / CUDA_SILU_BLOCK_SIZE;
+    silu_f32<<<num_blocks, CUDA_SILU_BLOCK_SIZE, 0, stream>>>(x, dst, k);
+}
+
+static void tanh_f32_cuda(const float * x, float * dst, const int k, cudaStream_t stream) {
+    const int num_blocks = (k + CUDA_TANH_BLOCK_SIZE - 1) / CUDA_TANH_BLOCK_SIZE;
+    tanh_f32<<<num_blocks, CUDA_TANH_BLOCK_SIZE, 0, stream>>>(x, dst, k);
+}
+
+static void relu_f32_cuda(const float * x, float * dst, const int k, cudaStream_t stream) {
+    const int num_blocks = (k + CUDA_RELU_BLOCK_SIZE - 1) / CUDA_RELU_BLOCK_SIZE;
+    relu_f32<<<num_blocks, CUDA_RELU_BLOCK_SIZE, 0, stream>>>(x, dst, k);
+}
+
+static void sigmoid_f32_cuda(const float * x, float * dst, const int k, cudaStream_t stream) {
+    const int num_blocks = (k + CUDA_SIGMOID_BLOCK_SIZE - 1) / CUDA_SIGMOID_BLOCK_SIZE;
+    sigmoid_f32<<<num_blocks, CUDA_SIGMOID_BLOCK_SIZE, 0, stream>>>(x, dst, k);
+}
+
+static void hardsigmoid_f32_cuda(const float * x, float * dst, const int k, cudaStream_t stream) {
+    const int num_blocks = (k + CUDA_HARDSIGMOID_BLOCK_SIZE - 1) / CUDA_HARDSIGMOID_BLOCK_SIZE;
+    hardsigmoid_f32<<<num_blocks, CUDA_HARDSIGMOID_BLOCK_SIZE, 0, stream>>>(x, dst, k);
+}
+
+static void hardswish_f32_cuda(const float * x, float * dst, const int k, cudaStream_t stream) {
+    const int num_blocks = (k + CUDA_HARDSWISH_BLOCK_SIZE - 1) / CUDA_HARDSWISH_BLOCK_SIZE;
+    hardswish_f32<<<num_blocks, CUDA_HARDSWISH_BLOCK_SIZE, 0, stream>>>(x, dst, k);
+}
+
+static void exp_f32_cuda(const float * x, float * dst, const int k, cudaStream_t stream) {
+    const int num_blocks = (k + CUDA_EXP_BLOCK_SIZE - 1) / CUDA_EXP_BLOCK_SIZE;
+    exp_f32<<<num_blocks, CUDA_EXP_BLOCK_SIZE, 0, stream>>>(x, dst, k);
+}
+
+static void leaky_relu_f32_cuda(const float * x, float * dst, const int k, const float negative_slope, cudaStream_t stream) {
+    const int num_blocks = (k + CUDA_RELU_BLOCK_SIZE - 1) / CUDA_RELU_BLOCK_SIZE;
+    leaky_relu_f32<<<num_blocks, CUDA_RELU_BLOCK_SIZE, 0, stream>>>(x, dst, k, negative_slope);
+}
+
+static void sqr_f32_cuda(const float * x, float * dst, const int k, cudaStream_t stream) {
+    const int num_blocks = (k + CUDA_SQR_BLOCK_SIZE - 1) / CUDA_SQR_BLOCK_SIZE;
+    sqr_f32<<<num_blocks, CUDA_SQR_BLOCK_SIZE, 0, stream>>>(x, dst, k);
+}
+
+static void sqrt_f32_cuda(const float * x, float * dst, const int k, cudaStream_t stream) {
+    const int num_blocks = (k + CUDA_SQRT_BLOCK_SIZE - 1) / CUDA_SQRT_BLOCK_SIZE;
+    sqrt_f32<<<num_blocks, CUDA_SQRT_BLOCK_SIZE, 0, stream>>>(x, dst, k);
+}
+
+static void sin_f32_cuda(const float * x, float * dst, const int k, cudaStream_t stream) {
+    const int num_blocks = (k + CUDA_SIN_BLOCK_SIZE - 1) / CUDA_SIN_BLOCK_SIZE;
+    sin_f32<<<num_blocks, CUDA_SIN_BLOCK_SIZE, 0, stream>>>(x, dst, k);
+}
+
+static void cos_f32_cuda(const float * x, float * dst, const int k, cudaStream_t stream) {
+    const int num_blocks = (k + CUDA_COS_BLOCK_SIZE - 1) / CUDA_COS_BLOCK_SIZE;
+    cos_f32<<<num_blocks, CUDA_COS_BLOCK_SIZE, 0, stream>>>(x, dst, k);
+}
+
+void ggml_cuda_op_neg(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
+    const ggml_tensor * src0 = dst->src[0];
+    const float * src0_d = (const float *)src0->data;
+    float * dst_d = (float *)dst->data;
+    cudaStream_t stream = ctx.stream();
+
+    GGML_ASSERT(ggml_is_contiguous(src0));
+
+    GGML_ASSERT(src0->type == GGML_TYPE_F32);
+    GGML_ASSERT( dst->type == GGML_TYPE_F32);
+
+    neg_f32_cuda(src0_d, dst_d, ggml_nelements(src0), stream);
+}
+
+void ggml_cuda_op_step(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
+    const ggml_tensor * src0 = dst->src[0];
+    const float * src0_d = (const float *)src0->data;
+    float * dst_d = (float *)dst->data;
+    cudaStream_t stream = ctx.stream();
+
+    GGML_ASSERT(ggml_is_contiguous(src0));
+
+    GGML_ASSERT(src0->type == GGML_TYPE_F32);
+    GGML_ASSERT( dst->type == GGML_TYPE_F32);
+
+    step_f32_cuda(src0_d, dst_d, ggml_nelements(src0), stream);
+}
+
+void ggml_cuda_op_gelu(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
+    const ggml_tensor * src0 = dst->src[0];
+    const float * src0_d = (const float *)src0->data;
+    float * dst_d = (float *)dst->data;
+    cudaStream_t stream = ctx.stream();
+
+    GGML_ASSERT(ggml_is_contiguous(src0));
+
+    GGML_ASSERT(src0->type == GGML_TYPE_F32);
+    GGML_ASSERT( dst->type == GGML_TYPE_F32);
+
+    gelu_f32_cuda(src0_d, dst_d, ggml_nelements(src0), stream);
+}
+
+void ggml_cuda_op_silu(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
+    const ggml_tensor * src0 = dst->src[0];
+    const float * src0_d = (const float *)src0->data;
+    float * dst_d = (float *)dst->data;
+    cudaStream_t stream = ctx.stream();
+
+    GGML_ASSERT(ggml_is_contiguous(src0));
+
+    GGML_ASSERT(src0->type == GGML_TYPE_F32);
+    GGML_ASSERT( dst->type == GGML_TYPE_F32);
+
+    silu_f32_cuda(src0_d, dst_d, ggml_nelements(src0), stream);
+}
+
+void ggml_cuda_op_gelu_quick(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
+    const ggml_tensor * src0 = dst->src[0];
+    const float * src0_d = (const float *)src0->data;
+    float * dst_d = (float *)dst->data;
+    cudaStream_t stream = ctx.stream();
+
+    GGML_ASSERT(ggml_is_contiguous(src0));
+
+    GGML_ASSERT(src0->type == GGML_TYPE_F32);
+    GGML_ASSERT( dst->type == GGML_TYPE_F32);
+
+    gelu_quick_f32_cuda(src0_d, dst_d, ggml_nelements(src0), stream);
+}
+
+void ggml_cuda_op_tanh(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
+    const ggml_tensor * src0 = dst->src[0];
+    const float * src0_d = (const float *)src0->data;
+    float * dst_d = (float *)dst->data;
+    cudaStream_t stream = ctx.stream();
+
+    GGML_ASSERT(ggml_is_contiguous(src0));
+
+    GGML_ASSERT(src0->type == GGML_TYPE_F32);
+    GGML_ASSERT( dst->type == GGML_TYPE_F32);
+
+    tanh_f32_cuda(src0_d, dst_d, ggml_nelements(src0), stream);
+}
+
+void ggml_cuda_op_relu(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
+    const ggml_tensor * src0 = dst->src[0];
+    const float * src0_d = (const float *)src0->data;
+    float * dst_d = (float *)dst->data;
+    cudaStream_t stream = ctx.stream();
+
+    GGML_ASSERT(ggml_is_contiguous(src0));
+
+    GGML_ASSERT(src0->type == GGML_TYPE_F32);
+    GGML_ASSERT( dst->type == GGML_TYPE_F32);
+
+    relu_f32_cuda(src0_d, dst_d, ggml_nelements(src0), stream);
+}
+
+void ggml_cuda_op_sigmoid(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
+    const ggml_tensor * src0 = dst->src[0];
+    const float * src0_d = (const float *)src0->data;
+    float * dst_d = (float *)dst->data;
+    cudaStream_t stream = ctx.stream();
+
+    GGML_ASSERT(ggml_is_contiguous(src0));
+
+    GGML_ASSERT(src0->type == GGML_TYPE_F32);
+    GGML_ASSERT( dst->type == GGML_TYPE_F32);
+
+    sigmoid_f32_cuda(src0_d, dst_d, ggml_nelements(src0), stream);
+}
+
+void ggml_cuda_op_hardsigmoid(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
+    const ggml_tensor * src0 = dst->src[0];
+    const float * src0_d = (const float *)src0->data;
+    float * dst_d = (float *)dst->data;
+    cudaStream_t stream = ctx.stream();
+
+    GGML_ASSERT(ggml_is_contiguous(src0));
+
+    GGML_ASSERT(src0->type == GGML_TYPE_F32);
+    GGML_ASSERT( dst->type == GGML_TYPE_F32);
+
+    hardsigmoid_f32_cuda(src0_d, dst_d, ggml_nelements(src0), stream);
+}
+
+void ggml_cuda_op_hardswish(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
+    const ggml_tensor * src0 = dst->src[0];
+    const float * src0_d = (const float *)src0->data;
+    float * dst_d = (float *)dst->data;
+    cudaStream_t stream = ctx.stream();
+
+    GGML_ASSERT(ggml_is_contiguous(src0));
+
+    GGML_ASSERT(src0->type == GGML_TYPE_F32);
+    GGML_ASSERT( dst->type == GGML_TYPE_F32);
+
+    hardswish_f32_cuda(src0_d, dst_d, ggml_nelements(src0), stream);
+}
+
+void ggml_cuda_op_exp(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
+    const ggml_tensor * src0 = dst->src[0];
+    const float * src0_d = (const float *)src0->data;
+    float * dst_d = (float *)dst->data;
+    cudaStream_t stream = ctx.stream();
+
+    GGML_ASSERT(ggml_is_contiguous(src0));
+
+    GGML_ASSERT(src0->type == GGML_TYPE_F32);
+    GGML_ASSERT( dst->type == GGML_TYPE_F32);
+
+    exp_f32_cuda(src0_d, dst_d, ggml_nelements(src0), stream);
+}
+
+void ggml_cuda_op_leaky_relu(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
+    const ggml_tensor * src0 = dst->src[0];
+    const float * src0_d = (const float *)src0->data;
+    float * dst_d = (float *)dst->data;
+    cudaStream_t stream = ctx.stream();
+
+    GGML_ASSERT(ggml_is_contiguous(src0));
+
+    GGML_ASSERT(src0->type == GGML_TYPE_F32);
+    GGML_ASSERT( dst->type == GGML_TYPE_F32);
+
+    float negative_slope;
+    memcpy(&negative_slope, dst->op_params, sizeof(float));
+
+    leaky_relu_f32_cuda(src0_d, dst_d, ggml_nelements(src0), negative_slope, stream);
+}
+
+void ggml_cuda_op_sqr(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
+    const ggml_tensor * src0 = dst->src[0];
+    const float * src0_d = (const float *)src0->data;
+    float * dst_d = (float *)dst->data;
+    cudaStream_t stream = ctx.stream();
+
+    GGML_ASSERT(ggml_is_contiguous(src0));
+
+    GGML_ASSERT(src0->type == GGML_TYPE_F32);
+    GGML_ASSERT( dst->type == GGML_TYPE_F32);
+
+    sqr_f32_cuda(src0_d, dst_d, ggml_nelements(src0), stream);
+}
+
+void ggml_cuda_op_sqrt(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
+    const ggml_tensor * src0 = dst->src[0];
+    const float * src0_d = (const float *)src0->data;
+    float * dst_d = (float *)dst->data;
+    cudaStream_t stream = ctx.stream();
+
+    GGML_ASSERT(ggml_is_contiguous(src0));
+
+    GGML_ASSERT(src0->type == GGML_TYPE_F32);
+    GGML_ASSERT( dst->type == GGML_TYPE_F32);
+
+    sqrt_f32_cuda(src0_d, dst_d, ggml_nelements(src0), stream);
+}
+
+void ggml_cuda_op_sin(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
+    const ggml_tensor * src0 = dst->src[0];
+    const float * src0_d = (const float *)src0->data;
+    float * dst_d = (float *)dst->data;
+    cudaStream_t stream = ctx.stream();
+
+    GGML_ASSERT(ggml_is_contiguous(src0));
+
+    GGML_ASSERT(src0->type == GGML_TYPE_F32);
+    GGML_ASSERT( dst->type == GGML_TYPE_F32);
+
+    sin_f32_cuda(src0_d, dst_d, ggml_nelements(src0), stream);
+}
+
+void ggml_cuda_op_cos(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
+    const ggml_tensor * src0 = dst->src[0];
+    const float * src0_d = (const float *)src0->data;
+    float * dst_d = (float *)dst->data;
+    cudaStream_t stream = ctx.stream();
+
+    GGML_ASSERT(ggml_is_contiguous(src0));
+
+    GGML_ASSERT(src0->type == GGML_TYPE_F32);
+    GGML_ASSERT( dst->type == GGML_TYPE_F32);
+
+    cos_f32_cuda(src0_d, dst_d, ggml_nelements(src0), stream);
+}
diff --git a/src/whisper_cpp/ggml/src/ggml-cuda/unary.cuh b/src/whisper_cpp/ggml/src/ggml-cuda/unary.cuh
new file mode 100644
index 00000000..c9193672
--- /dev/null
+++ b/src/whisper_cpp/ggml/src/ggml-cuda/unary.cuh
@@ -0,0 +1,48 @@
+#include "common.cuh"
+
+#define CUDA_NEG_BLOCK_SIZE 256
+#define CUDA_STEP_BLOCK_SIZE 256
+#define CUDA_GELU_BLOCK_SIZE 256
+#define CUDA_SILU_BLOCK_SIZE 256
+#define CUDA_TANH_BLOCK_SIZE 256
+#define CUDA_RELU_BLOCK_SIZE 256
+#define CUDA_SIGMOID_BLOCK_SIZE 256
+#define CUDA_HARDSIGMOID_BLOCK_SIZE 256
+#define CUDA_EXP_BLOCK_SIZE 256
+#define CUDA_HARDSWISH_BLOCK_SIZE 256
+#define CUDA_SQR_BLOCK_SIZE 256
+#define CUDA_SQRT_BLOCK_SIZE 256
+#define CUDA_SIN_BLOCK_SIZE 256
+#define CUDA_COS_BLOCK_SIZE 256
+
+void ggml_cuda_op_neg(ggml_backend_cuda_context & ctx, ggml_tensor * dst);
+
+void ggml_cuda_op_step(ggml_backend_cuda_context & ctx, ggml_tensor * dst);
+
+void ggml_cuda_op_gelu(ggml_backend_cuda_context & ctx, ggml_tensor * dst);
+
+void ggml_cuda_op_silu(ggml_backend_cuda_context & ctx, ggml_tensor * dst);
+
+void ggml_cuda_op_gelu_quick(ggml_backend_cuda_context & ctx, ggml_tensor * dst);
+
+void ggml_cuda_op_tanh(ggml_backend_cuda_context & ctx, ggml_tensor * dst);
+
+void ggml_cuda_op_relu(ggml_backend_cuda_context & ctx, ggml_tensor * dst);
+
+void ggml_cuda_op_sigmoid(ggml_backend_cuda_context & ctx, ggml_tensor * dst);
+
+void ggml_cuda_op_hardsigmoid(ggml_backend_cuda_context & ctx, ggml_tensor * dst);
+
+void ggml_cuda_op_exp(ggml_backend_cuda_context & ctx, ggml_tensor * dst);
+
+void ggml_cuda_op_hardswish(ggml_backend_cuda_context & ctx, ggml_tensor * dst);
+
+void ggml_cuda_op_leaky_relu(ggml_backend_cuda_context & ctx, ggml_tensor * dst);
+
+void ggml_cuda_op_sqr(ggml_backend_cuda_context & ctx, ggml_tensor * dst);
+
+void ggml_cuda_op_sqrt(ggml_backend_cuda_context & ctx, ggml_tensor * dst);
+
+void ggml_cuda_op_sin(ggml_backend_cuda_context & ctx, ggml_tensor * dst);
+
+void ggml_cuda_op_cos(ggml_backend_cuda_context & ctx, ggml_tensor * dst);
diff --git a/src/whisper_cpp/ggml/src/ggml-cuda/upscale.cu b/src/whisper_cpp/ggml/src/ggml-cuda/upscale.cu
new file mode 100644
index 00000000..cf513c3a
--- /dev/null
+++ b/src/whisper_cpp/ggml/src/ggml-cuda/upscale.cu
@@ -0,0 +1,51 @@
+#include "upscale.cuh"
+
+static __global__ void upscale_f32(const float * x, float * dst,
+        const int nb00, const int nb01, const int nb02, const int nb03,
+        const int ne10, const int ne11, const int ne12, const int ne13,
+        const float sf0, const float sf1, const float sf2, const float sf3) {
+    int index = threadIdx.x + blockIdx.x * blockDim.x;
+    if (index >= ne10 * ne11 * ne12 * ne13) {
+        return;
+    }
+
+    int i10 = index % ne10;
+    int i11 = (index / ne10) % ne11;
+    int i12 = (index / (ne10 * ne11)) % ne12;
+    int i13 = (index / (ne10 * ne11 * ne12)) % ne13;
+
+    int i00 = i10 / sf0;
+    int i01 = i11 / sf1;
+    int i02 = i12 / sf2;
+    int i03 = i13 / sf3;
+
+    dst[index] = *(float *)((char *)x + i03 * nb03 + i02 * nb02 + i01 * nb01 + i00 * nb00);
+}
+
+static void upscale_f32_cuda(const float * x, float * dst,
+        const int nb00, const int nb01, const int nb02, const int nb03,
+        const int ne10, const int ne11, const int ne12, const int ne13,
+        const float sf0, const float sf1, const float sf2, const float sf3,
+        cudaStream_t stream) {
+    int dst_size = ne10 * ne11 * ne12 * ne13;
+    int num_blocks = (dst_size + CUDA_UPSCALE_BLOCK_SIZE - 1) / CUDA_UPSCALE_BLOCK_SIZE;
+
+    upscale_f32<<<num_blocks, CUDA_UPSCALE_BLOCK_SIZE,0,stream>>>(x, dst, nb00, nb01, nb02, nb03, ne10, ne11, ne12, ne13, sf0, sf1, sf2, sf3);
+}
+
+void ggml_cuda_op_upscale(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
+    const ggml_tensor * src0 = dst->src[0];
+    const float * src0_d = (const float *)src0->data;
+    float * dst_d = (float *)dst->data;
+    cudaStream_t stream = ctx.stream();
+
+    GGML_ASSERT(src0->type == GGML_TYPE_F32);
+    GGML_ASSERT( dst->type == GGML_TYPE_F32);
+
+    const float sf0 = (float)dst->ne[0]/src0->ne[0];
+    const float sf1 = (float)dst->ne[1]/src0->ne[1];
+    const float sf2 = (float)dst->ne[2]/src0->ne[2];
+    const float sf3 = (float)dst->ne[3]/src0->ne[3];
+
+    upscale_f32_cuda(src0_d, dst_d, src0->nb[0], src0->nb[1], src0->nb[2], src0->nb[3], dst->ne[0], dst->ne[1], dst->ne[2], dst->ne[3], sf0, sf1, sf2, sf3, stream);
+}
diff --git a/src/whisper_cpp/ggml/src/ggml-cuda/upscale.cuh b/src/whisper_cpp/ggml/src/ggml-cuda/upscale.cuh
new file mode 100644
index 00000000..d4d76523
--- /dev/null
+++ b/src/whisper_cpp/ggml/src/ggml-cuda/upscale.cuh
@@ -0,0 +1,5 @@
+#include "common.cuh"
+
+#define CUDA_UPSCALE_BLOCK_SIZE 256
+
+void ggml_cuda_op_upscale(ggml_backend_cuda_context & ctx, ggml_tensor * dst);
diff --git a/src/whisper_cpp/ggml/src/ggml-cuda/vecdotq.cuh b/src/whisper_cpp/ggml/src/ggml-cuda/vecdotq.cuh
new file mode 100644
index 00000000..40091a0e
--- /dev/null
+++ b/src/whisper_cpp/ggml/src/ggml-cuda/vecdotq.cuh
@@ -0,0 +1,1133 @@
+#include "common.cuh"
+#include <cstdint>
+
+static __device__ __forceinline__ int get_int_b2(const void * x, const int & i32) {
+    const uint16_t * x16 = (const uint16_t *) x; // assume at least 2 byte alignment
+
+    int x32  = x16[2*i32 + 0] <<  0;
+    x32     |= x16[2*i32 + 1] << 16;
+
+    return x32;
+}
+
+static __device__ __forceinline__ int get_int_b4(const void * x, const int & i32) {
+    return ((const int *) x)[i32]; // assume at least 4 byte alignment
+}
+
+// VDR = vec dot ratio, how many contiguous integers each thread processes when the vec dot kernel is called
+// MMVQ = mul_mat_vec_q, MMQ = mul_mat_q
+
+#define VDR_Q4_0_Q8_1_MMVQ 2
+#define VDR_Q4_0_Q8_1_MMQ  4
+
+template <int vdr> static __device__ __forceinline__ float vec_dot_q4_0_q8_1_impl(
+    const int * v, const int * u, const float & d4, const half2 & ds8) {
+
+    int sumi = 0;
+
+#pragma unroll
+    for (int i = 0; i < vdr; ++i) {
+        const int vi0 = (v[i] >> 0) & 0x0F0F0F0F;
+        const int vi1 = (v[i] >> 4) & 0x0F0F0F0F;
+
+        // SIMD dot product of quantized values
+        sumi = ggml_cuda_dp4a(vi0, u[2*i+0], sumi);
+        sumi = ggml_cuda_dp4a(vi1, u[2*i+1], sumi);
+    }
+
+    const float2 ds8f = __half22float2(ds8);
+
+    // second part effectively subtracts 8 from each quant value
+    return d4 * (sumi * ds8f.x - (8*vdr/QI4_0) * ds8f.y);
+}
+
+#define VDR_Q4_1_Q8_1_MMVQ 2
+#define VDR_Q4_1_Q8_1_MMQ  4
+
+template <int vdr> static __device__ __forceinline__ float vec_dot_q4_1_q8_1_impl(
+    const int * v, const int * u, const half2 & dm4, const half2 & ds8) {
+
+    int sumi = 0;
+
+#pragma unroll
+    for (int i = 0; i < vdr; ++i) {
+        const int vi0 = (v[i] >> 0) & 0x0F0F0F0F;
+        const int vi1 = (v[i] >> 4) & 0x0F0F0F0F;
+
+        // SIMD dot product of quantized values
+        sumi = ggml_cuda_dp4a(vi0, u[2*i+0], sumi);
+        sumi = ggml_cuda_dp4a(vi1, u[2*i+1], sumi);
+    }
+
+#ifdef GGML_CUDA_F16
+    const float2 tmp = __half22float2(__hmul2(dm4, ds8));
+    const float d4d8 = tmp.x;
+    const float m4s8 = tmp.y;
+#else
+    const float2 dm4f = __half22float2(dm4);
+    const float2 ds8f = __half22float2(ds8);
+    const float d4d8 = dm4f.x * ds8f.x;
+    const float m4s8 = dm4f.y * ds8f.y;
+#endif // GGML_CUDA_F16
+
+    // scale second part of sum by QI8_1/(vdr * QR4_1) to compensate for multiple threads adding it
+    return sumi * d4d8 + m4s8 / (QI8_1 / (vdr * QR4_1));
+}
+
+#define VDR_Q5_0_Q8_1_MMVQ 2
+#define VDR_Q5_0_Q8_1_MMQ  4
+
+template <int vdr> static __device__ __forceinline__ float vec_dot_q5_0_q8_1_impl(
+    const int * vl, const int * vh, const int * u, const float & d5, const half2 & ds8) {
+
+    int sumi = 0;
+
+#pragma unroll
+    for (int i = 0; i < vdr; ++i) {
+        int vi0 = (vl[i] >>  0) & 0x0F0F0F0F; // lower 4 qs bits, still need qh as 5th bits
+        vi0    |= (vh[i] <<  4) & 0x00000010; // 0 ->  4
+        vi0    |= (vh[i] << 11) & 0x00001000; // 1 -> 12
+        vi0    |= (vh[i] << 18) & 0x00100000; // 2 -> 20
+        vi0    |= (vh[i] << 25) & 0x10000000; // 3 -> 28
+        sumi = ggml_cuda_dp4a(vi0, u[2*i+0], sumi); // SIMD dot product of quantized values
+
+        int vi1 = (vl[i] >>  4) & 0x0F0F0F0F; // upper 4 qs bits, still need qh as 5th bits
+        vi1    |= (vh[i] >> 12) & 0x00000010; // 16 ->  4
+        vi1    |= (vh[i] >>  5) & 0x00001000; // 17 -> 12
+        vi1    |= (vh[i] <<  2) & 0x00100000; // 18 -> 20
+        vi1    |= (vh[i] <<  9) & 0x10000000; // 19 -> 28
+        sumi = ggml_cuda_dp4a(vi1, u[2*i+1], sumi); // SIMD dot product of quantized values
+    }
+
+    const float2 ds8f = __half22float2(ds8);
+
+    // second part effectively subtracts 16 from each quant value
+    return d5 * (sumi * ds8f.x - (16*vdr/QI5_0) * ds8f.y);
+}
+
+#define VDR_Q5_1_Q8_1_MMVQ 2
+#define VDR_Q5_1_Q8_1_MMQ  4
+
+template <int vdr> static __device__ __forceinline__ float vec_dot_q5_1_q8_1_impl(
+    const int * vl, const int * vh, const int * u, const half2 & dm5, const half2 & ds8) {
+
+    int sumi = 0;
+
+#pragma unroll
+    for (int i = 0; i < vdr; ++i) {
+        int vi0 = (vl[i] >>  0) & 0x0F0F0F0F; // lower 4 qs bits, still need qh as 5th bits
+        vi0    |= (vh[i] <<  4) & 0x00000010; // 0 ->  4
+        vi0    |= (vh[i] << 11) & 0x00001000; // 1 -> 12
+        vi0    |= (vh[i] << 18) & 0x00100000; // 2 -> 20
+        vi0    |= (vh[i] << 25) & 0x10000000; // 3 -> 28
+        sumi = ggml_cuda_dp4a(vi0, u[2*i+0], sumi); // SIMD dot product of quantized values
+
+        int vi1 = (vl[i] >>  4) & 0x0F0F0F0F; // upper 4 qs bits, still need qh as 5th bits
+        vi1    |= (vh[i] >> 12) & 0x00000010; // 16 ->  4
+        vi1    |= (vh[i] >>  5) & 0x00001000; // 17 -> 12
+        vi1    |= (vh[i] <<  2) & 0x00100000; // 18 -> 20
+        vi1    |= (vh[i] <<  9) & 0x10000000; // 19 -> 28
+        sumi = ggml_cuda_dp4a(vi1, u[2*i+1], sumi); // SIMD dot product of quantized values
+    }
+
+#ifdef GGML_CUDA_F16
+    const float2 tmp = __half22float2(__hmul2(dm5, ds8));
+    const float d5d8 = tmp.x;
+    const float m5s8 = tmp.y;
+#else
+    const float2 dm5f = __half22float2(dm5);
+    const float2 ds8f = __half22float2(ds8);
+    const float d5d8 = dm5f.x * ds8f.x;
+    const float m5s8 = dm5f.y * ds8f.y;
+#endif // GGML_CUDA_F16
+
+    // scale second part of sum by QI5_1 / vdr to compensate for multiple threads adding it
+    return sumi*d5d8 + m5s8 / (QI5_1 / vdr);
+}
+
+#define VDR_Q8_0_Q8_1_MMVQ 2
+#define VDR_Q8_0_Q8_1_MMQ 8
+
+template <typename T, int vdr> static __device__ __forceinline__ T vec_dot_q8_0_q8_1_impl(
+    const int * v, const int * u, const T & d8_0, const T & d8_1) {
+
+    int sumi = 0;
+
+#pragma unroll
+    for (int i = 0; i < vdr; ++i) {
+        // SIMD dot product of quantized values
+        sumi = ggml_cuda_dp4a(v[i], u[i], sumi);
+    }
+
+    return d8_0*d8_1 * ((T) sumi);
+}
+
+template <int vdr> static __device__ __forceinline__ float vec_dot_q8_1_q8_1_impl(
+    const int * v, const int * u, const half2 & dm8, const half2 & ds8) {
+
+    int sumi = 0;
+
+#pragma unroll
+    for (int i = 0; i < vdr; ++i) {
+        // SIMD dot product of quantized values
+        sumi = ggml_cuda_dp4a(v[i], u[i], sumi);
+    }
+
+#ifdef GGML_CUDA_F16
+    const float2 tmp = __half22float2(__hmul2(dm8, ds8));
+    const float d8d8 = tmp.x;
+    const float m8s8 = tmp.y;
+#else
+    const float2 dm8f = __half22float2(dm8);
+    const float2 ds8f = __half22float2(ds8);
+    const float d8d8 = dm8f.x * ds8f.x;
+    const float m8s8 = dm8f.y * ds8f.y;
+#endif // GGML_CUDA_F16
+
+    // scale second part of sum by QI8_1/ vdr to compensate for multiple threads adding it
+    return sumi*d8d8 + m8s8 / (QI8_1 / vdr);
+}
+
+template <int vdr> static __device__ __forceinline__ float vec_dot_q8_0_16_q8_1_impl(
+    const int * v, const int * u, const float * d8_0, const float & d8_1) {
+
+    float sumf = 0.0f;
+
+#pragma unroll
+    for (int i0 = 0; i0 < vdr; i0 += QI8_0/2) {
+        int sumi = 0;
+
+#pragma unroll
+        for (int i = i0; i < i0 + QI8_0/2; ++i) {
+            // SIMD dot product of quantized values
+            sumi = ggml_cuda_dp4a(v[i], u[i], sumi);
+        }
+
+        sumf += d8_0[i0/(QI8_0/2)]*sumi;
+    }
+
+    return d8_1*sumf;
+}
+
+#define VDR_Q2_K_Q8_1_MMVQ 1
+#define VDR_Q2_K_Q8_1_MMQ  4
+
+// contiguous v/x values
+static __device__ __forceinline__ float vec_dot_q2_K_q8_1_impl_mmvq(
+    const int & v, const int * __restrict__ u, const uint8_t * __restrict__ scales,
+    const half2 & dm2, const float * __restrict__ d8) {
+
+    float sumf_d = 0.0f;
+    float sumf_m = 0.0f;
+
+#pragma unroll
+    for (int i = 0; i < QR2_K; ++i) {
+        const int sc = scales[2*i];
+
+        const int vi = (v >> (2*i)) & 0x03030303;
+
+        sumf_d += d8[i] * (ggml_cuda_dp4a(vi, u[i], 0) * (sc & 0xF)); // SIMD dot product
+
+        // fill int with 4x m
+        int m = sc >> 4;
+        m |= m <<  8;
+        m |= m << 16;
+        sumf_m += d8[i] * ggml_cuda_dp4a(m, u[i], 0); // multiply constant q2_K part with sum of q8_1 values
+    }
+
+    const float2 dm2f = __half22float2(dm2);
+
+    return dm2f.x*sumf_d - dm2f.y*sumf_m;
+}
+
+// contiguous v/x + u/y values
+template <int ns8>
+static __device__ __forceinline__ float vec_dot_q2_K_q8_1_impl_mmq(
+    const int * __restrict__ v, const int * __restrict__ u, const half2 * dm2, const float & d8, const half2 * s8) {
+
+    float sumf    = 0.0f;
+    float sumf_d8 = 0.0f;
+
+#pragma unroll
+    for (int i0 = 0; i0 < QR2_K*VDR_Q2_K_Q8_1_MMQ; i0 += QI8_1) {
+        const float2 dm2f0 = __half22float2(dm2[i0/(QI8_1/2) + 0]);
+        int sumi_d0 = 0;
+
+        const float2 dm2f1 = __half22float2(dm2[i0/(QI8_1/2) + 1]);
+        int sumi_d1 = 0;
+
+#pragma unroll
+        for (int i = i0; i < i0 + QI8_1/2; ++i) {
+            sumi_d0 = ggml_cuda_dp4a(v[i], u[i], sumi_d0);
+        }
+        sumf_d8 += dm2f0.x * sumi_d0;
+
+#pragma unroll
+        for (int i = i0 + QI8_1/2; i < i0 + QI8_1; ++i) {
+            sumi_d1 = ggml_cuda_dp4a(v[i], u[i], sumi_d1);
+        }
+        sumf_d8 += dm2f1.x * sumi_d1;
+
+        if (i0/QI8_1 < ns8) {
+            const float2 s8f = __half22float2(s8[i0/QI8_1]);
+            sumf -= dm2f0.y*s8f.x;
+            sumf -= dm2f1.y*s8f.y;
+        } else {
+            int sumi_m0 = 0;
+#pragma unroll
+            for (int i = i0; i < i0 + QI8_1/2; ++i) {
+                sumi_m0 = ggml_cuda_dp4a(0x01010101, u[i], sumi_m0);
+            }
+            sumf_d8 -= dm2f0.y * sumi_m0;
+
+            int sumi_m1 = 0;
+#pragma unroll
+            for (int i = i0 + QI8_1/2; i < i0 + QI8_1; ++i) {
+                sumi_m1 = ggml_cuda_dp4a(0x01010101, u[i], sumi_m1);
+            }
+            sumf_d8 -= dm2f1.y * sumi_m1;
+        }
+    }
+
+    return sumf + d8*sumf_d8;
+}
+
+#define VDR_Q3_K_Q8_1_MMVQ 1
+#define VDR_Q3_K_Q8_1_MMQ  2
+
+// contiguous v/x values
+static __device__ __forceinline__ float vec_dot_q3_K_q8_1_impl_mmvq(
+    const int & vl, const int & vh, const int * __restrict__ u, const uint8_t * __restrict__ scales,
+    const int & scale_offset, const float & d3, const float * __restrict__ d8) {
+
+    float sumf = 0.0f;
+
+#pragma unroll
+    for (int i = 0; i < QR3_K; ++i) {
+        const int isc = scale_offset + 2*i;
+
+        const int isc_low = isc % (QK_K/32);
+        const int sc_shift_low = 4 * (isc / (QK_K/32));
+        const int sc_low  = (scales[isc_low] >> sc_shift_low) & 0xF;
+
+        const int isc_high = isc % (QK_K/64);
+        const int sc_shift_high = 2 * (isc / (QK_K/64));
+        const int sc_high = ((scales[(QK_K/32) + isc_high] >> sc_shift_high) & 3) << 4;
+
+        const int sc = (sc_low | sc_high) - 32;
+
+        const int vil = (vl >> (2*i)) & 0x03030303;
+
+        const int vih = ((vh >> i) << 2) & 0x04040404;
+
+        const int vi = __vsubss4(vil, vih);
+
+        sumf += d8[i] * (ggml_cuda_dp4a(vi, u[i], 0) * sc); // SIMD dot product
+    }
+
+    return d3 * sumf;
+}
+
+// contiguous v/x + u/y values
+static __device__ __forceinline__ float vec_dot_q3_K_q8_1_impl_mmq(
+    const int * __restrict__ v, const int * __restrict__ u, const int8_t * __restrict__ scales,
+    const float & d3, const float & d8) {
+
+    int sumi = 0;
+
+#pragma unroll
+    for (int i0 = 0; i0 < QR3_K*VDR_Q3_K_Q8_1_MMQ; i0 += QI8_1/2) {
+        int sumi_sc = 0;
+
+#pragma unroll
+        for (int i = i0; i < i0 + QI8_1/2; ++i) {
+            sumi_sc = ggml_cuda_dp4a(v[i], u[i], sumi_sc); // SIMD dot product
+        }
+
+        sumi += sumi_sc * scales[i0 / (QI8_1/2)];
+    }
+
+    return d3*d8 * sumi;
+}
+
+#define VDR_Q4_K_Q8_1_MMVQ 2
+#define VDR_Q4_K_Q8_1_MMQ  8
+
+// contiguous v/x values
+static __device__ __forceinline__ float vec_dot_q4_K_q8_1_impl_vmmq(
+    const int * __restrict__ v, const int * __restrict__ u, const uint8_t * __restrict__ sc,
+    const uint8_t * __restrict__ m, const half2 & dm4, const float * __restrict__ d8) {
+
+    float sumf_d = 0.0f;
+    float sumf_m = 0.0f;
+
+#pragma unroll
+    for (int i = 0; i < QR4_K; ++i) {
+        const int v0i = (v[0] >> (4*i)) & 0x0F0F0F0F;
+        const int v1i = (v[1] >> (4*i)) & 0x0F0F0F0F;
+
+        const int dot1 = ggml_cuda_dp4a(v1i, u[2*i+1], ggml_cuda_dp4a(v0i, u[2*i+0], 0)); // SIMD dot product
+        const int dot2 = ggml_cuda_dp4a(0x01010101, u[2*i+1], ggml_cuda_dp4a(0x01010101, u[2*i+0], 0)); // sum of u
+
+        sumf_d += d8[i] * (dot1 * sc[i]);
+        sumf_m += d8[i] * (dot2 * m[i]);  // multiply constant part of q4_K with sum of q8_1 values
+    }
+
+    const float2 dm4f = __half22float2(dm4);
+
+    return dm4f.x*sumf_d - dm4f.y*sumf_m;
+}
+
+// contiguous v/x + u/y values
+static __device__ __forceinline__ float vec_dot_q4_K_q8_1_impl_mmq(
+    const int * __restrict__ v, const int * __restrict__ u, const uint8_t * __restrict__ sc,
+    const uint8_t * __restrict__ m, const half2 & dm4, const half2 * __restrict__ ds8) {
+
+    float sumf_d = 0.0f;
+    float sumf_m = 0.0f;
+
+#pragma unroll
+    for (int i = 0; i < QR4_K*VDR_Q4_K_Q8_1_MMQ/QI8_1; ++i) {
+        int sumi_d = 0;
+
+#pragma unroll
+        for (int j = 0; j < QI8_1; ++j) {
+            sumi_d = ggml_cuda_dp4a((v[j] >> (4*i)) & 0x0F0F0F0F, u[i*QI8_1 + j], sumi_d); // SIMD dot product
+        }
+
+        const float2 ds8f = __half22float2(ds8[i]);
+
+        sumf_d += ds8f.x * (sc[i] * sumi_d);
+        sumf_m += ds8f.y *   m[i]; // sum of q8_1 block * q4_K min val
+    }
+
+    const float2 dm4f = __half22float2(dm4);
+
+    return dm4f.x*sumf_d - dm4f.y*sumf_m;
+}
+
+#define VDR_Q5_K_Q8_1_MMVQ 2
+#define VDR_Q5_K_Q8_1_MMQ  8
+
+// contiguous v/x values
+static __device__ __forceinline__ float vec_dot_q5_K_q8_1_impl_vmmq(
+    const int * __restrict__ vl, const int * __restrict__ vh, const int * __restrict__ u, const uint8_t * __restrict__ sc,
+    const uint8_t * __restrict__ m, const half2 & dm5, const float * __restrict__ d8) {
+
+    float sumf_d = 0.0f;
+    float sumf_m = 0.0f;
+
+#pragma unroll
+    for (int i = 0; i < QR5_K; ++i) {
+        const int vl0i = (vl[0] >> (4*i)) & 0x0F0F0F0F;
+        const int vl1i = (vl[1] >> (4*i)) & 0x0F0F0F0F;
+
+        const int vh0i = ((vh[0] >> i) << 4) & 0x10101010;
+        const int vh1i = ((vh[1] >> i) << 4) & 0x10101010;
+
+        const int v0i = vl0i | vh0i;
+        const int v1i = vl1i | vh1i;
+
+        const int dot1 = ggml_cuda_dp4a(v0i, u[2*i+0], ggml_cuda_dp4a(v1i, u[2*i+1], 0)); // SIMD dot product
+        const int dot2 = ggml_cuda_dp4a(0x01010101, u[2*i+0], ggml_cuda_dp4a(0x01010101, u[2*i+1], 0)); // sum of u
+
+        sumf_d += d8[i] * (dot1 * sc[i]);
+        sumf_m += d8[i] * (dot2 * m[i]);
+
+    }
+
+    const float2 dm5f = __half22float2(dm5);
+
+    return dm5f.x*sumf_d - dm5f.y*sumf_m;
+}
+
+// contiguous v/x + u/y values
+static __device__ __forceinline__ float vec_dot_q5_K_q8_1_impl_mmq(
+    const int * __restrict__ v, const int * __restrict__ u, const uint8_t * __restrict__ sc,
+    const uint8_t * __restrict__ m, const half2 & dm4, const half2 * __restrict__ ds8) {
+
+    float sumf_d = 0.0f;
+    float sumf_m = 0.0f;
+
+#pragma unroll
+    for (int i = 0; i < QR5_K*VDR_Q5_K_Q8_1_MMQ/QI8_1; ++i) {
+        int sumi_d = 0;
+
+#pragma unroll
+        for (int j = 0; j < QI8_1; ++j) {
+            sumi_d = ggml_cuda_dp4a(v[i*QI8_1 + j], u[i*QI8_1 + j], sumi_d); // SIMD dot product
+        }
+
+        const float2 ds8f = __half22float2(ds8[i]);
+
+        sumf_d += ds8f.x * (sc[i] * sumi_d);
+        sumf_m += ds8f.y *   m[i]; // sum of q8_1 block * q4_K min val
+    }
+
+    const float2 dm4f = __half22float2(dm4);
+
+    return dm4f.x*sumf_d - dm4f.y*sumf_m;
+}
+
+#define VDR_Q6_K_Q8_1_MMVQ 1
+#define VDR_Q6_K_Q8_1_MMQ  8
+
+// contiguous v/x values
+static __device__ __forceinline__ float vec_dot_q6_K_q8_1_impl_mmvq(
+    const int & vl, const int & vh, const int * __restrict__ u, const int8_t * __restrict__ scales,
+    const float & d, const float * __restrict__ d8) {
+
+    float sumf = 0.0f;
+
+#pragma unroll
+    for (int i = 0; i < QR6_K; ++i) {
+        const int sc = scales[4*i];
+
+        const int vil = (vl >> (4*i)) & 0x0F0F0F0F;
+
+        const int vih = ((vh >> (4*i)) << 4) & 0x30303030;
+
+        const int vi = __vsubss4((vil | vih), 0x20202020); // vi = (vil | vih) - 32
+
+        sumf += d8[i] * (ggml_cuda_dp4a(vi, u[i], 0) * sc); // SIMD dot product
+    }
+
+    return d*sumf;
+}
+
+// contiguous v/x + u/y values
+static __device__ __forceinline__ float vec_dot_q6_K_q8_1_impl_mmq(
+    const int * __restrict__ v, const int * __restrict__ u, const int8_t * __restrict__ sc,
+    const float & d6, const float * __restrict__ d8) {
+
+    float sumf_d = 0.0f;
+
+    const int      sc_packed = get_int_b4(sc, 0);
+    const int8_t * sc_reg    = (const int8_t *) &sc_packed;
+
+#pragma unroll
+    for (int i0 = 0; i0 < VDR_Q6_K_Q8_1_MMQ; i0 += 4) {
+        int2 sumi_d = {0, 0}; // 2 q6_K scales per q8_1 scale
+
+#pragma unroll
+        for (int i = i0; i < i0 + 2; ++i) {
+            sumi_d.x = ggml_cuda_dp4a(v[2*i+0], u[2*i+0], sumi_d.x); // SIMD dot product
+            sumi_d.x = ggml_cuda_dp4a(v[2*i+1], u[2*i+1], sumi_d.x); // SIMD dot product
+
+            sumi_d.y = ggml_cuda_dp4a(v[2*i+4], u[2*i+4], sumi_d.y); // SIMD dot product
+            sumi_d.y = ggml_cuda_dp4a(v[2*i+5], u[2*i+5], sumi_d.y); // SIMD dot product
+        }
+
+        sumf_d += d8[i0/4] * (sc_reg[i0/2+0]*sumi_d.x + sc_reg[i0/2+1]*sumi_d.y);
+    }
+
+    return d6 * sumf_d;
+}
+
+static __device__ __forceinline__ float vec_dot_q4_0_q8_1(
+    const void * __restrict__ vbq, const block_q8_1 * __restrict__ bq8_1, const int & kbx, const int & iqs) {
+
+    const block_q4_0 * bq4_0 = (const block_q4_0 *) vbq + kbx;
+
+    int v[VDR_Q4_0_Q8_1_MMVQ];
+    int u[2*VDR_Q4_0_Q8_1_MMVQ];
+
+#pragma unroll
+    for (int i = 0; i < VDR_Q4_0_Q8_1_MMVQ; ++i) {
+        v[i]     = get_int_b2(bq4_0->qs, iqs + i);
+        u[2*i+0] = get_int_b4(bq8_1->qs, iqs + i);
+        u[2*i+1] = get_int_b4(bq8_1->qs, iqs + i + QI4_0);
+    }
+
+    return vec_dot_q4_0_q8_1_impl<VDR_Q4_0_Q8_1_MMVQ>(v, u, bq4_0->d, bq8_1->ds);
+}
+
+
+static __device__ __forceinline__ float vec_dot_q4_1_q8_1(
+    const void * __restrict__ vbq, const block_q8_1 * __restrict__ bq8_1, const int & kbx, const int & iqs) {
+
+    const block_q4_1 * bq4_1 = (const block_q4_1 *) vbq + kbx;
+
+    int v[VDR_Q4_1_Q8_1_MMVQ];
+    int u[2*VDR_Q4_1_Q8_1_MMVQ];
+
+#pragma unroll
+    for (int i = 0; i < VDR_Q4_1_Q8_1_MMVQ; ++i) {
+        v[i]     = get_int_b4(bq4_1->qs, iqs + i);
+        u[2*i+0] = get_int_b4(bq8_1->qs, iqs + i);
+        u[2*i+1] = get_int_b4(bq8_1->qs, iqs + i + QI4_1);
+    }
+
+    return vec_dot_q4_1_q8_1_impl<VDR_Q4_1_Q8_1_MMVQ>(v, u, bq4_1->dm, bq8_1->ds);
+}
+
+static __device__ __forceinline__ float vec_dot_q5_0_q8_1(
+    const void * __restrict__ vbq, const block_q8_1 * __restrict__ bq8_1, const int & kbx, const int & iqs) {
+
+    const block_q5_0 * bq5_0 = (const block_q5_0 *) vbq + kbx;
+
+    int vl[VDR_Q5_0_Q8_1_MMVQ];
+    int vh[VDR_Q5_0_Q8_1_MMVQ];
+    int  u[2*VDR_Q5_0_Q8_1_MMVQ];
+
+#pragma unroll
+    for (int i = 0; i < VDR_Q5_0_Q8_1_MMVQ; ++i) {
+        vl[i]    = get_int_b2(bq5_0->qs, iqs + i);
+        vh[i]    = get_int_b2(bq5_0->qh, 0) >> (4 * (iqs + i));
+        u[2*i+0] = get_int_b4(bq8_1->qs, iqs + i);
+        u[2*i+1] = get_int_b4(bq8_1->qs, iqs + i + QI5_0);
+    }
+
+    return vec_dot_q5_0_q8_1_impl<VDR_Q5_0_Q8_1_MMVQ>(vl, vh, u, bq5_0->d, bq8_1->ds);
+}
+
+static __device__ __forceinline__ float vec_dot_q5_1_q8_1(
+    const void * __restrict__ vbq, const block_q8_1 * __restrict__ bq8_1, const int & kbx, const int & iqs) {
+
+    const block_q5_1 * bq5_1 = (const block_q5_1 *) vbq + kbx;
+
+    int vl[VDR_Q5_1_Q8_1_MMVQ];
+    int vh[VDR_Q5_1_Q8_1_MMVQ];
+    int  u[2*VDR_Q5_1_Q8_1_MMVQ];
+
+#pragma unroll
+    for (int i = 0; i < VDR_Q5_1_Q8_1_MMVQ; ++i) {
+        vl[i]    = get_int_b4(bq5_1->qs, iqs + i);
+        vh[i]    = get_int_b4(bq5_1->qh, 0) >> (4 * (iqs + i));
+        u[2*i+0] = get_int_b4(bq8_1->qs, iqs + i);
+        u[2*i+1] = get_int_b4(bq8_1->qs, iqs + i + QI5_1);
+    }
+
+    return vec_dot_q5_1_q8_1_impl<VDR_Q5_1_Q8_1_MMVQ>(vl, vh, u, bq5_1->dm, bq8_1->ds);
+}
+
+static __device__ __forceinline__ float vec_dot_q8_0_q8_1(
+    const void * __restrict__ vbq, const block_q8_1 * __restrict__ bq8_1, const int & kbx, const int & iqs) {
+
+    const block_q8_0 * bq8_0 = (const block_q8_0 *) vbq + kbx;
+
+    int v[VDR_Q8_0_Q8_1_MMVQ];
+    int u[VDR_Q8_0_Q8_1_MMVQ];
+
+#pragma unroll
+    for (int i = 0; i < VDR_Q8_0_Q8_1_MMVQ; ++i) {
+        v[i] = get_int_b2(bq8_0->qs, iqs + i);
+        u[i] = get_int_b4(bq8_1->qs, iqs + i);
+    }
+
+    return vec_dot_q8_0_q8_1_impl<float, VDR_Q8_0_Q8_1_MMVQ>(v, u, bq8_0->d, __low2half(bq8_1->ds));
+}
+
+static __device__ __forceinline__ float vec_dot_q2_K_q8_1(
+    const void * __restrict__ vbq, const block_q8_1 * __restrict__ bq8_1, const int & kbx, const int & iqs) {
+
+    const block_q2_K * bq2_K = (const block_q2_K *) vbq + kbx;
+
+    const int bq8_offset = QR2_K * (iqs / QI8_1);
+    const int scale_offset = iqs - iqs % QI8_1 + (iqs % QI8_1) / (QI8_1/2);
+
+    const uint8_t * scales = bq2_K->scales + scale_offset;
+
+    const int v = get_int_b4(bq2_K->qs, iqs);
+    int    u[QR2_K];
+    float d8[QR2_K];
+
+#pragma unroll
+    for (int i = 0; i < QR2_K; ++ i) {
+        u[i]  = get_int_b4(bq8_1[bq8_offset + i].qs, iqs % QI8_1);
+        d8[i] = __low2float(bq8_1[bq8_offset + i].ds);
+    }
+
+    return vec_dot_q2_K_q8_1_impl_mmvq(v, u, scales, bq2_K->dm, d8);
+}
+
+static __device__ __forceinline__ float vec_dot_q3_K_q8_1(
+    const void * __restrict__ vbq, const block_q8_1 * __restrict__ bq8_1, const int & kbx, const int & iqs) {
+
+    const block_q3_K * bq3_K = (const block_q3_K *) vbq + kbx;
+
+    const int bq8_offset = QR3_K * (iqs / (QI3_K/2));
+    const int scale_offset = iqs - iqs % QI8_1 + (iqs % QI8_1) / (QI8_1/2);
+
+    const float d = bq3_K->d;
+
+    const int vl = get_int_b2(bq3_K->qs, iqs);
+
+    // invert the mask with ~ so that a 0/1 results in 4/0 being subtracted
+    const int vh = ~get_int_b2(bq3_K->hmask, iqs % (QI3_K/2)) >> bq8_offset;
+
+    int    u[QR3_K];
+    float d8[QR3_K];
+
+#pragma unroll
+    for (int i = 0; i < QR3_K; ++i) {
+        u[i]  = get_int_b4(bq8_1[bq8_offset + i].qs, iqs % QI8_1);
+        d8[i] = __low2float(bq8_1[bq8_offset + i].ds);
+    }
+
+    return vec_dot_q3_K_q8_1_impl_mmvq(vl, vh, u, bq3_K->scales, scale_offset, d, d8);
+}
+
+static __device__ __forceinline__ float vec_dot_q4_K_q8_1(
+    const void * __restrict__ vbq, const block_q8_1 * __restrict__ bq8_1, const int & kbx, const int & iqs) {
+
+    const block_q4_K * bq4_K = (const block_q4_K *) vbq + kbx;
+
+    int    v[2];
+    int    u[2*QR4_K];
+    float d8[QR4_K];
+
+    // iqs is in 0,2..30. bq8_offset = iqs/4 -> bq8_offset = 0, 2, 4, 6
+    const int bq8_offset = QR4_K * ((iqs/2) / (QI8_1/2));
+
+    // iqs = 0....3 -> bq8_offset = 0, want q4_offset = 0, 4, 8, 12
+    // iqs = 4....7 -> bq8_offset = 2, want q4_offset = 32, 36, 40, 44
+    // iqs = 8...11 -> bq8_offset = 4, want q4_offset = 64, 68, 72, 76
+    // iqs = 12..15 -> bq8_offset = 6, want q4_offset = 96, 100, 104, 108
+
+    const int * q4 = (const int *)(bq4_K->qs + 16 * bq8_offset + 4 * ((iqs/2)%4));
+    v[0] = q4[0];
+    v[1] = q4[4];
+
+    const uint16_t * scales = (const uint16_t *)bq4_K->scales;
+    uint16_t aux[2];
+    const int j = bq8_offset/2;
+    if (j < 2) {
+        aux[0] = scales[j+0] & 0x3f3f;
+        aux[1] = scales[j+2] & 0x3f3f;
+    } else {
+        aux[0] = ((scales[j+2] >> 0) & 0x0f0f) | ((scales[j-2] & 0xc0c0) >> 2);
+        aux[1] = ((scales[j+2] >> 4) & 0x0f0f) | ((scales[j-0] & 0xc0c0) >> 2);
+    }
+    const uint8_t * sc = (const uint8_t *)aux;
+    const uint8_t * m  = sc + 2;
+
+    for (int i = 0; i < QR4_K; ++i) {
+        const block_q8_1 * bq8i = bq8_1 + bq8_offset + i;
+        d8[i] = __low2float(bq8i->ds);
+
+        const int * q8 = (const int *)bq8i->qs + ((iqs/2)%4);
+        u[2*i+0] = q8[0];
+        u[2*i+1] = q8[4];
+    }
+
+    return vec_dot_q4_K_q8_1_impl_vmmq(v, u, sc, m, bq4_K->dm, d8);
+}
+
+static __device__ __forceinline__ float vec_dot_q5_K_q8_1(
+    const void * __restrict__ vbq, const block_q8_1 * __restrict__ bq8_1, const int & kbx, const int & iqs) {
+
+    const block_q5_K * bq5_K = (const block_q5_K *) vbq + kbx;
+
+    int   vl[2];
+    int   vh[2];
+    int    u[2*QR5_K];
+    float d8[QR5_K];
+
+    const int bq8_offset = QR5_K * ((iqs/2) / (QI8_1/2));
+    const int * ql = (const int *)(bq5_K->qs + 16 * bq8_offset + 4 * ((iqs/2)%4));
+    const int * qh = (const int *)(bq5_K->qh + 4 * ((iqs/2)%4));
+
+    vl[0] = ql[0];
+    vl[1] = ql[4];
+
+    vh[0] = qh[0] >> bq8_offset;
+    vh[1] = qh[4] >> bq8_offset;
+
+    const uint16_t * scales = (const uint16_t *)bq5_K->scales;
+    uint16_t aux[2];
+    const int j = bq8_offset/2;
+    if (j < 2) {
+        aux[0] = scales[j+0] & 0x3f3f;
+        aux[1] = scales[j+2] & 0x3f3f;
+    } else {
+        aux[0] = ((scales[j+2] >> 0) & 0x0f0f) | ((scales[j-2] & 0xc0c0) >> 2);
+        aux[1] = ((scales[j+2] >> 4) & 0x0f0f) | ((scales[j-0] & 0xc0c0) >> 2);
+    }
+    const uint8_t * sc = (const uint8_t *)aux;
+    const uint8_t * m  = sc + 2;
+
+#pragma unroll
+    for (int i = 0; i < QR5_K; ++i) {
+        const block_q8_1 * bq8i = bq8_1 + bq8_offset + i;
+        d8[i] = __low2float(bq8i->ds);
+
+        const int * q8 = (const int *)bq8i->qs + ((iqs/2)%4);
+        u[2*i+0] = q8[0];
+        u[2*i+1] = q8[4];
+    }
+
+    return vec_dot_q5_K_q8_1_impl_vmmq(vl, vh, u, sc, m, bq5_K->dm, d8);
+}
+
+static __device__ __forceinline__ float vec_dot_q6_K_q8_1(
+    const void * __restrict__ vbq, const block_q8_1 * __restrict__ bq8_1, const int & kbx, const int & iqs) {
+
+    const block_q6_K * bq6_K = (const block_q6_K *) vbq + kbx;
+
+    const int bq8_offset = 2 * QR6_K * (iqs / (QI6_K/2)) + (iqs % (QI6_K/2)) / (QI6_K/4);
+    const int scale_offset = (QI6_K/4) * (iqs / (QI6_K/2)) + (iqs % (QI6_K/2)) / (QI6_K/8);
+    const int vh_shift = 2 * ((iqs % (QI6_K/2)) / (QI6_K/4));
+
+    const int vl = get_int_b2(bq6_K->ql, iqs);
+    const int vh = get_int_b2(bq6_K->qh, (QI6_K/4) * (iqs / (QI6_K/2)) + iqs % (QI6_K/4)) >> vh_shift;
+
+    const int8_t * scales = bq6_K->scales + scale_offset;
+
+    int    u[QR6_K];
+    float d8[QR6_K];
+
+#pragma unroll
+    for (int i = 0; i < QR6_K; ++i) {
+        u[i]  = get_int_b4(bq8_1[bq8_offset + 2*i].qs, iqs % QI8_1);
+        d8[i] = __low2float(bq8_1[bq8_offset + 2*i].ds);
+    }
+
+    return vec_dot_q6_K_q8_1_impl_mmvq(vl, vh, u, scales, bq6_K->d, d8);
+}
+
+#define VDR_IQ2_XXS_Q8_1_MMVQ 2
+#define VDR_IQ2_XXS_Q8_1_MMQ  2
+
+static __device__ __forceinline__ float vec_dot_iq2_xxs_q8_1(
+    const void * __restrict__ vbq, const block_q8_1 * __restrict__ bq8_1, const int & kbx, const int & iqs) {
+
+    const block_iq2_xxs * bq2 = (const block_iq2_xxs *) vbq + kbx;
+
+    const int q2 = get_int_b2(bq2->qs, iqs);
+    const uint8_t * aux8 = (const uint8_t *) &q2;
+    const uint32_t aux32 = get_int_b2(bq2->qs, iqs + 1);
+
+    int sumi = 0;
+#pragma unroll
+    for (int k0 = 0; k0 < 8; k0 += 2) {
+        const int * grid_pos = (const int *) (iq2xxs_grid + aux8[k0/2]);
+        const int signs_packed = ksigns_iq2xs[(aux32 >> (7*k0/2)) & 0x7F];
+
+        const int signs0 = __vcmpne4(((signs_packed & 0x03) << 7) | ((signs_packed & 0x0C) << 21), 0x00000000);
+        const int grid0 = __vsub4(grid_pos[0] ^ signs0, signs0);
+        const int u0 = get_int_b4(bq8_1[iqs/2].qs, k0 + 0);
+        sumi = ggml_cuda_dp4a(grid0, u0, sumi);
+
+        const int signs1 = __vcmpne4(((signs_packed & 0x30) << 3) | ((signs_packed & 0xC0) << 17), 0x00000000);
+        const int grid1 = __vsub4(grid_pos[1] ^ signs1, signs1);
+        const int u1 = get_int_b4(bq8_1[iqs/2].qs, k0 + 1);
+        sumi = ggml_cuda_dp4a(grid1, u1, sumi);
+    }
+
+    const int ls = aux32 >> 28;
+    sumi = (ls*sumi + sumi/2)/4;
+    const float d = __half2float(bq2->d) * __low2float(bq8_1[iqs/2].ds);
+    return d * sumi;
+}
+
+#define VDR_IQ2_XS_Q8_1_MMVQ 2
+#define VDR_IQ2_XS_Q8_1_MMQ  2
+
+static __device__ __forceinline__ float vec_dot_iq2_xs_q8_1(
+    const void * __restrict__ vbq, const block_q8_1 * __restrict__ bq8_1, const int & kbx, const int & iqs) {
+
+    const block_iq2_xs * bq2 = (const block_iq2_xs *) vbq + kbx;
+
+    const int2 q2_packed = make_int2(get_int_b2(bq2->qs, iqs + 0), get_int_b2(bq2->qs, iqs + 1));
+    const uint16_t * q2 = (const uint16_t *) &q2_packed;
+    const int ls0 = bq2->scales[iqs/2] & 0x0F;
+    const int ls1 = bq2->scales[iqs/2] >> 4;
+
+    int sumi0 = 0;
+    int sumi1 = 0;
+#pragma unroll
+    for (int l0 = 0; l0 < 8; l0 += 2) {
+        const uint32_t * grid_pos = (const uint32_t *)(iq2xs_grid + (q2[l0/2] & 0x000001FF));
+        const uint32_t * signs    = (const uint32_t *)(ksigns64   + (q2[l0/2] >> 9));
+
+        const int grid_l = __vsub4(grid_pos[0] ^ signs[0], signs[0]);
+        const int grid_h = __vsub4(grid_pos[1] ^ signs[1], signs[1]);
+
+        const int u0 = get_int_b4(bq8_1[iqs/2].qs, l0 + 0);
+        const int u1 = get_int_b4(bq8_1[iqs/2].qs, l0 + 1);
+
+        if (l0 < 4) {
+            sumi0 = ggml_cuda_dp4a(grid_l, u0, sumi0);
+            sumi0 = ggml_cuda_dp4a(grid_h, u1, sumi0);
+        } else {
+            sumi1 = ggml_cuda_dp4a(grid_l, u0, sumi1);
+            sumi1 = ggml_cuda_dp4a(grid_h, u1, sumi1);
+        }
+    }
+    const int sumi = (sumi0*ls0 + sumi1*ls1 + (sumi0 + sumi1)/2)/4;
+    const float d = __half2float(bq2->d) * __low2float(bq8_1[iqs/2].ds);
+    return d * sumi;
+}
+
+#define VDR_IQ2_S_Q8_1_MMVQ 2
+#define VDR_IQ2_S_Q8_1_MMQ  2
+
+static __device__ __forceinline__ float vec_dot_iq2_s_q8_1(
+    const void * __restrict__ vbq, const block_q8_1 * __restrict__ bq8_1, const int & kbx, const int & iqs) {
+
+    const block_iq2_s * bq2 = (const block_iq2_s *) vbq + kbx;
+
+    const int       qs_packed = get_int_b2(bq2->qs, iqs/2);
+    const uint8_t * qs        = (const uint8_t *) &qs_packed;
+
+    const int qh = bq2->qh[iqs/2];
+
+    const int       signs_packed_32 = get_int_b2(bq2->qs, QK_K/32 + iqs/2);
+    const uint8_t * signs_packed_8  = (const uint8_t *) &signs_packed_32;
+
+    const int ls0 = bq2->scales[iqs/2] & 0x0F;
+    const int ls1 = bq2->scales[iqs/2] >> 4;
+
+    int sumi0 = 0;
+    int sumi1 = 0;
+#pragma unroll
+    for (int l0 = 0; l0 < 8; l0 += 2) {
+        const int * grid_pos = (const int *)(iq2s_grid + (qs[l0/2] | ((qh << (8-l0)) & 0x300)));
+
+        const int signs0 = __vcmpne4(((signs_packed_8[l0/2] & 0x03) << 7) | ((signs_packed_8[l0/2] & 0x0C) << 21), 0x00000000);
+        const int signs1 = __vcmpne4(((signs_packed_8[l0/2] & 0x30) << 3) | ((signs_packed_8[l0/2] & 0xC0) << 17), 0x00000000);
+
+        const int grid_l = __vsub4(grid_pos[0] ^ signs0, signs0);
+        const int grid_h = __vsub4(grid_pos[1] ^ signs1, signs1);
+
+        const int u0 = get_int_b4(bq8_1[iqs/2].qs, l0 + 0);
+        const int u1 = get_int_b4(bq8_1[iqs/2].qs, l0 + 1);
+
+        if (l0 < 4) {
+            sumi0 = ggml_cuda_dp4a(grid_l, u0, sumi0);
+            sumi0 = ggml_cuda_dp4a(grid_h, u1, sumi0);
+        } else {
+            sumi1 = ggml_cuda_dp4a(grid_l, u0, sumi1);
+            sumi1 = ggml_cuda_dp4a(grid_h, u1, sumi1);
+        }
+    }
+    const int sumi = (sumi0*ls0 + sumi1*ls1 + (sumi0 + sumi1)/2)/4;
+
+    const float d = __half2float(bq2->d) * __low2float(bq8_1[iqs/2].ds);
+    return d * sumi;
+}
+
+#define VDR_IQ3_XXS_Q8_1_MMVQ 2
+#define VDR_IQ3_XXS_Q8_1_MMQ  2
+
+static __device__ __forceinline__ float vec_dot_iq3_xxs_q8_1(
+    const void * __restrict__ vbq, const block_q8_1 * __restrict__ bq8_1, const int & kbx, const int & iqs) {
+
+    const block_iq3_xxs * bq3 = (const block_iq3_xxs *) vbq + kbx;
+
+    const int2 q3_packed = make_int2(get_int_b2(bq3->qs, iqs), get_int_b2(bq3->qs, iqs+1));
+    const uint8_t * q3 = (const uint8_t *) &q3_packed;
+    const uint32_t aux32 = get_int_b2(bq3->qs, QK_K/16 + iqs/2);
+
+    int sumi = 0;
+#pragma unroll
+    for (int l0 = 0; l0 < 8; l0 += 2) {
+        const int2 grid_pos = make_int2(iq3xxs_grid[q3[l0 + 0]], iq3xxs_grid[q3[l0 + 1]]);
+
+        const int * signs = (const int *)(ksigns64 + ((aux32 >> (7*l0/2)) & 0x7F));
+
+        const int grid_l = __vsub4(grid_pos.x ^ signs[0], signs[0]);
+        const int grid_h = __vsub4(grid_pos.y ^ signs[1], signs[1]);
+
+        const int u0 = get_int_b4(bq8_1[iqs/2].qs, l0 + 0);
+        const int u1 = get_int_b4(bq8_1[iqs/2].qs, l0 + 1);
+
+        sumi = ggml_cuda_dp4a(grid_l, u0, sumi);
+        sumi = ggml_cuda_dp4a(grid_h, u1, sumi);
+    }
+
+    const int ls = aux32 >> 28;
+    sumi = (ls*sumi + sumi/2)/2;
+    const float d = __half2float(bq3->d) * __low2float(bq8_1[iqs/2].ds);
+    return d * sumi;
+}
+
+#define VDR_IQ3_S_Q8_1_MMVQ 2
+#define VDR_IQ3_S_Q8_1_MMQ  2
+
+// TODO: don't use lookup table for signs
+static __device__ __forceinline__ float vec_dot_iq3_s_q8_1(
+    const void * __restrict__ vbq, const block_q8_1 * __restrict__ bq8_1, const int & kbx, const int & iqs) {
+
+    const block_iq3_s * bq3 = (const block_iq3_s *) vbq + kbx;
+
+    const int2      qs_packed = make_int2(get_int_b2(bq3->qs, iqs + 0), get_int_b2(bq3->qs, iqs + 1));
+    const uint8_t * qs        = (const uint8_t *) &qs_packed;
+
+    const int qh = bq3->qh[iqs/2];
+
+    const int       signs_packed_32 = get_int_b2(bq3->signs, iqs/2);
+    const uint8_t * signs_packed_8  = (const uint8_t *) &signs_packed_32;
+
+    int sumi = 0;
+#pragma unroll
+    for (int l0 = 0; l0 < 8; l0 += 2) {
+        const int2 grid_pos = make_int2(
+            iq3s_grid[qs[l0 + 0] | ((qh << (8 - l0)) & 0x100)],
+            iq3s_grid[qs[l0 + 1] | ((qh << (7 - l0)) & 0x100)]);
+
+        const int signs0 = __vcmpne4(((signs_packed_8[l0/2] & 0x03) << 7) | ((signs_packed_8[l0/2] & 0x0C) << 21), 0x00000000);
+        const int signs1 = __vcmpne4(((signs_packed_8[l0/2] & 0x30) << 3) | ((signs_packed_8[l0/2] & 0xC0) << 17), 0x00000000);
+
+        const int grid_l = __vsub4(grid_pos.x ^ signs0, signs0);
+        const int grid_h = __vsub4(grid_pos.y ^ signs1, signs1);
+
+        const int u0 = get_int_b4(bq8_1[iqs/2].qs, l0 + 0);
+        const int u1 = get_int_b4(bq8_1[iqs/2].qs, l0 + 1);
+
+        sumi = ggml_cuda_dp4a(grid_l, u0, sumi);
+        sumi = ggml_cuda_dp4a(grid_h, u1, sumi);
+    }
+
+    sumi *= 1 + 2*((bq3->scales[iqs/4] >> ((iqs << 1) & 0x04)) & 0x0F);
+
+    const float d = __half2float(bq3->d) * __low2float(bq8_1[iqs/2].ds);
+    return d * sumi;
+}
+
+#define VDR_IQ1_S_Q8_1_MMVQ 1
+#define VDR_IQ1_S_Q8_1_MMQ  1
+
+static __device__ __forceinline__ float vec_dot_iq1_s_q8_1(
+    const void * __restrict__ vbq, const block_q8_1 * __restrict__ bq8_1, const int & kbx, const int & iqs) {
+    const block_iq1_s * bq1 = (const block_iq1_s *) vbq + kbx;
+
+    const int       qs_packed = get_int_b2(bq1->qs, iqs);
+    const uint8_t * qs        = (const uint8_t *) &qs_packed;
+
+    const int qh = bq1->qh[iqs];
+
+    int sumi = 0;
+#pragma unroll
+    for (int l0 = 0; l0 < 8; l0 += 2) {
+        const int grid = iq1s_grid_gpu[qs[l0/2] | (((qh >> 3*(l0/2)) & 0x07) << 8)];
+
+        const int grid0 = (grid >> 0) & 0x0F0F0F0F;
+        const int grid1 = (grid >> 4) & 0x0F0F0F0F;
+
+        const int u0 = get_int_b4(bq8_1[iqs].qs, l0 + 0);
+        const int u1 = get_int_b4(bq8_1[iqs].qs, l0 + 1);
+
+        sumi = ggml_cuda_dp4a(grid0, u0, sumi);
+        sumi = ggml_cuda_dp4a(grid1, u1, sumi);
+    }
+
+    const float  d1q   = __half2float(bq1->d) * (((qh >> 11) & 0x0E) + 1);
+    const float  delta = -1.0f + IQ1S_DELTA - (qh & 0x8000) * (2.0f*IQ1S_DELTA/0x8000);
+    const float2 ds    = __half22float2(bq8_1[iqs].ds);
+    return d1q * (ds.x*sumi + ds.y*delta);
+}
+
+#define VDR_IQ1_M_Q8_1_MMVQ 1
+#define VDR_IQ1_M_Q8_1_MMQ  1
+
+static __device__ __forceinline__ float vec_dot_iq1_m_q8_1(
+    const void * __restrict__ vbq, const block_q8_1 * __restrict__ bq8_1, const int & kbx, const int & iqs) {
+
+    const block_iq1_m * bq1 = (const block_iq1_m *) vbq + kbx;
+
+    const int       qs_packed = get_int_b4(bq1->qs, iqs);
+    const uint8_t * qs        = (const uint8_t *) &qs_packed;
+
+    int   sumi[2] = {0};
+    float sumf[2] = {0.0f};
+#pragma unroll
+    for (int l0 = 0; l0 < 8; l0 += 2) {
+        const int qhl = bq1->qh[2*iqs + l0/4] >> (4 * ((l0/2) % 2));
+
+        const int grid = iq1s_grid_gpu[qs[l0/2] | ((qhl & 0x07) << 8)];
+
+        const int grid0 = (grid >> 0) & 0x0F0F0F0F;
+        const int grid1 = (grid >> 4) & 0x0F0F0F0F;
+
+        const int u0 = get_int_b4(bq8_1[iqs].qs, l0 + 0);
+        const int u1 = get_int_b4(bq8_1[iqs].qs, l0 + 1);
+
+        sumi[l0/4] = ggml_cuda_dp4a(grid0, u0, sumi[l0/4]);
+        sumi[l0/4] = ggml_cuda_dp4a(grid1, u1, sumi[l0/4]);
+
+        const float delta = -1.0f + IQ1M_DELTA - (qhl & 0x08) * (2.0f*IQ1M_DELTA/0x08);
+        int sumy = 0;
+        sumy = ggml_cuda_dp4a(u0, 0x01010101, sumy);
+        sumy = ggml_cuda_dp4a(u1, 0x01010101, sumy);
+        sumf[l0/4] += delta*sumy;
+    }
+
+    const uint16_t * sc = (const uint16_t *) bq1->scales;
+
+    iq1m_scale_t scale;
+    scale.u16 = (sc[0] >> 12) | ((sc[1] >> 8) & 0x00F0) | ((sc[2] >> 4) & 0x0F00) | (sc[3] & 0xF000);
+    const float d = __half2float(scale.f16) * __low2float(bq8_1[iqs].ds);
+
+    const int tmp = sc[iqs/2] >> (6*(iqs%2));
+    const int sc0 = 2*((tmp >> 0) & 0x07) + 1;
+    const int sc1 = 2*((tmp >> 3) & 0x07) + 1;
+    return d * ((sumi[0] + sumf[0]) * sc0 + (sumi[1] + sumf[1]) * sc1);
+}
+
+static __device__ __forceinline__ int2 get_int_from_table_16(const int & q4) {
+    const int      q0_32  = (q4 >> 0) & 0x0F0F0F0F;
+    const int8_t * q0_8   = (const int8_t *) &q0_32;
+    const char4    val0_8 = make_char4(
+        kvalues_iq4nl[q0_8[0]], kvalues_iq4nl[q0_8[1]], kvalues_iq4nl[q0_8[2]], kvalues_iq4nl[q0_8[3]]);
+
+    const int      q1_32  = (q4 >> 4) & 0x0F0F0F0F;
+    const int8_t * q1_8   = (const int8_t *) &q1_32;
+    const char4    val1_8 = make_char4(
+        kvalues_iq4nl[q1_8[0]], kvalues_iq4nl[q1_8[1]], kvalues_iq4nl[q1_8[2]], kvalues_iq4nl[q1_8[3]]);
+
+    return make_int2(*((const int *) &val0_8), *((const int *) &val1_8));
+}
+
+#define VDR_IQ4_NL_Q8_1_MMVQ 2
+#define VDR_IQ4_NL_Q8_1_MMQ  4
+
+static __device__ __forceinline__ float vec_dot_iq4_nl_q8_1(
+    const void * __restrict__ vbq, const block_q8_1 * __restrict__ bq8_1, const int & kbx, const int & iqs) {
+
+    const block_iq4_nl * bq4 = (const block_iq4_nl *) vbq + kbx;
+
+    const int * q8 = (const int *) bq8_1->qs + iqs;
+
+    int sumi = 0;
+#pragma unroll
+    for (int l = 0; l < VDR_Q4_0_Q8_1_MMVQ; ++l) {
+        const int aux_q4 = get_int_b2(bq4->qs, iqs + l);
+        const int2 v = get_int_from_table_16(aux_q4);
+
+        sumi = ggml_cuda_dp4a(v.x, q8[l + 0], sumi);
+        sumi = ggml_cuda_dp4a(v.y, q8[l + 4], sumi);
+    }
+
+    const float d = __half2float(bq4->d) * __low2float(bq8_1->ds);
+    return d * sumi;
+}
+
+#define VDR_IQ4_XS_Q8_1_MMVQ 4
+#define VDR_IQ4_XS_Q8_1_MMQ  4
+
+static __device__ __forceinline__ float vec_dot_iq4_xs_q8_1(
+    const void * __restrict__ vbq, const block_q8_1 * __restrict__ bq8_1, const int & kbx, const int & iqs) {
+
+    const block_iq4_xs * bq4 = (const block_iq4_xs *) vbq + kbx;
+
+    int sumi = 0;
+#pragma unroll
+    for (int j = 0; j < 4; ++j) {
+        const int aux_q4 = get_int_b4(bq4->qs, iqs + j);
+        const int2 v = get_int_from_table_16(aux_q4);
+
+        const int u0 = get_int_b4(bq8_1[iqs/4].qs, j + 0);
+        const int u1 = get_int_b4(bq8_1[iqs/4].qs, j + 4);
+
+        sumi = ggml_cuda_dp4a(v.x, u0, sumi);
+        sumi = ggml_cuda_dp4a(v.y, u1, sumi);
+    }
+
+    const int ls = ((bq4->scales_l[iqs/8] >> (iqs & 0x04)) & 0x0F) | (((bq4->scales_h >> (iqs/2)) & 0x03) << 4);
+    sumi *= ls - 32;
+
+    const float d = __half2float(bq4->d) * __low2float(bq8_1[iqs/4].ds);
+    return d * sumi;
+}
diff --git a/src/whisper_cpp/ggml/src/ggml-cuda/vendors/cuda.h b/src/whisper_cpp/ggml/src/ggml-cuda/vendors/cuda.h
new file mode 100644
index 00000000..db9f6a16
--- /dev/null
+++ b/src/whisper_cpp/ggml/src/ggml-cuda/vendors/cuda.h
@@ -0,0 +1,14 @@
+#pragma once
+
+#include <cuda_runtime.h>
+#include <cuda.h>
+#include <cublas_v2.h>
+#include <cuda_fp16.h>
+
+#if CUDART_VERSION < 11020
+#define CU_DEVICE_ATTRIBUTE_VIRTUAL_MEMORY_MANAGEMENT_SUPPORTED CU_DEVICE_ATTRIBUTE_VIRTUAL_ADDRESS_MANAGEMENT_SUPPORTED
+#define CUBLAS_TF32_TENSOR_OP_MATH CUBLAS_TENSOR_OP_MATH
+#define CUBLAS_COMPUTE_16F CUDA_R_16F
+#define CUBLAS_COMPUTE_32F CUDA_R_32F
+#define cublasComputeType_t cudaDataType_t
+#endif // CUDART_VERSION < 11020
diff --git a/src/whisper_cpp/ggml/src/ggml-cuda/vendors/hip.h b/src/whisper_cpp/ggml/src/ggml-cuda/vendors/hip.h
new file mode 100644
index 00000000..1f3c70c2
--- /dev/null
+++ b/src/whisper_cpp/ggml/src/ggml-cuda/vendors/hip.h
@@ -0,0 +1,178 @@
+#pragma once
+
+#include <hip/hip_runtime.h>
+#include <hipblas/hipblas.h>
+#include <hip/hip_fp16.h>
+#ifdef __HIP_PLATFORM_AMD__
+// for rocblas_initialize()
+#include "rocblas/rocblas.h"
+#endif // __HIP_PLATFORM_AMD__
+#define CUBLAS_COMPUTE_16F HIPBLAS_R_16F
+#define CUBLAS_COMPUTE_32F HIPBLAS_R_32F
+#define CUBLAS_COMPUTE_32F_FAST_16F HIPBLAS_R_32F
+#define CUBLAS_GEMM_DEFAULT HIPBLAS_GEMM_DEFAULT
+#define CUBLAS_GEMM_DEFAULT_TENSOR_OP HIPBLAS_GEMM_DEFAULT
+#define CUBLAS_OP_N HIPBLAS_OP_N
+#define CUBLAS_OP_T HIPBLAS_OP_T
+#define CUBLAS_STATUS_SUCCESS HIPBLAS_STATUS_SUCCESS
+#define CUBLAS_TF32_TENSOR_OP_MATH 0
+#define CUDA_R_16F  HIPBLAS_R_16F
+#define CUDA_R_32F  HIPBLAS_R_32F
+#define __shfl_xor_sync(mask, var, laneMask, width) __shfl_xor(var, laneMask, width)
+#define cublasComputeType_t hipblasDatatype_t //deprecated, new hipblasComputeType_t not in 5.6
+#define cublasCreate hipblasCreate
+#define cublasDestroy hipblasDestroy
+#define cublasGemmEx hipblasGemmEx
+#define cublasGemmBatchedEx hipblasGemmBatchedEx
+#define cublasGemmStridedBatchedEx hipblasGemmStridedBatchedEx
+#define cublasHandle_t hipblasHandle_t
+#define cublasSetMathMode(handle, mode) CUBLAS_STATUS_SUCCESS
+#define cublasSetStream hipblasSetStream
+#define cublasSgemm hipblasSgemm
+#define cublasStatus_t hipblasStatus_t
+#define cublasOperation_t hipblasOperation_t
+#define cudaDataType_t hipblasDatatype_t //deprecated, new hipblasDatatype not in 5.6
+#define cudaDeviceCanAccessPeer hipDeviceCanAccessPeer
+#define cudaDeviceDisablePeerAccess hipDeviceDisablePeerAccess
+#define cudaDeviceEnablePeerAccess hipDeviceEnablePeerAccess
+#define cudaDeviceProp hipDeviceProp_t
+#define cudaDeviceSynchronize hipDeviceSynchronize
+#define cudaError_t hipError_t
+#define cudaErrorPeerAccessAlreadyEnabled hipErrorPeerAccessAlreadyEnabled
+#define cudaErrorPeerAccessNotEnabled hipErrorPeerAccessNotEnabled
+#define cudaEventCreateWithFlags hipEventCreateWithFlags
+#define cudaEventDisableTiming hipEventDisableTiming
+#define cudaEventRecord hipEventRecord
+#define cudaEventSynchronize hipEventSynchronize
+#define cudaEvent_t hipEvent_t
+#define cudaEventDestroy hipEventDestroy
+#define cudaFree hipFree
+#define cudaFreeHost hipHostFree
+#define cudaGetDevice hipGetDevice
+#define cudaGetDeviceCount hipGetDeviceCount
+#define cudaGetDeviceProperties hipGetDeviceProperties
+#define cudaGetErrorString hipGetErrorString
+#define cudaGetLastError hipGetLastError
+#define cudaHostRegister hipHostRegister
+#define cudaHostRegisterPortable hipHostRegisterPortable
+#define cudaHostRegisterReadOnly hipHostRegisterReadOnly
+#define cudaHostUnregister hipHostUnregister
+#define cudaLaunchHostFunc hipLaunchHostFunc
+#define cudaMalloc hipMalloc
+#define cudaMallocHost(ptr, size) hipHostMalloc(ptr, size, hipHostMallocDefault)
+#define cudaMemcpy hipMemcpy
+#define cudaMemcpyAsync hipMemcpyAsync
+#define cudaMemcpyPeerAsync hipMemcpyPeerAsync
+#define cudaMemcpy2DAsync hipMemcpy2DAsync
+#define cudaMemcpyDeviceToDevice hipMemcpyDeviceToDevice
+#define cudaMemcpyDeviceToHost hipMemcpyDeviceToHost
+#define cudaMemcpyHostToDevice hipMemcpyHostToDevice
+#define cudaMemcpyKind hipMemcpyKind
+#define cudaMemset hipMemset
+#define cudaMemsetAsync hipMemsetAsync
+#define cudaMemGetInfo hipMemGetInfo
+#define cudaOccupancyMaxPotentialBlockSize hipOccupancyMaxPotentialBlockSize
+#define cudaSetDevice hipSetDevice
+#define cudaStreamCreateWithFlags hipStreamCreateWithFlags
+#define cudaStreamDestroy hipStreamDestroy
+#define cudaStreamFireAndForget hipStreamFireAndForget
+#define cudaStreamNonBlocking hipStreamNonBlocking
+#define cudaStreamPerThread hipStreamPerThread
+#define cudaStreamSynchronize hipStreamSynchronize
+#define cudaStreamWaitEvent(stream, event, flags) hipStreamWaitEvent(stream, event, flags)
+#define cudaStream_t hipStream_t
+#define cudaSuccess hipSuccess
+#define __trap() do { abort(); __builtin_unreachable(); } while(0)
+#define CUBLAS_STATUS_SUCCESS HIPBLAS_STATUS_SUCCESS
+#define CUBLAS_STATUS_NOT_INITIALIZED HIPBLAS_STATUS_NOT_INITIALIZED
+#define CUBLAS_STATUS_ALLOC_FAILED HIPBLAS_STATUS_ALLOC_FAILED
+#define CUBLAS_STATUS_INVALID_VALUE HIPBLAS_STATUS_INVALID_VALUE
+#define CUBLAS_STATUS_ARCH_MISMATCH HIPBLAS_STATUS_ARCH_MISMATCH
+#define CUBLAS_STATUS_MAPPING_ERROR HIPBLAS_STATUS_MAPPING_ERROR
+#define CUBLAS_STATUS_EXECUTION_FAILED HIPBLAS_STATUS_EXECUTION_FAILED
+#define CUBLAS_STATUS_INTERNAL_ERROR HIPBLAS_STATUS_INTERNAL_ERROR
+#define CUBLAS_STATUS_NOT_SUPPORTED HIPBLAS_STATUS_NOT_SUPPORTED
+
+#define __CUDA_ARCH__ 1300
+
+#if defined(__gfx1100__) || defined(__gfx1101__) || defined(__gfx1102__) || defined(__gfx1103__) || \
+    defined(__gfx1150__) || defined(__gfx1151__)
+#define RDNA3
+#endif
+
+#if defined(__gfx1030__) || defined(__gfx1031__) || defined(__gfx1032__) || defined(__gfx1033__) || \
+    defined(__gfx1034__) || defined(__gfx1035__) || defined(__gfx1036__) || defined(__gfx1037__)
+#define RDNA2
+#endif
+
+#if defined(__gfx1010__) || defined(__gfx1012__)
+#define RDNA1
+#endif
+
+#ifndef __has_builtin
+    #define __has_builtin(x) 0
+#endif
+
+typedef int8_t int8x4_t __attribute__((ext_vector_type(4)));
+typedef uint8_t uint8x4_t __attribute__((ext_vector_type(4)));
+static __device__ __forceinline__ int __vsubss4(const int a, const int b) {
+    const int8x4_t va = reinterpret_cast<const int8x4_t&>(a);
+    const int8x4_t vb = reinterpret_cast<const int8x4_t&>(b);
+#if __has_builtin(__builtin_elementwise_sub_sat)
+    const int8x4_t c = __builtin_elementwise_sub_sat(va, vb);
+    return reinterpret_cast<const int &>(c);
+#else
+    int8x4_t c;
+    int16_t tmp;
+#pragma unroll
+    for (int i = 0; i < 4; i++) {
+        tmp = va[i] - vb[i];
+        if(tmp > std::numeric_limits<int8_t>::max()) tmp = std::numeric_limits<int8_t>::max();
+        if(tmp < std::numeric_limits<int8_t>::min()) tmp = std::numeric_limits<int8_t>::min();
+        c[i] = tmp;
+    }
+    return reinterpret_cast<int &>(c);
+#endif // __has_builtin(__builtin_elementwise_sub_sat)
+}
+
+static __device__ __forceinline__ int __vsub4(const int a, const int b) {
+    return __vsubss4(a, b);
+}
+
+static __device__ __forceinline__ unsigned int __vcmpeq4(unsigned int a, unsigned int b) {
+    const uint8x4_t& va = reinterpret_cast<const uint8x4_t&>(a);
+    const uint8x4_t& vb = reinterpret_cast<const uint8x4_t&>(b);
+    unsigned int c;
+    uint8x4_t& vc = reinterpret_cast<uint8x4_t&>(c);
+#pragma unroll
+    for (int i = 0; i < 4; ++i) {
+        vc[i] = va[i] == vb[i] ? 0xff : 0x00;
+    }
+    return c;
+}
+
+static __device__ __forceinline__ unsigned int __vcmpne4(unsigned int a, unsigned int b) {
+    const uint8x4_t& va = reinterpret_cast<const uint8x4_t&>(a);
+    const uint8x4_t& vb = reinterpret_cast<const uint8x4_t&>(b);
+    unsigned int c;
+    uint8x4_t& vc = reinterpret_cast<uint8x4_t&>(c);
+#pragma unroll
+    for (int i = 0; i < 4; ++i) {
+        vc[i] = va[i] == vb[i] ? 0x00 : 0xff;
+    }
+    return c;
+}
+
+#if defined(__HIP_PLATFORM_AMD__) && HIP_VERSION < 50600000
+// __shfl_xor() for half2 was added in ROCm 5.6
+static __device__ __forceinline__ half2 __shfl_xor(half2 var, int laneMask, int width) {
+    typedef union half2_b32 {
+        half2 val;
+        int   b32;
+    } half2_b32_t;
+    half2_b32_t tmp;
+    tmp.val = var;
+    tmp.b32 = __shfl_xor(tmp.b32, laneMask, width);
+    return tmp.val;
+}
+#endif // defined(__HIP_PLATFORM_AMD__) && HIP_VERSION < 50600000
diff --git a/src/whisper_cpp/ggml/src/ggml-cuda/vendors/musa.h b/src/whisper_cpp/ggml/src/ggml-cuda/vendors/musa.h
new file mode 100644
index 00000000..1604b822
--- /dev/null
+++ b/src/whisper_cpp/ggml/src/ggml-cuda/vendors/musa.h
@@ -0,0 +1,134 @@
+#pragma once
+
+#include <musa_runtime.h>
+#include <musa.h>
+#include <mublas.h>
+#include <musa_fp16.h>
+#define CUBLAS_COMPUTE_16F CUDA_R_16F
+#define CUBLAS_COMPUTE_32F CUDA_R_32F
+#define CUBLAS_COMPUTE_32F_FAST_16F MUBLAS_COMPUTE_32F_FAST_16F
+#define CUBLAS_GEMM_DEFAULT MUBLAS_GEMM_DEFAULT
+#define CUBLAS_GEMM_DEFAULT_TENSOR_OP MUBLAS_GEMM_DEFAULT
+#define CUBLAS_OP_N MUBLAS_OP_N
+#define CUBLAS_OP_T MUBLAS_OP_T
+#define CUBLAS_STATUS_SUCCESS MUBLAS_STATUS_SUCCESS
+#define CUBLAS_TF32_TENSOR_OP_MATH MUBLAS_MATH_MODE_DEFAULT
+#define CUDA_R_16F  MUSA_R_16F
+#define CUDA_R_32F  MUSA_R_32F
+#define cublasComputeType_t cudaDataType_t
+#define cublasCreate mublasCreate
+#define cublasDestroy mublasDestroy
+#define cublasGemmEx mublasGemmEx
+#define cublasGemmBatchedEx mublasGemmBatchedEx
+#define cublasGemmStridedBatchedEx mublasGemmStridedBatchedEx
+#define cublasHandle_t mublasHandle_t
+#define cublasSetMathMode mublasSetMathMode
+#define cublasSetStream mublasSetStream
+#define cublasSgemm mublasSgemm
+#define cublasStatus_t mublasStatus_t
+#define cublasOperation_t mublasOperation_t
+#define cublasGetStatusString mublasStatus_to_string
+#define cudaDataType_t musaDataType_t
+#define cudaDeviceCanAccessPeer musaDeviceCanAccessPeer
+#define cudaDeviceDisablePeerAccess musaDeviceDisablePeerAccess
+#define cudaDeviceEnablePeerAccess musaDeviceEnablePeerAccess
+#define cudaDeviceProp musaDeviceProp
+#define cudaDeviceSynchronize musaDeviceSynchronize
+#define cudaError_t musaError_t
+#define cudaErrorPeerAccessAlreadyEnabled musaErrorPeerAccessAlreadyEnabled
+#define cudaErrorPeerAccessNotEnabled musaErrorPeerAccessNotEnabled
+#define cudaEventCreateWithFlags musaEventCreateWithFlags
+#define cudaEventDisableTiming musaEventDisableTiming
+#define cudaEventRecord musaEventRecord
+#define cudaEventSynchronize musaEventSynchronize
+#define cudaEvent_t musaEvent_t
+#define cudaEventDestroy musaEventDestroy
+#define cudaFree musaFree
+#define cudaFreeHost musaFreeHost
+#define cudaGetDevice musaGetDevice
+#define cudaGetDeviceCount musaGetDeviceCount
+#define cudaGetDeviceProperties musaGetDeviceProperties
+#define cudaGetErrorString musaGetErrorString
+#define cudaGetLastError musaGetLastError
+#define cudaHostRegister musaHostRegister
+#define cudaHostRegisterPortable musaHostRegisterPortable
+#define cudaHostRegisterReadOnly musaHostRegisterReadOnly
+#define cudaHostUnregister musaHostUnregister
+#define cudaLaunchHostFunc musaLaunchHostFunc
+#define cudaMalloc musaMalloc
+#define cudaMallocHost musaMallocHost
+#define cudaMallocManaged musaMallocManaged
+#define cudaMemcpy musaMemcpy
+#define cudaMemcpyAsync musaMemcpyAsync
+#define cudaMemcpyPeerAsync musaMemcpyPeerAsync
+#define cudaMemcpy2DAsync musaMemcpy2DAsync
+#define cudaMemcpyDeviceToDevice musaMemcpyDeviceToDevice
+#define cudaMemcpyDeviceToHost musaMemcpyDeviceToHost
+#define cudaMemcpyHostToDevice musaMemcpyHostToDevice
+#define cudaMemcpyKind musaMemcpyKind
+#define cudaMemset musaMemset
+#define cudaMemsetAsync musaMemsetAsync
+#define cudaMemGetInfo musaMemGetInfo
+#define cudaOccupancyMaxPotentialBlockSize musaOccupancyMaxPotentialBlockSize
+#define cudaSetDevice musaSetDevice
+#define cudaStreamCreateWithFlags musaStreamCreateWithFlags
+#define cudaStreamDestroy musaStreamDestroy
+#define cudaStreamFireAndForget musaStreamFireAndForget
+#define cudaStreamNonBlocking musaStreamNonBlocking
+#define cudaStreamPerThread musaStreamPerThread
+#define cudaStreamSynchronize musaStreamSynchronize
+#define cudaStreamWaitEvent musaStreamWaitEvent
+#define cudaStream_t musaStream_t
+#define cudaSuccess musaSuccess
+
+// Additional mappings for MUSA virtual memory pool
+#define CU_DEVICE_ATTRIBUTE_VIRTUAL_MEMORY_MANAGEMENT_SUPPORTED MU_DEVICE_ATTRIBUTE_VIRTUAL_ADDRESS_MANAGEMENT_SUPPORTED
+#define CU_MEM_ACCESS_FLAGS_PROT_READWRITE MU_MEM_ACCESS_FLAGS_PROT_READWRITE
+#define CU_MEM_ALLOC_GRANULARITY_RECOMMENDED MU_MEM_ALLOC_GRANULARITY_RECOMMENDED
+#define CU_MEM_ALLOCATION_TYPE_PINNED MU_MEM_ALLOCATION_TYPE_PINNED
+#define CU_MEM_LOCATION_TYPE_DEVICE MU_MEM_LOCATION_TYPE_DEVICE
+#define CUdevice MUdevice
+#define CUdeviceptr MUdeviceptr
+#define CUmemAccessDesc MUmemAccessDesc
+#define CUmemAllocationProp MUmemAllocationProp
+#define CUmemGenericAllocationHandle MUmemGenericAllocationHandle
+#define cuDeviceGet muDeviceGet
+#define cuDeviceGetAttribute muDeviceGetAttribute
+#define cuMemAddressFree muMemAddressFree
+#define cuMemAddressReserve muMemAddressReserve
+#define cuMemCreate muMemCreate
+#define cuMemGetAllocationGranularity muMemGetAllocationGranularity
+#define cuMemMap muMemMap
+#define cuMemRelease muMemRelease
+#define cuMemSetAccess muMemSetAccess
+#define cuMemUnmap muMemUnmap
+#define cudaFuncAttributeMaxDynamicSharedMemorySize musaFuncAttributeMaxDynamicSharedMemorySize
+#define cudaFuncSetAttribute musaFuncSetAttribute
+#define cudaMemcpy3DPeerParms musaMemcpy3DPeerParms
+#define make_cudaExtent make_musaExtent
+#define make_cudaPitchedPtr make_musaPitchedPtr
+
+// Additional mappings for MUSA graphs
+#define CUDA_SUCCESS MUSA_SUCCESS
+#define CUresult MUresult
+#define cuGetErrorString muGetErrorString
+#define cudaErrorGraphExecUpdateFailure musaErrorGraphExecUpdateFailure
+#define cudaErrorInvalidDeviceFunction musaErrorInvalidDeviceFunction
+#define cudaGraphDestroy musaGraphDestroy
+#define cudaGraphExecDestroy musaGraphExecDestroy
+#define cudaGraphExec_t musaGraphExec_t
+#define cudaGraphExecUpdate musaGraphExecUpdate
+#define cudaGraphExecUpdateResultInfo musaGraphExecUpdateResult
+#define cudaGraphGetNodes musaGraphGetNodes
+#define cudaGraphInstantiate musaGraphInstantiate
+#define cudaGraphKernelNodeGetParams musaGraphKernelNodeGetParams
+#define cudaGraphKernelNodeSetParams musaGraphKernelNodeSetParams
+#define cudaGraphLaunch musaGraphLaunch
+#define cudaGraphNodeGetType musaGraphNodeGetType
+#define cudaGraphNode_t musaGraphNode_t
+#define cudaGraphNodeType musaGraphNodeType
+#define cudaGraphNodeTypeKernel musaGraphNodeTypeKernel
+#define cudaGraph_t musaGraph_t
+#define cudaKernelNodeParams musaKernelNodeParams
+#define cudaStreamCaptureModeRelaxed musaStreamCaptureModeRelaxed
+#define cudaStreamEndCapture musaStreamEndCapture
diff --git a/src/whisper_cpp/ggml/src/ggml-impl.h b/src/whisper_cpp/ggml/src/ggml-impl.h
new file mode 100644
index 00000000..d3f4bad8
--- /dev/null
+++ b/src/whisper_cpp/ggml/src/ggml-impl.h
@@ -0,0 +1,201 @@
+#pragma once
+
+// GGML internal header
+
+#include "ggml.h"
+
+#include <assert.h>
+#include <stdlib.h> // load `stdlib.h` before other headers to work around MinGW bug: https://sourceforge.net/p/mingw-w64/bugs/192/
+#include <stdbool.h>
+#include <stdint.h>
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+#undef MIN
+#undef MAX
+
+#define MIN(a, b) ((a) < (b) ? (a) : (b))
+#define MAX(a, b) ((a) > (b) ? (a) : (b))
+
+// static_assert should be a #define, but if it's not,
+// fall back to the _Static_assert C11 keyword.
+// if C99 - static_assert is noop
+// ref: https://stackoverflow.com/a/53923785/4039976
+#ifndef __cplusplus
+#ifndef static_assert
+#if defined(__STDC_VERSION__) && (__STDC_VERSION__ >= 201100L)
+#define static_assert(cond, msg) _Static_assert(cond, msg)
+#else
+#define static_assert(cond, msg) struct global_scope_noop_trick
+#endif
+#endif
+#endif
+
+//
+// logging
+//
+
+GGML_ATTRIBUTE_FORMAT(2, 3)
+void ggml_log_internal        (enum ggml_log_level level, const char * format, ...);
+void ggml_log_callback_default(enum ggml_log_level level, const char * text, void * user_data);
+
+#define GGML_LOG(...)       ggml_log_internal(GGML_LOG_LEVEL_NONE , __VA_ARGS__)
+#define GGML_LOG_INFO(...)  ggml_log_internal(GGML_LOG_LEVEL_INFO , __VA_ARGS__)
+#define GGML_LOG_WARN(...)  ggml_log_internal(GGML_LOG_LEVEL_WARN , __VA_ARGS__)
+#define GGML_LOG_ERROR(...) ggml_log_internal(GGML_LOG_LEVEL_ERROR, __VA_ARGS__)
+#define GGML_LOG_DEBUG(...) ggml_log_internal(GGML_LOG_LEVEL_DEBUG, __VA_ARGS__)
+#define GGML_LOG_CONT(...)  ggml_log_internal(GGML_LOG_LEVEL_CONT , __VA_ARGS__)
+
+// bitset
+
+typedef uint32_t ggml_bitset_t;
+
+static_assert(sizeof(ggml_bitset_t) == 4, "bitset_t constants must be updated");
+#define BITSET_SHR 5 // log2(sizeof(ggml_bitset_t)*8)
+#define BITSET_MASK (sizeof(ggml_bitset_t)*8 - 1)
+
+static size_t ggml_bitset_size(size_t n) {
+    return (n + BITSET_MASK) >> BITSET_SHR;
+}
+
+static inline bool ggml_bitset_get(const ggml_bitset_t * bitset, size_t i) {
+    return !!(bitset[i >> BITSET_SHR] & (1u << (i & BITSET_MASK)));
+}
+
+static inline void ggml_bitset_set(ggml_bitset_t * bitset, size_t i) {
+    bitset[i >> BITSET_SHR] |= (1u << (i & BITSET_MASK));
+}
+
+static inline void ggml_bitset_clear(ggml_bitset_t * bitset, size_t i) {
+    bitset[i >> BITSET_SHR] &= ~(1u << (i & BITSET_MASK));
+}
+
+// hash set
+
+#define GGML_HASHSET_FULL ((size_t)-1)
+#define GGML_HASHSET_ALREADY_EXISTS ((size_t)-2)
+
+struct ggml_hash_set {
+    size_t size;
+    ggml_bitset_t * used;       // whether or not the keys are in use i.e. set
+    struct ggml_tensor ** keys; // actual tensors in the set, keys[i] is only defined if ggml_bitset_get(used, i)
+};
+
+struct ggml_hash_set ggml_hash_set_new(size_t size);
+void                 ggml_hash_set_free(struct ggml_hash_set * hash_set);
+
+// returns the minimum size for a hash set that can hold min_sz elements
+size_t ggml_hash_size(size_t min_sz);
+
+// remove all elements from the hash set
+void ggml_hash_set_reset(struct ggml_hash_set * hash_set);
+
+// returns true if key is in the hash set
+static bool ggml_hash_contains(const struct ggml_hash_set * hash_set, struct ggml_tensor * key);
+
+// returns GGML_HASHSET_FULL if table is full, otherwise the current index of the key or where it should be inserted
+static size_t ggml_hash_find(const struct ggml_hash_set * hash_set, struct ggml_tensor * key);
+
+// returns GGML_HASHSET_ALREADY_EXISTS if key already exists, index otherwise, asserts if table is full
+static size_t ggml_hash_insert(struct ggml_hash_set * hash_set, struct ggml_tensor * key);
+
+// return index, asserts if table is full
+static size_t ggml_hash_find_or_insert(struct ggml_hash_set * hash_set, struct ggml_tensor * key);
+
+// hash function for ggml_tensor
+static inline size_t ggml_hash(const struct ggml_tensor * p) {
+    // the last 4 bits are always zero due to alignment
+    return (size_t)(uintptr_t)p >> 4;
+}
+
+static size_t ggml_hash_find(const struct ggml_hash_set * hash_set, struct ggml_tensor * key) {
+    size_t h = ggml_hash(key) % hash_set->size;
+
+    // linear probing
+    size_t i = h;
+    while (ggml_bitset_get(hash_set->used, i) && hash_set->keys[i] != key) {
+        i = (i + 1) % hash_set->size;
+        if (i == h) {
+            // visited all hash table entries -> not found
+            return GGML_HASHSET_FULL;
+        }
+    }
+    return i;
+}
+
+static bool ggml_hash_contains(const struct ggml_hash_set * hash_set, struct ggml_tensor * key) {
+    size_t i = ggml_hash_find(hash_set, key);
+    return i != GGML_HASHSET_FULL && ggml_bitset_get(hash_set->used, i);
+}
+
+static size_t ggml_hash_insert(struct ggml_hash_set * hash_set, struct ggml_tensor * key) {
+    size_t h = ggml_hash(key) % hash_set->size;
+
+    // linear probing
+    size_t i = h;
+    do {
+        if (!ggml_bitset_get(hash_set->used, i)) {
+            ggml_bitset_set(hash_set->used, i);
+            hash_set->keys[i] = key;
+            return i;
+        }
+        if (hash_set->keys[i] == key) {
+            return GGML_HASHSET_ALREADY_EXISTS;
+        }
+        i = (i + 1) % hash_set->size;
+    } while (i != h);
+
+    // visited all hash table entries -> not found
+    GGML_ABORT("fatal error");
+}
+
+static size_t ggml_hash_find_or_insert(struct ggml_hash_set * hash_set, struct ggml_tensor * key) {
+    size_t h = ggml_hash(key) % hash_set->size;
+
+    // linear probing
+    size_t i = h;
+    do {
+        if (!ggml_bitset_get(hash_set->used, i)) {
+            ggml_bitset_set(hash_set->used, i);
+            hash_set->keys[i] = key;
+            return i;
+        }
+        if (hash_set->keys[i] == key) {
+            return i;
+        }
+        i = (i + 1) % hash_set->size;
+    } while (i != h);
+
+    // visited all hash table entries -> not found
+    GGML_ABORT("fatal error");
+}
+
+// computation graph
+
+enum ggml_cgraph_eval_order {
+    GGML_CGRAPH_EVAL_ORDER_LEFT_TO_RIGHT = 0,
+    GGML_CGRAPH_EVAL_ORDER_RIGHT_TO_LEFT,
+    GGML_CGRAPH_EVAL_ORDER_COUNT
+};
+
+struct ggml_cgraph {
+    int size;
+    int n_nodes;
+    int n_leafs;
+
+    struct ggml_tensor ** nodes;
+    struct ggml_tensor ** grads;
+    struct ggml_tensor ** leafs;
+
+    struct ggml_hash_set visited_hash_set;
+
+    enum ggml_cgraph_eval_order order;
+};
+
+struct ggml_cgraph ggml_graph_view(struct ggml_cgraph * cgraph, int i0, int i1);
+
+#ifdef __cplusplus
+}
+#endif
diff --git a/src/whisper_cpp/ggml/src/ggml-kompute.cpp b/src/whisper_cpp/ggml/src/ggml-kompute.cpp
new file mode 100644
index 00000000..2c926aae
--- /dev/null
+++ b/src/whisper_cpp/ggml/src/ggml-kompute.cpp
@@ -0,0 +1,2018 @@
+#include "ggml-impl.h"
+#include "ggml-backend.h"
+#include "ggml-backend-impl.h"
+#include "ggml-kompute.h"
+
+// These are generated at build time by cmake custom command
+#include "shaderop_scale.h"
+#include "shaderop_scale_8.h"
+#include "shaderop_add.h"
+#include "shaderop_addrow.h"
+#include "shaderop_mul.h"
+#include "shaderop_silu.h"
+#include "shaderop_relu.h"
+#include "shaderop_gelu.h"
+#include "shaderop_softmax.h"
+#include "shaderop_norm.h"
+#include "shaderop_rmsnorm.h"
+#include "shaderop_diagmask.h"
+#include "shaderop_mul_mat_f16.h"
+#include "shaderop_mul_mat_q8_0.h"
+#include "shaderop_mul_mat_q4_0.h"
+#include "shaderop_mul_mat_q4_1.h"
+#include "shaderop_mul_mat_q6_k.h"
+#include "shaderop_mul_mat_mat_f32.h"
+#include "shaderop_getrows_f32.h"
+#include "shaderop_getrows_f16.h"
+#include "shaderop_getrows_q4_0.h"
+#include "shaderop_getrows_q4_1.h"
+#include "shaderop_getrows_q6_k.h"
+#include "shaderop_rope_f16.h"
+#include "shaderop_rope_f32.h"
+#include "shaderop_cpy_f16_f16.h"
+#include "shaderop_cpy_f16_f32.h"
+#include "shaderop_cpy_f32_f16.h"
+#include "shaderop_cpy_f32_f32.h"
+
+#include <algorithm>
+#include <array>
+#include <cassert>
+#include <cstdint>
+#include <cstdio>
+#include <cstring>
+#include <iostream>
+#include <memory>
+#include <stdexcept>
+#include <string>
+#include <unordered_map>
+#include <utility>
+#include <vector>
+
+#include <kompute/Kompute.hpp>
+#include <vulkan/vulkan.hpp>
+
+#ifdef __linux__
+#include <cstdlib> // for setenv
+#endif
+
+#define QK4_0 32
+#define QR4_0 2
+#define QK4_1 32
+#define QK_NL 16
+
+typedef ggml_fp16_t half;
+
+static std::string ggml_kompute_format_name(int device) {
+    return "Kompute" + std::to_string(device);
+}
+
+struct ggml_kompute_context {
+    int device;
+    std::string name;
+    std::shared_ptr<vk::DescriptorPool> pool;
+
+    ggml_kompute_context(int device)
+        : device(device), name(ggml_kompute_format_name(device)) {}
+};
+
+// FIXME: It would be good to consolidate the kompute manager and the kompute context into one object
+// and consolidate the init functions and simplify object lifetime management. As it currently stands,
+// we *have* to have the kompute manager no matter what for device discovery, but the kompute context
+// is only created when a device is set and vulkan is explicitly turned on.
+static ggml_kompute_context *s_kompute_context = nullptr;
+
+class kompute_manager {
+    kp::Manager *s_mgr = nullptr;
+
+public:
+    kp::Manager *operator()() {
+        if (s_mgr && !s_mgr->hasInstance()) {
+            destroy();
+        }
+        if (!s_mgr) {
+            s_mgr = new kp::Manager;
+        }
+        return s_mgr;
+    }
+
+    void destroy() {
+        delete s_mgr;
+        s_mgr = nullptr;
+    }
+};
+
+static kompute_manager komputeManager;
+
+struct ggml_vk_memory {
+    void *data = nullptr;
+    size_t size = 0;
+    vk::DeviceMemory *primaryMemory = nullptr;
+    vk::Buffer *primaryBuffer = nullptr;
+    vk::DeviceMemory *stagingMemory = nullptr;
+    vk::Buffer *stagingBuffer = nullptr;
+};
+
+#ifdef __linux__
+__attribute__((constructor))
+static void enable_sam() {
+    setenv("RADV_PERFTEST", "sam", false);
+}
+#endif
+
+static bool ggml_vk_checkPhysicalDeviceFeatures(vk::PhysicalDevice physical_device) {
+    vk::PhysicalDeviceFeatures availableFeatures;
+    physical_device.getFeatures(&availableFeatures);
+
+    if (!availableFeatures.shaderInt16)
+        return false;
+
+    vk::PhysicalDeviceVulkan11Features availableFeatures11;
+    vk::PhysicalDeviceVulkan12Features availableFeatures12;
+
+    availableFeatures11.pNext = &availableFeatures12;
+    availableFeatures12.pNext = nullptr;
+
+    vk::PhysicalDeviceFeatures2 features2;
+    features2.pNext = &availableFeatures11;
+
+    physical_device.getFeatures2(&features2);
+
+    if (!availableFeatures11.uniformAndStorageBuffer16BitAccess ||
+        !availableFeatures11.storageBuffer16BitAccess) {
+        return false;
+    }
+
+    if (!availableFeatures12.storageBuffer8BitAccess ||
+        !availableFeatures12.uniformAndStorageBuffer8BitAccess ||
+        !availableFeatures12.shaderFloat16 ||
+        !availableFeatures12.shaderInt8) {
+        return false;
+    }
+
+    return true;
+}
+
+static const char * ggml_vk_getVendorName(uint32_t vendorID) {
+    switch (vendorID) {
+        case 0x10DE:
+            return "nvidia";
+        case 0x1002:
+            return "amd";
+        case 0x8086:
+            return "intel";
+        default:
+            return "unknown";
+    }
+}
+
+static std::vector<ggml_vk_device> ggml_vk_available_devices_internal(size_t memoryRequired) {
+    std::vector<ggml_vk_device> results;
+    if (!komputeManager()->hasVulkan() || !komputeManager()->hasInstance())
+        return results;
+
+    std::vector<vk::PhysicalDevice> physical_devices;
+    try {
+        physical_devices = komputeManager()->listDevices();
+    } catch (vk::SystemError & err) {
+        std::cerr << __func__ << ": ignoring Vulkan exception: " << err.what() << "\n";
+        return results;
+    }
+
+    uint32_t deviceCount = physical_devices.size();
+    if (deviceCount == 0)
+        return results;
+
+    std::unordered_map<std::string, size_t> count_by_name;
+
+    for (uint32_t i = 0; i < deviceCount; i++) {
+        const auto & physical_device = physical_devices[i];
+
+        VkPhysicalDeviceProperties dev_props = physical_device.getProperties();
+        VkPhysicalDeviceMemoryProperties memoryProperties = physical_device.getMemoryProperties();
+        const uint32_t major = VK_VERSION_MAJOR(dev_props.apiVersion);
+        const uint32_t minor = VK_VERSION_MINOR(dev_props.apiVersion);
+        if (major < 1 || minor < 2)
+            continue;
+
+        if (!ggml_vk_checkPhysicalDeviceFeatures(physical_device))
+            continue;
+
+        size_t heapSize = 0;
+        for (uint32_t j = 0; j < memoryProperties.memoryHeapCount; ++j) {
+            VkMemoryHeap heap = memoryProperties.memoryHeaps[j];
+            if (heap.flags & VK_MEMORY_HEAP_DEVICE_LOCAL_BIT) {
+                heapSize = heap.size;
+                break;
+            }
+        }
+
+        if (heapSize < memoryRequired)
+            continue;
+
+        auto ext_props = physical_device.enumerateDeviceExtensionProperties();
+        bool has_maintenance4 = false;
+
+        // Check if maintenance4 is supported
+        for (const auto & properties : ext_props) {
+            if (strcmp("VK_KHR_maintenance4", properties.extensionName) == 0) {
+                has_maintenance4 = true;
+            }
+        }
+
+        vk::PhysicalDeviceSubgroupProperties subgroup_props;
+        vk::PhysicalDeviceProperties2 dev_props2;
+        vk::PhysicalDeviceMaintenance3Properties dev_props3;
+        vk::PhysicalDeviceMaintenance4Properties dev_props4;
+        dev_props2.pNext = &dev_props3;
+        dev_props3.pNext = &subgroup_props;
+        if (has_maintenance4) {
+            subgroup_props.pNext = &dev_props4;
+        }
+        physical_device.getProperties2(&dev_props2);
+
+        if (subgroup_props.subgroupSize < 32)
+            continue;
+
+        ggml_vk_device d;
+        d.index = i;
+        d.type = dev_props.deviceType;
+        d.heapSize = heapSize;
+        d.vendor = strdup(ggml_vk_getVendorName(dev_props.vendorID));
+        d.subgroupSize = subgroup_props.subgroupSize;
+        d.bufferAlignment = dev_props.limits.minStorageBufferOffsetAlignment;
+
+        if (has_maintenance4) {
+            d.maxAlloc = std::min(dev_props3.maxMemoryAllocationSize, dev_props4.maxBufferSize);
+        } else {
+            d.maxAlloc = dev_props3.maxMemoryAllocationSize;
+        }
+
+        std::string name(dev_props.deviceName);
+        size_t n_idx = ++count_by_name[name];
+        if (n_idx > 1) {
+            name += " (" + std::to_string(n_idx) + ")";
+        }
+        d.name = strdup(name.c_str());
+
+        results.push_back(d);
+    }
+
+    std::stable_sort(results.begin(), results.end(),
+        [](const ggml_vk_device& lhs, const ggml_vk_device& rhs) -> bool {
+            if (lhs.type != rhs.type) {
+                if (lhs.type == VK_PHYSICAL_DEVICE_TYPE_DISCRETE_GPU) return true;
+                if (rhs.type == VK_PHYSICAL_DEVICE_TYPE_DISCRETE_GPU) return false;
+
+                if (lhs.type == VK_PHYSICAL_DEVICE_TYPE_INTEGRATED_GPU) return true;
+                if (rhs.type == VK_PHYSICAL_DEVICE_TYPE_INTEGRATED_GPU) return false;
+            }
+            return lhs.heapSize < rhs.heapSize;
+        }
+    );
+
+    return results;
+}
+
+// public API returns a C-style array
+ggml_vk_device * ggml_vk_available_devices(size_t memoryRequired, size_t * count) {
+    auto devices = ggml_vk_available_devices_internal(memoryRequired);
+    *count = devices.size();
+    if (devices.empty()) {
+        return nullptr;
+    }
+
+    size_t nbytes = sizeof (ggml_vk_device) * (devices.size());
+    auto * arr = static_cast<ggml_vk_device *>(malloc(nbytes));
+    memcpy(arr, devices.data(), nbytes);
+    return arr;
+}
+
+static void ggml_vk_filterByVendor(std::vector<ggml_vk_device>& devices, const std::string& targetVendor) {
+    devices.erase(
+        std::remove_if(devices.begin(), devices.end(),
+            [&targetVendor](const ggml_vk_device& device) {
+                return device.vendor != targetVendor;
+            }),
+        devices.end()
+    );
+}
+
+static void ggml_vk_filterByName(std::vector<ggml_vk_device>& devices, const std::string& targetName) {
+    devices.erase(
+        std::remove_if(devices.begin(), devices.end(),
+            [&targetName](const ggml_vk_device& device) {
+                return device.name != targetName;
+            }),
+        devices.end()
+    );
+}
+
+static bool ggml_vk_get_device(ggml_vk_device * device, size_t memoryRequired, const std::string & name) {
+    if (name.empty())
+        return false;
+
+    auto devices = ggml_vk_available_devices_internal(memoryRequired);
+    if (name == "amd" || name == "nvidia" || name == "intel") {
+        ggml_vk_filterByVendor(devices, name);
+    } else if (name != "gpu") {
+        ggml_vk_filterByName(devices, name);
+    }
+
+    if (devices.empty())
+        return false;
+
+    *device = devices.front();
+    return true;
+}
+
+bool ggml_vk_get_device(ggml_vk_device * device, size_t memoryRequired, const char * name) {
+    return ggml_vk_get_device(device, memoryRequired, std::string(name));
+}
+
+bool ggml_vk_has_vulkan() {
+    return komputeManager()->hasVulkan();
+}
+
+bool ggml_vk_has_device() {
+    return komputeManager()->hasDevice();
+}
+
+ggml_vk_device ggml_vk_current_device() {
+    if (!komputeManager()->hasDevice())
+        return ggml_vk_device();
+
+    auto devices = ggml_vk_available_devices_internal(0);
+    ggml_vk_filterByName(devices, komputeManager()->physicalDevice()->getProperties().deviceName.data());
+    GGML_ASSERT(!devices.empty());
+    return devices.front();
+}
+
+static
+void ggml_vk_allocate_descriptor_pool(struct ggml_kompute_context * ctx, size_t size) {
+    std::vector<vk::DescriptorPoolSize> descriptorPoolSizes = {
+        vk::DescriptorPoolSize(
+          vk::DescriptorType::eStorageBuffer,
+          3 * size // Descriptor count is number of possible tensors to pass into an algorithm
+          )
+    };
+
+    vk::DescriptorPoolCreateInfo descriptorPoolInfo(
+      vk::DescriptorPoolCreateFlags(),
+      size, // Max sets
+      static_cast<uint32_t>(descriptorPoolSizes.size()),
+      descriptorPoolSizes.data());
+
+    ctx->pool = std::make_shared<vk::DescriptorPool>();
+    vk::Result r = komputeManager()->device()->createDescriptorPool(
+      &descriptorPoolInfo, nullptr, ctx->pool.get());
+    if (r != vk::Result::eSuccess)
+        std::cerr << "Error allocating descriptor pool" << vk::to_string(r);
+}
+
+static
+void ggml_vk_free_descriptor_pool(struct ggml_kompute_context * ctx) {
+    if (ctx->pool) {
+        komputeManager()->device()->destroy(
+          *ctx->pool,
+          (vk::Optional<const vk::AllocationCallbacks>)nullptr);
+        ctx->pool = nullptr;
+    }
+}
+
+static
+vk::Buffer *ggml_vk_allocate_buffer(size_t size) {
+    vk::BufferCreateInfo bufferCreateInfo;
+    bufferCreateInfo.size = size;
+    bufferCreateInfo.usage = vk::BufferUsageFlagBits::eStorageBuffer |
+                             vk::BufferUsageFlagBits::eTransferSrc |
+                             vk::BufferUsageFlagBits::eTransferDst;
+    bufferCreateInfo.sharingMode = vk::SharingMode::eExclusive;
+
+    vk::Buffer *vkBuffer = new vk::Buffer;
+    vk::Result r = komputeManager()->device()->createBuffer(&bufferCreateInfo, nullptr, vkBuffer);
+    if (r != vk::Result::eSuccess)
+        std::cerr << "Error allocating buffer " << vk::to_string(r) << std::endl;
+    return vkBuffer;
+}
+
+static
+vk::DeviceMemory *ggml_vk_allocate(size_t size, vk::MemoryPropertyFlags flags, vk::MemoryRequirements requirements, bool *isHostVisible) {
+
+    uint32_t memoryTypeIndex = -1;
+    bool memoryTypeIndexFound = false;
+    vk::PhysicalDeviceMemoryProperties memoryProperties = komputeManager()->physicalDevice()->getMemoryProperties();
+    for (uint32_t i = 0; i < memoryProperties.memoryTypeCount; i++) {
+        const vk::MemoryType &memoryType = memoryProperties.memoryTypes[i];
+        const vk::MemoryHeap &memoryHeap = memoryProperties.memoryHeaps[memoryType.heapIndex];
+        if (memoryHeap.size < size) {
+            continue;
+        }
+
+        if (requirements.memoryTypeBits & (1 << i)) {
+            if (((memoryProperties.memoryTypes[i]).propertyFlags &
+                 flags) == flags) {
+                memoryTypeIndex = i;
+                memoryTypeIndexFound = true;
+                if (isHostVisible && (memoryProperties.memoryTypes[i].propertyFlags & vk::MemoryPropertyFlagBits::eHostVisible)) {
+                    *isHostVisible = true;
+                }
+                break;
+            }
+        }
+    }
+    if (!memoryTypeIndexFound) {
+        throw std::runtime_error(
+          "Memory type index for buffer creation not found");
+    }
+
+    vk::MemoryAllocateInfo allocInfo;
+    allocInfo.allocationSize = size;
+    allocInfo.memoryTypeIndex = memoryTypeIndex;
+    vk::DeviceMemory *vkDeviceMemory =  new vk::DeviceMemory;
+    vk::Result r = komputeManager()->device()->allocateMemory(&allocInfo, nullptr, vkDeviceMemory);
+    if (r != vk::Result::eSuccess) {
+        std::cerr << "Error allocating memory " << vk::to_string(r) << std::endl;
+        throw std::runtime_error("Error allocating vulkan memory.");
+    }
+    return vkDeviceMemory;
+}
+
+static size_t ggml_vk_aligned_offset(ggml_backend_buffer_t buffer, size_t offset) {
+    size_t minStorageBufferOffsetAlignment = ggml_backend_buffer_get_alignment(buffer);
+
+    // If offset is already aligned, return it directly
+    if (offset % minStorageBufferOffsetAlignment == 0) {
+        return offset;
+    }
+
+    // Otherwise, return the largest multiple of minStorageBufferOffsetAlignment less than offset
+    return (offset / minStorageBufferOffsetAlignment) * minStorageBufferOffsetAlignment;
+}
+
+static ggml_vk_memory ggml_vk_allocate(size_t size) {
+    ggml_vk_memory memory;
+    bool isHostVisible = false;
+    {
+        memory.primaryBuffer = ggml_vk_allocate_buffer(size);
+        vk::MemoryRequirements memoryRequirements = komputeManager()->device()->getBufferMemoryRequirements(*memory.primaryBuffer);
+        vk::MemoryPropertyFlags memoryPropertyFlags = vk::MemoryPropertyFlagBits::eDeviceLocal;
+        memory.primaryMemory = ggml_vk_allocate(size, memoryPropertyFlags, memoryRequirements, &isHostVisible);
+        komputeManager()->device()->bindBufferMemory(*memory.primaryBuffer, *memory.primaryMemory, 0);
+        if (isHostVisible) {
+            vk::Result r = komputeManager()->device()->mapMemory(*memory.primaryMemory, 0, size, vk::MemoryMapFlags(), &memory.data);
+            if (r != vk::Result::eSuccess)
+                std::cerr << "Error mapping memory" << vk::to_string(r);
+        }
+    }
+
+    if (!isHostVisible) {
+        memory.stagingBuffer = ggml_vk_allocate_buffer(size);
+        vk::MemoryRequirements memoryRequirements = komputeManager()->device()->getBufferMemoryRequirements(*memory.stagingBuffer);
+        vk::MemoryPropertyFlags memoryPropertyFlags = vk::MemoryPropertyFlagBits::eHostVisible |
+                                                      vk::MemoryPropertyFlagBits::eHostCoherent |
+                                                      vk::MemoryPropertyFlagBits::eHostCached;
+        memory.stagingMemory = ggml_vk_allocate(size, memoryPropertyFlags, memoryRequirements, &isHostVisible);
+        komputeManager()->device()->bindBufferMemory(*memory.stagingBuffer, *memory.stagingMemory, 0);
+        vk::Result r = komputeManager()->device()->mapMemory(*memory.stagingMemory, 0, size, vk::MemoryMapFlags(), &memory.data);
+        if (r != vk::Result::eSuccess)
+            std::cerr << "Error mapping memory" << vk::to_string(r);
+    }
+
+    memory.size = size;
+    return memory;
+}
+
+static void ggml_vk_free_memory(ggml_vk_memory &memory)
+{
+    komputeManager()->device()->destroy(
+      *memory.primaryBuffer,
+      (vk::Optional<const vk::AllocationCallbacks>)nullptr);
+    if (memory.stagingBuffer) {
+        komputeManager()->device()->destroy(
+          *memory.stagingBuffer,
+          (vk::Optional<const vk::AllocationCallbacks>)nullptr);
+    }
+    komputeManager()->device()->freeMemory(
+      *memory.primaryMemory,
+      (vk::Optional<const vk::AllocationCallbacks>)nullptr);
+    if (memory.stagingMemory) {
+        komputeManager()->device()->freeMemory(
+          *memory.stagingMemory,
+          (vk::Optional<const vk::AllocationCallbacks>)nullptr);
+    }
+}
+
+static const char * ggml_backend_kompute_buffer_type_get_name(ggml_backend_buffer_type_t buft);
+
+static
+ggml_vk_memory * ggml_vk_find_tensor(const struct ggml_tensor * t, uint64_t & offset) {
+    ggml_backend_buffer_t buffer = t->view_src ? t->view_src->buffer : t->buffer;
+
+    // compatibility with ggml-backend
+    GGML_ASSERT(buffer && buffer->buft->iface.get_name == ggml_backend_kompute_buffer_type_get_name);
+
+    ggml_vk_memory * buf_ctx = static_cast<ggml_vk_memory *>(buffer->context);
+
+    const intptr_t ioffs = intptr_t(t->data) - intptr_t(buf_ctx->data);
+
+    GGML_ASSERT(ioffs >= 0 && ioffs + int64_t(ggml_nbytes(t)) <= int64_t(buffer->size));
+
+    offset = uint64_t(ioffs);
+    return buf_ctx;
+}
+
+static
+const std::shared_ptr<kp::Tensor> ggml_vk_get_tensor(const struct ggml_tensor * t, uint32_t * alignedOffset = nullptr) {
+    uint64_t originalOffset = 0;
+    auto * res = ggml_vk_find_tensor(t, originalOffset);
+    if (!res) {
+        static std::shared_ptr<kp::Tensor> nullTensor = nullptr;
+        return nullTensor;
+    }
+
+    // Create a tensor whose memory will be composed of our buffers at the correct offset
+    const size_t nelements = ggml_nelements(t);
+    size_t nbytes = ggml_nbytes(t);
+
+    size_t vulkanOffset = ggml_vk_aligned_offset(t->buffer, originalOffset);
+    if (alignedOffset) {
+        *alignedOffset = originalOffset - vulkanOffset;
+        nbytes += *alignedOffset;
+    }
+
+    return komputeManager()->tensor(
+        t->data,
+        nelements,
+        nbytes, kp::Tensor::TensorDataTypes::eFloat,
+        res->primaryMemory, res->primaryBuffer,
+        res->stagingMemory, res->stagingBuffer,
+        vulkanOffset);
+}
+
+static std::vector<uint32_t> getSpirvShader(const unsigned char* rawData, size_t size) {
+    if (size % sizeof(uint32_t) != 0) {
+        throw std::runtime_error("Invalid size: must be divisible by sizeof(uint32_t)");
+    }
+
+    const uint32_t* data_ptr = reinterpret_cast<const uint32_t*>(rawData);
+    size_t count = size / sizeof(uint32_t);
+    return std::vector<uint32_t>(data_ptr, data_ptr + count);
+}
+
+inline static
+uint32_t safe_divide(uint32_t a, uint32_t b) {
+    if (b <= 1) {
+        return a;
+    }
+    if ((a % b) != 0) {
+        fprintf(stderr, "((%u %% %u) == %u) != 0\n", a, b, a % b);
+        GGML_ABORT("safe_divide result would've had remainder");
+    }
+    return a / b;
+}
+
+static void ggml_vk_add(
+    kp::Sequence& seq,
+    const std::shared_ptr<kp::Tensor>& inA,
+    const std::shared_ptr<kp::Tensor>& inB,
+    const std::shared_ptr<kp::Tensor>& out,
+    uint32_t inAOff, uint32_t inBOff, uint32_t outOff,
+    int32_t ne00, int32_t ne01, int32_t ne02, int32_t ne03,
+    int32_t nb00, int32_t nb01, int32_t nb02, int32_t nb03,
+    int32_t ne10, int32_t ne11, int32_t ne12, int32_t ne13,
+    int32_t nb10, int32_t nb11, int32_t nb12, int32_t nb13,
+    int32_t ne0,
+    int32_t nb0,  int32_t nb1,  int32_t nb2,  int32_t nb3
+) {
+    const static auto spirv = getSpirvShader(kp::shader_data::op_add_comp_spv,
+        kp::shader_data::op_add_comp_spv_len);
+
+    struct PushConstants {
+        uint32_t inAOff, inBOff, outOff;
+        int32_t ne00;
+        int32_t nb00, nb01, nb02, nb03;
+        int32_t ne10, ne11, ne12, ne13;
+        int32_t nb10, nb11, nb12, nb13;
+        int32_t ne0;
+        int32_t nb0, nb1, nb2, nb3;
+    } const pushConsts {
+        safe_divide(inAOff, 4), safe_divide(inBOff, 4), safe_divide(outOff, 4),
+        ne00,
+        nb00, nb01, nb02, nb03,
+        ne10, ne11, ne12, ne13,
+        nb10, nb11, nb12, nb13,
+        ne0,
+        nb0, nb1, nb2, nb3
+    };
+
+    std::shared_ptr<kp::Algorithm> s_algo = nullptr;
+    if (!komputeManager()->hasAlgorithm(__func__)) {
+        s_algo = komputeManager()->algorithm<float, PushConstants>(__func__, s_kompute_context->pool.get(), {inA, inB, out}, spirv, {unsigned(ne01), unsigned(ne02), unsigned(ne03)}, {}, {pushConsts});
+    } else {
+        s_algo = komputeManager()->getAlgorithm(__func__);
+        s_algo->setTensors({inA, inB, out});
+        s_algo->setWorkgroup({unsigned(ne01), unsigned(ne02), unsigned(ne03)});
+        s_algo->setPushConstants<PushConstants>({pushConsts});
+        s_algo->updateDescriptors(s_kompute_context->pool.get());
+    }
+    seq.record<kp::OpAlgoDispatch>(s_algo);
+}
+
+static void ggml_vk_addrow(kp::Sequence& seq,
+                 const std::shared_ptr<kp::Tensor>& inA,
+                 const std::shared_ptr<kp::Tensor>& inB,
+                 const std::shared_ptr<kp::Tensor>& out,
+                 uint32_t inAOff, uint32_t inBOff, uint32_t outOff,
+                 uint32_t size, uint32_t row = 0) {
+
+    const static auto spirv = getSpirvShader(kp::shader_data::op_addrow_comp_spv,
+        kp::shader_data::op_addrow_comp_spv_len);
+
+    struct PushConstants {
+        uint32_t inAOff, inBOff, outOff;
+        uint32_t row;
+    } const pushConsts {
+        safe_divide(inAOff, 4), safe_divide(inBOff, 4), safe_divide(outOff, 4),
+        row
+    };
+
+    std::shared_ptr<kp::Algorithm> s_algo = nullptr;
+    if (!komputeManager()->hasAlgorithm(__func__))
+        s_algo = komputeManager()->algorithm<float, PushConstants>(__func__, s_kompute_context->pool.get(), {inA, inB, out}, spirv, {size}, {}, {pushConsts});
+    else {
+        s_algo = komputeManager()->getAlgorithm(__func__);
+        s_algo->setTensors({inA, inB, out});
+        s_algo->setWorkgroup({size});
+        s_algo->setPushConstants<PushConstants>({pushConsts});
+        s_algo->updateDescriptors(s_kompute_context->pool.get());
+    }
+    seq.record<kp::OpAlgoDispatch>(s_algo);
+}
+
+static void ggml_vk_mul(
+    kp::Sequence& seq,
+    const std::shared_ptr<kp::Tensor>& inA,
+    const std::shared_ptr<kp::Tensor>& inB,
+    const std::shared_ptr<kp::Tensor>& out,
+    uint32_t inAOff, uint32_t inBOff, uint32_t outOff,
+    int32_t ne00, int32_t ne01, int32_t ne02, int32_t ne03,
+    int32_t nb00, int32_t nb01, int32_t nb02, int32_t nb03,
+    int32_t ne10, int32_t ne11, int32_t ne12, int32_t ne13,
+    int32_t nb10, int32_t nb11, int32_t nb12, int32_t nb13,
+    int32_t ne0,
+    int32_t nb0,  int32_t nb1,  int32_t nb2,  int32_t nb3
+) {
+    const static auto spirv = getSpirvShader(kp::shader_data::op_mul_comp_spv,
+        kp::shader_data::op_mul_comp_spv_len);
+
+    struct PushConstants {
+        uint32_t inAOff, inBOff, outOff;
+        int32_t ne00;
+        int32_t nb00, nb01, nb02, nb03;
+        int32_t ne10, ne11, ne12, ne13;
+        int32_t nb10, nb11, nb12, nb13;
+        int32_t ne0;
+        int32_t nb0, nb1, nb2, nb3;
+    } const pushConsts {
+        safe_divide(inAOff, 4), safe_divide(inBOff, 4), safe_divide(outOff, 4),
+        ne00,
+        nb00, nb01, nb02, nb03,
+        ne10, ne11, ne12, ne13,
+        nb10, nb11, nb12, nb13,
+        ne0,
+        nb0, nb1, nb2, nb3
+    };
+
+    std::shared_ptr<kp::Algorithm> s_algo = nullptr;
+    if (!komputeManager()->hasAlgorithm(__func__)) {
+        s_algo = komputeManager()->algorithm<float, PushConstants>(__func__, s_kompute_context->pool.get(), {inA, inB, out}, spirv, {unsigned(ne01), unsigned(ne02), unsigned(ne03)}, {}, {pushConsts});
+    } else {
+        s_algo = komputeManager()->getAlgorithm(__func__);
+        s_algo->setTensors({inA, inB, out});
+        s_algo->setWorkgroup({unsigned(ne01), unsigned(ne02), unsigned(ne03)});
+        s_algo->setPushConstants<PushConstants>({pushConsts});
+        s_algo->updateDescriptors(s_kompute_context->pool.get());
+    }
+    seq.record<kp::OpAlgoDispatch>(s_algo);
+}
+
+static void ggml_vk_scale(kp::Sequence& seq,
+                   const std::shared_ptr<kp::Tensor>& in,
+                   const std::shared_ptr<kp::Tensor>& out,
+                   uint32_t inOff, uint32_t outOff,
+                   uint32_t size, float scale) {
+    const static auto spirv_1 = getSpirvShader(
+        kp::shader_data::op_scale_comp_spv, kp::shader_data::op_scale_comp_spv_len
+    );
+    const static auto spirv_8 = getSpirvShader(
+        kp::shader_data::op_scale_8_comp_spv, kp::shader_data::op_scale_8_comp_spv_len
+    );
+
+    struct PushConstants {
+        uint32_t inOff, outOff;
+        float scale;
+    } const pushConsts {
+        safe_divide(inOff, 4), safe_divide(outOff, 4),
+        scale
+    };
+
+    const auto * spirv = &spirv_1;
+    std::string name(__func__);
+    if (size % 8 == 0) {
+        size /= 8;
+        name += "_8";
+        spirv = &spirv_8;
+    }
+
+    std::shared_ptr<kp::Algorithm> s_algo = nullptr;
+    if (!komputeManager()->hasAlgorithm(name)) {
+        s_algo = komputeManager()->algorithm<float, PushConstants>(name, s_kompute_context->pool.get(), {in, out}, *spirv, {size}, {}, {pushConsts});
+    } else {
+        s_algo = komputeManager()->getAlgorithm(name);
+        s_algo->setTensors({in, out});
+        s_algo->setWorkgroup({size});
+        s_algo->setPushConstants<PushConstants>({pushConsts});
+        s_algo->updateDescriptors(s_kompute_context->pool.get());
+    }
+    seq.record<kp::OpAlgoDispatch>(s_algo);
+}
+
+static void ggml_vk_xxlu(
+    const std::vector<uint32_t>& spirv, const char * suffix, kp::Sequence& seq,
+    const std::shared_ptr<kp::Tensor>& in,
+    const std::shared_ptr<kp::Tensor>& out,
+    uint32_t inOff, uint32_t outOff,
+    uint32_t size
+) {
+    struct PushConstants {
+        uint32_t inOff, outOff;
+    } const pushConsts {
+        safe_divide(inOff, 4), safe_divide(outOff, 4),
+    };
+
+    auto name = std::string(__func__) + "_" + suffix;
+    std::shared_ptr<kp::Algorithm> s_algo = nullptr;
+    if (!komputeManager()->hasAlgorithm(name)) {
+        s_algo = komputeManager()->algorithm<float, PushConstants>(name, s_kompute_context->pool.get(), {in, out}, spirv, {size}, {}, {pushConsts});
+    } else {
+        s_algo = komputeManager()->getAlgorithm(name);
+        s_algo->setTensors({in, out});
+        s_algo->setWorkgroup({size});
+        s_algo->setPushConstants<PushConstants>({pushConsts});
+        s_algo->updateDescriptors(s_kompute_context->pool.get());
+    }
+    seq.record<kp::OpAlgoDispatch>(s_algo);
+}
+
+template <typename... Args>
+static void ggml_vk_silu(Args&&... args) {
+    const static auto spirv = getSpirvShader(kp::shader_data::op_silu_comp_spv,
+        kp::shader_data::op_silu_comp_spv_len);
+
+    ggml_vk_xxlu(spirv, "silu", std::forward<Args>(args)...);
+}
+
+template <typename... Args>
+static void ggml_vk_relu(Args&&... args) {
+    const static auto spirv = getSpirvShader(kp::shader_data::op_relu_comp_spv,
+        kp::shader_data::op_relu_comp_spv_len);
+
+    ggml_vk_xxlu(spirv, "relu", std::forward<Args>(args)...);
+}
+
+template <typename... Args>
+static void ggml_vk_gelu(Args&&... args) {
+    const static auto spirv = getSpirvShader(kp::shader_data::op_gelu_comp_spv,
+        kp::shader_data::op_gelu_comp_spv_len);
+
+    ggml_vk_xxlu(spirv, "gelu", std::forward<Args>(args)...);
+}
+
+static void ggml_vk_soft_max(
+    kp::Sequence& seq,
+    const std::shared_ptr<kp::Tensor>& inA,
+    const std::shared_ptr<kp::Tensor>& inB,
+    const std::shared_ptr<kp::Tensor>& out,
+    uint32_t inAOff, uint32_t inBOff, uint32_t outOff,
+    int32_t ne00, int32_t ne01, int32_t ne02, uint32_t ne03,
+    float scale
+) {
+    const static auto spirv = getSpirvShader(kp::shader_data::op_softmax_comp_spv,
+        kp::shader_data::op_softmax_comp_spv_len);
+
+    struct PushConstants {
+        uint32_t inAOff, inBOff, outOff;
+        int32_t ne00, ne01, ne02;
+        float scale;
+        int32_t mask;
+    } pushConsts {
+        safe_divide(inAOff, 4), safe_divide(inBOff, 4), safe_divide(outOff, 4),
+        ne00, ne01, ne02,
+        scale,
+        bool(inB)
+    };
+
+    auto & inB_ = inB ? inB : inA;
+
+    std::shared_ptr<kp::Algorithm> s_algo = nullptr;
+    if (!komputeManager()->hasAlgorithm(__func__)) {
+        // FIXME: The softmax kernel needs to be fixed to use the subgroupsize which can vary by device
+        const uint32_t local_x = 32;
+        s_algo = komputeManager()->algorithm<uint32_t, PushConstants>(__func__, s_kompute_context->pool.get(), {inA, inB_, out}, spirv, {unsigned(ne01), unsigned(ne02), unsigned(ne03)}, {local_x}, {pushConsts});
+    } else {
+        s_algo = komputeManager()->getAlgorithm(__func__);
+        s_algo->setTensors({inA, inB_, out});
+        s_algo->setWorkgroup({unsigned(ne01), unsigned(ne02), unsigned(ne03)});
+        s_algo->setPushConstants<PushConstants>({pushConsts});
+        s_algo->updateDescriptors(s_kompute_context->pool.get());
+    }
+    seq.record<kp::OpAlgoDispatch>(s_algo);
+}
+
+static void ggml_vk_norm_(
+    const std::vector<uint32_t>& spirv, const char * suffix, kp::Sequence& seq,
+    const std::shared_ptr<kp::Tensor>& in,
+    const std::shared_ptr<kp::Tensor>& out,
+    uint32_t inOff, uint32_t outOff,
+    int32_t ne00, int32_t nb01,
+    int32_t nrows, float epsilon
+) {
+    GGML_ASSERT(nb01%sizeof(float) == 0);
+    GGML_ASSERT(ne00%sizeof(float) == 0);
+
+    struct PushConstants {
+        uint32_t inOff, outOff;
+        uint32_t ne00, nb01;
+        float eps;
+    } pushConsts {
+        safe_divide(inOff, 4), safe_divide(outOff, 4),
+        (uint32_t)ne00, (uint32_t)nb01, epsilon
+    };
+
+    auto name = std::string(__func__) + "_" + suffix;
+    std::shared_ptr<kp::Algorithm> s_algo = nullptr;
+    if (!komputeManager()->hasAlgorithm(name)) {
+        s_algo = komputeManager()->algorithm<float, PushConstants>(name, s_kompute_context->pool.get(), {in, out}, spirv, {(uint32_t)nrows}, {}, {pushConsts});
+    } else {
+        s_algo = komputeManager()->getAlgorithm(name);
+        s_algo->setTensors({in, out});
+        s_algo->setWorkgroup({(uint32_t)nrows});
+        s_algo->setPushConstants<PushConstants>({pushConsts});
+        s_algo->updateDescriptors(s_kompute_context->pool.get());
+    }
+    seq.record<kp::OpAlgoDispatch>(s_algo);
+}
+
+template <typename... Args>
+static void ggml_vk_norm(Args&&... args) {
+    const static auto spirv = getSpirvShader(kp::shader_data::op_norm_comp_spv,
+        kp::shader_data::op_norm_comp_spv_len);
+
+    ggml_vk_norm_(spirv, "norm", std::forward<Args>(args)...);
+}
+
+template <typename... Args>
+static void ggml_vk_rms_norm(Args&&... args) {
+    const static auto spirv = getSpirvShader(kp::shader_data::op_rmsnorm_comp_spv,
+        kp::shader_data::op_rmsnorm_comp_spv_len);
+
+    ggml_vk_norm_(spirv, "rms", std::forward<Args>(args)...);
+}
+
+static void ggml_vk_diag_mask_inf(kp::Sequence& seq,
+                           const std::shared_ptr<kp::Tensor>& in,
+                           const std::shared_ptr<kp::Tensor>& out,
+                           uint32_t inOff, uint32_t outOff,
+                           uint32_t n_past,
+                           int32_t ne00, int32_t ne01, int32_t ne02) {
+    const static auto spirv = getSpirvShader(kp::shader_data::op_diagmask_comp_spv,
+        kp::shader_data::op_diagmask_comp_spv_len);
+
+    struct PushConstants {
+        uint32_t inOff, outOff;
+        uint32_t n_past;
+        int32_t ne00, ne01;
+    } pushConsts {
+        safe_divide(inOff, 4), safe_divide(outOff, 4),
+        n_past,
+        ne00, ne01
+    };
+
+    std::shared_ptr<kp::Algorithm> s_algo = nullptr;
+    if (!komputeManager()->hasAlgorithm(__func__))
+        s_algo = komputeManager()->algorithm<float, PushConstants>(__func__, s_kompute_context->pool.get(), {in, out}, spirv, {unsigned(ne00), unsigned(ne01), unsigned(ne02)}, {}, {pushConsts});
+    else {
+        s_algo = komputeManager()->getAlgorithm(__func__);
+        s_algo->setTensors({in, out});
+        s_algo->setWorkgroup({unsigned(ne00), unsigned(ne01), unsigned(ne02)});
+        s_algo->setPushConstants<PushConstants>({pushConsts});
+        s_algo->updateDescriptors(s_kompute_context->pool.get());
+    }
+    seq.record<kp::OpAlgoDispatch>(s_algo);
+}
+
+static void ggml_vk_mul_mat_f16(
+    kp::Sequence& seq,
+    const std::shared_ptr<kp::Tensor>& inA,
+    const std::shared_ptr<kp::Tensor>& inB,
+    const std::shared_ptr<kp::Tensor>& out,
+    uint32_t inAOff, uint32_t inBOff, uint32_t outOff,
+    int32_t ne00, int32_t ne01, int32_t ne02,
+    uint32_t nb00, uint32_t nb01, uint32_t nb02,
+    int32_t ne10, int32_t ne11, int32_t ne12, int32_t ne13,
+    uint32_t nb10, uint32_t nb11, uint32_t nb12,
+    int32_t ne0, int32_t ne1,
+    uint32_t r2, uint32_t r3
+) {
+    const static auto spirv = getSpirvShader(kp::shader_data::op_mul_mat_f16_comp_spv,
+        kp::shader_data::op_mul_mat_f16_comp_spv_len);
+
+    struct PushConstants {
+        uint32_t inAOff, inBOff, outOff;
+        int32_t ne00, ne01, ne02;
+        uint32_t nb00, nb01, nb02;
+        int32_t ne10, ne11, ne12;
+        uint32_t nb10, nb11, nb12;
+        int32_t ne0, ne1;
+        uint32_t r2, r3;
+    } pushConsts {
+        safe_divide(inAOff, 2), safe_divide(inBOff, 4), safe_divide(outOff, 4),
+        ne00, ne01, ne02,
+        nb00, nb01, nb02,
+        ne10, ne11, ne12,
+        nb10, nb11, nb12,
+        ne0, ne1,
+        r2, r3
+    };
+
+    const unsigned ny = unsigned((ne11 + 4 - 1)/4);
+
+    std::shared_ptr<kp::Algorithm> s_algo = nullptr;
+    if (!komputeManager()->hasAlgorithm(__func__)) {
+        const uint32_t local_x = ggml_vk_current_device().subgroupSize * 2;
+        s_algo = komputeManager()->algorithm<uint32_t, PushConstants>(__func__, s_kompute_context->pool.get(), {inA, inB, out}, spirv, {unsigned(ne01), ny, unsigned(ne12*ne13)}, {local_x}, {pushConsts});
+    } else {
+        s_algo = komputeManager()->getAlgorithm(__func__);
+        s_algo->setTensors({inA, inB, out});
+        s_algo->setWorkgroup({unsigned(ne01), ny, unsigned(ne12*ne13)});
+        s_algo->setPushConstants<PushConstants>({pushConsts});
+        s_algo->updateDescriptors(s_kompute_context->pool.get());
+    }
+    seq.record<kp::OpAlgoDispatch>(s_algo);
+}
+
+static void ggml_vk_mul_mat_mat_f32(kp::Sequence& seq,
+                         const std::shared_ptr<kp::Tensor>& inA,
+                         const std::shared_ptr<kp::Tensor>& inB,
+                         const std::shared_ptr<kp::Tensor>& out,
+                         uint32_t inAOff, uint32_t inBOff, uint32_t outOff,
+                         int32_t ne00, int32_t ne01, int32_t ne02,
+                         uint32_t nb01, uint32_t nb02,
+                         int32_t ne11, int32_t ne12,
+                         uint32_t nb11, uint32_t nb12,
+                         uint32_t nb1, uint32_t nb2) {
+    const static auto spirv = getSpirvShader(kp::shader_data::op_mul_mat_mat_f32_comp_spv,
+        kp::shader_data::op_mul_mat_mat_f32_comp_spv_len);
+
+    struct PushConstants {
+        uint32_t inAOff, inBOff, outOff;
+        int32_t ne00, ne01, ne02, ne11, ne12;
+        uint32_t nb01, nb02;
+        uint32_t nb11, nb12;
+        uint32_t nb1, nb2;
+    } pushConsts {
+        safe_divide(inAOff, 4), safe_divide(inBOff, 4), safe_divide(outOff, 4),
+        ne00, ne01, ne02, ne11, ne12,
+        nb01, nb02, nb11, nb12,
+        nb1, nb2
+    };
+
+    const uint32_t local_x = ggml_vk_current_device().subgroupSize;
+    std::shared_ptr<kp::Algorithm> s_algo = nullptr;
+    if (!komputeManager()->hasAlgorithm(__func__)) {
+        s_algo = komputeManager()->algorithm<uint32_t, PushConstants>(__func__, s_kompute_context->pool.get(),
+        {inA, inB, out}, spirv,
+        {unsigned(ne01),
+         unsigned(ne11),
+         unsigned(std::max(ne12, ne02))
+         },
+        {local_x},
+        {pushConsts});
+    } else {
+        s_algo = komputeManager()->getAlgorithm(__func__);
+        s_algo->setTensors({inA, inB, out});
+        s_algo->setWorkgroup({unsigned(ne01),
+                              unsigned(ne11),
+                              unsigned(std::max(ne12, ne02)),
+                              });
+        s_algo->setPushConstants<PushConstants>({pushConsts});
+        s_algo->updateDescriptors(s_kompute_context->pool.get());
+    }
+    seq.record<kp::OpAlgoDispatch>(s_algo);
+}
+
+static void ggml_vk_mul_mat_impl(
+    const std::vector<uint32_t>& spirv, const char * suffix, uint32_t block_size, kp::Sequence& seq,
+    const std::shared_ptr<kp::Tensor>& inA,
+    const std::shared_ptr<kp::Tensor>& inB,
+    const std::shared_ptr<kp::Tensor>& out,
+    uint32_t inAOff, uint32_t inBOff, uint32_t outOff,
+    int32_t ne00, int32_t ne01, int32_t ne02,
+    int32_t ne10, int32_t ne11, int32_t ne12, int32_t ne13,
+    int32_t ne0, int32_t ne1,
+    uint32_t r2, uint32_t r3
+) {
+    struct PushConstants {
+        uint32_t inAOff, inBOff, outOff;
+        int32_t ne00, ne01, ne02;
+        int32_t ne10, ne12;
+        int32_t ne0, ne1;
+        uint32_t r2, r3;
+    } pushConsts {
+        safe_divide(inAOff, block_size), safe_divide(inBOff, 4), safe_divide(outOff, 4),
+        ne00, ne01, ne02,
+        ne10, ne12,
+        ne0, ne1,
+        r2, r3
+    };
+
+    auto name = std::string(__func__) + "_" + suffix;
+    std::shared_ptr<kp::Algorithm> s_algo = nullptr;
+    if (!komputeManager()->hasAlgorithm(name)) {
+        const uint32_t local_x = ggml_vk_current_device().subgroupSize * 2;
+        s_algo = komputeManager()->algorithm<uint32_t, PushConstants>(name, s_kompute_context->pool.get(), {inA, inB, out}, spirv, {unsigned((ne01 + 7)/8), unsigned(ne11), unsigned(ne12*ne13)}, {local_x}, {pushConsts});
+    } else {
+        s_algo = komputeManager()->getAlgorithm(name);
+        s_algo->setTensors({inA, inB, out});
+        s_algo->setWorkgroup({unsigned((ne01 + 7)/8), unsigned(ne11), unsigned(ne12*ne13)});
+        s_algo->setPushConstants<PushConstants>({pushConsts});
+        s_algo->updateDescriptors(s_kompute_context->pool.get());
+    }
+    seq.record<kp::OpAlgoDispatch>(s_algo);
+}
+
+template <typename... Args>
+static void ggml_vk_mul_mat_q4_0(Args&&... args) {
+    const static auto spirv = getSpirvShader(kp::shader_data::op_mul_mat_q4_0_comp_spv,
+        kp::shader_data::op_mul_mat_q4_0_comp_spv_len);
+
+    ggml_vk_mul_mat_impl(spirv, "q4_0", 1/*We access blocks unaligned*/, std::forward<Args>(args)...);
+}
+
+template <typename... Args>
+static void ggml_vk_mul_mat_q4_1(Args&&... args) {
+    const static auto spirv = getSpirvShader(kp::shader_data::op_mul_mat_q4_1_comp_spv,
+        kp::shader_data::op_mul_mat_q4_1_comp_spv_len);
+
+    ggml_vk_mul_mat_impl(spirv, "q4_1", 1/*We access blocks unaligned*/, std::forward<Args>(args)...);
+}
+
+template <typename... Args>
+static void ggml_vk_mul_mat_q8_0(Args&&... args) {
+    const static auto spirv = getSpirvShader(kp::shader_data::op_mul_mat_q8_0_comp_spv,
+        kp::shader_data::op_mul_mat_q8_0_comp_spv_len);
+
+    ggml_vk_mul_mat_impl(spirv, "q8_0", 1/*We access blocks unaligned*/, std::forward<Args>(args)...);
+}
+
+static void ggml_vk_mul_mat_q6_k(
+    kp::Sequence& seq,
+    const std::shared_ptr<kp::Tensor>& inA,
+    const std::shared_ptr<kp::Tensor>& inB,
+    const std::shared_ptr<kp::Tensor>& out,
+    uint32_t inAOff, uint32_t inBOff, uint32_t outOff,
+    int32_t ne00, int32_t ne10, int32_t ne0, int32_t ne1,
+    int32_t ne01, int32_t ne11, int32_t ne12, int32_t ne02
+) {
+    const static auto spirv = getSpirvShader(kp::shader_data::op_mul_mat_q6_k_comp_spv,
+        kp::shader_data::op_mul_mat_q6_k_comp_spv_len);
+
+    struct PushConstants {
+        uint32_t inAOff, inBOff, outOff;
+        int32_t ne00, ne10, ne0, ne1, ne01, gqa;
+    } pushConsts {
+        inAOff, safe_divide(inBOff, 4), safe_divide(outOff, 4),
+        ne00, ne10, ne0, ne1, ne01, ne12/ne02
+    };
+
+    std::shared_ptr<kp::Algorithm> s_algo = nullptr;
+    if (!komputeManager()->hasAlgorithm(__func__)) {
+        const uint32_t local_x = ggml_vk_current_device().subgroupSize * 2;
+        s_algo = komputeManager()->algorithm<uint32_t, PushConstants>(__func__, s_kompute_context->pool.get(), {inA, inB, out}, spirv, {unsigned((ne01 + 1)/2), unsigned(ne11), unsigned(ne12)}, {local_x}, {pushConsts});
+    } else {
+        s_algo = komputeManager()->getAlgorithm(__func__);
+        s_algo->setTensors({inA, inB, out});
+        s_algo->setWorkgroup({unsigned((ne01 + 1)/2), unsigned(ne11), unsigned(ne12)});
+        s_algo->setPushConstants<PushConstants>({pushConsts});
+        s_algo->updateDescriptors(s_kompute_context->pool.get());
+    }
+    seq.record<kp::OpAlgoDispatch>(s_algo);
+}
+
+static void ggml_vk_get_rows(
+    const std::vector<uint32_t>& spirv,
+    const char * suffix,
+    unsigned element_size, unsigned qk,
+    kp::Sequence& seq,
+    const std::shared_ptr<kp::Tensor>& inA,
+    const std::shared_ptr<kp::Tensor>& inB,
+    const std::shared_ptr<kp::Tensor>& out,
+    uint32_t inAOff, uint32_t inBOff, uint32_t outOff,
+    int32_t ne00, int32_t nb01, int32_t nb1,
+    uint32_t size
+) {
+    GGML_ASSERT(nb01%element_size == 0);
+    GGML_ASSERT(nb1%sizeof(float) == 0);
+    if (qk) GGML_ASSERT(ne00%qk == 0);
+
+    struct PushConstants {
+        uint32_t inAOff, inBOff, outOff;
+        int32_t ne00, nb01, nb1;
+    } pushConsts {
+        safe_divide(inAOff, element_size), safe_divide(inBOff, 4), safe_divide(outOff, 4),
+        ne00, nb01, nb1
+    };
+
+    auto name = std::string(__func__) + "_" + suffix;
+    std::shared_ptr<kp::Algorithm> s_algo = nullptr;
+    if (!komputeManager()->hasAlgorithm(name)) {
+        s_algo = komputeManager()->algorithm<float, PushConstants>(name, s_kompute_context->pool.get(), {inA, inB, out}, spirv, {size}, {}, {pushConsts});
+    } else {
+        s_algo = komputeManager()->getAlgorithm(name);
+        s_algo->setTensors({inA, inB, out});
+        s_algo->setWorkgroup({size});
+        s_algo->setPushConstants<PushConstants>({pushConsts});
+        s_algo->updateDescriptors(s_kompute_context->pool.get());
+    }
+    seq.record<kp::OpAlgoDispatch>(s_algo);
+}
+
+template <typename... Args>
+static void ggml_vk_get_rows_f32(Args&&... args) {
+    const static auto spirv = getSpirvShader(kp::shader_data::op_getrows_f32_comp_spv,
+        kp::shader_data::op_getrows_f32_comp_spv_len);
+
+    ggml_vk_get_rows(spirv, "f32", sizeof(float), 0, std::forward<Args>(args)...);
+}
+
+template <typename... Args>
+static void ggml_vk_get_rows_f16(Args&&... args) {
+    const static auto spirv = getSpirvShader(kp::shader_data::op_getrows_f16_comp_spv,
+        kp::shader_data::op_getrows_f16_comp_spv_len);
+
+    ggml_vk_get_rows(spirv, "f16", sizeof(half), 0, std::forward<Args>(args)...);
+}
+
+template <typename... Args>
+static void ggml_vk_get_rows_q4_0(Args&&... args) {
+    const static auto spirv = getSpirvShader(kp::shader_data::op_getrows_q4_0_comp_spv,
+        kp::shader_data::op_getrows_q4_0_comp_spv_len);
+
+    ggml_vk_get_rows(spirv, "q4_0", 1/*We access blocks unaligned*/, QK4_0, std::forward<Args>(args)...);
+}
+
+template <typename... Args>
+static void ggml_vk_get_rows_q4_1(Args&&... args) {
+    const static auto spirv = getSpirvShader(kp::shader_data::op_getrows_q4_1_comp_spv,
+        kp::shader_data::op_getrows_q4_1_comp_spv_len);
+
+    ggml_vk_get_rows(spirv, "q4_1", 1/*We access blocks unaligned*/, QK4_1, std::forward<Args>(args)...);
+}
+
+template <typename... Args>
+static void ggml_vk_get_rows_q6_k(Args&&... args) {
+    const static auto spirv = getSpirvShader(kp::shader_data::op_getrows_q6_k_comp_spv,
+        kp::shader_data::op_getrows_q6_k_comp_spv_len);
+    ggml_vk_get_rows(spirv, "q6_k", 1/*We access blocks unaligned*/, QK_NL, std::forward<Args>(args)...);
+}
+
+static void ggml_vk_rope(
+    kp::Sequence& seq,
+    const std::shared_ptr<kp::Tensor>& inA,
+    const std::shared_ptr<kp::Tensor>& inB,
+    const std::shared_ptr<kp::Tensor>& out,
+    uint32_t inAOff, uint32_t inBOff, uint32_t outOff,
+    ggml_type src0t, int32_t n_dims, int32_t mode, int32_t n_ctx_orig,
+    float freq_base, float freq_scale, float ext_factor, float attn_factor, float beta_fast, float beta_slow,
+    int32_t ne01, int32_t ne02, int32_t ne03,
+    uint32_t nb00, uint32_t nb01, uint32_t nb02, uint32_t nb03,
+    int32_t ne0,
+    uint32_t nb0, uint32_t nb1, uint32_t nb2, uint32_t nb3
+) {
+    GGML_ASSERT(src0t == GGML_TYPE_F16 || src0t == GGML_TYPE_F32);
+
+    static const auto spirv_f16 = getSpirvShader(
+        kp::shader_data::op_rope_f16_comp_spv, kp::shader_data::op_rope_f16_comp_spv_len
+    );
+    static const auto spirv_f32 = getSpirvShader(
+        kp::shader_data::op_rope_f32_comp_spv, kp::shader_data::op_rope_f32_comp_spv_len
+    );
+
+    int type_size = src0t == GGML_TYPE_F16 ? 2 : 4;
+
+    GGML_ASSERT(nb03 % type_size == 0);
+    GGML_ASSERT(nb02 % type_size == 0);
+    GGML_ASSERT(nb01 % type_size == 0);
+    GGML_ASSERT(nb00 % type_size == 0);
+    GGML_ASSERT(nb3  % type_size == 0);
+    GGML_ASSERT(nb2  % type_size == 0);
+    GGML_ASSERT(nb1  % type_size == 0);
+    GGML_ASSERT(nb0  % type_size == 0);
+
+    struct PushConstants {
+        uint32_t inAOff, inBOff, outOff;
+        int32_t n_dims, mode, n_ctx_orig;
+        float freq_base, freq_scale, ext_factor, attn_factor, beta_fast, beta_slow;
+        uint32_t nb00, nb01, nb02, nb03;
+        int32_t ne0;
+        uint32_t nb0, nb1, nb2, nb3;
+    } pushConsts {
+        safe_divide(inAOff, type_size), safe_divide(inBOff, 4), safe_divide(outOff, type_size),
+        n_dims, mode, n_ctx_orig,
+        freq_base, freq_scale, ext_factor, attn_factor, beta_fast, beta_slow,
+        nb00, nb01, nb02, nb03,
+        ne0,
+        nb0, nb1, nb2, nb3
+    };
+
+    auto name = std::string(__func__) + (src0t == GGML_TYPE_F16 ? "_f16" : "_f32");
+    std::shared_ptr<kp::Algorithm> s_algo = nullptr;
+    if (!komputeManager()->hasAlgorithm(name)) {
+        s_algo = komputeManager()->algorithm<float, PushConstants>(
+            name, s_kompute_context->pool.get(), {inA, inB, out},
+            src0t == GGML_TYPE_F16 ? spirv_f16 : spirv_f32,
+            {unsigned(ne01), unsigned(ne02), unsigned(ne03)}, {}, {pushConsts}
+        );
+    } else {
+        s_algo = komputeManager()->getAlgorithm(name);
+        s_algo->setTensors({inA, inB, out});
+        s_algo->setWorkgroup({unsigned(ne01), unsigned(ne02), unsigned(ne03)});
+        s_algo->setPushConstants<PushConstants>({pushConsts});
+        s_algo->updateDescriptors(s_kompute_context->pool.get());
+    }
+    seq.record<kp::OpAlgoDispatch>(s_algo);
+}
+
+static void ggml_vk_cpy(
+    const std::vector<uint32_t>& spirv,
+    uint32_t in_element_size, uint32_t out_element_size,
+    kp::Sequence& seq,
+    const std::shared_ptr<kp::Tensor>& in,
+    const std::shared_ptr<kp::Tensor>& out,
+    uint32_t inOff, uint32_t outOff,
+    int32_t ne00, int32_t ne01, int32_t ne02, int32_t ne03,
+    uint32_t nb00, uint32_t nb01, uint32_t nb02, uint32_t nb03,
+    int32_t ne0, int32_t ne1, int32_t ne2,
+    uint32_t nb0, uint32_t nb1, uint32_t nb2, uint32_t nb3
+) {
+    struct PushConstants {
+        uint32_t inOff, outOff;
+        int32_t ne00, ne01, ne02;
+        uint32_t nb00, nb01, nb02, nb03;
+        int32_t ne0, ne1, ne2;
+        uint32_t nb0, nb1, nb2, nb3;
+    } pushConsts {
+        safe_divide(inOff, in_element_size), safe_divide(outOff, out_element_size),
+        ne00, ne01, ne02,
+        nb00, nb01, nb02, nb03,
+        ne0, ne1, ne2,
+        nb0, nb1, nb2, nb3
+    };
+
+    std::string name = std::string(__func__)
+                       + "_i_" + std::to_string(in_element_size)
+                       + "_o_" + std::to_string(out_element_size);
+    std::shared_ptr<kp::Algorithm> s_algo = nullptr;
+    if (!komputeManager()->hasAlgorithm(name))
+        s_algo = komputeManager()->algorithm<float, PushConstants>(name, s_kompute_context->pool.get(), {in, out}, spirv, {unsigned(ne01), unsigned(ne02), unsigned(ne03)}, {}, {pushConsts});
+    else {
+        s_algo = komputeManager()->getAlgorithm(name);
+        s_algo->setTensors({in, out});
+        s_algo->setWorkgroup({unsigned(ne01), unsigned(ne02), unsigned(ne03)});
+        s_algo->setPushConstants<PushConstants>({pushConsts});
+        s_algo->updateDescriptors(s_kompute_context->pool.get());
+    }
+    seq.record<kp::OpAlgoDispatch>(s_algo);
+}
+
+template <typename... Args>
+static void ggml_vk_cpy_f32_f16(Args&&... args) {
+    const static auto spirv = getSpirvShader(kp::shader_data::op_cpy_f32_f16_comp_spv,
+        kp::shader_data::op_cpy_f32_f16_comp_spv_len);
+    ggml_vk_cpy(spirv, 4, 2, std::forward<Args>(args)...);
+}
+
+template <typename... Args>
+static void ggml_vk_cpy_f32_f32(Args&&... args) {
+    const static auto spirv = getSpirvShader(kp::shader_data::op_cpy_f32_f32_comp_spv,
+        kp::shader_data::op_cpy_f32_f32_comp_spv_len);
+    ggml_vk_cpy(spirv, 4, 4, std::forward<Args>(args)...);
+}
+
+template <typename... Args>
+static void ggml_vk_cpy_f16_f16(Args&&... args) {
+    const static auto spirv = getSpirvShader(kp::shader_data::op_cpy_f16_f16_comp_spv,
+        kp::shader_data::op_cpy_f16_f16_comp_spv_len);
+    ggml_vk_cpy(spirv, 2, 2, std::forward<Args>(args)...);
+}
+
+template <typename... Args>
+static void ggml_vk_cpy_f16_f32(Args&&... args) {
+    const static auto spirv = getSpirvShader(kp::shader_data::op_cpy_f16_f32_comp_spv,
+        kp::shader_data::op_cpy_f16_f32_comp_spv_len);
+    ggml_vk_cpy(spirv, 2, 4, std::forward<Args>(args)...);
+}
+
+static bool ggml_vk_supports_op(const struct ggml_tensor * op) {
+    switch (op->type) {
+        case GGML_TYPE_F16:
+        case GGML_TYPE_F32:
+        case GGML_TYPE_Q4_0:
+        case GGML_TYPE_Q4_1:
+            break;
+        default:
+            return false;
+    }
+
+    switch (op->op) {
+        case GGML_OP_UNARY:
+            switch (ggml_get_unary_op(op)) {
+                case GGML_UNARY_OP_RELU:
+                case GGML_UNARY_OP_GELU:
+                case GGML_UNARY_OP_SILU:
+                    return ggml_is_contiguous(op->src[0]);
+                default:
+                    ;
+            }
+            break;
+        case GGML_OP_NONE:
+        case GGML_OP_RESHAPE:
+        case GGML_OP_VIEW:
+        case GGML_OP_TRANSPOSE:
+        case GGML_OP_PERMUTE:
+        case GGML_OP_ADD:
+        case GGML_OP_MUL:
+        case GGML_OP_SCALE:
+        case GGML_OP_SOFT_MAX:
+        case GGML_OP_RMS_NORM:
+        case GGML_OP_NORM:
+        case GGML_OP_ROPE:
+            return true;
+        case GGML_OP_DUP:
+        case GGML_OP_CPY:
+        case GGML_OP_CONT:
+            switch (op->src[0]->type) {
+                case GGML_TYPE_F32:
+                case GGML_TYPE_F16:
+                    break;
+                default:
+                    return false;
+            }
+            switch (op->type) {
+                case GGML_TYPE_F32:
+                case GGML_TYPE_F16:
+                    break;
+                default:
+                    return false;
+            }
+            return true;
+        case GGML_OP_DIAG_MASK_INF:
+            return op->ne[3] == 1;
+        case GGML_OP_GET_ROWS:
+            switch (op->src[0]->type) {
+                case GGML_TYPE_F32:
+                case GGML_TYPE_F16:
+                case GGML_TYPE_Q4_0:
+                case GGML_TYPE_Q4_1:
+                case GGML_TYPE_Q6_K:
+                    return op->ne[2] == 1 && op->ne[3] == 1;
+                default:
+                    ;
+            }
+            return false;
+        case GGML_OP_MUL_MAT:
+            if (op->src[1]->type != GGML_TYPE_F32 || ggml_is_transposed(op->src[0]) || ggml_is_transposed(op->src[1]))
+                return false;
+
+            switch (op->src[0]->type) {
+                case GGML_TYPE_F32:
+                case GGML_TYPE_Q6_K:
+                    return op->ne[3] == 1;
+                case GGML_TYPE_F16:
+                case GGML_TYPE_Q8_0:
+                case GGML_TYPE_Q4_0:
+                case GGML_TYPE_Q4_1:
+                    return true;
+                default:
+                    ;
+            }
+        default:
+            ;
+    }
+    return false;
+}
+
+static void ggml_vk_graph_compute(struct ggml_kompute_context * ctx, struct ggml_cgraph * gf) {
+    const int n_seq = 8;
+
+    // FIXME: Figure out if we can somehow optimize the size of the pool... right now we're setting
+    // it to the size of the graph, but I think it can be made smaller?
+    ggml_vk_allocate_descriptor_pool(ctx, gf->n_nodes);
+
+    std::vector<std::shared_ptr<kp::Sequence>> sequences(n_seq);
+
+    for (auto& sequence : sequences) {
+        sequence = komputeManager()->sequence();
+    }
+    for (int seq_idx = 0; seq_idx < n_seq; ++seq_idx) {
+        const int n_nodes_per_seq = (gf->n_nodes + n_seq - 1) / n_seq;
+
+        auto& seq = *sequences[seq_idx];
+
+        const int node_start = (seq_idx + 0) * n_nodes_per_seq;
+        const int node_end   = std::min((seq_idx == n_seq - 1) ? gf->n_nodes : (seq_idx + 1) * n_nodes_per_seq, gf->n_nodes);
+
+        bool any_commands_recorded = false;
+
+        for (int i = node_start; i < node_end; ++i) {
+            struct ggml_tensor * src0 = gf->nodes[i]->src[0];
+            struct ggml_tensor * src1 = gf->nodes[i]->src[1];
+            struct ggml_tensor * src2 = gf->nodes[i]->src[2]; GGML_UNUSED(src2);
+            struct ggml_tensor * dst = gf->nodes[i];
+            GGML_ASSERT(dst->data != nullptr);
+
+            if (ggml_is_empty(dst)) {
+                continue;
+            }
+
+            switch (dst->op) {
+                case GGML_OP_NONE:
+                case GGML_OP_RESHAPE:
+                case GGML_OP_VIEW:
+                case GGML_OP_TRANSPOSE:
+                case GGML_OP_PERMUTE:
+                    continue; // noop -> next node
+                default:
+                    break;
+            }
+
+            any_commands_recorded = true;
+
+            if (!ggml_vk_supports_op(dst)) {
+                 fprintf(stderr, "%s: error: unsupported op '%s'\n", __func__, ggml_op_desc(dst));
+                 GGML_ABORT("unsupported op");
+            }
+
+            const int32_t ne00 = src0 ? src0->ne[0] : 0;
+            const int32_t ne01 = src0 ? src0->ne[1] : 0;
+            const int32_t ne02 = src0 ? src0->ne[2] : 0;
+            const int32_t ne03 = src0 ? src0->ne[3] : 0;
+
+            const uint32_t nb00 = src0 ? src0->nb[0] : 0;
+            const uint32_t nb01 = src0 ? src0->nb[1] : 0;
+            const uint32_t nb02 = src0 ? src0->nb[2] : 0;
+            const uint32_t nb03 = src0 ? src0->nb[3] : 0;
+
+            const int32_t ne10 = src1 ? src1->ne[0] : 0;
+            const int32_t ne11 = src1 ? src1->ne[1] : 0;
+            const int32_t ne12 = src1 ? src1->ne[2] : 0;
+            const int32_t ne13 = src1 ? src1->ne[3] : 0;
+
+            const uint32_t nb10 = src1 ? src1->nb[0] : 0;
+            const uint32_t nb11 = src1 ? src1->nb[1] : 0;
+            const uint32_t nb12 = src1 ? src1->nb[2] : 0;
+            const uint32_t nb13 = src1 ? src1->nb[3] : 0;
+
+            const int32_t ne0 = dst ? dst->ne[0] : 0;
+            const int32_t ne1 = dst ? dst->ne[1] : 0;
+            const int32_t ne2 = dst ? dst->ne[2] : 0;
+//            const int32_t ne3 = dst ? dst->ne[3] : 0;
+
+            const uint32_t nb0 = dst ? dst->nb[0] : 0;
+            const uint32_t nb1 = dst ? dst->nb[1] : 0;
+            const uint32_t nb2 = dst ? dst->nb[2] : 0;
+            const uint32_t nb3 = dst ? dst->nb[3] : 0;
+
+            const enum ggml_type src0t = src0 ? src0->type : GGML_TYPE_COUNT;
+            const enum ggml_type src1t = src1 ? src1->type : GGML_TYPE_COUNT;
+            const enum ggml_type dstt = dst ? dst->type : GGML_TYPE_COUNT;
+
+            const static std::shared_ptr<kp::Tensor> nullTensor = nullptr;
+            uint32_t off_src0 = 0;
+            uint32_t off_src1 = 0;
+            uint32_t off_dst  = 0;
+            const std::shared_ptr<kp::Tensor>& id_src0 = src0 ? ggml_vk_get_tensor(src0, &off_src0) : nullTensor;
+            const std::shared_ptr<kp::Tensor>& id_src1 = src1 ? ggml_vk_get_tensor(src1, &off_src1) : nullTensor;
+            const std::shared_ptr<kp::Tensor>& id_dst  = dst  ? ggml_vk_get_tensor(dst,  &off_dst)  : nullTensor;
+
+            switch (dst->op) {
+                case GGML_OP_ADD:
+                    {
+                        if (ggml_nelements(src1) == ne10 && ggml_is_contiguous(src1) && ne00 % 4 == 0 && ne10 % 4 == 0) {
+                            // src1 is a row
+                            ggml_vk_addrow(seq, id_src0, id_src1, id_dst, off_src0, off_src1, off_dst, ggml_nelements(dst)/4, ne00);
+                        } else {
+                            ggml_vk_add(
+                                seq, id_src0, id_src1, id_dst, off_src0, off_src1, off_dst,
+                                ne00, ne01, ne02, ne03,
+                                nb00, nb01, nb02, nb03,
+                                ne10, ne11, ne12, ne13,
+                                nb10, nb11, nb12, nb13,
+                                ne0,
+                                nb0, nb1, nb2, nb3
+                            );
+                        }
+                    } break;
+                case GGML_OP_MUL:
+                    {
+                        ggml_vk_mul(
+                            seq, id_src0, id_src1, id_dst, off_src0, off_src1, off_dst,
+                            ne00, ne01, ne02, ne03,
+                            nb00, nb01, nb02, nb03,
+                            ne10, ne11, ne12, ne13,
+                            nb10, nb11, nb12, nb13,
+                            ne0,
+                            nb0, nb1, nb2, nb3
+                        );
+                    } break;
+                case GGML_OP_SCALE:
+                    {
+                        float scale; memcpy(&scale, dst->op_params, sizeof(float));
+
+                        ggml_vk_scale(seq, id_src0, id_dst, off_src0, off_dst, ggml_nelements(dst), scale);
+                    } break;
+                case GGML_OP_UNARY:
+                    {
+                        int64_t n = ggml_nelements(dst);
+                        GGML_ASSERT(n % 4 == 0);
+                        switch (ggml_get_unary_op(gf->nodes[i])) {
+                            case GGML_UNARY_OP_SILU:
+                                {
+                                    ggml_vk_silu(seq, id_src0, id_dst, off_src0, off_dst, n/4);
+                                } break;
+                            case GGML_UNARY_OP_RELU:
+                                {
+                                    ggml_vk_relu(seq, id_src0, id_dst, off_src0, off_dst, n/4);
+                                } break;
+                            case GGML_UNARY_OP_GELU:
+                                {
+                                    GGML_ASSERT(n % 8 == 0);
+                                    ggml_vk_gelu(seq, id_src0, id_dst, off_src0, off_dst, n/8);
+                                } break;
+                            default:
+                                {
+                                    fprintf(stderr, "%s: node %3d, op = %8s not implemented\n", __func__, i, ggml_op_name(dst->op));
+                                    GGML_ABORT("fatal error");
+                                }
+                        }
+                    } break;
+                case GGML_OP_SOFT_MAX:
+                    {
+                        float scale;
+                        float max_bias;
+
+                        memcpy(&scale,    (float *)dst->op_params + 0, sizeof(float));
+                        memcpy(&max_bias, (float *)dst->op_params + 1, sizeof(float));
+
+#pragma message("TODO: add ggml_vk_soft_max() F16 src1 support")
+#pragma message("ref:  https://github.com/ggerganov/llama.cpp/pull/5021")
+                        GGML_ASSERT(!src1 || src1t == GGML_TYPE_F32);
+
+#pragma message("TODO: add ALiBi support")
+#pragma message("ref:  https://github.com/ggerganov/llama.cpp/pull/7192")
+                        GGML_ASSERT(max_bias == 0.0f);
+
+                        ggml_vk_soft_max(seq, id_src0, id_src1, id_dst, off_src0, off_src1, off_dst, ne00, ne01, ne02, ne03, scale);
+                    } break;
+                case GGML_OP_DIAG_MASK_INF:
+                    {
+                        const int n_past = ((int32_t *)(dst->op_params))[0];
+                        ggml_vk_diag_mask_inf(seq, id_src0, id_dst, off_src0, off_dst, n_past, ne00, ne01, ne02);
+                    } break;
+                case GGML_OP_NORM:
+                    {
+                        float eps;
+                        memcpy(&eps, dst->op_params, sizeof(float));
+                        ggml_vk_norm(seq, id_src0, id_dst, off_src0, off_dst, ne00, nb01, ggml_nrows(src0), eps);
+                    } break;
+                case GGML_OP_RMS_NORM:
+                    {
+                        GGML_ASSERT(ne00 % 4 == 0);
+
+                        float eps;
+                        memcpy(&eps, dst->op_params, sizeof(float));
+                        ggml_vk_rms_norm(seq, id_src0, id_dst, off_src0, off_dst, ne00, nb01, ggml_nrows(src0), eps);
+                    } break;
+                case GGML_OP_MUL_MAT:
+                    {
+                        GGML_ASSERT(ne00 == ne10);
+
+                        GGML_ASSERT(ne12 % ne02 == 0);
+                        GGML_ASSERT(ne13 % ne03 == 0);
+
+                        const uint32_t r2 = ne12/ne02;
+                        const uint32_t r3 = ne13/ne03;
+
+                        if (src1t != GGML_TYPE_F32) {
+                            fprintf(stderr, "%s: %s: Unsupported src1 type: %u/%u\n", __func__, ggml_op_name(dst->op), src0t, src1t);
+                            goto not_implemented;
+                        }
+
+                        if (ggml_is_transposed(src0) ||
+                            ggml_is_transposed(src1)) {
+                            fprintf(stderr, "%s: %s: matmul on tranposed tensor not supported: %u/%u\n", __func__, ggml_op_name(dst->op), src0t, src1t);
+                            goto not_implemented;
+                        }
+
+                        switch (src0t) {
+                            case GGML_TYPE_F32:
+                                ggml_vk_mul_mat_mat_f32(
+                                    seq, id_src0, id_src1, id_dst, off_src0, off_src1, off_dst,
+                                    ne00, ne01, ne02, nb01, nb02, ne11, ne12, nb11, nb12, nb1, nb2
+                                );
+                                break;
+                            case GGML_TYPE_F16:
+                                ggml_vk_mul_mat_f16(
+                                    seq, id_src0, id_src1, id_dst, off_src0, off_src1, off_dst,
+                                    ne00, ne01, ne02, nb00, nb01, nb02, ne10, ne11, ne12, ne13, nb10, nb11, nb12,
+                                    ne0, ne1, r2, r3
+                                );
+                                break;
+                            case GGML_TYPE_Q8_0:
+                                ggml_vk_mul_mat_q8_0(
+                                    seq, id_src0, id_src1, id_dst, off_src0, off_src1, off_dst,
+                                    ne00, ne01, ne02, ne10, ne11, ne12, ne13, ne0, ne1, r2, r3
+                                );
+                                break;
+                            case GGML_TYPE_Q4_0:
+                                ggml_vk_mul_mat_q4_0(
+                                    seq, id_src0, id_src1, id_dst, off_src0, off_src1, off_dst,
+                                    ne00, ne01, ne02, ne10, ne11, ne12, ne13, ne0, ne1, r2, r3
+                                );
+                                break;
+                            case GGML_TYPE_Q4_1:
+                                ggml_vk_mul_mat_q4_1(
+                                    seq, id_src0, id_src1, id_dst, off_src0, off_src1, off_dst,
+                                    ne00, ne01, ne02, ne10, ne11, ne12, ne13, ne0, ne1, r2, r3
+                                );
+                                break;
+                            case GGML_TYPE_Q6_K:
+                                ggml_vk_mul_mat_q6_k(
+                                    seq, id_src0, id_src1, id_dst, off_src0, off_src1, off_dst,
+                                    ne00, ne10, ne0, ne1, ne01, ne11, ne12, ne02
+                                );
+                                break;
+                            default: {
+                                fprintf(stderr, "%s: %s: Unsupported quantization: %u/%u\n", __func__, ggml_op_name(dst->op), src0t, src1t);
+                                goto not_implemented;
+                            }
+                        }
+
+                    } break;
+                case GGML_OP_GET_ROWS:
+                    {
+                        if (src0t == GGML_TYPE_F32) {
+                            ggml_vk_get_rows_f32(seq, id_src0, id_src1, id_dst, off_src0, off_src1, off_dst, ne00, nb01, nb1, ggml_nelements(src1));
+                        } else if (src0t == GGML_TYPE_F16) {
+                            ggml_vk_get_rows_f16(seq, id_src0, id_src1, id_dst, off_src0, off_src1, off_dst, ne00, nb01, nb1, ggml_nelements(src1));
+                        } else if (src0t == GGML_TYPE_Q4_0) {
+                            ggml_vk_get_rows_q4_0(seq, id_src0, id_src1, id_dst, off_src0, off_src1, off_dst, ne00, nb01, nb1, ggml_nelements(src1));
+                        } else if (src0t == GGML_TYPE_Q4_1) {
+                            ggml_vk_get_rows_q4_1(seq, id_src0, id_src1, id_dst, off_src0, off_src1, off_dst, ne00, nb01, nb1, ggml_nelements(src1));
+                        } else if (src0t == GGML_TYPE_Q6_K) {
+                            ggml_vk_get_rows_q6_k(seq, id_src0, id_src1, id_dst, off_src0, off_src1, off_dst, ne00, nb01, nb1, ggml_nelements(src1));
+                        } else {
+                            fprintf(stderr, "%s: %s: Unsupported quantization: %u\n", __func__, ggml_op_name(dst->op), src0t);
+                            goto not_implemented;
+                        }
+                    } break;
+                case GGML_OP_ROPE:
+                    {
+#pragma message("TODO: implement phi3 frequency factors support")
+#pragma message("      https://github.com/ggerganov/llama.cpp/pull/7225")
+                        GGML_ASSERT(dst->src[2] == nullptr && "phi3 frequency factors not implemented yet");
+
+#pragma message("TODO: update rope NORM mode to match NEOX mode")
+#pragma message("      https://github.com/ggerganov/llama.cpp/pull/7634")
+
+                        GGML_ASSERT(ne10 == ne02);
+                        GGML_ASSERT(src0t == dstt);
+                        // const int n_past = ((int32_t *) dst->op_params)[0];
+                        const int n_dims     = ((int32_t *) dst->op_params)[1];
+                        const int mode       = ((int32_t *) dst->op_params)[2];
+                        // skip 3, n_ctx used in GLM RoPE, unimplemented in Vulkan
+                        const int n_ctx_orig = ((int32_t *) dst->op_params)[4];
+
+                        float freq_base, freq_scale, ext_factor, attn_factor, beta_fast, beta_slow;
+                        memcpy(&freq_base,   (int32_t *) dst->op_params +  5, sizeof(float));
+                        memcpy(&freq_scale,  (int32_t *) dst->op_params +  6, sizeof(float));
+                        memcpy(&ext_factor,  (int32_t *) dst->op_params +  7, sizeof(float));
+                        memcpy(&attn_factor, (int32_t *) dst->op_params +  8, sizeof(float));
+                        memcpy(&beta_fast,   (int32_t *) dst->op_params +  9, sizeof(float));
+                        memcpy(&beta_slow,   (int32_t *) dst->op_params + 10, sizeof(float));
+                        ggml_vk_rope(
+                            seq, id_src0, id_src1, id_dst, off_src0, off_src1, off_dst, src0t, n_dims, mode, n_ctx_orig,
+                            freq_base, freq_scale, ext_factor, attn_factor, beta_fast, beta_slow,
+                            ne01, ne02, ne03, nb00, nb01, nb02, nb03, ne0, nb0, nb1, nb2, nb3
+                        );
+                    } break;
+                case GGML_OP_DUP:
+                case GGML_OP_CPY:
+                case GGML_OP_CONT:
+                    {
+                        switch (src0t) {
+                            case GGML_TYPE_F32:
+                                {
+                                    switch (dstt) {
+                                        case GGML_TYPE_F16: ggml_vk_cpy_f32_f16(seq, id_src0, id_dst, off_src0, off_dst, ne00, ne01, ne02, ne03, nb00, nb01, nb02, nb03, ne0, ne1, ne2, nb0, nb1, nb2, nb3); break;
+                                        case GGML_TYPE_F32: ggml_vk_cpy_f32_f32(seq, id_src0, id_dst, off_src0, off_dst, ne00, ne01, ne02, ne03, nb00, nb01, nb02, nb03, ne0, ne1, ne2, nb0, nb1, nb2, nb3); break;
+                                        default: goto not_implemented;
+                                    }
+                                } break;
+                            case GGML_TYPE_F16:
+                                {
+                                    switch (dstt) {
+                                        case GGML_TYPE_F16: ggml_vk_cpy_f16_f16(seq, id_src0, id_dst, off_src0, off_dst, ne00, ne01, ne02, ne03, nb00, nb01, nb02, nb03, ne0, ne1, ne2, nb0, nb1, nb2, nb3); break;
+                                        case GGML_TYPE_F32: ggml_vk_cpy_f16_f32(seq, id_src0, id_dst, off_src0, off_dst, ne00, ne01, ne02, ne03, nb00, nb01, nb02, nb03, ne0, ne1, ne2, nb0, nb1, nb2, nb3); break;
+                                    default: goto not_implemented;
+                                } break;
+                            default: goto not_implemented;
+                            }
+                        }
+                    } break;
+                default: goto not_implemented;
+            }
+            continue;
+            not_implemented: {}
+            fprintf(stderr, "%s: node %3d, op = %8s not implemented\n", __func__, i, ggml_op_name(dst->op));
+            //GGML_ABORT("fatal error");
+        }
+
+        // Evaluate sequence
+        if (any_commands_recorded) {
+            seq.evalAsync();
+        }
+    }
+
+    // Wait for all sequences to finish
+    for (auto& sequence : sequences) {
+        if (sequence->isRunning())
+            sequence->evalAwait();
+    }
+
+    ggml_vk_free_descriptor_pool(ctx);
+}
+
+template<>
+kp::Tensor::TensorDataTypes
+kp::TensorT<half>::dataType()
+{
+    return TensorDataTypes::eFloat;
+}
+
+template<>
+kp::Tensor::TensorDataTypes
+kp::TensorT<uint8_t>::dataType()
+{
+    return TensorDataTypes::eUnsignedInt;
+}
+
+////////////////////////////////////////////////////////////////////////////////
+
+// backend interface
+
+struct ggml_backend_kompute_buffer_type_context {
+    int         device;
+    int         device_ref = 0;
+    uint64_t    buffer_alignment;
+    uint64_t    max_alloc;
+    std::string name;
+
+    ggml_backend_kompute_buffer_type_context(int device, uint64_t buffer_alignment, uint64_t max_alloc)
+        : device(device), buffer_alignment(buffer_alignment), max_alloc(max_alloc), name(ggml_kompute_format_name(device)) {}
+};
+
+static void ggml_backend_kompute_device_ref(ggml_backend_buffer_type_t buft) {
+    auto * ctx = static_cast<ggml_backend_kompute_buffer_type_context *>(buft->context);
+
+    if (!ctx->device_ref) {
+        komputeManager()->initializeDevice(
+            ctx->device, {}, {
+                "VK_KHR_shader_float16_int8", "VK_KHR_8bit_storage",
+                "VK_KHR_16bit_storage", "VK_KHR_shader_non_semantic_info"
+            }
+        );
+    }
+
+    assert(ggml_vk_has_device());
+    ctx->device_ref++;
+}
+
+static void ggml_backend_kompute_device_unref(ggml_backend_buffer_type_t buft) {
+    auto * ctx = static_cast<ggml_backend_kompute_buffer_type_context *>(buft->context);
+
+    assert(ctx->device_ref > 0);
+
+    ctx->device_ref--;
+
+    if (!ctx->device_ref) {
+        komputeManager.destroy();
+    }
+}
+
+static const char * ggml_backend_kompute_buffer_get_name(ggml_backend_buffer_t buffer) {
+    auto * ctx = static_cast<ggml_backend_kompute_buffer_type_context *>(buffer->buft->context);
+    return ctx->name.c_str();
+}
+
+static void ggml_backend_kompute_buffer_free_buffer(ggml_backend_buffer_t buffer) {
+    auto * memory = (ggml_vk_memory *)buffer->context;
+    if (ggml_vk_has_device()) {
+        ggml_vk_free_memory(*memory);
+    }
+    delete memory;
+}
+
+static void * ggml_backend_kompute_buffer_get_base(ggml_backend_buffer_t buffer) {
+    return ((ggml_vk_memory *)buffer->context)->data;
+}
+
+static void ggml_backend_kompute_buffer_set_tensor(ggml_backend_buffer_t buffer, ggml_tensor * tensor, const void * data, size_t offset, size_t size) {
+    GGML_UNUSED(buffer);
+
+    const auto res = ggml_vk_get_tensor(tensor);
+    GGML_ASSERT(res);
+
+    memcpy((char *)tensor->data + offset, data, size);
+
+    komputeManager()->sequence()->eval<kp::OpTensorSyncDevice>({res});
+}
+
+static void ggml_backend_kompute_buffer_get_tensor(ggml_backend_buffer_t buffer, const ggml_tensor * tensor, void * data, size_t offset, size_t size) {
+    GGML_UNUSED(buffer);
+
+    const auto res = ggml_vk_get_tensor(tensor);
+    GGML_ASSERT(res);
+
+    komputeManager()->sequence()->eval<kp::OpTensorSyncLocal>({res});
+
+    memcpy(data, (const char *)tensor->data + offset, size);
+}
+
+static void ggml_backend_kompute_buffer_clear(ggml_backend_buffer_t buffer, uint8_t value) {
+    auto * memory = (ggml_vk_memory *)buffer->context;
+    memset(memory->data, value, buffer->size);
+
+    if (memory->stagingBuffer)
+        komputeManager()->sequence()->eval<kp::OpBufferSyncDevice>(memory->primaryBuffer, memory->stagingBuffer, memory->size);
+}
+
+static ggml_backend_buffer_i ggml_backend_kompute_buffer_i = {
+    /* .get_name        = */ ggml_backend_kompute_buffer_get_name,
+    /* .free_buffer     = */ ggml_backend_kompute_buffer_free_buffer,
+    /* .get_base        = */ ggml_backend_kompute_buffer_get_base,
+    /* .init_tensor     = */ NULL,
+    /* .memset_tensor   = */ NULL,
+    /* .set_tensor      = */ ggml_backend_kompute_buffer_set_tensor,
+    /* .get_tensor      = */ ggml_backend_kompute_buffer_get_tensor,
+    /* .cpy_tensor      = */ NULL,
+    /* .clear           = */ ggml_backend_kompute_buffer_clear,
+    /* .reset           = */ NULL,
+};
+
+// default buffer type
+
+static const char * ggml_backend_kompute_buffer_type_get_name(ggml_backend_buffer_type_t buft) {
+    auto * ctx = static_cast<ggml_backend_kompute_buffer_type_context *>(buft->context);
+    return ctx->name.c_str();
+}
+
+static ggml_backend_buffer_t ggml_backend_kompute_buffer_type_alloc_buffer(ggml_backend_buffer_type_t buft, size_t size) {
+    ggml_backend_kompute_device_ref(buft);
+    auto * ctx = new ggml_vk_memory(ggml_vk_allocate(size));
+    return ggml_backend_buffer_init(buft, ggml_backend_kompute_buffer_i, ctx, size);
+}
+
+static size_t ggml_backend_kompute_buffer_type_get_alignment(ggml_backend_buffer_type_t buft) {
+    auto * ctx = static_cast<ggml_backend_kompute_buffer_type_context *>(buft->context);
+    return ctx->buffer_alignment;
+}
+
+static size_t ggml_backend_vk_buffer_type_get_max_size(ggml_backend_buffer_type_t buft) {
+    auto * ctx = static_cast<ggml_backend_kompute_buffer_type_context *>(buft->context);
+    return ctx->max_alloc;
+}
+
+static ggml_backend_buffer_type_i ggml_backend_kompute_buffer_type_interface = {
+    /* .get_name         = */ ggml_backend_kompute_buffer_type_get_name,
+    /* .alloc_buffer     = */ ggml_backend_kompute_buffer_type_alloc_buffer,
+    /* .get_alignment    = */ ggml_backend_kompute_buffer_type_get_alignment,
+    /* .get_max_size     = */ ggml_backend_vk_buffer_type_get_max_size,
+    /* .get_alloc_size   = */ NULL, // defaults to ggml_nbytes
+    /* .is_host          = */ NULL,
+};
+
+ggml_backend_buffer_type_t ggml_backend_kompute_buffer_type(int device) {
+    static std::vector<ggml_backend_buffer_type> bufts = []() {
+        std::vector<ggml_backend_buffer_type> vec;
+        auto devices = ggml_vk_available_devices_internal(0);
+        vec.reserve(devices.size());
+
+        for (const auto & dev : devices) {
+            vec.push_back({
+                /* .iface   = */ ggml_backend_kompute_buffer_type_interface,
+                /* .device  = */ nullptr,
+                /* .context = */ new ggml_backend_kompute_buffer_type_context(dev.index, dev.bufferAlignment, dev.maxAlloc)
+            });
+        }
+        return vec;
+    }();
+
+    auto it = std::find_if(bufts.begin(), bufts.end(), [device](const ggml_backend_buffer_type & t) {
+        return device == static_cast<ggml_backend_kompute_buffer_type_context *>(t.context)->device;
+    });
+    return it < bufts.end() ? &*it : nullptr;
+}
+
+// backend
+
+static const char * ggml_backend_kompute_name(ggml_backend_t backend) {
+    auto * ctx = static_cast<ggml_kompute_context *>(backend->context);
+    return ctx->name.c_str();
+}
+
+static void ggml_backend_kompute_free(ggml_backend_t backend) {
+    auto * ctx = static_cast<ggml_kompute_context *>(backend->context);
+
+    assert(ctx == s_kompute_context);
+    s_kompute_context = nullptr;
+    if (ctx != nullptr) {
+        delete ctx;
+    }
+
+    delete backend;
+}
+
+static ggml_backend_buffer_type_t ggml_backend_kompute_get_default_buffer_type(ggml_backend_t backend) {
+    auto * ctx = static_cast<ggml_kompute_context *>(backend->context);
+    return ggml_backend_kompute_buffer_type(ctx->device);
+}
+
+static ggml_status ggml_backend_kompute_graph_compute(ggml_backend_t backend, struct ggml_cgraph * cgraph) {
+    auto * ctx = static_cast<ggml_kompute_context *>(backend->context);
+    ggml_vk_graph_compute(ctx, cgraph);
+    return GGML_STATUS_SUCCESS;
+}
+
+static bool ggml_backend_kompute_supports_op(ggml_backend_t backend, const struct ggml_tensor * op) {
+    GGML_UNUSED(backend);
+    return ggml_vk_supports_op(op);
+}
+
+static bool ggml_backend_kompute_supports_buft(ggml_backend_t backend, ggml_backend_buffer_type_t buft) {
+    GGML_UNUSED(backend);
+    return buft->iface.get_name == ggml_backend_kompute_buffer_type_get_name;
+}
+
+static struct ggml_backend_i kompute_backend_i = {
+    /* .get_name                = */ ggml_backend_kompute_name,
+    /* .free                    = */ ggml_backend_kompute_free,
+    /* .get_default_buffer_type = */ ggml_backend_kompute_get_default_buffer_type,
+    /* .set_tensor_async        = */ NULL,
+    /* .get_tensor_async        = */ NULL,
+    /* .cpy_tensor_async        = */ NULL,
+    /* .synchronize             = */ NULL,
+    /* .graph_plan_create       = */ NULL,
+    /* .graph_plan_free         = */ NULL,
+    /* .graph_plan_update       = */ NULL,
+    /* .graph_plan_compute      = */ NULL,
+    /* .graph_compute           = */ ggml_backend_kompute_graph_compute,
+    /* .supports_op             = */ ggml_backend_kompute_supports_op,
+    /* .supports_buft           = */ ggml_backend_kompute_supports_buft,
+    /* .offload_op              = */ NULL,
+    /* .event_record            = */ NULL,
+    /* .event_wait              = */ NULL,
+};
+
+static ggml_guid_t ggml_backend_kompute_guid() {
+    static ggml_guid guid = { 0x7b, 0x57, 0xdc, 0xaf, 0xde, 0x12, 0x1d, 0x49, 0xfb, 0x35, 0xfa, 0x9b, 0x18, 0x31, 0x1d, 0xca };
+    return &guid;
+}
+
+ggml_backend_t ggml_backend_kompute_init(int device) {
+    GGML_ASSERT(s_kompute_context == nullptr);
+    s_kompute_context = new ggml_kompute_context(device);
+
+    ggml_backend_t kompute_backend = new ggml_backend {
+        /* .guid      = */ ggml_backend_kompute_guid(),
+        /* .interface = */ kompute_backend_i,
+        /* .device    = */ nullptr,
+        /* .context   = */ s_kompute_context,
+    };
+
+    return kompute_backend;
+}
+
+bool ggml_backend_is_kompute(ggml_backend_t backend) {
+    return backend != NULL && ggml_guid_matches(backend->guid, ggml_backend_kompute_guid());
+}
diff --git a/src/whisper_cpp/ggml/src/ggml-metal.m b/src/whisper_cpp/ggml/src/ggml-metal.m
new file mode 100644
index 00000000..7baee417
--- /dev/null
+++ b/src/whisper_cpp/ggml/src/ggml-metal.m
@@ -0,0 +1,3558 @@
+#import "ggml-metal.h"
+
+#import "ggml-impl.h"
+#import "ggml-backend-impl.h"
+
+#import <Foundation/Foundation.h>
+
+#import <Metal/Metal.h>
+
+#undef MIN
+#undef MAX
+#define MIN(a, b) ((a) < (b) ? (a) : (b))
+#define MAX(a, b) ((a) > (b) ? (a) : (b))
+
+// max memory buffers that can be mapped to the device
+#define GGML_METAL_MAX_BUFFERS 64
+
+// max number of MTLCommandBuffer used to submit a graph for processing
+#define GGML_METAL_MAX_COMMAND_BUFFERS 8
+
+#define UNUSED(x) (void)(x)
+
+struct ggml_metal_kernel {
+    id<MTLComputePipelineState> pipeline;
+};
+
+enum ggml_metal_kernel_type {
+    GGML_METAL_KERNEL_TYPE_ADD,
+    GGML_METAL_KERNEL_TYPE_ADD_ROW,
+    GGML_METAL_KERNEL_TYPE_SUB,
+    GGML_METAL_KERNEL_TYPE_SUB_ROW,
+    GGML_METAL_KERNEL_TYPE_MUL,
+    GGML_METAL_KERNEL_TYPE_MUL_ROW,
+    GGML_METAL_KERNEL_TYPE_DIV,
+    GGML_METAL_KERNEL_TYPE_DIV_ROW,
+    GGML_METAL_KERNEL_TYPE_REPEAT_F32,
+    GGML_METAL_KERNEL_TYPE_REPEAT_F16,
+    GGML_METAL_KERNEL_TYPE_REPEAT_I32,
+    GGML_METAL_KERNEL_TYPE_REPEAT_I16,
+    GGML_METAL_KERNEL_TYPE_SCALE,
+    GGML_METAL_KERNEL_TYPE_SCALE_4,
+    GGML_METAL_KERNEL_TYPE_CLAMP,
+    GGML_METAL_KERNEL_TYPE_TANH,
+    GGML_METAL_KERNEL_TYPE_RELU,
+    GGML_METAL_KERNEL_TYPE_SIGMOID,
+    GGML_METAL_KERNEL_TYPE_GELU,
+    GGML_METAL_KERNEL_TYPE_GELU_4,
+    GGML_METAL_KERNEL_TYPE_GELU_QUICK,
+    GGML_METAL_KERNEL_TYPE_GELU_QUICK_4,
+    GGML_METAL_KERNEL_TYPE_SILU,
+    GGML_METAL_KERNEL_TYPE_SILU_4,
+    GGML_METAL_KERNEL_TYPE_SOFT_MAX_F16,
+    GGML_METAL_KERNEL_TYPE_SOFT_MAX_F16_4,
+    GGML_METAL_KERNEL_TYPE_SOFT_MAX_F32,
+    GGML_METAL_KERNEL_TYPE_SOFT_MAX_F32_4,
+    GGML_METAL_KERNEL_TYPE_DIAG_MASK_INF,
+    GGML_METAL_KERNEL_TYPE_DIAG_MASK_INF_8,
+    GGML_METAL_KERNEL_TYPE_GET_ROWS_F32,
+    GGML_METAL_KERNEL_TYPE_GET_ROWS_F16,
+    GGML_METAL_KERNEL_TYPE_GET_ROWS_Q4_0,
+    GGML_METAL_KERNEL_TYPE_GET_ROWS_Q4_1,
+    GGML_METAL_KERNEL_TYPE_GET_ROWS_Q5_0,
+    GGML_METAL_KERNEL_TYPE_GET_ROWS_Q5_1,
+    GGML_METAL_KERNEL_TYPE_GET_ROWS_Q8_0,
+    GGML_METAL_KERNEL_TYPE_GET_ROWS_Q2_K,
+    GGML_METAL_KERNEL_TYPE_GET_ROWS_Q3_K,
+    GGML_METAL_KERNEL_TYPE_GET_ROWS_Q4_K,
+    GGML_METAL_KERNEL_TYPE_GET_ROWS_Q5_K,
+    GGML_METAL_KERNEL_TYPE_GET_ROWS_Q6_K,
+    GGML_METAL_KERNEL_TYPE_GET_ROWS_IQ2_XXS,
+    GGML_METAL_KERNEL_TYPE_GET_ROWS_IQ2_XS,
+    GGML_METAL_KERNEL_TYPE_GET_ROWS_IQ3_XXS,
+    GGML_METAL_KERNEL_TYPE_GET_ROWS_IQ3_S,
+    GGML_METAL_KERNEL_TYPE_GET_ROWS_IQ2_S,
+    GGML_METAL_KERNEL_TYPE_GET_ROWS_IQ1_S,
+    GGML_METAL_KERNEL_TYPE_GET_ROWS_IQ1_M,
+    GGML_METAL_KERNEL_TYPE_GET_ROWS_IQ4_NL,
+    GGML_METAL_KERNEL_TYPE_GET_ROWS_IQ4_XS,
+    GGML_METAL_KERNEL_TYPE_GET_ROWS_I32,
+    GGML_METAL_KERNEL_TYPE_RMS_NORM,
+    GGML_METAL_KERNEL_TYPE_GROUP_NORM,
+    GGML_METAL_KERNEL_TYPE_NORM,
+    GGML_METAL_KERNEL_TYPE_SSM_CONV_F32,
+    GGML_METAL_KERNEL_TYPE_SSM_SCAN_F32,
+    GGML_METAL_KERNEL_TYPE_MUL_MV_F32_F32,
+    GGML_METAL_KERNEL_TYPE_MUL_MV_F16_F16,
+    GGML_METAL_KERNEL_TYPE_MUL_MV_F16_F32,
+    GGML_METAL_KERNEL_TYPE_MUL_MV_F16_F32_1ROW,
+    GGML_METAL_KERNEL_TYPE_MUL_MV_F16_F32_L4,
+    GGML_METAL_KERNEL_TYPE_MUL_MV_Q4_0_F32,
+    GGML_METAL_KERNEL_TYPE_MUL_MV_Q4_1_F32,
+    GGML_METAL_KERNEL_TYPE_MUL_MV_Q5_0_F32,
+    GGML_METAL_KERNEL_TYPE_MUL_MV_Q5_1_F32,
+    GGML_METAL_KERNEL_TYPE_MUL_MV_Q8_0_F32,
+    GGML_METAL_KERNEL_TYPE_MUL_MV_Q2_K_F32,
+    GGML_METAL_KERNEL_TYPE_MUL_MV_Q3_K_F32,
+    GGML_METAL_KERNEL_TYPE_MUL_MV_Q4_K_F32,
+    GGML_METAL_KERNEL_TYPE_MUL_MV_Q5_K_F32,
+    GGML_METAL_KERNEL_TYPE_MUL_MV_Q6_K_F32,
+    GGML_METAL_KERNEL_TYPE_MUL_MV_IQ2_XXS_F32,
+    GGML_METAL_KERNEL_TYPE_MUL_MV_IQ2_XS_F32,
+    GGML_METAL_KERNEL_TYPE_MUL_MV_IQ3_XXS_F32,
+    GGML_METAL_KERNEL_TYPE_MUL_MV_IQ3_S_F32,
+    GGML_METAL_KERNEL_TYPE_MUL_MV_IQ2_S_F32,
+    GGML_METAL_KERNEL_TYPE_MUL_MV_IQ1_S_F32,
+    GGML_METAL_KERNEL_TYPE_MUL_MV_IQ1_M_F32,
+    GGML_METAL_KERNEL_TYPE_MUL_MV_IQ4_NL_F32,
+    GGML_METAL_KERNEL_TYPE_MUL_MV_IQ4_XS_F32,
+    GGML_METAL_KERNEL_TYPE_MUL_MV_ID_F32_F32,
+  //GGML_METAL_KERNEL_TYPE_MUL_MV_ID_F16_F16,
+    GGML_METAL_KERNEL_TYPE_MUL_MV_ID_F16_F32,
+  //GGML_METAL_KERNEL_TYPE_MUL_MV_ID_F16_F32_1ROW,
+  //GGML_METAL_KERNEL_TYPE_MUL_MV_ID_F16_F32_L4,
+    GGML_METAL_KERNEL_TYPE_MUL_MV_ID_Q4_0_F32,
+    GGML_METAL_KERNEL_TYPE_MUL_MV_ID_Q4_1_F32,
+    GGML_METAL_KERNEL_TYPE_MUL_MV_ID_Q5_0_F32,
+    GGML_METAL_KERNEL_TYPE_MUL_MV_ID_Q5_1_F32,
+    GGML_METAL_KERNEL_TYPE_MUL_MV_ID_Q8_0_F32,
+    GGML_METAL_KERNEL_TYPE_MUL_MV_ID_Q2_K_F32,
+    GGML_METAL_KERNEL_TYPE_MUL_MV_ID_Q3_K_F32,
+    GGML_METAL_KERNEL_TYPE_MUL_MV_ID_Q4_K_F32,
+    GGML_METAL_KERNEL_TYPE_MUL_MV_ID_Q5_K_F32,
+    GGML_METAL_KERNEL_TYPE_MUL_MV_ID_Q6_K_F32,
+    GGML_METAL_KERNEL_TYPE_MUL_MV_ID_IQ2_XXS_F32,
+    GGML_METAL_KERNEL_TYPE_MUL_MV_ID_IQ2_XS_F32,
+    GGML_METAL_KERNEL_TYPE_MUL_MV_ID_IQ3_XXS_F32,
+    GGML_METAL_KERNEL_TYPE_MUL_MV_ID_IQ3_S_F32,
+    GGML_METAL_KERNEL_TYPE_MUL_MV_ID_IQ2_S_F32,
+    GGML_METAL_KERNEL_TYPE_MUL_MV_ID_IQ1_S_F32,
+    GGML_METAL_KERNEL_TYPE_MUL_MV_ID_IQ1_M_F32,
+    GGML_METAL_KERNEL_TYPE_MUL_MV_ID_IQ4_NL_F32,
+    GGML_METAL_KERNEL_TYPE_MUL_MV_ID_IQ4_XS_F32,
+    GGML_METAL_KERNEL_TYPE_MUL_MM_F32_F32,
+    GGML_METAL_KERNEL_TYPE_MUL_MM_F16_F32,
+    GGML_METAL_KERNEL_TYPE_MUL_MM_Q4_0_F32,
+    GGML_METAL_KERNEL_TYPE_MUL_MM_Q4_1_F32,
+    GGML_METAL_KERNEL_TYPE_MUL_MM_Q5_0_F32,
+    GGML_METAL_KERNEL_TYPE_MUL_MM_Q5_1_F32,
+    GGML_METAL_KERNEL_TYPE_MUL_MM_Q8_0_F32,
+    GGML_METAL_KERNEL_TYPE_MUL_MM_Q2_K_F32,
+    GGML_METAL_KERNEL_TYPE_MUL_MM_Q3_K_F32,
+    GGML_METAL_KERNEL_TYPE_MUL_MM_Q4_K_F32,
+    GGML_METAL_KERNEL_TYPE_MUL_MM_Q5_K_F32,
+    GGML_METAL_KERNEL_TYPE_MUL_MM_Q6_K_F32,
+    GGML_METAL_KERNEL_TYPE_MUL_MM_IQ2_XXS_F32,
+    GGML_METAL_KERNEL_TYPE_MUL_MM_IQ2_XS_F32,
+    GGML_METAL_KERNEL_TYPE_MUL_MM_IQ3_XXS_F32,
+    GGML_METAL_KERNEL_TYPE_MUL_MM_IQ3_S_F32,
+    GGML_METAL_KERNEL_TYPE_MUL_MM_IQ2_S_F32,
+    GGML_METAL_KERNEL_TYPE_MUL_MM_IQ1_S_F32,
+    GGML_METAL_KERNEL_TYPE_MUL_MM_IQ1_M_F32,
+    GGML_METAL_KERNEL_TYPE_MUL_MM_IQ4_NL_F32,
+    GGML_METAL_KERNEL_TYPE_MUL_MM_IQ4_XS_F32,
+    GGML_METAL_KERNEL_TYPE_MUL_MM_ID_F32_F32,
+    GGML_METAL_KERNEL_TYPE_MUL_MM_ID_F16_F32,
+    GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q4_0_F32,
+    GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q4_1_F32,
+    GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q5_0_F32,
+    GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q5_1_F32,
+    GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q8_0_F32,
+    GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q2_K_F32,
+    GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q3_K_F32,
+    GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q4_K_F32,
+    GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q5_K_F32,
+    GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q6_K_F32,
+    GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ2_XXS_F32,
+    GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ2_XS_F32,
+    GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ3_XXS_F32,
+    GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ3_S_F32,
+    GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ2_S_F32,
+    GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ1_S_F32,
+    GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ1_M_F32,
+    GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ4_NL_F32,
+    GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ4_XS_F32,
+    GGML_METAL_KERNEL_TYPE_ROPE_NORM_F32,
+    GGML_METAL_KERNEL_TYPE_ROPE_NORM_F16,
+    GGML_METAL_KERNEL_TYPE_ROPE_NEOX_F32,
+    GGML_METAL_KERNEL_TYPE_ROPE_NEOX_F16,
+    GGML_METAL_KERNEL_TYPE_IM2COL_F16,
+    GGML_METAL_KERNEL_TYPE_IM2COL_F32,
+    GGML_METAL_KERNEL_TYPE_UPSCALE_F32,
+    GGML_METAL_KERNEL_TYPE_PAD_F32,
+    GGML_METAL_KERNEL_TYPE_ARANGE_F32,
+    GGML_METAL_KERNEL_TYPE_TIMESTEP_EMBEDDING_F32,
+    GGML_METAL_KERNEL_TYPE_ARGSORT_F32_I32_ASC,
+    GGML_METAL_KERNEL_TYPE_ARGSORT_F32_I32_DESC,
+    GGML_METAL_KERNEL_TYPE_LEAKY_RELU_F32,
+    GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_F16_H64,
+    GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_F16_H80,
+    GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_F16_H96,
+    GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_F16_H112,
+    GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_F16_H128,
+  //GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_F16_H256,     // https://github.com/ggerganov/llama.cpp/issues/7261
+    GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_F16_H128,
+  //GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_F16_H256, // https://github.com/ggerganov/llama.cpp/issues/7261
+    GGML_METAL_KERNEL_TYPE_CPY_F32_F32,
+    GGML_METAL_KERNEL_TYPE_CPY_F32_F16,
+    GGML_METAL_KERNEL_TYPE_CPY_F16_F16,
+    GGML_METAL_KERNEL_TYPE_CPY_F16_F32,
+    GGML_METAL_KERNEL_TYPE_CPY_F32_Q8_0,
+    GGML_METAL_KERNEL_TYPE_CPY_F32_Q4_0,
+    GGML_METAL_KERNEL_TYPE_CPY_F32_Q4_1,
+    GGML_METAL_KERNEL_TYPE_CPY_F32_Q5_0,
+    GGML_METAL_KERNEL_TYPE_CPY_F32_Q5_1,
+    GGML_METAL_KERNEL_TYPE_CPY_F32_IQ4_NL,
+    GGML_METAL_KERNEL_TYPE_CONCAT,
+    GGML_METAL_KERNEL_TYPE_SQR,
+    GGML_METAL_KERNEL_TYPE_SQRT,
+    GGML_METAL_KERNEL_TYPE_SIN,
+    GGML_METAL_KERNEL_TYPE_COS,
+    GGML_METAL_KERNEL_TYPE_SUM_ROWS,
+
+    GGML_METAL_KERNEL_TYPE_COUNT
+};
+
+struct ggml_backend_metal_context {
+    id<MTLDevice>       device;
+    id<MTLCommandQueue> queue;
+
+    dispatch_queue_t d_queue;
+
+    struct ggml_metal_kernel kernels[GGML_METAL_KERNEL_TYPE_COUNT];
+
+    bool support_simdgroup_reduction;
+    bool support_simdgroup_mm;
+
+    // capture state
+    bool capture_next_compute;
+    bool capture_started;
+
+    id<MTLCaptureScope> capture_scope;
+
+    // command buffer state
+    int n_cb;           // number of extra threads used to submit the command buffers
+    int n_nodes_0;      // number of nodes submitted by the main thread
+    int n_nodes_1;      // remaining number of nodes submitted by the n_cb threads
+    int n_nodes_per_cb;
+
+    struct ggml_cgraph * gf;
+
+    // the callback given to the thread pool
+    // TODO: ideally, this should be created once, utilizing the command buffer state above
+    //       for some reason, doing it like this leads to a crash
+    void (^encode_async)(size_t ith);
+
+    // n_cb command buffers + 1 used by the main thread
+    id<MTLCommandBuffer> command_buffers[GGML_METAL_MAX_COMMAND_BUFFERS + 1];
+
+    // abort ggml_metal_graph_compute if callback returns true
+    ggml_abort_callback abort_callback;
+    void *              abort_callback_data;
+};
+
+// MSL code
+// TODO: move the contents here when ready
+//       for now it is easier to work in a separate file
+// static NSString * const msl_library_source = @"see metal.metal";
+
+// Here to assist with NSBundle Path Hack
+@interface GGMLMetalClass : NSObject
+@end
+@implementation GGMLMetalClass
+@end
+
+static void * ggml_metal_host_malloc(size_t n) {
+    void * data = NULL;
+
+#if TARGET_OS_OSX
+    kern_return_t err = vm_allocate((vm_map_t) mach_task_self(), (void *) &data, n, VM_FLAGS_ANYWHERE);
+    if (err != KERN_SUCCESS) {
+        GGML_LOG_ERROR("%s: error: vm_allocate failed\n", __func__);
+        return NULL;
+    }
+#else
+    const int result = posix_memalign((void **) &data, sysconf(_SC_PAGESIZE), n);
+    if (result != 0) {
+        GGML_LOG_ERROR("%s: error: posix_memalign failed\n", __func__);
+        return NULL;
+    }
+#endif
+
+    return data;
+}
+
+static struct ggml_backend_metal_context * ggml_metal_init(void) {
+    GGML_LOG_INFO("%s: allocating\n", __func__);
+
+#if TARGET_OS_OSX && !GGML_METAL_NDEBUG
+    // Show all the Metal device instances in the system
+    NSArray * devices = MTLCopyAllDevices();
+    for (id<MTLDevice> device in devices) {
+        GGML_LOG_INFO("%s: found device: %s\n", __func__, [[device name] UTF8String]);
+    }
+    [devices release]; // since it was created by a *Copy* C method
+#endif
+
+    // Pick and show default Metal device
+    id<MTLDevice> device = MTLCreateSystemDefaultDevice();
+    GGML_LOG_INFO("%s: picking default device: %s\n", __func__, [[device name] UTF8String]);
+
+    // Configure context
+    struct ggml_backend_metal_context * ctx = calloc(1, sizeof(struct ggml_backend_metal_context));
+    ctx->device = device;
+    ctx->queue  = [ctx->device newCommandQueue];
+    ctx->d_queue = dispatch_queue_create("ggml-metal", DISPATCH_QUEUE_CONCURRENT);
+
+    id<MTLLibrary> metal_library;
+
+    // load library
+    //
+    // - first check if the library is embedded
+    // - then check if the library is in the bundle
+    // - if not found, load the source and compile it
+    // - if that fails, return NULL
+    {
+        NSBundle * bundle = nil;
+#ifdef SWIFT_PACKAGE
+        bundle = SWIFTPM_MODULE_BUNDLE;
+#else
+        bundle = [NSBundle bundleForClass:[GGMLMetalClass class]];
+#endif
+
+        NSError * error = nil;
+
+#if GGML_METAL_EMBED_LIBRARY
+        const bool try_metallib = false;
+#else
+        const bool try_metallib = true;
+#endif
+
+        NSString * path_lib = [bundle pathForResource:@"default" ofType:@"metallib"];
+        if (try_metallib && path_lib != nil) {
+            // pre-compiled library found
+            NSURL * libURL = [NSURL fileURLWithPath:path_lib];
+            GGML_LOG_INFO("%s: loading '%s'\n", __func__, [path_lib UTF8String]);
+
+            metal_library = [ctx->device newLibraryWithURL:libURL error:&error];
+            if (error) {
+                GGML_LOG_ERROR("%s: error: %s\n", __func__, [[error description] UTF8String]);
+                return NULL;
+            }
+        } else {
+#if GGML_METAL_EMBED_LIBRARY
+            GGML_LOG_INFO("%s: using embedded metal library\n", __func__);
+
+            extern const char ggml_metallib_start[];
+            extern const char ggml_metallib_end[];
+
+            NSString * src = [[NSString alloc] initWithBytes:ggml_metallib_start length:(ggml_metallib_end-ggml_metallib_start) encoding:NSUTF8StringEncoding];
+#else
+            GGML_LOG_INFO("%s: default.metallib not found, loading from source\n", __func__);
+
+            NSString * path_source;
+            NSString * path_resource = [[NSProcessInfo processInfo].environment objectForKey:@"GGML_METAL_PATH_RESOURCES"];
+
+            GGML_LOG_INFO("%s: GGML_METAL_PATH_RESOURCES = %s\n", __func__, path_resource ? [path_resource UTF8String] : "nil");
+
+            if (path_resource) {
+                path_source = [path_resource stringByAppendingPathComponent:@"ggml-metal.metal"];
+            } else {
+                path_source = [bundle pathForResource:@"ggml-metal" ofType:@"metal"];
+            }
+
+            if (path_source == nil) {
+                GGML_LOG_WARN("%s: error: could not use bundle path to find ggml-metal.metal, falling back to trying cwd\n", __func__);
+                path_source = @"ggml-metal.metal";
+            }
+
+            GGML_LOG_INFO("%s: loading '%s'\n", __func__, [path_source UTF8String]);
+
+            NSString * src = [NSString stringWithContentsOfFile:path_source encoding:NSUTF8StringEncoding error:&error];
+            if (error) {
+                GGML_LOG_ERROR("%s: error: %s\n", __func__, [[error description] UTF8String]);
+                return NULL;
+            }
+#endif // GGML_METAL_EMBED_LIBRARY
+
+            @autoreleasepool {
+                // dictionary of preprocessor macros
+                NSMutableDictionary * prep = [NSMutableDictionary dictionary];
+
+                MTLCompileOptions* options = [MTLCompileOptions new];
+                options.preprocessorMacros = prep;
+
+                //[options setFastMathEnabled:false];
+
+                metal_library = [ctx->device newLibraryWithSource:src options:options error:&error];
+                if (error) {
+                    GGML_LOG_ERROR("%s: error: %s\n", __func__, [[error description] UTF8String]);
+                    return NULL;
+                }
+            }
+        }
+    }
+
+    // print MTL GPU family:
+    GGML_LOG_INFO("%s: GPU name:   %s\n", __func__, [[ctx->device name] UTF8String]);
+
+    const NSInteger MTLGPUFamilyMetal3 = 5001;
+
+    // determine max supported GPU family
+    // https://developer.apple.com/metal/Metal-Shading-Language-Specification.pdf
+    // https://developer.apple.com/metal/Metal-Feature-Set-Tables.pdf
+    {
+        for (int i = MTLGPUFamilyApple1 + 20; i >= MTLGPUFamilyApple1; --i) {
+            if ([ctx->device supportsFamily:i]) {
+                GGML_LOG_INFO("%s: GPU family: MTLGPUFamilyApple%d  (%d)\n", __func__, i - (int) MTLGPUFamilyApple1 + 1, i);
+                break;
+            }
+        }
+
+        for (int i = MTLGPUFamilyCommon1 + 5; i >= MTLGPUFamilyCommon1; --i) {
+            if ([ctx->device supportsFamily:i]) {
+                GGML_LOG_INFO("%s: GPU family: MTLGPUFamilyCommon%d (%d)\n", __func__, i - (int) MTLGPUFamilyCommon1 + 1, i);
+                break;
+            }
+        }
+
+        for (int i = MTLGPUFamilyMetal3 + 5; i >= MTLGPUFamilyMetal3; --i) {
+            if ([ctx->device supportsFamily:i]) {
+                GGML_LOG_INFO("%s: GPU family: MTLGPUFamilyMetal%d  (%d)\n", __func__, i - (int) MTLGPUFamilyMetal3 + 3, i);
+                break;
+            }
+        }
+    }
+
+    ctx->support_simdgroup_reduction  = [ctx->device supportsFamily:MTLGPUFamilyApple7];
+    ctx->support_simdgroup_reduction |= [ctx->device supportsFamily:MTLGPUFamilyMetal3];
+
+    ctx->support_simdgroup_mm = [ctx->device supportsFamily:MTLGPUFamilyApple7];
+
+    GGML_LOG_INFO("%s: simdgroup reduction support   = %s\n",       __func__, ctx->support_simdgroup_reduction ? "true" : "false");
+    GGML_LOG_INFO("%s: simdgroup matrix mul. support = %s\n",       __func__, ctx->support_simdgroup_mm ? "true" : "false");
+    GGML_LOG_INFO("%s: hasUnifiedMemory              = %s\n",       __func__, ctx->device.hasUnifiedMemory ? "true" : "false");
+
+    ctx->capture_next_compute = false;
+    ctx->capture_started = false;
+    ctx->capture_scope = nil;
+
+    ctx->gf = nil;
+    ctx->encode_async = nil;
+    for (int i = 0; i < GGML_METAL_MAX_COMMAND_BUFFERS; ++i) {
+        ctx->command_buffers[i] = nil;
+    }
+
+#if TARGET_OS_OSX || (TARGET_OS_IOS && __clang_major__ >= 15)
+    if (@available(macOS 10.12, iOS 16.0, *)) {
+        GGML_LOG_INFO("%s: recommendedMaxWorkingSetSize  = %8.2f MB\n", __func__, ctx->device.recommendedMaxWorkingSetSize / 1e6);
+    }
+#elif TARGET_OS_OSX
+    if (ctx->device.maxTransferRate != 0) {
+        GGML_LOG_INFO("%s: maxTransferRate               = %8.2f MB/s\n", __func__, ctx->device.maxTransferRate / 1e6);
+    } else {
+        GGML_LOG_INFO("%s: maxTransferRate               = built-in GPU\n", __func__);
+    }
+#endif
+
+    // load kernels
+    {
+        NSError * error = nil;
+
+        for (int i = 0; i < GGML_METAL_KERNEL_TYPE_COUNT; ++i) {
+            ctx->kernels[i].pipeline = nil;
+        }
+
+        /*
+            GGML_LOG_INFO("%s: loaded %-40s %16p | th_max = %4d | th_width = %4d\n", __func__, "kernel_"#name, (void *) kernel->pipeline, \
+                    (int) kernel->pipeline.maxTotalThreadsPerThreadgroup, \
+                    (int) kernel->pipeline.threadExecutionWidth); \
+        */
+#define GGML_METAL_ADD_KERNEL(e, name, supported) \
+        if (supported) { \
+            struct ggml_metal_kernel * kernel = &ctx->kernels[e]; \
+            id<MTLFunction> metal_function = [metal_library newFunctionWithName:@"kernel_"#name]; \
+            kernel->pipeline = [ctx->device newComputePipelineStateWithFunction:metal_function error:&error]; \
+            [metal_function release]; \
+            if (error) { \
+                GGML_LOG_ERROR("%s: error: load pipeline error: %s\n", __func__, [[error description] UTF8String]); \
+                [metal_library release]; \
+                return NULL; \
+            } \
+        } else { \
+            GGML_LOG_WARN("%s: skipping %-40s (not supported)\n", __func__, "kernel_"#name); \
+        }
+
+        // simd_sum and simd_max requires MTLGPUFamilyApple7
+
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_ADD,                           add,                            true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_ADD_ROW,                       add_row,                        true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_SUB,                           sub,                            true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_SUB_ROW,                       sub_row,                        true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL,                           mul,                            true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_ROW,                       mul_row,                        true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_DIV,                           div,                            true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_DIV_ROW,                       div_row,                        true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_REPEAT_F32,                    repeat_f32,                     true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_REPEAT_F16,                    repeat_f16,                     true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_REPEAT_I32,                    repeat_i32,                     true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_REPEAT_I16,                    repeat_i16,                     true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_SCALE,                         scale,                          true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_SCALE_4,                       scale_4,                        true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_CLAMP,                         clamp,                          true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_TANH,                          tanh,                           true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_RELU,                          relu,                           true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_SIGMOID,                       sigmoid,                        true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GELU,                          gelu,                           true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GELU_4,                        gelu_4,                         true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GELU_QUICK,                    gelu_quick,                     true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GELU_QUICK_4,                  gelu_quick_4,                   true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_SILU,                          silu,                           true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_SILU_4,                        silu_4,                         true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_SOFT_MAX_F16,                  soft_max_f16,                   ctx->support_simdgroup_reduction);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_SOFT_MAX_F16_4,                soft_max_f16_4,                 ctx->support_simdgroup_reduction);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_SOFT_MAX_F32,                  soft_max_f32,                   ctx->support_simdgroup_reduction);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_SOFT_MAX_F32_4,                soft_max_f32_4,                 ctx->support_simdgroup_reduction);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_DIAG_MASK_INF,                 diag_mask_inf,                  true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_DIAG_MASK_INF_8,               diag_mask_inf_8,                true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_F32,                  get_rows_f32,                   true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_F16,                  get_rows_f16,                   true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_Q4_0,                 get_rows_q4_0,                  true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_Q4_1,                 get_rows_q4_1,                  true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_Q5_0,                 get_rows_q5_0,                  true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_Q5_1,                 get_rows_q5_1,                  true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_Q8_0,                 get_rows_q8_0,                  true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_Q2_K,                 get_rows_q2_K,                  true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_Q3_K,                 get_rows_q3_K,                  true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_Q4_K,                 get_rows_q4_K,                  true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_Q5_K,                 get_rows_q5_K,                  true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_Q6_K,                 get_rows_q6_K,                  true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_IQ2_XXS,              get_rows_iq2_xxs,               true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_IQ2_XS,               get_rows_iq2_xs,                true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_IQ3_XXS,              get_rows_iq3_xxs,               true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_IQ3_S,                get_rows_iq3_s,                 true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_IQ2_S,                get_rows_iq2_s,                 true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_IQ1_S,                get_rows_iq1_s,                 true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_IQ1_M,                get_rows_iq1_m,                 true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_IQ4_NL,               get_rows_iq4_nl,                true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_IQ4_XS,               get_rows_iq4_xs,                true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_I32,                  get_rows_i32,                   true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_RMS_NORM,                      rms_norm,                       ctx->support_simdgroup_reduction);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GROUP_NORM,                    group_norm,                     ctx->support_simdgroup_reduction);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_NORM,                          norm,                           true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_SSM_CONV_F32,                  ssm_conv_f32,                   true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_SSM_SCAN_F32,                  ssm_scan_f32,                   true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_F32_F32,                mul_mv_f32_f32,                 ctx->support_simdgroup_reduction);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_F16_F16,                mul_mv_f16_f16,                 ctx->support_simdgroup_reduction);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_F16_F32,                mul_mv_f16_f32,                 ctx->support_simdgroup_reduction);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_F16_F32_1ROW,           mul_mv_f16_f32_1row,            ctx->support_simdgroup_reduction);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_F16_F32_L4,             mul_mv_f16_f32_l4,              ctx->support_simdgroup_reduction);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_Q4_0_F32,               mul_mv_q4_0_f32,                ctx->support_simdgroup_reduction);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_Q4_1_F32,               mul_mv_q4_1_f32,                ctx->support_simdgroup_reduction);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_Q5_0_F32,               mul_mv_q5_0_f32,                ctx->support_simdgroup_reduction);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_Q5_1_F32,               mul_mv_q5_1_f32,                ctx->support_simdgroup_reduction);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_Q8_0_F32,               mul_mv_q8_0_f32,                ctx->support_simdgroup_reduction);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_Q2_K_F32,               mul_mv_q2_K_f32,                ctx->support_simdgroup_reduction);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_Q3_K_F32,               mul_mv_q3_K_f32,                ctx->support_simdgroup_reduction);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_Q4_K_F32,               mul_mv_q4_K_f32,                ctx->support_simdgroup_reduction);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_Q5_K_F32,               mul_mv_q5_K_f32,                ctx->support_simdgroup_reduction);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_Q6_K_F32,               mul_mv_q6_K_f32,                ctx->support_simdgroup_reduction);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_IQ2_XXS_F32,            mul_mv_iq2_xxs_f32,             ctx->support_simdgroup_reduction);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_IQ2_XS_F32,             mul_mv_iq2_xs_f32,              ctx->support_simdgroup_reduction);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_IQ3_XXS_F32,            mul_mv_iq3_xxs_f32,             ctx->support_simdgroup_reduction);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_IQ3_S_F32,              mul_mv_iq3_s_f32,               ctx->support_simdgroup_reduction);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_IQ2_S_F32,              mul_mv_iq2_s_f32,               ctx->support_simdgroup_reduction);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_IQ1_S_F32,              mul_mv_iq1_s_f32,               ctx->support_simdgroup_reduction);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_IQ1_M_F32,              mul_mv_iq1_m_f32,               ctx->support_simdgroup_reduction);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_IQ4_NL_F32,             mul_mv_iq4_nl_f32,              ctx->support_simdgroup_reduction);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_IQ4_XS_F32,             mul_mv_iq4_xs_f32,              ctx->support_simdgroup_reduction);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_F32_F32,             mul_mv_id_f32_f32,              ctx->support_simdgroup_reduction);
+      //GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_F16_F16,             mul_mv_id_f16_f16,              ctx->support_simdgroup_reduction);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_F16_F32,             mul_mv_id_f16_f32,              ctx->support_simdgroup_reduction);
+      //GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_F16_F32_1ROW,        mul_mv_id_f16_f32_1row,         ctx->support_simdgroup_reduction);
+      //GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_F16_F32_L4,          mul_mv_id_f16_f32_l4,           ctx->support_simdgroup_reduction);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_Q4_0_F32,            mul_mv_id_q4_0_f32,             ctx->support_simdgroup_reduction);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_Q4_1_F32,            mul_mv_id_q4_1_f32,             ctx->support_simdgroup_reduction);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_Q5_0_F32,            mul_mv_id_q5_0_f32,             ctx->support_simdgroup_reduction);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_Q5_1_F32,            mul_mv_id_q5_1_f32,             ctx->support_simdgroup_reduction);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_Q8_0_F32,            mul_mv_id_q8_0_f32,             ctx->support_simdgroup_reduction);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_Q2_K_F32,            mul_mv_id_q2_K_f32,             ctx->support_simdgroup_reduction);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_Q3_K_F32,            mul_mv_id_q3_K_f32,             ctx->support_simdgroup_reduction);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_Q4_K_F32,            mul_mv_id_q4_K_f32,             ctx->support_simdgroup_reduction);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_Q5_K_F32,            mul_mv_id_q5_K_f32,             ctx->support_simdgroup_reduction);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_Q6_K_F32,            mul_mv_id_q6_K_f32,             ctx->support_simdgroup_reduction);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_IQ2_XXS_F32,         mul_mv_id_iq2_xxs_f32,          ctx->support_simdgroup_reduction);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_IQ2_XS_F32,          mul_mv_id_iq2_xs_f32,           ctx->support_simdgroup_reduction);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_IQ3_XXS_F32,         mul_mv_id_iq3_xxs_f32,          ctx->support_simdgroup_reduction);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_IQ3_S_F32,           mul_mv_id_iq3_s_f32,            ctx->support_simdgroup_reduction);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_IQ2_S_F32,           mul_mv_id_iq2_s_f32,            ctx->support_simdgroup_reduction);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_IQ1_S_F32,           mul_mv_id_iq1_s_f32,            ctx->support_simdgroup_reduction);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_IQ1_M_F32,           mul_mv_id_iq1_m_f32,            ctx->support_simdgroup_reduction);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_IQ4_NL_F32,          mul_mv_id_iq4_nl_f32,           ctx->support_simdgroup_reduction);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_IQ4_XS_F32,          mul_mv_id_iq4_xs_f32,           ctx->support_simdgroup_reduction);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_F32_F32,                mul_mm_f32_f32,                 ctx->support_simdgroup_mm);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_F16_F32,                mul_mm_f16_f32,                 ctx->support_simdgroup_mm);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_Q4_0_F32,               mul_mm_q4_0_f32,                ctx->support_simdgroup_mm);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_Q4_1_F32,               mul_mm_q4_1_f32,                ctx->support_simdgroup_mm);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_Q5_0_F32,               mul_mm_q5_0_f32,                ctx->support_simdgroup_mm);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_Q5_1_F32,               mul_mm_q5_1_f32,                ctx->support_simdgroup_mm);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_Q8_0_F32,               mul_mm_q8_0_f32,                ctx->support_simdgroup_mm);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_Q2_K_F32,               mul_mm_q2_K_f32,                ctx->support_simdgroup_mm);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_Q3_K_F32,               mul_mm_q3_K_f32,                ctx->support_simdgroup_mm);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_Q4_K_F32,               mul_mm_q4_K_f32,                ctx->support_simdgroup_mm);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_Q5_K_F32,               mul_mm_q5_K_f32,                ctx->support_simdgroup_mm);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_Q6_K_F32,               mul_mm_q6_K_f32,                ctx->support_simdgroup_mm);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_IQ2_XXS_F32,            mul_mm_iq2_xxs_f32,             ctx->support_simdgroup_mm);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_IQ2_XS_F32,             mul_mm_iq2_xs_f32,              ctx->support_simdgroup_mm);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_IQ3_XXS_F32,            mul_mm_iq3_xxs_f32,             ctx->support_simdgroup_mm);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_IQ3_S_F32,              mul_mm_iq3_s_f32,               ctx->support_simdgroup_mm);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_IQ2_S_F32,              mul_mm_iq2_s_f32,               ctx->support_simdgroup_mm);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_IQ1_S_F32,              mul_mm_iq1_s_f32,               ctx->support_simdgroup_mm);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_IQ1_M_F32,              mul_mm_iq1_m_f32,               ctx->support_simdgroup_mm);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_IQ4_NL_F32,             mul_mm_iq4_nl_f32,              ctx->support_simdgroup_mm);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_IQ4_XS_F32,             mul_mm_iq4_xs_f32,              ctx->support_simdgroup_mm);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_F32_F32,             mul_mm_id_f32_f32,              ctx->support_simdgroup_mm);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_F16_F32,             mul_mm_id_f16_f32,              ctx->support_simdgroup_mm);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q4_0_F32,            mul_mm_id_q4_0_f32,             ctx->support_simdgroup_mm);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q4_1_F32,            mul_mm_id_q4_1_f32,             ctx->support_simdgroup_mm);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q5_0_F32,            mul_mm_id_q5_0_f32,             ctx->support_simdgroup_mm);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q5_1_F32,            mul_mm_id_q5_1_f32,             ctx->support_simdgroup_mm);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q8_0_F32,            mul_mm_id_q8_0_f32,             ctx->support_simdgroup_mm);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q2_K_F32,            mul_mm_id_q2_K_f32,             ctx->support_simdgroup_mm);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q3_K_F32,            mul_mm_id_q3_K_f32,             ctx->support_simdgroup_mm);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q4_K_F32,            mul_mm_id_q4_K_f32,             ctx->support_simdgroup_mm);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q5_K_F32,            mul_mm_id_q5_K_f32,             ctx->support_simdgroup_mm);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q6_K_F32,            mul_mm_id_q6_K_f32,             ctx->support_simdgroup_mm);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ2_XXS_F32,         mul_mm_id_iq2_xxs_f32,          ctx->support_simdgroup_mm);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ2_XS_F32,          mul_mm_id_iq2_xs_f32,           ctx->support_simdgroup_mm);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ3_XXS_F32,         mul_mm_id_iq3_xxs_f32,          ctx->support_simdgroup_mm);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ3_S_F32,           mul_mm_id_iq3_s_f32,            ctx->support_simdgroup_mm);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ2_S_F32,           mul_mm_id_iq2_s_f32,            ctx->support_simdgroup_mm);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ1_S_F32,           mul_mm_id_iq1_s_f32,            ctx->support_simdgroup_mm);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ1_M_F32,           mul_mm_id_iq1_m_f32,            ctx->support_simdgroup_mm);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ4_NL_F32,          mul_mm_id_iq4_nl_f32,           ctx->support_simdgroup_mm);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ4_XS_F32,          mul_mm_id_iq4_xs_f32,           ctx->support_simdgroup_mm);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_ROPE_NORM_F32,                 rope_norm_f32,                  true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_ROPE_NORM_F16,                 rope_norm_f16,                  true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_ROPE_NEOX_F32,                 rope_neox_f32,                  true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_ROPE_NEOX_F16,                 rope_neox_f16,                  true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_IM2COL_F16,                    im2col_f16,                     true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_IM2COL_F32,                    im2col_f32,                     true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_UPSCALE_F32,                   upscale_f32,                    true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_PAD_F32,                       pad_f32,                        true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_TIMESTEP_EMBEDDING_F32,        timestep_embedding_f32,         true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_ARANGE_F32,                    arange_f32,                     true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_ARGSORT_F32_I32_ASC,           argsort_f32_i32_asc,            true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_ARGSORT_F32_I32_DESC,          argsort_f32_i32_desc,           true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_LEAKY_RELU_F32,                leaky_relu_f32,                 true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_F16_H64,        flash_attn_ext_f16_h64,         ctx->support_simdgroup_mm);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_F16_H80,        flash_attn_ext_f16_h80,         ctx->support_simdgroup_mm);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_F16_H96,        flash_attn_ext_f16_h96,         ctx->support_simdgroup_mm);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_F16_H112,       flash_attn_ext_f16_h112,        ctx->support_simdgroup_mm);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_F16_H128,       flash_attn_ext_f16_h128,        ctx->support_simdgroup_mm);
+      //GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_F16_H256,       flash_attn_ext_f16_h256,        ctx->support_simdgroup_mm);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_F16_H128,   flash_attn_ext_vec_f16_h128,    ctx->support_simdgroup_reduction);
+      //GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_F16_H256,   flash_attn_ext_vec_f16_h256,    ctx->support_simdgroup_reduction);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_CPY_F32_F16,                   cpy_f32_f16,                    true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_CPY_F32_F32,                   cpy_f32_f32,                    true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_CPY_F16_F16,                   cpy_f16_f16,                    true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_CPY_F16_F32,                   cpy_f16_f32,                    true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_CPY_F32_Q8_0,                  cpy_f32_q8_0,                   true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_CPY_F32_Q4_0,                  cpy_f32_q4_0,                   true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_CPY_F32_Q4_1,                  cpy_f32_q4_1,                   true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_CPY_F32_Q5_0,                  cpy_f32_q5_0,                   true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_CPY_F32_Q5_1,                  cpy_f32_q5_1,                   true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_CPY_F32_IQ4_NL,                cpy_f32_iq4_nl,                 true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_CONCAT,                        concat,                         true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_SQR,                           sqr,                            true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_SQRT,                          sqrt,                           true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_SIN,                           sin,                            true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_COS,                           cos,                            true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_SUM_ROWS,                      sum_rows,                       true);
+    }
+
+    [metal_library release];
+
+    return ctx;
+}
+
+static void ggml_metal_free(struct ggml_backend_metal_context * ctx) {
+    GGML_LOG_INFO("%s: deallocating\n", __func__);
+
+    for (int i = 0; i < GGML_METAL_KERNEL_TYPE_COUNT; ++i) {
+        [ctx->kernels[i].pipeline release];
+    }
+
+    [ctx->queue release];
+    [ctx->device release];
+
+    dispatch_release(ctx->d_queue);
+
+    free(ctx);
+}
+
+// temporarily defined here for compatibility between ggml-backend and the old API
+
+struct ggml_backend_metal_buffer {
+    void   * data;
+    size_t   size;
+
+    id<MTLBuffer> metal;
+};
+
+struct ggml_backend_metal_buffer_context {
+    void * all_data;
+    size_t all_size;
+    bool owned;
+
+    // multiple buffers are used only to avoid the maximum buffer size limitation when using mmap
+    int n_buffers;
+    struct ggml_backend_metal_buffer buffers[GGML_METAL_MAX_BUFFERS];
+};
+
+// finds the Metal buffer that contains the tensor data on the GPU device
+// the assumption is that there is 1-to-1 mapping between the host and device memory buffers, so we can find the
+// Metal buffer based on the host memory pointer
+//
+static id<MTLBuffer> ggml_metal_get_buffer(struct ggml_tensor * t, size_t * offs) {
+    //GGML_LOG_INFO("%s: data tensor '%16s', offs_data = %8ld, offs_eval = %8ld, offs_cach = %8ld\n", __func__, t->name, offs_data, offs_eval, offs_cach);
+
+    const int64_t tsize = ggml_nbytes(t);
+
+    ggml_backend_buffer_t buffer = t->view_src ? t->view_src->buffer : t->buffer;
+
+    struct ggml_backend_metal_buffer_context * buf_ctx = (struct ggml_backend_metal_buffer_context *) buffer->context;
+
+    // find the view that contains the tensor fully
+    for (int i = 0; i < buf_ctx->n_buffers; ++i) {
+        const int64_t ioffs = (int64_t) t->data - (int64_t) buf_ctx->buffers[i].data;
+
+        //GGML_LOG_INFO("ioffs = %10ld, tsize = %10ld, sum = %10ld, buf_ctx->buffers[%d].size = %10ld\n", ioffs, tsize, ioffs + tsize, i, buf_ctx->buffers[i].size);
+        if (ioffs >= 0 && ioffs + tsize <= (int64_t) buf_ctx->buffers[i].size) {
+            *offs = (size_t) ioffs;
+
+            //GGML_LOG_INFO("%s: tensor '%16s', offs = %8ld\n", __func__, t->name, *offs);
+
+            return buf_ctx->buffers[i].metal;
+        }
+    }
+
+    GGML_LOG_ERROR("%s: error: tensor '%s' buffer is nil\n", __func__, t->name);
+
+    return nil;
+}
+
+static bool ggml_metal_supports_op(const struct ggml_backend_metal_context * ctx, const struct ggml_tensor * op) {
+    for (size_t i = 0, n = 3; i < n; ++i) {
+        if (op->src[i] != NULL && op->src[i]->type == GGML_TYPE_BF16) {
+            return false;
+        }
+    }
+
+    switch (op->op) {
+        case GGML_OP_UNARY:
+            switch (ggml_get_unary_op(op)) {
+                case GGML_UNARY_OP_TANH:
+                case GGML_UNARY_OP_RELU:
+                case GGML_UNARY_OP_SIGMOID:
+                case GGML_UNARY_OP_GELU:
+                case GGML_UNARY_OP_GELU_QUICK:
+                case GGML_UNARY_OP_SILU:
+                    return ggml_is_contiguous(op->src[0]);
+                default:
+                    return false;
+            }
+        case GGML_OP_NONE:
+        case GGML_OP_RESHAPE:
+        case GGML_OP_VIEW:
+        case GGML_OP_TRANSPOSE:
+        case GGML_OP_PERMUTE:
+        case GGML_OP_CONCAT:
+        case GGML_OP_ADD:
+        case GGML_OP_SUB:
+        case GGML_OP_ACC:
+        case GGML_OP_MUL:
+        case GGML_OP_DIV:
+        case GGML_OP_REPEAT:
+        case GGML_OP_SCALE:
+        case GGML_OP_CLAMP:
+            return true;
+        case GGML_OP_SQR:
+        case GGML_OP_SQRT:
+        case GGML_OP_SIN:
+        case GGML_OP_COS:
+            return ggml_is_contiguous(op->src[0]);
+        case GGML_OP_SUM_ROWS:
+        case GGML_OP_SOFT_MAX:
+        case GGML_OP_RMS_NORM:
+        case GGML_OP_GROUP_NORM:
+            return ctx->support_simdgroup_reduction;
+        case GGML_OP_NORM:
+        case GGML_OP_ROPE:
+            return true;
+        case GGML_OP_IM2COL:
+            return op->src[0]->type == GGML_TYPE_F16;
+        case GGML_OP_POOL_1D:
+        case GGML_OP_POOL_2D:
+            return false;
+        case GGML_OP_UPSCALE:
+        case GGML_OP_PAD:
+        case GGML_OP_ARANGE:
+        case GGML_OP_TIMESTEP_EMBEDDING:
+        case GGML_OP_ARGSORT:
+        case GGML_OP_LEAKY_RELU:
+            return true;
+        case GGML_OP_FLASH_ATTN_EXT:
+            if (op->src[1]->type != GGML_TYPE_F16) {
+                return false;
+            }
+            if (op->src[2]->type != GGML_TYPE_F16) {
+                return false;
+            }
+            if (op->src[0]->ne[0] == 256) {
+                return false;
+            }
+            return ctx->support_simdgroup_mm; // TODO: over-restricted for vec-kernels
+        case GGML_OP_SSM_CONV:
+        case GGML_OP_SSM_SCAN:
+            return true;
+        case GGML_OP_MUL_MAT:
+        case GGML_OP_MUL_MAT_ID:
+            return ctx->support_simdgroup_reduction &&
+                (op->src[0]->type != GGML_TYPE_F32 || op->src[1]->type == GGML_TYPE_F32);
+        case GGML_OP_CPY:
+        case GGML_OP_DUP:
+        case GGML_OP_CONT:
+            {
+                switch (op->src[0]->type) {
+                    case GGML_TYPE_F32:
+                        switch (op->type) {
+                           case GGML_TYPE_F32:
+                           case GGML_TYPE_F16:
+                           case GGML_TYPE_Q8_0:
+                           case GGML_TYPE_Q4_0:
+                           case GGML_TYPE_Q4_1:
+                           case GGML_TYPE_Q5_0:
+                           case GGML_TYPE_Q5_1:
+                           case GGML_TYPE_IQ4_NL:
+                                return true;
+                           default:
+                                return false;
+                        }
+                    case GGML_TYPE_F16:
+                        switch (op->type) {
+                           case GGML_TYPE_F32:
+                           case GGML_TYPE_F16:
+                                return true;
+                           default:
+                                return false;
+                        }
+                    default:
+                        return false;
+                };
+            }
+        case GGML_OP_DIAG_MASK_INF:
+        case GGML_OP_GET_ROWS:
+            {
+                return op->ne[3] == 1;
+            }
+        default:
+            return false;
+    }
+}
+
+static void ggml_metal_encode_node(
+     struct ggml_backend_metal_context * ctx,
+                                   int   idx,
+          id<MTLComputeCommandEncoder>   encoder) {
+    struct ggml_cgraph * gf = ctx->gf;
+
+    struct ggml_tensor * node = ggml_graph_node(gf, idx);
+
+    //GGML_LOG_INFO("%s: encoding node %3d, op = %8s\n", __func__, idx, ggml_op_name(node->op));
+
+    struct ggml_tensor * src0 = node->src[0];
+    struct ggml_tensor * src1 = node->src[1];
+    struct ggml_tensor * src2 = node->src[2];
+    struct ggml_tensor * dst  = node;
+
+    if (ggml_is_empty(dst)) {
+        return;
+    }
+
+    switch (dst->op) {
+        case GGML_OP_NONE:
+        case GGML_OP_RESHAPE:
+        case GGML_OP_VIEW:
+        case GGML_OP_TRANSPOSE:
+        case GGML_OP_PERMUTE:
+            {
+                // noop -> next node
+            } return;
+        default:
+            {
+            } break;
+    }
+
+    if (!ggml_metal_supports_op(ctx, dst)) {
+        GGML_LOG_ERROR("%s: error: unsupported op '%s'\n", __func__, ggml_op_desc(dst));
+        GGML_ABORT("unsupported op");
+    }
+
+    const int64_t  ne00 = src0 ? src0->ne[0] : 0;
+    const int64_t  ne01 = src0 ? src0->ne[1] : 0;
+    const int64_t  ne02 = src0 ? src0->ne[2] : 0;
+    const int64_t  ne03 = src0 ? src0->ne[3] : 0;
+
+    const uint64_t nb00 = src0 ? src0->nb[0] : 0;
+    const uint64_t nb01 = src0 ? src0->nb[1] : 0;
+    const uint64_t nb02 = src0 ? src0->nb[2] : 0;
+    const uint64_t nb03 = src0 ? src0->nb[3] : 0;
+
+    const int64_t  ne10 = src1 ? src1->ne[0] : 0;
+    const int64_t  ne11 = src1 ? src1->ne[1] : 0;
+    const int64_t  ne12 = src1 ? src1->ne[2] : 0;
+    const int64_t  ne13 = src1 ? src1->ne[3] : 0;
+
+    const uint64_t nb10 = src1 ? src1->nb[0] : 0;
+    const uint64_t nb11 = src1 ? src1->nb[1] : 0;
+    const uint64_t nb12 = src1 ? src1->nb[2] : 0;
+    const uint64_t nb13 = src1 ? src1->nb[3] : 0;
+
+    const int64_t  ne20 = src2 ? src2->ne[0] : 0;
+    const int64_t  ne21 = src2 ? src2->ne[1] : 0;
+    const int64_t  ne22 = src2 ? src2->ne[2] : 0; GGML_UNUSED(ne22);
+    const int64_t  ne23 = src2 ? src2->ne[3] : 0; GGML_UNUSED(ne23);
+
+    const uint64_t nb20 = src2 ? src2->nb[0] : 0; GGML_UNUSED(nb20);
+    const uint64_t nb21 = src2 ? src2->nb[1] : 0;
+    const uint64_t nb22 = src2 ? src2->nb[2] : 0;
+    const uint64_t nb23 = src2 ? src2->nb[3] : 0;
+
+    const int64_t  ne0  =  dst ?  dst->ne[0] : 0;
+    const int64_t  ne1  =  dst ?  dst->ne[1] : 0;
+    const int64_t  ne2  =  dst ?  dst->ne[2] : 0;
+    const int64_t  ne3  =  dst ?  dst->ne[3] : 0;
+
+    const uint64_t nb0  =  dst ?  dst->nb[0] : 0;
+    const uint64_t nb1  =  dst ?  dst->nb[1] : 0;
+    const uint64_t nb2  =  dst ?  dst->nb[2] : 0;
+    const uint64_t nb3  =  dst ?  dst->nb[3] : 0;
+
+    const enum ggml_type src0t = src0 ? src0->type : GGML_TYPE_COUNT;
+    const enum ggml_type src1t = src1 ? src1->type : GGML_TYPE_COUNT;
+    const enum ggml_type dstt  = dst  ? dst->type  : GGML_TYPE_COUNT;
+
+    size_t offs_src0 = 0;
+    size_t offs_src1 = 0;
+    size_t offs_src2 = 0;
+    size_t offs_dst  = 0;
+
+    id<MTLBuffer> id_src0 = src0 ? ggml_metal_get_buffer(src0, &offs_src0) : nil;
+    id<MTLBuffer> id_src1 = src1 ? ggml_metal_get_buffer(src1, &offs_src1) : nil;
+    id<MTLBuffer> id_src2 = src2 ? ggml_metal_get_buffer(src2, &offs_src2) : nil;
+    id<MTLBuffer> id_dst  = dst  ? ggml_metal_get_buffer(dst,  &offs_dst)  : nil;
+
+    //GGML_LOG_INFO("%s: op - %s\n", __func__, ggml_op_name(dst->op));
+    //if (src0) {
+    //    GGML_LOG_INFO("%s: src0 - %4s [%5lld, %5lld, %5lld], %d, %s\n", __func__, ggml_type_name(src0t), ne00, ne01, ne02,
+    //            ggml_is_contiguous(src0), src0->name);
+    //}
+    //if (src1) {
+    //    GGML_LOG_INFO("%s: src1 - %4s [%5lld, %5lld, %5lld], %d, %s\n", __func__, ggml_type_name(src1t), ne10, ne11, ne12,
+    //            ggml_is_contiguous(src1), src1->name);
+    //}
+    //if (dst) {
+    //    GGML_LOG_INFO("%s: dst  - %4s [%5lld, %5lld, %5lld], 1, %s\n",  __func__, ggml_type_name(dstt),  ne0,  ne1,  ne2,
+    //            dst->name);
+    //}
+
+    switch (dst->op) {
+        case GGML_OP_CONCAT:
+            {
+                id<MTLComputePipelineState> pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_CONCAT].pipeline;
+
+                const int32_t dim = ((const int32_t *) dst->op_params)[0];
+
+                [encoder setComputePipelineState:pipeline];
+                [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
+                [encoder setBuffer:id_src1 offset:offs_src1 atIndex:1];
+                [encoder setBuffer:id_dst  offset:offs_dst  atIndex:2];
+                [encoder setBytes:&ne00 length:sizeof(ne00) atIndex:3];
+                [encoder setBytes:&ne01 length:sizeof(ne01) atIndex:4];
+                [encoder setBytes:&ne02 length:sizeof(ne02) atIndex:5];
+                [encoder setBytes:&ne03 length:sizeof(ne03) atIndex:6];
+                [encoder setBytes:&nb00 length:sizeof(nb00) atIndex:7];
+                [encoder setBytes:&nb01 length:sizeof(nb01) atIndex:8];
+                [encoder setBytes:&nb02 length:sizeof(nb02) atIndex:9];
+                [encoder setBytes:&nb03 length:sizeof(nb03) atIndex:10];
+                [encoder setBytes:&ne10 length:sizeof(ne10) atIndex:11];
+                [encoder setBytes:&ne11 length:sizeof(ne11) atIndex:12];
+                [encoder setBytes:&ne12 length:sizeof(ne12) atIndex:13];
+                [encoder setBytes:&ne13 length:sizeof(ne13) atIndex:14];
+                [encoder setBytes:&nb10 length:sizeof(nb10) atIndex:15];
+                [encoder setBytes:&nb11 length:sizeof(nb11) atIndex:16];
+                [encoder setBytes:&nb12 length:sizeof(nb12) atIndex:17];
+                [encoder setBytes:&nb13 length:sizeof(nb13) atIndex:18];
+                [encoder setBytes:&ne0  length:sizeof(ne0)  atIndex:19];
+                [encoder setBytes:&ne1  length:sizeof(ne1)  atIndex:20];
+                [encoder setBytes:&ne2  length:sizeof(ne2)  atIndex:21];
+                [encoder setBytes:&ne3  length:sizeof(ne3)  atIndex:22];
+                [encoder setBytes:&nb0  length:sizeof(nb0)  atIndex:23];
+                [encoder setBytes:&nb1  length:sizeof(nb1)  atIndex:24];
+                [encoder setBytes:&nb2  length:sizeof(nb2)  atIndex:25];
+                [encoder setBytes:&nb3  length:sizeof(nb3)  atIndex:26];
+                [encoder setBytes:&dim  length:sizeof(dim)  atIndex:27];
+
+                const int nth = MIN(1024, ne0);
+
+                [encoder dispatchThreadgroups:MTLSizeMake(ne1, ne2, ne3) threadsPerThreadgroup:MTLSizeMake(nth, 1, 1)];
+            } break;
+        case GGML_OP_ADD:
+        case GGML_OP_SUB:
+        case GGML_OP_MUL:
+        case GGML_OP_DIV:
+            {
+                GGML_ASSERT(src0t == GGML_TYPE_F32);
+                GGML_ASSERT(src1t == GGML_TYPE_F32);
+
+                const size_t offs = 0;
+
+                bool bcast_row = false;
+
+                int64_t nb = ne00; // used by the "row" kernels
+
+                id<MTLComputePipelineState> pipeline = nil;
+
+                if (ggml_nelements(src1) == ne10 && ggml_is_contiguous(src1) && ne00 % 4 == 0 && ne10 % 4 == 0) {
+                    GGML_ASSERT(ggml_is_contiguous(src0));
+
+                    // src1 is a row
+                    GGML_ASSERT(ne11 == 1);
+
+                    nb = ne00 / 4;
+                    switch (dst->op) {
+                        case GGML_OP_ADD: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_ADD_ROW].pipeline; break;
+                        case GGML_OP_SUB: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_SUB_ROW].pipeline; break;
+                        case GGML_OP_MUL: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_ROW].pipeline; break;
+                        case GGML_OP_DIV: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_DIV_ROW].pipeline; break;
+                        default: GGML_ABORT("fatal error");
+                    }
+
+                    bcast_row = true;
+                } else {
+                    switch (dst->op) {
+                        case GGML_OP_ADD: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_ADD].pipeline; break;
+                        case GGML_OP_SUB: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_SUB].pipeline; break;
+                        case GGML_OP_MUL: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL].pipeline; break;
+                        case GGML_OP_DIV: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_DIV].pipeline; break;
+                        default: GGML_ABORT("fatal error");
+                    }
+                }
+
+                [encoder setComputePipelineState:pipeline];
+                [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
+                [encoder setBuffer:id_src1 offset:offs_src1 atIndex:1];
+                [encoder setBuffer:id_dst  offset:offs_dst  atIndex:2];
+                [encoder setBytes:&ne00 length:sizeof(ne00) atIndex:3];
+                [encoder setBytes:&ne01 length:sizeof(ne01) atIndex:4];
+                [encoder setBytes:&ne02 length:sizeof(ne02) atIndex:5];
+                [encoder setBytes:&ne03 length:sizeof(ne03) atIndex:6];
+                [encoder setBytes:&nb00 length:sizeof(nb00) atIndex:7];
+                [encoder setBytes:&nb01 length:sizeof(nb01) atIndex:8];
+                [encoder setBytes:&nb02 length:sizeof(nb02) atIndex:9];
+                [encoder setBytes:&nb03 length:sizeof(nb03) atIndex:10];
+                [encoder setBytes:&ne10 length:sizeof(ne10) atIndex:11];
+                [encoder setBytes:&ne11 length:sizeof(ne11) atIndex:12];
+                [encoder setBytes:&ne12 length:sizeof(ne12) atIndex:13];
+                [encoder setBytes:&ne13 length:sizeof(ne13) atIndex:14];
+                [encoder setBytes:&nb10 length:sizeof(nb10) atIndex:15];
+                [encoder setBytes:&nb11 length:sizeof(nb11) atIndex:16];
+                [encoder setBytes:&nb12 length:sizeof(nb12) atIndex:17];
+                [encoder setBytes:&nb13 length:sizeof(nb13) atIndex:18];
+                [encoder setBytes:&ne0  length:sizeof(ne0)  atIndex:19];
+                [encoder setBytes:&ne1  length:sizeof(ne1)  atIndex:20];
+                [encoder setBytes:&ne2  length:sizeof(ne2)  atIndex:21];
+                [encoder setBytes:&ne3  length:sizeof(ne3)  atIndex:22];
+                [encoder setBytes:&nb0  length:sizeof(nb0)  atIndex:23];
+                [encoder setBytes:&nb1  length:sizeof(nb1)  atIndex:24];
+                [encoder setBytes:&nb2  length:sizeof(nb2)  atIndex:25];
+                [encoder setBytes:&nb3  length:sizeof(nb3)  atIndex:26];
+                [encoder setBytes:&offs length:sizeof(offs) atIndex:27];
+                [encoder setBytes:&nb   length:sizeof(nb)   atIndex:28];
+
+                if (bcast_row) {
+                    const int64_t n = ggml_nelements(dst)/4;
+
+                    [encoder dispatchThreadgroups:MTLSizeMake(n, 1, 1) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)];
+                } else {
+                    const int nth = MIN((int) pipeline.maxTotalThreadsPerThreadgroup, ne0);
+
+                    [encoder dispatchThreadgroups:MTLSizeMake(ne01, ne02, ne03) threadsPerThreadgroup:MTLSizeMake(nth, 1, 1)];
+                }
+            } break;
+        case GGML_OP_REPEAT:
+            {
+                id<MTLComputePipelineState> pipeline;
+
+                switch (src0t) {
+                    case GGML_TYPE_F32: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_REPEAT_F32].pipeline; break;
+                    case GGML_TYPE_F16: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_REPEAT_F16].pipeline; break;
+                    case GGML_TYPE_I32: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_REPEAT_I32].pipeline; break;
+                    case GGML_TYPE_I16: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_REPEAT_I16].pipeline; break;
+                    default: GGML_ABORT("fatal error");
+                }
+
+                [encoder setComputePipelineState:pipeline];
+                [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
+                [encoder setBuffer:id_dst  offset:offs_dst  atIndex:1];
+                [encoder setBytes:&ne00 length:sizeof(ne00) atIndex:2];
+                [encoder setBytes:&ne01 length:sizeof(ne01) atIndex:3];
+                [encoder setBytes:&ne02 length:sizeof(ne02) atIndex:4];
+                [encoder setBytes:&ne03 length:sizeof(ne03) atIndex:5];
+                [encoder setBytes:&nb00 length:sizeof(nb00) atIndex:6];
+                [encoder setBytes:&nb01 length:sizeof(nb01) atIndex:7];
+                [encoder setBytes:&nb02 length:sizeof(nb02) atIndex:8];
+                [encoder setBytes:&nb03 length:sizeof(nb03) atIndex:9];
+                [encoder setBytes:&ne0  length:sizeof(ne0)  atIndex:10];
+                [encoder setBytes:&ne1  length:sizeof(ne1)  atIndex:11];
+                [encoder setBytes:&ne2  length:sizeof(ne2)  atIndex:12];
+                [encoder setBytes:&ne3  length:sizeof(ne3)  atIndex:13];
+                [encoder setBytes:&nb0  length:sizeof(nb0)  atIndex:14];
+                [encoder setBytes:&nb1  length:sizeof(nb1)  atIndex:15];
+                [encoder setBytes:&nb2  length:sizeof(nb2)  atIndex:16];
+                [encoder setBytes:&nb3  length:sizeof(nb3)  atIndex:17];
+
+                const int nth = MIN((int) pipeline.maxTotalThreadsPerThreadgroup, ne0);
+
+                [encoder dispatchThreadgroups:MTLSizeMake(ne1, ne2, ne3) threadsPerThreadgroup:MTLSizeMake(nth, 1, 1)];
+            } break;
+        case GGML_OP_ACC:
+            {
+                GGML_ASSERT(src0t == GGML_TYPE_F32);
+                GGML_ASSERT(src1t == GGML_TYPE_F32);
+                GGML_ASSERT(dstt  == GGML_TYPE_F32);
+
+                GGML_ASSERT(ggml_is_contiguous(src0));
+                GGML_ASSERT(ggml_is_contiguous(src1));
+
+                const size_t pnb1 = ((const int32_t *) dst->op_params)[0];
+                const size_t pnb2 = ((const int32_t *) dst->op_params)[1];
+                const size_t pnb3 = ((const int32_t *) dst->op_params)[2];
+                const size_t offs = ((const int32_t *) dst->op_params)[3];
+
+                const bool inplace = (bool) ((const int32_t *) dst->op_params)[4];
+
+                if (!inplace) {
+                    // run a separete kernel to cpy src->dst
+                    // not sure how to avoid this
+                    // TODO: make a simpler cpy_bytes kernel
+
+                    const id<MTLComputePipelineState> pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_CPY_F32_F32].pipeline;
+
+                    [encoder setComputePipelineState:pipeline];
+                    [encoder setBuffer:id_src0 offset:offs_src0        atIndex:0];
+                    [encoder setBuffer:id_dst  offset:offs_dst         atIndex:1];
+                    [encoder setBytes:&ne00    length:sizeof( int64_t) atIndex:2];
+                    [encoder setBytes:&ne01    length:sizeof( int64_t) atIndex:3];
+                    [encoder setBytes:&ne02    length:sizeof( int64_t) atIndex:4];
+                    [encoder setBytes:&ne03    length:sizeof( int64_t) atIndex:5];
+                    [encoder setBytes:&nb00    length:sizeof(uint64_t) atIndex:6];
+                    [encoder setBytes:&nb01    length:sizeof(uint64_t) atIndex:7];
+                    [encoder setBytes:&nb02    length:sizeof(uint64_t) atIndex:8];
+                    [encoder setBytes:&nb03    length:sizeof(uint64_t) atIndex:9];
+                    [encoder setBytes:&ne0     length:sizeof( int64_t) atIndex:10];
+                    [encoder setBytes:&ne1     length:sizeof( int64_t) atIndex:11];
+                    [encoder setBytes:&ne2     length:sizeof( int64_t) atIndex:12];
+                    [encoder setBytes:&ne3     length:sizeof( int64_t) atIndex:13];
+                    [encoder setBytes:&nb0     length:sizeof(uint64_t) atIndex:14];
+                    [encoder setBytes:&nb1     length:sizeof(uint64_t) atIndex:15];
+                    [encoder setBytes:&nb2     length:sizeof(uint64_t) atIndex:16];
+                    [encoder setBytes:&nb3     length:sizeof(uint64_t) atIndex:17];
+
+                    const int nth = MIN((int) pipeline.maxTotalThreadsPerThreadgroup, ne00);
+
+                    [encoder dispatchThreadgroups:MTLSizeMake(ne01, ne02, ne03) threadsPerThreadgroup:MTLSizeMake(nth, 1, 1)];
+                }
+
+                const id<MTLComputePipelineState> pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_ADD].pipeline;
+
+                [encoder setComputePipelineState:pipeline];
+                [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
+                [encoder setBuffer:id_src1 offset:offs_src1 atIndex:1];
+                [encoder setBuffer:id_dst  offset:offs_dst  atIndex:2];
+                [encoder setBytes:&ne00 length:sizeof(ne00) atIndex:3];
+                [encoder setBytes:&ne01 length:sizeof(ne01) atIndex:4];
+                [encoder setBytes:&ne02 length:sizeof(ne02) atIndex:5];
+                [encoder setBytes:&ne03 length:sizeof(ne03) atIndex:6];
+                [encoder setBytes:&nb00 length:sizeof(nb00) atIndex:7];
+                [encoder setBytes:&pnb1 length:sizeof(pnb1) atIndex:8];
+                [encoder setBytes:&pnb2 length:sizeof(pnb2) atIndex:9];
+                [encoder setBytes:&pnb3 length:sizeof(pnb3) atIndex:10];
+                [encoder setBytes:&ne10 length:sizeof(ne10) atIndex:11];
+                [encoder setBytes:&ne11 length:sizeof(ne11) atIndex:12];
+                [encoder setBytes:&ne12 length:sizeof(ne12) atIndex:13];
+                [encoder setBytes:&ne13 length:sizeof(ne13) atIndex:14];
+                [encoder setBytes:&nb10 length:sizeof(nb10) atIndex:15];
+                [encoder setBytes:&nb11 length:sizeof(nb11) atIndex:16];
+                [encoder setBytes:&nb12 length:sizeof(nb12) atIndex:17];
+                [encoder setBytes:&nb13 length:sizeof(nb13) atIndex:18];
+                [encoder setBytes:&ne0  length:sizeof(ne0)  atIndex:19];
+                [encoder setBytes:&ne1  length:sizeof(ne1)  atIndex:20];
+                [encoder setBytes:&ne2  length:sizeof(ne2)  atIndex:21];
+                [encoder setBytes:&ne3  length:sizeof(ne3)  atIndex:22];
+                [encoder setBytes:&nb0  length:sizeof(nb0)  atIndex:23];
+                [encoder setBytes:&pnb1 length:sizeof(pnb1) atIndex:24];
+                [encoder setBytes:&pnb2 length:sizeof(pnb2) atIndex:25];
+                [encoder setBytes:&pnb3 length:sizeof(pnb3) atIndex:26];
+                [encoder setBytes:&offs length:sizeof(offs) atIndex:27];
+
+                const int nth = MIN((int) pipeline.maxTotalThreadsPerThreadgroup, ne00);
+
+                [encoder dispatchThreadgroups:MTLSizeMake(ne11, ne12, ne13) threadsPerThreadgroup:MTLSizeMake(nth, 1, 1)];
+            } break;
+        case GGML_OP_SCALE:
+            {
+                GGML_ASSERT(ggml_is_contiguous(src0));
+
+                float scale;
+                memcpy(&scale, dst->op_params, sizeof(scale));
+
+                int64_t n = ggml_nelements(dst);
+
+                id<MTLComputePipelineState> pipeline = nil;
+
+                if (n % 4 == 0) {
+                    n /= 4;
+                    pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_SCALE_4].pipeline;
+                } else {
+                    pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_SCALE].pipeline;
+                }
+
+                [encoder setComputePipelineState:pipeline];
+                [encoder setBuffer:id_src0   offset:offs_src0 atIndex:0];
+                [encoder setBuffer:id_dst    offset:offs_dst  atIndex:1];
+                [encoder setBytes:&scale length:sizeof(scale) atIndex:2];
+
+                [encoder dispatchThreadgroups:MTLSizeMake(n, 1, 1) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)];
+            } break;
+        case GGML_OP_CLAMP:
+            {
+                id<MTLComputePipelineState> pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_CLAMP].pipeline;
+
+                float min;
+                float max;
+                memcpy(&min, ((const int32_t *) dst->op_params) + 0, sizeof(float));
+                memcpy(&max, ((const int32_t *) dst->op_params) + 1, sizeof(float));
+
+                [encoder setComputePipelineState:pipeline];
+                [encoder setBuffer:id_src0   offset:offs_src0 atIndex:0];
+                [encoder setBuffer:id_dst    offset:offs_dst  atIndex:1];
+                [encoder setBytes:&min length:sizeof(min) atIndex:2];
+                [encoder setBytes:&max length:sizeof(max) atIndex:3];
+
+                const int64_t n = ggml_nelements(dst);
+
+                [encoder dispatchThreadgroups:MTLSizeMake(n, 1, 1) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)];
+            } break;
+        case GGML_OP_UNARY:
+            switch (ggml_get_unary_op(node)) {
+                // we are not taking into account the strides, so for now require contiguous tensors
+                GGML_ASSERT(ggml_is_contiguous(src0));
+
+                case GGML_UNARY_OP_TANH:
+                {
+                    id<MTLComputePipelineState> pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_TANH].pipeline;
+
+                    [encoder setComputePipelineState:pipeline];
+                    [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
+                    [encoder setBuffer:id_dst  offset:offs_dst  atIndex:1];
+
+                    const int64_t n = ggml_nelements(dst);
+
+                    [encoder dispatchThreadgroups:MTLSizeMake(n, 1, 1) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)];
+                } break;
+                case GGML_UNARY_OP_RELU:
+                {
+                    id<MTLComputePipelineState> pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_RELU].pipeline;
+
+                    [encoder setComputePipelineState:pipeline];
+                    [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
+                    [encoder setBuffer:id_dst  offset:offs_dst  atIndex:1];
+
+                    const int64_t n = ggml_nelements(dst);
+
+                    [encoder dispatchThreadgroups:MTLSizeMake(n, 1, 1) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)];
+                } break;
+                case GGML_UNARY_OP_SIGMOID:
+                {
+                    id<MTLComputePipelineState> pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_SIGMOID].pipeline;
+
+                    [encoder setComputePipelineState:pipeline];
+                    [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
+                    [encoder setBuffer:id_dst  offset:offs_dst  atIndex:1];
+
+                    const int64_t n = ggml_nelements(dst);
+
+                    [encoder dispatchThreadgroups:MTLSizeMake(n, 1, 1) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)];
+                } break;
+                case GGML_UNARY_OP_GELU:
+                {
+                    int64_t n = ggml_nelements(dst);
+
+                    id<MTLComputePipelineState> pipeline = nil;
+
+                    if (n % 4 == 0) {
+                        pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_GELU_4].pipeline;
+                        n /= 4;
+                    } else {
+                        pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_GELU].pipeline;
+                    }
+
+                    [encoder setComputePipelineState:pipeline];
+                    [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
+                    [encoder setBuffer:id_dst  offset:offs_dst  atIndex:1];
+
+                    [encoder dispatchThreadgroups:MTLSizeMake(n, 1, 1) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)];
+                } break;
+                case GGML_UNARY_OP_GELU_QUICK:
+                {
+                    int64_t n = ggml_nelements(dst);
+
+                    id<MTLComputePipelineState> pipeline = nil;
+
+                    if (n % 4 == 0) {
+                        pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_GELU_QUICK_4].pipeline;
+                        n /= 4;
+                    } else {
+                        pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_GELU_QUICK].pipeline;
+                    }
+
+                    [encoder setComputePipelineState:pipeline];
+                    [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
+                    [encoder setBuffer:id_dst  offset:offs_dst  atIndex:1];
+
+                    [encoder dispatchThreadgroups:MTLSizeMake(n, 1, 1) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)];
+                } break;
+                case GGML_UNARY_OP_SILU:
+                {
+                    int64_t n = ggml_nelements(dst);
+
+                    id<MTLComputePipelineState> pipeline = nil;
+
+                    if (n % 4 == 0) {
+                        pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_SILU_4].pipeline;
+                        n /= 4;
+                    } else {
+                        pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_SILU].pipeline;
+                    }
+
+                    [encoder setComputePipelineState:pipeline];
+                    [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
+                    [encoder setBuffer:id_dst  offset:offs_dst  atIndex:1];
+
+                    [encoder dispatchThreadgroups:MTLSizeMake(n, 1, 1) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)];
+                } break;
+                default:
+                {
+                    GGML_LOG_WARN("%s: node %3d, op = %8s not implemented\n", __func__, idx, ggml_op_name(dst->op));
+                    GGML_ABORT("fatal error");
+                }
+            } break;
+        case GGML_OP_SQR:
+            {
+                GGML_ASSERT(ggml_is_contiguous(src0));
+
+                id<MTLComputePipelineState> pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_SQR].pipeline;
+
+                [encoder setComputePipelineState:pipeline];
+                [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
+                [encoder setBuffer:id_dst  offset:offs_dst atIndex:1];
+
+                const int64_t n = ggml_nelements(dst);
+
+                [encoder dispatchThreadgroups:MTLSizeMake(n, 1, 1) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)];
+            } break;
+        case GGML_OP_SQRT:
+            {
+                GGML_ASSERT(ggml_is_contiguous(src0));
+
+                id<MTLComputePipelineState> pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_SQRT].pipeline;
+
+                [encoder setComputePipelineState:pipeline];
+                [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
+                [encoder setBuffer:id_dst  offset:offs_dst atIndex:1];
+
+                const int64_t n = ggml_nelements(dst);
+
+                [encoder dispatchThreadgroups:MTLSizeMake(n, 1, 1) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)];
+            } break;
+        case GGML_OP_SIN:
+            {
+                GGML_ASSERT(ggml_is_contiguous(src0));
+
+                id<MTLComputePipelineState> pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_SIN].pipeline;
+
+                [encoder setComputePipelineState:pipeline];
+                [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
+                [encoder setBuffer:id_dst  offset:offs_dst atIndex:1];
+
+                const int64_t n = ggml_nelements(dst);
+
+                [encoder dispatchThreadgroups:MTLSizeMake(n, 1, 1) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)];
+            } break;
+        case GGML_OP_COS:
+            {
+                GGML_ASSERT(ggml_is_contiguous(src0));
+
+                id<MTLComputePipelineState> pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_COS].pipeline;
+
+                [encoder setComputePipelineState:pipeline];
+                [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
+                [encoder setBuffer:id_dst  offset:offs_dst atIndex:1];
+
+                const int64_t n = ggml_nelements(dst);
+
+                [encoder dispatchThreadgroups:MTLSizeMake(n, 1, 1) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)];
+            } break;
+        case GGML_OP_SUM_ROWS:
+            {
+                GGML_ASSERT(src0->nb[0] == ggml_type_size(src0->type));
+
+                id<MTLComputePipelineState> pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_SUM_ROWS].pipeline;
+
+                [encoder setComputePipelineState:pipeline];
+                [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
+                [encoder setBuffer:id_dst  offset:offs_dst  atIndex:1];
+                [encoder setBytes:&ne00 length:sizeof(ne00) atIndex:2];
+                [encoder setBytes:&ne01 length:sizeof(ne01) atIndex:3];
+                [encoder setBytes:&ne02 length:sizeof(ne02) atIndex:4];
+                [encoder setBytes:&ne03 length:sizeof(ne03) atIndex:5];
+                [encoder setBytes:&nb00 length:sizeof(nb00) atIndex:6];
+                [encoder setBytes:&nb01 length:sizeof(nb01) atIndex:7];
+                [encoder setBytes:&nb02 length:sizeof(nb02) atIndex:8];
+                [encoder setBytes:&nb03 length:sizeof(nb03) atIndex:9];
+                [encoder setBytes:&ne10 length:sizeof(ne10) atIndex:10];
+                [encoder setBytes:&ne11 length:sizeof(ne11) atIndex:11];
+                [encoder setBytes:&ne12 length:sizeof(ne12) atIndex:12];
+                [encoder setBytes:&ne13 length:sizeof(ne13) atIndex:13];
+                [encoder setBytes:&nb10 length:sizeof(nb10) atIndex:14];
+                [encoder setBytes:&nb11 length:sizeof(nb11) atIndex:15];
+                [encoder setBytes:&nb12 length:sizeof(nb12) atIndex:16];
+                [encoder setBytes:&nb13 length:sizeof(nb13) atIndex:17];
+                [encoder setBytes:&ne0  length:sizeof(ne0)  atIndex:18];
+                [encoder setBytes:&ne1  length:sizeof(ne1)  atIndex:19];
+                [encoder setBytes:&ne2  length:sizeof(ne2)  atIndex:20];
+                [encoder setBytes:&ne3  length:sizeof(ne3)  atIndex:21];
+                [encoder setBytes:&nb0  length:sizeof(nb0)  atIndex:22];
+                [encoder setBytes:&nb1  length:sizeof(nb1)  atIndex:23];
+                [encoder setBytes:&nb2  length:sizeof(nb2)  atIndex:24];
+                [encoder setBytes:&nb3  length:sizeof(nb3)  atIndex:25];
+
+                [encoder dispatchThreadgroups:MTLSizeMake(ne01, ne02, ne03) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)];
+            } break;
+        case GGML_OP_SOFT_MAX:
+            {
+                GGML_ASSERT(!src1 || src1->type == GGML_TYPE_F16 || src1->type == GGML_TYPE_F32);
+
+                int nth = 32; // SIMD width
+
+                id<MTLComputePipelineState> pipeline = nil;
+
+                const bool use_f16 = (src1 && src1->type == GGML_TYPE_F16);
+
+                if (ne00%4 == 0) {
+                    while (nth < ne00/4 && nth*ne01*ne02*ne03 < 256) {
+                        nth *= 2;
+                    }
+                    if (use_f16) {
+                        pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_SOFT_MAX_F16_4].pipeline;
+                    } else {
+                        pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_SOFT_MAX_F32_4].pipeline;
+                    }
+                } else {
+                    while (nth < ne00 && nth*ne01*ne02*ne03 < 256) {
+                        nth *= 2;
+                    }
+                    if (use_f16) {
+                        pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_SOFT_MAX_F16].pipeline;
+                    } else {
+                        pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_SOFT_MAX_F32].pipeline;
+                    }
+                }
+
+                float scale;
+                float max_bias;
+
+                memcpy(&scale,    ((const int32_t *) dst->op_params) + 0, sizeof(scale));
+                memcpy(&max_bias, ((const int32_t *) dst->op_params) + 1, sizeof(max_bias));
+
+                const int64_t nrows_x = ggml_nrows(src0);
+                const int64_t nrows_y = src0->ne[1];
+
+                const uint32_t n_head      = nrows_x/nrows_y;
+                const uint32_t n_head_log2 = 1u << (uint32_t) floorf(log2f((float) n_head));
+
+                const float m0 = powf(2.0f, -(max_bias       ) / n_head_log2);
+                const float m1 = powf(2.0f, -(max_bias / 2.0f) / n_head_log2);
+
+                [encoder setComputePipelineState:pipeline];
+                [encoder setBuffer:id_src0 offset:offs_src0   atIndex:0];
+                if (id_src1) {
+                    [encoder setBuffer:id_src1 offset:offs_src1   atIndex:1];
+                } else {
+                    [encoder setBuffer:id_src0 offset:offs_src0   atIndex:1];
+                }
+                [encoder setBuffer:id_dst      offset:offs_dst            atIndex:2];
+                [encoder setBytes:&ne00        length:sizeof(ne00)        atIndex:3];
+                [encoder setBytes:&ne01        length:sizeof(ne01)        atIndex:4];
+                [encoder setBytes:&ne02        length:sizeof(ne02)        atIndex:5];
+                [encoder setBytes:&scale       length:sizeof(scale)       atIndex:6];
+                [encoder setBytes:&max_bias    length:sizeof(max_bias)    atIndex:7];
+                [encoder setBytes:&m0          length:sizeof(m0)          atIndex:8];
+                [encoder setBytes:&m1          length:sizeof(m1)          atIndex:9];
+                [encoder setBytes:&n_head_log2 length:sizeof(n_head_log2) atIndex:10];
+                [encoder setThreadgroupMemoryLength:32*sizeof(float) atIndex:0];
+
+                [encoder dispatchThreadgroups:MTLSizeMake(ne01*ne02*ne03, 1, 1) threadsPerThreadgroup:MTLSizeMake(nth, 1, 1)];
+            } break;
+        case GGML_OP_DIAG_MASK_INF:
+            {
+                const int n_past = ((const int32_t *)(dst->op_params))[0];
+
+                id<MTLComputePipelineState> pipeline = nil;
+
+                if (ne00%8 == 0) {
+                    pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_DIAG_MASK_INF_8].pipeline;
+                } else {
+                    pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_DIAG_MASK_INF].pipeline;
+                }
+
+                [encoder setComputePipelineState:pipeline];
+                [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
+                [encoder setBuffer:id_dst  offset:offs_dst  atIndex:1];
+                [encoder setBytes:&ne00   length:sizeof(ne00) atIndex:2];
+                [encoder setBytes:&ne01   length:sizeof(ne01) atIndex:3];
+                [encoder setBytes:&n_past length:sizeof(int)  atIndex:4];
+
+                if (ne00%8 == 0) {
+                    [encoder dispatchThreadgroups:MTLSizeMake(ne00*ne01*ne02/8, 1, 1) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)];
+                }
+                else {
+                    [encoder dispatchThreadgroups:MTLSizeMake(ne00, ne01, ne02) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)];
+                }
+            } break;
+        case GGML_OP_SSM_CONV:
+            {
+                GGML_ASSERT(src0t == GGML_TYPE_F32);
+                GGML_ASSERT(src1t == GGML_TYPE_F32);
+
+                GGML_ASSERT(ggml_is_contiguous(src0));
+                GGML_ASSERT(ggml_is_contiguous(src1));
+
+                id<MTLComputePipelineState> pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_SSM_CONV_F32].pipeline;
+
+                [encoder setComputePipelineState:pipeline];
+                [encoder setBuffer:id_src0 offset:offs_src0    atIndex:0];
+                [encoder setBuffer:id_src1 offset:offs_src1    atIndex:1];
+                [encoder setBuffer:id_dst  offset:offs_dst     atIndex:2];
+                [encoder setBytes:&ne00    length:sizeof(ne00) atIndex:3];
+                [encoder setBytes:&ne01    length:sizeof(ne01) atIndex:4];
+                [encoder setBytes:&ne02    length:sizeof(ne02) atIndex:5];
+                [encoder setBytes:&nb00    length:sizeof(nb00) atIndex:6];
+                [encoder setBytes:&nb01    length:sizeof(nb01) atIndex:7];
+                [encoder setBytes:&nb02    length:sizeof(nb02) atIndex:8];
+                [encoder setBytes:&ne10    length:sizeof(ne10) atIndex:9];
+                [encoder setBytes:&ne11    length:sizeof(ne11) atIndex:10];
+                [encoder setBytes:&nb10    length:sizeof(nb10) atIndex:11];
+                [encoder setBytes:&nb11    length:sizeof(nb11) atIndex:12];
+                [encoder setBytes:&ne0     length:sizeof(ne0)  atIndex:13];
+                [encoder setBytes:&ne1     length:sizeof(ne1)  atIndex:14];
+                [encoder setBytes:&ne2     length:sizeof(ne2)  atIndex:15];
+                [encoder setBytes:&nb0     length:sizeof(nb0)  atIndex:16];
+                [encoder setBytes:&nb1     length:sizeof(nb1)  atIndex:17];
+                [encoder setBytes:&nb2     length:sizeof(nb2)  atIndex:18];
+
+                [encoder dispatchThreadgroups:MTLSizeMake(ne01, ne1, ne02) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)];
+            } break;
+        case GGML_OP_SSM_SCAN:
+            {
+                struct ggml_tensor * src3 = node->src[3];
+                struct ggml_tensor * src4 = node->src[4];
+                struct ggml_tensor * src5 = node->src[5];
+
+                GGML_ASSERT(src3);
+                GGML_ASSERT(src4);
+                GGML_ASSERT(src5);
+
+                size_t offs_src3 = 0;
+                size_t offs_src4 = 0;
+                size_t offs_src5 = 0;
+
+                id<MTLBuffer> id_src3 = src3 ? ggml_metal_get_buffer(src3, &offs_src3) : nil;
+                id<MTLBuffer> id_src4 = src4 ? ggml_metal_get_buffer(src4, &offs_src4) : nil;
+                id<MTLBuffer> id_src5 = src5 ? ggml_metal_get_buffer(src5, &offs_src5) : nil;
+
+                const int64_t  ne30 = src3->ne[0]; GGML_UNUSED(ne30);
+                const int64_t  ne31 = src3->ne[1]; GGML_UNUSED(ne31);
+
+                const uint64_t nb30 = src3->nb[0];
+                const uint64_t nb31 = src3->nb[1];
+
+                const int64_t  ne40 = src4->ne[0]; GGML_UNUSED(ne40);
+                const int64_t  ne41 = src4->ne[1]; GGML_UNUSED(ne41);
+                const int64_t  ne42 = src4->ne[2]; GGML_UNUSED(ne42);
+
+                const uint64_t nb40 = src4->nb[0];
+                const uint64_t nb41 = src4->nb[1];
+                const uint64_t nb42 = src4->nb[2];
+
+                const int64_t  ne50 = src5->ne[0]; GGML_UNUSED(ne50);
+                const int64_t  ne51 = src5->ne[1]; GGML_UNUSED(ne51);
+                const int64_t  ne52 = src5->ne[2]; GGML_UNUSED(ne52);
+
+                const uint64_t nb50 = src5->nb[0];
+                const uint64_t nb51 = src5->nb[1];
+                const uint64_t nb52 = src5->nb[2];
+
+                const int64_t d_state      = ne00;
+                const int64_t d_inner      = ne01;
+                const int64_t n_seq_tokens = ne11;
+                const int64_t n_seqs       = ne02;
+
+                id<MTLComputePipelineState> pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_SSM_SCAN_F32].pipeline;
+
+                [encoder setComputePipelineState:pipeline];
+                [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
+                [encoder setBuffer:id_src1 offset:offs_src1 atIndex:1];
+                [encoder setBuffer:id_src2 offset:offs_src2 atIndex:2];
+                [encoder setBuffer:id_src3 offset:offs_src3 atIndex:3];
+                [encoder setBuffer:id_src4 offset:offs_src4 atIndex:4];
+                [encoder setBuffer:id_src5 offset:offs_src5 atIndex:5];
+                [encoder setBuffer:id_dst  offset:offs_dst  atIndex:6];
+
+                [encoder setBytes:&d_state      length:sizeof(d_state)      atIndex:7];
+                [encoder setBytes:&d_inner      length:sizeof(d_inner)      atIndex:8];
+                [encoder setBytes:&n_seq_tokens length:sizeof(n_seq_tokens) atIndex:9];
+                [encoder setBytes:&n_seqs       length:sizeof(n_seqs)       atIndex:10];
+
+                [encoder setBytes:&nb00 length:sizeof(nb00) atIndex:11];
+                [encoder setBytes:&nb01 length:sizeof(nb01) atIndex:12];
+                [encoder setBytes:&nb02 length:sizeof(nb02) atIndex:13];
+                [encoder setBytes:&nb10 length:sizeof(nb10) atIndex:14];
+                [encoder setBytes:&nb11 length:sizeof(nb11) atIndex:15];
+                [encoder setBytes:&nb12 length:sizeof(nb12) atIndex:16];
+                [encoder setBytes:&nb13 length:sizeof(nb13) atIndex:17];
+                [encoder setBytes:&nb20 length:sizeof(nb20) atIndex:18];
+                [encoder setBytes:&nb21 length:sizeof(nb21) atIndex:19];
+                [encoder setBytes:&nb22 length:sizeof(nb22) atIndex:20];
+                [encoder setBytes:&nb30 length:sizeof(nb30) atIndex:21];
+                [encoder setBytes:&nb31 length:sizeof(nb31) atIndex:22];
+                [encoder setBytes:&nb40 length:sizeof(nb40) atIndex:23];
+                [encoder setBytes:&nb41 length:sizeof(nb41) atIndex:24];
+                [encoder setBytes:&nb42 length:sizeof(nb42) atIndex:25];
+                [encoder setBytes:&nb50 length:sizeof(nb50) atIndex:26];
+                [encoder setBytes:&nb51 length:sizeof(nb51) atIndex:27];
+                [encoder setBytes:&nb52 length:sizeof(nb52) atIndex:28];
+
+                [encoder dispatchThreadgroups:MTLSizeMake(d_inner, n_seqs, 1) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)];
+            } break;
+        case GGML_OP_MUL_MAT:
+            {
+                GGML_ASSERT(ne00 == ne10);
+
+                GGML_ASSERT(ne12 % ne02 == 0);
+                GGML_ASSERT(ne13 % ne03 == 0);
+
+                const uint r2 = ne12/ne02;
+                const uint r3 = ne13/ne03;
+
+                // find the break-even point where the matrix-matrix kernel becomes more efficient compared
+                // to the matrix-vector kernel
+                int ne11_mm_min = 1;
+
+#if 0
+                // the numbers below are measured on M2 Ultra for 7B and 13B models
+                // these numbers do not translate to other devices or model sizes
+                // TODO: need to find a better approach
+                        if ([ctx->device.name isEqualToString:@"Apple M2 Ultra"]) {
+                            switch (src0t) {
+                                case GGML_TYPE_F16:  ne11_mm_min = 2;  break;
+                                case GGML_TYPE_Q8_0: ne11_mm_min = 7;  break;
+                                case GGML_TYPE_Q2_K: ne11_mm_min = 15; break;
+                                case GGML_TYPE_Q3_K: ne11_mm_min = 7;  break;
+                                case GGML_TYPE_Q4_0:
+                                case GGML_TYPE_Q4_1: ne11_mm_min = 15; break;
+                                case GGML_TYPE_Q4_K: ne11_mm_min = 11; break;
+                                case GGML_TYPE_Q5_0:                          // not tested yet
+                                case GGML_TYPE_Q5_1: ne11_mm_min = 13; break; // not tested yet
+                                case GGML_TYPE_Q5_K: ne11_mm_min = 7;  break;
+                                case GGML_TYPE_Q6_K: ne11_mm_min = 7;  break;
+                                default:             ne11_mm_min = 1;  break;
+                            }
+                        }
+#endif
+
+                        // for now the matrix-matrix multiplication kernel only works on A14+/M1+ SoCs
+                        // AMD GPU and older A-chips will reuse matrix-vector multiplication kernel
+                        if ([ctx->device supportsFamily:MTLGPUFamilyApple7] &&
+                                !ggml_is_transposed(src0) &&
+                                !ggml_is_transposed(src1) &&
+                                src1t == GGML_TYPE_F32 &&
+                                ne00 % 32 == 0 && ne00 >= 64 &&
+                                (ne11 > ne11_mm_min || (ggml_is_quantized(src0t) && ne12 > 1))) {
+                            //printf("matrix: ne00 = %6d, ne01 = %6d, ne02 = %6d, ne11 = %6d, ne12 = %6d\n", ne00, ne01, ne02, ne11, ne12);
+
+                            // some Metal matrix data types require aligned pointers
+                            // ref: https://developer.apple.com/metal/Metal-Shading-Language-Specification.pdf (Table 2.5)
+                            switch (src0->type) {
+                                case GGML_TYPE_F32:  GGML_ASSERT(nb01 % 16 == 0); break;
+                                case GGML_TYPE_F16:  GGML_ASSERT(nb01 % 8  == 0); break;
+                                default: break;
+                            }
+
+                            id<MTLComputePipelineState> pipeline = nil;
+
+                            switch (src0->type) {
+                                case GGML_TYPE_F32:     pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_F32_F32    ].pipeline; break;
+                                case GGML_TYPE_F16:     pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_F16_F32    ].pipeline; break;
+                                case GGML_TYPE_Q4_0:    pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_Q4_0_F32   ].pipeline; break;
+                                case GGML_TYPE_Q4_1:    pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_Q4_1_F32   ].pipeline; break;
+                                case GGML_TYPE_Q5_0:    pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_Q5_0_F32   ].pipeline; break;
+                                case GGML_TYPE_Q5_1:    pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_Q5_1_F32   ].pipeline; break;
+                                case GGML_TYPE_Q8_0:    pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_Q8_0_F32   ].pipeline; break;
+                                case GGML_TYPE_Q2_K:    pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_Q2_K_F32   ].pipeline; break;
+                                case GGML_TYPE_Q3_K:    pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_Q3_K_F32   ].pipeline; break;
+                                case GGML_TYPE_Q4_K:    pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_Q4_K_F32   ].pipeline; break;
+                                case GGML_TYPE_Q5_K:    pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_Q5_K_F32   ].pipeline; break;
+                                case GGML_TYPE_Q6_K:    pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_Q6_K_F32   ].pipeline; break;
+                                case GGML_TYPE_IQ2_XXS: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_IQ2_XXS_F32].pipeline; break;
+                                case GGML_TYPE_IQ2_XS:  pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_IQ2_XS_F32 ].pipeline; break;
+                                case GGML_TYPE_IQ3_XXS: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_IQ3_XXS_F32].pipeline; break;
+                                case GGML_TYPE_IQ3_S:   pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_IQ3_S_F32  ].pipeline; break;
+                                case GGML_TYPE_IQ2_S:   pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_IQ2_S_F32  ].pipeline; break;
+                                case GGML_TYPE_IQ1_S:   pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_IQ1_S_F32  ].pipeline; break;
+                                case GGML_TYPE_IQ1_M:   pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_IQ1_M_F32  ].pipeline; break;
+                                case GGML_TYPE_IQ4_NL:  pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_IQ4_NL_F32 ].pipeline; break;
+                                case GGML_TYPE_IQ4_XS:  pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_IQ4_XS_F32 ].pipeline; break;
+                                default: GGML_ABORT("MUL MAT-MAT not implemented");
+                            }
+
+                            [encoder setComputePipelineState:pipeline];
+                            [encoder setBuffer:id_src0 offset:offs_src0    atIndex:0];
+                            [encoder setBuffer:id_src1 offset:offs_src1    atIndex:1];
+                            [encoder setBuffer:id_dst  offset:offs_dst     atIndex:2];
+                            [encoder setBytes:&ne00    length:sizeof(ne00) atIndex:3];
+                            [encoder setBytes:&ne02    length:sizeof(ne02) atIndex:4];
+                            [encoder setBytes:&nb01    length:sizeof(nb01) atIndex:5];
+                            [encoder setBytes:&nb02    length:sizeof(nb02) atIndex:6];
+                            [encoder setBytes:&ne12    length:sizeof(ne12) atIndex:7];
+                            [encoder setBytes:&nb10    length:sizeof(nb10) atIndex:8];
+                            [encoder setBytes:&nb11    length:sizeof(nb11) atIndex:9];
+                            [encoder setBytes:&nb12    length:sizeof(nb12) atIndex:10];
+                            [encoder setBytes:&ne0     length:sizeof(ne0)  atIndex:11];
+                            [encoder setBytes:&ne1     length:sizeof(ne1)  atIndex:12];
+                            [encoder setBytes:&r2      length:sizeof(r2)   atIndex:13];
+                            [encoder setBytes:&r3      length:sizeof(r3)   atIndex:14];
+                            [encoder setThreadgroupMemoryLength:8192 atIndex:0];
+                            [encoder dispatchThreadgroups:MTLSizeMake( (ne11 + 31)/32, (ne01 + 63)/64, ne12*ne13) threadsPerThreadgroup:MTLSizeMake(128, 1, 1)];
+                        } else {
+                            int nth0 = 32;
+                            int nth1 = 1;
+                            int nrows = 1;
+                            //printf("vector: ne00 = %6d, ne01 = %6d, ne02 = %6d, ne11 = %6d, ne12 = %6d\n", ne00, ne01, ne02, ne11, ne12);
+
+                            id<MTLComputePipelineState> pipeline = nil;
+
+                            // use custom matrix x vector kernel
+                            switch (src0t) {
+                                case GGML_TYPE_F32:
+                                    {
+                                        GGML_ASSERT(src1t == GGML_TYPE_F32);
+                                        pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_F32_F32].pipeline;
+                                        nrows = 4;
+                                    } break;
+                                case GGML_TYPE_F16:
+                                    {
+                                        nth0 = 32;
+                                        nth1 = 1;
+                                        if (src1t == GGML_TYPE_F32) {
+                                            if (ne11 * ne12 < 4) {
+                                                pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_F16_F32_1ROW].pipeline;
+                                            } else if (ne00 >= 128 && ne01 >= 8 && ne00%4 == 0) {
+                                                pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_F16_F32_L4].pipeline;
+                                                nrows = ne11;
+                                            } else {
+                                                pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_F16_F32].pipeline;
+                                                nrows = 4;
+                                            }
+                                        } else {
+                                            pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_F16_F16].pipeline;
+                                            nrows = 4;
+                                        }
+                                    } break;
+                                case GGML_TYPE_Q4_0:
+                                    {
+                                        nth0 = 8;
+                                        nth1 = 8;
+                                        pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_Q4_0_F32].pipeline;
+                                    } break;
+                                case GGML_TYPE_Q4_1:
+                                    {
+                                        nth0 = 8;
+                                        nth1 = 8;
+                                        pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_Q4_1_F32].pipeline;
+                                    } break;
+                                case GGML_TYPE_Q5_0:
+                                    {
+                                        nth0 = 8;
+                                        nth1 = 8;
+                                        pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_Q5_0_F32].pipeline;
+                                    } break;
+                                case GGML_TYPE_Q5_1:
+                                    {
+                                        nth0 = 8;
+                                        nth1 = 8;
+                                        pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_Q5_1_F32].pipeline;
+                                    } break;
+                                case GGML_TYPE_Q8_0:
+                                    {
+                                        nth0 = 8;
+                                        nth1 = 8;
+                                        pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_Q8_0_F32].pipeline;
+                                    } break;
+                                case GGML_TYPE_Q2_K:
+                                    {
+                                        nth0 = 2;
+                                        nth1 = 32;
+                                        pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_Q2_K_F32].pipeline;
+                                    } break;
+                                case GGML_TYPE_Q3_K:
+                                    {
+                                        nth0 = 2;
+                                        nth1 = 32;
+                                        pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_Q3_K_F32].pipeline;
+                                    } break;
+                                case GGML_TYPE_Q4_K:
+                                    {
+                                        nth0 = 4; //1;
+                                        nth1 = 8; //32;
+                                        pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_Q4_K_F32].pipeline;
+                                    } break;
+                                case GGML_TYPE_Q5_K:
+                                    {
+                                        nth0 = 2;
+                                        nth1 = 32;
+                                        pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_Q5_K_F32].pipeline;
+                                    } break;
+                                case GGML_TYPE_Q6_K:
+                                    {
+                                        nth0 = 2;
+                                        nth1 = 32;
+                                        pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_Q6_K_F32].pipeline;
+                                    } break;
+                                case GGML_TYPE_IQ2_XXS:
+                                    {
+                                        nth0 = 4;
+                                        nth1 = 16;
+                                        pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_IQ2_XXS_F32].pipeline;
+                                    } break;
+                                case GGML_TYPE_IQ2_XS:
+                                    {
+                                        nth0 = 4;
+                                        nth1 = 16;
+                                        pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_IQ2_XS_F32].pipeline;
+                                    } break;
+                                case GGML_TYPE_IQ3_XXS:
+                                    {
+                                        nth0 = 4;
+                                        nth1 = 16;
+                                        pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_IQ3_XXS_F32].pipeline;
+                                    } break;
+                                case GGML_TYPE_IQ3_S:
+                                    {
+                                        nth0 = 4;
+                                        nth1 = 16;
+                                        pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_IQ3_S_F32].pipeline;
+                                    } break;
+                                case GGML_TYPE_IQ2_S:
+                                    {
+                                        nth0 = 4;
+                                        nth1 = 16;
+                                        pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_IQ2_S_F32].pipeline;
+                                    } break;
+                                case GGML_TYPE_IQ1_S:
+                                    {
+                                        nth0 = 4;
+                                        nth1 = 16;
+                                        pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_IQ1_S_F32].pipeline;
+                                    } break;
+                                case GGML_TYPE_IQ1_M:
+                                    {
+                                        nth0 = 4;
+                                        nth1 = 16;
+                                        pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_IQ1_M_F32].pipeline;
+                                    } break;
+                                case GGML_TYPE_IQ4_NL:
+                                    {
+                                        nth0 = 4;
+                                        nth1 = 16;
+                                        pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_IQ4_NL_F32].pipeline;
+                                    } break;
+                                case GGML_TYPE_IQ4_XS:
+                                    {
+                                        nth0 = 4;
+                                        nth1 = 16;
+                                        pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_IQ4_XS_F32].pipeline;
+                                    } break;
+                                default:
+                                    {
+                                        GGML_LOG_ERROR("Asserting on type %d\n", (int)src0t);
+                                        GGML_ABORT("not implemented");
+                                    }
+                            };
+
+                            [encoder setComputePipelineState:pipeline];
+                            [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
+                            [encoder setBuffer:id_src1 offset:offs_src1 atIndex:1];
+                            [encoder setBuffer:id_dst  offset:offs_dst  atIndex:2];
+                            [encoder setBytes:&ne00 length:sizeof(ne00) atIndex:3];
+                            [encoder setBytes:&ne01 length:sizeof(ne01) atIndex:4];
+                            [encoder setBytes:&ne02 length:sizeof(ne02) atIndex:5];
+                            [encoder setBytes:&nb00 length:sizeof(nb00) atIndex:6];
+                            [encoder setBytes:&nb01 length:sizeof(nb01) atIndex:7];
+                            [encoder setBytes:&nb02 length:sizeof(nb02) atIndex:8];
+                            [encoder setBytes:&ne10 length:sizeof(ne10) atIndex:9];
+                            [encoder setBytes:&ne11 length:sizeof(ne11) atIndex:10];
+                            [encoder setBytes:&ne12 length:sizeof(ne12) atIndex:11];
+                            [encoder setBytes:&nb10 length:sizeof(nb10) atIndex:12];
+                            [encoder setBytes:&nb11 length:sizeof(nb11) atIndex:13];
+                            [encoder setBytes:&nb12 length:sizeof(nb12) atIndex:14];
+                            [encoder setBytes:&ne0  length:sizeof(ne0)  atIndex:15];
+                            [encoder setBytes:&ne1  length:sizeof(ne1)  atIndex:16];
+                            [encoder setBytes:&r2   length:sizeof(r2)   atIndex:17];
+                            [encoder setBytes:&r3   length:sizeof(r3)   atIndex:18];
+
+                            if (src0t == GGML_TYPE_Q4_0  || src0t == GGML_TYPE_Q4_1  || src0t == GGML_TYPE_Q5_0 ||
+                                    src0t == GGML_TYPE_Q5_1  || src0t == GGML_TYPE_Q8_0  || src0t == GGML_TYPE_Q2_K ||
+                                    src0t == GGML_TYPE_IQ1_S || src0t == GGML_TYPE_IQ1_M || src0t == GGML_TYPE_IQ2_S) {
+                                [encoder dispatchThreadgroups:MTLSizeMake((ne01 + 7)/8, ne11, ne12*ne13) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
+                            }
+                            else if (src0t == GGML_TYPE_IQ2_XXS || src0t == GGML_TYPE_IQ2_XS) {
+                                const int mem_size = src0t == GGML_TYPE_IQ2_XXS ? 256*8+128 : 512*8+128;
+                                [encoder setThreadgroupMemoryLength:mem_size atIndex:0];
+                                [encoder dispatchThreadgroups:MTLSizeMake((ne01 + 7)/8, ne11, ne12*ne13) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
+                            }
+                            else if (src0t == GGML_TYPE_IQ3_XXS || src0t == GGML_TYPE_IQ3_S) {
+                                const int mem_size = src0t == GGML_TYPE_IQ3_XXS ? 256*4+128 : 512*4;
+                                [encoder setThreadgroupMemoryLength:mem_size atIndex:0];
+                                [encoder dispatchThreadgroups:MTLSizeMake((ne01 + 7)/8, ne11, ne12*ne13) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
+                            }
+                            else if (src0t == GGML_TYPE_IQ4_NL || src0t == GGML_TYPE_IQ4_XS) {
+                                const int mem_size = 32*sizeof(float);
+                                [encoder setThreadgroupMemoryLength:mem_size atIndex:0];
+                                [encoder dispatchThreadgroups:MTLSizeMake((ne01 + 3)/4, ne11, ne12*ne13) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
+                            }
+                            else if (src0t == GGML_TYPE_Q4_K) {
+                                [encoder dispatchThreadgroups:MTLSizeMake((ne01 + 3)/4, ne11, ne12*ne13) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
+                            }
+                            else if (src0t == GGML_TYPE_Q3_K) {
+                                [encoder dispatchThreadgroups:MTLSizeMake((ne01 + 3)/4, ne11, ne12*ne13) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
+                            }
+                            else if (src0t == GGML_TYPE_Q5_K) {
+                                [encoder dispatchThreadgroups:MTLSizeMake((ne01 + 3)/4, ne11, ne12*ne13) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
+                            }
+                            else if (src0t == GGML_TYPE_Q6_K) {
+                                [encoder dispatchThreadgroups:MTLSizeMake((ne01 + 1)/2, ne11, ne12*ne13) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
+                            } else {
+                                const int64_t ny = (ne11 + nrows - 1)/nrows;
+                                [encoder dispatchThreadgroups:MTLSizeMake(ne01, ny, ne12*ne13) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
+                            }
+                        }
+            } break;
+        case GGML_OP_MUL_MAT_ID:
+            {
+                const int n_as = src0->ne[2];
+
+                // src2 = ids
+                const enum ggml_type src2t = src2->type; GGML_UNUSED(src2t);
+
+                GGML_ASSERT(src2t == GGML_TYPE_I32);
+
+                GGML_ASSERT(!ggml_is_transposed(src0));
+                GGML_ASSERT(!ggml_is_transposed(src1));
+
+                GGML_ASSERT(src1t == GGML_TYPE_F32);
+
+                // find the break-even point where the matrix-matrix kernel becomes more efficient compared
+                // to the matrix-vector kernel
+                // ne20 = n_used_experts
+                // ne21 = n_rows
+                const int dst_rows = ne20*ne21;
+                const int dst_rows_min = n_as;
+                const int dst_rows_max = (ctx->device.maxThreadgroupMemoryLength - 32 - 8192)/4;
+
+                // max size of the rowids array in the kernel shared buffer
+                GGML_ASSERT(dst_rows <= dst_rows_max);
+
+                // for now the matrix-matrix multiplication kernel only works on A14+/M1+ SoCs
+                // AMD GPU and older A-chips will reuse matrix-vector multiplication kernel
+                // !!!
+                // TODO: for now, always use mat-vec kernels until we figure out how to improve the
+                //       indirect matrix multiplication
+                // !!!
+                if ([ctx->device supportsFamily:MTLGPUFamilyApple7] &&
+                        ne00 % 32 == 0 && ne00 >= 64 &&
+                        dst_rows > dst_rows_min) {
+
+                    // some Metal matrix data types require aligned pointers
+                    // ref: https://developer.apple.com/metal/Metal-Shading-Language-Specification.pdf (Table 2.5)
+                    switch (src0->type) {
+                        case GGML_TYPE_F32: GGML_ASSERT(nb01 % 16 == 0); break;
+                        case GGML_TYPE_F16: GGML_ASSERT(nb01 % 8  == 0); break;
+                        default: break;
+                    }
+
+                    id<MTLComputePipelineState> pipeline = nil;
+
+                    switch (src0->type) {
+                        case GGML_TYPE_F32:     pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_ID_F32_F32    ].pipeline; break;
+                        case GGML_TYPE_F16:     pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_ID_F16_F32    ].pipeline; break;
+                        case GGML_TYPE_Q4_0:    pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q4_0_F32   ].pipeline; break;
+                        case GGML_TYPE_Q4_1:    pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q4_1_F32   ].pipeline; break;
+                        case GGML_TYPE_Q5_0:    pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q5_0_F32   ].pipeline; break;
+                        case GGML_TYPE_Q5_1:    pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q5_1_F32   ].pipeline; break;
+                        case GGML_TYPE_Q8_0:    pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q8_0_F32   ].pipeline; break;
+                        case GGML_TYPE_Q2_K:    pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q2_K_F32   ].pipeline; break;
+                        case GGML_TYPE_Q3_K:    pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q3_K_F32   ].pipeline; break;
+                        case GGML_TYPE_Q4_K:    pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q4_K_F32   ].pipeline; break;
+                        case GGML_TYPE_Q5_K:    pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q5_K_F32   ].pipeline; break;
+                        case GGML_TYPE_Q6_K:    pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q6_K_F32   ].pipeline; break;
+                        case GGML_TYPE_IQ2_XXS: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ2_XXS_F32].pipeline; break;
+                        case GGML_TYPE_IQ2_XS:  pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ2_XS_F32 ].pipeline; break;
+                        case GGML_TYPE_IQ3_XXS: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ3_XXS_F32].pipeline; break;
+                        case GGML_TYPE_IQ3_S:   pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ3_S_F32  ].pipeline; break;
+                        case GGML_TYPE_IQ2_S:   pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ2_S_F32  ].pipeline; break;
+                        case GGML_TYPE_IQ1_S:   pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ1_S_F32  ].pipeline; break;
+                        case GGML_TYPE_IQ1_M:   pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ1_M_F32  ].pipeline; break;
+                        case GGML_TYPE_IQ4_NL:  pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ4_NL_F32 ].pipeline; break;
+                        case GGML_TYPE_IQ4_XS:  pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ4_XS_F32 ].pipeline; break;
+                        default: GGML_ABORT("MUL_MAT_ID not implemented");
+                    }
+
+                    [encoder setComputePipelineState:pipeline];
+                    [encoder setBuffer:id_src0 offset:offs_src0    atIndex:0];
+                    [encoder setBuffer:id_src1 offset:offs_src1    atIndex:1];
+                    [encoder setBuffer:id_dst  offset:offs_dst     atIndex:2];
+                    [encoder setBuffer:id_src2 offset:offs_src2    atIndex:3];
+                    [encoder setBytes:&ne20    length:sizeof(ne20) atIndex:4];
+                    [encoder setBytes:&ne21    length:sizeof(ne21) atIndex:5];
+                    [encoder setBytes:&nb21    length:sizeof(nb21) atIndex:6];
+                    [encoder setBytes:&ne00    length:sizeof(ne00) atIndex:7];
+                    [encoder setBytes:&ne02    length:sizeof(ne02) atIndex:8];
+                    [encoder setBytes:&nb01    length:sizeof(nb01) atIndex:9];
+                    [encoder setBytes:&nb02    length:sizeof(nb02) atIndex:10];
+                    [encoder setBytes:&ne11    length:sizeof(ne11) atIndex:11];
+                    [encoder setBytes:&ne12    length:sizeof(ne12) atIndex:12];
+                    [encoder setBytes:&ne13    length:sizeof(ne13) atIndex:13];
+                    [encoder setBytes:&nb10    length:sizeof(nb10) atIndex:14];
+                    [encoder setBytes:&nb11    length:sizeof(nb11) atIndex:15];
+                    [encoder setBytes:&nb12    length:sizeof(nb12) atIndex:16];
+                    [encoder setBytes:&ne0     length:sizeof(ne0)  atIndex:17];
+                    [encoder setBytes:&ne1     length:sizeof(ne1)  atIndex:18];
+                    [encoder setBytes:&nb1     length:sizeof(nb1)  atIndex:19];
+
+                    [encoder setThreadgroupMemoryLength:GGML_PAD(8192 + dst_rows*4/*sizeof(ushort2)*/, 16) atIndex:0];
+
+                    [encoder dispatchThreadgroups:MTLSizeMake((ne21 + 31)/32, (ne01 + 63)/64, n_as) threadsPerThreadgroup:MTLSizeMake(128, 1, 1)];
+                } else {
+                    int nth0 = 32;
+                    int nth1 = 1;
+                    int nrows = 1;
+                    //printf("vector: ne00 = %6d, ne01 = %6d, ne02 = %6d, ne11 = %6d, ne12 = %6d\n", ne00, ne01, ne02, ne11, ne12);
+
+                    id<MTLComputePipelineState> pipeline = nil;
+
+                    // use custom matrix x vector kernel
+                    switch (src0t) {
+                        case GGML_TYPE_F32:
+                            {
+                                GGML_ASSERT(src1t == GGML_TYPE_F32);
+                                pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_ID_F32_F32].pipeline;
+                            } break;
+                        case GGML_TYPE_F16:
+                            {
+                                GGML_ASSERT(src1t == GGML_TYPE_F32);
+                                nth0 = 32;
+                                nth1 = 1;
+                                pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_ID_F16_F32].pipeline;
+                            } break;
+                        case GGML_TYPE_Q4_0:
+                            {
+                                nth0 = 8;
+                                nth1 = 8;
+                                pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_ID_Q4_0_F32].pipeline;
+                            } break;
+                        case GGML_TYPE_Q4_1:
+                            {
+                                nth0 = 8;
+                                nth1 = 8;
+                                pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_ID_Q4_1_F32].pipeline;
+                            } break;
+                        case GGML_TYPE_Q5_0:
+                            {
+                                nth0 = 8;
+                                nth1 = 8;
+                                pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_ID_Q5_0_F32].pipeline;
+                            } break;
+                        case GGML_TYPE_Q5_1:
+                            {
+                                nth0 = 8;
+                                nth1 = 8;
+                                pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_ID_Q5_1_F32].pipeline;
+                            } break;
+                        case GGML_TYPE_Q8_0:
+                            {
+                                nth0 = 8;
+                                nth1 = 8;
+                                pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_ID_Q8_0_F32].pipeline;
+                            } break;
+                        case GGML_TYPE_Q2_K:
+                            {
+                                nth0 = 2;
+                                nth1 = 32;
+                                pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_ID_Q2_K_F32].pipeline;
+                            } break;
+                        case GGML_TYPE_Q3_K:
+                            {
+                                nth0 = 2;
+                                nth1 = 32;
+                                pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_ID_Q3_K_F32].pipeline;
+                            } break;
+                        case GGML_TYPE_Q4_K:
+                            {
+                                nth0 = 4; //1;
+                                nth1 = 8; //32;
+                                pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_ID_Q4_K_F32].pipeline;
+                            } break;
+                        case GGML_TYPE_Q5_K:
+                            {
+                                nth0 = 2;
+                                nth1 = 32;
+                                pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_ID_Q5_K_F32].pipeline;
+                            } break;
+                        case GGML_TYPE_Q6_K:
+                            {
+                                nth0 = 2;
+                                nth1 = 32;
+                                pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_ID_Q6_K_F32].pipeline;
+                            } break;
+                        case GGML_TYPE_IQ2_XXS:
+                            {
+                                nth0 = 4;
+                                nth1 = 16;
+                                pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_ID_IQ2_XXS_F32].pipeline;
+                            } break;
+                        case GGML_TYPE_IQ2_XS:
+                            {
+                                nth0 = 4;
+                                nth1 = 16;
+                                pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_ID_IQ2_XS_F32].pipeline;
+                            } break;
+                        case GGML_TYPE_IQ3_XXS:
+                            {
+                                nth0 = 4;
+                                nth1 = 16;
+                                pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_ID_IQ3_XXS_F32].pipeline;
+                            } break;
+                        case GGML_TYPE_IQ3_S:
+                            {
+                                nth0 = 4;
+                                nth1 = 16;
+                                pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_ID_IQ3_S_F32].pipeline;
+                            } break;
+                        case GGML_TYPE_IQ2_S:
+                            {
+                                nth0 = 4;
+                                nth1 = 16;
+                                pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_ID_IQ2_S_F32].pipeline;
+                            } break;
+                        case GGML_TYPE_IQ1_S:
+                            {
+                                nth0 = 4;
+                                nth1 = 16;
+                                pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_ID_IQ1_S_F32].pipeline;
+                            } break;
+                        case GGML_TYPE_IQ1_M:
+                            {
+                                nth0 = 4;
+                                nth1 = 16;
+                                pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_ID_IQ1_M_F32].pipeline;
+                            } break;
+                        case GGML_TYPE_IQ4_NL:
+                            {
+                                nth0 = 4;
+                                nth1 = 16;
+                                pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_ID_IQ4_NL_F32].pipeline;
+                            } break;
+                        case GGML_TYPE_IQ4_XS:
+                            {
+                                nth0 = 4;
+                                nth1 = 16;
+                                pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_ID_IQ4_XS_F32].pipeline;
+                            } break;
+                        default:
+                            {
+                                GGML_LOG_ERROR("Asserting on type %d\n", (int)src2t);
+                                GGML_ABORT("not implemented");
+                            }
+                    };
+
+                    if (ggml_is_quantized(src0t)) {
+                        GGML_ASSERT(ne00 >= nth0*nth1);
+                    }
+
+                    [encoder setComputePipelineState:pipeline];
+                    [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
+                    [encoder setBuffer:id_src1 offset:offs_src1 atIndex:1];
+                    [encoder setBuffer:id_dst  offset:offs_dst  atIndex:2];
+                    [encoder setBuffer:id_src2 offset:offs_src2 atIndex:3];
+                    [encoder setBytes:&ne20 length:sizeof(ne20) atIndex:4];
+                    [encoder setBytes:&ne21 length:sizeof(ne21) atIndex:5];
+                    [encoder setBytes:&nb21 length:sizeof(nb21) atIndex:6];
+                    [encoder setBytes:&ne00 length:sizeof(ne00) atIndex:7];
+                    [encoder setBytes:&ne01 length:sizeof(ne01) atIndex:8];
+                    [encoder setBytes:&ne02 length:sizeof(ne02) atIndex:9];
+                    [encoder setBytes:&nb00 length:sizeof(nb00) atIndex:10];
+                    [encoder setBytes:&nb01 length:sizeof(nb01) atIndex:11];
+                    [encoder setBytes:&nb02 length:sizeof(nb02) atIndex:12];
+                    [encoder setBytes:&ne10 length:sizeof(ne10) atIndex:13];
+                    [encoder setBytes:&ne11 length:sizeof(ne11) atIndex:14];
+                    [encoder setBytes:&ne12 length:sizeof(ne12) atIndex:15];
+                    [encoder setBytes:&ne13 length:sizeof(ne13) atIndex:16];
+                    [encoder setBytes:&nb10 length:sizeof(nb10) atIndex:17];
+                    [encoder setBytes:&nb11 length:sizeof(nb11) atIndex:18];
+                    [encoder setBytes:&nb12 length:sizeof(nb12) atIndex:19];
+                    [encoder setBytes:&ne0  length:sizeof(ne0)  atIndex:20];
+                    [encoder setBytes:&ne1  length:sizeof(ne1)  atIndex:21];
+                    [encoder setBytes:&nb1  length:sizeof(nb1)  atIndex:22];
+
+                    const int64_t _ne1 = 1;
+                    const int tgz = dst_rows;
+
+                    if (src0t == GGML_TYPE_Q4_0  || src0t == GGML_TYPE_Q4_1  || src0t == GGML_TYPE_Q5_0 ||
+                            src0t == GGML_TYPE_Q5_1  || src0t == GGML_TYPE_Q8_0  || src0t == GGML_TYPE_Q2_K ||
+                            src0t == GGML_TYPE_IQ1_S || src0t == GGML_TYPE_IQ1_M || src0t == GGML_TYPE_IQ2_S) {
+                        [encoder dispatchThreadgroups:MTLSizeMake((ne01 + 7)/8, _ne1, tgz) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
+                    }
+                    else if (src0t == GGML_TYPE_IQ2_XXS || src0t == GGML_TYPE_IQ2_XS) {
+                        const int mem_size = src0t == GGML_TYPE_IQ2_XXS ? 256*8+128 : 512*8+128;
+                        [encoder setThreadgroupMemoryLength:mem_size atIndex:0];
+                        [encoder dispatchThreadgroups:MTLSizeMake((ne01 + 7)/8, _ne1, tgz) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
+                    }
+                    else if (src0t == GGML_TYPE_IQ3_XXS || src0t == GGML_TYPE_IQ3_S) {
+                        const int mem_size = src0t == GGML_TYPE_IQ3_XXS ? 256*4+128 : 512*4;
+                        [encoder setThreadgroupMemoryLength:mem_size atIndex:0];
+                        [encoder dispatchThreadgroups:MTLSizeMake((ne01 + 7)/8, _ne1, tgz) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
+                    }
+                    else if (src0t == GGML_TYPE_IQ4_NL || src0t == GGML_TYPE_IQ4_XS) {
+                        const int mem_size = 32*sizeof(float);
+                        [encoder setThreadgroupMemoryLength:mem_size atIndex:0];
+                        [encoder dispatchThreadgroups:MTLSizeMake((ne01 + 3)/4, _ne1, tgz) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
+                    }
+                    else if (src0t == GGML_TYPE_Q4_K) {
+                        [encoder dispatchThreadgroups:MTLSizeMake((ne01 + 3)/4, _ne1, tgz) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
+                    }
+                    else if (src0t == GGML_TYPE_Q3_K) {
+                        [encoder dispatchThreadgroups:MTLSizeMake((ne01 + 3)/4, _ne1, tgz) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
+                    }
+                    else if (src0t == GGML_TYPE_Q5_K) {
+                        [encoder dispatchThreadgroups:MTLSizeMake((ne01 + 3)/4, _ne1, tgz) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
+                    }
+                    else if (src0t == GGML_TYPE_Q6_K) {
+                        [encoder dispatchThreadgroups:MTLSizeMake((ne01 + 1)/2, _ne1, tgz) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
+                    } else {
+                        const int64_t ny = (_ne1 + nrows - 1)/nrows; // = _ne1
+                        [encoder dispatchThreadgroups:MTLSizeMake(ne01, ny, tgz) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
+                    }
+                }
+            } break;
+        case GGML_OP_GET_ROWS:
+            {
+                id<MTLComputePipelineState> pipeline = nil;
+
+                switch (src0->type) {
+                    case GGML_TYPE_F32:     pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_GET_ROWS_F32    ].pipeline; break;
+                    case GGML_TYPE_F16:     pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_GET_ROWS_F16    ].pipeline; break;
+                    case GGML_TYPE_Q4_0:    pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_GET_ROWS_Q4_0   ].pipeline; break;
+                    case GGML_TYPE_Q4_1:    pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_GET_ROWS_Q4_1   ].pipeline; break;
+                    case GGML_TYPE_Q5_0:    pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_GET_ROWS_Q5_0   ].pipeline; break;
+                    case GGML_TYPE_Q5_1:    pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_GET_ROWS_Q5_1   ].pipeline; break;
+                    case GGML_TYPE_Q8_0:    pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_GET_ROWS_Q8_0   ].pipeline; break;
+                    case GGML_TYPE_Q2_K:    pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_GET_ROWS_Q2_K   ].pipeline; break;
+                    case GGML_TYPE_Q3_K:    pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_GET_ROWS_Q3_K   ].pipeline; break;
+                    case GGML_TYPE_Q4_K:    pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_GET_ROWS_Q4_K   ].pipeline; break;
+                    case GGML_TYPE_Q5_K:    pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_GET_ROWS_Q5_K   ].pipeline; break;
+                    case GGML_TYPE_Q6_K:    pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_GET_ROWS_Q6_K   ].pipeline; break;
+                    case GGML_TYPE_IQ2_XXS: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_GET_ROWS_IQ2_XXS].pipeline; break;
+                    case GGML_TYPE_IQ2_XS:  pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_GET_ROWS_IQ2_XS ].pipeline; break;
+                    case GGML_TYPE_IQ3_XXS: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_GET_ROWS_IQ3_XXS].pipeline; break;
+                    case GGML_TYPE_IQ3_S:   pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_GET_ROWS_IQ3_S  ].pipeline; break;
+                    case GGML_TYPE_IQ2_S:   pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_GET_ROWS_IQ2_S  ].pipeline; break;
+                    case GGML_TYPE_IQ1_S:   pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_GET_ROWS_IQ1_S  ].pipeline; break;
+                    case GGML_TYPE_IQ1_M:   pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_GET_ROWS_IQ1_M  ].pipeline; break;
+                    case GGML_TYPE_IQ4_NL:  pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_GET_ROWS_IQ4_NL ].pipeline; break;
+                    case GGML_TYPE_IQ4_XS:  pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_GET_ROWS_IQ4_XS ].pipeline; break;
+                    case GGML_TYPE_I32:     pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_GET_ROWS_I32    ].pipeline; break;
+                    default: GGML_ABORT("not implemented");
+                }
+
+                [encoder setComputePipelineState:pipeline];
+                [encoder setBuffer:id_src0     offset:offs_src0 atIndex:0];
+                [encoder setBuffer:id_src1     offset:offs_src1 atIndex:1];
+                [encoder setBuffer:id_dst      offset:offs_dst  atIndex:2];
+                [encoder setBytes:&ne00 length:sizeof( int64_t) atIndex:3];
+                [encoder setBytes:&nb01 length:sizeof(uint64_t) atIndex:4];
+                [encoder setBytes:&nb02 length:sizeof(uint64_t) atIndex:5];
+                [encoder setBytes:&ne10 length:sizeof( int64_t) atIndex:6];
+                [encoder setBytes:&nb10 length:sizeof( int64_t) atIndex:7];
+                [encoder setBytes:&nb11 length:sizeof( int64_t) atIndex:8];
+                [encoder setBytes:&nb1  length:sizeof(uint64_t) atIndex:9];
+                [encoder setBytes:&nb2  length:sizeof(uint64_t) atIndex:10];
+
+                [encoder dispatchThreadgroups:MTLSizeMake(ne10, ne11, 1) threadsPerThreadgroup:MTLSizeMake(32, 1, 1)];
+            } break;
+        case GGML_OP_RMS_NORM:
+            {
+                GGML_ASSERT(ne00 % 4 == 0);
+                GGML_ASSERT(ggml_is_contiguous_1(src0));
+
+                float eps;
+                memcpy(&eps, dst->op_params, sizeof(float));
+
+                int nth = 32; // SIMD width
+
+                while (nth < ne00/4 && nth < 1024) {
+                    nth *= 2;
+                }
+
+                id<MTLComputePipelineState> pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_RMS_NORM].pipeline;
+
+                [encoder setComputePipelineState:pipeline];
+                [encoder setBuffer:id_src0 offset:offs_src0        atIndex:0];
+                [encoder setBuffer:id_dst  offset:offs_dst         atIndex:1];
+                [encoder setBytes:&ne00    length:sizeof( int64_t) atIndex:2];
+                [encoder setBytes:&nb01    length:sizeof(uint64_t) atIndex:3];
+                [encoder setBytes:&eps     length:sizeof(   float) atIndex:4];
+                [encoder setThreadgroupMemoryLength:32*sizeof(float) atIndex:0];
+
+                const int64_t nrows = ggml_nrows(src0);
+
+                [encoder dispatchThreadgroups:MTLSizeMake(nrows, 1, 1) threadsPerThreadgroup:MTLSizeMake(nth, 1, 1)];
+            } break;
+        case GGML_OP_GROUP_NORM:
+            {
+                GGML_ASSERT(ne00 % 4 == 0);
+                GGML_ASSERT(ggml_is_contiguous(src0));
+
+                float eps;
+                memcpy(&eps, dst->op_params + 1, sizeof(float));
+
+                const int32_t n_groups = ((const int32_t *) dst->op_params)[0];
+
+                int nth = 32; // SIMD width
+
+                //while (nth < ne00/4 && nth < 1024) {
+                //    nth *= 2;
+                //}
+
+                id<MTLComputePipelineState> pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_GROUP_NORM].pipeline;
+
+                [encoder setComputePipelineState:pipeline];
+                [encoder setBuffer:id_src0  offset:offs_src0        atIndex:0];
+                [encoder setBuffer:id_dst   offset:offs_dst         atIndex:1];
+                [encoder setBytes:&ne00     length:sizeof( int64_t) atIndex:2];
+                [encoder setBytes:&ne01     length:sizeof( int64_t) atIndex:3];
+                [encoder setBytes:&ne02     length:sizeof( int64_t) atIndex:4];
+                [encoder setBytes:&nb00     length:sizeof(uint64_t) atIndex:5];
+                [encoder setBytes:&nb01     length:sizeof(uint64_t) atIndex:6];
+                [encoder setBytes:&nb02     length:sizeof(uint64_t) atIndex:7];
+                [encoder setBytes:&n_groups length:sizeof( int32_t) atIndex:8];
+                [encoder setBytes:&eps      length:sizeof(   float) atIndex:9];
+                [encoder setThreadgroupMemoryLength:32*sizeof(float) atIndex:0];
+
+                [encoder dispatchThreadgroups:MTLSizeMake(n_groups, 1, 1) threadsPerThreadgroup:MTLSizeMake(nth, 1, 1)];
+            } break;
+        case GGML_OP_NORM:
+            {
+                GGML_ASSERT(ggml_is_contiguous_1(src0));
+
+                float eps;
+                memcpy(&eps, dst->op_params, sizeof(float));
+
+                const int nth = MIN(256, ne00);
+
+                id<MTLComputePipelineState> pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_NORM].pipeline;
+
+                [encoder setComputePipelineState:pipeline];
+                [encoder setBuffer:id_src0 offset:offs_src0        atIndex:0];
+                [encoder setBuffer:id_dst  offset:offs_dst         atIndex:1];
+                [encoder setBytes:&ne00    length:sizeof( int64_t) atIndex:2];
+                [encoder setBytes:&nb01    length:sizeof(uint64_t) atIndex:3];
+                [encoder setBytes:&eps     length:sizeof(   float) atIndex:4];
+                [encoder setThreadgroupMemoryLength:GGML_PAD(nth*sizeof(float), 16) atIndex:0];
+
+                const int64_t nrows = ggml_nrows(src0);
+
+                [encoder dispatchThreadgroups:MTLSizeMake(nrows, 1, 1) threadsPerThreadgroup:MTLSizeMake(nth, 1, 1)];
+            } break;
+        case GGML_OP_ROPE:
+            {
+                GGML_ASSERT(ne10 == ne02);
+
+                const int nth = MIN(1024, ne00);
+
+                const int n_past     = ((const int32_t *) dst->op_params)[0];
+                const int n_dims     = ((const int32_t *) dst->op_params)[1];
+                const int mode       = ((const int32_t *) dst->op_params)[2];
+                // skip 3, n_ctx, used in GLM RoPE, unimplemented in metal
+                const int n_ctx_orig = ((const int32_t *) dst->op_params)[4];
+
+                float freq_base;
+                float freq_scale;
+                float ext_factor;
+                float attn_factor;
+                float beta_fast;
+                float beta_slow;
+
+                memcpy(&freq_base,   (const int32_t *) dst->op_params +  5, sizeof(float));
+                memcpy(&freq_scale,  (const int32_t *) dst->op_params +  6, sizeof(float));
+                memcpy(&ext_factor,  (const int32_t *) dst->op_params +  7, sizeof(float));
+                memcpy(&attn_factor, (const int32_t *) dst->op_params +  8, sizeof(float));
+                memcpy(&beta_fast,   (const int32_t *) dst->op_params +  9, sizeof(float));
+                memcpy(&beta_slow,   (const int32_t *) dst->op_params + 10, sizeof(float));
+
+                const bool is_neox = mode & GGML_ROPE_TYPE_NEOX;
+
+                id<MTLComputePipelineState> pipeline = nil;
+
+                if (!is_neox) {
+                    switch (src0->type) {
+                        case GGML_TYPE_F32: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_ROPE_NORM_F32].pipeline; break;
+                        case GGML_TYPE_F16: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_ROPE_NORM_F16].pipeline; break;
+                        default: GGML_ABORT("fatal error");
+                    };
+                } else {
+                    switch (src0->type) {
+                        case GGML_TYPE_F32: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_ROPE_NEOX_F32].pipeline; break;
+                        case GGML_TYPE_F16: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_ROPE_NEOX_F16].pipeline; break;
+                        default: GGML_ABORT("fatal error");
+                    };
+                }
+
+                [encoder setComputePipelineState:pipeline];
+                [encoder setBuffer:id_src0     offset:offs_src0        atIndex:0];
+                [encoder setBuffer:id_src1     offset:offs_src1        atIndex:1];
+                if (id_src2 != nil) {
+                    [encoder setBuffer:id_src2 offset:offs_src2        atIndex:2];
+                } else {
+                    [encoder setBuffer:id_src0 offset:offs_src0        atIndex:2];
+                }
+                [encoder setBuffer:id_dst      offset:offs_dst         atIndex:3];
+                [encoder setBytes:&ne00        length:sizeof( int64_t) atIndex:4];
+                [encoder setBytes:&ne01        length:sizeof( int64_t) atIndex:5];
+                [encoder setBytes:&ne02        length:sizeof( int64_t) atIndex:6];
+                [encoder setBytes:&ne03        length:sizeof( int64_t) atIndex:7];
+                [encoder setBytes:&nb00        length:sizeof(uint64_t) atIndex:8];
+                [encoder setBytes:&nb01        length:sizeof(uint64_t) atIndex:9];
+                [encoder setBytes:&nb02        length:sizeof(uint64_t) atIndex:10];
+                [encoder setBytes:&nb03        length:sizeof(uint64_t) atIndex:11];
+                [encoder setBytes:&ne0         length:sizeof( int64_t) atIndex:12];
+                [encoder setBytes:&ne1         length:sizeof( int64_t) atIndex:13];
+                [encoder setBytes:&ne2         length:sizeof( int64_t) atIndex:14];
+                [encoder setBytes:&ne3         length:sizeof( int64_t) atIndex:15];
+                [encoder setBytes:&nb0         length:sizeof(uint64_t) atIndex:16];
+                [encoder setBytes:&nb1         length:sizeof(uint64_t) atIndex:17];
+                [encoder setBytes:&nb2         length:sizeof(uint64_t) atIndex:18];
+                [encoder setBytes:&nb3         length:sizeof(uint64_t) atIndex:19];
+                [encoder setBytes:&n_past      length:sizeof(     int) atIndex:20];
+                [encoder setBytes:&n_dims      length:sizeof(     int) atIndex:21];
+                [encoder setBytes:&n_ctx_orig  length:sizeof(     int) atIndex:22];
+                [encoder setBytes:&freq_base   length:sizeof(   float) atIndex:23];
+                [encoder setBytes:&freq_scale  length:sizeof(   float) atIndex:24];
+                [encoder setBytes:&ext_factor  length:sizeof(   float) atIndex:25];
+                [encoder setBytes:&attn_factor length:sizeof(   float) atIndex:26];
+                [encoder setBytes:&beta_fast   length:sizeof(   float) atIndex:27];
+                [encoder setBytes:&beta_slow   length:sizeof(   float) atIndex:28];
+
+                [encoder dispatchThreadgroups:MTLSizeMake(ne01, ne02, ne03) threadsPerThreadgroup:MTLSizeMake(nth, 1, 1)];
+            } break;
+        case GGML_OP_IM2COL:
+            {
+                GGML_ASSERT(src0->type == GGML_TYPE_F16);
+                GGML_ASSERT(src1->type == GGML_TYPE_F32);
+                GGML_ASSERT( dst->type == GGML_TYPE_F16 || dst->type == GGML_TYPE_F32);
+
+                const int32_t s0 = ((const int32_t *)(dst->op_params))[0];
+                const int32_t s1 = ((const int32_t *)(dst->op_params))[1];
+                const int32_t p0 = ((const int32_t *)(dst->op_params))[2];
+                const int32_t p1 = ((const int32_t *)(dst->op_params))[3];
+                const int32_t d0 = ((const int32_t *)(dst->op_params))[4];
+                const int32_t d1 = ((const int32_t *)(dst->op_params))[5];
+
+                const bool is_2D = ((const int32_t *)(dst->op_params))[6] == 1;
+
+                const int32_t N  = src1->ne[is_2D ? 3 : 2];
+                const int32_t IC = src1->ne[is_2D ? 2 : 1];
+                const int32_t IH = is_2D ? src1->ne[1] : 1;
+                const int32_t IW =         src1->ne[0];
+
+                const int32_t KH = is_2D ? src0->ne[1] : 1;
+                const int32_t KW =         src0->ne[0];
+
+                const int32_t OH = is_2D ? dst->ne[2] : 1;
+                const int32_t OW =         dst->ne[1];
+
+                const int32_t CHW = IC * KH * KW;
+
+                const int32_t ofs0 = src1->nb[is_2D ? 3 : 2] / 4;
+                const int32_t ofs1 = src1->nb[is_2D ? 2 : 1] / 4;
+
+                id<MTLComputePipelineState> pipeline = nil;
+
+                switch (dst->type) {
+                    case GGML_TYPE_F32: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_IM2COL_F32].pipeline; break;
+                    case GGML_TYPE_F16: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_IM2COL_F16].pipeline; break;
+                    default: GGML_ABORT("fatal error");
+                };
+
+                [encoder setComputePipelineState:pipeline];
+                [encoder setBuffer:id_src1 offset:offs_src1        atIndex:0];
+                [encoder setBuffer:id_dst  offset:offs_dst         atIndex:1];
+                [encoder setBytes:&ofs0    length:sizeof( int32_t) atIndex:2];
+                [encoder setBytes:&ofs1    length:sizeof( int32_t) atIndex:3];
+                [encoder setBytes:&IW      length:sizeof( int32_t) atIndex:4];
+                [encoder setBytes:&IH      length:sizeof( int32_t) atIndex:5];
+                [encoder setBytes:&CHW     length:sizeof( int32_t) atIndex:6];
+                [encoder setBytes:&s0      length:sizeof( int32_t) atIndex:7];
+                [encoder setBytes:&s1      length:sizeof( int32_t) atIndex:8];
+                [encoder setBytes:&p0      length:sizeof( int32_t) atIndex:9];
+                [encoder setBytes:&p1      length:sizeof( int32_t) atIndex:10];
+                [encoder setBytes:&d0      length:sizeof( int32_t) atIndex:11];
+                [encoder setBytes:&d1      length:sizeof( int32_t) atIndex:12];
+
+                [encoder dispatchThreadgroups:MTLSizeMake(IC, OH, OW) threadsPerThreadgroup:MTLSizeMake(N, KH, KW)];
+            } break;
+        case GGML_OP_UPSCALE:
+            {
+                GGML_ASSERT(src0->type == GGML_TYPE_F32);
+
+                const float sf0 = (float)ne0/src0->ne[0];
+                const float sf1 = (float)ne1/src0->ne[1];
+                const float sf2 = (float)ne2/src0->ne[2];
+                const float sf3 = (float)ne3/src0->ne[3];
+
+                const id<MTLComputePipelineState> pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_UPSCALE_F32].pipeline;
+
+                [encoder setComputePipelineState:pipeline];
+                [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
+                [encoder setBuffer:id_dst  offset:offs_dst  atIndex:1];
+                [encoder setBytes:&ne00 length:sizeof(ne00) atIndex:2];
+                [encoder setBytes:&ne01 length:sizeof(ne01) atIndex:3];
+                [encoder setBytes:&ne02 length:sizeof(ne02) atIndex:4];
+                [encoder setBytes:&ne03 length:sizeof(ne03) atIndex:5];
+                [encoder setBytes:&nb00 length:sizeof(nb00) atIndex:6];
+                [encoder setBytes:&nb01 length:sizeof(nb01) atIndex:7];
+                [encoder setBytes:&nb02 length:sizeof(nb02) atIndex:8];
+                [encoder setBytes:&nb03 length:sizeof(nb03) atIndex:9];
+                [encoder setBytes:&ne0  length:sizeof(ne0)  atIndex:10];
+                [encoder setBytes:&ne1  length:sizeof(ne1)  atIndex:11];
+                [encoder setBytes:&ne2  length:sizeof(ne2)  atIndex:12];
+                [encoder setBytes:&ne3  length:sizeof(ne3)  atIndex:13];
+                [encoder setBytes:&nb0  length:sizeof(nb0)  atIndex:14];
+                [encoder setBytes:&nb1  length:sizeof(nb1)  atIndex:15];
+                [encoder setBytes:&nb2  length:sizeof(nb2)  atIndex:16];
+                [encoder setBytes:&nb3  length:sizeof(nb3)  atIndex:17];
+                [encoder setBytes:&sf0  length:sizeof(sf0)  atIndex:18];
+                [encoder setBytes:&sf1  length:sizeof(sf1)  atIndex:19];
+                [encoder setBytes:&sf2  length:sizeof(sf2)  atIndex:20];
+                [encoder setBytes:&sf3  length:sizeof(sf3)  atIndex:21];
+
+                const int nth = MIN((int) pipeline.maxTotalThreadsPerThreadgroup, ne0);
+
+                [encoder dispatchThreadgroups:MTLSizeMake(ne1, ne2, ne3) threadsPerThreadgroup:MTLSizeMake(nth, 1, 1)];
+            } break;
+        case GGML_OP_PAD:
+            {
+                GGML_ASSERT(src0->type == GGML_TYPE_F32);
+
+                id<MTLComputePipelineState> pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_PAD_F32].pipeline;
+
+                [encoder setComputePipelineState:pipeline];
+                [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
+                [encoder setBuffer:id_dst  offset:offs_dst  atIndex:1];
+                [encoder setBytes:&ne00 length:sizeof(ne00) atIndex:2];
+                [encoder setBytes:&ne01 length:sizeof(ne01) atIndex:3];
+                [encoder setBytes:&ne02 length:sizeof(ne02) atIndex:4];
+                [encoder setBytes:&ne03 length:sizeof(ne03) atIndex:5];
+                [encoder setBytes:&nb00 length:sizeof(nb00) atIndex:6];
+                [encoder setBytes:&nb01 length:sizeof(nb01) atIndex:7];
+                [encoder setBytes:&nb02 length:sizeof(nb02) atIndex:8];
+                [encoder setBytes:&nb03 length:sizeof(nb03) atIndex:9];
+                [encoder setBytes:&ne0  length:sizeof(ne0)  atIndex:10];
+                [encoder setBytes:&ne1  length:sizeof(ne1)  atIndex:11];
+                [encoder setBytes:&ne2  length:sizeof(ne2)  atIndex:12];
+                [encoder setBytes:&ne3  length:sizeof(ne3)  atIndex:13];
+                [encoder setBytes:&nb0  length:sizeof(nb0)  atIndex:14];
+                [encoder setBytes:&nb1  length:sizeof(nb1)  atIndex:15];
+                [encoder setBytes:&nb2  length:sizeof(nb2)  atIndex:16];
+                [encoder setBytes:&nb3  length:sizeof(nb3)  atIndex:17];
+
+                const int nth = MIN(1024, ne0);
+
+                [encoder dispatchThreadgroups:MTLSizeMake(ne1, ne2, ne3) threadsPerThreadgroup:MTLSizeMake(nth, 1, 1)];
+            } break;
+        case GGML_OP_ARANGE:
+            {
+                GGML_ASSERT(dst->type == GGML_TYPE_F32);
+
+                float start;
+                float step;
+
+                memcpy(&start, ((const int32_t *) dst->op_params) + 0, sizeof(float));
+                memcpy(&step,  ((const int32_t *) dst->op_params) + 2, sizeof(float));
+
+                id<MTLComputePipelineState> pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_ARANGE_F32].pipeline;
+
+                [encoder setComputePipelineState:pipeline];
+                [encoder setBuffer:id_dst  offset:offs_dst    atIndex:0];
+                [encoder setBytes:&ne0   length:sizeof(ne0)   atIndex:1];
+                [encoder setBytes:&start length:sizeof(start) atIndex:2];
+                [encoder setBytes:&step  length:sizeof(step)  atIndex:3];
+
+                const int nth = MIN(1024, ne0);
+
+                [encoder dispatchThreadgroups:MTLSizeMake(1, 1, 1) threadsPerThreadgroup:MTLSizeMake(nth, 1, 1)];
+            } break;
+        case GGML_OP_TIMESTEP_EMBEDDING:
+            {
+                GGML_ASSERT(src0->type == GGML_TYPE_F32);
+
+                const int dim        = dst->op_params[0];
+                const int max_period = dst->op_params[1];
+
+                const int half = dim / 2;
+
+                id<MTLComputePipelineState> pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_TIMESTEP_EMBEDDING_F32].pipeline;
+
+                [encoder setComputePipelineState:pipeline];
+                [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
+                [encoder setBuffer:id_dst  offset:offs_dst  atIndex:1];
+                [encoder setBytes:&nb1   length:sizeof(nb1) atIndex:2];
+                [encoder setBytes:&dim   length:sizeof(dim) atIndex:3];
+                [encoder setBytes:&max_period length:sizeof(max_period) atIndex:4];
+
+                const int nth = MIN(1024, half);
+
+                [encoder dispatchThreadgroups:MTLSizeMake(ne00, 1, 1) threadsPerThreadgroup:MTLSizeMake(nth, 1, 1)];
+            } break;
+        case GGML_OP_ARGSORT:
+            {
+                GGML_ASSERT(src0->type == GGML_TYPE_F32);
+                GGML_ASSERT( dst->type == GGML_TYPE_I32);
+
+                const int nrows = ggml_nrows(src0);
+
+                enum ggml_sort_order order = (enum ggml_sort_order) dst->op_params[0];
+
+                // bitonic sort requires the number of elements to be power of 2
+                int64_t ne00_padded = 1;
+                while (ne00_padded < ne00) {
+                    ne00_padded *= 2;
+                }
+
+                // Metal kernels require the buffer size to be multiple of 16 bytes
+                // https://developer.apple.com/documentation/metal/mtlcomputecommandencoder/1443142-setthreadgroupmemorylength
+                const int mem_size = GGML_PAD(ne00_padded*sizeof(int32_t), 16);
+
+                id<MTLComputePipelineState> pipeline = nil;
+
+                switch (order) {
+                    case GGML_SORT_ORDER_ASC:  pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_ARGSORT_F32_I32_ASC].pipeline;  break;
+                    case GGML_SORT_ORDER_DESC: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_ARGSORT_F32_I32_DESC].pipeline; break;
+                    default: GGML_ABORT("fatal error");
+                };
+
+                [encoder setComputePipelineState:pipeline];
+                [encoder setBuffer:id_src0     offset:offs_src0        atIndex:0];
+                [encoder setBuffer:id_dst      offset:offs_dst         atIndex:1];
+                [encoder setBytes:&ne00        length:sizeof( int64_t) atIndex:2];
+                [encoder setBytes:&ne00_padded length:sizeof( int64_t) atIndex:3];
+                [encoder setThreadgroupMemoryLength:mem_size atIndex:0];
+
+                [encoder dispatchThreadgroups:MTLSizeMake(1, nrows, 1) threadsPerThreadgroup:MTLSizeMake(ne00_padded, 1, 1)];
+            } break;
+        case GGML_OP_LEAKY_RELU:
+            {
+                GGML_ASSERT(src0->type == GGML_TYPE_F32);
+
+                float slope;
+                memcpy(&slope, dst->op_params, sizeof(float));
+
+                id<MTLComputePipelineState> pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_LEAKY_RELU_F32].pipeline;
+
+                [encoder setComputePipelineState:pipeline];
+                [encoder setBuffer:id_src0 offset:offs_src0   atIndex:0];
+                [encoder setBuffer:id_dst  offset:offs_dst    atIndex:1];
+                [encoder setBytes:&slope length:sizeof(slope) atIndex:2];
+
+                const int64_t n = ggml_nelements(dst);
+
+                [encoder dispatchThreadgroups:MTLSizeMake(n, 1, 1) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)];
+            } break;
+        case GGML_OP_FLASH_ATTN_EXT:
+            {
+                GGML_ASSERT(ne00 % 4  == 0);
+                GGML_ASSERT(ne11 % 32 == 0);
+
+                GGML_ASSERT(src0->type == GGML_TYPE_F32);
+
+                GGML_ASSERT(ggml_are_same_shape (src1, src2));
+
+                struct ggml_tensor * src3 = node->src[3];
+
+                size_t offs_src3 = 0;
+
+                id<MTLBuffer> id_src3 = src3 ? ggml_metal_get_buffer(src3, &offs_src3) : nil;
+
+                GGML_ASSERT(!src3 || src3->type == GGML_TYPE_F16);
+                GGML_ASSERT(!src3 || src3->ne[1] >= GGML_PAD(src0->ne[1], 8) &&
+                        "the Flash-Attention Metal kernel requires the mask to be padded to 8 and at least n_queries big");
+
+                const int64_t  ne30 = src3 ? src3->ne[0] : 0; GGML_UNUSED(ne30);
+                //const int64_t  ne31 = src3 ? src3->ne[1] : 0;
+                const int64_t  ne32 = src3 ? src3->ne[2] : 0; GGML_UNUSED(ne32);
+                const int64_t  ne33 = src3 ? src3->ne[3] : 0; GGML_UNUSED(ne33);
+
+                const uint64_t nb30 = src3 ? src3->nb[0] : 0; GGML_UNUSED(nb30);
+                const uint64_t nb31 = src3 ? src3->nb[1] : 0;
+                const uint64_t nb32 = src3 ? src3->nb[2] : 0; GGML_UNUSED(nb32);
+                const uint64_t nb33 = src3 ? src3->nb[3] : 0; GGML_UNUSED(nb33);
+
+                const enum ggml_type src2t = src2 ? src2->type : GGML_TYPE_COUNT; GGML_UNUSED(src2t);
+
+                float scale;
+                float max_bias;
+                float logit_softcap;
+                memcpy(&scale,         ((const int32_t *) dst->op_params) + 0, sizeof(scale));
+                memcpy(&max_bias,      ((const int32_t *) dst->op_params) + 1, sizeof(max_bias));
+                memcpy(&logit_softcap, ((const int32_t *) dst->op_params) + 2, sizeof(logit_softcap));
+
+                if (logit_softcap != 0.0f) {
+                    scale /= logit_softcap;
+                }
+
+                const uint32_t n_head      = src0->ne[2];
+                const uint32_t n_head_log2 = 1u << (uint32_t) floorf(log2f((float) n_head));
+
+                const float m0 = powf(2.0f, -(max_bias       ) / n_head_log2);
+                const float m1 = powf(2.0f, -(max_bias / 2.0f) / n_head_log2);
+
+                id<MTLComputePipelineState> pipeline = nil;
+
+                bool use_vec_kernel = false;
+
+                if (ne01 >= 4 || (ne00%128 != 0)) {
+                    switch (ne00) {
+                        case 64:  pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_F16_H64 ].pipeline; break;
+                        case 80:  pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_F16_H80 ].pipeline; break;
+                        case 96:  pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_F16_H96 ].pipeline; break;
+                        case 112: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_F16_H112].pipeline; break;
+                        case 128: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_F16_H128].pipeline; break;
+                                  //case 256: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_F16_H256].pipeline; break;
+                        default:
+                                  {
+                                      GGML_LOG_ERROR("unsupported size: %lld\n", ne00);
+                                      GGML_LOG_ERROR("add template specialization for this size\n");
+                                      GGML_ABORT("add template specialization for this size");
+                                  }
+                    }
+                } else {
+                    use_vec_kernel = true;
+
+                    switch (ne00) {
+                        case 128: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_F16_H128].pipeline; break;
+                                  //case 256: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_VEC_F16_H256].pipeline; break;
+                        default:
+                                  {
+                                      GGML_LOG_ERROR("unsupported size: %lld\n", ne00);
+                                      GGML_LOG_ERROR("add template specialization for this size\n");
+                                      GGML_ABORT("add template specialization for this size");
+                                  }
+                    }
+                }
+
+                [encoder setComputePipelineState:pipeline];
+                [encoder setBuffer:id_src0     offset:offs_src0           atIndex:0];
+                [encoder setBuffer:id_src1     offset:offs_src1           atIndex:1];
+                [encoder setBuffer:id_src2     offset:offs_src2           atIndex:2];
+                if (id_src3) {
+                    [encoder setBuffer:id_src3     offset:offs_src3           atIndex:3];
+                } else {
+                    [encoder setBuffer:id_src0     offset:offs_src0           atIndex:3];
+                }
+                [encoder setBuffer:id_dst        offset:offs_dst              atIndex:4];
+                [encoder setBytes:&ne01          length:sizeof( int64_t)      atIndex:5];
+                [encoder setBytes:&ne02          length:sizeof( int64_t)      atIndex:6];
+                [encoder setBytes:&ne03          length:sizeof( int64_t)      atIndex:7];
+                [encoder setBytes:&nb01          length:sizeof(uint64_t)      atIndex:8];
+                [encoder setBytes:&nb02          length:sizeof(uint64_t)      atIndex:9];
+                [encoder setBytes:&nb03          length:sizeof(uint64_t)      atIndex:10];
+                [encoder setBytes:&ne11          length:sizeof( int64_t)      atIndex:11];
+                [encoder setBytes:&ne12          length:sizeof( int64_t)      atIndex:12];
+                [encoder setBytes:&ne13          length:sizeof( int64_t)      atIndex:13];
+                [encoder setBytes:&nb11          length:sizeof(uint64_t)      atIndex:14];
+                [encoder setBytes:&nb12          length:sizeof(uint64_t)      atIndex:15];
+                [encoder setBytes:&nb13          length:sizeof(uint64_t)      atIndex:16];
+                [encoder setBytes:&nb21          length:sizeof(uint64_t)      atIndex:17];
+                [encoder setBytes:&nb22          length:sizeof(uint64_t)      atIndex:18];
+                [encoder setBytes:&nb23          length:sizeof(uint64_t)      atIndex:19];
+                [encoder setBytes:&nb31          length:sizeof(uint64_t)      atIndex:20];
+                [encoder setBytes:&ne1           length:sizeof( int64_t)      atIndex:21];
+                [encoder setBytes:&ne2           length:sizeof( int64_t)      atIndex:22];
+                [encoder setBytes:&scale         length:sizeof(   float)      atIndex:23];
+                [encoder setBytes:&max_bias      length:sizeof(   float)      atIndex:24];
+                [encoder setBytes:&m0            length:sizeof(m0)            atIndex:25];
+                [encoder setBytes:&m1            length:sizeof(m1)            atIndex:26];
+                [encoder setBytes:&n_head_log2   length:sizeof(n_head_log2)   atIndex:27];
+                [encoder setBytes:&logit_softcap length:sizeof(logit_softcap) atIndex:28];
+
+                if (!use_vec_kernel) {
+                    // half8x8 kernel
+                    const int64_t nqptg = 8;  // queries per threadgroup    !! sync with kernel template arguments !!
+                    const int64_t ncpsg = 32; // cache values per simdgroup !! sync with kernel template arguments !!
+
+                    GGML_ASSERT(nqptg <= 32);
+                    GGML_ASSERT(nqptg  % 8  == 0);
+                    GGML_ASSERT(ncpsg  % 32 == 0);
+
+                    int64_t nsgmax = 2;
+
+                    while (true) {
+                        const size_t smem = nqptg*(ne00 + 2*nsgmax*(ncpsg + nqptg))*(sizeof(float)/2);
+                        if (smem > ctx->device.maxThreadgroupMemoryLength) {
+                            break;
+                        }
+                        nsgmax *= 2;
+                    }
+                    nsgmax /= 2;
+
+                    // simdgroups per threadgroup (a.k.a. warps)
+                    const int64_t nsg = ne01 <= nqptg ? MAX(4, MIN(nsgmax, MIN(ne11/ncpsg, (int64_t) pipeline.maxTotalThreadsPerThreadgroup/32))) : 4;
+
+                    const size_t smem = nqptg*(ne00 + 2*nsg*(ncpsg + nqptg))*(sizeof(float)/2);
+
+                    //printf("smem: %zu, max: %zu\n", smem, ctx->device.maxThreadgroupMemoryLength);
+                    GGML_ASSERT(smem <= ctx->device.maxThreadgroupMemoryLength);
+
+                    [encoder setThreadgroupMemoryLength:GGML_PAD(smem, 16) atIndex:0];
+
+                    [encoder dispatchThreadgroups:MTLSizeMake((ne01 + nqptg - 1)/nqptg, ne02, ne03) threadsPerThreadgroup:MTLSizeMake(32, nsg, 1)];
+                } else {
+                    // half1x4 kernel
+                    const int64_t nqptg = 1;  // queries per threadgroup    !! sync with kernel template arguments !!
+                    const int64_t ncpsg = 32; // cache values per simdgroup !! sync with kernel template arguments !!
+
+                    GGML_ASSERT(nqptg <= 32);
+                    GGML_ASSERT(nqptg  % 1  == 0);
+                    GGML_ASSERT(ncpsg  % 32 == 0);
+
+                    // simdgroups per threadgroup (a.k.a. warps)
+                    const int64_t nsgt = MAX(2, MIN(ne11/ncpsg, (int64_t) pipeline.maxTotalThreadsPerThreadgroup/32));
+
+                    int64_t nsg = 1;
+                    while (nsg <= nsgt) {
+                        nsg *= 2;
+                    }
+                    nsg /= 2;
+
+                    const size_t smem = (nqptg*(ne00 + 2*nsg*(ncpsg + nqptg)) + nsg*ne00)*(sizeof(float)/2);
+
+                    //printf("smem: %zu, max: %zu\n", smem, ctx->device.maxThreadgroupMemoryLength);
+                    GGML_ASSERT(smem <= ctx->device.maxThreadgroupMemoryLength);
+                    [encoder setThreadgroupMemoryLength:GGML_PAD(smem, 16) atIndex:0];
+
+                    [encoder dispatchThreadgroups:MTLSizeMake((ne01 + nqptg - 1)/nqptg, ne02, ne03) threadsPerThreadgroup:MTLSizeMake(32, nsg, 1)];
+                }
+            } break;
+        case GGML_OP_DUP:
+        case GGML_OP_CPY:
+        case GGML_OP_CONT:
+            {
+                GGML_ASSERT(ne00 % ggml_blck_size(src0->type) == 0);
+
+                int nth = MIN(1024, ne00/ggml_blck_size(src0->type));
+
+                id<MTLComputePipelineState> pipeline = nil;
+
+                switch (src0t) {
+                    case GGML_TYPE_F32:
+                        {
+                            GGML_ASSERT(ne0 % ggml_blck_size(dst->type) == 0);
+
+                            switch (dstt) {
+                                case GGML_TYPE_F32:    pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_CPY_F32_F32].pipeline; break;
+                                case GGML_TYPE_F16:    pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_CPY_F32_F16].pipeline; break;
+                                case GGML_TYPE_Q8_0:   pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_CPY_F32_Q8_0].pipeline; break;
+                                case GGML_TYPE_Q4_0:   pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_CPY_F32_Q4_0].pipeline; break;
+                                case GGML_TYPE_Q4_1:   pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_CPY_F32_Q4_1].pipeline; break;
+                                case GGML_TYPE_Q5_0:   pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_CPY_F32_Q5_0].pipeline; break;
+                                case GGML_TYPE_Q5_1:   pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_CPY_F32_Q5_1].pipeline; break;
+                                case GGML_TYPE_IQ4_NL: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_CPY_F32_IQ4_NL].pipeline; break;
+                                default: GGML_ABORT("not implemented");
+                            };
+                        } break;
+                    case GGML_TYPE_F16:
+                        {
+                            switch (dstt) {
+                                case GGML_TYPE_F32:  pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_CPY_F16_F32].pipeline; break;
+                                case GGML_TYPE_F16:  pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_CPY_F16_F16].pipeline; break;
+                                default: GGML_ABORT("not implemented");
+                            };
+                        } break;
+                    default: GGML_ABORT("not implemented");
+                }
+
+                [encoder setComputePipelineState:pipeline];
+                [encoder setBuffer:id_src0 offset:offs_src0        atIndex:0];
+                [encoder setBuffer:id_dst  offset:offs_dst         atIndex:1];
+                [encoder setBytes:&ne00    length:sizeof( int64_t) atIndex:2];
+                [encoder setBytes:&ne01    length:sizeof( int64_t) atIndex:3];
+                [encoder setBytes:&ne02    length:sizeof( int64_t) atIndex:4];
+                [encoder setBytes:&ne03    length:sizeof( int64_t) atIndex:5];
+                [encoder setBytes:&nb00    length:sizeof(uint64_t) atIndex:6];
+                [encoder setBytes:&nb01    length:sizeof(uint64_t) atIndex:7];
+                [encoder setBytes:&nb02    length:sizeof(uint64_t) atIndex:8];
+                [encoder setBytes:&nb03    length:sizeof(uint64_t) atIndex:9];
+                [encoder setBytes:&ne0     length:sizeof( int64_t) atIndex:10];
+                [encoder setBytes:&ne1     length:sizeof( int64_t) atIndex:11];
+                [encoder setBytes:&ne2     length:sizeof( int64_t) atIndex:12];
+                [encoder setBytes:&ne3     length:sizeof( int64_t) atIndex:13];
+                [encoder setBytes:&nb0     length:sizeof(uint64_t) atIndex:14];
+                [encoder setBytes:&nb1     length:sizeof(uint64_t) atIndex:15];
+                [encoder setBytes:&nb2     length:sizeof(uint64_t) atIndex:16];
+                [encoder setBytes:&nb3     length:sizeof(uint64_t) atIndex:17];
+
+                [encoder dispatchThreadgroups:MTLSizeMake(ne01, ne02, ne03) threadsPerThreadgroup:MTLSizeMake(nth, 1, 1)];
+            } break;
+       default:
+            {
+                GGML_LOG_ERROR("%s: error: node %3d, op = %8s not implemented\n", __func__, idx, ggml_op_name(dst->op));
+                GGML_ABORT("fatal error");
+            }
+    }
+}
+
+static enum ggml_status ggml_metal_graph_compute(
+        struct ggml_backend_metal_context * ctx,
+                       struct ggml_cgraph * gf) {
+    // number of nodes encoded by the main thread (empirically determined)
+    const int n_main = 128;
+
+    // number of threads in addition to the main thread
+    const int n_cb = ctx->n_cb;
+
+    // submit the ggml compute graph to the GPU by creating command buffers and encoding the ops in them
+    // the first n_nodes_0 are encoded and submitted for processing directly by the calling thread
+    // while these nodes are processing, we start n_cb threads to enqueue the rest of the nodes
+    // each thread creates it's own command buffer and enqueues the ops in parallel
+    //
+    // tests on M1 Pro and M2 Ultra using LLaMA models, show that optimal values for n_cb are 1 or 2
+
+    @autoreleasepool {
+        ctx->gf = gf;
+
+        ctx->n_nodes_0 = MIN(n_main, gf->n_nodes);
+        ctx->n_nodes_1 = gf->n_nodes - ctx->n_nodes_0;
+
+        ctx->n_nodes_per_cb = (ctx->n_nodes_1 + ctx->n_cb - 1) / ctx->n_cb;
+
+        const bool should_capture = ctx->capture_next_compute;
+        if (should_capture) {
+            ctx->capture_next_compute = false;
+
+            if (!ctx->capture_started) {
+                // create capture scope
+                ctx->capture_scope = [[MTLCaptureManager sharedCaptureManager] newCaptureScopeWithDevice:ctx->device];
+
+                MTLCaptureDescriptor * descriptor = [MTLCaptureDescriptor new];
+                descriptor.captureObject = ctx->capture_scope;
+                descriptor.destination = MTLCaptureDestinationGPUTraceDocument;
+                descriptor.outputURL = [NSURL fileURLWithPath:[NSString stringWithFormat:@"/tmp/perf-metal.gputrace"]];
+
+                NSError * error = nil;
+                if (![[MTLCaptureManager sharedCaptureManager] startCaptureWithDescriptor:descriptor error:&error]) {
+                    GGML_LOG_ERROR("%s: error: unable to start capture '%s'\n", __func__, [[error localizedDescription] UTF8String]);
+                } else {
+                    [ctx->capture_scope beginScope];
+                    ctx->capture_started = true;
+                }
+            }
+        }
+
+        // TODO: how to avoid this allocation? I tried initializing it in ggml_backend_metal_set_n_cb but it crashes.
+        ctx->encode_async = ^(size_t iter) {
+            const int cb_idx = iter;
+            const int n_cb_l = ctx->n_cb;
+
+            const int n_nodes_0 = ctx->n_nodes_0;
+            const int n_nodes_1 = ctx->n_nodes_1;
+
+            const int n_nodes_per_cb = ctx->n_nodes_per_cb;
+
+            id<MTLCommandBuffer> command_buffer  = ctx->command_buffers[cb_idx];
+            id<MTLComputeCommandEncoder> encoder = [command_buffer computeCommandEncoder];
+
+            int node_start = 0;
+            int node_end   = n_nodes_0;
+
+            if (cb_idx < n_cb_l) {
+                node_start = n_nodes_0 + (                                         (cb_idx + 0) * n_nodes_per_cb);
+                node_end   = n_nodes_0 + (MIN((cb_idx == n_cb_l - 1) ? n_nodes_1 : (cb_idx + 1) * n_nodes_per_cb, n_nodes_1));
+            }
+
+            for (int idx = node_start; idx < node_end; ++idx) {
+                if (should_capture) {
+                    [encoder pushDebugGroup:[NSString stringWithCString:ggml_op_desc(ggml_graph_node(gf, idx)) encoding:NSUTF8StringEncoding]];
+                }
+
+                ggml_metal_encode_node(ctx, idx, encoder);
+
+                if (should_capture) {
+                    [encoder popDebugGroup];
+                }
+            }
+
+            [encoder endEncoding];
+
+            if (cb_idx < 2 || ctx->abort_callback == NULL) {
+                [command_buffer commit];
+            }
+        };
+
+        // the main thread commits the first few commands immediately
+        // command_buffer[n_cb]
+        {
+            id<MTLCommandBuffer> command_buffer = [ctx->queue commandBufferWithUnretainedReferences];
+            ctx->command_buffers[n_cb] = command_buffer;
+
+            [command_buffer enqueue];
+            ctx->encode_async(n_cb);
+        }
+
+        // prepare the rest of the command buffers asynchronously
+        // command_buffer[0.. n_cb)
+        for (int cb_idx = 0; cb_idx < n_cb; ++cb_idx) {
+            id<MTLCommandBuffer> command_buffer = [ctx->queue commandBufferWithUnretainedReferences];
+            ctx->command_buffers[cb_idx] = command_buffer;
+
+            // always enqueue the first two command buffers
+            // enqueue all of the command buffers if we don't need to abort
+            if (cb_idx < 2 || ctx->abort_callback == NULL) {
+                [command_buffer enqueue];
+            }
+        }
+
+        dispatch_apply(n_cb, ctx->d_queue, ctx->encode_async);
+
+        // wait for completion and check status of each command buffer
+        // needed to detect if the device ran out-of-memory for example (#1881)
+        {
+            id<MTLCommandBuffer> command_buffer = ctx->command_buffers[n_cb];
+            [command_buffer waitUntilCompleted];
+
+            MTLCommandBufferStatus status = [command_buffer status];
+            if (status != MTLCommandBufferStatusCompleted) {
+                GGML_LOG_INFO("%s: command buffer %d failed with status %lu\n", __func__, n_cb, status);
+                if (status == MTLCommandBufferStatusError) {
+                    GGML_LOG_INFO("error: %s\n", [[command_buffer error].localizedDescription UTF8String]);
+                }
+
+                return GGML_STATUS_FAILED;
+            }
+        }
+
+        for (int i = 0; i < n_cb; ++i) {
+            id<MTLCommandBuffer> command_buffer = ctx->command_buffers[i];
+            [command_buffer waitUntilCompleted];
+
+            MTLCommandBufferStatus status = [command_buffer status];
+            if (status != MTLCommandBufferStatusCompleted) {
+                GGML_LOG_INFO("%s: command buffer %d failed with status %lu\n", __func__, i, status);
+                if (status == MTLCommandBufferStatusError) {
+                    GGML_LOG_INFO("error: %s\n", [[command_buffer error].localizedDescription UTF8String]);
+                }
+
+                return GGML_STATUS_FAILED;
+            }
+
+            id<MTLCommandBuffer> next_buffer = (i + 1 < n_cb ? ctx->command_buffers[i + 1] : nil);
+            if (!next_buffer) {
+                continue;
+            }
+
+            const bool next_queued = ([next_buffer status] != MTLCommandBufferStatusNotEnqueued);
+            if (next_queued) {
+                continue;
+            }
+
+            if (ctx->abort_callback && ctx->abort_callback(ctx->abort_callback_data)) {
+                GGML_LOG_INFO("%s: command buffer %d aborted", __func__, i);
+                return GGML_STATUS_ABORTED;
+            }
+
+            [next_buffer commit];
+        }
+
+        if (!should_capture && ctx->capture_started) {
+            [ctx->capture_scope endScope];
+            [[MTLCaptureManager sharedCaptureManager] stopCapture];
+        }
+    }
+
+    return GGML_STATUS_SUCCESS;
+}
+
+////////////////////////////////////////////////////////////////////////////////
+
+// backend interface
+
+// default buffer
+static id<MTLDevice> g_backend_device = nil;
+static int g_backend_device_ref_count = 0;
+
+static id<MTLDevice> ggml_backend_metal_get_device(void) {
+    if (g_backend_device == nil) {
+        g_backend_device = MTLCreateSystemDefaultDevice();
+    }
+
+    g_backend_device_ref_count++;
+
+    return g_backend_device;
+}
+
+static void ggml_backend_metal_free_device(void) {
+    assert(g_backend_device_ref_count > 0);
+
+    g_backend_device_ref_count--;
+
+    if (g_backend_device_ref_count == 0) {
+        [g_backend_device release];
+        g_backend_device = nil;
+    }
+}
+
+static const char * ggml_backend_metal_buffer_get_name(ggml_backend_buffer_t buffer) {
+    return "Metal";
+
+    UNUSED(buffer);
+}
+
+static void ggml_backend_metal_buffer_free_buffer(ggml_backend_buffer_t buffer) {
+    struct ggml_backend_metal_buffer_context * ctx = (struct ggml_backend_metal_buffer_context *)buffer->context;
+
+    for (int i = 0; i < ctx->n_buffers; i++) {
+        [ctx->buffers[i].metal release];
+    }
+    ggml_backend_metal_free_device();
+
+    if (ctx->owned) {
+#if TARGET_OS_OSX
+        vm_deallocate((vm_map_t)mach_task_self(), (vm_address_t)ctx->all_data, ctx->all_size);
+#else
+        free(ctx->all_data);
+#endif
+    }
+
+    free(ctx);
+}
+
+static void * ggml_backend_metal_buffer_get_base(ggml_backend_buffer_t buffer) {
+    struct ggml_backend_metal_buffer_context * ctx = (struct ggml_backend_metal_buffer_context *)buffer->context;
+
+    return ctx->all_data;
+}
+
+static void ggml_backend_metal_buffer_set_tensor(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor, const void * data, size_t offset, size_t size) {
+    memcpy((char *)tensor->data + offset, data, size);
+
+    UNUSED(buffer);
+}
+
+static void ggml_backend_metal_buffer_get_tensor(ggml_backend_buffer_t buffer, const struct ggml_tensor * tensor, void * data, size_t offset, size_t size) {
+    memcpy(data, (const char *)tensor->data + offset, size);
+
+    UNUSED(buffer);
+}
+
+static bool ggml_backend_metal_buffer_cpy_tensor(ggml_backend_buffer_t buffer, const struct ggml_tensor * src, struct ggml_tensor * dst) {
+    if (ggml_backend_buffer_is_host(src->buffer)) {
+        memcpy(dst->data, src->data, ggml_nbytes(src));
+        return true;
+    }
+    return false;
+
+    UNUSED(buffer);
+}
+
+static void ggml_backend_metal_buffer_clear(ggml_backend_buffer_t buffer, uint8_t value) {
+    struct ggml_backend_metal_buffer_context * ctx = (struct ggml_backend_metal_buffer_context *)buffer->context;
+
+    memset(ctx->all_data, value, ctx->all_size);
+}
+
+static struct ggml_backend_buffer_i ggml_backend_metal_buffer_i = {
+    /* .get_name        = */ ggml_backend_metal_buffer_get_name,
+    /* .free_buffer     = */ ggml_backend_metal_buffer_free_buffer,
+    /* .get_base        = */ ggml_backend_metal_buffer_get_base,
+    /* .init_tensor     = */ NULL,
+    /* .memset_tensor   = */ NULL,
+    /* .set_tensor      = */ ggml_backend_metal_buffer_set_tensor,
+    /* .get_tensor      = */ ggml_backend_metal_buffer_get_tensor,
+    /* .cpy_tensor      = */ ggml_backend_metal_buffer_cpy_tensor,
+    /* .clear           = */ ggml_backend_metal_buffer_clear,
+    /* .reset           = */ NULL,
+};
+
+// default buffer type
+
+static const char * ggml_backend_metal_buffer_type_get_name(ggml_backend_buffer_type_t buft) {
+    return "Metal";
+
+    UNUSED(buft);
+}
+
+static void ggml_backend_metal_log_allocated_size(id<MTLDevice> device, size_t size_aligned) {
+#ifndef GGML_METAL_NDEBUG
+#if TARGET_OS_OSX || (TARGET_OS_IOS && __clang_major__ >= 15)
+    if (@available(macOS 10.12, iOS 16.0, *)) {
+        GGML_LOG_DEBUG("%s: allocated buffer, size = %8.2f MiB, (%8.2f / %8.2f)\n",
+                __func__,
+                size_aligned / 1024.0 / 1024.0,
+                device.currentAllocatedSize / 1024.0 / 1024.0,
+                device.recommendedMaxWorkingSetSize / 1024.0 / 1024.0);
+
+        if (device.currentAllocatedSize > device.recommendedMaxWorkingSetSize) {
+            GGML_LOG_WARN("%s: warning: current allocated size is greater than the recommended max working set size\n", __func__);
+        }
+    } else {
+        GGML_LOG_INFO("%s: allocated buffer, size = %8.2f MiB, (%8.2f)\n",
+                __func__,
+                size_aligned / 1024.0 / 1024.0,
+                device.currentAllocatedSize / 1024.0 / 1024.0);
+    }
+#endif
+#endif
+    UNUSED(device);
+    UNUSED(size_aligned);
+}
+
+static ggml_backend_buffer_t ggml_backend_metal_buffer_type_alloc_buffer(ggml_backend_buffer_type_t buft, size_t size) {
+    struct ggml_backend_metal_buffer_context * ctx = calloc(1, sizeof(struct ggml_backend_metal_buffer_context));
+
+    const size_t size_page = sysconf(_SC_PAGESIZE);
+
+    size_t size_aligned = size;
+    if ((size_aligned % size_page) != 0) {
+        size_aligned += (size_page - (size_aligned % size_page));
+    }
+
+    id<MTLDevice> device = ggml_backend_metal_get_device();
+
+    ctx->all_data = ggml_metal_host_malloc(size_aligned);
+    ctx->all_size = size_aligned;
+    ctx->owned = true;
+    ctx->n_buffers = 1;
+
+    if (ctx->all_data != NULL) {
+        ctx->buffers[0].data  = ctx->all_data;
+        ctx->buffers[0].size  = size;
+        ctx->buffers[0].metal = nil;
+
+        if (size_aligned > 0) {
+            ctx->buffers[0].metal = [device newBufferWithBytesNoCopy:ctx->all_data
+                            length:size_aligned
+                            options:MTLResourceStorageModeShared
+                            deallocator:nil];
+        }
+    }
+
+    if (size_aligned > 0 && (ctx->all_data == NULL || ctx->buffers[0].metal == nil)) {
+        GGML_LOG_ERROR("%s: error: failed to allocate buffer, size = %8.2f MiB\n", __func__, size_aligned / 1024.0 / 1024.0);
+        free(ctx);
+        ggml_backend_metal_free_device();
+        return NULL;
+    }
+
+    //ggml_backend_metal_log_allocated_size(device, size_aligned);
+
+    return ggml_backend_buffer_init(buft, ggml_backend_metal_buffer_i, ctx, size);
+}
+
+static size_t ggml_backend_metal_buffer_type_get_alignment(ggml_backend_buffer_type_t buft) {
+    return 32;
+    UNUSED(buft);
+}
+
+static size_t ggml_backend_metal_buffer_type_get_max_size(ggml_backend_buffer_type_t buft) {
+    id<MTLDevice> device = ggml_backend_metal_get_device();
+    size_t max_size = device.maxBufferLength;
+    ggml_backend_metal_free_device();
+
+    return max_size;
+
+    UNUSED(buft);
+}
+
+static bool ggml_backend_metal_buffer_type_is_host(ggml_backend_buffer_type_t buft) {
+    return true;
+
+    UNUSED(buft);
+}
+
+ggml_backend_buffer_type_t ggml_backend_metal_buffer_type(void) {
+    static struct ggml_backend_buffer_type ggml_backend_buffer_type_metal = {
+        /* .iface = */ {
+            /* .get_name         = */ ggml_backend_metal_buffer_type_get_name,
+            /* .alloc_buffer     = */ ggml_backend_metal_buffer_type_alloc_buffer,
+            /* .get_alignment    = */ ggml_backend_metal_buffer_type_get_alignment,
+            /* .get_max_size     = */ ggml_backend_metal_buffer_type_get_max_size,
+            /* .get_alloc_size   = */ NULL, // defaults to ggml_nbytes
+            /* .is_host          = */ ggml_backend_metal_buffer_type_is_host,
+        },
+        /* .device  = */ NULL,
+        /* .context = */ NULL,
+    };
+
+    return &ggml_backend_buffer_type_metal;
+}
+
+// buffer from ptr
+
+ggml_backend_buffer_t ggml_backend_metal_buffer_from_ptr(void * data, size_t size, size_t max_size) {
+    struct ggml_backend_metal_buffer_context * ctx = calloc(1, sizeof(struct ggml_backend_metal_buffer_context));
+
+    ctx->all_data = data;
+    ctx->all_size = size;
+    ctx->owned = false;
+    ctx->n_buffers = 0;
+
+    const size_t size_page = sysconf(_SC_PAGESIZE);
+
+    // page-align the data ptr
+    {
+        const uintptr_t offs = (uintptr_t) data % size_page;
+        data  = (void *) ((char *) data - offs);
+        size += offs;
+    }
+
+    size_t size_aligned = size;
+    if ((size_aligned % size_page) != 0) {
+        size_aligned += (size_page - (size_aligned % size_page));
+    }
+
+    id<MTLDevice> device = ggml_backend_metal_get_device();
+
+    // the buffer fits into the max buffer size allowed by the device
+    if (size_aligned <= device.maxBufferLength) {
+        ctx->buffers[ctx->n_buffers].data  = data;
+        ctx->buffers[ctx->n_buffers].size  = size;
+        ctx->buffers[ctx->n_buffers].metal = nil;
+
+        if (size_aligned > 0) {
+            ctx->buffers[ctx->n_buffers].metal = [device newBufferWithBytesNoCopy:data length:size_aligned options:MTLResourceStorageModeShared deallocator:nil];
+
+            if (ctx->buffers[ctx->n_buffers].metal == nil) {
+                GGML_LOG_ERROR("%s: error: failed to allocate buffer, size = %8.2f MiB\n", __func__, size_aligned / 1024.0 / 1024.0);
+                return false;
+            }
+        }
+
+        ggml_backend_metal_log_allocated_size(device, size_aligned);
+
+        ++ctx->n_buffers;
+    } else {
+        // this overlap between the views will guarantee that the tensor with the maximum size will fully fit into
+        // one of the views
+        const size_t size_ovlp = ((max_size + size_page - 1) / size_page + 1) * size_page; // round-up 2 pages just in case
+        const size_t size_step = device.maxBufferLength - size_ovlp;
+        const size_t size_view = device.maxBufferLength;
+
+        for (size_t i = 0; i < size; i += size_step) {
+            const size_t size_step_aligned = (i + size_view <= size) ? size_view : (size_aligned - i);
+
+            ctx->buffers[ctx->n_buffers].data  = (void *) ((uint8_t *) data + i);
+            ctx->buffers[ctx->n_buffers].size  = size_step_aligned;
+            ctx->buffers[ctx->n_buffers].metal = nil;
+
+            if (size_step_aligned > 0) {
+                ctx->buffers[ctx->n_buffers].metal = [device newBufferWithBytesNoCopy:(void *) ((uint8_t *) data + i) length:size_step_aligned options:MTLResourceStorageModeShared deallocator:nil];
+
+                if (ctx->buffers[ctx->n_buffers].metal == nil) {
+                    GGML_LOG_ERROR("%s: error: failed to allocate buffer, size = %8.2f MiB\n", __func__, size_step_aligned / 1024.0 / 1024.0);
+                    return false;
+                }
+            }
+
+            ggml_backend_metal_log_allocated_size(device, size_step_aligned);
+
+            if (i + size_step < size) {
+                GGML_LOG_INFO("\n");
+            }
+
+            ++ctx->n_buffers;
+        }
+    }
+
+    return ggml_backend_buffer_init(ggml_backend_metal_buffer_type(), ggml_backend_metal_buffer_i, ctx, size);
+}
+
+// backend
+
+static const char * ggml_backend_metal_name(ggml_backend_t backend) {
+    return "Metal";
+
+    UNUSED(backend);
+}
+
+static void ggml_backend_metal_free(ggml_backend_t backend) {
+    struct ggml_backend_metal_context * ctx = (struct ggml_backend_metal_context *)backend->context;
+    ggml_metal_free(ctx);
+    free(backend);
+}
+
+static ggml_backend_buffer_type_t ggml_backend_metal_get_default_buffer_type(ggml_backend_t backend) {
+    return ggml_backend_metal_buffer_type();
+
+    UNUSED(backend);
+}
+
+static enum ggml_status ggml_backend_metal_graph_compute(ggml_backend_t backend, struct ggml_cgraph * cgraph) {
+    struct ggml_backend_metal_context * metal_ctx = (struct ggml_backend_metal_context *)backend->context;
+
+    return ggml_metal_graph_compute(metal_ctx, cgraph);
+}
+
+static bool ggml_backend_metal_supports_op(ggml_backend_t backend, const struct ggml_tensor * op) {
+    struct ggml_backend_metal_context * metal_ctx = (struct ggml_backend_metal_context *)backend->context;
+
+    return ggml_metal_supports_op(metal_ctx, op);
+}
+
+static bool ggml_backend_metal_supports_buft(ggml_backend_t backend, ggml_backend_buffer_type_t buft) {
+    return buft->iface.get_name == ggml_backend_metal_buffer_type_get_name;
+
+    UNUSED(backend);
+}
+
+static void ggml_backend_metal_set_n_cb(ggml_backend_t backend, int n_cb) {
+    GGML_ASSERT(ggml_backend_is_metal(backend));
+
+    struct ggml_backend_metal_context * ctx = (struct ggml_backend_metal_context *)backend->context;
+
+    if (ctx->n_cb != n_cb) {
+        ctx->n_cb = MIN(n_cb, GGML_METAL_MAX_COMMAND_BUFFERS);
+
+        if (ctx->n_cb > 2) {
+            GGML_LOG_WARN("%s: n_cb = %d, using n_cb > 2 is not recommended and can degrade the performance in some cases\n", __func__, n_cb);
+        }
+    }
+
+    // TODO: setting encode_async here causes crash during the next ggml_metal_graph_compute call. why?
+    //ctx->encode_async = ^(size_t iter) {
+    //    ...
+    //};
+}
+
+static struct ggml_backend_i ggml_backend_metal_i = {
+    /* .get_name                = */ ggml_backend_metal_name,
+    /* .free                    = */ ggml_backend_metal_free,
+    /* .get_default_buffer_type = */ ggml_backend_metal_get_default_buffer_type,
+    /* .set_tensor_async        = */ NULL,
+    /* .get_tensor_async        = */ NULL,
+    /* .cpy_tensor_async        = */ NULL,
+    /* .synchronize             = */ NULL,
+    /* .graph_plan_create       = */ NULL,
+    /* .graph_plan_free         = */ NULL,
+    /* .graph_plan_update       = */ NULL,
+    /* .graph_plan_compute      = */ NULL,
+    /* .graph_compute           = */ ggml_backend_metal_graph_compute,
+    /* .supports_op             = */ ggml_backend_metal_supports_op,
+    /* .supports_buft           = */ ggml_backend_metal_supports_buft,
+    /* .offload_op              = */ NULL,
+    /* .event_record            = */ NULL,
+    /* .event_wait              = */ NULL,
+};
+
+static ggml_guid_t ggml_backend_metal_guid(void) {
+    static ggml_guid guid = { 0x81, 0xa1, 0x8b, 0x1e, 0x71, 0xec, 0x79, 0xed, 0x2b, 0x85, 0xdc, 0x8a, 0x61, 0x98, 0x30, 0xe6 };
+    return &guid;
+}
+
+ggml_backend_t ggml_backend_metal_init(void) {
+    struct ggml_backend_metal_context * ctx = ggml_metal_init();
+    if (ctx == NULL) {
+        GGML_LOG_ERROR("%s: error: failed to allocate context\n", __func__);
+        return NULL;
+    }
+
+    ggml_backend_t backend = malloc(sizeof(struct ggml_backend));
+
+    *backend = (struct ggml_backend) {
+        /* .guid      = */ ggml_backend_metal_guid(),
+        /* .interface = */ ggml_backend_metal_i,
+        /* .device    = */ NULL,
+        /* .context   = */ ctx,
+    };
+
+    ggml_backend_metal_set_n_cb(backend, 1);
+
+    return backend;
+}
+
+bool ggml_backend_is_metal(ggml_backend_t backend) {
+    return backend != NULL && ggml_guid_matches(backend->guid, ggml_backend_metal_guid());
+}
+
+void ggml_backend_metal_set_abort_callback(ggml_backend_t backend, ggml_abort_callback abort_callback, void * user_data) {
+    GGML_ASSERT(ggml_backend_is_metal(backend));
+
+    struct ggml_backend_metal_context * ctx = (struct ggml_backend_metal_context *)backend->context;
+
+    ctx->abort_callback = abort_callback;
+    ctx->abort_callback_data = user_data;
+}
+
+bool ggml_backend_metal_supports_family(ggml_backend_t backend, int family) {
+    GGML_ASSERT(ggml_backend_is_metal(backend));
+
+    struct ggml_backend_metal_context * ctx = (struct ggml_backend_metal_context *)backend->context;
+
+    return [ctx->device supportsFamily:(MTLGPUFamilyApple1 + family - 1)];
+}
+
+void ggml_backend_metal_capture_next_compute(ggml_backend_t backend) {
+    GGML_ASSERT(ggml_backend_is_metal(backend));
+
+    struct ggml_backend_metal_context * ctx = (struct ggml_backend_metal_context *)backend->context;
+    ctx->capture_next_compute = true;
+}
+
+ggml_backend_t ggml_backend_reg_metal_init(const char * params, void * user_data); // silence warning
+
+ggml_backend_t ggml_backend_reg_metal_init(const char * params, void * user_data) {
+    return ggml_backend_metal_init();
+
+    GGML_UNUSED(params);
+    GGML_UNUSED(user_data);
+}
diff --git a/src/whisper_cpp/ggml/src/ggml-metal.metal b/src/whisper_cpp/ggml/src/ggml-metal.metal
new file mode 100644
index 00000000..2b200032
--- /dev/null
+++ b/src/whisper_cpp/ggml/src/ggml-metal.metal
@@ -0,0 +1,6374 @@
+#define GGML_COMMON_DECL_METAL
+#define GGML_COMMON_IMPL_METAL
+#include "ggml-common.h"
+
+#include <metal_stdlib>
+
+using namespace metal;
+
+#define MAX(x, y) ((x) > (y) ? (x) : (y))
+#define MIN(x, y) ((x) < (y) ? (x) : (y))
+#define SWAP(x, y) { auto tmp = (x); (x) = (y); (y) = tmp; }
+
+#define N_SIMDWIDTH 32 // assuming SIMD group size is 32
+
+enum ggml_sort_order {
+    GGML_SORT_ORDER_ASC,
+    GGML_SORT_ORDER_DESC,
+};
+
+// general-purpose kernel for addition, subtraction, multiplication and division of two tensors
+// pros: works for non-contiguous tensors, supports broadcast across all dims
+// cons: not very efficient
+kernel void kernel_add(
+        device const char * src0,
+        device const char * src1,
+        device       char * dst,
+        constant  int64_t & ne00,
+        constant  int64_t & ne01,
+        constant  int64_t & ne02,
+        constant  int64_t & ne03,
+        constant uint64_t & nb00,
+        constant uint64_t & nb01,
+        constant uint64_t & nb02,
+        constant uint64_t & nb03,
+        constant  int64_t & ne10,
+        constant  int64_t & ne11,
+        constant  int64_t & ne12,
+        constant  int64_t & ne13,
+        constant uint64_t & nb10,
+        constant uint64_t & nb11,
+        constant uint64_t & nb12,
+        constant uint64_t & nb13,
+        constant  int64_t & ne0,
+        constant  int64_t & ne1,
+        constant  int64_t & ne2,
+        constant  int64_t & ne3,
+        constant uint64_t & nb0,
+        constant uint64_t & nb1,
+        constant uint64_t & nb2,
+        constant uint64_t & nb3,
+        constant  int64_t & offs,
+        uint3 tgpig[[threadgroup_position_in_grid]],
+        uint3 tpitg[[thread_position_in_threadgroup]],
+        uint3   ntg[[threads_per_threadgroup]]) {
+    const int64_t i03 = tgpig.z;
+    const int64_t i02 = tgpig.y;
+    const int64_t i01 = tgpig.x;
+
+    const int64_t i13 = i03 % ne13;
+    const int64_t i12 = i02 % ne12;
+    const int64_t i11 = i01 % ne11;
+
+    device const char * src0_ptr = src0 + i03*nb03 + i02*nb02 + i01*nb01 + offs;
+    device const char * src1_ptr = src1 + i13*nb13 + i12*nb12 + i11*nb11;
+    device       char * dst_ptr  = dst  + i03*nb3  + i02*nb2  + i01*nb1  + offs;
+
+    for (int i0 = tpitg.x; i0 < ne0; i0 += ntg.x) {
+        const int i10 = i0 % ne10;
+        *((device float *)(dst_ptr + i0*nb0)) = *((device float *)(src0_ptr + i0*nb00)) + *((device float *)(src1_ptr + i10*nb10));
+    }
+}
+
+kernel void kernel_sub(
+        device const char * src0,
+        device const char * src1,
+        device       char * dst,
+        constant  int64_t & ne00,
+        constant  int64_t & ne01,
+        constant  int64_t & ne02,
+        constant  int64_t & ne03,
+        constant uint64_t & nb00,
+        constant uint64_t & nb01,
+        constant uint64_t & nb02,
+        constant uint64_t & nb03,
+        constant  int64_t & ne10,
+        constant  int64_t & ne11,
+        constant  int64_t & ne12,
+        constant  int64_t & ne13,
+        constant uint64_t & nb10,
+        constant uint64_t & nb11,
+        constant uint64_t & nb12,
+        constant uint64_t & nb13,
+        constant  int64_t & ne0,
+        constant  int64_t & ne1,
+        constant  int64_t & ne2,
+        constant  int64_t & ne3,
+        constant uint64_t & nb0,
+        constant uint64_t & nb1,
+        constant uint64_t & nb2,
+        constant uint64_t & nb3,
+        constant  int64_t & offs,
+        uint3 tgpig[[threadgroup_position_in_grid]],
+        uint3 tpitg[[thread_position_in_threadgroup]],
+        uint3   ntg[[threads_per_threadgroup]]) {
+    const int64_t i03 = tgpig.z;
+    const int64_t i02 = tgpig.y;
+    const int64_t i01 = tgpig.x;
+
+    const int64_t i13 = i03 % ne13;
+    const int64_t i12 = i02 % ne12;
+    const int64_t i11 = i01 % ne11;
+
+    device const char * src0_ptr = src0 + i03*nb03 + i02*nb02 + i01*nb01 + offs;
+    device const char * src1_ptr = src1 + i13*nb13 + i12*nb12 + i11*nb11;
+    device       char * dst_ptr  = dst  + i03*nb3  + i02*nb2  + i01*nb1  + offs;
+
+    for (int i0 = tpitg.x; i0 < ne0; i0 += ntg.x) {
+        const int i10 = i0 % ne10;
+        *((device float *)(dst_ptr + i0*nb0)) = *((device float *)(src0_ptr + i0*nb00)) - *((device float *)(src1_ptr + i10*nb10));
+    }
+}
+
+kernel void kernel_mul(
+        device const char * src0,
+        device const char * src1,
+        device       char * dst,
+        constant  int64_t & ne00,
+        constant  int64_t & ne01,
+        constant  int64_t & ne02,
+        constant  int64_t & ne03,
+        constant uint64_t & nb00,
+        constant uint64_t & nb01,
+        constant uint64_t & nb02,
+        constant uint64_t & nb03,
+        constant  int64_t & ne10,
+        constant  int64_t & ne11,
+        constant  int64_t & ne12,
+        constant  int64_t & ne13,
+        constant uint64_t & nb10,
+        constant uint64_t & nb11,
+        constant uint64_t & nb12,
+        constant uint64_t & nb13,
+        constant  int64_t & ne0,
+        constant  int64_t & ne1,
+        constant  int64_t & ne2,
+        constant  int64_t & ne3,
+        constant uint64_t & nb0,
+        constant uint64_t & nb1,
+        constant uint64_t & nb2,
+        constant uint64_t & nb3,
+        uint3 tgpig[[threadgroup_position_in_grid]],
+        uint3 tpitg[[thread_position_in_threadgroup]],
+        uint3   ntg[[threads_per_threadgroup]]) {
+    const int64_t i03 = tgpig.z;
+    const int64_t i02 = tgpig.y;
+    const int64_t i01 = tgpig.x;
+
+    const int64_t i13 = i03 % ne13;
+    const int64_t i12 = i02 % ne12;
+    const int64_t i11 = i01 % ne11;
+
+    device const char * src0_ptr = src0 + i03*nb03 + i02*nb02 + i01*nb01;
+    device const char * src1_ptr = src1 + i13*nb13 + i12*nb12 + i11*nb11;
+    device       char * dst_ptr  = dst  + i03*nb3  + i02*nb2  + i01*nb1;
+
+    for (int i0 = tpitg.x; i0 < ne0; i0 += ntg.x) {
+        const int i10 = i0 % ne10;
+        *((device float *)(dst_ptr + i0*nb0)) = *((device float *)(src0_ptr + i0*nb00)) * *((device float *)(src1_ptr + i10*nb10));
+    }
+}
+
+kernel void kernel_div(
+        device const char * src0,
+        device const char * src1,
+        device       char * dst,
+        constant  int64_t & ne00,
+        constant  int64_t & ne01,
+        constant  int64_t & ne02,
+        constant  int64_t & ne03,
+        constant uint64_t & nb00,
+        constant uint64_t & nb01,
+        constant uint64_t & nb02,
+        constant uint64_t & nb03,
+        constant  int64_t & ne10,
+        constant  int64_t & ne11,
+        constant  int64_t & ne12,
+        constant  int64_t & ne13,
+        constant uint64_t & nb10,
+        constant uint64_t & nb11,
+        constant uint64_t & nb12,
+        constant uint64_t & nb13,
+        constant  int64_t & ne0,
+        constant  int64_t & ne1,
+        constant  int64_t & ne2,
+        constant  int64_t & ne3,
+        constant uint64_t & nb0,
+        constant uint64_t & nb1,
+        constant uint64_t & nb2,
+        constant uint64_t & nb3,
+        uint3 tgpig[[threadgroup_position_in_grid]],
+        uint3 tpitg[[thread_position_in_threadgroup]],
+        uint3   ntg[[threads_per_threadgroup]]) {
+    const int64_t i03 = tgpig.z;
+    const int64_t i02 = tgpig.y;
+    const int64_t i01 = tgpig.x;
+
+    const int64_t i13 = i03 % ne13;
+    const int64_t i12 = i02 % ne12;
+    const int64_t i11 = i01 % ne11;
+
+    device const char * src0_ptr = src0 + i03*nb03 + i02*nb02 + i01*nb01;
+    device const char * src1_ptr = src1 + i13*nb13 + i12*nb12 + i11*nb11;
+    device       char * dst_ptr  = dst  + i03*nb3  + i02*nb2  + i01*nb1;
+
+    for (int i0 = tpitg.x; i0 < ne0; i0 += ntg.x) {
+        const int i10 = i0 % ne10;
+        *((device float *)(dst_ptr + i0*nb0)) = *((device float *)(src0_ptr + i0*nb00)) / *((device float *)(src1_ptr + i10*nb10));
+    }
+}
+
+template<typename T>
+kernel void kernel_repeat(
+        device const char * src0,
+        device       char * dst,
+        constant  int64_t & ne00,
+        constant  int64_t & ne01,
+        constant  int64_t & ne02,
+        constant  int64_t & ne03,
+        constant uint64_t & nb00,
+        constant uint64_t & nb01,
+        constant uint64_t & nb02,
+        constant uint64_t & nb03,
+        constant  int64_t & ne0,
+        constant  int64_t & ne1,
+        constant  int64_t & ne2,
+        constant  int64_t & ne3,
+        constant uint64_t & nb0,
+        constant uint64_t & nb1,
+        constant uint64_t & nb2,
+        constant uint64_t & nb3,
+        uint3 tgpig[[threadgroup_position_in_grid]],
+        uint3 tpitg[[thread_position_in_threadgroup]],
+        uint3   ntg[[threads_per_threadgroup]]) {
+    const int64_t i3 = tgpig.z;
+    const int64_t i2 = tgpig.y;
+    const int64_t i1 = tgpig.x;
+
+    const int64_t i03 = i3 % ne03;
+    const int64_t i02 = i2 % ne02;
+    const int64_t i01 = i1 % ne01;
+
+    device const char * src0_ptr = src0 + i03*nb03 + i02*nb02 + i01*nb01;
+    device       char * dst_ptr  = dst  +  i3*nb3  +  i2*nb2  +  i1*nb1 ;
+
+    for (int i0 = tpitg.x; i0 < ne0; i0 += ntg.x) {
+        const int i00 = i0 % ne00;
+        *((device T *)(dst_ptr + i0*nb0)) = *((device T *)(src0_ptr + i00*nb00));
+    }
+}
+
+typedef decltype(kernel_repeat<float>) kernel_repeat_t;
+
+template [[host_name("kernel_repeat_f32")]] kernel kernel_repeat_t kernel_repeat<float>;
+template [[host_name("kernel_repeat_f16")]] kernel kernel_repeat_t kernel_repeat<half>;
+template [[host_name("kernel_repeat_i32")]] kernel kernel_repeat_t kernel_repeat<int>;
+template [[host_name("kernel_repeat_i16")]] kernel kernel_repeat_t kernel_repeat<short>;
+
+// assumption: src1 is a row
+// broadcast src1 into src0
+kernel void kernel_add_row(
+        device const float4 * src0,
+        device const float4 * src1,
+        device       float4 * dst,
+        constant   uint64_t & nb [[buffer(28)]],
+        uint tpig[[thread_position_in_grid]]) {
+    dst[tpig] = src0[tpig] + src1[tpig % nb];
+}
+
+kernel void kernel_sub_row(
+        device const float4 * src0,
+        device const float4 * src1,
+        device       float4 * dst,
+        constant   uint64_t & nb [[buffer(28)]],
+        uint tpig[[thread_position_in_grid]]) {
+    dst[tpig] = src0[tpig] - src1[tpig % nb];
+}
+
+kernel void kernel_mul_row(
+        device const float4 * src0,
+        device const float4 * src1,
+        device       float4 * dst,
+        constant   uint64_t & nb  [[buffer(28)]],
+        uint tpig[[thread_position_in_grid]]) {
+    dst[tpig] = src0[tpig] * src1[tpig % nb];
+}
+
+kernel void kernel_div_row(
+        device const float4 * src0,
+        device const float4 * src1,
+        device       float4 * dst,
+        constant   uint64_t & nb  [[buffer(28)]],
+        uint tpig[[thread_position_in_grid]]) {
+    dst[tpig] = src0[tpig] / src1[tpig % nb];
+}
+
+kernel void kernel_scale(
+        device const float * src0,
+        device       float * dst,
+        constant     float & scale,
+        uint tpig[[thread_position_in_grid]]) {
+    dst[tpig] = src0[tpig] * scale;
+}
+
+kernel void kernel_scale_4(
+        device const float4 * src0,
+        device       float4 * dst,
+        constant     float  & scale,
+        uint tpig[[thread_position_in_grid]]) {
+    dst[tpig] = src0[tpig] * scale;
+}
+
+kernel void kernel_clamp(
+        device const float * src0,
+        device       float * dst,
+        constant     float & min,
+        constant     float & max,
+        uint tpig[[thread_position_in_grid]]) {
+    dst[tpig] = src0[tpig] < min ? min : (src0[tpig] > max ? max : src0[tpig]);
+}
+
+kernel void kernel_relu(
+        device const float * src0,
+        device       float * dst,
+        uint tpig[[thread_position_in_grid]]) {
+    dst[tpig] = max(0.0f, src0[tpig]);
+}
+
+kernel void kernel_sigmoid(
+        device const float * src0,
+        device       float * dst,
+        uint tpig[[thread_position_in_grid]]) {
+    dst[tpig] = 1.0f / (1.0f + exp(-src0[tpig]));
+}
+
+kernel void kernel_tanh(
+        device const float * src0,
+        device       float * dst,
+        uint tpig[[thread_position_in_grid]]) {
+    device const float & x = src0[tpig];
+    dst[tpig] = precise::tanh(x);
+}
+
+constant float GELU_COEF_A     = 0.044715f;
+constant float GELU_QUICK_COEF = -1.702f;
+constant float SQRT_2_OVER_PI  = 0.79788456080286535587989211986876f;
+
+kernel void kernel_gelu(
+    device const float * src0,
+    device       float * dst,
+    uint tpig[[thread_position_in_grid]]) {
+    device const float & x = src0[tpig];
+
+    dst[tpig] = 0.5f*x*(1.0f + precise::tanh(SQRT_2_OVER_PI*x*(1.0f + GELU_COEF_A*x*x)));
+}
+
+kernel void kernel_gelu_4(
+    device const float4 * src0,
+    device       float4 * dst,
+    uint tpig[[thread_position_in_grid]]) {
+    device const float4 & x = src0[tpig];
+
+    // BEWARE !!!
+    // Simply using "tanh" instead of "precise::tanh" will sometimes results in NaNs!
+    // This was observed with Falcon 7B and 40B models
+    //
+    dst[tpig] = 0.5f*x*(1.0f + precise::tanh(SQRT_2_OVER_PI*x*(1.0f + GELU_COEF_A*x*x)));
+}
+
+kernel void kernel_gelu_quick(
+    device const float * src0,
+    device       float * dst,
+    uint tpig[[thread_position_in_grid]]) {
+    device const float & x = src0[tpig];
+
+    dst[tpig] = x*(1.0f/(1.0f+exp(GELU_QUICK_COEF*x)));
+}
+
+kernel void kernel_gelu_quick_4(
+    device const float4 * src0,
+    device       float4 * dst,
+    uint tpig[[thread_position_in_grid]]) {
+    device const float4 & x = src0[tpig];
+
+    dst[tpig] = x*(1.0f/(1.0f+exp(GELU_QUICK_COEF*x)));
+}
+
+kernel void kernel_silu(
+        device const float * src0,
+        device       float * dst,
+        uint tpig[[thread_position_in_grid]]) {
+    device const float & x = src0[tpig];
+    dst[tpig] = x / (1.0f + exp(-x));
+}
+
+kernel void kernel_silu_4(
+        device const float4 * src0,
+        device       float4 * dst,
+        uint tpig[[thread_position_in_grid]]) {
+    device const float4 & x = src0[tpig];
+    dst[tpig] = x / (1.0f + exp(-x));
+}
+
+kernel void kernel_sqr(
+        device const float * src0,
+        device       float * dst,
+        uint tpig[[thread_position_in_grid]]) {
+    dst[tpig] = src0[tpig] * src0[tpig];
+}
+
+kernel void kernel_sqrt(
+        device const float * src0,
+        device       float * dst,
+        uint tpig[[thread_position_in_grid]]) {
+    dst[tpig] = sqrt(src0[tpig]);
+}
+
+kernel void kernel_sin(
+        device const float * src0,
+        device       float * dst,
+        uint tpig[[thread_position_in_grid]]) {
+    dst[tpig] = sin(src0[tpig]);
+}
+
+kernel void kernel_cos(
+        device const float * src0,
+        device       float * dst,
+        uint tpig[[thread_position_in_grid]]) {
+    dst[tpig] = cos(src0[tpig]);
+}
+
+kernel void kernel_sum_rows(
+        device const float * src0,
+        device       float * dst,
+        constant  int64_t & ne00,
+        constant  int64_t & ne01,
+        constant  int64_t & ne02,
+        constant  int64_t & ne03,
+        constant uint64_t & nb00,
+        constant uint64_t & nb01,
+        constant uint64_t & nb02,
+        constant uint64_t & nb03,
+        constant  int64_t & ne10,
+        constant  int64_t & ne11,
+        constant  int64_t & ne12,
+        constant  int64_t & ne13,
+        constant uint64_t & nb10,
+        constant uint64_t & nb11,
+        constant uint64_t & nb12,
+        constant uint64_t & nb13,
+        constant  int64_t & ne0,
+        constant  int64_t & ne1,
+        constant  int64_t & ne2,
+        constant  int64_t & ne3,
+        constant uint64_t & nb0,
+        constant uint64_t & nb1,
+        constant uint64_t & nb2,
+        constant uint64_t & nb3,
+        uint3 tpig[[thread_position_in_grid]]) {
+    int64_t i3 = tpig.z;
+    int64_t i2 = tpig.y;
+    int64_t i1 = tpig.x;
+
+    if (i3 >= ne03 || i2 >= ne02 || i1 >= ne01) {
+        return;
+    }
+
+    device const float * src_row = (device const float *) ((device const char *) src0 + i1*nb01 + i2*nb02 + i3*nb03);
+    device       float * dst_row = (device       float *) ((device       char *) dst  + i1*nb1  + i2*nb2  + i3*nb3);
+
+    float row_sum = 0;
+
+    for (int64_t i0 = 0; i0 < ne00; i0++) {
+        row_sum += src_row[i0];
+    }
+
+    dst_row[0] = row_sum;
+}
+
+template<typename T>
+kernel void kernel_soft_max(
+        device const  char * src0,
+        device const  char * src1,
+        device        char * dst,
+        constant   int64_t & ne00,
+        constant   int64_t & ne01,
+        constant   int64_t & ne02,
+        constant     float & scale,
+        constant     float & max_bias,
+        constant     float & m0,
+        constant     float & m1,
+        constant  uint32_t & n_head_log2,
+        threadgroup  float * buf [[threadgroup(0)]],
+        uint  tgpig[[threadgroup_position_in_grid]],
+        uint  tpitg[[thread_position_in_threadgroup]],
+        uint  sgitg[[simdgroup_index_in_threadgroup]],
+        uint  tiisg[[thread_index_in_simdgroup]],
+        uint    ntg[[threads_per_threadgroup]]) {
+    const int64_t i03 = (tgpig) / (ne02*ne01);
+    const int64_t i02 = (tgpig - i03*ne02*ne01) / ne01;
+    const int64_t i01 = (tgpig - i03*ne02*ne01 - i02*ne01);
+
+    device const float * psrc0 = (device const float *) src0 + (i03*ne02*ne01*ne00 + i02*ne01*ne00 + i01*ne00);
+    device const     T * pmask = src1 != src0 ? (device const    T *) src1         + i01*ne00 : nullptr;
+    device       float * pdst  = (device       float *) dst  + (i03*ne02*ne01*ne00 + i02*ne01*ne00 + i01*ne00);
+
+    float slope = 1.0f;
+
+    // ALiBi
+    if (max_bias > 0.0f) {
+        const int64_t h = i02;
+
+        const float base = h < n_head_log2 ? m0 : m1;
+        const int   exp  = h < n_head_log2 ? h + 1 : 2*(h - n_head_log2) + 1;
+
+        slope = pow(base, exp);
+    }
+
+    // parallel max
+    float lmax = -INFINITY;
+
+    for (int i00 = tpitg; i00 < ne00; i00 += ntg) {
+        lmax = MAX(lmax, psrc0[i00]*scale + (pmask ? slope*pmask[i00] : 0.0f));
+    }
+
+    // find the max value in the block
+    float max_val = simd_max(lmax);
+    if (ntg > N_SIMDWIDTH) {
+        if (sgitg == 0) {
+            buf[tiisg] = -INFINITY;
+        }
+
+        threadgroup_barrier(mem_flags::mem_threadgroup);
+
+        if (tiisg == 0) {
+            buf[sgitg] = max_val;
+        }
+
+        threadgroup_barrier(mem_flags::mem_threadgroup);
+
+        max_val = buf[tiisg];
+        max_val = simd_max(max_val);
+    }
+
+    // parallel sum
+    float lsum = 0.0f;
+    for (int i00 = tpitg; i00 < ne00; i00 += ntg) {
+        const float exp_psrc0 = exp((psrc0[i00]*scale + (pmask ? slope*pmask[i00] : 0.0f)) - max_val);
+        lsum += exp_psrc0;
+        pdst[i00] = exp_psrc0;
+    }
+
+    // This barrier fixes a failing test
+    // ref: https://github.com/ggerganov/ggml/pull/621#discussion_r1425156335
+    threadgroup_barrier(mem_flags::mem_none);
+
+    float sum = simd_sum(lsum);
+
+    if (ntg > N_SIMDWIDTH) {
+        if (sgitg == 0) {
+            buf[tiisg] = 0.0f;
+        }
+
+        threadgroup_barrier(mem_flags::mem_threadgroup);
+
+        if (tiisg == 0) {
+            buf[sgitg] = sum;
+        }
+
+        threadgroup_barrier(mem_flags::mem_threadgroup);
+
+        sum = buf[tiisg];
+        sum = simd_sum(sum);
+    }
+
+    const float inv_sum = 1.0f/sum;
+
+    for (int i00 = tpitg; i00 < ne00; i00 += ntg) {
+        pdst[i00] *= inv_sum;
+    }
+}
+
+template<typename T>
+kernel void kernel_soft_max_4(
+        device const  char * src0,
+        device const  char * src1,
+        device        char * dst,
+        constant   int64_t & ne00,
+        constant   int64_t & ne01,
+        constant   int64_t & ne02,
+        constant     float & scale,
+        constant     float & max_bias,
+        constant     float & m0,
+        constant     float & m1,
+        constant  uint32_t & n_head_log2,
+        threadgroup  float * buf [[threadgroup(0)]],
+        uint  tgpig[[threadgroup_position_in_grid]],
+        uint  tpitg[[thread_position_in_threadgroup]],
+        uint  sgitg[[simdgroup_index_in_threadgroup]],
+        uint  tiisg[[thread_index_in_simdgroup]],
+        uint    ntg[[threads_per_threadgroup]]) {
+    const int64_t i03 = (tgpig) / (ne02*ne01);
+    const int64_t i02 = (tgpig - i03*ne02*ne01) / ne01;
+    const int64_t i01 = (tgpig - i03*ne02*ne01 - i02*ne01);
+
+    device const float4 * psrc4 = (device const float4 *) src0 + (i03*ne02*ne01*ne00 + i02*ne01*ne00 + i01*ne00)/4;
+    device const      T * pmask = src1 != src0 ? (device const     T *) src1         + i01*ne00/4 : nullptr;
+    device       float4 * pdst4 = (device       float4 *) dst  + (i03*ne02*ne01*ne00 + i02*ne01*ne00 + i01*ne00)/4;
+
+    float slope = 1.0f;
+
+    if (max_bias > 0.0f) {
+        const int64_t h = i02;
+
+        const float base = h < n_head_log2 ? m0 : m1;
+        const int   exp  = h < n_head_log2 ? h + 1 : 2*(h - n_head_log2) + 1;
+
+        slope = pow(base, exp);
+    }
+
+    // parallel max
+    float4 lmax4 = -INFINITY;
+
+    for (int i00 = tpitg; i00 < ne00/4; i00 += ntg) {
+        lmax4 = fmax(lmax4, psrc4[i00]*scale + (float4)((pmask ? slope*pmask[i00] : 0.0f)));
+    }
+
+    const float lmax = MAX(MAX(lmax4[0], lmax4[1]), MAX(lmax4[2], lmax4[3]));
+
+    float max_val = simd_max(lmax);
+    if (ntg > N_SIMDWIDTH) {
+        if (sgitg == 0) {
+            buf[tiisg] = -INFINITY;
+        }
+
+        threadgroup_barrier(mem_flags::mem_threadgroup);
+
+        if (tiisg == 0) {
+            buf[sgitg] = max_val;
+        }
+
+        threadgroup_barrier(mem_flags::mem_threadgroup);
+
+        max_val = buf[tiisg];
+        max_val = simd_max(max_val);
+    }
+
+    // parallel sum
+    float4 lsum4 = 0.0f;
+    for (int i00 = tpitg; i00 < ne00/4; i00 += ntg) {
+        const float4 exp_psrc4 = exp((psrc4[i00]*scale + (float4)((pmask ? slope*pmask[i00] : 0.0f))) - max_val);
+        lsum4 += exp_psrc4;
+        pdst4[i00] = exp_psrc4;
+    }
+
+    const float lsum = lsum4[0] + lsum4[1] + lsum4[2] + lsum4[3];
+
+    // This barrier fixes a failing test
+    // ref: https://github.com/ggerganov/ggml/pull/621#discussion_r1425156335
+    threadgroup_barrier(mem_flags::mem_none);
+
+    float sum = simd_sum(lsum);
+
+    if (ntg > N_SIMDWIDTH) {
+        if (sgitg == 0) {
+            buf[tiisg] = 0.0f;
+        }
+
+        threadgroup_barrier(mem_flags::mem_threadgroup);
+
+        if (tiisg == 0) {
+            buf[sgitg] = sum;
+        }
+
+        threadgroup_barrier(mem_flags::mem_threadgroup);
+
+        sum = buf[tiisg];
+        sum = simd_sum(sum);
+    }
+
+    const float inv_sum = 1.0f/sum;
+
+    for (int i00 = tpitg; i00 < ne00/4; i00 += ntg) {
+        pdst4[i00] *= inv_sum;
+    }
+}
+
+typedef decltype(kernel_soft_max<float>)    kernel_soft_max_t;
+typedef decltype(kernel_soft_max_4<float4>) kernel_soft_max_4_t;
+
+template [[host_name("kernel_soft_max_f16")]]   kernel kernel_soft_max_t   kernel_soft_max<half>;
+template [[host_name("kernel_soft_max_f32")]]   kernel kernel_soft_max_t   kernel_soft_max<float>;
+template [[host_name("kernel_soft_max_f16_4")]] kernel kernel_soft_max_4_t kernel_soft_max_4<half4>;
+template [[host_name("kernel_soft_max_f32_4")]] kernel kernel_soft_max_4_t kernel_soft_max_4<float4>;
+
+kernel void kernel_diag_mask_inf(
+        device const float * src0,
+        device       float * dst,
+        constant   int64_t & ne00,
+        constant   int64_t & ne01,
+        constant       int & n_past,
+        uint3 tpig[[thread_position_in_grid]]) {
+    const int64_t i02 = tpig[2];
+    const int64_t i01 = tpig[1];
+    const int64_t i00 = tpig[0];
+
+    if (i00 > n_past + i01) {
+        dst[i02*ne01*ne00 + i01*ne00 + i00] = -INFINITY;
+    } else {
+        dst[i02*ne01*ne00 + i01*ne00 + i00] = src0[i02*ne01*ne00 + i01*ne00 + i00];
+    }
+}
+
+kernel void kernel_diag_mask_inf_8(
+        device const float4 * src0,
+        device       float4 * dst,
+        constant    int64_t & ne00,
+        constant    int64_t & ne01,
+        constant        int & n_past,
+        uint3 tpig[[thread_position_in_grid]]) {
+
+    const int64_t i = 2*tpig[0];
+
+    dst[i+0] = src0[i+0];
+    dst[i+1] = src0[i+1];
+    int64_t i4 = 4*i;
+    const int64_t i02 = i4/(ne00*ne01); i4 -= i02*ne00*ne01;
+    const int64_t i01 = i4/(ne00);      i4 -= i01*ne00;
+    const int64_t i00 = i4;
+    for (int k = 3; k >= 0; --k) {
+        if (i00 + 4 + k <= n_past + i01) {
+            break;
+        }
+        dst[i+1][k] = -INFINITY;
+        if (i00 + k > n_past + i01) {
+            dst[i][k] = -INFINITY;
+        }
+    }
+}
+
+// ref: ggml.c:ggml_compute_forward_ssm_conv_f32
+// TODO: optimize
+kernel void kernel_ssm_conv_f32(
+        device const  void * src0,
+        device const  void * src1,
+        device       float * dst,
+        constant   int64_t & ne00,
+        constant   int64_t & ne01,
+        constant   int64_t & ne02,
+        constant  uint64_t & nb00,
+        constant  uint64_t & nb01,
+        constant  uint64_t & nb02,
+        constant   int64_t & ne10,
+        constant   int64_t & ne11,
+        constant  uint64_t & nb10,
+        constant  uint64_t & nb11,
+        constant   int64_t & ne0,
+        constant   int64_t & ne1,
+        constant   int64_t & ne2,
+        constant  uint64_t & nb0,
+        constant  uint64_t & nb1,
+        constant  uint64_t & nb2,
+        uint3 tgpig[[threadgroup_position_in_grid]],
+        uint3 tpitg[[thread_position_in_threadgroup]],
+        uint3   ntg[[threads_per_threadgroup]]) {
+    const int64_t ir = tgpig.x;
+    const int64_t i2 = tgpig.y;
+    const int64_t i3 = tgpig.z;
+
+    const int64_t nc  = ne10;
+    const int64_t ncs = ne00;
+    const int64_t nr  = ne01;
+    const int64_t n_t = ne1;
+    const int64_t n_s = ne2;
+
+    device const float * s = (device const float *) ((device const char *) src0 + ir*nb01 + i2*nb00 + i3*nb02);
+    device const float * c = (device const float *) ((device const char *) src1 + ir*nb11);
+    device       float * x = (device       float *) ((device       char *) dst  + ir*nb0  + i2*nb1  + i3*nb2);
+
+    float sumf = 0.0f;
+
+    for (int64_t i0 = 0; i0 < nc; ++i0) {
+        sumf += s[i0] * c[i0];
+    }
+
+    x[0] = sumf;
+}
+
+// ref: ggml.c:ggml_compute_forward_ssm_scan_f32
+// TODO: optimize
+kernel void kernel_ssm_scan_f32(
+        device const void * src0,
+        device const void * src1,
+        device const void * src2,
+        device const void * src3,
+        device const void * src4,
+        device const void * src5,
+        device      float * dst,
+        constant  int64_t & d_state,
+        constant  int64_t & d_inner,
+        constant  int64_t & n_seq_tokens,
+        constant  int64_t & n_seqs,
+        constant uint64_t & nb00,
+        constant uint64_t & nb01,
+        constant uint64_t & nb02,
+        constant uint64_t & nb10,
+        constant uint64_t & nb11,
+        constant uint64_t & nb12,
+        constant uint64_t & nb13,
+        constant uint64_t & nb20,
+        constant uint64_t & nb21,
+        constant uint64_t & nb22,
+        constant uint64_t & nb30,
+        constant uint64_t & nb31,
+        constant uint64_t & nb40,
+        constant uint64_t & nb41,
+        constant uint64_t & nb42,
+        constant uint64_t & nb50,
+        constant uint64_t & nb51,
+        constant uint64_t & nb52,
+        uint3 tgpig[[threadgroup_position_in_grid]],
+        uint3 tpitg[[thread_position_in_threadgroup]],
+        uint3   ntg[[threads_per_threadgroup]]) {
+    const int64_t ir = tgpig.x;
+    const int64_t i3 = tgpig.y;
+
+    const int64_t nc  = d_state;
+    const int64_t nr  = d_inner;
+    const int64_t n_t = n_seq_tokens;
+    const int64_t n_s = n_seqs;
+
+    for (int64_t i2 = 0; i2 < n_t; ++i2) {
+        device const float * s0 = (device const float *) ((device const char *) src0 + ir*nb01 + i3*nb02);
+        device const float * x  = (device const float *) ((device const char *) src1 + ir*nb10 + i2*nb11 + i3*nb12);
+        device const float * dt = (device const float *) ((device const char *) src2 + ir*nb20 + i2*nb21 + i3*nb22);
+        device const float * A  = (device const float *) ((device const char *) src3 + ir*nb31);
+        device const float * B  = (device const float *) ((device const char *) src4 + i2*nb41 + i3*nb42);
+        device const float * C  = (device const float *) ((device const char *) src5 + i2*nb51 + i3*nb52);
+        device       float * y  = (device       float *) ((device       char *) dst  + ir*nb10 + i2*nb11 + i3*nb12); // TODO: do not use src1 strides
+        device       float * s  = (device       float *) ((device       char *) dst  + ir*nb01 + i3*nb02 +    nb13);
+
+        if (i2 > 0) {
+            s0 = s;
+        }
+
+        // i1 == 0
+        float dt_soft_plus = dt[0] <= 20.0f ? log(1.0f + exp(dt[0])) : dt[0];
+        float x_dt = x[0] * dt_soft_plus;
+        float sumf = 0.0f;
+
+        for (int64_t i0 = 0; i0 < nc; ++i0) {
+            int64_t i = i0;
+            float state = (s0[i] * exp(dt_soft_plus * A[i])) + (B[i0] * x_dt);
+            sumf += state * C[i0];
+            s[i] = state;
+        }
+
+        y[0] = sumf;
+    }
+}
+
+kernel void kernel_norm(
+        device const  void * src0,
+        device       float * dst,
+        constant   int64_t & ne00,
+        constant  uint64_t & nb01,
+        constant     float & eps,
+        threadgroup float  * sum [[threadgroup(0)]],
+        uint tgpig[[threadgroup_position_in_grid]],
+        uint tpitg[[thread_position_in_threadgroup]],
+        uint   ntg[[threads_per_threadgroup]]) {
+    device const float * x = (device const float *) ((device const char *) src0 + tgpig*nb01);
+    // MEAN
+    // parallel sum
+    sum[tpitg] = 0.0f;
+    for (int i00 = tpitg; i00 < ne00; i00 += ntg) {
+        sum[tpitg] += x[i00];
+    }
+    // reduce
+    threadgroup_barrier(mem_flags::mem_threadgroup);
+    for (uint i = ntg/2; i > 0; i /= 2) {
+        if (tpitg < i) {
+            sum[tpitg] += sum[tpitg + i];
+        }
+        threadgroup_barrier(mem_flags::mem_threadgroup);
+    }
+    const float mean  = sum[0] / ne00;
+
+    // recenter and VARIANCE
+    threadgroup_barrier(mem_flags::mem_threadgroup);
+    device float * y = dst + tgpig*ne00;
+    sum[tpitg] = 0.0f;
+    for (int i00 = tpitg; i00 < ne00; i00 += ntg) {
+        y[i00] = x[i00] - mean;
+        sum[tpitg] += y[i00] * y[i00];
+    }
+
+    // reduce
+    threadgroup_barrier(mem_flags::mem_threadgroup);
+    for (uint i = ntg/2; i > 0; i /= 2) {
+        if (tpitg < i) {
+            sum[tpitg] += sum[tpitg + i];
+        }
+        threadgroup_barrier(mem_flags::mem_threadgroup);
+    }
+    const float variance = sum[0] / ne00;
+
+    const float scale = 1.0f/sqrt(variance + eps);
+    for (int i00 = tpitg; i00 < ne00; i00 += ntg) {
+        y[i00] = y[i00] * scale;
+    }
+}
+
+kernel void kernel_rms_norm(
+        device const  void * src0,
+        device       float * dst,
+        constant   int64_t & ne00,
+        constant  uint64_t & nb01,
+        constant     float & eps,
+        threadgroup float  * buf [[threadgroup(0)]],
+        uint tgpig[[threadgroup_position_in_grid]],
+        uint tpitg[[thread_position_in_threadgroup]],
+        uint sgitg[[simdgroup_index_in_threadgroup]],
+        uint tiisg[[thread_index_in_simdgroup]],
+        uint   ntg[[threads_per_threadgroup]]) {
+    device const float4 * x = (device const float4 *) ((device const char *) src0 + tgpig*nb01);
+
+    float4 sumf = 0;
+    float all_sum = 0;
+
+    // parallel sum
+    for (int i00 = tpitg; i00 < ne00/4; i00 += ntg) {
+        sumf += x[i00] * x[i00];
+    }
+    all_sum = sumf[0] + sumf[1] + sumf[2] + sumf[3];
+    all_sum = simd_sum(all_sum);
+    if (ntg > N_SIMDWIDTH) {
+        if (sgitg == 0) {
+            buf[tiisg] = 0.0f;
+        }
+
+        threadgroup_barrier(mem_flags::mem_threadgroup);
+
+        if (tiisg == 0) {
+            buf[sgitg] = all_sum;
+        }
+
+        threadgroup_barrier(mem_flags::mem_threadgroup);
+
+        all_sum = buf[tiisg];
+        all_sum = simd_sum(all_sum);
+    }
+
+    const float mean  = all_sum/ne00;
+    const float scale = 1.0f/sqrt(mean + eps);
+
+    device float4 * y = (device float4 *) (dst + tgpig*ne00);
+    for (int i00 = tpitg; i00 < ne00/4; i00 += ntg) {
+        y[i00] = x[i00] * scale;
+    }
+}
+
+kernel void kernel_group_norm(
+        device const float * src0,
+        device       float * dst,
+        constant   int64_t & ne00,
+        constant   int64_t & ne01,
+        constant   int64_t & ne02,
+        constant  uint64_t & nb00,
+        constant  uint64_t & nb01,
+        constant  uint64_t & nb02,
+        constant   int32_t & n_groups,
+        constant     float & eps,
+        threadgroup float  * buf [[threadgroup(0)]],
+        uint tgpig[[threadgroup_position_in_grid]],
+        uint tpitg[[thread_position_in_threadgroup]],
+        uint sgitg[[simdgroup_index_in_threadgroup]],
+        uint tiisg[[thread_index_in_simdgroup]],
+        uint   ntg[[threads_per_threadgroup]]) {
+    const int64_t ne = ne00*ne01*ne02;
+    const int64_t gs = ne00*ne01*((ne02 + n_groups - 1) / n_groups);
+
+    int start = tgpig * gs;
+    int end   = start + gs;
+
+    start += tpitg;
+
+    if (end >= ne) {
+        end = ne;
+    }
+
+    float tmp = 0.0f; // partial sum for thread in warp
+
+    for (int j = start; j < end; j += ntg) {
+        tmp += src0[j];
+    }
+
+    threadgroup_barrier(mem_flags::mem_threadgroup);
+    tmp = simd_sum(tmp);
+    if (ntg > N_SIMDWIDTH) {
+        if (sgitg == 0) {
+            buf[tiisg] = 0.0f;
+        }
+
+        threadgroup_barrier(mem_flags::mem_threadgroup);
+
+        if (tiisg == 0) {
+            buf[sgitg] = tmp;
+        }
+
+        threadgroup_barrier(mem_flags::mem_threadgroup);
+
+        tmp = buf[tiisg];
+        tmp = simd_sum(tmp);
+    }
+
+    const float mean = tmp / gs;
+    tmp = 0.0f;
+
+    for (int j = start; j < end; j += ntg) {
+        float xi = src0[j] - mean;
+        dst[j] = xi;
+        tmp += xi * xi;
+    }
+
+    tmp = simd_sum(tmp);
+    if (ntg > N_SIMDWIDTH) {
+        if (sgitg == 0) {
+            buf[tiisg] = 0.0f;
+        }
+
+        threadgroup_barrier(mem_flags::mem_threadgroup);
+
+        if (tiisg == 0) {
+            buf[sgitg] = tmp;
+        }
+
+        threadgroup_barrier(mem_flags::mem_threadgroup);
+
+        tmp = buf[tiisg];
+        tmp = simd_sum(tmp);
+    }
+
+    const float variance = tmp / gs;
+    const float scale = 1.0f/sqrt(variance + eps);
+    for (int j = start; j < end; j += ntg) {
+        dst[j] *= scale;
+    }
+}
+
+// function for calculate inner product between half a q4_0 block and 16 floats (yl), sumy is SUM(yl[i])
+// il indicates where the q4 quants begin (0 or QK4_0/4)
+// we assume that the yl's have been multiplied with the appropriate scale factor
+// that corresponds to the missing bit shifts (1, 1/16, 1/256, 1/4096)
+inline float block_q_n_dot_y(device const block_q4_0 * qb_curr, float sumy, thread float * yl, int il) {
+    float d = qb_curr->d;
+
+    float2 acc = 0.f;
+
+    device const uint16_t * qs = ((device const uint16_t *)qb_curr + 1 + il/2);
+
+    for (int i = 0; i < 8; i+=2) {
+        acc[0] += yl[i + 0] * (qs[i / 2] & 0x000F)
+                + yl[i + 1] * (qs[i / 2] & 0x0F00);
+        acc[1] += yl[i + 8] * (qs[i / 2] & 0x00F0)
+                + yl[i + 9] * (qs[i / 2] & 0xF000);
+    }
+    return d * (sumy * -8.f + acc[0] + acc[1]);
+}
+
+// function for calculate inner product between half a q4_1 block and 16 floats (yl), sumy is SUM(yl[i])
+// il indicates where the q4 quants begin (0 or QK4_0/4)
+// we assume that the yl's have been multiplied with the appropriate scale factor
+// that corresponds to the missing bit shifts (1, 1/16, 1/256, 1/4096)
+inline float block_q_n_dot_y(device const block_q4_1 * qb_curr, float sumy, thread float * yl, int il) {
+    float d = qb_curr->d;
+    float m = qb_curr->m;
+
+    float2 acc = 0.f;
+
+    device const uint16_t * qs = ((device const uint16_t *)qb_curr + 2 + il/2);
+
+    for (int i = 0; i < 8; i+=2) {
+        acc[0] += yl[i + 0] * (qs[i / 2] & 0x000F)
+                + yl[i + 1] * (qs[i / 2] & 0x0F00);
+        acc[1] += yl[i + 8] * (qs[i / 2] & 0x00F0)
+                + yl[i + 9] * (qs[i / 2] & 0xF000);
+    }
+    return d * (acc[0] + acc[1]) + sumy * m;
+}
+
+// function for calculate inner product between half a q5_0 block and 16 floats (yl), sumy is SUM(yl[i])
+// il indicates where the q5 quants begin (0 or QK5_0/4)
+// we assume that the yl's have been multiplied with the appropriate scale factor
+// that corresponds to the missing bit shifts (1, 1/16, 1/256, 1/4096)
+inline float block_q_n_dot_y(device const block_q5_0 * qb_curr, float sumy, thread float * yl, int il) {
+    float d = qb_curr->d;
+
+    float2 acc = 0.f;
+
+    device const uint16_t * qs =  ((device const uint16_t *)qb_curr + 3 + il/2);
+           const uint32_t   qh = *((device const uint32_t *)qb_curr->qh);
+
+    for (int i = 0; i < 8; i+=2) {
+        acc[0] += yl[i + 0] * ((qs[i / 2] & 0x000F) | ((qh >> (i+0+il        ) << 4 ) & 0x00010))
+                + yl[i + 1] * ((qs[i / 2] & 0x0F00) | ((qh >> (i+1+il        ) << 12) & 0x01000));
+        acc[1] += yl[i + 8] * ((qs[i / 2] & 0x00F0) | ((qh >> (i+0+il+QK5_0/2) << 8 ) & 0x00100))
+                + yl[i + 9] * ((qs[i / 2] & 0xF000) | ((qh >> (i+1+il+QK5_0/2) << 16) & 0x10000));
+    }
+    return d * (sumy * -16.f + acc[0] + acc[1]);
+}
+
+// function for calculate inner product between half a q5_1 block and 16 floats (yl), sumy is SUM(yl[i])
+// il indicates where the q5 quants begin (0 or QK5_1/4)
+// we assume that the yl's have been multiplied with the appropriate scale factor
+// that corresponds to the missing bit shifts (1, 1/16, 1/256, 1/4096)
+inline float block_q_n_dot_y(device const block_q5_1 * qb_curr, float sumy, thread float * yl, int il) {
+    float d = qb_curr->d;
+    float m = qb_curr->m;
+
+    float2 acc = 0.f;
+
+    device const uint16_t * qs =  ((device const uint16_t *)qb_curr + 4 + il/2);
+           const uint32_t   qh = *((device const uint32_t *)qb_curr->qh);
+
+    for (int i = 0; i < 8; i+=2) {
+        acc[0] += yl[i + 0] * ((qs[i / 2] & 0x000F) | ((qh >> (i+0+il        ) << 4 ) & 0x00010))
+                + yl[i + 1] * ((qs[i / 2] & 0x0F00) | ((qh >> (i+1+il        ) << 12) & 0x01000));
+        acc[1] += yl[i + 8] * ((qs[i / 2] & 0x00F0) | ((qh >> (i+0+il+QK5_0/2) << 8 ) & 0x00100))
+                + yl[i + 9] * ((qs[i / 2] & 0xF000) | ((qh >> (i+1+il+QK5_0/2) << 16) & 0x10000));
+    }
+    return d * (acc[0] + acc[1]) + sumy * m;
+}
+
+// putting them in the kernel cause a significant performance penalty
+#define N_DST 4        // each SIMD group works on 4 rows
+#define N_SIMDGROUP 2  // number of SIMD groups in a thread group
+//Note: This is a template, but strictly speaking it only applies to
+//      quantizations where the block size is 32. It also does not
+//      guard against the number of rows not being divisible by
+//      N_DST, so this is another explicit assumption of the implementation.
+template<typename block_q_type, int nr, int nsg, int nw>
+void mul_vec_q_n_f32_impl(
+        device const void  * src0,
+        device const float * src1,
+        device       float * dst,
+                   int64_t   ne00,
+                   int64_t   ne01,
+                   int64_t   ne02,
+                   int64_t   ne10,
+                   int64_t   ne12,
+                   int64_t   ne0,
+                   int64_t   ne1,
+                   uint      r2,
+                   uint      r3,
+        threadgroup int8_t * shared_values,
+                   uint3 tgpig, uint tiisg, uint sgitg) {
+    const int nb = ne00/QK4_0;
+
+    const int r0 = tgpig.x;
+    const int r1 = tgpig.y;
+    const int im = tgpig.z;
+
+    const int first_row = (r0 * nsg + sgitg) * nr;
+
+    const uint i12 = im%ne12;
+    const uint i13 = im/ne12;
+
+    const uint offset0 = first_row * nb + (i12/r2)*(nb*ne01) + (i13/r3)*(nb*ne01*ne02);
+
+    device const block_q_type * x = (device const block_q_type *) src0 + offset0;
+    device const float        * y = (device const float        *) src1 + r1*ne10 + im*ne00*ne1;
+
+    float yl[16]; // src1 vector cache
+    float sumf[nr] = {0.f};
+
+    const int ix = (tiisg/2);
+    const int il = (tiisg%2)*8;
+
+    device const float * yb = y + ix * QK4_0 + il;
+
+    // each thread in a SIMD group deals with half a block.
+    for (int ib = ix; ib < nb; ib += nw/2) {
+        float sumy = 0;
+        for (int i = 0; i < 8; i += 2) {
+            sumy += yb[i] + yb[i+1];
+            yl[i+0] = yb[i+ 0];
+            yl[i+1] = yb[i+ 1]/256.f;
+
+            sumy += yb[i+16] + yb[i+17];
+            yl[i+8] = yb[i+16]/16.f;
+            yl[i+9] = yb[i+17]/4096.f;
+        }
+
+        for (int row = 0; row < nr; row++) {
+            sumf[row] += block_q_n_dot_y(x+ib+row*nb, sumy, yl, il);
+        }
+
+        yb += QK4_0 * 16;
+    }
+
+    for (int row = 0; row < nr; ++row) {
+        const float tot = simd_sum(sumf[row]);
+        if (tiisg == 0 && first_row + row < ne01) {
+            dst[im*ne0*ne1 + r1*ne0 + first_row + row] = tot;
+        }
+    }
+}
+
+kernel void kernel_mul_mv_q4_0_f32(
+        device const  void * src0,
+        device const float * src1,
+        device       float * dst,
+        constant   int64_t & ne00,
+        constant   int64_t & ne01,
+        constant   int64_t & ne02,
+        constant  uint64_t & nb00,
+        constant  uint64_t & nb01,
+        constant  uint64_t & nb02,
+        constant   int64_t & ne10,
+        constant   int64_t & ne11,
+        constant   int64_t & ne12,
+        constant  uint64_t & nb10,
+        constant  uint64_t & nb11,
+        constant  uint64_t & nb12,
+        constant   int64_t & ne0,
+        constant   int64_t & ne1,
+        constant   uint    & r2,
+        constant   uint    & r3,
+        uint3 tgpig[[threadgroup_position_in_grid]],
+        uint  tiisg[[thread_index_in_simdgroup]],
+        uint  sgitg[[simdgroup_index_in_threadgroup]]) {
+    mul_vec_q_n_f32_impl<block_q4_0, N_DST, N_SIMDGROUP, N_SIMDWIDTH>(src0,src1,dst,ne00,ne01,ne02,ne10,ne12,ne0,ne1,r2,r3,nullptr,tgpig,tiisg,sgitg);
+}
+
+kernel void kernel_mul_mv_q4_1_f32(
+        device const  void * src0,
+        device const float * src1,
+        device       float * dst,
+        constant   int64_t & ne00,
+        constant   int64_t & ne01,
+        constant   int64_t & ne02,
+        constant  uint64_t & nb00,
+        constant  uint64_t & nb01,
+        constant  uint64_t & nb02,
+        constant   int64_t & ne10,
+        constant   int64_t & ne11,
+        constant   int64_t & ne12,
+        constant  uint64_t & nb10,
+        constant  uint64_t & nb11,
+        constant  uint64_t & nb12,
+        constant   int64_t & ne0,
+        constant   int64_t & ne1,
+        constant   uint    & r2,
+        constant   uint    & r3,
+        uint3 tgpig[[threadgroup_position_in_grid]],
+        uint tiisg[[thread_index_in_simdgroup]],
+        uint sgitg[[simdgroup_index_in_threadgroup]]) {
+     mul_vec_q_n_f32_impl<block_q4_1, N_DST, N_SIMDGROUP, N_SIMDWIDTH>(src0,src1,dst,ne00,ne01,ne02,ne10,ne12,ne0,ne1,r2,r3,nullptr,tgpig,tiisg,sgitg);
+}
+
+kernel void kernel_mul_mv_q5_0_f32(
+        device const  void * src0,
+        device const float * src1,
+        device       float * dst,
+        constant   int64_t & ne00,
+        constant   int64_t & ne01,
+        constant   int64_t & ne02,
+        constant  uint64_t & nb00,
+        constant  uint64_t & nb01,
+        constant  uint64_t & nb02,
+        constant   int64_t & ne10,
+        constant   int64_t & ne11,
+        constant   int64_t & ne12,
+        constant  uint64_t & nb10,
+        constant  uint64_t & nb11,
+        constant  uint64_t & nb12,
+        constant   int64_t & ne0,
+        constant   int64_t & ne1,
+        constant   uint    & r2,
+        constant   uint    & r3,
+        uint3 tgpig[[threadgroup_position_in_grid]],
+        uint  tiisg[[thread_index_in_simdgroup]],
+        uint  sgitg[[simdgroup_index_in_threadgroup]]) {
+    mul_vec_q_n_f32_impl<block_q5_0, N_DST, N_SIMDGROUP, N_SIMDWIDTH>(src0,src1,dst,ne00,ne01,ne02,ne10,ne12,ne0,ne1,r2,r3,nullptr,tgpig,tiisg,sgitg);
+}
+
+kernel void kernel_mul_mv_q5_1_f32(
+        device const  void * src0,
+        device const float * src1,
+        device       float * dst,
+        constant   int64_t & ne00,
+        constant   int64_t & ne01,
+        constant   int64_t & ne02,
+        constant  uint64_t & nb00,
+        constant  uint64_t & nb01,
+        constant  uint64_t & nb02,
+        constant   int64_t & ne10,
+        constant   int64_t & ne11,
+        constant   int64_t & ne12,
+        constant  uint64_t & nb10,
+        constant  uint64_t & nb11,
+        constant  uint64_t & nb12,
+        constant   int64_t & ne0,
+        constant   int64_t & ne1,
+        constant   uint    & r2,
+        constant   uint    & r3,
+        uint3 tgpig[[threadgroup_position_in_grid]],
+        uint  tiisg[[thread_index_in_simdgroup]],
+        uint  sgitg[[simdgroup_index_in_threadgroup]]) {
+    mul_vec_q_n_f32_impl<block_q5_1, N_DST, N_SIMDGROUP, N_SIMDWIDTH>(src0,src1,dst,ne00,ne01,ne02,ne10,ne12,ne0,ne1,r2,r3,nullptr,tgpig,tiisg,sgitg);
+}
+
+
+#define NB_Q8_0 8
+
+void kernel_mul_mv_q8_0_f32_impl(
+        device const  void * src0,
+        device const float * src1,
+        device       float * dst,
+                   int64_t   ne00,
+                   int64_t   ne01,
+                   int64_t   ne02,
+                   int64_t   ne10,
+                   int64_t   ne12,
+                   int64_t   ne0,
+                   int64_t   ne1,
+                   uint      r2,
+                   uint      r3,
+        threadgroup int8_t * shared_values,
+                   uint3     tgpig,
+                   uint      tiisg,
+                   uint      sgitg) {
+    const int nr  = N_DST;
+    const int nsg = N_SIMDGROUP;
+    const int nw  = N_SIMDWIDTH;
+
+    const int nb = ne00/QK8_0;
+    const int r0 = tgpig.x;
+    const int r1 = tgpig.y;
+    const int im = tgpig.z;
+
+    const int first_row = (r0 * nsg + sgitg) * nr;
+
+    const uint i12 = im%ne12;
+    const uint i13 = im/ne12;
+
+    const uint offset0 = first_row * nb + (i12/r2)*(nb*ne01) + (i13/r3)*(nb*ne01*ne02);
+
+    device const block_q8_0 * x = (device const block_q8_0 *) src0 + offset0;
+    device const float      * y = (device const float      *) src1 + r1*ne10 + im*ne00*ne1;
+
+    float yl[NB_Q8_0];
+    float sumf[nr]={0.f};
+
+    const int ix = tiisg/4;
+    const int il = tiisg%4;
+
+    device const float * yb = y + ix * QK8_0 + NB_Q8_0*il;
+
+    // each thread in a SIMD group deals with NB_Q8_0 quants at a time
+    for (int ib = ix; ib < nb; ib += nw/4) {
+        for (int i = 0; i < NB_Q8_0; ++i) {
+            yl[i] = yb[i];
+        }
+
+        for (int row = 0; row < nr; row++) {
+            device const int8_t * qs = x[ib+row*nb].qs + NB_Q8_0*il;
+            float sumq = 0.f;
+            for (int iq = 0; iq < NB_Q8_0; ++iq) {
+                sumq += qs[iq] * yl[iq];
+            }
+            sumf[row] += sumq*x[ib+row*nb].d;
+        }
+
+        yb += NB_Q8_0 * nw;
+    }
+
+    for (int row = 0; row < nr; ++row) {
+        const float tot = simd_sum(sumf[row]);
+        if (tiisg == 0 && first_row + row < ne01) {
+            dst[r1*ne0 + im*ne0*ne1 + first_row + row] = tot;
+        }
+    }
+}
+
+[[host_name("kernel_mul_mv_q8_0_f32")]]
+kernel void kernel_mul_mv_q8_0_f32(
+        device const  void * src0,
+        device const float * src1,
+        device       float * dst,
+        constant   int64_t & ne00,
+        constant   int64_t & ne01,
+        constant   int64_t & ne02,
+        constant  uint64_t & nb00,
+        constant  uint64_t & nb01,
+        constant  uint64_t & nb02,
+        constant   int64_t & ne10,
+        constant   int64_t & ne11,
+        constant   int64_t & ne12,
+        constant  uint64_t & nb10,
+        constant  uint64_t & nb11,
+        constant  uint64_t & nb12,
+        constant   int64_t & ne0,
+        constant   int64_t & ne1,
+        constant   uint    & r2,
+        constant   uint    & r3,
+        uint3 tgpig[[threadgroup_position_in_grid]],
+        uint  tiisg[[thread_index_in_simdgroup]],
+        uint  sgitg[[simdgroup_index_in_threadgroup]]) {
+    kernel_mul_mv_q8_0_f32_impl(src0,src1,dst,ne00,ne01,ne02,ne10,ne12,ne0,ne1,r2,r3,nullptr,tgpig,tiisg,sgitg);
+}
+
+#define N_MV_T_T 4
+
+template<typename T0, typename T04, typename T1, typename T14>
+void kernel_mul_mv_impl(
+        device const  char * src0,
+        device const  char * src1,
+        device       float * dst,
+                   int64_t   ne00,
+                   int64_t   ne01,
+                   int64_t   ne02,
+                  uint64_t   nb00,
+                  uint64_t   nb01,
+                  uint64_t   nb02,
+                   int64_t   ne10,
+                   int64_t   ne11,
+                   int64_t   ne12,
+                  uint64_t   nb10,
+                  uint64_t   nb11,
+                  uint64_t   nb12,
+                   int64_t   ne0,
+                   int64_t   ne1,
+                   uint      r2,
+                   uint      r3,
+                   uint3     tgpig,
+                   uint      tiisg) {
+    const int64_t r0 = tgpig.x;
+    const int64_t rb = tgpig.y*N_MV_T_T;
+    const int64_t im = tgpig.z;
+
+    const uint i12 = im%ne12;
+    const uint i13 = im/ne12;
+
+    const uint offset0 = r0*nb01 + (i12/r2)*nb02 + (i13/r3)*nb02*ne02;
+
+    device const T0 * x = (device const T0 *) (src0 + offset0);
+
+    if (ne00 < 128) {
+        for (int row = 0; row < N_MV_T_T; ++row) {
+            int r1 = rb + row;
+            if (r1 >= ne11) {
+                break;
+            }
+
+            device const T1 * y = (device const T1 *) (src1 + r1*nb11 + im*nb12);
+
+            float sumf = 0;
+            for (int i = tiisg; i < ne00; i += 32) {
+                sumf += (T0) x[i] * (T1) y[i];
+            }
+
+            float all_sum = simd_sum(sumf);
+            if (tiisg == 0) {
+                dst[im*ne1*ne0 + r1*ne0 + r0] = all_sum;
+            }
+        }
+    } else {
+        device const T04 * x4 = (device const T04 *) x;
+        for (int row = 0; row < N_MV_T_T; ++row) {
+            int r1 = rb + row;
+            if (r1 >= ne11) {
+                break;
+            }
+
+            device const T1  * y  = (device const T1  *) (src1 + r1*nb11 + im*nb12);
+            device const T14 * y4 = (device const T14 *) y;
+
+            float sumf = 0;
+            for (int i = tiisg; i < ne00/4; i += 32) {
+                for (int k = 0; k < 4; ++k) sumf += (float) (x4[i][k] * y4[i][k]);
+            }
+
+            float all_sum = simd_sum(sumf);
+            if (tiisg == 0) {
+                for (int i = 4*(ne00/4); i < ne00; ++i) all_sum += (float) (x[i] * y[i]);
+                dst[im*ne1*ne0 + r1*ne0 + r0] = all_sum;
+            }
+        }
+    }
+}
+
+template<typename T0, typename T04, typename T1, typename T14>
+kernel void kernel_mul_mv(
+        device const  char * src0,
+        device const  char * src1,
+        device       float * dst,
+        constant   int64_t & ne00,
+        constant   int64_t & ne01,
+        constant   int64_t & ne02,
+        constant  uint64_t & nb00,
+        constant  uint64_t & nb01,
+        constant  uint64_t & nb02,
+        constant   int64_t & ne10,
+        constant   int64_t & ne11,
+        constant   int64_t & ne12,
+        constant  uint64_t & nb10,
+        constant  uint64_t & nb11,
+        constant  uint64_t & nb12,
+        constant   int64_t & ne0,
+        constant   int64_t & ne1,
+        constant   uint    & r2,
+        constant   uint    & r3,
+        uint3 tgpig[[threadgroup_position_in_grid]],
+        uint  tiisg[[thread_index_in_simdgroup]]) {
+    kernel_mul_mv_impl<T0, T04, T1, T14>(
+        src0,
+        src1,
+        dst,
+        ne00,
+        ne01,
+        ne02,
+        nb00,
+        nb01,
+        nb02,
+        ne10,
+        ne11,
+        ne12,
+        nb10,
+        nb11,
+        nb12,
+        ne0,
+        ne1,
+        r2,
+        r3,
+        tgpig,
+        tiisg);
+}
+
+typedef decltype(kernel_mul_mv<half, half4, half, half4>) mul_mv_t;
+
+template [[host_name("kernel_mul_mv_f32_f32")]]   kernel mul_mv_t kernel_mul_mv<float,  float4,  float,  float4>;
+template [[host_name("kernel_mul_mv_f16_f32")]]   kernel mul_mv_t kernel_mul_mv<half,   half4,   float,  float4>;
+template [[host_name("kernel_mul_mv_f16_f16")]]   kernel mul_mv_t kernel_mul_mv<half,   half4,   half,   half4>;
+
+template<typename T, typename T4>
+kernel void kernel_mul_mv_1row(
+        device const  char * src0,
+        device const  char * src1,
+        device       float * dst,
+        constant   int64_t & ne00,
+        constant   int64_t & ne01,
+        constant   int64_t & ne02,
+        constant  uint64_t & nb00,
+        constant  uint64_t & nb01,
+        constant  uint64_t & nb02,
+        constant   int64_t & ne10,
+        constant   int64_t & ne11,
+        constant   int64_t & ne12,
+        constant  uint64_t & nb10,
+        constant  uint64_t & nb11,
+        constant  uint64_t & nb12,
+        constant   int64_t & ne0,
+        constant   int64_t & ne1,
+        constant   uint    & r2,
+        constant   uint    & r3,
+        uint3 tgpig[[threadgroup_position_in_grid]],
+        uint  tiisg[[thread_index_in_simdgroup]]) {
+
+    const int64_t r0 = tgpig.x;
+    const int64_t r1 = tgpig.y;
+    const int64_t im = tgpig.z;
+
+    const uint i12 = im%ne12;
+    const uint i13 = im/ne12;
+
+    const uint offset0 = r0*nb01 + (i12/r2)*nb02 + (i13/r3)*nb02*ne02;
+
+    device const T     * x = (device const T     *) (src0 + offset0);
+    device const float * y = (device const float *) (src1 + r1*nb11 + im*nb12);
+
+    float sumf = 0;
+    if (ne00 < 128) {
+        for (int i = tiisg; i < ne00; i += 32) {
+            sumf += (float) x[i] * (float) y[i];
+        }
+        float all_sum = simd_sum(sumf);
+        if (tiisg == 0) {
+            dst[im*ne1*ne0 + r1*ne0 + r0] = all_sum;
+        }
+    } else {
+        device const T4     * x4 = (device const T4     *) x;
+        device const float4 * y4 = (device const float4 *) y;
+
+        for (int i = tiisg; i < ne00/4; i += 32) {
+            for (int k = 0; k < 4; ++k) sumf += (float) (x4[i][k] * y4[i][k]);
+        }
+
+        float all_sum = simd_sum(sumf);
+
+        if (tiisg == 0) {
+            for (int i = 4*(ne00/4); i < ne00; ++i) all_sum += (float) (x[i] * y[i]);
+            dst[im*ne1*ne0 + r1*ne0 + r0] = all_sum;
+        }
+    }
+}
+
+typedef decltype(kernel_mul_mv_1row<half, half4>) mul_mv_1row_t;
+
+template [[host_name("kernel_mul_mv_f16_f32_1row")]]  kernel mul_mv_1row_t kernel_mul_mv_1row<half,   half4>;
+
+// Assumes row size (ne00) is a multiple of 4
+template<typename T, typename T4>
+kernel void kernel_mul_mv_l4(
+        device const  char * src0,
+        device const  char * src1,
+        device       float * dst,
+        constant   int64_t & ne00,
+        constant   int64_t & ne01,
+        constant   int64_t & ne02,
+        constant  uint64_t & nb00,
+        constant  uint64_t & nb01,
+        constant  uint64_t & nb02,
+        constant   int64_t & ne10,
+        constant   int64_t & ne11,
+        constant   int64_t & ne12,
+        constant  uint64_t & nb10,
+        constant  uint64_t & nb11,
+        constant  uint64_t & nb12,
+        constant   int64_t & ne0,
+        constant   int64_t & ne1,
+        constant   uint    & r2,
+        constant   uint    & r3,
+        uint3 tgpig[[threadgroup_position_in_grid]],
+        uint tiisg[[thread_index_in_simdgroup]]) {
+
+    const int nrows = ne11;
+    const int64_t r0 = tgpig.x;
+    const int64_t im = tgpig.z;
+
+    const uint i12 = im%ne12;
+    const uint i13 = im/ne12;
+
+    const uint offset0 = r0*nb01 + (i12/r2)*nb02 + (i13/r3)*nb02*ne02;
+
+    device const T4 * x4 = (device const T4 *) (src0 + offset0);
+
+    for (int r1 = 0; r1 < nrows; ++r1) {
+        device const float4 * y4 = (device const float4 *) (src1 + r1*nb11 + im*nb12);
+
+        float sumf = 0;
+        for (int i = tiisg; i < ne00/4; i += 32) {
+            for (int k = 0; k < 4; ++k) sumf += (float) (x4[i][k] * y4[i][k]);
+        }
+
+        float all_sum = simd_sum(sumf);
+        if (tiisg == 0) {
+            dst[im*ne1*ne0 + r1*ne0 + r0] = all_sum;
+        }
+    }
+}
+
+typedef decltype(kernel_mul_mv_l4<half, half4>) mul_mv_l4_t;
+
+template [[host_name("kernel_mul_mv_f16_f32_l4")]]  kernel mul_mv_l4_t kernel_mul_mv_l4<half, half4>;
+
+static float rope_yarn_ramp(const float low, const float high, const int i0) {
+    const float y = (i0 / 2 - low) / max(0.001f, high - low);
+    return 1.0f - min(1.0f, max(0.0f, y));
+}
+
+// YaRN algorithm based on LlamaYaRNScaledRotaryEmbedding.py from https://github.com/jquesnelle/yarn
+// MIT licensed. Copyright (c) 2023 Jeffrey Quesnelle and Bowen Peng.
+static void rope_yarn(
+    float theta_extrap, float freq_scale, float corr_dims[2], int64_t i0, float ext_factor, float mscale,
+    thread float * cos_theta, thread float * sin_theta) {
+    // Get n-d rotational scaling corrected for extrapolation
+    float theta_interp = freq_scale * theta_extrap;
+    float theta = theta_interp;
+    if (ext_factor != 0.0f) {
+        float ramp_mix = rope_yarn_ramp(corr_dims[0], corr_dims[1], i0) * ext_factor;
+        theta = theta_interp * (1 - ramp_mix) + theta_extrap * ramp_mix;
+
+        // Get n-d magnitude scaling corrected for interpolation
+        mscale *= 1.0f + 0.1f * log(1.0f / freq_scale);
+    }
+    *cos_theta = cos(theta) * mscale;
+    *sin_theta = sin(theta) * mscale;
+}
+
+// Apparently solving `n_rot = 2pi * x * base^((2 * max_pos_emb) / n_dims)` for x, we get
+// `corr_fac(n_rot) = n_dims * log(max_pos_emb / (n_rot * 2pi)) / (2 * log(base))`
+static float rope_yarn_corr_factor(int n_dims, int n_ctx_orig, float n_rot, float base) {
+    return n_dims * log(n_ctx_orig / (n_rot * 2 * M_PI_F)) / (2 * log(base));
+}
+
+static void rope_yarn_corr_dims(
+    int n_dims, int n_ctx_orig, float freq_base, float beta_fast, float beta_slow, float dims[2]
+) {
+    // start and end correction dims
+    dims[0] = max(0.0f,         floor(rope_yarn_corr_factor(n_dims, n_ctx_orig, beta_fast, freq_base)));
+    dims[1] = min(n_dims - 1.0f, ceil(rope_yarn_corr_factor(n_dims, n_ctx_orig, beta_slow, freq_base)));
+}
+
+template<typename T>
+kernel void kernel_rope_norm(
+        device const    void * src0,
+        device const int32_t * src1,
+        device const   float * src2,
+        device         float * dst,
+        constant     int64_t & ne00,
+        constant     int64_t & ne01,
+        constant     int64_t & ne02,
+        constant     int64_t & ne03,
+        constant    uint64_t & nb00,
+        constant    uint64_t & nb01,
+        constant    uint64_t & nb02,
+        constant    uint64_t & nb03,
+        constant     int64_t & ne0,
+        constant     int64_t & ne1,
+        constant     int64_t & ne2,
+        constant     int64_t & ne3,
+        constant    uint64_t & nb0,
+        constant    uint64_t & nb1,
+        constant    uint64_t & nb2,
+        constant    uint64_t & nb3,
+        constant         int & n_past,
+        constant         int & n_dims,
+        constant         int & n_ctx_orig,
+        constant       float & freq_base,
+        constant       float & freq_scale,
+        constant       float & ext_factor,
+        constant       float & attn_factor,
+        constant       float & beta_fast,
+        constant       float & beta_slow,
+        uint  tiitg[[thread_index_in_threadgroup]],
+        uint3 tptg[[threads_per_threadgroup]],
+        uint3 tgpig[[threadgroup_position_in_grid]]) {
+    const int64_t i3 = tgpig[2];
+    const int64_t i2 = tgpig[1];
+    const int64_t i1 = tgpig[0];
+
+    float corr_dims[2];
+    rope_yarn_corr_dims(n_dims, n_ctx_orig, freq_base, beta_fast, beta_slow, corr_dims);
+
+    device const int32_t * pos = src1;
+
+    const float theta_base = (float) pos[i2];
+    const float inv_ndims = -1.f/n_dims;
+
+    float cos_theta;
+    float sin_theta;
+
+    for (int64_t i0 = 2*tiitg; i0 < ne0; i0 += 2*tptg.x) {
+        if (i0 < n_dims) {
+            const int64_t ic = i0/2;
+
+            const float theta = theta_base * pow(freq_base, inv_ndims*i0);
+
+            const float freq_factor = src2 != src0 ? src2[ic] : 1.0f;
+
+            rope_yarn(theta/freq_factor, freq_scale, corr_dims, i0, ext_factor, attn_factor, &cos_theta, &sin_theta);
+
+            device const T * const src = (device T *)((device char *) src0 + i3*nb03 + i2*nb02 + i1*nb01 + i0*nb00);
+            device       T * dst_data  = (device T *)((device char *)  dst + i3*nb3  + i2*nb2  + i1*nb1  + i0*nb0);
+
+            const float x0 = src[0];
+            const float x1 = src[1];
+
+            dst_data[0] = x0*cos_theta - x1*sin_theta;
+            dst_data[1] = x0*sin_theta + x1*cos_theta;
+        } else {
+            device const T * const src = (device T *)((device char *) src0 + i3*nb03 + i2*nb02 + i1*nb01 + i0*nb00);
+            device       T * dst_data  = (device T *)((device char *)  dst + i3*nb3  + i2*nb2  + i1*nb1  + i0*nb0);
+
+            dst_data[0] = src[0];
+            dst_data[1] = src[1];
+        }
+    }
+}
+
+template<typename T>
+kernel void kernel_rope_neox(
+        device const    void * src0,
+        device const int32_t * src1,
+        device const   float * src2,
+        device         float * dst,
+        constant     int64_t & ne00,
+        constant     int64_t & ne01,
+        constant     int64_t & ne02,
+        constant     int64_t & ne03,
+        constant    uint64_t & nb00,
+        constant    uint64_t & nb01,
+        constant    uint64_t & nb02,
+        constant    uint64_t & nb03,
+        constant     int64_t & ne0,
+        constant     int64_t & ne1,
+        constant     int64_t & ne2,
+        constant     int64_t & ne3,
+        constant    uint64_t & nb0,
+        constant    uint64_t & nb1,
+        constant    uint64_t & nb2,
+        constant    uint64_t & nb3,
+        constant         int & n_past,
+        constant         int & n_dims,
+        constant         int & n_ctx_orig,
+        constant       float & freq_base,
+        constant       float & freq_scale,
+        constant       float & ext_factor,
+        constant       float & attn_factor,
+        constant       float & beta_fast,
+        constant       float & beta_slow,
+        uint  tiitg[[thread_index_in_threadgroup]],
+        uint3 tptg[[threads_per_threadgroup]],
+        uint3 tgpig[[threadgroup_position_in_grid]]) {
+    const int64_t i3 = tgpig[2];
+    const int64_t i2 = tgpig[1];
+    const int64_t i1 = tgpig[0];
+
+    float corr_dims[2];
+    rope_yarn_corr_dims(n_dims, n_ctx_orig, freq_base, beta_fast, beta_slow, corr_dims);
+
+    device const int32_t * pos = src1;
+
+    const float theta_base = (float) pos[i2];
+    const float inv_ndims = -1.f/n_dims;
+
+    float cos_theta;
+    float sin_theta;
+
+    for (int64_t i0 = 2*tiitg; i0 < ne0; i0 += 2*tptg.x) {
+        if (i0 < n_dims) {
+            const int64_t ic = i0/2;
+
+            const float theta = theta_base * pow(freq_base, inv_ndims*i0);
+
+            const float freq_factor = src2 != src0 ? src2[ic] : 1.0f;
+
+            rope_yarn(theta/freq_factor, freq_scale, corr_dims, i0, ext_factor, attn_factor, &cos_theta, &sin_theta);
+
+            device const T * const src = (device T *)((device char *) src0 + i3*nb03 + i2*nb02 + i1*nb01 + ic*nb00);
+            device       T * dst_data  = (device T *)((device char *)  dst + i3*nb3  + i2*nb2  + i1*nb1  + ic*nb0);
+
+            const float x0 = src[0];
+            const float x1 = src[n_dims/2];
+
+            dst_data[0]        = x0*cos_theta - x1*sin_theta;
+            dst_data[n_dims/2] = x0*sin_theta + x1*cos_theta;
+        } else {
+            device const T * const src = (device T *)((device char *) src0 + i3*nb03 + i2*nb02 + i1*nb01 + i0*nb00);
+            device       T * dst_data  = (device T *)((device char *)  dst + i3*nb3  + i2*nb2  + i1*nb1  + i0*nb0);
+
+            dst_data[0] = src[0];
+            dst_data[1] = src[1];
+        }
+    }
+}
+
+typedef decltype(kernel_rope_norm<float>) kernel_rope_norm_t;
+typedef decltype(kernel_rope_neox<float>) kernel_rope_neox_t;
+
+template [[host_name("kernel_rope_norm_f32")]] kernel kernel_rope_norm_t kernel_rope_norm<float>;
+template [[host_name("kernel_rope_norm_f16")]] kernel kernel_rope_norm_t kernel_rope_norm<half>;
+
+template [[host_name("kernel_rope_neox_f32")]] kernel kernel_rope_neox_t kernel_rope_neox<float>;
+template [[host_name("kernel_rope_neox_f16")]] kernel kernel_rope_neox_t kernel_rope_neox<half>;
+
+typedef void (im2col_t)(
+        device const float * x,
+        device        char * dst,
+        constant   int32_t & ofs0,
+        constant   int32_t & ofs1,
+        constant   int32_t & IW,
+        constant   int32_t & IH,
+        constant   int32_t & CHW,
+        constant   int32_t & s0,
+        constant   int32_t & s1,
+        constant   int32_t & p0,
+        constant   int32_t & p1,
+        constant   int32_t & d0,
+        constant   int32_t & d1,
+        uint3 tgpig[[threadgroup_position_in_grid]],
+        uint3  tgpg[[threadgroups_per_grid]],
+        uint3 tpitg[[thread_position_in_threadgroup]],
+        uint3   ntg[[threads_per_threadgroup]]);
+
+template <typename T>
+kernel void kernel_im2col(
+        device const float * x,
+        device        char * dst,
+        constant   int32_t & ofs0,
+        constant   int32_t & ofs1,
+        constant   int32_t & IW,
+        constant   int32_t & IH,
+        constant   int32_t & CHW,
+        constant   int32_t & s0,
+        constant   int32_t & s1,
+        constant   int32_t & p0,
+        constant   int32_t & p1,
+        constant   int32_t & d0,
+        constant   int32_t & d1,
+        uint3 tgpig[[threadgroup_position_in_grid]],
+        uint3  tgpg[[threadgroups_per_grid]],
+        uint3 tpitg[[thread_position_in_threadgroup]],
+        uint3   ntg[[threads_per_threadgroup]]) {
+    const int32_t iiw = tgpig[2] * s0 + tpitg[2] * d0 - p0;
+    const int32_t iih = tgpig[1] * s1 + tpitg[1] * d1 - p1;
+
+    const int32_t offset_dst =
+        (tpitg[0] * tgpg[1] * tgpg[2] + tgpig[1] * tgpg[2] + tgpig[2]) * CHW +
+        (tgpig[0] * (ntg[1] * ntg[2]) + tpitg[1] * ntg[2] + tpitg[2]);
+
+    device T * pdst = (device T *) (dst);
+
+    if (iih < 0 || iih >= IH || iiw < 0 || iiw >= IW) {
+        pdst[offset_dst] = 0.0f;
+    } else {
+        const int32_t offset_src = tpitg[0] * ofs0 + tgpig[0] * ofs1;
+        pdst[offset_dst] = x[offset_src + iih * IW + iiw];
+    }
+}
+
+template [[host_name("kernel_im2col_f32")]] kernel im2col_t kernel_im2col<float>;
+template [[host_name("kernel_im2col_f16")]] kernel im2col_t kernel_im2col<half>;
+
+kernel void kernel_upscale_f32(
+    device  const char * src0,
+    device        char * dst,
+    constant   int64_t & ne00,
+    constant   int64_t & ne01,
+    constant   int64_t & ne02,
+    constant   int64_t & ne03,
+    constant  uint64_t & nb00,
+    constant  uint64_t & nb01,
+    constant  uint64_t & nb02,
+    constant  uint64_t & nb03,
+    constant   int64_t & ne0,
+    constant   int64_t & ne1,
+    constant   int64_t & ne2,
+    constant   int64_t & ne3,
+    constant  uint64_t & nb0,
+    constant  uint64_t & nb1,
+    constant  uint64_t & nb2,
+    constant  uint64_t & nb3,
+    constant     float & sf0,
+    constant     float & sf1,
+    constant     float & sf2,
+    constant     float & sf3,
+    uint3 tgpig[[threadgroup_position_in_grid]],
+    uint3 tpitg[[thread_position_in_threadgroup]],
+    uint3   ntg[[threads_per_threadgroup]]) {
+
+    const int64_t i3 = tgpig.z;
+    const int64_t i2 = tgpig.y;
+    const int64_t i1 = tgpig.x;
+
+    const int64_t i03 = i3/sf3;
+    const int64_t i02 = i2/sf2;
+    const int64_t i01 = i1/sf1;
+
+    for (int i0 = tpitg.x; i0 < ne0; i0 += ntg.x) {
+        const int64_t i00 = i0/sf0;
+
+        device const float * src0_ptr = (device const float *) (src0 + i03*nb03 + i02*nb02 + i01*nb01 + i00*nb00);
+        device       float * dst_ptr  = (device       float *) (dst  +  i3*nb3  +  i2*nb2  +  i1*nb1  +  i0*nb0);
+
+        dst_ptr[0] = src0_ptr[0];
+    }
+}
+
+kernel void kernel_pad_f32(
+    device  const char * src0,
+    device        char * dst,
+    constant   int64_t & ne00,
+    constant   int64_t & ne01,
+    constant   int64_t & ne02,
+    constant   int64_t & ne03,
+    constant  uint64_t & nb00,
+    constant  uint64_t & nb01,
+    constant  uint64_t & nb02,
+    constant  uint64_t & nb03,
+    constant   int64_t & ne0,
+    constant   int64_t & ne1,
+    constant   int64_t & ne2,
+    constant   int64_t & ne3,
+    constant  uint64_t & nb0,
+    constant  uint64_t & nb1,
+    constant  uint64_t & nb2,
+    constant  uint64_t & nb3,
+    uint3 tgpig[[threadgroup_position_in_grid]],
+    uint3 tpitg[[thread_position_in_threadgroup]],
+    uint3   ntg[[threads_per_threadgroup]]) {
+
+    const int64_t i3 = tgpig.z;
+    const int64_t i2 = tgpig.y;
+    const int64_t i1 = tgpig.x;
+
+    const int64_t i03 = i3;
+    const int64_t i02 = i2;
+    const int64_t i01 = i1;
+
+    device const float * src0_ptr = (device const float *) (src0 + i03*nb03 + i02*nb02 + i01*nb01);
+    device       float * dst_ptr  = (device       float *) (dst  +  i3*nb3  +  i2*nb2  +  i1*nb1);
+
+    if (i1 < ne01 && i2 < ne02 && i3 < ne03) {
+        for (int i0 = tpitg.x; i0 < ne0; i0 += ntg.x) {
+            if (i0 < ne00) {
+                dst_ptr[i0] = src0_ptr[i0];
+            } else {
+                dst_ptr[i0] = 0.0f;
+            }
+        }
+
+        return;
+    }
+
+    for (int i0 = tpitg.x; i0 < ne0; i0 += ntg.x) {
+        dst_ptr[i0] = 0.0f;
+    }
+}
+
+kernel void kernel_arange_f32(
+    device        char * dst,
+    constant   int64_t & ne0,
+    constant   float   & start,
+    constant   float   & step,
+    uint3 tgpig[[threadgroup_position_in_grid]],
+    uint3 tpitg[[thread_position_in_threadgroup]],
+    uint3   ntg[[threads_per_threadgroup]]) {
+
+    device float * dst_ptr = (device float *) dst;
+
+    for (int i0 = tpitg.x; i0 < ne0; i0 += ntg.x) {
+        dst_ptr[i0] = start + step * i0;
+    }
+}
+
+kernel void kernel_timestep_embedding_f32(
+    device  const char * src0,
+    device        char * dst,
+    constant  uint64_t & nb1,
+    constant  int      & dim,
+    constant  int      & max_period,
+    uint3 tgpig[[threadgroup_position_in_grid]],
+    uint3 tpitg[[thread_position_in_threadgroup]],
+    uint3   ntg[[threads_per_threadgroup]]) {
+
+    int i = tgpig.x;
+    device float * embed_data = (device float *)(dst +  i*nb1);
+
+    int half_ = dim / 2;
+    for (int j = tpitg.x; j < half_; j += ntg.x) {
+        float timestep = ((device float *)src0)[i];
+        float freq = (float)exp(-log((float)max_period) * j / half_);
+        float arg = timestep * freq;
+        embed_data[j        ] = cos(arg);
+        embed_data[j + half_] = sin(arg);
+    }
+
+    if (dim % 2 != 0 && tpitg.x == 0) {
+        embed_data[dim] = 0.f;
+    }
+}
+
+// bitonic sort implementation following the CUDA kernels as reference
+typedef void (argsort_t)(
+        device const float  * x,
+        device     int32_t  * dst,
+        constant   int64_t  & ncols,
+        constant   int64_t  & ncols_pad,
+        threadgroup int32_t * shared_values [[threadgroup(0)]],
+        uint3 tgpig[[threadgroup_position_in_grid]],
+        uint3 tpitg[[thread_position_in_threadgroup]]);
+
+template<ggml_sort_order order>
+kernel void kernel_argsort_f32_i32(
+        device const float   * x,
+        device       int32_t * dst,
+        constant     int64_t & ncols,
+        constant     int64_t & ncols_pad,
+        threadgroup int32_t  * shared_values [[threadgroup(0)]],
+        uint3 tgpig[[threadgroup_position_in_grid]],
+        uint3 tpitg[[thread_position_in_threadgroup]]) {
+    // bitonic sort
+    int col = tpitg[0];
+    int row = tgpig[1];
+
+    if (col >= ncols_pad) return;
+
+    device const float   * x_row   = x + row * ncols;
+    threadgroup int32_t  * dst_row = shared_values;
+
+    // initialize indices
+    dst_row[col] = col;
+
+    threadgroup_barrier(mem_flags::mem_threadgroup);
+
+    for (int k = 2; k <= ncols_pad; k *= 2) {
+        for (int j = k / 2; j > 0; j /= 2) {
+            int ixj = col ^ j;
+            if (ixj > col) {
+                if ((col & k) == 0) {
+                    if (dst_row[col] >= ncols ||
+                        (dst_row[ixj] < ncols && (order == GGML_SORT_ORDER_ASC ?
+                            x_row[dst_row[col]] > x_row[dst_row[ixj]] :
+                            x_row[dst_row[col]] < x_row[dst_row[ixj]]))
+                    ) {
+                        SWAP(dst_row[col], dst_row[ixj]);
+                    }
+                } else {
+                    if (dst_row[ixj] >= ncols ||
+                        (dst_row[col] < ncols && (order == GGML_SORT_ORDER_ASC ?
+                            x_row[dst_row[col]] < x_row[dst_row[ixj]] :
+                            x_row[dst_row[col]] > x_row[dst_row[ixj]]))
+                    ) {
+                        SWAP(dst_row[col], dst_row[ixj]);
+                    }
+                }
+            }
+            threadgroup_barrier(mem_flags::mem_threadgroup);
+        }
+    }
+
+    // copy the result to dst without the padding
+    if (col < ncols) {
+        dst[row * ncols + col] = dst_row[col];
+    }
+}
+
+template [[host_name("kernel_argsort_f32_i32_asc")]]  kernel argsort_t kernel_argsort_f32_i32<GGML_SORT_ORDER_ASC>;
+template [[host_name("kernel_argsort_f32_i32_desc")]] kernel argsort_t kernel_argsort_f32_i32<GGML_SORT_ORDER_DESC>;
+
+kernel void kernel_leaky_relu_f32(
+        device const float * src0,
+        device       float * dst,
+        constant     float & slope,
+        uint tpig[[thread_position_in_grid]]) {
+    dst[tpig] = src0[tpig] > 0.0f ? src0[tpig] : src0[tpig] * slope;
+}
+
+typedef void (flash_attn_ext_f16_t)(
+        device const  char * q,
+        device const  char * k,
+        device const  char * v,
+        device const  char * mask,
+        device       float * dst,
+        constant   int64_t & ne01,
+        constant   int64_t & ne02,
+        constant   int64_t & ne03,
+        constant  uint64_t & nb01,
+        constant  uint64_t & nb02,
+        constant  uint64_t & nb03,
+        constant   int64_t & ne11,
+        constant   int64_t & ne12,
+        constant   int64_t & ne13,
+        constant  uint64_t & nb11,
+        constant  uint64_t & nb12,
+        constant  uint64_t & nb13,
+        constant  uint64_t & nb21,
+        constant  uint64_t & nb22,
+        constant  uint64_t & nb23,
+        constant  uint64_t & nb31,
+        constant   int64_t & ne1,
+        constant   int64_t & ne2,
+        constant     float & scale,
+        constant     float & max_bias,
+        constant     float & m0,
+        constant     float & m1,
+        constant  uint32_t & n_head_log2,
+        constant     float & logit_softcap,
+        threadgroup   half * shared,
+        uint3  tgpig[[threadgroup_position_in_grid]],
+        uint3  tpitg[[thread_position_in_threadgroup]],
+        uint3    ntg[[threads_per_threadgroup]],
+        ushort tiisg[[thread_index_in_simdgroup]],
+        ushort sgitg[[simdgroup_index_in_threadgroup]]);
+
+// ref: https://arxiv.org/pdf/2307.08691.pdf
+template<int64_t D, int64_t Q = 8, int64_t C = 32> // head size, queries per threadgroup, cache items per threadgroup
+kernel void kernel_flash_attn_ext_f16(
+        device const  char * q,
+        device const  char * k,
+        device const  char * v,
+        device const  char * mask,
+        device       float * dst,
+        constant   int64_t & ne01,
+        constant   int64_t & ne02,
+        constant   int64_t & ne03,
+        constant  uint64_t & nb01,
+        constant  uint64_t & nb02,
+        constant  uint64_t & nb03,
+        constant   int64_t & ne11,
+        constant   int64_t & ne12,
+        constant   int64_t & ne13,
+        constant  uint64_t & nb11,
+        constant  uint64_t & nb12,
+        constant  uint64_t & nb13,
+        constant  uint64_t & nb21,
+        constant  uint64_t & nb22,
+        constant  uint64_t & nb23,
+        constant  uint64_t & nb31,
+        constant   int64_t & ne1,
+        constant   int64_t & ne2,
+        constant     float & scale,
+        constant     float & max_bias,
+        constant     float & m0,
+        constant     float & m1,
+        constant  uint32_t & n_head_log2,
+        constant     float & logit_softcap,
+        threadgroup   half * shared [[threadgroup(0)]],
+        uint3  tgpig[[threadgroup_position_in_grid]],
+        uint3  tpitg[[thread_position_in_threadgroup]],
+        uint3    ntg[[threads_per_threadgroup]],
+        ushort tiisg[[thread_index_in_simdgroup]],
+        ushort sgitg[[simdgroup_index_in_threadgroup]]) {
+    const short nsg = ntg.y; // number of simdgroups
+
+    const short iq3 = tgpig[2];
+    const short iq2 = tgpig[1];
+    const short iq1 = tgpig[0]*Q;
+
+    const short D4 = D/4;
+    const short D8 = D/8;
+  //const short Q8 = Q/8;
+    const short NW = N_SIMDWIDTH;
+    const short SH = (C + Q); // shared memory per simdgroup in (half)
+
+    const short T  = D + 2*nsg*SH; // shared memory size per query in (half)
+    const short TF = T/2;        // shared memory size per query in (float)
+    const short T4 = T/4;        // shared memory size per query in (half4)
+
+    threadgroup half  * sq  = (threadgroup half  *) (shared +              0*D); // holds the query data
+    threadgroup half4 * sq4 = (threadgroup half4 *) (shared +              0*D); // same as above but in half4
+    threadgroup float * ss  = (threadgroup float *) (shared + 2*sgitg*SH + 1*D); // scratch buffer for attention and diagonal matrix
+
+    // store the result for all queries in local memory in 8x8 matrices (the O matrix from the paper)
+    simdgroup_half8x8 lo[D8];
+
+    // load heads from Q to shared memory
+    for (short j = sgitg; j < Q; j += nsg) {
+        device const float4 * q4 = (device const float4 *) ((device const char *) q + ((iq1 + j)*nb01 + iq2*nb02 + iq3*nb03));
+
+        for (short i = tiisg; i < D4; i += NW) {
+            if (iq1 + j < ne01) {
+                sq4[j*T4 + i] = (half4) q4[i];
+            } else {
+                sq4[j*T4 + i] = 0.0h;
+            }
+        }
+    }
+
+    // zero out lo
+    for (short i = 0; i < D8; ++i) {
+        lo[i] = make_filled_simdgroup_matrix<half, 8>(0.0h);
+    }
+
+    // zero out shared memory SH
+    for (short j = 0; j < Q; ++j) {
+        for (short i = tiisg; i < SH; i += NW) {
+            ss[j*TF + i] = 0.0f;
+        }
+    }
+
+    threadgroup_barrier(mem_flags::mem_threadgroup);
+
+    {
+        float S[Q] = { [0 ... Q-1] = 0.0h };
+        float M[Q] = { [0 ... Q-1] = -FLT_MAX/2 };
+
+        // assume K and V are same shape
+        const short ne22 = ne12;
+        const short ne23 = ne13;
+
+        // broadcast
+        const short rk2 = ne02/ne12;
+        const short rk3 = ne03/ne13;
+
+        const short rv2 = ne02/ne22;
+        const short rv3 = ne03/ne23;
+
+        // k indices
+        const short ik2 = iq2/rk2;
+        const short ik3 = iq3/rk3;
+
+        // v indices
+        const short iv2 = iq2/rv2;
+        const short iv3 = iq3/rv3;
+
+        // load the queries from shared memory into local memory
+        simdgroup_half8x8 mq[D8];
+
+        for (short i = 0; i < D8; ++i) {
+            simdgroup_load(mq[i], sq + i*8, T);
+        }
+
+        // pointer to the mask
+        device const half * mp = (device const half *) (mask + iq1*nb31);
+
+        float slope = 1.0f;
+
+        // ALiBi
+        if (max_bias > 0.0f) {
+            const uint32_t h = iq2;
+
+            const float base = h < n_head_log2 ? m0 : m1;
+            const int   exph = h < n_head_log2 ? h + 1 : 2*(h - n_head_log2) + 1;
+
+            slope = pow(base, exph);
+        }
+
+        // loop over the KV cache
+        // each simdgroup handles blocks of Q rows and C columns
+        for (int ic0 = 0; ic0 < ne11; ic0 += C*nsg) {
+            const int ic = ic0 + C*sgitg;
+            if (ic >= ne11) {
+                break;
+            }
+
+            // Q*K^T
+            {
+                for (short cc = 0; cc < C/8; ++cc) {
+                    simdgroup_float8x8 mqk = make_filled_simdgroup_matrix<float, 8>(0.h);
+
+                    device const half * pk = (device const half *) ((device const char *) k + ((ic + 8*cc)*nb11 + ik2*nb12 + ik3*nb13));
+
+                    for (short i = 0; i < D8; ++i) {
+                        simdgroup_half8x8 mk;
+                        simdgroup_load(mk, pk + i*8, nb11/sizeof(half), 0, true); // transpose
+
+                        simdgroup_multiply_accumulate(mqk, mq[i], mk, mqk);
+                    }
+
+                    simdgroup_store(mqk, ss + 8*cc, TF, 0, false);
+                }
+            }
+
+            // used to detect blocks full of -INF
+            float smax = -INFINITY;
+
+            // online softmax
+            {
+                float ms[Q];
+
+                for (short j = 0; j < Q; ++j) {
+                    const float m = M[j];
+
+                    // scale and apply the logitcap / mask
+                    float s = ss[j*TF + tiisg]*scale;
+
+                    if (logit_softcap != 0.0f) {
+                        s = logit_softcap*precise::tanh(s);
+                    }
+
+                    if (mask != q) {
+                        // mqk = mqk + mask*slope
+                        s += slope*mp[ic + j*nb31/sizeof(half) + tiisg];
+                    }
+
+                    smax = simd_max(max(smax, s));
+                    M[j] = simd_max(max(M[j], s));
+
+                                ms[j] = exp(m - M[j]);
+                    const float vs    = exp(s - M[j]);
+
+                    S[j] = S[j]*ms[j] + simd_sum(vs);
+
+                    // the P matrix from the paper (Q rows, C columns)
+                    ss[j*TF + tiisg] = vs;
+                }
+
+                // create a QxQ diagonal matrix for rescaling the output
+                if (tiisg < Q) {
+                    ss[tiisg*TF + C + tiisg] = ms[tiisg];
+                }
+            }
+
+            // skip -INF blocks
+            if (smax == -INFINITY) {
+                continue;
+            }
+
+            // O = diag(ms)*O
+            {
+                simdgroup_float8x8 mm;
+                simdgroup_load(mm, ss + C, TF, 0, false);
+
+                for (short i = 0; i < D8; ++i) {
+                    simdgroup_multiply(lo[i], mm, lo[i]);
+                }
+            }
+
+            // O = O + (Q*K^T)*V
+            {
+                for (short cc = 0; cc < C/8; ++cc) {
+                    device const half * pv = (device const half *) ((device const char *) v + ((ic + 8*cc)*nb21 + iv2*nb22 + iv3*nb23));
+
+                    for (short i = 0; i < D8; ++i) {
+                        simdgroup_half8x8 mk;
+                        simdgroup_load(mk, pv + i*8, nb21/sizeof(half), 0, false);
+
+                        simdgroup_float8x8 mv;
+                        simdgroup_load(mv, ss + 8*cc, TF, 0, false);
+
+                        simdgroup_multiply_accumulate(lo[i], mv, mk, lo[i]);
+                    }
+                }
+            }
+        }
+
+        // these are needed for reducing the results from the simdgroups (reuse the ss buffer)
+        for (short j = 0; j < Q; ++j) {
+            if (tiisg == 0) {
+                ss[j*TF + 0] = S[j];
+                ss[j*TF + 1] = M[j];
+            }
+        }
+    }
+
+    // reduce the warps sequentially
+    for (short sg = 1; sg < nsg; ++sg) {
+        float S = { 0.0h };
+        float M = { -FLT_MAX/2 };
+
+        threadgroup_barrier(mem_flags::mem_threadgroup);
+
+        // each simdgroup stores its output to shared memory, reusing sq
+        if (sgitg == sg) {
+            for (short i = 0; i < D8; ++i) {
+                simdgroup_store(lo[i], sq + i*8, T, 0, false);
+            }
+        }
+
+        threadgroup_barrier(mem_flags::mem_threadgroup);
+
+        // the first simdgroup accumulates the results from the other simdgroups
+        if (sgitg == 0) {
+            for (short j = 0; j < Q; ++j) {
+                const float S0 = ss[j*TF +         0];
+                const float S1 = ss[j*TF + sg*SH + 0];
+
+                const float M0 = ss[j*TF +         1];
+                const float M1 = ss[j*TF + sg*SH + 1];
+
+                M = max(M0, M1);
+
+                const float ms0 = exp(M0 - M);
+                const float ms1 = exp(M1 - M);
+
+                S = S0*ms0 + S1*ms1;
+
+                if (tiisg == 0) {
+                    ss[j*TF + 0] = S;
+                    ss[j*TF + 1] = M;
+
+                    ss[j*TF + C + j        ] = ms0;
+                    ss[j*TF + C + j + sg*SH] = ms1;
+                }
+            }
+
+            // O_0 = diag(ms0)*O_0 + diag(ms1)*O_1
+            {
+                simdgroup_half8x8 t;
+                simdgroup_float8x8 ms0;
+                simdgroup_float8x8 ms1;
+
+                simdgroup_load(ms0, ss + C,         TF, 0, false);
+                simdgroup_load(ms1, ss + C + sg*SH, TF, 0, false);
+
+                for (short i = 0; i < D8; ++i) {
+                    simdgroup_load    (t, sq + i*8, T, 0, false);
+                    simdgroup_multiply(t, ms1, t);
+
+                    simdgroup_multiply_accumulate(lo[i], ms0, lo[i], t);
+                }
+            }
+        }
+    }
+
+    // store result to shared memory (reuse sq)
+    if (sgitg == 0) {
+        for (short i = 0; i < D8; ++i) {
+            simdgroup_store(lo[i], sq + i*8, T, 0, false);
+        }
+    }
+
+    device float4 * dst4 = (device float4 *) dst;
+
+    // final rescale with 1/S and store to global memory
+    if (sgitg == 0) {
+        for (short j = 0; j < Q && iq1 + j < ne01; ++j) {
+            const float S = ss[j*TF + 0];
+
+            for (short i = tiisg; i < D4; i += NW) {
+                dst4[(iq3*ne2*ne1 + iq2 + (iq1 + j)*ne1)*D4 + i] = (float4) sq4[j*T4 + i]/S;
+            }
+        }
+    }
+}
+
+template [[host_name("kernel_flash_attn_ext_f16_h64" )]] kernel flash_attn_ext_f16_t kernel_flash_attn_ext_f16<64>;
+template [[host_name("kernel_flash_attn_ext_f16_h80" )]] kernel flash_attn_ext_f16_t kernel_flash_attn_ext_f16<80>;
+template [[host_name("kernel_flash_attn_ext_f16_h96" )]] kernel flash_attn_ext_f16_t kernel_flash_attn_ext_f16<96>;
+template [[host_name("kernel_flash_attn_ext_f16_h112")]] kernel flash_attn_ext_f16_t kernel_flash_attn_ext_f16<112>;
+template [[host_name("kernel_flash_attn_ext_f16_h128")]] kernel flash_attn_ext_f16_t kernel_flash_attn_ext_f16<128>;
+//template [[host_name("kernel_flash_attn_ext_f16_h256")]] kernel flash_attn_ext_f16_t kernel_flash_attn_ext_f16<256>;
+
+template<int64_t D, int64_t Q = 1, int64_t C = 32> // head size, queries per threadgroup, cache items per threadgroup
+kernel void kernel_flash_attn_ext_vec_f16(
+        device const  char * q,
+        device const  char * k,
+        device const  char * v,
+        device const  char * mask,
+        device       float * dst,
+        constant   int64_t & ne01,
+        constant   int64_t & ne02,
+        constant   int64_t & ne03,
+        constant  uint64_t & nb01,
+        constant  uint64_t & nb02,
+        constant  uint64_t & nb03,
+        constant   int64_t & ne11,
+        constant   int64_t & ne12,
+        constant   int64_t & ne13,
+        constant  uint64_t & nb11,
+        constant  uint64_t & nb12,
+        constant  uint64_t & nb13,
+        constant  uint64_t & nb21,
+        constant  uint64_t & nb22,
+        constant  uint64_t & nb23,
+        constant  uint64_t & nb31,
+        constant   int64_t & ne1,
+        constant   int64_t & ne2,
+        constant     float & scale,
+        constant     float & max_bias,
+        constant     float & m0,
+        constant     float & m1,
+        constant  uint32_t & n_head_log2,
+        constant     float & logit_softcap,
+        threadgroup   half * shared [[threadgroup(0)]],
+        uint3  tgpig[[threadgroup_position_in_grid]],
+        uint3  tpitg[[thread_position_in_threadgroup]],
+        uint3    ntg[[threads_per_threadgroup]],
+        ushort tiisg[[thread_index_in_simdgroup]],
+        ushort sgitg[[simdgroup_index_in_threadgroup]]) {
+    const short nsg = ntg.y; // number of simdgroups
+
+    const short iq3 = tgpig[2];
+    const short iq2 = tgpig[1];
+    const short iq1 = tgpig[0];
+
+    const short D4 = D/4;
+    const short NW = N_SIMDWIDTH;
+    const short SH = (C + Q); // shared memory per simdgroup in (half)
+
+    const short T  = D + 2*nsg*SH; // shared memory size per query in (half)
+
+    float slope = 1.0f;
+
+    // ALiBi
+    if (max_bias > 0.0f) {
+        const uint32_t h = iq2;
+
+        const float base = h < n_head_log2 ? m0 : m1;
+        const int   exp  = h < n_head_log2 ? h + 1 : 2*(h - n_head_log2) + 1;
+
+        slope = pow(base, exp);
+    }
+
+  //threadgroup half   * sq  = (threadgroup half   *) (shared +              0*D); // holds the query data
+    threadgroup half4  * sq4 = (threadgroup half4  *) (shared +              0*D); // same as above but in half4
+    threadgroup float  * ss  = (threadgroup float  *) (shared + 2*sgitg*SH + 1*D); // scratch buffer for attention and diagonal matrix
+    threadgroup float4 * ss4 = (threadgroup float4 *) (shared + 2*sgitg*SH + 1*D); // same as above but in half4
+    threadgroup half4  * sr4 = (threadgroup half4  *) (shared +   sgitg*D  + 1*T); // scratch buffer for the results
+
+    // store the result for all queries in local memory in 8x8 matrices (the O matrix from the paper)
+    half4 lo[D4/NW];
+
+    // load heads from Q to shared memory
+    device const float4 * q4 = (device const float4 *) ((device const char *) q + (iq1*nb01 + iq2*nb02 + iq3*nb03));
+
+    for (short i = tiisg; i < D4; i += NW) {
+        if (iq1 < ne01) {
+            sq4[i] = (half4) q4[i];
+        } else {
+            sq4[i] = 0.0h;
+        }
+    }
+
+    // zero out lo
+    for (short i = tiisg; i < D4; i += NW) {
+        lo[i/NW] = 0.0h;
+    }
+
+    // zero out shared memory SH
+    for (short i = tiisg; i < SH/4; i += NW) {
+        ss4[i] = 0.0h;
+    }
+
+    threadgroup_barrier(mem_flags::mem_threadgroup);
+
+    {
+        float S = { 0.0h };
+        float M = { -FLT_MAX/2 };
+
+        // assume K and V are same shape
+        const short ne22 = ne12;
+        const short ne23 = ne13;
+
+        // broadcast
+        const short rk2 = ne02/ne12;
+        const short rk3 = ne03/ne13;
+
+        const short rv2 = ne02/ne22;
+        const short rv3 = ne03/ne23;
+
+        // k indices
+        const short ik2 = iq2 / rk2;
+        const short ik3 = iq3 / rk3;
+
+        // v indices
+        const short iv2 = iq2 / rv2;
+        const short iv3 = iq3 / rv3;
+
+        // load the queries from shared memory into local memory
+        float4 mq[D4];
+
+        for (short ii = 0; ii < D4; ii += NW) {
+            short i = ii + tiisg;
+            mq[i] = (float4) sq4[i];
+        }
+
+        // pointer to the mask
+        device const half4 * mp4 = (device const half4 *) (mask + iq1*nb31);
+
+        // loop over the KV cache
+        // each simdgroup handles blocks of Q rows and C columns
+        for (int ic0 = 0; ic0 < ne11; ic0 += C*nsg) {
+            const int ic = ic0 + C*sgitg;
+            if (ic >= ne11) {
+                break;
+            }
+
+            // Q*K^T
+            {
+#pragma unroll
+                for (short cc = 0; cc < C/4; ++cc) {
+                    float4 mqk = { 0.0h };
+
+                    device const half4 * pk4 = (device const half4 *) ((device const char *) k + ((ic + 4*cc)*nb11 + ik2*nb12 + ik3*nb13));
+
+#pragma unroll
+                    for (short ii = 0; ii < D4; ii += NW) {
+                        const short i = ii + tiisg;
+
+                        float4x4 mk;
+                        mk[0] = (float4) pk4[i + 0*(nb11/8)];
+                        mk[1] = (float4) pk4[i + 1*(nb11/8)];
+                        mk[2] = (float4) pk4[i + 2*(nb11/8)];
+                        mk[3] = (float4) pk4[i + 3*(nb11/8)];
+
+                        mqk += (float4) (mq[i] * mk);
+                    }
+
+                    // reduce the results from the threads in the simdgroup
+                    mqk += simd_shuffle_down(mqk, 16);
+                    mqk += simd_shuffle_down(mqk,  8);
+                    mqk += simd_shuffle_down(mqk,  4);
+                    mqk += simd_shuffle_down(mqk,  2);
+                    mqk += simd_shuffle_down(mqk,  1);
+
+                    // mqk = mqk*scale + mask*slope
+                    if (tiisg == 0) {
+                        mqk *= scale;
+
+                        if (logit_softcap != 0.0f) {
+                            mqk = logit_softcap*precise::tanh(mqk);
+                        }
+
+                        mqk += (mask != q) ? ((float4) mp4[ic/4 + cc])*slope : (float4) 0.0f;
+
+                        ss4[cc] = mqk;
+                    }
+                }
+            }
+
+            // online softmax
+            {
+                const short p = tiisg;
+
+                const float m = M;
+                const float s = ss[p];
+
+                M = simd_max(max(M, s));
+
+                const float ms = exp(m - M);
+                const float vs = exp(s - M);
+
+                S = S*ms + simd_sum(vs);
+
+                // the P matrix from the paper (Q rows, C columns)
+                ss[p] = vs;
+
+                // O = diag(ms)*O
+#pragma unroll
+                for (short ii = 0; ii < D4; ii += NW) {
+                    const short i = ii + tiisg;
+                    lo[i/NW] *= ms;
+                }
+            }
+
+            // O = O + (Q*K^T)*V
+            {
+#pragma unroll
+                for (short cc = 0; cc < C/4; ++cc) {
+                    device const half4 * pv4 = (device const half4 *) ((device const char *) v + ((ic + 4*cc)*nb21 + iv2*nb22 + iv3*nb23));
+
+#pragma unroll
+                    for (short ii = 0; ii < D4; ii += NW) {
+                        const short i = ii + tiisg;
+
+                        lo[i/NW] += pv4[i + 0*(nb21/8)] * ss[4*cc + 0];
+                        lo[i/NW] += pv4[i + 1*(nb21/8)] * ss[4*cc + 1];
+                        lo[i/NW] += pv4[i + 2*(nb21/8)] * ss[4*cc + 2];
+                        lo[i/NW] += pv4[i + 3*(nb21/8)] * ss[4*cc + 3];
+                    }
+                }
+            }
+
+        }
+
+        // these are needed for reducing the results from the simdgroups (reuse the ss buffer)
+        if (tiisg == 0) {
+            ss[0] = S;
+            ss[1] = M;
+        }
+    }
+
+    // store results to shared memory
+    for (short ii = 0; ii < D4; ii += NW) {
+        short i = ii + tiisg;
+        sr4[i] = lo[ii/NW];
+    }
+
+    threadgroup_barrier(mem_flags::mem_threadgroup);
+
+    // parallel reduce
+    for (short r = nsg/2; r > 0; r >>= 1) {
+        if (sgitg < r) {
+            const float S0 = ss[       0];
+            const float S1 = ss[r*SH + 0];
+
+            const float M0 = ss[       1];
+            const float M1 = ss[r*SH + 1];
+
+            const float M = max(M0, M1);
+
+            const float ms0 = exp(M0 - M);
+            const float ms1 = exp(M1 - M);
+
+            const float S = S0*ms0 + S1*ms1;
+
+            if (tiisg == 0) {
+                ss[0] = S;
+                ss[1] = M;
+            }
+
+            // O_0 = diag(ms0)*O_0 + diag(ms1)*O_1
+            for (short ii = 0; ii < D4; ii += NW) {
+                short i = ii + tiisg;
+                sr4[i] = sr4[i]*ms0 + sr4[i + r*D4]*ms1;
+            }
+        }
+
+        threadgroup_barrier(mem_flags::mem_threadgroup);
+    }
+
+    device float4 * dst4 = (device float4 *) dst;
+
+    // final rescale with 1/S and store to global memory
+    if (sgitg == 0) {
+        const float S = ss[0];
+
+        for (short ii = 0; ii < D4; ii += NW) {
+            short i = ii + tiisg;
+            dst4[(iq3*ne2*ne1 + iq2 + (iq1)*ne1)*D4 + i] = (float4) sr4[i]/S;
+        }
+    }
+}
+
+template [[host_name("kernel_flash_attn_ext_vec_f16_h128")]] kernel flash_attn_ext_f16_t kernel_flash_attn_ext_vec_f16<128>;
+//template [[host_name("kernel_flash_attn_ext_vec_f16_h256")]] kernel flash_attn_ext_f16_t kernel_flash_attn_ext_vec_f16<256>;
+
+template<typename T0, typename T1>
+kernel void kernel_cpy(
+        device  const void * src0,
+        device        void * dst,
+        constant   int64_t & ne00,
+        constant   int64_t & ne01,
+        constant   int64_t & ne02,
+        constant   int64_t & ne03,
+        constant  uint64_t & nb00,
+        constant  uint64_t & nb01,
+        constant  uint64_t & nb02,
+        constant  uint64_t & nb03,
+        constant   int64_t & ne0,
+        constant   int64_t & ne1,
+        constant   int64_t & ne2,
+        constant   int64_t & ne3,
+        constant  uint64_t & nb0,
+        constant  uint64_t & nb1,
+        constant  uint64_t & nb2,
+        constant  uint64_t & nb3,
+        uint3 tgpig[[threadgroup_position_in_grid]],
+        uint3 tpitg[[thread_position_in_threadgroup]],
+        uint3   ntg[[threads_per_threadgroup]]) {
+    const int64_t i03 = tgpig[2];
+    const int64_t i02 = tgpig[1];
+    const int64_t i01 = tgpig[0];
+
+    const int64_t n = i03*ne02*ne01*ne00 + i02*ne01*ne00 + i01*ne00;
+
+    const int64_t i3 = n / (ne2*ne1*ne0);
+    const int64_t i2 = (n - i3*ne2*ne1*ne0) / (ne1*ne0);
+    const int64_t i1 = (n - i3*ne2*ne1*ne0 - i2*ne1*ne0) / ne0;
+    const int64_t i0 = (n - i3*ne2*ne1*ne0 - i2*ne1*ne0 - i1*ne0);
+
+    device T1 * dst_data = (device T1 *) ((device char *) dst + i3*nb3 + i2*nb2 + i1*nb1 + i0*nb0);
+
+    for (int64_t i00 = tpitg.x; i00 < ne00; i00 += ntg.x) {
+        device const T0 * src = (device T0 *)((device char *) src0 + i03*nb03 + i02*nb02 + i01*nb01 + i00*nb00);
+        dst_data[i00] = (T1) src[0];
+    }
+}
+
+typedef decltype(kernel_cpy<float, float>) kernel_cpy_t;
+
+template [[host_name("kernel_cpy_f32_f32")]]  kernel kernel_cpy_t kernel_cpy<float,  float>;
+template [[host_name("kernel_cpy_f32_f16")]]  kernel kernel_cpy_t kernel_cpy<float,  half>;
+template [[host_name("kernel_cpy_f16_f16")]]  kernel kernel_cpy_t kernel_cpy<half,   half>;
+template [[host_name("kernel_cpy_f16_f32")]]  kernel kernel_cpy_t kernel_cpy<half,   float>;
+
+kernel void kernel_cpy_f32_q8_0(
+        device const float * src0,
+        device        void * dst,
+        constant   int64_t & ne00,
+        constant   int64_t & ne01,
+        constant   int64_t & ne02,
+        constant   int64_t & ne03,
+        constant  uint64_t & nb00,
+        constant  uint64_t & nb01,
+        constant  uint64_t & nb02,
+        constant  uint64_t & nb03,
+        constant   int64_t & ne0,
+        constant   int64_t & ne1,
+        constant   int64_t & ne2,
+        constant   int64_t & ne3,
+        constant  uint64_t & nb0,
+        constant  uint64_t & nb1,
+        constant  uint64_t & nb2,
+        constant  uint64_t & nb3,
+        uint3 tgpig[[threadgroup_position_in_grid]],
+        uint3 tpitg[[thread_position_in_threadgroup]],
+        uint3   ntg[[threads_per_threadgroup]]) {
+    const int64_t i03 = tgpig[2];
+    const int64_t i02 = tgpig[1];
+    const int64_t i01 = tgpig[0];
+
+    const int64_t n = i03*ne02*ne01*ne00 + i02*ne01*ne00 + i01*ne00;
+
+    const int64_t i3 = n / (ne2*ne1*ne0);
+    const int64_t i2 = (n - i3*ne2*ne1*ne0) / (ne1*ne0);
+    const int64_t i1 = (n - i3*ne2*ne1*ne0 - i2*ne1*ne0) / ne0;
+    const int64_t i0 = (n - i3*ne2*ne1*ne0 - i2*ne1*ne0 - i1*ne0)/QK8_0;
+
+    device block_q8_0 * dst_data = (device block_q8_0 *) ((device char *) dst + i3*nb3 + i2*nb2 + i1*nb1 + i0*nb0);
+
+    for (int64_t i00 = tpitg.x*QK8_0; i00 < ne00; i00 += ntg.x*QK8_0) {
+        device const float * src = (device float *)((device char *) src0 + i03*nb03 + i02*nb02 + i01*nb01 + i00*nb00);
+
+        float amax = 0.0f; // absolute max
+
+        for (int j = 0; j < QK8_0; j++) {
+            const float v = src[j];
+            amax = MAX(amax, fabs(v));
+        }
+
+        const float d = amax / ((1 << 7) - 1);
+        const float id = d ? 1.0f/d : 0.0f;
+
+        dst_data[i00/QK8_0].d = d;
+
+        for (int j = 0; j < QK8_0; ++j) {
+            const float x0 = src[j]*id;
+
+            dst_data[i00/QK8_0].qs[j] = round(x0);
+        }
+    }
+}
+
+kernel void kernel_cpy_f32_q4_0(
+        device const float * src0,
+        device        void * dst,
+        constant   int64_t & ne00,
+        constant   int64_t & ne01,
+        constant   int64_t & ne02,
+        constant   int64_t & ne03,
+        constant  uint64_t & nb00,
+        constant  uint64_t & nb01,
+        constant  uint64_t & nb02,
+        constant  uint64_t & nb03,
+        constant   int64_t & ne0,
+        constant   int64_t & ne1,
+        constant   int64_t & ne2,
+        constant   int64_t & ne3,
+        constant  uint64_t & nb0,
+        constant  uint64_t & nb1,
+        constant  uint64_t & nb2,
+        constant  uint64_t & nb3,
+        uint3 tgpig[[threadgroup_position_in_grid]],
+        uint3 tpitg[[thread_position_in_threadgroup]],
+        uint3   ntg[[threads_per_threadgroup]]) {
+    const int64_t i03 = tgpig[2];
+    const int64_t i02 = tgpig[1];
+    const int64_t i01 = tgpig[0];
+
+    const int64_t n = i03*ne02*ne01*ne00 + i02*ne01*ne00 + i01*ne00;
+
+    const int64_t i3 = n / (ne2*ne1*ne0);
+    const int64_t i2 = (n - i3*ne2*ne1*ne0) / (ne1*ne0);
+    const int64_t i1 = (n - i3*ne2*ne1*ne0 - i2*ne1*ne0) / ne0;
+    const int64_t i0 = (n - i3*ne2*ne1*ne0 - i2*ne1*ne0 - i1*ne0)/QK4_0;
+
+    device block_q4_0 * dst_data = (device block_q4_0 *) ((device char *) dst + i3*nb3 + i2*nb2 + i1*nb1 + i0*nb0);
+
+    for (int64_t i00 = tpitg.x*QK4_0; i00 < ne00; i00 += ntg.x*QK4_0) {
+        device const float * src = (device float *)((device char *) src0 + i03*nb03 + i02*nb02 + i01*nb01 + i00*nb00);
+
+        float amax = 0.0f; // absolute max
+        float max  = 0.0f;
+
+        for (int j = 0; j < QK4_0; j++) {
+            const float v = src[j];
+            if (amax < fabs(v)) {
+                amax = fabs(v);
+                max  = v;
+            }
+        }
+
+        const float d = max / -8;
+        const float id = d ? 1.0f/d : 0.0f;
+
+        dst_data[i00/QK4_0].d = d;
+
+        for (int j = 0; j < QK4_0/2; ++j) {
+            const float x0 = src[0       + j]*id;
+            const float x1 = src[QK4_0/2 + j]*id;
+
+            const uint8_t xi0 = MIN(15, (int8_t)(x0 + 8.5f));
+            const uint8_t xi1 = MIN(15, (int8_t)(x1 + 8.5f));
+
+            dst_data[i00/QK4_0].qs[j]  = xi0;
+            dst_data[i00/QK4_0].qs[j] |= xi1 << 4;
+        }
+    }
+}
+
+kernel void kernel_cpy_f32_q4_1(
+        device const float * src0,
+        device        void * dst,
+        constant   int64_t & ne00,
+        constant   int64_t & ne01,
+        constant   int64_t & ne02,
+        constant   int64_t & ne03,
+        constant  uint64_t & nb00,
+        constant  uint64_t & nb01,
+        constant  uint64_t & nb02,
+        constant  uint64_t & nb03,
+        constant   int64_t & ne0,
+        constant   int64_t & ne1,
+        constant   int64_t & ne2,
+        constant   int64_t & ne3,
+        constant  uint64_t & nb0,
+        constant  uint64_t & nb1,
+        constant  uint64_t & nb2,
+        constant  uint64_t & nb3,
+        uint3 tgpig[[threadgroup_position_in_grid]],
+        uint3 tpitg[[thread_position_in_threadgroup]],
+        uint3   ntg[[threads_per_threadgroup]]) {
+    const int64_t i03 = tgpig[2];
+    const int64_t i02 = tgpig[1];
+    const int64_t i01 = tgpig[0];
+
+    const int64_t n = i03*ne02*ne01*ne00 + i02*ne01*ne00 + i01*ne00;
+
+    const int64_t i3 = n / (ne2*ne1*ne0);
+    const int64_t i2 = (n - i3*ne2*ne1*ne0) / (ne1*ne0);
+    const int64_t i1 = (n - i3*ne2*ne1*ne0 - i2*ne1*ne0) / ne0;
+    const int64_t i0 = (n - i3*ne2*ne1*ne0 - i2*ne1*ne0 - i1*ne0)/QK4_1;
+
+    device block_q4_1 * dst_data = (device block_q4_1 *) ((device char *) dst + i3*nb3 + i2*nb2 + i1*nb1 + i0*nb0);
+
+    for (int64_t i00 = tpitg.x*QK4_1; i00 < ne00; i00 += ntg.x*QK4_1) {
+        device const float * src = (device float *)((device char *) src0 + i03*nb03 + i02*nb02 + i01*nb01 + i00*nb00);
+
+        float min = FLT_MAX;
+        float max = -FLT_MAX;
+
+        for (int j = 0; j < QK4_1; j++) {
+            const float v = src[j];
+            if (min > v) min = v;
+            if (max < v) max = v;
+        }
+
+        const float d = (max - min) / ((1 << 4) - 1);
+        const float id = d ? 1.0f/d : 0.0f;
+
+        dst_data[i00/QK4_1].d = d;
+        dst_data[i00/QK4_1].m = min;
+
+        for (int j = 0; j < QK4_1/2; ++j) {
+            const float x0 = (src[0       + j] - min)*id;
+            const float x1 = (src[QK4_1/2 + j] - min)*id;
+
+            const uint8_t xi0 = MIN(15, (int8_t)(x0 + 0.5f));
+            const uint8_t xi1 = MIN(15, (int8_t)(x1 + 0.5f));
+
+            dst_data[i00/QK4_1].qs[j]  = xi0;
+            dst_data[i00/QK4_1].qs[j] |= xi1 << 4;
+        }
+    }
+}
+
+kernel void kernel_cpy_f32_q5_0(
+        device const float * src0,
+        device        void * dst,
+        constant   int64_t & ne00,
+        constant   int64_t & ne01,
+        constant   int64_t & ne02,
+        constant   int64_t & ne03,
+        constant  uint64_t & nb00,
+        constant  uint64_t & nb01,
+        constant  uint64_t & nb02,
+        constant  uint64_t & nb03,
+        constant   int64_t & ne0,
+        constant   int64_t & ne1,
+        constant   int64_t & ne2,
+        constant   int64_t & ne3,
+        constant  uint64_t & nb0,
+        constant  uint64_t & nb1,
+        constant  uint64_t & nb2,
+        constant  uint64_t & nb3,
+        uint3 tgpig[[threadgroup_position_in_grid]],
+        uint3 tpitg[[thread_position_in_threadgroup]],
+        uint3   ntg[[threads_per_threadgroup]]) {
+    const int64_t i03 = tgpig[2];
+    const int64_t i02 = tgpig[1];
+    const int64_t i01 = tgpig[0];
+
+    const int64_t n = i03*ne02*ne01*ne00 + i02*ne01*ne00 + i01*ne00;
+
+    const int64_t i3 = n / (ne2*ne1*ne0);
+    const int64_t i2 = (n - i3*ne2*ne1*ne0) / (ne1*ne0);
+    const int64_t i1 = (n - i3*ne2*ne1*ne0 - i2*ne1*ne0) / ne0;
+    const int64_t i0 = (n - i3*ne2*ne1*ne0 - i2*ne1*ne0 - i1*ne0)/QK5_0;
+
+    device block_q5_0 * dst_data = (device block_q5_0 *) ((device char *) dst + i3*nb3 + i2*nb2 + i1*nb1 + i0*nb0);
+
+    for (int64_t i00 = tpitg.x*QK5_0; i00 < ne00; i00 += ntg.x*QK5_0) {
+        device const float * src = (device float *)((device char *) src0 + i03*nb03 + i02*nb02 + i01*nb01 + i00*nb00);
+
+        float amax = 0.0f; // absolute max
+        float max  = 0.0f;
+
+        for (int j = 0; j < QK5_0; j++) {
+            const float v = src[j];
+            if (amax < fabs(v)) {
+                amax = fabs(v);
+                max  = v;
+            }
+        }
+
+        const float d = max / -16;
+        const float id = d ? 1.0f/d : 0.0f;
+
+        dst_data[i00/QK5_0].d = d;
+
+        uint32_t qh = 0;
+        for (int j = 0; j < QK5_0/2; ++j) {
+            const float x0 = src[0       + j]*id;
+            const float x1 = src[QK5_0/2 + j]*id;
+
+            const uint8_t xi0 = MIN(31, (int8_t)(x0 + 16.5f));
+            const uint8_t xi1 = MIN(31, (int8_t)(x1 + 16.5f));
+
+            dst_data[i00/QK5_0].qs[j] = (xi0 & 0xf) | ((xi1 & 0xf) << 4);
+            qh |= ((xi0 & 0x10u) >> 4) << (j + 0);
+            qh |= ((xi1 & 0x10u) >> 4) << (j + QK5_0/2);
+        }
+        thread const uint8_t * qh8 = (thread const uint8_t *)&qh;
+        for (int j = 0; j < 4; ++j) {
+            dst_data[i00/QK5_0].qh[j] = qh8[j];
+        }
+    }
+}
+
+kernel void kernel_cpy_f32_q5_1(
+        device const float * src0,
+        device        void * dst,
+        constant   int64_t & ne00,
+        constant   int64_t & ne01,
+        constant   int64_t & ne02,
+        constant   int64_t & ne03,
+        constant  uint64_t & nb00,
+        constant  uint64_t & nb01,
+        constant  uint64_t & nb02,
+        constant  uint64_t & nb03,
+        constant   int64_t & ne0,
+        constant   int64_t & ne1,
+        constant   int64_t & ne2,
+        constant   int64_t & ne3,
+        constant  uint64_t & nb0,
+        constant  uint64_t & nb1,
+        constant  uint64_t & nb2,
+        constant  uint64_t & nb3,
+        uint3 tgpig[[threadgroup_position_in_grid]],
+        uint3 tpitg[[thread_position_in_threadgroup]],
+        uint3   ntg[[threads_per_threadgroup]]) {
+    const int64_t i03 = tgpig[2];
+    const int64_t i02 = tgpig[1];
+    const int64_t i01 = tgpig[0];
+
+    const int64_t n = i03*ne02*ne01*ne00 + i02*ne01*ne00 + i01*ne00;
+
+    const int64_t i3 = n / (ne2*ne1*ne0);
+    const int64_t i2 = (n - i3*ne2*ne1*ne0) / (ne1*ne0);
+    const int64_t i1 = (n - i3*ne2*ne1*ne0 - i2*ne1*ne0) / ne0;
+    const int64_t i0 = (n - i3*ne2*ne1*ne0 - i2*ne1*ne0 - i1*ne0)/QK5_1;
+
+    device block_q5_1 * dst_data = (device block_q5_1 *) ((device char *) dst + i3*nb3 + i2*nb2 + i1*nb1 + i0*nb0);
+
+    for (int64_t i00 = tpitg.x*QK5_1; i00 < ne00; i00 += ntg.x*QK5_1) {
+        device const float * src = (device float *)((device char *) src0 + i03*nb03 + i02*nb02 + i01*nb01 + i00*nb00);
+
+        float max = src[0];
+        float min = src[0];
+
+        for (int j = 1; j < QK5_1; j++) {
+            const float v = src[j];
+            min = v < min ? v : min;
+            max = v > max ? v : max;
+        }
+
+        const float d = (max - min) / 31;
+        const float id = d ? 1.0f/d : 0.0f;
+
+        dst_data[i00/QK5_1].d = d;
+        dst_data[i00/QK5_1].m = min;
+
+        uint32_t qh = 0;
+        for (int j = 0; j < QK5_1/2; ++j) {
+            const float x0 = (src[0       + j] - min)*id;
+            const float x1 = (src[QK5_1/2 + j] - min)*id;
+
+            const uint8_t xi0 = (uint8_t)(x0 + 0.5f);
+            const uint8_t xi1 = (uint8_t)(x1 + 0.5f);
+
+            dst_data[i00/QK5_1].qs[j] = (xi0 & 0xf) | ((xi1 & 0xf) << 4);
+            qh |= ((xi0 & 0x10u) >> 4) << (j + 0);
+            qh |= ((xi1 & 0x10u) >> 4) << (j + QK5_1/2);
+        }
+        thread const uint8_t * qh8 = (thread const uint8_t *)&qh;
+        for (int j = 0; j < 4; ++j) {
+            dst_data[i00/QK5_1].qh[j] = qh8[j];
+        }
+    }
+}
+
+static inline int best_index_int8(int n, constant float * val, float x) {
+    if (x <= val[0]) return 0;
+    if (x >= val[n-1]) return n-1;
+    int ml = 0, mu = n-1;
+    while (mu-ml > 1) {
+        int mav = (ml+mu)/2;
+        if (x < val[mav]) mu = mav; else ml = mav;
+    }
+    return x - val[mu-1] < val[mu] - x ? mu-1 : mu;
+}
+
+constexpr constant static float kvalues_iq4nl_f[16] = {
+    -127.f, -104.f, -83.f, -65.f, -49.f, -35.f, -22.f, -10.f, 1.f, 13.f, 25.f, 38.f, 53.f, 69.f, 89.f, 113.f
+};
+
+kernel void kernel_cpy_f32_iq4_nl(
+        device const float * src0,
+        device        void * dst,
+        constant   int64_t & ne00,
+        constant   int64_t & ne01,
+        constant   int64_t & ne02,
+        constant   int64_t & ne03,
+        constant  uint64_t & nb00,
+        constant  uint64_t & nb01,
+        constant  uint64_t & nb02,
+        constant  uint64_t & nb03,
+        constant   int64_t & ne0,
+        constant   int64_t & ne1,
+        constant   int64_t & ne2,
+        constant   int64_t & ne3,
+        constant  uint64_t & nb0,
+        constant  uint64_t & nb1,
+        constant  uint64_t & nb2,
+        constant  uint64_t & nb3,
+        uint3 tgpig[[threadgroup_position_in_grid]],
+        uint3 tpitg[[thread_position_in_threadgroup]],
+        uint3   ntg[[threads_per_threadgroup]]) {
+    const int64_t i03 = tgpig[2];
+    const int64_t i02 = tgpig[1];
+    const int64_t i01 = tgpig[0];
+
+    const int64_t n = i03*ne02*ne01*ne00 + i02*ne01*ne00 + i01*ne00;
+
+    const int64_t i3 = n / (ne2*ne1*ne0);
+    const int64_t i2 = (n - i3*ne2*ne1*ne0) / (ne1*ne0);
+    const int64_t i1 = (n - i3*ne2*ne1*ne0 - i2*ne1*ne0) / ne0;
+    const int64_t i0 = (n - i3*ne2*ne1*ne0 - i2*ne1*ne0 - i1*ne0)/QK4_NL;
+
+    device block_iq4_nl * dst_data = (device block_iq4_nl *) ((device char *) dst + i3*nb3 + i2*nb2 + i1*nb1 + i0*nb0);
+
+    for (int64_t i00 = tpitg.x*QK4_NL; i00 < ne00; i00 += ntg.x*QK4_NL) {
+        device const float * src = (device float *)((device char *) src0 + i03*nb03 + i02*nb02 + i01*nb01 + i00*nb00);
+
+        float amax = 0.0f; // absolute max
+        float max  = 0.0f;
+
+        for (int j = 0; j < QK4_0; j++) {
+            const float v = src[j];
+            if (amax < fabs(v)) {
+                amax = fabs(v);
+                max  = v;
+            }
+        }
+
+        const float d = max / kvalues_iq4nl_f[0];
+        const float id = d ? 1.0f/d : 0.0f;
+
+        float sumqx = 0, sumq2 = 0;
+        for (int j = 0; j < QK4_NL/2; ++j) {
+            const float x0 = src[0        + j]*id;
+            const float x1 = src[QK4_NL/2 + j]*id;
+
+            const uint8_t xi0 = best_index_int8(16, kvalues_iq4nl_f, x0);
+            const uint8_t xi1 = best_index_int8(16, kvalues_iq4nl_f, x1);
+
+            dst_data[i00/QK4_NL].qs[j] = xi0 | (xi1 << 4);
+
+            const float v0 = kvalues_iq4nl_f[xi0];
+            const float v1 = kvalues_iq4nl_f[xi1];
+            const float w0 = src[0        + j]*src[0        + j];
+            const float w1 = src[QK4_NL/2 + j]*src[QK4_NL/2 + j];
+            sumqx += w0*v0*src[j] + w1*v1*src[QK4_NL/2 + j];
+            sumq2 += w0*v0*v0 + w1*v1*v1;
+
+        }
+
+        dst_data[i00/QK4_NL].d = sumq2 > 0 ? sumqx/sumq2 : d;
+
+    }
+}
+
+kernel void kernel_concat(
+    device  const char * src0,
+    device  const char * src1,
+    device        char * dst,
+    constant   int64_t & ne00,
+    constant   int64_t & ne01,
+    constant   int64_t & ne02,
+    constant   int64_t & ne03,
+    constant  uint64_t & nb00,
+    constant  uint64_t & nb01,
+    constant  uint64_t & nb02,
+    constant  uint64_t & nb03,
+    constant   int64_t & ne10,
+    constant   int64_t & ne11,
+    constant   int64_t & ne12,
+    constant   int64_t & ne13,
+    constant  uint64_t & nb10,
+    constant  uint64_t & nb11,
+    constant  uint64_t & nb12,
+    constant  uint64_t & nb13,
+    constant   int64_t & ne0,
+    constant   int64_t & ne1,
+    constant   int64_t & ne2,
+    constant   int64_t & ne3,
+    constant  uint64_t & nb0,
+    constant  uint64_t & nb1,
+    constant  uint64_t & nb2,
+    constant  uint64_t & nb3,
+    constant   int32_t & dim,
+    uint3 tgpig[[threadgroup_position_in_grid]],
+    uint3 tpitg[[thread_position_in_threadgroup]],
+    uint3   ntg[[threads_per_threadgroup]]) {
+
+    const int64_t i3 = tgpig.z;
+    const int64_t i2 = tgpig.y;
+    const int64_t i1 = tgpig.x;
+
+    int64_t o[4] = {0, 0, 0, 0};
+    o[dim] = dim == 0 ? ne00 : (dim == 1 ? ne01 : (dim == 2 ? ne02 : ne03));
+
+    device const float * x;
+
+    for (int i0 = tpitg.x; i0 < ne0; i0 += ntg.x) {
+        if (i0 < ne00 && i1 < ne01 && i2 < ne02 && i3 < ne03) {
+            x = (device const float *)(src0 + (i3       )*nb03 + (i2       )*nb02 + (i1       )*nb01 + (i0       )*nb00);
+        } else {
+            x = (device const float *)(src1 + (i3 - o[3])*nb13 + (i2 - o[2])*nb12 + (i1 - o[1])*nb11 + (i0 - o[0])*nb10);
+        }
+
+        device float * y = (device float *)(dst + i3*nb3 + i2*nb2 + i1*nb1 + i0*nb0);
+
+        *y = *x;
+    }
+}
+
+void kernel_mul_mv_q2_K_f32_impl(
+        device const  void * src0,
+        device const float * src1,
+        device       float * dst,
+                   int64_t   ne00,
+                   int64_t   ne01,
+                   int64_t   ne02,
+                   int64_t   ne10,
+                   int64_t   ne12,
+                   int64_t   ne0,
+                   int64_t   ne1,
+                   uint      r2,
+                   uint      r3,
+        threadgroup int8_t * shared_values,
+                   uint3     tgpig,
+                   uint      tiisg,
+                   uint      sgitg) {
+
+    const int nb = ne00/QK_K;
+    const int r0 = tgpig.x;
+    const int r1 = tgpig.y;
+    const int im = tgpig.z;
+
+    const int first_row = (r0 * N_SIMDGROUP + sgitg) * N_DST;
+    const int ib_row = first_row * nb;
+
+    const uint i12 = im%ne12;
+    const uint i13 = im/ne12;
+
+    const uint offset0 = (i12/r2)*(nb*ne01) + (i13/r3)*(nb*ne01*ne02);
+
+    device const block_q2_K * x = (device const block_q2_K *) src0 + ib_row + offset0;
+    device const float      * y = (device const float      *) src1 + r1*ne10 + im*ne00*ne1;
+
+    float yl[32];
+    float sumf[N_DST]={0.f}, all_sum;
+
+    const int step = sizeof(block_q2_K) * nb;
+
+    const int ix = tiisg/8;  // 0...3
+    const int it = tiisg%8;  // 0...7
+    const int iq = it/4;     // 0 or 1
+    const int ir = it%4;     // 0...3
+    const int is = (8*ir)/16;// 0 or 1
+
+    device const float * y4 = y + ix * QK_K + 128 * iq + 8 * ir;
+
+    for (int ib = ix; ib < nb; ib += 4) {
+
+        float4 sumy = {0.f, 0.f, 0.f, 0.f};
+        for (int i = 0; i < 8; ++i) {
+            yl[i+ 0] = y4[i+ 0]; sumy[0] += yl[i+ 0];
+            yl[i+ 8] = y4[i+32]; sumy[1] += yl[i+ 8];
+            yl[i+16] = y4[i+64]; sumy[2] += yl[i+16];
+            yl[i+24] = y4[i+96]; sumy[3] += yl[i+24];
+        }
+
+        device const uint8_t  * sc = (device const uint8_t  *)x[ib].scales + 8*iq + is;
+        device const uint16_t * qs = (device const uint16_t *)x[ib].qs + 16 * iq + 4 * ir;
+        device const half     * dh = &x[ib].d;
+
+        for (int row = 0; row < N_DST; row++) {
+
+            float4 acc1 = {0.f, 0.f, 0.f, 0.f};
+            float4 acc2 = {0.f, 0.f, 0.f, 0.f};
+            for (int i = 0; i < 8; i += 2) {
+                acc1[0] += yl[i+ 0] * (qs[i/2] & 0x0003);
+                acc2[0] += yl[i+ 1] * (qs[i/2] & 0x0300);
+                acc1[1] += yl[i+ 8] * (qs[i/2] & 0x000c);
+                acc2[1] += yl[i+ 9] * (qs[i/2] & 0x0c00);
+                acc1[2] += yl[i+16] * (qs[i/2] & 0x0030);
+                acc2[2] += yl[i+17] * (qs[i/2] & 0x3000);
+                acc1[3] += yl[i+24] * (qs[i/2] & 0x00c0);
+                acc2[3] += yl[i+25] * (qs[i/2] & 0xc000);
+            }
+            float dall = dh[0];
+            float dmin = dh[1] * 1.f/16.f;
+            sumf[row] += dall * ((acc1[0] + 1.f/256.f * acc2[0]) * (sc[0] & 0xF) * 1.f/ 1.f +
+                                 (acc1[1] + 1.f/256.f * acc2[1]) * (sc[2] & 0xF) * 1.f/ 4.f +
+                                 (acc1[2] + 1.f/256.f * acc2[2]) * (sc[4] & 0xF) * 1.f/16.f +
+                                 (acc1[3] + 1.f/256.f * acc2[3]) * (sc[6] & 0xF) * 1.f/64.f) -
+                         dmin * (sumy[0] * (sc[0] & 0xF0) + sumy[1] * (sc[2] & 0xF0) + sumy[2] * (sc[4] & 0xF0) + sumy[3] * (sc[6] & 0xF0));
+
+            qs += step/2;
+            sc += step;
+            dh += step/2;
+        }
+
+        y4 += 4 * QK_K;
+    }
+
+    for (int row = 0; row < N_DST; ++row) {
+        all_sum = simd_sum(sumf[row]);
+        if (tiisg == 0) {
+            dst[r1*ne0 + im*ne0*ne1 + first_row + row] = all_sum;
+        }
+    }
+}
+
+[[host_name("kernel_mul_mv_q2_K_f32")]]
+kernel void kernel_mul_mv_q2_K_f32(
+        device const  void * src0,
+        device const float * src1,
+        device       float * dst,
+        constant   int64_t & ne00,
+        constant   int64_t & ne01,
+        constant   int64_t & ne02,
+        constant  uint64_t & nb00,
+        constant  uint64_t & nb01,
+        constant  uint64_t & nb02,
+        constant   int64_t & ne10,
+        constant   int64_t & ne11,
+        constant   int64_t & ne12,
+        constant  uint64_t & nb10,
+        constant  uint64_t & nb11,
+        constant  uint64_t & nb12,
+        constant   int64_t & ne0,
+        constant   int64_t & ne1,
+        constant   uint    & r2,
+        constant   uint    & r3,
+        uint3 tgpig[[threadgroup_position_in_grid]],
+        uint  tiisg[[thread_index_in_simdgroup]],
+        uint  sgitg[[simdgroup_index_in_threadgroup]]) {
+
+    kernel_mul_mv_q2_K_f32_impl(src0, src1, dst, ne00, ne01, ne02, ne10, ne12, ne0, ne1, r2, r3, nullptr, tgpig, tiisg, sgitg);
+}
+
+void kernel_mul_mv_q3_K_f32_impl(
+        device const  void * src0,
+        device const float * src1,
+        device       float * dst,
+                   int64_t   ne00,
+                   int64_t   ne01,
+                   int64_t   ne02,
+                   int64_t   ne10,
+                   int64_t   ne12,
+                   int64_t   ne0,
+                   int64_t   ne1,
+                   uint      r2,
+                   uint      r3,
+        threadgroup int8_t * shared_values,
+                   uint3     tgpig,
+                   uint      tiisg,
+                   uint      sgitg) {
+
+    const int nb = ne00/QK_K;
+
+    const int64_t r0 = tgpig.x;
+    const int64_t r1 = tgpig.y;
+    const int64_t im = tgpig.z;
+
+    const int first_row = (r0 * N_SIMDGROUP + sgitg) * 2;
+
+    const uint i12 = im%ne12;
+    const uint i13 = im/ne12;
+
+    const uint offset0 = (i12/r2)*(nb*ne01) + (i13/r3)*(nb*ne01*ne02);
+
+    device const block_q3_K * x = (device const block_q3_K *) src0 + first_row*nb + offset0;
+    device const float     * yy = (device const float      *) src1 + r1*ne10 + im*ne00*ne1;
+
+    float yl[32];
+
+    //const uint16_t kmask1 = 0x3030;
+    //const uint16_t kmask2 = 0x0f0f;
+
+    const int tid = tiisg/4;
+    const int ix  = tiisg%4;
+    const int ip  = tid/4;          // 0 or 1
+    const int il  = 2*((tid%4)/2);  // 0 or 2
+    const int ir  = tid%2;
+    const int n   = 8;
+    const int l0  = n*ir;
+
+    // One would think that the Metal compiler would figure out that ip and il can only have
+    // 4 possible states, and optimize accordingly. Well, no. It needs help, and we do it
+    // with these two tales.
+    //
+    // Possible masks for the high bit
+    const ushort4 mm[4] = {{0x0001, 0x0100, 0x0002, 0x0200},  // ip = 0, il = 0
+                           {0x0004, 0x0400, 0x0008, 0x0800},  // ip = 0, il = 2
+                           {0x0010, 0x1000, 0x0020, 0x2000},  // ip = 1, il = 0
+                           {0x0040, 0x4000, 0x0080, 0x8000}}; // ip = 1, il = 2
+
+    // Possible masks for the low 2 bits
+    const int4 qm[2] = {{0x0003, 0x0300, 0x000c, 0x0c00}, {0x0030, 0x3000, 0x00c0, 0xc000}};
+
+    const ushort4 hm = mm[2*ip + il/2];
+
+    const int shift = 2*il;
+    const float    v1 = il == 0 ? 4.f : 64.f;
+    const float    v2 = 4.f * v1;
+
+    const uint16_t s_shift1 = 4*ip;
+    const uint16_t s_shift2 = s_shift1 + il;
+
+    const int q_offset = 32*ip + l0;
+    const int y_offset = 128*ip + 32*il + l0;
+
+    const int step = sizeof(block_q3_K) * nb / 2;
+
+    device const float * y1 = yy + ix*QK_K + y_offset;
+
+    uint32_t scales32, aux32;
+    thread uint16_t * scales16 = (thread uint16_t *)&scales32;
+    thread const int8_t * scales = (thread const int8_t *)&scales32;
+
+    float sumf1[2] = {0.f};
+    float sumf2[2] = {0.f};
+    for (int i = ix; i < nb; i += 4) {
+
+        for (int l = 0; l < 8; ++l) {
+            yl[l+ 0] = y1[l+ 0];
+            yl[l+ 8] = y1[l+16];
+            yl[l+16] = y1[l+32];
+            yl[l+24] = y1[l+48];
+        }
+
+        device const uint16_t * q = (device const uint16_t *)(x[i].qs + q_offset);
+        device const uint16_t * h = (device const uint16_t *)(x[i].hmask + l0);
+        device const uint16_t * a = (device const uint16_t *)(x[i].scales);
+        device const half * dh = &x[i].d;
+
+        for (int row = 0; row < 2; ++row) {
+
+            const float d_all = (float)dh[0];
+
+            scales16[0] = a[4];
+            scales16[1] = a[5];
+            aux32 = ((scales32 >> s_shift2) << 4) & 0x30303030;
+            scales16[0] = a[il+0];
+            scales16[1] = a[il+1];
+            scales32 = ((scales32 >> s_shift1) & 0x0f0f0f0f) | aux32;
+
+            float s1 = 0, s2 = 0, s3 = 0, s4 = 0, s5 = 0, s6 = 0;
+            for (int l = 0; l < n; l += 2) {
+                const int32_t qs = q[l/2];
+                s1 += yl[l+0] * (qs & qm[il/2][0]);
+                s2 += yl[l+1] * (qs & qm[il/2][1]);
+                s3 += ((h[l/2] & hm[0]) ? 0.f : yl[l+0]) + ((h[l/2] & hm[1]) ? 0.f : yl[l+1]);
+                s4 += yl[l+16] * (qs & qm[il/2][2]);
+                s5 += yl[l+17] * (qs & qm[il/2][3]);
+                s6 += ((h[l/2] & hm[2]) ? 0.f : yl[l+16]) + ((h[l/2] & hm[3]) ? 0.f : yl[l+17]);
+            }
+            float d1 = d_all * (s1 + 1.f/256.f * s2 - s3*v1);
+            float d2 = d_all * (s4 + 1.f/256.f * s5 - s6*v2);
+            sumf1[row] += d1 * (scales[0] - 32);
+            sumf2[row] += d2 * (scales[2] - 32);
+
+            s1 = s2 = s3 = s4 = s5 = s6 = 0;
+            for (int l = 0; l < n; l += 2) {
+                const int32_t qs = q[l/2+8];
+                s1 += yl[l+8] * (qs & qm[il/2][0]);
+                s2 += yl[l+9] * (qs & qm[il/2][1]);
+                s3 += ((h[l/2+8] & hm[0]) ? 0.f : yl[l+8]) + ((h[l/2+8] & hm[1]) ? 0.f : yl[l+9]);
+                s4 += yl[l+24] * (qs & qm[il/2][2]);
+                s5 += yl[l+25] * (qs & qm[il/2][3]);
+                s6 += ((h[l/2+8] & hm[2]) ? 0.f : yl[l+24]) + ((h[l/2+8] & hm[3]) ? 0.f : yl[l+25]);
+            }
+            d1 = d_all * (s1 + 1.f/256.f * s2 - s3*v1);
+            d2 = d_all * (s4 + 1.f/256.f * s5 - s6*v2);
+            sumf1[row] += d1 * (scales[1] - 32);
+            sumf2[row] += d2 * (scales[3] - 32);
+
+            q  += step;
+            h  += step;
+            a  += step;
+            dh += step;
+
+        }
+
+        y1 += 4 * QK_K;
+
+    }
+
+    for (int row = 0; row < 2; ++row) {
+        const float sumf = (sumf1[row] + 0.25f * sumf2[row]) / (1 << shift);
+        sumf1[row] = simd_sum(sumf);
+    }
+    if (tiisg == 0) {
+        for (int row = 0; row < 2; ++row) {
+            dst[r1*ne0 + im*ne0*ne1 + first_row + row] = sumf1[row];
+        }
+    }
+}
+
+[[host_name("kernel_mul_mv_q3_K_f32")]]
+kernel void kernel_mul_mv_q3_K_f32(
+        device const  void * src0,
+        device const float * src1,
+        device       float * dst,
+        constant   int64_t & ne00,
+        constant   int64_t & ne01,
+        constant   int64_t & ne02,
+        constant  uint64_t & nb00,
+        constant  uint64_t & nb01,
+        constant  uint64_t & nb02,
+        constant   int64_t & ne10,
+        constant   int64_t & ne11,
+        constant   int64_t & ne12,
+        constant  uint64_t & nb10,
+        constant  uint64_t & nb11,
+        constant  uint64_t & nb12,
+        constant   int64_t & ne0,
+        constant   int64_t & ne1,
+        constant   uint    & r2,
+        constant   uint    & r3,
+        uint3 tgpig[[threadgroup_position_in_grid]],
+        uint  tiisg[[thread_index_in_simdgroup]],
+        uint  sgitg[[simdgroup_index_in_threadgroup]]) {
+
+    kernel_mul_mv_q3_K_f32_impl(src0, src1, dst, ne00, ne01, ne02, ne10, ne12, ne0, ne1, r2, r3, nullptr, tgpig, tiisg, sgitg);
+}
+
+void kernel_mul_mv_q4_K_f32_impl(
+        device const  void * src0,
+        device const float * src1,
+        device       float * dst,
+                   int64_t   ne00,
+                   int64_t   ne01,
+                   int64_t   ne02,
+                   int64_t   ne10,
+                   int64_t   ne12,
+                   int64_t   ne0,
+                   int64_t   ne1,
+                   uint      r2,
+                   uint      r3,
+        threadgroup int8_t * shared_values,
+                   uint3     tgpig,
+                   uint      tiisg,
+                   uint      sgitg) {
+
+    const uint16_t kmask1 = 0x3f3f;
+    const uint16_t kmask2 = 0x0f0f;
+    const uint16_t kmask3 = 0xc0c0;
+
+    const int ix = tiisg/8;  // 0...3
+    const int it = tiisg%8;  // 0...7
+    const int iq = it/4;     // 0 or 1
+    const int ir = it%4;     // 0...3
+
+    const int nb = ne00/QK_K;
+    const int r0 = tgpig.x;
+    const int r1 = tgpig.y;
+    const int im = tgpig.z;
+    //const int first_row = (r0 * N_SIMDGROUP + sgitg) * N_DST;
+    const int first_row = r0 * N_DST;
+    const int ib_row = first_row * nb;
+
+    const uint i12 = im%ne12;
+    const uint i13 = im/ne12;
+
+    const uint offset0 = (i12/r2)*(nb*ne01) + (i13/r3)*(nb*ne01*ne02);
+
+    device const block_q4_K * x = (device const block_q4_K *) src0 + ib_row + offset0;
+    device const float      * y = (device const float      *) src1 + r1*ne10 + im*ne00*ne1;
+
+    float yl[16];
+    float yh[16];
+    float sumf[N_DST]={0.f}, all_sum;
+
+    const int step = sizeof(block_q4_K) * nb / 2;
+
+    device const float * y4 = y + ix * QK_K + 64 * iq + 8 * ir;
+
+    uint16_t sc16[4];
+    thread const uint8_t * sc8 = (thread const uint8_t *)sc16;
+
+    for (int ib = ix; ib < nb; ib += 4) {
+
+        float4 sumy = {0.f, 0.f, 0.f, 0.f};
+        for (int i = 0; i < 8; ++i) {
+            yl[i+0] = y4[i+  0]; sumy[0] += yl[i+0];
+            yl[i+8] = y4[i+ 32]; sumy[1] += yl[i+8];
+            yh[i+0] = y4[i+128]; sumy[2] += yh[i+0];
+            yh[i+8] = y4[i+160]; sumy[3] += yh[i+8];
+        }
+
+        device const uint16_t * sc = (device const uint16_t *)x[ib].scales + iq;
+        device const uint16_t * q1 = (device const uint16_t *)x[ib].qs + 16 * iq + 4 * ir;
+        device const half     * dh = &x[ib].d;
+
+        for (int row = 0; row < N_DST; row++) {
+
+            sc16[0] = sc[0] & kmask1;
+            sc16[1] = sc[2] & kmask1;
+            sc16[2] = ((sc[4] >> 0) & kmask2) | ((sc[0] & kmask3) >> 2);
+            sc16[3] = ((sc[4] >> 4) & kmask2) | ((sc[2] & kmask3) >> 2);
+
+            device const uint16_t * q2 = q1 + 32;
+
+            float4 acc1 = {0.f, 0.f, 0.f, 0.f};
+            float4 acc2 = {0.f, 0.f, 0.f, 0.f};
+            for (int i = 0; i < 8; i += 2) {
+                acc1[0] += yl[i+0] * (q1[i/2] & 0x000F);
+                acc1[1] += yl[i+1] * (q1[i/2] & 0x0F00);
+                acc1[2] += yl[i+8] * (q1[i/2] & 0x00F0);
+                acc1[3] += yl[i+9] * (q1[i/2] & 0xF000);
+                acc2[0] += yh[i+0] * (q2[i/2] & 0x000F);
+                acc2[1] += yh[i+1] * (q2[i/2] & 0x0F00);
+                acc2[2] += yh[i+8] * (q2[i/2] & 0x00F0);
+                acc2[3] += yh[i+9] * (q2[i/2] & 0xF000);
+            }
+
+            float dall = dh[0];
+            float dmin = dh[1];
+            sumf[row] += dall * ((acc1[0] + 1.f/256.f * acc1[1]) * sc8[0] +
+                                 (acc1[2] + 1.f/256.f * acc1[3]) * sc8[1] * 1.f/16.f +
+                                 (acc2[0] + 1.f/256.f * acc2[1]) * sc8[4] +
+                                 (acc2[2] + 1.f/256.f * acc2[3]) * sc8[5] * 1.f/16.f) -
+                         dmin * (sumy[0] * sc8[2] + sumy[1] * sc8[3] + sumy[2] * sc8[6] + sumy[3] * sc8[7]);
+
+            q1 += step;
+            sc += step;
+            dh += step;
+        }
+
+        y4 += 4 * QK_K;
+    }
+
+    for (int row = 0; row < N_DST; ++row) {
+        all_sum = simd_sum(sumf[row]);
+        if (tiisg == 0) {
+            dst[r1*ne0 + im*ne0*ne1 + first_row + row] = all_sum;
+        }
+    }
+}
+
+[[host_name("kernel_mul_mv_q4_K_f32")]]
+kernel void kernel_mul_mv_q4_K_f32(
+        device const  void * src0,
+        device const float * src1,
+        device       float * dst,
+        constant   int64_t & ne00,
+        constant   int64_t & ne01,
+        constant   int64_t & ne02,
+        constant  uint64_t & nb00,
+        constant  uint64_t & nb01,
+        constant  uint64_t & nb02,
+        constant   int64_t & ne10,
+        constant   int64_t & ne11,
+        constant   int64_t & ne12,
+        constant  uint64_t & nb10,
+        constant  uint64_t & nb11,
+        constant  uint64_t & nb12,
+        constant   int64_t & ne0,
+        constant   int64_t & ne1,
+        constant   uint    & r2,
+        constant   uint    & r3,
+        uint3 tgpig[[threadgroup_position_in_grid]],
+        uint tiisg[[thread_index_in_simdgroup]],
+        uint sgitg[[simdgroup_index_in_threadgroup]]) {
+
+    kernel_mul_mv_q4_K_f32_impl(src0, src1, dst, ne00, ne01, ne02, ne10, ne12, ne0, ne1, r2, r3, nullptr, tgpig, tiisg, sgitg);
+}
+
+void kernel_mul_mv_q5_K_f32_impl(
+        device const  void * src0,
+        device const float * src1,
+        device       float * dst,
+                   int64_t   ne00,
+                   int64_t   ne01,
+                   int64_t   ne02,
+                   int64_t   ne10,
+                   int64_t   ne12,
+                   int64_t   ne0,
+                   int64_t   ne1,
+                   uint      r2,
+                   uint      r3,
+        threadgroup int8_t * shared_values,
+                   uint3     tgpig,
+                   uint      tiisg,
+                   uint      sgitg) {
+
+    const int nb = ne00/QK_K;
+
+    const int64_t r0 = tgpig.x;
+    const int64_t r1 = tgpig.y;
+    const int im = tgpig.z;
+
+    const int first_row = (r0 * N_SIMDGROUP + sgitg) * 2;
+
+    const uint i12 = im%ne12;
+    const uint i13 = im/ne12;
+
+    const uint offset0 = (i12/r2)*(nb*ne01) + (i13/r3)*(nb*ne01*ne02);
+
+    device const block_q5_K * x = (device const block_q5_K *) src0 + first_row*nb + offset0;
+    device const float     * yy = (device const float      *) src1 + r1*ne10 + im*ne00*ne1;
+
+    float sumf[2]={0.f};
+
+    const int step = sizeof(block_q5_K) * nb;
+
+    float yl[16], yh[16];
+
+    const uint16_t kmask1 = 0x3f3f;
+    const uint16_t kmask2 = 0x0f0f;
+    const uint16_t kmask3 = 0xc0c0;
+
+    const int tid = tiisg/4;
+    const int ix  = tiisg%4;
+    const int iq  = tid/4;
+    const int ir  = tid%4;
+    const int n   = 8;
+
+    const int l0 = n*ir;
+    const int q_offset = 32*iq + l0;
+    const int y_offset = 64*iq + l0;
+
+    const uint8_t hm1 = 1u << (2*iq);
+    const uint8_t hm2 = hm1 << 1;
+    const uint8_t hm3 = hm1 << 4;
+    const uint8_t hm4 = hm2 << 4;
+
+    uint16_t sc16[4];
+    thread const uint8_t * sc8 = (thread const uint8_t *)sc16;
+
+    device const float * y1 = yy + ix*QK_K + y_offset;
+
+    for (int i = ix; i < nb; i += 4) {
+
+        device const uint8_t * q1 = x[i].qs + q_offset;
+        device const uint8_t * qh = x[i].qh + l0;
+        device const half * dh = &x[i].d;
+        device const uint16_t * a = (device const uint16_t *)x[i].scales + iq;
+
+        device const float * y2 = y1 + 128;
+        float4 sumy = {0.f, 0.f, 0.f, 0.f};
+        for (int l = 0; l < 8; ++l) {
+            yl[l+0] = y1[l+ 0]; sumy[0] += yl[l+0];
+            yl[l+8] = y1[l+32]; sumy[1] += yl[l+8];
+            yh[l+0] = y2[l+ 0]; sumy[2] += yh[l+0];
+            yh[l+8] = y2[l+32]; sumy[3] += yh[l+8];
+        }
+
+        for (int row = 0; row < 2; ++row) {
+
+            device const uint8_t * q2 = q1 + 64;
+
+            sc16[0] = a[0] & kmask1;
+            sc16[1] = a[2] & kmask1;
+            sc16[2] = ((a[4] >> 0) & kmask2) | ((a[0] & kmask3) >> 2);
+            sc16[3] = ((a[4] >> 4) & kmask2) | ((a[2] & kmask3) >> 2);
+
+            float4 acc1 = {0.f};
+            float4 acc2 = {0.f};
+            for (int l = 0; l < n; ++l) {
+                uint8_t h = qh[l];
+                acc1[0] += yl[l+0] * (q1[l] & 0x0F);
+                acc1[1] += yl[l+8] * (q1[l] & 0xF0);
+                acc1[2] += yh[l+0] * (q2[l] & 0x0F);
+                acc1[3] += yh[l+8] * (q2[l] & 0xF0);
+                acc2[0] += h & hm1 ? yl[l+0] : 0.f;
+                acc2[1] += h & hm2 ? yl[l+8] : 0.f;
+                acc2[2] += h & hm3 ? yh[l+0] : 0.f;
+                acc2[3] += h & hm4 ? yh[l+8] : 0.f;
+            }
+            const float dall = dh[0];
+            const float dmin = dh[1];
+            sumf[row] += dall * (sc8[0] * (acc1[0] +  16.f*acc2[0]) +
+                                 sc8[1] * (acc1[1]/16.f + 16.f*acc2[1]) +
+                                 sc8[4] * (acc1[2] +  16.f*acc2[2]) +
+                                 sc8[5] * (acc1[3]/16.f + 16.f*acc2[3])) -
+                         dmin * (sumy[0] * sc8[2] + sumy[1] * sc8[3] + sumy[2] * sc8[6] + sumy[3] * sc8[7]);
+
+            q1 += step;
+            qh += step;
+            dh += step/2;
+            a  += step/2;
+
+        }
+
+        y1 += 4 * QK_K;
+
+    }
+
+    for (int row = 0; row < 2; ++row) {
+        const float tot = simd_sum(sumf[row]);
+        if (tiisg == 0) {
+            dst[r1*ne0 + im*ne0*ne1 + first_row + row] = tot;
+        }
+    }
+}
+
+[[host_name("kernel_mul_mv_q5_K_f32")]]
+kernel void kernel_mul_mv_q5_K_f32(
+        device const  void * src0,
+        device const float * src1,
+        device       float * dst,
+        constant   int64_t & ne00,
+        constant   int64_t & ne01,
+        constant   int64_t & ne02,
+        constant  uint64_t & nb00,
+        constant  uint64_t & nb01,
+        constant  uint64_t & nb02,
+        constant   int64_t & ne10,
+        constant   int64_t & ne11,
+        constant   int64_t & ne12,
+        constant  uint64_t & nb10,
+        constant  uint64_t & nb11,
+        constant  uint64_t & nb12,
+        constant   int64_t & ne0,
+        constant   int64_t & ne1,
+        constant   uint    & r2,
+        constant   uint    & r3,
+        uint3 tgpig[[threadgroup_position_in_grid]],
+        uint  tiisg[[thread_index_in_simdgroup]],
+        uint  sgitg[[simdgroup_index_in_threadgroup]]) {
+
+    kernel_mul_mv_q5_K_f32_impl(src0, src1, dst, ne00, ne01, ne02, ne10, ne12, ne0, ne1, r2, r3, nullptr, tgpig, tiisg, sgitg);
+}
+
+void kernel_mul_mv_q6_K_f32_impl(
+        device const  void * src0,
+        device const float * src1,
+        device       float * dst,
+                   int64_t   ne00,
+                   int64_t   ne01,
+                   int64_t   ne02,
+                   int64_t   ne10,
+                   int64_t   ne12,
+                   int64_t   ne0,
+                   int64_t   ne1,
+                   uint      r2,
+                   uint      r3,
+        threadgroup int8_t * shared_values,
+                   uint3     tgpig,
+                   uint      tiisg,
+                   uint      sgitg) {
+
+    const uint8_t kmask1 = 0x03;
+    const uint8_t kmask2 = 0x0C;
+    const uint8_t kmask3 = 0x30;
+    const uint8_t kmask4 = 0xC0;
+
+    const int nb = ne00/QK_K;
+
+    const int64_t r0 = tgpig.x;
+    const int64_t r1 = tgpig.y;
+    const int     im = tgpig.z;
+
+    const int row = 2 * r0 + sgitg;
+
+    const uint i12 = im%ne12;
+    const uint i13 = im/ne12;
+
+    const uint offset0 = (i12/r2)*(nb*ne01) + (i13/r3)*(nb*ne01*ne02);
+
+    device const block_q6_K * x = (device const block_q6_K *) src0 + row * nb + offset0;
+    device const float     * yy = (device const float      *) src1 + r1*ne10 + im*ne00*ne1;
+
+    float sumf = 0;
+
+    const int tid  = tiisg/2;
+    const int ix   = tiisg%2;
+    const int ip   = tid/8;         // 0 or 1
+    const int il   = tid%8;
+    const int n    = 4;
+    const int l0   = n*il;
+    const int is   = 8*ip + l0/16;
+
+    const int y_offset = 128*ip + l0;
+    const int q_offset_l = 64*ip + l0;
+    const int q_offset_h = 32*ip + l0;
+
+    for (int i = ix; i < nb; i += 2) {
+
+        device const uint8_t * q1 = x[i].ql + q_offset_l;
+        device const uint8_t * q2 = q1 + 32;
+        device const uint8_t * qh = x[i].qh + q_offset_h;
+        device const int8_t  * sc = x[i].scales + is;
+
+        device const float * y = yy + i * QK_K + y_offset;
+
+        const float dall = x[i].d;
+
+        float4 sums = {0.f, 0.f, 0.f, 0.f};
+        for (int l = 0; l < n; ++l) {
+            sums[0] += y[l+ 0] * ((int8_t)((q1[l] & 0xF) | ((qh[l] & kmask1) << 4)) - 32);
+            sums[1] += y[l+32] * ((int8_t)((q2[l] & 0xF) | ((qh[l] & kmask2) << 2)) - 32);
+            sums[2] += y[l+64] * ((int8_t)((q1[l]  >> 4) | ((qh[l] & kmask3) << 0)) - 32);
+            sums[3] += y[l+96] * ((int8_t)((q2[l]  >> 4) | ((qh[l] & kmask4) >> 2)) - 32);
+        }
+
+        sumf += dall * (sums[0] * sc[0] + sums[1] * sc[2] + sums[2] * sc[4] + sums[3] * sc[6]);
+
+    }
+
+    const float tot = simd_sum(sumf);
+    if (tiisg == 0) {
+        dst[r1*ne0 + im*ne0*ne1 + row] = tot;
+    }
+}
+
+[[host_name("kernel_mul_mv_q6_K_f32")]]
+kernel void kernel_mul_mv_q6_K_f32(
+        device const  void * src0,
+        device const float * src1,
+        device       float * dst,
+        constant   int64_t & ne00,
+        constant   int64_t & ne01,
+        constant   int64_t & ne02,
+        constant  uint64_t & nb00,
+        constant  uint64_t & nb01,
+        constant  uint64_t & nb02,
+        constant   int64_t & ne10,
+        constant   int64_t & ne11,
+        constant   int64_t & ne12,
+        constant  uint64_t & nb10,
+        constant  uint64_t & nb11,
+        constant  uint64_t & nb12,
+        constant   int64_t & ne0,
+        constant   int64_t & ne1,
+        constant   uint    & r2,
+        constant   uint    & r3,
+        uint3 tgpig[[threadgroup_position_in_grid]],
+        uint  tiisg[[thread_index_in_simdgroup]],
+        uint  sgitg[[simdgroup_index_in_threadgroup]]) {
+
+    kernel_mul_mv_q6_K_f32_impl(src0, src1, dst, ne00, ne01, ne02, ne10, ne12, ne0, ne1, r2, r3, nullptr, tgpig, tiisg, sgitg);
+}
+
+// ======================= "True" 2-bit
+
+void kernel_mul_mv_iq2_xxs_f32_impl(
+        device const  void * src0,
+        device const float * src1,
+        device       float * dst,
+                   int64_t   ne00,
+                   int64_t   ne01,
+                   int64_t   ne02,
+                   int64_t   ne10,
+                   int64_t   ne12,
+                   int64_t   ne0,
+                   int64_t   ne1,
+                   uint      r2,
+                   uint      r3,
+        threadgroup int8_t * shared_values,
+                   uint3     tgpig,
+                   uint      tiisg,
+                   uint      sgitg) {
+
+    const int nb = ne00/QK_K;
+    const int r0 = tgpig.x;
+    const int r1 = tgpig.y;
+    const int im = tgpig.z;
+
+    const int first_row = (r0 * N_SIMDGROUP + sgitg) * N_DST;
+    const int ib_row = first_row * nb;
+
+    const uint i12 = im%ne12;
+    const uint i13 = im/ne12;
+
+    const uint offset0 = (i12/r2)*(nb*ne01) + (i13/r3)*(nb*ne01*ne02);
+
+    device const block_iq2_xxs * x = (device const block_iq2_xxs *) src0 + ib_row + offset0;
+    device const float         * y = (device const float         *) src1 + r1*ne10 + im*ne00*ne1;
+
+    float yl[32];
+    float sumf[N_DST]={0.f}, all_sum;
+
+    const int nb32 = nb * (QK_K / 32);
+
+    threadgroup uint64_t * values = (threadgroup uint64_t *)shared_values;
+    threadgroup uint8_t  * shared_signs = (threadgroup uint8_t *)(values + 256);
+    {
+        int nval = 4;
+        int pos  = (32*sgitg + tiisg)*nval;
+        for (int i = 0; i < nval; ++i) values[pos + i] = iq2xxs_grid[pos + i];
+        nval = 2;
+        pos  = (32*sgitg + tiisg)*nval;
+        for (int i = 0; i < nval; ++i) shared_signs[pos+i] = ksigns_iq2xs[pos+i];
+        threadgroup_barrier(mem_flags::mem_threadgroup);
+    }
+
+    const int ix = tiisg;
+
+    device const float * y4 = y + 32 * ix;
+
+    for (int ib32 = ix; ib32 < nb32; ib32 += 32) {
+
+        for (int i = 0; i < 32; ++i) {
+            yl[i] = y4[i];
+        }
+
+        const int ibl = ib32 / (QK_K / 32);
+        const int ib  = ib32 % (QK_K / 32);
+
+        device const block_iq2_xxs * xr = x + ibl;
+        device const uint16_t * q2 = xr->qs + 4 * ib;
+        device const half * dh = &xr->d;
+
+        for (int row = 0; row < N_DST; row++) {
+
+            const float db = dh[0];
+            device const uint8_t * aux8 = (device const uint8_t *)q2;
+            const uint32_t aux32 = q2[2] | (q2[3] << 16);
+            const float d = db * (0.5f + (aux32 >> 28));
+
+            float sum = 0;
+            for (int l = 0; l < 4; ++l) {
+                const threadgroup uint8_t * grid = (const threadgroup uint8_t *)(values + aux8[l]);
+                const uint8_t signs = shared_signs[(aux32 >> 7*l) & 127];
+                for (int j = 0; j < 8; ++j) {
+                    sum += yl[8*l + j] * grid[j] * (signs & kmask_iq2xs[j] ? -1.f : 1.f);
+                }
+            }
+            sumf[row] += d * sum;
+
+            dh += nb*sizeof(block_iq2_xxs)/2;
+            q2 += nb*sizeof(block_iq2_xxs)/2;
+        }
+
+        y4 += 32 * 32;
+    }
+
+    for (int row = 0; row < N_DST; ++row) {
+        all_sum = simd_sum(sumf[row]);
+        if (tiisg == 0) {
+            dst[r1*ne0 + im*ne0*ne1 + first_row + row] = all_sum * 0.25f;
+        }
+    }
+}
+
+[[host_name("kernel_mul_mv_iq2_xxs_f32")]]
+kernel void kernel_mul_mv_iq2_xxs_f32(
+        device const  void * src0,
+        device const float * src1,
+        device       float * dst,
+        constant   int64_t & ne00,
+        constant   int64_t & ne01,
+        constant   int64_t & ne02,
+        constant  uint64_t & nb00,
+        constant  uint64_t & nb01,
+        constant  uint64_t & nb02,
+        constant   int64_t & ne10,
+        constant   int64_t & ne11,
+        constant   int64_t & ne12,
+        constant  uint64_t & nb10,
+        constant  uint64_t & nb11,
+        constant  uint64_t & nb12,
+        constant   int64_t & ne0,
+        constant   int64_t & ne1,
+        constant   uint    & r2,
+        constant   uint    & r3,
+        threadgroup int8_t * shared_values [[threadgroup(0)]],
+        uint3 tgpig[[threadgroup_position_in_grid]],
+        uint  tiisg[[thread_index_in_simdgroup]],
+        uint  sgitg[[simdgroup_index_in_threadgroup]]) {
+
+    kernel_mul_mv_iq2_xxs_f32_impl(src0, src1, dst, ne00, ne01, ne02, ne10, ne12, ne0, ne1, r2, r3, shared_values, tgpig, tiisg, sgitg);
+}
+
+void kernel_mul_mv_iq2_xs_f32_impl(
+        device const  void * src0,
+        device const float * src1,
+        device       float * dst,
+                   int64_t   ne00,
+                   int64_t   ne01,
+                   int64_t   ne02,
+                   int64_t   ne10,
+                   int64_t   ne12,
+                   int64_t   ne0,
+                   int64_t   ne1,
+                   uint      r2,
+                   uint      r3,
+        threadgroup int8_t * shared_values,
+                   uint3     tgpig,
+                   uint      tiisg,
+                   uint      sgitg) {
+
+    const int nb = ne00/QK_K;
+    const int r0 = tgpig.x;
+    const int r1 = tgpig.y;
+    const int im = tgpig.z;
+
+    const int first_row = (r0 * N_SIMDGROUP + sgitg) * N_DST;
+    const int ib_row = first_row * nb;
+
+    const uint i12 = im%ne12;
+    const uint i13 = im/ne12;
+
+    const uint offset0 = (i12/r2)*(nb*ne01) + (i13/r3)*(nb*ne01*ne02);
+
+    device const block_iq2_xs * x = (device const block_iq2_xs *) src0 + ib_row + offset0;
+    device const float        * y = (device const float        *) src1 + r1*ne10 + im*ne00*ne1;
+
+    float yl[32];
+    float sumf[N_DST]={0.f}, all_sum;
+
+    const int nb32 = nb * (QK_K / 32);
+
+    threadgroup uint64_t * values = (threadgroup uint64_t *)shared_values;
+    threadgroup uint8_t  * shared_signs = (threadgroup uint8_t *)(values + 512);
+    {
+        int nval = 8;
+        int pos  = (32*sgitg + tiisg)*nval;
+        for (int i = 0; i < nval; ++i) values[pos + i] = iq2xs_grid[pos + i];
+        nval = 2;
+        pos  = (32*sgitg + tiisg)*nval;
+        for (int i = 0; i < nval; ++i) shared_signs[pos+i] = ksigns_iq2xs[pos+i];
+        threadgroup_barrier(mem_flags::mem_threadgroup);
+    }
+
+    const int ix = tiisg;
+
+    device const float * y4 = y + 32 * ix;
+
+    for (int ib32 = ix; ib32 < nb32; ib32 += 32) {
+
+        for (int i = 0; i < 32; ++i) {
+            yl[i] = y4[i];
+        }
+
+        const int ibl = ib32 / (QK_K / 32);
+        const int ib  = ib32 % (QK_K / 32);
+
+        device const block_iq2_xs * xr = x + ibl;
+        device const uint16_t * q2 = xr->qs + 4 * ib;
+        device const uint8_t  * sc = xr->scales + ib;
+        device const half * dh = &xr->d;
+
+        for (int row = 0; row < N_DST; row++) {
+
+            const float db = dh[0];
+            const uint8_t ls1 = sc[0] & 0xf;
+            const uint8_t ls2 = sc[0] >>  4;
+            const float d1 = db * (0.5f + ls1);
+            const float d2 = db * (0.5f + ls2);
+
+            float sum1 = 0, sum2 = 0;
+            for (int l = 0; l < 2; ++l) {
+                const threadgroup uint8_t * grid = (const threadgroup uint8_t *)(values + (q2[l] & 511));
+                const uint8_t signs = shared_signs[(q2[l] >> 9)];
+                for (int j = 0; j < 8; ++j) {
+                    sum1 += yl[8*l + j] * grid[j] * (signs & kmask_iq2xs[j] ? -1.f : 1.f);
+                }
+            }
+            for (int l = 2; l < 4; ++l) {
+                const threadgroup uint8_t * grid = (const threadgroup uint8_t *)(values + (q2[l] & 511));
+                const uint8_t signs = shared_signs[(q2[l] >> 9)];
+                for (int j = 0; j < 8; ++j) {
+                    sum2 += yl[8*l + j] * grid[j] * (signs & kmask_iq2xs[j] ? -1.f : 1.f);
+                }
+            }
+            sumf[row] += d1 * sum1 + d2 * sum2;
+
+            dh += nb*sizeof(block_iq2_xs)/2;
+            q2 += nb*sizeof(block_iq2_xs)/2;
+            sc += nb*sizeof(block_iq2_xs);
+        }
+
+        y4 += 32 * 32;
+    }
+
+    for (int row = 0; row < N_DST; ++row) {
+        all_sum = simd_sum(sumf[row]);
+        if (tiisg == 0) {
+            dst[r1*ne0 + im*ne0*ne1 + first_row + row] = all_sum * 0.25f;
+        }
+    }
+}
+
+[[host_name("kernel_mul_mv_iq2_xs_f32")]]
+kernel void kernel_mul_mv_iq2_xs_f32(
+        device const  void * src0,
+        device const float * src1,
+        device       float * dst,
+        constant   int64_t & ne00,
+        constant   int64_t & ne01,
+        constant   int64_t & ne02,
+        constant  uint64_t & nb00,
+        constant  uint64_t & nb01,
+        constant  uint64_t & nb02,
+        constant   int64_t & ne10,
+        constant   int64_t & ne11,
+        constant   int64_t & ne12,
+        constant  uint64_t & nb10,
+        constant  uint64_t & nb11,
+        constant  uint64_t & nb12,
+        constant   int64_t & ne0,
+        constant   int64_t & ne1,
+        constant   uint    & r2,
+        constant   uint    & r3,
+        threadgroup int8_t * shared_values [[threadgroup(0)]],
+        uint3 tgpig[[threadgroup_position_in_grid]],
+        uint  tiisg[[thread_index_in_simdgroup]],
+        uint  sgitg[[simdgroup_index_in_threadgroup]]) {
+
+    kernel_mul_mv_iq2_xs_f32_impl(src0, src1, dst, ne00, ne01, ne02, ne10, ne12, ne0, ne1, r2, r3, shared_values, tgpig, tiisg, sgitg);
+}
+
+void kernel_mul_mv_iq3_xxs_f32_impl(
+        device const  void * src0,
+        device const float * src1,
+        device       float * dst,
+                   int64_t   ne00,
+                   int64_t   ne01,
+                   int64_t   ne02,
+                   int64_t   ne10,
+                   int64_t   ne12,
+                   int64_t   ne0,
+                   int64_t   ne1,
+                   uint      r2,
+                   uint      r3,
+        threadgroup int8_t * shared_values,
+                   uint3     tgpig,
+                   uint      tiisg,
+                   uint      sgitg) {
+
+    const int nb = ne00/QK_K;
+    const int r0 = tgpig.x;
+    const int r1 = tgpig.y;
+    const int im = tgpig.z;
+
+    const int first_row = (r0 * N_SIMDGROUP + sgitg) * N_DST;
+    const int ib_row = first_row * nb;
+
+    const uint i12 = im%ne12;
+    const uint i13 = im/ne12;
+
+    const uint offset0 = (i12/r2)*(nb*ne01) + (i13/r3)*(nb*ne01*ne02);
+
+    device const block_iq3_xxs * x = (device const block_iq3_xxs *) src0 + ib_row + offset0;
+    device const float         * y = (device const float         *) src1 + r1*ne10 + im*ne00*ne1;
+
+    float yl[32];
+    float sumf[N_DST]={0.f}, all_sum;
+
+    const int nb32 = nb * (QK_K / 32);
+
+    threadgroup uint32_t * values = (threadgroup uint32_t *)shared_values;
+    threadgroup uint8_t  * shared_signs = (threadgroup uint8_t *)(values + 256);
+    {
+        int nval = 4;
+        int pos  = (32*sgitg + tiisg)*nval;
+        for (int i = 0; i < nval; ++i) values[pos + i] = iq3xxs_grid[pos + i];
+        nval = 2;
+        pos  = (32*sgitg + tiisg)*nval;
+        for (int i = 0; i < nval; ++i) shared_signs[pos+i] = ksigns_iq2xs[pos+i];
+        threadgroup_barrier(mem_flags::mem_threadgroup);
+    }
+
+    const int ix = tiisg;
+
+    device const float * y4 = y + 32 * ix;
+
+    for (int ib32 = ix; ib32 < nb32; ib32 += 32) {
+
+        for (int i = 0; i < 32; ++i) {
+            yl[i] = y4[i];
+        }
+
+        const int ibl = ib32 / (QK_K / 32);
+        const int ib  = ib32 % (QK_K / 32);
+
+        device const block_iq3_xxs * xr = x + ibl;
+        device const uint8_t  * q3 = xr->qs + 8 * ib;
+        device const uint16_t * gas = (device const uint16_t *)(xr->qs + QK_K/4) + 2 * ib;
+        device const half * dh = &xr->d;
+
+        for (int row = 0; row < N_DST; row++) {
+
+            const float db = dh[0];
+            const uint32_t aux32 = gas[0] | (gas[1] << 16);
+            const float d = db * (0.5f + (aux32 >> 28));
+
+            float2 sum = {0};
+            for (int l = 0; l < 4; ++l) {
+                const threadgroup uint8_t * grid1 = (const threadgroup uint8_t *)(values + q3[2*l+0]);
+                const threadgroup uint8_t * grid2 = (const threadgroup uint8_t *)(values + q3[2*l+1]);
+                const uint8_t signs = shared_signs[(aux32 >> 7*l) & 127];
+                for (int j = 0; j < 4; ++j) {
+                    sum[0] += yl[8*l + j + 0] * grid1[j] * (signs & kmask_iq2xs[j+0] ? -1.f : 1.f);
+                    sum[1] += yl[8*l + j + 4] * grid2[j] * (signs & kmask_iq2xs[j+4] ? -1.f : 1.f);
+                }
+            }
+            sumf[row] += d * (sum[0] + sum[1]);
+
+            dh  += nb*sizeof(block_iq3_xxs)/2;
+            q3  += nb*sizeof(block_iq3_xxs);
+            gas += nb*sizeof(block_iq3_xxs)/2;
+        }
+
+        y4 += 32 * 32;
+    }
+
+    for (int row = 0; row < N_DST; ++row) {
+        all_sum = simd_sum(sumf[row]);
+        if (tiisg == 0) {
+            dst[r1*ne0 + im*ne0*ne1 + first_row + row] = all_sum * 0.5f;
+        }
+    }
+}
+
+[[host_name("kernel_mul_mv_iq3_xxs_f32")]]
+kernel void kernel_mul_mv_iq3_xxs_f32(
+        device const  void * src0,
+        device const float * src1,
+        device       float * dst,
+        constant   int64_t & ne00,
+        constant   int64_t & ne01,
+        constant   int64_t & ne02,
+        constant  uint64_t & nb00,
+        constant  uint64_t & nb01,
+        constant  uint64_t & nb02,
+        constant   int64_t & ne10,
+        constant   int64_t & ne11,
+        constant   int64_t & ne12,
+        constant  uint64_t & nb10,
+        constant  uint64_t & nb11,
+        constant  uint64_t & nb12,
+        constant   int64_t & ne0,
+        constant   int64_t & ne1,
+        constant   uint    & r2,
+        constant   uint    & r3,
+        threadgroup int8_t * shared_values [[threadgroup(0)]],
+        uint3 tgpig[[threadgroup_position_in_grid]],
+        uint  tiisg[[thread_index_in_simdgroup]],
+        uint  sgitg[[simdgroup_index_in_threadgroup]]) {
+
+    kernel_mul_mv_iq3_xxs_f32_impl(src0, src1, dst, ne00, ne01, ne02, ne10, ne12, ne0, ne1, r2, r3, shared_values, tgpig, tiisg, sgitg);
+}
+
+void kernel_mul_mv_iq3_s_f32_impl(
+        device const  void * src0,
+        device const float * src1,
+        device       float * dst,
+                   int64_t   ne00,
+                   int64_t   ne01,
+                   int64_t   ne02,
+                   int64_t   ne10,
+                   int64_t   ne12,
+                   int64_t   ne0,
+                   int64_t   ne1,
+                   uint      r2,
+                   uint      r3,
+        threadgroup int8_t * shared_values,
+                   uint3     tgpig,
+                   uint      tiisg,
+                   uint      sgitg) {
+
+    const int nb = ne00/QK_K;
+    const int r0 = tgpig.x;
+    const int r1 = tgpig.y;
+    const int im = tgpig.z;
+
+    const int first_row = (r0 * N_SIMDGROUP + sgitg) * N_DST;
+    const int ib_row = first_row * nb;
+
+    const uint i12 = im%ne12;
+    const uint i13 = im/ne12;
+
+    const uint offset0 = (i12/r2)*(nb*ne01) + (i13/r3)*(nb*ne01*ne02);
+
+    device const block_iq3_s * x = (device const block_iq3_s *) src0 + ib_row + offset0;
+    device const float       * y = (device const float       *) src1 + r1*ne10 + im*ne00*ne1;
+
+    float yl[32];
+    float sumf[N_DST]={0.f}, all_sum;
+
+    const int nb32 = nb * (QK_K / 32);
+
+    threadgroup uint32_t * values = (threadgroup uint32_t *)shared_values;
+    {
+        int nval = 8;
+        int pos  = (32*sgitg + tiisg)*nval;
+        for (int i = 0; i < nval; ++i) values[pos + i] = iq3s_grid[pos + i];
+        threadgroup_barrier(mem_flags::mem_threadgroup);
+    }
+
+    const int ix = tiisg;
+
+    device const float * y4 = y + 32 * ix;
+
+    for (int ib32 = ix; ib32 < nb32; ib32 += 32) {
+
+        for (int i = 0; i < 32; ++i) {
+            yl[i] = y4[i];
+        }
+
+        const int ibl = ib32 / (QK_K / 32);
+        const int ib  = ib32 % (QK_K / 32);
+
+        device const block_iq3_s * xr = x + ibl;
+        device const uint8_t * qs = xr->qs + 8 * ib;
+        device const uint8_t * qh = xr->qh + ib;
+        device const uint8_t * sc = xr->scales + (ib/2);
+        device const uint8_t * signs = xr->signs + 4 * ib;
+        device const half * dh = &xr->d;
+
+        for (int row = 0; row < N_DST; row++) {
+
+            const float db = dh[0];
+            const float d = db * (1 + 2*((sc[0] >> 4*(ib%2)) & 0xf));
+
+            float2 sum = {0};
+            for (int l = 0; l < 4; ++l) {
+                const threadgroup uint32_t * table1 = qh[0] & kmask_iq2xs[2*l+0] ? values + 256 : values;
+                const threadgroup uint32_t * table2 = qh[0] & kmask_iq2xs[2*l+1] ? values + 256 : values;
+                const threadgroup uint8_t * grid1 = (const threadgroup uint8_t *)(table1 + qs[2*l+0]);
+                const threadgroup uint8_t * grid2 = (const threadgroup uint8_t *)(table2 + qs[2*l+1]);
+                for (int j = 0; j < 4; ++j) {
+                    sum[0] += yl[8*l + j + 0] * grid1[j] * select(1, -1, signs[l] & kmask_iq2xs[j+0]);
+                    sum[1] += yl[8*l + j + 4] * grid2[j] * select(1, -1, signs[l] & kmask_iq2xs[j+4]);
+                }
+            }
+            sumf[row] += d * (sum[0] + sum[1]);
+
+            dh  += nb*sizeof(block_iq3_s)/2;
+            qs  += nb*sizeof(block_iq3_s);
+            qh  += nb*sizeof(block_iq3_s);
+            sc  += nb*sizeof(block_iq3_s);
+            signs += nb*sizeof(block_iq3_s);
+        }
+
+        y4 += 32 * 32;
+    }
+
+    for (int row = 0; row < N_DST; ++row) {
+        all_sum = simd_sum(sumf[row]);
+        if (tiisg == 0) {
+            dst[r1*ne0 + im*ne0*ne1 + first_row + row] = all_sum;
+        }
+    }
+}
+
+[[host_name("kernel_mul_mv_iq3_s_f32")]]
+kernel void kernel_mul_mv_iq3_s_f32(
+        device const  void * src0,
+        device const float * src1,
+        device       float * dst,
+        constant   int64_t & ne00,
+        constant   int64_t & ne01,
+        constant   int64_t & ne02,
+        constant  uint64_t & nb00,
+        constant  uint64_t & nb01,
+        constant  uint64_t & nb02,
+        constant   int64_t & ne10,
+        constant   int64_t & ne11,
+        constant   int64_t & ne12,
+        constant  uint64_t & nb10,
+        constant  uint64_t & nb11,
+        constant  uint64_t & nb12,
+        constant   int64_t & ne0,
+        constant   int64_t & ne1,
+        constant   uint    & r2,
+        constant   uint    & r3,
+        threadgroup int8_t * shared_values [[threadgroup(0)]],
+        uint3 tgpig[[threadgroup_position_in_grid]],
+        uint  tiisg[[thread_index_in_simdgroup]],
+        uint  sgitg[[simdgroup_index_in_threadgroup]]) {
+
+    kernel_mul_mv_iq3_s_f32_impl(src0, src1, dst, ne00, ne01, ne02, ne10, ne12, ne0, ne1, r2, r3, shared_values, tgpig, tiisg, sgitg);
+}
+
+void kernel_mul_mv_iq2_s_f32_impl(
+        device const  void * src0,
+        device const float * src1,
+        device       float * dst,
+                   int64_t   ne00,
+                   int64_t   ne01,
+                   int64_t   ne02,
+                   int64_t   ne10,
+                   int64_t   ne12,
+                   int64_t   ne0,
+                   int64_t   ne1,
+                   uint      r2,
+                   uint      r3,
+        threadgroup int8_t * shared_values,
+                   uint3     tgpig,
+                   uint      tiisg,
+                   uint      sgitg) {
+
+    const int nb = ne00/QK_K;
+    const int r0 = tgpig.x;
+    const int r1 = tgpig.y;
+    const int im = tgpig.z;
+
+    const int first_row = (r0 * N_SIMDGROUP + sgitg) * N_DST;
+    const int ib_row = first_row * nb;
+
+    const uint i12 = im%ne12;
+    const uint i13 = im/ne12;
+
+    const uint offset0 = (i12/r2)*(nb*ne01) + (i13/r3)*(nb*ne01*ne02);
+
+    device const block_iq2_s * x = (device const block_iq2_s *) src0 + ib_row + offset0;
+    device const float       * y = (device const float       *) src1 + r1*ne10 + im*ne00*ne1;
+
+    float yl[32];
+    float sumf[N_DST]={0.f}, all_sum;
+
+    const int nb32 = nb * (QK_K / 32);
+
+    //threadgroup uint64_t * values = (threadgroup uint64_t *)shared_values;
+    //{
+    //    int nval = 32;
+    //    int pos  = (32*sgitg + tiisg)*nval;
+    //    for (int i = 0; i < nval; ++i) values[pos + i] = iq2s_grid[pos + i];
+    //    threadgroup_barrier(mem_flags::mem_threadgroup);
+    //}
+
+    const int ix = tiisg;
+
+    device const float * y4 = y + 32 * ix;
+
+    for (int ib32 = ix; ib32 < nb32; ib32 += 32) {
+
+        for (int i = 0; i < 32; ++i) {
+            yl[i] = y4[i];
+        }
+
+        const int ibl = ib32 / (QK_K / 32);
+        const int ib  = ib32 % (QK_K / 32);
+
+        device const block_iq2_s * xr = x + ibl;
+        device const uint8_t * qs = xr->qs + 4 * ib;
+        device const uint8_t * qh = xr->qh + ib;
+        device const uint8_t * sc = xr->scales + ib;
+        device const uint8_t * signs = qs + QK_K/8;
+        device const half * dh = &xr->d;
+
+        for (int row = 0; row < N_DST; row++) {
+
+            const float db = dh[0];
+            const float d1 = db * (0.5f + (sc[0] & 0xf));
+            const float d2 = db * (0.5f + (sc[0] >>  4));
+
+            float2 sum = {0};
+            for (int l = 0; l < 2; ++l) {
+                //const threadgroup uint8_t * grid1 = (const threadgroup uint8_t *)(values + (qs[l+0] | ((qh[0] << (8-2*l)) & 0x300)));
+                //const threadgroup uint8_t * grid2 = (const threadgroup uint8_t *)(values + (qs[l+2] | ((qh[0] << (4-2*l)) & 0x300)));
+                constant uint8_t * grid1 = (constant uint8_t *)(iq2s_grid + (qs[l+0] | ((qh[0] << (8-2*l)) & 0x300)));
+                constant uint8_t * grid2 = (constant uint8_t *)(iq2s_grid + (qs[l+2] | ((qh[0] << (4-2*l)) & 0x300)));
+                for (int j = 0; j < 8; ++j) {
+                    sum[0] += yl[8*l + j +  0] * grid1[j] * select(1, -1, signs[l+0] & kmask_iq2xs[j]);
+                    sum[1] += yl[8*l + j + 16] * grid2[j] * select(1, -1, signs[l+2] & kmask_iq2xs[j]);
+                }
+            }
+            sumf[row] += d1 * sum[0] + d2 * sum[1];
+
+            dh  += nb*sizeof(block_iq2_s)/2;
+            qs  += nb*sizeof(block_iq2_s);
+            qh  += nb*sizeof(block_iq2_s);
+            sc  += nb*sizeof(block_iq2_s);
+            signs += nb*sizeof(block_iq2_s);
+        }
+
+        y4 += 32 * 32;
+    }
+
+    for (int row = 0; row < N_DST; ++row) {
+        all_sum = simd_sum(sumf[row]);
+        if (tiisg == 0) {
+            dst[r1*ne0 + im*ne0*ne1 + first_row + row] = all_sum * 0.25f;
+        }
+    }
+}
+
+[[host_name("kernel_mul_mv_iq2_s_f32")]]
+kernel void kernel_mul_mv_iq2_s_f32(
+        device const  void * src0,
+        device const float * src1,
+        device       float * dst,
+        constant   int64_t & ne00,
+        constant   int64_t & ne01,
+        constant   int64_t & ne02,
+        constant  uint64_t & nb00,
+        constant  uint64_t & nb01,
+        constant  uint64_t & nb02,
+        constant   int64_t & ne10,
+        constant   int64_t & ne11,
+        constant   int64_t & ne12,
+        constant  uint64_t & nb10,
+        constant  uint64_t & nb11,
+        constant  uint64_t & nb12,
+        constant   int64_t & ne0,
+        constant   int64_t & ne1,
+        constant   uint    & r2,
+        constant   uint    & r3,
+        threadgroup int8_t * shared_values [[threadgroup(0)]],
+        uint3 tgpig[[threadgroup_position_in_grid]],
+        uint  tiisg[[thread_index_in_simdgroup]],
+        uint  sgitg[[simdgroup_index_in_threadgroup]]) {
+
+    kernel_mul_mv_iq2_s_f32_impl(src0, src1, dst, ne00, ne01, ne02, ne10, ne12, ne0, ne1, r2, r3, shared_values, tgpig, tiisg, sgitg);
+}
+
+void kernel_mul_mv_iq1_s_f32_impl(
+        device const  void * src0,
+        device const float * src1,
+        device       float * dst,
+                   int64_t   ne00,
+                   int64_t   ne01,
+                   int64_t   ne02,
+                   int64_t   ne10,
+                   int64_t   ne12,
+                   int64_t   ne0,
+                   int64_t   ne1,
+                   uint      r2,
+                   uint      r3,
+        threadgroup int8_t * shared_value,
+                   uint3     tgpig,
+                   uint      tiisg,
+                   uint      sgitg) {
+
+    const int nb = ne00/QK_K;
+    const int r0 = tgpig.x;
+    const int r1 = tgpig.y;
+    const int im = tgpig.z;
+
+    const int first_row = (r0 * N_SIMDGROUP + sgitg) * N_DST;
+    const int ib_row = first_row * nb;
+
+    const uint i12 = im%ne12;
+    const uint i13 = im/ne12;
+
+    const uint offset0 = (i12/r2)*(nb*ne01) + (i13/r3)*(nb*ne01*ne02);
+    device const block_iq1_s * x = (device const block_iq1_s *) src0 + ib_row + offset0;
+    device const float       * y = (device const float       *) src1 + r1*ne10 + im*ne00*ne1;
+
+    float yl[32];
+    float sumf[N_DST]={0.f}, all_sum;
+
+    const int nb32 = nb * (QK_K / 32);
+
+    const int ix = tiisg;
+
+    device const float * y4 = y + 32 * ix;
+
+    for (int ib32 = ix; ib32 < nb32; ib32 += 32) {
+
+        float sumy = 0;
+        for (int i = 0; i < 32; ++i) {
+            yl[i] = y4[i];
+            sumy += yl[i];
+        }
+
+        const int ibl = ib32 / (QK_K / 32);
+        const int ib  = ib32 % (QK_K / 32);
+
+        device const block_iq1_s * xr = x + ibl;
+        device const uint8_t  * qs = xr->qs + 4 * ib;
+        device const uint16_t * qh = xr->qh + ib;
+        device const half     * dh = &xr->d;
+
+        for (int row = 0; row < N_DST; row++) {
+
+            constant uint8_t * grid1 = (constant uint8_t *)(iq1s_grid_gpu + (qs[0] | ((qh[0] << 8) & 0x700)));
+            constant uint8_t * grid2 = (constant uint8_t *)(iq1s_grid_gpu + (qs[1] | ((qh[0] << 5) & 0x700)));
+            constant uint8_t * grid3 = (constant uint8_t *)(iq1s_grid_gpu + (qs[2] | ((qh[0] << 2) & 0x700)));
+            constant uint8_t * grid4 = (constant uint8_t *)(iq1s_grid_gpu + (qs[3] | ((qh[0] >> 1) & 0x700)));
+
+            float sum = 0;
+            for (int j = 0; j < 4; ++j) {
+                sum += yl[j+ 0] * (grid1[j] & 0xf) + yl[j+ 4] * (grid1[j] >> 4)
+                     + yl[j+ 8] * (grid2[j] & 0xf) + yl[j+12] * (grid2[j] >> 4)
+                     + yl[j+16] * (grid3[j] & 0xf) + yl[j+20] * (grid3[j] >> 4)
+                     + yl[j+24] * (grid4[j] & 0xf) + yl[j+28] * (grid4[j] >> 4);
+            }
+            sumf[row] += (float)dh[0] * (sum + sumy * (qh[0] & 0x8000 ? -1 - IQ1S_DELTA : -1 + IQ1S_DELTA)) * (2*((qh[0] >> 12) & 7) + 1);
+
+            dh += nb*sizeof(block_iq1_s)/2;
+            qs += nb*sizeof(block_iq1_s);
+            qh += nb*sizeof(block_iq1_s)/2;
+        }
+
+        y4 += 32 * 32;
+    }
+
+    for (int row = 0; row < N_DST; ++row) {
+        all_sum = simd_sum(sumf[row]);
+        if (tiisg == 0) {
+            dst[r1*ne0 + im*ne0*ne1 + first_row + row] = all_sum;
+        }
+    }
+}
+
+void kernel_mul_mv_iq1_m_f32_impl(
+        device const  void * src0,
+        device const float * src1,
+        device       float * dst,
+                   int64_t   ne00,
+                   int64_t   ne01,
+                   int64_t   ne02,
+                   int64_t   ne10,
+                   int64_t   ne12,
+                   int64_t   ne0,
+                   int64_t   ne1,
+                   uint      r2,
+                   uint      r3,
+        threadgroup int8_t * shared_value,
+                   uint3     tgpig,
+                   uint      tiisg,
+                   uint      sgitg) {
+
+    const int nb = ne00/QK_K;
+    const int r0 = tgpig.x;
+    const int r1 = tgpig.y;
+    const int im = tgpig.z;
+
+    const int first_row = (r0 * N_SIMDGROUP + sgitg) * N_DST;
+    const int ib_row = first_row * nb;
+
+    const uint i12 = im%ne12;
+    const uint i13 = im/ne12;
+
+    const uint offset0 = (i12/r2)*(nb*ne01) + (i13/r3)*(nb*ne01*ne02);
+    device const block_iq1_m * x = (device const block_iq1_m *) src0 + ib_row + offset0;
+    device const float       * y = (device const float       *) src1 + r1*ne10 + im*ne00*ne1;
+
+    float yl[32];
+    float sumf[N_DST]={0.f}, all_sum;
+
+    const int nb32 = nb * (QK_K / 32);
+
+    const int ix = tiisg;
+
+    device const float * y4 = y + 32 * ix;
+
+    iq1m_scale_t scale;
+
+    for (int ib32 = ix; ib32 < nb32; ib32 += 32) {
+
+        float4 sumy = {0.f};
+        for (int i = 0; i < 8; ++i) {
+            yl[i+ 0] = y4[i+ 0]; sumy[0] += yl[i+ 0];
+            yl[i+ 8] = y4[i+ 8]; sumy[1] += yl[i+ 8];
+            yl[i+16] = y4[i+16]; sumy[2] += yl[i+16];
+            yl[i+24] = y4[i+24]; sumy[3] += yl[i+24];
+        }
+
+        const int ibl = ib32 / (QK_K / 32);
+        const int ib  = ib32 % (QK_K / 32);
+
+        device const block_iq1_m * xr = x + ibl;
+        device const uint8_t  * qs = xr->qs + 4 * ib;
+        device const uint8_t  * qh = xr->qh + 2 * ib;
+        device const uint16_t * sc = (device const uint16_t *)xr->scales;
+
+        for (int row = 0; row < N_DST; row++) {
+            scale.u16 = (sc[0] >> 12) | ((sc[1] >> 8) & 0x00f0) | ((sc[2] >> 4) & 0x0f00) | (sc[3] & 0xf000);
+
+            constant uint8_t * grid1 = (constant uint8_t *)(iq1s_grid_gpu + (qs[0] | ((qh[0] << 8) & 0x700)));
+            constant uint8_t * grid2 = (constant uint8_t *)(iq1s_grid_gpu + (qs[1] | ((qh[0] << 4) & 0x700)));
+            constant uint8_t * grid3 = (constant uint8_t *)(iq1s_grid_gpu + (qs[2] | ((qh[1] << 8) & 0x700)));
+            constant uint8_t * grid4 = (constant uint8_t *)(iq1s_grid_gpu + (qs[3] | ((qh[1] << 4) & 0x700)));
+
+            float2 sum = {0.f};
+            for (int j = 0; j < 4; ++j) {
+                sum[0] += yl[j+ 0] * (grid1[j] & 0xf) + yl[j+ 4] * (grid1[j] >> 4)
+                        + yl[j+ 8] * (grid2[j] & 0xf) + yl[j+12] * (grid2[j] >> 4);
+                sum[1] += yl[j+16] * (grid3[j] & 0xf) + yl[j+20] * (grid3[j] >> 4)
+                        + yl[j+24] * (grid4[j] & 0xf) + yl[j+28] * (grid4[j] >> 4);
+            }
+            const float delta1 = sumy[0] * (qh[0] & 0x08 ? -1 - IQ1M_DELTA : -1 + IQ1M_DELTA) + sumy[1] * (qh[0] & 0x80 ? -1 - IQ1M_DELTA : -1 + IQ1M_DELTA);
+            const float delta2 = sumy[2] * (qh[1] & 0x08 ? -1 - IQ1M_DELTA : -1 + IQ1M_DELTA) + sumy[3] * (qh[1] & 0x80 ? -1 - IQ1M_DELTA : -1 + IQ1M_DELTA);
+
+            sumf[row] += (float)scale.f16 * ((sum[0] + delta1) * (2*((sc[ib/2] >> (6*(ib%2)+0)) & 7) + 1) +
+                                             (sum[1] + delta2) * (2*((sc[ib/2] >> (6*(ib%2)+3)) & 7) + 1));
+
+            sc += nb*sizeof(block_iq1_m)/2;
+            qs += nb*sizeof(block_iq1_m);
+            qh += nb*sizeof(block_iq1_m);
+        }
+
+        y4 += 32 * 32;
+    }
+
+    for (int row = 0; row < N_DST; ++row) {
+        all_sum = simd_sum(sumf[row]);
+        if (tiisg == 0) {
+            dst[r1*ne0 + im*ne0*ne1 + first_row + row] = all_sum;
+        }
+    }
+}
+
+void kernel_mul_mv_iq4_nl_f32_impl(
+        device const  void * src0,
+        device const float * src1,
+        device       float * dst,
+                   int64_t   ne00,
+                   int64_t   ne01,
+                   int64_t   ne02,
+                   int64_t   ne10,
+                   int64_t   ne12,
+                   int64_t   ne0,
+                   int64_t   ne1,
+                   uint      r2,
+                   uint      r3,
+        threadgroup int8_t * shared_values_i8,
+                   uint3     tgpig,
+                   uint      tiisg,
+                   uint      sgitg) {
+
+    threadgroup float * shared_values = (threadgroup float *)shared_values_i8;
+    const int nb = ne00/QK4_NL;
+    const int r0 = tgpig.x;
+    const int r1 = tgpig.y;
+    const int im = tgpig.z;
+    const int first_row = (r0 * 2 + sgitg) * 2;
+    const int ib_row = first_row * nb;
+
+    const uint i12 = im%ne12;
+    const uint i13 = im/ne12;
+
+    const uint offset0 = (i12/r2)*(nb*ne01) + (i13/r3)*(nb*ne01*ne02);
+    device const block_iq4_nl * x = (device const block_iq4_nl *) src0 + ib_row + offset0;
+    device const float        * y = (device const float        *) src1 + r1*ne10 + im*ne00*ne1;
+
+    const int ix = tiisg/2;  // 0...15
+    const int it = tiisg%2;  // 0 or 1
+
+    shared_values[tiisg] = kvalues_iq4nl_f[tiisg%16];
+    threadgroup_barrier(mem_flags::mem_threadgroup);
+
+    float4 yl[4];
+    float sumf[2]={0.f}, all_sum;
+
+    device const float * yb = y + ix * QK4_NL + it * 8;
+
+    uint32_t aux32[2];
+    thread const uint8_t * q8 = (thread const uint8_t *)aux32;
+
+    float4 qf1, qf2;
+
+    for (int ib = ix; ib < nb; ib += 16) {
+
+        device const float4 * y4 = (device const float4 *)yb;
+        yl[0] = y4[0]; yl[1] = y4[4]; yl[2] = y4[1]; yl[3] = y4[5];
+
+        for (int row = 0; row < 2 && first_row + row < ne01; ++row) {
+
+            device const block_iq4_nl & xb = x[row*nb + ib];
+            device const uint16_t * q4 = (device const uint16_t *)(xb.qs + 8*it);
+
+            float4 acc1 = {0.f}, acc2 = {0.f};
+
+            aux32[0] = q4[0] | (q4[1] << 16);
+            aux32[1] = (aux32[0] >> 4) & 0x0f0f0f0f;
+            aux32[0] &= 0x0f0f0f0f;
+            qf1 = {shared_values[q8[0]], shared_values[q8[1]], shared_values[q8[2]], shared_values[q8[3]]};
+            qf2 = {shared_values[q8[4]], shared_values[q8[5]], shared_values[q8[6]], shared_values[q8[7]]};
+            acc1 += yl[0] * qf1;
+            acc2 += yl[1] * qf2;
+
+            aux32[0] = q4[2] | (q4[3] << 16);
+            aux32[1] = (aux32[0] >> 4) & 0x0f0f0f0f;
+            aux32[0] &= 0x0f0f0f0f;
+            qf1 = {shared_values[q8[0]], shared_values[q8[1]], shared_values[q8[2]], shared_values[q8[3]]};
+            qf2 = {shared_values[q8[4]], shared_values[q8[5]], shared_values[q8[6]], shared_values[q8[7]]};
+            acc1 += yl[2] * qf1;
+            acc2 += yl[3] * qf2;
+
+            acc1 += acc2;
+
+            sumf[row] += (float)xb.d * (acc1[0] + acc1[1] + acc1[2] + acc1[3]);
+
+        }
+
+        yb += 16 * QK4_NL;
+    }
+
+    for (int row = 0; row < 2 && first_row + row < ne01; ++row) {
+        all_sum = simd_sum(sumf[row]);
+        if (tiisg == 0) {
+            dst[r1*ne0 + im*ne0*ne1 + first_row + row] = all_sum;
+        }
+    }
+}
+
+void kernel_mul_mv_iq4_xs_f32_impl(
+        device const  void * src0,
+        device const float * src1,
+        device       float * dst,
+                   int64_t   ne00,
+                   int64_t   ne01,
+                   int64_t   ne02,
+                   int64_t   ne10,
+                   int64_t   ne12,
+                   int64_t   ne0,
+                   int64_t   ne1,
+                   uint      r2,
+                   uint      r3,
+        threadgroup int8_t * shared_values_i8,
+                   uint3     tgpig,
+                   uint      tiisg,
+                   uint      sgitg) {
+
+    threadgroup float * shared_values = (threadgroup float *)shared_values_i8;
+    const int nb = ne00/QK_K;
+    const int r0 = tgpig.x;
+    const int r1 = tgpig.y;
+    const int im = tgpig.z;
+    const int first_row = (r0 * 2 + sgitg) * 2;
+    const int ib_row = first_row * nb;
+
+    const uint i12 = im%ne12;
+    const uint i13 = im/ne12;
+
+    const uint offset0 = (i12/r2)*(nb*ne01) + (i13/r3)*(nb*ne01*ne02);
+    device const block_iq4_xs * x = (device const block_iq4_xs *) src0 + ib_row + offset0;
+    device const float        * y = (device const float        *) src1 + r1*ne10 + im*ne00*ne1;
+
+    const int ix = tiisg/16;  // 0 or 1
+    const int it = tiisg%16;  // 0...15
+    const int ib = it/2;
+    const int il = it%2;
+
+    shared_values[tiisg] = kvalues_iq4nl_f[tiisg%16];
+    threadgroup_barrier(mem_flags::mem_threadgroup);
+
+    float4 yl[4];
+    float sumf[2]={0.f}, all_sum;
+
+    device const float * yb = y + ix * QK_K + ib * 32 + il * 8;
+
+    uint32_t aux32[2];
+    thread const uint8_t * q8 = (thread const uint8_t *)aux32;
+
+    float4 qf1, qf2;
+
+    for (int ibl = ix; ibl < nb; ibl += 2) {
+
+        device const float4 * y4 = (device const float4 *)yb;
+        yl[0] = y4[0]; yl[1] = y4[4]; yl[2] = y4[1]; yl[3] = y4[5];
+
+        for (int row = 0; row < 2; ++row) {
+
+            device const block_iq4_xs & xb = x[row*nb + ibl];
+            device const uint32_t * q4 = (device const uint32_t *)(xb.qs + 16*ib + 8*il);
+
+            float4 acc1 = {0.f}, acc2 = {0.f};
+
+            aux32[0] = q4[0] & 0x0f0f0f0f;
+            aux32[1] = (q4[0] >> 4) & 0x0f0f0f0f;
+            qf1 = {shared_values[q8[0]], shared_values[q8[1]], shared_values[q8[2]], shared_values[q8[3]]};
+            qf2 = {shared_values[q8[4]], shared_values[q8[5]], shared_values[q8[6]], shared_values[q8[7]]};
+            acc1 += yl[0] * qf1;
+            acc2 += yl[1] * qf2;
+
+            aux32[0] = q4[1] & 0x0f0f0f0f;
+            aux32[1] = (q4[1] >> 4) & 0x0f0f0f0f;
+            qf1 = {shared_values[q8[0]], shared_values[q8[1]], shared_values[q8[2]], shared_values[q8[3]]};
+            qf2 = {shared_values[q8[4]], shared_values[q8[5]], shared_values[q8[6]], shared_values[q8[7]]};
+            acc1 += yl[2] * qf1;
+            acc2 += yl[3] * qf2;
+
+            acc1 += acc2;
+
+            const int ls = (((xb.scales_l[ib/2] >> 4*(ib%2)) & 0xf) | (((xb.scales_h >> 2*ib) & 3) << 4)) - 32;
+            sumf[row] += (float)xb.d * ls * (acc1[0] + acc1[1] + acc1[2] + acc1[3]);
+
+        }
+
+        yb += 2 * QK_K;
+    }
+
+    for (int row = 0; row < 2; ++row) {
+        all_sum = simd_sum(sumf[row]);
+        if (tiisg == 0) {
+            dst[r1*ne0 + im*ne0*ne1 + first_row + row] = all_sum;
+        }
+    }
+}
+
+[[host_name("kernel_mul_mv_iq1_s_f32")]]
+kernel void kernel_mul_mv_iq1_s_f32(
+        device const  void * src0,
+        device const float * src1,
+        device       float * dst,
+        constant   int64_t & ne00,
+        constant   int64_t & ne01,
+        constant   int64_t & ne02,
+        constant  uint64_t & nb00,
+        constant  uint64_t & nb01,
+        constant  uint64_t & nb02,
+        constant   int64_t & ne10,
+        constant   int64_t & ne11,
+        constant   int64_t & ne12,
+        constant  uint64_t & nb10,
+        constant  uint64_t & nb11,
+        constant  uint64_t & nb12,
+        constant   int64_t & ne0,
+        constant   int64_t & ne1,
+        constant   uint    & r2,
+        constant   uint    & r3,
+        uint3 tgpig[[threadgroup_position_in_grid]],
+        uint  tiisg[[thread_index_in_simdgroup]],
+        uint  sgitg[[simdgroup_index_in_threadgroup]]) {
+
+    kernel_mul_mv_iq1_s_f32_impl(src0, src1, dst, ne00, ne01, ne02, ne10, ne12, ne0, ne1, r2, r3, nullptr, tgpig, tiisg, sgitg);
+}
+
+[[host_name("kernel_mul_mv_iq1_m_f32")]]
+kernel void kernel_mul_mv_iq1_m_f32(
+        device const  void * src0,
+        device const float * src1,
+        device       float * dst,
+        constant   int64_t & ne00,
+        constant   int64_t & ne01,
+        constant   int64_t & ne02,
+        constant  uint64_t & nb00,
+        constant  uint64_t & nb01,
+        constant  uint64_t & nb02,
+        constant   int64_t & ne10,
+        constant   int64_t & ne11,
+        constant   int64_t & ne12,
+        constant  uint64_t & nb10,
+        constant  uint64_t & nb11,
+        constant  uint64_t & nb12,
+        constant   int64_t & ne0,
+        constant   int64_t & ne1,
+        constant   uint    & r2,
+        constant   uint    & r3,
+        uint3 tgpig[[threadgroup_position_in_grid]],
+        uint  tiisg[[thread_index_in_simdgroup]],
+        uint  sgitg[[simdgroup_index_in_threadgroup]]) {
+
+    kernel_mul_mv_iq1_m_f32_impl(src0, src1, dst, ne00, ne01, ne02, ne10, ne12, ne0, ne1, r2, r3, nullptr, tgpig, tiisg, sgitg);
+}
+
+[[host_name("kernel_mul_mv_iq4_nl_f32")]]
+kernel void kernel_mul_mv_iq4_nl_f32(
+        device const  void * src0,
+        device const float * src1,
+        device       float * dst,
+        constant   int64_t & ne00,
+        constant   int64_t & ne01,
+        constant   int64_t & ne02,
+        constant  uint64_t & nb00,
+        constant  uint64_t & nb01,
+        constant  uint64_t & nb02,
+        constant   int64_t & ne10,
+        constant   int64_t & ne11,
+        constant   int64_t & ne12,
+        constant  uint64_t & nb10,
+        constant  uint64_t & nb11,
+        constant  uint64_t & nb12,
+        constant   int64_t & ne0,
+        constant   int64_t & ne1,
+        constant   uint    & r2,
+        constant   uint    & r3,
+        threadgroup int8_t * shared_values [[threadgroup(0)]],
+        uint3 tgpig[[threadgroup_position_in_grid]],
+        uint tiisg[[thread_index_in_simdgroup]],
+        uint sgitg[[simdgroup_index_in_threadgroup]]) {
+
+    kernel_mul_mv_iq4_nl_f32_impl(src0, src1, dst, ne00, ne01, ne02, ne10, ne12, ne0, ne1, r2, r3, shared_values, tgpig, tiisg, sgitg);
+}
+
+[[host_name("kernel_mul_mv_iq4_xs_f32")]]
+kernel void kernel_mul_mv_iq4_xs_f32(
+        device const  void * src0,
+        device const float * src1,
+        device       float * dst,
+        constant   int64_t & ne00,
+        constant   int64_t & ne01,
+        constant   int64_t & ne02,
+        constant  uint64_t & nb00,
+        constant  uint64_t & nb01,
+        constant  uint64_t & nb02,
+        constant   int64_t & ne10,
+        constant   int64_t & ne11,
+        constant   int64_t & ne12,
+        constant  uint64_t & nb10,
+        constant  uint64_t & nb11,
+        constant  uint64_t & nb12,
+        constant   int64_t & ne0,
+        constant   int64_t & ne1,
+        constant   uint    & r2,
+        constant   uint    & r3,
+        threadgroup int8_t * shared_values [[threadgroup(0)]],
+        uint3 tgpig[[threadgroup_position_in_grid]],
+        uint tiisg[[thread_index_in_simdgroup]],
+        uint sgitg[[simdgroup_index_in_threadgroup]]) {
+
+    kernel_mul_mv_iq4_xs_f32_impl(src0, src1, dst, ne00, ne01, ne02, ne10, ne12, ne0, ne1, r2, r3, shared_values, tgpig, tiisg, sgitg);
+}
+
+//============================= templates and their specializations =============================
+
+// NOTE: this is not dequantizing - we are simply fitting the template
+template <typename type4x4>
+void dequantize_f32(device const float4x4 * src, short il, thread type4x4 & reg) {
+    float4x4 temp = *(((device float4x4 *)src));
+    for (int i = 0; i < 16; i++){
+        reg[i/4][i%4] = temp[i/4][i%4];
+    }
+}
+
+template <typename type4x4>
+void dequantize_f16(device const half4x4 * src, short il, thread type4x4 & reg) {
+    half4x4 temp = *(((device half4x4 *)src));
+    for (int i = 0; i < 16; i++){
+        reg[i/4][i%4] = temp[i/4][i%4];
+    }
+}
+
+template <typename type4x4>
+void dequantize_q4_0(device const block_q4_0 *xb, short il, thread type4x4 & reg) {
+    device const uint16_t * qs = ((device const uint16_t *)xb + 1);
+    const float d1 = il ? (xb->d / 16.h) : xb->d;
+    const float d2 = d1 / 256.f;
+    const float md = -8.h * xb->d;
+    const ushort mask0 = il ? 0x00F0 : 0x000F;
+    const ushort mask1 = mask0 << 8;
+
+    for (int i=0;i<8;i++) {
+        reg[i/2][2*(i%2)+0] = d1 * (qs[i] & mask0) + md;
+        reg[i/2][2*(i%2)+1] = d2 * (qs[i] & mask1) + md;
+    }
+}
+
+template <typename type4x4>
+void dequantize_q4_1(device const block_q4_1 *xb, short il, thread type4x4 & reg) {
+    device const uint16_t * qs = ((device const uint16_t *)xb + 2);
+    const float d1 = il ? (xb->d / 16.h) : xb->d;
+    const float d2 = d1 / 256.f;
+    const float  m = xb->m;
+    const ushort mask0 = il ? 0x00F0 : 0x000F;
+    const ushort mask1 = mask0 << 8;
+
+    for (int i=0;i<8;i++) {
+        reg[i/2][2*(i%2)+0] = ((qs[i] & mask0) * d1) + m;
+        reg[i/2][2*(i%2)+1] = ((qs[i] & mask1) * d2) + m;
+    }
+}
+
+template <typename type4x4>
+void dequantize_q5_0(device const block_q5_0 *xb, short il, thread type4x4 & reg) {
+    device const uint16_t * qs = ((device const uint16_t *)xb + 3);
+    const float d = xb->d;
+    const float md = -16.h * xb->d;
+    const ushort mask = il ? 0x00F0 : 0x000F;
+
+    const uint32_t qh = *((device const uint32_t *)xb->qh);
+
+    const int x_mv = il ? 4 : 0;
+
+    const int gh_mv = il ? 12 : 0;
+    const int gh_bk = il ?  0 : 4;
+
+    for (int i = 0; i < 8; i++) {
+        // extract the 5-th bits for x0 and x1
+        const uint8_t xh_0 = ((qh >> (gh_mv + 2*i  )) << gh_bk) & 0x10;
+        const uint8_t xh_1 = ((qh >> (gh_mv + 2*i+1)) << gh_bk) & 0x10;
+
+        // combine the 4-bits from qs with the 5th bit
+        const int32_t x0 = ((((qs[i]     ) & mask) >> x_mv) | xh_0);
+        const int32_t x1 = ((((qs[i] >> 8) & mask) >> x_mv) | xh_1);
+
+        reg[i/2][2*(i%2)+0] = d * x0 + md;
+        reg[i/2][2*(i%2)+1] = d * x1 + md;
+    }
+}
+
+template <typename type4x4>
+void dequantize_q5_1(device const block_q5_1 *xb, short il, thread type4x4 & reg) {
+    device const uint16_t * qs = ((device const uint16_t *)xb + 4);
+    const float d = xb->d;
+    const float m = xb->m;
+    const ushort mask = il ? 0x00F0 : 0x000F;
+
+    const uint32_t qh = *((device const uint32_t *)xb->qh);
+
+    const int x_mv = il ? 4 : 0;
+
+    const int gh_mv = il ? 12 : 0;
+    const int gh_bk = il ?  0 : 4;
+
+    for (int i = 0; i < 8; i++) {
+        // extract the 5-th bits for x0 and x1
+        const uint8_t xh_0 = ((qh >> (gh_mv + 2*i  )) << gh_bk) & 0x10;
+        const uint8_t xh_1 = ((qh >> (gh_mv + 2*i+1)) << gh_bk) & 0x10;
+
+        // combine the 4-bits from qs with the 5th bit
+        const int32_t x0 = ((((qs[i]     ) & mask) >> x_mv) | xh_0);
+        const int32_t x1 = ((((qs[i] >> 8) & mask) >> x_mv) | xh_1);
+
+        reg[i/2][2*(i%2)+0] = d * x0 + m;
+        reg[i/2][2*(i%2)+1] = d * x1 + m;
+    }
+}
+
+template <typename type4x4>
+void dequantize_q8_0(device const block_q8_0 *xb, short il, thread type4x4 & reg) {
+    device const int8_t * qs = ((device const int8_t *)xb->qs);
+    const half d = xb->d;
+
+    for (int i = 0; i < 16; i++) {
+        reg[i/4][i%4] = (qs[i + 16*il] * d);
+    }
+}
+
+template <typename type4x4>
+void dequantize_q2_K(device const block_q2_K *xb, short il, thread type4x4 & reg) {
+    const float d = xb->d;
+    const float min = xb->dmin;
+    device const uint8_t * q = (device const uint8_t *)xb->qs;
+    float dl, ml;
+    uint8_t sc = xb->scales[il];
+
+    q = q + 32*(il/8) + 16*(il&1);
+    il = (il/2)%4;
+
+    half  coef = il>1 ? (il>2 ? 1/64.h : 1/16.h) : (il>0 ? 1/4.h : 1.h);
+    uchar mask = il>1 ? (il>2 ? 192    : 48)     : (il>0 ? 12    : 3);
+    dl = d * (sc & 0xF) * coef, ml = min * (sc >> 4);
+    for (int i = 0; i < 16; ++i) {
+        reg[i/4][i%4] = dl * (q[i] & mask) - ml;
+    }
+}
+
+template <typename type4x4>
+void dequantize_q3_K(device const block_q3_K *xb, short il, thread type4x4 & reg) {
+    const half d_all = xb->d;
+    device const uint8_t * q = (device const uint8_t *)xb->qs;
+    device const uint8_t * h = (device const uint8_t *)xb->hmask;
+    device const int8_t * scales = (device const int8_t *)xb->scales;
+
+    q = q + 32 * (il/8) + 16 * (il&1);
+    h = h + 16 * (il&1);
+    uint8_t m = 1 << (il/2);
+    uint16_t kmask1 = (il/4)>1 ? ((il/4)>2 ? 192 : 48) : \
+                                 ((il/4)>0 ? 12  : 3);
+    uint16_t kmask2 = il/8 ? 0xF0 : 0x0F;
+    uint16_t scale_2 = scales[il%8], scale_1 = scales[8 + il%4];
+    int16_t  dl_int = (il/4)&1 ? (scale_2&kmask2) | ((scale_1&kmask1) << 2)
+                               : (scale_2&kmask2) | ((scale_1&kmask1) << 4);
+    float dl = il<8 ? d_all * (dl_int - 32.f) : d_all * (dl_int / 16.f - 32.f);
+    const float ml = 4.f * dl;
+
+    il = (il/2) & 3;
+    const half    coef = il>1 ? (il>2 ? 1/64.h : 1/16.h) : (il>0 ? 1/4.h : 1.h);
+    const uint8_t mask = il>1 ? (il>2 ? 192    : 48)     : (il>0 ? 12    : 3);
+    dl *= coef;
+
+    for (int i = 0; i < 16; ++i) {
+        reg[i/4][i%4] = dl * (q[i] & mask) - (h[i] & m ? 0 : ml);
+    }
+}
+
+static inline uchar2 get_scale_min_k4_just2(int j, int k, device const uchar * q) {
+    return j < 4 ? uchar2{uchar(q[j+0+k] & 63), uchar(q[j+4+k] & 63)}
+                 : uchar2{uchar((q[j+4+k] & 0xF) | ((q[j-4+k] & 0xc0) >> 2)), uchar((q[j+4+k] >> 4) | ((q[j-0+k] & 0xc0) >> 2))};
+}
+
+template <typename type4x4>
+void dequantize_q4_K(device const block_q4_K *xb, short il, thread type4x4 & reg) {
+    device const uchar * q = xb->qs;
+
+    short is = (il/4) * 2;
+    q = q + (il/4) * 32 + 16 * (il&1);
+    il = il & 3;
+    const uchar2 sc = get_scale_min_k4_just2(is, il/2, xb->scales);
+    const float d   = il < 2 ? xb->d : xb->d / 16.h;
+    const float min = xb->dmin;
+    const float dl = d * sc[0];
+    const float ml = min * sc[1];
+
+    const ushort mask = il<2 ? 0x0F : 0xF0;
+    for (int i = 0; i < 16; ++i) {
+        reg[i/4][i%4] = dl * (q[i] & mask) - ml;
+    }
+}
+
+template <typename type4x4>
+void dequantize_q5_K(device const block_q5_K *xb, short il, thread type4x4 & reg) {
+    device const uint8_t * q  = xb->qs;
+    device const uint8_t * qh = xb->qh;
+
+    short is = (il/4) * 2;
+    q  = q + 32 * (il/4) + 16 * (il&1);
+    qh = qh + 16 * (il&1);
+    uint8_t ul = 1 << (il/2);
+    il = il & 3;
+    const uchar2 sc = get_scale_min_k4_just2(is, il/2, xb->scales);
+    const float d = il < 2 ? xb->d : xb->d / 16.f;
+    const float min = xb->dmin;
+    const float dl = d * sc[0];
+    const float ml = min * sc[1];
+
+    const ushort mask  = il<2 ? 0x0F : 0xF0;
+    const float qh_val = il<2 ? 16.f : 256.f;
+    for (int i = 0; i < 16; ++i) {
+        reg[i/4][i%4] = dl * ((q[i] & mask) + (qh[i] & ul ? qh_val : 0)) - ml;
+    }
+}
+
+template <typename type4x4>
+void dequantize_q6_K(device const block_q6_K *xb, short il, thread type4x4 & reg) {
+    const half d_all = xb->d;
+    device const uint8_t * ql = (device const uint8_t *)xb->ql;
+    device const uint8_t * qh = (device const uint8_t *)xb->qh;
+    device const int8_t * scales = (device const int8_t *)xb->scales;
+
+    ql = ql + 64*(il/8) + 32*((il/2)&1) + 16*(il&1);
+    qh = qh + 32*(il/8) + 16*(il&1);
+    float sc = scales[(il%2) + 2 * ((il/2))];
+    il = (il/2) & 3;
+
+    const uint16_t  kmask1 = il>1 ? (il>2 ? 192 : 48) : (il>0 ? 12 : 3);
+    const uint16_t  kmask2 = il>1 ? 0xF0              : 0x0F;
+    const float       coef = il>1 ? 1.f/16.f          : 1.f;
+    const float ml = d_all * sc * 32.f;
+    const float dl = d_all * sc * coef;
+    for (int i = 0; i < 16; ++i) {
+        const half q = il&1 ? ((ql[i] & kmask2) | ((qh[i] & kmask1) << 2))
+                            : ((ql[i] & kmask2) | ((qh[i] & kmask1) << 4));
+        reg[i/4][i%4] = dl * q - ml;
+    }
+}
+
+template <typename type4x4>
+void dequantize_iq2_xxs(device const block_iq2_xxs * xb, short il, thread type4x4 & reg) {
+    // il is 0...15 for QK_K = 256 => index of block of 32 is il/2
+    const float d = xb->d;
+    const int ib32 = il/2;
+    il = il%2;
+    // il = 0 or 1. il = 0 processes the first 16 quants in a block of 32, il = 1 the second 16
+    // each block of 32 needs 2 uint32_t's for the quants & scale, so 4 uint16_t's.
+    device const uint16_t * q2 = xb->qs + 4*ib32;
+    const uint32_t aux32_g = q2[0] | (q2[1] << 16);
+    const uint32_t aux32_s = q2[2] | (q2[3] << 16);
+    thread const uint8_t * aux8 = (thread const uint8_t *)&aux32_g;
+    const float dl = d * (0.5f + (aux32_s >> 28)) * 0.25f;
+    constant uint8_t * grid = (constant uint8_t *)(iq2xxs_grid + aux8[2*il+0]);
+    uint8_t signs = ksigns_iq2xs[(aux32_s >> 14*il) & 127];
+    for (int i = 0; i < 8; ++i) {
+        reg[i/4][i%4] = dl * grid[i] * (signs & kmask_iq2xs[i] ? -1.f : 1.f);
+    }
+    grid = (constant uint8_t *)(iq2xxs_grid + aux8[2*il+1]);
+    signs = ksigns_iq2xs[(aux32_s >> (14*il+7)) & 127];
+    for (int i = 0; i < 8; ++i) {
+        reg[2+i/4][i%4] = dl * grid[i] * (signs & kmask_iq2xs[i] ? -1.f : 1.f);
+    }
+}
+
+template <typename type4x4>
+void dequantize_iq2_xs(device const block_iq2_xs * xb, short il, thread type4x4 & reg) {
+    // il is 0...15 for QK_K = 256 => index of block of 32 is il/2
+    const float d = xb->d;
+    const int ib32 = il/2;
+    il = il%2;
+    // il = 0 or 1. il = 0 processes the first 16 quants in a block of 32, il = 1 the second 16
+    device const uint16_t * q2 = xb->qs + 4*ib32;
+    const float dl = d * (0.5f + ((xb->scales[ib32] >> 4*il) & 0xf)) * 0.25f;
+    constant uint8_t * grid = (constant uint8_t *)(iq2xs_grid + (q2[2*il+0] & 511));
+    uint8_t signs = ksigns_iq2xs[q2[2*il+0] >> 9];
+    for (int i = 0; i < 8; ++i) {
+        reg[i/4][i%4] = dl * grid[i] * (signs & kmask_iq2xs[i] ? -1.f : 1.f);
+    }
+    grid = (constant uint8_t *)(iq2xs_grid + (q2[2*il+1] & 511));
+    signs = ksigns_iq2xs[q2[2*il+1] >> 9];
+    for (int i = 0; i < 8; ++i) {
+        reg[2+i/4][i%4] = dl * grid[i] * (signs & kmask_iq2xs[i] ? -1.f : 1.f);
+    }
+}
+
+template <typename type4x4>
+void dequantize_iq3_xxs(device const block_iq3_xxs * xb, short il, thread type4x4 & reg) {
+    // il is 0...15 for QK_K = 256 => index of block of 32 is il/2
+    const float d = xb->d;
+    const int ib32 = il/2;
+    il = il%2;
+    // il = 0 or 1. il = 0 processes the first 16 quants in a block of 32, il = 1 the second 16
+    device const uint8_t * q3 = xb->qs + 8*ib32;
+    device const uint16_t * gas = (device const uint16_t *)(xb->qs + QK_K/4) + 2*ib32;
+    const uint32_t aux32 = gas[0] | (gas[1] << 16);
+    const float dl = d * (0.5f + (aux32 >> 28)) * 0.5f;
+    constant uint8_t * grid1 = (constant uint8_t *)(iq3xxs_grid + q3[4*il+0]);
+    constant uint8_t * grid2 = (constant uint8_t *)(iq3xxs_grid + q3[4*il+1]);
+    uint8_t signs = ksigns_iq2xs[(aux32 >> 14*il) & 127];
+    for (int i = 0; i < 4; ++i) {
+        reg[0][i] = dl * grid1[i] * (signs & kmask_iq2xs[i+0] ? -1.f : 1.f);
+        reg[1][i] = dl * grid2[i] * (signs & kmask_iq2xs[i+4] ? -1.f : 1.f);
+    }
+    grid1 = (constant uint8_t *)(iq3xxs_grid + q3[4*il+2]);
+    grid2 = (constant uint8_t *)(iq3xxs_grid + q3[4*il+3]);
+    signs = ksigns_iq2xs[(aux32 >> (14*il+7)) & 127];
+    for (int i = 0; i < 4; ++i) {
+        reg[2][i] = dl * grid1[i] * (signs & kmask_iq2xs[i+0] ? -1.f : 1.f);
+        reg[3][i] = dl * grid2[i] * (signs & kmask_iq2xs[i+4] ? -1.f : 1.f);
+    }
+}
+
+template <typename type4x4>
+void dequantize_iq3_s(device const block_iq3_s * xb, short il, thread type4x4 & reg) {
+    // il is 0...15 for QK_K = 256 => index of block of 32 is il/2
+    const float d = xb->d;
+    const int ib32 = il/2;
+    il = il%2;
+    // il = 0 or 1. il = 0 processes the first 16 quants in a block of 32, il = 1 the second 16
+    device const uint8_t * qs = xb->qs + 8*ib32;
+    device const uint8_t * signs = xb->signs + 4*ib32 + 2*il;
+    const uint8_t qh = xb->qh[ib32] >> 4*il;
+    const float dl = d * (1 + 2*((xb->scales[ib32/2] >> 4*(ib32%2)) & 0xf));
+    constant uint8_t * grid1 = (constant uint8_t *)(iq3s_grid + (qs[4*il+0] | ((qh << 8) & 256)));
+    constant uint8_t * grid2 = (constant uint8_t *)(iq3s_grid + (qs[4*il+1] | ((qh << 7) & 256)));
+    for (int i = 0; i < 4; ++i) {
+        reg[0][i] = dl * grid1[i] * select(1, -1, signs[0] & kmask_iq2xs[i+0]);
+        reg[1][i] = dl * grid2[i] * select(1, -1, signs[0] & kmask_iq2xs[i+4]);
+    }
+    grid1 = (constant uint8_t *)(iq3s_grid + (qs[4*il+2] | ((qh << 6) & 256)));
+    grid2 = (constant uint8_t *)(iq3s_grid + (qs[4*il+3] | ((qh << 5) & 256)));
+    for (int i = 0; i < 4; ++i) {
+        reg[2][i] = dl * grid1[i] * select(1, -1, signs[1] & kmask_iq2xs[i+0]);
+        reg[3][i] = dl * grid2[i] * select(1, -1, signs[1] & kmask_iq2xs[i+4]);
+    }
+}
+
+template <typename type4x4>
+void dequantize_iq2_s(device const block_iq2_s * xb, short il, thread type4x4 & reg) {
+    // il is 0...15 for QK_K = 256 => index of block of 32 is il/2
+    const float d = xb->d;
+    const int ib32 = il/2;
+    il = il%2;
+    // il = 0 or 1. il = 0 processes the first 16 quants in a block of 32, il = 1 the second 16
+    device const uint8_t * qs = xb->qs + 4*ib32 + 2*il;
+    device const uint8_t * signs = qs + QK_K/8;
+    const uint8_t qh = xb->qh[ib32] >> 4*il;
+    const float dl = d * (0.5f + ((xb->scales[ib32] >> 4*il) & 0xf)) * 0.25f;
+    constant uint8_t * grid1 = (constant uint8_t *)(iq2s_grid + (qs[0] | ((qh << 8) & 0x300)));
+    constant uint8_t * grid2 = (constant uint8_t *)(iq2s_grid + (qs[1] | ((qh << 6) & 0x300)));
+    for (int i = 0; i < 8; ++i) {
+        reg[i/4+0][i%4] = dl * grid1[i] * select(1, -1, signs[0] & kmask_iq2xs[i]);
+        reg[i/4+2][i%4] = dl * grid2[i] * select(1, -1, signs[1] & kmask_iq2xs[i]);
+    }
+}
+
+template <typename type4x4>
+void dequantize_iq1_s(device const block_iq1_s * xb, short il, thread type4x4 & reg) {
+    // il is 0...15 for QK_K = 256 => index of block of 32 is il/2
+    const int ib32 = il/2;
+    il = il%2;
+    const float d = xb->d;
+    device const uint8_t  * qs = xb->qs + 4*ib32 + 2*il;
+    device const uint16_t * qh = xb->qh;
+    const float dl = d * (2*((qh[ib32] >> 12) & 7) + 1);
+    const float ml = dl * (qh[ib32] & 0x8000 ? -1 - IQ1S_DELTA : -1 + IQ1S_DELTA);
+    const uint16_t h = qh[ib32] >> 6*il;
+    constant uint8_t * grid1 = (constant uint8_t *)(iq1s_grid_gpu + (qs[0] | ((h << 8) & 0x700)));
+    constant uint8_t * grid2 = (constant uint8_t *)(iq1s_grid_gpu + (qs[1] | ((h << 5) & 0x700)));
+    for (int i = 0; i < 4; ++i) {
+        reg[0][i] = dl * (grid1[i] & 0xf) + ml;
+        reg[1][i] = dl * (grid1[i] >>  4) + ml;
+        reg[2][i] = dl * (grid2[i] & 0xf) + ml;
+        reg[3][i] = dl * (grid2[i] >>  4) + ml;
+    }
+}
+
+template <typename type4x4>
+void dequantize_iq1_m(device const block_iq1_m * xb, short il, thread type4x4 & reg) {
+    // il is 0...15 for QK_K = 256 => index of block of 32 is il/2
+    const int ib32 = il/2;
+    il = il%2;
+    device const uint16_t * sc = (device const uint16_t *)xb->scales;
+
+    iq1m_scale_t scale;
+    scale.u16 = (sc[0] >> 12) | ((sc[1] >> 8) & 0x00f0) | ((sc[2] >> 4) & 0x0f00) | (sc[3] & 0xf000);
+    const float d = scale.f16;
+
+    device const uint8_t * qs = xb->qs + 4*ib32 + 2*il;
+    device const uint8_t * qh = xb->qh + 2*ib32 + il;
+
+    const float dl  = d * (2*((sc[ib32/2] >> (6*(ib32%2)+3*il)) & 7) + 1);
+    const float ml1 = dl * (qh[0] & 0x08 ? -1 - IQ1M_DELTA : -1 + IQ1M_DELTA);
+    const float ml2 = dl * (qh[0] & 0x80 ? -1 - IQ1M_DELTA : -1 + IQ1M_DELTA);
+    constant uint8_t * grid1 = (constant uint8_t *)(iq1s_grid_gpu + (qs[0] | ((qh[0] << 8) & 0x700)));
+    constant uint8_t * grid2 = (constant uint8_t *)(iq1s_grid_gpu + (qs[1] | ((qh[0] << 4) & 0x700)));
+    for (int i = 0; i < 4; ++i) {
+        reg[0][i] = dl * (grid1[i] & 0xf) + ml1;
+        reg[1][i] = dl * (grid1[i] >>  4) + ml1;
+        reg[2][i] = dl * (grid2[i] & 0xf) + ml2;
+        reg[3][i] = dl * (grid2[i] >>  4) + ml2;
+    }
+}
+
+template <typename type4x4>
+void dequantize_iq4_nl(device const block_iq4_nl * xb, short il, thread type4x4 & reg) {
+    device const uint16_t * q4 = (device const uint16_t *)xb->qs;
+    const float d = xb->d;
+    uint32_t aux32;
+    thread const uint8_t * q8 = (thread const uint8_t *)&aux32;
+    for (int i = 0; i < 4; ++i) {
+        aux32 = ((q4[2*i] | (q4[2*i+1] << 16)) >> 4*il) & 0x0f0f0f0f;
+        reg[i][0] = d * kvalues_iq4nl_f[q8[0]];
+        reg[i][1] = d * kvalues_iq4nl_f[q8[1]];
+        reg[i][2] = d * kvalues_iq4nl_f[q8[2]];
+        reg[i][3] = d * kvalues_iq4nl_f[q8[3]];
+    }
+}
+
+template <typename type4x4>
+void dequantize_iq4_xs(device const block_iq4_xs * xb, short il, thread type4x4 & reg) {
+    // il is 0...15 for QK_K = 256 => index of block of 32 is il/2
+    const int ib32 = il/2;
+    il = il%2;
+    // il = 0 or 1. il = 0 processes the first 16 quants in a block of 32, il = 1 the second 16
+    device const uint32_t * q4 = (device const uint32_t *)xb->qs + 4*ib32;
+    const int ls = ((xb->scales_l[ib32/2] >> 4*(ib32%2)) & 0xf) | (((xb->scales_h >> 2*ib32) & 3) << 4);
+    const float d = (float)xb->d * (ls - 32);
+    uint32_t aux32;
+    thread const uint8_t * q8 = (thread const uint8_t *)&aux32;
+    for (int i = 0; i < 4; ++i) {
+        aux32 = (q4[i] >> 4*il) & 0x0f0f0f0f;
+        reg[i][0] = d * kvalues_iq4nl_f[q8[0]];
+        reg[i][1] = d * kvalues_iq4nl_f[q8[1]];
+        reg[i][2] = d * kvalues_iq4nl_f[q8[2]];
+        reg[i][3] = d * kvalues_iq4nl_f[q8[3]];
+    }
+}
+
+template<typename block_q, short nl, void (*dequantize_func)(device const block_q *, short, thread float4x4 &)>
+kernel void kernel_get_rows_q(
+        device const  void * src0,
+        device const  void * src1,
+        device       float * dst,
+        constant   int64_t & ne00,
+        constant  uint64_t & nb01,
+        constant  uint64_t & nb02,
+        constant   int64_t & ne10,
+        constant  uint64_t & nb10,
+        constant  uint64_t & nb11,
+        constant  uint64_t & nb1,
+        constant  uint64_t & nb2,
+        uint3                tgpig[[threadgroup_position_in_grid]],
+        uint                 tiitg[[thread_index_in_threadgroup]],
+        uint3                tptg [[threads_per_threadgroup]]) {
+    const int64_t i10 = tgpig.x;
+    const int64_t i11 = tgpig.y;
+
+    const int64_t r = ((const device int32_t *) ((const device char *) src1 + i11*nb11 + i10*nb10))[0];
+
+    const int64_t i02 = i11;
+
+    for (int64_t ind = tiitg; ind < ne00/16; ind += tptg.x) {
+        float4x4 temp;
+        dequantize_func(((device const block_q *) ((const device char *) src0 + r*nb01 + i02*nb02)) + ind/nl, ind%nl, temp);
+        *(((device float4x4 *) ((device char *) dst + i11*nb2 + i10*nb1)) + ind) = temp;
+    }
+}
+
+template<typename T>
+kernel void kernel_get_rows_f(
+        device const  void * src0,
+        device const  void * src1,
+        device       float * dst,
+        constant   int64_t & ne00,
+        constant  uint64_t & nb01,
+        constant  uint64_t & nb02,
+        constant   int64_t & ne10,
+        constant  uint64_t & nb10,
+        constant  uint64_t & nb11,
+        constant  uint64_t & nb1,
+        constant  uint64_t & nb2,
+        uint3                tgpig[[threadgroup_position_in_grid]],
+        uint                 tiitg[[thread_index_in_threadgroup]],
+        uint3                tptg [[threads_per_threadgroup]]) {
+    const int64_t i10 = tgpig.x;
+    const int64_t i11 = tgpig.y;
+
+    const int64_t r = ((const device int32_t *) ((const device char *) src1 + i11*nb11 + i10*nb10))[0];
+
+    const int64_t i02 = i11;
+
+    for (int ind = tiitg; ind < ne00; ind += tptg.x) {
+        ((      device float *) ((      device char *)  dst + i11*nb2  + i10*nb1))[ind] =
+        ((const device T     *) ((const device char *) src0 + i02*nb02 +  r*nb01))[ind];
+    }
+}
+
+kernel void kernel_get_rows_i32(
+        device const  void * src0,
+        device const  void * src1,
+        device     int32_t * dst,
+        constant   int64_t & ne00,
+        constant  uint64_t & nb01,
+        constant  uint64_t & nb02,
+        constant   int64_t & ne10,
+        constant  uint64_t & nb10,
+        constant  uint64_t & nb11,
+        constant  uint64_t & nb1,
+        constant  uint64_t & nb2,
+        uint3                tgpig[[threadgroup_position_in_grid]],
+        uint                 tiitg[[thread_index_in_threadgroup]],
+        uint3                tptg [[threads_per_threadgroup]]) {
+    const int64_t i10 = tgpig.x;
+    const int64_t i11 = tgpig.y;
+
+    const int64_t r = ((const device int32_t *) ((const device char *) src1 + i11*nb11 + i10*nb10))[0];
+
+    const int64_t i02 = i11;
+
+    for (int ind = tiitg; ind < ne00; ind += tptg.x) {
+        ((      device int32_t *) ((      device char *) dst  + i11*nb2 + i10*nb1))[ind] =
+        ((const device int32_t *) ((const device char *) src0 + i02*nb02 + r*nb01))[ind];
+    }
+}
+
+
+#define BLOCK_SIZE_M 64 // 8 simdgroup matrices from matrix A
+#define BLOCK_SIZE_N 32 // 4 simdgroup matrices from matrix B
+#define BLOCK_SIZE_K 32
+#define THREAD_MAT_M 4 // each thread take 4 simdgroup matrices from matrix A
+#define THREAD_MAT_N 2 // each thread take 2 simdgroup matrices from matrix B
+#define THREAD_PER_BLOCK 128
+#define THREAD_PER_ROW 2 // 2 thread for each row in matrix A to load numbers
+#define THREAD_PER_COL 4 // 4 thread for each row in matrix B to load numbers
+#define SG_MAT_SIZE 64 // simdgroup matrix is of shape 8x8
+#define SG_MAT_ROW 8
+
+// each block_q contains 16*nl weights
+template<typename T, typename T4x4, typename simdgroup_T8x8, typename block_q, short nl, void (*dequantize_func)(device const block_q *, short, thread T4x4 &)>
+kernel void kernel_mul_mm(device const  uchar * src0,
+                          device const  uchar * src1,
+                          device        float * dst,
+                          constant    int64_t & ne00,
+                          constant    int64_t & ne02,
+                          constant   uint64_t & nb01,
+                          constant   uint64_t & nb02,
+                          constant    int64_t & ne12,
+                          constant   uint64_t & nb10,
+                          constant   uint64_t & nb11,
+                          constant   uint64_t & nb12,
+                          constant    int64_t & ne0,
+                          constant    int64_t & ne1,
+                          constant       uint & r2,
+                          constant       uint & r3,
+                          threadgroup   uchar * shared_memory [[threadgroup(0)]],
+                          uint3                 tgpig[[threadgroup_position_in_grid]],
+                          uint                  tiitg[[thread_index_in_threadgroup]],
+                          uint                  sgitg[[simdgroup_index_in_threadgroup]]) {
+
+    threadgroup T     * sa = (threadgroup T     *)(shared_memory);
+    threadgroup float * sb = (threadgroup float *)(shared_memory + 4096);
+
+    const uint r0 = tgpig.y;
+    const uint r1 = tgpig.x;
+    const uint im = tgpig.z;
+
+    // if this block is of 64x32 shape or smaller
+    short n_rows = (ne0 - r0 * BLOCK_SIZE_M < BLOCK_SIZE_M) ? (ne0 - r0 * BLOCK_SIZE_M) : BLOCK_SIZE_M;
+    short n_cols = (ne1 - r1 * BLOCK_SIZE_N < BLOCK_SIZE_N) ? (ne1 - r1 * BLOCK_SIZE_N) : BLOCK_SIZE_N;
+
+    // a thread shouldn't load data outside of the matrix
+    short thread_row = ((short)tiitg/THREAD_PER_ROW) < n_rows ? ((short)tiitg/THREAD_PER_ROW) : n_rows - 1;
+    short thread_col = ((short)tiitg/THREAD_PER_COL) < n_cols ? ((short)tiitg/THREAD_PER_COL) : n_cols - 1;
+
+    simdgroup_T8x8     ma[4];
+    simdgroup_float8x8 mb[2];
+    simdgroup_float8x8 c_res[8];
+    for (int i = 0; i < 8; i++){
+        c_res[i] = make_filled_simdgroup_matrix<float, 8>(0.f);
+    }
+
+    short il = (tiitg % THREAD_PER_ROW);
+
+    const uint i12 = im%ne12;
+    const uint i13 = im/ne12;
+
+    uint   offset0 = (i12/r2)*nb02 + (i13/r3)*(nb02*ne02);
+    ushort offset1 = il/nl;
+
+    device const block_q * x = (device const block_q *)(src0 + (r0 * BLOCK_SIZE_M + thread_row) * nb01 + offset0) + offset1;
+    device const float   * y = (device const float   *)(src1
+        + nb12 * im
+        + nb11 * (r1 * BLOCK_SIZE_N + thread_col)
+        + nb10 * (BLOCK_SIZE_K / THREAD_PER_COL * (tiitg % THREAD_PER_COL)));
+
+    for (int loop_k = 0; loop_k < ne00; loop_k += BLOCK_SIZE_K) {
+        // load data and store to threadgroup memory
+        T4x4 temp_a;
+        dequantize_func(x, il, temp_a);
+        threadgroup_barrier(mem_flags::mem_threadgroup);
+
+        #pragma unroll(16)
+        for (int i = 0; i < 16; i++) {
+            *(sa + SG_MAT_SIZE * ((tiitg / THREAD_PER_ROW / 8) \
+            +                     (tiitg % THREAD_PER_ROW) * 16 + (i / 8) * 8) \
+            +                     (tiitg / THREAD_PER_ROW) % 8  + (i & 7) * 8) = temp_a[i/4][i%4];
+        }
+
+        *(threadgroup float2x4 *)(sb + (tiitg % THREAD_PER_COL) * 8 * 32 + 8 * (tiitg / THREAD_PER_COL)) = *((device float2x4 *)y);
+
+        il = (il + 2 < nl) ? il + 2 : il % 2;
+        x  = (il < 2) ? x + (2+nl-1)/nl : x;
+        y += BLOCK_SIZE_K;
+
+        threadgroup_barrier(mem_flags::mem_threadgroup);
+
+        // load matrices from threadgroup memory and conduct outer products
+        threadgroup T     * lsma = (sa + THREAD_MAT_M * SG_MAT_SIZE * (sgitg % 2));
+        threadgroup float * lsmb = (sb + THREAD_MAT_N * SG_MAT_SIZE * (sgitg / 2));
+
+        #pragma unroll(4)
+        for (int ik = 0; ik < BLOCK_SIZE_K / 8; ik++) {
+            #pragma unroll(4)
+            for (int i = 0; i < 4; i++) {
+                simdgroup_load(ma[i],lsma + SG_MAT_SIZE * i);
+            }
+            simdgroup_barrier(mem_flags::mem_none);
+            #pragma unroll(2)
+            for (int i = 0; i < 2; i++) {
+                simdgroup_load(mb[i],lsmb + SG_MAT_SIZE * i);
+            }
+
+            lsma += BLOCK_SIZE_M / SG_MAT_ROW * SG_MAT_SIZE;
+            lsmb += BLOCK_SIZE_N / SG_MAT_ROW * SG_MAT_SIZE;
+
+            #pragma unroll(8)
+            for (int i = 0; i < 8; i++){
+                simdgroup_multiply_accumulate(c_res[i], mb[i/4], ma[i%4], c_res[i]);
+            }
+        }
+    }
+
+    if ((r0 + 1) * BLOCK_SIZE_M <= ne0 && (r1 + 1) * BLOCK_SIZE_N <= ne1) {
+        device float * C = dst + (BLOCK_SIZE_M * r0 + 32 * (sgitg &  1)) \
+                               + (BLOCK_SIZE_N * r1 + 16 * (sgitg >> 1)) * ne0 + im*ne1*ne0;
+        for (int i = 0; i < 8; i++) {
+            simdgroup_store(c_res[i], C + 8 * (i%4) + 8 * ne0 * (i/4), ne0);
+        }
+    } else {
+        // block is smaller than 64x32, we should avoid writing data outside of the matrix
+        threadgroup_barrier(mem_flags::mem_threadgroup);
+        threadgroup float * temp_str = ((threadgroup float *)shared_memory) \
+                                      + 32 * (sgitg&1) + (16 * (sgitg>>1)) * BLOCK_SIZE_M;
+        for (int i = 0; i < 8; i++) {
+            simdgroup_store(c_res[i], temp_str + 8 * (i%4) + 8 * BLOCK_SIZE_M * (i/4), BLOCK_SIZE_M);
+        }
+
+        threadgroup_barrier(mem_flags::mem_threadgroup);
+
+        device float * C = dst + (BLOCK_SIZE_M * r0) + (BLOCK_SIZE_N * r1) * ne0 + im*ne1*ne0;
+        if (sgitg == 0) {
+            for (int i = 0; i < n_rows; i++) {
+                for (int j = tiitg; j < n_cols; j += BLOCK_SIZE_N) {
+                    *(C + i + j * ne0) = *(temp_str + i + j * BLOCK_SIZE_M);
+                }
+            }
+        }
+    }
+}
+
+// same as kernel_mul_mm_impl, but src1 and dst are accessed via indices stored in rowids
+template<typename block_q, short nl, void (*dequantize_func)(device const block_q *, short, thread half4x4 &)>
+void kernel_mul_mm_id_impl(
+        device const  uchar * src0,
+        device const  uchar * src1,
+        threadgroup ushort2 * rowids,
+        device        float * dst,
+        constant    int64_t & ne00,
+        constant    int64_t & ne02,
+        constant   uint64_t & nb01,
+        constant   uint64_t & nb02,
+        constant    int64_t & ne11,
+        constant    int64_t & ne12,
+        constant   uint64_t & nb10,
+        constant   uint64_t & nb11,
+        constant   uint64_t & nb12,
+        constant    int64_t & ne0,
+                    int64_t   ne1,
+                    int64_t   ne0ne1,
+        threadgroup   uchar * shared_memory,
+        uint3                 tgpig[[threadgroup_position_in_grid]],
+        uint                  tiitg[[thread_index_in_threadgroup]],
+        uint                  sgitg[[simdgroup_index_in_threadgroup]]) {
+
+    threadgroup half  * sa = (threadgroup half  *)(shared_memory);
+    threadgroup float * sb = (threadgroup float *)(shared_memory + 4096);
+
+    const uint r0 = tgpig.y;
+    const uint r1 = tgpig.x;
+
+    if (r1 * BLOCK_SIZE_N >= ne1) return;
+
+    // if this block is of 64x32 shape or smaller
+    short n_rows = (ne0 - r0 * BLOCK_SIZE_M < BLOCK_SIZE_M) ? (ne0 - r0 * BLOCK_SIZE_M) : BLOCK_SIZE_M;
+    short n_cols = (ne1 - r1 * BLOCK_SIZE_N < BLOCK_SIZE_N) ? (ne1 - r1 * BLOCK_SIZE_N) : BLOCK_SIZE_N;
+
+    // a thread shouldn't load data outside of the matrix
+    short thread_row = ((short)tiitg/THREAD_PER_ROW) < n_rows ? ((short)tiitg/THREAD_PER_ROW) : n_rows - 1;
+    short thread_col = ((short)tiitg/THREAD_PER_COL) < n_cols ? ((short)tiitg/THREAD_PER_COL) : n_cols - 1;
+
+    simdgroup_half8x8  ma[4];
+    simdgroup_float8x8 mb[2];
+    simdgroup_float8x8 c_res[8];
+    for (int i = 0; i < 8; i++){
+        c_res[i] = make_filled_simdgroup_matrix<float, 8>(0.f);
+    }
+    short il = (tiitg % THREAD_PER_ROW);
+
+    ushort offset1 = il/nl;
+
+    threadgroup const auto & id = rowids[r1 * BLOCK_SIZE_N + thread_col];
+
+    device const block_q * x = (device const block_q *)(src0 + (r0 * BLOCK_SIZE_M + thread_row) * nb01) + offset1;
+    device const float   * y = (device const float   *)(src1
+        + nb12 * id[1]
+        + nb11 * (id[0] % ne11)
+        + nb10 * (BLOCK_SIZE_K / THREAD_PER_COL * (tiitg % THREAD_PER_COL)));
+
+    for (int loop_k = 0; loop_k < ne00; loop_k += BLOCK_SIZE_K) {
+        // load data and store to threadgroup memory
+        half4x4 temp_a;
+        dequantize_func(x, il, temp_a);
+        threadgroup_barrier(mem_flags::mem_threadgroup);
+
+        for (int i = 0; i < 16; i++) {
+            *(sa + SG_MAT_SIZE * ((tiitg / THREAD_PER_ROW / 8) \
+            +                     (tiitg % THREAD_PER_ROW) * 16 + (i / 8) * 8) \
+            +                     (tiitg / THREAD_PER_ROW) % 8  + (i & 7) * 8) = temp_a[i/4][i%4];
+        }
+
+        *(threadgroup float2x4 *)(sb + (tiitg % THREAD_PER_COL) * 8 * 32 + 8 * (tiitg / THREAD_PER_COL)) = *((device float2x4 *)y);
+
+        il = (il + 2 < nl) ? il + 2 : il % 2;
+        x  = (il < 2) ? x + (2+nl-1)/nl : x;
+        y += BLOCK_SIZE_K;
+
+        threadgroup_barrier(mem_flags::mem_threadgroup);
+
+        // load matrices from threadgroup memory and conduct outer products
+        threadgroup half  * lsma = (sa + THREAD_MAT_M * SG_MAT_SIZE * (sgitg % 2));
+        threadgroup float * lsmb = (sb + THREAD_MAT_N * SG_MAT_SIZE * (sgitg / 2));
+
+        for (int ik = 0; ik < BLOCK_SIZE_K / 8; ik++) {
+            for (int i = 0; i < 4; i++) {
+                simdgroup_load(ma[i], lsma + SG_MAT_SIZE * i);
+            }
+            simdgroup_barrier(mem_flags::mem_none);
+            for (int i = 0; i < 2; i++) {
+                simdgroup_load(mb[i], lsmb + SG_MAT_SIZE * i);
+            }
+
+            lsma += BLOCK_SIZE_M / SG_MAT_ROW * SG_MAT_SIZE;
+            lsmb += BLOCK_SIZE_N / SG_MAT_ROW * SG_MAT_SIZE;
+
+            for (int i = 0; i < 8; i++){
+                simdgroup_multiply_accumulate(c_res[i], mb[i/4], ma[i%4], c_res[i]);
+            }
+        }
+    }
+
+    {
+        threadgroup_barrier(mem_flags::mem_threadgroup);
+        threadgroup float * temp_str = ((threadgroup float *)shared_memory) \
+                                      + 32 * (sgitg&1) + (16 * (sgitg>>1)) * BLOCK_SIZE_M;
+        for (int i = 0; i < 8; i++) {
+            simdgroup_store(c_res[i], temp_str + 8 * (i%4) + 8 * BLOCK_SIZE_M * (i/4), BLOCK_SIZE_M);
+        }
+
+        threadgroup_barrier(mem_flags::mem_threadgroup);
+
+        device float * C = dst + (BLOCK_SIZE_M * r0);
+        if (sgitg == 0) {
+            for (int j = tiitg; j < n_cols; j += BLOCK_SIZE_N) {
+                threadgroup const auto & jid = rowids[r1 * BLOCK_SIZE_N + j];
+                int joff =  jid[0] * ne0 + jid[1] * ne0ne1;
+                for (int i = 0; i < n_rows; i++) {
+                    *(C + i + joff) = *(temp_str + i + j * BLOCK_SIZE_M);
+                }
+            }
+        }
+    }
+}
+
+template<typename block_q, short nl, void (*dequantize_func)(device const block_q *, short, thread half4x4 &)>
+kernel void kernel_mul_mm_id(
+        device const   uchar * src0s,
+        device const   uchar * src1,
+        device         float * dst,
+        device const   uchar * ids,
+        constant     int64_t & nei0,
+        constant     int64_t & nei1,
+        constant    uint64_t & nbi1,
+        constant     int64_t & ne00,
+        constant     int64_t & ne02,
+        constant    uint64_t & nb01,
+        constant    uint64_t & nb02,
+        constant     int64_t & ne11,
+        constant     int64_t & ne12,
+        constant     int64_t & ne13,
+        constant    uint64_t & nb10,
+        constant    uint64_t & nb11,
+        constant    uint64_t & nb12,
+        constant     int64_t & ne0,
+        constant     int64_t & ne1,
+        constant    uint64_t & nb1,
+        threadgroup    uchar * shared_memory [[threadgroup(0)]],
+        uint3                  tgpig[[threadgroup_position_in_grid]],
+        uint                   tiitg[[thread_index_in_threadgroup]],
+        uint                   sgitg[[simdgroup_index_in_threadgroup]]) {
+
+    const int32_t i02 = tgpig.z;
+    tgpig.z = 0;
+
+    device const uchar * src0 = src0s + i02*nb02;
+
+    // row indices
+    threadgroup ushort2 * rowids = (threadgroup ushort2 *)(shared_memory + 8192);
+
+    // TODO: parallelize this loop
+    int64_t _ne1 = 0;
+    for (ushort ii1 = 0; ii1 < nei1; ii1++) {
+        for (ushort ii0 = 0; ii0 < nei0; ii0++) {
+            int32_t id = ((device int32_t *) (ids + ii1*nbi1))[ii0];
+            if (id == i02) {
+                //if (tiitg == 0) {
+                    rowids[_ne1] = ushort2(ii0, ii1);
+                //}
+                _ne1++;
+            }
+        }
+    }
+
+    threadgroup_barrier(mem_flags::mem_threadgroup);
+
+    kernel_mul_mm_id_impl<block_q, nl, dequantize_func>(
+        src0,
+        src1,
+        rowids,
+        dst,
+        ne00,
+        ne02,
+        nb01,
+        nb02,
+        ne11,
+        ne12,
+        nb10,
+        nb11,
+        nb12,
+        ne0,
+        _ne1,
+        ne0*ne1,
+        shared_memory,
+        tgpig,
+        tiitg,
+        sgitg);
+}
+
+#define QK_NL 16
+
+//
+// get rows
+//
+
+typedef decltype(kernel_get_rows_f<float>) get_rows_f_t;
+
+template [[host_name("kernel_get_rows_f32")]]  kernel get_rows_f_t kernel_get_rows_f<float>;
+template [[host_name("kernel_get_rows_f16")]]  kernel get_rows_f_t kernel_get_rows_f<half>;
+
+typedef decltype(kernel_get_rows_q<block_q4_0, 2, dequantize_q4_0>) get_rows_q_t;
+
+template [[host_name("kernel_get_rows_q4_0")]]    kernel get_rows_q_t kernel_get_rows_q<block_q4_0,    2, dequantize_q4_0>;
+template [[host_name("kernel_get_rows_q4_1")]]    kernel get_rows_q_t kernel_get_rows_q<block_q4_1,    2, dequantize_q4_1>;
+template [[host_name("kernel_get_rows_q5_0")]]    kernel get_rows_q_t kernel_get_rows_q<block_q5_0,    2, dequantize_q5_0>;
+template [[host_name("kernel_get_rows_q5_1")]]    kernel get_rows_q_t kernel_get_rows_q<block_q5_1,    2, dequantize_q5_1>;
+template [[host_name("kernel_get_rows_q8_0")]]    kernel get_rows_q_t kernel_get_rows_q<block_q8_0,    2, dequantize_q8_0>;
+template [[host_name("kernel_get_rows_q2_K")]]    kernel get_rows_q_t kernel_get_rows_q<block_q2_K,    QK_NL, dequantize_q2_K>;
+template [[host_name("kernel_get_rows_q3_K")]]    kernel get_rows_q_t kernel_get_rows_q<block_q3_K,    QK_NL, dequantize_q3_K>;
+template [[host_name("kernel_get_rows_q4_K")]]    kernel get_rows_q_t kernel_get_rows_q<block_q4_K,    QK_NL, dequantize_q4_K>;
+template [[host_name("kernel_get_rows_q5_K")]]    kernel get_rows_q_t kernel_get_rows_q<block_q5_K,    QK_NL, dequantize_q5_K>;
+template [[host_name("kernel_get_rows_q6_K")]]    kernel get_rows_q_t kernel_get_rows_q<block_q6_K,    QK_NL, dequantize_q6_K>;
+template [[host_name("kernel_get_rows_iq2_xxs")]] kernel get_rows_q_t kernel_get_rows_q<block_iq2_xxs, QK_NL, dequantize_iq2_xxs>;
+template [[host_name("kernel_get_rows_iq2_xs")]]  kernel get_rows_q_t kernel_get_rows_q<block_iq2_xs,  QK_NL, dequantize_iq2_xs>;
+template [[host_name("kernel_get_rows_iq3_xxs")]] kernel get_rows_q_t kernel_get_rows_q<block_iq3_xxs, QK_NL, dequantize_iq3_xxs>;
+template [[host_name("kernel_get_rows_iq3_s")]]   kernel get_rows_q_t kernel_get_rows_q<block_iq3_s,   QK_NL, dequantize_iq3_s>;
+template [[host_name("kernel_get_rows_iq2_s")]]   kernel get_rows_q_t kernel_get_rows_q<block_iq2_s,   QK_NL, dequantize_iq2_s>;
+template [[host_name("kernel_get_rows_iq1_s")]]   kernel get_rows_q_t kernel_get_rows_q<block_iq1_s,   QK_NL, dequantize_iq1_s>;
+template [[host_name("kernel_get_rows_iq1_m")]]   kernel get_rows_q_t kernel_get_rows_q<block_iq1_m,   QK_NL, dequantize_iq1_m>;
+template [[host_name("kernel_get_rows_iq4_nl")]]  kernel get_rows_q_t kernel_get_rows_q<block_iq4_nl,  2,     dequantize_iq4_nl>;
+template [[host_name("kernel_get_rows_iq4_xs")]]  kernel get_rows_q_t kernel_get_rows_q<block_iq4_xs,  QK_NL, dequantize_iq4_xs>;
+
+//
+// matrix-matrix multiplication
+//
+
+typedef decltype(kernel_mul_mm<half, half4x4, simdgroup_half8x8, float4x4, 1, dequantize_f32>) mat_mm_t;
+
+template [[host_name("kernel_mul_mm_f32_f32")]]     kernel mat_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   float4x4,      1,     dequantize_f32>;
+template [[host_name("kernel_mul_mm_f16_f32")]]     kernel mat_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   half4x4,       1,     dequantize_f16>;
+template [[host_name("kernel_mul_mm_q4_0_f32")]]    kernel mat_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   block_q4_0,    2,     dequantize_q4_0>;
+template [[host_name("kernel_mul_mm_q4_1_f32")]]    kernel mat_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   block_q4_1,    2,     dequantize_q4_1>;
+template [[host_name("kernel_mul_mm_q5_0_f32")]]    kernel mat_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   block_q5_0,    2,     dequantize_q5_0>;
+template [[host_name("kernel_mul_mm_q5_1_f32")]]    kernel mat_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   block_q5_1,    2,     dequantize_q5_1>;
+template [[host_name("kernel_mul_mm_q8_0_f32")]]    kernel mat_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   block_q8_0,    2,     dequantize_q8_0>;
+template [[host_name("kernel_mul_mm_q2_K_f32")]]    kernel mat_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   block_q2_K,    QK_NL, dequantize_q2_K>;
+template [[host_name("kernel_mul_mm_q3_K_f32")]]    kernel mat_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   block_q3_K,    QK_NL, dequantize_q3_K>;
+template [[host_name("kernel_mul_mm_q4_K_f32")]]    kernel mat_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   block_q4_K,    QK_NL, dequantize_q4_K>;
+template [[host_name("kernel_mul_mm_q5_K_f32")]]    kernel mat_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   block_q5_K,    QK_NL, dequantize_q5_K>;
+template [[host_name("kernel_mul_mm_q6_K_f32")]]    kernel mat_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   block_q6_K,    QK_NL, dequantize_q6_K>;
+template [[host_name("kernel_mul_mm_iq2_xxs_f32")]] kernel mat_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   block_iq2_xxs, QK_NL, dequantize_iq2_xxs>;
+template [[host_name("kernel_mul_mm_iq2_xs_f32")]]  kernel mat_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   block_iq2_xs,  QK_NL, dequantize_iq2_xs>;
+template [[host_name("kernel_mul_mm_iq3_xxs_f32")]] kernel mat_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   block_iq3_xxs, QK_NL, dequantize_iq3_xxs>;
+template [[host_name("kernel_mul_mm_iq3_s_f32")]]   kernel mat_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   block_iq3_s,   QK_NL, dequantize_iq3_s>;
+template [[host_name("kernel_mul_mm_iq2_s_f32")]]   kernel mat_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   block_iq2_s,   QK_NL, dequantize_iq2_s>;
+template [[host_name("kernel_mul_mm_iq1_s_f32")]]   kernel mat_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   block_iq1_s,   QK_NL, dequantize_iq1_s>;
+template [[host_name("kernel_mul_mm_iq1_m_f32")]]   kernel mat_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   block_iq1_m,   QK_NL, dequantize_iq1_m>;
+template [[host_name("kernel_mul_mm_iq4_nl_f32")]]  kernel mat_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   block_iq4_nl,  2,     dequantize_iq4_nl>;
+template [[host_name("kernel_mul_mm_iq4_xs_f32")]]  kernel mat_mm_t kernel_mul_mm<half,   half4x4,   simdgroup_half8x8,   block_iq4_xs,  QK_NL, dequantize_iq4_xs>;
+
+//
+// indirect matrix-matrix multiplication
+//
+
+typedef decltype(kernel_mul_mm_id<float4x4, 1, dequantize_f32>) mat_mm_id_t;
+
+template [[host_name("kernel_mul_mm_id_f32_f32")]]     kernel mat_mm_id_t kernel_mul_mm_id<float4x4,      1,     dequantize_f32>;
+template [[host_name("kernel_mul_mm_id_f16_f32")]]     kernel mat_mm_id_t kernel_mul_mm_id<half4x4,       1,     dequantize_f16>;
+template [[host_name("kernel_mul_mm_id_q4_0_f32")]]    kernel mat_mm_id_t kernel_mul_mm_id<block_q4_0,    2,     dequantize_q4_0>;
+template [[host_name("kernel_mul_mm_id_q4_1_f32")]]    kernel mat_mm_id_t kernel_mul_mm_id<block_q4_1,    2,     dequantize_q4_1>;
+template [[host_name("kernel_mul_mm_id_q5_0_f32")]]    kernel mat_mm_id_t kernel_mul_mm_id<block_q5_0,    2,     dequantize_q5_0>;
+template [[host_name("kernel_mul_mm_id_q5_1_f32")]]    kernel mat_mm_id_t kernel_mul_mm_id<block_q5_1,    2,     dequantize_q5_1>;
+template [[host_name("kernel_mul_mm_id_q8_0_f32")]]    kernel mat_mm_id_t kernel_mul_mm_id<block_q8_0,    2,     dequantize_q8_0>;
+template [[host_name("kernel_mul_mm_id_q2_K_f32")]]    kernel mat_mm_id_t kernel_mul_mm_id<block_q2_K,    QK_NL, dequantize_q2_K>;
+template [[host_name("kernel_mul_mm_id_q3_K_f32")]]    kernel mat_mm_id_t kernel_mul_mm_id<block_q3_K,    QK_NL, dequantize_q3_K>;
+template [[host_name("kernel_mul_mm_id_q4_K_f32")]]    kernel mat_mm_id_t kernel_mul_mm_id<block_q4_K,    QK_NL, dequantize_q4_K>;
+template [[host_name("kernel_mul_mm_id_q5_K_f32")]]    kernel mat_mm_id_t kernel_mul_mm_id<block_q5_K,    QK_NL, dequantize_q5_K>;
+template [[host_name("kernel_mul_mm_id_q6_K_f32")]]    kernel mat_mm_id_t kernel_mul_mm_id<block_q6_K,    QK_NL, dequantize_q6_K>;
+template [[host_name("kernel_mul_mm_id_iq2_xxs_f32")]] kernel mat_mm_id_t kernel_mul_mm_id<block_iq2_xxs, QK_NL, dequantize_iq2_xxs>;
+template [[host_name("kernel_mul_mm_id_iq2_xs_f32")]]  kernel mat_mm_id_t kernel_mul_mm_id<block_iq2_xs,  QK_NL, dequantize_iq2_xs>;
+template [[host_name("kernel_mul_mm_id_iq3_xxs_f32")]] kernel mat_mm_id_t kernel_mul_mm_id<block_iq3_xxs, QK_NL, dequantize_iq3_xxs>;
+template [[host_name("kernel_mul_mm_id_iq3_s_f32")]]   kernel mat_mm_id_t kernel_mul_mm_id<block_iq3_s,   QK_NL, dequantize_iq3_s>;
+template [[host_name("kernel_mul_mm_id_iq2_s_f32")]]   kernel mat_mm_id_t kernel_mul_mm_id<block_iq2_s,   QK_NL, dequantize_iq2_s>;
+template [[host_name("kernel_mul_mm_id_iq1_s_f32")]]   kernel mat_mm_id_t kernel_mul_mm_id<block_iq1_s,   QK_NL, dequantize_iq1_s>;
+template [[host_name("kernel_mul_mm_id_iq1_m_f32")]]   kernel mat_mm_id_t kernel_mul_mm_id<block_iq1_m,   QK_NL, dequantize_iq1_m>;
+template [[host_name("kernel_mul_mm_id_iq4_nl_f32")]]  kernel mat_mm_id_t kernel_mul_mm_id<block_iq4_nl,  2,     dequantize_iq4_nl>;
+template [[host_name("kernel_mul_mm_id_iq4_xs_f32")]]  kernel mat_mm_id_t kernel_mul_mm_id<block_iq4_xs,  QK_NL, dequantize_iq4_xs>;
+
+//
+// matrix-vector multiplication
+//
+
+typedef void (kernel_mul_mv_impl_t)(
+        device const  char * src0,
+        device const  char * src1,
+        device       float * dst,
+                   int64_t   ne00,
+                   int64_t   ne01,
+                   int64_t   ne02,
+                  uint64_t   nb00,
+                  uint64_t   nb01,
+                  uint64_t   nb02,
+                   int64_t   ne10,
+                   int64_t   ne11,
+                   int64_t   ne12,
+                  uint64_t   nb10,
+                  uint64_t   nb11,
+                  uint64_t   nb12,
+                   int64_t   ne0,
+                   int64_t   ne1,
+                   uint      r2,
+                   uint      r3,
+                   uint3     tgpig,
+                   uint      tiisg);
+
+typedef void (kernel_mul_mv2_impl_t)(
+        device const  void * src0,
+        device const float * src1,
+        device       float * dst,
+                   int64_t   ne00,
+                   int64_t   ne01,
+                   int64_t   ne02,
+                   int64_t   ne10,
+                   int64_t   ne12,
+                   int64_t   ne0,
+                   int64_t   ne1,
+                   uint      r2,
+                   uint      r3,
+        threadgroup int8_t * shared_values,
+                   uint3     tgpig,
+                   uint      tiisg,
+                   uint      sgitg);
+
+template<kernel_mul_mv_impl_t impl_fn>
+void mmv_fn(
+        device const    char * src0,
+        device const    char * src1,
+        device         float * dst,
+                     int64_t   ne00,
+                     int64_t   ne01,
+                     int64_t   ne02,
+                    uint64_t   nb00,
+                    uint64_t   nb01,
+                    uint64_t   nb02,
+                     int64_t   ne10,
+                     int64_t   ne11,
+                     int64_t   ne12,
+                     int64_t   ne13,
+                    uint64_t   nb10,
+                    uint64_t   nb11,
+                    uint64_t   nb12,
+                     int64_t   ne0,
+                     int64_t   ne1,
+                    uint64_t   nb1,
+                        uint   r2,
+                        uint   r3,
+        threadgroup int8_t   * shared_values,
+        uint3                  tgpig,
+        uint                   tiitg,
+        uint                   tiisg,
+        uint                   sgitg) {
+    impl_fn(src0,src1,dst,ne00,ne01,ne02,nb00,nb01,nb02,ne10,ne11,ne12,nb10,nb11,nb12,ne0,ne1,r2,r3,tgpig,tiisg);
+}
+
+template<kernel_mul_mv2_impl_t impl_fn>
+void mmv_fn(
+        device const    char * src0,
+        device const    char * src1,
+        device         float * dst,
+                     int64_t   ne00,
+                     int64_t   ne01,
+                     int64_t   ne02,
+                    uint64_t   nb00,
+                    uint64_t   nb01,
+                    uint64_t   nb02,
+                     int64_t   ne10,
+                     int64_t   ne11,
+                     int64_t   ne12,
+                     int64_t   ne13,
+                    uint64_t   nb10,
+                    uint64_t   nb11,
+                    uint64_t   nb12,
+                     int64_t   ne0,
+                     int64_t   ne1,
+                    uint64_t   nb1,
+                        uint   r2,
+                        uint   r3,
+        threadgroup int8_t   * shared_values,
+        uint3                  tgpig,
+        uint                   tiitg,
+        uint                   tiisg,
+        uint                   sgitg) {
+    impl_fn(src0,(const device float *)src1,dst,ne00,ne01,ne02,ne10,ne12,ne0,ne1,r2,r3,shared_values,tgpig,tiisg,sgitg);
+}
+
+typedef decltype(mmv_fn<kernel_mul_mv_impl<half, half4, half, half4>>) mul_mv_impl_fn_t;
+
+template<mul_mv_impl_fn_t impl_fn>
+kernel void kernel_mul_mv_id(
+        device const    char * src0s,
+        device const    char * src1,
+        device         float * dst,
+        device const    char * ids,
+        constant     int64_t & nei0,
+        constant     int64_t & nei1,
+        constant    uint64_t & nbi1,
+        constant     int64_t & ne00,
+        constant     int64_t & ne01,
+        constant     int64_t & ne02,
+        constant    uint64_t & nb00,
+        constant    uint64_t & nb01,
+        constant    uint64_t & nb02,
+        constant     int64_t & ne10,
+        constant     int64_t & ne11,
+        constant     int64_t & ne12,
+        constant     int64_t & ne13,
+        constant    uint64_t & nb10,
+        constant    uint64_t & nb11,
+        constant    uint64_t & nb12,
+        constant     int64_t & ne0,
+        constant     int64_t & ne1,
+        constant    uint64_t & nb1,
+        threadgroup int8_t   * shared_values [[threadgroup(0)]],
+        uint3                  tgpig[[threadgroup_position_in_grid]],
+        uint                   tiitg[[thread_index_in_threadgroup]],
+        uint                   tiisg[[thread_index_in_simdgroup]],
+        uint                   sgitg[[simdgroup_index_in_threadgroup]]) {
+    const int iid1 = tgpig.z/nei0;
+    const int idx = tgpig.z%nei0;
+
+    tgpig.z = 0;
+
+    const int32_t i02 = ((device const int32_t *) (ids + iid1*nbi1))[idx];
+
+    const int64_t i11 = idx % ne11;
+    const int64_t i12 = iid1;
+
+    const int64_t i1 = idx;
+    const int64_t i2 = i12;
+
+    device const char * src0_cur = src0s + i02*nb02;
+    device const char * src1_cur = src1 + i11*nb11 + i12*nb12;
+    device      float * dst_cur  = dst + i1*ne0 + i2*ne1*ne0;
+
+    impl_fn(
+        /* src0 */ src0_cur,
+        /* src1 */ src1_cur,
+        /* dst  */ dst_cur,
+        /* ne00 */ ne00,
+        /* ne01 */ ne01,
+        /* ne02 */ 1,//ne02,
+        /* nb00 */ nb00,
+        /* nb01 */ nb01,
+        /* nb02 */ nb02,
+        /* ne10 */ ne10,
+        /* ne11 */ 1,//ne11,
+        /* ne12 */ 1,//ne12,
+        /* ne13 */ 1,//ne13,
+        /* nb10 */ nb10,
+        /* nb11 */ nb11,
+        /* nb12 */ nb12,
+        /* ne0  */ ne0,
+        /* ne1  */ 1,//ne1,
+        /* nb1  */ nb1,
+        /* r2   */ 1,
+        /* r3   */ 1,
+        shared_values,
+        tgpig,
+        tiitg,
+        tiisg,
+        sgitg);
+}
+
+typedef decltype(kernel_mul_mv_id<mmv_fn<kernel_mul_mv_impl<float, float4, float, float4>>>) kernel_mul_mv_id_t;
+
+template [[host_name("kernel_mul_mv_id_f32_f32")]]     kernel kernel_mul_mv_id_t kernel_mul_mv_id<mmv_fn<kernel_mul_mv_impl<float, float4, float, float4>>>;
+template [[host_name("kernel_mul_mv_id_f16_f32")]]     kernel kernel_mul_mv_id_t kernel_mul_mv_id<mmv_fn<kernel_mul_mv_impl<half, half4, float, float4>>>;
+template [[host_name("kernel_mul_mv_id_q8_0_f32")]]    kernel kernel_mul_mv_id_t kernel_mul_mv_id<mmv_fn<kernel_mul_mv_q8_0_f32_impl>>;
+template [[host_name("kernel_mul_mv_id_q4_0_f32")]]    kernel kernel_mul_mv_id_t kernel_mul_mv_id<mmv_fn<mul_vec_q_n_f32_impl<block_q4_0, N_DST, N_SIMDGROUP, N_SIMDWIDTH>>>;
+template [[host_name("kernel_mul_mv_id_q4_1_f32")]]    kernel kernel_mul_mv_id_t kernel_mul_mv_id<mmv_fn<mul_vec_q_n_f32_impl<block_q4_1, N_DST, N_SIMDGROUP, N_SIMDWIDTH>>>;
+template [[host_name("kernel_mul_mv_id_q5_0_f32")]]    kernel kernel_mul_mv_id_t kernel_mul_mv_id<mmv_fn<mul_vec_q_n_f32_impl<block_q5_0, N_DST, N_SIMDGROUP, N_SIMDWIDTH>>>;
+template [[host_name("kernel_mul_mv_id_q5_1_f32")]]    kernel kernel_mul_mv_id_t kernel_mul_mv_id<mmv_fn<mul_vec_q_n_f32_impl<block_q5_1, N_DST, N_SIMDGROUP, N_SIMDWIDTH>>>;
+template [[host_name("kernel_mul_mv_id_q2_K_f32")]]    kernel kernel_mul_mv_id_t kernel_mul_mv_id<mmv_fn<kernel_mul_mv_q2_K_f32_impl>>;
+template [[host_name("kernel_mul_mv_id_q3_K_f32")]]    kernel kernel_mul_mv_id_t kernel_mul_mv_id<mmv_fn<kernel_mul_mv_q3_K_f32_impl>>;
+template [[host_name("kernel_mul_mv_id_q4_K_f32")]]    kernel kernel_mul_mv_id_t kernel_mul_mv_id<mmv_fn<kernel_mul_mv_q4_K_f32_impl>>;
+template [[host_name("kernel_mul_mv_id_q5_K_f32")]]    kernel kernel_mul_mv_id_t kernel_mul_mv_id<mmv_fn<kernel_mul_mv_q5_K_f32_impl>>;
+template [[host_name("kernel_mul_mv_id_q6_K_f32")]]    kernel kernel_mul_mv_id_t kernel_mul_mv_id<mmv_fn<kernel_mul_mv_q6_K_f32_impl>>;
+template [[host_name("kernel_mul_mv_id_iq1_s_f32")]]   kernel kernel_mul_mv_id_t kernel_mul_mv_id<mmv_fn<kernel_mul_mv_iq1_s_f32_impl>>;
+template [[host_name("kernel_mul_mv_id_iq1_m_f32")]]   kernel kernel_mul_mv_id_t kernel_mul_mv_id<mmv_fn<kernel_mul_mv_iq1_m_f32_impl>>;
+template [[host_name("kernel_mul_mv_id_iq2_xxs_f32")]] kernel kernel_mul_mv_id_t kernel_mul_mv_id<mmv_fn<kernel_mul_mv_iq2_xxs_f32_impl>>;
+template [[host_name("kernel_mul_mv_id_iq2_xs_f32")]]  kernel kernel_mul_mv_id_t kernel_mul_mv_id<mmv_fn<kernel_mul_mv_iq2_xs_f32_impl>>;
+template [[host_name("kernel_mul_mv_id_iq3_xxs_f32")]] kernel kernel_mul_mv_id_t kernel_mul_mv_id<mmv_fn<kernel_mul_mv_iq3_xxs_f32_impl>>;
+template [[host_name("kernel_mul_mv_id_iq3_s_f32")]]   kernel kernel_mul_mv_id_t kernel_mul_mv_id<mmv_fn<kernel_mul_mv_iq3_s_f32_impl>>;
+template [[host_name("kernel_mul_mv_id_iq2_s_f32")]]   kernel kernel_mul_mv_id_t kernel_mul_mv_id<mmv_fn<kernel_mul_mv_iq2_s_f32_impl>>;
+template [[host_name("kernel_mul_mv_id_iq4_nl_f32")]]  kernel kernel_mul_mv_id_t kernel_mul_mv_id<mmv_fn<kernel_mul_mv_iq4_nl_f32_impl>>;
+template [[host_name("kernel_mul_mv_id_iq4_xs_f32")]]  kernel kernel_mul_mv_id_t kernel_mul_mv_id<mmv_fn<kernel_mul_mv_iq4_xs_f32_impl>>;
diff --git a/src/whisper_cpp/ggml/src/ggml-quants.c b/src/whisper_cpp/ggml/src/ggml-quants.c
new file mode 100644
index 00000000..7aa6dce8
--- /dev/null
+++ b/src/whisper_cpp/ggml/src/ggml-quants.c
@@ -0,0 +1,15752 @@
+#define GGML_COMMON_IMPL_C
+#include "ggml-common.h"
+
+#include "ggml-quants.h"
+#include "ggml-impl.h"
+#include "ggml-cpu-impl.h"
+
+
+#include <math.h>
+#include <string.h>
+#include <assert.h>
+#include <float.h>
+#include <stdlib.h> // for qsort
+#include <stdio.h>  // for GGML_ASSERT
+
+#define GROUP_MAX_EPS 1e-15f
+#define GROUP_MAX_EPS_IQ3_XXS 1e-8f
+#define GROUP_MAX_EPS_IQ2_S 1e-8f
+#define GROUP_MAX_EPS_IQ1_M 1e-7f
+#define GROUP_MAX_EPS_IQ1_S 1e-12f
+
+#if defined(_MSC_VER)
+// disable "possible loss of data" to avoid warnings for hundreds of casts
+// we should just be careful :)
+#pragma warning(disable: 4244 4267)
+#endif
+
+#define UNUSED GGML_UNUSED
+
+// some compilers don't provide _mm256_set_m128i, e.g. gcc 7
+#define MM256_SET_M128I(a, b) _mm256_insertf128_si256(_mm256_castsi128_si256(b), (a), 1)
+
+#if defined(__AVX__) || defined(__AVX2__) || defined(__AVX512F__) || defined(__SSSE3__)
+// multiply int8_t, add results pairwise twice
+static inline __m128i mul_sum_i8_pairs(const __m128i x, const __m128i y) {
+    // Get absolute values of x vectors
+    const __m128i ax = _mm_sign_epi8(x, x);
+    // Sign the values of the y vectors
+    const __m128i sy = _mm_sign_epi8(y, x);
+    // Perform multiplication and create 16-bit values
+    const __m128i dot = _mm_maddubs_epi16(ax, sy);
+    const __m128i ones = _mm_set1_epi16(1);
+    return _mm_madd_epi16(ones, dot);
+}
+
+#if __AVX__ || __AVX2__ || __AVX512F__
+// horizontally add 8 floats
+static inline float hsum_float_8(const __m256 x) {
+    __m128 res = _mm256_extractf128_ps(x, 1);
+    res = _mm_add_ps(res, _mm256_castps256_ps128(x));
+    res = _mm_add_ps(res, _mm_movehl_ps(res, res));
+    res = _mm_add_ss(res, _mm_movehdup_ps(res));
+    return _mm_cvtss_f32(res);
+}
+
+// horizontally add 8 int32_t
+static inline int hsum_i32_8(const __m256i a) {
+    const __m128i sum128 = _mm_add_epi32(_mm256_castsi256_si128(a), _mm256_extractf128_si256(a, 1));
+    const __m128i hi64 = _mm_unpackhi_epi64(sum128, sum128);
+    const __m128i sum64 = _mm_add_epi32(hi64, sum128);
+    const __m128i hi32  = _mm_shuffle_epi32(sum64, _MM_SHUFFLE(2, 3, 0, 1));
+    return _mm_cvtsi128_si32(_mm_add_epi32(sum64, hi32));
+}
+
+// horizontally add 4 int32_t
+static inline int hsum_i32_4(const __m128i a) {
+    const __m128i hi64 = _mm_unpackhi_epi64(a, a);
+    const __m128i sum64 = _mm_add_epi32(hi64, a);
+    const __m128i hi32  = _mm_shuffle_epi32(sum64, _MM_SHUFFLE(2, 3, 0, 1));
+    return _mm_cvtsi128_si32(_mm_add_epi32(sum64, hi32));
+}
+
+#if defined(__AVX2__) || defined(__AVX512F__)
+// spread 32 bits to 32 bytes { 0x00, 0xFF }
+static inline __m256i bytes_from_bits_32(const uint8_t * x) {
+    uint32_t x32;
+    memcpy(&x32, x, sizeof(uint32_t));
+    const __m256i shuf_mask = _mm256_set_epi64x(
+            0x0303030303030303, 0x0202020202020202,
+            0x0101010101010101, 0x0000000000000000);
+    __m256i bytes = _mm256_shuffle_epi8(_mm256_set1_epi32(x32), shuf_mask);
+    const __m256i bit_mask = _mm256_set1_epi64x(0x7fbfdfeff7fbfdfe);
+    bytes = _mm256_or_si256(bytes, bit_mask);
+    return _mm256_cmpeq_epi8(bytes, _mm256_set1_epi64x(-1));
+}
+
+// Unpack 32 4-bit fields into 32 bytes
+// The output vector contains 32 bytes, each one in [ 0 .. 15 ] interval
+static inline __m256i bytes_from_nibbles_32(const uint8_t * rsi)
+{
+    const __m128i tmp = _mm_loadu_si128((const __m128i *)rsi);
+    const __m256i bytes = MM256_SET_M128I(_mm_srli_epi16(tmp, 4), tmp);
+    const __m256i lowMask = _mm256_set1_epi8( 0xF );
+    return _mm256_and_si256(lowMask, bytes);
+}
+
+// add int16_t pairwise and return as float vector
+static inline __m256 sum_i16_pairs_float(const __m256i x) {
+    const __m256i ones = _mm256_set1_epi16(1);
+    const __m256i summed_pairs = _mm256_madd_epi16(ones, x);
+    return _mm256_cvtepi32_ps(summed_pairs);
+}
+
+static inline __m256 mul_sum_us8_pairs_float(const __m256i ax, const __m256i sy) {
+#if defined(__AVXVNNI__) || (defined(__AVX512VNNI__) && defined(__AVX512VL__))
+    const __m256i zero = _mm256_setzero_si256();
+    const __m256i summed_pairs = _mm256_dpbusd_epi32(zero, ax, sy);
+    return _mm256_cvtepi32_ps(summed_pairs);
+#else
+    // Perform multiplication and create 16-bit values
+    const __m256i dot = _mm256_maddubs_epi16(ax, sy);
+    return sum_i16_pairs_float(dot);
+#endif
+}
+
+// multiply int8_t, add results pairwise twice and return as float vector
+static inline __m256 mul_sum_i8_pairs_float(const __m256i x, const __m256i y) {
+#if __AVXVNNIINT8__
+    const __m256i zero = _mm256_setzero_si256();
+    const __m256i summed_pairs = _mm256_dpbssd_epi32(zero, x, y);
+    return _mm256_cvtepi32_ps(summed_pairs);
+#else
+    // Get absolute values of x vectors
+    const __m256i ax = _mm256_sign_epi8(x, x);
+    // Sign the values of the y vectors
+    const __m256i sy = _mm256_sign_epi8(y, x);
+    return mul_sum_us8_pairs_float(ax, sy);
+#endif
+}
+
+static inline __m128i packNibbles( __m256i bytes )
+{
+    // Move bits within 16-bit lanes from 0000_abcd_0000_efgh into 0000_0000_abcd_efgh
+#if __AVX512F__
+    const __m256i bytes_srli_4 = _mm256_srli_epi16(bytes, 4);   // 0000_0000_abcd_0000
+    bytes = _mm256_or_si256(bytes, bytes_srli_4);               // 0000_abcd_abcd_efgh
+    return _mm256_cvtepi16_epi8(bytes);                         // abcd_efgh
+#else
+    const __m256i lowByte = _mm256_set1_epi16( 0xFF );
+    __m256i high = _mm256_andnot_si256( lowByte, bytes );
+    __m256i low = _mm256_and_si256( lowByte, bytes );
+    high = _mm256_srli_epi16( high, 4 );
+    bytes = _mm256_or_si256( low, high );
+
+    // Compress uint16_t lanes into bytes
+    __m128i r0 = _mm256_castsi256_si128( bytes );
+    __m128i r1 = _mm256_extracti128_si256( bytes, 1 );
+    return _mm_packus_epi16( r0, r1 );
+#endif
+}
+#elif defined(__AVX__)
+// spread 32 bits to 32 bytes { 0x00, 0xFF }
+static inline __m256i bytes_from_bits_32(const uint8_t * x) {
+    uint32_t x32;
+    memcpy(&x32, x, sizeof(uint32_t));
+    const __m128i shuf_maskl = _mm_set_epi64x(0x0101010101010101, 0x0000000000000000);
+    const __m128i shuf_maskh = _mm_set_epi64x(0x0303030303030303, 0x0202020202020202);
+    __m128i bytesl = _mm_shuffle_epi8(_mm_set1_epi32(x32), shuf_maskl);
+    __m128i bytesh = _mm_shuffle_epi8(_mm_set1_epi32(x32), shuf_maskh);
+    const __m128i bit_mask = _mm_set1_epi64x(0x7fbfdfeff7fbfdfe);
+    bytesl = _mm_or_si128(bytesl, bit_mask);
+    bytesh = _mm_or_si128(bytesh, bit_mask);
+    bytesl = _mm_cmpeq_epi8(bytesl, _mm_set1_epi64x(-1));
+    bytesh = _mm_cmpeq_epi8(bytesh, _mm_set1_epi64x(-1));
+    return MM256_SET_M128I(bytesh, bytesl);
+}
+
+// Unpack 32 4-bit fields into 32 bytes
+// The output vector contains 32 bytes, each one in [ 0 .. 15 ] interval
+static inline __m256i bytes_from_nibbles_32(const uint8_t * rsi)
+{
+    // Load 16 bytes from memory
+    __m128i tmpl = _mm_loadu_si128((const __m128i *)rsi);
+    __m128i tmph = _mm_srli_epi16(tmpl, 4);
+    const __m128i lowMask = _mm_set1_epi8(0xF);
+    tmpl = _mm_and_si128(lowMask, tmpl);
+    tmph = _mm_and_si128(lowMask, tmph);
+    return MM256_SET_M128I(tmph, tmpl);
+}
+
+// add int16_t pairwise and return as float vector
+static inline __m256 sum_i16_pairs_float(const __m128i xh, const __m128i xl) {
+    const __m128i ones = _mm_set1_epi16(1);
+    const __m128i summed_pairsl = _mm_madd_epi16(ones, xl);
+    const __m128i summed_pairsh = _mm_madd_epi16(ones, xh);
+    const __m256i summed_pairs = MM256_SET_M128I(summed_pairsh, summed_pairsl);
+    return _mm256_cvtepi32_ps(summed_pairs);
+}
+
+static inline __m256 mul_sum_us8_pairs_float(const __m256i ax, const __m256i sy) {
+    const __m128i axl = _mm256_castsi256_si128(ax);
+    const __m128i axh = _mm256_extractf128_si256(ax, 1);
+    const __m128i syl = _mm256_castsi256_si128(sy);
+    const __m128i syh = _mm256_extractf128_si256(sy, 1);
+    // Perform multiplication and create 16-bit values
+    const __m128i dotl = _mm_maddubs_epi16(axl, syl);
+    const __m128i doth = _mm_maddubs_epi16(axh, syh);
+    return sum_i16_pairs_float(doth, dotl);
+}
+
+// multiply int8_t, add results pairwise twice and return as float vector
+static inline __m256 mul_sum_i8_pairs_float(const __m256i x, const __m256i y) {
+    const __m128i xl = _mm256_castsi256_si128(x);
+    const __m128i xh = _mm256_extractf128_si256(x, 1);
+    const __m128i yl = _mm256_castsi256_si128(y);
+    const __m128i yh = _mm256_extractf128_si256(y, 1);
+    // Get absolute values of x vectors
+    const __m128i axl = _mm_sign_epi8(xl, xl);
+    const __m128i axh = _mm_sign_epi8(xh, xh);
+    // Sign the values of the y vectors
+    const __m128i syl = _mm_sign_epi8(yl, xl);
+    const __m128i syh = _mm_sign_epi8(yh, xh);
+    // Perform multiplication and create 16-bit values
+    const __m128i dotl = _mm_maddubs_epi16(axl, syl);
+    const __m128i doth = _mm_maddubs_epi16(axh, syh);
+    return sum_i16_pairs_float(doth, dotl);
+}
+
+static inline __m128i packNibbles( __m128i bytes1, __m128i bytes2 )
+{
+    // Move bits within 16-bit lanes from 0000_abcd_0000_efgh into 0000_0000_abcd_efgh
+    const __m128i lowByte = _mm_set1_epi16( 0xFF );
+    __m128i high = _mm_andnot_si128( lowByte, bytes1 );
+    __m128i low = _mm_and_si128( lowByte, bytes1 );
+    high = _mm_srli_epi16( high, 4 );
+    bytes1 = _mm_or_si128( low, high );
+    high = _mm_andnot_si128( lowByte, bytes2 );
+    low = _mm_and_si128( lowByte, bytes2 );
+    high = _mm_srli_epi16( high, 4 );
+    bytes2 = _mm_or_si128( low, high );
+
+    return _mm_packus_epi16( bytes1, bytes2);
+}
+
+static inline __m128i mul_add_epi8_sse(const __m128i x, const __m128i y) {
+    const __m128i ax = _mm_sign_epi8(x, x);
+    const __m128i sy = _mm_sign_epi8(y, x);
+    return _mm_maddubs_epi16(ax, sy);
+}
+#endif
+#elif defined(__SSSE3__)
+// horizontally add 4x4 floats
+static inline float hsum_float_4x4(const __m128 a, const __m128 b, const __m128 c, const __m128 d) {
+    __m128 res_0 =_mm_hadd_ps(a, b);
+    __m128 res_1 =_mm_hadd_ps(c, d);
+    __m128 res =_mm_hadd_ps(res_0, res_1);
+    res =_mm_hadd_ps(res, res);
+    res =_mm_hadd_ps(res, res);
+
+    return _mm_cvtss_f32(res);
+}
+#endif // __AVX__ || __AVX2__ || __AVX512F__
+#endif // defined(__AVX__) || defined(__AVX2__) || defined(__AVX512F__) || defined(__SSSE3__)
+
+#if defined(__ARM_NEON) || defined(__wasm_simd128__) || defined(__POWER9_VECTOR__)
+#define B1(c,s,n)  0x ## n ## c ,  0x ## n ## s
+#define B2(c,s,n) B1(c,s,n ## c), B1(c,s,n ## s)
+#define B3(c,s,n) B2(c,s,n ## c), B2(c,s,n ## s)
+#define B4(c,s,n) B3(c,s,n ## c), B3(c,s,n ## s)
+#define B5(c,s,n) B4(c,s,n ## c), B4(c,s,n ## s)
+#define B6(c,s,n) B5(c,s,n ## c), B5(c,s,n ## s)
+#define B7(c,s,n) B6(c,s,n ## c), B6(c,s,n ## s)
+#define B8(c,s  ) B7(c,s,     c), B7(c,s,     s)
+
+// precomputed tables for expanding 8bits to 8 bytes:
+static const uint64_t table_b2b_0[1 << 8] = { B8(00, 10) }; // ( b) << 4
+static const uint64_t table_b2b_1[1 << 8] = { B8(10, 00) }; // (!b) << 4
+#endif
+
+#if defined(__loongarch_asx)
+
+#ifdef __clang__
+#define VREGS_PREFIX "$vr"
+#define XREGS_PREFIX "$xr"
+#else // GCC
+#define VREGS_PREFIX "$f"
+#define XREGS_PREFIX "$f"
+#endif
+#define __ALL_REGS "0,1,2,3,4,5,6,7,8,9,10,11,12,13,14,15,16,17,18,19,20,21,22,23,24,25,26,27,28,29,30,31"
+// Convert __m128i to __m256i
+static inline __m256i ____m256i(__m128i in) {
+    __m256i out = __lasx_xvldi(0);
+    __asm__ volatile (
+        ".irp i," __ALL_REGS                "\n\t"
+        " .ifc %[out], " XREGS_PREFIX"\\i    \n\t"
+        "  .irp j," __ALL_REGS              "\n\t"
+        "   .ifc %[in], " VREGS_PREFIX "\\j  \n\t"
+        "    xvpermi.q $xr\\i, $xr\\j, 0x20  \n\t"
+        "   .endif                           \n\t"
+        "  .endr                             \n\t"
+        " .endif                             \n\t"
+        ".endr                               \n\t"
+        : [out] "+f" (out) : [in] "f" (in)
+    );
+    return out;
+}
+// Convert two __m128i to __m256i
+static inline __m256i lasx_set_q(__m128i inhi, __m128i inlo) {
+    __m256i out;
+    __asm__ volatile (
+        ".irp i," __ALL_REGS                "\n\t"
+        " .ifc %[hi], " VREGS_PREFIX "\\i    \n\t"
+        "  .irp j," __ALL_REGS              "\n\t"
+        "   .ifc %[lo], " VREGS_PREFIX "\\j  \n\t"
+        "    xvpermi.q $xr\\i, $xr\\j, 0x20  \n\t"
+        "   .endif                           \n\t"
+        "  .endr                             \n\t"
+        " .endif                             \n\t"
+        ".endr                               \n\t"
+        ".ifnc %[out], %[hi]                 \n\t"
+        ".irp i," __ALL_REGS                "\n\t"
+        " .ifc %[out], " XREGS_PREFIX "\\i   \n\t"
+        "  .irp j," __ALL_REGS              "\n\t"
+        "   .ifc %[hi], " VREGS_PREFIX "\\j  \n\t"
+        "    xvori.b $xr\\i, $xr\\j, 0       \n\t"
+        "   .endif                           \n\t"
+        "  .endr                             \n\t"
+        " .endif                             \n\t"
+        ".endr                               \n\t"
+        ".endif                              \n\t"
+        : [out] "=f" (out), [hi] "+f" (inhi)
+        : [lo] "f" (inlo)
+    );
+    return out;
+}
+// Convert __m256i low part to __m128i
+static inline __m128i lasx_extracti128_lo(__m256i in) {
+    __m128i out;
+    __asm__ volatile (
+        ".ifnc %[out], %[in]                 \n\t"
+        ".irp i," __ALL_REGS                "\n\t"
+        " .ifc %[out], " VREGS_PREFIX "\\i   \n\t"
+        "  .irp j," __ALL_REGS              "\n\t"
+        "   .ifc %[in], " XREGS_PREFIX "\\j  \n\t"
+        "    vori.b $vr\\i, $vr\\j, 0        \n\t"
+        "   .endif                           \n\t"
+        "  .endr                             \n\t"
+        " .endif                             \n\t"
+        ".endr                               \n\t"
+        ".endif                              \n\t"
+        : [out] "=f" (out) : [in] "f" (in)
+    );
+    return out;
+}
+// Convert __m256i high part to __m128i
+static inline __m128i lasx_extracti128_hi(__m256i in) {
+    __m128i out;
+    __asm__ volatile (
+        ".irp i," __ALL_REGS                "\n\t"
+        " .ifc %[out], " VREGS_PREFIX "\\i   \n\t"
+        "  .irp j," __ALL_REGS              "\n\t"
+        "   .ifc %[in], " XREGS_PREFIX "\\j  \n\t"
+        "    xvpermi.q $xr\\i, $xr\\j, 0x11  \n\t"
+        "   .endif                           \n\t"
+        "  .endr                             \n\t"
+        " .endif                             \n\t"
+        ".endr                               \n\t"
+        : [out] "=f" (out) : [in] "f" (in)
+    );
+    return out;
+}
+
+static __m256i lasx_set_w(int e7, int e6, int e5, int e4, int e3, int e2, int e1, int e0) {
+    v8i32 __ret = {e0, e1, e2, e3, e4, e5, e6, e7};
+    return (__m256i)__ret;
+}
+
+static __m128i lsx_set_w(int32_t a, int32_t b, int32_t c, int32_t d) {
+    v4i32 __ret = {d, c, b, a};
+    return (__m128i)__ret;
+}
+
+static __m256i lasx_set_d(int64_t a, int64_t b, int64_t c, int64_t d) {
+    v4i64 __ret = {d, c, b, a};
+    return (__m256i)__ret;
+}
+
+static __m256i lasx_insertf128( __m128i x, __m128i y) {
+    return lasx_set_q(x, y);
+}
+
+static __m128i lsx_shuffle_b(__m128i a, __m128i b) {
+    __m128i mask_f, zero, tmp0, tmp2, mask;
+    int f = 0x8f;
+    mask_f = __lsx_vreplgr2vr_b(f);
+    zero = __lsx_vldi(0);
+    tmp0 = __lsx_vand_v(b, mask_f); // get mask with low 4 bit and sign bits
+    tmp0 = __lsx_vori_b(tmp0, 0x10); // make each mask or  with 0x10 prepare for positive
+    mask = __lsx_vsle_b(zero, tmp0); // if mask >= 0, set mask
+    tmp2 = __lsx_vand_v(tmp0, mask); // maskout the in2 < ones
+    return __lsx_vshuf_b(a, zero, tmp2);
+}
+
+static __m256i lasx_shuffle_b(__m256i a, __m256i b) {
+    __m256i mask_f, zero, tmp0, tmp2, mask;
+    int f = 0x8f;
+    mask_f = __lasx_xvreplgr2vr_b(f);
+    zero = __lasx_xvldi(0);
+    tmp0 = __lasx_xvand_v(b, mask_f); // get mask with low 4 bit and sign bits
+    tmp0 = __lasx_xvori_b(tmp0, 0x10); // make each mask or  with 0x10 prepare for positive
+    mask = __lasx_xvsle_b(zero, tmp0); // if mask >= 0, set mask
+    tmp2 = __lasx_xvand_v(tmp0, mask); // maskout the in2 < ones
+    return __lasx_xvshuf_b(a, zero, tmp2);
+}
+
+static __m256i lasx_extu8_16(__m128i a) {
+    __m128i zero = __lsx_vldi(0);
+    __m128i vlo = __lsx_vilvl_b(zero, a);
+    __m128i vhi = __lsx_vilvh_b(zero, a);
+    return lasx_set_q(vhi, vlo);
+}
+
+static __m256i lasx_ext8_16(__m128i a) {
+     __m128i sign = __lsx_vslti_b(a, 0);
+     __m128i vlo = __lsx_vilvl_b(sign, a);
+     __m128i vhi = __lsx_vilvh_b(sign, a);
+     return lasx_set_q(vhi, vlo);
+}
+
+static __m256i lasx_ext16_32(__m128i a) {
+    __m256i tmp1;
+    tmp1 = __lasx_xvinsgr2vr_w(tmp1, __lsx_vpickve2gr_h(a, 0), 0);
+    tmp1 = __lasx_xvinsgr2vr_w(tmp1, __lsx_vpickve2gr_h(a, 1), 1);
+    tmp1 = __lasx_xvinsgr2vr_w(tmp1, __lsx_vpickve2gr_h(a, 2), 2);
+    tmp1 = __lasx_xvinsgr2vr_w(tmp1, __lsx_vpickve2gr_h(a, 3), 3);
+    tmp1 = __lasx_xvinsgr2vr_w(tmp1, __lsx_vpickve2gr_h(a, 4), 4);
+    tmp1 = __lasx_xvinsgr2vr_w(tmp1, __lsx_vpickve2gr_h(a, 5), 5);
+    tmp1 = __lasx_xvinsgr2vr_w(tmp1, __lsx_vpickve2gr_h(a, 6), 6);
+    tmp1 = __lasx_xvinsgr2vr_w(tmp1, __lsx_vpickve2gr_h(a, 7), 7);
+    return tmp1;
+}
+
+static __m128i lasx_extracti128( __m256i a, int pos) {
+    __m128i ret;
+    if( pos == 0)
+    {
+       ret = lasx_extracti128_lo(a);
+    } else {
+       ret = lasx_extracti128_hi(a);
+    }
+    return ret;
+}
+
+static __m128 lasx_extractf128( __m256 a, int pos) {
+    __m128 ret;
+    if( pos == 0)
+    {
+       ret = (__m128)lasx_extracti128_lo((__m256i)a);
+    } else {
+       ret = (__m128)lasx_extracti128_hi((__m256i)a);
+    }
+    return ret;
+}
+
+static __m128i lsx_hadd_h(__m128i a, __m128i b) {
+    __m128i tmp1 = __lsx_vpickev_h(b, a);
+    __m128i tmp2 = __lsx_vpickod_h(b, a);
+    return __lsx_vadd_h(tmp1, tmp2);
+}
+
+static __m128i lsx_hadd_w(__m128i a, __m128i b) {
+    __m128i tmp1 = __lsx_vpickev_w(b, a);
+    __m128i tmp2 = __lsx_vpickod_w(b, a);
+    return __lsx_vadd_w(tmp1, tmp2);
+}
+
+static __m128 lsx_hadd_s(__m128 a, __m128 b) {
+    __m128 tmp1 = (__m128)__lsx_vpickev_w((__m128i)b, (__m128i)a);
+    __m128 tmp2 = (__m128)__lsx_vpickod_w((__m128i)b, (__m128i)a);
+
+    return __lsx_vfadd_s(tmp1, tmp2);
+}
+
+static __m256i lasx_maddubs_h(__m256i a, __m256i b) {
+    __m256i tmp1, tmp2;
+    tmp1 = __lasx_xvmulwev_h_b(a, b);
+    tmp2 = __lasx_xvmulwod_h_b(a, b);
+    return __lasx_xvsadd_h(tmp1, tmp2);
+}
+
+static __m256i lasx_madd_h(__m256i a, __m256i b) {
+    __m256i tmp1, tmp2;
+    tmp1 = __lasx_xvmulwev_w_h(a, b);
+    tmp2 = __lasx_xvmulwod_w_h(a, b);
+    return __lasx_xvadd_w(tmp1, tmp2);
+}
+
+static __m256i lasx_packs_w(__m256i a, __m256i b) {
+    __m256i tmp, tmp1;
+    tmp = __lasx_xvsat_w(a, 15);
+    tmp1 = __lasx_xvsat_w(b, 15);
+    return __lasx_xvpickev_h(tmp1, tmp);
+}
+
+static __m256i lasx_packs_h(__m256i a, __m256i b) {
+    __m256i tmp, tmp1;
+    tmp = __lasx_xvsat_h(a, 7);
+    tmp1 = __lasx_xvsat_h(b, 7);
+    return __lasx_xvpickev_b(tmp1, tmp);
+}
+
+static __m128i lsx_packs_w(__m128i a, __m128i b) {
+    __m128i tmp, tmp1;
+    tmp = __lsx_vsat_w(a, 15);
+    tmp1 = __lsx_vsat_w(b, 15);
+    return __lsx_vpickev_h(tmp1, tmp);
+}
+
+static __m128i lsx_packs_h(__m128i a, __m128i b) {
+    __m128i tmp, tmp1;
+    tmp = __lsx_vsat_h(a, 7);
+    tmp1 = __lsx_vsat_h(b, 7);
+    return __lsx_vpickev_b(tmp1, tmp);
+}
+
+static __m128i lsx_packus_h(__m128i a, __m128i b) {
+    __m128i tmp, tmp1;
+    tmp = __lsx_vsat_hu(a, 7);
+    tmp1 = __lsx_vsat_hu(b, 7);
+    return __lsx_vpickev_b(tmp1, tmp);
+}
+
+
+static __m128i lsx_maddubs_h(__m128i a, __m128i b) {
+    __m128i tmp1, tmp2;
+    tmp1 = __lsx_vmulwev_h_b(a, b);
+    tmp2 = __lsx_vmulwod_h_b(a, b);
+    return __lsx_vsadd_h(tmp1, tmp2);
+}
+
+static __m128i lsx_madd_h(__m128i a, __m128i b) {
+    __m128i tmp1, tmp2;
+    tmp1 = __lsx_vmulwev_w_h(a, b);
+    tmp2 = __lsx_vmulwod_w_h(a, b);
+    return __lsx_vadd_w(tmp1, tmp2);
+}
+
+// multiply int8_t, add results pairwise twice
+static inline __m128i mul_sum_i8_pairs(const __m128i x, const __m128i y) {
+    // Get absolute values of x vectors
+    const __m128i ax = __lsx_vsigncov_b(x, x);
+    // Sign the values of the y vectors
+    const __m128i sy = __lsx_vsigncov_b(x, y);
+    // Perform multiplication and create 16-bit values
+    const __m128i dot = lsx_maddubs_h(ax, sy);
+    const __m128i ones = __lsx_vreplgr2vr_h(1);
+    return lsx_madd_h(ones, dot);
+}
+
+// horizontally add 8 floats
+static inline float hsum_float_8(const __m256 x) {
+    __m128 res = lasx_extractf128(x, 1);
+    ft_union tmp;
+    res = __lsx_vfadd_s(res, lasx_extractf128(x, 0));
+    res = __lsx_vfadd_s(res, (__m128)__lsx_vpickod_d((__m128i)res, (__m128i)res));
+    res = __lsx_vfadd_s(res, (__m128)__lsx_vinsgr2vr_w(__lsx_vldi(0), __lsx_vpickve2gr_w(res, 1), 0));
+    tmp.i = __lsx_vpickve2gr_w(res, 0);
+    return tmp.f;
+}
+
+// horizontally add 8 int32_t
+static inline int hsum_i32_8(const __m256i a) {
+
+    __m256i tmp1 = __lasx_xvpermi_q(a, a, 0x11);
+    __m256i tmp2 = __lasx_xvpermi_q(a, a, 0x00);
+
+    __m128i  tmp1_128 = lasx_extracti128_lo(tmp1);
+    __m128i  tmp2_128 = lasx_extracti128_lo(tmp2);
+
+    __m128i sum128 = __lsx_vadd_w(tmp1_128, tmp2_128);
+
+    __m128i ev = __lsx_vpickev_w(sum128, sum128);
+    __m128i od = __lsx_vpickod_w(sum128, sum128);
+    __m128i sum64 = __lsx_vadd_w(ev, od);
+
+    int sum64_1, sum64_2;
+    sum64_1 = __lsx_vpickve2gr_w(sum64, 0);
+    sum64_2 = __lsx_vpickve2gr_w(sum64, 1);
+
+    return  sum64_1 + sum64_2;
+}
+
+// horizontally add 4 int32_t
+static inline int hsum_i32_4(const __m128i a) {
+    __m128i ev = __lsx_vpickev_w(a, a);
+    __m128i od = __lsx_vpickod_w(a, a);
+    __m128i sum64 = __lsx_vadd_w(ev, od);
+
+    int sum64_1, sum64_2;
+    sum64_1 = __lsx_vpickve2gr_w(sum64, 0);
+    sum64_2 = __lsx_vpickve2gr_w(sum64, 1);
+
+    return  sum64_1 + sum64_2;
+}
+
+// spread 32 bits to 32 bytes { 0x00, 0xFF }
+static inline __m256i bytes_from_bits_32(const uint8_t * x) {
+
+    uint32_t x32;
+    memcpy(&x32, x, sizeof(uint32_t));
+    const __m256i shuf_mask = lasx_set_d(
+            0x0303030303030303, 0x0202020202020202,
+            0x0101010101010101, 0x0000000000000000);
+
+    __m256i bytes = lasx_shuffle_b(__lasx_xvreplgr2vr_w(x32), shuf_mask);
+    const __m256i bit_mask = __lasx_xvreplgr2vr_d(0x7fbfdfeff7fbfdfe);
+    bytes = __lasx_xvor_v(bytes, bit_mask);
+    return __lasx_xvseq_b(bytes, __lasx_xvreplgr2vr_d(-1));
+}
+
+// Unpack 32 4-bit fields into 32 bytes
+// The output vector contains 32 bytes, each one in [ 0 .. 15 ] interval
+static inline __m256i bytes_from_nibbles_32(const uint8_t * rsi) {
+    const __m128i lo = __lsx_vld((const __m128i *)rsi, 0);
+    __m128i hi = __lsx_vsrli_h(lo, 4);
+    return __lasx_xvandi_b(lasx_insertf128(hi, lo), 0xf);
+}
+
+// add int16_t pairwise and return as float vector
+static inline __m256 sum_i16_pairs_float(const __m256i x) {
+    __m256i v = __lasx_xvpackod_h(x, x);
+    __m256i summed_pairs = __lasx_xvaddwev_w_h(x, v);
+    return __lasx_xvffint_s_w(summed_pairs);
+}
+
+static inline __m256 mul_sum_us8_pairs_float(const __m256i ax, const __m256i sy) {
+    // Perform multiplication and create 16-bit values
+    const __m256i dot = lasx_maddubs_h(ax, sy);
+    return sum_i16_pairs_float(dot);
+}
+
+// multiply int8_t, add results pairwise twice and return as float vector
+static inline __m256 mul_sum_i8_pairs_float(const __m256i x, const __m256i y) {
+
+    // Get absolute values of x vectors
+    const __m256i ax = __lasx_xvsigncov_b(x, x);
+    // Sign the values of the y vectors
+    const __m256i sy = __lasx_xvsigncov_b(x, y);
+
+    return mul_sum_us8_pairs_float(ax, sy);
+}
+
+static inline __m128i packNibbles( __m256i bytes ) {
+    // Move bits within 16-bit lanes from 0000_abcd_0000_efgh into 0000_0000_abcd_efgh
+    const __m256i lowByte = __lasx_xvreplgr2vr_h(0xFF);
+     __m256i high = __lasx_xvandn_v(lowByte, bytes);
+    __m256i low = __lasx_xvand_v(lowByte, bytes);
+    high = __lasx_xvsrli_h(high, 4);
+    bytes = __lasx_xvor_v(low, high);
+    // Compress uint16_t lanes into bytes
+    __m128i *r0 = (__m128i *)&bytes;
+    __m256i tmp_h128 = __lasx_xvpermi_q(bytes, bytes, 0x11);
+    __m128i *r1 = (__m128i *)&tmp_h128;
+
+    __m128i zero = __lsx_vldi(0);
+    __m128i tmp, tmp2, tmp3;
+
+    tmp = __lsx_vmax_h(zero, *r0);
+    tmp2 = __lsx_vsat_hu(tmp, 7);
+
+    tmp = __lsx_vmax_h(zero, *r1);
+    tmp3 = __lsx_vsat_hu(tmp, 7);
+    return  __lsx_vpickev_b(tmp3, tmp2);
+}
+#endif  //__loongarch_asx
+
+// reference implementation for deterministic creation of model files
+void quantize_row_q4_0_ref(const float * restrict x, block_q4_0 * restrict y, int64_t k) {
+    static const int qk = QK4_0;
+
+    assert(k % qk == 0);
+
+    const int nb = k / qk;
+
+    for (int i = 0; i < nb; i++) {
+        float amax = 0.0f; // absolute max
+        float max  = 0.0f;
+
+        for (int j = 0; j < qk; j++) {
+            const float v = x[i*qk + j];
+            if (amax < fabsf(v)) {
+                amax = fabsf(v);
+                max  = v;
+            }
+        }
+
+        const float d  = max / -8;
+        const float id = d ? 1.0f/d : 0.0f;
+
+        y[i].d = GGML_FP32_TO_FP16(d);
+
+        for (int j = 0; j < qk/2; ++j) {
+            const float x0 = x[i*qk + 0    + j]*id;
+            const float x1 = x[i*qk + qk/2 + j]*id;
+
+            const uint8_t xi0 = MIN(15, (int8_t)(x0 + 8.5f));
+            const uint8_t xi1 = MIN(15, (int8_t)(x1 + 8.5f));
+
+            y[i].qs[j]  = xi0;
+            y[i].qs[j] |= xi1 << 4;
+        }
+    }
+}
+
+void quantize_row_q4_0(const float * restrict x, void * restrict y, int64_t k) {
+    quantize_row_q4_0_ref(x, y, k);
+}
+
+
+void quantize_row_q4_1_ref(const float * restrict x, block_q4_1 * restrict y, int64_t k) {
+    const int qk = QK4_1;
+
+    assert(k % qk == 0);
+
+    const int nb = k / qk;
+
+    for (int i = 0; i < nb; i++) {
+        float min = FLT_MAX;
+        float max = -FLT_MAX;
+
+        for (int j = 0; j < qk; j++) {
+            const float v = x[i*qk + j];
+
+            if (v < min) min = v;
+            if (v > max) max = v;
+        }
+
+        const float d  = (max - min) / ((1 << 4) - 1);
+        const float id = d ? 1.0f/d : 0.0f;
+
+        y[i].d = GGML_FP32_TO_FP16(d);
+        y[i].m = GGML_FP32_TO_FP16(min);
+
+        for (int j = 0; j < qk/2; ++j) {
+            const float x0 = (x[i*qk + 0    + j] - min)*id;
+            const float x1 = (x[i*qk + qk/2 + j] - min)*id;
+
+            const uint8_t xi0 = MIN(15, (int8_t)(x0 + 0.5f));
+            const uint8_t xi1 = MIN(15, (int8_t)(x1 + 0.5f));
+
+            y[i].qs[j]  = xi0;
+            y[i].qs[j] |= xi1 << 4;
+        }
+    }
+}
+
+void quantize_row_q4_1(const float * restrict x, void * restrict y, int64_t k) {
+    quantize_row_q4_1_ref(x, y, k);
+}
+
+void quantize_row_q5_0_ref(const float * restrict x, block_q5_0 * restrict y, int64_t k) {
+    static const int qk = QK5_0;
+
+    assert(k % qk == 0);
+
+    const int nb = k / qk;
+
+    for (int i = 0; i < nb; i++) {
+        float amax = 0.0f; // absolute max
+        float max  = 0.0f;
+
+        for (int j = 0; j < qk; j++) {
+            const float v = x[i*qk + j];
+            if (amax < fabsf(v)) {
+                amax = fabsf(v);
+                max  = v;
+            }
+        }
+
+        const float d  = max / -16;
+        const float id = d ? 1.0f/d : 0.0f;
+
+        y[i].d = GGML_FP32_TO_FP16(d);
+
+        uint32_t qh = 0;
+
+        for (int j = 0; j < qk/2; ++j) {
+            const float x0 = x[i*qk + 0    + j]*id;
+            const float x1 = x[i*qk + qk/2 + j]*id;
+
+            const uint8_t xi0 = MIN(31, (int8_t)(x0 + 16.5f));
+            const uint8_t xi1 = MIN(31, (int8_t)(x1 + 16.5f));
+
+            y[i].qs[j] = (xi0 & 0x0F) | ((xi1 & 0x0F) << 4);
+
+            // get the 5-th bit and store it in qh at the right position
+            qh |= ((xi0 & 0x10u) >> 4) << (j + 0);
+            qh |= ((xi1 & 0x10u) >> 4) << (j + qk/2);
+        }
+
+        memcpy(&y[i].qh, &qh, sizeof(qh));
+    }
+}
+
+void quantize_row_q5_0(const float * restrict x, void * restrict y, int64_t k) {
+    quantize_row_q5_0_ref(x, y, k);
+}
+
+void quantize_row_q5_1_ref(const float * restrict x, block_q5_1 * restrict y, int64_t k) {
+    const int qk = QK5_1;
+
+    assert(k % qk == 0);
+
+    const int nb = k / qk;
+
+    for (int i = 0; i < nb; i++) {
+        float min = FLT_MAX;
+        float max = -FLT_MAX;
+
+        for (int j = 0; j < qk; j++) {
+            const float v = x[i*qk + j];
+
+            if (v < min) min = v;
+            if (v > max) max = v;
+        }
+
+        const float d  = (max - min) / ((1 << 5) - 1);
+        const float id = d ? 1.0f/d : 0.0f;
+
+        y[i].d = GGML_FP32_TO_FP16(d);
+        y[i].m = GGML_FP32_TO_FP16(min);
+
+        uint32_t qh = 0;
+
+        for (int j = 0; j < qk/2; ++j) {
+            const float x0 = (x[i*qk + 0    + j] - min)*id;
+            const float x1 = (x[i*qk + qk/2 + j] - min)*id;
+
+            const uint8_t xi0 = (uint8_t)(x0 + 0.5f);
+            const uint8_t xi1 = (uint8_t)(x1 + 0.5f);
+
+            y[i].qs[j] = (xi0 & 0x0F) | ((xi1 & 0x0F) << 4);
+
+            // get the 5-th bit and store it in qh at the right position
+            qh |= ((xi0 & 0x10u) >> 4) << (j + 0);
+            qh |= ((xi1 & 0x10u) >> 4) << (j + qk/2);
+        }
+
+        memcpy(&y[i].qh, &qh, sizeof(y[i].qh));
+    }
+}
+
+void quantize_row_q5_1(const float * restrict x, void * restrict y, int64_t k) {
+    quantize_row_q5_1_ref(x, y, k);
+}
+
+// reference implementation for deterministic creation of model files
+void quantize_row_q8_0_ref(const float * restrict x, block_q8_0 * restrict y, int64_t k) {
+    assert(k % QK8_0 == 0);
+    const int nb = k / QK8_0;
+
+    for (int i = 0; i < nb; i++) {
+        float amax = 0.0f; // absolute max
+
+        for (int j = 0; j < QK8_0; j++) {
+            const float v = x[i*QK8_0 + j];
+            amax = MAX(amax, fabsf(v));
+        }
+
+        const float d = amax / ((1 << 7) - 1);
+        const float id = d ? 1.0f/d : 0.0f;
+
+        y[i].d = GGML_FP32_TO_FP16(d);
+
+        for (int j = 0; j < QK8_0; ++j) {
+            const float x0 = x[i*QK8_0 + j]*id;
+
+            y[i].qs[j] = roundf(x0);
+        }
+    }
+}
+
+void quantize_row_q8_0(const float * restrict x, void * restrict vy, int64_t k) {
+    assert(QK8_0 == 32);
+    assert(k % QK8_0 == 0);
+    const int nb = k / QK8_0;
+
+    block_q8_0 * restrict y = vy;
+
+#if defined(__ARM_NEON)
+    for (int i = 0; i < nb; i++) {
+        float32x4_t srcv [8];
+        float32x4_t asrcv[8];
+        float32x4_t amaxv[8];
+
+        for (int j = 0; j < 8; j++) srcv[j]  = vld1q_f32(x + i*32 + 4*j);
+        for (int j = 0; j < 8; j++) asrcv[j] = vabsq_f32(srcv[j]);
+
+        for (int j = 0; j < 4; j++) amaxv[2*j] = vmaxq_f32(asrcv[2*j], asrcv[2*j+1]);
+        for (int j = 0; j < 2; j++) amaxv[4*j] = vmaxq_f32(amaxv[4*j], amaxv[4*j+2]);
+        for (int j = 0; j < 1; j++) amaxv[8*j] = vmaxq_f32(amaxv[8*j], amaxv[8*j+4]);
+
+        const float amax = vmaxvq_f32(amaxv[0]);
+
+        const float d = amax / ((1 << 7) - 1);
+        const float id = d ? 1.0f/d : 0.0f;
+
+        y[i].d = GGML_FP32_TO_FP16(d);
+
+        for (int j = 0; j < 8; j++) {
+            const float32x4_t v  = vmulq_n_f32(srcv[j], id);
+            const int32x4_t   vi = vcvtnq_s32_f32(v);
+
+            y[i].qs[4*j + 0] = vgetq_lane_s32(vi, 0);
+            y[i].qs[4*j + 1] = vgetq_lane_s32(vi, 1);
+            y[i].qs[4*j + 2] = vgetq_lane_s32(vi, 2);
+            y[i].qs[4*j + 3] = vgetq_lane_s32(vi, 3);
+        }
+    }
+#elif defined(__wasm_simd128__)
+    for (int i = 0; i < nb; i++) {
+        v128_t srcv [8];
+        v128_t asrcv[8];
+        v128_t amaxv[8];
+
+        for (int j = 0; j < 8; j++) srcv[j]  = wasm_v128_load(x + i*32 + 4*j);
+        for (int j = 0; j < 8; j++) asrcv[j] = wasm_f32x4_abs(srcv[j]);
+
+        for (int j = 0; j < 4; j++) amaxv[2*j] = wasm_f32x4_max(asrcv[2*j], asrcv[2*j+1]);
+        for (int j = 0; j < 2; j++) amaxv[4*j] = wasm_f32x4_max(amaxv[4*j], amaxv[4*j+2]);
+        for (int j = 0; j < 1; j++) amaxv[8*j] = wasm_f32x4_max(amaxv[8*j], amaxv[8*j+4]);
+
+        const float amax = MAX(MAX(wasm_f32x4_extract_lane(amaxv[0], 0),
+                                   wasm_f32x4_extract_lane(amaxv[0], 1)),
+                               MAX(wasm_f32x4_extract_lane(amaxv[0], 2),
+                                   wasm_f32x4_extract_lane(amaxv[0], 3)));
+
+        const float d = amax / ((1 << 7) - 1);
+        const float id = d ? 1.0f/d : 0.0f;
+
+        y[i].d = GGML_FP32_TO_FP16(d);
+
+        for (int j = 0; j < 8; j++) {
+            const v128_t v  = wasm_f32x4_mul(srcv[j], wasm_f32x4_splat(id));
+            const v128_t vi = wasm_i32x4_trunc_sat_f32x4(v);
+
+            y[i].qs[4*j + 0] = wasm_i32x4_extract_lane(vi, 0);
+            y[i].qs[4*j + 1] = wasm_i32x4_extract_lane(vi, 1);
+            y[i].qs[4*j + 2] = wasm_i32x4_extract_lane(vi, 2);
+            y[i].qs[4*j + 3] = wasm_i32x4_extract_lane(vi, 3);
+        }
+    }
+#elif defined(__AVX2__) || defined(__AVX__)
+    for (int i = 0; i < nb; i++) {
+        // Load elements into 4 AVX vectors
+        __m256 v0 = _mm256_loadu_ps( x );
+        __m256 v1 = _mm256_loadu_ps( x + 8 );
+        __m256 v2 = _mm256_loadu_ps( x + 16 );
+        __m256 v3 = _mm256_loadu_ps( x + 24 );
+        x += 32;
+
+        // Compute max(abs(e)) for the block
+        const __m256 signBit = _mm256_set1_ps( -0.0f );
+        __m256 maxAbs = _mm256_andnot_ps( signBit, v0 );
+        maxAbs = _mm256_max_ps( maxAbs, _mm256_andnot_ps( signBit, v1 ) );
+        maxAbs = _mm256_max_ps( maxAbs, _mm256_andnot_ps( signBit, v2 ) );
+        maxAbs = _mm256_max_ps( maxAbs, _mm256_andnot_ps( signBit, v3 ) );
+
+        __m128 max4 = _mm_max_ps( _mm256_extractf128_ps( maxAbs, 1 ), _mm256_castps256_ps128( maxAbs ) );
+        max4 = _mm_max_ps( max4, _mm_movehl_ps( max4, max4 ) );
+        max4 = _mm_max_ss( max4, _mm_movehdup_ps( max4 ) );
+        const float maxScalar = _mm_cvtss_f32( max4 );
+
+        // Quantize these floats
+        const float d = maxScalar / 127.f;
+        y[i].d = GGML_FP32_TO_FP16(d);
+        const float id = ( maxScalar != 0.0f ) ? 127.f / maxScalar : 0.0f;
+        const __m256 mul = _mm256_set1_ps( id );
+
+        // Apply the multiplier
+        v0 = _mm256_mul_ps( v0, mul );
+        v1 = _mm256_mul_ps( v1, mul );
+        v2 = _mm256_mul_ps( v2, mul );
+        v3 = _mm256_mul_ps( v3, mul );
+
+        // Round to nearest integer
+        v0 = _mm256_round_ps( v0, _MM_ROUND_NEAREST );
+        v1 = _mm256_round_ps( v1, _MM_ROUND_NEAREST );
+        v2 = _mm256_round_ps( v2, _MM_ROUND_NEAREST );
+        v3 = _mm256_round_ps( v3, _MM_ROUND_NEAREST );
+
+        // Convert floats to integers
+        __m256i i0 = _mm256_cvtps_epi32( v0 );
+        __m256i i1 = _mm256_cvtps_epi32( v1 );
+        __m256i i2 = _mm256_cvtps_epi32( v2 );
+        __m256i i3 = _mm256_cvtps_epi32( v3 );
+
+#if defined(__AVX2__)
+        // Convert int32 to int16
+        i0 = _mm256_packs_epi32( i0, i1 );	// 0, 1, 2, 3,  8, 9, 10, 11,  4, 5, 6, 7, 12, 13, 14, 15
+        i2 = _mm256_packs_epi32( i2, i3 );	// 16, 17, 18, 19,  24, 25, 26, 27,  20, 21, 22, 23, 28, 29, 30, 31
+                                            // Convert int16 to int8
+        i0 = _mm256_packs_epi16( i0, i2 );	// 0, 1, 2, 3,  8, 9, 10, 11,  16, 17, 18, 19,  24, 25, 26, 27,  4, 5, 6, 7, 12, 13, 14, 15, 20, 21, 22, 23, 28, 29, 30, 31
+
+        // We got our precious signed bytes, but the order is now wrong
+        // These AVX2 pack instructions process 16-byte pieces independently
+        // The following instruction is fixing the order
+        const __m256i perm = _mm256_setr_epi32( 0, 4, 1, 5, 2, 6, 3, 7 );
+        i0 = _mm256_permutevar8x32_epi32( i0, perm );
+
+        _mm256_storeu_si256((__m256i *)y[i].qs, i0);
+#else
+        // Since we don't have in AVX some necessary functions,
+        // we split the registers in half and call AVX2 analogs from SSE
+        __m128i ni0 = _mm256_castsi256_si128( i0 );
+        __m128i ni1 = _mm256_extractf128_si256( i0, 1);
+        __m128i ni2 = _mm256_castsi256_si128( i1 );
+        __m128i ni3 = _mm256_extractf128_si256( i1, 1);
+        __m128i ni4 = _mm256_castsi256_si128( i2 );
+        __m128i ni5 = _mm256_extractf128_si256( i2, 1);
+        __m128i ni6 = _mm256_castsi256_si128( i3 );
+        __m128i ni7 = _mm256_extractf128_si256( i3, 1);
+
+        // Convert int32 to int16
+        ni0 = _mm_packs_epi32( ni0, ni1 );
+        ni2 = _mm_packs_epi32( ni2, ni3 );
+        ni4 = _mm_packs_epi32( ni4, ni5 );
+        ni6 = _mm_packs_epi32( ni6, ni7 );
+        // Convert int16 to int8
+        ni0 = _mm_packs_epi16( ni0, ni2 );
+        ni4 = _mm_packs_epi16( ni4, ni6 );
+
+        _mm_storeu_si128((__m128i *)(y[i].qs +  0), ni0);
+        _mm_storeu_si128((__m128i *)(y[i].qs + 16), ni4);
+#endif
+    }
+#elif defined(__riscv_v_intrinsic)
+
+    size_t vl = __riscv_vsetvl_e32m4(QK8_0);
+
+    for (int i = 0; i < nb; i++) {
+        // load elements
+        vfloat32m4_t v_x   = __riscv_vle32_v_f32m4(x+i*QK8_0, vl);
+
+        vfloat32m4_t vfabs = __riscv_vfabs_v_f32m4(v_x, vl);
+        vfloat32m1_t tmp   = __riscv_vfmv_v_f_f32m1(0.0f, vl);
+        vfloat32m1_t vmax  = __riscv_vfredmax_vs_f32m4_f32m1(vfabs, tmp, vl);
+        float amax = __riscv_vfmv_f_s_f32m1_f32(vmax);
+
+        const float d = amax / ((1 << 7) - 1);
+        const float id = d ? 1.0f/d : 0.0f;
+
+        y[i].d = GGML_FP32_TO_FP16(d);
+
+        vfloat32m4_t x0 = __riscv_vfmul_vf_f32m4(v_x, id, vl);
+
+        // convert to integer
+        vint16m2_t   vi = __riscv_vfncvt_x_f_w_i16m2(x0, vl);
+        vint8m1_t    vs = __riscv_vncvt_x_x_w_i8m1(vi, vl);
+
+        // store result
+        __riscv_vse8_v_i8m1(y[i].qs , vs, vl);
+    }
+
+#elif defined(__POWER9_VECTOR__)
+    for (int i = 0; i < nb; i++) {
+        vector float srcv [8];
+        vector float asrcv[8];
+        vector float amaxv[8];
+        vector signed int vi[8];
+
+        for (int j = 0; j < 8; j++) srcv[j]  = vec_xl(0, x + i*32 + 4*j);
+        for (int j = 0; j < 8; j++) asrcv[j] = vec_abs(srcv[j]);
+
+        for (int j = 0; j < 4; j++) amaxv[2*j] = vec_max(asrcv[2*j], asrcv[2*j+1]);
+        for (int j = 0; j < 2; j++) amaxv[4*j] = vec_max(amaxv[4*j], amaxv[4*j+2]);
+        for (int j = 0; j < 1; j++) amaxv[8*j] = vec_max(amaxv[8*j], amaxv[8*j+4]);
+
+        const float amax = MAX(MAX(vec_extract(amaxv[0], 0),
+                                   vec_extract(amaxv[0], 1)),
+                               MAX(vec_extract(amaxv[0], 2),
+                                   vec_extract(amaxv[0], 3)));
+
+        const float d = amax / ((1 << 7) - 1);
+        const float id = d ? 1.0f/d : 0.0f;
+        const vector float vid = vec_splats(id);
+
+        y[i].d = GGML_FP32_TO_FP16(d);
+
+        for (int j = 0; j < 8; j++) {
+            const vector float v  = vec_round(vec_mul(srcv[j], vid));
+            vi[j] = vec_cts(v, 0);
+        }
+        vec_xst(vec_pack(vec_pack(vi[0], vi[1]), vec_pack(vi[2], vi[3])),  0, &y[i].qs[0]);
+        vec_xst(vec_pack(vec_pack(vi[4], vi[5]), vec_pack(vi[6], vi[7])), 16, &y[i].qs[0]);
+    }
+
+#elif defined(__loongarch_asx)
+    for (int i = 0; i < nb; i++) {
+        ft_union fi;
+        __m256 v0 = (__m256)__lasx_xvld( x , 0);
+        __m256 v1 = (__m256)__lasx_xvld( x , 32);
+        __m256 v2 = (__m256)__lasx_xvld( x , 64);
+        __m256 v3 = (__m256)__lasx_xvld( x , 96);
+        x += 32;
+
+        // Compute max(abs(e)) for the block
+        const __m256 sign_bit = __lasx_xvreplfr2vr_s( -0.0f );
+        __m256 max_abs = (__m256)__lasx_xvandn_v( (__m256i)sign_bit, (__m256i)v0 );
+        max_abs = __lasx_xvfmax_s( max_abs, (__m256)__lasx_xvandn_v( (__m256i)sign_bit, (__m256i)v1 ) );
+        max_abs = __lasx_xvfmax_s( max_abs, (__m256)__lasx_xvandn_v( (__m256i)sign_bit, (__m256i)v2 ) );
+        max_abs = __lasx_xvfmax_s( max_abs, (__m256)__lasx_xvandn_v( (__m256i)sign_bit, (__m256i)v3 ) );
+
+        __m128 max4 = __lsx_vfmax_s( lasx_extractf128( max_abs, 1 ), lasx_extractf128( max_abs , 0) );
+        max4 = __lsx_vfmax_s( max4, (__m128)__lsx_vpickod_d((__m128i) max4, (__m128i)max4 ) );
+        __m128 tmp = max4;
+        max4 = __lsx_vfmax_s( max4, (__m128)__lsx_vinsgr2vr_w(tmp, __lsx_vpickve2gr_w( max4, 1 ), 0 ));
+        fi.i = __lsx_vpickve2gr_w( (__m128i)max4, 0 );
+        const float max_scalar = fi.f;
+
+        // Quantize these floats
+        const float d = max_scalar / 127.f;
+        y[i].d = GGML_FP32_TO_FP16(d);
+        const float id = ( max_scalar != 0.0f ) ? 127.f / max_scalar : 0.0f;
+        const __m256 mul = (__m256)__lasx_xvreplfr2vr_s( id );
+
+        // Apply the multiplier
+        v0 = __lasx_xvfmul_s( v0, mul );
+        v1 = __lasx_xvfmul_s( v1, mul );
+        v2 = __lasx_xvfmul_s( v2, mul );
+        v3 = __lasx_xvfmul_s( v3, mul );
+
+        // Round to nearest integer
+        __m256i i0 = __lasx_xvftintrne_w_s( v0 );
+        __m256i i1 = __lasx_xvftintrne_w_s( v1 );
+        __m256i i2 = __lasx_xvftintrne_w_s( v2 );
+        __m256i i3 = __lasx_xvftintrne_w_s( v3 );
+
+        __m128i ni0 = lasx_extracti128( i0, 0 );
+        __m128i ni1 = lasx_extracti128( i0, 1);
+        __m128i ni2 = lasx_extracti128( i1, 0);
+        __m128i ni3 = lasx_extracti128( i1, 1);
+        __m128i ni4 = lasx_extracti128( i2, 0);
+        __m128i ni5 = lasx_extracti128( i2, 1);
+        __m128i ni6 = lasx_extracti128( i3, 0);
+        __m128i ni7 = lasx_extracti128( i3, 1);
+
+        // Convert int32 to int16
+        ni0 = lsx_packs_w( ni0, ni1 );
+        ni2 = lsx_packs_w( ni2, ni3 );
+        ni4 = lsx_packs_w( ni4, ni5 );
+        ni6 = lsx_packs_w( ni6, ni7 );
+        // Convert int16 to int8
+        ni0 = lsx_packs_h( ni0, ni2 );
+        ni4 = lsx_packs_h( ni4, ni6 );
+
+        __lsx_vst(ni0, (__m128i *)(y[i].qs +  0), 0);
+        __lsx_vst(ni4, (__m128i *)(y[i].qs + 16), 0);
+
+    }
+#else
+    GGML_UNUSED(nb);
+    // scalar
+    quantize_row_q8_0_ref(x, y, k);
+#endif
+}
+
+// reference implementation for deterministic creation of model files
+void quantize_row_q8_1_ref(const float * restrict x, block_q8_1 * restrict y, int64_t k) {
+    assert(QK8_1 == 32);
+    assert(k % QK8_1 == 0);
+    const int nb = k / QK8_1;
+
+    for (int i = 0; i < nb; i++) {
+        float amax = 0.0f; // absolute max
+
+        for (int j = 0; j < QK8_1; j++) {
+            const float v = x[i*QK8_1 + j];
+            amax = MAX(amax, fabsf(v));
+        }
+
+        const float d = amax / ((1 << 7) - 1);
+        const float id = d ? 1.0f/d : 0.0f;
+
+        y[i].d = GGML_FP32_TO_FP16(d);
+
+        int sum = 0;
+
+        for (int j = 0; j < QK8_1/2; ++j) {
+            const float v0 = x[i*QK8_1           + j]*id;
+            const float v1 = x[i*QK8_1 + QK8_1/2 + j]*id;
+
+            y[i].qs[          j] = roundf(v0);
+            y[i].qs[QK8_1/2 + j] = roundf(v1);
+
+            sum += y[i].qs[          j];
+            sum += y[i].qs[QK8_1/2 + j];
+        }
+
+        y[i].s = GGML_FP32_TO_FP16(sum*d);
+    }
+}
+
+void quantize_row_q8_1(const float * restrict x, void * restrict vy, int64_t k) {
+    assert(k % QK8_1 == 0);
+    const int nb = k / QK8_1;
+
+    block_q8_1 * restrict y = vy;
+
+#if defined(__ARM_NEON)
+    for (int i = 0; i < nb; i++) {
+        float32x4_t srcv [8];
+        float32x4_t asrcv[8];
+        float32x4_t amaxv[8];
+
+        for (int j = 0; j < 8; j++) srcv[j]  = vld1q_f32(x + i*32 + 4*j);
+        for (int j = 0; j < 8; j++) asrcv[j] = vabsq_f32(srcv[j]);
+
+        for (int j = 0; j < 4; j++) amaxv[2*j] = vmaxq_f32(asrcv[2*j], asrcv[2*j+1]);
+        for (int j = 0; j < 2; j++) amaxv[4*j] = vmaxq_f32(amaxv[4*j], amaxv[4*j+2]);
+        for (int j = 0; j < 1; j++) amaxv[8*j] = vmaxq_f32(amaxv[8*j], amaxv[8*j+4]);
+
+        const float amax = vmaxvq_f32(amaxv[0]);
+
+        const float d = amax / ((1 << 7) - 1);
+        const float id = d ? 1.0f/d : 0.0f;
+
+        y[i].d = GGML_FP32_TO_FP16(d);
+
+        int32x4_t accv = vdupq_n_s32(0);
+
+        for (int j = 0; j < 8; j++) {
+            const float32x4_t v  = vmulq_n_f32(srcv[j], id);
+            const int32x4_t   vi = vcvtnq_s32_f32(v);
+
+            y[i].qs[4*j + 0] = vgetq_lane_s32(vi, 0);
+            y[i].qs[4*j + 1] = vgetq_lane_s32(vi, 1);
+            y[i].qs[4*j + 2] = vgetq_lane_s32(vi, 2);
+            y[i].qs[4*j + 3] = vgetq_lane_s32(vi, 3);
+
+            accv = vaddq_s32(accv, vi);
+        }
+
+        y[i].s = GGML_FP32_TO_FP16(d * vaddvq_s32(accv));
+    }
+#elif defined(__wasm_simd128__)
+    for (int i = 0; i < nb; i++) {
+        v128_t srcv [8];
+        v128_t asrcv[8];
+        v128_t amaxv[8];
+
+        for (int j = 0; j < 8; j++) srcv[j]  = wasm_v128_load(x + i*32 + 4*j);
+        for (int j = 0; j < 8; j++) asrcv[j] = wasm_f32x4_abs(srcv[j]);
+
+        for (int j = 0; j < 4; j++) amaxv[2*j] = wasm_f32x4_max(asrcv[2*j], asrcv[2*j+1]);
+        for (int j = 0; j < 2; j++) amaxv[4*j] = wasm_f32x4_max(amaxv[4*j], amaxv[4*j+2]);
+        for (int j = 0; j < 1; j++) amaxv[8*j] = wasm_f32x4_max(amaxv[8*j], amaxv[8*j+4]);
+
+        const float amax = MAX(MAX(wasm_f32x4_extract_lane(amaxv[0], 0),
+                                   wasm_f32x4_extract_lane(amaxv[0], 1)),
+                               MAX(wasm_f32x4_extract_lane(amaxv[0], 2),
+                                   wasm_f32x4_extract_lane(amaxv[0], 3)));
+
+        const float d = amax / ((1 << 7) - 1);
+        const float id = d ? 1.0f/d : 0.0f;
+
+        y[i].d = GGML_FP32_TO_FP16(d);
+
+        v128_t accv = wasm_i32x4_splat(0);
+
+        for (int j = 0; j < 8; j++) {
+            const v128_t v  = wasm_f32x4_mul(srcv[j], wasm_f32x4_splat(id));
+            const v128_t vi = wasm_i32x4_trunc_sat_f32x4(v);
+
+            y[i].qs[4*j + 0] = wasm_i32x4_extract_lane(vi, 0);
+            y[i].qs[4*j + 1] = wasm_i32x4_extract_lane(vi, 1);
+            y[i].qs[4*j + 2] = wasm_i32x4_extract_lane(vi, 2);
+            y[i].qs[4*j + 3] = wasm_i32x4_extract_lane(vi, 3);
+
+            accv = wasm_i32x4_add(accv, vi);
+        }
+
+        y[i].s = GGML_FP32_TO_FP16(
+                d * (wasm_i32x4_extract_lane(accv, 0) +
+                     wasm_i32x4_extract_lane(accv, 1) +
+                     wasm_i32x4_extract_lane(accv, 2) +
+                     wasm_i32x4_extract_lane(accv, 3)));
+    }
+#elif defined(__AVX2__) || defined(__AVX__)
+    for (int i = 0; i < nb; i++) {
+        // Load elements into 4 AVX vectors
+        __m256 v0 = _mm256_loadu_ps( x );
+        __m256 v1 = _mm256_loadu_ps( x + 8 );
+        __m256 v2 = _mm256_loadu_ps( x + 16 );
+        __m256 v3 = _mm256_loadu_ps( x + 24 );
+        x += 32;
+
+        // Compute max(abs(e)) for the block
+        const __m256 signBit = _mm256_set1_ps( -0.0f );
+        __m256 maxAbs = _mm256_andnot_ps( signBit, v0 );
+        maxAbs = _mm256_max_ps( maxAbs, _mm256_andnot_ps( signBit, v1 ) );
+        maxAbs = _mm256_max_ps( maxAbs, _mm256_andnot_ps( signBit, v2 ) );
+        maxAbs = _mm256_max_ps( maxAbs, _mm256_andnot_ps( signBit, v3 ) );
+
+        __m128 max4 = _mm_max_ps( _mm256_extractf128_ps( maxAbs, 1 ), _mm256_castps256_ps128( maxAbs ) );
+        max4 = _mm_max_ps( max4, _mm_movehl_ps( max4, max4 ) );
+        max4 = _mm_max_ss( max4, _mm_movehdup_ps( max4 ) );
+        const float max_scalar = _mm_cvtss_f32( max4 );
+
+        // Quantize these floats
+        const float d = max_scalar / 127.f;
+        y[i].d = GGML_FP32_TO_FP16(d);
+        const float id = ( max_scalar != 0.0f ) ? 127.f / max_scalar : 0.0f;
+        const __m256 mul = _mm256_set1_ps( id );
+
+        // Apply the multiplier
+        v0 = _mm256_mul_ps( v0, mul );
+        v1 = _mm256_mul_ps( v1, mul );
+        v2 = _mm256_mul_ps( v2, mul );
+        v3 = _mm256_mul_ps( v3, mul );
+
+        // Round to nearest integer
+        v0 = _mm256_round_ps( v0, _MM_ROUND_NEAREST );
+        v1 = _mm256_round_ps( v1, _MM_ROUND_NEAREST );
+        v2 = _mm256_round_ps( v2, _MM_ROUND_NEAREST );
+        v3 = _mm256_round_ps( v3, _MM_ROUND_NEAREST );
+
+        // Convert floats to integers
+        __m256i i0 = _mm256_cvtps_epi32( v0 );
+        __m256i i1 = _mm256_cvtps_epi32( v1 );
+        __m256i i2 = _mm256_cvtps_epi32( v2 );
+        __m256i i3 = _mm256_cvtps_epi32( v3 );
+
+#if defined(__AVX2__)
+        // Compute the sum of the quants and set y[i].s
+        y[i].s = GGML_FP32_TO_FP16(d * hsum_i32_8(_mm256_add_epi32(_mm256_add_epi32(i0, i1), _mm256_add_epi32(i2, i3))));
+
+        // Convert int32 to int16
+        i0 = _mm256_packs_epi32( i0, i1 );	// 0, 1, 2, 3,  8, 9, 10, 11,  4, 5, 6, 7, 12, 13, 14, 15
+        i2 = _mm256_packs_epi32( i2, i3 );	// 16, 17, 18, 19,  24, 25, 26, 27,  20, 21, 22, 23, 28, 29, 30, 31
+                                            // Convert int16 to int8
+        i0 = _mm256_packs_epi16( i0, i2 );	// 0, 1, 2, 3,  8, 9, 10, 11,  16, 17, 18, 19,  24, 25, 26, 27,  4, 5, 6, 7, 12, 13, 14, 15, 20, 21, 22, 23, 28, 29, 30, 31
+
+        // We got our precious signed bytes, but the order is now wrong
+        // These AVX2 pack instructions process 16-byte pieces independently
+        // The following instruction is fixing the order
+        const __m256i perm = _mm256_setr_epi32( 0, 4, 1, 5, 2, 6, 3, 7 );
+        i0 = _mm256_permutevar8x32_epi32( i0, perm );
+
+        _mm256_storeu_si256((__m256i *)y[i].qs, i0);
+#else
+        // Since we don't have in AVX some necessary functions,
+        // we split the registers in half and call AVX2 analogs from SSE
+        __m128i ni0 = _mm256_castsi256_si128( i0 );
+        __m128i ni1 = _mm256_extractf128_si256( i0, 1);
+        __m128i ni2 = _mm256_castsi256_si128( i1 );
+        __m128i ni3 = _mm256_extractf128_si256( i1, 1);
+        __m128i ni4 = _mm256_castsi256_si128( i2 );
+        __m128i ni5 = _mm256_extractf128_si256( i2, 1);
+        __m128i ni6 = _mm256_castsi256_si128( i3 );
+        __m128i ni7 = _mm256_extractf128_si256( i3, 1);
+
+        // Compute the sum of the quants and set y[i].s
+        const __m128i s0 = _mm_add_epi32(_mm_add_epi32(ni0, ni1), _mm_add_epi32(ni2, ni3));
+        const __m128i s1 = _mm_add_epi32(_mm_add_epi32(ni4, ni5), _mm_add_epi32(ni6, ni7));
+        y[i].s = GGML_FP32_TO_FP16(d * hsum_i32_4(_mm_add_epi32(s0, s1)));
+
+        // Convert int32 to int16
+        ni0 = _mm_packs_epi32( ni0, ni1 );
+        ni2 = _mm_packs_epi32( ni2, ni3 );
+        ni4 = _mm_packs_epi32( ni4, ni5 );
+        ni6 = _mm_packs_epi32( ni6, ni7 );
+        // Convert int16 to int8
+        ni0 = _mm_packs_epi16( ni0, ni2 );
+        ni4 = _mm_packs_epi16( ni4, ni6 );
+
+        _mm_storeu_si128((__m128i *)(y[i].qs +  0), ni0);
+        _mm_storeu_si128((__m128i *)(y[i].qs + 16), ni4);
+#endif
+    }
+#elif defined(__riscv_v_intrinsic)
+
+    size_t vl = __riscv_vsetvl_e32m4(QK8_1);
+
+    for (int i = 0; i < nb; i++) {
+        // load elements
+        vfloat32m4_t v_x   = __riscv_vle32_v_f32m4(x+i*QK8_1, vl);
+
+        vfloat32m4_t vfabs = __riscv_vfabs_v_f32m4(v_x, vl);
+        vfloat32m1_t tmp   = __riscv_vfmv_v_f_f32m1(0.0, vl);
+        vfloat32m1_t vmax  = __riscv_vfredmax_vs_f32m4_f32m1(vfabs, tmp, vl);
+        float amax = __riscv_vfmv_f_s_f32m1_f32(vmax);
+
+        const float d  = amax / ((1 << 7) - 1);
+        const float id = d ? 1.0f/d : 0.0f;
+
+        y[i].d = GGML_FP32_TO_FP16(d);
+
+        vfloat32m4_t x0 = __riscv_vfmul_vf_f32m4(v_x, id, vl);
+
+        // convert to integer
+        vint16m2_t   vi = __riscv_vfncvt_x_f_w_i16m2(x0, vl);
+        vint8m1_t    vs = __riscv_vncvt_x_x_w_i8m1(vi, vl);
+
+        // store result
+        __riscv_vse8_v_i8m1(y[i].qs , vs, vl);
+
+        // compute sum for y[i].s
+        vint16m1_t tmp2 = __riscv_vmv_v_x_i16m1(0, vl);
+        vint16m1_t vwrs = __riscv_vwredsum_vs_i8m1_i16m1(vs, tmp2, vl);
+
+        // set y[i].s
+        int sum = __riscv_vmv_x_s_i16m1_i16(vwrs);
+        y[i].s = GGML_FP32_TO_FP16(sum*d);
+    }
+
+#elif defined(__POWER9_VECTOR__)
+    for (int i = 0; i < nb; i++) {
+        vector float srcv [8];
+        vector float asrcv[8];
+        vector float amaxv[8];
+        vector signed int vi[8];
+
+        for (int j = 0; j < 8; j++) srcv[j]  = vec_xl(0, x + i*32 + 4*j);
+        for (int j = 0; j < 8; j++) asrcv[j] = vec_abs(srcv[j]);
+
+        for (int j = 0; j < 4; j++) amaxv[2*j] = vec_max(asrcv[2*j], asrcv[2*j+1]);
+        for (int j = 0; j < 2; j++) amaxv[4*j] = vec_max(amaxv[4*j], amaxv[4*j+2]);
+        for (int j = 0; j < 1; j++) amaxv[8*j] = vec_max(amaxv[8*j], amaxv[8*j+4]);
+
+        const float amax = MAX(MAX(vec_extract(amaxv[0], 0),
+                                   vec_extract(amaxv[0], 1)),
+                               MAX(vec_extract(amaxv[0], 2),
+                                   vec_extract(amaxv[0], 3)));
+
+        const float d = amax / ((1 << 7) - 1);
+        const float id = d ? 1.0f/d : 0.0f;
+        const vector float vid = vec_splats(id);
+
+        y[i].d = GGML_FP32_TO_FP16(d);
+
+        vector int accv = vec_splats(0);
+
+        for (int j = 0; j < 8; j++) {
+            const vector float v  = vec_round(vec_mul(srcv[j], vid));
+            vi[j] = vec_cts(v, 0);
+
+            accv = vec_add(accv, vi[j]);
+        }
+        vec_xst(vec_pack(vec_pack(vi[0], vi[1]), vec_pack(vi[2], vi[3])),  0, &y[i].qs[0]);
+        vec_xst(vec_pack(vec_pack(vi[4], vi[5]), vec_pack(vi[6], vi[7])), 16, &y[i].qs[0]);
+
+        accv = vec_add(accv, vec_sld(accv, accv, 4));
+        accv = vec_add(accv, vec_sld(accv, accv, 8));
+        y[i].s = GGML_FP32_TO_FP16(d * vec_extract(accv, 0));
+    }
+
+#elif defined(__loongarch_asx)
+    for (int i = 0; i < nb; i++) {
+        ft_union ft;
+        __m256 v0 = (__m256)__lasx_xvld( x , 0 );
+        __m256 v1 = (__m256)__lasx_xvld( x , 32 );
+        __m256 v2 = (__m256)__lasx_xvld( x , 64 );
+        __m256 v3 = (__m256)__lasx_xvld( x , 96 );
+        x += 32;
+
+        // Compute max(abs(e)) for the block
+        const __m256 sign_bit = __lasx_xvreplfr2vr_s( -0.0f );
+        __m256 max_abs = (__m256)__lasx_xvandn_v( (__m256i)sign_bit, (__m256i)v0 );
+        max_abs = __lasx_xvfmax_s( max_abs, (__m256)__lasx_xvandn_v( (__m256i)sign_bit, (__m256i)v1 ) );
+        max_abs = __lasx_xvfmax_s( max_abs, (__m256)__lasx_xvandn_v( (__m256i)sign_bit, (__m256i)v2 ) );
+        max_abs = __lasx_xvfmax_s( max_abs, (__m256)__lasx_xvandn_v( (__m256i)sign_bit, (__m256i)v3 ) );
+
+        __m128 max4 = __lsx_vfmax_s( lasx_extractf128( max_abs, 1 ), lasx_extractf128( max_abs, 0) );
+        max4 = __lsx_vfmax_s( max4, (__m128)__lsx_vpickod_d((__m128i) max4, (__m128i)max4 ) );
+        __m128 tmp = max4;
+        max4 = __lsx_vfmax_s( max4, (__m128)__lsx_vextrins_w((__m128i)tmp, (__m128i)max4, 0x10 ));
+        ft.i = __lsx_vpickve2gr_w( (__m128i)max4, 0 );
+        const float max_scalar = ft.f;
+
+        // Quantize these floats
+        const float d = max_scalar / 127.f;
+        y[i].d = GGML_FP32_TO_FP16(d);
+        const float id = ( max_scalar != 0.0f ) ? 127.f / max_scalar : 0.0f;
+        const __m256 mul = __lasx_xvreplfr2vr_s( id );
+
+        // Apply the multiplier
+        v0 = __lasx_xvfmul_s( v0, mul );
+        v1 = __lasx_xvfmul_s( v1, mul );
+        v2 = __lasx_xvfmul_s( v2, mul );
+        v3 = __lasx_xvfmul_s( v3, mul );
+
+        // Round to nearest integer
+        __m256i i0 = __lasx_xvftintrne_w_s( v0 );
+        __m256i i1 = __lasx_xvftintrne_w_s( v1 );
+        __m256i i2 = __lasx_xvftintrne_w_s( v2 );
+        __m256i i3 = __lasx_xvftintrne_w_s( v3 );
+
+        __m128i ni0 = lasx_extracti128(i0, 0);
+        __m128i ni1 = lasx_extracti128( i0, 1);
+        __m128i ni2 = lasx_extracti128( i1, 0);
+        __m128i ni3 = lasx_extracti128( i1, 1);
+        __m128i ni4 = lasx_extracti128( i2, 0 );
+        __m128i ni5 = lasx_extracti128( i2, 1);
+        __m128i ni6 = lasx_extracti128( i3, 0);
+        __m128i ni7 = lasx_extracti128( i3, 1);
+
+        // Compute the sum of the quants and set y[i].s
+        const __m128i s0 = __lsx_vadd_w(__lsx_vadd_w(ni0, ni1), __lsx_vadd_w(ni2, ni3));
+        const __m128i s1 = __lsx_vadd_w(__lsx_vadd_w(ni4, ni5), __lsx_vadd_w(ni6, ni7));
+        y[i].s = GGML_FP32_TO_FP16(d * hsum_i32_4(__lsx_vadd_w(s0, s1)));
+
+        // Convert int32 to int16
+        ni0 = lsx_packs_w( ni0, ni1 );
+        ni2 = lsx_packs_w( ni2, ni3 );
+        ni4 = lsx_packs_w( ni4, ni5 );
+        ni6 = lsx_packs_w( ni6, ni7 );
+        // Convert int16 to int8
+        ni0 = lsx_packs_h( ni0, ni2 );
+        ni4 = lsx_packs_h( ni4, ni6 );
+
+        __lsx_vst(ni0, (__m128i *)(y[i].qs +  0), 0);
+        __lsx_vst(ni4, (__m128i *)(y[i].qs + 16), 0);
+    }
+#else
+    GGML_UNUSED(nb);
+    // scalar
+    quantize_row_q8_1_ref(x, y, k);
+#endif
+}
+
+void dequantize_row_q4_0(const block_q4_0 * restrict x, float * restrict y, int64_t k) {
+    static const int qk = QK4_0;
+
+    assert(k % qk == 0);
+
+    const int nb = k / qk;
+
+    for (int i = 0; i < nb; i++) {
+        const float d = GGML_FP16_TO_FP32(x[i].d);
+
+        for (int j = 0; j < qk/2; ++j) {
+            const int x0 = (x[i].qs[j] & 0x0F) - 8;
+            const int x1 = (x[i].qs[j] >>   4) - 8;
+
+            y[i*qk + j + 0   ] = x0*d;
+            y[i*qk + j + qk/2] = x1*d;
+        }
+    }
+}
+
+void dequantize_row_q4_1(const block_q4_1 * restrict x, float * restrict y, int64_t k) {
+    static const int qk = QK4_1;
+
+    assert(k % qk == 0);
+
+    const int nb = k / qk;
+
+    for (int i = 0; i < nb; i++) {
+        const float d = GGML_FP16_TO_FP32(x[i].d);
+        const float m = GGML_FP16_TO_FP32(x[i].m);
+
+        for (int j = 0; j < qk/2; ++j) {
+            const int x0 = (x[i].qs[j] & 0x0F);
+            const int x1 = (x[i].qs[j] >>   4);
+
+            y[i*qk + j + 0   ] = x0*d + m;
+            y[i*qk + j + qk/2] = x1*d + m;
+        }
+    }
+}
+
+void dequantize_row_q5_0(const block_q5_0 * restrict x, float * restrict y, int64_t k) {
+    static const int qk = QK5_0;
+
+    assert(k % qk == 0);
+
+    const int nb = k / qk;
+
+    for (int i = 0; i < nb; i++) {
+        const float d = GGML_FP16_TO_FP32(x[i].d);
+
+        uint32_t qh;
+        memcpy(&qh, x[i].qh, sizeof(qh));
+
+        for (int j = 0; j < qk/2; ++j) {
+            const uint8_t xh_0 = ((qh >> (j +  0)) << 4) & 0x10;
+            const uint8_t xh_1 = ((qh >> (j + 12))     ) & 0x10;
+
+            const int32_t x0 = ((x[i].qs[j] & 0x0F) | xh_0) - 16;
+            const int32_t x1 = ((x[i].qs[j] >>   4) | xh_1) - 16;
+
+            y[i*qk + j + 0   ] = x0*d;
+            y[i*qk + j + qk/2] = x1*d;
+        }
+    }
+}
+
+void dequantize_row_q5_1(const block_q5_1 * restrict x, float * restrict y, int64_t k) {
+    static const int qk = QK5_1;
+
+    assert(k % qk == 0);
+
+    const int nb = k / qk;
+
+    for (int i = 0; i < nb; i++) {
+        const float d = GGML_FP16_TO_FP32(x[i].d);
+        const float m = GGML_FP16_TO_FP32(x[i].m);
+
+        uint32_t qh;
+        memcpy(&qh, x[i].qh, sizeof(qh));
+
+        for (int j = 0; j < qk/2; ++j) {
+            const uint8_t xh_0 = ((qh >> (j +  0)) << 4) & 0x10;
+            const uint8_t xh_1 = ((qh >> (j + 12))     ) & 0x10;
+
+            const int x0 = (x[i].qs[j] & 0x0F) | xh_0;
+            const int x1 = (x[i].qs[j] >>   4) | xh_1;
+
+            y[i*qk + j + 0   ] = x0*d + m;
+            y[i*qk + j + qk/2] = x1*d + m;
+        }
+    }
+}
+
+void dequantize_row_q8_0(const block_q8_0 * restrict x, float * restrict y, int64_t k) {
+    static const int qk = QK8_0;
+
+    assert(k % qk == 0);
+
+    const int nb = k / qk;
+
+    for (int i = 0; i < nb; i++) {
+        const float d = GGML_FP16_TO_FP32(x[i].d);
+
+        for (int j = 0; j < qk; ++j) {
+            y[i*qk + j] = x[i].qs[j]*d;
+        }
+    }
+}
+
+//
+// 2-6 bit quantization in super-blocks
+//
+
+//
+// ===================== Helper functions
+//
+static inline int nearest_int(float fval) {
+    assert(fabsf(fval) <= 4194303.f);
+    float val = fval + 12582912.f;
+    int i; memcpy(&i, &val, sizeof(int));
+    return (i & 0x007fffff) - 0x00400000;
+}
+
+static float make_qx_quants(int n, int nmax, const float * restrict x, int8_t * restrict L, int rmse_type,
+        const float * restrict qw) {
+    float max = 0;
+    float amax = 0;
+    for (int i = 0; i < n; ++i) {
+        float ax = fabsf(x[i]);
+        if (ax > amax) { amax = ax; max = x[i]; }
+    }
+    if (amax < GROUP_MAX_EPS) { // all zero
+        for (int i = 0; i < n; ++i) {
+            L[i] = 0;
+        }
+        return 0.f;
+    }
+    float iscale = -nmax / max;
+    if (rmse_type == 0) {
+        for (int i = 0; i < n; ++i) {
+            int l = nearest_int(iscale * x[i]);
+            L[i] = nmax + MAX(-nmax, MIN(nmax-1, l));
+        }
+        return 1/iscale;
+    }
+    bool return_early = false;
+    if (rmse_type < 0) {
+        rmse_type = -rmse_type;
+        return_early = true;
+    }
+    float sumlx = 0;
+    float suml2 = 0;
+#ifdef HAVE_BUGGY_APPLE_LINKER
+    // use 'volatile' to prevent unroll and work around a bug in Apple ld64 1015.7
+    for (volatile int i = 0; i < n; ++i) {
+#else
+    for (int i = 0; i < n; ++i) {
+#endif
+        int l = nearest_int(iscale * x[i]);
+        l = MAX(-nmax, MIN(nmax-1, l));
+        L[i] = l + nmax;
+        float w = qw ? qw[i] : rmse_type == 1 ? x[i] * x[i] : rmse_type == 2 ? 1 : rmse_type == 3 ? fabsf(x[i]) : sqrtf(fabsf(x[i]));
+        sumlx += w*x[i]*l;
+        suml2 += w*l*l;
+    }
+    float scale = suml2 ? sumlx/suml2 : 0.0f;
+    if (return_early) return suml2 > 0 ? 0.5f*(scale + 1/iscale) : 1/iscale;
+    float best = scale * sumlx;
+    for (int is = -9; is <= 9; ++is) {
+        if (is == 0) {
+            continue;
+        }
+        iscale = -(nmax + 0.1f*is) / max;
+        sumlx = suml2 = 0;
+        for (int i = 0; i < n; ++i) {
+            int l = nearest_int(iscale * x[i]);
+            l = MAX(-nmax, MIN(nmax-1, l));
+            float w = qw ? qw[i] : rmse_type == 1 ? x[i] * x[i] : rmse_type == 2 ? 1 : rmse_type == 3 ? fabsf(x[i]) : sqrtf(fabsf(x[i]));
+            sumlx += w*x[i]*l;
+            suml2 += w*l*l;
+        }
+        if (suml2 > 0 && sumlx*sumlx > best*suml2) {
+            for (int i = 0; i < n; ++i) {
+                int l = nearest_int(iscale * x[i]);
+                L[i] = nmax + MAX(-nmax, MIN(nmax-1, l));
+            }
+            scale = sumlx/suml2; best = scale*sumlx;
+        }
+    }
+    return scale;
+}
+
+static float make_q3_quants(int n, int nmax, const float * restrict x, int8_t * restrict L, bool do_rmse) {
+    float max = 0;
+    float amax = 0;
+    for (int i = 0; i < n; ++i) {
+        float ax = fabsf(x[i]);
+        if (ax > amax) { amax = ax; max = x[i]; }
+    }
+    if (amax < GROUP_MAX_EPS) { // all zero
+        for (int i = 0; i < n; ++i) { L[i] = 0; }
+        return 0.f;
+    }
+    float iscale = -nmax / max;
+    if (do_rmse) {
+        float sumlx = 0;
+        float suml2 = 0;
+        for (int i = 0; i < n; ++i) {
+            int l = nearest_int(iscale * x[i]);
+            l = MAX(-nmax, MIN(nmax-1, l));
+            L[i] = l;
+            float w = x[i]*x[i];
+            sumlx += w*x[i]*l;
+            suml2 += w*l*l;
+        }
+        for (int itry = 0; itry < 5; ++itry) {
+            int n_changed = 0;
+            for (int i = 0; i < n; ++i) {
+                float w = x[i]*x[i];
+                float slx = sumlx - w*x[i]*L[i];
+                if (slx > 0) {
+                    float sl2 = suml2 - w*L[i]*L[i];
+                    int new_l = nearest_int(x[i] * sl2 / slx);
+                    new_l = MAX(-nmax, MIN(nmax-1, new_l));
+                    if (new_l != L[i]) {
+                        slx += w*x[i]*new_l;
+                        sl2 += w*new_l*new_l;
+                        if (sl2 > 0 && slx*slx*suml2 > sumlx*sumlx*sl2) {
+                            L[i] = new_l; sumlx = slx; suml2 = sl2;
+                            ++n_changed;
+                        }
+                    }
+                }
+            }
+            if (!n_changed) {
+                break;
+            }
+        }
+        for (int i = 0; i < n; ++i) {
+            L[i] += nmax;
+        }
+        return sumlx / suml2;
+    }
+    for (int i = 0; i < n; ++i) {
+        int l = nearest_int(iscale * x[i]);
+        l = MAX(-nmax, MIN(nmax-1, l));
+        L[i] = l + nmax;
+    }
+    return 1/iscale;
+}
+
+static float make_qkx1_quants(int n, int nmax, const float * restrict x, uint8_t * restrict L, float * restrict the_min,
+        int ntry, float alpha) {
+    float min = x[0];
+    float max = x[0];
+    for (int i = 1; i < n; ++i) {
+        if (x[i] < min) min = x[i];
+        if (x[i] > max) max = x[i];
+    }
+    if (max == min) {
+        for (int i = 0; i < n; ++i) L[i] = 0;
+        *the_min = 0;
+        return 0.f;
+    }
+    if (min > 0) min = 0;
+    float iscale = nmax/(max - min);
+    float scale = 1/iscale;
+    for (int itry = 0; itry < ntry; ++itry) {
+        float sumlx = 0; int suml2 = 0;
+        bool did_change = false;
+        for (int i = 0; i < n; ++i) {
+            int l = nearest_int(iscale*(x[i] - min));
+            l = MAX(0, MIN(nmax, l));
+            if (l != L[i]) {
+                L[i] = l;
+                did_change = true;
+            }
+            sumlx += (x[i] - min)*l;
+            suml2 += l*l;
+        }
+        scale = sumlx/suml2;
+        float sum = 0;
+        for (int i = 0; i < n; ++i) {
+            sum += x[i] - scale*L[i];
+        }
+        min = alpha*min + (1 - alpha)*sum/n;
+        if (min > 0) min = 0;
+        iscale = 1/scale;
+        if (!did_change) break;
+    }
+    *the_min = -min;
+    return scale;
+}
+
+static float make_qkx2_quants(int n, int nmax, const float * restrict x, const float * restrict weights,
+        uint8_t * restrict L, float * restrict the_min, uint8_t * restrict Laux,
+        float rmin, float rdelta, int nstep, bool use_mad) {
+    float min = x[0];
+    float max = x[0];
+    float sum_w = weights[0];
+    float sum_x = sum_w * x[0];
+#ifdef HAVE_BUGGY_APPLE_LINKER
+    // use 'volatile' to prevent unroll and work around a bug in Apple ld64 1015.7
+    for (volatile int i = 1; i < n; ++i) {
+#else
+    for (int i = 1; i < n; ++i) {
+#endif
+        if (x[i] < min) min = x[i];
+        if (x[i] > max) max = x[i];
+        float w = weights[i];
+        sum_w += w;
+        sum_x += w * x[i];
+    }
+    if (min > 0) min = 0;
+    if (max == min) {
+        for (int i = 0; i < n; ++i) L[i] = 0;
+        *the_min = -min;
+        return 0.f;
+    }
+    float iscale = nmax/(max - min);
+    float scale = 1/iscale;
+    float best_mad = 0;
+    for (int i = 0; i < n; ++i) {
+        int l = nearest_int(iscale*(x[i] - min));
+        L[i] = MAX(0, MIN(nmax, l));
+        float diff = scale * L[i] + min - x[i];
+        diff = use_mad ? fabsf(diff) : diff * diff;
+        float w = weights[i];
+        best_mad += w * diff;
+    }
+    if (nstep < 1) {
+        *the_min = -min;
+        return scale;
+    }
+    for (int is = 0; is <= nstep; ++is) {
+        iscale = (rmin + rdelta*is + nmax)/(max - min);
+        float sum_l = 0, sum_l2 = 0, sum_xl = 0;
+        for (int i = 0; i < n; ++i) {
+            int l = nearest_int(iscale*(x[i] - min));
+            l = MAX(0, MIN(nmax, l));
+            Laux[i] = l;
+            float w = weights[i];
+            sum_l += w*l;
+            sum_l2 += w*l*l;
+            sum_xl += w*l*x[i];
+        }
+        float D = sum_w * sum_l2 - sum_l * sum_l;
+        if (D > 0) {
+            float this_scale = (sum_w * sum_xl - sum_x * sum_l)/D;
+            float this_min   = (sum_l2 * sum_x - sum_l * sum_xl)/D;
+            if (this_min > 0) {
+                this_min = 0;
+                this_scale = sum_xl / sum_l2;
+            }
+            float mad = 0;
+            for (int i = 0; i < n; ++i) {
+                float diff = this_scale * Laux[i] + this_min - x[i];
+                diff = use_mad ? fabsf(diff) : diff * diff;
+                float w = weights[i];
+                mad += w * diff;
+            }
+            if (mad < best_mad) {
+                for (int i = 0; i < n; ++i) {
+                    L[i] = Laux[i];
+                }
+                best_mad = mad;
+                scale = this_scale;
+                min = this_min;
+            }
+        }
+    }
+    *the_min = -min;
+    return scale;
+}
+
+static inline void get_scale_min_k4(int j, const uint8_t * restrict q, uint8_t * restrict d, uint8_t * restrict m) {
+    if (j < 4) {
+        *d = q[j] & 63; *m = q[j + 4] & 63;
+    } else {
+        *d = (q[j+4] & 0xF) | ((q[j-4] >> 6) << 4);
+        *m = (q[j+4] >>  4) | ((q[j-0] >> 6) << 4);
+    }
+}
+
+//========================- 2-bit (de)-quantization
+
+void quantize_row_q2_K_ref(const float * restrict x, block_q2_K * restrict y, int64_t k) {
+    assert(k % QK_K == 0);
+    const int nb = k / QK_K;
+
+    uint8_t L[QK_K];
+    uint8_t Laux[16];
+    float   weights[16];
+    float mins[QK_K/16];
+    float scales[QK_K/16];
+
+    const float q4scale = 15.f;
+
+    for (int i = 0; i < nb; i++) {
+        float max_scale = 0; // as we are deducting the min, scales are always positive
+        float max_min = 0;
+        for (int j = 0; j < QK_K/16; ++j) {
+            for (int l = 0; l < 16; ++l) weights[l] = fabsf(x[16*j + l]);
+            scales[j] = make_qkx2_quants(16, 3, x + 16*j, weights, L + 16*j, &mins[j], Laux, -0.5f, 0.1f, 15, true);
+            float scale = scales[j];
+            if (scale > max_scale) {
+                max_scale = scale;
+            }
+            float min = mins[j];
+            if (min > max_min) {
+                max_min = min;
+            }
+        }
+
+        if (max_scale > 0) {
+            float iscale = q4scale/max_scale;
+            for (int j = 0; j < QK_K/16; ++j) {
+                int l = nearest_int(iscale*scales[j]);
+                y[i].scales[j] = l;
+            }
+            y[i].d = GGML_FP32_TO_FP16(max_scale/q4scale);
+        } else {
+            for (int j = 0; j < QK_K/16; ++j) y[i].scales[j] = 0;
+            y[i].d = GGML_FP32_TO_FP16(0.f);
+        }
+        if (max_min > 0) {
+            float iscale = q4scale/max_min;
+            for (int j = 0; j < QK_K/16; ++j) {
+                int l = nearest_int(iscale*mins[j]);
+                y[i].scales[j] |= (l << 4);
+            }
+            y[i].dmin = GGML_FP32_TO_FP16(max_min/q4scale);
+        } else {
+            y[i].dmin = GGML_FP32_TO_FP16(0.f);
+        }
+        for (int j = 0; j < QK_K/16; ++j) {
+            const float d = GGML_FP16_TO_FP32(y[i].d) * (y[i].scales[j] & 0xF);
+            if (!d) continue;
+            const float dm = GGML_FP16_TO_FP32(y[i].dmin) * (y[i].scales[j] >> 4);
+            for (int ii = 0; ii < 16; ++ii) {
+                int l = nearest_int((x[16*j + ii] + dm)/d);
+                l = MAX(0, MIN(3, l));
+                L[16*j + ii] = l;
+            }
+        }
+
+        for (int j = 0; j < QK_K; j += 128) {
+            for (int l = 0; l < 32; ++l) {
+                y[i].qs[j/4 + l] = L[j + l] | (L[j + l + 32] << 2) | (L[j + l + 64] << 4) | (L[j + l + 96] << 6);
+            }
+        }
+
+        x += QK_K;
+    }
+}
+
+void dequantize_row_q2_K(const block_q2_K * restrict x, float * restrict y, int64_t k) {
+    assert(k % QK_K == 0);
+    const int nb = k / QK_K;
+
+    for (int i = 0; i < nb; i++) {
+
+        const float d = GGML_FP16_TO_FP32(x[i].d);
+        const float min = GGML_FP16_TO_FP32(x[i].dmin);
+
+        const uint8_t * q = x[i].qs;
+
+        int is = 0;
+        float dl, ml;
+        for (int n = 0; n < QK_K; n += 128) {
+            int shift = 0;
+            for (int j = 0; j < 4; ++j) {
+
+                uint8_t sc = x[i].scales[is++];
+                dl = d * (sc & 0xF); ml = min * (sc >> 4);
+                for (int l = 0; l < 16; ++l) *y++ = dl * ((int8_t)((q[l] >> shift) & 3)) - ml;
+
+                sc = x[i].scales[is++];
+                dl = d * (sc & 0xF); ml = min * (sc >> 4);
+                for (int l = 0; l < 16; ++l) *y++ = dl * ((int8_t)((q[l+16] >> shift) & 3)) - ml;
+
+                shift += 2;
+            }
+            q += 32;
+        }
+    }
+}
+
+void quantize_row_q2_K(const float * restrict x, void * restrict vy, int64_t k) {
+    quantize_row_q2_K_ref(x, vy, k);
+}
+
+static float make_qkx3_quants(int n, int nmax, const float * restrict x, const float * restrict weights,
+        uint8_t * restrict L, float * restrict the_min, uint8_t * restrict Laux,
+        float rmin, float rdelta, int nstep, bool use_mad) {
+    float min = x[0];
+    float max = x[0];
+    float sum_w = weights ? weights[0] : x[0]*x[0];
+    float sum_x = sum_w * x[0];
+#ifdef HAVE_BUGGY_APPLE_LINKER
+    // use 'volatile' to prevent unroll and work around a bug in Apple ld64 1015.7
+    for (volatile int i = 1; i < n; ++i) {
+#else
+    for (int i = 1; i < n; ++i) {
+#endif
+        if (x[i] < min) min = x[i];
+        if (x[i] > max) max = x[i];
+        float w = weights ? weights[i] : x[i]*x[i];
+        sum_w += w;
+        sum_x += w * x[i];
+    }
+    if (min > 0) {
+        min = 0;
+    }
+    if (max <= min) {
+        memset(L, 0, n);
+        *the_min = -min;
+        return 0.f;
+    }
+    float iscale = nmax/(max - min);
+    float scale = 1/iscale;
+    float best_mad = 0;
+    for (int i = 0; i < n; ++i) {
+        int l = nearest_int(iscale*(x[i] - min));
+        L[i] = MAX(0, MIN(nmax, l));
+        float diff = scale * L[i] + min - x[i];
+        diff = use_mad ? fabsf(diff) : diff*diff;
+        float w = weights ? weights[i] : x[i]*x[i];
+        best_mad += w * diff;
+    }
+    if (nstep < 1) {
+        *the_min = -min;
+        return scale;
+    }
+    for (int is = 0; is <= nstep; ++is) {
+        iscale = (rmin + rdelta*is + nmax)/(max - min);
+        float sum_l = 0, sum_l2 = 0, sum_xl = 0;
+        for (int i = 0; i < n; ++i) {
+            int l = nearest_int(iscale*(x[i] - min));
+            l = MAX(0, MIN(nmax, l));
+            Laux[i] = l;
+            float w = weights ? weights[i] : x[i]*x[i];
+            sum_l  += w*l;
+            sum_l2 += w*l*l;
+            sum_xl += w*l*x[i];
+        }
+        float D = sum_w * sum_l2 - sum_l * sum_l;
+        if (D > 0) {
+            float this_scale = (sum_w * sum_xl - sum_x * sum_l)/D;
+            float this_min   = (sum_l2 * sum_x - sum_l * sum_xl)/D;
+            if (this_min > 0) {
+                this_min = 0;
+                this_scale = sum_xl / sum_l2;
+            }
+            float mad = 0;
+            for (int i = 0; i < n; ++i) {
+                float diff = this_scale * Laux[i] + this_min - x[i];
+                diff = use_mad ? fabsf(diff) : diff*diff;
+                float w = weights ? weights[i] : x[i]*x[i];
+                mad += w * diff;
+            }
+            if (mad < best_mad) {
+                for (int i = 0; i < n; ++i) {
+                    L[i] = Laux[i];
+                }
+                best_mad = mad;
+                scale = this_scale;
+                min = this_min;
+            }
+        }
+    }
+    *the_min = -min;
+    return scale;
+}
+
+static float make_qp_quants(int n, int nmax, const float * restrict x, uint8_t * restrict L, const float * quant_weights) {
+    float max = 0;
+    for (int i = 0; i < n; ++i) {
+        max = MAX(max, x[i]);
+    }
+    if (!max) { // all zero
+        for (int i = 0; i < n; ++i) { L[i] = 0; }
+        return 0.f;
+    }
+    float iscale = nmax / max;
+    for (int i = 0; i < n; ++i) {
+        L[i] = nearest_int(iscale * x[i]);
+    }
+    float scale = 1/iscale;
+    float best_mse = 0;
+    for (int i = 0; i < n; ++i) {
+        float diff = x[i] - scale*L[i];
+        float w = quant_weights[i];
+        best_mse += w*diff*diff;
+    }
+    for (int is = -4; is <= 4; ++is) {
+        if (is == 0) continue;
+        float iscale_is = (0.1f*is + nmax)/max;
+        float scale_is = 1/iscale_is;
+        float mse = 0;
+        for (int i = 0; i < n; ++i) {
+            int l = nearest_int(iscale_is*x[i]);
+            l = MIN(nmax, l);
+            float diff = x[i] - scale_is*l;
+            float w = quant_weights[i];
+            mse += w*diff*diff;
+        }
+        if (mse < best_mse) {
+            best_mse = mse;
+            iscale = iscale_is;
+        }
+    }
+    float sumlx = 0;
+    float suml2 = 0;
+    for (int i = 0; i < n; ++i) {
+        int l = nearest_int(iscale * x[i]);
+        l = MIN(nmax, l);
+        L[i] = l;
+        float w = quant_weights[i];
+        sumlx += w*x[i]*l;
+        suml2 += w*l*l;
+    }
+    for (int itry = 0; itry < 5; ++itry) {
+        int n_changed = 0;
+        for (int i = 0; i < n; ++i) {
+            float w = quant_weights[i];
+            float slx = sumlx - w*x[i]*L[i];
+            float sl2 = suml2 - w*L[i]*L[i];
+            if (slx > 0 && sl2 > 0) {
+                int new_l = nearest_int(x[i] * sl2 / slx);
+                new_l = MIN(nmax, new_l);
+                if (new_l != L[i]) {
+                    slx += w*x[i]*new_l;
+                    sl2 += w*new_l*new_l;
+                    if (slx*slx*suml2 > sumlx*sumlx*sl2) {
+                        L[i] = new_l; sumlx = slx; suml2 = sl2;
+                        ++n_changed;
+                    }
+                }
+            }
+        }
+        if (!n_changed) {
+            break;
+        }
+    }
+    return sumlx/suml2;
+}
+
+static void quantize_row_q2_K_impl(const float * restrict x, block_q2_K * restrict y, int k, const float * restrict quant_weights) {
+    GGML_ASSERT(quant_weights);
+    assert(k % QK_K == 0);
+    const int nb = k / QK_K;
+    const bool requantize = true;
+
+    uint8_t L[QK_K];
+    uint8_t Laux[16];
+    float mins[QK_K/16];
+    float scales[QK_K/16];
+    float sw[QK_K/16];
+    float weight[16];
+    uint8_t Ls[QK_K/16], Lm[QK_K/16];
+
+    for (int i = 0; i < nb; i++) {
+        memset(sw, 0, QK_K/16*sizeof(float));
+        float sumx2 = 0;
+        for (int j = 0; j < QK_K; ++j) sumx2 += x[j]*x[j];
+        float sigma2 = sumx2/QK_K;
+        for (int j = 0; j < QK_K/16; ++j) {
+            const float * restrict qw = quant_weights + QK_K * i + 16*j;
+            for (int l = 0; l < 16; ++l) weight[l] = qw[l] * sqrtf(sigma2 + x[16*j + l]*x[16*j + l]);
+            for (int l = 0; l < QK_K/16; ++l) sw[j] += weight[l];
+            scales[j] = make_qkx3_quants(16, 3, x + 16*j, weight, L + 16*j, &mins[j], Laux, -0.9f, 0.05f, 36, false);
+        }
+
+        float dm, mm;
+        dm  = make_qp_quants(QK_K/16, 15, scales, Ls, sw);
+        mm  = make_qp_quants(QK_K/16, 15, mins,   Lm, sw);
+
+        y[i].d    = GGML_FP32_TO_FP16(dm);
+        y[i].dmin = GGML_FP32_TO_FP16(mm);
+        dm        = GGML_FP16_TO_FP32(y[i].d);
+        mm        = GGML_FP16_TO_FP32(y[i].dmin);
+
+        for (int j = 0; j < QK_K/16; ++j) {
+            y[i].scales[j] = Ls[j] | (Lm[j] << 4);
+        }
+
+        if (requantize) {
+            for (int j = 0; j < QK_K/16; ++j) {
+                const float d = dm * (y[i].scales[j] & 0xF);
+                if (!d) continue;
+                const float m = mm * (y[i].scales[j] >> 4);
+                for (int ii = 0; ii < 16; ++ii) {
+                    int l = nearest_int((x[16*j + ii] + m)/d);
+                    l = MAX(0, MIN(3, l));
+                    L[16*j + ii] = l;
+                }
+            }
+        }
+
+        for (int j = 0; j < QK_K; j += 128) {
+            for (int l = 0; l < 32; ++l) {
+                y[i].qs[j/4 + l] = L[j + l] | (L[j + l + 32] << 2) | (L[j + l + 64] << 4) | (L[j + l + 96] << 6);
+            }
+        }
+
+        x += QK_K;
+    }
+}
+
+size_t quantize_q2_K(const float * restrict src, void * restrict dst, int64_t nrow, int64_t n_per_row, const float * quant_weights) {
+    size_t row_size = ggml_row_size(GGML_TYPE_Q2_K, n_per_row);
+    if (!quant_weights) {
+        quantize_row_q2_K_ref(src, dst, (int64_t)nrow*n_per_row);
+    }
+    else {
+        char * qrow = (char *)dst;
+        for (int64_t row = 0; row < nrow; ++row) {
+            quantize_row_q2_K_impl(src, (block_q2_K*)qrow, n_per_row, quant_weights);
+            src += n_per_row;
+            qrow += row_size;
+        }
+    }
+    return nrow * row_size;
+}
+
+//========================= 3-bit (de)-quantization
+
+void quantize_row_q3_K_ref(const float * restrict x, block_q3_K * restrict y, int64_t k) {
+    assert(k % QK_K == 0);
+    const int nb = k / QK_K;
+
+    int8_t L[QK_K];
+    float scales[QK_K / 16];
+
+    for (int i = 0; i < nb; i++) {
+
+        float max_scale = 0;
+        float amax = 0;
+        for (int j = 0; j < QK_K/16; ++j) {
+            scales[j] = make_q3_quants(16, 4, x + 16*j, L + 16*j, true);
+            float scale = fabsf(scales[j]);
+            if (scale > amax) {
+                amax = scale; max_scale = scales[j];
+            }
+        }
+
+        memset(y[i].scales, 0, 12);
+        if (max_scale) {
+            float iscale = -32.f/max_scale;
+            for (int j = 0; j < QK_K/16; ++j) {
+                int8_t l = nearest_int(iscale*scales[j]);
+                l = MAX(-32, MIN(31, l)) + 32;
+                if (j < 8) {
+                    y[i].scales[j] = l & 0xF;
+                } else {
+                    y[i].scales[j-8] |= ((l & 0xF) << 4);
+                }
+                l >>= 4;
+                y[i].scales[j%4 + 8] |= (l << (2*(j/4)));
+            }
+            y[i].d = GGML_FP32_TO_FP16(1/iscale);
+        } else {
+            y[i].d = GGML_FP32_TO_FP16(0.f);
+        }
+
+        int8_t sc;
+        for (int j = 0; j < QK_K/16; ++j) {
+            sc = j < 8 ? y[i].scales[j] & 0xF : y[i].scales[j-8] >> 4;
+            sc = (sc | (((y[i].scales[8 + j%4] >> (2*(j/4))) & 3) << 4)) - 32;
+            float d = GGML_FP16_TO_FP32(y[i].d) * sc;
+            if (!d) {
+                continue;
+            }
+            for (int ii = 0; ii < 16; ++ii) {
+                int l = nearest_int(x[16*j + ii]/d);
+                l = MAX(-4, MIN(3, l));
+                L[16*j + ii] = l + 4;
+            }
+        }
+
+        memset(y[i].hmask, 0, QK_K/8);
+        // We put the high-bit for the 1st 8 quants into bit 0, the next 8 into bit 1, etc.
+        int m = 0;
+        uint8_t hm = 1;
+        for (int j = 0; j < QK_K; ++j) {
+            if (L[j] > 3) {
+                y[i].hmask[m] |= hm;
+                L[j] -= 4;
+            }
+            if (++m == QK_K/8) {
+                m = 0; hm <<= 1;
+            }
+        }
+        for (int j = 0; j < QK_K; j += 128) {
+            for (int l = 0; l < 32; ++l) {
+                y[i].qs[j/4 + l] = L[j + l] | (L[j + l + 32] << 2) | (L[j + l + 64] << 4) | (L[j + l + 96] << 6);
+            }
+        }
+
+        x += QK_K;
+    }
+}
+
+void dequantize_row_q3_K(const block_q3_K * restrict x, float * restrict y, int64_t k) {
+    assert(k % QK_K == 0);
+    const int nb = k / QK_K;
+
+    const uint32_t kmask1 = 0x03030303;
+    const uint32_t kmask2 = 0x0f0f0f0f;
+
+    uint32_t aux[4];
+    const int8_t * scales = (const int8_t*)aux;
+
+    for (int i = 0; i < nb; i++) {
+
+        const float d_all = GGML_FP16_TO_FP32(x[i].d);
+
+        const uint8_t * restrict q = x[i].qs;
+        const uint8_t * restrict hm = x[i].hmask;
+        uint8_t m = 1;
+
+        memcpy(aux, x[i].scales, 12);
+        uint32_t tmp = aux[2];
+        aux[2] = ((aux[0] >> 4) & kmask2) | (((tmp >> 4) & kmask1) << 4);
+        aux[3] = ((aux[1] >> 4) & kmask2) | (((tmp >> 6) & kmask1) << 4);
+        aux[0] = (aux[0] & kmask2) | (((tmp >> 0) & kmask1) << 4);
+        aux[1] = (aux[1] & kmask2) | (((tmp >> 2) & kmask1) << 4);
+
+        int is = 0;
+        float dl;
+        for (int n = 0; n < QK_K; n += 128) {
+            int shift = 0;
+            for (int j = 0; j < 4; ++j) {
+
+                dl = d_all * (scales[is++] - 32);
+                for (int l = 0; l < 16; ++l) {
+                    *y++ = dl * ((int8_t)((q[l+ 0] >> shift) & 3) - ((hm[l+ 0] & m) ? 0 : 4));
+                }
+
+                dl = d_all * (scales[is++] - 32);
+                for (int l = 0; l < 16; ++l) {
+                    *y++ = dl * ((int8_t)((q[l+16] >> shift) & 3) - ((hm[l+16] & m) ? 0 : 4));
+                }
+
+                shift += 2;
+                m <<= 1;
+            }
+            q += 32;
+        }
+
+    }
+}
+
+void quantize_row_q3_K(const float * restrict x, void * restrict vy, int64_t k) {
+    quantize_row_q3_K_ref(x, vy, k);
+}
+
+static void quantize_row_q3_K_impl(const float * restrict x, block_q3_K * restrict y, int64_t n_per_row, const float * restrict quant_weights) {
+    assert(n_per_row % QK_K == 0);
+    const int nb = n_per_row / QK_K;
+
+    int8_t L[QK_K];
+    float scales[QK_K / 16];
+    float weight[16];
+    float sw[QK_K / 16];
+    int8_t Ls[QK_K / 16];
+
+    for (int i = 0; i < nb; i++) {
+
+        float sumx2 = 0;
+        for (int j = 0; j < QK_K; ++j) sumx2 += x[j]*x[j];
+        float sigma2 = 2*sumx2/QK_K;
+
+        for (int j = 0; j < QK_K/16; ++j) {
+            if (quant_weights) {
+                const float * qw = quant_weights + QK_K * i + 16*j;
+                for (int l = 0; l < 16; ++l) weight[l] = qw[l] * sqrtf(sigma2 + x[16*j+l]*x[16*j+l]);
+            } else {
+                for (int l = 0; l < 16; ++l) weight[l] = x[16*j+l]*x[16*j+l];
+            }
+            float sumw = 0;
+            for (int l = 0; l < 16; ++l) sumw += weight[l];
+            sw[j] = sumw;
+
+            scales[j] = make_qx_quants(16, 4, x + 16*j, L + 16*j, 1, weight);
+
+        }
+
+        memset(y[i].scales, 0, 12);
+
+        float d_block = make_qx_quants(QK_K/16, 32, scales, Ls, 1, sw);
+        for (int j = 0; j < QK_K/16; ++j) {
+            int l = Ls[j];
+            if (j < 8) {
+                y[i].scales[j] = l & 0xF;
+            } else {
+                y[i].scales[j-8] |= ((l & 0xF) << 4);
+            }
+            l >>= 4;
+            y[i].scales[j%4 + 8] |= (l << (2*(j/4)));
+        }
+        y[i].d = GGML_FP32_TO_FP16(d_block);
+
+        int8_t sc;
+        for (int j = 0; j < QK_K/16; ++j) {
+            sc = j < 8 ? y[i].scales[j] & 0xF : y[i].scales[j-8] >> 4;
+            sc = (sc | (((y[i].scales[8 + j%4] >> (2*(j/4))) & 3) << 4)) - 32;
+            float d = GGML_FP16_TO_FP32(y[i].d) * sc;
+            if (!d) {
+                continue;
+            }
+            for (int ii = 0; ii < 16; ++ii) {
+                int l = nearest_int(x[16*j + ii]/d);
+                l = MAX(-4, MIN(3, l));
+                L[16*j + ii] = l + 4;
+            }
+        }
+
+        memset(y[i].hmask, 0, QK_K/8);
+        // We put the high-bit for the 1st 8 quants into bit 0, the next 8 into bit 1, etc.
+        int m = 0;
+        uint8_t hm = 1;
+        for (int j = 0; j < QK_K; ++j) {
+            if (L[j] > 3) {
+                y[i].hmask[m] |= hm;
+                L[j] -= 4;
+            }
+            if (++m == QK_K/8) {
+                m = 0; hm <<= 1;
+            }
+        }
+        for (int j = 0; j < QK_K; j += 128) {
+            for (int l = 0; l < 32; ++l) {
+                y[i].qs[j/4 + l] = L[j + l] | (L[j + l + 32] << 2) | (L[j + l + 64] << 4) | (L[j + l + 96] << 6);
+            }
+        }
+
+        x += QK_K;
+    }
+}
+
+size_t quantize_q3_K(const float * restrict src, void * restrict dst, int64_t nrow, int64_t n_per_row, const float * quant_weights) {
+    size_t row_size = ggml_row_size(GGML_TYPE_Q3_K, n_per_row);
+    if (!quant_weights) {
+        quantize_row_q3_K_ref(src, dst, (int64_t)nrow*n_per_row);
+    }
+    else {
+        char * qrow = (char *)dst;
+        for (int64_t row = 0; row < nrow; ++row) {
+            quantize_row_q3_K_impl(src, (block_q3_K*)qrow, n_per_row, quant_weights);
+            src += n_per_row;
+            qrow += row_size;
+        }
+    }
+    return nrow * row_size;
+}
+
+// ====================== 4-bit (de)-quantization
+
+void quantize_row_q4_K_ref(const float * restrict x, block_q4_K * restrict y, int64_t k) {
+    assert(k % QK_K == 0);
+    const int nb = k / QK_K;
+
+    uint8_t L[QK_K];
+    uint8_t Laux[32];
+    float   weights[32];
+    float mins[QK_K/32];
+    float scales[QK_K/32];
+
+    for (int i = 0; i < nb; i++) {
+        float max_scale = 0; // as we are deducting the min, scales are always positive
+        float max_min = 0;
+        for (int j = 0; j < QK_K/32; ++j) {
+            //scales[j] = make_qkx1_quants(32, 15, x + 32*j, L + 32*j, &mins[j], 9, 0.5f);
+            float sum_x2 = 0;
+            for (int l = 0; l < 32; ++l) sum_x2 += x[32*j + l] * x[32*j + l];
+            float av_x = sqrtf(sum_x2/32);
+            for (int l = 0; l < 32; ++l) weights[l] = av_x + fabsf(x[32*j + l]);
+            scales[j] = make_qkx2_quants(32, 15, x + 32*j, weights, L + 32*j, &mins[j], Laux, -1.f, 0.1f, 20, false);
+            float scale = scales[j];
+            if (scale > max_scale) {
+                max_scale = scale;
+            }
+            float min = mins[j];
+            if (min > max_min) {
+                max_min = min;
+            }
+        }
+
+        float inv_scale = max_scale > 0 ? 63.f/max_scale : 0.f;
+        float inv_min   = max_min   > 0 ? 63.f/max_min   : 0.f;
+        for (int j = 0; j < QK_K/32; ++j) {
+            uint8_t ls = nearest_int(inv_scale*scales[j]);
+            uint8_t lm = nearest_int(inv_min*mins[j]);
+            ls = MIN(63, ls);
+            lm = MIN(63, lm);
+            if (j < 4) {
+                y[i].scales[j] = ls;
+                y[i].scales[j+4] = lm;
+            } else {
+                y[i].scales[j+4] = (ls & 0xF) | ((lm & 0xF) << 4);
+                y[i].scales[j-4] |= ((ls >> 4) << 6);
+                y[i].scales[j-0] |= ((lm >> 4) << 6);
+            }
+        }
+        y[i].d = GGML_FP32_TO_FP16(max_scale/63.f);
+        y[i].dmin = GGML_FP32_TO_FP16(max_min/63.f);
+
+        uint8_t sc, m;
+        for (int j = 0; j < QK_K/32; ++j) {
+            get_scale_min_k4(j, y[i].scales, &sc, &m);
+            const float d = GGML_FP16_TO_FP32(y[i].d) * sc;
+            if (!d) continue;
+            const float dm = GGML_FP16_TO_FP32(y[i].dmin) * m;
+            for (int ii = 0; ii < 32; ++ii) {
+                int l = nearest_int((x[32*j + ii] + dm)/d);
+                l = MAX(0, MIN(15, l));
+                L[32*j + ii] = l;
+            }
+        }
+
+        uint8_t * q = y[i].qs;
+        for (int j = 0; j < QK_K; j += 64) {
+            for (int l = 0; l < 32; ++l) q[l] = L[j + l] | (L[j + l + 32] << 4);
+            q += 32;
+        }
+
+        x += QK_K;
+    }
+}
+
+void dequantize_row_q4_K(const block_q4_K * restrict x, float * restrict y, int64_t k) {
+    assert(k % QK_K == 0);
+    const int nb = k / QK_K;
+
+    for (int i = 0; i < nb; i++) {
+        const uint8_t * q = x[i].qs;
+
+        const float d   = GGML_FP16_TO_FP32(x[i].d);
+        const float min = GGML_FP16_TO_FP32(x[i].dmin);
+
+        int is = 0;
+        uint8_t sc, m;
+        for (int j = 0; j < QK_K; j += 64) {
+            get_scale_min_k4(is + 0, x[i].scales, &sc, &m);
+            const float d1 = d * sc; const float m1 = min * m;
+            get_scale_min_k4(is + 1, x[i].scales, &sc, &m);
+            const float d2 = d * sc; const float m2 = min * m;
+            for (int l = 0; l < 32; ++l) *y++ = d1 * (q[l] & 0xF) - m1;
+            for (int l = 0; l < 32; ++l) *y++ = d2 * (q[l]  >> 4) - m2;
+            q += 32; is += 2;
+        }
+    }
+}
+
+void quantize_row_q4_K(const float * restrict x, void * restrict vy, int64_t k) {
+    assert(k % QK_K == 0);
+    block_q4_K * restrict y = vy;
+    quantize_row_q4_K_ref(x, y, k);
+}
+
+static void quantize_row_q4_K_impl(const float * restrict x, block_q4_K * restrict y, int64_t n_per_row, const float * quant_weights) {
+    assert(n_per_row % QK_K == 0);
+    const int64_t nb = n_per_row / QK_K;
+
+    uint8_t L[QK_K];
+    uint8_t Laux[32];
+    uint8_t Ls[QK_K/32];
+    uint8_t Lm[QK_K/32];
+    float   weights[32];
+    float   sw[QK_K/32];
+    float   mins[QK_K/32];
+    float   scales[QK_K/32];
+
+    for (int i = 0; i < nb; i++) {
+
+        float sum_x2 = 0;
+        for (int l = 0; l < QK_K; ++l) sum_x2 += x[l] * x[l];
+        float sigma2 = 2*sum_x2/QK_K;
+        float av_x = sqrtf(sigma2);
+
+        for (int j = 0; j < QK_K/32; ++j) {
+            if (quant_weights) {
+                const float * qw = quant_weights + QK_K*i + 32*j;
+                for (int l = 0; l < 32; ++l) weights[l] = qw[l] * sqrtf(sigma2 + x[32*j + l]*x[32*j + l]);
+            } else {
+                for (int l = 0; l < 32; ++l) weights[l] = av_x + fabsf(x[32*j + l]);
+            }
+            float sumw = 0;
+            for (int l = 0; l < 32; ++l) sumw += weights[l];
+            sw[j] = sumw;
+            scales[j] = make_qkx3_quants(32, 15, x + 32*j, weights, L + 32*j, &mins[j], Laux, -0.9f, 0.05f, 36, false);
+        }
+
+        float d_block = make_qp_quants(QK_K/32, 63, scales, Ls, sw);
+        float m_block = make_qp_quants(QK_K/32, 63, mins,   Lm, sw);
+        for (int j = 0; j < QK_K/32; ++j) {
+            uint8_t ls = Ls[j];
+            uint8_t lm = Lm[j];
+            if (j < 4) {
+                y[i].scales[j] = ls;
+                y[i].scales[j+4] = lm;
+            } else {
+                y[i].scales[j+4] = (ls & 0xF) | ((lm & 0xF) << 4);
+                y[i].scales[j-4] |= ((ls >> 4) << 6);
+                y[i].scales[j-0] |= ((lm >> 4) << 6);
+            }
+        }
+        y[i].d = GGML_FP32_TO_FP16(d_block);
+        y[i].dmin = GGML_FP32_TO_FP16(m_block);
+
+        uint8_t sc, m;
+        for (int j = 0; j < QK_K/32; ++j) {
+            get_scale_min_k4(j, y[i].scales, &sc, &m);
+            const float d = GGML_FP16_TO_FP32(y[i].d) * sc;
+            if (!d) continue;
+            const float dm = GGML_FP16_TO_FP32(y[i].dmin) * m;
+            for (int ii = 0; ii < 32; ++ii) {
+                int l = nearest_int((x[32*j + ii] + dm)/d);
+                l = MAX(0, MIN(15, l));
+                L[32*j + ii] = l;
+            }
+        }
+        uint8_t * q = y[i].qs;
+        for (int j = 0; j < QK_K; j += 64) {
+            for (int l = 0; l < 32; ++l) q[l] = L[j + l] | (L[j + l + 32] << 4);
+            q += 32;
+        }
+
+        x += QK_K;
+
+    }
+}
+
+size_t quantize_q4_K(const float * restrict src, void * restrict dst, int64_t nrow, int64_t n_per_row, const float * quant_weights) {
+    size_t row_size = ggml_row_size(GGML_TYPE_Q4_K, n_per_row);
+    if (!quant_weights) {
+        quantize_row_q4_K_ref(src, dst, (int64_t)nrow*n_per_row);
+    }
+    else {
+        char * qrow = (char *)dst;
+        for (int64_t row = 0; row < nrow; ++row) {
+            quantize_row_q4_K_impl(src, (block_q4_K*)qrow, n_per_row, quant_weights);
+            src += n_per_row;
+            qrow += row_size;
+        }
+    }
+    return nrow * row_size;
+}
+
+// ====================== 5-bit (de)-quantization
+
+void quantize_row_q5_K_ref(const float * restrict x, block_q5_K * restrict y, int64_t k) {
+    assert(k % QK_K == 0);
+    const int64_t nb = k / QK_K;
+
+    uint8_t L[QK_K];
+    float mins[QK_K/32];
+    float scales[QK_K/32];
+    float weights[32];
+    uint8_t Laux[32];
+
+    for (int i = 0; i < nb; i++) {
+        float max_scale = 0; // as we are deducting the min, scales are always positive
+        float max_min = 0;
+        for (int j = 0; j < QK_K/32; ++j) {
+            //scales[j] = make_qkx1_quants(32, 31, x + 32*j, L + 32*j, &mins[j], 9, 0.5f);
+            float sum_x2 = 0;
+            for (int l = 0; l < 32; ++l) sum_x2 += x[32*j + l] * x[32*j + l];
+            float av_x = sqrtf(sum_x2/32);
+            for (int l = 0; l < 32; ++l) weights[l] = av_x + fabsf(x[32*j + l]);
+            scales[j] = make_qkx2_quants(32, 31, x + 32*j, weights, L + 32*j, &mins[j], Laux, -0.5f, 0.1f, 15, false);
+            float scale = scales[j];
+            if (scale > max_scale) {
+                max_scale = scale;
+            }
+            float min = mins[j];
+            if (min > max_min) {
+                max_min = min;
+            }
+        }
+
+        float inv_scale = max_scale > 0 ? 63.f/max_scale : 0.f;
+        float inv_min   = max_min   > 0 ? 63.f/max_min   : 0.f;
+        for (int j = 0; j < QK_K/32; ++j) {
+            uint8_t ls = nearest_int(inv_scale*scales[j]);
+            uint8_t lm = nearest_int(inv_min*mins[j]);
+            ls = MIN(63, ls);
+            lm = MIN(63, lm);
+            if (j < 4) {
+                y[i].scales[j] = ls;
+                y[i].scales[j+4] = lm;
+            } else {
+                y[i].scales[j+4] = (ls & 0xF) | ((lm & 0xF) << 4);
+                y[i].scales[j-4] |= ((ls >> 4) << 6);
+                y[i].scales[j-0] |= ((lm >> 4) << 6);
+            }
+        }
+        y[i].d = GGML_FP32_TO_FP16(max_scale/63.f);
+        y[i].dmin = GGML_FP32_TO_FP16(max_min/63.f);
+
+        uint8_t sc, m;
+        for (int j = 0; j < QK_K/32; ++j) {
+            get_scale_min_k4(j, y[i].scales, &sc, &m);
+            const float d = GGML_FP16_TO_FP32(y[i].d) * sc;
+            if (!d) continue;
+            const float dm = GGML_FP16_TO_FP32(y[i].dmin) * m;
+            for (int ii = 0; ii < 32; ++ii) {
+                int l = nearest_int((x[32*j + ii] + dm)/d);
+                l = MAX(0, MIN(31, l));
+                L[32*j + ii] = l;
+            }
+        }
+
+        uint8_t * restrict qh = y[i].qh;
+        uint8_t * restrict ql = y[i].qs;
+        memset(qh, 0, QK_K/8);
+
+        uint8_t m1 = 1, m2 = 2;
+        for (int n = 0; n < QK_K; n += 64) {
+            for (int j = 0; j < 32; ++j) {
+                int l1 = L[n + j];
+                if (l1 > 15) {
+                    l1 -= 16; qh[j] |= m1;
+                }
+                int l2 = L[n + j + 32];
+                if (l2 > 15) {
+                    l2 -= 16; qh[j] |= m2;
+                }
+                ql[j] = l1 | (l2 << 4);
+            }
+            m1 <<= 2; m2 <<= 2;
+            ql += 32;
+        }
+
+        x += QK_K;
+    }
+}
+
+void dequantize_row_q5_K(const block_q5_K * restrict x, float * restrict y, int64_t k) {
+    assert(k % QK_K == 0);
+    const int64_t nb = k / QK_K;
+
+    for (int i = 0; i < nb; i++) {
+        const uint8_t * ql = x[i].qs;
+        const uint8_t * qh = x[i].qh;
+
+        const float d = GGML_FP16_TO_FP32(x[i].d);
+        const float min = GGML_FP16_TO_FP32(x[i].dmin);
+
+        int is = 0;
+        uint8_t sc, m;
+        uint8_t u1 = 1, u2 = 2;
+        for (int j = 0; j < QK_K; j += 64) {
+            get_scale_min_k4(is + 0, x[i].scales, &sc, &m);
+            const float d1 = d * sc; const float m1 = min * m;
+            get_scale_min_k4(is + 1, x[i].scales, &sc, &m);
+            const float d2 = d * sc; const float m2 = min * m;
+            for (int l = 0; l < 32; ++l) *y++ = d1 * ((ql[l] & 0xF) + (qh[l] & u1 ? 16 : 0)) - m1;
+            for (int l = 0; l < 32; ++l) *y++ = d2 * ((ql[l]  >> 4) + (qh[l] & u2 ? 16 : 0)) - m2;
+            ql += 32; is += 2;
+            u1 <<= 2; u2 <<= 2;
+        }
+    }
+}
+
+void quantize_row_q5_K(const float * restrict x, void * restrict vy, int64_t k) {
+    assert(k % QK_K == 0);
+    block_q5_K * restrict y = vy;
+    quantize_row_q5_K_ref(x, y, k);
+}
+
+static void quantize_row_q5_K_impl(const float * restrict x, block_q5_K * restrict y, int64_t n_per_row, const float * quant_weights) {
+    assert(n_per_row % QK_K == 0);
+    const int64_t nb = n_per_row / QK_K;
+
+    uint8_t L[QK_K];
+    uint8_t Laux[32];
+    uint8_t Ls[QK_K/32];
+    uint8_t Lm[QK_K/32];
+    float   mins[QK_K/32];
+    float   scales[QK_K/32];
+    float   sw[QK_K/32];
+    float   weights[32];
+
+    for (int i = 0; i < nb; i++) {
+
+        float sum_x2 = 0;
+        for (int l = 0; l < QK_K; ++l) sum_x2 += x[l] * x[l];
+        float sigma2 = 2*sum_x2/QK_K;
+        float av_x = sqrtf(sigma2);
+
+        for (int j = 0; j < QK_K/32; ++j) {
+            if (quant_weights) {
+                const float * qw = quant_weights + QK_K*i + 32*j;
+                for (int l = 0; l < 32; ++l) weights[l] = qw[l] * sqrtf(sigma2 + x[32*j + l]*x[32*j + l]);
+            } else {
+                for (int l = 0; l < 32; ++l) weights[l] = av_x + fabsf(x[32*j + l]);
+            }
+            float sumw = 0;
+            for (int l = 0; l < 32; ++l) sumw += weights[l];
+            sw[j] = sumw;
+
+            scales[j] = make_qkx3_quants(32, 31, x + 32*j, weights, L + 32*j, &mins[j], Laux, -0.9f, 0.05f, 36, false);
+        }
+
+        float d_block = make_qp_quants(QK_K/32, 63, scales, Ls, sw);
+        float m_block = make_qp_quants(QK_K/32, 63, mins,   Lm, sw);
+
+        for (int j = 0; j < QK_K/32; ++j) {
+            uint8_t ls = Ls[j];
+            uint8_t lm = Lm[j];
+            ls = MIN(63, ls);
+            lm = MIN(63, lm);
+            if (j < 4) {
+                y[i].scales[j] = ls;
+                y[i].scales[j+4] = lm;
+            } else {
+                y[i].scales[j+4] = (ls & 0xF) | ((lm & 0xF) << 4);
+                y[i].scales[j-4] |= ((ls >> 4) << 6);
+                y[i].scales[j-0] |= ((lm >> 4) << 6);
+            }
+        }
+        y[i].d = GGML_FP32_TO_FP16(d_block);
+        y[i].dmin = GGML_FP32_TO_FP16(m_block);
+
+        uint8_t sc, m;
+        for (int j = 0; j < QK_K/32; ++j) {
+            get_scale_min_k4(j, y[i].scales, &sc, &m);
+            const float d = GGML_FP16_TO_FP32(y[i].d) * sc;
+            if (!d) continue;
+            const float dm = GGML_FP16_TO_FP32(y[i].dmin) * m;
+            for (int ii = 0; ii < 32; ++ii) {
+                int l = nearest_int((x[32*j + ii] + dm)/d);
+                l = MAX(0, MIN(31, l));
+                L[32*j + ii] = l;
+            }
+        }
+
+        uint8_t * restrict qh = y[i].qh;
+        uint8_t * restrict ql = y[i].qs;
+        memset(qh, 0, QK_K/8);
+
+        uint8_t m1 = 1, m2 = 2;
+        for (int n = 0; n < QK_K; n += 64) {
+            for (int j = 0; j < 32; ++j) {
+                int l1 = L[n + j];
+                if (l1 > 15) {
+                    l1 -= 16; qh[j] |= m1;
+                }
+                int l2 = L[n + j + 32];
+                if (l2 > 15) {
+                    l2 -= 16; qh[j] |= m2;
+                }
+                ql[j] = l1 | (l2 << 4);
+            }
+            m1 <<= 2; m2 <<= 2;
+            ql += 32;
+        }
+
+        x += QK_K;
+
+    }
+}
+
+size_t quantize_q5_K(const float * restrict src, void * restrict dst, int64_t nrow, int64_t n_per_row, const float * quant_weights) {
+    size_t row_size = ggml_row_size(GGML_TYPE_Q5_K, n_per_row);
+    if (!quant_weights) {
+        quantize_row_q5_K_ref(src, dst, (int64_t)nrow*n_per_row);
+    }
+    else {
+        char * qrow = (char *)dst;
+        for (int64_t row = 0; row < nrow; ++row) {
+            quantize_row_q5_K_impl(src, (block_q5_K*)qrow, n_per_row, quant_weights);
+            src += n_per_row;
+            qrow += row_size;
+        }
+    }
+    return nrow * row_size;
+}
+
+// ====================== 6-bit (de)-quantization
+
+void quantize_row_q6_K_ref(const float * restrict x, block_q6_K * restrict y, int64_t k) {
+    assert(k % QK_K == 0);
+    const int64_t nb = k / QK_K;
+
+    int8_t L[QK_K];
+    float   scales[QK_K/16];
+
+    for (int i = 0; i < nb; i++) {
+
+        float max_scale = 0;
+        float max_abs_scale = 0;
+
+        for (int ib = 0; ib < QK_K/16; ++ib) {
+
+            const float scale = make_qx_quants(16, 32, x + 16*ib, L + 16*ib, 1, NULL);
+            scales[ib] = scale;
+
+            const float abs_scale = fabsf(scale);
+            if (abs_scale > max_abs_scale) {
+                max_abs_scale = abs_scale;
+                max_scale = scale;
+            }
+
+        }
+
+        if (max_abs_scale < GROUP_MAX_EPS) {
+            memset(&y[i], 0, sizeof(block_q6_K));
+            y[i].d = GGML_FP32_TO_FP16(0.f);
+            x += QK_K;
+            continue;
+        }
+
+        float iscale = -128.f/max_scale;
+        y[i].d = GGML_FP32_TO_FP16(1/iscale);
+        for (int ib = 0; ib < QK_K/16; ++ib) {
+            y[i].scales[ib] = MIN(127, nearest_int(iscale*scales[ib]));
+        }
+
+        for (int j = 0; j < QK_K/16; ++j) {
+            float d = GGML_FP16_TO_FP32(y[i].d) * y[i].scales[j];
+            if (!d) {
+                continue;
+            }
+            for (int ii = 0; ii < 16; ++ii) {
+                int l = nearest_int(x[16*j + ii]/d);
+                l = MAX(-32, MIN(31, l));
+                L[16*j + ii] = l + 32;
+            }
+        }
+
+        uint8_t * restrict ql = y[i].ql;
+        uint8_t * restrict qh = y[i].qh;
+        for (int j = 0; j < QK_K; j += 128) {
+            for (int l = 0; l < 32; ++l) {
+                const uint8_t q1 = L[j + l +  0] & 0xF;
+                const uint8_t q2 = L[j + l + 32] & 0xF;
+                const uint8_t q3 = L[j + l + 64] & 0xF;
+                const uint8_t q4 = L[j + l + 96] & 0xF;
+                ql[l+ 0] = q1 | (q3 << 4);
+                ql[l+32] = q2 | (q4 << 4);
+                qh[l] = (L[j + l] >> 4) | ((L[j + l + 32] >> 4) << 2) | ((L[j + l + 64] >> 4) << 4) | ((L[j + l + 96] >> 4) << 6);
+            }
+            ql += 64;
+            qh += 32;
+        }
+
+        x += QK_K;
+    }
+}
+
+void dequantize_row_q6_K(const block_q6_K * restrict x, float * restrict y, int64_t k) {
+    assert(k % QK_K == 0);
+    const int64_t nb = k / QK_K;
+
+    for (int i = 0; i < nb; i++) {
+        const float d = GGML_FP16_TO_FP32(x[i].d);
+
+        const uint8_t * restrict ql = x[i].ql;
+        const uint8_t * restrict qh = x[i].qh;
+        const int8_t  * restrict sc = x[i].scales;
+
+        for (int n = 0; n < QK_K; n += 128) {
+            for (int l = 0; l < 32; ++l) {
+                int is = l/16;
+                const int8_t q1 = (int8_t)((ql[l +  0] & 0xF) | (((qh[l] >> 0) & 3) << 4)) - 32;
+                const int8_t q2 = (int8_t)((ql[l + 32] & 0xF) | (((qh[l] >> 2) & 3) << 4)) - 32;
+                const int8_t q3 = (int8_t)((ql[l +  0]  >> 4) | (((qh[l] >> 4) & 3) << 4)) - 32;
+                const int8_t q4 = (int8_t)((ql[l + 32]  >> 4) | (((qh[l] >> 6) & 3) << 4)) - 32;
+                y[l +  0] = d * sc[is + 0] * q1;
+                y[l + 32] = d * sc[is + 2] * q2;
+                y[l + 64] = d * sc[is + 4] * q3;
+                y[l + 96] = d * sc[is + 6] * q4;
+            }
+            y  += 128;
+            ql += 64;
+            qh += 32;
+            sc += 8;
+        }
+    }
+}
+
+void quantize_row_q6_K(const float * restrict x, void * restrict vy, int64_t k) {
+    assert(k % QK_K == 0);
+    block_q6_K * restrict y = vy;
+    quantize_row_q6_K_ref(x, y, k);
+}
+
+static void quantize_row_q6_K_impl(const float * restrict x, block_q6_K * restrict y, int64_t n_per_row, const float * quant_weights) {
+    assert(n_per_row % QK_K == 0);
+    const int64_t nb = n_per_row / QK_K;
+
+    int8_t L[QK_K];
+    float   scales[QK_K/16];
+    //float   weights[16];
+
+    for (int i = 0; i < nb; i++) {
+
+        //float sum_x2 = 0;
+        //for (int j = 0; j < QK_K; ++j) sum_x2 += x[j]*x[j];
+        //float sigma2 = sum_x2/QK_K;
+
+        float max_scale = 0;
+        float max_abs_scale = 0;
+
+        for (int ib = 0; ib < QK_K/16; ++ib) {
+
+            float scale;
+            if (quant_weights) {
+                const float * qw = quant_weights + QK_K*i + 16*ib;
+                //for (int j = 0; j < 16; ++j) weights[j] = qw[j] * sqrtf(sigma2 + x[16*ib + j]*x[16*ib + j]);
+                //scale = make_qx_quants(16, 32, x + 16*ib, L + 16*ib, 1, weights);
+                scale = make_qx_quants(16, 32, x + 16*ib, L + 16*ib, 1, qw);
+            } else {
+                scale = make_qx_quants(16, 32, x + 16*ib, L + 16*ib, 1, NULL);
+            }
+            scales[ib] = scale;
+
+            const float abs_scale = fabsf(scale);
+            if (abs_scale > max_abs_scale) {
+                max_abs_scale = abs_scale;
+                max_scale = scale;
+            }
+
+        }
+
+        if (max_abs_scale < GROUP_MAX_EPS) {
+            memset(&y[i], 0, sizeof(block_q6_K));
+            y[i].d = GGML_FP32_TO_FP16(0.f);
+            x += QK_K;
+            continue;
+        }
+
+        float iscale = -128.f/max_scale;
+        y[i].d = GGML_FP32_TO_FP16(1/iscale);
+        for (int ib = 0; ib < QK_K/16; ++ib) {
+            y[i].scales[ib] = MIN(127, nearest_int(iscale*scales[ib]));
+        }
+
+        for (int j = 0; j < QK_K/16; ++j) {
+            float d = GGML_FP16_TO_FP32(y[i].d) * y[i].scales[j];
+            if (!d) {
+                continue;
+            }
+            for (int ii = 0; ii < 16; ++ii) {
+                int l = nearest_int(x[16*j + ii]/d);
+                l = MAX(-32, MIN(31, l));
+                L[16*j + ii] = l + 32;
+            }
+        }
+
+        uint8_t * restrict ql = y[i].ql;
+        uint8_t * restrict qh = y[i].qh;
+        for (int j = 0; j < QK_K; j += 128) {
+            for (int l = 0; l < 32; ++l) {
+                const uint8_t q1 = L[j + l +  0] & 0xF;
+                const uint8_t q2 = L[j + l + 32] & 0xF;
+                const uint8_t q3 = L[j + l + 64] & 0xF;
+                const uint8_t q4 = L[j + l + 96] & 0xF;
+                ql[l+ 0] = q1 | (q3 << 4);
+                ql[l+32] = q2 | (q4 << 4);
+                qh[l] = (L[j + l] >> 4) | ((L[j + l + 32] >> 4) << 2) | ((L[j + l + 64] >> 4) << 4) | ((L[j + l + 96] >> 4) << 6);
+            }
+            ql += 64;
+            qh += 32;
+        }
+
+        x += QK_K;
+
+    }
+}
+
+size_t quantize_q6_K(const float * restrict src, void * restrict dst, int64_t nrow, int64_t n_per_row, const float * quant_weights) {
+    size_t row_size = ggml_row_size(GGML_TYPE_Q6_K, n_per_row);
+    if (!quant_weights) {
+        quantize_row_q6_K_ref(src, dst, (int64_t)nrow*n_per_row);
+    }
+    else {
+        char * qrow = (char *)dst;
+        for (int64_t row = 0; row < nrow; ++row) {
+            quantize_row_q6_K_impl(src, (block_q6_K*)qrow, n_per_row, quant_weights);
+            src += n_per_row;
+            qrow += row_size;
+        }
+    }
+    return nrow * row_size;
+}
+
+static void quantize_row_q4_0_impl(const float * restrict x, block_q4_0 * restrict y, int64_t n_per_row, const float * quant_weights) {
+    static_assert(QK4_0 == 32, "QK4_0 must be 32");
+
+    if (!quant_weights) {
+        quantize_row_q4_0_ref(x, y, n_per_row);
+        return;
+    }
+
+    float weight[QK4_0];
+    int8_t L[QK4_0];
+
+    float sum_x2 = 0;
+    for (int j = 0; j < n_per_row; ++j) sum_x2 += x[j]*x[j];
+    float sigma2 = sum_x2/n_per_row;
+
+    const int64_t nb = n_per_row/QK4_0;
+    for (int ib = 0; ib < nb; ++ib) {
+        const float * xb = x + QK4_0 * ib;
+        const float * qw = quant_weights + QK4_0 * ib;
+        for (int j = 0; j < QK4_0; ++j) weight[j] = qw[j] * sqrtf(sigma2 + xb[j]*xb[j]);
+        float d = make_qx_quants(QK4_0, 8, xb, L, 1, weight);
+        y[ib].d = GGML_FP32_TO_FP16(d);
+        for (int j = 0; j < 16; ++j) {
+            y[ib].qs[j] = L[j] | (L[j+16] << 4);
+        }
+    }
+}
+
+size_t quantize_q4_0(const float * restrict src, void * restrict dst, int64_t nrow, int64_t n_per_row, const float * quant_weights) {
+    if (!quant_weights) {
+        quantize_row_q4_0_ref(src, dst, (int64_t)nrow*n_per_row);
+        return nrow * ggml_row_size(GGML_TYPE_Q4_0, n_per_row);
+    }
+    size_t row_size = ggml_row_size(GGML_TYPE_Q4_0, n_per_row);
+    char * qrow = (char *)dst;
+    for (int64_t row = 0; row < nrow; ++row) {
+        quantize_row_q4_0_impl(src, (block_q4_0*)qrow, n_per_row, quant_weights);
+        src += n_per_row;
+        qrow += row_size;
+    }
+    return nrow * row_size;
+}
+
+static void quantize_row_q4_1_impl(const float * restrict x, block_q4_1 * restrict y, int64_t n_per_row, const float * quant_weights) {
+    static_assert(QK4_1 == 32, "QK4_1 must be 32");
+
+    if (!quant_weights) {
+        quantize_row_q4_1_ref(x, y, n_per_row);
+        return;
+    }
+
+    float weight[QK4_1];
+    uint8_t L[QK4_1], Laux[QK4_1];
+
+    float sum_x2 = 0;
+    for (int j = 0; j < n_per_row; ++j) sum_x2 += x[j]*x[j];
+    float sigma2 = sum_x2/n_per_row;
+
+    const int64_t nb = n_per_row/QK4_1;
+    for (int ib = 0; ib < nb; ++ib) {
+        const float * xb = x + QK4_1 * ib;
+        const float * qw = quant_weights + QK4_1 * ib;
+        for (int j = 0; j < QK4_1; ++j) weight[j] = qw[j] * sqrtf(sigma2 + xb[j]*xb[j]);
+        float min;
+        float d = make_qkx3_quants(QK4_1, 15, xb, weight, L, &min, Laux, -0.9f, 0.05f, 36, false);
+        y[ib].d = GGML_FP32_TO_FP16(d);
+        y[ib].m = GGML_FP32_TO_FP16(-min);
+        for (int j = 0; j < 16; ++j) {
+            y[ib].qs[j] = L[j] | (L[j+16] << 4);
+        }
+    }
+}
+
+size_t quantize_q4_1(const float * restrict src, void * restrict dst, int64_t nrow, int64_t n_per_row, const float * quant_weights) {
+    if (!quant_weights) {
+        quantize_row_q4_1_ref(src, dst, (int64_t)nrow*n_per_row);
+        return nrow * ggml_row_size(GGML_TYPE_Q4_1, n_per_row);
+    }
+    size_t row_size = ggml_row_size(GGML_TYPE_Q4_1, n_per_row);
+    char * qrow = (char *)dst;
+    for (int64_t row = 0; row < nrow; ++row) {
+        quantize_row_q4_1_impl(src, (block_q4_1*)qrow, n_per_row, quant_weights);
+        src += n_per_row;
+        qrow += row_size;
+    }
+    return nrow * row_size;
+}
+
+static void quantize_row_q5_0_impl(const float * restrict x, block_q5_0 * restrict y, int64_t n_per_row, const float * quant_weights) {
+    static_assert(QK5_0 == 32, "QK5_0 must be 32");
+
+    if (!quant_weights) {
+        quantize_row_q5_0_ref(x, y, n_per_row);
+        return;
+    }
+
+    float weight[QK5_0];
+    int8_t L[QK5_0];
+
+    float sum_x2 = 0;
+    for (int j = 0; j < n_per_row; ++j) sum_x2 += x[j]*x[j];
+    float sigma2 = sum_x2/n_per_row;
+
+    const int64_t nb = n_per_row/QK5_0;
+    for (int ib = 0; ib < nb; ++ib) {
+        const float * xb = x + QK5_0 * ib;
+        const float * qw = quant_weights + QK5_0 * ib;
+        for (int j = 0; j < QK5_0; ++j) weight[j] = qw[j] * sqrtf(sigma2 + xb[j]*xb[j]);
+        float d = make_qx_quants(QK5_0, 16, xb, L, 1, weight);
+        y[ib].d = GGML_FP32_TO_FP16(d);
+
+        uint32_t qh = 0;
+
+        for (int j = 0; j < 16; ++j) {
+            const uint8_t xi0 = L[j];
+            const uint8_t xi1 = L[j+16];
+            y[ib].qs[j] = (xi0 & 0x0F) | ((xi1 & 0x0F) << 4);
+
+            // get the 5-th bit and store it in qh at the right position
+            qh |= ((xi0 & 0x10u) >> 4) << (j + 0);
+            qh |= ((xi1 & 0x10u) >> 4) << (j + QK5_0/2);
+        }
+
+        memcpy(&y[ib].qh, &qh, sizeof(qh));
+    }
+}
+
+size_t quantize_q5_0(const float * restrict src, void * restrict dst, int64_t nrow, int64_t n_per_row, const float * quant_weights) {
+    if (!quant_weights) {
+        quantize_row_q5_0_ref(src, dst, (int64_t)nrow*n_per_row);
+        return nrow * ggml_row_size(GGML_TYPE_Q5_0, n_per_row);
+    }
+    size_t row_size = ggml_row_size(GGML_TYPE_Q5_0, n_per_row);
+    char * qrow = (char *)dst;
+    for (int64_t row = 0; row < nrow; ++row) {
+        quantize_row_q5_0_impl(src, (block_q5_0*)qrow, n_per_row, quant_weights);
+        src += n_per_row;
+        qrow += row_size;
+    }
+    return nrow * row_size;
+}
+
+static void quantize_row_q5_1_impl(const float * restrict x, block_q5_1 * restrict y, int64_t n_per_row, const float * quant_weights) {
+    static_assert(QK5_1 == 32, "QK5_1 must be 32");
+
+    if (!quant_weights) {
+        quantize_row_q5_1_ref(x, y, n_per_row);
+        return;
+    }
+
+    float weight[QK5_1];
+    uint8_t L[QK5_1], Laux[QK5_1];
+
+    float sum_x2 = 0;
+    for (int j = 0; j < n_per_row; ++j) sum_x2 += x[j]*x[j];
+    float sigma2 = sum_x2/n_per_row;
+
+    const int64_t nb = n_per_row/QK5_1;
+    for (int ib = 0; ib < nb; ++ib) {
+        const float * xb = x + QK5_1 * ib;
+        const float * qw = quant_weights + QK5_1 * ib;
+        for (int j = 0; j < QK5_1; ++j) weight[j] = qw[j] * sqrtf(sigma2 + xb[j]*xb[j]);
+        float min;
+        float d = make_qkx3_quants(QK5_1, 31, xb, weight, L, &min, Laux, -0.9f, 0.05f, 36, false);
+        y[ib].d = GGML_FP32_TO_FP16(d);
+        y[ib].m = GGML_FP32_TO_FP16(-min);
+
+        uint32_t qh = 0;
+        for (int j = 0; j < 16; ++j) {
+            const uint8_t xi0 = L[j];
+            const uint8_t xi1 = L[j+16];
+            y[ib].qs[j] = (xi0 & 0x0F) | ((xi1 & 0x0F) << 4);
+            // get the 5-th bit and store it in qh at the right position
+            qh |= ((xi0 & 0x10u) >> 4) << (j + 0);
+            qh |= ((xi1 & 0x10u) >> 4) << (j + QK5_0/2);
+        }
+        memcpy(&y[ib].qh, &qh, sizeof(qh));
+    }
+}
+
+size_t quantize_q5_1(const float * restrict src, void * restrict dst, int64_t nrow, int64_t n_per_row, const float * quant_weights) {
+    if (!quant_weights) {
+        quantize_row_q5_1_ref(src, dst, (int64_t)nrow*n_per_row);
+        return nrow * ggml_row_size(GGML_TYPE_Q5_1, n_per_row);
+    }
+    size_t row_size = ggml_row_size(GGML_TYPE_Q5_1, n_per_row);
+    char * qrow = (char *)dst;
+    for (int64_t row = 0; row < nrow; ++row) {
+        quantize_row_q5_1_impl(src, (block_q5_1*)qrow, n_per_row, quant_weights);
+        src += n_per_row;
+        qrow += row_size;
+    }
+    return nrow * row_size;
+}
+
+size_t quantize_q8_0(const float * restrict src, void * restrict dst, int64_t nrow, int64_t n_per_row, const float * quant_weights) {
+    (void)quant_weights; // not used
+    const size_t row_size = ggml_row_size(GGML_TYPE_Q8_0, n_per_row);
+    quantize_row_q8_0_ref(src, dst, (int64_t)nrow*n_per_row);
+    return nrow * row_size;
+}
+
+// ====================== Ternary (de)-quantization (BitNet b1.58 and TriLMs)
+
+void quantize_row_tq1_0_ref(const float * restrict x, block_tq1_0 * restrict y, int64_t k) {
+    assert(k % QK_K == 0);
+    const int64_t nb = k / QK_K;
+
+    for (int64_t i = 0; i < nb; i++) {
+        float amax = 0.0f; // absolute max
+
+        for (int j = 0; j < QK_K; j++) {
+            const float v = x[j];
+            amax = MAX(amax, fabsf(v));
+        }
+
+        const float d = amax;
+        const float id = d ? 1.0f/d : 0.0f;
+
+        y[i].d = GGML_FP32_TO_FP16(d);
+
+        // 5 elements per byte, along 32 bytes
+        for (size_t j = 0; j < sizeof(y->qs) - sizeof(y->qs) % 32; j += 32) {
+            for (size_t m = 0; m < 32; ++m) {
+                uint8_t q = 0;
+                for (size_t n = 0; n < 5; ++n) {
+                    int xi = lroundf(x[m + n*32] * id) + 1; // -1, 0, 1 -> 0, 1, 2
+                    q *= 3;
+                    q += xi;
+                }
+                // ceiling division (243 == pow(3, 5))
+                q = ((uint16_t)q * 256 + (243 - 1)) / 243;
+                y[i].qs[j + m] = q;
+            }
+            x += 5*32;
+        }
+        // along 16 bytes
+        for (size_t j = sizeof(y->qs) - sizeof(y->qs) % 32; j < sizeof(y->qs); j += 16) {
+            for (size_t m = 0; m < 16; ++m) {
+                uint8_t q = 0;
+                for (size_t n = 0; n < 5; ++n) {
+                    int xi = lroundf(x[m + n*16] * id) + 1; // -1, 0, 1 -> 0, 1, 2
+                    q *= 3;
+                    q += xi;
+                }
+                // ceiling division (243 == pow(3, 5))
+                q = ((uint16_t)q * 256 + (243 - 1)) / 243;
+                y[i].qs[j + m] = q;
+            }
+            x += 5*16;
+        }
+        // 4 elements per byte
+        for (size_t j = 0; j < sizeof(y->qh); ++j) {
+            uint8_t q = 0;
+            for (size_t m = 0; m < 4; ++m) {
+                // -1, 0, 1 -> 0, 1, 2
+                int xi = lroundf(x[j + m*sizeof(y->qh)] * id) + 1;
+                q *= 3;
+                q += xi;
+            }
+            // shift the first value to the most significant trit
+            q *= 3;
+            // ceiling division (243 == pow(3, 5))
+            q = ((uint16_t)q * 256 + (243 - 1)) / 243;
+            y[i].qh[j] = q;
+        }
+        x += 4*sizeof(y->qh);
+    }
+}
+
+void quantize_row_tq2_0_ref(const float * restrict x, block_tq2_0 * restrict y, int64_t k) {
+    assert(k % QK_K == 0);
+    const int64_t nb = k / QK_K;
+
+    for (int64_t i = 0; i < nb; i++) {
+        float amax = 0.0f; // absolute max
+
+        for (int j = 0; j < QK_K; j++) {
+            const float v = x[j];
+            amax = MAX(amax, fabsf(v));
+        }
+
+        const float d = amax;
+        const float id = d ? 1.0f/d : 0.0f;
+
+        y[i].d = GGML_FP32_TO_FP16(d);
+
+        for (size_t j = 0; j < sizeof(y->qs); j += 32) {
+            for (size_t m = 0; m < 32; ++m) {
+                uint8_t q = 0;
+                for (size_t n = 0; n < 4; ++n) {
+                    // -1, 0, 1 -> 0, 1, 2
+                    int xi = lroundf(x[m + n*32] * id) + 1;
+                    q += (xi & 3) << (2*n);
+                }
+                y[i].qs[j + m] = q;
+            }
+            x += 4*32;
+        }
+    }
+}
+
+void quantize_row_tq1_0(const float * restrict x, void * restrict vy, int64_t k) {
+    assert(k % QK_K == 0);
+    block_tq1_0 * restrict y = vy;
+    quantize_row_tq1_0_ref(x, y, k);
+}
+
+void quantize_row_tq2_0(const float * restrict x, void * restrict vy, int64_t k) {
+    assert(k % QK_K == 0);
+    block_tq2_0 * restrict y = vy;
+    quantize_row_tq2_0_ref(x, y, k);
+}
+
+size_t quantize_tq1_0(const float * restrict src, void * restrict dst, int64_t nrow, int64_t n_per_row, const float * quant_weights) {
+    (void)quant_weights; // not used
+    const size_t row_size = ggml_row_size(GGML_TYPE_TQ1_0, n_per_row);
+    quantize_row_tq1_0(src, dst, (int64_t)nrow*n_per_row);
+    return nrow * row_size;
+}
+
+size_t quantize_tq2_0(const float * restrict src, void * restrict dst, int64_t nrow, int64_t n_per_row, const float * quant_weights) {
+    (void)quant_weights; // not used
+    const size_t row_size = ggml_row_size(GGML_TYPE_TQ2_0, n_per_row);
+    quantize_row_tq2_0(src, dst, (int64_t)nrow*n_per_row);
+    return nrow * row_size;
+}
+
+
+void dequantize_row_tq1_0(const block_tq1_0 * restrict x, float * restrict y, int64_t k) {
+    assert(k % QK_K == 0);
+    const int64_t nb = k / QK_K;
+
+    const uint8_t pow3[6] = {1, 3, 9, 27, 81, 243};
+
+    for (int64_t i = 0; i < nb; ++i) {
+
+        const float d = GGML_FP16_TO_FP32(x[i].d);
+
+        for (size_t j = 0; j < sizeof(x->qs) - sizeof(x->qs) % 32; j += 32) {
+            for (size_t n = 0; n < 5; ++n) {
+                for (size_t m = 0; m < 32; ++m) {
+                    uint8_t q = x[i].qs[j + m] * pow3[n];
+                    int16_t xi = ((uint16_t) q * 3) >> 8;
+                    *y++ = (float) (xi - 1) * d;
+                }
+            }
+        }
+        for (size_t j = sizeof(x->qs) - sizeof(x->qs) % 32; j < sizeof(x->qs); j += 16) {
+            for (size_t n = 0; n < 5; ++n) {
+                for (size_t m = 0; m < 16; ++m) {
+                    uint8_t q = x[i].qs[j + m] * pow3[n];
+                    int16_t xi = ((uint16_t) q * 3) >> 8;
+                    *y++ = (float) (xi - 1) * d;
+                }
+            }
+        }
+
+        for (size_t n = 0; n < 4; ++n) {
+            for (size_t j = 0; j < sizeof(x->qh); ++j) {
+                uint8_t q = x[i].qh[j] * pow3[n];
+                int16_t xi = ((uint16_t) q * 3) >> 8;
+                *y++ = (float) (xi - 1) * d;
+            }
+        }
+    }
+}
+
+void dequantize_row_tq2_0(const block_tq2_0 * restrict x, float * restrict y, int64_t k) {
+    assert(k % QK_K == 0);
+    const int64_t nb = k / QK_K;
+
+    for (int64_t i = 0; i < nb; ++i) {
+
+        const float d = GGML_FP16_TO_FP32(x[i].d);
+
+        for (size_t j = 0; j < sizeof(x->qs); j += 32) {
+            for (size_t l = 0; l < 4; ++l) {
+                for (size_t m = 0; m < 32; ++m) {
+                    int8_t q = (x[i].qs[j + m] >> (l*2)) & 3;
+                    *y++ = (float) (q - 1) * d;
+                }
+            }
+        }
+    }
+}
+
+// ====================== "True" 2-bit (de)-quantization
+
+void dequantize_row_iq2_xxs(const block_iq2_xxs * restrict x, float * restrict y, int64_t k) {
+    assert(k % QK_K == 0);
+    const int64_t nb = k / QK_K;
+
+    uint32_t aux32[2];
+    const uint8_t * aux8 = (const uint8_t *)aux32;
+
+    for (int i = 0; i < nb; i++) {
+
+        const float d = GGML_FP16_TO_FP32(x[i].d);
+
+        for (int ib32 = 0; ib32 < QK_K/32; ++ib32) {
+            memcpy(aux32, x[i].qs + 4*ib32, 2*sizeof(uint32_t));
+            const float db = d * (0.5f + (aux32[1] >> 28)) * 0.25f;
+            for (int l = 0; l < 4; ++l) {
+                const uint8_t * grid = (const uint8_t *)(iq2xxs_grid + aux8[l]);
+                const uint8_t  signs = ksigns_iq2xs[(aux32[1] >> 7*l) & 127];
+                for (int j = 0; j < 8; ++j) {
+                    y[j] = db * grid[j] * (signs & kmask_iq2xs[j] ? -1.f : 1.f);
+                }
+                y += 8;
+            }
+        }
+    }
+}
+
+// ====================== 2.3125 bpw (de)-quantization
+
+void dequantize_row_iq2_xs(const block_iq2_xs * restrict x, float * restrict y, int64_t k) {
+    assert(k % QK_K == 0);
+    const int64_t nb = k / QK_K;
+
+    float db[2];
+
+    for (int i = 0; i < nb; i++) {
+
+        const float d = GGML_FP16_TO_FP32(x[i].d);
+
+        for (int ib32 = 0; ib32 < QK_K/32; ++ib32) {
+            db[0] = d * (0.5f + (x[i].scales[ib32] & 0xf)) * 0.25f;
+            db[1] = d * (0.5f + (x[i].scales[ib32] >>  4)) * 0.25f;
+            for (int l = 0; l < 4; ++l) {
+                const uint8_t * grid = (const uint8_t *)(iq2xs_grid + (x[i].qs[4*ib32 + l] & 511));
+                const uint8_t  signs = ksigns_iq2xs[x[i].qs[4*ib32 + l] >> 9];
+                for (int j = 0; j < 8; ++j) {
+                    y[j] = db[l/2] * grid[j] * (signs & kmask_iq2xs[j] ? -1.f : 1.f);
+                }
+                y += 8;
+            }
+        }
+    }
+}
+
+// ====================== 2.5625 bpw (de)-quantization
+
+void dequantize_row_iq2_s(const block_iq2_s * restrict x, float * restrict y, int64_t k) {
+    assert(k % QK_K == 0);
+    const int64_t nb = k / QK_K;
+
+    float db[2];
+
+    for (int i = 0; i < nb; i++) {
+
+        const float d = GGML_FP16_TO_FP32(x[i].d);
+        const uint8_t * qs = x[i].qs;
+        const uint8_t * qh = x[i].qh;
+        const uint8_t * signs = qs + QK_K/8;
+
+        for (int ib32 = 0; ib32 < QK_K/32; ++ib32) {
+            db[0] = d * (0.5f + (x[i].scales[ib32] & 0xf)) * 0.25f;
+            db[1] = d * (0.5f + (x[i].scales[ib32] >>  4)) * 0.25f;
+            for (int l = 0; l < 4; ++l) {
+                const float dl = db[l/2];
+                const uint8_t * grid = (const uint8_t *)(iq2s_grid + (qs[l] | (qh[ib32] << (8-2*l) & 0x300)));
+                for (int j = 0; j < 8; ++j) {
+                    y[j] = dl * grid[j] * (signs[l] & kmask_iq2xs[j] ? -1.f : 1.f);
+                }
+                y += 8;
+            }
+            qs += 4;
+            signs += 4;
+        }
+    }
+}
+
+// ====================== 3.0625 bpw (de)-quantization
+
+void dequantize_row_iq3_xxs(const block_iq3_xxs * restrict x, float * restrict y, int64_t k) {
+    assert(k % QK_K == 0);
+    const int64_t nb = k / QK_K;
+
+    uint32_t aux32;
+
+    for (int i = 0; i < nb; i++) {
+
+        const float d = GGML_FP16_TO_FP32(x[i].d);
+        const uint8_t * qs = x[i].qs;
+        const uint8_t * scales_and_signs = qs + QK_K/4;
+
+        for (int ib32 = 0; ib32 < QK_K/32; ++ib32) {
+            memcpy(&aux32, scales_and_signs + 4*ib32, sizeof(uint32_t));
+            const float db = d * (0.5f + (aux32 >> 28)) * 0.5f;
+            for (int l = 0; l < 4; ++l) {
+                const uint8_t  signs = ksigns_iq2xs[(aux32 >> 7*l) & 127];
+                const uint8_t * grid1 = (const uint8_t *)(iq3xxs_grid + qs[2*l+0]);
+                const uint8_t * grid2 = (const uint8_t *)(iq3xxs_grid + qs[2*l+1]);
+                for (int j = 0; j < 4; ++j) {
+                    y[j+0] = db * grid1[j] * (signs & kmask_iq2xs[j+0] ? -1.f : 1.f);
+                    y[j+4] = db * grid2[j] * (signs & kmask_iq2xs[j+4] ? -1.f : 1.f);
+                }
+                y += 8;
+            }
+            qs += 8;
+        }
+    }
+}
+
+// ====================== 3.3125 bpw (de)-quantization
+
+void dequantize_row_iq3_s(const block_iq3_s * restrict x, float * restrict y, int64_t k) {
+    assert(k % QK_K == 0);
+    const int64_t nb = k / QK_K;
+
+    for (int i = 0; i < nb; i++) {
+
+        const float d = GGML_FP16_TO_FP32(x[i].d);
+        const uint8_t * qs = x[i].qs;
+        const uint8_t * qh = x[i].qh;
+        const uint8_t * signs = x[i].signs;
+
+        for (int ib32 = 0; ib32 < QK_K/32; ib32 += 2) {
+            const float db1 = d * (1 + 2*(x[i].scales[ib32/2] & 0xf));
+            const float db2 = d * (1 + 2*(x[i].scales[ib32/2] >>  4));
+            for (int l = 0; l < 4; ++l) {
+                const uint8_t * grid1 = (const uint8_t *)(iq3s_grid + (qs[2*l+0] | ((qh[0] << (8-2*l)) & 256)));
+                const uint8_t * grid2 = (const uint8_t *)(iq3s_grid + (qs[2*l+1] | ((qh[0] << (7-2*l)) & 256)));
+                for (int j = 0; j < 4; ++j) {
+                    y[j+0] = db1 * grid1[j] * (signs[l] & kmask_iq2xs[j+0] ? -1.f : 1.f);
+                    y[j+4] = db1 * grid2[j] * (signs[l] & kmask_iq2xs[j+4] ? -1.f : 1.f);
+                }
+                y += 8;
+            }
+            qs += 8;
+            signs += 4;
+            for (int l = 0; l < 4; ++l) {
+                const uint8_t * grid1 = (const uint8_t *)(iq3s_grid + (qs[2*l+0] | ((qh[1] << (8-2*l)) & 256)));
+                const uint8_t * grid2 = (const uint8_t *)(iq3s_grid + (qs[2*l+1] | ((qh[1] << (7-2*l)) & 256)));
+                for (int j = 0; j < 4; ++j) {
+                    y[j+0] = db2 * grid1[j] * (signs[l] & kmask_iq2xs[j+0] ? -1.f : 1.f);
+                    y[j+4] = db2 * grid2[j] * (signs[l] & kmask_iq2xs[j+4] ? -1.f : 1.f);
+                }
+                y += 8;
+            }
+            qh += 2;
+            qs += 8;
+            signs += 4;
+        }
+    }
+}
+
+// ====================== 1.5625 bpw (de)-quantization
+
+void dequantize_row_iq1_s(const block_iq1_s * restrict x, float * restrict y, int64_t k) {
+    assert(k % QK_K == 0);
+    const int64_t nb = k / QK_K;
+
+    for (int i = 0; i < nb; i++) {
+
+        const float d = GGML_FP16_TO_FP32(x[i].d);
+        const uint8_t  * qs = x[i].qs;
+        const uint16_t * qh = x[i].qh;
+
+        for (int ib = 0; ib < QK_K/32; ++ib) {
+            const float dl = d * (2*((qh[ib] >> 12) & 7) + 1);
+            const float delta = qh[ib] & 0x8000 ? -IQ1S_DELTA : IQ1S_DELTA;
+            for (int l = 0; l < 4; ++l) {
+                const int8_t * grid = (const int8_t *)(iq1s_grid + (qs[l] | (((qh[ib] >> 3*l) & 7) << 8)));
+                for (int j = 0; j < 8; ++j) {
+                    y[j] = dl * (grid[j] + delta);
+                }
+                y += 8;
+            }
+            qs += 4;
+        }
+    }
+}
+
+void dequantize_row_iq1_m(const block_iq1_m * restrict x, float * restrict y, int64_t k) {
+    assert(k % QK_K == 0);
+    const int64_t nb = k / QK_K;
+
+    float delta[4];
+    uint16_t idx[4];
+
+    iq1m_scale_t scale;
+
+    for (int i = 0; i < nb; i++) {
+
+        const uint16_t * sc = (const uint16_t *)x[i].scales;
+        scale.u16 = (sc[0] >> 12) | ((sc[1] >> 8) & 0x00f0) | ((sc[2] >> 4) & 0x0f00) | (sc[3] & 0xf000);
+        const float d = GGML_FP16_TO_FP32(scale.f16);
+
+        const uint8_t * qs = x[i].qs;
+        const uint8_t * qh = x[i].qh;
+
+        for (int ib = 0; ib < QK_K/32; ++ib) {
+            const float dl1 = d * (2*((sc[ib/2] >> (6*(ib%2)+0)) & 0x7) + 1);
+            const float dl2 = d * (2*((sc[ib/2] >> (6*(ib%2)+3)) & 0x7) + 1);
+
+            idx[0] = qs[0] | ((qh[0] << 8) & 0x700);
+            idx[1] = qs[1] | ((qh[0] << 4) & 0x700);
+            idx[2] = qs[2] | ((qh[1] << 8) & 0x700);
+            idx[3] = qs[3] | ((qh[1] << 4) & 0x700);
+            delta[0] = qh[0] & 0x08 ? -IQ1S_DELTA : IQ1S_DELTA;
+            delta[1] = qh[0] & 0x80 ? -IQ1S_DELTA : IQ1S_DELTA;
+            delta[2] = qh[1] & 0x08 ? -IQ1S_DELTA : IQ1S_DELTA;
+            delta[3] = qh[1] & 0x80 ? -IQ1S_DELTA : IQ1S_DELTA;
+            for (int l = 0; l < 2; ++l) {
+                const int8_t * grid = (const int8_t *)(iq1s_grid + idx[l]);
+                for (int j = 0; j < 8; ++j) {
+                    y[j] = dl1 * (grid[j] + delta[l]);
+                }
+                y += 8;
+            }
+            for (int l = 2; l < 4; ++l) {
+                const int8_t * grid = (const int8_t *)(iq1s_grid + idx[l]);
+                for (int j = 0; j < 8; ++j) {
+                    y[j] = dl2 * (grid[j] + delta[l]);
+                }
+                y += 8;
+            }
+            qs += 4;
+            qh += 2;
+        }
+    }
+}
+
+static const int8_t kvalues_iq4nl[16] = {-127, -104, -83, -65, -49, -35, -22, -10, 1, 13, 25, 38, 53, 69, 89, 113};
+
+void dequantize_row_iq4_nl(const block_iq4_nl * restrict x, float * restrict y, int64_t k) {
+    assert(k % QK4_NL == 0);
+    const int64_t nb = k / QK4_NL;
+
+    for (int i = 0; i < nb; i++) {
+
+        const uint8_t * qs = x[i].qs;
+
+        const float d = GGML_FP16_TO_FP32(x[i].d);
+        for (int j = 0; j < QK4_NL/2; ++j) {
+            y[j+       0] = d * kvalues_iq4nl[qs[j] & 0xf];
+            y[j+QK4_NL/2] = d * kvalues_iq4nl[qs[j] >>  4];
+        }
+        y  += QK4_NL;
+        qs += QK4_NL/2;
+    }
+}
+
+void dequantize_row_iq4_xs(const block_iq4_xs * restrict x, float * restrict y, int64_t k) {
+    assert(k % QK_K == 0);
+    const int64_t nb = k / QK_K;
+
+    for (int i = 0; i < nb; i++) {
+
+        const uint8_t * qs = x[i].qs;
+
+        const float d = GGML_FP16_TO_FP32(x[i].d);
+
+        for (int ib = 0; ib < QK_K/32; ++ib) {
+            const int ls = ((x[i].scales_l[ib/2] >> 4*(ib%2)) & 0xf) | (((x[i].scales_h >> 2*ib) & 3) << 4);
+            const float dl = d * (ls - 32);
+            for (int j = 0; j < 16; ++j) {
+                y[j+ 0] = dl * kvalues_iq4nl[qs[j] & 0xf];
+                y[j+16] = dl * kvalues_iq4nl[qs[j] >>  4];
+            }
+            y  += 32;
+            qs += 16;
+        }
+    }
+}
+
+//===================================== Q8_K ==============================================
+
+void quantize_row_q8_K_ref(const float * restrict x, block_q8_K * restrict y, int64_t k) {
+    assert(k % QK_K == 0);
+    const int64_t nb = k / QK_K;
+
+    for (int i = 0; i < nb; i++) {
+
+        float max = 0;
+        float amax = 0;
+        for (int j = 0; j < QK_K; ++j) {
+            float ax = fabsf(x[j]);
+            if (ax > amax) {
+                amax = ax; max = x[j];
+            }
+        }
+        if (!amax) {
+            y[i].d = 0;
+            memset(y[i].qs, 0, QK_K);
+            x += QK_K;
+            continue;
+        }
+        //const float iscale = -128.f/max;
+        // We need this change for IQ2_XXS, else the AVX implementation becomes very awkward
+        const float iscale = -127.f/max;
+        for (int j = 0; j < QK_K; ++j) {
+            int v = nearest_int(iscale*x[j]);
+            y[i].qs[j] = MIN(127, v);
+        }
+        for (int j = 0; j < QK_K/16; ++j) {
+            int sum = 0;
+            for (int ii = 0; ii < 16; ++ii) {
+                sum += y[i].qs[j*16 + ii];
+            }
+            y[i].bsums[j] = sum;
+        }
+        y[i].d = 1/iscale;
+        x += QK_K;
+    }
+}
+
+void dequantize_row_q8_K(const block_q8_K * restrict x, float * restrict y, int64_t k) {
+    assert(k % QK_K == 0);
+    const int64_t nb = k / QK_K;
+
+    for (int i = 0; i < nb; i++) {
+        for (int j = 0; j < QK_K; ++j) {
+            *y++ = x[i].d * x[i].qs[j];
+        }
+    }
+}
+
+void quantize_row_q8_K(const float * restrict x, void * restrict y, int64_t k) {
+    quantize_row_q8_K_ref(x, y, k);
+}
+
+//===================================== Dot products =================================
+
+//
+// Helper functions
+//
+#if __AVX__ || __AVX2__ || __AVX512F__
+
+// shuffles to pick the required scales in dot products
+static inline __m256i get_scale_shuffle_q3k(int i) {
+    static const uint8_t k_shuffle[128] = {
+         0, 1, 0, 1, 0, 1, 0, 1, 0, 1, 0, 1, 0, 1, 0, 1,     2, 3, 2, 3, 2, 3, 2, 3, 2, 3, 2, 3, 2, 3, 2, 3,
+         4, 5, 4, 5, 4, 5, 4, 5, 4, 5, 4, 5, 4, 5, 4, 5,     6, 7, 6, 7, 6, 7, 6, 7, 6, 7, 6, 7, 6, 7, 6, 7,
+         8, 9, 8, 9, 8, 9, 8, 9, 8, 9, 8, 9, 8, 9, 8, 9,    10,11,10,11,10,11,10,11,10,11,10,11,10,11,10,11,
+        12,13,12,13,12,13,12,13,12,13,12,13,12,13,12,13,    14,15,14,15,14,15,14,15,14,15,14,15,14,15,14,15,
+    };
+    return _mm256_loadu_si256((const __m256i*)k_shuffle + i);
+}
+static inline __m256i get_scale_shuffle_k4(int i) {
+    static const uint8_t k_shuffle[256] = {
+         0, 1, 0, 1, 0, 1, 0, 1, 0, 1, 0, 1, 0, 1, 0, 1, 0, 1, 0, 1, 0, 1, 0, 1, 0, 1, 0, 1, 0, 1, 0, 1,
+         2, 3, 2, 3, 2, 3, 2, 3, 2, 3, 2, 3, 2, 3, 2, 3, 2, 3, 2, 3, 2, 3, 2, 3, 2, 3, 2, 3, 2, 3, 2, 3,
+         4, 5, 4, 5, 4, 5, 4, 5, 4, 5, 4, 5, 4, 5, 4, 5, 4, 5, 4, 5, 4, 5, 4, 5, 4, 5, 4, 5, 4, 5, 4, 5,
+         6, 7, 6, 7, 6, 7, 6, 7, 6, 7, 6, 7, 6, 7, 6, 7, 6, 7, 6, 7, 6, 7, 6, 7, 6, 7, 6, 7, 6, 7, 6, 7,
+         8, 9, 8, 9, 8, 9, 8, 9, 8, 9, 8, 9, 8, 9, 8, 9, 8, 9, 8, 9, 8, 9, 8, 9, 8, 9, 8, 9, 8, 9, 8, 9,
+        10,11,10,11,10,11,10,11,10,11,10,11,10,11,10,11,10,11,10,11,10,11,10,11,10,11,10,11,10,11,10,11,
+        12,13,12,13,12,13,12,13,12,13,12,13,12,13,12,13,12,13,12,13,12,13,12,13,12,13,12,13,12,13,12,13,
+        14,15,14,15,14,15,14,15,14,15,14,15,14,15,14,15,14,15,14,15,14,15,14,15,14,15,14,15,14,15,14,15
+    };
+    return _mm256_loadu_si256((const __m256i*)k_shuffle + i);
+}
+static inline __m128i get_scale_shuffle(int i) {
+    static const uint8_t k_shuffle[128] = {
+         0, 0, 0, 0, 0, 0, 0, 0, 1, 1, 1, 1, 1, 1, 1, 1,
+         2, 2, 2, 2, 2, 2, 2, 2, 3, 3, 3, 3, 3, 3, 3, 3,
+         4, 4, 4, 4, 4, 4, 4, 4, 5, 5, 5, 5, 5, 5, 5, 5,
+         6, 6, 6, 6, 6, 6, 6, 6, 7, 7, 7, 7, 7, 7, 7, 7,
+         8, 8, 8, 8, 8, 8, 8, 8, 9, 9, 9, 9, 9, 9, 9, 9,
+        10,10,10,10,10,10,10,10, 11,11,11,11,11,11,11,11,
+        12,12,12,12,12,12,12,12, 13,13,13,13,13,13,13,13,
+        14,14,14,14,14,14,14,14, 15,15,15,15,15,15,15,15
+    };
+    return _mm_loadu_si128((const __m128i*)k_shuffle + i);
+}
+#elif defined(__loongarch_asx)
+// shuffles to pick the required scales in dot products
+static inline __m256i get_scale_shuffle_q3k(int i) {
+    static const uint8_t k_shuffle[128] = {
+         0, 1, 0, 1, 0, 1, 0, 1, 0, 1, 0, 1, 0, 1, 0, 1,     2, 3, 2, 3, 2, 3, 2, 3, 2, 3, 2, 3, 2, 3, 2, 3,
+         4, 5, 4, 5, 4, 5, 4, 5, 4, 5, 4, 5, 4, 5, 4, 5,     6, 7, 6, 7, 6, 7, 6, 7, 6, 7, 6, 7, 6, 7, 6, 7,
+         8, 9, 8, 9, 8, 9, 8, 9, 8, 9, 8, 9, 8, 9, 8, 9,    10,11,10,11,10,11,10,11,10,11,10,11,10,11,10,11,
+        12,13,12,13,12,13,12,13,12,13,12,13,12,13,12,13,    14,15,14,15,14,15,14,15,14,15,14,15,14,15,14,15,
+    };
+    return __lasx_xvld((const __m256i*)k_shuffle + i, 0);
+}
+static inline __m256i get_scale_shuffle_k4(int i) {
+    static const uint8_t k_shuffle[256] = {
+         0, 1, 0, 1, 0, 1, 0, 1, 0, 1, 0, 1, 0, 1, 0, 1, 0, 1, 0, 1, 0, 1, 0, 1, 0, 1, 0, 1, 0, 1, 0, 1,
+         2, 3, 2, 3, 2, 3, 2, 3, 2, 3, 2, 3, 2, 3, 2, 3, 2, 3, 2, 3, 2, 3, 2, 3, 2, 3, 2, 3, 2, 3, 2, 3,
+         4, 5, 4, 5, 4, 5, 4, 5, 4, 5, 4, 5, 4, 5, 4, 5, 4, 5, 4, 5, 4, 5, 4, 5, 4, 5, 4, 5, 4, 5, 4, 5,
+         6, 7, 6, 7, 6, 7, 6, 7, 6, 7, 6, 7, 6, 7, 6, 7, 6, 7, 6, 7, 6, 7, 6, 7, 6, 7, 6, 7, 6, 7, 6, 7,
+         8, 9, 8, 9, 8, 9, 8, 9, 8, 9, 8, 9, 8, 9, 8, 9, 8, 9, 8, 9, 8, 9, 8, 9, 8, 9, 8, 9, 8, 9, 8, 9,
+        10,11,10,11,10,11,10,11,10,11,10,11,10,11,10,11,10,11,10,11,10,11,10,11,10,11,10,11,10,11,10,11,
+        12,13,12,13,12,13,12,13,12,13,12,13,12,13,12,13,12,13,12,13,12,13,12,13,12,13,12,13,12,13,12,13,
+        14,15,14,15,14,15,14,15,14,15,14,15,14,15,14,15,14,15,14,15,14,15,14,15,14,15,14,15,14,15,14,15
+    };
+    return __lasx_xvld((const __m256i*)k_shuffle + i, 0);
+}
+static inline __m128i get_scale_shuffle(int i) {
+    static const uint8_t k_shuffle[128] = {
+         0, 0, 0, 0, 0, 0, 0, 0, 1, 1, 1, 1, 1, 1, 1, 1,
+         2, 2, 2, 2, 2, 2, 2, 2, 3, 3, 3, 3, 3, 3, 3, 3,
+         4, 4, 4, 4, 4, 4, 4, 4, 5, 5, 5, 5, 5, 5, 5, 5,
+         6, 6, 6, 6, 6, 6, 6, 6, 7, 7, 7, 7, 7, 7, 7, 7,
+         8, 8, 8, 8, 8, 8, 8, 8, 9, 9, 9, 9, 9, 9, 9, 9,
+        10,10,10,10,10,10,10,10, 11,11,11,11,11,11,11,11,
+        12,12,12,12,12,12,12,12, 13,13,13,13,13,13,13,13,
+        14,14,14,14,14,14,14,14, 15,15,15,15,15,15,15,15
+    };
+    return __lsx_vld((const __m128i*)k_shuffle + i, 0);
+}
+#endif
+
+void ggml_vec_dot_q4_0_q8_0(int n, float * restrict s, size_t bs, const void * restrict vx, size_t bx, const void * restrict vy, size_t by, int nrc) {
+    const int qk = QK8_0;
+    const int nb = n / qk;
+
+    assert(n % qk == 0);
+#if defined(__ARM_FEATURE_MATMUL_INT8)
+    assert((nrc == 2) || (nrc == 1));
+#else
+    assert(nrc == 1);
+#endif
+    UNUSED(nrc);
+    UNUSED(bx);
+    UNUSED(by);
+    UNUSED(bs);
+
+    const block_q4_0 * restrict x = vx;
+    const block_q8_0 * restrict y = vy;
+
+#if defined(__ARM_FEATURE_MATMUL_INT8)
+    if (nrc == 2) {
+        const block_q4_0 * restrict vx0 = vx;
+        const block_q4_0 * restrict vx1 = (const block_q4_0 *) ((const uint8_t*)vx + bx);
+        const block_q8_0 * restrict vy0 = vy;
+        const block_q8_0 * restrict vy1 = (const block_q8_0 *) ((const uint8_t*)vy + by);
+
+        float32x4_t sumv0 = vdupq_n_f32(0.0f);
+
+        for (int i = 0; i < nb; i++) {
+            const block_q4_0 * restrict b_x0 = &vx0[i];
+            const block_q4_0 * restrict b_x1 = &vx1[i];
+            const block_q8_0 * restrict b_y0 = &vy0[i];
+            const block_q8_0 * restrict b_y1 = &vy1[i];
+
+            const uint8x16_t m4b = vdupq_n_u8(0x0F);
+            const int8x16_t  s8b = vdupq_n_s8(0x8);
+
+            const uint8x16_t v0_0 = vld1q_u8(b_x0->qs);
+            const uint8x16_t v0_1 = vld1q_u8(b_x1->qs);
+
+            // 4-bit -> 8-bit
+            const int8x16_t v0_0l = vreinterpretq_s8_u8(vandq_u8  (v0_0, m4b));
+            const int8x16_t v0_0h = vreinterpretq_s8_u8(vshrq_n_u8(v0_0, 4));
+            const int8x16_t v0_1l = vreinterpretq_s8_u8(vandq_u8  (v0_1, m4b));
+            const int8x16_t v0_1h = vreinterpretq_s8_u8(vshrq_n_u8(v0_1, 4));
+
+            // sub 8
+            const int8x16_t x0_l = vsubq_s8(v0_0l, s8b);
+            const int8x16_t x0_h = vsubq_s8(v0_0h, s8b);
+            const int8x16_t x1_l = vsubq_s8(v0_1l, s8b);
+            const int8x16_t x1_h = vsubq_s8(v0_1h, s8b);
+
+            // load y
+            const int8x16_t y0_l = vld1q_s8(b_y0->qs);
+            const int8x16_t y0_h = vld1q_s8(b_y0->qs + 16);
+            const int8x16_t y1_l = vld1q_s8(b_y1->qs);
+            const int8x16_t y1_h = vld1q_s8(b_y1->qs + 16);
+
+            float32_t _scale[4] = { GGML_FP16_TO_FP32(b_x0->d)*GGML_FP16_TO_FP32(b_y0->d),
+                                    GGML_FP16_TO_FP32(b_x0->d)*GGML_FP16_TO_FP32(b_y1->d),
+                                    GGML_FP16_TO_FP32(b_x1->d)*GGML_FP16_TO_FP32(b_y0->d),
+                                    GGML_FP16_TO_FP32(b_x1->d)*GGML_FP16_TO_FP32(b_y1->d)};
+
+            float32x4_t scale = vld1q_f32(_scale);
+
+            int8x16_t l0 = vreinterpretq_s8_s64(vzip1q_s64(vreinterpretq_s64_s8(x0_l), vreinterpretq_s64_s8(x1_l)));
+            int8x16_t l1 = vreinterpretq_s8_s64(vzip2q_s64(vreinterpretq_s64_s8(x0_l), vreinterpretq_s64_s8(x1_l)));
+
+            int8x16_t l2 = vreinterpretq_s8_s64(vzip1q_s64(vreinterpretq_s64_s8(x0_h), vreinterpretq_s64_s8(x1_h)));
+            int8x16_t l3 = vreinterpretq_s8_s64(vzip2q_s64(vreinterpretq_s64_s8(x0_h), vreinterpretq_s64_s8(x1_h)));
+
+            int8x16_t r0 = vreinterpretq_s8_s64(vzip1q_s64(vreinterpretq_s64_s8(y0_l), vreinterpretq_s64_s8(y1_l)));
+            int8x16_t r1 = vreinterpretq_s8_s64(vzip2q_s64(vreinterpretq_s64_s8(y0_l), vreinterpretq_s64_s8(y1_l)));
+
+            int8x16_t r2 = vreinterpretq_s8_s64(vzip1q_s64(vreinterpretq_s64_s8(y0_h), vreinterpretq_s64_s8(y1_h)));
+            int8x16_t r3 = vreinterpretq_s8_s64(vzip2q_s64(vreinterpretq_s64_s8(y0_h), vreinterpretq_s64_s8(y1_h)));
+
+            sumv0 = vmlaq_f32(sumv0,(vcvtq_f32_s32(vmmlaq_s32((vmmlaq_s32((vmmlaq_s32((vmmlaq_s32(vdupq_n_s32(0), l0, r0)),
+                                                                                l1, r1)), l2, r2)), l3, r3))), scale);
+        }
+        float32x4_t sumv1 = vextq_f32(sumv0, sumv0, 2);
+        float32x4_t sumv2 = vzip1q_f32(sumv0, sumv1);
+
+        vst1_f32(s,      vget_low_f32(sumv2));
+        vst1_f32(s + bs, vget_high_f32(sumv2));
+        return;
+    }
+#endif
+
+    int ib = 0;
+    float sumf = 0;
+
+#if defined(__ARM_FEATURE_SVE)
+    svfloat32_t sumv0 = svdup_n_f32(0.0f);
+    svfloat32_t sumv1 = svdup_n_f32(0.0f);
+
+    const int vector_length = ggml_cpu_get_sve_cnt()*8;
+
+    // VLA Implementation using switch case
+    switch (vector_length) {
+        case 128:
+            {
+                // predicate for activating higher lanes for 4 float32 elements
+                const svbool_t ph4 = svptrue_pat_b32(SV_VL4);
+
+                for (; ib + 1 < nb; ib += 2) {
+                    const block_q4_0 * restrict x0 = &x[ib + 0];
+                    const block_q4_0 * restrict x1 = &x[ib + 1];
+                    const block_q8_0 * restrict y0 = &y[ib + 0];
+                    const block_q8_0 * restrict y1 = &y[ib + 1];
+
+                    // load x
+                    const svuint8_t qx0r = svld1rq_u8(svptrue_b8(), x0->qs);
+                    const svuint8_t qx1r = svld1rq_u8(svptrue_b8(), x1->qs);
+
+                    // 4-bit -> 8-bit
+                    const svint8_t qx0l = svreinterpret_s8_u8(svand_n_u8_m(svptrue_b8(), qx0r, 0x0F));
+                    const svint8_t qx0h = svreinterpret_s8_u8(svlsr_n_u8_m(svptrue_b8(), qx0r, 0x04));
+                    const svint8_t qx1l = svreinterpret_s8_u8(svand_n_u8_m(svptrue_b8(), qx1r, 0x0F));
+                    const svint8_t qx1h = svreinterpret_s8_u8(svlsr_n_u8_m(svptrue_b8(), qx1r, 0x04));
+
+                    // sub 8
+                    const svint8_t qx0ls = svsub_n_s8_x(svptrue_b8(), qx0h, 8);
+                    const svint8_t qx0hs = svsub_n_s8_x(svptrue_b8(), qx0l, 8);
+                    const svint8_t qx1ls = svsub_n_s8_x(svptrue_b8(), qx1h, 8);
+                    const svint8_t qx1hs = svsub_n_s8_x(svptrue_b8(), qx1l, 8);
+
+                    // load y
+                    const svint8_t qy0h = svld1_s8(svptrue_b8(), y0->qs);
+                    const svint8_t qy0l = svld1_s8(svptrue_b8(), y0->qs + 16);
+                    const svint8_t qy1h = svld1_s8(svptrue_b8(), y1->qs);
+                    const svint8_t qy1l = svld1_s8(svptrue_b8(), y1->qs + 16);
+
+                    // dot product
+                    sumv0 = svmla_n_f32_x(ph4, sumv0, svcvt_f32_s32_x(ph4, svadd_x(ph4,
+                                    svdot_s32(svdup_n_s32(0), qx0ls, qy0l),
+                                    svdot_s32(svdup_n_s32(0), qx0hs, qy0h))), GGML_FP16_TO_FP32(x0->d)*GGML_FP16_TO_FP32(y0->d));
+                    sumv1 = svmla_n_f32_x(ph4, sumv1, svcvt_f32_s32_x(ph4, svadd_x(ph4,
+                                    svdot_s32(svdup_n_s32(0), qx1ls, qy1l),
+                                    svdot_s32(svdup_n_s32(0), qx1hs, qy1h))), GGML_FP16_TO_FP32(x1->d)*GGML_FP16_TO_FP32(y1->d));
+                }
+
+                sumf = svaddv_f32(svptrue_b32(), svadd_f32_x(svptrue_b32(), sumv0, sumv1));
+            } break;
+        case 256:
+            {
+                // predicate for activating higher lanes for 16 int8 elements
+                const svbool_t ph16 = svptrue_pat_b8(SV_VL16);
+                // predicate for activating lower lanes for  16 int8 elements
+                const svbool_t pl16 = svnot_b_z(svptrue_b8(), ph16);
+
+                for (; ib + 1 < nb; ib += 2) {
+                    const block_q4_0 * restrict x0 = &x[ib + 0];
+                    const block_q4_0 * restrict x1 = &x[ib + 1];
+                    const block_q8_0 * restrict y0 = &y[ib + 0];
+                    const block_q8_0 * restrict y1 = &y[ib + 1];
+
+                    // load x
+                    const svuint8_t qx0r = svld1rq_u8(svptrue_b8(), x0->qs);
+                    const svuint8_t qx1r = svld1rq_u8(svptrue_b8(), x1->qs);
+
+                    // 4-bit -> 8-bit
+                    const svint8_t qx0 = svreinterpret_s8_u8(svlsr_n_u8_m(pl16, svand_n_u8_m(ph16, qx0r, 0x0F), 0x04));
+                    const svint8_t qx1 = svreinterpret_s8_u8(svlsr_n_u8_m(pl16, svand_n_u8_m(ph16, qx1r, 0x0F), 0x04));
+
+                    // sub 8
+                    const svint8_t qx0s = svsub_n_s8_x(svptrue_b8(), qx0, 8);
+                    const svint8_t qx1s = svsub_n_s8_x(svptrue_b8(), qx1, 8);
+
+                    // load y
+                    const svint8_t qy0 = svld1_s8(svptrue_b8(), y0->qs);
+                    const svint8_t qy1 = svld1_s8(svptrue_b8(), y1->qs);
+
+                    // dot product
+                    sumv0 = svmla_n_f32_x(svptrue_b32(), sumv0, svcvt_f32_s32_x(svptrue_b32(),
+                                svdot_s32(svdup_n_s32(0), qx0s, qy0)), GGML_FP16_TO_FP32(x0->d)*GGML_FP16_TO_FP32(y0->d));
+                    sumv1 = svmla_n_f32_x(svptrue_b32(), sumv1, svcvt_f32_s32_x(svptrue_b32(),
+                                svdot_s32(svdup_n_s32(0), qx1s, qy1)), GGML_FP16_TO_FP32(x1->d)*GGML_FP16_TO_FP32(y1->d));
+                }
+
+                sumf = svaddv_f32(svptrue_b32(), svadd_f32_x(svptrue_b32(), sumv0, sumv1));
+            } break;
+        case 512:
+            {
+                // predicate for activating higher lanes for 32 int8 elements
+                const svbool_t ph32 = svptrue_pat_b8(SV_VL32);
+
+                // predicate for activating higher lanes for 16 int8 elements
+                const svbool_t ph16 = svptrue_pat_b8(SV_VL16);
+                // predicate for activating lower lanes for 16 int8 elements from first 32 int8 activated lanes
+                const svbool_t pl16 = svnot_b_z(ph32, ph16);
+
+                for (; ib + 1 < nb; ib += 2) {
+                    const block_q4_0 * restrict x0 = &x[ib + 0];
+                    const block_q4_0 * restrict x1 = &x[ib + 1];
+                    const block_q8_0 * restrict y0 = &y[ib + 0];
+                    const block_q8_0 * restrict y1 = &y[ib + 1];
+
+                    // load x
+                    const svuint8_t qx0r = svld1rq_u8(ph32, x0->qs);
+                    const svuint8_t qx1r = svld1rq_u8(ph32, x1->qs);
+
+                    // 4-bit -> 8-bit
+                    const svint8_t qx0 = svreinterpret_s8_u8(svlsr_n_u8_m(pl16, svand_n_u8_m(ph16, qx0r, 0x0F), 0x04));
+                    const svint8_t qx1 = svreinterpret_s8_u8(svlsr_n_u8_m(pl16, svand_n_u8_m(ph16, qx1r, 0x0F), 0x04));
+
+                    // sub 8
+                    const svint8_t qx0s = svsub_n_s8_x(ph32, qx0, 8);
+                    const svint8_t qx1s = svsub_n_s8_x(ph32, qx1, 8);
+
+                    // load y
+                    const svint8_t qy0 = svld1_s8(ph32, y0->qs);
+                    const svint8_t qy1 = svld1_s8(ph32, y1->qs);
+
+                    // dot product
+                    sumv0 = svmla_n_f32_x(ph32, sumv0, svcvt_f32_s32_x(ph32,
+                                svdot_s32(svdup_n_s32(0), qx0s, qy0)), GGML_FP16_TO_FP32(x0->d)*GGML_FP16_TO_FP32(y0->d));
+                    sumv1 = svmla_n_f32_x(ph32, sumv1, svcvt_f32_s32_x(ph32,
+                                svdot_s32(svdup_n_s32(0), qx1s, qy1)), GGML_FP16_TO_FP32(x1->d)*GGML_FP16_TO_FP32(y1->d));
+                }
+
+                sumf = svaddv_f32(ph32, svadd_f32_x(ph32, sumv0, sumv1));
+            } break;
+        default:
+            assert(false && "Unsupported vector length");
+            break;
+    }
+
+#elif defined(__ARM_NEON)
+    float32x4_t sumv0 = vdupq_n_f32(0.0f);
+    float32x4_t sumv1 = vdupq_n_f32(0.0f);
+
+    for (; ib + 1 < nb; ib += 2) {
+        const block_q4_0 * restrict x0 = &x[ib + 0];
+        const block_q4_0 * restrict x1 = &x[ib + 1];
+        const block_q8_0 * restrict y0 = &y[ib + 0];
+        const block_q8_0 * restrict y1 = &y[ib + 1];
+
+        const uint8x16_t m4b = vdupq_n_u8(0x0F);
+        const int8x16_t  s8b = vdupq_n_s8(0x8);
+
+        const uint8x16_t v0_0 = vld1q_u8(x0->qs);
+        const uint8x16_t v0_1 = vld1q_u8(x1->qs);
+
+        // 4-bit -> 8-bit
+        const int8x16_t v0_0l = vreinterpretq_s8_u8(vandq_u8  (v0_0, m4b));
+        const int8x16_t v0_0h = vreinterpretq_s8_u8(vshrq_n_u8(v0_0, 4));
+        const int8x16_t v0_1l = vreinterpretq_s8_u8(vandq_u8  (v0_1, m4b));
+        const int8x16_t v0_1h = vreinterpretq_s8_u8(vshrq_n_u8(v0_1, 4));
+
+        // sub 8
+        const int8x16_t v0_0ls = vsubq_s8(v0_0l, s8b);
+        const int8x16_t v0_0hs = vsubq_s8(v0_0h, s8b);
+        const int8x16_t v0_1ls = vsubq_s8(v0_1l, s8b);
+        const int8x16_t v0_1hs = vsubq_s8(v0_1h, s8b);
+
+        // load y
+        const int8x16_t v1_0l = vld1q_s8(y0->qs);
+        const int8x16_t v1_0h = vld1q_s8(y0->qs + 16);
+        const int8x16_t v1_1l = vld1q_s8(y1->qs);
+        const int8x16_t v1_1h = vld1q_s8(y1->qs + 16);
+
+        // dot product into int32x4_t
+        const int32x4_t p_0 = ggml_vdotq_s32(ggml_vdotq_s32(vdupq_n_s32(0), v0_0ls, v1_0l), v0_0hs, v1_0h);
+        const int32x4_t p_1 = ggml_vdotq_s32(ggml_vdotq_s32(vdupq_n_s32(0), v0_1ls, v1_1l), v0_1hs, v1_1h);
+
+        sumv0 = vmlaq_n_f32(sumv0, vcvtq_f32_s32(p_0), GGML_FP16_TO_FP32(x0->d)*GGML_FP16_TO_FP32(y0->d));
+        sumv1 = vmlaq_n_f32(sumv1, vcvtq_f32_s32(p_1), GGML_FP16_TO_FP32(x1->d)*GGML_FP16_TO_FP32(y1->d));
+    }
+
+    sumf = vaddvq_f32(sumv0) + vaddvq_f32(sumv1);
+#elif defined(__AVX2__)
+    // Initialize accumulator with zeros
+    __m256 acc = _mm256_setzero_ps();
+
+    // Main loop
+    for (; ib < nb; ++ib) {
+        /* Compute combined scale for the block */
+        const __m256 d = _mm256_set1_ps( GGML_FP16_TO_FP32(x[ib].d) * GGML_FP16_TO_FP32(y[ib].d) );
+
+        __m256i qx = bytes_from_nibbles_32(x[ib].qs);
+
+        // Now we have a vector with bytes in [ 0 .. 15 ] interval. Offset them into [ -8 .. +7 ] interval.
+        const __m256i off = _mm256_set1_epi8( 8 );
+        qx = _mm256_sub_epi8( qx, off );
+
+        __m256i qy = _mm256_loadu_si256((const __m256i *)y[ib].qs);
+
+        const __m256 q = mul_sum_i8_pairs_float(qx, qy);
+
+        /* Multiply q with scale and accumulate */
+        acc = _mm256_fmadd_ps( d, q, acc );
+    }
+
+    sumf = hsum_float_8(acc);
+#elif defined(__AVX__)
+    const __m128i mone = _mm_set1_epi16(1);
+
+    __m256 accum1 = _mm256_setzero_ps();
+    __m256 accum2 = _mm256_setzero_ps();
+    for (; ib + 1 < nb; ib += 2) {
+        const __m128i q4bits_1 = _mm_loadu_si128((const __m128i *)x[ib + 0].qs);
+        const __m128i q4bits_2 = _mm_loadu_si128((const __m128i *)x[ib + 1].qs);
+        const __m128i q8b_1_0 = _mm_loadu_si128((const __m128i *)y[ib + 0].qs);
+        const __m128i q8b_1_1 = _mm_loadu_si128((const __m128i *)y[ib + 0].qs + 1);
+        const __m128i q8b_2_0 = _mm_loadu_si128((const __m128i *)y[ib + 1].qs);
+        const __m128i q8b_2_1 = _mm_loadu_si128((const __m128i *)y[ib + 1].qs + 1);
+
+        const __m128i q4b_1_0 = _mm_sub_epi8(_mm_and_si128(_mm_set1_epi8(15), q4bits_1), _mm_set1_epi8(8));
+        const __m128i q4b_1_1 = _mm_sub_epi8(_mm_and_si128(_mm_set1_epi8(15), _mm_srli_epi16(q4bits_1, 4)), _mm_set1_epi8(8));
+        const __m128i q4b_2_0 = _mm_sub_epi8(_mm_and_si128(_mm_set1_epi8(15), q4bits_2), _mm_set1_epi8(8));
+        const __m128i q4b_2_1 = _mm_sub_epi8(_mm_and_si128(_mm_set1_epi8(15), _mm_srli_epi16(q4bits_2, 4)), _mm_set1_epi8(8));
+        const __m128i p16_1_0 = mul_add_epi8_sse(q4b_1_0, q8b_1_0);
+        const __m128i p16_1_1 = mul_add_epi8_sse(q4b_1_1, q8b_1_1);
+        const __m128i p16_2_0 = mul_add_epi8_sse(q4b_2_0, q8b_2_0);
+        const __m128i p16_2_1 = mul_add_epi8_sse(q4b_2_1, q8b_2_1);
+        const __m128i p_1_0 = _mm_madd_epi16(p16_1_0, mone);
+        const __m128i p_1_1 = _mm_madd_epi16(p16_1_1, mone);
+        const __m128i p_2_0 = _mm_madd_epi16(p16_2_0, mone);
+        const __m128i p_2_1 = _mm_madd_epi16(p16_2_1, mone);
+        accum1 = _mm256_add_ps(_mm256_mul_ps(_mm256_set1_ps(GGML_FP16_TO_FP32(y[ib + 0].d)*GGML_FP16_TO_FP32(x[ib + 0].d)),
+                _mm256_cvtepi32_ps(MM256_SET_M128I(p_1_1, p_1_0))), accum1);
+        accum2 = _mm256_add_ps(_mm256_mul_ps(_mm256_set1_ps(GGML_FP16_TO_FP32(y[ib + 1].d)*GGML_FP16_TO_FP32(x[ib + 1].d)),
+                _mm256_cvtepi32_ps(MM256_SET_M128I(p_2_1, p_2_0))), accum2);
+    }
+
+    sumf = hsum_float_8(_mm256_add_ps(accum1, accum2));
+#elif defined(__SSSE3__)
+    // set constants
+    const __m128i lowMask = _mm_set1_epi8(0xF);
+    const __m128i off = _mm_set1_epi8(8);
+
+    // Initialize accumulator with zeros
+    __m128 acc_0 = _mm_setzero_ps();
+    __m128 acc_1 = _mm_setzero_ps();
+    __m128 acc_2 = _mm_setzero_ps();
+    __m128 acc_3 = _mm_setzero_ps();
+
+    for (; ib + 1 < nb; ib += 2) {
+        _mm_prefetch(&x[ib] + sizeof(block_q4_0), _MM_HINT_T0);
+        _mm_prefetch(&y[ib] + sizeof(block_q8_0), _MM_HINT_T0);
+
+        // Compute combined scale for the block 0 and 1
+        const __m128 d_0_1 = _mm_set1_ps( GGML_FP16_TO_FP32(x[ib].d) * GGML_FP16_TO_FP32(y[ib].d) );
+
+        const __m128i tmp_0_1 = _mm_loadu_si128((const __m128i *)x[ib].qs);
+
+        __m128i bx_0 = _mm_and_si128(lowMask, tmp_0_1);
+        __m128i by_0 = _mm_loadu_si128((const __m128i *)y[ib].qs);
+        bx_0 = _mm_sub_epi8(bx_0, off);
+        const __m128i i32_0 = mul_sum_i8_pairs(bx_0, by_0);
+
+        __m128i bx_1 = _mm_and_si128(lowMask, _mm_srli_epi64(tmp_0_1, 4));
+        __m128i by_1 = _mm_loadu_si128((const __m128i *)(y[ib].qs + 16));
+        bx_1 = _mm_sub_epi8(bx_1, off);
+        const __m128i i32_1 = mul_sum_i8_pairs(bx_1, by_1);
+
+        _mm_prefetch(&x[ib] + 2 * sizeof(block_q4_0), _MM_HINT_T0);
+        _mm_prefetch(&y[ib] + 2 * sizeof(block_q8_0), _MM_HINT_T0);
+
+        // Compute combined scale for the block 2 and 3
+        const __m128 d_2_3 = _mm_set1_ps( GGML_FP16_TO_FP32(x[ib + 1].d) * GGML_FP16_TO_FP32(y[ib + 1].d) );
+
+        const __m128i tmp_2_3 = _mm_loadu_si128((const __m128i *)x[ib + 1].qs);
+
+        __m128i bx_2 = _mm_and_si128(lowMask, tmp_2_3);
+        __m128i by_2 = _mm_loadu_si128((const __m128i *)y[ib + 1].qs);
+        bx_2 = _mm_sub_epi8(bx_2, off);
+        const __m128i i32_2 = mul_sum_i8_pairs(bx_2, by_2);
+
+        __m128i bx_3 = _mm_and_si128(lowMask, _mm_srli_epi64(tmp_2_3, 4));
+        __m128i by_3 = _mm_loadu_si128((const __m128i *)(y[ib + 1].qs + 16));
+        bx_3 = _mm_sub_epi8(bx_3, off);
+        const __m128i i32_3 = mul_sum_i8_pairs(bx_3, by_3);
+
+        // Convert int32_t to float
+        __m128 p0 = _mm_cvtepi32_ps(i32_0);
+        __m128 p1 = _mm_cvtepi32_ps(i32_1);
+        __m128 p2 = _mm_cvtepi32_ps(i32_2);
+        __m128 p3 = _mm_cvtepi32_ps(i32_3);
+
+        // Apply the scale
+        __m128 p0_d = _mm_mul_ps( d_0_1, p0 );
+        __m128 p1_d = _mm_mul_ps( d_0_1, p1 );
+        __m128 p2_d = _mm_mul_ps( d_2_3, p2 );
+        __m128 p3_d = _mm_mul_ps( d_2_3, p3 );
+
+        // Acummulate
+        acc_0 = _mm_add_ps(p0_d, acc_0);
+        acc_1 = _mm_add_ps(p1_d, acc_1);
+        acc_2 = _mm_add_ps(p2_d, acc_2);
+        acc_3 = _mm_add_ps(p3_d, acc_3);
+    }
+
+    sumf = hsum_float_4x4(acc_0, acc_1, acc_2, acc_3);
+#elif defined(__riscv_v_intrinsic)
+    size_t vl = __riscv_vsetvl_e8m1(qk/2);
+
+    for (; ib < nb; ++ib) {
+        // load elements
+        vuint8mf2_t tx = __riscv_vle8_v_u8mf2(x[ib].qs, vl);
+
+        vint8mf2_t y0 = __riscv_vle8_v_i8mf2(y[ib].qs, vl);
+        vint8mf2_t y1 = __riscv_vle8_v_i8mf2(y[ib].qs+16, vl);
+
+        // mask and store lower part of x, and then upper part
+        vuint8mf2_t x_a = __riscv_vand_vx_u8mf2(tx, 0x0F, vl);
+        vuint8mf2_t x_l = __riscv_vsrl_vx_u8mf2(tx, 0x04, vl);
+
+        vint8mf2_t x_ai = __riscv_vreinterpret_v_u8mf2_i8mf2(x_a);
+        vint8mf2_t x_li = __riscv_vreinterpret_v_u8mf2_i8mf2(x_l);
+
+        // subtract offset
+        vint8mf2_t v0 = __riscv_vsub_vx_i8mf2(x_ai, 8, vl);
+        vint8mf2_t v1 = __riscv_vsub_vx_i8mf2(x_li, 8, vl);
+
+        vint16m1_t vec_mul1 = __riscv_vwmul_vv_i16m1(v0, y0, vl);
+        vint16m1_t vec_mul2 = __riscv_vwmul_vv_i16m1(v1, y1, vl);
+
+        vint32m1_t vec_zero = __riscv_vmv_v_x_i32m1(0, vl);
+
+        vint32m1_t vs1 = __riscv_vwredsum_vs_i16m1_i32m1(vec_mul1, vec_zero, vl);
+        vint32m1_t vs2 = __riscv_vwredsum_vs_i16m1_i32m1(vec_mul2, vs1, vl);
+
+        int sumi = __riscv_vmv_x_s_i32m1_i32(vs2);
+
+        sumf += sumi*GGML_FP16_TO_FP32(x[ib].d)*GGML_FP16_TO_FP32(y[ib].d);
+    }
+
+#elif defined(__POWER9_VECTOR__)
+    const vector signed char lowMask = vec_splats((signed char)0xF);
+    const vector signed int v0 = vec_splats((int32_t)0);
+    const vector unsigned char v4 = vec_splats((unsigned char)0x4);
+    const vector signed char v8 = vec_splats((signed char)0x8);
+
+    vector float vsumf0 = vec_splats(0.0f);
+
+#pragma GCC unroll 8
+    for (; ib < nb; ++ib) {
+        __builtin_prefetch(x[ib].qs, 0, 1);
+        __builtin_prefetch(y[ib].qs, 0, 1);
+
+        vector float vxd = vec_splats(GGML_FP16_TO_FP32(x[ib].d));
+        vector float vyd = vec_splats(GGML_FP16_TO_FP32(y[ib].d));
+        vector float vd = vec_mul(vxd, vyd);
+
+        vector signed char qxs = (vector signed char)vec_xl( 0, x[ib].qs);
+        vector signed char q8y0 = vec_xl( 0, y[ib].qs);
+        vector signed char q8y1 = vec_xl(16, y[ib].qs);
+
+        vector signed char q4x0 = vec_and(qxs, lowMask);
+        vector signed char q4x1 = vec_sr(qxs, v4);
+
+        q4x0 = vec_sub(q4x0, v8);
+        q4x1 = vec_sub(q4x1, v8);
+
+        vector signed short qv0 = vec_add(vec_mule(q4x0, q8y0), vec_mulo(q4x0, q8y0));
+        vector signed short qv1 = vec_add(vec_mule(q4x1, q8y1), vec_mulo(q4x1, q8y1));
+
+        vector signed int vsumi0 = v0;
+
+        vsumi0 = vec_sum4s(qv0, vsumi0);
+        vsumi0 = vec_sum4s(qv1, vsumi0);
+
+        vsumf0 = vec_madd(vec_ctf(vsumi0, 0), vd, vsumf0);
+    }
+
+    vsumf0 = vec_add(vsumf0, vec_sld(vsumf0, vsumf0, 4));
+    vsumf0 = vec_add(vsumf0, vec_sld(vsumf0, vsumf0, 8));
+
+    sumf = vec_extract(vsumf0, 0);
+
+#elif defined(__loongarch_asx)
+    // Initialize accumulator with zeros
+    __m256 acc = (__m256)__lasx_xvldi(0);
+
+    // Main loop
+    for (; ib < nb; ++ib) {
+        /* Compute combined scale for the block */
+        const __m256 d = __lasx_xvreplfr2vr_s( GGML_FP16_TO_FP32(x[ib].d) * GGML_FP16_TO_FP32(y[ib].d) );
+
+        __m256i qx = bytes_from_nibbles_32(x[ib].qs);
+
+        // Now we have a vector with bytes in [ 0 .. 15 ] interval. Offset them into [ -8 .. +7 ] interval.
+        const __m256i off = __lasx_xvreplgr2vr_b( 8 );
+        qx = __lasx_xvsub_b( qx, off );
+
+        __m256i qy = __lasx_xvld((const __m256i *)y[ib].qs, 0);
+
+        const __m256 q = mul_sum_i8_pairs_float(qx, qy);
+
+        /* Multiply q with scale and accumulate */
+        acc = __lasx_xvfmadd_s( d, q, acc );
+    }
+
+    sumf = hsum_float_8(acc);
+#elif defined(__loongarch_sx)
+    // set constants
+    const __m128i low_mask = __lsx_vreplgr2vr_b(0xF);
+    const __m128i off = __lsx_vreplgr2vr_b(8);
+
+    // Initialize accumulator with zeros
+    __m128 acc_0 = __lsx_vldi(0);
+    __m128 acc_1 = __lsx_vldi(0);
+    __m128 acc_2 = __lsx_vldi(0);
+    __m128 acc_3 = __lsx_vldi(0);
+
+    for (; ib + 1 < nb; ib += 2) {
+
+        // Compute combined scale for the block 0 and 1
+        const __m128 d_0_1 = __lsx_vreplgr2vr_w( GGML_FP16_TO_FP32(x[ib].d) * GGML_FP16_TO_FP32(y[ib].d) );
+
+        const __m128i tmp_0_1 = __lsx_vld((const __m128i *)x[ib].qs, 0);
+
+        __m128i bx_0 = __lsx_vand_v(low_mask, tmp_0_1);
+        __m128i by_0 = __lsx_vld((const __m128i *)y[ib].qs, 0);
+        bx_0 = __lsx_vsub_b(bx_0, off);
+        const __m128i i32_0 = mul_sum_i8_pairs(bx_0, by_0);
+
+        __m128i bx_1 = __lsx_vand_v(low_mask, __lsx_vsrli_d(tmp_0_1, 4));
+        __m128i by_1 = __lsx_vld((const __m128i *)(y[ib].qs + 16), 0);
+        bx_1 = __lsx_vsub_b(bx_1, off);
+        const __m128i i32_1 = mul_sum_i8_pairs(bx_1, by_1);
+
+        //_mm_prefetch(&x[ib] + 2 * sizeof(block_q4_0), _MM_HINT_T0);
+        //_mm_prefetch(&y[ib] + 2 * sizeof(block_q8_0), _MM_HINT_T0);
+
+        // Compute combined scale for the block 2 and 3
+        const __m128 d_2_3 = __lsx_vreplgr2vr_w( GGML_FP16_TO_FP32(x[ib + 1].d) * GGML_FP16_TO_FP32(y[ib + 1].d) );
+
+        const __m128i tmp_2_3 = __lsx_vld((const __m128i *)x[ib + 1].qs, 0);
+
+        __m128i bx_2 = __lsx_vand_v(low_mask, tmp_2_3);
+        __m128i by_2 = __lsx_vld((const __m128i *)y[ib + 1].qs, 0);
+        bx_2 = __lsx_vsub_b(bx_2, off);
+        const __m128i i32_2 = mul_sum_i8_pairs(bx_2, by_2);
+
+        __m128i bx_3 = __lsx_vand_v(low_mask, __lsx_vsrli_d(tmp_2_3, 4));
+        __m128i by_3 = __lsx_vld((const __m128i *)(y[ib + 1].qs + 16), 0);
+        bx_3 = __lsx_vsub_b(bx_3, off);
+        const __m128i i32_3 = mul_sum_i8_pairs(bx_3, by_3);
+
+        // Convert int32_t to float
+        __m128 p0 = __lsx_vffint_s_w(i32_0);
+        __m128 p1 = __lsx_vffint_s_w(i32_1);
+        __m128 p2 = __lsx_vffint_s_w(i32_2);
+        __m128 p3 = __lsx_vffint_s_w(i32_3);
+
+        // Apply the scale
+        __m128 p0_d = __lsx_vfmul_s( d_0_1, p0 );
+        __m128 p1_d = __lsx_vfmul_s( d_0_1, p1 );
+        __m128 p2_d = __lsx_vfmul_s( d_2_3, p2 );
+        __m128 p3_d = __lsx_vfmul_s( d_2_3, p3 );
+
+        // Acummulate
+        acc_0 = __lsx_vfadd_s(p0_d, acc_0);
+        acc_1 = __lsx_vfadd_s(p1_d, acc_1);
+        acc_2 = __lsx_vfadd_s(p2_d, acc_2);
+        acc_3 = __lsx_vfadd_s(p3_d, acc_3);
+    }
+
+    sumf = hsum_float_4x4(acc_0, acc_1, acc_2, acc_3);
+#endif
+    for (; ib < nb; ++ib) {
+        int sumi0 = 0;
+        int sumi1 = 0;
+
+        for (int j = 0; j < qk/2; ++j) {
+            const int v0 = (x[ib].qs[j] & 0x0F) - 8;
+            const int v1 = (x[ib].qs[j] >>   4) - 8;
+
+            sumi0 += (v0 * y[ib].qs[j]);
+            sumi1 += (v1 * y[ib].qs[j + qk/2]);
+        }
+
+        int sumi = sumi0 + sumi1;
+        sumf += sumi*GGML_FP16_TO_FP32(x[ib].d)*GGML_FP16_TO_FP32(y[ib].d);
+    }
+
+    *s = sumf;
+}
+
+void ggml_vec_dot_q4_1_q8_1(int n, float * restrict s, size_t bs, const void * restrict vx, size_t bx, const void * restrict vy, size_t by, int nrc) {
+    const int qk = QK8_1;
+    const int nb = n / qk;
+
+    assert(n % qk == 0);
+#if defined(__ARM_FEATURE_MATMUL_INT8)
+    assert((nrc == 2) || (nrc == 1));
+#else
+    assert(nrc == 1);
+#endif
+    UNUSED(nrc);
+    UNUSED(bx);
+    UNUSED(by);
+    UNUSED(bs);
+
+    const block_q4_1 * restrict x = vx;
+    const block_q8_1 * restrict y = vy;
+
+#if defined(__ARM_FEATURE_MATMUL_INT8)
+    if (nrc == 2) {
+        const block_q4_1 * restrict vx0 = vx;
+        const block_q4_1 * restrict vx1 = (const block_q4_1 *) ((const uint8_t*)vx + bx);
+        const block_q8_1 * restrict vy0 = vy;
+        const block_q8_1 * restrict vy1 = (const block_q8_1 *) ((const uint8_t*)vy + by);
+
+        float32x4_t sumv0 = vdupq_n_f32(0.0f);
+        float32x4_t summs0 = vdupq_n_f32(0.0f);
+
+        for (int i = 0; i < nb; i++) {
+            const block_q4_1 * restrict b_x0 = &vx0[i];
+            const block_q4_1 * restrict b_x1 = &vx1[i];
+            const block_q8_1 * restrict b_y0 = &vy0[i];
+            const block_q8_1 * restrict b_y1 = &vy1[i];
+
+            float32_t summs_t[4] = {GGML_FP16_TO_FP32(b_x0->m) * GGML_FP16_TO_FP32(b_y0->s),
+                                    GGML_FP16_TO_FP32(b_x1->m) * GGML_FP16_TO_FP32(b_y0->s),
+                                    GGML_FP16_TO_FP32(b_x0->m) * GGML_FP16_TO_FP32(b_y1->s),
+                                    GGML_FP16_TO_FP32(b_x1->m) * GGML_FP16_TO_FP32(b_y1->s)};
+            summs0 = vaddq_f32(summs0, vld1q_f32(summs_t));
+
+            const uint8x16_t m4b = vdupq_n_u8(0x0F);
+
+            const uint8x16_t v0_0 = vld1q_u8(b_x0->qs);
+            const uint8x16_t v0_1 = vld1q_u8(b_x1->qs);
+
+            // 4-bit -> 8-bit
+            const int8x16_t x0_l = vreinterpretq_s8_u8(vandq_u8  (v0_0, m4b));
+            const int8x16_t x0_h = vreinterpretq_s8_u8(vshrq_n_u8(v0_0, 4));
+            const int8x16_t x1_l = vreinterpretq_s8_u8(vandq_u8  (v0_1, m4b));
+            const int8x16_t x1_h = vreinterpretq_s8_u8(vshrq_n_u8(v0_1, 4));
+
+            // load y
+            const int8x16_t y0_l = vld1q_s8(b_y0->qs);
+            const int8x16_t y0_h = vld1q_s8(b_y0->qs + 16);
+            const int8x16_t y1_l = vld1q_s8(b_y1->qs);
+            const int8x16_t y1_h = vld1q_s8(b_y1->qs + 16);
+
+            // mmla into int32x4_t
+            float32_t _scale[4] = {GGML_FP16_TO_FP32(b_x0->d)*b_y0->d,
+                                   GGML_FP16_TO_FP32(b_x0->d)*b_y1->d,
+                                   GGML_FP16_TO_FP32(b_x1->d)*b_y0->d,
+                                   GGML_FP16_TO_FP32(b_x1->d)*b_y1->d};
+            float32x4_t scale = vld1q_f32(_scale);
+
+            int8x16_t l0 = vreinterpretq_s8_s64(vzip1q_s64(vreinterpretq_s64_s8(x0_l), vreinterpretq_s64_s8(x1_l)));
+            int8x16_t l1 = vreinterpretq_s8_s64(vzip2q_s64(vreinterpretq_s64_s8(x0_l), vreinterpretq_s64_s8(x1_l)));
+
+            int8x16_t l2 = vreinterpretq_s8_s64(vzip1q_s64(vreinterpretq_s64_s8(x0_h), vreinterpretq_s64_s8(x1_h)));
+            int8x16_t l3 = vreinterpretq_s8_s64(vzip2q_s64(vreinterpretq_s64_s8(x0_h), vreinterpretq_s64_s8(x1_h)));
+
+            int8x16_t r0 = vreinterpretq_s8_s64(vzip1q_s64(vreinterpretq_s64_s8(y0_l), vreinterpretq_s64_s8(y1_l)));
+            int8x16_t r1 = vreinterpretq_s8_s64(vzip2q_s64(vreinterpretq_s64_s8(y0_l), vreinterpretq_s64_s8(y1_l)));
+
+            int8x16_t r2 = vreinterpretq_s8_s64(vzip1q_s64(vreinterpretq_s64_s8(y0_h), vreinterpretq_s64_s8(y1_h)));
+            int8x16_t r3 = vreinterpretq_s8_s64(vzip2q_s64(vreinterpretq_s64_s8(y0_h), vreinterpretq_s64_s8(y1_h)));
+            sumv0 = vmlaq_f32(sumv0,(vcvtq_f32_s32(vmmlaq_s32((vmmlaq_s32((vmmlaq_s32((vmmlaq_s32(vdupq_n_s32(0), l0, r0)),
+                                                                                l1, r1)), l2, r2)), l3, r3))), scale);
+        }
+
+        float32x4_t sumv1 = vextq_f32(sumv0, sumv0, 2);
+        float32x4_t sumv2 = vzip1q_f32(sumv0, sumv1);
+        sumv2 = vaddq_f32(sumv2, summs0);
+
+        vst1_f32(s,      vget_low_f32 (sumv2));
+        vst1_f32(s + bs, vget_high_f32(sumv2));
+        return;
+    }
+#endif
+
+    int ib = 0;
+    float sumf = 0;
+
+    // TODO: add WASM SIMD
+#if defined(__ARM_NEON)
+    float32x4_t sumv0 = vdupq_n_f32(0.0f);
+    float32x4_t sumv1 = vdupq_n_f32(0.0f);
+
+    float summs = 0;
+
+    for (; ib + 1 < nb; ib += 2) {
+        const block_q4_1 * restrict x0 = &x[ib + 0];
+        const block_q4_1 * restrict x1 = &x[ib + 1];
+        const block_q8_1 * restrict y0 = &y[ib + 0];
+        const block_q8_1 * restrict y1 = &y[ib + 1];
+
+        summs += GGML_FP16_TO_FP32(x0->m) * GGML_FP16_TO_FP32(y0->s) + GGML_FP16_TO_FP32(x1->m) * GGML_FP16_TO_FP32(y1->s);
+
+        const uint8x16_t m4b = vdupq_n_u8(0x0F);
+
+        const uint8x16_t v0_0 = vld1q_u8(x0->qs);
+        const uint8x16_t v0_1 = vld1q_u8(x1->qs);
+
+        // 4-bit -> 8-bit
+        const int8x16_t v0_0l = vreinterpretq_s8_u8(vandq_u8  (v0_0, m4b));
+        const int8x16_t v0_0h = vreinterpretq_s8_u8(vshrq_n_u8(v0_0, 4));
+        const int8x16_t v0_1l = vreinterpretq_s8_u8(vandq_u8  (v0_1, m4b));
+        const int8x16_t v0_1h = vreinterpretq_s8_u8(vshrq_n_u8(v0_1, 4));
+
+        // load y
+        const int8x16_t v1_0l = vld1q_s8(y0->qs);
+        const int8x16_t v1_0h = vld1q_s8(y0->qs + 16);
+        const int8x16_t v1_1l = vld1q_s8(y1->qs);
+        const int8x16_t v1_1h = vld1q_s8(y1->qs + 16);
+
+        // dot product into int32x4_t
+        const int32x4_t p_0 = ggml_vdotq_s32(ggml_vdotq_s32(vdupq_n_s32(0), v0_0l, v1_0l), v0_0h, v1_0h);
+        const int32x4_t p_1 = ggml_vdotq_s32(ggml_vdotq_s32(vdupq_n_s32(0), v0_1l, v1_1l), v0_1h, v1_1h);
+
+        sumv0 = vmlaq_n_f32(sumv0, vcvtq_f32_s32(p_0), GGML_FP16_TO_FP32(x0->d)*GGML_FP16_TO_FP32(y0->d));
+        sumv1 = vmlaq_n_f32(sumv1, vcvtq_f32_s32(p_1), GGML_FP16_TO_FP32(x1->d)*GGML_FP16_TO_FP32(y1->d));
+    }
+
+    sumf = vaddvq_f32(sumv0) + vaddvq_f32(sumv1) + summs;
+#elif defined(__AVX2__) || defined(__AVX__)
+    // Initialize accumulator with zeros
+    __m256 acc = _mm256_setzero_ps();
+
+    float summs = 0;
+
+    // Main loop
+    for (; ib < nb; ++ib) {
+        const float d0 = GGML_FP16_TO_FP32(x[ib].d);
+        const float d1 = GGML_FP16_TO_FP32(y[ib].d);
+
+        summs += GGML_FP16_TO_FP32(x[ib].m) * GGML_FP16_TO_FP32(y[ib].s);
+
+        const __m256 d0v = _mm256_set1_ps( d0 );
+        const __m256 d1v = _mm256_set1_ps( d1 );
+
+        // Compute combined scales
+        const __m256 d0d1 = _mm256_mul_ps( d0v, d1v );
+
+        // Load 16 bytes, and unpack 4 bit fields into bytes, making 32 bytes
+        const __m256i qx = bytes_from_nibbles_32(x[ib].qs);
+        const __m256i qy = _mm256_loadu_si256( (const __m256i *)y[ib].qs );
+
+        const __m256 xy = mul_sum_us8_pairs_float(qx, qy);
+
+        // Accumulate d0*d1*x*y
+#if defined(__AVX2__)
+        acc = _mm256_fmadd_ps( d0d1, xy, acc );
+#else
+        acc = _mm256_add_ps( _mm256_mul_ps( d0d1, xy ), acc );
+#endif
+    }
+
+    sumf = hsum_float_8(acc) + summs;
+#elif defined(__riscv_v_intrinsic)
+    size_t vl = __riscv_vsetvl_e8m1(qk/2);
+
+    for (; ib < nb; ++ib) {
+        // load elements
+        vuint8mf2_t tx = __riscv_vle8_v_u8mf2(x[ib].qs, vl);
+
+        vint8mf2_t y0 = __riscv_vle8_v_i8mf2(y[ib].qs, vl);
+        vint8mf2_t y1 = __riscv_vle8_v_i8mf2(y[ib].qs+16, vl);
+
+        // mask and store lower part of x, and then upper part
+        vuint8mf2_t x_a = __riscv_vand_vx_u8mf2(tx, 0x0F, vl);
+        vuint8mf2_t x_l = __riscv_vsrl_vx_u8mf2(tx, 0x04, vl);
+
+        vint8mf2_t v0 = __riscv_vreinterpret_v_u8mf2_i8mf2(x_a);
+        vint8mf2_t v1 = __riscv_vreinterpret_v_u8mf2_i8mf2(x_l);
+
+        vint16m1_t vec_mul1 = __riscv_vwmul_vv_i16m1(v0, y0, vl);
+        vint16m1_t vec_mul2 = __riscv_vwmul_vv_i16m1(v1, y1, vl);
+
+        vint32m1_t vec_zero = __riscv_vmv_v_x_i32m1(0, vl);
+
+        vint32m1_t vs1 = __riscv_vwredsum_vs_i16m1_i32m1(vec_mul1, vec_zero, vl);
+        vint32m1_t vs2 = __riscv_vwredsum_vs_i16m1_i32m1(vec_mul2, vs1, vl);
+
+        int sumi = __riscv_vmv_x_s_i32m1_i32(vs2);
+
+        sumf += (GGML_FP16_TO_FP32(x[ib].d)*GGML_FP16_TO_FP32(y[ib].d))*sumi + GGML_FP16_TO_FP32(x[ib].m)*GGML_FP16_TO_FP32(y[ib].s);
+    }
+
+#elif defined(__POWER9_VECTOR__)
+    const vector signed char lowMask = vec_splats((signed char)0xF);
+    const vector signed int v0 = vec_splats((int32_t)0);
+    const vector unsigned char v4 = vec_splats((unsigned char)0x4);
+
+    vector float vsumf0 = vec_splats(0.0f);
+
+#pragma GCC unroll 4
+    for (; ib < nb; ++ib) {
+        __builtin_prefetch(x[ib].qs, 0, 1);
+        __builtin_prefetch(y[ib].qs, 0, 1);
+
+        vector float vxd = vec_splats(GGML_FP16_TO_FP32(x[ib].d));
+        vector float vyd = vec_splats(GGML_FP16_TO_FP32(y[ib].d));
+        vector float vd = vec_mul(vxd, vyd);
+
+        vector float vxmin = vec_splats(GGML_FP16_TO_FP32(x[ib].m));
+        vector float vys = {GGML_FP16_TO_FP32(y[ib].s), 0.0f, 0.0f, 0.0f};
+        vsumf0 = vec_madd(vxmin, vys, vsumf0);
+
+        vector signed char qxs = (vector signed char)vec_xl( 0, x[ib].qs);
+        vector signed char q8y0 = vec_xl( 0, y[ib].qs);
+        vector signed char q8y1 = vec_xl(16, y[ib].qs);
+
+        vector unsigned char q4x0 = (vector unsigned char)vec_and(qxs, lowMask);
+        vector unsigned char q4x1 = (vector unsigned char)vec_sr(qxs, v4);
+
+        vector signed int vsumi0 = v0;
+
+        vsumi0 = vec_msum(q8y0, q4x0, vsumi0);
+        vsumi0 = vec_msum(q8y1, q4x1, vsumi0);
+
+        vsumf0 = vec_madd(vec_ctf(vsumi0, 0), vd, vsumf0);
+    }
+
+    vsumf0 = vec_add(vsumf0, vec_sld(vsumf0, vsumf0, 4));
+    vsumf0 = vec_add(vsumf0, vec_sld(vsumf0, vsumf0, 8));
+
+    sumf = vec_extract(vsumf0, 0);
+
+#elif defined(__loongarch_asx)
+    // Initialize accumulator with zeros
+    __m256 acc = (__m256)__lasx_xvldi(0);
+
+    float summs = 0;
+
+    // Main loop
+    for (; ib < nb; ++ib) {
+        const float d0 = GGML_FP16_TO_FP32(x[ib].d);
+        const float d1 = GGML_FP16_TO_FP32(y[ib].d);
+
+        summs += GGML_FP16_TO_FP32(x[ib].m) * GGML_FP16_TO_FP32(y[ib].s);
+
+        const __m256 d0v = __lasx_xvreplfr2vr_s( d0 );
+        const __m256 d1v = __lasx_xvreplfr2vr_s( d1 );
+
+        // Compute combined scales
+        const __m256 d0d1 = __lasx_xvfmul_s( d0v, d1v );
+
+        // Load 16 bytes, and unpack 4 bit fields into bytes, making 32 bytes
+        const __m256i qx = bytes_from_nibbles_32(x[ib].qs);
+        const __m256i qy = __lasx_xvld( (const __m256i *)y[ib].qs, 0);
+
+        const __m256 xy = mul_sum_us8_pairs_float(qx, qy);
+
+        // Accumulate d0*d1*x*y
+        acc = __lasx_xvfmadd_s( d0d1, xy, acc );
+    }
+
+    sumf = hsum_float_8(acc) + summs;
+#endif
+    for (; ib < nb; ++ib) {
+        int sumi0 = 0;
+        int sumi1 = 0;
+
+        for (int j = 0; j < qk/2; ++j) {
+            const int v0 = (x[ib].qs[j] & 0x0F);
+            const int v1 = (x[ib].qs[j] >>   4);
+
+            sumi0 += (v0 * y[ib].qs[j]);
+            sumi1 += (v1 * y[ib].qs[j + qk/2]);
+        }
+
+        int sumi = sumi0 + sumi1;
+        sumf += (GGML_FP16_TO_FP32(x[ib].d)*GGML_FP16_TO_FP32(y[ib].d))*sumi + GGML_FP16_TO_FP32(x[ib].m)*GGML_FP16_TO_FP32(y[ib].s);
+    }
+
+    *s = sumf;
+}
+
+void ggml_vec_dot_q5_0_q8_0(int n, float * restrict s, size_t bs, const void * restrict vx, size_t bx, const void * restrict vy, size_t by, int nrc) {
+    const int qk = QK8_0;
+    const int nb = n / qk;
+
+    int ib = 0;
+    float sumf = 0;
+
+    assert(n % qk == 0);
+    assert(qk == QK5_0);
+    assert(nrc == 1);
+    UNUSED(nrc);
+    UNUSED(bx);
+    UNUSED(by);
+    UNUSED(bs);
+
+    const block_q5_0 * restrict x = vx;
+    const block_q8_0 * restrict y = vy;
+
+#if defined(__ARM_NEON)
+    float32x4_t sumv0 = vdupq_n_f32(0.0f);
+    float32x4_t sumv1 = vdupq_n_f32(0.0f);
+
+    uint32_t qh0;
+    uint32_t qh1;
+
+    uint64_t tmp0[4];
+    uint64_t tmp1[4];
+
+    for (; ib + 1 < nb; ib += 2) {
+        const block_q5_0 * restrict x0 = &x[ib];
+        const block_q5_0 * restrict x1 = &x[ib + 1];
+        const block_q8_0 * restrict y0 = &y[ib];
+        const block_q8_0 * restrict y1 = &y[ib + 1];
+
+        const uint8x16_t m4b = vdupq_n_u8(0x0F);
+
+        // extract the 5th bit via lookup table ((!b) << 4)
+        memcpy(&qh0, x0->qh, sizeof(qh0));
+        memcpy(&qh1, x1->qh, sizeof(qh1));
+
+        tmp0[0] = table_b2b_1[(qh0 >>  0) & 0xFF];
+        tmp0[1] = table_b2b_1[(qh0 >>  8) & 0xFF];
+        tmp0[2] = table_b2b_1[(qh0 >> 16) & 0xFF];
+        tmp0[3] = table_b2b_1[(qh0 >> 24)       ];
+
+        tmp1[0] = table_b2b_1[(qh1 >>  0) & 0xFF];
+        tmp1[1] = table_b2b_1[(qh1 >>  8) & 0xFF];
+        tmp1[2] = table_b2b_1[(qh1 >> 16) & 0xFF];
+        tmp1[3] = table_b2b_1[(qh1 >> 24)       ];
+
+        const int8x16_t qhl0 = vld1q_s8((const int8_t *)(tmp0 + 0));
+        const int8x16_t qhh0 = vld1q_s8((const int8_t *)(tmp0 + 2));
+        const int8x16_t qhl1 = vld1q_s8((const int8_t *)(tmp1 + 0));
+        const int8x16_t qhh1 = vld1q_s8((const int8_t *)(tmp1 + 2));
+
+        const uint8x16_t v0_0 = vld1q_u8(x0->qs);
+        const uint8x16_t v0_1 = vld1q_u8(x1->qs);
+
+        // 4-bit -> 8-bit
+        int8x16_t v0_0l = vreinterpretq_s8_u8(vandq_u8  (v0_0, m4b));
+        int8x16_t v0_0h = vreinterpretq_s8_u8(vshrq_n_u8(v0_0, 4));
+        int8x16_t v0_1l = vreinterpretq_s8_u8(vandq_u8  (v0_1, m4b));
+        int8x16_t v0_1h = vreinterpretq_s8_u8(vshrq_n_u8(v0_1, 4));
+
+        // add high bit and sub 16 (equivalent to sub 0x10 when bit is zero)
+        const int8x16_t v0_0lf = vsubq_s8(v0_0l, qhl0);
+        const int8x16_t v0_0hf = vsubq_s8(v0_0h, qhh0);
+        const int8x16_t v0_1lf = vsubq_s8(v0_1l, qhl1);
+        const int8x16_t v0_1hf = vsubq_s8(v0_1h, qhh1);
+
+        // load y
+        const int8x16_t v1_0l = vld1q_s8(y0->qs);
+        const int8x16_t v1_0h = vld1q_s8(y0->qs + 16);
+        const int8x16_t v1_1l = vld1q_s8(y1->qs);
+        const int8x16_t v1_1h = vld1q_s8(y1->qs + 16);
+
+        sumv0 = vmlaq_n_f32(sumv0, vcvtq_f32_s32(vaddq_s32(
+                        ggml_vdotq_s32(vdupq_n_s32(0), v0_0lf, v1_0l),
+                        ggml_vdotq_s32(vdupq_n_s32(0), v0_0hf, v1_0h))), GGML_FP16_TO_FP32(x0->d)*GGML_FP16_TO_FP32(y0->d));
+        sumv1 = vmlaq_n_f32(sumv1, vcvtq_f32_s32(vaddq_s32(
+                        ggml_vdotq_s32(vdupq_n_s32(0), v0_1lf, v1_1l),
+                        ggml_vdotq_s32(vdupq_n_s32(0), v0_1hf, v1_1h))), GGML_FP16_TO_FP32(x1->d)*GGML_FP16_TO_FP32(y1->d));
+    }
+
+    sumf = vaddvq_f32(sumv0) + vaddvq_f32(sumv1);
+#elif defined(__wasm_simd128__)
+    v128_t sumv = wasm_f32x4_splat(0.0f);
+
+    uint32_t qh;
+    uint64_t tmp[4];
+
+    // TODO: check if unrolling this is better
+    for (; ib < nb; ++ib) {
+        const block_q5_0 * restrict x0 = &x[ib];
+        const block_q8_0 * restrict y0 = &y[ib];
+
+        const v128_t m4b  = wasm_i8x16_splat(0x0F);
+
+        // extract the 5th bit
+        memcpy(&qh, x0->qh, sizeof(qh));
+
+        tmp[0] = table_b2b_1[(qh >>  0) & 0xFF];
+        tmp[1] = table_b2b_1[(qh >>  8) & 0xFF];
+        tmp[2] = table_b2b_1[(qh >> 16) & 0xFF];
+        tmp[3] = table_b2b_1[(qh >> 24)       ];
+
+        const v128_t qhl = wasm_v128_load(tmp + 0);
+        const v128_t qhh = wasm_v128_load(tmp + 2);
+
+        const v128_t v0 = wasm_v128_load(x0->qs);
+
+        // 4-bit -> 8-bit
+        const v128_t v0l = wasm_v128_and (v0, m4b);
+        const v128_t v0h = wasm_u8x16_shr(v0, 4);
+
+        // add high bit and sub 16 (equivalent to sub 0x10 when bit is zero)
+        const v128_t v0lf = wasm_i8x16_sub(v0l, qhl);
+        const v128_t v0hf = wasm_i8x16_sub(v0h, qhh);
+
+        // load y
+        const v128_t v1l = wasm_v128_load(y0->qs);
+        const v128_t v1h = wasm_v128_load(y0->qs + 16);
+
+        // int8x16 -> int16x8
+        const v128_t v0lfl = wasm_i16x8_extend_low_i8x16 (v0lf);
+        const v128_t v0lfh = wasm_i16x8_extend_high_i8x16(v0lf);
+        const v128_t v0hfl = wasm_i16x8_extend_low_i8x16 (v0hf);
+        const v128_t v0hfh = wasm_i16x8_extend_high_i8x16(v0hf);
+
+        const v128_t v1ll = wasm_i16x8_extend_low_i8x16 (v1l);
+        const v128_t v1lh = wasm_i16x8_extend_high_i8x16(v1l);
+        const v128_t v1hl = wasm_i16x8_extend_low_i8x16 (v1h);
+        const v128_t v1hh = wasm_i16x8_extend_high_i8x16(v1h);
+
+        // dot product
+        sumv = wasm_f32x4_add(sumv, wasm_f32x4_mul(wasm_f32x4_convert_i32x4(
+                        wasm_i32x4_add(
+                            wasm_i32x4_add(wasm_i32x4_dot_i16x8(v0lfl, v1ll),
+                                           wasm_i32x4_dot_i16x8(v0lfh, v1lh)),
+                            wasm_i32x4_add(wasm_i32x4_dot_i16x8(v0hfl, v1hl),
+                                           wasm_i32x4_dot_i16x8(v0hfh, v1hh)))),
+                    wasm_f32x4_splat(GGML_FP16_TO_FP32(x0->d) * GGML_FP16_TO_FP32(y0->d))));
+    }
+
+    sumf = wasm_f32x4_extract_lane(sumv, 0) + wasm_f32x4_extract_lane(sumv, 1) +
+           wasm_f32x4_extract_lane(sumv, 2) + wasm_f32x4_extract_lane(sumv, 3);
+#elif defined(__AVX2__)
+    // Initialize accumulator with zeros
+    __m256 acc = _mm256_setzero_ps();
+
+    // Main loop
+    for (; ib < nb; ++ib) {
+        /* Compute combined scale for the block */
+        const __m256 d = _mm256_set1_ps(GGML_FP16_TO_FP32(x[ib].d) * GGML_FP16_TO_FP32(y[ib].d));
+
+        __m256i qx = bytes_from_nibbles_32(x[ib].qs);
+        __m256i bxhi = bytes_from_bits_32(x[ib].qh);
+        bxhi = _mm256_andnot_si256(bxhi, _mm256_set1_epi8((char)0xF0));
+        qx = _mm256_or_si256(qx, bxhi);
+
+        __m256i qy = _mm256_loadu_si256((const __m256i *)y[ib].qs);
+
+        const __m256 q = mul_sum_i8_pairs_float(qx, qy);
+
+        /* Multiply q with scale and accumulate */
+        acc = _mm256_fmadd_ps(d, q, acc);
+    }
+
+    sumf = hsum_float_8(acc);
+#elif defined(__AVX__)
+    // Initialize accumulator with zeros
+    __m256 acc = _mm256_setzero_ps();
+    __m128i mask = _mm_set1_epi8((char)0xF0);
+
+    // Main loop
+    for (; ib < nb; ++ib) {
+        /* Compute combined scale for the block */
+        const __m256 d = _mm256_set1_ps(GGML_FP16_TO_FP32(x[ib].d) * GGML_FP16_TO_FP32(y[ib].d));
+
+        __m256i bx_0 = bytes_from_nibbles_32(x[ib].qs);
+        const __m256i bxhi = bytes_from_bits_32(x[ib].qh);
+        __m128i bxhil = _mm256_castsi256_si128(bxhi);
+        __m128i bxhih = _mm256_extractf128_si256(bxhi, 1);
+        bxhil = _mm_andnot_si128(bxhil, mask);
+        bxhih = _mm_andnot_si128(bxhih, mask);
+        __m128i bxl = _mm256_castsi256_si128(bx_0);
+        __m128i bxh = _mm256_extractf128_si256(bx_0, 1);
+        bxl = _mm_or_si128(bxl, bxhil);
+        bxh = _mm_or_si128(bxh, bxhih);
+        bx_0 = MM256_SET_M128I(bxh, bxl);
+
+        const __m256i by_0 = _mm256_loadu_si256((const __m256i *)y[ib].qs);
+
+        const __m256 q = mul_sum_i8_pairs_float(bx_0, by_0);
+
+        /* Multiply q with scale and accumulate */
+        acc = _mm256_add_ps(_mm256_mul_ps(d, q), acc);
+    }
+
+    sumf = hsum_float_8(acc);
+#elif defined(__riscv_v_intrinsic)
+    uint32_t qh;
+
+    size_t vl = __riscv_vsetvl_e8m1(qk/2);
+
+    // These temporary registers are for masking and shift operations
+    vuint32m2_t vt_1 = __riscv_vid_v_u32m2(vl);
+    vuint32m2_t vt_2 = __riscv_vsll_vv_u32m2(__riscv_vmv_v_x_u32m2(1, vl), vt_1, vl);
+
+    vuint32m2_t vt_3 = __riscv_vsll_vx_u32m2(vt_2, 16, vl);
+    vuint32m2_t vt_4 = __riscv_vadd_vx_u32m2(vt_1, 12, vl);
+
+    for (; ib < nb; ++ib) {
+        memcpy(&qh, x[ib].qh, sizeof(uint32_t));
+
+        // ((qh & (1u << (j + 0 ))) >> (j + 0 )) << 4;
+        vuint32m2_t xha_0 = __riscv_vand_vx_u32m2(vt_2, qh, vl);
+        vuint32m2_t xhr_0 = __riscv_vsrl_vv_u32m2(xha_0, vt_1, vl);
+        vuint32m2_t xhl_0 = __riscv_vsll_vx_u32m2(xhr_0, 4, vl);
+
+        // ((qh & (1u << (j + 16))) >> (j + 12));
+        vuint32m2_t xha_1 = __riscv_vand_vx_u32m2(vt_3, qh, vl);
+        vuint32m2_t xhl_1 = __riscv_vsrl_vv_u32m2(xha_1, vt_4, vl);
+
+        // narrowing
+        vuint16m1_t xhc_0 = __riscv_vncvt_x_x_w_u16m1(xhl_0, vl);
+        vuint8mf2_t xh_0 = __riscv_vncvt_x_x_w_u8mf2(xhc_0, vl);
+
+        vuint16m1_t xhc_1 = __riscv_vncvt_x_x_w_u16m1(xhl_1, vl);
+        vuint8mf2_t xh_1 = __riscv_vncvt_x_x_w_u8mf2(xhc_1, vl);
+
+        // load
+        vuint8mf2_t tx = __riscv_vle8_v_u8mf2(x[ib].qs, vl);
+
+        vint8mf2_t y0 = __riscv_vle8_v_i8mf2(y[ib].qs, vl);
+        vint8mf2_t y1 = __riscv_vle8_v_i8mf2(y[ib].qs+16, vl);
+
+        vuint8mf2_t x_at = __riscv_vand_vx_u8mf2(tx, 0x0F, vl);
+        vuint8mf2_t x_lt = __riscv_vsrl_vx_u8mf2(tx, 0x04, vl);
+
+        vuint8mf2_t x_a = __riscv_vor_vv_u8mf2(x_at, xh_0, vl);
+        vuint8mf2_t x_l = __riscv_vor_vv_u8mf2(x_lt, xh_1, vl);
+
+        vint8mf2_t x_ai = __riscv_vreinterpret_v_u8mf2_i8mf2(x_a);
+        vint8mf2_t x_li = __riscv_vreinterpret_v_u8mf2_i8mf2(x_l);
+
+        vint8mf2_t v0 = __riscv_vsub_vx_i8mf2(x_ai, 16, vl);
+        vint8mf2_t v1 = __riscv_vsub_vx_i8mf2(x_li, 16, vl);
+
+        vint16m1_t vec_mul1 = __riscv_vwmul_vv_i16m1(v0, y0, vl);
+        vint16m1_t vec_mul2 = __riscv_vwmul_vv_i16m1(v1, y1, vl);
+
+        vint32m1_t vec_zero = __riscv_vmv_v_x_i32m1(0, vl);
+
+        vint32m1_t vs1 = __riscv_vwredsum_vs_i16m1_i32m1(vec_mul1, vec_zero, vl);
+        vint32m1_t vs2 = __riscv_vwredsum_vs_i16m1_i32m1(vec_mul2, vs1, vl);
+
+        int sumi = __riscv_vmv_x_s_i32m1_i32(vs2);
+
+        sumf += (GGML_FP16_TO_FP32(x[ib].d)*GGML_FP16_TO_FP32(y[ib].d)) * sumi;
+    }
+
+#elif defined(__POWER9_VECTOR__)
+    const vector signed char lowMask = vec_splats((signed char)0xF);
+    const vector unsigned char v4 = vec_splats((unsigned char)4);
+
+    vector float vsumf0 = vec_splats(0.0f);
+
+#pragma GCC unroll 4
+    for (; ib < nb; ++ib) {
+        __builtin_prefetch(x[ib].qs, 0, 1);
+        __builtin_prefetch(y[ib].qs, 0, 1);
+
+        vector float vxd = vec_splats(GGML_FP16_TO_FP32(x[ib].d));
+        vector float vyd = vec_splats(GGML_FP16_TO_FP32(y[ib].d));
+        vector float vd = vec_mul(vxd, vyd);
+
+        vector signed long long aux64x2_0 = {(uint64_t)(table_b2b_1[x[ib].qh[0]]), (uint64_t)(table_b2b_1[x[ib].qh[1]])};
+        vector signed long long aux64x2_1 = {(uint64_t)(table_b2b_1[x[ib].qh[2]]), (uint64_t)(table_b2b_1[x[ib].qh[3]])};
+
+        vector signed char qh0 = (vector signed char)aux64x2_0;
+        vector signed char qh1 = (vector signed char)aux64x2_1;
+
+        vector signed char qxs = (vector signed char)vec_xl( 0, x[ib].qs);
+
+        vector signed char q5x0 = vec_sub(vec_and (qxs, lowMask), qh0);
+        vector signed char q5x1 = vec_sub(vec_sr(qxs, v4), qh1);
+
+        vector signed char q8y0 = vec_xl(  0, y[ib].qs);
+        vector signed char q8y1 = vec_xl( 16, y[ib].qs);
+
+        vector signed short qv0 = vec_add(vec_mule(q5x0, q8y0), vec_mulo(q5x0, q8y0));
+        vector signed short qv1 = vec_add(vec_mule(q5x1, q8y1), vec_mulo(q5x1, q8y1));
+
+        qv0 = vec_add(qv0, qv1);
+
+        vector signed int vsumi0 = vec_add(vec_unpackh(qv0), vec_unpackl(qv0));
+
+        vsumf0 = vec_madd(vec_ctf(vsumi0, 0), vd, vsumf0);
+    }
+
+    vsumf0 = vec_add(vsumf0, vec_sld(vsumf0, vsumf0, 4));
+    vsumf0 = vec_add(vsumf0, vec_sld(vsumf0, vsumf0, 8));
+
+    sumf = vec_extract(vsumf0, 0);
+
+#elif defined(__loongarch_asx)
+    // Initialize accumulator with zeros
+    __m256 acc = (__m256)__lasx_xvldi(0);
+
+    // Main loop
+    for (; ib < nb; ++ib) {
+        /* Compute combined scale for the block */
+        const __m256 d = __lasx_xvreplfr2vr_s(GGML_FP16_TO_FP32(x[ib].d) * GGML_FP16_TO_FP32(y[ib].d)); //FIXME
+
+        __m256i qx = bytes_from_nibbles_32(x[ib].qs);
+        __m256i bxhi = bytes_from_bits_32(x[ib].qh);
+        bxhi = __lasx_xvandn_v(bxhi, __lasx_xvreplgr2vr_b((char)0xF0));
+        qx = __lasx_xvor_v(qx, bxhi);
+
+        __m256i qy = __lasx_xvld((const __m256i *)y[ib].qs, 0);
+
+        const __m256 q = mul_sum_i8_pairs_float(qx, qy);
+
+        /* Multiply q with scale and accumulate */
+        acc = __lasx_xvfmadd_s(d, q, acc);
+    }
+
+    sumf = hsum_float_8(acc);
+#endif
+    for (; ib < nb; ++ib) {
+        uint32_t qh;
+        memcpy(&qh, x[ib].qh, sizeof(qh));
+
+        int sumi0 = 0;
+        int sumi1 = 0;
+
+        for (int j = 0; j < qk/2; ++j) {
+            const uint8_t xh_0 = ((qh & (1u << (j + 0 ))) >> (j + 0 )) << 4;
+            const uint8_t xh_1 = ((qh & (1u << (j + 16))) >> (j + 12));
+
+            const int32_t x0 = (int8_t)(((x[ib].qs[j] & 0x0F) | xh_0) - 16);
+            const int32_t x1 = (int8_t)(((x[ib].qs[j] >>   4) | xh_1) - 16);
+
+            sumi0 += (x0 * y[ib].qs[j]);
+            sumi1 += (x1 * y[ib].qs[j + qk/2]);
+        }
+
+        int sumi = sumi0 + sumi1;
+        sumf += (GGML_FP16_TO_FP32(x[ib].d)*GGML_FP16_TO_FP32(y[ib].d)) * sumi;
+    }
+
+    *s = sumf;
+}
+
+void ggml_vec_dot_q5_1_q8_1(int n, float * restrict s, size_t bs, const void * restrict vx, size_t bx, const void * restrict vy, size_t by, int nrc) {
+    const int qk = QK8_1;
+    const int nb = n / qk;
+
+    int ib = 0;
+    float sumf = 0;
+
+    assert(n % qk == 0);
+    assert(qk == QK5_1);
+    assert(nrc == 1);
+    UNUSED(nrc);
+    UNUSED(bx);
+    UNUSED(by);
+    UNUSED(bs);
+
+    const block_q5_1 * restrict x = vx;
+    const block_q8_1 * restrict y = vy;
+
+#if defined(__ARM_NEON)
+    float32x4_t sumv0 = vdupq_n_f32(0.0f);
+    float32x4_t sumv1 = vdupq_n_f32(0.0f);
+
+    float summs0 = 0.0f;
+    float summs1 = 0.0f;
+
+    uint32_t qh0;
+    uint32_t qh1;
+
+    uint64_t tmp0[4];
+    uint64_t tmp1[4];
+
+    for (; ib + 1 < nb; ib += 2) {
+        const block_q5_1 * restrict x0 = &x[ib];
+        const block_q5_1 * restrict x1 = &x[ib + 1];
+        const block_q8_1 * restrict y0 = &y[ib];
+        const block_q8_1 * restrict y1 = &y[ib + 1];
+
+        const uint8x16_t m4b = vdupq_n_u8(0x0F);
+
+        summs0 += GGML_FP16_TO_FP32(x0->m) * GGML_FP16_TO_FP32(y0->s);
+        summs1 += GGML_FP16_TO_FP32(x1->m) * GGML_FP16_TO_FP32(y1->s);
+
+        // extract the 5th bit via lookup table ((b) << 4)
+        memcpy(&qh0, x0->qh, sizeof(qh0));
+        memcpy(&qh1, x1->qh, sizeof(qh1));
+
+        tmp0[0] = table_b2b_0[(qh0 >>  0) & 0xFF];
+        tmp0[1] = table_b2b_0[(qh0 >>  8) & 0xFF];
+        tmp0[2] = table_b2b_0[(qh0 >> 16) & 0xFF];
+        tmp0[3] = table_b2b_0[(qh0 >> 24)       ];
+
+        tmp1[0] = table_b2b_0[(qh1 >>  0) & 0xFF];
+        tmp1[1] = table_b2b_0[(qh1 >>  8) & 0xFF];
+        tmp1[2] = table_b2b_0[(qh1 >> 16) & 0xFF];
+        tmp1[3] = table_b2b_0[(qh1 >> 24)       ];
+
+        const int8x16_t qhl0 = vld1q_s8((const int8_t *)(tmp0 + 0));
+        const int8x16_t qhh0 = vld1q_s8((const int8_t *)(tmp0 + 2));
+        const int8x16_t qhl1 = vld1q_s8((const int8_t *)(tmp1 + 0));
+        const int8x16_t qhh1 = vld1q_s8((const int8_t *)(tmp1 + 2));
+
+        const uint8x16_t v0_0 = vld1q_u8(x0->qs);
+        const uint8x16_t v0_1 = vld1q_u8(x1->qs);
+
+        // 4-bit -> 8-bit
+        const int8x16_t v0_0l = vreinterpretq_s8_u8(vandq_u8  (v0_0, m4b));
+        const int8x16_t v0_0h = vreinterpretq_s8_u8(vshrq_n_u8(v0_0, 4));
+        const int8x16_t v0_1l = vreinterpretq_s8_u8(vandq_u8  (v0_1, m4b));
+        const int8x16_t v0_1h = vreinterpretq_s8_u8(vshrq_n_u8(v0_1, 4));
+
+        // add high bit
+        const int8x16_t v0_0lf = vorrq_s8(v0_0l, qhl0);
+        const int8x16_t v0_0hf = vorrq_s8(v0_0h, qhh0);
+        const int8x16_t v0_1lf = vorrq_s8(v0_1l, qhl1);
+        const int8x16_t v0_1hf = vorrq_s8(v0_1h, qhh1);
+
+        // load y
+        const int8x16_t v1_0l = vld1q_s8(y0->qs);
+        const int8x16_t v1_0h = vld1q_s8(y0->qs + 16);
+        const int8x16_t v1_1l = vld1q_s8(y1->qs);
+        const int8x16_t v1_1h = vld1q_s8(y1->qs + 16);
+
+        sumv0 = vmlaq_n_f32(sumv0, vcvtq_f32_s32(vaddq_s32(
+                        ggml_vdotq_s32(vdupq_n_s32(0), v0_0lf, v1_0l),
+                        ggml_vdotq_s32(vdupq_n_s32(0), v0_0hf, v1_0h))), GGML_FP16_TO_FP32(x0->d)*GGML_FP16_TO_FP32(y0->d));
+        sumv1 = vmlaq_n_f32(sumv1, vcvtq_f32_s32(vaddq_s32(
+                        ggml_vdotq_s32(vdupq_n_s32(0), v0_1lf, v1_1l),
+                        ggml_vdotq_s32(vdupq_n_s32(0), v0_1hf, v1_1h))), GGML_FP16_TO_FP32(x1->d)*GGML_FP16_TO_FP32(y1->d));
+    }
+
+    sumf = vaddvq_f32(sumv0) + vaddvq_f32(sumv1) + summs0 + summs1;
+#elif defined(__wasm_simd128__)
+    v128_t sumv = wasm_f32x4_splat(0.0f);
+
+    float summs = 0.0f;
+
+    uint32_t qh;
+    uint64_t tmp[4];
+
+    // TODO: check if unrolling this is better
+    for (; ib < nb; ++ib) {
+        const block_q5_1 * restrict x0 = &x[ib];
+        const block_q8_1 * restrict y0 = &y[ib];
+
+        summs += GGML_FP16_TO_FP32(x0->m) * GGML_FP16_TO_FP32(y0->s);
+
+        const v128_t m4b = wasm_i8x16_splat(0x0F);
+
+        // extract the 5th bit
+        memcpy(&qh, x0->qh, sizeof(qh));
+
+        tmp[0] = table_b2b_0[(qh >>  0) & 0xFF];
+        tmp[1] = table_b2b_0[(qh >>  8) & 0xFF];
+        tmp[2] = table_b2b_0[(qh >> 16) & 0xFF];
+        tmp[3] = table_b2b_0[(qh >> 24)       ];
+
+        const v128_t qhl = wasm_v128_load(tmp + 0);
+        const v128_t qhh = wasm_v128_load(tmp + 2);
+
+        const v128_t v0 = wasm_v128_load(x0->qs);
+
+        // 4-bit -> 8-bit
+        const v128_t v0l = wasm_v128_and (v0, m4b);
+        const v128_t v0h = wasm_u8x16_shr(v0, 4);
+
+        // add high bit
+        const v128_t v0lf = wasm_v128_or(v0l, qhl);
+        const v128_t v0hf = wasm_v128_or(v0h, qhh);
+
+        // load y
+        const v128_t v1l = wasm_v128_load(y0->qs);
+        const v128_t v1h = wasm_v128_load(y0->qs + 16);
+
+        // int8x16 -> int16x8
+        const v128_t v0lfl = wasm_i16x8_extend_low_i8x16 (v0lf);
+        const v128_t v0lfh = wasm_i16x8_extend_high_i8x16(v0lf);
+        const v128_t v0hfl = wasm_i16x8_extend_low_i8x16 (v0hf);
+        const v128_t v0hfh = wasm_i16x8_extend_high_i8x16(v0hf);
+
+        const v128_t v1ll = wasm_i16x8_extend_low_i8x16 (v1l);
+        const v128_t v1lh = wasm_i16x8_extend_high_i8x16(v1l);
+        const v128_t v1hl = wasm_i16x8_extend_low_i8x16 (v1h);
+        const v128_t v1hh = wasm_i16x8_extend_high_i8x16(v1h);
+
+        // dot product
+        sumv = wasm_f32x4_add(sumv,
+                wasm_f32x4_mul(wasm_f32x4_convert_i32x4(wasm_i32x4_add(
+                            wasm_i32x4_add(wasm_i32x4_dot_i16x8(v0lfl, v1ll),
+                                           wasm_i32x4_dot_i16x8(v0lfh, v1lh)),
+                            wasm_i32x4_add(wasm_i32x4_dot_i16x8(v0hfl, v1hl),
+                                           wasm_i32x4_dot_i16x8(v0hfh, v1hh)))),
+                    wasm_f32x4_splat(GGML_FP16_TO_FP32(x0->d) * GGML_FP16_TO_FP32(y0->d))));
+    }
+
+    sumf = wasm_f32x4_extract_lane(sumv, 0) + wasm_f32x4_extract_lane(sumv, 1) +
+           wasm_f32x4_extract_lane(sumv, 2) + wasm_f32x4_extract_lane(sumv, 3) + summs;
+#elif defined(__AVX2__)
+    // Initialize accumulator with zeros
+    __m256 acc = _mm256_setzero_ps();
+
+    float summs = 0.0f;
+
+    // Main loop
+    for (; ib < nb; ++ib) {
+        const __m256 dx = _mm256_set1_ps(GGML_FP16_TO_FP32(x[ib].d));
+
+        summs += GGML_FP16_TO_FP32(x[ib].m) * GGML_FP16_TO_FP32(y[ib].s);
+
+        __m256i qx = bytes_from_nibbles_32(x[ib].qs);
+        __m256i bxhi = bytes_from_bits_32(x[ib].qh);
+        bxhi = _mm256_and_si256(bxhi, _mm256_set1_epi8(0x10));
+        qx = _mm256_or_si256(qx, bxhi);
+
+        const __m256 dy = _mm256_set1_ps(GGML_FP16_TO_FP32(y[ib].d));
+        const __m256i qy = _mm256_loadu_si256((const __m256i *)y[ib].qs);
+
+        const __m256 q = mul_sum_us8_pairs_float(qx, qy);
+
+        acc = _mm256_fmadd_ps(q, _mm256_mul_ps(dx, dy), acc);
+    }
+
+    sumf = hsum_float_8(acc) + summs;
+#elif defined(__AVX__)
+    // Initialize accumulator with zeros
+    __m256 acc = _mm256_setzero_ps();
+    __m128i mask = _mm_set1_epi8(0x10);
+
+    float summs = 0.0f;
+
+    // Main loop
+    for (; ib < nb; ++ib) {
+        const __m256 dx = _mm256_set1_ps(GGML_FP16_TO_FP32(x[ib].d));
+
+        summs += GGML_FP16_TO_FP32(x[ib].m) * GGML_FP16_TO_FP32(y[ib].s);
+
+        __m256i bx_0 = bytes_from_nibbles_32(x[ib].qs);
+        const __m256i bxhi = bytes_from_bits_32(x[ib].qh);
+        __m128i bxhil = _mm256_castsi256_si128(bxhi);
+        __m128i bxhih = _mm256_extractf128_si256(bxhi, 1);
+        bxhil = _mm_and_si128(bxhil, mask);
+        bxhih = _mm_and_si128(bxhih, mask);
+        __m128i bxl = _mm256_castsi256_si128(bx_0);
+        __m128i bxh = _mm256_extractf128_si256(bx_0, 1);
+        bxl = _mm_or_si128(bxl, bxhil);
+        bxh = _mm_or_si128(bxh, bxhih);
+        bx_0 = MM256_SET_M128I(bxh, bxl);
+
+        const __m256 dy = _mm256_set1_ps(GGML_FP16_TO_FP32(y[ib].d));
+        const __m256i by_0 = _mm256_loadu_si256((const __m256i *)y[ib].qs);
+
+        const __m256 q = mul_sum_us8_pairs_float(bx_0, by_0);
+
+        acc = _mm256_add_ps(_mm256_mul_ps(q, _mm256_mul_ps(dx, dy)), acc);
+    }
+
+    sumf = hsum_float_8(acc) + summs;
+#elif defined(__riscv_v_intrinsic)
+    uint32_t qh;
+
+    size_t vl = __riscv_vsetvl_e8m1(qk/2);
+
+    // temporary registers for shift operations
+    vuint32m2_t vt_1 = __riscv_vid_v_u32m2(vl);
+    vuint32m2_t vt_2 = __riscv_vadd_vx_u32m2(vt_1, 12, vl);
+
+    for (; ib < nb; ++ib) {
+        memcpy(&qh, x[ib].qh, sizeof(uint32_t));
+
+        // load qh
+        vuint32m2_t vqh = __riscv_vmv_v_x_u32m2(qh, vl);
+
+        // ((qh >> (j +  0)) << 4) & 0x10;
+        vuint32m2_t xhr_0 = __riscv_vsrl_vv_u32m2(vqh, vt_1, vl);
+        vuint32m2_t xhl_0 = __riscv_vsll_vx_u32m2(xhr_0, 4, vl);
+        vuint32m2_t xha_0 = __riscv_vand_vx_u32m2(xhl_0, 0x10, vl);
+
+        // ((qh >> (j + 12))     ) & 0x10;
+        vuint32m2_t xhr_1 = __riscv_vsrl_vv_u32m2(vqh, vt_2, vl);
+        vuint32m2_t xha_1 = __riscv_vand_vx_u32m2(xhr_1, 0x10, vl);
+
+        // narrowing
+        vuint16m1_t xhc_0 = __riscv_vncvt_x_x_w_u16m1(xha_0, vl);
+        vuint8mf2_t xh_0 = __riscv_vncvt_x_x_w_u8mf2(xhc_0, vl);
+
+        vuint16m1_t xhc_1 = __riscv_vncvt_x_x_w_u16m1(xha_1, vl);
+        vuint8mf2_t xh_1 = __riscv_vncvt_x_x_w_u8mf2(xhc_1, vl);
+
+        // load
+        vuint8mf2_t tx = __riscv_vle8_v_u8mf2(x[ib].qs, vl);
+
+        vint8mf2_t y0 = __riscv_vle8_v_i8mf2(y[ib].qs, vl);
+        vint8mf2_t y1 = __riscv_vle8_v_i8mf2(y[ib].qs+16, vl);
+
+        vuint8mf2_t x_at = __riscv_vand_vx_u8mf2(tx, 0x0F, vl);
+        vuint8mf2_t x_lt = __riscv_vsrl_vx_u8mf2(tx, 0x04, vl);
+
+        vuint8mf2_t x_a = __riscv_vor_vv_u8mf2(x_at, xh_0, vl);
+        vuint8mf2_t x_l = __riscv_vor_vv_u8mf2(x_lt, xh_1, vl);
+
+        vint8mf2_t v0 = __riscv_vreinterpret_v_u8mf2_i8mf2(x_a);
+        vint8mf2_t v1 = __riscv_vreinterpret_v_u8mf2_i8mf2(x_l);
+
+        vint16m1_t vec_mul1 = __riscv_vwmul_vv_i16m1(v0, y0, vl);
+        vint16m1_t vec_mul2 = __riscv_vwmul_vv_i16m1(v1, y1, vl);
+
+        vint32m1_t vec_zero = __riscv_vmv_v_x_i32m1(0, vl);
+
+        vint32m1_t vs1 = __riscv_vwredsum_vs_i16m1_i32m1(vec_mul1, vec_zero, vl);
+        vint32m1_t vs2 = __riscv_vwredsum_vs_i16m1_i32m1(vec_mul2, vs1, vl);
+
+        int sumi = __riscv_vmv_x_s_i32m1_i32(vs2);
+
+        sumf += (GGML_FP16_TO_FP32(x[ib].d)*GGML_FP16_TO_FP32(y[ib].d))*sumi + GGML_FP16_TO_FP32(x[ib].m)*GGML_FP16_TO_FP32(y[ib].s);
+    }
+
+#elif defined(__POWER9_VECTOR__)
+    const vector signed char lowMask = vec_splats((signed char)0xF);
+    const vector signed int v0 = vec_splats((int32_t)0);
+    const vector unsigned char v4 = vec_splats((unsigned char)0x4);
+
+    vector float vsumf0 = vec_splats(0.0f);
+
+#pragma GCC unroll 4
+    for (; ib < nb; ++ib) {
+        __builtin_prefetch(x[ib].qs, 0, 1);
+        __builtin_prefetch(y[ib].qs, 0, 1);
+
+        vector float vxd = vec_splats(GGML_FP16_TO_FP32(x[ib].d));
+        vector float vyd = vec_splats(GGML_FP16_TO_FP32(y[ib].d));
+        vector float vd = vec_mul(vxd, vyd);
+
+        vector float vxmin = vec_splats(GGML_FP16_TO_FP32(x[ib].m));
+        vector float vys = {GGML_FP16_TO_FP32(y[ib].s), 0.f, 0.f, 0.f};
+        vsumf0 = vec_madd(vxmin, vys, vsumf0);
+
+        vector unsigned long long aux64x2_0 = {(uint64_t)(table_b2b_0[x[ib].qh[0]]), (uint64_t)(table_b2b_0[x[ib].qh[1]])};
+        vector unsigned long long aux64x2_1 = {(uint64_t)(table_b2b_0[x[ib].qh[2]]), (uint64_t)(table_b2b_0[x[ib].qh[3]])};
+
+        vector signed char qh0 = (vector signed char)aux64x2_0;
+        vector signed char qh1 = (vector signed char)aux64x2_1;
+
+        vector signed char qxs = (vector signed char)vec_xl( 0, x[ib].qs);
+
+        vector unsigned char q5x0 = (vector unsigned char)vec_or(vec_and(qxs, lowMask), qh0);
+        vector unsigned char q5x1 = (vector unsigned char)vec_or(vec_sr(qxs, v4), qh1);
+
+        vector signed char q8y0 = vec_xl(  0, y[ib].qs);
+        vector signed char q8y1 = vec_xl( 16, y[ib].qs);
+
+        vector signed int vsumi0 = v0;
+
+        vsumi0 = vec_msum(q8y0, q5x0, vsumi0);
+        vsumi0 = vec_msum(q8y1, q5x1, vsumi0);
+
+        vsumf0 = vec_madd(vec_ctf(vsumi0, 0), vd, vsumf0);
+    }
+
+    vsumf0 = vec_add(vsumf0, vec_sld(vsumf0, vsumf0, 4));
+    vsumf0 = vec_add(vsumf0, vec_sld(vsumf0, vsumf0, 8));
+
+    sumf = vec_extract(vsumf0, 0);
+
+#elif defined(__loongarch_asx)
+    // Initialize accumulator with zeros
+    __m256 acc = (__m256)__lasx_xvldi(0);
+
+    float summs = 0.0f;
+
+    // Main loop
+    for (; ib < nb; ++ib) {
+        const __m256 dx = __lasx_xvreplfr2vr_s(GGML_FP16_TO_FP32(x[ib].d));
+
+        summs += GGML_FP16_TO_FP32(x[ib].m) * GGML_FP16_TO_FP32(y[ib].s);
+
+        __m256i qx = bytes_from_nibbles_32(x[ib].qs);
+        __m256i bxhi = bytes_from_bits_32(x[ib].qh);
+        bxhi = __lasx_xvand_v(bxhi, __lasx_xvreplgr2vr_b(0x10));
+        qx = __lasx_xvor_v(qx, bxhi);
+
+        const __m256 dy = __lasx_xvreplfr2vr_s(GGML_FP16_TO_FP32(y[ib].d));
+        const __m256i qy = __lasx_xvld((const __m256i *)y[ib].qs, 0);
+
+        const __m256 q = mul_sum_us8_pairs_float(qx, qy);
+
+        acc = __lasx_xvfmadd_s(q, __lasx_xvfmul_s(dx, dy), acc);
+    }
+
+    sumf = hsum_float_8(acc) + summs;
+#endif
+    for (; ib < nb; ++ib) {
+        uint32_t qh;
+        memcpy(&qh, x[ib].qh, sizeof(qh));
+
+        int sumi0 = 0;
+        int sumi1 = 0;
+
+        for (int j = 0; j < qk/2; ++j) {
+            const uint8_t xh_0 = ((qh >> (j +  0)) << 4) & 0x10;
+            const uint8_t xh_1 = ((qh >> (j + 12))     ) & 0x10;
+
+            const int32_t x0 = (x[ib].qs[j] & 0xF) | xh_0;
+            const int32_t x1 = (x[ib].qs[j] >>  4) | xh_1;
+
+            sumi0 += (x0 * y[ib].qs[j]);
+            sumi1 += (x1 * y[ib].qs[j + qk/2]);
+        }
+
+        int sumi = sumi0 + sumi1;
+        sumf += (GGML_FP16_TO_FP32(x[ib].d)*GGML_FP16_TO_FP32(y[ib].d))*sumi + GGML_FP16_TO_FP32(x[ib].m)*GGML_FP16_TO_FP32(y[ib].s);
+    }
+
+    *s = sumf;
+}
+
+void ggml_vec_dot_q8_0_q8_0(int n, float * restrict s, size_t bs, const void * restrict vx, size_t bx, const void * restrict vy, size_t by, int nrc) {
+    const int qk = QK8_0;
+    const int nb = n / qk;
+
+    assert(n % qk == 0);
+#if defined(__ARM_FEATURE_MATMUL_INT8)
+    assert((nrc == 2) || (nrc == 1));
+#else
+    assert(nrc == 1);
+#endif
+    UNUSED(nrc);
+    UNUSED(bx);
+    UNUSED(by);
+    UNUSED(bs);
+
+    const block_q8_0 * restrict x = vx;
+    const block_q8_0 * restrict y = vy;
+
+#if defined(__ARM_FEATURE_MATMUL_INT8)
+    if (nrc == 2) {
+        const block_q8_0 * restrict vx0 = vx;
+        const block_q8_0 * restrict vx1 = (const block_q8_0 *) ((const uint8_t*)vx + bx);
+        const block_q8_0 * restrict vy0 = vy;
+        const block_q8_0 * restrict vy1 = (const block_q8_0 *) ((const uint8_t*)vy + by);
+
+        float32x4_t sumv0 = vdupq_n_f32(0.0f);
+
+        for (int i = 0; i < nb; i++) {
+            const block_q8_0 * restrict b_x0 = &vx0[i];
+            const block_q8_0 * restrict b_y0 = &vy0[i];
+
+            const block_q8_0 * restrict b_x1 = &vx1[i];
+            const block_q8_0 * restrict b_y1 = &vy1[i];
+
+            const int8x16_t x0_l = vld1q_s8(b_x0->qs);
+            const int8x16_t x0_h = vld1q_s8(b_x0->qs + 16);
+            const int8x16_t x1_l = vld1q_s8(b_x1->qs);
+            const int8x16_t x1_h = vld1q_s8(b_x1->qs + 16);
+
+            // load y
+            const int8x16_t y0_l = vld1q_s8(b_y0->qs);
+            const int8x16_t y0_h = vld1q_s8(b_y0->qs + 16);
+            const int8x16_t y1_l = vld1q_s8(b_y1->qs);
+            const int8x16_t y1_h = vld1q_s8(b_y1->qs + 16);
+
+            float32_t _scale[4] = {GGML_FP16_TO_FP32(b_x0->d)*GGML_FP16_TO_FP32(b_y0->d),
+                                   GGML_FP16_TO_FP32(b_x0->d)*GGML_FP16_TO_FP32(b_y1->d),
+                                   GGML_FP16_TO_FP32(b_x1->d)*GGML_FP16_TO_FP32(b_y0->d),
+                                   GGML_FP16_TO_FP32(b_x1->d)*GGML_FP16_TO_FP32(b_y1->d)};
+            float32x4_t scale = vld1q_f32(_scale);
+
+            int8x16_t l0 = vreinterpretq_s8_s64(vzip1q_s64(vreinterpretq_s64_s8(x0_l), vreinterpretq_s64_s8(x1_l)));
+            int8x16_t l1 = vreinterpretq_s8_s64(vzip2q_s64(vreinterpretq_s64_s8(x0_l), vreinterpretq_s64_s8(x1_l)));
+
+            int8x16_t l2 = vreinterpretq_s8_s64(vzip1q_s64(vreinterpretq_s64_s8(x0_h), vreinterpretq_s64_s8(x1_h)));
+            int8x16_t l3 = vreinterpretq_s8_s64(vzip2q_s64(vreinterpretq_s64_s8(x0_h), vreinterpretq_s64_s8(x1_h)));
+
+            int8x16_t r0 = vreinterpretq_s8_s64(vzip1q_s64(vreinterpretq_s64_s8(y0_l), vreinterpretq_s64_s8(y1_l)));
+            int8x16_t r1 = vreinterpretq_s8_s64(vzip2q_s64(vreinterpretq_s64_s8(y0_l), vreinterpretq_s64_s8(y1_l)));
+
+            int8x16_t r2 = vreinterpretq_s8_s64(vzip1q_s64(vreinterpretq_s64_s8(y0_h), vreinterpretq_s64_s8(y1_h)));
+            int8x16_t r3 = vreinterpretq_s8_s64(vzip2q_s64(vreinterpretq_s64_s8(y0_h), vreinterpretq_s64_s8(y1_h)));
+
+            sumv0 = vmlaq_f32(sumv0,(vcvtq_f32_s32(vmmlaq_s32((vmmlaq_s32((vmmlaq_s32((vmmlaq_s32(vdupq_n_s32(0), l0, r0)),
+                                                                                       l1, r1)), l2, r2)), l3, r3))), scale);
+        }
+        float32x4_t sumv1 = vextq_f32(sumv0, sumv0, 2);
+        float32x4_t sumv2 = vzip1q_f32(sumv0, sumv1);
+
+        vst1_f32(s, vget_low_f32(sumv2));
+        vst1_f32(s + bs, vget_high_f32(sumv2));
+        return;
+    }
+#endif
+
+    int ib = 0;
+    float sumf = 0;
+
+#if defined(__ARM_FEATURE_SVE)
+    svfloat32_t sumv0 = svdup_n_f32(0.0f);
+    svfloat32_t sumv1 = svdup_n_f32(0.0f);
+
+    const int vector_length = ggml_cpu_get_sve_cnt()*8;
+
+    //VLA Implemenation for SVE
+    switch (vector_length) {
+        case 128:
+            {
+                // predicate for activating lanes for 16 Int8 elements
+                const svbool_t ph16 = svptrue_pat_b8 (SV_VL16);
+                const svbool_t pl16 = svptrue_pat_b32(SV_VL4);
+
+                for (; ib + 1 < nb; ib += 2) {
+                    const block_q8_0 * restrict x0 = &x[ib + 0];
+                    const block_q8_0 * restrict x1 = &x[ib + 1];
+                    const block_q8_0 * restrict y0 = &y[ib + 0];
+                    const block_q8_0 * restrict y1 = &y[ib + 1];
+
+                    // load x
+                    const svint8_t qx0_0 = svld1_s8(ph16, x0->qs);
+                    const svint8_t qx0_1 = svld1_s8(ph16, x0->qs+16);
+                    const svint8_t qx1_0 = svld1_s8(ph16, x1->qs);
+                    const svint8_t qx1_1 = svld1_s8(ph16, x1->qs+16);
+
+                    // load y
+                    const svint8_t qy0_0 = svld1_s8(ph16, y0->qs);
+                    const svint8_t qy0_1 = svld1_s8(ph16, y0->qs+16);
+                    const svint8_t qy1_0 = svld1_s8(ph16, y1->qs);
+                    const svint8_t qy1_1 = svld1_s8(ph16, y1->qs+16);
+
+                    sumv0 = svmla_n_f32_x(pl16, sumv0, svcvt_f32_s32_x(pl16, svadd_x(pl16,
+                                    svdot_s32(svdup_n_s32(0), qx0_0, qy0_0),
+                                    svdot_s32(svdup_n_s32(0), qx0_1, qy0_1))), GGML_FP16_TO_FP32(x0->d)*GGML_FP16_TO_FP32(y0->d));
+                    sumv1 = svmla_n_f32_x(pl16, sumv1, svcvt_f32_s32_x(pl16, svadd_x(pl16,
+                                    svdot_s32(svdup_n_s32(0), qx1_0, qy1_0),
+                                    svdot_s32(svdup_n_s32(0), qx1_1, qy1_1))), GGML_FP16_TO_FP32(x1->d)*GGML_FP16_TO_FP32(y1->d));
+                }
+
+                sumf = svaddv_f32(pl16, svadd_f32_x(pl16, sumv0, sumv1));
+            } break;
+        case 256:
+            {
+                //printf("sve256");
+                for (; ib + 1 < nb; ib += 2) {
+                    const block_q8_0 * restrict x0 = &x[ib + 0];
+                    const block_q8_0 * restrict x1 = &x[ib + 1];
+                    const block_q8_0 * restrict y0 = &y[ib + 0];
+                    const block_q8_0 * restrict y1 = &y[ib + 1];
+
+                    // load x
+                    const svint8_t qx0 = svld1_s8(svptrue_b8(), x0->qs);
+                    const svint8_t qx1 = svld1_s8(svptrue_b8(), x1->qs);
+
+                    // load y
+                    const svint8_t qy0 = svld1_s8(svptrue_b8(), y0->qs);
+                    const svint8_t qy1 = svld1_s8(svptrue_b8(), y1->qs);
+
+                    sumv0 = svmla_n_f32_x(svptrue_b32(), sumv0, svcvt_f32_s32_x(svptrue_b32(),
+                                svdot_s32(svdup_n_s32(0), qx0, qy0)), GGML_FP16_TO_FP32(x0->d)*GGML_FP16_TO_FP32(y0->d));
+                    sumv1 = svmla_n_f32_x(svptrue_b32(), sumv1, svcvt_f32_s32_x(svptrue_b32(),
+                                svdot_s32(svdup_n_s32(0), qx1, qy1)), GGML_FP16_TO_FP32(x1->d)*GGML_FP16_TO_FP32(y1->d));
+                }
+
+                sumf = svaddv_f32(svptrue_b32(), svadd_f32_x(svptrue_b32(), sumv0, sumv1));
+            } break;
+        case 512:
+            {
+                // predicate for activating high 256 bit
+                const svbool_t ph32 = svptrue_pat_b8(SV_VL32);
+                // predicate for activating low 256 bit
+                const svbool_t pl32 = svnot_b_z(svptrue_b8(), ph32);
+
+                // predicate for activating high lanes for 8 float32 elements
+                const svbool_t ph8 = svptrue_pat_b32(SV_VL8);
+                // predicate for activating low lanes for 8 float32 elements
+                const svbool_t pl8 = svnot_b_z(svptrue_b32(), ph8);
+
+                svfloat32_t sumv00 = svdup_n_f32(0.0f);
+
+                for (; ib + 1 < nb; ib += 2) {
+                    const block_q8_0 * restrict x0 = &x[ib + 0];
+                    const block_q8_0 * restrict x1 = &x[ib + 1];
+                    const block_q8_0 * restrict y0 = &y[ib + 0];
+                    const block_q8_0 * restrict y1 = &y[ib + 1];
+
+                    //load 32 int8_t in first half of vector and put another 32 int8_t in second vector lower bits
+                    // and add them to make one 64 element vector
+                    // load x
+                    const svint8_t qx_32 = svld1_s8(ph32, x0->qs);
+                          svint8_t qx_64 = svld1_s8(pl32, x0->qs + 2);
+
+                    qx_64 = svadd_s8_x(svptrue_b8(), qx_32, qx_64);
+
+                    // load y
+                    const svint8_t qy_32 = svld1_s8(ph32, y0->qs);
+                          svint8_t qy_64 = svld1_s8(pl32, y0->qs + 2);
+
+                    qy_64 = svadd_s8_x(svptrue_b8(), qy_32, qy_64);
+
+                    // scale creation
+                    const float32_t deq1 = GGML_FP16_TO_FP32(x0->d)*GGML_FP16_TO_FP32(y0->d);
+                    const float32_t deq2 = GGML_FP16_TO_FP32(x1->d)*GGML_FP16_TO_FP32(y1->d);
+
+                    // duplicate deq1 in first half of vector and deq2 in second half of vector
+                    const svfloat32_t temp = svdup_f32_m(svdup_f32_z(ph8, deq1), pl8, deq2);
+
+                    const svfloat32_t sumvt = svcvt_f32_s32_x(svptrue_b32(), svdot_s32(svdup_n_s32(0), qx_64, qy_64));
+
+                    sumv00 = svmla_f32_m(svptrue_b32(), sumv00, sumvt, temp);
+                }
+
+                sumf = svaddv_f32(svptrue_b32(), sumv00);
+                break;
+            }
+        default:
+            assert(false && "Unsupported vector length");
+            break;
+    }
+#elif defined(__ARM_NEON)
+    float32x4_t sumv0 = vdupq_n_f32(0.0f);
+    float32x4_t sumv1 = vdupq_n_f32(0.0f);
+
+    for (; ib + 1 < nb; ib += 2) {
+        const block_q8_0 * restrict x0 = &x[ib + 0];
+        const block_q8_0 * restrict x1 = &x[ib + 1];
+        const block_q8_0 * restrict y0 = &y[ib + 0];
+        const block_q8_0 * restrict y1 = &y[ib + 1];
+
+        const int8x16_t x0_0 = vld1q_s8(x0->qs);
+        const int8x16_t x0_1 = vld1q_s8(x0->qs + 16);
+        const int8x16_t x1_0 = vld1q_s8(x1->qs);
+        const int8x16_t x1_1 = vld1q_s8(x1->qs + 16);
+
+        // load y
+        const int8x16_t y0_0 = vld1q_s8(y0->qs);
+        const int8x16_t y0_1 = vld1q_s8(y0->qs + 16);
+        const int8x16_t y1_0 = vld1q_s8(y1->qs);
+        const int8x16_t y1_1 = vld1q_s8(y1->qs + 16);
+
+        sumv0 = vmlaq_n_f32(sumv0, vcvtq_f32_s32(vaddq_s32(
+                        ggml_vdotq_s32(vdupq_n_s32(0), x0_0, y0_0),
+                        ggml_vdotq_s32(vdupq_n_s32(0), x0_1, y0_1))), GGML_FP16_TO_FP32(x0->d)*GGML_FP16_TO_FP32(y0->d));
+
+        sumv1 = vmlaq_n_f32(sumv1, vcvtq_f32_s32(vaddq_s32(
+                        ggml_vdotq_s32(vdupq_n_s32(0), x1_0, y1_0),
+                        ggml_vdotq_s32(vdupq_n_s32(0), x1_1, y1_1))), GGML_FP16_TO_FP32(x1->d)*GGML_FP16_TO_FP32(y1->d));
+    }
+
+    sumf = vaddvq_f32(sumv0) + vaddvq_f32(sumv1);
+#elif defined(__AVX2__) || defined(__AVX__)
+    // Initialize accumulator with zeros
+    __m256 acc = _mm256_setzero_ps();
+
+    // Main loop
+    for (; ib < nb; ++ib) {
+        // Compute combined scale for the block
+        const __m256 d = _mm256_set1_ps(GGML_FP16_TO_FP32(x[ib].d) * GGML_FP16_TO_FP32(y[ib].d));
+        __m256i qx = _mm256_loadu_si256((const __m256i *)x[ib].qs);
+        __m256i qy = _mm256_loadu_si256((const __m256i *)y[ib].qs);
+
+        const __m256 q = mul_sum_i8_pairs_float(qx, qy);
+
+        // Multiply q with scale and accumulate
+#if defined(__AVX2__)
+        acc = _mm256_fmadd_ps( d, q, acc );
+#else
+        acc = _mm256_add_ps( _mm256_mul_ps( d, q ), acc );
+#endif
+    }
+
+    sumf = hsum_float_8(acc);
+#elif defined(__riscv_v_intrinsic)
+    size_t vl = __riscv_vsetvl_e8m1(qk);
+
+    for (; ib < nb; ++ib) {
+        // load elements
+        vint8m1_t bx_0 = __riscv_vle8_v_i8m1(x[ib].qs, vl);
+        vint8m1_t by_0 = __riscv_vle8_v_i8m1(y[ib].qs, vl);
+
+        vint16m2_t vw_mul = __riscv_vwmul_vv_i16m2(bx_0, by_0, vl);
+
+        vint32m1_t v_zero = __riscv_vmv_v_x_i32m1(0, vl);
+        vint32m1_t v_sum = __riscv_vwredsum_vs_i16m2_i32m1(vw_mul, v_zero, vl);
+
+        int sumi = __riscv_vmv_x_s_i32m1_i32(v_sum);
+
+        sumf += sumi*(GGML_FP16_TO_FP32(x[ib].d)*GGML_FP16_TO_FP32(y[ib].d));
+    }
+#elif defined(__POWER9_VECTOR__)
+    const vector signed int v0 = vec_splats((int32_t)0);
+    vector float vsumf0 = vec_splats(0.0f);
+
+#pragma GCC unroll 8
+    for (; ib < nb; ++ib) {
+        __builtin_prefetch(x[ib].qs, 0, 1);
+        __builtin_prefetch(y[ib].qs, 0, 1);
+
+        vector float vxd = vec_splats(GGML_FP16_TO_FP32(x[ib].d));
+        vector float vyd = vec_splats(GGML_FP16_TO_FP32(y[ib].d));
+        vector float vd = vec_mul(vxd, vyd);
+
+        vector signed char q8x0 = vec_xl( 0, x[ib].qs);
+        vector signed char q8x1 = vec_xl(16, x[ib].qs);
+        vector signed char q8y0 = vec_xl( 0, y[ib].qs);
+        vector signed char q8y1 = vec_xl(16, y[ib].qs);
+
+        vector signed short qv0 = vec_mule(q8x0, q8y0);
+        vector signed short qv1 = vec_mulo(q8x0, q8y0);
+        vector signed short qv2 = vec_mule(q8x1, q8y1);
+        vector signed short qv3 = vec_mulo(q8x1, q8y1);
+
+        vector signed int vsumi0 = v0;
+        vector signed int vsumi1 = v0;
+
+        vsumi0 = vec_sum4s(qv0, vsumi0);
+        vsumi1 = vec_sum4s(qv1, vsumi1);
+        vsumi0 = vec_sum4s(qv2, vsumi0);
+        vsumi1 = vec_sum4s(qv3, vsumi1);
+
+        vsumi0 = vec_add(vsumi0, vsumi1);
+
+        vsumf0 = vec_madd(vec_ctf(vsumi0, 0), vd, vsumf0);
+    }
+
+    vsumf0 = vec_add(vsumf0, vec_sld(vsumf0, vsumf0, 4));
+    vsumf0 = vec_add(vsumf0, vec_sld(vsumf0, vsumf0, 8));
+
+    sumf = vec_extract(vsumf0, 0);
+
+#elif defined(__loongarch_asx)
+    // Initialize accumulator with zeros
+    __m256 acc = (__m256)__lasx_xvldi(0);
+
+    // Main loop
+    for (; ib < nb; ++ib) {
+        // Compute combined scale for the block
+        const __m256 d = __lasx_xvreplfr2vr_s(GGML_FP16_TO_FP32(x[ib].d) * GGML_FP16_TO_FP32(y[ib].d));
+        __m256i qx = __lasx_xvld((const __m256i *)x[ib].qs, 0);
+        __m256i qy = __lasx_xvld((const __m256i *)y[ib].qs, 0);
+
+        const __m256 q = mul_sum_i8_pairs_float(qx, qy);
+
+        // Multiply q with scale and accumulate
+        acc = __lasx_xvfmadd_s( d, q, acc );
+    }
+
+    sumf = hsum_float_8(acc);
+#endif
+    for (; ib < nb; ++ib) {
+        int sumi = 0;
+
+        for (int j = 0; j < qk; j++) {
+            sumi += x[ib].qs[j]*y[ib].qs[j];
+        }
+
+        sumf += sumi*(GGML_FP16_TO_FP32(x[ib].d)*GGML_FP16_TO_FP32(y[ib].d));
+    }
+
+    *s = sumf;
+}
+
+void ggml_vec_dot_tq1_0_q8_K(int n, float * restrict s, size_t bs, const void * restrict vx, size_t bx, const void * restrict vy, size_t by, int nrc) {
+    assert(nrc == 1);
+    UNUSED(nrc);
+    UNUSED(bx);
+    UNUSED(by);
+    UNUSED(bs);
+
+    const block_tq1_0 * restrict x = vx;
+    const block_q8_K  * restrict y = vy;
+
+    const int nb = n / QK_K;
+
+#if defined(__ARM_NEON)
+    float sumf = 0.0f;
+
+    uint8_t k_shift[16] = {1, 1, 1, 1, 3, 3, 3, 3, 9, 9, 9, 9, 27, 27, 27, 27};
+
+    const uint8x16_t shift = vld1q_u8(k_shift);
+
+    for (int i = 0; i < nb; ++i) {
+#if defined(__ARM_FEATURE_DOTPROD)
+        int32x4_t sumi0 = vdupq_n_s32(0);
+        int32x4_t sumi1 = vdupq_n_s32(0);
+#else
+        int16x8_t sumi0 = vdupq_n_s16(0);
+        int16x8_t sumi1 = vdupq_n_s16(0);
+#endif
+
+        // first 32 bytes of 5 elements
+        {
+            uint8x16_t qx0 = vld1q_u8(x[i].qs + 0);
+            uint8x16_t qx1 = vld1q_u8(x[i].qs + 16);
+            uint8x16_t qx2 = vmulq_u8(qx0, vdupq_n_u8(3));
+            uint8x16_t qx3 = vmulq_u8(qx1, vdupq_n_u8(3));
+            uint8x16_t qx4 = vmulq_u8(qx0, vdupq_n_u8(9));
+            uint8x16_t qx5 = vmulq_u8(qx1, vdupq_n_u8(9));
+            uint8x16_t qx6 = vmulq_u8(qx0, vdupq_n_u8(27));
+            uint8x16_t qx7 = vmulq_u8(qx1, vdupq_n_u8(27));
+            uint8x16_t qx8 = vmulq_u8(qx0, vdupq_n_u8(81));
+            uint8x16_t qx9 = vmulq_u8(qx1, vdupq_n_u8(81));
+
+            // multiply by 3 and keep the 2 bits above 8 bits
+            int8x16_t sqx0 = vreinterpretq_s8_u8(vshrq_n_u8(vhaddq_u8(qx0, vshrq_n_u8(qx0, 1)), 6));
+            int8x16_t sqx1 = vreinterpretq_s8_u8(vshrq_n_u8(vhaddq_u8(qx1, vshrq_n_u8(qx1, 1)), 6));
+            int8x16_t sqx2 = vreinterpretq_s8_u8(vshrq_n_u8(vhaddq_u8(qx2, vshrq_n_u8(qx2, 1)), 6));
+            int8x16_t sqx3 = vreinterpretq_s8_u8(vshrq_n_u8(vhaddq_u8(qx3, vshrq_n_u8(qx3, 1)), 6));
+            int8x16_t sqx4 = vreinterpretq_s8_u8(vshrq_n_u8(vhaddq_u8(qx4, vshrq_n_u8(qx4, 1)), 6));
+            int8x16_t sqx5 = vreinterpretq_s8_u8(vshrq_n_u8(vhaddq_u8(qx5, vshrq_n_u8(qx5, 1)), 6));
+            int8x16_t sqx6 = vreinterpretq_s8_u8(vshrq_n_u8(vhaddq_u8(qx6, vshrq_n_u8(qx6, 1)), 6));
+            int8x16_t sqx7 = vreinterpretq_s8_u8(vshrq_n_u8(vhaddq_u8(qx7, vshrq_n_u8(qx7, 1)), 6));
+            int8x16_t sqx8 = vreinterpretq_s8_u8(vshrq_n_u8(vhaddq_u8(qx8, vshrq_n_u8(qx8, 1)), 6));
+            int8x16_t sqx9 = vreinterpretq_s8_u8(vshrq_n_u8(vhaddq_u8(qx9, vshrq_n_u8(qx9, 1)), 6));
+
+            const int8x16_t qy0 = vld1q_s8(y[i].qs +   0);
+            const int8x16_t qy1 = vld1q_s8(y[i].qs +  16);
+            const int8x16_t qy2 = vld1q_s8(y[i].qs +  32);
+            const int8x16_t qy3 = vld1q_s8(y[i].qs +  48);
+            const int8x16_t qy4 = vld1q_s8(y[i].qs +  64);
+            const int8x16_t qy5 = vld1q_s8(y[i].qs +  80);
+            const int8x16_t qy6 = vld1q_s8(y[i].qs +  96);
+            const int8x16_t qy7 = vld1q_s8(y[i].qs + 112);
+            const int8x16_t qy8 = vld1q_s8(y[i].qs + 128);
+            const int8x16_t qy9 = vld1q_s8(y[i].qs + 144);
+
+#if defined(__ARM_FEATURE_DOTPROD)
+            sumi0 = vdotq_s32(sumi0, sqx0, qy0);
+            sumi1 = vdotq_s32(sumi1, sqx1, qy1);
+            sumi0 = vdotq_s32(sumi0, sqx2, qy2);
+            sumi1 = vdotq_s32(sumi1, sqx3, qy3);
+            sumi0 = vdotq_s32(sumi0, sqx4, qy4);
+            sumi1 = vdotq_s32(sumi1, sqx5, qy5);
+            sumi0 = vdotq_s32(sumi0, sqx6, qy6);
+            sumi1 = vdotq_s32(sumi1, sqx7, qy7);
+            sumi0 = vdotq_s32(sumi0, sqx8, qy8);
+            sumi1 = vdotq_s32(sumi1, sqx9, qy9);
+#else
+            sumi0 = vmlal_s8(sumi0, vget_low_s8(sqx0), vget_low_s8(qy0));
+            sumi1 = vmlal_s8(sumi1, vget_high_s8(sqx0), vget_high_s8(qy0));
+            sumi0 = vmlal_s8(sumi0, vget_low_s8(sqx1), vget_low_s8(qy1));
+            sumi1 = vmlal_s8(sumi1, vget_high_s8(sqx1), vget_high_s8(qy1));
+            sumi0 = vmlal_s8(sumi0, vget_low_s8(sqx2), vget_low_s8(qy2));
+            sumi1 = vmlal_s8(sumi1, vget_high_s8(sqx2), vget_high_s8(qy2));
+            sumi0 = vmlal_s8(sumi0, vget_low_s8(sqx3), vget_low_s8(qy3));
+            sumi1 = vmlal_s8(sumi1, vget_high_s8(sqx3), vget_high_s8(qy3));
+            sumi0 = vmlal_s8(sumi0, vget_low_s8(sqx4), vget_low_s8(qy4));
+            sumi1 = vmlal_s8(sumi1, vget_high_s8(sqx4), vget_high_s8(qy4));
+            sumi0 = vmlal_s8(sumi0, vget_low_s8(sqx5), vget_low_s8(qy5));
+            sumi1 = vmlal_s8(sumi1, vget_high_s8(sqx5), vget_high_s8(qy5));
+            sumi0 = vmlal_s8(sumi0, vget_low_s8(sqx6), vget_low_s8(qy6));
+            sumi1 = vmlal_s8(sumi1, vget_high_s8(sqx6), vget_high_s8(qy6));
+            sumi0 = vmlal_s8(sumi0, vget_low_s8(sqx7), vget_low_s8(qy7));
+            sumi1 = vmlal_s8(sumi1, vget_high_s8(sqx7), vget_high_s8(qy7));
+            sumi0 = vmlal_s8(sumi0, vget_low_s8(sqx8), vget_low_s8(qy8));
+            sumi1 = vmlal_s8(sumi1, vget_high_s8(sqx8), vget_high_s8(qy8));
+            sumi0 = vmlal_s8(sumi0, vget_low_s8(sqx9), vget_low_s8(qy9));
+            sumi1 = vmlal_s8(sumi1, vget_high_s8(sqx9), vget_high_s8(qy9));
+#endif
+        }
+
+        // last 16 bytes of 5-element, along with the 4 bytes of 4 elements
+        {
+            uint8x16_t qx0 = vld1q_u8(x[i].qs + 32);
+            uint8x16_t qx1 = vmulq_u8(qx0, vdupq_n_u8(3));
+            uint8x16_t qx2 = vmulq_u8(qx0, vdupq_n_u8(9));
+            uint8x16_t qx3 = vmulq_u8(qx0, vdupq_n_u8(27));
+            uint8x16_t qx4 = vmulq_u8(qx0, vdupq_n_u8(81));
+            uint32_t qh;
+            memcpy(&qh, x[i].qh, sizeof(qh)); // potentially unaligned
+            uint8x16_t qx5 = vreinterpretq_u8_u32(vdupq_n_u32(qh));
+            qx5 = vmulq_u8(qx5, shift);
+
+            // multiply by 3 and keep the 2 bits above 8 bits
+            int8x16_t sqx0 = vreinterpretq_s8_u8(vshrq_n_u8(vhaddq_u8(qx0, vshrq_n_u8(qx0, 1)), 6));
+            int8x16_t sqx1 = vreinterpretq_s8_u8(vshrq_n_u8(vhaddq_u8(qx1, vshrq_n_u8(qx1, 1)), 6));
+            int8x16_t sqx2 = vreinterpretq_s8_u8(vshrq_n_u8(vhaddq_u8(qx2, vshrq_n_u8(qx2, 1)), 6));
+            int8x16_t sqx3 = vreinterpretq_s8_u8(vshrq_n_u8(vhaddq_u8(qx3, vshrq_n_u8(qx3, 1)), 6));
+            int8x16_t sqx4 = vreinterpretq_s8_u8(vshrq_n_u8(vhaddq_u8(qx4, vshrq_n_u8(qx4, 1)), 6));
+            int8x16_t sqx5 = vreinterpretq_s8_u8(vshrq_n_u8(vhaddq_u8(qx5, vshrq_n_u8(qx5, 1)), 6));
+
+            const int8x16_t qy0 = vld1q_s8(y[i].qs + 160);
+            const int8x16_t qy1 = vld1q_s8(y[i].qs + 176);
+            const int8x16_t qy2 = vld1q_s8(y[i].qs + 192);
+            const int8x16_t qy3 = vld1q_s8(y[i].qs + 208);
+            const int8x16_t qy4 = vld1q_s8(y[i].qs + 224);
+            const int8x16_t qy5 = vld1q_s8(y[i].qs + 240);
+
+#if defined(__ARM_FEATURE_DOTPROD)
+            sumi0 = vdotq_s32(sumi0, sqx0, qy0);
+            sumi1 = vdotq_s32(sumi1, sqx1, qy1);
+            sumi0 = vdotq_s32(sumi0, sqx2, qy2);
+            sumi1 = vdotq_s32(sumi1, sqx3, qy3);
+            sumi0 = vdotq_s32(sumi0, sqx4, qy4);
+            sumi1 = vdotq_s32(sumi1, sqx5, qy5);
+#else
+            sumi0 = vmlal_s8(sumi0, vget_low_s8(sqx0), vget_low_s8(qy0));
+            sumi1 = vmlal_s8(sumi1, vget_high_s8(sqx0), vget_high_s8(qy0));
+            sumi0 = vmlal_s8(sumi0, vget_low_s8(sqx1), vget_low_s8(qy1));
+            sumi1 = vmlal_s8(sumi1, vget_high_s8(sqx1), vget_high_s8(qy1));
+            sumi0 = vmlal_s8(sumi0, vget_low_s8(sqx2), vget_low_s8(qy2));
+            sumi1 = vmlal_s8(sumi1, vget_high_s8(sqx2), vget_high_s8(qy2));
+            sumi0 = vmlal_s8(sumi0, vget_low_s8(sqx3), vget_low_s8(qy3));
+            sumi1 = vmlal_s8(sumi1, vget_high_s8(sqx3), vget_high_s8(qy3));
+            sumi0 = vmlal_s8(sumi0, vget_low_s8(sqx4), vget_low_s8(qy4));
+            sumi1 = vmlal_s8(sumi1, vget_high_s8(sqx4), vget_high_s8(qy4));
+            sumi0 = vmlal_s8(sumi0, vget_low_s8(sqx5), vget_low_s8(qy5));
+            sumi1 = vmlal_s8(sumi1, vget_high_s8(sqx5), vget_high_s8(qy5));
+#endif
+        }
+
+        const int16x8_t ysum0 = vld1q_s16(y[i].bsums);
+        const int16x8_t ysum1 = vld1q_s16(y[i].bsums + 8);
+
+        const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
+
+#if defined(__ARM_FEATURE_DOTPROD)
+        sumi0 = vaddq_s32(sumi0, sumi1);
+        sumi0 = vsubq_s32(sumi0, vpaddlq_s16(vaddq_s16(ysum0, ysum1)));
+
+        sumf += d * (float) vaddvq_s32(sumi0);
+#else
+        sumi0 = vaddq_s16(sumi0, sumi1);
+        sumi0 = vsubq_s16(sumi0, vaddq_s16(ysum0, ysum1));
+
+        sumf += d * (float) vaddlvq_s16(sumi0);
+#endif
+    }
+
+    *s = sumf;
+
+#elif defined(__AVX2__)
+    __m256 sumf = _mm256_setzero_ps();
+
+    for (int i = 0; i < nb; ++i) {
+        // 16-bit sums
+        __m256i sumi0 = _mm256_setzero_si256();
+        __m256i sumi1 = _mm256_setzero_si256();
+        __m256i sumi2 = _mm256_setzero_si256();
+
+        // first 32 bytes of 5 elements
+        {
+            __m256i qx0 = _mm256_loadu_si256((const __m256i *) (x[i].qs));
+            // 8-bit multiplies with shifts, masks and adds
+            __m256i qx1 = _mm256_add_epi8(qx0, _mm256_add_epi8(qx0, qx0)); // 1 * 3
+            __m256i qx2 = _mm256_add_epi8(_mm256_and_si256(_mm256_slli_epi16(qx0, 3), _mm256_set1_epi8(-8)), qx0); // 1 * 9
+            __m256i qx3 = _mm256_add_epi8(_mm256_and_si256(_mm256_slli_epi16(qx1, 3), _mm256_set1_epi8(-8)), qx1); // 3 * 9
+            __m256i qx4 = _mm256_add_epi8(_mm256_and_si256(_mm256_slli_epi16(qx2, 3), _mm256_set1_epi8(-8)), qx2); // 9 * 9
+
+            // TODO: can _mm256_mulhi_epu16 be faster even if 16-bits?
+
+            // Cancel the +1 from avg so that it behaves like a halving add
+            qx0 = _mm256_subs_epu8(qx0, _mm256_set1_epi8(1));
+            qx1 = _mm256_subs_epu8(qx1, _mm256_set1_epi8(1));
+            qx2 = _mm256_subs_epu8(qx2, _mm256_set1_epi8(1));
+            qx3 = _mm256_subs_epu8(qx3, _mm256_set1_epi8(1));
+            qx4 = _mm256_subs_epu8(qx4, _mm256_set1_epi8(1));
+            // Multiply by 3 and get the top 2 bits
+            qx0 = _mm256_avg_epu8(qx0, _mm256_avg_epu8(qx0, _mm256_setzero_si256()));
+            qx1 = _mm256_avg_epu8(qx1, _mm256_avg_epu8(qx1, _mm256_setzero_si256()));
+            qx2 = _mm256_avg_epu8(qx2, _mm256_avg_epu8(qx2, _mm256_setzero_si256()));
+            qx3 = _mm256_avg_epu8(qx3, _mm256_avg_epu8(qx3, _mm256_setzero_si256()));
+            qx4 = _mm256_avg_epu8(qx4, _mm256_avg_epu8(qx4, _mm256_setzero_si256()));
+            qx0 = _mm256_and_si256(_mm256_srli_epi16(qx0, 6), _mm256_set1_epi8(3));
+            qx1 = _mm256_and_si256(_mm256_srli_epi16(qx1, 6), _mm256_set1_epi8(3));
+            qx2 = _mm256_and_si256(_mm256_srli_epi16(qx2, 6), _mm256_set1_epi8(3));
+            qx3 = _mm256_and_si256(_mm256_srli_epi16(qx3, 6), _mm256_set1_epi8(3));
+            qx4 = _mm256_and_si256(_mm256_srli_epi16(qx4, 6), _mm256_set1_epi8(3));
+
+            const __m256i qy0 = _mm256_loadu_si256((const __m256i *) (y[i].qs +   0));
+            const __m256i qy1 = _mm256_loadu_si256((const __m256i *) (y[i].qs +  32));
+            const __m256i qy2 = _mm256_loadu_si256((const __m256i *) (y[i].qs +  64));
+            const __m256i qy3 = _mm256_loadu_si256((const __m256i *) (y[i].qs +  96));
+            const __m256i qy4 = _mm256_loadu_si256((const __m256i *) (y[i].qs + 128));
+
+            qx0 = _mm256_maddubs_epi16(qx0, qy0);
+            qx1 = _mm256_maddubs_epi16(qx1, qy1);
+            qx2 = _mm256_maddubs_epi16(qx2, qy2);
+            qx3 = _mm256_maddubs_epi16(qx3, qy3);
+            qx4 = _mm256_maddubs_epi16(qx4, qy4);
+
+            sumi0 = _mm256_add_epi16(sumi0, _mm256_add_epi16(qx0, qx1));
+            sumi1 = _mm256_add_epi16(sumi1, _mm256_add_epi16(qx2, qx3));
+            sumi2 = _mm256_add_epi16(sumi2, qx4);
+        }
+
+        // last 16 bytes of 5-element, along with the 4 bytes of 4 elements
+        {
+            __m128i qx0 = _mm_loadu_si128((const __m128i *) (x[i].qs + 32));
+            uint32_t qh;
+            memcpy(&qh, x[i].qh, sizeof(qh)); // potentially unaligned
+            __m256i qx5_l = _mm256_cvtepu8_epi16(_mm_set1_epi32(qh));
+            __m128i qx1 = _mm_add_epi8(qx0, _mm_add_epi8(qx0, qx0)); // 1 * 3
+            __m128i qx2 = _mm_add_epi8(_mm_and_si128(_mm_slli_epi16(qx0, 3), _mm_set1_epi8(-8)), qx0); // 1 * 9
+            __m128i qx3 = _mm_add_epi8(_mm_and_si128(_mm_slli_epi16(qx1, 3), _mm_set1_epi8(-8)), qx1); // 3 * 9
+            __m128i qx4 = _mm_add_epi8(_mm_and_si128(_mm_slli_epi16(qx2, 3), _mm_set1_epi8(-8)), qx2); // 9 * 9
+            __m256i qx01 = MM256_SET_M128I(qx1, qx0);
+            __m256i qx23 = MM256_SET_M128I(qx3, qx2);
+
+            // avx2 does not have 8-bit multiplies, so 16-bit it is.
+            qx5_l = _mm256_mullo_epi16(qx5_l, _mm256_set_epi16(27, 27, 27, 27, 9, 9, 9, 9, 3, 3, 3, 3, 1, 1, 1, 1));
+            qx5_l = _mm256_and_si256(qx5_l, _mm256_set1_epi16(0xFF));
+            __m128i qx5 = _mm_packus_epi16(_mm256_castsi256_si128(qx5_l), _mm256_extracti128_si256(qx5_l, 1));
+
+            __m256i qx45 = MM256_SET_M128I(qx5, qx4);
+
+            // Cancel the +1 from avg so that it behaves like a halving add
+            qx01 = _mm256_subs_epu8(qx01, _mm256_set1_epi8(1));
+            qx23 = _mm256_subs_epu8(qx23, _mm256_set1_epi8(1));
+            qx45 = _mm256_subs_epu8(qx45, _mm256_set1_epi8(1));
+            // Multiply by 3 and get the top 2 bits
+            qx01 = _mm256_avg_epu8(qx01, _mm256_avg_epu8(qx01, _mm256_setzero_si256()));
+            qx23 = _mm256_avg_epu8(qx23, _mm256_avg_epu8(qx23, _mm256_setzero_si256()));
+            qx45 = _mm256_avg_epu8(qx45, _mm256_avg_epu8(qx45, _mm256_setzero_si256()));
+            qx01 = _mm256_and_si256(_mm256_srli_epi16(qx01, 6), _mm256_set1_epi8(3));
+            qx23 = _mm256_and_si256(_mm256_srli_epi16(qx23, 6), _mm256_set1_epi8(3));
+            qx45 = _mm256_and_si256(_mm256_srli_epi16(qx45, 6), _mm256_set1_epi8(3));
+
+            const __m256i qy01 = _mm256_loadu_si256((const __m256i *) (y[i].qs + 160));
+            const __m256i qy23 = _mm256_loadu_si256((const __m256i *) (y[i].qs + 192));
+            const __m256i qy45 = _mm256_loadu_si256((const __m256i *) (y[i].qs + 224));
+
+            qx01 = _mm256_maddubs_epi16(qx01, qy01);
+            qx23 = _mm256_maddubs_epi16(qx23, qy23);
+            qx45 = _mm256_maddubs_epi16(qx45, qy45);
+
+            sumi0 = _mm256_add_epi16(sumi0, qx01);
+            sumi1 = _mm256_add_epi16(sumi1, qx23);
+            sumi2 = _mm256_add_epi16(sumi2, qx45);
+        }
+
+        const __m256i ysum = _mm256_loadu_si256((const __m256i *) y[i].bsums);
+        const __m256 d = _mm256_set1_ps(y[i].d * GGML_FP16_TO_FP32(x[i].d));
+
+        sumi0 = _mm256_sub_epi16(sumi0, ysum);
+        sumi0 = _mm256_add_epi16(sumi0, _mm256_add_epi16(sumi1, sumi2));
+        sumi0 = _mm256_madd_epi16(sumi0, _mm256_set1_epi16(1));
+
+        sumf = _mm256_add_ps(_mm256_mul_ps(_mm256_cvtepi32_ps(sumi0), d), sumf);
+    }
+
+    *s = hsum_float_8(sumf);
+
+#else
+    const uint8_t pow3[6] = {1, 3, 9, 27, 81, 243};
+
+    float sumf = 0.0f;
+
+    for (int i = 0; i < nb; ++i) {
+        int sum = 0;
+
+        for (size_t j = 0; j < sizeof(x->qs) - sizeof(x->qs) % 32; j += 32) {
+            for (size_t l = 0; l < 5; ++l) {
+                for (size_t m = 0; m < 32; ++m) {
+                    uint8_t q = x[i].qs[j + m] * pow3[l];
+                    uint16_t xi = ((uint16_t) q * 3) >> 8;
+                    sum += (xi - 1) * y[i].qs[j*5 + l*32 + m];
+                }
+            }
+        }
+        for (size_t j = sizeof(x->qs) - sizeof(x->qs) % 32; j < sizeof(x->qs); j += 16) {
+            for (size_t l = 0; l < 5; ++l) {
+                for (size_t m = 0; m < 16; ++m) {
+                    uint8_t q = x[i].qs[j + m] * pow3[l];
+                    uint16_t xi = ((uint16_t) q * 3) >> 8;
+                    sum += (xi - 1) * y[i].qs[j*5 + l*16 + m];
+                }
+            }
+        }
+
+        for (size_t l = 0; l < 4; ++l) {
+            for (size_t j = 0; j < sizeof(x->qh); ++j) {
+                uint8_t q = x[i].qh[j] * pow3[l];
+                uint16_t xi = ((uint16_t) q * 3) >> 8;
+                sum += (xi - 1) * y[i].qs[sizeof(x->qs)*5 + l*sizeof(x->qh) + j];
+            }
+        }
+
+        sumf += (float) sum * (GGML_FP16_TO_FP32(x[i].d) * y[i].d);
+    }
+
+    *s = sumf;
+#endif
+}
+
+void ggml_vec_dot_tq2_0_q8_K(int n, float * restrict s, size_t bs, const void * restrict vx, size_t bx, const void * restrict vy, size_t by, int nrc) {
+    assert(nrc == 1);
+    UNUSED(nrc);
+    UNUSED(bx);
+    UNUSED(by);
+    UNUSED(bs);
+
+    const block_tq2_0 * restrict x = vx;
+    const block_q8_K  * restrict y = vy;
+
+    const int nb = n / QK_K;
+
+#if defined(__ARM_NEON)
+    float sumf = 0.0f;
+
+    const uint8x16_t m3 = vdupq_n_u8(3);
+
+    for (int i = 0; i < nb; ++i) {
+#if defined(__ARM_FEATURE_DOTPROD)
+        int32x4_t sumi0 = vdupq_n_s32(0);
+        int32x4_t sumi1 = vdupq_n_s32(0);
+#else
+        int16x8_t sumi0 = vdupq_n_s16(0);
+        int16x8_t sumi1 = vdupq_n_s16(0);
+#endif
+
+        for (size_t j = 0; j < sizeof(x->qs); j += 32) {
+            uint8x16_t qx0 = vld1q_u8(x[i].qs + j);
+            uint8x16_t qx1 = vld1q_u8(x[i].qs + j + 16);
+            uint8x16_t qx2 = vshrq_n_u8(qx0, 2);
+            uint8x16_t qx3 = vshrq_n_u8(qx1, 2);
+            uint8x16_t qx4 = vshrq_n_u8(qx0, 4);
+            uint8x16_t qx5 = vshrq_n_u8(qx1, 4);
+            uint8x16_t qx6 = vshrq_n_u8(qx0, 6);
+            uint8x16_t qx7 = vshrq_n_u8(qx1, 6);
+
+            int8x16_t sqx0 = vreinterpretq_s8_u8(vandq_u8(qx0, m3));
+            int8x16_t sqx1 = vreinterpretq_s8_u8(vandq_u8(qx1, m3));
+            int8x16_t sqx2 = vreinterpretq_s8_u8(vandq_u8(qx2, m3));
+            int8x16_t sqx3 = vreinterpretq_s8_u8(vandq_u8(qx3, m3));
+            int8x16_t sqx4 = vreinterpretq_s8_u8(vandq_u8(qx4, m3));
+            int8x16_t sqx5 = vreinterpretq_s8_u8(vandq_u8(qx5, m3));
+            int8x16_t sqx6 = vreinterpretq_s8_u8(vandq_u8(qx6, m3));
+            int8x16_t sqx7 = vreinterpretq_s8_u8(vandq_u8(qx7, m3));
+
+            const int8x16_t qy0 = vld1q_s8(y[i].qs + j*4 +   0);
+            const int8x16_t qy1 = vld1q_s8(y[i].qs + j*4 +  16);
+            const int8x16_t qy2 = vld1q_s8(y[i].qs + j*4 +  32);
+            const int8x16_t qy3 = vld1q_s8(y[i].qs + j*4 +  48);
+            const int8x16_t qy4 = vld1q_s8(y[i].qs + j*4 +  64);
+            const int8x16_t qy5 = vld1q_s8(y[i].qs + j*4 +  80);
+            const int8x16_t qy6 = vld1q_s8(y[i].qs + j*4 +  96);
+            const int8x16_t qy7 = vld1q_s8(y[i].qs + j*4 + 112);
+
+#if defined(__ARM_FEATURE_DOTPROD)
+            sumi0 = vdotq_s32(sumi0, sqx0, qy0);
+            sumi1 = vdotq_s32(sumi1, sqx1, qy1);
+            sumi0 = vdotq_s32(sumi0, sqx2, qy2);
+            sumi1 = vdotq_s32(sumi1, sqx3, qy3);
+            sumi0 = vdotq_s32(sumi0, sqx4, qy4);
+            sumi1 = vdotq_s32(sumi1, sqx5, qy5);
+            sumi0 = vdotq_s32(sumi0, sqx6, qy6);
+            sumi1 = vdotq_s32(sumi1, sqx7, qy7);
+#else
+            sumi0 = vmlal_s8(sumi0, vget_low_s8(sqx0), vget_low_s8(qy0));
+            sumi1 = vmlal_s8(sumi1, vget_high_s8(sqx0), vget_high_s8(qy0));
+            sumi0 = vmlal_s8(sumi0, vget_low_s8(sqx1), vget_low_s8(qy1));
+            sumi1 = vmlal_s8(sumi1, vget_high_s8(sqx1), vget_high_s8(qy1));
+            sumi0 = vmlal_s8(sumi0, vget_low_s8(sqx2), vget_low_s8(qy2));
+            sumi1 = vmlal_s8(sumi1, vget_high_s8(sqx2), vget_high_s8(qy2));
+            sumi0 = vmlal_s8(sumi0, vget_low_s8(sqx3), vget_low_s8(qy3));
+            sumi1 = vmlal_s8(sumi1, vget_high_s8(sqx3), vget_high_s8(qy3));
+            sumi0 = vmlal_s8(sumi0, vget_low_s8(sqx4), vget_low_s8(qy4));
+            sumi1 = vmlal_s8(sumi1, vget_high_s8(sqx4), vget_high_s8(qy4));
+            sumi0 = vmlal_s8(sumi0, vget_low_s8(sqx5), vget_low_s8(qy5));
+            sumi1 = vmlal_s8(sumi1, vget_high_s8(sqx5), vget_high_s8(qy5));
+            sumi0 = vmlal_s8(sumi0, vget_low_s8(sqx6), vget_low_s8(qy6));
+            sumi1 = vmlal_s8(sumi1, vget_high_s8(sqx6), vget_high_s8(qy6));
+            sumi0 = vmlal_s8(sumi0, vget_low_s8(sqx7), vget_low_s8(qy7));
+            sumi1 = vmlal_s8(sumi1, vget_high_s8(sqx7), vget_high_s8(qy7));
+#endif
+        }
+
+        const int16x8_t ysum0 = vld1q_s16(y[i].bsums);
+        const int16x8_t ysum1 = vld1q_s16(y[i].bsums + 8);
+
+        const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
+
+#if defined(__ARM_FEATURE_DOTPROD)
+        sumi0 = vaddq_s32(sumi0, sumi1);
+        sumi0 = vsubq_s32(sumi0, vpaddlq_s16(vaddq_s16(ysum0, ysum1)));
+
+        sumf += d * (float) vaddvq_s32(sumi0);
+#else
+        sumi0 = vaddq_s16(sumi0, sumi1);
+        sumi0 = vsubq_s16(sumi0, vaddq_s16(ysum0, ysum1));
+
+        sumf += d * (float) vaddlvq_s16(sumi0);
+#endif
+    }
+
+    *s = sumf;
+
+#elif defined(__AVX2__)
+    __m256 sumf = _mm256_setzero_ps();
+
+    for (int i = 0; i < nb; ++i) {
+        // 16-bit sums, because 256*127 still fits
+        __m256i sumi0 = _mm256_setzero_si256();
+        __m256i sumi1 = _mm256_setzero_si256();
+
+        for (size_t j = 0; j < sizeof(x->qs); j += 32) {
+            __m256i qx0 = _mm256_loadu_si256((const __m256i *) (x[i].qs + j));
+            __m256i qx1 = _mm256_srli_epi16(qx0, 2);
+            __m256i qx2 = _mm256_srli_epi16(qx0, 4);
+            __m256i qx3 = _mm256_srli_epi16(qx0, 6);
+
+            // 0, 1, 2 (should not be 3)
+            qx0 = _mm256_and_si256(qx0, _mm256_set1_epi8(3));
+            qx1 = _mm256_and_si256(qx1, _mm256_set1_epi8(3));
+            qx2 = _mm256_and_si256(qx2, _mm256_set1_epi8(3));
+            qx3 = _mm256_and_si256(qx3, _mm256_set1_epi8(3));
+
+            const __m256i qy0 = _mm256_loadu_si256((const __m256i *) (y[i].qs + j*4 +  0));
+            const __m256i qy1 = _mm256_loadu_si256((const __m256i *) (y[i].qs + j*4 + 32));
+            const __m256i qy2 = _mm256_loadu_si256((const __m256i *) (y[i].qs + j*4 + 64));
+            const __m256i qy3 = _mm256_loadu_si256((const __m256i *) (y[i].qs + j*4 + 96));
+
+            qx0 = _mm256_maddubs_epi16(qx0, qy0);
+            qx1 = _mm256_maddubs_epi16(qx1, qy1);
+            qx2 = _mm256_maddubs_epi16(qx2, qy2);
+            qx3 = _mm256_maddubs_epi16(qx3, qy3);
+
+            sumi0 = _mm256_add_epi16(sumi0, _mm256_add_epi16(qx0, qx1));
+            sumi1 = _mm256_add_epi16(sumi1, _mm256_add_epi16(qx2, qx3));
+        }
+
+        const __m256i ysum = _mm256_loadu_si256((const __m256i *) y[i].bsums);
+        const __m256 d = _mm256_set1_ps(y[i].d * GGML_FP16_TO_FP32(x[i].d));
+
+        sumi0 = _mm256_add_epi16(sumi0, sumi1);
+        sumi0 = _mm256_sub_epi16(sumi0, ysum);
+        sumi0 = _mm256_madd_epi16(sumi0, _mm256_set1_epi16(1));
+
+        sumf = _mm256_add_ps(_mm256_mul_ps(_mm256_cvtepi32_ps(sumi0), d), sumf);
+    }
+
+    *s = hsum_float_8(sumf);
+
+#else
+    float sumf = 0.0f;
+
+    for (int i = 0; i < nb; ++i) {
+        int32_t sumi = 0;
+
+        for (size_t j = 0; j < sizeof(x->qs); j += 32) {
+            for (size_t l = 0; l < 4; ++l) {
+                for (size_t k = 0; k < 32; ++k) {
+                    sumi += y[i].qs[j*4 + l*32 + k] * (((x[i].qs[j + k] >> (l*2)) & 3) - 1);
+                }
+            }
+        }
+
+        const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d);
+
+        sumf += (float) sumi * d;
+    }
+
+    *s = sumf;
+#endif
+}
+
+void ggml_vec_dot_q2_K_q8_K(int n, float * restrict s, size_t bs, const void * restrict vx, size_t bx, const void * restrict vy, size_t by, int nrc) {
+    assert(nrc == 1);
+    UNUSED(nrc);
+    UNUSED(bx);
+    UNUSED(by);
+    UNUSED(bs);
+
+    const block_q2_K * restrict x = vx;
+    const block_q8_K * restrict y = vy;
+
+    const int nb = n / QK_K;
+
+#ifdef __ARM_NEON
+    const uint8x16_t m3 = vdupq_n_u8(0x3);
+    const uint8x16_t m4 = vdupq_n_u8(0xF);
+
+    const int32x4_t vzero = vdupq_n_s32(0);
+
+    ggml_int8x16x2_t q2bytes;
+    uint8_t aux[16];
+
+    float sum = 0;
+
+    for (int i = 0; i < nb; ++i) {
+        const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d);
+        const float dmin = -y[i].d * GGML_FP16_TO_FP32(x[i].dmin);
+
+        const uint8_t * restrict q2 = x[i].qs;
+        const int8_t  * restrict q8 = y[i].qs;
+        const uint8_t * restrict sc = x[i].scales;
+
+        const uint8x16_t mins_and_scales = vld1q_u8(sc);
+        const uint8x16_t scales = vandq_u8(mins_and_scales, m4);
+        vst1q_u8(aux, scales);
+
+        const uint8x16_t mins = vshrq_n_u8(mins_and_scales, 4);
+        const ggml_int16x8x2_t q8sums = ggml_vld1q_s16_x2(y[i].bsums);
+        const ggml_int16x8x2_t mins16 = {{vreinterpretq_s16_u16(vmovl_u8(vget_low_u8(mins))), vreinterpretq_s16_u16(vmovl_u8(vget_high_u8(mins)))}};
+        const int32x4_t s0 = vaddq_s32(vmull_s16(vget_low_s16 (mins16.val[0]), vget_low_s16 (q8sums.val[0])),
+                                       vmull_s16(vget_high_s16(mins16.val[0]), vget_high_s16(q8sums.val[0])));
+        const int32x4_t s1 = vaddq_s32(vmull_s16(vget_low_s16 (mins16.val[1]), vget_low_s16 (q8sums.val[1])),
+                                       vmull_s16(vget_high_s16(mins16.val[1]), vget_high_s16(q8sums.val[1])));
+        sum += dmin * vaddvq_s32(vaddq_s32(s0, s1));
+
+        int isum = 0;
+        int is = 0;
+
+// We use this macro instead of a function call because for some reason
+// the code runs 2-3% slower, even if the function is declared inline
+#define MULTIPLY_ACCUM_WITH_SCALE(index)\
+        isum += vaddvq_s32(ggml_vdotq_s32(vzero, q2bytes.val[0], q8bytes.val[0])) * aux[is+(index)];\
+        isum += vaddvq_s32(ggml_vdotq_s32(vzero, q2bytes.val[1], q8bytes.val[1])) * aux[is+1+(index)];
+
+#define SHIFT_MULTIPLY_ACCUM_WITH_SCALE(shift, index)\
+        q8bytes = ggml_vld1q_s8_x2(q8); q8 += 32;\
+        q2bytes.val[0] = vreinterpretq_s8_u8(vandq_u8(vshrq_n_u8(q2bits.val[0], (shift)), m3));\
+        q2bytes.val[1] = vreinterpretq_s8_u8(vandq_u8(vshrq_n_u8(q2bits.val[1], (shift)), m3));\
+        MULTIPLY_ACCUM_WITH_SCALE((index));
+
+        for (int j = 0; j < QK_K/128; ++j) {
+            const ggml_uint8x16x2_t q2bits = ggml_vld1q_u8_x2(q2); q2 += 32;
+
+            ggml_int8x16x2_t q8bytes = ggml_vld1q_s8_x2(q8); q8 += 32;
+            q2bytes.val[0] = vreinterpretq_s8_u8(vandq_u8(q2bits.val[0], m3));
+            q2bytes.val[1] = vreinterpretq_s8_u8(vandq_u8(q2bits.val[1], m3));
+
+            MULTIPLY_ACCUM_WITH_SCALE(0);
+
+            SHIFT_MULTIPLY_ACCUM_WITH_SCALE(2, 2);
+            SHIFT_MULTIPLY_ACCUM_WITH_SCALE(4, 4);
+            SHIFT_MULTIPLY_ACCUM_WITH_SCALE(6, 6);
+
+            is += 8;
+        }
+
+        sum += d * isum;
+    }
+
+    *s = sum;
+
+#elif defined __AVX2__
+
+    const __m256i m3 = _mm256_set1_epi8(3);
+    const __m128i m4 = _mm_set1_epi8(0xF);
+
+    __m256 acc = _mm256_setzero_ps();
+
+    for (int i = 0; i < nb; ++i) {
+
+        const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d);
+        const float dmin = -y[i].d * GGML_FP16_TO_FP32(x[i].dmin);
+
+        const uint8_t * restrict q2 = x[i].qs;
+        const int8_t  * restrict q8 = y[i].qs;
+
+        const __m128i mins_and_scales = _mm_loadu_si128((const __m128i*)x[i].scales);
+        const __m128i scales8 = _mm_and_si128(mins_and_scales, m4);
+        const __m128i mins8 = _mm_and_si128(_mm_srli_epi16(mins_and_scales, 4), m4);
+        const __m256i mins = _mm256_cvtepi8_epi16(mins8);
+        const __m256i prod = _mm256_madd_epi16(mins, _mm256_loadu_si256((const __m256i*)y[i].bsums));
+
+        acc = _mm256_fmadd_ps(_mm256_broadcast_ss(&dmin), _mm256_cvtepi32_ps(prod), acc);
+
+        const __m256i all_scales = _mm256_cvtepi8_epi16(scales8);
+        const __m128i l_scales = _mm256_extracti128_si256(all_scales, 0);
+        const __m128i h_scales = _mm256_extracti128_si256(all_scales, 1);
+        const __m256i scales[2] = {MM256_SET_M128I(l_scales, l_scales), MM256_SET_M128I(h_scales, h_scales)};
+
+        __m256i sumi = _mm256_setzero_si256();
+
+        for (int j = 0; j < QK_K/128; ++j) {
+
+            const __m256i q2bits = _mm256_loadu_si256((const __m256i*)q2); q2 += 32;
+
+            const __m256i q8_0 = _mm256_loadu_si256((const __m256i*)q8); q8 += 32;
+            const __m256i q8_1 = _mm256_loadu_si256((const __m256i*)q8); q8 += 32;
+            const __m256i q8_2 = _mm256_loadu_si256((const __m256i*)q8); q8 += 32;
+            const __m256i q8_3 = _mm256_loadu_si256((const __m256i*)q8); q8 += 32;
+
+            const __m256i q2_0 = _mm256_and_si256(q2bits, m3);
+            const __m256i q2_1 = _mm256_and_si256(_mm256_srli_epi16(q2bits, 2), m3);
+            const __m256i q2_2 = _mm256_and_si256(_mm256_srli_epi16(q2bits, 4), m3);
+            const __m256i q2_3 = _mm256_and_si256(_mm256_srli_epi16(q2bits, 6), m3);
+
+            __m256i p0 = _mm256_maddubs_epi16(q2_0, q8_0);
+            __m256i p1 = _mm256_maddubs_epi16(q2_1, q8_1);
+            __m256i p2 = _mm256_maddubs_epi16(q2_2, q8_2);
+            __m256i p3 = _mm256_maddubs_epi16(q2_3, q8_3);
+
+            p0 = _mm256_madd_epi16(_mm256_shuffle_epi8(scales[j], get_scale_shuffle_q3k(0)), p0);
+            p1 = _mm256_madd_epi16(_mm256_shuffle_epi8(scales[j], get_scale_shuffle_q3k(1)), p1);
+            p2 = _mm256_madd_epi16(_mm256_shuffle_epi8(scales[j], get_scale_shuffle_q3k(2)), p2);
+            p3 = _mm256_madd_epi16(_mm256_shuffle_epi8(scales[j], get_scale_shuffle_q3k(3)), p3);
+
+            p0 = _mm256_add_epi32(p0, p1);
+            p2 = _mm256_add_epi32(p2, p3);
+
+            sumi = _mm256_add_epi32(sumi, _mm256_add_epi32(p0, p2));
+        }
+
+        acc = _mm256_fmadd_ps(_mm256_broadcast_ss(&d), _mm256_cvtepi32_ps(sumi), acc);
+
+    }
+
+    *s = hsum_float_8(acc);
+
+#elif defined __AVX__
+
+    const __m128i m3 = _mm_set1_epi8(0x3);
+    const __m128i m4 = _mm_set1_epi8(0xF);
+    const __m128i m2 = _mm_set1_epi8(0x2);
+
+    __m256 acc = _mm256_setzero_ps();
+
+    for (int i = 0; i < nb; ++i) {
+
+        const float dall = y[i].d * GGML_FP16_TO_FP32(x[i].d);
+        const float dmin = -y[i].d * GGML_FP16_TO_FP32(x[i].dmin);
+
+        const uint8_t * restrict q2 = x[i].qs;
+        const int8_t  * restrict q8 = y[i].qs;
+
+        // load mins and scales from block_q2_K.scales[QK_K/16]
+        const __m128i mins_and_scales = _mm_loadu_si128((const __m128i*)x[i].scales);
+        const __m128i scales16 = _mm_and_si128(mins_and_scales, m4);
+        const __m128i mins16 = _mm_and_si128(_mm_srli_epi16(mins_and_scales, 4), m4);
+        const __m128i mins_0 = _mm_cvtepi8_epi16(mins16);
+        const __m128i mins_1 = _mm_cvtepi8_epi16(_mm_unpackhi_epi64(mins16, mins16));
+
+        // summs = y[i].bsums * (x[i].scales >> 4) in 16bits*8*2 to 32bits*4*2
+        const __m128i summs_0 = _mm_madd_epi16(mins_0, _mm_loadu_si128((const __m128i*)&y[i].bsums[0]));
+        const __m128i summs_1 = _mm_madd_epi16(mins_1, _mm_loadu_si128((const __m128i*)&y[i].bsums[8]));
+
+        // sumf += -dmin * summs in 32bits*8
+        acc = _mm256_add_ps(_mm256_mul_ps(_mm256_broadcast_ss(&dmin), _mm256_cvtepi32_ps(MM256_SET_M128I(summs_1, summs_0))), acc);
+
+        const __m128i scales_0 = _mm_cvtepi8_epi16(scales16);
+        const __m128i scales_1 = _mm_cvtepi8_epi16(_mm_unpackhi_epi64(scales16, scales16));
+        const __m128i scales[2] = { scales_0, scales_1 };
+
+        __m128i sumi_0 = _mm_setzero_si128();
+        __m128i sumi_1 = _mm_setzero_si128();
+
+        for (int j = 0; j < QK_K/128; ++j) {
+
+            // load Q8 quants int8*16*8 from block_q8_K.qs[QK_K]
+            const __m128i q8_0 = _mm_loadu_si128((const __m128i*)q8); q8 += 16;
+            const __m128i q8_1 = _mm_loadu_si128((const __m128i*)q8); q8 += 16;
+            const __m128i q8_2 = _mm_loadu_si128((const __m128i*)q8); q8 += 16;
+            const __m128i q8_3 = _mm_loadu_si128((const __m128i*)q8); q8 += 16;
+            const __m128i q8_4 = _mm_loadu_si128((const __m128i*)q8); q8 += 16;
+            const __m128i q8_5 = _mm_loadu_si128((const __m128i*)q8); q8 += 16;
+            const __m128i q8_6 = _mm_loadu_si128((const __m128i*)q8); q8 += 16;
+            const __m128i q8_7 = _mm_loadu_si128((const __m128i*)q8); q8 += 16;
+
+            // load 2bits*16*8 from block_q2_K.qs[QK_K/4]
+            __m128i q2bits = _mm_loadu_si128((const __m128i*)q2); q2 += 16;
+            const __m128i q2_0 = _mm_and_si128(q2bits, m3);
+            const __m128i q2_2 = _mm_and_si128(_mm_srli_epi16(q2bits, 2), m3);
+            const __m128i q2_4 = _mm_and_si128(_mm_srli_epi16(q2bits, 4), m3);
+            const __m128i q2_6 = _mm_and_si128(_mm_srli_epi16(q2bits, 6), m3);
+            q2bits = _mm_loadu_si128((const __m128i*)q2); q2 += 16;
+            const __m128i q2_1 = _mm_and_si128(q2bits, m3);
+            const __m128i q2_3 = _mm_and_si128(_mm_srli_epi16(q2bits, 2), m3);
+            const __m128i q2_5 = _mm_and_si128(_mm_srli_epi16(q2bits, 4), m3);
+            const __m128i q2_7 = _mm_and_si128(_mm_srli_epi16(q2bits, 6), m3);
+
+            // isuml = q8[l] * ((q2[l] >> shift) & 3) in 8bits*16*8 to 16bits*8*8
+            __m128i p0 = _mm_maddubs_epi16(q2_0, q8_0);
+            __m128i p1 = _mm_maddubs_epi16(q2_1, q8_1);
+            __m128i p2 = _mm_maddubs_epi16(q2_2, q8_2);
+            __m128i p3 = _mm_maddubs_epi16(q2_3, q8_3);
+            __m128i p4 = _mm_maddubs_epi16(q2_4, q8_4);
+            __m128i p5 = _mm_maddubs_epi16(q2_5, q8_5);
+            __m128i p6 = _mm_maddubs_epi16(q2_6, q8_6);
+            __m128i p7 = _mm_maddubs_epi16(q2_7, q8_7);
+
+            // isum += (x[i].scales[is++] & 0xF) * isuml in 16bits*8*8 to 32bits*4*8
+            __m128i shuffle = _mm_set1_epi16(0x0100);
+            p0 = _mm_madd_epi16(_mm_shuffle_epi8(scales[j], shuffle), p0);
+            shuffle = _mm_add_epi16(shuffle, m2);
+            p1 = _mm_madd_epi16(_mm_shuffle_epi8(scales[j], shuffle), p1);
+            shuffle = _mm_add_epi16(shuffle, m2);
+            p2 = _mm_madd_epi16(_mm_shuffle_epi8(scales[j], shuffle), p2);
+            shuffle = _mm_add_epi16(shuffle, m2);
+            p3 = _mm_madd_epi16(_mm_shuffle_epi8(scales[j], shuffle), p3);
+            shuffle = _mm_add_epi16(shuffle, m2);
+            p4 = _mm_madd_epi16(_mm_shuffle_epi8(scales[j], shuffle), p4);
+            shuffle = _mm_add_epi16(shuffle, m2);
+            p5 = _mm_madd_epi16(_mm_shuffle_epi8(scales[j], shuffle), p5);
+            shuffle = _mm_add_epi16(shuffle, m2);
+            p6 = _mm_madd_epi16(_mm_shuffle_epi8(scales[j], shuffle), p6);
+            shuffle = _mm_add_epi16(shuffle, m2);
+            p7 = _mm_madd_epi16(_mm_shuffle_epi8(scales[j], shuffle), p7);
+
+            p0 = _mm_add_epi32(p0, p1);
+            p2 = _mm_add_epi32(p2, p3);
+            p4 = _mm_add_epi32(p4, p5);
+            p6 = _mm_add_epi32(p6, p7);
+
+            // isum in 32bits*4*2
+            sumi_0 = _mm_add_epi32(sumi_0, _mm_add_epi32(p0, p2));
+            sumi_1 = _mm_add_epi32(sumi_1, _mm_add_epi32(p4, p6));
+        }
+
+        // sumf += dall * isum - dmin * summs in 32bits
+        __m256i sumi = MM256_SET_M128I(sumi_1, sumi_0);
+        acc = _mm256_add_ps(_mm256_mul_ps(_mm256_broadcast_ss(&dall), _mm256_cvtepi32_ps(sumi)), acc);
+    }
+
+    *s = hsum_float_8(acc);
+
+#elif defined __riscv_v_intrinsic
+
+    float sumf = 0;
+    uint8_t temp_01[32] = {0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
+                            1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1};
+
+    for (int i = 0; i < nb; ++i) {
+
+        const uint8_t * q2 = x[i].qs;
+        const  int8_t * q8 = y[i].qs;
+        const uint8_t * sc = x[i].scales;
+
+        const float dall = y[i].d * GGML_FP16_TO_FP32(x[i].d);
+        const float dmin = -y[i].d * GGML_FP16_TO_FP32(x[i].dmin);
+
+        size_t vl = 16;
+
+        vuint8m1_t scales = __riscv_vle8_v_u8m1(sc, vl);
+        vuint8m1_t aux = __riscv_vand_vx_u8m1(scales, 0x0F, vl);
+
+        vint16m1_t q8sums = __riscv_vle16_v_i16m1(y[i].bsums, vl);
+
+        vuint8mf2_t scales_2 = __riscv_vle8_v_u8mf2(sc, vl);
+        vuint8mf2_t mins8 = __riscv_vsrl_vx_u8mf2(scales_2, 0x4, vl);
+        vint16m1_t mins = __riscv_vreinterpret_v_u16m1_i16m1(__riscv_vzext_vf2_u16m1(mins8, vl));
+        vint32m2_t prod = __riscv_vwmul_vv_i32m2(q8sums, mins, vl);
+        vint32m1_t vsums = __riscv_vredsum_vs_i32m2_i32m1(prod, __riscv_vmv_v_x_i32m1(0, 1), vl);
+
+        sumf  += dmin * __riscv_vmv_x_s_i32m1_i32(vsums);
+
+        vl = 32;
+
+        vint32m1_t vzero = __riscv_vmv_v_x_i32m1(0, 1);
+        vuint8m1_t v_b = __riscv_vle8_v_u8m1(temp_01, vl);
+
+        uint8_t is=0;
+        int isum=0;
+
+        for (int j = 0; j < QK_K/128; ++j) {
+            // load Q2
+            vuint8m1_t q2_x = __riscv_vle8_v_u8m1(q2, vl);
+
+            vuint8m1_t q2_0 = __riscv_vand_vx_u8m1(q2_x, 0x03, vl);
+            vuint8m1_t q2_1 = __riscv_vand_vx_u8m1(__riscv_vsrl_vx_u8m1(q2_x, 0x2, vl), 0x03 , vl);
+            vuint8m1_t q2_2 = __riscv_vand_vx_u8m1(__riscv_vsrl_vx_u8m1(q2_x, 0x4, vl), 0x03 , vl);
+            vuint8m1_t q2_3 = __riscv_vand_vx_u8m1(__riscv_vsrl_vx_u8m1(q2_x, 0x6, vl), 0x03 , vl);
+
+            // duplicate scale elements for product
+            vuint8m1_t sc0 = __riscv_vrgather_vv_u8m1(aux, __riscv_vadd_vx_u8m1(v_b, 0+is, vl), vl);
+            vuint8m1_t sc1 = __riscv_vrgather_vv_u8m1(aux, __riscv_vadd_vx_u8m1(v_b, 2+is, vl), vl);
+            vuint8m1_t sc2 = __riscv_vrgather_vv_u8m1(aux, __riscv_vadd_vx_u8m1(v_b, 4+is, vl), vl);
+            vuint8m1_t sc3 = __riscv_vrgather_vv_u8m1(aux, __riscv_vadd_vx_u8m1(v_b, 6+is, vl), vl);
+
+            vint16m2_t p0 = __riscv_vreinterpret_v_u16m2_i16m2(__riscv_vwmulu_vv_u16m2(q2_0, sc0, vl));
+            vint16m2_t p1 = __riscv_vreinterpret_v_u16m2_i16m2(__riscv_vwmulu_vv_u16m2(q2_1, sc1, vl));
+            vint16m2_t p2 = __riscv_vreinterpret_v_u16m2_i16m2(__riscv_vwmulu_vv_u16m2(q2_2, sc2, vl));
+            vint16m2_t p3 = __riscv_vreinterpret_v_u16m2_i16m2(__riscv_vwmulu_vv_u16m2(q2_3, sc3, vl));
+
+            // load Q8
+            vint8m1_t q8_0 = __riscv_vle8_v_i8m1(q8, vl);
+            vint8m1_t q8_1 = __riscv_vle8_v_i8m1(q8+32, vl);
+            vint8m1_t q8_2 = __riscv_vle8_v_i8m1(q8+64, vl);
+            vint8m1_t q8_3 = __riscv_vle8_v_i8m1(q8+96, vl);
+
+            vint32m4_t s0 = __riscv_vwmul_vv_i32m4(p0, __riscv_vwcvt_x_x_v_i16m2(q8_0, vl), vl);
+            vint32m4_t s1 = __riscv_vwmul_vv_i32m4(p1, __riscv_vwcvt_x_x_v_i16m2(q8_1, vl), vl);
+            vint32m4_t s2 = __riscv_vwmul_vv_i32m4(p2, __riscv_vwcvt_x_x_v_i16m2(q8_2, vl), vl);
+            vint32m4_t s3 = __riscv_vwmul_vv_i32m4(p3, __riscv_vwcvt_x_x_v_i16m2(q8_3, vl), vl);
+
+            vint32m1_t isum0 = __riscv_vredsum_vs_i32m4_i32m1(__riscv_vadd_vv_i32m4(s0, s1, vl), vzero, vl);
+            vint32m1_t isum1 = __riscv_vredsum_vs_i32m4_i32m1(__riscv_vadd_vv_i32m4(s2, s3, vl), isum0, vl);
+
+            isum += __riscv_vmv_x_s_i32m1_i32(isum1);
+
+            q2+=32;  q8+=128;  is=8;
+
+        }
+
+        sumf += dall * isum;
+
+    }
+
+    *s = sumf;
+
+#elif defined(__POWER9_VECTOR__)
+    const vector signed char lowMask = vec_splats((signed char)0x3);
+    const vector signed char lowScaleMask = vec_splats((signed char)0xF);
+    const vector int v0 = vec_splats((int32_t)0);
+    const vector unsigned char v2 = vec_splats((unsigned char)0x2);
+    const vector unsigned char v6 = vec_splats((unsigned char)0x6);
+    const vector unsigned char v4 = vec_splats((unsigned char)0x4);
+
+    vector float vsumf0 = vec_splats(0.0f);
+    vector float vsumf1 = vec_splats(0.0f);
+    vector float vsumf2 = vec_splats(0.0f);
+    vector float vsumf3 = vec_splats(0.0f);
+
+    for (int i = 0; i < nb; ++i) {
+        vector float vxd = vec_splats(GGML_FP16_TO_FP32(x[i].d));
+        vector float vyd = vec_splats(y[i].d);
+        vector float vd = vec_mul(vxd, vyd);
+
+        vector float vxmin = vec_splats(GGML_FP16_TO_FP32(x[i].dmin));
+        vector float vdmin = vec_mul(vxmin, vyd);
+
+        vector signed short q8ysums0 = vec_xl( 0, y[i].bsums);
+        vector signed short q8ysums1 = vec_xl(16, y[i].bsums);
+
+        vector signed char q2xmins = (vector signed char)vec_xl( 0, x[i].scales);
+        vector signed char vscales = vec_and(q2xmins, lowScaleMask);
+
+        q2xmins = vec_sr(q2xmins, v4);
+        vector signed short q2xmins0 = vec_unpackh(q2xmins);
+        vector signed short q2xmins1 = vec_unpackl(q2xmins);
+
+        vector signed int prod0 = vec_mule(q2xmins0, q8ysums0);
+        vector signed int prod1 = vec_mulo(q2xmins0, q8ysums0);
+        vector signed int prod2 = vec_mule(q2xmins1, q8ysums1);
+        vector signed int prod3 = vec_mulo(q2xmins1, q8ysums1);
+
+        vsumf0 = vec_nmsub(vec_ctf(prod0, 0), vdmin, vsumf0);
+        vsumf1 = vec_nmsub(vec_ctf(prod1, 0), vdmin, vsumf1);
+        vsumf2 = vec_nmsub(vec_ctf(prod2, 0), vdmin, vsumf2);
+        vsumf3 = vec_nmsub(vec_ctf(prod3, 0), vdmin, vsumf3);
+
+        vector signed int vsumi0 = v0;
+        vector signed int vsumi1 = v0;
+        vector signed int vsumi2 = v0;
+        vector signed int vsumi3 = v0;
+        vector signed int vsumi4 = v0;
+        vector signed int vsumi5 = v0;
+        vector signed int vsumi6 = v0;
+        vector signed int vsumi7 = v0;
+
+        const uint8_t * restrict q2 = x[i].qs;
+        const int8_t  * restrict q8 = y[i].qs;
+
+        for (int j = 0; j < QK_K/128; ++j) {
+            __builtin_prefetch(q2, 0, 1);
+            __builtin_prefetch(q8, 0, 1);
+
+            vector signed char qxs0 = (vector signed char)vec_xl( 0, q2);
+            vector signed char qxs1 = (vector signed char)vec_xl(16, q2);
+            q2 += 32;
+
+            vector unsigned char q2x00 = (vector unsigned char)vec_and(qxs0, lowMask);
+            vector unsigned char q2x01 = (vector unsigned char)vec_and(vec_sr(qxs0, v2), lowMask);
+            vector unsigned char q2x02 = (vector unsigned char)vec_and(vec_sr(qxs0, v4), lowMask);
+            vector unsigned char q2x03 = (vector unsigned char)vec_and(vec_sr(qxs0, v6), lowMask);
+            vector unsigned char q2x10 = (vector unsigned char)vec_and(qxs1, lowMask);
+            vector unsigned char q2x11 = (vector unsigned char)vec_and(vec_sr(qxs1, v2), lowMask);
+            vector unsigned char q2x12 = (vector unsigned char)vec_and(vec_sr(qxs1, v4), lowMask);
+            vector unsigned char q2x13 = (vector unsigned char)vec_and(vec_sr(qxs1, v6), lowMask);
+
+            vector signed char q8y00 = vec_xl(  0, q8);
+            vector signed char q8y10 = vec_xl( 16, q8);
+            vector signed char q8y01 = vec_xl( 32, q8);
+            vector signed char q8y11 = vec_xl( 48, q8);
+            vector signed char q8y02 = vec_xl( 64, q8);
+            vector signed char q8y12 = vec_xl( 80, q8);
+            vector signed char q8y03 = vec_xl( 96, q8);
+            vector signed char q8y13 = vec_xl(112, q8);
+            q8 += 128;
+
+            vector signed int qv0 = vec_msum(q8y00, q2x00, v0);
+            vector signed int qv1 = vec_msum(q8y01, q2x01, v0);
+            vector signed int qv2 = vec_msum(q8y02, q2x02, v0);
+            vector signed int qv3 = vec_msum(q8y03, q2x03, v0);
+            vector signed int qv4 = vec_msum(q8y10, q2x10, v0);
+            vector signed int qv5 = vec_msum(q8y11, q2x11, v0);
+            vector signed int qv6 = vec_msum(q8y12, q2x12, v0);
+            vector signed int qv7 = vec_msum(q8y13, q2x13, v0);
+
+            vector signed short vscales_07 = vec_unpackh(vscales);
+            vector signed int vscales_03 = vec_unpackh(vscales_07);
+            vector signed int vscales_47 = vec_unpackl(vscales_07);
+            vector signed int vs0 = vec_splat(vscales_03, 0);
+            vector signed int vs1 = vec_splat(vscales_03, 1);
+            vector signed int vs2 = vec_splat(vscales_03, 2);
+            vector signed int vs3 = vec_splat(vscales_03, 3);
+            vector signed int vs4 = vec_splat(vscales_47, 0);
+            vector signed int vs5 = vec_splat(vscales_47, 1);
+            vector signed int vs6 = vec_splat(vscales_47, 2);
+            vector signed int vs7 = vec_splat(vscales_47, 3);
+            vscales = vec_sld(vscales, vscales, 8);
+
+            vsumi0 = vec_add(vec_mul(qv0, vs0), vsumi0);
+            vsumi1 = vec_add(vec_mul(qv1, vs2), vsumi1);
+            vsumi2 = vec_add(vec_mul(qv2, vs4), vsumi2);
+            vsumi3 = vec_add(vec_mul(qv3, vs6), vsumi3);
+            vsumi4 = vec_add(vec_mul(qv4, vs1), vsumi4);
+            vsumi5 = vec_add(vec_mul(qv5, vs3), vsumi5);
+            vsumi6 = vec_add(vec_mul(qv6, vs5), vsumi6);
+            vsumi7 = vec_add(vec_mul(qv7, vs7), vsumi7);
+        }
+
+        vsumi0 = vec_add(vsumi0, vsumi4);
+        vsumi1 = vec_add(vsumi1, vsumi5);
+        vsumi2 = vec_add(vsumi2, vsumi6);
+        vsumi3 = vec_add(vsumi3, vsumi7);
+
+        vsumf0 = vec_madd(vec_ctf(vsumi0, 0), vd, vsumf0);
+        vsumf1 = vec_madd(vec_ctf(vsumi1, 0), vd, vsumf1);
+        vsumf2 = vec_madd(vec_ctf(vsumi2, 0), vd, vsumf2);
+        vsumf3 = vec_madd(vec_ctf(vsumi3, 0), vd, vsumf3);
+    }
+
+    vsumf0 = vec_add(vsumf0, vsumf2);
+    vsumf1 = vec_add(vsumf1, vsumf3);
+
+    vsumf0 = vec_add(vsumf0, vsumf1);
+
+    vsumf0 = vec_add(vsumf0, vec_sld(vsumf0, vsumf0, 4));
+    vsumf0 = vec_add(vsumf0, vec_sld(vsumf0, vsumf0, 8));
+
+    *s = vec_extract(vsumf0, 0);
+
+#elif defined __loongarch_asx
+
+    const __m256i m3 = __lasx_xvreplgr2vr_b(3);
+    const __m128i m4 = __lsx_vreplgr2vr_b(0xF);
+
+    __m256 acc = (__m256)__lasx_xvldi(0);
+
+    for (int i = 0; i < nb; ++i) {
+
+        const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d);
+        const float dmin = -y[i].d * GGML_FP16_TO_FP32(x[i].dmin);
+
+        const uint8_t * restrict q2 = x[i].qs;
+        const int8_t  * restrict q8 = y[i].qs;
+
+        const __m128i mins_and_scales = __lsx_vld((const __m128i*)x[i].scales, 0);
+        const __m128i scales8 = __lsx_vand_v(mins_and_scales, m4);
+        const __m128i mins8 = __lsx_vand_v(__lsx_vsrli_h(mins_and_scales, 4), m4);
+        const __m256i mins = lasx_ext8_16(mins8);
+        const __m256i prod = lasx_madd_h(mins, __lasx_xvld((const __m256i*)y[i].bsums, 0));
+
+        acc = __lasx_xvfmadd_s(__lasx_xvreplfr2vr_s(dmin), __lasx_xvffint_s_w(prod), acc);
+
+        const __m256i all_scales = lasx_ext8_16(scales8);
+        const __m128i l_scales = lasx_extracti128(all_scales, 0);
+        const __m128i h_scales = lasx_extracti128(all_scales, 1);
+        const __m256i scales[2] = {lasx_insertf128(l_scales, l_scales), lasx_insertf128(h_scales, h_scales)};
+
+        __m256i sumi = __lasx_xvldi(0);
+
+        for (int j = 0; j < QK_K/128; ++j) {
+
+            const __m256i q2bits = __lasx_xvld((const __m256i*)q2, 0); q2 += 32;
+
+            const __m256i q8_0 = __lasx_xvld((const __m256i*)q8, 0); q8 += 32;
+            const __m256i q8_1 = __lasx_xvld((const __m256i*)q8, 0); q8 += 32;
+            const __m256i q8_2 = __lasx_xvld((const __m256i*)q8, 0); q8 += 32;
+            const __m256i q8_3 = __lasx_xvld((const __m256i*)q8, 0); q8 += 32;
+
+            const __m256i q2_0 = __lasx_xvand_v(q2bits, m3);
+            const __m256i q2_1 = __lasx_xvand_v(__lasx_xvsrli_h(q2bits, 2), m3);
+            const __m256i q2_2 = __lasx_xvand_v(__lasx_xvsrli_h(q2bits, 4), m3);
+            const __m256i q2_3 = __lasx_xvand_v(__lasx_xvsrli_h(q2bits, 6), m3);
+
+            __m256i p0 = lasx_maddubs_h(q2_0, q8_0);
+            __m256i p1 = lasx_maddubs_h(q2_1, q8_1);
+            __m256i p2 = lasx_maddubs_h(q2_2, q8_2);
+            __m256i p3 = lasx_maddubs_h(q2_3, q8_3);
+
+            p0 = lasx_madd_h(lasx_shuffle_b(scales[j], get_scale_shuffle_q3k(0)), p0);
+            p1 = lasx_madd_h(lasx_shuffle_b(scales[j], get_scale_shuffle_q3k(1)), p1);
+            p2 = lasx_madd_h(lasx_shuffle_b(scales[j], get_scale_shuffle_q3k(2)), p2);
+            p3 = lasx_madd_h(lasx_shuffle_b(scales[j], get_scale_shuffle_q3k(3)), p3);
+
+            p0 = __lasx_xvadd_w(p0, p1);
+            p2 = __lasx_xvadd_w(p2, p3);
+
+            sumi = __lasx_xvadd_w(sumi, __lasx_xvadd_w(p0, p2));
+        }
+
+        acc = __lasx_xvfmadd_s(__lasx_xvreplfr2vr_s(d), __lasx_xvffint_s_w(sumi), acc);
+
+    }
+
+    *s = hsum_float_8(acc);
+
+#else
+
+    float sumf = 0;
+
+    for (int i = 0; i < nb; ++i) {
+
+        const uint8_t * q2 = x[i].qs;
+        const  int8_t * q8 = y[i].qs;
+        const uint8_t * sc = x[i].scales;
+
+        int summs = 0;
+        for (int j = 0; j < 16; ++j) {
+            summs += y[i].bsums[j] * (sc[j] >> 4);
+        }
+
+        const float dall = y[i].d * GGML_FP16_TO_FP32(x[i].d);
+        const float dmin = y[i].d * GGML_FP16_TO_FP32(x[i].dmin);
+
+        int isum = 0;
+        int is = 0;
+        int d;
+        for (int k = 0; k < QK_K/128; ++k) {
+            int shift = 0;
+            for (int j = 0; j < 4; ++j) {
+                d = sc[is++] & 0xF;
+                int isuml = 0;
+                for (int l =  0; l < 16; ++l) isuml += q8[l] * ((q2[l] >> shift) & 3);
+                isum += d * isuml;
+                d = sc[is++] & 0xF;
+                isuml = 0;
+                for (int l = 16; l < 32; ++l) isuml += q8[l] * ((q2[l] >> shift) & 3);
+                isum += d * isuml;
+                shift += 2;
+                q8 += 32;
+            }
+            q2 += 32;
+        }
+        sumf += dall * isum - dmin * summs;
+    }
+    *s = sumf;
+#endif
+}
+
+void ggml_vec_dot_q3_K_q8_K(int n, float * restrict s, size_t bs, const void * restrict vx, size_t bx, const void * restrict vy, size_t by, int nrc) {
+    assert(n % QK_K == 0);
+    assert(nrc == 1);
+    UNUSED(nrc);
+    UNUSED(bx);
+    UNUSED(by);
+    UNUSED(bs);
+
+    const uint32_t kmask1 = 0x03030303;
+    const uint32_t kmask2 = 0x0f0f0f0f;
+
+    const block_q3_K * restrict x = vx;
+    const block_q8_K * restrict y = vy;
+
+    const int nb = n / QK_K;
+
+#ifdef __ARM_NEON
+
+    uint32_t aux[3];
+    uint32_t utmp[4];
+
+    const uint8x16_t m3b = vdupq_n_u8(0x3);
+    const int32x4_t  vzero = vdupq_n_s32(0);
+
+    const uint8x16_t m0 = vdupq_n_u8(1);
+    const uint8x16_t m1 = vshlq_n_u8(m0, 1);
+    const uint8x16_t m2 = vshlq_n_u8(m0, 2);
+    const uint8x16_t m3 = vshlq_n_u8(m0, 3);
+    const int8_t m32 = 32;
+
+    ggml_int8x16x4_t q3bytes;
+
+    float sum = 0;
+
+    for (int i = 0; i < nb; ++i) {
+
+        const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d);
+
+        const uint8_t * restrict q3 = x[i].qs;
+        const uint8_t * restrict qh = x[i].hmask;
+        const int8_t  * restrict q8 = y[i].qs;
+
+        ggml_uint8x16x2_t qhbits = ggml_vld1q_u8_x2(qh);
+
+        ggml_uint8x16x4_t q3h;
+
+        int32_t isum = 0;
+
+        // Set up scales
+        memcpy(aux, x[i].scales, 12);
+        utmp[3] = ((aux[1] >> 4) & kmask2) | (((aux[2] >> 6) & kmask1) << 4);
+        utmp[2] = ((aux[0] >> 4) & kmask2) | (((aux[2] >> 4) & kmask1) << 4);
+        utmp[1] = (aux[1] & kmask2) | (((aux[2] >> 2) & kmask1) << 4);
+        utmp[0] = (aux[0] & kmask2) | (((aux[2] >> 0) & kmask1) << 4);
+
+        int8_t * scale = (int8_t *)utmp;
+        for (int j = 0; j < 16; ++j) scale[j] -= m32;
+
+        for (int j = 0; j < QK_K/128; ++j) {
+
+            const ggml_uint8x16x2_t q3bits = ggml_vld1q_u8_x2(q3); q3 += 32;
+            const ggml_int8x16x4_t q8bytes_1 = ggml_vld1q_s8_x4(q8); q8 += 64;
+            const ggml_int8x16x4_t q8bytes_2 = ggml_vld1q_s8_x4(q8); q8 += 64;
+
+            q3h.val[0] = vshlq_n_u8(vbicq_u8(m0, qhbits.val[0]), 2);
+            q3h.val[1] = vshlq_n_u8(vbicq_u8(m0, qhbits.val[1]), 2);
+            q3h.val[2] = vshlq_n_u8(vbicq_u8(m1, qhbits.val[0]), 1);
+            q3h.val[3] = vshlq_n_u8(vbicq_u8(m1, qhbits.val[1]), 1);
+
+            q3bytes.val[0] = vsubq_s8(vreinterpretq_s8_u8(vandq_u8(q3bits.val[0], m3b)), vreinterpretq_s8_u8(q3h.val[0]));
+            q3bytes.val[1] = vsubq_s8(vreinterpretq_s8_u8(vandq_u8(q3bits.val[1], m3b)), vreinterpretq_s8_u8(q3h.val[1]));
+            q3bytes.val[2] = vsubq_s8(vreinterpretq_s8_u8(vandq_u8(vshrq_n_u8(q3bits.val[0], 2), m3b)), vreinterpretq_s8_u8(q3h.val[2]));
+            q3bytes.val[3] = vsubq_s8(vreinterpretq_s8_u8(vandq_u8(vshrq_n_u8(q3bits.val[1], 2), m3b)), vreinterpretq_s8_u8(q3h.val[3]));
+
+            isum += vaddvq_s32(ggml_vdotq_s32(vzero, q3bytes.val[0], q8bytes_1.val[0])) * scale[0];
+            isum += vaddvq_s32(ggml_vdotq_s32(vzero, q3bytes.val[1], q8bytes_1.val[1])) * scale[1];
+            isum += vaddvq_s32(ggml_vdotq_s32(vzero, q3bytes.val[2], q8bytes_1.val[2])) * scale[2];
+            isum += vaddvq_s32(ggml_vdotq_s32(vzero, q3bytes.val[3], q8bytes_1.val[3])) * scale[3];
+
+            scale += 4;
+
+            q3h.val[0] = vbicq_u8(m2, qhbits.val[0]);
+            q3h.val[1] = vbicq_u8(m2, qhbits.val[1]);
+            q3h.val[2] = vshrq_n_u8(vbicq_u8(m3, qhbits.val[0]), 1);
+            q3h.val[3] = vshrq_n_u8(vbicq_u8(m3, qhbits.val[1]), 1);
+
+            q3bytes.val[0] = vsubq_s8(vreinterpretq_s8_u8(vandq_u8(vshrq_n_u8(q3bits.val[0], 4), m3b)), vreinterpretq_s8_u8(q3h.val[0]));
+            q3bytes.val[1] = vsubq_s8(vreinterpretq_s8_u8(vandq_u8(vshrq_n_u8(q3bits.val[1], 4), m3b)), vreinterpretq_s8_u8(q3h.val[1]));
+            q3bytes.val[2] = vsubq_s8(vreinterpretq_s8_u8(vandq_u8(vshrq_n_u8(q3bits.val[0], 6), m3b)), vreinterpretq_s8_u8(q3h.val[2]));
+            q3bytes.val[3] = vsubq_s8(vreinterpretq_s8_u8(vandq_u8(vshrq_n_u8(q3bits.val[1], 6), m3b)), vreinterpretq_s8_u8(q3h.val[3]));
+
+            isum += vaddvq_s32(ggml_vdotq_s32(vzero, q3bytes.val[0], q8bytes_2.val[0])) * scale[0];
+            isum += vaddvq_s32(ggml_vdotq_s32(vzero, q3bytes.val[1], q8bytes_2.val[1])) * scale[1];
+            isum += vaddvq_s32(ggml_vdotq_s32(vzero, q3bytes.val[2], q8bytes_2.val[2])) * scale[2];
+            isum += vaddvq_s32(ggml_vdotq_s32(vzero, q3bytes.val[3], q8bytes_2.val[3])) * scale[3];
+
+            scale += 4;
+
+            if (j == 0) {
+                qhbits.val[0] = vshrq_n_u8(qhbits.val[0], 4);
+                qhbits.val[1] = vshrq_n_u8(qhbits.val[1], 4);
+            }
+
+        }
+        sum += d * isum;
+
+    }
+
+    *s = sum;
+
+#elif defined __AVX2__
+
+    const __m256i m3 = _mm256_set1_epi8(3);
+    const __m256i mone = _mm256_set1_epi8(1);
+    const __m128i m32 = _mm_set1_epi8(32);
+
+    __m256 acc = _mm256_setzero_ps();
+
+    uint32_t aux[3];
+
+    for (int i = 0; i < nb; ++i) {
+
+        const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d);
+
+        const uint8_t * restrict q3 = x[i].qs;
+        const int8_t  * restrict q8 = y[i].qs;
+
+        // Set up scales
+        memcpy(aux, x[i].scales, 12);
+        __m128i scales128 = _mm_set_epi32(
+                ((aux[1] >> 4) & kmask2) | (((aux[2] >> 6) & kmask1) << 4),
+                ((aux[0] >> 4) & kmask2) | (((aux[2] >> 4) & kmask1) << 4),
+                (aux[1] & kmask2) | (((aux[2] >> 2) & kmask1) << 4),
+                (aux[0] & kmask2) | (((aux[2] >> 0) & kmask1) << 4));
+        scales128 = _mm_sub_epi8(scales128, m32);
+        const __m256i all_scales = _mm256_cvtepi8_epi16(scales128);
+        const __m128i l_scales = _mm256_extracti128_si256(all_scales, 0);
+        const __m128i h_scales = _mm256_extracti128_si256(all_scales, 1);
+        const __m256i scales[2] = {MM256_SET_M128I(l_scales, l_scales), MM256_SET_M128I(h_scales, h_scales)};
+
+        // high bit
+        const __m256i hbits = _mm256_loadu_si256((const __m256i*)x[i].hmask);
+
+        // integer accumulator
+        __m256i sumi = _mm256_setzero_si256();
+
+        int bit = 0;
+        int is  = 0;
+
+        for (int j = 0; j < QK_K/128; ++j) {
+            // load low 2 bits
+            const __m256i q3bits = _mm256_loadu_si256((const __m256i*)q3); q3 += 32;
+
+            // prepare low and high bits
+            const __m256i q3l_0 = _mm256_and_si256(q3bits, m3);
+            const __m256i q3h_0 = _mm256_slli_epi16(_mm256_srli_epi16(_mm256_andnot_si256(hbits, _mm256_slli_epi16(mone, bit)), bit), 2);
+            ++bit;
+
+            const __m256i q3l_1 = _mm256_and_si256(_mm256_srli_epi16(q3bits, 2), m3);
+            const __m256i q3h_1 = _mm256_slli_epi16(_mm256_srli_epi16(_mm256_andnot_si256(hbits, _mm256_slli_epi16(mone, bit)), bit), 2);
+            ++bit;
+
+            const __m256i q3l_2 = _mm256_and_si256(_mm256_srli_epi16(q3bits, 4), m3);
+            const __m256i q3h_2 = _mm256_slli_epi16(_mm256_srli_epi16(_mm256_andnot_si256(hbits, _mm256_slli_epi16(mone, bit)), bit), 2);
+            ++bit;
+
+            const __m256i q3l_3 = _mm256_and_si256(_mm256_srli_epi16(q3bits, 6), m3);
+            const __m256i q3h_3 = _mm256_slli_epi16(_mm256_srli_epi16(_mm256_andnot_si256(hbits, _mm256_slli_epi16(mone, bit)), bit), 2);
+            ++bit;
+
+            // load Q8 quants
+            const __m256i q8_0 = _mm256_loadu_si256((const __m256i*)q8); q8 += 32;
+            const __m256i q8_1 = _mm256_loadu_si256((const __m256i*)q8); q8 += 32;
+            const __m256i q8_2 = _mm256_loadu_si256((const __m256i*)q8); q8 += 32;
+            const __m256i q8_3 = _mm256_loadu_si256((const __m256i*)q8); q8 += 32;
+
+            // Dot product: we multiply the 2 low bits and 1 high bit part separately, so we can use _mm256_maddubs_epi16,
+            // and then subtract. The high bit part has the 2 already subtracted (and so, it is zero if the high bit was not set,
+            // and 2 if the high bit was set)
+            __m256i q8s_0 = _mm256_maddubs_epi16(q3h_0, q8_0);
+            __m256i q8s_1 = _mm256_maddubs_epi16(q3h_1, q8_1);
+            __m256i q8s_2 = _mm256_maddubs_epi16(q3h_2, q8_2);
+            __m256i q8s_3 = _mm256_maddubs_epi16(q3h_3, q8_3);
+
+            __m256i p16_0 = _mm256_maddubs_epi16(q3l_0, q8_0);
+            __m256i p16_1 = _mm256_maddubs_epi16(q3l_1, q8_1);
+            __m256i p16_2 = _mm256_maddubs_epi16(q3l_2, q8_2);
+            __m256i p16_3 = _mm256_maddubs_epi16(q3l_3, q8_3);
+
+            p16_0 = _mm256_sub_epi16(p16_0, q8s_0);
+            p16_1 = _mm256_sub_epi16(p16_1, q8s_1);
+            p16_2 = _mm256_sub_epi16(p16_2, q8s_2);
+            p16_3 = _mm256_sub_epi16(p16_3, q8s_3);
+
+            // multiply with scales
+            p16_0 = _mm256_madd_epi16(_mm256_shuffle_epi8(scales[j], get_scale_shuffle_q3k(is + 0)), p16_0);
+            p16_1 = _mm256_madd_epi16(_mm256_shuffle_epi8(scales[j], get_scale_shuffle_q3k(is + 1)), p16_1);
+            p16_2 = _mm256_madd_epi16(_mm256_shuffle_epi8(scales[j], get_scale_shuffle_q3k(is + 2)), p16_2);
+            p16_3 = _mm256_madd_epi16(_mm256_shuffle_epi8(scales[j], get_scale_shuffle_q3k(is + 3)), p16_3);
+
+            // accumulate
+            p16_0 = _mm256_add_epi32(p16_0, p16_1);
+            p16_2 = _mm256_add_epi32(p16_2, p16_3);
+            sumi  = _mm256_add_epi32(sumi, _mm256_add_epi32(p16_0, p16_2));
+
+        }
+
+        // multiply with block scale and accumulate
+        acc = _mm256_fmadd_ps(_mm256_broadcast_ss(&d), _mm256_cvtepi32_ps(sumi), acc);
+
+    }
+
+    *s = hsum_float_8(acc);
+
+#elif defined __AVX__
+
+    const __m128i m3 = _mm_set1_epi8(3);
+    const __m128i mone = _mm_set1_epi8(1);
+    const __m128i m32 = _mm_set1_epi8(32);
+    const __m128i m2 = _mm_set1_epi8(2);
+
+    __m256 acc = _mm256_setzero_ps();
+
+    const uint32_t *aux;
+
+    for (int i = 0; i < nb; ++i) {
+
+        const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d);
+
+        const uint8_t * restrict q3 = x[i].qs;
+        const int8_t  * restrict q8 = y[i].qs;
+
+        // Set up scales
+        aux = (const uint32_t *)x[i].scales;
+        __m128i scales128 = _mm_set_epi32(
+                ((aux[1] >> 4) & kmask2) | (((aux[2] >> 6) & kmask1) << 4),
+                ((aux[0] >> 4) & kmask2) | (((aux[2] >> 4) & kmask1) << 4),
+                (aux[1] & kmask2) | (((aux[2] >> 2) & kmask1) << 4),
+                (aux[0] & kmask2) | (((aux[2] >> 0) & kmask1) << 4));
+        scales128 = _mm_sub_epi8(scales128, m32);
+        const __m128i scales_0 = _mm_cvtepi8_epi16(scales128);
+        const __m128i scales_1 = _mm_cvtepi8_epi16(_mm_unpackhi_epi64(scales128, scales128));
+        const __m128i scales[2] = { scales_0, scales_1 };
+
+        // high bit *128*2 from block_q3_K.hmask[QK_K/8]
+        const __m128i hbits_0 = _mm_loadu_si128((const __m128i*)&x[i].hmask[0]);
+        const __m128i hbits_1 = _mm_loadu_si128((const __m128i*)&x[i].hmask[16]);
+
+        // integer accumulator
+        __m128i sumi_0 = _mm_setzero_si128();
+        __m128i sumi_1 = _mm_setzero_si128();
+
+        for (int j = 0; j < QK_K/128; ++j) {
+            // load low 2 bits *64*2 from block_q3_K.qs[QK_K/4]
+            const __m128i q3bits_0 = _mm_loadu_si128((const __m128i*)q3); q3 += 16;
+            const __m128i q3bits_1 = _mm_loadu_si128((const __m128i*)q3); q3 += 16;
+
+            // prepare low and high bits
+            const int bit = j << 2;
+
+            const __m128i q3l_0 = _mm_and_si128(q3bits_0, m3);
+            const __m128i q3l_1 = _mm_and_si128(q3bits_1, m3);
+            const __m128i q3h_0 = _mm_slli_epi16(_mm_srli_epi16(_mm_andnot_si128(hbits_0, _mm_slli_epi16(mone, bit)), bit), 2);
+            const __m128i q3h_1 = _mm_slli_epi16(_mm_srli_epi16(_mm_andnot_si128(hbits_1, _mm_slli_epi16(mone, bit)), bit), 2);
+
+            const __m128i q3l_2 = _mm_and_si128(_mm_srli_epi16(q3bits_0, 2), m3);
+            const __m128i q3l_3 = _mm_and_si128(_mm_srli_epi16(q3bits_1, 2), m3);
+            const __m128i q3h_2 = _mm_slli_epi16(_mm_srli_epi16(_mm_andnot_si128(hbits_0, _mm_slli_epi16(mone, bit+1)), bit+1), 2);
+            const __m128i q3h_3 = _mm_slli_epi16(_mm_srli_epi16(_mm_andnot_si128(hbits_1, _mm_slli_epi16(mone, bit+1)), bit+1), 2);
+
+            const __m128i q3l_4 = _mm_and_si128(_mm_srli_epi16(q3bits_0, 4), m3);
+            const __m128i q3l_5 = _mm_and_si128(_mm_srli_epi16(q3bits_1, 4), m3);
+            const __m128i q3h_4 = _mm_slli_epi16(_mm_srli_epi16(_mm_andnot_si128(hbits_0, _mm_slli_epi16(mone, bit+2)), bit+2), 2);
+            const __m128i q3h_5 = _mm_slli_epi16(_mm_srli_epi16(_mm_andnot_si128(hbits_1, _mm_slli_epi16(mone, bit+2)), bit+2), 2);
+
+            const __m128i q3l_6 = _mm_and_si128(_mm_srli_epi16(q3bits_0, 6), m3);
+            const __m128i q3l_7 = _mm_and_si128(_mm_srli_epi16(q3bits_1, 6), m3);
+            const __m128i q3h_6 = _mm_slli_epi16(_mm_srli_epi16(_mm_andnot_si128(hbits_0, _mm_slli_epi16(mone, bit+3)), bit+3), 2);
+            const __m128i q3h_7 = _mm_slli_epi16(_mm_srli_epi16(_mm_andnot_si128(hbits_1, _mm_slli_epi16(mone, bit+3)), bit+3), 2);
+
+            // load Q8 quants from block_q8_K.qs[QK_K]
+            const __m128i q8_0 = _mm_loadu_si128((const __m128i*)q8); q8 += 16;
+            const __m128i q8_1 = _mm_loadu_si128((const __m128i*)q8); q8 += 16;
+            const __m128i q8_2 = _mm_loadu_si128((const __m128i*)q8); q8 += 16;
+            const __m128i q8_3 = _mm_loadu_si128((const __m128i*)q8); q8 += 16;
+            const __m128i q8_4 = _mm_loadu_si128((const __m128i*)q8); q8 += 16;
+            const __m128i q8_5 = _mm_loadu_si128((const __m128i*)q8); q8 += 16;
+            const __m128i q8_6 = _mm_loadu_si128((const __m128i*)q8); q8 += 16;
+            const __m128i q8_7 = _mm_loadu_si128((const __m128i*)q8); q8 += 16;
+
+            // Dot product: we multiply the 2 low bits and 1 high bit part separately, so we can use _mm256_maddubs_epi16,
+            // and then subtract. The high bit part has the 2 already subtracted (and so, it is zero if the high bit was not set,
+            // and 2 if the high bit was set)
+            __m128i q8s_0 = _mm_maddubs_epi16(q3h_0, q8_0);
+            __m128i q8s_1 = _mm_maddubs_epi16(q3h_1, q8_1);
+            __m128i q8s_2 = _mm_maddubs_epi16(q3h_2, q8_2);
+            __m128i q8s_3 = _mm_maddubs_epi16(q3h_3, q8_3);
+            __m128i q8s_4 = _mm_maddubs_epi16(q3h_4, q8_4);
+            __m128i q8s_5 = _mm_maddubs_epi16(q3h_5, q8_5);
+            __m128i q8s_6 = _mm_maddubs_epi16(q3h_6, q8_6);
+            __m128i q8s_7 = _mm_maddubs_epi16(q3h_7, q8_7);
+
+            __m128i p16_0 = _mm_maddubs_epi16(q3l_0, q8_0);
+            __m128i p16_1 = _mm_maddubs_epi16(q3l_1, q8_1);
+            __m128i p16_2 = _mm_maddubs_epi16(q3l_2, q8_2);
+            __m128i p16_3 = _mm_maddubs_epi16(q3l_3, q8_3);
+            __m128i p16_4 = _mm_maddubs_epi16(q3l_4, q8_4);
+            __m128i p16_5 = _mm_maddubs_epi16(q3l_5, q8_5);
+            __m128i p16_6 = _mm_maddubs_epi16(q3l_6, q8_6);
+            __m128i p16_7 = _mm_maddubs_epi16(q3l_7, q8_7);
+
+            p16_0 = _mm_sub_epi16(p16_0, q8s_0);
+            p16_1 = _mm_sub_epi16(p16_1, q8s_1);
+            p16_2 = _mm_sub_epi16(p16_2, q8s_2);
+            p16_3 = _mm_sub_epi16(p16_3, q8s_3);
+            p16_4 = _mm_sub_epi16(p16_4, q8s_4);
+            p16_5 = _mm_sub_epi16(p16_5, q8s_5);
+            p16_6 = _mm_sub_epi16(p16_6, q8s_6);
+            p16_7 = _mm_sub_epi16(p16_7, q8s_7);
+
+            // multiply with scales
+            __m128i shuffle = _mm_set1_epi16(0x0100);
+            p16_0 = _mm_madd_epi16(_mm_shuffle_epi8(scales[j], shuffle), p16_0);
+            shuffle = _mm_add_epi16(shuffle, m2);
+            p16_1 = _mm_madd_epi16(_mm_shuffle_epi8(scales[j], shuffle), p16_1);
+            shuffle = _mm_add_epi16(shuffle, m2);
+            p16_2 = _mm_madd_epi16(_mm_shuffle_epi8(scales[j], shuffle), p16_2);
+            shuffle = _mm_add_epi16(shuffle, m2);
+            p16_3 = _mm_madd_epi16(_mm_shuffle_epi8(scales[j], shuffle), p16_3);
+            shuffle = _mm_add_epi16(shuffle, m2);
+            p16_4 = _mm_madd_epi16(_mm_shuffle_epi8(scales[j], shuffle), p16_4);
+            shuffle = _mm_add_epi16(shuffle, m2);
+            p16_5 = _mm_madd_epi16(_mm_shuffle_epi8(scales[j], shuffle), p16_5);
+            shuffle = _mm_add_epi16(shuffle, m2);
+            p16_6 = _mm_madd_epi16(_mm_shuffle_epi8(scales[j], shuffle), p16_6);
+            shuffle = _mm_add_epi16(shuffle, m2);
+            p16_7 = _mm_madd_epi16(_mm_shuffle_epi8(scales[j], shuffle), p16_7);
+
+            // accumulate
+            p16_0 = _mm_add_epi32(p16_0, p16_1);
+            p16_2 = _mm_add_epi32(p16_2, p16_3);
+            p16_4 = _mm_add_epi32(p16_4, p16_5);
+            p16_6 = _mm_add_epi32(p16_6, p16_7);
+            sumi_0 = _mm_add_epi32(sumi_0, _mm_add_epi32(p16_0, p16_2));
+            sumi_1 = _mm_add_epi32(sumi_1, _mm_add_epi32(p16_4, p16_6));
+
+        }
+
+        // multiply with block scale and accumulate
+        __m256i sumi = MM256_SET_M128I(sumi_1, sumi_0);
+        acc = _mm256_add_ps(_mm256_mul_ps(_mm256_broadcast_ss(&d), _mm256_cvtepi32_ps(sumi)), acc);
+
+    }
+
+    *s = hsum_float_8(acc);
+
+#elif defined __riscv_v_intrinsic
+
+    uint32_t aux[3];
+    uint32_t utmp[4];
+
+    float sumf = 0;
+    for (int i = 0; i < nb; ++i) {
+
+        const uint8_t * restrict q3 = x[i].qs;
+        const uint8_t * restrict qh = x[i].hmask;
+        const  int8_t * restrict q8 = y[i].qs;
+
+        memcpy(aux, x[i].scales, 12);
+        utmp[3] = ((aux[1] >> 4) & kmask2) | (((aux[2] >> 6) & kmask1) << 4);
+        utmp[2] = ((aux[0] >> 4) & kmask2) | (((aux[2] >> 4) & kmask1) << 4);
+        utmp[1] = (aux[1] & kmask2) | (((aux[2] >> 2) & kmask1) << 4);
+        utmp[0] = (aux[0] & kmask2) | (((aux[2] >> 0) & kmask1) << 4);
+
+        int8_t * scale = (int8_t *)utmp;
+        for (int j = 0; j < 16; ++j) scale[j] -= 32;
+
+
+        size_t vl = 32;
+        uint8_t m =  1;
+
+        vint32m1_t vzero = __riscv_vmv_v_x_i32m1(0, 1);
+        vuint8m1_t vqh = __riscv_vle8_v_u8m1(qh, vl);
+
+        int sum_t = 0;
+
+        for (int j = 0; j < QK_K; j += 128) {
+
+            vl = 32;
+
+            // load Q3
+            vuint8m1_t q3_x = __riscv_vle8_v_u8m1(q3, vl);
+
+            vint8m1_t q3_0 = __riscv_vreinterpret_v_u8m1_i8m1(__riscv_vand_vx_u8m1(q3_x, 0x03, vl));
+            vint8m1_t q3_1 = __riscv_vreinterpret_v_u8m1_i8m1(__riscv_vand_vx_u8m1(__riscv_vsrl_vx_u8m1(q3_x, 0x2, vl), 0x03 , vl));
+            vint8m1_t q3_2 = __riscv_vreinterpret_v_u8m1_i8m1(__riscv_vand_vx_u8m1(__riscv_vsrl_vx_u8m1(q3_x, 0x4, vl), 0x03 , vl));
+            vint8m1_t q3_3 = __riscv_vreinterpret_v_u8m1_i8m1(__riscv_vand_vx_u8m1(__riscv_vsrl_vx_u8m1(q3_x, 0x6, vl), 0x03 , vl));
+
+            // compute mask for subtraction
+            vuint8m1_t qh_m0 = __riscv_vand_vx_u8m1(vqh, m, vl);
+            vbool8_t vmask_0 = __riscv_vmseq_vx_u8m1_b8(qh_m0, 0, vl);
+            vint8m1_t q3_m0 = __riscv_vsub_vx_i8m1_mu(vmask_0, q3_0, q3_0, 0x4, vl);
+            m <<= 1;
+
+            vuint8m1_t qh_m1 = __riscv_vand_vx_u8m1(vqh, m, vl);
+            vbool8_t vmask_1 = __riscv_vmseq_vx_u8m1_b8(qh_m1, 0, vl);
+            vint8m1_t q3_m1 = __riscv_vsub_vx_i8m1_mu(vmask_1, q3_1, q3_1, 0x4, vl);
+            m <<= 1;
+
+            vuint8m1_t qh_m2 = __riscv_vand_vx_u8m1(vqh, m, vl);
+            vbool8_t vmask_2 = __riscv_vmseq_vx_u8m1_b8(qh_m2, 0, vl);
+            vint8m1_t q3_m2 = __riscv_vsub_vx_i8m1_mu(vmask_2, q3_2, q3_2, 0x4, vl);
+            m <<= 1;
+
+            vuint8m1_t qh_m3 = __riscv_vand_vx_u8m1(vqh, m, vl);
+            vbool8_t vmask_3 = __riscv_vmseq_vx_u8m1_b8(qh_m3, 0, vl);
+            vint8m1_t q3_m3 = __riscv_vsub_vx_i8m1_mu(vmask_3, q3_3, q3_3, 0x4, vl);
+            m <<= 1;
+
+            // load Q8 and take product with Q3
+            vint16m2_t a0 = __riscv_vwmul_vv_i16m2(q3_m0, __riscv_vle8_v_i8m1(q8, vl), vl);
+            vint16m2_t a1 = __riscv_vwmul_vv_i16m2(q3_m1, __riscv_vle8_v_i8m1(q8+32, vl), vl);
+            vint16m2_t a2 = __riscv_vwmul_vv_i16m2(q3_m2, __riscv_vle8_v_i8m1(q8+64, vl), vl);
+            vint16m2_t a3 = __riscv_vwmul_vv_i16m2(q3_m3, __riscv_vle8_v_i8m1(q8+96, vl), vl);
+
+            vl = 16;
+
+            // retrieve lane to multiply with scale
+            vint32m2_t aux0_0 = __riscv_vwmul_vx_i32m2(__riscv_vget_v_i16m2_i16m1(a0, 0), (scale[0]), vl);
+            vint32m2_t aux0_1 = __riscv_vwmul_vx_i32m2(__riscv_vget_v_i16m2_i16m1(a0, 1), (scale[1]), vl);
+            vint32m2_t aux1_0 = __riscv_vwmul_vx_i32m2(__riscv_vget_v_i16m2_i16m1(a1, 0), (scale[2]), vl);
+            vint32m2_t aux1_1 = __riscv_vwmul_vx_i32m2(__riscv_vget_v_i16m2_i16m1(a1, 1), (scale[3]), vl);
+            vint32m2_t aux2_0 = __riscv_vwmul_vx_i32m2(__riscv_vget_v_i16m2_i16m1(a2, 0), (scale[4]), vl);
+            vint32m2_t aux2_1 = __riscv_vwmul_vx_i32m2(__riscv_vget_v_i16m2_i16m1(a2, 1), (scale[5]), vl);
+            vint32m2_t aux3_0 = __riscv_vwmul_vx_i32m2(__riscv_vget_v_i16m2_i16m1(a3, 0), (scale[6]), vl);
+            vint32m2_t aux3_1 = __riscv_vwmul_vx_i32m2(__riscv_vget_v_i16m2_i16m1(a3, 1), (scale[7]), vl);
+
+            vint32m1_t isum0 = __riscv_vredsum_vs_i32m2_i32m1(__riscv_vadd_vv_i32m2(aux0_0, aux0_1, vl), vzero, vl);
+            vint32m1_t isum1 = __riscv_vredsum_vs_i32m2_i32m1(__riscv_vadd_vv_i32m2(aux1_0, aux1_1, vl), isum0, vl);
+            vint32m1_t isum2 = __riscv_vredsum_vs_i32m2_i32m1(__riscv_vadd_vv_i32m2(aux2_0, aux2_1, vl), isum1, vl);
+            vint32m1_t isum3 = __riscv_vredsum_vs_i32m2_i32m1(__riscv_vadd_vv_i32m2(aux3_0, aux3_1, vl), isum2, vl);
+
+            sum_t +=  __riscv_vmv_x_s_i32m1_i32(isum3);
+
+            q3 += 32;    q8 += 128;   scale += 8;
+
+        }
+
+        const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
+
+        sumf += d*sum_t;
+
+    }
+
+    *s = sumf;
+
+#elif defined(__POWER9_VECTOR__)
+    const vector signed char lowMask = vec_splats((signed char)0x3);
+    const vector signed char lowMask1 = vec_splats((int8_t)0xf);
+    const vector signed char lowMask2 = vec_splats((int8_t)0x30);
+    const vector int v0 = vec_splats((int32_t)0);
+    const vector signed char v1 = vec_splats((signed char)0x1);
+    const vector unsigned char v2 = vec_splats((unsigned char)0x2);
+    const vector unsigned char v3 = vec_splats((unsigned char)0x3);
+    const vector unsigned char v4 = vec_splats((unsigned char)0x4);
+    const vector unsigned char v6 = vec_splats((unsigned char)0x6);
+    const vector signed char off = vec_splats((signed char)0x20);
+
+    vector float vsumf0 = vec_splats(0.0f);
+    vector float vsumf1 = vec_splats(0.0f);
+    vector float vsumf2 = vec_splats(0.0f);
+    vector float vsumf3 = vec_splats(0.0f);
+
+    for (int i = 0; i < nb; ++i) {
+        vector float vxd = vec_splats(GGML_FP16_TO_FP32(x[i].d));
+        vector float vyd = vec_splats(y[i].d);
+        vector float vd = vec_mul(vxd, vyd);
+
+        UNUSED(kmask1);
+        UNUSED(kmask2);
+
+        vector signed char u0 = (vector signed char)vec_xl_len(x[i].scales, 8);
+        vector signed char u1 = vec_and(u0, lowMask1);
+        vector signed char u2 = (vector signed char)vec_xl_len(x[i].scales + 8, 4);
+        vector signed char u3 = (vector signed char)vec_mergeh((vector signed int)u2, (vector signed int)vec_sr(u2, v2));
+        vector signed char u30 = vec_sl(vec_and(u3, lowMask), v4);
+        vector signed char u31 = vec_and(u3, lowMask2);
+
+        u1 = vec_or(u1, u30);
+        u2 = vec_or(vec_sr(u0, v4), u31);
+
+        vector signed char vscales = (vector signed char)vec_mergeh((vector signed long long)u1, (vector signed long long)u2);
+        vector signed char qxhs0 = (vector signed char)vec_xl( 0, x[i].hmask);
+        vector signed char qxhs1 = (vector signed char)vec_xl(16, x[i].hmask);
+
+        vscales = vec_sub(vscales, off);
+
+        vector signed int vsumi0 = v0;
+        vector signed int vsumi1 = v0;
+        vector signed int vsumi2 = v0;
+        vector signed int vsumi3 = v0;
+        vector signed int vsumi4 = v0;
+        vector signed int vsumi5 = v0;
+        vector signed int vsumi6 = v0;
+        vector signed int vsumi7 = v0;
+
+        const uint8_t * restrict q3 = x[i].qs;
+        const int8_t  * restrict q8 = y[i].qs;
+
+        for (int j = 0; j < QK_K/128; ++j) {
+            __builtin_prefetch(q3, 0, 1);
+            __builtin_prefetch(q8, 0, 1);
+
+            vector signed char qxs0 = (vector signed char)vec_xl( 0, q3);
+            vector signed char qxs1 = (vector signed char)vec_xl(16, q3);
+            q3 += 32;
+
+            //the low 2 bits
+            vector signed char qxs00 = vec_and(qxs0, lowMask);
+            vector signed char qxs01 = vec_and(vec_sr(qxs0, v2), lowMask);
+            vector signed char qxs02 = vec_and(vec_sr(qxs0, v4), lowMask);
+            vector signed char qxs03 = vec_and(vec_sr(qxs0, v6), lowMask);
+            vector signed char qxs10 = vec_and(qxs1, lowMask);
+            vector signed char qxs11 = vec_and(vec_sr(qxs1, v2), lowMask);
+            vector signed char qxs12 = vec_and(vec_sr(qxs1, v4), lowMask);
+            vector signed char qxs13 = vec_and(vec_sr(qxs1, v6), lowMask);
+
+            //the 3rd bit
+            vector signed char qxh00 = vec_sl(vec_andc(v1, qxhs0), v2);
+            vector signed char qxh01 = vec_sl(vec_andc(v1, vec_sr(qxhs0, (vector unsigned char)v1)), v2);
+            vector signed char qxh02 = vec_sl(vec_andc(v1, vec_sr(qxhs0, v2)), v2);
+            vector signed char qxh03 = vec_sl(vec_andc(v1, vec_sr(qxhs0, v3)), v2);
+            vector signed char qxh10 = vec_sl(vec_andc(v1, qxhs1), v2);
+            vector signed char qxh11 = vec_sl(vec_andc(v1, vec_sr(qxhs1, (vector unsigned char)v1)), v2);
+            vector signed char qxh12 = vec_sl(vec_andc(v1, vec_sr(qxhs1, v2)), v2);
+            vector signed char qxh13 = vec_sl(vec_andc(v1, vec_sr(qxhs1, v3)), v2);
+            qxhs0 = vec_sr(qxhs0, v4);
+            qxhs1 = vec_sr(qxhs1, v4);
+
+            vector signed char q3x00 = vec_sub(qxs00, qxh00);
+            vector signed char q3x01 = vec_sub(qxs01, qxh01);
+            vector signed char q3x02 = vec_sub(qxs02, qxh02);
+            vector signed char q3x03 = vec_sub(qxs03, qxh03);
+            vector signed char q3x10 = vec_sub(qxs10, qxh10);
+            vector signed char q3x11 = vec_sub(qxs11, qxh11);
+            vector signed char q3x12 = vec_sub(qxs12, qxh12);
+            vector signed char q3x13 = vec_sub(qxs13, qxh13);
+
+            vector signed char q8y00 = vec_xl(  0, q8);
+            vector signed char q8y10 = vec_xl( 16, q8);
+            vector signed char q8y01 = vec_xl( 32, q8);
+            vector signed char q8y11 = vec_xl( 48, q8);
+            vector signed char q8y02 = vec_xl( 64, q8);
+            vector signed char q8y12 = vec_xl( 80, q8);
+            vector signed char q8y03 = vec_xl( 96, q8);
+            vector signed char q8y13 = vec_xl(112, q8);
+            q8 += 128;
+
+            vector signed short vscales_h = vec_unpackh(vscales);
+            vector signed short vs0 = vec_splat(vscales_h, 0);
+            vector signed short vs1 = vec_splat(vscales_h, 1);
+            vector signed short vs2 = vec_splat(vscales_h, 2);
+            vector signed short vs3 = vec_splat(vscales_h, 3);
+            vector signed short vs4 = vec_splat(vscales_h, 4);
+            vector signed short vs5 = vec_splat(vscales_h, 5);
+            vector signed short vs6 = vec_splat(vscales_h, 6);
+            vector signed short vs7 = vec_splat(vscales_h, 7);
+            vscales = vec_sld(vscales, vscales, 8);
+
+            vector signed short qv00 = vec_add(vec_mule(q3x00, q8y00), vec_mulo(q3x00, q8y00));
+            vector signed short qv01 = vec_add(vec_mule(q3x01, q8y01), vec_mulo(q3x01, q8y01));
+            vector signed short qv02 = vec_add(vec_mule(q3x02, q8y02), vec_mulo(q3x02, q8y02));
+            vector signed short qv03 = vec_add(vec_mule(q3x03, q8y03), vec_mulo(q3x03, q8y03));
+            vector signed short qv10 = vec_add(vec_mule(q3x10, q8y10), vec_mulo(q3x10, q8y10));
+            vector signed short qv11 = vec_add(vec_mule(q3x11, q8y11), vec_mulo(q3x11, q8y11));
+            vector signed short qv12 = vec_add(vec_mule(q3x12, q8y12), vec_mulo(q3x12, q8y12));
+            vector signed short qv13 = vec_add(vec_mule(q3x13, q8y13), vec_mulo(q3x13, q8y13));
+
+            vsumi0 = vec_msum(qv00, vs0, vsumi0);
+            vsumi1 = vec_msum(qv01, vs2, vsumi1);
+            vsumi2 = vec_msum(qv02, vs4, vsumi2);
+            vsumi3 = vec_msum(qv03, vs6, vsumi3);
+            vsumi4 = vec_msum(qv10, vs1, vsumi4);
+            vsumi5 = vec_msum(qv11, vs3, vsumi5);
+            vsumi6 = vec_msum(qv12, vs5, vsumi6);
+            vsumi7 = vec_msum(qv13, vs7, vsumi7);
+        }
+
+        vsumi0 = vec_add(vsumi0, vsumi4);
+        vsumi1 = vec_add(vsumi1, vsumi5);
+        vsumi2 = vec_add(vsumi2, vsumi6);
+        vsumi3 = vec_add(vsumi3, vsumi7);
+
+        vsumf0 = vec_madd(vec_ctf(vsumi0, 0), vd, vsumf0);
+        vsumf1 = vec_madd(vec_ctf(vsumi1, 0), vd, vsumf1);
+        vsumf2 = vec_madd(vec_ctf(vsumi2, 0), vd, vsumf2);
+        vsumf3 = vec_madd(vec_ctf(vsumi3, 0), vd, vsumf3);
+    }
+
+    vsumf0 = vec_add(vsumf0, vsumf2);
+    vsumf1 = vec_add(vsumf1, vsumf3);
+
+    vsumf0 = vec_add(vsumf0, vsumf1);
+
+    vsumf0 = vec_add(vsumf0, vec_sld(vsumf0, vsumf0, 4));
+    vsumf0 = vec_add(vsumf0, vec_sld(vsumf0, vsumf0, 8));
+
+    *s = vec_extract(vsumf0, 0);
+
+#elif defined __loongarch_asx
+
+    const __m256i m3 = __lasx_xvreplgr2vr_b(3);
+    const __m256i mone = __lasx_xvreplgr2vr_b(1);
+    const __m128i m32 = __lsx_vreplgr2vr_b(32);
+
+    __m256 acc = (__m256)__lasx_xvldi(0);
+
+    uint32_t aux[3];
+
+    for (int i = 0; i < nb; ++i) {
+
+        const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d);
+        const uint8_t * restrict q3 = x[i].qs;
+        const int8_t  * restrict q8 = y[i].qs;
+        // Set up scales
+        memcpy(aux, x[i].scales, 12);
+        __m128i scales128 = lsx_set_w(
+                ((aux[1] >> 4) & kmask2) | (((aux[2] >> 6) & kmask1) << 4),
+                ((aux[0] >> 4) & kmask2) | (((aux[2] >> 4) & kmask1) << 4),
+                (aux[1] & kmask2) | (((aux[2] >> 2) & kmask1) << 4),
+                (aux[0] & kmask2) | (((aux[2] >> 0) & kmask1) << 4));
+        scales128 = __lsx_vsub_b(scales128, m32);
+        const __m256i all_scales = lasx_ext8_16(scales128);
+        const __m128i l_scales = lasx_extracti128(all_scales, 0);
+        const __m128i h_scales = lasx_extracti128(all_scales, 1);
+        const __m256i scales[2] = {lasx_insertf128(l_scales, l_scales), lasx_insertf128(h_scales, h_scales)};
+
+        // high bit
+        const __m256i hbits = __lasx_xvld((const __m256i*)x[i].hmask, 0);
+
+        // integer accumulator
+        __m256i sumi = __lasx_xvldi(0);
+
+        int bit = 0;
+        int is  = 0;
+        __m256i xvbit;
+
+
+        for (int j = 0; j < QK_K/128; ++j) {
+            // load low 2 bits
+            const __m256i q3bits = __lasx_xvld((const __m256i*)q3, 0); q3 += 32;
+
+            xvbit = __lasx_xvreplgr2vr_h(bit);
+            // prepare low and high bits
+            const __m256i q3l_0 = __lasx_xvand_v(q3bits, m3);
+            const __m256i q3h_0 = __lasx_xvslli_h(__lasx_xvsrl_h(__lasx_xvandn_v(hbits, __lasx_xvsll_h(mone, xvbit)), xvbit), 2);
+            ++bit;
+
+            xvbit = __lasx_xvreplgr2vr_h(bit);
+            const __m256i q3l_1 = __lasx_xvand_v(__lasx_xvsrli_h(q3bits, 2), m3);
+            const __m256i q3h_1 = __lasx_xvslli_h(__lasx_xvsrl_h(__lasx_xvandn_v(hbits, __lasx_xvsll_h(mone, xvbit)), xvbit), 2);
+            ++bit;
+
+            xvbit = __lasx_xvreplgr2vr_h(bit);
+            const __m256i q3l_2 = __lasx_xvand_v(__lasx_xvsrli_h(q3bits, 4), m3);
+            const __m256i q3h_2 = __lasx_xvslli_h(__lasx_xvsrl_h(__lasx_xvandn_v(hbits, __lasx_xvsll_h(mone, xvbit)), xvbit), 2);
+            ++bit;
+
+            xvbit = __lasx_xvreplgr2vr_h(bit);
+            const __m256i q3l_3 = __lasx_xvand_v(__lasx_xvsrli_h(q3bits, 6), m3);
+            const __m256i q3h_3 = __lasx_xvslli_h(__lasx_xvsrl_h(__lasx_xvandn_v(hbits, __lasx_xvsll_h(mone, xvbit)), xvbit), 2);
+            ++bit;
+
+            // load Q8 quants
+            const __m256i q8_0 = __lasx_xvld((const __m256i*)q8, 0); q8 += 32;
+            const __m256i q8_1 = __lasx_xvld((const __m256i*)q8, 0); q8 += 32;
+            const __m256i q8_2 = __lasx_xvld((const __m256i*)q8, 0); q8 += 32;
+            const __m256i q8_3 = __lasx_xvld((const __m256i*)q8, 0); q8 += 32;
+
+            // Dot product: we multiply the 2 low bits and 1 high bit part separately, so we can use lasx_maddubs_h,
+            // and then subtract. The high bit part has the 2 already subtracted (and so, it is zero if the high bit was not set,
+            // and 2 if the high bit was set)
+            __m256i q8s_0 = lasx_maddubs_h(q3h_0, q8_0);
+            __m256i q8s_1 = lasx_maddubs_h(q3h_1, q8_1);
+            __m256i q8s_2 = lasx_maddubs_h(q3h_2, q8_2);
+            __m256i q8s_3 = lasx_maddubs_h(q3h_3, q8_3);
+
+            __m256i p16_0 = lasx_maddubs_h(q3l_0, q8_0);
+            __m256i p16_1 = lasx_maddubs_h(q3l_1, q8_1);
+            __m256i p16_2 = lasx_maddubs_h(q3l_2, q8_2);
+            __m256i p16_3 = lasx_maddubs_h(q3l_3, q8_3);
+
+            p16_0 = __lasx_xvsub_h(p16_0, q8s_0);
+            p16_1 = __lasx_xvsub_h(p16_1, q8s_1);
+            p16_2 = __lasx_xvsub_h(p16_2, q8s_2);
+            p16_3 = __lasx_xvsub_h(p16_3, q8s_3);
+
+            // multiply with scales
+            p16_0 = lasx_madd_h(lasx_shuffle_b(scales[j], get_scale_shuffle_q3k(is + 0)), p16_0);
+            p16_1 = lasx_madd_h(lasx_shuffle_b(scales[j], get_scale_shuffle_q3k(is + 1)), p16_1);
+            p16_2 = lasx_madd_h(lasx_shuffle_b(scales[j], get_scale_shuffle_q3k(is + 2)), p16_2);
+            p16_3 = lasx_madd_h(lasx_shuffle_b(scales[j], get_scale_shuffle_q3k(is + 3)), p16_3);
+
+            // accumulate
+            p16_0 = __lasx_xvadd_w(p16_0, p16_1);
+            p16_2 = __lasx_xvadd_w(p16_2, p16_3);
+            sumi  = __lasx_xvadd_w(sumi, __lasx_xvadd_w(p16_0, p16_2));
+        }
+        // multiply with block scale and accumulate
+        acc = __lasx_xvfmadd_s(__lasx_xvreplfr2vr_s(d), __lasx_xvffint_s_w(sumi), acc);//FIXME
+    }
+
+    *s = hsum_float_8(acc);
+
+#else
+    // scalar version
+    // This function is written like this so the compiler can manage to vectorize most of it
+    // Using -Ofast, GCC and clang manage to produce code that is within a factor of 2 or so from the
+    // manually vectorized version above. Every other version I tried would run at least 4 times slower.
+    // The ideal situation would be if we could just write the code once, and the compiler would
+    // automatically produce the best possible set of machine instructions, instead of us having to manually
+    // write vectorized versions for AVX, ARM_NEON, etc.
+
+    int8_t  aux8[QK_K];
+    int16_t aux16[8];
+    float   sums [8];
+    int32_t aux32[8];
+    memset(sums, 0, 8*sizeof(float));
+
+    uint32_t auxs[4];
+    const int8_t * scales = (const int8_t*)auxs;
+
+    float sumf = 0;
+    for (int i = 0; i < nb; ++i) {
+        const uint8_t * restrict q3 = x[i].qs;
+        const uint8_t * restrict hm = x[i].hmask;
+        const  int8_t * restrict q8 = y[i].qs;
+        memset(aux32, 0, 8*sizeof(int32_t));
+        int8_t * restrict a = aux8;
+        uint8_t m = 1;
+        for (int j = 0; j < QK_K; j += 128) {
+            for (int l = 0; l < 32; ++l) a[l] = q3[l] & 3;
+            for (int l = 0; l < 32; ++l) a[l] -= (hm[l] & m ? 0 : 4);
+            a += 32; m <<= 1;
+            for (int l = 0; l < 32; ++l) a[l] = (q3[l] >> 2) & 3;
+            for (int l = 0; l < 32; ++l) a[l] -= (hm[l] & m ? 0 : 4);
+            a += 32; m <<= 1;
+            for (int l = 0; l < 32; ++l) a[l] = (q3[l] >> 4) & 3;
+            for (int l = 0; l < 32; ++l) a[l] -= (hm[l] & m ? 0 : 4);
+            a += 32; m <<= 1;
+            for (int l = 0; l < 32; ++l) a[l] = (q3[l] >> 6) & 3;
+            for (int l = 0; l < 32; ++l) a[l] -= (hm[l] & m ? 0 : 4);
+            a += 32; m <<= 1;
+            q3 += 32;
+        }
+        a = aux8;
+
+        memcpy(auxs, x[i].scales, 12);
+        uint32_t tmp = auxs[2];
+        auxs[2] = ((auxs[0] >> 4) & kmask2) | (((tmp >> 4) & kmask1) << 4);
+        auxs[3] = ((auxs[1] >> 4) & kmask2) | (((tmp >> 6) & kmask1) << 4);
+        auxs[0] = (auxs[0] & kmask2) | (((tmp >> 0) & kmask1) << 4);
+        auxs[1] = (auxs[1] & kmask2) | (((tmp >> 2) & kmask1) << 4);
+        for (int j = 0; j < QK_K/16; ++j) {
+            for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
+            for (int l = 0; l < 8; ++l) aux32[l] += (scales[j] - 32) * aux16[l];
+            q8 += 8; a += 8;
+            for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
+            for (int l = 0; l < 8; ++l) aux32[l] += (scales[j] - 32) * aux16[l];
+            q8 += 8; a += 8;
+        }
+        const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
+        for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l];
+    }
+    for (int l = 0; l < 8; ++l) sumf += sums[l];
+    *s = sumf;
+
+#endif
+
+}
+
+void ggml_vec_dot_q4_K_q8_K(int n, float * restrict s, size_t bs, const void * restrict vx, size_t bx, const void * restrict vy, size_t by, int nrc) {
+    assert(n % QK_K == 0);
+    assert(nrc == 1);
+    UNUSED(nrc);
+    UNUSED(bx);
+    UNUSED(by);
+    UNUSED(bs);
+
+    const block_q4_K * restrict x = vx;
+    const block_q8_K * restrict y = vy;
+
+    const int nb = n / QK_K;
+
+    static const uint32_t kmask1 = 0x3f3f3f3f;
+    static const uint32_t kmask2 = 0x0f0f0f0f;
+    static const uint32_t kmask3 = 0x03030303;
+
+    uint32_t utmp[4];
+
+#ifdef __ARM_NEON
+    const uint8x16_t m4b = vdupq_n_u8(0xf);
+    const int32x4_t mzero = vdupq_n_s32(0);
+
+    ggml_int8x16x2_t q4bytes;
+    ggml_int8x16x2_t q8bytes;
+
+    float sumf = 0;
+
+    for (int i = 0; i < nb; ++i) {
+
+        const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d);
+        const float dmin = y[i].d * GGML_FP16_TO_FP32(x[i].dmin);
+
+        const int16x8_t q8sums = vpaddq_s16(vld1q_s16(y[i].bsums), vld1q_s16(y[i].bsums + 8));
+
+        memcpy(utmp, x[i].scales, 12);
+
+        uint32x2_t mins8 = { 0 };
+        mins8 = vset_lane_u32(utmp[1] & kmask1, mins8, 0);
+        mins8 = vset_lane_u32(((utmp[2] >> 4) & kmask2) | (((utmp[1] >> 6) & kmask3) << 4), mins8, 1);
+
+        utmp[1] = (utmp[2] & kmask2) | (((utmp[0] >> 6) & kmask3) << 4);
+        utmp[0] &= kmask1;
+
+        const int16x8_t mins = vreinterpretq_s16_u16(vmovl_u8(vreinterpret_u8_u32(mins8)));
+        const int32x4_t prod = vaddq_s32(vmull_s16(vget_low_s16 (q8sums), vget_low_s16 (mins)),
+                                         vmull_s16(vget_high_s16(q8sums), vget_high_s16(mins)));
+        sumf -= dmin * vaddvq_s32(prod);
+
+        const uint8_t * scales = (const uint8_t *)utmp;
+
+        const uint8_t * restrict q4 = x[i].qs;
+        const int8_t  * restrict q8 = y[i].qs;
+
+        int32_t sumi1 = 0;
+        int32_t sumi2 = 0;
+
+        for (int j = 0; j < QK_K/64; ++j) {
+            const ggml_uint8x16x2_t q4bits = ggml_vld1q_u8_x2(q4); q4 += 32;
+
+            q8bytes = ggml_vld1q_s8_x2(q8); q8 += 32;
+            q4bytes.val[0] = vreinterpretq_s8_u8(vandq_u8  (q4bits.val[0], m4b));
+            q4bytes.val[1] = vreinterpretq_s8_u8(vandq_u8  (q4bits.val[1], m4b));
+
+            const int32x4_t p1 = ggml_vdotq_s32(ggml_vdotq_s32(mzero, q4bytes.val[0], q8bytes.val[0]), q4bytes.val[1], q8bytes.val[1]);
+            sumi1 += vaddvq_s32(p1) * scales[2*j+0];
+
+            q8bytes = ggml_vld1q_s8_x2(q8); q8 += 32;
+            q4bytes.val[0] = vreinterpretq_s8_u8(vshrq_n_u8(q4bits.val[0], 4));
+            q4bytes.val[1] = vreinterpretq_s8_u8(vshrq_n_u8(q4bits.val[1], 4));
+
+            const int32x4_t p2 = ggml_vdotq_s32(ggml_vdotq_s32(mzero, q4bytes.val[0], q8bytes.val[0]), q4bytes.val[1], q8bytes.val[1]);
+
+            sumi2 += vaddvq_s32(p2) * scales[2*j+1];
+        }
+
+        sumf += d * (sumi1 + sumi2);
+
+    }
+
+    *s = sumf;
+
+#elif defined __AVX2__
+
+    const __m256i m4 = _mm256_set1_epi8(0xF);
+
+    __m256 acc = _mm256_setzero_ps();
+    __m128 acc_m = _mm_setzero_ps();
+
+   for (int i = 0; i < nb; ++i) {
+
+        const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d);
+        const float dmin = -y[i].d * GGML_FP16_TO_FP32(x[i].dmin);
+
+        memcpy(utmp, x[i].scales, 12);
+        utmp[3] = ((utmp[2] >> 4) & kmask2) | (((utmp[1] >> 6) & kmask3) << 4);
+        const uint32_t uaux = utmp[1] & kmask1;
+        utmp[1] = (utmp[2] & kmask2) | (((utmp[0] >> 6) & kmask3) << 4);
+        utmp[2] = uaux;
+        utmp[0] &= kmask1;
+
+        const uint8_t * restrict q4 = x[i].qs;
+        const int8_t  * restrict q8 = y[i].qs;
+
+        const __m256i mins_and_scales = _mm256_cvtepu8_epi16(_mm_set_epi32(utmp[3], utmp[2], utmp[1], utmp[0]));
+
+        const __m256i q8sums = _mm256_loadu_si256((const __m256i*)y[i].bsums);
+        const __m128i q8s = _mm_hadd_epi16(_mm256_extracti128_si256(q8sums, 0), _mm256_extracti128_si256(q8sums, 1));
+        const __m128i prod = _mm_madd_epi16(_mm256_extracti128_si256(mins_and_scales, 1), q8s);
+        acc_m = _mm_fmadd_ps(_mm_set1_ps(dmin), _mm_cvtepi32_ps(prod), acc_m);
+
+        const __m128i sc128  = _mm256_extracti128_si256(mins_and_scales, 0);
+        const __m256i scales = MM256_SET_M128I(sc128, sc128);
+
+        __m256i sumi = _mm256_setzero_si256();
+
+        for (int j = 0; j < QK_K/64; ++j) {
+
+            const __m256i scale_l = _mm256_shuffle_epi8(scales, get_scale_shuffle_k4(2*j+0));
+            const __m256i scale_h = _mm256_shuffle_epi8(scales, get_scale_shuffle_k4(2*j+1));
+
+            const __m256i q4bits = _mm256_loadu_si256((const __m256i*)q4); q4 += 32;
+            const __m256i q4l = _mm256_and_si256(q4bits, m4);
+            const __m256i q4h = _mm256_and_si256(_mm256_srli_epi16(q4bits, 4), m4);
+
+            const __m256i q8l = _mm256_loadu_si256((const __m256i*)q8); q8 += 32;
+            __m256i p16l = _mm256_maddubs_epi16(q4l, q8l);
+            p16l = _mm256_madd_epi16(scale_l, p16l);
+
+            const __m256i q8h = _mm256_loadu_si256((const __m256i*)q8); q8 += 32;
+            __m256i p16h = _mm256_maddubs_epi16(q4h, q8h);
+            p16h = _mm256_madd_epi16(scale_h, p16h);
+            const __m256i sumj = _mm256_add_epi32(p16l, p16h);
+
+            sumi = _mm256_add_epi32(sumi, sumj);
+        }
+
+        __m256 vd = _mm256_set1_ps(d);
+        acc = _mm256_fmadd_ps(vd, _mm256_cvtepi32_ps(sumi), acc);
+
+    }
+
+    acc_m = _mm_add_ps(acc_m, _mm_movehl_ps(acc_m, acc_m));
+    acc_m = _mm_add_ss(acc_m, _mm_movehdup_ps(acc_m));
+
+    *s = hsum_float_8(acc) + _mm_cvtss_f32(acc_m);
+
+#elif defined __AVX__
+
+    const __m128i m4 = _mm_set1_epi8(0xF);
+    const __m128i m2 = _mm_set1_epi8(0x2);
+
+    __m256 acc = _mm256_setzero_ps();
+    __m128 acc_m = _mm_setzero_ps();
+
+   for (int i = 0; i < nb; ++i) {
+
+        const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d);
+        const float dmin = -y[i].d * GGML_FP16_TO_FP32(x[i].dmin);
+
+        const uint8_t * restrict q4 = x[i].qs;
+        const int8_t  * restrict q8 = y[i].qs;
+
+        memcpy(utmp, x[i].scales, 12);
+        utmp[3] = ((utmp[2] >> 4) & kmask2) | (((utmp[1] >> 6) & kmask3) << 4);
+        const uint32_t uaux = utmp[1] & kmask1;
+        utmp[1] = (utmp[2] & kmask2) | (((utmp[0] >> 6) & kmask3) << 4);
+        utmp[2] = uaux;
+        utmp[0] &= kmask1;
+
+        const __m128i utmps = _mm_set_epi32(utmp[3], utmp[2], utmp[1], utmp[0]);
+        const __m128i scales = _mm_cvtepu8_epi16(utmps);
+        const __m128i mins = _mm_cvtepu8_epi16(_mm_unpackhi_epi64(utmps, utmps));
+
+        const __m128i q8sums_0 = _mm_loadu_si128((const __m128i*)&y[i].bsums[0]);
+        const __m128i q8sums_1 = _mm_loadu_si128((const __m128i*)&y[i].bsums[8]);
+        const __m128i q8s = _mm_hadd_epi16(q8sums_0, q8sums_1);
+        const __m128i prod = _mm_madd_epi16(mins, q8s);
+        acc_m = _mm_add_ps(_mm_mul_ps(_mm_set1_ps(dmin), _mm_cvtepi32_ps(prod)), acc_m);
+
+        __m128i sumi_0 = _mm_setzero_si128();
+        __m128i sumi_1 = _mm_setzero_si128();
+
+        __m128i shuffle = _mm_set1_epi16(0x0100);
+        for (int j = 0; j < QK_K/64; ++j) {
+
+            const __m128i scale_l = _mm_shuffle_epi8(scales, shuffle);
+            shuffle = _mm_add_epi16(shuffle, m2);
+            const __m128i scale_h = _mm_shuffle_epi8(scales, shuffle);
+            shuffle = _mm_add_epi16(shuffle, m2);
+
+            __m128i q4bits = _mm_loadu_si128((const __m128i*)q4); q4 += 16;
+            const __m128i q4l_0 = _mm_and_si128(q4bits, m4);
+            const __m128i q4h_0 = _mm_and_si128(_mm_srli_epi16(q4bits, 4), m4);
+            q4bits = _mm_loadu_si128((const __m128i*)q4); q4 += 16;
+            const __m128i q4l_1 = _mm_and_si128(q4bits, m4);
+            const __m128i q4h_1 = _mm_and_si128(_mm_srli_epi16(q4bits, 4), m4);
+
+            const __m128i q8l_0 = _mm_loadu_si128((const __m128i*)q8); q8 += 16;
+            __m128i p16l = _mm_maddubs_epi16(q4l_0, q8l_0);
+            p16l = _mm_madd_epi16(scale_l, p16l);
+            sumi_0 = _mm_add_epi32(sumi_0, p16l);
+            const __m128i q8l_1 = _mm_loadu_si128((const __m128i*)q8); q8 += 16;
+            p16l = _mm_maddubs_epi16(q4l_1, q8l_1);
+            p16l = _mm_madd_epi16(scale_l, p16l);
+            sumi_1 = _mm_add_epi32(sumi_1, p16l);
+
+            const __m128i q8h_0 = _mm_loadu_si128((const __m128i*)q8); q8 += 16;
+            __m128i p16h = _mm_maddubs_epi16(q4h_0, q8h_0);
+            p16h = _mm_madd_epi16(scale_h, p16h);
+            sumi_0 = _mm_add_epi32(sumi_0, p16h);
+            const __m128i q8h_1 = _mm_loadu_si128((const __m128i*)q8); q8 += 16;
+            p16h = _mm_maddubs_epi16(q4h_1, q8h_1);
+            p16h = _mm_madd_epi16(scale_h, p16h);
+            sumi_1 = _mm_add_epi32(sumi_1, p16h);
+
+        }
+
+        __m256 vd = _mm256_set1_ps(d);
+        __m256i sumi = MM256_SET_M128I(sumi_1, sumi_0);
+        acc = _mm256_add_ps(_mm256_mul_ps(vd, _mm256_cvtepi32_ps(sumi)), acc);
+
+    }
+
+    acc_m = _mm_add_ps(acc_m, _mm_movehl_ps(acc_m, acc_m));
+    acc_m = _mm_add_ss(acc_m, _mm_movehdup_ps(acc_m));
+
+    *s = hsum_float_8(acc) + _mm_cvtss_f32(acc_m);
+
+#elif defined __riscv_v_intrinsic
+
+    const uint8_t * scales = (const uint8_t*)&utmp[0];
+    const uint8_t * mins   = (const uint8_t*)&utmp[2];
+
+    float sumf = 0;
+
+    for (int i = 0; i < nb; ++i) {
+
+        size_t vl = 8;
+
+        const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d);
+        const float dmin = y[i].d * GGML_FP16_TO_FP32(x[i].dmin);
+
+        vint16mf2_t q8sums_0 = __riscv_vlse16_v_i16mf2(y[i].bsums, 4, vl);
+        vint16mf2_t q8sums_1 = __riscv_vlse16_v_i16mf2(y[i].bsums+1, 4, vl);
+        vint16mf2_t q8sums   = __riscv_vadd_vv_i16mf2(q8sums_0, q8sums_1, vl);
+
+        memcpy(utmp, x[i].scales, 12);
+        utmp[3] = ((utmp[2] >> 4) & kmask2) | (((utmp[1] >> 6) & kmask3) << 4);
+        const uint32_t uaux = utmp[1] & kmask1;
+        utmp[1] = (utmp[2] & kmask2) | (((utmp[0] >> 6) & kmask3) << 4);
+        utmp[2] = uaux;
+        utmp[0] &= kmask1;
+
+        vuint8mf4_t mins8  = __riscv_vle8_v_u8mf4(mins, vl);
+        vint16mf2_t v_mins = __riscv_vreinterpret_v_u16mf2_i16mf2(__riscv_vzext_vf2_u16mf2(mins8, vl));
+        vint32m1_t  prod   = __riscv_vwmul_vv_i32m1(q8sums, v_mins, vl);
+
+        vint32m1_t sumi = __riscv_vredsum_vs_i32m1_i32m1(prod, __riscv_vmv_v_x_i32m1(0, 1), vl);
+        sumf -= dmin * __riscv_vmv_x_s_i32m1_i32(sumi);
+
+        const uint8_t * restrict q4 = x[i].qs;
+        const int8_t  * restrict q8 = y[i].qs;
+
+        vl = 32;
+
+        int32_t sum_1 = 0;
+        int32_t sum_2 = 0;
+
+        vint16m1_t vzero = __riscv_vmv_v_x_i16m1(0, 1);
+
+        for (int j = 0; j < QK_K/64; ++j) {
+            // load Q4
+            vuint8m1_t q4_x = __riscv_vle8_v_u8m1(q4, vl);
+
+            // load Q8 and multiply it with lower Q4 nibble
+            vint8m1_t  q8_0 = __riscv_vle8_v_i8m1(q8, vl);
+            vint8m1_t  q4_0 = __riscv_vreinterpret_v_u8m1_i8m1(__riscv_vand_vx_u8m1(q4_x, 0x0F, vl));
+            vint16m2_t qv_0 = __riscv_vwmul_vv_i16m2(q4_0, q8_0, vl);
+            vint16m1_t vs_0 = __riscv_vredsum_vs_i16m2_i16m1(qv_0, vzero, vl);
+
+            sum_1 += __riscv_vmv_x_s_i16m1_i16(vs_0) * scales[2*j+0];
+
+            // load Q8 and multiply it with upper Q4 nibble
+            vint8m1_t  q8_1 = __riscv_vle8_v_i8m1(q8+32, vl);
+            vint8m1_t  q4_1 = __riscv_vreinterpret_v_u8m1_i8m1(__riscv_vsrl_vx_u8m1(q4_x, 0x04, vl));
+            vint16m2_t qv_1 = __riscv_vwmul_vv_i16m2(q4_1, q8_1, vl);
+            vint16m1_t vs_1 = __riscv_vredsum_vs_i16m2_i16m1(qv_1, vzero, vl);
+
+            sum_2 += __riscv_vmv_x_s_i16m1_i16(vs_1) * scales[2*j+1];
+
+            q4 += 32;    q8 += 64;
+
+        }
+
+        sumf += d*(sum_1 + sum_2);
+
+    }
+
+    *s = sumf;
+
+#elif defined(__POWER9_VECTOR__)
+    const vector signed char lowMask = vec_splats((signed char)0xF);
+    const vector signed char lowMask1 = vec_splats((int8_t)0x3f);
+    const vector signed char lowMask2 = vec_splats((int8_t)0x30);
+    const vector int v0 = vec_splats((int32_t)0);
+    const vector unsigned char v2 = vec_splats((uint8_t)2);
+    const vector unsigned char v4 = vec_splats((unsigned char)0x4);
+
+    vector float vsumf0 = vec_splats(0.0f);
+    vector float vsumf1 = vec_splats(0.0f);
+    vector float vsumf2 = vec_splats(0.0f);
+    vector float vsumf3 = vec_splats(0.0f);
+
+    for (int i = 0; i < nb; ++i) {
+        vector float vxd = vec_splats(GGML_FP16_TO_FP32(x[i].d));
+        vector float vyd = vec_splats(y[i].d);
+        vector float vd = vec_mul(vxd, vyd);
+
+        vector float vxmin = vec_splats(GGML_FP16_TO_FP32(x[i].dmin));
+        vector float vdmin = vec_mul(vxmin, vyd);
+
+        vector signed short q8ysums0 = vec_xl( 0, y[i].bsums);
+        vector signed short q8ysums1 = vec_xl(16, y[i].bsums);
+
+        UNUSED(kmask1);
+        UNUSED(kmask2);
+        UNUSED(kmask3);
+        UNUSED(utmp);
+
+        vector signed char u0 = (vector signed char)vec_xl_len(x[i].scales, 8);
+        vector signed char u1 = vec_and(vec_sr(u0, v2), lowMask2);
+        vector signed char u2 = (vector signed char)vec_xl_len(x[i].scales + 8, 4);
+        vector signed char u3 = vec_sr(u2, v4);
+
+        vector signed char u30 = u1;
+        vector signed char u31 = (vector signed char)vec_mergeh((vector signed int)vec_and(u2, lowMask), (vector signed int)u3);
+
+        u1 = vec_and(u0, lowMask1);
+        u2 = vec_or(u30, u31);
+
+        vector signed char utmps = (vector signed char)vec_mergeh((vector signed int)u1, (vector signed int)u2);
+
+        vector signed short vscales = vec_unpackh(utmps);
+        vector signed short q4xmins = vec_unpackl(utmps);
+        vector signed short q4xmins0 = vec_mergeh(q4xmins, q4xmins);
+        vector signed short q4xmins1 = vec_mergel(q4xmins, q4xmins);
+
+        vector signed int prod0 = vec_mule(q4xmins0, q8ysums0);
+        vector signed int prod1 = vec_mule(q4xmins1, q8ysums1);
+        vector signed int prod2 = vec_mulo(q4xmins0, q8ysums0);
+        vector signed int prod3 = vec_mulo(q4xmins1, q8ysums1);
+
+        vsumf0 = vec_nmsub(vec_ctf(prod0, 0), vdmin, vsumf0);
+        vsumf1 = vec_nmsub(vec_ctf(prod1, 0), vdmin, vsumf1);
+        vsumf2 = vec_nmsub(vec_ctf(prod2, 0), vdmin, vsumf2);
+        vsumf3 = vec_nmsub(vec_ctf(prod3, 0), vdmin, vsumf3);
+
+        vector signed int vsumi0 = v0;
+        vector signed int vsumi1 = v0;
+        vector signed int vsumi2 = v0;
+        vector signed int vsumi3 = v0;
+
+        const uint8_t * restrict q4 = x[i].qs;
+        const int8_t  * restrict q8 = y[i].qs;
+
+        for (int j = 0; j < QK_K/64; j+=2) {
+            __builtin_prefetch(q4, 0, 1);
+            __builtin_prefetch(q8, 0, 1);
+
+            vector signed char qxs0 = (vector signed char)vec_xl( 0, q4);
+            vector signed char qxs1 = (vector signed char)vec_xl(16, q4);
+            vector signed char qxs2 = (vector signed char)vec_xl(32, q4);
+            vector signed char qxs3 = (vector signed char)vec_xl(48, q4);
+            q4 += 64;
+
+            vector unsigned char q4x00 = (vector unsigned char)vec_and(qxs0, lowMask);
+            vector unsigned char q4x01 = (vector unsigned char)vec_sr(qxs0, v4);
+            vector unsigned char q4x10 = (vector unsigned char)vec_and(qxs1, lowMask);
+            vector unsigned char q4x11 = (vector unsigned char)vec_sr(qxs1, v4);
+            vector unsigned char q4x20 = (vector unsigned char)vec_and(qxs2, lowMask);
+            vector unsigned char q4x21 = (vector unsigned char)vec_sr(qxs2, v4);
+            vector unsigned char q4x30 = (vector unsigned char)vec_and(qxs3, lowMask);
+            vector unsigned char q4x31 = (vector unsigned char)vec_sr(qxs3, v4);
+
+            vector signed char q8y00 = vec_xl(  0, q8);
+            vector signed char q8y10 = vec_xl( 16, q8);
+            vector signed char q8y01 = vec_xl( 32, q8);
+            vector signed char q8y11 = vec_xl( 48, q8);
+            vector signed char q8y20 = vec_xl( 64, q8);
+            vector signed char q8y30 = vec_xl( 80, q8);
+            vector signed char q8y21 = vec_xl( 96, q8);
+            vector signed char q8y31 = vec_xl(112, q8);
+            q8 += 128;
+
+            vector signed int qv00 = vec_msum(q8y00, q4x00, v0);
+            vector signed int qv01 = vec_msum(q8y01, q4x01, v0);
+            vector signed int qv10 = vec_msum(q8y10, q4x10, v0);
+            vector signed int qv11 = vec_msum(q8y11, q4x11, v0);
+            vector signed int qv20 = vec_msum(q8y20, q4x20, v0);
+            vector signed int qv21 = vec_msum(q8y21, q4x21, v0);
+            vector signed int qv30 = vec_msum(q8y30, q4x30, v0);
+            vector signed int qv31 = vec_msum(q8y31, q4x31, v0);
+
+            vector signed int vscales_h = vec_unpackh(vscales);
+            vector signed int vs0 = vec_splat(vscales_h, 0);
+            vector signed int vs1 = vec_splat(vscales_h, 1);
+            vector signed int vs2 = vec_splat(vscales_h, 2);
+            vector signed int vs3 = vec_splat(vscales_h, 3);
+            vscales = vec_sld(vscales, vscales, 8);
+
+            vsumi0 = vec_add(vec_mul(qv00, vs0), vsumi0);
+            vsumi1 = vec_add(vec_mul(qv01, vs1), vsumi1);
+            vsumi2 = vec_add(vec_mul(qv20, vs2), vsumi2);
+            vsumi3 = vec_add(vec_mul(qv21, vs3), vsumi3);
+
+            vsumi0 = vec_add(vec_mul(qv10, vs0), vsumi0);
+            vsumi1 = vec_add(vec_mul(qv11, vs1), vsumi1);
+            vsumi2 = vec_add(vec_mul(qv30, vs2), vsumi2);
+            vsumi3 = vec_add(vec_mul(qv31, vs3), vsumi3);
+        }
+
+        vsumf0 = vec_madd(vec_ctf(vsumi0, 0), vd, vsumf0);
+        vsumf1 = vec_madd(vec_ctf(vsumi1, 0), vd, vsumf1);
+        vsumf2 = vec_madd(vec_ctf(vsumi2, 0), vd, vsumf2);
+        vsumf3 = vec_madd(vec_ctf(vsumi3, 0), vd, vsumf3);
+    }
+
+    vsumf0 = vec_add(vsumf0, vsumf2);
+    vsumf1 = vec_add(vsumf1, vsumf3);
+
+    vsumf0 = vec_add(vsumf0, vsumf1);
+
+    vsumf0 = vec_add(vsumf0, vec_sld(vsumf0, vsumf0, 4));
+    vsumf0 = vec_add(vsumf0, vec_sld(vsumf0, vsumf0, 8));
+
+    *s = vec_extract(vsumf0, 0);
+
+#elif defined __loongarch_asx
+    GGML_UNUSED(kmask1);
+    GGML_UNUSED(kmask2);
+    GGML_UNUSED(kmask3);
+
+    const __m256i m4 = __lasx_xvreplgr2vr_b(0xF);
+
+    __m256 acc = (__m256)__lasx_xvldi(0);
+    __m128 acc_m = (__m128)__lsx_vldi(0);
+
+   for (int i = 0; i < nb; ++i) {
+
+        const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d);
+        const float dmin = -y[i].d * GGML_FP16_TO_FP32(x[i].dmin);
+
+        memcpy(utmp, x[i].scales, 12);
+        utmp[3] = ((utmp[2] >> 4) & kmask2) | (((utmp[1] >> 6) & kmask3) << 4);
+        const uint32_t uaux = utmp[1] & kmask1;
+        utmp[1] = (utmp[2] & kmask2) | (((utmp[0] >> 6) & kmask3) << 4);
+        utmp[2] = uaux;
+        utmp[0] &= kmask1;
+
+        const uint8_t * restrict q4 = x[i].qs;
+        const int8_t  * restrict q8 = y[i].qs;
+
+        const __m256i mins_and_scales = lasx_extu8_16(lsx_set_w(utmp[3], utmp[2], utmp[1], utmp[0]));
+
+        const __m256i q8sums = __lasx_xvld((const __m256i*)y[i].bsums, 0);
+        const __m128i q8s = lsx_hadd_h(lasx_extracti128(q8sums, 0), lasx_extracti128(q8sums, 1));
+        const __m128i prod = lsx_madd_h(lasx_extracti128(mins_and_scales, 1), q8s);
+        acc_m = __lsx_vfmadd_s(__lsx_vreplfr2vr_s(dmin), __lsx_vffint_s_w(prod), acc_m);
+
+        const __m128i sc128  = lasx_extracti128(mins_and_scales, 0);
+        const __m256i scales = lasx_insertf128(sc128, sc128);
+
+        __m256i sumi = __lasx_xvldi(0);
+
+        for (int j = 0; j < QK_K/64; ++j) {
+
+            const __m256i scale_l = lasx_shuffle_b(scales, get_scale_shuffle_k4(2*j+0));
+            const __m256i scale_h = lasx_shuffle_b(scales, get_scale_shuffle_k4(2*j+1));
+
+            const __m256i q4bits = __lasx_xvld((const __m256i*)q4, 0); q4 += 32;
+            const __m256i q4l = __lasx_xvand_v(q4bits, m4);
+            const __m256i q4h = __lasx_xvand_v(__lasx_xvsrli_h(q4bits, 4), m4);
+
+            const __m256i q8l = __lasx_xvld((const __m256i*)q8, 0); q8 += 32;
+            __m256i p16l = lasx_maddubs_h(q4l, q8l);
+            p16l = lasx_madd_h(scale_l, p16l);
+
+            const __m256i q8h = __lasx_xvld((const __m256i*)q8, 0); q8 += 32;
+            __m256i p16h = lasx_maddubs_h(q4h, q8h);
+            p16h = lasx_madd_h(scale_h, p16h);
+            const __m256i sumj = __lasx_xvadd_w(p16l, p16h);
+
+            sumi = __lasx_xvadd_w(sumi, sumj);
+        }
+
+        __m256 vd = __lasx_xvreplfr2vr_s(d);
+        acc = __lasx_xvfmadd_s(vd, __lasx_xvffint_s_w(sumi), acc);
+
+    }
+
+    acc_m = __lsx_vfadd_s(acc_m, (__m128)__lsx_vpermi_w((__m128i)acc_m, (__m128i)acc_m, 0xee));
+    __m128i tmp1 = __lsx_vinsgr2vr_w(__lsx_vldi(0), __lsx_vpickve2gr_w((__m128i)acc_m, 1), 0);
+    acc_m = __lsx_vfadd_s(acc_m, (__m128)tmp1);
+
+
+    ft_union fi;
+    fi.i = __lsx_vpickve2gr_w(acc_m, 0);
+    *s = hsum_float_8(acc) + fi.f ;
+#else
+
+    const uint8_t * scales = (const uint8_t*)&utmp[0];
+    const uint8_t * mins   = (const uint8_t*)&utmp[2];
+
+    int8_t  aux8[QK_K];
+    int16_t aux16[8];
+    float   sums [8];
+    int32_t aux32[8];
+    memset(sums, 0, 8*sizeof(float));
+
+    float sumf = 0;
+    for (int i = 0; i < nb; ++i) {
+        const uint8_t * restrict q4 = x[i].qs;
+        const  int8_t * restrict q8 = y[i].qs;
+        memset(aux32, 0, 8*sizeof(int32_t));
+        int8_t * restrict a = aux8;
+        for (int j = 0; j < QK_K/64; ++j) {
+            for (int l = 0; l < 32; ++l) a[l] = (int8_t)(q4[l] & 0xF);
+            a += 32;
+            for (int l = 0; l < 32; ++l) a[l] = (int8_t)(q4[l]  >> 4);
+            a += 32; q4 += 32;
+        }
+        memcpy(utmp, x[i].scales, 12);
+        utmp[3] = ((utmp[2] >> 4) & kmask2) | (((utmp[1] >> 6) & kmask3) << 4);
+        const uint32_t uaux = utmp[1] & kmask1;
+        utmp[1] = (utmp[2] & kmask2) | (((utmp[0] >> 6) & kmask3) << 4);
+        utmp[2] = uaux;
+        utmp[0] &= kmask1;
+
+        int sumi = 0;
+        for (int j = 0; j < QK_K/16; ++j) sumi += y[i].bsums[j] * mins[j/2];
+        a = aux8;
+        int is = 0;
+        for (int j = 0; j < QK_K/32; ++j) {
+            int32_t scale = scales[is++];
+            for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
+            for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
+            q8 += 8; a += 8;
+            for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
+            for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
+            q8 += 8; a += 8;
+            for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
+            for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
+            q8 += 8; a += 8;
+            for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
+            for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
+            q8 += 8; a += 8;
+        }
+        const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
+        for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l];
+        const float dmin = GGML_FP16_TO_FP32(x[i].dmin) * y[i].d;
+        sumf -= dmin * sumi;
+    }
+    for (int l = 0; l < 8; ++l) sumf += sums[l];
+    *s = sumf;
+#endif
+}
+
+void ggml_vec_dot_q5_K_q8_K(int n, float * restrict s, size_t bs, const void * restrict vx, size_t bx, const void * restrict vy,  size_t by, int nrc) {
+    assert(n % QK_K == 0);
+    assert(nrc == 1);
+    UNUSED(nrc);
+    UNUSED(bx);
+    UNUSED(by);
+    UNUSED(bs);
+
+    const block_q5_K * restrict x = vx;
+    const block_q8_K * restrict y = vy;
+
+    const int nb = n / QK_K;
+
+    static const uint32_t kmask1 = 0x3f3f3f3f;
+    static const uint32_t kmask2 = 0x0f0f0f0f;
+    static const uint32_t kmask3 = 0x03030303;
+
+    uint32_t utmp[4];
+
+#ifdef __ARM_NEON
+    const uint8x16_t m4b = vdupq_n_u8(0xf);
+    const uint8x16_t mone = vdupq_n_u8(1);
+    const uint8x16_t mtwo = vdupq_n_u8(2);
+    const int32x4_t mzero = vdupq_n_s32(0);
+
+    ggml_int8x16x4_t q5bytes;
+
+    float sumf = 0;
+
+    for (int i = 0; i < nb; ++i) {
+
+        const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d);
+        const float dmin = y[i].d * GGML_FP16_TO_FP32(x[i].dmin);
+
+        const int16x8_t q8sums = vpaddq_s16(vld1q_s16(y[i].bsums), vld1q_s16(y[i].bsums + 8));
+
+        memcpy(utmp, x[i].scales, 12);
+        utmp[3] = ((utmp[2] >> 4) & kmask2) | (((utmp[1] >> 6) & kmask3) << 4);
+        const uint32_t uaux = utmp[1] & kmask1;
+        utmp[1] = (utmp[2] & kmask2) | (((utmp[0] >> 6) & kmask3) << 4);
+        utmp[2] = uaux;
+        utmp[0] &= kmask1;
+
+        const uint8x8_t mins8 = vld1_u8((const uint8_t*)utmp + 8);
+        const int16x8_t mins = vreinterpretq_s16_u16(vmovl_u8(mins8));
+        const int32x4_t prod = vaddq_s32(vmull_s16(vget_low_s16 (q8sums), vget_low_s16 (mins)),
+                                         vmull_s16(vget_high_s16(q8sums), vget_high_s16(mins)));
+        int32_t sumi_mins = vaddvq_s32(prod);
+
+        const uint8_t * scales = (const uint8_t *)utmp;
+
+        const uint8_t * restrict q5 = x[i].qs;
+        const uint8_t * restrict qh = x[i].qh;
+        const int8_t  * restrict q8 = y[i].qs;
+
+        ggml_uint8x16x2_t qhbits = ggml_vld1q_u8_x2(qh);
+
+        ggml_uint8x16x4_t q5h;
+
+        int32_t sumi = 0;
+
+        for (int j = 0; j < QK_K/64; ++j) {
+
+            const ggml_uint8x16x2_t q5bits = ggml_vld1q_u8_x2(q5); q5 += 32;
+            const ggml_int8x16x4_t q8bytes = ggml_vld1q_s8_x4(q8); q8 += 64;
+
+            q5h.val[0] = vshlq_n_u8(vandq_u8(mone, qhbits.val[0]), 4);
+            q5h.val[1] = vshlq_n_u8(vandq_u8(mone, qhbits.val[1]), 4);
+            q5h.val[2] = vshlq_n_u8(vandq_u8(mtwo, qhbits.val[0]), 3);
+            q5h.val[3] = vshlq_n_u8(vandq_u8(mtwo, qhbits.val[1]), 3);
+            qhbits.val[0] = vshrq_n_u8(qhbits.val[0], 2);
+            qhbits.val[1] = vshrq_n_u8(qhbits.val[1], 2);
+
+            q5bytes.val[0] = vreinterpretq_s8_u8(vorrq_u8(vandq_u8(q5bits.val[0], m4b), q5h.val[0]));
+            q5bytes.val[1] = vreinterpretq_s8_u8(vorrq_u8(vandq_u8(q5bits.val[1], m4b), q5h.val[1]));
+            q5bytes.val[2] = vreinterpretq_s8_u8(vorrq_u8(vshrq_n_u8(q5bits.val[0], 4), q5h.val[2]));
+            q5bytes.val[3] = vreinterpretq_s8_u8(vorrq_u8(vshrq_n_u8(q5bits.val[1], 4), q5h.val[3]));
+
+            sumi += vaddvq_s32(ggml_vdotq_s32(ggml_vdotq_s32(mzero, q5bytes.val[0], q8bytes.val[0]), q5bytes.val[1], q8bytes.val[1])) * *scales++;
+            sumi += vaddvq_s32(ggml_vdotq_s32(ggml_vdotq_s32(mzero, q5bytes.val[2], q8bytes.val[2]), q5bytes.val[3], q8bytes.val[3])) * *scales++;
+        }
+
+        sumf += d * sumi - dmin * sumi_mins;
+    }
+
+    *s = sumf;
+
+#elif defined __AVX2__
+
+    const __m256i m4 = _mm256_set1_epi8(0xF);
+    const __m128i mzero = _mm_setzero_si128();
+    const __m256i mone  = _mm256_set1_epi8(1);
+
+    __m256 acc = _mm256_setzero_ps();
+
+    float summs = 0.f;
+
+    for (int i = 0; i < nb; ++i) {
+        const uint8_t * restrict q5 = x[i].qs;
+        const int8_t  * restrict q8 = y[i].qs;
+
+        const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d);
+        const float dmin = -y[i].d * GGML_FP16_TO_FP32(x[i].dmin);
+
+        memcpy(utmp, x[i].scales, 12);
+        utmp[3] = ((utmp[2] >> 4) & kmask2) | (((utmp[1] >> 6) & kmask3) << 4);
+        const uint32_t uaux = utmp[1] & kmask1;
+        utmp[1] = (utmp[2] & kmask2) | (((utmp[0] >> 6) & kmask3) << 4);
+        utmp[2] = uaux;
+        utmp[0] &= kmask1;
+
+        const __m256i mins_and_scales = _mm256_cvtepu8_epi16(_mm_set_epi32(utmp[3], utmp[2], utmp[1], utmp[0]));
+
+        const __m256i q8sums = _mm256_loadu_si256((const __m256i*)y[i].bsums);
+        const __m128i q8s = _mm_hadd_epi16(_mm256_extracti128_si256(q8sums, 0), _mm256_extracti128_si256(q8sums, 1));
+        const __m128i prod = _mm_madd_epi16(_mm256_extracti128_si256(mins_and_scales, 1), q8s);
+        const __m128i hsum = _mm_hadd_epi32(_mm_hadd_epi32(prod, mzero), mzero);
+        summs += dmin * _mm_extract_epi32(hsum, 0);
+
+        const __m128i sc128  = _mm256_extracti128_si256(mins_and_scales, 0);
+        const __m256i scales = MM256_SET_M128I(sc128, sc128);
+
+        const __m256i hbits = _mm256_loadu_si256((const __m256i*)x[i].qh);
+        __m256i hmask = mone;
+
+        __m256i sumi = _mm256_setzero_si256();
+
+        int bit = 0;
+
+        for (int j = 0; j < QK_K/64; ++j) {
+
+            const __m256i scale_0 = _mm256_shuffle_epi8(scales, get_scale_shuffle_k4(2*j+0));
+            const __m256i scale_1 = _mm256_shuffle_epi8(scales, get_scale_shuffle_k4(2*j+1));
+
+            const __m256i q5bits = _mm256_loadu_si256((const __m256i*)q5); q5 += 32;
+
+            const __m256i q5l_0 = _mm256_and_si256(q5bits, m4);
+            const __m256i q5h_0 = _mm256_slli_epi16(_mm256_srli_epi16(_mm256_and_si256(hbits, hmask), bit++), 4);
+            const __m256i q5_0  = _mm256_add_epi8(q5l_0, q5h_0);
+            hmask = _mm256_slli_epi16(hmask, 1);
+
+            const __m256i q5l_1 = _mm256_and_si256(_mm256_srli_epi16(q5bits, 4), m4);
+            const __m256i q5h_1 = _mm256_slli_epi16(_mm256_srli_epi16(_mm256_and_si256(hbits, hmask), bit++), 4);
+            const __m256i q5_1  = _mm256_add_epi8(q5l_1, q5h_1);
+            hmask = _mm256_slli_epi16(hmask, 1);
+
+            const __m256i q8_0 = _mm256_loadu_si256((const __m256i*)q8); q8 += 32;
+            const __m256i q8_1 = _mm256_loadu_si256((const __m256i*)q8); q8 += 32;
+
+            __m256i p16_0 = _mm256_maddubs_epi16(q5_0, q8_0);
+            __m256i p16_1 = _mm256_maddubs_epi16(q5_1, q8_1);
+
+            p16_0 = _mm256_madd_epi16(scale_0, p16_0);
+            p16_1 = _mm256_madd_epi16(scale_1, p16_1);
+
+            sumi = _mm256_add_epi32(sumi, _mm256_add_epi32(p16_0, p16_1));
+
+        }
+
+        __m256 vd = _mm256_set1_ps(d);
+        acc = _mm256_fmadd_ps(vd, _mm256_cvtepi32_ps(sumi), acc);
+
+    }
+
+    *s = hsum_float_8(acc) + summs;
+
+#elif defined __AVX__
+
+    const __m128i m4 = _mm_set1_epi8(0xF);
+    const __m128i mzero = _mm_setzero_si128();
+    const __m128i mone  = _mm_set1_epi8(1);
+    const __m128i m2 = _mm_set1_epi8(2);
+
+    __m256 acc = _mm256_setzero_ps();
+
+    float summs = 0.f;
+
+    for (int i = 0; i < nb; ++i) {
+
+        const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d);
+        const float dmin = -y[i].d * GGML_FP16_TO_FP32(x[i].dmin);
+
+        const uint8_t * restrict q5 = x[i].qs;
+        const int8_t  * restrict q8 = y[i].qs;
+
+        memcpy(utmp, x[i].scales, 12);
+        utmp[3] = ((utmp[2] >> 4) & kmask2) | (((utmp[1] >> 6) & kmask3) << 4);
+        const uint32_t uaux = utmp[1] & kmask1;
+        utmp[1] = (utmp[2] & kmask2) | (((utmp[0] >> 6) & kmask3) << 4);
+        utmp[2] = uaux;
+        utmp[0] &= kmask1;
+
+        const __m128i utmps = _mm_set_epi32(utmp[3], utmp[2], utmp[1], utmp[0]);
+        const __m128i scales = _mm_cvtepu8_epi16(utmps);
+        const __m128i mins = _mm_cvtepu8_epi16(_mm_unpackhi_epi64(utmps, utmps));
+
+        const __m128i q8sums_0 = _mm_loadu_si128((const __m128i*)&y[i].bsums[0]);
+        const __m128i q8sums_1 = _mm_loadu_si128((const __m128i*)&y[i].bsums[8]);
+        const __m128i q8s = _mm_hadd_epi16(q8sums_0, q8sums_1);
+        const __m128i prod = _mm_madd_epi16(mins, q8s);
+        const __m128i hsum = _mm_hadd_epi32(_mm_hadd_epi32(prod, mzero), mzero);
+        summs += dmin * _mm_extract_epi32(hsum, 0);
+
+        const __m128i hbits_0 = _mm_loadu_si128((const __m128i*)&x[i].qh[0]);
+        const __m128i hbits_1 = _mm_loadu_si128((const __m128i*)&x[i].qh[16]);
+        __m128i hmask = mone;
+
+        __m128i sumi_0 = _mm_setzero_si128();
+        __m128i sumi_1 = _mm_setzero_si128();
+
+        int bit = 0;
+
+        __m128i shuffle = _mm_set1_epi16(0x0100);
+        for (int j = 0; j < QK_K/64; ++j) {
+
+            const __m128i scale_0 = _mm_shuffle_epi8(scales, shuffle);
+            shuffle = _mm_add_epi16(shuffle, m2);
+            const __m128i scale_1 = _mm_shuffle_epi8(scales, shuffle);
+            shuffle = _mm_add_epi16(shuffle, m2);
+
+            const __m128i q5bits_0 = _mm_loadu_si128((const __m128i*)q5); q5 += 16;
+            const __m128i q5bits_1 = _mm_loadu_si128((const __m128i*)q5); q5 += 16;
+
+            __m128i q5l_0 = _mm_and_si128(q5bits_0, m4);
+            __m128i q5l_1 = _mm_and_si128(q5bits_1, m4);
+            __m128i q5h_0 = _mm_slli_epi16(_mm_srli_epi16(_mm_and_si128(hbits_0, hmask), bit), 4);
+            __m128i q5h_1 = _mm_slli_epi16(_mm_srli_epi16(_mm_and_si128(hbits_1, hmask), bit++), 4);
+            __m128i q5_0  = _mm_add_epi8(q5l_0, q5h_0);
+            __m128i q5_1  = _mm_add_epi8(q5l_1, q5h_1);
+            hmask = _mm_slli_epi16(hmask, 1);
+
+            __m128i q8_0 = _mm_loadu_si128((const __m128i*)q8); q8 += 16;
+            __m128i q8_1 = _mm_loadu_si128((const __m128i*)q8); q8 += 16;
+            __m128i p16_0 = _mm_maddubs_epi16(q5_0, q8_0);
+            __m128i p16_1 = _mm_maddubs_epi16(q5_1, q8_1);
+            p16_0 = _mm_madd_epi16(scale_0, p16_0);
+            p16_1 = _mm_madd_epi16(scale_0, p16_1);
+
+            q5l_0 = _mm_and_si128(_mm_srli_epi16(q5bits_0, 4), m4);
+            q5l_1 = _mm_and_si128(_mm_srli_epi16(q5bits_1, 4), m4);
+            q5h_0 = _mm_slli_epi16(_mm_srli_epi16(_mm_and_si128(hbits_0, hmask), bit), 4);
+            q5h_1 = _mm_slli_epi16(_mm_srli_epi16(_mm_and_si128(hbits_1, hmask), bit++), 4);
+            q5_0  = _mm_add_epi8(q5l_0, q5h_0);
+            q5_1  = _mm_add_epi8(q5l_1, q5h_1);
+            hmask = _mm_slli_epi16(hmask, 1);
+
+            q8_0 = _mm_loadu_si128((const __m128i*)q8); q8 += 16;
+            q8_1 = _mm_loadu_si128((const __m128i*)q8); q8 += 16;
+            __m128i p16_2 = _mm_maddubs_epi16(q5_0, q8_0);
+            __m128i p16_3 = _mm_maddubs_epi16(q5_1, q8_1);
+            p16_2 = _mm_madd_epi16(scale_1, p16_2);
+            p16_3 = _mm_madd_epi16(scale_1, p16_3);
+
+            sumi_0 = _mm_add_epi32(sumi_0, _mm_add_epi32(p16_0, p16_2));
+            sumi_1 = _mm_add_epi32(sumi_1, _mm_add_epi32(p16_1, p16_3));
+
+        }
+
+        __m256 vd = _mm256_set1_ps(d);
+        __m256i sumi = MM256_SET_M128I(sumi_1, sumi_0);
+        acc = _mm256_add_ps(_mm256_mul_ps(vd, _mm256_cvtepi32_ps(sumi)), acc);
+
+    }
+
+    *s = hsum_float_8(acc) + summs;
+
+#elif defined __riscv_v_intrinsic
+
+    const uint8_t * scales = (const uint8_t*)&utmp[0];
+    const uint8_t * mins   = (const uint8_t*)&utmp[2];
+
+    float sumf = 0;
+    float sums = 0.0;
+
+    size_t vl;
+
+    for (int i = 0; i < nb; ++i) {
+
+        vl = 8;
+
+        const uint8_t * restrict q5 = x[i].qs;
+        const uint8_t * restrict hm = x[i].qh;
+        const  int8_t * restrict q8 = y[i].qs;
+
+        const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
+        const float dmin = GGML_FP16_TO_FP32(x[i].dmin) * y[i].d;
+
+        vint16mf2_t q8sums_0 = __riscv_vlse16_v_i16mf2(y[i].bsums, 4, vl);
+        vint16mf2_t q8sums_1 = __riscv_vlse16_v_i16mf2(y[i].bsums+1, 4, vl);
+        vint16mf2_t q8sums = __riscv_vadd_vv_i16mf2(q8sums_0, q8sums_1, vl);
+
+        memcpy(utmp, x[i].scales, 12);
+        utmp[3] = ((utmp[2] >> 4) & kmask2) | (((utmp[1] >> 6) & kmask3) << 4);
+        const uint32_t uaux = utmp[1] & kmask1;
+        utmp[1] = (utmp[2] & kmask2) | (((utmp[0] >> 6) & kmask3) << 4);
+        utmp[2] = uaux;
+        utmp[0] &= kmask1;
+
+        vuint8mf4_t mins8 = __riscv_vle8_v_u8mf4(mins, vl);
+        vint16mf2_t v_mins = __riscv_vreinterpret_v_u16mf2_i16mf2(__riscv_vzext_vf2_u16mf2(mins8, vl));
+        vint32m1_t prod = __riscv_vwmul_vv_i32m1(q8sums, v_mins, vl);
+
+        vint32m1_t sumi = __riscv_vredsum_vs_i32m1_i32m1(prod, __riscv_vmv_v_x_i32m1(0, 1), vl);
+        sumf -= dmin * __riscv_vmv_x_s_i32m1_i32(sumi);
+
+        vl = 32;
+        int32_t aux32 = 0;
+        int is = 0;
+
+        uint8_t m = 1;
+        vint32m1_t vzero = __riscv_vmv_v_x_i32m1(0, 1);
+        vuint8m1_t vqh = __riscv_vle8_v_u8m1(hm, vl);
+
+        for (int j = 0; j < QK_K/64; ++j) {
+            // load Q5 and Q8
+            vuint8m1_t q5_x = __riscv_vle8_v_u8m1(q5, vl);
+            vint8m1_t  q8_y1 = __riscv_vle8_v_i8m1(q8, vl);
+            vint8m1_t  q8_y2 = __riscv_vle8_v_i8m1(q8+32, vl);
+
+            // compute mask for addition
+            vint8m1_t q5_a = __riscv_vreinterpret_v_u8m1_i8m1(__riscv_vand_vx_u8m1(q5_x, 0x0F, vl));
+            vuint8m1_t qh_m1 = __riscv_vand_vx_u8m1(vqh, m, vl);
+            vbool8_t vmask_1 = __riscv_vmsne_vx_u8m1_b8(qh_m1, 0, vl);
+            vint8m1_t q5_m1 = __riscv_vadd_vx_i8m1_mu(vmask_1, q5_a, q5_a, 16, vl);
+            m <<= 1;
+
+            vint8m1_t q5_l = __riscv_vreinterpret_v_u8m1_i8m1(__riscv_vsrl_vx_u8m1(q5_x, 0x04, vl));
+            vuint8m1_t qh_m2 = __riscv_vand_vx_u8m1(vqh, m, vl);
+            vbool8_t vmask_2 = __riscv_vmsne_vx_u8m1_b8(qh_m2, 0, vl);
+            vint8m1_t q5_m2 = __riscv_vadd_vx_i8m1_mu(vmask_2, q5_l, q5_l, 16, vl);
+            m <<= 1;
+
+            vint16m2_t v0 = __riscv_vwmul_vv_i16m2(q5_m1, q8_y1, vl);
+            vint16m2_t v1 = __riscv_vwmul_vv_i16m2(q5_m2, q8_y2, vl);
+
+            vint32m4_t vs1 = __riscv_vwmul_vx_i32m4(v0, scales[is++], vl);
+            vint32m4_t vs2 = __riscv_vwmul_vx_i32m4(v1, scales[is++], vl);
+
+            vint32m1_t vacc1 = __riscv_vredsum_vs_i32m4_i32m1(vs1, vzero, vl);
+            vint32m1_t vacc2 = __riscv_vredsum_vs_i32m4_i32m1(vs2, vzero, vl);
+
+            aux32 += __riscv_vmv_x_s_i32m1_i32(vacc1) + __riscv_vmv_x_s_i32m1_i32(vacc2);
+            q5 += 32;    q8 += 64;
+
+        }
+
+        vfloat32m1_t vaux = __riscv_vfmul_vf_f32m1(__riscv_vfmv_v_f_f32m1(aux32, 1), d, 1);
+        sums += __riscv_vfmv_f_s_f32m1_f32(vaux);
+
+    }
+
+    *s = sumf+sums;
+
+#elif defined(__POWER9_VECTOR__)
+    const vector signed char lowMask = vec_splats((signed char)0xF);
+    const vector signed char lowMask1 = vec_splats((int8_t)0x3f);
+    const vector signed char lowMask2 = vec_splats((int8_t)0x30);
+    const vector int v0 = vec_splats((int32_t)0);
+    const vector unsigned char v1 = vec_splats((unsigned char)0x1);
+    const vector unsigned char v2 = vec_splats((unsigned char)0x2);
+    const vector unsigned char v3 = vec_splats((unsigned char)0x3);
+    const vector unsigned char v4 = vec_splats((unsigned char)0x4);
+
+    vector float vsumf0 = vec_splats(0.0f);
+    vector float vsumf1 = vec_splats(0.0f);
+    vector float vsumf2 = vec_splats(0.0f);
+    vector float vsumf3 = vec_splats(0.0f);
+
+    for (int i = 0; i < nb; ++i) {
+        vector float vxd = vec_splats(GGML_FP16_TO_FP32(x[i].d));
+        vector float vyd = vec_splats(y[i].d);
+        vector float vd = vec_mul(vxd, vyd);
+
+        vector float vxmin = vec_splats(GGML_FP16_TO_FP32(x[i].dmin));
+        vector float vdmin = vec_mul(vxmin, vyd);
+
+        UNUSED(kmask1);
+        UNUSED(kmask2);
+        UNUSED(kmask3);
+        UNUSED(utmp);
+
+        vector signed char u0 = (vector signed char)vec_xl_len(x[i].scales, 8);
+        vector signed char u1 = vec_and(vec_sr(u0, v2), lowMask2);
+        vector signed char u2 = (vector signed char)vec_xl_len(x[i].scales + 8, 4);
+        vector signed char u3 = vec_sr(u2, v4);
+
+        vector signed char u30 = u1;
+        vector signed char u31 = (vector signed char)vec_mergeh((vector signed int)vec_and(u2, lowMask), (vector signed int)u3);
+
+        u1 = vec_and(u0, lowMask1);
+        u2 = vec_or(u30, u31);
+
+        vector signed char utmps = (vector signed char)vec_mergeh((vector signed int)u1, (vector signed int)u2);
+
+        vector signed short q8ysums0 = vec_xl( 0, y[i].bsums);
+        vector signed short q8ysums1 = vec_xl(16, y[i].bsums);
+
+        vector signed short vscales = vec_unpackh(utmps);
+
+        vector signed short q5xmins = vec_unpackl(utmps);
+        vector signed short q5xmins0 = vec_mergeh(q5xmins, q5xmins);
+        vector signed short q5xmins1 = vec_mergel(q5xmins, q5xmins);
+
+        vector signed int prod0 = vec_mule(q5xmins0, q8ysums0);
+        vector signed int prod1 = vec_mule(q5xmins1, q8ysums1);
+        vector signed int prod2 = vec_mulo(q5xmins0, q8ysums0);
+        vector signed int prod3 = vec_mulo(q5xmins1, q8ysums1);
+
+        vsumf0 = vec_nmsub(vec_ctf(prod0, 0), vdmin, vsumf0);
+        vsumf1 = vec_nmsub(vec_ctf(prod1, 0), vdmin, vsumf1);
+        vsumf2 = vec_nmsub(vec_ctf(prod2, 0), vdmin, vsumf2);
+        vsumf3 = vec_nmsub(vec_ctf(prod3, 0), vdmin, vsumf3);
+
+        vector signed char qxhs0 = (vector signed char)vec_xl( 0, x[i].qh);
+        vector signed char qxhs1 = (vector signed char)vec_xl(16, x[i].qh);
+
+        vector signed int vsumi0 = v0;
+        vector signed int vsumi1 = v0;
+        vector signed int vsumi2 = v0;
+        vector signed int vsumi3 = v0;
+
+        const uint8_t * restrict q5 = x[i].qs;
+        const int8_t  * restrict q8 = y[i].qs;
+
+        for (int j = 0; j < QK_K/64; ++j) {
+            __builtin_prefetch(q5, 0, 1);
+            __builtin_prefetch(q8, 0, 1);
+
+            vector signed char qxs0 = (vector signed char)vec_xl( 0, q5);
+            vector signed char qxs1 = (vector signed char)vec_xl(16, q5);
+            q5 += 32;
+
+            vector signed char qxs00 = vec_and(qxs0, lowMask);
+            vector signed char qxs01 = vec_sr(qxs0, v4);
+            vector signed char qxs10 = vec_and(qxs1, lowMask);
+            vector signed char qxs11 = vec_sr(qxs1, v4);
+
+            vector signed char q5h00 = vec_sl(vec_and((vector signed char)v1, qxhs0), v4);
+            vector signed char q5h01 = vec_sl(vec_and((vector signed char)v2, qxhs0), v3);
+            vector signed char q5h10 = vec_sl(vec_and((vector signed char)v1, qxhs1), v4);
+            vector signed char q5h11 = vec_sl(vec_and((vector signed char)v2, qxhs1), v3);
+            qxhs0 = vec_sr(qxhs0, v2);
+            qxhs1 = vec_sr(qxhs1, v2);
+
+            vector unsigned char q5x00 = (vector unsigned char)vec_or(q5h00, qxs00);
+            vector unsigned char q5x01 = (vector unsigned char)vec_or(q5h01, qxs01);
+            vector unsigned char q5x10 = (vector unsigned char)vec_or(q5h10, qxs10);
+            vector unsigned char q5x11 = (vector unsigned char)vec_or(q5h11, qxs11);
+
+            vector signed char q8y00 = vec_xl( 0, q8);
+            vector signed char q8y10 = vec_xl(16, q8);
+            vector signed char q8y01 = vec_xl(32, q8);
+            vector signed char q8y11 = vec_xl(48, q8);
+            q8 += 64;
+
+            vector signed int qv00 = vec_msum(q8y00, q5x00, v0);
+            vector signed int qv01 = vec_msum(q8y01, q5x01, v0);
+            vector signed int qv10 = vec_msum(q8y10, q5x10, v0);
+            vector signed int qv11 = vec_msum(q8y11, q5x11, v0);
+
+            vector signed int vscales_h = vec_unpackh(vscales);
+            vector signed int vs0 = vec_splat(vscales_h, 0);
+            vector signed int vs1 = vec_splat(vscales_h, 1);
+            vscales = vec_sld(vscales, vscales, 12);
+
+            vsumi0 = vec_add(vec_mul(qv00, vs0), vsumi0);
+            vsumi1 = vec_add(vec_mul(qv10, vs0), vsumi1);
+            vsumi2 = vec_add(vec_mul(qv01, vs1), vsumi2);
+            vsumi3 = vec_add(vec_mul(qv11, vs1), vsumi3);
+        }
+
+        vsumf0 = vec_madd(vec_ctf(vsumi0, 0), vd, vsumf0);
+        vsumf1 = vec_madd(vec_ctf(vsumi1, 0), vd, vsumf1);
+        vsumf2 = vec_madd(vec_ctf(vsumi2, 0), vd, vsumf2);
+        vsumf3 = vec_madd(vec_ctf(vsumi3, 0), vd, vsumf3);
+    }
+
+    vsumf0 = vec_add(vsumf0, vsumf2);
+    vsumf1 = vec_add(vsumf1, vsumf3);
+
+    vsumf0 = vec_add(vsumf0, vsumf1);
+
+    vsumf0 = vec_add(vsumf0, vec_sld(vsumf0, vsumf0, 4));
+    vsumf0 = vec_add(vsumf0, vec_sld(vsumf0, vsumf0, 8));
+
+    *s = vec_extract(vsumf0, 0);
+
+#elif defined __loongarch_asx
+    GGML_UNUSED(kmask1);
+    GGML_UNUSED(kmask2);
+    GGML_UNUSED(kmask3);
+
+    const __m256i m4 = __lasx_xvreplgr2vr_b(0xF);
+    const __m128i mzero = __lsx_vldi(0);
+    const __m256i mone  = __lasx_xvreplgr2vr_b(1);
+
+    __m256 acc = (__m256)__lasx_xvldi(0);
+
+    float summs = 0.f;
+
+   for (int i = 0; i < nb; ++i) {
+
+        const uint8_t * restrict q5 = x[i].qs;
+        const int8_t  * restrict q8 = y[i].qs;
+
+        const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d);
+        const float dmin = -y[i].d * GGML_FP16_TO_FP32(x[i].dmin);
+
+        memcpy(utmp, x[i].scales, 12);
+        utmp[3] = ((utmp[2] >> 4) & kmask2) | (((utmp[1] >> 6) & kmask3) << 4);
+        const uint32_t uaux = utmp[1] & kmask1;
+        utmp[1] = (utmp[2] & kmask2) | (((utmp[0] >> 6) & kmask3) << 4);
+        utmp[2] = uaux;
+        utmp[0] &= kmask1;
+
+        const __m256i mins_and_scales = lasx_extu8_16(lsx_set_w(utmp[3], utmp[2], utmp[1], utmp[0]));
+
+        const __m256i q8sums = __lasx_xvld((const __m256i*)y[i].bsums, 0);
+        const __m128i q8s = lsx_hadd_h(lasx_extracti128(q8sums, 0), lasx_extracti128(q8sums, 1));
+        const __m128i prod = lsx_madd_h(lasx_extracti128(mins_and_scales, 1), q8s);
+        const __m128i hsum = lsx_hadd_w(lsx_hadd_w(prod, mzero), mzero);
+        summs += dmin * __lsx_vpickve2gr_w(hsum, 0);    //TODO check
+
+        const __m128i sc128  = lasx_extracti128(mins_and_scales, 0);
+        const __m256i scales = lasx_insertf128(sc128, sc128);
+
+        const __m256i hbits = __lasx_xvld((const __m256i*)x[i].qh, 0);
+        __m256i hmask = mone;
+
+        __m256i sumi = __lasx_xvldi(0);
+
+        int bit = 0;
+        __m256i xvbit;
+
+        for (int j = 0; j < QK_K/64; ++j) {
+
+            const __m256i scale_0 = lasx_shuffle_b(scales, get_scale_shuffle_k4(2*j+0));
+            const __m256i scale_1 = lasx_shuffle_b(scales, get_scale_shuffle_k4(2*j+1));
+
+            const __m256i q5bits = __lasx_xvld((const __m256i*)q5, 0); q5 += 32;
+
+            xvbit = __lasx_xvreplgr2vr_h(bit++);
+            const __m256i q5l_0 = __lasx_xvand_v(q5bits, m4);
+            const __m256i q5h_0 = __lasx_xvslli_h(__lasx_xvsrl_h(__lasx_xvand_v(hbits, hmask), xvbit), 4);
+            const __m256i q5_0  = __lasx_xvadd_b(q5l_0, q5h_0);
+            hmask = __lasx_xvslli_h(hmask, 1);
+
+            xvbit = __lasx_xvreplgr2vr_h(bit++);
+            const __m256i q5l_1 = __lasx_xvand_v(__lasx_xvsrli_h(q5bits, 4), m4);
+            const __m256i q5h_1 = __lasx_xvslli_h(__lasx_xvsrl_h(__lasx_xvand_v(hbits, hmask), xvbit), 4);
+            const __m256i q5_1  = __lasx_xvadd_b(q5l_1, q5h_1);
+            hmask = __lasx_xvslli_h(hmask, 1);
+
+            const __m256i q8_0 = __lasx_xvld((const __m256i*)q8, 0); q8 += 32;
+            const __m256i q8_1 = __lasx_xvld((const __m256i*)q8, 0); q8 += 32;
+
+            __m256i p16_0 = lasx_maddubs_h(q5_0, q8_0);
+            __m256i p16_1 = lasx_maddubs_h(q5_1, q8_1);
+
+            p16_0 = lasx_madd_h(scale_0, p16_0);
+            p16_1 = lasx_madd_h(scale_1, p16_1);
+
+            sumi = __lasx_xvadd_w(sumi, __lasx_xvadd_w(p16_0, p16_1));
+
+        }
+
+        __m256 vd = __lasx_xvreplfr2vr_s(d);
+        acc = __lasx_xvfmadd_s(vd, __lasx_xvffint_s_w(sumi), acc);
+
+    }
+
+    *s = hsum_float_8(acc) + summs;
+
+#else
+
+    const uint8_t * scales = (const uint8_t*)&utmp[0];
+    const uint8_t * mins   = (const uint8_t*)&utmp[2];
+
+    int8_t  aux8[QK_K];
+    int16_t aux16[8];
+    float   sums [8];
+    int32_t aux32[8];
+    memset(sums, 0, 8*sizeof(float));
+
+    float sumf = 0;
+    for (int i = 0; i < nb; ++i) {
+        const uint8_t * restrict q4 = x[i].qs;
+        const uint8_t * restrict hm = x[i].qh;
+        const  int8_t * restrict q8 = y[i].qs;
+        memset(aux32, 0, 8*sizeof(int32_t));
+        int8_t * restrict a = aux8;
+        uint8_t m = 1;
+        for (int j = 0; j < QK_K/64; ++j) {
+            for (int l = 0; l < 32; ++l) a[l] = (int8_t)(q4[l] & 0xF);
+            for (int l = 0; l < 32; ++l) a[l] += (hm[l] & m ? 16 : 0);
+            a += 32; m <<= 1;
+            for (int l = 0; l < 32; ++l) a[l] = (int8_t)(q4[l]  >> 4);
+            for (int l = 0; l < 32; ++l) a[l] += (hm[l] & m ? 16 : 0);
+            a += 32; m <<= 1;
+            q4 += 32;
+        }
+        memcpy(utmp, x[i].scales, 12);
+        utmp[3] = ((utmp[2] >> 4) & kmask2) | (((utmp[1] >> 6) & kmask3) << 4);
+        const uint32_t uaux = utmp[1] & kmask1;
+        utmp[1] = (utmp[2] & kmask2) | (((utmp[0] >> 6) & kmask3) << 4);
+        utmp[2] = uaux;
+        utmp[0] &= kmask1;
+
+        int sumi = 0;
+        for (int j = 0; j < QK_K/16; ++j) sumi += y[i].bsums[j] * mins[j/2];
+        a = aux8;
+        int is = 0;
+        for (int j = 0; j < QK_K/32; ++j) {
+            int32_t scale = scales[is++];
+            for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
+            for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
+            q8 += 8; a += 8;
+            for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
+            for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
+            q8 += 8; a += 8;
+            for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
+            for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
+            q8 += 8; a += 8;
+            for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
+            for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
+            q8 += 8; a += 8;
+        }
+        const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
+        for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l];
+        const float dmin = GGML_FP16_TO_FP32(x[i].dmin) * y[i].d;
+        sumf -= dmin * sumi;
+    }
+    for (int l = 0; l < 8; ++l) sumf += sums[l];
+    *s = sumf;
+#endif
+}
+
+void ggml_vec_dot_q6_K_q8_K(int n, float * restrict s, size_t bs, const void * restrict vx, size_t bx, const void * restrict vy, size_t by, int nrc) {
+    assert(n % QK_K == 0);
+    assert(nrc == 1);
+    UNUSED(nrc);
+    UNUSED(bx);
+    UNUSED(by);
+    UNUSED(bs);
+
+    const block_q6_K * restrict x = vx;
+    const block_q8_K * restrict y = vy;
+
+    const int nb = n / QK_K;
+
+#ifdef __ARM_NEON
+    float sum = 0;
+
+    const uint8x16_t m4b = vdupq_n_u8(0xF);
+    const int32x4_t  vzero = vdupq_n_s32(0);
+    //const int8x16_t  m32s = vdupq_n_s8(32);
+
+    const uint8x16_t mone = vdupq_n_u8(3);
+
+    ggml_int8x16x4_t q6bytes;
+    ggml_uint8x16x4_t q6h;
+
+    for (int i = 0; i < nb; ++i) {
+
+        const float d_all = GGML_FP16_TO_FP32(x[i].d);
+
+        const uint8_t * restrict q6 = x[i].ql;
+        const uint8_t * restrict qh = x[i].qh;
+        const int8_t  * restrict q8 = y[i].qs;
+
+        const int8_t * restrict scale = x[i].scales;
+
+        const ggml_int16x8x2_t q8sums = ggml_vld1q_s16_x2(y[i].bsums);
+        const int8x16_t scales = vld1q_s8(scale);
+        const ggml_int16x8x2_t q6scales = {{vmovl_s8(vget_low_s8(scales)), vmovl_s8(vget_high_s8(scales))}};
+
+        const int32x4_t prod = vaddq_s32(vaddq_s32(vmull_s16(vget_low_s16 (q8sums.val[0]), vget_low_s16 (q6scales.val[0])),
+                                                   vmull_s16(vget_high_s16(q8sums.val[0]), vget_high_s16(q6scales.val[0]))),
+                                         vaddq_s32(vmull_s16(vget_low_s16 (q8sums.val[1]), vget_low_s16 (q6scales.val[1])),
+                                                   vmull_s16(vget_high_s16(q8sums.val[1]), vget_high_s16(q6scales.val[1]))));
+        int32_t isum_mins = vaddvq_s32(prod);
+
+        int32_t isum = 0;
+
+        for (int j = 0; j < QK_K/128; ++j) {
+
+            ggml_uint8x16x2_t qhbits = ggml_vld1q_u8_x2(qh); qh += 32;
+            ggml_uint8x16x4_t q6bits = ggml_vld1q_u8_x4(q6); q6 += 64;
+            ggml_int8x16x4_t q8bytes = ggml_vld1q_s8_x4(q8); q8 += 64;
+
+            q6h.val[0] = vshlq_n_u8(vandq_u8(mone, qhbits.val[0]), 4);
+            q6h.val[1] = vshlq_n_u8(vandq_u8(mone, qhbits.val[1]), 4);
+            uint8x16_t shifted = vshrq_n_u8(qhbits.val[0], 2);
+            q6h.val[2] = vshlq_n_u8(vandq_u8(mone, shifted), 4);
+            shifted = vshrq_n_u8(qhbits.val[1], 2);
+            q6h.val[3] = vshlq_n_u8(vandq_u8(mone, shifted), 4);
+
+            //q6bytes.val[0] = vsubq_s8(vreinterpretq_s8_u8(vorrq_u8(vandq_u8(q6bits.val[0], m4b), q6h.val[0])), m32s);
+            //q6bytes.val[1] = vsubq_s8(vreinterpretq_s8_u8(vorrq_u8(vandq_u8(q6bits.val[1], m4b), q6h.val[1])), m32s);
+            //q6bytes.val[2] = vsubq_s8(vreinterpretq_s8_u8(vorrq_u8(vandq_u8(q6bits.val[2], m4b), q6h.val[2])), m32s);
+            //q6bytes.val[3] = vsubq_s8(vreinterpretq_s8_u8(vorrq_u8(vandq_u8(q6bits.val[3], m4b), q6h.val[3])), m32s);
+            q6bytes.val[0] = vreinterpretq_s8_u8(vorrq_u8(vandq_u8(q6bits.val[0], m4b), q6h.val[0]));
+            q6bytes.val[1] = vreinterpretq_s8_u8(vorrq_u8(vandq_u8(q6bits.val[1], m4b), q6h.val[1]));
+            q6bytes.val[2] = vreinterpretq_s8_u8(vorrq_u8(vandq_u8(q6bits.val[2], m4b), q6h.val[2]));
+            q6bytes.val[3] = vreinterpretq_s8_u8(vorrq_u8(vandq_u8(q6bits.val[3], m4b), q6h.val[3]));
+
+            isum += vaddvq_s32(ggml_vdotq_s32(vzero, q6bytes.val[0], q8bytes.val[0])) * scale[0] +
+                    vaddvq_s32(ggml_vdotq_s32(vzero, q6bytes.val[1], q8bytes.val[1])) * scale[1] +
+                    vaddvq_s32(ggml_vdotq_s32(vzero, q6bytes.val[2], q8bytes.val[2])) * scale[2] +
+                    vaddvq_s32(ggml_vdotq_s32(vzero, q6bytes.val[3], q8bytes.val[3])) * scale[3];
+
+            scale += 4;
+
+            q8bytes = ggml_vld1q_s8_x4(q8); q8 += 64;
+
+            shifted = vshrq_n_u8(qhbits.val[0], 4);
+            q6h.val[0] = vshlq_n_u8(vandq_u8(mone, shifted), 4);
+            shifted = vshrq_n_u8(qhbits.val[1], 4);
+            q6h.val[1] = vshlq_n_u8(vandq_u8(mone, shifted), 4);
+            shifted = vshrq_n_u8(qhbits.val[0], 6);
+            q6h.val[2] = vshlq_n_u8(vandq_u8(mone, shifted), 4);
+            shifted = vshrq_n_u8(qhbits.val[1], 6);
+            q6h.val[3] = vshlq_n_u8(vandq_u8(mone, shifted), 4);
+
+            //q6bytes.val[0] = vsubq_s8(vreinterpretq_s8_u8(vorrq_u8(vshrq_n_u8(q6bits.val[0], 4), q6h.val[0])), m32s);
+            //q6bytes.val[1] = vsubq_s8(vreinterpretq_s8_u8(vorrq_u8(vshrq_n_u8(q6bits.val[1], 4), q6h.val[1])), m32s);
+            //q6bytes.val[2] = vsubq_s8(vreinterpretq_s8_u8(vorrq_u8(vshrq_n_u8(q6bits.val[2], 4), q6h.val[2])), m32s);
+            //q6bytes.val[3] = vsubq_s8(vreinterpretq_s8_u8(vorrq_u8(vshrq_n_u8(q6bits.val[3], 4), q6h.val[3])), m32s);
+            q6bytes.val[0] = vreinterpretq_s8_u8(vorrq_u8(vshrq_n_u8(q6bits.val[0], 4), q6h.val[0]));
+            q6bytes.val[1] = vreinterpretq_s8_u8(vorrq_u8(vshrq_n_u8(q6bits.val[1], 4), q6h.val[1]));
+            q6bytes.val[2] = vreinterpretq_s8_u8(vorrq_u8(vshrq_n_u8(q6bits.val[2], 4), q6h.val[2]));
+            q6bytes.val[3] = vreinterpretq_s8_u8(vorrq_u8(vshrq_n_u8(q6bits.val[3], 4), q6h.val[3]));
+
+            isum += vaddvq_s32(ggml_vdotq_s32(vzero, q6bytes.val[0], q8bytes.val[0])) * scale[0] +
+                    vaddvq_s32(ggml_vdotq_s32(vzero, q6bytes.val[1], q8bytes.val[1])) * scale[1] +
+                    vaddvq_s32(ggml_vdotq_s32(vzero, q6bytes.val[2], q8bytes.val[2])) * scale[2] +
+                    vaddvq_s32(ggml_vdotq_s32(vzero, q6bytes.val[3], q8bytes.val[3])) * scale[3];
+            scale += 4;
+        }
+        //sum += isum * d_all * y[i].d;
+        sum += d_all * y[i].d * (isum - 32 * isum_mins);
+
+    }
+    *s = sum;
+
+#elif defined __AVX2__
+
+    const __m256i m4 = _mm256_set1_epi8(0xF);
+    const __m256i m2 = _mm256_set1_epi8(3);
+    const __m256i m32s = _mm256_set1_epi8(32);
+
+    __m256 acc = _mm256_setzero_ps();
+
+    for (int i = 0; i < nb; ++i) {
+
+        const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d);
+
+        const uint8_t * restrict q4 = x[i].ql;
+        const uint8_t * restrict qh = x[i].qh;
+        const int8_t  * restrict q8 = y[i].qs;
+
+        const __m128i scales = _mm_loadu_si128((const __m128i*)x[i].scales);
+
+        __m256i sumi = _mm256_setzero_si256();
+
+        int is = 0;
+
+        for (int j = 0; j < QK_K/128; ++j) {
+
+            const __m128i scale_0 = _mm_shuffle_epi8(scales, get_scale_shuffle(is + 0));
+            const __m128i scale_1 = _mm_shuffle_epi8(scales, get_scale_shuffle(is + 1));
+            const __m128i scale_2 = _mm_shuffle_epi8(scales, get_scale_shuffle(is + 2));
+            const __m128i scale_3 = _mm_shuffle_epi8(scales, get_scale_shuffle(is + 3));
+            is += 4;
+
+            const __m256i q4bits1 = _mm256_loadu_si256((const __m256i*)q4); q4 += 32;
+            const __m256i q4bits2 = _mm256_loadu_si256((const __m256i*)q4); q4 += 32;
+            const __m256i q4bitsH = _mm256_loadu_si256((const __m256i*)qh); qh += 32;
+
+            const __m256i q4h_0 = _mm256_slli_epi16(_mm256_and_si256(q4bitsH, m2), 4);
+            const __m256i q4h_1 = _mm256_slli_epi16(_mm256_and_si256(_mm256_srli_epi16(q4bitsH, 2), m2), 4);
+            const __m256i q4h_2 = _mm256_slli_epi16(_mm256_and_si256(_mm256_srli_epi16(q4bitsH, 4), m2), 4);
+            const __m256i q4h_3 = _mm256_slli_epi16(_mm256_and_si256(_mm256_srli_epi16(q4bitsH, 6), m2), 4);
+
+            const __m256i q4_0 = _mm256_or_si256(_mm256_and_si256(q4bits1, m4), q4h_0);
+            const __m256i q4_1 = _mm256_or_si256(_mm256_and_si256(q4bits2, m4), q4h_1);
+            const __m256i q4_2 = _mm256_or_si256(_mm256_and_si256(_mm256_srli_epi16(q4bits1, 4), m4), q4h_2);
+            const __m256i q4_3 = _mm256_or_si256(_mm256_and_si256(_mm256_srli_epi16(q4bits2, 4), m4), q4h_3);
+
+            const __m256i q8_0 = _mm256_loadu_si256((const __m256i*)q8); q8 += 32;
+            const __m256i q8_1 = _mm256_loadu_si256((const __m256i*)q8); q8 += 32;
+            const __m256i q8_2 = _mm256_loadu_si256((const __m256i*)q8); q8 += 32;
+            const __m256i q8_3 = _mm256_loadu_si256((const __m256i*)q8); q8 += 32;
+
+            __m256i q8s_0 = _mm256_maddubs_epi16(m32s, q8_0);
+            __m256i q8s_1 = _mm256_maddubs_epi16(m32s, q8_1);
+            __m256i q8s_2 = _mm256_maddubs_epi16(m32s, q8_2);
+            __m256i q8s_3 = _mm256_maddubs_epi16(m32s, q8_3);
+
+            __m256i p16_0 = _mm256_maddubs_epi16(q4_0, q8_0);
+            __m256i p16_1 = _mm256_maddubs_epi16(q4_1, q8_1);
+            __m256i p16_2 = _mm256_maddubs_epi16(q4_2, q8_2);
+            __m256i p16_3 = _mm256_maddubs_epi16(q4_3, q8_3);
+
+            p16_0 = _mm256_sub_epi16(p16_0, q8s_0);
+            p16_1 = _mm256_sub_epi16(p16_1, q8s_1);
+            p16_2 = _mm256_sub_epi16(p16_2, q8s_2);
+            p16_3 = _mm256_sub_epi16(p16_3, q8s_3);
+
+            p16_0 = _mm256_madd_epi16(_mm256_cvtepi8_epi16(scale_0), p16_0);
+            p16_1 = _mm256_madd_epi16(_mm256_cvtepi8_epi16(scale_1), p16_1);
+            p16_2 = _mm256_madd_epi16(_mm256_cvtepi8_epi16(scale_2), p16_2);
+            p16_3 = _mm256_madd_epi16(_mm256_cvtepi8_epi16(scale_3), p16_3);
+
+            sumi = _mm256_add_epi32(sumi, _mm256_add_epi32(p16_0, p16_1));
+            sumi = _mm256_add_epi32(sumi, _mm256_add_epi32(p16_2, p16_3));
+
+        }
+
+        acc = _mm256_fmadd_ps(_mm256_broadcast_ss(&d), _mm256_cvtepi32_ps(sumi), acc);
+    }
+
+    *s = hsum_float_8(acc);
+
+#elif defined __AVX__
+
+    const __m128i m4 = _mm_set1_epi8(0xF);
+    const __m128i m3 = _mm_set1_epi8(3);
+    const __m128i m32s = _mm_set1_epi8(32);
+    const __m128i m2 = _mm_set1_epi8(2);
+
+    __m256 acc = _mm256_setzero_ps();
+
+    for (int i = 0; i < nb; ++i) {
+
+        const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d);
+
+        const uint8_t * restrict q4 = x[i].ql;
+        const uint8_t * restrict qh = x[i].qh;
+        const int8_t  * restrict q8 = y[i].qs;
+
+        const __m128i scales = _mm_loadu_si128((const __m128i*)x[i].scales);
+
+        __m128i sumi_0 = _mm_setzero_si128();
+        __m128i sumi_1 = _mm_setzero_si128();
+
+        __m128i shuffle = _mm_set_epi64x(0x0101010101010101, 0x0000000000000000);
+        for (int j = 0; j < QK_K/128; ++j) {
+
+            const __m128i q4bitsH_0 = _mm_loadu_si128((const __m128i*)qh); qh += 16;
+            const __m128i q4bitsH_1 = _mm_loadu_si128((const __m128i*)qh); qh += 16;
+
+            const __m128i q4h_0 = _mm_slli_epi16(_mm_and_si128(q4bitsH_0, m3), 4);
+            const __m128i q4h_1 = _mm_slli_epi16(_mm_and_si128(q4bitsH_1, m3), 4);
+            const __m128i q4h_2 = _mm_slli_epi16(_mm_and_si128(_mm_srli_epi16(q4bitsH_0, 2), m3), 4);
+            const __m128i q4h_3 = _mm_slli_epi16(_mm_and_si128(_mm_srli_epi16(q4bitsH_1, 2), m3), 4);
+            const __m128i q4h_4 = _mm_slli_epi16(_mm_and_si128(_mm_srli_epi16(q4bitsH_0, 4), m3), 4);
+            const __m128i q4h_5 = _mm_slli_epi16(_mm_and_si128(_mm_srli_epi16(q4bitsH_1, 4), m3), 4);
+            const __m128i q4h_6 = _mm_slli_epi16(_mm_and_si128(_mm_srli_epi16(q4bitsH_0, 6), m3), 4);
+            const __m128i q4h_7 = _mm_slli_epi16(_mm_and_si128(_mm_srli_epi16(q4bitsH_1, 6), m3), 4);
+
+            const __m128i q4bits1_0 = _mm_loadu_si128((const __m128i*)q4); q4 += 16;
+            const __m128i q4bits1_1 = _mm_loadu_si128((const __m128i*)q4); q4 += 16;
+            const __m128i q4bits2_0 = _mm_loadu_si128((const __m128i*)q4); q4 += 16;
+            const __m128i q4bits2_1 = _mm_loadu_si128((const __m128i*)q4); q4 += 16;
+
+            const __m128i q4_0 = _mm_or_si128(_mm_and_si128(q4bits1_0, m4), q4h_0);
+            const __m128i q4_1 = _mm_or_si128(_mm_and_si128(q4bits1_1, m4), q4h_1);
+            const __m128i q4_2 = _mm_or_si128(_mm_and_si128(q4bits2_0, m4), q4h_2);
+            const __m128i q4_3 = _mm_or_si128(_mm_and_si128(q4bits2_1, m4), q4h_3);
+            const __m128i q4_4 = _mm_or_si128(_mm_and_si128(_mm_srli_epi16(q4bits1_0, 4), m4), q4h_4);
+            const __m128i q4_5 = _mm_or_si128(_mm_and_si128(_mm_srli_epi16(q4bits1_1, 4), m4), q4h_5);
+            const __m128i q4_6 = _mm_or_si128(_mm_and_si128(_mm_srli_epi16(q4bits2_0, 4), m4), q4h_6);
+            const __m128i q4_7 = _mm_or_si128(_mm_and_si128(_mm_srli_epi16(q4bits2_1, 4), m4), q4h_7);
+
+            const __m128i q8_0 = _mm_loadu_si128((const __m128i*)q8); q8 += 16;
+            const __m128i q8_1 = _mm_loadu_si128((const __m128i*)q8); q8 += 16;
+            const __m128i q8_2 = _mm_loadu_si128((const __m128i*)q8); q8 += 16;
+            const __m128i q8_3 = _mm_loadu_si128((const __m128i*)q8); q8 += 16;
+            const __m128i q8_4 = _mm_loadu_si128((const __m128i*)q8); q8 += 16;
+            const __m128i q8_5 = _mm_loadu_si128((const __m128i*)q8); q8 += 16;
+            const __m128i q8_6 = _mm_loadu_si128((const __m128i*)q8); q8 += 16;
+            const __m128i q8_7 = _mm_loadu_si128((const __m128i*)q8); q8 += 16;
+
+            __m128i q8s_0 = _mm_maddubs_epi16(m32s, q8_0);
+            __m128i q8s_1 = _mm_maddubs_epi16(m32s, q8_1);
+            __m128i q8s_2 = _mm_maddubs_epi16(m32s, q8_2);
+            __m128i q8s_3 = _mm_maddubs_epi16(m32s, q8_3);
+            __m128i q8s_4 = _mm_maddubs_epi16(m32s, q8_4);
+            __m128i q8s_5 = _mm_maddubs_epi16(m32s, q8_5);
+            __m128i q8s_6 = _mm_maddubs_epi16(m32s, q8_6);
+            __m128i q8s_7 = _mm_maddubs_epi16(m32s, q8_7);
+
+            __m128i p16_0 = _mm_maddubs_epi16(q4_0, q8_0);
+            __m128i p16_1 = _mm_maddubs_epi16(q4_1, q8_1);
+            __m128i p16_2 = _mm_maddubs_epi16(q4_2, q8_2);
+            __m128i p16_3 = _mm_maddubs_epi16(q4_3, q8_3);
+            __m128i p16_4 = _mm_maddubs_epi16(q4_4, q8_4);
+            __m128i p16_5 = _mm_maddubs_epi16(q4_5, q8_5);
+            __m128i p16_6 = _mm_maddubs_epi16(q4_6, q8_6);
+            __m128i p16_7 = _mm_maddubs_epi16(q4_7, q8_7);
+
+            p16_0 = _mm_sub_epi16(p16_0, q8s_0);
+            p16_1 = _mm_sub_epi16(p16_1, q8s_1);
+            p16_2 = _mm_sub_epi16(p16_2, q8s_2);
+            p16_3 = _mm_sub_epi16(p16_3, q8s_3);
+            p16_4 = _mm_sub_epi16(p16_4, q8s_4);
+            p16_5 = _mm_sub_epi16(p16_5, q8s_5);
+            p16_6 = _mm_sub_epi16(p16_6, q8s_6);
+            p16_7 = _mm_sub_epi16(p16_7, q8s_7);
+
+            const __m128i scale_0 = _mm_shuffle_epi8(scales, shuffle);
+            shuffle = _mm_add_epi8(shuffle, m2);
+            const __m128i scale_1 = _mm_shuffle_epi8(scales, shuffle);
+            shuffle = _mm_add_epi8(shuffle, m2);
+            const __m128i scale_2 = _mm_shuffle_epi8(scales, shuffle);
+            shuffle = _mm_add_epi8(shuffle, m2);
+            const __m128i scale_3 = _mm_shuffle_epi8(scales, shuffle);
+            shuffle = _mm_add_epi8(shuffle, m2);
+
+            p16_0 = _mm_madd_epi16(_mm_cvtepi8_epi16(scale_0), p16_0);
+            p16_1 = _mm_madd_epi16(_mm_cvtepi8_epi16(_mm_unpackhi_epi64(scale_0, scale_0)), p16_1);
+            p16_2 = _mm_madd_epi16(_mm_cvtepi8_epi16(scale_1), p16_2);
+            p16_3 = _mm_madd_epi16(_mm_cvtepi8_epi16(_mm_unpackhi_epi64(scale_1, scale_1)), p16_3);
+            p16_4 = _mm_madd_epi16(_mm_cvtepi8_epi16(scale_2), p16_4);
+            p16_5 = _mm_madd_epi16(_mm_cvtepi8_epi16(_mm_unpackhi_epi64(scale_2, scale_2)), p16_5);
+            p16_6 = _mm_madd_epi16(_mm_cvtepi8_epi16(scale_3), p16_6);
+            p16_7 = _mm_madd_epi16(_mm_cvtepi8_epi16(_mm_unpackhi_epi64(scale_3, scale_3)), p16_7);
+
+            sumi_0 = _mm_add_epi32(sumi_0, _mm_add_epi32(p16_0, p16_2));
+            sumi_1 = _mm_add_epi32(sumi_1, _mm_add_epi32(p16_1, p16_3));
+            sumi_0 = _mm_add_epi32(sumi_0, _mm_add_epi32(p16_4, p16_6));
+            sumi_1 = _mm_add_epi32(sumi_1, _mm_add_epi32(p16_5, p16_7));
+
+        }
+
+        __m256i sumi = MM256_SET_M128I(sumi_1, sumi_0);
+        acc = _mm256_add_ps(_mm256_mul_ps(_mm256_broadcast_ss(&d), _mm256_cvtepi32_ps(sumi)), acc);
+    }
+
+    *s = hsum_float_8(acc);
+
+#elif defined __riscv_v_intrinsic
+
+    float sumf = 0;
+    for (int i = 0; i < nb; ++i) {
+
+        const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
+
+        const uint8_t * restrict q6 = x[i].ql;
+        const uint8_t * restrict qh = x[i].qh;
+        const  int8_t * restrict q8 = y[i].qs;
+
+        const int8_t * restrict scale = x[i].scales;
+
+        size_t vl;
+
+        vint32m1_t vzero = __riscv_vmv_v_x_i32m1(0, 1);
+
+        int sum_t = 0;
+        int is = 0;
+
+        for (int j = 0; j < QK_K/128; ++j) {
+
+            vl = 32;
+
+            // load qh
+            vuint8m1_t qh_x = __riscv_vle8_v_u8m1(qh, vl);
+
+            // load Q6
+            vuint8m1_t q6_0 = __riscv_vle8_v_u8m1(q6, vl);
+            vuint8m1_t q6_1 = __riscv_vle8_v_u8m1(q6+32, vl);
+
+            vuint8m1_t q6a_0 = __riscv_vand_vx_u8m1(q6_0, 0x0F, vl);
+            vuint8m1_t q6a_1 = __riscv_vand_vx_u8m1(q6_1, 0x0F, vl);
+            vuint8m1_t q6s_0 = __riscv_vsrl_vx_u8m1(q6_0, 0x04, vl);
+            vuint8m1_t q6s_1 = __riscv_vsrl_vx_u8m1(q6_1, 0x04, vl);
+
+            vuint8m1_t qh_0 = __riscv_vand_vx_u8m1(qh_x, 0x03, vl);
+            vuint8m1_t qh_1 = __riscv_vand_vx_u8m1(__riscv_vsrl_vx_u8m1(qh_x, 0x2, vl), 0x03 , vl);
+            vuint8m1_t qh_2 = __riscv_vand_vx_u8m1(__riscv_vsrl_vx_u8m1(qh_x, 0x4, vl), 0x03 , vl);
+            vuint8m1_t qh_3 = __riscv_vand_vx_u8m1(__riscv_vsrl_vx_u8m1(qh_x, 0x6, vl), 0x03 , vl);
+
+            vuint8m1_t qhi_0 = __riscv_vor_vv_u8m1(q6a_0, __riscv_vsll_vx_u8m1(qh_0, 0x04, vl), vl);
+            vuint8m1_t qhi_1 = __riscv_vor_vv_u8m1(q6a_1, __riscv_vsll_vx_u8m1(qh_1, 0x04, vl), vl);
+            vuint8m1_t qhi_2 = __riscv_vor_vv_u8m1(q6s_0, __riscv_vsll_vx_u8m1(qh_2, 0x04, vl), vl);
+            vuint8m1_t qhi_3 = __riscv_vor_vv_u8m1(q6s_1, __riscv_vsll_vx_u8m1(qh_3, 0x04, vl), vl);
+
+            vint8m1_t a_0 = __riscv_vsub_vx_i8m1(__riscv_vreinterpret_v_u8m1_i8m1(qhi_0), 32, vl);
+            vint8m1_t a_1 = __riscv_vsub_vx_i8m1(__riscv_vreinterpret_v_u8m1_i8m1(qhi_1), 32, vl);
+            vint8m1_t a_2 = __riscv_vsub_vx_i8m1(__riscv_vreinterpret_v_u8m1_i8m1(qhi_2), 32, vl);
+            vint8m1_t a_3 = __riscv_vsub_vx_i8m1(__riscv_vreinterpret_v_u8m1_i8m1(qhi_3), 32, vl);
+
+            // load Q8 and take product
+            vint16m2_t va_q_0 = __riscv_vwmul_vv_i16m2(a_0, __riscv_vle8_v_i8m1(q8, vl), vl);
+            vint16m2_t va_q_1 = __riscv_vwmul_vv_i16m2(a_1, __riscv_vle8_v_i8m1(q8+32, vl), vl);
+            vint16m2_t va_q_2 = __riscv_vwmul_vv_i16m2(a_2, __riscv_vle8_v_i8m1(q8+64, vl), vl);
+            vint16m2_t va_q_3 = __riscv_vwmul_vv_i16m2(a_3, __riscv_vle8_v_i8m1(q8+96, vl), vl);
+
+            vl = 16;
+
+            vint32m2_t vaux_0 = __riscv_vwmul_vx_i32m2(__riscv_vget_v_i16m2_i16m1(va_q_0, 0), scale[is+0], vl);
+            vint32m2_t vaux_1 = __riscv_vwmul_vx_i32m2(__riscv_vget_v_i16m2_i16m1(va_q_0, 1), scale[is+1], vl);
+            vint32m2_t vaux_2 = __riscv_vwmul_vx_i32m2(__riscv_vget_v_i16m2_i16m1(va_q_1, 0), scale[is+2], vl);
+            vint32m2_t vaux_3 = __riscv_vwmul_vx_i32m2(__riscv_vget_v_i16m2_i16m1(va_q_1, 1), scale[is+3], vl);
+            vint32m2_t vaux_4 = __riscv_vwmul_vx_i32m2(__riscv_vget_v_i16m2_i16m1(va_q_2, 0), scale[is+4], vl);
+            vint32m2_t vaux_5 = __riscv_vwmul_vx_i32m2(__riscv_vget_v_i16m2_i16m1(va_q_2, 1), scale[is+5], vl);
+            vint32m2_t vaux_6 = __riscv_vwmul_vx_i32m2(__riscv_vget_v_i16m2_i16m1(va_q_3, 0), scale[is+6], vl);
+            vint32m2_t vaux_7 = __riscv_vwmul_vx_i32m2(__riscv_vget_v_i16m2_i16m1(va_q_3, 1), scale[is+7], vl);
+
+            vint32m1_t isum0 = __riscv_vredsum_vs_i32m2_i32m1(__riscv_vadd_vv_i32m2(vaux_0, vaux_1, vl), vzero, vl);
+            vint32m1_t isum1 = __riscv_vredsum_vs_i32m2_i32m1(__riscv_vadd_vv_i32m2(vaux_2, vaux_3, vl), isum0, vl);
+            vint32m1_t isum2 = __riscv_vredsum_vs_i32m2_i32m1(__riscv_vadd_vv_i32m2(vaux_4, vaux_5, vl), isum1, vl);
+            vint32m1_t isum3 = __riscv_vredsum_vs_i32m2_i32m1(__riscv_vadd_vv_i32m2(vaux_6, vaux_7, vl), isum2, vl);
+
+            sum_t += __riscv_vmv_x_s_i32m1_i32(isum3);
+
+            q6 += 64;   qh += 32;   q8 += 128;   is=8;
+
+        }
+
+        sumf += d * sum_t;
+
+    }
+
+    *s = sumf;
+
+#elif defined(__POWER9_VECTOR__)
+    const vector signed char lowMask = vec_splats((signed char)0xF);
+    const vector int v0 = vec_splats((int32_t)0);
+    const vector unsigned char v2 = vec_splats((unsigned char)0x2);
+    const vector unsigned char v3 = vec_splats((unsigned char)0x3);
+    const vector unsigned char v4 = vec_splats((unsigned char)0x4);
+    const vector unsigned char v6 = vec_splats((unsigned char)0x6);
+    const vector signed char off = vec_splats((signed char)0x20);
+
+    vector float vsumf0 = vec_splats(0.0f);
+    vector float vsumf1 = vec_splats(0.0f);
+    vector float vsumf2 = vec_splats(0.0f);
+    vector float vsumf3 = vec_splats(0.0f);
+
+    for (int i = 0; i < nb; ++i) {
+        vector float vxd = vec_splats(GGML_FP16_TO_FP32(x[i].d));
+        vector float vyd = vec_splats(y[i].d);
+        vector float vd = vec_mul(vxd, vyd);
+
+        vector signed int vsumi0 = v0;
+        vector signed int vsumi1 = v0;
+        vector signed int vsumi2 = v0;
+        vector signed int vsumi3 = v0;
+        vector signed int vsumi4 = v0;
+        vector signed int vsumi5 = v0;
+        vector signed int vsumi6 = v0;
+        vector signed int vsumi7 = v0;
+
+        const uint8_t * restrict q6 = x[i].ql;
+        const uint8_t * restrict qh = x[i].qh;
+        const int8_t  * restrict qs = x[i].scales;
+        const int8_t  * restrict q8 = y[i].qs;
+
+        for (int j = 0; j < QK_K/128; ++j) {
+            __builtin_prefetch(q6, 0, 0);
+            __builtin_prefetch(qh, 0, 0);
+            __builtin_prefetch(q8, 0, 0);
+
+            vector signed char qxs0 = (vector signed char)vec_xl( 0, q6);
+            vector signed char qxs1 = (vector signed char)vec_xl(16, q6);
+            vector signed char qxs2 = (vector signed char)vec_xl(32, q6);
+            vector signed char qxs3 = (vector signed char)vec_xl(48, q6);
+            q6 += 64;
+
+            vector signed char qxs00 = vec_and(qxs0, lowMask);
+            vector signed char qxs01 = vec_sr(qxs0, v4);
+            vector signed char qxs10 = vec_and(qxs1, lowMask);
+            vector signed char qxs11 = vec_sr(qxs1, v4);
+            vector signed char qxs20 = vec_and(qxs2, lowMask);
+            vector signed char qxs21 = vec_sr(qxs2, v4);
+            vector signed char qxs30 = vec_and(qxs3, lowMask);
+            vector signed char qxs31 = vec_sr(qxs3, v4);
+
+            vector signed char qxhs0 = (vector signed char)vec_xl( 0, qh);
+            vector signed char qxhs1 = (vector signed char)vec_xl(16, qh);
+            qh += 32;
+
+            vector signed char qxh00 = vec_sl(vec_and((vector signed char)v3, qxhs0), v4);
+            vector signed char qxh01 = vec_sl(vec_and((vector signed char)v3, vec_sr(qxhs0, v4)), v4);
+            vector signed char qxh10 = vec_sl(vec_and((vector signed char)v3, qxhs1), v4);
+            vector signed char qxh11 = vec_sl(vec_and((vector signed char)v3, vec_sr(qxhs1, v4)), v4);
+            vector signed char qxh20 = vec_sl(vec_and((vector signed char)v3, vec_sr(qxhs0, v2)), v4);
+            vector signed char qxh21 = vec_sl(vec_and((vector signed char)v3, vec_sr(qxhs0, v6)), v4);
+            vector signed char qxh30 = vec_sl(vec_and((vector signed char)v3, vec_sr(qxhs1, v2)), v4);
+            vector signed char qxh31 = vec_sl(vec_and((vector signed char)v3, vec_sr(qxhs1, v6)), v4);
+
+            vector signed char q6x00 = vec_sub(vec_or(qxh00, qxs00), off);
+            vector signed char q6x01 = vec_sub(vec_or(qxh01, qxs01), off);
+            vector signed char q6x10 = vec_sub(vec_or(qxh10, qxs10), off);
+            vector signed char q6x11 = vec_sub(vec_or(qxh11, qxs11), off);
+            vector signed char q6x20 = vec_sub(vec_or(qxh20, qxs20), off);
+            vector signed char q6x21 = vec_sub(vec_or(qxh21, qxs21), off);
+            vector signed char q6x30 = vec_sub(vec_or(qxh30, qxs30), off);
+            vector signed char q6x31 = vec_sub(vec_or(qxh31, qxs31), off);
+
+            vector signed char q8y00 = vec_xl(  0, q8);
+            vector signed char q8y10 = vec_xl( 16, q8);
+            vector signed char q8y20 = vec_xl( 32, q8);
+            vector signed char q8y30 = vec_xl( 48, q8);
+            vector signed char q8y01 = vec_xl( 64, q8);
+            vector signed char q8y11 = vec_xl( 80, q8);
+            vector signed char q8y21 = vec_xl( 96, q8);
+            vector signed char q8y31 = vec_xl(112, q8);
+            q8 += 128;
+
+            vector signed short qv00 = vec_add(vec_mule(q6x00, q8y00), vec_mulo(q6x00, q8y00));
+            vector signed short qv10 = vec_add(vec_mule(q6x10, q8y10), vec_mulo(q6x10, q8y10));
+            vector signed short qv20 = vec_add(vec_mule(q6x20, q8y20), vec_mulo(q6x20, q8y20));
+            vector signed short qv30 = vec_add(vec_mule(q6x30, q8y30), vec_mulo(q6x30, q8y30));
+            vector signed short qv01 = vec_add(vec_mule(q6x01, q8y01), vec_mulo(q6x01, q8y01));
+            vector signed short qv11 = vec_add(vec_mule(q6x11, q8y11), vec_mulo(q6x11, q8y11));
+            vector signed short qv21 = vec_add(vec_mule(q6x21, q8y21), vec_mulo(q6x21, q8y21));
+            vector signed short qv31 = vec_add(vec_mule(q6x31, q8y31), vec_mulo(q6x31, q8y31));
+
+            vector signed short vscales = vec_unpackh(vec_xl_len(qs, 8));
+            qs += 8;
+
+            vector signed short vs0 = vec_splat(vscales, 0);
+            vector signed short vs1 = vec_splat(vscales, 1);
+            vector signed short vs2 = vec_splat(vscales, 2);
+            vector signed short vs3 = vec_splat(vscales, 3);
+            vector signed short vs4 = vec_splat(vscales, 4);
+            vector signed short vs5 = vec_splat(vscales, 5);
+            vector signed short vs6 = vec_splat(vscales, 6);
+            vector signed short vs7 = vec_splat(vscales, 7);
+
+            vsumi0 = vec_msum(qv00, vs0, vsumi0);
+            vsumi1 = vec_msum(qv01, vs4, vsumi1);
+            vsumi2 = vec_msum(qv10, vs1, vsumi2);
+            vsumi3 = vec_msum(qv11, vs5, vsumi3);
+            vsumi4 = vec_msum(qv20, vs2, vsumi4);
+            vsumi5 = vec_msum(qv21, vs6, vsumi5);
+            vsumi6 = vec_msum(qv30, vs3, vsumi6);
+            vsumi7 = vec_msum(qv31, vs7, vsumi7);
+        }
+
+        vsumi0 = vec_add(vsumi0, vsumi4);
+        vsumi1 = vec_add(vsumi1, vsumi5);
+        vsumi2 = vec_add(vsumi2, vsumi6);
+        vsumi3 = vec_add(vsumi3, vsumi7);
+
+        vsumf0 = vec_madd(vec_ctf(vsumi0, 0), vd, vsumf0);
+        vsumf1 = vec_madd(vec_ctf(vsumi1, 0), vd, vsumf1);
+        vsumf2 = vec_madd(vec_ctf(vsumi2, 0), vd, vsumf2);
+        vsumf3 = vec_madd(vec_ctf(vsumi3, 0), vd, vsumf3);
+    }
+
+    vsumf0 = vec_add(vsumf0, vsumf2);
+    vsumf1 = vec_add(vsumf1, vsumf3);
+
+    vsumf0 = vec_add(vsumf0, vsumf1);
+
+    vsumf0 = vec_add(vsumf0, vec_sld(vsumf0, vsumf0, 4));
+    vsumf0 = vec_add(vsumf0, vec_sld(vsumf0, vsumf0, 8));
+
+    *s = vec_extract(vsumf0, 0);
+
+#elif defined __loongarch_asx
+
+    const __m256i m4 = __lasx_xvreplgr2vr_b(0xF);
+    const __m256i m2 = __lasx_xvreplgr2vr_b(3);
+    const __m256i m32s = __lasx_xvreplgr2vr_b(32);
+
+    __m256 acc = (__m256)__lasx_xvldi(0);
+
+    for (int i = 0; i < nb; ++i) {
+
+        const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d);
+
+        const uint8_t * restrict q4 = x[i].ql;
+        const uint8_t * restrict qh = x[i].qh;
+        const int8_t  * restrict q8 = y[i].qs;
+
+        const __m128i scales = __lsx_vld((const __m128i*)x[i].scales, 0);
+
+        __m256i sumi = __lasx_xvldi(0);
+
+        int is = 0;
+
+        for (int j = 0; j < QK_K/128; ++j) {
+
+            const __m128i scale_0 = lsx_shuffle_b(scales, get_scale_shuffle(is + 0));
+            const __m128i scale_1 = lsx_shuffle_b(scales, get_scale_shuffle(is + 1));
+            const __m128i scale_2 = lsx_shuffle_b(scales, get_scale_shuffle(is + 2));
+            const __m128i scale_3 = lsx_shuffle_b(scales, get_scale_shuffle(is + 3));
+            is += 4;
+
+            const __m256i q4bits1 = __lasx_xvld((const __m256i*)q4, 0); q4 += 32;
+            const __m256i q4bits2 = __lasx_xvld((const __m256i*)q4, 0); q4 += 32;
+            const __m256i q4bitsH = __lasx_xvld((const __m256i*)qh, 0); qh += 32;
+
+            const __m256i q4h_0 = __lasx_xvslli_h(__lasx_xvand_v(q4bitsH, m2), 4);
+            const __m256i q4h_1 = __lasx_xvslli_h(__lasx_xvand_v(__lasx_xvsrli_h(q4bitsH, 2), m2), 4);
+            const __m256i q4h_2 = __lasx_xvslli_h(__lasx_xvand_v(__lasx_xvsrli_h(q4bitsH, 4), m2), 4);
+            const __m256i q4h_3 = __lasx_xvslli_h(__lasx_xvand_v(__lasx_xvsrli_h(q4bitsH, 6), m2), 4);
+
+            const __m256i q4_0 = __lasx_xvor_v(__lasx_xvand_v(q4bits1, m4), q4h_0);
+            const __m256i q4_1 = __lasx_xvor_v(__lasx_xvand_v(q4bits2, m4), q4h_1);
+            const __m256i q4_2 = __lasx_xvor_v(__lasx_xvand_v(__lasx_xvsrli_h(q4bits1, 4), m4), q4h_2);
+            const __m256i q4_3 = __lasx_xvor_v(__lasx_xvand_v(__lasx_xvsrli_h(q4bits2, 4), m4), q4h_3);
+
+            const __m256i q8_0 = __lasx_xvld((const __m256i*)q8, 0); q8 += 32;
+            const __m256i q8_1 = __lasx_xvld((const __m256i*)q8, 0); q8 += 32;
+            const __m256i q8_2 = __lasx_xvld((const __m256i*)q8, 0); q8 += 32;
+            const __m256i q8_3 = __lasx_xvld((const __m256i*)q8, 0); q8 += 32;
+
+            __m256i q8s_0 = lasx_maddubs_h(m32s, q8_0);
+            __m256i q8s_1 = lasx_maddubs_h(m32s, q8_1);
+            __m256i q8s_2 = lasx_maddubs_h(m32s, q8_2);
+            __m256i q8s_3 = lasx_maddubs_h(m32s, q8_3);
+
+            __m256i p16_0 = lasx_maddubs_h(q4_0, q8_0);
+            __m256i p16_1 = lasx_maddubs_h(q4_1, q8_1);
+            __m256i p16_2 = lasx_maddubs_h(q4_2, q8_2);
+            __m256i p16_3 = lasx_maddubs_h(q4_3, q8_3);
+
+            p16_0 = __lasx_xvsub_h(p16_0, q8s_0);
+            p16_1 = __lasx_xvsub_h(p16_1, q8s_1);
+            p16_2 = __lasx_xvsub_h(p16_2, q8s_2);
+            p16_3 = __lasx_xvsub_h(p16_3, q8s_3);
+
+            p16_0 = lasx_madd_h(lasx_ext8_16(scale_0), p16_0);
+            p16_1 = lasx_madd_h(lasx_ext8_16(scale_1), p16_1);
+            p16_2 = lasx_madd_h(lasx_ext8_16(scale_2), p16_2);
+            p16_3 = lasx_madd_h(lasx_ext8_16(scale_3), p16_3);
+
+            sumi = __lasx_xvadd_w(sumi, __lasx_xvadd_w(p16_0, p16_1));
+            sumi = __lasx_xvadd_w(sumi, __lasx_xvadd_w(p16_2, p16_3));
+        }
+
+        acc = __lasx_xvfmadd_s((__m256)__lasx_xvreplfr2vr_s(d), __lasx_xvffint_s_w(sumi), acc);
+    }
+
+    *s = hsum_float_8(acc);
+
+#else
+
+    int8_t  aux8[QK_K];
+    int16_t aux16[8];
+    float   sums [8];
+    int32_t aux32[8];
+    memset(sums, 0, 8*sizeof(float));
+
+    float sumf = 0;
+    for (int i = 0; i < nb; ++i) {
+        const uint8_t * restrict q4 = x[i].ql;
+        const uint8_t * restrict qh = x[i].qh;
+        const  int8_t * restrict q8 = y[i].qs;
+        memset(aux32, 0, 8*sizeof(int32_t));
+        int8_t * restrict a = aux8;
+        for (int j = 0; j < QK_K; j += 128) {
+            for (int l = 0; l < 32; ++l) {
+                a[l +  0] = (int8_t)((q4[l +  0] & 0xF) | (((qh[l] >> 0) & 3) << 4)) - 32;
+                a[l + 32] = (int8_t)((q4[l + 32] & 0xF) | (((qh[l] >> 2) & 3) << 4)) - 32;
+                a[l + 64] = (int8_t)((q4[l +  0] >>  4) | (((qh[l] >> 4) & 3) << 4)) - 32;
+                a[l + 96] = (int8_t)((q4[l + 32] >>  4) | (((qh[l] >> 6) & 3) << 4)) - 32;
+            }
+            a  += 128;
+            q4 += 64;
+            qh += 32;
+        }
+        a = aux8;
+        int is = 0;
+        for (int j = 0; j < QK_K/16; ++j) {
+            int scale = x[i].scales[is++];
+            for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
+            for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
+            q8 += 8; a += 8;
+            for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
+            for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
+            q8 += 8; a += 8;
+        }
+        const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
+        for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l];
+    }
+    for (int l = 0; l < 8; ++l) sumf += sums[l];
+    *s = sumf;
+#endif
+}
+
+#if defined (__AVX__) || defined (__AVX2__) || defined (__ARM_NEON) || defined (__POWER9_VECTOR__) || defined(__loongarch_asx)
+static const int8_t keven_signs_q2xs[1024] = {
+     1,  1,  1,  1,  1,  1,  1,  1, -1,  1,  1,  1,  1,  1,  1, -1,  1, -1,  1,  1,  1,  1,  1, -1, -1, -1,  1,  1,  1,  1,  1,  1,
+     1,  1, -1,  1,  1,  1,  1, -1, -1,  1, -1,  1,  1,  1,  1,  1,  1, -1, -1,  1,  1,  1,  1,  1, -1, -1, -1,  1,  1,  1,  1, -1,
+     1,  1,  1, -1,  1,  1,  1, -1, -1,  1,  1, -1,  1,  1,  1,  1,  1, -1,  1, -1,  1,  1,  1,  1, -1, -1,  1, -1,  1,  1,  1, -1,
+     1,  1, -1, -1,  1,  1,  1,  1, -1,  1, -1, -1,  1,  1,  1, -1,  1, -1, -1, -1,  1,  1,  1, -1, -1, -1, -1, -1,  1,  1,  1,  1,
+     1,  1,  1,  1, -1,  1,  1, -1, -1,  1,  1,  1, -1,  1,  1,  1,  1, -1,  1,  1, -1,  1,  1,  1, -1, -1,  1,  1, -1,  1,  1, -1,
+     1,  1, -1,  1, -1,  1,  1,  1, -1,  1, -1,  1, -1,  1,  1, -1,  1, -1, -1,  1, -1,  1,  1, -1, -1, -1, -1,  1, -1,  1,  1,  1,
+     1,  1,  1, -1, -1,  1,  1,  1, -1,  1,  1, -1, -1,  1,  1, -1,  1, -1,  1, -1, -1,  1,  1, -1, -1, -1,  1, -1, -1,  1,  1,  1,
+     1,  1, -1, -1, -1,  1,  1, -1, -1,  1, -1, -1, -1,  1,  1,  1,  1, -1, -1, -1, -1,  1,  1,  1, -1, -1, -1, -1, -1,  1,  1, -1,
+     1,  1,  1,  1,  1, -1,  1, -1, -1,  1,  1,  1,  1, -1,  1,  1,  1, -1,  1,  1,  1, -1,  1,  1, -1, -1,  1,  1,  1, -1,  1, -1,
+     1,  1, -1,  1,  1, -1,  1,  1, -1,  1, -1,  1,  1, -1,  1, -1,  1, -1, -1,  1,  1, -1,  1, -1, -1, -1, -1,  1,  1, -1,  1,  1,
+     1,  1,  1, -1,  1, -1,  1,  1, -1,  1,  1, -1,  1, -1,  1, -1,  1, -1,  1, -1,  1, -1,  1, -1, -1, -1,  1, -1,  1, -1,  1,  1,
+     1,  1, -1, -1,  1, -1,  1, -1, -1,  1, -1, -1,  1, -1,  1,  1,  1, -1, -1, -1,  1, -1,  1,  1, -1, -1, -1, -1,  1, -1,  1, -1,
+     1,  1,  1,  1, -1, -1,  1,  1, -1,  1,  1,  1, -1, -1,  1, -1,  1, -1,  1,  1, -1, -1,  1, -1, -1, -1,  1,  1, -1, -1,  1,  1,
+     1,  1, -1,  1, -1, -1,  1, -1, -1,  1, -1,  1, -1, -1,  1,  1,  1, -1, -1,  1, -1, -1,  1,  1, -1, -1, -1,  1, -1, -1,  1, -1,
+     1,  1,  1, -1, -1, -1,  1, -1, -1,  1,  1, -1, -1, -1,  1,  1,  1, -1,  1, -1, -1, -1,  1,  1, -1, -1,  1, -1, -1, -1,  1, -1,
+     1,  1, -1, -1, -1, -1,  1,  1, -1,  1, -1, -1, -1, -1,  1, -1,  1, -1, -1, -1, -1, -1,  1, -1, -1, -1, -1, -1, -1, -1,  1,  1,
+     1,  1,  1,  1,  1,  1, -1, -1, -1,  1,  1,  1,  1,  1, -1,  1,  1, -1,  1,  1,  1,  1, -1,  1, -1, -1,  1,  1,  1,  1, -1, -1,
+     1,  1, -1,  1,  1,  1, -1,  1, -1,  1, -1,  1,  1,  1, -1, -1,  1, -1, -1,  1,  1,  1, -1, -1, -1, -1, -1,  1,  1,  1, -1,  1,
+     1,  1,  1, -1,  1,  1, -1,  1, -1,  1,  1, -1,  1,  1, -1, -1,  1, -1,  1, -1,  1,  1, -1, -1, -1, -1,  1, -1,  1,  1, -1,  1,
+     1,  1, -1, -1,  1,  1, -1, -1, -1,  1, -1, -1,  1,  1, -1,  1,  1, -1, -1, -1,  1,  1, -1,  1, -1, -1, -1, -1,  1,  1, -1, -1,
+     1,  1,  1,  1, -1,  1, -1,  1, -1,  1,  1,  1, -1,  1, -1, -1,  1, -1,  1,  1, -1,  1, -1, -1, -1, -1,  1,  1, -1,  1, -1,  1,
+     1,  1, -1,  1, -1,  1, -1, -1, -1,  1, -1,  1, -1,  1, -1,  1,  1, -1, -1,  1, -1,  1, -1,  1, -1, -1, -1,  1, -1,  1, -1, -1,
+     1,  1,  1, -1, -1,  1, -1, -1, -1,  1,  1, -1, -1,  1, -1,  1,  1, -1,  1, -1, -1,  1, -1,  1, -1, -1,  1, -1, -1,  1, -1, -1,
+     1,  1, -1, -1, -1,  1, -1,  1, -1,  1, -1, -1, -1,  1, -1, -1,  1, -1, -1, -1, -1,  1, -1, -1, -1, -1, -1, -1, -1,  1, -1,  1,
+     1,  1,  1,  1,  1, -1, -1,  1, -1,  1,  1,  1,  1, -1, -1, -1,  1, -1,  1,  1,  1, -1, -1, -1, -1, -1,  1,  1,  1, -1, -1,  1,
+     1,  1, -1,  1,  1, -1, -1, -1, -1,  1, -1,  1,  1, -1, -1,  1,  1, -1, -1,  1,  1, -1, -1,  1, -1, -1, -1,  1,  1, -1, -1, -1,
+     1,  1,  1, -1,  1, -1, -1, -1, -1,  1,  1, -1,  1, -1, -1,  1,  1, -1,  1, -1,  1, -1, -1,  1, -1, -1,  1, -1,  1, -1, -1, -1,
+     1,  1, -1, -1,  1, -1, -1,  1, -1,  1, -1, -1,  1, -1, -1, -1,  1, -1, -1, -1,  1, -1, -1, -1, -1, -1, -1, -1,  1, -1, -1,  1,
+     1,  1,  1,  1, -1, -1, -1, -1, -1,  1,  1,  1, -1, -1, -1,  1,  1, -1,  1,  1, -1, -1, -1,  1, -1, -1,  1,  1, -1, -1, -1, -1,
+     1,  1, -1,  1, -1, -1, -1,  1, -1,  1, -1,  1, -1, -1, -1, -1,  1, -1, -1,  1, -1, -1, -1, -1, -1, -1, -1,  1, -1, -1, -1,  1,
+     1,  1,  1, -1, -1, -1, -1,  1, -1,  1,  1, -1, -1, -1, -1, -1,  1, -1,  1, -1, -1, -1, -1, -1, -1, -1,  1, -1, -1, -1, -1,  1,
+     1,  1, -1, -1, -1, -1, -1, -1, -1,  1, -1, -1, -1, -1, -1,  1,  1, -1, -1, -1, -1, -1, -1,  1, -1, -1, -1, -1, -1, -1, -1, -1,
+};
+#endif
+
+void ggml_vec_dot_iq2_xxs_q8_K(int n, float * restrict s, size_t bs, const void * restrict vx, size_t bx, const void * restrict vy, size_t by, int nrc) {
+    assert(n % QK_K == 0);
+    assert(nrc == 1);
+    UNUSED(nrc);
+    UNUSED(bx);
+    UNUSED(by);
+    UNUSED(bs);
+
+    const block_iq2_xxs * restrict x = vx;
+    const block_q8_K    * restrict y = vy;
+
+    const int nb = n / QK_K;
+
+#if defined(__ARM_NEON)
+
+    const uint64_t * signs64 = (const uint64_t *)keven_signs_q2xs;
+
+    uint32_t aux32[4];
+    const uint8_t * aux8 = (const uint8_t *)aux32;
+
+    ggml_int8x16x4_t q2u;
+    ggml_int8x16x4_t q2s;
+    ggml_int8x16x4_t q8b;
+
+    float sumf = 0;
+    for (int i = 0; i < nb; ++i) {
+        const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
+        const uint16_t * restrict q2 = x[i].qs;
+        const int8_t   * restrict q8 = y[i].qs;
+        float sumf1 = 0, sumf2 = 0;
+        for (int ib32 = 0; ib32 < QK_K/32; ib32 += 2) {
+            q8b = ggml_vld1q_s8_x4(q8); q8 += 64;
+            memcpy(aux32, q2, 4*sizeof(uint32_t)); q2 += 8;
+            q2u.val[0] = vcombine_s8(vld1_s8((const void *)(iq2xxs_grid + aux8[ 0])), vld1_s8((const void *)(iq2xxs_grid + aux8[ 1])));
+            q2u.val[1] = vcombine_s8(vld1_s8((const void *)(iq2xxs_grid + aux8[ 2])), vld1_s8((const void *)(iq2xxs_grid + aux8[ 3])));
+            q2u.val[2] = vcombine_s8(vld1_s8((const void *)(iq2xxs_grid + aux8[ 8])), vld1_s8((const void *)(iq2xxs_grid + aux8[ 9])));
+            q2u.val[3] = vcombine_s8(vld1_s8((const void *)(iq2xxs_grid + aux8[10])), vld1_s8((const void *)(iq2xxs_grid + aux8[11])));
+            q2s.val[0] = vcombine_s8(vld1_s8((const void *)(signs64 + ((aux32[1] >>  0) & 127))), vld1_s8((const void *)(signs64 + ((aux32[1] >>  7) & 127))));
+            q2s.val[1] = vcombine_s8(vld1_s8((const void *)(signs64 + ((aux32[1] >> 14) & 127))), vld1_s8((const void *)(signs64 + ((aux32[1] >> 21) & 127))));
+            q2s.val[2] = vcombine_s8(vld1_s8((const void *)(signs64 + ((aux32[3] >>  0) & 127))), vld1_s8((const void *)(signs64 + ((aux32[3] >>  7) & 127))));
+            q2s.val[3] = vcombine_s8(vld1_s8((const void *)(signs64 + ((aux32[3] >> 14) & 127))), vld1_s8((const void *)(signs64 + ((aux32[3] >> 21) & 127))));
+            q2u.val[0] = vmulq_s8(q2u.val[0], q2s.val[0]);
+            q2u.val[1] = vmulq_s8(q2u.val[1], q2s.val[1]);
+            q2u.val[2] = vmulq_s8(q2u.val[2], q2s.val[2]);
+            q2u.val[3] = vmulq_s8(q2u.val[3], q2s.val[3]);
+            const int32x4_t p1 = ggml_vdotq_s32(ggml_vdotq_s32(vdupq_n_s32(0), q2u.val[0], q8b.val[0]), q2u.val[1], q8b.val[1]);
+            const int32x4_t p2 = ggml_vdotq_s32(ggml_vdotq_s32(vdupq_n_s32(0), q2u.val[2], q8b.val[2]), q2u.val[3], q8b.val[3]);
+            sumf1 += vaddvq_s32(p1) * (0.5f + (aux32[1] >> 28));
+            sumf2 += vaddvq_s32(p2) * (0.5f + (aux32[3] >> 28));
+        }
+        sumf += d*(sumf1 + sumf2);
+    }
+    *s = 0.25f * sumf;
+
+#elif defined(__AVX2__)
+
+    const uint64_t * signs64 = (const uint64_t *)keven_signs_q2xs;
+
+    uint32_t aux32[4];
+    const uint8_t * aux8 = (const uint8_t *)aux32;
+
+    __m256 accumf = _mm256_setzero_ps();
+    for (int i = 0; i < nb; ++i) {
+        const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
+        const uint16_t * restrict q2 = x[i].qs;
+        const int8_t   * restrict q8 = y[i].qs;
+        __m256i sumi1 = _mm256_setzero_si256();
+        __m256i sumi2 = _mm256_setzero_si256();
+        for (int ib32 = 0; ib32 < QK_K/32; ib32 += 2) {
+            const __m256i q8_1 = _mm256_loadu_si256((const __m256i *)q8); q8 += 32;
+            const __m256i q8_2 = _mm256_loadu_si256((const __m256i *)q8); q8 += 32;
+            memcpy(aux32, q2, 4*sizeof(uint32_t)); q2 += 8;
+            const __m256i q2_1 = _mm256_set_epi64x(iq2xxs_grid[aux8[ 3]], iq2xxs_grid[aux8[ 2]], iq2xxs_grid[aux8[1]], iq2xxs_grid[aux8[0]]);
+            const __m256i q2_2 = _mm256_set_epi64x(iq2xxs_grid[aux8[11]], iq2xxs_grid[aux8[10]], iq2xxs_grid[aux8[9]], iq2xxs_grid[aux8[8]]);
+            const __m256i s2_1 = _mm256_set_epi64x(signs64[(aux32[1] >> 21) & 127], signs64[(aux32[1] >> 14) & 127],
+                                                   signs64[(aux32[1] >>  7) & 127], signs64[(aux32[1] >>  0) & 127]);
+            const __m256i s2_2 = _mm256_set_epi64x(signs64[(aux32[3] >> 21) & 127], signs64[(aux32[3] >> 14) & 127],
+                                                   signs64[(aux32[3] >>  7) & 127], signs64[(aux32[3] >>  0) & 127]);
+            const __m256i q8s_1 = _mm256_sign_epi8(q8_1, s2_1);
+            const __m256i q8s_2 = _mm256_sign_epi8(q8_2, s2_2);
+            const __m256i dot1  = _mm256_maddubs_epi16(q2_1, q8s_1);
+            const __m256i dot2  = _mm256_maddubs_epi16(q2_2, q8s_2);
+            const uint16_t ls1 = aux32[1] >> 28;
+            const uint16_t ls2 = aux32[3] >> 28;
+            const __m256i p1 = _mm256_madd_epi16(dot1, _mm256_set1_epi16(2*ls1+1));
+            const __m256i p2 = _mm256_madd_epi16(dot2, _mm256_set1_epi16(2*ls2+1));
+            sumi1 = _mm256_add_epi32(sumi1, p1);
+            sumi2 = _mm256_add_epi32(sumi2, p2);
+        }
+
+        accumf = _mm256_fmadd_ps(_mm256_set1_ps(d), _mm256_cvtepi32_ps(_mm256_add_epi32(sumi1, sumi2)), accumf);
+
+    }
+
+    *s = 0.125f * hsum_float_8(accumf);
+
+#elif defined(__AVX__)
+    const uint64_t * signs64 = (const uint64_t *)keven_signs_q2xs;
+
+    uint32_t aux32[4];
+    const uint8_t * aux8 = (const uint8_t *)aux32;
+
+    __m256 accumf = _mm256_setzero_ps();
+    for (int i = 0; i < nb; ++i) {
+        const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
+        const uint16_t * restrict q2 = x[i].qs;
+        const int8_t   * restrict q8 = y[i].qs;
+        __m128i sumi1_0 = _mm_setzero_si128();
+        __m128i sumi1_1 = _mm_setzero_si128();
+        __m128i sumi2_0 = _mm_setzero_si128();
+        __m128i sumi2_1 = _mm_setzero_si128();
+        for (int ib32 = 0; ib32 < QK_K/32; ib32 += 2) {
+            const __m128i q8_1_0 = _mm_loadu_si128((const __m128i *)q8); q8 += 16;
+            const __m128i q8_1_1 = _mm_loadu_si128((const __m128i *)q8); q8 += 16;
+            const __m128i q8_2_0 = _mm_loadu_si128((const __m128i *)q8); q8 += 16;
+            const __m128i q8_2_1 = _mm_loadu_si128((const __m128i *)q8); q8 += 16;
+            memcpy(aux32, q2, 4*sizeof(uint32_t)); q2 += 8;
+            const __m128i q2_1_0 = _mm_set_epi64x(iq2xxs_grid[aux8[1]], iq2xxs_grid[aux8[0]]);
+            const __m128i q2_1_1 = _mm_set_epi64x(iq2xxs_grid[aux8[3]], iq2xxs_grid[aux8[2]]);
+            const __m128i q2_2_0 = _mm_set_epi64x(iq2xxs_grid[aux8[9]], iq2xxs_grid[aux8[8]]);
+            const __m128i q2_2_1 = _mm_set_epi64x(iq2xxs_grid[aux8[11]], iq2xxs_grid[aux8[10]]);
+            const __m128i s2_1_0 = _mm_set_epi64x(signs64[(aux32[1] >>  7) & 127], signs64[(aux32[1] >>  0) & 127]);
+            const __m128i s2_1_1 = _mm_set_epi64x(signs64[(aux32[1] >> 21) & 127], signs64[(aux32[1] >> 14) & 127]);
+            const __m128i s2_2_0 = _mm_set_epi64x(signs64[(aux32[3] >>  7) & 127], signs64[(aux32[3] >>  0) & 127]);
+            const __m128i s2_2_1 = _mm_set_epi64x(signs64[(aux32[3] >> 21) & 127], signs64[(aux32[3] >> 14) & 127]);
+            const __m128i q8s_1_0 = _mm_sign_epi8(q8_1_0, s2_1_0);
+            const __m128i q8s_1_1 = _mm_sign_epi8(q8_1_1, s2_1_1);
+            const __m128i q8s_2_0 = _mm_sign_epi8(q8_2_0, s2_2_0);
+            const __m128i q8s_2_1 = _mm_sign_epi8(q8_2_1, s2_2_1);
+            const __m128i dot1_0  = _mm_maddubs_epi16(q2_1_0, q8s_1_0);
+            const __m128i dot1_1  = _mm_maddubs_epi16(q2_1_1, q8s_1_1);
+            const __m128i dot2_0  = _mm_maddubs_epi16(q2_2_0, q8s_2_0);
+            const __m128i dot2_1  = _mm_maddubs_epi16(q2_2_1, q8s_2_1);
+            const uint16_t ls1 = aux32[1] >> 28;
+            const uint16_t ls2 = aux32[3] >> 28;
+            const __m128i p1_0 = _mm_madd_epi16(dot1_0, _mm_set1_epi16(2*ls1+1));
+            const __m128i p1_1 = _mm_madd_epi16(dot1_1, _mm_set1_epi16(2*ls1+1));
+            const __m128i p2_0 = _mm_madd_epi16(dot2_0, _mm_set1_epi16(2*ls2+1));
+            const __m128i p2_1 = _mm_madd_epi16(dot2_1, _mm_set1_epi16(2*ls2+1));
+            sumi1_0 = _mm_add_epi32(sumi1_0, p1_0);
+            sumi1_1 = _mm_add_epi32(sumi1_1, p1_1);
+            sumi2_0 = _mm_add_epi32(sumi2_0, p2_0);
+            sumi2_1 = _mm_add_epi32(sumi2_1, p2_1);
+        }
+
+        accumf = _mm256_add_ps(_mm256_mul_ps(_mm256_set1_ps(d), _mm256_cvtepi32_ps(MM256_SET_M128I(_mm_add_epi32(sumi1_1, sumi2_1), _mm_add_epi32(sumi1_0, sumi2_0)))), accumf);
+
+    }
+
+    *s = 0.125f * hsum_float_8(accumf);
+
+#elif defined(__POWER9_VECTOR__)
+    const vector int v0 = vec_splats((int32_t)0);
+    vector float vsumf0 = vec_splats(0.0f);
+    vector float vsumf1 = vec_splats(0.0f);
+    vector float vsumf2 = vec_splats(0.0f);
+    vector float vsumf3 = vec_splats(0.0f);
+
+    const uint64_t * signs64 = (const uint64_t *)keven_signs_q2xs;
+
+    for (int i = 0; i < nb; ++i) {
+        vector float vxd = vec_splats(GGML_FP16_TO_FP32(x[i].d));
+        vector float vyd = vec_splats(y[i].d);
+        vector float vd = vec_mul(vxd, vyd);
+
+        vector signed int vsumi0 = v0;
+        vector signed int vsumi1 = v0;
+        vector signed int vsumi2 = v0;
+        vector signed int vsumi3 = v0;
+
+        const uint16_t * restrict q2 = x[i].qs;
+        const int8_t  *  restrict q8 = y[i].qs;
+
+        for (int j = 0; j < QK_K/32; j += 2) {
+            __builtin_prefetch(q2, 0, 1);
+            __builtin_prefetch(q8, 0, 1);
+
+            uint32_t aux32[4];
+            const uint8_t * aux8 = (const uint8_t *)aux32;
+
+            memcpy(aux32, q2, 4*sizeof(uint32_t));
+            q2 += 8;
+
+            vector signed long long aux64x2_0 = {*(const int64_t *)(iq2xxs_grid + aux8[ 0]), *(const int64_t *)(iq2xxs_grid + aux8[ 1])};
+            vector signed long long aux64x2_1 = {*(const int64_t *)(iq2xxs_grid + aux8[ 2]), *(const int64_t *)(iq2xxs_grid + aux8[ 3])};
+            vector signed long long aux64x2_2 = {*(const int64_t *)(iq2xxs_grid + aux8[ 8]), *(const int64_t *)(iq2xxs_grid + aux8[ 9])};
+            vector signed long long aux64x2_3 = {*(const int64_t *)(iq2xxs_grid + aux8[10]), *(const int64_t *)(iq2xxs_grid + aux8[11])};
+
+            vector signed long long vsigns0 = {*(const int64_t *)(signs64 + ((aux32[1] >>  0) & 127)), *(const int64_t *)(signs64 + ((aux32[1] >>  7) & 127))};
+            vector signed long long vsigns1 = {*(const int64_t *)(signs64 + ((aux32[1] >> 14) & 127)), *(const int64_t *)(signs64 + ((aux32[1] >> 21) & 127))};
+            vector signed long long vsigns2 = {*(const int64_t *)(signs64 + ((aux32[3] >>  0) & 127)), *(const int64_t *)(signs64 + ((aux32[3] >>  7) & 127))};
+            vector signed long long vsigns3 = {*(const int64_t *)(signs64 + ((aux32[3] >> 14) & 127)), *(const int64_t *)(signs64 + ((aux32[3] >> 21) & 127))};
+
+            vector signed char q2x0 = (vector signed char)vec_mul((vector signed char)vsigns0, (vector signed char)aux64x2_0);
+            vector signed char q2x1 = (vector signed char)vec_mul((vector signed char)vsigns1, (vector signed char)aux64x2_1);
+            vector signed char q2x2 = (vector signed char)vec_mul((vector signed char)vsigns2, (vector signed char)aux64x2_2);
+            vector signed char q2x3 = (vector signed char)vec_mul((vector signed char)vsigns3, (vector signed char)aux64x2_3);
+
+            vector signed char q8y0 = vec_xl( 0, q8);
+            vector signed char q8y1 = vec_xl(16, q8);
+            vector signed char q8y2 = vec_xl(32, q8);
+            vector signed char q8y3 = vec_xl(48, q8);
+            q8 += 64;
+
+            vector signed short qv0 = vec_add(vec_mule(q2x0, q8y0), vec_mulo(q2x0, q8y0));
+            vector signed short qv1 = vec_add(vec_mule(q2x1, q8y1), vec_mulo(q2x1, q8y1));
+            vector signed short qv2 = vec_add(vec_mule(q2x2, q8y2), vec_mulo(q2x2, q8y2));
+            vector signed short qv3 = vec_add(vec_mule(q2x3, q8y3), vec_mulo(q2x3, q8y3));
+
+            const uint16_t ls0 = aux32[1] >> 28;
+            const uint16_t ls1 = aux32[3] >> 28;
+
+            vector signed short vscales01 = vec_splats((int16_t)(2*ls0+1));
+            vector signed short vscales23 = vec_splats((int16_t)(2*ls1+1));
+
+            vsumi0 = vec_msum(qv0, vscales01, vsumi0);
+            vsumi1 = vec_msum(qv1, vscales01, vsumi1);
+            vsumi2 = vec_msum(qv2, vscales23, vsumi2);
+            vsumi3 = vec_msum(qv3, vscales23, vsumi3);
+        }
+
+        vsumf0 = vec_madd(vec_ctf(vsumi0, 0), vd, vsumf0);
+        vsumf1 = vec_madd(vec_ctf(vsumi1, 0), vd, vsumf1);
+        vsumf2 = vec_madd(vec_ctf(vsumi2, 0), vd, vsumf2);
+        vsumf3 = vec_madd(vec_ctf(vsumi3, 0), vd, vsumf3);
+    }
+
+    vsumf0 = vec_add(vsumf0, vsumf2);
+    vsumf1 = vec_add(vsumf1, vsumf3);
+
+    vsumf0 = vec_add(vsumf0, vsumf1);
+
+    vsumf0 = vec_add(vsumf0, vec_sld(vsumf0, vsumf0, 4));
+    vsumf0 = vec_add(vsumf0, vec_sld(vsumf0, vsumf0, 8));
+
+    *s = 0.125f * vec_extract(vsumf0, 0);
+
+#elif defined(__loongarch_asx)
+
+    const uint64_t * signs64 = (const uint64_t *)keven_signs_q2xs;
+
+    uint32_t aux32[4];
+    const uint8_t * aux8 = (const uint8_t *)aux32;
+
+    __m256 accumf = (__m256)__lasx_xvldi(0);
+    for (int i = 0; i < nb; ++i) {
+        const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
+        const uint16_t * restrict q2 = x[i].qs;
+        const int8_t   * restrict q8 = y[i].qs;
+        __m256i sumi1 = __lasx_xvldi(0);
+        __m256i sumi2 = __lasx_xvldi(0);
+        for (int ib32 = 0; ib32 < QK_K/32; ib32 += 2) {
+            const __m256i q8_1 = __lasx_xvld((const __m256i *)q8, 0); q8 += 32;
+            const __m256i q8_2 = __lasx_xvld((const __m256i *)q8, 0); q8 += 32;
+            memcpy(aux32, q2, 4*sizeof(uint32_t)); q2 += 8;
+
+            const __m256i q2_1 = lasx_set_d(iq2xxs_grid[aux8[ 3]], iq2xxs_grid[aux8[ 2]], iq2xxs_grid[aux8[1]], iq2xxs_grid[aux8[0]]);
+            const __m256i q2_2 = lasx_set_d(iq2xxs_grid[aux8[11]], iq2xxs_grid[aux8[10]], iq2xxs_grid[aux8[9]], iq2xxs_grid[aux8[8]]);
+            const __m256i s2_1 = lasx_set_d(signs64[(aux32[1] >> 21) & 127], signs64[(aux32[1] >> 14) & 127],
+                                                   signs64[(aux32[1] >>  7) & 127], signs64[(aux32[1] >>  0) & 127]);
+            const __m256i s2_2 = lasx_set_d(signs64[(aux32[3] >> 21) & 127], signs64[(aux32[3] >> 14) & 127],
+                                                   signs64[(aux32[3] >>  7) & 127], signs64[(aux32[3] >>  0) & 127]);
+            const __m256i q8s_1 = __lasx_xvsigncov_b(s2_1, q8_1);
+            const __m256i q8s_2 = __lasx_xvsigncov_b(s2_2, q8_2);
+            const __m256i dot1  = lasx_maddubs_h(q2_1, q8s_1);
+            const __m256i dot2  = lasx_maddubs_h(q2_2, q8s_2);
+            const uint16_t ls1 = aux32[1] >> 28;
+            const uint16_t ls2 = aux32[3] >> 28;
+            const __m256i p1 = lasx_madd_h(dot1, __lasx_xvreplgr2vr_h(2*ls1+1));
+            const __m256i p2 = lasx_madd_h(dot2, __lasx_xvreplgr2vr_h(2*ls2+1));
+            sumi1 = __lasx_xvadd_w(sumi1, p1);
+            sumi2 = __lasx_xvadd_w(sumi2, p2);
+        }
+
+        accumf = __lasx_xvfmadd_s(__lasx_xvreplfr2vr_s(d), __lasx_xvffint_s_w(__lasx_xvadd_w(sumi1, sumi2)), accumf);
+    }
+
+    *s = 0.125f * hsum_float_8(accumf);
+
+#else
+
+    uint32_t aux32[2];
+    const uint8_t * aux8 = (const uint8_t *)aux32;
+
+    float sumf = 0.f;
+    for (int i = 0; i < nb; ++i) {
+        const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
+        const uint16_t * restrict q2 = x[i].qs;
+        const int8_t   * restrict q8 = y[i].qs;
+        int32_t bsum = 0;
+        for (int ib32 = 0; ib32 < QK_K/32; ++ib32) {
+            memcpy(aux32, q2, 2*sizeof(uint32_t));
+            q2 += 4;
+            const uint32_t ls = 2*(aux32[1] >> 28) + 1;
+            int32_t sumi = 0;
+            for (int l = 0; l < 4; ++l) {
+                const uint8_t * grid = (const uint8_t *)(iq2xxs_grid + aux8[l]);
+                const uint8_t  signs = ksigns_iq2xs[(aux32[1] >> 7*l) & 127];
+                for (int j = 0; j < 8; ++j) {
+                    sumi += grid[j] * q8[j] * (signs & kmask_iq2xs[j] ? -1 : 1);
+                }
+                q8 += 8;
+            }
+            bsum += sumi * ls;
+        }
+        sumf += d * bsum;
+    }
+    *s = 0.125f * sumf;
+#endif
+}
+
+void ggml_vec_dot_iq2_xs_q8_K(int n, float * restrict s, size_t bs, const void * restrict vx, size_t bx, const void * restrict vy, size_t by, int nrc) {
+    assert(n % QK_K == 0);
+    assert(nrc == 1);
+    UNUSED(nrc);
+    UNUSED(bx);
+    UNUSED(by);
+    UNUSED(bs);
+
+    const block_iq2_xs * restrict x = vx;
+    const block_q8_K   * restrict y = vy;
+
+    const int nb = n / QK_K;
+
+#if defined(__ARM_NEON)
+
+    const uint64_t * signs64 = (const uint64_t *)keven_signs_q2xs;
+
+    ggml_int8x16x4_t q2u;
+    ggml_int8x16x4_t q2s;
+    ggml_int8x16x4_t q8b;
+
+    int32x4x4_t scales32;
+
+    float sumf = 0;
+    for (int i = 0; i < nb; ++i) {
+        const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
+        const uint16_t * restrict q2 = x[i].qs;
+        const int8_t   * restrict q8 = y[i].qs;
+        const uint8x8_t scales8 = vld1_u8(x[i].scales);
+        const uint8x8_t scales_l = vand_u8(scales8, vdup_n_u8(0xf));
+        const uint8x8_t scales_h = vshr_n_u8(scales8, 4);
+        uint8x16_t scales = vcombine_u8(vzip1_u8(scales_l, scales_h), vzip2_u8(scales_l, scales_h));
+        scales = vaddq_u8(vshlq_n_u8(scales, 1), vdupq_n_u8(1));
+        const uint16x8_t scales1 = vmovl_u8(vget_low_u8(scales));
+        const uint16x8_t scales2 = vmovl_u8(vget_high_u8(scales));
+        scales32.val[0] = vreinterpretq_s32_u32(vmovl_u16(vget_low_u16(scales1)));
+        scales32.val[1] = vreinterpretq_s32_u32(vmovl_u16(vget_high_u16(scales1)));
+        scales32.val[2] = vreinterpretq_s32_u32(vmovl_u16(vget_low_u16(scales2)));
+        scales32.val[3] = vreinterpretq_s32_u32(vmovl_u16(vget_high_u16(scales2)));
+        int32x4_t sumi = vdupq_n_s32(0);
+        for (int ib64 = 0; ib64 < QK_K/64; ++ib64) {
+            q8b = ggml_vld1q_s8_x4(q8); q8 += 64;
+            q2u.val[0] = vcombine_s8(vld1_s8((const void *)(iq2xs_grid + (q2[0] & 511))), vld1_s8((const void *)(iq2xs_grid + (q2[1] & 511))));
+            q2u.val[1] = vcombine_s8(vld1_s8((const void *)(iq2xs_grid + (q2[2] & 511))), vld1_s8((const void *)(iq2xs_grid + (q2[3] & 511))));
+            q2u.val[2] = vcombine_s8(vld1_s8((const void *)(iq2xs_grid + (q2[4] & 511))), vld1_s8((const void *)(iq2xs_grid + (q2[5] & 511))));
+            q2u.val[3] = vcombine_s8(vld1_s8((const void *)(iq2xs_grid + (q2[6] & 511))), vld1_s8((const void *)(iq2xs_grid + (q2[7] & 511))));
+            q2s.val[0] = vcombine_s8(vld1_s8((const void *)(signs64 + (q2[0] >> 9))), vld1_s8((const void *)(signs64 + (q2[1] >> 9))));
+            q2s.val[1] = vcombine_s8(vld1_s8((const void *)(signs64 + (q2[2] >> 9))), vld1_s8((const void *)(signs64 + (q2[3] >> 9))));
+            q2s.val[2] = vcombine_s8(vld1_s8((const void *)(signs64 + (q2[4] >> 9))), vld1_s8((const void *)(signs64 + (q2[5] >> 9))));
+            q2s.val[3] = vcombine_s8(vld1_s8((const void *)(signs64 + (q2[6] >> 9))), vld1_s8((const void *)(signs64 + (q2[7] >> 9))));
+            q2u.val[0] = vmulq_s8(q2u.val[0], q2s.val[0]);
+            q2u.val[1] = vmulq_s8(q2u.val[1], q2s.val[1]);
+            q2u.val[2] = vmulq_s8(q2u.val[2], q2s.val[2]);
+            q2u.val[3] = vmulq_s8(q2u.val[3], q2s.val[3]);
+            const int32x4_t p1 = ggml_vdotq_s32(vdupq_n_s32(0), q2u.val[0], q8b.val[0]);
+            const int32x4_t p2 = ggml_vdotq_s32(vdupq_n_s32(0), q2u.val[1], q8b.val[1]);
+            const int32x4_t p3 = ggml_vdotq_s32(vdupq_n_s32(0), q2u.val[2], q8b.val[2]);
+            const int32x4_t p4 = ggml_vdotq_s32(vdupq_n_s32(0), q2u.val[3], q8b.val[3]);
+            const int32x4_t p = vpaddq_s32(vpaddq_s32(p1, p2), vpaddq_s32(p3, p4));
+            sumi = vmlaq_s32(sumi, p, scales32.val[ib64]);
+            q2 += 8;
+        }
+        sumf += d*vaddvq_s32(sumi);
+    }
+    *s = 0.125f * sumf;
+
+#elif defined(__AVX2__)
+
+    const __m256i mone = _mm256_set1_epi8(1);
+    static const char block_sign_shuffle_mask_1[32] = {
+        0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x02, 0x02, 0x02, 0x02, 0x02, 0x02, 0x02, 0x02,
+        0x04, 0x04, 0x04, 0x04, 0x04, 0x04, 0x04, 0x04, 0x06, 0x06, 0x06, 0x06, 0x06, 0x06, 0x06, 0x06,
+    };
+    static const char block_sign_shuffle_mask_2[32] = {
+        0x08, 0x08, 0x08, 0x08, 0x08, 0x08, 0x08, 0x08, 0x0a, 0x0a, 0x0a, 0x0a, 0x0a, 0x0a, 0x0a, 0x0a,
+        0x0c, 0x0c, 0x0c, 0x0c, 0x0c, 0x0c, 0x0c, 0x0c, 0x0e, 0x0e, 0x0e, 0x0e, 0x0e, 0x0e, 0x0e, 0x0e,
+    };
+    static const uint8_t bit_selector_mask_bytes[32] = {
+        0x01, 0x02, 0x04, 0x08, 0x10, 0x20, 0x40, 0x80, 0x01, 0x02, 0x04, 0x08, 0x10, 0x20, 0x40, 0x80,
+        0x01, 0x02, 0x04, 0x08, 0x10, 0x20, 0x40, 0x80, 0x01, 0x02, 0x04, 0x08, 0x10, 0x20, 0x40, 0x80,
+    };
+
+    const __m256i bit_selector_mask = _mm256_loadu_si256((const __m256i*)bit_selector_mask_bytes);
+    const __m256i block_sign_shuffle_1 = _mm256_loadu_si256((const __m256i*)block_sign_shuffle_mask_1);
+    const __m256i block_sign_shuffle_2 = _mm256_loadu_si256((const __m256i*)block_sign_shuffle_mask_2);
+
+    static const uint8_t k_bit_helper[32] = {
+        0x00, 0x80, 0x80, 0x00, 0x80, 0x00, 0x00, 0x80, 0x80, 0x00, 0x00, 0x80, 0x00, 0x80, 0x80, 0x00,
+        0x00, 0x80, 0x80, 0x00, 0x80, 0x00, 0x00, 0x80, 0x80, 0x00, 0x00, 0x80, 0x00, 0x80, 0x80, 0x00,
+    };
+    const __m256i bit_helper = _mm256_loadu_si256((const __m256i*)k_bit_helper);
+    const __m256i m511 = _mm256_set1_epi16(511);
+    const __m128i m4 = _mm_set1_epi8(0xf);
+    const __m128i m1 = _mm_set1_epi8(1);
+
+    uint64_t aux64;
+
+    // somewhat hacky, but gives a significant boost in performance
+    __m256i aux_gindex;
+    const uint16_t * gindex = (const uint16_t *)&aux_gindex;
+
+    __m256 accumf = _mm256_setzero_ps();
+    for (int i = 0; i < nb; ++i) {
+        const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
+        const uint16_t * restrict q2 = x[i].qs;
+        const int8_t   * restrict q8 = y[i].qs;
+
+        memcpy(&aux64, x[i].scales, 8);
+        __m128i stmp = _mm_set1_epi64x(aux64);
+        stmp = _mm_unpacklo_epi8(_mm_and_si128(stmp, m4), _mm_and_si128(_mm_srli_epi16(stmp, 4), m4));
+        const __m128i scales = _mm_add_epi8(_mm_slli_epi16(stmp, 1), m1);
+
+        __m256i sumi1 = _mm256_setzero_si256();
+        __m256i sumi2 = _mm256_setzero_si256();
+        for (int ib32 = 0; ib32 < QK_K/32; ib32 += 4) {
+
+            const __m256i q2_data = _mm256_loadu_si256((const __m256i*)q2);  q2 += 16;
+            aux_gindex = _mm256_and_si256(q2_data, m511);
+
+            const __m256i partial_sign_bits = _mm256_srli_epi16(q2_data, 9);
+            const __m256i partial_sign_bits_upper = _mm256_srli_epi16(q2_data, 13);
+            const __m256i partial_sign_bits_for_counting = _mm256_xor_si256(partial_sign_bits, partial_sign_bits_upper);
+
+            const __m256i odd_bits = _mm256_shuffle_epi8(bit_helper, partial_sign_bits_for_counting);
+            const __m256i full_sign_bits = _mm256_or_si256(partial_sign_bits, odd_bits);
+
+            const __m256i q8_1 = _mm256_loadu_si256((const __m256i *)q8); q8 += 32;
+            const __m256i q8_2 = _mm256_loadu_si256((const __m256i *)q8); q8 += 32;
+            const __m256i q8_3 = _mm256_loadu_si256((const __m256i *)q8); q8 += 32;
+            const __m256i q8_4 = _mm256_loadu_si256((const __m256i *)q8); q8 += 32;
+
+            const __m256i q2_1 = _mm256_set_epi64x(iq2xs_grid[gindex[ 3]], iq2xs_grid[gindex[ 2]],
+                                                   iq2xs_grid[gindex[ 1]], iq2xs_grid[gindex[ 0]]);
+            const __m256i q2_2 = _mm256_set_epi64x(iq2xs_grid[gindex[ 7]], iq2xs_grid[gindex[ 6]],
+                                                   iq2xs_grid[gindex[ 5]], iq2xs_grid[gindex[ 4]]);
+            const __m256i q2_3 = _mm256_set_epi64x(iq2xs_grid[gindex[11]], iq2xs_grid[gindex[10]],
+                                                   iq2xs_grid[gindex[ 9]], iq2xs_grid[gindex[ 8]]);
+            const __m256i q2_4 = _mm256_set_epi64x(iq2xs_grid[gindex[15]], iq2xs_grid[gindex[14]],
+                                                   iq2xs_grid[gindex[13]], iq2xs_grid[gindex[12]]);
+
+            const __m128i full_signs_l = _mm256_castsi256_si128(full_sign_bits);
+            const __m128i full_signs_h = _mm256_extractf128_si256(full_sign_bits, 1);
+            const __m256i full_signs_1 = MM256_SET_M128I(full_signs_l, full_signs_l);
+            const __m256i full_signs_2 = MM256_SET_M128I(full_signs_h, full_signs_h);
+
+            __m256i signs;
+            signs = _mm256_shuffle_epi8(full_signs_1, block_sign_shuffle_1);
+            signs = _mm256_cmpeq_epi8(_mm256_and_si256(signs, bit_selector_mask), bit_selector_mask);
+            const __m256i q8s_1 = _mm256_sign_epi8(q8_1, _mm256_or_si256(signs, mone));
+
+            signs = _mm256_shuffle_epi8(full_signs_1, block_sign_shuffle_2);
+            signs = _mm256_cmpeq_epi8(_mm256_and_si256(signs, bit_selector_mask), bit_selector_mask);
+            const __m256i q8s_2 = _mm256_sign_epi8(q8_2, _mm256_or_si256(signs, mone));
+
+            signs = _mm256_shuffle_epi8(full_signs_2, block_sign_shuffle_1);
+            signs = _mm256_cmpeq_epi8(_mm256_and_si256(signs, bit_selector_mask), bit_selector_mask);
+            const __m256i q8s_3 = _mm256_sign_epi8(q8_3, _mm256_or_si256(signs, mone));
+
+            signs = _mm256_shuffle_epi8(full_signs_2, block_sign_shuffle_2);
+            signs = _mm256_cmpeq_epi8(_mm256_and_si256(signs, bit_selector_mask), bit_selector_mask);
+            const __m256i q8s_4 = _mm256_sign_epi8(q8_4, _mm256_or_si256(signs, mone));
+
+            const __m256i dot1  = _mm256_maddubs_epi16(q2_1, q8s_1);
+            const __m256i dot2  = _mm256_maddubs_epi16(q2_2, q8s_2);
+            const __m256i dot3  = _mm256_maddubs_epi16(q2_3, q8s_3);
+            const __m256i dot4  = _mm256_maddubs_epi16(q2_4, q8s_4);
+
+            const __m256i sc1 = _mm256_cvtepi8_epi16(_mm_shuffle_epi8(scales, get_scale_shuffle(ib32+0)));
+            const __m256i sc2 = _mm256_cvtepi8_epi16(_mm_shuffle_epi8(scales, get_scale_shuffle(ib32+1)));
+            const __m256i sc3 = _mm256_cvtepi8_epi16(_mm_shuffle_epi8(scales, get_scale_shuffle(ib32+2)));
+            const __m256i sc4 = _mm256_cvtepi8_epi16(_mm_shuffle_epi8(scales, get_scale_shuffle(ib32+3)));
+
+            sumi1 = _mm256_add_epi32(sumi1, _mm256_madd_epi16(dot1, sc1));
+            sumi2 = _mm256_add_epi32(sumi2, _mm256_madd_epi16(dot2, sc2));
+            sumi1 = _mm256_add_epi32(sumi1, _mm256_madd_epi16(dot3, sc3));
+            sumi2 = _mm256_add_epi32(sumi2, _mm256_madd_epi16(dot4, sc4));
+        }
+
+        accumf = _mm256_fmadd_ps(_mm256_set1_ps(d), _mm256_cvtepi32_ps(_mm256_add_epi32(sumi1, sumi2)), accumf);
+
+    }
+
+    *s = 0.125f * hsum_float_8(accumf);
+
+#elif defined(__AVX__)
+    const __m128i mone = _mm_set1_epi8(1);
+    static const char block_sign_shuffle_mask_1[32] = {
+        0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x02, 0x02, 0x02, 0x02, 0x02, 0x02, 0x02, 0x02,
+        0x04, 0x04, 0x04, 0x04, 0x04, 0x04, 0x04, 0x04, 0x06, 0x06, 0x06, 0x06, 0x06, 0x06, 0x06, 0x06,
+    };
+    static const char block_sign_shuffle_mask_2[32] = {
+        0x08, 0x08, 0x08, 0x08, 0x08, 0x08, 0x08, 0x08, 0x0a, 0x0a, 0x0a, 0x0a, 0x0a, 0x0a, 0x0a, 0x0a,
+        0x0c, 0x0c, 0x0c, 0x0c, 0x0c, 0x0c, 0x0c, 0x0c, 0x0e, 0x0e, 0x0e, 0x0e, 0x0e, 0x0e, 0x0e, 0x0e,
+    };
+    static const uint8_t bit_selector_mask_bytes[32] = {
+        0x01, 0x02, 0x04, 0x08, 0x10, 0x20, 0x40, 0x80, 0x01, 0x02, 0x04, 0x08, 0x10, 0x20, 0x40, 0x80,
+        0x01, 0x02, 0x04, 0x08, 0x10, 0x20, 0x40, 0x80, 0x01, 0x02, 0x04, 0x08, 0x10, 0x20, 0x40, 0x80,
+    };
+
+    const __m128i bit_selector_mask_0 = _mm_loadu_si128((const __m128i*)bit_selector_mask_bytes);
+    const __m128i bit_selector_mask_1 = _mm_loadu_si128((const __m128i*)bit_selector_mask_bytes + 1);
+    const __m128i block_sign_shuffle_1_0 = _mm_loadu_si128((const __m128i*)block_sign_shuffle_mask_1);
+    const __m128i block_sign_shuffle_1_1 = _mm_loadu_si128((const __m128i*)block_sign_shuffle_mask_1 + 1);
+    const __m128i block_sign_shuffle_2_0 = _mm_loadu_si128((const __m128i*)block_sign_shuffle_mask_2);
+    const __m128i block_sign_shuffle_2_1 = _mm_loadu_si128((const __m128i*)block_sign_shuffle_mask_2 + 1);
+
+    static const uint8_t k_bit_helper[32] = {
+        0x00, 0x80, 0x80, 0x00, 0x80, 0x00, 0x00, 0x80, 0x80, 0x00, 0x00, 0x80, 0x00, 0x80, 0x80, 0x00,
+        0x00, 0x80, 0x80, 0x00, 0x80, 0x00, 0x00, 0x80, 0x80, 0x00, 0x00, 0x80, 0x00, 0x80, 0x80, 0x00,
+    };
+    const __m128i bit_helper_0 = _mm_loadu_si128((const __m128i*)k_bit_helper);
+    const __m128i bit_helper_1 = _mm_loadu_si128((const __m128i*)k_bit_helper + 1);
+    const __m128i m511 = _mm_set1_epi16(511);
+    const __m128i m4 = _mm_set1_epi8(0xf);
+    const __m128i m1 = _mm_set1_epi8(1);
+
+    uint64_t aux64;
+
+    // somewhat hacky, but gives a significant boost in performance
+    __m256i aux_gindex;
+    const uint16_t * gindex = (const uint16_t *)&aux_gindex;
+
+    __m256 accumf = _mm256_setzero_ps();
+    for (int i = 0; i < nb; ++i) {
+        const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
+        const uint16_t * restrict q2 = x[i].qs;
+        const int8_t   * restrict q8 = y[i].qs;
+
+        memcpy(&aux64, x[i].scales, 8);
+        __m128i stmp = _mm_set1_epi64x(aux64);
+        stmp = _mm_unpacklo_epi8(_mm_and_si128(stmp, m4), _mm_and_si128(_mm_srli_epi16(stmp, 4), m4));
+        const __m128i scales = _mm_add_epi8(_mm_slli_epi16(stmp, 1), m1);
+
+        __m128i sumi1_0 = _mm_setzero_si128();
+        __m128i sumi1_1 = _mm_setzero_si128();
+        __m128i sumi2_0 = _mm_setzero_si128();
+        __m128i sumi2_1 = _mm_setzero_si128();
+        for (int ib32 = 0; ib32 < QK_K/32; ib32 += 4) {
+
+            const __m128i q2_data_0 = _mm_loadu_si128((const __m128i*)q2);
+            const __m128i q2_data_1 = _mm_loadu_si128((const __m128i*)q2 + 1);  q2 += 16;
+            aux_gindex = MM256_SET_M128I(_mm_and_si128(q2_data_1, m511), _mm_and_si128(q2_data_0, m511));
+
+            const __m128i partial_sign_bits_0 = _mm_srli_epi16(q2_data_0, 9);
+            const __m128i partial_sign_bits_1 = _mm_srli_epi16(q2_data_1, 9);
+            const __m128i partial_sign_bits_upper_0 = _mm_srli_epi16(q2_data_0, 13);
+            const __m128i partial_sign_bits_upper_1 = _mm_srli_epi16(q2_data_1, 13);
+            const __m128i partial_sign_bits_for_counting_0 = _mm_xor_si128(partial_sign_bits_0, partial_sign_bits_upper_0);
+            const __m128i partial_sign_bits_for_counting_1 = _mm_xor_si128(partial_sign_bits_1, partial_sign_bits_upper_1);
+
+            const __m128i odd_bits_0 = _mm_shuffle_epi8(bit_helper_0, partial_sign_bits_for_counting_0);
+            const __m128i odd_bits_1 = _mm_shuffle_epi8(bit_helper_1, partial_sign_bits_for_counting_1);
+            const __m128i full_sign_bits_0 = _mm_or_si128(partial_sign_bits_0, odd_bits_0);
+            const __m128i full_sign_bits_1 = _mm_or_si128(partial_sign_bits_1, odd_bits_1);
+
+            const __m128i q8_1_0 = _mm_loadu_si128((const __m128i *)q8); q8 += 16;
+            const __m128i q8_1_1 = _mm_loadu_si128((const __m128i *)q8); q8 += 16;
+            const __m128i q8_2_0 = _mm_loadu_si128((const __m128i *)q8); q8 += 16;
+            const __m128i q8_2_1 = _mm_loadu_si128((const __m128i *)q8); q8 += 16;
+            const __m128i q8_3_0 = _mm_loadu_si128((const __m128i *)q8); q8 += 16;
+            const __m128i q8_3_1 = _mm_loadu_si128((const __m128i *)q8); q8 += 16;
+            const __m128i q8_4_0 = _mm_loadu_si128((const __m128i *)q8); q8 += 16;
+            const __m128i q8_4_1 = _mm_loadu_si128((const __m128i *)q8); q8 += 16;
+
+            const __m128i q2_1_0 = _mm_set_epi64x(iq2xs_grid[gindex[1]], iq2xs_grid[gindex[0]]);
+            const __m128i q2_1_1 = _mm_set_epi64x(iq2xs_grid[gindex[3]], iq2xs_grid[gindex[2]]);
+            const __m128i q2_2_0 = _mm_set_epi64x(iq2xs_grid[gindex[5]], iq2xs_grid[gindex[4]]);
+            const __m128i q2_2_1 = _mm_set_epi64x(iq2xs_grid[gindex[7]], iq2xs_grid[gindex[6]]);
+            const __m128i q2_3_0 = _mm_set_epi64x(iq2xs_grid[gindex[9]], iq2xs_grid[gindex[8]]);
+            const __m128i q2_3_1 = _mm_set_epi64x(iq2xs_grid[gindex[11]], iq2xs_grid[gindex[10]]);
+            const __m128i q2_4_0 = _mm_set_epi64x(iq2xs_grid[gindex[13]], iq2xs_grid[gindex[12]]);
+            const __m128i q2_4_1 = _mm_set_epi64x(iq2xs_grid[gindex[15]], iq2xs_grid[gindex[14]]);
+
+            // AVX2 full_signs_1 is full_sign_bits_0 here
+            // AVX2 full_signs_2 is full_sign_bits_1 here
+            __m128i signs_0, signs_1;
+            signs_0 = _mm_shuffle_epi8(full_sign_bits_0, block_sign_shuffle_1_0);
+            signs_1 = _mm_shuffle_epi8(full_sign_bits_0, block_sign_shuffle_1_1);
+            signs_0 = _mm_cmpeq_epi8(_mm_and_si128(signs_0, bit_selector_mask_0), bit_selector_mask_0);
+            signs_1 = _mm_cmpeq_epi8(_mm_and_si128(signs_1, bit_selector_mask_1), bit_selector_mask_1);
+            const __m128i q8s_1_0 = _mm_sign_epi8(q8_1_0, _mm_or_si128(signs_0, mone));
+            const __m128i q8s_1_1 = _mm_sign_epi8(q8_1_1, _mm_or_si128(signs_1, mone));
+
+            signs_0 = _mm_shuffle_epi8(full_sign_bits_0, block_sign_shuffle_2_0);
+            signs_1 = _mm_shuffle_epi8(full_sign_bits_0, block_sign_shuffle_2_1);
+            signs_0 = _mm_cmpeq_epi8(_mm_and_si128(signs_0, bit_selector_mask_0), bit_selector_mask_0);
+            signs_1 = _mm_cmpeq_epi8(_mm_and_si128(signs_1, bit_selector_mask_1), bit_selector_mask_1);
+            const __m128i q8s_2_0 = _mm_sign_epi8(q8_2_0, _mm_or_si128(signs_0, mone));
+            const __m128i q8s_2_1 = _mm_sign_epi8(q8_2_1, _mm_or_si128(signs_1, mone));
+
+            signs_0 = _mm_shuffle_epi8(full_sign_bits_1, block_sign_shuffle_1_0);
+            signs_1 = _mm_shuffle_epi8(full_sign_bits_1, block_sign_shuffle_1_1);
+            signs_0 = _mm_cmpeq_epi8(_mm_and_si128(signs_0, bit_selector_mask_0), bit_selector_mask_0);
+            signs_1 = _mm_cmpeq_epi8(_mm_and_si128(signs_1, bit_selector_mask_1), bit_selector_mask_1);
+            const __m128i q8s_3_0 = _mm_sign_epi8(q8_3_0, _mm_or_si128(signs_0, mone));
+            const __m128i q8s_3_1 = _mm_sign_epi8(q8_3_1, _mm_or_si128(signs_1, mone));
+
+            signs_0 = _mm_shuffle_epi8(full_sign_bits_1, block_sign_shuffle_2_0);
+            signs_1 = _mm_shuffle_epi8(full_sign_bits_1, block_sign_shuffle_2_1);
+            signs_0 = _mm_cmpeq_epi8(_mm_and_si128(signs_0, bit_selector_mask_0), bit_selector_mask_0);
+            signs_1 = _mm_cmpeq_epi8(_mm_and_si128(signs_1, bit_selector_mask_1), bit_selector_mask_1);
+            const __m128i q8s_4_0 = _mm_sign_epi8(q8_4_0, _mm_or_si128(signs_0, mone));
+            const __m128i q8s_4_1 = _mm_sign_epi8(q8_4_1, _mm_or_si128(signs_1, mone));
+
+            const __m128i dot1_0  = _mm_maddubs_epi16(q2_1_0, q8s_1_0);
+            const __m128i dot1_1  = _mm_maddubs_epi16(q2_1_1, q8s_1_1);
+            const __m128i dot2_0  = _mm_maddubs_epi16(q2_2_0, q8s_2_0);
+            const __m128i dot2_1  = _mm_maddubs_epi16(q2_2_1, q8s_2_1);
+            const __m128i dot3_0  = _mm_maddubs_epi16(q2_3_0, q8s_3_0);
+            const __m128i dot3_1  = _mm_maddubs_epi16(q2_3_1, q8s_3_1);
+            const __m128i dot4_0  = _mm_maddubs_epi16(q2_4_0, q8s_4_0);
+            const __m128i dot4_1  = _mm_maddubs_epi16(q2_4_1, q8s_4_1);
+
+            __m128i sc_tmp = _mm_shuffle_epi8(scales, get_scale_shuffle(ib32+0));
+            const __m128i sc1_0 = _mm_cvtepi8_epi16(sc_tmp);
+            const __m128i sc1_1 = _mm_cvtepi8_epi16(_mm_srli_si128(sc_tmp, 8));
+            sc_tmp = _mm_shuffle_epi8(scales, get_scale_shuffle(ib32+1));
+            const __m128i sc2_0 = _mm_cvtepi8_epi16(sc_tmp);
+            const __m128i sc2_1 = _mm_cvtepi8_epi16(_mm_srli_si128(sc_tmp, 8));
+            sc_tmp = _mm_shuffle_epi8(scales, get_scale_shuffle(ib32+2));
+            const __m128i sc3_0 = _mm_cvtepi8_epi16(sc_tmp);
+            const __m128i sc3_1 = _mm_cvtepi8_epi16(_mm_srli_si128(sc_tmp, 8));
+            sc_tmp = _mm_shuffle_epi8(scales, get_scale_shuffle(ib32+3));
+            const __m128i sc4_0 = _mm_cvtepi8_epi16(sc_tmp);
+            const __m128i sc4_1 = _mm_cvtepi8_epi16(_mm_srli_si128(sc_tmp, 8));
+
+            sumi1_0 = _mm_add_epi32(sumi1_0, _mm_madd_epi16(dot1_0, sc1_0));
+            sumi1_1 = _mm_add_epi32(sumi1_1, _mm_madd_epi16(dot1_1, sc1_1));
+            sumi2_0 = _mm_add_epi32(sumi2_0, _mm_madd_epi16(dot2_0, sc2_0));
+            sumi2_1 = _mm_add_epi32(sumi2_1, _mm_madd_epi16(dot2_1, sc2_1));
+            sumi1_0 = _mm_add_epi32(sumi1_0, _mm_madd_epi16(dot3_0, sc3_0));
+            sumi1_1 = _mm_add_epi32(sumi1_1, _mm_madd_epi16(dot3_1, sc3_1));
+            sumi2_0 = _mm_add_epi32(sumi2_0, _mm_madd_epi16(dot4_0, sc4_0));
+            sumi2_1 = _mm_add_epi32(sumi2_1, _mm_madd_epi16(dot4_1, sc4_1));
+        }
+
+        accumf = _mm256_add_ps(_mm256_mul_ps(_mm256_set1_ps(d), _mm256_cvtepi32_ps(MM256_SET_M128I(_mm_add_epi32(sumi1_1, sumi2_1), _mm_add_epi32(sumi1_0, sumi2_0)))), accumf);
+
+    }
+
+    *s = 0.125f * hsum_float_8(accumf);
+
+#elif defined(__loongarch_asx)
+
+    const __m256i mone = __lasx_xvreplgr2vr_b(1);
+    static const char block_sign_shuffle_mask_1[32] = {
+        0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x02, 0x02, 0x02, 0x02, 0x02, 0x02, 0x02, 0x02,
+        0x04, 0x04, 0x04, 0x04, 0x04, 0x04, 0x04, 0x04, 0x06, 0x06, 0x06, 0x06, 0x06, 0x06, 0x06, 0x06,
+    };
+    static const char block_sign_shuffle_mask_2[32] = {
+        0x08, 0x08, 0x08, 0x08, 0x08, 0x08, 0x08, 0x08, 0x0a, 0x0a, 0x0a, 0x0a, 0x0a, 0x0a, 0x0a, 0x0a,
+        0x0c, 0x0c, 0x0c, 0x0c, 0x0c, 0x0c, 0x0c, 0x0c, 0x0e, 0x0e, 0x0e, 0x0e, 0x0e, 0x0e, 0x0e, 0x0e,
+    };
+    static const uint8_t bit_selector_mask_bytes[32] = {
+        0x01, 0x02, 0x04, 0x08, 0x10, 0x20, 0x40, 0x80, 0x01, 0x02, 0x04, 0x08, 0x10, 0x20, 0x40, 0x80,
+        0x01, 0x02, 0x04, 0x08, 0x10, 0x20, 0x40, 0x80, 0x01, 0x02, 0x04, 0x08, 0x10, 0x20, 0x40, 0x80,
+    };
+
+    const __m256i bit_selector_mask = __lasx_xvld((const __m256i*)bit_selector_mask_bytes, 0);
+    const __m256i block_sign_shuffle_1 = __lasx_xvld((const __m256i*)block_sign_shuffle_mask_1, 0);
+    const __m256i block_sign_shuffle_2 = __lasx_xvld((const __m256i*)block_sign_shuffle_mask_2, 0);
+
+    static const uint8_t k_bit_helper[32] = {
+        0x00, 0x80, 0x80, 0x00, 0x80, 0x00, 0x00, 0x80, 0x80, 0x00, 0x00, 0x80, 0x00, 0x80, 0x80, 0x00,
+        0x00, 0x80, 0x80, 0x00, 0x80, 0x00, 0x00, 0x80, 0x80, 0x00, 0x00, 0x80, 0x00, 0x80, 0x80, 0x00,
+    };
+    const __m256i bit_helper = __lasx_xvld((const __m256i*)k_bit_helper, 0);
+    const __m256i m511 = __lasx_xvreplgr2vr_h(511);
+    const __m128i m4 = __lsx_vreplgr2vr_b(0xf);
+    const __m128i m1 = __lsx_vreplgr2vr_b(1);
+
+    uint64_t aux64;
+
+    // somewhat hacky, but gives a significant boost in performance
+    __m256i aux_gindex;
+    const uint16_t * gindex = (const uint16_t *)&aux_gindex;
+
+    __m256 accumf = (__m256)__lasx_xvldi(0);
+    for (int i = 0; i < nb; ++i) {
+        const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
+        const uint16_t * restrict q2 = x[i].qs;
+        const int8_t   * restrict q8 = y[i].qs;
+
+        memcpy(&aux64, x[i].scales, 8);
+        __m128i stmp = __lsx_vreplgr2vr_d(aux64);
+        stmp = __lsx_vilvl_b( __lsx_vand_v(__lsx_vsrli_h(stmp, 4), m4), __lsx_vand_v(stmp, m4));
+        const __m128i scales = __lsx_vadd_b(__lsx_vslli_h(stmp, 1), m1);
+
+        __m256i sumi1 = __lasx_xvldi(0);
+        __m256i sumi2 = __lasx_xvldi(0);
+        for (int ib32 = 0; ib32 < QK_K/32; ib32 += 4) {
+
+            const __m256i q2_data = __lasx_xvld((const __m256i*)q2, 0);  q2 += 16;
+            aux_gindex = __lasx_xvand_v(q2_data, m511);
+
+            const __m256i partial_sign_bits = __lasx_xvsrli_h(q2_data, 9);
+            const __m256i partial_sign_bits_upper = __lasx_xvsrli_h(q2_data, 13);
+            const __m256i partial_sign_bits_for_counting = __lasx_xvxor_v(partial_sign_bits, partial_sign_bits_upper);
+
+            const __m256i odd_bits = lasx_shuffle_b(bit_helper, partial_sign_bits_for_counting);
+            const __m256i full_sign_bits = __lasx_xvor_v(partial_sign_bits, odd_bits);
+
+            const __m256i q8_1 = __lasx_xvld((const __m256i *)q8, 0); q8 += 32;
+            const __m256i q8_2 = __lasx_xvld((const __m256i *)q8, 0); q8 += 32;
+            const __m256i q8_3 = __lasx_xvld((const __m256i *)q8, 0); q8 += 32;
+            const __m256i q8_4 = __lasx_xvld((const __m256i *)q8, 0); q8 += 32;
+
+            const __m256i q2_1 = lasx_set_d(iq2xs_grid[gindex[ 3]], iq2xs_grid[gindex[ 2]],
+                                                   iq2xs_grid[gindex[ 1]], iq2xs_grid[gindex[ 0]]);
+            const __m256i q2_2 = lasx_set_d(iq2xs_grid[gindex[ 7]], iq2xs_grid[gindex[ 6]],
+                                                   iq2xs_grid[gindex[ 5]], iq2xs_grid[gindex[ 4]]);
+            const __m256i q2_3 = lasx_set_d(iq2xs_grid[gindex[11]], iq2xs_grid[gindex[10]],
+                                                   iq2xs_grid[gindex[ 9]], iq2xs_grid[gindex[ 8]]);
+            const __m256i q2_4 = lasx_set_d(iq2xs_grid[gindex[15]], iq2xs_grid[gindex[14]],
+                                                   iq2xs_grid[gindex[13]], iq2xs_grid[gindex[12]]);
+
+            const __m128i full_signs_l = lasx_extracti128(full_sign_bits, 0);
+            const __m128i full_signs_h = lasx_extracti128(full_sign_bits, 1);
+            const __m256i full_signs_1 = lasx_insertf128(full_signs_l, full_signs_l);
+            const __m256i full_signs_2 = lasx_insertf128(full_signs_h, full_signs_h);
+
+            __m256i signs;
+            signs = lasx_shuffle_b(full_signs_1, block_sign_shuffle_1);
+            signs = __lasx_xvseq_b(__lasx_xvand_v(signs, bit_selector_mask), bit_selector_mask);
+            const __m256i q8s_1 = __lasx_xvsigncov_b(__lasx_xvor_v(signs, mone), q8_1);
+
+            signs = lasx_shuffle_b(full_signs_1, block_sign_shuffle_2);
+            signs = __lasx_xvseq_b(__lasx_xvand_v(signs, bit_selector_mask), bit_selector_mask);
+            const __m256i q8s_2 = __lasx_xvsigncov_b(__lasx_xvor_v(signs, mone), q8_2);
+
+            signs = lasx_shuffle_b(full_signs_2, block_sign_shuffle_1);
+            signs = __lasx_xvseq_b(__lasx_xvand_v(signs, bit_selector_mask), bit_selector_mask);
+            const __m256i q8s_3 = __lasx_xvsigncov_b(__lasx_xvor_v(signs, mone), q8_3);
+
+            signs = lasx_shuffle_b(full_signs_2, block_sign_shuffle_2);
+            signs = __lasx_xvseq_b(__lasx_xvand_v(signs, bit_selector_mask), bit_selector_mask);
+            const __m256i q8s_4 = __lasx_xvsigncov_b(__lasx_xvor_v(signs, mone), q8_4);
+
+            const __m256i dot1  = lasx_maddubs_h(q2_1, q8s_1);
+            const __m256i dot2  = lasx_maddubs_h(q2_2, q8s_2);
+            const __m256i dot3  = lasx_maddubs_h(q2_3, q8s_3);
+            const __m256i dot4  = lasx_maddubs_h(q2_4, q8s_4);
+
+            const __m256i sc1 = lasx_ext8_16(lsx_shuffle_b(scales, get_scale_shuffle(ib32+0)));
+            const __m256i sc2 = lasx_ext8_16(lsx_shuffle_b(scales, get_scale_shuffle(ib32+1)));
+            const __m256i sc3 = lasx_ext8_16(lsx_shuffle_b(scales, get_scale_shuffle(ib32+2)));
+            const __m256i sc4 = lasx_ext8_16(lsx_shuffle_b(scales, get_scale_shuffle(ib32+3)));
+
+            sumi1 = __lasx_xvadd_w(sumi1, lasx_madd_h(dot1, sc1));
+            sumi2 = __lasx_xvadd_w(sumi2, lasx_madd_h(dot2, sc2));
+            sumi1 = __lasx_xvadd_w(sumi1, lasx_madd_h(dot3, sc3));
+            sumi2 = __lasx_xvadd_w(sumi2, lasx_madd_h(dot4, sc4));
+        }
+
+        accumf = __lasx_xvfmadd_s(__lasx_xvreplfr2vr_s(d), __lasx_xvffint_s_w(__lasx_xvadd_w(sumi1, sumi2)), accumf);
+
+    }
+
+    *s = 0.125f * hsum_float_8(accumf);
+#elif defined(__POWER9_VECTOR__)
+    const vector int v0 = vec_splats((int32_t)0);
+    vector float vsumf0 = vec_splats(0.0f);
+    vector float vsumf1 = vec_splats(0.0f);
+    vector float vsumf2 = vec_splats(0.0f);
+    vector float vsumf3 = vec_splats(0.0f);
+
+    const uint64_t * signs64 = (const uint64_t *)keven_signs_q2xs;
+
+    for (int i = 0; i < nb; ++i) {
+        vector float vxd = vec_splats(GGML_FP16_TO_FP32(x[i].d));
+        vector float vyd = vec_splats(y[i].d);
+        vector float vd = vec_mul(vxd, vyd);
+
+        vector signed int vsumi0 = v0;
+        vector signed int vsumi1 = v0;
+        vector signed int vsumi2 = v0;
+        vector signed int vsumi3 = v0;
+
+        const uint16_t * restrict q2 = x[i].qs;
+        const uint8_t  * restrict sc = x[i].scales;
+        const int8_t  *  restrict q8 = y[i].qs;
+
+        for (int j = 0; j < QK_K/64; ++j) {
+            __builtin_prefetch(q2, 0, 1);
+            __builtin_prefetch(q8, 0, 1);
+
+            vector signed long long aux64x2_0 = {*(const int64_t *)(iq2xs_grid + (q2[0] & 511)), *(const int64_t *)(iq2xs_grid + (q2[1] & 511))};
+            vector signed long long aux64x2_1 = {*(const int64_t *)(iq2xs_grid + (q2[2] & 511)), *(const int64_t *)(iq2xs_grid + (q2[3] & 511))};
+            vector signed long long aux64x2_2 = {*(const int64_t *)(iq2xs_grid + (q2[4] & 511)), *(const int64_t *)(iq2xs_grid + (q2[5] & 511))};
+            vector signed long long aux64x2_3 = {*(const int64_t *)(iq2xs_grid + (q2[6] & 511)), *(const int64_t *)(iq2xs_grid + (q2[7] & 511))};
+
+            vector signed long long vsigns0 = {*(const int64_t *)(signs64 + ((q2[0] >> 9))), *(const int64_t *)(signs64 + ((q2[1] >> 9)))};
+            vector signed long long vsigns1 = {*(const int64_t *)(signs64 + ((q2[2] >> 9))), *(const int64_t *)(signs64 + ((q2[3] >> 9)))};
+            vector signed long long vsigns2 = {*(const int64_t *)(signs64 + ((q2[4] >> 9))), *(const int64_t *)(signs64 + ((q2[5] >> 9)))};
+            vector signed long long vsigns3 = {*(const int64_t *)(signs64 + ((q2[6] >> 9))), *(const int64_t *)(signs64 + ((q2[7] >> 9)))};
+            q2 += 8;
+
+            vector signed char q2x0 = (vector signed char)vec_mul((vector signed char)vsigns0, (vector signed char)aux64x2_0);
+            vector signed char q2x1 = (vector signed char)vec_mul((vector signed char)vsigns1, (vector signed char)aux64x2_1);
+            vector signed char q2x2 = (vector signed char)vec_mul((vector signed char)vsigns2, (vector signed char)aux64x2_2);
+            vector signed char q2x3 = (vector signed char)vec_mul((vector signed char)vsigns3, (vector signed char)aux64x2_3);
+
+            vector signed char q8y0 = vec_xl( 0, q8);
+            vector signed char q8y1 = vec_xl(16, q8);
+            vector signed char q8y2 = vec_xl(32, q8);
+            vector signed char q8y3 = vec_xl(48, q8);
+            q8 += 64;
+
+            vector signed short qv0 = vec_add(vec_mule(q2x0, q8y0), vec_mulo(q2x0, q8y0));
+            vector signed short qv1 = vec_add(vec_mule(q2x1, q8y1), vec_mulo(q2x1, q8y1));
+            vector signed short qv2 = vec_add(vec_mule(q2x2, q8y2), vec_mulo(q2x2, q8y2));
+            vector signed short qv3 = vec_add(vec_mule(q2x3, q8y3), vec_mulo(q2x3, q8y3));
+
+            const uint16_t ls0 = (uint16_t)(sc[0] & 0xf);
+            const uint16_t ls1 = (uint16_t)(sc[0] >>  4);
+            const uint16_t ls2 = (uint16_t)(sc[1] & 0xf);
+            const uint16_t ls3 = (uint16_t)(sc[1] >>  4);
+            sc += 2;
+
+            vector signed short vscales0 = vec_splats((int16_t)(2*ls0+1));
+            vector signed short vscales1 = vec_splats((int16_t)(2*ls1+1));
+            vector signed short vscales2 = vec_splats((int16_t)(2*ls2+1));
+            vector signed short vscales3 = vec_splats((int16_t)(2*ls3+1));
+
+            vsumi0 = vec_msum(qv0, vscales0, vsumi0);
+            vsumi1 = vec_msum(qv1, vscales1, vsumi1);
+            vsumi2 = vec_msum(qv2, vscales2, vsumi2);
+            vsumi3 = vec_msum(qv3, vscales3, vsumi3);
+        }
+
+        vsumf0 = vec_madd(vec_ctf(vsumi0, 0), vd, vsumf0);
+        vsumf1 = vec_madd(vec_ctf(vsumi1, 0), vd, vsumf1);
+        vsumf2 = vec_madd(vec_ctf(vsumi2, 0), vd, vsumf2);
+        vsumf3 = vec_madd(vec_ctf(vsumi3, 0), vd, vsumf3);
+    }
+
+    vsumf0 = vec_add(vsumf0, vsumf2);
+    vsumf1 = vec_add(vsumf1, vsumf3);
+
+    vsumf0 = vec_add(vsumf0, vsumf1);
+
+    vsumf0 = vec_add(vsumf0, vec_sld(vsumf0, vsumf0, 4));
+    vsumf0 = vec_add(vsumf0, vec_sld(vsumf0, vsumf0, 8));
+
+    *s = 0.125f * vec_extract(vsumf0, 0);
+#else
+
+    float sumf = 0.f;
+    for (int i = 0; i < nb; ++i) {
+        const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
+        const uint16_t * restrict q2 = x[i].qs;
+        const uint8_t  * restrict sc = x[i].scales;
+        const int8_t   * restrict q8 = y[i].qs;
+        int32_t bsum = 0;
+        for (int ib32 = 0; ib32 < QK_K/32; ++ib32) {
+            const uint16_t ls1 = 2*(sc[ib32] & 0xf) + 1;
+            const uint16_t ls2 = 2*(sc[ib32] >>  4) + 1;
+            int32_t sumi = 0;
+            for (int l = 0; l < 2; ++l) {
+                const uint8_t * grid = (const uint8_t *)(iq2xs_grid + (q2[l] & 511));
+                const uint8_t  signs = ksigns_iq2xs[q2[l] >> 9];
+                for (int j = 0; j < 8; ++j) {
+                    sumi += grid[j] * q8[j] * (signs & kmask_iq2xs[j] ? -1 : 1);
+                }
+                q8 += 8;
+            }
+            bsum += sumi * ls1;
+            sumi = 0;
+            for (int l = 2; l < 4; ++l) {
+                const uint8_t * grid = (const uint8_t *)(iq2xs_grid + (q2[l] & 511));
+                const uint8_t  signs = ksigns_iq2xs[q2[l] >> 9];
+                for (int j = 0; j < 8; ++j) {
+                    sumi += grid[j] * q8[j] * (signs & kmask_iq2xs[j] ? -1 : 1);
+                }
+                q8 += 8;
+            }
+            bsum += sumi * ls2;
+            q2 += 4;
+        }
+        sumf += d * bsum;
+    }
+    *s = 0.125f * sumf;
+#endif
+}
+
+void ggml_vec_dot_iq2_s_q8_K(int n, float * restrict s, size_t bs, const void * restrict vx, size_t bx, const void * restrict vy, size_t by, int nrc) {
+    assert(n % QK_K == 0);
+    assert(nrc == 1);
+    UNUSED(nrc);
+    UNUSED(bx);
+    UNUSED(by);
+    UNUSED(bs);
+
+    const block_iq2_s * restrict x = vx;
+    const block_q8_K  * restrict y = vy;
+
+    const int nb = n / QK_K;
+
+#if defined(__ARM_NEON)
+
+   static const uint8_t k_mask1[32] = {0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x01, 0x01, 0x01, 0x01, 0x01, 0x01, 0x01, 0x01,
+                                       0x02, 0x02, 0x02, 0x02, 0x02, 0x02, 0x02, 0x02, 0x03, 0x03, 0x03, 0x03, 0x03, 0x03, 0x03, 0x03
+   };
+
+    static const uint8_t k_mask2[16] = {0x01, 0x02, 0x04, 0x08, 0x10, 0x20, 0x40, 0x80, 0x01, 0x02, 0x04, 0x08, 0x10, 0x20, 0x40, 0x80,};
+
+    const ggml_uint8x16x2_t mask1 = ggml_vld1q_u8_x2(k_mask1);
+    const uint8x16_t        mask2 = vld1q_u8(k_mask2);
+    const uint8x16_t m1 = vdupq_n_u8(1);
+    const int32x4_t vzero = vdupq_n_s32(0);
+
+    uint8x16x2_t vs;
+    ggml_int8x16x4_t q2s;
+    ggml_int8x16x4_t q8b;
+
+    float sumf = 0;
+    for (int i = 0; i < nb; ++i) {
+
+        const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
+
+        const uint8_t * restrict qs = x[i].qs;
+        const uint8_t * restrict qh = x[i].qh;
+        const uint16_t * restrict signs = (const uint16_t *)(x[i].qs + QK_K/8);
+        const int8_t  * restrict q8 = y[i].qs;
+
+        int sumi1 = 0, sumi2 = 0;
+        for (int ib32 = 0; ib32 < QK_K/32; ib32 += 2) {
+            q8b = ggml_vld1q_s8_x4(q8); q8 += 64;
+            q2s.val[0] = vcombine_s8(vld1_s8((const int8_t *)(iq2s_grid + (qs[0] | ((qh[ib32+0] << 8) & 0x300)))),
+                                     vld1_s8((const int8_t *)(iq2s_grid + (qs[1] | ((qh[ib32+0] << 6) & 0x300)))));
+            q2s.val[1] = vcombine_s8(vld1_s8((const int8_t *)(iq2s_grid + (qs[2] | ((qh[ib32+0] << 4) & 0x300)))),
+                                     vld1_s8((const int8_t *)(iq2s_grid + (qs[3] | ((qh[ib32+0] << 2) & 0x300)))));
+            q2s.val[2] = vcombine_s8(vld1_s8((const int8_t *)(iq2s_grid + (qs[4] | ((qh[ib32+1] << 8) & 0x300)))),
+                                     vld1_s8((const int8_t *)(iq2s_grid + (qs[5] | ((qh[ib32+1] << 6) & 0x300)))));
+            q2s.val[3] = vcombine_s8(vld1_s8((const int8_t *)(iq2s_grid + (qs[6] | ((qh[ib32+1] << 4) & 0x300)))),
+                                     vld1_s8((const int8_t *)(iq2s_grid + (qs[7] | ((qh[ib32+1] << 2) & 0x300)))));
+            qs += 8;
+
+            vs.val[0] = vreinterpretq_u8_u32(vdupq_n_u32(signs[0] | ((uint32_t) signs[1] << 16)));
+            vs.val[1] = vandq_u8(ggml_vqtbl1q_u8(vs.val[0], mask1.val[1]), mask2);
+            vs.val[0] = vandq_u8(ggml_vqtbl1q_u8(vs.val[0], mask1.val[0]), mask2);
+            vs.val[0] = vceqq_u8(vs.val[0], mask2);
+            vs.val[1] = vceqq_u8(vs.val[1], mask2);
+
+            q2s.val[0] = vmulq_s8(vreinterpretq_s8_u8(vorrq_u8(vs.val[0], m1)), q2s.val[0]);
+            q2s.val[1] = vmulq_s8(vreinterpretq_s8_u8(vorrq_u8(vs.val[1], m1)), q2s.val[1]);
+
+            vs.val[0] = vreinterpretq_u8_u32(vdupq_n_u32(signs[2] | ((uint32_t) signs[3] << 16)));
+            vs.val[1] = vandq_u8(ggml_vqtbl1q_u8(vs.val[0], mask1.val[1]), mask2);
+            vs.val[0] = vandq_u8(ggml_vqtbl1q_u8(vs.val[0], mask1.val[0]), mask2);
+            vs.val[0] = vceqq_u8(vs.val[0], mask2);
+            vs.val[1] = vceqq_u8(vs.val[1], mask2);
+
+            signs += 4;
+
+            q2s.val[2] = vmulq_s8(vreinterpretq_s8_u8(vorrq_u8(vs.val[0], m1)), q2s.val[2]);
+            q2s.val[3] = vmulq_s8(vreinterpretq_s8_u8(vorrq_u8(vs.val[1], m1)), q2s.val[3]);
+
+            const int32x4_t p1 = ggml_vdotq_s32(vzero, q2s.val[0], q8b.val[0]);
+            const int32x4_t p2 = ggml_vdotq_s32(vzero, q2s.val[1], q8b.val[1]);
+            const int32x4_t p3 = ggml_vdotq_s32(vzero, q2s.val[2], q8b.val[2]);
+            const int32x4_t p4 = ggml_vdotq_s32(vzero, q2s.val[3], q8b.val[3]);
+
+            sumi1 += vaddvq_s32(p1) * (1 + 2*(x[i].scales[ib32+0] & 0xf));
+            sumi2 += vaddvq_s32(p2) * (1 + 2*(x[i].scales[ib32+0] >>  4));
+            sumi1 += vaddvq_s32(p3) * (1 + 2*(x[i].scales[ib32+1] & 0xf));
+            sumi2 += vaddvq_s32(p4) * (1 + 2*(x[i].scales[ib32+1] >>  4));
+        }
+        sumf += d*(sumi1 + sumi2);
+    }
+
+    *s = 0.125f * sumf;
+
+#elif defined(__AVX2__)
+
+   static const uint8_t k_mask1[32] = {0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x01, 0x01, 0x01, 0x01, 0x01, 0x01, 0x01, 0x01,
+                                       0x02, 0x02, 0x02, 0x02, 0x02, 0x02, 0x02, 0x02, 0x03, 0x03, 0x03, 0x03, 0x03, 0x03, 0x03, 0x03
+   };
+
+    static const uint8_t k_mask2[32] = {0x01, 0x02, 0x04, 0x08, 0x10, 0x20, 0x40, 0x80, 0x01, 0x02, 0x04, 0x08, 0x10, 0x20, 0x40, 0x80,
+                                        0x01, 0x02, 0x04, 0x08, 0x10, 0x20, 0x40, 0x80, 0x01, 0x02, 0x04, 0x08, 0x10, 0x20, 0x40, 0x80,
+    };
+
+    const __m128i m4 = _mm_set1_epi8(0xf);
+    const __m128i m1 = _mm_set1_epi8(1);
+
+    const __m256i mask1 = _mm256_loadu_si256((const __m256i*)k_mask1);
+    const __m256i mask2 = _mm256_loadu_si256((const __m256i*)k_mask2);
+
+    uint64_t aux64;
+
+    __m256 accumf = _mm256_setzero_ps();
+    for (int i = 0; i < nb; ++i) {
+        const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
+        const uint8_t * restrict qs = x[i].qs;
+        const uint8_t * restrict qh = x[i].qh;
+        const uint16_t * restrict signs = (const uint16_t *)(x[i].qs + QK_K/8);
+        const int8_t  * restrict q8 = y[i].qs;
+
+        memcpy(&aux64, x[i].scales, 8);
+        const __m128i scales8 = _mm_add_epi8(_mm_slli_epi16(_mm_and_si128(_mm_set_epi64x(aux64 >> 4, aux64), m4), 1), m1);
+        const __m256i scales16 = _mm256_cvtepi8_epi16(scales8); // 0 2 4 6 8 10 12 14 1 3 5 7 9 11 13 15
+
+        __m256i sumi1 = _mm256_setzero_si256();
+        __m256i sumi2 = _mm256_setzero_si256();
+        for (int ib32 = 0; ib32 < QK_K/32; ib32 += 2) {
+            const __m256i q8_1 = _mm256_loadu_si256((const __m256i *)q8); q8 += 32;
+            const __m256i q8_2 = _mm256_loadu_si256((const __m256i *)q8); q8 += 32;
+            const __m256i q2_1 = _mm256_set_epi64x(iq2s_grid[qs[3] | ((qh[ib32+0] << 2) & 0x300)],
+                                                   iq2s_grid[qs[2] | ((qh[ib32+0] << 4) & 0x300)],
+                                                   iq2s_grid[qs[1] | ((qh[ib32+0] << 6) & 0x300)],
+                                                   iq2s_grid[qs[0] | ((qh[ib32+0] << 8) & 0x300)]);
+            const __m256i q2_2 = _mm256_set_epi64x(iq2s_grid[qs[7] | ((qh[ib32+1] << 2) & 0x300)],
+                                                   iq2s_grid[qs[6] | ((qh[ib32+1] << 4) & 0x300)],
+                                                   iq2s_grid[qs[5] | ((qh[ib32+1] << 6) & 0x300)],
+                                                   iq2s_grid[qs[4] | ((qh[ib32+1] << 8) & 0x300)]);
+            qs += 8;
+
+            __m256i aux256 = _mm256_set1_epi32(signs[0] | ((uint32_t) signs[1] << 16));
+            aux256 = _mm256_and_si256(_mm256_shuffle_epi8(aux256,mask1), mask2);
+            const __m256i s2_1 = _mm256_cmpeq_epi8(aux256, mask2);
+            const __m256i q8s_1 = _mm256_sub_epi8(_mm256_xor_si256(s2_1, q8_1), s2_1);
+
+            aux256 = _mm256_set1_epi32(signs[2] | ((uint32_t) signs[3] << 16));
+            aux256 = _mm256_and_si256(_mm256_shuffle_epi8(aux256,mask1), mask2);
+            const __m256i s2_2 = _mm256_cmpeq_epi8(aux256, mask2);
+            const __m256i q8s_2 = _mm256_sub_epi8(_mm256_xor_si256(s2_2, q8_2), s2_2);
+
+            signs += 4;
+
+            const __m256i dot1  = _mm256_maddubs_epi16(q2_1, q8s_1); // blocks 2*ib32+0, 2*ib32+1
+            const __m256i dot2  = _mm256_maddubs_epi16(q2_2, q8s_2); // blocks 2*ib32+2, 2*ib32+3
+
+            const __m256i p1 = _mm256_madd_epi16(dot1, _mm256_shuffle_epi8(scales16, get_scale_shuffle_k4(ib32+0)));
+            const __m256i p2 = _mm256_madd_epi16(dot2, _mm256_shuffle_epi8(scales16, get_scale_shuffle_k4(ib32+1)));
+            sumi1 = _mm256_add_epi32(sumi1, p1);
+            sumi2 = _mm256_add_epi32(sumi2, p2);
+        }
+
+        accumf = _mm256_fmadd_ps(_mm256_set1_ps(d), _mm256_cvtepi32_ps(_mm256_add_epi32(sumi1, sumi2)), accumf);
+
+    }
+
+    *s = 0.125f * hsum_float_8(accumf);
+
+#elif defined(__AVX__)
+   static const uint8_t k_mask1[32] = {0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x01, 0x01, 0x01, 0x01, 0x01, 0x01, 0x01, 0x01,
+                                       0x02, 0x02, 0x02, 0x02, 0x02, 0x02, 0x02, 0x02, 0x03, 0x03, 0x03, 0x03, 0x03, 0x03, 0x03, 0x03
+   };
+
+    static const uint8_t k_mask2[32] = {0x01, 0x02, 0x04, 0x08, 0x10, 0x20, 0x40, 0x80, 0x01, 0x02, 0x04, 0x08, 0x10, 0x20, 0x40, 0x80,
+                                        0x01, 0x02, 0x04, 0x08, 0x10, 0x20, 0x40, 0x80, 0x01, 0x02, 0x04, 0x08, 0x10, 0x20, 0x40, 0x80,
+    };
+
+    const __m128i m4 = _mm_set1_epi8(0xf);
+    const __m128i m1 = _mm_set1_epi8(1);
+
+    const __m128i mask1_0 = _mm_loadu_si128((const __m128i*)k_mask1);
+    const __m128i mask1_1 = _mm_loadu_si128((const __m128i*)k_mask1 + 1);
+    const __m128i mask2_0 = _mm_loadu_si128((const __m128i*)k_mask2);
+    const __m128i mask2_1 = _mm_loadu_si128((const __m128i*)k_mask2 + 1);
+
+    uint64_t aux64;
+
+    __m256 accumf = _mm256_setzero_ps();
+    for (int i = 0; i < nb; ++i) {
+        const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
+        const uint8_t * restrict qs = x[i].qs;
+        const uint8_t * restrict qh = x[i].qh;
+        const uint16_t * restrict signs = (const uint16_t *)(x[i].qs + QK_K/8);
+        const int8_t  * restrict q8 = y[i].qs;
+
+        memcpy(&aux64, x[i].scales, 8);
+        const __m128i scales8 = _mm_add_epi8(_mm_slli_epi16(_mm_and_si128(_mm_set_epi64x(aux64 >> 4, aux64), m4), 1), m1);
+        const __m128i scales16_0 = _mm_cvtepi8_epi16(scales8);
+        const __m128i scales16_1 = _mm_cvtepi8_epi16(_mm_srli_si128(scales8, 8));
+
+        __m128i sumi1_0 = _mm_setzero_si128();
+        __m128i sumi1_1 = _mm_setzero_si128();
+        __m128i sumi2_0 = _mm_setzero_si128();
+        __m128i sumi2_1 = _mm_setzero_si128();
+        for (int ib32 = 0; ib32 < QK_K/32; ib32 += 2) {
+            const __m128i q8_1_0 = _mm_loadu_si128((const __m128i *)q8); q8 += 16;
+            const __m128i q8_1_1 = _mm_loadu_si128((const __m128i *)q8); q8 += 16;
+            const __m128i q8_2_0 = _mm_loadu_si128((const __m128i *)q8); q8 += 16;
+            const __m128i q8_2_1 = _mm_loadu_si128((const __m128i *)q8); q8 += 16;
+            const __m128i q2_1_0 = _mm_set_epi64x(iq2s_grid[qs[1] | ((qh[ib32+0] << 6) & 0x300)],
+                                                  iq2s_grid[qs[0] | ((qh[ib32+0] << 8) & 0x300)]);
+            const __m128i q2_1_1 = _mm_set_epi64x(iq2s_grid[qs[3] | ((qh[ib32+0] << 2) & 0x300)],
+                                                  iq2s_grid[qs[2] | ((qh[ib32+0] << 4) & 0x300)]);
+            const __m128i q2_2_0 = _mm_set_epi64x(iq2s_grid[qs[5] | ((qh[ib32+1] << 6) & 0x300)],
+                                                  iq2s_grid[qs[4] | ((qh[ib32+1] << 8) & 0x300)]);
+            const __m128i q2_2_1 = _mm_set_epi64x(iq2s_grid[qs[7] | ((qh[ib32+1] << 2) & 0x300)],
+                                                  iq2s_grid[qs[6] | ((qh[ib32+1] << 4) & 0x300)]);
+            qs += 8;
+
+            __m128i aux128_0 = _mm_set1_epi32(signs[0] | ((uint32_t) signs[1] << 16));
+            __m128i aux128_1 = aux128_0;
+            aux128_0 = _mm_and_si128(_mm_shuffle_epi8(aux128_0,mask1_0), mask2_0);
+            aux128_1 = _mm_and_si128(_mm_shuffle_epi8(aux128_1,mask1_1), mask2_1);
+            const __m128i s2_1_0 = _mm_cmpeq_epi8(aux128_0, mask2_0);
+            const __m128i s2_1_1 = _mm_cmpeq_epi8(aux128_1, mask2_1);
+            const __m128i q8s_1_0 = _mm_sub_epi8(_mm_xor_si128(s2_1_0, q8_1_0), s2_1_0);
+            const __m128i q8s_1_1 = _mm_sub_epi8(_mm_xor_si128(s2_1_1, q8_1_1), s2_1_1);
+
+            aux128_0 = _mm_set1_epi32(signs[2] | ((uint32_t) signs[3] << 16));
+            aux128_1 = aux128_0;
+            aux128_0 = _mm_and_si128(_mm_shuffle_epi8(aux128_0,mask1_0), mask2_0);
+            aux128_1 = _mm_and_si128(_mm_shuffle_epi8(aux128_1,mask1_1), mask2_1);
+            const __m128i s2_2_0 = _mm_cmpeq_epi8(aux128_0, mask2_0);
+            const __m128i s2_2_1 = _mm_cmpeq_epi8(aux128_1, mask2_1);
+            const __m128i q8s_2_0 = _mm_sub_epi8(_mm_xor_si128(s2_2_0, q8_2_0), s2_2_0);
+            const __m128i q8s_2_1 = _mm_sub_epi8(_mm_xor_si128(s2_2_1, q8_2_1), s2_2_1);
+
+            signs += 4;
+
+            const __m128i dot1_0  = _mm_maddubs_epi16(q2_1_0, q8s_1_0);
+            const __m128i dot1_1  = _mm_maddubs_epi16(q2_1_1, q8s_1_1);
+            const __m128i dot2_0  = _mm_maddubs_epi16(q2_2_0, q8s_2_0);
+            const __m128i dot2_1  = _mm_maddubs_epi16(q2_2_1, q8s_2_1);
+
+            const __m128i p1_0 = _mm_madd_epi16(dot1_0, _mm_shuffle_epi8(scales16_0, _mm256_extractf128_si256(get_scale_shuffle_k4(ib32+0), 0)));
+            const __m128i p1_1 = _mm_madd_epi16(dot1_1, _mm_shuffle_epi8(scales16_1, _mm256_extractf128_si256(get_scale_shuffle_k4(ib32+0), 1)));
+            const __m128i p2_0 = _mm_madd_epi16(dot2_0, _mm_shuffle_epi8(scales16_0, _mm256_extractf128_si256(get_scale_shuffle_k4(ib32+1), 0)));
+            const __m128i p2_1 = _mm_madd_epi16(dot2_1, _mm_shuffle_epi8(scales16_1, _mm256_extractf128_si256(get_scale_shuffle_k4(ib32+1), 1)));
+            sumi1_0 = _mm_add_epi32(sumi1_0, p1_0);
+            sumi1_1 = _mm_add_epi32(sumi1_1, p1_1);
+            sumi2_0 = _mm_add_epi32(sumi2_0, p2_0);
+            sumi2_1 = _mm_add_epi32(sumi2_1, p2_1);
+        }
+
+        accumf = _mm256_add_ps(_mm256_mul_ps(_mm256_set1_ps(d), _mm256_cvtepi32_ps(MM256_SET_M128I(_mm_add_epi32(sumi1_1, sumi2_1), _mm_add_epi32(sumi1_0, sumi2_0)))), accumf);
+
+    }
+
+    *s = 0.125f * hsum_float_8(accumf);
+
+#elif defined(__POWER9_VECTOR__)
+    static const uint8_t k_mask1[32] = {0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x01, 0x01, 0x01, 0x01, 0x01, 0x01, 0x01, 0x01,
+                                        0x02, 0x02, 0x02, 0x02, 0x02, 0x02, 0x02, 0x02, 0x03, 0x03, 0x03, 0x03, 0x03, 0x03, 0x03, 0x03
+    };
+
+    static const uint8_t k_mask2[16] = {0x01, 0x02, 0x04, 0x08, 0x10, 0x20, 0x40, 0x80, 0x01, 0x02, 0x04, 0x08, 0x10, 0x20, 0x40, 0x80,};
+
+    const vector int v0 = vec_splats((int32_t)0);
+
+    vector float vsumf0 = vec_splats(0.0f);
+    vector float vsumf1 = vec_splats(0.0f);
+    vector float vsumf2 = vec_splats(0.0f);
+    vector float vsumf3 = vec_splats(0.0f);
+
+    const vector unsigned char mask0 = vec_xl( 0, k_mask1);
+    const vector unsigned char mask1 = vec_xl(16, k_mask1);
+    const vector signed char mask2 = (vector signed char)vec_xl( 0, k_mask2);
+
+    for (int i = 0; i < nb; ++i) {
+        vector float vxd = vec_splats(GGML_FP16_TO_FP32(x[i].d));
+        vector float vyd = vec_splats(y[i].d);
+        vector float vd = vec_mul(vxd, vyd);
+
+        vector signed int vsumi0 = v0;
+        vector signed int vsumi1 = v0;
+        vector signed int vsumi2 = v0;
+        vector signed int vsumi3 = v0;
+
+        const uint8_t *  restrict q2 = x[i].qs;
+        const uint8_t *  restrict qh = x[i].qh;
+        const uint16_t * restrict signs = (const uint16_t *)(x[i].qs + QK_K/8);
+        const uint8_t *  restrict sc = x[i].scales;
+        const int8_t  *  restrict q8 = y[i].qs;
+
+        for (int j = 0; j < QK_K/32; j += 2) {
+            __builtin_prefetch(q2, 0, 1);
+            __builtin_prefetch(q8, 0, 1);
+
+            vector signed long long aux64x2_0 = {*(const int64_t *)(iq2s_grid + (q2[0] | ((qh[0] << 8) & 0x300))), *(const int64_t *)(iq2s_grid + (q2[1] | ((qh[0] << 6) & 0x300)))};
+            vector signed long long aux64x2_1 = {*(const int64_t *)(iq2s_grid + (q2[2] | ((qh[0] << 4) & 0x300))), *(const int64_t *)(iq2s_grid + (q2[3] | ((qh[0] << 2) & 0x300)))};
+            vector signed long long aux64x2_2 = {*(const int64_t *)(iq2s_grid + (q2[4] | ((qh[1] << 8) & 0x300))), *(const int64_t *)(iq2s_grid + (q2[5] | ((qh[1] << 6) & 0x300)))};
+            vector signed long long aux64x2_3 = {*(const int64_t *)(iq2s_grid + (q2[6] | ((qh[1] << 4) & 0x300))), *(const int64_t *)(iq2s_grid + (q2[7] | ((qh[1] << 2) & 0x300)))};
+            q2 += 8;
+            qh += 2;
+
+            vector signed char vsigns01 = (vector signed char)vec_splats(*(const uint32_t *)&signs[0]);
+            vector signed char vsigns23 = (vector signed char)vec_splats(*(const uint32_t *)&signs[2]);
+            signs += 4;
+
+            vector signed char vsigns0 = vec_perm(vsigns01, vsigns01, mask0);
+            vector signed char vsigns1 = vec_perm(vsigns01, vsigns01, mask1);
+            vector signed char vsigns2 = vec_perm(vsigns23, vsigns23, mask0);
+            vector signed char vsigns3 = vec_perm(vsigns23, vsigns23, mask1);
+
+            vsigns0 = (vector signed char)vec_cmpeq(vec_and(vsigns0, mask2), mask2);
+            vsigns1 = (vector signed char)vec_cmpeq(vec_and(vsigns1, mask2), mask2);
+            vsigns2 = (vector signed char)vec_cmpeq(vec_and(vsigns2, mask2), mask2);
+            vsigns3 = (vector signed char)vec_cmpeq(vec_and(vsigns3, mask2), mask2);
+
+            vector signed char q2x0 = vec_sub(vec_xor(vsigns0, (vector signed char)aux64x2_0), vsigns0);
+            vector signed char q2x1 = vec_sub(vec_xor(vsigns1, (vector signed char)aux64x2_1), vsigns1);
+            vector signed char q2x2 = vec_sub(vec_xor(vsigns2, (vector signed char)aux64x2_2), vsigns2);
+            vector signed char q2x3 = vec_sub(vec_xor(vsigns3, (vector signed char)aux64x2_3), vsigns3);
+
+            vector signed char q8y0 = vec_xl( 0, q8);
+            vector signed char q8y1 = vec_xl(16, q8);
+            vector signed char q8y2 = vec_xl(32, q8);
+            vector signed char q8y3 = vec_xl(48, q8);
+            q8 += 64;
+
+            vector signed short qv0 = vec_add(vec_mule(q2x0, q8y0), vec_mulo(q2x0, q8y0));
+            vector signed short qv1 = vec_add(vec_mule(q2x1, q8y1), vec_mulo(q2x1, q8y1));
+            vector signed short qv2 = vec_add(vec_mule(q2x2, q8y2), vec_mulo(q2x2, q8y2));
+            vector signed short qv3 = vec_add(vec_mule(q2x3, q8y3), vec_mulo(q2x3, q8y3));
+
+            const uint16_t ls0 = (uint16_t)(sc[0] & 0xf);
+            const uint16_t ls1 = (uint16_t)(sc[0] >>  4);
+            const uint16_t ls2 = (uint16_t)(sc[1] & 0xf);
+            const uint16_t ls3 = (uint16_t)(sc[1] >>  4);
+            sc += 2;
+
+            vector signed short vscales0 = vec_splats((int16_t)(2*ls0+1));
+            vector signed short vscales1 = vec_splats((int16_t)(2*ls1+1));
+            vector signed short vscales2 = vec_splats((int16_t)(2*ls2+1));
+            vector signed short vscales3 = vec_splats((int16_t)(2*ls3+1));
+
+            vsumi0 = vec_msum(qv0, vscales0, vsumi0);
+            vsumi1 = vec_msum(qv1, vscales1, vsumi1);
+            vsumi2 = vec_msum(qv2, vscales2, vsumi2);
+            vsumi3 = vec_msum(qv3, vscales3, vsumi3);
+        }
+
+        vsumf0 = vec_madd(vec_ctf(vsumi0, 0), vd, vsumf0);
+        vsumf1 = vec_madd(vec_ctf(vsumi1, 0), vd, vsumf1);
+        vsumf2 = vec_madd(vec_ctf(vsumi2, 0), vd, vsumf2);
+        vsumf3 = vec_madd(vec_ctf(vsumi3, 0), vd, vsumf3);
+    }
+
+    vsumf0 = vec_add(vsumf0, vsumf2);
+    vsumf1 = vec_add(vsumf1, vsumf3);
+
+    vsumf0 = vec_add(vsumf0, vsumf1);
+
+    vsumf0 = vec_add(vsumf0, vec_sld(vsumf0, vsumf0, 4));
+    vsumf0 = vec_add(vsumf0, vec_sld(vsumf0, vsumf0, 8));
+
+    *s = 0.125f * vec_extract(vsumf0, 0);
+
+#elif defined(__loongarch_asx)
+
+   static const uint8_t k_mask1[32] = {0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x01, 0x01, 0x01, 0x01, 0x01, 0x01, 0x01, 0x01,
+                                       0x02, 0x02, 0x02, 0x02, 0x02, 0x02, 0x02, 0x02, 0x03, 0x03, 0x03, 0x03, 0x03, 0x03, 0x03, 0x03
+   };
+
+    static const uint8_t k_mask2[32] = {0x01, 0x02, 0x04, 0x08, 0x10, 0x20, 0x40, 0x80, 0x01, 0x02, 0x04, 0x08, 0x10, 0x20, 0x40, 0x80,
+                                        0x01, 0x02, 0x04, 0x08, 0x10, 0x20, 0x40, 0x80, 0x01, 0x02, 0x04, 0x08, 0x10, 0x20, 0x40, 0x80,
+    };
+
+
+    const __m128i m4 = __lsx_vreplgr2vr_b(0xf);
+    const __m128i m1 = __lsx_vreplgr2vr_b(1);
+
+    const __m256i mask1 = __lasx_xvld((const __m256i*)k_mask1, 0);
+    const __m256i mask2 = __lasx_xvld((const __m256i*)k_mask2, 0);
+    uint64_t aux64;
+
+    __m256 accumf = (__m256)__lasx_xvldi(0);
+    for (int i = 0; i < nb; ++i) {
+        const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
+        const uint8_t * restrict qs = x[i].qs;
+        const uint8_t * restrict qh = x[i].qh;
+        const uint16_t * restrict signs = (const uint16_t *)(x[i].qs + QK_K/8);
+        const int8_t  * restrict q8 = y[i].qs;
+
+        __m128i tmp1;
+        memcpy(&aux64, x[i].scales, 8);
+        tmp1 = __lsx_vinsgr2vr_d(tmp1, aux64, 0);
+        tmp1 = __lsx_vinsgr2vr_d(tmp1, aux64 >> 4, 1);
+        const __m128i scales8 = __lsx_vadd_b(__lsx_vslli_h(__lsx_vand_v(tmp1, m4), 1), m1);
+        const __m256i scales16 = lasx_ext8_16(scales8); // 0 2 4 6 8 10 12 14 1 3 5 7 9 11 13 15
+
+        __m256i sumi1 = __lasx_xvldi(0);
+        __m256i sumi2 = __lasx_xvldi(0);
+        for (int ib32 = 0; ib32 < QK_K/32; ib32 += 2) {
+            const __m256i q8_1 = __lasx_xvld((const __m256i *)q8, 0); q8 += 32;
+            const __m256i q8_2 = __lasx_xvld((const __m256i *)q8, 0); q8 += 32;
+            const __m256i q2_1 = lasx_set_d(iq2s_grid[qs[3] | ((qh[ib32+0] << 2) & 0x300)],
+                                                   iq2s_grid[qs[2] | ((qh[ib32+0] << 4) & 0x300)],
+                                                   iq2s_grid[qs[1] | ((qh[ib32+0] << 6) & 0x300)],
+                                                   iq2s_grid[qs[0] | ((qh[ib32+0] << 8) & 0x300)]);
+            const __m256i q2_2 = lasx_set_d(iq2s_grid[qs[7] | ((qh[ib32+1] << 2) & 0x300)],
+                                                   iq2s_grid[qs[6] | ((qh[ib32+1] << 4) & 0x300)],
+                                                   iq2s_grid[qs[5] | ((qh[ib32+1] << 6) & 0x300)],
+                                                   iq2s_grid[qs[4] | ((qh[ib32+1] << 8) & 0x300)]);
+            qs += 8;
+
+            __m256i aux256 = __lasx_xvreplgr2vr_w(signs[0] | ((uint32_t) signs[1] << 16));
+            aux256 = __lasx_xvand_v(lasx_shuffle_b(aux256,mask1), mask2);
+            const __m256i s2_1 = __lasx_xvseq_b(aux256, mask2);
+            const __m256i q8s_1 = __lasx_xvsub_b(__lasx_xvxor_v(s2_1, q8_1), s2_1);
+
+            aux256 = __lasx_xvreplgr2vr_w(signs[2] | ((uint32_t) signs[3] << 16));
+            aux256 = __lasx_xvand_v(lasx_shuffle_b(aux256,mask1), mask2);
+            const __m256i s2_2 = __lasx_xvseq_b(aux256, mask2);
+            const __m256i q8s_2 = __lasx_xvsub_b(__lasx_xvxor_v(s2_2, q8_2), s2_2);
+
+            signs += 4;
+
+            const __m256i dot1  = lasx_maddubs_h(q2_1, q8s_1); // blocks 2*ib32+0, 2*ib32+1
+            const __m256i dot2  = lasx_maddubs_h(q2_2, q8s_2); // blocks 2*ib32+2, 2*ib32+3
+
+            const __m256i p1 = lasx_madd_h(dot1, lasx_shuffle_b(scales16, get_scale_shuffle_k4(ib32+0)));
+            const __m256i p2 = lasx_madd_h(dot2, lasx_shuffle_b(scales16, get_scale_shuffle_k4(ib32+1)));
+            sumi1 = __lasx_xvadd_w(sumi1, p1);
+            sumi2 = __lasx_xvadd_w(sumi2, p2);
+        }
+
+        accumf = __lasx_xvfmadd_s(__lasx_xvreplfr2vr_s(d), __lasx_xvffint_s_w(__lasx_xvadd_w(sumi1, sumi2)), accumf);
+    }
+
+    *s = 0.125f * hsum_float_8(accumf);
+
+#else
+
+    float sumf = 0;
+    for (int i = 0; i < nb; i++) {
+
+        const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
+        const int8_t  * q8 = y[i].qs;
+        const uint8_t * qs = x[i].qs;
+        const uint8_t * qh = x[i].qh;
+        const uint8_t * signs = qs + QK_K/8;
+
+        int bsum = 0;
+        for (int ib32 = 0; ib32 < QK_K/32; ++ib32) {
+            int ls1 = 1 + 2*(x[i].scales[ib32] & 0xf);
+            int ls2 = 1 + 2*(x[i].scales[ib32] >>  4);
+            int sumi1 = 0, sumi2 = 0;
+            for (int l = 0; l < 2; ++l) {
+                const uint8_t * grid = (const uint8_t *)(iq2s_grid + (qs[l] | (qh[ib32] << (8-2*l) & 0x300)));
+                for (int j = 0; j < 8; ++j) {
+                    sumi1 += q8[j] * grid[j] * (signs[l] & kmask_iq2xs[j] ? -1 : 1);
+                }
+                q8 += 8;
+            }
+            for (int l = 2; l < 4; ++l) {
+                const uint8_t * grid = (const uint8_t *)(iq2s_grid + (qs[l] | (qh[ib32] << (8-2*l) & 0x300)));
+                for (int j = 0; j < 8; ++j) {
+                    sumi2 += q8[j] * grid[j] * (signs[l] & kmask_iq2xs[j] ? -1 : 1);
+                }
+                q8 += 8;
+            }
+            bsum += ls1 * sumi1 + ls2 * sumi2;
+            qs += 4;
+            signs += 4;
+        }
+
+        sumf += d * bsum;
+    }
+
+    *s = 0.125f * sumf;
+
+#endif
+
+}
+
+void ggml_vec_dot_iq3_xxs_q8_K(int n, float * restrict s, size_t bs, const void * restrict vx, size_t bx, const void * restrict vy, size_t by, int nrc) {
+    assert(n % QK_K == 0);
+    assert(nrc == 1);
+    UNUSED(nrc);
+    UNUSED(bx);
+    UNUSED(by);
+    UNUSED(bs);
+
+    const block_iq3_xxs * restrict x = vx;
+    const block_q8_K    * restrict y = vy;
+
+    const int nb = n / QK_K;
+
+#if defined(__ARM_NEON)
+
+    const uint64_t * signs64 = (const uint64_t *)keven_signs_q2xs;
+
+    uint32_t aux32[2];
+
+    ggml_int8x16x4_t q3s;
+    ggml_int8x16x4_t q8b;
+
+    float sumf = 0;
+    for (int i = 0; i < nb; ++i) {
+        const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
+        const uint8_t * restrict q3 = x[i].qs;
+        const uint8_t * restrict gas = x[i].qs + QK_K/4;
+        const int8_t   * restrict q8 = y[i].qs;
+        float sumf1 = 0, sumf2 = 0;
+        for (int ib32 = 0; ib32 < QK_K/32; ib32 += 2) {
+            q8b = ggml_vld1q_s8_x4(q8); q8 += 64;
+            memcpy(aux32, gas, 2*sizeof(uint32_t)); gas += 2*sizeof(uint32_t);
+            const uint32x4_t aux32x4_0 = ggml_vld1q_u32(iq3xxs_grid[q3[ 0]], iq3xxs_grid[q3[ 1]], iq3xxs_grid[q3[ 2]], iq3xxs_grid[q3[ 3]]);
+            const uint32x4_t aux32x4_1 = ggml_vld1q_u32(iq3xxs_grid[q3[ 4]], iq3xxs_grid[q3[ 5]], iq3xxs_grid[q3[ 6]], iq3xxs_grid[q3[ 7]]);
+            const uint32x4_t aux32x4_2 = ggml_vld1q_u32(iq3xxs_grid[q3[ 8]], iq3xxs_grid[q3[ 9]], iq3xxs_grid[q3[10]], iq3xxs_grid[q3[11]]);
+            const uint32x4_t aux32x4_3 = ggml_vld1q_u32(iq3xxs_grid[q3[12]], iq3xxs_grid[q3[13]], iq3xxs_grid[q3[14]], iq3xxs_grid[q3[15]]);
+            q3 += 16;
+            q3s.val[0] = vcombine_s8(vld1_s8((const void *)(signs64 + ((aux32[0] >>  0) & 127))), vld1_s8((const void *)(signs64 + ((aux32[0] >>  7) & 127))));
+            q3s.val[1] = vcombine_s8(vld1_s8((const void *)(signs64 + ((aux32[0] >> 14) & 127))), vld1_s8((const void *)(signs64 + ((aux32[0] >> 21) & 127))));
+            q3s.val[2] = vcombine_s8(vld1_s8((const void *)(signs64 + ((aux32[1] >>  0) & 127))), vld1_s8((const void *)(signs64 + ((aux32[1] >>  7) & 127))));
+            q3s.val[3] = vcombine_s8(vld1_s8((const void *)(signs64 + ((aux32[1] >> 14) & 127))), vld1_s8((const void *)(signs64 + ((aux32[1] >> 21) & 127))));
+            q3s.val[0] = vmulq_s8(q3s.val[0], vreinterpretq_s8_u32(aux32x4_0));
+            q3s.val[1] = vmulq_s8(q3s.val[1], vreinterpretq_s8_u32(aux32x4_1));
+            q3s.val[2] = vmulq_s8(q3s.val[2], vreinterpretq_s8_u32(aux32x4_2));
+            q3s.val[3] = vmulq_s8(q3s.val[3], vreinterpretq_s8_u32(aux32x4_3));
+            const int32x4_t p1 = ggml_vdotq_s32(ggml_vdotq_s32(vdupq_n_s32(0), q3s.val[0], q8b.val[0]), q3s.val[1], q8b.val[1]);
+            const int32x4_t p2 = ggml_vdotq_s32(ggml_vdotq_s32(vdupq_n_s32(0), q3s.val[2], q8b.val[2]), q3s.val[3], q8b.val[3]);
+            sumf1 += vaddvq_s32(p1) * (0.5f + (aux32[0] >> 28));
+            sumf2 += vaddvq_s32(p2) * (0.5f + (aux32[1] >> 28));
+        }
+        sumf += d*(sumf1 + sumf2);
+    }
+    *s = 0.5f * sumf;
+
+#elif defined(__AVX2__)
+
+    const uint64_t * signs64 = (const uint64_t *)keven_signs_q2xs;
+
+    uint32_t aux32[2];
+
+    __m256 accumf = _mm256_setzero_ps();
+    for (int i = 0; i < nb; ++i) {
+        const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
+        const uint8_t * restrict q3 = x[i].qs;
+        const uint8_t * restrict gas = x[i].qs + QK_K/4;
+        const int8_t  * restrict q8 = y[i].qs;
+        __m256i sumi1 = _mm256_setzero_si256();
+        __m256i sumi2 = _mm256_setzero_si256();
+        for (int ib32 = 0; ib32 < QK_K/32; ib32 += 2) {
+            const __m256i q8_1 = _mm256_loadu_si256((const __m256i *)q8); q8 += 32;
+            const __m256i q8_2 = _mm256_loadu_si256((const __m256i *)q8); q8 += 32;
+            const __m256i q2_1 = _mm256_set_epi32(iq3xxs_grid[q3[7]], iq3xxs_grid[q3[6]], iq3xxs_grid[q3[5]], iq3xxs_grid[q3[4]],
+                                                  iq3xxs_grid[q3[3]], iq3xxs_grid[q3[2]], iq3xxs_grid[q3[1]], iq3xxs_grid[q3[0]]);
+            q3 += 8;
+            const __m256i q2_2 = _mm256_set_epi32(iq3xxs_grid[q3[7]], iq3xxs_grid[q3[6]], iq3xxs_grid[q3[5]], iq3xxs_grid[q3[4]],
+                                                  iq3xxs_grid[q3[3]], iq3xxs_grid[q3[2]], iq3xxs_grid[q3[1]], iq3xxs_grid[q3[0]]);
+            q3 += 8;
+            memcpy(aux32, gas, 8); gas += 8;
+            const __m256i s2_1 = _mm256_set_epi64x(signs64[(aux32[0] >> 21) & 127], signs64[(aux32[0] >> 14) & 127],
+                                                   signs64[(aux32[0] >>  7) & 127], signs64[(aux32[0] >>  0) & 127]);
+            const __m256i s2_2 = _mm256_set_epi64x(signs64[(aux32[1] >> 21) & 127], signs64[(aux32[1] >> 14) & 127],
+                                                   signs64[(aux32[1] >>  7) & 127], signs64[(aux32[1] >>  0) & 127]);
+            const __m256i q8s_1 = _mm256_sign_epi8(q8_1, s2_1);
+            const __m256i q8s_2 = _mm256_sign_epi8(q8_2, s2_2);
+            const __m256i dot1  = _mm256_maddubs_epi16(q2_1, q8s_1);
+            const __m256i dot2  = _mm256_maddubs_epi16(q2_2, q8s_2);
+            const uint16_t ls1 = aux32[0] >> 28;
+            const uint16_t ls2 = aux32[1] >> 28;
+            const __m256i p1 = _mm256_madd_epi16(dot1, _mm256_set1_epi16(2*ls1+1));
+            const __m256i p2 = _mm256_madd_epi16(dot2, _mm256_set1_epi16(2*ls2+1));
+            sumi1 = _mm256_add_epi32(sumi1, p1);
+            sumi2 = _mm256_add_epi32(sumi2, p2);
+        }
+
+        accumf = _mm256_fmadd_ps(_mm256_set1_ps(d), _mm256_cvtepi32_ps(_mm256_add_epi32(sumi1, sumi2)), accumf);
+
+    }
+
+    *s = 0.25f * hsum_float_8(accumf);
+
+#elif defined(__AVX__)
+    const uint64_t * signs64 = (const uint64_t *)keven_signs_q2xs;
+
+    uint32_t aux32[2];
+
+    __m256 accumf = _mm256_setzero_ps();
+    for (int i = 0; i < nb; ++i) {
+        const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
+        const uint8_t * restrict q3 = x[i].qs;
+        const uint8_t * restrict gas = x[i].qs + QK_K/4;
+        const int8_t  * restrict q8 = y[i].qs;
+        __m128i sumi1_0 = _mm_setzero_si128();
+        __m128i sumi1_1 = _mm_setzero_si128();
+        __m128i sumi2_0 = _mm_setzero_si128();
+        __m128i sumi2_1 = _mm_setzero_si128();
+        for (int ib32 = 0; ib32 < QK_K/32; ib32 += 2) {
+            const __m128i q8_1_0 = _mm_loadu_si128((const __m128i *)q8); q8 += 16;
+            const __m128i q8_1_1 = _mm_loadu_si128((const __m128i *)q8); q8 += 16;
+            const __m128i q8_2_0 = _mm_loadu_si128((const __m128i *)q8); q8 += 16;
+            const __m128i q8_2_1 = _mm_loadu_si128((const __m128i *)q8); q8 += 16;
+            const __m128i q2_1_0 = _mm_set_epi32(iq3xxs_grid[q3[3]], iq3xxs_grid[q3[2]], iq3xxs_grid[q3[1]], iq3xxs_grid[q3[0]]);
+            const __m128i q2_1_1 = _mm_set_epi32(iq3xxs_grid[q3[7]], iq3xxs_grid[q3[6]], iq3xxs_grid[q3[5]], iq3xxs_grid[q3[4]]);
+            q3 += 8;
+            const __m128i q2_2_0 = _mm_set_epi32(iq3xxs_grid[q3[3]], iq3xxs_grid[q3[2]], iq3xxs_grid[q3[1]], iq3xxs_grid[q3[0]]);
+            const __m128i q2_2_1 = _mm_set_epi32(iq3xxs_grid[q3[7]], iq3xxs_grid[q3[6]], iq3xxs_grid[q3[5]], iq3xxs_grid[q3[4]]);
+            q3 += 8;
+            memcpy(aux32, gas, 8); gas += 8;
+            const __m128i s2_1_0 = _mm_set_epi64x(signs64[(aux32[0] >>  7) & 127], signs64[(aux32[0] >>  0) & 127]);
+            const __m128i s2_1_1 = _mm_set_epi64x(signs64[(aux32[0] >> 21) & 127], signs64[(aux32[0] >> 14) & 127]);
+            const __m128i s2_2_0 = _mm_set_epi64x(signs64[(aux32[1] >>  7) & 127], signs64[(aux32[1] >>  0) & 127]);
+            const __m128i s2_2_1 = _mm_set_epi64x(signs64[(aux32[1] >> 21) & 127], signs64[(aux32[1] >> 14) & 127]);
+            const __m128i q8s_1_0 = _mm_sign_epi8(q8_1_0, s2_1_0);
+            const __m128i q8s_1_1 = _mm_sign_epi8(q8_1_1, s2_1_1);
+            const __m128i q8s_2_0 = _mm_sign_epi8(q8_2_0, s2_2_0);
+            const __m128i q8s_2_1 = _mm_sign_epi8(q8_2_1, s2_2_1);
+            const __m128i dot1_0  = _mm_maddubs_epi16(q2_1_0, q8s_1_0);
+            const __m128i dot1_1  = _mm_maddubs_epi16(q2_1_1, q8s_1_1);
+            const __m128i dot2_0  = _mm_maddubs_epi16(q2_2_0, q8s_2_0);
+            const __m128i dot2_1  = _mm_maddubs_epi16(q2_2_1, q8s_2_1);
+            const uint16_t ls1 = aux32[0] >> 28;
+            const uint16_t ls2 = aux32[1] >> 28;
+            const __m128i p1_0 = _mm_madd_epi16(dot1_0, _mm_set1_epi16(2*ls1+1));
+            const __m128i p1_1 = _mm_madd_epi16(dot1_1, _mm_set1_epi16(2*ls1+1));
+            const __m128i p2_0 = _mm_madd_epi16(dot2_0, _mm_set1_epi16(2*ls2+1));
+            const __m128i p2_1 = _mm_madd_epi16(dot2_1, _mm_set1_epi16(2*ls2+1));
+            sumi1_0 = _mm_add_epi32(sumi1_0, p1_0);
+            sumi1_1 = _mm_add_epi32(sumi1_1, p1_1);
+            sumi2_0 = _mm_add_epi32(sumi2_0, p2_0);
+            sumi2_1 = _mm_add_epi32(sumi2_1, p2_1);
+        }
+
+        accumf = _mm256_add_ps(_mm256_mul_ps(_mm256_set1_ps(d), _mm256_cvtepi32_ps(MM256_SET_M128I(_mm_add_epi32(sumi1_1, sumi2_1), _mm_add_epi32(sumi1_0, sumi2_0)))), accumf);
+
+    }
+
+    *s = 0.25f * hsum_float_8(accumf);
+
+#elif defined(__POWER9_VECTOR__)
+    const uint64_t * signs64 = (const uint64_t *)keven_signs_q2xs;
+
+    const vector int v0 = vec_splats((int32_t)0);
+
+    vector float vsumf0 = vec_splats(0.0f);
+    vector float vsumf1 = vec_splats(0.0f);
+    vector float vsumf2 = vec_splats(0.0f);
+    vector float vsumf3 = vec_splats(0.0f);
+
+    for (int i = 0; i < nb; ++i) {
+        vector float vxd = vec_splats(GGML_FP16_TO_FP32(x[i].d));
+        vector float vyd = vec_splats(y[i].d);
+        vector float vd = vec_mul(vxd, vyd);
+
+        vector signed int vsumi0 = v0;
+        vector signed int vsumi1 = v0;
+        vector signed int vsumi2 = v0;
+        vector signed int vsumi3 = v0;
+
+        const uint8_t * restrict q3 = x[i].qs;
+        const uint32_t * restrict signs = (const uint32_t *)(x[i].qs + QK_K/4);
+        const int8_t  * restrict q8 = y[i].qs;
+
+#pragma GCC unroll 1
+        for (int j = 0; j < QK_K/32; j += 2) {
+            __builtin_prefetch(q3, 0, 1);
+            __builtin_prefetch(q8, 0, 1);
+
+            vector unsigned int aux32x4_0 = {iq3xxs_grid[q3[ 0]], iq3xxs_grid[q3[ 1]], iq3xxs_grid[q3[ 2]], iq3xxs_grid[q3[ 3]]};
+            vector unsigned int aux32x4_1 = {iq3xxs_grid[q3[ 4]], iq3xxs_grid[q3[ 5]], iq3xxs_grid[q3[ 6]], iq3xxs_grid[q3[ 7]]};
+            vector unsigned int aux32x4_2 = {iq3xxs_grid[q3[ 8]], iq3xxs_grid[q3[ 9]], iq3xxs_grid[q3[10]], iq3xxs_grid[q3[11]]};
+            vector unsigned int aux32x4_3 = {iq3xxs_grid[q3[12]], iq3xxs_grid[q3[13]], iq3xxs_grid[q3[14]], iq3xxs_grid[q3[15]]};
+            q3 += 16;
+
+            vector unsigned long long aux64x2_0 = {(uint64_t)(signs64[(signs[0] >>  0) & 127]), (uint64_t)(signs64[(signs[0] >>  7) & 127])};
+            vector unsigned long long aux64x2_1 = {(uint64_t)(signs64[(signs[0] >> 14) & 127]), (uint64_t)(signs64[(signs[0] >> 21) & 127])};
+            vector unsigned long long aux64x2_2 = {(uint64_t)(signs64[(signs[1] >>  0) & 127]), (uint64_t)(signs64[(signs[1] >>  7) & 127])};
+            vector unsigned long long aux64x2_3 = {(uint64_t)(signs64[(signs[1] >> 14) & 127]), (uint64_t)(signs64[(signs[1] >> 21) & 127])};
+
+            vector signed char q3x0 = vec_mul((vector signed char)aux64x2_0, (vector signed char)aux32x4_0);
+            vector signed char q3x1 = vec_mul((vector signed char)aux64x2_1, (vector signed char)aux32x4_1);
+            vector signed char q3x2 = vec_mul((vector signed char)aux64x2_2, (vector signed char)aux32x4_2);
+            vector signed char q3x3 = vec_mul((vector signed char)aux64x2_3, (vector signed char)aux32x4_3);
+
+            vector signed char q8y0 = vec_xl( 0, q8);
+            vector signed char q8y1 = vec_xl(16, q8);
+            vector signed char q8y2 = vec_xl(32, q8);
+            vector signed char q8y3 = vec_xl(48, q8);
+            q8 += 64;
+
+            vector signed short qv0 = vec_add(vec_mule(q3x0, q8y0), vec_mulo(q3x0, q8y0));
+            vector signed short qv1 = vec_add(vec_mule(q3x1, q8y1), vec_mulo(q3x1, q8y1));
+            vector signed short qv2 = vec_add(vec_mule(q3x2, q8y2), vec_mulo(q3x2, q8y2));
+            vector signed short qv3 = vec_add(vec_mule(q3x3, q8y3), vec_mulo(q3x3, q8y3));
+
+            const uint16_t ls0 = (uint16_t)(signs[0] >> 28);
+            const uint16_t ls1 = (uint16_t)(signs[1] >> 28);
+            signs += 2;
+
+            vector signed short vscales01 = (vector signed short)vec_splats((uint16_t)(2*ls0+1));
+            vector signed short vscales23 = (vector signed short)vec_splats((uint16_t)(2*ls1+1));
+
+            vsumi0 = vec_msum(qv0, vscales01, vsumi0);
+            vsumi1 = vec_msum(qv1, vscales01, vsumi1);
+            vsumi2 = vec_msum(qv2, vscales23, vsumi2);
+            vsumi3 = vec_msum(qv3, vscales23, vsumi3);
+        }
+
+        vsumf0 = vec_madd(vec_ctf(vsumi0, 0), vd, vsumf0);
+        vsumf1 = vec_madd(vec_ctf(vsumi1, 0), vd, vsumf1);
+        vsumf2 = vec_madd(vec_ctf(vsumi2, 0), vd, vsumf2);
+        vsumf3 = vec_madd(vec_ctf(vsumi3, 0), vd, vsumf3);
+    }
+
+    vsumf0 = vec_add(vsumf0, vsumf2);
+    vsumf1 = vec_add(vsumf1, vsumf3);
+
+    vsumf0 = vec_add(vsumf0, vsumf1);
+
+    vsumf0 = vec_add(vsumf0, vec_sld(vsumf0, vsumf0, 4));
+    vsumf0 = vec_add(vsumf0, vec_sld(vsumf0, vsumf0, 8));
+
+    *s = 0.25f * vec_extract(vsumf0, 0);
+
+#elif defined(__loongarch_asx)
+
+    const uint64_t * signs64 = (const uint64_t *)keven_signs_q2xs;
+
+    uint32_t aux32[2];
+
+    __m256 accumf = (__m256)__lasx_xvldi(0);
+    for (int i = 0; i < nb; ++i) {
+        const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
+        const uint8_t * restrict q3 = x[i].qs;
+        const uint8_t * restrict gas = x[i].qs + QK_K/4;
+        const int8_t  * restrict q8 = y[i].qs;
+        __m256i sumi1 = __lasx_xvldi(0);
+        __m256i sumi2 = __lasx_xvldi(0);
+        for (int ib32 = 0; ib32 < QK_K/32; ib32 += 2) {
+            const __m256i q8_1 = __lasx_xvld((const __m256i *)q8, 0); q8 += 32;
+            const __m256i q8_2 = __lasx_xvld((const __m256i *)q8, 0); q8 += 32;
+            const __m256i q2_1 = lasx_set_w(iq3xxs_grid[q3[7]], iq3xxs_grid[q3[6]], iq3xxs_grid[q3[5]], iq3xxs_grid[q3[4]],
+                                                iq3xxs_grid[q3[3]], iq3xxs_grid[q3[2]], iq3xxs_grid[q3[1]], iq3xxs_grid[q3[0]]);
+            q3 += 8;
+            const __m256i q2_2 = lasx_set_w(iq3xxs_grid[q3[7]], iq3xxs_grid[q3[6]], iq3xxs_grid[q3[5]], iq3xxs_grid[q3[4]],
+                                                iq3xxs_grid[q3[3]], iq3xxs_grid[q3[2]], iq3xxs_grid[q3[1]], iq3xxs_grid[q3[0]]);
+            q3 += 8;
+            memcpy(aux32, gas, 8); gas += 8;
+
+            const __m256i s2_1 = lasx_set_d(signs64[(aux32[0] >> 21) & 127], signs64[(aux32[0] >> 14) & 127],
+                                                   signs64[(aux32[0] >>  7) & 127], signs64[(aux32[0] >>  0) & 127]);
+            const __m256i s2_2 = lasx_set_d(signs64[(aux32[1] >> 21) & 127], signs64[(aux32[1] >> 14) & 127],
+                                                   signs64[(aux32[1] >>  7) & 127], signs64[(aux32[1] >>  0) & 127]);
+            const __m256i q8s_1 = __lasx_xvsigncov_b(s2_1, q8_1);
+            const __m256i q8s_2 = __lasx_xvsigncov_b(s2_2, q8_2);
+            const __m256i dot1  = lasx_maddubs_h(q2_1, q8s_1);
+            const __m256i dot2  = lasx_maddubs_h(q2_2, q8s_2);
+            const uint16_t ls1 = aux32[0] >> 28;
+            const uint16_t ls2 = aux32[1] >> 28;
+
+            const __m256i p1 = lasx_madd_h(dot1, __lasx_xvreplgr2vr_h(2*ls1+1));
+            const __m256i p2 = lasx_madd_h(dot2, __lasx_xvreplgr2vr_h(2*ls2+1));
+            sumi1 = __lasx_xvadd_w(sumi1, p1);
+            sumi2 = __lasx_xvadd_w(sumi2, p2);
+        }
+
+        accumf = __lasx_xvfmadd_s(__lasx_xvreplfr2vr_s(d), __lasx_xvffint_s_w(__lasx_xvadd_w(sumi1, sumi2)), accumf);
+    }
+
+    *s = 0.25f * hsum_float_8(accumf);
+
+#else
+
+    uint32_t aux32;
+
+    float sumf = 0.f;
+    for (int i = 0; i < nb; ++i) {
+        const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
+        const uint8_t * restrict q3 = x[i].qs;
+        const uint8_t * restrict gas = x[i].qs + QK_K/4;
+        const int8_t  * restrict q8 = y[i].qs;
+        int32_t bsum = 0;
+        for (int ib32 = 0; ib32 < QK_K/32; ++ib32) {
+            memcpy(&aux32, gas, sizeof(uint32_t)); gas += sizeof(uint32_t);
+            const uint32_t ls = 2*(aux32 >> 28) + 1;
+            int32_t sumi = 0;
+            for (int l = 0; l < 4; ++l) {
+                const uint8_t * grid1 = (const uint8_t *)(iq3xxs_grid + q3[2*l+0]);
+                const uint8_t * grid2 = (const uint8_t *)(iq3xxs_grid + q3[2*l+1]);
+                const uint8_t  signs = ksigns_iq2xs[(aux32 >> 7*l) & 127];
+                for (int j = 0; j < 4; ++j) {
+                    sumi += grid1[j] * q8[j+0] * (signs & kmask_iq2xs[j+0] ? -1 : 1);
+                    sumi += grid2[j] * q8[j+4] * (signs & kmask_iq2xs[j+4] ? -1 : 1);
+                }
+                q8 += 8;
+            }
+            q3 += 8;
+            bsum += sumi * ls;
+        }
+        sumf += d * bsum;
+    }
+    *s = 0.25f * sumf;
+#endif
+}
+
+void ggml_vec_dot_iq3_s_q8_K (int n, float * restrict s, size_t bs, const void * restrict vx, size_t bx, const void * restrict vy, size_t by, int nrc) {
+    assert(n % QK_K == 0);
+    assert(nrc == 1);
+    UNUSED(nrc);
+    UNUSED(bx);
+    UNUSED(by);
+    UNUSED(bs);
+
+    const block_iq3_s * restrict x = vx;
+    const block_q8_K  * restrict y = vy;
+
+    const int nb = n / QK_K;
+
+#if defined(__ARM_NEON)
+
+    typedef union {
+        uint16x8_t vec_index;
+        uint16_t   index[8];
+    } vec_index_t;
+
+   static const uint8_t k_mask1[32] = {0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x01, 0x01, 0x01, 0x01, 0x01, 0x01, 0x01, 0x01,
+                                       0x02, 0x02, 0x02, 0x02, 0x02, 0x02, 0x02, 0x02, 0x03, 0x03, 0x03, 0x03, 0x03, 0x03, 0x03, 0x03
+   };
+
+    static const uint8_t k_mask2[16] = {0x01, 0x02, 0x04, 0x08, 0x10, 0x20, 0x40, 0x80, 0x01, 0x02, 0x04, 0x08, 0x10, 0x20, 0x40, 0x80,};
+
+    static const int16_t k_shift[8] = {8, 7, 6, 5, 4, 3, 2, 1};
+
+    const ggml_uint8x16x2_t mask1 = ggml_vld1q_u8_x2(k_mask1);
+    const uint8x16_t        mask2 = vld1q_u8(k_mask2);
+
+    const int16x8_t  hshift = vld1q_s16(k_shift);
+    const uint16x8_t m256   = vdupq_n_u16(256);
+    const uint8x16_t m1     = vdupq_n_u8(1);
+
+    uint8x16x2_t vs;
+    ggml_int8x16x4_t q3s;
+    ggml_int8x16x4_t q8b;
+    vec_index_t idx;
+
+    uint32_t scales32[2];
+    const uint8_t * scales8 = (const uint8_t *)scales32;
+
+    float sumf = 0;
+    for (int i = 0; i < nb; ++i) {
+        const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
+        const uint8_t * restrict qs = x[i].qs;
+        const uint8_t * restrict qh = x[i].qh;
+        const uint16_t * restrict signs = (const uint16_t *)x[i].signs;
+        const int8_t   * restrict q8 = y[i].qs;
+
+        memcpy(scales32, x[i].scales, 4);
+        scales32[1] = (((scales32[0] >> 4) & 0x0f0f0f0f) << 1) | 0x01010101;
+        scales32[0] = ((scales32[0] & 0x0f0f0f0f) << 1) | 0x01010101;
+
+        int sumi1 = 0, sumi2 = 0;
+        for (int ib32 = 0; ib32 < QK_K/32; ib32 += 2) {
+            q8b = ggml_vld1q_s8_x4(q8); q8 += 64;
+
+            const uint8x16_t idx_l = vld1q_u8(qs); qs += 16;
+            idx.vec_index = vorrq_u16(vmovl_u8(vget_low_u8 (idx_l)), vandq_u16(vshlq_u16(vdupq_n_u16(qh[ib32+0]), hshift), m256));
+            const uint32x4_t aux32x4_0 = ggml_vld1q_u32(iq3s_grid[idx.index[0]], iq3s_grid[idx.index[1]],
+                                                        iq3s_grid[idx.index[2]], iq3s_grid[idx.index[3]]);
+            const uint32x4_t aux32x4_1 = ggml_vld1q_u32(iq3s_grid[idx.index[4]], iq3s_grid[idx.index[5]],
+                                                        iq3s_grid[idx.index[6]], iq3s_grid[idx.index[7]]);
+            idx.vec_index = vorrq_u16(vmovl_u8(vget_high_u8(idx_l)), vandq_u16(vshlq_u16(vdupq_n_u16(qh[ib32+1]), hshift), m256));
+            const uint32x4_t aux32x4_2 = ggml_vld1q_u32(iq3s_grid[idx.index[0]], iq3s_grid[idx.index[1]],
+                                                        iq3s_grid[idx.index[2]], iq3s_grid[idx.index[3]]);
+            const uint32x4_t aux32x4_3 = ggml_vld1q_u32(iq3s_grid[idx.index[4]], iq3s_grid[idx.index[5]],
+                                                        iq3s_grid[idx.index[6]], iq3s_grid[idx.index[7]]);
+
+
+            vs.val[0] = vreinterpretq_u8_u32(vdupq_n_u32(signs[0] | ((uint32_t) signs[1] << 16)));
+            vs.val[1] = vandq_u8(ggml_vqtbl1q_u8(vs.val[0], mask1.val[1]), mask2);
+            vs.val[0] = vandq_u8(ggml_vqtbl1q_u8(vs.val[0], mask1.val[0]), mask2);
+            vs.val[0] = vorrq_u8(vceqq_u8(vs.val[0], mask2), m1);
+            vs.val[1] = vorrq_u8(vceqq_u8(vs.val[1], mask2), m1);
+
+            q3s.val[0] = vmulq_s8(vreinterpretq_s8_u8(vs.val[0]), vreinterpretq_s8_u32(aux32x4_0));
+            q3s.val[1] = vmulq_s8(vreinterpretq_s8_u8(vs.val[1]), vreinterpretq_s8_u32(aux32x4_1));
+
+            vs.val[0] = vreinterpretq_u8_u32(vdupq_n_u32(signs[2] | ((uint32_t) signs[3] << 16)));
+            vs.val[1] = vandq_u8(ggml_vqtbl1q_u8(vs.val[0], mask1.val[1]), mask2);
+            vs.val[0] = vandq_u8(ggml_vqtbl1q_u8(vs.val[0], mask1.val[0]), mask2);
+            vs.val[0] = vorrq_u8(vceqq_u8(vs.val[0], mask2), m1);
+            vs.val[1] = vorrq_u8(vceqq_u8(vs.val[1], mask2), m1);
+
+            signs += 4;
+
+            q3s.val[2] = vmulq_s8(vreinterpretq_s8_u8(vs.val[0]), vreinterpretq_s8_u32(aux32x4_2));
+            q3s.val[3] = vmulq_s8(vreinterpretq_s8_u8(vs.val[1]), vreinterpretq_s8_u32(aux32x4_3));
+
+            const int32x4_t p1 = ggml_vdotq_s32(ggml_vdotq_s32(vdupq_n_s32(0), q3s.val[0], q8b.val[0]), q3s.val[1], q8b.val[1]);
+            const int32x4_t p2 = ggml_vdotq_s32(ggml_vdotq_s32(vdupq_n_s32(0), q3s.val[2], q8b.val[2]), q3s.val[3], q8b.val[3]);
+
+            sumi1 += vaddvq_s32(p1) * scales8[ib32/2+0];
+            sumi2 += vaddvq_s32(p2) * scales8[ib32/2+4];
+        }
+        sumf += d*(sumi1 + sumi2);
+    }
+    *s = sumf;
+
+#elif defined(__AVX2__)
+
+   static const uint8_t k_mask1[32] = {0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x01, 0x01, 0x01, 0x01, 0x01, 0x01, 0x01, 0x01,
+                                       0x02, 0x02, 0x02, 0x02, 0x02, 0x02, 0x02, 0x02, 0x03, 0x03, 0x03, 0x03, 0x03, 0x03, 0x03, 0x03
+   };
+
+    static const uint8_t k_mask2[32] = {0x01, 0x02, 0x04, 0x08, 0x10, 0x20, 0x40, 0x80, 0x01, 0x02, 0x04, 0x08, 0x10, 0x20, 0x40, 0x80,
+                                        0x01, 0x02, 0x04, 0x08, 0x10, 0x20, 0x40, 0x80, 0x01, 0x02, 0x04, 0x08, 0x10, 0x20, 0x40, 0x80,
+    };
+
+    const __m256i mask1 = _mm256_loadu_si256((const __m256i*)k_mask1);
+    const __m256i mask2 = _mm256_loadu_si256((const __m256i*)k_mask2);
+
+    const __m256i idx_shift = _mm256_set_epi32(1, 2, 3, 4, 5, 6, 7, 8);
+    const __m256i idx_mask  = _mm256_set1_epi32(256);
+
+    typedef union {
+        __m256i  vec[2];
+        uint32_t index[16];
+    } index_t;
+
+    index_t idx;
+
+    __m256 accumf = _mm256_setzero_ps();
+    for (int i = 0; i < nb; ++i) {
+        const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
+        const uint8_t * restrict qs = x[i].qs;
+        const uint8_t * restrict qh = x[i].qh;
+        const uint16_t * restrict signs = (const uint16_t *)x[i].signs;
+        const int8_t  * restrict q8 = y[i].qs;
+        __m256i sumi1 = _mm256_setzero_si256();
+        __m256i sumi2 = _mm256_setzero_si256();
+        for (int ib32 = 0; ib32 < QK_K/32; ib32 += 2) {
+            const __m256i q8_1 = _mm256_loadu_si256((const __m256i *)q8); q8 += 32;
+            const __m256i q8_2 = _mm256_loadu_si256((const __m256i *)q8); q8 += 32;
+            const __m256i idx_l = _mm256_cvtepu8_epi16(_mm_loadu_si128((const __m128i *)qs)); qs += 16;
+            idx.vec[0] = _mm256_set1_epi32(qh[ib32+0]);
+            idx.vec[1] = _mm256_set1_epi32(qh[ib32+1]);
+            idx.vec[0] = _mm256_and_si256(_mm256_sllv_epi32(idx.vec[0], idx_shift), idx_mask);
+            idx.vec[1] = _mm256_and_si256(_mm256_sllv_epi32(idx.vec[1], idx_shift), idx_mask);
+            idx.vec[0] = _mm256_or_si256(idx.vec[0], _mm256_cvtepi16_epi32(_mm256_castsi256_si128(idx_l)));
+            idx.vec[1] = _mm256_or_si256(idx.vec[1], _mm256_cvtepi16_epi32(_mm256_extractf128_si256(idx_l, 1)));
+
+            // At leat on my CPU (Ryzen 7950X), using _mm256_i32gather_epi32 is slower than _mm256_set_epi32. Strange.
+            //const __m256i q2_1 = _mm256_i32gather_epi32((const int *)iq3s_grid, idx.vec[0], 4);
+            //const __m256i q2_2 = _mm256_i32gather_epi32((const int *)iq3s_grid, idx.vec[1], 4);
+            const __m256i q2_1 = _mm256_set_epi32(
+                    iq3s_grid[idx.index[7]], iq3s_grid[idx.index[6]], iq3s_grid[idx.index[5]], iq3s_grid[idx.index[4]],
+                    iq3s_grid[idx.index[3]], iq3s_grid[idx.index[2]], iq3s_grid[idx.index[1]], iq3s_grid[idx.index[0]]
+            );
+            const __m256i q2_2 = _mm256_set_epi32(
+                    iq3s_grid[idx.index[15]], iq3s_grid[idx.index[14]], iq3s_grid[idx.index[13]], iq3s_grid[idx.index[12]],
+                    iq3s_grid[idx.index[11]], iq3s_grid[idx.index[10]], iq3s_grid[idx.index[ 9]], iq3s_grid[idx.index[ 8]]
+            );
+
+            __m256i aux256 = _mm256_set1_epi32(signs[0] | (signs[1] << 16));
+            aux256 = _mm256_and_si256(_mm256_shuffle_epi8(aux256,mask1), mask2);
+            const __m256i s2_1 = _mm256_cmpeq_epi8(aux256, mask2);
+            const __m256i q8s_1 = _mm256_sub_epi8(_mm256_xor_si256(s2_1, q8_1), s2_1);
+
+            aux256 = _mm256_set1_epi32(signs[2] | (signs[3] << 16));
+            aux256 = _mm256_and_si256(_mm256_shuffle_epi8(aux256,mask1), mask2);
+            const __m256i s2_2 = _mm256_cmpeq_epi8(aux256, mask2);
+            const __m256i q8s_2 = _mm256_sub_epi8(_mm256_xor_si256(s2_2, q8_2), s2_2);
+
+            signs += 4;
+
+            const __m256i dot1  = _mm256_maddubs_epi16(q2_1, q8s_1);
+            const __m256i dot2  = _mm256_maddubs_epi16(q2_2, q8s_2);
+            const uint16_t ls1 = x[i].scales[ib32/2] & 0xf;
+            const uint16_t ls2 = x[i].scales[ib32/2] >>  4;
+            const __m256i p1 = _mm256_madd_epi16(dot1, _mm256_set1_epi16(2*ls1+1));
+            const __m256i p2 = _mm256_madd_epi16(dot2, _mm256_set1_epi16(2*ls2+1));
+            sumi1 = _mm256_add_epi32(sumi1, p1);
+            sumi2 = _mm256_add_epi32(sumi2, p2);
+        }
+
+        accumf = _mm256_fmadd_ps(_mm256_set1_ps(d), _mm256_cvtepi32_ps(_mm256_add_epi32(sumi1, sumi2)), accumf);
+
+    }
+
+    *s = hsum_float_8(accumf);
+
+#elif defined(__AVX__)
+   static const uint8_t k_mask1[32] = {0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x01, 0x01, 0x01, 0x01, 0x01, 0x01, 0x01, 0x01,
+                                       0x02, 0x02, 0x02, 0x02, 0x02, 0x02, 0x02, 0x02, 0x03, 0x03, 0x03, 0x03, 0x03, 0x03, 0x03, 0x03
+   };
+
+    static const uint8_t k_mask2[32] = {0x01, 0x02, 0x04, 0x08, 0x10, 0x20, 0x40, 0x80, 0x01, 0x02, 0x04, 0x08, 0x10, 0x20, 0x40, 0x80,
+                                        0x01, 0x02, 0x04, 0x08, 0x10, 0x20, 0x40, 0x80, 0x01, 0x02, 0x04, 0x08, 0x10, 0x20, 0x40, 0x80,
+    };
+
+    const __m128i mask1_0 = _mm_loadu_si128((const __m128i*)k_mask1);
+    const __m128i mask1_1 = _mm_loadu_si128((const __m128i*)k_mask1 + 1);
+    const __m128i mask2_0 = _mm_loadu_si128((const __m128i*)k_mask2);
+    const __m128i mask2_1 = _mm_loadu_si128((const __m128i*)k_mask2 + 1);
+
+    const __m128i idx_mul_0 = _mm_set_epi32(32, 64, 128, 256);
+    const __m128i idx_mul_1 = _mm_set_epi32(2, 4, 8, 16);
+    const __m128i idx_mask  = _mm_set1_epi32(256);
+
+    typedef union {
+        __m128i  vec[4];
+        uint32_t index[16];
+    } index_t;
+
+    index_t idx;
+
+    __m256 accumf = _mm256_setzero_ps();
+    for (int i = 0; i < nb; ++i) {
+        const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
+        const uint8_t * restrict qs = x[i].qs;
+        const uint8_t * restrict qh = x[i].qh;
+        const uint16_t * restrict signs = (const uint16_t *)x[i].signs;
+        const int8_t  * restrict q8 = y[i].qs;
+        __m128i sumi1_0 = _mm_setzero_si128();
+        __m128i sumi1_1 = _mm_setzero_si128();
+        __m128i sumi2_0 = _mm_setzero_si128();
+        __m128i sumi2_1 = _mm_setzero_si128();
+        for (int ib32 = 0; ib32 < QK_K/32; ib32 += 2) {
+            const __m128i q8_1_0 = _mm_loadu_si128((const __m128i *)q8); q8 += 16;
+            const __m128i q8_1_1 = _mm_loadu_si128((const __m128i *)q8); q8 += 16;
+            const __m128i q8_2_0 = _mm_loadu_si128((const __m128i *)q8); q8 += 16;
+            const __m128i q8_2_1 = _mm_loadu_si128((const __m128i *)q8); q8 += 16;
+            const __m128i qs_tmp = _mm_loadu_si128((const __m128i *)qs);
+            const __m128i idx_l_0 = _mm_cvtepu8_epi16(qs_tmp);
+            const __m128i idx_l_1 = _mm_cvtepu8_epi16(_mm_srli_si128(qs_tmp, 8)); qs += 16;
+            idx.vec[0] = _mm_set1_epi32(qh[ib32+0]);
+            idx.vec[1] = idx.vec[0];
+            idx.vec[2] = _mm_set1_epi32(qh[ib32+1]);
+            idx.vec[3] = idx.vec[2];
+
+            idx.vec[0] = _mm_and_si128(_mm_mullo_epi32(idx.vec[0], idx_mul_0), idx_mask);
+            idx.vec[1] = _mm_and_si128(_mm_mullo_epi32(idx.vec[1], idx_mul_1), idx_mask);
+            idx.vec[2] = _mm_and_si128(_mm_mullo_epi32(idx.vec[2], idx_mul_0), idx_mask);
+            idx.vec[3] = _mm_and_si128(_mm_mullo_epi32(idx.vec[3], idx_mul_1), idx_mask);
+
+            idx.vec[0] = _mm_or_si128(idx.vec[0], _mm_cvtepi16_epi32(idx_l_0));
+            idx.vec[1] = _mm_or_si128(idx.vec[1], _mm_cvtepi16_epi32(_mm_srli_si128(idx_l_0, 8)));
+            idx.vec[2] = _mm_or_si128(idx.vec[2], _mm_cvtepi16_epi32(idx_l_1));
+            idx.vec[3] = _mm_or_si128(idx.vec[3], _mm_cvtepi16_epi32(_mm_srli_si128(idx_l_1, 8)));
+
+            const __m128i q2_1_0 = _mm_set_epi32(iq3s_grid[idx.index[3]], iq3s_grid[idx.index[2]], iq3s_grid[idx.index[1]], iq3s_grid[idx.index[0]]);
+            const __m128i q2_1_1 = _mm_set_epi32(iq3s_grid[idx.index[7]], iq3s_grid[idx.index[6]], iq3s_grid[idx.index[5]], iq3s_grid[idx.index[4]]);
+            const __m128i q2_2_0 = _mm_set_epi32(iq3s_grid[idx.index[11]], iq3s_grid[idx.index[10]], iq3s_grid[idx.index[9]], iq3s_grid[idx.index[8]]);
+            const __m128i q2_2_1 = _mm_set_epi32(iq3s_grid[idx.index[15]], iq3s_grid[idx.index[14]], iq3s_grid[idx.index[13]], iq3s_grid[idx.index[12]]);
+
+            __m128i aux128_0 = _mm_set1_epi32(signs[0] | (signs[1] << 16));
+            __m128i aux128_1 = aux128_0;
+            aux128_0 = _mm_and_si128(_mm_shuffle_epi8(aux128_0,mask1_0), mask2_0);
+            aux128_1 = _mm_and_si128(_mm_shuffle_epi8(aux128_1,mask1_1), mask2_1);
+            const __m128i s2_1_0 = _mm_cmpeq_epi8(aux128_0, mask2_0);
+            const __m128i s2_1_1 = _mm_cmpeq_epi8(aux128_1, mask2_1);
+            const __m128i q8s_1_0 = _mm_sub_epi8(_mm_xor_si128(s2_1_0, q8_1_0), s2_1_0);
+            const __m128i q8s_1_1 = _mm_sub_epi8(_mm_xor_si128(s2_1_1, q8_1_1), s2_1_1);
+
+            aux128_0 = _mm_set1_epi32(signs[2] | (signs[3] << 16));
+            aux128_1 = aux128_0;
+            aux128_0 = _mm_and_si128(_mm_shuffle_epi8(aux128_0,mask1_0), mask2_0);
+            aux128_1 = _mm_and_si128(_mm_shuffle_epi8(aux128_1,mask1_1), mask2_1);
+            const __m128i s2_2_0 = _mm_cmpeq_epi8(aux128_0, mask2_0);
+            const __m128i s2_2_1 = _mm_cmpeq_epi8(aux128_1, mask2_1);
+            const __m128i q8s_2_0 = _mm_sub_epi8(_mm_xor_si128(s2_2_0, q8_2_0), s2_2_0);
+            const __m128i q8s_2_1 = _mm_sub_epi8(_mm_xor_si128(s2_2_1, q8_2_1), s2_2_1);
+
+            signs += 4;
+
+            const __m128i dot1_0  = _mm_maddubs_epi16(q2_1_0, q8s_1_0);
+            const __m128i dot1_1  = _mm_maddubs_epi16(q2_1_1, q8s_1_1);
+            const __m128i dot2_0  = _mm_maddubs_epi16(q2_2_0, q8s_2_0);
+            const __m128i dot2_1  = _mm_maddubs_epi16(q2_2_1, q8s_2_1);
+            const uint16_t ls1 = x[i].scales[ib32/2] & 0xf;
+            const uint16_t ls2 = x[i].scales[ib32/2] >>  4;
+            const __m128i p1_0 = _mm_madd_epi16(dot1_0, _mm_set1_epi16(2*ls1+1));
+            const __m128i p1_1 = _mm_madd_epi16(dot1_1, _mm_set1_epi16(2*ls1+1));
+            const __m128i p2_0 = _mm_madd_epi16(dot2_0, _mm_set1_epi16(2*ls2+1));
+            const __m128i p2_1 = _mm_madd_epi16(dot2_1, _mm_set1_epi16(2*ls2+1));
+            sumi1_0 = _mm_add_epi32(sumi1_0, p1_0);
+            sumi1_1 = _mm_add_epi32(sumi1_1, p1_1);
+            sumi2_0 = _mm_add_epi32(sumi2_0, p2_0);
+            sumi2_1 = _mm_add_epi32(sumi2_1, p2_1);
+        }
+
+        accumf = _mm256_add_ps(_mm256_mul_ps(_mm256_set1_ps(d), _mm256_cvtepi32_ps(MM256_SET_M128I(_mm_add_epi32(sumi1_1, sumi2_1), _mm_add_epi32(sumi1_0, sumi2_0)))), accumf);
+
+    }
+
+    *s = hsum_float_8(accumf);
+
+#elif defined(__POWER9_VECTOR__)
+    static const uint8_t k_mask1[32] = {0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x01, 0x01, 0x01, 0x01, 0x01, 0x01, 0x01, 0x01,
+                                        0x02, 0x02, 0x02, 0x02, 0x02, 0x02, 0x02, 0x02, 0x03, 0x03, 0x03, 0x03, 0x03, 0x03, 0x03, 0x03
+    };
+
+    static const uint8_t k_mask2[16] = {0x01, 0x02, 0x04, 0x08, 0x10, 0x20, 0x40, 0x80, 0x01, 0x02, 0x04, 0x08, 0x10, 0x20, 0x40, 0x80,};
+
+    const vector int v0 = vec_splats((int32_t)0);
+
+    vector float vsumf0 = vec_splats(0.0f);
+    vector float vsumf1 = vec_splats(0.0f);
+    vector float vsumf2 = vec_splats(0.0f);
+    vector float vsumf3 = vec_splats(0.0f);
+
+    const vector unsigned char mask0 = vec_xl( 0, k_mask1);
+    const vector unsigned char mask1 = vec_xl(16, k_mask1);
+    const vector signed char mask2 = (vector signed char)vec_xl( 0, k_mask2);
+
+    for (int i = 0; i < nb; ++i) {
+        vector float vxd = vec_splats(GGML_FP16_TO_FP32(x[i].d));
+        vector float vyd = vec_splats(y[i].d);
+        vector float vd = vec_mul(vxd, vyd);
+
+        const uint8_t *  restrict q3 = x[i].qs;
+        const uint8_t *  restrict qh = x[i].qh;
+        const uint16_t * restrict signs = (const uint16_t *)(x[i].signs);
+        const uint8_t *  restrict sc = x[i].scales;
+        const int8_t  *  restrict q8 = y[i].qs;
+
+        vector signed int vsumi0 = v0;
+        vector signed int vsumi1 = v0;
+        vector signed int vsumi2 = v0;
+        vector signed int vsumi3 = v0;
+
+        for (int j = 0; j < QK_K/32; j += 2) {
+            __builtin_prefetch(q3, 0, 1);
+            __builtin_prefetch(q8, 0, 1);
+
+            vector unsigned int aux32x4_0 = {iq3s_grid[q3[ 0] | ((qh[0] << 8) & 256)], iq3s_grid[q3[ 1] | ((qh[0] << 7) & 256)],
+                                             iq3s_grid[q3[ 2] | ((qh[0] << 6) & 256)], iq3s_grid[q3[ 3] | ((qh[0] << 5) & 256)]};
+            vector unsigned int aux32x4_1 = {iq3s_grid[q3[ 4] | ((qh[0] << 4) & 256)], iq3s_grid[q3[ 5] | ((qh[0] << 3) & 256)],
+                                             iq3s_grid[q3[ 6] | ((qh[0] << 2) & 256)], iq3s_grid[q3[ 7] | ((qh[0] << 1) & 256)]};
+            vector unsigned int aux32x4_2 = {iq3s_grid[q3[ 8] | ((qh[1] << 8) & 256)], iq3s_grid[q3[ 9] | ((qh[1] << 7) & 256)],
+                                             iq3s_grid[q3[10] | ((qh[1] << 6) & 256)], iq3s_grid[q3[11] | ((qh[1] << 5) & 256)]};
+            vector unsigned int aux32x4_3 = {iq3s_grid[q3[12] | ((qh[1] << 4) & 256)], iq3s_grid[q3[13] | ((qh[1] << 3) & 256)],
+                                             iq3s_grid[q3[14] | ((qh[1] << 2) & 256)], iq3s_grid[q3[15] | ((qh[1] << 1) & 256)]};
+            q3 += 16;
+            qh += 2;
+
+            vector signed char vsigns01 = (vector signed char)vec_splats(*(const uint32_t *)&signs[0]);
+            vector signed char vsigns02 = (vector signed char)vec_splats(*(const uint32_t *)&signs[2]);
+            signs += 4;
+
+            vector signed char vsigns0 = vec_perm(vsigns01, vsigns01, mask0);
+            vector signed char vsigns1 = vec_perm(vsigns01, vsigns01, mask1);
+            vector signed char vsigns2 = vec_perm(vsigns02, vsigns02, mask0);
+            vector signed char vsigns3 = vec_perm(vsigns02, vsigns02, mask1);
+
+            vsigns0 = (vector signed char)vec_cmpeq(vec_and(vsigns0, mask2), mask2);
+            vsigns1 = (vector signed char)vec_cmpeq(vec_and(vsigns1, mask2), mask2);
+            vsigns2 = (vector signed char)vec_cmpeq(vec_and(vsigns2, mask2), mask2);
+            vsigns3 = (vector signed char)vec_cmpeq(vec_and(vsigns3, mask2), mask2);
+
+            vector signed char q3x0 = vec_sub(vec_xor(vsigns0, (vector signed char)aux32x4_0), vsigns0);
+            vector signed char q3x1 = vec_sub(vec_xor(vsigns1, (vector signed char)aux32x4_1), vsigns1);
+            vector signed char q3x2 = vec_sub(vec_xor(vsigns2, (vector signed char)aux32x4_2), vsigns2);
+            vector signed char q3x3 = vec_sub(vec_xor(vsigns3, (vector signed char)aux32x4_3), vsigns3);
+
+            vector signed char q8y0 = vec_xl( 0, q8);
+            vector signed char q8y1 = vec_xl(16, q8);
+            vector signed char q8y2 = vec_xl(32, q8);
+            vector signed char q8y3 = vec_xl(48, q8);
+            q8 += 64;
+
+            vector signed short qv0 = vec_add(vec_mule(q3x0, q8y0), vec_mulo(q3x0, q8y0));
+            vector signed short qv1 = vec_add(vec_mule(q3x1, q8y1), vec_mulo(q3x1, q8y1));
+            vector signed short qv2 = vec_add(vec_mule(q3x2, q8y2), vec_mulo(q3x2, q8y2));
+            vector signed short qv3 = vec_add(vec_mule(q3x3, q8y3), vec_mulo(q3x3, q8y3));
+
+            const uint16_t ls0 = (uint16_t)(sc[0] & 0xf);
+            const uint16_t ls1 = (uint16_t)(sc[0] >>  4);
+            sc ++;
+
+            vector signed short vscales01 = (vector signed short)vec_splats((uint16_t)(2*ls0+1));
+            vector signed short vscales23 = (vector signed short)vec_splats((uint16_t)(2*ls1+1));
+
+            vsumi0 = vec_msum(qv0, vscales01, vsumi0);
+            vsumi1 = vec_msum(qv1, vscales01, vsumi1);
+            vsumi2 = vec_msum(qv2, vscales23, vsumi2);
+            vsumi3 = vec_msum(qv3, vscales23, vsumi3);
+        }
+
+        vsumf0 = vec_madd(vec_ctf(vsumi0, 0), vd, vsumf0);
+        vsumf1 = vec_madd(vec_ctf(vsumi1, 0), vd, vsumf1);
+        vsumf2 = vec_madd(vec_ctf(vsumi2, 0), vd, vsumf2);
+        vsumf3 = vec_madd(vec_ctf(vsumi3, 0), vd, vsumf3);
+    }
+
+    vsumf0 = vec_add(vsumf0, vsumf2);
+    vsumf1 = vec_add(vsumf1, vsumf3);
+
+    vsumf0 = vec_add(vsumf0, vsumf1);
+
+    vsumf0 = vec_add(vsumf0, vec_sld(vsumf0, vsumf0, 4));
+    vsumf0 = vec_add(vsumf0, vec_sld(vsumf0, vsumf0, 8));
+
+    *s = vec_extract(vsumf0, 0);
+
+#elif defined(__loongarch_asx)
+
+   static const uint8_t k_mask1[32] = {0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x01, 0x01, 0x01, 0x01, 0x01, 0x01, 0x01, 0x01,
+                                       0x02, 0x02, 0x02, 0x02, 0x02, 0x02, 0x02, 0x02, 0x03, 0x03, 0x03, 0x03, 0x03, 0x03, 0x03, 0x03
+   };
+
+    static const uint8_t k_mask2[32] = {0x01, 0x02, 0x04, 0x08, 0x10, 0x20, 0x40, 0x80, 0x01, 0x02, 0x04, 0x08, 0x10, 0x20, 0x40, 0x80,
+                                        0x01, 0x02, 0x04, 0x08, 0x10, 0x20, 0x40, 0x80, 0x01, 0x02, 0x04, 0x08, 0x10, 0x20, 0x40, 0x80,
+    };
+
+    const __m256i mask1 = __lasx_xvld((const __m256i*)k_mask1, 0);
+    const __m256i mask2 = __lasx_xvld((const __m256i*)k_mask2, 0);
+
+    __m256i idx_shift = lasx_set_w(1, 2, 3, 4, 5, 6, 7, 8);
+    const __m256i idx_mask  = __lasx_xvreplgr2vr_w(256);
+
+    typedef union {
+        __m256i  vec[2];
+        uint32_t index[16];
+    } index_t;
+
+    index_t idx;
+
+    __m256 accumf = (__m256)__lasx_xvldi(0);
+    for (int i = 0; i < nb; ++i) {
+        const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
+        const uint8_t * restrict qs = x[i].qs;
+        const uint8_t * restrict qh = x[i].qh;
+        const uint16_t * restrict signs = (const uint16_t *)x[i].signs;
+        const int8_t  * restrict q8 = y[i].qs;
+        __m256i sumi1 = __lasx_xvldi(0);
+        __m256i sumi2 = __lasx_xvldi(0);
+        for (int ib32 = 0; ib32 < QK_K/32; ib32 += 2) {
+            const __m256i q8_1 = __lasx_xvld((const __m256i *)q8, 0); q8 += 32;
+            const __m256i q8_2 = __lasx_xvld((const __m256i *)q8, 0); q8 += 32;
+            const __m256i idx_l = lasx_extu8_16(__lsx_vld(qs, 0)); qs += 16;
+            idx.vec[0] = __lasx_xvreplgr2vr_w(qh[ib32+0]);
+            idx.vec[1] = __lasx_xvreplgr2vr_w(qh[ib32+1]);
+            idx.vec[0] = __lasx_xvand_v(__lasx_xvsll_w(idx.vec[0], idx_shift), idx_mask);
+            idx.vec[1] = __lasx_xvand_v(__lasx_xvsll_w(idx.vec[1], idx_shift), idx_mask);
+            idx.vec[0] = __lasx_xvor_v(idx.vec[0], lasx_ext16_32(lasx_extracti128(idx_l, 0)));
+            idx.vec[1] = __lasx_xvor_v(idx.vec[1], lasx_ext16_32(lasx_extracti128(idx_l, 1)));
+
+            // At leat on my CPU (Ryzen 7950X), using _mm256_i32gather_epi32 is slower than _mm256_set_epi32. Strange.
+            //const __m256i q2_1 = _mm256_i32gather_epi32((const int *)iq3s_grid, idx.vec[0], 4);
+            //const __m256i q2_2 = _mm256_i32gather_epi32((const int *)iq3s_grid, idx.vec[1], 4);
+            const __m256i q2_1 = lasx_set_w(
+                    iq3s_grid[idx.index[7]], iq3s_grid[idx.index[6]], iq3s_grid[idx.index[5]], iq3s_grid[idx.index[4]],
+                    iq3s_grid[idx.index[3]], iq3s_grid[idx.index[2]], iq3s_grid[idx.index[1]], iq3s_grid[idx.index[0]]
+            );
+            const __m256i q2_2 = lasx_set_w(
+                    iq3s_grid[idx.index[15]], iq3s_grid[idx.index[14]], iq3s_grid[idx.index[13]], iq3s_grid[idx.index[12]],
+                    iq3s_grid[idx.index[11]], iq3s_grid[idx.index[10]], iq3s_grid[idx.index[ 9]], iq3s_grid[idx.index[ 8]]
+            );
+
+            __m256i aux256 = __lasx_xvreplgr2vr_w(signs[0] | (signs[1] << 16));
+            aux256 = __lasx_xvand_v(lasx_shuffle_b(aux256,mask1), mask2);
+            const __m256i s2_1 = __lasx_xvseq_b(aux256, mask2);
+            const __m256i q8s_1 = __lasx_xvsub_b(__lasx_xvxor_v(s2_1, q8_1), s2_1);
+
+            aux256 = __lasx_xvreplgr2vr_w(signs[2] | (signs[3] << 16));
+            aux256 = __lasx_xvand_v(lasx_shuffle_b(aux256,mask1), mask2);
+            const __m256i s2_2 = __lasx_xvseq_b(aux256, mask2);
+            const __m256i q8s_2 = __lasx_xvsub_b(__lasx_xvxor_v(s2_2, q8_2), s2_2);
+
+            signs += 4;
+
+            const __m256i dot1 = lasx_maddubs_h(q2_1, q8s_1);
+            const __m256i dot2  = lasx_maddubs_h(q2_2, q8s_2);
+            const uint16_t ls1 = x[i].scales[ib32/2] & 0xf;
+            const uint16_t ls2 = x[i].scales[ib32/2] >>  4;
+            const __m256i p1 = lasx_madd_h(dot1, __lasx_xvreplgr2vr_h(2*ls1+1));
+            const __m256i p2 = lasx_madd_h(dot2, __lasx_xvreplgr2vr_h(2*ls2+1));
+            sumi1 = __lasx_xvadd_w(sumi1, p1);
+            sumi2 = __lasx_xvadd_w(sumi2, p2);
+        }
+
+        accumf = __lasx_xvfmadd_s(__lasx_xvreplfr2vr_s(d), __lasx_xvffint_s_w(__lasx_xvadd_w(sumi1, sumi2)), accumf);
+    }
+
+    *s = hsum_float_8(accumf);
+
+#else
+
+    float sumf = 0.f;
+    for (int i = 0; i < nb; ++i) {
+        const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
+        const uint8_t * restrict qs = x[i].qs;
+        const uint8_t * restrict qh = x[i].qh;
+        const uint8_t * restrict signs = x[i].signs;
+        const int8_t  * restrict q8 = y[i].qs;
+        int32_t bsum = 0;
+        for (int ib32 = 0; ib32 < QK_K/32; ib32 += 2) {
+            const uint32_t ls1 = 2*(x[i].scales[ib32/2] & 0xf) + 1;
+            const uint32_t ls2 = 2*(x[i].scales[ib32/2] >>  4) + 1;
+            int32_t sumi = 0;
+            for (int l = 0; l < 4; ++l) {
+                const uint8_t * grid1 = (const uint8_t *)(iq3s_grid + (qs[2*l+0] | ((qh[ib32+0] << (8-2*l)) & 256)));
+                const uint8_t * grid2 = (const uint8_t *)(iq3s_grid + (qs[2*l+1] | ((qh[ib32+0] << (7-2*l)) & 256)));
+                for (int j = 0; j < 4; ++j) {
+                    sumi += grid1[j] * q8[j+0] * (signs[l] & kmask_iq2xs[j+0] ? -1 : 1);
+                    sumi += grid2[j] * q8[j+4] * (signs[l] & kmask_iq2xs[j+4] ? -1 : 1);
+                }
+                q8 += 8;
+            }
+            qs += 8;
+            signs += 4;
+            bsum += sumi * ls1;
+            sumi = 0;
+            for (int l = 0; l < 4; ++l) {
+                const uint8_t * grid1 = (const uint8_t *)(iq3s_grid + (qs[2*l+0] | ((qh[ib32+1] << (8-2*l)) & 256)));
+                const uint8_t * grid2 = (const uint8_t *)(iq3s_grid + (qs[2*l+1] | ((qh[ib32+1] << (7-2*l)) & 256)));
+                for (int j = 0; j < 4; ++j) {
+                    sumi += grid1[j] * q8[j+0] * (signs[l] & kmask_iq2xs[j+0] ? -1 : 1);
+                    sumi += grid2[j] * q8[j+4] * (signs[l] & kmask_iq2xs[j+4] ? -1 : 1);
+                }
+                q8 += 8;
+            }
+            qs += 8;
+            signs += 4;
+            bsum += sumi * ls2;
+        }
+        sumf += d * bsum;
+    }
+    *s = sumf;
+#endif
+}
+
+#if defined(__AVX2__)
+static inline __m256i mul_add_epi8(const __m256i x, const __m256i y) {
+    const __m256i ax = _mm256_sign_epi8(x, x);
+    const __m256i sy = _mm256_sign_epi8(y, x);
+    return _mm256_maddubs_epi16(ax, sy);
+}
+#elif defined(__loongarch_asx)
+static inline __m256i mul_add_epi8(const __m256i x, const __m256i y) {
+    const __m256i ax = __lasx_xvsigncov_b(x, x);
+    const __m256i sy = __lasx_xvsigncov_b(x, y);
+    __m256i tmp1, tmp2, tmp3;
+    tmp1 = __lasx_xvmulwev_h_bu_b(ax, sy);
+    tmp2 = __lasx_xvmulwod_h_bu_b(ax, sy);
+    tmp3 = __lasx_xvadd_h(tmp1, tmp2);
+    return __lasx_xvsat_h(tmp3, 15);
+}
+#endif
+
+void ggml_vec_dot_iq1_s_q8_K  (int n, float * restrict s, size_t bs, const void * restrict vx, size_t bx, const void * restrict vy, size_t by, int nrc) {
+    assert(n % QK_K == 0);
+    assert(nrc == 1);
+    UNUSED(nrc);
+    UNUSED(bx);
+    UNUSED(by);
+    UNUSED(bs);
+
+    const block_iq1_s * restrict x = vx;
+    const block_q8_K  * restrict y = vy;
+
+    const int nb = n / QK_K;
+
+#if defined __ARM_NEON
+
+    ggml_int8x16x4_t q1b;
+    ggml_int8x16x4_t q8b;
+
+    float sumf = 0;
+    for (int i = 0; i < nb; ++i) {
+
+        const int8_t   * q8 = y[i].qs;
+        const uint8_t  * qs = x[i].qs;
+        const uint16_t * qh = x[i].qh;
+
+        int sumi1 = 0, sumi2 = 0, sumi3 = 0;
+
+        for (int ib = 0; ib < QK_K/32; ib += 2) {
+
+            q1b.val[0] = vcombine_s8(vld1_s8((const int8_t *)(iq1s_grid + (qs[0] | ((qh[ib+0] << 8) & 0x700)))),
+                                     vld1_s8((const int8_t *)(iq1s_grid + (qs[1] | ((qh[ib+0] << 5) & 0x700)))));
+            q1b.val[1] = vcombine_s8(vld1_s8((const int8_t *)(iq1s_grid + (qs[2] | ((qh[ib+0] << 2) & 0x700)))),
+                                     vld1_s8((const int8_t *)(iq1s_grid + (qs[3] | ((qh[ib+0] >> 1) & 0x700)))));
+            q1b.val[2] = vcombine_s8(vld1_s8((const int8_t *)(iq1s_grid + (qs[4] | ((qh[ib+1] << 8) & 0x700)))),
+                                     vld1_s8((const int8_t *)(iq1s_grid + (qs[5] | ((qh[ib+1] << 5) & 0x700)))));
+            q1b.val[3] = vcombine_s8(vld1_s8((const int8_t *)(iq1s_grid + (qs[6] | ((qh[ib+1] << 2) & 0x700)))),
+                                     vld1_s8((const int8_t *)(iq1s_grid + (qs[7] | ((qh[ib+1] >> 1) & 0x700)))));
+            qs += 8;
+
+            q8b = ggml_vld1q_s8_x4(q8); q8 += 64;
+
+            const int32x4_t p1 = ggml_vdotq_s32(ggml_vdotq_s32(vdupq_n_s32(0), q1b.val[0], q8b.val[0]), q1b.val[1], q8b.val[1]);
+            const int32x4_t p2 = ggml_vdotq_s32(ggml_vdotq_s32(vdupq_n_s32(0), q1b.val[2], q8b.val[2]), q1b.val[3], q8b.val[3]);
+
+            const int ls1 = 2*((qh[ib+0] >> 12) & 7) + 1;
+            const int ls2 = 2*((qh[ib+1] >> 12) & 7) + 1;
+            sumi1 += vaddvq_s32(p1) * ls1;
+            sumi2 += vaddvq_s32(p2) * ls2;
+            sumi3 += (y[i].bsums[2*ib+0] + y[i].bsums[2*ib+1]) * ls1 * (qh[ib+0] & 0x8000 ? -1 : 1)
+                   + (y[i].bsums[2*ib+2] + y[i].bsums[2*ib+3]) * ls2 * (qh[ib+1] & 0x8000 ? -1 : 1);
+
+        }
+
+        sumf += y[i].d * GGML_FP16_TO_FP32(x[i].d) * (sumi1 + sumi2 + IQ1S_DELTA * sumi3);
+    }
+
+    *s = sumf;
+
+#elif defined __AVX2__
+
+    __m256 accum = _mm256_setzero_ps();
+    float accum1 = 0;
+    for (int i = 0; i < nb; ++i) {
+
+        const int8_t   * q8 = y[i].qs;
+        const uint8_t  * qs = x[i].qs;
+        const uint16_t * qh = x[i].qh;
+
+        __m256i sumi = _mm256_setzero_si256();
+        int sumi1 = 0;
+        for (int ib = 0; ib < QK_K/32; ib += 2) {
+            const __m256i q1b_1 = _mm256_set_epi64x(iq1s_grid[qs[3] | ((qh[ib+0] >> 1) & 0x700)], iq1s_grid[qs[2] | ((qh[ib+0] << 2) & 0x700)],
+                                                    iq1s_grid[qs[1] | ((qh[ib+0] << 5) & 0x700)], iq1s_grid[qs[0] | ((qh[ib+0] << 8) & 0x700)]);
+            const __m256i q1b_2 = _mm256_set_epi64x(iq1s_grid[qs[7] | ((qh[ib+1] >> 1) & 0x700)], iq1s_grid[qs[6] | ((qh[ib+1] << 2) & 0x700)],
+                                                    iq1s_grid[qs[5] | ((qh[ib+1] << 5) & 0x700)], iq1s_grid[qs[4] | ((qh[ib+1] << 8) & 0x700)]);
+            qs += 8;
+            const __m256i q8b_1 = _mm256_loadu_si256((const __m256i*)q8); q8 += 32;
+            const __m256i q8b_2 = _mm256_loadu_si256((const __m256i*)q8); q8 += 32;
+
+            const __m256i dot1 = mul_add_epi8(q1b_1, q8b_1);
+            const __m256i dot2 = mul_add_epi8(q1b_2, q8b_2);
+            const int16_t ls1 = 2*((qh[ib+0] >> 12) & 7) + 1;
+            const int16_t ls2 = 2*((qh[ib+1] >> 12) & 7) + 1;
+            const __m256i p1 = _mm256_madd_epi16(dot1, _mm256_set1_epi16(ls1));
+            const __m256i p2 = _mm256_madd_epi16(dot2, _mm256_set1_epi16(ls2));
+
+            sumi = _mm256_add_epi32(sumi, _mm256_add_epi32(p1, p2));
+            sumi1 += (y[i].bsums[2*ib+0] + y[i].bsums[2*ib+1]) * (qh[ib+0] & 0x8000 ? -1 : 1) * ls1
+                   + (y[i].bsums[2*ib+2] + y[i].bsums[2*ib+3]) * (qh[ib+1] & 0x8000 ? -1 : 1) * ls2;
+        }
+
+        const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d);
+        accum = _mm256_fmadd_ps(_mm256_set1_ps(d), _mm256_cvtepi32_ps(sumi), accum);
+        accum1 += d * sumi1;
+
+    }
+
+    *s = hsum_float_8(accum) + IQ1S_DELTA * accum1;
+
+#elif defined __AVX__
+    __m256 accum = _mm256_setzero_ps();
+    float accum1 = 0;
+    for (int i = 0; i < nb; ++i) {
+
+        const int8_t   * q8 = y[i].qs;
+        const uint8_t  * qs = x[i].qs;
+        const uint16_t * qh = x[i].qh;
+
+        __m128i sumi1_0 = _mm_setzero_si128();
+        __m128i sumi1_1 = _mm_setzero_si128();
+        int sumi1 = 0;
+        for (int ib = 0; ib < QK_K/32; ib += 2) {
+            const __m128i q1b_1_0 = _mm_set_epi64x(iq1s_grid[qs[1] | ((qh[ib+0] << 5) & 0x700)], iq1s_grid[qs[0] | ((qh[ib+0] << 8) & 0x700)]);
+            const __m128i q1b_1_1 = _mm_set_epi64x(iq1s_grid[qs[3] | ((qh[ib+0] >> 1) & 0x700)], iq1s_grid[qs[2] | ((qh[ib+0] << 2) & 0x700)]);
+            const __m128i q1b_2_0 = _mm_set_epi64x(iq1s_grid[qs[5] | ((qh[ib+1] << 5) & 0x700)], iq1s_grid[qs[4] | ((qh[ib+1] << 8) & 0x700)]);
+            const __m128i q1b_2_1 = _mm_set_epi64x(iq1s_grid[qs[7] | ((qh[ib+1] >> 1) & 0x700)], iq1s_grid[qs[6] | ((qh[ib+1] << 2) & 0x700)]);
+            qs += 8;
+            const __m128i q8b_1_0 = _mm_loadu_si128((const __m128i *)q8); q8 += 16;
+            const __m128i q8b_1_1 = _mm_loadu_si128((const __m128i *)q8); q8 += 16;
+            const __m128i q8b_2_0 = _mm_loadu_si128((const __m128i *)q8); q8 += 16;
+            const __m128i q8b_2_1 = _mm_loadu_si128((const __m128i *)q8); q8 += 16;
+
+            const __m128i dot1_0 = mul_add_epi8_sse(q1b_1_0, q8b_1_0);
+            const __m128i dot1_1 = mul_add_epi8_sse(q1b_1_1, q8b_1_1);
+            const __m128i dot2_0 = mul_add_epi8_sse(q1b_2_0, q8b_2_0);
+            const __m128i dot2_1 = mul_add_epi8_sse(q1b_2_1, q8b_2_1);
+            const int16_t ls1 = 2*((qh[ib+0] >> 12) & 7) + 1;
+            const int16_t ls2 = 2*((qh[ib+1] >> 12) & 7) + 1;
+            const __m128i p1_0 = _mm_madd_epi16(dot1_0, _mm_set1_epi16(ls1));
+            const __m128i p1_1 = _mm_madd_epi16(dot1_1, _mm_set1_epi16(ls1));
+            const __m128i p2_0 = _mm_madd_epi16(dot2_0, _mm_set1_epi16(ls2));
+            const __m128i p2_1 = _mm_madd_epi16(dot2_1, _mm_set1_epi16(ls2));
+
+            sumi1_0 = _mm_add_epi32(sumi1_0, _mm_add_epi32(p1_0, p2_0));
+            sumi1_1 = _mm_add_epi32(sumi1_1, _mm_add_epi32(p1_1, p2_1));
+            sumi1 += (y[i].bsums[2*ib+0] + y[i].bsums[2*ib+1]) * (qh[ib+0] & 0x8000 ? -1 : 1) * ls1
+                   + (y[i].bsums[2*ib+2] + y[i].bsums[2*ib+3]) * (qh[ib+1] & 0x8000 ? -1 : 1) * ls2;
+        }
+
+        const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d);
+        accum = _mm256_add_ps(_mm256_mul_ps(_mm256_set1_ps(d), _mm256_cvtepi32_ps(MM256_SET_M128I(sumi1_1, sumi1_0))), accum);
+        accum1 += d * sumi1;
+
+    }
+
+    *s = hsum_float_8(accum) + IQ1S_DELTA * accum1;
+
+#elif defined(__POWER9_VECTOR__)
+    const vector unsigned char v0 = vec_splats((unsigned char)0x0);
+    const vector unsigned short vsign = vec_splats((unsigned short)0x8000);
+
+    vector float vsumf0 = vec_splats(0.0f);
+    vector float vsumf1 = vec_splats(0.0f);
+    vector float vsumf2 = vec_splats(0.0f);
+    vector float vsumf3 = vec_splats(0.0f);
+
+    for (int i = 0; i < nb; ++i) {
+        vector float vxd = vec_splats(GGML_FP16_TO_FP32(x[i].d));
+        vector float vyd = vec_splats(y[i].d);
+        vector float vd = vec_mul(vxd, vyd);
+
+        vector signed int vsumi0 = vec_splats((int32_t)0);
+        vector signed int vsumi1 = vec_splats((int32_t)0);
+        vector signed int vsumi2 = vec_splats((int32_t)0);
+        vector signed int vsumi3 = vec_splats((int32_t)0);
+        vector signed int vsumi8 = vec_splats((int32_t)0);
+
+        const uint8_t  * restrict q1 = x[i].qs;
+        const uint16_t * restrict qh = x[i].qh;
+        const int8_t   * restrict q8 = y[i].qs;
+        const int16_t  * restrict qs = y[i].bsums;
+
+        for (int j = 0; j < QK_K/32; j += 2) {
+            __builtin_prefetch(q1, 0, 1);
+            __builtin_prefetch(qh, 0, 1);
+            __builtin_prefetch(q8, 0, 1);
+
+            vector signed long long aux64x2_0 = {*(const int64_t *)(iq1s_grid + (q1[0] | ((qh[0] << 8) & 0x700))), *(const int64_t *)(iq1s_grid + (q1[1] | ((qh[0] << 5) & 0x700)))};
+            vector signed long long aux64x2_1 = {*(const int64_t *)(iq1s_grid + (q1[2] | ((qh[0] << 2) & 0x700))), *(const int64_t *)(iq1s_grid + (q1[3] | ((qh[0] >> 1) & 0x700)))};
+            vector signed long long aux64x2_2 = {*(const int64_t *)(iq1s_grid + (q1[4] | ((qh[1] << 8) & 0x700))), *(const int64_t *)(iq1s_grid + (q1[5] | ((qh[1] << 5) & 0x700)))};
+            vector signed long long aux64x2_3 = {*(const int64_t *)(iq1s_grid + (q1[6] | ((qh[1] << 2) & 0x700))), *(const int64_t *)(iq1s_grid + (q1[7] | ((qh[1] >> 1) & 0x700)))};
+            q1 += 8;
+
+            vector signed char q1x0 = (vector signed char)aux64x2_0;
+            vector signed char q1x1 = (vector signed char)aux64x2_1;
+            vector signed char q1x2 = (vector signed char)aux64x2_2;
+            vector signed char q1x3 = (vector signed char)aux64x2_3;
+
+            vector signed char q8y0 = vec_xl( 0, q8);
+            vector signed char q8y1 = vec_xl(16, q8);
+            vector signed char q8y2 = vec_xl(32, q8);
+            vector signed char q8y3 = vec_xl(48, q8);
+            q8 += 64;
+
+            vector signed short qv0 = vec_add(vec_mule(q1x0, q8y0), vec_mulo(q1x0, q8y0));
+            vector signed short qv1 = vec_add(vec_mule(q1x1, q8y1), vec_mulo(q1x1, q8y1));
+            vector signed short qv2 = vec_add(vec_mule(q1x2, q8y2), vec_mulo(q1x2, q8y2));
+            vector signed short qv3 = vec_add(vec_mule(q1x3, q8y3), vec_mulo(q1x3, q8y3));
+
+            const uint16_t ls0 = (uint16_t)((qh[0] >> 12) & 7);
+            const uint16_t ls1 = (uint16_t)((qh[1] >> 12) & 7);
+
+            vector signed short vscales01 = (vector signed short)vec_splats((uint16_t)(2*ls0+1));
+            vector signed short vscales23 = (vector signed short)vec_splats((uint16_t)(2*ls1+1));
+            vector signed short vscales = vec_sld(vscales23, vscales01, 8);
+
+            vsumi0 = vec_msum(qv0, vscales01, vsumi0);
+            vsumi1 = vec_msum(qv1, vscales01, vsumi1);
+            vsumi2 = vec_msum(qv2, vscales23, vsumi2);
+            vsumi3 = vec_msum(qv3, vscales23, vsumi3);
+
+            vector signed short q8ysums = vec_xl_len(qs, 8);
+            qs += 4;
+            q8ysums = vec_mergeh(q8ysums, (vector signed short)v0);
+
+            vector signed short qxh = (vector signed short)vec_sld(vec_splats(qh[1]), vec_splats(qh[0]), 8);
+            qh += 2;
+            vector __bool short vsel = vec_cmpge(qxh, (vector signed short)v0);
+
+            vector signed short q8ysum = vec_sel((vector signed short)vec_xor((vector unsigned short)q8ysums, vsign), q8ysums, vsel);
+
+            vsumi8 = vec_add(vec_mule(q8ysum, vscales), vsumi8);
+        }
+
+        vsumf0 = vec_madd(vec_ctf(vsumi0, 0), vd, vsumf0);
+        vsumf1 = vec_madd(vec_ctf(vsumi1, 0), vd, vsumf1);
+        vsumf2 = vec_madd(vec_ctf(vsumi2, 0), vd, vsumf2);
+        vsumf3 = vec_madd(vec_ctf(vsumi3, 0), vd, vsumf3);
+
+        vsumf0 = vec_madd(vec_ctf(vsumi8, 0), vec_mul(vd, vec_splats(IQ1S_DELTA)), vsumf0);
+    }
+
+    vsumf0 = vec_add(vsumf0, vsumf2);
+    vsumf1 = vec_add(vsumf1, vsumf3);
+
+    vsumf0 = vec_add(vsumf0, vsumf1);
+
+    vsumf0 = vec_add(vsumf0, vec_sld(vsumf0, vsumf0, 4));
+    vsumf0 = vec_add(vsumf0, vec_sld(vsumf0, vsumf0, 8));
+
+    *s = vec_extract(vsumf0, 0);
+
+#elif defined(__loongarch_asx)
+
+    __m256 accum = (__m256)__lasx_xvldi(0);
+    float accum1 = 0;
+    for (int i = 0; i < nb; ++i) {
+
+        const int8_t   * q8 = y[i].qs;
+        const uint8_t  * qs = x[i].qs;
+        const uint16_t * qh = x[i].qh;
+
+        __m256i sumi = __lasx_xvldi(0);
+        int sumi1 = 0;
+        for (int ib = 0; ib < QK_K/32; ib += 2) {
+            __m256i q1b_1 = __lasx_xvinsgr2vr_d(q1b_1, iq1s_grid[qs[0] | ((qh[ib+0] << 8) & 0x700)], 0);
+            q1b_1 = __lasx_xvinsgr2vr_d(q1b_1, iq1s_grid[qs[1] | ((qh[ib+0] << 5) & 0x700)], 1);
+            q1b_1 = __lasx_xvinsgr2vr_d(q1b_1, iq1s_grid[qs[2] | ((qh[ib+0] << 2) & 0x700)], 2);
+            q1b_1 = __lasx_xvinsgr2vr_d(q1b_1, iq1s_grid[qs[3] | ((qh[ib+0] >> 1) & 0x700)], 3);
+
+            __m256i q1b_2 = __lasx_xvinsgr2vr_d(q1b_2, iq1s_grid[qs[4] | ((qh[ib+1] << 8) & 0x700)], 0);
+            q1b_2 = __lasx_xvinsgr2vr_d(q1b_2, iq1s_grid[qs[5] | ((qh[ib+1] << 5) & 0x700)], 1);
+            q1b_2 = __lasx_xvinsgr2vr_d(q1b_2, iq1s_grid[qs[6] | ((qh[ib+1] << 2) & 0x700)], 2);
+            q1b_2 = __lasx_xvinsgr2vr_d(q1b_2, iq1s_grid[qs[7] | ((qh[ib+1] >> 1) & 0x700)], 3);
+
+            qs += 8;
+            const __m256i q8b_1 = __lasx_xvld((const __m256i*)q8, 0); q8 += 32;
+            const __m256i q8b_2 = __lasx_xvld((const __m256i*)q8, 0); q8 += 32;
+
+            const __m256i dot1 = mul_add_epi8(q1b_1, q8b_1);
+            const __m256i dot2 = mul_add_epi8(q1b_2, q8b_2);
+            const int16_t ls1 = 2*((qh[ib+0] >> 12) & 7) + 1;
+            const int16_t ls2 = 2*((qh[ib+1] >> 12) & 7) + 1;
+
+            __m256i tmp1, tmp5, tmp6;
+            tmp1 = __lasx_xvreplgr2vr_h(ls1);
+            tmp5 = __lasx_xvmulwev_w_h(dot1, tmp1);
+            tmp6 = __lasx_xvmulwod_w_h(dot1, tmp1);
+            const __m256i p1 = __lasx_xvadd_w(tmp5, tmp6);
+
+            tmp1 = __lasx_xvreplgr2vr_h(ls2);
+            tmp5 = __lasx_xvmulwev_w_h(dot2, tmp1);
+            tmp6 = __lasx_xvmulwod_w_h(dot2, tmp1);
+            const __m256i p2 = __lasx_xvadd_w(tmp5, tmp6);
+
+            sumi = __lasx_xvadd_w(sumi, __lasx_xvadd_w(p1, p2));
+            sumi1 += (y[i].bsums[2*ib+0] + y[i].bsums[2*ib+1]) * (qh[ib+0] & 0x8000 ? -1 : 1) * ls1
+                   + (y[i].bsums[2*ib+2] + y[i].bsums[2*ib+3]) * (qh[ib+1] & 0x8000 ? -1 : 1) * ls2;
+        }
+
+        const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d);
+        accum = __lasx_xvfmadd_s(__lasx_xvreplfr2vr_s(d), __lasx_xvffint_s_w(sumi), accum);
+        accum1 += d * sumi1;
+    }
+
+    *s = hsum_float_8(accum) + IQ1S_DELTA * accum1;
+
+#else
+
+    float sumf = 0;
+    for (int i = 0; i < nb; i++) {
+
+        const int8_t   * q8 = y[i].qs;
+        const uint8_t  * qs = x[i].qs;
+        const uint16_t * qh = x[i].qh;
+
+        int sumi = 0, sumi1 = 0;
+        for (int ib = 0; ib < QK_K/32; ++ib) {
+            const int ls = 2*((qh[ib] >> 12) & 7) + 1;
+            const int delta = qh[ib] & 0x8000 ? -1 : 1;
+            int lsum = 0;
+            for (int l = 0; l < 4; ++l) {
+                const int8_t * grid = (const int8_t *)(iq1s_grid + (qs[l] | (((qh[ib] >> 3*l) & 7) << 8)));
+                for (int j = 0; j < 8; ++j) {
+                    lsum += q8[j] * grid[j];
+                }
+                q8 += 8;
+            }
+            sumi  += ls * lsum;
+            sumi1 += ls * delta * (y[i].bsums[2*ib+0] + y[i].bsums[2*ib+1]);
+            qs += 4;
+        }
+
+        sumf += GGML_FP16_TO_FP32(x[i].d) * y[i].d * (sumi + IQ1S_DELTA * sumi1);
+    }
+
+    *s = sumf;
+
+#endif
+}
+
+void ggml_vec_dot_iq1_m_q8_K  (int n, float * restrict s, size_t bs, const void * restrict vx, size_t bx, const void * restrict vy, size_t by, int nrc) {
+    assert(n % QK_K == 0);
+    assert(nrc == 1);
+    UNUSED(nrc);
+    UNUSED(bx);
+    UNUSED(by);
+    UNUSED(bs);
+
+    const block_iq1_m * restrict x = vx;
+    const block_q8_K  * restrict y = vy;
+
+    const int nb = n / QK_K;
+
+    iq1m_scale_t scale;
+
+#if defined __ARM_NEON
+    const int32x4_t mask  = vdupq_n_s32(0x7);
+    const int32x4_t mone  = vdupq_n_s32(1);
+    const int32x4_t mzero = vdupq_n_s32(0);
+
+    ggml_int8x16x4_t deltas;
+    deltas.val[0] = vcombine_s8(vdup_n_s8(+1), vdup_n_s8(+1));
+    deltas.val[1] = vcombine_s8(vdup_n_s8(-1), vdup_n_s8(+1));
+    deltas.val[2] = vcombine_s8(vdup_n_s8(+1), vdup_n_s8(-1));
+    deltas.val[3] = vcombine_s8(vdup_n_s8(-1), vdup_n_s8(-1));
+
+    ggml_int8x16x4_t q1b;
+    ggml_int8x16x4_t q8b;
+
+    uint32_t aux32;
+    const uint8_t * aux8 = (const uint8_t *)&aux32;
+
+    float sumf = 0;
+    for (int i = 0; i < nb; ++i) {
+
+        const int8_t   * q8 = y[i].qs;
+        const uint8_t  * qs = x[i].qs;
+        const uint8_t  * qh = x[i].qh;
+        const uint16_t * sc = (const uint16_t *)x[i].scales;
+
+        scale.u16 = (sc[0] >> 12) | ((sc[1] >> 8) & 0x00f0) | ((sc[2] >> 4) & 0x0f00) | (sc[3] & 0xf000);
+
+        int32x4_t sumi1 = mzero;
+        int32x4_t sumi2 = mzero;
+
+        for (int ib = 0; ib < QK_K/32; ib += 2) {
+
+            q1b.val[0] = vcombine_s8(vld1_s8((const int8_t *)(iq1s_grid + (qs[0] | ((qh[0] << 8) & 0x700)))),
+                                     vld1_s8((const int8_t *)(iq1s_grid + (qs[1] | ((qh[0] << 4) & 0x700)))));
+            q1b.val[1] = vcombine_s8(vld1_s8((const int8_t *)(iq1s_grid + (qs[2] | ((qh[1] << 8) & 0x700)))),
+                                     vld1_s8((const int8_t *)(iq1s_grid + (qs[3] | ((qh[1] << 4) & 0x700)))));
+            q1b.val[2] = vcombine_s8(vld1_s8((const int8_t *)(iq1s_grid + (qs[4] | ((qh[2] << 8) & 0x700)))),
+                                     vld1_s8((const int8_t *)(iq1s_grid + (qs[5] | ((qh[2] << 4) & 0x700)))));
+            q1b.val[3] = vcombine_s8(vld1_s8((const int8_t *)(iq1s_grid + (qs[6] | ((qh[3] << 8) & 0x700)))),
+                                     vld1_s8((const int8_t *)(iq1s_grid + (qs[7] | ((qh[3] << 4) & 0x700)))));
+
+            q8b = ggml_vld1q_s8_x4(q8); q8 += 64;
+
+            const int32x4_t p1 = vpaddq_s32(ggml_vdotq_s32(mzero, q1b.val[0], q8b.val[0]), ggml_vdotq_s32(mzero, q1b.val[1], q8b.val[1]));
+            const int32x4_t p2 = vpaddq_s32(ggml_vdotq_s32(mzero, q1b.val[2], q8b.val[2]), ggml_vdotq_s32(mzero, q1b.val[3], q8b.val[3]));
+            const int32x4_t p12 = vpaddq_s32(p1, p2);
+
+            const uint32_t * qh32 = (const uint32_t *)qh; // we are 4-byte aligned, so we can do that
+            aux32 = ((qh32[0] >> 3) & 0x01010101) | ((qh32[0] >> 6) & 0x02020202);
+
+            const int32x4_t p3 = vpaddq_s32(ggml_vdotq_s32(mzero, deltas.val[aux8[0]], q8b.val[0]), ggml_vdotq_s32(mzero, deltas.val[aux8[1]], q8b.val[1]));
+            const int32x4_t p4 = vpaddq_s32(ggml_vdotq_s32(mzero, deltas.val[aux8[2]], q8b.val[2]), ggml_vdotq_s32(mzero, deltas.val[aux8[3]], q8b.val[3]));
+            const int32x4_t p34 = vpaddq_s32(p3, p4);
+
+            int32x4_t scales_4 = ggml_vld1q_u32(sc[ib/2] >> 0, sc[ib/2] >> 3, sc[ib/2] >> 6, sc[ib/2] >> 9);
+
+            scales_4 = vaddq_s32(vshlq_n_s32(vandq_s32(scales_4, mask), 1), mone);
+
+            sumi1 = vmlaq_s32(sumi1, scales_4, p12);
+            sumi2 = vmlaq_s32(sumi2, scales_4, p34);
+
+            qs += 8; qh += 4;
+
+        }
+
+        sumf += y[i].d * GGML_FP16_TO_FP32(scale.f16) * (vaddvq_s32(sumi1) + IQ1M_DELTA * vaddvq_s32(sumi2));
+    }
+
+    *s = sumf;
+
+#elif defined __AVX2__
+
+    const __m256i mask = _mm256_set1_epi16(0x7);
+    const __m256i mone = _mm256_set1_epi16(1);
+
+    __m256 accum1 = _mm256_setzero_ps();
+    __m256 accum2 = _mm256_setzero_ps();
+    for (int i = 0; i < nb; ++i) {
+
+        const int8_t   * q8 = y[i].qs;
+        const uint8_t  * qs = x[i].qs;
+        const uint8_t  * qh = x[i].qh;
+        const uint16_t * sc = (const uint16_t *)x[i].scales;
+
+        scale.u16 = (sc[0] >> 12) | ((sc[1] >> 8) & 0x00f0) | ((sc[2] >> 4) & 0x0f00) | (sc[3] & 0xf000);
+
+        __m256i sumi1 = _mm256_setzero_si256();
+        __m256i sumi2 = _mm256_setzero_si256();
+        for (int ib = 0; ib < QK_K/32; ib += 2) {
+            const __m256i q1b_1 = _mm256_set_epi64x(
+                    iq1s_grid[qs[3] | (((uint16_t)qh[1] << 4) & 0x700)], iq1s_grid[qs[2] | (((uint16_t)qh[1] << 8) & 0x700)],
+                    iq1s_grid[qs[1] | (((uint16_t)qh[0] << 4) & 0x700)], iq1s_grid[qs[0] | (((uint16_t)qh[0] << 8) & 0x700)]
+            );
+            const __m256i q1b_2 = _mm256_set_epi64x(
+                    iq1s_grid[qs[7] | (((uint16_t)qh[3] << 4) & 0x700)], iq1s_grid[qs[6] | (((uint16_t)qh[3] << 8) & 0x700)],
+                    iq1s_grid[qs[5] | (((uint16_t)qh[2] << 4) & 0x700)], iq1s_grid[qs[4] | (((uint16_t)qh[2] << 8) & 0x700)]
+            );
+            const __m256i q8b_1 = _mm256_loadu_si256((const __m256i*)q8); q8 += 32;
+            const __m256i q8b_2 = _mm256_loadu_si256((const __m256i*)q8); q8 += 32;
+
+            const __m256i dot1 = mul_add_epi8(q1b_1, q8b_1);
+            const __m256i dot2 = mul_add_epi8(q1b_2, q8b_2);
+
+            const __m256i delta1 = _mm256_set_epi64x(qh[1] & 0x80 ? 0xffffffffffffffff : 0x0101010101010101,
+                                                     qh[1] & 0x08 ? 0xffffffffffffffff : 0x0101010101010101,
+                                                     qh[0] & 0x80 ? 0xffffffffffffffff : 0x0101010101010101,
+                                                     qh[0] & 0x08 ? 0xffffffffffffffff : 0x0101010101010101);
+            const __m256i delta2 = _mm256_set_epi64x(qh[3] & 0x80 ? 0xffffffffffffffff : 0x0101010101010101,
+                                                     qh[3] & 0x08 ? 0xffffffffffffffff : 0x0101010101010101,
+                                                     qh[2] & 0x80 ? 0xffffffffffffffff : 0x0101010101010101,
+                                                     qh[2] & 0x08 ? 0xffffffffffffffff : 0x0101010101010101);
+
+            const __m256i dot3 = mul_add_epi8(delta1, q8b_1);
+            const __m256i dot4 = mul_add_epi8(delta2, q8b_2);
+
+            __m256i scale1 = MM256_SET_M128I(_mm_set1_epi16(sc[ib/2] >> 3), _mm_set1_epi16(sc[ib/2] >> 0));
+            __m256i scale2 = MM256_SET_M128I(_mm_set1_epi16(sc[ib/2] >> 9), _mm_set1_epi16(sc[ib/2] >> 6));
+
+            scale1 = _mm256_add_epi16(_mm256_slli_epi16(_mm256_and_si256(scale1, mask), 1), mone);
+            scale2 = _mm256_add_epi16(_mm256_slli_epi16(_mm256_and_si256(scale2, mask), 1), mone);
+            const __m256i p1 = _mm256_madd_epi16(dot1, scale1);
+            const __m256i p2 = _mm256_madd_epi16(dot2, scale2);
+            const __m256i p3 = _mm256_madd_epi16(dot3, scale1);
+            const __m256i p4 = _mm256_madd_epi16(dot4, scale2);
+
+            sumi1 = _mm256_add_epi32(sumi1, _mm256_add_epi32(p1, p2));
+            sumi2 = _mm256_add_epi32(sumi2, _mm256_add_epi32(p3, p4));
+
+            qs += 8; qh += 4;
+        }
+
+        const __m256 d = _mm256_set1_ps(y[i].d * GGML_FP16_TO_FP32(scale.f16));
+
+        accum1 = _mm256_fmadd_ps(d, _mm256_cvtepi32_ps(sumi1), accum1);
+        accum2 = _mm256_fmadd_ps(d, _mm256_cvtepi32_ps(sumi2), accum2);
+    }
+
+    *s = hsum_float_8(accum1) + IQ1M_DELTA * hsum_float_8(accum2);
+
+#elif defined __AVX__
+    const __m128i mask = _mm_set1_epi16(0x7);
+    const __m128i mone = _mm_set1_epi16(1);
+
+    __m256 accum1 = _mm256_setzero_ps();
+    __m256 accum2 = _mm256_setzero_ps();
+    for (int i = 0; i < nb; ++i) {
+
+        const int8_t   * q8 = y[i].qs;
+        const uint8_t  * qs = x[i].qs;
+        const uint8_t  * qh = x[i].qh;
+        const uint16_t * sc = (const uint16_t *)x[i].scales;
+
+        scale.u16 = (sc[0] >> 12) | ((sc[1] >> 8) & 0x00f0) | ((sc[2] >> 4) & 0x0f00) | (sc[3] & 0xf000);
+
+        __m128i sumi1_0 = _mm_setzero_si128();
+        __m128i sumi1_1 = _mm_setzero_si128();
+        __m128i sumi2_0 = _mm_setzero_si128();
+        __m128i sumi2_1 = _mm_setzero_si128();
+        for (int ib = 0; ib < QK_K/32; ib += 2) {
+            const __m128i q1b_1_0 = _mm_set_epi64x(
+                    iq1s_grid[qs[1] | (((uint16_t)qh[0] << 4) & 0x700)], iq1s_grid[qs[0] | (((uint16_t)qh[0] << 8) & 0x700)]);
+            const __m128i q1b_1_1 = _mm_set_epi64x(
+                    iq1s_grid[qs[3] | (((uint16_t)qh[1] << 4) & 0x700)], iq1s_grid[qs[2] | (((uint16_t)qh[1] << 8) & 0x700)]);
+            const __m128i q1b_2_0 = _mm_set_epi64x(
+                    iq1s_grid[qs[5] | (((uint16_t)qh[2] << 4) & 0x700)], iq1s_grid[qs[4] | (((uint16_t)qh[2] << 8) & 0x700)]);
+            const __m128i q1b_2_1 = _mm_set_epi64x(
+                    iq1s_grid[qs[7] | (((uint16_t)qh[3] << 4) & 0x700)], iq1s_grid[qs[6] | (((uint16_t)qh[3] << 8) & 0x700)]);
+            const __m128i q8b_1_0 = _mm_loadu_si128((const __m128i *)q8); q8 += 16;
+            const __m128i q8b_1_1 = _mm_loadu_si128((const __m128i *)q8); q8 += 16;
+            const __m128i q8b_2_0 = _mm_loadu_si128((const __m128i *)q8); q8 += 16;
+            const __m128i q8b_2_1 = _mm_loadu_si128((const __m128i *)q8); q8 += 16;
+
+            const __m128i dot1_0 = mul_add_epi8_sse(q1b_1_0, q8b_1_0);
+            const __m128i dot1_1 = mul_add_epi8_sse(q1b_1_1, q8b_1_1);
+            const __m128i dot2_0 = mul_add_epi8_sse(q1b_2_0, q8b_2_0);
+            const __m128i dot2_1 = mul_add_epi8_sse(q1b_2_1, q8b_2_1);
+
+            const __m128i delta1_0 = _mm_set_epi64x(qh[0] & 0x80 ? 0xffffffffffffffff : 0x0101010101010101,
+                                                     qh[0] & 0x08 ? 0xffffffffffffffff : 0x0101010101010101);
+            const __m128i delta1_1 = _mm_set_epi64x(qh[1] & 0x80 ? 0xffffffffffffffff : 0x0101010101010101,
+                                                     qh[1] & 0x08 ? 0xffffffffffffffff : 0x0101010101010101);
+            const __m128i delta2_0 = _mm_set_epi64x(qh[2] & 0x80 ? 0xffffffffffffffff : 0x0101010101010101,
+                                                     qh[2] & 0x08 ? 0xffffffffffffffff : 0x0101010101010101);
+            const __m128i delta2_1 = _mm_set_epi64x(qh[3] & 0x80 ? 0xffffffffffffffff : 0x0101010101010101,
+                                                     qh[3] & 0x08 ? 0xffffffffffffffff : 0x0101010101010101);
+
+            const __m128i dot3_0 = mul_add_epi8_sse(delta1_0, q8b_1_0);
+            const __m128i dot3_1 = mul_add_epi8_sse(delta1_1, q8b_1_1);
+            const __m128i dot4_0 = mul_add_epi8_sse(delta2_0, q8b_2_0);
+            const __m128i dot4_1 = mul_add_epi8_sse(delta2_1, q8b_2_1);
+
+            __m128i scale1_0 = _mm_set1_epi16(sc[ib/2] >> 0);
+            __m128i scale1_1 = _mm_set1_epi16(sc[ib/2] >> 3);
+            __m128i scale2_0 = _mm_set1_epi16(sc[ib/2] >> 6);
+            __m128i scale2_1 = _mm_set1_epi16(sc[ib/2] >> 9);
+
+            scale1_0 = _mm_add_epi16(_mm_slli_epi16(_mm_and_si128(scale1_0, mask), 1), mone);
+            scale1_1 = _mm_add_epi16(_mm_slli_epi16(_mm_and_si128(scale1_1, mask), 1), mone);
+            scale2_0 = _mm_add_epi16(_mm_slli_epi16(_mm_and_si128(scale2_0, mask), 1), mone);
+            scale2_1 = _mm_add_epi16(_mm_slli_epi16(_mm_and_si128(scale2_1, mask), 1), mone);
+            const __m128i p1_0 = _mm_madd_epi16(dot1_0, scale1_0);
+            const __m128i p1_1 = _mm_madd_epi16(dot1_1, scale1_1);
+            const __m128i p2_0 = _mm_madd_epi16(dot2_0, scale2_0);
+            const __m128i p2_1 = _mm_madd_epi16(dot2_1, scale2_1);
+            const __m128i p3_0 = _mm_madd_epi16(dot3_0, scale1_0);
+            const __m128i p3_1 = _mm_madd_epi16(dot3_1, scale1_1);
+            const __m128i p4_0 = _mm_madd_epi16(dot4_0, scale2_0);
+            const __m128i p4_1 = _mm_madd_epi16(dot4_1, scale2_1);
+
+            sumi1_0 = _mm_add_epi32(sumi1_0, _mm_add_epi32(p1_0, p2_0));
+            sumi1_1 = _mm_add_epi32(sumi1_1, _mm_add_epi32(p1_1, p2_1));
+            sumi2_0 = _mm_add_epi32(sumi2_0, _mm_add_epi32(p3_0, p4_0));
+            sumi2_1 = _mm_add_epi32(sumi2_1, _mm_add_epi32(p3_1, p4_1));
+
+            qs += 8; qh += 4;
+        }
+
+        const __m256 d = _mm256_set1_ps(y[i].d * GGML_FP16_TO_FP32(scale.f16));
+
+        accum1 = _mm256_add_ps(_mm256_mul_ps(d, _mm256_cvtepi32_ps(MM256_SET_M128I(sumi1_1, sumi1_0))), accum1);
+        accum2 = _mm256_add_ps(_mm256_mul_ps(d, _mm256_cvtepi32_ps(MM256_SET_M128I(sumi2_1, sumi2_0))), accum2);
+    }
+
+    *s = hsum_float_8(accum1) + IQ1M_DELTA * hsum_float_8(accum2);
+
+#else
+
+    int sum1[2], sum2[2], delta[4];
+
+    float sumf = 0;
+    for (int i = 0; i < nb; i++) {
+
+        const int8_t   * q8 = y[i].qs;
+        const uint8_t  * qs = x[i].qs;
+        const uint8_t  * qh = x[i].qh;
+        const uint16_t * sc = (const uint16_t *)x[i].scales;
+
+        scale.u16 = (sc[0] >> 12) | ((sc[1] >> 8) & 0x00f0) | ((sc[2] >> 4) & 0x0f00) | (sc[3] & 0xf000);
+
+        int sumi1 = 0, sumi2 = 0;
+        for (int ib = 0; ib < QK_K/32; ++ib) {
+            delta[0] = qh[0] & 0x08 ? -1 : 1;
+            delta[1] = qh[0] & 0x80 ? -1 : 1;
+            delta[2] = qh[1] & 0x08 ? -1 : 1;
+            delta[3] = qh[1] & 0x80 ? -1 : 1;
+            sum1[0] = sum1[1] = sum2[0] = sum2[1] = 0;
+            for (int l = 0; l < 4; ++l) {
+                const int8_t * grid = (const int8_t *)(iq1s_grid + (qs[l] | (((uint16_t)qh[l/2] << (8 - 4*(l%2))) & 0x700)));
+                int lsum1 = 0, lsum2 = 0;
+                for (int j = 0; j < 8; ++j) {
+                    lsum1 += q8[j] * grid[j];
+                    lsum2 += q8[j];
+                }
+                q8 += 8;
+                sum1[l/2] += lsum1;
+                sum2[l/2] += lsum2*delta[l];
+            }
+
+            const int ls1 = 2*((sc[ib/2] >> (6*(ib%2)+0)) & 0x7) + 1;
+            const int ls2 = 2*((sc[ib/2] >> (6*(ib%2)+3)) & 0x7) + 1;
+
+            sumi1 += sum1[0] * ls1 + sum1[1] * ls2;
+            sumi2 += sum2[0] * ls1 + sum2[1] * ls2;
+            qs += 4;
+            qh += 2;
+        }
+
+        sumf += GGML_FP16_TO_FP32(scale.f16) * y[i].d * (sumi1 + IQ1M_DELTA * sumi2);
+    }
+
+    *s = sumf;
+
+#endif
+}
+
+void ggml_vec_dot_iq4_nl_q8_0(int n, float * restrict s, size_t bs, const void * restrict vx, size_t bx, const void * restrict vy, size_t by, int nrc) {
+    assert(nrc == 1);
+    UNUSED(nrc);
+    UNUSED(bx);
+    UNUSED(by);
+    UNUSED(bs);
+    assert(n % QK4_NL == 0);
+    static_assert(QK4_NL == QK8_0, "QK4_NL and QK8_0 must be the same");
+
+    const block_iq4_nl * restrict x = vx;
+    const block_q8_0   * restrict y = vy;
+
+    const int nb = n / QK4_NL;
+
+    int ib = 0;
+    float sumf = 0;
+
+#if defined __ARM_NEON
+    const int8x16_t values = vld1q_s8(kvalues_iq4nl);
+    const uint8x16_t m4b = vdupq_n_u8(0x0f);
+    uint8x16x2_t q4bits;
+    int8x16x4_t q4b;
+    int8x16x4_t q8b;
+    int32x4_t prod_1, prod_2;
+
+    for (; ib + 1 < nb; ib += 2) {
+
+        q4bits.val[0] = vld1q_u8(x[ib + 0].qs);
+        q4bits.val[1] = vld1q_u8(x[ib + 1].qs);
+        q8b.val[0]    = vld1q_s8(y[ib + 0].qs);
+        q8b.val[1]    = vld1q_s8(y[ib + 0].qs + 16);
+        q8b.val[2]    = vld1q_s8(y[ib + 1].qs);
+        q8b.val[3]    = vld1q_s8(y[ib + 1].qs + 16);
+
+        q4b.val[0] = ggml_vqtbl1q_s8(values, vandq_u8  (q4bits.val[0], m4b));
+        q4b.val[1] = ggml_vqtbl1q_s8(values, vshrq_n_u8(q4bits.val[0], 4));
+        q4b.val[2] = ggml_vqtbl1q_s8(values, vandq_u8  (q4bits.val[1], m4b));
+        q4b.val[3] = ggml_vqtbl1q_s8(values, vshrq_n_u8(q4bits.val[1], 4));
+
+        prod_1 = ggml_vdotq_s32(ggml_vdotq_s32(vdupq_n_s32(0), q4b.val[0], q8b.val[0]), q4b.val[1], q8b.val[1]);
+        prod_2 = ggml_vdotq_s32(ggml_vdotq_s32(vdupq_n_s32(0), q4b.val[2], q8b.val[2]), q4b.val[3], q8b.val[3]);
+
+        sumf +=
+            GGML_FP16_TO_FP32(x[ib+0].d) * GGML_FP16_TO_FP32(y[ib + 0].d) * vaddvq_s32(prod_1) +
+            GGML_FP16_TO_FP32(x[ib+1].d) * GGML_FP16_TO_FP32(y[ib + 1].d) * vaddvq_s32(prod_2);
+    }
+
+#elif defined __AVX2__
+
+    const __m128i values128 = _mm_loadu_si128((const __m128i*)kvalues_iq4nl);
+    const __m128i m4b  = _mm_set1_epi8(0x0f);
+    const __m256i mone = _mm256_set1_epi16(1);
+
+    __m256 accum1 = _mm256_setzero_ps();
+    __m256 accum2 = _mm256_setzero_ps();
+    for (; ib + 1 < nb; ib += 2) {
+        const __m128i q4bits_1 = _mm_loadu_si128((const __m128i*)x[ib + 0].qs);
+        const __m128i q4bits_2 = _mm_loadu_si128((const __m128i*)x[ib + 1].qs);
+        const __m256i q8b_1 = _mm256_loadu_si256((const __m256i *)y[ib + 0].qs);
+        const __m256i q8b_2 = _mm256_loadu_si256((const __m256i *)y[ib + 1].qs);
+        const __m256i q4b_1 = MM256_SET_M128I(_mm_shuffle_epi8(values128, _mm_and_si128(_mm_srli_epi16(q4bits_1, 4), m4b)),
+                                              _mm_shuffle_epi8(values128, _mm_and_si128(q4bits_1, m4b)));
+        const __m256i q4b_2 = MM256_SET_M128I(_mm_shuffle_epi8(values128, _mm_and_si128(_mm_srli_epi16(q4bits_2, 4), m4b)),
+                                              _mm_shuffle_epi8(values128, _mm_and_si128(q4bits_2, m4b)));
+        const __m256i p16_1 = mul_add_epi8(q4b_1, q8b_1);
+        const __m256i p16_2 = mul_add_epi8(q4b_2, q8b_2);
+        const __m256i p_1 = _mm256_madd_epi16(p16_1, mone);
+        const __m256i p_2 = _mm256_madd_epi16(p16_2, mone);
+        accum1 = _mm256_fmadd_ps(_mm256_set1_ps(GGML_FP16_TO_FP32(y[ib + 0].d)*GGML_FP16_TO_FP32(x[ib + 0].d)),
+                _mm256_cvtepi32_ps(p_1), accum1);
+        accum2 = _mm256_fmadd_ps(_mm256_set1_ps(GGML_FP16_TO_FP32(y[ib + 1].d)*GGML_FP16_TO_FP32(x[ib + 1].d)),
+                _mm256_cvtepi32_ps(p_2), accum2);
+    }
+
+    sumf = hsum_float_8(_mm256_add_ps(accum1, accum2));
+
+#elif defined __AVX__
+    const __m128i values128 = _mm_loadu_si128((const __m128i*)kvalues_iq4nl);
+    const __m128i m4b  = _mm_set1_epi8(0x0f);
+    const __m128i mone = _mm_set1_epi16(1);
+
+    __m256 accum1 = _mm256_setzero_ps();
+    __m256 accum2 = _mm256_setzero_ps();
+    for (; ib + 1 < nb; ib += 2) {
+        const __m128i q4bits_1 = _mm_loadu_si128((const __m128i *)x[ib + 0].qs);
+        const __m128i q4bits_2 = _mm_loadu_si128((const __m128i *)x[ib + 1].qs);
+        const __m128i q8b_1_0 = _mm_loadu_si128((const __m128i *)y[ib + 0].qs);
+        const __m128i q8b_1_1 = _mm_loadu_si128((const __m128i *)y[ib + 0].qs + 1);
+        const __m128i q8b_2_0 = _mm_loadu_si128((const __m128i *)y[ib + 1].qs);
+        const __m128i q8b_2_1 = _mm_loadu_si128((const __m128i *)y[ib + 1].qs + 1);
+
+        const __m128i q4b_1_0 = _mm_shuffle_epi8(values128, _mm_and_si128(q4bits_1, m4b));
+        const __m128i q4b_1_1 = _mm_shuffle_epi8(values128, _mm_and_si128(_mm_srli_epi16(q4bits_1, 4), m4b));
+        const __m128i q4b_2_0 = _mm_shuffle_epi8(values128, _mm_and_si128(q4bits_2, m4b));
+        const __m128i q4b_2_1 = _mm_shuffle_epi8(values128, _mm_and_si128(_mm_srli_epi16(q4bits_2, 4), m4b));
+        const __m128i p16_1_0 = mul_add_epi8_sse(q4b_1_0, q8b_1_0);
+        const __m128i p16_1_1 = mul_add_epi8_sse(q4b_1_1, q8b_1_1);
+        const __m128i p16_2_0 = mul_add_epi8_sse(q4b_2_0, q8b_2_0);
+        const __m128i p16_2_1 = mul_add_epi8_sse(q4b_2_1, q8b_2_1);
+        const __m128i p_1_0 = _mm_madd_epi16(p16_1_0, mone);
+        const __m128i p_1_1 = _mm_madd_epi16(p16_1_1, mone);
+        const __m128i p_2_0 = _mm_madd_epi16(p16_2_0, mone);
+        const __m128i p_2_1 = _mm_madd_epi16(p16_2_1, mone);
+        accum1 = _mm256_add_ps(_mm256_mul_ps(_mm256_set1_ps(GGML_FP16_TO_FP32(y[ib + 0].d)*GGML_FP16_TO_FP32(x[ib + 0].d)),
+                _mm256_cvtepi32_ps(MM256_SET_M128I(p_1_1, p_1_0))), accum1);
+        accum2 = _mm256_add_ps(_mm256_mul_ps(_mm256_set1_ps(GGML_FP16_TO_FP32(y[ib + 1].d)*GGML_FP16_TO_FP32(x[ib + 1].d)),
+                _mm256_cvtepi32_ps(MM256_SET_M128I(p_2_1, p_2_0))), accum2);
+    }
+
+    sumf = hsum_float_8(_mm256_add_ps(accum1, accum2));
+
+#elif defined(__POWER9_VECTOR__)
+    const vector signed char lowMask = vec_splats((signed char)0xF);
+    const vector signed int v0 = vec_splats((int32_t)0);
+    const vector unsigned char v4 = vec_splats((unsigned char)0x4);
+
+    vector float vsumf0 = vec_splats(0.0f);
+    vector float vsumf1 = vec_splats(0.0f);
+
+    const vector signed char values = vec_xl( 0, kvalues_iq4nl);
+
+#pragma GCC unroll 4
+    for (; ib < nb; ++ib) {
+        __builtin_prefetch(x[ib].qs, 0, 1);
+        __builtin_prefetch(y[ib].qs, 0, 1);
+
+
+        vector float vxd = vec_splats(GGML_FP16_TO_FP32(x[ib].d));
+        vector float vyd = vec_splats(GGML_FP16_TO_FP32(y[ib].d));
+        vector float vd = vec_mul(vxd, vyd);
+
+        vector signed char qxs = (vector signed char)vec_xl( 0, x[ib].qs);
+        vector signed char q4x0 = vec_and(qxs, lowMask);
+        vector signed char q4x1 = vec_sr(qxs, v4);
+
+        q4x0 = vec_perm(values, values, (vector unsigned char)q4x0);
+        q4x1 = vec_perm(values, values, (vector unsigned char)q4x1);
+
+        vector signed char q8y0 = vec_xl( 0, y[ib].qs);
+        vector signed char q8y1 = vec_xl(16, y[ib].qs);
+
+        vector signed short qv0 = vec_add(vec_mule(q4x0, q8y0), vec_mulo(q4x0, q8y0));
+        vector signed short qv1 = vec_add(vec_mule(q4x1, q8y1), vec_mulo(q4x1, q8y1));
+
+        vector signed int vsumi0 = v0;
+        vector signed int vsumi1 = v0;
+
+        vsumi0 = vec_sum4s(qv0, vsumi0);
+        vsumi1 = vec_sum4s(qv1, vsumi1);
+
+        vsumf0 = vec_madd(vec_ctf(vsumi0, 0), vd, vsumf0);
+        vsumf1 = vec_madd(vec_ctf(vsumi1, 0), vd, vsumf1);
+    }
+
+    vsumf0 = vec_add(vsumf0, vsumf1);
+
+    vsumf0 = vec_add(vsumf0, vec_sld(vsumf0, vsumf0, 4));
+    vsumf0 = vec_add(vsumf0, vec_sld(vsumf0, vsumf0, 8));
+
+    sumf = vec_extract(vsumf0, 0);
+
+#elif defined (__loongarch_asx)
+
+    const __m128i values128 = __lsx_vld((const __m128i*)kvalues_iq4nl, 0);
+    const __m128i m4b  = __lsx_vreplgr2vr_b(0x0f);
+    const __m256i mone = __lasx_xvreplgr2vr_h(1);
+
+    __m256 accum1 = (__m256)__lasx_xvldi(0);
+    __m256 accum2 = (__m256)__lasx_xvldi(0);
+    for (; ib + 1 < nb; ib += 2) {
+        const __m128i q4bits_1 = __lsx_vld((const __m128i*)x[ib + 0].qs, 0);
+        const __m128i q4bits_2 = __lsx_vld((const __m128i*)x[ib + 1].qs, 0);
+        const __m256i q8b_1 = __lasx_xvld((const __m256i *)y[ib + 0].qs, 0);
+        const __m256i q8b_2 = __lasx_xvld((const __m256i *)y[ib + 1].qs, 0);
+        const __m256i q4b_1 = lasx_insertf128(lsx_shuffle_b(values128, __lsx_vand_v(__lsx_vsrli_h(q4bits_1, 4), m4b)),
+                                              lsx_shuffle_b(values128, __lsx_vand_v(q4bits_1, m4b)));
+        const __m256i q4b_2 = lasx_insertf128(lsx_shuffle_b(values128, __lsx_vand_v(__lsx_vsrli_h(q4bits_2, 4), m4b)),
+                                              lsx_shuffle_b(values128, __lsx_vand_v(q4bits_2, m4b)));
+        const __m256i p16_1 = mul_add_epi8(q4b_1, q8b_1);
+        const __m256i p16_2 = mul_add_epi8(q4b_2, q8b_2);
+        const __m256i p_1 = lasx_madd_h(p16_1, mone);
+        const __m256i p_2 = lasx_madd_h(p16_2, mone);
+        accum1 = __lasx_xvfmadd_s(__lasx_xvreplfr2vr_s(GGML_FP16_TO_FP32(y[ib + 0].d)*GGML_FP16_TO_FP32(x[ib + 0].d)),
+                __lasx_xvffint_s_w(p_1), accum1);
+        accum2 = __lasx_xvfmadd_s(__lasx_xvreplfr2vr_s(GGML_FP16_TO_FP32(y[ib + 1].d)*GGML_FP16_TO_FP32(x[ib + 1].d)),
+                __lasx_xvffint_s_w(p_2), accum2);
+    }
+
+    sumf = hsum_float_8(__lasx_xvfadd_s(accum1, accum2));
+
+#endif
+    for (; ib < nb; ++ib) {
+        const float d = GGML_FP16_TO_FP32(y[ib].d)*GGML_FP16_TO_FP32(x[ib].d);
+        int sumi1 = 0, sumi2 = 0;
+        for (int j = 0; j < QK4_NL/2; ++j) {
+            sumi1 += y[ib].qs[j+       0] * kvalues_iq4nl[x[ib].qs[j] & 0xf];
+            sumi2 += y[ib].qs[j+QK4_NL/2] * kvalues_iq4nl[x[ib].qs[j] >>  4];
+        }
+        sumf += d * (sumi1 + sumi2);
+    }
+    *s = sumf;
+}
+
+void ggml_vec_dot_iq4_xs_q8_K(int n, float * restrict s, size_t bs, const void * restrict vx, size_t bx, const void * restrict vy, size_t by, int nrc) {
+    assert(nrc == 1);
+    UNUSED(nrc);
+    UNUSED(bx);
+    UNUSED(by);
+    UNUSED(bs);
+    assert(n % QK_K == 0);
+
+    const block_iq4_xs * restrict x = vx;
+    const block_q8_K   * restrict y = vy;
+
+    const int nb = n / QK_K;
+
+#if defined __ARM_NEON
+    const int8x16_t values = vld1q_s8(kvalues_iq4nl);
+    const uint8x16_t m4b = vdupq_n_u8(0x0f);
+    ggml_uint8x16x2_t q4bits;
+    ggml_int8x16x4_t q4b;
+    ggml_int8x16x4_t q8b;
+    int32x4_t prod_1, prod_2;
+
+    float sumf = 0;
+
+    for (int ibl = 0; ibl < nb; ++ibl) {
+
+        const int8_t  * q8 = y[ibl].qs;
+        const uint8_t * q4 = x[ibl].qs;
+        uint16_t h = x[ibl].scales_h;
+
+        int sumi1 = 0, sumi2 = 0;
+        for (int ib = 0; ib < QK_K/64; ++ib) {
+
+            q4bits = ggml_vld1q_u8_x2(q4); q4 += 32;
+            q8b    = ggml_vld1q_s8_x4(q8); q8 += 64;
+
+            q4b.val[0] = ggml_vqtbl1q_s8(values, vandq_u8  (q4bits.val[0], m4b));
+            q4b.val[1] = ggml_vqtbl1q_s8(values, vshrq_n_u8(q4bits.val[0], 4));
+            q4b.val[2] = ggml_vqtbl1q_s8(values, vandq_u8  (q4bits.val[1], m4b));
+            q4b.val[3] = ggml_vqtbl1q_s8(values, vshrq_n_u8(q4bits.val[1], 4));
+
+            prod_1 = ggml_vdotq_s32(ggml_vdotq_s32(vdupq_n_s32(0), q4b.val[0], q8b.val[0]), q4b.val[1], q8b.val[1]);
+            prod_2 = ggml_vdotq_s32(ggml_vdotq_s32(vdupq_n_s32(0), q4b.val[2], q8b.val[2]), q4b.val[3], q8b.val[3]);
+
+            int ls1 = ((x[ibl].scales_l[ib] & 0xf) | ((h << 4) & 0x30)) - 32;
+            int ls2 = ((x[ibl].scales_l[ib] >>  4) | ((h << 2) & 0x30)) - 32;
+            h >>= 4;
+            sumi1 += vaddvq_s32(prod_1) * ls1;
+            sumi2 += vaddvq_s32(prod_2) * ls2;
+
+        }
+
+        sumf += GGML_FP16_TO_FP32(x[ibl].d) * y[ibl].d * (sumi1 + sumi2);
+    }
+
+    *s = sumf;
+
+#elif defined __AVX2__
+
+    const __m128i values128 = _mm_loadu_si128((const __m128i*)kvalues_iq4nl);
+    const __m128i m4b  = _mm_set1_epi8(0x0f);
+
+    __m256 accum = _mm256_setzero_ps();
+    for (int ibl = 0; ibl < nb; ++ibl) {
+        const uint8_t * qs = x[ibl].qs;
+        const int8_t  * q8 = y[ibl].qs;
+        uint16_t sh = x[ibl].scales_h;
+        __m256i sumi1 = _mm256_setzero_si256();
+        __m256i sumi2 = _mm256_setzero_si256();
+        for (int ib = 0; ib < QK_K/32; ib += 2) {
+            const __m128i q4bits_1 = _mm_loadu_si128((const __m128i*)qs);  qs += 16;
+            const __m128i q4bits_2 = _mm_loadu_si128((const __m128i*)qs);  qs += 16;
+            const __m256i q8b_1 = _mm256_loadu_si256((const __m256i *)q8); q8 += 32;
+            const __m256i q8b_2 = _mm256_loadu_si256((const __m256i *)q8); q8 += 32;
+            const __m256i q4b_1 = MM256_SET_M128I(_mm_shuffle_epi8(values128, _mm_and_si128(_mm_srli_epi16(q4bits_1, 4), m4b)),
+                                                  _mm_shuffle_epi8(values128, _mm_and_si128(q4bits_1, m4b)));
+            const __m256i q4b_2 = MM256_SET_M128I(_mm_shuffle_epi8(values128, _mm_and_si128(_mm_srli_epi16(q4bits_2, 4), m4b)),
+                                                  _mm_shuffle_epi8(values128, _mm_and_si128(q4bits_2, m4b)));
+            const __m256i p16_1 = mul_add_epi8(q4b_1, q8b_1);
+            const __m256i p16_2 = mul_add_epi8(q4b_2, q8b_2);
+            const int16_t ls1 = ((x[ibl].scales_l[ib/2] & 0xf) | ((sh << 4) & 0x30)) - 32;
+            const int16_t ls2 = ((x[ibl].scales_l[ib/2] >>  4) | ((sh << 2) & 0x30)) - 32;
+            sh >>= 4;
+            const __m256i p_1 = _mm256_madd_epi16(p16_1, _mm256_set1_epi16(ls1));
+            const __m256i p_2 = _mm256_madd_epi16(p16_2, _mm256_set1_epi16(ls2));
+            sumi1 = _mm256_add_epi32(p_1, sumi1);
+            sumi2 = _mm256_add_epi32(p_2, sumi2);
+        }
+        accum = _mm256_fmadd_ps(_mm256_set1_ps(GGML_FP16_TO_FP32(x[ibl].d)*y[ibl].d),
+                _mm256_cvtepi32_ps(_mm256_add_epi32(sumi1, sumi2)), accum);
+    }
+
+    *s = hsum_float_8(accum);
+
+#elif defined __AVX__
+    const __m128i values128 = _mm_loadu_si128((const __m128i*)kvalues_iq4nl);
+    const __m128i m4b  = _mm_set1_epi8(0x0f);
+
+    __m256 accum = _mm256_setzero_ps();
+    for (int ibl = 0; ibl < nb; ++ibl) {
+        const uint8_t * qs = x[ibl].qs;
+        const int8_t  * q8 = y[ibl].qs;
+        uint16_t sh = x[ibl].scales_h;
+        __m128i sumi1_0 = _mm_setzero_si128();
+        __m128i sumi1_1 = _mm_setzero_si128();
+        __m128i sumi2_0 = _mm_setzero_si128();
+        __m128i sumi2_1 = _mm_setzero_si128();
+        for (int ib = 0; ib < QK_K/32; ib += 2) {
+            const __m128i q4bits_1 = _mm_loadu_si128((const __m128i *)qs); qs += 16;
+            const __m128i q4bits_2 = _mm_loadu_si128((const __m128i *)qs); qs += 16;
+            const __m128i q8b_1_0 = _mm_loadu_si128((const __m128i *)q8); q8 += 16;
+            const __m128i q8b_1_1 = _mm_loadu_si128((const __m128i *)q8); q8 += 16;
+            const __m128i q8b_2_0 = _mm_loadu_si128((const __m128i *)q8); q8 += 16;
+            const __m128i q8b_2_1 = _mm_loadu_si128((const __m128i *)q8); q8 += 16;
+            const __m128i q4b_1_0 = _mm_shuffle_epi8(values128, _mm_and_si128(q4bits_1, m4b));
+            const __m128i q4b_1_1 = _mm_shuffle_epi8(values128, _mm_and_si128(_mm_srli_epi16(q4bits_1, 4), m4b));
+            const __m128i q4b_2_0 = _mm_shuffle_epi8(values128, _mm_and_si128(q4bits_2, m4b));
+            const __m128i q4b_2_1 = _mm_shuffle_epi8(values128, _mm_and_si128(_mm_srli_epi16(q4bits_2, 4), m4b));
+            const __m128i p16_1_0 = mul_add_epi8_sse(q4b_1_0, q8b_1_0);
+            const __m128i p16_1_1 = mul_add_epi8_sse(q4b_1_1, q8b_1_1);
+            const __m128i p16_2_0 = mul_add_epi8_sse(q4b_2_0, q8b_2_0);
+            const __m128i p16_2_1 = mul_add_epi8_sse(q4b_2_1, q8b_2_1);
+            const int16_t ls1 = ((x[ibl].scales_l[ib/2] & 0xf) | ((sh << 4) & 0x30)) - 32;
+            const int16_t ls2 = ((x[ibl].scales_l[ib/2] >>  4) | ((sh << 2) & 0x30)) - 32;
+            sh >>= 4;
+            const __m128i p_1_0 = _mm_madd_epi16(p16_1_0, _mm_set1_epi16(ls1));
+            const __m128i p_1_1 = _mm_madd_epi16(p16_1_1, _mm_set1_epi16(ls1));
+            const __m128i p_2_0 = _mm_madd_epi16(p16_2_0, _mm_set1_epi16(ls2));
+            const __m128i p_2_1 = _mm_madd_epi16(p16_2_1, _mm_set1_epi16(ls2));
+            sumi1_0 = _mm_add_epi32(p_1_0, sumi1_0);
+            sumi1_1 = _mm_add_epi32(p_1_1, sumi1_1);
+            sumi2_0 = _mm_add_epi32(p_2_0, sumi2_0);
+            sumi2_1 = _mm_add_epi32(p_2_1, sumi2_1);
+        }
+        __m128i sumi12_0 = _mm_add_epi32(sumi1_0, sumi2_0);
+        __m128i sumi12_1 = _mm_add_epi32(sumi1_1, sumi2_1);
+        accum = _mm256_add_ps(_mm256_mul_ps(_mm256_set1_ps(GGML_FP16_TO_FP32(x[ibl].d)*y[ibl].d),
+                _mm256_cvtepi32_ps(MM256_SET_M128I(sumi12_1, sumi12_0))), accum);
+    }
+
+    *s = hsum_float_8(accum);
+
+#elif defined(__POWER9_VECTOR__)
+    const vector signed char lowMask = vec_splats((signed char)0xF);
+    const vector int v0 = vec_splats((int32_t)0);
+    const vector unsigned char v4 = vec_splats((unsigned char)0x4);
+
+    vector float vsumf0 = vec_splats(0.0f);
+    vector float vsumf1 = vec_splats(0.0f);
+    vector float vsumf2 = vec_splats(0.0f);
+    vector float vsumf3 = vec_splats(0.0f);
+
+    const vector signed char values = vec_xl( 0, kvalues_iq4nl);
+
+    for (int ibl = 0; ibl < nb; ++ibl) {
+
+        vector float vxd = vec_splats(GGML_FP16_TO_FP32(x[ibl].d));
+        vector float vyd = vec_splats(y[ibl].d);
+        vector float vd = vec_mul(vxd, vyd);
+
+        vector signed int vsumi0 = v0;
+        vector signed int vsumi1 = v0;
+        vector signed int vsumi2 = v0;
+        vector signed int vsumi3 = v0;
+
+        uint16_t h = x[ibl].scales_h;
+
+        const uint8_t * restrict q4 = x[ibl].qs;
+        const uint8_t * restrict sc = x[ibl].scales_l;
+        const int8_t  * restrict q8 = y[ibl].qs;
+
+        for (int ib = 0; ib < QK_K/64; ib ++ ) {
+            __builtin_prefetch(q4, 0, 1);
+            __builtin_prefetch(q8, 0, 1);
+
+            vector signed char qxs0 = (vector signed char)vec_xl( 0, q4);
+            vector signed char qxs1 = (vector signed char)vec_xl(16, q4);
+            q4 += 32;
+
+            vector signed char q4x00 = (vector signed char)vec_and(qxs0, lowMask);
+            vector signed char q4x01 = (vector signed char)vec_sr(qxs0, v4);
+            vector signed char q4x10 = (vector signed char)vec_and(qxs1, lowMask);
+            vector signed char q4x11 = (vector signed char)vec_sr(qxs1, v4);
+
+            q4x00 = vec_perm(values, values, (vector unsigned char)q4x00);
+            q4x01 = vec_perm(values, values, (vector unsigned char)q4x01);
+            q4x10 = vec_perm(values, values, (vector unsigned char)q4x10);
+            q4x11 = vec_perm(values, values, (vector unsigned char)q4x11);
+
+            vector signed char q8y0 = vec_xl( 0, q8);
+            vector signed char q8y1 = vec_xl(16, q8);
+            vector signed char q8y2 = vec_xl(32, q8);
+            vector signed char q8y3 = vec_xl(48, q8);
+            q8 += 64;
+
+            vector signed short qv0 = vec_add(vec_mule(q4x00, q8y0), vec_mulo(q4x00, q8y0));
+            vector signed short qv1 = vec_add(vec_mule(q4x01, q8y1), vec_mulo(q4x01, q8y1));
+            vector signed short qv2 = vec_add(vec_mule(q4x10, q8y2), vec_mulo(q4x10, q8y2));
+            vector signed short qv3 = vec_add(vec_mule(q4x11, q8y3), vec_mulo(q4x11, q8y3));
+
+            const uint16_t ls0 = (uint16_t)(((sc[0] & 0xf) | ((h << 4) & 0x30)) - 32);
+            const uint16_t ls1 = (uint16_t)(((sc[0] >>  4) | ((h << 2) & 0x30)) - 32);
+            h >>= 4;
+            sc ++;
+
+            vector signed short vscales01 = vec_splats((int16_t)ls0);
+            vector signed short vscales23 = vec_splats((int16_t)ls1);
+
+            vsumi0 = vec_msum(qv0, vscales01, vsumi0);
+            vsumi1 = vec_msum(qv1, vscales01, vsumi1);
+            vsumi2 = vec_msum(qv2, vscales23, vsumi2);
+            vsumi3 = vec_msum(qv3, vscales23, vsumi3);
+        }
+
+        vsumf0 = vec_madd(vec_ctf(vsumi0, 0), vd, vsumf0);
+        vsumf1 = vec_madd(vec_ctf(vsumi1, 0), vd, vsumf1);
+        vsumf2 = vec_madd(vec_ctf(vsumi2, 0), vd, vsumf2);
+        vsumf3 = vec_madd(vec_ctf(vsumi3, 0), vd, vsumf3);
+    }
+
+    vsumf0 = vec_add(vsumf0, vsumf2);
+    vsumf1 = vec_add(vsumf1, vsumf3);
+
+    vsumf0 = vec_add(vsumf0, vsumf1);
+
+    vsumf0 = vec_add(vsumf0, vec_sld(vsumf0, vsumf0, 4));
+    vsumf0 = vec_add(vsumf0, vec_sld(vsumf0, vsumf0, 8));
+
+    *s = vec_extract(vsumf0, 0);
+
+#elif defined(__loongarch_asx)
+
+    const __m128i values128 = __lsx_vld((const __m128i*)kvalues_iq4nl, 0);
+    const __m128i m4b  = __lsx_vreplgr2vr_b(0x0f);
+
+    __m256 accum = (__m256)__lasx_xvldi(0);
+    __m256i tmp1;
+    __m128i tmp0, tmp2, tmp3, tmp4, mask_8f, mask;
+
+    mask_8f = __lsx_vreplgr2vr_b(0x8f);
+    for (int ibl = 0; ibl < nb; ++ibl) {
+        const uint8_t * qs = x[ibl].qs;
+        const int8_t  * q8 = y[ibl].qs;
+        uint16_t sh = x[ibl].scales_h;
+        __m256i sumi1 = __lasx_xvldi(0);
+        __m256i sumi2 = __lasx_xvldi(0);
+        __m128i zero = __lsx_vldi(0);
+        for (int ib = 0; ib < QK_K/32; ib += 2) {
+            const __m128i q4bits_1 = __lsx_vld((const __m128i*)qs, 0);  qs += 16;
+            const __m128i q4bits_2 = __lsx_vld((const __m128i*)qs, 0);  qs += 16;
+            const __m256i q8b_1 = __lasx_xvld((const __m256i *)q8, 0); q8 += 32;
+            const __m256i q8b_2 = __lasx_xvld((const __m256i *)q8, 0); q8 += 32;
+            tmp2 = __lsx_vand_v(__lsx_vand_v(__lsx_vsrli_h(q4bits_1, 4), m4b), mask_8f);
+            tmp0 = __lsx_vori_b(tmp2, 0x10);
+            mask = __lsx_vsle_b(zero, tmp2);
+            tmp3 = __lsx_vand_v(tmp0, mask);
+            tmp3 = __lsx_vshuf_b(values128, zero, tmp3);
+
+            tmp2 = __lsx_vand_v(__lsx_vand_v(q4bits_1, m4b), mask_8f);
+            tmp0 = __lsx_vori_b(tmp2, 0x10);
+            mask = __lsx_vsle_b(zero, tmp2);
+            tmp4 = __lsx_vand_v(tmp0, mask);
+            tmp4 = __lsx_vshuf_b(values128, zero, tmp4);
+
+            const __m256i q4b_1 = lasx_insertf128(tmp3, tmp4);
+
+            tmp2 = __lsx_vand_v(__lsx_vand_v(__lsx_vsrli_h(q4bits_2, 4), m4b), mask_8f);
+            tmp0 = __lsx_vori_b(tmp2, 0x10);
+            mask = __lsx_vsle_b(zero, tmp2);
+            tmp3 = __lsx_vand_v(tmp0, mask);
+            tmp3 = __lsx_vshuf_b(values128, zero, tmp3);
+
+            tmp2 = __lsx_vand_v(__lsx_vand_v(q4bits_2, m4b), mask_8f);
+            tmp0 = __lsx_vori_b(tmp2, 0x10);
+            mask = __lsx_vsle_b(zero, tmp2);
+            tmp4 = __lsx_vand_v(tmp0, mask);
+            tmp4 = __lsx_vshuf_b(values128, zero, tmp4);
+
+            const __m256i q4b_2 = lasx_insertf128(tmp3, tmp4);
+
+            const __m256i p16_1 = mul_add_epi8(q4b_1, q8b_1);
+            const __m256i p16_2 = mul_add_epi8(q4b_2, q8b_2);
+            const int16_t ls1 = ((x[ibl].scales_l[ib/2] & 0xf) | ((sh << 4) & 0x30)) - 32;
+            const int16_t ls2 = ((x[ibl].scales_l[ib/2] >>  4) | ((sh << 2) & 0x30)) - 32;
+            sh >>= 4;
+            __m256i tmp5, tmp6;
+            tmp1 = __lasx_xvreplgr2vr_h(ls1);
+            tmp5 = __lasx_xvmulwev_w_h(p16_1, tmp1);
+            tmp6 = __lasx_xvmulwod_w_h(p16_1, tmp1);
+            const __m256i p_1 = __lasx_xvadd_w(tmp5, tmp6);
+            tmp1 = __lasx_xvreplgr2vr_h(ls2);
+            tmp5 = __lasx_xvmulwev_w_h(p16_2, tmp1);
+            tmp6 = __lasx_xvmulwod_w_h(p16_2, tmp1);
+            const __m256i p_2 = __lasx_xvadd_w(tmp5, tmp6);
+            sumi1 = __lasx_xvadd_w(p_1, sumi1);
+            sumi2 = __lasx_xvadd_w(p_2, sumi2);
+        }
+        accum = __lasx_xvfmadd_s(__lasx_xvreplfr2vr_s(GGML_FP16_TO_FP32(x[ibl].d)*y[ibl].d),
+                __lasx_xvffint_s_w(__lasx_xvadd_w(sumi1, sumi2)), accum);
+    }
+
+    *s = hsum_float_8(accum);
+
+#else
+    float sumf = 0;
+    for (int ibl = 0; ibl < nb; ++ibl) {
+        const float d4d8 = GGML_FP16_TO_FP32(x[ibl].d) * y[ibl].d;
+        uint16_t h = x[ibl].scales_h;
+        const uint8_t * qs = x[ibl].qs;
+        const int8_t  * q8 = y[ibl].qs;
+        for (int ib = 0; ib < QK_K/32; ib += 2) {
+            const uint8_t ls1 = (x[ibl].scales_l[ib/2] & 0xf) | ((h << 4) & 0x30);
+            const uint8_t ls2 = (x[ibl].scales_l[ib/2] >>  4) | ((h << 2) & 0x30);
+            h >>= 4;
+            const float d1 = d4d8*(ls1 - 32);
+            const float d2 = d4d8*(ls2 - 32);
+            int sumi1 = 0, sumi2 = 0;
+            for (int j = 0; j < 16; ++j) {
+                sumi1 += q8[j+ 0] * kvalues_iq4nl[qs[j] & 0xf];
+                sumi2 += q8[j+16] * kvalues_iq4nl[qs[j] >>  4];
+            }
+            sumf += d1 * (sumi1 + sumi2);
+            qs += 16;
+            q8 += 32;
+            sumi1 = sumi2 = 0;
+            for (int j = 0; j < 16; ++j) {
+                sumi1 += q8[j+ 0] * kvalues_iq4nl[qs[j] & 0xf];
+                sumi2 += q8[j+16] * kvalues_iq4nl[qs[j] >>  4];
+            }
+            sumf += d2 * (sumi1 + sumi2);
+            qs += 16;
+            q8 += 32;
+        }
+    }
+    *s = sumf;
+#endif
+}
+
+// ================================ IQ2 quantization =============================================
+
+typedef struct {
+    uint64_t * grid;
+    int      * map;
+    uint16_t * neighbours;
+} iq2_entry_t;
+
+static iq2_entry_t iq2_data[4] = {
+    {NULL, NULL, NULL},
+    {NULL, NULL, NULL},
+    {NULL, NULL, NULL},
+    {NULL, NULL, NULL},
+};
+
+static inline int iq2_data_index(enum ggml_type type) {
+    GGML_ASSERT(type == GGML_TYPE_IQ2_XXS || type == GGML_TYPE_IQ2_XS || type == GGML_TYPE_IQ1_S || type == GGML_TYPE_IQ1_M || type == GGML_TYPE_IQ2_S);
+    return type == GGML_TYPE_IQ2_XXS ? 0 :
+           type == GGML_TYPE_IQ2_XS  ? 1 :
+           type == GGML_TYPE_IQ1_S || type == GGML_TYPE_IQ1_M ? 2 : 3;
+}
+
+static inline int iq2_grid_size(enum ggml_type type) {
+    GGML_ASSERT(type == GGML_TYPE_IQ2_XXS || type == GGML_TYPE_IQ2_XS || type == GGML_TYPE_IQ1_S || type == GGML_TYPE_IQ1_M || type == GGML_TYPE_IQ2_S);
+    return type == GGML_TYPE_IQ2_XXS ? 256 :
+           type == GGML_TYPE_IQ2_XS  ? 512 :
+           type == GGML_TYPE_IQ1_S || type == GGML_TYPE_IQ1_M ? NGRID_IQ1S : 1024;
+}
+
+static int iq2_compare_func(const void * left, const void * right) {
+    const int * l = (const int *)left;
+    const int * r = (const int *)right;
+    return l[0] < r[0] ? -1 : l[0] > r[0] ? 1 : l[1] < r[1] ? -1 : l[1] > r[1] ? 1 : 0;
+}
+
+void iq2xs_init_impl(enum ggml_type type) {
+    const int gindex = iq2_data_index(type);
+    const int grid_size = iq2_grid_size(type);
+    if (iq2_data[gindex].grid) {
+        return;
+    }
+    static const uint16_t kgrid_2bit_256[256] = {
+            0,     2,     5,     8,    10,    17,    20,    32,    34,    40,    42,    65,    68,    80,    88,    97,
+          100,   128,   130,   138,   162,   257,   260,   272,   277,   320,   388,   408,   512,   514,   546,   642,
+         1025,  1028,  1040,  1057,  1060,  1088,  1090,  1096,  1120,  1153,  1156,  1168,  1188,  1280,  1282,  1288,
+         1312,  1350,  1385,  1408,  1425,  1545,  1552,  1600,  1668,  1700,  2048,  2053,  2056,  2068,  2088,  2113,
+         2116,  2128,  2130,  2184,  2308,  2368,  2562,  2580,  4097,  4100,  4112,  4129,  4160,  4192,  4228,  4240,
+         4245,  4352,  4360,  4384,  4432,  4442,  4480,  4644,  4677,  5120,  5128,  5152,  5157,  5193,  5248,  5400,
+         5474,  5632,  5654,  6145,  6148,  6160,  6208,  6273,  6400,  6405,  6560,  6737,  8192,  8194,  8202,  8260,
+         8289,  8320,  8322,  8489,  8520,  8704,  8706,  9217,  9220,  9232,  9280,  9302,  9472,  9537,  9572,  9872,
+        10248, 10272, 10388, 10820, 16385, 16388, 16400, 16408, 16417, 16420, 16448, 16456, 16470, 16480, 16513, 16516,
+        16528, 16640, 16672, 16737, 16768, 16773, 16897, 16912, 16968, 16982, 17000, 17408, 17416, 17440, 17536, 17561,
+        17682, 17700, 17920, 18433, 18436, 18448, 18496, 18501, 18688, 18776, 18785, 18818, 19013, 19088, 20480, 20488,
+        20497, 20505, 20512, 20608, 20616, 20740, 20802, 20900, 21137, 21648, 21650, 21770, 22017, 22100, 22528, 22545,
+        22553, 22628, 22848, 23048, 24580, 24592, 24640, 24680, 24832, 24917, 25112, 25184, 25600, 25605, 25872, 25874,
+        25988, 26690, 32768, 32770, 32778, 32833, 32898, 33028, 33048, 33088, 33297, 33793, 33796, 33808, 33813, 33856,
+        33888, 34048, 34118, 34196, 34313, 34368, 34400, 34818, 35076, 35345, 36868, 36880, 36900, 36928, 37025, 37142,
+        37248, 37445, 37888, 37922, 37956, 38225, 39041, 39200, 40962, 41040, 41093, 41225, 41472, 42008, 43088, 43268,
+    };
+    static const uint16_t kgrid_2bit_512[512] = {
+            0,     2,     5,     8,    10,    17,    20,    22,    25,    32,    34,    37,    40,    65,    68,    70,
+           73,    80,    82,    85,    88,    97,   100,   128,   130,   133,   136,   145,   148,   153,   160,   257,
+          260,   262,   265,   272,   274,   277,   280,   282,   289,   292,   320,   322,   325,   328,   337,   340,
+          352,   360,   385,   388,   400,   512,   514,   517,   520,   529,   532,   544,   577,   580,   592,   597,
+          640,   650,  1025,  1028,  1030,  1033,  1040,  1042,  1045,  1048,  1057,  1060,  1088,  1090,  1093,  1096,
+         1105,  1108,  1110,  1120,  1153,  1156,  1168,  1280,  1282,  1285,  1288,  1297,  1300,  1312,  1345,  1348,
+         1360,  1377,  1408,  1537,  1540,  1552,  1574,  1600,  1602,  1668,  2048,  2050,  2053,  2056,  2058,  2065,
+         2068,  2080,  2085,  2113,  2116,  2128,  2136,  2176,  2208,  2218,  2305,  2308,  2320,  2368,  2433,  2441,
+         2560,  2592,  2600,  2710,  2720,  4097,  4100,  4102,  4105,  4112,  4114,  4117,  4120,  4129,  4132,  4160,
+         4162,  4165,  4168,  4177,  4180,  4192,  4202,  4225,  4228,  4240,  4352,  4354,  4357,  4360,  4369,  4372,
+         4384,  4417,  4420,  4432,  4480,  4500,  4502,  4609,  4612,  4614,  4624,  4672,  4704,  5120,  5122,  5125,
+         5128,  5137,  5140,  5152,  5185,  5188,  5193,  5200,  5220,  5248,  5377,  5380,  5392,  5440,  5632,  5652,
+         5705,  6145,  6148,  6160,  6162,  6208,  6228,  6278,  6400,  6405,  6502,  6737,  6825,  8192,  8194,  8197,
+         8200,  8202,  8209,  8212,  8224,  8257,  8260,  8272,  8320,  8352,  8449,  8452,  8464,  8512,  8520,  8549,
+         8704,  8738,  8832,  8872,  9217,  9220,  9232,  9257,  9280,  9472,  9537,  9554,  9625,  9729,  9754,  9894,
+        10240, 10248, 10250, 10272, 10325, 10376, 10402, 10600, 10640, 10760, 10784, 10882, 10888, 10890, 16385, 16388,
+        16390, 16393, 16400, 16402, 16405, 16408, 16417, 16420, 16448, 16450, 16453, 16456, 16458, 16465, 16468, 16480,
+        16485, 16513, 16516, 16528, 16640, 16642, 16645, 16648, 16657, 16660, 16672, 16705, 16708, 16720, 16768, 16773,
+        16802, 16897, 16900, 16912, 16914, 16937, 16960, 17408, 17410, 17413, 17416, 17425, 17428, 17433, 17440, 17473,
+        17476, 17488, 17536, 17556, 17665, 17668, 17680, 17700, 17728, 17818, 17920, 17930, 17988, 18000, 18433, 18436,
+        18448, 18496, 18501, 18516, 18530, 18688, 18705, 18756, 18768, 18793, 18948, 20480, 20482, 20485, 20488, 20497,
+        20500, 20512, 20520, 20545, 20548, 20560, 20608, 20737, 20740, 20752, 20757, 20800, 20802, 20992, 21060, 21162,
+        21505, 21508, 21520, 21537, 21568, 21600, 21633, 21665, 21760, 21768, 21888, 21896, 22049, 22120, 22177, 22528,
+        22548, 22593, 22608, 22681, 22810, 22848, 22850, 23173, 24577, 24580, 24592, 24640, 24660, 24674, 24710, 24745,
+        24832, 25124, 25162, 25234, 25600, 25622, 25872, 25920, 25925, 26020, 26625, 26730, 26917, 27142, 27220, 27234,
+        32768, 32770, 32773, 32776, 32785, 32788, 32800, 32810, 32833, 32836, 32848, 32896, 32898, 32936, 32938, 33025,
+        33028, 33030, 33040, 33088, 33105, 33113, 33280, 33312, 33408, 33410, 33440, 33448, 33793, 33796, 33808, 33810,
+        33813, 33856, 33888, 33929, 34048, 34116, 34213, 34328, 34410, 34816, 34824, 34853, 34906, 34944, 34946, 34984,
+        35078, 35362, 35456, 35464, 35478, 35496, 36865, 36868, 36880, 36928, 36950, 36996, 37120, 37154, 37220, 37462,
+        37513, 37888, 37893, 37956, 37968, 37976, 38185, 38288, 38290, 38465, 38993, 39078, 39241, 39445, 39520, 40960,
+        40962, 40968, 40970, 40992, 41002, 41120, 41297, 41305, 41382, 41472, 41474, 41480, 41514, 41600, 41632, 42048,
+        42133, 42597, 42648, 43018, 43040, 43042, 43048, 43168, 43176, 43268, 43396, 43398, 43560, 43562, 43665, 43690,
+    };
+    static const uint16_t kgrid_1bit_2048[NGRID_IQ1S] = {
+            0,     2,     5,     8,    10,    17,    21,    32,    34,    40,    42,    69,    81,    84,    86,   101,
+          128,   130,   136,   138,   149,   160,   162,   168,   170,   260,   261,   273,   276,   278,   281,   282,
+          293,   321,   326,   329,   338,   341,   346,   353,   356,   358,   360,   389,   401,   404,   406,   421,
+          512,   514,   520,   522,   533,   544,   546,   552,   554,   581,   593,   601,   612,   617,   640,   642,
+          648,   650,   657,   661,   665,   672,   674,   680,   682,  1041,  1044,  1046,  1061,  1089,  1097,  1109,
+         1114,  1124,  1125,  1169,  1177,  1189,  1281,  1284,  1285,  1286,  1301,  1304,  1306,  1321,  1344,  1349,
+         1354,  1360,  1361,  1364,  1365,  1366,  1369,  1376,  1378,  1381,  1384,  1386,  1409,  1425,  1429,  1432,
+         1434,  1441,  1444,  1445,  1446,  1449,  1556,  1561,  1601,  1604,  1616,  1618,  1621,  1624,  1632,  1633,
+         1638,  1641,  1669,  1681,  1684,  1689,  2048,  2050,  2056,  2058,  2069,  2080,  2082,  2088,  2090,  2117,
+         2129,  2134,  2149,  2176,  2178,  2184,  2186,  2197,  2208,  2210,  2216,  2218,  2309,  2321,  2324,  2329,
+         2340,  2341,  2369,  2384,  2385,  2389,  2401,  2404,  2409,  2449,  2452,  2454,  2457,  2469,  2560,  2562,
+         2568,  2570,  2581,  2592,  2594,  2600,  2602,  2629,  2641,  2649,  2657,  2661,  2688,  2690,  2693,  2696,
+         2698,  2709,  2720,  2722,  2728,  2730,  4112,  4113,  4116,  4121,  4132,  4133,  4161,  4164,  4176,  4181,
+         4184,  4193,  4196,  4197,  4201,  4241,  4244,  4246,  4257,  4261,  4353,  4356,  4358,  4361,  4368,  4370,
+         4373,  4376,  4385,  4388,  4393,  4421,  4426,  4432,  4433,  4434,  4436,  4437,  4438,  4441,  4448,  4453,
+         4484,  4498,  4501,  4513,  4516,  4625,  4628,  4630,  4645,  4672,  4678,  4681,  4690,  4693,  4696,  4698,
+         4708,  4710,  4741,  4753,  4756,  4758,  4773,  5121,  5126,  5129,  5140,  5141,  5144,  5145,  5153,  5158,
+         5185,  5189,  5190,  5192,  5194,  5201,  5204,  5205,  5206,  5209,  5218,  5221,  5224,  5252,  5257,  5264,
+         5268,  5269,  5272,  5273,  5274,  5281,  5284,  5285,  5289,  5378,  5381,  5386,  5393,  5396,  5397,  5398,
+         5401,  5408,  5410,  5413,  5416,  5418,  5441,  5444,  5445,  5446,  5457,  5458,  5460,  5461,  5462,  5465,
+         5466,  5473,  5476,  5477,  5478,  5481,  5504,  5506,  5508,  5509,  5512,  5514,  5520,  5521,  5524,  5525,
+         5526,  5529,  5530,  5536,  5538,  5541,  5633,  5636,  5637,  5638,  5653,  5654,  5656,  5658,  5665,  5670,
+         5696,  5698,  5700,  5701,  5704,  5706,  5713,  5717,  5718,  5720,  5721,  5729,  5732,  5733,  5736,  5737,
+         5738,  5766,  5770,  5778,  5781,  5796,  5801,  6161,  6166,  6181,  6209,  6212,  6214,  6217,  6224,  6229,
+         6232,  6234,  6240,  6241,  6244,  6246,  6249,  6277,  6289,  6292,  6309,  6416,  6418,  6421,  6426,  6433,
+         6437,  6466,  6468,  6469,  6472,  6481,  6484,  6485,  6486,  6489,  6490,  6496,  6501,  6506,  6537,  6545,
+         6546,  6549,  6552,  6561,  6566,  6569,  6665,  6678,  6692,  6694,  6724,  6726,  6729,  6736,  6738,  6741,
+         6744,  6753,  6758,  6761,  6789,  6801,  6806,  6810,  8192,  8194,  8200,  8202,  8213,  8224,  8226,  8229,
+         8232,  8234,  8261,  8273,  8281,  8289,  8293,  8320,  8322,  8328,  8330,  8341,  8352,  8354,  8357,  8360,
+         8362,  8453,  8465,  8468,  8473,  8485,  8514,  8516,  8521,  8533,  8536,  8538,  8545,  8548,  8549,  8550,
+         8581,  8592,  8598,  8601,  8613,  8705,  8712,  8714,  8721,  8725,  8736,  8738,  8744,  8746,  8773,  8785,
+         8790,  8793,  8805,  8833,  8840,  8842,  8849,  8853,  8864,  8866,  8872,  8874,  9221,  9236,  9238,  9241,
+         9253,  9284,  9285,  9286,  9289,  9298,  9301,  9304,  9306,  9318,  9349,  9361,  9364,  9369,  9377,  9381,
+         9481,  9493,  9505,  9513,  9536,  9541,  9544,  9553,  9556,  9557,  9561,  9570,  9573,  9576,  9609,  9616,
+         9620,  9621,  9624,  9626,  9633,  9636,  9638,  9641,  9733,  9744,  9746,  9753,  9765,  9793,  9801,  9813,
+         9824,  9825,  9833,  9860,  9862,  9872,  9882, 10240, 10242, 10248, 10250, 10261, 10272, 10274, 10280, 10282,
+        10309, 10321, 10324, 10341, 10368, 10370, 10376, 10378, 10400, 10402, 10408, 10410, 10505, 10513, 10516, 10521,
+        10533, 10566, 10569, 10578, 10581, 10593, 10596, 10598, 10601, 10629, 10640, 10646, 10649, 10660, 10661, 10752,
+        10754, 10760, 10762, 10784, 10786, 10792, 10794, 10821, 10833, 10838, 10841, 10853, 10880, 10882, 10888, 10890,
+        10901, 10912, 10914, 10920, 10922, 16389, 16401, 16406, 16421, 16457, 16466, 16469, 16472, 16474, 16481, 16484,
+        16486, 16532, 16537, 16545, 16550, 16640, 16641, 16644, 16646, 16649, 16658, 16661, 16662, 16664, 16666, 16673,
+        16678, 16681, 16709, 16712, 16714, 16721, 16724, 16725, 16726, 16729, 16730, 16741, 16744, 16746, 16769, 16772,
+        16774, 16784, 16786, 16789, 16800, 16801, 16802, 16901, 16913, 16916, 16918, 16933, 16961, 16978, 16981, 16986,
+        16996, 17001, 17033, 17044, 17061, 17409, 17429, 17433, 17449, 17477, 17480, 17482, 17489, 17492, 17493, 17494,
+        17505, 17506, 17509, 17512, 17514, 17537, 17542, 17545, 17552, 17554, 17557, 17568, 17569, 17577, 17665, 17666,
+        17669, 17674, 17681, 17684, 17685, 17686, 17689, 17696, 17701, 17706, 17729, 17732, 17733, 17734, 17737, 17744,
+        17745, 17748, 17749, 17750, 17752, 17753, 17761, 17764, 17765, 17766, 17769, 17794, 17796, 17797, 17800, 17809,
+        17812, 17813, 17814, 17817, 17818, 17829, 17832, 17834, 17921, 17925, 17929, 17940, 17941, 17944, 17946, 17953,
+        17956, 17961, 17984, 17986, 17989, 17992, 18000, 18001, 18002, 18005, 18006, 18009, 18018, 18021, 18024, 18049,
+        18053, 18058, 18068, 18069, 18081, 18084, 18086, 18437, 18449, 18453, 18458, 18469, 18498, 18505, 18512, 18517,
+        18520, 18529, 18532, 18534, 18537, 18565, 18577, 18580, 18582, 18585, 18597, 18689, 18693, 18694, 18698, 18704,
+        18708, 18709, 18712, 18721, 18724, 18726, 18752, 18757, 18762, 18769, 18770, 18772, 18773, 18774, 18777, 18784,
+        18786, 18789, 18790, 18794, 18822, 18825, 18834, 18837, 18838, 18840, 18849, 18852, 18854, 18857, 18966, 19012,
+        19014, 19017, 19029, 19032, 19034, 19044, 19049, 19092, 19109, 20481, 20484, 20485, 20486, 20489, 20498, 20501,
+        20506, 20513, 20516, 20521, 20544, 20549, 20552, 20561, 20564, 20565, 20566, 20569, 20581, 20584, 20614, 20617,
+        20629, 20632, 20640, 20641, 20646, 20649, 20741, 20744, 20745, 20746, 20753, 20756, 20757, 20758, 20760, 20761,
+        20768, 20773, 20774, 20776, 20778, 20801, 20804, 20805, 20806, 20809, 20816, 20817, 20818, 20820, 20821, 20822,
+        20824, 20825, 20826, 20833, 20836, 20837, 20838, 20841, 20866, 20869, 20881, 20884, 20885, 20886, 20889, 20896,
+        20901, 20906, 20993, 20998, 21010, 21013, 21018, 21025, 21028, 21058, 21061, 21066, 21073, 21076, 21077, 21078,
+        21081, 21090, 21093, 21125, 21136, 21138, 21141, 21145, 21146, 21156, 21508, 21509, 21521, 21524, 21525, 21526,
+        21528, 21529, 21537, 21541, 21544, 21546, 21569, 21572, 21573, 21574, 21577, 21578, 21584, 21585, 21588, 21589,
+        21590, 21592, 21593, 21594, 21601, 21602, 21604, 21605, 21606, 21609, 21632, 21640, 21642, 21649, 21652, 21653,
+        21654, 21657, 21665, 21668, 21669, 21674, 21761, 21762, 21764, 21765, 21766, 21769, 21776, 21777, 21778, 21780,
+        21781, 21782, 21785, 21786, 21793, 21796, 21797, 21798, 21801, 21824, 21825, 21826, 21828, 21829, 21830, 21832,
+        21833, 21840, 21841, 21842, 21844, 21845, 21846, 21848, 21849, 21850, 21856, 21857, 21860, 21861, 21862, 21864,
+        21865, 21866, 21889, 21892, 21893, 21897, 21898, 21904, 21905, 21908, 21909, 21910, 21912, 21913, 21921, 21924,
+        21925, 21926, 21929, 22016, 22017, 22018, 22020, 22022, 22024, 22025, 22033, 22036, 22037, 22040, 22041, 22048,
+        22049, 22050, 22052, 22053, 22054, 22056, 22057, 22081, 22085, 22086, 22088, 22089, 22090, 22096, 22097, 22098,
+        22100, 22101, 22102, 22104, 22105, 22106, 22113, 22116, 22117, 22121, 22146, 22149, 22150, 22152, 22153, 22154,
+        22161, 22165, 22170, 22178, 22181, 22182, 22184, 22185, 22532, 22533, 22534, 22537, 22544, 22549, 22552, 22561,
+        22570, 22597, 22600, 22602, 22609, 22612, 22613, 22614, 22616, 22617, 22624, 22626, 22628, 22629, 22658, 22665,
+        22672, 22674, 22677, 22680, 22689, 22697, 22785, 22786, 22789, 22794, 22801, 22804, 22805, 22806, 22809, 22821,
+        22849, 22852, 22853, 22854, 22857, 22864, 22865, 22866, 22868, 22869, 22870, 22872, 22873, 22874, 22881, 22884,
+        22885, 22886, 22889, 22913, 22917, 22921, 22929, 22932, 22933, 22934, 22936, 22937, 22949, 23044, 23048, 23061,
+        23066, 23072, 23077, 23078, 23081, 23109, 23112, 23113, 23121, 23125, 23126, 23128, 23129, 23138, 23141, 23144,
+        23146, 23169, 23178, 23186, 23189, 23190, 23192, 23194, 23201, 24581, 24596, 24598, 24601, 24613, 24644, 24656,
+        24661, 24662, 24664, 24666, 24673, 24676, 24678, 24681, 24705, 24726, 24741, 24833, 24836, 24838, 24841, 24850,
+        24853, 24865, 24866, 24870, 24873, 24901, 24905, 24913, 24917, 24918, 24921, 24933, 24934, 24938, 24964, 24970,
+        24978, 24981, 24993, 24998, 25001, 25105, 25110, 25113, 25152, 25153, 25158, 25173, 25174, 25176, 25184, 25221,
+        25233, 25238, 25253, 25617, 25618, 25621, 25622, 25626, 25633, 25638, 25641, 25664, 25666, 25669, 25672, 25674,
+        25681, 25684, 25685, 25686, 25689, 25690, 25696, 25698, 25701, 25732, 25733, 25737, 25744, 25746, 25748, 25749,
+        25750, 25752, 25754, 25761, 25764, 25769, 25861, 25864, 25866, 25873, 25877, 25878, 25881, 25924, 25925, 25926,
+        25929, 25936, 25937, 25940, 25941, 25942, 25945, 25953, 25956, 25957, 25958, 25961, 25990, 25993, 25994, 26001,
+        26005, 26006, 26009, 26010, 26018, 26021, 26022, 26024, 26114, 26121, 26133, 26144, 26150, 26152, 26153, 26176,
+        26181, 26184, 26186, 26193, 26196, 26197, 26198, 26200, 26202, 26208, 26213, 26216, 26240, 26242, 26245, 26250,
+        26260, 26262, 26264, 26265, 26272, 26276, 26278, 26282, 26646, 26649, 26661, 26689, 26706, 26709, 26714, 26721,
+        26729, 26757, 26769, 26776, 26790, 26881, 26884, 26896, 26901, 26913, 26916, 26918, 26921, 26944, 26945, 26949,
+        26950, 26952, 26961, 26964, 26965, 26966, 26969, 26976, 26981, 26986, 27010, 27012, 27018, 27029, 27041, 27044,
+        27045, 27049, 27153, 27158, 27160, 27201, 27204, 27209, 27216, 27221, 27224, 27226, 27236, 27237, 27241, 27270,
+        27284, 27288, 27290, 27302, 32768, 32770, 32776, 32778, 32800, 32802, 32808, 32810, 32837, 32848, 32849, 32852,
+        32854, 32857, 32869, 32896, 32898, 32904, 32906, 32917, 32928, 32930, 32936, 32938, 33029, 33041, 33044, 33046,
+        33049, 33061, 33089, 33092, 33097, 33104, 33106, 33109, 33110, 33112, 33113, 33124, 33126, 33129, 33157, 33161,
+        33172, 33174, 33177, 33189, 33280, 33282, 33288, 33290, 33301, 33312, 33314, 33320, 33322, 33361, 33364, 33369,
+        33381, 33408, 33410, 33416, 33418, 33429, 33440, 33442, 33448, 33450, 33812, 33817, 33857, 33860, 33873, 33877,
+        33882, 33889, 33892, 33897, 33940, 33945, 34049, 34057, 34066, 34069, 34074, 34086, 34089, 34112, 34113, 34117,
+        34120, 34129, 34132, 34133, 34134, 34137, 34138, 34149, 34150, 34152, 34154, 34177, 34180, 34182, 34185, 34192,
+        34194, 34197, 34200, 34214, 34321, 34326, 34329, 34341, 34369, 34372, 34377, 34378, 34384, 34389, 34393, 34394,
+        34401, 34406, 34410, 34437, 34449, 34458, 34468, 34816, 34818, 34824, 34826, 34837, 34848, 34850, 34856, 34858,
+        34881, 34885, 34897, 34900, 34905, 34917, 34921, 34944, 34946, 34952, 34954, 34965, 34976, 34978, 34984, 34986,
+        35077, 35078, 35089, 35092, 35094, 35109, 35137, 35140, 35142, 35145, 35152, 35154, 35157, 35162, 35169, 35172,
+        35205, 35222, 35225, 35237, 35328, 35330, 35336, 35338, 35349, 35360, 35362, 35368, 35370, 35397, 35409, 35412,
+        35414, 35456, 35458, 35464, 35466, 35477, 35488, 35490, 35496, 35498, 36869, 36881, 36886, 36888, 36889, 36901,
+        36929, 36934, 36937, 36949, 36952, 36954, 36969, 36970, 36997, 37009, 37012, 37014, 37017, 37029, 37121, 37124,
+        37126, 37129, 37136, 37141, 37144, 37146, 37153, 37156, 37158, 37161, 37184, 37189, 37200, 37201, 37204, 37205,
+        37206, 37209, 37218, 37221, 37252, 37254, 37266, 37269, 37272, 37281, 37284, 37286, 37289, 37381, 37393, 37396,
+        37401, 37413, 37444, 37446, 37449, 37456, 37458, 37461, 37464, 37478, 37481, 37509, 37524, 37526, 37545, 37889,
+        37892, 37894, 37904, 37909, 37912, 37926, 37952, 37962, 37969, 37972, 37973, 37974, 37976, 37977, 37984, 37985,
+        37986, 37989, 38020, 38022, 38034, 38036, 38037, 38040, 38049, 38057, 38144, 38149, 38152, 38154, 38160, 38161,
+        38164, 38165, 38166, 38169, 38177, 38181, 38185, 38186, 38209, 38212, 38213, 38214, 38217, 38224, 38225, 38226,
+        38228, 38229, 38230, 38232, 38233, 38234, 38241, 38244, 38245, 38246, 38249, 38273, 38277, 38280, 38289, 38290,
+        38292, 38293, 38294, 38297, 38298, 38304, 38306, 38309, 38312, 38314, 38401, 38404, 38416, 38421, 38425, 38432,
+        38438, 38441, 38469, 38472, 38473, 38481, 38482, 38485, 38486, 38489, 38501, 38504, 38530, 38532, 38537, 38538,
+        38546, 38548, 38549, 38564, 38566, 38569, 38917, 38934, 38937, 38949, 38977, 38982, 38992, 38994, 38997, 38998,
+        39002, 39012, 39013, 39045, 39057, 39062, 39065, 39077, 39172, 39174, 39177, 39184, 39186, 39189, 39192, 39194,
+        39200, 39201, 39204, 39206, 39232, 39234, 39237, 39240, 39242, 39249, 39252, 39253, 39254, 39257, 39266, 39269,
+        39270, 39274, 39297, 39300, 39312, 39314, 39317, 39322, 39329, 39334, 39429, 39445, 39461, 39492, 39494, 39497,
+        39504, 39509, 39512, 39521, 39557, 39569, 39572, 39573, 39574, 40960, 40962, 40968, 40970, 40981, 40992, 40994,
+        41000, 41002, 41029, 41041, 41044, 41046, 41049, 41088, 41090, 41096, 41098, 41109, 41120, 41122, 41128, 41130,
+        41221, 41225, 41233, 41236, 41238, 41241, 41242, 41286, 41289, 41297, 41301, 41304, 41306, 41313, 41316, 41349,
+        41360, 41362, 41366, 41369, 41474, 41480, 41482, 41488, 41497, 41506, 41512, 41514, 41541, 41553, 41558, 41561,
+        41573, 41600, 41602, 41608, 41610, 41621, 41632, 41634, 41640, 41642, 42009, 42021, 42049, 42052, 42064, 42068,
+        42069, 42072, 42074, 42081, 42085, 42086, 42088, 42089, 42117, 42246, 42249, 42256, 42258, 42261, 42264, 42278,
+        42281, 42306, 42309, 42321, 42324, 42325, 42326, 42329, 42341, 42346, 42369, 42372, 42373, 42374, 42377, 42386,
+        42389, 42392, 42501, 42513, 42518, 42522, 42529, 42533, 42564, 42566, 42570, 42578, 42581, 42582, 42584, 42592,
+        42594, 42630, 42640, 42645, 42646, 42649, 42657, 42660, 42662, 43008, 43010, 43016, 43018, 43040, 43042, 43048,
+        43050, 43089, 43092, 43094, 43097, 43136, 43138, 43144, 43146, 43157, 43168, 43170, 43176, 43178, 43269, 43284,
+        43289, 43297, 43301, 43329, 43344, 43349, 43354, 43361, 43366, 43369, 43408, 43414, 43520, 43522, 43528, 43530,
+        43552, 43554, 43560, 43562, 43601, 43604, 43606, 43648, 43650, 43656, 43658, 43669, 43680, 43682, 43688, 43690,
+    };
+    static const uint16_t kgrid_2bit_1024[1024] = {
+            0,     2,     5,     8,    10,    17,    20,    22,    25,    32,    34,    37,    40,    65,    68,    70,
+           73,    80,    82,    85,    88,    97,   100,   102,   105,   128,   130,   133,   136,   145,   148,   160,
+          165,   170,   257,   260,   262,   265,   272,   274,   277,   280,   289,   292,   320,   322,   325,   328,
+          337,   340,   342,   345,   352,   357,   360,   385,   388,   400,   402,   405,   417,   420,   512,   514,
+          517,   520,   529,   532,   544,   554,   577,   580,   582,   585,   592,   597,   640,   645,   650,   660,
+          674,  1025,  1028,  1030,  1033,  1040,  1042,  1045,  1048,  1057,  1060,  1062,  1065,  1088,  1090,  1093,
+         1096,  1098,  1105,  1108,  1110,  1113,  1120,  1122,  1125,  1153,  1156,  1158,  1161,  1168,  1173,  1176,
+         1185,  1188,  1280,  1282,  1285,  1288,  1290,  1297,  1300,  1302,  1305,  1312,  1317,  1320,  1345,  1348,
+         1350,  1353,  1360,  1362,  1365,  1368,  1377,  1380,  1408,  1410,  1413,  1416,  1425,  1428,  1440,  1537,
+         1540,  1542,  1545,  1552,  1557,  1600,  1605,  1608,  1617,  1620,  1632,  1665,  1668,  1680,  2048,  2050,
+         2053,  2056,  2065,  2068,  2070,  2073,  2080,  2085,  2090,  2113,  2116,  2118,  2121,  2128,  2130,  2133,
+         2136,  2145,  2148,  2176,  2181,  2196,  2218,  2305,  2308,  2320,  2322,  2325,  2328,  2337,  2368,  2373,
+         2376,  2385,  2388,  2400,  2433,  2448,  2560,  2577,  2580,  2594,  2600,  2602,  2640,  2713,  4097,  4100,
+         4102,  4105,  4112,  4114,  4117,  4120,  4129,  4132,  4134,  4160,  4162,  4165,  4168,  4177,  4180,  4182,
+         4185,  4192,  4194,  4197,  4200,  4225,  4228,  4230,  4240,  4245,  4248,  4257,  4260,  4352,  4354,  4357,
+         4360,  4362,  4369,  4372,  4374,  4377,  4384,  4386,  4389,  4392,  4417,  4420,  4422,  4425,  4432,  4434,
+         4437,  4440,  4449,  4452,  4480,  4482,  4485,  4488,  4497,  4500,  4609,  4612,  4617,  4624,  4629,  4641,
+         4644,  4672,  4677,  4689,  4692,  4737,  4740,  4752,  5120,  5122,  5125,  5128,  5137,  5140,  5142,  5145,
+         5152,  5157,  5160,  5185,  5188,  5190,  5193,  5200,  5202,  5205,  5208,  5217,  5220,  5248,  5250,  5253,
+         5256,  5265,  5268,  5280,  5377,  5380,  5382,  5385,  5392,  5394,  5397,  5400,  5409,  5412,  5440,  5442,
+         5445,  5448,  5457,  5460,  5472,  5505,  5508,  5520,  5632,  5637,  5640,  5649,  5652,  5664,  5697,  5700,
+         5712,  5760,  5802,  6145,  6148,  6150,  6153,  6160,  6165,  6168,  6177,  6208,  6210,  6213,  6216,  6225,
+         6228,  6240,  6273,  6276,  6400,  6402,  6405,  6408,  6417,  6420,  6432,  6465,  6468,  6480,  6505,  6562,
+         6660,  6672,  6720,  6742,  8192,  8194,  8197,  8200,  8209,  8212,  8214,  8217,  8224,  8229,  8234,  8257,
+         8260,  8272,  8274,  8277,  8292,  8320,  8330,  8340,  8362,  8449,  8452,  8464,  8466,  8469,  8481,  8512,
+         8514,  8517,  8529,  8532,  8544,  8577,  8580,  8592,  8704,  8714,  8738,  8744,  8746,  8772,  8784,  8840,
+         8842,  8872,  9217,  9220,  9222,  9225,  9232,  9237,  9240,  9249,  9252,  9280,  9282,  9285,  9288,  9297,
+         9300,  9312,  9345,  9348,  9360,  9472,  9477,  9480,  9489,  9492,  9504,  9537,  9540,  9552,  9574,  9600,
+         9729,  9732,  9744,  9792,  9817, 10240, 10245, 10257, 10260, 10305, 10308, 10320, 10378, 10410, 10497, 10500,
+        10512, 10645, 10762, 10786, 10852, 10888, 10890, 16385, 16388, 16390, 16393, 16400, 16402, 16405, 16408, 16410,
+        16417, 16420, 16422, 16448, 16450, 16453, 16456, 16458, 16465, 16468, 16470, 16473, 16480, 16482, 16485, 16513,
+        16516, 16528, 16533, 16536, 16545, 16548, 16640, 16642, 16645, 16648, 16657, 16660, 16662, 16665, 16672, 16674,
+        16677, 16705, 16708, 16710, 16713, 16720, 16722, 16725, 16728, 16737, 16740, 16768, 16770, 16773, 16776, 16785,
+        16788, 16800, 16897, 16900, 16912, 16914, 16917, 16920, 16932, 16960, 16965, 16968, 16977, 16980, 16992, 17025,
+        17028, 17408, 17410, 17413, 17416, 17418, 17425, 17428, 17430, 17433, 17440, 17442, 17445, 17448, 17473, 17476,
+        17478, 17481, 17488, 17490, 17493, 17496, 17505, 17508, 17536, 17538, 17541, 17544, 17553, 17556, 17568, 17665,
+        17668, 17670, 17673, 17680, 17682, 17685, 17688, 17697, 17700, 17728, 17730, 17733, 17736, 17745, 17748, 17760,
+        17770, 17793, 17796, 17808, 17920, 17922, 17925, 17928, 17937, 17940, 17952, 17985, 17988, 18000, 18048, 18085,
+        18433, 18436, 18441, 18448, 18450, 18453, 18456, 18465, 18468, 18496, 18498, 18501, 18504, 18513, 18516, 18528,
+        18564, 18576, 18688, 18690, 18693, 18696, 18705, 18708, 18720, 18753, 18756, 18768, 18816, 18838, 18945, 18948,
+        18960, 19008, 20480, 20482, 20485, 20488, 20497, 20500, 20502, 20505, 20512, 20514, 20517, 20520, 20545, 20548,
+        20550, 20553, 20560, 20562, 20565, 20568, 20577, 20580, 20608, 20610, 20613, 20616, 20625, 20628, 20737, 20740,
+        20742, 20745, 20752, 20754, 20757, 20760, 20769, 20772, 20800, 20802, 20805, 20808, 20817, 20820, 20832, 20865,
+        20868, 20880, 20992, 20997, 21000, 21009, 21012, 21024, 21057, 21060, 21072, 21097, 21120, 21505, 21508, 21510,
+        21513, 21520, 21522, 21525, 21528, 21537, 21540, 21568, 21570, 21573, 21576, 21585, 21588, 21600, 21633, 21636,
+        21648, 21760, 21762, 21765, 21768, 21777, 21780, 21792, 21825, 21828, 21840, 21888, 22017, 22020, 22032, 22054,
+        22080, 22528, 22530, 22533, 22536, 22545, 22548, 22560, 22593, 22596, 22608, 22618, 22656, 22785, 22788, 22800,
+        22848, 23040, 23065, 23173, 23208, 24577, 24580, 24582, 24592, 24594, 24597, 24600, 24609, 24612, 24640, 24645,
+        24648, 24657, 24660, 24672, 24708, 24720, 24832, 24834, 24837, 24840, 24849, 24852, 24864, 24897, 24900, 24912,
+        24960, 24985, 25092, 25104, 25152, 25174, 25249, 25600, 25605, 25608, 25617, 25620, 25632, 25665, 25668, 25680,
+        25728, 25857, 25860, 25872, 25920, 25930, 25960, 26002, 26112, 26260, 26625, 26628, 26640, 26725, 26776, 26880,
+        26922, 27202, 27297, 32768, 32770, 32773, 32776, 32785, 32788, 32793, 32800, 32805, 32833, 32836, 32848, 32850,
+        32853, 32856, 32865, 32896, 32901, 32913, 32916, 33025, 33028, 33033, 33040, 33042, 33045, 33048, 33057, 33060,
+        33088, 33090, 33093, 33096, 33105, 33108, 33153, 33156, 33168, 33193, 33280, 33285, 33290, 33297, 33300, 33345,
+        33348, 33360, 33793, 33796, 33798, 33801, 33808, 33810, 33813, 33816, 33825, 33856, 33858, 33861, 33864, 33873,
+        33876, 33888, 33921, 33924, 33936, 34048, 34050, 34053, 34056, 34065, 34068, 34080, 34113, 34116, 34128, 34176,
+        34186, 34305, 34308, 34320, 34345, 34368, 34816, 34821, 34833, 34836, 34881, 34884, 34896, 34978, 35073, 35076,
+        35136, 35173, 35362, 35416, 35418, 35458, 35490, 36865, 36868, 36873, 36880, 36882, 36885, 36888, 36900, 36928,
+        36930, 36933, 36936, 36945, 36948, 36960, 36993, 36996, 37008, 37120, 37125, 37137, 37140, 37185, 37188, 37200,
+        37210, 37377, 37380, 37392, 37440, 37542, 37888, 37890, 37893, 37896, 37905, 37908, 37920, 37953, 37956, 37968,
+        38016, 38038, 38145, 38148, 38160, 38208, 38296, 38305, 38400, 38470, 38500, 38913, 38916, 38928, 38950, 38976,
+        39081, 39168, 39241, 39250, 39568, 40960, 40965, 40970, 40980, 40994, 41002, 41025, 41028, 41040, 41122, 41130,
+        41280, 41317, 41474, 41482, 41506, 41512, 41514, 41602, 41608, 41610, 41640, 41985, 41988, 42000, 42048, 42121,
+        42148, 42240, 42265, 42577, 43018, 43048, 43170, 43348, 43398, 43528, 43530, 43552, 43554, 43560, 43656, 43690,
+    };
+
+    const int kmap_size = 43692;
+    //const int nwant = type == GGML_TYPE_IQ1_S ? 3 : 2;
+    const int nwant = type == GGML_TYPE_IQ1_S || type == GGML_TYPE_IQ1_M ? 3 : type == GGML_TYPE_IQ2_S ? 1 : 2;
+    const uint16_t * kgrid = type == GGML_TYPE_IQ2_XXS ? kgrid_2bit_256 :
+                             type == GGML_TYPE_IQ2_XS  ? kgrid_2bit_512 :
+                             type == GGML_TYPE_IQ1_S || type == GGML_TYPE_IQ1_M ? kgrid_1bit_2048 : kgrid_2bit_1024;
+    uint64_t * kgrid_q2xs;
+    int      * kmap_q2xs;
+    uint16_t * kneighbors_q2xs;
+
+    //printf("================================================================= %s(grid_size = %d)\n", __func__, grid_size);
+    uint64_t * the_grid = (uint64_t *)malloc(grid_size*sizeof(uint64_t));
+    for (int k = 0; k < grid_size; ++k) {
+        int8_t * pos = (int8_t *)(the_grid + k);
+        for (int i = 0; i < 8; ++i) {
+            int l = (kgrid[k] >> 2*i) & 0x3;
+            pos[i] = 2*l + 1;
+        }
+    }
+    kgrid_q2xs = the_grid;
+    iq2_data[gindex].grid = the_grid;
+    kmap_q2xs = (int *)malloc(kmap_size*sizeof(int));
+    iq2_data[gindex].map = kmap_q2xs;
+    for (int i = 0; i < kmap_size; ++i) kmap_q2xs[i] = -1;
+    uint64_t aux64;
+    uint8_t * aux8 = (uint8_t *)&aux64;
+    for (int i = 0; i < grid_size; ++i) {
+        aux64 = kgrid_q2xs[i];
+        uint16_t index = 0;
+        for (int k=0; k<8; ++k) {
+            uint16_t q = (aux8[k] - 1)/2;
+            index |= (q << 2*k);
+        }
+        kmap_q2xs[index] = i;
+    }
+    int8_t pos[8];
+    int * dist2 = (int *)malloc(2*grid_size*sizeof(int));
+    int num_neighbors = 0, num_not_in_map = 0;
+    for (int i = 0; i < kmap_size; ++i) {
+        if (kmap_q2xs[i] >= 0) continue;
+        ++num_not_in_map;
+        for (int k = 0; k < 8; ++k) {
+            int l = (i >> 2*k) & 0x3;
+            pos[k] = 2*l + 1;
+        }
+        for (int j = 0; j < grid_size; ++j) {
+            const int8_t * pg = (const int8_t *)(kgrid_q2xs + j);
+            int d2 = 0;
+            for (int k = 0; k < 8; ++k) d2 += (pg[k] - pos[k])*(pg[k] - pos[k]);
+            dist2[2*j+0] = d2;
+            dist2[2*j+1] = j;
+        }
+        qsort(dist2, grid_size, 2*sizeof(int), iq2_compare_func);
+        int n = 0; int d2 = dist2[0];
+        int nhave = 1;
+        for (int j = 0; j < grid_size; ++j) {
+            if (dist2[2*j] > d2) {
+                if (nhave == nwant) break;
+                d2 = dist2[2*j];
+                ++nhave;
+            }
+            ++n;
+        }
+        num_neighbors += n;
+    }
+    //printf("%s: %d neighbours in total\n", __func__, num_neighbors);
+    kneighbors_q2xs = (uint16_t *)malloc((num_neighbors + num_not_in_map)*sizeof(uint16_t));
+    iq2_data[gindex].neighbours = kneighbors_q2xs;
+    int counter = 0;
+    for (int i = 0; i < kmap_size; ++i) {
+        if (kmap_q2xs[i] >= 0) continue;
+        for (int k = 0; k < 8; ++k) {
+            int l = (i >> 2*k) & 0x3;
+            pos[k] = 2*l + 1;
+        }
+        for (int j = 0; j < grid_size; ++j) {
+            const int8_t * pg = (const int8_t *)(kgrid_q2xs + j);
+            int d2 = 0;
+            for (int k = 0; k < 8; ++k) d2 += (pg[k] - pos[k])*(pg[k] - pos[k]);
+            dist2[2*j+0] = d2;
+            dist2[2*j+1] = j;
+        }
+        qsort(dist2, grid_size, 2*sizeof(int), iq2_compare_func);
+        kmap_q2xs[i] = -(counter + 1);
+        int d2 = dist2[0];
+        uint16_t * start = &kneighbors_q2xs[counter++];
+        int n = 0, nhave = 1;
+        for (int j = 0; j < grid_size; ++j) {
+            if (dist2[2*j] > d2) {
+                if (nhave == nwant) break;
+                d2 = dist2[2*j];
+                ++nhave;
+            }
+            kneighbors_q2xs[counter++] = dist2[2*j+1];
+            ++n;
+        }
+        *start = n;
+    }
+    free(dist2);
+}
+
+void iq2xs_free_impl(enum ggml_type type) {
+    GGML_ASSERT(type == GGML_TYPE_IQ2_XXS || type == GGML_TYPE_IQ2_XS || type == GGML_TYPE_IQ1_S || type == GGML_TYPE_IQ1_M || type == GGML_TYPE_IQ2_S);
+    const int gindex = iq2_data_index(type);
+    if (iq2_data[gindex].grid) {
+        free(iq2_data[gindex].grid);       iq2_data[gindex].grid = NULL;
+        free(iq2_data[gindex].map);        iq2_data[gindex].map  = NULL;
+        free(iq2_data[gindex].neighbours); iq2_data[gindex].neighbours = NULL;
+    }
+}
+
+static int iq2_find_best_neighbour(const uint16_t * restrict neighbours, const uint64_t * restrict grid,
+        const float * restrict xval, const float * restrict weight, float scale, int8_t * restrict L) {
+    int num_neighbors = neighbours[0];
+    GGML_ASSERT(num_neighbors > 0);
+    float best_d2 = FLT_MAX;
+    int grid_index = -1;
+    for (int j = 1; j <= num_neighbors; ++j) {
+        const int8_t * pg = (const int8_t *)(grid + neighbours[j]);
+        float d2 = 0;
+        for (int i = 0; i < 8; ++i) {
+            float q = pg[i];
+            float diff = scale*q - xval[i];
+            d2 += weight[i]*diff*diff;
+        }
+        if (d2 < best_d2) {
+            best_d2 = d2; grid_index = neighbours[j];
+        }
+    }
+    GGML_ASSERT(grid_index >= 0);
+    const int8_t * pg = (const int8_t *)(grid + grid_index);
+    for (int i = 0; i < 8; ++i) L[i] = (pg[i] - 1)/2;
+    return grid_index;
+}
+
+static void quantize_row_iq2_xxs_impl(const float * restrict x, void * restrict vy, int64_t n, const float * restrict quant_weights) {
+
+    const int gindex = iq2_data_index(GGML_TYPE_IQ2_XXS);
+
+    const uint64_t * kgrid_q2xs      = iq2_data[gindex].grid;
+    const int      * kmap_q2xs       = iq2_data[gindex].map;
+    const uint16_t * kneighbors_q2xs = iq2_data[gindex].neighbours;
+
+    GGML_ASSERT(quant_weights   && "missing quantization weights");
+    GGML_ASSERT(kgrid_q2xs      && "forgot to call ggml_quantize_init()?");
+    GGML_ASSERT(kmap_q2xs       && "forgot to call ggml_quantize_init()?");
+    GGML_ASSERT(kneighbors_q2xs && "forgot to call ggml_quantize_init()?");
+    GGML_ASSERT(n%QK_K == 0);
+
+    const int kMaxQ = 3;
+
+    const int64_t nbl = n/QK_K;
+
+    block_iq2_xxs * y = vy;
+
+    float scales[QK_K/32];
+    float weight[32];
+    float xval[32];
+    int8_t L[32];
+    int8_t Laux[32];
+    float  waux[32];
+    uint8_t block_signs[4];
+    uint32_t q2[2*(QK_K/32)];
+
+    for (int ibl = 0; ibl < nbl; ++ibl) {
+
+        y[ibl].d = GGML_FP32_TO_FP16(0.f);
+        memset(q2, 0, QK_K/4);
+
+        float max_scale = 0;
+
+        const float * xbl = x + QK_K*ibl;
+        float sumx2 = 0;
+        for (int i = 0; i < QK_K; ++i) sumx2 += xbl[i]*xbl[i];
+        float sigma2 = sumx2/QK_K;
+
+        for (int ib = 0; ib < QK_K/32; ++ib) {
+            const float * xb = xbl + 32*ib;
+            const float * qw = quant_weights + QK_K*ibl + 32*ib;
+            for (int i = 0; i < 32; ++i) weight[i] = qw[i] * sqrtf(sigma2 + xb[i]*xb[i]);
+            for (int i = 0; i < 32; ++i) waux[i] = sqrtf(weight[i]);
+            for (int k = 0; k < 4; ++k) {
+                int nflip = 0;
+                uint8_t s = 0;
+                for (int i = 0; i < 8; ++i) {
+                    if (xb[8*k + i] >= 0) xval[8*k + i] = xb[8*k + i];
+                    else {
+                        xval[8*k + i] = -xb[8*k + i]; ++nflip; s |= (1 << i);
+                    }
+                }
+                if (nflip%2) {
+                    int imin = 0; float min = weight[8*k+imin]*xb[8*k+imin]*xb[8*k+imin];
+                    for (int i = 1; i < 8; ++i) {
+                        float ax = weight[8*k+i]*xb[8*k+i]*xb[8*k+i];
+                        if (ax < min) {
+                            min = ax; imin = i;
+                        }
+                    }
+                    xval[8*k+imin] = -xval[8*k+imin];
+                    s ^= (1 << imin);
+                }
+                block_signs[k] = s & 127;
+            }
+            float max = xval[0];
+            for (int i = 1; i < 32; ++i) max = MAX(max, xval[i]);
+            if (max < GROUP_MAX_EPS) {
+                scales[ib] = 0;
+                memset(L, 0, 32);
+                continue;
+            }
+            float scale = make_qp_quants(32, kMaxQ+1, xval, (uint8_t*)L, weight);
+            float eff_max = scale*kMaxQ;
+            float best = 0;
+            for (int is = -6; is <= 6; ++is) {
+                float id = (2*kMaxQ-1+is*0.1f)/eff_max;
+                float this_scale = 1/id;
+                for (int k = 0; k < 4; ++k) {
+                    for (int i = 0; i < 8; ++i) {
+                        int l = nearest_int(0.5f*(id*xval[8*k+i]-1));
+                        Laux[8*k+i] = MAX(0, MIN(kMaxQ-1, l));
+                    }
+                    uint16_t u = 0;
+                    for (int i = 0; i < 8; ++i) u |= (Laux[8*k+i] << 2*i);
+                    int grid_index = kmap_q2xs[u];
+                    if (grid_index < 0) {
+                        const uint16_t * neighbours = kneighbors_q2xs - kmap_q2xs[u] - 1;
+                        grid_index = iq2_find_best_neighbour(neighbours, kgrid_q2xs, xval + 8*k, waux + 8*k, this_scale, Laux + 8*k);
+                    }
+                }
+                float sumqx = 0, sumq2 = 0;
+                for (int i = 0; i < 32; ++i) {
+                    float w = weight[i];
+                    float q = 2*Laux[i] + 1;
+                    sumqx += w*xval[i]*q;
+                    sumq2 += w*q*q;
+                }
+                if (sumq2 > 0 && sumqx*sumqx > best*sumq2) {
+                    scale = sumqx/sumq2; best = scale*sumqx;
+                    memcpy(L, Laux, 32);
+                }
+            }
+            if (scale > 0) {
+                float id = 1/scale;
+                for (int k = 0; k < 4; ++k) {
+                    uint16_t u = 0;
+                    for (int i = 0; i < 8; ++i) {
+                        int l = nearest_int(0.5f*(id*xval[8*k+i]-1));
+                        l = MAX(0, MIN(kMaxQ-1, l));
+                        u |= (l << 2*i);
+                    }
+                    int grid_index = kmap_q2xs[u];
+                    if (grid_index < 0) {
+                        const uint16_t * neighbours = kneighbors_q2xs - kmap_q2xs[u] - 1;
+                        grid_index = iq2_find_best_neighbour(neighbours, kgrid_q2xs, xval + 8*k, waux + 8*k, scale, L + 8*k);
+                    }
+                    const int8_t * pg = (const int8_t *)(kgrid_q2xs + grid_index);
+                    for (int i = 0; i < 8; ++i) L[8*k+i] = (pg[i] - 1)/2;
+                }
+                float sumqx = 0, sumq2 = 0;
+                for (int i = 0; i < 32; ++i) {
+                    float w = weight[i];
+                    float q = 2*L[i] + 1;
+                    sumqx += w*xval[i]*q;
+                    sumq2 += w*q*q;
+                }
+                if (sumq2 > 0) scale = sumqx/sumq2;
+            }
+            if (scale < 0) {
+                // This should never happen, but just in case, flip scale so that it is positive (we use uint's to encode the scale)
+                // and correspondingly flip quant signs.
+                scale = -scale;
+                for (int k = 0; k < 4; ++k) block_signs[k] = (~block_signs[k]) & 127;
+            }
+            for (int k = 0; k < 4; ++k) {
+                uint16_t u = 0;
+                for (int i = 0; i < 8; ++i) u |= (L[8*k+i] << 2*i);
+                int grid_index = kmap_q2xs[u];
+                if (grid_index < 0) {
+                    printf("Oops: found point %u not on grid:", u);
+                    for (int i = 0; i < 8; ++i) printf(" %d", L[8*k+i]);
+                    printf("\n");
+                    GGML_ABORT("fatal error");
+                }
+                q2[2*ib+0] |= ((uint32_t) grid_index << 8*k);
+                q2[2*ib+1] |= (block_signs[k] << 7*k);
+            }
+            GGML_ASSERT(scale >= 0);
+            scales[ib] = scale;
+            max_scale = MAX(max_scale, scale);
+        }
+
+        if (!max_scale) {
+            memset(y[ibl].qs, 0, QK_K/4);
+            continue;
+        }
+
+        float d = max_scale/31;
+        y[ibl].d = GGML_FP32_TO_FP16(d);
+        float id = 1/d;
+        for (int ib = 0; ib < QK_K/32; ++ib) {
+            int l = nearest_int(0.5f*(id*scales[ib]-1));
+            l = MAX(0, MIN(15, l));
+            q2[2*ib+1] |= ((uint32_t)l << 28);
+        }
+        memcpy(y[ibl].qs, q2, QK_K/4);
+    }
+}
+
+static void quantize_row_iq2_xs_impl(const float * restrict x, void * restrict vy, int64_t n, const float * restrict quant_weights) {
+
+    const int gindex = iq2_data_index(GGML_TYPE_IQ2_XS);
+
+    const uint64_t * kgrid_q2xs      = iq2_data[gindex].grid;
+    const int      * kmap_q2xs       = iq2_data[gindex].map;
+    const uint16_t * kneighbors_q2xs = iq2_data[gindex].neighbours;
+
+    GGML_ASSERT(quant_weights   && "missing quantization weights");
+    GGML_ASSERT(kmap_q2xs       && "forgot to call ggml_quantize_init()?");
+    GGML_ASSERT(kgrid_q2xs      && "forgot to call ggml_quantize_init()?");
+    GGML_ASSERT(kneighbors_q2xs && "forgot to call ggml_quantize_init()?");
+    GGML_ASSERT(n%QK_K == 0);
+
+    const int kMaxQ = 3;
+
+    const int64_t nbl = n/QK_K;
+
+    block_iq2_xs * y = vy;
+
+    float scales[QK_K/16];
+    float weight[16];
+    float xval[16];
+    int8_t L[16];
+    int8_t Laux[16];
+    float  waux[16];
+    bool   is_on_grid[2];
+    bool   is_on_grid_aux[2];
+    uint8_t block_signs[2];
+    uint16_t q2[2*(QK_K/16)];
+
+    for (int ibl = 0; ibl < nbl; ++ibl) {
+
+        y[ibl].d = GGML_FP32_TO_FP16(0.f);
+        memset(q2, 0, QK_K/4);
+        memset(y[ibl].scales, 0, QK_K/32);
+
+        float max_scale = 0;
+
+        const float * xbl = x + QK_K*ibl;
+        float sumx2 = 0;
+        for (int i = 0; i < QK_K; ++i) sumx2 += xbl[i]*xbl[i];
+        float sigma2 = sumx2/QK_K;
+
+        for (int ib = 0; ib < QK_K/16; ++ib) {
+            const float * xb = xbl + 16*ib;
+            const float * qw = quant_weights + QK_K*ibl + 16*ib;
+            for (int i = 0; i < 16; ++i) weight[i] = qw[i] * sqrtf(sigma2 + xb[i]*xb[i]);
+            for (int i = 0; i < 16; ++i) waux[i] = sqrtf(weight[i]);
+            for (int k = 0; k < 2; ++k) {
+                int nflip = 0;
+                uint8_t s = 0;
+                for (int i = 0; i < 8; ++i) {
+                    if (xb[8*k + i] >= 0) xval[8*k + i] = xb[8*k + i];
+                    else {
+                        xval[8*k + i] = -xb[8*k + i]; ++nflip; s |= (1 << i);
+                    }
+                }
+                if (nflip%2) {
+                    int imin = 0; float min = weight[8*k+imin]*xb[8*k+imin]*xb[8*k+imin];
+                    for (int i = 1; i < 8; ++i) {
+                        float ax = weight[8*k+i]*xb[8*k+i]*xb[8*k+i];
+                        if (ax < min) {
+                            min = ax; imin = i;
+                        }
+                    }
+                    xval[8*k+imin] = -xval[8*k+imin];
+                    s ^= (1 << imin);
+                }
+                block_signs[k] = s & 127;
+            }
+            float max = xval[0];
+            for (int i = 1; i < 16; ++i) max = MAX(max, xval[i]);
+            if (max < GROUP_MAX_EPS) {
+                scales[ib] = 0;
+                memset(L, 0, 16);
+                continue;
+            }
+            float best = 0;
+            float scale = max/(2*kMaxQ-1);
+            is_on_grid[0] = is_on_grid[1] = true;
+            for (int is = -9; is <= 9; ++is) {
+                float id = (2*kMaxQ-1+is*0.1f)/max;
+                float this_scale = 1/id;
+                for (int k = 0; k < 2; ++k) {
+                    for (int i = 0; i < 8; ++i) {
+                        int l = nearest_int(0.5f*(id*xval[8*k+i]-1));
+                        Laux[8*k+i] = MAX(0, MIN(kMaxQ-1, l));
+                    }
+                    uint16_t u = 0;
+                    for (int i = 0; i < 8; ++i) u |= (Laux[8*k+i] << 2*i);
+                    int grid_index = kmap_q2xs[u];
+                    is_on_grid_aux[k] = true;
+                    if (grid_index < 0) {
+                        is_on_grid_aux[k] = false;
+                        const uint16_t * neighbours = kneighbors_q2xs - kmap_q2xs[u] - 1;
+                        grid_index = iq2_find_best_neighbour(neighbours, kgrid_q2xs, xval + 8*k, waux + 8*k, this_scale, Laux + 8*k);
+                    }
+                }
+                float sumqx = 0, sumq2 = 0;
+                for (int i = 0; i < 16; ++i) {
+                    float w = weight[i];
+                    float q = 2*Laux[i] + 1;
+                    sumqx += w*xval[i]*q;
+                    sumq2 += w*q*q;
+                }
+                if (sumq2 > 0 && sumqx*sumqx > best*sumq2) {
+                    scale = sumqx/sumq2; best = scale*sumqx;
+                    for (int i = 0; i < 16; ++i) L[i] = Laux[i];
+                    for (int k = 0; k <  2; ++k) is_on_grid[k] = is_on_grid_aux[k];
+                }
+            }
+            int n_not_ongrid = 0;
+            for (int k = 0; k < 2; ++k) if (!is_on_grid[k]) ++n_not_ongrid;
+            if (n_not_ongrid > 0 && scale > 0) {
+                float id = 1/scale;
+                for (int k = 0; k < 2; ++k) {
+                    if (is_on_grid[k]) continue;
+                    uint16_t u = 0;
+                    for (int i = 0; i < 8; ++i) {
+                        int l = nearest_int(0.5f*(id*xval[8*k+i]-1));
+                        l = MAX(0, MIN(kMaxQ-1, l));
+                        u |= (l << 2*i);
+                        L[8*k + i] = l;
+                    }
+                    int grid_index = kmap_q2xs[u];
+                    if (grid_index < 0) {
+                        const uint16_t * neighbours = kneighbors_q2xs - kmap_q2xs[u] - 1;
+                        grid_index = iq2_find_best_neighbour(neighbours, kgrid_q2xs, xval + 8*k, waux + 8*k, scale, L + 8*k);
+                    }
+                }
+                float sumqx = 0, sumq2 = 0;
+                for (int i = 0; i < 16; ++i) {
+                    float w = weight[i];
+                    float q = 2*L[i] + 1;
+                    sumqx += w*xval[i]*q;
+                    sumq2 += w*q*q;
+                }
+                if (sumq2 > 0) scale = sumqx/sumq2;
+            }
+            if (scale < 0) {
+                scale = -scale;
+                for (int k = 0; k < 2; ++k) block_signs[k] = (~block_signs[k]) & 127;
+            }
+            for (int k = 0; k < 2; ++k) {
+                uint16_t u = 0;
+                for (int i = 0; i < 8; ++i) u |= (L[8*k+i] << 2*i);
+                int grid_index = kmap_q2xs[u];
+                if (grid_index < 0) {
+                    printf("Oops: found point %u not on grid:", u);
+                    for (int i = 0; i < 8; ++i) printf(" %d", L[8*k+i]);
+                    printf("\n");
+                    GGML_ABORT("fatal error");
+                }
+                q2[2*ib+k] = grid_index | (block_signs[k] << 9);
+            }
+            GGML_ASSERT(scale >= 0);
+            scales[ib] = scale;
+            max_scale = MAX(max_scale, scale);
+        }
+
+        if (!max_scale) {
+            memset(y[ibl].qs, 0, QK_K/4);
+            continue;
+        }
+
+        float d = max_scale/31;
+        y[ibl].d = GGML_FP32_TO_FP16(d);
+        float id = 1/d;
+        for (int ib = 0; ib < QK_K/16; ++ib) {
+            int l = nearest_int(0.5f*(id*scales[ib]-1));
+            l = MAX(0, MIN(15, l));
+            if (ib%2 == 0) y[ibl].scales[ib/2] = l;
+            else y[ibl].scales[ib/2] |= (l << 4);
+        }
+        memcpy(y[ibl].qs, q2, QK_K/4);
+
+    }
+}
+
+size_t quantize_iq2_xxs(const float * restrict src, void * restrict dst, int64_t nrow, int64_t n_per_row, const float * quant_weights) {
+    GGML_ASSERT(n_per_row%QK_K == 0);
+    int64_t nblock = n_per_row/QK_K;
+    char * qrow = (char *)dst;
+    for (int64_t row = 0; row < nrow; ++row) {
+        quantize_row_iq2_xxs_impl(src, qrow, n_per_row, quant_weights);
+        src += n_per_row;
+        qrow += nblock*sizeof(block_iq2_xxs);
+    }
+    return nrow * nblock * sizeof(block_iq2_xxs);
+}
+
+size_t quantize_iq2_xs(const float * restrict src, void * restrict dst, int64_t nrow, int64_t n_per_row, const float * quant_weights) {
+    GGML_ASSERT(n_per_row%QK_K == 0);
+    int64_t nblock = n_per_row/QK_K;
+    char * qrow = (char *)dst;
+    for (int64_t row = 0; row < nrow; ++row) {
+        quantize_row_iq2_xs_impl(src, qrow, n_per_row, quant_weights);
+        src += n_per_row;
+        qrow += nblock*sizeof(block_iq2_xs);
+    }
+    return nrow * nblock * sizeof(block_iq2_xs);
+}
+
+//
+// ============================================= 3-bit using D4 lattice
+//
+
+typedef struct {
+    uint32_t * grid;
+    int      * map;
+    uint16_t * neighbours;
+} iq3_entry_t;
+
+static iq3_entry_t iq3_data[2] = {
+    {NULL, NULL, NULL},
+    {NULL, NULL, NULL},
+};
+
+static inline int iq3_data_index(int grid_size) {
+    (void)grid_size;
+    GGML_ASSERT(grid_size == 256 || grid_size == 512);
+    return grid_size == 256 ? 0 : 1;
+}
+
+static int iq3_compare_func(const void * left, const void * right) {
+    const int * l = (const int *)left;
+    const int * r = (const int *)right;
+    return l[0] < r[0] ? -1 : l[0] > r[0] ? 1 : l[1] < r[1] ? -1 : l[1] > r[1] ? 1 : 0;
+}
+
+void iq3xs_init_impl(int grid_size) {
+    const int gindex = iq3_data_index(grid_size);
+    if (iq3_data[gindex].grid) {
+        return;
+    }
+    static const uint16_t kgrid_256[256] = {
+            0,     2,     4,     9,    11,    15,    16,    18,    25,    34,    59,    61,    65,    67,    72,    74,
+           81,    85,    88,    90,    97,   108,   120,   128,   130,   132,   137,   144,   146,   153,   155,   159,
+          169,   175,   189,   193,   199,   200,   202,   213,   248,   267,   287,   292,   303,   315,   317,   321,
+          327,   346,   362,   413,   436,   456,   460,   462,   483,   497,   513,   515,   520,   522,   529,   531,
+          536,   538,   540,   551,   552,   576,   578,   585,   592,   594,   641,   643,   648,   650,   657,   664,
+          698,   704,   706,   720,   729,   742,   758,   769,   773,   808,   848,   852,   870,   889,   901,   978,
+          992,  1024,  1026,  1033,  1035,  1040,  1042,  1046,  1049,  1058,  1089,  1091,  1093,  1096,  1098,  1105,
+         1112,  1139,  1143,  1144,  1152,  1154,  1161,  1167,  1168,  1170,  1183,  1184,  1197,  1217,  1224,  1228,
+         1272,  1276,  1309,  1323,  1347,  1367,  1377,  1404,  1473,  1475,  1486,  1509,  1537,  1544,  1546,  1553,
+         1555,  1576,  1589,  1594,  1600,  1602,  1616,  1625,  1636,  1638,  1665,  1667,  1672,  1685,  1706,  1722,
+         1737,  1755,  1816,  1831,  1850,  1856,  1862,  1874,  1901,  1932,  1950,  1971,  2011,  2032,  2052,  2063,
+         2077,  2079,  2091,  2095,  2172,  2192,  2207,  2208,  2224,  2230,  2247,  2277,  2308,  2345,  2356,  2389,
+         2403,  2424,  2501,  2504,  2506,  2520,  2570,  2593,  2616,  2624,  2630,  2646,  2669,  2700,  2714,  2746,
+         2754,  2795,  2824,  2835,  2839,  2874,  2882,  2905,  2984,  3028,  3042,  3092,  3108,  3110,  3124,  3153,
+         3185,  3215,  3252,  3288,  3294,  3364,  3397,  3434,  3483,  3523,  3537,  3587,  3589,  3591,  3592,  3610,
+         3626,  3670,  3680,  3722,  3749,  3754,  3776,  3789,  3803,  3824,  3857,  3873,  3904,  3906,  3924,  3992,
+    };
+    static const uint16_t kgrid_512[512] = {
+            0,     1,     2,     5,     7,     8,     9,    10,    12,    14,    16,    17,    21,    27,    32,    34,
+           37,    39,    41,    43,    48,    50,    57,    60,    63,    64,    65,    66,    68,    72,    73,    77,
+           80,    83,    87,    89,    93,   100,   113,   117,   122,   128,   129,   133,   135,   136,   139,   142,
+          145,   149,   152,   156,   162,   165,   167,   169,   171,   184,   187,   195,   201,   205,   208,   210,
+          217,   219,   222,   228,   232,   234,   247,   249,   253,   256,   267,   271,   273,   276,   282,   288,
+          291,   297,   312,   322,   324,   336,   338,   342,   347,   353,   357,   359,   374,   379,   390,   393,
+          395,   409,   426,   441,   448,   450,   452,   464,   466,   470,   475,   488,   492,   512,   513,   514,
+          516,   520,   521,   523,   525,   527,   528,   530,   537,   540,   542,   556,   558,   561,   570,   576,
+          577,   579,   582,   584,   588,   593,   600,   603,   609,   616,   618,   632,   638,   640,   650,   653,
+          655,   656,   660,   666,   672,   675,   685,   688,   698,   705,   708,   711,   712,   715,   721,   727,
+          728,   732,   737,   754,   760,   771,   773,   778,   780,   793,   795,   802,   806,   808,   812,   833,
+          840,   843,   849,   856,   858,   873,   912,   916,   919,   932,   934,   961,   963,   968,   970,   977,
+          989,   993,  1010,  1016,  1024,  1025,  1027,  1029,  1031,  1032,  1034,  1036,  1038,  1041,  1043,  1047,
+         1048,  1050,  1057,  1059,  1061,  1064,  1066,  1079,  1080,  1083,  1085,  1088,  1090,  1096,  1099,  1103,
+         1106,  1109,  1113,  1116,  1122,  1129,  1153,  1156,  1159,  1169,  1171,  1176,  1183,  1185,  1195,  1199,
+         1209,  1212,  1216,  1218,  1221,  1225,  1234,  1236,  1241,  1243,  1250,  1256,  1270,  1281,  1287,  1296,
+         1299,  1306,  1309,  1313,  1338,  1341,  1348,  1353,  1362,  1375,  1376,  1387,  1400,  1408,  1410,  1415,
+         1425,  1453,  1457,  1477,  1481,  1494,  1496,  1507,  1512,  1538,  1545,  1547,  1549,  1551,  1554,  1561,
+         1563,  1565,  1570,  1572,  1575,  1577,  1587,  1593,  1601,  1603,  1605,  1612,  1617,  1619,  1632,  1648,
+         1658,  1662,  1664,  1674,  1680,  1690,  1692,  1704,  1729,  1736,  1740,  1745,  1747,  1751,  1752,  1761,
+         1763,  1767,  1773,  1787,  1795,  1801,  1806,  1810,  1817,  1834,  1840,  1844,  1857,  1864,  1866,  1877,
+         1882,  1892,  1902,  1915,  1934,  1953,  1985,  1987,  2000,  2002,  2013,  2048,  2052,  2058,  2064,  2068,
+         2071,  2074,  2081,  2088,  2104,  2114,  2119,  2121,  2123,  2130,  2136,  2141,  2147,  2153,  2157,  2177,
+         2179,  2184,  2189,  2193,  2203,  2208,  2223,  2226,  2232,  2244,  2249,  2251,  2256,  2258,  2265,  2269,
+         2304,  2306,  2324,  2335,  2336,  2361,  2373,  2375,  2385,  2418,  2443,  2460,  2480,  2504,  2509,  2520,
+         2531,  2537,  2562,  2568,  2572,  2578,  2592,  2596,  2599,  2602,  2614,  2620,  2625,  2627,  2629,  2634,
+         2641,  2650,  2682,  2688,  2697,  2707,  2712,  2718,  2731,  2754,  2759,  2760,  2775,  2788,  2793,  2805,
+         2811,  2817,  2820,  2832,  2842,  2854,  2890,  2902,  2921,  2923,  2978,  3010,  3012,  3026,  3081,  3083,
+         3085,  3097,  3099,  3120,  3136,  3152,  3159,  3188,  3210,  3228,  3234,  3245,  3250,  3256,  3264,  3276,
+         3281,  3296,  3349,  3363,  3378,  3392,  3395,  3420,  3440,  3461,  3488,  3529,  3531,  3584,  3588,  3591,
+         3600,  3602,  3614,  3616,  3628,  3634,  3650,  3657,  3668,  3683,  3685,  3713,  3716,  3720,  3726,  3729,
+         3736,  3753,  3778,  3802,  3805,  3819,  3841,  3845,  3851,  3856,  3880,  3922,  3938,  3970,  3993,  4032,
+    };
+
+    const int kmap_size = 4096;
+    const int nwant = grid_size == 256 ? 2 : 3;
+    const uint16_t * kgrid = grid_size == 256 ? kgrid_256 : kgrid_512;
+    uint32_t * kgrid_q3xs;
+    int      * kmap_q3xs;
+    uint16_t * kneighbors_q3xs;
+
+    //printf("================================================================= %s(grid_size = %d)\n", __func__, grid_size);
+    uint32_t * the_grid = (uint32_t *)malloc(grid_size*sizeof(uint32_t));
+    for (int k = 0; k < grid_size; ++k) {
+        int8_t * pos = (int8_t *)(the_grid + k);
+        for (int i = 0; i < 4; ++i) {
+            int l = (kgrid[k] >> 3*i) & 0x7;
+            pos[i] = 2*l + 1;
+        }
+    }
+    kgrid_q3xs = the_grid;
+    iq3_data[gindex].grid = the_grid;
+    kmap_q3xs = (int *)malloc(kmap_size*sizeof(int));
+    iq3_data[gindex].map = kmap_q3xs;
+    for (int i = 0; i < kmap_size; ++i) kmap_q3xs[i] = -1;
+    uint32_t aux32;
+    uint8_t * aux8 = (uint8_t *)&aux32;
+    for (int i = 0; i < grid_size; ++i) {
+        aux32 = kgrid_q3xs[i];
+        uint16_t index = 0;
+        for (int k=0; k<4; ++k) {
+            uint16_t q = (aux8[k] - 1)/2;
+            index |= (q << 3*k);
+        }
+        kmap_q3xs[index] = i;
+    }
+    int8_t pos[4];
+    int * dist2 = (int *)malloc(2*grid_size*sizeof(int));
+    int num_neighbors = 0, num_not_in_map = 0;
+    for (int i = 0; i < kmap_size; ++i) {
+        if (kmap_q3xs[i] >= 0) continue;
+        ++num_not_in_map;
+        for (int k = 0; k < 4; ++k) {
+            int l = (i >> 3*k) & 0x7;
+            pos[k] = 2*l + 1;
+        }
+        for (int j = 0; j < grid_size; ++j) {
+            const int8_t * pg = (const int8_t *)(kgrid_q3xs + j);
+            int d2 = 0;
+            for (int k = 0; k < 4; ++k) d2 += (pg[k] - pos[k])*(pg[k] - pos[k]);
+            dist2[2*j+0] = d2;
+            dist2[2*j+1] = j;
+        }
+        qsort(dist2, grid_size, 2*sizeof(int), iq3_compare_func);
+        int n = 0; int d2 = dist2[0];
+        int nhave = 1;
+        for (int j = 0; j < grid_size; ++j) {
+            if (dist2[2*j] > d2) {
+                if (nhave == nwant) break;
+                d2 = dist2[2*j];
+                ++nhave;
+            }
+            ++n;
+        }
+        num_neighbors += n;
+    }
+    //printf("%s: %d neighbours in total\n", __func__, num_neighbors);
+    kneighbors_q3xs = (uint16_t *)malloc((num_neighbors + num_not_in_map)*sizeof(uint16_t));
+    iq3_data[gindex].neighbours = kneighbors_q3xs;
+    int counter = 0;
+    for (int i = 0; i < kmap_size; ++i) {
+        if (kmap_q3xs[i] >= 0) continue;
+        for (int k = 0; k < 4; ++k) {
+            int l = (i >> 3*k) & 0x7;
+            pos[k] = 2*l + 1;
+        }
+        for (int j = 0; j < grid_size; ++j) {
+            const int8_t * pg = (const int8_t *)(kgrid_q3xs + j);
+            int d2 = 0;
+            for (int k = 0; k < 4; ++k) d2 += (pg[k] - pos[k])*(pg[k] - pos[k]);
+            dist2[2*j+0] = d2;
+            dist2[2*j+1] = j;
+        }
+        qsort(dist2, grid_size, 2*sizeof(int), iq3_compare_func);
+        kmap_q3xs[i] = -(counter + 1);
+        int d2 = dist2[0];
+        uint16_t * start = &kneighbors_q3xs[counter++];
+        int n = 0, nhave = 1;
+        for (int j = 0; j < grid_size; ++j) {
+            if (dist2[2*j] > d2) {
+                if (nhave == nwant) break;
+                d2 = dist2[2*j];
+                ++nhave;
+            }
+            kneighbors_q3xs[counter++] = dist2[2*j+1];
+            ++n;
+        }
+        *start = n;
+    }
+    free(dist2);
+}
+
+void iq3xs_free_impl(int grid_size) {
+    GGML_ASSERT(grid_size == 256 || grid_size == 512);
+    const int gindex = iq3_data_index(grid_size);
+    if (iq3_data[gindex].grid) {
+        free(iq3_data[gindex].grid);       iq3_data[gindex].grid = NULL;
+        free(iq3_data[gindex].map);        iq3_data[gindex].map  = NULL;
+        free(iq3_data[gindex].neighbours); iq3_data[gindex].neighbours = NULL;
+    }
+}
+
+static int iq3_find_best_neighbour(const uint16_t * restrict neighbours, const uint32_t * restrict grid,
+        const float * restrict xval, const float * restrict weight, float scale, int8_t * restrict L) {
+    int num_neighbors = neighbours[0];
+    GGML_ASSERT(num_neighbors > 0);
+    float best_d2 = FLT_MAX;
+    int grid_index = -1;
+    for (int j = 1; j <= num_neighbors; ++j) {
+        const int8_t * pg = (const int8_t *)(grid + neighbours[j]);
+        float d2 = 0;
+        for (int i = 0; i < 4; ++i) {
+            float q = pg[i];
+            float diff = scale*q - xval[i];
+            d2 += weight[i]*diff*diff;
+        }
+        if (d2 < best_d2) {
+            best_d2 = d2; grid_index = neighbours[j];
+        }
+    }
+    GGML_ASSERT(grid_index >= 0);
+    const int8_t * pg = (const int8_t *)(grid + grid_index);
+    for (int i = 0; i < 4; ++i) L[i] = (pg[i] - 1)/2;
+    return grid_index;
+}
+
+static void quantize_row_iq3_xxs_impl(int grid_size, const float * restrict x, void * restrict vy, int64_t n,
+        const float * restrict quant_weights) {
+
+    const int gindex = iq3_data_index(grid_size);
+
+    const uint32_t * kgrid_q3xs      = iq3_data[gindex].grid;
+    const int      * kmap_q3xs       = iq3_data[gindex].map;
+    const uint16_t * kneighbors_q3xs = iq3_data[gindex].neighbours;
+
+    //GGML_ASSERT(quant_weights   && "missing quantization weights");
+    GGML_ASSERT(kgrid_q3xs      && "forgot to call ggml_quantize_init()?");
+    GGML_ASSERT(kmap_q3xs       && "forgot to call ggml_quantize_init()?");
+    GGML_ASSERT(kneighbors_q3xs && "forgot to call ggml_quantize_init()?");
+    GGML_ASSERT(n%QK_K == 0);
+
+    const int kMaxQ = 8;
+
+    const int64_t nbl = n/QK_K;
+
+    ggml_fp16_t * dh;
+    uint8_t * qs;
+    int block_size;
+    if (grid_size == 256) {
+        block_iq3_xxs * y = vy;
+        dh = &y->d;
+        qs = y->qs;
+        block_size = sizeof(block_iq3_xxs);
+    } else {
+        block_iq3_s * y = vy;
+        dh = &y->d;
+        qs = y->qs;
+        block_size = sizeof(block_iq3_s);
+    }
+    int quant_size = block_size - sizeof(ggml_fp16_t);
+
+    float scales[QK_K/32];
+    float weight[32];
+    float xval[32];
+    int8_t L[32];
+    int8_t Laux[32];
+    float  waux[32];
+    bool   is_on_grid[8];
+    bool   is_on_grid_aux[8];
+    uint8_t block_signs[8];
+    uint8_t q3[3*(QK_K/8)+QK_K/32];
+    uint32_t * scales_and_signs = (uint32_t *)(q3 + QK_K/4);
+    uint8_t  * qh = q3 + 3*(QK_K/8);
+
+    for (int ibl = 0; ibl < nbl; ++ibl) {
+
+        dh[0] = GGML_FP32_TO_FP16(0.f);
+        memset(q3, 0, 3*QK_K/8+QK_K/32);
+
+        float max_scale = 0;
+
+        const float * xbl = x + QK_K*ibl;
+        float sumx2 = 0;
+        for (int i = 0; i < QK_K; ++i) sumx2 += xbl[i]*xbl[i];
+        float sigma2 = 2*sumx2/QK_K;
+
+        for (int ib = 0; ib < QK_K/32; ++ib) {
+            const float * xb = xbl + 32*ib;
+            if (quant_weights) {
+                const float * qw = quant_weights + QK_K*ibl + 32*ib;
+                for (int i = 0; i < 32; ++i) weight[i] = qw[i] * sqrtf(sigma2 + xb[i]*xb[i]);
+            } else {
+                for (int i = 0; i < 32; ++i) weight[i] = xb[i]*xb[i];
+            }
+            for (int i = 0; i < 32; ++i) waux[i] = sqrtf(weight[i]);
+            for (int k = 0; k < 4; ++k) {
+                int nflip = 0;
+                uint8_t s = 0;
+                for (int i = 0; i < 8; ++i) {
+                    if (xb[8*k + i] >= 0) xval[8*k + i] = xb[8*k + i];
+                    else {
+                        xval[8*k + i] = -xb[8*k + i]; ++nflip; s |= (1 << i);
+                    }
+                }
+                if (nflip%2) {
+                    int imin = 0; float min = weight[8*k+imin]*xb[8*k+imin]*xb[8*k+imin];
+                    for (int i = 1; i < 8; ++i) {
+                        float ax = weight[8*k+i]*xb[8*k+i]*xb[8*k+i];
+                        if (ax < min) {
+                            min = ax; imin = i;
+                        }
+                    }
+                    xval[8*k+imin] = -xval[8*k+imin];
+                    s ^= (1 << imin);
+                }
+                block_signs[k] = s & 127;
+            }
+            float max = xval[0];
+            for (int i = 1; i < 32; ++i) max = MAX(max, xval[i]);
+            if (max < GROUP_MAX_EPS_IQ3_XXS) {
+                scales[ib] = 0;
+                memset(L, 0, 32);
+                continue;
+            }
+            float best = 0;
+            float scale = max/(2*kMaxQ-1);
+            for (int is = -15; is <= 15; ++is) {
+                float id = (2*kMaxQ-1+is*0.2f)/max;
+                float this_scale = 1/id;
+                for (int k = 0; k < 8; ++k) {
+                    for (int i = 0; i < 4; ++i) {
+                        int l = nearest_int(0.5f*(id*xval[4*k+i]-1));
+                        Laux[4*k+i] = MAX(0, MIN(kMaxQ-1, l));
+                    }
+                    uint16_t u = 0;
+                    for (int i = 0; i < 4; ++i) u |= (Laux[4*k+i] << 3*i);
+                    int grid_index = kmap_q3xs[u];
+                    is_on_grid_aux[k] = true;
+                    if (grid_index < 0) {
+                        is_on_grid_aux[k] = false;
+                        const uint16_t * neighbours = kneighbors_q3xs - kmap_q3xs[u] - 1;
+                        grid_index = iq3_find_best_neighbour(neighbours, kgrid_q3xs, xval + 4*k, waux + 4*k, this_scale, Laux + 4*k);
+                    }
+                }
+                float sumqx = 0, sumq2 = 0;
+                for (int i = 0; i < 32; ++i) {
+                    float w = weight[i];
+                    float q = 2*Laux[i] + 1;
+                    sumqx += w*xval[i]*q;
+                    sumq2 += w*q*q;
+                }
+                if (sumq2 > 0 && sumqx*sumqx > best*sumq2) {
+                    scale = sumqx/sumq2; best = scale*sumqx;
+                    for (int i = 0; i < 32; ++i) L[i] = Laux[i];
+                    for (int k = 0; k <  8; ++k) is_on_grid[k] = is_on_grid_aux[k];
+                }
+            }
+            int n_not_ongrid = 0;
+            for (int k = 0; k < 8; ++k) if (!is_on_grid[k]) ++n_not_ongrid;
+            if (n_not_ongrid > 0 && scale > 0) {
+                float id = 1/scale;
+                for (int k = 0; k < 8; ++k) {
+                    if (is_on_grid[k]) continue;
+                    uint16_t u = 0;
+                    for (int i = 0; i < 4; ++i) {
+                        int l = nearest_int(0.5f*(id*xval[4*k+i]-1));
+                        l = MAX(0, MIN(kMaxQ-1, l));
+                        u |= (l << 3*i);
+                    }
+                    int grid_index = kmap_q3xs[u];
+                    if (grid_index < 0) {
+                        const uint16_t * neighbours = kneighbors_q3xs - kmap_q3xs[u] - 1;
+                        grid_index = iq3_find_best_neighbour(neighbours, kgrid_q3xs, xval + 4*k, waux + 4*k, scale, L + 4*k);
+                    }
+                    const int8_t * pg = (const int8_t *)(kgrid_q3xs + grid_index);
+                    for (int i = 0; i < 4; ++i) L[4*k+i] = (pg[i] - 1)/2;
+                }
+                float sumqx = 0, sumq2 = 0;
+                for (int i = 0; i < 32; ++i) {
+                    float w = weight[i];
+                    float q = 2*L[i] + 1;
+                    sumqx += w*xval[i]*q;
+                    sumq2 += w*q*q;
+                }
+                if (sumq2 > 0) scale = sumqx/sumq2;
+            }
+            if (scale < 0) {
+                // This should never happen, but just in case, flip scale so that it is positive (we use uint's to encode the scale)
+                // and correspondingly flip quant signs.
+                scale = -scale;
+                for (int k = 0; k < 4; ++k) block_signs[k] = (~block_signs[k]) & 127;
+            }
+            for (int k = 0; k < 8; ++k) {
+                uint16_t u = 0;
+                for (int i = 0; i < 4; ++i) u |= (L[4*k+i] << 3*i);
+                int grid_index = kmap_q3xs[u];
+                if (grid_index < 0) {
+                    printf("Oops: found point %u not on grid:", u);
+                    for (int i = 0; i < 4; ++i) printf(" %d", L[4*k+i]);
+                    printf("\n");
+                    GGML_ABORT("fatal error");
+                }
+                if (grid_size == 256) {
+                    q3[8*ib+k] = grid_index;
+                } else {
+                    q3[8*ib+k] = grid_index & 255;
+                    qh[ib] |= ((grid_index >> 8) << k);
+                }
+
+            }
+            scales_and_signs[ib] = block_signs[0] | (block_signs[1] << 7) | (block_signs[2] << 14) | (block_signs[3] << 21);
+            GGML_ASSERT(scale >= 0);
+            scales[ib] = scale;
+            max_scale = MAX(max_scale, scale);
+        }
+
+        if (!max_scale) {
+            memset(qs, 0, quant_size);
+            dh += block_size/sizeof(ggml_fp16_t);
+            qs += block_size;
+            continue;
+        }
+
+        float d = max_scale/31;
+        dh[0] = GGML_FP32_TO_FP16(d * 1.0125f);  // small improvement via this fudge factor
+        float id = 1/d;
+        for (int ib = 0; ib < QK_K/32; ++ib) {
+            int l = nearest_int(0.5f*(id*scales[ib]-1));
+            l = MAX(0, MIN(15, l));
+            scales_and_signs[ib] |= ((uint32_t)l << 28);
+        }
+        memcpy(qs, q3, quant_size);
+
+        dh += block_size/sizeof(ggml_fp16_t);
+        qs += block_size;
+
+    }
+}
+
+size_t quantize_iq3_xxs(const float * restrict src, void * restrict dst, int64_t nrow, int64_t n_per_row, const float * quant_weights) {
+    GGML_ASSERT(n_per_row%QK_K == 0);
+    int64_t nblock = n_per_row/QK_K;
+    char * qrow = (char *)dst;
+    for (int64_t row = 0; row < nrow; ++row) {
+        quantize_row_iq3_xxs_impl(256, src, qrow, n_per_row, quant_weights);
+        src += n_per_row;
+        qrow += nblock*sizeof(block_iq3_xxs);
+    }
+    return nrow * nblock * sizeof(block_iq3_xxs);
+}
+
+void quantize_row_iq3_xxs(const float * restrict x, void * restrict vy, int64_t k) {
+    assert(k % QK_K == 0);
+    block_iq3_xxs * restrict y = vy;
+    quantize_row_iq3_xxs_ref(x, y, k);
+}
+
+void quantize_row_iq3_xxs_ref(const float * restrict x, block_iq3_xxs * restrict y, int64_t k) {
+    assert(k % QK_K == 0);
+    quantize_row_iq3_xxs_impl(256, x, y, k, NULL);
+}
+
+static void quantize_row_iq3_s_impl(int block_size, const float * restrict x, void * restrict vy, int n,
+        const float * restrict quant_weights,
+        float   * scales,
+        float   * weight,
+        float   * xval,
+        int8_t  * L,
+        int8_t  * Laux,
+        float   * waux,
+        bool    * is_on_grid,
+        bool    * is_on_grid_aux,
+        uint8_t * block_signs) {
+
+    const int gindex = iq3_data_index(512);
+
+    const uint32_t * kgrid_q3xs      = iq3_data[gindex].grid;
+    const int      * kmap_q3xs       = iq3_data[gindex].map;
+    const uint16_t * kneighbors_q3xs = iq3_data[gindex].neighbours;
+
+    //GGML_ASSERT(quant_weights   && "missing quantization weights");
+    GGML_ASSERT(kgrid_q3xs      && "forgot to call ggml_quantize_init()?");
+    GGML_ASSERT(kmap_q3xs       && "forgot to call ggml_quantize_init()?");
+    GGML_ASSERT(kneighbors_q3xs && "forgot to call ggml_quantize_init()?");
+    GGML_ASSERT(n%QK_K == 0);
+
+    const int kMaxQ = 8;
+
+    const int64_t nbl = n/QK_K;
+
+    block_iq3_s * y = vy;
+
+    const int bs4 = block_size/4;
+    const int bs8 = block_size/8;
+
+    for (int ibl = 0; ibl < nbl; ++ibl) {
+
+        memset(&y[ibl], 0, sizeof(block_iq3_s));
+        y[ibl].d = GGML_FP32_TO_FP16(0.f);
+
+        uint8_t * qs = y[ibl].qs;
+        uint8_t * qh = y[ibl].qh;
+        uint8_t * signs = y[ibl].signs;
+
+        float max_scale = 0;
+
+        const float * xbl = x + QK_K*ibl;
+        float sumx2 = 0;
+        for (int i = 0; i < QK_K; ++i) sumx2 += xbl[i]*xbl[i];
+        float sigma2 = 2*sumx2/QK_K;
+
+        for (int ib = 0; ib < QK_K/block_size; ++ib) {
+            const float * xb = xbl + block_size*ib;
+            if (quant_weights) {
+                const float * qw = quant_weights + QK_K*ibl + block_size*ib;
+                for (int i = 0; i < block_size; ++i) weight[i] = qw[i] * sqrtf(sigma2 + xb[i]*xb[i]);
+            } else {
+                for (int i = 0; i < block_size; ++i) weight[i] = xb[i]*xb[i];
+            }
+            for (int i = 0; i < block_size; ++i) waux[i] = sqrtf(weight[i]);
+            for (int k = 0; k < bs8; ++k) {
+                uint8_t s = 0;
+                for (int i = 0; i < 8; ++i) {
+                    if (xb[8*k + i] >= 0) xval[8*k + i] = xb[8*k + i];
+                    else {
+                        xval[8*k + i] = -xb[8*k + i]; s |= (1 << i);
+                    }
+                }
+                block_signs[k] = s;
+            }
+            float max = xval[0];
+            for (int i = 1; i < block_size; ++i) max = MAX(max, xval[i]);
+            if (!max) {
+                scales[ib] = 0;
+                continue;
+            }
+            float best = 0;
+            float scale = max/(2*kMaxQ-1);
+            for (int k = 0; k < bs4; ++k) is_on_grid[k] = false;
+            for (int is = -9; is <= 9; ++is) {
+                float id = (2*kMaxQ-1+is*0.2f)/max;
+                float this_scale = 1/id;
+                for (int k = 0; k < bs4; ++k) {
+                    for (int i = 0; i < 4; ++i) {
+                        int l = nearest_int(0.5f*(id*xval[4*k+i]-1));
+                        Laux[4*k+i] = MAX(0, MIN(kMaxQ-1, l));
+                    }
+                    uint16_t u = 0;
+                    for (int i = 0; i < 4; ++i) u |= (Laux[4*k+i] << 3*i);
+                    int grid_index = kmap_q3xs[u];
+                    is_on_grid_aux[k] = true;
+                    if (grid_index < 0) {
+                        is_on_grid_aux[k] = false;
+                        const uint16_t * neighbours = kneighbors_q3xs - kmap_q3xs[u] - 1;
+                        grid_index = iq3_find_best_neighbour(neighbours, kgrid_q3xs, xval + 4*k, waux + 4*k, this_scale, Laux + 4*k);
+                    }
+                }
+                float sumqx = 0, sumq2 = 0;
+                for (int i = 0; i < block_size; ++i) {
+                    float w = weight[i];
+                    float q = 2*Laux[i] + 1;
+                    sumqx += w*xval[i]*q;
+                    sumq2 += w*q*q;
+                }
+                if (sumq2 > 0 && sumqx*sumqx > best*sumq2) {
+                    scale = sumqx/sumq2; best = scale*sumqx;
+                    for (int i = 0; i < block_size; ++i) L[i] = Laux[i];
+                    for (int k = 0; k < bs4; ++k) is_on_grid[k] = is_on_grid_aux[k];
+                }
+            }
+            int n_not_ongrid = 0;
+            for (int k = 0; k < bs4; ++k) if (!is_on_grid[k]) ++n_not_ongrid;
+            if (n_not_ongrid > 0 && scale > 0) {
+                float id = 1/scale;
+                for (int k = 0; k < bs4; ++k) {
+                    //if (is_on_grid[k]) continue;
+                    uint16_t u = 0;
+                    for (int i = 0; i < 4; ++i) {
+                        int l = nearest_int(0.5f*(id*xval[4*k+i]-1));
+                        l = MAX(0, MIN(kMaxQ-1, l));
+                        u |= (l << 3*i);
+                    }
+                    int grid_index = kmap_q3xs[u];
+                    if (grid_index < 0) {
+                        const uint16_t * neighbours = kneighbors_q3xs - kmap_q3xs[u] - 1;
+                        grid_index = iq3_find_best_neighbour(neighbours, kgrid_q3xs, xval + 4*k, waux + 4*k, scale, L + 4*k);
+                    }
+                    const int8_t * pg = (const int8_t *)(kgrid_q3xs + grid_index);
+                    for (int i = 0; i < 4; ++i) L[4*k+i] = (pg[i] - 1)/2;
+                }
+                float sumqx = 0, sumq2 = 0;
+                for (int i = 0; i < block_size; ++i) {
+                    float w = weight[i];
+                    float q = 2*L[i] + 1;
+                    sumqx += w*xval[i]*q;
+                    sumq2 += w*q*q;
+                }
+                if (sumq2 > 0) scale = sumqx/sumq2;
+            }
+            if (scale < 0) {
+                // This should never happen, but just in case, flip scale so that it is positive (we use uint's to encode the scale)
+                // and correspondingly flip quant signs.
+                scale = -scale;
+                for (int k = 0; k < bs8; ++k) block_signs[k] = ~block_signs[k];
+            }
+            for (int k = 0; k < bs4; ++k) {
+                uint16_t u = 0;
+                for (int i = 0; i < 4; ++i) u |= (L[4*k+i] << 3*i);
+                int grid_index = kmap_q3xs[u];
+                if (grid_index < 0) {
+                    printf("Oops: found point %u not on grid:", u);
+                    for (int i = 0; i < 4; ++i) printf(" %d", L[4*k+i]);
+                    printf("\n");
+                    GGML_ABORT("fatal error");
+                }
+                qs[k] = grid_index & 255;
+                qh[(ib*bs4+k)/8] |= ((grid_index >> 8) << ((ib*bs4+k)%8));
+            }
+            qs += bs4;
+            for (int k = 0; k < bs8; ++k) signs[k] = block_signs[k];
+            signs += bs8;
+            GGML_ASSERT(scale >= 0);
+            scales[ib] = scale;
+            max_scale = MAX(max_scale, scale);
+        }
+
+        if (!max_scale) {
+            continue;
+        }
+
+        float d = max_scale/31;
+        y[ibl].d = GGML_FP32_TO_FP16(d * 1.033f);
+        float id = 1/d;
+        for (int ib = 0; ib < QK_K/block_size; ib += 2) {
+            int l1 = nearest_int(0.5f*(id*scales[ib+0]-1));
+            l1 = MAX(0, MIN(15, l1));
+            int l2 = nearest_int(0.5f*(id*scales[ib+1]-1));
+            l2 = MAX(0, MIN(15, l2));
+            y[ibl].scales[ib/2] = l1 | (l2 << 4);
+        }
+
+    }
+}
+
+#define IQ3S_BLOCK_SIZE 32
+size_t quantize_iq3_s(const float * restrict src, void * restrict dst, int64_t nrow, int64_t n_per_row, const float * quant_weights) {
+    GGML_ASSERT(n_per_row%QK_K == 0);
+    int64_t nblock = n_per_row/QK_K;
+    float scales[QK_K/IQ3S_BLOCK_SIZE];
+    float weight[IQ3S_BLOCK_SIZE];
+    float xval[IQ3S_BLOCK_SIZE];
+    int8_t L[IQ3S_BLOCK_SIZE];
+    int8_t Laux[IQ3S_BLOCK_SIZE];
+    float  waux[IQ3S_BLOCK_SIZE];
+    bool   is_on_grid[IQ3S_BLOCK_SIZE/4];
+    bool   is_on_grid_aux[IQ3S_BLOCK_SIZE/4];
+    uint8_t block_signs[IQ3S_BLOCK_SIZE/8];
+    char * qrow = (char *)dst;
+    for (int64_t row = 0; row < nrow; ++row) {
+        quantize_row_iq3_s_impl(IQ3S_BLOCK_SIZE, src, qrow, n_per_row, quant_weights,
+                scales, weight, xval, L, Laux, waux, is_on_grid, is_on_grid_aux, block_signs);
+        src += n_per_row;
+        qrow += nblock*sizeof(block_iq3_s);
+    }
+    return nrow * nblock * sizeof(block_iq3_s);
+}
+
+void quantize_row_iq3_s(const float * restrict x, void * restrict vy, int64_t k) {
+    assert(k % QK_K == 0);
+    block_iq3_s * restrict y = vy;
+    quantize_row_iq3_s_ref(x, y, k);
+}
+
+void quantize_row_iq3_s_ref(const float * restrict x, block_iq3_s * restrict y, int64_t k) {
+    assert(k % QK_K == 0);
+    quantize_iq3_s(x, y, 1, k, NULL);
+}
+
+
+// =================================== 1.5 bpw ===================================================
+
+static int iq1_find_best_neighbour(const uint16_t * restrict neighbours, const uint64_t * restrict grid,
+        const float * restrict xval, const float * restrict weight, float * scale, int8_t * restrict L, int ngrid) {
+    int num_neighbors = neighbours[0];
+    GGML_ASSERT(num_neighbors > 0);
+    float best_score = -FLT_MAX;
+    int grid_index = -1;
+    for (int j = 1; j <= num_neighbors; ++j) {
+        const int8_t * pg = (const int8_t *)(grid + neighbours[j]);
+        float sumqx = 0, sumq2 = 0;
+        for (int i = 0; i < 8; ++i) {
+            float q = (pg[i] - 3)/2;
+            float w = weight[i];
+            sumqx += w*q*xval[i];
+            sumq2 += w*q*q;
+        }
+        if (sumqx > 0 && sumq2 > 0 && sumqx*sumqx > best_score*sumq2) {
+            *scale = sumqx/sumq2; best_score = *scale * sumqx;
+            grid_index = neighbours[j];
+        }
+    }
+    if (grid_index < 0) {
+        for (int i = 0; i < ngrid; ++i) {
+            const int8_t * grid_i = (const int8_t *)(grid + i);
+            float sumqx = 0, sumq2 = 0;
+            for (int j = 0; j < 8; ++j) {
+                float w = weight[j];
+                float q = (grid_i[j] - 3)/2;
+                sumqx += w*q*xval[j];
+                sumq2 += w*q*q;
+            }
+            if (sumqx > 0 && sumq2 > 0 && sumqx*sumqx > best_score*sumq2) {
+                *scale = sumqx/sumq2; best_score = *scale*sumqx;
+                grid_index = i;
+            }
+        }
+    }
+    if (grid_index < 0) {
+        printf("Oops, did not find grid point\n");
+        printf("Have %d neighbours\n", num_neighbors);
+        for (int j = 1; j <= num_neighbors; ++j) {
+            const int8_t * pg = (const int8_t *)(grid + neighbours[j]);
+            float sumqx = 0, sumq2 = 0;
+            for (int i = 0; i < 8; ++i) {
+                float q = (pg[i] - 3)/2;
+                float w = weight[i];
+                sumqx += w*q*xval[i];
+                sumq2 += w*q*q;
+            }
+            printf("    neighbour %d: sumqx = %g sumq2 = %g\n", j, (double)sumqx, (double)sumq2);
+        }
+    }
+    GGML_ASSERT(grid_index >= 0);
+    //!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!
+    *scale *= 1.05f;  // This is a fudge factor. Don't ask me why it improves the result.
+    //!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!
+    const int8_t * pg = (const int8_t *)(grid + grid_index);
+    for (int i = 0; i < 8; ++i) L[i] = (pg[i] - 1)/2;
+    return grid_index;
+}
+
+static int iq1_find_best_neighbour2(const uint16_t * restrict neighbours, const uint64_t * restrict grid,
+        const float * restrict xval, const float * restrict weight, float scale, const float * restrict xg, int8_t * restrict L, int ngrid) {
+    int num_neighbors = neighbours[0];
+    GGML_ASSERT(num_neighbors > 0);
+    float best_score = FLT_MAX;
+    int grid_index = -1;
+    for (int j = 1; j <= num_neighbors; ++j) {
+        const int8_t * pg = (const int8_t *)(grid + neighbours[j]);
+        float d2 = 0;
+        for (int i = 0; i < 8; ++i) {
+            float q = xg[(pg[i] - 1)/2];
+            float w = weight[i];
+            float diff = scale*q - xval[i];
+            d2 += w*diff*diff;
+        }
+        if (d2 < best_score) {
+            best_score = d2;
+            grid_index = neighbours[j];
+        }
+    }
+    if (grid_index < 0) {
+        for (int i = 0; i < ngrid; ++i) {
+            const int8_t * grid_i = (const int8_t *)(grid + i);
+            float d2 = 0;
+            for (int j = 0; j < 8; ++j) {
+                float w = weight[j];
+                float q = xg[(grid_i[j] - 1)/2];
+                float diff = scale*q - xval[i];
+                d2 += w*diff*diff;
+            }
+            if (d2 < best_score) {
+                best_score = d2;
+                grid_index = i;
+            }
+        }
+    }
+    if (grid_index < 0) {
+        printf("Oops, did not find grid point\n");
+        printf("Have %d neighbours\n", num_neighbors);
+        for (int j = 1; j <= num_neighbors; ++j) {
+            const int8_t * pg = (const int8_t *)(grid + neighbours[j]);
+            float sumqx = 0, sumq2 = 0;
+            for (int i = 0; i < 8; ++i) {
+                float q = xg[(pg[i] - 1)/2];
+                float w = weight[i];
+                sumqx += w*q*xval[i];
+                sumq2 += w*q*q;
+            }
+            printf("    neighbour %d: sumqx = %g sumq2 = %g\n", j, (double)sumqx, (double)sumq2);
+        }
+    }
+    GGML_ASSERT(grid_index >= 0);
+    const int8_t * pg = (const int8_t *)(grid + grid_index);
+    for (int i = 0; i < 8; ++i) L[i] = (pg[i] - 1)/2;
+    return grid_index;
+}
+
+static int iq1_sort_helper(const void * left, const void * right) {
+    const float * l = left;
+    const float * r = right;
+    return *l < *r ? -1 : *l > *r ? 1 : 0;
+}
+
+#define IQ1S_BLOCK_SIZE 32
+#define IQ1M_BLOCK_SIZE 16
+static void quantize_row_iq1_s_impl(const float * restrict x, void * restrict vy, int64_t n, const float * restrict quant_weights,
+        float    * scales,
+        float    * weight,
+        float    * sumx,
+        float    * sumw,
+        float    * pairs,
+        int8_t   * L,
+        uint16_t * index,
+        int8_t   * shifts) {
+
+    const int gindex = iq2_data_index(GGML_TYPE_IQ1_S);
+
+    const uint64_t * kgrid_q2xs      = iq2_data[gindex].grid;
+    const int      * kmap_q2xs       = iq2_data[gindex].map;
+    const uint16_t * kneighbors_q2xs = iq2_data[gindex].neighbours;
+
+    GGML_ASSERT(quant_weights   && "missing quantization weights");
+    GGML_ASSERT(kgrid_q2xs      && "forgot to call ggml_quantize_init()?");
+    GGML_ASSERT(kmap_q2xs       && "forgot to call ggml_quantize_init()?");
+    GGML_ASSERT(kneighbors_q2xs && "forgot to call ggml_quantize_init()?");
+    GGML_ASSERT(n%QK_K == 0);
+
+    block_iq1_s * y = vy;
+
+    const int64_t nbl = n/QK_K;
+
+    const int block_size = IQ1S_BLOCK_SIZE;
+
+    const float x_p[3] = {-1 + IQ1S_DELTA,  IQ1S_DELTA, 1 + IQ1S_DELTA};
+    const float x_m[3] = {-1 - IQ1S_DELTA, -IQ1S_DELTA, 1 - IQ1S_DELTA};
+
+
+    int * idx = (int *)(pairs + 1);
+
+    for (int ibl = 0; ibl < nbl; ++ibl) {
+
+        y[ibl].d = GGML_FP32_TO_FP16(0.f);
+        memset(y[ibl].qs, 0, QK_K/8);
+        memset(y[ibl].qh, 0, QK_K/16);
+
+        float max_scale = 0;
+
+        const float * xbl = x + QK_K*ibl;
+        float sumx2 = 0;
+        for (int i = 0; i < QK_K; ++i) sumx2 += xbl[i]*xbl[i];
+        float sigma2 = 2*sumx2/QK_K;
+
+        for (int ib = 0; ib < QK_K/block_size; ++ib) {
+            const float * xb = xbl + block_size*ib;
+            const float * qw = quant_weights + QK_K*ibl + block_size*ib;
+            for (int i = 0; i < block_size; ++i) weight[i] = qw[i] * sqrtf(sigma2 + xb[i]*xb[i]);
+            float max = fabsf(xb[0]);
+            for (int i = 1; i < block_size; ++i) max = MAX(max, fabsf(xb[i]));
+            if (max < GROUP_MAX_EPS_IQ1_S) {
+                scales[ib] = 0;
+                memset(L, 1, block_size);
+                continue;
+            }
+            // Here we solve exactly the sum of squared difference (SSD) weighted minimization problem.
+            // With just 3 allowed quant values (-1, 0, 1), we can search exhaustively for the two
+            // boundaries that split the weights xb[i] into 3 groups. To do so, we sort the weights
+            // in ascending order, compute Si = sum[weight[j] xb[j], j = 0...i] and
+            // Wi = sum[weight[j], j = 0...i], and use these to quckly get get the optimum scale
+            // for each possible and score for each split.
+            for (int j = 0; j < block_size; ++j) {
+                pairs[2*j] = xb[j];
+                idx[2*j] = j;
+            }
+            qsort(pairs, block_size, 2*sizeof(float), iq1_sort_helper);
+            {
+                sumx[0] = sumw[0] = 0;
+                for (int j = 0; j < block_size; ++j) {
+                    int i = idx[2*j];
+                    sumx[j+1] = sumx[j] + weight[i]*xb[i];
+                    sumw[j+1] = sumw[j] + weight[i];
+                }
+            }
+            float best_score = -FLT_MIN, scale = max;
+            int besti1 = -1, besti2 = -1, best_shift = 0;
+            for (int i1 = 0; i1 <= block_size; ++i1) {
+                for (int i2 = i1; i2 <= block_size; ++i2) {
+                    float sumqx = (sumx[i1] - sumx[0])*x_p[0] + (sumx[i2] - sumx[i1])*x_p[1] + (sumx[block_size] - sumx[i2])*x_p[2];
+                    float sumq2 = (sumw[i1] - sumw[0])*x_p[0]*x_p[0] + (sumw[i2] - sumw[i1])*x_p[1]*x_p[1] + (sumw[block_size] - sumw[i2])*x_p[2]*x_p[2];
+                    if (sumq2 > 0 && sumqx*sumqx > best_score*sumq2) {
+                        scale = sumqx/sumq2; best_score = scale*sumqx;
+                        besti1 = i1; besti2 = i2; best_shift = 1;
+                    }
+                    sumqx = (sumx[i1] - sumx[0])*x_m[0] + (sumx[i2] - sumx[i1])*x_m[1] + (sumx[block_size] - sumx[i2])*x_m[2];
+                    sumq2 = (sumw[i1] - sumw[0])*x_m[0]*x_m[0] + (sumw[i2] - sumw[i1])*x_m[1]*x_m[1] + (sumw[block_size] - sumw[i2])*x_m[2]*x_m[2];
+                    if (sumq2 > 0 && sumqx*sumqx > best_score*sumq2) {
+                        scale = sumqx/sumq2; best_score = scale*sumqx;
+                        besti1 = i1; besti2 = i2; best_shift = -1;
+                    }
+                }
+            }
+            GGML_ASSERT(besti1 >= 0 && besti2 >= 0 && best_shift != 0);
+            for (int j =      0; j < besti1; ++j) L[idx[2*j]] = 0;
+            for (int j = besti1; j < besti2; ++j) L[idx[2*j]] = 1;
+            for (int j = besti2; j < block_size; ++j) L[idx[2*j]] = 2;
+            if (scale < 0) {
+                for (int j = 0; j < block_size; ++j) L[j] = 2 - L[j];
+                scale = -scale; best_shift = -best_shift;
+            }
+            bool all_on_grid = true;
+            const float * xx = best_shift == 1 ? x_p : x_m;
+            for (int k = 0; k < block_size/8; ++k) {
+                uint16_t u = 0;
+                for (int j = 0; j < 8; ++j) u |= (L[8*k+j] << 2*j);
+                int grid_index = kmap_q2xs[u];
+                if (grid_index < 0) {
+                    all_on_grid = false;
+                    const uint16_t * neighbours = kneighbors_q2xs - kmap_q2xs[u] - 1;
+                    grid_index = iq1_find_best_neighbour2(neighbours, kgrid_q2xs, xb + 8*k, weight + 8*k, scale, xx, L + 8*k, NGRID_IQ1S);
+                    GGML_ASSERT(grid_index >= 0);
+                }
+                index[k] = grid_index;
+            }
+            if (!all_on_grid) {
+                float sumqx = 0, sumq2 = 0;
+                for (int k = 0; k < block_size/8; ++k) {
+                    const int8_t * pg = (const int8_t *)(kgrid_q2xs + index[k]);
+                    for (int j = 0; j < 8; ++j) {
+                        float w = weight[8*k + j];
+                        float q = xx[(pg[j] - 1)/2];
+                        sumqx += w*q*xb[8*k+j];
+                        sumq2 += w*q*q;
+                    }
+                }
+                if (sumqx > 0 && sumq2 > 0) scale = sumqx/sumq2;
+            }
+            uint16_t h = 0;
+            for (int k = 0; k < block_size/8; ++k) {
+                y[ibl].qs[(block_size/8)*ib + k] = index[k] & 255;
+                h |= (index[k] >> 8) << 3*k;
+            }
+            y[ibl].qh[ib] = h;
+            GGML_ASSERT(scale >= 0);
+            scales[ib] = scale;
+            shifts[ib] = best_shift;
+            max_scale = MAX(max_scale, scale);
+        }
+
+        if (!max_scale) {
+            continue;
+        }
+
+        float d = max_scale/15;
+        y[ibl].d = GGML_FP32_TO_FP16(d*1.125f); // 1.125f is another fudge factor. Don't ask me why it is needed.
+        float id = 1/d;
+        for (int ib = 0; ib < QK_K/block_size; ++ib) {
+            int l = nearest_int(0.5f*(id*scales[ib]-1));
+            l = MAX(0, MIN(7, l));
+            if (shifts[ib] == -1) l |= 8;
+            y[ibl].qh[ib] |= (l << 12);
+        }
+    }
+}
+
+size_t quantize_iq1_s(const float * restrict src, void * restrict dst, int64_t nrow, int64_t n_per_row, const float * quant_weights) {
+    GGML_ASSERT(n_per_row%QK_K == 0);
+    float  scales[QK_K/IQ1S_BLOCK_SIZE];
+    float  weight[IQ1S_BLOCK_SIZE];
+    int8_t L[IQ1S_BLOCK_SIZE];
+    float  sumx[IQ1S_BLOCK_SIZE+1];
+    float  sumw[IQ1S_BLOCK_SIZE+1];
+    float  pairs[2*IQ1S_BLOCK_SIZE];
+    uint16_t index[IQ1S_BLOCK_SIZE/8];
+    int8_t shifts[QK_K/IQ1S_BLOCK_SIZE];
+    int64_t nblock = n_per_row/QK_K;
+    char * qrow = (char *)dst;
+    for (int64_t row = 0; row < nrow; ++row) {
+        quantize_row_iq1_s_impl(src, qrow, n_per_row, quant_weights, scales, weight, sumx, sumw, pairs, L, index, shifts);
+        src += n_per_row;
+        qrow += nblock*sizeof(block_iq1_s);
+    }
+    return nrow * nblock * sizeof(block_iq1_s);
+}
+
+static void quantize_row_iq1_m_impl(const float * restrict x, void * restrict vy, int64_t n, const float * restrict quant_weights,
+        float    * scales,
+        float    * weight,
+        float    * pairs,
+        int8_t   * L,
+        uint16_t * index,
+        int8_t   * shifts) {
+
+    const int gindex = iq2_data_index(GGML_TYPE_IQ1_M);
+
+    const uint64_t * kgrid_q2xs      = iq2_data[gindex].grid;
+    const int      * kmap_q2xs       = iq2_data[gindex].map;
+    const uint16_t * kneighbors_q2xs = iq2_data[gindex].neighbours;
+
+    //GGML_ASSERT(quant_weights   && "missing quantization weights");
+    GGML_ASSERT(kgrid_q2xs      && "forgot to call ggml_quantize_init()?");
+    GGML_ASSERT(kmap_q2xs       && "forgot to call ggml_quantize_init()?");
+    GGML_ASSERT(kneighbors_q2xs && "forgot to call ggml_quantize_init()?");
+    GGML_ASSERT(n%QK_K == 0);
+
+    block_iq1_m * y = vy;
+
+    const int64_t nbl = n/QK_K;
+
+    const int block_size = IQ1M_BLOCK_SIZE;
+
+    const float x_p[3] = {-1 + IQ1M_DELTA,  IQ1M_DELTA, 1 + IQ1M_DELTA};
+    const float x_m[3] = {-1 - IQ1M_DELTA, -IQ1M_DELTA, 1 - IQ1M_DELTA};
+    const uint8_t masks[4] = {0x00, 0x80, 0x08, 0x88};
+
+    int * idx = (int *)(pairs + 1);
+
+    float sumqx[4], sumq2[4];
+
+    iq1m_scale_t s;
+    const float * xx;
+
+    for (int ibl = 0; ibl < nbl; ++ibl) {
+        memset(y[ibl].qs, 0, QK_K/8);
+        memset(y[ibl].qh, 0, QK_K/16);
+        memset(y[ibl].scales, 0, QK_K/32);
+
+        float max_scale = 0;
+
+        const float * xbl = x + QK_K*ibl;
+        float sumx2 = 0;
+        for (int i = 0; i < QK_K; ++i) sumx2 += xbl[i]*xbl[i];
+        float sigma2 = 2*sumx2/QK_K;
+
+        for (int ib = 0; ib < QK_K/block_size; ++ib) {
+            const float * xb = xbl + block_size*ib;
+            if (quant_weights) {
+                const float * qw = quant_weights + QK_K*ibl + block_size*ib;
+                for (int i = 0; i < block_size; ++i) weight[i] = qw[i] * sqrtf(sigma2 + xb[i]*xb[i]);
+            } else {
+                for (int i = 0; i < block_size; ++i) weight[i] = xb[i]*xb[i];
+            }
+            float max = fabsf(xb[0]);
+            for (int i = 1; i < block_size; ++i) max = MAX(max, fabsf(xb[i]));
+            if (max < GROUP_MAX_EPS_IQ1_M) {
+                scales[ib] = 0;
+                memset(L, 1, block_size);
+                continue;
+            }
+            // Here we solve exactly the sum of squared difference (SSD) weighted minimization problem.
+            // With just 3 allowed quant values (-1, 0, 1), we can search exhaustively for the two
+            // boundaries that split the weights xb[i] into 3 groups. To do so, we sort the weights
+            // in ascending order, compute Si = sum[weight[j] xb[j], j = 0...i] and
+            // Wi = sum[weight[j], j = 0...i], and use these to quckly get get the optimum scale
+            // for each possible and score for each split.
+            for (int j = 0; j < block_size; ++j) {
+                pairs[2*j] = xb[j];
+                idx[2*j] = j;
+            }
+            qsort(pairs, block_size, 2*sizeof(float), iq1_sort_helper);
+            float best_score = -FLT_MIN, scale = max;
+            int besti1 = -1, besti2 = -1, best_k = -1;
+            // 0: +, +
+            // 1: +, -
+            // 2: -, +
+            // 3: -, -
+            for (int i1 = 0; i1 <= block_size; ++i1) {
+                for (int i2 = i1; i2 <= block_size; ++i2) {
+                    memset(sumqx, 0, 4*sizeof(float));
+                    memset(sumq2, 0, 4*sizeof(float));
+                    for (int j = 0; j < i1; ++j) {
+                        int i = idx[2*j];
+                        if (i < block_size/2) {
+                            sumqx[0] += weight[i]*x_p[0]*xb[i];
+                            sumqx[1] += weight[i]*x_p[0]*xb[i];
+                            sumqx[2] += weight[i]*x_m[0]*xb[i];
+                            sumqx[3] += weight[i]*x_m[0]*xb[i];
+                            sumq2[0] += weight[i]*x_p[0]*x_p[0];
+                            sumq2[1] += weight[i]*x_p[0]*x_p[0];
+                            sumq2[2] += weight[i]*x_m[0]*x_m[0];
+                            sumq2[3] += weight[i]*x_m[0]*x_m[0];
+                        } else {
+                            sumqx[0] += weight[i]*x_p[0]*xb[i];
+                            sumqx[2] += weight[i]*x_p[0]*xb[i];
+                            sumqx[1] += weight[i]*x_m[0]*xb[i];
+                            sumqx[3] += weight[i]*x_m[0]*xb[i];
+                            sumq2[0] += weight[i]*x_p[0]*x_p[0];
+                            sumq2[2] += weight[i]*x_p[0]*x_p[0];
+                            sumq2[1] += weight[i]*x_m[0]*x_m[0];
+                            sumq2[3] += weight[i]*x_m[0]*x_m[0];
+                        }
+                    }
+                    for (int j = i1; j < i2; ++j) {
+                        int i = idx[2*j];
+                        if (i < block_size/2) {
+                            sumqx[0] += weight[i]*x_p[1]*xb[i];
+                            sumqx[1] += weight[i]*x_p[1]*xb[i];
+                            sumqx[2] += weight[i]*x_m[1]*xb[i];
+                            sumqx[3] += weight[i]*x_m[1]*xb[i];
+                            sumq2[0] += weight[i]*x_p[1]*x_p[1];
+                            sumq2[1] += weight[i]*x_p[1]*x_p[1];
+                            sumq2[2] += weight[i]*x_m[1]*x_m[1];
+                            sumq2[3] += weight[i]*x_m[1]*x_m[1];
+                        } else {
+                            sumqx[0] += weight[i]*x_p[1]*xb[i];
+                            sumqx[2] += weight[i]*x_p[1]*xb[i];
+                            sumqx[1] += weight[i]*x_m[1]*xb[i];
+                            sumqx[3] += weight[i]*x_m[1]*xb[i];
+                            sumq2[0] += weight[i]*x_p[1]*x_p[1];
+                            sumq2[2] += weight[i]*x_p[1]*x_p[1];
+                            sumq2[1] += weight[i]*x_m[1]*x_m[1];
+                            sumq2[3] += weight[i]*x_m[1]*x_m[1];
+                        }
+                    }
+                    for (int j = i2; j < block_size; ++j) {
+                        int i = idx[2*j];
+                        if (i < block_size/2) {
+                            sumqx[0] += weight[i]*x_p[2]*xb[i];
+                            sumqx[1] += weight[i]*x_p[2]*xb[i];
+                            sumqx[2] += weight[i]*x_m[2]*xb[i];
+                            sumqx[3] += weight[i]*x_m[2]*xb[i];
+                            sumq2[0] += weight[i]*x_p[2]*x_p[2];
+                            sumq2[1] += weight[i]*x_p[2]*x_p[2];
+                            sumq2[2] += weight[i]*x_m[2]*x_m[2];
+                            sumq2[3] += weight[i]*x_m[2]*x_m[2];
+                        } else {
+                            sumqx[0] += weight[i]*x_p[2]*xb[i];
+                            sumqx[2] += weight[i]*x_p[2]*xb[i];
+                            sumqx[1] += weight[i]*x_m[2]*xb[i];
+                            sumqx[3] += weight[i]*x_m[2]*xb[i];
+                            sumq2[0] += weight[i]*x_p[2]*x_p[2];
+                            sumq2[2] += weight[i]*x_p[2]*x_p[2];
+                            sumq2[1] += weight[i]*x_m[2]*x_m[2];
+                            sumq2[3] += weight[i]*x_m[2]*x_m[2];
+                        }
+                    }
+                    for (int k = 0; k < 4; ++k) {
+                        if (sumq2[k] > 0 && sumqx[k]*sumqx[k] > best_score*sumq2[k]) {
+                            scale = sumqx[k]/sumq2[k]; best_score = scale*sumqx[k];
+                            besti1 = i1; besti2 = i2; best_k = k;
+                        }
+                    }
+                }
+            }
+            GGML_ASSERT(besti1 >= 0 && besti2 >= 0 && best_k >= 0);
+            for (int j =      0; j < besti1; ++j) L[idx[2*j]] = 0;
+            for (int j = besti1; j < besti2; ++j) L[idx[2*j]] = 1;
+            for (int j = besti2; j < block_size; ++j) L[idx[2*j]] = 2;
+            if (scale < 0) {
+                for (int j = 0; j < block_size; ++j) L[j] = 2 - L[j];
+                scale = -scale;
+                best_k = best_k == 0 ? 3 : best_k == 1 ? 2 : best_k == 2 ? 1 : 0;
+            }
+            bool all_on_grid = true;
+            for (int k = 0; k < block_size/8; ++k) {
+                if (k == 0) xx = best_k < 2 ? x_p : x_m;
+                else xx = best_k%2 == 0 ? x_p : x_m;
+                uint16_t u = 0;
+                for (int j = 0; j < 8; ++j) u |= (L[8*k+j] << 2*j);
+                int grid_index = kmap_q2xs[u];
+                if (grid_index < 0) {
+                    all_on_grid = false;
+                    const uint16_t * neighbours = kneighbors_q2xs - kmap_q2xs[u] - 1;
+                    grid_index = iq1_find_best_neighbour2(neighbours, kgrid_q2xs, xb + 8*k, weight + 8*k, scale, xx, L + 8*k, NGRID_IQ1S);
+                    GGML_ASSERT(grid_index >= 0);
+                }
+                index[k] = grid_index;
+            }
+            if (!all_on_grid) {
+                float sumqx_f = 0, sumq2_f = 0;
+                for (int k = 0; k < block_size/8; ++k) {
+                    if (k == 0) xx = best_k < 2 ? x_p : x_m;
+                    else xx = best_k%2 == 0 ? x_p : x_m;
+                    const int8_t * pg = (const int8_t *)(kgrid_q2xs + index[k]);
+                    for (int j = 0; j < 8; ++j) {
+                        float w = weight[8*k + j];
+                        float q = xx[(pg[j] - 1)/2];
+                        sumqx_f += w*q*xb[8*k+j];
+                        sumq2_f += w*q*q;
+                    }
+                }
+                if (sumqx_f > 0 && sumq2_f > 0) scale = sumqx_f/sumq2_f;
+            }
+            y[ibl].qs[2*ib + 0] = index[0] & 255;
+            y[ibl].qs[2*ib + 1] = index[1] & 255;
+            y[ibl].qh[ib] = (index[0] >> 8) | ((index[1] >> 8) << 4);
+            GGML_ASSERT(scale >= 0);
+            scales[ib] = scale;
+            shifts[ib] = best_k;
+            max_scale = MAX(max_scale, scale);
+        }
+
+        if (!max_scale) {
+            continue;
+        }
+
+        uint16_t * sc = (uint16_t *)y[ibl].scales;
+        float d = max_scale/15;
+        float id = 1/d;
+        float sumqx_f = 0, sumq2_f = 0;
+        for (int ib = 0; ib < QK_K/block_size; ++ib) {
+            int l = nearest_int(0.5f*(id*scales[ib+0]-1));
+            l = MAX(0, MIN(7, l));
+            sc[ib/4] |= (l << 3*(ib%4));
+            y[ibl].qh[ib] |= masks[shifts[ib]];
+            const float * xb = xbl + block_size*ib;
+            if (quant_weights) {
+                const float * qw = quant_weights + QK_K*ibl + block_size*ib;
+                for (int i = 0; i < block_size; ++i) weight[i] = qw[i] * sqrtf(sigma2 + xb[i]*xb[i]);
+            } else {
+                for (int i = 0; i < block_size; ++i) weight[i] = xb[i]*xb[i];
+            }
+            for (int k = 0; k < block_size/8; ++k) {
+                if (k == 0) xx = shifts[ib] < 2 ? x_p : x_m;
+                else xx = shifts[ib]%2 == 0 ? x_p : x_m;
+                const int8_t * pg = (const int8_t *)(kgrid_q2xs + y[ibl].qs[2*ib+k] + ((y[ibl].qh[ib] << (8 - 4*k)) & 0x700));
+                for (int j = 0; j < 8; ++j) {
+                    float w = weight[8*k + j];
+                    float q = xx[(pg[j] - 1)/2]*(2*l+1);
+                    sumqx_f += w*q*xb[8*k+j];
+                    sumq2_f += w*q*q;
+                }
+            }
+        }
+        if (sumq2_f > 0) d = sumqx_f/sumq2_f;
+        s.f16 = GGML_FP32_TO_FP16(d*1.1125f); // 1.1125f is another fudge factor. Don't ask me why it is needed.
+        sc[0] |= ((s.u16 & 0x000f) << 12);
+        sc[1] |= ((s.u16 & 0x00f0) <<  8);
+        sc[2] |= ((s.u16 & 0x0f00) <<  4);
+        sc[3] |= ((s.u16 & 0xf000) <<  0);
+    }
+}
+
+size_t quantize_iq1_m(const float * restrict src, void * restrict dst, int64_t nrow, int64_t n_per_row, const float * quant_weights) {
+    GGML_ASSERT(n_per_row%QK_K == 0);
+    float  scales[QK_K/IQ1M_BLOCK_SIZE];
+    float  weight[IQ1M_BLOCK_SIZE];
+    int8_t L[IQ1M_BLOCK_SIZE];
+    float  pairs[2*IQ1M_BLOCK_SIZE];
+    uint16_t index[IQ1M_BLOCK_SIZE/8];
+    int8_t shifts[QK_K/IQ1M_BLOCK_SIZE];
+    int64_t nblock = n_per_row/QK_K;
+    char * qrow = (char *)dst;
+    for (int64_t row = 0; row < nrow; ++row) {
+        quantize_row_iq1_m_impl(src, qrow, n_per_row, quant_weights, scales, weight, pairs, L, index, shifts);
+        src += n_per_row;
+        qrow += nblock*sizeof(block_iq1_m);
+    }
+    return nrow * nblock * sizeof(block_iq1_m);
+}
+
+// ============================ 4-bit non-linear quants
+
+static inline int best_index_int8(int n, const int8_t * val, float x) {
+    if (x <= val[0]) return 0;
+    if (x >= val[n-1]) return n-1;
+    int ml = 0, mu = n-1;
+    while (mu-ml > 1) {
+        int mav = (ml+mu)/2;
+        if (x < val[mav]) mu = mav; else ml = mav;
+    }
+    return x - val[mu-1] < val[mu] - x ? mu-1 : mu;
+}
+
+static void quantize_row_iq4_nl_impl(const int super_block_size, const int block_size, const float * restrict x,
+        ggml_fp16_t * dh, uint8_t * q4, uint16_t * scales_h, uint8_t * scales_l,
+        float * scales, float * weight, uint8_t * L,
+        const int8_t * values,
+        const float * quant_weights,
+        const int ntry) {
+
+    float sigma2 = 0;
+    for (int j = 0; j < super_block_size; ++j) sigma2 += x[j]*x[j];
+    sigma2 *= 2.f/super_block_size;
+
+    memset(q4, 0, super_block_size/2);
+    dh[0] = GGML_FP32_TO_FP16(0.f);
+
+    float max_scale = 0, amax_scale = 0;
+    for (int ib = 0; ib < super_block_size/block_size; ++ib) {
+        const float * xb = x + ib*block_size;
+        uint8_t * Lb = L + ib*block_size;
+        if (quant_weights) {
+            const float * qw = quant_weights + ib*block_size;
+            for (int j = 0; j < block_size; ++j) weight[j] = qw[j] * sqrtf(sigma2 + xb[j]*xb[j]);
+        } else {
+            for (int j = 0; j < block_size; ++j) weight[j] = xb[j]*xb[j];
+        }
+        float amax = 0, max = 0;
+        for (int j = 0; j < block_size; ++j) {
+            float ax = fabsf(xb[j]);
+            if (ax > amax) {
+                amax = ax; max = xb[j];
+            }
+        }
+        if (amax < GROUP_MAX_EPS) {
+            scales[ib] = 0;
+            continue;
+        }
+        float d = ntry > 0 ? -max/values[0] : max/values[0];
+        float id = 1/d;
+        float sumqx = 0, sumq2 = 0;
+        for (int j = 0; j < block_size; ++j) {
+            float al = id*xb[j];
+            int l = best_index_int8(16, values, al);
+            Lb[j] = l;
+            float q = values[l];
+            float w = weight[j];
+            sumqx += w*q*xb[j];
+            sumq2 += w*q*q;
+        }
+        d = sumqx/sumq2;
+        float best = d*sumqx;
+        for (int itry = -ntry; itry <= ntry; ++itry) {
+            id = (itry + values[0])/max;
+            sumqx = sumq2 = 0;
+            for (int j = 0; j < block_size; ++j) {
+                float al = id*xb[j];
+                int l = best_index_int8(16, values, al);
+                float q = values[l];
+                float w = weight[j];
+                sumqx += w*q*xb[j];
+                sumq2 += w*q*q;
+            }
+            if (sumq2 > 0 && sumqx*sumqx > best*sumq2) {
+                d = sumqx/sumq2; best = d * sumqx;
+            }
+        }
+        scales[ib] = d;
+        float abs_d = fabsf(d);
+        if (abs_d > amax_scale) {
+            amax_scale = abs_d; max_scale = d;
+        }
+    }
+
+    if (super_block_size/block_size > 1) {
+        int nb = super_block_size/block_size;
+        memset(scales_h, 0, ((nb+7)/8)*sizeof(uint16_t));
+        float d = -max_scale/32;
+        dh[0] = GGML_FP32_TO_FP16(d);
+        float id = d ? 1/d : 0.f;
+        for (int ib = 0; ib < super_block_size/block_size; ++ib) {
+            int l = nearest_int(id*scales[ib]);
+            l = MAX(-32, MIN(31, l));
+            float dl = d * l;
+            float idl = dl ? 1/dl : 0.f;
+            uint8_t * Lb = L + ib*block_size;
+            const float * xb = x + ib*block_size;
+            for (int j = 0; j < block_size; ++j) {
+                Lb[j] = best_index_int8(16, values, idl*xb[j]);
+            }
+            l += 32;
+            uint8_t l_l = l & 0xf;
+            uint8_t l_h = l >>  4;
+            if (ib%2 == 0) scales_l[ib/2] = l_l;
+            else scales_l[ib/2] |= (l_l << 4);
+            scales_h[ib/8] |= (l_h << 2*(ib%8));
+        }
+    } else {
+        dh[0] = GGML_FP32_TO_FP16(scales[0]);
+        if (ntry > 0) {
+            float id = scales[0] ? 1/scales[0] : 0;
+            for (int j = 0; j < super_block_size; ++j) {
+                L[j] = best_index_int8(16, values, id*x[j]);
+            }
+        }
+    }
+
+    for (int i = 0; i < super_block_size/32; ++i) {
+        for (int j = 0; j < 16; ++j) {
+            q4[16*i + j] = L[32*i + j] | (L[32*i + 16 + j] << 4);
+        }
+    }
+}
+
+size_t quantize_iq4_nl(const float * restrict src, void * restrict dst, int64_t nrow, int64_t n_per_row, const float * quant_weights) {
+    GGML_ASSERT(n_per_row%QK4_NL == 0);
+    int64_t nblock = n_per_row/QK4_NL;
+    char * qrow = (char *)dst;
+    uint8_t L[QK4_NL];
+    float weight[QK4_NL];
+    uint16_t unused_h;
+    uint8_t * unused_l = NULL;
+    float scale;
+    for (int64_t row = 0; row < nrow; ++row) {
+        block_iq4_nl * iq4 = (block_iq4_nl *)qrow;
+        for (int ibl = 0; ibl < nblock; ++ibl) {
+            const float * qw = quant_weights ? quant_weights + QK4_NL*ibl : NULL;
+            quantize_row_iq4_nl_impl(QK4_NL, 32, src + QK4_NL*ibl, &iq4[ibl].d, iq4[ibl].qs, &unused_h, unused_l,
+                    &scale, weight, L, kvalues_iq4nl, qw, 7);
+        }
+        src += n_per_row;
+        qrow += nblock*sizeof(block_iq4_nl);
+    }
+    return nrow * nblock * sizeof(block_iq4_nl);
+}
+
+void quantize_row_iq4_nl(const float * restrict x, void * restrict vy, int64_t k) {
+    GGML_ASSERT(k%QK4_NL == 0);
+    int64_t nblock = k/QK4_NL;
+    uint8_t L[QK4_NL];
+    float weight[QK4_NL];
+    uint16_t unused_h;
+    uint8_t * unused_l = NULL;
+    float scale;
+    block_iq4_nl * iq4 = (block_iq4_nl *)vy;
+    for (int ibl = 0; ibl < nblock; ++ibl) {
+        quantize_row_iq4_nl_impl(QK4_NL, 32, x + QK4_NL*ibl, &iq4[ibl].d, iq4[ibl].qs, &unused_h, unused_l,
+                &scale, weight, L, kvalues_iq4nl, NULL, -1);
+    }
+}
+
+void quantize_row_iq4_nl_ref(const float * restrict x, block_iq4_nl * restrict y, int64_t k) {
+    assert(k % QK4_NL == 0);
+    quantize_row_iq4_nl(x, y, k);
+}
+
+size_t quantize_iq4_xs(const float * restrict src, void * restrict dst, int64_t nrow, int64_t n_per_row, const float * quant_weights) {
+    GGML_ASSERT(n_per_row%QK_K == 0);
+    int64_t nblock = n_per_row/QK_K;
+    char * qrow = (char *)dst;
+    uint8_t L[QK_K];
+    float weight[32];
+    float scales[QK_K/32];
+    for (int64_t row = 0; row < nrow; ++row) {
+        block_iq4_xs * iq4 = (block_iq4_xs *)qrow;
+        for (int ibl = 0; ibl < nblock; ++ibl) {
+            const float * qw = quant_weights ? quant_weights + QK_K*ibl : NULL;
+            quantize_row_iq4_nl_impl(QK_K, 32, src + QK_K*ibl, &iq4[ibl].d, iq4[ibl].qs, &iq4[ibl].scales_h, iq4[ibl].scales_l,
+                    scales, weight, L, kvalues_iq4nl, qw, 7);
+        }
+        src += n_per_row;
+        qrow += nblock*sizeof(block_iq4_xs);
+    }
+    return nrow * nblock * sizeof(block_iq4_xs);
+}
+
+void quantize_row_iq4_xs(const float * restrict x, void * restrict vy, int64_t k) {
+    assert(k % QK_K == 0);
+    block_iq4_xs * restrict y = vy;
+    quantize_row_iq4_xs_ref(x, y, k);
+}
+
+void quantize_row_iq4_xs_ref(const float * restrict x, block_iq4_xs * restrict y, int64_t k) {
+    assert(k % QK_K == 0);
+    quantize_iq4_xs(x, y, 1, k, NULL);
+}
+
+// =============================== 2.5625 bpw
+
+static void quantize_row_iq2_s_impl(const float * restrict x, void * restrict vy, int64_t n, const float * restrict quant_weights) {
+
+    const int gindex = iq2_data_index(GGML_TYPE_IQ2_S);
+
+    const uint64_t * kgrid_q2xs      = iq2_data[gindex].grid;
+    const int      * kmap_q2xs       = iq2_data[gindex].map;
+    const uint16_t * kneighbors_q2xs = iq2_data[gindex].neighbours;
+
+    GGML_ASSERT(kmap_q2xs       && "forgot to call ggml_quantize_init()?");
+    GGML_ASSERT(kgrid_q2xs      && "forgot to call ggml_quantize_init()?");
+    GGML_ASSERT(kneighbors_q2xs && "forgot to call ggml_quantize_init()?");
+    GGML_ASSERT(n%QK_K == 0);
+
+    const int kMaxQ = 3;
+
+    const int64_t nbl = n/QK_K;
+
+    block_iq2_s * y = vy;
+
+    float scales[QK_K/16];
+    float weight[16];
+    float xval[16];
+    int8_t L[16];
+    int8_t Laux[16];
+    float  waux[16];
+    bool   is_on_grid[2];
+    bool   is_on_grid_aux[2];
+    uint8_t block_signs[2];
+
+    for (int ibl = 0; ibl < nbl; ++ibl) {
+
+        memset(&y[ibl], 0, sizeof(block_iq2_s));
+        y[ibl].d = GGML_FP32_TO_FP16(0.f);
+
+        float max_scale = 0;
+
+        const float * xbl = x + QK_K*ibl;
+        float sumx2 = 0;
+        for (int i = 0; i < QK_K; ++i) sumx2 += xbl[i]*xbl[i];
+        float sigma2 = 2*sumx2/QK_K;
+
+        for (int ib = 0; ib < QK_K/16; ++ib) {
+            const float * xb = xbl + 16*ib;
+            if (quant_weights) {
+                const float * qw = quant_weights + QK_K*ibl + 16*ib;
+                for (int i = 0; i < 16; ++i) weight[i] = qw[i] * sqrtf(sigma2 + xb[i]*xb[i]);
+            } else {
+                for (int i = 0; i < 16; ++i) weight[i] = 0.25f*sigma2 + xb[i]*xb[i];
+            }
+            for (int i = 0; i < 16; ++i) waux[i] = sqrtf(weight[i]);
+            for (int k = 0; k < 2; ++k) {
+                uint8_t s = 0;
+                for (int i = 0; i < 8; ++i) {
+                    if (xb[8*k + i] >= 0) xval[8*k + i] = xb[8*k + i];
+                    else {
+                        xval[8*k + i] = -xb[8*k + i]; s |= (1 << i);
+                    }
+                }
+                block_signs[k] = s;
+            }
+            float max = xval[0];
+            for (int i = 1; i < 16; ++i) max = MAX(max, xval[i]);
+            if (max < GROUP_MAX_EPS_IQ2_S) {
+                scales[ib] = 0;
+                continue;
+            }
+            float best = 0;
+            float scale = max/(2*kMaxQ-1);
+            is_on_grid[0] = is_on_grid[1] = true;
+            for (int is = -9; is <= 9; ++is) {
+                float id = (2*kMaxQ-1+is*0.1f)/max;
+                float this_scale = 1/id;
+                for (int k = 0; k < 2; ++k) {
+                    for (int i = 0; i < 8; ++i) {
+                        int l = nearest_int(0.5f*(id*xval[8*k+i]-1));
+                        Laux[8*k+i] = MAX(0, MIN(kMaxQ-1, l));
+                    }
+                    uint16_t u = 0;
+                    for (int i = 0; i < 8; ++i) u |= (Laux[8*k+i] << 2*i);
+                    int grid_index = kmap_q2xs[u];
+                    is_on_grid_aux[k] = true;
+                    if (grid_index < 0) {
+                        is_on_grid_aux[k] = false;
+                        const uint16_t * neighbours = kneighbors_q2xs - kmap_q2xs[u] - 1;
+                        grid_index = iq2_find_best_neighbour(neighbours, kgrid_q2xs, xval + 8*k, waux + 8*k, this_scale, Laux + 8*k);
+                    }
+                }
+                float sumqx = 0, sumq2 = 0;
+                for (int i = 0; i < 16; ++i) {
+                    float w = weight[i];
+                    float q = 2*Laux[i] + 1;
+                    sumqx += w*xval[i]*q;
+                    sumq2 += w*q*q;
+                }
+                if (sumq2 > 0 && sumqx*sumqx > best*sumq2) {
+                    scale = sumqx/sumq2; best = scale*sumqx;
+                    for (int i = 0; i < 16; ++i) L[i] = Laux[i];
+                    for (int k = 0; k <  2; ++k) is_on_grid[k] = is_on_grid_aux[k];
+                }
+            }
+            int n_not_ongrid = 0;
+            for (int k = 0; k < 2; ++k) if (!is_on_grid[k]) ++n_not_ongrid;
+            if (n_not_ongrid > 0 && scale > 0) {
+                float id = 1/scale;
+                for (int k = 0; k < 2; ++k) {
+                    if (is_on_grid[k]) continue;
+                    uint16_t u = 0;
+                    for (int i = 0; i < 8; ++i) {
+                        int l = nearest_int(0.5f*(id*xval[8*k+i]-1));
+                        l = MAX(0, MIN(kMaxQ-1, l));
+                        u |= (l << 2*i);
+                        L[8*k + i] = l;
+                    }
+                    int grid_index = kmap_q2xs[u];
+                    if (grid_index < 0) {
+                        const uint16_t * neighbours = kneighbors_q2xs - kmap_q2xs[u] - 1;
+                        grid_index = iq2_find_best_neighbour(neighbours, kgrid_q2xs, xval + 8*k, waux + 8*k, scale, L + 8*k);
+                    }
+                }
+                float sumqx = 0, sumq2 = 0;
+                for (int i = 0; i < 16; ++i) {
+                    float w = weight[i];
+                    float q = 2*L[i] + 1;
+                    sumqx += w*xval[i]*q;
+                    sumq2 += w*q*q;
+                }
+                if (sumq2 > 0) scale = sumqx/sumq2;
+            }
+            if (scale < 0) {
+                scale = -scale;
+                for (int k = 0; k < 2; ++k) block_signs[k] = ~block_signs[k];
+            }
+            for (int k = 0; k < 2; ++k) {
+                uint16_t u = 0;
+                for (int i = 0; i < 8; ++i) u |= (L[8*k+i] << 2*i);
+                int grid_index = kmap_q2xs[u];
+                if (grid_index < 0) {
+                    printf("Oops: found point %u not on grid:", u);
+                    for (int i = 0; i < 8; ++i) printf(" %d", L[8*k+i]);
+                    printf("\n");
+                    GGML_ABORT("fatal error");
+                }
+                const int i8 = 2*ib + k;
+                y[ibl].qs[i8] = grid_index & 255;
+                y[ibl].qh[i8/4] |= ((grid_index >> 8) << 2*(i8%4));
+                y[ibl].qs[QK_K/8 + i8] = block_signs[k];
+            }
+            GGML_ASSERT(scale >= 0);
+            scales[ib] = scale;
+            max_scale = MAX(max_scale, scale);
+        }
+
+        if (!max_scale) {
+            continue;
+        }
+
+        float d = max_scale/31;
+        y[ibl].d = GGML_FP32_TO_FP16(d * 0.9875f);
+        float id = 1/d;
+        for (int ib = 0; ib < QK_K/16; ++ib) {
+            int l = nearest_int(0.5f*(id*scales[ib]-1));
+            l = MAX(0, MIN(15, l));
+            if (ib%2 == 0) y[ibl].scales[ib/2] = l;
+            else y[ibl].scales[ib/2] |= (l << 4);
+        }
+    }
+}
+
+size_t quantize_iq2_s(const float * restrict src, void * restrict dst, int64_t nrow, int64_t n_per_row, const float * quant_weights) {
+    GGML_ASSERT(n_per_row%QK_K == 0);
+    int64_t nblock = n_per_row/QK_K;
+    char * qrow = (char *)dst;
+    for (int64_t row = 0; row < nrow; ++row) {
+        quantize_row_iq2_s_impl(src, qrow, n_per_row, quant_weights);
+        src += n_per_row;
+        qrow += nblock*sizeof(block_iq2_s);
+    }
+    return nrow * nblock * sizeof(block_iq2_s);
+}
+
+void quantize_row_iq2_s_ref(const float * restrict x, block_iq2_s * restrict y, int64_t k) {
+    assert(k % QK_K == 0);
+    quantize_iq2_s(x, y, 1, k, NULL);
+}
+
+void quantize_row_iq2_s(const float * restrict x, void * restrict vy, int64_t k) {
+    assert(k % QK_K == 0);
+    block_iq2_s * restrict y = vy;
+    quantize_row_iq2_s_ref(x, y, k);
+}
+
+static bool validate_float(float f, size_t i) {
+    if (isinf(f)) {
+        fprintf(stderr, "ggml_validate_row_data: found inf value at block %zu\n", i);
+        return false;
+    }
+
+    if (isnan(f)) {
+        fprintf(stderr, "ggml_validate_row_data: found nan value at block %zu\n", i);
+        return false;
+    }
+
+    return true;
+}
+
+static bool isinf_fp16(ggml_fp16_t f) {
+    return (f & 0x7c00) == 0x7c00 && (f & 0x03ff) == 0;
+}
+
+static bool isnan_fp16(ggml_fp16_t f) {
+    return (f & 0x7c00) == 0x7c00 && (f & 0x03ff) != 0;
+}
+
+static bool validate_fp16(ggml_fp16_t f, size_t i) {
+    if (isinf_fp16(f)) {
+        fprintf(stderr, "ggml_validate_row_data: found inf value at block %zu\n", i);
+        return false;
+    }
+
+    if (isnan_fp16(f)) {
+        fprintf(stderr, "ggml_validate_row_data: found nan value at block %zu\n", i);
+        return false;
+    }
+
+    return true;
+}
+
+#define VALIDATE_ROW_DATA_D_F16_IMPL(type, data, nb) \
+    const type * q = (const type *) (data); \
+    for (size_t i = 0; i < (nb); ++i) { \
+        if (!validate_fp16(q[i].d, i)) { \
+            return false; \
+        } \
+    }
+
+#define VALIDATE_ROW_DATA_DM_F16_IMPL(type, data, nb, d, m) \
+    const type * q = (const type *) (data); \
+    for (size_t i = 0; i < (nb); ++i) { \
+        if (!validate_fp16(q[i].d, i) || !validate_fp16(q[i].m, i)) { \
+            return false; \
+        } \
+    }
+
+#define VALIDATE_ROW_DATA_DVEC_F16_IMPL(type, data, nb, nr) \
+    const type * q = (const type *) (data); \
+    for (size_t i = 0; i < (nb); ++i) { \
+        for (size_t j = 0; j < (nr); ++j) { \
+            if (!validate_fp16(q[i].d[j], i)) { \
+                return false; \
+            } \
+        } \
+    }
+
+bool ggml_validate_row_data(enum ggml_type type, const void * data, size_t nbytes) {
+    if (type < 0 || type >= GGML_TYPE_COUNT) {
+        fprintf(stderr, "%s: invalid type %d\n", __func__, type);
+        return false;
+    }
+
+    if (nbytes % ggml_type_size(type) != 0) {
+        fprintf(stderr, "%s: invalid size %zu for type %s (type size = %zu)\n", __func__, nbytes, ggml_type_name(type), ggml_type_size(type));
+        return false;
+    }
+
+    const size_t nb = nbytes/ggml_type_size(type);
+
+    switch (type) {
+        case GGML_TYPE_BF16:
+            {
+                int nans = 0;
+                int infs = 0;
+                const unsigned short * f = (const unsigned short *) data;
+                for (size_t i = 0; i < nb; ++i) {
+                    nans += (f[i] & 0x7fff) > 0x7f80;
+                    infs += (f[i] & 0x7fff) == 0x7f80;
+                }
+                if (nans) {
+                    fprintf(stderr, "%s: found %d NaNs in row of %zu BF16 values\n", __func__, nans, nb);
+                    return false;
+                }
+                if (infs) {
+                    fprintf(stderr, "%s: found %d infinities in row of %zu BF16 values\n", __func__, infs, nb);
+                    return false;
+                }
+            } break;
+        case GGML_TYPE_F16:
+            {
+                const ggml_fp16_t * f = (const ggml_fp16_t *) data;
+                size_t i = 0;
+#if defined(__AVX2__)
+                for (; i + 15 < nb; i += 16) {
+                    __m256i v = _mm256_loadu_si256((const __m256i *)(f + i));
+                    __m256i vexp = _mm256_and_si256(v, _mm256_set1_epi16(0x7c00));
+                    __m256i cmp = _mm256_cmpeq_epi16(vexp, _mm256_set1_epi16(0x7c00));
+                    int mask = _mm256_movemask_epi8(cmp);
+                    if (mask) {
+                        for (size_t j = 0; j < 16; ++j) {
+                            if (!validate_fp16(f[i + j], i + j)) {
+                                return false;
+                            }
+                        }
+                        GGML_UNREACHABLE();
+                    }
+                }
+#elif defined(__ARM_NEON)
+                for (; i + 7 < nb; i += 8) {
+                    uint16x8_t v = vld1q_u16(f + i);
+                    uint16x8_t vexp = vandq_u16(v, vdupq_n_u16(0x7c00));
+                    uint16x8_t cmp = vceqq_u16(vexp, vdupq_n_u16(0x7c00));
+                    uint64_t mask = vget_lane_u64(vreinterpret_u64_u8(vshrn_n_u16(cmp, 4)), 0);
+                    if (mask) {
+                        for (size_t j = 0; j < 8; ++j) {
+                            if (!validate_fp16(f[i + j], i + j)) {
+                                return false;
+                            }
+                        }
+                        GGML_UNREACHABLE();
+                    }
+                }
+#endif
+                for (; i < nb; ++i) {
+                    if (!validate_fp16(f[i], i)) {
+                        return false;
+                    }
+                }
+            } break;
+        case GGML_TYPE_F32:
+            {
+                const float * f = (const float *) data;
+                size_t i = 0;
+#if defined(__AVX2__)
+                for (; i + 7 < nb; i += 8) {
+                    __m256i v = _mm256_loadu_si256((const __m256i *)(f + i));
+                    __m256i vexp = _mm256_and_si256(v, _mm256_set1_epi32(0x7f800000));
+                    __m256i cmp = _mm256_cmpeq_epi32(vexp, _mm256_set1_epi32(0x7f800000));
+                    int mask = _mm256_movemask_epi8(cmp);
+                    if (mask) {
+                        for (size_t j = 0; j < 8; ++j) {
+                            if (!validate_float(f[i + j], i + j)) {
+                                return false;
+                            }
+                        }
+                        GGML_UNREACHABLE();
+                    }
+                }
+#elif defined(__ARM_NEON)
+                for (; i + 3 < nb; i += 4) {
+                    uint32x4_t v = vld1q_u32((const uint32_t *)f + i);
+                    uint32x4_t vexp = vandq_u32(v, vdupq_n_u32(0x7f800000));
+                    uint32x4_t cmp = vceqq_u32(vexp, vdupq_n_u32(0x7f800000));
+                    uint64_t mask = vget_lane_u64(vreinterpret_u64_u16(vshrn_n_u32(cmp, 8)), 0);
+                    if (mask) {
+                        for (size_t j = 0; j < 4; ++j) {
+                            if (!validate_float(f[i + j], i + j)) {
+                                return false;
+                            }
+                        }
+                        GGML_UNREACHABLE();
+                    }
+                }
+#endif
+                for (; i < nb; ++i) {
+                    if (!validate_float(f[i], i)) {
+                        return false;
+                    }
+                }
+            } break;
+        case GGML_TYPE_F64:
+            {
+                const double * f = (const double *) data;
+                for (size_t i = 0; i < nb; ++i) {
+                    if (!validate_float(f[i], i)) {
+                        return false;
+                    }
+                }
+            } break;
+        case GGML_TYPE_Q4_0:
+            {
+                VALIDATE_ROW_DATA_D_F16_IMPL(block_q4_0, data, nb);
+            } break;
+        case GGML_TYPE_Q4_1:
+            {
+                VALIDATE_ROW_DATA_DM_F16_IMPL(block_q4_1, data, nb, d, m);
+            } break;
+        case GGML_TYPE_Q5_0:
+            {
+                VALIDATE_ROW_DATA_D_F16_IMPL(block_q5_0, data, nb);
+            } break;
+        case GGML_TYPE_Q5_1:
+            {
+                VALIDATE_ROW_DATA_DM_F16_IMPL(block_q5_1, data, nb, d, m);
+            } break;
+        case GGML_TYPE_Q8_0:
+            {
+                VALIDATE_ROW_DATA_D_F16_IMPL(block_q8_0, data, nb);
+            } break;
+        case GGML_TYPE_Q2_K:
+            {
+                VALIDATE_ROW_DATA_DM_F16_IMPL(block_q2_K, data, nb, d, dmin);
+            } break;
+        case GGML_TYPE_Q3_K:
+            {
+                VALIDATE_ROW_DATA_D_F16_IMPL(block_q3_K, data, nb);
+            } break;
+        case GGML_TYPE_Q4_K:
+            {
+                VALIDATE_ROW_DATA_DM_F16_IMPL(block_q4_K, data, nb, d, dmin);
+            } break;
+        case GGML_TYPE_Q5_K:
+            {
+                VALIDATE_ROW_DATA_DM_F16_IMPL(block_q5_K, data, nb, d, dmin);
+            } break;
+        case GGML_TYPE_Q6_K:
+            {
+                VALIDATE_ROW_DATA_D_F16_IMPL(block_q6_K, data, nb);
+            } break;
+        case GGML_TYPE_Q8_K:
+            {
+                const block_q8_K * q = (const block_q8_K *) data;
+                for (size_t i = 0; i < nb; ++i) {
+                    if (!validate_float(q[i].d, i)) {
+                        return false;
+                    }
+                }
+            } break;
+        case GGML_TYPE_TQ1_0:
+            {
+                VALIDATE_ROW_DATA_D_F16_IMPL(block_tq1_0, data, nb);
+            } break;
+        case GGML_TYPE_TQ2_0:
+            {
+                VALIDATE_ROW_DATA_D_F16_IMPL(block_tq2_0, data, nb);
+            } break;
+        case GGML_TYPE_IQ1_S:
+            {
+                VALIDATE_ROW_DATA_D_F16_IMPL(block_iq1_s, data, nb);
+            } break;
+        case GGML_TYPE_IQ1_M:
+            {
+                const block_iq1_m * q = (const block_iq1_m *) data;
+                for (size_t i = 0; i < nb; ++i) {
+                    iq1m_scale_t scale;
+                    const uint16_t * sc = (const uint16_t *)q[i].scales;
+                    scale.u16 = (sc[0] >> 12) | ((sc[1] >> 8) & 0x00f0) | ((sc[2] >> 4) & 0x0f00) | (sc[3] & 0xf000);
+                    if (!validate_fp16(scale.f16, i)) {
+                        return false;
+                    }
+                }
+            } break;
+        case GGML_TYPE_IQ2_XXS:
+            {
+                VALIDATE_ROW_DATA_D_F16_IMPL(block_iq2_xxs, data, nb);
+            } break;
+        case GGML_TYPE_IQ2_XS:
+            {
+                VALIDATE_ROW_DATA_D_F16_IMPL(block_iq2_xs, data, nb);
+            } break;
+        case GGML_TYPE_IQ2_S:
+            {
+                VALIDATE_ROW_DATA_D_F16_IMPL(block_iq2_s, data, nb);
+            } break;
+        case GGML_TYPE_IQ3_XXS:
+            {
+                VALIDATE_ROW_DATA_D_F16_IMPL(block_iq3_xxs, data, nb);
+            } break;
+
+        case GGML_TYPE_IQ3_S:
+            {
+                VALIDATE_ROW_DATA_D_F16_IMPL(block_iq3_s, data, nb);
+            } break;
+        case GGML_TYPE_IQ4_XS:
+            {
+                VALIDATE_ROW_DATA_D_F16_IMPL(block_iq4_xs, data, nb);
+            } break;
+        case GGML_TYPE_IQ4_NL:
+            {
+                VALIDATE_ROW_DATA_D_F16_IMPL(block_iq4_nl, data, nb);
+            } break;
+        case GGML_TYPE_Q4_0_4_4:
+        case GGML_TYPE_Q4_0_4_8:
+            {
+                VALIDATE_ROW_DATA_DVEC_F16_IMPL(block_q4_0x4, data, nbytes / sizeof(block_q4_0x4), 4);
+            } break;
+        case GGML_TYPE_Q4_0_8_8:
+            {
+                VALIDATE_ROW_DATA_DVEC_F16_IMPL(block_q4_0x8, data, nbytes / sizeof(block_q4_0x8), 8);
+            } break;
+
+        case GGML_TYPE_I8:
+        case GGML_TYPE_I16:
+        case GGML_TYPE_I32:
+        case GGML_TYPE_I64:
+            // nothing to validate
+            break;
+        default:
+            {
+                fprintf(stderr, "%s: invalid type %d\n", __func__, type);
+                return false;
+            }
+    }
+
+    return true;
+}
diff --git a/src/whisper_cpp/ggml/src/ggml-quants.h b/src/whisper_cpp/ggml/src/ggml-quants.h
new file mode 100644
index 00000000..df9c4b24
--- /dev/null
+++ b/src/whisper_cpp/ggml/src/ggml-quants.h
@@ -0,0 +1,147 @@
+#pragma once
+
+#define GGML_COMMON_DECL_C
+#include "ggml-common.h"
+
+#include "ggml.h"
+
+// GGML internal header
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+// Quantization
+void quantize_row_q4_0_ref(const float * GGML_RESTRICT x, block_q4_0 * GGML_RESTRICT y, int64_t k);
+void quantize_row_q4_1_ref(const float * GGML_RESTRICT x, block_q4_1 * GGML_RESTRICT y, int64_t k);
+void quantize_row_q5_0_ref(const float * GGML_RESTRICT x, block_q5_0 * GGML_RESTRICT y, int64_t k);
+void quantize_row_q5_1_ref(const float * GGML_RESTRICT x, block_q5_1 * GGML_RESTRICT y, int64_t k);
+void quantize_row_q8_0_ref(const float * GGML_RESTRICT x, block_q8_0 * GGML_RESTRICT y, int64_t k);
+void quantize_row_q8_1_ref(const float * GGML_RESTRICT x, block_q8_1 * GGML_RESTRICT y, int64_t k);
+
+void quantize_row_q2_K_ref(const float * GGML_RESTRICT x, block_q2_K * GGML_RESTRICT y, int64_t k);
+void quantize_row_q3_K_ref(const float * GGML_RESTRICT x, block_q3_K * GGML_RESTRICT y, int64_t k);
+void quantize_row_q4_K_ref(const float * GGML_RESTRICT x, block_q4_K * GGML_RESTRICT y, int64_t k);
+void quantize_row_q5_K_ref(const float * GGML_RESTRICT x, block_q5_K * GGML_RESTRICT y, int64_t k);
+void quantize_row_q6_K_ref(const float * GGML_RESTRICT x, block_q6_K * GGML_RESTRICT y, int64_t k);
+void quantize_row_q8_K_ref(const float * GGML_RESTRICT x, block_q8_K * GGML_RESTRICT y, int64_t k);
+
+void quantize_row_tq1_0_ref(const float * GGML_RESTRICT x, block_tq1_0 * GGML_RESTRICT y, int64_t k);
+void quantize_row_tq2_0_ref(const float * GGML_RESTRICT x, block_tq2_0 * GGML_RESTRICT y, int64_t k);
+
+void quantize_row_iq3_xxs_ref(const float * GGML_RESTRICT x, block_iq3_xxs * GGML_RESTRICT y, int64_t k);
+void quantize_row_iq4_nl_ref (const float * GGML_RESTRICT x, block_iq4_nl  * GGML_RESTRICT y, int64_t k);
+void quantize_row_iq4_xs_ref (const float * GGML_RESTRICT x, block_iq4_xs  * GGML_RESTRICT y, int64_t k);
+void quantize_row_iq3_s_ref  (const float * GGML_RESTRICT x, block_iq3_s   * GGML_RESTRICT y, int64_t k);
+void quantize_row_iq2_s_ref  (const float * GGML_RESTRICT x, block_iq2_s   * GGML_RESTRICT y, int64_t k);
+
+void quantize_row_q4_0(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
+void quantize_row_q4_1(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
+void quantize_row_q5_0(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
+void quantize_row_q5_1(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
+void quantize_row_q8_0(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
+void quantize_row_q8_1(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
+
+void quantize_row_q2_K(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
+void quantize_row_q3_K(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
+void quantize_row_q4_K(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
+void quantize_row_q5_K(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
+void quantize_row_q6_K(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
+void quantize_row_q8_K(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
+
+void quantize_row_tq1_0(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
+void quantize_row_tq2_0(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
+
+void quantize_row_iq3_xxs(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
+void quantize_row_iq4_nl (const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
+void quantize_row_iq4_xs (const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
+void quantize_row_iq3_s  (const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
+void quantize_row_iq2_s  (const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
+
+// Dequantization
+void dequantize_row_q4_0(const block_q4_0 * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
+void dequantize_row_q4_1(const block_q4_1 * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
+void dequantize_row_q5_0(const block_q5_0 * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
+void dequantize_row_q5_1(const block_q5_1 * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
+void dequantize_row_q8_0(const block_q8_0 * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
+//void dequantize_row_q8_1(const block_q8_1 * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
+
+void dequantize_row_q2_K(const block_q2_K * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
+void dequantize_row_q3_K(const block_q3_K * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
+void dequantize_row_q4_K(const block_q4_K * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
+void dequantize_row_q5_K(const block_q5_K * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
+void dequantize_row_q6_K(const block_q6_K * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
+void dequantize_row_q8_K(const block_q8_K * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
+
+void dequantize_row_tq1_0(const block_tq1_0 * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
+void dequantize_row_tq2_0(const block_tq2_0 * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
+
+void dequantize_row_iq2_xxs(const block_iq2_xxs * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
+void dequantize_row_iq2_xs (const block_iq2_xs  * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
+void dequantize_row_iq2_s  (const block_iq2_s   * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
+void dequantize_row_iq3_xxs(const block_iq3_xxs * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
+void dequantize_row_iq1_s  (const block_iq1_s   * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
+void dequantize_row_iq1_m  (const block_iq1_m   * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
+void dequantize_row_iq4_nl (const block_iq4_nl  * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
+void dequantize_row_iq4_xs (const block_iq4_xs  * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
+void dequantize_row_iq3_s  (const block_iq3_s   * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
+
+// Dot product
+void ggml_vec_dot_q4_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
+void ggml_vec_dot_q4_1_q8_1(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
+void ggml_vec_dot_q5_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
+void ggml_vec_dot_q5_1_q8_1(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
+void ggml_vec_dot_q8_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
+
+void ggml_vec_dot_q2_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
+void ggml_vec_dot_q3_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
+void ggml_vec_dot_q4_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
+void ggml_vec_dot_q5_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
+void ggml_vec_dot_q6_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
+
+void ggml_vec_dot_tq1_0_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
+void ggml_vec_dot_tq2_0_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
+
+void ggml_vec_dot_iq2_xxs_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
+void ggml_vec_dot_iq2_xs_q8_K (int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
+void ggml_vec_dot_iq2_s_q8_K  (int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
+void ggml_vec_dot_iq3_xxs_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
+void ggml_vec_dot_iq1_s_q8_K  (int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
+void ggml_vec_dot_iq1_m_q8_K  (int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
+void ggml_vec_dot_iq4_nl_q8_0 (int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
+void ggml_vec_dot_iq4_xs_q8_K (int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
+void ggml_vec_dot_iq3_s_q8_K  (int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
+
+// Quantization utilizing an importance matrix (a.k.a. "Activation aWare Quantization")
+size_t quantize_iq2_xxs(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
+size_t quantize_iq2_xs (const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
+size_t quantize_iq2_s  (const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
+size_t quantize_iq3_xxs(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
+size_t quantize_iq1_s  (const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
+size_t quantize_iq1_m  (const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
+size_t quantize_iq4_nl (const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
+size_t quantize_iq4_xs (const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
+size_t quantize_iq3_s  (const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
+
+size_t quantize_tq1_0(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
+size_t quantize_tq2_0(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
+
+size_t quantize_q2_K(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
+size_t quantize_q3_K(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
+size_t quantize_q4_K(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
+size_t quantize_q5_K(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
+size_t quantize_q6_K(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
+size_t quantize_q4_0(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
+size_t quantize_q4_1(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
+size_t quantize_q5_0(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
+size_t quantize_q5_1(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
+size_t quantize_q8_0(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
+
+void iq2xs_init_impl(enum ggml_type type);
+void iq2xs_free_impl(enum ggml_type type);
+void iq3xs_init_impl(int grid_size);
+void iq3xs_free_impl(int grid_size);
+
+#ifdef __cplusplus
+}
+#endif
diff --git a/src/whisper_cpp/ggml/src/ggml-rpc.cpp b/src/whisper_cpp/ggml/src/ggml-rpc.cpp
new file mode 100644
index 00000000..ab7298cb
--- /dev/null
+++ b/src/whisper_cpp/ggml/src/ggml-rpc.cpp
@@ -0,0 +1,1228 @@
+#include "ggml-rpc.h"
+#include "ggml-impl.h"
+#include "ggml-backend-impl.h"
+
+#include <cinttypes>
+#include <string>
+#include <vector>
+#include <memory>
+#include <mutex>
+#include <unordered_map>
+#include <unordered_set>
+#ifdef _WIN32
+#  define WIN32_LEAN_AND_MEAN
+#  ifndef NOMINMAX
+#     define NOMINMAX
+#  endif
+#  include <windows.h>
+#  include <winsock2.h>
+#else
+#  include <arpa/inet.h>
+#  include <sys/socket.h>
+#  include <sys/types.h>
+#  include <netinet/in.h>
+#  include <netinet/tcp.h>
+#  include <netdb.h>
+#  include <unistd.h>
+#endif
+#include <string.h>
+
+#define UNUSED GGML_UNUSED
+
+#define GGML_DEBUG 0
+#if (GGML_DEBUG >= 1)
+#define GGML_PRINT_DEBUG(...) printf(__VA_ARGS__)
+#else
+#define GGML_PRINT_DEBUG(...)
+#endif
+
+#ifdef _WIN32
+typedef SOCKET sockfd_t;
+using ssize_t = __int64;
+#else
+typedef int sockfd_t;
+#endif
+
+// cross-platform socket
+struct socket_t {
+    sockfd_t fd;
+    socket_t(sockfd_t fd) : fd(fd) {}
+    ~socket_t() {
+        GGML_PRINT_DEBUG("[%s] closing socket %d\n", __func__, this->fd);
+#ifdef _WIN32
+        closesocket(this->fd);
+#else
+        close(this->fd);
+#endif
+    }
+};
+
+// ggml_tensor is serialized into rpc_tensor
+#pragma pack(push, 1)
+struct rpc_tensor {
+    uint64_t id;
+    uint32_t type;
+    uint64_t buffer;
+    uint32_t ne[GGML_MAX_DIMS];
+    uint32_t nb[GGML_MAX_DIMS];
+    uint32_t op;
+    int32_t  op_params[GGML_MAX_OP_PARAMS / sizeof(int32_t)];
+    int32_t  flags;
+    uint64_t src[GGML_MAX_SRC];
+    uint64_t view_src;
+    uint64_t view_offs;
+    uint64_t data;
+    char name[GGML_MAX_NAME];
+
+    char padding[4];
+};
+#pragma pack(pop)
+
+static_assert(sizeof(rpc_tensor) % 8 == 0, "rpc_tensor size must be multiple of 8");
+
+// RPC commands
+enum rpc_cmd {
+    RPC_CMD_ALLOC_BUFFER = 0,
+    RPC_CMD_GET_ALIGNMENT,
+    RPC_CMD_GET_MAX_SIZE,
+    RPC_CMD_BUFFER_GET_BASE,
+    RPC_CMD_FREE_BUFFER,
+    RPC_CMD_BUFFER_CLEAR,
+    RPC_CMD_SET_TENSOR,
+    RPC_CMD_GET_TENSOR,
+    RPC_CMD_COPY_TENSOR,
+    RPC_CMD_GRAPH_COMPUTE,
+    RPC_CMD_GET_DEVICE_MEMORY,
+    RPC_CMD_COUNT,
+};
+
+// RPC data structures
+
+static ggml_guid_t ggml_backend_rpc_guid() {
+    static ggml_guid guid = {0x99, 0x68, 0x5b, 0x6c, 0xd2, 0x83, 0x3d, 0x24, 0x25, 0x36, 0x72, 0xe1, 0x5b, 0x0e, 0x14, 0x03};
+    return &guid;
+}
+
+struct ggml_backend_rpc_buffer_type_context {
+    std::string endpoint;
+    std::string name;
+    size_t alignment;
+    size_t max_size;
+};
+
+struct ggml_backend_rpc_context {
+    std::string endpoint;
+    std::string name;
+};
+
+struct ggml_backend_rpc_buffer_context {
+    std::shared_ptr<socket_t> sock;
+    std::unordered_map<ggml_backend_buffer_t, void *> base_cache;
+    uint64_t remote_ptr;
+    std::string name;
+};
+
+// RPC helper functions
+
+static std::shared_ptr<socket_t> make_socket(sockfd_t fd) {
+#ifdef _WIN32
+    if (fd == INVALID_SOCKET) {
+        return nullptr;
+    }
+#else
+    if (fd < 0) {
+        return nullptr;
+    }
+#endif
+    return std::make_shared<socket_t>(fd);
+}
+
+static bool set_no_delay(sockfd_t sockfd) {
+    int flag = 1;
+    // set TCP_NODELAY to disable Nagle's algorithm
+    int ret = setsockopt(sockfd, IPPROTO_TCP, TCP_NODELAY, (char *)&flag, sizeof(int));
+    return ret == 0;
+}
+
+static bool set_reuse_addr(sockfd_t sockfd) {
+    int flag = 1;
+    int ret = setsockopt(sockfd, SOL_SOCKET, SO_REUSEADDR, (char *)&flag, sizeof(int));
+    return ret == 0;
+}
+
+static std::shared_ptr<socket_t> socket_connect(const char * host, int port) {
+    struct sockaddr_in addr;
+    auto sockfd = socket(AF_INET, SOCK_STREAM, 0);
+    auto sock_ptr = make_socket(sockfd);
+    if (sock_ptr == nullptr) {
+        return nullptr;
+    }
+    if (!set_no_delay(sockfd)) {
+        fprintf(stderr, "Failed to set TCP_NODELAY\n");
+        return nullptr;
+    }
+    addr.sin_family = AF_INET;
+    addr.sin_port = htons(port);
+    struct hostent * server = gethostbyname(host);
+    if (server == NULL) {
+        fprintf(stderr, "Cannot resolve host '%s'\n", host);
+        return nullptr;
+    }
+    memcpy(&addr.sin_addr.s_addr, server->h_addr, server->h_length);
+    if (connect(sock_ptr->fd, (struct sockaddr *)&addr, sizeof(addr)) < 0) {
+        return nullptr;
+    }
+    return sock_ptr;
+}
+
+static std::shared_ptr<socket_t> socket_accept(sockfd_t srv_sockfd) {
+    auto client_socket_fd = accept(srv_sockfd, NULL, NULL);
+    auto client_socket = make_socket(client_socket_fd);
+    if (client_socket == nullptr) {
+        return nullptr;
+    }
+    if (!set_no_delay(client_socket_fd)) {
+        fprintf(stderr, "Failed to set TCP_NODELAY\n");
+        return nullptr;
+    }
+    return client_socket;
+}
+
+static std::shared_ptr<socket_t> create_server_socket(const char * host, int port) {
+    auto sockfd = socket(AF_INET, SOCK_STREAM, 0);
+    auto sock = make_socket(sockfd);
+    if (sock == nullptr) {
+        return nullptr;
+    }
+    if (!set_reuse_addr(sockfd)) {
+        fprintf(stderr, "Failed to set SO_REUSEADDR\n");
+        return nullptr;
+    }
+    if (inet_addr(host) == INADDR_NONE) {
+        fprintf(stderr, "Invalid host address: %s\n", host);
+        return nullptr;
+    }
+    struct sockaddr_in serv_addr;
+    serv_addr.sin_family = AF_INET;
+    serv_addr.sin_addr.s_addr = inet_addr(host);
+    serv_addr.sin_port = htons(port);
+
+    if (bind(sockfd, (struct sockaddr *) &serv_addr, sizeof(serv_addr)) < 0) {
+        return nullptr;
+    }
+    if (listen(sockfd, 1) < 0) {
+        return nullptr;
+    }
+    return sock;
+}
+
+static bool send_data(sockfd_t sockfd, const void * data, size_t size) {
+    size_t bytes_sent = 0;
+    while (bytes_sent < size) {
+        ssize_t n = send(sockfd, (const char *)data + bytes_sent, size - bytes_sent, 0);
+        if (n < 0) {
+            return false;
+        }
+        bytes_sent += n;
+    }
+    return true;
+}
+
+static bool recv_data(sockfd_t sockfd, void * data, size_t size) {
+    size_t bytes_recv = 0;
+    while (bytes_recv < size) {
+        ssize_t n = recv(sockfd, (char *)data + bytes_recv, size - bytes_recv, 0);
+        if (n <= 0) {
+            return false;
+        }
+        bytes_recv += n;
+    }
+    return true;
+}
+
+static bool parse_endpoint(const std::string & endpoint, std::string & host, int & port) {
+    size_t pos = endpoint.find(':');
+    if (pos == std::string::npos) {
+        return false;
+    }
+    host = endpoint.substr(0, pos);
+    port = std::stoi(endpoint.substr(pos + 1));
+    return true;
+}
+
+// RPC request : | rpc_cmd (1 byte) | request_size (8 bytes) | request_data (request_size bytes) |
+// RPC response: | response_size (8 bytes) | response_data (response_size bytes) |
+static bool send_rpc_cmd(const std::shared_ptr<socket_t> & sock, enum rpc_cmd cmd, const std::vector<uint8_t> & input, std::vector<uint8_t> & output) {
+    uint8_t cmd_byte = cmd;
+    if (!send_data(sock->fd, &cmd_byte, sizeof(cmd_byte))) {
+        return false;
+    }
+    uint64_t input_size = input.size();
+    if (!send_data(sock->fd, &input_size, sizeof(input_size))) {
+        return false;
+    }
+    if (!send_data(sock->fd, input.data(), input.size())) {
+        return false;
+    }
+    uint64_t output_size;
+    if (!recv_data(sock->fd, &output_size, sizeof(output_size))) {
+        return false;
+    }
+    if (output_size == 0) {
+        output.clear();
+        return true;
+    }
+    output.resize(output_size);
+    if (!recv_data(sock->fd, output.data(), output_size)) {
+        return false;
+    }
+    return true;
+}
+
+// RPC client-side implementation
+
+static std::shared_ptr<socket_t> get_socket(const std::string & endpoint) {
+    static std::mutex mutex;
+    std::lock_guard<std::mutex> lock(mutex);
+    static std::unordered_map<std::string, std::weak_ptr<socket_t>> sockets;
+    static bool initialized = false;
+
+    auto it = sockets.find(endpoint);
+    if (it != sockets.end()) {
+        if (auto sock = it->second.lock()) {
+            return sock;
+        }
+    }
+    std::string host;
+    int port;
+    if (!parse_endpoint(endpoint, host, port)) {
+        return nullptr;
+    }
+#ifdef _WIN32
+    if (!initialized) {
+        WSADATA wsaData;
+        int res = WSAStartup(MAKEWORD(2, 2), &wsaData);
+        if (res != 0) {
+            return nullptr;
+        }
+        initialized = true;
+    }
+#else
+    UNUSED(initialized);
+#endif
+    auto sock = socket_connect(host.c_str(), port);
+    if (sock == nullptr) {
+        return nullptr;
+    }
+    GGML_PRINT_DEBUG("[%s] connected to %s, sockfd=%d\n", __func__, endpoint.c_str(), sock->fd);
+    sockets[endpoint] = sock;
+    return sock;
+}
+
+static const char * ggml_backend_rpc_buffer_get_name(ggml_backend_buffer_t buffer) {
+    ggml_backend_rpc_buffer_context * ctx = (ggml_backend_rpc_buffer_context *)buffer->context;
+    return ctx->name.c_str();
+}
+
+static void ggml_backend_rpc_buffer_free_buffer(ggml_backend_buffer_t buffer) {
+    ggml_backend_rpc_buffer_context * ctx = (ggml_backend_rpc_buffer_context *)buffer->context;
+    // input serialization format: | remote_ptr (8 bytes) |
+    std::vector<uint8_t> input(sizeof(uint64_t), 0);
+    uint64_t remote_ptr = ctx->remote_ptr;
+    memcpy(input.data(), &remote_ptr, sizeof(remote_ptr));
+    std::vector<uint8_t> output;
+    bool status = send_rpc_cmd(ctx->sock, RPC_CMD_FREE_BUFFER, input, output);
+    GGML_ASSERT(status);
+    GGML_ASSERT(output.empty());
+    delete ctx;
+}
+
+static void * ggml_backend_rpc_buffer_get_base(ggml_backend_buffer_t buffer) {
+    ggml_backend_rpc_buffer_context * ctx = (ggml_backend_rpc_buffer_context *)buffer->context;
+    if (ctx->base_cache.find(buffer) != ctx->base_cache.end()) {
+        return ctx->base_cache[buffer];
+    }
+    // input serialization format: | remote_ptr (8 bytes) |
+    std::vector<uint8_t> input(sizeof(uint64_t), 0);
+    uint64_t remote_ptr = ctx->remote_ptr;
+    memcpy(input.data(), &remote_ptr, sizeof(remote_ptr));
+    std::vector<uint8_t> output;
+    bool status = send_rpc_cmd(ctx->sock, RPC_CMD_BUFFER_GET_BASE, input, output);
+    GGML_ASSERT(status);
+    GGML_ASSERT(output.size() == sizeof(uint64_t));
+    // output serialization format: | base_ptr (8 bytes) |
+    uint64_t base_ptr;
+    memcpy(&base_ptr, output.data(), sizeof(base_ptr));
+    void * base = reinterpret_cast<void *>(base_ptr);
+    ctx->base_cache[buffer] = base;
+    return base;
+}
+
+static rpc_tensor serialize_tensor(const ggml_tensor * tensor) {
+    rpc_tensor result;
+    result.id = reinterpret_cast<uint64_t>(tensor);
+    result.type = tensor->type;
+    if (tensor->buffer) {
+        ggml_backend_buffer_t buffer = tensor->buffer;
+        ggml_backend_rpc_buffer_context * ctx = (ggml_backend_rpc_buffer_context *)buffer->context;
+        result.buffer = ctx->remote_ptr;
+    } else {
+        result.buffer = 0;
+    }
+    for (uint32_t i = 0; i < GGML_MAX_DIMS; i++) {
+        result.ne[i] = tensor->ne[i];
+        result.nb[i] = tensor->nb[i];
+    }
+    result.op = tensor->op;
+    for (uint32_t i = 0; i < GGML_MAX_OP_PARAMS / sizeof(int32_t); i++) {
+        result.op_params[i] = tensor->op_params[i];
+    }
+    result.flags = tensor->flags;
+    for (uint32_t i = 0; i < GGML_MAX_SRC; i++) {
+        result.src[i] = reinterpret_cast<uint64_t>(tensor->src[i]);
+    }
+    result.view_src = reinterpret_cast<uint64_t>(tensor->view_src);
+    result.view_offs = tensor->view_offs;
+    result.data = reinterpret_cast<uint64_t>(tensor->data);
+    snprintf(result.name, GGML_MAX_NAME, "%s", tensor->name);
+    return result;
+}
+
+static void ggml_backend_rpc_buffer_init_tensor(ggml_backend_buffer_t buffer, ggml_tensor * tensor) {
+    UNUSED(buffer);
+    if (ggml_is_quantized(tensor->type)) {
+        // TODO: this check is due to MATRIX_ROW_PADDING in CUDA and should be generalized
+        GGML_ASSERT(tensor->ne[0] % 512 == 0 && "unsupported quantized tensor");
+    }
+}
+
+static void ggml_backend_rpc_buffer_set_tensor(ggml_backend_buffer_t buffer, ggml_tensor * tensor, const void * data, size_t offset, size_t size) {
+    ggml_backend_rpc_buffer_context * ctx = (ggml_backend_rpc_buffer_context *)buffer->context;
+    // input serialization format: | rpc_tensor | offset (8 bytes) | data (size bytes) |
+    size_t input_size = sizeof(rpc_tensor) + sizeof(uint64_t) + size;
+    std::vector<uint8_t> input(input_size, 0);
+    rpc_tensor rpc_tensor = serialize_tensor(tensor);
+    memcpy(input.data(), &rpc_tensor, sizeof(rpc_tensor));
+    memcpy(input.data() + sizeof(rpc_tensor), &offset, sizeof(offset));
+    memcpy(input.data() + sizeof(rpc_tensor) + sizeof(offset), data, size);
+    std::vector<uint8_t> output;
+    bool status = send_rpc_cmd(ctx->sock, RPC_CMD_SET_TENSOR, input, output);
+    GGML_ASSERT(status);
+}
+
+static void ggml_backend_rpc_buffer_get_tensor(ggml_backend_buffer_t buffer, const ggml_tensor * tensor, void * data, size_t offset, size_t size) {
+    ggml_backend_rpc_buffer_context * ctx = (ggml_backend_rpc_buffer_context *)buffer->context;
+    // input serialization format: | rpc_tensor | offset (8 bytes) | size (8 bytes) |
+    int input_size = sizeof(rpc_tensor) + 2*sizeof(uint64_t);
+    std::vector<uint8_t> input(input_size, 0);
+    rpc_tensor rpc_tensor = serialize_tensor(tensor);
+    memcpy(input.data(), &rpc_tensor, sizeof(rpc_tensor));
+    memcpy(input.data() + sizeof(rpc_tensor), &offset, sizeof(offset));
+    memcpy(input.data() + sizeof(rpc_tensor) + sizeof(offset), &size, sizeof(size));
+    std::vector<uint8_t> output;
+    bool status = send_rpc_cmd(ctx->sock, RPC_CMD_GET_TENSOR, input, output);
+    GGML_ASSERT(status);
+    GGML_ASSERT(output.size() == size);
+    // output serialization format: | data (size bytes) |
+    memcpy(data, output.data(), size);
+}
+
+static bool ggml_backend_rpc_buffer_cpy_tensor(ggml_backend_buffer_t buffer, const ggml_tensor * src, ggml_tensor * dst) {
+    // check if src and dst are on the same server
+    ggml_backend_buffer_t src_buffer = src->buffer;
+    ggml_backend_rpc_buffer_context * src_ctx = (ggml_backend_rpc_buffer_context *)src_buffer->context;
+    ggml_backend_buffer_t dst_buffer = dst->buffer;
+    ggml_backend_rpc_buffer_context * dst_ctx = (ggml_backend_rpc_buffer_context *)dst_buffer->context;
+    if (src_ctx->sock != dst_ctx->sock) {
+        return false;
+    }
+    ggml_backend_rpc_buffer_context * ctx = (ggml_backend_rpc_buffer_context *)buffer->context;
+    // input serialization format: | rpc_tensor src | rpc_tensor dst |
+    int input_size = 2*sizeof(rpc_tensor);
+    std::vector<uint8_t> input(input_size, 0);
+    rpc_tensor rpc_src = serialize_tensor(src);
+    rpc_tensor rpc_dst = serialize_tensor(dst);
+    memcpy(input.data(), &rpc_src, sizeof(rpc_src));
+    memcpy(input.data() + sizeof(rpc_src), &rpc_dst, sizeof(rpc_dst));
+    std::vector<uint8_t> output;
+    bool status = send_rpc_cmd(ctx->sock, RPC_CMD_COPY_TENSOR, input, output);
+    GGML_ASSERT(status);
+    // output serialization format: | result (1 byte) |
+    GGML_ASSERT(output.size() == 1);
+    return output[0];
+}
+
+static void ggml_backend_rpc_buffer_clear(ggml_backend_buffer_t buffer, uint8_t value) {
+    ggml_backend_rpc_buffer_context * ctx = (ggml_backend_rpc_buffer_context *)buffer->context;
+    // serialization format: | bufptr (8 bytes) | value (1 byte) |
+    int input_size = sizeof(uint64_t) + sizeof(uint8_t);
+    std::vector<uint8_t> input(input_size, 0);
+    memcpy(input.data(), &ctx->remote_ptr, sizeof(ctx->remote_ptr));
+    memcpy(input.data() + sizeof(ctx->remote_ptr), &value, sizeof(value));
+    std::vector<uint8_t> output;
+    bool status = send_rpc_cmd(ctx->sock, RPC_CMD_BUFFER_CLEAR, input, output);
+    GGML_ASSERT(status);
+}
+
+static ggml_backend_buffer_i ggml_backend_rpc_buffer_interface = {
+    /* .get_name        = */ ggml_backend_rpc_buffer_get_name,
+    /* .free_buffer     = */ ggml_backend_rpc_buffer_free_buffer,
+    /* .get_base        = */ ggml_backend_rpc_buffer_get_base,
+    /* .init_tensor     = */ ggml_backend_rpc_buffer_init_tensor,
+    /* .memset_tensor   = */ NULL,
+    /* .set_tensor      = */ ggml_backend_rpc_buffer_set_tensor,
+    /* .get_tensor      = */ ggml_backend_rpc_buffer_get_tensor,
+    /* .cpy_tensor      = */ ggml_backend_rpc_buffer_cpy_tensor,
+    /* .clear           = */ ggml_backend_rpc_buffer_clear,
+    /* .reset           = */ NULL,
+};
+
+static const char * ggml_backend_rpc_buffer_type_name(ggml_backend_buffer_type_t buft) {
+    ggml_backend_rpc_buffer_type_context * buft_ctx = (ggml_backend_rpc_buffer_type_context *)buft->context;
+    return buft_ctx->name.c_str();
+}
+
+static ggml_backend_buffer_t ggml_backend_rpc_buffer_type_alloc_buffer(ggml_backend_buffer_type_t buft, size_t size) {
+    ggml_backend_rpc_buffer_type_context * buft_ctx = (ggml_backend_rpc_buffer_type_context *)buft->context;
+    // input serialization format: | size (8 bytes) |
+    int input_size = sizeof(uint64_t);
+    std::vector<uint8_t> input(input_size, 0);
+    memcpy(input.data(), &size, sizeof(size));
+    std::vector<uint8_t> output;
+    auto sock = get_socket(buft_ctx->endpoint);
+    bool status = send_rpc_cmd(sock, RPC_CMD_ALLOC_BUFFER, input, output);
+    GGML_ASSERT(status);
+    GGML_ASSERT(output.size() == 2*sizeof(uint64_t));
+    // output serialization format: | remote_ptr (8 bytes) | remote_size (8 bytes) |
+    uint64_t remote_ptr;
+    memcpy(&remote_ptr, output.data(), sizeof(remote_ptr));
+    size_t remote_size;
+    memcpy(&remote_size, output.data() + sizeof(uint64_t), sizeof(remote_size));
+    if (remote_ptr != 0) {
+        ggml_backend_buffer_t buffer = ggml_backend_buffer_init(buft,
+            ggml_backend_rpc_buffer_interface,
+            new ggml_backend_rpc_buffer_context{sock, {}, remote_ptr, "RPC[" + std::string(buft_ctx->endpoint) + "]"},
+            remote_size);
+        return buffer;
+    } else {
+        return nullptr;
+    }
+}
+
+static size_t get_alignment(const std::shared_ptr<socket_t> & sock) {
+    // input serialization format: | 0 bytes |
+    std::vector<uint8_t> input;
+    std::vector<uint8_t> output;
+    bool status = send_rpc_cmd(sock, RPC_CMD_GET_ALIGNMENT, input, output);
+    GGML_ASSERT(status);
+    GGML_ASSERT(output.size() == sizeof(uint64_t));
+    // output serialization format: | alignment (8 bytes) |
+    uint64_t alignment;
+    memcpy(&alignment, output.data(), sizeof(alignment));
+    return alignment;
+}
+
+static size_t ggml_backend_rpc_buffer_type_get_alignment(ggml_backend_buffer_type_t buft) {
+    ggml_backend_rpc_buffer_type_context * buft_ctx = (ggml_backend_rpc_buffer_type_context *)buft->context;
+    return buft_ctx->alignment;
+}
+
+static size_t get_max_size(const std::shared_ptr<socket_t> & sock) {
+    // input serialization format: | 0 bytes |
+    std::vector<uint8_t> input;
+    std::vector<uint8_t> output;
+    bool status = send_rpc_cmd(sock, RPC_CMD_GET_MAX_SIZE, input, output);
+    GGML_ASSERT(status);
+    GGML_ASSERT(output.size() == sizeof(uint64_t));
+    // output serialization format: | max_size (8 bytes) |
+    uint64_t max_size;
+    memcpy(&max_size, output.data(), sizeof(max_size));
+    return max_size;
+}
+
+static size_t ggml_backend_rpc_get_max_size(ggml_backend_buffer_type_t buft) {
+    ggml_backend_rpc_buffer_type_context * buft_ctx = (ggml_backend_rpc_buffer_type_context *)buft->context;
+    return buft_ctx->max_size;
+}
+
+static size_t ggml_backend_rpc_buffer_type_get_alloc_size(ggml_backend_buffer_type_t buft, const ggml_tensor * tensor) {
+    UNUSED(buft);
+    return ggml_nbytes(tensor);
+}
+
+static ggml_backend_buffer_type_i ggml_backend_rpc_buffer_type_interface = {
+    /* .get_name         = */ ggml_backend_rpc_buffer_type_name,
+    /* .alloc_buffer     = */ ggml_backend_rpc_buffer_type_alloc_buffer,
+    /* .get_alignment    = */ ggml_backend_rpc_buffer_type_get_alignment,
+    /* .get_max_size     = */ ggml_backend_rpc_get_max_size,
+    /* .get_alloc_size   = */ ggml_backend_rpc_buffer_type_get_alloc_size,
+    /* .is_host          = */ NULL,
+};
+
+static const char * ggml_backend_rpc_name(ggml_backend_t backend) {
+    ggml_backend_rpc_context * rpc_ctx = (ggml_backend_rpc_context *)backend->context;
+
+    return rpc_ctx->name.c_str();
+}
+
+static void ggml_backend_rpc_free(ggml_backend_t backend) {
+    ggml_backend_rpc_context * rpc_ctx = (ggml_backend_rpc_context *)backend->context;
+    delete rpc_ctx;
+    delete backend;
+}
+
+static ggml_backend_buffer_type_t ggml_backend_rpc_get_default_buffer_type(ggml_backend_t backend) {
+    ggml_backend_rpc_context * ctx = (ggml_backend_rpc_context *)backend->context;
+    return ggml_backend_rpc_buffer_type(ctx->endpoint.c_str());
+}
+
+static void ggml_backend_rpc_synchronize(ggml_backend_t backend) {
+    UNUSED(backend);
+    // this is no-op because we don't have any async operations
+}
+
+static void add_tensor(ggml_tensor * tensor, std::vector<rpc_tensor> & tensors, std::unordered_set<ggml_tensor*> & visited) {
+    if (tensor == nullptr) {
+        return;
+    }
+    if (visited.find(tensor) != visited.end()) {
+        return;
+    }
+    visited.insert(tensor);
+    for (int i = 0; i < GGML_MAX_SRC; i++) {
+        add_tensor(tensor->src[i], tensors, visited);
+    }
+    add_tensor(tensor->view_src, tensors, visited);
+    tensors.push_back(serialize_tensor(tensor));
+}
+
+static void serialize_graph(const ggml_cgraph * cgraph, std::vector<uint8_t> & output) {
+    uint32_t n_nodes = cgraph->n_nodes;
+    std::vector<rpc_tensor> tensors;
+    std::unordered_set<ggml_tensor*> visited;
+    for (uint32_t i = 0; i < n_nodes; i++) {
+        add_tensor(cgraph->nodes[i], tensors, visited);
+    }
+    // serialization format:
+    // | n_nodes (4 bytes) | nodes (n_nodes * sizeof(uint64_t) | n_tensors (4 bytes) | tensors (n_tensors * sizeof(rpc_tensor)) |
+    uint32_t n_tensors = tensors.size();
+    int output_size = sizeof(uint32_t) + n_nodes * sizeof(uint64_t) + sizeof(uint32_t) + n_tensors * sizeof(rpc_tensor);
+    output.resize(output_size, 0);
+    memcpy(output.data(), &n_nodes, sizeof(n_nodes));
+    for (uint32_t i = 0; i < n_nodes; i++) {
+        memcpy(output.data() + sizeof(n_nodes) + i * sizeof(uint64_t), &cgraph->nodes[i], sizeof(uint64_t));
+    }
+    uint32_t * out_ntensors = (uint32_t *)(output.data() + sizeof(n_nodes) + n_nodes * sizeof(uint64_t));
+    *out_ntensors = n_tensors;
+    rpc_tensor * out_tensors = (rpc_tensor *)(output.data() + sizeof(n_nodes) + n_nodes * sizeof(uint64_t) + sizeof(uint32_t));
+    memcpy(out_tensors, tensors.data(), n_tensors * sizeof(rpc_tensor));
+}
+
+static enum ggml_status ggml_backend_rpc_graph_compute(ggml_backend_t backend, ggml_cgraph * cgraph) {
+    ggml_backend_rpc_context * rpc_ctx = (ggml_backend_rpc_context *)backend->context;
+    std::vector<uint8_t> input;
+    serialize_graph(cgraph, input);
+    std::vector<uint8_t> output;
+    auto sock = get_socket(rpc_ctx->endpoint);
+    bool status = send_rpc_cmd(sock, RPC_CMD_GRAPH_COMPUTE, input, output);
+    GGML_ASSERT(status);
+    GGML_ASSERT(output.size() == 1);
+    return (enum ggml_status)output[0];
+}
+
+static bool ggml_backend_rpc_supports_op(ggml_backend_t backend, const ggml_tensor * op) {
+    UNUSED(backend);
+    UNUSED(op);
+    //TODO: call the remote backend and cache the results
+    return true;
+}
+
+static bool ggml_backend_rpc_supports_buft(ggml_backend_t backend, ggml_backend_buffer_type_t buft) {
+    if (!buft || buft->iface.get_name != ggml_backend_rpc_buffer_type_name) {
+        return false;
+    }
+    ggml_backend_rpc_buffer_type_context * buft_ctx = (ggml_backend_rpc_buffer_type_context *)buft->context;
+    ggml_backend_rpc_context * rpc_ctx = (ggml_backend_rpc_context *)backend->context;
+    return buft_ctx->endpoint == rpc_ctx->endpoint;
+}
+
+static ggml_backend_i ggml_backend_rpc_interface = {
+    /* .get_name                = */ ggml_backend_rpc_name,
+    /* .free                    = */ ggml_backend_rpc_free,
+    /* .get_default_buffer_type = */ ggml_backend_rpc_get_default_buffer_type,
+    /* .set_tensor_async        = */ NULL,
+    /* .get_tensor_async        = */ NULL,
+    /* .cpy_tensor_async        = */ NULL,
+    /* .synchronize             = */ ggml_backend_rpc_synchronize,
+    /* .graph_plan_create       = */ NULL,
+    /* .graph_plan_free         = */ NULL,
+    /* .graph_plan_update       = */ NULL,
+    /* .graph_plan_compute      = */ NULL,
+    /* .graph_compute           = */ ggml_backend_rpc_graph_compute,
+    /* .supports_op             = */ ggml_backend_rpc_supports_op,
+    /* .supports_buft           = */ ggml_backend_rpc_supports_buft,
+    /* .offload_op              = */ NULL,
+    /* .event_record            = */ NULL,
+    /* .event_wait              = */ NULL,
+};
+
+GGML_API ggml_backend_buffer_type_t ggml_backend_rpc_buffer_type(const char * endpoint) {
+    static std::mutex mutex;
+    std::lock_guard<std::mutex> lock(mutex);
+    // NOTE: buffer types are allocated and never freed; this is by design
+    static std::unordered_map<std::string, ggml_backend_buffer_type_t> buft_map;
+    auto it = buft_map.find(endpoint);
+    if (it != buft_map.end()) {
+        return it->second;
+    }
+    auto sock = get_socket(endpoint);
+    if (sock == nullptr) {
+        fprintf(stderr, "Failed to connect to %s\n", endpoint);
+        return nullptr;
+    }
+    size_t alignment = get_alignment(sock);
+    size_t max_size = get_max_size(sock);
+    ggml_backend_rpc_buffer_type_context * buft_ctx = new ggml_backend_rpc_buffer_type_context {
+        /* .endpoint  = */ endpoint,
+        /* .name      = */ "RPC[" + std::string(endpoint) + "]",
+        /* .alignment = */ alignment,
+        /* .max_size  = */ max_size
+    };
+
+    ggml_backend_buffer_type_t buft = new ggml_backend_buffer_type {
+        /* .iface   = */ ggml_backend_rpc_buffer_type_interface,
+        /* .device  = */ nullptr,
+        /* .context = */ buft_ctx
+    };
+    buft_map[endpoint] = buft;
+    return buft;
+}
+
+ggml_backend_t ggml_backend_rpc_init(const char * endpoint) {
+    ggml_backend_rpc_context * ctx = new ggml_backend_rpc_context {
+        /* .endpoint  = */ endpoint,
+        /* .name      = */ "RPC[" + std::string(endpoint) + "]",
+    };
+
+    ggml_backend_t backend = new ggml_backend {
+        /* .guid      = */ ggml_backend_rpc_guid(),
+        /* .interface = */ ggml_backend_rpc_interface,
+        /* .device    = */ nullptr,
+        /* .context   = */ ctx
+    };
+    return backend;
+}
+
+GGML_API bool ggml_backend_is_rpc(ggml_backend_t backend) {
+    return backend != NULL && ggml_guid_matches(backend->guid, ggml_backend_rpc_guid());
+}
+
+static void get_device_memory(const std::shared_ptr<socket_t> & sock, size_t * free, size_t * total) {
+    // input serialization format: | 0 bytes |
+    std::vector<uint8_t> input;
+    std::vector<uint8_t> output;
+    bool status = send_rpc_cmd(sock, RPC_CMD_GET_DEVICE_MEMORY, input, output);
+    GGML_ASSERT(status);
+    GGML_ASSERT(output.size() == 2*sizeof(uint64_t));
+    // output serialization format: | free (8 bytes) | total (8 bytes) |
+    uint64_t free_mem;
+    memcpy(&free_mem, output.data(), sizeof(free_mem));
+    uint64_t total_mem;
+    memcpy(&total_mem, output.data() + sizeof(uint64_t), sizeof(total_mem));
+    *free = free_mem;
+    *total = total_mem;
+}
+
+GGML_API void ggml_backend_rpc_get_device_memory(const char * endpoint, size_t * free, size_t * total) {
+    auto sock = get_socket(endpoint);
+    if (sock == nullptr) {
+        *free = 0;
+        *total = 0;
+        return;
+    }
+    get_device_memory(sock, free, total);
+}
+
+// RPC server-side implementation
+
+class rpc_server {
+public:
+    rpc_server(ggml_backend_t backend) : backend(backend) {}
+    ~rpc_server();
+
+    bool alloc_buffer(const std::vector<uint8_t> & input, std::vector<uint8_t> & output);
+    void get_alignment(std::vector<uint8_t> & output);
+    void get_max_size(std::vector<uint8_t> & output);
+    bool buffer_get_base(const std::vector<uint8_t> & input, std::vector<uint8_t> & output);
+    bool free_buffer(const std::vector<uint8_t> & input);
+    bool buffer_clear(const std::vector<uint8_t> & input);
+    bool set_tensor(const std::vector<uint8_t> & input);
+    bool get_tensor(const std::vector<uint8_t> & input, std::vector<uint8_t> & output);
+    bool copy_tensor(const std::vector<uint8_t> & input, std::vector<uint8_t> & output);
+    bool graph_compute(const std::vector<uint8_t> & input, std::vector<uint8_t> & output);
+
+private:
+    ggml_tensor * deserialize_tensor(struct ggml_context * ctx, const rpc_tensor * tensor);
+    ggml_tensor * create_node(uint64_t id,
+                              struct ggml_context * ctx,
+                              const std::unordered_map<uint64_t, const rpc_tensor*> & tensor_ptrs,
+                              std::unordered_map<uint64_t, struct ggml_tensor*> & tensor_map);
+
+
+    ggml_backend_t backend;
+    std::unordered_set<ggml_backend_buffer_t> buffers;
+};
+
+bool rpc_server::alloc_buffer(const std::vector<uint8_t> & input, std::vector<uint8_t> & output) {
+    // input serialization format: | size (8 bytes) |
+    if (input.size() != sizeof(uint64_t)) {
+        return false;
+    }
+    uint64_t size;
+    memcpy(&size, input.data(), sizeof(size));
+    ggml_backend_buffer_type_t buft = ggml_backend_get_default_buffer_type(backend);
+    ggml_backend_buffer_t buffer = ggml_backend_buft_alloc_buffer(buft, size);
+    uint64_t remote_ptr = 0;
+    uint64_t remote_size = 0;
+    if (buffer != nullptr) {
+        remote_ptr = reinterpret_cast<uint64_t>(buffer);
+        remote_size = buffer->size;
+        GGML_PRINT_DEBUG("[%s] size: %" PRIu64 " -> remote_ptr: %" PRIx64 ", remote_size: %" PRIu64 "\n", __func__, size, remote_ptr, remote_size);
+        buffers.insert(buffer);
+    } else {
+        GGML_PRINT_DEBUG("[%s] size: %" PRIu64 " -> failed\n", __func__, size);
+    }
+    // output serialization format: | remote_ptr (8 bytes) | remote_size (8 bytes) |
+    output.resize(2*sizeof(uint64_t), 0);
+    memcpy(output.data(), &remote_ptr, sizeof(remote_ptr));
+    memcpy(output.data() + sizeof(uint64_t), &remote_size, sizeof(remote_size));
+    return true;
+}
+
+void rpc_server::get_alignment(std::vector<uint8_t> & output) {
+    ggml_backend_buffer_type_t buft = ggml_backend_get_default_buffer_type(backend);
+    size_t alignment = ggml_backend_buft_get_alignment(buft);
+    GGML_PRINT_DEBUG("[%s] alignment: %lu\n", __func__, alignment);
+    // output serialization format: | alignment (8 bytes) |
+    output.resize(sizeof(uint64_t), 0);
+    memcpy(output.data(), &alignment, sizeof(alignment));
+}
+
+void rpc_server::get_max_size(std::vector<uint8_t> & output) {
+    ggml_backend_buffer_type_t buft = ggml_backend_get_default_buffer_type(backend);
+    size_t max_size = ggml_backend_buft_get_max_size(buft);
+    GGML_PRINT_DEBUG("[%s] max_size: %lu\n", __func__, max_size);
+    // output serialization format: | max_size (8 bytes) |
+    output.resize(sizeof(uint64_t), 0);
+    memcpy(output.data(), &max_size, sizeof(max_size));
+}
+
+bool rpc_server::buffer_get_base(const std::vector<uint8_t> & input, std::vector<uint8_t> & output) {
+    // input serialization format: | remote_ptr (8 bytes) |
+    if (input.size() != sizeof(uint64_t)) {
+        return false;
+    }
+    uint64_t remote_ptr;
+    memcpy(&remote_ptr, input.data(), sizeof(remote_ptr));
+    GGML_PRINT_DEBUG("[%s] remote_ptr: %" PRIx64 "\n", __func__, remote_ptr);
+    ggml_backend_buffer_t buffer = reinterpret_cast<ggml_backend_buffer_t>(remote_ptr);
+    if (buffers.find(buffer) == buffers.end()) {
+        GGML_PRINT_DEBUG("[%s] buffer not found\n", __func__);
+        return false;
+    }
+    void * base = ggml_backend_buffer_get_base(buffer);
+    // output serialization format: | base_ptr (8 bytes) |
+    uint64_t base_ptr = reinterpret_cast<uint64_t>(base);
+    output.resize(sizeof(uint64_t), 0);
+    memcpy(output.data(), &base_ptr, sizeof(base_ptr));
+    return true;
+}
+
+bool rpc_server::free_buffer(const std::vector<uint8_t> & input) {
+    // input serialization format: | remote_ptr (8 bytes) |
+    if (input.size() != sizeof(uint64_t)) {
+        return false;
+    }
+    uint64_t remote_ptr;
+    memcpy(&remote_ptr, input.data(), sizeof(remote_ptr));
+    GGML_PRINT_DEBUG("[%s] remote_ptr: %" PRIx64 "\n", __func__, remote_ptr);
+    ggml_backend_buffer_t buffer = reinterpret_cast<ggml_backend_buffer_t>(remote_ptr);
+    if (buffers.find(buffer) == buffers.end()) {
+        GGML_PRINT_DEBUG("[%s] buffer not found\n", __func__);
+        return false;
+    }
+    ggml_backend_buffer_free(buffer);
+    buffers.erase(buffer);
+    return true;
+}
+
+bool rpc_server::buffer_clear(const std::vector<uint8_t> & input) {
+    // input serialization format: | remote_ptr (8 bytes) | value (1 byte) |
+    if (input.size() != sizeof(uint64_t) + sizeof(uint8_t)) {
+        return false;
+    }
+    uint64_t remote_ptr;
+    memcpy(&remote_ptr, input.data(), sizeof(remote_ptr));
+    uint8_t value;
+    memcpy(&value, input.data() + sizeof(uint64_t), sizeof(value));
+    GGML_PRINT_DEBUG("[%s] remote_ptr: %" PRIx64 ", value: %u\n", __func__, remote_ptr, value);
+    ggml_backend_buffer_t buffer = reinterpret_cast<ggml_backend_buffer_t>(remote_ptr);
+    if (buffers.find(buffer) == buffers.end()) {
+        GGML_PRINT_DEBUG("[%s] buffer not found\n", __func__);
+        return false;
+    }
+    ggml_backend_buffer_clear(buffer, value);
+    return true;
+}
+
+ggml_tensor * rpc_server::deserialize_tensor(struct ggml_context * ctx, const rpc_tensor * tensor) {
+    ggml_tensor * result = ggml_new_tensor_4d(ctx, (ggml_type) tensor->type,
+        tensor->ne[0], tensor->ne[1], tensor->ne[2], tensor->ne[3]);
+    for (uint32_t i = 0; i < GGML_MAX_DIMS; i++) {
+        result->nb[i] = tensor->nb[i];
+    }
+    result->buffer = reinterpret_cast<ggml_backend_buffer_t>(tensor->buffer);
+    if (result->buffer && buffers.find(result->buffer) == buffers.end()) {
+        result->buffer = nullptr;
+    }
+
+    if (result->buffer) {
+        // require that the tensor data does not go beyond the buffer end
+        uint64_t tensor_size = (uint64_t) ggml_nbytes(result);
+        uint64_t buffer_start = (uint64_t) ggml_backend_buffer_get_base(result->buffer);
+        uint64_t buffer_size = (uint64_t) ggml_backend_buffer_get_size(result->buffer);
+        GGML_ASSERT(tensor->data + tensor_size >= tensor->data); // check for overflow
+        GGML_ASSERT(tensor->data >= buffer_start && tensor->data + tensor_size <= buffer_start + buffer_size);
+    }
+
+    result->op = (ggml_op) tensor->op;
+    for (uint32_t i = 0; i < GGML_MAX_OP_PARAMS / sizeof(int32_t); i++) {
+        result->op_params[i] = tensor->op_params[i];
+    }
+    result->flags = tensor->flags;
+    result->data = reinterpret_cast<void *>(tensor->data);
+    ggml_set_name(result, tensor->name);
+    return result;
+}
+
+
+bool rpc_server::set_tensor(const std::vector<uint8_t> & input) {
+    // serialization format: | rpc_tensor | offset (8 bytes) | data (size bytes) |
+    if (input.size() < sizeof(rpc_tensor) + sizeof(uint64_t)) {
+        return false;
+    }
+    const rpc_tensor * in_tensor = (const rpc_tensor *)input.data();
+    uint64_t offset;
+    memcpy(&offset, input.data() + sizeof(rpc_tensor), sizeof(offset));
+    const size_t size = input.size() - sizeof(rpc_tensor) - sizeof(offset);
+
+    struct ggml_init_params params {
+        /*.mem_size   =*/ ggml_tensor_overhead(),
+        /*.mem_buffer =*/ NULL,
+        /*.no_alloc   =*/ true,
+    };
+    struct ggml_context * ctx = ggml_init(params);
+    ggml_tensor * tensor = deserialize_tensor(ctx, in_tensor);
+    if (tensor == nullptr) {
+        GGML_PRINT_DEBUG("[%s] error deserializing tensor\n", __func__);
+        ggml_free(ctx);
+        return false;
+    }
+    GGML_PRINT_DEBUG("[%s] buffer: %p, data: %p, offset: %" PRIu64 ", size: %zu\n", __func__, (void*)tensor->buffer, tensor->data, offset, size);
+
+    // sanitize tensor->data
+    {
+        const size_t p0 = (size_t) ggml_backend_buffer_get_base(tensor->buffer);
+        const size_t p1 = p0 + ggml_backend_buffer_get_size(tensor->buffer);
+
+        if (in_tensor->data + offset < p0 || in_tensor->data + offset >= p1 || size > (p1 - in_tensor->data - offset)) {
+            GGML_ABORT("[%s] tensor->data out of bounds\n", __func__);
+        }
+    }
+
+    const void * data = input.data() + sizeof(rpc_tensor) + sizeof(offset);
+    ggml_backend_tensor_set(tensor, data, offset, size);
+    ggml_free(ctx);
+    return true;
+}
+
+bool rpc_server::get_tensor(const std::vector<uint8_t> & input, std::vector<uint8_t> & output) {
+    // serialization format: | rpc_tensor | offset (8 bytes) | size (8 bytes) |
+    if (input.size() != sizeof(rpc_tensor) + 2*sizeof(uint64_t)) {
+        return false;
+    }
+    const rpc_tensor * in_tensor = (const rpc_tensor *)input.data();
+    uint64_t offset;
+    memcpy(&offset, input.data() + sizeof(rpc_tensor), sizeof(offset));
+    uint64_t size;
+    memcpy(&size, input.data() + sizeof(rpc_tensor) + sizeof(offset), sizeof(size));
+
+    struct ggml_init_params params {
+        /*.mem_size   =*/ ggml_tensor_overhead(),
+        /*.mem_buffer =*/ NULL,
+        /*.no_alloc   =*/ true,
+    };
+    struct ggml_context * ctx = ggml_init(params);
+    ggml_tensor * tensor = deserialize_tensor(ctx, in_tensor);
+    if (tensor == nullptr) {
+        GGML_PRINT_DEBUG("[%s] error deserializing tensor\n", __func__);
+        ggml_free(ctx);
+        return false;
+    }
+    GGML_PRINT_DEBUG("[%s] buffer: %p, data: %p, offset: %" PRIu64 ", size: %" PRIu64 "\n", __func__, (void*)tensor->buffer, tensor->data, offset, size);
+
+    // sanitize tensor->data
+    {
+        const size_t p0 = (size_t) ggml_backend_buffer_get_base(tensor->buffer);
+        const size_t p1 = p0 + ggml_backend_buffer_get_size(tensor->buffer);
+
+        if (in_tensor->data + offset < p0 || in_tensor->data + offset >= p1 || size > (p1 - in_tensor->data - offset)) {
+            GGML_ABORT("[%s] tensor->data out of bounds\n", __func__);
+        }
+    }
+
+    // output serialization format: | data (size bytes) |
+    output.resize(size, 0);
+    ggml_backend_tensor_get(tensor, output.data(), offset, size);
+    ggml_free(ctx);
+    return true;
+}
+
+bool rpc_server::copy_tensor(const std::vector<uint8_t> & input, std::vector<uint8_t> & output) {
+    // serialization format: | rpc_tensor src | rpc_tensor dst |
+    if (input.size() != 2*sizeof(rpc_tensor)) {
+        return false;
+    }
+    const rpc_tensor * rpc_src = (const rpc_tensor *)input.data();
+    const rpc_tensor * rpc_dst = (const rpc_tensor *)(input.data() + sizeof(rpc_src));
+
+    struct ggml_init_params params {
+        /*.mem_size   =*/ 2*ggml_tensor_overhead(),
+        /*.mem_buffer =*/ NULL,
+        /*.no_alloc   =*/ true,
+    };
+    struct ggml_context * ctx = ggml_init(params);
+    ggml_tensor * src = deserialize_tensor(ctx, rpc_src);
+    ggml_tensor * dst = deserialize_tensor(ctx, rpc_dst);
+    if (src == nullptr || dst == nullptr) {
+        GGML_PRINT_DEBUG("[%s] error deserializing tensors\n", __func__);
+        ggml_free(ctx);
+        return false;
+    }
+    GGML_PRINT_DEBUG("[%s] src->buffer: %p, dst->buffer: %p\n", __func__, (void*)src->buffer, (void*)dst->buffer);
+    bool result = ggml_backend_buffer_copy_tensor(src, dst);
+    // output serialization format: | result (1 byte) |
+    output.resize(1, 0);
+    output[0] = result;
+    ggml_free(ctx);
+    return true;
+}
+
+ggml_tensor * rpc_server::create_node(uint64_t id,
+                                      struct ggml_context * ctx,
+                                      const std::unordered_map<uint64_t, const rpc_tensor*> & tensor_ptrs,
+                                      std::unordered_map<uint64_t, struct ggml_tensor*> & tensor_map) {
+    if (id == 0) {
+        return nullptr;
+    }
+    if (tensor_map.find(id) != tensor_map.end()) {
+        return tensor_map[id];
+    }
+    const rpc_tensor * tensor = tensor_ptrs.at(id);
+    struct ggml_tensor * result = deserialize_tensor(ctx, tensor);
+    if (result == nullptr) {
+        return nullptr;
+    }
+    tensor_map[id] = result;
+    for (int i = 0; i < GGML_MAX_SRC; i++) {
+        result->src[i] = create_node(tensor->src[i], ctx, tensor_ptrs, tensor_map);
+    }
+    result->view_src = create_node(tensor->view_src, ctx, tensor_ptrs, tensor_map);
+    result->view_offs = tensor->view_offs;
+    return result;
+}
+
+bool rpc_server::graph_compute(const std::vector<uint8_t> & input, std::vector<uint8_t> & output) {
+    // serialization format:
+    // | n_nodes (4 bytes) | nodes (n_nodes * sizeof(uint64_t) | n_tensors (4 bytes) | tensors (n_tensors * sizeof(rpc_tensor)) |
+    if (input.size() < sizeof(uint32_t)) {
+        return false;
+    }
+    uint32_t n_nodes;
+    memcpy(&n_nodes, input.data(), sizeof(n_nodes));
+    if (input.size() < sizeof(uint32_t) + n_nodes*sizeof(uint64_t) + sizeof(uint32_t)) {
+        return false;
+    }
+    const uint64_t * nodes = (const uint64_t *)(input.data() + sizeof(n_nodes));
+    uint32_t n_tensors;
+    memcpy(&n_tensors, input.data() + sizeof(n_nodes) + n_nodes*sizeof(uint64_t), sizeof(n_tensors));
+    if (input.size() < sizeof(uint32_t) + n_nodes*sizeof(uint64_t) + sizeof(uint32_t) + n_tensors*sizeof(rpc_tensor)) {
+        return false;
+    }
+    const rpc_tensor * tensors = (const rpc_tensor *)(input.data() + sizeof(n_nodes) + n_nodes*sizeof(uint64_t) + sizeof(n_tensors));
+    GGML_PRINT_DEBUG("[%s] n_nodes: %u, n_tensors: %u\n", __func__, n_nodes, n_tensors);
+
+    size_t buf_size = ggml_tensor_overhead()*(n_nodes + n_tensors) + ggml_graph_overhead_custom(n_nodes, false);
+    struct ggml_init_params params = {
+        /*.mem_size   =*/ buf_size,
+        /*.mem_buffer =*/ NULL,
+        /*.no_alloc   =*/ true,
+    };
+    struct ggml_context * ctx = ggml_init(params);
+    struct ggml_cgraph * graph = ggml_new_graph_custom(ctx, n_nodes, false);
+    graph->n_nodes = n_nodes;
+    std::unordered_map<uint64_t, const rpc_tensor*> tensor_ptrs;
+    for (uint32_t i = 0; i < n_tensors; i++) {
+        tensor_ptrs[tensors[i].id] = &tensors[i];
+    }
+    std::unordered_map<uint64_t, ggml_tensor*> tensor_map;
+    for (uint32_t i = 0; i < n_nodes; i++) {
+        int64_t id;
+        memcpy(&id, &nodes[i], sizeof(id));
+        graph->nodes[i] = create_node(id, ctx, tensor_ptrs, tensor_map);
+    }
+    ggml_status status = ggml_backend_graph_compute(backend, graph);
+    // output serialization format: | status (1 byte) |
+    output.resize(1, 0);
+    output[0] = status;
+    ggml_free(ctx);
+    return true;
+}
+
+rpc_server::~rpc_server() {
+    for (auto buffer : buffers) {
+        ggml_backend_buffer_free(buffer);
+    }
+}
+
+static void rpc_serve_client(ggml_backend_t backend, sockfd_t sockfd, size_t free_mem, size_t total_mem) {
+    rpc_server server(backend);
+    while (true) {
+        uint8_t cmd;
+        if (!recv_data(sockfd, &cmd, 1)) {
+            break;
+        }
+        if (cmd >= RPC_CMD_COUNT) {
+            // fail fast if the command is invalid
+            fprintf(stderr, "Unknown command: %d\n", cmd);
+            break;
+        }
+        std::vector<uint8_t> input;
+        std::vector<uint8_t> output;
+        uint64_t input_size;
+        if (!recv_data(sockfd, &input_size, sizeof(input_size))) {
+            break;
+        }
+        try {
+            input.resize(input_size);
+        } catch (const std::bad_alloc & e) {
+            fprintf(stderr, "Failed to allocate input buffer of size %" PRIu64 "\n", input_size);
+            break;
+        }
+        if (!recv_data(sockfd, input.data(), input_size)) {
+            break;
+        }
+        bool ok = true;
+        switch (cmd) {
+            case RPC_CMD_ALLOC_BUFFER: {
+                ok = server.alloc_buffer(input, output);
+                break;
+            }
+            case RPC_CMD_GET_ALIGNMENT: {
+                server.get_alignment(output);
+                break;
+            }
+            case RPC_CMD_GET_MAX_SIZE: {
+                server.get_max_size(output);
+                break;
+            }
+            case RPC_CMD_BUFFER_GET_BASE: {
+                ok = server.buffer_get_base(input, output);
+                break;
+            }
+            case RPC_CMD_FREE_BUFFER: {
+                ok = server.free_buffer(input);
+                break;
+            }
+            case RPC_CMD_BUFFER_CLEAR: {
+                ok = server.buffer_clear(input);
+                break;
+            }
+            case RPC_CMD_SET_TENSOR: {
+                ok = server.set_tensor(input);
+                break;
+            }
+            case RPC_CMD_GET_TENSOR: {
+                ok = server.get_tensor(input, output);
+                break;
+            }
+            case RPC_CMD_COPY_TENSOR: {
+                ok = server.copy_tensor(input, output);
+                break;
+            }
+            case RPC_CMD_GRAPH_COMPUTE: {
+                ok = server.graph_compute(input, output);
+                break;
+            }
+            case RPC_CMD_GET_DEVICE_MEMORY: {
+                // output serialization format: | free (8 bytes) | total (8 bytes) |
+                output.resize(2*sizeof(uint64_t), 0);
+                memcpy(output.data(), &free_mem, sizeof(free_mem));
+                memcpy(output.data() + sizeof(uint64_t), &total_mem, sizeof(total_mem));
+                break;
+            }
+            default: {
+                fprintf(stderr, "Unknown command: %d\n", cmd);
+                ok = false;
+            }
+        }
+        if (!ok) {
+            break;
+        }
+        uint64_t output_size = output.size();
+        if (!send_data(sockfd, &output_size, sizeof(output_size))) {
+            break;
+        }
+        if (!send_data(sockfd, output.data(), output_size)) {
+            break;
+        }
+    }
+}
+
+void start_rpc_server(ggml_backend_t backend, const char * endpoint, size_t free_mem, size_t total_mem) {
+    std::string host;
+    int port;
+    if (!parse_endpoint(endpoint, host, port)) {
+        return;
+    }
+#ifdef _WIN32
+    {
+        WSADATA wsaData;
+        int res = WSAStartup(MAKEWORD(2, 2), &wsaData);
+        if (res != 0) {
+            fprintf(stderr, "WSAStartup failed: %d\n", res);
+            return;
+        }
+    }
+#endif
+    auto server_socket = create_server_socket(host.c_str(), port);
+    if (server_socket == nullptr) {
+        fprintf(stderr, "Failed to create server socket\n");
+        return;
+    }
+    while (true) {
+        auto client_socket = socket_accept(server_socket->fd);
+        if (client_socket == nullptr) {
+            fprintf(stderr, "Failed to accept client connection\n");
+            return;
+        }
+        printf("Accepted client connection, free_mem=%zu, total_mem=%zu\n", free_mem, total_mem);
+        fflush(stdout);
+        rpc_serve_client(backend, client_socket->fd, free_mem, total_mem);
+        printf("Client connection closed\n");
+        fflush(stdout);
+    }
+#ifdef _WIN32
+    WSACleanup();
+#endif
+}
diff --git a/src/whisper_cpp/ggml/src/ggml-sycl.cpp b/src/whisper_cpp/ggml/src/ggml-sycl.cpp
new file mode 100644
index 00000000..4d3f1c5c
--- /dev/null
+++ b/src/whisper_cpp/ggml/src/ggml-sycl.cpp
@@ -0,0 +1,5242 @@
+//
+// MIT license
+// Copyright (C) 2024 Intel Corporation
+// SPDX-License-Identifier: MIT
+//
+
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+//
+
+#include <algorithm>
+#include <assert.h>
+#include <atomic>
+#include <cinttypes>
+#include <cstddef>
+#include <cstdint>
+#include <cstdlib>
+#include <float.h>
+#include <limits>
+#include <stdint.h>
+#include <stdio.h>
+#include <vector>
+#include <cmath>
+#include <iostream>
+#include <fstream>
+#include <stdio.h>
+#include <stdlib.h>
+#include <regex>
+
+#include <sycl/sycl.hpp>
+#include <sycl/half_type.hpp>
+
+#include "ggml-sycl.h"
+#include "ggml-impl.h"
+#include "ggml-backend-impl.h"
+
+#include "ggml-sycl/backend.hpp"
+#include "ggml-sycl/presets.hpp"
+#include "ggml-sycl/gemm.hpp"
+
+bool   ggml_sycl_loaded(void);
+void   ggml_sycl_free_data(struct ggml_tensor * tensor);
+void   ggml_sycl_copy_to_device(struct ggml_tensor * tensor);
+void   ggml_sycl_set_main_device(int main_device);
+void   ggml_sycl_set_mul_mat_q(bool mul_mat_q);
+void   ggml_sycl_get_device_description(int device, char * description, size_t description_size);
+bool   ggml_backend_is_sycl(ggml_backend_t backend);
+int    ggml_backend_sycl_get_device(ggml_backend_t backend);
+static bool ggml_backend_buffer_is_sycl_split(ggml_backend_buffer_t buffer);
+static inline int get_sycl_env(const char *env_name, int default_val);
+
+
+void dev2dev_memcpy(sycl::queue &q_dst, sycl::queue &q_src, void *ptr_dst,
+                    const void *ptr_src, size_t size) {
+    char *host_buf = (char *)malloc(size);
+    q_src.memcpy(host_buf, (const char *)ptr_src, size).wait();
+    q_dst.memcpy((char *)ptr_dst, host_buf, size).wait();
+    free(host_buf);
+}
+
+typedef void (*cpy_kernel_t)(const char * cx, char * cdst);
+typedef void (*ggml_sycl_func_t)(ggml_backend_sycl_context & ctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst);
+typedef void (*ggml_sycl_op_mul_mat_t)(
+    ggml_backend_sycl_context & ctx,
+    const ggml_tensor *src0, const ggml_tensor *src1, ggml_tensor *dst,
+    const char *src0_dd_i, const float *src1_ddf_i, const char *src1_ddq_i,
+    float *dst_dd_i, const int64_t row_low, const int64_t row_high,
+    const int64_t src1_ncols, const int64_t src1_padded_row_size,
+    const queue_ptr &stream);
+typedef void (*ggml_sycl_op_flatten_t)(ggml_backend_sycl_context & ctx, const ggml_tensor *src0,
+                                       const ggml_tensor *src1,
+                                       ggml_tensor *dst, const float *src0_dd,
+                                       const float *src1_dd, float *dst_dd,
+                                       const queue_ptr &main_stream);
+
+static __dpct_inline__ float op_repeat(const float a, const float b) {
+    return b;
+    GGML_UNUSED(a);
+}
+
+static __dpct_inline__ float op_add(const float a, const float b) {
+    return a + b;
+}
+
+static __dpct_inline__ float op_mul(const float a, const float b) {
+    return a * b;
+}
+
+static __dpct_inline__ float op_div(const float a, const float b) {
+    return a / b;
+}
+
+template<float (*bin_op)(const float, const float), typename src0_t, typename src1_t, typename dst_t>
+static void k_bin_bcast(const src0_t * src0, const src1_t * src1, dst_t * dst,
+        int ne0, int ne1, int ne2, int ne3,
+        int ne10, int ne11, int ne12, int ne13,
+        /*int s0, */ int s1,  int s2,  int s3,
+        /*int s10,*/ int s11, int s12, int s13,
+        const sycl::nd_item<3> &item_ct1) {
+    const int i0s = item_ct1.get_local_range(2) * item_ct1.get_group(2) +
+                    item_ct1.get_local_id(2);
+    const int i1 = (item_ct1.get_local_range(1) * item_ct1.get_group(1) +
+                    item_ct1.get_local_id(1));
+    const int i2 = (item_ct1.get_local_range(0) * item_ct1.get_group(0) +
+                    item_ct1.get_local_id(0)) /
+                   ne3;
+    const int i3 = (item_ct1.get_local_range(0) * item_ct1.get_group(0) +
+                    item_ct1.get_local_id(0)) %
+                   ne3;
+
+    if (i0s >= ne0 || i1 >= ne1 || i2 >= ne2 || i3 >= ne3) {
+        return;
+    }
+
+    const int i11 = i1 % ne11;
+    const int i12 = i2 % ne12;
+    const int i13 = i3 % ne13;
+
+    const size_t i_src0 = i3*s3 + i2*s2 + i1*s1;
+    const size_t i_src1 = i13*s13 + i12*s12 + i11*s11;
+    const size_t i_dst  = i_src0;
+
+    const src0_t * src0_row = src0 + i_src0;
+    const src1_t * src1_row = src1 + i_src1;
+    dst_t * dst_row = dst + i_dst;
+
+    for (int i0 = i0s; i0 < ne0;
+         i0 += item_ct1.get_local_range(2) * item_ct1.get_group_range(2)) {
+        const int i10 = i0 % ne10;
+        dst_row[i0] = (dst_t)bin_op(src0 ? (float)src0_row[i0] : 0.0f, (float)src1_row[i10]);
+    }
+}
+
+template<float (*bin_op)(const float, const float), typename src0_t, typename src1_t, typename dst_t>
+static void k_bin_bcast_unravel(const src0_t * src0, const src1_t * src1, dst_t * dst,
+        int ne0, int ne1, int ne2, int ne3,
+        int ne10, int ne11, int ne12, int ne13,
+        /*int s0, */ int s1,  int s2,  int s3,
+        /*int s10,*/ int s11, int s12, int s13,
+        const sycl::nd_item<3> &item_ct1) {
+
+    const int i = item_ct1.get_local_range(2) * item_ct1.get_group(2) +
+                  item_ct1.get_local_id(2);
+
+    const int i3 = i/(ne2*ne1*ne0);
+    const int i2 = (i/(ne1*ne0)) % ne2;
+    const int i1 = (i/ne0) % ne1;
+    const int i0 = i % ne0;
+
+    if (i0 >= ne0 || i1 >= ne1 || i2 >= ne2 || i3 >= ne3) {
+        return;
+    }
+
+    const int i11 = i1 % ne11;
+    const int i12 = i2 % ne12;
+    const int i13 = i3 % ne13;
+
+    const size_t i_src0 = i3*s3 + i2*s2 + i1*s1;
+    const size_t i_src1 = i13*s13 + i12*s12 + i11*s11;
+    const size_t i_dst  = i_src0;
+
+    const src0_t * src0_row = src0 + i_src0;
+    const src1_t * src1_row = src1 + i_src1;
+    dst_t * dst_row = dst + i_dst;
+
+    const int i10 = i0 % ne10;
+    dst_row[i0] = (dst_t)bin_op(src0 ? (float)src0_row[i0] : 0.0f, (float)src1_row[i10]);
+}
+
+static void acc_f32(const float * x, const float * y, float * dst, const int ne,
+    const int ne10, const int ne11, const int ne12,
+    const int nb1, const int nb2, int offset, const sycl::nd_item<3> &item_ct1) {
+    const int i = item_ct1.get_local_range(2) * item_ct1.get_group(2) +
+                  item_ct1.get_local_id(2);
+    if (i >= ne) {
+        return;
+    }
+    int src1_idx = i - offset;
+    int oz = src1_idx / nb2;
+    int oy = (src1_idx - (oz * nb2)) / nb1;
+    int ox = src1_idx % nb1;
+    if (src1_idx >= 0 && ox < ne10 && oy < ne11 && oz < ne12) {
+        dst[i] = x[i] + y[ox + oy * ne10 + oz * ne10 * ne11];
+    } else {
+        dst[i] = x[i];
+    }
+}
+
+static void gelu_f32(const float * x, float * dst, const int k,
+                     const sycl::nd_item<3> &item_ct1) {
+    const float GELU_COEF_A    = 0.044715f;
+    const float SQRT_2_OVER_PI = 0.79788456080286535587989211986876f;
+    const int i = item_ct1.get_local_range(2) * item_ct1.get_group(2) +
+                  item_ct1.get_local_id(2);
+
+    if (i >= k) {
+        return;
+    }
+
+    float xi = x[i];
+    dst[i] = 0.5f * xi *
+             (1.0f +
+              sycl::tanh(SQRT_2_OVER_PI * xi * (1.0f + GELU_COEF_A * xi * xi)));
+}
+
+static void silu_f32(const float * x, float * dst, const int k,
+                     const sycl::nd_item<3> &item_ct1) {
+    const int i = item_ct1.get_local_range(2) * item_ct1.get_group(2) +
+                  item_ct1.get_local_id(2);
+
+    if (i >= k) {
+        return;
+    }
+    dst[i] = x[i] / (1.0f + sycl::native::exp(-x[i]));
+}
+
+static void gelu_quick_f32(const float *x, float *dst, int k,
+                           const sycl::nd_item<3> &item_ct1) {
+    const float GELU_QUICK_COEF = -1.702f;
+    const int i = item_ct1.get_local_range(2) * item_ct1.get_group(2) +
+                  item_ct1.get_local_id(2);
+    if (i >= k) {
+        return;
+    }
+    dst[i] = x[i] * (1.0f / (1.0f + sycl::native::exp(GELU_QUICK_COEF * x[i])));
+}
+
+static void tanh_f32(const float *x, float *dst, int k,
+                     const sycl::nd_item<3> &item_ct1) {
+    const int i = item_ct1.get_local_range(2) * item_ct1.get_group(2) +
+                  item_ct1.get_local_id(2);
+    if (i >= k) {
+        return;
+    }
+    dst[i] = sycl::tanh((float)(x[i]));
+}
+
+static void relu_f32(const float * x, float * dst, const int k,
+                     const sycl::nd_item<3> &item_ct1) {
+    const int i = item_ct1.get_local_range(2) * item_ct1.get_group(2) +
+                  item_ct1.get_local_id(2);
+
+    if (i >= k) {
+        return;
+    }
+    dst[i] = sycl::fmax((float)(x[i]), (float)0);
+}
+
+static void hardsigmoid_f32(const float * x, float * dst, const int k,
+                            const sycl::nd_item<3> &item_ct1) {
+    const int i = item_ct1.get_local_range(2) * item_ct1.get_group(2) +
+                  item_ct1.get_local_id(2);
+
+    if (i >= k) {
+        return;
+    }
+    dst[i] = sycl::fmin(1.0f, sycl::fmax(0.0f, (x[i] + 3.0f) / 6.0f));
+}
+
+static void hardswish_f32(const float * x, float * dst, const int k,
+                          const sycl::nd_item<3> &item_ct1) {
+    const int i = item_ct1.get_local_range(2) * item_ct1.get_group(2) +
+                  item_ct1.get_local_id(2);
+
+    if (i >= k) {
+        return;
+    }
+    dst[i] = x[i] * sycl::fmin(1.0f, sycl::fmax(0.0f, (x[i] + 3.0f) / 6.0f));
+}
+
+static void leaky_relu_f32(const float *x, float *dst, const int k, const float negative_slope,
+                           const sycl::nd_item<3> &item_ct1) {
+    const int i = item_ct1.get_local_range(2) * item_ct1.get_group(2) +
+                  item_ct1.get_local_id(2);
+    if (i >= k) {
+        return;
+    }
+    dst[i] = sycl::fmax((float)(x[i]), (float)0) +
+             sycl::fmin((float)(x[i]), 0.0f) * negative_slope;
+}
+
+static void sqr_f32(const float * x, float * dst, const int k,
+                    const sycl::nd_item<3> &item_ct1) {
+    const int i = item_ct1.get_local_range(2) * item_ct1.get_group(2) +
+                  item_ct1.get_local_id(2);
+
+    if (i >= k) {
+        return;
+    }
+    dst[i] = x[i] * x[i];
+}
+
+static void upscale_f32(const float  *x, float *dst, const int nb00, const int nb01,
+                        const int nb02, const int nb03, const int ne10, const int ne11,
+                        const int ne12, const int ne13, const float sf0, const float sf1,
+                        const float sf2, const float sf3, const sycl::nd_item<1> &item_ct1) {
+    int index = item_ct1.get_local_id(0) +
+               item_ct1.get_group(0) * item_ct1.get_local_range(0);
+    if (index >= ne10 * ne11 * ne12 * ne13) {
+        return;
+    }
+    // operation
+    int i10 = index % ne10;
+    int i11 = (index / ne10) % ne11;
+    int i12 = (index / (ne10 * ne11)) % ne12;
+    int i13 = (index / (ne10 * ne11 * ne12)) % ne13;
+
+    int i00 = i10 / sf0;
+    int i01 = i11 / sf1;
+    int i02 = i12 / sf2;
+    int i03 = i13 / sf3;
+
+    dst[index] = *(float *)((char *)x + i03 * nb03 + i02 * nb02 + i01 * nb01 + i00 * nb00);
+}
+
+static void pad_f32(const float  *x, float *dst, const int ne0, const int ne00, const int ne01, const int ne02,
+                    const sycl::nd_item<3> &item_ct1) {
+    int nidx = item_ct1.get_local_id(2) +
+               item_ct1.get_group(2) * item_ct1.get_local_range(2);
+    if (nidx >= ne0) {
+        return;
+    }
+
+    // operation
+    int offset_dst = nidx + item_ct1.get_group(1) * ne0 +
+                     item_ct1.get_group(0) * ne0 * item_ct1.get_group_range(1);
+    if (nidx < ne00 && item_ct1.get_group(1) < ne01 &&
+        item_ct1.get_group(0) < ne02) {
+        int offset_src = nidx + item_ct1.get_group(1) * ne00 +
+                         item_ct1.get_group(0) * ne00 * ne01;
+            dst[offset_dst] = x[offset_src];
+    } else {
+        dst[offset_dst] = 0.0f;
+    }
+}
+
+template<int QUANT_BLOCK_TILE>
+static void quantize_q8_1(const float * __restrict__ x, void * __restrict__ vy, const int kx, const int kx_padded,
+                          const sycl::nd_item<3> &item_ct1) {
+    const int ix = (item_ct1.get_local_range(2) * item_ct1.get_group(2) +
+                    item_ct1.get_local_id(2)) * QUANT_BLOCK_TILE;
+
+    if (ix >= kx_padded) {
+        return;
+    }
+
+    const int iy = item_ct1.get_local_range(1) * item_ct1.get_group(1) +
+                   item_ct1.get_local_id(1);
+
+    const int i_padded = iy*kx_padded + ix;
+
+    block_q8_1 * y = (block_q8_1 *) vy;
+
+    const int ib = i_padded / QK8_1; // block index
+    const int iqs = i_padded % QK8_1; // quant index
+    typedef  sycl::vec<float, QUANT_BLOCK_TILE> TC;
+    typedef  sycl::vec<int8_t, QUANT_BLOCK_TILE> TQ;
+    TC zeros;
+    TQ qzeros;
+#pragma unroll
+    for (int i = 0; i < QUANT_BLOCK_TILE; i++)
+    {
+        zeros[i] = 0.f;
+        qzeros[i] = 0;
+    }
+    const TC xi = ix < kx ? *(TC *)&x[iy * kx + ix] : zeros;
+    float sum = xi[0];
+    float amax = sycl::fabs(xi[0]);
+#pragma unroll
+    for (int i = 1; i < QUANT_BLOCK_TILE; i++)
+    {
+        sum += xi[i];
+        amax = sycl::fmax(sycl::fabs(xi[i]), amax);
+    }
+    sum = warp_reduce_sum(sum, item_ct1);
+    amax = warp_reduce_max(amax, item_ct1);
+
+    const float d = amax / 127;
+    TQ q = qzeros;
+    if (amax != 0.0f)
+    {
+#pragma unroll
+        for (int i = 0; i < QUANT_BLOCK_TILE; i++) {
+            q[i] = sycl::round(xi[i] / d);
+        }
+    }
+
+    *(TQ *)&y[ib].qs[iqs] = q;
+
+    if (iqs > 0) {
+        return;
+    }
+
+    reinterpret_cast<sycl::half &>(y[ib].ds.x()) = d;
+    reinterpret_cast<sycl::half &>(y[ib].ds.y()) = sum;
+}
+
+template<int qk, int qr, dequantize_kernel_t dequantize_kernel, typename dst_t>
+static void k_get_rows(
+            const void * src0, const int32_t * src1, dst_t * dst,
+            int64_t ne00, /*int64_t ne01, int64_t ne02, int64_t ne03,*/
+            /*int64_t ne10, int64_t ne11,*/ int64_t ne12, /*int64_t ne13,*/
+            /*size_t s0,*/ size_t s1, size_t s2, size_t s3,
+            /*size_t nb00,*/ size_t nb01, size_t nb02, size_t nb03,
+            size_t s10, size_t s11, size_t s12,
+            const sycl::nd_item<3> &item_ct1/*, size_t s13*/) {
+
+    const int i00 = (item_ct1.get_group(2) * item_ct1.get_local_range(2) +
+                     item_ct1.get_local_id(2)) *
+                    2;
+    const int i10 = item_ct1.get_local_range(1) * item_ct1.get_group(1) +
+                    item_ct1.get_local_id(1);
+    const int i11 = (item_ct1.get_group(0) * item_ct1.get_local_range(0) +
+                     item_ct1.get_local_id(0)) /
+                    ne12;
+    const int i12 = (item_ct1.get_group(0) * item_ct1.get_local_range(0) +
+                     item_ct1.get_local_id(0)) %
+                    ne12;
+
+    if (i00 >= ne00) {
+        return;
+    }
+
+    const int i01 = src1[i10*s10 + i11*s11 + i12*s12];
+
+    dst_t * dst_row = dst + i10*s1 + i11*s2 + i12*s3;
+    const void * src0_row = (const char *)src0 + i01*nb01 + i11*nb02 + i12*nb03;
+
+    const int ib = i00/qk; // block index
+    const int iqs = (i00%qk)/qr; // quant index
+    const int iybs = i00 - i00%qk; // dst block start index
+    const int y_offset = qr == 1 ? 1 : qk/2;
+
+    // dequantize
+    dfloat2 v;
+    dequantize_kernel(src0_row, ib, iqs, v);
+
+    dst_row[iybs + iqs + 0] = v.x();
+    dst_row[iybs + iqs + y_offset] = v.y();
+}
+
+template<typename src0_t, typename dst_t>
+static void k_get_rows_float(
+            const src0_t * src0, const int32_t * src1, dst_t * dst,
+            int64_t ne00, /*int64_t ne01, int64_t ne02, int64_t ne03,*/
+            /*int64_t ne10, int64_t ne11,*/ int64_t ne12, /*int64_t ne13,*/
+            /*size_t s0,*/ size_t s1, size_t s2, size_t s3,
+            /*size_t nb00,*/ size_t nb01, size_t nb02, size_t nb03,
+            size_t s10, size_t s11, size_t s12,
+            const sycl::nd_item<3> &item_ct1/*, size_t s13*/) {
+
+    const int i00 = item_ct1.get_group(2) * item_ct1.get_local_range(2) +
+                    item_ct1.get_local_id(2);
+    const int i10 = item_ct1.get_local_range(1) * item_ct1.get_group(1) +
+                    item_ct1.get_local_id(1);
+    const int i11 = (item_ct1.get_group(0) * item_ct1.get_local_range(0) +
+                     item_ct1.get_local_id(0)) /
+                    ne12;
+    const int i12 = (item_ct1.get_group(0) * item_ct1.get_local_range(0) +
+                     item_ct1.get_local_id(0)) %
+                    ne12;
+
+    if (i00 >= ne00) {
+        return;
+    }
+
+    const int i01 = src1[i10*s10 + i11*s11 + i12*s12];
+
+    dst_t * dst_row = dst + i10*s1 + i11*s2 + i12*s3;
+    const src0_t * src0_row = (const src0_t *)((const char *)src0 + i01*nb01 + i11*nb02 + i12*nb03);
+
+    dst_row[i00] = src0_row[i00];
+}
+
+static void mul_mat_p021_f16_f32(
+    const void * __restrict__ vx, const float * __restrict__ y, float * __restrict__ dst,
+    const int ncols_x, const int nrows_x, const int nchannels_x, const int nchannels_y,
+    const sycl::nd_item<3> &item_ct1) {
+
+    const sycl::half *x = (const sycl::half *)vx;
+
+    const int row_x = item_ct1.get_local_range(1) * item_ct1.get_group(1) +
+                      item_ct1.get_local_id(1);
+    const int channel = item_ct1.get_local_range(0) * item_ct1.get_group(0) +
+                        item_ct1.get_local_id(0);
+    const int channel_x = channel / (nchannels_y / nchannels_x);
+
+    const int nrows_y = ncols_x;
+    const int nrows_dst = nrows_x;
+    const int row_dst = row_x;
+
+    float tmp = 0.0f;
+
+    for (int col_x0 = 0; col_x0 < ncols_x;
+         col_x0 += item_ct1.get_local_range(2)) {
+        const int col_x = col_x0 + item_ct1.get_local_id(2);
+
+        if (col_x >= ncols_x) {
+            break;
+        }
+
+        // x is transposed and permuted
+        const int ix = row_x*nchannels_x*ncols_x + channel_x*ncols_x + col_x;
+        const float xi =
+            sycl::vec<sycl::half, 1>(x[ix])
+                .convert<float, sycl::rounding_mode::automatic>()[0];
+
+        const int row_y = col_x;
+
+
+        // y is not transposed but permuted
+        const int iy = channel*nrows_y + row_y;
+
+        tmp += xi * y[iy];
+    }
+
+    // dst is not transposed and not permuted
+    const int idst = channel*nrows_dst + row_dst;
+
+    // sum up partial sums and write back result
+#pragma unroll
+    for (int mask = WARP_SIZE / 2; mask > 0; mask >>= 1) {
+        tmp +=
+            dpct::permute_sub_group_by_xor(item_ct1.get_sub_group(), tmp, mask);
+    }
+
+    if (item_ct1.get_local_id(2) == 0) {
+        dst[idst] = tmp;
+    }
+}
+
+static void mul_mat_vec_nc_f16_f32( // nc == non-contiguous
+    const void * __restrict__ vx, const float * __restrict__ y, float * __restrict__ dst, const int ncols_x, const int nrows_x,
+    const int row_stride_x, const int channel_stride_x, const int channel_x_divisor,
+    const sycl::nd_item<3> &item_ct1) {
+
+    const sycl::half *x = (const sycl::half *)vx;
+
+    const int row_x = item_ct1.get_local_range(1) * item_ct1.get_group(1) +
+                      item_ct1.get_local_id(1);
+    const int channel = item_ct1.get_local_range(0) * item_ct1.get_group(0) +
+                        item_ct1.get_local_id(0);
+    const int channel_x = channel / channel_x_divisor;
+
+    const int nrows_y   = ncols_x;
+    const int nrows_dst = nrows_x;
+    const int row_dst   = row_x;
+
+    const int idst = channel*nrows_dst + row_dst;
+
+    float tmp = 0.0f;
+
+    for (int col_x0 = 0; col_x0 < ncols_x;
+         col_x0 += item_ct1.get_local_range(2)) {
+        const int col_x = col_x0 + item_ct1.get_local_id(2);
+
+        if (col_x >= ncols_x) {
+            break;
+        }
+
+        const int row_y = col_x;
+
+        const int ix = channel_x*channel_stride_x + row_x*row_stride_x + col_x;
+        const int iy = channel*nrows_y + row_y;
+
+        const float xi =
+            sycl::vec<sycl::half, 1>(x[ix])
+                .convert<float, sycl::rounding_mode::automatic>()[0];
+
+        tmp += xi * y[iy];
+    }
+
+    // sum up partial sums and write back result
+#pragma unroll
+    for (int mask = WARP_SIZE / 2; mask > 0; mask >>= 1) {
+        tmp +=
+            dpct::permute_sub_group_by_xor(item_ct1.get_sub_group(), tmp, mask);
+    }
+
+    if (item_ct1.get_local_id(2) == 0) {
+        dst[idst] = tmp;
+    }
+}
+
+static void cpy_1_f32_f32(const char * cxi, char * cdsti) {
+    const float * xi = (const float *) cxi;
+    float * dsti = (float *) cdsti;
+
+    *dsti = *xi;
+}
+
+static void cpy_1_f32_f16(const char * cxi, char * cdsti) {
+    const float * xi = (const float *) cxi;
+    sycl::half *dsti = (sycl::half *)cdsti;
+
+    *dsti = sycl::vec<float, 1>(*xi)
+                .convert<sycl::half, sycl::rounding_mode::automatic>()[0];
+}
+
+static void cpy_1_f16_f16(const char * cxi, char * cdsti) {
+    const sycl::half *xi = (const sycl::half *)cxi;
+    sycl::half *dsti = (sycl::half *)cdsti;
+
+    *dsti = *xi;
+}
+
+static void cpy_1_f16_f32(const char * cxi, char * cdsti) {
+    const sycl::half *xi = (const sycl::half *)cxi;
+    float * dsti = (float *) cdsti;
+
+    *dsti = *xi;
+}
+
+static void cpy_1_i16_i16(const char * cxi, char * cdsti) {
+    const int16_t *xi = (const int16_t *)cxi;
+    int16_t *dsti = (int16_t *)cdsti;
+
+    *dsti = *xi;
+}
+
+static void cpy_1_i32_i32(const char * cxi, char * cdsti) {
+    const int32_t *xi = (const int32_t *)cxi;
+    int32_t *dsti = (int32_t *)cdsti;
+
+    *dsti = *xi;
+}
+
+template <cpy_kernel_t cpy_1>
+static void cpy_f32_f16(const char * cx, char * cdst, const int ne,
+                        const int ne00, const int ne01, const int ne02, const int nb00, const int nb01, const int nb02,
+                        const int nb03, const int ne10, const int ne11, const int ne12, const int nb10, const int nb11,
+                        const int nb12, const int nb13, const sycl::nd_item<3> &item_ct1) {
+    const int i = item_ct1.get_local_range(2) * item_ct1.get_group(2) +
+                  item_ct1.get_local_id(2);
+
+    if (i >= ne) {
+        return;
+    }
+
+    // determine indices i02/i12, i01/i11, i00/i10 as a function of index i of flattened tensor
+    // then combine those indices with the corresponding byte offsets to get the total offsets
+    const int i03 = i/(ne00 * ne01 * ne02);
+    const int i02 = (i - i03*ne00*ne01*ne02 )/ (ne00*ne01);
+    const int i01 = (i - i03*ne00*ne01*ne02  -  i02*ne01*ne00) / ne00;
+    const int i00 = i - i03*ne00*ne01*ne02 - i02*ne01*ne00 - i01*ne00;
+    const int x_offset = i00*nb00 + i01*nb01 + i02*nb02 + i03 * nb03;
+
+    const int i13 = i/(ne10 * ne11 * ne12);
+    const int i12 = (i - i13*ne10*ne11*ne12) / (ne10*ne11);
+    const int i11 = (i - i13*ne10*ne11*ne12 - i12*ne10*ne11) / ne10;
+    const int i10 = i - i13*ne10*ne11*ne12 - i12*ne10*ne11 - i11*ne10;
+    const int dst_offset = i10*nb10 + i11*nb11 + i12*nb12 + i13 * nb13;
+
+    cpy_1(cx + x_offset, cdst + dst_offset);
+}
+
+static void cpy_blck_f32_q8_0(const char * cxi, char * cdsti) {
+    const float * xi = (const float *) cxi;
+    block_q8_0 * dsti = (block_q8_0 *) cdsti;
+
+    float amax = 0.0f; // absolute max
+
+    for (int j = 0; j < QK8_0; j++) {
+        const float v = xi[j];
+        amax = sycl::fmax(amax, sycl::fabs((float)v));
+    }
+
+    const float d = amax / ((1 << 7) - 1);
+    const float id = d ? 1.0f/d : 0.0f;
+
+    dsti->d = d;
+
+    for (int j = 0; j < QK8_0; ++j) {
+        const float x0 = xi[j]*id;
+
+        dsti->qs[j] = sycl::round((float)x0);
+    }
+}
+
+static void cpy_blck_f32_q4_0(const char * cxi, char * cdsti) {
+    const float * xi = (const float *) cxi;
+    block_q4_0 * dsti = (block_q4_0 *) cdsti;
+
+    float amax = 0.0f;
+    float vmax = 0.0f;
+
+    for (int j = 0; j < QK4_0; ++j) {
+        const float v = xi[j];
+        if (amax < sycl::fabs((float)v)) {
+            amax = sycl::fabs((float)v);
+            vmax = v;
+        }
+    }
+
+    const float d  = vmax / -8;
+    const float id = d ? 1.0f/d : 0.0f;
+
+    dsti->d = d;
+
+    for (int j = 0; j < QK4_0/2; ++j) {
+        const float x0 = xi[0       + j]*id;
+        const float x1 = xi[QK4_0/2 + j]*id;
+
+        const uint8_t xi0 = dpct::min(15, (int8_t)(x0 + 8.5f));
+        const uint8_t xi1 = dpct::min(15, (int8_t)(x1 + 8.5f));
+
+        dsti->qs[j]  = xi0;
+        dsti->qs[j] |= xi1 << 4;
+    }
+}
+
+static void cpy_blck_f32_q4_1(const char * cxi, char * cdsti) {
+    const float * xi = (const float *) cxi;
+    block_q4_1 * dsti = (block_q4_1 *) cdsti;
+
+    float vmin = FLT_MAX;
+    float vmax = -FLT_MAX;
+
+    for (int j = 0; j < QK4_1; ++j) {
+        const float v = xi[j];
+
+        if (v < vmin) vmin = v;
+        if (v > vmax) vmax = v;
+    }
+
+    const float d  = (vmax - vmin) / ((1 << 4) - 1);
+    const float id = d ? 1.0f/d : 0.0f;
+
+    dsti->dm.x() = d;
+    dsti->dm.y() = vmin;
+
+    for (int j = 0; j < QK4_1/2; ++j) {
+        const float x0 = (xi[0       + j] - vmin)*id;
+        const float x1 = (xi[QK4_1/2 + j] - vmin)*id;
+
+        const uint8_t xi0 = dpct::min(15, (int8_t)(x0 + 0.5f));
+        const uint8_t xi1 = dpct::min(15, (int8_t)(x1 + 0.5f));
+
+        dsti->qs[j]  = xi0;
+        dsti->qs[j] |= xi1 << 4;
+    }
+}
+
+template <cpy_kernel_t cpy_blck, int qk>
+static void cpy_f32_q(const char * cx, char * cdst, const int ne,
+                      const int ne00, const int ne01, const int ne02, const int nb00, const int nb01, const int nb02,
+                      const int nb03, const int ne10, const int ne11, const int ne12, const int nb10, const int nb11,
+                      const int nb12, const int nb13, const sycl::nd_item<3> &item_ct1) {
+    const int i = (item_ct1.get_local_range(2) * item_ct1.get_group(2) +
+                   item_ct1.get_local_id(2)) *
+                  qk;
+
+    if (i >= ne) {
+        return;
+    }
+
+    const int i03 = i/(ne00 * ne01 * ne02);
+    const int i02 = (i - i03*ne00*ne01*ne02 )/ (ne00*ne01);
+    const int i01 = (i - i03*ne00*ne01*ne02  -  i02*ne01*ne00) / ne00;
+    const int i00 = i - i03*ne00*ne01*ne02 - i02*ne01*ne00 - i01*ne00;
+    const int x_offset = i00*nb00 + i01*nb01 + i02*nb02 + i03 * nb03;
+
+    const int i13 = i/(ne10 * ne11 * ne12);
+    const int i12 = (i - i13*ne10*ne11*ne12) / (ne10*ne11);
+    const int i11 = (i - i13*ne10*ne11*ne12 - i12*ne10*ne11) / ne10;
+    const int i10 = i - i13*ne10*ne11*ne12 - i12*ne10*ne11 - i11*ne10;
+    const int dst_offset = (i10/qk)*nb10 + i11*nb11 + i12*nb12 + i13*nb13;
+
+    cpy_blck(cx + x_offset, cdst + dst_offset);
+}
+
+static void k_sum_rows_f32(const float * x, float * dst, const int ncols,
+                           const sycl::nd_item<3> &item_ct1) {
+    const int row = item_ct1.get_group(1);
+    const int col = item_ct1.get_local_id(2);
+
+    float sum = 0.0f;
+    for (int i = col; i < ncols; i += item_ct1.get_local_range(2)) {
+        sum += x[row * ncols + i];
+    }
+
+    sum = warp_reduce_sum(sum, item_ct1);
+
+    if (col == 0) {
+        dst[row] = sum;
+    }
+}
+
+
+template<typename T>
+static inline void ggml_sycl_swap(T & a, T & b) {
+    T tmp = a;
+    a = b;
+    b = tmp;
+}
+
+template <ggml_sort_order order>
+__dpct_inline__ static void
+k_argsort_f32_i32(const float *x, int *dst, const int ncols, int ncols_pad,
+                  const sycl::nd_item<3> &item_ct1, uint8_t *dpct_local) {
+    // bitonic sort
+    int col = item_ct1.get_local_id(2);
+    int row = item_ct1.get_group(1);
+
+    if (col >= ncols_pad) {
+        return;
+    }
+
+    const float * x_row = x + row * ncols;
+    auto dst_row = (int *)dpct_local;
+
+    // initialize indices
+    dst_row[col] = col;
+
+    item_ct1.barrier(sycl::access::fence_space::local_space);
+
+    for (int k = 2; k <= ncols_pad; k *= 2) {
+        for (int j = k / 2; j > 0; j /= 2) {
+            int ixj = col ^ j;
+            if (ixj > col) {
+                if ((col & k) == 0) {
+                    if (dst_row[col] >= ncols ||
+                        (dst_row[ixj] < ncols && (order == GGML_SORT_ORDER_ASC ?
+                            x_row[dst_row[col]] > x_row[dst_row[ixj]] :
+                            x_row[dst_row[col]] < x_row[dst_row[ixj]]))
+                    ) {
+                        ggml_sycl_swap(dst_row[col], dst_row[ixj]);
+                    }
+                } else {
+                    if (dst_row[ixj] >= ncols ||
+                        (dst_row[col] < ncols && (order == GGML_SORT_ORDER_ASC ?
+                            x_row[dst_row[col]] < x_row[dst_row[ixj]] :
+                            x_row[dst_row[col]] > x_row[dst_row[ixj]]))
+                    ) {
+                        ggml_sycl_swap(dst_row[col], dst_row[ixj]);
+                    }
+                }
+            }
+            /*
+            DPCT1118:1: SYCL group functions and algorithms must be encountered
+            in converged control flow. You may need to adjust the code.
+            */
+            item_ct1.barrier(sycl::access::fence_space::local_space);
+        }
+    }
+
+    // copy the result to dst without the padding
+    if (col < ncols) {
+        dst[row * ncols + col] = dst_row[col];
+    }
+}
+
+
+static void diag_mask_inf_f32(const float * x, float * dst, const int ncols, const int rows_per_channel, const int n_past,
+                              const sycl::nd_item<3> &item_ct1) {
+    const int col = item_ct1.get_local_range(1) * item_ct1.get_group(1) +
+                    item_ct1.get_local_id(1);
+    const int row = item_ct1.get_local_range(2) * item_ct1.get_group(2) +
+                    item_ct1.get_local_id(2);
+
+    if (col >= ncols) {
+        return;
+    }
+
+    const int i = row*ncols + col;
+    //dst[i] = col > (n_past + row % rows_per_channel) ? -INFINITY : x[i];
+    //dst[i] = x[i] - (col > n_past + row % rows_per_channel) * INT_MAX; // equivalent within rounding error but slightly faster on GPU
+    dst[i] = x[i] - (col > n_past + row % rows_per_channel) * FLT_MAX;
+}
+
+static void scale_f32(const float * x, float * dst, const float scale, const int k,
+                      const sycl::nd_item<3> &item_ct1) {
+    const int i = item_ct1.get_local_range(2) * item_ct1.get_group(2) +
+                  item_ct1.get_local_id(2);
+
+    if (i >= k) {
+        return;
+    }
+
+    dst[i] = scale * x[i];
+}
+
+static void clamp_f32(const float * x, float * dst, const float min, const float max, const int k,
+                      const sycl::nd_item<3> &item_ct1) {
+    const int i = item_ct1.get_local_range(2) * item_ct1.get_group(2) +
+                  item_ct1.get_local_id(2);
+
+    if (i >= k) {
+        return;
+    }
+
+    dst[i] = x[i] < min ? min : (x[i] > max ? max : x[i]);
+}
+
+template <typename Ti, typename To>
+static  void pool2d_nchw_kernel(
+        const int ih, const int iw, const int oh, const int ow,
+        const int kh, const int kw, const int sh, const int sw,
+        const int ph, const int pw, const int parallel_elements,
+        const Ti* src, To* dst, const enum ggml_op_pool op,
+        const sycl::nd_item<3> &item_ct1) {
+        int idx = item_ct1.get_local_id(2) +
+                  item_ct1.get_group(2) * item_ct1.get_local_range(2);
+        if (idx >= parallel_elements) {
+            return;
+        }
+
+        const int I_HW = ih * iw;
+        const int O_HW = oh * ow;
+        const int nc = idx / O_HW;
+        const int cur_oh = idx % O_HW / ow;
+        const int cur_ow = idx % O_HW % ow;
+        const Ti* i_ptr = src + nc * I_HW;
+        To* o_ptr = dst + nc * O_HW;
+        const int start_h = cur_oh * sh - ph;
+        const int bh = sycl::max(0, start_h);
+        const int eh = sycl::min(ih, start_h + kh);
+        const int start_w = cur_ow * sw - pw;
+        const int bw = sycl::max(0, start_w);
+        const int ew = sycl::min(iw, start_w + kw);
+
+        To res = 0;
+
+        switch (op) {
+            case GGML_OP_POOL_AVG: res = 0; break;
+            case GGML_OP_POOL_MAX: res = -FLT_MAX; break;
+        }
+
+        for (int i = bh; i < eh; i += 1) {
+            for (int j = bw; j < ew; j += 1) {
+#if DPCT_COMPATIBILITY_TEMP >= 350
+                /*
+                DPCT1098:106: The '*' expression is used instead of the __ldg
+                call. These two expressions do not provide the exact same
+                functionality. Check the generated code for potential precision
+                and/or performance issues.
+                */
+                Ti cur = *(i_ptr + i * iw + j);
+#else
+                Ti cur = i_ptr[i * iw + j];
+#endif
+                switch (op) {
+                    case GGML_OP_POOL_AVG: res += (cur / (kh * kw)); break;
+                    case GGML_OP_POOL_MAX: res = sycl::max(res, (To)cur); break;
+                }
+            }
+        }
+        o_ptr[cur_oh * ow + cur_ow] = res;
+}
+
+template <int qk, int qr, dequantize_kernel_t dq>
+static void get_rows_sycl(ggml_backend_sycl_context & ctx, const ggml_tensor *src0, const ggml_tensor *src1,
+                          ggml_tensor *dst, const void *src0_dd,
+                          const int32_t *src1_dd, float *dst_dd,
+                          queue_ptr stream) {
+
+    GGML_TENSOR_BINARY_OP_LOCALS
+
+    const sycl::range<3> block_dims(1, 1, SYCL_GET_ROWS_BLOCK_SIZE);
+    const int block_num_x = (ne00 + 2*SYCL_GET_ROWS_BLOCK_SIZE - 1) / (2*SYCL_GET_ROWS_BLOCK_SIZE);
+    const sycl::range<3> block_nums(ne11 * ne12, ne10, block_num_x);
+
+    // strides in elements
+    //const size_t s0 = nb0 / ggml_element_size(dst);
+    const size_t s1 = nb1 / ggml_element_size(dst);
+    const size_t s2 = nb2 / ggml_element_size(dst);
+    const size_t s3 = nb3 / ggml_element_size(dst);
+
+    const size_t s10 = nb10 / ggml_element_size(src1);
+    const size_t s11 = nb11 / ggml_element_size(src1);
+    const size_t s12 = nb12 / ggml_element_size(src1);
+    //const size_t s13 = nb13 / ggml_element_size(src1);
+
+    GGML_ASSERT(ne00 % 2 == 0);
+
+    stream->parallel_for(sycl::nd_range<3>(block_nums * block_dims, block_dims),
+                         [=](sycl::nd_item<3> item_ct1) {
+                             k_get_rows<qk, qr, dq>(
+                                 src0_dd, src1_dd, dst_dd, ne00, ne12, s1, s2,
+                                 s3, nb01, nb02, nb03, s10, s11, s12, item_ct1);
+                         });
+
+    (void) dst;
+}
+
+template <typename src0_t>
+static void get_rows_sycl_float(ggml_backend_sycl_context & ctx, const ggml_tensor *src0,
+                                const ggml_tensor *src1, ggml_tensor *dst,
+                                const src0_t *src0_dd, const int32_t *src1_dd,
+                                float *dst_dd, queue_ptr stream) {
+
+    GGML_TENSOR_BINARY_OP_LOCALS
+
+    const sycl::range<3> block_dims(1, 1, SYCL_GET_ROWS_BLOCK_SIZE);
+    const int block_num_x = (ne00 + SYCL_GET_ROWS_BLOCK_SIZE - 1) / SYCL_GET_ROWS_BLOCK_SIZE;
+    const sycl::range<3> block_nums(ne11 * ne12, ne10, block_num_x);
+
+    // strides in elements
+    //const size_t s0 = nb0 / ggml_element_size(dst);
+    const size_t s1 = nb1 / ggml_element_size(dst);
+    const size_t s2 = nb2 / ggml_element_size(dst);
+    const size_t s3 = nb3 / ggml_element_size(dst);
+
+    const size_t s10 = nb10 / ggml_element_size(src1);
+    const size_t s11 = nb11 / ggml_element_size(src1);
+    const size_t s12 = nb12 / ggml_element_size(src1);
+    //const size_t s13 = nb13 / ggml_element_size(src1);
+
+    {
+        dpct::has_capability_or_fail(stream->get_device(),
+                                     {sycl::aspect::fp16});
+
+        stream->parallel_for(
+            sycl::nd_range<3>(block_nums * block_dims, block_dims),
+            [=](sycl::nd_item<3> item_ct1) {
+                k_get_rows_float(src0_dd, src1_dd, dst_dd, ne00, ne12, s1, s2,
+                                 s3, nb01, nb02, nb03, s10, s11, s12, item_ct1);
+            });
+    }
+
+    (void) dst;
+}
+
+template<float (*bin_op)(const float, const float)>
+struct bin_bcast_sycl {
+    template <typename src0_t, typename src1_t, typename dst_t>
+    void operator()(ggml_backend_sycl_context & ctx,
+                    const struct ggml_tensor *src0,
+                    const struct ggml_tensor *src1, struct ggml_tensor *dst,
+                    const src0_t *src0_dd, const src1_t *src1_dd, dst_t *dst_dd,
+                    queue_ptr stream) {
+
+        GGML_TENSOR_BINARY_OP_LOCALS
+
+        int nr0 = ne10/ne0;
+        int nr1 = ne11/ne1;
+        int nr2 = ne12/ne2;
+        int nr3 = ne13/ne3;
+
+        int nr[4] = { nr0, nr1, nr2, nr3 };
+
+        // collapse dimensions until first broadcast dimension
+        int64_t cne0[] = {ne0, ne1, ne2, ne3};
+        int64_t cne1[] = {ne10, ne11, ne12, ne13};
+        size_t cnb0[] = {nb0, nb1, nb2, nb3};
+        size_t cnb1[] = {nb10, nb11, nb12, nb13};
+        auto collapse = [](int64_t cne[]) {
+            cne[0] *= cne[1];
+            cne[1] = cne[2];
+            cne[2] = cne[3];
+            cne[3] = 1;
+        };
+
+        auto collapse_nb = [](size_t cnb[], int64_t cne[]) {
+            cnb[1] *= cne[1];
+            cnb[2] *= cne[2];
+            cnb[3] *= cne[3];
+        };
+
+        for (int i = 0; i < 4; i++) {
+            if (nr[i] != 1) {
+                break;
+            }
+            if (i > 0) {
+                collapse_nb(cnb0, cne0);
+                collapse_nb(cnb1, cne1);
+                collapse(cne0);
+                collapse(cne1);
+            }
+        }
+        {
+            int64_t ne0 = cne0[0];
+            int64_t ne1 = cne0[1];
+            int64_t ne2 = cne0[2];
+            int64_t ne3 = cne0[3];
+
+            int64_t ne10 = cne1[0];
+            int64_t ne11 = cne1[1];
+            int64_t ne12 = cne1[2];
+            int64_t ne13 = cne1[3];
+
+            size_t nb0 = cnb0[0];
+            size_t nb1 = cnb0[1];
+            size_t nb2 = cnb0[2];
+            size_t nb3 = cnb0[3];
+
+            size_t nb10 = cnb1[0];
+            size_t nb11 = cnb1[1];
+            size_t nb12 = cnb1[2];
+            size_t nb13 = cnb1[3];
+
+            size_t s0 = nb0 / sizeof(dst_t);
+            size_t s1 = nb1 / sizeof(dst_t);
+            size_t s2 = nb2 / sizeof(dst_t);
+            size_t s3 = nb3 / sizeof(dst_t);
+
+            size_t s10 = nb10 / sizeof(src1_t);
+            size_t s11 = nb11 / sizeof(src1_t);
+            size_t s12 = nb12 / sizeof(src1_t);
+            size_t s13 = nb13 / sizeof(src1_t);
+
+            GGML_ASSERT(s0 == 1);
+            GGML_ASSERT(s10 == 1);
+
+            const int block_size = 128;
+
+            int64_t hne0 = std::max(ne0/2LL, 1LL);
+
+            sycl::range<3> block_dims(1, 1, 1);
+            block_dims[2] = std::min<unsigned int>(hne0, block_size);
+            block_dims[1] = std::min<unsigned int>(
+                ne1, block_size / (unsigned int)block_dims[2]);
+            block_dims[0] = std::min(
+                std::min<unsigned int>(
+                    ne2 * ne3, block_size / (unsigned int)block_dims[2] /
+                                   (unsigned int)block_dims[1]),
+                64U);
+
+            sycl::range<3> block_nums(
+                (ne2 * ne3 + block_dims[0] - 1) / block_dims[0],
+                (ne1 + block_dims[1] - 1) / block_dims[1],
+                (hne0 + block_dims[2] - 1) / block_dims[2]);
+
+            if (block_nums[0] > 65535) {
+                // this is the maximum number of blocks in z direction, fallback to 1D grid kernel
+                int block_num = (ne0*ne1*ne2*ne3 + block_size - 1) / block_size;
+                {
+                    dpct::has_capability_or_fail(stream->get_device(),
+                                                 {sycl::aspect::fp16});
+
+                    stream->parallel_for(
+                        sycl::nd_range<3>(sycl::range<3>(1, 1, block_num) *
+                                              sycl::range<3>(1, 1, block_size),
+                                          sycl::range<3>(1, 1, block_size)),
+                        [=](sycl::nd_item<3> item_ct1) {
+                            k_bin_bcast_unravel<bin_op>(
+                                src0_dd, src1_dd, dst_dd, ne0, ne1, ne2, ne3,
+                                ne10, ne11, ne12, ne13, s1, s2, s3, s11, s12,
+                                s13, item_ct1);
+                        });
+                }
+            } else {
+                /*
+                DPCT1049:16: The work-group size passed to the SYCL kernel may
+                exceed the limit. To get the device limit, query
+                info::device::max_work_group_size. Adjust the work-group size if
+                needed.
+                */
+                dpct::has_capability_or_fail(stream->get_device(),
+                                             {sycl::aspect::fp16});
+
+                stream->parallel_for(
+                    sycl::nd_range<3>(block_nums * block_dims, block_dims),
+                    [=](sycl::nd_item<3> item_ct1) {
+                        k_bin_bcast<bin_op>(src0_dd, src1_dd, dst_dd, ne0, ne1,
+                                            ne2, ne3, ne10, ne11, ne12, ne13,
+                                            s1, s2, s3, s11, s12, s13,
+                                            item_ct1);
+                    });
+            }
+        }
+    }
+};
+
+static void acc_f32_sycl(const float *x, const float *y, float *dst,
+                         const int n_elements, const int ne10, const int ne11,
+                         const int ne12, const int nb1, const int nb2,
+                         const int offset, queue_ptr stream) {
+    int num_blocks = (n_elements + SYCL_ACC_BLOCK_SIZE - 1) / SYCL_ACC_BLOCK_SIZE;
+    stream->parallel_for(
+        sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) *
+                              sycl::range<3>(1, 1, SYCL_ACC_BLOCK_SIZE),
+                          sycl::range<3>(1, 1, SYCL_ACC_BLOCK_SIZE)),
+        [=](sycl::nd_item<3> item_ct1) {
+            acc_f32(x, y, dst, n_elements, ne10, ne11, ne12, nb1, nb2, offset,
+                    item_ct1);
+        });
+}
+
+static void gelu_f32_sycl(const float *x, float *dst, const int k,
+                          queue_ptr stream) {
+    const int num_blocks = (k + SYCL_GELU_BLOCK_SIZE - 1) / SYCL_GELU_BLOCK_SIZE;
+    stream->parallel_for(
+        sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) *
+                              sycl::range<3>(1, 1, SYCL_GELU_BLOCK_SIZE),
+                          sycl::range<3>(1, 1, SYCL_GELU_BLOCK_SIZE)),
+        [=](sycl::nd_item<3> item_ct1) {
+            gelu_f32(x, dst, k, item_ct1);
+        });
+}
+
+static void silu_f32_sycl(const float *x, float *dst, const int k,
+                          queue_ptr stream) {
+    const int num_blocks = (k + SYCL_SILU_BLOCK_SIZE - 1) / SYCL_SILU_BLOCK_SIZE;
+    stream->parallel_for(
+        sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) *
+                              sycl::range<3>(1, 1, SYCL_SILU_BLOCK_SIZE),
+                          sycl::range<3>(1, 1, SYCL_SILU_BLOCK_SIZE)),
+        [=](sycl::nd_item<3> item_ct1) {
+            silu_f32(x, dst, k, item_ct1);
+        });
+}
+
+static void gelu_quick_f32_sycl(const float *x, float *dst, const int k,
+                                queue_ptr stream) {
+    const int num_blocks = (k + SYCL_GELU_BLOCK_SIZE - 1) / SYCL_GELU_BLOCK_SIZE;
+    stream->parallel_for(
+        sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) *
+                              sycl::range<3>(1, 1, SYCL_GELU_BLOCK_SIZE),
+                          sycl::range<3>(1, 1, SYCL_GELU_BLOCK_SIZE)),
+        [=](sycl::nd_item<3> item_ct1) {
+            gelu_quick_f32(x, dst, k, item_ct1);
+        });
+}
+
+static void tanh_f32_sycl(const float *x, float *dst, const int k,
+                          queue_ptr stream) {
+    const int num_blocks = (k + SYCL_TANH_BLOCK_SIZE - 1) / SYCL_TANH_BLOCK_SIZE;
+    stream->parallel_for(
+        sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) *
+                              sycl::range<3>(1, 1, SYCL_TANH_BLOCK_SIZE),
+                          sycl::range<3>(1, 1, SYCL_TANH_BLOCK_SIZE)),
+        [=](sycl::nd_item<3> item_ct1) {
+            tanh_f32(x, dst, k, item_ct1);
+        });
+}
+
+static void relu_f32_sycl(const float *x, float *dst, const int k,
+                          queue_ptr stream) {
+    const int num_blocks = (k + SYCL_RELU_BLOCK_SIZE - 1) / SYCL_RELU_BLOCK_SIZE;
+    stream->parallel_for(
+        sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) *
+                              sycl::range<3>(1, 1, SYCL_RELU_BLOCK_SIZE),
+                          sycl::range<3>(1, 1, SYCL_RELU_BLOCK_SIZE)),
+        [=](sycl::nd_item<3> item_ct1) {
+            relu_f32(x, dst, k, item_ct1);
+        });
+}
+
+static void hardsigmoid_f32_sycl(const float *x, float *dst, const int k,
+                                 queue_ptr stream) {
+    const int num_blocks = (k + SYCL_HARDSIGMOID_BLOCK_SIZE - 1) / SYCL_HARDSIGMOID_BLOCK_SIZE;
+    stream->parallel_for(
+        sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) *
+                              sycl::range<3>(1, 1, SYCL_HARDSIGMOID_BLOCK_SIZE),
+                          sycl::range<3>(1, 1, SYCL_HARDSIGMOID_BLOCK_SIZE)),
+        [=](sycl::nd_item<3> item_ct1) {
+            hardsigmoid_f32(x, dst, k, item_ct1);
+        });
+}
+
+static void hardswish_f32_sycl(const float *x, float *dst, const int k,
+                               queue_ptr stream) {
+    const int num_blocks = (k + SYCL_HARDSWISH_BLOCK_SIZE - 1) / SYCL_HARDSWISH_BLOCK_SIZE;
+    stream->parallel_for(
+        sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) *
+                              sycl::range<3>(1, 1, SYCL_HARDSWISH_BLOCK_SIZE),
+                          sycl::range<3>(1, 1, SYCL_HARDSWISH_BLOCK_SIZE)),
+        [=](sycl::nd_item<3> item_ct1) {
+            hardswish_f32(x, dst, k, item_ct1);
+        });
+}
+
+static void leaky_relu_f32_sycl(const float *x, float *dst, const int k,
+                                const float negative_slope,
+                                queue_ptr stream) {
+    const int num_blocks = (k + SYCL_RELU_BLOCK_SIZE - 1) / SYCL_RELU_BLOCK_SIZE;
+    stream->parallel_for(
+        sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) *
+                              sycl::range<3>(1, 1, SYCL_RELU_BLOCK_SIZE),
+                          sycl::range<3>(1, 1, SYCL_RELU_BLOCK_SIZE)),
+        [=](sycl::nd_item<3> item_ct1) {
+            leaky_relu_f32(x, dst, k, negative_slope, item_ct1);
+        });
+}
+
+static void sqr_f32_sycl(const float *x, float *dst, const int k,
+                         queue_ptr stream) {
+    const int num_blocks = (k + SYCL_SQR_BLOCK_SIZE - 1) / SYCL_SQR_BLOCK_SIZE;
+    stream->parallel_for(
+        sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) *
+                              sycl::range<3>(1, 1, SYCL_SQR_BLOCK_SIZE),
+                          sycl::range<3>(1, 1, SYCL_SQR_BLOCK_SIZE)),
+        [=](sycl::nd_item<3> item_ct1) {
+            sqr_f32(x, dst, k, item_ct1);
+        });
+}
+
+static void upscale_f32_sycl(const float *x, float *dst, const int nb00, const int nb01,
+                             const int nb02, const int nb03, const int ne10, const int ne11,
+                             const int ne12, const int ne13, const float sf0, const float sf1,
+                             const float sf2, const float sf3, queue_ptr stream) {
+    int dst_size = ne10 * ne11 * ne12 * ne13;
+    int num_blocks = (dst_size + SYCL_UPSCALE_BLOCK_SIZE - 1) / SYCL_UPSCALE_BLOCK_SIZE;
+    sycl::range<1> gridDim(num_blocks * SYCL_UPSCALE_BLOCK_SIZE);
+    stream->parallel_for(
+        sycl::nd_range<1>(gridDim, sycl::range<1>(SYCL_UPSCALE_BLOCK_SIZE)),
+        [=](sycl::nd_item<1> item_ct1) {
+            upscale_f32(x, dst, nb00, nb01, nb02, nb03, ne10, ne11, ne12, ne13, sf0, sf1, sf2, sf3, item_ct1);
+        });
+}
+
+static void pad_f32_sycl(const float *x, float *dst, const int ne00,
+                         const int ne01, const int ne02, const int ne0,
+                         const int ne1, const int ne2, queue_ptr stream) {
+    int num_blocks = (ne0 + SYCL_PAD_BLOCK_SIZE - 1) / SYCL_PAD_BLOCK_SIZE;
+    sycl::range<3> gridDim(ne2, ne1, num_blocks);
+    stream->parallel_for(
+        sycl::nd_range<3>(gridDim * sycl::range<3>(1, 1, SYCL_PAD_BLOCK_SIZE),
+                          sycl::range<3>(1, 1, SYCL_PAD_BLOCK_SIZE)),
+        [=](sycl::nd_item<3> item_ct1) {
+            pad_f32(x, dst, ne0, ne00, ne01, ne02, item_ct1);
+        });
+}
+
+static void quantize_row_q8_1_sycl(const float *x, void *vy, const int kx,
+                                   const int ky, const int kx_padded,
+                                   queue_ptr stream) {
+    const int block_num_x = (kx_padded + SYCL_QUANTIZE_BLOCK_SIZE - 1) / SYCL_QUANTIZE_BLOCK_SIZE;
+    const sycl::range<3> num_blocks(1, ky, block_num_x);
+    int constexpr QUANT_BLOCK_TILE = QK8_1 / WARP_SIZE;
+    static_assert(QK8_1 % WARP_SIZE == 0);
+    const sycl::range<3> block_size(1, 1, SYCL_QUANTIZE_BLOCK_SIZE / QUANT_BLOCK_TILE);
+    {
+        dpct::has_capability_or_fail(stream->get_device(),
+                                     {sycl::aspect::fp16});
+
+        stream->parallel_for(
+            sycl::nd_range<3>(num_blocks * block_size, block_size),
+            [=](sycl::nd_item<3> item_ct1) [[intel::reqd_sub_group_size(WARP_SIZE)]] {
+                quantize_q8_1<QUANT_BLOCK_TILE>(x, vy, kx, kx_padded, item_ct1);
+            });
+    }
+}
+
+static void ggml_mul_mat_p021_f16_f32_sycl(const void *vx, const float *y,
+                                           float *dst, const int ncols_x,
+                                           const int nrows_x,
+                                           const int nchannels_x,
+                                           const int nchannels_y,
+                                           queue_ptr stream) {
+
+    const sycl::range<3> block_nums(nchannels_y, nrows_x, 1);
+    const sycl::range<3> block_dims(1, 1, WARP_SIZE);
+    {
+        dpct::has_capability_or_fail(stream->get_device(),
+                                     {sycl::aspect::fp16});
+
+        stream->parallel_for(
+            sycl::nd_range<3>(block_nums * block_dims, block_dims),
+            [=](sycl::nd_item<3> item_ct1) [[intel::reqd_sub_group_size(WARP_SIZE)]] {
+                mul_mat_p021_f16_f32(vx, y, dst, ncols_x, nrows_x, nchannels_x,
+                                     nchannels_y, item_ct1);
+            });
+    }
+}
+
+static void ggml_mul_mat_vec_nc_f16_f32_sycl(
+    const void *vx, const float *y, float *dst, const int ncols_x,
+    const int nrows_x, const int row_stride_x, const int nchannels_x,
+    const int nchannels_y, const int channel_stride_x, queue_ptr stream) {
+
+    const sycl::range<3> block_nums(nchannels_y, nrows_x, 1);
+    const sycl::range<3> block_dims(1, 1, WARP_SIZE);
+    {
+        dpct::has_capability_or_fail(stream->get_device(),
+                                     {sycl::aspect::fp16});
+
+        stream->parallel_for(
+            sycl::nd_range<3>(block_nums * block_dims, block_dims),
+            [=](sycl::nd_item<3> item_ct1) [[intel::reqd_sub_group_size(WARP_SIZE)]] {
+                mul_mat_vec_nc_f16_f32(vx, y, dst, ncols_x, nrows_x,
+                                       row_stride_x, channel_stride_x,
+                                       nchannels_y / nchannels_x, item_ct1);
+            });
+    }
+}
+
+static void
+ggml_cpy_f16_f32_sycl(const char *cx, char *cdst, const int ne, const int ne00,
+                      const int ne01, const int ne02, const int nb00,
+                      const int nb01, const int nb02, const int nb03,
+                      const int ne10, const int ne11, const int ne12,
+                      const int nb10, const int nb11, const int nb12,
+                      const int nb13, queue_ptr stream) {
+
+    const int num_blocks = (ne + SYCL_CPY_BLOCK_SIZE - 1) / SYCL_CPY_BLOCK_SIZE;
+    {
+        dpct::has_capability_or_fail(stream->get_device(),
+                                     {sycl::aspect::fp16});
+
+        stream->parallel_for(
+            sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) *
+                                  sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE),
+                              sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE)),
+            [=](sycl::nd_item<3> item_ct1) {
+                cpy_f32_f16<cpy_1_f16_f32>(cx, cdst, ne, ne00, ne01, ne02, nb00,
+                                           nb01, nb02, nb03, ne10, ne11, ne12,
+                                           nb10, nb11, nb12, nb13, item_ct1);
+            });
+    }
+}
+
+static void ggml_cpy_f32_f32_sycl(const char *cx, char *cdst, const int ne,
+                                  const int ne00, const int ne01,
+                                  const int ne02, const int nb00,
+                                  const int nb01, const int nb02,
+                                  const int nb03, const int ne10,
+                                  const int ne11, const int ne12,
+                                  const int nb10, const int nb11,
+                                  const int nb12, const int nb13,
+                                  queue_ptr stream) {
+
+    const int num_blocks = (ne + SYCL_CPY_BLOCK_SIZE - 1) / SYCL_CPY_BLOCK_SIZE;
+    {
+        dpct::has_capability_or_fail(stream->get_device(),
+                                     {sycl::aspect::fp16});
+
+        stream->parallel_for(
+            sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) *
+                                  sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE),
+                              sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE)),
+            [=](sycl::nd_item<3> item_ct1) {
+                cpy_f32_f16<cpy_1_f32_f32>(cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02,
+                                           nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13,
+                                           item_ct1);
+            });
+    }
+}
+
+static void ggml_cpy_f32_f16_sycl(const char *cx, char *cdst, const int ne,
+                                  const int ne00, const int ne01,
+                                  const int ne02, const int nb00,
+                                  const int nb01, const int nb02,
+                                  const int nb03, const int ne10,
+                                  const int ne11, const int ne12,
+                                  const int nb10, const int nb11,
+                                  const int nb12, const int nb13,
+                                  queue_ptr stream) {
+
+    const int num_blocks = (ne + SYCL_CPY_BLOCK_SIZE - 1) / SYCL_CPY_BLOCK_SIZE;
+    {
+        dpct::has_capability_or_fail(stream->get_device(),
+                                     {sycl::aspect::fp16});
+
+        stream->parallel_for(
+            sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) *
+                                  sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE),
+                              sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE)),
+            [=](sycl::nd_item<3> item_ct1) {
+                cpy_f32_f16<cpy_1_f32_f16>(cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02,
+                                           nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13,
+                                           item_ct1);
+            });
+    }
+}
+
+static void ggml_cpy_f32_q8_0_sycl(const char *cx, char *cdst, const int ne,
+                                   const int ne00, const int ne01,
+                                   const int ne02, const int nb00,
+                                   const int nb01, const int nb02,
+                                   const int nb03, const int ne10,
+                                   const int ne11, const int ne12,
+                                   const int nb10, const int nb11,
+                                   const int nb12, const int nb13,
+                                   queue_ptr stream) {
+
+    GGML_ASSERT(ne % QK8_0 == 0);
+    const int num_blocks = ne / QK8_0;
+    stream->parallel_for(sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks),
+                                           sycl::range<3>(1, 1, 1)),
+                         [=](sycl::nd_item<3> item_ct1) {
+                             cpy_f32_q<cpy_blck_f32_q8_0, QK8_0>(
+                                 cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02,
+                                 nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13,
+                                 item_ct1);
+                         });
+}
+
+static void ggml_cpy_f32_q4_0_sycl(const char *cx, char *cdst, const int ne,
+                                   const int ne00, const int ne01,
+                                   const int ne02, const int nb00,
+                                   const int nb01, const int nb02,
+                                   const int nb03, const int ne10,
+                                   const int ne11, const int ne12,
+                                   const int nb10, const int nb11,
+                                   const int nb12, const int nb13,
+                                   queue_ptr stream) {
+
+    GGML_ASSERT(ne % QK4_0 == 0);
+    const int num_blocks = ne / QK4_0;
+    stream->parallel_for(sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks),
+                                           sycl::range<3>(1, 1, 1)),
+                         [=](sycl::nd_item<3> item_ct1) {
+                             cpy_f32_q<cpy_blck_f32_q4_0, QK4_0>(
+                                 cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02,
+                                 nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13,
+                                 item_ct1);
+                         });
+}
+
+static void ggml_cpy_f32_q4_1_sycl(const char *cx, char *cdst, const int ne,
+                                   const int ne00, const int ne01,
+                                   const int ne02, const int nb00,
+                                   const int nb01, const int nb02,
+                                   const int nb03, const int ne10,
+                                   const int ne11, const int ne12,
+                                   const int nb10, const int nb11,
+                                   const int nb12, const int nb13,
+                                   queue_ptr stream) {
+
+    GGML_ASSERT(ne % QK4_1 == 0);
+    const int num_blocks = ne / QK4_1;
+    stream->parallel_for(sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks),
+                                           sycl::range<3>(1, 1, 1)),
+                         [=](sycl::nd_item<3> item_ct1) {
+                             cpy_f32_q<cpy_blck_f32_q4_1, QK4_1>(
+                                 cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02,
+                                 nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13,
+                                 item_ct1);
+                         });
+}
+
+static void ggml_cpy_f16_f16_sycl(const char *cx, char *cdst, const int ne,
+                                  const int ne00, const int ne01,
+                                  const int ne02, const int nb00,
+                                  const int nb01, const int nb02,
+                                  const int nb03, const int ne10,
+                                  const int ne11, const int ne12,
+                                  const int nb10, const int nb11,
+                                  const int nb12, const int nb13,
+                                  queue_ptr stream) {
+
+    const int num_blocks = (ne + SYCL_CPY_BLOCK_SIZE - 1) / SYCL_CPY_BLOCK_SIZE;
+    {
+        dpct::has_capability_or_fail(stream->get_device(),
+                                     {sycl::aspect::fp16});
+
+        stream->parallel_for(
+            sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) *
+                                  sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE),
+                              sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE)),
+            [=](sycl::nd_item<3> item_ct1) {
+                cpy_f32_f16<cpy_1_f16_f16>(cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02,
+                                           nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13,
+                                           item_ct1);
+            });
+    }
+}
+
+static void ggml_cpy_i16_i16_sycl(const char *cx, char *cdst, const int ne,
+                                  const int ne00, const int ne01,
+                                  const int ne02, const int nb00,
+                                  const int nb01, const int nb02,
+                                  const int nb03, const int ne10,
+                                  const int ne11, const int ne12,
+                                  const int nb10, const int nb11,
+                                  const int nb12, const int nb13,
+                                  queue_ptr stream) {
+
+    const int num_blocks = (ne + SYCL_CPY_BLOCK_SIZE - 1) / SYCL_CPY_BLOCK_SIZE;
+    {
+        // dpct::has_capability_or_fail(stream->get_device(),
+        //                              {sycl::aspect::fp16});
+
+        stream->parallel_for(
+            sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) *
+                                  sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE),
+                              sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE)),
+            [=](sycl::nd_item<3> item_ct1) {
+                cpy_f32_f16<cpy_1_i16_i16>(cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02,
+                                           nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13,
+                                           item_ct1);
+            });
+    }
+}
+
+static void ggml_cpy_i32_i32_sycl(const char *cx, char *cdst, const int ne,
+                                  const int ne00, const int ne01,
+                                  const int ne02, const int nb00,
+                                  const int nb01, const int nb02,
+                                  const int nb03, const int ne10,
+                                  const int ne11, const int ne12,
+                                  const int nb10, const int nb11,
+                                  const int nb12, const int nb13,
+                                  queue_ptr stream) {
+
+    const int num_blocks = (ne + SYCL_CPY_BLOCK_SIZE - 1) / SYCL_CPY_BLOCK_SIZE;
+    {
+        // dpct::has_capability_or_fail(stream->get_device(),
+        //                              {sycl::aspect::fp16});
+
+        stream->parallel_for(
+            sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) *
+                                  sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE),
+                              sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE)),
+            [=](sycl::nd_item<3> item_ct1) {
+                cpy_f32_f16<cpy_1_i32_i32>(cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02,
+                                           nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13,
+                                           item_ct1);
+            });
+    }
+}
+
+static void scale_f32_sycl(const float *x, float *dst, const float scale,
+                           const int k, queue_ptr stream) {
+    const int num_blocks = (k + SYCL_SCALE_BLOCK_SIZE - 1) / SYCL_SCALE_BLOCK_SIZE;
+    stream->parallel_for(
+        sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) *
+                              sycl::range<3>(1, 1, SYCL_SCALE_BLOCK_SIZE),
+                          sycl::range<3>(1, 1, SYCL_SCALE_BLOCK_SIZE)),
+        [=](sycl::nd_item<3> item_ct1) {
+            scale_f32(x, dst, scale, k, item_ct1);
+        });
+}
+
+static void clamp_f32_sycl(const float *x, float *dst, const float min,
+                           const float max, const int k,
+                           queue_ptr stream) {
+    const int num_blocks = (k + SYCL_CLAMP_BLOCK_SIZE - 1) / SYCL_CLAMP_BLOCK_SIZE;
+    stream->parallel_for(
+        sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) *
+                              sycl::range<3>(1, 1, SYCL_CLAMP_BLOCK_SIZE),
+                          sycl::range<3>(1, 1, SYCL_CLAMP_BLOCK_SIZE)),
+        [=](sycl::nd_item<3> item_ct1) {
+            clamp_f32(x, dst, min, max, k, item_ct1);
+        });
+}
+
+static void sum_rows_f32_sycl(const float *x, float *dst, const int ncols,
+                              const int nrows, queue_ptr stream) {
+    const sycl::range<3> block_dims(1, 1, WARP_SIZE);
+    const sycl::range<3> block_nums(1, nrows, 1);
+    stream->parallel_for(sycl::nd_range<3>(block_nums * block_dims, block_dims),
+                         [=](sycl::nd_item<3> item_ct1)
+                             [[intel::reqd_sub_group_size(WARP_SIZE)]] {
+                                 k_sum_rows_f32(x, dst, ncols, item_ct1);
+                             });
+}
+
+static int next_power_of_2(int x) {
+    int n = 1;
+    while (n < x) {
+        n *= 2;
+    }
+    return n;
+}
+
+static void argsort_f32_i32_sycl(const float *x, int *dst, const int ncols,
+                                 const int nrows, ggml_sort_order order,
+                                 queue_ptr stream) {
+    // bitonic sort requires ncols to be power of 2
+    const int ncols_pad = next_power_of_2(ncols);
+
+    const sycl::range<3> block_dims(1, 1, ncols_pad);
+    const sycl::range<3> block_nums(1, nrows, 1);
+    const size_t shared_mem = ncols_pad * sizeof(int);
+
+    if (order == GGML_SORT_ORDER_ASC) {
+        stream->submit([&](sycl::handler &cgh) {
+            sycl::local_accessor<uint8_t, 1> dpct_local_acc_ct1(
+                sycl::range<1>(shared_mem), cgh);
+
+            cgh.parallel_for(
+                sycl::nd_range<3>(block_nums * block_dims, block_dims),
+                [=](sycl::nd_item<3> item_ct1) {
+                    k_argsort_f32_i32<GGML_SORT_ORDER_ASC>(
+                        x, dst, ncols, ncols_pad, item_ct1,
+                        dpct_local_acc_ct1.get_multi_ptr<sycl::access::decorated::no>()
+                            .get());
+                });
+        });
+    } else if (order == GGML_SORT_ORDER_DESC) {
+        stream->submit([&](sycl::handler &cgh) {
+            sycl::local_accessor<uint8_t, 1> dpct_local_acc_ct1(
+                sycl::range<1>(shared_mem), cgh);
+
+            cgh.parallel_for(
+                sycl::nd_range<3>(block_nums * block_dims, block_dims),
+                [=](sycl::nd_item<3> item_ct1) {
+                    k_argsort_f32_i32<GGML_SORT_ORDER_DESC>(
+                        x, dst, ncols, ncols_pad, item_ct1,
+                        dpct_local_acc_ct1.get_multi_ptr<sycl::access::decorated::no>()
+                            .get());
+                });
+        });
+    } else {
+        GGML_ABORT("fatal error");
+    }
+}
+
+static void diag_mask_inf_f32_sycl(const float *x, float *dst,
+                                   const int ncols_x, const int nrows_x,
+                                   const int rows_per_channel, const int n_past,
+                                   queue_ptr stream) {
+    const sycl::range<3> block_dims(1, SYCL_DIAG_MASK_INF_BLOCK_SIZE, 1);
+    const int block_num_x = (ncols_x + SYCL_DIAG_MASK_INF_BLOCK_SIZE - 1) / SYCL_DIAG_MASK_INF_BLOCK_SIZE;
+    const sycl::range<3> block_nums(1, block_num_x, nrows_x);
+    stream->parallel_for(sycl::nd_range<3>(block_nums * block_dims, block_dims),
+                         [=](sycl::nd_item<3> item_ct1) {
+                             diag_mask_inf_f32(x, dst, ncols_x,
+                                               rows_per_channel, n_past,
+                                               item_ct1);
+                         });
+}
+
+static bool g_sycl_loaded = false;
+
+bool ggml_sycl_loaded(void) {
+    return g_sycl_loaded;
+}
+
+void print_device_detail(int id, sycl::device &device, std::string device_type) {
+
+    dpct::device_info prop;
+    SYCL_CHECK(CHECK_TRY_ERROR(
+        dpct::get_device_info(prop, device)));
+
+    std::string version;
+    version += std::to_string(prop.get_major_version());
+    version += ".";
+    version += std::to_string(prop.get_minor_version());
+
+    device_type = std::regex_replace(device_type, std::regex("ext_oneapi_"), "");
+    std::string name = std::string(prop.get_name());
+    name = std::regex_replace(name, std::regex("\\(R\\)"), "");
+    name = std::regex_replace(name, std::regex("\\(TM\\)"), "");
+
+    auto global_mem_size = prop.get_global_mem_size()/1000000;
+
+    fprintf(stderr, "|%2d|%19s|%39s|%7s|%7d|%8d|%5d|%6luM|%21s|\n", id, device_type.c_str(),
+            name.c_str(), version.c_str(), prop.get_max_compute_units(),
+            prop.get_max_work_group_size(), prop.get_max_sub_group_size(),
+            global_mem_size, device.get_info<sycl::info::device::driver_version>().c_str());
+}
+
+void ggml_backend_sycl_print_sycl_devices() {
+    GGML_SYCL_DEBUG("[SYCL] call ggml_backend_sycl_print_sycl_devices\n");
+    int device_count = dpct::dev_mgr::instance().device_count();
+    std::map<std::string, size_t> DeviceNums;
+    fprintf(stderr, "found %d SYCL devices:\n", device_count);
+    fprintf(stderr, "|  |                   |                                       |       |Max    |        |Max  |Global |                     |\n");
+    fprintf(stderr, "|  |                   |                                       |       |compute|Max work|sub  |mem    |                     |\n");
+    fprintf(stderr, "|ID|        Device Type|                                   Name|Version|units  |group   |group|size   |       Driver version|\n");
+    fprintf(stderr, "|--|-------------------|---------------------------------------|-------|-------|--------|-----|-------|---------------------|\n");
+    for (int id = 0; id < device_count; ++id) {
+        sycl::device device = dpct::dev_mgr::instance().get_device(id);
+        sycl::backend backend = device.get_backend();
+        std::string backend_type = get_device_backend_and_type(device);
+        int type_id=DeviceNums[backend_type]++;
+        std::stringstream device_type;
+        device_type << "[" <<  backend_type << ":" << std::to_string(type_id) << "]";
+        print_device_detail(id, device, device_type.str());
+    }
+}
+
+static inline int get_sycl_env(const char *env_name, int default_val) {
+    char *user_device_string = getenv(env_name);
+    int user_number = default_val;
+
+    unsigned n;
+    if (user_device_string != NULL &&
+        sscanf(user_device_string, " %u", &n) == 1) {
+        user_number = (int)n;
+    } else {
+        user_number = default_val;
+    }
+    return user_number;
+}
+
+static void ggml_check_sycl() try {
+    static bool initialized = false;
+
+    if (!initialized) {
+        fprintf(stderr, "[SYCL] call ggml_check_sycl\n");
+        g_ggml_sycl_debug = get_sycl_env("GGML_SYCL_DEBUG", 0);
+
+        fprintf(stderr, "%s: GGML_SYCL_DEBUG: %d\n", __func__, g_ggml_sycl_debug);
+
+#if defined(GGML_SYCL_F16)
+        fprintf(stderr, "%s: GGML_SYCL_F16: yes\n", __func__);
+#else
+        fprintf(stderr, "%s: GGML_SYCL_F16: no\n", __func__);
+#endif
+
+/* NOT REMOVE, keep it for next optimize for XMX.
+#if defined(SYCL_USE_XMX)
+        fprintf(stderr, "%s: SYCL_USE_XMX: yes\n", __func__);
+#else
+        fprintf(stderr, "%s: SYCL_USE_XMX: no\n", __func__);
+#endif
+*/
+
+        if (CHECK_TRY_ERROR(g_all_sycl_device_count =
+                            dpct::dev_mgr::instance().device_count()) != 0) {
+            initialized = true;
+            g_sycl_loaded = false;
+            return;
+        }
+        GGML_ASSERT(g_all_sycl_device_count <= GGML_SYCL_MAX_DEVICES);
+        ggml_backend_sycl_print_sycl_devices();
+        initialized = true;
+        g_sycl_loaded = true;
+    }
+}
+catch (sycl::exception const &exc) {
+  std::cerr << exc.what() << "Exception caught at file:" << __FILE__
+            << ", line:" << __LINE__ << std::endl;
+  std::exit(1);
+}
+
+static ggml_sycl_device_info ggml_sycl_init() {
+    ggml_sycl_device_info info = {};
+
+    info.device_count = dpct::dev_mgr::instance().device_count();
+    if (info.device_count == 0) {
+        fprintf(stderr, "%s: failed to initialize " GGML_SYCL_NAME ": %s\n", __func__);
+        return info;
+    }
+
+    GGML_ASSERT(info.device_count <= GGML_SYCL_MAX_DEVICES);
+
+    int64_t total_vram = 0;
+#if defined(GGML_SYCL_FORCE_MMQ)
+    fprintf(stderr, "%s: GGML_SYCL_FORCE_MMQ:   yes\n", __func__);
+#else
+    fprintf(stderr, "%s: GGML_SYCL_FORCE_MMQ:   no\n", __func__);
+#endif
+#if defined(SYCL_USE_XMX)
+    fprintf(stderr, "%s: SYCL_USE_XMX: yes\n", __func__);
+#else
+    fprintf(stderr, "%s: SYCL_USE_XMX: no\n", __func__);
+#endif
+    fprintf(stderr, "%s: found %d " GGML_SYCL_NAME " devices:\n", __func__, info.device_count);
+
+    for (int i = 0; i < info.device_count; ++i) {
+        info.devices[i].vmm = 0;
+        dpct::device_info prop;
+        SYCL_CHECK(CHECK_TRY_ERROR(dpct::get_device_info(
+            prop, dpct::dev_mgr::instance().get_device(i))));
+
+        info.default_tensor_split[i] = total_vram;
+        total_vram += prop.get_global_mem_size();
+
+        info.devices[i].cc =
+            100 * prop.get_major_version() + 10 * prop.get_minor_version();
+
+        info.max_work_group_sizes[i] = prop.get_max_work_group_size();
+    }
+
+    for (int id = 0; id < info.device_count; ++id) {
+        info.default_tensor_split[id] /= total_vram;
+    }
+    return info;
+}
+
+const ggml_sycl_device_info & ggml_sycl_info() {
+    static ggml_sycl_device_info info = ggml_sycl_init();
+    return info;
+}
+
+/*
+device_index: device index from 0 to n (continue numbers).
+    It is used for device select/set in SYCL backend internal data structure.
+*/
+inline void check_allow_gpu_index(const int device_index) {
+  if (device_index >= ggml_sycl_info().device_count) {
+    char error_buf[256];
+    snprintf(
+        error_buf,
+        sizeof(error_buf),
+        "%s error: device_index:%d is out of range: [0-%d]",
+        __func__,
+        device_index,
+        ggml_sycl_info().device_count - 1);
+    fprintf(stderr, "%s\n", error_buf);
+    assert(false);
+  }
+}
+
+// buffer pool for sycl (legacy)
+struct ggml_sycl_pool_leg : public ggml_sycl_pool {
+    static const int MAX_SYCL_BUFFERS = 256;
+
+    int device;
+    queue_ptr qptr;
+    struct ggml_sycl_buffer {
+        void * ptr = nullptr;
+        size_t size = 0;
+    };
+
+    ggml_sycl_buffer buffer_pool[MAX_SYCL_BUFFERS] = {};
+    size_t pool_size = 0;
+
+    explicit ggml_sycl_pool_leg(queue_ptr qptr_, int device_) :
+        qptr(qptr_),
+        device(device_) {
+    }
+
+    ~ggml_sycl_pool_leg() {
+        for (int i = 0; i < MAX_SYCL_BUFFERS; ++i) {
+            ggml_sycl_buffer & b = buffer_pool[i];
+            if (b.ptr != nullptr) {
+                SYCL_CHECK(CHECK_TRY_ERROR(sycl::free(b.ptr, *qptr)));
+                pool_size -= b.size;
+            }
+        }
+        GGML_ASSERT(pool_size == 0);
+    }
+
+    void * alloc(size_t size, size_t * actual_size) override {
+#ifdef DEBUG_sycl_MALLOC
+        int nnz = 0;
+        size_t max_size = 0;
+#endif
+        size_t best_diff = 1ull << 36;
+        int ibest = -1;
+        for (int i = 0; i < MAX_SYCL_BUFFERS; ++i) {
+            ggml_sycl_buffer& b = buffer_pool[i];
+            if (b.ptr != nullptr) {
+#ifdef DEBUG_sycl_MALLOC
+                ++nnz;
+                if (b.size > max_size) max_size = b.size;
+#endif
+                if (b.size >= size) {
+                    size_t diff = b.size - size;
+                    if (diff < best_diff) {
+                        best_diff = diff;
+                        ibest = i;
+                        if (!best_diff) {
+                            void * ptr = b.ptr;
+                            *actual_size = b.size;
+                            b.ptr = nullptr;
+                            b.size = 0;
+                            return ptr;
+                        }
+                    }
+                }
+            }
+        }
+        if (ibest >= 0) {
+            ggml_sycl_buffer& b = buffer_pool[ibest];
+            void * ptr = b.ptr;
+            *actual_size = b.size;
+            b.ptr = nullptr;
+            b.size = 0;
+            return ptr;
+        }
+        void * ptr;
+        size_t look_ahead_size = (size_t) (1.05 * size);
+
+        SYCL_CHECK(
+            CHECK_TRY_ERROR(ptr = (void *)sycl::malloc_device(
+                                look_ahead_size, *qptr)));
+        if (!ptr) {
+            fprintf(stderr, "%s: can't malloc %lu Bytes memory on device", __func__, look_ahead_size);
+            return nullptr;
+        }
+
+        *actual_size = look_ahead_size;
+        pool_size += look_ahead_size;
+
+    #ifdef DEBUG_SYCL_MALLOC
+        fprintf(stderr, "%s[%d]: %d buffers, max_size = %u MB, pool_size = %u MB, requested %u MB\n", __func__, id, nnz,
+                (uint32_t)(max_size/1024/1024), (uint32_t)(g_sycl_pool_size[id]/1024/1024), (uint32_t)(size/1024/1024));
+    #endif
+        // GGML_SYCL_DEBUG("ggml_sycl_pool_malloc_leg look_ahead_size=%lu, return %p\n", look_ahead_size, ptr);
+        return ptr;
+    }
+
+    void free(void * ptr, size_t size) override {
+        for (int i = 0; i < MAX_SYCL_BUFFERS; ++i) {
+            ggml_sycl_buffer& b = buffer_pool[i];
+            if (b.ptr == nullptr) {
+                b.ptr = ptr;
+                b.size = size;
+                return;
+            }
+        }
+        fprintf(stderr, "WARNING: sycl buffer pool full, increase MAX_sycl_BUFFERS\n");
+        SYCL_CHECK(CHECK_TRY_ERROR(sycl::free(ptr, *qptr)));
+        pool_size -= size;
+    }
+};
+
+std::unique_ptr<ggml_sycl_pool> ggml_backend_sycl_context::new_pool_for_device(queue_ptr qptr, int device) {
+    // TBD: NO VMM support
+    // if (ggml_sycl_info().devices[device].vmm) {
+    //     return std::unique_ptr<ggml_sycl_pool>(new ggml_sycl_pool_vmm(device));
+    // }
+   return std::unique_ptr<ggml_sycl_pool>(new ggml_sycl_pool_leg(qptr, device));
+}
+
+// TBD pool with virtual memory management
+// struct ggml_sycl_pool_vmm : public ggml_sycl_pool
+
+static dpct::err0 ggml_sycl_cpy_tensor_2d(void *dst,
+                                          const struct ggml_tensor *src,
+                                          int64_t i3, int64_t i2,
+                                          int64_t i1_low, int64_t i1_high,
+                                          queue_ptr stream) try {
+
+    dpct::memcpy_direction kind;
+    char * src_ptr;
+    if (src->backend == GGML_BACKEND_TYPE_CPU) {
+        kind = dpct::host_to_device;
+        src_ptr = (char *) src->data;
+        // GGML_SYCL_DEBUG("ggml_sycl_cpy_tensor_2d  GGML_BACKEND_TYPE_CPU src_ptr %p\n", src_ptr);
+    } else if (src->backend == GGML_BACKEND_TYPE_GPU || src->backend == GGML_BACKEND_TYPE_GPU_SPLIT) {
+        GGML_ASSERT(src->backend != GGML_BACKEND_TYPE_GPU_SPLIT || (i1_low == 0 && i1_high == src->ne[1]));
+        kind = dpct::device_to_device;
+        ggml_tensor_extra_gpu * extra = (ggml_tensor_extra_gpu *) src->extra;
+        int id;
+        SYCL_CHECK(CHECK_TRY_ERROR(
+            id = get_current_device_id()));
+        // GGML_SYCL_DEBUG("current device index %d\n", id);
+        src_ptr = (char *) extra->data_device[id];
+    } else {
+        // GGML_SYCL_DEBUG("GGML_ABORT("fatal error")\n");
+        GGML_ABORT("fatal error");
+    }
+    char * dst_ptr = (char *) dst;
+
+    GGML_TENSOR_LOCALS_1(int64_t, ne, src, ne);
+    GGML_TENSOR_LOCALS(int64_t, nb, src, nb);
+    const enum ggml_type type = src->type;
+    const int64_t ts = ggml_type_size(type);
+    const int64_t bs = ggml_blck_size(type);
+    int64_t i1_diff = i1_high - i1_low;
+
+    const char * x = src_ptr + i1_low*nb1 + i2*nb2 + i3*nb3;
+    if (nb0 == ts && nb1 == ts*ne0/bs) {
+        // GGML_SYCL_DEBUG("stream->memcpy: dst_ptr=%p, x=%p, size=%lu\n", dst_ptr, x, i1_diff * nb1);
+        // return CHECK_TRY_ERROR(stream->memcpy(dst_ptr, x, i1_diff * nb1));
+        return CHECK_TRY_ERROR(dpct::async_dpct_memcpy(dst_ptr, x, i1_diff * nb1,
+                                    kind, *stream));
+
+    } else if (nb0 == ts) {
+        return CHECK_TRY_ERROR(
+            dpct::async_dpct_memcpy(dst_ptr, ts * ne0 / bs, x, nb1,
+                                    ts * ne0 / bs, i1_diff, kind, *stream));
+    } else {
+        for (int64_t i1 = 0; i1 < i1_diff; i1++) {
+            const void * rx = (const void *) ((const char *) x + i1*nb1);
+            void * rd = (void *) (dst_ptr + i1*ts*ne0/bs);
+            // pretend the row is a matrix with cols=1
+            dpct::err0 r = CHECK_TRY_ERROR(dpct::async_dpct_memcpy(
+                rd, ts / bs, rx, nb0, ts / bs, ne0, kind, *stream));
+            /*
+            DPCT1001:85: The statement could not be removed.
+            */
+            /*
+            DPCT1000:86: Error handling if-stmt was detected but could not be
+            rewritten.
+            */
+            if (r != 0) return r;
+        }
+        return 0;
+    }
+}
+catch (sycl::exception const &exc) {
+  std::cerr << exc.what() << "Exception caught at file:" << __FILE__
+            << ", line:" << __LINE__ << std::endl;
+  std::exit(1);
+}
+
+static void ggml_sycl_op_get_rows(ggml_backend_sycl_context & ctx, const ggml_tensor *src0,
+                                  const ggml_tensor *src1, ggml_tensor *dst,
+                                  const float *src0_d, const float *src1_d,
+                                  float *dst_d, const queue_ptr &stream) {
+
+    GGML_ASSERT(src1->type == GGML_TYPE_I32);
+    GGML_ASSERT(dst->type == GGML_TYPE_F32);
+
+    GGML_ASSERT(src0->nb[0] == ggml_type_size(src0->type));
+    GGML_ASSERT(src1->nb[0] == ggml_type_size(src1->type));
+    GGML_ASSERT(dst->nb[0] == ggml_type_size(dst->type));
+
+    const int32_t * src1_i32 = (const int32_t *) src1_d;
+
+    switch (src0->type) {
+        case GGML_TYPE_F16:
+            get_rows_sycl_float(ctx, src0, src1, dst, (const sycl::half *)src0_d,
+                                src1_i32, dst_d, stream);
+            break;
+        case GGML_TYPE_F32:
+            get_rows_sycl_float(ctx, src0, src1, dst, src0_d, src1_i32, dst_d, stream);
+            break;
+        case GGML_TYPE_Q4_0:
+            get_rows_sycl<QK4_0, QR4_0, dequantize_q4_0>(ctx, src0, src1, dst, src0_d, src1_i32, dst_d, stream);
+            break;
+        case GGML_TYPE_Q4_1:
+            get_rows_sycl<QK4_1, QR4_1, dequantize_q4_1>(ctx, src0, src1, dst, src0_d, src1_i32, dst_d, stream);
+            break;
+        case GGML_TYPE_Q5_0:
+            get_rows_sycl<QK5_0, QR5_0, dequantize_q5_0>(ctx, src0, src1, dst, src0_d, src1_i32, dst_d, stream);
+            break;
+        case GGML_TYPE_Q5_1:
+            get_rows_sycl<QK5_1, QR5_1, dequantize_q5_1>(ctx, src0, src1, dst, src0_d, src1_i32, dst_d, stream);
+            break;
+        case GGML_TYPE_Q8_0:
+            get_rows_sycl<QK8_0, QR8_0, dequantize_q8_0>(ctx, src0, src1, dst, src0_d, src1_i32, dst_d, stream);
+            break;
+        default:
+            // TODO: k-quants
+            fprintf(stderr, "%s: unsupported type: %s\n", __func__, ggml_type_name(src0->type));
+            GGML_ABORT("fatal error");
+            break;
+    }
+}
+
+template <class op>
+inline void ggml_sycl_op_bin_bcast(ggml_backend_sycl_context & ctx, const ggml_tensor *src0,
+                                   const ggml_tensor *src1, ggml_tensor *dst,
+                                   const float *src0_dd, const float *src1_dd,
+                                   float *dst_dd,
+                                   const queue_ptr &main_stream) {
+
+    if (src0->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32) {
+        op()(ctx, src0, src1, dst, src0_dd, src1_dd, dst_dd, main_stream);
+    } else if (src0->type == GGML_TYPE_F16 && dst->type == GGML_TYPE_F16) {
+        op()(ctx, src0, src1, dst, (const sycl::half *)src0_dd, src1_dd,
+             (sycl::half *)dst_dd, main_stream);
+    } else if (src0->type == GGML_TYPE_F16 && dst->type == GGML_TYPE_F32) {
+        op()(ctx, src0, src1, dst, (const sycl::half *)src0_dd, src1_dd, dst_dd,
+             main_stream);
+    } else if (src0->type == GGML_TYPE_I32 && dst->type == GGML_TYPE_I32) {
+        op()(ctx, src0, src1, dst, (const int32_t *)src0_dd, (const int32_t *)src1_dd, (int32_t *)dst_dd,
+             main_stream);
+    } else if (src0->type == GGML_TYPE_I16 && dst->type == GGML_TYPE_I16) {
+        op()(ctx, src0, src1, dst, (const int16_t *)src0_dd, (const int16_t *)src1_dd, (int16_t *)dst_dd,
+             main_stream);
+    } else {
+        fprintf(stderr, "%s: unsupported types: dst: %s, src0: %s, src1: %s\n", __func__,
+            ggml_type_name(dst->type), ggml_type_name(src0->type), ggml_type_name(src1->type));
+        GGML_ABORT("fatal error");
+    }
+}
+
+static void ggml_sycl_op_repeat(ggml_backend_sycl_context & ctx, const ggml_tensor *src0,
+                                const ggml_tensor *src1, ggml_tensor *dst,
+                                const float *src0_d, const float *src1_d,
+                                float *dst_d,
+                                const queue_ptr &main_stream) {
+
+    ggml_sycl_op_bin_bcast<bin_bcast_sycl<op_repeat>>(ctx, dst, src0, dst, nullptr, src0_d, dst_d, main_stream);
+
+    (void) src1;
+    (void) src1_d;
+}
+
+inline void ggml_sycl_op_add(ggml_backend_sycl_context & ctx, const ggml_tensor *src0, const ggml_tensor *src1,
+                             ggml_tensor *dst, const float *src0_dd,
+                             const float *src1_dd, float *dst_dd,
+                             const queue_ptr &main_stream) {
+
+    ggml_sycl_op_bin_bcast<bin_bcast_sycl<op_add>>(ctx, src0, src1, dst, src0_dd, src1_dd, dst_dd, main_stream);
+}
+
+inline void ggml_sycl_op_acc(ggml_backend_sycl_context & ctx, const ggml_tensor *src0, const ggml_tensor *src1,
+                             ggml_tensor *dst, const float *src0_dd,
+                             const float *src1_dd, float *dst_dd,
+                             const queue_ptr &main_stream) {
+
+    GGML_ASSERT(src0->type == GGML_TYPE_F32);
+    GGML_ASSERT(src1->type == GGML_TYPE_F32);
+    GGML_ASSERT( dst->type == GGML_TYPE_F32);
+    GGML_ASSERT(dst->ne[3] == 1); // just 3D tensors supported
+
+    int nb1 = dst->op_params[0] / 4; // 4 bytes of float32
+    int nb2 = dst->op_params[1] / 4; // 4 bytes of float32
+    // int nb3 = dst->op_params[2] / 4; // 4 bytes of float32 - unused
+    int offset = dst->op_params[3] / 4; // offset in bytes
+
+    acc_f32_sycl(src0_dd, src1_dd, dst_dd, ggml_nelements(dst), src1->ne[0], src1->ne[1], src1->ne[2], nb1, nb2, offset, main_stream);
+
+    (void) dst;
+}
+
+inline void ggml_sycl_op_mul(ggml_backend_sycl_context & ctx, const ggml_tensor *src0, const ggml_tensor *src1,
+                             ggml_tensor *dst, const float *src0_dd,
+                             const float *src1_dd, float *dst_dd,
+                             const queue_ptr &main_stream) {
+
+    ggml_sycl_op_bin_bcast<bin_bcast_sycl<op_mul>>(ctx, src0, src1, dst, src0_dd, src1_dd, dst_dd, main_stream);
+}
+
+inline void ggml_sycl_op_div(ggml_backend_sycl_context & ctx, const ggml_tensor *src0, const ggml_tensor *src1,
+                             ggml_tensor *dst, const float *src0_dd,
+                             const float *src1_dd, float *dst_dd,
+                             const queue_ptr &main_stream) {
+
+    ggml_sycl_op_bin_bcast<bin_bcast_sycl<op_div>>(ctx, src0, src1, dst, src0_dd, src1_dd, dst_dd, main_stream);
+}
+
+inline void ggml_sycl_op_gelu(ggml_backend_sycl_context & ctx, const ggml_tensor *src0, const ggml_tensor *src1,
+                              ggml_tensor *dst, const float *src0_dd,
+                              const float *src1_dd, float *dst_dd,
+                              const queue_ptr &main_stream) {
+
+    GGML_ASSERT(src0->type == GGML_TYPE_F32);
+    GGML_ASSERT( dst->type == GGML_TYPE_F32);
+
+    gelu_f32_sycl(src0_dd, dst_dd, ggml_nelements(src0), main_stream);
+
+    (void) src1;
+    (void) dst;
+    (void) src1_dd;
+}
+
+inline void ggml_sycl_op_silu(ggml_backend_sycl_context & ctx, const ggml_tensor *src0, const ggml_tensor *src1,
+                              ggml_tensor *dst, const float *src0_dd,
+                              const float *src1_dd, float *dst_dd,
+                              const queue_ptr &main_stream) {
+
+    GGML_ASSERT(src0->type == GGML_TYPE_F32);
+    GGML_ASSERT( dst->type == GGML_TYPE_F32);
+
+    silu_f32_sycl(src0_dd, dst_dd, ggml_nelements(src0), main_stream);
+
+    (void) src1;
+    (void) dst;
+    (void) src1_dd;
+}
+
+inline void ggml_sycl_op_gelu_quick(ggml_backend_sycl_context & ctx, const ggml_tensor *src0,
+                                    const ggml_tensor *src1, ggml_tensor *dst,
+                                    const float *src0_dd, const float *src1_dd,
+                                    float *dst_dd,
+                                    const queue_ptr &main_stream) {
+
+    GGML_ASSERT(src0->type == GGML_TYPE_F32);
+    GGML_ASSERT( dst->type == GGML_TYPE_F32);
+
+    gelu_quick_f32_sycl(src0_dd, dst_dd, ggml_nelements(src0), main_stream);
+
+    (void) src1;
+    (void) dst;
+    (void) src1_dd;
+}
+
+inline void ggml_sycl_op_tanh(ggml_backend_sycl_context & ctx, const ggml_tensor *src0, const ggml_tensor *src1,
+                              ggml_tensor *dst, const float *src0_dd,
+                              const float *src1_dd, float *dst_dd,
+                              const queue_ptr &main_stream) {
+
+    GGML_ASSERT(src0->type == GGML_TYPE_F32);
+    GGML_ASSERT( dst->type == GGML_TYPE_F32);
+    tanh_f32_sycl(src0_dd, dst_dd, ggml_nelements(src0), main_stream);
+
+    (void) src1;
+    (void) dst;
+    (void) src1_dd;
+}
+
+inline void ggml_sycl_op_relu(ggml_backend_sycl_context & ctx, const ggml_tensor *src0, const ggml_tensor *src1,
+                              ggml_tensor *dst, const float *src0_dd,
+                              const float *src1_dd, float *dst_dd,
+                              const queue_ptr &main_stream) {
+
+    GGML_ASSERT(src0->type == GGML_TYPE_F32);
+    GGML_ASSERT( dst->type == GGML_TYPE_F32);
+
+    relu_f32_sycl(src0_dd, dst_dd, ggml_nelements(src0), main_stream);
+
+    (void) src1;
+    (void) dst;
+    (void) src1_dd;
+}
+
+static void ggml_sycl_op_hardsigmoid(ggml_backend_sycl_context & ctx, const ggml_tensor *src0,
+                                     const ggml_tensor *src1, ggml_tensor *dst,
+                                     const float *src0_dd, const float *src1_dd,
+                                     float *dst_dd,
+                                     const queue_ptr &main_stream) {
+
+    GGML_ASSERT(src0->type == GGML_TYPE_F32);
+    GGML_ASSERT( dst->type == GGML_TYPE_F32);
+
+    hardsigmoid_f32_sycl(src0_dd, dst_dd, ggml_nelements(src0), main_stream);
+
+    (void) src1;
+    (void) dst;
+    (void) src1_dd;
+}
+
+static void ggml_sycl_op_hardswish(ggml_backend_sycl_context & ctx, const ggml_tensor *src0,
+                                   const ggml_tensor *src1, ggml_tensor *dst,
+                                   const float *src0_dd, const float *src1_dd,
+                                   float *dst_dd, const queue_ptr &main_stream) {
+
+    GGML_ASSERT(src0->type == GGML_TYPE_F32);
+    GGML_ASSERT( dst->type == GGML_TYPE_F32);
+
+    hardswish_f32_sycl(src0_dd, dst_dd, ggml_nelements(src0), main_stream);
+
+    (void) src1;
+    (void) dst;
+    (void) src1_dd;
+}
+
+inline void ggml_sycl_op_leaky_relu(ggml_backend_sycl_context & ctx, const ggml_tensor *src0,
+                                    const ggml_tensor *src1, ggml_tensor *dst,
+                                    const float *src0_dd, const float *src1_dd,
+                                    float *dst_dd,
+                                    const queue_ptr &main_stream) {
+
+    GGML_ASSERT(src0->type == GGML_TYPE_F32);
+    GGML_ASSERT( dst->type == GGML_TYPE_F32);
+
+    float negative_slope;
+    memcpy(&negative_slope, dst->op_params, sizeof(float));
+
+    leaky_relu_f32_sycl(src0_dd, dst_dd, ggml_nelements(src0), negative_slope, main_stream);
+
+    (void) src1;
+    (void) dst;
+    (void) src1_dd;
+}
+
+inline void ggml_sycl_op_sqr(ggml_backend_sycl_context & ctx, const ggml_tensor *src0, const ggml_tensor *src1,
+                             ggml_tensor *dst, const float *src0_dd,
+                             const float *src1_dd, float *dst_dd,
+                             const queue_ptr &main_stream) {
+
+    GGML_ASSERT(src0->type == GGML_TYPE_F32);
+    GGML_ASSERT( dst->type == GGML_TYPE_F32);
+
+    sqr_f32_sycl(src0_dd, dst_dd, ggml_nelements(src0), main_stream);
+
+    (void) src1;
+    (void) dst;
+    (void) src1_dd;
+}
+
+inline void ggml_sycl_op_upscale(ggml_backend_sycl_context & ctx, const ggml_tensor *src0,
+                                 const ggml_tensor *src1, ggml_tensor *dst,
+                                 const float *src0_dd, const float *src1_dd,
+                                 float *dst_dd,
+                                 const queue_ptr &main_stream) {
+
+    GGML_ASSERT(src0->type == GGML_TYPE_F32);
+    GGML_ASSERT(dst->type == GGML_TYPE_F32);
+
+    const float sf0 = (float)dst->ne[0]/src0->ne[0];
+    const float sf1 = (float)dst->ne[1]/src0->ne[1];
+    const float sf2 = (float)dst->ne[2]/src0->ne[2];
+    const float sf3 = (float)dst->ne[3]/src0->ne[3];
+
+    upscale_f32_sycl(src0_dd, dst_dd, src0->nb[0], src0->nb[1], src0->nb[2], src0->nb[3],
+                     dst->ne[0], dst->ne[1], dst->ne[2], dst->ne[3], sf0, sf1, sf2, sf3,
+                     main_stream);
+
+    (void) src1;
+    (void) dst;
+    (void) src1_dd;
+}
+
+inline void ggml_sycl_op_pad(ggml_backend_sycl_context & ctx, const ggml_tensor *src0, const ggml_tensor *src1,
+                             ggml_tensor *dst, const float *src0_dd,
+                             const float *src1_dd, float *dst_dd,
+                             const queue_ptr &main_stream) {
+
+    GGML_ASSERT(src0->type == GGML_TYPE_F32);
+    GGML_ASSERT(dst->type == GGML_TYPE_F32);
+    GGML_ASSERT(src0->ne[3] == 1 && dst->ne[3] == 1); // just 3D tensors
+
+    pad_f32_sycl(src0_dd, dst_dd,
+        src0->ne[0], src0->ne[1], src0->ne[2],
+        dst->ne[0], dst->ne[1], dst->ne[2], main_stream);
+
+    (void) src1;
+    (void) dst;
+    (void) src1_dd;
+}
+
+static int64_t get_row_rounding(ggml_type type, const std::array<float, GGML_SYCL_MAX_DEVICES> & tensor_split) {
+    int64_t min_compute_capability = INT_MAX;
+    int64_t max_compute_capability = INT_MIN;
+    for (int i = 0; i < ggml_sycl_info().device_count; ++i) {
+        if (tensor_split[i] < (i + 1 < ggml_sycl_info().device_count ? tensor_split[i + 1] : 1.0f)) {
+            if (min_compute_capability > ggml_sycl_info().devices[i].cc) {
+                min_compute_capability = ggml_sycl_info().devices[i].cc;
+            }
+            if (max_compute_capability < ggml_sycl_info().devices[i].cc) {
+                max_compute_capability = ggml_sycl_info().devices[i].cc;
+            }
+        }
+    }
+
+    switch(type) {
+        case GGML_TYPE_Q4_0:
+        case GGML_TYPE_Q4_1:
+            return max_compute_capability >= VER_GEN9 ? 128 : 64;
+        case GGML_TYPE_Q5_0:
+        case GGML_TYPE_Q5_1:
+        case GGML_TYPE_Q8_0:
+            return 64;
+        case GGML_TYPE_F16:
+        case GGML_TYPE_F32:
+            return 1;
+        case GGML_TYPE_Q2_K:
+        case GGML_TYPE_Q3_K:
+        case GGML_TYPE_Q4_K:
+        case GGML_TYPE_Q5_K:
+        case GGML_TYPE_IQ2_XXS:
+        case GGML_TYPE_IQ2_XS:
+        case GGML_TYPE_IQ2_S:
+        case GGML_TYPE_IQ1_S:
+        case GGML_TYPE_IQ1_M:
+        case GGML_TYPE_IQ3_XXS:
+        case GGML_TYPE_IQ4_XS:
+        case GGML_TYPE_IQ4_NL:
+            return max_compute_capability >= VER_GEN9 ? 128 : 64;
+        case GGML_TYPE_IQ3_S:
+            return max_compute_capability >= VER_GEN9 ? 128 : 64;
+        case GGML_TYPE_Q6_K:
+            return 64;
+        default:
+            GGML_ABORT("fatal error");
+    }
+
+}
+
+inline void ggml_sycl_op_mul_mat_sycl(
+    ggml_backend_sycl_context & ctx,
+    const ggml_tensor *src0, const ggml_tensor *src1, ggml_tensor *dst,
+    const char *src0_dd_i, const float *src1_ddf_i, const char *src1_ddq_i,
+    float *dst_dd_i, const int64_t row_low, const int64_t row_high,
+    const int64_t src1_ncols, const int64_t src1_padded_row_size,
+    const queue_ptr &stream) try {
+
+    GGML_ASSERT(src0_dd_i  != nullptr);
+    GGML_ASSERT(src1_ddf_i != nullptr);
+    GGML_ASSERT(dst_dd_i   != nullptr);
+
+    const int64_t ne00 = src0->ne[0];
+    const int64_t ne10 = src1->ne[0];
+
+    const int64_t ne0 = dst->ne[0];
+
+    const int64_t row_diff = row_high - row_low;
+
+    int id;
+    SYCL_CHECK(
+        CHECK_TRY_ERROR(id = get_current_device_id()));
+
+    // the main device has a larger memory buffer to hold the results from all GPUs
+    // ldc == nrows of the matrix that cuBLAS writes into
+    int ldc = id == ctx.device ? ne0 : row_diff;
+
+#ifdef GGML_SYCL_F16
+    bool use_fp16 = true;  // TODO(Yu) SYCL capability check
+#else
+    bool use_fp16 = false;
+#endif
+    if ((src0->type == GGML_TYPE_F16 || ggml_is_quantized(src0->type)) &&
+        use_fp16 && ggml_is_contiguous(src0) && row_diff == src0->ne[1] &&
+        dst->op_params[0] == GGML_PREC_DEFAULT) {
+
+        // GGML_SYCL_DEBUG("ggml_sycl_op_mul_mat_sycl - fp16 path\n");
+        ggml_sycl_pool_alloc<sycl::half> src0_as_f16(ctx.pool());
+        if (src0->type != GGML_TYPE_F16) {
+            const to_fp16_sycl_t to_fp16_sycl = ggml_get_to_fp16_sycl(src0->type);
+            GGML_ASSERT(to_fp16_sycl != nullptr);
+            size_t ne = row_diff*ne00;
+            src0_as_f16.alloc(ne);
+            to_fp16_sycl(src0_dd_i, src0_as_f16.get(), ne, stream);
+        }
+        const sycl::half *src0_ptr = src0->type == GGML_TYPE_F16
+                                         ? (const sycl::half *)src0_dd_i
+                                         : src0_as_f16.get();
+
+        ggml_sycl_pool_alloc<sycl::half> src1_as_f16(ctx.pool());
+        if (src1->type != GGML_TYPE_F16) {
+            const to_fp16_sycl_t to_fp16_sycl = ggml_get_to_fp16_sycl(src1->type);
+            GGML_ASSERT(to_fp16_sycl != nullptr);
+            size_t ne = src1_ncols*ne10;
+            src1_as_f16.alloc(ne);
+            to_fp16_sycl(src1_ddf_i, src1_as_f16.get(), ne, stream);
+        }
+        const sycl::half *src1_ptr = src1->type == GGML_TYPE_F16
+                ? (const sycl::half *)src1->data + src1_padded_row_size
+                                         : src1_as_f16.get();
+        ggml_sycl_pool_alloc<sycl::half> dst_f16(ctx.pool(), row_diff * src1_ncols);
+
+        const sycl::half alpha_f16 = 1.0f;
+        const sycl::half beta_f16 = 0.0f;
+#if !GGML_SYCL_DNNL
+        SYCL_CHECK(CHECK_TRY_ERROR(dpct::gemm(
+            *stream, oneapi::mkl::transpose::trans,
+            oneapi::mkl::transpose::nontrans, row_diff, src1_ncols, ne10,
+            &alpha_f16, src0_ptr, dpct::library_data_t::real_half, ne00,
+            src1_ptr, dpct::library_data_t::real_half, ne10, &beta_f16,
+            dst_f16.get(), dpct::library_data_t::real_half, ldc,
+            dpct::library_data_t::real_half)));
+        const to_fp32_sycl_t to_fp32_sycl = ggml_get_to_fp32_sycl(GGML_TYPE_F16);
+        to_fp32_sycl(dst_f16.get(), dst_dd_i, row_diff*src1_ncols, stream);
+#else
+        auto dnnl_stream = ctx.stream_dnnl(stream);
+        DnnlGemmWrapper::row_gemm(dnnl_stream, false, true, src1_ncols, row_diff, ne10, src1_ptr, DnnlGemmWrapper::to_dt<sycl::half>(),
+            src0_ptr, DnnlGemmWrapper::to_dt<sycl::half>(), dst_f16.get(), DnnlGemmWrapper::to_dt<sycl::half>());
+        const to_fp32_sycl_t to_fp32_sycl = ggml_get_to_fp32_sycl(GGML_TYPE_F16);
+        to_fp32_sycl(dst_f16.get(), dst_dd_i, row_diff* src1_ncols, stream);
+#endif
+    }
+    else {
+        // GGML_SYCL_DEBUG("ggml_sycl_op_mul_mat_sycl - fp32 path\n");
+        ggml_sycl_pool_alloc<float> src0_ddq_as_f32(ctx.pool());
+        ggml_sycl_pool_alloc<float> src1_ddq_as_f32(ctx.pool());
+        if (src0->type != GGML_TYPE_F32) {
+            const to_fp32_sycl_t to_fp32_sycl = ggml_get_to_fp32_sycl(src0->type);
+            GGML_ASSERT(to_fp32_sycl != nullptr);
+            src0_ddq_as_f32.alloc(row_diff*ne00);
+            to_fp32_sycl(src0_dd_i, src0_ddq_as_f32.get(), row_diff*ne00, stream);
+        }
+        if (src1->type != GGML_TYPE_F32) {
+            const to_fp32_sycl_t to_fp32_sycl = ggml_get_to_fp32_sycl(src1->type);
+            GGML_ASSERT(to_fp32_sycl != nullptr);
+            src1_ddq_as_f32.alloc(src1_ncols*ne10);
+            to_fp32_sycl(src1_ddf_i, src1_ddq_as_f32.get(), src1_ncols*ne10, stream);
+        }
+        const float * src0_ddf_i = src0->type == GGML_TYPE_F32 ? (const float *) src0_dd_i : src0_ddq_as_f32.get();
+        const float * src1_ddf1_i = src1->type == GGML_TYPE_F32 ? (const float *) src1_ddf_i : src1_ddq_as_f32.get();
+
+        const float alpha = 1.0f;
+        const float beta = 0.0f;
+#if !GGML_SYCL_DNNL
+        SYCL_CHECK(CHECK_TRY_ERROR(oneapi::mkl::blas::column_major::gemm(
+            *stream, oneapi::mkl::transpose::trans,
+            oneapi::mkl::transpose::nontrans, row_diff, src1_ncols, ne10,
+            dpct::get_value(&alpha, *stream), src0_ddf_i, ne00,
+            src1_ddf1_i, ne10, dpct::get_value(&beta, *stream),
+            dst_dd_i, ldc)));
+#else
+        auto dnnl_stream = ctx.stream_dnnl(stream);
+         DnnlGemmWrapper::row_gemm(dnnl_stream, false, true, src1_ncols, row_diff, ne10, src1_ddf1_i, DnnlGemmWrapper::to_dt<float>(),
+            src0_ddf_i, DnnlGemmWrapper::to_dt<float>(), dst_dd_i, DnnlGemmWrapper::to_dt<float>());
+#endif
+    }
+    (void) dst;
+    (void) src1_ddq_i;
+    (void) src1_padded_row_size;
+}
+catch (sycl::exception const &exc) {
+  std::cerr << exc.what() << "Exception caught at file:" << __FILE__
+            << ", line:" << __LINE__ << std::endl;
+  std::exit(1);
+}
+
+static void ggml_sycl_op_pool2d(ggml_backend_sycl_context & ctx, const ggml_tensor *src0,
+                                const ggml_tensor *src1, ggml_tensor *dst,
+                                const float *src0_dd, const float *src1_dd,
+                                float *dst_dd, const queue_ptr &main_stream) {
+
+    GGML_ASSERT(src0->type == GGML_TYPE_F32);
+    GGML_ASSERT( dst->type == GGML_TYPE_F32);
+
+    const int32_t * opts = (const int32_t *)dst->op_params;
+    enum ggml_op_pool op = static_cast<ggml_op_pool>(opts[0]);
+    const int k0 = opts[1];
+    const int k1 = opts[2];
+    const int s0 = opts[3];
+    const int s1 = opts[4];
+    const int p0 = opts[5];
+    const int p1 = opts[6];
+
+    const int64_t IH = src0->ne[1];
+    const int64_t IW = src0->ne[0];
+
+    const int64_t N = dst->ne[3];
+    const int64_t OC = dst->ne[2];
+    const int64_t OH = dst->ne[1];
+    const int64_t OW = dst->ne[0];
+
+    const int parallel_elements = N * OC * OH * OW;
+    const int num_blocks = (parallel_elements + SYCL_POOL2D_BLOCK_SIZE - 1) / SYCL_POOL2D_BLOCK_SIZE;
+    sycl::range<3> block_nums(1, 1, num_blocks);
+    main_stream->parallel_for(
+        sycl::nd_range<3>(block_nums *
+                              sycl::range<3>(1, 1, SYCL_IM2COL_BLOCK_SIZE),
+                          sycl::range<3>(1, 1, SYCL_IM2COL_BLOCK_SIZE)),
+        [=](sycl::nd_item<3> item_ct1) {
+            pool2d_nchw_kernel(IH, IW, OH, OW, k1, k0, s1, s0, p1, p0,
+                               parallel_elements, src0_dd, dst_dd, op,
+                               item_ct1);
+        });
+
+    (void) src1;
+    (void) src1_dd;
+}
+
+inline void ggml_sycl_op_sum_rows(ggml_backend_sycl_context & ctx, const ggml_tensor *src0,
+                                  const ggml_tensor *src1, ggml_tensor *dst,
+                                  const float *src0_dd, const float *src1_dd,
+                                  float *dst_dd,
+                                  const queue_ptr &main_stream) {
+
+    GGML_ASSERT(src0->type == GGML_TYPE_F32);
+    GGML_ASSERT( dst->type == GGML_TYPE_F32);
+
+    const int64_t ncols = src0->ne[0];
+    const int64_t nrows = ggml_nrows(src0);
+
+    sum_rows_f32_sycl(src0_dd, dst_dd, ncols, nrows, main_stream);
+
+    (void) src1;
+    (void) dst;
+    (void) src1_dd;
+}
+
+inline void ggml_sycl_op_argsort(ggml_backend_sycl_context & ctx, const ggml_tensor *src0,
+                                 const ggml_tensor *src1, ggml_tensor *dst,
+                                 const float *src0_dd, const float *src1_dd,
+                                 float *dst_dd,
+                                 const queue_ptr &main_stream) {
+
+    GGML_ASSERT(src0->type == GGML_TYPE_F32);
+    GGML_ASSERT( dst->type == GGML_TYPE_I32);
+
+    const int64_t ncols = src0->ne[0];
+    const int64_t nrows = ggml_nrows(src0);
+
+    enum ggml_sort_order order = (enum ggml_sort_order) dst->op_params[0];
+
+    argsort_f32_i32_sycl(src0_dd, (int *)dst_dd, ncols, nrows, order, main_stream);
+
+    (void) src1;
+    (void) dst;
+    (void) src1_dd;
+}
+
+inline void ggml_sycl_op_diag_mask_inf(ggml_backend_sycl_context & ctx, const ggml_tensor *src0,
+                                       const ggml_tensor *src1,
+                                       ggml_tensor *dst, const float *src0_dd,
+                                       const float *src1_dd, float *dst_dd,
+                                       const queue_ptr &main_stream) {
+
+    GGML_ASSERT(src0->type == GGML_TYPE_F32);
+    GGML_ASSERT( dst->type == GGML_TYPE_F32);
+
+    const int64_t ne00 = src0->ne[0];
+    const int64_t ne01 = src0->ne[1];
+    const int nrows0 = ggml_nrows(src0);
+
+    const int n_past = ((int32_t *) dst->op_params)[0];
+
+    diag_mask_inf_f32_sycl(src0_dd, dst_dd, ne00, nrows0, ne01, n_past, main_stream);
+
+    (void) src1;
+    (void) dst;
+    (void) src1_dd;
+}
+
+inline void ggml_sycl_op_scale(ggml_backend_sycl_context & ctx, const ggml_tensor *src0, const ggml_tensor *src1,
+                               ggml_tensor *dst, const float *src0_dd,
+                               const float *src1_dd, float *dst_dd,
+                               const queue_ptr &main_stream) {
+
+    GGML_ASSERT(src0->type == GGML_TYPE_F32);
+    GGML_ASSERT( dst->type == GGML_TYPE_F32);
+
+    float scale;
+    memcpy(&scale, dst->op_params, sizeof(float));
+
+    scale_f32_sycl(src0_dd, dst_dd, scale, ggml_nelements(src0), main_stream);
+    /*
+    DPCT1010:87: SYCL uses exceptions to report errors and does not use the
+    error codes. The call was replaced with 0. You need to rewrite this code.
+    */
+    SYCL_CHECK(0);
+
+    (void) src1;
+    (void) dst;
+    (void) src1_dd;
+}
+
+inline void ggml_sycl_op_clamp(ggml_backend_sycl_context & ctx, const ggml_tensor *src0, const ggml_tensor *src1,
+                               ggml_tensor *dst, const float *src0_dd,
+                               const float *src1_dd, float *dst_dd,
+                               const queue_ptr &main_stream) {
+
+    GGML_ASSERT(src0->type == GGML_TYPE_F32);
+    GGML_ASSERT( dst->type == GGML_TYPE_F32);
+
+    float min;
+    float max;
+    memcpy(&min, dst->op_params, sizeof(float));
+    memcpy(&max, (float *) dst->op_params + 1, sizeof(float));
+
+    clamp_f32_sycl(src0_dd, dst_dd, min, max, ggml_nelements(src0), main_stream);
+    /*
+    DPCT1010:88: SYCL uses exceptions to report errors and does not use the
+    error codes. The call was replaced with 0. You need to rewrite this code.
+    */
+    SYCL_CHECK(0);
+
+    (void) src1;
+    (void) dst;
+    (void) src1_dd;
+}
+
+static void ggml_sycl_op_flatten(ggml_backend_sycl_context & ctx, const ggml_tensor *src0,
+                                 const ggml_tensor *src1, ggml_tensor *dst,
+                                 const ggml_sycl_op_flatten_t op) try {
+    const int64_t nrows0 = ggml_nrows(src0);
+
+    const bool use_src1 = src1 != nullptr;
+    const int64_t nrows1 = use_src1 ? ggml_nrows(src1) : 1;
+
+    GGML_ASSERT(!use_src1 || src1->backend != GGML_BACKEND_TYPE_GPU_SPLIT);
+    GGML_ASSERT(              dst->backend != GGML_BACKEND_TYPE_GPU_SPLIT);
+
+    ggml_tensor_extra_gpu * src0_extra =            (ggml_tensor_extra_gpu *) src0->extra;
+    ggml_tensor_extra_gpu * src1_extra = use_src1 ? (ggml_tensor_extra_gpu *) src1->extra : nullptr;
+    ggml_tensor_extra_gpu * dst_extra  =            (ggml_tensor_extra_gpu *)  dst->extra;
+
+    // dd = data device
+    float * src0_ddf = (float *) src0->data;
+    float * src1_ddf = use_src1 ? (float *) src1->data : nullptr;
+    float *  dst_ddf = (float *) dst->data;
+
+    ggml_sycl_pool_alloc<float> src0_f(ctx.pool());
+    ggml_sycl_pool_alloc<float> src1_f(ctx.pool());
+    ggml_sycl_pool_alloc<float>  dst_f(ctx.pool());
+
+    ggml_sycl_set_device(ctx.device);
+    queue_ptr main_stream = ctx.stream();
+    // GGML_SYCL_DEBUG("ctx.device=%d, main_stream=%p src0_on_device=%d, src1_on_device=%d, dst_on_device=%d\n",
+        // ctx.device, main_stream, src0_on_device, src1_on_device, dst_on_device);
+
+    // do the computation
+    op(ctx, src0, src1, dst, src0_ddf, src1_ddf, dst_ddf, main_stream);
+    // print_ggml_tensor("tensor", dst);
+}
+catch (sycl::exception const &exc) {
+
+  std::cerr << exc.what() << "Exception caught at file:" << __FILE__
+            << ", line:" << __LINE__ << std::endl;
+  std::exit(1);
+}
+
+static void ggml_sycl_set_peer_access(const int n_tokens, int main_device) {
+    static bool peer_access_enabled = false;
+
+    const bool enable_peer_access = n_tokens <= GGML_SYCL_PEER_MAX_BATCH_SIZE;
+
+    if (peer_access_enabled == enable_peer_access) {
+        return;
+    }
+
+#ifdef NDEBUG
+    for (int i = 0; i < ggml_sycl_info().device_count; ++i) {
+        SYCL_CHECK(ggml_sycl_set_device(i));
+    }
+
+    for (int i = 0; i < ggml_sycl_info().device_count; ++i) {
+        SYCL_CHECK(ggml_sycl_set_device(i));
+
+        for (int id_other = 0; id_other < ggml_sycl_info().device_count; ++id_other) {
+            if (i == id_other) {
+                continue;
+            }
+            if (i != main_device && id_other != main_device) {
+                continue;
+            }
+
+            // int can_access_peer;
+            // SYCL_CHECK(syclDeviceCanAccessPeer(&can_access_peer, id, id_other));
+            // if (can_access_peer) {
+            //     if (enable_peer_access) {
+            //         SYCL_CHECK(syclDeviceEnablePeerAccess(id_other, 0));
+            //     } else {
+            //         SYCL_CHECK(syclDeviceDisablePeerAccess(id_other));
+            //     }
+            // }
+        }
+    }
+#endif // NDEBUG
+
+    peer_access_enabled = enable_peer_access;
+}
+
+struct ggml_backend_sycl_split_buffer_type_context {
+    std::array<float, GGML_SYCL_MAX_DEVICES> tensor_split;
+};
+
+static void ggml_sycl_op_mul_mat(ggml_backend_sycl_context & ctx, const ggml_tensor *src0,
+                                 const ggml_tensor *src1, ggml_tensor *dst,
+                                 ggml_sycl_op_mul_mat_t op,
+                                 const bool convert_src1_to_q8_1) try {
+
+    GGML_TENSOR_LOCALS(int64_t, ne0, src0, ne);
+
+    GGML_TENSOR_LOCALS(int64_t, ne1, src1, ne);
+    const int64_t nrows1 = ggml_nrows(src1);
+
+    GGML_ASSERT(ne03 == ne13);
+
+    const int64_t ne0 = dst->ne[0];
+    const int64_t ne1 = dst->ne[1];
+
+    const int nb2 = dst->nb[2];
+    const int nb3 = dst->nb[3];
+
+    GGML_ASSERT(dst->backend != GGML_BACKEND_TYPE_GPU_SPLIT);
+    GGML_ASSERT(src1->backend != GGML_BACKEND_TYPE_GPU_SPLIT);
+    GGML_ASSERT(src1->type == GGML_TYPE_F32 || (src1->ne[2] == 1 && src1->ne[3] == 1));
+
+    GGML_ASSERT(ne12 >= ne02 && ne12 % ne02 == 0);
+
+    const int64_t i02_divisor = ne12 / ne02;
+
+    const size_t src0_ts = ggml_type_size(src0->type);
+    const size_t src0_bs = ggml_blck_size(src0->type);
+    const size_t q8_1_ts = sizeof(block_q8_1);
+    const size_t q8_1_bs = QK8_1;
+
+    ggml_tensor_extra_gpu * src0_extra = (ggml_tensor_extra_gpu *) src0->extra;
+    ggml_tensor_extra_gpu * src1_extra = (ggml_tensor_extra_gpu *) src1->extra;
+    ggml_tensor_extra_gpu *  dst_extra = (ggml_tensor_extra_gpu *)  dst->extra;
+
+    const bool src0_is_contiguous = ggml_is_contiguous(src0);
+    const bool src1_is_contiguous = ggml_is_contiguous(src1);
+
+    int64_t src1_padded_col_size = GGML_PAD(ne10, MATRIX_ROW_PADDING);
+
+    const bool split = src0->backend == GGML_BACKEND_TYPE_GPU_SPLIT;
+    GGML_ASSERT(!(split && ne02 > 1));
+    GGML_ASSERT(!(split && ne03 > 1));
+    GGML_ASSERT(!(split && ne02 < ne12));
+
+    std::array<float, GGML_SYCL_MAX_DEVICES> tensor_split;
+    if (split) {
+        // TODO: check that src0->buffer->buft is a split buffer type, replace GGML_BACKEND_TYPE_GPU_SPLIT check
+        // GGML_ASSERT(src0->buffer != nullptr && src0->buffer->buft == ...);
+        ggml_backend_sycl_split_buffer_type_context * buft_ctx = (ggml_backend_sycl_split_buffer_type_context *) src0->buffer->buft->context;
+        tensor_split = buft_ctx->tensor_split;
+    }
+
+    struct dev_data {
+        ggml_sycl_pool_alloc<char> src0_dd_alloc;
+        ggml_sycl_pool_alloc<float> src1_ddf_alloc;
+        ggml_sycl_pool_alloc<char> src1_ddq_alloc;
+        ggml_sycl_pool_alloc<float> dst_dd_alloc;
+
+        char *src0_dd = nullptr;
+        float *src1_ddf = nullptr; // float
+        char *src1_ddq = nullptr;  // q8_1
+        float *dst_dd = nullptr;
+
+        int64_t row_low;
+        int64_t row_high;
+    };
+
+    dev_data dev[GGML_SYCL_MAX_DEVICES];
+
+    int used_devices = 0;
+    queue_ptr main_stream = ctx.stream();
+
+    for (int i = 0; i < ggml_sycl_info().device_count; ++i) {
+        // by default, use all rows
+        dev[i].row_low  = 0;
+        dev[i].row_high = ne01;
+
+        // for multi GPU, get the row boundaries from tensor split
+        // and round to mul_mat_q tile sizes
+        if (split) {
+            const int64_t rounding = get_row_rounding(src0->type, tensor_split);
+
+            if (i != 0) {
+                dev[i].row_low  = ne01*tensor_split[i];
+                if (dev[i].row_low < ne01) {
+                    dev[i].row_low -= dev[i].row_low % rounding;
+                }
+            }
+
+            if (i != ggml_sycl_info().device_count - 1) {
+                dev[i].row_high  = ne01*tensor_split[i + 1];
+                if (dev[i].row_high < ne01) {
+                    dev[i].row_high -= dev[i].row_high % rounding;
+                }
+            }
+        }
+    }
+
+    for (int i = 0; i < ggml_sycl_info().device_count; ++i) {
+        if ((!split && i != ctx.device) || dev[i].row_low == dev[i].row_high) {
+            continue;
+        }
+
+        used_devices++;
+
+        const bool src1_on_device = i == ctx.device;
+        const bool  dst_on_device = i == ctx.device;
+
+        ggml_sycl_set_device(i);
+        queue_ptr stream = ctx.stream(i, 0);
+
+        if (src0_is_contiguous) {
+            dev[i].src0_dd = (char *) src0->data;
+        } else {
+            dev[i].src0_dd = dev[i].src0_dd_alloc.alloc(ctx.pool(i), ggml_nbytes(src0));
+        }
+
+        if (src1_on_device && src1_is_contiguous) {
+            dev[i].src1_ddf = (float *) src1->data;
+        } else {
+            dev[i].src1_ddf = dev[i].src1_ddf_alloc.alloc(ctx.pool(i), ggml_nelements(src1));
+        }
+
+        if (convert_src1_to_q8_1) {
+            dev[i].src1_ddq = dev[i].src1_ddq_alloc.alloc(ctx.pool(i), nrows1*src1_padded_col_size*q8_1_ts/q8_1_bs);
+
+            if (src1_on_device && src1_is_contiguous) {
+                quantize_row_q8_1_sycl(dev[i].src1_ddf, dev[i].src1_ddq, ne10, nrows1, src1_padded_col_size, stream);
+                /*
+                DPCT1010:90: SYCL uses exceptions to report errors and does not
+                use the error codes. The call was replaced with 0. You need to
+                rewrite this code.
+                */
+                SYCL_CHECK(0);
+            }
+        }
+
+        if (dst_on_device) {
+            dev[i].dst_dd = (float *) dst->data;
+        } else {
+            const size_t size_dst_ddf = split ? (dev[i].row_high - dev[i].row_low)*ne1 : ggml_nelements(dst);
+            dev[i].dst_dd = dev[i].dst_dd_alloc.alloc(ctx.pool(i), size_dst_ddf);
+        }
+    }
+
+    // if multiple devices are used they need to wait for the main device
+    // here an event is recorded that signals that the main device has finished calculating the input data
+    if (split && used_devices > 1) {
+        ggml_sycl_set_device(ctx.device);
+        /*
+        DPCT1024:91: The original code returned the error code that was further
+        consumed by the program logic. This original code was replaced with 0.
+        You may need to rewrite the program logic consuming the error code.
+        */
+        SYCL_CHECK(CHECK_TRY_ERROR(
+            *src0_extra->events[ctx.device][0] =
+                ctx.stream()->ext_oneapi_submit_barrier()));
+    }
+
+    const int64_t src1_col_stride = split && used_devices > 1 ? MUL_MAT_SRC1_COL_STRIDE : ne11;
+    for (int64_t src1_col_0 = 0; src1_col_0 < ne11; src1_col_0 += src1_col_stride) {
+        const int64_t is = split ? (src1_col_0/src1_col_stride) % GGML_SYCL_MAX_STREAMS : 0;
+        const int64_t src1_ncols = src1_col_0 + src1_col_stride > ne11 ? ne11 - src1_col_0 : src1_col_stride;
+
+        for (int i = 0; i < ggml_sycl_info().device_count; ++i) {
+            if ((!split && i != ctx.device) || dev[i].row_low == dev[i].row_high) {
+                continue;
+            }
+
+            const bool src1_on_device = i == ctx.device;
+            const bool  dst_on_device = i == ctx.device;
+            const int64_t row_diff = dev[i].row_high - dev[i].row_low;
+
+            ggml_sycl_set_device(i);
+            queue_ptr stream = ctx.stream(i, is);
+
+            // wait for main GPU data if necessary
+            if (split && (i != ctx.device || is != 0)) {
+                /*
+                DPCT1009:163: SYCL uses exceptions to report errors and does not
+                use the error codes. The original code was commented out and a
+                warning string was inserted. You need to rewrite this code.
+                */
+                SYCL_CHECK(CHECK_TRY_ERROR(stream->ext_oneapi_submit_barrier(
+                    {*src0_extra->events[ctx.device][0]})));
+            }
+
+            for (int64_t i0 = 0; i0 < ne13*ne12; ++i0) {
+                const int64_t i03 = i0 / ne12;
+                const int64_t i02 = i0 % ne12;
+
+                const size_t src1_ddq_i_offset = (i0*ne11 + src1_col_0) * src1_padded_col_size*q8_1_ts/q8_1_bs;
+
+                // for split tensors the data begins at i0 == i0_offset_low
+                char  *  src0_dd_i =  dev[i].src0_dd + (i0/i02_divisor) * (ne01*ne00*src0_ts)/src0_bs;
+                float * src1_ddf_i = dev[i].src1_ddf + (i0*ne11 + src1_col_0) * ne10;
+                char  * src1_ddq_i = dev[i].src1_ddq +  src1_ddq_i_offset;
+                float *   dst_dd_i =   dev[i].dst_dd + (i0*ne1  + src1_col_0) * (dst_on_device ? ne0 : row_diff);
+
+                // the main device memory buffer can be on VRAM scratch, with space for all partial results
+                // in that case an offset on dst_ddf_i is needed
+                if (i == ctx.device) {
+                    dst_dd_i += dev[i].row_low; // offset is 0 if no tensor split
+                }
+
+                // copy src0, src1 to device if necessary
+                if (src1_is_contiguous) {
+                    if (i != ctx.device) {
+                        if (convert_src1_to_q8_1) {
+                            char * src1_ddq_i_source = dev[ctx.device].src1_ddq + src1_ddq_i_offset;
+                          SYCL_CHECK(CHECK_TRY_ERROR(stream->memcpy(
+                                src1_ddq_i, src1_ddq_i_source,
+                                src1_ncols * src1_padded_col_size * q8_1_ts /
+                                    q8_1_bs).wait()));
+                        } else {
+
+                            float * src1_ddf_i_source = (float *) src1_extra->data_device[ctx.device];
+                            src1_ddf_i_source += (i0*ne11 + src1_col_0) * ne10;
+
+                            SYCL_CHECK(CHECK_TRY_ERROR(dev2dev_memcpy(*stream, *main_stream,
+                                src1_ddf_i, src1_ddf_i_source,
+                                src1_ncols * ne10 * sizeof(float))));
+                        }
+                    }
+                } else if (src1_on_device && !src1_is_contiguous) {
+                    SYCL_CHECK(ggml_sycl_cpy_tensor_2d(
+                                   src1_ddf_i, src1, i03, i02, src1_col_0, src1_col_0+src1_ncols, stream));
+                } else {
+                    GGML_ABORT("fatal error");
+                }
+
+                if (convert_src1_to_q8_1 && !src1_is_contiguous) {
+                    quantize_row_q8_1_sycl(src1_ddf_i, src1_ddq_i, ne10, src1_ncols, src1_padded_col_size, stream);
+                    /*
+                    DPCT1010:92: SYCL uses exceptions to report errors and does
+                    not use the error codes. The call was replaced with 0. You
+                    need to rewrite this code.
+                    */
+                    SYCL_CHECK(0);
+                }
+
+                if (src1_col_0 == 0 && !src0_is_contiguous && i02 % i02_divisor == 0) {
+                    SYCL_CHECK(ggml_sycl_cpy_tensor_2d(src0_dd_i, src0, i03, i02/i02_divisor, dev[i].row_low, dev[i].row_high, stream));
+                }
+                if (src1->type == GGML_TYPE_F16) {
+                    src1_padded_col_size = (i0 * ne11 + src1_col_0) * ne10;
+                }
+                // do the computation
+                SYCL_CHECK(CHECK_TRY_ERROR(op(ctx, src0, src1, dst, src0_dd_i, src1_ddf_i, src1_ddq_i, dst_dd_i,
+                    dev[i].row_low, dev[i].row_high, src1_ncols, src1_padded_col_size, stream)));
+                /*
+                DPCT1010:93: SYCL uses exceptions to report errors and does not
+                use the error codes. The call was replaced with 0. You need to
+                rewrite this code.
+                */
+                SYCL_CHECK(0);
+
+                // copy dst to host or other device if necessary
+                if (!dst_on_device) {
+                    void * dst_off_device = dst->data;
+                    if (split) {
+                        // src0 = weight matrix is saved as a transposed matrix for better memory layout.
+                        // dst is NOT transposed.
+                        // The outputs of matrix matrix multiplications can therefore NOT simply be concatenated for >1 GPU.
+                        // Instead they need to be copied to the correct slice in ne0 = dst row index.
+                        // If dst is a vector with ne0 == 1 then you don't have to do this but it still produces correct results.
+                        float * dhf_dst_i = (float *) ((char *) dst_off_device + i02*nb2 + i03*nb3);
+                        GGML_ASSERT(dst->nb[1] == ne0*sizeof(float));
+                        dhf_dst_i += src1_col_0*ne0 + dev[i].row_low;
+
+                        SYCL_CHECK(CHECK_TRY_ERROR(dpct::async_dpct_memcpy(
+                            dhf_dst_i, ne0 * sizeof(float), dst_dd_i,
+                            row_diff * sizeof(float), row_diff * sizeof(float),
+                            src1_ncols, dpct::device_to_device, *stream)));
+                    } else {
+                        float * dhf_dst_i = (float *) ((char *) dst_off_device + i02*nb2 + i03*nb3);
+                        GGML_ASSERT(dst->nb[1] == ne0*sizeof(float));
+                        dhf_dst_i += src1_col_0*ne0;
+                        SYCL_CHECK(CHECK_TRY_ERROR(
+                            stream->memcpy(dhf_dst_i, dst_dd_i,
+                                           src1_ncols * ne0 * sizeof(float)).wait()));
+                    }
+                }
+
+                // add event for the main device to wait on until other device is done
+                if (split && (i != ctx.device || is != 0)) {
+                    /*
+                    DPCT1024:94: The original code returned the error code that
+                    was further consumed by the program logic. This original
+                    code was replaced with 0. You may need to rewrite the
+                    program logic consuming the error code.
+                    */
+                    SYCL_CHECK(CHECK_TRY_ERROR(
+                        *src0_extra->events[i][is] =
+                            stream->ext_oneapi_submit_barrier()));
+                }
+            }
+        }
+    }
+
+    // main device waits for all other devices to be finished
+    if (split && ggml_sycl_info().device_count > 1) {
+        int64_t is_max = (ne11 + MUL_MAT_SRC1_COL_STRIDE - 1) / MUL_MAT_SRC1_COL_STRIDE;
+        is_max = is_max <= GGML_SYCL_MAX_STREAMS ? is_max : GGML_SYCL_MAX_STREAMS;
+
+        ggml_sycl_set_device(ctx.device);
+        for (int i = 0; i < ggml_sycl_info().device_count; ++i) {
+            if (dev[i].row_low == dev[i].row_high) {
+                continue;
+            }
+            for (int64_t is = 0; is < is_max; ++is) {
+                SYCL_CHECK(CHECK_TRY_ERROR(
+                    ctx.stream()->ext_oneapi_submit_barrier(
+                        {*src0_extra->events[i][is]})));
+            }
+        }
+    }
+}
+catch (sycl::exception const &exc) {
+  std::cerr << exc.what() << "Exception caught at file:" << __FILE__
+            << ", line:" << __LINE__ << std::endl;
+  std::exit(1);
+}
+
+
+static void ggml_sycl_repeat(ggml_backend_sycl_context & ctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
+    GGML_SYCL_DEBUG("call %s\n", __func__);
+    ggml_sycl_op_flatten(ctx, src0, src1, dst, ggml_sycl_op_repeat);
+    GGML_SYCL_DEBUG("call %s done\n", __func__);
+}
+
+static void ggml_sycl_get_rows(ggml_backend_sycl_context & ctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
+    GGML_SYCL_DEBUG("call %s\n", __func__);
+    ggml_sycl_op_flatten(ctx, src0, src1, dst, ggml_sycl_op_get_rows);
+    GGML_SYCL_DEBUG("call %s done\n", __func__);
+}
+
+static void ggml_sycl_add(ggml_backend_sycl_context & ctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
+    GGML_SYCL_DEBUG("call %s\n", __func__);
+    ggml_sycl_op_flatten(ctx, src0, src1, dst, ggml_sycl_op_add);
+    GGML_SYCL_DEBUG("call %s done\n", __func__);
+}
+
+static void ggml_sycl_acc(ggml_backend_sycl_context & ctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
+    GGML_SYCL_DEBUG("call %s\n", __func__);
+    ggml_sycl_op_flatten(ctx, src0, src1, dst, ggml_sycl_op_acc);
+    GGML_SYCL_DEBUG("call %s done\n", __func__);
+}
+
+static void ggml_sycl_mul(ggml_backend_sycl_context & ctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
+    GGML_SYCL_DEBUG("call %s\n", __func__);
+    ggml_sycl_op_flatten(ctx, src0, src1, dst, ggml_sycl_op_mul);
+    GGML_SYCL_DEBUG("call %s done\n", __func__);
+}
+
+static void ggml_sycl_div(ggml_backend_sycl_context & ctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
+    GGML_SYCL_DEBUG("call %s\n", __func__);
+    ggml_sycl_op_flatten(ctx, src0, src1, dst, ggml_sycl_op_div);
+    GGML_SYCL_DEBUG("call %s done\n", __func__);
+}
+
+static void ggml_sycl_gelu(ggml_backend_sycl_context & ctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
+    GGML_SYCL_DEBUG("call %s\n", __func__);
+    ggml_sycl_op_flatten(ctx, src0, src1, dst, ggml_sycl_op_gelu);
+    GGML_SYCL_DEBUG("call %s done\n", __func__);
+}
+
+static void ggml_sycl_silu(ggml_backend_sycl_context & ctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
+    GGML_SYCL_DEBUG("call %s\n", __func__);
+    ggml_sycl_op_flatten(ctx, src0, src1, dst, ggml_sycl_op_silu);
+    GGML_SYCL_DEBUG("call %s done\n", __func__);
+}
+
+static void ggml_sycl_gelu_quick(ggml_backend_sycl_context & ctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
+    GGML_SYCL_DEBUG("call %s\n", __func__);
+    ggml_sycl_op_flatten(ctx, src0, src1, dst, ggml_sycl_op_gelu_quick);
+    GGML_SYCL_DEBUG("call %s done\n", __func__);
+}
+
+static void ggml_sycl_tanh(ggml_backend_sycl_context & ctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
+    GGML_SYCL_DEBUG("call %s\n", __func__);
+    ggml_sycl_op_flatten(ctx, src0, src1, dst, ggml_sycl_op_tanh);
+    GGML_SYCL_DEBUG("call %s done\n", __func__);
+}
+
+static void ggml_sycl_relu(ggml_backend_sycl_context & ctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
+    GGML_SYCL_DEBUG("call %s\n", __func__);
+    ggml_sycl_op_flatten(ctx, src0, src1, dst, ggml_sycl_op_relu);
+    GGML_SYCL_DEBUG("call %s done\n", __func__);
+}
+
+static void ggml_sycl_hardsigmoid(ggml_backend_sycl_context & ctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
+    GGML_SYCL_DEBUG("call %s\n", __func__);
+    ggml_sycl_op_flatten(ctx, src0, src1, dst, ggml_sycl_op_hardsigmoid);
+    GGML_SYCL_DEBUG("call %s done\n", __func__);
+}
+
+static void ggml_sycl_hardswish(ggml_backend_sycl_context & ctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
+    GGML_SYCL_DEBUG("call %s\n", __func__);
+    ggml_sycl_op_flatten(ctx, src0, src1, dst, ggml_sycl_op_hardswish);
+    GGML_SYCL_DEBUG("call %s done\n", __func__);
+}
+
+static void ggml_sycl_leaky_relu(ggml_backend_sycl_context & ctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
+    GGML_SYCL_DEBUG("call %s\n", __func__);
+    ggml_sycl_op_flatten(ctx, src0, src1, dst, ggml_sycl_op_leaky_relu);
+    GGML_SYCL_DEBUG("call %s done\n", __func__);
+}
+
+static void ggml_sycl_sqr(ggml_backend_sycl_context & ctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
+    GGML_SYCL_DEBUG("call %s\n", __func__);
+    ggml_sycl_op_flatten(ctx, src0, src1, dst, ggml_sycl_op_sqr);
+    GGML_SYCL_DEBUG("call %s done\n", __func__);
+}
+
+static void ggml_sycl_norm(ggml_backend_sycl_context & ctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
+    GGML_SYCL_DEBUG("call %s\n", __func__);
+    ggml_sycl_op_flatten(ctx, src0, src1, dst, ggml_sycl_op_norm);
+    GGML_SYCL_DEBUG("call %s done\n", __func__);
+}
+
+static void ggml_sycl_group_norm(ggml_backend_sycl_context & ctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
+    GGML_SYCL_DEBUG("call %s\n", __func__);
+    ggml_sycl_op_flatten(ctx, src0, src1, dst, ggml_sycl_op_group_norm);
+    GGML_SYCL_DEBUG("call %s done\n", __func__);
+}
+
+static void ggml_sycl_upscale(ggml_backend_sycl_context & ctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
+    GGML_SYCL_DEBUG("call %s\n", __func__);
+    ggml_sycl_op_flatten(ctx, src0, src1, dst, ggml_sycl_op_upscale);
+    GGML_SYCL_DEBUG("call %s done\n", __func__);
+}
+
+static void ggml_sycl_pad(ggml_backend_sycl_context & ctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
+    GGML_SYCL_DEBUG("call %s\n", __func__);
+    ggml_sycl_op_flatten(ctx, src0, src1, dst, ggml_sycl_op_pad);
+    GGML_SYCL_DEBUG("call %s done\n", __func__);
+}
+
+
+static void ggml_sycl_rms_norm(ggml_backend_sycl_context & ctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
+    GGML_SYCL_DEBUG("call %s\n", __func__);
+    ggml_sycl_op_flatten(ctx, src0, src1, dst, ggml_sycl_op_rms_norm);
+    GGML_SYCL_DEBUG("call %s done\n", __func__);
+}
+
+static void ggml_sycl_mul_mat_vec_p021(ggml_backend_sycl_context & ctx, const ggml_tensor *src0,
+                                       const ggml_tensor *src1,
+                                       ggml_tensor *dst) try {
+    GGML_ASSERT(ggml_is_permuted(src0) && ggml_is_permuted(src1));
+    GGML_ASSERT(src0->backend != GGML_BACKEND_TYPE_GPU_SPLIT);
+    GGML_ASSERT(src0->nb[0] <= src0->nb[1] && src0->nb[2] <= src0->nb[3]); // 0213 permutation
+    GGML_ASSERT(src1->nb[0] <= src1->nb[1] && src1->nb[2] <= src1->nb[3]); // 0213 permutation
+    GGML_ASSERT(src0->type == GGML_TYPE_F16);
+    GGML_ASSERT(src1->type == GGML_TYPE_F32);
+
+    const int64_t ne00 = src0->ne[0];
+    const int64_t ne01 = src0->ne[1];
+    const int64_t ne02 = src0->ne[2];
+
+    const int64_t ne12 = src1->ne[2];
+
+    SYCL_CHECK(ggml_sycl_set_device(ctx.device));
+    queue_ptr main_stream = ctx.stream();
+
+    void  * src0_ddq = src0->data;
+    float * src1_ddf = (float *) src1->data;
+    float * dst_ddf  = (float *) dst->data;
+
+    ggml_mul_mat_p021_f16_f32_sycl(src0_ddq, src1_ddf, dst_ddf, ne00, ne01, ne02, ne12, main_stream);
+}
+catch (sycl::exception const &exc) {
+  std::cerr << exc.what() << "Exception caught at file:" << __FILE__
+            << ", line:" << __LINE__ << std::endl;
+  std::exit(1);
+}
+
+static void ggml_sycl_mul_mat_vec_nc(ggml_backend_sycl_context & ctx, const ggml_tensor *src0,
+                                     const ggml_tensor *src1,
+                                     ggml_tensor *dst) try {
+    GGML_ASSERT(!ggml_is_transposed(src0));
+    GGML_ASSERT(!ggml_is_transposed(src1));
+    GGML_ASSERT(!ggml_is_permuted(src0));
+    GGML_ASSERT(src0->backend != GGML_BACKEND_TYPE_GPU_SPLIT);
+    GGML_ASSERT(src0->type == GGML_TYPE_F16);
+    GGML_ASSERT(src1->type == GGML_TYPE_F32);
+
+    const int64_t ne00 = src0->ne[0];
+    const int64_t ne01 = src0->ne[1];
+    const int64_t ne02 = src0->ne[2];
+
+    const int64_t nb01 = src0->nb[1];
+    const int64_t nb02 = src0->nb[2];
+
+    const int64_t ne12 = src1->ne[2];
+
+    SYCL_CHECK(ggml_sycl_set_device(ctx.device));
+    queue_ptr main_stream = ctx.stream();
+
+    void  * src0_ddq = src0->data;
+    float * src1_ddf = (float *) src1->data;
+    float * dst_ddf  = (float *) dst->data;
+
+    const int64_t row_stride_x = nb01 / sizeof(sycl::half);
+    const int64_t channel_stride_x = nb02 / sizeof(sycl::half);
+
+    ggml_mul_mat_vec_nc_f16_f32_sycl(src0_ddq, src1_ddf, dst_ddf, ne00, ne01, row_stride_x, ne02, ne12, channel_stride_x, main_stream);
+}
+catch (sycl::exception const &exc) {
+  std::cerr << exc.what() << "Exception caught at file:" << __FILE__
+            << ", line:" << __LINE__ << std::endl;
+  std::exit(1);
+}
+
+static void k_compute_batched_ptrs(const sycl::half *src0_as_f16,
+                                   const sycl::half *src1_as_f16, char *dst,
+                                   const void **ptrs_src, void **ptrs_dst,
+                                   int64_t ne12, int64_t ne13, int64_t ne23,
+                                   size_t nb02, size_t nb03, size_t nb12,
+                                   size_t nb13, size_t nbd2, size_t nbd3,
+                                   int64_t r2, int64_t r3,
+                                   const sycl::nd_item<3> &item_ct1) {
+    int64_t i13 = item_ct1.get_group(2) * item_ct1.get_local_range(2) +
+                  item_ct1.get_local_id(2);
+    int64_t i12 = item_ct1.get_group(1) * item_ct1.get_local_range(1) +
+                  item_ct1.get_local_id(1);
+
+    if (i13 >= ne13 || i12 >= ne12) {
+        return;
+    }
+
+    int64_t i03 = i13 / r3;
+    int64_t i02 = i12 / r2;
+
+    ptrs_src[0*ne23 + i12 + i13*ne12] = (const char *) src0_as_f16 + i02*nb02 + i03*nb03;
+    ptrs_src[1*ne23 + i12 + i13*ne12] = (const char *) src1_as_f16 + i12*nb12 + i13*nb13;
+    ptrs_dst[0*ne23 + i12 + i13*ne12] = (      char *)         dst + i12*nbd2 + i13*nbd3;
+}
+
+static void ggml_sycl_mul_mat_batched_sycl(ggml_backend_sycl_context & ctx,
+                                             const ggml_tensor *src0,
+                                             const ggml_tensor *src1,
+                                             ggml_tensor *dst) try {
+    GGML_ASSERT(!ggml_is_transposed(src0));
+    GGML_ASSERT(!ggml_is_transposed(src1));
+    GGML_ASSERT(src0->backend != GGML_BACKEND_TYPE_GPU_SPLIT);
+    GGML_ASSERT(src0->type == GGML_TYPE_F16);
+
+    GGML_TENSOR_BINARY_OP_LOCALS
+
+    const int64_t ne_dst = ggml_nelements(dst);
+
+    SYCL_CHECK(ggml_sycl_set_device(ctx.device));
+    queue_ptr main_stream = ctx.stream();;
+
+    void * src0_ddq = src0->data;
+    sycl::half *src0_as_f16 = (sycl::half *)src0_ddq;
+    float * src1_ddf = (float *) src1->data;
+    float * dst_ddf = (float *) dst->data;
+
+    // convert src1 to fp16
+    ggml_sycl_pool_alloc<sycl::half> src1_f16_alloc(ctx.pool());
+    if (src1->type != GGML_TYPE_F16) {
+        const to_fp16_sycl_t to_fp16_sycl = ggml_get_to_fp16_sycl(src1->type);
+        const int64_t ne_src1 = ggml_nelements(src1);
+        src1_f16_alloc.alloc(ne_src1);
+        GGML_ASSERT(to_fp16_sycl != nullptr);
+        to_fp16_sycl(src1_ddf, src1_f16_alloc.get(), ne_src1, main_stream);
+    }
+    sycl::half *src1_f16 = src1->type == GGML_TYPE_F16 ? (sycl::half *)src1_ddf
+                                                       : src1_f16_alloc.get();
+
+    char * dst_t;
+
+    dpct::library_data_t cu_compute_type = dpct::library_data_t::real_float;
+    dpct::library_data_t cu_data_type = dpct::library_data_t::real_float;
+
+    // dst strides
+    size_t nbd2 = dst->nb[2];
+    size_t nbd3 = dst->nb[3];
+
+    const float alpha_f32 = 1.0f;
+    const float beta_f32 = 0.0f;
+
+    const void * alpha = &alpha_f32;
+    const void * beta  = &beta_f32;
+
+    dst_t = (char *) dst_ddf;
+
+    GGML_ASSERT(ne12 % ne02 == 0);
+    GGML_ASSERT(ne13 % ne03 == 0);
+
+    // broadcast factors
+    const int64_t r2 = ne12/ne02;
+    const int64_t r3 = ne13/ne03;
+
+    if (r2 == 1 && r3 == 1 && ggml_is_contiguous_2(src0) && ggml_is_contiguous_2(src1)) {
+        // there is no broadcast and src0, src1 are contiguous across dims 2, 3
+        SYCL_CHECK(CHECK_TRY_ERROR(dpct::gemm_batch(
+            *main_stream, oneapi::mkl::transpose::trans,
+            oneapi::mkl::transpose::nontrans, ne01, ne11, ne10, alpha,
+            (const char *)src0_as_f16, dpct::library_data_t::real_half,
+            nb01 / nb00, nb02 / nb00,
+            (const char *)src1_f16, dpct::library_data_t::real_half,
+            nb11 / nb10, nb12 / nb10, beta,
+            (char *)dst_t, cu_data_type, ne01, nb2 / nb0,
+            ne12 * ne13, cu_compute_type)));
+    } else {
+        const int ne23 = ne12*ne13;
+
+        ggml_sycl_pool_alloc<const void *> ptrs_src(ctx.pool(), 2*ne23);
+        ggml_sycl_pool_alloc<      void *> ptrs_dst(ctx.pool(), 1*ne23);
+
+        sycl::range<3> block_dims(1, ne12, ne13);
+        /*
+        DPCT1049:47: The work-group size passed to the SYCL kernel may exceed
+        the limit. To get the device limit, query
+        info::device::max_work_group_size. Adjust the work-group size if needed.
+        */
+        {
+            dpct::has_capability_or_fail(main_stream->get_device(),
+                                         {sycl::aspect::fp16});
+
+            main_stream->submit([&](sycl::handler &cgh) {
+                const void **ptrs_src_get = ptrs_src.get();
+                void **ptrs_dst_get = ptrs_dst.get();
+                size_t nb12_scaled = src1->type == GGML_TYPE_F16 ? nb12 : nb12 / 2;
+                size_t nb13_scaled = src1->type == GGML_TYPE_F16 ? nb13 : nb13 / 2;
+                cgh.parallel_for(sycl::nd_range<3>(block_dims, block_dims),
+                                 [=](sycl::nd_item<3> item_ct1) {
+                                     k_compute_batched_ptrs(
+                                         src0_as_f16, src1_f16,
+                                         dst_t, ptrs_src_get,
+                                         ptrs_dst_get, ne12, ne13, ne23,
+                                         nb02, nb03, nb12_scaled, nb13_scaled,
+                                         nbd2, nbd3, r2, r3, item_ct1);
+                                 });
+            });
+        }
+        SYCL_CHECK(CHECK_TRY_ERROR(dpct::gemm_batch(
+            *main_stream, oneapi::mkl::transpose::trans,
+            oneapi::mkl::transpose::nontrans, ne01, ne11, ne10, alpha,
+            (const void **)(ptrs_src.get() + 0 * ne23),
+            dpct::library_data_t::real_half, nb01 / nb00,
+            (const void **)(ptrs_src.get() + 1 * ne23),
+            dpct::library_data_t::real_half, nb11 / nb10, beta,
+            (void **)(ptrs_dst.get() + 0 * ne23), cu_data_type, ne01, ne23,
+            cu_compute_type)));
+    }
+}
+catch (sycl::exception const &exc) {
+  std::cerr << exc.what() << "Exception caught at file:" << __FILE__
+            << ", line:" << __LINE__ << std::endl;
+  std::exit(1);
+}
+
+inline bool ggml_sycl_supports_mmq(enum ggml_type type) {
+    // TODO: accuracy issues in MMQ
+    return false;
+}
+
+bool ggml_sycl_supports_dmmv(enum ggml_type type) {
+    switch (type) {
+        case GGML_TYPE_Q4_0:
+        case GGML_TYPE_Q4_1:
+        case GGML_TYPE_Q5_0:
+        case GGML_TYPE_Q5_1:
+        case GGML_TYPE_Q8_0:
+        case GGML_TYPE_Q2_K:
+        case GGML_TYPE_Q3_K:
+        case GGML_TYPE_Q4_K:
+        case GGML_TYPE_Q5_K:
+        case GGML_TYPE_Q6_K:
+        case GGML_TYPE_F16:
+            return true;
+        default:
+            return false;
+    }
+}
+
+static void ggml_sycl_mul_mat(ggml_backend_sycl_context & ctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
+    const bool split = ggml_backend_buffer_is_sycl_split(src0->buffer);
+    int64_t min_compute_capability = INT_MAX;
+
+    if (split) {
+        ggml_backend_sycl_split_buffer_type_context * buft_ctx = (ggml_backend_sycl_split_buffer_type_context *) src0->buffer->buft->context;
+        auto & tensor_split = buft_ctx->tensor_split;
+        for (int id = 0; id < ggml_sycl_info().device_count; ++id) {
+            // skip devices that are not going to do any work:
+            if (tensor_split[id] >= (id + 1 < ggml_sycl_info().device_count ? tensor_split[id + 1] : 1.0f)) {
+                continue;
+            }
+
+            if (min_compute_capability > ggml_sycl_info().devices[id].cc) {
+                min_compute_capability = ggml_sycl_info().devices[id].cc;
+            }
+        }
+    } else {
+        min_compute_capability    = ggml_sycl_info().devices[ctx.device].cc;
+    }
+
+    // check data types and tensor shapes for custom matrix multiplication kernels:
+    bool use_dequantize_mul_mat_vec = ggml_sycl_supports_dmmv(src0->type)
+        && src1->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32
+        && src0->ne[0] % GGML_SYCL_DMMV_X == 0 && src1->ne[1] == 1;
+
+    bool use_mul_mat_vec_q =  ggml_is_quantized(src0->type)
+        && src1->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32
+        && src1->ne[1] <= MMVQ_MAX_BATCH_SIZE;
+
+    bool use_mul_mat_q =  ggml_sycl_supports_mmq(src0->type)
+        && src1->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32;
+
+    // mmvq and mmq need the __dp4a instruction which is available for gen12+
+    // Workaround in https://github.com/ggerganov/llama.cpp/commit/95f84d5ce8b449a9b16009434aca800df504a02e
+    use_mul_mat_q = use_mul_mat_q && (src0->type != GGML_TYPE_IQ2_XXS);
+#ifdef SYCL_USE_XMX
+    use_mul_mat_q = use_mul_mat_q && (src1->ne[1] <= MMQ_MAX_BATCH_SIZE);
+#endif // SYCL_USE_XMX
+
+    // mmvq path is faster in the CUDA backend.
+    if (ctx.stream()->get_backend() == sycl::backend::ext_oneapi_cuda)
+        use_dequantize_mul_mat_vec = use_dequantize_mul_mat_vec && !use_mul_mat_vec_q;
+
+    if (!split && src0->type == GGML_TYPE_F16 && ggml_is_permuted(src0) && ggml_is_permuted(src1) && src1->ne[1] == 1) {
+        // KQ single-batch
+        ggml_sycl_mul_mat_vec_p021(ctx, src0, src1, dst);
+    } else if (!split && src0->type == GGML_TYPE_F16 && !ggml_is_contiguous(src0) && !ggml_is_transposed(src1) && src1->ne[1] == 1) {
+        // KQV single-batch
+        ggml_sycl_mul_mat_vec_nc(ctx, src0, src1, dst);
+    } else if (!split && src0->type == GGML_TYPE_F16 && !ggml_is_transposed(src0) && !ggml_is_transposed(src1) && src1->ne[2]*src1->ne[3] > 1) {
+        // KQ + KQV multi-batch
+        ggml_sycl_mul_mat_batched_sycl(ctx, src0, src1, dst);
+    } else if (use_dequantize_mul_mat_vec) {
+        ggml_sycl_op_mul_mat(ctx, src0, src1, dst, ggml_sycl_op_dequantize_mul_mat_vec, false);
+    } else if (use_mul_mat_vec_q) {
+        ggml_sycl_op_mul_mat(ctx, src0, src1, dst, ggml_sycl_op_mul_mat_vec_q, true);
+    } else if (use_mul_mat_q) {
+        ggml_sycl_op_mul_mat(ctx, src0, src1, dst, ggml_sycl_op_mul_mat_q, true);
+    } else {
+        ggml_sycl_op_mul_mat(ctx, src0, src1, dst, ggml_sycl_op_mul_mat_sycl, false);
+    }
+}
+
+
+struct mmid_row_mapping {
+    int32_t i1;
+    int32_t i2;
+};
+
+__dpct_inline__ static void k_copy_src1_to_contiguous(
+    const char *__restrict__ src1_original, char *__restrict__ src1_contiguous,
+    int *__restrict__ cur_src1_row, mmid_row_mapping *__restrict__ row_mapping,
+    const char *__restrict ids, int64_t i02, size_t ids_nb1, size_t ids_nb0,
+    int64_t ne11, int64_t ne10, size_t nb11, size_t nb12,
+    const sycl::nd_item<3> &item_ct1, int &src1_row) {
+    int32_t iid1 = item_ct1.get_group(2);
+    int32_t id = item_ct1.get_group(1);
+
+    const int32_t row_id_i = *(const int32_t *) (ids + iid1*ids_nb1 + id*ids_nb0);
+
+    if (row_id_i != i02) {
+        return;
+    }
+
+    const int64_t i11 = id % ne11;
+    const int64_t i12 = iid1;
+
+    if (item_ct1.get_local_id(2) == 0) {
+        src1_row =
+            dpct::atomic_fetch_add<sycl::access::address_space::generic_space>(
+                cur_src1_row, 1);
+        row_mapping[src1_row] = {id, iid1};
+    }
+    /*
+    DPCT1065:194: Consider replacing sycl::nd_item::barrier() with
+    sycl::nd_item::barrier(sycl::access::fence_space::local_space) for better
+    performance if there is no access to global memory.
+    */
+    item_ct1.barrier();
+
+    const float * src1_row_original = (const float *)(src1_original + i11*nb11 + i12*nb12);
+    float * src1_row_contiguous = (float *)(src1_contiguous + src1_row*nb11);
+
+#pragma unroll
+    for (int i = item_ct1.get_local_id(2); i < ne10;
+         i += item_ct1.get_local_range(2)) {
+        src1_row_contiguous[i] = src1_row_original[i];
+    }
+}
+
+__dpct_inline__ static void k_copy_dst_from_contiguous(
+    char *__restrict__ dst_original, const char *__restrict__ dst_contiguous,
+    const mmid_row_mapping *__restrict__ row_mapping, int64_t ne0, size_t nb1,
+    size_t nb2, const sycl::nd_item<3> &item_ct1) {
+    int32_t i = item_ct1.get_group(2);
+
+    const int32_t i1 = row_mapping[i].i1;
+    const int32_t i2 = row_mapping[i].i2;
+
+    const float * dst_row_contiguous = (const float *)(dst_contiguous + i*nb1);
+    float * dst_row_original = (float *)(dst_original + i1*nb1 + i2*nb2);
+
+#pragma unroll
+    for (int j = item_ct1.get_local_id(2); j < ne0;
+         j += item_ct1.get_local_range(2)) {
+        dst_row_original[j] = dst_row_contiguous[j];
+    }
+}
+
+static void ggml_sycl_mul_mat_id(ggml_backend_sycl_context & ctx, const ggml_tensor *src0,
+                                 const ggml_tensor *src1,
+                                 ggml_tensor *dst) try {
+    GGML_ASSERT(!ggml_backend_buffer_is_sycl_split(src0->buffer) && "mul_mat_id does not support split buffers");
+
+    const ggml_tensor *ids = dst->src[2];
+    GGML_TENSOR_BINARY_OP_LOCALS
+
+    const queue_ptr stream = ctx.stream();
+
+    const int64_t n_as = ne02;
+    const int64_t n_ids = ids->ne[0];
+
+    std::vector<char> ids_host(ggml_nbytes(ids));
+    const char * ids_dev = (const char *) ids->data;
+
+    SYCL_CHECK(CHECK_TRY_ERROR(
+        stream->memcpy(ids_host.data(), ids_dev, ggml_nbytes(ids))));
+    SYCL_CHECK(CHECK_TRY_ERROR(stream->wait()));
+
+    ggml_tensor src0_row = *src0;
+    ggml_tensor src1_row = *src1;
+    ggml_tensor dst_row = *dst;
+
+    char *src0_original = (char *)src0->data;
+    char *src1_original = (char *)src1->data;
+    char *dst_original = (char *)dst->data;
+
+    src0_row.ne[2] = 1;
+    src0_row.ne[3] = 1;
+    src0_row.nb[3] = nb02;
+
+    src1_row.ne[1] = 1;
+    src1_row.ne[2] = 1;
+    src1_row.ne[3] = 1;
+    src1_row.nb[2] = nb11;
+    src1_row.nb[3] = nb11;
+
+    dst_row.ne[1] = 1;
+    dst_row.ne[2] = 1;
+    dst_row.ne[3] = 1;
+    dst_row.nb[2] = nb1;
+    dst_row.nb[3] = nb1;
+    if (ne12 == 1) {
+        for (int64_t iid1 = 0; iid1 < ids->ne[1]; iid1++) {
+            for (int64_t id = 0; id < n_ids; id++) {
+                const int32_t i02 = *(const int32_t *) (ids_host.data() + iid1*ids->nb[1] + id*ids->nb[0]);
+                GGML_ASSERT(i02 >= 0 && i02 < n_as);
+
+                const int64_t i11 = id % ne11;
+                const int64_t i12 = iid1;
+
+                const int64_t i1 = id;
+                const int64_t i2 = i12;
+
+            src0_row.data = src0_original + i02*nb02;
+            src1_row.data = src1_original + + i11*nb11 + i12*nb12;
+            dst_row.data = dst_original + i1*nb1   + i2*nb2;
+
+            ggml_sycl_mul_mat(ctx, &src0_row, &src1_row, &dst_row);
+            }
+        }
+    } else {
+        ggml_sycl_pool_alloc<char> src1_contiguous(ctx.pool(), sizeof(float)*ggml_nelements(src1));
+        ggml_sycl_pool_alloc<char>  dst_contiguous(ctx.pool(), sizeof(float)*ggml_nelements(dst));
+
+        src1_row.data = src1_contiguous.get();
+        dst_row.data  =  dst_contiguous.get();
+
+        for (int64_t i02 = 0; i02 < n_as; i02++) {
+            int64_t num_src1_rows = 0;
+            for (int64_t iid1 = 0; iid1 < ids->ne[1]; iid1++) {
+                for (int64_t id = 0; id < n_ids; id++) {
+                    const int32_t row_id_i = *(const int32_t *) (ids_host.data() + iid1*ids->nb[1] + id*ids->nb[0]);
+
+                    GGML_ASSERT(row_id_i >= 0 && row_id_i < n_as);
+
+                    if (row_id_i != i02) {
+                        continue;
+                    }
+
+                    num_src1_rows++;
+                }
+            }
+
+            if (num_src1_rows == 0) {
+                continue;
+            }
+
+
+            ggml_sycl_pool_alloc<int> dev_cur_src1_row(ctx.pool(), 1);
+            ggml_sycl_pool_alloc<mmid_row_mapping> dev_row_mapping(ctx.pool(), num_src1_rows);
+            SYCL_CHECK(CHECK_TRY_ERROR(
+                stream->memset(dev_cur_src1_row.get(), 0, sizeof(int))));
+
+            {
+                sycl::range<3> block_dims(1, 1, std::min((unsigned int)ne10, 768u));
+                sycl::range<3> grid_dims(1, n_ids, ids->ne[1]);
+                stream->submit([&](sycl::handler &cgh) {
+                    sycl::local_accessor<int, 0> src1_row_acc(cgh);
+
+                    char *__restrict src1_contiguous_get =
+                        src1_contiguous.get();
+                    int *__restrict dev_cur_src1_row_get =
+                        dev_cur_src1_row.get();
+                    mmid_row_mapping *__restrict dev_row_mapping_get =
+                        dev_row_mapping.get();
+                    size_t ids_nb_ct6 = ids->nb[1];
+                    size_t ids_nb_ct7 = ids->nb[0];
+
+                    cgh.parallel_for(
+                        sycl::nd_range<3>(grid_dims * block_dims, block_dims),
+                        [=](sycl::nd_item<3> item_ct1) {
+                            k_copy_src1_to_contiguous(
+                                src1_original, src1_contiguous_get,
+                                dev_cur_src1_row_get,
+                                dev_row_mapping_get, ids_dev, i02,
+                                ids_nb_ct6, ids_nb_ct7, ne11, ne10, nb11, nb12,
+                                item_ct1, src1_row_acc);
+                        });
+                });
+            }
+
+            src0_row.data = src0_original + i02*nb02;
+
+            GGML_ASSERT(nb11 == sizeof(float)*ne10);
+            GGML_ASSERT(nb1 == sizeof(float)*ne0);
+            src1_row.ne[1] = num_src1_rows;
+
+            src1_row.nb[1] = nb11;
+            src1_row.nb[2] = num_src1_rows*nb11;
+            src1_row.nb[3] = num_src1_rows*nb11;
+
+            dst_row.ne[1] = num_src1_rows;
+            dst_row.nb[1] = nb1;
+            dst_row.nb[2] = num_src1_rows*nb1;
+            dst_row.nb[3] = num_src1_rows*nb1;
+
+            ggml_sycl_mul_mat(ctx, &src0_row, &src1_row, &dst_row);
+
+            {
+                sycl::range<3> block_dims(1, 1, std::min((unsigned int)ne0, 768u));
+                sycl::range<3> grid_dims(1, 1, num_src1_rows);
+                stream->submit([&](sycl::handler &cgh) {
+                    const char *__restrict dst_contiguous_get =
+                        dst_contiguous.get();
+                    const mmid_row_mapping *__restrict dev_row_mapping_get =
+                        dev_row_mapping.get();
+
+                    cgh.parallel_for(
+                        sycl::nd_range<3>(grid_dims * block_dims, block_dims),
+                        [=](sycl::nd_item<3> item_ct1) {
+                            k_copy_dst_from_contiguous(dst_original,
+                                                       dst_contiguous_get,
+                                                       dev_row_mapping_get,
+                                                       ne0, nb1, nb2, item_ct1);
+                        });
+                });
+            }
+        }
+    }
+}
+catch (sycl::exception const &exc) {
+  std::cerr << exc.what() << "Exception caught at file:" << __FILE__
+            << ", line:" << __LINE__ << std::endl;
+  std::exit(1);
+}
+
+static void ggml_sycl_scale(ggml_backend_sycl_context & ctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
+    ggml_sycl_op_flatten(ctx, src0, src1, dst, ggml_sycl_op_scale);
+}
+
+static void ggml_sycl_clamp(ggml_backend_sycl_context & ctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
+    ggml_sycl_op_flatten(ctx, src0, src1, dst, ggml_sycl_op_clamp);
+}
+
+static void ggml_sycl_cpy(ggml_backend_sycl_context & ctx, const ggml_tensor *src0, const ggml_tensor *src1,
+                          ggml_tensor *dst) try {
+    const int64_t ne = ggml_nelements(src0);
+    GGML_ASSERT(ne == ggml_nelements(src1));
+
+    GGML_ASSERT(ggml_nbytes(src0) <= INT_MAX);
+    GGML_ASSERT(ggml_nbytes(src1) <= INT_MAX);
+
+    GGML_TENSOR_BINARY_OP_LOCALS01;
+
+    SYCL_CHECK(ggml_sycl_set_device(ctx.device));
+    queue_ptr main_stream = ctx.stream();
+
+    char * src0_ddc = (char *) src0->data;
+    char * src1_ddc = (char *) src1->data;
+
+    if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_F32) {
+        ggml_cpy_f32_f32_sycl (src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream);
+    } else if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_F16) {
+        ggml_cpy_f32_f16_sycl (src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream);
+    } else if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_Q8_0) {
+        ggml_cpy_f32_q8_0_sycl(src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream);
+    } else if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_Q4_0) {
+        ggml_cpy_f32_q4_0_sycl(src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream);
+    } else if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_Q4_1) {
+        ggml_cpy_f32_q4_1_sycl(src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream);
+    } else if (src0->type == GGML_TYPE_F16 && src1->type == GGML_TYPE_F32) {
+        ggml_cpy_f16_f32_sycl (src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream);
+    } else if (src0->type == GGML_TYPE_F16 && src1->type == GGML_TYPE_F16) {
+        ggml_cpy_f16_f16_sycl (src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream);
+    } else if (src0->type == GGML_TYPE_I16 && src1->type == GGML_TYPE_I16) {
+        ggml_cpy_i16_i16_sycl (src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream);
+    } else if (src0->type == GGML_TYPE_I32 && src1->type == GGML_TYPE_I32) {
+        ggml_cpy_i32_i32_sycl (src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream);
+    } else {
+        fprintf(stderr, "%s: unsupported type combination (%s to %s)\n", __func__,
+                ggml_type_name(src0->type), ggml_type_name(src1->type));
+        GGML_ABORT("fatal error");
+    }
+
+    (void) dst;
+}
+catch (sycl::exception const &exc) {
+  std::cerr << exc.what() << "Exception caught at file:" << __FILE__
+            << ", line:" << __LINE__ << std::endl;
+  std::exit(1);
+}
+
+static void ggml_sycl_dup(ggml_backend_sycl_context & ctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
+    // TODO: why do we pass dst as src1 here?
+    ggml_sycl_cpy(ctx, src0, dst, nullptr);
+    (void) src1;
+}
+
+static void ggml_sycl_diag_mask_inf(ggml_backend_sycl_context & ctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
+    ggml_sycl_op_flatten(ctx, src0, src1, dst, ggml_sycl_op_diag_mask_inf);
+}
+
+static void ggml_sycl_soft_max(ggml_backend_sycl_context & ctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
+    ggml_sycl_op_flatten(ctx, src0, src1, dst, ggml_sycl_op_soft_max);
+}
+
+static void ggml_sycl_rope(ggml_backend_sycl_context & ctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
+    GGML_ASSERT(ggml_is_contiguous(src0)); // TODO: this restriction is temporary until non-cont support is implemented
+    ggml_sycl_op_flatten(ctx, src0, src1, dst, ggml_sycl_op_rope);
+}
+
+static void ggml_sycl_pool2d(ggml_backend_sycl_context & ctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
+    ggml_sycl_op_flatten(ctx, src0, src1, dst, ggml_sycl_op_pool2d);
+}
+
+static void ggml_sycl_im2col(ggml_backend_sycl_context & ctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
+    ggml_sycl_op_flatten(ctx, src0, src1, dst, ggml_sycl_op_im2col);
+}
+
+static void ggml_sycl_sum_rows(ggml_backend_sycl_context & ctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
+    GGML_ASSERT(ggml_is_contiguous(src0));
+    ggml_sycl_op_flatten(ctx, src0, src1, dst, ggml_sycl_op_sum_rows);
+}
+
+static void ggml_sycl_argsort(ggml_backend_sycl_context & ctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
+    GGML_ASSERT(ggml_is_contiguous(src0));
+    ggml_sycl_op_flatten(ctx, src0, src1, dst, ggml_sycl_op_argsort);
+}
+
+static void ggml_sycl_nop(ggml_backend_sycl_context & ctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
+    (void) src0;
+    (void) src1;
+    (void) dst;
+}
+
+static size_t ggml_nbytes_split(const struct ggml_tensor * tensor, int nrows_split) {
+    static_assert(GGML_MAX_DIMS == 4, "GGML_MAX_DIMS is not 4 - update this function");
+
+    return nrows_split*ggml_row_size(tensor->type, tensor->ne[0]);
+}
+
+void ggml_sycl_set_main_device(const int main_device) try {
+    if (dpct::get_current_device_id() == main_device) return;
+    check_allow_gpu_index(main_device);
+    dpct::select_device(main_device);
+
+    if (g_ggml_sycl_debug) {
+        dpct::device_info prop;
+        SYCL_CHECK(CHECK_TRY_ERROR(dpct::get_device_info(
+            prop, dpct::dev_mgr::instance().get_device(main_device))));
+        fprintf(stderr, "Using device %d (%s) as main device\n",
+                main_device, prop.get_name());
+    }
+}
+catch (sycl::exception const &exc) {
+  std::cerr << exc.what() << "Exception caught at file:" << __FILE__
+            << ", line:" << __LINE__ << std::endl;
+  std::exit(1);
+}
+
+bool ggml_sycl_compute_forward(ggml_backend_sycl_context & ctx, struct ggml_tensor * tensor) {
+    if (!g_sycl_loaded) return false;
+
+    ggml_sycl_func_t func;
+
+    switch (tensor->op) {
+        case GGML_OP_CONV_TRANSPOSE_1D:
+            func = ggml_sycl_op_conv_transpose_1d;
+            break;
+        case GGML_OP_REPEAT:
+            func = ggml_sycl_repeat;
+            break;
+        case GGML_OP_GET_ROWS:
+            func = ggml_sycl_get_rows;
+            break;
+        case GGML_OP_DUP:
+            func = ggml_sycl_dup;
+            break;
+        case GGML_OP_ADD:
+            func = ggml_sycl_add;
+            break;
+        case GGML_OP_ACC:
+            func = ggml_sycl_acc;
+            break;
+        case GGML_OP_MUL:
+            func = ggml_sycl_mul;
+            break;
+        case GGML_OP_DIV:
+            func = ggml_sycl_div;
+            break;
+        case GGML_OP_UNARY:
+            switch (ggml_get_unary_op(tensor)) {
+                case GGML_UNARY_OP_GELU:
+                    func = ggml_sycl_gelu;
+                    break;
+                case GGML_UNARY_OP_SILU:
+                    func = ggml_sycl_silu;
+                    break;
+                case GGML_UNARY_OP_GELU_QUICK:
+                    func = ggml_sycl_gelu_quick;
+                    break;
+                case GGML_UNARY_OP_TANH:
+                    func = ggml_sycl_tanh;
+                    break;
+                case GGML_UNARY_OP_RELU:
+                    func = ggml_sycl_relu;
+                    break;
+                case GGML_UNARY_OP_HARDSIGMOID:
+                    func = ggml_sycl_hardsigmoid;
+                    break;
+                case GGML_UNARY_OP_HARDSWISH:
+                    func = ggml_sycl_hardswish;
+                    break;
+                default:
+                    return false;
+            }
+            break;
+        case GGML_OP_NORM:
+            func = ggml_sycl_norm;
+            break;
+        case GGML_OP_GROUP_NORM:
+            func = ggml_sycl_group_norm;
+            break;
+        case GGML_OP_CONCAT:
+            func = ggml_sycl_op_concat;
+            break;
+        case GGML_OP_UPSCALE:
+            func = ggml_sycl_upscale;
+            break;
+        case GGML_OP_PAD:
+            func = ggml_sycl_pad;
+            break;
+        case GGML_OP_LEAKY_RELU:
+            func = ggml_sycl_leaky_relu;
+            break;
+        case GGML_OP_RMS_NORM:
+            func = ggml_sycl_rms_norm;
+            break;
+        case GGML_OP_MUL_MAT:
+            if (tensor->src[0]->ne[3] != tensor->src[1]->ne[3]) {
+                return false;
+            }
+            func = ggml_sycl_mul_mat;
+            break;
+        case GGML_OP_MUL_MAT_ID:
+            if (tensor->src[0]->ne[3] != tensor->src[1]->ne[3]) {
+                return false;
+            }
+            func = ggml_sycl_mul_mat_id;
+            break;
+        case GGML_OP_SCALE:
+            func = ggml_sycl_scale;
+            break;
+        case GGML_OP_SQR:
+            func = ggml_sycl_sqr;
+            break;
+        case GGML_OP_CLAMP:
+            func = ggml_sycl_clamp;
+            break;
+        case GGML_OP_CPY:
+            func = ggml_sycl_cpy;
+            break;
+        case GGML_OP_CONT:
+            func = ggml_sycl_dup;
+            break;
+        case GGML_OP_NONE:
+        case GGML_OP_RESHAPE:
+        case GGML_OP_VIEW:
+        case GGML_OP_PERMUTE:
+        case GGML_OP_TRANSPOSE:
+            func = ggml_sycl_nop;
+            break;
+        case GGML_OP_DIAG_MASK_INF:
+            func = ggml_sycl_diag_mask_inf;
+            break;
+        case GGML_OP_SOFT_MAX:
+            func = ggml_sycl_soft_max;
+            break;
+        case GGML_OP_ROPE:
+            func = ggml_sycl_rope;
+            break;
+        case GGML_OP_IM2COL:
+            func = ggml_sycl_im2col;
+            break;
+        case GGML_OP_POOL_2D:
+            func = ggml_sycl_pool2d;
+            break;
+        case GGML_OP_SUM_ROWS:
+            func = ggml_sycl_sum_rows;
+            break;
+        case GGML_OP_ARGSORT:
+            func = ggml_sycl_argsort;
+            break;
+        case GGML_OP_TIMESTEP_EMBEDDING:
+            func = ggml_sycl_op_timestep_embedding;
+            break;
+        default:
+            return false;
+    }
+
+    if (tensor->src[0] != nullptr && ggml_backend_buffer_is_sycl_split(tensor->src[0]->buffer)) {
+        ggml_sycl_set_peer_access(tensor->src[1]->ne[1], ctx.device);
+    }
+
+    func(ctx, tensor->src[0], tensor->src[1], tensor);
+    return true;
+}
+
+GGML_API void   ggml_sycl_get_gpu_list(int *id_list, int max_len) try {
+    GGML_SYCL_DEBUG("[SYCL] call ggml_sycl_get_gpu_list\n");
+    for(int i=0;i<max_len;i++) id_list[i] = -1;
+
+    for (int i=0;i< ggml_sycl_info().device_count;i++){
+        if (i>=max_len) break;
+        id_list[i] = i;
+    }
+    return;
+}
+catch (sycl::exception const &exc) {
+  std::cerr << exc.what() << "Exception caught at file:" << __FILE__
+            << ", line:" << __LINE__ << std::endl;
+  std::exit(1);
+}
+
+int ggml_sycl_get_device_count() try {
+    int device_count;
+    if (CHECK_TRY_ERROR(device_count =
+                             dpct::dev_mgr::instance().device_count()) != 0) {
+        return 0;
+    }
+    return device_count;
+}
+catch (sycl::exception const &exc) {
+  std::cerr << exc.what() << "Exception caught at file:" << __FILE__
+            << ", line:" << __LINE__ << std::endl;
+  std::exit(1);
+}
+
+GGML_API void ggml_sycl_get_device_description(int device, char *description,
+                                      size_t description_size) try {
+    GGML_SYCL_DEBUG("[SYCL] call ggml_sycl_get_device_description\n");
+    dpct::device_info prop;
+    SYCL_CHECK(CHECK_TRY_ERROR(dpct::get_device_info(
+        prop, dpct::dev_mgr::instance().get_device(device))));
+    snprintf(description, description_size, "%s", prop.get_name());
+}
+catch (sycl::exception const &exc) {
+  std::cerr << exc.what() << "Exception caught at file:" << __FILE__
+            << ", line:" << __LINE__ << std::endl;
+  std::exit(1);
+}
+
+void ggml_backend_sycl_get_device_memory(int device, size_t *free,
+                                                   size_t *total) try {
+    GGML_SYCL_DEBUG("[SYCL] call ggml_backend_sycl_get_device_memory\n");
+    ggml_sycl_set_device(device);
+
+    /*
+    DPCT1009:218: SYCL uses exceptions to report errors and does not use the
+    error codes. The original code was commented out and a warning string was
+    inserted. You need to rewrite this code.
+    */
+    /*
+    DPCT1106:217: 'cudaMemGetInfo' was migrated with the Intel extensions for
+    device information which may not be supported by all compilers or runtimes.
+    You may need to adjust the code.
+    */
+    SYCL_CHECK(CHECK_TRY_ERROR(
+        dpct::dev_mgr::instance().get_device(device).get_memory_info(*free, *total)));
+}
+catch (sycl::exception const &exc) {
+  std::cerr << exc.what() << "Exception caught at file:" << __FILE__
+            << ", line:" << __LINE__ << std::endl;
+  std::exit(1);
+}
+
+////////////////////////////////////////////////////////////////////////////////
+
+// backend interface
+
+#define UNUSED GGML_UNUSED
+
+// sycl buffer
+
+struct ggml_backend_sycl_buffer_context {
+    int device;
+    void * dev_ptr = nullptr;
+    queue_ptr stream;
+    std::string name;
+
+     ggml_backend_sycl_buffer_context(int device, void * dev_ptr, queue_ptr stream) :
+        device(device), dev_ptr(dev_ptr), stream(stream) {
+            check_allow_gpu_index(device);
+            name = (GGML_SYCL_NAME + std::to_string(device));
+        }
+
+
+    ~ggml_backend_sycl_buffer_context() {
+        if (dev_ptr != nullptr) {
+            ggml_sycl_set_device(device);
+            SYCL_CHECK(CHECK_TRY_ERROR(sycl::free(dev_ptr, *stream)));
+        }
+    }
+};
+
+static const char * ggml_backend_sycl_buffer_get_name(ggml_backend_buffer_t buffer) {
+    ggml_backend_sycl_buffer_context * ctx = (ggml_backend_sycl_buffer_context *)buffer->context;
+    return ctx->name.c_str();
+}
+
+static bool ggml_backend_buffer_is_sycl(ggml_backend_buffer_t buffer) {
+    return buffer->iface.get_name == ggml_backend_sycl_buffer_get_name;
+}
+
+static void
+ggml_backend_sycl_buffer_free_buffer(ggml_backend_buffer_t buffer) try {
+    ggml_backend_sycl_buffer_context * ctx = ( ggml_backend_sycl_buffer_context *)buffer->context;
+    ggml_sycl_set_device(ctx->device);
+
+    delete ctx;
+}
+catch (sycl::exception const &exc) {
+  std::cerr << exc.what() << "Exception caught at file:" << __FILE__
+            << ", line:" << __LINE__ << std::endl;
+  std::exit(1);
+}
+
+static void * ggml_backend_sycl_buffer_get_base(ggml_backend_buffer_t buffer) {
+    ggml_backend_sycl_buffer_context * ctx = ( ggml_backend_sycl_buffer_context *)buffer->context;
+    return ctx->dev_ptr;
+}
+
+static void
+ggml_backend_sycl_buffer_init_tensor(ggml_backend_buffer_t buffer,
+                                     ggml_tensor *tensor) try {
+    ggml_backend_sycl_buffer_context * ctx = (ggml_backend_sycl_buffer_context *)buffer->context;
+
+    if (tensor->view_src != NULL && tensor->view_offs == 0) {
+        assert(tensor->view_src->buffer->buft == buffer->buft);
+        tensor->backend = tensor->view_src->backend;
+        tensor->extra = tensor->view_src->extra;
+        return;
+    }
+
+
+    if (ggml_is_quantized(tensor->type)) {
+        // initialize padding to 0 to avoid possible NaN values
+        size_t original_size = ggml_nbytes(tensor);
+        size_t padded_size = ggml_backend_buft_get_alloc_size(buffer->buft, tensor);
+
+        if (padded_size > original_size && tensor->view_src == nullptr) {
+            SYCL_CHECK(CHECK_TRY_ERROR(ctx->stream->memset(
+                (char *)tensor->data + original_size, 0,
+                padded_size - original_size).wait()));
+        }
+    }
+}
+catch (sycl::exception const &exc) {
+  std::cerr << exc.what() << "Exception caught at file:" << __FILE__
+            << ", line:" << __LINE__ << std::endl;
+  std::exit(1);
+}
+
+static void ggml_backend_sycl_buffer_set_tensor(ggml_backend_buffer_t buffer,
+                                                ggml_tensor *tensor,
+                                                const void *data, size_t offset,
+                                                size_t size) try {
+
+    ggml_backend_sycl_buffer_context * ctx = ( ggml_backend_sycl_buffer_context *)buffer->context;
+
+    ggml_sycl_set_device(ctx->device);
+    auto stream = &(dpct::dev_mgr::instance().get_device(ctx->device).default_queue());
+    SYCL_CHECK(
+        CHECK_TRY_ERROR(dpct::dev_mgr::instance().get_device(ctx->device).queues_wait_and_throw()));
+    char* host_buf = (char*)malloc(size);
+    memcpy(host_buf, data, size);
+    SYCL_CHECK(
+        CHECK_TRY_ERROR((*stream).memcpy((char *)tensor->data + offset, host_buf, size)
+                             .wait()));
+    free(host_buf);
+}
+catch (sycl::exception const &exc) {
+  std::cerr << exc.what() << "Exception caught at file:" << __FILE__
+            << ", line:" << __LINE__ << std::endl;
+  std::exit(1);
+}
+
+static void ggml_backend_sycl_buffer_get_tensor(ggml_backend_buffer_t buffer,
+                                                const ggml_tensor *tensor,
+                                                void *data, size_t offset,
+                                                size_t size) try {
+
+    ggml_backend_sycl_buffer_context * ctx = ( ggml_backend_sycl_buffer_context *)buffer->context;
+
+    ggml_sycl_set_device(ctx->device);
+    auto stream = dpct::dev_mgr::instance().get_device(ctx->device).default_queue();
+
+    SYCL_CHECK(CHECK_TRY_ERROR(
+        stream.memcpy(data, (const char *)tensor->data + offset, size)
+            .wait()));
+}
+catch (sycl::exception const &exc) {
+  std::cerr << exc.what() << "Exception caught at file:" << __FILE__
+            << ", line:" << __LINE__ << std::endl;
+  std::exit(1);
+}
+
+static bool
+ggml_backend_sycl_buffer_cpy_tensor(ggml_backend_buffer_t buffer,
+                                    const ggml_tensor *src,
+                                    ggml_tensor *dst) try {
+    if (ggml_backend_buffer_is_sycl(src->buffer)) {
+        ggml_backend_sycl_buffer_context * src_ctx = (ggml_backend_sycl_buffer_context *)src->buffer->context;
+        ggml_backend_sycl_buffer_context * dst_ctx = (ggml_backend_sycl_buffer_context *)dst->buffer->context;
+
+        ggml_sycl_set_device(src_ctx->device);
+        /*
+        DPCT1009:198: SYCL uses exceptions to report errors and does not use the
+        error codes. The original code was commented out and a warning string
+        was inserted. You need to rewrite this code.
+        */
+        SYCL_CHECK(CHECK_TRY_ERROR(
+            dpct::dev_mgr::instance().get_device(src_ctx->device).queues_wait_and_throw()));
+        ggml_sycl_set_device(dst_ctx->device);
+        /*
+        DPCT1009:199: SYCL uses exceptions to report errors and does not use the
+        error codes. The original code was commented out and a warning string
+        was inserted. You need to rewrite this code.
+        */
+        SYCL_CHECK(CHECK_TRY_ERROR(
+            dpct::dev_mgr::instance().get_device(dst_ctx->device).queues_wait_and_throw()));
+        /*
+        DPCT1009:200: SYCL uses exceptions to report errors and does not use the
+        error codes. The original code was commented out and a warning string
+        was inserted. You need to rewrite this code.
+        */
+
+        queue_ptr stream_dst = dst_ctx->stream;
+        queue_ptr stream_src = src_ctx->stream;
+        size_t size = ggml_nbytes(src);
+
+        //todo. it's dirty solutino to walkaroud known issue:device2device cross GPUs.
+        dev2dev_memcpy(*stream_dst, *stream_src, dst->data, src->data, size);
+
+//todo, it's known issue：error in device2device cross GPUs. reused when the issue is fixed. DON"T remove
+#if 0
+        SYCL_CHECK(CHECK_TRY_ERROR((*stream).memcpy(
+            (char *)dst->data, (const char *)src->data, size).wait()));
+
+        /*
+        DPCT1009:201: SYCL uses exceptions to report errors and does not use the
+        error codes. The original code was commented out and a warning string
+        was inserted. You need to rewrite this code.
+        */
+        SYCL_CHECK(CHECK_TRY_ERROR(
+            dpct::dev_mgr::instance().get_device(dst_ctx->device).queues_wait_and_throw()));
+#endif
+        return true;
+    }
+    return false;
+}
+catch (sycl::exception const &exc) {
+  std::cerr << exc.what() << "Exception caught at file:" << __FILE__
+            << ", line:" << __LINE__ << std::endl;
+  std::exit(1);
+}
+
+
+static void ggml_backend_sycl_buffer_clear(ggml_backend_buffer_t buffer,
+                                           uint8_t value) try {
+     ggml_backend_sycl_buffer_context * ctx = ( ggml_backend_sycl_buffer_context *)buffer->context;
+
+    ggml_sycl_set_device(ctx->device);
+    queue_ptr stream = ctx->stream;
+    SYCL_CHECK(
+        CHECK_TRY_ERROR(dpct::get_current_device().queues_wait_and_throw()));
+
+    SYCL_CHECK(CHECK_TRY_ERROR((*stream)
+                                    .memset(ctx->dev_ptr, value, buffer->size)
+                                    .wait()));
+}
+catch (sycl::exception const &exc) {
+  std::cerr << exc.what() << "Exception caught at file:" << __FILE__
+            << ", line:" << __LINE__ << std::endl;
+  std::exit(1);
+}
+
+static struct ggml_backend_buffer_i ggml_backend_sycl_buffer_interface = {
+    /* .get_name        = */ ggml_backend_sycl_buffer_get_name,
+    /* .free_buffer     = */ ggml_backend_sycl_buffer_free_buffer,
+    /* .get_base        = */ ggml_backend_sycl_buffer_get_base,
+    /* .init_tensor     = */ ggml_backend_sycl_buffer_init_tensor,
+    /* .memset_tensor   = */ NULL,
+    /* .set_tensor      = */ ggml_backend_sycl_buffer_set_tensor,
+    /* .get_tensor      = */ ggml_backend_sycl_buffer_get_tensor,
+    /* .cpy_tensor      = */ ggml_backend_sycl_buffer_cpy_tensor,
+    /* .clear           = */ ggml_backend_sycl_buffer_clear,
+    /* .reset           = */ NULL,
+};
+
+// sycl buffer type
+struct ggml_backend_sycl_buffer_type_context {
+    int device;
+    std::string name;
+
+    // each buffer type has its own stream
+    queue_ptr stream = nullptr;
+};
+
+static const char * ggml_backend_sycl_buffer_type_name(ggml_backend_buffer_type_t buft) {
+    ggml_backend_sycl_buffer_type_context * ctx = (ggml_backend_sycl_buffer_type_context *)buft->context;
+
+    return ctx->name.c_str();
+}
+static ggml_backend_buffer_t
+ggml_backend_sycl_buffer_type_alloc_buffer(ggml_backend_buffer_type_t buft,
+                                           size_t size) try {
+    ggml_backend_sycl_buffer_type_context * buft_ctx = (ggml_backend_sycl_buffer_type_context *)buft->context;
+    ggml_sycl_set_device(buft_ctx->device);
+    const queue_ptr stream = buft_ctx->stream;
+    size = std::max(size, (size_t)1); // syclMalloc returns null for size 0
+
+    void * dev_ptr;
+    SYCL_CHECK(CHECK_TRY_ERROR(dev_ptr = (void *)sycl::malloc_device(
+                                    size, *stream)));
+    if (!dev_ptr) {
+        fprintf(stderr, "%s: can't malloc %lu Bytes memory on device", __func__, size);
+        return nullptr;
+    }
+    ggml_backend_sycl_buffer_context * ctx = new  ggml_backend_sycl_buffer_context(buft_ctx->device, dev_ptr, buft_ctx->stream);
+    return ggml_backend_buffer_init(buft, ggml_backend_sycl_buffer_interface, ctx, size);
+}
+catch (sycl::exception const &exc) {
+  std::cerr << exc.what() << "Exception caught at file:" << __FILE__
+            << ", line:" << __LINE__ << std::endl;
+  std::exit(1);
+}
+
+static size_t ggml_backend_sycl_buffer_type_get_alignment(ggml_backend_buffer_type_t buft) {
+    return 128;
+    UNUSED(buft);
+}
+
+static size_t ggml_backend_sycl_buffer_type_get_max_size(ggml_backend_buffer_type_t buft) {
+    return dpct::get_current_device().get_max_mem_alloc_size();
+
+    UNUSED(buft);
+}
+
+static size_t ggml_backend_sycl_buffer_type_get_alloc_size(ggml_backend_buffer_type_t buft, const ggml_tensor * tensor) {
+    size_t size = ggml_nbytes(tensor);
+    int64_t ne0 = tensor->ne[0];
+
+    if (ggml_is_quantized(tensor->type)) {
+        if (ne0 % MATRIX_ROW_PADDING != 0) {
+            size += ggml_row_size(tensor->type, MATRIX_ROW_PADDING - ne0 % MATRIX_ROW_PADDING);
+        }
+    }
+
+    return size;
+
+    UNUSED(buft);
+}
+
+static ggml_backend_buffer_type_i ggml_backend_sycl_buffer_type_interface = {
+    /* .get_name         = */ ggml_backend_sycl_buffer_type_name,
+    /* .alloc_buffer     = */ ggml_backend_sycl_buffer_type_alloc_buffer,
+    /* .get_alignment    = */ ggml_backend_sycl_buffer_type_get_alignment,
+    /* .get_max_size     = */ ggml_backend_sycl_buffer_type_get_max_size,
+    /* .get_alloc_size   = */ ggml_backend_sycl_buffer_type_get_alloc_size,
+    /* .is_host          = */ nullptr,
+};
+
+ggml_backend_buffer_type_t ggml_backend_sycl_buffer_type(int device) {
+    static std::mutex mutex;
+    std::lock_guard<std::mutex> lock(mutex);
+
+    GGML_SYCL_DEBUG("[SYCL] call ggml_backend_sycl_buffer_type\n");
+
+    if (device>=ggml_sycl_info().device_count or device<0) {
+        printf("ggml_backend_sycl_buffer_type error: device_index:%d is out of range [0, %d], miss to call ggml_backend_sycl_set_single_device()\n",
+            device, ggml_sycl_info().device_count-1);
+        GGML_ASSERT(device<ggml_sycl_info().device_count);
+    }
+    static struct ggml_backend_buffer_type ggml_backend_sycl_buffer_types[GGML_SYCL_MAX_DEVICES];
+
+    static bool ggml_backend_sycl_buffer_type_initialized = false;
+
+    if (!ggml_backend_sycl_buffer_type_initialized) {
+        for (int i = 0; i < ggml_sycl_info().device_count; i++) {
+            auto & device_i = dpct::dev_mgr::instance().get_device(i);
+            queue_ptr stream = &(device_i.default_queue());
+            ggml_backend_sycl_buffer_types[i] = {
+                /* .iface    = */ ggml_backend_sycl_buffer_type_interface,
+                /* .device   = */ nullptr,
+                /* .context  = */ new ggml_backend_sycl_buffer_type_context{i, GGML_SYCL_NAME + std::to_string(i), stream},
+            };
+        }
+        ggml_backend_sycl_buffer_type_initialized = true;
+    }
+    return &ggml_backend_sycl_buffer_types[device];
+}
+
+ggml_backend_buffer_type_t ggml_backend_sycl_buffer_type(ggml_backend_sycl_context * ctx) {
+    GGML_SYCL_DEBUG("[SYCL] call ggml_backend_sycl_buffer_type\n");
+
+    int device = ctx->device;
+    if (device>=ggml_sycl_info().device_count or device<0) {
+        printf("ggml_backend_sycl_buffer_type error: device_index:%d is out of range [0, %d], miss to call ggml_backend_sycl_set_single_device()\n",
+            device, ggml_sycl_info().device_count-1);
+        GGML_ASSERT(device<ggml_sycl_info().device_count);
+    }
+    static struct ggml_backend_buffer_type ggml_backend_sycl_buffer_types[GGML_SYCL_MAX_DEVICES];
+
+    static bool ggml_backend_sycl_buffer_type_initialized = false;
+
+    if (!ggml_backend_sycl_buffer_type_initialized) {
+        for (int i = 0; i < ggml_sycl_info().device_count; i++) {
+            ggml_backend_sycl_buffer_types[i] = {
+                /* .iface    = */ ggml_backend_sycl_buffer_type_interface,
+                /* .device   = */ nullptr,
+                /* .context  = */ new ggml_backend_sycl_buffer_type_context{i, GGML_SYCL_NAME + std::to_string(i), ctx->stream(i, 0)},
+            };
+        }
+        ggml_backend_sycl_buffer_type_initialized = true;
+    }
+    return &ggml_backend_sycl_buffer_types[device];
+}
+
+// sycl split buffer type
+static void get_row_split(int64_t * row_low, int64_t * row_high, const ggml_tensor * tensor, const std::array<float, GGML_SYCL_MAX_DEVICES> & tensor_split, int id) {
+    const int64_t nrows = ggml_nrows(tensor);
+    const int64_t rounding = get_row_rounding(tensor->type, tensor_split);
+
+    *row_low = id == 0 ? 0 : nrows*tensor_split[id];
+    *row_low -= *row_low % rounding;
+    if (id == ggml_sycl_info().device_count - 1) {
+        *row_high = nrows;
+    } else {
+        *row_high = nrows*tensor_split[id + 1];
+        *row_high -= *row_high % rounding;
+    }
+}
+
+struct ggml_backend_sycl_split_buffer_context {
+    ~ggml_backend_sycl_split_buffer_context() try {
+        for (ggml_tensor_extra_gpu * extra : tensor_extras) {
+            for (int i = 0; i < ggml_sycl_info().device_count; ++i) {
+                for (int64_t is = 0; is < GGML_SYCL_MAX_STREAMS; ++is) {
+                    if (extra->events[i][is] != nullptr) {
+                        /*
+                        DPCT1009:206: SYCL uses exceptions to report errors and
+                        does not use the error codes. The original code was
+                        commented out and a warning string was inserted. You
+                        need to rewrite this code.
+                        */
+                        SYCL_CHECK(CHECK_TRY_ERROR(
+                            dpct::destroy_event(extra->events[i][is])));
+                    }
+                }
+                if (extra->data_device[i] != nullptr) {
+                    /*
+                    DPCT1009:207: SYCL uses exceptions to report errors and does
+                    not use the error codes. The original code was commented out
+                    and a warning string was inserted. You need to rewrite this
+                    code.
+                    */
+                    ggml_sycl_set_device(i);
+                    SYCL_CHECK(CHECK_TRY_ERROR(sycl::free(
+                        extra->data_device[i], *(streams[i]))));
+                }
+            }
+            delete extra;
+        }
+    }
+    catch (sycl::exception const &exc) {
+      std::cerr << exc.what() << "Exception caught at file:" << __FILE__
+                << ", line:" << __LINE__ << std::endl;
+      std::exit(1);
+    }
+
+    std::vector<ggml_tensor_extra_gpu *> tensor_extras;
+    std::vector<queue_ptr> streams;
+};
+
+static const char * ggml_backend_sycl_split_buffer_get_name(ggml_backend_buffer_t buffer) {
+    return GGML_SYCL_NAME "_Split";
+
+    UNUSED(buffer);
+}
+
+static bool ggml_backend_buffer_is_sycl_split(ggml_backend_buffer_t buffer) {
+   return buffer->iface.get_name == ggml_backend_sycl_split_buffer_get_name;
+}
+
+static void ggml_backend_sycl_split_buffer_free_buffer(ggml_backend_buffer_t buffer) {
+    ggml_backend_sycl_split_buffer_context * ctx = (ggml_backend_sycl_split_buffer_context *)buffer->context;
+    delete ctx;
+}
+
+static void * ggml_backend_sycl_split_buffer_get_base(ggml_backend_buffer_t buffer) {
+    // the pointers are stored in the tensor extras, this is just a dummy address and never dereferenced
+    return (void *)0x1000;
+
+    UNUSED(buffer);
+}
+
+static void
+ggml_backend_sycl_split_buffer_init_tensor(ggml_backend_buffer_t buffer,
+                                           ggml_tensor *tensor) try {
+    GGML_ASSERT(tensor->view_src == nullptr); // views of split tensors are not supported
+
+    ggml_backend_sycl_split_buffer_context * ctx = (ggml_backend_sycl_split_buffer_context *)buffer->context;
+    ggml_backend_sycl_split_buffer_type_context * buft_ctx = (ggml_backend_sycl_split_buffer_type_context *)buffer->buft->context;
+
+    const int64_t ne0 = tensor->ne[0];
+
+    ggml_tensor_extra_gpu * extra = new ggml_tensor_extra_gpu{};
+
+    ctx->tensor_extras.push_back(extra);
+        ctx->streams.push_back(&(dpct::get_current_device().default_queue()));
+
+    for (int i = 0; i < ggml_sycl_info().device_count; ++i) {
+        int64_t row_low, row_high;
+        get_row_split(&row_low, &row_high, tensor, buft_ctx->tensor_split, i);
+
+        int64_t nrows_split = row_high - row_low;
+        if (nrows_split == 0) {
+            continue;
+        }
+
+        size_t size = ggml_nbytes_split(tensor, nrows_split);
+        const size_t original_size = size;
+
+        // pad last row to a multiple of 512 elements to avoid out-of-bounds memory accesses
+        if (ne0 % MATRIX_ROW_PADDING != 0) {
+            size += ggml_row_size(tensor->type, MATRIX_ROW_PADDING - ne0 % MATRIX_ROW_PADDING);
+        }
+
+        // FIXME: do not crash if cudaMalloc fails
+        // currently, init_tensor cannot fail, it needs to be fixed in ggml-backend first
+        ggml_sycl_set_device(i);
+        const queue_ptr stream = ctx->streams[i];
+        char * buf;
+        /*
+        DPCT1009:208: SYCL uses exceptions to report errors and does not use the
+        error codes. The original code was commented out and a warning string
+        was inserted. You need to rewrite this code.
+        */
+        SYCL_CHECK(CHECK_TRY_ERROR(buf = (char *)sycl::malloc_device(
+                                        size, *stream)));
+        if (!buf) {
+            char err_buf[1024];
+            snprintf(err_buf, 1023, "%s: can't malloc %lu Bytes memory on device", __func__, size);
+            throw std::runtime_error(err_buf);
+        }
+        // set padding to 0 to avoid possible NaN values
+        if (size > original_size) {
+            /*
+            DPCT1009:209: SYCL uses exceptions to report errors and does not use
+            the error codes. The original code was commented out and a warning
+            string was inserted. You need to rewrite this code.
+            */
+            SYCL_CHECK(CHECK_TRY_ERROR(
+                (*stream)
+                    .memset(buf + original_size, 0, size - original_size)
+                    .wait()));
+        }
+
+        extra->data_device[i] = buf;
+
+        for (int64_t is = 0; is < GGML_SYCL_MAX_STREAMS; ++is) {
+            /*
+            DPCT1009:210: SYCL uses exceptions to report errors and does not use
+            the error codes. The original code was commented out and a warning
+            string was inserted. You need to rewrite this code.
+            */
+            SYCL_CHECK(
+                CHECK_TRY_ERROR(extra->events[i][is] = new sycl::event()));
+        }
+    }
+    tensor->backend = GGML_BACKEND_TYPE_GPU_SPLIT;
+    tensor->extra = extra;
+}
+catch (sycl::exception const &exc) {
+  std::cerr << exc.what() << "Exception caught at file:" << __FILE__
+            << ", line:" << __LINE__ << std::endl;
+  std::exit(1);
+}
+
+static void
+ggml_backend_sycl_split_buffer_set_tensor(ggml_backend_buffer_t buffer,
+                                          ggml_tensor *tensor, const void *data,
+                                          size_t offset, size_t size) try {
+    // split tensors must always be set in their entirety at once
+    GGML_ASSERT(offset == 0);
+    GGML_ASSERT(size == ggml_nbytes(tensor));
+
+    ggml_backend_sycl_split_buffer_context * ctx = (ggml_backend_sycl_split_buffer_context *)buffer->context;
+    ggml_backend_sycl_split_buffer_type_context * buft_ctx = (ggml_backend_sycl_split_buffer_type_context *)buffer->buft->context;
+
+    const int64_t ne0 = tensor->ne[0];
+    const size_t nb1 = tensor->nb[1];
+    ggml_tensor_extra_gpu * extra = (ggml_tensor_extra_gpu *)tensor->extra;
+
+    for (int i = 0; i < ggml_sycl_info().device_count; ++i) {
+        int64_t row_low, row_high;
+        get_row_split(&row_low, &row_high, tensor, buft_ctx->tensor_split, i);
+
+        int64_t nrows_split = row_high - row_low;
+        if (nrows_split == 0) {
+            continue;
+        }
+
+        const size_t offset_split = row_low*nb1;
+        size_t size = ggml_nbytes_split(tensor, nrows_split);
+        const size_t original_size = size;
+
+        // pad last row to a multiple of 512 elements to avoid out-of-bounds memory accesses
+        if (ne0 % MATRIX_ROW_PADDING != 0) {
+            size += ggml_row_size(tensor->type, MATRIX_ROW_PADDING - ne0 % MATRIX_ROW_PADDING);
+        }
+
+        const char * buf_host = (const char *)data + offset_split;
+        /*
+        DPCT1009:211: SYCL uses exceptions to report errors and does not use the
+        error codes. The original code was commented out and a warning string
+        was inserted. You need to rewrite this code.
+        */
+        ggml_sycl_set_device(i);
+        const queue_ptr stream = ctx->streams[i];
+        SYCL_CHECK(CHECK_TRY_ERROR(
+            (*stream)
+                .memcpy(extra->data_device[i], buf_host, original_size)
+                .wait()));
+    }
+}
+catch (sycl::exception const &exc) {
+  std::cerr << exc.what() << "Exception caught at file:" << __FILE__
+            << ", line:" << __LINE__ << std::endl;
+  std::exit(1);
+}
+
+static void
+ggml_backend_sycl_split_buffer_get_tensor(ggml_backend_buffer_t buffer,
+                                          const ggml_tensor *tensor, void *data,
+                                          size_t offset, size_t size) try {
+    // split tensors must always be set in their entirety at once
+    GGML_ASSERT(offset == 0);
+    GGML_ASSERT(size == ggml_nbytes(tensor));
+
+    ggml_backend_sycl_split_buffer_context * ctx = (ggml_backend_sycl_split_buffer_context *)buffer->context;
+    ggml_backend_sycl_split_buffer_type_context * buft_ctx = (ggml_backend_sycl_split_buffer_type_context *)buffer->buft->context;
+
+    const int64_t ne0 = tensor->ne[0];
+    const size_t nb1 = tensor->nb[1];
+    ggml_tensor_extra_gpu * extra = (ggml_tensor_extra_gpu *)tensor->extra;
+
+    for (int i = 0; i < ggml_sycl_info().device_count; ++i) {
+        int64_t row_low, row_high;
+        get_row_split(&row_low, &row_high, tensor, buft_ctx->tensor_split, i);
+
+        int64_t nrows_split = row_high - row_low;
+        if (nrows_split == 0) {
+            continue;
+        }
+
+        const size_t offset_split = row_low*nb1;
+        size_t size = ggml_nbytes_split(tensor, nrows_split);
+        const size_t original_size = size;
+
+        // pad last row to a multiple of 512 elements to avoid out-of-bounds memory accesses
+        if (ne0 % MATRIX_ROW_PADDING != 0) {
+            size += ggml_row_size(tensor->type, MATRIX_ROW_PADDING - ne0 % MATRIX_ROW_PADDING);
+        }
+
+        char * buf_host = (char *)data + offset_split;
+        /*
+        DPCT1009:212: SYCL uses exceptions to report errors and does not use the
+        error codes. The original code was commented out and a warning string
+        was inserted. You need to rewrite this code.
+        */
+        ggml_sycl_set_device(i);
+        const queue_ptr stream = ctx->streams[i];
+        SYCL_CHECK(CHECK_TRY_ERROR(
+            (*stream)
+                .memcpy(buf_host, extra->data_device[i], original_size)
+                .wait()));
+    }
+}
+catch (sycl::exception const &exc) {
+  std::cerr << exc.what() << "Exception caught at file:" << __FILE__
+            << ", line:" << __LINE__ << std::endl;
+  std::exit(1);
+}
+
+static void ggml_backend_sycl_split_buffer_clear(ggml_backend_buffer_t buffer, uint8_t value) {
+    UNUSED(buffer);
+    UNUSED(value);
+}
+
+static struct ggml_backend_buffer_i ggml_backend_sycl_split_buffer_interface = {
+    /* .get_name        = */ ggml_backend_sycl_split_buffer_get_name,
+    /* .free_buffer     = */ ggml_backend_sycl_split_buffer_free_buffer,
+    /* .get_base        = */ ggml_backend_sycl_split_buffer_get_base,
+    /* .init_tensor     = */ ggml_backend_sycl_split_buffer_init_tensor,
+    /* .memset_tensor   = */ NULL,
+    /* .set_tensor      = */ ggml_backend_sycl_split_buffer_set_tensor,
+    /* .get_tensor      = */ ggml_backend_sycl_split_buffer_get_tensor,
+    /* .cpy_tensor      = */ NULL,
+    /* .clear           = */ ggml_backend_sycl_split_buffer_clear,
+    /* .reset           = */ NULL,
+};
+
+static const char * ggml_backend_sycl_split_buffer_type_name(ggml_backend_buffer_type_t buft) {
+    return GGML_SYCL_NAME "_Split";
+
+    UNUSED(buft);
+}
+
+static ggml_backend_buffer_t ggml_backend_sycl_split_buffer_type_alloc_buffer(ggml_backend_buffer_type_t buft, size_t size) {
+    // since we don't know the exact split after rounding, we cannot allocate the device buffers at this point
+    // instead, we allocate them for each tensor separately in init_tensor
+    // however, the size still represents the maximum cumulative size of all the device buffers after the tensors are allocated,
+    // as returned by get_alloc_size. this limit is enforced during tensor allocation by ggml-alloc, so it must be correct.
+    ggml_backend_sycl_split_buffer_context * ctx = new ggml_backend_sycl_split_buffer_context();
+
+    return ggml_backend_buffer_init(buft, ggml_backend_sycl_split_buffer_interface, ctx, size);
+}
+
+static size_t ggml_backend_sycl_split_buffer_type_get_alignment(ggml_backend_buffer_type_t buft) {
+    return 128;
+    UNUSED(buft);
+}
+
+static size_t ggml_backend_sycl_split_buffer_type_get_alloc_size(ggml_backend_buffer_type_t buft, const ggml_tensor * tensor) {
+    ggml_backend_sycl_split_buffer_type_context * ctx = (ggml_backend_sycl_split_buffer_type_context *)buft->context;
+
+    size_t total_size = 0;
+
+    const int64_t ne0 = tensor->ne[0];
+
+    for (int i = 0; i < ggml_sycl_info().device_count; ++i) {
+        int64_t row_low, row_high;
+        get_row_split(&row_low, &row_high, tensor, ctx->tensor_split, i);
+
+        int64_t nrows_split = row_high - row_low;
+        if (nrows_split == 0) {
+            continue;
+        }
+
+        total_size += ggml_nbytes_split(tensor, nrows_split);
+
+        // pad last row to a multiple of 512 elements to avoid out-of-bounds memory accesses
+        if (ne0 % MATRIX_ROW_PADDING != 0) {
+            total_size += ggml_row_size(tensor->type, MATRIX_ROW_PADDING - ne0 % MATRIX_ROW_PADDING);
+        }
+    }
+
+    return total_size;
+}
+
+static bool ggml_backend_sycl_split_buffer_type_is_host(ggml_backend_buffer_type_t buft) {
+    return false;
+
+    UNUSED(buft);
+}
+
+static ggml_backend_buffer_type_i ggml_backend_sycl_split_buffer_type_interface = {
+    /* .get_name         = */ ggml_backend_sycl_split_buffer_type_name,
+    /* .alloc_buffer     = */ ggml_backend_sycl_split_buffer_type_alloc_buffer,
+    /* .get_alignment    = */ ggml_backend_sycl_split_buffer_type_get_alignment,
+    /* .get_max_size     = */ NULL, // defaults to SIZE_MAX
+    /* .get_alloc_size   = */ ggml_backend_sycl_split_buffer_type_get_alloc_size,
+    /* .is_host          = */ ggml_backend_sycl_split_buffer_type_is_host,
+};
+
+ggml_backend_buffer_type_t ggml_backend_sycl_split_buffer_type(const float * tensor_split) {
+    static std::mutex mutex;
+    std::lock_guard<std::mutex> lock(mutex);
+
+    GGML_SYCL_DEBUG("[SYCL] call ggml_backend_sycl_split_buffer_type\n");
+    ggml_check_sycl();
+    // FIXME: this is not thread safe
+    static std::map<std::array<float, GGML_SYCL_MAX_DEVICES>, struct ggml_backend_buffer_type> buft_map;
+
+    std::array<float, GGML_SYCL_MAX_DEVICES> tensor_split_arr = {};
+
+    bool all_zero = tensor_split == nullptr || std::all_of(tensor_split, tensor_split + GGML_SYCL_MAX_DEVICES, [](float x) { return x == 0.0f; });
+    if (all_zero) {
+        tensor_split_arr = ggml_sycl_info().default_tensor_split;
+    } else {
+        float split_sum = 0.0f;
+        for (int i = 0; i < ggml_sycl_info().device_count; ++i) {
+            tensor_split_arr[i] = split_sum;
+            split_sum += tensor_split[i];
+        }
+        for (int i = 0; i < ggml_sycl_info().device_count; ++i) {
+            tensor_split_arr[i] /= split_sum;
+        }
+    }
+
+    auto it = buft_map.find(tensor_split_arr);
+    if (it != buft_map.end()) {
+        return &it->second;
+    }
+
+    struct ggml_backend_buffer_type buft {
+        /* .iface   = */ ggml_backend_sycl_split_buffer_type_interface,
+        /* .device  = */ nullptr,
+        /* .context = */ new ggml_backend_sycl_split_buffer_type_context{tensor_split_arr},
+    };
+
+    auto result = buft_map.emplace(tensor_split_arr, buft);
+    return &result.first->second;
+}
+
+// host buffer type
+
+static const char * ggml_backend_sycl_host_buffer_type_name(ggml_backend_buffer_type_t buft) {
+    return GGML_SYCL_NAME "_Host";
+
+    UNUSED(buft);
+}
+
+static const char * ggml_backend_sycl_host_buffer_name(ggml_backend_buffer_t buffer) {
+    return GGML_SYCL_NAME "_Host";
+
+    UNUSED(buffer);
+}
+
+static void ggml_backend_sycl_host_buffer_free_buffer(ggml_backend_buffer_t buffer) {
+    ggml_sycl_host_free(buffer->context);
+}
+
+static ggml_backend_buffer_t ggml_backend_sycl_host_buffer_type_alloc_buffer(ggml_backend_buffer_type_t buft, size_t size) {
+    void * ptr = ggml_sycl_host_malloc(size);
+
+    if (ptr == nullptr) {
+        // fallback to cpu buffer
+        return ggml_backend_buft_alloc_buffer(ggml_backend_cpu_buffer_type(), size);
+    }
+
+    // FIXME: this is a hack to avoid having to implement a new buffer type
+    ggml_backend_buffer_t buffer = ggml_backend_cpu_buffer_from_ptr(ptr, size);
+    buffer->buft = buft;
+    buffer->iface.get_name = ggml_backend_sycl_host_buffer_name;
+    buffer->iface.free_buffer = ggml_backend_sycl_host_buffer_free_buffer;
+
+    return buffer;
+}
+
+ggml_backend_buffer_type_t ggml_backend_sycl_host_buffer_type() {
+    GGML_SYCL_DEBUG("[SYCL] call ggml_backend_sycl_host_buffer_type\n");
+    static struct ggml_backend_buffer_type ggml_backend_sycl_buffer_type_host = {
+        /* .iface    = */ {
+            /* .get_name         = */ ggml_backend_sycl_host_buffer_type_name,
+            /* .alloc_buffer     = */ ggml_backend_sycl_host_buffer_type_alloc_buffer,
+            /* .get_alignment    = */ ggml_backend_cpu_buffer_type()->iface.get_alignment,
+            /* .get_max_size     = */ NULL, // TODO: return device.maxBufferLength
+            /* .get_alloc_size   = */ ggml_backend_cpu_buffer_type()->iface.get_alloc_size,
+            /* .is_host          = */ ggml_backend_cpu_buffer_type()->iface.is_host,
+        },
+        /* .device   = */ nullptr,
+        /* .context  = */ nullptr,
+    };
+
+    return &ggml_backend_sycl_buffer_type_host;
+}
+
+// backend
+
+static const char * ggml_backend_sycl_name(ggml_backend_t backend) {
+
+    ggml_backend_sycl_context * sycl_ctx = (ggml_backend_sycl_context *)backend->context;
+
+    return sycl_ctx->name.c_str();
+}
+
+static void ggml_backend_sycl_free(ggml_backend_t backend) {
+    ggml_backend_sycl_context * sycl_ctx = (ggml_backend_sycl_context *)backend->context;
+
+    delete sycl_ctx;
+    delete backend;
+}
+
+
+static ggml_backend_buffer_type_t ggml_backend_sycl_get_default_buffer_type(ggml_backend_t backend) {
+    ggml_backend_sycl_context * sycl_ctx = (ggml_backend_sycl_context *)backend->context;
+    return ggml_backend_sycl_buffer_type(sycl_ctx->device);
+}
+
+static void ggml_backend_sycl_set_tensor_async(ggml_backend_t backend,
+                                               ggml_tensor *tensor,
+                                               const void *data, size_t offset,
+                                               size_t size) try {
+    ggml_backend_sycl_context * sycl_ctx = (ggml_backend_sycl_context *)backend->context;
+    ggml_backend_buffer_t buf = tensor->view_src ? tensor->view_src->buffer : tensor->buffer;
+
+    GGML_ASSERT(buf->buft == ggml_backend_sycl_buffer_type(sycl_ctx->device) && "unsupported buffer type");
+    const queue_ptr stream = sycl_ctx->stream(sycl_ctx->device, 0);
+    SYCL_CHECK(CHECK_TRY_ERROR((stream)->memcpy(
+        (char *)tensor->data + offset, data, size).wait()));
+}
+catch (sycl::exception const &exc) {
+  std::cerr << exc.what() << "Exception caught at file:" << __FILE__
+            << ", line:" << __LINE__ << std::endl;
+  std::exit(1);
+}
+
+static void ggml_backend_sycl_get_tensor_async(ggml_backend_t backend,
+                                               const ggml_tensor *tensor,
+                                               void *data, size_t offset,
+                                               size_t size) try {
+    ggml_backend_sycl_context * sycl_ctx = (ggml_backend_sycl_context *)backend->context;
+    ggml_backend_buffer_t buf = tensor->view_src ? tensor->view_src->buffer : tensor->buffer;
+
+    GGML_ASSERT(buf->buft == ggml_backend_sycl_buffer_type(sycl_ctx->device) && "unsupported buffer type");
+    const queue_ptr stream = sycl_ctx->stream(sycl_ctx->device, 0);
+    SYCL_CHECK(CHECK_TRY_ERROR((stream)->memcpy(
+        data, (const char *)tensor->data + offset, size).wait()));
+}
+catch (sycl::exception const &exc) {
+  std::cerr << exc.what() << "Exception caught at file:" << __FILE__
+            << ", line:" << __LINE__ << std::endl;
+  std::exit(1);
+}
+
+static bool ggml_backend_sycl_cpy_tensor_async(ggml_backend_t backend,
+                                               const ggml_tensor *src,
+                                               ggml_tensor *dst) try {
+    ggml_backend_sycl_context * sycl_ctx = (ggml_backend_sycl_context *)backend->context;
+    if (dst->buffer->buft == ggml_backend_sycl_buffer_type(sycl_ctx->device) && ggml_backend_buffer_is_sycl(src->buffer)) {
+        /*
+        DPCT1009:215: SYCL uses exceptions to report errors and does not use the
+        error codes. The original code was commented out and a warning string
+        was inserted. You need to rewrite this code.
+        */
+        const queue_ptr stream = sycl_ctx->stream(sycl_ctx->device, 0);
+        SYCL_CHECK(CHECK_TRY_ERROR((stream)->memcpy(
+            dst->data, src->data, ggml_nbytes(dst)).wait()));
+        return true;
+    }
+
+    return false;
+}
+catch (sycl::exception const &exc) {
+  std::cerr << exc.what() << "Exception caught at file:" << __FILE__
+            << ", line:" << __LINE__ << std::endl;
+  std::exit(1);
+}
+
+static void ggml_backend_sycl_synchronize(ggml_backend_t backend) try {
+    ggml_backend_sycl_context * sycl_ctx = (ggml_backend_sycl_context *)backend->context;
+    const queue_ptr stream = sycl_ctx->stream(sycl_ctx->device, 0);
+    SYCL_CHECK(CHECK_TRY_ERROR((stream)->wait()));
+
+    UNUSED(backend);
+}
+catch (sycl::exception const &exc) {
+  std::cerr << exc.what() << "Exception caught at file:" << __FILE__
+            << ", line:" << __LINE__ << std::endl;
+  std::exit(1);
+}
+
+static ggml_status ggml_backend_sycl_graph_compute(ggml_backend_t backend, ggml_cgraph * cgraph) {
+    ggml_backend_sycl_context * sycl_ctx = (ggml_backend_sycl_context *)backend->context;
+    ggml_sycl_set_main_device(sycl_ctx->device);
+
+
+    for (int i = 0; i < cgraph->n_nodes; i++) {
+        ggml_tensor * node = cgraph->nodes[i];
+        if (ggml_is_empty(node) || node->op == GGML_OP_RESHAPE || node->op == GGML_OP_TRANSPOSE || node->op == GGML_OP_VIEW || node->op == GGML_OP_PERMUTE || node->op == GGML_OP_NONE) {
+            continue;
+        }
+#ifndef NDEBUG
+        assert(node->buffer->buft == ggml_backend_sycl_buffer_type(sycl_ctx->device));
+        for (int j = 0; j < GGML_MAX_SRC; j++) {
+            if (node->src[j] != nullptr) {
+                assert(node->src[j]->buffer->buft == ggml_backend_sycl_buffer_type(sycl_ctx->device));
+            }
+        }
+#endif
+        bool ok = ggml_sycl_compute_forward(*sycl_ctx, node);
+        if (!ok) {
+            fprintf(stderr, "%s: error: op not supported %s (%s)\n", __func__, node->name, ggml_op_name(node->op));
+        }
+        GGML_ASSERT(ok);
+    }
+
+    return GGML_STATUS_SUCCESS;
+}
+
+static bool ggml_backend_sycl_supports_op(ggml_backend_t backend, const ggml_tensor * op) {
+    switch (op->op) {
+        case GGML_OP_CONV_TRANSPOSE_1D:
+            {
+                ggml_type src0_type = op->src[0]->type;
+                ggml_type src1_type = op->src[1]->type;
+                if (src0_type == GGML_TYPE_F32 && src1_type == GGML_TYPE_F32) {
+                    return true;
+                }
+                return false;
+            } break;
+        case GGML_OP_UNARY:
+            switch (ggml_get_unary_op(op)) {
+                case GGML_UNARY_OP_GELU:
+                case GGML_UNARY_OP_SILU:
+                case GGML_UNARY_OP_RELU:
+                case GGML_UNARY_OP_HARDSIGMOID:
+                case GGML_UNARY_OP_HARDSWISH:
+                case GGML_UNARY_OP_GELU_QUICK:
+                case GGML_UNARY_OP_TANH:
+                    return ggml_is_contiguous(op->src[0]);
+                default:
+                    return false;
+            }
+            break;
+        case GGML_OP_MUL_MAT:
+        case GGML_OP_MUL_MAT_ID:
+            {
+                struct ggml_tensor * a;
+                struct ggml_tensor * b;
+                if (op->op == GGML_OP_MUL_MAT) {
+                    a = op->src[0];
+                    b = op->src[1];
+                } else {
+                    a = op->src[2];
+                    b = op->src[1];
+                }
+                if (a->ne[3] != b->ne[3]) {
+                    return false;
+                }
+                ggml_type a_type = a->type;
+                if (a_type == GGML_TYPE_IQ4_NL  || a_type == GGML_TYPE_IQ4_XS ||
+                    a_type == GGML_TYPE_IQ3_XXS || a_type == GGML_TYPE_IQ3_S  ||
+                    a_type == GGML_TYPE_IQ2_XXS || a_type == GGML_TYPE_IQ2_XS || a_type == GGML_TYPE_IQ2_S ||
+                    a_type == GGML_TYPE_IQ1_S || a_type == GGML_TYPE_IQ1_M
+                    ) {
+                    if (b->ne[1] == 1 && ggml_nrows(b) > 1) {
+                        return false;
+                    }
+                }
+                ggml_type src0_type = op->src[0]->type;
+                if (src0_type == GGML_TYPE_BF16) {
+                    return false;
+                }
+                return true;
+            } break;
+        case GGML_OP_GET_ROWS:
+            {
+                switch (op->src[0]->type) {
+                    case GGML_TYPE_F16:
+                    case GGML_TYPE_F32:
+                    case GGML_TYPE_Q4_0:
+                    case GGML_TYPE_Q4_1:
+                    case GGML_TYPE_Q5_0:
+                    case GGML_TYPE_Q5_1:
+                    case GGML_TYPE_Q8_0:
+                        return true;
+                    default:
+                        return false;
+                }
+            } break;
+        case GGML_OP_CPY:
+            {
+                ggml_type src0_type = op->src[0]->type;
+                ggml_type src1_type = op->src[1]->type;
+                if (src0_type == GGML_TYPE_F32 && src1_type == GGML_TYPE_F32) {
+                    return true;
+                }
+                if (src0_type == GGML_TYPE_F32 && src1_type == GGML_TYPE_F16) {
+                    return true;
+                }
+                if (src0_type == GGML_TYPE_F32 && src1_type == GGML_TYPE_Q8_0) {
+                    return true;
+                }
+                if (src0_type == GGML_TYPE_F32 && src1_type == GGML_TYPE_Q4_0) {
+                    return true;
+                }
+                if (src0_type == GGML_TYPE_F32 && src1_type == GGML_TYPE_Q4_1) {
+                    return true;
+                }
+                if (src0_type == GGML_TYPE_F16 && src1_type == GGML_TYPE_F16) {
+                    return true;
+                }
+                if (src0_type == GGML_TYPE_F16 && src1_type == GGML_TYPE_F32) {
+                    return true;
+                }
+                return false;
+            } break;
+        case GGML_OP_CONCAT:
+            {
+                ggml_type src0_type = op->src[0]->type;
+                int dim = op->op_params[0];
+                return ggml_is_contiguous(op->src[0]) && ggml_is_contiguous(op->src[1]) && src0_type != GGML_TYPE_I32 && src0_type != GGML_TYPE_I16 && dim == 2;
+            } break;
+        case GGML_OP_DUP:
+        case GGML_OP_NONE:
+        case GGML_OP_RESHAPE:
+        case GGML_OP_REPEAT:
+        case GGML_OP_VIEW:
+        case GGML_OP_PERMUTE:
+        case GGML_OP_TRANSPOSE:
+        case GGML_OP_NORM:
+        case GGML_OP_ADD:
+        case GGML_OP_MUL:
+        case GGML_OP_DIV:
+        case GGML_OP_RMS_NORM:
+        case GGML_OP_SCALE:
+        case GGML_OP_SQR:
+        case GGML_OP_CLAMP:
+            return true;
+        case GGML_OP_CONT:
+            return op->src[0]->type != GGML_TYPE_BF16;
+        case GGML_OP_DIAG_MASK_INF:
+        case GGML_OP_SOFT_MAX:
+            return true;
+        case GGML_OP_ROPE:
+            return ggml_is_contiguous(op->src[0]);
+        case GGML_OP_IM2COL:
+            // TODO: add support for the new F32 operations
+            return op->src[0]->type == GGML_TYPE_F16;
+        case GGML_OP_POOL_2D:
+        case GGML_OP_SUM_ROWS:
+        case GGML_OP_ARGSORT:
+        case GGML_OP_ACC:
+        case GGML_OP_GROUP_NORM:
+        case GGML_OP_UPSCALE:
+        case GGML_OP_PAD:
+        case GGML_OP_LEAKY_RELU:
+        case GGML_OP_TIMESTEP_EMBEDDING:
+            return true;
+        default:
+            return false;
+    }
+
+    UNUSED(backend);
+}
+
+static bool ggml_backend_sycl_offload_op(ggml_backend_t backend, const ggml_tensor * op) {
+    const int min_batch_size = 32;
+    return op->ne[1] >= min_batch_size && op->op != GGML_OP_GET_ROWS && op->op != GGML_OP_MUL_MAT_ID;
+    GGML_UNUSED(backend);
+}
+
+static bool ggml_backend_sycl_supports_buft(ggml_backend_t backend, ggml_backend_buffer_type_t buft) {
+    if (buft->iface.get_name != ggml_backend_sycl_buffer_type_name) {
+        return false;
+    }
+    ggml_backend_sycl_buffer_type_context * buft_ctx = (ggml_backend_sycl_buffer_type_context *)buft->context;
+    ggml_backend_sycl_context * sycl_ctx = (ggml_backend_sycl_context *)backend->context;
+    return buft_ctx->device == sycl_ctx->device;
+}
+
+static ggml_backend_i ggml_backend_sycl_interface = {
+    /* .get_name                = */ ggml_backend_sycl_name,
+    /* .free                    = */ ggml_backend_sycl_free,
+    /* .get_default_buffer_type = */ ggml_backend_sycl_get_default_buffer_type,
+    /* .set_tensor_async        = */ ggml_backend_sycl_set_tensor_async,
+    /* .get_tensor_async        = */ ggml_backend_sycl_get_tensor_async,
+    /* .cpy_tensor_async        = */ NULL, //ggml_backend_sycl_cpy_tensor_async, // TODO: update for the new interface
+    /* .synchronize             = */ ggml_backend_sycl_synchronize,
+    /* .graph_plan_create       = */ NULL,
+    /* .graph_plan_free         = */ NULL,
+    /* .graph_plan_update       = */ NULL,
+    /* .graph_plan_compute      = */ NULL,
+    /* .graph_compute           = */ ggml_backend_sycl_graph_compute,
+    /* .supports_op             = */ ggml_backend_sycl_supports_op,
+    /* .supports_buft           = */ ggml_backend_sycl_supports_buft,
+    /* .offload_op              = */ ggml_backend_sycl_offload_op,
+    /* .event_record            = */ NULL,
+    /* .event_wait              = */ NULL,
+};
+
+static ggml_guid_t ggml_backend_sycl_guid() {
+    static ggml_guid guid = { 0x58, 0x05, 0x13, 0x8f, 0xcd, 0x3a, 0x61, 0x9d, 0xe7, 0xcd, 0x98, 0xa9, 0x03, 0xfd, 0x7c, 0x53 };
+    return &guid;
+}
+
+ggml_backend_t ggml_backend_sycl_init(int device) {
+    GGML_SYCL_DEBUG("[SYCL] call ggml_backend_sycl_init\n");
+    ggml_check_sycl();
+
+    check_allow_gpu_index(device);
+
+    ggml_backend_sycl_context * ctx = new ggml_backend_sycl_context(device);
+    if (ctx == nullptr) {
+        fprintf(stderr, "%s: error: failed to allocate context\n", __func__);
+        return nullptr;
+    };
+
+    ggml_backend_t sycl_backend = new ggml_backend {
+        /* .guid      = */ ggml_backend_sycl_guid(),
+        /* .interface = */ ggml_backend_sycl_interface,
+        /* .device    = */ nullptr,
+        /* .context   = */ ctx
+    };
+
+    return sycl_backend;
+}
+
+bool ggml_backend_is_sycl(ggml_backend_t backend) {
+    return backend != NULL && ggml_guid_matches(backend->guid, ggml_backend_sycl_guid());
+}
+
+int ggml_backend_sycl_get_device_count() {
+    GGML_SYCL_DEBUG("[SYCL] call ggml_backend_sycl_get_device_count\n");
+    return ggml_sycl_info().device_count;
+}
diff --git a/src/whisper_cpp/ggml/src/ggml-sycl/backend.hpp b/src/whisper_cpp/ggml/src/ggml-sycl/backend.hpp
new file mode 100644
index 00000000..d21b5f8d
--- /dev/null
+++ b/src/whisper_cpp/ggml/src/ggml-sycl/backend.hpp
@@ -0,0 +1,30 @@
+//
+// MIT license
+// Copyright (C) 2024 Intel Corporation
+// SPDX-License-Identifier: MIT
+//
+
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+//
+
+#ifndef GGML_SYCL_BACKEND_HPP
+#define GGML_SYCL_BACKEND_HPP
+
+#include "concat.hpp"
+#include "common.hpp"
+#include "conv.hpp"
+#include "convert.hpp"
+#include "dequantize.hpp"
+#include "dmmv.hpp"
+#include "mmq.hpp"
+#include "mmvq.hpp"
+#include "rope.hpp"
+#include "norm.hpp"
+#include "softmax.hpp"
+#include "tsembd.hpp"
+#include "im2col.hpp"
+
+#endif // GGML_SYCL_BACKEND_HPP
diff --git a/src/whisper_cpp/ggml/src/ggml-sycl/common.cpp b/src/whisper_cpp/ggml/src/ggml-sycl/common.cpp
new file mode 100644
index 00000000..cf5291b3
--- /dev/null
+++ b/src/whisper_cpp/ggml/src/ggml-sycl/common.cpp
@@ -0,0 +1,64 @@
+//
+// MIT license
+// Copyright (C) 2024 Intel Corporation
+// SPDX-License-Identifier: MIT
+//
+
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+//
+
+#include "common.hpp"
+
+int get_current_device_id() {
+  return dpct::dev_mgr::instance().current_device_id();
+}
+
+void* ggml_sycl_host_malloc(size_t size) try {
+  if (getenv("GGML_SYCL_NO_PINNED") != nullptr) {
+    return nullptr;
+  }
+
+  void* ptr = nullptr;
+  // allow to use dpct::get_in_order_queue() for host malloc
+  dpct::err0 err = CHECK_TRY_ERROR(
+      ptr = (void*)sycl::malloc_host(size, dpct::get_in_order_queue()));
+
+  if (err != 0) {
+    // clear the error
+    fprintf(
+        stderr,
+        "WARNING: failed to allocate %.2f MB of pinned memory: %s\n",
+        size / 1024.0 / 1024.0,
+        "syclGetErrorString is not supported");
+    return nullptr;
+  }
+
+  return ptr;
+} catch (sycl::exception const& exc) {
+  std::cerr << exc.what() << "Exception caught at file:" << __FILE__
+            << ", line:" << __LINE__ << std::endl;
+  std::exit(1);
+}
+
+void ggml_sycl_host_free(void* ptr) try {
+  // allow to use dpct::get_in_order_queue() for host malloc
+  SYCL_CHECK(CHECK_TRY_ERROR(sycl::free(ptr, dpct::get_in_order_queue())));
+} catch (sycl::exception const& exc) {
+  std::cerr << exc.what() << "Exception caught at file:" << __FILE__
+            << ", line:" << __LINE__ << std::endl;
+  std::exit(1);
+}
+
+int64_t downsample_sycl_global_range(int64_t accumulate_block_num, int64_t block_size) {
+  const int64_t max_range = std::numeric_limits<int>::max();
+  int64_t sycl_down_blk_size = block_size;
+  int64_t global_range = accumulate_block_num * sycl_down_blk_size;
+  while(global_range > max_range) {
+      sycl_down_blk_size /= 2;
+      global_range = accumulate_block_num * sycl_down_blk_size;
+  }
+  return sycl_down_blk_size;
+}
diff --git a/src/whisper_cpp/ggml/src/ggml-sycl/common.hpp b/src/whisper_cpp/ggml/src/ggml-sycl/common.hpp
new file mode 100644
index 00000000..bc0faa86
--- /dev/null
+++ b/src/whisper_cpp/ggml/src/ggml-sycl/common.hpp
@@ -0,0 +1,407 @@
+//
+// MIT license
+// Copyright (C) 2024 Intel Corporation
+// SPDX-License-Identifier: MIT
+//
+
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+//
+
+#ifndef GGML_SYCL_COMMON_HPP
+#define GGML_SYCL_COMMON_HPP
+
+#include <fstream>
+#include <iostream>
+
+#include "dpct/helper.hpp"
+#include "ggml-sycl.h"
+#include "presets.hpp"
+#if GGML_SYCL_DNNL
+#include "dnnl.hpp"
+#include "dnnl_sycl.hpp"
+#endif
+
+#define GGML_COMMON_DECL_SYCL
+#define GGML_COMMON_IMPL_SYCL
+#include "ggml-common.h"
+
+void* ggml_sycl_host_malloc(size_t size);
+void ggml_sycl_host_free(void* ptr);
+
+static int g_ggml_sycl_debug = 0;
+#define GGML_SYCL_DEBUG(...)        \
+  do {                              \
+    if (g_ggml_sycl_debug)          \
+      fprintf(stderr, __VA_ARGS__); \
+  } while (0)
+
+#define CHECK_TRY_ERROR(expr)                                            \
+  [&]() {                                                                \
+    try {                                                                \
+      expr;                                                              \
+      return dpct::success;                                              \
+    } catch (std::exception const& e) {                                  \
+      std::cerr << e.what() << "\nException caught at file:" << __FILE__ \
+                << ", line:" << __LINE__ << ", func:" << __func__        \
+                << std::endl;                                            \
+      return dpct::default_error;                                        \
+    }                                                                    \
+  }()
+
+
+#define __SYCL_ARCH__ DPCT_COMPATIBILITY_TEMP
+#define VER_4VEC 610 // todo for hardward optimize.
+#define VER_GEN9 700 // todo for hardward optimize.
+#define VER_GEN12 1000000 // todo for hardward optimize.
+#define VER_GEN13 (VER_GEN12 + 1030) // todo for hardward optimize.
+
+#define GGML_SYCL_MAX_NODES 8192 // TODO: adapt to hardwares
+
+// define for XMX in Intel GPU
+// TODO: currently, it's not used for XMX really.
+#if !defined(GGML_SYCL_FORCE_MMQ)
+    #define SYCL_USE_XMX
+#endif
+
+// max batch size to use MMQ kernels when tensor cores are available
+#define MMQ_MAX_BATCH_SIZE 32
+
+#if defined(_MSC_VER)
+#pragma warning(disable : 4244 4267) // possible loss of data
+#endif
+
+// dmmv = dequantize_mul_mat_vec
+#ifndef GGML_SYCL_DMMV_X
+#define GGML_SYCL_DMMV_X 32
+#endif
+#ifndef GGML_SYCL_MMV_Y
+#define GGML_SYCL_MMV_Y 1
+#endif
+
+typedef sycl::queue *queue_ptr;
+
+enum ggml_sycl_backend_gpu_mode {
+  SYCL_UNSET_GPU_MODE = -1,
+  SYCL_SINGLE_GPU_MODE = 0,
+  SYCL_MUL_GPU_MODE
+};
+
+static_assert(sizeof(sycl::half) == sizeof(ggml_fp16_t), "wrong fp16 size");
+
+static void crash() {
+  int* ptr = NULL;
+  *ptr = 0;
+}
+
+[[noreturn]] static void ggml_sycl_error(
+    const char* stmt,
+    const char* func,
+    const char* file,
+    const int line,
+    const char* msg) {
+  fprintf(stderr, "SYCL error: %s: %s\n", stmt, msg);
+  fprintf(stderr, "  in function %s at %s:%d\n", func, file, line);
+  GGML_ABORT("SYCL error");
+}
+
+#define SYCL_CHECK(err)                     \
+  do {                                      \
+    auto err_ = (err);                      \
+    if (err_ != 0)                          \
+      ggml_sycl_error(                      \
+          #err,                             \
+          __func__,                         \
+          __FILE__,                         \
+          __LINE__,                         \
+          "Meet error in this line code!"); \
+  } while (0)
+
+#if DPCT_COMPAT_RT_VERSION >= 11100
+#define GGML_SYCL_ASSUME(x) __builtin_assume(x)
+#else
+#define GGML_SYCL_ASSUME(x)
+#endif // DPCT_COMPAT_RT_VERSION >= 11100
+
+#ifdef GGML_SYCL_F16
+typedef sycl::half dfloat; // dequantize float
+typedef sycl::half2 dfloat2;
+#else
+typedef float dfloat; // dequantize float
+typedef sycl::float2 dfloat2;
+#endif // GGML_SYCL_F16
+
+#define MMVQ_MAX_BATCH_SIZE  8
+
+static const int8_t kvalues_iq4nl[16]={-127, -104, -83, -65, -49, -35, -22, -10, 1, 13, 25, 38, 53, 69, 89, 113};
+
+static int g_all_sycl_device_count = -1;
+static bool g_ggml_backend_sycl_buffer_type_initialized = false;
+
+static ggml_sycl_backend_gpu_mode g_ggml_sycl_backend_gpu_mode =
+    SYCL_UNSET_GPU_MODE;
+
+static void* g_scratch_buffer = nullptr;
+static size_t g_scratch_size = 0; // disabled by default
+static size_t g_scratch_offset = 0;
+
+[[noreturn]] static inline void bad_arch(const sycl::stream& stream_ct1) {
+  stream_ct1 << "ERROR: ggml-sycl was compiled without support for the "
+                "current GPU architecture.\n";
+  // __trap();
+  std::exit(1);
+
+  (void)bad_arch; // suppress unused function warning
+}
+
+int get_current_device_id();
+
+inline dpct::err0 ggml_sycl_set_device(const int device) try {
+
+  int current_device_id;
+  SYCL_CHECK(CHECK_TRY_ERROR(current_device_id = get_current_device_id()));
+
+  // GGML_SYCL_DEBUG("ggml_sycl_set_device device_id=%d,
+  // current_device_id=%d\n", device, current_device);
+  if (device == current_device_id) {
+    return 0;
+  }
+
+  return CHECK_TRY_ERROR(dpct::select_device(device));
+} catch (sycl::exception const& exc) {
+  std::cerr << exc.what() << "Exception caught at file:" << __FILE__
+            << ", line:" << __LINE__ << std::endl;
+  crash();
+  std::exit(1);
+}
+
+//////////////////////
+
+struct ggml_sycl_device_info {
+    int device_count;
+
+    struct sycl_device_info {
+        int     cc;                 // compute capability
+        // int     nsm;                // number of streaming multiprocessors
+        // size_t  smpb;               // max. shared memory per block
+        bool    vmm;                // virtual memory support
+        size_t  total_vram;
+    };
+
+    sycl_device_info devices[GGML_SYCL_MAX_DEVICES] = {};
+
+    std::array<float, GGML_SYCL_MAX_DEVICES> default_tensor_split = {};
+
+    int max_work_group_sizes[GGML_SYCL_MAX_DEVICES] = {0};
+};
+
+const ggml_sycl_device_info & ggml_sycl_info();
+
+struct ggml_sycl_pool {
+    virtual ~ggml_sycl_pool() = default;
+
+    virtual void * alloc(size_t size, size_t * actual_size) = 0;
+    virtual void free(void * ptr, size_t size) = 0;
+};
+
+template<typename T>
+struct ggml_sycl_pool_alloc {
+    ggml_sycl_pool * pool = nullptr;
+    T * ptr = nullptr;
+    size_t actual_size = 0;
+
+    explicit ggml_sycl_pool_alloc(ggml_sycl_pool & pool) : pool(&pool) {
+    }
+
+    ggml_sycl_pool_alloc(ggml_sycl_pool & pool, size_t size) : pool(&pool) {
+        alloc(size);
+    }
+
+    ~ggml_sycl_pool_alloc() {
+        if (ptr != nullptr) {
+            pool->free(ptr, actual_size);
+        }
+    }
+
+    // size is in number of elements
+    T * alloc(size_t size) {
+        GGML_ASSERT(pool != nullptr);
+        GGML_ASSERT(ptr == nullptr);
+        ptr = (T *) pool->alloc(size * sizeof(T), &this->actual_size);
+        return ptr;
+    }
+
+    T * alloc(ggml_sycl_pool & pool, size_t size) {
+        this->pool = &pool;
+        return alloc(size);
+    }
+
+    T * get() {
+        return ptr;
+    }
+
+    ggml_sycl_pool_alloc() = default;
+    ggml_sycl_pool_alloc(const ggml_sycl_pool_alloc &) = delete;
+    ggml_sycl_pool_alloc(ggml_sycl_pool_alloc &&) = delete;
+    ggml_sycl_pool_alloc& operator=(const ggml_sycl_pool_alloc &) = delete;
+    ggml_sycl_pool_alloc& operator=(ggml_sycl_pool_alloc &&) = delete;
+};
+
+// backend interface
+
+struct ggml_tensor_extra_gpu {
+  void* data_device[GGML_SYCL_MAX_DEVICES]; // 1 pointer for each device for split
+                                       // tensors
+  dpct::event_ptr events[GGML_SYCL_MAX_DEVICES]
+                        [GGML_SYCL_MAX_STREAMS]; // events for synchronizing multiple GPUs
+};
+
+struct ggml_backend_sycl_context {
+    int device;
+    std::string name;
+
+    queue_ptr qptrs[GGML_SYCL_MAX_DEVICES][GGML_SYCL_MAX_STREAMS] = { { nullptr } };
+
+    explicit ggml_backend_sycl_context(int device) :
+        device(device),
+        name(GGML_SYCL_NAME + std::to_string(device)) {
+    }
+
+    queue_ptr stream(int device, int stream) {
+        if (qptrs[device][stream] == nullptr) {
+            qptrs[device][stream] = &(dpct::get_device(device).default_queue());
+        }
+        return qptrs[device][stream];
+    }
+
+    queue_ptr stream() {
+        return stream(device, 0);
+    }
+
+#if GGML_SYCL_DNNL
+    dnnl::engine make_engine(sycl::queue* q) {
+        // Get the device associated with the queue
+        sycl::device dev = q->get_device();
+        // Get the context associated with the queue
+        sycl::context ctx = q->get_context();
+        const dnnl::engine eng = dnnl::sycl_interop::make_engine(dev, ctx);
+        return eng;
+    }
+
+    std::unordered_map<sycl::queue*, dnnl::stream> stream_map;
+    std::unordered_map<sycl::queue*, dnnl::engine> engine_map;
+    dnnl::stream stream_dnnl(int device, int _stream) {
+        auto q = stream(device, _stream);
+        return stream_dnnl(q);
+    }
+    dnnl::engine engine_dnnl(sycl::queue* qptr) {
+        auto it = engine_map.find(qptr);
+        if (it == engine_map.end()) {
+            auto eng = make_engine(qptr);
+            engine_map[qptr] = eng;
+            return eng;
+        }
+        else
+        {
+            return it->second;
+        }
+    }
+    dnnl::stream stream_dnnl(sycl::queue* qptr) {
+        auto it = stream_map.find(qptr);
+        if (it == stream_map.end()) {
+            auto eng = engine_dnnl(qptr);
+            auto stream = dnnl::sycl_interop::make_stream(eng, *qptr);
+            stream_map[qptr] = stream;
+            return stream;
+        }
+        else
+        {
+            return it->second;
+        }
+    }
+    dnnl::stream stream_dnnl() {
+        return stream_dnnl(device, 0);
+    }
+#endif
+
+    // pool
+    std::unique_ptr<ggml_sycl_pool> pools[GGML_SYCL_MAX_DEVICES];
+
+    static std::unique_ptr<ggml_sycl_pool> new_pool_for_device(queue_ptr qptr, int device);
+
+    ggml_sycl_pool & pool(int device) {
+        if (pools[device] == nullptr) {
+            pools[device] = new_pool_for_device(stream(device,0), device);
+        }
+        return *pools[device];
+    }
+
+    ggml_sycl_pool & pool() {
+        return pool(device);
+    }
+};
+
+// common device functions
+
+static __dpct_inline__ float warp_reduce_sum(float x,
+    const sycl::nd_item<3>& item_ct1) {
+#pragma unroll
+    for (int mask = WARP_SIZE / 2; mask > 0; mask >>= 1) {
+        /*
+        DPCT1096:98: The right-most dimension of the work-group used in the SYCL
+        kernel that calls this function may be less than "32". The function
+        "dpct::permute_sub_group_by_xor" may return an unexpected result on the
+        CPU device. Modify the size of the work-group to ensure that the value
+        of the right-most dimension is a multiple of "32".
+        */
+        x += dpct::permute_sub_group_by_xor(item_ct1.get_sub_group(), x, mask);
+    }
+    return x;
+}
+
+static __dpct_inline__ sycl::float2
+warp_reduce_sum(sycl::float2 a, const sycl::nd_item<3>& item_ct1) {
+#pragma unroll
+    for (int mask = WARP_SIZE / 2; mask > 0; mask >>= 1) {
+        a.x() += dpct::permute_sub_group_by_xor(item_ct1.get_sub_group(), a.x(),
+            mask);
+        a.y() += dpct::permute_sub_group_by_xor(item_ct1.get_sub_group(), a.y(),
+            mask);
+    }
+    return a;
+}
+
+static __dpct_inline__ float warp_reduce_max(float x,
+    const sycl::nd_item<3>& item_ct1) {
+#pragma unroll
+    for (int mask = WARP_SIZE / 2; mask > 0; mask >>= 1) {
+        /*
+        DPCT1096:97: The right-most dimension of the work-group used in the SYCL
+        kernel that calls this function may be less than "32". The function
+        "dpct::permute_sub_group_by_xor" may return an unexpected result on the
+        CPU device. Modify the size of the work-group to ensure that the value
+        of the right-most dimension is a multiple of "32".
+        */
+        x = sycl::fmax(x, dpct::permute_sub_group_by_xor(
+            item_ct1.get_sub_group(), x, mask));
+    }
+    return x;
+}
+
+// Helper for vec loading aligned data
+template <typename Tp, int n>
+inline sycl::vec<Tp, n> vec_aligned_load(const Tp* aligned_ptr) {
+    return *reinterpret_cast<const sycl::vec<Tp, n>*>(aligned_ptr);
+}
+
+// Helper for accessing pointers with no warnings
+template <typename Tp, int dim>
+static __dpct_inline__ Tp* get_pointer(sycl::local_accessor<Tp, dim> acc) {
+    return acc.template get_multi_ptr<sycl::access::decorated::no>().get();
+}
+
+int64_t downsample_sycl_global_range(int64_t accumulate_block_num, int64_t block_size);
+
+#endif // GGML_SYCL_COMMON_HPP
diff --git a/src/whisper_cpp/ggml/src/ggml-sycl/concat.cpp b/src/whisper_cpp/ggml/src/ggml-sycl/concat.cpp
new file mode 100644
index 00000000..632eedb9
--- /dev/null
+++ b/src/whisper_cpp/ggml/src/ggml-sycl/concat.cpp
@@ -0,0 +1,195 @@
+//
+// MIT license
+// Copyright (C) 2024 Intel Corporation
+// SPDX-License-Identifier: MIT
+//
+
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+//
+
+#include "concat.hpp"
+#include "common.hpp"
+
+static void concat_f32_dim0(const float *x, const float *y, float *dst,
+                            const int ne0, const int ne00,
+                            const sycl::nd_item<3> &item_ct1) {
+  int nidx = item_ct1.get_local_id(2) +
+             item_ct1.get_group(2) * item_ct1.get_local_range(2);
+  if (nidx >= ne0) {
+    return;
+  }
+  // operation
+  int offset_dst = nidx + item_ct1.get_group(1) * ne0 +
+                   item_ct1.get_group(0) * ne0 * item_ct1.get_group_range(1);
+  if (nidx < ne00) { // src0
+    int offset_src = nidx + item_ct1.get_group(1) * ne00 +
+                     item_ct1.get_group(0) * ne00 * item_ct1.get_group_range(1);
+    dst[offset_dst] = x[offset_src];
+  } else {
+    int offset_src =
+        nidx - ne00 + item_ct1.get_group(1) * (ne0 - ne00) +
+        item_ct1.get_group(0) * (ne0 - ne00) * item_ct1.get_group_range(1);
+    dst[offset_dst] = y[offset_src];
+  }
+}
+
+static void concat_f32_dim1(const float *x, const float *y, float *dst,
+                            const int ne0, const int ne01,
+                            const sycl::nd_item<3> &item_ct1) {
+  int nidx = item_ct1.get_local_id(2) +
+             item_ct1.get_group(2) * item_ct1.get_local_range(2);
+  if (nidx >= ne0) {
+    return;
+  }
+  // operation
+  int offset_dst = nidx + item_ct1.get_group(1) * ne0 +
+                   item_ct1.get_group(0) * ne0 * item_ct1.get_group_range(1);
+  if (item_ct1.get_group(1) < ne01) { // src0
+    int offset_src =
+        nidx + item_ct1.get_group(1) * ne0 + item_ct1.get_group(0) * ne0 * ne01;
+    dst[offset_dst] = x[offset_src];
+  } else {
+    int offset_src =
+        nidx + (item_ct1.get_group(1) - ne01) * ne0 +
+        item_ct1.get_group(0) * ne0 * (item_ct1.get_group_range(1) - ne01);
+    dst[offset_dst] = y[offset_src];
+  }
+}
+
+static void concat_f32_dim2(const float *x, const float *y, float *dst,
+                            const int ne0, const int ne02,
+                            const sycl::nd_item<3> &item_ct1) {
+  int nidx = item_ct1.get_local_id(2) +
+             item_ct1.get_group(2) * item_ct1.get_local_range(2);
+  if (nidx >= ne0) {
+    return;
+  }
+  // operation
+  int offset_dst = nidx + item_ct1.get_group(1) * ne0 +
+                   item_ct1.get_group(0) * ne0 * item_ct1.get_group_range(1);
+  if (item_ct1.get_group(0) < ne02) { // src0
+    int offset_src = nidx + item_ct1.get_group(1) * ne0 +
+                     item_ct1.get_group(0) * ne0 * item_ct1.get_group_range(1);
+    dst[offset_dst] = x[offset_src];
+  } else {
+    int offset_src =
+        nidx + item_ct1.get_group(1) * ne0 +
+        (item_ct1.get_group(0) - ne02) * ne0 * item_ct1.get_group_range(1);
+    dst[offset_dst] = y[offset_src];
+  }
+}
+
+static void concat_f32_sycl(const float *x, const float *y, float *dst,
+                            int ne00, int ne01, int ne02, int ne0, int ne1,
+                            int ne2, int dim, queue_ptr stream) {
+  int num_blocks = (ne0 + SYCL_CONCAT_BLOCK_SIZE - 1) / SYCL_CONCAT_BLOCK_SIZE;
+  sycl::range<3> gridDim(ne2, ne1, num_blocks);
+  switch (dim) {
+  case 0:
+    stream->parallel_for(
+        sycl::nd_range<3>(gridDim *
+                              sycl::range<3>(1, 1, SYCL_CONCAT_BLOCK_SIZE),
+                          sycl::range<3>(1, 1, SYCL_CONCAT_BLOCK_SIZE)),
+        [=](sycl::nd_item<3> item_ct1) {
+          concat_f32_dim0(x, y, dst, ne0, ne00, item_ct1);
+        });
+    break;
+  case 1:
+    stream->parallel_for(
+        sycl::nd_range<3>(gridDim *
+                              sycl::range<3>(1, 1, SYCL_CONCAT_BLOCK_SIZE),
+                          sycl::range<3>(1, 1, SYCL_CONCAT_BLOCK_SIZE)),
+        [=](sycl::nd_item<3> item_ct1) {
+          concat_f32_dim1(x, y, dst, ne0, ne01, item_ct1);
+        });
+    break;
+  default:
+    stream->parallel_for(
+        sycl::nd_range<3>(gridDim *
+                              sycl::range<3>(1, 1, SYCL_CONCAT_BLOCK_SIZE),
+                          sycl::range<3>(1, 1, SYCL_CONCAT_BLOCK_SIZE)),
+        [=](sycl::nd_item<3> item_ct1) {
+          concat_f32_dim2(x, y, dst, ne0, ne02, item_ct1);
+        });
+    break;
+  }
+}
+
+// non-contiguous kernel (slow)
+static void concat_f32_sycl_non_cont(
+    queue_ptr stream, const char *src0, const char *src1, char *dst,
+    int64_t ne00, int64_t ne01, int64_t ne02, int64_t ne03, uint64_t nb00,
+    uint64_t nb01, uint64_t nb02, uint64_t nb03, int64_t /*ne10*/,
+    int64_t /*ne11*/, int64_t /*ne12*/, int64_t /*ne13*/, uint64_t nb10,
+    uint64_t nb11, uint64_t nb12, uint64_t nb13, int64_t ne0, int64_t ne1,
+    int64_t ne2, int64_t ne3, uint64_t nb0, uint64_t nb1, uint64_t nb2,
+    uint64_t nb3, int32_t dim) {
+  sycl::range<3> gridDim(ne3, ne2, ne1);
+  stream->parallel_for(
+      sycl::nd_range<3>(gridDim, sycl::range<3>(1, 1, 1)),
+      [=](sycl::nd_item<3> item_ct1) {
+        int64_t i3 = item_ct1.get_group(0);
+        int64_t i2 = item_ct1.get_group(1);
+        int64_t i1 = item_ct1.get_group(2);
+
+        int64_t o[4] = {0, 0, 0, 0};
+        o[dim] = dim == 0 ? ne00 : (dim == 1 ? ne01 : (dim == 2 ? ne02 : ne03));
+
+        const float *x;
+
+        for (int i0 = item_ct1.get_local_id(2); i0 < ne0;
+             i0 += item_ct1.get_local_range(2)) {
+          if (i0 < ne00 && i1 < ne01 && i2 < ne02 && i3 < ne03) {
+            x = (const float *)(src0 + (i3)*nb03 + (i2)*nb02 + (i1)*nb01 +
+                                (i0)*nb00);
+          } else {
+            x = (const float *)(src1 + (i3 - o[3]) * nb13 + (i2 - o[2]) * nb12 +
+                                (i1 - o[1]) * nb11 + (i0 - o[0]) * nb10);
+          }
+
+          float *y = (float *)(dst + i3 * nb3 + i2 * nb2 + i1 * nb1 + i0 * nb0);
+
+          *y = *x;
+        }
+      });
+}
+
+void ggml_sycl_op_concat(ggml_backend_sycl_context & ctx, const ggml_tensor *src0,
+                                const ggml_tensor *src1, ggml_tensor *dst) {
+  queue_ptr stream = ctx.stream();
+
+  const int32_t dim = ((int32_t *)dst->op_params)[0];
+
+  if (ggml_is_contiguous(src0) && ggml_is_contiguous(src1)) {
+    const float *src0_d = (const float *)src0->data;
+    const float *src1_d = (const float *)src1->data;
+
+    float *dst_d = (float *)dst->data;
+
+    if (dim != 3) {
+      for (int i3 = 0; i3 < dst->ne[3]; i3++) {
+        concat_f32_sycl(
+            src0_d + i3 * (src0->nb[3] / 4), src1_d + i3 * (src1->nb[3] / 4),
+            dst_d + i3 * (dst->nb[3] / 4), src0->ne[0], src0->ne[1],
+            src0->ne[2], dst->ne[0], dst->ne[1], dst->ne[2], dim, stream);
+      }
+    } else {
+      const size_t size0 = ggml_nbytes(src0);
+      const size_t size1 = ggml_nbytes(src1);
+
+      SYCL_CHECK(CHECK_TRY_ERROR(stream->memcpy(dst_d, src0_d, size0).wait()));
+      SYCL_CHECK(CHECK_TRY_ERROR(
+          stream->memcpy(dst_d + size0 / 4, src1_d, size1).wait()));
+    }
+  } else
+    concat_f32_sycl_non_cont(
+        stream, (const char *)src0->data, (const char *)src1->data,
+        (char *)dst->data, src0->ne[0], src0->ne[1], src0->ne[2], src0->ne[3],
+        src0->nb[0], src0->nb[1], src0->nb[2], src0->nb[3], src1->ne[0],
+        src1->ne[1], src1->ne[2], src1->ne[3], src1->nb[0], src1->nb[1],
+        src1->nb[2], src1->nb[3], dst->ne[0], dst->ne[1], dst->ne[2],
+        dst->ne[3], dst->nb[0], dst->nb[1], dst->nb[2], dst->nb[3], dim);
+}
diff --git a/src/whisper_cpp/ggml/src/ggml-sycl/concat.hpp b/src/whisper_cpp/ggml/src/ggml-sycl/concat.hpp
new file mode 100644
index 00000000..5a04feaa
--- /dev/null
+++ b/src/whisper_cpp/ggml/src/ggml-sycl/concat.hpp
@@ -0,0 +1,21 @@
+//
+// MIT license
+// Copyright (C) 2024 Intel Corporation
+// SPDX-License-Identifier: MIT
+//
+
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+//
+
+#ifndef GGML_SYCL_CONCAT_HPP
+#define GGML_SYCL_CONCAT_HPP
+
+#include "common.hpp"
+
+void ggml_sycl_op_concat(ggml_backend_sycl_context & ctx, const ggml_tensor *src0,
+                                const ggml_tensor *src1, ggml_tensor *dst);
+
+#endif // GGML_SYCL_CONCAT_HPP
diff --git a/src/whisper_cpp/ggml/src/ggml-sycl/conv.cpp b/src/whisper_cpp/ggml/src/ggml-sycl/conv.cpp
new file mode 100644
index 00000000..bc4ab1dd
--- /dev/null
+++ b/src/whisper_cpp/ggml/src/ggml-sycl/conv.cpp
@@ -0,0 +1,99 @@
+//
+// MIT license
+// Copyright (C) 2024 Intel Corporation
+// SPDX-License-Identifier: MIT
+//
+
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+//
+
+#include "conv.hpp"
+
+static  void conv_transpose_1d_kernel(
+        const int s0, const int output_size,
+        const int src0_ne0, const int src0_ne1, const int src0_ne2,
+        const int src1_ne0, const int dst_ne0,
+        const float * src0, const float * src1,  float * dst,
+        const sycl::nd_item<3> &item_ct1) {
+    int global_index = item_ct1.get_local_id(2) +
+                       item_ct1.get_group(2) * item_ct1.get_local_range(2);
+    if (global_index >= output_size) {
+        return;
+    }
+
+    int out_index = global_index / dst_ne0;
+
+    float accumulator = 0;
+
+    for (int c = 0; c < src0_ne2; c++) {
+        int idx = global_index % dst_ne0;
+
+        int kernel_offset = (src0_ne0 * src0_ne1 * c) + (out_index * src0_ne0);
+        int input_offset = src1_ne0 * c;
+
+        for (int i = 0; i < src1_ne0; i++) {
+            if (!(idx >= i*s0 && idx < i*s0 + src0_ne0)) {
+                continue;
+            }
+            int weight_idx = idx - i*s0;
+
+            float kernel_weight = src0[kernel_offset + weight_idx];
+            float input_value =  src1[input_offset+i];
+
+            accumulator += kernel_weight * input_value;
+        }
+    }
+    dst[global_index] = accumulator;
+}
+
+static void conv_transpose_1d_f32_f32_sycl(
+    const int s0, const int output_size,
+    const int src0_ne0, const int src0_ne1, const int src0_ne2,
+    const int src1_ne0, const int dst_ne0,
+    const float *src0, const float *src1, float *dst,
+    const queue_ptr& stream) {
+
+    const int num_blocks = (output_size + SYCL_CONV_TRANPOSE_1D_BLOCK_SIZE - 1) / SYCL_CONV_TRANPOSE_1D_BLOCK_SIZE;
+    const sycl::range<3> block_dims(1, 1, SYCL_CONV_TRANPOSE_1D_BLOCK_SIZE);
+    const sycl::range<3> block_nums(1, 1, num_blocks);
+    stream->parallel_for(
+        sycl::nd_range<3>(
+            block_nums * block_dims, block_dims),
+        [=](sycl::nd_item<3> item_ct1) {
+            conv_transpose_1d_kernel(
+                s0, output_size,
+                src0_ne0, src0_ne1, src0_ne2,
+                src1_ne0, dst_ne0,
+                src0, src1, dst, item_ct1);
+        });
+}
+
+void ggml_sycl_op_conv_transpose_1d(ggml_backend_sycl_context & ctx, const ggml_tensor *src0,
+    const ggml_tensor *src1, ggml_tensor *dst) {
+    const float * src0_d = (const float *)src0->data;
+    const float * src1_d = (const float *)src1->data;
+
+    float * dst_d = (float *)dst->data;
+    dpct::queue_ptr stream = ctx.stream();
+
+    GGML_ASSERT(src0->type == GGML_TYPE_F32);
+    GGML_ASSERT( dst->type == GGML_TYPE_F32);
+
+    GGML_ASSERT(ggml_is_contiguous(src0));
+    GGML_ASSERT(ggml_is_contiguous(src1));
+
+    const int32_t * opts = (const int32_t *)dst->op_params;
+
+    const int s0 = opts[0];
+
+    const int64_t output_size = ggml_nelements(dst);
+
+    conv_transpose_1d_f32_f32_sycl(s0, output_size,
+        src0->ne[0], src0->ne[1], src0->ne[2],
+        src1->ne[0], dst->ne[0],
+        src0_d, src1_d, dst_d, stream);
+}
+
diff --git a/src/whisper_cpp/ggml/src/ggml-sycl/conv.hpp b/src/whisper_cpp/ggml/src/ggml-sycl/conv.hpp
new file mode 100644
index 00000000..eb20730f
--- /dev/null
+++ b/src/whisper_cpp/ggml/src/ggml-sycl/conv.hpp
@@ -0,0 +1,21 @@
+//
+// MIT license
+// Copyright (C) 2024 Intel Corporation
+// SPDX-License-Identifier: MIT
+//
+
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+//
+
+#ifndef GGML_SYCL_CONV_HPP
+#define GGML_SYCL_CONV_HPP
+
+#include "common.hpp"
+
+void ggml_sycl_op_conv_transpose_1d(ggml_backend_sycl_context & ctx, const ggml_tensor *src0,
+  const ggml_tensor *src1, ggml_tensor *dst);
+
+#endif // GGML_SYCL_CONV_HPP
diff --git a/src/whisper_cpp/ggml/src/ggml-sycl/convert.cpp b/src/whisper_cpp/ggml/src/ggml-sycl/convert.cpp
new file mode 100644
index 00000000..5fd15e6c
--- /dev/null
+++ b/src/whisper_cpp/ggml/src/ggml-sycl/convert.cpp
@@ -0,0 +1,547 @@
+#include "convert.hpp"
+#include "dequantize.hpp"
+#include "presets.hpp"
+
+template <int qk, int qr, dequantize_kernel_t dequantize_kernel, typename dst_t>
+static void dequantize_block(const void * __restrict__ vx, dst_t * __restrict__ y, const int64_t k,
+                             const sycl::nd_item<3> &item_ct1) {
+    const int64_t i = 2 * (item_ct1.get_local_range(2) * item_ct1.get_group(2) +
+                       item_ct1.get_local_id(2));
+
+    if (i >= k) {
+        return;
+    }
+
+    const int64_t ib = i/qk; // block index
+    const int64_t iqs = (i%qk)/qr; // quant index
+    const int64_t iybs = i - i%qk; // y block start index
+    const int64_t y_offset = qr == 1 ? 1 : qk/2;
+
+    // dequantize
+    dfloat2 v;
+    dequantize_kernel(vx, ib, iqs, v);
+
+    y[iybs + iqs + 0] = v.x();
+    y[iybs + iqs + y_offset] = v.y();
+}
+
+template <int qk, int qr, dequantize_kernel_t dequantize_kernel, typename dst_t>
+static void dequantize_block_sycl(const void *__restrict__ vx,
+                                  dst_t *__restrict__ y, const int64_t k,
+                                  dpct::queue_ptr stream) {
+    const int64_t num_blocks = (k + 2*SYCL_DEQUANTIZE_BLOCK_SIZE - 1) / (2*SYCL_DEQUANTIZE_BLOCK_SIZE);
+    {
+        dpct::has_capability_or_fail(stream->get_device(),
+                                     {sycl::aspect::fp16});
+        stream->parallel_for(
+            sycl::nd_range<3>(
+                sycl::range<3>(1, 1, num_blocks) *
+                    sycl::range<3>(1, 1, SYCL_DEQUANTIZE_BLOCK_SIZE),
+                sycl::range<3>(1, 1, SYCL_DEQUANTIZE_BLOCK_SIZE)),
+            [=](sycl::nd_item<3> item_ct1) {
+                dequantize_block<qk, qr, dequantize_kernel>(vx, y, k, item_ct1);
+            });
+    }
+}
+
+template <typename dst_t>
+static void dequantize_row_q2_K_sycl(const void *vx, dst_t *y, const int64_t k,
+                                     dpct::queue_ptr stream) {
+    const int64_t nb = k / QK_K;
+#if QK_K == 256
+    {
+        dpct::has_capability_or_fail(stream->get_device(),
+                                     {sycl::aspect::fp16});
+
+        stream->parallel_for(sycl::nd_range<3>(sycl::range<3>(1, 1, nb) *
+                                                   sycl::range<3>(1, 1, 64),
+                                               sycl::range<3>(1, 1, 64)),
+                             [=](sycl::nd_item<3> item_ct1) {
+                                 dequantize_block_q2_K(vx, y, item_ct1);
+                             });
+    }
+#else
+    {
+        dpct::has_capability_or_fail(stream->get_device(),
+                                     {sycl::aspect::fp16});
+
+        stream->parallel_for(sycl::nd_range<3>(sycl::range<3>(1, 1, nb) *
+                                                   sycl::range<3>(1, 1, 32),
+                                               sycl::range<3>(1, 1, 32)),
+                             [=](sycl::nd_item<3> item_ct1) {
+                                 dequantize_block_q2_K(vx, y, item_ct1);
+                             });
+    }
+
+#endif
+}
+
+template <typename dst_t>
+static void dequantize_row_q3_K_sycl(const void *vx, dst_t *y, const int64_t k,
+                                     dpct::queue_ptr stream) {
+    const int64_t nb = k / QK_K;
+#if QK_K == 256
+    {
+        dpct::has_capability_or_fail(stream->get_device(),
+                                     {sycl::aspect::fp16});
+
+        stream->parallel_for(sycl::nd_range<3>(sycl::range<3>(1, 1, nb) *
+                                                   sycl::range<3>(1, 1, 64),
+                                               sycl::range<3>(1, 1, 64)),
+                             [=](sycl::nd_item<3> item_ct1) {
+                                 dequantize_block_q3_K(vx, y, item_ct1);
+                             });
+    }
+#else
+    {
+        dpct::has_capability_or_fail(stream->get_device(),
+                                     {sycl::aspect::fp16});
+
+        stream->parallel_for(sycl::nd_range<3>(sycl::range<3>(1, 1, nb) *
+                                                   sycl::range<3>(1, 1, 32),
+                                               sycl::range<3>(1, 1, 32)),
+                             [=](sycl::nd_item<3> item_ct1) {
+                                 dequantize_block_q3_K(vx, y, item_ct1);
+                             });
+    }
+#endif
+}
+
+template <typename dst_t>
+static void dequantize_row_q4_0_sycl(const void *vx, dst_t *y, const int64_t k,
+                                     dpct::queue_ptr stream) {
+    const int64_t nb32 = k / 32;
+    const int64_t nb = (k + 255) / 256;
+    {
+        dpct::has_capability_or_fail(stream->get_device(),
+                                     {sycl::aspect::fp16});
+
+        stream->parallel_for(sycl::nd_range<3>(sycl::range<3>(1, 1, nb) *
+                                                   sycl::range<3>(1, 1, 32),
+                                               sycl::range<3>(1, 1, 32)),
+                             [=](sycl::nd_item<3> item_ct1) {
+                                 dequantize_block_q4_0(vx, y, nb32, item_ct1);
+                             });
+    }
+}
+
+template <typename dst_t>
+static void dequantize_row_q4_1_sycl(const void *vx, dst_t *y, const int64_t k,
+                                     dpct::queue_ptr stream) {
+    const int64_t nb32 = k / 32;
+    const int64_t nb = (k + 255) / 256;
+    {
+        dpct::has_capability_or_fail(stream->get_device(),
+                                     {sycl::aspect::fp16});
+
+        stream->parallel_for(sycl::nd_range<3>(sycl::range<3>(1, 1, nb) *
+                                                   sycl::range<3>(1, 1, 32),
+                                               sycl::range<3>(1, 1, 32)),
+                             [=](sycl::nd_item<3> item_ct1) {
+                                 dequantize_block_q4_1(vx, y, nb32, item_ct1);
+                             });
+    }
+}
+
+
+template <typename dst_t>
+static void dequantize_row_q4_K_sycl(const void *vx, dst_t *y, const int64_t k,
+                                     dpct::queue_ptr stream) {
+    const int64_t nb = k / QK_K;
+    {
+        dpct::has_capability_or_fail(stream->get_device(),
+                                     {sycl::aspect::fp16});
+
+        stream->submit([&](sycl::handler &cgh) {
+            sycl::local_accessor<uint8_t, 1> scale_local_acc(sycl::range<1>(12), cgh);
+            cgh.parallel_for(sycl::nd_range<3>(sycl::range<3>(1, 1, nb) *
+                                                   sycl::range<3>(1, 1, 32),
+                                               sycl::range<3>(1, 1, 32)),
+                             [=](sycl::nd_item<3> item_ct1) {
+                                 dequantize_block_q4_K(vx, y, get_pointer(scale_local_acc), item_ct1);
+                             });
+        });
+    }
+}
+
+template <typename dst_t>
+static void dequantize_row_q5_K_sycl(const void *vx, dst_t *y, const int64_t k,
+                                     dpct::queue_ptr stream) {
+    const int64_t nb = k / QK_K;
+#if QK_K == 256
+    {
+        dpct::has_capability_or_fail(stream->get_device(),
+                                     {sycl::aspect::fp16});
+
+        stream->parallel_for(sycl::nd_range<3>(sycl::range<3>(1, 1, nb) *
+                                                   sycl::range<3>(1, 1, 64),
+                                               sycl::range<3>(1, 1, 64)),
+                             [=](sycl::nd_item<3> item_ct1) {
+                                 dequantize_block_q5_K(vx, y, item_ct1);
+                             });
+    }
+#else
+    {
+        dpct::has_capability_or_fail(stream->get_device(),
+                                     {sycl::aspect::fp16});
+
+        stream->parallel_for(sycl::nd_range<3>(sycl::range<3>(1, 1, nb) *
+                                                   sycl::range<3>(1, 1, 32),
+                                               sycl::range<3>(1, 1, 32)),
+                             [=](sycl::nd_item<3> item_ct1) {
+                                 dequantize_block_q5_K(vx, y, item_ct1);
+                             });
+    }
+
+#endif
+}
+
+template <typename dst_t>
+static void dequantize_row_q6_K_sycl(const void *vx, dst_t *y, const int64_t k,
+                                     dpct::queue_ptr stream) {
+    const int64_t nb = k / QK_K;
+#if QK_K == 256
+    {
+        dpct::has_capability_or_fail(stream->get_device(),
+                                     {sycl::aspect::fp16});
+
+        stream->parallel_for(sycl::nd_range<3>(sycl::range<3>(1, 1, nb) *
+                                                   sycl::range<3>(1, 1, 64),
+                                               sycl::range<3>(1, 1, 64)),
+                             [=](sycl::nd_item<3> item_ct1) {
+                                 dequantize_block_q6_K(vx, y, item_ct1);
+                             });
+    }
+#else
+    {
+        dpct::has_capability_or_fail(stream->get_device(),
+                                     {sycl::aspect::fp16});
+
+        stream->parallel_for(sycl::nd_range<3>(sycl::range<3>(1, 1, nb) *
+                                                   sycl::range<3>(1, 1, 32),
+                                               sycl::range<3>(1, 1, 32)),
+                             [=](sycl::nd_item<3> item_ct1) {
+                                 dequantize_block_q6_K(vx, y, item_ct1);
+                             });
+    }
+
+#endif
+}
+
+template <typename dst_t>
+static void dequantize_row_iq1_s_sycl(const void *vx, dst_t *y, const int64_t k,
+                                        dpct::queue_ptr stream) {
+    const int64_t nb = k / QK_K;
+    {
+        dpct::has_capability_or_fail(stream->get_device(),
+                                     {sycl::aspect::fp16});
+
+        stream->submit([&](sycl::handler &cgh) {
+            cgh.parallel_for(sycl::nd_range<3>(sycl::range<3>(1, 1, nb) *
+                                                   sycl::range<3>(1, 1, 32),
+                                               sycl::range<3>(1, 1, 32)),
+                             [=](sycl::nd_item<3> item_ct1) {
+                                 dequantize_block_iq1_s(
+                                     vx, y, item_ct1, iq1s_grid_gpu
+                                     );
+                             });
+        });
+    }
+}
+
+template <typename dst_t>
+static void dequantize_row_iq1_m_sycl(const void *vx, dst_t *y, const int64_t k,
+                                        dpct::queue_ptr stream) {
+    const int64_t nb = k / QK_K;
+    {
+        dpct::has_capability_or_fail(stream->get_device(),
+                                     {sycl::aspect::fp16});
+
+        stream->submit([&](sycl::handler &cgh) {
+            cgh.parallel_for(sycl::nd_range<3>(sycl::range<3>(1, 1, nb) *
+                                                   sycl::range<3>(1, 1, 32),
+                                               sycl::range<3>(1, 1, 32)),
+                             [=](sycl::nd_item<3> item_ct1) {
+                                 dequantize_block_iq1_m(
+                                     vx, y, item_ct1, iq1s_grid_gpu
+                                     );
+                             });
+        });
+    }
+}
+
+template <typename dst_t>
+static void dequantize_row_iq2_xxs_sycl(const void *vx, dst_t *y, const int64_t k,
+                                        dpct::queue_ptr stream) {
+    const int64_t nb = k / QK_K;
+    {
+        dpct::has_capability_or_fail(stream->get_device(),
+                                     {sycl::aspect::fp16});
+
+        stream->submit([&](sycl::handler &cgh) {
+            cgh.parallel_for(sycl::nd_range<3>(sycl::range<3>(1, 1, nb) *
+                                                   sycl::range<3>(1, 1, 32),
+                                               sycl::range<3>(1, 1, 32)),
+                             [=](sycl::nd_item<3> item_ct1) {
+                                 dequantize_block_iq2_xxs(
+                                     vx, y, item_ct1, iq2xxs_grid,
+                                     ksigns_iq2xs, kmask_iq2xs);
+                             });
+        });
+    }
+}
+
+template <typename dst_t>
+static void dequantize_row_iq2_xs_sycl(const void *vx, dst_t *y, const int64_t k,
+                                       dpct::queue_ptr stream) {
+    const int64_t nb = k / QK_K;
+    {
+        dpct::has_capability_or_fail(stream->get_device(),
+                                     {sycl::aspect::fp16});
+
+        stream->submit([&](sycl::handler &cgh) {
+            cgh.parallel_for(sycl::nd_range<3>(sycl::range<3>(1, 1, nb) *
+                                                   sycl::range<3>(1, 1, 32),
+                                               sycl::range<3>(1, 1, 32)),
+                             [=](sycl::nd_item<3> item_ct1) {
+                                 dequantize_block_iq2_xs(
+                                     vx, y, item_ct1, iq2xs_grid,
+                                     ksigns_iq2xs, kmask_iq2xs);
+                             });
+        });
+    }
+}
+
+template <typename dst_t>
+static void dequantize_row_iq2_s_sycl(const void *vx, dst_t *y, const int64_t k,
+                                      dpct::queue_ptr stream) {
+    const int64_t nb = k / QK_K;
+    {
+        dpct::has_capability_or_fail(stream->get_device(),
+                                     {sycl::aspect::fp16});
+
+        stream->submit([&](sycl::handler &cgh) {
+            cgh.parallel_for(sycl::nd_range<3>(sycl::range<3>(1, 1, nb) *
+                                                   sycl::range<3>(1, 1, 32),
+                                               sycl::range<3>(1, 1, 32)),
+                             [=](sycl::nd_item<3> item_ct1) {
+                                 dequantize_block_iq2_s(vx, y, item_ct1);
+                             });
+        });
+    }
+}
+
+
+template <typename dst_t>
+static void dequantize_row_iq3_xxs_sycl(const void *vx, dst_t *y, const int64_t k,
+                                        dpct::queue_ptr stream) {
+    const int64_t nb = k / QK_K;
+    {
+        dpct::has_capability_or_fail(stream->get_device(),
+                                     {sycl::aspect::fp16});
+
+        stream->submit([&](sycl::handler &cgh) {
+            cgh.parallel_for(sycl::nd_range<3>(sycl::range<3>(1, 1, nb) *
+                                                   sycl::range<3>(1, 1, 32),
+                                               sycl::range<3>(1, 1, 32)),
+                             [=](sycl::nd_item<3> item_ct1) {
+                                 dequantize_block_iq3_xxs(
+                                     vx, y, item_ct1, iq3xxs_grid,
+                                     ksigns_iq2xs, kmask_iq2xs);
+                             });
+        });
+    }
+}
+
+template <typename dst_t>
+static void dequantize_row_iq3_s_sycl(const void *vx, dst_t *y, const int64_t k,
+                                        dpct::queue_ptr stream) {
+    const int64_t nb = k / QK_K;
+    {
+        dpct::has_capability_or_fail(stream->get_device(),
+                                     {sycl::aspect::fp16});
+
+        stream->submit([&](sycl::handler &cgh) {
+            cgh.parallel_for(sycl::nd_range<3>(sycl::range<3>(1, 1, nb) *
+                                                   sycl::range<3>(1, 1, 32),
+                                               sycl::range<3>(1, 1, 32)),
+                             [=](sycl::nd_item<3> item_ct1) {
+                                 dequantize_block_iq3_s(
+                                     vx, y, item_ct1, kmask_iq2xs, iq3s_grid);
+                             });
+        });
+    }
+}
+
+template <typename dst_t>
+static void dequantize_row_iq4_xs_sycl(const void *vx, dst_t *y, const int64_t k,
+                                       dpct::queue_ptr stream) {
+    const int64_t nb = (k + QK_K - 1) / QK_K;
+#if QK_K == 64
+    dequantize_row_iq4_nl_sycl(vx, y, k, stream);
+#else
+      {
+            dpct::has_capability_or_fail(stream->get_device(),
+                                         {sycl::aspect::fp16});
+
+            stream->submit([&](sycl::handler &cgh) {
+                  cgh.parallel_for(
+                      sycl::nd_range<3>(sycl::range<3>(1, 1, nb) *
+                                            sycl::range<3>(1, 1, 32),
+                                        sycl::range<3>(1, 1, 32)),
+                      [=](sycl::nd_item<3> item_ct1) {
+                            dequantize_block_iq4_xs(vx, y, item_ct1);
+                      });
+            });
+      }
+#endif
+}
+
+template <typename dst_t>
+static void dequantize_row_iq4_nl_sycl(const void *vx, dst_t *y, const int64_t k,
+                                       dpct::queue_ptr stream) {
+    const int64_t nb = (k + QK_K - 1) / QK_K;
+      {
+            dpct::has_capability_or_fail(stream->get_device(),
+                                         {sycl::aspect::fp16});
+
+            stream->submit([&](sycl::handler &cgh) {
+                  cgh.parallel_for(
+                      sycl::nd_range<3>(sycl::range<3>(1, 1, nb) *
+                                            sycl::range<3>(1, 1, 32),
+                                        sycl::range<3>(1, 1, 32)),
+                      [=](sycl::nd_item<3> item_ct1) {
+                            dequantize_block_iq4_nl(vx, y, item_ct1);
+                      });
+            });
+      }
+}
+
+template <typename src_t, typename dst_t>
+static void convert_unary(const void * __restrict__ vx, dst_t * __restrict__ y, const int64_t k,
+                          const sycl::nd_item<3> &item_ct1) {
+    const int64_t work_group_size = item_ct1.get_local_range(2);
+    const int64_t global_id = item_ct1.get_local_id(2) + work_group_size * item_ct1.get_group(2);
+
+    // make each work-item deal with more elements since sycl global range can not exceed max int
+    const src_t * x = (src_t *) vx;
+    for (int64_t i = global_id; i < k; i += work_group_size * item_ct1.get_group_range(2)) {
+        y[i] = x[i];
+    }
+}
+
+template <typename src_t, typename dst_t>
+static void convert_unary_sycl(const void *__restrict__ vx,
+                               dst_t *__restrict__ y, const int64_t k,
+                               dpct::queue_ptr stream) {
+    const int64_t num_blocks = (k + SYCL_DEQUANTIZE_BLOCK_SIZE - 1) / SYCL_DEQUANTIZE_BLOCK_SIZE;
+
+    // decrease global range when it exceeds the max int
+    int64_t local_size = downsample_sycl_global_range(num_blocks, SYCL_DEQUANTIZE_BLOCK_SIZE);
+    sycl::range<3> block_nums(1, 1, num_blocks);
+    sycl::range<3> local_range(1, 1, local_size);
+    {
+        dpct::has_capability_or_fail(stream->get_device(),
+                                     {sycl::aspect::fp16});
+
+        stream->parallel_for(
+            sycl::nd_range<3>(block_nums * local_range, local_range),
+            [=](sycl::nd_item<3> item_ct1) {
+                convert_unary<src_t>(vx, y, k, item_ct1);
+            });
+    }
+}
+
+to_fp16_sycl_t ggml_get_to_fp16_sycl(ggml_type type) {
+    switch (type) {
+        case GGML_TYPE_Q4_0:
+            return dequantize_block_sycl<QK4_0, QR4_0, dequantize_q4_0>;
+        case GGML_TYPE_Q4_1:
+            return dequantize_block_sycl<QK4_1, QR4_1, dequantize_q4_1>;
+        case GGML_TYPE_Q5_0:
+            return dequantize_block_sycl<QK5_0, QR5_0, dequantize_q5_0>;
+        case GGML_TYPE_Q5_1:
+            return dequantize_block_sycl<QK5_1, QR5_1, dequantize_q5_1>;
+        case GGML_TYPE_Q8_0:
+            return dequantize_block_sycl<QK8_0, QR8_0, dequantize_q8_0>;
+        case GGML_TYPE_Q2_K:
+            return dequantize_row_q2_K_sycl;
+        case GGML_TYPE_Q3_K:
+            return dequantize_row_q3_K_sycl;
+        case GGML_TYPE_Q4_K:
+            return dequantize_row_q4_K_sycl;
+        case GGML_TYPE_Q5_K:
+            return dequantize_row_q5_K_sycl;
+        case GGML_TYPE_Q6_K:
+            return dequantize_row_q6_K_sycl;
+        case GGML_TYPE_IQ1_S:
+            return dequantize_row_iq1_s_sycl;
+        case GGML_TYPE_IQ1_M:
+            return dequantize_row_iq1_m_sycl;
+        case GGML_TYPE_IQ2_XXS:
+            return dequantize_row_iq2_xxs_sycl;
+        case GGML_TYPE_IQ2_XS:
+            return dequantize_row_iq2_xs_sycl;
+        case GGML_TYPE_IQ2_S:
+            return dequantize_row_iq2_s_sycl;
+        case GGML_TYPE_IQ3_XXS:
+            return dequantize_row_iq3_xxs_sycl;
+        case GGML_TYPE_IQ3_S:
+            return dequantize_row_iq3_s_sycl;
+        case GGML_TYPE_IQ4_XS:
+            return dequantize_row_iq4_xs_sycl;
+        case GGML_TYPE_IQ4_NL:
+            return dequantize_row_iq4_nl_sycl;
+        case GGML_TYPE_F32:
+            return convert_unary_sycl<float>;
+        default:
+            return nullptr;
+    }
+}
+
+to_fp32_sycl_t ggml_get_to_fp32_sycl(ggml_type type) {
+    switch (type) {
+        case GGML_TYPE_Q4_0:
+            return dequantize_row_q4_0_sycl;
+        case GGML_TYPE_Q4_1:
+            return dequantize_row_q4_1_sycl;
+        case GGML_TYPE_Q5_0:
+            return dequantize_block_sycl<QK5_0, QR5_0, dequantize_q5_0>;
+        case GGML_TYPE_Q5_1:
+            return dequantize_block_sycl<QK5_1, QR5_1, dequantize_q5_1>;
+        case GGML_TYPE_Q8_0:
+            return dequantize_block_sycl<QK8_0, QR8_0, dequantize_q8_0>;
+        case GGML_TYPE_Q2_K:
+            return dequantize_row_q2_K_sycl;
+        case GGML_TYPE_Q3_K:
+            return dequantize_row_q3_K_sycl;
+        case GGML_TYPE_Q4_K:
+            return dequantize_row_q4_K_sycl;
+        case GGML_TYPE_Q5_K:
+            return dequantize_row_q5_K_sycl;
+        case GGML_TYPE_Q6_K:
+            return dequantize_row_q6_K_sycl;
+        case GGML_TYPE_IQ1_S:
+            return dequantize_row_iq1_s_sycl;
+        case GGML_TYPE_IQ1_M:
+            return dequantize_row_iq1_m_sycl;
+        case GGML_TYPE_IQ2_XXS:
+            return dequantize_row_iq2_xxs_sycl;
+        case GGML_TYPE_IQ2_XS:
+            return dequantize_row_iq2_xs_sycl;
+        case GGML_TYPE_IQ2_S:
+            return dequantize_row_iq2_s_sycl;
+        case GGML_TYPE_IQ3_XXS:
+            return dequantize_row_iq3_xxs_sycl;
+        case GGML_TYPE_IQ3_S:
+            return dequantize_row_iq3_s_sycl;
+        case GGML_TYPE_IQ4_XS:
+            return dequantize_row_iq4_xs_sycl;
+        case GGML_TYPE_IQ4_NL:
+            return dequantize_row_iq4_nl_sycl;
+        case GGML_TYPE_F16:
+            return convert_unary_sycl<sycl::half>;
+        default:
+            return nullptr;
+    }
+}
diff --git a/src/whisper_cpp/ggml/src/ggml-sycl/convert.hpp b/src/whisper_cpp/ggml/src/ggml-sycl/convert.hpp
new file mode 100644
index 00000000..0ce2874a
--- /dev/null
+++ b/src/whisper_cpp/ggml/src/ggml-sycl/convert.hpp
@@ -0,0 +1,27 @@
+//
+// MIT license
+// Copyright (C) 2024 Intel Corporation
+// SPDX-License-Identifier: MIT
+//
+
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+//
+
+#ifndef GGML_SYCL_CONVERT_HPP
+#define GGML_SYCL_CONVERT_HPP
+
+#include "common.hpp"
+
+template <typename T>
+using to_t_sycl_t = void (*)(const void *__restrict__ x, T *__restrict__ y,
+                             int64_t k, dpct::queue_ptr stream);
+typedef to_t_sycl_t<float> to_fp32_sycl_t;
+typedef to_t_sycl_t<sycl::half> to_fp16_sycl_t;
+
+to_fp16_sycl_t ggml_get_to_fp16_sycl(ggml_type type);
+to_fp32_sycl_t ggml_get_to_fp32_sycl(ggml_type type);
+
+#endif // GGML_SYCL_CONVERT_HPP
diff --git a/src/whisper_cpp/ggml/src/ggml-sycl/dequantize.hpp b/src/whisper_cpp/ggml/src/ggml-sycl/dequantize.hpp
new file mode 100644
index 00000000..b8304c3a
--- /dev/null
+++ b/src/whisper_cpp/ggml/src/ggml-sycl/dequantize.hpp
@@ -0,0 +1,698 @@
+//
+// MIT license
+// Copyright (C) 2024 Intel Corporation
+// SPDX-License-Identifier: MIT
+//
+
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+//
+
+#ifndef GGML_SYCL_DEQUANTIZE_HPP
+#define GGML_SYCL_DEQUANTIZE_HPP
+
+#include "common.hpp"
+
+typedef void (*dequantize_kernel_t)(const void * vx, const int64_t ib, const int iqs, dfloat2 & v);
+
+static __dpct_inline__ void dequantize_q4_0(const void *vx, const int64_t ib,
+                                            const int iqs, dfloat2 &v) {
+    const block_q4_0 * x = (const block_q4_0 *) vx;
+
+    const dfloat d = x[ib].d;
+
+    const int vui = x[ib].qs[iqs];
+
+    v.x() = vui & 0xF;
+    v.y() = vui >> 4;
+
+#ifdef GGML_SYCL_F16
+    // v = v - {8.0f, 8.0f};
+    // v = v * {d, d};
+    v.s0() = (v.s0() - 8.0f) * d;
+    v.s1() = (v.s1() - 8.0f) * d;
+
+#else
+    v.x() = (v.x() - 8.0f) * d;
+    v.y() = (v.y() - 8.0f) * d;
+#endif // GGML_SYCL_F16
+}
+
+static __dpct_inline__ void dequantize_q4_1(const void *vx, const int64_t ib,
+                                            const int iqs, dfloat2 &v) {
+    const block_q4_1 * x = (const block_q4_1 *) vx;
+
+    const dfloat d = x[ib].dm[0];
+    const dfloat m = x[ib].dm[1];
+
+    const int vui = x[ib].qs[iqs];
+
+    v.x() = vui & 0xF;
+    v.y() = vui >> 4;
+
+#ifdef GGML_SYCL_F16
+    // v = v * {d, d};
+    // v = v + {m, m};
+    v.s0() = sycl::fma(v.s0(), d, m);
+    v.s1() = sycl::fma(v.s1(), d, m);
+
+#else
+    v.x() = sycl::fma(v.x(), d, m);
+    v.y() = sycl::fma(v.y(), d, m);
+#endif // GGML_SYCL_F16
+}
+
+static __dpct_inline__ void dequantize_q5_0(const void *vx, const int64_t ib,
+                                            const int iqs, dfloat2 &v) {
+    const block_q5_0 * x = (const block_q5_0 *) vx;
+
+    const dfloat d = x[ib].d;
+
+    uint32_t qh;
+    memcpy(&qh, x[ib].qh, sizeof(qh));
+
+    const int xh_0 = ((qh >> (iqs +  0)) << 4) & 0x10;
+    const int xh_1 = ((qh >> (iqs + 12))     ) & 0x10;
+
+    v.x() = ((x[ib].qs[iqs] & 0xf) | xh_0);
+    v.y() = ((x[ib].qs[iqs] >> 4) | xh_1);
+
+#ifdef GGML_SYCL_F16
+    // v = v - {16.0f, 16.0f};
+    // v = v * {d, d};
+    v.s0() = (v.s0() - 16.0f) * d;
+    v.s1() = (v.s1() - 16.0f) * d;
+
+#else
+    v.x() = (v.x() - 16.0f) * d;
+    v.y() = (v.y() - 16.0f) * d;
+#endif // GGML_SYCL_F16
+}
+
+static __dpct_inline__ void dequantize_q5_1(const void *vx, const int64_t ib,
+                                            const int iqs, dfloat2 &v) {
+    const block_q5_1 * x = (const block_q5_1 *) vx;
+
+    const dfloat d = x[ib].dm[0];
+    const dfloat m = x[ib].dm[1];
+
+    uint32_t qh;
+    memcpy(&qh, x[ib].qh, sizeof(qh));
+
+    const int xh_0 = ((qh >> (iqs +  0)) << 4) & 0x10;
+    const int xh_1 = ((qh >> (iqs + 12))     ) & 0x10;
+
+    v.x() = ((x[ib].qs[iqs] & 0xf) | xh_0);
+    v.y() = ((x[ib].qs[iqs] >> 4) | xh_1);
+
+#ifdef GGML_SYCL_F16
+    // v = v * {d, d};
+    // v = v + {m, m};
+    v.s0() = sycl::fma(v.s0(), d, m);
+    v.s1() = sycl::fma(v.s1(), d, m);
+#else
+    v.x() = sycl::fma(v.x(), d, m);
+    v.y() = sycl::fma(v.y(), d, m);
+#endif // GGML_SYCL_F16
+}
+
+static __dpct_inline__ void dequantize_q8_0(const void *vx, const int64_t ib,
+                                            const int iqs, dfloat2 &v) {
+    const block_q8_0 * x = (const block_q8_0 *) vx;
+
+    const dfloat d = x[ib].d;
+
+    v.x() = x[ib].qs[iqs + 0];
+    v.y() = x[ib].qs[iqs + 1];
+
+#ifdef GGML_SYCL_F16
+    // v = v * {d, d};
+    v.s0() *= d;
+    v.s1() *= d;
+#else
+    v.x() *= d;
+    v.y() *= d;
+#endif // GGML_SYCL_F16
+}
+
+template<typename dst_t>
+static void dequantize_block_q4_0(const void * __restrict__ vx, dst_t * __restrict__ yy, int64_t nb32,
+                                  const sycl::nd_item<3> &item_ct1) {
+
+    const int64_t i = item_ct1.get_group(2);
+
+    // assume 32 threads
+    const int64_t tid = item_ct1.get_local_id(2);
+    const int64_t il  = tid/8;
+    const int64_t ir  = tid%8;
+    const int64_t ib = 8*i + ir;
+    if (ib >= nb32) {
+        return;
+    }
+
+    dst_t * y = yy + 256*i + 32*ir + 4*il;
+
+    const block_q4_0 * x = (const block_q4_0 *)vx + ib;
+    const float d = sycl::vec<sycl::half, 1>(x->d)
+                        .convert<float, sycl::rounding_mode::automatic>()[0];
+    const float dm = -8*d;
+
+    const uint8_t * q = x->qs + 4*il;
+
+    for (int l = 0; l < 4; ++l) {
+        y[l+ 0] = d * (q[l] & 0xF) + dm;
+        y[l+16] = d * (q[l] >>  4) + dm;
+    }
+}
+
+template<typename dst_t>
+static void dequantize_block_q4_1(const void * __restrict__ vx, dst_t * __restrict__ yy, int64_t nb32,
+                                  const sycl::nd_item<3> &item_ct1) {
+
+    const int64_t i = item_ct1.get_group(2);
+
+    // assume 32 threads
+    const int64_t tid = item_ct1.get_local_id(2);
+    const int64_t il  = tid/8;
+    const int64_t ir  = tid%8;
+    const int64_t ib = 8*i + ir;
+    if (ib >= nb32) {
+        return;
+    }
+
+    dst_t * y = yy + 256*i + 32*ir + 4*il;
+
+    const block_q4_1 * x = (const block_q4_1 *)vx + ib;
+    const sycl::float2 d =
+        x->dm.convert<float, sycl::rounding_mode::automatic>();
+
+    const uint8_t * q = x->qs + 4*il;
+
+    for (int l = 0; l < 4; ++l) {
+        y[l + 0] = d.x() * (q[l] & 0xF) + d.y();
+        y[l + 16] = d.x() * (q[l] >> 4) + d.y();
+    }
+}
+
+
+//================================== k-quants
+
+template<typename dst_t>
+static void dequantize_block_q2_K(const void * __restrict__ vx, dst_t * __restrict__ yy,
+                                  const sycl::nd_item<3> &item_ct1) {
+
+    const int64_t i = item_ct1.get_group(2);
+    const block_q2_K * x = (const block_q2_K *) vx;
+
+    const int64_t tid = item_ct1.get_local_id(2);
+#if QK_K == 256
+    const int64_t n   = tid/32;
+    const int64_t l   = tid - 32*n;
+    const int64_t is  = 8*n + l/16;
+
+    const uint8_t q = x[i].qs[32*n + l];
+    dst_t * y = yy + i*QK_K + 128*n;
+
+    float dall = x[i].dm[0];
+    float dmin = x[i].dm[1];
+    y[l+ 0] = dall * (x[i].scales[is+0] & 0xF) * ((q >> 0) & 3) - dmin * (x[i].scales[is+0] >> 4);
+    y[l+32] = dall * (x[i].scales[is+2] & 0xF) * ((q >> 2) & 3) - dmin * (x[i].scales[is+2] >> 4);
+    y[l+64] = dall * (x[i].scales[is+4] & 0xF) * ((q >> 4) & 3) - dmin * (x[i].scales[is+4] >> 4);
+    y[l+96] = dall * (x[i].scales[is+6] & 0xF) * ((q >> 6) & 3) - dmin * (x[i].scales[is+6] >> 4);
+#else
+    const int64_t is = tid/16;  // 0 or 1
+    const int64_t il = tid%16;  // 0...15
+    const uint8_t q = x[i].qs[il] >> (2*is);
+    dst_t * y = yy + i*QK_K + 16*is + il;
+
+    float dall = x[i].dm[0];
+    float dmin = x[i].dm[1];
+    y[ 0] = dall * (x[i].scales[is+0] & 0xF) * ((q >> 0) & 3) - dmin * (x[i].scales[is+0] >> 4);
+    y[32] = dall * (x[i].scales[is+2] & 0xF) * ((q >> 4) & 3) - dmin * (x[i].scales[is+2] >> 4);
+#endif
+
+}
+
+template<typename dst_t>
+static void dequantize_block_q3_K(const void * __restrict__ vx, dst_t * __restrict__ yy,
+                                  const sycl::nd_item<3> &item_ct1) {
+
+    const int64_t i = item_ct1.get_group(2);
+    const block_q3_K * x = (const block_q3_K *) vx;
+
+#if QK_K == 256
+    const int64_t r = item_ct1.get_local_id(2) / 4;
+    const int64_t tid = r/2;
+    const int64_t is0 = r%2;
+    const int64_t l0 = 16 * is0 + 4 * (item_ct1.get_local_id(2) % 4);
+    const int64_t n = tid / 4;
+    const int64_t j = tid - 4*n;
+
+    uint8_t m = 1 << (4*n + j);
+    int64_t is = 8*n + 2*j + is0;
+    int shift = 2*j;
+
+    int8_t us = is <  4 ? (x[i].scales[is-0] & 0xF) | (((x[i].scales[is+8] >> 0) & 3) << 4) :
+                is <  8 ? (x[i].scales[is-0] & 0xF) | (((x[i].scales[is+4] >> 2) & 3) << 4) :
+                is < 12 ? (x[i].scales[is-8] >>  4) | (((x[i].scales[is+0] >> 4) & 3) << 4) :
+                          (x[i].scales[is-8] >>  4) | (((x[i].scales[is-4] >> 6) & 3) << 4);
+    float d_all = x[i].d;
+    float dl = d_all * (us - 32);
+
+    dst_t * y = yy + i*QK_K + 128*n + 32*j;
+    const uint8_t * q = x[i].qs + 32*n;
+    const uint8_t * hm = x[i].hmask;
+
+    for (int l = l0; l < l0+4; ++l) y[l] = dl * ((int8_t)((q[l] >> shift) & 3) - ((hm[l] & m) ? 0 : 4));
+#else
+    const int64_t tid = item_ct1.get_local_id(2);
+    const int64_t is  = tid/16;  // 0 or 1
+    const int64_t il  = tid%16;  // 0...15
+    const int64_t im  = il/8;    // 0...1
+    const int64_t in  = il%8;    // 0...7
+
+    dst_t * y = yy + i*QK_K + 16*is + il;
+
+    const uint8_t q = x[i].qs[il] >> (2*is);
+    const uint8_t h = x[i].hmask[in] >> (2*is + im);
+    const float   d = (float)x[i].d;
+
+    if (is == 0) {
+        y[ 0] = d * ((x[i].scales[0] & 0xF) - 8) * ((int8_t)((q >> 0) & 3) - ((h >> 0) & 1 ? 0 : 4));
+        y[32] = d * ((x[i].scales[1] & 0xF) - 8) * ((int8_t)((q >> 4) & 3) - ((h >> 4) & 1 ? 0 : 4));
+    } else {
+        y[ 0] = d * ((x[i].scales[0] >>  4) - 8) * ((int8_t)((q >> 0) & 3) - ((h >> 0) & 1 ? 0 : 4));
+        y[32] = d * ((x[i].scales[1] >>  4) - 8) * ((int8_t)((q >> 4) & 3) - ((h >> 4) & 1 ? 0 : 4));
+    }
+#endif
+
+}
+
+#if QK_K == 256
+static inline void get_scale_min_k4(int j, const uint8_t * q, uint8_t & d, uint8_t & m) {
+    if (j < 4) {
+        d = q[j] & 63;
+        m = q[j + 4] & 63;
+    } else {
+        d = (q[j+4] & 0xF) | ((q[j-4] >> 6) << 4);
+        m = (q[j+4] >>  4) | ((q[j-0] >> 6) << 4);
+    }
+}
+#endif
+
+template<typename dst_t>
+static void dequantize_block_q4_K(const void * __restrict__ vx, dst_t * __restrict__ yy,
+                                  uint8_t* scales_local, const sycl::nd_item<3> &item_ct1) {
+    const block_q4_K * x = (const block_q4_K *) vx;
+
+    const int64_t i = item_ct1.get_group(2);
+
+#if QK_K == 256
+    // assume 32 threads
+    const int64_t tid = item_ct1.get_local_id(2);
+    const int64_t il  = tid/8;
+    const int64_t ir  = tid%8;
+    const int64_t is  = 2*il;
+    const int64_t n   = 4;
+
+    dst_t * y = yy + i*QK_K + 64*il + n*ir;
+
+    const sycl::half2 dm = x[i].dm;
+    const float dall = dm[0];
+    const float dmin = dm[1];
+
+    if (tid < 12)
+        scales_local[tid] = x[i].scales[tid];
+    item_ct1.barrier(sycl::access::fence_space::local_space);
+
+    uint8_t sc, m;
+    get_scale_min_k4(is + 0, scales_local, sc, m);
+    const float d1 = dall * sc;
+    const float m1 = dmin * m;
+    get_scale_min_k4(is + 1, scales_local, sc, m);
+    const float d2 = dall * sc;
+    const float m2 = dmin * m;
+
+    sycl::vec<uint8_t, n> q_vec = vec_aligned_load<uint8_t, n>(x[i].qs + 32*il + n*ir);
+    for (int l = 0; l < n; ++l) {
+        y[l + 0] = d1 * (q_vec[l] & 0xF) - m1;
+        y[l +32] = d2 * (q_vec[l] >>  4) - m2;
+    }
+#else
+    const int64_t tid = item_ct1.get_local_id(2);
+    const uint8_t * q = x[i].qs;
+    dst_t * y = yy + i*QK_K;
+    const float d = (float)x[i].dm[0];
+    const float m = (float)x[i].dm[1];
+    y[tid+ 0] = d * (x[i].scales[0] & 0xF) * (q[tid] & 0xF) - m * (x[i].scales[0] >> 4);
+    y[tid+32] = d * (x[i].scales[1] & 0xF) * (q[tid] >>  4) - m * (x[i].scales[1] >> 4);
+#endif
+}
+
+template<typename dst_t>
+static void dequantize_block_q5_K(const void * __restrict__ vx, dst_t * __restrict__ yy,
+                                  const sycl::nd_item<3> &item_ct1) {
+    const block_q5_K * x = (const block_q5_K *) vx;
+
+    const int64_t i = item_ct1.get_group(2);
+
+#if QK_K == 256
+    // assume 64 threads - this is very slightly better than the one below
+    const int64_t tid = item_ct1.get_local_id(2);
+    const int64_t il  = tid/16;   // il is in 0...3
+    const int64_t ir  = tid%16;   // ir is in 0...15
+    const int64_t is  = 2*il;     // is is in 0...6
+
+    dst_t * y = yy + i*QK_K + 64*il + 2*ir;
+
+    const float dall = x[i].dm[0];
+    const float dmin = x[i].dm[1];
+
+    const uint8_t * ql = x[i].qs + 32*il + 2*ir;
+    const uint8_t * qh = x[i].qh + 2*ir;
+
+    uint8_t sc, m;
+    get_scale_min_k4(is + 0, x[i].scales, sc, m);
+    const float d1 = dall * sc; const float m1 = dmin * m;
+    get_scale_min_k4(is + 1, x[i].scales, sc, m);
+    const float d2 = dall * sc; const float m2 = dmin * m;
+
+    uint8_t   hm  = 1 << (2*il);
+    y[ 0] = d1 * ((ql[ 0] & 0xF) + (qh[ 0] & hm ? 16 : 0)) - m1;
+    y[ 1] = d1 * ((ql[ 1] & 0xF) + (qh[ 1] & hm ? 16 : 0)) - m1;
+    hm <<= 1;
+    y[32] = d2 * ((ql[ 0] >>  4) + (qh[ 0] & hm ? 16 : 0)) - m2;
+    y[33] = d2 * ((ql[ 1] >>  4) + (qh[ 1] & hm ? 16 : 0)) - m2;
+#else
+    const int64_t tid = item_ct1.get_local_id(2);
+    const uint8_t q = x[i].qs[tid];
+    const int64_t im = tid/8;  // 0...3
+    const int64_t in = tid%8;  // 0...7
+    const int64_t is = tid/16; // 0 or 1
+    const uint8_t h = x[i].qh[in] >> im;
+    const float d = x[i].d;
+    dst_t * y = yy + i*QK_K + tid;
+    y[ 0] = d * x[i].scales[is+0] * ((q & 0xF) - ((h >> 0) & 1 ? 0 : 16));
+    y[32] = d * x[i].scales[is+2] * ((q >>  4) - ((h >> 4) & 1 ? 0 : 16));
+#endif
+}
+
+template<typename dst_t>
+static void dequantize_block_q6_K(const void * __restrict__ vx, dst_t * __restrict__ yy,
+                                  const sycl::nd_item<3> &item_ct1) {
+    const block_q6_K * x = (const block_q6_K *) vx;
+
+    const int64_t i = item_ct1.get_group(2);
+#if QK_K == 256
+
+    // assume 64 threads - this is very slightly better than the one below
+    const int64_t tid = item_ct1.get_local_id(2);
+    const int64_t ip  = tid/32;   // ip is 0 or 1
+    const int64_t il  = tid - 32*ip; // 0...32
+    const int64_t is  = 8*ip + il/16;
+
+    dst_t * y = yy + i*QK_K + 128*ip + il;
+
+    const float d = x[i].d;
+
+    const uint8_t * ql = x[i].ql + 64*ip + il;
+    const uint8_t   qh = x[i].qh[32*ip + il];
+    const int8_t  * sc = x[i].scales + is;
+
+    y[ 0] = d * sc[0] * ((int8_t)((ql[ 0] & 0xF) | (((qh >> 0) & 3) << 4)) - 32);
+    y[32] = d * sc[2] * ((int8_t)((ql[32] & 0xF) | (((qh >> 2) & 3) << 4)) - 32);
+    y[64] = d * sc[4] * ((int8_t)((ql[ 0]  >> 4) | (((qh >> 4) & 3) << 4)) - 32);
+    y[96] = d * sc[6] * ((int8_t)((ql[32]  >> 4) | (((qh >> 6) & 3) << 4)) - 32);
+#else
+
+    // assume 32 threads
+    const int64_t tid = item_ct1.get_local_id(2);
+    const int64_t ip  = tid/16;         // 0 or 1
+    const int64_t il  = tid - 16*ip;    // 0...15
+
+    dst_t * y = yy + i*QK_K + 16*ip + il;
+
+    const float d = x[i].d;
+
+    const uint8_t   ql = x[i].ql[16*ip + il];
+    const uint8_t   qh = x[i].qh[il] >> (2*ip);
+    const int8_t  * sc = x[i].scales;
+
+    y[ 0] = d * sc[ip+0] * ((int8_t)((ql & 0xF) | (((qh >> 0) & 3) << 4)) - 32);
+    y[32] = d * sc[ip+2] * ((int8_t)((ql  >> 4) | (((qh >> 4) & 3) << 4)) - 32);
+#endif
+}
+
+template<typename dst_t>
+static void dequantize_block_iq2_xxs(const void * __restrict__ vx, dst_t * __restrict__ yy,
+                                     const sycl::nd_item<3> &item_ct1,
+                                     const uint64_t *iq2xxs_grid_ptr,
+                                     const uint8_t *ksigns_iq2xs_ptr,
+                                     const uint8_t *kmask_iq2xs_ptr) {
+
+    const int64_t i = item_ct1.get_group(2);
+    const block_iq2_xxs * x = (const block_iq2_xxs  *) vx;
+
+    const int64_t tid = item_ct1.get_local_id(2);
+#if QK_K == 256
+    const int64_t il = tid/8; // 0...3
+    const int64_t ib = tid%8; // 0...7
+    dst_t * y = yy + i*QK_K + 32*ib + 8*il;
+    const uint16_t * q2 = x[i].qs + 4*ib;
+    const uint8_t  * aux8 = (const uint8_t *)q2;
+    const uint8_t  * grid = (const uint8_t *)(iq2xxs_grid_ptr + aux8[il]);
+    const uint32_t aux32 = q2[2] | (q2[3] << 16);
+    const float d = (float)x[i].d * (0.5f + (aux32 >> 28)) * 0.25f;
+    const uint8_t signs = ksigns_iq2xs_ptr[(aux32 >> 7*il) & 127];
+    for (int j = 0; j < 8; ++j) y[j] = d * grid[j] * (signs & kmask_iq2xs_ptr[j] ? -1.f : 1.f);
+#else
+    assert(false);
+#endif
+
+}
+
+template<typename dst_t>
+static void dequantize_block_iq2_xs(const void * __restrict__ vx, dst_t * __restrict__ yy,
+                                    const sycl::nd_item<3> &item_ct1,
+                                    const uint64_t *iq2xs_grid,
+                                    const uint8_t *ksigns_iq2xs,
+                                    const uint8_t *kmask_iq2xs) {
+
+    const int64_t i = item_ct1.get_group(2);
+    const block_iq2_xs * x = (const block_iq2_xs *) vx;
+
+    const int64_t tid = item_ct1.get_local_id(2);
+#if QK_K == 256
+    const int64_t il = tid/8; // 0...3
+    const int64_t ib = tid%8; // 0...7
+    dst_t * y = yy + i*QK_K + 32*ib + 8*il;
+    const uint16_t * q2 = x[i].qs + 4*ib;
+    const uint8_t  * grid = (const uint8_t *)(iq2xs_grid + (q2[il] & 511));
+    const float d = (float)x[i].d * (0.5f + ((x[i].scales[ib] >> 4*(il/2)) & 0xf)) * 0.25f;
+    const uint8_t signs = ksigns_iq2xs[q2[il] >> 9];
+    for (int j = 0; j < 8; ++j) y[j] = d * grid[j] * (signs & kmask_iq2xs[j] ? -1.f : 1.f);
+#else
+    assert(false);
+#endif
+
+}
+
+template <typename dst_t>
+__dpct_inline__ static void
+dequantize_block_iq2_s(const void *__restrict__ vx, dst_t *__restrict__ yy,
+                       const sycl::nd_item<3> &item_ct1) {
+
+    const int64_t i = item_ct1.get_group(2);
+    const block_iq2_s * x = (const block_iq2_s *) vx;
+
+    const int64_t tid = item_ct1.get_local_id(2);
+#if QK_K == 256
+    const int64_t il = tid/8; // 0...3
+    const int64_t ib = tid%8; // 0...7
+    dst_t * y = yy + i*QK_K + 32*ib + 8*il;
+    const uint8_t * grid = (const uint8_t *)(iq2s_grid + (x[i].qs[4*ib+il] | ((x[i].qh[ib] << (8-2*il)) & 0x300)));
+    const float d = (float)x[i].d * (0.5f + ((x[i].scales[ib] >> 4*(il/2)) & 0xf)) * 0.25f;
+    const uint8_t signs = x[i].qs[QK_K/8+4*ib+il];
+#pragma unroll
+    for (int j = 0; j < 8; ++j)
+        y[j] = d * grid[j] * (signs & kmask_iq2xs[j] ? -1.f : 1.f);
+#else
+    assert(false);
+
+#endif
+
+}
+
+template<typename dst_t>
+static void dequantize_block_iq3_xxs(const void * __restrict__ vx, dst_t * __restrict__ yy,
+                                     const sycl::nd_item<3> &item_ct1,
+                                     const uint32_t *iq3xxs_grid,
+                                     const uint8_t *ksigns_iq2xs,
+                                     const uint8_t *kmask_iq2xs) {
+
+    const int64_t i = item_ct1.get_group(2);
+    const block_iq3_xxs * x = (const block_iq3_xxs  *) vx;
+
+    const int64_t tid = item_ct1.get_local_id(2);
+#if QK_K == 256
+    const int64_t il = tid/8; // 0...3
+    const int64_t ib = tid%8; // 0...7
+    dst_t * y = yy + i*QK_K + 32*ib + 8*il;
+    const uint8_t  * q3 = x[i].qs + 8*ib;
+    const uint16_t * gas = (const uint16_t *)(x[i].qs + QK_K/4) + 2*ib;
+    const uint8_t  * grid1 = (const uint8_t *)(iq3xxs_grid + q3[2*il+0]);
+    const uint8_t  * grid2 = (const uint8_t *)(iq3xxs_grid + q3[2*il+1]);
+    const uint32_t aux32 = gas[0] | (gas[1] << 16);
+    const float d = (float)x[i].d * (0.5f + (aux32 >> 28)) * 0.5f;
+    const uint8_t signs = ksigns_iq2xs[(aux32 >> 7*il) & 127];
+    for (int j = 0; j < 4; ++j) {
+        y[j+0] = d * grid1[j] * (signs & kmask_iq2xs[j+0] ? -1.f : 1.f);
+        y[j+4] = d * grid2[j] * (signs & kmask_iq2xs[j+4] ? -1.f : 1.f);
+    }
+#else
+    assert(false);
+#endif
+
+}
+
+template <typename dst_t>
+__dpct_inline__ static void
+dequantize_block_iq3_s(const void *__restrict__ vx, dst_t *__restrict__ yy,
+                       const sycl::nd_item<3> &item_ct1,
+                       const uint8_t *kmask_iq2xs, const uint32_t *iq3s_grid) {
+
+    const int64_t i = item_ct1.get_group(2);
+    const block_iq3_s * x = (const block_iq3_s *) vx;
+
+    const int64_t tid = item_ct1.get_local_id(2);
+#if QK_K == 256
+    const int64_t il = tid/8; // 0...3
+    const int64_t ib = tid%8; // 0...7
+    dst_t * y = yy + i*QK_K + 32*ib + 8*il;
+    const uint8_t * qs = x[i].qs + 8*ib;
+    const uint8_t * grid1 = (const uint8_t *)(iq3s_grid + (qs[2*il+0] | ((x[i].qh[ib] << (8-2*il)) & 256)));
+    const uint8_t * grid2 = (const uint8_t *)(iq3s_grid + (qs[2*il+1] | ((x[i].qh[ib] << (7-2*il)) & 256)));
+    const float d = (float)x[i].d * (1 + 2*((x[i].scales[ib/2] >> 4*(ib%2)) & 0xf));
+    const uint8_t signs = x[i].signs[4*ib + il];
+#pragma unroll
+    for (int j = 0; j < 4; ++j) {
+        y[j+0] = d * grid1[j] * (signs & kmask_iq2xs[j+0] ? -1.f : 1.f);
+        y[j+4] = d * grid2[j] * (signs & kmask_iq2xs[j+4] ? -1.f : 1.f);
+    }
+#else
+    assert(false);
+#endif
+
+}
+
+template <typename dst_t>
+__dpct_inline__ static void
+dequantize_block_iq1_s(const void *__restrict__ vx, dst_t *__restrict__ yy,
+                       const sycl::nd_item<3> &item_ct1,
+                       const uint32_t *iq1s_grid_gpu) {
+
+    const int64_t i = item_ct1.get_group(2);
+    const block_iq1_s * x = (const block_iq1_s  *) vx;
+
+    const int64_t tid = item_ct1.get_local_id(2);
+#if QK_K == 256
+    const int64_t il = tid/8; // 0...3
+    const int64_t ib = tid%8; // 0...7
+    dst_t * y = yy + i*QK_K + 32*ib + 8*il;
+    const float delta = x[i].qh[ib] & 0x8000 ? -1 - IQ1S_DELTA : -1 + IQ1S_DELTA;
+    const float d = (float)x[i].d * (2*((x[i].qh[ib] >> 12) & 7) + 1);
+    uint32_t grid32[2]; const int8_t * q = (const int8_t *)grid32;
+    grid32[0] = iq1s_grid_gpu[x[i].qs[4*ib+il] | (((x[i].qh[ib] >> 3*il) & 7) << 8)];
+    grid32[1] = (grid32[0] >> 4) & 0x0f0f0f0f;
+    grid32[0] &= 0x0f0f0f0f;
+#pragma unroll
+    for (int j = 0; j < 8; ++j) {
+        y[j] = d * (q[j] + delta);
+    }
+#else
+    assert(false);
+#endif
+
+}
+
+template <typename dst_t>
+__dpct_inline__ static void
+dequantize_block_iq1_m(const void *__restrict__ vx, dst_t *__restrict__ yy,
+                       const sycl::nd_item<3> &item_ct1,
+                       const uint32_t *iq1s_grid_gpu) {
+
+    const int64_t i = item_ct1.get_group(2);
+    const block_iq1_m * x = (const block_iq1_m  *) vx;
+
+    const int64_t tid = item_ct1.get_local_id(2);
+#if QK_K == 256
+    const int64_t il = tid/8; // 0...3
+    const int64_t ib = tid%8; // 0...7
+    dst_t * y = yy + i*QK_K + 32*ib + 8*il;
+    const uint16_t * sc = (const uint16_t *)x[i].scales;
+    iq1m_scale_t scale;
+    scale.u16 = (sc[0] >> 12) | ((sc[1] >> 8) & 0x00f0) | ((sc[2] >> 4) & 0x0f00) | (sc[3] & 0xf000);
+    const int ib16 = 2*ib + il/2; // sc[ib16/4] >> 3*(ib16%4) -> sc[ib/2] >> 3*((2*ib+il/2)%4);
+    const float d = (float)scale.f16 * (2*((sc[ib16/4] >> 3*(ib16%4)) & 0x7) + 1);
+    const float delta = x[i].qh[2*ib+il/2] & (0x08 << 4*(il%2)) ? -1 - IQ1M_DELTA : -1 + IQ1M_DELTA;
+    uint32_t grid32[2]; const int8_t * q = (const int8_t *)grid32;
+    grid32[0] = iq1s_grid_gpu[x[i].qs[4*ib+il] | (((x[i].qh[2*ib+il/2] >> 4*(il%2)) & 7) << 8)];
+    grid32[1] = (grid32[0] >> 4) & 0x0f0f0f0f;
+    grid32[0] &= 0x0f0f0f0f;
+#pragma unroll
+    for (int j = 0; j < 8; ++j) {
+        y[j] = d * (q[j] + delta);
+    }
+#else
+    assert(false);
+#endif
+
+}
+
+template <typename dst_t>
+__dpct_inline__ static void
+dequantize_block_iq4_nl(const void *__restrict__ vx, dst_t *__restrict__ yy,
+                        const sycl::nd_item<3> &item_ct1) {
+
+    const int64_t i = item_ct1.get_group(2);
+    const block_iq4_nl * x = (const block_iq4_nl *) vx + i*(QK_K/QK4_NL);
+
+    const int64_t tid = item_ct1.get_local_id(2);
+    const int64_t il = tid/8; // 0...3
+    const int64_t ib = tid%8; // 0...7
+    dst_t * y = yy + i*QK_K + 32*ib + 4*il;
+    const uint8_t  * q4 = x[ib].qs + 4*il;
+    const float d = (float)x[ib].d;
+#pragma unroll
+    for (int j = 0; j < 4; ++j) {
+        y[j+ 0] = d * kvalues_iq4nl[q4[j] & 0xf];
+        y[j+16] = d * kvalues_iq4nl[q4[j] >>  4];
+    }
+
+}
+
+
+template <typename dst_t>
+__dpct_inline__ static void
+dequantize_block_iq4_xs(const void *__restrict__ vx, dst_t *__restrict__ yy,
+                        const sycl::nd_item<3> &item_ct1) {
+    const int64_t i = item_ct1.get_group(2);
+    const block_iq4_xs * x = (const block_iq4_xs *)vx;
+
+    const int64_t tid = item_ct1.get_local_id(2);
+    const int64_t il = tid/8; // 0...3
+    const int64_t ib = tid%8; // 0...7
+    dst_t * y = yy + i*QK_K + 32*ib + 4*il;
+    const uint8_t  * q4 = x[i].qs + 16*ib + 4*il;
+    const float d = (float)x[i].d * ((((x[i].scales_l[ib/2] >> 4*(ib%2)) & 0xf) | (((x[i].scales_h >> 2*ib) & 3) << 4)) - 32);
+#pragma unroll
+    for (int j = 0; j < 4; ++j) {
+        y[j+ 0] = d * kvalues_iq4nl[q4[j] & 0xf];
+        y[j+16] = d * kvalues_iq4nl[q4[j] >>  4];
+    }
+}
+
+
+#endif // GGML_SYCL_DEQUANTIZE_HPP
diff --git a/src/whisper_cpp/ggml/src/ggml-sycl/dmmv.cpp b/src/whisper_cpp/ggml/src/ggml-sycl/dmmv.cpp
new file mode 100644
index 00000000..0c3dfaa3
--- /dev/null
+++ b/src/whisper_cpp/ggml/src/ggml-sycl/dmmv.cpp
@@ -0,0 +1,1023 @@
+#include "convert.hpp"
+#include "dmmv.hpp"
+#include "dequantize.hpp"
+#include "presets.hpp"
+
+
+static void convert_f16(const void * vx, const int64_t ib, const int iqs, dfloat2 & v){
+    const sycl::half *x = (const sycl::half *)vx;
+
+    // automatic half -> float type cast if dfloat == float
+    v.x() = x[ib + iqs + 0];
+    v.y() = x[ib + iqs + 1];
+}
+
+static void convert_f32(const void * vx, const int64_t ib, const int iqs, dfloat2 & v){
+    const float * x = (const float *) vx;
+
+    // automatic half -> float type cast if dfloat == float
+    v.x() = x[ib + iqs + 0];
+    v.y() = x[ib + iqs + 1];
+}
+
+template <int qk, int qr, dequantize_kernel_t dequantize_kernel>
+static void dequantize_mul_mat_vec(const void * __restrict__ vx, const dfloat * __restrict__ y, float * __restrict__ dst, const int ncols, const int nrows,
+                                   const sycl::nd_item<3> &item_ct1) {
+    // qk = quantized weights per x block
+    // qr = number of quantized weights per data value in x block
+    const int row = item_ct1.get_group(2) * item_ct1.get_local_range(1) +
+                    item_ct1.get_local_id(1);
+
+    if (row >= nrows) {
+        return;
+    }
+
+    const int tid = item_ct1.get_local_id(2);
+
+    const int iter_stride = 2*GGML_SYCL_DMMV_X;
+    const int vals_per_iter = iter_stride / WARP_SIZE; // num quantized vals per thread and i iter
+    const int y_offset = qr == 1 ? 1 : qk/2;
+
+// partial sum for each thread
+#ifdef GGML_SYCL_F16
+    sycl::half2 tmp = {0.0f, 0.0f}; // two sums for f16 to take advantage of half2 intrinsics
+#else
+    float tmp = 0.0f;
+#endif // GGML_SYCL_F16
+
+    for (int i = 0; i < ncols; i += iter_stride) {
+        const int col = i + vals_per_iter*tid;
+        const int ib = (row*ncols + col)/qk; // x block index
+        const int iqs = (col%qk)/qr; // x quant index
+        const int iybs = col - col%qk; // y block start index
+
+// processing >2 values per i iter is faster for fast GPUs
+#pragma unroll
+        for (int j = 0; j < vals_per_iter; j += 2) {
+            // process 2 vals per j iter
+
+            // dequantize
+            // for qr = 2 the iqs needs to increase by 1 per j iter because 2 weights per data val
+            dfloat2 v;
+            dequantize_kernel(vx, ib, iqs + j/qr, v);
+
+            // matrix multiplication
+            // for qr = 2 the y index needs to increase by 1 per j iter because of y_offset = qk/2
+#ifdef GGML_SYCL_F16
+            dfloat2 t1{y[iybs + iqs + j / qr + 0],
+                        y[iybs + iqs + j / qr + y_offset]};
+
+            tmp += v * t1;
+#else
+            tmp += v.x() * y[iybs + iqs + j / qr + 0];
+            tmp += v.y() * y[iybs + iqs + j / qr + y_offset];
+#endif // GGML_SYCL_F16
+        }
+    }
+
+    // sum up partial sums and write back result
+    const int mask_start = ncols > GGML_SYCL_DMMV_X ? WARP_SIZE >> 1 : WARP_SIZE >> 2;
+    for (int mask = mask_start; mask > 0; mask >>= 1) {
+        tmp +=
+            dpct::permute_sub_group_by_xor(item_ct1.get_sub_group(), tmp, mask);
+    }
+
+    if (tid == 0) {
+#ifdef GGML_SYCL_F16
+        dst[row] = tmp.x() + tmp.y();
+#else
+        dst[row] = tmp;
+#endif // GGML_SYCL_F16
+    }
+}
+
+static void convert_mul_mat_vec_f16_sycl(const void *vx, const dfloat *y,
+                                         float *dst, const int ncols,
+                                         const int nrows,
+                                         dpct::queue_ptr stream) {
+    GGML_ASSERT(ncols % GGML_SYCL_DMMV_X == 0);
+    const int block_num_y = (nrows + GGML_SYCL_MMV_Y - 1) / GGML_SYCL_MMV_Y;
+    const sycl::range<3> block_nums(1, 1, block_num_y);
+    const sycl::range<3> block_dims(1, GGML_SYCL_MMV_Y, WARP_SIZE);
+    {
+        dpct::has_capability_or_fail(stream->get_device(),
+                                     {sycl::aspect::fp16});
+
+        stream->parallel_for(
+            sycl::nd_range<3>(block_nums * block_dims, block_dims),
+            [=](sycl::nd_item<3> item_ct1) [[intel::reqd_sub_group_size(WARP_SIZE)]] {
+                dequantize_mul_mat_vec<1, 1, convert_f16>(vx, y, dst, ncols,
+                                                          nrows, item_ct1);
+            });
+    }
+}
+
+/*
+DPCT1110:4: The total declared local variable size in device function
+dequantize_mul_mat_vec_q2_k exceeds 128 bytes and may cause high register
+pressure. Consult with your hardware vendor to find the total register size
+available and adjust the code, or use smaller sub-group size to avoid high
+register pressure.
+*/
+static void dequantize_mul_mat_vec_q2_k(const void *__restrict__ vx,
+                                        const float *__restrict__ yy,
+                                        float *__restrict__ dst,
+                                        const int ncols, int nrows,
+                                        const sycl::nd_item<3> &item_ct1) {
+
+    static_assert(16%K_QUANTS_PER_ITERATION == 0, "16 must be divisible by K_QUANTS_PER_ITERATION");
+
+    const int row = item_ct1.get_group(2) * item_ct1.get_local_range(1) +
+                    item_ct1.get_local_id(1);
+    if (row > nrows) return;
+
+    const int num_blocks_per_row = ncols / QK_K;
+    const int ib0 = row*num_blocks_per_row;
+
+    const block_q2_K * x = (const block_q2_K *)vx + ib0;
+
+    float tmp = 0; // partial sum for thread in warp
+
+#if QK_K == 256
+    const int tid =
+        item_ct1.get_local_id(2) / K_QUANTS_PER_ITERATION; // 0...31 or 0...15
+    const int ix =
+        item_ct1.get_local_id(2) % K_QUANTS_PER_ITERATION; // 0 or 0,1
+
+    const int step = 16/K_QUANTS_PER_ITERATION;
+
+    const int im = tid/step;                             // 0 or 1. 0 computes 0..., 1 computes 128...
+    const int in = tid - step*im;                        // 0...15 or 0...7
+
+    const int l0 = K_QUANTS_PER_ITERATION*in;            // 0...15 or 0...14 in steps of 2
+    const int q_offset = 32*im + l0;
+    const int s_offset = 8*im;
+    const int y_offset = 128*im + l0;
+
+    uint32_t aux[4];
+    const uint8_t * d = (const uint8_t *)aux;
+    const uint8_t * m = (const uint8_t *)(aux + 2);
+
+    for (int i = ix; i < num_blocks_per_row; i += K_QUANTS_PER_ITERATION) {
+
+        const float   * y = yy + i * QK_K + y_offset;
+        const uint8_t * q = x[i].qs + q_offset;
+
+        const float dall = x[i].dm[0];
+        const float dmin = x[i].dm[1];
+
+        const uint32_t * a = (const uint32_t *)(x[i].scales + s_offset);
+        aux[0] = a[0] & 0x0f0f0f0f;
+        aux[1] = a[1] & 0x0f0f0f0f;
+        aux[2] = (a[0] >> 4) & 0x0f0f0f0f;
+        aux[3] = (a[1] >> 4) & 0x0f0f0f0f;
+
+        float sum1 = 0, sum2 = 0;
+        for (int l = 0; l < K_QUANTS_PER_ITERATION; ++l) {
+            sum1 += y[l+ 0] * d[0] * ((q[l+ 0] >> 0) & 3)
+                  + y[l+32] * d[2] * ((q[l+ 0] >> 2) & 3)
+                  + y[l+64] * d[4] * ((q[l+ 0] >> 4) & 3)
+                  + y[l+96] * d[6] * ((q[l+ 0] >> 6) & 3)
+                  + y[l+16] * d[1] * ((q[l+16] >> 0) & 3)
+                  + y[l+48] * d[3] * ((q[l+16] >> 2) & 3)
+                  + y[l+80] * d[5] * ((q[l+16] >> 4) & 3)
+                  +y[l+112] * d[7] * ((q[l+16] >> 6) & 3);
+            sum2 += y[l+ 0] * m[0] + y[l+32] * m[2] + y[l+64] * m[4] + y[ l+96] * m[6]
+                  + y[l+16] * m[1] + y[l+48] * m[3] + y[l+80] * m[5] + y[l+112] * m[7];
+
+        }
+        tmp += dall * sum1 - dmin * sum2;
+
+    }
+#else
+    const int tid = item_ct1.get_local_id(2) /
+                    (2 * K_QUANTS_PER_ITERATION); // 0...15 or 0...7
+    const int ix = item_ct1.get_local_id(2) %
+                   (2 * K_QUANTS_PER_ITERATION); // 0....1 or 0...3
+    const int offset = tid * K_QUANTS_PER_ITERATION;
+
+    uint32_t uaux[2];
+    const uint8_t * d = (const uint8_t *)uaux;
+
+
+    for (int i = ix; i < num_blocks_per_row; i += 2*K_QUANTS_PER_ITERATION) {
+
+        const float   * y = yy + i * QK_K + offset;
+        const uint8_t * q = x[i].qs + offset;
+        const uint32_t * s = (const uint32_t *)x[i].scales;
+
+        uaux[0] = s[0] & 0x0f0f0f0f;
+        uaux[1] = (s[0] >> 4) & 0x0f0f0f0f;
+
+        const sycl::float2 dall =
+            x[i].dm.convert<float, sycl::rounding_mode::automatic>();
+
+        float sum1 = 0, sum2 = 0;
+        for (int l = 0; l < K_QUANTS_PER_ITERATION; ++l) {
+            const uint8_t ql = q[l];
+            sum1 += y[l+ 0] * d[0] * ((ql >> 0) & 3)
+                  + y[l+16] * d[1] * ((ql >> 2) & 3)
+                  + y[l+32] * d[2] * ((ql >> 4) & 3)
+                  + y[l+48] * d[3] * ((ql >> 6) & 3);
+            sum2 += y[l+0] * d[4] + y[l+16] * d[5] + y[l+32] * d[6] + y[l+48] * d[7];
+        }
+        tmp += dall.x() * sum1 - dall.y() * sum2;
+    }
+
+#endif
+
+    // sum up partial sums and write back result
+#pragma unroll
+    for (int mask = QK_WARP_SIZE / 2; mask > 0; mask >>= 1) {
+        tmp +=
+            dpct::permute_sub_group_by_xor(item_ct1.get_sub_group(), tmp, mask);
+    }
+
+    if (item_ct1.get_local_id(2) == 0) {
+        dst[row] = tmp;
+    }
+}
+
+/*
+DPCT1110:5: The total declared local variable size in device function
+dequantize_mul_mat_vec_q3_k exceeds 128 bytes and may cause high register
+pressure. Consult with your hardware vendor to find the total register size
+available and adjust the code, or use smaller sub-group size to avoid high
+register pressure.
+*/
+static void dequantize_mul_mat_vec_q3_k(const void *__restrict__ vx,
+                                        const float *__restrict__ yy,
+                                        float *__restrict__ dst,
+                                        const int ncols, int nrows,
+                                        const sycl::nd_item<3> &item_ct1) {
+
+    const int row = item_ct1.get_group(2) * item_ct1.get_local_range(1) +
+                    item_ct1.get_local_id(1);
+    if (row > nrows) return;
+
+    const int num_blocks_per_row = ncols / QK_K;
+    const int ib0 = row*num_blocks_per_row;
+
+    const block_q3_K * x = (const block_q3_K *)vx + ib0;
+
+    float tmp = 0; // partial sum for thread in warp
+
+#if QK_K == 256
+
+    const uint16_t kmask1 = 0x0303;
+    const uint16_t kmask2 = 0x0f0f;
+
+    const int tid =
+        item_ct1.get_local_id(2) / K_QUANTS_PER_ITERATION; // 0...31 or 0...16
+    const int ix =
+        item_ct1.get_local_id(2) % K_QUANTS_PER_ITERATION; // 0 or 0,1
+
+    const int n  = K_QUANTS_PER_ITERATION;               // iterations in the inner loop
+    const int step = 16/K_QUANTS_PER_ITERATION;
+    const int im = tid/step;                             // 0 or 1. 0 computes 0..., 1 computes 128...
+    const int in = tid - step*im;                        // 0....15 or 0...7
+
+    const uint8_t m = 1 << (4*im);
+
+    const int l0 = n*in;                                 // 0...15 or 0...14 in steps of 2
+    const int q_offset =  32*im + l0;
+    const int y_offset = 128*im + l0;
+
+    uint16_t utmp[4];
+    const int8_t * s = (const int8_t *)utmp;
+
+    const uint16_t s_shift = 4*im;
+
+    for (int i = ix; i < num_blocks_per_row; i += K_QUANTS_PER_ITERATION) {
+
+        const float   * y  = yy + i * QK_K + y_offset;
+        const uint8_t * q = x[i].qs + q_offset;
+        const uint8_t * h = x[i].hmask + l0;
+
+        const uint16_t * a = (const uint16_t *)x[i].scales;
+        utmp[0] = ((a[0] >> s_shift) & kmask2) | (((a[4] >> (s_shift + 0)) & kmask1) << 4);
+        utmp[1] = ((a[1] >> s_shift) & kmask2) | (((a[5] >> (s_shift + 0)) & kmask1) << 4);
+        utmp[2] = ((a[2] >> s_shift) & kmask2) | (((a[4] >> (s_shift + 2)) & kmask1) << 4);
+        utmp[3] = ((a[3] >> s_shift) & kmask2) | (((a[5] >> (s_shift + 2)) & kmask1) << 4);
+
+        const float d = x[i].d;
+
+        float sum = 0;
+        for (int l = 0; l < n; ++l) {
+            sum += y[l+ 0] * (s[0] - 32) * (((q[l] >> 0) & 3) - (h[l] & (m << 0) ? 0 : 4))
+                 + y[l+32] * (s[2] - 32) * (((q[l] >> 2) & 3) - (h[l] & (m << 1) ? 0 : 4))
+                 + y[l+64] * (s[4] - 32) * (((q[l] >> 4) & 3) - (h[l] & (m << 2) ? 0 : 4))
+                 + y[l+96] * (s[6] - 32) * (((q[l] >> 6) & 3) - (h[l] & (m << 3) ? 0 : 4));
+            sum += y[l+16] * (s[1] - 32) * (((q[l+16] >> 0) & 3) - (h[l+16] & (m << 0) ? 0 : 4))
+                 + y[l+48] * (s[3] - 32) * (((q[l+16] >> 2) & 3) - (h[l+16] & (m << 1) ? 0 : 4))
+                 + y[l+80] * (s[5] - 32) * (((q[l+16] >> 4) & 3) - (h[l+16] & (m << 2) ? 0 : 4))
+                + y[l+112] * (s[7] - 32) * (((q[l+16] >> 6) & 3) - (h[l+16] & (m << 3) ? 0 : 4));
+        }
+        tmp += d * sum;
+
+    }
+#else
+
+    const int tid = item_ct1.get_local_id(2)/(2*K_QUANTS_PER_ITERATION);  // 0...15 or 0...7
+    const int ix  = item_ct1.get_local_id(2)%(2*K_QUANTS_PER_ITERATION);  // 0....1 or 0...3
+    const int offset = tid * K_QUANTS_PER_ITERATION;         // 0...15 or 0...14
+    const int in = offset/8;                                 // 0 or 1
+    const int im = offset%8;                                 // 0...7
+
+    for (int i = ix; i < num_blocks_per_row; i += 2*K_QUANTS_PER_ITERATION) {
+
+        const float   * y = yy + i * QK_K + offset;
+        const uint8_t * q = x[i].qs + offset;
+        const uint8_t * s = x[i].scales;
+
+        const float dall = (float)x[i].d;
+
+        float sum = 0;
+        for (int l = 0; l < K_QUANTS_PER_ITERATION; ++l) {
+            const uint8_t hl = x[i].hmask[im+l] >> in;
+            const uint8_t ql = q[l];
+            sum += y[l+ 0] * dall * ((s[0] & 0xF) - 8) * ((int8_t)((ql >> 0) & 3) - ((hl >> 0) & 1 ? 0 : 4))
+                 + y[l+16] * dall * ((s[0] >>  4) - 8) * ((int8_t)((ql >> 2) & 3) - ((hl >> 2) & 1 ? 0 : 4))
+                 + y[l+32] * dall * ((s[1] & 0xF) - 8) * ((int8_t)((ql >> 4) & 3) - ((hl >> 4) & 1 ? 0 : 4))
+                 + y[l+48] * dall * ((s[1] >>  4) - 8) * ((int8_t)((ql >> 6) & 3) - ((hl >> 6) & 1 ? 0 : 4));
+        }
+        tmp += sum;
+    }
+#endif
+
+    // sum up partial sums and write back result
+#pragma unroll
+    for (int mask = QK_WARP_SIZE / 2; mask > 0; mask >>= 1) {
+        tmp +=
+            dpct::permute_sub_group_by_xor(item_ct1.get_sub_group(), tmp, mask);
+    }
+
+    if (item_ct1.get_local_id(2) == 0) {
+        dst[row] = tmp;
+    }
+}
+
+/*
+DPCT1110:6: The total declared local variable size in device function
+dequantize_mul_mat_vec_q4_k exceeds 128 bytes and may cause high register
+pressure. Consult with your hardware vendor to find the total register size
+available and adjust the code, or use smaller sub-group size to avoid high
+register pressure.
+*/
+static void dequantize_mul_mat_vec_q4_k(const void *__restrict__ vx,
+                                        const float *__restrict__ yy,
+                                        float *__restrict__ dst,
+                                        const int ncols, int nrows,
+                                        const sycl::nd_item<3> &item_ct1) {
+
+    const int row = item_ct1.get_group(2) * item_ct1.get_local_range(1) +
+                    item_ct1.get_local_id(1);
+    if (row > nrows) return;
+    const int num_blocks_per_row = ncols / QK_K;
+    const int ib0 = row*num_blocks_per_row;
+
+    const block_q4_K * x = (const block_q4_K *)vx + ib0;
+
+#if QK_K == 256
+    const uint16_t kmask1 = 0x3f3f;
+    const uint16_t kmask2 = 0x0f0f;
+    const uint16_t kmask3 = 0xc0c0;
+
+    const int tid =
+        item_ct1.get_local_id(2) / K_QUANTS_PER_ITERATION; // 0...31 or 0...16
+    const int ix =
+        item_ct1.get_local_id(2) % K_QUANTS_PER_ITERATION; // 0 or 0,1
+
+    const int step = 8/K_QUANTS_PER_ITERATION;           // 8 or 4
+
+    const int il  = tid/step;                            // 0...3
+    const int ir  = tid - step*il;                       // 0...7 or 0...3
+    const int n   = 2 * K_QUANTS_PER_ITERATION;          // 2 or 4
+
+    const int im = il/2;  // 0 or 1. 0 computes 0,32 + 128,160, 1 computes 64,96 + 192,224
+    const int in = il%2;
+
+    const int l0 = n*(2*ir + in);
+    const int q_offset = 32*im + l0;
+    const int y_offset = 64*im + l0;
+
+    uint16_t aux[4];
+    const uint8_t * sc = (const uint8_t *)aux;
+
+#if K_QUANTS_PER_ITERATION == 2
+    uint32_t q32[4];
+    const uint8_t * q4 = (const uint8_t *)q32;
+#else
+    uint16_t q16[4];
+    const uint8_t * q4 = (const uint8_t *)q16;
+#endif
+
+    float tmp = 0; // partial sum for thread in warp
+
+    for (int i = ix; i < num_blocks_per_row; i += K_QUANTS_PER_ITERATION) {
+
+        const float   * y1 = yy + i*QK_K + y_offset;
+        const float   * y2 = y1 + 128;
+
+        const float dall = x[i].dm[0];
+        const float dmin = x[i].dm[1];
+
+        const uint16_t * a = (const uint16_t *)x[i].scales;
+        aux[0] = a[im+0] & kmask1;
+        aux[1] = a[im+2] & kmask1;
+        aux[2] = ((a[im+4] >> 0) & kmask2) | ((a[im+0] & kmask3) >> 2);
+        aux[3] = ((a[im+4] >> 4) & kmask2) | ((a[im+2] & kmask3) >> 2);
+
+#if K_QUANTS_PER_ITERATION == 2
+        const uint32_t * q1 = (const uint32_t *)(x[i].qs + q_offset);
+        const uint32_t * q2 = q1 + 16;
+
+        q32[0] = q1[0] & 0x0f0f0f0f;
+        q32[1] = q1[0] & 0xf0f0f0f0;
+        q32[2] = q2[0] & 0x0f0f0f0f;
+        q32[3] = q2[0] & 0xf0f0f0f0;
+
+        sycl::float4 s = {0.f, 0.f, 0.f, 0.f};
+        float smin = 0;
+        for (int l = 0; l < 4; ++l) {
+            s.x() += y1[l] * q4[l + 0]; s.y() += y1[l + 32] * q4[l + 4];
+            s.z() += y2[l] * q4[l + 8]; s.w() += y2[l + 32] * q4[l + 12];
+            smin += y1[l] * sc[2] + y1[l+32] * sc[3] + y2[l] * sc[6] + y2[l+32] * sc[7];
+        }
+        tmp += dall * (s.x() * sc[0] + s.y() * sc[1] * 1.f / 16.f +
+                       s.z() * sc[4] + s.w() * sc[5] * 1.f / 16.f) -
+               dmin * smin;
+#else
+        const uint16_t * q1 = (const uint16_t *)(x[i].qs + q_offset);
+        const uint16_t * q2 = q1 + 32;
+
+        q16[0] = q1[0] & 0x0f0f;
+        q16[1] = q1[0] & 0xf0f0;
+        q16[2] = q2[0] & 0x0f0f;
+        q16[3] = q2[0] & 0xf0f0;
+
+        float4 s = {0.f, 0.f, 0.f, 0.f};
+        float smin = 0;
+        for (int l = 0; l < 2; ++l) {
+            s.x += y1[l] * q4[l+0]; s.y += y1[l+32] * q4[l+2];
+            s.z += y2[l] * q4[l+4]; s.w += y2[l+32] * q4[l+6];
+            smin += y1[l] * sc[2] + y1[l+32] * sc[3] + y2[l] * sc[6] + y2[l+32] * sc[7];
+        }
+        tmp += dall * (s.x * sc[0] + s.y * sc[1] * 1.f/16.f + s.z * sc[4] + s.w * sc[5] * 1.f/16.f) - dmin * smin;
+#endif
+
+    }
+#else
+    const int tid = item_ct1.get_local_id(2)/(2*K_QUANTS_PER_ITERATION);  // 0...15
+    const int ix  = item_ct1.get_local_id(2)%(2*K_QUANTS_PER_ITERATION);
+
+    const int step = tid * K_QUANTS_PER_ITERATION;
+
+    uint16_t aux16[2];
+    const uint8_t * s = (const uint8_t *)aux16;
+
+    float tmp = 0;
+
+    for (int i = ix; i < num_blocks_per_row; i += 2*K_QUANTS_PER_ITERATION) {
+        const uint8_t * q = x[i].qs + step;
+        const float   * y = yy + i*QK_K + step;
+        const uint16_t * a = (const uint16_t *)x[i].scales;
+        aux16[0] = a[0] & 0x0f0f;
+        aux16[1] = (a[0] >> 4) & 0x0f0f;
+        const float d = (float)x[i].dm[0];
+        const float m = (float)x[i].dm[1];
+        float sum = 0.f;
+        for (int j = 0; j < K_QUANTS_PER_ITERATION; ++j) {
+            sum += y[j+ 0] * (d * s[0] * (q[j+ 0] & 0xF) - m * s[2])
+                 + y[j+16] * (d * s[0] * (q[j+16] & 0xF) - m * s[2])
+                 + y[j+32] * (d * s[1] * (q[j+ 0] >>  4) - m * s[3])
+                 + y[j+48] * (d * s[1] * (q[j+16] >>  4) - m * s[3]);
+        }
+        tmp += sum;
+    }
+
+#endif
+
+    // sum up partial sums and write back result
+#pragma unroll
+    for (int mask = QK_WARP_SIZE / 2; mask > 0; mask >>= 1) {
+        tmp +=
+            dpct::permute_sub_group_by_xor(item_ct1.get_sub_group(), tmp, mask);
+    }
+
+    if (tid == 0) {
+        dst[row] = tmp;
+    }
+}
+
+/*
+DPCT1110:7: The total declared local variable size in device function
+dequantize_mul_mat_vec_q5_k exceeds 128 bytes and may cause high register
+pressure. Consult with your hardware vendor to find the total register size
+available and adjust the code, or use smaller sub-group size to avoid high
+register pressure.
+*/
+static void dequantize_mul_mat_vec_q5_k(const void *__restrict__ vx,
+                                        const float *__restrict__ yy,
+                                        float *__restrict__ dst,
+                                        const int ncols,
+                                        const sycl::nd_item<3> &item_ct1) {
+
+    const int row = item_ct1.get_group(2);
+    const int num_blocks_per_row = ncols / QK_K;
+    const int ib0 = row*num_blocks_per_row;
+
+    const block_q5_K * x = (const block_q5_K *)vx + ib0;
+
+    float tmp = 0; // partial sum for thread in warp
+
+#if QK_K == 256
+    const uint16_t kmask1 = 0x3f3f;
+    const uint16_t kmask2 = 0x0f0f;
+    const uint16_t kmask3 = 0xc0c0;
+
+    const int tid = item_ct1.get_local_id(2) / 2; // 0...15
+    const int ix = item_ct1.get_local_id(2) % 2;
+
+    const int il  = tid/4;     // 0...3
+    const int ir  = tid - 4*il;// 0...3
+    const int n   = 2;
+
+    const int im = il/2;  // 0 or 1. 0 computes 0,32 + 128,160, 1 computes 64,96 + 192,224
+    const int in = il%2;
+
+    const int l0 = n*(2*ir + in);
+    const int q_offset = 32*im + l0;
+    const int y_offset = 64*im + l0;
+
+    const uint8_t hm1  = 1 << (2*im);
+    const uint8_t hm2  = hm1 << 4;
+
+    uint16_t aux[4];
+    const uint8_t * sc = (const uint8_t *)aux;
+
+    uint16_t q16[8];
+    const uint8_t * q4 = (const uint8_t *)q16;
+
+    for (int i = ix; i < num_blocks_per_row; i += 2) {
+
+        const uint8_t * ql1 = x[i].qs + q_offset;
+        const uint8_t * qh  = x[i].qh + l0;
+        const float   * y1  = yy + i*QK_K + y_offset;
+        const float   * y2  = y1 + 128;
+
+        const float dall = x[i].dm[0];
+        const float dmin = x[i].dm[1];
+
+        const uint16_t * a = (const uint16_t *)x[i].scales;
+        aux[0] = a[im+0] & kmask1;
+        aux[1] = a[im+2] & kmask1;
+        aux[2] = ((a[im+4] >> 0) & kmask2) | ((a[im+0] & kmask3) >> 2);
+        aux[3] = ((a[im+4] >> 4) & kmask2) | ((a[im+2] & kmask3) >> 2);
+
+        sycl::float4 sum = {0.f, 0.f, 0.f, 0.f};
+        float smin = 0;
+        const uint16_t * q1 = (const uint16_t *)ql1;
+        const uint16_t * q2 = q1 + 32;
+        q16[0] = q1[0] & 0x0f0f;
+        q16[1] = q1[8] & 0x0f0f;
+        q16[2] = (q1[0] >> 4) & 0x0f0f;
+        q16[3] = (q1[8] >> 4) & 0x0f0f;
+        q16[4] = q2[0] & 0x0f0f;
+        q16[5] = q2[8] & 0x0f0f;
+        q16[6] = (q2[0] >> 4) & 0x0f0f;
+        q16[7] = (q2[8] >> 4) & 0x0f0f;
+        for (int l = 0; l < n; ++l) {
+            sum.x() +=
+                y1[l + 0] * (q4[l + 0] + (qh[l + 0] & (hm1 << 0) ? 16 : 0)) +
+                y1[l + 16] * (q4[l + 2] + (qh[l + 16] & (hm1 << 0) ? 16 : 0));
+            sum.y() +=
+                y1[l + 32] * (q4[l + 4] + (qh[l + 0] & (hm1 << 1) ? 16 : 0)) +
+                y1[l + 48] * (q4[l + 6] + (qh[l + 16] & (hm1 << 1) ? 16 : 0));
+            sum.z() +=
+                y2[l + 0] * (q4[l + 8] + (qh[l + 0] & (hm2 << 0) ? 16 : 0)) +
+                y2[l + 16] * (q4[l + 10] + (qh[l + 16] & (hm2 << 0) ? 16 : 0));
+            sum.w() +=
+                y2[l + 32] * (q4[l + 12] + (qh[l + 0] & (hm2 << 1) ? 16 : 0)) +
+                y2[l + 48] * (q4[l + 14] + (qh[l + 16] & (hm2 << 1) ? 16 : 0));
+            smin += (y1[l] + y1[l+16]) * sc[2] + (y1[l+32] + y1[l+48]) * sc[3]
+                  + (y2[l] + y2[l+16]) * sc[6] + (y2[l+32] + y2[l+48]) * sc[7];
+        }
+        tmp += dall * (sum.x() * sc[0] + sum.y() * sc[1] + sum.z() * sc[4] +
+                       sum.w() * sc[5]) -
+               dmin * smin;
+    }
+
+#else
+    const int tid = item_ct1.get_local_id(2)/(2*K_QUANTS_PER_ITERATION);  // 0...15
+    const int ix  = item_ct1.get_local_id(2)%(2*K_QUANTS_PER_ITERATION);
+    const int step = tid * K_QUANTS_PER_ITERATION;
+    const int im = step/8;
+    const int in = step%8;
+
+    for (int i = ix; i < num_blocks_per_row; i += 2*K_QUANTS_PER_ITERATION) {
+        const uint8_t * q = x[i].qs + step;
+        const int8_t  * s = x[i].scales;
+        const float   * y = yy + i*QK_K + step;
+        const float     d = x[i].d;
+        float sum = 0.f;
+        for (int j = 0; j < K_QUANTS_PER_ITERATION; ++j) {
+            const uint8_t h = x[i].qh[in+j] >> im;
+            sum += y[j+ 0] * d * s[0] * ((q[j+ 0] & 0xF) - ((h >> 0) & 1 ? 0 : 16))
+                 + y[j+16] * d * s[1] * ((q[j+16] & 0xF) - ((h >> 2) & 1 ? 0 : 16))
+                 + y[j+32] * d * s[2] * ((q[j+ 0] >>  4) - ((h >> 4) & 1 ? 0 : 16))
+                 + y[j+48] * d * s[3] * ((q[j+16] >>  4) - ((h >> 6) & 1 ? 0 : 16));
+        }
+        tmp += sum;
+    }
+#endif
+
+    // sum up partial sums and write back result
+#pragma unroll
+    for (int mask = QK_WARP_SIZE / 2; mask > 0; mask >>= 1) {
+        tmp +=
+            dpct::permute_sub_group_by_xor(item_ct1.get_sub_group(), tmp, mask);
+    }
+
+    if (item_ct1.get_local_id(2) == 0) {
+        dst[row] = tmp;
+    }
+}
+
+static void dequantize_mul_mat_vec_q6_k(const void * __restrict__ vx, const float * __restrict__ yy, float * __restrict__ dst, const int ncols, int nrows,
+                                        const sycl::nd_item<3> &item_ct1) {
+
+    static_assert(16%K_QUANTS_PER_ITERATION == 0, "16 must be divisible by K_QUANTS_PER_ITERATION");
+
+    const int row = item_ct1.get_group(2) * item_ct1.get_local_range(1) +
+                    item_ct1.get_local_id(1);
+    if (row > nrows) return;
+
+    const int num_blocks_per_row = ncols / QK_K;
+    const int ib0 = row*num_blocks_per_row;
+
+    const block_q6_K * x = (const block_q6_K *)vx + ib0;
+
+#if QK_K == 256
+
+    const int tid =
+        item_ct1.get_local_id(2) / K_QUANTS_PER_ITERATION; // 0...31 or 0...16
+    const int ix =
+        item_ct1.get_local_id(2) % K_QUANTS_PER_ITERATION; // 0 or 0, 1
+
+    const int step = 16/K_QUANTS_PER_ITERATION;          // 16 or 8
+
+    const int im = tid/step;                             // 0 or 1. 0 computes 0..., 1 computes 128...
+    const int in = tid - step*im;                        // 0...15 or 0...7
+
+#if K_QUANTS_PER_ITERATION == 1
+    const int l0 = K_QUANTS_PER_ITERATION*in;            // 0...15
+    const int is = 0;
+#else
+    const int l0 = 4 * in;                               // 0, 4, 8, ..., 28
+    const int is = in / 4;
+#endif
+    const int ql_offset = 64*im + l0;
+    const int qh_offset = 32*im + l0;
+    const int s_offset  =  8*im + is;
+    const int y_offset = 128*im + l0;
+
+    float tmp = 0; // partial sum for thread in warp
+
+    for (int i = ix; i < num_blocks_per_row; i += K_QUANTS_PER_ITERATION) {
+
+        const float   * y  = yy + i * QK_K + y_offset;
+        const uint8_t * ql = x[i].ql + ql_offset;
+        const uint8_t * qh = x[i].qh + qh_offset;
+        const int8_t  * s  = x[i].scales + s_offset;
+
+        const float d = x[i].d;
+
+#if K_QUANTS_PER_ITERATION == 1
+        float sum = y[ 0] * s[0] * d * ((int8_t)((ql[ 0] & 0xF) | ((qh[ 0] & 0x03) << 4)) - 32)
+                  + y[16] * s[1] * d * ((int8_t)((ql[16] & 0xF) | ((qh[16] & 0x03) << 4)) - 32)
+                  + y[32] * s[2] * d * ((int8_t)((ql[32] & 0xF) | ((qh[ 0] & 0x0c) << 2)) - 32)
+                  + y[48] * s[3] * d * ((int8_t)((ql[48] & 0xF) | ((qh[16] & 0x0c) << 2)) - 32)
+                  + y[64] * s[4] * d * ((int8_t)((ql[ 0]  >> 4) | ((qh[ 0] & 0x30) >> 0)) - 32)
+                  + y[80] * s[5] * d * ((int8_t)((ql[16]  >> 4) | ((qh[16] & 0x30) >> 0)) - 32)
+                  + y[96] * s[6] * d * ((int8_t)((ql[32]  >> 4) | ((qh[ 0] & 0xc0) >> 2)) - 32)
+                  +y[112] * s[7] * d * ((int8_t)((ql[48]  >> 4) | ((qh[16] & 0xc0) >> 2)) - 32);
+        tmp += sum;
+#else
+        float sum = 0;
+        for (int l = 0; l < 4; ++l) {
+            sum += y[l+ 0] * s[0] * d * ((int8_t)((ql[l+ 0] & 0xF) | (((qh[l] >> 0) & 3) << 4)) - 32)
+                 + y[l+32] * s[2] * d * ((int8_t)((ql[l+32] & 0xF) | (((qh[l] >> 2) & 3) << 4)) - 32)
+                 + y[l+64] * s[4] * d * ((int8_t)((ql[l+ 0]  >> 4) | (((qh[l] >> 4) & 3) << 4)) - 32)
+                 + y[l+96] * s[6] * d * ((int8_t)((ql[l+32]  >> 4) | (((qh[l] >> 6) & 3) << 4)) - 32);
+        }
+        tmp += sum;
+#endif
+
+    }
+
+#else
+
+    const int tid = item_ct1.get_local_id(2)/(2*K_QUANTS_PER_ITERATION);  // 0...7
+    const int ix  = item_ct1.get_local_id(2)%(2*K_QUANTS_PER_ITERATION);  // 0...3
+
+    const int step = tid * K_QUANTS_PER_ITERATION;
+
+    float tmp = 0; // partial sum for thread in warp
+
+    for (int i = ix; i < num_blocks_per_row; i += 2*K_QUANTS_PER_ITERATION) {
+
+        const float   * y  = yy + i * QK_K + step;
+        const uint8_t * ql = x[i].ql + step;
+        const uint8_t * qh = x[i].qh + step;
+        const int8_t  * s  = x[i].scales;
+
+        const float d = x[i+0].d;
+
+        float sum = 0;
+        for (int j = 0; j < K_QUANTS_PER_ITERATION; ++j) {
+            sum += y[j+ 0] * s[0] * d * ((int8_t)((ql[j+ 0] & 0xF) | ((qh[j] & 0x03) << 4)) - 32)
+                 + y[j+16] * s[1] * d * ((int8_t)((ql[j+16] & 0xF) | ((qh[j] & 0x0c) << 2)) - 32)
+                 + y[j+32] * s[2] * d * ((int8_t)((ql[j+ 0] >>  4) | ((qh[j] & 0x30) >> 0)) - 32)
+                 + y[j+48] * s[3] * d * ((int8_t)((ql[j+16] >>  4) | ((qh[j] & 0xc0) >> 2)) - 32);
+        }
+        tmp += sum;
+
+    }
+
+#endif
+
+    // sum up partial sums and write back result
+#pragma unroll
+    for (int mask = QK_WARP_SIZE / 2; mask > 0; mask >>= 1) {
+        tmp +=
+            dpct::permute_sub_group_by_xor(item_ct1.get_sub_group(), tmp, mask);
+    }
+
+    if (tid == 0) {
+        dst[row] = tmp;
+    }
+}
+
+
+static void dequantize_mul_mat_vec_q4_0_sycl(const void *vx, const dfloat *y,
+                                             float *dst, const int ncols,
+                                             const int nrows,
+                                             dpct::queue_ptr stream) {
+    GGML_ASSERT(ncols % GGML_SYCL_DMMV_X == 0);
+    const int block_num_y = (nrows + GGML_SYCL_MMV_Y - 1) / GGML_SYCL_MMV_Y;
+    // the number of rows may exceed maximum grid size in the y or z dimensions, use the x dimension instead
+    const sycl::range<3> block_nums(1, 1, block_num_y);
+    const sycl::range<3> block_dims(1, GGML_SYCL_MMV_Y, WARP_SIZE);
+    {
+        dpct::has_capability_or_fail(stream->get_device(),
+                                     {sycl::aspect::fp16});
+
+        stream->parallel_for(
+            sycl::nd_range<3>(block_nums * block_dims, block_dims),
+            [=](sycl::nd_item<3> item_ct1) [[intel::reqd_sub_group_size(WARP_SIZE)]] {
+                dequantize_mul_mat_vec<QK4_0, QR4_0, dequantize_q4_0>(
+                    vx, y, dst, ncols, nrows, item_ct1);
+            });
+    }
+}
+
+static void dequantize_mul_mat_vec_q4_1_sycl(const void *vx, const dfloat *y,
+                                             float *dst, const int ncols,
+                                             const int nrows,
+                                             dpct::queue_ptr stream) {
+    GGML_ASSERT(ncols % GGML_SYCL_DMMV_X == 0);
+    const int block_num_y = (nrows + GGML_SYCL_MMV_Y - 1) / GGML_SYCL_MMV_Y;
+    const sycl::range<3> block_nums(1, 1, block_num_y);
+    const sycl::range<3> block_dims(1, GGML_SYCL_MMV_Y, WARP_SIZE);
+    {
+        dpct::has_capability_or_fail(stream->get_device(),
+                                     {sycl::aspect::fp16});
+
+        stream->parallel_for(
+            sycl::nd_range<3>(block_nums * block_dims, block_dims),
+            [=](sycl::nd_item<3> item_ct1) [[intel::reqd_sub_group_size(WARP_SIZE)]] {
+                dequantize_mul_mat_vec<QK4_1, QR4_1, dequantize_q4_1>(
+                    vx, y, dst, ncols, nrows, item_ct1);
+            });
+    }
+}
+
+static void dequantize_mul_mat_vec_q5_0_sycl(const void *vx, const dfloat *y,
+                                             float *dst, const int ncols,
+                                             const int nrows,
+                                             dpct::queue_ptr stream) {
+    GGML_ASSERT(ncols % GGML_SYCL_DMMV_X == 0);
+    const int block_num_y = (nrows + GGML_SYCL_MMV_Y - 1) / GGML_SYCL_MMV_Y;
+    const sycl::range<3> block_nums(1, 1, block_num_y);
+    const sycl::range<3> block_dims(1, GGML_SYCL_MMV_Y, WARP_SIZE);
+    {
+        dpct::has_capability_or_fail(stream->get_device(),
+                                     {sycl::aspect::fp16});
+
+        stream->parallel_for(
+            sycl::nd_range<3>(block_nums * block_dims, block_dims),
+            [=](sycl::nd_item<3> item_ct1) [[intel::reqd_sub_group_size(WARP_SIZE)]] {
+                dequantize_mul_mat_vec<QK5_0, QR5_0, dequantize_q5_0>(
+                    vx, y, dst, ncols, nrows, item_ct1);
+            });
+    }
+}
+
+static void dequantize_mul_mat_vec_q5_1_sycl(const void *vx, const dfloat *y,
+                                             float *dst, const int ncols,
+                                             const int nrows,
+                                             dpct::queue_ptr stream) {
+    GGML_ASSERT(ncols % GGML_SYCL_DMMV_X == 0);
+    const int block_num_y = (nrows + GGML_SYCL_MMV_Y - 1) / GGML_SYCL_MMV_Y;
+    const sycl::range<3> block_nums(1, 1, block_num_y);
+    const sycl::range<3> block_dims(1, GGML_SYCL_MMV_Y, WARP_SIZE);
+    {
+        dpct::has_capability_or_fail(stream->get_device(),
+                                     {sycl::aspect::fp16});
+
+        stream->parallel_for(
+            sycl::nd_range<3>(block_nums * block_dims, block_dims),
+            [=](sycl::nd_item<3> item_ct1) [[intel::reqd_sub_group_size(WARP_SIZE)]] {
+                dequantize_mul_mat_vec<QK5_1, QR5_1, dequantize_q5_1>(
+                    vx, y, dst, ncols, nrows, item_ct1);
+            });
+    }
+}
+
+static void dequantize_mul_mat_vec_q8_0_sycl(const void *vx, const dfloat *y,
+                                             float *dst, const int ncols,
+                                             const int nrows,
+                                             dpct::queue_ptr stream) {
+    GGML_ASSERT(ncols % GGML_SYCL_DMMV_X == 0);
+    const int block_num_y = (nrows + GGML_SYCL_MMV_Y - 1) / GGML_SYCL_MMV_Y;
+    const sycl::range<3> block_nums(1, 1, block_num_y);
+    const sycl::range<3> block_dims(1, GGML_SYCL_MMV_Y, WARP_SIZE);
+    {
+        dpct::has_capability_or_fail(stream->get_device(),
+                                     {sycl::aspect::fp16});
+
+        stream->parallel_for(
+            sycl::nd_range<3>(block_nums * block_dims, block_dims),
+            [=](sycl::nd_item<3> item_ct1) [[intel::reqd_sub_group_size(WARP_SIZE)]] {
+                dequantize_mul_mat_vec<QK8_0, QR8_0, dequantize_q8_0>(
+                    vx, y, dst, ncols, nrows, item_ct1);
+            });
+    }
+}
+
+static void dequantize_mul_mat_vec_q2_K_sycl(const void *vx, const float *y,
+                                             float *dst, const int ncols,
+                                             const int nrows,
+                                             dpct::queue_ptr stream) {
+    GGML_ASSERT(ncols % QK_K == 0);
+    const int ny = 2; // very slightly faster than 1 even when K_QUANTS_PER_ITERATION = 2
+    const int block_num_y = (nrows + ny - 1) / ny;
+    const sycl::range<3> block_nums(1, 1, block_num_y);
+    const sycl::range<3> block_dims(1, ny, QK_WARP_SIZE);
+    stream->parallel_for(
+        sycl::nd_range<3>(block_nums * block_dims, block_dims),
+        [=](sycl::nd_item<3> item_ct1) [[intel::reqd_sub_group_size(QK_WARP_SIZE)]] {
+            dequantize_mul_mat_vec_q2_k(vx, y, dst, ncols, nrows, item_ct1);
+        });
+}
+
+static void dequantize_mul_mat_vec_q3_K_sycl(const void *vx, const float *y,
+                                             float *dst, const int ncols,
+                                             const int nrows,
+                                             dpct::queue_ptr stream) {
+    GGML_ASSERT(ncols % QK_K == 0);
+    const int ny = 2 / K_QUANTS_PER_ITERATION;
+    const int block_num_y = (nrows + ny - 1) / ny;
+    const sycl::range<3> block_nums(1, 1, block_num_y);
+    const sycl::range<3> block_dims(1, ny, QK_WARP_SIZE);
+    stream->parallel_for(
+        sycl::nd_range<3>(block_nums * block_dims, block_dims),
+        [=](sycl::nd_item<3> item_ct1) [[intel::reqd_sub_group_size(QK_WARP_SIZE)]] {
+            dequantize_mul_mat_vec_q3_k(vx, y, dst, ncols, nrows, item_ct1);
+        });
+}
+
+static void dequantize_mul_mat_vec_q4_K_sycl(const void *vx, const float *y,
+                                             float *dst, const int ncols,
+                                             const int nrows,
+                                             dpct::queue_ptr stream) {
+    GGML_ASSERT(ncols % QK_K == 0);
+    const int ny = 2 / K_QUANTS_PER_ITERATION;
+    const int block_num_y = (nrows + ny - 1) / ny;
+    const sycl::range<3> block_nums(1, 1, block_num_y);
+    const sycl::range<3> block_dims(1, ny, QK_WARP_SIZE);
+    stream->parallel_for(
+        sycl::nd_range<3>(block_nums * block_dims, block_dims),
+        [=](sycl::nd_item<3> item_ct1) [[intel::reqd_sub_group_size(QK_WARP_SIZE)]] {
+            dequantize_mul_mat_vec_q4_k(vx, y, dst, ncols, nrows, item_ct1);
+        });
+}
+
+static void dequantize_mul_mat_vec_q5_K_sycl(const void *vx, const float *y,
+                                             float *dst, const int ncols,
+                                             const int nrows,
+                                             dpct::queue_ptr stream) {
+    GGML_ASSERT(ncols % QK_K == 0);
+    const sycl::range<3> block_dims(1, 1, QK_WARP_SIZE);
+    stream->parallel_for(
+        sycl::nd_range<3>(sycl::range<3>(1, 1, nrows) * block_dims, block_dims),
+        [=](sycl::nd_item<3> item_ct1) [[intel::reqd_sub_group_size(QK_WARP_SIZE)]] {
+            dequantize_mul_mat_vec_q5_k(vx, y, dst, ncols, item_ct1);
+        });
+}
+
+static void dequantize_mul_mat_vec_q6_K_sycl(const void *vx, const float *y,
+                                             float *dst, const int ncols,
+                                             const int nrows,
+                                             dpct::queue_ptr stream) {
+    GGML_ASSERT(ncols % QK_K == 0);
+    const int ny = 2 / K_QUANTS_PER_ITERATION;
+    const int block_num_y = (nrows + ny - 1) / ny;
+    const sycl::range<3> block_nums(1, 1, block_num_y);
+    const sycl::range<3> block_dims(1, ny, QK_WARP_SIZE);
+    stream->parallel_for(
+        sycl::nd_range<3>(block_nums * block_dims, block_dims),
+        [=](sycl::nd_item<3> item_ct1) [[intel::reqd_sub_group_size(QK_WARP_SIZE)]] {
+            dequantize_mul_mat_vec_q6_k(vx, y, dst, ncols, nrows, item_ct1);
+        });
+}
+
+void ggml_sycl_op_dequantize_mul_mat_vec(
+    ggml_backend_sycl_context & ctx,
+    const ggml_tensor *src0, const ggml_tensor *src1, ggml_tensor *dst,
+    const char *src0_dd_i, const float *src1_ddf_i, const char *src1_ddq_i,
+    float *dst_dd_i, const int64_t row_low, const int64_t row_high,
+    const int64_t src1_ncols, const int64_t src1_padded_row_size,
+    const dpct::queue_ptr &stream) {
+
+    const int64_t ne00 = src0->ne[0];
+    const int64_t row_diff = row_high - row_low;
+
+    GGML_ASSERT(src1->type == GGML_TYPE_F32);
+    // on some GPUs it is faster to convert src1 to half and to use half precision intrinsics
+#ifdef GGML_SYCL_F16
+    ggml_sycl_pool_alloc<sycl::half> src1_dfloat_a(ctx.pool());
+    sycl::half *src1_dfloat = nullptr; // dfloat == half
+
+    bool src1_convert_f16 =
+        src0->type == GGML_TYPE_Q4_0 || src0->type == GGML_TYPE_Q4_1 ||
+        src0->type == GGML_TYPE_Q5_0 || src0->type == GGML_TYPE_Q5_1 ||
+        src0->type == GGML_TYPE_Q8_0 || src0->type == GGML_TYPE_F16;
+
+    if (src1_convert_f16) {
+        src1_dfloat = src1_dfloat_a.alloc(ne00);
+        const to_fp16_sycl_t to_fp16_sycl = ggml_get_to_fp16_sycl(src1->type);
+        GGML_ASSERT(to_fp16_sycl != nullptr);
+        to_fp16_sycl(src1_ddf_i, src1_dfloat, ne00, stream);
+    }
+#else
+    const dfloat * src1_dfloat = (const dfloat *) src1_ddf_i; // dfloat == float, no conversion
+#endif // GGML_SYCL_F16
+
+    switch (src0->type) {
+        case GGML_TYPE_Q4_0:
+            dequantize_mul_mat_vec_q4_0_sycl(src0_dd_i, src1_dfloat, dst_dd_i, ne00, row_diff, stream);
+            break;
+        case GGML_TYPE_Q4_1:
+            dequantize_mul_mat_vec_q4_1_sycl(src0_dd_i, src1_dfloat, dst_dd_i, ne00, row_diff, stream);
+            break;
+        case GGML_TYPE_Q5_0:
+            dequantize_mul_mat_vec_q5_0_sycl(src0_dd_i, src1_dfloat, dst_dd_i, ne00, row_diff, stream);
+            break;
+        case GGML_TYPE_Q5_1:
+            dequantize_mul_mat_vec_q5_1_sycl(src0_dd_i, src1_dfloat, dst_dd_i, ne00, row_diff, stream);
+            break;
+        case GGML_TYPE_Q8_0:
+            dequantize_mul_mat_vec_q8_0_sycl(src0_dd_i, src1_dfloat, dst_dd_i, ne00, row_diff, stream);
+            break;
+        case GGML_TYPE_Q2_K:
+            dequantize_mul_mat_vec_q2_K_sycl(src0_dd_i, src1_ddf_i, dst_dd_i, ne00, row_diff, stream);
+            break;
+        case GGML_TYPE_Q3_K:
+            dequantize_mul_mat_vec_q3_K_sycl(src0_dd_i, src1_ddf_i, dst_dd_i, ne00, row_diff, stream);
+            break;
+        case GGML_TYPE_Q4_K:
+            dequantize_mul_mat_vec_q4_K_sycl(src0_dd_i, src1_ddf_i, dst_dd_i, ne00, row_diff, stream);
+            break;
+        case GGML_TYPE_Q5_K:
+            dequantize_mul_mat_vec_q5_K_sycl(src0_dd_i, src1_ddf_i, dst_dd_i, ne00, row_diff, stream);
+            break;
+        case GGML_TYPE_Q6_K:
+            dequantize_mul_mat_vec_q6_K_sycl(src0_dd_i, src1_ddf_i, dst_dd_i, ne00, row_diff, stream);
+            break;
+        case GGML_TYPE_F16:
+            convert_mul_mat_vec_f16_sycl(src0_dd_i, src1_dfloat, dst_dd_i, ne00, row_diff, stream);
+            break;
+        default:
+            printf("ggml_sycl_op_dequantize_mul_mat_vec unsupported GGML_TYPE %d\n", src0->type);
+            GGML_ABORT("fatal error");
+            break;
+    }
+
+    (void) src1;
+    (void) dst;
+    (void) src1_ddq_i;
+    (void) src1_ncols;
+    (void) src1_padded_row_size;
+}
diff --git a/src/whisper_cpp/ggml/src/ggml-sycl/dmmv.hpp b/src/whisper_cpp/ggml/src/ggml-sycl/dmmv.hpp
new file mode 100644
index 00000000..bd837356
--- /dev/null
+++ b/src/whisper_cpp/ggml/src/ggml-sycl/dmmv.hpp
@@ -0,0 +1,27 @@
+//
+// MIT license
+// Copyright (C) 2024 Intel Corporation
+// SPDX-License-Identifier: MIT
+//
+
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+//
+
+#ifndef GGML_SYCL_DMMV_HPP
+#define GGML_SYCL_DMMV_HPP
+
+#include "common.hpp"
+
+
+void ggml_sycl_op_dequantize_mul_mat_vec(
+    ggml_backend_sycl_context & ctx,
+    const ggml_tensor *src0, const ggml_tensor *src1, ggml_tensor *dst,
+    const char *src0_dd_i, const float *src1_ddf_i, const char *src1_ddq_i,
+    float *dst_dd_i, const int64_t row_low, const int64_t row_high,
+    const int64_t src1_ncols, const int64_t src1_padded_row_size,
+    const dpct::queue_ptr &stream);
+
+#endif // GGML_SYCL_DMMV_HPP
diff --git a/src/whisper_cpp/ggml/src/ggml-sycl/dpct/helper.hpp b/src/whisper_cpp/ggml/src/ggml-sycl/dpct/helper.hpp
new file mode 100644
index 00000000..fe4a8f74
--- /dev/null
+++ b/src/whisper_cpp/ggml/src/ggml-sycl/dpct/helper.hpp
@@ -0,0 +1,3024 @@
+//
+// MIT license
+// Copyright (C) 2024 Intel Corporation
+// SPDX-License-Identifier: MIT
+//
+
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+//
+
+#ifndef GGML_SYCL_DPCT_HELPER_HPP
+#define GGML_SYCL_DPCT_HELPER_HPP
+
+#include <sycl/sycl.hpp>
+#include <sycl/half_type.hpp>
+#include <oneapi/mkl.hpp>
+#include <map>
+
+#include "ggml.h"
+
+#if defined(__linux__)
+#include <sys/mman.h>
+#elif defined(_WIN64)
+#ifndef NOMINMAX
+#define NOMINMAX
+#endif
+#include <windows.h>
+#else
+#error "Only support Windows and Linux."
+#endif
+
+#if defined(__linux__)
+#include <unistd.h>
+#include <sys/syscall.h>
+#endif
+#if defined(_WIN64)
+#ifndef NOMINMAX
+#define NOMINMAX
+#endif
+#include <windows.h>
+#endif
+
+#define DPCT_COMPATIBILITY_TEMP (900)
+
+#if defined(_MSC_VER)
+#define __dpct_align__(n) __declspec(align(n))
+#define __dpct_inline__ __forceinline
+#else
+#define __dpct_align__(n) __attribute__((aligned(n)))
+#define __dpct_inline__ __inline__ __attribute__((always_inline))
+#endif
+
+#if defined(_MSC_VER)
+#define __dpct_noinline__ __declspec(noinline)
+#else
+#define __dpct_noinline__ __attribute__((noinline))
+#endif
+
+inline std::string get_device_type_name(const sycl::device &Device) {
+    auto DeviceType = Device.get_info<sycl::info::device::device_type>();
+    switch (DeviceType) {
+    case sycl::info::device_type::cpu:
+        return "cpu";
+    case sycl::info::device_type::gpu:
+        return "gpu";
+    case sycl::info::device_type::host:
+        return "host";
+    case sycl::info::device_type::accelerator:
+        return "acc";
+    default:
+        return "unknown";
+    }
+}
+
+inline std::string get_device_backend_and_type(const sycl::device &device) {
+    std::stringstream device_type;
+    sycl::backend backend = device.get_backend();
+    device_type <<  backend << ":" << get_device_type_name(device);
+    return device_type.str();
+}
+
+namespace dpct
+{
+    typedef sycl::queue *queue_ptr;
+    typedef sycl::event *event_ptr;
+    typedef char *device_ptr;
+    typedef uint8_t byte_t;
+    typedef sycl::buffer<byte_t> buffer_t;
+
+    /// SYCL default exception handler
+    inline auto exception_handler = [](sycl::exception_list exceptions)
+    {
+        for (std::exception_ptr const &e : exceptions)
+        {
+            try
+            {
+                std::rethrow_exception(e);
+            }
+            catch (sycl::exception const &e)
+            {
+                std::cerr << "Caught asynchronous SYCL exception:" << std::endl
+                          << e.what() << std::endl
+                          << "Exception caught at file:" << __FILE__
+                          << ", line:" << __LINE__ << std::endl;
+            }
+        }
+    };
+
+    enum error_code
+    {
+        success = 0,
+        default_error = 999
+    };
+
+    enum memcpy_direction
+    {
+        host_to_host,
+        host_to_device,
+        device_to_host,
+        device_to_device,
+        automatic
+    };
+
+    enum memory_region
+    {
+        global = 0, // device global memory
+        constant,   // device constant memory
+        local,      // device local memory
+        shared,     // memory which can be accessed by host and device
+    };
+
+    enum class library_data_t : unsigned char
+    {
+        real_float = 0,
+        complex_float,
+        real_double,
+        complex_double,
+        real_half,
+        complex_half,
+        real_bfloat16,
+        complex_bfloat16,
+        real_int4,
+        complex_int4,
+        real_uint4,
+        complex_uint4,
+        real_int8,
+        complex_int8,
+        real_uint8,
+        complex_uint8,
+        real_int16,
+        complex_int16,
+        real_uint16,
+        complex_uint16,
+        real_int32,
+        complex_int32,
+        real_uint32,
+        complex_uint32,
+        real_int64,
+        complex_int64,
+        real_uint64,
+        complex_uint64,
+        real_int8_4,
+        real_int8_32,
+        real_uint8_4,
+        library_data_t_size
+    };
+
+    template <typename T>
+    struct DataType
+    {
+        using T2 = T;
+    };
+    template <typename T>
+    struct DataType<sycl::vec<T, 2>>
+    {
+        using T2 = std::complex<T>;
+    };
+
+    static void destroy_event(event_ptr event)
+    {
+        delete event;
+    }
+
+    static inline unsigned int get_tid()
+    {
+#if defined(__linux__)
+        return syscall(SYS_gettid);
+#elif defined(_WIN64)
+        return GetCurrentThreadId();
+#else
+#error "Only support Windows and Linux."
+#endif
+    }
+
+    namespace detail
+    {
+        static void get_version(const sycl::device &dev, int &major, int &minor)
+        {
+            // Version string has the following format:
+            // a. OpenCL<space><major.minor><space><vendor-specific-information>
+            // b. <major.minor>
+            // c. <AmdGcnArchName> e.g gfx1030
+            std::string ver;
+            ver = dev.get_info<sycl::info::device::version>();
+            std::string::size_type i = 0;
+            while (i < ver.size()) {
+              if (isdigit(ver[i]))
+                break;
+              i++;
+            }
+            major = std::stoi(&(ver[i]));
+            while (i < ver.size()) {
+              if (ver[i] == '.')
+                break;
+              i++;
+            }
+            if (i < ver.size()) {
+              // a. and b.
+              i++;
+              minor = std::stoi(&(ver[i]));
+            } else {
+              // c.
+              minor = 0;
+            }
+        }
+
+        template <typename tag, typename T>
+        class generic_error_type
+        {
+        public:
+            generic_error_type() = default;
+            generic_error_type(T value) : value{value} {}
+            operator T() const { return value; }
+
+        private:
+            T value;
+        };
+
+    } // namespace detail
+
+    /// Pitched 2D/3D memory data.
+    class pitched_data
+    {
+    public:
+        pitched_data() : pitched_data(nullptr, 0, 0, 0) {}
+        pitched_data(void *data, size_t pitch, size_t x, size_t y)
+            : _data(data), _pitch(pitch), _x(x), _y(y) {}
+
+        void *get_data_ptr() { return _data; }
+        void set_data_ptr(void *data) { _data = data; }
+
+        size_t get_pitch() { return _pitch; }
+        void set_pitch(size_t pitch) { _pitch = pitch; }
+
+        size_t get_x() { return _x; }
+        void set_x(size_t x) { _x = x; }
+
+        size_t get_y() { return _y; }
+        void set_y(size_t y) { _y = y; }
+
+    private:
+        void *_data;
+        size_t _pitch, _x, _y;
+    };
+
+    class device_info
+    {
+    public:
+        // get interface
+        const char *get_name() const { return _name; }
+        char *get_name() { return _name; }
+        template <typename WorkItemSizesTy = sycl::range<3>,
+                  std::enable_if_t<std::is_same_v<WorkItemSizesTy, sycl::range<3>> ||
+                                       std::is_same_v<WorkItemSizesTy, int *>,
+                                   int> = 0>
+        auto get_max_work_item_sizes() const
+        {
+            if constexpr (std::is_same_v<WorkItemSizesTy, sycl::range<3>>)
+                return sycl::range<3>(_max_work_item_sizes_i[0],
+                                      _max_work_item_sizes_i[1],
+                                      _max_work_item_sizes_i[2]);
+            else
+            {
+                return _max_work_item_sizes_i;
+            }
+        }
+        template <typename WorkItemSizesTy = sycl::range<3>,
+                  std::enable_if_t<std::is_same_v<WorkItemSizesTy, sycl::range<3>> ||
+                                       std::is_same_v<WorkItemSizesTy, int *>,
+                                   int> = 0>
+        auto get_max_work_item_sizes()
+        {
+            if constexpr (std::is_same_v<WorkItemSizesTy, sycl::range<3>>)
+                return sycl::range<3>(_max_work_item_sizes_i[0],
+                                      _max_work_item_sizes_i[1],
+                                      _max_work_item_sizes_i[2]);
+            else
+            {
+                return _max_work_item_sizes_i;
+            }
+        }
+        bool get_host_unified_memory() const { return _host_unified_memory; }
+        int get_major_version() const { return _major; }
+        int get_minor_version() const { return _minor; }
+        int get_integrated() const { return _integrated; }
+        int get_max_clock_frequency() const { return _frequency; }
+        int get_max_compute_units() const { return _max_compute_units; }
+        int get_max_work_group_size() const { return _max_work_group_size; }
+        int get_max_sub_group_size() const { return _max_sub_group_size; }
+        int get_max_work_items_per_compute_unit() const
+        {
+            return _max_work_items_per_compute_unit;
+        }
+        int get_max_register_size_per_work_group() const
+        {
+            return _max_register_size_per_work_group;
+        }
+        template <typename NDRangeSizeTy = size_t *,
+                  std::enable_if_t<std::is_same_v<NDRangeSizeTy, size_t *> ||
+                                       std::is_same_v<NDRangeSizeTy, int *>,
+                                   int> = 0>
+        auto get_max_nd_range_size() const
+        {
+            if constexpr (std::is_same_v<NDRangeSizeTy, size_t *>)
+                return _max_nd_range_size;
+            else
+                return _max_nd_range_size_i;
+        }
+        template <typename NDRangeSizeTy = size_t *,
+                  std::enable_if_t<std::is_same_v<NDRangeSizeTy, size_t *> ||
+                                       std::is_same_v<NDRangeSizeTy, int *>,
+                                   int> = 0>
+        auto get_max_nd_range_size()
+        {
+            if constexpr (std::is_same_v<NDRangeSizeTy, size_t *>)
+                return _max_nd_range_size;
+            else
+                return _max_nd_range_size_i;
+        }
+        size_t get_global_mem_size() const { return _global_mem_size; }
+        size_t get_local_mem_size() const { return _local_mem_size; }
+        size_t get_max_mem_alloc_size() const { return _max_mem_alloc_size; }
+        /// Returns the maximum clock rate of device's global memory in kHz. If
+        /// compiler does not support this API then returns default value 3200000 kHz.
+        unsigned int get_memory_clock_rate() const { return _memory_clock_rate; }
+        /// Returns the maximum bus width between device and memory in bits. If
+        /// compiler does not support this API then returns default value 64 bits.
+        unsigned int get_memory_bus_width() const { return _memory_bus_width; }
+        uint32_t get_device_id() const { return _device_id; }
+        std::array<unsigned char, 16> get_uuid() const { return _uuid; }
+        /// Returns global memory cache size in bytes.
+        unsigned int get_global_mem_cache_size() const
+        {
+            return _global_mem_cache_size;
+        }
+
+        // set interface
+        void set_name(const char *name)
+        {
+            size_t length = strlen(name);
+            if (length < 256)
+            {
+                std::memcpy(_name, name, length + 1);
+            }
+            else
+            {
+                std::memcpy(_name, name, 255);
+                _name[255] = '\0';
+            }
+        }
+        void set_max_work_item_sizes(const sycl::range<3> max_work_item_sizes)
+        {
+            for (int i = 0; i < 3; ++i)
+                _max_work_item_sizes_i[i] = max_work_item_sizes[i];
+        }
+        [[deprecated]] void
+        set_max_work_item_sizes(const sycl::id<3> max_work_item_sizes)
+        {
+            for (int i = 0; i < 3; ++i)
+            {
+                _max_work_item_sizes_i[i] = max_work_item_sizes[i];
+            }
+        }
+        void set_host_unified_memory(bool host_unified_memory)
+        {
+            _host_unified_memory = host_unified_memory;
+        }
+        void set_major_version(int major) { _major = major; }
+        void set_minor_version(int minor) { _minor = minor; }
+        void set_integrated(int integrated) { _integrated = integrated; }
+        void set_max_clock_frequency(int frequency) { _frequency = frequency; }
+        void set_max_compute_units(int max_compute_units)
+        {
+            _max_compute_units = max_compute_units;
+        }
+        void set_global_mem_size(size_t global_mem_size)
+        {
+            _global_mem_size = global_mem_size;
+        }
+        void set_local_mem_size(size_t local_mem_size)
+        {
+            _local_mem_size = local_mem_size;
+        }
+        void set_max_mem_alloc_size(size_t max_mem_alloc_size)
+        {
+            _max_mem_alloc_size = max_mem_alloc_size;
+        }
+        void set_max_work_group_size(int max_work_group_size)
+        {
+            _max_work_group_size = max_work_group_size;
+        }
+        void set_max_sub_group_size(int max_sub_group_size)
+        {
+            _max_sub_group_size = max_sub_group_size;
+        }
+        void
+        set_max_work_items_per_compute_unit(int max_work_items_per_compute_unit)
+        {
+            _max_work_items_per_compute_unit = max_work_items_per_compute_unit;
+        }
+        void set_max_nd_range_size(int max_nd_range_size[])
+        {
+            for (int i = 0; i < 3; i++)
+            {
+                _max_nd_range_size[i] = max_nd_range_size[i];
+                _max_nd_range_size_i[i] = max_nd_range_size[i];
+            }
+        }
+        void set_memory_clock_rate(unsigned int memory_clock_rate)
+        {
+            _memory_clock_rate = memory_clock_rate;
+        }
+        void set_memory_bus_width(unsigned int memory_bus_width)
+        {
+            _memory_bus_width = memory_bus_width;
+        }
+        void
+        set_max_register_size_per_work_group(int max_register_size_per_work_group)
+        {
+            _max_register_size_per_work_group = max_register_size_per_work_group;
+        }
+        void set_device_id(uint32_t device_id)
+        {
+            _device_id = device_id;
+        }
+        void set_uuid(std::array<unsigned char, 16> uuid)
+        {
+            _uuid = std::move(uuid);
+        }
+        void set_global_mem_cache_size(unsigned int global_mem_cache_size)
+        {
+            _global_mem_cache_size = global_mem_cache_size;
+        }
+
+    private:
+        char _name[256];
+        int _max_work_item_sizes_i[3];
+        bool _host_unified_memory = false;
+        int _major;
+        int _minor;
+        int _integrated = 0;
+        int _frequency;
+        // Set estimated value 3200000 kHz as default value.
+        unsigned int _memory_clock_rate = 3200000;
+        // Set estimated value 64 bits as default value.
+        unsigned int _memory_bus_width = 64;
+        unsigned int _global_mem_cache_size;
+        int _max_compute_units;
+        int _max_work_group_size;
+        int _max_sub_group_size;
+        int _max_work_items_per_compute_unit;
+        int _max_register_size_per_work_group;
+        size_t _global_mem_size;
+        size_t _local_mem_size;
+        size_t _max_mem_alloc_size;
+        size_t _max_nd_range_size[3];
+        int _max_nd_range_size_i[3];
+        uint32_t _device_id;
+        std::array<unsigned char, 16> _uuid;
+    };
+
+    static int get_major_version(const sycl::device &dev)
+    {
+        int major, minor;
+        detail::get_version(dev, major, minor);
+        return major;
+    }
+
+    static int get_minor_version(const sycl::device &dev)
+    {
+        int major, minor;
+        detail::get_version(dev, major, minor);
+        return minor;
+    }
+
+    static void get_device_info(device_info &out, const sycl::device &dev)
+    {
+        device_info prop;
+        prop.set_name(dev.get_info<sycl::info::device::name>().c_str());
+
+        int major, minor;
+        detail::get_version(dev, major, minor);
+        prop.set_major_version(major);
+        prop.set_minor_version(minor);
+
+        prop.set_max_work_item_sizes(
+#if (__SYCL_COMPILER_VERSION && __SYCL_COMPILER_VERSION < 20220902)
+            // oneAPI DPC++ compiler older than 2022/09/02, where max_work_item_sizes
+            // is an enum class element
+            dev.get_info<sycl::info::device::max_work_item_sizes>());
+#else
+            // SYCL 2020-conformant code, max_work_item_sizes is a struct templated by
+            // an int
+            dev.get_info<sycl::info::device::max_work_item_sizes<3>>());
+#endif
+        prop.set_host_unified_memory(dev.has(sycl::aspect::usm_host_allocations));
+
+        prop.set_max_clock_frequency(
+            dev.get_info<sycl::info::device::max_clock_frequency>() * 1000);
+
+        prop.set_max_compute_units(
+            dev.get_info<sycl::info::device::max_compute_units>());
+        prop.set_max_work_group_size(
+            dev.get_info<sycl::info::device::max_work_group_size>());
+        prop.set_global_mem_size(dev.get_info<sycl::info::device::global_mem_size>());
+        prop.set_local_mem_size(dev.get_info<sycl::info::device::local_mem_size>());
+        prop.set_max_mem_alloc_size(dev.get_info<sycl::info::device::max_mem_alloc_size>());
+
+#if (defined(SYCL_EXT_INTEL_DEVICE_INFO) && SYCL_EXT_INTEL_DEVICE_INFO >= 6)
+        if (dev.has(sycl::aspect::ext_intel_memory_clock_rate))
+        {
+            unsigned int tmp =
+                dev.get_info<sycl::ext::intel::info::device::memory_clock_rate>();
+            if (tmp != 0)
+                prop.set_memory_clock_rate(1000 * tmp);
+        }
+        if (dev.has(sycl::aspect::ext_intel_memory_bus_width))
+        {
+            prop.set_memory_bus_width(
+                dev.get_info<sycl::ext::intel::info::device::memory_bus_width>());
+        }
+        if (dev.has(sycl::aspect::ext_intel_device_id))
+        {
+            prop.set_device_id(
+                dev.get_info<sycl::ext::intel::info::device::device_id>());
+        }
+        if (dev.has(sycl::aspect::ext_intel_device_info_uuid))
+        {
+            prop.set_uuid(dev.get_info<sycl::ext::intel::info::device::uuid>());
+        }
+#elif defined(_MSC_VER) && !defined(__clang__)
+#pragma message("get_device_info: querying memory_clock_rate and \
+        memory_bus_width are not supported by the compiler used. \
+        Use 3200000 kHz as memory_clock_rate default value. \
+        Use 64 bits as memory_bus_width default value.")
+#else
+#warning "get_device_info: querying memory_clock_rate and \
+        memory_bus_width are not supported by the compiler used. \
+        Use 3200000 kHz as memory_clock_rate default value. \
+        Use 64 bits as memory_bus_width default value."
+#endif
+
+        size_t max_sub_group_size = 1;
+        std::vector<size_t> sub_group_sizes =
+            dev.get_info<sycl::info::device::sub_group_sizes>();
+
+        for (const auto &sub_group_size : sub_group_sizes)
+        {
+            if (max_sub_group_size < sub_group_size)
+                max_sub_group_size = sub_group_size;
+        }
+
+        prop.set_max_sub_group_size(max_sub_group_size);
+
+        prop.set_max_work_items_per_compute_unit(
+            dev.get_info<sycl::info::device::max_work_group_size>());
+        int max_nd_range_size[] = {0x7FFFFFFF, 0x7FFFFFFF, 0x7FFFFFFF};
+        prop.set_max_nd_range_size(max_nd_range_size);
+
+        // Estimates max register size per work group, feel free to update the value
+        // according to device properties.
+        prop.set_max_register_size_per_work_group(65536);
+
+        prop.set_global_mem_cache_size(
+            dev.get_info<sycl::info::device::global_mem_cache_size>());
+        out = prop;
+    }
+
+    /// dpct device extension
+    class device_ext : public sycl::device {
+      typedef std::mutex mutex_type;
+
+     public:
+      device_ext() : sycl::device() {}
+      ~device_ext() {
+        std::lock_guard<mutex_type> lock(m_mutex);
+        clear_queues();
+      }
+      device_ext(const sycl::device &base) : sycl::device(base) {
+        std::lock_guard<mutex_type> lock(m_mutex);
+        init_queues();
+      }
+
+      int is_native_atomic_supported() { return 0; }
+      int get_major_version() const { return dpct::get_major_version(*this); }
+
+      int get_minor_version() const { return dpct::get_minor_version(*this); }
+
+      int get_max_compute_units() const {
+        return get_device_info().get_max_compute_units();
+      }
+
+      /// Return the maximum clock frequency of this device in KHz.
+      int get_max_clock_frequency() const {
+        return get_device_info().get_max_clock_frequency();
+      }
+
+      int get_integrated() const { return get_device_info().get_integrated(); }
+
+      int get_max_sub_group_size() const {
+        return get_device_info().get_max_sub_group_size();
+      }
+
+      int get_max_register_size_per_work_group() const {
+        return get_device_info().get_max_register_size_per_work_group();
+      }
+
+      int get_max_work_group_size() const {
+        return get_device_info().get_max_work_group_size();
+      }
+
+      int get_mem_base_addr_align() const {
+        return get_info<sycl::info::device::mem_base_addr_align>();
+      }
+
+      size_t get_global_mem_size() const {
+        return get_device_info().get_global_mem_size();
+      }
+
+      size_t get_max_mem_alloc_size() const {
+        return get_device_info().get_max_mem_alloc_size();
+      }
+
+      /// Get the number of bytes of free and total memory on the SYCL device.
+      /// \param [out] free_memory The number of bytes of free memory on the
+      /// SYCL device. \param [out] total_memory The number of bytes of total
+      /// memory on the SYCL device.
+      void get_memory_info(size_t &free_memory, size_t &total_memory) {
+        total_memory = get_device_info().get_global_mem_size();
+        const char *warning_info =
+            "get_memory_info: [warning] ext_intel_free_memory is not "
+            "supported (export/set ZES_ENABLE_SYSMAN=1 to support), "
+            "use total memory as free memory";
+#if (defined(__SYCL_COMPILER_VERSION) && __SYCL_COMPILER_VERSION >= 20221105)
+        if (!has(sycl::aspect::ext_intel_free_memory)) {
+          std::cerr << warning_info << std::endl;
+          free_memory = total_memory;
+        } else {
+          free_memory = get_info<sycl::ext::intel::info::device::free_memory>();
+        }
+#else
+        std::cerr << warning_info << std::endl;
+        free_memory = total_memory;
+#if defined(_MSC_VER) && !defined(__clang__)
+#pragma message("Querying the number of bytes of free memory is not supported")
+#else
+#warning "Querying the number of bytes of free memory is not supported"
+#endif
+#endif
+      }
+
+      void get_device_info(device_info &out) const {
+        dpct::get_device_info(out, *this);
+      }
+
+      device_info get_device_info() const {
+        device_info prop;
+        dpct::get_device_info(prop, *this);
+        return prop;
+      }
+
+      void reset() {
+        std::lock_guard<mutex_type> lock(m_mutex);
+        clear_queues();
+        init_queues();
+      }
+
+      sycl::queue &in_order_queue() { return _q_in_order; }
+
+      sycl::queue &out_of_order_queue() { return _q_out_of_order; }
+
+      sycl::queue &default_queue() { return in_order_queue(); }
+
+      void queues_wait_and_throw() {
+        std::unique_lock<mutex_type> lock(m_mutex);
+        lock.unlock();
+        for (auto &q : _queues) {
+            q.wait_and_throw();
+        }
+        // Guard the destruct of current_queues to make sure the ref count is
+        // safe.
+        lock.lock();
+      }
+
+      sycl::queue create_queue(bool enable_exception_handler = false) {
+        return create_in_order_queue(enable_exception_handler);
+      }
+
+      sycl::queue create_queue(sycl::device device,
+                               bool enable_exception_handler = false) {
+        return create_in_order_queue(device, enable_exception_handler);
+      }
+
+      sycl::queue create_in_order_queue(bool enable_exception_handler = false) {
+        std::lock_guard<mutex_type> lock(m_mutex);
+        return create_queue_impl(enable_exception_handler,
+                                 sycl::property::queue::in_order());
+      }
+
+      sycl::queue create_in_order_queue(sycl::device device,
+                                        bool enable_exception_handler = false) {
+        std::lock_guard<mutex_type> lock(m_mutex);
+        return create_queue_impl(device, enable_exception_handler,
+                                 sycl::property::queue::in_order());
+      }
+
+      sycl::queue create_out_of_order_queue(
+          bool enable_exception_handler = false) {
+        std::lock_guard<mutex_type> lock(m_mutex);
+        return create_queue_impl(enable_exception_handler);
+      }
+
+      void destroy_queue(sycl::queue queue) {
+        std::lock_guard<mutex_type> lock(m_mutex);
+        _queues.erase(std::remove_if(_queues.begin(), _queues.end(),
+                                    [=](const sycl::queue &q) -> bool
+                                    {
+                                        return q == queue;
+                                    }),
+                    _queues.end());
+      }
+      void set_saved_queue(sycl::queue q) {
+        std::lock_guard<mutex_type> lock(m_mutex);
+        _saved_queue = q;
+      }
+      sycl::queue get_saved_queue() const {
+        std::lock_guard<mutex_type> lock(m_mutex);
+        return _saved_queue;
+      }
+
+     private:
+      void clear_queues() { _queues.clear(); }
+
+      void init_queues() {
+        _q_in_order =
+            create_queue_impl(true, sycl::property::queue::in_order());
+        _q_out_of_order = create_queue_impl(true);
+        _saved_queue = default_queue();
+      }
+
+      /// Caller should acquire resource \p m_mutex before calling this
+      /// function.
+      template <class... Properties>
+      sycl::queue create_queue_impl(bool enable_exception_handler,
+                                    Properties... properties) {
+        sycl::async_handler eh = {};
+        if (enable_exception_handler) {
+          eh = exception_handler;
+        }
+        _queues.push_back(sycl::queue(
+            *this, eh,
+            sycl::property_list(
+#ifdef DPCT_PROFILING_ENABLED
+                sycl::property::queue::enable_profiling(),
+#endif
+                properties...)));
+
+        return _queues.back();
+      }
+
+      template <class... Properties>
+      sycl::queue create_queue_impl(sycl::device device,
+                                    bool enable_exception_handler,
+                                    Properties... properties) {
+        sycl::async_handler eh = {};
+        if (enable_exception_handler) {
+          eh = exception_handler;
+        }
+        _queues.push_back(sycl::queue(
+            device, eh,
+                        sycl::property_list(
+#ifdef DPCT_PROFILING_ENABLED
+                            sycl::property::queue::enable_profiling(),
+#endif
+                            properties...)));
+
+        return _queues.back();
+      }
+
+      void get_version(int &major, int &minor) const {
+        detail::get_version(*this, major, minor);
+      }
+      sycl::queue _q_in_order, _q_out_of_order;
+      sycl::queue _saved_queue;
+      std::vector<sycl::queue> _queues;
+      mutable mutex_type m_mutex;
+    };
+
+
+    /// device manager
+    class dev_mgr
+    {
+    public:
+        device_ext &current_device()
+        {
+            unsigned int dev_id = current_device_id();
+            check_id(dev_id);
+            return *_devs[dev_id];
+        }
+        device_ext &cpu_device() const
+        {
+            std::lock_guard<std::recursive_mutex> lock(m_mutex);
+            if (_cpu_device == -1)
+            {
+                throw std::runtime_error("no valid cpu device");
+            }
+            else
+            {
+                return *_devs[_cpu_device];
+            }
+        }
+        device_ext &get_device(unsigned int id) const
+        {
+            std::lock_guard<std::recursive_mutex> lock(m_mutex);
+            check_id(id);
+            return *_devs[id];
+        }
+        unsigned int current_device_id() const
+        {
+            std::lock_guard<std::recursive_mutex> lock(m_mutex);
+            auto it = _thread2dev_map.find(get_tid());
+            if (it != _thread2dev_map.end())
+                return it->second;
+            return DEFAULT_DEVICE_ID;
+        }
+
+        /// Select device with a device ID.
+        /// \param [in] id The id of the device which can
+        /// be obtained through get_device_id(const sycl::device).
+        void select_device(unsigned int id)
+        {
+            std::lock_guard<std::recursive_mutex> lock(m_mutex);
+            check_id(id);
+            _thread2dev_map[get_tid()] = id;
+        }
+        unsigned int device_count() { return _devs.size(); }
+
+        unsigned int get_device_id(const sycl::device &dev)
+        {
+            unsigned int id = 0;
+            for (auto &dev_item : _devs)
+            {
+                if (*dev_item == dev)
+                {
+                    return id;
+                }
+                id++;
+            }
+            return -1;
+        }
+
+        inline std::string get_preferred_gpu_platform_name() {
+            std::string result;
+
+            std::string filter = "";
+            char* env = getenv("ONEAPI_DEVICE_SELECTOR");
+            if (env) {
+                if (std::strstr(env, "level_zero")) {
+                    filter = "level-zero";
+                }
+                else if (std::strstr(env, "opencl")) {
+                    filter = "opencl";
+                }
+                else if (std::strstr(env, "cuda")) {
+                    filter = "cuda";
+                }
+                else if (std::strstr(env, "hip")) {
+                    filter = "hip";
+                }
+                else {
+                    throw std::runtime_error("invalid device filter: " + std::string(env));
+                }
+            } else {
+                auto default_device = sycl::device(sycl::default_selector_v);
+                auto default_platform_name = default_device.get_platform().get_info<sycl::info::platform::name>();
+
+                if (std::strstr(default_platform_name.c_str(), "Level-Zero") || default_device.is_cpu()) {
+                    filter = "level-zero";
+                }
+                else if (std::strstr(default_platform_name.c_str(), "CUDA")) {
+                    filter = "cuda";
+                }
+                else if (std::strstr(default_platform_name.c_str(), "HIP")) {
+                    filter = "hip";
+                }
+            }
+
+            auto platform_list = sycl::platform::get_platforms();
+
+            for (const auto& platform : platform_list) {
+                auto devices = platform.get_devices();
+                auto gpu_dev = std::find_if(devices.begin(), devices.end(), [](const sycl::device& d) {
+                    return d.is_gpu();
+                });
+
+                if (gpu_dev == devices.end()) {
+                    // cout << "platform [" << platform_name
+                    //      << "] does not contain GPU devices, skipping\n";
+                    continue;
+                }
+
+                auto platform_name = platform.get_info<sycl::info::platform::name>();
+                std::string platform_name_low_case;
+                platform_name_low_case.resize(platform_name.size());
+
+                std::transform(
+                    platform_name.begin(), platform_name.end(), platform_name_low_case.begin(), ::tolower);
+
+                if (platform_name_low_case.find(filter) == std::string::npos) {
+                    // cout << "platform [" << platform_name
+                    //      << "] does not match with requested "
+                    //      << filter << ", skipping\n";
+                    continue;
+                }
+
+                result = platform_name;
+            }
+
+            if (result.empty())
+                throw std::runtime_error("can not find preferred GPU platform");
+
+            return result;
+        }
+
+        template <class DeviceSelector>
+        std::enable_if_t<
+            std::is_invocable_r_v<int, DeviceSelector, const sycl::device &>>
+        select_device(const DeviceSelector &selector = sycl::gpu_selector_v)
+        {
+            sycl::device selected_device = sycl::device(selector);
+            unsigned int selected_device_id = get_device_id(selected_device);
+            select_device(selected_device_id);
+        }
+
+        /// Returns the instance of device manager singleton.
+        static dev_mgr &instance()
+        {
+            static dev_mgr d_m;
+            return d_m;
+        }
+        dev_mgr(const dev_mgr &) = delete;
+        dev_mgr &operator=(const dev_mgr &) = delete;
+        dev_mgr(dev_mgr &&) = delete;
+        dev_mgr &operator=(dev_mgr &&) = delete;
+
+    private:
+        mutable std::recursive_mutex m_mutex;
+        static bool compare_dev(sycl::device &device1, sycl::device &device2)
+        {
+            sycl::backend backend1 = device1.get_backend();
+            sycl::backend backend2 = device2.get_backend();
+            // levelzero backends always come first
+            if(backend1 == sycl::backend::ext_oneapi_level_zero && backend2 != sycl::backend::ext_oneapi_level_zero) return true;
+            if(backend1 != sycl::backend::ext_oneapi_level_zero && backend2 == sycl::backend::ext_oneapi_level_zero) return false;
+            dpct::device_info prop1;
+            dpct::get_device_info(prop1, device1);
+            dpct::device_info prop2;
+            dpct::get_device_info(prop2, device2);
+            return prop1.get_max_compute_units() > prop2.get_max_compute_units();
+        }
+        static int convert_backend_index(std::string & backend) {
+            if (backend == "ext_oneapi_level_zero:gpu") return 0;
+            if (backend == "opencl:gpu") return 1;
+            if (backend == "ext_oneapi_cuda:gpu") return 2;
+            if (backend == "ext_oneapi_hip:gpu") return 3;
+            if (backend == "opencl:cpu") return 4;
+            if (backend == "opencl:acc") return 5;
+            printf("convert_backend_index: can't handle backend=%s\n", backend.c_str());
+            GGML_ABORT("fatal error");
+        }
+        static bool compare_backend(std::string &backend1, std::string &backend2) {
+            return convert_backend_index(backend1) < convert_backend_index(backend2);
+        }
+        dev_mgr()
+        {
+            sycl::device default_device =
+                sycl::device(sycl::default_selector_v);
+            _devs.push_back(std::make_shared<device_ext>(default_device));
+
+            std::vector<sycl::device> sycl_all_devs;
+            // Collect other devices except for the default device.
+            if (default_device.is_cpu())
+                _cpu_device = 0;
+
+            auto Platforms = sycl::platform::get_platforms();
+            // Keep track of the number of devices per backend
+            std::map<sycl::backend, size_t> DeviceNums;
+            std::map<std::string, std::vector<sycl::device>> backend_devices;
+            auto preferred_platform_name = get_preferred_gpu_platform_name();
+
+            while (!Platforms.empty()) {
+                auto Platform = Platforms.back();
+                Platforms.pop_back();
+                auto platform_name = Platform.get_info<sycl::info::platform::name>();
+                if (platform_name.compare(preferred_platform_name) != 0) {
+                    continue;
+                }
+                auto devices = Platform.get_devices();
+                std::string backend_type = get_device_backend_and_type(devices[0]);
+                for (const auto &device : devices) {
+                    backend_devices[backend_type].push_back(device);
+                }
+            }
+
+            std::vector<std::string> keys;
+            for(auto it = backend_devices.begin(); it != backend_devices.end(); ++it) {
+                keys.push_back(it->first);
+            }
+            std::sort(keys.begin(), keys.end(), compare_backend);
+
+            for (auto &key : keys) {
+                std::vector<sycl::device> devs = backend_devices[key];
+                std::sort(devs.begin(), devs.end(), compare_dev);
+                for (const auto &dev : devs) {
+                    sycl_all_devs.push_back(dev);
+                }
+            }
+
+            for (auto &dev : sycl_all_devs)
+            {
+                if (dev == default_device)
+                {
+                    continue;
+                }
+                _devs.push_back(std::make_shared<device_ext>(dev));
+                if (_cpu_device == -1 && dev.is_cpu())
+                {
+                    _cpu_device = _devs.size() - 1;
+                }
+            }
+        }
+        void check_id(unsigned int id) const
+        {
+            if (id >= _devs.size())
+            {
+                throw std::runtime_error("invalid device id");
+            }
+        }
+        std::vector<std::shared_ptr<device_ext>> _devs;
+        /// DEFAULT_DEVICE_ID is used, if current_device_id() can not find current
+        /// thread id in _thread2dev_map, which means default device should be used
+        /// for the current thread.
+        const unsigned int DEFAULT_DEVICE_ID = 0;
+        /// thread-id to device-id map.
+        std::map<unsigned int, unsigned int> _thread2dev_map;
+        int _cpu_device = -1;
+    };
+
+    static inline sycl::queue &get_default_queue()
+    {
+        return dev_mgr::instance().current_device().default_queue();
+    }
+
+    namespace detail
+    {
+        enum class pointer_access_attribute
+        {
+            host_only = 0,
+            device_only,
+            host_device,
+            end
+        };
+
+        static pointer_access_attribute get_pointer_attribute(sycl::queue &q,
+                                                              const void *ptr)
+        {
+            switch (sycl::get_pointer_type(ptr, q.get_context()))
+            {
+            case sycl::usm::alloc::unknown:
+                return pointer_access_attribute::host_only;
+            case sycl::usm::alloc::device:
+                return pointer_access_attribute::device_only;
+            case sycl::usm::alloc::shared:
+            case sycl::usm::alloc::host:
+                return pointer_access_attribute::host_device;
+            }
+        }
+
+        template <typename ArgT>
+        inline constexpr std::uint64_t get_type_combination_id(ArgT Val)
+        {
+            static_assert((unsigned char)library_data_t::library_data_t_size <=
+                              std::numeric_limits<unsigned char>::max() &&
+                          "library_data_t size exceeds limit.");
+            static_assert(std::is_same_v<ArgT, library_data_t>, "Unsupported ArgT");
+            return (std::uint64_t)Val;
+        }
+
+        template <typename FirstT, typename... RestT>
+        inline constexpr std::uint64_t get_type_combination_id(FirstT FirstVal,
+                                                               RestT... RestVal)
+        {
+            static_assert((std::uint8_t)library_data_t::library_data_t_size <=
+                              std::numeric_limits<unsigned char>::max() &&
+                          "library_data_t size exceeds limit.");
+            static_assert(sizeof...(RestT) <= 8 && "Too many parameters");
+            static_assert(std::is_same_v<FirstT, library_data_t>, "Unsupported FirstT");
+            return get_type_combination_id(RestVal...) << 8 | ((std::uint64_t)FirstVal);
+        }
+
+        class mem_mgr
+        {
+            mem_mgr()
+            {
+                // Reserved address space, no real memory allocation happens here.
+#if defined(__linux__)
+                mapped_address_space =
+                    (byte_t *)mmap(nullptr, mapped_region_size, PROT_NONE,
+                                   MAP_PRIVATE | MAP_ANONYMOUS, -1, 0);
+#elif defined(_WIN64)
+                mapped_address_space = (byte_t *)VirtualAlloc(
+                    NULL,               // NULL specified as the base address parameter
+                    mapped_region_size, // Size of allocation
+                    MEM_RESERVE,        // Allocate reserved pages
+                    PAGE_NOACCESS);     // Protection = no access
+#else
+#error "Only support Windows and Linux."
+#endif
+                next_free = mapped_address_space;
+            }
+
+        public:
+            using buffer_id_t = int;
+
+            struct allocation
+            {
+                buffer_t buffer;
+                byte_t *alloc_ptr;
+                size_t size;
+            };
+
+            ~mem_mgr()
+            {
+#if defined(__linux__)
+                munmap(mapped_address_space, mapped_region_size);
+#elif defined(_WIN64)
+                VirtualFree(mapped_address_space, 0, MEM_RELEASE);
+#else
+#error "Only support Windows and Linux."
+#endif
+            }
+
+            mem_mgr(const mem_mgr &) = delete;
+            mem_mgr &operator=(const mem_mgr &) = delete;
+            mem_mgr(mem_mgr &&) = delete;
+            mem_mgr &operator=(mem_mgr &&) = delete;
+
+            /// Allocate
+            void *mem_alloc(size_t size)
+            {
+                if (!size)
+                    return nullptr;
+                std::lock_guard<std::mutex> lock(m_mutex);
+                if (next_free + size > mapped_address_space + mapped_region_size)
+                {
+                    throw std::runtime_error("dpct_malloc: out of memory for virtual memory pool");
+                }
+                // Allocation
+                sycl::range<1> r(size);
+                buffer_t buf(r);
+                allocation A{buf, next_free, size};
+                // Map allocation to device pointer
+                void *result = next_free;
+                m_map.emplace(next_free + size, A);
+                // Update pointer to the next free space.
+                next_free += (size + extra_padding + alignment - 1) & ~(alignment - 1);
+
+                return result;
+            }
+
+            /// Deallocate
+            void mem_free(const void *ptr)
+            {
+                if (!ptr)
+                    return;
+                std::lock_guard<std::mutex> lock(m_mutex);
+                auto it = get_map_iterator(ptr);
+                m_map.erase(it);
+            }
+
+            /// map: device pointer -> allocation(buffer, alloc_ptr, size)
+            allocation translate_ptr(const void *ptr)
+            {
+                std::lock_guard<std::mutex> lock(m_mutex);
+                auto it = get_map_iterator(ptr);
+                return it->second;
+            }
+
+            /// Check if the pointer represents device pointer or not.
+            bool is_device_ptr(const void *ptr) const
+            {
+                std::lock_guard<std::mutex> lock(m_mutex);
+                return (mapped_address_space <= ptr) &&
+                       (ptr < mapped_address_space + mapped_region_size);
+            }
+
+            /// Returns the instance of memory manager singleton.
+            static mem_mgr &instance()
+            {
+                static mem_mgr m;
+                return m;
+            }
+
+        private:
+            std::map<byte_t *, allocation> m_map;
+            mutable std::mutex m_mutex;
+            byte_t *mapped_address_space;
+            byte_t *next_free;
+            const size_t mapped_region_size = 128ull * 1024 * 1024 * 1024;
+            const size_t alignment = 256;
+            /// This padding may be defined to some positive value to debug
+            /// out of bound accesses.
+            const size_t extra_padding = 0;
+
+            std::map<byte_t *, allocation>::iterator get_map_iterator(const void *ptr)
+            {
+                auto it = m_map.upper_bound((byte_t *)ptr);
+                if (it == m_map.end())
+                {
+                    // Not a virtual pointer.
+                    throw std::runtime_error("can not get buffer from non-virtual pointer");
+                }
+                const allocation &alloc = it->second;
+                if (ptr < alloc.alloc_ptr)
+                {
+                    // Out of bound.
+                    // This may happen if there's a gap between allocations due to alignment
+                    // or extra padding and pointer points to this gap.
+                    throw std::runtime_error("invalid virtual pointer");
+                }
+                return it;
+            }
+        };
+
+        template <class T, memory_region Memory, size_t Dimension>
+        class accessor;
+        template <memory_region Memory, class T = byte_t>
+        class memory_traits
+        {
+        public:
+            static constexpr sycl::access::target target =
+                sycl::access::target::device;
+            static constexpr sycl::access_mode mode =
+                (Memory == constant) ? sycl::access_mode::read
+                                     : sycl::access_mode::read_write;
+            static constexpr size_t type_size = sizeof(T);
+            using element_t =
+                typename std::conditional<Memory == constant, const T, T>::type;
+            using value_t = typename std::remove_cv<T>::type;
+            template <size_t Dimension = 1>
+            using accessor_t = typename std::conditional<
+                Memory == local, sycl::local_accessor<value_t, Dimension>,
+                sycl::accessor<T, Dimension, mode, target>>::type;
+            using pointer_t = T *;
+        };
+
+        static inline void *dpct_malloc(size_t size, sycl::queue &q)
+        {
+            return sycl::malloc_device(size, q.get_device(), q.get_context());
+        }
+
+#define PITCH_DEFAULT_ALIGN(x) (((x) + 31) & ~(0x1F))
+        static inline void *dpct_malloc(size_t &pitch, size_t x, size_t y, size_t z,
+                                        sycl::queue &q)
+        {
+            pitch = PITCH_DEFAULT_ALIGN(x);
+            return dpct_malloc(pitch * y * z, q);
+        }
+
+        /**
+         * @brief Sets \p value to the first \p size elements starting from \p dev_ptr in \p q.
+         * @tparam valueT The type of the element to be set.
+         * @param [in] q The queue in which the operation is done.
+         * @param [in] dev_ptr Pointer to the virtual device memory address.
+         * @param [in] value The value to be set.
+         * @param [in] size Number of elements to be set to the value.
+         * @return An event representing the memset operation.
+         */
+        template <typename valueT>
+        static inline sycl::event dpct_memset(sycl::queue &q, void *dev_ptr,
+                                              valueT value, size_t size)
+        {
+            return q.fill(dev_ptr, value, size);
+        }
+
+        /**
+         * @brief Sets \p value to the 3D memory region pointed by \p data in \p q.
+         * @tparam valueT The type of the element to be set.
+         * @param [in] q The queue in which the operation is done.
+         * @param [in] data Pointer to the pitched device memory region.
+         * @param [in] value The value to be set.
+         * @param [in] size 3D memory region by number of elements.
+         * @return An event list representing the memset operations.
+         */
+        template <typename valueT>
+        static inline std::vector<sycl::event>
+        dpct_memset(sycl::queue &q, pitched_data data, valueT value,
+                    sycl::range<3> size)
+        {
+            std::vector<sycl::event> event_list;
+            size_t slice = data.get_pitch() * data.get_y();
+            unsigned char *data_surface = (unsigned char *)data.get_data_ptr();
+            for (size_t z = 0; z < size.get(2); ++z)
+            {
+                unsigned char *data_ptr = data_surface;
+                for (size_t y = 0; y < size.get(1); ++y)
+                {
+                    event_list.push_back(dpct_memset(q, data_ptr, value, size.get(0)));
+                    data_ptr += data.get_pitch();
+                }
+                data_surface += slice;
+            }
+            return event_list;
+        }
+
+        /**
+         * @brief Sets \p val to the pitched 2D memory region pointed by \p ptr in \p q.
+         * @tparam valueT The type of the element to be set.
+         * @param [in] q The queue in which the operation is done.
+         * @param [in] ptr Pointer to the virtual device memory.
+         * @param [in] pitch The pitch size by number of elements, including padding.
+         * @param [in] val The value to be set.
+         * @param [in] x The width of memory region by number of elements.
+         * @param [in] y The height of memory region by number of elements.
+         * @return An event list representing the memset operations.
+         */
+        template <typename valueT>
+        static inline std::vector<sycl::event>
+        dpct_memset(sycl::queue &q, void *ptr, size_t pitch, valueT val, size_t x,
+                    size_t y)
+        {
+            return dpct_memset(q, pitched_data(ptr, pitch, x, 1), val,
+                               sycl::range<3>(x, y, 1));
+        }
+
+        static memcpy_direction deduce_memcpy_direction(sycl::queue &q, void *to_ptr,
+                                                        const void *from_ptr,
+                                                        memcpy_direction dir)
+        {
+            switch (dir)
+            {
+            case memcpy_direction::host_to_host:
+            case memcpy_direction::host_to_device:
+            case memcpy_direction::device_to_host:
+            case memcpy_direction::device_to_device:
+                return dir;
+            case memcpy_direction::automatic:
+            {
+                // table[to_attribute][from_attribute]
+                static const memcpy_direction
+                    direction_table[static_cast<unsigned>(pointer_access_attribute::end)]
+                                   [static_cast<unsigned>(pointer_access_attribute::end)] =
+                                       {{memcpy_direction::host_to_host,
+                                         memcpy_direction::device_to_host,
+                                         memcpy_direction::host_to_host},
+                                        {memcpy_direction::host_to_device,
+                                         memcpy_direction::device_to_device,
+                                         memcpy_direction::device_to_device},
+                                        {memcpy_direction::host_to_host,
+                                         memcpy_direction::device_to_device,
+                                         memcpy_direction::device_to_device}};
+                return direction_table[static_cast<unsigned>(get_pointer_attribute(
+                    q, to_ptr))][static_cast<unsigned>(get_pointer_attribute(q, from_ptr))];
+            }
+            default:
+                throw std::runtime_error("dpct_memcpy: invalid direction value");
+            }
+        }
+
+        static sycl::event
+        dpct_memcpy(sycl::queue &q, void *to_ptr, const void *from_ptr, size_t size,
+                    memcpy_direction direction,
+                    const std::vector<sycl::event> &dep_events = {})
+        {
+            if (!size)
+                return sycl::event{};
+            return q.memcpy(to_ptr, from_ptr, size, dep_events);
+            GGML_UNUSED(direction);
+        }
+
+        // Get actual copy range and make sure it will not exceed range.
+        static inline size_t get_copy_range(sycl::range<3> size, size_t slice,
+                                            size_t pitch)
+        {
+            return slice * (size.get(2) - 1) + pitch * (size.get(1) - 1) + size.get(0);
+        }
+
+        static inline size_t get_offset(sycl::id<3> id, size_t slice,
+                                        size_t pitch)
+        {
+            return slice * id.get(2) + pitch * id.get(1) + id.get(0);
+        }
+
+        /// copy 3D matrix specified by \p size from 3D matrix specified by \p from_ptr
+        /// and \p from_range to another specified by \p to_ptr and \p to_range.
+        static inline std::vector<sycl::event>
+        dpct_memcpy(sycl::queue &q, void *to_ptr, const void *from_ptr,
+                    sycl::range<3> to_range, sycl::range<3> from_range,
+                    sycl::id<3> to_id, sycl::id<3> from_id,
+                    sycl::range<3> size, memcpy_direction direction,
+                    const std::vector<sycl::event> &dep_events = {})
+        {
+            // RAII for host pointer
+            class host_buffer
+            {
+                void *_buf;
+                size_t _size;
+                sycl::queue &_q;
+                const std::vector<sycl::event> &_deps; // free operation depends
+
+            public:
+                host_buffer(size_t size, sycl::queue &q,
+                            const std::vector<sycl::event> &deps)
+                    : _buf(std::malloc(size)), _size(size), _q(q), _deps(deps) {}
+                void *get_ptr() const { return _buf; }
+                size_t get_size() const { return _size; }
+                ~host_buffer()
+                {
+                    if (_buf)
+                    {
+                        _q.submit([&](sycl::handler &cgh)
+                                  {
+        cgh.depends_on(_deps);
+        cgh.host_task([buf = _buf] { std::free(buf); }); });
+                    }
+                }
+            };
+            std::vector<sycl::event> event_list;
+
+            size_t to_slice = to_range.get(1) * to_range.get(0),
+                   from_slice = from_range.get(1) * from_range.get(0);
+            unsigned char *to_surface =
+                (unsigned char *)to_ptr + get_offset(to_id, to_slice, to_range.get(0));
+            const unsigned char *from_surface =
+                (const unsigned char *)from_ptr +
+                get_offset(from_id, from_slice, from_range.get(0));
+
+            if (to_slice == from_slice && to_slice == size.get(1) * size.get(0))
+            {
+                return {dpct_memcpy(q, to_surface, from_surface, to_slice * size.get(2),
+                                    direction, dep_events)};
+            }
+            direction = deduce_memcpy_direction(q, to_ptr, from_ptr, direction);
+            size_t size_slice = size.get(1) * size.get(0);
+            switch (direction)
+            {
+            case host_to_host:
+                for (size_t z = 0; z < size.get(2); ++z)
+                {
+                    unsigned char *to_ptr = to_surface;
+                    const unsigned char *from_ptr = from_surface;
+                    if (to_range.get(0) == from_range.get(0) &&
+                        to_range.get(0) == size.get(0))
+                    {
+                        event_list.push_back(dpct_memcpy(q, to_ptr, from_ptr, size_slice,
+                                                         direction, dep_events));
+                    }
+                    else
+                    {
+                        for (size_t y = 0; y < size.get(1); ++y)
+                        {
+                            event_list.push_back(dpct_memcpy(q, to_ptr, from_ptr, size.get(0),
+                                                             direction, dep_events));
+                            to_ptr += to_range.get(0);
+                            from_ptr += from_range.get(0);
+                        }
+                    }
+                    to_surface += to_slice;
+                    from_surface += from_slice;
+                }
+                break;
+            case host_to_device:
+            {
+                host_buffer buf(get_copy_range(size, to_slice, to_range.get(0)), q,
+                                event_list);
+                std::vector<sycl::event> host_events;
+                if (to_slice == size_slice)
+                {
+                    // Copy host data to a temp host buffer with the shape of target.
+                    host_events =
+                        dpct_memcpy(q, buf.get_ptr(), from_surface, to_range, from_range,
+                                    sycl::id<3>(0, 0, 0), sycl::id<3>(0, 0, 0), size,
+                                    host_to_host, dep_events);
+                }
+                else
+                {
+                    // Copy host data to a temp host buffer with the shape of target.
+                    host_events = dpct_memcpy(
+                        q, buf.get_ptr(), from_surface, to_range, from_range,
+                        sycl::id<3>(0, 0, 0), sycl::id<3>(0, 0, 0), size, host_to_host,
+                        // If has padding data, not sure whether it is useless. So fill temp
+                        // buffer with it.
+                        std::vector<sycl::event>{
+                            dpct_memcpy(q, buf.get_ptr(), to_surface, buf.get_size(),
+                                        device_to_host, dep_events)});
+                }
+                // Copy from temp host buffer to device with only one submit.
+                event_list.push_back(dpct_memcpy(q, to_surface, buf.get_ptr(),
+                                                 buf.get_size(), host_to_device,
+                                                 host_events));
+                break;
+            }
+            case device_to_host:
+            {
+                host_buffer buf(get_copy_range(size, from_slice, from_range.get(0)), q,
+                                event_list);
+                // Copy from host temp buffer to host target with reshaping.
+                event_list = dpct_memcpy(
+                    q, to_surface, buf.get_ptr(), to_range, from_range, sycl::id<3>(0, 0, 0),
+                    sycl::id<3>(0, 0, 0), size, host_to_host,
+                    // Copy from device to temp host buffer with only one submit.
+                    std::vector<sycl::event>{dpct_memcpy(q, buf.get_ptr(), from_surface,
+                                                         buf.get_size(),
+                                                         device_to_host, dep_events)});
+                break;
+            }
+            case device_to_device:
+                event_list.push_back(q.submit([&](sycl::handler &cgh){
+                cgh.depends_on(dep_events);
+                cgh.parallel_for<class dpct_memcpy_3d_detail>(
+                    size,
+                    [=](sycl::id<3> id) {
+                        to_surface[get_offset(id, to_slice, to_range.get(0))] =
+                            from_surface[get_offset(id, from_slice, from_range.get(0))];
+                    }); }));
+                break;
+            default:
+                throw std::runtime_error("dpct_memcpy: invalid direction value");
+            }
+            return event_list;
+        }
+
+        /// memcpy 2D/3D matrix specified by pitched_data.
+        static inline std::vector<sycl::event>
+        dpct_memcpy(sycl::queue &q, pitched_data to, sycl::id<3> to_id,
+                    pitched_data from, sycl::id<3> from_id, sycl::range<3> size,
+                    memcpy_direction direction = automatic)
+        {
+            return dpct_memcpy(q, to.get_data_ptr(), from.get_data_ptr(),
+                               sycl::range<3>(to.get_pitch(), to.get_y(), 1),
+                               sycl::range<3>(from.get_pitch(), from.get_y(), 1), to_id, from_id,
+                               size, direction);
+        }
+
+        /// memcpy 2D matrix with pitch.
+        static inline std::vector<sycl::event>
+        dpct_memcpy(sycl::queue &q, void *to_ptr, const void *from_ptr,
+                    size_t to_pitch, size_t from_pitch, size_t x, size_t y,
+                    memcpy_direction direction = automatic)
+        {
+            return dpct_memcpy(q, to_ptr, from_ptr, sycl::range<3>(to_pitch, y, 1),
+                               sycl::range<3>(from_pitch, y, 1),
+                               sycl::id<3>(0, 0, 0), sycl::id<3>(0, 0, 0),
+                               sycl::range<3>(x, y, 1), direction);
+        }
+
+        namespace deprecated
+        {
+
+            template <typename T, sycl::usm::alloc AllocKind>
+            class usm_allocator
+            {
+            private:
+                using Alloc = sycl::usm_allocator<T, AllocKind>;
+                Alloc _impl;
+
+            public:
+                using value_type = typename std::allocator_traits<Alloc>::value_type;
+                using pointer = typename std::allocator_traits<Alloc>::pointer;
+                using const_pointer = typename std::allocator_traits<Alloc>::const_pointer;
+                using void_pointer = typename std::allocator_traits<Alloc>::void_pointer;
+                using const_void_pointer =
+                    typename std::allocator_traits<Alloc>::const_void_pointer;
+                using reference = typename std::allocator_traits<Alloc>::value_type &;
+                using const_reference =
+                    const typename std::allocator_traits<Alloc>::value_type &;
+                using difference_type =
+                    typename std::allocator_traits<Alloc>::difference_type;
+                using size_type = typename std::allocator_traits<Alloc>::size_type;
+                using propagate_on_container_copy_assignment = typename std::allocator_traits<
+                    Alloc>::propagate_on_container_copy_assignment;
+                using propagate_on_container_move_assignment = typename std::allocator_traits<
+                    Alloc>::propagate_on_container_move_assignment;
+                using propagate_on_container_swap =
+                    typename std::allocator_traits<Alloc>::propagate_on_container_swap;
+                using is_always_equal =
+                    typename std::allocator_traits<Alloc>::is_always_equal;
+
+                template <typename U>
+                struct rebind
+                {
+                    typedef usm_allocator<U, AllocKind> other;
+                };
+
+                usm_allocator() : _impl(dpct::get_default_queue()) {}
+                ~usm_allocator() {}
+                usm_allocator(const usm_allocator &other) : _impl(other._impl) {}
+                usm_allocator(usm_allocator &&other) : _impl(std::move(other._impl)) {}
+                pointer address(reference r) { return &r; }
+                const_pointer address(const_reference r) { return &r; }
+                pointer allocate(size_type cnt, const_void_pointer hint = nullptr)
+                {
+                    return std::allocator_traits<Alloc>::allocate(_impl, cnt, hint);
+                }
+                void deallocate(pointer p, size_type cnt)
+                {
+                    std::allocator_traits<Alloc>::deallocate(_impl, p, cnt);
+                }
+                size_type max_size() const
+                {
+                    return std::allocator_traits<Alloc>::max_size(_impl);
+                }
+                bool operator==(const usm_allocator &other) const { return _impl == other._impl; }
+                bool operator!=(const usm_allocator &other) const { return _impl != other._impl; }
+            };
+
+        } // namespace deprecated
+
+        inline void dpct_free(void *ptr,
+                              const sycl::queue &q)
+        {
+            if (ptr)
+            {
+                sycl::free(ptr, q.get_context());
+            }
+        }
+
+        template <typename T>
+        inline auto get_memory(const void *x)
+        {
+            T *new_x = reinterpret_cast<T *>(const_cast<void *>(x));
+            return new_x;
+        }
+
+        template <typename T>
+        inline typename DataType<T>::T2 get_value(const T *s, sycl::queue &q)
+        {
+            using Ty = typename DataType<T>::T2;
+            Ty s_h;
+            if (get_pointer_attribute(q, s) == pointer_access_attribute::device_only)
+                detail::dpct_memcpy(q, (void *)&s_h, (const void *)s, sizeof(T), device_to_host)
+                    .wait();
+            else
+                s_h = *reinterpret_cast<const Ty *>(s);
+            return s_h;
+        }
+
+    } // namespace detail
+
+    template <typename T>
+    inline auto get_value(const T *s, sycl::queue &q)
+    {
+        return detail::get_value(s, q);
+    }
+
+    namespace detail
+    {
+        template <class Ta, class Tb, class Tc, class Ts>
+        inline void gemm_impl(sycl::queue &q, oneapi::mkl::transpose a_trans,
+                              oneapi::mkl::transpose b_trans, int m, int n, int k,
+                              const void *alpha, const void *a, int lda, const void *b,
+                              int ldb, const void *beta, void *c, int ldc)
+        {
+            Ts alpha_value = dpct::get_value(reinterpret_cast<const Ts *>(alpha), q);
+            Ts beta_value = dpct::get_value(reinterpret_cast<const Ts *>(beta), q);
+            auto data_a = get_memory<const Ta>(a);
+            auto data_b = get_memory<const Tb>(b);
+            auto data_c = get_memory<Tc>(c);
+            oneapi::mkl::blas::column_major::gemm(
+                q, a_trans, b_trans, m, n, k, alpha_value, data_a, lda,
+                data_b, ldb, beta_value, data_c, ldc);
+        }
+
+        template <typename VecT, class BinaryOperation, class = void>
+        class vectorized_binary
+        {
+        public:
+            inline VecT operator()(VecT a, VecT b, const BinaryOperation binary_op)
+            {
+                VecT v4;
+                for (size_t i = 0; i < v4.size(); ++i)
+                {
+                    v4[i] = binary_op(a[i], b[i]);
+                }
+                return v4;
+            }
+        };
+
+        template <typename VecT, class BinaryOperation>
+        class vectorized_binary<
+            VecT, BinaryOperation,
+            std::void_t<std::invoke_result_t<BinaryOperation, VecT, VecT>>>
+        {
+        public:
+            inline VecT operator()(VecT a, VecT b, const BinaryOperation binary_op)
+            {
+                return binary_op(a, b).template as<VecT>();
+            }
+        };
+
+        template <class Ta, class Tb, class Tc, class Ts>
+        inline void gemm_batch_impl(sycl::queue &q, oneapi::mkl::transpose a_trans,
+                                    oneapi::mkl::transpose b_trans, int m, int n, int k,
+                                    const void *alpha, const void **a, int lda,
+                                    const void **b, int ldb, const void *beta, void **c,
+                                    int ldc, int batch_size)
+        {
+            struct matrix_info_t
+            {
+                oneapi::mkl::transpose transpose_info[2];
+                Ts value_info[2];
+                std::int64_t size_info[3];
+                std::int64_t ld_info[3];
+                std::int64_t groupsize_info;
+            };
+
+            Ts alpha_value = dpct::get_value(reinterpret_cast<const Ts *>(alpha), q);
+            Ts beta_value = dpct::get_value(reinterpret_cast<const Ts *>(beta), q);
+
+            matrix_info_t *matrix_info =
+                (matrix_info_t *)std::malloc(sizeof(matrix_info_t));
+            matrix_info->transpose_info[0] = a_trans;
+            matrix_info->transpose_info[1] = b_trans;
+            matrix_info->value_info[0] = alpha_value;
+            matrix_info->value_info[1] = beta_value;
+            matrix_info->size_info[0] = m;
+            matrix_info->size_info[1] = n;
+            matrix_info->size_info[2] = k;
+            matrix_info->ld_info[0] = lda;
+            matrix_info->ld_info[1] = ldb;
+            matrix_info->ld_info[2] = ldc;
+            matrix_info->groupsize_info = batch_size;
+
+            sycl::event e = oneapi::mkl::blas::column_major::gemm_batch(
+                q, matrix_info->transpose_info, matrix_info->transpose_info + 1,
+                matrix_info->size_info, matrix_info->size_info + 1,
+                matrix_info->size_info + 2, matrix_info->value_info,
+                reinterpret_cast<const Ta **>(a), matrix_info->ld_info,
+                reinterpret_cast<const Tb **>(b), matrix_info->ld_info + 1,
+                matrix_info->value_info + 1, reinterpret_cast<Tc **>(c),
+                matrix_info->ld_info + 2, 1, &(matrix_info->groupsize_info));
+
+            q.submit([&](sycl::handler &cgh)
+                     {
+    cgh.depends_on(e);
+    cgh.host_task([=] { std::free(matrix_info); }); });
+        }
+
+        template <class Ta, class Tb, class Tc, class Ts>
+        inline void
+        gemm_batch_impl(sycl::queue &q, oneapi::mkl::transpose a_trans,
+                        oneapi::mkl::transpose b_trans, int m, int n,
+                        int k, const void *alpha, const void *a, int lda,
+                        long long int stride_a, const void *b, int ldb,
+                        long long int stride_b, const void *beta, void *c,
+                        int ldc, long long int stride_c, int batch_size)
+        {
+            Ts alpha_value = dpct::get_value(reinterpret_cast<const Ts *>(alpha), q);
+            Ts beta_value = dpct::get_value(reinterpret_cast<const Ts *>(beta), q);
+            auto data_a = get_memory<const Ta>(a);
+            auto data_b = get_memory<const Tb>(b);
+            auto data_c = get_memory<Tc>(c);
+            oneapi::mkl::blas::column_major::gemm_batch(
+                q, a_trans, b_trans, m, n, k, alpha_value, data_a, lda,
+                stride_a, data_b, ldb, stride_b, beta_value,
+                data_c, ldc, stride_c, batch_size);
+        }
+
+    } // namespace detail
+
+    template <typename VecT, class BinaryOperation>
+    inline unsigned vectorized_binary(unsigned a, unsigned b,
+                                      const BinaryOperation binary_op)
+    {
+        sycl::vec<unsigned, 1> v0{a}, v1{b};
+        auto v2 = v0.as<VecT>();
+        auto v3 = v1.as<VecT>();
+        auto v4 =
+            detail::vectorized_binary<VecT, BinaryOperation>()(v2, v3, binary_op);
+        v0 = v4.template as<sycl::vec<unsigned, 1>>();
+        return v0;
+    }
+
+    static void async_dpct_memcpy(void *to_ptr, const void *from_ptr, size_t size,
+                                  memcpy_direction direction = automatic,
+                                  sycl::queue &q = dpct::get_default_queue())
+    {
+        detail::dpct_memcpy(q, to_ptr, from_ptr, size, direction);
+    }
+
+    static inline unsigned int select_device(unsigned int id)
+    {
+        dev_mgr::instance().select_device(id);
+        return id;
+    }
+
+    template <typename T>
+    T permute_sub_group_by_xor(sycl::sub_group g, T x, unsigned int mask,
+                               unsigned int logical_sub_group_size = 32)
+    {
+        unsigned int id = g.get_local_linear_id();
+        unsigned int start_index =
+            id / logical_sub_group_size * logical_sub_group_size;
+        unsigned int target_offset = (id % logical_sub_group_size) ^ mask;
+        return sycl::select_from_group(g, x,
+                                       target_offset < logical_sub_group_size
+                                           ? start_index + target_offset
+                                           : id);
+    }
+
+    template <typename T>
+    sycl::vec<T, 4> extract_and_sign_or_zero_extend4(T val)
+    {
+        return sycl::vec<T, 1>(val)
+            .template as<sycl::vec<
+                std::conditional_t<std::is_signed_v<T>, int8_t, uint8_t>, 4>>()
+            .template convert<T>();
+    }
+
+    template <typename T1, typename T2>
+    using dot_product_acc_t =
+        std::conditional_t<std::is_unsigned_v<T1> && std::is_unsigned_v<T2>,
+                           uint32_t, int32_t>;
+
+    template <typename T1, typename T2, typename T3>
+    inline auto dp4a(T1 a, T2 b, T3 c)
+    {
+        dot_product_acc_t<T1, T2> res = c;
+        auto va = extract_and_sign_or_zero_extend4(a);
+        auto vb = extract_and_sign_or_zero_extend4(b);
+        res += va[0] * vb[0];
+        res += va[1] * vb[1];
+        res += va[2] * vb[2];
+        res += va[3] * vb[3];
+        return res;
+    }
+
+    struct sub_sat
+    {
+        template <typename T>
+        auto operator()(const T x, const T y) const
+        {
+            return sycl::sub_sat(x, y);
+        }
+    };
+
+    template <typename S, typename T>
+    inline T vectorized_min(T a, T b)
+    {
+        sycl::vec<T, 1> v0{a}, v1{b};
+        auto v2 = v0.template as<S>();
+        auto v3 = v1.template as<S>();
+        auto v4 = sycl::min(v2, v3);
+        v0 = v4.template as<sycl::vec<T, 1>>();
+        return v0;
+    }
+
+    inline float pow(const float a, const int b) { return sycl::pown(a, b); }
+    inline double pow(const double a, const int b) { return sycl::pown(a, b); }
+    inline float pow(const float a, const float b) { return sycl::pow(a, b); }
+    inline double pow(const double a, const double b) { return sycl::pow(a, b); }
+    template <typename T, typename U>
+    inline typename std::enable_if_t<std::is_floating_point_v<T>, T>
+    pow(const T a, const U b)
+    {
+        return sycl::pow(a, static_cast<T>(b));
+    }
+    template <typename T, typename U>
+    inline typename std::enable_if_t<!std::is_floating_point_v<T>, double>
+    pow(const T a, const U b)
+    {
+        return sycl::pow(static_cast<double>(a), static_cast<double>(b));
+    }
+
+    inline double min(const double a, const float b)
+    {
+        return sycl::fmin(a, static_cast<double>(b));
+    }
+    inline double min(const float a, const double b)
+    {
+        return sycl::fmin(static_cast<double>(a), b);
+    }
+    inline float min(const float a, const float b) { return sycl::fmin(a, b); }
+    inline double min(const double a, const double b) { return sycl::fmin(a, b); }
+    inline std::uint32_t min(const std::uint32_t a, const std::int32_t b)
+    {
+        return sycl::min(a, static_cast<std::uint32_t>(b));
+    }
+    inline std::uint32_t min(const std::int32_t a, const std::uint32_t b)
+    {
+        return sycl::min(static_cast<std::uint32_t>(a), b);
+    }
+    inline std::int32_t min(const std::int32_t a, const std::int32_t b)
+    {
+        return sycl::min(a, b);
+    }
+    inline std::uint32_t min(const std::uint32_t a, const std::uint32_t b)
+    {
+        return sycl::min(a, b);
+    }
+    inline std::uint64_t min(const std::uint64_t a, const std::int64_t b)
+    {
+        return sycl::min(a, static_cast<std::uint64_t>(b));
+    }
+    inline std::uint64_t min(const std::int64_t a, const std::uint64_t b)
+    {
+        return sycl::min(static_cast<std::uint64_t>(a), b);
+    }
+    inline std::int64_t min(const std::int64_t a, const std::int64_t b)
+    {
+        return sycl::min(a, b);
+    }
+    inline std::uint64_t min(const std::uint64_t a, const std::uint64_t b)
+    {
+        return sycl::min(a, b);
+    }
+    inline std::uint64_t min(const std::uint64_t a, const std::int32_t b)
+    {
+        return sycl::min(a, static_cast<std::uint64_t>(b));
+    }
+    inline std::uint64_t min(const std::int32_t a, const std::uint64_t b)
+    {
+        return sycl::min(static_cast<std::uint64_t>(a), b);
+    }
+    inline std::uint64_t min(const std::uint64_t a, const std::uint32_t b)
+    {
+        return sycl::min(a, static_cast<std::uint64_t>(b));
+    }
+    inline std::uint64_t min(const std::uint32_t a, const std::uint64_t b)
+    {
+        return sycl::min(static_cast<std::uint64_t>(a), b);
+    }
+    // max function overloads.
+    // For floating-point types, `float` or `double` arguments are acceptable.
+    // For integer types, `std::uint32_t`, `std::int32_t`, `std::uint64_t` or
+    // `std::int64_t` type arguments are acceptable.
+    inline double max(const double a, const float b)
+    {
+        return sycl::fmax(a, static_cast<double>(b));
+    }
+    inline double max(const float a, const double b)
+    {
+        return sycl::fmax(static_cast<double>(a), b);
+    }
+    inline float max(const float a, const float b) { return sycl::fmax(a, b); }
+    inline double max(const double a, const double b) { return sycl::fmax(a, b); }
+    inline std::uint32_t max(const std::uint32_t a, const std::int32_t b)
+    {
+        return sycl::max(a, static_cast<std::uint32_t>(b));
+    }
+    inline std::uint32_t max(const std::int32_t a, const std::uint32_t b)
+    {
+        return sycl::max(static_cast<std::uint32_t>(a), b);
+    }
+    inline std::int32_t max(const std::int32_t a, const std::int32_t b)
+    {
+        return sycl::max(a, b);
+    }
+    inline std::uint32_t max(const std::uint32_t a, const std::uint32_t b)
+    {
+        return sycl::max(a, b);
+    }
+    inline std::uint64_t max(const std::uint64_t a, const std::int64_t b)
+    {
+        return sycl::max(a, static_cast<std::uint64_t>(b));
+    }
+    inline std::uint64_t max(const std::int64_t a, const std::uint64_t b)
+    {
+        return sycl::max(static_cast<std::uint64_t>(a), b);
+    }
+    inline std::int64_t max(const std::int64_t a, const std::int64_t b)
+    {
+        return sycl::max(a, b);
+    }
+    inline std::uint64_t max(const std::uint64_t a, const std::uint64_t b)
+    {
+        return sycl::max(a, b);
+    }
+    inline std::uint64_t max(const std::uint64_t a, const std::int32_t b)
+    {
+        return sycl::max(a, static_cast<std::uint64_t>(b));
+    }
+    inline std::uint64_t max(const std::int32_t a, const std::uint64_t b)
+    {
+        return sycl::max(static_cast<std::uint64_t>(a), b);
+    }
+    inline std::uint64_t max(const std::uint64_t a, const std::uint32_t b)
+    {
+        return sycl::max(a, static_cast<std::uint64_t>(b));
+    }
+    inline std::uint64_t max(const std::uint32_t a, const std::uint64_t b)
+    {
+        return sycl::max(static_cast<std::uint64_t>(a), b);
+    }
+
+    inline void
+    has_capability_or_fail(const sycl::device &dev,
+                           const std::initializer_list<sycl::aspect> &props)
+    {
+        for (const auto &it : props)
+        {
+            if (dev.has(it))
+                continue;
+            switch (it)
+            {
+            case sycl::aspect::fp64:
+                throw std::runtime_error("'double' is not supported in '" +
+                                         dev.get_info<sycl::info::device::name>() +
+                                         "' device");
+                break;
+            case sycl::aspect::fp16:
+                throw std::runtime_error("'half' is not supported in '" +
+                                         dev.get_info<sycl::info::device::name>() +
+                                         "' device");
+                break;
+            default:
+#define __SYCL_ASPECT(ASPECT, ID) \
+    case sycl::aspect::ASPECT:    \
+        return #ASPECT;
+#define __SYCL_ASPECT_DEPRECATED(ASPECT, ID, MESSAGE) __SYCL_ASPECT(ASPECT, ID)
+#define __SYCL_ASPECT_DEPRECATED_ALIAS(ASPECT, ID, MESSAGE)
+                auto getAspectNameStr = [](sycl::aspect AspectNum) -> std::string
+                {
+                    switch (AspectNum)
+                    {
+#include <sycl/info/aspects.def>
+#include <sycl/info/aspects_deprecated.def>
+                    default:
+                        return "unknown aspect";
+                    }
+                };
+#undef __SYCL_ASPECT_DEPRECATED_ALIAS
+#undef __SYCL_ASPECT_DEPRECATED
+#undef __SYCL_ASPECT
+                throw std::runtime_error(
+                    "'" + getAspectNameStr(it) + "' is not supported in '" +
+                    dev.get_info<sycl::info::device::name>() + "' device");
+            }
+            break;
+        }
+    }
+
+    static inline unsigned int get_current_device_id()
+    {
+        return dev_mgr::instance().current_device_id();
+    }
+
+    static inline device_ext &get_current_device()
+    {
+        return dev_mgr::instance().current_device();
+    }
+
+    static inline device_ext &get_device(unsigned int id)
+    {
+        return dev_mgr::instance().get_device(id);
+    }
+
+    static inline sycl::queue &get_in_order_queue()
+    {
+        return dev_mgr::instance().current_device().in_order_queue();
+    }
+
+    static sycl::event
+    dpct_memcpy(sycl::queue &q, void *to_ptr, const void *from_ptr, size_t size,
+                memcpy_direction direction,
+                const std::vector<sycl::event> &dep_events = {})
+    {
+        if (!size)
+            return sycl::event{};
+        return q.memcpy(to_ptr, from_ptr, size, dep_events);
+        GGML_UNUSED(direction);
+    }
+
+    // Get actual copy range and make sure it will not exceed range.
+    static inline size_t get_copy_range(sycl::range<3> size, size_t slice,
+                                        size_t pitch)
+    {
+        return slice * (size.get(2) - 1) + pitch * (size.get(1) - 1) + size.get(0);
+    }
+
+    static inline size_t get_offset(sycl::id<3> id, size_t slice,
+                                    size_t pitch)
+    {
+        return slice * id.get(2) + pitch * id.get(1) + id.get(0);
+    }
+
+    /// copy 3D matrix specified by \p size from 3D matrix specified by \p from_ptr
+    /// and \p from_range to another specified by \p to_ptr and \p to_range.
+    static inline std::vector<sycl::event>
+    dpct_memcpy(sycl::queue &q, void *to_ptr, const void *from_ptr,
+                sycl::range<3> to_range, sycl::range<3> from_range,
+                sycl::id<3> to_id, sycl::id<3> from_id,
+                sycl::range<3> size, memcpy_direction direction,
+                const std::vector<sycl::event> &dep_events = {})
+    {
+        // RAII for host pointer
+        class host_buffer
+        {
+            void *_buf;
+            size_t _size;
+            sycl::queue &_q;
+            const std::vector<sycl::event> &_deps; // free operation depends
+
+        public:
+            host_buffer(size_t size, sycl::queue &q,
+                        const std::vector<sycl::event> &deps)
+                : _buf(std::malloc(size)), _size(size), _q(q), _deps(deps) {}
+            void *get_ptr() const { return _buf; }
+            size_t get_size() const { return _size; }
+            ~host_buffer()
+            {
+                if (_buf)
+                {
+                    _q.submit([&](sycl::handler &cgh)
+                              {
+            cgh.depends_on(_deps);
+            cgh.host_task([buf = _buf] { std::free(buf); }); });
+                }
+            }
+        };
+        std::vector<sycl::event> event_list;
+
+        size_t to_slice = to_range.get(1) * to_range.get(0),
+               from_slice = from_range.get(1) * from_range.get(0);
+        unsigned char *to_surface =
+            (unsigned char *)to_ptr + get_offset(to_id, to_slice, to_range.get(0));
+        const unsigned char *from_surface =
+            (const unsigned char *)from_ptr +
+            get_offset(from_id, from_slice, from_range.get(0));
+
+        if (to_slice == from_slice && to_slice == size.get(1) * size.get(0))
+        {
+            return {dpct_memcpy(q, to_surface, from_surface, to_slice * size.get(2),
+                                direction, dep_events)};
+        }
+        direction = detail::deduce_memcpy_direction(q, to_ptr, from_ptr, direction);
+        size_t size_slice = size.get(1) * size.get(0);
+        switch (direction)
+        {
+        case host_to_host:
+            for (size_t z = 0; z < size.get(2); ++z)
+            {
+                unsigned char *to_ptr = to_surface;
+                const unsigned char *from_ptr = from_surface;
+                if (to_range.get(0) == from_range.get(0) &&
+                    to_range.get(0) == size.get(0))
+                {
+                    event_list.push_back(dpct_memcpy(q, to_ptr, from_ptr, size_slice,
+                                                     direction, dep_events));
+                }
+                else
+                {
+                    for (size_t y = 0; y < size.get(1); ++y)
+                    {
+                        event_list.push_back(dpct_memcpy(q, to_ptr, from_ptr, size.get(0),
+                                                         direction, dep_events));
+                        to_ptr += to_range.get(0);
+                        from_ptr += from_range.get(0);
+                    }
+                }
+                to_surface += to_slice;
+                from_surface += from_slice;
+            }
+            break;
+        case host_to_device:
+        {
+            host_buffer buf(get_copy_range(size, to_slice, to_range.get(0)), q,
+                            event_list);
+            std::vector<sycl::event> host_events;
+            if (to_slice == size_slice)
+            {
+                // Copy host data to a temp host buffer with the shape of target.
+                host_events =
+                    dpct_memcpy(q, buf.get_ptr(), from_surface, to_range, from_range,
+                                sycl::id<3>(0, 0, 0), sycl::id<3>(0, 0, 0), size,
+                                host_to_host, dep_events);
+            }
+            else
+            {
+                // Copy host data to a temp host buffer with the shape of target.
+                host_events = dpct_memcpy(
+                    q, buf.get_ptr(), from_surface, to_range, from_range,
+                    sycl::id<3>(0, 0, 0), sycl::id<3>(0, 0, 0), size, host_to_host,
+                    // If has padding data, not sure whether it is useless. So fill temp
+                    // buffer with it.
+                    std::vector<sycl::event>{
+                        dpct_memcpy(q, buf.get_ptr(), to_surface, buf.get_size(),
+                                    device_to_host, dep_events)});
+            }
+            // Copy from temp host buffer to device with only one submit.
+            event_list.push_back(dpct_memcpy(q, to_surface, buf.get_ptr(),
+                                             buf.get_size(), host_to_device,
+                                             host_events));
+            break;
+        }
+        case device_to_host:
+        {
+            host_buffer buf(get_copy_range(size, from_slice, from_range.get(0)), q,
+                            event_list);
+            // Copy from host temp buffer to host target with reshaping.
+            event_list = dpct_memcpy(
+                q, to_surface, buf.get_ptr(), to_range, from_range, sycl::id<3>(0, 0, 0),
+                sycl::id<3>(0, 0, 0), size, host_to_host,
+                // Copy from device to temp host buffer with only one submit.
+                std::vector<sycl::event>{dpct_memcpy(q, buf.get_ptr(), from_surface,
+                                                     buf.get_size(),
+                                                     device_to_host, dep_events)});
+            break;
+        }
+        case device_to_device:
+            event_list.push_back(q.submit([&](sycl::handler &cgh)
+                                          {
+        cgh.depends_on(dep_events);
+        cgh.parallel_for<class dpct_memcpy_3d_detail>(
+            size,
+            [=](sycl::id<3> id) {
+                to_surface[get_offset(id, to_slice, to_range.get(0))] =
+                    from_surface[get_offset(id, from_slice, from_range.get(0))];
+            }); }));
+        break;
+        default:
+            throw std::runtime_error("dpct_memcpy: invalid direction value");
+        }
+        return event_list;
+    }
+
+    /// memcpy 2D/3D matrix specified by pitched_data.
+    static inline std::vector<sycl::event>
+    dpct_memcpy(sycl::queue &q, pitched_data to, sycl::id<3> to_id,
+                pitched_data from, sycl::id<3> from_id, sycl::range<3> size,
+                memcpy_direction direction = automatic)
+    {
+        return dpct_memcpy(q, to.get_data_ptr(), from.get_data_ptr(),
+                           sycl::range<3>(to.get_pitch(), to.get_y(), 1),
+                           sycl::range<3>(from.get_pitch(), from.get_y(), 1), to_id, from_id,
+                           size, direction);
+    }
+
+    /// memcpy 2D matrix with pitch.
+    static inline std::vector<sycl::event>
+    dpct_memcpy(sycl::queue &q, void *to_ptr, const void *from_ptr,
+                size_t to_pitch, size_t from_pitch, size_t x, size_t y,
+                memcpy_direction direction = automatic)
+    {
+        return dpct_memcpy(q, to_ptr, from_ptr, sycl::range<3>(to_pitch, y, 1),
+                           sycl::range<3>(from_pitch, y, 1),
+                           sycl::id<3>(0, 0, 0), sycl::id<3>(0, 0, 0),
+                           sycl::range<3>(x, y, 1), direction);
+    }
+
+    inline void gemm(sycl::queue &q, oneapi::mkl::transpose a_trans,
+                     oneapi::mkl::transpose b_trans, int m, int n, int k,
+                     const void *alpha, const void *a, library_data_t a_type,
+                     int lda, const void *b, library_data_t b_type, int ldb,
+                     const void *beta, void *c, library_data_t c_type, int ldc,
+                     library_data_t scaling_type)
+    {
+        if (scaling_type == library_data_t::real_float &&
+            c_type == library_data_t::complex_float)
+        {
+            scaling_type = library_data_t::complex_float;
+        }
+        else if (scaling_type == library_data_t::real_double &&
+                 c_type == library_data_t::complex_double)
+        {
+            scaling_type = library_data_t::complex_double;
+        }
+
+        std::uint64_t key =
+            detail::get_type_combination_id(a_type, b_type, c_type, scaling_type);
+        switch (key)
+        {
+        case detail::get_type_combination_id(
+            library_data_t::real_float, library_data_t::real_float,
+            library_data_t::real_float, library_data_t::real_float):
+        {
+            detail::gemm_impl<float, float, float, float>(
+                q, a_trans, b_trans, m, n, k, alpha, a, lda, b, ldb, beta, c, ldc);
+            break;
+        }
+        case detail::get_type_combination_id(
+            library_data_t::real_double, library_data_t::real_double,
+            library_data_t::real_double, library_data_t::real_double):
+        {
+            detail::gemm_impl<double, double, double, double>(
+                q, a_trans, b_trans, m, n, k, alpha, a, lda, b, ldb, beta, c, ldc);
+            break;
+        }
+        case detail::get_type_combination_id(
+            library_data_t::complex_float, library_data_t::complex_float,
+            library_data_t::complex_float, library_data_t::complex_float):
+        {
+            detail::gemm_impl<std::complex<float>, std::complex<float>,
+                              std::complex<float>, std::complex<float>>(
+                q, a_trans, b_trans, m, n, k, alpha, a, lda, b, ldb, beta, c, ldc);
+            break;
+        }
+        case detail::get_type_combination_id(
+            library_data_t::complex_double, library_data_t::complex_double,
+            library_data_t::complex_double, library_data_t::complex_double):
+        {
+            detail::gemm_impl<std::complex<double>, std::complex<double>,
+                              std::complex<double>, std::complex<double>>(
+                q, a_trans, b_trans, m, n, k, alpha, a, lda, b, ldb, beta, c, ldc);
+            break;
+        }
+        case detail::get_type_combination_id(
+            library_data_t::real_half, library_data_t::real_half,
+            library_data_t::real_half, library_data_t::real_half):
+        {
+            detail::gemm_impl<sycl::half, sycl::half, sycl::half,
+                              sycl::half>(q, a_trans, b_trans, m, n, k, alpha, a,
+                                          lda, b, ldb, beta, c, ldc);
+            break;
+        }
+#ifdef __INTEL_MKL__
+        case detail::get_type_combination_id(
+            library_data_t::real_bfloat16, library_data_t::real_bfloat16,
+            library_data_t::real_float, library_data_t::real_float):
+        {
+            detail::gemm_impl<oneapi::mkl::bfloat16, oneapi::mkl::bfloat16, float,
+                              float>(q, a_trans, b_trans, m, n, k, alpha, a, lda, b,
+                                     ldb, beta, c, ldc);
+            break;
+        }
+        case detail::get_type_combination_id(
+            library_data_t::real_half, library_data_t::real_half,
+            library_data_t::real_float, library_data_t::real_float):
+        {
+            detail::gemm_impl<sycl::half, sycl::half, float, float>(
+                q, a_trans, b_trans, m, n, k, alpha, a, lda, b, ldb, beta, c, ldc);
+            break;
+        }
+        case detail::get_type_combination_id(
+            library_data_t::real_half, library_data_t::real_half,
+            library_data_t::real_half, library_data_t::real_float):
+        {
+            float alpha_value =
+                dpct::get_value(reinterpret_cast<const float *>(alpha), q);
+            float beta_value =
+                dpct::get_value(reinterpret_cast<const float *>(beta), q);
+            sycl::half alpha_half(alpha_value);
+            sycl::half beta_half(beta_value);
+            detail::gemm_impl<sycl::half, sycl::half, sycl::half,
+                              sycl::half>(q, a_trans, b_trans, m, n, k, &alpha_half,
+                                          a, lda, b, ldb, &beta_half, c, ldc);
+            break;
+        }
+        case detail::get_type_combination_id(
+            library_data_t::real_int8, library_data_t::real_int8,
+            library_data_t::real_float, library_data_t::real_float):
+        {
+            detail::gemm_impl<std::int8_t, std::int8_t, float, float>(
+                q, a_trans, b_trans, m, n, k, alpha, a, lda, b, ldb, beta, c, ldc);
+            break;
+        }
+        case detail::get_type_combination_id(
+            library_data_t::real_bfloat16, library_data_t::real_bfloat16,
+            library_data_t::real_bfloat16, library_data_t::real_float):
+        {
+            detail::gemm_impl<oneapi::mkl::bfloat16, oneapi::mkl::bfloat16,
+                              oneapi::mkl::bfloat16, float>(
+                q, a_trans, b_trans, m, n, k, alpha, a, lda, b, ldb, beta, c, ldc);
+            break;
+        }
+        case detail::get_type_combination_id(
+            library_data_t::real_int8, library_data_t::real_int8,
+            library_data_t::real_int32, library_data_t::real_int32):
+        {
+            float alpha_float =
+                dpct::get_value(reinterpret_cast<const std::int32_t *>(alpha), q);
+            float beta_float =
+                dpct::get_value(reinterpret_cast<const std::int32_t *>(beta), q);
+            detail::gemm_impl<std::int8_t, std::int8_t, std::int32_t, float>(
+                q, a_trans, b_trans, m, n, k, &alpha_float, a, lda, b, ldb, &beta_float, c, ldc);
+            break;
+        }
+#endif // __INTEL_MKL__
+        default:
+            throw std::runtime_error("the combination of data type is unsupported");
+        }
+    } // gemm()
+
+    /// Computes a batch of matrix-matrix product with general matrices.
+    /// \param [in] q The queue where the routine should be executed.
+    /// \param [in] a_trans Specifies the operation applied to A.
+    /// \param [in] b_trans Specifies the operation applied to B.
+    /// \param [in] m Specifies the number of rows of the matrix op(A) and of the matrix C.
+    /// \param [in] n Specifies the number of columns of the matrix op(B) and of the matrix C.
+    /// \param [in] k Specifies the number of columns of the matrix op(A) and the number of rows of the matrix op(B).
+    /// \param [in] alpha Scaling factor for the matrix-matrix product.
+    /// \param [in] a Input matrix A.
+    /// \param [in] a_type Data type of the matrix A.
+    /// \param [in] lda Leading dimension of A.
+    /// \param [in] b Input matrix B.
+    /// \param [in] b_type Data type of the matrix B.
+    /// \param [in] ldb Leading dimension of B.
+    /// \param [in] beta Scaling factor for matrix C.
+    /// \param [in, out] c Input/Output matrix C.
+    /// \param [in] c_type Data type of the matrix C.
+    /// \param [in] ldc Leading dimension of C.
+    /// \param [in] batch_size Specifies the number of matrix multiply operations to perform.
+    /// \param [in] scaling_type Data type of the scaling factors.
+    inline void gemm_batch(sycl::queue &q, oneapi::mkl::transpose a_trans,
+                           oneapi::mkl::transpose b_trans, int m, int n, int k,
+                           const void *alpha, const void *a[],
+                           library_data_t a_type, int lda, const void *b[],
+                           library_data_t b_type, int ldb, const void *beta,
+                           void *c[], library_data_t c_type, int ldc,
+                           int batch_size, library_data_t scaling_type)
+    {
+        if (scaling_type == library_data_t::real_float &&
+            c_type == library_data_t::complex_float)
+        {
+            scaling_type = library_data_t::complex_float;
+        }
+        else if (scaling_type == library_data_t::real_double &&
+                 c_type == library_data_t::complex_double)
+        {
+            scaling_type = library_data_t::complex_double;
+        }
+
+        std::uint64_t key =
+            detail::get_type_combination_id(a_type, b_type, c_type, scaling_type);
+        switch (key)
+        {
+        case detail::get_type_combination_id(
+            library_data_t::real_float, library_data_t::real_float,
+            library_data_t::real_float, library_data_t::real_float):
+        {
+            detail::gemm_batch_impl<float, float, float, float>(
+                q, a_trans, b_trans, m, n, k, alpha, a, lda, b, ldb, beta, c, ldc,
+                batch_size);
+            break;
+        }
+        case detail::get_type_combination_id(
+            library_data_t::real_double, library_data_t::real_double,
+            library_data_t::real_double, library_data_t::real_double):
+        {
+            detail::gemm_batch_impl<double, double, double, double>(
+                q, a_trans, b_trans, m, n, k, alpha, a, lda, b, ldb, beta, c, ldc,
+                batch_size);
+            break;
+        }
+        case detail::get_type_combination_id(
+            library_data_t::complex_float, library_data_t::complex_float,
+            library_data_t::complex_float, library_data_t::complex_float):
+        {
+            detail::gemm_batch_impl<std::complex<float>, std::complex<float>,
+                                    std::complex<float>, std::complex<float>>(
+                q, a_trans, b_trans, m, n, k, alpha, a, lda, b, ldb, beta, c, ldc,
+                batch_size);
+            break;
+        }
+        case detail::get_type_combination_id(
+            library_data_t::complex_double, library_data_t::complex_double,
+            library_data_t::complex_double, library_data_t::complex_double):
+        {
+            detail::gemm_batch_impl<std::complex<double>, std::complex<double>,
+                                    std::complex<double>, std::complex<double>>(
+                q, a_trans, b_trans, m, n, k, alpha, a, lda, b, ldb, beta, c, ldc,
+                batch_size);
+            break;
+        }
+        case detail::get_type_combination_id(
+            library_data_t::real_half, library_data_t::real_half,
+            library_data_t::real_half, library_data_t::real_half):
+        {
+            detail::gemm_batch_impl<sycl::half, sycl::half, sycl::half,
+                                    sycl::half>(q, a_trans, b_trans, m, n, k, alpha,
+                                                a, lda, b, ldb, beta, c, ldc,
+                                                batch_size);
+            break;
+        }
+#ifdef __INTEL_MKL__
+        case detail::get_type_combination_id(
+            library_data_t::real_bfloat16, library_data_t::real_bfloat16,
+            library_data_t::real_bfloat16, library_data_t::real_float):
+        {
+            detail::gemm_batch_impl<oneapi::mkl::bfloat16, oneapi::mkl::bfloat16,
+                                    oneapi::mkl::bfloat16, float>(
+                q, a_trans, b_trans, m, n, k, alpha, a, lda, b, ldb, beta, c, ldc,
+                batch_size);
+            break;
+        }
+        case detail::get_type_combination_id(
+            library_data_t::real_bfloat16, library_data_t::real_bfloat16,
+            library_data_t::real_float, library_data_t::real_float):
+        {
+            detail::gemm_batch_impl<oneapi::mkl::bfloat16, oneapi::mkl::bfloat16, float,
+                                    float>(q, a_trans, b_trans, m, n, k, alpha, a, lda,
+                                           b, ldb, beta, c, ldc, batch_size);
+            break;
+        }
+#endif
+        case detail::get_type_combination_id(
+            library_data_t::real_int8, library_data_t::real_int8,
+            library_data_t::real_int32, library_data_t::real_int32):
+        {
+            float alpha_float =
+                dpct::get_value(reinterpret_cast<const std::int32_t *>(alpha), q);
+            float beta_float =
+                dpct::get_value(reinterpret_cast<const std::int32_t *>(beta), q);
+            detail::gemm_batch_impl<std::int8_t, std::int8_t, std::int32_t,
+                                    float>(q, a_trans, b_trans, m, n, k, &alpha_float,
+                                           a, lda, b, ldb, &beta_float, c, ldc,
+                                           batch_size);
+            break;
+        }
+        case detail::get_type_combination_id(
+            library_data_t::real_int8, library_data_t::real_int8,
+            library_data_t::real_float, library_data_t::real_float):
+        {
+            detail::gemm_batch_impl<std::int8_t, std::int8_t, float, float>(
+                q, a_trans, b_trans, m, n, k, alpha, a, lda, b, ldb, beta, c, ldc,
+                batch_size);
+            break;
+        }
+        case detail::get_type_combination_id(
+            library_data_t::real_half, library_data_t::real_half,
+            library_data_t::real_float, library_data_t::real_float):
+        {
+            detail::gemm_batch_impl<sycl::half, sycl::half, float, float>(
+                q, a_trans, b_trans, m, n, k, alpha, a, lda, b, ldb, beta, c, ldc,
+                batch_size);
+            break;
+        }
+        case detail::get_type_combination_id(
+            library_data_t::real_half, library_data_t::real_half,
+            library_data_t::real_half, library_data_t::real_float):
+        {
+            float alpha_value =
+                dpct::get_value(reinterpret_cast<const float *>(alpha), q);
+            float beta_value =
+                dpct::get_value(reinterpret_cast<const float *>(beta), q);
+            sycl::half alpha_half(alpha_value);
+            sycl::half beta_half(beta_value);
+            detail::gemm_batch_impl<sycl::half, sycl::half, sycl::half, sycl::half>(
+                q, a_trans, b_trans, m, n, k, &alpha_half, a, lda, b, ldb, &beta_half, c, ldc,
+                batch_size);
+            break;
+        }
+        default:
+            throw std::runtime_error("the combination of data type is unsupported");
+        }
+    }
+
+    /// Computes a batch of matrix-matrix product with general matrices.
+    /// \param [in] q The queue where the routine should be executed.
+    /// \param [in] a_trans Specifies the operation applied to A.
+    /// \param [in] b_trans Specifies the operation applied to B.
+    /// \param [in] m Specifies the number of rows of the matrix op(A) and of the matrix C.
+    /// \param [in] n Specifies the number of columns of the matrix op(B) and of the matrix C.
+    /// \param [in] k Specifies the number of columns of the matrix op(A) and the number of rows of the matrix op(B).
+    /// \param [in] alpha Scaling factor for the matrix-matrix product.
+    /// \param [in] a Input matrix A.
+    /// \param [in] a_type Data type of the matrix A.
+    /// \param [in] lda Leading dimension of A.
+    /// \param [in] stride_a Stride between the different A matrices.
+    /// \param [in] b Input matrix B.
+    /// \param [in] b_type Data type of the matrix B.
+    /// \param [in] ldb Leading dimension of B.
+    /// \param [in] stride_b Stride between the different B matrices.
+    /// \param [in] beta Scaling factor for matrix C.
+    /// \param [in, out] c Input/Output matrix C.
+    /// \param [in] c_type Data type of the matrix C.
+    /// \param [in] ldc Leading dimension of C.
+    /// \param [in] stride_c Stride between the different C matrices.
+    /// \param [in] batch_size Specifies the number of matrix multiply operations to perform.
+    /// \param [in] scaling_type Data type of the scaling factors.
+    inline void gemm_batch(sycl::queue &q, oneapi::mkl::transpose a_trans,
+                           oneapi::mkl::transpose b_trans, int m, int n, int k,
+                           const void *alpha, const void *a, library_data_t a_type,
+                           int lda, long long int stride_a, const void *b,
+                           library_data_t b_type, int ldb, long long int stride_b,
+                           const void *beta, void *c, library_data_t c_type,
+                           int ldc, long long int stride_c, int batch_size,
+                           library_data_t scaling_type)
+    {
+        if (scaling_type == library_data_t::real_float &&
+            c_type == library_data_t::complex_float)
+        {
+            scaling_type = library_data_t::complex_float;
+        }
+        else if (scaling_type == library_data_t::real_double &&
+                 c_type == library_data_t::complex_double)
+        {
+            scaling_type = library_data_t::complex_double;
+        }
+
+        std::uint64_t key =
+            detail::get_type_combination_id(a_type, b_type, c_type, scaling_type);
+        switch (key)
+        {
+        case detail::get_type_combination_id(
+            library_data_t::real_float, library_data_t::real_float,
+            library_data_t::real_float, library_data_t::real_float):
+        {
+            detail::gemm_batch_impl<float, float, float, float>(
+                q, a_trans, b_trans, m, n, k, alpha, a, lda, stride_a, b, ldb, stride_b,
+                beta, c, ldc, stride_c, batch_size);
+            break;
+        }
+        case detail::get_type_combination_id(
+            library_data_t::real_double, library_data_t::real_double,
+            library_data_t::real_double, library_data_t::real_double):
+        {
+            detail::gemm_batch_impl<double, double, double, double>(
+                q, a_trans, b_trans, m, n, k, alpha, a, lda, stride_a, b, ldb, stride_b,
+                beta, c, ldc, stride_c, batch_size);
+            break;
+        }
+        case detail::get_type_combination_id(
+            library_data_t::complex_float, library_data_t::complex_float,
+            library_data_t::complex_float, library_data_t::complex_float):
+        {
+            detail::gemm_batch_impl<std::complex<float>, std::complex<float>,
+                                    std::complex<float>, std::complex<float>>(
+                q, a_trans, b_trans, m, n, k, alpha, a, lda, stride_a, b, ldb, stride_b,
+                beta, c, ldc, stride_c, batch_size);
+            break;
+        }
+        case detail::get_type_combination_id(
+            library_data_t::complex_double, library_data_t::complex_double,
+            library_data_t::complex_double, library_data_t::complex_double):
+        {
+            detail::gemm_batch_impl<std::complex<double>, std::complex<double>,
+                                    std::complex<double>, std::complex<double>>(
+                q, a_trans, b_trans, m, n, k, alpha, a, lda, stride_a, b, ldb, stride_b,
+                beta, c, ldc, stride_c, batch_size);
+            break;
+        }
+        case detail::get_type_combination_id(
+            library_data_t::real_half, library_data_t::real_half,
+            library_data_t::real_half, library_data_t::real_half):
+        {
+            detail::gemm_batch_impl<sycl::half, sycl::half, sycl::half,
+                                    sycl::half>(q, a_trans, b_trans, m, n, k, alpha,
+                                                a, lda, stride_a, b, ldb, stride_b,
+                                                beta, c, ldc, stride_c, batch_size);
+            break;
+        }
+#ifdef __INTEL_MKL__
+        case detail::get_type_combination_id(
+            library_data_t::real_bfloat16, library_data_t::real_bfloat16,
+            library_data_t::real_bfloat16, library_data_t::real_float):
+        {
+            detail::gemm_batch_impl<oneapi::mkl::bfloat16, oneapi::mkl::bfloat16,
+                                    oneapi::mkl::bfloat16, float>(
+                q, a_trans, b_trans, m, n, k, alpha, a, lda, stride_a, b, ldb, stride_b,
+                beta, c, ldc, stride_c, batch_size);
+            break;
+        }
+        case detail::get_type_combination_id(
+            library_data_t::real_bfloat16, library_data_t::real_bfloat16,
+            library_data_t::real_float, library_data_t::real_float):
+        {
+            detail::gemm_batch_impl<oneapi::mkl::bfloat16, oneapi::mkl::bfloat16, float,
+                                    float>(q, a_trans, b_trans, m, n, k, alpha, a, lda,
+                                           stride_a, b, ldb, stride_b, beta, c, ldc,
+                                           stride_c, batch_size);
+            break;
+        }
+#endif
+        case detail::get_type_combination_id(
+            library_data_t::real_int8, library_data_t::real_int8,
+            library_data_t::real_int32, library_data_t::real_int32):
+        {
+            detail::gemm_batch_impl<std::int8_t, std::int8_t, std::int32_t,
+                                    std::int32_t>(q, a_trans, b_trans, m, n, k, alpha,
+                                                  a, lda, stride_a, b, ldb, stride_b,
+                                                  beta, c, ldc, stride_c, batch_size);
+            break;
+        }
+        case detail::get_type_combination_id(
+            library_data_t::real_int8, library_data_t::real_int8,
+            library_data_t::real_float, library_data_t::real_float):
+        {
+            detail::gemm_batch_impl<std::int8_t, std::int8_t, float, float>(
+                q, a_trans, b_trans, m, n, k, alpha, a, lda, stride_a, b, ldb, stride_b,
+                beta, c, ldc, stride_c, batch_size);
+            break;
+        }
+        case detail::get_type_combination_id(
+            library_data_t::real_half, library_data_t::real_half,
+            library_data_t::real_float, library_data_t::real_float):
+        {
+            detail::gemm_batch_impl<sycl::half, sycl::half, float, float>(
+                q, a_trans, b_trans, m, n, k, alpha, a, lda, stride_a, b, ldb, stride_b,
+                beta, c, ldc, stride_c, batch_size);
+            break;
+        }
+        case detail::get_type_combination_id(
+            library_data_t::real_half, library_data_t::real_half,
+            library_data_t::real_half, library_data_t::real_float):
+        {
+            float alpha_value =
+                dpct::get_value(reinterpret_cast<const float *>(alpha), q);
+            float beta_value =
+                dpct::get_value(reinterpret_cast<const float *>(beta), q);
+            sycl::half alpha_half(alpha_value);
+            sycl::half beta_half(beta_value);
+            detail::gemm_batch_impl<sycl::half, sycl::half, sycl::half, sycl::half>(
+                q, a_trans, b_trans, m, n, k, &alpha_half, a, lda, stride_a, b, ldb, stride_b,
+                &beta_half, c, ldc, stride_c, batch_size);
+            break;
+        }
+        default:
+            throw std::runtime_error("the combination of data type is unsupported");
+        }
+    }
+
+    static inline void
+    async_dpct_memcpy(void *to_ptr, size_t to_pitch, const void *from_ptr,
+                      size_t from_pitch, size_t x, size_t y,
+                      memcpy_direction direction = automatic,
+                      sycl::queue &q = get_default_queue())
+    {
+        detail::dpct_memcpy(q, to_ptr, from_ptr, to_pitch, from_pitch, x, y,
+                            direction);
+    }
+
+    using err0 = detail::generic_error_type<struct err0_tag, int>;
+    using err1 = detail::generic_error_type<struct err1_tag, int>;
+
+    static inline void dpct_free(void *ptr, sycl::queue &q = get_default_queue()) {
+        detail::dpct_free(ptr, q);
+    }
+
+    /// dpct accessor used as device function parameter.
+    template <class T, memory_region Memory, size_t Dimension> class accessor;
+    template <class T, memory_region Memory> class accessor<T, Memory, 3> {
+    public:
+        using memory_t = detail::memory_traits<Memory, T>;
+        using element_t = typename memory_t::element_t;
+        using pointer_t = typename memory_t::pointer_t;
+        using accessor_t = typename memory_t::template accessor_t<3>;
+        accessor(pointer_t data, const sycl::range<3> &in_range)
+            : _data(data), _range(in_range) {}
+        template <memory_region M = Memory>
+        accessor(typename std::enable_if<M != local, const accessor_t>::type &acc)
+            : accessor(acc, acc.get_range()) {}
+        accessor(const accessor_t &acc, const sycl::range<3> &in_range)
+            : accessor(acc.get_pointer(), in_range) {}
+        accessor<T, Memory, 2> operator[](size_t index) const {
+            sycl::range<2> sub(_range.get(1), _range.get(2));
+            return accessor<T, Memory, 2>(_data + index * sub.size(), sub);
+        }
+
+        pointer_t get_ptr() const { return _data; }
+
+    private:
+        pointer_t _data;
+        sycl::range<3> _range;
+    };
+    template <class T, memory_region Memory> class accessor<T, Memory, 2> {
+    public:
+        using memory_t = detail::memory_traits<Memory, T>;
+        using element_t = typename memory_t::element_t;
+        using pointer_t = typename memory_t::pointer_t;
+        using accessor_t = typename memory_t::template accessor_t<2>;
+        accessor(pointer_t data, const sycl::range<2> &in_range)
+            : _data(data), _range(in_range) {}
+        template <memory_region M = Memory>
+        accessor(typename std::enable_if<M != local, const accessor_t>::type &acc)
+            : accessor(acc, acc.get_range()) {}
+        accessor(const accessor_t &acc, const sycl::range<2> &in_range)
+            : accessor(acc.get_pointer(), in_range) {}
+
+        pointer_t operator[](size_t index) const {
+            return _data + _range.get(1) * index;
+        }
+
+        pointer_t get_ptr() const { return _data; }
+
+    private:
+        pointer_t _data;
+        sycl::range<2> _range;
+    };
+
+    namespace detail {
+        /// Device variable with address space of shared, global or constant.
+        template <class T, memory_region Memory, size_t Dimension> class device_memory {
+        public:
+            using accessor_t =
+                typename detail::memory_traits<Memory,
+                                            T>::template accessor_t<Dimension>;
+            using value_t = typename detail::memory_traits<Memory, T>::value_t;
+            using dpct_accessor_t = dpct::accessor<T, Memory, Dimension>;
+
+            device_memory() : device_memory(sycl::range<Dimension>(1)) {}
+
+            /// Constructor of 1-D array with initializer list
+            device_memory(const sycl::range<Dimension> &in_range,
+                        std::initializer_list<value_t> &&init_list)
+                : device_memory(in_range) {
+                assert(init_list.size() <= in_range.size());
+                _host_ptr = (value_t *)std::malloc(_size);
+                std::memset(_host_ptr, 0, _size);
+                std::memcpy(_host_ptr, init_list.begin(), init_list.size() * sizeof(T));
+            }
+
+            /// Constructor of 2-D array with initializer list
+            template <size_t D = Dimension>
+            device_memory(
+                const typename std::enable_if<D == 2, sycl::range<2>>::type &in_range,
+                std::initializer_list<std::initializer_list<value_t>> &&init_list)
+                : device_memory(in_range) {
+                assert(init_list.size() <= in_range[0]);
+                _host_ptr = (value_t *)std::malloc(_size);
+                std::memset(_host_ptr, 0, _size);
+                auto tmp_data = _host_ptr;
+                for (auto sub_list : init_list) {
+                    assert(sub_list.size() <= in_range[1]);
+                    std::memcpy(tmp_data, sub_list.begin(),
+                                sub_list.size() * sizeof(T));
+                    tmp_data += in_range[1];
+                }
+            }
+
+            /// Constructor with range
+            device_memory(const sycl::range<Dimension> &range_in)
+                : _size(range_in.size() * sizeof(T)), _range(range_in),
+                _reference(false), _host_ptr(nullptr), _device_ptr(nullptr) {
+                static_assert(
+                    (Memory == global) || (Memory == constant) || (Memory == shared),
+                    "device memory region should be global, constant or shared");
+                // Make sure that singleton class mem_mgr and dev_mgr will destruct
+                // later than this.
+                detail::mem_mgr::instance();
+                dev_mgr::instance();
+            }
+
+            /// Constructor with range
+            template <class... Args>
+            device_memory(Args... Arguments)
+                : device_memory(sycl::range<Dimension>(Arguments...)) {}
+
+            ~device_memory() {
+                if (_device_ptr && !_reference)
+                    dpct::dpct_free(_device_ptr);
+                if (_host_ptr)
+                    std::free(_host_ptr);
+            }
+
+            /// Allocate memory with default queue, and init memory if has initial
+            /// value.
+            void init() { init(dpct::get_default_queue()); }
+            /// Allocate memory with specified queue, and init memory if has initial
+            /// value.
+            void init(sycl::queue &q) {
+                if (_device_ptr)
+                    return;
+                if (!_size)
+                    return;
+                allocate_device(q);
+                if (_host_ptr)
+                    detail::dpct_memcpy(q, _device_ptr, _host_ptr, _size,
+                                        host_to_device);
+            }
+
+            /// The variable is assigned to a device pointer.
+            void assign(value_t *src, size_t size) {
+                this->~device_memory();
+                new (this) device_memory(src, size);
+            }
+
+            /// Get memory pointer of the memory object, which is virtual pointer when
+            /// usm is not used, and device pointer when usm is used.
+            value_t *get_ptr() { return get_ptr(get_default_queue()); }
+            /// Get memory pointer of the memory object, which is virtual pointer when
+            /// usm is not used, and device pointer when usm is used.
+            value_t *get_ptr(sycl::queue &q) {
+                init(q);
+                return _device_ptr;
+            }
+
+            /// Get the device memory object size in bytes.
+            size_t get_size() { return _size; }
+
+            template <size_t D = Dimension>
+            typename std::enable_if<D == 1, T>::type &operator[](size_t index) {
+                init();
+                return _device_ptr[index];
+            }
+
+            /// Get dpct::accessor with dimension info for the device memory object
+            /// when usm is used and dimension is greater than 1.
+            template <size_t D = Dimension>
+            typename std::enable_if<D != 1, dpct_accessor_t>::type
+            get_access([[maybe_unused]] sycl::handler &cgh) {
+                return dpct_accessor_t((T *)_device_ptr, _range);
+            }
+
+        private:
+            device_memory(value_t *memory_ptr, size_t size)
+                : _size(size), _range(size / sizeof(T)), _reference(true),
+                _device_ptr(memory_ptr) {}
+
+            void allocate_device(sycl::queue &q) {
+        #ifndef DPCT_USM_LEVEL_NONE
+                if (Memory == shared) {
+                    _device_ptr = (value_t *)sycl::malloc_shared(_size, q.get_device(),
+                                                                q.get_context());
+                    return;
+                }
+        #ifdef SYCL_EXT_ONEAPI_USM_DEVICE_READ_ONLY
+                if (Memory == constant) {
+                    _device_ptr = (value_t *)sycl::malloc_device(
+                        _size, q.get_device(), q.get_context(),
+                        sycl::ext::oneapi::property::usm::device_read_only());
+                    return;
+                }
+        #endif
+        #endif
+                _device_ptr = (value_t *)detail::dpct_malloc(_size, q);
+            }
+
+            size_t _size;
+            sycl::range<Dimension> _range;
+            bool _reference;
+            value_t *_host_ptr;
+            value_t *_device_ptr;
+        };
+        template <class T, memory_region Memory>
+        class device_memory<T, Memory, 0> : public device_memory<T, Memory, 1> {
+        public:
+            using base = device_memory<T, Memory, 1>;
+            using value_t = typename base::value_t;
+            using accessor_t =
+                typename detail::memory_traits<Memory, T>::template accessor_t<0>;
+
+            /// Constructor with initial value.
+            device_memory(const value_t &val) : base(sycl::range<1>(1), {val}) {}
+
+            /// Default constructor
+            device_memory() : base(1) {}
+        };
+        } // namespace detail
+
+    template <class T, size_t Dimension>
+    using global_memory = detail::device_memory<T, global, Dimension>;
+    template <class T, size_t Dimension>
+    using constant_memory = detail::device_memory<T, constant, Dimension>;
+    template <class T, size_t Dimension>
+    using shared_memory = detail::device_memory<T, shared, Dimension>;
+
+
+    template <typename T,
+            sycl::access::address_space addressSpace =
+                sycl::access::address_space::global_space,
+            sycl::memory_order memoryOrder = sycl::memory_order::relaxed,
+            sycl::memory_scope memoryScope = sycl::memory_scope::device>
+    inline T atomic_fetch_add(T *addr, T operand) {
+    auto atm =
+        sycl::atomic_ref<T, memoryOrder, memoryScope, addressSpace>(addr[0]);
+    return atm.fetch_add(operand);
+    }
+
+    template <sycl::access::address_space addressSpace =
+                sycl::access::address_space::global_space,
+            sycl::memory_order memoryOrder = sycl::memory_order::relaxed,
+            sycl::memory_scope memoryScope = sycl::memory_scope::device,
+            typename T1, typename T2>
+    inline T1 atomic_fetch_add(T1 *addr, T2 operand) {
+    auto atm =
+        sycl::atomic_ref<T1, memoryOrder, memoryScope, addressSpace>(addr[0]);
+    return atm.fetch_add(operand);
+    }
+
+    template <typename T, sycl::access::address_space addressSpace =
+                            sycl::access::address_space::global_space>
+    inline T atomic_fetch_add(T *addr, T operand,
+                            sycl::memory_order memoryOrder) {
+    switch (memoryOrder) {
+        case sycl::memory_order::relaxed:
+            return atomic_fetch_add<T, addressSpace, sycl::memory_order::relaxed,
+                                    sycl::memory_scope::device>(addr, operand);
+        case sycl::memory_order::acq_rel:
+            return atomic_fetch_add<T, addressSpace, sycl::memory_order::acq_rel,
+                                    sycl::memory_scope::device>(addr, operand);
+        case sycl::memory_order::seq_cst:
+            return atomic_fetch_add<T, addressSpace, sycl::memory_order::seq_cst,
+                                    sycl::memory_scope::device>(addr, operand);
+        default:
+            assert(false && "Invalid memory_order for atomics. Valid memory_order for "
+                            "atomics are: sycl::memory_order::relaxed, "
+                            "sycl::memory_order::acq_rel, sycl::memory_order::seq_cst!");
+        }
+    }
+
+    template <sycl::access::address_space addressSpace =
+                sycl::access::address_space::global_space,
+            typename T1, typename T2>
+    inline T1 atomic_fetch_add(T1 *addr, T2 operand,
+                            sycl::memory_order memoryOrder) {
+    atomic_fetch_add<T1, addressSpace>(addr, operand, memoryOrder);
+    }
+
+} // COPY from DPCT head files
+
+#endif // GGML_SYCL_DPCT_HELPER_HPP
diff --git a/src/whisper_cpp/ggml/src/ggml-sycl/gemm.hpp b/src/whisper_cpp/ggml/src/ggml-sycl/gemm.hpp
new file mode 100644
index 00000000..2ad9b36f
--- /dev/null
+++ b/src/whisper_cpp/ggml/src/ggml-sycl/gemm.hpp
@@ -0,0 +1,101 @@
+//
+// MIT license
+// Copyright (C) 2024 Intel Corporation
+// SPDX-License-Identifier: MIT
+//
+
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+//
+
+#ifndef GGML_SYCL_GEMM_HPP
+#define GGML_SYCL_GEMM_HPP
+
+#include <fstream>
+#include <iostream>
+
+#include "ggml-sycl.h"
+
+#if GGML_SYCL_DNNL
+
+#include "dnnl.hpp"
+#include "dnnl_sycl.hpp"
+
+class DnnlGemmWrapper {
+public:
+    using dt = dnnl::memory::data_type;
+    using tag = dnnl::memory::format_tag;
+
+    template<typename T>
+    static constexpr dt to_dt() {
+        if constexpr (std::is_same_v<T, float>) return dt::f32;
+        else if constexpr (std::is_same_v<T, sycl::half>) return dt::f16;
+        else static_assert(0);
+    }
+
+    static inline void row_gemm(sycl::queue& q, bool a_trans,
+        bool b_trans, int m, int n, int k,
+        const void* a, dt at, const void* b, dt bt, void* c, dt ct)
+    {
+        // Get the device associated with the queue
+        sycl::device dev = q.get_device();
+        // Get the context associated with the queue
+        sycl::context ctx = q.get_context();
+        const dnnl::engine eng = dnnl::sycl_interop::make_engine(dev, ctx);
+        const dnnl::stream stream = dnnl::sycl_interop::make_stream(eng, q);
+        dnnl::memory::dims a_dims = { m, k };
+        dnnl::memory::dims b_dims = { k, n };
+        dnnl::memory::dims c_dims = { m, n };
+        const auto a_in_md = dnnl::memory::desc(a_dims, at, a_trans ? tag::ba : tag::ab);
+        const auto b_in_md = dnnl::memory::desc(b_dims, bt, b_trans ? tag::ba : tag::ab);
+        const auto c_md = dnnl::memory::desc(c_dims, ct, tag::ab);
+        auto a_mem = dnnl::memory(a_in_md, eng, (void*)a);
+        auto b_mem = dnnl::memory(b_in_md, eng, (void*)b);
+        auto matmul_pd = dnnl::matmul::primitive_desc(eng, a_in_md, b_in_md, c_md);
+        auto c_mem = dnnl::memory(matmul_pd.dst_desc(), eng, c);
+
+        // Create the primitive.
+        auto matmul_prim = dnnl::matmul(matmul_pd);
+        // Primitive arguments.
+        std::unordered_map<int, dnnl::memory> matmul_args;
+        matmul_args.insert({ DNNL_ARG_SRC, a_mem });
+        matmul_args.insert({ DNNL_ARG_WEIGHTS, b_mem });
+        matmul_args.insert({ DNNL_ARG_DST, c_mem });
+
+        matmul_prim.execute(stream, matmul_args);
+    }
+
+
+    static inline void row_gemm(const dnnl::stream& stream, bool a_trans,
+        bool b_trans, int m, int n, int k,
+        const void* a, dt at, const void* b, dt bt, void* c, dt ct)
+    {
+        auto const eng = stream.get_engine();
+        dnnl::memory::dims a_dims = { m, k };
+        dnnl::memory::dims b_dims = { k, n };
+        dnnl::memory::dims c_dims = { m, n };
+        const auto a_in_md = dnnl::memory::desc(a_dims, at, a_trans ? tag::ba : tag::ab);
+        const auto b_in_md = dnnl::memory::desc(b_dims, bt, b_trans ? tag::ba : tag::ab);
+        const auto c_md = dnnl::memory::desc(c_dims, ct, tag::ab);
+        auto a_mem = dnnl::memory(a_in_md, eng, (void*)a);
+        auto b_mem = dnnl::memory(b_in_md, eng, (void*)b);
+        auto matmul_pd = dnnl::matmul::primitive_desc(eng, a_in_md, b_in_md, c_md);
+        auto c_mem = dnnl::memory(matmul_pd.dst_desc(), eng, c);
+
+        // Create the primitive.
+        auto matmul_prim = dnnl::matmul(matmul_pd);
+        // Primitive arguments.
+        std::unordered_map<int, dnnl::memory> matmul_args;
+        matmul_args.insert({ DNNL_ARG_SRC, a_mem });
+        matmul_args.insert({ DNNL_ARG_WEIGHTS, b_mem });
+        matmul_args.insert({ DNNL_ARG_DST, c_mem });
+
+        matmul_prim.execute(stream, matmul_args);
+    }
+};
+
+#endif
+
+#endif // GGML_SYCL_GEMM_HPP
diff --git a/src/whisper_cpp/ggml/src/ggml-sycl/im2col.cpp b/src/whisper_cpp/ggml/src/ggml-sycl/im2col.cpp
new file mode 100644
index 00000000..6a0a0fcd
--- /dev/null
+++ b/src/whisper_cpp/ggml/src/ggml-sycl/im2col.cpp
@@ -0,0 +1,125 @@
+//
+// MIT license
+// Copyright (C) 2024 Intel Corporation
+// SPDX-License-Identifier: MIT
+//
+
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+//
+
+#include "im2col.hpp"
+
+template <typename T>
+static void im2col_kernel(
+        const float *x, T *dst, int64_t batch_offset, int64_t offset_delta,
+        int64_t IC, int64_t IW, int64_t IH, int64_t OH, int64_t OW, int64_t KW, int64_t KH,
+        int64_t pelements, int64_t CHW, int s0, int s1, int p0, int p1, int d0, int d1,
+        const sycl::nd_item<3> &item_ct1) {
+    const int64_t work_group_size = item_ct1.get_local_range(2);
+    const int64_t global_id = item_ct1.get_local_id(2) + work_group_size * item_ct1.get_group(2);
+
+    // make each work-item deal with more elements since sycl global range can not exceed max int
+    for (int64_t i = global_id; i < pelements; i += work_group_size * item_ct1.get_group_range(2)) {
+
+        const int64_t ksize = OW * (KH > 1 ? KW : 1);
+        const int64_t kx = i / ksize;
+        const int64_t kd = kx * ksize;
+        const int64_t ky = (i - kd) / OW;
+        const int64_t ix = i % OW;
+
+        const int64_t  oh = item_ct1.get_group(1);
+        const int64_t  batch = item_ct1.get_group(0) / IC;
+        const int64_t  ic = item_ct1.get_group(0) % IC;
+
+        const int64_t iiw = ix * s0 + kx * d0 - p0;
+        const int64_t iih = oh * s1 + ky * d1 - p1;
+
+        const int64_t offset_dst =
+            ((batch * OH + oh) * OW + ix) * CHW +
+            (ic * (KW * KH) + ky * KW + kx);
+
+        if (iih < 0 || iih >= IH || iiw < 0 || iiw >= IW) {
+            dst[offset_dst] =
+                sycl::vec<float, 1>(0.0f)
+                    .convert<sycl::half, sycl::rounding_mode::automatic>()[0];
+        } else {
+            const int64_t offset_src = ic * offset_delta + batch * batch_offset;
+            dst[offset_dst] =
+                sycl::vec<float, 1>(x[offset_src + iih * IW + iiw])
+                    .convert<sycl::half, sycl::rounding_mode::automatic>()[0];
+        }
+    }
+}
+
+template <typename T>
+static void im2col_sycl(
+        const float *x, T *dst, int64_t IW, int64_t IH, int64_t OW, int64_t OH, int64_t KW,
+        int64_t KH, int64_t IC, int64_t batch, int64_t batch_offset, int64_t offset_delta,
+        int s0, int s1, int p0, int p1, int d0, int d1,
+        queue_ptr stream) {
+    const int64_t parallel_elements = OW * KW * KH;
+    const int64_t num_blocks = (parallel_elements + SYCL_IM2COL_BLOCK_SIZE - 1) / SYCL_IM2COL_BLOCK_SIZE;
+
+    // decrease global range when it exceeds the max int
+    int64_t local_size = downsample_sycl_global_range(batch * IC * OH * num_blocks, SYCL_IM2COL_BLOCK_SIZE);
+    sycl::range<3> block_nums(batch * IC, OH, num_blocks);
+    sycl::range<3> local_range(1, 1, local_size);
+
+    {
+        dpct::has_capability_or_fail(stream->get_device(),
+                                     {sycl::aspect::fp16});
+
+        stream->parallel_for(
+            sycl::nd_range<3>(block_nums * local_range, local_range),
+            [=](sycl::nd_item<3> item_ct1) {
+                im2col_kernel(x, dst, batch_offset, offset_delta, IC, IW, IH, OH, OW, KW, KH,
+                               parallel_elements, (IC * KH * KW), s0, s1, p0,
+                               p1, d0, d1, item_ct1);
+            });
+    }
+}
+
+void ggml_sycl_op_im2col(
+        ggml_backend_sycl_context & ctx, const ggml_tensor *src0, const ggml_tensor *src1,
+        ggml_tensor *dst, const float *src0_dd, const float *src1_dd, float *dst_dd,
+        const queue_ptr &main_stream) {
+
+    GGML_ASSERT(src0->type == GGML_TYPE_F16);
+    GGML_ASSERT(src1->type == GGML_TYPE_F32);
+    GGML_ASSERT(dst->type == GGML_TYPE_F16 || dst->type == GGML_TYPE_F32);
+
+    const int32_t s0 = ((const int32_t*)(dst->op_params))[0];
+    const int32_t s1 = ((const int32_t*)(dst->op_params))[1];
+    const int32_t p0 = ((const int32_t*)(dst->op_params))[2];
+    const int32_t p1 = ((const int32_t*)(dst->op_params))[3];
+    const int32_t d0 = ((const int32_t*)(dst->op_params))[4];
+    const int32_t d1 = ((const int32_t*)(dst->op_params))[5];
+
+    const bool is_2D = ((const int32_t*)(dst->op_params))[6] == 1;
+
+    const int64_t IC = src1->ne[is_2D ? 2 : 1];
+    const int64_t IH = is_2D ? src1->ne[1] : 1;
+    const int64_t IW =         src1->ne[0];
+
+    const int64_t KH = is_2D ? src0->ne[1] : 1;
+    const int64_t KW =         src0->ne[0];
+
+    const int64_t OH = is_2D ? dst->ne[2] : 1;
+    const int64_t OW =         dst->ne[1];
+
+    const size_t delta_offset = src1->nb[is_2D ? 2 : 1] / 4; // nb is byte offset, src is type float32
+    const int64_t batch = src1->ne[3];
+    const size_t batch_offset = src1->nb[3] / 4; // nb is byte offset, src is type float32
+
+    if (dst->type == GGML_TYPE_F16) {
+        im2col_sycl(src1_dd, (sycl::half *)dst_dd, IW, IH, OW, OH, KW, KH, IC, batch, batch_offset, delta_offset, s0, s1, p0, p1, d0, d1, main_stream);
+    } else {
+        im2col_sycl(src1_dd, (float *)dst_dd, IW, IH, OW, OH, KW, KH, IC, batch, batch_offset, delta_offset, s0, s1, p0, p1, d0, d1, main_stream);
+    }
+
+    (void) src0;
+    (void) src0_dd;
+}
diff --git a/src/whisper_cpp/ggml/src/ggml-sycl/im2col.hpp b/src/whisper_cpp/ggml/src/ggml-sycl/im2col.hpp
new file mode 100644
index 00000000..7db144fb
--- /dev/null
+++ b/src/whisper_cpp/ggml/src/ggml-sycl/im2col.hpp
@@ -0,0 +1,23 @@
+//
+// MIT license
+// Copyright (C) 2024 Intel Corporation
+// SPDX-License-Identifier: MIT
+//
+
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+//
+
+#ifndef GGML_SYCL_IM2COL_HPP
+#define GGML_SYCL_IM2COL_HPP
+
+#include "common.hpp"
+
+void ggml_sycl_op_im2col(
+        ggml_backend_sycl_context & ctx, const ggml_tensor *src0, const ggml_tensor *src1,
+        ggml_tensor *dst, const float *src0_dd, const float *src1_dd, float *dst_dd,
+        const queue_ptr &main_stream);
+
+#endif // GGML_SYCL_IM2COL_HPP
diff --git a/src/whisper_cpp/ggml/src/ggml-sycl/mmq.cpp b/src/whisper_cpp/ggml/src/ggml-sycl/mmq.cpp
new file mode 100644
index 00000000..e952533d
--- /dev/null
+++ b/src/whisper_cpp/ggml/src/ggml-sycl/mmq.cpp
@@ -0,0 +1,3031 @@
+//
+// MIT license
+// Copyright (C) 2024 Intel Corporation
+// SPDX-License-Identifier: MIT
+//
+
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+//
+
+#include "mmq.hpp"
+#include "vecdotq.hpp"
+
+typedef void (*allocate_tiles_sycl_t)(
+    int** x_ql,
+    sycl::half2** x_dm,
+    int** x_qh,
+    int** x_sc);
+typedef void (*load_tiles_sycl_t)(
+    const void* __restrict__ vx,
+    int* __restrict__ x_ql,
+    sycl::half2* __restrict__ x_dm,
+    int* __restrict__ x_qh,
+    int* __restrict__ x_sc,
+    const int& i_offset,
+    const int& i_max,
+    const int& k,
+    const int& blocks_per_row);
+typedef float (*vec_dot_q_mul_mat_sycl_t)(
+    const int* __restrict__ x_ql,
+    const sycl::half2* __restrict__ x_dm,
+    const int* __restrict__ x_qh,
+    const int* __restrict__ x_sc,
+    const int* __restrict__ y_qs,
+    const sycl::half2* __restrict__ y_ms,
+    const int& i,
+    const int& j,
+    const int& k);
+
+
+template <int mmq_y>
+static __dpct_inline__ void
+allocate_tiles_q4_0(int **x_ql, sycl::half2 **x_dm, int **x_qh, int **x_sc,
+                    int *tile_x_qs_q4_0, float *tile_x_d_q4_0) {
+    (void)x_qh; (void)x_sc;
+
+    *x_ql = tile_x_qs_q4_0;
+    *x_dm = (sycl::half2 *)tile_x_d_q4_0;
+}
+
+template <int mmq_y, int nwarps, bool need_check>
+static __dpct_inline__ void
+load_tiles_q4_0(const void *__restrict__ vx, int *__restrict__ x_ql,
+                sycl::half2 *__restrict__ x_dm, int *__restrict__ x_qh,
+                int *__restrict__ x_sc, const int &i_offset, const int &i_max,
+                const int &k, const int &blocks_per_row) {
+    (void)x_qh; (void)x_sc;
+    GGML_SYCL_ASSUME(i_offset >= 0);
+    GGML_SYCL_ASSUME(i_offset <  nwarps);
+    GGML_SYCL_ASSUME(k >= 0);
+    GGML_SYCL_ASSUME(k <  WARP_SIZE);
+
+    const int kbx  = k / QI4_0;
+    const int kqsx = k % QI4_0;
+
+    const block_q4_0 * bx0 = (const block_q4_0 *) vx;
+
+    float * x_dmf = (float *) x_dm;
+
+#pragma unroll
+    for (int i0 = 0; i0 < mmq_y; i0 += nwarps) {
+        int i = i0 + i_offset;
+
+        if (need_check) {
+            i = sycl::min(i, i_max);
+        }
+
+        const block_q4_0 * bxi = bx0 + i*blocks_per_row + kbx;
+
+        x_ql[i * (WARP_SIZE + 1) + k] = get_int_from_uint8(bxi->qs, kqsx);
+        // x_dmf[i * (WARP_SIZE/QI4_0) + i / QI4_0 + kbx] = bxi->d;
+    }
+
+    const int blocks_per_tile_x_row = WARP_SIZE / QI4_0;
+    const int kbxd = k % blocks_per_tile_x_row;
+
+#pragma unroll
+    for (int i0 = 0; i0 < mmq_y; i0 += nwarps * QI4_0) {
+        int i = i0 + i_offset * QI4_0 + k / blocks_per_tile_x_row;
+
+        if (need_check) {
+            i = sycl::min(i, i_max);
+        }
+
+        const block_q4_0 * bxi = bx0 + i*blocks_per_row + kbxd;
+
+        x_dmf[i * (WARP_SIZE/QI4_0) + i / QI4_0 + kbxd] = bxi->d;
+    }
+}
+
+static __dpct_inline__ float vec_dot_q4_0_q8_1_mul_mat(
+    const int *__restrict__ x_ql, const sycl::half2 *__restrict__ x_dm,
+    const int *__restrict__ x_qh, const int *__restrict__ x_sc,
+    const int *__restrict__ y_qs, const sycl::half2 *__restrict__ y_ds,
+    const int &i, const int &j, const int &k) {
+    (void)x_qh; (void)x_sc;
+
+    const int kyqs = k % (QI8_1/2) + QI8_1 * (k / (QI8_1/2));
+    const float * x_dmf = (const float *) x_dm;
+
+    int u[2*VDR_Q4_0_Q8_1_MMQ];
+
+#pragma unroll
+    for (int l = 0; l < VDR_Q4_0_Q8_1_MMQ; ++l) {
+        u[2*l+0] = y_qs[j * WARP_SIZE + (kyqs + l)         % WARP_SIZE];
+        u[2*l+1] = y_qs[j * WARP_SIZE + (kyqs + l + QI4_0) % WARP_SIZE];
+    }
+
+    return vec_dot_q4_0_q8_1_impl<VDR_Q4_0_Q8_1_MMQ>
+        (&x_ql[i * (WARP_SIZE + 1) + k], u, x_dmf[i * (WARP_SIZE/QI4_0) + i/QI4_0 + k/QI4_0],
+         y_ds[j * (WARP_SIZE/QI8_1) + (2*k/QI8_1) % (WARP_SIZE/QI8_1)]);
+}
+
+template <int mmq_y>
+static __dpct_inline__ void
+allocate_tiles_q4_1(int **x_ql, sycl::half2 **x_dm, int **x_qh, int **x_sc,
+                    int *tile_x_qs_q4_1, sycl::half2 *tile_x_dm_q4_1) {
+    (void)x_qh; (void)x_sc;
+
+    *x_ql = tile_x_qs_q4_1;
+    *x_dm = tile_x_dm_q4_1;
+}
+
+
+template <int mmq_y, int nwarps, bool need_check>
+static __dpct_inline__ void
+load_tiles_q4_1(const void *__restrict__ vx, int *__restrict__ x_ql,
+                sycl::half2 *__restrict__ x_dm, int *__restrict__ x_qh,
+                int *__restrict__ x_sc, const int &i_offset, const int &i_max,
+                const int &k, const int &blocks_per_row) {
+    (void)x_qh; (void)x_sc;
+
+    GGML_SYCL_ASSUME(i_offset >= 0);
+    GGML_SYCL_ASSUME(i_offset <  nwarps);
+    GGML_SYCL_ASSUME(k >= 0);
+    GGML_SYCL_ASSUME(k <  WARP_SIZE);
+
+    const int kbx  = k / QI4_1;
+    const int kqsx = k % QI4_1;
+
+    const block_q4_1 * bx0 = (const block_q4_1 *) vx;
+
+#pragma unroll
+    for (int i0 = 0; i0 < mmq_y; i0 += nwarps) {
+        int i = i0 + i_offset;
+
+        if (need_check) {
+            i = sycl::min(i, i_max);
+        }
+
+        const block_q4_1 * bxi = bx0 + i*blocks_per_row + kbx;
+
+        x_ql[i * (WARP_SIZE + 1) + k] = get_int_from_uint8_aligned(bxi->qs, kqsx);
+    }
+
+    const int blocks_per_tile_x_row = WARP_SIZE / QI4_1;
+    const int kbxd = k % blocks_per_tile_x_row;
+
+#pragma unroll
+    for (int i0 = 0; i0 < mmq_y; i0 += nwarps * QI4_1) {
+        int i = i0 + i_offset * QI4_1 + k / blocks_per_tile_x_row;
+
+        if (need_check) {
+            i = sycl::min(i, i_max);
+        }
+
+        const block_q4_1 * bxi = bx0 + i*blocks_per_row + kbxd;
+
+        x_dm[i * (WARP_SIZE/QI4_1) + i / QI4_1 + kbxd] = bxi->dm;
+    }
+}
+
+static __dpct_inline__ float vec_dot_q4_1_q8_1_mul_mat(
+    const int *__restrict__ x_ql, const sycl::half2 *__restrict__ x_dm,
+    const int *__restrict__ x_qh, const int *__restrict__ x_sc,
+    const int *__restrict__ y_qs, const sycl::half2 *__restrict__ y_ds,
+    const int &i, const int &j, const int &k) {
+    (void)x_qh; (void)x_sc;
+
+    const int kyqs = k % (QI8_1/2) + QI8_1 * (k / (QI8_1/2));
+
+    int u[2*VDR_Q4_1_Q8_1_MMQ];
+
+#pragma unroll
+    for (int l = 0; l < VDR_Q4_1_Q8_1_MMQ; ++l) {
+        u[2*l+0] = y_qs[j * WARP_SIZE + (kyqs + l)         % WARP_SIZE];
+        u[2*l+1] = y_qs[j * WARP_SIZE + (kyqs + l + QI4_1) % WARP_SIZE];
+    }
+
+    return vec_dot_q4_1_q8_1_impl<VDR_Q4_1_Q8_1_MMQ>
+        (&x_ql[i * (WARP_SIZE + 1) + k], u, x_dm[i * (WARP_SIZE/QI4_1) + i/QI4_1 + k/QI4_1],
+         y_ds[j * (WARP_SIZE/QI8_1) + (2*k/QI8_1) % (WARP_SIZE/QI8_1)]);
+}
+
+template <int mmq_y>
+static __dpct_inline__ void
+allocate_tiles_q5_0(int **x_ql, sycl::half2 **x_dm, int **x_qh, int **x_sc,
+                    int *tile_x_ql_q5_0, float *tile_x_d_q5_0) {
+    (void)x_qh; (void)x_sc;
+
+    *x_ql = tile_x_ql_q5_0;
+    *x_dm = (sycl::half2 *)tile_x_d_q5_0;
+}
+
+template <int mmq_y, int nwarps, bool need_check>
+static __dpct_inline__ void
+load_tiles_q5_0(const void *__restrict__ vx, int *__restrict__ x_ql,
+                sycl::half2 *__restrict__ x_dm, int *__restrict__ x_qh,
+                int *__restrict__ x_sc, const int &i_offset, const int &i_max,
+                const int &k, const int &blocks_per_row) {
+    (void)x_qh; (void)x_sc;
+
+    GGML_SYCL_ASSUME(i_offset >= 0);
+    GGML_SYCL_ASSUME(i_offset <  nwarps);
+    GGML_SYCL_ASSUME(k >= 0);
+    GGML_SYCL_ASSUME(k <  WARP_SIZE);
+
+    const int kbx  = k / QI5_0;
+    const int kqsx = k % QI5_0;
+
+    const block_q5_0 * bx0 = (const block_q5_0 *) vx;
+
+#pragma unroll
+    for (int i0 = 0; i0 < mmq_y; i0 += nwarps) {
+        int i = i0 + i_offset;
+
+        if (need_check) {
+            i = sycl::min(i, i_max);
+        }
+
+        const block_q5_0 * bxi = bx0 + i*blocks_per_row + kbx;
+
+        const int ql = get_int_from_uint8(bxi->qs, kqsx);
+        const int qh = get_int_from_uint8(bxi->qh, 0) >> (4 * (k % QI5_0));
+
+        int qs0 = (ql >>  0)   & 0x0F0F0F0F;
+        qs0    |= (qh <<  4)   & 0x00000010;  // 0 ->  4
+        qs0    |= (qh << 11)   & 0x00001000;  // 1 -> 12
+        qs0    |= (qh << 18)   & 0x00100000;  // 2 -> 20
+        qs0    |= (qh << 25)   & 0x10000000;  // 3 -> 28
+        qs0 = dpct::vectorized_binary<sycl::char4>(
+            qs0, 0x10101010, dpct::sub_sat()); // subtract 16
+
+        x_ql[i * (2*WARP_SIZE + 1) + 2*k+0] = qs0;
+
+        int qs1 = (ql >>  4)   & 0x0F0F0F0F;
+        qs1    |= (qh >> 12)   & 0x00000010;  // 16 ->  4
+        qs1    |= (qh >>  5)   & 0x00001000;  // 17 -> 12
+        qs1    |= (qh <<  2)   & 0x00100000;  // 18 -> 20
+        qs1    |= (qh <<  9)   & 0x10000000;  // 19 -> 28
+        qs1 = dpct::vectorized_binary<sycl::char4>(
+            qs1, 0x10101010, dpct::sub_sat()); // subtract 16
+
+        x_ql[i * (2*WARP_SIZE + 1) + 2*k+1] = qs1;
+    }
+
+    const int blocks_per_tile_x_row = WARP_SIZE / QI5_0;
+    const int kbxd = k % blocks_per_tile_x_row;
+    float * x_dmf = (float *) x_dm;
+
+#pragma unroll
+    for (int i0 = 0; i0 < mmq_y; i0 += nwarps * QI5_0) {
+        int i = i0 + i_offset * QI5_0 + k / blocks_per_tile_x_row;
+
+        if (need_check) {
+            i = sycl::min(i, i_max);
+        }
+
+        const block_q5_0 * bxi = bx0 + i*blocks_per_row + kbxd;
+
+        x_dmf[i * (WARP_SIZE/QI5_0) + i / QI5_0 + kbxd] = bxi->d;
+    }
+}
+
+static __dpct_inline__ float vec_dot_q5_0_q8_1_mul_mat(
+    const int *__restrict__ x_ql, const sycl::half2 *__restrict__ x_dm,
+    const int *__restrict__ x_qh, const int *__restrict__ x_sc,
+    const int *__restrict__ y_qs, const sycl::half2 *__restrict__ y_ds,
+    const int &i, const int &j, const int &k) {
+    (void)x_qh; (void)x_sc;
+
+    const int kyqs = k % (QI8_1/2) + QI8_1 * (k / (QI8_1/2));
+    const int index_bx = i * (WARP_SIZE/QI5_0) + i/QI5_0 + k/QI5_0;
+    const float * x_dmf = (const float *) x_dm;
+    const float * y_df  = (const float *) y_ds;
+
+    int u[2*VDR_Q5_0_Q8_1_MMQ];
+
+#pragma unroll
+    for (int l = 0; l < VDR_Q5_0_Q8_1_MMQ; ++l) {
+        u[2*l+0] = y_qs[j * WARP_SIZE + (kyqs + l)         % WARP_SIZE];
+        u[2*l+1] = y_qs[j * WARP_SIZE + (kyqs + l + QI5_0) % WARP_SIZE];
+    }
+
+    return vec_dot_q8_0_q8_1_impl<QR5_0*VDR_Q5_0_Q8_1_MMQ>
+        (&x_ql[i * (2*WARP_SIZE + 1) + 2 * k], u, x_dmf[index_bx], y_df[j * (WARP_SIZE/QI8_1) + (2*k/QI8_1) % (WARP_SIZE/QI8_1)]);
+}
+
+template <int mmq_y>
+static __dpct_inline__ void
+allocate_tiles_q5_1(int **x_ql, sycl::half2 **x_dm, int **x_qh, int **x_sc,
+                    int *tile_x_ql_q5_1, sycl::half2 *tile_x_dm_q5_1) {
+    (void)x_qh; (void)x_sc;
+
+    *x_ql = tile_x_ql_q5_1;
+    *x_dm = tile_x_dm_q5_1;
+}
+
+template <int mmq_y, int nwarps, bool need_check>
+static __dpct_inline__ void
+load_tiles_q5_1(const void *__restrict__ vx, int *__restrict__ x_ql,
+                sycl::half2 *__restrict__ x_dm, int *__restrict__ x_qh,
+                int *__restrict__ x_sc, const int &i_offset, const int &i_max,
+                const int &k, const int &blocks_per_row) {
+    (void)x_qh; (void)x_sc;
+
+    GGML_SYCL_ASSUME(i_offset >= 0);
+    GGML_SYCL_ASSUME(i_offset < nwarps);
+    GGML_SYCL_ASSUME(k >= 0);
+    GGML_SYCL_ASSUME(k <  WARP_SIZE);
+
+    const int kbx  = k / QI5_1;
+    const int kqsx = k % QI5_1;
+
+    const block_q5_1 * bx0 = (const block_q5_1 *) vx;
+
+#pragma unroll
+    for (int i0 = 0; i0 < mmq_y; i0 += nwarps) {
+        int i = i0 + i_offset;
+
+        if (need_check) {
+            i = sycl::min(i, i_max);
+        }
+
+        const block_q5_1 * bxi = bx0 + i*blocks_per_row + kbx;
+
+        const int ql = get_int_from_uint8_aligned(bxi->qs, kqsx);
+        const int qh = get_int_from_uint8_aligned(bxi->qh, 0) >> (4 * (k % QI5_1));
+
+        int qs0 = (ql >>  0) & 0x0F0F0F0F;
+        qs0    |= (qh <<  4) & 0x00000010; // 0 ->  4
+        qs0    |= (qh << 11) & 0x00001000; // 1 -> 12
+        qs0    |= (qh << 18) & 0x00100000; // 2 -> 20
+        qs0    |= (qh << 25) & 0x10000000; // 3 -> 28
+
+        x_ql[i * (2*WARP_SIZE + 1) + 2*k+0] = qs0;
+
+        int qs1 = (ql >>  4) & 0x0F0F0F0F;
+        qs1    |= (qh >> 12) & 0x00000010; // 16 ->  4
+        qs1    |= (qh >>  5) & 0x00001000; // 17 -> 12
+        qs1    |= (qh <<  2) & 0x00100000; // 18 -> 20
+        qs1    |= (qh <<  9) & 0x10000000; // 19 -> 28
+
+        x_ql[i * (2*WARP_SIZE + 1) + 2*k+1] = qs1;
+    }
+
+    const int blocks_per_tile_x_row = WARP_SIZE / QI5_1;
+    const int kbxd = k % blocks_per_tile_x_row;
+
+#pragma unroll
+    for (int i0 = 0; i0 < mmq_y; i0 += nwarps * QI5_1) {
+        int i = i0 + i_offset * QI5_1 + k / blocks_per_tile_x_row;
+
+        if (need_check) {
+            i = sycl::min(i, i_max);
+        }
+
+        const block_q5_1 * bxi = bx0 + i*blocks_per_row + kbxd;
+
+        x_dm[i * (WARP_SIZE/QI5_1) + i / QI5_1 + kbxd] = bxi->dm;
+    }
+}
+
+static __dpct_inline__ float vec_dot_q5_1_q8_1_mul_mat(
+    const int *__restrict__ x_ql, const sycl::half2 *__restrict__ x_dm,
+    const int *__restrict__ x_qh, const int *__restrict__ x_sc,
+    const int *__restrict__ y_qs, const sycl::half2 *__restrict__ y_ds,
+    const int &i, const int &j, const int &k) {
+    (void)x_qh; (void)x_sc;
+
+    const int kyqs = k % (QI8_1/2) + QI8_1 * (k / (QI8_1/2));
+    const int index_bx = i * (WARP_SIZE/QI5_1) + + i/QI5_1 + k/QI5_1;
+
+    int u[2*VDR_Q5_1_Q8_1_MMQ];
+
+#pragma unroll
+    for (int l = 0; l < VDR_Q5_1_Q8_1_MMQ; ++l) {
+        u[2*l+0] = y_qs[j * WARP_SIZE + (kyqs + l)         % WARP_SIZE];
+        u[2*l+1] = y_qs[j * WARP_SIZE + (kyqs + l + QI5_1) % WARP_SIZE];
+    }
+
+    return vec_dot_q8_1_q8_1_impl<QR5_1*VDR_Q5_1_Q8_1_MMQ>
+        (&x_ql[i * (2*WARP_SIZE + 1) + 2 * k], u, x_dm[index_bx], y_ds[j * (WARP_SIZE/QI8_1) + (2*k/QI8_1) % (WARP_SIZE/QI8_1)]);
+}
+
+template <int mmq_y>
+static __dpct_inline__ void
+allocate_tiles_q8_0(int **x_ql, sycl::half2 **x_dm, int **x_qh, int **x_sc,
+                    int *tile_x_qs_q8_0, float *tile_x_d_q8_0) {
+    (void)x_qh; (void)x_sc;
+
+    *x_ql = tile_x_qs_q8_0;
+    *x_dm = (sycl::half2 *)tile_x_d_q8_0;
+}
+
+template <int mmq_y, int nwarps, bool need_check>
+static __dpct_inline__ void
+load_tiles_q8_0(const void *__restrict__ vx, int *__restrict__ x_ql,
+                sycl::half2 *__restrict__ x_dm, int *__restrict__ x_qh,
+                int *__restrict__ x_sc, const int &i_offset, const int &i_max,
+                const int &k, const int &blocks_per_row) {
+    (void)x_qh; (void)x_sc;
+
+    GGML_SYCL_ASSUME(i_offset >= 0);
+    GGML_SYCL_ASSUME(i_offset <  nwarps);
+    GGML_SYCL_ASSUME(k >= 0);
+    GGML_SYCL_ASSUME(k <  WARP_SIZE);
+
+    const int kbx  = k / QI8_0;
+    const int kqsx = k % QI8_0;
+    float * x_dmf = (float *) x_dm;
+
+    const block_q8_0 * bx0 = (const block_q8_0 *) vx;
+
+#pragma unroll
+    for (int i0 = 0; i0 < mmq_y; i0 += nwarps) {
+        int i = i0 + i_offset;
+
+        if (need_check) {
+            i = sycl::min(i, i_max);
+        }
+
+        const block_q8_0 * bxi = bx0 + i*blocks_per_row + kbx;
+
+        x_ql[i * (WARP_SIZE + 1) + k] = get_int_from_int8(bxi->qs, kqsx);
+    }
+
+    const int blocks_per_tile_x_row = WARP_SIZE / QI8_0;
+    const int kbxd = k % blocks_per_tile_x_row;
+
+#pragma unroll
+    for (int i0 = 0; i0 < mmq_y; i0 += nwarps * QI8_0) {
+        int i = i0 + i_offset * QI8_0 + k / blocks_per_tile_x_row;
+
+        if (need_check) {
+            i = sycl::min(i, i_max);
+        }
+
+        const block_q8_0 * bxi = bx0 + i*blocks_per_row + kbxd;
+
+        x_dmf[i * (WARP_SIZE/QI8_0) + i / QI8_0 + kbxd] = bxi->d;
+    }
+}
+
+static __dpct_inline__ float vec_dot_q8_0_q8_1_mul_mat(
+    const int *__restrict__ x_ql, const sycl::half2 *__restrict__ x_dm,
+    const int *__restrict__ x_qh, const int *__restrict__ x_sc,
+    const int *__restrict__ y_qs, const sycl::half2 *__restrict__ y_ds,
+    const int &i, const int &j, const int &k) {
+    (void)x_qh; (void)x_sc;
+
+    const float * x_dmf = (const float *) x_dm;
+    const float * y_df  = (const float *) y_ds;
+
+    return vec_dot_q8_0_q8_1_impl<VDR_Q8_0_Q8_1_MMQ>
+        (&x_ql[i * (WARP_SIZE + 1) + k], &y_qs[j * WARP_SIZE + k], x_dmf[i * (WARP_SIZE/QI8_0) + i/QI8_0 + k/QI8_0],
+         y_df[j * (WARP_SIZE/QI8_1) + k/QI8_1]);
+}
+
+template <int mmq_y>
+static __dpct_inline__ void
+allocate_tiles_q2_K(int **x_ql, sycl::half2 **x_dm, int **x_qh, int **x_sc,
+                    int *tile_x_ql_q2_K, sycl::half2 *tile_x_dm_q2_K,
+                    int *tile_x_sc_q2_K) {
+    (void)x_qh;
+
+    *x_ql = tile_x_ql_q2_K;
+    *x_dm = tile_x_dm_q2_K;
+    *x_sc = tile_x_sc_q2_K;
+}
+
+template <int mmq_y, int nwarps, bool need_check>
+static __dpct_inline__ void
+load_tiles_q2_K(const void *__restrict__ vx, int *__restrict__ x_ql,
+                sycl::half2 *__restrict__ x_dm, int *__restrict__ x_qh,
+                int *__restrict__ x_sc, const int &i_offset, const int &i_max,
+                const int &k, const int &blocks_per_row) {
+    (void)x_qh;
+
+    GGML_SYCL_ASSUME(i_offset >= 0);
+    GGML_SYCL_ASSUME(i_offset <  nwarps);
+    GGML_SYCL_ASSUME(k >= 0);
+    GGML_SYCL_ASSUME(k <  WARP_SIZE);
+
+    const int kbx  = k / QI2_K;
+    const int kqsx = k % QI2_K;
+
+    const block_q2_K * bx0 = (const block_q2_K *) vx;
+
+#pragma unroll
+    for (int i0 = 0; i0 < mmq_y; i0 += nwarps) {
+        int i = i0 + i_offset;
+
+        if (need_check) {
+            i = sycl::min(i, i_max);
+        }
+
+        const block_q2_K * bxi = bx0 + i*blocks_per_row + kbx;
+
+        x_ql[i * (WARP_SIZE + 1) + k] = get_int_from_uint8_aligned(bxi->qs, kqsx);
+    }
+
+    const int blocks_per_tile_x_row = WARP_SIZE / QI2_K;
+    const int kbxd = k % blocks_per_tile_x_row;
+
+#pragma unroll
+    for (int i0 = 0; i0 < mmq_y; i0 += nwarps * QI2_K) {
+        int i = (i0 + i_offset * QI2_K + k / blocks_per_tile_x_row) % mmq_y;
+
+        if (need_check) {
+            i = sycl::min(i, i_max);
+        }
+
+        const block_q2_K * bxi = bx0 + i*blocks_per_row + kbxd;
+
+        x_dm[i * (WARP_SIZE/QI2_K) + i / QI2_K + kbxd] = bxi->dm;
+    }
+
+#pragma unroll
+    for (int i0 = 0; i0 < mmq_y; i0 += nwarps * 4) {
+        int i = i0 + i_offset * 4 + k / (WARP_SIZE/4);
+
+        if (need_check) {
+            i = sycl::min(i, i_max);
+        }
+
+        const block_q2_K * bxi = bx0 + i*blocks_per_row + (k % (WARP_SIZE/4)) / (QI2_K/4);
+
+        x_sc[i * (WARP_SIZE/4) + i / 4 + k % (WARP_SIZE/4)] = get_int_from_uint8_aligned(bxi->scales, k % (QI2_K/4));
+    }
+}
+
+#define VDR_Q2_K_Q8_1_MMQ  2
+// contiguous u/y values
+static __dpct_inline__ float
+vec_dot_q2_K_q8_1_impl_mmq(const int *__restrict__ v, const int *__restrict__ u,
+                           const uint8_t *__restrict__ scales,
+                           const sycl::half2 &dm2, const float &d8) {
+
+    int sumi_d = 0;
+    int sumi_m = 0;
+
+#pragma unroll
+    for (int i0 = 0; i0 < QI8_1; i0 += QI8_1/2) {
+        int sumi_d_sc = 0;
+
+        const int sc = scales[i0 / (QI8_1/2)];
+
+        // fill int with 4x m
+        int m = sc >> 4;
+        m |= m <<  8;
+        m |= m << 16;
+
+#pragma unroll
+        for (int i = i0; i < i0 + QI8_1/2; ++i) {
+            sumi_d_sc = dpct::dp4a(v[i], u[i], sumi_d_sc); // SIMD dot product
+            sumi_m = dpct::dp4a(m, u[i],
+                                sumi_m); // multiply sum of q8_1 values with m
+        }
+
+        sumi_d += sumi_d_sc * (sc & 0xF);
+    }
+
+    const sycl::float2 dm2f =
+        dm2.convert<float, sycl::rounding_mode::automatic>();
+
+    return d8 * (dm2f.x() * sumi_d - dm2f.y() * sumi_m);
+}
+
+static __dpct_inline__ float vec_dot_q2_K_q8_1_mul_mat(
+    const int *__restrict__ x_ql, const sycl::half2 *__restrict__ x_dm,
+    const int *__restrict__ x_qh, const int *__restrict__ x_sc,
+    const int *__restrict__ y_qs, const sycl::half2 *__restrict__ y_ds,
+    const int &i, const int &j, const int &k) {
+    (void)x_qh;
+
+    const int kbx = k / QI2_K;
+    const int ky  = (k % QI2_K) * QR2_K;
+    const float * y_df = (const float *) y_ds;
+
+    int v[QR2_K*VDR_Q2_K_Q8_1_MMQ];
+
+    const int kqsx = i * (WARP_SIZE + 1) + kbx*QI2_K + (QI2_K/2) * (ky/(2*QI2_K)) + ky % (QI2_K/2);
+    const int shift = 2 * ((ky % (2*QI2_K)) / (QI2_K/2));
+
+#pragma unroll
+    for (int l = 0; l < QR2_K*VDR_Q2_K_Q8_1_MMQ; ++l) {
+        v[l] = (x_ql[kqsx + l] >> shift) & 0x03030303;
+    }
+
+    const uint8_t * scales = ((const uint8_t *) &x_sc[i * (WARP_SIZE/4) + i/4 + kbx*4]) + ky/4;
+
+    const int index_y = j * WARP_SIZE + (QR2_K*k) % WARP_SIZE;
+    return vec_dot_q2_K_q8_1_impl_mmq(v, &y_qs[index_y], scales, x_dm[i * (WARP_SIZE/QI2_K) + i/QI2_K + kbx], y_df[index_y/QI8_1]);
+}
+
+template <int mmq_y>
+static __dpct_inline__ void
+allocate_tiles_q3_K(int **x_ql, sycl::half2 **x_dm, int **x_qh, int **x_sc,
+                    int *tile_x_ql_q3_K, sycl::half2 *tile_x_dm_q3_K,
+                    int *tile_x_qh_q3_K, int *tile_x_sc_q3_K) {
+
+    *x_ql = tile_x_ql_q3_K;
+    *x_dm = tile_x_dm_q3_K;
+    *x_qh = tile_x_qh_q3_K;
+    *x_sc = tile_x_sc_q3_K;
+}
+
+template <int mmq_y, int nwarps, bool need_check>
+static __dpct_inline__ void
+load_tiles_q3_K(const void *__restrict__ vx, int *__restrict__ x_ql,
+                sycl::half2 *__restrict__ x_dm, int *__restrict__ x_qh,
+                int *__restrict__ x_sc, const int &i_offset, const int &i_max,
+                const int &k, const int &blocks_per_row) {
+
+    GGML_SYCL_ASSUME(i_offset >= 0);
+    GGML_SYCL_ASSUME(i_offset <  nwarps);
+    GGML_SYCL_ASSUME(k >= 0);
+    GGML_SYCL_ASSUME(k <  WARP_SIZE);
+
+    const int kbx  = k / QI3_K;
+    const int kqsx = k % QI3_K;
+
+    const block_q3_K * bx0 = (const block_q3_K *) vx;
+
+#pragma unroll
+    for (int i0 = 0; i0 < mmq_y; i0 += nwarps) {
+        int i = i0 + i_offset;
+
+        if (need_check) {
+            i = sycl::min(i, i_max);
+        }
+
+        const block_q3_K * bxi = bx0 + i*blocks_per_row + kbx;
+
+        x_ql[i * (WARP_SIZE + 1) + k] = get_int_from_uint8(bxi->qs, kqsx);
+    }
+
+    const int blocks_per_tile_x_row = WARP_SIZE / QI3_K;
+    const int kbxd = k % blocks_per_tile_x_row;
+    float * x_dmf = (float *) x_dm;
+
+#pragma unroll
+    for (int i0 = 0; i0 < mmq_y; i0 += nwarps * QI3_K) {
+        int i = (i0 + i_offset * QI3_K + k / blocks_per_tile_x_row) % mmq_y;
+
+        if (need_check) {
+            i = sycl::min(i, i_max);
+        }
+
+        const block_q3_K * bxi = bx0 + i*blocks_per_row + kbxd;
+
+        x_dmf[i * (WARP_SIZE/QI3_K) + i / QI3_K + kbxd] = bxi->d;
+    }
+
+#pragma unroll
+    for (int i0 = 0; i0 < mmq_y; i0 += nwarps * 2) {
+        int i = i0 + i_offset * 2 + k / (WARP_SIZE/2);
+
+        if (need_check) {
+            i = sycl::min(i, i_max);
+        }
+
+        const block_q3_K * bxi = bx0 + i*blocks_per_row + (k % (WARP_SIZE/2)) / (QI3_K/2);
+
+        // invert the mask with ~ so that a 0/1 results in 4/0 being subtracted
+        x_qh[i * (WARP_SIZE/2) + i / 2 + k % (WARP_SIZE/2)] = ~get_int_from_uint8(bxi->hmask, k % (QI3_K/2));
+    }
+
+#pragma unroll
+    for (int i0 = 0; i0 < mmq_y; i0 += nwarps * 4) {
+        int i = i0 + i_offset * 4 + k / (WARP_SIZE/4);
+
+        if (need_check) {
+            i = sycl::min(i, i_max);
+        }
+
+        const block_q3_K * bxi = bx0 + i*blocks_per_row + (k % (WARP_SIZE/4)) / (QI3_K/4);
+
+        const int ksc = k % (QI3_K/4);
+
+        const int ksc_low = ksc % (QI3_K/8);
+        const int shift_low = 4 * (ksc / (QI3_K/8));
+        const int sc_low = (get_int_from_uint8(bxi->scales, ksc_low) >> shift_low) & 0x0F0F0F0F;
+
+        const int ksc_high = QI3_K/8;
+        const int shift_high = 2 * ksc;
+        const int sc_high = ((get_int_from_uint8(bxi->scales, ksc_high) >> shift_high) << 4) & 0x30303030;
+
+        const int sc = dpct::vectorized_binary<sycl::char4>(
+            sc_low | sc_high, 0x20202020, dpct::sub_sat());
+
+        x_sc[i * (WARP_SIZE/4) + i / 4 + k % (WARP_SIZE/4)] = sc;
+    }
+}
+
+#define VDR_Q3_K_Q8_1_MMQ  2
+// contiguous u/y values
+static __dpct_inline__ float
+vec_dot_q3_K_q8_1_impl_mmq(const int *__restrict__ v, const int *__restrict__ u,
+                           const int8_t *__restrict__ scales, const float &d3,
+                           const float &d8) {
+
+    int sumi = 0;
+
+#pragma unroll
+    for (int i0 = 0; i0 < QR3_K*VDR_Q3_K_Q8_1_MMQ; i0 += QI8_1/2) {
+        int sumi_sc = 0;
+
+        for (int i = i0; i < i0 + QI8_1/2; ++i) {
+            sumi_sc = dpct::dp4a(v[i], u[i], sumi_sc); // SIMD dot product
+        }
+
+        sumi += sumi_sc * scales[i0 / (QI8_1/2)];
+    }
+
+    return d3*d8 * sumi;
+}
+
+static __dpct_inline__ float vec_dot_q3_K_q8_1_mul_mat(
+    const int *__restrict__ x_ql, const sycl::half2 *__restrict__ x_dm,
+    const int *__restrict__ x_qh, const int *__restrict__ x_sc,
+    const int *__restrict__ y_qs, const sycl::half2 *__restrict__ y_ds,
+    const int &i, const int &j, const int &k) {
+
+    const int kbx  = k / QI3_K;
+    const int ky  = (k % QI3_K) * QR3_K;
+    const float * x_dmf = (const float *) x_dm;
+    const float * y_df  = (const float *) y_ds;
+
+    const int8_t * scales = ((const int8_t *) (x_sc + i * (WARP_SIZE/4) + i/4 + kbx*4)) + ky/4;
+
+    int v[QR3_K*VDR_Q3_K_Q8_1_MMQ];
+
+#pragma unroll
+    for (int l = 0; l < QR3_K*VDR_Q3_K_Q8_1_MMQ; ++l) {
+        const int kqsx = i * (WARP_SIZE + 1) + kbx*QI3_K + (QI3_K/2) * (ky/(2*QI3_K)) + ky % (QI3_K/2);
+        const int shift = 2 * ((ky % 32) / 8);
+        const int vll = (x_ql[kqsx + l] >> shift) & 0x03030303;
+
+        const int vh = x_qh[i * (WARP_SIZE/2) + i/2 + kbx * (QI3_K/2) + (ky+l)%8] >> ((ky+l) / 8);
+        const int vlh = (vh << 2) & 0x04040404;
+
+        v[l] = dpct::vectorized_binary<sycl::char4>(vll, vlh, dpct::sub_sat());
+    }
+
+    const int index_y = j * WARP_SIZE + (k*QR3_K) % WARP_SIZE;
+    return vec_dot_q3_K_q8_1_impl_mmq(v, &y_qs[index_y], scales, x_dmf[i * (WARP_SIZE/QI3_K) + i/QI3_K + kbx], y_df[index_y/QI8_1]);
+}
+
+template <int mmq_y>
+static __dpct_inline__ void
+allocate_tiles_q4_K(int **x_ql, sycl::half2 **x_dm, int **x_qh, int **x_sc,
+                    int *tile_x_ql_q4_K, sycl::half2 *tile_x_dm_q4_K,
+                    int *tile_x_sc_q4_K) {
+    (void)x_qh;
+
+    *x_ql = tile_x_ql_q4_K;
+    *x_dm = tile_x_dm_q4_K;
+    *x_sc = tile_x_sc_q4_K;
+}
+
+template <int mmq_y, int nwarps, bool need_check>
+static __dpct_inline__ void
+load_tiles_q4_K(const void *__restrict__ vx, int *__restrict__ x_ql,
+                sycl::half2 *__restrict__ x_dm, int *__restrict__ x_qh,
+                int *__restrict__ x_sc, const int &i_offset, const int &i_max,
+                const int &k, const int &blocks_per_row) {
+    (void)x_qh;
+
+    GGML_SYCL_ASSUME(i_offset >= 0);
+    GGML_SYCL_ASSUME(i_offset <  nwarps);
+    GGML_SYCL_ASSUME(k >= 0);
+    GGML_SYCL_ASSUME(k <  WARP_SIZE);
+
+    const int kbx  = k / QI4_K; // == 0 if QK_K == 256
+    const int kqsx = k % QI4_K; // == k if QK_K == 256
+
+    const block_q4_K * bx0 = (const block_q4_K *) vx;
+
+#pragma unroll
+    for (int i0 = 0; i0 < mmq_y; i0 += nwarps) {
+        int i = i0 + i_offset;
+
+        if (need_check) {
+            i = sycl::min(i, i_max);
+        }
+
+        const block_q4_K * bxi = bx0 + i*blocks_per_row + kbx;
+
+        x_ql[i * (WARP_SIZE + 1) + k] = get_int_from_uint8_aligned(bxi->qs, kqsx);
+    }
+
+    const int blocks_per_tile_x_row = WARP_SIZE / QI4_K; // == 1 if QK_K == 256
+    const int kbxd = k % blocks_per_tile_x_row;          // == 0 if QK_K == 256
+
+#pragma unroll
+    for (int i0 = 0; i0 < mmq_y; i0 += nwarps * QI4_K) {
+        int i = (i0 + i_offset * QI4_K + k / blocks_per_tile_x_row) % mmq_y;
+
+        if (need_check) {
+            i = sycl::min(i, i_max);
+        }
+
+        const block_q4_K * bxi = bx0 + i*blocks_per_row + kbxd;
+
+#if QK_K == 256
+        x_dm[i * (WARP_SIZE/QI4_K) + i / QI4_K + kbxd] = bxi->dm;
+#else
+        x_dm[i * (WARP_SIZE/QI4_K) + i / QI4_K + kbxd] = {bxi->dm[0], bxi->dm[1]};
+#endif
+    }
+
+#pragma unroll
+    for (int i0 = 0; i0 < mmq_y; i0 += nwarps * 8) {
+        int i = (i0 + i_offset * 8 + k / (WARP_SIZE/8)) % mmq_y;
+
+        if (need_check) {
+            i = sycl::min(i, i_max);
+        }
+
+        const block_q4_K * bxi = bx0 + i*blocks_per_row + (k % (WARP_SIZE/8)) / (QI4_K/8);
+
+        const int * scales = (const int *) bxi->scales;
+
+        const int ksc = k % (WARP_SIZE/8);
+
+        // scale arrangement after the following two lines: sc0,...,sc3, sc4,...,sc7, m0,...,m3, m4,...,m8
+        int scales8 = (scales[(ksc%2) + (ksc!=0)] >> (4 * (ksc & (ksc/2)))) & 0x0F0F0F0F; // lower 4 bits
+        scales8    |= (scales[ksc/2]              >> (2 * (ksc % 2)))       & 0x30303030; // upper 2 bits
+
+        x_sc[i * (WARP_SIZE/8) + i / 8 + ksc] = scales8;
+    }
+}
+
+
+#define VDR_Q4_K_Q8_1_MMQ  8
+
+// contiguous u/y values
+static __dpct_inline__ float vec_dot_q4_K_q8_1_impl_mmq(
+    const int *__restrict__ v, const int *__restrict__ u,
+    const uint8_t *__restrict__ sc, const uint8_t *__restrict__ m,
+    const sycl::half2 &dm4, const sycl::half2 *__restrict__ ds8) {
+
+    float sumf_d = 0.0f;
+    float sumf_m = 0.0f;
+
+#pragma unroll
+    for (int i = 0; i < QR4_K*VDR_Q4_K_Q8_1_MMQ/QI8_1; ++i) {
+        int sumi_d = 0;
+
+#pragma unroll
+        for (int j = 0; j < QI8_1; ++j) {
+            sumi_d = dpct::dp4a((v[j] >> (4 * i)) & 0x0F0F0F0F,
+                                u[i * QI8_1 + j], sumi_d); // SIMD dot product
+        }
+
+        const sycl::float2 ds8f =
+            ds8[i].convert<float, sycl::rounding_mode::automatic>();
+
+        sumf_d += ds8f.x() * (sc[i] * sumi_d);
+        sumf_m += ds8f.y() * m[i]; // sum of q8_1 block * q4_K min val
+    }
+
+    const sycl::float2 dm4f =
+        dm4.convert<float, sycl::rounding_mode::automatic>();
+
+    return dm4f.x() * sumf_d - dm4f.y() * sumf_m;
+}
+
+
+static __dpct_inline__ float vec_dot_q4_K_q8_1_mul_mat(
+    const int *__restrict__ x_ql, const sycl::half2 *__restrict__ x_dm,
+    const int *__restrict__ x_qh, const int *__restrict__ x_sc,
+    const int *__restrict__ y_qs, const sycl::half2 *__restrict__ y_ds,
+    const int &i, const int &j, const int &k) {
+    (void)x_qh;
+
+    const uint8_t * sc = ((const uint8_t *) &x_sc[i * (WARP_SIZE/8) + i/8 + k/16]) + 2*((k % 16) / 8);
+
+    const int index_y = j * WARP_SIZE + (QR4_K*k) % WARP_SIZE;
+    return vec_dot_q4_K_q8_1_impl_mmq(&x_ql[i * (WARP_SIZE + 1) + k], &y_qs[index_y], sc, sc+8,
+                                      x_dm[i * (WARP_SIZE/QI4_K) + i/QI4_K], &y_ds[index_y/QI8_1]);
+}
+
+template <int mmq_y>
+static __dpct_inline__ void
+allocate_tiles_q5_K(int **x_ql, sycl::half2 **x_dm, int **x_qh, int **x_sc,
+                    int *tile_x_ql_q5_K, sycl::half2 *tile_x_dm_q5_K,
+                    int *tile_x_sc_q5_K) {
+    (void)x_qh;
+
+    *x_ql = tile_x_ql_q5_K;
+    *x_dm = tile_x_dm_q5_K;
+    *x_sc = tile_x_sc_q5_K;
+}
+
+template <int mmq_y, int nwarps, bool need_check>
+static __dpct_inline__ void
+load_tiles_q5_K(const void *__restrict__ vx, int *__restrict__ x_ql,
+                sycl::half2 *__restrict__ x_dm, int *__restrict__ x_qh,
+                int *__restrict__ x_sc, const int &i_offset, const int &i_max,
+                const int &k, const int &blocks_per_row) {
+    (void)x_qh;
+
+    GGML_SYCL_ASSUME(i_offset >= 0);
+    GGML_SYCL_ASSUME(i_offset <  nwarps);
+    GGML_SYCL_ASSUME(k >= 0);
+    GGML_SYCL_ASSUME(k <  WARP_SIZE);
+
+    const int kbx  = k / QI5_K; // == 0 if QK_K == 256
+    const int kqsx = k % QI5_K; // == k if QK_K == 256
+
+    const block_q5_K * bx0 = (const block_q5_K *) vx;
+
+#pragma unroll
+    for (int i0 = 0; i0 < mmq_y; i0 += nwarps) {
+        int i = i0 + i_offset;
+
+        if (need_check) {
+            i = sycl::min(i, i_max);
+        }
+
+        const block_q5_K * bxi = bx0 + i*blocks_per_row + kbx;
+        const int ky = QR5_K*kqsx;
+
+        const int ql = get_int_from_uint8_aligned(bxi->qs, kqsx);
+        const int ql0 = (ql >> 0) & 0x0F0F0F0F;
+        const int ql1 = (ql >> 4) & 0x0F0F0F0F;
+
+        const int qh = get_int_from_uint8_aligned(bxi->qh, kqsx % (QI5_K/4));
+        const int qh0 = ((qh >> (2 * (kqsx / (QI5_K/4)) + 0)) << 4) & 0x10101010;
+        const int qh1 = ((qh >> (2 * (kqsx / (QI5_K/4)) + 1)) << 4) & 0x10101010;
+
+        const int kq0 = ky - ky % (QI5_K/2) + k % (QI5_K/4) + 0;
+        const int kq1 = ky - ky % (QI5_K/2) + k % (QI5_K/4) + (QI5_K/4);
+
+        x_ql[i * (2*WARP_SIZE + 1) + kq0] = ql0 | qh0;
+        x_ql[i * (2*WARP_SIZE + 1) + kq1] = ql1 | qh1;
+    }
+
+    const int blocks_per_tile_x_row = WARP_SIZE / QI5_K; // == 1 if QK_K == 256
+    const int kbxd = k % blocks_per_tile_x_row;          // == 0 if QK_K == 256
+
+#pragma unroll
+    for (int i0 = 0; i0 < mmq_y; i0 += nwarps * QI5_K) {
+        int i = (i0 + i_offset * QI5_K + k / blocks_per_tile_x_row) % mmq_y;
+
+        if (need_check) {
+            i = sycl::min(i, i_max);
+        }
+
+        const block_q5_K * bxi = bx0 + i*blocks_per_row + kbxd;
+
+#if QK_K == 256
+        x_dm[i * (WARP_SIZE/QI5_K) + i / QI5_K + kbxd] = bxi->dm;
+#endif
+    }
+
+#pragma unroll
+    for (int i0 = 0; i0 < mmq_y; i0 += nwarps * 8) {
+        int i = (i0 + i_offset * 8 + k / (WARP_SIZE/8)) % mmq_y;
+
+        if (need_check) {
+            i = sycl::min(i, i_max);
+        }
+
+        const block_q5_K * bxi = bx0 + i*blocks_per_row + (k % (WARP_SIZE/8)) / (QI5_K/8);
+
+        const int * scales = (const int *) bxi->scales;
+
+        const int ksc = k % (WARP_SIZE/8);
+
+        // scale arrangement after the following two lines: sc0,...,sc3, sc4,...,sc7, m0,...,m3, m4,...,m8
+        int scales8 = (scales[(ksc%2) + (ksc!=0)] >> (4 * (ksc & (ksc/2)))) & 0x0F0F0F0F; // lower 4 bits
+        scales8    |= (scales[ksc/2]              >> (2 * (ksc % 2)))       & 0x30303030; // upper 2 bits
+
+        x_sc[i * (WARP_SIZE/8) + i / 8 + ksc] = scales8;
+    }
+}
+
+#define VDR_Q5_K_Q8_1_MMQ  8
+
+// contiguous u/y values
+static __dpct_inline__ float vec_dot_q5_K_q8_1_impl_mmq(
+    const int *__restrict__ v, const int *__restrict__ u,
+    const uint8_t *__restrict__ sc, const uint8_t *__restrict__ m,
+    const sycl::half2 &dm4, const sycl::half2 *__restrict__ ds8) {
+
+    float sumf_d = 0.0f;
+    float sumf_m = 0.0f;
+
+#pragma unroll
+    for (int i = 0; i < QR5_K*VDR_Q5_K_Q8_1_MMQ/QI8_1; ++i) {
+        int sumi_d = 0;
+
+#pragma unroll
+        for (int j = 0; j < QI8_1; ++j) {
+            sumi_d = dpct::dp4a(v[i * QI8_1 + j], u[i * QI8_1 + j],
+                                sumi_d); // SIMD dot product
+        }
+
+        const sycl::float2 ds8f =
+            ds8[i].convert<float, sycl::rounding_mode::automatic>();
+
+        sumf_d += ds8f.x() * (sc[i] * sumi_d);
+        sumf_m += ds8f.y() * m[i]; // sum of q8_1 block * q4_K min val
+    }
+
+    const sycl::float2 dm4f =
+        dm4.convert<float, sycl::rounding_mode::automatic>();
+
+    return dm4f.x() * sumf_d - dm4f.y() * sumf_m;
+}
+
+static __dpct_inline__ float vec_dot_q5_K_q8_1_mul_mat(
+    const int *__restrict__ x_ql, const sycl::half2 *__restrict__ x_dm,
+    const int *__restrict__ x_qh, const int *__restrict__ x_sc,
+    const int *__restrict__ y_qs, const sycl::half2 *__restrict__ y_ds,
+    const int &i, const int &j, const int &k) {
+    (void)x_qh;
+
+    const uint8_t * sc = ((const uint8_t *) &x_sc[i * (WARP_SIZE/8) + i/8 + k/16]) + 2 * ((k % 16) / 8);
+
+    const int index_x = i * (QR5_K*WARP_SIZE + 1) +  QR5_K*k;
+    const int index_y = j * WARP_SIZE             + (QR5_K*k) % WARP_SIZE;
+    return vec_dot_q5_K_q8_1_impl_mmq(&x_ql[index_x], &y_qs[index_y], sc, sc+8,
+                                      x_dm[i * (WARP_SIZE/QI5_K) + i/QI5_K], &y_ds[index_y/QI8_1]);
+}
+
+template <int mmq_y>
+static __dpct_inline__ void
+allocate_tiles_q6_K(int **x_ql, sycl::half2 **x_dm, int **x_qh, int **x_sc,
+                    int *tile_x_ql, sycl::half2 *tile_x_dm, int *tile_x_sc) {
+    (void)x_qh;
+
+    *x_ql = tile_x_ql;
+    *x_dm = tile_x_dm;
+    *x_sc = tile_x_sc;
+}
+
+template <int mmq_y, int nwarps, bool need_check>
+static __dpct_inline__ void
+load_tiles_q6_K(const void *__restrict__ vx, int *__restrict__ x_ql,
+                sycl::half2 *__restrict__ x_dm, int *__restrict__ x_qh,
+                int *__restrict__ x_sc, const int &i_offset, const int &i_max,
+                const int &k, const int &blocks_per_row) {
+    (void)x_qh;
+
+    GGML_SYCL_ASSUME(i_offset >= 0);
+    GGML_SYCL_ASSUME(i_offset <  nwarps);
+    GGML_SYCL_ASSUME(k >= 0);
+    GGML_SYCL_ASSUME(k <  WARP_SIZE);
+
+    const int kbx  = k / QI6_K; // == 0 if QK_K == 256
+    const int kqsx = k % QI6_K; // == k if QK_K == 256
+
+    const block_q6_K * bx0 = (const block_q6_K *) vx;
+
+#pragma unroll
+    for (int i0 = 0; i0 < mmq_y; i0 += nwarps) {
+        int i = i0 + i_offset;
+
+        if (need_check) {
+            i = sycl::min(i, i_max);
+        }
+
+        const block_q6_K * bxi = bx0 + i*blocks_per_row + kbx;
+        const int ky = QR6_K*kqsx;
+
+        const int ql = get_int_from_uint8(bxi->ql, kqsx);
+        const int ql0 = (ql >> 0) & 0x0F0F0F0F;
+        const int ql1 = (ql >> 4) & 0x0F0F0F0F;
+
+        const int qh = get_int_from_uint8(bxi->qh, (QI6_K/4) * (kqsx / (QI6_K/2)) + kqsx % (QI6_K/4));
+        const int qh0 = ((qh >> (2 * ((kqsx % (QI6_K/2)) / (QI6_K/4)))) << 4) & 0x30303030;
+        const int qh1 =  (qh >> (2 * ((kqsx % (QI6_K/2)) / (QI6_K/4))))       & 0x30303030;
+
+        const int kq0 = ky - ky % QI6_K + k % (QI6_K/2) + 0;
+        const int kq1 = ky - ky % QI6_K + k % (QI6_K/2) + (QI6_K/2);
+
+        x_ql[i * (2 * WARP_SIZE + 1) + kq0] =
+            dpct::vectorized_binary<sycl::char4>(ql0 | qh0, 0x20202020,
+                                                 dpct::sub_sat());
+        x_ql[i * (2 * WARP_SIZE + 1) + kq1] =
+            dpct::vectorized_binary<sycl::char4>(ql1 | qh1, 0x20202020,
+                                                 dpct::sub_sat());
+    }
+
+    const int blocks_per_tile_x_row = WARP_SIZE / QI6_K; // == 1 if QK_K == 256
+    const int kbxd = k % blocks_per_tile_x_row;          // == 0 if QK_K == 256
+    float * x_dmf = (float *) x_dm;
+
+#pragma unroll
+    for (int i0 = 0; i0 < mmq_y; i0 += nwarps * QI6_K) {
+        int i = (i0 + i_offset * QI6_K + k / blocks_per_tile_x_row) % mmq_y;
+
+        if (need_check) {
+            i = sycl::min(i, i_max);
+        }
+
+        const block_q6_K * bxi = bx0 + i*blocks_per_row + kbxd;
+
+        x_dmf[i * (WARP_SIZE/QI6_K) + i / QI6_K + kbxd] = bxi->d;
+    }
+
+#pragma unroll
+    for (int i0 = 0; i0 < mmq_y; i0 += nwarps * 8) {
+        int i = (i0 + i_offset * 8 + k / (WARP_SIZE/8)) % mmq_y;
+
+        if (need_check) {
+            i = sycl::min(i, i_max);
+        }
+
+        const block_q6_K * bxi = bx0 + i*blocks_per_row + (k % (WARP_SIZE/8)) / 4;
+
+        x_sc[i * (WARP_SIZE/8) + i / 8 + k % (WARP_SIZE/8)] = get_int_from_int8(bxi->scales, k % (QI6_K/8));
+    }
+}
+
+#define VDR_Q6_K_Q8_1_MMQ  8
+
+// contiguous u/y values
+static __dpct_inline__ float
+vec_dot_q6_K_q8_1_impl_mmq(const int *__restrict__ v, const int *__restrict__ u,
+                           const int8_t *__restrict__ sc, const float &d6,
+                           const float *__restrict__ d8) {
+
+    float sumf_d = 0.0f;
+
+#pragma unroll
+    for (int i0 = 0; i0 < VDR_Q6_K_Q8_1_MMQ; i0 += 4) {
+        sycl::int2 sumi_d = {0, 0}; // 2 q6_K scales per q8_1 scale
+
+#pragma unroll
+        for (int i = i0; i < i0 + 2; ++i) {
+            sumi_d.x() = dpct::dp4a(v[2 * i + 0], u[2 * i + 0],
+                                    sumi_d.x()); // SIMD dot product
+            sumi_d.x() = dpct::dp4a(v[2 * i + 1], u[2 * i + 1],
+                                    sumi_d.x()); // SIMD dot product
+
+            sumi_d.y() = dpct::dp4a(v[2 * i + 4], u[2 * i + 4],
+                                    sumi_d.y()); // SIMD dot product
+            sumi_d.y() = dpct::dp4a(v[2 * i + 5], u[2 * i + 5],
+                                    sumi_d.y()); // SIMD dot product
+        }
+
+        sumf_d += d8[i0 / 4] *
+                  (sc[i0 / 2 + 0] * sumi_d.x() + sc[i0 / 2 + 1] * sumi_d.y());
+    }
+
+    return d6 * sumf_d;
+}
+
+static __dpct_inline__ float vec_dot_q6_K_q8_1_mul_mat(
+    const int *__restrict__ x_ql, const sycl::half2 *__restrict__ x_dm,
+    const int *__restrict__ x_qh, const int *__restrict__ x_sc,
+    const int *__restrict__ y_qs, const sycl::half2 *__restrict__ y_ds,
+    const int &i, const int &j, const int &k) {
+    (void)x_qh;
+
+    const float * x_dmf = (const float *) x_dm;
+    const float * y_df  = (const float *) y_ds;
+
+    const int8_t * sc = ((const int8_t *) &x_sc[i * (WARP_SIZE/8) + i/8 + k/8]);
+
+    const int index_x = i * (QR6_K*WARP_SIZE + 1) +  QR6_K*k;
+    const int index_y = j * WARP_SIZE             + (QR6_K*k) % WARP_SIZE;
+    return vec_dot_q6_K_q8_1_impl_mmq(&x_ql[index_x], &y_qs[index_y], sc, x_dmf[i * (WARP_SIZE/QI6_K) + i/QI6_K], &y_df[index_y/QI8_1]);
+}
+
+template <int qk, int qr, int qi, bool need_sum, typename block_q_t, int mmq_x,
+          int mmq_y, int nwarps, load_tiles_sycl_t load_tiles, int vdr,
+          vec_dot_q_mul_mat_sycl_t vec_dot>
+/*
+DPCT1110:8: The total declared local variable size in device function mul_mat_q
+exceeds 128 bytes and may cause high register pressure. Consult with your
+hardware vendor to find the total register size available and adjust the code,
+or use smaller sub-group size to avoid high register pressure.
+*/
+static __dpct_inline__ void
+mul_mat_q(const void *__restrict__ vx, const void *__restrict__ vy,
+          float *__restrict__ dst, const int ncols_x, const int nrows_x,
+          const int ncols_y, const int nrows_y, const int nrows_dst,
+          int *tile_x_ql, sycl::half2 *tile_x_dm, int *tile_x_qh,
+          int *tile_x_sc, const sycl::nd_item<3> &item_ct1, int *tile_y_qs,
+          sycl::half2 *tile_y_ds) {
+
+    const block_q_t  * x = (const block_q_t  *) vx;
+    const block_q8_1 * y = (const block_q8_1 *) vy;
+
+    const int blocks_per_row_x = ncols_x / qk;
+    const int blocks_per_col_y = nrows_y / QK8_1;
+    const int blocks_per_warp = WARP_SIZE / qi;
+
+    const int & ncols_dst = ncols_y;
+
+    const int row_dst_0 = item_ct1.get_group(2) * mmq_y;
+    const int & row_x_0 = row_dst_0;
+
+    const int col_dst_0 = item_ct1.get_group(1) * mmq_x;
+    const int & col_y_0 = col_dst_0;
+
+    float sum[mmq_y/WARP_SIZE][mmq_x/nwarps] = {{0.0f}};
+
+    for (int ib0 = 0; ib0 < blocks_per_row_x; ib0 += blocks_per_warp) {
+
+        load_tiles(x + row_x_0 * blocks_per_row_x + ib0, tile_x_ql, tile_x_dm,
+                   tile_x_qh, tile_x_sc, item_ct1.get_local_id(1),
+                   nrows_x - row_x_0 - 1, item_ct1.get_local_id(2),
+                   blocks_per_row_x);
+
+#pragma unroll
+        for (int ir = 0; ir < qr; ++ir) {
+            const int kqs = ir * WARP_SIZE + item_ct1.get_local_id(2);
+            const int kbxd = kqs / QI8_1;
+
+#pragma unroll
+            for (int i = 0; i < mmq_x; i += nwarps) {
+                const int col_y_eff = dpct::min(
+                    (unsigned int)(col_y_0 + item_ct1.get_local_id(1) + i),
+                    ncols_y - 1); // to prevent out-of-bounds memory accesses
+
+                const block_q8_1 * by0 = &y[col_y_eff*blocks_per_col_y + ib0 * (qk/QK8_1) + kbxd];
+
+                const int index_y = (item_ct1.get_local_id(1) + i) * WARP_SIZE +
+                                    kqs % WARP_SIZE;
+                tile_y_qs[index_y] = get_int_from_int8_aligned(
+                    by0->qs, item_ct1.get_local_id(2) % QI8_1);
+            }
+
+#pragma unroll
+            for (int ids0 = 0; ids0 < mmq_x; ids0 += nwarps * QI8_1) {
+                const int ids =
+                    (ids0 + item_ct1.get_local_id(1) * QI8_1 +
+                     item_ct1.get_local_id(2) / (WARP_SIZE / QI8_1)) %
+                    mmq_x;
+                const int kby = item_ct1.get_local_id(2) % (WARP_SIZE / QI8_1);
+                const int col_y_eff = sycl::min(col_y_0 + ids, ncols_y - 1);
+
+                // if the sum is not needed it's faster to transform the scale to f32 ahead of time
+                const sycl::half2 *dsi_src =
+                    &y[col_y_eff * blocks_per_col_y + ib0 * (qk / QK8_1) +
+                       ir * (WARP_SIZE / QI8_1) + kby]
+                         .ds;
+                sycl::half2 *dsi_dst =
+                    &tile_y_ds[ids * (WARP_SIZE / QI8_1) + kby];
+                if (need_sum) {
+                    *dsi_dst = *dsi_src;
+                } else {
+                    float * dfi_dst = (float *) dsi_dst;
+                    *dfi_dst = (*dsi_src)[0];
+                }
+            }
+
+            /*
+            DPCT1118:9: SYCL group functions and algorithms must be encountered
+            in converged control flow. You may need to adjust the code.
+            */
+            /*
+            DPCT1065:56: Consider replacing sycl::nd_item::barrier() with
+            sycl::nd_item::barrier(sycl::access::fence_space::local_space) for
+            better performance if there is no access to global memory.
+            */
+            item_ct1.barrier();
+
+// #pragma unroll // unrolling this loop causes too much register pressure
+            for (int k = ir*WARP_SIZE/qr; k < (ir+1)*WARP_SIZE/qr; k += vdr) {
+#pragma unroll
+                for (int j = 0; j < mmq_x; j += nwarps) {
+#pragma unroll
+                    for (int i = 0; i < mmq_y; i += WARP_SIZE) {
+                        sum[i / WARP_SIZE][j / nwarps] += vec_dot(
+                            tile_x_ql, tile_x_dm, tile_x_qh, tile_x_sc,
+                            tile_y_qs, tile_y_ds, item_ct1.get_local_id(2) + i,
+                            item_ct1.get_local_id(1) + j, k);
+                    }
+                }
+            }
+
+            /*
+            DPCT1118:10: SYCL group functions and algorithms must be encountered
+            in converged control flow. You may need to adjust the code.
+            */
+            /*
+            DPCT1065:57: Consider replacing sycl::nd_item::barrier() with
+            sycl::nd_item::barrier(sycl::access::fence_space::local_space) for
+            better performance if there is no access to global memory.
+            */
+            item_ct1.barrier();
+        }
+    }
+
+#pragma unroll
+    for (int j = 0; j < mmq_x; j += nwarps) {
+        const int col_dst = col_dst_0 + j + item_ct1.get_local_id(1);
+
+        if (col_dst >= ncols_dst) {
+            return;
+        }
+
+#pragma unroll
+        for (int i = 0; i < mmq_y; i += WARP_SIZE) {
+            const int row_dst = row_dst_0 + item_ct1.get_local_id(2) + i;
+
+            if (row_dst >= nrows_dst) {
+                continue;
+            }
+
+            dst[col_dst*nrows_dst + row_dst] = sum[i/WARP_SIZE][j/nwarps];
+        }
+    }
+}
+
+#define  MMQ_X_Q4_0_RDNA2  64
+#define  MMQ_Y_Q4_0_RDNA2  128
+#define NWARPS_Q4_0_RDNA2  8
+#define  MMQ_X_Q4_0_RDNA1  64
+#define  MMQ_Y_Q4_0_RDNA1  64
+#define NWARPS_Q4_0_RDNA1  8
+#if defined(SYCL_USE_XMX)
+#define  MMQ_X_Q4_0_AMPERE 4
+#define  MMQ_Y_Q4_0_AMPERE 32
+#define NWARPS_Q4_0_AMPERE 4
+#else
+#define  MMQ_X_Q4_0_AMPERE 64
+#define  MMQ_Y_Q4_0_AMPERE 128
+#define NWARPS_Q4_0_AMPERE 4
+#endif
+#define  MMQ_X_Q4_0_PASCAL 64
+#define  MMQ_Y_Q4_0_PASCAL 64
+#define NWARPS_Q4_0_PASCAL 8
+
+template <bool need_check> static void
+    mul_mat_q4_0(
+    const void * __restrict__ vx, const void * __restrict__ vy, float * __restrict__ dst,
+    const int ncols_x, const int nrows_x, const int ncols_y, const int nrows_y, const int nrows_dst,
+    const sycl::nd_item<3> &item_ct1, int *tile_x_qs_q4_0, float *tile_x_d_q4_0,
+    int *tile_y_qs, sycl::half2 *tile_y_ds) {
+    int   * tile_x_ql = nullptr;
+    sycl::half2 *tile_x_dm = nullptr;
+    int   * tile_x_qh = nullptr;
+    int   * tile_x_sc = nullptr;
+
+//sycl_todo: change according to hardware
+
+    const int mmq_x  =  MMQ_X_Q4_0_AMPERE;
+    const int mmq_y  =  MMQ_Y_Q4_0_AMPERE;
+    const int nwarps = NWARPS_Q4_0_AMPERE;
+    allocate_tiles_q4_0<mmq_y>(&tile_x_ql, &tile_x_dm, &tile_x_qh, &tile_x_sc,
+                               tile_x_qs_q4_0, tile_x_d_q4_0);
+    mul_mat_q<QK4_0, QR4_0, QI4_0, true, block_q4_0, mmq_x, mmq_y, nwarps,
+              load_tiles_q4_0<mmq_y, nwarps, need_check>, VDR_Q4_0_Q8_1_MMQ,
+              vec_dot_q4_0_q8_1_mul_mat>(
+        vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst, tile_x_ql,
+        tile_x_dm, tile_x_qh, tile_x_sc, item_ct1, tile_y_qs, tile_y_ds);
+}
+
+#define  MMQ_X_Q4_1_RDNA2  64
+#define  MMQ_Y_Q4_1_RDNA2  128
+#define NWARPS_Q4_1_RDNA2  8
+#define  MMQ_X_Q4_1_RDNA1  64
+#define  MMQ_Y_Q4_1_RDNA1  64
+#define NWARPS_Q4_1_RDNA1  8
+#if defined(SYCL_USE_XMX)
+#define  MMQ_X_Q4_1_AMPERE 4
+#define  MMQ_Y_Q4_1_AMPERE 32
+#define NWARPS_Q4_1_AMPERE 4
+#else
+#define  MMQ_X_Q4_1_AMPERE 64
+#define  MMQ_Y_Q4_1_AMPERE 128
+#define NWARPS_Q4_1_AMPERE 4
+#endif
+#define  MMQ_X_Q4_1_PASCAL 64
+#define  MMQ_Y_Q4_1_PASCAL 64
+#define NWARPS_Q4_1_PASCAL 8
+
+template <bool need_check> static void
+    mul_mat_q4_1(
+    const void * __restrict__ vx, const void * __restrict__ vy, float * __restrict__ dst,
+    const int ncols_x, const int nrows_x, const int ncols_y, const int nrows_y, const int nrows_dst,
+    const sycl::nd_item<3> &item_ct1, int *tile_x_qs_q4_1,
+    sycl::half2 *tile_x_dm_q4_1, int *tile_y_qs, sycl::half2 *tile_y_ds) {
+    int   * tile_x_ql = nullptr;
+    sycl::half2 *tile_x_dm = nullptr;
+    int   * tile_x_qh = nullptr;
+    int   * tile_x_sc = nullptr;
+
+//sycl_todo: change according to hardware
+    const int mmq_x  =  MMQ_X_Q4_1_AMPERE;
+    const int mmq_y  =  MMQ_Y_Q4_1_AMPERE;
+    const int nwarps = NWARPS_Q4_1_AMPERE;
+    allocate_tiles_q4_1<mmq_y>(&tile_x_ql, &tile_x_dm, &tile_x_qh, &tile_x_sc,
+                               tile_x_qs_q4_1, tile_x_dm_q4_1);
+    mul_mat_q<QK4_1, QR4_1, QI4_1, true, block_q4_1, mmq_x, mmq_y, nwarps,
+              load_tiles_q4_1<mmq_y, nwarps, need_check>, VDR_Q4_1_Q8_1_MMQ,
+              vec_dot_q4_1_q8_1_mul_mat>(
+        vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst, tile_x_ql,
+        tile_x_dm, tile_x_qh, tile_x_sc, item_ct1, tile_y_qs, tile_y_ds);
+}
+
+#define  MMQ_X_Q5_0_RDNA2  64
+#define  MMQ_Y_Q5_0_RDNA2  128
+#define NWARPS_Q5_0_RDNA2  8
+#define  MMQ_X_Q5_0_RDNA1  64
+#define  MMQ_Y_Q5_0_RDNA1  64
+#define NWARPS_Q5_0_RDNA1  8
+#if defined(SYCL_USE_XMX)
+#define  MMQ_X_Q5_0_AMPERE 4
+#define  MMQ_Y_Q5_0_AMPERE 32
+#define NWARPS_Q5_0_AMPERE 4
+#else
+#define  MMQ_X_Q5_0_AMPERE 128
+#define  MMQ_Y_Q5_0_AMPERE 64
+#define NWARPS_Q5_0_AMPERE 4
+#endif
+#define  MMQ_X_Q5_0_PASCAL 64
+#define  MMQ_Y_Q5_0_PASCAL 64
+#define NWARPS_Q5_0_PASCAL 8
+
+template <bool need_check> static void
+    mul_mat_q5_0(
+    const void * __restrict__ vx, const void * __restrict__ vy, float * __restrict__ dst,
+    const int ncols_x, const int nrows_x, const int ncols_y, const int nrows_y, const int nrows_dst,
+    const sycl::nd_item<3> &item_ct1, int *tile_x_ql_q5_0, float *tile_x_d_q5_0,
+    int *tile_y_qs, sycl::half2 *tile_y_ds) {
+    int   * tile_x_ql = nullptr;
+    sycl::half2 *tile_x_dm = nullptr;
+    int   * tile_x_qh = nullptr;
+    int   * tile_x_sc = nullptr;
+
+//sycl_todo: change according to hardware
+    const int mmq_x  =  MMQ_X_Q5_0_AMPERE;
+    const int mmq_y  =  MMQ_Y_Q5_0_AMPERE;
+    const int nwarps = NWARPS_Q5_0_AMPERE;
+    allocate_tiles_q5_0<mmq_y>(&tile_x_ql, &tile_x_dm, &tile_x_qh, &tile_x_sc,
+                               tile_x_ql_q5_0, tile_x_d_q5_0);
+    mul_mat_q<QK5_0, QR5_0, QI5_0, false, block_q5_0, mmq_x, mmq_y, nwarps,
+              load_tiles_q5_0<mmq_y, nwarps, need_check>, VDR_Q5_0_Q8_1_MMQ,
+              vec_dot_q5_0_q8_1_mul_mat>(
+        vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst, tile_x_ql,
+        tile_x_dm, tile_x_qh, tile_x_sc, item_ct1, tile_y_qs, tile_y_ds);
+}
+
+#define  MMQ_X_Q5_1_RDNA2  64
+#define  MMQ_Y_Q5_1_RDNA2  128
+#define NWARPS_Q5_1_RDNA2  8
+#define  MMQ_X_Q5_1_RDNA1  64
+#define  MMQ_Y_Q5_1_RDNA1  64
+#define NWARPS_Q5_1_RDNA1  8
+#if defined(SYCL_USE_XMX)
+#define  MMQ_X_Q5_1_AMPERE 4
+#define  MMQ_Y_Q5_1_AMPERE 32
+#define NWARPS_Q5_1_AMPERE 4
+#else
+#define  MMQ_X_Q5_1_AMPERE 128
+#define  MMQ_Y_Q5_1_AMPERE 64
+#define NWARPS_Q5_1_AMPERE 4
+#endif
+#define  MMQ_X_Q5_1_PASCAL 64
+#define  MMQ_Y_Q5_1_PASCAL 64
+#define NWARPS_Q5_1_PASCAL 8
+
+template <bool need_check> static void
+mul_mat_q5_1(
+    const void * __restrict__ vx, const void * __restrict__ vy, float * __restrict__ dst,
+    const int ncols_x, const int nrows_x, const int ncols_y, const int nrows_y, const int nrows_dst,
+    const sycl::nd_item<3> &item_ct1, int *tile_x_ql_q5_1,
+    sycl::half2 *tile_x_dm_q5_1, int *tile_y_qs, sycl::half2 *tile_y_ds) {
+    int   * tile_x_ql = nullptr;
+    sycl::half2 *tile_x_dm = nullptr;
+    int   * tile_x_qh = nullptr;
+    int   * tile_x_sc = nullptr;
+
+//sycl_todo: change according to hardware
+    const int mmq_x  =  MMQ_X_Q5_1_AMPERE;
+    const int mmq_y  =  MMQ_Y_Q5_1_AMPERE;
+    const int nwarps = NWARPS_Q5_1_AMPERE;
+    allocate_tiles_q5_1<mmq_y>(&tile_x_ql, &tile_x_dm, &tile_x_qh, &tile_x_sc,
+                               tile_x_ql_q5_1, tile_x_dm_q5_1);
+    mul_mat_q<QK5_1, QR5_1, QI5_1, true, block_q5_1, mmq_x, mmq_y, nwarps,
+              load_tiles_q5_1<mmq_y, nwarps, need_check>, VDR_Q5_1_Q8_1_MMQ,
+              vec_dot_q5_1_q8_1_mul_mat>(
+        vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst, tile_x_ql,
+        tile_x_dm, tile_x_qh, tile_x_sc, item_ct1, tile_y_qs, tile_y_ds);
+}
+
+#define  MMQ_X_Q8_0_RDNA2  64
+#define  MMQ_Y_Q8_0_RDNA2  128
+#define NWARPS_Q8_0_RDNA2  8
+#define  MMQ_X_Q8_0_RDNA1  64
+#define  MMQ_Y_Q8_0_RDNA1  64
+#define NWARPS_Q8_0_RDNA1  8
+#if defined(SYCL_USE_XMX)
+#define  MMQ_X_Q8_0_AMPERE 4
+#define  MMQ_Y_Q8_0_AMPERE 32
+#define NWARPS_Q8_0_AMPERE 4
+#else
+#define  MMQ_X_Q8_0_AMPERE 128
+#define  MMQ_Y_Q8_0_AMPERE 64
+#define NWARPS_Q8_0_AMPERE 4
+#endif
+#define  MMQ_X_Q8_0_PASCAL 64
+#define  MMQ_Y_Q8_0_PASCAL 64
+#define NWARPS_Q8_0_PASCAL 8
+
+template <bool need_check> static void
+    mul_mat_q8_0(
+    const void * __restrict__ vx, const void * __restrict__ vy, float * __restrict__ dst,
+    const int ncols_x, const int nrows_x, const int ncols_y, const int nrows_y, const int nrows_dst,
+    const sycl::nd_item<3> &item_ct1, int *tile_x_qs_q8_0, float *tile_x_d_q8_0,
+    int *tile_y_qs, sycl::half2 *tile_y_ds) {
+    int   * tile_x_ql = nullptr;
+    sycl::half2 *tile_x_dm = nullptr;
+    int   * tile_x_qh = nullptr;
+    int   * tile_x_sc = nullptr;
+
+//sycl_todo: change according to hardware
+    const int mmq_x  =  MMQ_X_Q8_0_AMPERE;
+    const int mmq_y  =  MMQ_Y_Q8_0_AMPERE;
+    const int nwarps = NWARPS_Q8_0_AMPERE;
+    allocate_tiles_q8_0<mmq_y>(&tile_x_ql, &tile_x_dm, &tile_x_qh, &tile_x_sc,
+                               tile_x_qs_q8_0, tile_x_d_q8_0);
+    mul_mat_q<QK8_0, QR8_0, QI8_0, false, block_q8_0, mmq_x, mmq_y, nwarps,
+              load_tiles_q8_0<mmq_y, nwarps, need_check>, VDR_Q8_0_Q8_1_MMQ,
+              vec_dot_q8_0_q8_1_mul_mat>(
+        vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst, tile_x_ql,
+        tile_x_dm, tile_x_qh, tile_x_sc, item_ct1, tile_y_qs, tile_y_ds);
+}
+
+#define  MMQ_X_Q2_K_RDNA2  64
+#define  MMQ_Y_Q2_K_RDNA2  128
+#define NWARPS_Q2_K_RDNA2  8
+#define  MMQ_X_Q2_K_RDNA1  128
+#define  MMQ_Y_Q2_K_RDNA1  32
+#define NWARPS_Q2_K_RDNA1  8
+#if defined(SYCL_USE_XMX)
+#define  MMQ_X_Q2_K_AMPERE 4
+#define  MMQ_Y_Q2_K_AMPERE 32
+#define NWARPS_Q2_K_AMPERE 4
+#else
+#define  MMQ_X_Q2_K_AMPERE 64
+#define  MMQ_Y_Q2_K_AMPERE 128
+#define NWARPS_Q2_K_AMPERE 4
+#endif
+#define  MMQ_X_Q2_K_PASCAL 64
+#define  MMQ_Y_Q2_K_PASCAL 64
+#define NWARPS_Q2_K_PASCAL 8
+
+template <bool need_check> static void
+mul_mat_q2_K(
+    const void * __restrict__ vx, const void * __restrict__ vy, float * __restrict__ dst,
+    const int ncols_x, const int nrows_x, const int ncols_y, const int nrows_y, const int nrows_dst,
+    const sycl::nd_item<3> &item_ct1, int *tile_x_ql_q2_K,
+    sycl::half2 *tile_x_dm_q2_K, int *tile_x_sc_q2_K, int *tile_y_qs,
+    sycl::half2 *tile_y_ds) {
+    int   * tile_x_ql = nullptr;
+    sycl::half2 *tile_x_dm = nullptr;
+    int   * tile_x_qh = nullptr;
+    int   * tile_x_sc = nullptr;
+
+//sycl_todo: change according to hardware
+    const int mmq_x  =  MMQ_X_Q2_K_AMPERE;
+    const int mmq_y  =  MMQ_Y_Q2_K_AMPERE;
+    const int nwarps = NWARPS_Q2_K_AMPERE;
+    allocate_tiles_q2_K<mmq_y>(&tile_x_ql, &tile_x_dm, &tile_x_qh, &tile_x_sc,
+                               tile_x_ql_q2_K, tile_x_dm_q2_K, tile_x_sc_q2_K);
+    mul_mat_q<QK_K, QR2_K, QI2_K, false, block_q2_K, mmq_x, mmq_y, nwarps,
+              load_tiles_q2_K<mmq_y, nwarps, need_check>, VDR_Q2_K_Q8_1_MMQ,
+              vec_dot_q2_K_q8_1_mul_mat>(
+        vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst, tile_x_ql,
+        tile_x_dm, tile_x_qh, tile_x_sc, item_ct1, tile_y_qs, tile_y_ds);
+}
+
+#define  MMQ_X_Q3_K_RDNA2  128
+#define  MMQ_Y_Q3_K_RDNA2  64
+#define NWARPS_Q3_K_RDNA2  8
+#define  MMQ_X_Q3_K_RDNA1  32
+#define  MMQ_Y_Q3_K_RDNA1  128
+#define NWARPS_Q3_K_RDNA1  8
+#if defined(SYCL_USE_XMX)
+#define  MMQ_X_Q3_K_AMPERE 4
+#define  MMQ_Y_Q3_K_AMPERE 32
+#define NWARPS_Q3_K_AMPERE 4
+#else
+#define  MMQ_X_Q3_K_AMPERE 128
+#define  MMQ_Y_Q3_K_AMPERE 128
+#define NWARPS_Q3_K_AMPERE 4
+#endif
+#define  MMQ_X_Q3_K_PASCAL 64
+#define  MMQ_Y_Q3_K_PASCAL 64
+#define NWARPS_Q3_K_PASCAL 8
+
+template <bool need_check> static void
+mul_mat_q3_K(
+    const void * __restrict__ vx, const void * __restrict__ vy, float * __restrict__ dst,
+    const int ncols_x, const int nrows_x, const int ncols_y, const int nrows_y, const int nrows_dst,
+    const sycl::nd_item<3> &item_ct1, int *tile_x_ql_q3_K,
+    sycl::half2 *tile_x_dm_q3_K, int *tile_x_qh_q3_K, int *tile_x_sc_q3_K,
+    int *tile_y_qs, sycl::half2 *tile_y_ds) {
+    int   * tile_x_ql = nullptr;
+    sycl::half2 *tile_x_dm = nullptr;
+    int   * tile_x_qh = nullptr;
+    int   * tile_x_sc = nullptr;
+
+//sycl_todo: change according to hardware
+    const int mmq_x  =  MMQ_X_Q3_K_AMPERE;
+    const int mmq_y  =  MMQ_Y_Q3_K_AMPERE;
+    const int nwarps = NWARPS_Q3_K_AMPERE;
+    allocate_tiles_q3_K<mmq_y>(&tile_x_ql, &tile_x_dm, &tile_x_qh, &tile_x_sc,
+                               tile_x_ql_q3_K, tile_x_dm_q3_K, tile_x_qh_q3_K,
+                               tile_x_sc_q3_K);
+    mul_mat_q<QK_K, QR3_K, QI3_K, false, block_q3_K, mmq_x, mmq_y, nwarps,
+              load_tiles_q3_K<mmq_y, nwarps, need_check>, VDR_Q3_K_Q8_1_MMQ,
+              vec_dot_q3_K_q8_1_mul_mat>(
+        vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst, tile_x_ql,
+        tile_x_dm, tile_x_qh, tile_x_sc, item_ct1, tile_y_qs, tile_y_ds);
+}
+
+#define  MMQ_X_Q4_K_RDNA2  64
+#define  MMQ_Y_Q4_K_RDNA2  128
+#define NWARPS_Q4_K_RDNA2  8
+#define  MMQ_X_Q4_K_RDNA1  32
+#define  MMQ_Y_Q4_K_RDNA1  64
+#define NWARPS_Q4_K_RDNA1  8
+#if defined(SYCL_USE_XMX)
+#define  MMQ_X_Q4_K_AMPERE 4
+#define  MMQ_Y_Q4_K_AMPERE 32
+#define NWARPS_Q4_K_AMPERE 4
+#else
+#define  MMQ_X_Q4_K_AMPERE 64
+#define  MMQ_Y_Q4_K_AMPERE 128
+#define NWARPS_Q4_K_AMPERE 4
+#endif
+#define  MMQ_X_Q4_K_PASCAL 64
+#define  MMQ_Y_Q4_K_PASCAL 64
+#define NWARPS_Q4_K_PASCAL 8
+
+template <bool need_check> static void
+    mul_mat_q4_K(
+    const void * __restrict__ vx, const void * __restrict__ vy, float * __restrict__ dst,
+    const int ncols_x, const int nrows_x, const int ncols_y, const int nrows_y, const int nrows_dst,
+    const sycl::nd_item<3> &item_ct1, int *tile_x_ql_q4_K,
+    sycl::half2 *tile_x_dm_q4_K, int *tile_x_sc_q4_K, int *tile_y_qs,
+    sycl::half2 *tile_y_ds) {
+    int   * tile_x_ql = nullptr;
+    sycl::half2 *tile_x_dm = nullptr;
+    int   * tile_x_qh = nullptr;
+    int   * tile_x_sc = nullptr;
+
+//sycl_todo: change according to hardware
+    const int mmq_x  =  MMQ_X_Q4_K_AMPERE;
+    const int mmq_y  =  MMQ_Y_Q4_K_AMPERE;
+    const int nwarps = NWARPS_Q4_K_AMPERE;
+    allocate_tiles_q4_K<mmq_y>(&tile_x_ql, &tile_x_dm, &tile_x_qh, &tile_x_sc,
+                               tile_x_ql_q4_K, tile_x_dm_q4_K, tile_x_sc_q4_K);
+    mul_mat_q<QK_K, QR4_K, QI4_K, true, block_q4_K, mmq_x, mmq_y, nwarps,
+              load_tiles_q4_K<mmq_y, nwarps, need_check>, VDR_Q4_K_Q8_1_MMQ,
+              vec_dot_q4_K_q8_1_mul_mat>(
+        vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst, tile_x_ql,
+        tile_x_dm, tile_x_qh, tile_x_sc, item_ct1, tile_y_qs, tile_y_ds);
+}
+
+#define  MMQ_X_Q5_K_RDNA2  64
+#define  MMQ_Y_Q5_K_RDNA2  128
+#define NWARPS_Q5_K_RDNA2  8
+#define  MMQ_X_Q5_K_RDNA1  32
+#define  MMQ_Y_Q5_K_RDNA1  64
+#define NWARPS_Q5_K_RDNA1  8
+#if defined(SYCL_USE_XMX)
+#define  MMQ_X_Q5_K_AMPERE 4
+#define  MMQ_Y_Q5_K_AMPERE 32
+#define NWARPS_Q5_K_AMPERE 4
+#else
+#define  MMQ_X_Q5_K_AMPERE 64
+#define  MMQ_Y_Q5_K_AMPERE 128
+#define NWARPS_Q5_K_AMPERE 4
+#endif
+#define  MMQ_X_Q5_K_PASCAL 64
+#define  MMQ_Y_Q5_K_PASCAL 64
+#define NWARPS_Q5_K_PASCAL 8
+
+template <bool need_check> static void
+mul_mat_q5_K(
+    const void * __restrict__ vx, const void * __restrict__ vy, float * __restrict__ dst,
+    const int ncols_x, const int nrows_x, const int ncols_y, const int nrows_y, const int nrows_dst,
+    const sycl::nd_item<3> &item_ct1, int *tile_x_ql_q5_K,
+    sycl::half2 *tile_x_dm_q5_K, int *tile_x_sc_q5_K, int *tile_y_qs,
+    sycl::half2 *tile_y_ds) {
+    int   * tile_x_ql = nullptr;
+    sycl::half2 *tile_x_dm = nullptr;
+    int   * tile_x_qh = nullptr;
+    int   * tile_x_sc = nullptr;
+
+//sycl_todo: change according to hardware
+    const int mmq_x  =  MMQ_X_Q5_K_AMPERE;
+    const int mmq_y  =  MMQ_Y_Q5_K_AMPERE;
+    const int nwarps = NWARPS_Q5_K_AMPERE;
+    allocate_tiles_q5_K<mmq_y>(&tile_x_ql, &tile_x_dm, &tile_x_qh, &tile_x_sc,
+                               tile_x_ql_q5_K, tile_x_dm_q5_K, tile_x_sc_q5_K);
+    mul_mat_q<QK_K, QR5_K, QI5_K, true, block_q5_K, mmq_x, mmq_y, nwarps,
+              load_tiles_q5_K<mmq_y, nwarps, need_check>, VDR_Q5_K_Q8_1_MMQ,
+              vec_dot_q5_K_q8_1_mul_mat>(
+        vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst, tile_x_ql,
+        tile_x_dm, tile_x_qh, tile_x_sc, item_ct1, tile_y_qs, tile_y_ds);
+}
+
+#define  MMQ_X_Q6_K_RDNA2  64
+#define  MMQ_Y_Q6_K_RDNA2  128
+#define NWARPS_Q6_K_RDNA2  8
+#define  MMQ_X_Q6_K_RDNA1  32
+#define  MMQ_Y_Q6_K_RDNA1  64
+#define NWARPS_Q6_K_RDNA1  8
+#if defined(SYCL_USE_XMX)
+#define  MMQ_X_Q6_K_AMPERE 4
+#define  MMQ_Y_Q6_K_AMPERE 32
+#define NWARPS_Q6_K_AMPERE 4
+#else
+#define  MMQ_X_Q6_K_AMPERE 64
+#define  MMQ_Y_Q6_K_AMPERE 64
+#define NWARPS_Q6_K_AMPERE 4
+#endif
+#define  MMQ_X_Q6_K_PASCAL 64
+#define  MMQ_Y_Q6_K_PASCAL 64
+#define NWARPS_Q6_K_PASCAL 8
+
+template <bool need_check> static void
+    mul_mat_q6_K(
+    const void * __restrict__ vx, const void * __restrict__ vy, float * __restrict__ dst,
+    const int ncols_x, const int nrows_x, const int ncols_y, const int nrows_y, const int nrows_dst,
+    const sycl::nd_item<3> &item_ct1, int *tile_x_ql, sycl::half2 *tile_x_dm,
+    int *tile_x_sc, int *tile_y_qs, sycl::half2 *tile_y_ds) {
+    // int   * tile_x_ql = nullptr;
+    // sycl::half2 *tile_x_dm = nullptr;
+    int   * tile_x_qh = nullptr;
+    // int   * tile_x_sc = nullptr;
+
+//sycl_todo: change according to hardware
+    const int mmq_x  =  MMQ_X_Q6_K_AMPERE;
+    const int mmq_y  =  MMQ_Y_Q6_K_AMPERE;
+    const int nwarps = NWARPS_Q6_K_AMPERE;
+    allocate_tiles_q6_K<mmq_y>(&tile_x_ql, &tile_x_dm, &tile_x_qh, &tile_x_sc,
+                               tile_x_ql, tile_x_dm, tile_x_sc);
+    mul_mat_q<QK_K, QR6_K, QI6_K, false, block_q6_K, mmq_x, mmq_y, nwarps,
+              load_tiles_q6_K<mmq_y, nwarps, need_check>, VDR_Q6_K_Q8_1_MMQ,
+              vec_dot_q6_K_q8_1_mul_mat>(
+        vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y, nrows_dst, tile_x_ql,
+        tile_x_dm, tile_x_qh, tile_x_sc, item_ct1, tile_y_qs, tile_y_ds);
+}
+
+static void ggml_mul_mat_q4_0_q8_1_sycl(const void *vx, const void *vy,
+                                        float *dst, const int ncols_x,
+                                        const int nrows_x, const int ncols_y,
+                                        const int nrows_y, const int nrows_dst,
+                                        dpct::queue_ptr stream) try {
+
+    int id;
+    SYCL_CHECK(
+        CHECK_TRY_ERROR(id = get_current_device_id()));
+    const int compute_capability = ggml_sycl_info().devices[id].cc;
+
+    int mmq_x, mmq_y, nwarps;
+    if (compute_capability >= VER_GEN13) {
+        mmq_x  =  MMQ_X_Q4_0_RDNA2;
+        mmq_y  =  MMQ_Y_Q4_0_RDNA2;
+        nwarps = NWARPS_Q4_0_RDNA2;
+    } else if (compute_capability >= VER_GEN12) {
+        mmq_x  =  MMQ_X_Q4_0_RDNA1;
+        mmq_y  =  MMQ_Y_Q4_0_RDNA1;
+        nwarps = NWARPS_Q4_0_RDNA1;
+    } else if (compute_capability >= VER_GEN9) {
+        mmq_x  =  MMQ_X_Q4_0_AMPERE;
+        mmq_y  =  MMQ_Y_Q4_0_AMPERE;
+        nwarps = NWARPS_Q4_0_AMPERE;
+    } else if (compute_capability >= VER_4VEC) {
+        mmq_x  =  MMQ_X_Q4_0_PASCAL;
+        mmq_y  =  MMQ_Y_Q4_0_PASCAL;
+        nwarps = NWARPS_Q4_0_PASCAL;
+    } else {
+        GGML_ABORT("fatal error");
+    }
+
+    const int block_num_x = (nrows_x + mmq_y - 1) / mmq_y;
+    const int block_num_y = (ncols_y + mmq_x - 1) / mmq_x;
+    const sycl::range<3> block_nums(1, block_num_y, block_num_x);
+    const sycl::range<3> block_dims(1, nwarps, WARP_SIZE);
+
+    if (nrows_x % mmq_y == 0) {
+        const bool need_check = false;
+        /*
+        DPCT1049:20: The work-group size passed to the SYCL kernel may exceed
+        the limit. To get the device limit, query
+        info::device::max_work_group_size. Adjust the work-group size if needed.
+        */
+        {
+            dpct::has_capability_or_fail(stream->get_device(),
+                                         {sycl::aspect::fp16});
+
+            stream->submit([&](sycl::handler &cgh) {
+                sycl::local_accessor<int, 1> tile_x_qs_q4_0_acc_ct1(
+                    sycl::range<1>(mmq_y * (WARP_SIZE) + mmq_y), cgh);
+                sycl::local_accessor<float, 1> tile_x_d_q4_0_acc_ct1(
+                    sycl::range<1>(mmq_y * (WARP_SIZE / QI4_0) + mmq_y / QI4_0),
+                    cgh);
+                sycl::local_accessor<int, 1> tile_y_qs_acc_ct1(
+                    sycl::range<1>(mmq_x * WARP_SIZE), cgh);
+                sycl::local_accessor<sycl::half2, 1> tile_y_ds_acc_ct1(
+                    sycl::range<1>(mmq_x * WARP_SIZE / QI8_1), cgh);
+
+                cgh.parallel_for(
+                    sycl::nd_range<3>(block_nums * block_dims, block_dims),
+                    [=](sycl::nd_item<3> item_ct1) {
+                        mul_mat_q4_0<need_check>(
+                            vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y,
+                            nrows_dst, item_ct1,
+                            get_pointer(tile_x_qs_q4_0_acc_ct1),
+                            get_pointer(tile_x_d_q4_0_acc_ct1),
+                            get_pointer(tile_y_qs_acc_ct1),
+                            get_pointer(tile_y_ds_acc_ct1));
+                    });
+            });
+        }
+    } else {
+        const bool need_check = true;
+        /*
+        DPCT1049:21: The work-group size passed to the SYCL kernel may exceed
+        the limit. To get the device limit, query
+        info::device::max_work_group_size. Adjust the work-group size if needed.
+        */
+        {
+            dpct::has_capability_or_fail(stream->get_device(),
+                                         {sycl::aspect::fp16});
+
+            stream->submit([&](sycl::handler &cgh) {
+                sycl::local_accessor<int, 1> tile_x_qs_q4_0_acc_ct1(
+                    sycl::range<1>(mmq_y * (WARP_SIZE) + mmq_y), cgh);
+                sycl::local_accessor<float, 1> tile_x_d_q4_0_acc_ct1(
+                    sycl::range<1>(mmq_y * (WARP_SIZE / QI4_0) + mmq_y / QI4_0),
+                    cgh);
+                sycl::local_accessor<int, 1> tile_y_qs_acc_ct1(
+                    sycl::range<1>(mmq_x * WARP_SIZE), cgh);
+                sycl::local_accessor<sycl::half2, 1> tile_y_ds_acc_ct1(
+                    sycl::range<1>(mmq_x * WARP_SIZE / QI8_1), cgh);
+
+                cgh.parallel_for(
+                    sycl::nd_range<3>(block_nums * block_dims, block_dims),
+                    [=](sycl::nd_item<3> item_ct1) {
+                        mul_mat_q4_0<need_check>(
+                            vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y,
+                            nrows_dst, item_ct1,
+                            get_pointer(tile_x_qs_q4_0_acc_ct1),
+                            get_pointer(tile_x_d_q4_0_acc_ct1),
+                            get_pointer(tile_y_qs_acc_ct1),
+                            get_pointer(tile_y_ds_acc_ct1));
+                    });
+            });
+        }
+    }
+}
+catch (sycl::exception const &exc) {
+  std::cerr << exc.what() << "Exception caught at file:" << __FILE__
+            << ", line:" << __LINE__ << std::endl;
+  std::exit(1);
+}
+
+static void ggml_mul_mat_q4_1_q8_1_sycl(const void *vx, const void *vy,
+                                        float *dst, const int ncols_x,
+                                        const int nrows_x, const int ncols_y,
+                                        const int nrows_y, const int nrows_dst,
+                                        dpct::queue_ptr stream) try {
+
+    int id;
+    SYCL_CHECK(
+        CHECK_TRY_ERROR(id = get_current_device_id()));
+    const int compute_capability = ggml_sycl_info().devices[id].cc;
+
+    int mmq_x, mmq_y, nwarps;
+    if (compute_capability >= VER_GEN13) {
+        mmq_x  =  MMQ_X_Q4_1_RDNA2;
+        mmq_y  =  MMQ_Y_Q4_1_RDNA2;
+        nwarps = NWARPS_Q4_1_RDNA2;
+    } else if (compute_capability >= VER_GEN12) {
+        mmq_x  =  MMQ_X_Q4_1_RDNA1;
+        mmq_y  =  MMQ_Y_Q4_1_RDNA1;
+        nwarps = NWARPS_Q4_1_RDNA1;
+    } else if (compute_capability >= VER_GEN9) {
+        mmq_x  =  MMQ_X_Q4_1_AMPERE;
+        mmq_y  =  MMQ_Y_Q4_1_AMPERE;
+        nwarps = NWARPS_Q4_1_AMPERE;
+    } else if (compute_capability >= VER_4VEC) {
+        mmq_x  =  MMQ_X_Q4_1_PASCAL;
+        mmq_y  =  MMQ_Y_Q4_1_PASCAL;
+        nwarps = NWARPS_Q4_1_PASCAL;
+    } else {
+        GGML_ABORT("fatal error");
+    }
+
+    const int block_num_x = (nrows_x + mmq_y - 1) / mmq_y;
+    const int block_num_y = (ncols_y + mmq_x - 1) / mmq_x;
+    const sycl::range<3> block_nums(1, block_num_y, block_num_x);
+    const sycl::range<3> block_dims(1, nwarps, WARP_SIZE);
+
+    if (nrows_x % mmq_y == 0) {
+        const bool need_check = false;
+        /*
+        DPCT1049:22: The work-group size passed to the SYCL kernel may exceed
+        the limit. To get the device limit, query
+        info::device::max_work_group_size. Adjust the work-group size if needed.
+        */
+        {
+            dpct::has_capability_or_fail(stream->get_device(),
+                                         {sycl::aspect::fp16});
+
+            stream->submit([&](sycl::handler &cgh) {
+                sycl::local_accessor<int, 1> tile_x_qs_q4_1_acc_ct1(
+                    sycl::range<1>(mmq_y * (WARP_SIZE) + +mmq_y), cgh);
+                sycl::local_accessor<sycl::half2, 1> tile_x_dm_q4_1_acc_ct1(
+                    sycl::range<1>(mmq_y * (WARP_SIZE / QI4_1) + mmq_y / QI4_1),
+                    cgh);
+                sycl::local_accessor<int, 1> tile_y_qs_acc_ct1(
+                    sycl::range<1>(mmq_x * WARP_SIZE), cgh);
+                sycl::local_accessor<sycl::half2, 1> tile_y_ds_acc_ct1(
+                    sycl::range<1>(mmq_x * WARP_SIZE / QI8_1), cgh);
+
+                cgh.parallel_for(
+                    sycl::nd_range<3>(block_nums * block_dims, block_dims),
+                    [=](sycl::nd_item<3> item_ct1) {
+                        mul_mat_q4_1<need_check>(
+                            vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y,
+                            nrows_dst, item_ct1,
+                            get_pointer(tile_x_qs_q4_1_acc_ct1),
+                            get_pointer(tile_x_dm_q4_1_acc_ct1),
+                            get_pointer(tile_y_qs_acc_ct1),
+                            get_pointer(tile_y_ds_acc_ct1));
+                    });
+            });
+        }
+    } else {
+        const bool need_check = true;
+        /*
+        DPCT1049:23: The work-group size passed to the SYCL kernel may exceed
+        the limit. To get the device limit, query
+        info::device::max_work_group_size. Adjust the work-group size if needed.
+        */
+        {
+            dpct::has_capability_or_fail(stream->get_device(),
+                                         {sycl::aspect::fp16});
+
+            stream->submit([&](sycl::handler &cgh) {
+                sycl::local_accessor<int, 1> tile_x_qs_q4_1_acc_ct1(
+                    sycl::range<1>(mmq_y * (WARP_SIZE) + +mmq_y), cgh);
+                sycl::local_accessor<sycl::half2, 1> tile_x_dm_q4_1_acc_ct1(
+                    sycl::range<1>(mmq_y * (WARP_SIZE / QI4_1) + mmq_y / QI4_1),
+                    cgh);
+                sycl::local_accessor<int, 1> tile_y_qs_acc_ct1(
+                    sycl::range<1>(mmq_x * WARP_SIZE), cgh);
+                sycl::local_accessor<sycl::half2, 1> tile_y_ds_acc_ct1(
+                    sycl::range<1>(mmq_x * WARP_SIZE / QI8_1), cgh);
+
+                cgh.parallel_for(
+                    sycl::nd_range<3>(block_nums * block_dims, block_dims),
+                    [=](sycl::nd_item<3> item_ct1) {
+                        mul_mat_q4_1<need_check>(
+                            vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y,
+                            nrows_dst, item_ct1,
+                            get_pointer(tile_x_qs_q4_1_acc_ct1),
+                            get_pointer(tile_x_dm_q4_1_acc_ct1),
+                            get_pointer(tile_y_qs_acc_ct1),
+                            get_pointer(tile_y_ds_acc_ct1));
+                    });
+            });
+        }
+    }
+}
+catch (sycl::exception const &exc) {
+  std::cerr << exc.what() << "Exception caught at file:" << __FILE__
+            << ", line:" << __LINE__ << std::endl;
+  std::exit(1);
+}
+
+static void ggml_mul_mat_q5_0_q8_1_sycl(const void *vx, const void *vy,
+                                        float *dst, const int ncols_x,
+                                        const int nrows_x, const int ncols_y,
+                                        const int nrows_y, const int nrows_dst,
+                                        dpct::queue_ptr stream) try {
+
+    int id;
+    SYCL_CHECK(
+        CHECK_TRY_ERROR(id = get_current_device_id()));
+    const int compute_capability = ggml_sycl_info().devices[id].cc;
+
+    int mmq_x, mmq_y, nwarps;
+    if (compute_capability >= VER_GEN13) {
+        mmq_x  =  MMQ_X_Q5_0_RDNA2;
+        mmq_y  =  MMQ_Y_Q5_0_RDNA2;
+        nwarps = NWARPS_Q5_0_RDNA2;
+    } else if (compute_capability >= VER_GEN12) {
+        mmq_x  =  MMQ_X_Q5_0_RDNA1;
+        mmq_y  =  MMQ_Y_Q5_0_RDNA1;
+        nwarps = NWARPS_Q5_0_RDNA1;
+    } else if (compute_capability >= VER_GEN9) {
+        mmq_x  =  MMQ_X_Q5_0_AMPERE;
+        mmq_y  =  MMQ_Y_Q5_0_AMPERE;
+        nwarps = NWARPS_Q5_0_AMPERE;
+    } else if (compute_capability >= VER_4VEC) {
+        mmq_x  =  MMQ_X_Q5_0_PASCAL;
+        mmq_y  =  MMQ_Y_Q5_0_PASCAL;
+        nwarps = NWARPS_Q5_0_PASCAL;
+    } else {
+        GGML_ABORT("fatal error");
+    }
+
+    const int block_num_x = (nrows_x + mmq_y - 1) / mmq_y;
+    const int block_num_y = (ncols_y + mmq_x - 1) / mmq_x;
+    const sycl::range<3> block_nums(1, block_num_y, block_num_x);
+    const sycl::range<3> block_dims(1, nwarps, WARP_SIZE);
+
+    if (nrows_x % mmq_y == 0) {
+        const bool need_check = false;
+        /*
+        DPCT1049:24: The work-group size passed to the SYCL kernel may exceed
+        the limit. To get the device limit, query
+        info::device::max_work_group_size. Adjust the work-group size if needed.
+        */
+        {
+            dpct::has_capability_or_fail(stream->get_device(),
+                                         {sycl::aspect::fp16});
+
+            stream->submit([&](sycl::handler &cgh) {
+                sycl::local_accessor<int, 1> tile_x_ql_q5_0_acc_ct1(
+                    sycl::range<1>(mmq_y * (2 * WARP_SIZE) + mmq_y), cgh);
+                sycl::local_accessor<float, 1> tile_x_d_q5_0_acc_ct1(
+                    sycl::range<1>(mmq_y * (WARP_SIZE / QI5_0) + mmq_y / QI5_0),
+                    cgh);
+                sycl::local_accessor<int, 1> tile_y_qs_acc_ct1(
+                    sycl::range<1>(mmq_x * WARP_SIZE), cgh);
+                sycl::local_accessor<sycl::half2, 1> tile_y_ds_acc_ct1(
+                    sycl::range<1>(mmq_x * WARP_SIZE / QI8_1), cgh);
+
+                cgh.parallel_for(
+                    sycl::nd_range<3>(block_nums * block_dims, block_dims),
+                    [=](sycl::nd_item<3> item_ct1) {
+                        mul_mat_q5_0<need_check>(
+                            vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y,
+                            nrows_dst, item_ct1,
+                            get_pointer(tile_x_ql_q5_0_acc_ct1),
+                            get_pointer(tile_x_d_q5_0_acc_ct1),
+                            get_pointer(tile_y_qs_acc_ct1),
+                            get_pointer(tile_y_ds_acc_ct1));
+                    });
+            });
+        }
+    } else {
+        const bool need_check = true;
+        /*
+        DPCT1049:25: The work-group size passed to the SYCL kernel may exceed
+        the limit. To get the device limit, query
+        info::device::max_work_group_size. Adjust the work-group size if needed.
+        */
+        {
+            dpct::has_capability_or_fail(stream->get_device(),
+                                         {sycl::aspect::fp16});
+
+            stream->submit([&](sycl::handler &cgh) {
+                sycl::local_accessor<int, 1> tile_x_ql_q5_0_acc_ct1(
+                    sycl::range<1>(mmq_y * (2 * WARP_SIZE) + mmq_y), cgh);
+                sycl::local_accessor<float, 1> tile_x_d_q5_0_acc_ct1(
+                    sycl::range<1>(mmq_y * (WARP_SIZE / QI5_0) + mmq_y / QI5_0),
+                    cgh);
+                sycl::local_accessor<int, 1> tile_y_qs_acc_ct1(
+                    sycl::range<1>(mmq_x * WARP_SIZE), cgh);
+                sycl::local_accessor<sycl::half2, 1> tile_y_ds_acc_ct1(
+                    sycl::range<1>(mmq_x * WARP_SIZE / QI8_1), cgh);
+
+                cgh.parallel_for(
+                    sycl::nd_range<3>(block_nums * block_dims, block_dims),
+                    [=](sycl::nd_item<3> item_ct1) {
+                        mul_mat_q5_0<need_check>(
+                            vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y,
+                            nrows_dst, item_ct1,
+                            get_pointer(tile_x_ql_q5_0_acc_ct1),
+                            get_pointer(tile_x_d_q5_0_acc_ct1),
+                            get_pointer(tile_y_qs_acc_ct1),
+                            get_pointer(tile_y_ds_acc_ct1));
+                    });
+            });
+        }
+    }
+}
+catch (sycl::exception const &exc) {
+  std::cerr << exc.what() << "Exception caught at file:" << __FILE__
+            << ", line:" << __LINE__ << std::endl;
+  std::exit(1);
+}
+
+static void ggml_mul_mat_q5_1_q8_1_sycl(const void *vx, const void *vy,
+                                        float *dst, const int ncols_x,
+                                        const int nrows_x, const int ncols_y,
+                                        const int nrows_y, const int nrows_dst,
+                                        dpct::queue_ptr stream) try {
+
+    int id;
+    SYCL_CHECK(
+        CHECK_TRY_ERROR(id = get_current_device_id()));
+    const int compute_capability = ggml_sycl_info().devices[id].cc;
+
+    int mmq_x, mmq_y, nwarps;
+    if (compute_capability >= VER_GEN13) {
+        mmq_x  =  MMQ_X_Q5_1_RDNA2;
+        mmq_y  =  MMQ_Y_Q5_1_RDNA2;
+        nwarps = NWARPS_Q5_1_RDNA2;
+    } else if (compute_capability >= VER_GEN12) {
+        mmq_x  =  MMQ_X_Q5_1_RDNA1;
+        mmq_y  =  MMQ_Y_Q5_1_RDNA1;
+        nwarps = NWARPS_Q5_1_RDNA1;
+    } else if (compute_capability >= VER_GEN9) {
+        mmq_x  =  MMQ_X_Q5_1_AMPERE;
+        mmq_y  =  MMQ_Y_Q5_1_AMPERE;
+        nwarps = NWARPS_Q5_1_AMPERE;
+    } else if (compute_capability >= VER_4VEC) {
+        mmq_x  =  MMQ_X_Q5_1_PASCAL;
+        mmq_y  =  MMQ_Y_Q5_1_PASCAL;
+        nwarps = NWARPS_Q5_1_PASCAL;
+    } else {
+        GGML_ABORT("fatal error");
+    }
+
+    const int block_num_x = (nrows_x + mmq_y - 1) / mmq_y;
+    const int block_num_y = (ncols_y + mmq_x - 1) / mmq_x;
+    const sycl::range<3> block_nums(1, block_num_y, block_num_x);
+    const sycl::range<3> block_dims(1, nwarps, WARP_SIZE);
+
+    if (nrows_x % mmq_y == 0) {
+        const bool need_check = false;
+        /*
+        DPCT1049:26: The work-group size passed to the SYCL kernel may exceed
+        the limit. To get the device limit, query
+        info::device::max_work_group_size. Adjust the work-group size if needed.
+        */
+        {
+            dpct::has_capability_or_fail(stream->get_device(),
+                                         {sycl::aspect::fp16});
+
+            stream->submit([&](sycl::handler &cgh) {
+                sycl::local_accessor<int, 1> tile_x_ql_q5_1_acc_ct1(
+                    sycl::range<1>(mmq_y * (2 * WARP_SIZE) + mmq_y), cgh);
+                sycl::local_accessor<sycl::half2, 1> tile_x_dm_q5_1_acc_ct1(
+                    sycl::range<1>(mmq_y * (WARP_SIZE / QI5_1) + mmq_y / QI5_1),
+                    cgh);
+                sycl::local_accessor<int, 1> tile_y_qs_acc_ct1(
+                    sycl::range<1>(mmq_x * WARP_SIZE), cgh);
+                sycl::local_accessor<sycl::half2, 1> tile_y_ds_acc_ct1(
+                    sycl::range<1>(mmq_x * WARP_SIZE / QI8_1), cgh);
+
+                cgh.parallel_for(
+                    sycl::nd_range<3>(block_nums * block_dims, block_dims),
+                    [=](sycl::nd_item<3> item_ct1) {
+                        mul_mat_q5_1<need_check>(
+                            vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y,
+                            nrows_dst, item_ct1,
+                            get_pointer(tile_x_ql_q5_1_acc_ct1),
+                            get_pointer(tile_x_dm_q5_1_acc_ct1),
+                            get_pointer(tile_y_qs_acc_ct1),
+                            get_pointer(tile_y_ds_acc_ct1));
+                    });
+            });
+        }
+    } else {
+        const bool need_check = true;
+        /*
+        DPCT1049:27: The work-group size passed to the SYCL kernel may exceed
+        the limit. To get the device limit, query
+        info::device::max_work_group_size. Adjust the work-group size if needed.
+        */
+        {
+            dpct::has_capability_or_fail(stream->get_device(),
+                                         {sycl::aspect::fp16});
+
+            stream->submit([&](sycl::handler &cgh) {
+                sycl::local_accessor<int, 1> tile_x_ql_q5_1_acc_ct1(
+                    sycl::range<1>(mmq_y * (2 * WARP_SIZE) + mmq_y), cgh);
+                sycl::local_accessor<sycl::half2, 1> tile_x_dm_q5_1_acc_ct1(
+                    sycl::range<1>(mmq_y * (WARP_SIZE / QI5_1) + mmq_y / QI5_1),
+                    cgh);
+                sycl::local_accessor<int, 1> tile_y_qs_acc_ct1(
+                    sycl::range<1>(mmq_x * WARP_SIZE), cgh);
+                sycl::local_accessor<sycl::half2, 1> tile_y_ds_acc_ct1(
+                    sycl::range<1>(mmq_x * WARP_SIZE / QI8_1), cgh);
+
+                cgh.parallel_for(
+                    sycl::nd_range<3>(block_nums * block_dims, block_dims),
+                    [=](sycl::nd_item<3> item_ct1) {
+                        mul_mat_q5_1<need_check>(
+                            vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y,
+                            nrows_dst, item_ct1,
+                            get_pointer(tile_x_ql_q5_1_acc_ct1),
+                            get_pointer(tile_x_dm_q5_1_acc_ct1),
+                            get_pointer(tile_y_qs_acc_ct1),
+                            get_pointer(tile_y_ds_acc_ct1));
+                    });
+            });
+        }
+    }
+}
+catch (sycl::exception const &exc) {
+  std::cerr << exc.what() << "Exception caught at file:" << __FILE__
+            << ", line:" << __LINE__ << std::endl;
+  std::exit(1);
+}
+
+static void ggml_mul_mat_q8_0_q8_1_sycl(const void *vx, const void *vy,
+                                        float *dst, const int ncols_x,
+                                        const int nrows_x, const int ncols_y,
+                                        const int nrows_y, const int nrows_dst,
+                                        dpct::queue_ptr stream) try {
+
+    int id;
+    SYCL_CHECK(
+        CHECK_TRY_ERROR(id = get_current_device_id()));
+    const int compute_capability = ggml_sycl_info().devices[id].cc;
+
+    int mmq_x, mmq_y, nwarps;
+    if (compute_capability >= VER_GEN13) {
+        mmq_x  =  MMQ_X_Q8_0_RDNA2;
+        mmq_y  =  MMQ_Y_Q8_0_RDNA2;
+        nwarps = NWARPS_Q8_0_RDNA2;
+    } else if (compute_capability >= VER_GEN12) {
+        mmq_x  =  MMQ_X_Q8_0_RDNA1;
+        mmq_y  =  MMQ_Y_Q8_0_RDNA1;
+        nwarps = NWARPS_Q8_0_RDNA1;
+    } else if (compute_capability >= VER_GEN9) {
+        mmq_x  =  MMQ_X_Q8_0_AMPERE;
+        mmq_y  =  MMQ_Y_Q8_0_AMPERE;
+        nwarps = NWARPS_Q8_0_AMPERE;
+    } else if (compute_capability >= VER_4VEC) {
+        mmq_x  =  MMQ_X_Q8_0_PASCAL;
+        mmq_y  =  MMQ_Y_Q8_0_PASCAL;
+        nwarps = NWARPS_Q8_0_PASCAL;
+    } else {
+        GGML_ABORT("fatal error");
+    }
+
+    const int block_num_x = (nrows_x + mmq_y - 1) / mmq_y;
+    const int block_num_y = (ncols_y + mmq_x - 1) / mmq_x;
+    const sycl::range<3> block_nums(1, block_num_y, block_num_x);
+    const sycl::range<3> block_dims(1, nwarps, WARP_SIZE);
+
+    if (nrows_x % mmq_y == 0) {
+        const bool need_check = false;
+        /*
+        DPCT1049:28: The work-group size passed to the SYCL kernel may exceed
+        the limit. To get the device limit, query
+        info::device::max_work_group_size. Adjust the work-group size if needed.
+        */
+        {
+            dpct::has_capability_or_fail(stream->get_device(),
+                                         {sycl::aspect::fp16});
+
+            stream->submit([&](sycl::handler &cgh) {
+                sycl::local_accessor<int, 1> tile_x_qs_q8_0_acc_ct1(
+                    sycl::range<1>(mmq_y * (WARP_SIZE) + mmq_y), cgh);
+                sycl::local_accessor<float, 1> tile_x_d_q8_0_acc_ct1(
+                    sycl::range<1>(mmq_y * (WARP_SIZE / QI8_0) + mmq_y / QI8_0),
+                    cgh);
+                sycl::local_accessor<int, 1> tile_y_qs_acc_ct1(
+                    sycl::range<1>(mmq_x * WARP_SIZE), cgh);
+                sycl::local_accessor<sycl::half2, 1> tile_y_ds_acc_ct1(
+                    sycl::range<1>(mmq_x * WARP_SIZE / QI8_1), cgh);
+
+                cgh.parallel_for(
+                    sycl::nd_range<3>(block_nums * block_dims, block_dims),
+                    [=](sycl::nd_item<3> item_ct1) {
+                        mul_mat_q8_0<need_check>(
+                            vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y,
+                            nrows_dst, item_ct1,
+                            get_pointer(tile_x_qs_q8_0_acc_ct1),
+                            get_pointer(tile_x_d_q8_0_acc_ct1),
+                            get_pointer(tile_y_qs_acc_ct1),
+                            get_pointer(tile_y_ds_acc_ct1));
+                    });
+            });
+        }
+    } else {
+        const bool need_check = true;
+        /*
+        DPCT1049:29: The work-group size passed to the SYCL kernel may exceed
+        the limit. To get the device limit, query
+        info::device::max_work_group_size. Adjust the work-group size if needed.
+        */
+        {
+            dpct::has_capability_or_fail(stream->get_device(),
+                                         {sycl::aspect::fp16});
+
+            stream->submit([&](sycl::handler &cgh) {
+                sycl::local_accessor<int, 1> tile_x_qs_q8_0_acc_ct1(
+                    sycl::range<1>(mmq_y * (WARP_SIZE) + mmq_y), cgh);
+                sycl::local_accessor<float, 1> tile_x_d_q8_0_acc_ct1(
+                    sycl::range<1>(mmq_y * (WARP_SIZE / QI8_0) + mmq_y / QI8_0),
+                    cgh);
+                sycl::local_accessor<int, 1> tile_y_qs_acc_ct1(
+                    sycl::range<1>(mmq_x * WARP_SIZE), cgh);
+                sycl::local_accessor<sycl::half2, 1> tile_y_ds_acc_ct1(
+                    sycl::range<1>(mmq_x * WARP_SIZE / QI8_1), cgh);
+
+                cgh.parallel_for(
+                    sycl::nd_range<3>(block_nums * block_dims, block_dims),
+                    [=](sycl::nd_item<3> item_ct1) {
+                        mul_mat_q8_0<need_check>(
+                            vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y,
+                            nrows_dst, item_ct1,
+                            get_pointer(tile_x_qs_q8_0_acc_ct1),
+                            get_pointer(tile_x_d_q8_0_acc_ct1),
+                            get_pointer(tile_y_qs_acc_ct1),
+                            get_pointer(tile_y_ds_acc_ct1));
+                    });
+            });
+        }
+    }
+}
+catch (sycl::exception const &exc) {
+  std::cerr << exc.what() << "Exception caught at file:" << __FILE__
+            << ", line:" << __LINE__ << std::endl;
+  std::exit(1);
+}
+
+static void ggml_mul_mat_q2_K_q8_1_sycl(const void *vx, const void *vy,
+                                        float *dst, const int ncols_x,
+                                        const int nrows_x, const int ncols_y,
+                                        const int nrows_y, const int nrows_dst,
+                                        dpct::queue_ptr stream) try {
+
+    int id;
+    SYCL_CHECK(
+        CHECK_TRY_ERROR(id = get_current_device_id()));
+    const int compute_capability = ggml_sycl_info().devices[id].cc;
+
+    int mmq_x, mmq_y, nwarps;
+    if (compute_capability >= VER_GEN13) {
+        mmq_x  =  MMQ_X_Q2_K_RDNA2;
+        mmq_y  =  MMQ_Y_Q2_K_RDNA2;
+        nwarps = NWARPS_Q2_K_RDNA2;
+    } else if (compute_capability >= VER_GEN12) {
+        mmq_x  =  MMQ_X_Q2_K_RDNA1;
+        mmq_y  =  MMQ_Y_Q2_K_RDNA1;
+        nwarps = NWARPS_Q2_K_RDNA1;
+    } else if (compute_capability >= VER_GEN9) {
+        mmq_x  =  MMQ_X_Q2_K_AMPERE;
+        mmq_y  =  MMQ_Y_Q2_K_AMPERE;
+        nwarps = NWARPS_Q2_K_AMPERE;
+    } else if (compute_capability >= VER_4VEC) {
+        mmq_x  =  MMQ_X_Q2_K_PASCAL;
+        mmq_y  =  MMQ_Y_Q2_K_PASCAL;
+        nwarps = NWARPS_Q2_K_PASCAL;
+    } else {
+        GGML_ABORT("fatal error");
+    }
+
+    const int block_num_x = (nrows_x + mmq_y - 1) / mmq_y;
+    const int block_num_y = (ncols_y + mmq_x - 1) / mmq_x;
+    const sycl::range<3> block_nums(1, block_num_y, block_num_x);
+    const sycl::range<3> block_dims(1, nwarps, WARP_SIZE);
+
+    if (nrows_x % mmq_y == 0) {
+        const bool need_check = false;
+        /*
+        DPCT1049:30: The work-group size passed to the SYCL kernel may exceed
+        the limit. To get the device limit, query
+        info::device::max_work_group_size. Adjust the work-group size if needed.
+        */
+        {
+            dpct::has_capability_or_fail(stream->get_device(),
+                                         {sycl::aspect::fp16});
+
+            stream->submit([&](sycl::handler &cgh) {
+                sycl::local_accessor<int, 1> tile_x_ql_q2_K_acc_ct1(
+                    sycl::range<1>(mmq_y * (WARP_SIZE) + mmq_y), cgh);
+                sycl::local_accessor<sycl::half2, 1> tile_x_dm_q2_K_acc_ct1(
+                    sycl::range<1>(mmq_y * (WARP_SIZE / QI2_K) + mmq_y / QI2_K),
+                    cgh);
+                sycl::local_accessor<int, 1> tile_x_sc_q2_K_acc_ct1(
+                    sycl::range<1>(mmq_y * (WARP_SIZE / 4) + mmq_y / 4), cgh);
+                sycl::local_accessor<int, 1> tile_y_qs_acc_ct1(
+                    sycl::range<1>(mmq_x * WARP_SIZE), cgh);
+                sycl::local_accessor<sycl::half2, 1> tile_y_ds_acc_ct1(
+                    sycl::range<1>(mmq_x * WARP_SIZE / QI8_1), cgh);
+
+                cgh.parallel_for(
+                    sycl::nd_range<3>(block_nums * block_dims, block_dims),
+                    [=](sycl::nd_item<3> item_ct1) {
+                        mul_mat_q2_K<need_check>(
+                            vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y,
+                            nrows_dst, item_ct1,
+                            get_pointer(tile_x_ql_q2_K_acc_ct1),
+                            get_pointer(tile_x_dm_q2_K_acc_ct1),
+                            get_pointer(tile_x_sc_q2_K_acc_ct1),
+                            get_pointer(tile_y_qs_acc_ct1),
+                            get_pointer(tile_y_ds_acc_ct1));
+                    });
+            });
+        }
+    } else {
+        const bool need_check = true;
+        /*
+        DPCT1049:31: The work-group size passed to the SYCL kernel may exceed
+        the limit. To get the device limit, query
+        info::device::max_work_group_size. Adjust the work-group size if needed.
+        */
+        {
+            dpct::has_capability_or_fail(stream->get_device(),
+                                         {sycl::aspect::fp16});
+
+            stream->submit([&](sycl::handler &cgh) {
+                sycl::local_accessor<int, 1> tile_x_ql_q2_K_acc_ct1(
+                    sycl::range<1>(mmq_y * (WARP_SIZE) + mmq_y), cgh);
+                sycl::local_accessor<sycl::half2, 1> tile_x_dm_q2_K_acc_ct1(
+                    sycl::range<1>(mmq_y * (WARP_SIZE / QI2_K) + mmq_y / QI2_K),
+                    cgh);
+                sycl::local_accessor<int, 1> tile_x_sc_q2_K_acc_ct1(
+                    sycl::range<1>(mmq_y * (WARP_SIZE / 4) + mmq_y / 4), cgh);
+                sycl::local_accessor<int, 1> tile_y_qs_acc_ct1(
+                    sycl::range<1>(mmq_x * WARP_SIZE), cgh);
+                sycl::local_accessor<sycl::half2, 1> tile_y_ds_acc_ct1(
+                    sycl::range<1>(mmq_x * WARP_SIZE / QI8_1), cgh);
+
+                cgh.parallel_for(
+                    sycl::nd_range<3>(block_nums * block_dims, block_dims),
+                    [=](sycl::nd_item<3> item_ct1) {
+                        mul_mat_q2_K<need_check>(
+                            vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y,
+                            nrows_dst, item_ct1,
+                            get_pointer(tile_x_ql_q2_K_acc_ct1),
+                            get_pointer(tile_x_dm_q2_K_acc_ct1),
+                            get_pointer(tile_x_sc_q2_K_acc_ct1),
+                            get_pointer(tile_y_qs_acc_ct1),
+                            get_pointer(tile_y_ds_acc_ct1));
+                    });
+            });
+        }
+    }
+}
+catch (sycl::exception const &exc) {
+  std::cerr << exc.what() << "Exception caught at file:" << __FILE__
+            << ", line:" << __LINE__ << std::endl;
+  std::exit(1);
+}
+
+static void ggml_mul_mat_q3_K_q8_1_sycl(const void *vx, const void *vy,
+                                        float *dst, const int ncols_x,
+                                        const int nrows_x, const int ncols_y,
+                                        const int nrows_y, const int nrows_dst,
+                                        dpct::queue_ptr stream) try {
+
+#if QK_K == 256
+
+    int id;
+    SYCL_CHECK(
+        CHECK_TRY_ERROR(id = get_current_device_id()));
+    const int compute_capability = ggml_sycl_info().devices[id].cc;
+
+    int mmq_x, mmq_y, nwarps;
+    if (compute_capability >= VER_GEN13) {
+        mmq_x  =  MMQ_X_Q3_K_RDNA2;
+        mmq_y  =  MMQ_Y_Q3_K_RDNA2;
+        nwarps = NWARPS_Q3_K_RDNA2;
+    } else if (compute_capability >= VER_GEN12) {
+        mmq_x  =  MMQ_X_Q3_K_RDNA1;
+        mmq_y  =  MMQ_Y_Q3_K_RDNA1;
+        nwarps = NWARPS_Q3_K_RDNA1;
+    } else if (compute_capability >= VER_GEN9) {
+        mmq_x  =  MMQ_X_Q3_K_AMPERE;
+        mmq_y  =  MMQ_Y_Q3_K_AMPERE;
+        nwarps = NWARPS_Q3_K_AMPERE;
+    } else if (compute_capability >= VER_4VEC) {
+        mmq_x  =  MMQ_X_Q3_K_PASCAL;
+        mmq_y  =  MMQ_Y_Q3_K_PASCAL;
+        nwarps = NWARPS_Q3_K_PASCAL;
+    } else {
+        GGML_ABORT("fatal error");
+    }
+
+    const int block_num_x = (nrows_x + mmq_y - 1) / mmq_y;
+    const int block_num_y = (ncols_y + mmq_x - 1) / mmq_x;
+    const sycl::range<3> block_nums(1, block_num_y, block_num_x);
+    const sycl::range<3> block_dims(1, nwarps, WARP_SIZE);
+
+    if (nrows_x % mmq_y == 0) {
+        const bool need_check = false;
+        /*
+        DPCT1049:32: The work-group size passed to the SYCL kernel may exceed
+        the limit. To get the device limit, query
+        info::device::max_work_group_size. Adjust the work-group size if needed.
+        */
+        {
+            dpct::has_capability_or_fail(stream->get_device(),
+                                         {sycl::aspect::fp16});
+
+            stream->submit([&](sycl::handler &cgh) {
+                sycl::local_accessor<int, 1> tile_x_ql_q3_K_acc_ct1(
+                    sycl::range<1>(mmq_y * (WARP_SIZE) + mmq_y), cgh);
+                sycl::local_accessor<sycl::half2, 1> tile_x_dm_q3_K_acc_ct1(
+                    sycl::range<1>(mmq_y * (WARP_SIZE / QI3_K) + mmq_y / QI3_K),
+                    cgh);
+                sycl::local_accessor<int, 1> tile_x_qh_q3_K_acc_ct1(
+                    sycl::range<1>(mmq_y * (WARP_SIZE / 2) + mmq_y / 2), cgh);
+                sycl::local_accessor<int, 1> tile_x_sc_q3_K_acc_ct1(
+                    sycl::range<1>(mmq_y * (WARP_SIZE / 4) + mmq_y / 4), cgh);
+                sycl::local_accessor<int, 1> tile_y_qs_acc_ct1(
+                    sycl::range<1>(mmq_x * WARP_SIZE), cgh);
+                sycl::local_accessor<sycl::half2, 1> tile_y_ds_acc_ct1(
+                    sycl::range<1>(mmq_x * WARP_SIZE / QI8_1), cgh);
+
+                cgh.parallel_for(
+                    sycl::nd_range<3>(block_nums * block_dims, block_dims),
+                    [=](sycl::nd_item<3> item_ct1) {
+                        mul_mat_q3_K<need_check>(
+                            vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y,
+                            nrows_dst, item_ct1,
+                            get_pointer(tile_x_ql_q3_K_acc_ct1),
+                            get_pointer(tile_x_dm_q3_K_acc_ct1),
+                            get_pointer(tile_x_qh_q3_K_acc_ct1),
+                            get_pointer(tile_x_sc_q3_K_acc_ct1),
+                            get_pointer(tile_y_qs_acc_ct1),
+                            get_pointer(tile_y_ds_acc_ct1));
+                    });
+            });
+        }
+    } else {
+        const bool need_check = true;
+        /*
+        DPCT1049:33: The work-group size passed to the SYCL kernel may exceed
+        the limit. To get the device limit, query
+        info::device::max_work_group_size. Adjust the work-group size if needed.
+        */
+        {
+            dpct::has_capability_or_fail(stream->get_device(),
+                                         {sycl::aspect::fp16});
+
+            stream->submit([&](sycl::handler &cgh) {
+                sycl::local_accessor<int, 1> tile_x_ql_q3_K_acc_ct1(
+                    sycl::range<1>(mmq_y * (WARP_SIZE) + mmq_y), cgh);
+                sycl::local_accessor<sycl::half2, 1> tile_x_dm_q3_K_acc_ct1(
+                    sycl::range<1>(mmq_y * (WARP_SIZE / QI3_K) + mmq_y / QI3_K),
+                    cgh);
+                sycl::local_accessor<int, 1> tile_x_qh_q3_K_acc_ct1(
+                    sycl::range<1>(mmq_y * (WARP_SIZE / 2) + mmq_y / 2), cgh);
+                sycl::local_accessor<int, 1> tile_x_sc_q3_K_acc_ct1(
+                    sycl::range<1>(mmq_y * (WARP_SIZE / 4) + mmq_y / 4), cgh);
+                sycl::local_accessor<int, 1> tile_y_qs_acc_ct1(
+                    sycl::range<1>(mmq_x * WARP_SIZE), cgh);
+                sycl::local_accessor<sycl::half2, 1> tile_y_ds_acc_ct1(
+                    sycl::range<1>(mmq_x * WARP_SIZE / QI8_1), cgh);
+
+                cgh.parallel_for(
+                    sycl::nd_range<3>(block_nums * block_dims, block_dims),
+                    [=](sycl::nd_item<3> item_ct1) {
+                        mul_mat_q3_K<need_check>(
+                            vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y,
+                            nrows_dst, item_ct1,
+                            get_pointer(tile_x_ql_q3_K_acc_ct1),
+                            get_pointer(tile_x_dm_q3_K_acc_ct1),
+                            get_pointer(tile_x_qh_q3_K_acc_ct1),
+                            get_pointer(tile_x_sc_q3_K_acc_ct1),
+                            get_pointer(tile_y_qs_acc_ct1),
+                            get_pointer(tile_y_ds_acc_ct1));
+                    });
+            });
+        }
+    }
+#endif
+}
+catch (sycl::exception const &exc) {
+  std::cerr << exc.what() << "Exception caught at file:" << __FILE__
+            << ", line:" << __LINE__ << std::endl;
+  std::exit(1);
+}
+
+static void ggml_mul_mat_q4_K_q8_1_sycl(const void *vx, const void *vy,
+                                        float *dst, const int ncols_x,
+                                        const int nrows_x, const int ncols_y,
+                                        const int nrows_y, const int nrows_dst,
+                                        dpct::queue_ptr stream) try {
+
+    int id;
+    SYCL_CHECK(
+        CHECK_TRY_ERROR(id = get_current_device_id()));
+    const int compute_capability = ggml_sycl_info().devices[id].cc;
+
+    int mmq_x, mmq_y, nwarps;
+    if (compute_capability >= VER_GEN13) {
+        mmq_x  =  MMQ_X_Q4_K_RDNA2;
+        mmq_y  =  MMQ_Y_Q4_K_RDNA2;
+        nwarps = NWARPS_Q4_K_RDNA2;
+    } else if (compute_capability >= VER_GEN12) {
+        mmq_x  =  MMQ_X_Q4_K_RDNA1;
+        mmq_y  =  MMQ_Y_Q4_K_RDNA1;
+        nwarps = NWARPS_Q4_K_RDNA1;
+    } else if (compute_capability >= VER_GEN9) {
+        mmq_x  =  MMQ_X_Q4_K_AMPERE;
+        mmq_y  =  MMQ_Y_Q4_K_AMPERE;
+        nwarps = NWARPS_Q4_K_AMPERE;
+    } else if (compute_capability >= VER_4VEC) {
+        mmq_x  =  MMQ_X_Q4_K_PASCAL;
+        mmq_y  =  MMQ_Y_Q4_K_PASCAL;
+        nwarps = NWARPS_Q4_K_PASCAL;
+    } else {
+        GGML_ABORT("fatal error");
+    }
+
+    const int block_num_x = (nrows_x + mmq_y - 1) / mmq_y;
+    const int block_num_y = (ncols_y + mmq_x - 1) / mmq_x;
+    const sycl::range<3> block_nums(1, block_num_y, block_num_x);
+    const sycl::range<3> block_dims(1, nwarps, WARP_SIZE);
+
+    if (nrows_x % mmq_y == 0) {
+        const bool need_check = false;
+        /*
+        DPCT1049:34: The work-group size passed to the SYCL kernel may exceed
+        the limit. To get the device limit, query
+        info::device::max_work_group_size. Adjust the work-group size if needed.
+        */
+        {
+            dpct::has_capability_or_fail(stream->get_device(),
+                                         {sycl::aspect::fp16});
+
+            stream->submit([&](sycl::handler &cgh) {
+                sycl::local_accessor<int, 1> tile_x_ql_q4_K_acc_ct1(
+                    sycl::range<1>(mmq_y * (WARP_SIZE) + mmq_y), cgh);
+                sycl::local_accessor<sycl::half2, 1> tile_x_dm_q4_K_acc_ct1(
+                    sycl::range<1>(mmq_y * (WARP_SIZE / QI4_K) + mmq_y / QI4_K),
+                    cgh);
+                sycl::local_accessor<int, 1> tile_x_sc_q4_K_acc_ct1(
+                    sycl::range<1>(mmq_y * (WARP_SIZE / 8) + mmq_y / 8), cgh);
+                sycl::local_accessor<int, 1> tile_y_qs_acc_ct1(
+                    sycl::range<1>(mmq_x * WARP_SIZE), cgh);
+                sycl::local_accessor<sycl::half2, 1> tile_y_ds_acc_ct1(
+                    sycl::range<1>(mmq_x * WARP_SIZE / QI8_1), cgh);
+
+                cgh.parallel_for(
+                    sycl::nd_range<3>(block_nums * block_dims, block_dims),
+                    [=](sycl::nd_item<3> item_ct1) {
+                        mul_mat_q4_K<need_check>(
+                            vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y,
+                            nrows_dst, item_ct1,
+                            get_pointer(tile_x_ql_q4_K_acc_ct1),
+                            get_pointer(tile_x_dm_q4_K_acc_ct1),
+                            get_pointer(tile_x_sc_q4_K_acc_ct1),
+                            get_pointer(tile_y_qs_acc_ct1),
+                            get_pointer(tile_y_ds_acc_ct1));
+                    });
+            });
+        }
+    } else {
+        const bool need_check = true;
+        /*
+        DPCT1049:35: The work-group size passed to the SYCL kernel may exceed
+        the limit. To get the device limit, query
+        info::device::max_work_group_size. Adjust the work-group size if needed.
+        */
+        {
+            dpct::has_capability_or_fail(stream->get_device(),
+                                         {sycl::aspect::fp16});
+
+            stream->submit([&](sycl::handler &cgh) {
+                sycl::local_accessor<int, 1> tile_x_ql_q4_K_acc_ct1(
+                    sycl::range<1>(mmq_y * (WARP_SIZE) + mmq_y), cgh);
+                sycl::local_accessor<sycl::half2, 1> tile_x_dm_q4_K_acc_ct1(
+                    sycl::range<1>(mmq_y * (WARP_SIZE / QI4_K) + mmq_y / QI4_K),
+                    cgh);
+                sycl::local_accessor<int, 1> tile_x_sc_q4_K_acc_ct1(
+                    sycl::range<1>(mmq_y * (WARP_SIZE / 8) + mmq_y / 8), cgh);
+                sycl::local_accessor<int, 1> tile_y_qs_acc_ct1(
+                    sycl::range<1>(mmq_x * WARP_SIZE), cgh);
+                sycl::local_accessor<sycl::half2, 1> tile_y_ds_acc_ct1(
+                    sycl::range<1>(mmq_x * WARP_SIZE / QI8_1), cgh);
+
+                cgh.parallel_for(
+                    sycl::nd_range<3>(block_nums * block_dims, block_dims),
+                    [=](sycl::nd_item<3> item_ct1) {
+                        mul_mat_q4_K<need_check>(
+                            vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y,
+                            nrows_dst, item_ct1,
+                            get_pointer(tile_x_ql_q4_K_acc_ct1),
+                            get_pointer(tile_x_dm_q4_K_acc_ct1),
+                            get_pointer(tile_x_sc_q4_K_acc_ct1),
+                            get_pointer(tile_y_qs_acc_ct1),
+                            get_pointer(tile_y_ds_acc_ct1));
+                    });
+            });
+        }
+    }
+}
+catch (sycl::exception const &exc) {
+  std::cerr << exc.what() << "Exception caught at file:" << __FILE__
+            << ", line:" << __LINE__ << std::endl;
+  std::exit(1);
+}
+
+static void ggml_mul_mat_q5_K_q8_1_sycl(const void *vx, const void *vy,
+                                        float *dst, const int ncols_x,
+                                        const int nrows_x, const int ncols_y,
+                                        const int nrows_y, const int nrows_dst,
+                                        dpct::queue_ptr stream) try {
+
+    int id;
+    SYCL_CHECK(
+        CHECK_TRY_ERROR(id = get_current_device_id()));
+    const int compute_capability = ggml_sycl_info().devices[id].cc;
+
+    int mmq_x, mmq_y, nwarps;
+    if (compute_capability >= VER_GEN13) {
+        mmq_x  =  MMQ_X_Q5_K_RDNA2;
+        mmq_y  =  MMQ_Y_Q5_K_RDNA2;
+        nwarps = NWARPS_Q5_K_RDNA2;
+    } else if (compute_capability >= VER_GEN12) {
+        mmq_x  =  MMQ_X_Q5_K_RDNA1;
+        mmq_y  =  MMQ_Y_Q5_K_RDNA1;
+        nwarps = NWARPS_Q5_K_RDNA1;
+    } else if (compute_capability >= VER_GEN9) {
+        mmq_x  =  MMQ_X_Q5_K_AMPERE;
+        mmq_y  =  MMQ_Y_Q5_K_AMPERE;
+        nwarps = NWARPS_Q5_K_AMPERE;
+    } else if (compute_capability >= VER_4VEC) {
+        mmq_x  =  MMQ_X_Q5_K_PASCAL;
+        mmq_y  =  MMQ_Y_Q5_K_PASCAL;
+        nwarps = NWARPS_Q5_K_PASCAL;
+    } else {
+        GGML_ABORT("fatal error");
+    }
+
+    const int block_num_x = (nrows_x + mmq_y - 1) / mmq_y;
+    const int block_num_y = (ncols_y + mmq_x - 1) / mmq_x;
+    const sycl::range<3> block_nums(1, block_num_y, block_num_x);
+    const sycl::range<3> block_dims(1, nwarps, WARP_SIZE);
+
+    if (nrows_x % mmq_y == 0) {
+        const bool need_check = false;
+        /*
+        DPCT1049:36: The work-group size passed to the SYCL kernel may exceed
+        the limit. To get the device limit, query
+        info::device::max_work_group_size. Adjust the work-group size if needed.
+        */
+        {
+            dpct::has_capability_or_fail(stream->get_device(),
+                                         {sycl::aspect::fp16});
+
+            stream->submit([&](sycl::handler &cgh) {
+                sycl::local_accessor<int, 1> tile_x_ql_q5_K_acc_ct1(
+                    sycl::range<1>(mmq_y * (2 * WARP_SIZE) + mmq_y), cgh);
+                sycl::local_accessor<sycl::half2, 1> tile_x_dm_q5_K_acc_ct1(
+                    sycl::range<1>(mmq_y * (WARP_SIZE / QI5_K) + mmq_y / QI5_K),
+                    cgh);
+                sycl::local_accessor<int, 1> tile_x_sc_q5_K_acc_ct1(
+                    sycl::range<1>(mmq_y * (WARP_SIZE / 8) + mmq_y / 8), cgh);
+                sycl::local_accessor<int, 1> tile_y_qs_acc_ct1(
+                    sycl::range<1>(mmq_x * WARP_SIZE), cgh);
+                sycl::local_accessor<sycl::half2, 1> tile_y_ds_acc_ct1(
+                    sycl::range<1>(mmq_x * WARP_SIZE / QI8_1), cgh);
+
+                cgh.parallel_for(
+                    sycl::nd_range<3>(block_nums * block_dims, block_dims),
+                    [=](sycl::nd_item<3> item_ct1) {
+                        mul_mat_q5_K<need_check>(
+                            vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y,
+                            nrows_dst, item_ct1,
+                            get_pointer(tile_x_ql_q5_K_acc_ct1),
+                            get_pointer(tile_x_dm_q5_K_acc_ct1),
+                            get_pointer(tile_x_sc_q5_K_acc_ct1),
+                            get_pointer(tile_y_qs_acc_ct1),
+                            get_pointer(tile_y_ds_acc_ct1));
+                    });
+            });
+        }
+    } else {
+        const bool need_check = true;
+        /*
+        DPCT1049:37: The work-group size passed to the SYCL kernel may exceed
+        the limit. To get the device limit, query
+        info::device::max_work_group_size. Adjust the work-group size if needed.
+        */
+        {
+            dpct::has_capability_or_fail(stream->get_device(),
+                                         {sycl::aspect::fp16});
+
+            stream->submit([&](sycl::handler &cgh) {
+                sycl::local_accessor<int, 1> tile_x_ql_q5_K_acc_ct1(
+                    sycl::range<1>(mmq_y * (2 * WARP_SIZE) + mmq_y), cgh);
+                sycl::local_accessor<sycl::half2, 1> tile_x_dm_q5_K_acc_ct1(
+                    sycl::range<1>(mmq_y * (WARP_SIZE / QI5_K) + mmq_y / QI5_K),
+                    cgh);
+                sycl::local_accessor<int, 1> tile_x_sc_q5_K_acc_ct1(
+                    sycl::range<1>(mmq_y * (WARP_SIZE / 8) + mmq_y / 8), cgh);
+                sycl::local_accessor<int, 1> tile_y_qs_acc_ct1(
+                    sycl::range<1>(mmq_x * WARP_SIZE), cgh);
+                sycl::local_accessor<sycl::half2, 1> tile_y_ds_acc_ct1(
+                    sycl::range<1>(mmq_x * WARP_SIZE / QI8_1), cgh);
+
+                cgh.parallel_for(
+                    sycl::nd_range<3>(block_nums * block_dims, block_dims),
+                    [=](sycl::nd_item<3> item_ct1) {
+                        mul_mat_q5_K<need_check>(
+                            vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y,
+                            nrows_dst, item_ct1,
+                            get_pointer(tile_x_ql_q5_K_acc_ct1),
+                            get_pointer(tile_x_dm_q5_K_acc_ct1),
+                            get_pointer(tile_x_sc_q5_K_acc_ct1),
+                            get_pointer(tile_y_qs_acc_ct1),
+                            get_pointer(tile_y_ds_acc_ct1));
+                    });
+            });
+        }
+    }
+}
+catch (sycl::exception const &exc) {
+  std::cerr << exc.what() << "Exception caught at file:" << __FILE__
+            << ", line:" << __LINE__ << std::endl;
+  std::exit(1);
+}
+
+static void ggml_mul_mat_q6_K_q8_1_sycl(const void *vx, const void *vy,
+                                        float *dst, const int ncols_x,
+                                        const int nrows_x, const int ncols_y,
+                                        const int nrows_y, const int nrows_dst,
+                                        dpct::queue_ptr stream) try {
+
+    int id;
+    SYCL_CHECK(
+        CHECK_TRY_ERROR(id = get_current_device_id()));
+    const int compute_capability = ggml_sycl_info().devices[id].cc;
+
+    int mmq_x, mmq_y, nwarps;
+    if (compute_capability >= VER_GEN13) {
+        mmq_x  =  MMQ_X_Q6_K_RDNA2;
+        mmq_y  =  MMQ_Y_Q6_K_RDNA2;
+        nwarps = NWARPS_Q6_K_RDNA2;
+    } else if (compute_capability >= VER_GEN12) {
+        mmq_x  =  MMQ_X_Q6_K_RDNA1;
+        mmq_y  =  MMQ_Y_Q6_K_RDNA1;
+        nwarps = NWARPS_Q6_K_RDNA1;
+    } else if (compute_capability >= VER_GEN9) {
+        mmq_x  =  MMQ_X_Q6_K_AMPERE;
+        mmq_y  =  MMQ_Y_Q6_K_AMPERE;
+        nwarps = NWARPS_Q6_K_AMPERE;
+    } else if (compute_capability >= VER_4VEC) {
+        mmq_x  =  MMQ_X_Q6_K_PASCAL;
+        mmq_y  =  MMQ_Y_Q6_K_PASCAL;
+        nwarps = NWARPS_Q6_K_PASCAL;
+    } else {
+        GGML_ABORT("fatal error");
+    }
+
+    const int block_num_x = (nrows_x + mmq_y - 1) / mmq_y;
+    const int block_num_y = (ncols_y + mmq_x - 1) / mmq_x;
+    const sycl::range<3> block_nums(1, block_num_y, block_num_x);
+    const sycl::range<3> block_dims(1, nwarps, WARP_SIZE);
+
+    if (nrows_x % mmq_y == 0) {
+        const bool need_check = false;
+        /*
+        DPCT1049:38: The work-group size passed to the SYCL kernel may exceed
+        the limit. To get the device limit, query
+        info::device::max_work_group_size. Adjust the work-group size if needed.
+        */
+        {
+            dpct::has_capability_or_fail(stream->get_device(),
+                                         {sycl::aspect::fp16});
+
+            stream->submit([&](sycl::handler &cgh) {
+                sycl::local_accessor<int, 1> tile_x_ql_acc_ct1(
+                    sycl::range<1>(mmq_y * (2 * WARP_SIZE) + mmq_y), cgh);
+                sycl::local_accessor<sycl::half2, 1> tile_x_dm_acc_ct1(
+                    sycl::range<1>(mmq_y * (WARP_SIZE / QI6_K) + mmq_y / QI6_K),
+                    cgh);
+                sycl::local_accessor<int, 1> tile_x_sc_acc_ct1(
+                    sycl::range<1>(mmq_y * (WARP_SIZE / 8) + mmq_y / 8), cgh);
+                sycl::local_accessor<int, 1> tile_y_qs_acc_ct1(
+                    sycl::range<1>(mmq_x * WARP_SIZE), cgh);
+                sycl::local_accessor<sycl::half2, 1> tile_y_ds_acc_ct1(
+                    sycl::range<1>(mmq_x * WARP_SIZE / QI8_1), cgh);
+
+                cgh.parallel_for(
+                    sycl::nd_range<3>(block_nums * block_dims, block_dims),
+                    [=](sycl::nd_item<3> item_ct1) {
+                        mul_mat_q6_K<need_check>(
+                            vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y,
+                            nrows_dst, item_ct1,
+                            get_pointer(tile_x_ql_acc_ct1),
+                            get_pointer(tile_x_dm_acc_ct1),
+                            get_pointer(tile_x_sc_acc_ct1),
+                            get_pointer(tile_y_qs_acc_ct1),
+                            get_pointer(tile_y_ds_acc_ct1));
+                    });
+            });
+        }
+    } else {
+        const bool need_check = true;
+        /*
+        DPCT1049:39: The work-group size passed to the SYCL kernel may exceed
+        the limit. To get the device limit, query
+        info::device::max_work_group_size. Adjust the work-group size if needed.
+        */
+        {
+            dpct::has_capability_or_fail(stream->get_device(),
+                                         {sycl::aspect::fp16});
+
+            stream->submit([&](sycl::handler &cgh) {
+                sycl::local_accessor<int, 1> tile_x_ql_acc_ct1(
+                    sycl::range<1>(mmq_y * (2 * WARP_SIZE) + mmq_y), cgh);
+                sycl::local_accessor<sycl::half2, 1> tile_x_dm_acc_ct1(
+                    sycl::range<1>(mmq_y * (WARP_SIZE / QI6_K) + mmq_y / QI6_K),
+                    cgh);
+                sycl::local_accessor<int, 1> tile_x_sc_acc_ct1(
+                    sycl::range<1>(mmq_y * (WARP_SIZE / 8) + mmq_y / 8), cgh);
+                sycl::local_accessor<int, 1> tile_y_qs_acc_ct1(
+                    sycl::range<1>(mmq_x * WARP_SIZE), cgh);
+                sycl::local_accessor<sycl::half2, 1> tile_y_ds_acc_ct1(
+                    sycl::range<1>(mmq_x * WARP_SIZE / QI8_1), cgh);
+
+                cgh.parallel_for(
+                    sycl::nd_range<3>(block_nums * block_dims, block_dims),
+                    [=](sycl::nd_item<3> item_ct1) {
+                        mul_mat_q6_K<need_check>(
+                            vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y,
+                            nrows_dst, item_ct1,
+                            get_pointer(tile_x_ql_acc_ct1),
+                            get_pointer(tile_x_dm_acc_ct1),
+                            get_pointer(tile_x_sc_acc_ct1),
+                            get_pointer(tile_y_qs_acc_ct1),
+                            get_pointer(tile_y_ds_acc_ct1));
+                    });
+            });
+        }
+    }
+}
+catch (sycl::exception const &exc) {
+  std::cerr << exc.what() << "Exception caught at file:" << __FILE__
+            << ", line:" << __LINE__ << std::endl;
+  std::exit(1);
+}
+
+void ggml_sycl_op_mul_mat_q(
+    ggml_backend_sycl_context & ctx,
+    const ggml_tensor *src0, const ggml_tensor *src1, ggml_tensor *dst,
+    const char *src0_dd_i, const float *src1_ddf_i, const char *src1_ddq_i,
+    float *dst_dd_i, const int64_t row_low, const int64_t row_high,
+    const int64_t src1_ncols, const int64_t src1_padded_row_size,
+    const dpct::queue_ptr &stream) try {
+
+    const int64_t ne00 = src0->ne[0];
+
+    const int64_t ne10 = src1->ne[0];
+    GGML_ASSERT(ne10 % QK8_1 == 0);
+
+    const int64_t ne0 = dst->ne[0];
+
+    const int64_t row_diff = row_high - row_low;
+
+    int device_id;
+    SYCL_CHECK(
+        CHECK_TRY_ERROR(device_id = get_current_device_id()));
+
+    // the main device has a larger memory buffer to hold the results from all GPUs
+    // nrows_dst == nrows of the matrix that the dequantize_mul_mat kernel writes into
+    const int64_t nrows_dst = device_id == ctx.device ? ne0 : row_diff;
+
+    switch (src0->type) {
+        case GGML_TYPE_Q4_0:
+            ggml_mul_mat_q4_0_q8_1_sycl(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, src1_ncols, src1_padded_row_size, nrows_dst, stream);
+            break;
+        case GGML_TYPE_Q4_1:
+            ggml_mul_mat_q4_1_q8_1_sycl(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, src1_ncols, src1_padded_row_size, nrows_dst, stream);
+            break;
+        case GGML_TYPE_Q5_0:
+            ggml_mul_mat_q5_0_q8_1_sycl(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, src1_ncols, src1_padded_row_size, nrows_dst, stream);
+            break;
+        case GGML_TYPE_Q5_1:
+            ggml_mul_mat_q5_1_q8_1_sycl(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, src1_ncols, src1_padded_row_size, nrows_dst, stream);
+            break;
+        case GGML_TYPE_Q8_0:
+            ggml_mul_mat_q8_0_q8_1_sycl(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, src1_ncols, src1_padded_row_size, nrows_dst, stream);
+            break;
+        case GGML_TYPE_Q2_K:
+            ggml_mul_mat_q2_K_q8_1_sycl(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, src1_ncols, src1_padded_row_size, nrows_dst, stream);
+            break;
+        case GGML_TYPE_Q3_K:
+            ggml_mul_mat_q3_K_q8_1_sycl(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, src1_ncols, src1_padded_row_size, nrows_dst, stream);
+            break;
+        case GGML_TYPE_Q4_K:
+            ggml_mul_mat_q4_K_q8_1_sycl(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, src1_ncols, src1_padded_row_size, nrows_dst, stream);
+            break;
+        case GGML_TYPE_Q5_K:
+            ggml_mul_mat_q5_K_q8_1_sycl(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, src1_ncols, src1_padded_row_size, nrows_dst, stream);
+            break;
+        case GGML_TYPE_Q6_K:
+            ggml_mul_mat_q6_K_q8_1_sycl(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, src1_ncols, src1_padded_row_size, nrows_dst, stream);
+            break;
+        default:
+            GGML_ABORT("fatal error");
+            break;
+    }
+
+    (void) src1;
+    (void) dst;
+    (void) src1_ddf_i;
+}
+catch (sycl::exception const &exc) {
+  std::cerr << exc.what() << "Exception caught at file:" << __FILE__
+            << ", line:" << __LINE__ << std::endl;
+  std::exit(1);
+}
diff --git a/src/whisper_cpp/ggml/src/ggml-sycl/mmq.hpp b/src/whisper_cpp/ggml/src/ggml-sycl/mmq.hpp
new file mode 100644
index 00000000..3f5297aa
--- /dev/null
+++ b/src/whisper_cpp/ggml/src/ggml-sycl/mmq.hpp
@@ -0,0 +1,33 @@
+//
+// MIT license
+// Copyright (C) 2024 Intel Corporation
+// SPDX-License-Identifier: MIT
+//
+
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+//
+
+#ifndef GGML_SYCL_MMQ_HPP
+#define GGML_SYCL_MMQ_HPP
+
+#include "common.hpp"
+
+void ggml_sycl_op_mul_mat_q(
+    ggml_backend_sycl_context & ctx,
+    const ggml_tensor* src0,
+    const ggml_tensor* src1,
+    ggml_tensor* dst,
+    const char* src0_dd_i,
+    const float* src1_ddf_i,
+    const char* src1_ddq_i,
+    float* dst_dd_i,
+    const int64_t row_low,
+    const int64_t row_high,
+    const int64_t src1_ncols,
+    const int64_t src1_padded_row_size,
+    const dpct::queue_ptr& stream);
+
+#endif // GGML_SYCL_MMQ_HPP
diff --git a/src/whisper_cpp/ggml/src/ggml-sycl/mmvq.cpp b/src/whisper_cpp/ggml/src/ggml-sycl/mmvq.cpp
new file mode 100644
index 00000000..1b96925e
--- /dev/null
+++ b/src/whisper_cpp/ggml/src/ggml-sycl/mmvq.cpp
@@ -0,0 +1,1027 @@
+#include "mmvq.hpp"
+#include "vecdotq.hpp"
+
+
+template <int qk, int qi, typename block_q_t, int vdr, vec_dot_q_sycl_t vec_dot_q_sycl>
+static void mul_mat_vec_q(const void * __restrict__ vx, const void * __restrict__ vy, float * __restrict__ dst, const int ncols, const int nrows,
+                          const sycl::nd_item<3> &item_ct1) {
+    const int row = item_ct1.get_group(2) * item_ct1.get_local_range(1) +
+                    item_ct1.get_local_id(1);
+
+    if (row >= nrows) {
+        return;
+    }
+
+    const int blocks_per_row = ncols / qk;
+    const int blocks_per_warp = vdr * WARP_SIZE / qi;
+
+// partial sum for each thread
+    float tmp = 0.0f;
+
+    const block_q_t  * x = (const block_q_t  *) vx;
+    const block_q8_1 * y = (const block_q8_1 *) vy;
+
+    for (int i = item_ct1.get_local_id(2) / (qi / vdr); i < blocks_per_row;
+         i += blocks_per_warp) {
+        const int ibx = row*blocks_per_row + i; // x block index
+
+        const int iby = i * (qk/QK8_1); // y block index that aligns with ibx
+
+        const int iqs =
+            vdr *
+            (item_ct1.get_local_id(2) %
+             (qi / vdr)); // x block quant index when casting the quants to int
+
+        tmp += vec_dot_q_sycl(&x[ibx], &y[iby], iqs);
+    }
+
+    // sum up partial sums and write back result
+#pragma unroll
+    for (int mask = WARP_SIZE / 2; mask > 0; mask >>= 1) {
+        tmp +=
+            dpct::permute_sub_group_by_xor(item_ct1.get_sub_group(), tmp, mask);
+    }
+
+    if (item_ct1.get_local_id(2) == 0) {
+        dst[row] = tmp;
+    }
+}
+
+template <int qk, int qi, typename block_q_t, int vdr>
+static void mul_mat_vec_q_iq2_xxs_q8_1(const void *__restrict__ vx,
+                                       const void *__restrict__ vy,
+                                       float *__restrict__ dst, const int ncols,
+                                       const int nrows,
+                                       const sycl::nd_item<3> &item_ct1) {
+    const int row = item_ct1.get_group(2) * item_ct1.get_local_range(1) +
+                    item_ct1.get_local_id(1);
+
+    if (row >= nrows) {
+        return;
+    }
+
+    const int blocks_per_row = ncols / qk;
+    const int blocks_per_warp = vdr * WARP_SIZE / qi;
+
+// partial sum for each thread
+    float tmp = 0.0f;
+
+    const block_q_t  * x = (const block_q_t  *) vx;
+    const block_q8_1 * y = (const block_q8_1 *) vy;
+
+    for (int i = item_ct1.get_local_id(2) / (qi / vdr); i < blocks_per_row;
+         i += blocks_per_warp) {
+        const int ibx = row*blocks_per_row + i; // x block index
+
+        const int iby = i * (qk/QK8_1); // y block index that aligns with ibx
+
+        const int iqs =
+            vdr *
+            (item_ct1.get_local_id(2) %
+             (qi / vdr)); // x block quant index when casting the quants to int
+
+        tmp += vec_dot_iq2_xxs_q8_1(&x[ibx], &y[iby], iqs, iq2xxs_grid, ksigns_iq2xs, kmask_iq2xs);
+    }
+
+    // sum up partial sums and write back result
+#pragma unroll
+    for (int mask = WARP_SIZE / 2; mask > 0; mask >>= 1) {
+        tmp +=
+            dpct::permute_sub_group_by_xor(item_ct1.get_sub_group(), tmp, mask);
+    }
+
+    if (item_ct1.get_local_id(2) == 0) {
+        dst[row] = tmp;
+    }
+}
+
+template <int qk, int qi, typename block_q_t, int vdr>
+static void mul_mat_vec_q_iq2_xs_q8_1(const void *__restrict__ vx,
+                                      const void *__restrict__ vy,
+                                      float *__restrict__ dst, const int ncols,
+                                      const int nrows,
+                                      const sycl::nd_item<3> &item_ct1) {
+    const int row = item_ct1.get_group(2) * item_ct1.get_local_range(1) +
+                    item_ct1.get_local_id(1);
+
+    if (row >= nrows) {
+        return;
+    }
+
+    const int blocks_per_row = ncols / qk;
+    const int blocks_per_warp = vdr * WARP_SIZE / qi;
+
+// partial sum for each thread
+    float tmp = 0.0f;
+
+    const block_q_t  * x = (const block_q_t  *) vx;
+    const block_q8_1 * y = (const block_q8_1 *) vy;
+
+    for (int i = item_ct1.get_local_id(2) / (qi / vdr); i < blocks_per_row;
+         i += blocks_per_warp) {
+        const int ibx = row*blocks_per_row + i; // x block index
+
+        const int iby = i * (qk/QK8_1); // y block index that aligns with ibx
+
+        const int iqs =
+            vdr *
+            (item_ct1.get_local_id(2) %
+             (qi / vdr)); // x block quant index when casting the quants to int
+
+        tmp += vec_dot_iq2_xs_q8_1(&x[ibx], &y[iby], iqs, iq2xs_grid, ksigns64);
+    }
+
+    // sum up partial sums and write back result
+#pragma unroll
+    for (int mask = WARP_SIZE / 2; mask > 0; mask >>= 1) {
+        tmp +=
+            dpct::permute_sub_group_by_xor(item_ct1.get_sub_group(), tmp, mask);
+    }
+
+    if (item_ct1.get_local_id(2) == 0) {
+        dst[row] = tmp;
+    }
+}
+
+template <int qk, int qi, typename block_q_t, int vdr>
+static void mul_mat_vec_q_iq2_s_q8_1(const void *__restrict__ vx,
+                                     const void *__restrict__ vy,
+                                     float *__restrict__ dst, const int ncols,
+                                     const int nrows,
+                                     const sycl::nd_item<3> &item_ct1) {
+    const int row = item_ct1.get_group(2) * item_ct1.get_local_range(1) +
+                    item_ct1.get_local_id(1);
+
+    if (row >= nrows) {
+        return;
+    }
+
+    const int blocks_per_row = ncols / qk;
+    const int blocks_per_warp = vdr * WARP_SIZE / qi;
+
+// partial sum for each thread
+    float tmp = 0.0f;
+
+    const block_q_t  * x = (const block_q_t  *) vx;
+    const block_q8_1 * y = (const block_q8_1 *) vy;
+
+    for (int i = item_ct1.get_local_id(2) / (qi / vdr); i < blocks_per_row;
+         i += blocks_per_warp) {
+        const int ibx = row*blocks_per_row + i; // x block index
+
+        const int iby = i * (qk/QK8_1); // y block index that aligns with ibx
+
+        const int iqs =
+            vdr *
+            (item_ct1.get_local_id(2) %
+             (qi / vdr)); // x block quant index when casting the quants to int
+
+        tmp += vec_dot_iq2_s_q8_1(&x[ibx], &y[iby], iqs);
+    }
+
+    // sum up partial sums and write back result
+#pragma unroll
+    for (int mask = WARP_SIZE / 2; mask > 0; mask >>= 1) {
+        tmp +=
+            dpct::permute_sub_group_by_xor(item_ct1.get_sub_group(), tmp, mask);
+    }
+
+    if (item_ct1.get_local_id(2) == 0) {
+        dst[row] = tmp;
+    }
+}
+
+template <int qk, int qi, typename block_q_t, int vdr>
+static void mul_mat_vec_q_iq3_xxs_q8_1(const void *__restrict__ vx,
+                                       const void *__restrict__ vy,
+                                       float *__restrict__ dst, const int ncols,
+                                       const int nrows,
+                                       const sycl::nd_item<3> &item_ct1) {
+    const int row = item_ct1.get_group(2) * item_ct1.get_local_range(1) +
+                    item_ct1.get_local_id(1);
+
+    if (row >= nrows) {
+        return;
+    }
+
+    const int blocks_per_row = ncols / qk;
+    const int blocks_per_warp = vdr * WARP_SIZE / qi;
+
+// partial sum for each thread
+    float tmp = 0.0f;
+
+    const block_q_t  * x = (const block_q_t  *) vx;
+    const block_q8_1 * y = (const block_q8_1 *) vy;
+
+    for (int i = item_ct1.get_local_id(2) / (qi / vdr); i < blocks_per_row;
+         i += blocks_per_warp) {
+        const int ibx = row*blocks_per_row + i; // x block index
+
+        const int iby = i * (qk/QK8_1); // y block index that aligns with ibx
+
+        const int iqs =
+            vdr *
+            (item_ct1.get_local_id(2) %
+             (qi / vdr)); // x block quant index when casting the quants to int
+
+        tmp += vec_dot_iq3_xxs_q8_1(&x[ibx], &y[iby], iqs, iq3xxs_grid, ksigns64);
+    }
+
+    // sum up partial sums and write back result
+#pragma unroll
+    for (int mask = WARP_SIZE / 2; mask > 0; mask >>= 1) {
+        tmp +=
+            dpct::permute_sub_group_by_xor(item_ct1.get_sub_group(), tmp, mask);
+    }
+
+    if (item_ct1.get_local_id(2) == 0) {
+        dst[row] = tmp;
+    }
+}
+
+template <int qk, int qi, typename block_q_t, int vdr>
+static void mul_mat_vec_q_iq3_s_q8_1(const void *__restrict__ vx,
+                                     const void *__restrict__ vy,
+                                     float *__restrict__ dst, const int ncols,
+                                     const int nrows,
+                                     const sycl::nd_item<3> &item_ct1) {
+    const int row = item_ct1.get_group(2) * item_ct1.get_local_range(1) +
+                    item_ct1.get_local_id(1);
+
+    if (row >= nrows) {
+        return;
+    }
+
+    const int blocks_per_row = ncols / qk;
+    const int blocks_per_warp = vdr * WARP_SIZE / qi;
+
+// partial sum for each thread
+    float tmp = 0.0f;
+
+    const block_q_t  * x = (const block_q_t  *) vx;
+    const block_q8_1 * y = (const block_q8_1 *) vy;
+
+    for (int i = item_ct1.get_local_id(2) / (qi / vdr); i < blocks_per_row;
+         i += blocks_per_warp) {
+        const int ibx = row*blocks_per_row + i; // x block index
+
+        const int iby = i * (qk/QK8_1); // y block index that aligns with ibx
+
+        const int iqs =
+            vdr *
+            (item_ct1.get_local_id(2) %
+             (qi / vdr)); // x block quant index when casting the quants to int
+
+        tmp += vec_dot_iq3_s_q8_1(&x[ibx], &y[iby], iqs, iq3s_grid);
+    }
+
+    // sum up partial sums and write back result
+#pragma unroll
+    for (int mask = WARP_SIZE / 2; mask > 0; mask >>= 1) {
+        tmp +=
+            dpct::permute_sub_group_by_xor(item_ct1.get_sub_group(), tmp, mask);
+    }
+
+    if (item_ct1.get_local_id(2) == 0) {
+        dst[row] = tmp;
+    }
+}
+
+template <int qk, int qi, typename block_q_t, int vdr>
+static void mul_mat_vec_q_iq1_s_q8_1(const void *__restrict__ vx,
+                                     const void *__restrict__ vy,
+                                     float *__restrict__ dst, const int ncols,
+                                     const int nrows,
+                                     const sycl::nd_item<3> &item_ct1) {
+    const int row = item_ct1.get_group(2) * item_ct1.get_local_range(1) +
+                    item_ct1.get_local_id(1);
+
+    if (row >= nrows) {
+        return;
+    }
+
+    const int blocks_per_row = ncols / qk;
+    const int blocks_per_warp = vdr * WARP_SIZE / qi;
+
+// partial sum for each thread
+    float tmp = 0.0f;
+
+    const block_q_t  * x = (const block_q_t  *) vx;
+    const block_q8_1 * y = (const block_q8_1 *) vy;
+
+    for (int i = item_ct1.get_local_id(2) / (qi / vdr); i < blocks_per_row;
+         i += blocks_per_warp) {
+        const int ibx = row*blocks_per_row + i; // x block index
+
+        const int iby = i * (qk/QK8_1); // y block index that aligns with ibx
+
+        const int iqs =
+            vdr *
+            (item_ct1.get_local_id(2) %
+             (qi / vdr)); // x block quant index when casting the quants to int
+
+        tmp += vec_dot_iq1_s_q8_1(&x[ibx], &y[iby], iqs, iq1s_grid_gpu);
+    }
+
+    // sum up partial sums and write back result
+#pragma unroll
+    for (int mask = WARP_SIZE / 2; mask > 0; mask >>= 1) {
+        tmp +=
+            dpct::permute_sub_group_by_xor(item_ct1.get_sub_group(), tmp, mask);
+    }
+
+    if (item_ct1.get_local_id(2) == 0) {
+        dst[row] = tmp;
+    }
+}
+
+template <int qk, int qi, typename block_q_t, int vdr>
+static void mul_mat_vec_q_iq1_m_q8_1(const void *__restrict__ vx,
+                                     const void *__restrict__ vy,
+                                     float *__restrict__ dst, const int ncols,
+                                     const int nrows,
+                                     const sycl::nd_item<3> &item_ct1) {
+    const int row = item_ct1.get_group(2) * item_ct1.get_local_range(1) +
+                    item_ct1.get_local_id(1);
+
+    if (row >= nrows) {
+        return;
+    }
+
+    const int blocks_per_row = ncols / qk;
+    const int blocks_per_warp = vdr * WARP_SIZE / qi;
+
+// partial sum for each thread
+    float tmp = 0.0f;
+
+    const block_q_t  * x = (const block_q_t  *) vx;
+    const block_q8_1 * y = (const block_q8_1 *) vy;
+
+    for (int i = item_ct1.get_local_id(2) / (qi / vdr); i < blocks_per_row;
+         i += blocks_per_warp) {
+        const int ibx = row*blocks_per_row + i; // x block index
+
+        const int iby = i * (qk/QK8_1); // y block index that aligns with ibx
+
+        const int iqs =
+            vdr *
+            (item_ct1.get_local_id(2) %
+             (qi / vdr)); // x block quant index when casting the quants to int
+
+        tmp += vec_dot_iq1_m_q8_1(&x[ibx], &y[iby], iqs);
+    }
+
+    // sum up partial sums and write back result
+#pragma unroll
+    for (int mask = WARP_SIZE / 2; mask > 0; mask >>= 1) {
+        tmp +=
+            dpct::permute_sub_group_by_xor(item_ct1.get_sub_group(), tmp, mask);
+    }
+
+    if (item_ct1.get_local_id(2) == 0) {
+        dst[row] = tmp;
+    }
+}
+
+template <int qk, int qi, typename block_q_t, int vdr>
+static void mul_mat_vec_q_iq4_nl_q8_1(const void *__restrict__ vx,
+                                      const void *__restrict__ vy,
+                                      float *__restrict__ dst, const int ncols,
+                                      const int nrows,
+                                      const sycl::nd_item<3> &item_ct1) {
+    const int row = item_ct1.get_group(2) * item_ct1.get_local_range(1) +
+                    item_ct1.get_local_id(1);
+
+    if (row >= nrows) {
+        return;
+    }
+
+    const int blocks_per_row = ncols / qk;
+    const int blocks_per_warp = vdr * WARP_SIZE / qi;
+
+// partial sum for each thread
+    float tmp = 0.0f;
+
+    const block_q_t  * x = (const block_q_t  *) vx;
+    const block_q8_1 * y = (const block_q8_1 *) vy;
+
+    for (int i = item_ct1.get_local_id(2) / (qi / vdr); i < blocks_per_row;
+         i += blocks_per_warp) {
+        const int ibx = row*blocks_per_row + i; // x block index
+
+        const int iby = i * (qk/QK8_1); // y block index that aligns with ibx
+
+        const int iqs =
+            vdr *
+            (item_ct1.get_local_id(2) %
+             (qi / vdr)); // x block quant index when casting the quants to int
+
+        tmp += vec_dot_iq4_nl_q8_1(&x[ibx], &y[iby], iqs);
+    }
+
+    // sum up partial sums and write back result
+#pragma unroll
+    for (int mask = WARP_SIZE / 2; mask > 0; mask >>= 1) {
+        tmp +=
+            dpct::permute_sub_group_by_xor(item_ct1.get_sub_group(), tmp, mask);
+    }
+
+    if (item_ct1.get_local_id(2) == 0) {
+        dst[row] = tmp;
+    }
+}
+
+
+template <int qk, int qi, typename block_q_t, int vdr>
+static void mul_mat_vec_q_iq4_xs_q8_1(const void *__restrict__ vx,
+                                      const void *__restrict__ vy,
+                                      float *__restrict__ dst, const int ncols,
+                                      const int nrows,
+                                      const sycl::nd_item<3> &item_ct1) {
+    const int row = item_ct1.get_group(2) * item_ct1.get_local_range(1) +
+                    item_ct1.get_local_id(1);
+
+    if (row >= nrows) {
+        return;
+    }
+
+    const int blocks_per_row = ncols / qk;
+    const int blocks_per_warp = vdr * WARP_SIZE / qi;
+
+// partial sum for each thread
+    float tmp = 0.0f;
+
+    const block_q_t  * x = (const block_q_t  *) vx;
+    const block_q8_1 * y = (const block_q8_1 *) vy;
+
+    for (int i = item_ct1.get_local_id(2) / (qi / vdr); i < blocks_per_row;
+         i += blocks_per_warp) {
+        const int ibx = row*blocks_per_row + i; // x block index
+
+        const int iby = i * (qk/QK8_1); // y block index that aligns with ibx
+
+        const int iqs =
+            vdr *
+            (item_ct1.get_local_id(2) %
+             (qi / vdr)); // x block quant index when casting the quants to int
+
+        tmp += vec_dot_iq4_xs_q8_1(&x[ibx], &y[iby], iqs);
+    }
+
+    // sum up partial sums and write back result
+#pragma unroll
+    for (int mask = WARP_SIZE / 2; mask > 0; mask >>= 1) {
+        tmp +=
+            dpct::permute_sub_group_by_xor(item_ct1.get_sub_group(), tmp, mask);
+    }
+
+    if (item_ct1.get_local_id(2) == 0) {
+        dst[row] = tmp;
+    }
+}
+
+static void mul_mat_vec_q4_0_q8_1_sycl(const void *vx, const void *vy,
+                                       float *dst, const int ncols,
+                                       const int nrows,
+                                       dpct::queue_ptr stream) {
+    GGML_ASSERT(ncols % QK4_0 == 0);
+    const int block_num_y = (nrows + GGML_SYCL_MMV_Y - 1) / GGML_SYCL_MMV_Y;
+    const sycl::range<3> block_nums(1, 1, block_num_y);
+    const sycl::range<3> block_dims(1, GGML_SYCL_MMV_Y, WARP_SIZE);
+    {
+
+        stream->submit([&](sycl::handler &cgh) {
+
+            cgh.parallel_for(
+                sycl::nd_range<3>(block_nums * block_dims, block_dims),
+                [=](sycl::nd_item<3> item_ct1)
+                    [[intel::reqd_sub_group_size(WARP_SIZE)]] {
+                        mul_mat_vec_q<QK4_0, QI4_0, block_q4_0,
+                                      VDR_Q4_0_Q8_1_MMVQ, vec_dot_q4_0_q8_1>(
+                            vx, vy, dst, ncols, nrows, item_ct1);
+                    });
+        });
+    }
+}
+
+static void mul_mat_vec_q4_1_q8_1_sycl(const void *vx, const void *vy,
+                                       float *dst, const int ncols,
+                                       const int nrows,
+                                       dpct::queue_ptr stream) {
+    GGML_ASSERT(ncols % QK4_1 == 0);
+    const int block_num_y = (nrows + GGML_SYCL_MMV_Y - 1) / GGML_SYCL_MMV_Y;
+    const sycl::range<3> block_nums(1, 1, block_num_y);
+    const sycl::range<3> block_dims(1, GGML_SYCL_MMV_Y, WARP_SIZE);
+    {
+
+        stream->submit([&](sycl::handler &cgh) {
+
+            cgh.parallel_for(
+                sycl::nd_range<3>(block_nums * block_dims, block_dims),
+                [=](sycl::nd_item<3> item_ct1)
+                    [[intel::reqd_sub_group_size(WARP_SIZE)]] {
+                        mul_mat_vec_q<QK4_0, QI4_1, block_q4_1,
+                                      VDR_Q4_1_Q8_1_MMVQ, vec_dot_q4_1_q8_1>(
+                            vx, vy, dst, ncols, nrows, item_ct1);
+                    });
+        });
+    }
+}
+
+static void mul_mat_vec_q5_0_q8_1_sycl(const void *vx, const void *vy,
+                                       float *dst, const int ncols,
+                                       const int nrows,
+                                       dpct::queue_ptr stream) {
+    GGML_ASSERT(ncols % QK5_0 == 0);
+    const int block_num_y = (nrows + GGML_SYCL_MMV_Y - 1) / GGML_SYCL_MMV_Y;
+    const sycl::range<3> block_nums(1, 1, block_num_y);
+    const sycl::range<3> block_dims(1, GGML_SYCL_MMV_Y, WARP_SIZE);
+    {
+
+        stream->submit([&](sycl::handler &cgh) {
+
+            cgh.parallel_for(
+                sycl::nd_range<3>(block_nums * block_dims, block_dims),
+                [=](sycl::nd_item<3> item_ct1)
+                    [[intel::reqd_sub_group_size(WARP_SIZE)]] {
+                        mul_mat_vec_q<QK5_0, QI5_0, block_q5_0,
+                                      VDR_Q5_0_Q8_1_MMVQ, vec_dot_q5_0_q8_1>(
+                            vx, vy, dst, ncols, nrows, item_ct1);
+                    });
+        });
+    }
+}
+
+static void mul_mat_vec_q5_1_q8_1_sycl(const void *vx, const void *vy,
+                                       float *dst, const int ncols,
+                                       const int nrows,
+                                       dpct::queue_ptr stream) {
+    GGML_ASSERT(ncols % QK5_1 == 0);
+    const int block_num_y = (nrows + GGML_SYCL_MMV_Y - 1) / GGML_SYCL_MMV_Y;
+    const sycl::range<3> block_nums(1, 1, block_num_y);
+    const sycl::range<3> block_dims(1, GGML_SYCL_MMV_Y, WARP_SIZE);
+    {
+
+        stream->submit([&](sycl::handler &cgh) {
+
+            cgh.parallel_for(
+                sycl::nd_range<3>(block_nums * block_dims, block_dims),
+                [=](sycl::nd_item<3> item_ct1)
+                    [[intel::reqd_sub_group_size(WARP_SIZE)]] {
+                        mul_mat_vec_q<QK5_1, QI5_1, block_q5_1,
+                                      VDR_Q5_1_Q8_1_MMVQ, vec_dot_q5_1_q8_1>(
+                            vx, vy, dst, ncols, nrows, item_ct1);
+                    });
+        });
+    }
+}
+
+static void mul_mat_vec_q8_0_q8_1_sycl(const void *vx, const void *vy,
+                                       float *dst, const int ncols,
+                                       const int nrows,
+                                       dpct::queue_ptr stream) {
+    GGML_ASSERT(ncols % QK8_0 == 0);
+    const int block_num_y = (nrows + GGML_SYCL_MMV_Y - 1) / GGML_SYCL_MMV_Y;
+    const sycl::range<3> block_nums(1, 1, block_num_y);
+    const sycl::range<3> block_dims(1, GGML_SYCL_MMV_Y, WARP_SIZE);
+    {
+
+        stream->submit([&](sycl::handler &cgh) {
+
+            cgh.parallel_for(
+                sycl::nd_range<3>(block_nums * block_dims, block_dims),
+                [=](sycl::nd_item<3> item_ct1)
+                    [[intel::reqd_sub_group_size(WARP_SIZE)]] {
+                        mul_mat_vec_q<QK8_0, QI8_0, block_q8_0,
+                                      VDR_Q8_0_Q8_1_MMVQ, vec_dot_q8_0_q8_1>(
+                            vx, vy, dst, ncols, nrows, item_ct1);
+                    });
+        });
+    }
+}
+
+static void mul_mat_vec_q2_K_q8_1_sycl(const void *vx, const void *vy,
+                                       float *dst, const int ncols,
+                                       const int nrows,
+                                       dpct::queue_ptr stream) {
+    GGML_ASSERT(ncols % QK_K == 0);
+    const int block_num_y = (nrows + GGML_SYCL_MMV_Y - 1) / GGML_SYCL_MMV_Y;
+    const sycl::range<3> block_nums(1, 1, block_num_y);
+    const sycl::range<3> block_dims(1, GGML_SYCL_MMV_Y, WARP_SIZE);
+    {
+
+        stream->submit([&](sycl::handler &cgh) {
+
+            cgh.parallel_for(
+                sycl::nd_range<3>(block_nums * block_dims, block_dims),
+                [=](sycl::nd_item<3> item_ct1)
+                    [[intel::reqd_sub_group_size(WARP_SIZE)]] {
+                        mul_mat_vec_q<QK_K, QI2_K, block_q2_K,
+                                      VDR_Q2_K_Q8_1_MMVQ, vec_dot_q2_K_q8_1>(
+                            vx, vy, dst, ncols, nrows, item_ct1);
+                    });
+        });
+    }
+}
+
+static void mul_mat_vec_q3_K_q8_1_sycl(const void *vx, const void *vy,
+                                       float *dst, const int ncols,
+                                       const int nrows,
+                                       dpct::queue_ptr stream) {
+    GGML_ASSERT(ncols % QK_K == 0);
+    const int block_num_y = (nrows + GGML_SYCL_MMV_Y - 1) / GGML_SYCL_MMV_Y;
+    const sycl::range<3> block_nums(1, 1, block_num_y);
+    const sycl::range<3> block_dims(1, GGML_SYCL_MMV_Y, WARP_SIZE);
+    {
+
+        stream->submit([&](sycl::handler &cgh) {
+
+            cgh.parallel_for(
+                sycl::nd_range<3>(block_nums * block_dims, block_dims),
+                [=](sycl::nd_item<3> item_ct1)
+                    [[intel::reqd_sub_group_size(WARP_SIZE)]] {
+                        mul_mat_vec_q<QK_K, QI3_K, block_q3_K,
+                                      VDR_Q3_K_Q8_1_MMVQ, vec_dot_q3_K_q8_1>(
+                            vx, vy, dst, ncols, nrows, item_ct1);
+                    });
+        });
+    }
+}
+
+static void mul_mat_vec_q4_K_q8_1_sycl(const void *vx, const void *vy,
+                                       float *dst, const int ncols,
+                                       const int nrows,
+                                       dpct::queue_ptr stream) {
+    GGML_ASSERT(ncols % QK_K == 0);
+    const int block_num_y = (nrows + GGML_SYCL_MMV_Y - 1) / GGML_SYCL_MMV_Y;
+    const sycl::range<3> block_nums(1, 1, block_num_y);
+    const sycl::range<3> block_dims(1, GGML_SYCL_MMV_Y, WARP_SIZE);
+    {
+
+        stream->submit([&](sycl::handler &cgh) {
+
+            cgh.parallel_for(
+                sycl::nd_range<3>(block_nums * block_dims, block_dims),
+                [=](sycl::nd_item<3> item_ct1)
+                    [[intel::reqd_sub_group_size(WARP_SIZE)]] {
+                        mul_mat_vec_q<QK_K, QI4_K, block_q4_K,
+                                      VDR_Q4_K_Q8_1_MMVQ, vec_dot_q4_K_q8_1>(
+                            vx, vy, dst, ncols, nrows, item_ct1);
+                    });
+        });
+    }
+}
+
+static void mul_mat_vec_q5_K_q8_1_sycl(const void *vx, const void *vy,
+                                       float *dst, const int ncols,
+                                       const int nrows,
+                                       dpct::queue_ptr stream) {
+    GGML_ASSERT(ncols % QK_K == 0);
+    const int block_num_y = (nrows + GGML_SYCL_MMV_Y - 1) / GGML_SYCL_MMV_Y;
+    const sycl::range<3> block_nums(1, 1, block_num_y);
+    const sycl::range<3> block_dims(1, GGML_SYCL_MMV_Y, WARP_SIZE);
+    {
+
+        stream->submit([&](sycl::handler &cgh) {
+
+            cgh.parallel_for(
+                sycl::nd_range<3>(block_nums * block_dims, block_dims),
+                [=](sycl::nd_item<3> item_ct1)
+                    [[intel::reqd_sub_group_size(WARP_SIZE)]] {
+                        mul_mat_vec_q<QK_K, QI5_K, block_q5_K,
+                                      VDR_Q5_K_Q8_1_MMVQ, vec_dot_q5_K_q8_1>(
+                            vx, vy, dst, ncols, nrows, item_ct1);
+                    });
+        });
+    }
+}
+
+static void mul_mat_vec_q6_K_q8_1_sycl(const void *vx, const void *vy,
+                                       float *dst, const int ncols,
+                                       const int nrows,
+                                       dpct::queue_ptr stream) {
+    GGML_ASSERT(ncols % QK_K == 0);
+    const int block_num_y = (nrows + GGML_SYCL_MMV_Y - 1) / GGML_SYCL_MMV_Y;
+    const sycl::range<3> block_nums(1, 1, block_num_y);
+    const sycl::range<3> block_dims(1, GGML_SYCL_MMV_Y, WARP_SIZE);
+    {
+
+        stream->submit([&](sycl::handler &cgh) {
+
+            cgh.parallel_for(
+                sycl::nd_range<3>(block_nums * block_dims, block_dims),
+                [=](sycl::nd_item<3> item_ct1)
+                    [[intel::reqd_sub_group_size(WARP_SIZE)]] {
+                        mul_mat_vec_q<QK_K, QI6_K, block_q6_K,
+                                      VDR_Q6_K_Q8_1_MMVQ, vec_dot_q6_K_q8_1>(
+                            vx, vy, dst, ncols, nrows, item_ct1);
+                    });
+        });
+    }
+}
+
+
+static void mul_mat_vec_iq2_xxs_q8_1_sycl(const void *vx, const void *vy,
+                                          float *dst, const int ncols,
+                                          const int nrows,
+                                          dpct::queue_ptr stream) {
+    GGML_ASSERT(ncols % QK_K == 0);
+    const int block_num_y = (nrows + GGML_SYCL_MMV_Y - 1) / GGML_SYCL_MMV_Y;
+    const sycl::range<3> block_nums(1, 1, block_num_y);
+    const sycl::range<3> block_dims(1, GGML_SYCL_MMV_Y, WARP_SIZE);
+    {
+        stream->submit([&](sycl::handler &cgh) {
+            cgh.parallel_for(
+                sycl::nd_range<3>(block_nums * block_dims, block_dims),
+                [=](sycl::nd_item<3> item_ct1)
+                    [[intel::reqd_sub_group_size(WARP_SIZE)]] {
+                        mul_mat_vec_q_iq2_xxs_q8_1<QK_K, QI2_XXS/2, block_iq2_xxs, 1>(
+                            vx, vy, dst, ncols, nrows, item_ct1);
+                    });
+        });
+    }
+}
+
+static void mul_mat_vec_iq2_xs_q8_1_sycl(const void *vx, const void *vy,
+                                         float *dst, const int ncols,
+                                         const int nrows,
+                                         dpct::queue_ptr stream) {
+    GGML_ASSERT(ncols % QK_K == 0);
+    const int block_num_y = (nrows + GGML_SYCL_MMV_Y - 1) / GGML_SYCL_MMV_Y;
+    const sycl::range<3> block_nums(1, 1, block_num_y);
+    const sycl::range<3> block_dims(1, GGML_SYCL_MMV_Y, WARP_SIZE);
+    {
+
+        stream->submit([&](sycl::handler &cgh) {
+            auto iq2xs_grid_ptr_ct1 = &iq2xs_grid[0];
+            auto ksigns64_ptr_ct1 = &ksigns64[0];
+
+            cgh.parallel_for(
+                sycl::nd_range<3>(block_nums * block_dims, block_dims),
+                [=](sycl::nd_item<3> item_ct1)
+                    [[intel::reqd_sub_group_size(WARP_SIZE)]] {
+                        mul_mat_vec_q_iq2_xs_q8_1<QK_K, QI2_XS/2, block_iq2_xs, 1>(
+                            vx, vy, dst, ncols, nrows, item_ct1);
+                    });
+        });
+    }
+}
+
+static void mul_mat_vec_iq2_s_q8_1_sycl(const void *vx, const void *vy,
+                                         float *dst, const int ncols,
+                                         const int nrows,
+                                         dpct::queue_ptr stream) {
+    GGML_ASSERT(ncols % QK_K == 0);
+    const int block_num_y = (nrows + GGML_SYCL_MMV_Y - 1) / GGML_SYCL_MMV_Y;
+    const sycl::range<3> block_nums(1, 1, block_num_y);
+    const sycl::range<3> block_dims(1, GGML_SYCL_MMV_Y, WARP_SIZE);
+    {
+
+        stream->submit([&](sycl::handler &cgh) {
+            auto iq2xs_grid_ptr_ct1 = &iq2xs_grid[0];
+            auto ksigns64_ptr_ct1 = &ksigns64[0];
+
+            cgh.parallel_for(
+                sycl::nd_range<3>(block_nums * block_dims, block_dims),
+                [=](sycl::nd_item<3> item_ct1)
+                    [[intel::reqd_sub_group_size(WARP_SIZE)]] {
+                        mul_mat_vec_q_iq2_s_q8_1<QK_K, QI2_S/2, block_iq2_s, 1>(
+                            vx, vy, dst, ncols, nrows, item_ct1);
+                    });
+        });
+    }
+}
+
+static void mul_mat_vec_iq3_xxs_q8_1_sycl(const void *vx, const void *vy,
+                                          float *dst, const int ncols,
+                                          const int nrows,
+                                          dpct::queue_ptr stream) {
+    GGML_ASSERT(ncols % QK_K == 0);
+    const int block_num_y = (nrows + GGML_SYCL_MMV_Y - 1) / GGML_SYCL_MMV_Y;
+    const sycl::range<3> block_nums(1, 1, block_num_y);
+    const sycl::range<3> block_dims(1, GGML_SYCL_MMV_Y, WARP_SIZE);
+    {
+
+        stream->submit([&](sycl::handler &cgh) {
+            auto iq3xxs_grid_ptr_ct1 = &iq3xxs_grid[0];
+            auto ksigns64_ptr_ct1 = &ksigns64[0];
+
+            cgh.parallel_for(
+                sycl::nd_range<3>(block_nums * block_dims, block_dims),
+                [=](sycl::nd_item<3> item_ct1)
+                    [[intel::reqd_sub_group_size(WARP_SIZE)]] {
+                        mul_mat_vec_q_iq3_xxs_q8_1<QK_K, QI3_XXS/2, block_iq3_xxs, 1>(
+                            vx, vy, dst, ncols, nrows, item_ct1);
+                    });
+        });
+    }
+}
+
+static void mul_mat_vec_iq3_s_q8_1_sycl(const void *vx, const void *vy,
+                                          float *dst, const int ncols,
+                                          const int nrows,
+                                          dpct::queue_ptr stream) {
+    GGML_ASSERT(ncols % QK_K == 0);
+    const int block_num_y = (nrows + GGML_SYCL_MMV_Y - 1) / GGML_SYCL_MMV_Y;
+    const sycl::range<3> block_nums(1, 1, block_num_y);
+    const sycl::range<3> block_dims(1, GGML_SYCL_MMV_Y, WARP_SIZE);
+    {
+
+        stream->submit([&](sycl::handler &cgh) {
+            auto iq3s_grid_ptr_ct1 = &iq3s_grid[0];
+
+            cgh.parallel_for(
+                sycl::nd_range<3>(block_nums * block_dims, block_dims),
+                [=](sycl::nd_item<3> item_ct1)
+                    [[intel::reqd_sub_group_size(WARP_SIZE)]] {
+                        mul_mat_vec_q_iq3_s_q8_1<QK_K, QI3_S/2, block_iq3_s, 1>(
+                            vx, vy, dst, ncols, nrows, item_ct1);
+                    });
+        });
+    }
+}
+
+static void mul_mat_vec_iq1_s_q8_1_sycl(const void *vx, const void *vy,
+                                          float *dst, const int ncols,
+                                          const int nrows,
+                                          dpct::queue_ptr stream) {
+    GGML_ASSERT(ncols % QK_K == 0);
+    const int block_num_y = (nrows + GGML_SYCL_MMV_Y - 1) / GGML_SYCL_MMV_Y;
+    const sycl::range<3> block_nums(1, 1, block_num_y);
+    const sycl::range<3> block_dims(1, GGML_SYCL_MMV_Y, WARP_SIZE);
+    {
+
+        stream->submit([&](sycl::handler &cgh) {
+            auto iq1s_grid_ptr_ct1 = &iq1s_grid_gpu[0];
+            auto ksigns64_ptr_ct1 = &ksigns64[0];
+
+            cgh.parallel_for(
+                sycl::nd_range<3>(block_nums * block_dims, block_dims),
+                [=](sycl::nd_item<3> item_ct1)
+                    [[intel::reqd_sub_group_size(WARP_SIZE)]] {
+                        mul_mat_vec_q_iq1_s_q8_1<QK_K, QI1_S, block_iq1_s, 1>(
+                            vx, vy, dst, ncols, nrows, item_ct1);
+                    });
+        });
+    }
+}
+
+static void mul_mat_vec_iq1_m_q8_1_sycl(const void *vx, const void *vy,
+                                          float *dst, const int ncols,
+                                          const int nrows,
+                                          dpct::queue_ptr stream) {
+    GGML_ASSERT(ncols % QK_K == 0);
+    const int block_num_y = (nrows + GGML_SYCL_MMV_Y - 1) / GGML_SYCL_MMV_Y;
+    const sycl::range<3> block_nums(1, 1, block_num_y);
+    const sycl::range<3> block_dims(1, GGML_SYCL_MMV_Y, WARP_SIZE);
+    {
+        stream->submit([&](sycl::handler &cgh) {
+            cgh.parallel_for(
+                sycl::nd_range<3>(block_nums * block_dims, block_dims),
+                [=](sycl::nd_item<3> item_ct1)
+                    [[intel::reqd_sub_group_size(WARP_SIZE)]] {
+                        mul_mat_vec_q_iq1_m_q8_1<QK_K, QI1_S, block_iq1_m, 1>(
+                            vx, vy, dst, ncols, nrows, item_ct1);
+                    });
+        });
+    }
+}
+
+static void mul_mat_vec_iq4_nl_q8_1_sycl(const void *vx, const void *vy,
+                                          float *dst, const int ncols,
+                                          const int nrows,
+                                          dpct::queue_ptr stream) {
+    GGML_ASSERT(ncols % QK4_NL == 0);
+    const int block_num_y = (nrows + GGML_SYCL_MMV_Y - 1) / GGML_SYCL_MMV_Y;
+    const sycl::range<3> block_nums(1, 1, block_num_y);
+    const sycl::range<3> block_dims(1, GGML_SYCL_MMV_Y, WARP_SIZE);
+    {
+
+        stream->submit([&](sycl::handler &cgh) {
+            cgh.parallel_for(
+                sycl::nd_range<3>(block_nums * block_dims, block_dims),
+                [=](sycl::nd_item<3> item_ct1)
+                    [[intel::reqd_sub_group_size(WARP_SIZE)]] {
+                        mul_mat_vec_q_iq4_nl_q8_1<QK4_NL, QI4_NL, block_iq4_nl, 2>(
+                            vx, vy, dst, ncols, nrows, item_ct1);
+                    });
+        });
+    }
+}
+
+static void mul_mat_vec_iq4_xs_q8_1_sycl(const void *vx, const void *vy,
+                                          float *dst, const int ncols,
+                                          const int nrows,
+                                          dpct::queue_ptr stream) {
+    GGML_ASSERT(ncols % QK_K == 0);
+    const int block_num_y = (nrows + GGML_SYCL_MMV_Y - 1) / GGML_SYCL_MMV_Y;
+    const sycl::range<3> block_nums(1, 1, block_num_y);
+    const sycl::range<3> block_dims(1, GGML_SYCL_MMV_Y, WARP_SIZE);
+    {
+
+        stream->submit([&](sycl::handler &cgh) {
+            cgh.parallel_for(
+                sycl::nd_range<3>(block_nums * block_dims, block_dims),
+                [=](sycl::nd_item<3> item_ct1)
+                    [[intel::reqd_sub_group_size(WARP_SIZE)]] {
+                        mul_mat_vec_q_iq4_xs_q8_1<QK_K, QI4_XS/4, block_iq4_xs, 1>(
+                            vx, vy, dst, ncols, nrows, item_ct1);
+                    });
+        });
+    }
+}
+
+void ggml_sycl_op_mul_mat_vec_q(
+    ggml_backend_sycl_context & ctx,
+    const ggml_tensor *src0, const ggml_tensor *src1, ggml_tensor *dst,
+    const char *src0_dd_i, const float *src1_ddf_i, const char *src1_ddq_i,
+    float *dst_dd_i, const int64_t row_low, const int64_t row_high,
+    const int64_t src1_ncols, const int64_t src1_padded_col_size,
+    const dpct::queue_ptr &stream) {
+
+    const int64_t ne10 = src1->ne[0];
+    GGML_ASSERT(ne10 % QK8_1 == 0);
+
+    const int64_t ne00 = src0->ne[0];
+    const int64_t row_diff = row_high - row_low;
+
+    int id;
+    SYCL_CHECK(
+        CHECK_TRY_ERROR(id = get_current_device_id()));
+    const size_t q8_1_ts = sizeof(block_q8_1);
+    const size_t q8_1_bs = QK8_1;
+    // the main device has a larger memory buffer to hold the results from all GPUs
+    // nrows_dst == nrows of the matrix that the kernel writes into
+    const int64_t nrows_dst = id == ctx.device ? ne00 : row_diff;
+    for (int i = 0; i < src1_ncols; i++)
+    {
+        const size_t src1_ddq_i_offset = i * src1_padded_col_size * q8_1_ts / q8_1_bs;
+        const char* src1_ddq_i_bs = src1_ddq_i + src1_ddq_i_offset;
+        float* dst_dd_i_bs = dst_dd_i + i * dst->ne[0];
+        switch (src0->type) {
+        case GGML_TYPE_Q4_0:
+            mul_mat_vec_q4_0_q8_1_sycl(src0_dd_i, src1_ddq_i_bs, dst_dd_i_bs, ne00, row_diff, stream);
+            break;
+        case GGML_TYPE_Q4_1:
+            mul_mat_vec_q4_1_q8_1_sycl(src0_dd_i, src1_ddq_i_bs, dst_dd_i_bs, ne00, row_diff, stream);
+            break;
+        case GGML_TYPE_Q5_0:
+            mul_mat_vec_q5_0_q8_1_sycl(src0_dd_i, src1_ddq_i_bs, dst_dd_i_bs, ne00, row_diff, stream);
+            break;
+        case GGML_TYPE_Q5_1:
+            mul_mat_vec_q5_1_q8_1_sycl(src0_dd_i, src1_ddq_i_bs, dst_dd_i_bs, ne00, row_diff, stream);
+            break;
+        case GGML_TYPE_Q8_0:
+            mul_mat_vec_q8_0_q8_1_sycl(src0_dd_i, src1_ddq_i_bs, dst_dd_i_bs, ne00, row_diff, stream);
+            break;
+        case GGML_TYPE_Q2_K:
+            mul_mat_vec_q2_K_q8_1_sycl(src0_dd_i, src1_ddq_i_bs, dst_dd_i_bs, ne00, row_diff, stream);
+            break;
+        case GGML_TYPE_Q3_K:
+            mul_mat_vec_q3_K_q8_1_sycl(src0_dd_i, src1_ddq_i_bs, dst_dd_i_bs, ne00, row_diff, stream);
+            break;
+        case GGML_TYPE_Q4_K:
+            mul_mat_vec_q4_K_q8_1_sycl(src0_dd_i, src1_ddq_i_bs, dst_dd_i_bs, ne00, row_diff, stream);
+            break;
+        case GGML_TYPE_Q5_K:
+            mul_mat_vec_q5_K_q8_1_sycl(src0_dd_i, src1_ddq_i_bs, dst_dd_i_bs, ne00, row_diff, stream);
+            break;
+        case GGML_TYPE_Q6_K:
+            mul_mat_vec_q6_K_q8_1_sycl(src0_dd_i, src1_ddq_i_bs, dst_dd_i_bs, ne00, row_diff, stream);
+            break;
+        case GGML_TYPE_IQ1_S:
+            mul_mat_vec_iq1_s_q8_1_sycl(src0_dd_i, src1_ddq_i_bs, dst_dd_i_bs, ne00, row_diff, stream);
+            break;
+        case GGML_TYPE_IQ1_M:
+            mul_mat_vec_iq1_m_q8_1_sycl(src0_dd_i, src1_ddq_i_bs, dst_dd_i_bs, ne00, row_diff, stream);
+            break;
+        case GGML_TYPE_IQ2_XXS:
+            mul_mat_vec_iq2_xxs_q8_1_sycl(src0_dd_i, src1_ddq_i_bs, dst_dd_i_bs, ne00, row_diff, stream);
+            break;
+        case GGML_TYPE_IQ2_XS:
+            mul_mat_vec_iq2_xs_q8_1_sycl(src0_dd_i, src1_ddq_i_bs, dst_dd_i_bs, ne00, row_diff, stream);
+            break;
+        case GGML_TYPE_IQ2_S:
+            mul_mat_vec_iq2_s_q8_1_sycl(src0_dd_i, src1_ddq_i_bs, dst_dd_i_bs, ne00, row_diff, stream);
+            break;
+        case GGML_TYPE_IQ3_XXS:
+            mul_mat_vec_iq3_xxs_q8_1_sycl(src0_dd_i, src1_ddq_i_bs, dst_dd_i_bs, ne00, row_diff, stream);
+            break;
+        case GGML_TYPE_IQ3_S:
+            mul_mat_vec_iq3_s_q8_1_sycl(src0_dd_i, src1_ddq_i_bs, dst_dd_i_bs, ne00, row_diff, stream);
+            break;
+        case GGML_TYPE_IQ4_NL:
+            mul_mat_vec_iq4_nl_q8_1_sycl(src0_dd_i, src1_ddq_i_bs, dst_dd_i_bs, ne00, row_diff, stream);
+            break;
+        case GGML_TYPE_IQ4_XS:
+            mul_mat_vec_iq4_xs_q8_1_sycl(src0_dd_i, src1_ddq_i_bs, dst_dd_i_bs, ne00, row_diff, stream);
+            break;
+        default:
+            GGML_ABORT("fatal error");
+            break;
+        }
+    }
+    (void) src1;
+    (void) dst;
+    (void) src1_ddf_i;
+}
diff --git a/src/whisper_cpp/ggml/src/ggml-sycl/mmvq.hpp b/src/whisper_cpp/ggml/src/ggml-sycl/mmvq.hpp
new file mode 100644
index 00000000..049b43d4
--- /dev/null
+++ b/src/whisper_cpp/ggml/src/ggml-sycl/mmvq.hpp
@@ -0,0 +1,27 @@
+//
+// MIT license
+// Copyright (C) 2024 Intel Corporation
+// SPDX-License-Identifier: MIT
+//
+
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+//
+
+#ifndef GGML_SYCL_MMVQ_HPP
+#define GGML_SYCL_MMVQ_HPP
+
+#include "common.hpp"
+
+
+void ggml_sycl_op_mul_mat_vec_q(
+    ggml_backend_sycl_context & ctx,
+    const ggml_tensor *src0, const ggml_tensor *src1, ggml_tensor *dst,
+    const char *src0_dd_i, const float *src1_ddf_i, const char *src1_ddq_i,
+    float *dst_dd_i, const int64_t row_low, const int64_t row_high,
+    const int64_t src1_ncols, const int64_t src1_padded_row_size,
+    const dpct::queue_ptr &stream);
+
+#endif // GGML_SYCL_MMVQ_HPP
diff --git a/src/whisper_cpp/ggml/src/ggml-sycl/norm.cpp b/src/whisper_cpp/ggml/src/ggml-sycl/norm.cpp
new file mode 100644
index 00000000..b3159b9d
--- /dev/null
+++ b/src/whisper_cpp/ggml/src/ggml-sycl/norm.cpp
@@ -0,0 +1,377 @@
+#include "norm.hpp"
+
+static void norm_f32(const float* x, float* dst, const int ncols, const float eps,
+    const sycl::nd_item<3>& item_ct1, sycl::float2* s_sum, int block_size) {
+    const int row = item_ct1.get_group(2) * item_ct1.get_local_range(1) +
+        item_ct1.get_local_id(1);
+    const int tid = item_ct1.get_local_id(2);
+
+    const int nthreads = item_ct1.get_local_range(2);
+    const int nwarps = nthreads / WARP_SIZE;
+    assert(nwarps % WARP_SIZE == 0);
+    sycl::float2 mean_var = sycl::float2(0.f, 0.f);
+
+    for (int col = tid; col < ncols; col += block_size) {
+        const float xi = x[row * ncols + col];
+        mean_var.x() += xi;
+        mean_var.y() += xi * xi;
+    }
+
+    // sum up partial sums
+    mean_var = warp_reduce_sum(mean_var, item_ct1);
+    if (block_size > WARP_SIZE) {
+
+        int warp_id = item_ct1.get_local_id(2) / WARP_SIZE;
+        int lane_id = item_ct1.get_local_id(2) % WARP_SIZE;
+        if (lane_id == 0) {
+            s_sum[warp_id] = mean_var;
+        }
+        /*
+        DPCT1118:0: SYCL group functions and algorithms must be encountered in
+        converged control flow. You may need to adjust the code.
+        */
+        item_ct1.barrier(sycl::access::fence_space::local_space);
+        mean_var = 0.f;
+        int nreduce = nwarps / WARP_SIZE;
+        for (size_t i = 0; i < nreduce; i += 1)
+        {
+            mean_var += s_sum[lane_id + i * WARP_SIZE];
+        }
+        mean_var = warp_reduce_sum(mean_var, item_ct1);
+    }
+
+    const float mean = mean_var.x() / ncols;
+    const float var = mean_var.y() / ncols - mean * mean;
+    const float inv_std = sycl::rsqrt(var + eps);
+
+    for (int col = tid; col < ncols; col += block_size) {
+        dst[row * ncols + col] = (x[row * ncols + col] - mean) * inv_std;
+    }
+}
+
+static void group_norm_f32(const float* x, float* dst, const int group_size, const int ne_elements, const float eps,
+    const sycl::nd_item<3>& item_ct1, float* s_sum, int block_size) {
+    int start = item_ct1.get_group(2) * group_size;
+    int end = start + group_size;
+    const int nthreads = item_ct1.get_local_range(2);
+    const int nwarps = nthreads / WARP_SIZE;
+    assert(nwarps % WARP_SIZE == 0);
+    start += item_ct1.get_local_id(2);
+    int nreduce = nwarps / WARP_SIZE;
+
+    if (end >= ne_elements) {
+        end = ne_elements;
+    }
+
+    float tmp = 0.0f; // partial sum for thread in warp
+
+    for (int j = start; j < end; j += block_size) {
+        tmp += x[j];
+    }
+
+    tmp = warp_reduce_sum(tmp, item_ct1);
+    if (block_size > WARP_SIZE) {
+
+        int warp_id = item_ct1.get_local_id(2) / WARP_SIZE;
+        int lane_id = item_ct1.get_local_id(2) % WARP_SIZE;
+        if (lane_id == 0) {
+            s_sum[warp_id] = tmp;
+        }
+        /*
+        DPCT1118:1: SYCL group functions and algorithms must be encountered in
+        converged control flow. You may need to adjust the code.
+        */
+        /*
+        DPCT1065:54: Consider replacing sycl::nd_item::barrier() with
+        sycl::nd_item::barrier(sycl::access::fence_space::local_space) for
+        better performance if there is no access to global memory.
+        */
+        item_ct1.barrier();
+        tmp = 0.f;
+        for (size_t i = 0; i < nreduce; i += 1)
+        {
+            tmp += s_sum[lane_id + i * WARP_SIZE];
+        }
+        tmp = warp_reduce_sum(tmp, item_ct1);
+    }
+
+    float mean = tmp / group_size;
+    tmp = 0.0f;
+
+    for (int j = start; j < end; j += block_size) {
+        float xi = x[j] - mean;
+        dst[j] = xi;
+        tmp += xi * xi;
+    }
+
+    tmp = warp_reduce_sum(tmp, item_ct1);
+    if (block_size > WARP_SIZE) {
+
+        int warp_id = item_ct1.get_local_id(2) / WARP_SIZE;
+        int lane_id = item_ct1.get_local_id(2) % WARP_SIZE;
+        if (lane_id == 0) {
+            s_sum[warp_id] = tmp;
+        }
+        /*
+        DPCT1118:2: SYCL group functions and algorithms must be encountered in
+        converged control flow. You may need to adjust the code.
+        */
+        /*
+        DPCT1065:55: Consider replacing sycl::nd_item::barrier() with
+        sycl::nd_item::barrier(sycl::access::fence_space::local_space) for
+        better performance if there is no access to global memory.
+        */
+        item_ct1.barrier();
+        tmp = 0.f;
+        for (size_t i = 0; i < nreduce; i += 1)
+        {
+            tmp += s_sum[lane_id + i * WARP_SIZE];
+        }
+        tmp = warp_reduce_sum(tmp, item_ct1);
+    }
+
+    float variance = tmp / group_size;
+    float scale = sycl::rsqrt(variance + eps);
+    for (int j = start; j < end; j += block_size) {
+        dst[j] *= scale;
+    }
+}
+
+static void rms_norm_f32(const float* x, float* dst, const int ncols, const float eps,
+    const sycl::nd_item<3>& item_ct1, float* s_sum, int block_size) {
+    const int row = item_ct1.get_group(2) * item_ct1.get_local_range(1) +
+        item_ct1.get_local_id(1);
+    const int tid = item_ct1.get_local_id(2);
+    const int nthreads = item_ct1.get_local_range(2);
+    const int nwarps = nthreads / WARP_SIZE;
+    assert(nwarps % WARP_SIZE == 0);
+    float tmp = 0.0f; // partial sum for thread in warp
+
+    for (int col = tid; col < ncols; col += block_size) {
+        const float xi = x[row * ncols + col];
+        tmp += xi * xi;
+    }
+
+    // sum up partial sums
+    tmp = warp_reduce_sum(tmp, item_ct1);
+    if (block_size > WARP_SIZE) {
+
+        int warp_id = item_ct1.get_local_id(2) / WARP_SIZE;
+        int lane_id = item_ct1.get_local_id(2) % WARP_SIZE;
+        if (lane_id == 0) {
+            s_sum[warp_id] = tmp;
+        }
+        /*
+        DPCT1118:3: SYCL group functions and algorithms must be encountered in
+        converged control flow. You may need to adjust the code.
+        */
+        item_ct1.barrier(sycl::access::fence_space::local_space);
+        int nreduce = nwarps / WARP_SIZE;
+        tmp = 0.f;
+        for (size_t i = 0; i < nreduce; i += 1)
+        {
+            tmp += s_sum[lane_id + i * WARP_SIZE];
+        }
+        tmp = warp_reduce_sum(tmp, item_ct1);
+    }
+
+    const float mean = tmp / ncols;
+    const float scale = sycl::rsqrt(mean + eps);
+
+    for (int col = tid; col < ncols; col += block_size) {
+        dst[row * ncols + col] = scale * x[row * ncols + col];
+    }
+}
+
+static void norm_f32_sycl(const float* x, float* dst, const int ncols,
+    const int nrows, const float eps,
+    queue_ptr stream, int device) {
+    GGML_ASSERT(ncols % WARP_SIZE == 0);
+    if (ncols < 1024) {
+        const sycl::range<3> block_dims(1, 1, WARP_SIZE);
+        stream->submit([&](sycl::handler& cgh) {
+            cgh.parallel_for(
+                sycl::nd_range<3>(sycl::range<3>(1, 1, nrows) * block_dims,
+                    block_dims),
+                [=](sycl::nd_item<3> item_ct1)
+                [[intel::reqd_sub_group_size(WARP_SIZE)]] {
+                    norm_f32(x, dst, ncols, eps, item_ct1,
+                        nullptr, WARP_SIZE);
+                });
+            });
+    }
+    else {
+        const int work_group_size = ggml_sycl_info().max_work_group_sizes[device];
+        const sycl::range<3> block_dims(1, 1, work_group_size);
+        /*
+        DPCT1049:17: The work-group size passed to the SYCL kernel may exceed
+        the limit. To get the device limit, query
+        info::device::max_work_group_size. Adjust the work-group size if needed.
+        */
+        stream->submit([&](sycl::handler& cgh) {
+            sycl::local_accessor<sycl::float2, 1> s_sum_acc_ct1(
+                sycl::range<1>(work_group_size / WARP_SIZE), cgh);
+
+            cgh.parallel_for(
+                sycl::nd_range<3>(sycl::range<3>(1, 1, nrows) * block_dims,
+                    block_dims),
+                [=](sycl::nd_item<3> item_ct1)
+                [[intel::reqd_sub_group_size(WARP_SIZE)]] {
+                    norm_f32(x, dst, ncols, eps, item_ct1,
+                        get_pointer(s_sum_acc_ct1), work_group_size);
+                });
+            });
+    }
+}
+
+static void group_norm_f32_sycl(const float* x, float* dst,
+    const int num_groups, const float eps, const int group_size,
+    const int ne_elements, queue_ptr stream, int device) {
+    if (group_size < 1024) {
+        const sycl::range<3> block_dims(1, 1, WARP_SIZE);
+        stream->submit([&](sycl::handler& cgh) {
+            const float eps_ct4 = eps;
+            cgh.parallel_for(
+                sycl::nd_range<3>(sycl::range<3>(1, 1, num_groups) * block_dims,
+                    block_dims),
+                [=](sycl::nd_item<3> item_ct1)
+                [[intel::reqd_sub_group_size(WARP_SIZE)]] {
+                    group_norm_f32(
+                        x, dst, group_size, ne_elements, eps_ct4, item_ct1,
+                        nullptr, WARP_SIZE);
+                });
+            });
+    }
+    else {
+        const int work_group_size = ggml_sycl_info().max_work_group_sizes[device];
+        const sycl::range<3> block_dims(1, 1, work_group_size);
+        /*
+        DPCT1049:18: The work-group size passed to the SYCL kernel may exceed
+        the limit. To get the device limit, query
+        info::device::max_work_group_size. Adjust the work-group size if needed.
+        */
+
+        stream->submit([&](sycl::handler& cgh) {
+            sycl::local_accessor<float, 1> s_sum_acc_ct1(sycl::range<1>(work_group_size / WARP_SIZE),
+                cgh);
+
+            const float eps_ct4 = eps;
+
+            cgh.parallel_for(
+                sycl::nd_range<3>(sycl::range<3>(1, 1, num_groups) * block_dims,
+                    block_dims),
+                [=](sycl::nd_item<3> item_ct1)
+                [[intel::reqd_sub_group_size(WARP_SIZE)]] {
+                    group_norm_f32(x, dst, group_size, ne_elements,
+                        eps_ct4, item_ct1,
+                        get_pointer(s_sum_acc_ct1), work_group_size);
+                });
+            });
+    }
+}
+
+static void rms_norm_f32_sycl(const float* x, float* dst, const int ncols,
+    const int nrows, const float eps,
+    queue_ptr stream, int device) {
+    GGML_ASSERT(ncols % WARP_SIZE == 0);
+    // printf("%s ncols=%d, nrows=%d, WARP_SIZE=%d\n", __func__, ncols, nrows, WARP_SIZE);
+    if (ncols < 1024) {
+        const sycl::range<3> block_dims(1, 1, WARP_SIZE);
+        stream->submit([&](sycl::handler& cgh) {
+            cgh.parallel_for(
+                sycl::nd_range<3>(sycl::range<3>(1, 1, nrows) * block_dims,
+                    block_dims),
+                [=](sycl::nd_item<3> item_ct1)
+                [[intel::reqd_sub_group_size(WARP_SIZE)]] {
+                    rms_norm_f32(x, dst, ncols, eps, item_ct1,
+                        nullptr, WARP_SIZE);
+                });
+            });
+    }
+    else {
+        const int work_group_size = ggml_sycl_info().max_work_group_sizes[device];
+        const sycl::range<3> block_dims(1, 1, work_group_size);
+        /*
+        DPCT1049:19: The work-group size passed to the SYCL kernel may exceed
+        the limit. To get the device limit, query
+        info::device::max_work_group_size. Adjust the work-group size if needed.
+        */
+        stream->submit([&](sycl::handler& cgh) {
+            sycl::local_accessor<float, 1> s_sum_acc_ct1(sycl::range<1>(work_group_size / WARP_SIZE),
+                cgh);
+            cgh.parallel_for(
+                sycl::nd_range<3>(sycl::range<3>(1, 1, nrows) * block_dims,
+                    block_dims),
+                [=](sycl::nd_item<3> item_ct1)
+                [[intel::reqd_sub_group_size(WARP_SIZE)]] {
+                    rms_norm_f32(x, dst, ncols, eps, item_ct1,
+                        get_pointer(s_sum_acc_ct1), work_group_size);
+                });
+            });
+    }
+}
+
+void ggml_sycl_op_norm(ggml_backend_sycl_context& ctx, const ggml_tensor* src0, const ggml_tensor* src1,
+    ggml_tensor* dst, const float* src0_dd,
+    const float* src1_dd, float* dst_dd,
+    const queue_ptr& main_stream) {
+
+    GGML_ASSERT(src0->type == GGML_TYPE_F32);
+    GGML_ASSERT(dst->type == GGML_TYPE_F32);
+
+    const int64_t ne00 = src0->ne[0];
+    const int64_t nrows = ggml_nrows(src0);
+
+    float eps;
+    memcpy(&eps, dst->op_params, sizeof(float));
+
+    norm_f32_sycl(src0_dd, dst_dd, ne00, nrows, eps, main_stream, ctx.device);
+
+    (void)src1;
+    (void)dst;
+    (void)src1_dd;
+}
+
+void ggml_sycl_op_group_norm(ggml_backend_sycl_context& ctx, const ggml_tensor* src0,
+    const ggml_tensor* src1, ggml_tensor* dst,
+    const float* src0_dd, const float* src1_dd,
+    float* dst_dd,
+    const queue_ptr& main_stream) {
+
+    GGML_ASSERT(src0->type == GGML_TYPE_F32);
+    GGML_ASSERT(dst->type == GGML_TYPE_F32);
+
+    int num_groups = dst->op_params[0];
+
+    float eps;
+    memcpy(&eps, dst->op_params + 1, sizeof(float));
+
+    int group_size = src0->ne[0] * src0->ne[1] * ((src0->ne[2] + num_groups - 1) / num_groups);
+    group_norm_f32_sycl(src0_dd, dst_dd, num_groups, eps, group_size, src0->ne[0] * src0->ne[1] * src0->ne[2], main_stream, ctx.device);
+
+    (void)src1;
+    (void)dst;
+    (void)src1_dd;
+}
+
+void ggml_sycl_op_rms_norm(ggml_backend_sycl_context& ctx, const ggml_tensor* src0,
+    const ggml_tensor* src1, ggml_tensor* dst,
+    const float* src0_dd, const float* src1_dd,
+    float* dst_dd,
+    const queue_ptr& main_stream) {
+
+    GGML_ASSERT(src0->type == GGML_TYPE_F32);
+    GGML_ASSERT(dst->type == GGML_TYPE_F32);
+
+    const int64_t ne00 = src0->ne[0];
+    const int64_t nrows = ggml_nrows(src0);
+
+    float eps;
+    memcpy(&eps, dst->op_params, sizeof(float));
+
+    rms_norm_f32_sycl(src0_dd, dst_dd, ne00, nrows, eps, main_stream, ctx.device);
+
+    (void)src1;
+    (void)dst;
+    (void)src1_dd;
+}
diff --git a/src/whisper_cpp/ggml/src/ggml-sycl/norm.hpp b/src/whisper_cpp/ggml/src/ggml-sycl/norm.hpp
new file mode 100644
index 00000000..a9ad9156
--- /dev/null
+++ b/src/whisper_cpp/ggml/src/ggml-sycl/norm.hpp
@@ -0,0 +1,35 @@
+//
+// MIT license
+// Copyright (C) 2024 Intel Corporation
+// SPDX-License-Identifier: MIT
+//
+
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+//
+
+#ifndef GGML_SYCL_NORM_HPP
+#define GGML_SYCL_NORM_HPP
+
+#include "common.hpp"
+
+void ggml_sycl_op_norm(ggml_backend_sycl_context& ctx, const ggml_tensor* src0, const ggml_tensor* src1,
+    ggml_tensor* dst, const float* src0_dd,
+    const float* src1_dd, float* dst_dd,
+    const queue_ptr& main_stream);
+
+void ggml_sycl_op_rms_norm(ggml_backend_sycl_context& ctx, const ggml_tensor* src0,
+    const ggml_tensor* src1, ggml_tensor* dst,
+    const float* src0_dd, const float* src1_dd,
+    float* dst_dd,
+    const queue_ptr& main_stream);
+
+void ggml_sycl_op_group_norm(ggml_backend_sycl_context& ctx, const ggml_tensor* src0,
+    const ggml_tensor* src1, ggml_tensor* dst,
+    const float* src0_dd, const float* src1_dd,
+    float* dst_dd,
+    const queue_ptr& main_stream);
+
+#endif // GGML_SYCL_NORM_HPP
diff --git a/src/whisper_cpp/ggml/src/ggml-sycl/presets.hpp b/src/whisper_cpp/ggml/src/ggml-sycl/presets.hpp
new file mode 100644
index 00000000..340ab8e9
--- /dev/null
+++ b/src/whisper_cpp/ggml/src/ggml-sycl/presets.hpp
@@ -0,0 +1,68 @@
+//
+// MIT license
+// Copyright (C) 2024 Intel Corporation
+// SPDX-License-Identifier: MIT
+//
+
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+//
+
+#ifndef GGML_SYCL_PRESETS_HPP
+#define GGML_SYCL_PRESETS_HPP
+
+#define GGML_SYCL_MAX_STREAMS       8
+#define GGML_SYCL_MAX_BUFFERS       256
+
+#define WARP_SIZE GGML_SYCL_WARP_SIZE
+#define MATRIX_ROW_PADDING 512 // last row of quant. matrices is a multiple of this to avoid out-of-bounds memory accesses
+
+#define SYCL_GELU_BLOCK_SIZE 256
+#define SYCL_SILU_BLOCK_SIZE 256
+#define SYCL_TANH_BLOCK_SIZE 256
+#define SYCL_RELU_BLOCK_SIZE 256
+#define SYCL_HARDSIGMOID_BLOCK_SIZE 256
+#define SYCL_HARDSWISH_BLOCK_SIZE 256
+#define SYCL_SQR_BLOCK_SIZE 256
+#define SYCL_CPY_BLOCK_SIZE 32
+#define SYCL_SCALE_BLOCK_SIZE 256
+#define SYCL_CLAMP_BLOCK_SIZE 256
+#define SYCL_ROPE_BLOCK_SIZE 256
+#define SYCL_ALIBI_BLOCK_SIZE 32
+#define SYCL_DIAG_MASK_INF_BLOCK_SIZE 32
+#define SYCL_QUANTIZE_BLOCK_SIZE 256
+#define SYCL_DEQUANTIZE_BLOCK_SIZE 256
+#define SYCL_GET_ROWS_BLOCK_SIZE 256
+#define SYCL_UPSCALE_BLOCK_SIZE 256
+#define SYCL_CONCAT_BLOCK_SIZE 256
+#define SYCL_PAD_BLOCK_SIZE 256
+#define SYCL_ACC_BLOCK_SIZE 256
+#define SYCL_IM2COL_BLOCK_SIZE 256
+#define SYCL_POOL2D_BLOCK_SIZE 256
+#define SYCL_CONV_TRANPOSE_1D_BLOCK_SIZE 256
+#define SYCL_TIMESTEP_EMBEDDING_BLOCK_SIZE 256
+
+// dmmv = dequantize_mul_mat_vec
+#ifndef GGML_SYCL_DMMV_X
+#define GGML_SYCL_DMMV_X 32
+#endif
+#ifndef GGML_SYCL_MMV_Y
+#define GGML_SYCL_MMV_Y 1
+#endif
+
+#ifndef K_QUANTS_PER_ITERATION
+#define K_QUANTS_PER_ITERATION 2
+#else
+static_assert(K_QUANTS_PER_ITERATION == 1 || K_QUANTS_PER_ITERATION == 2, "K_QUANTS_PER_ITERATION must be 1 or 2");
+#endif
+
+#ifndef GGML_SYCL_PEER_MAX_BATCH_SIZE
+#define GGML_SYCL_PEER_MAX_BATCH_SIZE 128
+#endif // GGML_SYCL_PEER_MAX_BATCH_SIZE
+
+#define MUL_MAT_SRC1_COL_STRIDE 128
+
+#define QK_WARP_SIZE 32
+#endif // GGML_SYCL_PRESETS_HPP
diff --git a/src/whisper_cpp/ggml/src/ggml-sycl/rope.cpp b/src/whisper_cpp/ggml/src/ggml-sycl/rope.cpp
new file mode 100644
index 00000000..1f06f78f
--- /dev/null
+++ b/src/whisper_cpp/ggml/src/ggml-sycl/rope.cpp
@@ -0,0 +1,275 @@
+#include "rope.hpp"
+
+struct rope_corr_dims {
+    float v[2];
+};
+
+static float rope_yarn_ramp(const float low, const float high, const int i0) {
+    const float y = (i0 / 2 - low) / sycl::max(0.001f, high - low);
+    return 1.0f - sycl::min(1.0f, sycl::max(0.0f, y));
+}
+
+// YaRN algorithm based on LlamaYaRNScaledRotaryEmbedding.py from https://github.com/jquesnelle/yarn
+// MIT licensed. Copyright (c) 2023 Jeffrey Quesnelle and Bowen Peng.
+static void rope_yarn(
+    float theta_extrap, float freq_scale, rope_corr_dims corr_dims, int64_t i0, float ext_factor, float mscale,
+    float * cos_theta, float * sin_theta) {
+    // Get n-d rotational scaling corrected for extrapolation
+    float theta_interp = freq_scale * theta_extrap;
+    float theta = theta_interp;
+    if (ext_factor != 0.0f) {
+        float ramp_mix = rope_yarn_ramp(corr_dims.v[0], corr_dims.v[1], i0) * ext_factor;
+        theta = theta_interp * (1 - ramp_mix) + theta_extrap * ramp_mix;
+
+        // Get n-d magnitude scaling corrected for interpolation
+        mscale *= 1.0f + 0.1f * sycl::log(1.0f / freq_scale);
+    }
+    *cos_theta = sycl::cos(theta) * mscale;
+    *sin_theta = sycl::sin(theta) * mscale;
+}
+
+template<typename T, bool has_ff>
+static void rope_norm(
+    const T * x, T * dst, int ne0, int n_dims, const int32_t * pos, float freq_scale, int p_delta_rows,
+    float ext_factor, float attn_factor, rope_corr_dims corr_dims, float theta_scale, const float * freq_factors,
+    const sycl::nd_item<3> &item_ct1) {
+    const int i0 = 2 * (item_ct1.get_local_range(1) * item_ct1.get_group(1) +
+                         item_ct1.get_local_id(1));
+
+    if (i0 >= ne0) {
+        return;
+    }
+
+    const int row = item_ct1.get_local_range(2) * item_ct1.get_group(2) +
+                    item_ct1.get_local_id(2);
+
+    if (i0 >= n_dims) {
+        const int i = row*ne0 + i0;
+
+        dst[i + 0] = x[i + 0];
+        dst[i + 1] = x[i + 1];
+
+        return;
+    }
+
+    const int i = row*ne0 + i0;
+    const int i2 = row/p_delta_rows;
+
+    const float theta_base = pos[i2] * sycl::pow(theta_scale, i0 / 2.0f);
+
+    const float freq_factor = has_ff ? freq_factors[i0/2] : 1.0f;
+
+    float cos_theta;
+    float sin_theta;
+
+    rope_yarn(theta_base/freq_factor, freq_scale, corr_dims, i0, ext_factor, attn_factor, &cos_theta, &sin_theta);
+
+    const float x0 = x[i + 0];
+    const float x1 = x[i + 1];
+
+    dst[i + 0] = x0*cos_theta - x1*sin_theta;
+    dst[i + 1] = x0*sin_theta + x1*cos_theta;
+}
+
+template<typename T, bool has_ff>
+static void rope_neox(
+    const T * x, T * dst, int ne0, int n_dims, const int32_t * pos, float freq_scale, int p_delta_rows,
+    float ext_factor, float attn_factor, rope_corr_dims corr_dims, float theta_scale, const float * freq_factors,
+    const sycl::nd_item<3> &item_ct1) {
+    const int i0 = 2 * (item_ct1.get_local_range(1) * item_ct1.get_group(1) +
+                         item_ct1.get_local_id(1));
+
+    if (i0 >= ne0) {
+        return;
+    }
+
+    const int row = item_ct1.get_local_range(2) * item_ct1.get_group(2) +
+                    item_ct1.get_local_id(2);
+
+    if (i0 >= n_dims) {
+        const int i = row*ne0 + i0;
+
+        dst[i + 0] = x[i + 0];
+        dst[i + 1] = x[i + 1];
+
+        return;
+    }
+
+    const int i  = row*ne0 + i0/2;
+    const int i2 = row/p_delta_rows;
+
+    const float theta_base = pos[i2] * sycl::pow(theta_scale, i0 / 2.0f);
+
+    const float freq_factor = has_ff ? freq_factors[i0/2] : 1.0f;
+
+    float cos_theta;
+    float sin_theta;
+
+    rope_yarn(theta_base/freq_factor, freq_scale, corr_dims, i0, ext_factor, attn_factor, &cos_theta, &sin_theta);
+
+    const float x0 = x[i + 0];
+    const float x1 = x[i + n_dims/2];
+
+    dst[i + 0]        = x0*cos_theta - x1*sin_theta;
+    dst[i + n_dims/2] = x0*sin_theta + x1*cos_theta;
+}
+
+template <typename T>
+static void rope_norm_sycl(
+    const T *x, T *dst, int ne0, int n_dims, int nr, const int32_t *pos, float freq_scale, int p_delta_rows,
+    float freq_base, float ext_factor, float attn_factor, rope_corr_dims corr_dims, const float * freq_factors, queue_ptr stream) {
+    GGML_ASSERT(ne0 % 2 == 0);
+    const sycl::range<3> block_dims(1, SYCL_ROPE_BLOCK_SIZE, 1);
+    const int num_blocks_x = (ne0 + 2*SYCL_ROPE_BLOCK_SIZE - 1) / (2*SYCL_ROPE_BLOCK_SIZE);
+    const sycl::range<3> block_nums(1, num_blocks_x, nr);
+
+    const float theta_scale = powf(freq_base, -2.0f/n_dims);
+
+    dpct::has_capability_or_fail(stream->get_device(),
+                                     {sycl::aspect::fp16});
+
+    if (freq_factors == nullptr) {
+        /*
+        DPCT1049:40: The work-group size passed to the SYCL kernel may exceed
+        the limit. To get the device limit, query
+        info::device::max_work_group_size. Adjust the work-group size if needed.
+        */
+        stream->parallel_for(
+            sycl::nd_range<3>(block_nums * block_dims, block_dims),
+            [=](sycl::nd_item<3> item_ct1) {
+                rope_norm<T, false>(x, dst, ne0, n_dims, pos, freq_scale, p_delta_rows,
+                               ext_factor, attn_factor, corr_dims, theta_scale, freq_factors,
+                               item_ct1);
+            });
+    } else {
+        /*
+        DPCT1049:41: The work-group size passed to the SYCL kernel may exceed
+        the limit. To get the device limit, query
+        info::device::max_work_group_size. Adjust the work-group size if needed.
+        */
+        stream->parallel_for(
+            sycl::nd_range<3>(block_nums * block_dims, block_dims),
+            [=](sycl::nd_item<3> item_ct1) {
+                rope_norm<T, true>(x, dst, ne0, n_dims, pos, freq_scale, p_delta_rows,
+                              ext_factor, attn_factor, corr_dims, theta_scale, freq_factors,
+                              item_ct1);
+            });
+    }
+}
+
+template <typename T>
+static void rope_neox_sycl(
+    const T *x, T *dst, int ne0, int n_dims, int nr, const int32_t *pos, float freq_scale, int p_delta_rows,
+    float freq_base, float ext_factor, float attn_factor, rope_corr_dims corr_dims, const float * freq_factors, queue_ptr stream) {
+    GGML_ASSERT(ne0 % 2 == 0);
+    const sycl::range<3> block_dims(1, SYCL_ROPE_BLOCK_SIZE, 1);
+    const int num_blocks_x = (ne0 + 2*SYCL_ROPE_BLOCK_SIZE - 1) / (2*SYCL_ROPE_BLOCK_SIZE);
+    const sycl::range<3> block_nums(1, num_blocks_x, nr);
+
+    const float theta_scale = powf(freq_base, -2.0f/n_dims);
+
+    dpct::has_capability_or_fail(stream->get_device(),
+                                    {sycl::aspect::fp16});
+
+    if (freq_factors == nullptr) {
+        stream->parallel_for(
+            sycl::nd_range<3>(block_nums * block_dims, block_dims),
+            [=](sycl::nd_item<3> item_ct1) {
+                rope_neox<T, false>(x, dst, ne0, n_dims, pos, freq_scale,
+                                    p_delta_rows, ext_factor, attn_factor,
+                                    corr_dims, theta_scale, freq_factors,
+                                    item_ct1);
+            });
+    } else {
+        stream->parallel_for(
+            sycl::nd_range<3>(block_nums * block_dims, block_dims),
+            [=](sycl::nd_item<3> item_ct1) {
+                rope_neox<T, true>(x, dst, ne0, n_dims, pos, freq_scale,
+                                    p_delta_rows, ext_factor, attn_factor,
+                                    corr_dims, theta_scale, freq_factors,
+                                    item_ct1);
+            });
+    }
+}
+
+void ggml_sycl_op_rope(
+    ggml_backend_sycl_context & ctx, const ggml_tensor *src0, const ggml_tensor *src1, ggml_tensor *dst,
+    const float *src0_dd, const float *src1_dd, float *dst_dd, const queue_ptr &main_stream) {
+    const ggml_tensor * src2 = dst->src[2];
+
+    GGML_ASSERT(src0->type == GGML_TYPE_F32 || src0->type == GGML_TYPE_F16);
+    GGML_ASSERT( dst->type == GGML_TYPE_F32 ||  dst->type == GGML_TYPE_F16);
+    GGML_ASSERT(src0->type == dst->type);
+
+    const int64_t ne00 = src0->ne[0];
+    const int64_t ne01 = src0->ne[1];
+    const int64_t nr = ggml_nrows(src0);
+
+    //const int n_past      = ((int32_t *) dst->op_params)[0];
+    const int n_dims      = ((int32_t *) dst->op_params)[1];
+    const int mode        = ((int32_t *) dst->op_params)[2];
+    //const int n_ctx       = ((int32_t *) dst->op_params)[3];
+    const int n_ctx_orig  = ((int32_t *) dst->op_params)[4];
+
+    // RoPE alteration for extended context
+    float freq_base;
+    float freq_scale;
+    float ext_factor;
+    float attn_factor;
+    float beta_fast;
+    float beta_slow;
+
+    memcpy(&freq_base,   (int32_t *) dst->op_params +  5, sizeof(float));
+    memcpy(&freq_scale,  (int32_t *) dst->op_params +  6, sizeof(float));
+    memcpy(&ext_factor,  (int32_t *) dst->op_params +  7, sizeof(float));
+    memcpy(&attn_factor, (int32_t *) dst->op_params +  8, sizeof(float));
+    memcpy(&beta_fast,   (int32_t *) dst->op_params +  9, sizeof(float));
+    memcpy(&beta_slow,   (int32_t *) dst->op_params + 10, sizeof(float));
+
+    const bool is_neox = mode & GGML_ROPE_TYPE_NEOX;
+
+    const int32_t * pos = (const int32_t *) src1_dd;
+
+    const float * freq_factors = nullptr;
+    if (src2 != nullptr) {
+        freq_factors = (const float *) src2->data;
+    }
+
+    rope_corr_dims corr_dims;
+    ggml_rope_yarn_corr_dims(n_dims, n_ctx_orig, freq_base, beta_fast, beta_slow, corr_dims.v);
+
+    // compute
+    if (is_neox) {
+        if (src0->type == GGML_TYPE_F32) {
+            rope_neox_sycl(
+                (const float *)src0_dd, (float *)dst_dd, ne00, n_dims, nr, pos, freq_scale, ne01, freq_base, ext_factor,
+                attn_factor, corr_dims, freq_factors, main_stream
+            );
+        } else if (src0->type == GGML_TYPE_F16) {
+            rope_neox_sycl(
+                (const sycl::half *)src0_dd, (sycl::half *)dst_dd, ne00, n_dims, nr, pos, freq_scale, ne01, freq_base, ext_factor,
+                attn_factor, corr_dims, freq_factors, main_stream
+            );
+        } else {
+            GGML_ABORT("fatal error");
+        }
+    } else {
+        if (src0->type == GGML_TYPE_F32) {
+            rope_norm_sycl(
+                (const float *)src0_dd, (float *)dst_dd, ne00, n_dims, nr, pos, freq_scale, ne01, freq_base, ext_factor,
+                attn_factor, corr_dims, freq_factors, main_stream
+            );
+        } else if (src0->type == GGML_TYPE_F16) {
+            rope_norm_sycl(
+                (const sycl::half *)src0_dd, (sycl::half *)dst_dd, ne00, n_dims, nr, pos, freq_scale, ne01, freq_base, ext_factor,
+                attn_factor, corr_dims, freq_factors, main_stream
+            );
+        } else {
+            GGML_ABORT("fatal error");
+        }
+    }
+
+    (void) src1;
+    (void) dst;
+    (void) src1_dd;
+}
diff --git a/src/whisper_cpp/ggml/src/ggml-sycl/rope.hpp b/src/whisper_cpp/ggml/src/ggml-sycl/rope.hpp
new file mode 100644
index 00000000..00354c31
--- /dev/null
+++ b/src/whisper_cpp/ggml/src/ggml-sycl/rope.hpp
@@ -0,0 +1,22 @@
+//
+// MIT license
+// Copyright (C) 2024 Intel Corporation
+// SPDX-License-Identifier: MIT
+//
+
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+//
+
+#ifndef GGML_SYCL_ROPE_HPP
+#define GGML_SYCL_ROPE_HPP
+
+#include "common.hpp"
+
+void ggml_sycl_op_rope(
+    ggml_backend_sycl_context & ctx, const ggml_tensor *src0, const ggml_tensor *src1, ggml_tensor *dst,
+    const float *src0_dd, const float *src1_dd, float *dst_dd, const queue_ptr &main_stream);
+
+#endif // GGML_SYCL_ROPE_HPP
diff --git a/src/whisper_cpp/ggml/src/ggml-sycl/softmax.cpp b/src/whisper_cpp/ggml/src/ggml-sycl/softmax.cpp
new file mode 100644
index 00000000..17a542e4
--- /dev/null
+++ b/src/whisper_cpp/ggml/src/ggml-sycl/softmax.cpp
@@ -0,0 +1,251 @@
+#include "norm.hpp"
+
+template <bool vals_smem, int ncols_template, int block_size_template>
+static void soft_max_f32(const float * x, const float * mask, float * dst, const int ncols_par,
+                         const int nrows_y, const float scale, const float max_bias, const float m0,
+                         const float m1, uint32_t n_head_log2, const sycl::nd_item<3> &item_ct1, float *buf) {
+    const int ncols = ncols_template == 0 ? ncols_par : ncols_template;
+
+    const int tid = item_ct1.get_local_id(2);
+    const int rowx = item_ct1.get_group(2);
+    const int rowy = rowx % nrows_y; // broadcast the mask (y) in the row dimension
+
+    const int block_size = block_size_template == 0 ? item_ct1.get_local_range(2) : block_size_template;
+
+    const int warp_id = item_ct1.get_local_id(2) / WARP_SIZE;
+    const int lane_id = item_ct1.get_local_id(2) % WARP_SIZE;
+    const int nthreads = block_size;
+    const int nwarps = nthreads / WARP_SIZE;
+    int nreduce = nwarps / WARP_SIZE;
+    float slope = 1.0f;
+
+    // ALiBi
+    if (max_bias > 0.0f) {
+        const uint32_t h = rowx/nrows_y; // head index
+
+        const float base = h < n_head_log2 ? m0 : m1;
+        const int   exp  = h < n_head_log2 ? h + 1 : 2*(h - n_head_log2) + 1;
+
+        slope = sycl::pow(base, float(exp));
+    }
+
+    float *vals = vals_smem ? buf + std::max(nwarps, WARP_SIZE) : dst + rowx * ncols;
+    float max_val = -INFINITY;
+
+    for (int col0 = 0; col0 < ncols; col0 += block_size) {
+        const int col = col0 + tid;
+
+        if (ncols_template == 0 && col >= ncols) {
+            break;
+        }
+
+        const int ix = rowx*ncols + col;
+        const int iy = rowy*ncols + col;
+
+        const float val = x[ix]*scale + (mask ? slope*mask[iy] : 0.0f);
+
+        vals[col] = val;
+        max_val = sycl::max(max_val, val);
+    }
+
+    // find the max value in the block
+    max_val = warp_reduce_max(max_val, item_ct1);
+    if (block_size > WARP_SIZE) {
+        if (warp_id == 0) {
+            buf[lane_id] = -INFINITY;
+            for (size_t i = 1; i < nreduce; i += 1)
+                buf[lane_id + i * WARP_SIZE] = -INFINITY;
+        }
+        item_ct1.barrier(sycl::access::fence_space::local_space);
+
+        if (lane_id == 0) {
+            buf[warp_id] = max_val;
+        }
+        item_ct1.barrier(sycl::access::fence_space::local_space);
+        max_val = buf[lane_id];
+        for (size_t i = 1; i < nreduce; i += 1)
+        {
+            max_val = std::max(max_val, buf[lane_id + i * WARP_SIZE]);
+        }
+        max_val = warp_reduce_max(max_val, item_ct1);
+    }
+
+    float tmp = 0.f;
+#pragma unroll
+    for (int col0 = 0; col0 < ncols; col0 += block_size) {
+        const int col = col0 + tid;
+                if (ncols_template == 0 && col >= ncols) {
+            break;
+        }
+
+        const float val = sycl::native::exp(vals[col] - max_val);
+        tmp += val;
+        vals[col] = val;
+    }
+
+    // find the sum of exps in the block
+    tmp = warp_reduce_sum(tmp, item_ct1);
+    if (block_size > WARP_SIZE) {
+        item_ct1.barrier(sycl::access::fence_space::local_space);
+        if (warp_id == 0) {
+            buf[lane_id] = 0.f;
+            for (size_t i = 1; i < nreduce; i += 1)
+                buf[lane_id + i * WARP_SIZE] = 0.f;
+        }
+        item_ct1.barrier(sycl::access::fence_space::local_space);
+
+        if (lane_id == 0) {
+            buf[warp_id] = tmp;
+        }
+        item_ct1.barrier(sycl::access::fence_space::local_space);
+
+        tmp = buf[lane_id];
+        for (size_t i = 1; i < nreduce; i += 1)
+        {
+            tmp += buf[lane_id + i * WARP_SIZE];
+        }
+        tmp = warp_reduce_sum(tmp, item_ct1);
+    }
+
+    const float inv_sum = 1.f / tmp;
+
+#pragma unroll
+    for (int col0 = 0; col0 < ncols; col0 += block_size) {
+        const int col = col0 + tid;
+
+        if (ncols_template == 0 && col >= ncols) {
+            return;
+        }
+
+        const int idst = rowx*ncols + col;
+        dst[idst] = vals[col] * inv_sum;
+    }
+}
+
+template <bool vals_smem, int ncols_template, int block_size_template>
+static void soft_max_f32_submitter(const float * x, const float * mask, float * dst, const int ncols_par,
+                                   const int nrows_y, const float scale, const float max_bias, const float m0,
+                                   const float m1, uint32_t n_head_log2, sycl::range<3> block_nums, sycl::range<3> block_dims,
+                                   const size_t n_local_scratch, queue_ptr stream) {
+    stream->submit([&](sycl::handler &cgh) {
+        sycl::local_accessor<float, 1> local_buf_acc(n_local_scratch, cgh);
+
+        cgh.parallel_for(
+            sycl::nd_range<3>(block_nums * block_dims, block_dims),
+            [=](sycl::nd_item<3> item_ct1) [[intel::reqd_sub_group_size(WARP_SIZE)]] {
+                soft_max_f32<vals_smem, ncols_template, block_size_template>(x, mask, dst, ncols_par,
+                                                                             nrows_y, scale, max_bias, m0,
+                                                                             m1, n_head_log2, item_ct1,
+                                                                             get_pointer(local_buf_acc));
+            });
+    });
+}
+
+static void soft_max_f32_sycl(const float * x, const float * mask,
+                              float * dst, const int ncols_x, const int nrows_x,
+                              const int nrows_y, const float scale, const float max_bias,
+                              queue_ptr stream, int device) {
+    int nth = WARP_SIZE;
+    int max_block_size = ggml_sycl_info().max_work_group_sizes[device];
+    while (nth < ncols_x && nth < max_block_size) nth *= 2;
+    if (nth>max_block_size) nth = max_block_size;
+
+    const sycl::range<3> block_dims(1, 1, nth);
+    const sycl::range<3> block_nums(1, 1, nrows_x);
+    const size_t n_val_tmp = nth / WARP_SIZE;
+    const size_t n_local_scratch = (GGML_PAD(ncols_x, WARP_SIZE) + n_val_tmp);
+
+    const uint32_t n_head_kv   = nrows_x/nrows_y;
+    const uint32_t n_head_log2 = 1u << (uint32_t) floorf(log2f((float) n_head_kv));
+
+    const float m0 = powf(2.0f, -(max_bias       ) / n_head_log2);
+    const float m1 = powf(2.0f, -(max_bias / 2.0f) / n_head_log2);
+
+    const size_t local_mem_size = stream->get_device().get_info<sycl::info::device::local_mem_size>();
+    if (n_local_scratch*sizeof(float) < local_mem_size) {
+        if (ncols_x > max_block_size) {
+            soft_max_f32_submitter<true, 0, 0>(x, mask, dst, ncols_x, nrows_y, scale,
+                                               max_bias, m0, m1, n_head_log2, block_nums,
+                                               block_dims, n_local_scratch, stream);
+            return;
+        }
+        switch (ncols_x) {
+            case 32:
+                soft_max_f32_submitter<true, 32, 32>(x, mask, dst, ncols_x, nrows_y, scale,
+                                                     max_bias, m0, m1, n_head_log2, block_nums,
+                                                     block_dims, n_local_scratch, stream);
+                break;
+            case 64:
+                soft_max_f32_submitter<true, 64, 64>(x, mask, dst, ncols_x, nrows_y, scale,
+                                                     max_bias, m0, m1, n_head_log2, block_nums,
+                                                     block_dims, n_local_scratch, stream);
+                break;
+            case 128:
+                soft_max_f32_submitter<true, 128, 128>(x, mask, dst, ncols_x, nrows_y, scale,
+                                                       max_bias, m0, m1, n_head_log2, block_nums,
+                                                       block_dims, n_local_scratch, stream);
+                break;
+            case 256:
+                soft_max_f32_submitter<true, 256, 256>(x, mask, dst, ncols_x, nrows_y, scale,
+                                                       max_bias, m0, m1, n_head_log2, block_nums,
+                                                       block_dims, n_local_scratch, stream);
+                break;
+            case 512:
+                soft_max_f32_submitter<true, 512, 512>(x, mask, dst, ncols_x, nrows_y, scale,
+                                                       max_bias, m0, m1, n_head_log2, block_nums,
+                                                       block_dims, n_local_scratch, stream);
+                break;
+            case 1024:
+                soft_max_f32_submitter<true, 1024, 1024>(x, mask, dst, ncols_x, nrows_y, scale,
+                                                         max_bias, m0, m1, n_head_log2, block_nums,
+                                                         block_dims, n_local_scratch, stream);
+                break;
+            case 2048:
+                soft_max_f32_submitter<true, 2048, 1024>(x, mask, dst, ncols_x, nrows_y, scale,
+                                                         max_bias, m0, m1, n_head_log2, block_nums,
+                                                         block_dims, n_local_scratch, stream);
+                break;
+            case 4096:
+                soft_max_f32_submitter<true, 4096, 1024>(x, mask, dst, ncols_x, nrows_y, scale,
+                                                         max_bias, m0, m1, n_head_log2, block_nums,
+                                                         block_dims, n_local_scratch, stream);
+                break;
+            default:
+                soft_max_f32_submitter<true, 0, 0>(x, mask, dst, ncols_x, nrows_y, scale,
+                                                   max_bias, m0, m1, n_head_log2, block_nums,
+                                                   block_dims, n_local_scratch, stream);
+                break;
+        }
+    } else {
+        soft_max_f32_submitter<false, 0, 0>(x, mask, dst, ncols_x, nrows_y, scale,
+                                            max_bias, m0, m1, n_head_log2, block_nums,
+                                            block_dims, WARP_SIZE, stream);
+    }
+}
+
+void ggml_sycl_op_soft_max(ggml_backend_sycl_context & ctx, const ggml_tensor *src0,
+                                  const ggml_tensor *src1, ggml_tensor *dst,
+                                  const float *src0_dd, const float *src1_dd,
+                                  float *dst_dd,
+                                  const queue_ptr &main_stream) {
+
+    GGML_ASSERT(src0->type == GGML_TYPE_F32);
+    GGML_ASSERT( dst->type == GGML_TYPE_F32);
+
+#pragma message("TODO: add ggml_sycl_op_soft_max() F16 src1 support")
+#pragma message("ref:  https://github.com/ggerganov/llama.cpp/pull/5021")
+    GGML_ASSERT(!src1 || src1->type == GGML_TYPE_F32); // src1 contains mask and it is optional
+
+    const int64_t ne00 = src0->ne[0];
+    const int64_t nrows_x = ggml_nrows(src0);
+    const int64_t nrows_y = src0->ne[1];
+
+    float scale = 1.0f;
+    float max_bias = 0.0f;
+
+    memcpy(&scale, dst->op_params + 0, sizeof(float));
+    memcpy(&max_bias, dst->op_params + 1, sizeof(float));
+
+    soft_max_f32_sycl(src0_dd, src1 ? src1_dd : nullptr, dst_dd, ne00,
+        nrows_x, nrows_y, scale, max_bias, main_stream, ctx.device);
+}
diff --git a/src/whisper_cpp/ggml/src/ggml-sycl/softmax.hpp b/src/whisper_cpp/ggml/src/ggml-sycl/softmax.hpp
new file mode 100644
index 00000000..bdb8f712
--- /dev/null
+++ b/src/whisper_cpp/ggml/src/ggml-sycl/softmax.hpp
@@ -0,0 +1,24 @@
+//
+// MIT license
+// Copyright (C) 2024 Intel Corporation
+// SPDX-License-Identifier: MIT
+//
+
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+//
+
+#ifndef GGML_SYCL_SOFTMAX_HPP
+#define GGML_SYCL_SOFTMAX_HPP
+
+#include "common.hpp"
+
+void ggml_sycl_op_soft_max(ggml_backend_sycl_context &ctx, const ggml_tensor *src0,
+    const ggml_tensor *src1, ggml_tensor *dst,
+    const float *src0_dd, const float *src1_dd,
+    float *dst_dd,
+    const queue_ptr &main_stream);
+
+#endif // GGML_SYCL_SOFTMAX_HPP
diff --git a/src/whisper_cpp/ggml/src/ggml-sycl/tsembd.cpp b/src/whisper_cpp/ggml/src/ggml-sycl/tsembd.cpp
new file mode 100644
index 00000000..d5c227cd
--- /dev/null
+++ b/src/whisper_cpp/ggml/src/ggml-sycl/tsembd.cpp
@@ -0,0 +1,71 @@
+//
+// MIT license
+// Copyright (C) 2024 Intel Corporation
+// SPDX-License-Identifier: MIT
+//
+
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+//
+
+#include "tsembd.hpp"
+
+static void timestep_embedding_f32(
+        const float * timesteps, float * dst, const int nb1,
+        const int dim, const int max_period, const sycl::nd_item<3> &item_ct1) {
+    // item_ct1.get_group(1)(blockIDx.y): idx of timesteps->ne[0]
+    // item_ct1.get_group(2) (blockIDx.x): idx of ((dim + 1) / 2) / BLOCK_SIZE
+    int i = item_ct1.get_group(1);
+    int j = item_ct1.get_local_id(2) + item_ct1.get_group(2) * item_ct1.get_local_range(2);
+    float * embed_data = (float *)((char *)dst +  i*nb1);
+
+    if (dim % 2 != 0 && j == ((dim + 1) / 2)) {
+        embed_data[dim] = 0.f;
+    }
+
+    int half = dim / 2;
+    if (j >= half) {
+        return;
+    }
+
+    float timestep = timesteps[i];
+    float freq = (float)sycl::native::exp(-(sycl::log((float)max_period)) * j / half);
+    float arg = timestep * freq;
+    embed_data[j] = sycl::cos(arg);
+    embed_data[j + half] = sycl::sin(arg);
+}
+
+static void timestep_embedding_f32_sycl(
+        const float * x, float * dst, const int ne00, const int nb1,
+        const int dim, const int max_period, const queue_ptr& stream) {
+    // As the kernel returns when thread.idx is larger than dim/2, the half_ceil does not need to pad
+    int half_ceil = dim / 2;
+    int num_blocks = (half_ceil + SYCL_TIMESTEP_EMBEDDING_BLOCK_SIZE - 1) / SYCL_TIMESTEP_EMBEDDING_BLOCK_SIZE;
+    sycl::range<3> block_dims(1, 1, SYCL_TIMESTEP_EMBEDDING_BLOCK_SIZE);
+    sycl::range<3> gridDim(1, ne00, num_blocks);
+    stream->parallel_for(
+        sycl::nd_range<3>(
+            gridDim * block_dims, block_dims),
+        [=](sycl::nd_item<3> item_ct1) {
+            timestep_embedding_f32(
+                x, dst, nb1, dim, max_period, item_ct1
+            );
+        });
+}
+
+void ggml_sycl_op_timestep_embedding(ggml_backend_sycl_context & ctx, const ggml_tensor *src0,
+    const ggml_tensor *src1, ggml_tensor * dst) {
+    const float * src0_d = (const float *)src0->data;
+    float * dst_d = (float *)dst->data;
+    dpct::queue_ptr stream = ctx.stream();
+
+    GGML_ASSERT(src0->type == GGML_TYPE_F32);
+    GGML_ASSERT(dst->type == GGML_TYPE_F32);
+
+    const int dim = dst->op_params[0];
+    const int max_period = dst->op_params[1];
+
+    timestep_embedding_f32_sycl(src0_d, dst_d, src0->ne[0], dst->nb[1], dim, max_period, stream);
+}
diff --git a/src/whisper_cpp/ggml/src/ggml-sycl/tsembd.hpp b/src/whisper_cpp/ggml/src/ggml-sycl/tsembd.hpp
new file mode 100644
index 00000000..ff854c33
--- /dev/null
+++ b/src/whisper_cpp/ggml/src/ggml-sycl/tsembd.hpp
@@ -0,0 +1,21 @@
+//
+// MIT license
+// Copyright (C) 2024 Intel Corporation
+// SPDX-License-Identifier: MIT
+//
+
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+//
+
+#ifndef GGML_SYCL_TSEMBD_HPP
+#define GGML_SYCL_TSEMBD_HPP
+
+#include "common.hpp"
+
+void ggml_sycl_op_timestep_embedding(ggml_backend_sycl_context & ctx, const ggml_tensor *src0,
+    const ggml_tensor *src1, ggml_tensor * dst);
+
+#endif // GGML_SYCL_TSEMBD_HPP
diff --git a/src/whisper_cpp/ggml/src/ggml-sycl/vecdotq.hpp b/src/whisper_cpp/ggml/src/ggml-sycl/vecdotq.hpp
new file mode 100644
index 00000000..d2dccade
--- /dev/null
+++ b/src/whisper_cpp/ggml/src/ggml-sycl/vecdotq.hpp
@@ -0,0 +1,1140 @@
+//
+// MIT license
+// Copyright (C) 2024 Intel Corporation
+// SPDX-License-Identifier: MIT
+//
+
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+//
+
+#ifndef GGML_SYCL_VECDOTQ_HPP
+#define GGML_SYCL_VECDOTQ_HPP
+
+#include "dpct/helper.hpp"
+
+typedef float (*vec_dot_q_sycl_t)(const void * __restrict__ vbq, const block_q8_1 * __restrict__ bq8_1, const int & iqs);
+
+static __dpct_inline__ int get_int_from_int8(const int8_t* x8, const int& i32) {
+  const uint16_t* x16 =
+      (const uint16_t*)(x8 + sizeof(int) * i32); // assume at least 2 byte
+                                                 // alignment
+
+  int x32 = 0;
+  x32 |= x16[0] << 0;
+  x32 |= x16[1] << 16;
+
+  return x32;
+}
+
+static __dpct_inline__ int get_int_from_uint8(
+    const uint8_t* x8,
+    const int& i32) {
+  const uint16_t* x16 =
+      (const uint16_t*)(x8 + sizeof(int) * i32); // assume at least 2 byte
+                                                 // alignment
+
+  int x32 = 0;
+  x32 |= x16[0] << 0;
+  x32 |= x16[1] << 16;
+
+  return x32;
+}
+
+static __dpct_inline__ int get_int_from_int8_aligned(
+    const int8_t* x8,
+    const int& i32) {
+  return *(
+      (const int*)(x8 + sizeof(int) * i32)); // assume at least 4 byte alignment
+}
+
+static __dpct_inline__ int get_int_from_uint8_aligned(
+    const uint8_t* x8,
+    const int& i32) {
+  return *(
+      (const int*)(x8 + sizeof(int) * i32)); // assume at least 4 byte alignment
+}
+
+static __dpct_inline__ void get_int_from_table_16(const uint32_t &q4,
+                                                  const uint8_t *values,
+                                                  int &val1, int &val2) {
+
+    uint32_t aux32; const uint8_t * q8 = (const uint8_t *)&aux32;
+    aux32 = q4 & 0x0f0f0f0f;
+    uint16_t v1 = values[q8[0]] | (values[q8[1]] << 8);
+    uint16_t v2 = values[q8[2]] | (values[q8[3]] << 8);
+    val1 = v1 | (v2 << 16);
+    aux32 = (q4 >> 4) & 0x0f0f0f0f;
+    v1 = values[q8[0]] | (values[q8[1]] << 8);
+    v2 = values[q8[2]] | (values[q8[3]] << 8);
+    val2 = v1 | (v2 << 16);
+}
+
+#define VDR_Q2_K_Q8_1_MMVQ 1
+
+// contiguous v/x values
+static __dpct_inline__ float vec_dot_q2_K_q8_1_impl_mmvq(
+    const int &v, const int *__restrict__ u, const uint8_t *__restrict__ scales,
+    const sycl::half2 &dm2, const float *__restrict__ d8) {
+
+    float sumf_d = 0.0f;
+    float sumf_m = 0.0f;
+
+#pragma unroll
+    for (int i = 0; i < QR2_K; ++i) {
+        const int sc = scales[2*i];
+
+        const int vi = (v >> (2*i)) & 0x03030303;
+
+        sumf_d +=
+            d8[i] * (dpct::dp4a(vi, u[i], 0) * (sc & 0xF)); // SIMD dot product
+
+        // fill int with 4x m
+        int m = sc >> 4;
+        m |= m <<  8;
+        m |= m << 16;
+        sumf_m += d8[i] *
+                  dpct::dp4a(
+                      m, u[i],
+                      0); // multiply constant q2_K part with sum of q8_1 values
+    }
+
+    const sycl::float2 dm2f =
+        dm2.convert<float, sycl::rounding_mode::automatic>();
+
+    return dm2f.x() * sumf_d - dm2f.y() * sumf_m;
+}
+
+
+#define VDR_Q3_K_Q8_1_MMVQ 1
+
+// contiguous v/x values
+static __dpct_inline__ float vec_dot_q3_K_q8_1_impl_mmvq(
+    const int &vl, const int &vh, const int *__restrict__ u,
+    const uint8_t *__restrict__ scales, const int &scale_offset,
+    const float &d3, const float *__restrict__ d8) {
+
+    float sumf = 0.0f;
+
+#pragma unroll
+    for (int i = 0; i < QR3_K; ++i) {
+        const int isc = scale_offset + 2*i;
+
+        const int isc_low = isc % (QK_K/32);
+        const int sc_shift_low = 4 * (isc / (QK_K/32));
+        const int sc_low  = (scales[isc_low] >> sc_shift_low) & 0xF;
+
+        const int isc_high = isc % (QK_K/64);
+        const int sc_shift_high = 2 * (isc / (QK_K/64));
+        const int sc_high = ((scales[(QK_K/32) + isc_high] >> sc_shift_high) & 3) << 4;
+
+        const int sc = (sc_low | sc_high) - 32;
+
+        const int vil = (vl >> (2*i)) & 0x03030303;
+
+        const int vih = ((vh >> i) << 2) & 0x04040404;
+
+        const int vi =
+            dpct::vectorized_binary<sycl::char4>(vil, vih, dpct::sub_sat());
+
+        sumf += d8[i] * (dpct::dp4a(vi, u[i], 0) * sc); // SIMD dot product
+    }
+
+    return d3 * sumf;
+}
+
+#define VDR_Q4_K_Q8_1_MMVQ 2
+
+// contiguous v/x values
+static __dpct_inline__ float vec_dot_q4_K_q8_1_impl_vmmq(
+    const int *__restrict__ v, const int *__restrict__ u,
+    const uint8_t *__restrict__ sc, const uint8_t *__restrict__ m,
+    const sycl::half2 &dm4, const float *__restrict__ d8) {
+
+    float sumf_d = 0.0f;
+    float sumf_m = 0.0f;
+
+#pragma unroll
+    for (int i = 0; i < QR4_K; ++i) {
+        const int v0i = (v[0] >> (4*i)) & 0x0F0F0F0F;
+        const int v1i = (v[1] >> (4*i)) & 0x0F0F0F0F;
+
+        const int dot1 =
+            dpct::dp4a(v1i, u[2 * i + 1],
+                       dpct::dp4a(v0i, u[2 * i + 0], 0)); // SIMD dot product
+        const int dot2 =
+            dpct::dp4a(0x01010101, u[2 * i + 1],
+                       dpct::dp4a(0x01010101, u[2 * i + 0], 0)); // sum of u
+
+        sumf_d += d8[i] * (dot1 * sc[i]);
+        sumf_m += d8[i] * (dot2 * m[i]);  // multiply constant part of q4_K with sum of q8_1 values
+    }
+
+    const sycl::float2 dm4f =
+        dm4.convert<float, sycl::rounding_mode::automatic>();
+
+    return dm4f.x() * sumf_d - dm4f.y() * sumf_m;
+}
+
+
+#define VDR_Q5_K_Q8_1_MMVQ 2
+
+// contiguous v/x values
+static __dpct_inline__ float vec_dot_q5_K_q8_1_impl_vmmq(
+    const int *__restrict__ vl, const int *__restrict__ vh,
+    const int *__restrict__ u, const uint8_t *__restrict__ sc,
+    const uint8_t *__restrict__ m, const sycl::half2 &dm5,
+    const float *__restrict__ d8) {
+
+    float sumf_d = 0.0f;
+    float sumf_m = 0.0f;
+
+#pragma unroll
+    for (int i = 0; i < QR5_K; ++i) {
+        const int vl0i = (vl[0] >> (4*i)) & 0x0F0F0F0F;
+        const int vl1i = (vl[1] >> (4*i)) & 0x0F0F0F0F;
+
+        const int vh0i = ((vh[0] >> i) << 4) & 0x10101010;
+        const int vh1i = ((vh[1] >> i) << 4) & 0x10101010;
+
+        const int v0i = vl0i | vh0i;
+        const int v1i = vl1i | vh1i;
+
+        const int dot1 =
+            dpct::dp4a(v0i, u[2 * i + 0],
+                       dpct::dp4a(v1i, u[2 * i + 1], 0)); // SIMD dot product
+        const int dot2 =
+            dpct::dp4a(0x01010101, u[2 * i + 0],
+                       dpct::dp4a(0x01010101, u[2 * i + 1], 0)); // sum of u
+
+        sumf_d += d8[i] * (dot1 * sc[i]);
+        sumf_m += d8[i] * (dot2 * m[i]);
+
+    }
+
+    const sycl::float2 dm5f =
+        dm5.convert<float, sycl::rounding_mode::automatic>();
+
+    return dm5f.x() * sumf_d - dm5f.y() * sumf_m;
+}
+
+
+#define VDR_Q6_K_Q8_1_MMVQ 1
+
+// contiguous v/x values
+static __dpct_inline__ float
+vec_dot_q6_K_q8_1_impl_mmvq(const int &vl, const int &vh,
+                            const int *__restrict__ u,
+                            const int8_t *__restrict__ scales, const float &d,
+                            const float *__restrict__ d8) {
+
+    float sumf = 0.0f;
+
+#pragma unroll
+    for (int i = 0; i < QR6_K; ++i) {
+        const int sc = scales[4*i];
+
+        const int vil = (vl >> (4*i)) & 0x0F0F0F0F;
+
+        const int vih = ((vh >> (4*i)) << 4) & 0x30303030;
+
+        const int vi = dpct::vectorized_binary<sycl::char4>(
+            (vil | vih), 0x20202020, dpct::sub_sat()); // vi = (vil | vih) - 32
+
+        sumf += d8[i] * (dpct::dp4a(vi, u[i], 0) * sc); // SIMD dot product
+    }
+
+    return d*sumf;
+}
+
+// VDR = vec dot ratio, how many contiguous integers each thread processes when the vec dot kernel is called
+// MMVQ = mul_mat_vec_q, MMQ = mul_mat_q
+
+#define VDR_Q4_0_Q8_1_MMVQ 2
+#define VDR_Q4_0_Q8_1_MMQ  4
+
+template <int vdr>
+static __dpct_inline__ float vec_dot_q4_0_q8_1_impl(const int *v, const int *u,
+                                                    const float &d4,
+                                                    const sycl::half2 &ds8) {
+    int sumi = 0;
+#pragma unroll
+    for (int i = 0; i < vdr; ++i) {
+        const int vi0 = (v[i] >> 0) & 0x0F0F0F0F;
+        const int vi1 = (v[i] >> 4) & 0x0F0F0F0F;
+
+        // SIMD dot product of quantized values
+        sumi = dpct::dp4a(vi0, u[2 * i + 0], sumi);
+        sumi = dpct::dp4a(vi1, u[2 * i + 1], sumi);
+    }
+
+    const sycl::float2 ds8f =
+        ds8.convert<float, sycl::rounding_mode::automatic>();
+
+    // second part effectively subtracts 8 from each quant value
+    return d4 * (sumi * ds8f.x() - (8 * vdr / QI4_0) * ds8f.y());
+}
+
+#define VDR_Q4_1_Q8_1_MMVQ 2
+#define VDR_Q4_1_Q8_1_MMQ  4
+
+template <int vdr>
+static __dpct_inline__ float vec_dot_q4_1_q8_1_impl(const int *v, const int *u,
+                                                    const sycl::half2 &dm4,
+                                                    const sycl::half2 &ds8) {
+
+    int sumi = 0;
+
+#pragma unroll
+    for (int i = 0; i < vdr; ++i) {
+        const int vi0 = (v[i] >> 0) & 0x0F0F0F0F;
+        const int vi1 = (v[i] >> 4) & 0x0F0F0F0F;
+
+        // SIMD dot product of quantized values
+        sumi = dpct::dp4a(vi0, u[2 * i + 0], sumi);
+        sumi = dpct::dp4a(vi1, u[2 * i + 1], sumi);
+    }
+
+#ifdef GGML_SYCL_F16
+    const sycl::float2 tmp =
+        (dm4 * ds8).convert<float, sycl::rounding_mode::automatic>();
+    const float d4d8 = tmp.x();
+    const float m4s8 = tmp.y();
+#else
+    const sycl::float2 dm4f =
+        dm4.convert<float, sycl::rounding_mode::automatic>();
+    const sycl::float2 ds8f =
+        ds8.convert<float, sycl::rounding_mode::automatic>();
+    const float d4d8 = dm4f.x() * ds8f.x();
+    const float m4s8 = dm4f.y() * ds8f.y();
+#endif // GGML_SYCL_F16
+
+    // scale second part of sum by QI8_1/(vdr * QR4_1) to compensate for multiple threads adding it
+    return sumi * d4d8 + m4s8 / (QI8_1 / (vdr * QR4_1));
+}
+
+#define VDR_Q5_0_Q8_1_MMVQ 2
+#define VDR_Q5_0_Q8_1_MMQ  4
+
+template <int vdr>
+static __dpct_inline__ float
+vec_dot_q5_0_q8_1_impl(const int *vl, const int *vh, const int *u,
+                       const float &d5, const sycl::half2 &ds8) {
+    int sumi = 0;
+
+#pragma unroll
+    for (int i = 0; i < vdr; ++i) {
+        int vi0 = (vl[i] >>  0) & 0x0F0F0F0F; // lower 4 qs bits, still need qh as 5th bits
+        vi0    |= (vh[i] <<  4) & 0x00000010; // 0 ->  4
+        vi0    |= (vh[i] << 11) & 0x00001000; // 1 -> 12
+        vi0    |= (vh[i] << 18) & 0x00100000; // 2 -> 20
+        vi0    |= (vh[i] << 25) & 0x10000000; // 3 -> 28
+        sumi = dpct::dp4a(vi0, u[2 * i + 0],
+                          sumi); // SIMD dot product of quantized values
+
+        int vi1 = (vl[i] >>  4) & 0x0F0F0F0F; // upper 4 qs bits, still need qh as 5th bits
+        vi1    |= (vh[i] >> 12) & 0x00000010; // 16 ->  4
+        vi1    |= (vh[i] >>  5) & 0x00001000; // 17 -> 12
+        vi1    |= (vh[i] <<  2) & 0x00100000; // 18 -> 20
+        vi1    |= (vh[i] <<  9) & 0x10000000; // 19 -> 28
+        sumi = dpct::dp4a(vi1, u[2 * i + 1],
+                          sumi); // SIMD dot product of quantized values
+    }
+
+    const sycl::float2 ds8f =
+        ds8.convert<float, sycl::rounding_mode::automatic>();
+
+    // second part effectively subtracts 16 from each quant value
+    return d5 * (sumi * ds8f.x() - (16 * vdr / QI5_0) * ds8f.y());
+}
+
+#define VDR_Q5_1_Q8_1_MMVQ 2
+#define VDR_Q5_1_Q8_1_MMQ  4
+
+template <int vdr>
+static __dpct_inline__ float
+vec_dot_q5_1_q8_1_impl(const int *vl, const int *vh, const int *u,
+                       const sycl::half2 &dm5, const sycl::half2 &ds8) {
+
+    int sumi = 0;
+
+#pragma unroll
+    for (int i = 0; i < vdr; ++i) {
+        int vi0 = (vl[i] >>  0) & 0x0F0F0F0F; // lower 4 qs bits, still need qh as 5th bits
+        vi0    |= (vh[i] <<  4) & 0x00000010; // 0 ->  4
+        vi0    |= (vh[i] << 11) & 0x00001000; // 1 -> 12
+        vi0    |= (vh[i] << 18) & 0x00100000; // 2 -> 20
+        vi0    |= (vh[i] << 25) & 0x10000000; // 3 -> 28
+        sumi = dpct::dp4a(vi0, u[2 * i + 0],
+                          sumi); // SIMD dot product of quantized values
+
+        int vi1 = (vl[i] >>  4) & 0x0F0F0F0F; // upper 4 qs bits, still need qh as 5th bits
+        vi1    |= (vh[i] >> 12) & 0x00000010; // 16 ->  4
+        vi1    |= (vh[i] >>  5) & 0x00001000; // 17 -> 12
+        vi1    |= (vh[i] <<  2) & 0x00100000; // 18 -> 20
+        vi1    |= (vh[i] <<  9) & 0x10000000; // 19 -> 28
+        sumi = dpct::dp4a(vi1, u[2 * i + 1],
+                          sumi); // SIMD dot product of quantized values
+    }
+
+#ifdef GGML_SYCL_F16
+     const sycl::float2 tmp =
+        (dm5 * ds8).convert<float, sycl::rounding_mode::automatic>();
+    const float d5d8 = tmp.x();
+    const float m5s8 = tmp.y();
+
+
+#else
+    const sycl::float2 dm5f =
+        dm5.convert<float, sycl::rounding_mode::automatic>();
+    const sycl::float2 ds8f =
+        ds8.convert<float, sycl::rounding_mode::automatic>();
+    const float d5d8 = dm5f.x() * ds8f.x();
+    const float m5s8 = dm5f.y() * ds8f.y();
+#endif // GGML_SYCL_F16
+
+    // scale second part of sum by QI5_1 / vdr to compensate for multiple threads adding it
+    return sumi*d5d8 + m5s8 / (QI5_1 / vdr);
+}
+
+#define VDR_Q8_0_Q8_1_MMVQ 2
+#define VDR_Q8_0_Q8_1_MMQ 8
+
+template <int vdr>
+static __dpct_inline__ float vec_dot_q8_0_q8_1_impl(const int *v, const int *u,
+                                                    const float &d8_0,
+                                                    const float &d8_1) {
+
+    int sumi = 0;
+
+#pragma unroll
+    for (int i = 0; i < vdr; ++i) {
+        // SIMD dot product of quantized values
+        sumi = dpct::dp4a(v[i], u[i], sumi);
+    }
+
+    return d8_0*d8_1 * sumi;
+}
+
+template <int vdr>
+static __dpct_inline__ float vec_dot_q8_1_q8_1_impl(const int *v, const int *u,
+                                                    const sycl::half2 &dm8,
+                                                    const sycl::half2 &ds8) {
+
+    int sumi = 0;
+
+#pragma unroll
+    for (int i = 0; i < vdr; ++i) {
+        // SIMD dot product of quantized values
+        sumi = dpct::dp4a(v[i], u[i], sumi);
+    }
+
+#ifdef GGML_SYCL_F16
+    const sycl::float2 tmp =
+        (dm8 * ds8).convert<float, sycl::rounding_mode::automatic>();
+    const float d8d8 = tmp.x();
+    const float m8s8 = tmp.y();
+#else
+    const sycl::float2 dm8f =
+        dm8.convert<float, sycl::rounding_mode::automatic>();
+    const sycl::float2 ds8f =
+        ds8.convert<float, sycl::rounding_mode::automatic>();
+    const float d8d8 = dm8f.x() * ds8f.x();
+    const float m8s8 = dm8f.y() * ds8f.y();
+#endif // GGML_SYCL_F16
+
+    // scale second part of sum by QI8_1/ vdr to compensate for multiple threads adding it
+    return sumi*d8d8 + m8s8 / (QI8_1 / vdr);
+}
+
+static __dpct_inline__ float
+vec_dot_q4_0_q8_1(const void *__restrict__ vbq,
+                  const block_q8_1 *__restrict__ bq8_1, const int &iqs) {
+
+    const block_q4_0 * bq4_0 = (const block_q4_0 *) vbq;
+
+    int v[VDR_Q4_0_Q8_1_MMVQ];
+    int u[2*VDR_Q4_0_Q8_1_MMVQ];
+
+#pragma unroll
+    for (int i = 0; i < VDR_Q4_0_Q8_1_MMVQ; ++i) {
+        v[i]     = get_int_from_uint8(bq4_0->qs, iqs + i);
+        u[2*i+0] = get_int_from_int8_aligned(bq8_1->qs, iqs + i);
+        u[2*i+1] = get_int_from_int8_aligned(bq8_1->qs, iqs + i + QI4_0);
+    }
+
+    return vec_dot_q4_0_q8_1_impl<VDR_Q4_0_Q8_1_MMVQ>(v, u, bq4_0->d, bq8_1->ds);
+}
+
+static __dpct_inline__ float
+vec_dot_q4_1_q8_1(const void *__restrict__ vbq,
+                  const block_q8_1 *__restrict__ bq8_1, const int &iqs) {
+
+    const block_q4_1 * bq4_1 = (const block_q4_1 *) vbq;
+
+    int v[VDR_Q4_1_Q8_1_MMVQ];
+    int u[2*VDR_Q4_1_Q8_1_MMVQ];
+
+#pragma unroll
+    for (int i = 0; i < VDR_Q4_1_Q8_1_MMVQ; ++i) {
+        v[i]    = get_int_from_uint8_aligned(bq4_1->qs, iqs + i);
+        u[2*i+0] = get_int_from_int8_aligned(bq8_1->qs, iqs + i);
+        u[2*i+1] = get_int_from_int8_aligned(bq8_1->qs, iqs + i + QI4_1);
+    }
+
+    return vec_dot_q4_1_q8_1_impl<VDR_Q4_1_Q8_1_MMVQ>(v, u, bq4_1->dm, bq8_1->ds);
+}
+
+static __dpct_inline__ float
+vec_dot_q5_0_q8_1(const void *__restrict__ vbq,
+                  const block_q8_1 *__restrict__ bq8_1, const int &iqs) {
+
+    const block_q5_0 * bq5_0 = (const block_q5_0 *) vbq;
+
+    int vl[VDR_Q5_0_Q8_1_MMVQ];
+    int vh[VDR_Q5_0_Q8_1_MMVQ];
+    int  u[2*VDR_Q5_0_Q8_1_MMVQ];
+
+#pragma unroll
+    for (int i = 0; i < VDR_Q5_0_Q8_1_MMVQ; ++i) {
+        vl[i]    = get_int_from_uint8(bq5_0->qs, iqs + i);
+        vh[i]    = get_int_from_uint8(bq5_0->qh, 0) >> (4 * (iqs + i));
+        u[2*i+0] = get_int_from_int8_aligned(bq8_1->qs, iqs + i);
+        u[2*i+1] = get_int_from_int8_aligned(bq8_1->qs, iqs + i + QI5_0);
+    }
+
+    return vec_dot_q5_0_q8_1_impl<VDR_Q5_0_Q8_1_MMVQ>(vl, vh, u, bq5_0->d, bq8_1->ds);
+}
+
+static __dpct_inline__ float
+vec_dot_q5_1_q8_1(const void *__restrict__ vbq,
+                  const block_q8_1 *__restrict__ bq8_1, const int &iqs) {
+
+    const block_q5_1 * bq5_1 = (const block_q5_1 *) vbq;
+
+    int vl[VDR_Q5_1_Q8_1_MMVQ];
+    int vh[VDR_Q5_1_Q8_1_MMVQ];
+    int  u[2*VDR_Q5_1_Q8_1_MMVQ];
+
+#pragma unroll
+    for (int i = 0; i < VDR_Q5_1_Q8_1_MMVQ; ++i) {
+        vl[i]   = get_int_from_uint8_aligned(bq5_1->qs, iqs + i);
+        vh[i]   = get_int_from_uint8_aligned(bq5_1->qh, 0) >> (4 * (iqs + i));
+        u[2*i+0] = get_int_from_int8_aligned(bq8_1->qs, iqs + i);
+        u[2*i+1] = get_int_from_int8_aligned(bq8_1->qs, iqs + i + QI5_1);
+    }
+
+    return vec_dot_q5_1_q8_1_impl<VDR_Q5_1_Q8_1_MMVQ>(vl, vh, u, bq5_1->dm, bq8_1->ds);
+}
+
+static __dpct_inline__ float
+vec_dot_q8_0_q8_1(const void *__restrict__ vbq,
+                  const block_q8_1 *__restrict__ bq8_1, const int &iqs) {
+
+    const block_q8_0 * bq8_0 = (const block_q8_0 *) vbq;
+
+    int v[VDR_Q8_0_Q8_1_MMVQ];
+    int u[VDR_Q8_0_Q8_1_MMVQ];
+
+#pragma unroll
+    for (int i = 0; i < VDR_Q8_0_Q8_1_MMVQ; ++i) {
+        v[i] = get_int_from_int8(bq8_0->qs, iqs + i);
+        u[i] = get_int_from_int8_aligned(bq8_1->qs, iqs + i);
+    }
+
+    return vec_dot_q8_0_q8_1_impl<VDR_Q8_0_Q8_1_MMVQ>(v, u, bq8_0->d,
+                                                      bq8_1->ds[0]);
+}
+
+static __dpct_inline__ float
+vec_dot_q2_K_q8_1(const void *__restrict__ vbq,
+                  const block_q8_1 *__restrict__ bq8_1, const int &iqs) {
+
+    const block_q2_K * bq2_K = (const block_q2_K *) vbq;
+
+    const int bq8_offset = QR2_K * (iqs / QI8_1);
+    const int scale_offset = iqs - iqs % QI8_1 + (iqs % QI8_1) / (QI8_1/2);
+
+    const uint8_t * scales = bq2_K->scales + scale_offset;
+
+    const int v = get_int_from_uint8_aligned(bq2_K->qs, iqs);
+    int    u[QR2_K];
+    float d8[QR2_K];
+
+#pragma unroll
+    for (int i = 0; i < QR2_K; ++ i) {
+        u[i]  = get_int_from_int8_aligned(bq8_1[bq8_offset + i].qs, iqs % QI8_1);
+        d8[i] = bq8_1[bq8_offset + i].ds[0];
+    }
+
+    return vec_dot_q2_K_q8_1_impl_mmvq(v, u, scales, bq2_K->dm, d8);
+}
+
+static __dpct_inline__ float
+vec_dot_q3_K_q8_1(const void *__restrict__ vbq,
+                  const block_q8_1 *__restrict__ bq8_1, const int &iqs) {
+
+    const block_q3_K * bq3_K = (const block_q3_K *) vbq;
+
+    const int bq8_offset = QR3_K * (iqs / (QI3_K/2));
+    const int scale_offset = iqs - iqs % QI8_1 + (iqs % QI8_1) / (QI8_1/2);
+
+    const float d = bq3_K->d;
+
+    const int vl = get_int_from_uint8(bq3_K->qs, iqs);
+
+    // invert the mask with ~ so that a 0/1 results in 4/0 being subtracted
+    const int vh = ~get_int_from_uint8(bq3_K->hmask, iqs % (QI3_K/2)) >> bq8_offset;
+
+    int    u[QR3_K];
+    float d8[QR3_K];
+
+#pragma unroll
+    for (int i = 0; i < QR3_K; ++i) {
+        u[i]  = get_int_from_int8_aligned(bq8_1[bq8_offset + i].qs, iqs % QI8_1);
+        d8[i] = bq8_1[bq8_offset + i].ds[0];
+    }
+
+    return vec_dot_q3_K_q8_1_impl_mmvq(vl, vh, u, bq3_K->scales, scale_offset, d, d8);
+}
+
+static __dpct_inline__ float
+vec_dot_q4_K_q8_1(const void *__restrict__ vbq,
+                  const block_q8_1 *__restrict__ bq8_1, const int &iqs) {
+
+#ifndef GGML_QKK_64
+    const block_q4_K * bq4_K = (const block_q4_K *) vbq;
+
+    int    v[2];
+    int    u[2*QR4_K];
+    float d8[QR4_K];
+
+    // iqs is in 0,2..30. bq8_offset = iqs/4 -> bq8_offset = 0, 2, 4, 6
+    const int bq8_offset = QR4_K * ((iqs/2) / (QI8_1/2));
+
+    // iqs = 0....3 -> bq8_offset = 0, want q4_offset = 0, 4, 8, 12
+    // iqs = 4....7 -> bq8_offset = 2, want q4_offset = 32, 36, 40, 44
+    // iqs = 8...11 -> bq8_offset = 4, want q4_offset = 64, 68, 72, 76
+    // iqs = 12..15 -> bq8_offset = 6, want q4_offset = 96, 100, 104, 108
+
+    const int * q4 = (const int *)(bq4_K->qs + 16 * bq8_offset + 4 * ((iqs/2)%4));
+    v[0] = q4[0];
+    v[1] = q4[4];
+
+    const uint16_t * scales = (const uint16_t *)bq4_K->scales;
+    uint16_t aux[2];
+    const int j = bq8_offset/2;
+    if (j < 2) {
+        aux[0] = scales[j+0] & 0x3f3f;
+        aux[1] = scales[j+2] & 0x3f3f;
+    } else {
+        aux[0] = ((scales[j+2] >> 0) & 0x0f0f) | ((scales[j-2] & 0xc0c0) >> 2);
+        aux[1] = ((scales[j+2] >> 4) & 0x0f0f) | ((scales[j-0] & 0xc0c0) >> 2);
+    }
+    const uint8_t * sc = (const uint8_t *)aux;
+    const uint8_t * m  = sc + 2;
+
+    for (int i = 0; i < QR4_K; ++i) {
+        const block_q8_1 * bq8i = bq8_1 + bq8_offset + i;
+        d8[i] = bq8i->ds[0];
+
+        const int * q8 = (const int *)bq8i->qs + ((iqs/2)%4);
+        u[2*i+0] = q8[0];
+        u[2*i+1] = q8[4];
+    }
+
+    return vec_dot_q4_K_q8_1_impl_vmmq(v, u, sc, m, bq4_K->dm, d8);
+
+#else
+
+#if __SYCL_ARCH__ >= VER_4VEC // lowest compute capability for integer intrinsics
+    const block_q4_K * bq4_K = (const block_q4_K *) vbq;
+
+    float sumf_d = 0.0f;
+    float sumf_m = 0.0f;
+
+    uint16_t aux16[2];
+    const uint8_t * s = (const uint8_t *)aux16;
+
+    const uint16_t * a = (const uint16_t *)bq4_K->scales;
+    aux16[0] = a[0] & 0x0f0f;
+    aux16[1] = (a[0] >> 4) & 0x0f0f;
+
+    const float dall = bq4_K->dm[0];
+    const float dmin = bq4_K->dm[1];
+
+    const float d8_1 = bq8_1[0].ds[0];
+    const float d8_2 = bq8_1[1].ds[1];
+
+    const int ui1 = *((const int *)bq8_1[0].qs + (iqs/2));
+    const int ui2 = *((const int *)bq8_1[0].qs + (iqs/2) + 4);
+    const int ui3 = *((const int *)bq8_1[1].qs + (iqs/2));
+    const int ui4 = *((const int *)bq8_1[1].qs + (iqs/2) + 4);
+
+    const int * q4 = (const int *)bq4_K->qs + (iqs/2);
+    const int v1 = q4[0];
+    const int v2 = q4[4];
+
+    const int dot1 = dpct::dp4a(ui2, v2 & 0x0f0f0f0f, dpct::dp4a(ui1, v1 & 0x0f0f0f0f, 0));
+    const int dot2 = dpct::dp4a(ui4, (v2 >> 4) & 0x0f0f0f0f, dpct::dp4a(ui3, (v1 >> 4) & 0x0f0f0f0f, 0));
+    const int dot3 = dpct::dp4a(0x01010101, ui2, dpct::dp4a(0x01010101, ui1, 0));
+    const int dot4 = dpct::dp4a(0x01010101, ui4, dpct::dp4a(0x01010101, ui3, 0));
+
+    sumf_d += d8_1 * (dot1 * s[0]) + d8_2 * (dot2 * s[1]);
+    sumf_m += d8_1 * (dot3 * s[2]) + d8_2 * (dot4 * s[3]);
+
+    return dall * sumf_d - dmin * sumf_m;
+
+#else
+    bad_arch();
+#endif // __SYCL_ARCH__ >= VER_4VEC
+
+#endif
+}
+
+static __dpct_inline__ float
+vec_dot_q5_K_q8_1(const void *__restrict__ vbq,
+                  const block_q8_1 *__restrict__ bq8_1, const int &iqs) {
+
+#ifndef GGML_QKK_64
+    const block_q5_K * bq5_K = (const block_q5_K *) vbq;
+
+    int   vl[2];
+    int   vh[2];
+    int    u[2*QR5_K];
+    float d8[QR5_K];
+
+    const int bq8_offset = QR5_K * ((iqs/2) / (QI8_1/2));
+    const int * ql = (const int *)(bq5_K->qs + 16 * bq8_offset + 4 * ((iqs/2)%4));
+    const int * qh = (const int *)(bq5_K->qh + 4 * ((iqs/2)%4));
+
+    vl[0] = ql[0];
+    vl[1] = ql[4];
+
+    vh[0] = qh[0] >> bq8_offset;
+    vh[1] = qh[4] >> bq8_offset;
+
+    const uint16_t * scales = (const uint16_t *)bq5_K->scales;
+    uint16_t aux[2];
+    const int j = bq8_offset/2;
+    if (j < 2) {
+        aux[0] = scales[j+0] & 0x3f3f;
+        aux[1] = scales[j+2] & 0x3f3f;
+    } else {
+        aux[0] = ((scales[j+2] >> 0) & 0x0f0f) | ((scales[j-2] & 0xc0c0) >> 2);
+        aux[1] = ((scales[j+2] >> 4) & 0x0f0f) | ((scales[j-0] & 0xc0c0) >> 2);
+    }
+    const uint8_t * sc = (const uint8_t *)aux;
+    const uint8_t * m  = sc + 2;
+
+#pragma unroll
+    for (int i = 0; i < QR5_K; ++i) {
+        const block_q8_1 * bq8i = bq8_1 + bq8_offset + i;
+        d8[i] = bq8i->ds[0];
+
+        const int * q8 = (const int *)bq8i->qs + ((iqs/2)%4);
+        u[2*i+0] = q8[0];
+        u[2*i+1] = q8[4];
+    }
+
+    return vec_dot_q5_K_q8_1_impl_vmmq(vl, vh, u, sc, m, bq5_K->dm, d8);
+
+#else
+
+#if __SYCL_ARCH__ >= VER_4VEC // lowest compute capability for integer intrinsics
+    const block_q5_K * bq5_K = (const block_q5_K *) vbq;
+
+    const int8_t * s = bq5_K->scales;
+
+    const float d = bq5_K->d;
+
+    const float d8_1 = bq8_1[0].ds[0];
+    const float d8_2 = bq8_1[1].ds[1];
+
+    const int ui1 = *((const int *)bq8_1[0].qs + (iqs/2));
+    const int ui2 = *((const int *)bq8_1[0].qs + (iqs/2) + 4);
+    const int ui3 = *((const int *)bq8_1[1].qs + (iqs/2));
+    const int ui4 = *((const int *)bq8_1[1].qs + (iqs/2) + 4);
+
+    const int * ql = (const int *)bq5_K->qs + (iqs/2);
+    const int vl1 = ql[0];
+    const int vl2 = ql[4];
+
+    const int step = 4 * (iqs/2); // 0, 4, 8, 12
+    const int im = step/8; // = 0 for iqs = 0, 2, = 1 for iqs = 4, 6
+    const int in = step%8; // 0, 4, 0, 4
+    const int vh = (*((const int *)(bq5_K->qh + in))) >> im;
+
+    const int v1 = (((vh << 4) & 0x10101010) ^ 0x10101010) | ((vl1 >> 0) & 0x0f0f0f0f);
+    const int v2 = (((vh << 2) & 0x10101010) ^ 0x10101010) | ((vl2 >> 0) & 0x0f0f0f0f);
+    const int v3 = (((vh >> 0) & 0x10101010) ^ 0x10101010) | ((vl1 >> 4) & 0x0f0f0f0f);
+    const int v4 = (((vh >> 2) & 0x10101010) ^ 0x10101010) | ((vl2 >> 4) & 0x0f0f0f0f);
+
+    const float sumf_d = d8_1 * (dpct::dp4a(ui1, v1, 0) * s[0] + dpct::dp4a(ui2, v2, 0) * s[1])
+                       + d8_2 * (dpct::dp4a(ui3, v3, 0) * s[2] + dpct::dp4a(ui4, v4, 0) * s[3]);
+
+    return d * sumf_d;
+
+#else
+    bad_arch();
+#endif // __SYCL_ARCH__ >= VER_4VEC
+
+#endif
+}
+
+static __dpct_inline__ float
+vec_dot_q6_K_q8_1(const void *__restrict__ vbq,
+                  const block_q8_1 *__restrict__ bq8_1, const int &iqs) {
+
+    const block_q6_K * bq6_K = (const block_q6_K *) vbq;
+
+    const int bq8_offset = 2 * QR6_K * (iqs / (QI6_K/2)) + (iqs % (QI6_K/2)) / (QI6_K/4);
+    const int scale_offset = (QI6_K/4) * (iqs / (QI6_K/2)) + (iqs % (QI6_K/2)) / (QI6_K/8);
+    const int vh_shift = 2 * ((iqs % (QI6_K/2)) / (QI6_K/4));
+
+    const int vl = get_int_from_uint8(bq6_K->ql, iqs);
+    const int vh = get_int_from_uint8(bq6_K->qh, (QI6_K/4) * (iqs / (QI6_K/2)) + iqs % (QI6_K/4)) >> vh_shift;
+
+    const int8_t * scales = bq6_K->scales + scale_offset;
+
+    int    u[QR6_K];
+    float d8[QR6_K];
+
+#pragma unroll
+    for (int i = 0; i < QR6_K; ++i) {
+        u[i]  = get_int_from_int8_aligned(bq8_1[bq8_offset + 2*i].qs, iqs % QI8_1);
+        d8[i] = bq8_1[bq8_offset + 2 * i].ds[0];
+    }
+
+    return vec_dot_q6_K_q8_1_impl_mmvq(vl, vh, u, scales, bq6_K->d, d8);
+}
+
+
+static __dpct_inline__ float
+vec_dot_iq2_xxs_q8_1(const void *__restrict__ vbq,
+                     const block_q8_1 *__restrict__ bq8_1, const int &iqs,
+                     const uint64_t *iq2xxs_grid, const uint8_t *ksigns_iq2xs,
+                     const uint8_t *kmask_iq2xs) {
+#if QK_K == 256
+    const block_iq2_xxs * bq2 = (const block_iq2_xxs *) vbq;
+
+    const int ib32 = iqs;
+    const uint16_t * q2 = bq2->qs + 4*ib32;
+    const uint8_t  * aux8 = (const uint8_t *)q2;
+    const int8_t   * q8 = bq8_1[ib32].qs;
+    uint32_t aux32 = q2[2] | (q2[3] << 16);
+    int sumi = 0;
+    for (int l = 0; l < 4; ++l) {
+        const uint8_t * grid = (const uint8_t *)(iq2xxs_grid + aux8[l]);
+        const uint8_t  signs = ksigns_iq2xs[aux32 & 127];
+        for (int j = 0; j < 8; ++j) {
+            sumi += q8[j] * grid[j] * (signs & kmask_iq2xs[j] ? -1 : 1);
+        }
+        q8 += 8;
+        aux32 >>= 7;
+    }
+    const float d = (float)bq2->d * (0.5f + aux32) * bq8_1[ib32].ds[0] * 0.25f;
+    return d * sumi;
+#else
+    assert(false);
+    return 0.f;
+#endif
+}
+
+static __dpct_inline__ float
+vec_dot_iq2_xs_q8_1(const void *__restrict__ vbq,
+                    const block_q8_1 *__restrict__ bq8_1, const int &iqs,
+                    const uint64_t *iq2xs_grid, const uint64_t *ksigns64) {
+#if DPCT_COMPATIBILITY_TEMP >=                                                 \
+    MIN_CC_DP4A // lowest compute capability for integer intrinsics
+#if QK_K == 256
+    const block_iq2_xs * bq2 = (const block_iq2_xs *) vbq;
+
+    const int ib32 = iqs;
+    const uint16_t * q2 = bq2->qs + 4*ib32;
+    const int8_t   * q8 = bq8_1[ib32].qs;
+    const uint8_t ls1 = bq2->scales[ib32] & 0xf;
+    const uint8_t ls2 = bq2->scales[ib32] >>  4;
+    int sumi1 = 0;
+    for (int l = 0; l < 2; ++l) {
+        const uint32_t * grid = (const uint32_t *)(iq2xs_grid + (q2[l] & 511));
+        const uint32_t * signs = (const uint32_t *)(ksigns64 + (q2[l] >> 9));
+        const int grid_l = dpct::vectorized_binary<sycl::uchar4>(
+            grid[0] ^ signs[0], signs[0], std::minus<>());
+        const int grid_h = dpct::vectorized_binary<sycl::uchar4>(
+            grid[1] ^ signs[1], signs[1], std::minus<>());
+        sumi1 = dpct::dp4a(grid_l, *((const int *)q8 + 0), sumi1);
+        sumi1 = dpct::dp4a(grid_h, *((const int *)q8 + 1), sumi1);
+        q8 += 8;
+    }
+    int sumi2 = 0;
+    for (int l = 2; l < 4; ++l) {
+        const uint32_t * grid = (const uint32_t *)(iq2xs_grid + (q2[l] & 511));
+        const uint32_t * signs = (const uint32_t *)(ksigns64 + (q2[l] >> 9));
+        const int grid_l = dpct::vectorized_binary<sycl::uchar4>(
+            grid[0] ^ signs[0], signs[0], std::minus<>());
+        const int grid_h = dpct::vectorized_binary<sycl::uchar4>(
+            grid[1] ^ signs[1], signs[1], std::minus<>());
+        sumi2 = dpct::dp4a(grid_l, *((const int *)q8 + 0), sumi2);
+        sumi2 = dpct::dp4a(grid_h, *((const int *)q8 + 1), sumi2);
+        q8 += 8;
+    }
+    const float d = (float)bq2->d * bq8_1[ib32].ds[0] * 0.25f;
+    return d * ((0.5f + ls1) * sumi1 + (0.5f + ls2) * sumi2);
+#else
+    assert(false);
+    return 0.f;
+#endif
+#else
+    assert(false);
+    return 0.f;
+#endif
+}
+
+static __dpct_inline__ float
+vec_dot_iq2_s_q8_1(const void *__restrict__ vbq,
+                   const block_q8_1 *__restrict__ bq8_1, const int &iqs) {
+#if QK_K == 256
+    const block_iq2_s * bq2 = (const block_iq2_s *) vbq;
+
+    const int ib32 = iqs;
+    const int8_t  * q8 = bq8_1[ib32].qs;
+    const uint8_t * signs = bq2->qs + QK_K/8 + 4*ib32;
+    const uint8_t ls1 = bq2->scales[ib32] & 0xf;
+    const uint8_t ls2 = bq2->scales[ib32] >>  4;
+    int sumi1 = 0;
+    for (int l = 0; l < 2; ++l) {
+        const uint32_t * grid = (const uint32_t *)(iq2s_grid + (bq2->qs[4*ib32+l] | ((bq2->qh[ib32] << (8-2*l)) & 0x300)));
+        const uint32_t signs0 = dpct::vectorized_binary<sycl::uchar4>(
+            ((signs[l] & 0xf) * 0x01010101) & 0x08040201, 0x08040201,
+            std::equal_to<>());
+        const uint32_t signs1 = dpct::vectorized_binary<sycl::uchar4>(
+            ((signs[l] >> 4) * 0x01010101) & 0x08040201, 0x08040201,
+            std::equal_to<>());
+        const int grid_l = dpct::vectorized_binary<sycl::uchar4>(
+            grid[0] ^ signs0, signs0, std::minus<>());
+        const int grid_h = dpct::vectorized_binary<sycl::uchar4>(
+            grid[1] ^ signs1, signs1, std::minus<>());
+        sumi1 = dpct::dp4a(grid_l, *((const int *)q8 + 0), sumi1);
+        sumi1 = dpct::dp4a(grid_h, *((const int *)q8 + 1), sumi1);
+        q8 += 8;
+    }
+    int sumi2 = 0;
+    for (int l = 2; l < 4; ++l) {
+        const uint32_t * grid = (const uint32_t *)(iq2s_grid + (bq2->qs[4*ib32+l] | ((bq2->qh[ib32] << (8-2*l)) & 0x300)));
+        const uint32_t signs0 = dpct::vectorized_binary<sycl::uchar4>(
+            ((signs[l] & 0xf) * 0x01010101) & 0x08040201, 0x08040201,
+            std::equal_to<>());
+        const uint32_t signs1 = dpct::vectorized_binary<sycl::uchar4>(
+            ((signs[l] >> 4) * 0x01010101) & 0x08040201, 0x08040201,
+            std::equal_to<>());
+        const int grid_l = dpct::vectorized_binary<sycl::uchar4>(
+            grid[0] ^ signs0, signs0, std::minus<>());
+        const int grid_h = dpct::vectorized_binary<sycl::uchar4>(
+            grid[1] ^ signs1, signs1, std::minus<>());
+        sumi2 = dpct::dp4a(grid_l, *((const int *)q8 + 0), sumi2);
+        sumi2 = dpct::dp4a(grid_h, *((const int *)q8 + 1), sumi2);
+        q8 += 8;
+    }
+    const float d = (float)bq2->d * bq8_1[ib32].ds[0] * 0.25f;
+    return d * ((0.5f + ls1) * sumi1 + (0.5f + ls2) * sumi2);
+#else
+    assert(false);
+#endif
+}
+
+static __dpct_inline__ float
+vec_dot_iq3_xxs_q8_1(const void *__restrict__ vbq,
+                     const block_q8_1 *__restrict__ bq8_1, const int &iqs,
+                     const uint32_t *iq3xxs_grid, const uint64_t *ksigns64) {
+#if DPCT_COMPATIBILITY_TEMP >=                                                 \
+    MIN_CC_DP4A // lowest compute capability for integer intrinsics
+#if QK_K == 256
+    const block_iq3_xxs * bq2 = (const block_iq3_xxs *) vbq;
+
+    const int ib32 = iqs;
+    const uint8_t  * q3 = bq2->qs + 8*ib32;
+    const uint16_t * gas = (const uint16_t *)(bq2->qs + QK_K/4) + 2*ib32;
+    const int8_t   * q8 = bq8_1[ib32].qs;
+    uint32_t aux32 = gas[0] | (gas[1] << 16);
+    int sumi = 0;
+    for (int l = 0; l < 4; ++l) {
+        const uint32_t * grid1 = iq3xxs_grid + q3[2*l+0];
+        const uint32_t * grid2 = iq3xxs_grid + q3[2*l+1];
+        const uint32_t * signs = (const uint32_t *)(ksigns64 + (aux32 & 127));
+        const int grid_l = dpct::vectorized_binary<sycl::uchar4>(
+            grid1[0] ^ signs[0], signs[0], std::minus<>());
+        const int grid_h = dpct::vectorized_binary<sycl::uchar4>(
+            grid2[0] ^ signs[1], signs[1], std::minus<>());
+        sumi = dpct::dp4a(grid_l, *((int *)q8 + 0), sumi);
+        sumi = dpct::dp4a(grid_h, *((int *)q8 + 1), sumi);
+        q8 += 8;
+        aux32 >>= 7;
+    }
+    const float d = (float)bq2->d * (0.5f + aux32) * bq8_1[ib32].ds[0] * 0.5f;
+    return d * sumi;
+#else
+    assert(false);
+    return 0.f;
+#endif
+#else
+    assert(false);
+    return 0.f;
+#endif
+}
+
+static __dpct_inline__ float
+vec_dot_iq3_s_q8_1(const void *__restrict__ vbq,
+                   const block_q8_1 *__restrict__ bq8_1, const int &iqs,
+                   const uint32_t *iq3s_grid) {
+#if QK_K == 256
+    const block_iq3_s * bq2 = (const block_iq3_s *) vbq;
+
+    const int ib32 = iqs;
+    const uint8_t  * qs = bq2->qs + 8*ib32;
+    const int8_t   * q8 = bq8_1[ib32].qs;
+    int sumi = 0;
+    for (int l = 0; l < 4; ++l) {
+        const uint32_t * grid1 = iq3s_grid + (qs[2*l+0] | ((bq2->qh[ib32] << (8 - 2*l)) & 256));
+        const uint32_t * grid2 = iq3s_grid + (qs[2*l+1] | ((bq2->qh[ib32] << (7 - 2*l)) & 256));
+        uint32_t signs0 = dpct::vectorized_binary<sycl::uchar4>(
+            ((bq2->signs[4 * ib32 + l] & 0xf) * 0x01010101) & 0x08040201,
+            0x08040201, std::equal_to<>());
+        uint32_t signs1 = dpct::vectorized_binary<sycl::uchar4>(
+            ((bq2->signs[4 * ib32 + l] >> 4) * 0x01010101) & 0x08040201,
+            0x08040201, std::equal_to<>());
+        const int grid_l = dpct::vectorized_binary<sycl::uchar4>(
+            grid1[0] ^ signs0, signs0, std::minus<>());
+        const int grid_h = dpct::vectorized_binary<sycl::uchar4>(
+            grid2[0] ^ signs1, signs1, std::minus<>());
+        sumi = dpct::dp4a(grid_l, *((int *)q8 + 0), sumi);
+        sumi = dpct::dp4a(grid_h, *((int *)q8 + 1), sumi);
+        q8 += 8;
+    }
+    const float d =
+        (float)bq2->d *
+        (1 + 2 * ((bq2->scales[ib32 / 2] >> 4 * (ib32 % 2)) & 0xf)) *
+        bq8_1[ib32].ds[0];
+    return d * sumi;
+#else
+    assert(false);
+#endif
+}
+
+static __dpct_inline__ float
+vec_dot_iq1_s_q8_1(const void *__restrict__ vbq,
+                   const block_q8_1 *__restrict__ bq8_1, const int &iqs,
+                   const uint32_t *iq1s_grid_gpu) {
+#if QK_K == 256
+    const block_iq1_s * bq1 = (const block_iq1_s *) vbq;
+
+    const int ib32 = iqs;
+    int sumi = 0;
+    const int * q8 = (const int *)bq8_1[ib32].qs;
+    for (int l = 0; l < 4; ++l) {
+        const int * grid = (const int *)(iq1s_grid_gpu + (bq1->qs[4*ib32+l] | (((bq1->qh[ib32] >> 3*l) & 7) << 8)));
+        int grid0 = grid[0] & 0x0f0f0f0f;
+        int grid1 = (grid[0] >> 4) & 0x0f0f0f0f;
+        sumi = dpct::dp4a(q8[2 * l + 1], grid1,
+                          dpct::dp4a(q8[2 * l + 0], grid0, sumi));
+    }
+
+    const float delta = bq1->qh[ib32] & 0x8000 ? -1-IQ1S_DELTA : -1+IQ1S_DELTA;
+    const float d1q = (float)bq1->d * (2*((bq1->qh[ib32] >> 12) & 7) + 1);
+    const float d = d1q * bq8_1[ib32].ds[0];
+    const float m = d1q * bq8_1[ib32].ds[1];
+    return d * sumi + m * delta;
+#else
+    assert(false);
+#endif
+}
+
+static __dpct_inline__ float
+vec_dot_iq1_m_q8_1(const void *__restrict__ vbq,
+                   const block_q8_1 *__restrict__ bq8_1, const int &iqs) {
+#if QK_K == 256
+    const block_iq1_m * bq1 = (const block_iq1_m *) vbq;
+
+    const int ib32 = iqs;
+    int   sumi[2] = {0, 0};
+    float sumf[2] = {0.f, 0.f};
+
+    const int * q8 = (const int *)bq8_1[ib32].qs;
+    for (int l = 0; l < 4; ++l) {
+        const int * grid = (const int *)(iq1s_grid_gpu + (bq1->qs[4*ib32+l] | (((bq1->qh[2*ib32+l/2] >> 4*(l%2)) & 7) << 8)));
+        int grid0 = grid[0] & 0x0f0f0f0f;
+        int grid1 = (grid[0] >> 4) & 0x0f0f0f0f;
+        sumi[l / 2] = dpct::dp4a(q8[2 * l + 1], grid1,
+                                 dpct::dp4a(q8[2 * l + 0], grid0, sumi[l / 2]));
+        const float delta = (bq1->qh[2*ib32+l/2] >> 4*(l%2)) & 0x08 ? -1-IQ1M_DELTA : -1+IQ1M_DELTA;
+        const int sumy = dpct::dp4a(q8[2 * l + 1], 0x01010101,
+                                    dpct::dp4a(q8[2 * l + 0], 0x01010101, 0));
+        sumf[l/2] += delta*sumy;
+    }
+
+    iq1m_scale_t scale;
+    const uint16_t * sc = (const uint16_t *)bq1->scales;
+    scale.u16 = (sc[0] >> 12) | ((sc[1] >> 8) & 0x00f0) | ((sc[2] >> 4) & 0x0f00) | (sc[3] & 0xf000);
+    const float d = (float)scale.f16 * bq8_1[ib32].ds[0];
+    return d * ((sumi[0] + sumf[0]) * (2*((sc[ib32/2] >> 6*(ib32%2)) & 0x7) + 1) + (sumi[1] + sumf[1]) * (2*((sc[ib32/2] >> (6*(ib32%2)+3)) & 0x7) + 1));
+#else
+    assert(false);
+#endif
+}
+
+
+static __dpct_inline__ float
+vec_dot_iq4_nl_q8_1(const void *__restrict__ vbq,
+                    const block_q8_1 *__restrict__ bq8_1, const int &iqs) {
+
+    const block_iq4_nl * bq = (const block_iq4_nl *) vbq;
+
+    const uint16_t * q4 = (const uint16_t *)bq->qs + 2*iqs;
+    const int32_t  * q8 = (const int32_t  *)bq8_1->qs + iqs;
+
+    const uint8_t * values = (const uint8_t *)kvalues_iq4nl;
+
+    int v1, v2;
+    int sumi1 = 0, sumi2 = 0;
+    for (int l = 0; l < VDR_Q4_0_Q8_1_MMVQ; ++l) {
+        const uint32_t aux = q4[2*l] | (q4[2*l+1] << 16);
+        get_int_from_table_16(aux, values, v1, v2);
+        sumi1 = dpct::dp4a(v1, q8[l + 0], sumi1);
+        sumi2 = dpct::dp4a(v2, q8[l + 4], sumi2);
+    }
+
+    const float d = (float)bq->d * bq8_1->ds[0];
+    return d * (sumi1 + sumi2);
+}
+
+
+static __dpct_inline__ float
+vec_dot_iq4_xs_q8_1(const void *__restrict__ vbq,
+                    const block_q8_1 *__restrict__ bq8_1, const int &iqs) {
+
+#if QK_K == 256
+    const block_iq4_xs * bq4 = (const block_iq4_xs *) vbq;
+    const uint8_t * values = (const uint8_t *)kvalues_iq4nl;
+
+    // iqs is 0...7
+    const int ib32 = iqs;
+    const int32_t  * q8 = (const int *)bq8_1[ib32].qs;
+    const uint32_t * q4 = (const uint32_t *)bq4->qs + 4*ib32;
+    const int8_t ls = ((bq4->scales_l[ib32/2] >> 4*(ib32%2)) & 0xf) | (((bq4->scales_h >> 2*ib32) & 3) << 4);
+    const float d = (float)bq4->d * (ls - 32) * bq8_1[ib32].ds[0];
+    int v1, v2;
+    int sumi1 = 0, sumi2 = 0;
+    for (int j = 0; j < 4; ++j) {
+        get_int_from_table_16(q4[j], values, v1, v2);
+        sumi1 = dpct::dp4a(v1, q8[j + 0], sumi1);
+        sumi2 = dpct::dp4a(v2, q8[j + 4], sumi2);
+    }
+    return d * (sumi1 + sumi2);
+#else
+    assert(false);
+#endif
+}
+
+#endif // GGML_SYCL_VECDOTQ_HPP
diff --git a/src/whisper_cpp/ggml/src/ggml-vulkan.cpp b/src/whisper_cpp/ggml/src/ggml-vulkan.cpp
new file mode 100644
index 00000000..30bd376d
--- /dev/null
+++ b/src/whisper_cpp/ggml/src/ggml-vulkan.cpp
@@ -0,0 +1,7406 @@
+#include "ggml-vulkan.h"
+#include <vulkan/vulkan_core.h>
+#if defined(GGML_VULKAN_RUN_TESTS) || defined(GGML_VULKAN_PERF)
+#include <chrono>
+#endif
+
+#include <vulkan/vulkan.hpp>
+
+#include <algorithm>
+#include <cmath>
+#include <iomanip>
+#include <iostream>
+#include <tuple>
+#include <vector>
+#include <sstream>
+#include <utility>
+#include <memory>
+#include <limits>
+#include <map>
+#include <unordered_map>
+#include <memory>
+#include <mutex>
+#include <future>
+#include <thread>
+
+#include "ggml-impl.h"
+#include "ggml-backend-impl.h"
+
+#include "ggml-vulkan-shaders.hpp"
+
+#define VK_API_VERSION VK_API_VERSION_1_2
+
+#define CEIL_DIV(M, N) (((M) + (N)-1) / (N))
+
+#define VK_VENDOR_ID_AMD 0x1002
+#define VK_VENDOR_ID_APPLE 0x106b
+#define VK_VENDOR_ID_INTEL 0x8086
+#define VK_VENDOR_ID_NVIDIA 0x10de
+
+#define VK_DEVICE_DESCRIPTOR_POOL_SIZE 32
+
+#define GGML_VK_MAX_NODES 8192
+
+#define MAX_VK_BUFFERS 256
+
+#ifndef K_QUANTS_PER_ITERATION
+#define K_QUANTS_PER_ITERATION 1
+#else
+static_assert(K_QUANTS_PER_ITERATION == 1 || K_QUANTS_PER_ITERATION == 2, "K_QUANTS_PER_ITERATION must be 1 or 2");
+#endif
+
+#define VK_CHECK(err, msg)                                          \
+    do {                                                            \
+        vk::Result err_ = (err);                                    \
+        if (err_ != vk::Result::eSuccess) {                         \
+            fprintf(stderr, "ggml_vulkan: %s error %s at %s:%d\n",  \
+                #err, to_string(err_).c_str(), __FILE__, __LINE__); \
+            exit(1);                                                \
+        }                                                           \
+    } while (0)
+
+#ifdef GGML_VULKAN_DEBUG
+#define VK_LOG_DEBUG(msg) std::cerr << msg << std::endl
+#else
+#define VK_LOG_DEBUG(msg) ((void) 0)
+#endif // GGML_VULKAN_DEBUG
+
+struct ggml_backend_vk_context;
+
+struct vk_queue {
+    uint32_t queue_family_index;
+    vk::Queue queue;
+    vk::CommandPool pool;
+    uint32_t cmd_buffer_idx;
+    std::vector<vk::CommandBuffer> cmd_buffers;
+
+    vk::PipelineStageFlags stage_flags;
+
+    bool transfer_only;
+};
+
+struct vk_pipeline_struct {
+    std::string name;
+    vk::ShaderModule shader_module;
+    vk::DescriptorSetLayout dsl;
+    std::vector<vk::DescriptorPool> descriptor_pools;
+    std::vector<vk::DescriptorSet> descriptor_sets;
+    uint32_t descriptor_set_idx;
+    vk::PipelineLayout layout;
+    vk::Pipeline pipeline;
+    uint32_t push_constant_size;
+    uint32_t parameter_count;
+    std::array<uint32_t, 3> wg_denoms;
+    uint32_t align;
+};
+
+typedef std::shared_ptr<vk_pipeline_struct> vk_pipeline;
+typedef std::weak_ptr<vk_pipeline_struct> vk_pipeline_ref;
+
+static void ggml_vk_destroy_pipeline(vk::Device& device, vk_pipeline& pipeline);
+
+struct vk_matmul_pipeline_struct {
+    vk_pipeline l, m, s;
+    vk_pipeline a_l, a_m, a_s;
+};
+
+typedef std::shared_ptr<vk_matmul_pipeline_struct> vk_matmul_pipeline;
+
+struct vk_device_struct;
+typedef std::shared_ptr<vk_device_struct> vk_device;
+typedef std::weak_ptr<vk_device_struct> vk_device_ref;
+
+struct vk_buffer_struct;
+typedef std::shared_ptr<vk_buffer_struct> vk_buffer;
+typedef std::weak_ptr<vk_buffer_struct> vk_buffer_ref;
+
+struct ggml_backend_vk_buffer_type_context {
+    std::string name;
+    vk_device device;
+};
+
+static const char * ggml_backend_vk_buffer_type_name(ggml_backend_buffer_type_t buft);
+static ggml_backend_buffer_t ggml_backend_vk_buffer_type_alloc_buffer(ggml_backend_buffer_type_t buft, size_t size);
+static size_t ggml_backend_vk_buffer_type_get_alignment(ggml_backend_buffer_type_t buft);
+static size_t ggml_backend_vk_buffer_type_get_max_size(ggml_backend_buffer_type_t buft);
+static size_t ggml_backend_vk_buffer_type_get_alloc_size(ggml_backend_buffer_type_t buft, const ggml_tensor * tensor);
+static ggml_backend_buffer_type_i ggml_backend_vk_buffer_type_interface = {
+    /* .get_name         = */ ggml_backend_vk_buffer_type_name,
+    /* .alloc_buffer     = */ ggml_backend_vk_buffer_type_alloc_buffer,
+    /* .get_alignment    = */ ggml_backend_vk_buffer_type_get_alignment,
+    /* .get_max_size     = */ ggml_backend_vk_buffer_type_get_max_size,
+    /* .get_alloc_size   = */ ggml_backend_vk_buffer_type_get_alloc_size,
+    /* .is_host          = */ NULL,
+};
+
+#ifdef GGML_VULKAN_MEMORY_DEBUG
+class vk_memory_logger;
+#endif
+#ifdef GGML_VULKAN_PERF
+class vk_perf_logger;
+#endif
+static void ggml_vk_destroy_buffer(vk_buffer& buf);
+
+struct vk_device_struct {
+    std::mutex mutex;
+
+    vk::PhysicalDevice physical_device;
+    vk::PhysicalDeviceProperties properties;
+    std::string name;
+    uint64_t max_memory_allocation_size;
+    bool fp16;
+    vk::Device device;
+    uint32_t vendor_id;
+    vk_queue compute_queue;
+    vk_queue transfer_queue;
+    bool single_queue;
+    uint32_t subgroup_size;
+    bool uma;
+
+    size_t idx;
+
+    vk_matmul_pipeline pipeline_matmul_f32;
+    vk_matmul_pipeline pipeline_matmul_f32_f16;
+    vk_matmul_pipeline pipeline_matmul_f16;
+    vk_matmul_pipeline pipeline_matmul_f16_f32;
+    vk_pipeline pipeline_matmul_split_k_reduce;
+
+    vk_matmul_pipeline pipeline_dequant_mul_mat_mat[GGML_TYPE_COUNT];
+
+    vk_matmul_pipeline pipeline_matmul_id_f32;
+    vk_matmul_pipeline pipeline_matmul_id_f16;
+    vk_matmul_pipeline pipeline_matmul_id_f16_f32;
+
+    vk_matmul_pipeline pipeline_dequant_mul_mat_mat_id[GGML_TYPE_COUNT];
+
+    vk_pipeline pipeline_dequant[GGML_TYPE_COUNT];
+    vk_pipeline pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_COUNT];
+    vk_pipeline pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_COUNT];
+    vk_pipeline pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_COUNT];
+
+    vk_pipeline pipeline_mul_mat_vec_p021_f16_f32;
+    vk_pipeline pipeline_mul_mat_vec_nc_f16_f32;
+    vk_pipeline pipeline_get_rows[GGML_TYPE_COUNT];
+    vk_pipeline pipeline_get_rows_f32[GGML_TYPE_COUNT];
+    vk_pipeline pipeline_acc_f32;
+    vk_pipeline pipeline_add_f32, pipeline_add_f16_f32_f16;
+    vk_pipeline pipeline_mul_f32;
+    vk_pipeline pipeline_div_f32;
+    vk_pipeline pipeline_concat_f32, pipeline_concat_f16, pipeline_concat_i32;
+    vk_pipeline pipeline_upscale_f32;
+    vk_pipeline pipeline_scale_f32;
+    vk_pipeline pipeline_sqr_f32;
+    vk_pipeline pipeline_sin_f32;
+    vk_pipeline pipeline_cos_f32;
+    vk_pipeline pipeline_clamp_f32;
+    vk_pipeline pipeline_pad_f32;
+    vk_pipeline pipeline_repeat_f32;
+    vk_pipeline pipeline_cpy_f32_f32, pipeline_cpy_f32_f16, pipeline_cpy_f16_f16;
+    vk_pipeline pipeline_norm_f32;
+    vk_pipeline pipeline_group_norm_f32;
+    vk_pipeline pipeline_rms_norm_f32;
+    vk_pipeline pipeline_gelu_f32;
+    vk_pipeline pipeline_gelu_quick_f32;
+    vk_pipeline pipeline_silu_f32;
+    vk_pipeline pipeline_relu_f32;
+    vk_pipeline pipeline_leaky_relu_f32;
+    vk_pipeline pipeline_tanh_f32;
+    vk_pipeline pipeline_diag_mask_inf_f32;
+    vk_pipeline pipeline_soft_max_f32, pipeline_soft_max_f32_f16;
+    vk_pipeline pipeline_rope_norm_f32, pipeline_rope_norm_f16;
+    vk_pipeline pipeline_rope_neox_f32, pipeline_rope_neox_f16;
+    vk_pipeline pipeline_argsort_f32;
+    vk_pipeline pipeline_sum_rows_f32;
+    vk_pipeline pipeline_im2col_f32, pipeline_im2col_f32_f16;
+    vk_pipeline pipeline_timestep_embedding_f32;
+
+    std::unordered_map<std::string, vk_pipeline_ref> pipelines;
+    std::unordered_map<std::string, uint64_t> pipeline_descriptor_set_requirements;
+
+    std::vector<std::tuple<void*, size_t, vk_buffer>> pinned_memory;
+
+    vk::Fence fence;
+    vk_buffer sync_staging;
+
+    ggml_backend_buffer_type buffer_type;
+
+#ifdef GGML_VULKAN_MEMORY_DEBUG
+    std::unique_ptr<vk_memory_logger> memory_logger;
+#endif
+#ifdef GGML_VULKAN_PERF
+    std::unique_ptr<vk_perf_logger> perf_logger;
+#endif
+
+    ~vk_device_struct() {
+        VK_LOG_DEBUG("destroy device " << name);
+
+        device.destroyFence(fence);
+
+        ggml_vk_destroy_buffer(sync_staging);
+
+        device.destroyCommandPool(compute_queue.pool);
+        if (!single_queue) {
+            device.destroyCommandPool(transfer_queue.pool);
+        }
+
+        for (auto& pipeline : pipelines) {
+            if (pipeline.second.expired()) {
+                continue;
+            }
+
+            vk_pipeline pl = pipeline.second.lock();
+            ggml_vk_destroy_pipeline(device, pl);
+        }
+        pipelines.clear();
+
+        device.destroy();
+    }
+};
+
+struct vk_buffer_struct {
+    vk::Buffer buffer = VK_NULL_HANDLE;
+    vk::DeviceMemory device_memory = VK_NULL_HANDLE;
+    vk::MemoryPropertyFlags memory_property_flags;
+    void * ptr;
+    size_t size = 0;
+
+    vk_device device;
+
+    ~vk_buffer_struct() {
+        if (size == 0) {
+            return;
+        }
+        VK_LOG_DEBUG("~vk_buffer_struct(" << buffer << ", " << size << ")");
+
+        device->device.freeMemory(device_memory);
+        device->device.destroyBuffer(buffer);
+    }
+};
+
+struct vk_subbuffer {
+    vk_buffer buffer;
+    uint64_t offset;
+    uint64_t size;
+
+    operator vk::DescriptorBufferInfo() const {
+        return { buffer->buffer, offset, size };
+    }
+};
+
+struct vk_semaphore {
+    vk::Semaphore s;
+    uint64_t value;
+};
+
+struct vk_submission {
+    vk::CommandBuffer buffer;
+    std::vector<vk_semaphore> wait_semaphores;
+    std::vector<vk_semaphore> signal_semaphores;
+};
+
+typedef std::vector<vk_submission> vk_sequence;
+
+struct vk_mat_mat_push_constants {
+    uint32_t M; uint32_t N; uint32_t K;
+    uint32_t stride_a; uint32_t stride_b; uint32_t stride_d;
+    uint32_t batch_stride_a; uint32_t batch_stride_b; uint32_t batch_stride_d;
+    uint32_t k_split;
+    uint32_t ne02; uint32_t ne12; uint32_t broadcast2; uint32_t broadcast3;
+};
+struct vk_mat_vec_push_constants {
+    uint32_t ncols; uint32_t stride_a; uint32_t stride_b; uint32_t stride_d;
+    uint32_t batch_stride_a; uint32_t batch_stride_b; uint32_t batch_stride_d;
+    uint32_t ne02; uint32_t ne12; uint32_t broadcast2; uint32_t broadcast3;
+};
+
+struct vk_mat_mat_id_push_constants {
+    uint32_t M; uint32_t N; uint32_t K;
+    uint32_t stride_a; uint32_t stride_b; uint32_t stride_d;
+    uint32_t batch_stride_a; uint32_t batch_stride_b; uint32_t batch_stride_d;
+    uint32_t nei0; uint32_t nei1; uint32_t nbi1; uint32_t ne11;
+};
+struct vk_mat_vec_id_push_constants {
+    uint32_t ncols; uint32_t stride_a; uint32_t stride_b; uint32_t stride_d;
+    uint32_t batch_stride_a; uint32_t batch_stride_b; uint32_t batch_stride_d;
+    uint32_t nei0; uint32_t ne11;
+};
+
+struct vk_op_push_constants {
+    uint32_t KX;
+    uint32_t KY;
+    float param1;
+    float param2;
+};
+
+struct vk_op_unary_push_constants {
+    uint32_t ne;
+    uint32_t ne00; uint32_t ne01; uint32_t ne02; uint32_t ne03; uint32_t nb00; uint32_t nb01; uint32_t nb02; uint32_t nb03;
+    uint32_t ne10; uint32_t ne11; uint32_t ne12; uint32_t ne13; uint32_t nb10; uint32_t nb11; uint32_t nb12; uint32_t nb13;
+    uint32_t d_offset;
+    float param1; float param2;
+};
+
+struct vk_op_binary_push_constants {
+    uint32_t ne;
+    uint32_t ne00; uint32_t ne01; uint32_t ne02; uint32_t ne03; uint32_t nb00; uint32_t nb01; uint32_t nb02; uint32_t nb03;
+    uint32_t ne10; uint32_t ne11; uint32_t ne12; uint32_t ne13; uint32_t nb10; uint32_t nb11; uint32_t nb12; uint32_t nb13;
+    uint32_t ne20; uint32_t ne21; uint32_t ne22; uint32_t ne23; uint32_t nb20; uint32_t nb21; uint32_t nb22; uint32_t nb23;
+    uint32_t d_offset;
+    float param1; float param2; int32_t param3;
+};
+
+struct vk_op_diag_mask_push_constants {
+    uint32_t ncols;
+    uint32_t rows_per_channel;
+    int32_t n_past;
+};
+
+struct vk_op_rope_push_constants {
+    uint32_t ncols;
+    uint32_t n_dims;
+    float freq_scale;
+    uint32_t p_delta_rows;
+    float freq_base;
+    float ext_factor;
+    float attn_factor;
+    float corr_dims[2];
+    float theta_scale;
+    uint32_t has_ff;
+};
+
+struct vk_op_soft_max_push_constants {
+    uint32_t KX;
+    uint32_t KY;
+    float scale;
+    float max_bias;
+    float m0;
+    float m1;
+    uint32_t n_head_log2;
+};
+
+struct vk_op_argsort_push_constants {
+    uint32_t ncols;
+    uint32_t ncols_pad;
+    int32_t order;
+};
+
+struct vk_op_im2col_push_constants {
+    uint32_t batch_offset; uint32_t offset_delta;
+    uint32_t IC;
+    uint32_t IW; uint32_t IH;
+    uint32_t OW; uint32_t OH;
+    uint32_t KW; uint32_t KH;
+    uint32_t pelements;
+    uint32_t CHW;
+    int32_t s0; int32_t s1;
+    int32_t p0; int32_t p1;
+    int32_t d0; int32_t d1;
+};
+
+struct vk_op_timestep_embedding_push_constants {
+    uint32_t nb1;
+    uint32_t dim;
+    uint32_t max_period;
+};
+
+// Allow pre-recording command buffers
+struct vk_staging_memcpy {
+    vk_staging_memcpy(void * _dst, const void * _src, size_t _n) : dst(_dst), src(_src), n(_n) {}
+
+    void * dst;
+    const void * src;
+    size_t n;
+};
+
+struct vk_op_upscale_push_constants {
+    uint32_t ne; uint32_t d_offset;
+    uint32_t nb00; uint32_t nb01; uint32_t nb02; uint32_t nb03;
+    uint32_t ne10; uint32_t ne11; uint32_t ne12; uint32_t ne13;
+    float sf0; float sf1; float sf2; float sf3;
+};
+
+struct vk_context_struct {
+    vk_submission * s;
+    std::vector<vk_sequence> seqs;
+
+    int exit_tensor_idx;
+
+    std::vector<vk_staging_memcpy> in_memcpys;
+    std::vector<vk_staging_memcpy> out_memcpys;
+
+    vk_queue * q;
+};
+typedef std::shared_ptr<vk_context_struct> vk_context;
+typedef std::weak_ptr<vk_context_struct> vk_context_ref;
+
+struct ggml_vk_garbage_collector {
+    std::vector<vk_semaphore> tl_semaphores;
+    std::vector<vk_semaphore> semaphores;
+    std::vector<vk::Event> events;
+    std::vector<vk_buffer> temp_buffers;
+    std::vector<vk_context> contexts;
+};
+
+#if defined(GGML_VULKAN_MEMORY_DEBUG) || defined(GGML_VULKAN_DEBUG)
+#define VK_LOG_MEMORY(msg) std::cerr << "ggml_vulkan memory: " << msg << std::endl
+
+static std::string format_size(size_t size) {
+    const size_t kib = 1024;
+    const size_t mib = kib * 1024;
+    const size_t gib = mib * 1024;
+
+    std::ostringstream oss;
+    oss << std::fixed << std::setprecision(2);
+
+    if (size >= gib) {
+        oss << static_cast<double>(size) / gib << " GiB";
+    } else if (size >= mib) {
+        oss << static_cast<double>(size) / mib << " MiB";
+    } else if (size >= kib) {
+        oss << static_cast<double>(size) / kib << " KiB";
+    } else {
+        oss << size << " B";
+    }
+
+    return oss.str();
+}
+
+static std::mutex log_mutex;
+
+class vk_memory_logger {
+public:
+    vk_memory_logger(): total_device(0), total_host(0) {}
+    void log_allocation(vk_buffer_ref buf_ref, size_t size);
+    void log_deallocation(vk_buffer_ref buf_ref);
+
+private:
+    std::map<vk::Buffer, size_t> allocations; // Track allocations
+    size_t total_device;
+    size_t total_host;
+};
+#else
+#define VK_LOG_MEMORY(msg) ((void) 0)
+#endif // GGML_VULKAN_MEMORY_DEBUG
+
+#if defined(GGML_VULKAN_PERF)
+
+class vk_perf_logger {
+public:
+    void print_timings() {
+        std::cerr << "----------------\nVulkan Timings:" << std::endl;
+        for (const auto& t : timings) {
+            uint64_t total = 0;
+            for (const auto& time : t.second) {
+                total += time;
+            }
+            std::cerr << t.first << ": " << t.second.size() << " x " << (total / t.second.size() / 1000.0) << " ms" << std::endl;
+        }
+
+        timings.clear();
+    }
+
+    void log_timing(const ggml_tensor * node, uint64_t time) {
+        if (node->op == GGML_OP_UNARY) {
+            timings[ggml_unary_op_name(ggml_get_unary_op(node))].push_back(time);
+            return;
+        }
+        if (node->op == GGML_OP_MUL_MAT || node->op == GGML_OP_MUL_MAT_ID) {
+            const uint64_t m = node->src[0]->ne[1];
+            const uint64_t n = node->src[1]->ne[1];
+            const uint64_t k = node->src[1]->ne[0];
+            std::string name = ggml_op_name(node->op);
+            if (n == 1) {
+                name += "_VEC m=" + std::to_string(m) + " k=" + std::to_string(k);
+            } else {
+                name += " m=" + std::to_string(m) + " n=" + std::to_string(n) + " k=" + std::to_string(k);
+            }
+            timings[name].push_back(time);
+            return;
+        }
+        timings[ggml_op_name(node->op)].push_back(time);
+    }
+private:
+    std::map<std::string, std::vector<uint64_t>> timings;
+};
+#endif // GGML_VULKAN_PERF
+
+struct ggml_backend_vk_context {
+    std::string name;
+
+    vk_device device;
+
+    size_t semaphore_idx, event_idx;
+    ggml_vk_garbage_collector gc;
+    size_t prealloc_size_x, prealloc_size_y, prealloc_size_split_k;
+    vk_buffer prealloc_x, prealloc_y, prealloc_split_k;
+    vk::Fence fence;
+
+    vk_buffer buffer_pool[MAX_VK_BUFFERS];
+
+    vk_context_ref compute_ctx;
+    vk_context_ref transfer_ctx;
+
+    std::vector<vk_context_ref> tensor_ctxs;
+};
+
+static void * const vk_ptr_base = (void *)(uintptr_t) 0x1000;  // NOLINT
+
+static uint64_t vk_tensor_offset(const ggml_tensor * tensor) {
+    if (tensor->view_src) {
+        return (uint8_t *) tensor->view_src->data - (uint8_t *) vk_ptr_base;
+    }
+    return (uint8_t *) tensor->data - (uint8_t *) vk_ptr_base;
+}
+
+struct ggml_backend_vk_buffer_context {
+    vk_device_ref device;
+    vk_buffer dev_buffer;
+    std::string name;
+
+    ggml_backend_vk_buffer_context(vk_device_ref device, vk_buffer&& dev_buffer, std::string& name) :
+        device(device),
+        dev_buffer(dev_buffer),
+        name(name) {
+    }
+
+    ~ggml_backend_vk_buffer_context() {
+        ggml_vk_destroy_buffer(dev_buffer);
+    }
+};
+
+#ifdef GGML_VULKAN_MEMORY_DEBUG
+void vk_memory_logger::log_allocation(vk_buffer_ref buf_ref, size_t size) {
+    std::lock_guard<std::mutex> guard(log_mutex);
+    vk_buffer buf = buf_ref.lock();
+    const bool device = bool(buf->memory_property_flags & vk::MemoryPropertyFlagBits::eDeviceLocal);
+    const std::string type = device ? "device" : "host";
+    allocations[buf->buffer] = size;
+    total_device += device ? size : 0;
+    total_host += device ? 0 : size;
+    VK_LOG_MEMORY(buf->device->name << ": +" << format_size(size) << " " << type << " at " << buf->buffer << ". Total device: " << format_size(total_device) << ", total host: " << format_size(total_host));
+}
+
+void vk_memory_logger::log_deallocation(vk_buffer_ref buf_ref) {
+    if (buf_ref.expired() || buf_ref.lock()->size == 0) {
+        return;
+    }
+
+    std::lock_guard<std::mutex> guard(log_mutex);
+    vk_buffer buf = buf_ref.lock();
+    const bool device = bool(buf->memory_property_flags & vk::MemoryPropertyFlagBits::eDeviceLocal);
+    std::string type = device ? "device" : "host";
+    auto it = allocations.find(buf->buffer);
+    total_device -= device ? it->second : 0;
+    total_host -= device ? 0 : it->second;
+    if (it != allocations.end()) {
+        VK_LOG_MEMORY(buf->device->name << ": -" << format_size(it->second) << " " << type << " at " << buf->buffer << ". Total device: " << format_size(total_device) << ", total host: " << format_size(total_host));
+        allocations.erase(it);
+    } else {
+        VK_LOG_MEMORY("ERROR " << buf->device->name << ": Attempted to deallocate unknown " << type << " memory at " << buf->buffer);
+    }
+}
+#endif // GGML_VULKAN_MEMORY_DEBUG
+
+struct vk_instance_t {
+    vk::Instance instance;
+
+    std::vector<size_t> device_indices;
+    vk_device devices[GGML_VK_MAX_DEVICES];
+};
+
+static bool vk_instance_initialized = false;
+static vk_instance_t vk_instance;
+
+#ifdef GGML_VULKAN_CHECK_RESULTS
+static size_t vk_skip_checks;
+static size_t vk_output_tensor;
+
+static void ggml_vk_print_tensor(const ggml_tensor * tensor, const char * name);
+static void ggml_vk_check_results_0(ggml_tensor * tensor);
+static void ggml_vk_check_results_1(ggml_tensor * tensor);
+#endif
+
+typedef void (*ggml_vk_func_t)(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst);
+
+static void ggml_backend_vk_free(ggml_backend_t backend);
+
+// variables to track number of compiles in progress
+static uint32_t compile_count = 0;
+static std::mutex compile_count_mutex;
+static std::condition_variable compile_count_cond;
+
+static void ggml_vk_create_pipeline_func(vk_device& device, vk_pipeline& pipeline, const std::string name, size_t spv_size, const void* spv_data, const std::string entrypoint, uint32_t parameter_count, uint32_t push_constant_size, std::array<uint32_t, 3> wg_denoms, std::vector<uint32_t> specialization_constants, uint32_t align) {
+    VK_LOG_DEBUG("ggml_vk_create_pipeline(" << device->name << ", " << name << ", " << entrypoint << ", " << parameter_count << ", " << push_constant_size << ", (" << wg_denoms[0] << "," << wg_denoms[1] << "," << wg_denoms[2] << "), specialization_constants, " << align << ")");
+    GGML_ASSERT(parameter_count > 0);
+    GGML_ASSERT(wg_denoms[0] > 0 && wg_denoms[1] > 0 && wg_denoms[2] > 0); // NOLINT
+
+    pipeline = std::make_shared<vk_pipeline_struct>();
+    pipeline->name = name;
+    pipeline->parameter_count = parameter_count;
+    pipeline->push_constant_size = push_constant_size;
+    pipeline->wg_denoms = wg_denoms;
+    pipeline->align = align;
+
+    vk::ShaderModuleCreateInfo shader_module_create_info({}, spv_size, reinterpret_cast<const uint32_t *>(spv_data));
+    pipeline->shader_module = device->device.createShaderModule(shader_module_create_info);
+
+    std::vector<vk::DescriptorSetLayoutBinding> dsl_binding;
+    std::vector<vk::DescriptorBindingFlags> dsl_binding_flags;
+    for (uint32_t i = 0; i < parameter_count; i++) {
+        dsl_binding.push_back({i, vk::DescriptorType::eStorageBuffer, 1, vk::ShaderStageFlagBits::eCompute});
+        dsl_binding_flags.push_back({});
+    }
+
+    vk::DescriptorSetLayoutBindingFlagsCreateInfo dslbfci = { dsl_binding_flags };
+
+    vk::PushConstantRange pcr(
+        vk::ShaderStageFlagBits::eCompute,
+        0,
+        pipeline->push_constant_size
+    );
+
+    vk::DescriptorSetLayoutCreateInfo descriptor_set_layout_create_info(
+        {},
+        dsl_binding);
+    descriptor_set_layout_create_info.setPNext(&dslbfci);
+    pipeline->dsl = device->device.createDescriptorSetLayout(descriptor_set_layout_create_info);
+
+    vk::DescriptorPoolSize descriptor_pool_size(vk::DescriptorType::eStorageBuffer, pipeline->parameter_count * VK_DEVICE_DESCRIPTOR_POOL_SIZE);
+    vk::DescriptorPoolCreateInfo descriptor_pool_create_info({}, VK_DEVICE_DESCRIPTOR_POOL_SIZE, descriptor_pool_size);
+    pipeline->descriptor_pools.push_back(device->device.createDescriptorPool(descriptor_pool_create_info));
+
+    pipeline->descriptor_set_idx = 0;
+
+    vk::PipelineLayoutCreateInfo pipeline_layout_create_info(vk::PipelineLayoutCreateFlags(), pipeline->dsl, pcr);
+    pipeline->layout = device->device.createPipelineLayout(pipeline_layout_create_info);
+
+    std::vector<vk::SpecializationMapEntry> specialization_entries(specialization_constants.size());
+
+    for (size_t i = 0; i < specialization_constants.size(); i++) {
+        specialization_entries[i].constantID = i;
+        specialization_entries[i].offset = i * sizeof(uint32_t);
+        specialization_entries[i].size = sizeof(uint32_t);
+    }
+
+    vk::SpecializationInfo specialization_info(
+        specialization_entries.size(),
+        specialization_entries.data(),
+        specialization_constants.size() * sizeof(uint32_t),
+        specialization_constants.data()
+    );
+
+    vk::PipelineShaderStageCreateInfo pipeline_shader_create_info(
+            vk::PipelineShaderStageCreateFlags(),
+            vk::ShaderStageFlagBits::eCompute,
+            pipeline->shader_module,
+            entrypoint.c_str(),
+            &specialization_info);
+    vk::ComputePipelineCreateInfo compute_pipeline_create_info(
+        vk::PipelineCreateFlags(),
+        pipeline_shader_create_info,
+        pipeline->layout);
+    pipeline->pipeline = device->device.createComputePipeline(VK_NULL_HANDLE, compute_pipeline_create_info).value;
+
+    {
+        std::lock_guard<std::mutex> guard(device->mutex);
+        device->pipelines.insert({ pipeline->name, pipeline });
+    }
+
+    {
+        std::lock_guard<std::mutex> guard(compile_count_mutex);
+        assert(compile_count > 0);
+        compile_count--;
+    }
+    compile_count_cond.notify_all();
+}
+
+static void ggml_vk_destroy_pipeline(vk::Device& device, vk_pipeline& pipeline) {
+    VK_LOG_DEBUG("ggml_pipeline_destroy_pipeline(" << pipeline->name << ")");
+    for (auto& pool : pipeline->descriptor_pools) {
+        device.destroyDescriptorPool(pool);
+    }
+    pipeline->descriptor_pools.clear();
+    pipeline->descriptor_sets.clear();
+    pipeline->descriptor_set_idx = 0;
+
+    device.destroyDescriptorSetLayout(pipeline->dsl);
+
+    device.destroyPipelineLayout(pipeline->layout);
+
+    device.destroyShaderModule(pipeline->shader_module);
+
+    device.destroyPipeline(pipeline->pipeline);
+}
+
+static void ggml_pipeline_request_descriptor_sets(vk_device& device, vk_pipeline& pipeline, uint32_t n) {
+    VK_LOG_DEBUG("ggml_pipeline_request_descriptor_sets(" << pipeline->name << ", " << n << ")");
+    device->pipeline_descriptor_set_requirements[pipeline->name] += n;
+}
+
+static void ggml_pipeline_allocate_descriptor_sets(vk_device& device) {
+    std::lock_guard<std::mutex> guard(device->mutex);
+
+    for (auto& pair : device->pipeline_descriptor_set_requirements) {
+        vk_pipeline pipeline = device->pipelines.at(pair.first).lock();
+        const uint64_t n = pair.second;
+
+        VK_LOG_DEBUG("ggml_pipeline_allocate_descriptor_sets(" << pipeline->name << ", " << n << ")");
+
+        if (pipeline->descriptor_sets.size() >= pipeline->descriptor_set_idx + n) {
+            // Enough descriptors are available
+            continue;
+        }
+
+        uint32_t to_alloc = pipeline->descriptor_set_idx + n - pipeline->descriptor_sets.size();
+        uint32_t pool_remaining = VK_DEVICE_DESCRIPTOR_POOL_SIZE - pipeline->descriptor_sets.size() % VK_DEVICE_DESCRIPTOR_POOL_SIZE;
+        uint32_t pool_idx = pipeline->descriptor_sets.size() / VK_DEVICE_DESCRIPTOR_POOL_SIZE;
+
+        while (to_alloc > 0) {
+            const uint32_t alloc_count = std::min(pool_remaining, to_alloc);
+            to_alloc -= alloc_count;
+            pool_remaining = VK_DEVICE_DESCRIPTOR_POOL_SIZE;
+
+            if (pool_idx >= pipeline->descriptor_pools.size()) {
+                vk::DescriptorPoolSize descriptor_pool_size(vk::DescriptorType::eStorageBuffer, pipeline->parameter_count * VK_DEVICE_DESCRIPTOR_POOL_SIZE);
+                vk::DescriptorPoolCreateInfo descriptor_pool_create_info({}, VK_DEVICE_DESCRIPTOR_POOL_SIZE, descriptor_pool_size);
+                pipeline->descriptor_pools.push_back(device->device.createDescriptorPool(descriptor_pool_create_info));
+            }
+
+            std::vector<vk::DescriptorSetLayout> layouts(alloc_count);
+            for (uint32_t i = 0; i < alloc_count; i++) {
+                layouts[i] = pipeline->dsl;
+            }
+            vk::DescriptorSetAllocateInfo descriptor_set_alloc_info(pipeline->descriptor_pools[pool_idx], alloc_count, layouts.data());
+            std::vector<vk::DescriptorSet> sets = device->device.allocateDescriptorSets(descriptor_set_alloc_info);
+            pipeline->descriptor_sets.insert(pipeline->descriptor_sets.end(), sets.begin(), sets.end());
+
+            pool_idx++;
+        }
+    }
+}
+
+static void ggml_pipeline_cleanup(vk_pipeline& pipeline) {
+    VK_LOG_DEBUG("ggml_pipeline_cleanup(" << pipeline->name << ")");
+    pipeline->descriptor_set_idx = 0;
+}
+
+static vk::CommandBuffer ggml_vk_create_cmd_buffer(vk_device& device, vk_queue& q) {
+    VK_LOG_DEBUG("ggml_vk_create_cmd_buffer()");
+    std::lock_guard<std::mutex> guard(device->mutex);
+
+    if (q.cmd_buffers.size() > q.cmd_buffer_idx) {
+        // Reuse command buffer
+        return q.cmd_buffers[q.cmd_buffer_idx++];
+    }
+
+    vk::CommandBufferAllocateInfo command_buffer_alloc_info(
+        q.pool,
+        vk::CommandBufferLevel::ePrimary,
+        1);
+    const std::vector<vk::CommandBuffer> cmd_buffers = device->device.allocateCommandBuffers(command_buffer_alloc_info);
+    auto buf = cmd_buffers.front();
+
+    q.cmd_buffers.push_back(buf);
+    q.cmd_buffer_idx++;
+
+    return buf;
+}
+
+static vk_submission ggml_vk_create_submission(vk_device& device, vk_queue& q, std::vector<vk_semaphore> wait_semaphores, std::vector<vk_semaphore> signal_semaphores) {
+    VK_LOG_DEBUG("ggml_vk_create_submission()");
+    vk_submission s;
+    s.buffer = ggml_vk_create_cmd_buffer(device, q);
+    s.wait_semaphores = std::move(wait_semaphores);
+    s.signal_semaphores = std::move(signal_semaphores);
+    return s;
+}
+
+static void ggml_vk_submit(vk_context& ctx, vk::Fence fence) {
+    if (ctx->seqs.empty()) {
+        if (fence) {
+            ctx->q->queue.submit({}, fence);
+        }
+        return;
+    }
+    VK_LOG_DEBUG("ggml_vk_submit(" << ctx << ", " << fence << ")");
+
+    std::vector<std::vector<uint64_t>> tl_wait_vals;
+    std::vector<std::vector<uint64_t>> tl_signal_vals;
+    std::vector<std::vector<vk::Semaphore>> tl_wait_semaphores;
+    std::vector<std::vector<vk::Semaphore>> tl_signal_semaphores;
+    std::vector<vk::TimelineSemaphoreSubmitInfo> tl_submit_infos;
+    std::vector<vk::SubmitInfo> submit_infos;
+    int idx = -1;
+    std::vector<std::vector<vk::PipelineStageFlags>> stage_flags;
+
+    size_t reserve = 0;
+
+    for (const auto& sequence : ctx->seqs) {
+        reserve += sequence.size();
+    }
+
+    // Pre-reserve vectors to prevent reallocation, which invalidates pointers
+    tl_wait_semaphores.reserve(reserve);
+    tl_wait_vals.reserve(reserve);
+    tl_signal_semaphores.reserve(reserve);
+    tl_signal_vals.reserve(reserve);
+    tl_submit_infos.reserve(reserve);
+    submit_infos.reserve(reserve);
+    stage_flags.reserve(reserve);
+
+    for (const auto& sequence : ctx->seqs) {
+        for (const auto& submission : sequence) {
+            stage_flags.push_back({});
+            idx++;
+            tl_wait_vals.push_back({});
+            tl_wait_semaphores.push_back({});
+            tl_signal_vals.push_back({});
+            tl_signal_semaphores.push_back({});
+            for (size_t i = 0; i < submission.wait_semaphores.size(); i++) {
+                stage_flags[idx].push_back(ctx->q->stage_flags);
+                tl_wait_vals[idx].push_back(submission.wait_semaphores[i].value);
+                tl_wait_semaphores[idx].push_back(submission.wait_semaphores[i].s);
+            }
+            for (size_t i = 0; i < submission.signal_semaphores.size(); i++) {
+                tl_signal_vals[idx].push_back(submission.signal_semaphores[i].value);
+                tl_signal_semaphores[idx].push_back(submission.signal_semaphores[i].s);
+            }
+            tl_submit_infos.push_back({
+                (uint32_t) submission.wait_semaphores.size(),
+                tl_wait_vals[idx].data(),
+                (uint32_t) submission.signal_semaphores.size(),
+                tl_signal_vals[idx].data(),
+            });
+            tl_submit_infos[idx].sType = vk::StructureType::eTimelineSemaphoreSubmitInfo;
+            tl_submit_infos[idx].pNext = nullptr;
+            vk::SubmitInfo si{
+                (uint32_t) submission.wait_semaphores.size(),
+                tl_wait_semaphores[idx].data(),
+                stage_flags[idx].data(),
+                1,
+                &submission.buffer,
+                (uint32_t) submission.signal_semaphores.size(),
+                tl_signal_semaphores[idx].data(),
+            };
+            si.setPNext(&tl_submit_infos[idx]);
+            submit_infos.push_back(si);
+        }
+    }
+
+    ctx->q->queue.submit(submit_infos, fence);
+
+    ctx->seqs.clear();
+}
+
+static uint32_t ggml_vk_find_queue_family_index(std::vector<vk::QueueFamilyProperties>& queue_family_props, const vk::QueueFlags& required, const vk::QueueFlags& avoid, int32_t compute_index, uint32_t min_num_queues) {
+    VK_LOG_DEBUG("ggml_vk_find_queue_family_index()");
+    const uint32_t qfsize = queue_family_props.size();
+
+    // Try with avoid preferences first
+    for (uint32_t i = 0; i < qfsize; i++) {
+        if (queue_family_props[i].queueCount >= min_num_queues && (compute_index < 0 || i != (uint32_t) compute_index) && queue_family_props[i].queueFlags & required && !(queue_family_props[i].queueFlags & avoid)) {
+            return i;
+        }
+    }
+
+    // Fall back to only required
+    for (size_t i = 0; i < qfsize; i++) {
+        if (queue_family_props[i].queueCount >= min_num_queues && (compute_index < 0 || i != (uint32_t) compute_index) && queue_family_props[i].queueFlags & required) {
+            return i;
+        }
+    }
+
+    // Fall back to reusing compute queue
+    for (size_t i = 0; i < qfsize; i++) {
+        if (queue_family_props[i].queueCount >= min_num_queues && queue_family_props[i].queueFlags & required) {
+            return i;
+        }
+    }
+
+    // Fall back to ignoring min_num_queries
+    for (size_t i = 0; i < qfsize; i++) {
+        if (queue_family_props[i].queueFlags & required) {
+            return i;
+        }
+    }
+
+    // All commands that are allowed on a queue that supports transfer operations are also allowed on a queue that supports either graphics or compute operations.
+    // Thus, if the capabilities of a queue family include VK_QUEUE_GRAPHICS_BIT or VK_QUEUE_COMPUTE_BIT, then reporting the VK_QUEUE_TRANSFER_BIT capability separately for that queue family is optional.
+    if (compute_index >= 0) {
+        return compute_index;
+    }
+
+    std::cerr << "ggml_vulkan: No suitable queue family index found." << std::endl;
+
+    for(auto &q_family : queue_family_props) {
+        std::cerr << "Queue number: "  + std::to_string(q_family.queueCount) << " flags: " + to_string(q_family.queueFlags) << std::endl;
+    }
+    abort();
+}
+
+static void ggml_vk_create_queue(vk_device& device, vk_queue& q, uint32_t queue_family_index, uint32_t queue_index, vk::PipelineStageFlags&& stage_flags, bool transfer_only) {
+    VK_LOG_DEBUG("ggml_vk_create_queue()");
+    std::lock_guard<std::mutex> guard(device->mutex);
+
+    q.queue_family_index = queue_family_index;
+    q.transfer_only = transfer_only;
+
+    vk::CommandPoolCreateInfo command_pool_create_info_compute(vk::CommandPoolCreateFlags(VK_COMMAND_POOL_CREATE_TRANSIENT_BIT), queue_family_index);
+    q.pool = device->device.createCommandPool(command_pool_create_info_compute);
+
+    q.cmd_buffer_idx = 0;
+
+    q.queue = device->device.getQueue(queue_family_index, queue_index);
+
+    q.stage_flags = stage_flags;
+}
+
+static vk_context ggml_vk_create_context(ggml_backend_vk_context * ctx, vk_queue& q) {
+    vk_context result = std::make_shared<vk_context_struct>();
+    VK_LOG_DEBUG("ggml_vk_create_context(" << result << ")");
+    ctx->gc.contexts.emplace_back(result);
+    result->q = &q;
+    return result;
+}
+
+static vk_context ggml_vk_create_temporary_context(vk_queue& q) {
+    vk_context result = std::make_shared<vk_context_struct>();
+    VK_LOG_DEBUG("ggml_vk_create_temporary_context(" << result << ")");
+    result->q = &q;
+    return result;
+}
+
+static vk_semaphore * ggml_vk_create_binary_semaphore(ggml_backend_vk_context * ctx) {
+    VK_LOG_DEBUG("ggml_vk_create_timeline_semaphore()");
+    vk::SemaphoreTypeCreateInfo tci{ vk::SemaphoreType::eBinary, 0 };
+    vk::SemaphoreCreateInfo ci{};
+    ci.setPNext(&tci);
+    vk::Semaphore semaphore = ctx->device->device.createSemaphore(ci);
+    ctx->gc.semaphores.push_back({ semaphore, 0 });
+    return &ctx->gc.semaphores[ctx->gc.semaphores.size() - 1];
+}
+
+static vk_semaphore * ggml_vk_create_timeline_semaphore(ggml_backend_vk_context * ctx) {
+    VK_LOG_DEBUG("ggml_vk_create_timeline_semaphore()");
+    if (ctx->semaphore_idx >= ctx->gc.tl_semaphores.size()) {
+        vk::SemaphoreTypeCreateInfo tci{ vk::SemaphoreType::eTimeline, 0 };
+        vk::SemaphoreCreateInfo ci{};
+        ci.setPNext(&tci);
+        vk::Semaphore semaphore = ctx->device->device.createSemaphore(ci);
+        ctx->gc.tl_semaphores.push_back({ semaphore, 0 });
+    }
+    return &ctx->gc.tl_semaphores[ctx->semaphore_idx++];
+}
+
+static vk::Event ggml_vk_create_event(ggml_backend_vk_context * ctx) {
+    if (ctx->event_idx >= ctx->gc.events.size()) {
+        ctx->gc.events.push_back(ctx->device->device.createEvent({}));
+    }
+    return ctx->gc.events[ctx->event_idx++];
+}
+
+static void ggml_vk_queue_cleanup(vk_device& device, vk_queue& q) {
+    VK_LOG_DEBUG("ggml_vk_queue_cleanup()");
+    std::lock_guard<std::mutex> guard(device->mutex);
+
+    // Requires command buffers to be done
+    device->device.resetCommandPool(q.pool);
+    q.cmd_buffer_idx = 0;
+}
+
+static uint32_t find_properties(const vk::PhysicalDeviceMemoryProperties* mem_props, vk::MemoryRequirements* mem_req, vk::MemoryPropertyFlags flags) {
+    for (uint32_t i = 0; i < mem_props->memoryTypeCount; ++i) {
+        vk::MemoryType memory_type = mem_props->memoryTypes[i];
+        if ((mem_req->memoryTypeBits & ((uint64_t)1 << i)) &&
+            (flags & memory_type.propertyFlags) == flags &&
+            mem_props->memoryHeaps[memory_type.heapIndex].size >= mem_req->size) {
+            return static_cast<int32_t>(i);
+        }
+    }
+    return UINT32_MAX;
+}
+
+static vk_buffer ggml_vk_create_buffer(vk_device& device, size_t size, vk::MemoryPropertyFlags req_flags, vk::MemoryPropertyFlags fallback_flags = vk::MemoryPropertyFlags(0)) {
+    VK_LOG_DEBUG("ggml_vk_create_buffer(" << device->name << ", " << size << ", " << to_string(req_flags) << ", " << to_string(fallback_flags) << ")");
+    if (size > device->max_memory_allocation_size) {
+        throw vk::OutOfDeviceMemoryError("Requested buffer size exceeds device memory allocation limit");
+    }
+
+    std::lock_guard<std::mutex> guard(device->mutex);
+
+    vk_buffer buf = std::make_shared<vk_buffer_struct>();
+
+    if (size == 0) {
+        buf->size = 0;
+        return buf;
+    }
+
+    buf->size = size;
+    vk::BufferCreateInfo buffer_create_info{
+        vk::BufferCreateFlags(),
+        size,
+        vk::BufferUsageFlagBits::eStorageBuffer | vk::BufferUsageFlagBits::eTransferSrc | vk::BufferUsageFlagBits::eTransferDst,
+        vk::SharingMode::eExclusive,
+        0,
+        nullptr,
+    };
+
+    buf->buffer = device->device.createBuffer(buffer_create_info);
+
+    vk::MemoryRequirements mem_req = device->device.getBufferMemoryRequirements(buf->buffer);
+
+    vk::PhysicalDeviceMemoryProperties mem_props = device->physical_device.getMemoryProperties();
+
+    uint32_t memory_type_index = UINT32_MAX;
+
+    memory_type_index = find_properties(&mem_props, &mem_req, req_flags);
+    buf->memory_property_flags = req_flags;
+
+    if (memory_type_index == UINT32_MAX && fallback_flags) {
+        memory_type_index = find_properties(&mem_props, &mem_req, fallback_flags);
+        buf->memory_property_flags = fallback_flags;
+    }
+
+    if (memory_type_index == UINT32_MAX) {
+        device->device.destroyBuffer(buf->buffer);
+        buf->size = 0;
+        throw vk::OutOfDeviceMemoryError("No suitable memory type found");
+    }
+
+    try {
+        buf->device_memory = device->device.allocateMemory({ mem_req.size, memory_type_index });
+    } catch (const vk::SystemError& e) {
+        if (buf->memory_property_flags != fallback_flags) {
+            // Try again with fallback flags
+            memory_type_index = find_properties(&mem_props, &mem_req, fallback_flags);
+            buf->memory_property_flags = fallback_flags;
+
+            try {
+                buf->device_memory = device->device.allocateMemory({ mem_req.size, memory_type_index });
+            }
+            catch (const vk::SystemError& e) {
+                device->device.destroyBuffer(buf->buffer);
+                buf->size = 0;
+                throw e;
+            }
+        } else {
+            // Out of Host/Device memory, clean up buffer
+            device->device.destroyBuffer(buf->buffer);
+            buf->size = 0;
+            throw e;
+        }
+    }
+    buf->ptr = nullptr;
+
+    if (buf->memory_property_flags & vk::MemoryPropertyFlagBits::eHostVisible) {
+        buf->ptr = device->device.mapMemory(buf->device_memory, 0, VK_WHOLE_SIZE);
+    }
+
+    device->device.bindBufferMemory(buf->buffer, buf->device_memory, 0);
+
+    buf->device = device;
+
+#ifdef GGML_VULKAN_MEMORY_DEBUG
+    device->memory_logger->log_allocation(buf, size);
+#endif
+
+    return buf;
+}
+
+static vk_buffer ggml_vk_create_buffer_check(vk_device& device, size_t size, vk::MemoryPropertyFlags req_flags, vk::MemoryPropertyFlags fallback_flags = vk::MemoryPropertyFlags(0)) {
+    try {
+        return ggml_vk_create_buffer(device, size, req_flags, fallback_flags);
+    } catch (const vk::SystemError& e) {
+        std::cerr << "ggml_vulkan: Memory allocation of size " << size << " failed." << std::endl;
+        std::cerr << "ggml_vulkan: " << e.what() << std::endl;
+        throw e;
+    }
+}
+
+static vk_buffer ggml_vk_create_buffer_device(vk_device& device, size_t size) {
+    vk_buffer buf;
+    try {
+        if (device->uma) {
+            // Fall back to host memory type
+            buf = ggml_vk_create_buffer(device, size, vk::MemoryPropertyFlagBits::eDeviceLocal, vk::MemoryPropertyFlagBits::eHostVisible | vk::MemoryPropertyFlagBits::eHostCoherent);
+        } else {
+            // use rebar if available, otherwise fallback to device only visible memory
+            buf = ggml_vk_create_buffer(device, size, vk::MemoryPropertyFlagBits::eDeviceLocal | vk::MemoryPropertyFlagBits::eHostVisible | vk::MemoryPropertyFlagBits::eHostCoherent, vk::MemoryPropertyFlagBits::eDeviceLocal);
+        }
+    } catch (const vk::SystemError& e) {
+        std::cerr << "ggml_vulkan: Device memory allocation of size " << size << " failed." << std::endl;
+        std::cerr << "ggml_vulkan: " << e.what() << std::endl;
+        throw e;
+    }
+
+    return buf;
+}
+
+static void ggml_vk_destroy_buffer(vk_buffer& buf) {
+    if (buf == nullptr) {
+        return;
+    }
+
+#ifdef GGML_VULKAN_MEMORY_DEBUG
+    if (buf->device != nullptr) {
+        buf->device->memory_logger->log_deallocation(buf);
+    }
+#endif
+
+    buf.reset();
+}
+
+static vk_subbuffer ggml_vk_subbuffer(vk_buffer& buf) {
+    return { buf, 0, VK_WHOLE_SIZE };
+}
+
+static void ggml_vk_sync_buffers(vk_context& ctx) {
+    VK_LOG_DEBUG("ggml_vk_sync_buffers()");
+
+    const bool transfer_queue = ctx->q->transfer_only;
+
+    ctx->s->buffer.pipelineBarrier(
+        ctx->q->stage_flags,
+        ctx->q->stage_flags,
+        {},
+        { {
+          { !transfer_queue ? (vk::AccessFlagBits::eShaderRead | vk::AccessFlagBits::eShaderWrite | vk::AccessFlagBits::eTransferRead | vk::AccessFlagBits::eTransferWrite) : (vk::AccessFlagBits::eTransferRead | vk::AccessFlagBits::eTransferWrite) },
+          { !transfer_queue ? (vk::AccessFlagBits::eShaderRead | vk::AccessFlagBits::eShaderWrite | vk::AccessFlagBits::eTransferRead | vk::AccessFlagBits::eTransferWrite) : (vk::AccessFlagBits::eTransferRead | vk::AccessFlagBits::eTransferWrite) }
+        } },
+        {},
+        {}
+    );
+}
+
+static void ggml_vk_wait_events(vk_context& ctx, std::vector<vk::Event>&& events) {
+    VK_LOG_DEBUG("ggml_vk_wait_events()");
+    if (events.empty()) {
+        return;
+    }
+
+    ctx->s->buffer.waitEvents(
+        events,
+        ctx->q->stage_flags,
+        ctx->q->stage_flags,
+        {},
+        {},
+        {}
+    );
+}
+
+static void ggml_vk_load_shaders(vk_device& device) {
+    VK_LOG_DEBUG("ggml_vk_load_shaders(" << device->name << ")");
+
+    // mulmat
+    std::initializer_list<uint32_t> warptile_l = { 128, 128, 128, 16, device->subgroup_size * 2, 64, 2, 4, 4, device->subgroup_size };
+    std::initializer_list<uint32_t> warptile_m = { 128,  64,  64, 16, device->subgroup_size, 32, 2, 4, 2, device->subgroup_size };
+    std::initializer_list<uint32_t> warptile_s = { std::max(device->subgroup_size, 16u),  32,  32, 16, 32, 32, 2, 2, 2, device->subgroup_size };
+
+    std::initializer_list<uint32_t> warptile_mmq_l = { 128, 128, 128, 32, device->subgroup_size * 2, 64, 2, 4, 4, device->subgroup_size };
+    std::initializer_list<uint32_t> warptile_mmq_m = { 128,  64,  64, 32, device->subgroup_size, 32, 2, 4, 2, device->subgroup_size };
+    std::initializer_list<uint32_t> warptile_mmq_s = { std::max(device->subgroup_size, 16u),  32,  32, 32, 32, 32, 2, 2, 2, device->subgroup_size };
+
+    std::array<uint32_t, 3> l_wg_denoms = {128, 128, 1 };
+    std::array<uint32_t, 3> m_wg_denoms = { 64,  64, 1 };
+    std::array<uint32_t, 3> s_wg_denoms = { 32,  32, 1 };
+
+    uint32_t l_align = 128;
+    uint32_t m_align =  64;
+    uint32_t s_align =  32;
+
+    device->pipeline_matmul_f32 = std::make_shared<vk_matmul_pipeline_struct>();
+    device->pipeline_matmul_f32_f16 = std::make_shared<vk_matmul_pipeline_struct>();
+    device->pipeline_matmul_f16_f32 = std::make_shared<vk_matmul_pipeline_struct>();
+    device->pipeline_matmul_f16 = std::make_shared<vk_matmul_pipeline_struct>();
+    device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_0] = std::make_shared<vk_matmul_pipeline_struct>();
+    device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_1] = std::make_shared<vk_matmul_pipeline_struct>();
+    device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_0] = std::make_shared<vk_matmul_pipeline_struct>();
+    device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_1] = std::make_shared<vk_matmul_pipeline_struct>();
+    device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q8_0] = std::make_shared<vk_matmul_pipeline_struct>();
+    device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q2_K] = std::make_shared<vk_matmul_pipeline_struct>();
+    device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q3_K] = std::make_shared<vk_matmul_pipeline_struct>();
+    device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_K] = std::make_shared<vk_matmul_pipeline_struct>();
+    device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_K] = std::make_shared<vk_matmul_pipeline_struct>();
+    device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q6_K] = std::make_shared<vk_matmul_pipeline_struct>();
+    device->pipeline_dequant_mul_mat_mat[GGML_TYPE_IQ4_NL] = std::make_shared<vk_matmul_pipeline_struct>();
+
+    device->pipeline_matmul_id_f32 = std::make_shared<vk_matmul_pipeline_struct>();
+    device->pipeline_matmul_id_f16_f32 = std::make_shared<vk_matmul_pipeline_struct>();
+    device->pipeline_matmul_id_f16 = std::make_shared<vk_matmul_pipeline_struct>();
+    device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_0] = std::make_shared<vk_matmul_pipeline_struct>();
+    device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_1] = std::make_shared<vk_matmul_pipeline_struct>();
+    device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_0] = std::make_shared<vk_matmul_pipeline_struct>();
+    device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_1] = std::make_shared<vk_matmul_pipeline_struct>();
+    device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q8_0] = std::make_shared<vk_matmul_pipeline_struct>();
+    device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q2_K] = std::make_shared<vk_matmul_pipeline_struct>();
+    device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q3_K] = std::make_shared<vk_matmul_pipeline_struct>();
+    device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_K] = std::make_shared<vk_matmul_pipeline_struct>();
+    device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_K] = std::make_shared<vk_matmul_pipeline_struct>();
+    device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q6_K] = std::make_shared<vk_matmul_pipeline_struct>();
+    device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ4_NL] = std::make_shared<vk_matmul_pipeline_struct>();
+
+    std::vector<std::future<void>> compiles;
+    auto const &ggml_vk_create_pipeline = [&](vk_device& device, vk_pipeline& pipeline, const std::string &name, size_t spv_size, const void* spv_data, const std::string &entrypoint, uint32_t parameter_count, uint32_t push_constant_size, std::array<uint32_t, 3> wg_denoms, std::vector<uint32_t>&& specialization_constants, uint32_t align) {
+        {
+            // wait until fewer than N compiles are in progress
+            uint32_t N = std::max(1u, std::thread::hardware_concurrency());
+            std::unique_lock<std::mutex> guard(compile_count_mutex);
+            while (compile_count >= N) {
+                compile_count_cond.wait(guard);
+            }
+            compile_count++;
+        }
+        compiles.push_back(std::async(ggml_vk_create_pipeline_func, std::ref(device), std::ref(pipeline), name, spv_size, spv_data, entrypoint, parameter_count, push_constant_size, wg_denoms, specialization_constants, align));
+    };
+
+    if (device->fp16) {
+        ggml_vk_create_pipeline(device, device->pipeline_matmul_f32->l, "matmul_f32_l", matmul_f32_f32_len, matmul_f32_f32_data, "main", 3, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_l, 1);
+        ggml_vk_create_pipeline(device, device->pipeline_matmul_f32->m, "matmul_f32_m", matmul_f32_f32_len, matmul_f32_f32_data, "main", 3, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_m, 1);
+        ggml_vk_create_pipeline(device, device->pipeline_matmul_f32->s, "matmul_f32_s", matmul_f32_f32_len, matmul_f32_f32_data, "main", 3, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_s, 1);
+        ggml_vk_create_pipeline(device, device->pipeline_matmul_f32->a_l, "matmul_f32_aligned_l", matmul_f32_f32_aligned_len, matmul_f32_f32_aligned_data, "main", 3, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_l, l_align);
+        ggml_vk_create_pipeline(device, device->pipeline_matmul_f32->a_m, "matmul_f32_aligned_m", matmul_f32_f32_aligned_len, matmul_f32_f32_aligned_data, "main", 3, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_m, m_align);
+        ggml_vk_create_pipeline(device, device->pipeline_matmul_f32->a_s, "matmul_f32_aligned_s", matmul_f32_f32_aligned_len, matmul_f32_f32_aligned_data, "main", 3, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_s, s_align);
+
+        ggml_vk_create_pipeline(device, device->pipeline_matmul_f32_f16->l, "matmul_f32_f16_l", matmul_f32_f16_len, matmul_f32_f16_data, "main", 3, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_l, 1);
+        ggml_vk_create_pipeline(device, device->pipeline_matmul_f32_f16->m, "matmul_f32_f16_m", matmul_f32_f16_len, matmul_f32_f16_data, "main", 3, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_m, 1);
+        ggml_vk_create_pipeline(device, device->pipeline_matmul_f32_f16->s, "matmul_f32_f16_s", matmul_f32_f16_len, matmul_f32_f16_data, "main", 3, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_s, 1);
+        ggml_vk_create_pipeline(device, device->pipeline_matmul_f32_f16->a_l, "matmul_f32_f16_aligned_l", matmul_f32_f16_aligned_len, matmul_f32_f16_aligned_data, "main", 3, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_l, l_align);
+        ggml_vk_create_pipeline(device, device->pipeline_matmul_f32_f16->a_m, "matmul_f32_f16_aligned_m", matmul_f32_f16_aligned_len, matmul_f32_f16_aligned_data, "main", 3, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_m, m_align);
+        ggml_vk_create_pipeline(device, device->pipeline_matmul_f32_f16->a_s, "matmul_f32_f16_aligned_s", matmul_f32_f16_aligned_len, matmul_f32_f16_aligned_data, "main", 3, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_s, s_align);
+
+        ggml_vk_create_pipeline(device, device->pipeline_matmul_f16->l, "matmul_f16_l", matmul_f16_len, matmul_f16_data, "main", 3, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_l, 1);
+        ggml_vk_create_pipeline(device, device->pipeline_matmul_f16->m, "matmul_f16_m", matmul_f16_len, matmul_f16_data, "main", 3, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_m, 1);
+        ggml_vk_create_pipeline(device, device->pipeline_matmul_f16->s, "matmul_f16_s", matmul_f16_len, matmul_f16_data, "main", 3, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_s, 1);
+        ggml_vk_create_pipeline(device, device->pipeline_matmul_f16->a_l, "matmul_f16_aligned_l", matmul_f16_aligned_len, matmul_f16_aligned_data, "main", 3, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_l, l_align);
+        ggml_vk_create_pipeline(device, device->pipeline_matmul_f16->a_m, "matmul_f16_aligned_m", matmul_f16_aligned_len, matmul_f16_aligned_data, "main", 3, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_m, m_align);
+        ggml_vk_create_pipeline(device, device->pipeline_matmul_f16->a_s, "matmul_f16_aligned_s", matmul_f16_aligned_len, matmul_f16_aligned_data, "main", 3, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_s, s_align);
+
+        ggml_vk_create_pipeline(device, device->pipeline_matmul_f16_f32->l, "matmul_f16_f32_l", matmul_f16_f32_len, matmul_f16_f32_data, "main", 3, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_l, 1);
+        ggml_vk_create_pipeline(device, device->pipeline_matmul_f16_f32->m, "matmul_f16_f32_m", matmul_f16_f32_len, matmul_f16_f32_data, "main", 3, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_m, 1);
+        ggml_vk_create_pipeline(device, device->pipeline_matmul_f16_f32->s, "matmul_f16_f32_s", matmul_f16_f32_len, matmul_f16_f32_data, "main", 3, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_s, 1);
+        ggml_vk_create_pipeline(device, device->pipeline_matmul_f16_f32->a_l, "matmul_f16_f32_aligned_l", matmul_f16_f32_aligned_len, matmul_f16_f32_aligned_data, "main", 3, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_l, l_align);
+        ggml_vk_create_pipeline(device, device->pipeline_matmul_f16_f32->a_m, "matmul_f16_f32_aligned_m", matmul_f16_f32_aligned_len, matmul_f16_f32_aligned_data, "main", 3, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_m, m_align);
+        ggml_vk_create_pipeline(device, device->pipeline_matmul_f16_f32->a_s, "matmul_f16_f32_aligned_s", matmul_f16_f32_aligned_len, matmul_f16_f32_aligned_data, "main", 3, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_s, s_align);
+
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_0]->l, "matmul_q4_0_f32_l", matmul_q4_0_f32_len, matmul_q4_0_f32_data, "main", 3, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_0]->m, "matmul_q4_0_f32_m", matmul_q4_0_f32_len, matmul_q4_0_f32_data, "main", 3, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_0]->s, "matmul_q4_0_f32_s", matmul_q4_0_f32_len, matmul_q4_0_f32_data, "main", 3, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_0]->a_l, "matmul_q4_0_f32_aligned_l", matmul_q4_0_f32_aligned_len, matmul_q4_0_f32_aligned_data, "main", 3, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_0]->a_m, "matmul_q4_0_f32_aligned_m", matmul_q4_0_f32_aligned_len, matmul_q4_0_f32_aligned_data, "main", 3, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_0]->a_s, "matmul_q4_0_f32_aligned_s", matmul_q4_0_f32_aligned_len, matmul_q4_0_f32_aligned_data, "main", 3, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
+
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_1]->l, "matmul_q4_1_f32_l", matmul_q4_1_f32_len, matmul_q4_1_f32_data, "main", 3, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_1]->m, "matmul_q4_1_f32_m", matmul_q4_1_f32_len, matmul_q4_1_f32_data, "main", 3, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_1]->s, "matmul_q4_1_f32_s", matmul_q4_1_f32_len, matmul_q4_1_f32_data, "main", 3, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_1]->a_l, "matmul_q4_1_f32_aligned_l", matmul_q4_1_f32_aligned_len, matmul_q4_1_f32_aligned_data, "main", 3, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_1]->a_m, "matmul_q4_1_f32_aligned_m", matmul_q4_1_f32_aligned_len, matmul_q4_1_f32_aligned_data, "main", 3, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_1]->a_s, "matmul_q4_1_f32_aligned_s", matmul_q4_1_f32_aligned_len, matmul_q4_1_f32_aligned_data, "main", 3, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
+
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_0]->l, "matmul_q5_0_f32_l", matmul_q5_0_f32_len, matmul_q5_0_f32_data, "main", 3, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_0]->m, "matmul_q5_0_f32_m", matmul_q5_0_f32_len, matmul_q5_0_f32_data, "main", 3, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_0]->s, "matmul_q5_0_f32_s", matmul_q5_0_f32_len, matmul_q5_0_f32_data, "main", 3, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_0]->a_l, "matmul_q5_0_f32_aligned_l", matmul_q5_0_f32_aligned_len, matmul_q5_0_f32_aligned_data, "main", 3, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_0]->a_m, "matmul_q5_0_f32_aligned_m", matmul_q5_0_f32_aligned_len, matmul_q5_0_f32_aligned_data, "main", 3, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_0]->a_s, "matmul_q5_0_f32_aligned_s", matmul_q5_0_f32_aligned_len, matmul_q5_0_f32_aligned_data, "main", 3, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
+
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_1]->l, "matmul_q5_1_f32_l", matmul_q5_1_f32_len, matmul_q5_1_f32_data, "main", 3, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_1]->m, "matmul_q5_1_f32_m", matmul_q5_1_f32_len, matmul_q5_1_f32_data, "main", 3, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_1]->s, "matmul_q5_1_f32_s", matmul_q5_1_f32_len, matmul_q5_1_f32_data, "main", 3, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_1]->a_l, "matmul_q5_1_f32_aligned_l", matmul_q5_1_f32_aligned_len, matmul_q5_1_f32_aligned_data, "main", 3, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_1]->a_m, "matmul_q5_1_f32_aligned_m", matmul_q5_1_f32_aligned_len, matmul_q5_1_f32_aligned_data, "main", 3, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_1]->a_s, "matmul_q5_1_f32_aligned_s", matmul_q5_1_f32_aligned_len, matmul_q5_1_f32_aligned_data, "main", 3, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
+
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q8_0]->l, "matmul_q8_0_f32_l", matmul_q8_0_f32_len, matmul_q8_0_f32_data, "main", 3, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q8_0]->m, "matmul_q8_0_f32_m", matmul_q8_0_f32_len, matmul_q8_0_f32_data, "main", 3, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q8_0]->s, "matmul_q8_0_f32_s", matmul_q8_0_f32_len, matmul_q8_0_f32_data, "main", 3, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q8_0]->a_l, "matmul_q8_0_f32_aligned_l", matmul_q8_0_f32_aligned_len, matmul_q8_0_f32_aligned_data, "main", 3, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q8_0]->a_m, "matmul_q8_0_f32_aligned_m", matmul_q8_0_f32_aligned_len, matmul_q8_0_f32_aligned_data, "main", 3, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q8_0]->a_s, "matmul_q8_0_f32_aligned_s", matmul_q8_0_f32_aligned_len, matmul_q8_0_f32_aligned_data, "main", 3, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
+
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q2_K]->l, "matmul_q2_k_f32_l", matmul_q2_k_f32_len, matmul_q2_k_f32_data, "main", 3, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q2_K]->m, "matmul_q2_k_f32_m", matmul_q2_k_f32_len, matmul_q2_k_f32_data, "main", 3, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q2_K]->s, "matmul_q2_k_f32_s", matmul_q2_k_f32_len, matmul_q2_k_f32_data, "main", 3, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q2_K]->a_l, "matmul_q2_k_f32_aligned_l", matmul_q2_k_f32_aligned_len, matmul_q2_k_f32_aligned_data, "main", 3, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q2_K]->a_m, "matmul_q2_k_f32_aligned_m", matmul_q2_k_f32_aligned_len, matmul_q2_k_f32_aligned_data, "main", 3, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q2_K]->a_s, "matmul_q2_k_f32_aligned_s", matmul_q2_k_f32_aligned_len, matmul_q2_k_f32_aligned_data, "main", 3, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
+
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q3_K]->l, "matmul_q3_k_f32_l", matmul_q3_k_f32_len, matmul_q3_k_f32_data, "main", 3, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q3_K]->m, "matmul_q3_k_f32_m", matmul_q3_k_f32_len, matmul_q3_k_f32_data, "main", 3, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q3_K]->s, "matmul_q3_k_f32_s", matmul_q3_k_f32_len, matmul_q3_k_f32_data, "main", 3, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q3_K]->a_l, "matmul_q3_k_f32_aligned_l", matmul_q3_k_f32_aligned_len, matmul_q3_k_f32_aligned_data, "main", 3, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q3_K]->a_m, "matmul_q3_k_f32_aligned_m", matmul_q3_k_f32_aligned_len, matmul_q3_k_f32_aligned_data, "main", 3, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q3_K]->a_s, "matmul_q3_k_f32_aligned_s", matmul_q3_k_f32_aligned_len, matmul_q3_k_f32_aligned_data, "main", 3, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
+
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_K]->l, "matmul_q4_k_f32_l", matmul_q4_k_f32_len, matmul_q4_k_f32_data, "main", 3, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_K]->m, "matmul_q4_k_f32_m", matmul_q4_k_f32_len, matmul_q4_k_f32_data, "main", 3, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_K]->s, "matmul_q4_k_f32_s", matmul_q4_k_f32_len, matmul_q4_k_f32_data, "main", 3, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_K]->a_l, "matmul_q4_k_f32_aligned_l", matmul_q4_k_f32_aligned_len, matmul_q4_k_f32_aligned_data, "main", 3, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_K]->a_m, "matmul_q4_k_f32_aligned_m", matmul_q4_k_f32_aligned_len, matmul_q4_k_f32_aligned_data, "main", 3, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_K]->a_s, "matmul_q4_k_f32_aligned_s", matmul_q4_k_f32_aligned_len, matmul_q4_k_f32_aligned_data, "main", 3, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
+
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_K]->l, "matmul_q5_k_f32_l", matmul_q5_k_f32_len, matmul_q5_k_f32_data, "main", 3, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_K]->m, "matmul_q5_k_f32_m", matmul_q5_k_f32_len, matmul_q5_k_f32_data, "main", 3, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_K]->s, "matmul_q5_k_f32_s", matmul_q5_k_f32_len, matmul_q5_k_f32_data, "main", 3, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_K]->a_l, "matmul_q5_k_f32_aligned_l", matmul_q5_k_f32_aligned_len, matmul_q5_k_f32_aligned_data, "main", 3, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_K]->a_m, "matmul_q5_k_f32_aligned_m", matmul_q5_k_f32_aligned_len, matmul_q5_k_f32_aligned_data, "main", 3, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_K]->a_s, "matmul_q5_k_f32_aligned_s", matmul_q5_k_f32_aligned_len, matmul_q5_k_f32_aligned_data, "main", 3, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
+
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q6_K]->l, "matmul_q6_k_f32_l", matmul_q6_k_f32_len, matmul_q6_k_f32_data, "main", 3, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q6_K]->m, "matmul_q6_k_f32_m", matmul_q6_k_f32_len, matmul_q6_k_f32_data, "main", 3, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q6_K]->s, "matmul_q6_k_f32_s", matmul_q6_k_f32_len, matmul_q6_k_f32_data, "main", 3, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q6_K]->a_l, "matmul_q6_k_f32_aligned_l", matmul_q6_k_f32_aligned_len, matmul_q6_k_f32_aligned_data, "main", 3, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q6_K]->a_m, "matmul_q6_k_f32_aligned_m", matmul_q6_k_f32_aligned_len, matmul_q6_k_f32_aligned_data, "main", 3, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q6_K]->a_s, "matmul_q6_k_f32_aligned_s", matmul_q6_k_f32_aligned_len, matmul_q6_k_f32_aligned_data, "main", 3, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
+
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_IQ4_NL]->l, "matmul_iq4_nl_f32_l", matmul_iq4_nl_f32_len, matmul_iq4_nl_f32_data, "main", 3, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_IQ4_NL]->m, "matmul_iq4_nl_f32_m", matmul_iq4_nl_f32_len, matmul_iq4_nl_f32_data, "main", 3, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_IQ4_NL]->s, "matmul_iq4_nl_f32_s", matmul_iq4_nl_f32_len, matmul_iq4_nl_f32_data, "main", 3, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_IQ4_NL]->a_l, "matmul_iq4_nl_f32_aligned_l", matmul_iq4_nl_f32_aligned_len, matmul_iq4_nl_f32_aligned_data, "main", 3, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_IQ4_NL]->a_m, "matmul_iq4_nl_f32_aligned_m", matmul_iq4_nl_f32_aligned_len, matmul_iq4_nl_f32_aligned_data, "main", 3, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_IQ4_NL]->a_s, "matmul_iq4_nl_f32_aligned_s", matmul_iq4_nl_f32_aligned_len, matmul_iq4_nl_f32_aligned_data, "main", 3, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
+
+        ggml_vk_create_pipeline(device, device->pipeline_matmul_id_f32->l, "matmul_id_f32_l", matmul_id_f32_f32_len, matmul_id_f32_f32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), l_wg_denoms, warptile_l, 1);
+        ggml_vk_create_pipeline(device, device->pipeline_matmul_id_f32->m, "matmul_id_f32_m", matmul_id_f32_f32_len, matmul_id_f32_f32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), m_wg_denoms, warptile_m, 1);
+        ggml_vk_create_pipeline(device, device->pipeline_matmul_id_f32->s, "matmul_id_f32_s", matmul_id_f32_f32_len, matmul_id_f32_f32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), s_wg_denoms, warptile_s, 1);
+        ggml_vk_create_pipeline(device, device->pipeline_matmul_id_f32->a_l, "matmul_id_f32_aligned_l", matmul_id_f32_f32_aligned_len, matmul_id_f32_f32_aligned_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), l_wg_denoms, warptile_l, l_align);
+        ggml_vk_create_pipeline(device, device->pipeline_matmul_id_f32->a_m, "matmul_id_f32_aligned_m", matmul_id_f32_f32_aligned_len, matmul_id_f32_f32_aligned_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), m_wg_denoms, warptile_m, m_align);
+        ggml_vk_create_pipeline(device, device->pipeline_matmul_id_f32->a_s, "matmul_id_f32_aligned_s", matmul_id_f32_f32_aligned_len, matmul_id_f32_f32_aligned_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), s_wg_denoms, warptile_s, s_align);
+
+        ggml_vk_create_pipeline(device, device->pipeline_matmul_id_f16->l, "matmul_id_f16_l", matmul_id_f16_len, matmul_id_f16_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), l_wg_denoms, warptile_l, 1);
+        ggml_vk_create_pipeline(device, device->pipeline_matmul_id_f16->m, "matmul_id_f16_m", matmul_id_f16_len, matmul_id_f16_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), m_wg_denoms, warptile_m, 1);
+        ggml_vk_create_pipeline(device, device->pipeline_matmul_id_f16->s, "matmul_id_f16_s", matmul_id_f16_len, matmul_id_f16_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), s_wg_denoms, warptile_s, 1);
+        ggml_vk_create_pipeline(device, device->pipeline_matmul_id_f16->a_l, "matmul_id_f16_aligned_l", matmul_id_f16_aligned_len, matmul_id_f16_aligned_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), l_wg_denoms, warptile_l, l_align);
+        ggml_vk_create_pipeline(device, device->pipeline_matmul_id_f16->a_m, "matmul_id_f16_aligned_m", matmul_id_f16_aligned_len, matmul_id_f16_aligned_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), m_wg_denoms, warptile_m, m_align);
+        ggml_vk_create_pipeline(device, device->pipeline_matmul_id_f16->a_s, "matmul_id_f16_aligned_s", matmul_id_f16_aligned_len, matmul_id_f16_aligned_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), s_wg_denoms, warptile_s, s_align);
+
+        ggml_vk_create_pipeline(device, device->pipeline_matmul_id_f16_f32->l, "matmul_id_f16_f32_l", matmul_id_f16_f32_len, matmul_id_f16_f32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), l_wg_denoms, warptile_l, 1);
+        ggml_vk_create_pipeline(device, device->pipeline_matmul_id_f16_f32->m, "matmul_id_f16_f32_m", matmul_id_f16_f32_len, matmul_id_f16_f32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), m_wg_denoms, warptile_m, 1);
+        ggml_vk_create_pipeline(device, device->pipeline_matmul_id_f16_f32->s, "matmul_id_f16_f32_s", matmul_id_f16_f32_len, matmul_id_f16_f32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), s_wg_denoms, warptile_s, 1);
+        ggml_vk_create_pipeline(device, device->pipeline_matmul_id_f16_f32->a_l, "matmul_id_f16_f32_aligned_l", matmul_id_f16_f32_aligned_len, matmul_id_f16_f32_aligned_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), l_wg_denoms, warptile_l, l_align);
+        ggml_vk_create_pipeline(device, device->pipeline_matmul_id_f16_f32->a_m, "matmul_id_f16_f32_aligned_m", matmul_id_f16_f32_aligned_len, matmul_id_f16_f32_aligned_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), m_wg_denoms, warptile_m, m_align);
+        ggml_vk_create_pipeline(device, device->pipeline_matmul_id_f16_f32->a_s, "matmul_id_f16_f32_aligned_s", matmul_id_f16_f32_aligned_len, matmul_id_f16_f32_aligned_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), s_wg_denoms, warptile_s, s_align);
+
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_0]->l, "matmul_id_q4_0_f32_l", matmul_id_q4_0_f32_len, matmul_id_q4_0_f32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_0]->m, "matmul_id_q4_0_f32_m", matmul_id_q4_0_f32_len, matmul_id_q4_0_f32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_0]->s, "matmul_id_q4_0_f32_s", matmul_id_q4_0_f32_len, matmul_id_q4_0_f32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_0]->a_l, "matmul_id_q4_0_f32_aligned_l", matmul_id_q4_0_f32_aligned_len, matmul_id_q4_0_f32_aligned_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_0]->a_m, "matmul_id_q4_0_f32_aligned_m", matmul_id_q4_0_f32_aligned_len, matmul_id_q4_0_f32_aligned_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_0]->a_s, "matmul_id_q4_0_f32_aligned_s", matmul_id_q4_0_f32_aligned_len, matmul_id_q4_0_f32_aligned_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
+
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_1]->l, "matmul_id_q4_1_f32_l", matmul_id_q4_1_f32_len, matmul_id_q4_1_f32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_1]->m, "matmul_id_q4_1_f32_m", matmul_id_q4_1_f32_len, matmul_id_q4_1_f32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_1]->s, "matmul_id_q4_1_f32_s", matmul_id_q4_1_f32_len, matmul_id_q4_1_f32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_1]->a_l, "matmul_id_q4_1_f32_aligned_l", matmul_id_q4_1_f32_aligned_len, matmul_id_q4_1_f32_aligned_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_1]->a_m, "matmul_id_q4_1_f32_aligned_m", matmul_id_q4_1_f32_aligned_len, matmul_id_q4_1_f32_aligned_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_1]->a_s, "matmul_id_q4_1_f32_aligned_s", matmul_id_q4_1_f32_aligned_len, matmul_id_q4_1_f32_aligned_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
+
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_0]->l, "matmul_id_q5_0_f32_l", matmul_id_q5_0_f32_len, matmul_id_q5_0_f32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_0]->m, "matmul_id_q5_0_f32_m", matmul_id_q5_0_f32_len, matmul_id_q5_0_f32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_0]->s, "matmul_id_q5_0_f32_s", matmul_id_q5_0_f32_len, matmul_id_q5_0_f32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_0]->a_l, "matmul_id_q5_0_f32_aligned_l", matmul_id_q5_0_f32_aligned_len, matmul_id_q5_0_f32_aligned_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_0]->a_m, "matmul_id_q5_0_f32_aligned_m", matmul_id_q5_0_f32_aligned_len, matmul_id_q5_0_f32_aligned_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_0]->a_s, "matmul_id_q5_0_f32_aligned_s", matmul_id_q5_0_f32_aligned_len, matmul_id_q5_0_f32_aligned_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
+
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_1]->l, "matmul_id_q5_1_f32_l", matmul_id_q5_1_f32_len, matmul_id_q5_1_f32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_1]->m, "matmul_id_q5_1_f32_m", matmul_id_q5_1_f32_len, matmul_id_q5_1_f32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_1]->s, "matmul_id_q5_1_f32_s", matmul_id_q5_1_f32_len, matmul_id_q5_1_f32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_1]->a_l, "matmul_id_q5_1_f32_aligned_l", matmul_id_q5_1_f32_aligned_len, matmul_id_q5_1_f32_aligned_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_1]->a_m, "matmul_id_q5_1_f32_aligned_m", matmul_id_q5_1_f32_aligned_len, matmul_id_q5_1_f32_aligned_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_1]->a_s, "matmul_id_q5_1_f32_aligned_s", matmul_id_q5_1_f32_aligned_len, matmul_id_q5_1_f32_aligned_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
+
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q8_0]->l, "matmul_id_q8_0_f32_l", matmul_id_q8_0_f32_len, matmul_id_q8_0_f32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q8_0]->m, "matmul_id_q8_0_f32_m", matmul_id_q8_0_f32_len, matmul_id_q8_0_f32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q8_0]->s, "matmul_id_q8_0_f32_s", matmul_id_q8_0_f32_len, matmul_id_q8_0_f32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q8_0]->a_l, "matmul_id_q8_0_f32_aligned_l", matmul_id_q8_0_f32_aligned_len, matmul_id_q8_0_f32_aligned_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q8_0]->a_m, "matmul_id_q8_0_f32_aligned_m", matmul_id_q8_0_f32_aligned_len, matmul_id_q8_0_f32_aligned_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q8_0]->a_s, "matmul_id_q8_0_f32_aligned_s", matmul_id_q8_0_f32_aligned_len, matmul_id_q8_0_f32_aligned_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
+
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q2_K]->l, "matmul_id_q2_k_f32_l", matmul_id_q2_k_f32_len, matmul_id_q2_k_f32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q2_K]->m, "matmul_id_q2_k_f32_m", matmul_id_q2_k_f32_len, matmul_id_q2_k_f32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q2_K]->s, "matmul_id_q2_k_f32_s", matmul_id_q2_k_f32_len, matmul_id_q2_k_f32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q2_K]->a_l, "matmul_id_q2_k_f32_aligned_l", matmul_id_q2_k_f32_aligned_len, matmul_id_q2_k_f32_aligned_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q2_K]->a_m, "matmul_id_q2_k_f32_aligned_m", matmul_id_q2_k_f32_aligned_len, matmul_id_q2_k_f32_aligned_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q2_K]->a_s, "matmul_id_q2_k_f32_aligned_s", matmul_id_q2_k_f32_aligned_len, matmul_id_q2_k_f32_aligned_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
+
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q3_K]->l, "matmul_id_q3_k_f32_l", matmul_id_q3_k_f32_len, matmul_id_q3_k_f32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q3_K]->m, "matmul_id_q3_k_f32_m", matmul_id_q3_k_f32_len, matmul_id_q3_k_f32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q3_K]->s, "matmul_id_q3_k_f32_s", matmul_id_q3_k_f32_len, matmul_id_q3_k_f32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q3_K]->a_l, "matmul_id_q3_k_f32_aligned_l", matmul_id_q3_k_f32_aligned_len, matmul_id_q3_k_f32_aligned_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q3_K]->a_m, "matmul_id_q3_k_f32_aligned_m", matmul_id_q3_k_f32_aligned_len, matmul_id_q3_k_f32_aligned_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q3_K]->a_s, "matmul_id_q3_k_f32_aligned_s", matmul_id_q3_k_f32_aligned_len, matmul_id_q3_k_f32_aligned_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
+
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_K]->l, "matmul_id_q4_k_f32_l", matmul_id_q4_k_f32_len, matmul_id_q4_k_f32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_K]->m, "matmul_id_q4_k_f32_m", matmul_id_q4_k_f32_len, matmul_id_q4_k_f32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_K]->s, "matmul_id_q4_k_f32_s", matmul_id_q4_k_f32_len, matmul_id_q4_k_f32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_K]->a_l, "matmul_id_q4_k_f32_aligned_l", matmul_id_q4_k_f32_aligned_len, matmul_id_q4_k_f32_aligned_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_K]->a_m, "matmul_id_q4_k_f32_aligned_m", matmul_id_q4_k_f32_aligned_len, matmul_id_q4_k_f32_aligned_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_K]->a_s, "matmul_id_q4_k_f32_aligned_s", matmul_id_q4_k_f32_aligned_len, matmul_id_q4_k_f32_aligned_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
+
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_K]->l, "matmul_id_q5_k_f32_l", matmul_id_q5_k_f32_len, matmul_id_q5_k_f32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_K]->m, "matmul_id_q5_k_f32_m", matmul_id_q5_k_f32_len, matmul_id_q5_k_f32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_K]->s, "matmul_id_q5_k_f32_s", matmul_id_q5_k_f32_len, matmul_id_q5_k_f32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_K]->a_l, "matmul_id_q5_k_f32_aligned_l", matmul_id_q5_k_f32_aligned_len, matmul_id_q5_k_f32_aligned_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_K]->a_m, "matmul_id_q5_k_f32_aligned_m", matmul_id_q5_k_f32_aligned_len, matmul_id_q5_k_f32_aligned_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_K]->a_s, "matmul_id_q5_k_f32_aligned_s", matmul_id_q5_k_f32_aligned_len, matmul_id_q5_k_f32_aligned_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
+
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q6_K]->l, "matmul_id_q6_k_f32_l", matmul_id_q6_k_f32_len, matmul_id_q6_k_f32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q6_K]->m, "matmul_id_q6_k_f32_m", matmul_id_q6_k_f32_len, matmul_id_q6_k_f32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q6_K]->s, "matmul_id_q6_k_f32_s", matmul_id_q6_k_f32_len, matmul_id_q6_k_f32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q6_K]->a_l, "matmul_id_q6_k_f32_aligned_l", matmul_id_q6_k_f32_aligned_len, matmul_id_q6_k_f32_aligned_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q6_K]->a_m, "matmul_id_q6_k_f32_aligned_m", matmul_id_q6_k_f32_aligned_len, matmul_id_q6_k_f32_aligned_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q6_K]->a_s, "matmul_id_q6_k_f32_aligned_s", matmul_id_q6_k_f32_aligned_len, matmul_id_q6_k_f32_aligned_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
+
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ4_NL]->l, "matmul_id_iq4_nl_f32_l", matmul_id_iq4_nl_f32_len, matmul_id_iq4_nl_f32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ4_NL]->m, "matmul_id_iq4_nl_f32_m", matmul_id_iq4_nl_f32_len, matmul_id_iq4_nl_f32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ4_NL]->s, "matmul_id_iq4_nl_f32_s", matmul_id_iq4_nl_f32_len, matmul_id_iq4_nl_f32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ4_NL]->a_l, "matmul_id_iq4_nl_f32_aligned_l", matmul_id_iq4_nl_f32_aligned_len, matmul_id_iq4_nl_f32_aligned_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ4_NL]->a_m, "matmul_id_iq4_nl_f32_aligned_m", matmul_id_iq4_nl_f32_aligned_len, matmul_id_iq4_nl_f32_aligned_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ4_NL]->a_s, "matmul_id_iq4_nl_f32_aligned_s", matmul_id_iq4_nl_f32_aligned_len, matmul_id_iq4_nl_f32_aligned_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
+    } else {
+        ggml_vk_create_pipeline(device, device->pipeline_matmul_f32->l, "matmul_f32_l", matmul_f32_f32_fp32_len, matmul_f32_f32_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_l, 1);
+        ggml_vk_create_pipeline(device, device->pipeline_matmul_f32->m, "matmul_f32_m", matmul_f32_f32_fp32_len, matmul_f32_f32_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_m, 1);
+        ggml_vk_create_pipeline(device, device->pipeline_matmul_f32->s, "matmul_f32_s", matmul_f32_f32_fp32_len, matmul_f32_f32_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_s, 1);
+        ggml_vk_create_pipeline(device, device->pipeline_matmul_f32->a_l, "matmul_f32_aligned_l", matmul_f32_f32_aligned_fp32_len, matmul_f32_f32_aligned_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_l, l_align);
+        ggml_vk_create_pipeline(device, device->pipeline_matmul_f32->a_m, "matmul_f32_aligned_m", matmul_f32_f32_aligned_fp32_len, matmul_f32_f32_aligned_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_m, m_align);
+        ggml_vk_create_pipeline(device, device->pipeline_matmul_f32->a_s, "matmul_f32_aligned_s", matmul_f32_f32_aligned_fp32_len, matmul_f32_f32_aligned_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_s, s_align);
+
+        ggml_vk_create_pipeline(device, device->pipeline_matmul_f32_f16->l, "matmul_f32_f16_l", matmul_f32_f16_fp32_len, matmul_f32_f16_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_l, 1);
+        ggml_vk_create_pipeline(device, device->pipeline_matmul_f32_f16->m, "matmul_f32_f16_m", matmul_f32_f16_fp32_len, matmul_f32_f16_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_m, 1);
+        ggml_vk_create_pipeline(device, device->pipeline_matmul_f32_f16->s, "matmul_f32_f16_s", matmul_f32_f16_fp32_len, matmul_f32_f16_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_s, 1);
+        ggml_vk_create_pipeline(device, device->pipeline_matmul_f32_f16->a_l, "matmul_f32_f16_aligned_l", matmul_f32_f16_aligned_fp32_len, matmul_f32_f16_aligned_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_l, l_align);
+        ggml_vk_create_pipeline(device, device->pipeline_matmul_f32_f16->a_m, "matmul_f32_f16_aligned_m", matmul_f32_f16_aligned_fp32_len, matmul_f32_f16_aligned_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_m, m_align);
+        ggml_vk_create_pipeline(device, device->pipeline_matmul_f32_f16->a_s, "matmul_f32_f16_aligned_s", matmul_f32_f16_aligned_fp32_len, matmul_f32_f16_aligned_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_s, s_align);
+
+        ggml_vk_create_pipeline(device, device->pipeline_matmul_f16->l, "matmul_f16_l", matmul_f16_fp32_len, matmul_f16_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_l, 1);
+        ggml_vk_create_pipeline(device, device->pipeline_matmul_f16->m, "matmul_f16_m", matmul_f16_fp32_len, matmul_f16_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_m, 1);
+        ggml_vk_create_pipeline(device, device->pipeline_matmul_f16->s, "matmul_f16_s", matmul_f16_fp32_len, matmul_f16_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_s, 1);
+        ggml_vk_create_pipeline(device, device->pipeline_matmul_f16->a_l, "matmul_f16_aligned_l", matmul_f16_aligned_fp32_len, matmul_f16_aligned_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_l, l_align);
+        ggml_vk_create_pipeline(device, device->pipeline_matmul_f16->a_m, "matmul_f16_aligned_m", matmul_f16_aligned_fp32_len, matmul_f16_aligned_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_m, m_align);
+        ggml_vk_create_pipeline(device, device->pipeline_matmul_f16->a_s, "matmul_f16_aligned_s", matmul_f16_aligned_fp32_len, matmul_f16_aligned_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_s, s_align);
+
+        ggml_vk_create_pipeline(device, device->pipeline_matmul_f16_f32->l, "matmul_f16_f32_l", matmul_f16_f32_fp32_len, matmul_f16_f32_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_l, 1);
+        ggml_vk_create_pipeline(device, device->pipeline_matmul_f16_f32->m, "matmul_f16_f32_m", matmul_f16_f32_fp32_len, matmul_f16_f32_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_m, 1);
+        ggml_vk_create_pipeline(device, device->pipeline_matmul_f16_f32->s, "matmul_f16_f32_s", matmul_f16_f32_fp32_len, matmul_f16_f32_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_s, 1);
+        ggml_vk_create_pipeline(device, device->pipeline_matmul_f16_f32->a_l, "matmul_f16_f32_aligned_l", matmul_f16_f32_aligned_fp32_len, matmul_f16_f32_aligned_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_l, l_align);
+        ggml_vk_create_pipeline(device, device->pipeline_matmul_f16_f32->a_m, "matmul_f16_f32_aligned_m", matmul_f16_f32_aligned_fp32_len, matmul_f16_f32_aligned_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_m, m_align);
+        ggml_vk_create_pipeline(device, device->pipeline_matmul_f16_f32->a_s, "matmul_f16_f32_aligned_s", matmul_f16_f32_aligned_fp32_len, matmul_f16_f32_aligned_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_s, s_align);
+
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_0]->l, "matmul_q4_0_f32_l", matmul_q4_0_f32_fp32_len, matmul_q4_0_f32_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_0]->m, "matmul_q4_0_f32_m", matmul_q4_0_f32_fp32_len, matmul_q4_0_f32_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_0]->s, "matmul_q4_0_f32_s", matmul_q4_0_f32_fp32_len, matmul_q4_0_f32_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_0]->a_l, "matmul_q4_0_f32_aligned_l", matmul_q4_0_f32_aligned_fp32_len, matmul_q4_0_f32_aligned_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_0]->a_m, "matmul_q4_0_f32_aligned_m", matmul_q4_0_f32_aligned_fp32_len, matmul_q4_0_f32_aligned_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_0]->a_s, "matmul_q4_0_f32_aligned_s", matmul_q4_0_f32_aligned_fp32_len, matmul_q4_0_f32_aligned_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
+
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_1]->l, "matmul_q4_1_f32_l", matmul_q4_1_f32_fp32_len, matmul_q4_1_f32_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_1]->m, "matmul_q4_1_f32_m", matmul_q4_1_f32_fp32_len, matmul_q4_1_f32_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_1]->s, "matmul_q4_1_f32_s", matmul_q4_1_f32_fp32_len, matmul_q4_1_f32_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_1]->a_l, "matmul_q4_1_f32_aligned_l", matmul_q4_1_f32_aligned_fp32_len, matmul_q4_1_f32_aligned_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_1]->a_m, "matmul_q4_1_f32_aligned_m", matmul_q4_1_f32_aligned_fp32_len, matmul_q4_1_f32_aligned_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_1]->a_s, "matmul_q4_1_f32_aligned_s", matmul_q4_1_f32_aligned_fp32_len, matmul_q4_1_f32_aligned_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
+
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_0]->l, "matmul_q5_0_f32_l", matmul_q5_0_f32_fp32_len, matmul_q5_0_f32_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_0]->m, "matmul_q5_0_f32_m", matmul_q5_0_f32_fp32_len, matmul_q5_0_f32_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_0]->s, "matmul_q5_0_f32_s", matmul_q5_0_f32_fp32_len, matmul_q5_0_f32_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_0]->a_l, "matmul_q5_0_f32_aligned_l", matmul_q5_0_f32_aligned_fp32_len, matmul_q5_0_f32_aligned_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_0]->a_m, "matmul_q5_0_f32_aligned_m", matmul_q5_0_f32_aligned_fp32_len, matmul_q5_0_f32_aligned_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_0]->a_s, "matmul_q5_0_f32_aligned_s", matmul_q5_0_f32_aligned_fp32_len, matmul_q5_0_f32_aligned_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
+
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_1]->l, "matmul_q5_1_f32_l", matmul_q5_1_f32_fp32_len, matmul_q5_1_f32_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_1]->m, "matmul_q5_1_f32_m", matmul_q5_1_f32_fp32_len, matmul_q5_1_f32_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_1]->s, "matmul_q5_1_f32_s", matmul_q5_1_f32_fp32_len, matmul_q5_1_f32_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_1]->a_l, "matmul_q5_1_f32_aligned_l", matmul_q5_1_f32_aligned_fp32_len, matmul_q5_1_f32_aligned_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_1]->a_m, "matmul_q5_1_f32_aligned_m", matmul_q5_1_f32_aligned_fp32_len, matmul_q5_1_f32_aligned_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_1]->a_s, "matmul_q5_1_f32_aligned_s", matmul_q5_1_f32_aligned_fp32_len, matmul_q5_1_f32_aligned_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
+
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q8_0]->l, "matmul_q8_0_f32_l", matmul_q8_0_f32_fp32_len, matmul_q8_0_f32_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q8_0]->m, "matmul_q8_0_f32_m", matmul_q8_0_f32_fp32_len, matmul_q8_0_f32_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q8_0]->s, "matmul_q8_0_f32_s", matmul_q8_0_f32_fp32_len, matmul_q8_0_f32_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q8_0]->a_l, "matmul_q8_0_f32_aligned_l", matmul_q8_0_f32_aligned_fp32_len, matmul_q8_0_f32_aligned_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q8_0]->a_m, "matmul_q8_0_f32_aligned_m", matmul_q8_0_f32_aligned_fp32_len, matmul_q8_0_f32_aligned_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q8_0]->a_s, "matmul_q8_0_f32_aligned_s", matmul_q8_0_f32_aligned_fp32_len, matmul_q8_0_f32_aligned_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
+
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q2_K]->l, "matmul_q2_k_f32_l", matmul_q2_k_f32_fp32_len, matmul_q2_k_f32_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q2_K]->m, "matmul_q2_k_f32_m", matmul_q2_k_f32_fp32_len, matmul_q2_k_f32_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q2_K]->s, "matmul_q2_k_f32_s", matmul_q2_k_f32_fp32_len, matmul_q2_k_f32_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q2_K]->a_l, "matmul_q2_k_f32_aligned_l", matmul_q2_k_f32_aligned_fp32_len, matmul_q2_k_f32_aligned_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q2_K]->a_m, "matmul_q2_k_f32_aligned_m", matmul_q2_k_f32_aligned_fp32_len, matmul_q2_k_f32_aligned_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q2_K]->a_s, "matmul_q2_k_f32_aligned_s", matmul_q2_k_f32_aligned_fp32_len, matmul_q2_k_f32_aligned_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
+
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q3_K]->l, "matmul_q3_k_f32_l", matmul_q3_k_f32_fp32_len, matmul_q3_k_f32_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q3_K]->m, "matmul_q3_k_f32_m", matmul_q3_k_f32_fp32_len, matmul_q3_k_f32_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q3_K]->s, "matmul_q3_k_f32_s", matmul_q3_k_f32_fp32_len, matmul_q3_k_f32_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q3_K]->a_l, "matmul_q3_k_f32_aligned_l", matmul_q3_k_f32_aligned_fp32_len, matmul_q3_k_f32_aligned_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q3_K]->a_m, "matmul_q3_k_f32_aligned_m", matmul_q3_k_f32_aligned_fp32_len, matmul_q3_k_f32_aligned_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q3_K]->a_s, "matmul_q3_k_f32_aligned_s", matmul_q3_k_f32_aligned_fp32_len, matmul_q3_k_f32_aligned_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
+
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_K]->l, "matmul_q4_k_f32_l", matmul_q4_k_f32_fp32_len, matmul_q4_k_f32_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_K]->m, "matmul_q4_k_f32_m", matmul_q4_k_f32_fp32_len, matmul_q4_k_f32_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_K]->s, "matmul_q4_k_f32_s", matmul_q4_k_f32_fp32_len, matmul_q4_k_f32_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_K]->a_l, "matmul_q4_k_f32_aligned_l", matmul_q4_k_f32_aligned_fp32_len, matmul_q4_k_f32_aligned_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_K]->a_m, "matmul_q4_k_f32_aligned_m", matmul_q4_k_f32_aligned_fp32_len, matmul_q4_k_f32_aligned_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q4_K]->a_s, "matmul_q4_k_f32_aligned_s", matmul_q4_k_f32_aligned_fp32_len, matmul_q4_k_f32_aligned_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
+
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_K]->l, "matmul_q5_k_f32_l", matmul_q5_k_f32_fp32_len, matmul_q5_k_f32_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_K]->m, "matmul_q5_k_f32_m", matmul_q5_k_f32_fp32_len, matmul_q5_k_f32_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_K]->s, "matmul_q5_k_f32_s", matmul_q5_k_f32_fp32_len, matmul_q5_k_f32_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_K]->a_l, "matmul_q5_k_f32_aligned_l", matmul_q5_k_f32_aligned_fp32_len, matmul_q5_k_f32_aligned_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_K]->a_m, "matmul_q5_k_f32_aligned_m", matmul_q5_k_f32_aligned_fp32_len, matmul_q5_k_f32_aligned_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q5_K]->a_s, "matmul_q5_k_f32_aligned_s", matmul_q5_k_f32_aligned_fp32_len, matmul_q5_k_f32_aligned_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
+
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q6_K]->l, "matmul_q6_k_f32_l", matmul_q6_k_f32_fp32_len, matmul_q6_k_f32_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q6_K]->m, "matmul_q6_k_f32_m", matmul_q6_k_f32_fp32_len, matmul_q6_k_f32_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q6_K]->s, "matmul_q6_k_f32_s", matmul_q6_k_f32_fp32_len, matmul_q6_k_f32_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q6_K]->a_l, "matmul_q6_k_f32_aligned_l", matmul_q6_k_f32_aligned_fp32_len, matmul_q6_k_f32_aligned_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q6_K]->a_m, "matmul_q6_k_f32_aligned_m", matmul_q6_k_f32_aligned_fp32_len, matmul_q6_k_f32_aligned_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_Q6_K]->a_s, "matmul_q6_k_f32_aligned_s", matmul_q6_k_f32_aligned_fp32_len, matmul_q6_k_f32_aligned_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
+
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_IQ4_NL]->l, "matmul_iq4_nl_f32_l", matmul_iq4_nl_f32_fp32_len, matmul_iq4_nl_f32_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_IQ4_NL]->m, "matmul_iq4_nl_f32_m", matmul_iq4_nl_f32_fp32_len, matmul_iq4_nl_f32_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_IQ4_NL]->s, "matmul_iq4_nl_f32_s", matmul_iq4_nl_f32_fp32_len, matmul_iq4_nl_f32_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_IQ4_NL]->a_l, "matmul_iq4_nl_f32_aligned_l", matmul_iq4_nl_f32_aligned_fp32_len, matmul_iq4_nl_f32_aligned_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_IQ4_NL]->a_m, "matmul_iq4_nl_f32_aligned_m", matmul_iq4_nl_f32_aligned_fp32_len, matmul_iq4_nl_f32_aligned_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat[GGML_TYPE_IQ4_NL]->a_s, "matmul_iq4_nl_f32_aligned_s", matmul_iq4_nl_f32_aligned_fp32_len, matmul_iq4_nl_f32_aligned_fp32_data, "main", 3, sizeof(vk_mat_mat_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
+
+        ggml_vk_create_pipeline(device, device->pipeline_matmul_id_f32->l, "matmul_id_f32_l", matmul_id_f32_f32_fp32_len, matmul_id_f32_f32_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), l_wg_denoms, warptile_l, 1);
+        ggml_vk_create_pipeline(device, device->pipeline_matmul_id_f32->m, "matmul_id_f32_m", matmul_id_f32_f32_fp32_len, matmul_id_f32_f32_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), m_wg_denoms, warptile_m, 1);
+        ggml_vk_create_pipeline(device, device->pipeline_matmul_id_f32->s, "matmul_id_f32_s", matmul_id_f32_f32_fp32_len, matmul_id_f32_f32_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), s_wg_denoms, warptile_s, 1);
+        ggml_vk_create_pipeline(device, device->pipeline_matmul_id_f32->a_l, "matmul_id_f32_aligned_l", matmul_id_f32_f32_aligned_fp32_len, matmul_id_f32_f32_aligned_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), l_wg_denoms, warptile_l, l_align);
+        ggml_vk_create_pipeline(device, device->pipeline_matmul_id_f32->a_m, "matmul_id_f32_aligned_m", matmul_id_f32_f32_aligned_fp32_len, matmul_id_f32_f32_aligned_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), m_wg_denoms, warptile_m, m_align);
+        ggml_vk_create_pipeline(device, device->pipeline_matmul_id_f32->a_s, "matmul_id_f32_aligned_s", matmul_id_f32_f32_aligned_fp32_len, matmul_id_f32_f32_aligned_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), s_wg_denoms, warptile_s, s_align);
+
+        ggml_vk_create_pipeline(device, device->pipeline_matmul_id_f16->l, "matmul_id_f16_l", matmul_id_f16_fp32_len, matmul_id_f16_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), l_wg_denoms, warptile_l, 1);
+        ggml_vk_create_pipeline(device, device->pipeline_matmul_id_f16->m, "matmul_id_f16_m", matmul_id_f16_fp32_len, matmul_id_f16_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), m_wg_denoms, warptile_m, 1);
+        ggml_vk_create_pipeline(device, device->pipeline_matmul_id_f16->s, "matmul_id_f16_s", matmul_id_f16_fp32_len, matmul_id_f16_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), s_wg_denoms, warptile_s, 1);
+        ggml_vk_create_pipeline(device, device->pipeline_matmul_id_f16->a_l, "matmul_id_f16_aligned_l", matmul_id_f16_aligned_fp32_len, matmul_id_f16_aligned_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), l_wg_denoms, warptile_l, l_align);
+        ggml_vk_create_pipeline(device, device->pipeline_matmul_id_f16->a_m, "matmul_id_f16_aligned_m", matmul_id_f16_aligned_fp32_len, matmul_id_f16_aligned_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), m_wg_denoms, warptile_m, m_align);
+        ggml_vk_create_pipeline(device, device->pipeline_matmul_id_f16->a_s, "matmul_id_f16_aligned_s", matmul_id_f16_aligned_fp32_len, matmul_id_f16_aligned_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), s_wg_denoms, warptile_s, s_align);
+
+        ggml_vk_create_pipeline(device, device->pipeline_matmul_id_f16_f32->l, "matmul_id_f16_f32_l", matmul_id_f16_f32_fp32_len, matmul_id_f16_f32_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), l_wg_denoms, warptile_l, 1);
+        ggml_vk_create_pipeline(device, device->pipeline_matmul_id_f16_f32->m, "matmul_id_f16_f32_m", matmul_id_f16_f32_fp32_len, matmul_id_f16_f32_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), m_wg_denoms, warptile_m, 1);
+        ggml_vk_create_pipeline(device, device->pipeline_matmul_id_f16_f32->s, "matmul_id_f16_f32_s", matmul_id_f16_f32_fp32_len, matmul_id_f16_f32_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), s_wg_denoms, warptile_s, 1);
+        ggml_vk_create_pipeline(device, device->pipeline_matmul_id_f16_f32->a_l, "matmul_id_f16_f32_aligned_l", matmul_id_f16_f32_aligned_fp32_len, matmul_id_f16_f32_aligned_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), l_wg_denoms, warptile_l, l_align);
+        ggml_vk_create_pipeline(device, device->pipeline_matmul_id_f16_f32->a_m, "matmul_id_f16_f32_aligned_m", matmul_id_f16_f32_aligned_fp32_len, matmul_id_f16_f32_aligned_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), m_wg_denoms, warptile_m, m_align);
+        ggml_vk_create_pipeline(device, device->pipeline_matmul_id_f16_f32->a_s, "matmul_id_f16_f32_aligned_s", matmul_id_f16_f32_aligned_fp32_len, matmul_id_f16_f32_aligned_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), s_wg_denoms, warptile_s, s_align);
+
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_0]->l, "matmul_id_q4_0_f32_l", matmul_id_q4_0_f32_fp32_len, matmul_id_q4_0_f32_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_0]->m, "matmul_id_q4_0_f32_m", matmul_id_q4_0_f32_fp32_len, matmul_id_q4_0_f32_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_0]->s, "matmul_id_q4_0_f32_s", matmul_id_q4_0_f32_fp32_len, matmul_id_q4_0_f32_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_0]->a_l, "matmul_id_q4_0_f32_aligned_l", matmul_id_q4_0_f32_aligned_fp32_len, matmul_id_q4_0_f32_aligned_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_0]->a_m, "matmul_id_q4_0_f32_aligned_m", matmul_id_q4_0_f32_aligned_fp32_len, matmul_id_q4_0_f32_aligned_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_0]->a_s, "matmul_id_q4_0_f32_aligned_s", matmul_id_q4_0_f32_aligned_fp32_len, matmul_id_q4_0_f32_aligned_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
+
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_1]->l, "matmul_id_q4_1_f32_l", matmul_id_q4_1_f32_fp32_len, matmul_id_q4_1_f32_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_1]->m, "matmul_id_q4_1_f32_m", matmul_id_q4_1_f32_fp32_len, matmul_id_q4_1_f32_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_1]->s, "matmul_id_q4_1_f32_s", matmul_id_q4_1_f32_fp32_len, matmul_id_q4_1_f32_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_1]->a_l, "matmul_id_q4_1_f32_aligned_l", matmul_id_q4_1_f32_aligned_fp32_len, matmul_id_q4_1_f32_aligned_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_1]->a_m, "matmul_id_q4_1_f32_aligned_m", matmul_id_q4_1_f32_aligned_fp32_len, matmul_id_q4_1_f32_aligned_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_1]->a_s, "matmul_id_q4_1_f32_aligned_s", matmul_id_q4_1_f32_aligned_fp32_len, matmul_id_q4_1_f32_aligned_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
+
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_0]->l, "matmul_id_q5_0_f32_l", matmul_id_q5_0_f32_fp32_len, matmul_id_q5_0_f32_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_0]->m, "matmul_id_q5_0_f32_m", matmul_id_q5_0_f32_fp32_len, matmul_id_q5_0_f32_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_0]->s, "matmul_id_q5_0_f32_s", matmul_id_q5_0_f32_fp32_len, matmul_id_q5_0_f32_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_0]->a_l, "matmul_id_q5_0_f32_aligned_l", matmul_id_q5_0_f32_aligned_fp32_len, matmul_id_q5_0_f32_aligned_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_0]->a_m, "matmul_id_q5_0_f32_aligned_m", matmul_id_q5_0_f32_aligned_fp32_len, matmul_id_q5_0_f32_aligned_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_0]->a_s, "matmul_id_q5_0_f32_aligned_s", matmul_id_q5_0_f32_aligned_fp32_len, matmul_id_q5_0_f32_aligned_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
+
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_1]->l, "matmul_id_q5_1_f32_l", matmul_id_q5_1_f32_fp32_len, matmul_id_q5_1_f32_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_1]->m, "matmul_id_q5_1_f32_m", matmul_id_q5_1_f32_fp32_len, matmul_id_q5_1_f32_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_1]->s, "matmul_id_q5_1_f32_s", matmul_id_q5_1_f32_fp32_len, matmul_id_q5_1_f32_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_1]->a_l, "matmul_id_q5_1_f32_aligned_l", matmul_id_q5_1_f32_aligned_fp32_len, matmul_id_q5_1_f32_aligned_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_1]->a_m, "matmul_id_q5_1_f32_aligned_m", matmul_id_q5_1_f32_aligned_fp32_len, matmul_id_q5_1_f32_aligned_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_1]->a_s, "matmul_id_q5_1_f32_aligned_s", matmul_id_q5_1_f32_aligned_fp32_len, matmul_id_q5_1_f32_aligned_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
+
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q8_0]->l, "matmul_id_q8_0_f32_l", matmul_id_q8_0_f32_fp32_len, matmul_id_q8_0_f32_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q8_0]->m, "matmul_id_q8_0_f32_m", matmul_id_q8_0_f32_fp32_len, matmul_id_q8_0_f32_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q8_0]->s, "matmul_id_q8_0_f32_s", matmul_id_q8_0_f32_fp32_len, matmul_id_q8_0_f32_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q8_0]->a_l, "matmul_id_q8_0_f32_aligned_l", matmul_id_q8_0_f32_aligned_fp32_len, matmul_id_q8_0_f32_aligned_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q8_0]->a_m, "matmul_id_q8_0_f32_aligned_m", matmul_id_q8_0_f32_aligned_fp32_len, matmul_id_q8_0_f32_aligned_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q8_0]->a_s, "matmul_id_q8_0_f32_aligned_s", matmul_id_q8_0_f32_aligned_fp32_len, matmul_id_q8_0_f32_aligned_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
+
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q2_K]->l, "matmul_id_q2_k_f32_l", matmul_id_q2_k_f32_fp32_len, matmul_id_q2_k_f32_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q2_K]->m, "matmul_id_q2_k_f32_m", matmul_id_q2_k_f32_fp32_len, matmul_id_q2_k_f32_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q2_K]->s, "matmul_id_q2_k_f32_s", matmul_id_q2_k_f32_fp32_len, matmul_id_q2_k_f32_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q2_K]->a_l, "matmul_id_q2_k_f32_aligned_l", matmul_id_q2_k_f32_aligned_fp32_len, matmul_id_q2_k_f32_aligned_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q2_K]->a_m, "matmul_id_q2_k_f32_aligned_m", matmul_id_q2_k_f32_aligned_fp32_len, matmul_id_q2_k_f32_aligned_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q2_K]->a_s, "matmul_id_q2_k_f32_aligned_s", matmul_id_q2_k_f32_aligned_fp32_len, matmul_id_q2_k_f32_aligned_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
+
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q3_K]->l, "matmul_id_q3_k_f32_l", matmul_id_q3_k_f32_fp32_len, matmul_id_q3_k_f32_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q3_K]->m, "matmul_id_q3_k_f32_m", matmul_id_q3_k_f32_fp32_len, matmul_id_q3_k_f32_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q3_K]->s, "matmul_id_q3_k_f32_s", matmul_id_q3_k_f32_fp32_len, matmul_id_q3_k_f32_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q3_K]->a_l, "matmul_id_q3_k_f32_aligned_l", matmul_id_q3_k_f32_aligned_fp32_len, matmul_id_q3_k_f32_aligned_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q3_K]->a_m, "matmul_id_q3_k_f32_aligned_m", matmul_id_q3_k_f32_aligned_fp32_len, matmul_id_q3_k_f32_aligned_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q3_K]->a_s, "matmul_id_q3_k_f32_aligned_s", matmul_id_q3_k_f32_aligned_fp32_len, matmul_id_q3_k_f32_aligned_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
+
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_K]->l, "matmul_id_q4_k_f32_l", matmul_id_q4_k_f32_fp32_len, matmul_id_q4_k_f32_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_K]->m, "matmul_id_q4_k_f32_m", matmul_id_q4_k_f32_fp32_len, matmul_id_q4_k_f32_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_K]->s, "matmul_id_q4_k_f32_s", matmul_id_q4_k_f32_fp32_len, matmul_id_q4_k_f32_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_K]->a_l, "matmul_id_q4_k_f32_aligned_l", matmul_id_q4_k_f32_aligned_fp32_len, matmul_id_q4_k_f32_aligned_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_K]->a_m, "matmul_id_q4_k_f32_aligned_m", matmul_id_q4_k_f32_aligned_fp32_len, matmul_id_q4_k_f32_aligned_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_K]->a_s, "matmul_id_q4_k_f32_aligned_s", matmul_id_q4_k_f32_aligned_fp32_len, matmul_id_q4_k_f32_aligned_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
+
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_K]->l, "matmul_id_q5_k_f32_l", matmul_id_q5_k_f32_fp32_len, matmul_id_q5_k_f32_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_K]->m, "matmul_id_q5_k_f32_m", matmul_id_q5_k_f32_fp32_len, matmul_id_q5_k_f32_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_K]->s, "matmul_id_q5_k_f32_s", matmul_id_q5_k_f32_fp32_len, matmul_id_q5_k_f32_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_K]->a_l, "matmul_id_q5_k_f32_aligned_l", matmul_id_q5_k_f32_aligned_fp32_len, matmul_id_q5_k_f32_aligned_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_K]->a_m, "matmul_id_q5_k_f32_aligned_m", matmul_id_q5_k_f32_aligned_fp32_len, matmul_id_q5_k_f32_aligned_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_K]->a_s, "matmul_id_q5_k_f32_aligned_s", matmul_id_q5_k_f32_aligned_fp32_len, matmul_id_q5_k_f32_aligned_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
+
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q6_K]->l, "matmul_id_q6_k_f32_l", matmul_id_q6_k_f32_fp32_len, matmul_id_q6_k_f32_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q6_K]->m, "matmul_id_q6_k_f32_m", matmul_id_q6_k_f32_fp32_len, matmul_id_q6_k_f32_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q6_K]->s, "matmul_id_q6_k_f32_s", matmul_id_q6_k_f32_fp32_len, matmul_id_q6_k_f32_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q6_K]->a_l, "matmul_id_q6_k_f32_aligned_l", matmul_id_q6_k_f32_aligned_fp32_len, matmul_id_q6_k_f32_aligned_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q6_K]->a_m, "matmul_id_q6_k_f32_aligned_m", matmul_id_q6_k_f32_aligned_fp32_len, matmul_id_q6_k_f32_aligned_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q6_K]->a_s, "matmul_id_q6_k_f32_aligned_s", matmul_id_q6_k_f32_aligned_fp32_len, matmul_id_q6_k_f32_aligned_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
+
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ4_NL]->l, "matmul_id_iq4_nl_f32_l", matmul_id_iq4_nl_f32_fp32_len, matmul_id_iq4_nl_f32_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ4_NL]->m, "matmul_id_iq4_nl_f32_m", matmul_id_iq4_nl_f32_fp32_len, matmul_id_iq4_nl_f32_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ4_NL]->s, "matmul_id_iq4_nl_f32_s", matmul_id_iq4_nl_f32_fp32_len, matmul_id_iq4_nl_f32_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ4_NL]->a_l, "matmul_id_iq4_nl_f32_aligned_l", matmul_id_iq4_nl_f32_aligned_fp32_len, matmul_id_iq4_nl_f32_aligned_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), l_wg_denoms, warptile_mmq_l, l_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ4_NL]->a_m, "matmul_id_iq4_nl_f32_aligned_m", matmul_id_iq4_nl_f32_aligned_fp32_len, matmul_id_iq4_nl_f32_aligned_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), m_wg_denoms, warptile_mmq_m, m_align);
+        ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ4_NL]->a_s, "matmul_id_iq4_nl_f32_aligned_s", matmul_id_iq4_nl_f32_aligned_fp32_len, matmul_id_iq4_nl_f32_aligned_fp32_data, "main", 4, sizeof(vk_mat_mat_id_push_constants), s_wg_denoms, warptile_mmq_s, s_align);
+    }
+
+    // mul mat vec
+    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_F32 ], "mul_mat_vec_f32_f32_f32",  mul_mat_vec_f32_f32_f32_len,  mul_mat_vec_f32_f32_f32_data,  "main", 3, sizeof(vk_mat_vec_push_constants), {1, 1, 1}, { device->subgroup_size }, 1);
+    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_F16 ], "mul_mat_vec_f16_f32_f32",  mul_mat_vec_f16_f32_f32_len,  mul_mat_vec_f16_f32_f32_data,  "main", 3, sizeof(vk_mat_vec_push_constants), {1, 1, 1}, { device->subgroup_size }, 1);
+    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_Q4_0], "mul_mat_vec_q4_0_f32_f32", mul_mat_vec_q4_0_f32_f32_len, mul_mat_vec_q4_0_f32_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {1, 1, 1}, { device->subgroup_size }, 1);
+    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_Q4_1], "mul_mat_vec_q4_1_f32_f32", mul_mat_vec_q4_1_f32_f32_len, mul_mat_vec_q4_1_f32_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {1, 1, 1}, { device->subgroup_size }, 1);
+    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_Q5_0], "mul_mat_vec_q5_0_f32_f32", mul_mat_vec_q5_0_f32_f32_len, mul_mat_vec_q5_0_f32_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {1, 1, 1}, { device->subgroup_size }, 1);
+    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_Q5_1], "mul_mat_vec_q5_1_f32_f32", mul_mat_vec_q5_1_f32_f32_len, mul_mat_vec_q5_1_f32_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {1, 1, 1}, { device->subgroup_size }, 1);
+    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_Q8_0], "mul_mat_vec_q8_0_f32_f32", mul_mat_vec_q8_0_f32_f32_len, mul_mat_vec_q8_0_f32_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {1, 1, 1}, { device->subgroup_size }, 1);
+    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_Q2_K], "mul_mat_vec_q2_k_f32_f32", mul_mat_vec_q2_k_f32_f32_len, mul_mat_vec_q2_k_f32_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {1, 1, 1}, { device->subgroup_size }, 1);
+    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_Q3_K], "mul_mat_vec_q3_k_f32_f32", mul_mat_vec_q3_k_f32_f32_len, mul_mat_vec_q3_k_f32_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {1, 1, 1}, { device->subgroup_size }, 1);
+    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_Q4_K], "mul_mat_vec_q4_k_f32_f32", mul_mat_vec_q4_k_f32_f32_len, mul_mat_vec_q4_k_f32_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {1, 1, 1}, { device->subgroup_size }, 1);
+    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_Q5_K], "mul_mat_vec_q5_k_f32_f32", mul_mat_vec_q5_k_f32_f32_len, mul_mat_vec_q5_k_f32_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {1, 1, 1}, { device->subgroup_size }, 1);
+    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_Q6_K], "mul_mat_vec_q6_k_f32_f32", mul_mat_vec_q6_k_f32_f32_len, mul_mat_vec_q6_k_f32_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {1, 1, 1}, { device->subgroup_size }, 1);
+    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_IQ4_NL], "mul_mat_vec_iq4_nl_f32_f32", mul_mat_vec_iq4_nl_f32_f32_len, mul_mat_vec_iq4_nl_f32_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {1, 1, 1}, { device->subgroup_size }, 1);
+
+    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_F32 ], "mul_mat_vec_f32_f16_f32",  mul_mat_vec_f32_f16_f32_len,  mul_mat_vec_f32_f16_f32_data,  "main", 3, sizeof(vk_mat_vec_push_constants), {1, 1, 1}, { device->subgroup_size }, 1);
+    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_F16 ], "mul_mat_vec_f16_f16_f32",  mul_mat_vec_f16_f16_f32_len,  mul_mat_vec_f16_f16_f32_data,  "main", 3, sizeof(vk_mat_vec_push_constants), {1, 1, 1}, { device->subgroup_size }, 1);
+    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_Q4_0], "mul_mat_vec_q4_0_f16_f32", mul_mat_vec_q4_0_f16_f32_len, mul_mat_vec_q4_0_f16_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {1, 1, 1}, { device->subgroup_size }, 1);
+    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_Q4_1], "mul_mat_vec_q4_1_f16_f32", mul_mat_vec_q4_1_f16_f32_len, mul_mat_vec_q4_1_f16_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {1, 1, 1}, { device->subgroup_size }, 1);
+    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_Q5_0], "mul_mat_vec_q5_0_f16_f32", mul_mat_vec_q5_0_f16_f32_len, mul_mat_vec_q5_0_f16_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {1, 1, 1}, { device->subgroup_size }, 1);
+    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_Q5_1], "mul_mat_vec_q5_1_f16_f32", mul_mat_vec_q5_1_f16_f32_len, mul_mat_vec_q5_1_f16_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {1, 1, 1}, { device->subgroup_size }, 1);
+    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_Q8_0], "mul_mat_vec_q8_0_f16_f32", mul_mat_vec_q8_0_f16_f32_len, mul_mat_vec_q8_0_f16_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {1, 1, 1}, { device->subgroup_size }, 1);
+    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_Q2_K], "mul_mat_vec_q2_k_f16_f32", mul_mat_vec_q2_k_f16_f32_len, mul_mat_vec_q2_k_f16_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {1, 1, 1}, { device->subgroup_size }, 1);
+    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_Q3_K], "mul_mat_vec_q3_k_f16_f32", mul_mat_vec_q3_k_f16_f32_len, mul_mat_vec_q3_k_f16_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {1, 1, 1}, { device->subgroup_size }, 1);
+    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_Q4_K], "mul_mat_vec_q4_k_f16_f32", mul_mat_vec_q4_k_f16_f32_len, mul_mat_vec_q4_k_f16_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {1, 1, 1}, { device->subgroup_size }, 1);
+    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_Q5_K], "mul_mat_vec_q5_k_f16_f32", mul_mat_vec_q5_k_f16_f32_len, mul_mat_vec_q5_k_f16_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {1, 1, 1}, { device->subgroup_size }, 1);
+    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_Q6_K], "mul_mat_vec_q6_k_f16_f32", mul_mat_vec_q6_k_f16_f32_len, mul_mat_vec_q6_k_f16_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {1, 1, 1}, { device->subgroup_size }, 1);
+    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_IQ4_NL], "mul_mat_vec_iq4_nl_f16_f32", mul_mat_vec_iq4_nl_f16_f32_len, mul_mat_vec_iq4_nl_f16_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {1, 1, 1}, { device->subgroup_size }, 1);
+
+    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_F32 ], "mul_mat_vec_id_f32_f32",  mul_mat_vec_id_f32_f32_len,  mul_mat_vec_id_f32_f32_data,  "main", 4, sizeof(vk_mat_vec_id_push_constants), {1, 1, 1}, { device->subgroup_size }, 1);
+    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_F16 ], "mul_mat_vec_id_f16_f32",  mul_mat_vec_id_f16_f32_len,  mul_mat_vec_id_f16_f32_data,  "main", 4, sizeof(vk_mat_vec_id_push_constants), {1, 1, 1}, { device->subgroup_size }, 1);
+    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_Q4_0], "mul_mat_vec_id_q4_0_f32", mul_mat_vec_id_q4_0_f32_len, mul_mat_vec_id_q4_0_f32_data, "main", 4, sizeof(vk_mat_vec_id_push_constants), {1, 1, 1}, { device->subgroup_size }, 1);
+    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_Q4_1], "mul_mat_vec_id_q4_1_f32", mul_mat_vec_id_q4_1_f32_len, mul_mat_vec_id_q4_1_f32_data, "main", 4, sizeof(vk_mat_vec_id_push_constants), {1, 1, 1}, { device->subgroup_size }, 1);
+    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_Q5_0], "mul_mat_vec_id_q5_0_f32", mul_mat_vec_id_q5_0_f32_len, mul_mat_vec_id_q5_0_f32_data, "main", 4, sizeof(vk_mat_vec_id_push_constants), {1, 1, 1}, { device->subgroup_size }, 1);
+    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_Q5_1], "mul_mat_vec_id_q5_1_f32", mul_mat_vec_id_q5_1_f32_len, mul_mat_vec_id_q5_1_f32_data, "main", 4, sizeof(vk_mat_vec_id_push_constants), {1, 1, 1}, { device->subgroup_size }, 1);
+    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_Q8_0], "mul_mat_vec_id_q8_0_f32", mul_mat_vec_id_q8_0_f32_len, mul_mat_vec_id_q8_0_f32_data, "main", 4, sizeof(vk_mat_vec_id_push_constants), {1, 1, 1}, { device->subgroup_size }, 1);
+    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_Q2_K], "mul_mat_vec_id_q2_k_f32", mul_mat_vec_id_q2_k_f32_len, mul_mat_vec_id_q2_k_f32_data, "main", 4, sizeof(vk_mat_vec_id_push_constants), {1, 1, 1}, { device->subgroup_size }, 1);
+    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_Q3_K], "mul_mat_vec_id_q3_k_f32", mul_mat_vec_id_q3_k_f32_len, mul_mat_vec_id_q3_k_f32_data, "main", 4, sizeof(vk_mat_vec_id_push_constants), {1, 1, 1}, { device->subgroup_size }, 1);
+    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_Q4_K], "mul_mat_vec_id_q4_k_f32", mul_mat_vec_id_q4_k_f32_len, mul_mat_vec_id_q4_k_f32_data, "main", 4, sizeof(vk_mat_vec_id_push_constants), {1, 1, 1}, { device->subgroup_size }, 1);
+    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_Q5_K], "mul_mat_vec_id_q5_k_f32", mul_mat_vec_id_q5_k_f32_len, mul_mat_vec_id_q5_k_f32_data, "main", 4, sizeof(vk_mat_vec_id_push_constants), {1, 1, 1}, { device->subgroup_size }, 1);
+    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_Q6_K], "mul_mat_vec_id_q6_k_f32", mul_mat_vec_id_q6_k_f32_len, mul_mat_vec_id_q6_k_f32_data, "main", 4, sizeof(vk_mat_vec_id_push_constants), {1, 1, 1}, { device->subgroup_size }, 1);
+    ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_IQ4_NL], "mul_mat_vec_id_iq4_nl_f32", mul_mat_vec_id_iq4_nl_f32_len, mul_mat_vec_id_iq4_nl_f32_data, "main", 4, sizeof(vk_mat_vec_id_push_constants), {1, 1, 1}, { device->subgroup_size }, 1);
+
+    // dequant shaders
+    ggml_vk_create_pipeline(device, device->pipeline_dequant[GGML_TYPE_F32 ], "f32_to_f16",   dequant_f32_len,  dequant_f32_data,  "main", 2, 5 * sizeof(uint32_t), {256 * 16, 1, 1}, {}, 1);
+    ggml_vk_create_pipeline(device, device->pipeline_dequant[GGML_TYPE_Q4_0], "dequant_q4_0", dequant_q4_0_len, dequant_q4_0_data, "main", 2, 5 * sizeof(uint32_t), {256 * 16, 1, 1}, {}, 1);
+    ggml_vk_create_pipeline(device, device->pipeline_dequant[GGML_TYPE_Q4_1], "dequant_q4_1", dequant_q4_1_len, dequant_q4_1_data, "main", 2, 5 * sizeof(uint32_t), {256 * 16, 1, 1}, {}, 1);
+    ggml_vk_create_pipeline(device, device->pipeline_dequant[GGML_TYPE_Q5_0], "dequant_q5_0", dequant_q5_0_len, dequant_q5_0_data, "main", 2, 5 * sizeof(uint32_t), {256 * 16, 1, 1}, {}, 1);
+    ggml_vk_create_pipeline(device, device->pipeline_dequant[GGML_TYPE_Q5_1], "dequant_q5_1", dequant_q5_1_len, dequant_q5_1_data, "main", 2, 5 * sizeof(uint32_t), {256 * 16, 1, 1}, {}, 1);
+    ggml_vk_create_pipeline(device, device->pipeline_dequant[GGML_TYPE_Q8_0], "dequant_q8_0", dequant_q8_0_len, dequant_q8_0_data, "main", 2, 5 * sizeof(uint32_t), {256 * 16, 1, 1}, {}, 1);
+    ggml_vk_create_pipeline(device, device->pipeline_dequant[GGML_TYPE_Q2_K], "dequant_q2_k", dequant_q2_k_len, dequant_q2_k_data, "main", 2, 5 * sizeof(uint32_t), {256 * 64, 1, 1}, {}, 1);
+    ggml_vk_create_pipeline(device, device->pipeline_dequant[GGML_TYPE_Q3_K], "dequant_q3_k", dequant_q3_k_len, dequant_q3_k_data, "main", 2, 5 * sizeof(uint32_t), {256 * 64, 1, 1}, {}, 1);
+    ggml_vk_create_pipeline(device, device->pipeline_dequant[GGML_TYPE_Q4_K], "dequant_q4_k", dequant_q4_k_len, dequant_q4_k_data, "main", 2, 5 * sizeof(uint32_t), {256 * 32, 1, 1}, {}, 1);
+    ggml_vk_create_pipeline(device, device->pipeline_dequant[GGML_TYPE_Q5_K], "dequant_q5_k", dequant_q5_k_len, dequant_q5_k_data, "main", 2, 5 * sizeof(uint32_t), {256 * 64, 1, 1}, {}, 1);
+    ggml_vk_create_pipeline(device, device->pipeline_dequant[GGML_TYPE_Q6_K], "dequant_q6_k", dequant_q6_k_len, dequant_q6_k_data, "main", 2, 5 * sizeof(uint32_t), {256 * 64, 1, 1}, {}, 1);
+    ggml_vk_create_pipeline(device, device->pipeline_dequant[GGML_TYPE_IQ4_NL], "dequant_iq4_nl", dequant_iq4_nl_len, dequant_iq4_nl_data, "main", 2, 5 * sizeof(uint32_t), {256 * 16, 1, 1}, {}, 1);
+
+    // get_rows
+    ggml_vk_create_pipeline(device, device->pipeline_get_rows[GGML_TYPE_F32 ], "get_rows_f32",  get_rows_f32_len,  get_rows_f32_data,  "main", 3, sizeof(vk_op_binary_push_constants), { 512, 1, 1}, {}, 1);
+    ggml_vk_create_pipeline(device, device->pipeline_get_rows[GGML_TYPE_F16 ], "get_rows_f16",  get_rows_f16_len,  get_rows_f16_data,  "main", 3, sizeof(vk_op_binary_push_constants), { 512, 1, 1}, {}, 1);
+    ggml_vk_create_pipeline(device, device->pipeline_get_rows[GGML_TYPE_Q4_0], "get_rows_q4_0", get_rows_q4_0_len, get_rows_q4_0_data, "main", 3, sizeof(vk_op_binary_push_constants), {1024, 1, 1}, {}, 1);
+    ggml_vk_create_pipeline(device, device->pipeline_get_rows[GGML_TYPE_Q4_1], "get_rows_q4_1", get_rows_q4_1_len, get_rows_q4_1_data, "main", 3, sizeof(vk_op_binary_push_constants), {1024, 1, 1}, {}, 1);
+    ggml_vk_create_pipeline(device, device->pipeline_get_rows[GGML_TYPE_Q5_0], "get_rows_q5_0", get_rows_q5_0_len, get_rows_q5_0_data, "main", 3, sizeof(vk_op_binary_push_constants), {1024, 1, 1}, {}, 1);
+    ggml_vk_create_pipeline(device, device->pipeline_get_rows[GGML_TYPE_Q5_1], "get_rows_q5_1", get_rows_q5_1_len, get_rows_q5_1_data, "main", 3, sizeof(vk_op_binary_push_constants), {1024, 1, 1}, {}, 1);
+    ggml_vk_create_pipeline(device, device->pipeline_get_rows[GGML_TYPE_Q8_0], "get_rows_q8_0", get_rows_q8_0_len, get_rows_q8_0_data, "main", 3, sizeof(vk_op_binary_push_constants), {1024, 1, 1}, {}, 1);
+    ggml_vk_create_pipeline(device, device->pipeline_get_rows[GGML_TYPE_IQ4_NL], "get_rows_iq4_nl", get_rows_iq4_nl_len, get_rows_iq4_nl_data, "main", 3, sizeof(vk_op_binary_push_constants), {1024, 1, 1}, {}, 1);
+
+    ggml_vk_create_pipeline(device, device->pipeline_get_rows_f32[GGML_TYPE_F32 ], "get_rows_f32_f32",  get_rows_f32_f32_len,  get_rows_f32_f32_data,  "main", 3, sizeof(vk_op_binary_push_constants), { 512, 1, 1}, {}, 1);
+    ggml_vk_create_pipeline(device, device->pipeline_get_rows_f32[GGML_TYPE_F16 ], "get_rows_f16_f32",  get_rows_f16_f32_len,  get_rows_f16_f32_data,  "main", 3, sizeof(vk_op_binary_push_constants), { 512, 1, 1}, {}, 1);
+    ggml_vk_create_pipeline(device, device->pipeline_get_rows_f32[GGML_TYPE_Q4_0], "get_rows_q4_0_f32", get_rows_q4_0_f32_len, get_rows_q4_0_f32_data, "main", 3, sizeof(vk_op_binary_push_constants), {1024, 1, 1}, {}, 1);
+    ggml_vk_create_pipeline(device, device->pipeline_get_rows_f32[GGML_TYPE_Q4_1], "get_rows_q4_1_f32", get_rows_q4_1_f32_len, get_rows_q4_1_f32_data, "main", 3, sizeof(vk_op_binary_push_constants), {1024, 1, 1}, {}, 1);
+    ggml_vk_create_pipeline(device, device->pipeline_get_rows_f32[GGML_TYPE_Q5_0], "get_rows_q5_0_f32", get_rows_q5_0_f32_len, get_rows_q5_0_f32_data, "main", 3, sizeof(vk_op_binary_push_constants), {1024, 1, 1}, {}, 1);
+    ggml_vk_create_pipeline(device, device->pipeline_get_rows_f32[GGML_TYPE_Q5_1], "get_rows_q5_1_f32", get_rows_q5_1_f32_len, get_rows_q5_1_f32_data, "main", 3, sizeof(vk_op_binary_push_constants), {1024, 1, 1}, {}, 1);
+    ggml_vk_create_pipeline(device, device->pipeline_get_rows_f32[GGML_TYPE_Q8_0], "get_rows_q8_0_f32", get_rows_q8_0_f32_len, get_rows_q8_0_f32_data, "main", 3, sizeof(vk_op_binary_push_constants), {1024, 1, 1}, {}, 1);
+    ggml_vk_create_pipeline(device, device->pipeline_get_rows_f32[GGML_TYPE_IQ4_NL], "get_rows_iq4_nl_f32", get_rows_iq4_nl_f32_len, get_rows_iq4_nl_f32_data, "main", 3, sizeof(vk_op_binary_push_constants), {1024, 1, 1}, {}, 1);
+
+    ggml_vk_create_pipeline(device, device->pipeline_matmul_split_k_reduce, "split_k_reduce", split_k_reduce_len, split_k_reduce_data, "main", 2, 2 * sizeof(uint32_t), {256, 1, 1}, {}, 1);
+
+    ggml_vk_create_pipeline(device, device->pipeline_mul_mat_vec_p021_f16_f32, "mul_mat_vec_p021_f16_f32", mul_mat_vec_p021_f16_f32_len, mul_mat_vec_p021_f16_f32_data, "main", 3, 6 * sizeof(uint32_t), {1, 1, 1}, {}, 1);
+    ggml_vk_create_pipeline(device, device->pipeline_mul_mat_vec_nc_f16_f32, "mul_mat_vec_nc_f16_f32", mul_mat_vec_nc_f16_f32_len, mul_mat_vec_nc_f16_f32_data, "main", 3, 7 * sizeof(uint32_t), {1, 1, 1}, {}, 1);
+
+    ggml_vk_create_pipeline(device, device->pipeline_norm_f32, "norm_f32", norm_f32_len, norm_f32_data, "main", 2, sizeof(vk_op_push_constants), {1, 1, 1}, {}, 1);
+    ggml_vk_create_pipeline(device, device->pipeline_group_norm_f32, "group_norm_f32", group_norm_f32_len, group_norm_f32_data, "main", 2, sizeof(vk_op_push_constants), {1, 1, 1}, {}, 1);
+    ggml_vk_create_pipeline(device, device->pipeline_rms_norm_f32, "rms_norm_f32", rms_norm_f32_len, rms_norm_f32_data, "main", 2, sizeof(vk_op_push_constants), {1, 1, 1}, {}, 1);
+
+    ggml_vk_create_pipeline(device, device->pipeline_cpy_f32_f32, "cpy_f32_f32", cpy_f32_f32_len, cpy_f32_f32_data, "main", 2, sizeof(vk_op_unary_push_constants), {512, 1, 1}, {}, 1);
+    ggml_vk_create_pipeline(device, device->pipeline_cpy_f32_f16, "cpy_f32_f16", cpy_f32_f16_len, cpy_f32_f16_data, "main", 2, sizeof(vk_op_unary_push_constants), {512, 1, 1}, {}, 1);
+    ggml_vk_create_pipeline(device, device->pipeline_cpy_f16_f16, "cpy_f16_f16", cpy_f16_f16_len, cpy_f16_f16_data, "main", 2, sizeof(vk_op_unary_push_constants), {512, 1, 1}, {}, 1);
+
+    ggml_vk_create_pipeline(device, device->pipeline_add_f32, "add_f32", add_f32_len, add_f32_data, "main", 3, sizeof(vk_op_binary_push_constants), {512, 1, 1}, {}, 1);
+    ggml_vk_create_pipeline(device, device->pipeline_add_f16_f32_f16, "add_f16_f32_f16", add_f16_f32_f16_len, add_f16_f32_f16_data, "main", 3, sizeof(vk_op_binary_push_constants), {512, 1, 1}, {}, 1);
+
+    ggml_vk_create_pipeline(device, device->pipeline_acc_f32, "acc_f32", acc_f32_len, acc_f32_data, "main", 3, sizeof(vk_op_binary_push_constants), {512, 1, 1}, {}, 1);
+
+    ggml_vk_create_pipeline(device, device->pipeline_mul_f32, "mul_f32", mul_f32_len, mul_f32_data, "main", 3, sizeof(vk_op_binary_push_constants), {512, 1, 1}, {}, 1);
+    ggml_vk_create_pipeline(device, device->pipeline_div_f32, "div_f32", div_f32_len, div_f32_data, "main", 3, sizeof(vk_op_binary_push_constants), {512, 1, 1}, {}, 1);
+
+    ggml_vk_create_pipeline(device, device->pipeline_concat_f32, "concat_f32", concat_f32_len, concat_f32_data, "main", 3, sizeof(vk_op_binary_push_constants), {512, 1, 1}, {}, 1);
+    ggml_vk_create_pipeline(device, device->pipeline_concat_f16, "concat_f16", concat_f16_len, concat_f16_data, "main", 3, sizeof(vk_op_binary_push_constants), {512, 1, 1}, {}, 1);
+    ggml_vk_create_pipeline(device, device->pipeline_concat_i32, "concat_i32", concat_i32_len, concat_i32_data, "main", 3, sizeof(vk_op_binary_push_constants), {512, 1, 1}, {}, 1);
+
+    ggml_vk_create_pipeline(device, device->pipeline_upscale_f32, "upscale_f32", upscale_f32_len, upscale_f32_data, "main", 2, sizeof(vk_op_upscale_push_constants), {512, 1, 1}, {}, 1);
+
+    ggml_vk_create_pipeline(device, device->pipeline_scale_f32, "scale_f32", scale_f32_len, scale_f32_data, "main", 2, sizeof(vk_op_unary_push_constants), {512, 1, 1}, {}, 1);
+
+    ggml_vk_create_pipeline(device, device->pipeline_sqr_f32, "sqr_f32", sqr_f32_len, sqr_f32_data, "main", 2, sizeof(vk_op_unary_push_constants), {512, 1, 1}, {}, 1);
+    ggml_vk_create_pipeline(device, device->pipeline_sin_f32, "sin_f32", sin_f32_len, sin_f32_data, "main", 2, sizeof(vk_op_unary_push_constants), {512, 1, 1}, {}, 1);
+    ggml_vk_create_pipeline(device, device->pipeline_cos_f32, "cos_f32", cos_f32_len, cos_f32_data, "main", 2, sizeof(vk_op_unary_push_constants), {512, 1, 1}, {}, 1);
+
+    ggml_vk_create_pipeline(device, device->pipeline_clamp_f32, "clamp_f32", clamp_f32_len, clamp_f32_data, "main", 2, sizeof(vk_op_unary_push_constants), {512, 1, 1}, {}, 1);
+
+    ggml_vk_create_pipeline(device, device->pipeline_pad_f32, "pad_f32", pad_f32_len, pad_f32_data, "main", 2, sizeof(vk_op_unary_push_constants), {512, 1, 1}, {}, 1);
+
+    ggml_vk_create_pipeline(device, device->pipeline_repeat_f32, "repeat_f32", repeat_f32_len, repeat_f32_data, "main", 2, sizeof(vk_op_unary_push_constants), {512, 1, 1}, {}, 1);
+
+    ggml_vk_create_pipeline(device, device->pipeline_gelu_f32, "gelu_f32", gelu_f32_len, gelu_f32_data, "main", 2, sizeof(vk_op_push_constants), {512, 1, 1}, {}, 1);
+    ggml_vk_create_pipeline(device, device->pipeline_gelu_quick_f32, "gelu_quick_f32", gelu_quick_f32_len, gelu_quick_f32_data, "main", 2, sizeof(vk_op_push_constants), {512, 1, 1}, {}, 1);
+    ggml_vk_create_pipeline(device, device->pipeline_silu_f32, "silu_f32", silu_f32_len, silu_f32_data, "main", 2, sizeof(vk_op_push_constants), {512, 1, 1}, {}, 1);
+    ggml_vk_create_pipeline(device, device->pipeline_relu_f32, "relu_f32", relu_f32_len, relu_f32_data, "main", 2, sizeof(vk_op_push_constants), {512, 1, 1}, {}, 1);
+    ggml_vk_create_pipeline(device, device->pipeline_leaky_relu_f32, "leaky_relu_f32", leaky_relu_f32_len, leaky_relu_f32_data, "main", 2, sizeof(vk_op_push_constants), {512, 1, 1}, {}, 1);
+    ggml_vk_create_pipeline(device, device->pipeline_tanh_f32, "tanh_f32", tanh_f32_len, tanh_f32_data, "main", 2, sizeof(vk_op_push_constants), {512, 1, 1}, {}, 1);
+
+    ggml_vk_create_pipeline(device, device->pipeline_diag_mask_inf_f32, "diag_mask_inf_f32", diag_mask_inf_f32_len, diag_mask_inf_f32_data, "main", 2, sizeof(vk_op_diag_mask_push_constants), {512, 1, 1}, {}, 1);
+
+    ggml_vk_create_pipeline(device, device->pipeline_soft_max_f32, "soft_max_f32", soft_max_f32_len, soft_max_f32_data, "main", 3, sizeof(vk_op_soft_max_push_constants), {1, 1, 1}, {}, 1);
+    ggml_vk_create_pipeline(device, device->pipeline_soft_max_f32_f16, "soft_max_f32_f16", soft_max_f32_f16_len, soft_max_f32_f16_data, "main", 3, sizeof(vk_op_soft_max_push_constants), {1, 1, 1}, {}, 1);
+
+    ggml_vk_create_pipeline(device, device->pipeline_rope_norm_f32, "rope_norm_f32", rope_norm_f32_len, rope_norm_f32_data, "main", 4, sizeof(vk_op_rope_push_constants), {1, 512, 1}, {}, 1);
+    ggml_vk_create_pipeline(device, device->pipeline_rope_norm_f16, "rope_norm_f16", rope_norm_f16_len, rope_norm_f16_data, "main", 4, sizeof(vk_op_rope_push_constants), {1, 512, 1}, {}, 1);
+
+    ggml_vk_create_pipeline(device, device->pipeline_rope_neox_f32, "rope_neox_f32", rope_neox_f32_len, rope_neox_f32_data, "main", 4, sizeof(vk_op_rope_push_constants), {1, 512, 1}, {}, 1);
+    ggml_vk_create_pipeline(device, device->pipeline_rope_neox_f16, "rope_neox_f16", rope_neox_f16_len, rope_neox_f16_data, "main", 4, sizeof(vk_op_rope_push_constants), {1, 512, 1}, {}, 1);
+
+    ggml_vk_create_pipeline(device, device->pipeline_argsort_f32, "argsort_f32", argsort_f32_len, argsort_f32_data, "main", 2, sizeof(vk_op_argsort_push_constants), {1024, 1, 1}, {}, 1);
+
+    ggml_vk_create_pipeline(device, device->pipeline_sum_rows_f32, "sum_rows_f32", sum_rows_f32_len, sum_rows_f32_data, "main", 2, sizeof(vk_op_push_constants), {1, 1, 1}, { device->subgroup_size }, 1);
+
+    ggml_vk_create_pipeline(device, device->pipeline_im2col_f32, "im2col_f32", im2col_f32_len, im2col_f32_data, "main", 2, sizeof(vk_op_im2col_push_constants), {256, 1, 1}, {}, 1);
+    ggml_vk_create_pipeline(device, device->pipeline_im2col_f32_f16, "im2col_f32_f16", im2col_f32_f16_len, im2col_f32_f16_data, "main", 2, sizeof(vk_op_im2col_push_constants), {256, 1, 1}, {}, 1);
+
+    ggml_vk_create_pipeline(device, device->pipeline_timestep_embedding_f32, "timestep_embedding_f32", timestep_embedding_f32_len, timestep_embedding_f32_data, "main", 2, sizeof(vk_op_timestep_embedding_push_constants), {256, 1, 1}, {}, 1);
+
+    for (auto &c : compiles) {
+        c.wait();
+    }
+}
+
+static vk_device ggml_vk_get_device(size_t idx) {
+    VK_LOG_DEBUG("ggml_vk_get_device(" << idx << ")");
+
+    if (vk_instance.devices[idx] == nullptr) {
+        VK_LOG_DEBUG("Initializing new vk_device");
+        vk_device device = std::make_shared<vk_device_struct>();
+        vk_instance.devices[idx] = device;
+
+#ifdef GGML_VULKAN_MEMORY_DEBUG
+        device->memory_logger = std::unique_ptr<vk_memory_logger>(new vk_memory_logger());
+#endif
+#ifdef GGML_VULKAN_PERF
+        device->perf_logger = std::unique_ptr<vk_perf_logger>(new vk_perf_logger());
+#endif
+
+        size_t dev_num = vk_instance.device_indices[idx];
+
+        std::vector<vk::PhysicalDevice> physical_devices = vk_instance.instance.enumeratePhysicalDevices();
+
+        if (dev_num >= physical_devices.size()) {
+            std::cerr << "ggml_vulkan: Device with index " << dev_num << " does not exist." << std::endl;
+            throw std::runtime_error("Device not found");
+        }
+
+        device->physical_device = physical_devices[dev_num];
+        const std::vector<vk::ExtensionProperties> ext_props = device->physical_device.enumerateDeviceExtensionProperties();
+
+        bool maintenance4_support = false;
+
+        // Check if maintenance4 is supported
+        for (const auto& properties : ext_props) {
+            if (strcmp("VK_KHR_maintenance4", properties.extensionName) == 0) {
+                maintenance4_support = true;
+            }
+        }
+
+        vk::PhysicalDeviceProperties2 props2;
+        vk::PhysicalDeviceMaintenance3Properties props3;
+        vk::PhysicalDeviceMaintenance4Properties props4;
+        vk::PhysicalDeviceSubgroupProperties subgroup_props;
+        props2.pNext = &props3;
+        props3.pNext = &subgroup_props;
+        if (maintenance4_support) {
+            subgroup_props.pNext = &props4;
+        }
+        device->physical_device.getProperties2(&props2);
+        device->properties = props2.properties;
+
+        const char* GGML_VK_FORCE_MAX_ALLOCATION_SIZE = getenv("GGML_VK_FORCE_MAX_ALLOCATION_SIZE");
+
+        if (GGML_VK_FORCE_MAX_ALLOCATION_SIZE != nullptr) {
+            device->max_memory_allocation_size = std::stoi(GGML_VK_FORCE_MAX_ALLOCATION_SIZE);
+        } else if (maintenance4_support) {
+            device->max_memory_allocation_size = std::min(props3.maxMemoryAllocationSize, props4.maxBufferSize);
+        } else {
+            device->max_memory_allocation_size = props3.maxMemoryAllocationSize;
+        }
+
+        device->vendor_id = device->properties.vendorID;
+        device->subgroup_size = subgroup_props.subgroupSize;
+        device->uma = device->properties.deviceType == vk::PhysicalDeviceType::eIntegratedGpu;
+
+        bool fp16_storage = false;
+        bool fp16_compute = false;
+
+        for (const auto& properties : ext_props) {
+            if (strcmp("VK_KHR_16bit_storage", properties.extensionName) == 0) {
+                fp16_storage = true;
+            } else if (strcmp("VK_KHR_shader_float16_int8", properties.extensionName) == 0) {
+                fp16_compute = true;
+            }
+        }
+
+        const char* GGML_VK_DISABLE_F16 = getenv("GGML_VK_DISABLE_F16");
+        const bool force_disable_f16 = GGML_VK_DISABLE_F16 != nullptr;
+
+        device->fp16 = !force_disable_f16 && fp16_storage && fp16_compute;
+
+        std::vector<vk::QueueFamilyProperties> queue_family_props = device->physical_device.getQueueFamilyProperties();
+
+        // Try to find a non-graphics compute queue and transfer-focused queues
+        const uint32_t compute_queue_family_index = ggml_vk_find_queue_family_index(queue_family_props, vk::QueueFlagBits::eCompute, vk::QueueFlagBits::eGraphics, -1, 1);
+        const uint32_t transfer_queue_family_index = ggml_vk_find_queue_family_index(queue_family_props, vk::QueueFlagBits::eTransfer, vk::QueueFlagBits::eCompute | vk::QueueFlagBits::eGraphics, compute_queue_family_index, 1);
+
+        const float priorities[] = { 1.0f, 1.0f };
+        device->single_queue = compute_queue_family_index == transfer_queue_family_index && queue_family_props[compute_queue_family_index].queueCount == 1;
+
+        std::vector<vk::DeviceQueueCreateInfo> device_queue_create_infos;
+        if (compute_queue_family_index != transfer_queue_family_index) {
+            device_queue_create_infos.push_back({vk::DeviceQueueCreateFlags(), compute_queue_family_index, 1, priorities});
+            device_queue_create_infos.push_back({vk::DeviceQueueCreateFlags(), transfer_queue_family_index, 1, priorities + 1});
+        } else if(!device->single_queue) {
+            device_queue_create_infos.push_back({vk::DeviceQueueCreateFlags(), compute_queue_family_index, 2, priorities});
+        } else {
+            device_queue_create_infos.push_back({vk::DeviceQueueCreateFlags(), compute_queue_family_index, 1, priorities});
+        }
+        vk::DeviceCreateInfo device_create_info;
+        std::vector<const char *> device_extensions;
+        vk::PhysicalDeviceFeatures device_features = device->physical_device.getFeatures();
+
+        VkPhysicalDeviceFeatures2 device_features2;
+        device_features2.sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_FEATURES_2;
+        device_features2.pNext = nullptr;
+        device_features2.features = (VkPhysicalDeviceFeatures)device_features;
+
+        VkPhysicalDeviceVulkan11Features vk11_features;
+        vk11_features.pNext = nullptr;
+        vk11_features.sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VULKAN_1_1_FEATURES;
+        device_features2.pNext = &vk11_features;
+
+        VkPhysicalDeviceVulkan12Features vk12_features;
+        vk12_features.pNext = nullptr;
+        vk12_features.sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VULKAN_1_2_FEATURES;
+        vk11_features.pNext = &vk12_features;
+
+        vkGetPhysicalDeviceFeatures2(device->physical_device, &device_features2);
+
+        device->fp16 = device->fp16 && vk12_features.shaderFloat16;
+
+        if (!vk11_features.storageBuffer16BitAccess) {
+            std::cerr << "ggml_vulkan: device " << GGML_VK_NAME << idx << " does not support 16-bit storage." << std::endl;
+            throw std::runtime_error("Unsupported device");
+        }
+
+        device_extensions.push_back("VK_KHR_16bit_storage");
+
+#ifdef GGML_VULKAN_VALIDATE
+        device_extensions.push_back("VK_KHR_shader_non_semantic_info");
+#endif
+
+        if (device->fp16) {
+            device_extensions.push_back("VK_KHR_shader_float16_int8");
+        }
+        device->name = device->properties.deviceName.data();
+
+        device_create_info = {
+            vk::DeviceCreateFlags(),
+            device_queue_create_infos,
+            {},
+            device_extensions
+        };
+        device_create_info.setPNext(&device_features2);
+        device->device = device->physical_device.createDevice(device_create_info);
+
+        // Queues
+        ggml_vk_create_queue(device, device->compute_queue, compute_queue_family_index, 0, { vk::PipelineStageFlagBits::eComputeShader | vk::PipelineStageFlagBits::eTransfer }, false);
+
+        // Shaders
+        ggml_vk_load_shaders(device);
+
+        if (!device->single_queue) {
+            const uint32_t transfer_queue_index = compute_queue_family_index == transfer_queue_family_index ? 1 : 0;
+            ggml_vk_create_queue(device, device->transfer_queue, transfer_queue_family_index, transfer_queue_index, { vk::PipelineStageFlagBits::eTransfer }, true);
+        } else {
+            // TODO: Use pointer or reference to avoid copy
+            device->transfer_queue = device->compute_queue;
+        }
+
+        device->buffer_type = {
+            /* .iface    = */ ggml_backend_vk_buffer_type_interface,
+            /* .device   = */ nullptr,
+            /* .context  = */ new ggml_backend_vk_buffer_type_context{ device->name, device },
+        };
+
+        device->fence = device->device.createFence({});
+
+        device->idx = idx;
+
+        return device;
+    }
+
+    return vk_instance.devices[idx];
+}
+
+
+static void ggml_vk_print_gpu_info(size_t idx) {
+    GGML_ASSERT(idx < vk_instance.device_indices.size());
+    size_t dev_num = vk_instance.device_indices[idx];
+    VK_LOG_DEBUG("ggml_vk_print_gpu_info(" << dev_num << ")");
+    GGML_ASSERT(vk_instance_initialized);
+
+    std::vector<vk::PhysicalDevice> devices = vk_instance.instance.enumeratePhysicalDevices();
+
+    if (dev_num >= devices.size()) {
+        std::cerr << "ggml_vulkan: Device with index " << dev_num << " does not exist." << std::endl;
+        throw std::runtime_error("Device not found");
+    }
+
+    vk::PhysicalDevice physical_device = devices[dev_num];
+    std::vector<vk::ExtensionProperties> ext_props = physical_device.enumerateDeviceExtensionProperties();
+
+    vk::PhysicalDeviceProperties2 props2;
+    vk::PhysicalDeviceMaintenance3Properties props3;
+    vk::PhysicalDeviceSubgroupProperties subgroup_props;
+    vk::PhysicalDeviceDriverProperties driver_props;
+    props2.pNext = &props3;
+    props3.pNext = &subgroup_props;
+    subgroup_props.pNext = &driver_props;
+    physical_device.getProperties2(&props2);
+
+    const size_t subgroup_size = subgroup_props.subgroupSize;
+    const bool uma = props2.properties.deviceType == vk::PhysicalDeviceType::eIntegratedGpu;
+
+    bool fp16_storage = false;
+    bool fp16_compute = false;
+
+    for (auto properties : ext_props) {
+        if (strcmp("VK_KHR_16bit_storage", properties.extensionName) == 0) {
+            fp16_storage = true;
+        } else if (strcmp("VK_KHR_shader_float16_int8", properties.extensionName) == 0) {
+            fp16_compute = true;
+        }
+    }
+
+    const char* GGML_VK_DISABLE_F16 = getenv("GGML_VK_DISABLE_F16");
+    bool force_disable_f16 = GGML_VK_DISABLE_F16 != nullptr;
+
+    bool fp16 = !force_disable_f16 && fp16_storage && fp16_compute;
+
+    vk::PhysicalDeviceFeatures device_features = physical_device.getFeatures();
+
+    VkPhysicalDeviceFeatures2 device_features2;
+    device_features2.sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_FEATURES_2;
+    device_features2.pNext = nullptr;
+    device_features2.features = (VkPhysicalDeviceFeatures)device_features;
+
+    VkPhysicalDeviceVulkan11Features vk11_features;
+    vk11_features.pNext = nullptr;
+    vk11_features.sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VULKAN_1_1_FEATURES;
+    device_features2.pNext = &vk11_features;
+
+    VkPhysicalDeviceVulkan12Features vk12_features;
+    vk12_features.pNext = nullptr;
+    vk12_features.sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_VULKAN_1_2_FEATURES;
+    vk11_features.pNext = &vk12_features;
+
+    vkGetPhysicalDeviceFeatures2(physical_device, &device_features2);
+
+    fp16 = fp16 && vk12_features.shaderFloat16;
+
+    std::string device_name = props2.properties.deviceName.data();
+    std::cerr << GGML_VK_NAME << idx << ": " << device_name << " (" << driver_props.driverName << ") | uma: " << uma << " | fp16: " << fp16 << " | warp size: " << subgroup_size << std::endl;
+
+    if (props2.properties.deviceType == vk::PhysicalDeviceType::eCpu) {
+        std::cerr << "ggml_vulkan: Warning: Device type is CPU. This is probably not the device you want." << std::endl;
+    }
+}
+
+static bool ggml_vk_instance_validation_ext_available(const std::vector<vk::ExtensionProperties>& instance_extensions);
+static bool ggml_vk_instance_portability_enumeration_ext_available(const std::vector<vk::ExtensionProperties>& instance_extensions);
+
+void ggml_vk_instance_init() {
+    if (vk_instance_initialized) {
+        return;
+    }
+    VK_LOG_DEBUG("ggml_vk_instance_init()");
+
+    vk_instance_initialized = true;
+
+    vk::ApplicationInfo app_info{ "ggml-vulkan", 1, nullptr, 0, VK_API_VERSION };
+
+    const std::vector<vk::ExtensionProperties> instance_extensions = vk::enumerateInstanceExtensionProperties();
+    const bool validation_ext = ggml_vk_instance_validation_ext_available(instance_extensions);
+#ifdef __APPLE__
+    const bool portability_enumeration_ext = ggml_vk_instance_portability_enumeration_ext_available(instance_extensions);
+#endif
+
+    std::vector<const char*> layers;
+
+    if (validation_ext) {
+        layers.push_back("VK_LAYER_KHRONOS_validation");
+    }
+    std::vector<const char*> extensions;
+    if (validation_ext) {
+        extensions.push_back("VK_EXT_validation_features");
+    }
+#ifdef __APPLE__
+    if (portability_enumeration_ext) {
+        extensions.push_back("VK_KHR_portability_enumeration");
+    }
+#endif
+    vk::InstanceCreateInfo instance_create_info(vk::InstanceCreateFlags{}, &app_info, layers, extensions);
+#ifdef __APPLE__
+    if (portability_enumeration_ext) {
+        instance_create_info.flags |= vk::InstanceCreateFlagBits::eEnumeratePortabilityKHR;
+    }
+#endif
+
+    std::vector<vk::ValidationFeatureEnableEXT> features_enable;
+    vk::ValidationFeaturesEXT validation_features;
+
+    if (validation_ext) {
+        features_enable = { vk::ValidationFeatureEnableEXT::eBestPractices };
+        validation_features = {
+            features_enable,
+            {},
+        };
+        validation_features.setPNext(nullptr);
+        instance_create_info.setPNext(&validation_features);
+
+        std::cerr << "ggml_vulkan: Validation layers enabled" << std::endl;
+    }
+    vk_instance.instance = vk::createInstance(instance_create_info);
+
+    size_t num_available_devices = vk_instance.instance.enumeratePhysicalDevices().size();
+
+    // Emulate behavior of CUDA_VISIBLE_DEVICES for Vulkan
+    char * devices_env = getenv("GGML_VK_VISIBLE_DEVICES");
+    if (devices_env != nullptr) {
+        std::string devices(devices_env);
+        std::replace(devices.begin(), devices.end(), ',', ' ');
+
+        std::stringstream ss(devices);
+        size_t tmp;
+        while (ss >> tmp) {
+            if(tmp >= num_available_devices) {
+                std::cerr << "ggml_vulkan: Invalid device index " << tmp << " in GGML_VK_VISIBLE_DEVICES." << std::endl;
+                throw std::runtime_error("Invalid Vulkan device index");
+            }
+            vk_instance.device_indices.push_back(tmp);
+        }
+    } else {
+        std::vector<vk::PhysicalDevice> devices = vk_instance.instance.enumeratePhysicalDevices();
+
+        // Make sure at least one device exists
+        if (devices.empty()) {
+            std::cerr << "ggml_vulkan: Error: No devices found." << std::endl;
+            GGML_ABORT("fatal error");
+        }
+
+        // Default to using all dedicated GPUs
+        for (size_t i = 0; i < devices.size(); i++) {
+            vk::PhysicalDeviceProperties2 new_props;
+            vk::PhysicalDeviceDriverProperties new_driver;
+            vk::PhysicalDeviceIDProperties new_id;
+            new_props.pNext = &new_driver;
+            new_driver.pNext = &new_id;
+            devices[i].getProperties2(&new_props);
+
+            if (new_props.properties.deviceType == vk::PhysicalDeviceType::eDiscreteGpu) {
+                // Check if there are two physical devices corresponding to the same GPU
+                auto old_device = std::find_if(
+                    vk_instance.device_indices.begin(),
+                    vk_instance.device_indices.end(),
+                    [&devices, &new_id](const size_t k){
+                        vk::PhysicalDeviceProperties2 old_props;
+                        vk::PhysicalDeviceIDProperties old_id;
+                        old_props.pNext = &old_id;
+                        devices[k].getProperties2(&old_props);
+                        return std::equal(std::begin(old_id.deviceUUID), std::end(old_id.deviceUUID), std::begin(new_id.deviceUUID));
+                    }
+                );
+                if (old_device == vk_instance.device_indices.end()) {
+                    vk_instance.device_indices.push_back(i);
+                } else {
+                    // There can be two physical devices corresponding to the same GPU if there are 2 different drivers
+                    // This can cause error when splitting layers aross the devices, need to keep only 1
+                    VK_LOG_DEBUG("Device " << i << " and device " << *old_device << " have the same deviceUUID");
+
+                    vk::PhysicalDeviceProperties2 old_props;
+                    vk::PhysicalDeviceDriverProperties old_driver;
+                    old_props.pNext = &old_driver;
+                    devices[*old_device].getProperties2(&old_props);
+
+                    std::map<vk::DriverId, int> driver_priorities {};
+                    int old_priority = std::numeric_limits<int>::max();
+                    int new_priority = std::numeric_limits<int>::max();
+
+                    // Check https://registry.khronos.org/vulkan/specs/1.3-extensions/man/html/VkDriverId.html for the list of driver id
+                    // Smaller number -> higher priority
+                    switch (old_props.properties.vendorID) {
+                        case VK_VENDOR_ID_AMD:
+                            driver_priorities[vk::DriverId::eMesaRadv] = 1;
+                            driver_priorities[vk::DriverId::eAmdOpenSource] = 2;
+                            driver_priorities[vk::DriverId::eAmdProprietary] = 3;
+                            break;
+                        case VK_VENDOR_ID_INTEL:
+                            driver_priorities[vk::DriverId::eIntelOpenSourceMESA] = 1;
+                            driver_priorities[vk::DriverId::eIntelProprietaryWindows] = 2;
+                            break;
+                        case VK_VENDOR_ID_NVIDIA:
+                            driver_priorities[vk::DriverId::eNvidiaProprietary] = 1;
+#if defined(VK_API_VERSION_1_3) && VK_HEADER_VERSION >= 235
+                            driver_priorities[vk::DriverId::eMesaNvk] = 2;
+#endif
+                            break;
+                    }
+
+                    if (driver_priorities.count(old_driver.driverID)) {
+                        old_priority = driver_priorities[old_driver.driverID];
+                    }
+                    if (driver_priorities.count(new_driver.driverID)) {
+                        new_priority = driver_priorities[new_driver.driverID];
+                    }
+
+                    if (new_priority < old_priority) {
+                        auto r = std::remove(vk_instance.device_indices.begin(), vk_instance.device_indices.end(), *old_device);
+                        vk_instance.device_indices.erase(r, vk_instance.device_indices.end());
+                        vk_instance.device_indices.push_back(i);
+
+                        VK_LOG_DEBUG("Prioritize device " << i << " driver " << new_driver.driverName << " over device " << *old_device << " driver " << old_driver.driverName);
+                    }
+                    else {
+                        VK_LOG_DEBUG("Prioritize device " << *old_device << " driver " << old_driver.driverName << " over device " << i << " driver " << new_driver.driverName << std::endl);
+                    }
+                }
+            }
+        }
+
+        // If no dedicated GPUs found, fall back to GPU 0
+        if (vk_instance.device_indices.empty()) {
+            vk_instance.device_indices.push_back(0);
+        }
+    }
+
+    std::cerr << "ggml_vulkan: Found " << vk_instance.device_indices.size() << " Vulkan devices:" << std::endl;
+
+    for (size_t i = 0; i < vk_instance.device_indices.size(); i++) {
+        ggml_vk_print_gpu_info(i);
+    }
+}
+
+static void ggml_vk_init(ggml_backend_vk_context * ctx, size_t idx) {
+    VK_LOG_DEBUG("ggml_vk_init(" << ctx->name << ", " << idx << ")");
+    ggml_vk_instance_init();
+    GGML_ASSERT(idx < vk_instance.device_indices.size());
+
+    ctx->name = GGML_VK_NAME + std::to_string(idx);
+
+    ctx->device = ggml_vk_get_device(idx);
+
+    ctx->semaphore_idx = 0;
+    ctx->event_idx = 0;
+
+    ctx->prealloc_size_x = 0;
+    ctx->prealloc_size_y = 0;
+    ctx->prealloc_size_split_k = 0;
+
+    ctx->fence = ctx->device->device.createFence({});
+
+#ifdef GGML_VULKAN_CHECK_RESULTS
+    const char* skip_checks = getenv("GGML_VULKAN_SKIP_CHECKS");
+    vk_skip_checks = (skip_checks == NULL ? 0 : atoi(skip_checks));
+    const char* output_tensor = getenv("GGML_VULKAN_OUTPUT_TENSOR");
+    vk_output_tensor = (output_tensor == NULL ? 0 : atoi(output_tensor));
+#endif
+}
+
+static vk_pipeline ggml_vk_get_to_fp16(ggml_backend_vk_context * ctx, ggml_type type) {
+    VK_LOG_DEBUG("ggml_vk_get_to_fp16()");
+    switch (type) {
+        case GGML_TYPE_F32:
+        case GGML_TYPE_Q4_0:
+        case GGML_TYPE_Q4_1:
+        case GGML_TYPE_Q5_0:
+        case GGML_TYPE_Q5_1:
+        case GGML_TYPE_Q8_0:
+        case GGML_TYPE_Q2_K:
+        case GGML_TYPE_Q3_K:
+        case GGML_TYPE_Q4_K:
+        case GGML_TYPE_Q5_K:
+        case GGML_TYPE_Q6_K:
+        case GGML_TYPE_IQ4_NL:
+            break;
+        default:
+            return nullptr;
+    }
+
+    return ctx->device->pipeline_dequant[type];
+}
+
+static vk_matmul_pipeline ggml_vk_get_mul_mat_mat_pipeline(ggml_backend_vk_context * ctx, ggml_type src0_type, ggml_type src1_type) {
+    VK_LOG_DEBUG("ggml_vk_get_mul_mat_mat_pipeline(" << ggml_type_name(src0_type) << ", " << ggml_type_name(src1_type) << ")");
+    if (src0_type == GGML_TYPE_F32 && src1_type == GGML_TYPE_F32) {
+        return ctx->device->pipeline_matmul_f32;
+    }
+    if (src0_type == GGML_TYPE_F32 && src1_type == GGML_TYPE_F16) {
+        return ctx->device->pipeline_matmul_f32_f16;
+    }
+    if (src0_type == GGML_TYPE_F16 && src1_type == GGML_TYPE_F32) {
+        return ctx->device->pipeline_matmul_f16_f32;
+    }
+    if (src0_type == GGML_TYPE_F16 && src1_type == GGML_TYPE_F16) {
+        return ctx->device->pipeline_matmul_f16;
+    }
+
+    if (src1_type != GGML_TYPE_F32) {
+        return nullptr;
+    }
+
+    switch (src0_type) {
+        case GGML_TYPE_Q4_0:
+        case GGML_TYPE_Q4_1:
+        case GGML_TYPE_Q5_0:
+        case GGML_TYPE_Q5_1:
+        case GGML_TYPE_Q8_0:
+        case GGML_TYPE_Q2_K:
+        case GGML_TYPE_Q3_K:
+        case GGML_TYPE_Q4_K:
+        case GGML_TYPE_Q5_K:
+        case GGML_TYPE_Q6_K:
+        case GGML_TYPE_IQ4_NL:
+            break;
+        default:
+            return nullptr;
+    }
+
+    return ctx->device->pipeline_dequant_mul_mat_mat[src0_type];
+}
+
+static vk_pipeline ggml_vk_get_dequantize_mul_mat_vec(ggml_backend_vk_context * ctx, ggml_type a_type, ggml_type b_type) {
+    VK_LOG_DEBUG("ggml_vk_get_dequantize_mul_mat_vec()");
+    GGML_ASSERT(b_type == GGML_TYPE_F32 || b_type == GGML_TYPE_F16);
+
+    switch (a_type) {
+        case GGML_TYPE_F32:
+        case GGML_TYPE_F16:
+        case GGML_TYPE_Q4_0:
+        case GGML_TYPE_Q4_1:
+        case GGML_TYPE_Q5_0:
+        case GGML_TYPE_Q5_1:
+        case GGML_TYPE_Q8_0:
+        case GGML_TYPE_Q2_K:
+        case GGML_TYPE_Q3_K:
+        case GGML_TYPE_Q4_K:
+        case GGML_TYPE_Q5_K:
+        case GGML_TYPE_Q6_K:
+        case GGML_TYPE_IQ4_NL:
+            break;
+        default:
+            return nullptr;
+    }
+
+    return b_type == GGML_TYPE_F32 ? ctx->device->pipeline_dequant_mul_mat_vec_f32_f32[a_type] : ctx->device->pipeline_dequant_mul_mat_vec_f16_f32[a_type];
+}
+
+static vk_matmul_pipeline ggml_vk_get_mul_mat_mat_id_pipeline(ggml_backend_vk_context * ctx, ggml_type src0_type, ggml_type src1_type) {
+    VK_LOG_DEBUG("ggml_vk_get_mul_mat_mat_id_pipeline()");
+    if (src0_type == GGML_TYPE_F32 && src1_type == GGML_TYPE_F32) {
+        return ctx->device->pipeline_matmul_id_f32;
+    }
+    if (src0_type == GGML_TYPE_F16 && src1_type == GGML_TYPE_F32) {
+        return ctx->device->pipeline_matmul_id_f16_f32;
+    }
+    if (src0_type == GGML_TYPE_F16 && src1_type == GGML_TYPE_F16) {
+        return ctx->device->pipeline_matmul_id_f16;
+    }
+
+    GGML_ASSERT(src1_type == GGML_TYPE_F32);
+
+    switch (src0_type) {
+        case GGML_TYPE_Q4_0:
+        case GGML_TYPE_Q4_1:
+        case GGML_TYPE_Q5_0:
+        case GGML_TYPE_Q5_1:
+        case GGML_TYPE_Q8_0:
+        case GGML_TYPE_Q2_K:
+        case GGML_TYPE_Q3_K:
+        case GGML_TYPE_Q4_K:
+        case GGML_TYPE_Q5_K:
+        case GGML_TYPE_Q6_K:
+        case GGML_TYPE_IQ4_NL:
+            break;
+        default:
+            return nullptr;
+    }
+
+    return ctx->device->pipeline_dequant_mul_mat_mat_id[src0_type];
+}
+
+static vk_pipeline ggml_vk_get_dequantize_mul_mat_vec_id(ggml_backend_vk_context * ctx, ggml_type a_type, ggml_type b_type) {
+    VK_LOG_DEBUG("ggml_vk_get_dequantize_mul_mat_vec()");
+    GGML_ASSERT(b_type == GGML_TYPE_F32);
+
+    switch (a_type) {
+        case GGML_TYPE_F32:
+        case GGML_TYPE_F16:
+        case GGML_TYPE_Q4_0:
+        case GGML_TYPE_Q4_1:
+        case GGML_TYPE_Q5_0:
+        case GGML_TYPE_Q5_1:
+        case GGML_TYPE_Q8_0:
+        case GGML_TYPE_Q2_K:
+        case GGML_TYPE_Q3_K:
+        case GGML_TYPE_Q4_K:
+        case GGML_TYPE_Q5_K:
+        case GGML_TYPE_Q6_K:
+        case GGML_TYPE_IQ4_NL:
+            break;
+        default:
+            return nullptr;
+    }
+
+    return ctx->device->pipeline_dequant_mul_mat_vec_id_f32[a_type];
+}
+
+static vk_buffer ggml_vk_pool_malloc(ggml_backend_vk_context * ctx, size_t size) {
+    VK_LOG_DEBUG("ggml_vk_pool_malloc(" << size << ")");
+    VK_LOG_MEMORY("ggml_vk_pool_malloc");
+
+    int best_i = -1;
+    size_t best_size = std::numeric_limits<size_t>::max(); //smallest unused buffer that fits our needs
+    int worst_i = -1;
+    size_t worst_size = 0; //largest unused buffer seen so far
+    for (int i = 0; i < MAX_VK_BUFFERS; ++i) {
+        vk_buffer &b = ctx->buffer_pool[i];
+        if (b != nullptr && b->size >= size && b->size < best_size) {
+            best_i = i;
+            best_size = b->size;
+        }
+        if (b != nullptr && b->size > worst_size) {
+            worst_i = i;
+            worst_size = b->size;
+        }
+    }
+    if(best_i != -1) {
+        //found the smallest buffer that fits our needs
+        vk_buffer b = ctx->buffer_pool[best_i];
+        ctx->buffer_pool[best_i].reset();
+        return b;
+    }
+    if(worst_i != -1) {
+        //no buffer that fits our needs, resize largest one to save memory
+        vk_buffer& b = ctx->buffer_pool[worst_i];
+        ggml_vk_destroy_buffer(b);
+    }
+
+    return ggml_vk_create_buffer_device(ctx->device, size);
+}
+
+static void ggml_vk_pool_free(ggml_backend_vk_context * ctx, vk_buffer& buffer) {
+    VK_LOG_DEBUG("ggml_vk_pool_free(" << buffer->size << ")");
+    for (int i = 0; i < MAX_VK_BUFFERS; ++i) {
+        vk_buffer& b = ctx->buffer_pool[i];
+        if (b == nullptr) {
+            b = buffer;
+            return;
+        }
+    }
+    std::cerr << "ggml_vulkan: WARNING: vk buffer pool full, increase MAX_VK_BUFFERS" << std::endl;
+    ggml_vk_destroy_buffer(buffer);
+}
+
+// Returns an available temporary buffer that may only be used temporarily, it will be reused
+static vk_buffer ggml_vk_create_buffer_temp(ggml_backend_vk_context * ctx, size_t size) {
+    // Try to find existing temp buffer with enough capacity
+    for (auto& buffer : ctx->gc.temp_buffers) {
+        if (buffer->size >= size) {
+            return buffer;
+        }
+    }
+
+    VK_LOG_MEMORY("ggml_vk_create_buffer_temp(" << size << ")");
+
+    // Otherwise create new buffer
+    vk_buffer buf = ggml_vk_pool_malloc(ctx, size);
+    ctx->gc.temp_buffers.push_back(buf);
+
+    return buf;
+}
+
+static void * ggml_vk_host_malloc(vk_device& device, size_t size) {
+    VK_LOG_MEMORY("ggml_vk_host_malloc(" << size << ")");
+    vk_buffer buf = ggml_vk_create_buffer(device, size,
+        vk::MemoryPropertyFlagBits::eHostVisible | vk::MemoryPropertyFlagBits::eHostCoherent | vk::MemoryPropertyFlagBits::eHostCached,
+        vk::MemoryPropertyFlagBits::eHostVisible | vk::MemoryPropertyFlagBits::eHostCoherent);
+
+    if(!(buf->memory_property_flags & vk::MemoryPropertyFlagBits::eHostVisible)) {
+        fprintf(stderr, "WARNING: failed to allocate %.2f MB of pinned memory\n",
+            size/1024.0/1024.0);
+        device->device.freeMemory(buf->device_memory);
+        device->device.destroyBuffer(buf->buffer);
+        return nullptr;
+    }
+
+    device->pinned_memory.push_back(std::make_tuple(buf->ptr, size, buf));
+
+    return buf->ptr;
+}
+
+static void ggml_vk_host_free(vk_device& device, void* ptr) {
+    if (ptr == nullptr) {
+        return;
+    }
+    VK_LOG_MEMORY("ggml_vk_host_free(" << ptr << ")");
+    vk_buffer buf;
+    size_t index;
+    for (size_t i = 0; i < device->pinned_memory.size(); i++) {
+        const uint8_t* addr = (const uint8_t*) std::get<0>(device->pinned_memory[i]);
+        const uint8_t* endr = addr + std::get<1>(device->pinned_memory[i]);
+        if (ptr >= addr && ptr < endr) {
+            buf = std::get<2>(device->pinned_memory[i]);
+            index = i;
+            break;
+        }
+    }
+    if (buf == nullptr) {
+        fprintf(stderr, "WARNING: failed to free pinned memory: memory not in map\n");
+        return;
+    }
+
+    ggml_vk_destroy_buffer(buf);
+
+    device->pinned_memory.erase(device->pinned_memory.begin() + index);
+}
+
+static void ggml_vk_host_get(vk_device& device, const void * ptr, vk_buffer& buf, size_t& buf_offset) {
+    buf = nullptr;
+    buf_offset = 0;
+    for (size_t i = 0; i < device->pinned_memory.size(); i++) {
+        const uint8_t* addr = (const uint8_t*) std::get<0>(device->pinned_memory[i]);
+        const uint8_t* endr = addr + std::get<1>(device->pinned_memory[i]);
+        if (ptr >= addr && ptr < endr) {
+            buf = std::get<2>(device->pinned_memory[i]);
+            buf_offset = ((const uint8_t *)ptr) - addr;
+            break;
+        }
+    }
+}
+
+static vk_submission ggml_vk_begin_submission(vk_device& device, vk_queue& q, bool one_time = true) {
+    vk_submission s;
+    s.buffer = ggml_vk_create_cmd_buffer(device, q);
+    if (one_time) {
+        s.buffer.begin({ vk::CommandBufferUsageFlagBits::eOneTimeSubmit });
+    } else {
+        s.buffer.begin({ vk::CommandBufferUsageFlags{} });
+    }
+
+    return s;
+}
+
+
+
+static void ggml_vk_dispatch_pipeline(ggml_backend_vk_context* ctx, vk_context& subctx, vk_pipeline& pipeline, std::initializer_list<vk::DescriptorBufferInfo> const& descriptor_buffer_infos, size_t push_constant_size, const void* push_constants, std::array<uint32_t, 3> elements) {
+    const uint32_t wg0 = CEIL_DIV(elements[0], pipeline->wg_denoms[0]);
+    const uint32_t wg1 = CEIL_DIV(elements[1], pipeline->wg_denoms[1]);
+    const uint32_t wg2 = CEIL_DIV(elements[2], pipeline->wg_denoms[2]);
+    VK_LOG_DEBUG("ggml_vk_dispatch_pipeline(" << pipeline->name << ", {";
+    for (auto& buffer : descriptor_buffer_infos) {
+        std::cerr << "(" << buffer.buffer << ", " << buffer.offset << ", " << buffer.range << "), ";
+    }
+    std::cerr << "}, (" << wg0 << "," << wg1 << "," << wg2 << "))");
+    GGML_ASSERT(pipeline->descriptor_set_idx < pipeline->descriptor_sets.size());
+    GGML_ASSERT(descriptor_buffer_infos.size() == pipeline->parameter_count);
+
+    vk::DescriptorSet& descriptor_set = pipeline->descriptor_sets[pipeline->descriptor_set_idx++];
+    vk::WriteDescriptorSet write_descriptor_set{ descriptor_set, 0, 0, pipeline->parameter_count, vk::DescriptorType::eStorageBuffer, nullptr, descriptor_buffer_infos.begin() };
+    ctx->device->device.updateDescriptorSets({ write_descriptor_set }, {});
+
+    subctx->s->buffer.pushConstants(pipeline->layout, vk::ShaderStageFlagBits::eCompute, 0, push_constant_size, push_constants);
+    subctx->s->buffer.bindPipeline(vk::PipelineBindPoint::eCompute, pipeline->pipeline);
+    subctx->s->buffer.bindDescriptorSets(vk::PipelineBindPoint::eCompute,
+                                pipeline->layout,
+                                0,
+                                { descriptor_set },
+                                {});
+    subctx->s->buffer.dispatch(wg0, wg1, wg2);
+}
+
+static void ggml_vk_end_submission(vk_submission& s, std::vector<vk_semaphore> wait_semaphores, std::vector<vk_semaphore> signal_semaphores) {
+    s.buffer.end();
+
+    s.wait_semaphores = std::move(wait_semaphores);
+    s.signal_semaphores = std::move(signal_semaphores);
+}
+
+static void ggml_vk_ctx_end(vk_context& ctx) {
+    VK_LOG_DEBUG("ggml_vk_ctx_end(" << ctx << ", " << ctx->seqs.size() << ")");
+    if (ctx->s == nullptr) {
+        return;
+    }
+
+    ctx->s->buffer.end();
+    ctx->s = nullptr;
+}
+
+static void ggml_vk_ctx_begin(vk_device& device, vk_context& subctx) {
+    VK_LOG_DEBUG("ggml_vk_ctx_begin(" << device->name << ")");
+    if (subctx->s != nullptr) {
+        ggml_vk_ctx_end(subctx);
+    }
+
+    subctx->seqs.push_back({ ggml_vk_begin_submission(device, *subctx->q) });
+    subctx->s = subctx->seqs[subctx->seqs.size() - 1].data();
+}
+
+static size_t ggml_vk_align_size(size_t width, size_t align) {
+    VK_LOG_DEBUG("ggml_vk_align_size(" << width << ", " << align << ")");
+    return CEIL_DIV(width, align) * align;
+}
+
+static void deferred_memcpy(void * dst, const void * src, size_t size, std::vector<vk_staging_memcpy>* memcpys = nullptr) {
+    if (memcpys == nullptr) {
+        memcpy(dst, src, size);
+    } else {
+        memcpys->emplace_back(dst, src, size);
+    }
+}
+
+static void ggml_vk_ensure_sync_staging_buffer(vk_device& device, size_t size) {
+    if (device->sync_staging == nullptr || device->sync_staging->size < size) {
+        VK_LOG_MEMORY("ggml_vk_ensure_sync_staging_buffer(" << size << ")");
+        ggml_vk_destroy_buffer(device->sync_staging);
+        device->sync_staging = ggml_vk_create_buffer_check(device, size,
+            vk::MemoryPropertyFlagBits::eHostVisible | vk::MemoryPropertyFlagBits::eHostCoherent | vk::MemoryPropertyFlagBits::eHostCached,
+            vk::MemoryPropertyFlagBits::eHostVisible | vk::MemoryPropertyFlagBits::eHostCoherent);
+    }
+}
+
+static void ggml_vk_buffer_write_nc_async(ggml_backend_vk_context * ctx, vk_context& subctx, vk_buffer& dst, size_t offset, const ggml_tensor * tensor, bool sync_staging = false) {
+    VK_LOG_DEBUG("ggml_vk_buffer_write_nc_async(" << tensor << ")");
+    GGML_ASSERT(!ggml_is_contiguous(tensor));
+    // Buffer is already mapped
+    if(dst->memory_property_flags & vk::MemoryPropertyFlagBits::eHostVisible) {
+        std::cerr << "ggml_vulkan: buffer_write_nc_async dst buffer is host_visible. Use synchronous write." << std::endl;
+        GGML_ABORT("fatal error");
+    }
+    // Check if src is pinned memory
+    vk_buffer buf;
+    size_t buf_offset;
+    ggml_vk_host_get(ctx->device, tensor->data, buf, buf_offset);
+
+    const uint64_t ne0 = tensor->ne[0];
+    const uint64_t ne1 = tensor->ne[1];
+    const uint64_t ne2 = tensor->ne[2];
+    const uint64_t ne3 = tensor->ne[3];
+    const uint64_t nb0 = tensor->nb[0];
+    const uint64_t nb1 = tensor->nb[1];
+    const uint64_t nb2 = tensor->nb[2];
+    const uint64_t nb3 = tensor->nb[3];
+    const ggml_type type = tensor->type;
+    const uint64_t ts = ggml_type_size(type);
+    const uint64_t bs = ggml_blck_size(type);
+
+    const uint64_t dstnb0 = ts;
+    const uint64_t dstnb1 = dstnb0*(ne0/bs);
+    const uint64_t dstnb2 = dstnb1*ne1;
+    const uint64_t dstnb3 = dstnb2*ne2;
+
+    const uint64_t ne = ggml_nelements(tensor);
+
+    if (buf != nullptr) {
+        // Memory is pinned, use as staging buffer
+        std::vector<vk::BufferCopy> slices;
+
+        for (uint64_t i3 = 0; i3 < ne3; i3++) {
+            for (uint64_t i2 = 0; i2 < ne2; i2++) {
+                // Find longest contiguous slice
+                if (ne1*nb1 == dstnb2) {
+                    slices.push_back({ buf_offset + i3*nb3 + i2*nb2, offset + i3*dstnb3 + i2*dstnb2, dstnb2 });
+                } else {
+                    for (uint64_t i1 = 0; i1 < ne1; i1++) {
+                        if (ne0*nb0/bs == dstnb1) {
+                            slices.push_back({ buf_offset + i3*nb3 + i2*nb2 + i1*nb1, offset + i3*dstnb3 + i2*dstnb2 + i1*dstnb1, dstnb1 });
+                        } else {
+                            const uint64_t s_off = buf_offset + i3*nb3 + i2*nb2 + i1*nb1;
+                            const uint64_t d_off = offset + i3*dstnb3 + i2*dstnb2 + i1*dstnb1;
+                            for (uint64_t i0 = 0; i0 < ne0; i0++) {
+                                slices.push_back({ s_off + i1*nb0, d_off + i0*dstnb0, dstnb0 });
+                            }
+                        }
+                    }
+                }
+            }
+        }
+
+        ggml_vk_sync_buffers(subctx);
+        subctx->s->buffer.copyBuffer(buf->buffer, dst->buffer, slices);
+        return;
+    }
+
+    if (!sync_staging) {
+        GGML_ABORT("Asynchronous write to non-pinned memory not supported");
+    }
+
+    // Staging buffer required
+    vk_buffer& staging = ctx->device->sync_staging;
+    const uint64_t copy_size = ts*ne/bs;
+    ggml_vk_ensure_sync_staging_buffer(ctx->device, copy_size);
+    VkBufferCopy buf_copy{ 0, offset, copy_size };
+
+    ggml_vk_sync_buffers(subctx);
+    vkCmdCopyBuffer(subctx->s->buffer, staging->buffer, dst->buffer, 1, &buf_copy);
+
+    for (uint64_t i3 = 0; i3 < ne3; i3++) {
+        for (uint64_t i2 = 0; i2 < ne2; i2++) {
+            // Find longest contiguous slice
+            if (ne1*nb1 == dstnb2) {
+                deferred_memcpy((uint8_t *)staging->ptr + i3*dstnb3 + i2*dstnb2, (const uint8_t *) tensor->data + buf_offset + i3*nb3 + i2*nb2, dstnb2, &subctx->in_memcpys);
+            } else {
+                for (uint64_t i1 = 0; i1 < ne1; i1++) {
+                    if (ne0*nb0/bs == dstnb1) {
+                        deferred_memcpy((uint8_t *)staging->ptr + i3*dstnb3 + i2*dstnb2 + i1*dstnb1, (const uint8_t *) tensor->data + buf_offset + i3*nb3 + i2*nb2 + i1*nb1, dstnb1, &subctx->in_memcpys);
+                    } else {
+                        const uint64_t s_off = buf_offset + i3*nb3 + i2*nb2 + i1*nb1;
+                        const uint64_t d_off = i3*dstnb3 + i2*dstnb2 + i1*dstnb1;
+                        for (uint64_t i0 = 0; i0 < ne0; i0++) {
+                            deferred_memcpy((uint8_t *)staging->ptr + d_off + i0*dstnb0, (const uint8_t *) tensor->data + s_off + i0*nb0, dstnb0, &subctx->in_memcpys);
+                        }
+                    }
+                }
+            }
+        }
+    }
+}
+
+static void ggml_vk_buffer_write_2d_async(vk_context subctx, vk_buffer& dst, size_t offset, const void * src, size_t spitch, size_t width, size_t height, bool sync_staging = false) {
+    VK_LOG_DEBUG("ggml_vk_buffer_write_2d_async(" << width << ", " << height << ")");
+    // Buffer is already mapped
+    if(dst->memory_property_flags & vk::MemoryPropertyFlagBits::eHostVisible) {
+        std::cerr << "ggml_vulkan: buffer_write_async dst buffer is host_visible. Use synchronous write." << std::endl;
+        GGML_ABORT("fatal error");
+    }
+    // Check if src is pinned memory
+    vk_buffer buf = nullptr;
+    size_t buf_offset;
+    ggml_vk_host_get(dst->device, src, buf, buf_offset);
+
+    if (buf != nullptr) {
+        // Memory is pinned, use as staging buffer
+        std::vector<vk::BufferCopy> slices(1);
+        if (width == spitch) {
+            // Only do single write if stride is equal
+            slices[0].srcOffset = buf_offset;
+            slices[0].dstOffset = offset;
+            slices[0].size = width * height;
+        } else {
+            slices.resize(height);
+            for (size_t i = 0; i < height; i++) {
+                slices[i].srcOffset = buf_offset + i * spitch;
+                slices[i].dstOffset = offset + i * width;
+                slices[i].size = width;
+            }
+        }
+
+        ggml_vk_sync_buffers(subctx);
+        subctx->s->buffer.copyBuffer(buf->buffer, dst->buffer, slices);
+        return;
+    }
+    VK_LOG_DEBUG("STAGING");
+
+    if (!sync_staging) {
+        GGML_ABORT("Asynchronous write to non-pinned memory not supported");
+    }
+
+    // Staging buffer required
+    const size_t copy_size = width*height;
+    ggml_vk_ensure_sync_staging_buffer(dst->device, copy_size);
+
+    vk_buffer& staging_buffer = dst->device->sync_staging;
+
+    VkBufferCopy buf_copy = {
+        0,
+        offset,
+        copy_size};
+
+    ggml_vk_sync_buffers(subctx);
+    vkCmdCopyBuffer(subctx->s->buffer, staging_buffer->buffer, dst->buffer, 1, &buf_copy);
+
+    if (width == spitch) {
+        deferred_memcpy((uint8_t *)staging_buffer->ptr, src, width * height, &subctx->in_memcpys);
+    } else {
+        for (size_t i = 0; i < height; i++) {
+            deferred_memcpy((uint8_t *)staging_buffer->ptr + i * width, (const uint8_t *) src + i * spitch, width, &subctx->in_memcpys);
+        }
+    }
+}
+
+static void ggml_vk_buffer_write_async(vk_context subctx, vk_buffer& dst, size_t offset, const void * src, size_t size, bool sync_staging = false) {
+    VK_LOG_DEBUG("ggml_vk_buffer_write_async(" << size << ")");
+    return ggml_vk_buffer_write_2d_async(subctx, dst, offset, src, size, size, 1, sync_staging);
+}
+
+static void ggml_vk_buffer_write_2d(vk_buffer& dst, size_t offset, const void * src, size_t spitch, size_t width, size_t height) {
+    VK_LOG_DEBUG("ggml_vk_buffer_write_2d(" << width << ", " << height << ")");
+    // Buffer is already mapped
+    if(dst->memory_property_flags & vk::MemoryPropertyFlagBits::eHostVisible) {
+        GGML_ASSERT(dst->memory_property_flags & vk::MemoryPropertyFlagBits::eHostCoherent);
+
+        for (size_t i = 0; i < height; i++) {
+            memcpy((uint8_t *)dst->ptr + offset + i * width, (const uint8_t *) src + i * spitch, width);
+        }
+    } else {
+        vk_context subctx = ggml_vk_create_temporary_context(dst->device->transfer_queue);
+        ggml_vk_ctx_begin(dst->device, subctx);
+        ggml_vk_buffer_write_2d_async(subctx, dst, offset, src, spitch, width, height, true);
+        ggml_vk_ctx_end(subctx);
+
+        for (auto& cpy : subctx->in_memcpys) {
+            memcpy(cpy.dst, cpy.src, cpy.n);
+        }
+
+        ggml_vk_submit(subctx, dst->device->fence);
+        VK_CHECK(dst->device->device.waitForFences({ dst->device->fence }, true, UINT64_MAX), "vk_buffer_write_2d waitForFences");
+        dst->device->device.resetFences({ dst->device->fence });
+    }
+}
+
+static void ggml_vk_buffer_write(vk_buffer& dst, size_t offset, const void * src, size_t size) {
+    VK_LOG_DEBUG("ggml_vk_buffer_write(" << size << ")");
+    ggml_vk_buffer_write_2d(dst, offset, src, 0, size, 1);
+}
+
+static void ggml_vk_buffer_read_2d_async(vk_context subctx, vk_buffer& src, size_t offset, void * dst, size_t spitch, size_t dpitch, size_t width, size_t height, bool sync_staging = false) {
+    VK_LOG_DEBUG("ggml_vk_buffer_read_2d_async(offset=" << offset << ", width=" << width << ", height=" << height << ")");
+    GGML_ASSERT(width > 0);
+    GGML_ASSERT(height > 0);
+    GGML_ASSERT(src != nullptr);
+
+    // TODO: staging_offset is not used
+
+    // Check if dst is pinned memory
+    vk_buffer buf = nullptr;
+    size_t buf_offset;
+    ggml_vk_host_get(src->device, dst, buf, buf_offset);
+
+    std::vector<vk::BufferCopy> slices(1);
+    if (width == spitch && width == dpitch) {
+        // Only do single write if stride is equal
+        slices[0].srcOffset = offset;
+        slices[0].dstOffset = buf_offset;
+        slices[0].size = width * height;
+    } else {
+        slices.resize(height);
+        for (size_t i = 0; i < height; i++) {
+            slices[i].srcOffset = offset + i * spitch;
+            slices[i].dstOffset = buf_offset + i * dpitch;
+            slices[i].size = width;
+        }
+    }
+
+    if (buf != nullptr) {
+        // Memory is pinned, use as staging buffer
+        ggml_vk_sync_buffers(subctx);
+        subctx->s->buffer.copyBuffer(src->buffer, buf->buffer, slices);
+
+        return;
+    }
+    VK_LOG_DEBUG("STAGING");
+
+    if (!sync_staging) {
+        GGML_ABORT("Asynchronous read from non-pinned memory not supported");
+    }
+
+    // Fall back to staging buffer
+    const size_t copy_size = dpitch * height;
+    ggml_vk_ensure_sync_staging_buffer(src->device, copy_size);
+
+    vk_buffer& staging_buffer = src->device->sync_staging;
+
+    ggml_vk_sync_buffers(subctx);
+    subctx->s->buffer.copyBuffer(src->buffer, staging_buffer->buffer, slices);
+
+    deferred_memcpy(dst, staging_buffer->ptr, copy_size, &subctx->out_memcpys);
+}
+
+static void ggml_vk_buffer_read_async(vk_context subctx, vk_buffer& src, size_t offset, void * dst, size_t size, bool sync_staging = false) {
+    return ggml_vk_buffer_read_2d_async(subctx, src, offset, dst, size, size, size, 1, sync_staging);
+}
+
+static void ggml_vk_buffer_read(vk_buffer& src, size_t offset, void * dst, size_t size) {
+    VK_LOG_DEBUG("ggml_vk_buffer_read(" << src->buffer << ", " << offset << ", " << size << ")");
+
+    // If the device is not an UMA device the memory is host-accessible through rebar. While writing
+    // through PCIe is sufficient fast reading back data from PCIe is slower than going through
+    // the HW device to host copy path.
+    if(src->memory_property_flags & vk::MemoryPropertyFlagBits::eHostVisible && src->device->uma) {
+        GGML_ASSERT(src->memory_property_flags & vk::MemoryPropertyFlagBits::eHostCoherent);
+
+        memcpy(dst, (uint8_t *) src->ptr + offset, size);
+    } else {
+        vk_context subctx = ggml_vk_create_temporary_context(src->device->transfer_queue);
+        ggml_vk_ctx_begin(src->device, subctx);
+        ggml_vk_buffer_read_async(subctx, src, offset, dst, size, true);
+        ggml_vk_ctx_end(subctx);
+
+        ggml_vk_submit(subctx, src->device->fence);
+        VK_CHECK(src->device->device.waitForFences({ src->device->fence }, true, UINT64_MAX), "vk_buffer_read waitForFences");
+        src->device->device.resetFences({ src->device->fence });
+
+        for (auto& cpy : subctx->out_memcpys) {
+            memcpy(cpy.dst, cpy.src, cpy.n);
+        }
+    }
+}
+
+static void ggml_vk_buffer_copy_async(vk_context& ctx, vk_buffer& dst, size_t dst_offset, vk_buffer& src, size_t src_offset, size_t size) {
+    VK_LOG_DEBUG("ggml_vk_buffer_copy_async(" << size << ")");
+    // Make sure both buffers are on same device
+    GGML_ASSERT(src->device == dst->device);
+
+    VkBufferCopy bc{ src_offset, dst_offset, size };
+
+    vkCmdCopyBuffer(ctx->s->buffer, src->buffer, dst->buffer, 1, &bc);
+}
+
+static void ggml_vk_buffer_copy(vk_buffer& dst, size_t dst_offset, vk_buffer& src, size_t src_offset, size_t size) {
+    if (src->device == dst->device) {
+        VK_LOG_DEBUG("ggml_vk_buffer_copy(SINGLE_DEVICE, " << size << ")");
+        // Copy within the device
+        vk_context subctx = ggml_vk_create_temporary_context(src->device->transfer_queue);
+        ggml_vk_ctx_begin(src->device, subctx);
+        ggml_vk_buffer_copy_async(subctx, dst, dst_offset, src, src_offset, size);
+        ggml_vk_ctx_end(subctx);
+        ggml_vk_submit(subctx, src->device->fence);
+        VK_CHECK(src->device->device.waitForFences({ src->device->fence }, true, UINT64_MAX), "vk_buffer_copy waitForFences");
+        src->device->device.resetFences({ src->device->fence });
+    } else {
+        VK_LOG_DEBUG("ggml_vk_buffer_copy(MULTI_DEVICE, " << size << ")");
+        // Copy device to device
+        ggml_vk_ensure_sync_staging_buffer(src->device, size);
+        ggml_vk_ensure_sync_staging_buffer(dst->device, size);
+
+        // Copy to src staging buffer
+        ggml_vk_buffer_copy(src->device->sync_staging, 0, src, src_offset, size);
+        // memcpy to dst staging buffer
+        memcpy(dst->device->sync_staging->ptr, src->device->sync_staging->ptr, size);
+        // Copy to dst buffer
+        ggml_vk_buffer_copy(dst, dst_offset, dst->device->sync_staging, 0, size);
+    }
+}
+
+static void ggml_vk_buffer_memset(vk_buffer& dst, size_t offset, uint32_t c, size_t size) {
+    VK_LOG_DEBUG("ggml_vk_buffer_memset(" << offset << ", " << c << ", " << size << ")");
+
+    vk_context subctx = ggml_vk_create_temporary_context(dst->device->transfer_queue);
+    ggml_vk_ctx_begin(dst->device, subctx);
+    subctx->s->buffer.fillBuffer(dst->buffer, offset, size, c);
+    ggml_vk_ctx_end(subctx);
+
+    ggml_vk_submit(subctx, dst->device->fence);
+    VK_CHECK(dst->device->device.waitForFences({ dst->device->fence }, true, UINT64_MAX), "vk_memset waitForFences");
+    dst->device->device.resetFences({ dst->device->fence });
+}
+
+static uint32_t ggml_vk_guess_split_k(int m, int n, int k) {
+    VK_LOG_DEBUG("ggml_vk_guess_split_k(" << m << ", " << n << ", " << k << ")");
+    // if (k > 128 && (m < 128 || n < 128) && m > 2 && n > 2) {
+    //     return 4;
+    // }
+
+    return 1;
+
+    GGML_UNUSED(m); GGML_UNUSED(n); GGML_UNUSED(k);
+}
+
+static vk_pipeline ggml_vk_guess_matmul_pipeline_amd(ggml_backend_vk_context * ctx, vk_matmul_pipeline& mmp, int m, int n, bool aligned) {
+    if (m <= 32 || n <= 32) {
+        return aligned ? mmp->a_s : mmp->s;
+    }
+    return aligned ? mmp->a_m : mmp->m;
+
+    GGML_UNUSED(ctx);
+}
+
+static vk_pipeline ggml_vk_guess_matmul_pipeline_apple(ggml_backend_vk_context * ctx, vk_matmul_pipeline& mmp, bool aligned) {
+    return aligned ? mmp->a_m : mmp->m;
+
+    GGML_UNUSED(ctx);
+}
+
+static vk_pipeline ggml_vk_guess_matmul_pipeline_intel(ggml_backend_vk_context * ctx, vk_matmul_pipeline& mmp, bool aligned) {
+    return aligned ? mmp->a_s : mmp->s;
+
+    GGML_UNUSED(ctx);
+}
+
+static vk_pipeline ggml_vk_guess_matmul_pipeline(ggml_backend_vk_context * ctx, vk_matmul_pipeline& mmp, int m, int n, bool aligned) {
+    VK_LOG_DEBUG("ggml_vk_guess_matmul_pipeline(" << m << ", " << n << ", " << aligned << ")");
+    switch (ctx->device->vendor_id) {
+    case VK_VENDOR_ID_AMD:
+        return ggml_vk_guess_matmul_pipeline_amd(ctx, mmp, m, n, aligned);
+    case VK_VENDOR_ID_APPLE:
+        return ggml_vk_guess_matmul_pipeline_apple(ctx, mmp, aligned);
+    case VK_VENDOR_ID_INTEL:
+        return ggml_vk_guess_matmul_pipeline_intel(ctx, mmp, aligned);
+    default:
+        break;
+    }
+
+    if (m <= 32 || n <= 32) {
+        return aligned ? mmp->a_s : mmp->s;
+    }
+    if (m <= 64 || n <= 64) {
+        return aligned ? mmp->a_m : mmp->m;
+    }
+    return aligned ? mmp->a_l : mmp->l;
+}
+
+static uint32_t ggml_vk_guess_matmul_pipeline_align(ggml_backend_vk_context * ctx, vk_matmul_pipeline& mmp, int m, int n) {
+    VK_LOG_DEBUG("ggml_vk_guess_matmul_pipeline_align(" << m << ", " << n << ")");
+    return ggml_vk_guess_matmul_pipeline(ctx, mmp, m, n, true)->align;
+}
+
+static void ggml_vk_matmul(
+        ggml_backend_vk_context * ctx, vk_context& subctx, vk_pipeline& pipeline,
+        vk_subbuffer&& a, vk_subbuffer&& b, vk_subbuffer&& d, vk_subbuffer&& split_k_buffer,
+        uint32_t m, uint32_t n, uint32_t k, uint32_t stride_a, uint32_t stride_b, uint32_t stride_d,
+        uint32_t batch_stride_a, uint32_t batch_stride_b, uint32_t batch_stride_d,
+        uint32_t split_k, uint32_t batch, uint32_t ne02, uint32_t ne12, uint32_t broadcast2, uint32_t broadcast3) {
+        VK_LOG_DEBUG("ggml_vk_matmul(a: (" << a.buffer->buffer << ", " << a.offset << ", " << a.size << "), b: (" << b.buffer->buffer << ", " << b.offset << ", " << b.size << "), d: (" << d.buffer->buffer << ", " << d.offset << ", " << d.size << "), split_k: (" << (split_k_buffer.buffer != nullptr ? split_k_buffer.buffer->buffer : VK_NULL_HANDLE) << ", " << split_k_buffer.offset << ", " << split_k_buffer.size << "), m: " << m << ", n: " << n << ", k: " << k << ", stride_a: " << stride_a << ", stride_b: " << stride_b << ", stride_d: " << stride_d << ", batch_stride_a: " << batch_stride_a << ", batch_stride_b: " << batch_stride_b << ", batch_stride_d: " << batch_stride_d << ", split_k: " << split_k << ", batch: " << batch << ", ne02: " << ne02 << ", ne12: " << ne12 << ", broadcast2: " << broadcast2 << ", broadcast3: " << broadcast3 << ")");
+    ggml_vk_sync_buffers(subctx);
+    if (split_k == 1) {
+        const vk_mat_mat_push_constants pc = { m, n, k, stride_a, stride_b, stride_d, batch_stride_a, batch_stride_b, batch_stride_d, k, ne02, ne12, broadcast2, broadcast3 };
+        ggml_vk_dispatch_pipeline(ctx, subctx, pipeline, { a, b, d }, sizeof(vk_mat_mat_push_constants), &pc, { m, n, batch });
+        return;
+    }
+
+    GGML_ASSERT(batch_stride_d == m * n);
+
+    const vk_mat_mat_push_constants pc1 = { m, n, k, stride_a, stride_b, stride_d, batch_stride_a, batch_stride_b, batch_stride_d, CEIL_DIV(k, split_k), ne02, ne12, broadcast2, broadcast3 };
+    // Make sure enough workgroups get assigned for split k to work
+    ggml_vk_dispatch_pipeline(ctx, subctx, pipeline, { a, b, split_k_buffer }, sizeof(vk_mat_mat_push_constants), &pc1, { (CEIL_DIV(m, pipeline->wg_denoms[0]) * pipeline->wg_denoms[0]) * split_k, n, batch });
+    ggml_vk_sync_buffers(subctx);
+    const std::array<uint32_t, 2> pc2 = { (uint32_t)(m * n * batch), split_k };
+    ggml_vk_dispatch_pipeline(ctx, subctx, ctx->device->pipeline_matmul_split_k_reduce, { split_k_buffer, d }, pc2.size() * sizeof(uint32_t), pc2.data(), { m * n * batch, 1, 1 });
+}
+
+static void ggml_vk_matmul_id(
+        ggml_backend_vk_context * ctx, vk_context& subctx, vk_pipeline& pipeline,
+        vk_subbuffer&& a, vk_subbuffer&& b, vk_subbuffer&& d, vk_subbuffer&& ids,
+        uint32_t m, uint32_t n, uint32_t k, uint32_t stride_a, uint32_t stride_b, uint32_t stride_d,
+        uint32_t batch_stride_a, uint32_t batch_stride_b, uint32_t batch_stride_d,
+        uint32_t n_as, uint32_t nei0, uint32_t nei1, uint32_t nbi1, uint32_t ne11) {
+    VK_LOG_DEBUG("ggml_vk_matmul_id(a: (" << a.buffer->buffer << ", " << a.offset << ", " << a.size << "), b: (" << b.buffer->buffer << ", " << b.offset << ", " << b.size << "), d: (" << d.buffer->buffer << ", " << d.offset << ", " << d.size << "), ids: (" << ids.buffer->buffer << ", " << ids.offset << ", " << ids.size << "), " <<
+        "m: " << m << ", n: " << n << ", k: " << k << ", stride_a: " << stride_a << ", stride_b: " << stride_b << ", stride_d: " << stride_d << ", " <<
+        "batch_stride_a: " << batch_stride_a << ", batch_stride_b: " << batch_stride_b << ", batch_stride_d: " << batch_stride_d << ", " <<
+        "n_as: " << n_as << ", nei0: " << nei0 << ", nei1: " << nei1 << ", nbi1: " << nbi1 << ", ne11: " << ne11 << ")");
+    ggml_vk_sync_buffers(subctx);
+    const vk_mat_mat_id_push_constants pc = { m, n, k, stride_a, stride_b, stride_d, batch_stride_a, batch_stride_b, batch_stride_d,
+                                              nei0, nei1, nbi1, ne11 };
+    ggml_vk_dispatch_pipeline(ctx, subctx, pipeline, { a, b, d, ids }, sizeof(vk_mat_mat_id_push_constants), &pc, { m, nei1, n_as });
+}
+
+static bool ggml_vk_dim01_contiguous(const ggml_tensor * tensor) {
+    return
+        tensor->nb[0] == ggml_type_size(tensor->type) &&
+        tensor->nb[1] == (tensor->nb[0]*tensor->ne[0])/ggml_blck_size(tensor->type) &&
+        tensor->nb[3] == tensor->nb[2]*tensor->ne[2];
+}
+
+static vk_pipeline ggml_vk_get_cpy_pipeline(ggml_backend_vk_context * ctx, ggml_type from, ggml_type to) {
+    if (from == GGML_TYPE_F32 && to == GGML_TYPE_F32) {
+        return ctx->device->pipeline_cpy_f32_f32;
+    }
+    if (from == GGML_TYPE_F32 && to == GGML_TYPE_F16) {
+        return ctx->device->pipeline_cpy_f32_f16;
+    }
+    if (from == GGML_TYPE_F16 && to == GGML_TYPE_F16) {
+        return ctx->device->pipeline_cpy_f16_f16;
+    }
+
+    std::cerr << "Missing CPY op for types: " << ggml_type_name(from) << " " << ggml_type_name(to) << std::endl;
+    GGML_ABORT("fatal error");
+}
+
+static void ggml_vk_cpy_to_contiguous(ggml_backend_vk_context * ctx, vk_context& subctx, vk_pipeline pipeline, const ggml_tensor * tensor, vk_subbuffer&& in, vk_subbuffer&& out) {
+    VK_LOG_DEBUG("ggml_vk_cpy_to_contiguous((" << tensor << ", type=" << tensor->type << ", ne0=" << tensor->ne[0] << ", ne1=" << tensor->ne[1] << ", ne2=" << tensor->ne[2] << ", ne3=" << tensor->ne[3] << ", nb0=" << tensor->nb[0] << ", nb1=" << tensor->nb[1] << ", nb2=" << tensor->nb[2] << ", nb3=" << tensor->nb[3] << "), ";
+    std::cerr << "buffer in size=" << in.buffer->size << ", buffer out size=" << out.buffer->size << ")");
+    const int tensor_type_size = ggml_type_size(tensor->type);
+
+    const uint32_t ne = ggml_nelements(tensor);
+
+    const vk_op_unary_push_constants pc = {
+        (uint32_t)ne,
+        (uint32_t)tensor->ne[0], (uint32_t)tensor->ne[1], (uint32_t)tensor->ne[2], (uint32_t)tensor->ne[3], (uint32_t)tensor->nb[0] / tensor_type_size, (uint32_t)tensor->nb[1] / tensor_type_size, (uint32_t)tensor->nb[2] / tensor_type_size, (uint32_t)tensor->nb[3] / tensor_type_size,
+        (uint32_t)tensor->ne[0], (uint32_t)tensor->ne[1], (uint32_t)tensor->ne[2], (uint32_t)tensor->ne[3],                       1                   , (uint32_t)tensor->ne[0]                   , (uint32_t)(tensor->ne[0] * tensor->ne[1]) , (uint32_t)(tensor->ne[0] * tensor->ne[1] * tensor->ne[2]),
+        0,
+        0.0f, 0.0f,
+    };
+    ggml_vk_sync_buffers(subctx);
+    ggml_vk_dispatch_pipeline(ctx, subctx, pipeline, { in, out }, sizeof(vk_op_unary_push_constants), &pc, { ne, 1, 1 });
+}
+
+static void ggml_vk_mul_mat_q_f16(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, bool dryrun = false) {
+    VK_LOG_DEBUG("ggml_vk_mul_mat_q_f16((" << src0 << ", name=" << src0->name << ", type=" << src0->type << ", ne0=" << src0->ne[0] << ", ne1=" << src0->ne[1] << ", ne2=" << src0->ne[2] << ", ne3=" << src0->ne[3] << ", nb0=" << src0->nb[0] << ", nb1=" << src0->nb[1] << ", nb2=" << src0->nb[2] << ", nb3=" << src0->nb[3];
+    std::cerr << "), (" << src1 << ", name=" << src1->name << ", type=" << src1->type << ", ne0=" << src1->ne[0] << ", ne1=" << src1->ne[1] << ", ne2=" << src1->ne[2] << ", ne3=" << src1->ne[3] << ", nb0=" << src1->nb[0] << ", nb1=" << src1->nb[1] << ", nb2=" << src1->nb[2] << ", nb3=" << src1->nb[3];
+    std::cerr << "), (" << dst << ", name=" << dst->name << ", type=" << dst->type << ", ne0=" << dst->ne[0] << ", ne1=" << dst->ne[1] << ", ne2=" << dst->ne[2] << ", ne3=" << dst->ne[3] << ", nb0=" << dst->nb[0] << ", nb1=" << dst->nb[1] << ", nb2=" << dst->nb[2] << ", nb3=" << dst->nb[3];
+    std::cerr << "), " << (dryrun ? "dryrun" : "") << ")");
+    GGML_ASSERT(ggml_vk_dim01_contiguous(src0) || src0->type == GGML_TYPE_F32 || src0->type == GGML_TYPE_F16);  // NOLINT
+    GGML_ASSERT(ggml_vk_dim01_contiguous(src1) || src1->type == GGML_TYPE_F32 || src1->type == GGML_TYPE_F16);  // NOLINT
+
+    const uint64_t ne00 = src0->ne[0];
+    const uint64_t ne01 = src0->ne[1];
+    const uint64_t ne02 = src0->ne[2];
+    const uint64_t ne03 = src0->ne[3];
+
+    const uint64_t ne10 = src1->ne[0];
+    const uint64_t ne11 = src1->ne[1];
+    const uint64_t ne12 = src1->ne[2];
+    const uint64_t ne13 = src1->ne[3];
+
+    const uint64_t ne20 = dst->ne[0];
+    const uint64_t ne21 = dst->ne[1];
+
+    const uint64_t r2 = ne12 / ne02;
+    const uint64_t r3 = ne13 / ne03;
+
+    ggml_backend_vk_buffer_context * dst_buf_ctx = (ggml_backend_vk_buffer_context *)dst->buffer->context;
+    ggml_backend_vk_buffer_context * src0_buf_ctx = (ggml_backend_vk_buffer_context *)src0->buffer->context;
+    ggml_backend_vk_buffer_context * src1_buf_ctx = (ggml_backend_vk_buffer_context *)src1->buffer->context;
+
+    vk_buffer d_Qx;
+    size_t qx_buf_offset = 0;
+    vk_buffer d_Qy;
+    size_t qy_buf_offset = 0;
+
+    bool src0_uma = false;
+    bool src1_uma = false;
+
+    if (ctx->device->uma) {
+        ggml_vk_host_get(ctx->device, src0->data, d_Qx, qx_buf_offset);
+        ggml_vk_host_get(ctx->device, src1->data, d_Qy, qy_buf_offset);
+        src0_uma = d_Qx != nullptr;
+        src1_uma = d_Qy != nullptr;
+    }
+
+    const bool x_non_contig = !ggml_vk_dim01_contiguous(src0);
+    const bool y_non_contig = !ggml_vk_dim01_contiguous(src1);
+
+    const bool y_f32_kernel = src1->type == GGML_TYPE_F32 && !y_non_contig;
+
+    vk_matmul_pipeline mmp = ggml_vk_get_mul_mat_mat_pipeline(ctx, src0->type, y_non_contig ? GGML_TYPE_F16 : src1->type);
+
+    const bool qx_needs_dequant = mmp == nullptr || x_non_contig;
+    const bool qy_needs_dequant = (src1->type != GGML_TYPE_F16 && !y_f32_kernel) || y_non_contig;
+
+    if (mmp == nullptr) {
+        // Fall back to dequant + f16 mulmat
+        mmp = ggml_vk_get_mul_mat_mat_pipeline(ctx, GGML_TYPE_F16, y_f32_kernel ? GGML_TYPE_F32 : GGML_TYPE_F16);
+    }
+
+    // Not implemented
+    GGML_ASSERT(y_non_contig || !qy_needs_dequant);  // NOLINT
+
+    const int x_ne = ne01 * ne00;
+    const int y_ne = ne11 * ne10;
+    const int d_ne = ne11 * ne01;
+
+    const uint32_t kpad = ggml_vk_align_size(ne10, ggml_vk_guess_matmul_pipeline_align(ctx, mmp, ne01, ne11));
+    const bool aligned = ne10 == kpad && ne01 > 8 && ne11 > 8;
+
+    const uint32_t split_k = ggml_vk_guess_split_k(ne01, ne11, ne10);
+
+    vk_pipeline pipeline = ggml_vk_guess_matmul_pipeline(ctx, mmp, ne01, ne11, aligned);
+
+    const uint64_t qx_sz = ggml_type_size(src0->type) * x_ne / ggml_blck_size(src0->type);
+    const uint64_t qy_sz = ggml_type_size(src1->type) * y_ne / ggml_blck_size(src1->type);
+    const uint64_t x_sz = !qx_needs_dequant ? qx_sz : sizeof(ggml_fp16_t) * x_ne;
+    const uint64_t y_sz = y_f32_kernel ? sizeof(float) * y_ne : sizeof(ggml_fp16_t) * y_ne;
+    const uint64_t d_sz = sizeof(float) * d_ne;
+
+    vk_pipeline to_fp16_vk_0 = nullptr;
+    vk_pipeline to_fp16_vk_1 = nullptr;
+
+    if (x_non_contig) {
+        to_fp16_vk_0 = ggml_vk_get_cpy_pipeline(ctx, src0->type, GGML_TYPE_F16);
+    } else {
+        to_fp16_vk_0 = ggml_vk_get_to_fp16(ctx, src0->type);
+    }
+    if (y_non_contig) {
+        to_fp16_vk_1 = ggml_vk_get_cpy_pipeline(ctx, src1->type, GGML_TYPE_F16);
+    } else {
+        to_fp16_vk_1 = ggml_vk_get_to_fp16(ctx, src1->type);
+    }
+    GGML_ASSERT(!qx_needs_dequant || to_fp16_vk_0 != nullptr);  // NOLINT
+    GGML_ASSERT(!qy_needs_dequant || to_fp16_vk_1 != nullptr);  // NOLINT
+
+    if (dryrun) {
+        const uint64_t x_sz_upd = x_sz * ne02 * ne03;
+        const uint64_t y_sz_upd = y_sz * ne12 * ne13;
+        const uint64_t split_k_size = split_k > 1 ? d_sz * ne12 * ne13 * 4 : 0;
+        if (
+                (qx_needs_dequant && x_sz_upd > ctx->device->max_memory_allocation_size) ||
+                (qy_needs_dequant && y_sz_upd > ctx->device->max_memory_allocation_size) ||
+                (split_k > 1 && split_k_size > ctx->device->max_memory_allocation_size)) {
+            GGML_ABORT("Requested preallocation size is too large");
+        }
+        if (qx_needs_dequant && ctx->prealloc_size_x < x_sz_upd) {
+            ctx->prealloc_size_x = x_sz_upd;
+        }
+        if (qy_needs_dequant && ctx->prealloc_size_y < y_sz_upd) {
+            ctx->prealloc_size_y = y_sz_upd;
+        }
+        if (split_k > 1 && ctx->prealloc_size_split_k < split_k_size) {
+            ctx->prealloc_size_split_k = split_k_size;
+        }
+
+        // Request descriptor sets
+        ggml_pipeline_request_descriptor_sets(ctx->device, pipeline, 1);
+        if (qx_needs_dequant) {
+            ggml_pipeline_request_descriptor_sets(ctx->device, to_fp16_vk_0, 1);
+        }
+        if (qy_needs_dequant) {
+            ggml_pipeline_request_descriptor_sets(ctx->device, to_fp16_vk_1, 1);
+        }
+        if (split_k > 1) {
+            ggml_pipeline_request_descriptor_sets(ctx->device, ctx->device->pipeline_matmul_split_k_reduce, 1);
+        }
+        return;
+    }
+
+    vk_buffer d_D = dst_buf_ctx->dev_buffer;
+    const uint64_t d_buf_offset = vk_tensor_offset(dst) + dst->view_offs;
+    GGML_ASSERT(d_D != nullptr);
+    GGML_ASSERT(d_D->size >= d_buf_offset + d_sz * ne02 * ne03);
+    vk_buffer d_X;
+    uint64_t x_buf_offset = 0;
+    vk_buffer d_Y;
+    uint64_t y_buf_offset = 0;
+    if (!src0_uma) {
+        d_Qx = src0_buf_ctx->dev_buffer;
+        qx_buf_offset = vk_tensor_offset(src0) + src0->view_offs;
+        GGML_ASSERT(d_Qx != nullptr);
+    }
+    if (!src1_uma) {
+        d_Qy = src1_buf_ctx->dev_buffer;
+        qy_buf_offset = vk_tensor_offset(src1) + src1->view_offs;
+        GGML_ASSERT(d_Qy != nullptr);
+    }
+    if (qx_needs_dequant) {
+        d_X = ctx->prealloc_x;
+        GGML_ASSERT(d_X->size >= x_sz * ne02 * ne03);
+    } else {
+        d_X = d_Qx;
+        x_buf_offset = qx_buf_offset;
+        GGML_ASSERT(qx_sz == x_sz);
+    }
+    if (qy_needs_dequant) {
+        d_Y = ctx->prealloc_y;
+        GGML_ASSERT(d_Y->size >= y_sz * ne02 * ne03);
+    } else {
+        d_Y = d_Qy;
+        y_buf_offset = qy_buf_offset;
+        GGML_ASSERT(qy_sz == y_sz);
+    }
+
+    if (x_non_contig) {
+        ggml_vk_cpy_to_contiguous(ctx, subctx, to_fp16_vk_0, src0, { d_Qx, qx_buf_offset, VK_WHOLE_SIZE }, { d_X, 0, VK_WHOLE_SIZE });
+    } else if (qx_needs_dequant) {
+        const std::vector<uint32_t> pc = { (uint32_t)ne01, (uint32_t)ne10, (uint32_t)ne10, (uint32_t)ne10, (uint32_t)(ggml_nelements(src0)) };
+        ggml_vk_sync_buffers(subctx);
+        ggml_vk_dispatch_pipeline(ctx, subctx, to_fp16_vk_0, { vk_subbuffer{ d_Qx, qx_buf_offset, qx_sz * ne02 * ne03 }, vk_subbuffer{ d_X, 0, x_sz * ne02 * ne03 } }, pc.size() * sizeof(uint32_t), pc.data(), { (uint32_t)(x_ne * ne02 * ne03), 1, 1});
+    }
+    if (y_non_contig) {
+        ggml_vk_cpy_to_contiguous(ctx, subctx, to_fp16_vk_1, src1, { d_Qy, qy_buf_offset, VK_WHOLE_SIZE }, { d_Y, 0, VK_WHOLE_SIZE });
+    }
+
+    uint32_t stride_batch_x = ne00*ne01;
+    uint32_t stride_batch_y = ne10*ne11;
+
+    if (!ggml_vk_dim01_contiguous(src0) && !qx_needs_dequant) {
+        stride_batch_x = src0->nb[0] / ggml_type_size(src0->type);
+    }
+
+    if (!ggml_vk_dim01_contiguous(src1) && !qy_needs_dequant) {
+        stride_batch_y = src1->nb[0] / ggml_type_size(src1->type);
+    }
+
+    // compute
+    ggml_vk_matmul(
+        ctx, subctx, pipeline,
+        { d_X, x_buf_offset, x_sz * ne02 * ne03 }, { d_Y, y_buf_offset, y_sz * ne12 * ne13 },
+        { d_D, d_buf_offset, d_sz * ne12 * ne13 }, { ctx->prealloc_split_k, 0, d_sz * ne12 * ne13 * split_k },
+        ne01, ne11, ne10,
+        ne10, ne10, ne01, stride_batch_x, stride_batch_y, ne20*ne21,
+        split_k, ne12*ne13, ne02, ne12, r2, r3
+    );  // NOLINT
+}
+
+static void ggml_vk_mul_mat_vec_q_f16(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, bool dryrun = false) {
+    VK_LOG_DEBUG("ggml_vk_mul_mat_vec_q_f16((" << src0 << ", name=" << src0->name << ", type=" << src0->type << ", ne0=" << src0->ne[0] << ", ne1=" << src0->ne[1] << ", ne2=" << src0->ne[2] << ", ne3=" << src0->ne[3] << ", nb0=" << src0->nb[0] << ", nb1=" << src0->nb[1] << ", nb2=" << src0->nb[2] << ", nb3=" << src0->nb[3];
+    std::cerr << "), (" << src1 << ", name=" << src1->name << ", type=" << src1->type << ", ne0=" << src1->ne[0] << ", ne1=" << src1->ne[1] << ", ne2=" << src1->ne[2] << ", ne3=" << src1->ne[3] << ", nb0=" << src1->nb[0] << ", nb1=" << src1->nb[1] << ", nb2=" << src1->nb[2] << ", nb3=" << src1->nb[3];
+    std::cerr << "), (" << dst << ", name=" << dst->name << ", type=" << dst->type << ", ne0=" << dst->ne[0] << ", ne1=" << dst->ne[1] << ", ne2=" << dst->ne[2] << ", ne3=" << dst->ne[3] << ", nb0=" << dst->nb[0] << ", nb1=" << dst->nb[1] << ", nb2=" << dst->nb[2] << ", nb3=" << dst->nb[3];
+    std::cerr << "), " << (dryrun ? "dryrun" : "") << "),)");
+    GGML_ASSERT(ggml_vk_dim01_contiguous(src0) || src0->type == GGML_TYPE_F32 || src0->type == GGML_TYPE_F16);  // NOLINT
+    GGML_ASSERT(ggml_vk_dim01_contiguous(src1) || src1->type == GGML_TYPE_F32 || src1->type == GGML_TYPE_F16);  // NOLINT
+
+    const uint64_t ne00 = src0->ne[0];
+    const uint64_t ne01 = src0->ne[1];
+    const uint64_t ne02 = src0->ne[2];
+    const uint64_t ne03 = src0->ne[3];
+
+    const uint64_t ne10 = src1->ne[0];
+    const uint64_t ne11 = src1->ne[1];
+    const uint64_t ne12 = src1->ne[2];
+    const uint64_t ne13 = src1->ne[3];
+
+    GGML_ASSERT(ne11 == 1);
+
+    const uint64_t ne20 = dst->ne[0];
+    const uint64_t ne21 = dst->ne[1];
+    const uint64_t ne22 = dst->ne[2];
+    const uint64_t ne23 = dst->ne[3];
+
+    const uint64_t r2 = ne12 / ne02;
+    const uint64_t r3 = ne13 / ne03;
+
+    ggml_backend_vk_buffer_context * dst_buf_ctx = (ggml_backend_vk_buffer_context *)dst->buffer->context;
+    ggml_backend_vk_buffer_context * src0_buf_ctx = (ggml_backend_vk_buffer_context *)src0->buffer->context;
+    ggml_backend_vk_buffer_context * src1_buf_ctx = (ggml_backend_vk_buffer_context *)src1->buffer->context;
+
+    vk_buffer d_Qx;
+    size_t qx_buf_offset = 0;
+    vk_buffer d_Qy;
+    size_t qy_buf_offset = 0;
+
+    bool src0_uma = false;
+    bool src1_uma = false;
+
+    if (ctx->device->uma) {
+        ggml_vk_host_get(ctx->device, src0->data, d_Qx, qx_buf_offset);
+        ggml_vk_host_get(ctx->device, src1->data, d_Qy, qy_buf_offset);
+        src0_uma = d_Qx != nullptr;
+        src1_uma = d_Qy != nullptr;
+    }
+
+    const bool x_non_contig = !ggml_vk_dim01_contiguous(src0);
+    const bool y_non_contig = !ggml_vk_dim01_contiguous(src1);
+
+    const bool f16_f32_kernel = src1->type == GGML_TYPE_F32;
+
+    const bool qx_needs_dequant = x_non_contig;
+    const bool qy_needs_dequant = (src1->type != GGML_TYPE_F16 && !f16_f32_kernel) || y_non_contig;
+
+    // Not implemented
+    GGML_ASSERT(y_non_contig || !qy_needs_dequant);  // NOLINT
+
+    const uint64_t x_ne = ne01 * ne00;
+    const uint64_t y_ne = ne11 * ne10;
+    const uint64_t d_ne = ne11 * ne01;
+
+    const uint64_t qx_sz = ggml_vk_align_size(ggml_type_size(src0->type) * x_ne / ggml_blck_size(src0->type), ctx->device->properties.limits.minStorageBufferOffsetAlignment);
+    const uint64_t qy_sz = ggml_type_size(src1->type) * y_ne / ggml_blck_size(src1->type);
+    const uint64_t x_sz = x_non_contig ? ggml_vk_align_size(ggml_type_size(src0->type) * x_ne, ctx->device->properties.limits.minStorageBufferOffsetAlignment) : qx_sz;
+    const uint64_t y_sz = f16_f32_kernel ? sizeof(float) * y_ne : sizeof(ggml_fp16_t) * y_ne;
+    const uint64_t d_sz = sizeof(float) * d_ne;
+
+    vk_pipeline to_fp16_vk_0 = nullptr;
+    vk_pipeline to_fp16_vk_1 = nullptr;
+    if (x_non_contig) {
+        to_fp16_vk_0 = ggml_vk_get_cpy_pipeline(ctx, src0->type, src0->type);
+    }
+    if (y_non_contig) {
+        to_fp16_vk_1 = ggml_vk_get_cpy_pipeline(ctx, src1->type, src1->type);
+    } else {
+        to_fp16_vk_1 = ggml_vk_get_to_fp16(ctx, src1->type);
+    }
+    vk_pipeline dmmv = ggml_vk_get_dequantize_mul_mat_vec(ctx, src0->type, src1->type);
+    GGML_ASSERT(!qx_needs_dequant || to_fp16_vk_0 != nullptr);  // NOLINT
+    GGML_ASSERT(!qy_needs_dequant || to_fp16_vk_1 != nullptr);  // NOLINT
+    GGML_ASSERT(dmmv != nullptr);
+
+    if (dryrun) {
+        const uint64_t x_sz_upd = x_sz * ne02 * ne03;
+        const uint64_t y_sz_upd = y_sz * ne12 * ne13;
+        if (
+                (qx_needs_dequant && x_sz_upd > ctx->device->max_memory_allocation_size) ||
+                (qy_needs_dequant && y_sz_upd > ctx->device->max_memory_allocation_size)) {
+            GGML_ABORT("Requested preallocation size is too large");
+        }
+        if (qx_needs_dequant && ctx->prealloc_size_x < x_sz_upd) {
+            ctx->prealloc_size_x = x_sz_upd;
+        }
+        if (qy_needs_dequant && ctx->prealloc_size_y < y_sz_upd) {
+            ctx->prealloc_size_y = y_sz_upd;
+        }
+
+        // Request descriptor sets
+        if (qx_needs_dequant) {
+            ggml_pipeline_request_descriptor_sets(ctx->device, to_fp16_vk_0, 1);
+        }
+        if (qy_needs_dequant) {
+            ggml_pipeline_request_descriptor_sets(ctx->device, to_fp16_vk_1, 1);
+        }
+        ggml_pipeline_request_descriptor_sets(ctx->device, dmmv, 1);
+        return;
+    }
+
+    vk_buffer d_D = dst_buf_ctx->dev_buffer;
+    const uint64_t d_buf_offset = vk_tensor_offset(dst) + dst->view_offs;
+    GGML_ASSERT(d_D != nullptr);
+    vk_buffer d_X;
+    uint64_t x_buf_offset = 0;
+    vk_buffer d_Y;
+    uint64_t y_buf_offset = 0;
+    if(!src0_uma) {
+        d_Qx = src0_buf_ctx->dev_buffer;
+        qx_buf_offset = vk_tensor_offset(src0) + src0->view_offs;
+        GGML_ASSERT(d_Qx != nullptr);
+    }
+    if(!src1_uma) {
+        d_Qy = src1_buf_ctx->dev_buffer;
+        qy_buf_offset = vk_tensor_offset(src1) + src1->view_offs;
+        GGML_ASSERT(d_Qy != nullptr);
+    }
+    if (qx_needs_dequant) {
+        d_X = ctx->prealloc_x;
+    } else {
+        d_X = d_Qx;
+        x_buf_offset = qx_buf_offset;
+        GGML_ASSERT(qx_sz == x_sz);
+    }
+    if (qy_needs_dequant) {
+        d_Y = ctx->prealloc_y;
+    } else {
+        d_Y = d_Qy;
+        y_buf_offset = qy_buf_offset;
+        GGML_ASSERT(qy_sz == y_sz);
+    }
+
+    if (x_non_contig) {
+        GGML_ASSERT(x_sz == ggml_vk_align_size(ggml_type_size(src0->type) * x_ne, ctx->device->properties.limits.minStorageBufferOffsetAlignment));
+        ggml_vk_cpy_to_contiguous(ctx, subctx, to_fp16_vk_0, src0, { d_Qx, qx_buf_offset, VK_WHOLE_SIZE }, { d_X, 0, VK_WHOLE_SIZE });
+    }
+    if (y_non_contig) {
+        GGML_ASSERT(y_sz == ggml_type_size(src1->type) * y_ne);
+        ggml_vk_cpy_to_contiguous(ctx, subctx, to_fp16_vk_1, src1, { d_Qy, qy_buf_offset, VK_WHOLE_SIZE }, { d_Y, 0, VK_WHOLE_SIZE });
+    }
+
+    uint32_t stride_batch_x = ne00*ne01;
+    uint32_t stride_batch_y = ne10*ne11;
+
+    if (!ggml_vk_dim01_contiguous(src0) && !qx_needs_dequant) {
+        stride_batch_x = src0->nb[0] / ggml_type_size(src0->type);
+    }
+
+    if (!ggml_vk_dim01_contiguous(src1) && !qy_needs_dequant) {
+        stride_batch_y = src1->nb[0] / ggml_type_size(src1->type);
+    }
+
+    const uint32_t max_groups_x = ctx->device->properties.limits.maxComputeWorkGroupCount[0];
+
+    uint32_t groups_x = ne01;
+    uint32_t groups_z = 1;
+
+    if (ne01 > max_groups_x) {
+        groups_z = 64;
+        groups_x /= groups_z;
+    }
+
+    // compute
+    const vk_mat_vec_push_constants pc = {
+        (uint32_t)ne00, (uint32_t)ne10, (uint32_t)ne10, (uint32_t)ne01,
+        stride_batch_x, stride_batch_y, (uint32_t)(ne20*ne21),
+        (uint32_t)ne02, (uint32_t)ne12, (uint32_t)r2, (uint32_t)r3,
+    };
+    ggml_vk_sync_buffers(subctx);
+    ggml_vk_dispatch_pipeline(ctx, subctx, dmmv,
+                              { vk_subbuffer{ d_X, x_buf_offset, x_sz * ne02 * ne03 }, vk_subbuffer{ d_Y, y_buf_offset, y_sz * ne12 * ne13 }, vk_subbuffer{ d_D, d_buf_offset, d_sz * ne22 * ne23} },
+                              sizeof(vk_mat_vec_push_constants), &pc, { groups_x, (uint32_t)(ne12 * ne13), groups_z });
+}
+
+static void ggml_vk_mul_mat_vec_p021_f16_f32(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, bool dryrun = false) {
+    VK_LOG_DEBUG("ggml_vk_mul_mat_p021_f16_f32(" << src0 << ", name=" << src0->name << ", type=" << src0->type << ", ne0=" << src0->ne[0] << ", ne1=" << src0->ne[1] << ", ne2=" << src0->ne[2] << ", ne3=" << src0->ne[3] << ", nb0=" << src0->nb[0] << ", nb1=" << src0->nb[1] << ", nb2=" << src0->nb[2] << ", nb3=" << src0->nb[3];
+    std::cerr << "), (" << src1 << ", name=" << src1->name << ", type=" << src1->type << ", ne0=" << src1->ne[0] << ", ne1=" << src1->ne[1] << ", ne2=" << src1->ne[2] << ", ne3=" << src1->ne[3] << ", nb0=" << src1->nb[0] << ", nb1=" << src1->nb[1] << ", nb2=" << src1->nb[2] << ", nb3=" << src1->nb[3];
+    std::cerr << "), (" << dst << ", name=" << dst->name << ", type=" << dst->type << ", ne0=" << dst->ne[0] << ", ne1=" << dst->ne[1] << ", ne2=" << dst->ne[2] << ", ne3=" << dst->ne[3] << ", nb0=" << dst->nb[0] << ", nb1=" << dst->nb[1] << ", nb2=" << dst->nb[2] << ", nb3=" << dst->nb[3];
+    std::cerr << "), " << (dryrun ? "dryrun" : "") << ")");
+    GGML_ASSERT(ggml_is_permuted(src0) && ggml_is_permuted(src1));
+    GGML_ASSERT(src0->nb[0] <= src0->nb[1] && src0->nb[2] <= src0->nb[3]);  // NOLINT
+    GGML_ASSERT(src1->nb[0] <= src1->nb[1] && src1->nb[2] <= src1->nb[3]);  // NOLINT
+    GGML_ASSERT(src0->type == GGML_TYPE_F16);
+    GGML_ASSERT(src1->type == GGML_TYPE_F32);
+
+    const uint64_t ne00 = src0->ne[0];
+    const uint64_t ne01 = src0->ne[1];
+    const uint64_t ne02 = src0->ne[2];
+    // const uint64_t ne03 = src0->ne[3];
+
+    const uint64_t ne10 = src1->ne[0];
+    const uint64_t ne11 = src1->ne[1];
+    const uint64_t ne12 = src1->ne[2];
+    // const uint64_t ne13 = src1->ne[3];
+
+    GGML_ASSERT(ne11 == 1);
+
+    ggml_backend_vk_buffer_context * dst_buf_ctx = (ggml_backend_vk_buffer_context *)dst->buffer->context;
+    ggml_backend_vk_buffer_context * src0_buf_ctx = (ggml_backend_vk_buffer_context *)src0->buffer->context;
+    ggml_backend_vk_buffer_context * src1_buf_ctx = (ggml_backend_vk_buffer_context *)src1->buffer->context;
+
+    vk_buffer d_Qy;
+    size_t qy_buf_offset = 0;
+
+    bool src1_uma = false;
+
+    if (ctx->device->uma) {
+        ggml_vk_host_get(ctx->device, src1->data, d_Qy, qy_buf_offset);
+        src1_uma = d_Qy != nullptr;
+    }
+
+    const uint64_t x_ne = ne00 * ne01 * ne02;
+    const uint64_t y_ne = ne10 * ne11 * ne12;
+    const uint64_t d_ne = ne01 * ne11 * ne12;
+
+    const uint64_t qx_sz = ggml_vk_align_size(ggml_type_size(src0->type) * x_ne / ggml_blck_size(src0->type), ctx->device->properties.limits.minStorageBufferOffsetAlignment);
+    const uint64_t qy_sz = ggml_type_size(src1->type) * y_ne / ggml_blck_size(src1->type);
+    const uint64_t d_sz = sizeof(float) * d_ne;
+
+    if (dryrun) {
+        // Request descriptor sets
+        ggml_pipeline_request_descriptor_sets(ctx->device, ctx->device->pipeline_mul_mat_vec_p021_f16_f32, 1);
+        return;
+    }
+
+    vk_buffer d_D = dst_buf_ctx->dev_buffer;
+    const uint64_t d_buf_offset = vk_tensor_offset(dst) + dst->view_offs;
+    GGML_ASSERT(d_D != nullptr);
+    vk_buffer d_Qx = src0_buf_ctx->dev_buffer;
+    const uint64_t qx_buf_offset = vk_tensor_offset(src0) + src0->view_offs;
+    GGML_ASSERT(d_Qx != nullptr);
+    if (!src1_uma) {
+        d_Qy = src1_buf_ctx->dev_buffer;
+        qy_buf_offset = vk_tensor_offset(src1) + src1->view_offs;
+        GGML_ASSERT(d_Qx != nullptr);
+    }
+
+    const uint64_t qy_buffer_offset = (qy_buf_offset / ctx->device->properties.limits.minStorageBufferOffsetAlignment) * ctx->device->properties.limits.minStorageBufferOffsetAlignment;
+    const uint64_t qy_shader_offset = qy_buf_offset - qy_buffer_offset;
+
+    const uint64_t d_buffer_offset = (d_buf_offset / ctx->device->properties.limits.minStorageBufferOffsetAlignment) * ctx->device->properties.limits.minStorageBufferOffsetAlignment;
+    const uint64_t d_shader_offset = d_buf_offset - d_buffer_offset;
+
+    // compute
+    const std::array<uint32_t, 6> pc = { (uint32_t)ne00, (uint32_t)ne01, (uint32_t)ne02, (uint32_t)ne12, (uint32_t)(qy_shader_offset / ggml_type_size(src1->type)), (uint32_t)(d_shader_offset / ggml_type_size(dst->type)) };
+    ggml_vk_sync_buffers(subctx);
+    ggml_vk_dispatch_pipeline(ctx, subctx, ctx->device->pipeline_mul_mat_vec_p021_f16_f32, { vk_subbuffer{ d_Qx, qx_buf_offset, qx_sz }, vk_subbuffer{ d_Qy, qy_buffer_offset, qy_sz + qy_shader_offset }, vk_subbuffer{ d_D, d_buffer_offset, d_sz + d_shader_offset } }, 6 * sizeof(uint32_t), &pc, { 1, (uint32_t)ne01, (uint32_t)ne12 });
+}
+
+static void ggml_vk_mul_mat_vec_nc_f16_f32(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, bool dryrun = false) {
+    VK_LOG_DEBUG("ggml_vk_mul_mat_nc_f16_f32((" << src0 << ", name=" << src0->name << ", type=" << src0->type << ", ne0=" << src0->ne[0] << ", ne1=" << src0->ne[1] << ", ne2=" << src0->ne[2] << ", ne3=" << src0->ne[3] << ", nb0=" << src0->nb[0] << ", nb1=" << src0->nb[1] << ", nb2=" << src0->nb[2] << ", nb3=" << src0->nb[3];
+    std::cerr << "), (" << src1 << ", name=" << src1->name << ", type=" << src1->type << ", ne0=" << src1->ne[0] << ", ne1=" << src1->ne[1] << ", ne2=" << src1->ne[2] << ", ne3=" << src1->ne[3] << ", nb0=" << src1->nb[0] << ", nb1=" << src1->nb[1] << ", nb2=" << src1->nb[2] << ", nb3=" << src1->nb[3];
+    std::cerr << "), (" << dst << ", name=" << dst->name << ", type=" << dst->type << ", ne0=" << dst->ne[0] << ", ne1=" << dst->ne[1] << ", ne2=" << dst->ne[2] << ", ne3=" << dst->ne[3] << ", nb0=" << dst->nb[0] << ", nb1=" << dst->nb[1] << ", nb2=" << dst->nb[2] << ", nb3=" << dst->nb[3];
+    std::cerr << "), " << (dryrun ? "dryrun" : "") << ")");
+    GGML_ASSERT(!ggml_is_transposed(src0));
+    GGML_ASSERT(!ggml_is_transposed(src1));
+    GGML_ASSERT(!ggml_is_permuted(src0));
+    GGML_ASSERT(src0->type == GGML_TYPE_F16);
+    GGML_ASSERT(src1->type == GGML_TYPE_F32);
+
+    const uint64_t ne00 = src0->ne[0];
+    const uint64_t ne01 = src0->ne[1];
+    const uint64_t ne02 = src0->ne[2];
+    // const uint64_t ne03 = src0->ne[3];
+
+    const uint64_t nb01 = src0->nb[1];
+    const uint64_t nb02 = src0->nb[2];
+
+    // const uint64_t ne10 = src1->ne[0];
+    const uint64_t ne11 = src1->ne[1];
+    const uint64_t ne12 = src1->ne[2];
+    // const uint64_t ne13 = src1->ne[3];
+
+    GGML_ASSERT(ne11 == 1);
+
+    ggml_backend_vk_buffer_context * dst_buf_ctx = (ggml_backend_vk_buffer_context *)dst->buffer->context;
+    ggml_backend_vk_buffer_context * src0_buf_ctx = (ggml_backend_vk_buffer_context *)src0->buffer->context;
+    ggml_backend_vk_buffer_context * src1_buf_ctx = (ggml_backend_vk_buffer_context *)src1->buffer->context;
+
+    vk_buffer d_Qy = nullptr;
+    size_t qy_buf_offset = 0;
+
+    bool src1_uma = false;
+
+    if (ctx->device->uma) {
+        ggml_vk_host_get(ctx->device, src1->data, d_Qy, qy_buf_offset);
+        src1_uma = d_Qy != nullptr;
+    }
+
+    const uint64_t d_ne = ne01 * ne11 * ne12;
+
+    const uint32_t row_stride_x = nb01 / sizeof(ggml_fp16_t);
+    const uint32_t channel_stride_x = nb02 / sizeof(ggml_fp16_t);
+
+    const uint64_t qx_sz = ggml_nbytes(src0);
+    const uint64_t qy_sz = ggml_nbytes(src1);
+    const uint64_t d_sz = sizeof(float) * d_ne;
+
+    if (dryrun) {
+        // Request descriptor sets
+        ggml_pipeline_request_descriptor_sets(ctx->device, ctx->device->pipeline_mul_mat_vec_nc_f16_f32, 1);
+        return;
+    }
+
+    vk_buffer d_D = dst_buf_ctx->dev_buffer;
+    const uint64_t d_buf_offset = vk_tensor_offset(dst) + dst->view_offs;
+    GGML_ASSERT(d_D != nullptr);
+    vk_buffer d_Qx = src0_buf_ctx->dev_buffer;
+    const uint64_t qx_buf_offset = vk_tensor_offset(src0) + src0->view_offs;
+    GGML_ASSERT(d_Qx != nullptr);
+    if (!src1_uma) {
+        d_Qy = src1_buf_ctx->dev_buffer;
+        qy_buf_offset = vk_tensor_offset(src1) + src1->view_offs;
+        GGML_ASSERT(d_Qx != nullptr);
+    }
+
+    const uint64_t qy_buffer_offset = (qy_buf_offset / ctx->device->properties.limits.minStorageBufferOffsetAlignment) * ctx->device->properties.limits.minStorageBufferOffsetAlignment;
+    const uint64_t qy_shader_offset = qy_buf_offset - qy_buffer_offset;
+
+    const uint64_t d_buffer_offset = (d_buf_offset / ctx->device->properties.limits.minStorageBufferOffsetAlignment) * ctx->device->properties.limits.minStorageBufferOffsetAlignment;
+    const uint64_t d_shader_offset = d_buf_offset - d_buffer_offset;
+
+    // compute
+    const std::array<uint32_t, 7> pc = { (uint32_t)ne00, (uint32_t)ne01, row_stride_x, channel_stride_x, (uint32_t)(ne12 / ne02), (uint32_t)(qy_shader_offset / ggml_type_size(src1->type)), (uint32_t)(d_shader_offset / ggml_type_size(dst->type)) };
+    ggml_vk_sync_buffers(subctx);
+    ggml_vk_dispatch_pipeline(ctx, subctx, ctx->device->pipeline_mul_mat_vec_nc_f16_f32,
+        { vk_subbuffer{ d_Qx, qx_buf_offset, qx_sz }, vk_subbuffer{ d_Qy, qy_buffer_offset, qy_sz + qy_shader_offset }, vk_subbuffer{ d_D, d_buffer_offset, d_sz + d_shader_offset } }, 7 * sizeof(uint32_t), &pc, { 1, (uint32_t)ne01, (uint32_t)ne12 });
+}
+
+static void ggml_vk_mul_mat(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, bool dryrun = false) {
+    VK_LOG_DEBUG("ggml_vk_mul_mat(" << src0 << ", " << src1 << ", " << dst << ")");
+    if (src0->type == GGML_TYPE_F16 && ggml_is_permuted(src0) && ggml_is_permuted(src1) && dst->ne[1] == 1) {
+        ggml_vk_mul_mat_vec_p021_f16_f32(ctx, subctx, src0, src1, dst, dryrun);
+    } else if (src0->type == GGML_TYPE_F16 && !ggml_is_contiguous(src0) && !ggml_is_transposed(src1) && dst->ne[1] == 1) {
+        ggml_vk_mul_mat_vec_nc_f16_f32(ctx, subctx, src0, src1, dst, dryrun);
+    } else if (dst->ne[1] == 1 && (src0->type == GGML_TYPE_F32 || src0->type == GGML_TYPE_F16 || ggml_is_quantized(src0->type))) {
+        ggml_vk_mul_mat_vec_q_f16(ctx, subctx, src0, src1, dst, dryrun);
+    } else {
+        ggml_vk_mul_mat_q_f16(ctx, subctx, src0, src1, dst, dryrun);
+    }
+}
+
+static void ggml_vk_mul_mat_id_q_f16(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, const ggml_tensor * src1, const ggml_tensor * ids, ggml_tensor * dst, bool dryrun = false) {
+    VK_LOG_DEBUG("ggml_vk_mul_mat_id_q_f16((" << src0 << ", name=" << src0->name << ", type=" << src0->type << ", ne0=" << src0->ne[0] << ", ne1=" << src0->ne[1] << ", ne2=" << src0->ne[2] << ", ne3=" << src0->ne[3] << ", nb0=" << src0->nb[0] << ", nb1=" << src0->nb[1] << ", nb2=" << src0->nb[2] << ", nb3=" << src0->nb[3];
+    std::cerr << "), (" << src1 << ", name=" << src1->name << ", type=" << src1->type << ", ne0=" << src1->ne[0] << ", ne1=" << src1->ne[1] << ", ne2=" << src1->ne[2] << ", ne3=" << src1->ne[3] << ", nb0=" << src1->nb[0] << ", nb1=" << src1->nb[1] << ", nb2=" << src1->nb[2] << ", nb3=" << src1->nb[3];
+    std::cerr << "), (" << ids << ", name=" << ids->name << ", type=" << ids->type << ", ne0=" << ids->ne[0] << ", ne1=" << ids->ne[1] << ", ne2=" << ids->ne[2] << ", ne3=" << ids->ne[3] << ", nb0=" << ids->nb[0] << ", nb1=" << ids->nb[1] << ", nb2=" << ids->nb[2] << ", nb3=" << ids->nb[3];
+    std::cerr << "), (" << dst << ", name=" << dst->name << ", type=" << dst->type << ", ne0=" << dst->ne[0] << ", ne1=" << dst->ne[1] << ", ne2=" << dst->ne[2] << ", ne3=" << dst->ne[3] << ", nb0=" << dst->nb[0] << ", nb1=" << dst->nb[1] << ", nb2=" << dst->nb[2] << ", nb3=" << dst->nb[3] << "),)");
+    GGML_ASSERT(ggml_vk_dim01_contiguous(src1) || src1->type == GGML_TYPE_F32 || src1->type == GGML_TYPE_F16);  // NOLINT
+    GGML_ASSERT(ids->type == GGML_TYPE_I32);
+
+    const uint64_t ne00 = src0->ne[0];
+    const uint64_t ne01 = src0->ne[1];
+    const uint64_t ne02 = src0->ne[2];
+    const uint64_t ne03 = src0->ne[3];
+
+    const uint64_t ne10 = src1->ne[0];
+    const uint64_t ne11 = src1->ne[1];
+    const uint64_t ne12 = src1->ne[2];
+    const uint64_t ne13 = src1->ne[3];
+
+    const uint64_t nei0 = ids->ne[0];
+    const uint64_t nei1 = ids->ne[1];
+    GGML_ASSERT(nei0 * nei1 <= 3072);
+
+    const uint32_t nbi1 = ids->nb[1];
+    const uint32_t nbi2 = ids->nb[2];
+
+    const uint64_t ne20 = dst->ne[0];
+    const uint64_t ne21 = dst->ne[1];
+    const uint64_t ne22 = dst->ne[2];
+    const uint64_t ne23 = dst->ne[3];
+
+    const uint64_t n_as = ne02;
+
+    ggml_backend_vk_buffer_context * dst_buf_ctx = (ggml_backend_vk_buffer_context *)dst->buffer->context;
+    ggml_backend_vk_buffer_context * src0_buf_ctx = (ggml_backend_vk_buffer_context *)src0->buffer->context;
+    ggml_backend_vk_buffer_context * src1_buf_ctx = (ggml_backend_vk_buffer_context *)src1->buffer->context;
+    ggml_backend_vk_buffer_context * ids_buf_ctx = (ggml_backend_vk_buffer_context *)ids->buffer->context;
+
+    vk_buffer d_Qx;
+    size_t qx_buf_offset = 0;
+    vk_buffer d_Qy;
+    size_t qy_buf_offset = 0;
+    vk_buffer d_ids;
+    size_t ids_buf_offset = 0;
+
+    bool src0_uma = false;
+    bool src1_uma = false;
+    bool ids_uma = false;
+
+    if (ctx->device->uma) {
+        ggml_vk_host_get(ctx->device, src0->data, d_Qx, qx_buf_offset);
+        ggml_vk_host_get(ctx->device, src1->data, d_Qy, qy_buf_offset);
+        ggml_vk_host_get(ctx->device, ids->data, d_ids, ids_buf_offset);
+        src0_uma = d_Qx != nullptr;
+        src1_uma = d_Qy != nullptr;
+        ids_uma = d_ids != nullptr;
+    }
+
+    const bool x_non_contig = !ggml_vk_dim01_contiguous(src0);
+    const bool y_non_contig = !ggml_vk_dim01_contiguous(src1);
+
+    const bool y_f32_kernel = src1->type == GGML_TYPE_F32 && !y_non_contig;
+
+    vk_matmul_pipeline mmp = ggml_vk_get_mul_mat_mat_id_pipeline(ctx, src0->type, y_non_contig ? GGML_TYPE_F16 : src1->type);
+
+    const bool qx_needs_dequant = mmp == nullptr || x_non_contig;
+    const bool qy_needs_dequant = (src1->type != GGML_TYPE_F16 && !y_f32_kernel) || y_non_contig;
+
+    if (mmp == nullptr) {
+        GGML_ABORT("fatal error");
+    }
+
+    // Not implemented
+    GGML_ASSERT(y_non_contig || !qy_needs_dequant);  // NOLINT
+
+    const uint64_t x_ne = ne01 * ne00;
+    const uint64_t y_ne = ne11 * ne10;
+    const uint64_t d_ne = ne21 * ne20;
+
+    const uint32_t kpad = ggml_vk_align_size(ne10, ggml_vk_guess_matmul_pipeline_align(ctx, mmp, ne01, nei1));
+    const bool aligned = ne10 == kpad && ne01 > 8 && nei1 > 8;
+
+    vk_pipeline pipeline = ggml_vk_guess_matmul_pipeline(ctx, mmp, ne01, nei1, aligned);
+
+    const uint64_t qx_sz = ggml_type_size(src0->type) * x_ne / ggml_blck_size(src0->type);
+    const uint64_t qy_sz = ggml_type_size(src1->type) * y_ne / ggml_blck_size(src1->type);
+    const uint64_t x_sz = !qx_needs_dequant ? qx_sz : sizeof(ggml_fp16_t) * x_ne;
+    const uint64_t y_sz = y_f32_kernel ? sizeof(float) * y_ne : sizeof(ggml_fp16_t) * y_ne;
+    const uint64_t ids_sz = nbi2;
+    const uint64_t d_sz = sizeof(float) * d_ne;
+
+    vk_pipeline to_fp16_vk_0 = nullptr;
+    vk_pipeline to_fp16_vk_1 = nullptr;
+
+    if (x_non_contig) {
+        to_fp16_vk_0 = ggml_vk_get_cpy_pipeline(ctx, src0->type, GGML_TYPE_F16);
+    } else {
+        to_fp16_vk_0 = ggml_vk_get_to_fp16(ctx, src0->type);
+    }
+    if (y_non_contig) {
+        to_fp16_vk_1 = ggml_vk_get_cpy_pipeline(ctx, src1->type, GGML_TYPE_F16);
+    } else {
+        to_fp16_vk_1 = ggml_vk_get_to_fp16(ctx, src1->type);
+    }
+    GGML_ASSERT(!qx_needs_dequant || to_fp16_vk_0 != nullptr);  // NOLINT
+    GGML_ASSERT(!qy_needs_dequant || to_fp16_vk_1 != nullptr);  // NOLINT
+
+    if (dryrun) {
+        const uint64_t x_sz_upd = x_sz * ne02 * ne03;
+        const uint64_t y_sz_upd = y_sz * ne12 * ne13;
+        if (
+                (qx_needs_dequant && x_sz_upd > ctx->device->max_memory_allocation_size) ||
+                (qy_needs_dequant && y_sz_upd > ctx->device->max_memory_allocation_size)) {
+            GGML_ABORT("Requested preallocation size is too large");
+        }
+        if (qx_needs_dequant && ctx->prealloc_size_x < x_sz_upd) {
+            ctx->prealloc_size_x = x_sz_upd;
+        }
+        if (qy_needs_dequant && ctx->prealloc_size_y < y_sz_upd) {
+            ctx->prealloc_size_y = y_sz_upd;
+        }
+
+        // Request descriptor sets
+        ggml_pipeline_request_descriptor_sets(ctx->device, pipeline, 1);
+        if (qx_needs_dequant) {
+            ggml_pipeline_request_descriptor_sets(ctx->device, to_fp16_vk_0, 1);
+        }
+        if (qy_needs_dequant) {
+            ggml_pipeline_request_descriptor_sets(ctx->device, to_fp16_vk_1, 1);
+        }
+        return;
+    }
+
+    vk_buffer d_D = dst_buf_ctx->dev_buffer;
+    const uint64_t d_buf_offset = vk_tensor_offset(dst) + dst->view_offs;
+    GGML_ASSERT(d_D != nullptr);
+    vk_buffer d_X;
+    uint64_t x_buf_offset = 0;
+    vk_buffer d_Y;
+    uint64_t y_buf_offset = 0;
+    if (!src0_uma) {
+        d_Qx = src0_buf_ctx->dev_buffer;
+        qx_buf_offset = vk_tensor_offset(src0) + src0->view_offs;
+        GGML_ASSERT(d_Qx != nullptr);
+    }
+    if (!src1_uma) {
+        d_Qy = src1_buf_ctx->dev_buffer;
+        qy_buf_offset = vk_tensor_offset(src1) + src1->view_offs;
+        GGML_ASSERT(d_Qy != nullptr);
+    }
+    if (!ids_uma) {
+        d_ids = ids_buf_ctx->dev_buffer;
+        ids_buf_offset = vk_tensor_offset(ids) + ids->view_offs;
+        GGML_ASSERT(d_ids != nullptr);
+    }
+    if (qx_needs_dequant) {
+        d_X = ctx->prealloc_x;
+        GGML_ASSERT(d_X->size >= x_sz * ne02 * ne03);
+    } else {
+        d_X = d_Qx;
+        x_buf_offset = qx_buf_offset;
+        GGML_ASSERT(qx_sz == x_sz);
+    }
+    if (qy_needs_dequant) {
+        d_Y = ctx->prealloc_y;
+        GGML_ASSERT(d_Y->size >= y_sz * ne02 * ne03);
+    } else {
+        d_Y = d_Qy;
+        y_buf_offset = qy_buf_offset;
+        GGML_ASSERT(qy_sz == y_sz);
+    }
+
+    if (x_non_contig) {
+        ggml_vk_cpy_to_contiguous(ctx, subctx, to_fp16_vk_0, src0, { d_Qx, qx_buf_offset, VK_WHOLE_SIZE }, { d_X, 0, VK_WHOLE_SIZE });
+    } else if (qx_needs_dequant) {
+        const std::vector<uint32_t> pc = { (uint32_t)ne01, (uint32_t)ne10, (uint32_t)ne10, (uint32_t)ne10, (uint32_t)(ggml_nelements(src0)) };
+        ggml_vk_sync_buffers(subctx);
+        ggml_vk_dispatch_pipeline(ctx, subctx, to_fp16_vk_0,
+            { vk_subbuffer{ d_Qx, qx_buf_offset, qx_sz * ne02 * ne03 }, vk_subbuffer{ d_X, 0, x_sz * ne02 * ne03 } }, pc.size() * sizeof(uint32_t), pc.data(), { (uint32_t)(x_ne * ne02 * ne03), 1, 1});
+    }
+    if (y_non_contig) {
+        ggml_vk_cpy_to_contiguous(ctx, subctx, to_fp16_vk_1, src1, { d_Qy, qy_buf_offset, VK_WHOLE_SIZE }, { d_Y, 0, VK_WHOLE_SIZE });
+    }
+
+    uint32_t stride_batch_x = ne00*ne01;
+    uint32_t stride_batch_y = ne10*ne11;
+
+    if (!ggml_vk_dim01_contiguous(src0) && !qx_needs_dequant) {
+        stride_batch_x = src0->nb[0] / ggml_type_size(src0->type);
+    }
+
+    if (!ggml_vk_dim01_contiguous(src1) && !qy_needs_dequant) {
+        stride_batch_y = src1->nb[0] / ggml_type_size(src1->type);
+    }
+
+    // compute
+    ggml_vk_matmul_id(
+        ctx, subctx, pipeline,
+        { d_X, x_buf_offset, x_sz * ne02 * ne03 }, { d_Y, y_buf_offset, y_sz * ne12 * ne13 },
+        { d_D, d_buf_offset, d_sz * ne22 * ne23 }, { d_ids, ids_buf_offset, ids_sz },
+        ne01, ne21, ne10, ne10, ne10, ne01,
+        stride_batch_x, stride_batch_y, ne20*ne21,
+        n_as, nei0, nei1, nbi1 / ggml_type_size(ids->type), ne11
+    );  // NOLINT
+}
+
+static void ggml_vk_mul_mat_vec_id_q_f16(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, const ggml_tensor * src1, const ggml_tensor * ids, ggml_tensor * dst, bool dryrun = false) {
+    VK_LOG_DEBUG("ggml_vk_mul_mat_vec_id_q_f16((" << src0 << ", name=" << src0->name << ", type=" << src0->type << ", ne0=" << src0->ne[0] << ", ne1=" << src0->ne[1] << ", ne2=" << src0->ne[2] << ", ne3=" << src0->ne[3] << ", nb0=" << src0->nb[0] << ", nb1=" << src0->nb[1] << ", nb2=" << src0->nb[2] << ", nb3=" << src0->nb[3];
+    std::cerr << "), (" << src1 << ", name=" << src1->name << ", type=" << src1->type << ", ne0=" << src1->ne[0] << ", ne1=" << src1->ne[1] << ", ne2=" << src1->ne[2] << ", ne3=" << src1->ne[3] << ", nb0=" << src1->nb[0] << ", nb1=" << src1->nb[1] << ", nb2=" << src1->nb[2] << ", nb3=" << src1->nb[3];
+    std::cerr << "), (" << ids << ", name=" << ids->name << ", type=" << ids->type << ", ne0=" << ids->ne[0] << ", ne1=" << ids->ne[1] << ", ne2=" << ids->ne[2] << ", ne3=" << ids->ne[3] << ", nb0=" << ids->nb[0] << ", nb1=" << ids->nb[1] << ", nb2=" << ids->nb[2] << ", nb3=" << ids->nb[3];
+    std::cerr << "), (" << dst << ", name=" << dst->name << ", type=" << dst->type << ", ne0=" << dst->ne[0] << ", ne1=" << dst->ne[1] << ", ne2=" << dst->ne[2] << ", ne3=" << dst->ne[3] << ", nb0=" << dst->nb[0] << ", nb1=" << dst->nb[1] << ", nb2=" << dst->nb[2] << ", nb3=" << dst->nb[3];
+    std::cerr << "), " << (dryrun ? "dryrun" : "") << ")");
+    GGML_ASSERT(ggml_vk_dim01_contiguous(src0) || src0->type == GGML_TYPE_F32 || src0->type == GGML_TYPE_F16);  // NOLINT
+    GGML_ASSERT(ggml_vk_dim01_contiguous(src1) || src1->type == GGML_TYPE_F32 || src1->type == GGML_TYPE_F16);  // NOLINT
+    GGML_ASSERT(ids->type == GGML_TYPE_I32);
+
+    const uint64_t ne00 = src0->ne[0];
+    const uint64_t ne01 = src0->ne[1];
+    const uint64_t ne02 = src0->ne[2];
+    const uint64_t ne03 = src0->ne[3];
+
+    const uint64_t ne10 = src1->ne[0];
+    const uint64_t ne11 = src1->ne[1];
+    const uint64_t ne12 = src1->ne[2];
+    const uint64_t ne13 = src1->ne[3];
+
+    const uint64_t nei0 = ids->ne[0];
+    const uint64_t nei1 = ids->ne[1];
+
+    const uint64_t nbi2 = ids->nb[2];
+
+    GGML_ASSERT(nei1 == 1);
+
+    const uint64_t ne20 = dst->ne[0];
+    const uint64_t ne21 = dst->ne[1];
+    const uint64_t ne22 = dst->ne[2];
+    const uint64_t ne23 = dst->ne[3];
+
+    ggml_backend_vk_buffer_context * dst_buf_ctx = (ggml_backend_vk_buffer_context *)dst->buffer->context;
+    ggml_backend_vk_buffer_context * src0_buf_ctx = (ggml_backend_vk_buffer_context *)src0->buffer->context;
+    ggml_backend_vk_buffer_context * src1_buf_ctx = (ggml_backend_vk_buffer_context *)src1->buffer->context;
+    ggml_backend_vk_buffer_context * ids_buf_ctx = (ggml_backend_vk_buffer_context *)ids->buffer->context;
+
+    vk_buffer d_Qx;
+    size_t qx_buf_offset = 0;
+    vk_buffer d_Qy;
+    size_t qy_buf_offset = 0;
+    vk_buffer d_ids;
+    size_t ids_buf_offset = 0;
+
+    bool src0_uma = false;
+    bool src1_uma = false;
+    bool ids_uma = false;
+
+    if (ctx->device->uma) {
+        ggml_vk_host_get(ctx->device, src0->data, d_Qx, qx_buf_offset);
+        ggml_vk_host_get(ctx->device, src1->data, d_Qy, qy_buf_offset);
+        ggml_vk_host_get(ctx->device, ids->data, d_ids, ids_buf_offset);
+        src0_uma = d_Qx != nullptr;
+        src1_uma = d_Qy != nullptr;
+        ids_uma = d_ids != nullptr;
+    }
+
+    const bool x_non_contig = !ggml_vk_dim01_contiguous(src0);
+    const bool y_non_contig = !ggml_vk_dim01_contiguous(src1);
+
+    const bool f16_f32_kernel = src1->type == GGML_TYPE_F32;
+
+    const bool qx_needs_dequant = x_non_contig;
+    const bool qy_needs_dequant = (src1->type != GGML_TYPE_F16 && !f16_f32_kernel) || y_non_contig;
+
+    // Not implemented
+    GGML_ASSERT(y_non_contig || !qy_needs_dequant);  // NOLINT
+
+    const uint64_t x_ne = ne01 * ne00;
+    const uint64_t y_ne = ne11 * ne10;
+    const uint64_t d_ne = ne21 * ne20;
+
+    const uint64_t qx_sz = ggml_vk_align_size(ggml_type_size(src0->type) * x_ne / ggml_blck_size(src0->type), ctx->device->properties.limits.minStorageBufferOffsetAlignment);
+    const uint64_t qy_sz = ggml_type_size(src1->type) * y_ne / ggml_blck_size(src1->type);
+    const uint64_t x_sz = x_non_contig ? ggml_vk_align_size(ggml_type_size(src0->type) * x_ne, ctx->device->properties.limits.minStorageBufferOffsetAlignment) : qx_sz;
+    const uint64_t y_sz = f16_f32_kernel ? sizeof(float) * y_ne : sizeof(ggml_fp16_t) * y_ne;
+    const uint64_t ids_sz = nbi2;
+    const uint64_t d_sz = sizeof(float) * d_ne;
+
+    vk_pipeline to_fp16_vk_0 = nullptr;
+    vk_pipeline to_fp16_vk_1 = nullptr;
+    if (x_non_contig) {
+        to_fp16_vk_0 = ggml_vk_get_cpy_pipeline(ctx, src0->type, src0->type);
+    }
+    if (y_non_contig) {
+        to_fp16_vk_1 = ggml_vk_get_cpy_pipeline(ctx, src1->type, src1->type);
+    } else {
+        to_fp16_vk_1 = ggml_vk_get_to_fp16(ctx, src1->type);
+    }
+    vk_pipeline dmmv = ggml_vk_get_dequantize_mul_mat_vec_id(ctx, src0->type, src1->type);
+    GGML_ASSERT(!qx_needs_dequant || to_fp16_vk_0 != nullptr);  // NOLINT
+    GGML_ASSERT(!qy_needs_dequant || to_fp16_vk_1 != nullptr);  // NOLINT
+    GGML_ASSERT(dmmv != nullptr);
+
+    if (dryrun) {
+        const uint64_t x_sz_upd = x_sz * ne02 * ne03;
+        const uint64_t y_sz_upd = y_sz * ne12 * ne13;
+        if (
+                (qx_needs_dequant && x_sz_upd > ctx->device->max_memory_allocation_size) ||
+                (qy_needs_dequant && y_sz_upd > ctx->device->max_memory_allocation_size)) {
+            GGML_ABORT("Requested preallocation size is too large");
+        }
+        if (qx_needs_dequant && ctx->prealloc_size_x < x_sz_upd) {
+            ctx->prealloc_size_x = x_sz_upd;
+        }
+        if (qy_needs_dequant && ctx->prealloc_size_y < y_sz_upd) {
+            ctx->prealloc_size_y = y_sz_upd;
+        }
+
+        // Request descriptor sets
+        if (qx_needs_dequant) {
+            ggml_pipeline_request_descriptor_sets(ctx->device, to_fp16_vk_0, 1);
+        }
+        if (qy_needs_dequant) {
+            ggml_pipeline_request_descriptor_sets(ctx->device, to_fp16_vk_1, 1);
+        }
+        ggml_pipeline_request_descriptor_sets(ctx->device, dmmv, 1);
+        return;
+    }
+
+    vk_buffer d_D = dst_buf_ctx->dev_buffer;
+    const uint64_t d_buf_offset = vk_tensor_offset(dst) + dst->view_offs;
+    GGML_ASSERT(d_D != nullptr);
+    vk_buffer d_X;
+    uint64_t x_buf_offset = 0;
+    vk_buffer d_Y;
+    uint64_t y_buf_offset = 0;
+    if(!src0_uma) {
+        d_Qx = src0_buf_ctx->dev_buffer;
+        qx_buf_offset = vk_tensor_offset(src0) + src0->view_offs;
+        GGML_ASSERT(d_Qx != nullptr);
+    }
+    if(!src1_uma) {
+        d_Qy = src1_buf_ctx->dev_buffer;
+        qy_buf_offset = vk_tensor_offset(src1) + src1->view_offs;
+        GGML_ASSERT(d_Qy != nullptr);
+    }
+    if(!ids_uma) {
+        d_ids = ids_buf_ctx->dev_buffer;
+        ids_buf_offset = vk_tensor_offset(ids) + ids->view_offs;
+        GGML_ASSERT(d_ids != nullptr);
+    }
+    if (qx_needs_dequant) {
+        d_X = ctx->prealloc_x;
+    } else {
+        d_X = d_Qx;
+        x_buf_offset = qx_buf_offset;
+        GGML_ASSERT(qx_sz == x_sz);
+    }
+    if (qy_needs_dequant) {
+        d_Y = ctx->prealloc_y;
+    } else {
+        d_Y = d_Qy;
+        y_buf_offset = qy_buf_offset;
+        GGML_ASSERT(qy_sz == y_sz);
+    }
+
+    if (x_non_contig) {
+        GGML_ASSERT(x_sz == ggml_vk_align_size(ggml_type_size(src0->type) * x_ne, ctx->device->properties.limits.minStorageBufferOffsetAlignment));
+        ggml_vk_cpy_to_contiguous(ctx, subctx, to_fp16_vk_0, src0, { d_Qx, qx_buf_offset, VK_WHOLE_SIZE }, { d_X, 0, VK_WHOLE_SIZE });
+    }
+    if (y_non_contig) {
+        GGML_ASSERT(y_sz == ggml_type_size(src1->type) * y_ne);
+        ggml_vk_cpy_to_contiguous(ctx, subctx, to_fp16_vk_1, src1, { d_Qy, qy_buf_offset, VK_WHOLE_SIZE }, { d_Y, 0, VK_WHOLE_SIZE });
+    }
+
+    uint32_t stride_batch_y = ne10*ne11;
+
+    if (!ggml_vk_dim01_contiguous(src1) && !qy_needs_dequant) {
+        stride_batch_y = src1->nb[0] / ggml_type_size(src1->type);
+    }
+
+    const uint32_t max_groups_x = ctx->device->properties.limits.maxComputeWorkGroupCount[0];
+
+    uint32_t groups_x = ne01;
+    uint32_t groups_z = 1;
+
+    if (ne01 > max_groups_x) {
+        groups_z = 64;
+        groups_x /= groups_z;
+    }
+
+    // compute
+    const vk_mat_vec_id_push_constants pc = {
+        (uint32_t)ne00, (uint32_t)ne10, (uint32_t)ne10, (uint32_t)ne01,
+        (uint32_t)x_ne, stride_batch_y, (uint32_t)(ne20*ne21),
+        (uint32_t)nei0, (uint32_t)ne11,
+    };
+    ggml_vk_sync_buffers(subctx);
+    ggml_vk_dispatch_pipeline(ctx, subctx, dmmv,
+        { vk_subbuffer{ d_X, x_buf_offset, x_sz * ne02 * ne03 },
+        vk_subbuffer{ d_Y, y_buf_offset, y_sz * ne12 * ne13 }, vk_subbuffer{ d_D, d_buf_offset, d_sz * ne22 * ne23}, vk_subbuffer{ d_ids, ids_buf_offset, ids_sz } },
+        sizeof(vk_mat_vec_id_push_constants), &pc, { groups_x, (uint32_t)nei0, groups_z });
+}
+
+static void ggml_vk_mul_mat_id(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, const ggml_tensor * src1, const ggml_tensor * src2, ggml_tensor * dst, bool dryrun = false) {
+    VK_LOG_DEBUG("ggml_vk_mul_mat_id(" << src0 << ", " << src1 << ", " << src2 << ", " << dst << ")");
+    if (src2->ne[1] == 1 && (src0->type == GGML_TYPE_F32 || src0->type == GGML_TYPE_F16 || ggml_is_quantized(src0->type))) {
+        ggml_vk_mul_mat_vec_id_q_f16(ctx, subctx, src0, src1, src2, dst, dryrun);
+    } else {
+        ggml_vk_mul_mat_id_q_f16(ctx, subctx, src0, src1, src2, dst, dryrun);
+    }
+}
+
+static vk_pipeline ggml_vk_op_get_pipeline(ggml_backend_vk_context * ctx, const ggml_tensor * src0, const ggml_tensor * src1, const ggml_tensor * src2, ggml_tensor * dst, ggml_op op) {
+    switch (op) {
+    case GGML_OP_GET_ROWS:
+        GGML_ASSERT(src1->type == GGML_TYPE_I32);
+        if (dst->type == GGML_TYPE_F16) {
+            return ctx->device->pipeline_get_rows[src0->type];
+        }
+        if (dst->type == GGML_TYPE_F32) {
+            return ctx->device->pipeline_get_rows_f32[src0->type];
+        }
+        return nullptr;
+    case GGML_OP_ACC:
+        if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32) {
+            return ctx->device->pipeline_acc_f32;
+        }
+        return nullptr;
+    case GGML_OP_ADD:
+        if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32) {
+            return ctx->device->pipeline_add_f32;
+        }
+        if (src0->type == GGML_TYPE_F16 && src1->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F16) {
+            return ctx->device->pipeline_add_f16_f32_f16;
+        }
+        return nullptr;
+    case GGML_OP_MUL:
+        if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32) {
+            return ctx->device->pipeline_mul_f32;
+        }
+        return nullptr;
+    case GGML_OP_DIV:
+        if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32) {
+            return ctx->device->pipeline_div_f32;
+        }
+        return nullptr;
+    case GGML_OP_CONCAT:
+        if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32) {
+            return ctx->device->pipeline_concat_f32;
+        }
+        if (src0->type == GGML_TYPE_F16 && src1->type == GGML_TYPE_F16 && dst->type == GGML_TYPE_F16) {
+            return ctx->device->pipeline_concat_f16;
+        }
+        if (src0->type == GGML_TYPE_I32 && src1->type == GGML_TYPE_I32 && dst->type == GGML_TYPE_I32) {
+            return ctx->device->pipeline_concat_i32;
+        }
+        return nullptr;
+    case GGML_OP_UPSCALE:
+        if (src0->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32) {
+            return ctx->device->pipeline_upscale_f32;
+        }
+        return nullptr;
+    case GGML_OP_SCALE:
+        if (src0->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32) {
+            return ctx->device->pipeline_scale_f32;
+        }
+        return nullptr;
+    case GGML_OP_SQR:
+        if (src0->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32) {
+            return ctx->device->pipeline_sqr_f32;
+        }
+        return nullptr;
+    case GGML_OP_SIN:
+        if (src0->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32) {
+            return ctx->device->pipeline_sin_f32;
+        }
+        return nullptr;
+    case GGML_OP_COS:
+        if (src0->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32) {
+            return ctx->device->pipeline_cos_f32;
+        }
+        return nullptr;
+    case GGML_OP_CLAMP:
+        if (src0->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32) {
+            return ctx->device->pipeline_clamp_f32;
+        }
+        return nullptr;
+    case GGML_OP_PAD:
+        if (src0->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32) {
+            return ctx->device->pipeline_pad_f32;
+        }
+        return nullptr;
+    case GGML_OP_REPEAT:
+        if (ggml_type_size(src0->type) == sizeof(float) && ggml_type_size(dst->type) == sizeof(float)) {
+            return ctx->device->pipeline_repeat_f32;
+        }
+        return nullptr;
+    case GGML_OP_CPY:
+    case GGML_OP_CONT:
+    case GGML_OP_DUP:
+        return ggml_vk_get_cpy_pipeline(ctx, src0->type, dst->type);
+    case GGML_OP_NORM:
+        if (src0->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32) {
+            return ctx->device->pipeline_norm_f32;
+        }
+        return nullptr;
+    case GGML_OP_GROUP_NORM:
+        if (src0->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32) {
+            return ctx->device->pipeline_group_norm_f32;
+        }
+        return nullptr;
+    case GGML_OP_RMS_NORM:
+        if (src0->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32) {
+            return ctx->device->pipeline_rms_norm_f32;
+        }
+        return nullptr;
+    case GGML_OP_UNARY:
+        switch (ggml_get_unary_op(dst)) {
+            case GGML_UNARY_OP_SILU:
+                if (src0->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32) {
+                    return ctx->device->pipeline_silu_f32;
+                }
+                break;
+            case GGML_UNARY_OP_GELU:
+                if (src0->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32) {
+                    return ctx->device->pipeline_gelu_f32;
+                }
+                break;
+            case GGML_UNARY_OP_GELU_QUICK:
+                if (src0->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32) {
+                    return ctx->device->pipeline_gelu_quick_f32;
+                }
+                break;
+            case GGML_UNARY_OP_RELU:
+                if (src0->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32) {
+                    return ctx->device->pipeline_relu_f32;
+                }
+                break;
+            case GGML_UNARY_OP_TANH:
+                if (src0->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32) {
+                    return ctx->device->pipeline_tanh_f32;
+                }
+                break;
+            default:
+                break;
+        }
+        return nullptr;
+    case GGML_OP_DIAG_MASK_INF:
+        if (src0->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32) {
+            return ctx->device->pipeline_diag_mask_inf_f32;
+        }
+        return nullptr;
+    case GGML_OP_SOFT_MAX:
+        GGML_ASSERT(!src1 || src1->type == GGML_TYPE_F32 || src1->type == GGML_TYPE_F16);
+
+        if (src0->type == GGML_TYPE_F32 && (src1 == nullptr || src1->type == GGML_TYPE_F32) && dst->type == GGML_TYPE_F32) {
+            return ctx->device->pipeline_soft_max_f32;
+        }
+        if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_F16 && dst->type == GGML_TYPE_F32) {
+            return ctx->device->pipeline_soft_max_f32_f16;
+        }
+        return nullptr;
+    case GGML_OP_ROPE:
+        {
+            const int mode = ((const int32_t *) dst->op_params)[2];
+            const bool is_neox = mode & GGML_ROPE_TYPE_NEOX;
+
+            if (is_neox) {
+                if (src0->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32) {
+                    return ctx->device->pipeline_rope_neox_f32;
+                }
+                if (src0->type == GGML_TYPE_F16 && dst->type == GGML_TYPE_F16) {
+                    return ctx->device->pipeline_rope_neox_f16;
+                }
+            } else {
+                if (src0->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32) {
+                    return ctx->device->pipeline_rope_norm_f32;
+                }
+                if (src0->type == GGML_TYPE_F16 && dst->type == GGML_TYPE_F16) {
+                    return ctx->device->pipeline_rope_norm_f16;
+                }
+            }
+            return nullptr;
+        }
+    case GGML_OP_ARGSORT:
+        if (src0->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_I32) {
+            return ctx->device->pipeline_argsort_f32;
+        }
+        return nullptr;
+    case GGML_OP_SUM_ROWS:
+        if (src0->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32) {
+            return ctx->device->pipeline_sum_rows_f32;
+        }
+        return nullptr;
+    case GGML_OP_IM2COL:
+        if (src1->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32) {
+            return ctx->device->pipeline_im2col_f32;
+        }
+        if (src1->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F16) {
+            return ctx->device->pipeline_im2col_f32_f16;
+        }
+        return nullptr;
+    case GGML_OP_TIMESTEP_EMBEDDING:
+        if (src0->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32) {
+            return ctx->device->pipeline_timestep_embedding_f32;
+        }
+        return nullptr;
+    case GGML_OP_LEAKY_RELU:
+        if (src0->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32) {
+            return ctx->device->pipeline_leaky_relu_f32;
+        }
+        return nullptr;
+    default:
+        return nullptr;
+    }
+
+    GGML_UNUSED(src2);
+}
+
+static bool ggml_vk_op_supports_incontiguous(ggml_op op) {
+    switch (op) {
+    case GGML_OP_CPY:
+    case GGML_OP_GET_ROWS:
+    case GGML_OP_ADD:
+    case GGML_OP_MUL:
+    case GGML_OP_DIV:
+    case GGML_OP_CONCAT:
+    case GGML_OP_UPSCALE:
+    case GGML_OP_SCALE:
+    case GGML_OP_SQR:
+    case GGML_OP_SIN:
+    case GGML_OP_COS:
+    case GGML_OP_CLAMP:
+    case GGML_OP_PAD:
+    case GGML_OP_REPEAT:
+        return true;
+    default:
+        return false;
+    }
+}
+
+template<typename PC>
+static void ggml_vk_op_f32(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, const ggml_tensor * src1, const ggml_tensor * src2, ggml_tensor * dst, ggml_op op, const PC&& pc, bool dryrun = false) {
+    VK_LOG_DEBUG("ggml_vk_op_f32((" << src0 << ", name=" << src0->name << ", type=" << src0->type << ", ne0=" << src0->ne[0] << ", ne1=" << src0->ne[1] << ", ne2=" << src0->ne[2] << ", ne3=" << src0->ne[3] << ", nb0=" << src0->nb[0] << ", nb1=" << src0->nb[1] << ", nb2=" << src0->nb[2] << ", nb3=" << src0->nb[3];
+    if (src1 != nullptr) {
+        std::cerr << "), (" << src1 << ", name=" << src1->name << ", type=" << src1->type << ", ne0=" << src1->ne[0] << ", ne1=" << src1->ne[1] << ", ne2=" << src1->ne[2] << ", ne3=" << src1->ne[3] << ", nb0=" << src1->nb[0] << ", nb1=" << src1->nb[1] << ", nb2=" << src1->nb[2] << ", nb3=" << src1->nb[3];
+    }
+    if (src2 != nullptr) {
+        std::cerr << "), (" << src2 << ", name=" << src2->name << ", type=" << src2->type << ", ne0=" << src2->ne[0] << ", ne1=" << src2->ne[1] << ", ne2=" << src2->ne[2] << ", ne3=" << src2->ne[3] << ", nb0=" << src2->nb[0] << ", nb1=" << src2->nb[1] << ", nb2=" << src2->nb[2] << ", nb3=" << src2->nb[3];
+    }
+    std::cerr << "), (" << dst << ", name=" << dst->name << ", type=" << dst->type << ", ne0=" << dst->ne[0] << ", ne1=" << dst->ne[1] << ", ne2=" << dst->ne[2] << ", ne3=" << dst->ne[3] << ", nb0=" << dst->nb[0] << ", nb1=" << dst->nb[1] << ", nb2=" << dst->nb[2] << ", nb3=" << dst->nb[3];
+    std::cerr << "), " << ggml_op_name(op) << ", " << (dryrun ? "dryrun" : "") << ")");
+    GGML_ASSERT(op == GGML_OP_GET_ROWS || (!ggml_is_quantized(src0->type) && (src1 == nullptr || !ggml_is_quantized(src1->type))));  // NOLINT
+    GGML_ASSERT(ggml_vk_op_supports_incontiguous(op) || ggml_vk_dim01_contiguous(src0));  // NOLINT
+    GGML_ASSERT(dst->buffer != nullptr);
+    const uint64_t ne00 = src0->ne[0];
+    const uint64_t ne01 = src0->ne[1];
+    const uint64_t ne02 = src0->ne[2];
+    const uint64_t ne03 = src0->ne[3];
+    const uint64_t ne0 = ne00 * ne01;
+
+    const bool use_src1 = src1 != nullptr;
+    const uint64_t ne10 = use_src1 ? src1->ne[0] : 0;
+    const uint64_t ne11 = use_src1 ? src1->ne[1] : 0;
+    const uint64_t ne12 = use_src1 ? src1->ne[2] : 0;
+    const uint64_t ne13 = use_src1 ? src1->ne[3] : 0;
+    const uint64_t ne1 = ne10 * ne11;
+    // const uint64_t nb10 = use_src1 ? src1->nb[0] : 0;
+
+    const bool use_src2 = src2 != nullptr;
+    const uint64_t ne20 = use_src2 ? src2->ne[0] : 0;
+    const uint64_t ne21 = use_src2 ? src2->ne[1] : 0;
+    const uint64_t ne22 = use_src2 ? src2->ne[2] : 0;
+    const uint64_t ne23 = use_src2 ? src2->ne[3] : 0;
+    const uint64_t ne2 = ne20 * ne21;
+
+    const uint64_t ned0 = dst->ne[0];
+    const uint64_t ned1 = dst->ne[1];
+    const uint64_t ned2 = dst->ne[2];
+    const uint64_t ned3 = dst->ne[3];
+    const uint64_t ned = ned0 * ned1;
+
+    vk_pipeline pipeline = ggml_vk_op_get_pipeline(ctx, src0, src1, src2, dst, op);
+
+    if (pipeline == nullptr) {
+        std::cerr << "ggml_vulkan: Error: Missing op: " << ggml_op_name(op) << " for " << ggml_type_name(src0->type);
+        if (src1 != nullptr) {
+            std::cerr << " and " << ggml_type_name(src1->type);
+        }
+        std::cerr << " to " << ggml_type_name(dst->type) << std::endl;
+        GGML_ABORT("fatal error");
+    }
+
+    if (dryrun) {
+        ggml_pipeline_request_descriptor_sets(ctx->device, pipeline, 1);
+        return;
+    }
+
+    const bool op_supports_incontiguous = ggml_vk_op_supports_incontiguous(op);
+
+    ggml_backend_vk_buffer_context * dst_buf_ctx = (ggml_backend_vk_buffer_context *)dst->buffer->context;
+    ggml_backend_vk_buffer_context * src0_buf_ctx = (ggml_backend_vk_buffer_context *)src0->buffer->context;
+    ggml_backend_vk_buffer_context * src1_buf_ctx = use_src1 ? (ggml_backend_vk_buffer_context *)src1->buffer->context : nullptr;
+    ggml_backend_vk_buffer_context * src2_buf_ctx = use_src2 ? (ggml_backend_vk_buffer_context *)src2->buffer->context : nullptr;
+
+    vk_buffer d_X = nullptr;
+    size_t x_buf_offset = 0;
+    vk_buffer d_Y = nullptr;
+    size_t y_buf_offset = 0;
+    vk_buffer d_Z = nullptr;
+    size_t z_buf_offset = 0;
+
+    bool src0_uma = false;
+    bool src1_uma = false;
+    bool src2_uma = false;
+
+    if (ctx->device->uma) {
+        ggml_vk_host_get(ctx->device, src0->data, d_X, x_buf_offset);
+        src0_uma = d_X != nullptr;
+        if (use_src1) {
+            ggml_vk_host_get(ctx->device, src1->data, d_Y, y_buf_offset);
+            src1_uma = d_Y != nullptr;
+        }
+        if (use_src2) {
+            ggml_vk_host_get(ctx->device, src2->data, d_Z, z_buf_offset);
+            src2_uma = d_Z != nullptr;
+        }
+    }
+
+    uint64_t x_sz = ggml_type_size(src0->type)/ggml_blck_size(src0->type) * ne0;
+    uint64_t y_sz = use_src1 ? ggml_type_size(src1->type) * ne1 : 0;
+    uint64_t z_sz = use_src2 ? ggml_type_size(src2->type) * ne2 : 0;
+    uint64_t d_sz = ggml_type_size(dst->type) * ned;
+
+    vk_buffer d_D = dst_buf_ctx->dev_buffer;
+
+    // Workaround for tiny tensor inputs on ROPE
+    if (op == GGML_OP_ROPE && use_src1 && y_sz > d_D->size) {
+        y_sz = VK_WHOLE_SIZE;
+    }
+
+    GGML_ASSERT(d_D != nullptr);
+    uint64_t d_buf_offset = ((vk_tensor_offset(dst) + dst->view_offs) / ctx->device->properties.limits.minStorageBufferOffsetAlignment) * ctx->device->properties.limits.minStorageBufferOffsetAlignment;
+    GGML_ASSERT(d_buf_offset == vk_tensor_offset(dst) || op == GGML_OP_CPY);  // NOLINT
+    if(!src0_uma) {
+        d_X = src0_buf_ctx->dev_buffer;
+        x_buf_offset = vk_tensor_offset(src0) + src0->view_offs;
+        GGML_ASSERT(d_X != nullptr);
+    }
+    if (use_src1 && !src1_uma) {
+        d_Y = src1_buf_ctx->dev_buffer;
+        y_buf_offset = vk_tensor_offset(src1) + src1->view_offs;
+        GGML_ASSERT(d_Y != nullptr);
+    }
+    if (use_src2 && !src2_uma) {
+        d_Z = src2_buf_ctx->dev_buffer;
+        z_buf_offset = vk_tensor_offset(src2) + src2->view_offs;
+        GGML_ASSERT(d_Z != nullptr);
+    }
+
+    if (op_supports_incontiguous) {
+        x_sz = ggml_nbytes(src0);
+        y_sz = use_src1 ? ggml_nbytes(src1) : 0;
+        z_sz = use_src2 ? ggml_nbytes(src2) : 0;
+        d_sz = ggml_nbytes(dst);
+
+        if (x_buf_offset + x_sz >= d_X->size) {
+            x_sz = VK_WHOLE_SIZE;
+        }
+        if (use_src1 && y_buf_offset + y_sz >= d_Y->size) {
+            y_sz = VK_WHOLE_SIZE;
+        }
+        if (use_src2 && z_buf_offset + z_sz >= d_Z->size) {
+            z_sz = VK_WHOLE_SIZE;
+        }
+        if (d_buf_offset + d_sz >= d_D->size) {
+            d_sz = VK_WHOLE_SIZE;
+        }
+    }
+
+    std::array<uint32_t, 3> elements;
+
+    // Single call if dimension 2 is contiguous
+    GGML_ASSERT(op_supports_incontiguous || (ggml_is_contiguous(src0) && (src1 == nullptr || ggml_is_contiguous(src1))));
+
+    switch (op) {
+    case GGML_OP_NORM:
+    case GGML_OP_RMS_NORM:
+    case GGML_OP_SOFT_MAX:
+    case GGML_OP_SUM_ROWS:
+        {
+            const uint32_t nr = ggml_nrows(src0);
+            if (nr > 262144) {
+                elements = { 512, 512, CEIL_DIV(nr, 262144) };
+            } else if (nr > 512) {
+                elements = { 512, CEIL_DIV(nr, 512), 1 };
+            } else {
+                elements = { nr, 1, 1 };
+            }
+        } break;
+    case GGML_OP_GROUP_NORM:
+        {
+            const uint32_t num_groups = dst->op_params[0];
+            elements = { num_groups * (uint32_t)src0->ne[3], 1, 1 };
+        } break;
+    case GGML_OP_DIAG_MASK_INF:
+    case GGML_OP_ROPE:
+        elements = { (uint32_t)ggml_nrows(src0), (uint32_t)ne00, 1 };
+        break;
+    case GGML_OP_GET_ROWS:
+        elements = { (uint32_t)ne00, (uint32_t)ne10, (uint32_t)(ne11 * ne12) };
+        break;
+    case GGML_OP_ARGSORT:
+        elements = { (uint32_t)ne00, (uint32_t)ggml_nrows(src0), 1 };
+        break;
+    case GGML_OP_IM2COL:
+        {
+            const bool is_2D = dst->op_params[6] == 1;
+
+            const uint32_t IC = src1->ne[is_2D ? 2 : 1];
+
+            const uint32_t KH = is_2D ? src0->ne[1] : 1;
+            const uint32_t KW =         src0->ne[0];
+
+            const uint32_t OH = is_2D ? dst->ne[2] : 1;
+            const uint32_t OW =         dst->ne[1];
+
+            const uint32_t batch = src1->ne[3];
+
+            elements = { OW * KW * KH, OH, batch * IC };
+        } break;
+    case GGML_OP_TIMESTEP_EMBEDDING:
+        {
+            const uint32_t dim = dst->op_params[0];
+            uint32_t half_ceil = (dim + 1) / 2;
+            elements = { half_ceil, (uint32_t)src0->ne[0], 1 };
+        } break;
+    case GGML_OP_ADD:
+    case GGML_OP_DIV:
+    case GGML_OP_MUL:
+    case GGML_OP_SCALE:
+    case GGML_OP_SQR:
+    case GGML_OP_SIN:
+    case GGML_OP_COS:
+    case GGML_OP_CLAMP:
+    case GGML_OP_PAD:
+    case GGML_OP_REPEAT:
+    case GGML_OP_CPY:
+    case GGML_OP_CONCAT:
+    case GGML_OP_UPSCALE:
+    case GGML_OP_UNARY:
+        {
+            const uint32_t ne = ggml_nelements(dst);
+            if (ne > 262144) {
+                elements = { 512, 512, CEIL_DIV(ne, 262144) };
+            } else if (ne > 512) {
+                elements = { 512, CEIL_DIV(ne, 512), 1 };
+            } else {
+                elements = { ne, 1, 1 };
+            }
+        } break;
+    default:
+        elements = { (uint32_t)ggml_nelements(src0), 1, 1 };
+        break;
+    }
+
+    if (!op_supports_incontiguous) {
+        if (x_sz != VK_WHOLE_SIZE) {
+            x_sz *= ne02 * ne03;
+        }
+        if (use_src1 && y_sz != VK_WHOLE_SIZE) {
+            y_sz *= ne12 * ne13;
+        }
+        if (use_src2 && z_sz != VK_WHOLE_SIZE) {
+            z_sz *= ne22 * ne23;
+        }
+        if (d_sz != VK_WHOLE_SIZE) {
+            d_sz *= ned2 * ned3;
+        }
+    }
+
+    if (op == GGML_OP_SOFT_MAX) {
+        // Empty src1 is possible in soft_max, but the shader needs a buffer
+        vk_subbuffer subbuf_y;
+        if (use_src1) {
+            subbuf_y = { d_Y, y_buf_offset, y_sz };
+        } else {
+            subbuf_y = { d_X, 0, x_sz };
+        }
+
+        ggml_vk_sync_buffers(subctx);
+        ggml_vk_dispatch_pipeline(ctx, subctx, pipeline, { vk_subbuffer{ d_X, x_buf_offset, x_sz }, subbuf_y, vk_subbuffer{ d_D, d_buf_offset, d_sz } }, sizeof(PC), &pc, elements);
+    } else if (op == GGML_OP_ROPE) {
+        // Empty src2 is possible in rope, but the shader needs a buffer
+        vk_subbuffer subbuf_z;
+        if (use_src2) {
+            subbuf_z = { d_Z, z_buf_offset, z_sz };
+        } else {
+            subbuf_z = { d_X, 0, x_sz };
+        }
+
+        ggml_vk_sync_buffers(subctx);
+        ggml_vk_dispatch_pipeline(ctx, subctx, pipeline, { vk_subbuffer{ d_X, x_buf_offset, x_sz }, vk_subbuffer{ d_Y, y_buf_offset, y_sz }, subbuf_z, vk_subbuffer{ d_D, d_buf_offset, d_sz } }, sizeof(PC), &pc, elements);
+    } else if (op == GGML_OP_IM2COL) {
+        // im2col uses only src1 and dst buffers
+        ggml_vk_sync_buffers(subctx);
+        ggml_vk_dispatch_pipeline(ctx, subctx, pipeline, { vk_subbuffer{ d_Y, y_buf_offset, y_sz }, vk_subbuffer{ d_D, d_buf_offset, d_sz } }, sizeof(PC), &pc, elements);
+    } else if (use_src2) {
+        ggml_vk_sync_buffers(subctx);
+        ggml_vk_dispatch_pipeline(ctx, subctx, pipeline, { vk_subbuffer{ d_X, x_buf_offset, x_sz }, vk_subbuffer{ d_Y, y_buf_offset, y_sz }, vk_subbuffer{ d_Z, z_buf_offset, z_sz }, vk_subbuffer{ d_D, d_buf_offset, d_sz } }, sizeof(PC), &pc, elements);
+    } else if (use_src1) {
+        ggml_vk_sync_buffers(subctx);
+        ggml_vk_dispatch_pipeline(ctx, subctx, pipeline, { vk_subbuffer{ d_X, x_buf_offset, x_sz }, vk_subbuffer{ d_Y, y_buf_offset, y_sz }, vk_subbuffer{ d_D, d_buf_offset, d_sz } }, sizeof(PC), &pc, elements);
+    } else {
+        ggml_vk_sync_buffers(subctx);
+        ggml_vk_dispatch_pipeline(ctx, subctx, pipeline, { vk_subbuffer{ d_X, x_buf_offset, x_sz }, vk_subbuffer{ d_D, d_buf_offset, d_sz } }, sizeof(PC), &pc, elements);
+    }
+}
+
+static void ggml_vk_get_rows(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, bool dryrun = false) {
+    const uint32_t src0_type_size = ggml_type_size(src0->type);
+    const uint32_t src1_type_size = ggml_type_size(src1->type);
+    const uint32_t dst_type_size = ggml_type_size(dst->type);
+
+    ggml_vk_op_f32<vk_op_binary_push_constants>(ctx, subctx, src0, src1, nullptr, dst, GGML_OP_GET_ROWS, {
+        (uint32_t)ggml_nelements(src0),
+        (uint32_t)src0->ne[0], (uint32_t)src0->ne[1], (uint32_t)src0->ne[2],(uint32_t)src0->ne[3], (uint32_t)src0->nb[0] / src0_type_size, (uint32_t)src0->nb[1] / src0_type_size, (uint32_t)src0->nb[2] / src0_type_size, (uint32_t)src0->nb[3] / src0_type_size,
+        (uint32_t)src1->ne[0], (uint32_t)src1->ne[1], (uint32_t)src1->ne[2],(uint32_t)src1->ne[3], (uint32_t)src1->nb[0] / src1_type_size, (uint32_t)src1->nb[1] / src1_type_size, (uint32_t)src1->nb[2] / src1_type_size, (uint32_t)src1->nb[3] / src1_type_size,
+        (uint32_t) dst->ne[0], (uint32_t) dst->ne[1], (uint32_t) dst->ne[2],(uint32_t) dst->ne[3], (uint32_t) dst->nb[0] /  dst_type_size, (uint32_t) dst->nb[1] /  dst_type_size, (uint32_t) dst->nb[2] /  dst_type_size, (uint32_t) dst->nb[3] /  dst_type_size,
+        0,
+        0.0f, 0.0f, 0,
+    }, dryrun);
+}
+
+static void ggml_vk_acc(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, bool dryrun = false) {
+    const uint32_t src0_type_size = ggml_type_size(src0->type);
+    const uint32_t src1_type_size = ggml_type_size(src1->type);
+    const uint32_t dst_type_size = ggml_type_size(dst->type);
+    const uint32_t d_offset = ((vk_tensor_offset(dst) + dst->view_offs) % ctx->device->properties.limits.minStorageBufferOffsetAlignment) / dst_type_size;
+
+    int nb1 = dst->op_params[0] / 4; // 4 bytes of float32
+    int nb2 = dst->op_params[1] / 4; // 4 bytes of float32
+    // int nb3 = dst->op_params[2] / 4; // 4 bytes of float32 - unused
+    int offset = dst->op_params[3] / 4; // offset in bytes
+
+    ggml_vk_op_f32<vk_op_binary_push_constants>(ctx, subctx, src0, src1, nullptr, dst, GGML_OP_ACC, {
+        (uint32_t)ggml_nelements(src0),
+        (uint32_t)src0->ne[0], (uint32_t)src0->ne[1], (uint32_t)src0->ne[2],(uint32_t)src0->ne[3], (uint32_t)src0->nb[0] / src0_type_size, (uint32_t)nb1, (uint32_t)nb2, (uint32_t)src0->nb[3] / src0_type_size,
+        (uint32_t)src1->ne[0], (uint32_t)src1->ne[1], (uint32_t)src1->ne[2],(uint32_t)src1->ne[3], (uint32_t)src1->nb[0] / src1_type_size, (uint32_t)src1->nb[1] / src1_type_size, (uint32_t)src1->nb[2] / src1_type_size, (uint32_t)src1->nb[3] / src1_type_size,
+        (uint32_t) dst->ne[0], (uint32_t) dst->ne[1], (uint32_t) dst->ne[2],(uint32_t) dst->ne[3], (uint32_t) dst->nb[0] /  dst_type_size, (uint32_t)nb1, (uint32_t)nb2, (uint32_t) dst->nb[3] /  dst_type_size,
+        d_offset,
+        0.0f, 0.0f, offset,
+    }, dryrun);
+}
+
+static void ggml_vk_add(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, bool dryrun = false) {
+    const uint32_t src0_type_size = ggml_type_size(src0->type);
+    const uint32_t src1_type_size = ggml_type_size(src1->type);
+    const uint32_t dst_type_size = ggml_type_size(dst->type);
+
+    ggml_vk_op_f32<vk_op_binary_push_constants>(ctx, subctx, src0, src1, nullptr, dst, GGML_OP_ADD, {
+        (uint32_t)ggml_nelements(src0),
+        (uint32_t)src0->ne[0], (uint32_t)src0->ne[1], (uint32_t)src0->ne[2],(uint32_t)src0->ne[3], (uint32_t)src0->nb[0] / src0_type_size, (uint32_t)src0->nb[1] / src0_type_size, (uint32_t)src0->nb[2] / src0_type_size, (uint32_t)src0->nb[3] / src0_type_size,
+        (uint32_t)src1->ne[0], (uint32_t)src1->ne[1], (uint32_t)src1->ne[2],(uint32_t)src1->ne[3], (uint32_t)src1->nb[0] / src1_type_size, (uint32_t)src1->nb[1] / src1_type_size, (uint32_t)src1->nb[2] / src1_type_size, (uint32_t)src1->nb[3] / src1_type_size,
+        (uint32_t) dst->ne[0], (uint32_t) dst->ne[1], (uint32_t) dst->ne[2],(uint32_t) dst->ne[3], (uint32_t) dst->nb[0] /  dst_type_size, (uint32_t) dst->nb[1] /  dst_type_size, (uint32_t) dst->nb[2] /  dst_type_size, (uint32_t) dst->nb[3] /  dst_type_size,
+        0,
+        0.0f, 0.0f, 0,
+    }, dryrun);
+}
+
+static void ggml_vk_mul(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, bool dryrun = false) {
+    const uint32_t src0_type_size = ggml_type_size(src0->type);
+    const uint32_t src1_type_size = ggml_type_size(src1->type);
+    const uint32_t dst_type_size = ggml_type_size(dst->type);
+
+    ggml_vk_op_f32<vk_op_binary_push_constants>(ctx, subctx, src0, src1, nullptr, dst, GGML_OP_MUL, {
+        (uint32_t)ggml_nelements(src0),
+        (uint32_t)src0->ne[0], (uint32_t)src0->ne[1], (uint32_t)src0->ne[2],(uint32_t)src0->ne[3], (uint32_t)src0->nb[0] / src0_type_size, (uint32_t)src0->nb[1] / src0_type_size, (uint32_t)src0->nb[2] / src0_type_size, (uint32_t)src0->nb[3] / src0_type_size,
+        (uint32_t)src1->ne[0], (uint32_t)src1->ne[1], (uint32_t)src1->ne[2],(uint32_t)src1->ne[3], (uint32_t)src1->nb[0] / src1_type_size, (uint32_t)src1->nb[1] / src1_type_size, (uint32_t)src1->nb[2] / src1_type_size, (uint32_t)src1->nb[3] / src1_type_size,
+        (uint32_t) dst->ne[0], (uint32_t) dst->ne[1], (uint32_t) dst->ne[2],(uint32_t) dst->ne[3], (uint32_t) dst->nb[0] /  dst_type_size, (uint32_t) dst->nb[1] /  dst_type_size, (uint32_t) dst->nb[2] /  dst_type_size, (uint32_t) dst->nb[3] /  dst_type_size,
+        0,
+        0.0f, 0.0f, 0,
+    }, dryrun);
+}
+
+static void ggml_vk_div(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, bool dryrun = false) {
+    const uint32_t src0_type_size = ggml_type_size(src0->type);
+    const uint32_t src1_type_size = ggml_type_size(src1->type);
+    const uint32_t dst_type_size = ggml_type_size(dst->type);
+
+    ggml_vk_op_f32<vk_op_binary_push_constants>(ctx, subctx, src0, src1, nullptr, dst, GGML_OP_DIV, {
+        (uint32_t)ggml_nelements(src0),
+        (uint32_t)src0->ne[0], (uint32_t)src0->ne[1], (uint32_t)src0->ne[2],(uint32_t)src0->ne[3], (uint32_t)src0->nb[0] / src0_type_size, (uint32_t)src0->nb[1] / src0_type_size, (uint32_t)src0->nb[2] / src0_type_size, (uint32_t)src0->nb[3] / src0_type_size,
+        (uint32_t)src1->ne[0], (uint32_t)src1->ne[1], (uint32_t)src1->ne[2],(uint32_t)src1->ne[3], (uint32_t)src1->nb[0] / src1_type_size, (uint32_t)src1->nb[1] / src1_type_size, (uint32_t)src1->nb[2] / src1_type_size, (uint32_t)src1->nb[3] / src1_type_size,
+        (uint32_t) dst->ne[0], (uint32_t) dst->ne[1], (uint32_t) dst->ne[2],(uint32_t) dst->ne[3], (uint32_t) dst->nb[0] /  dst_type_size, (uint32_t) dst->nb[1] /  dst_type_size, (uint32_t) dst->nb[2] /  dst_type_size, (uint32_t) dst->nb[3] /  dst_type_size,
+        0,
+        0.0f, 0.0f, 0,
+    }, dryrun);
+}
+
+static void ggml_vk_concat(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, bool dryrun = false) {
+    int * op_params = (int *)dst->op_params;
+
+    const uint32_t src0_type_size = ggml_type_size(src0->type);
+    const uint32_t src1_type_size = ggml_type_size(src1->type);
+    const uint32_t dst_type_size = ggml_type_size(dst->type);
+
+    ggml_vk_op_f32<vk_op_binary_push_constants>(ctx, subctx, src0, src1, nullptr, dst, GGML_OP_CONCAT, {
+        (uint32_t)ggml_nelements(dst),
+        (uint32_t)src0->ne[0], (uint32_t)src0->ne[1], (uint32_t)src0->ne[2],(uint32_t)src0->ne[3], (uint32_t)src0->nb[0] / src0_type_size, (uint32_t)src0->nb[1] / src0_type_size, (uint32_t)src0->nb[2] / src0_type_size, (uint32_t)src0->nb[3] / src0_type_size,
+        (uint32_t)src1->ne[0], (uint32_t)src1->ne[1], (uint32_t)src1->ne[2],(uint32_t)src1->ne[3], (uint32_t)src1->nb[0] / src1_type_size, (uint32_t)src1->nb[1] / src1_type_size, (uint32_t)src1->nb[2] / src1_type_size, (uint32_t)src1->nb[3] / src1_type_size,
+        (uint32_t) dst->ne[0], (uint32_t) dst->ne[1], (uint32_t) dst->ne[2],(uint32_t) dst->ne[3], (uint32_t) dst->nb[0] /  dst_type_size, (uint32_t) dst->nb[1] /  dst_type_size, (uint32_t) dst->nb[2] /  dst_type_size, (uint32_t) dst->nb[3] /  dst_type_size,
+        0,
+        0.0f, 0.0f, op_params[0],
+    }, dryrun);
+}
+
+static void ggml_vk_upscale(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, ggml_tensor * dst, bool dryrun = false) {
+    const uint32_t src0_type_size = ggml_type_size(src0->type);
+
+    const float sf0 = (float)dst->ne[0] / src0->ne[0];
+    const float sf1 = (float)dst->ne[1] / src0->ne[1];
+    const float sf2 = (float)dst->ne[2] / src0->ne[2];
+    const float sf3 = (float)dst->ne[3] / src0->ne[3];
+
+    ggml_vk_op_f32<vk_op_upscale_push_constants>(ctx, subctx, src0, nullptr, nullptr, dst, GGML_OP_UPSCALE, {
+        (uint32_t)ggml_nelements(dst), 0,
+        (uint32_t)src0->nb[0] / src0_type_size, (uint32_t)src0->nb[1] / src0_type_size, (uint32_t)src0->nb[2] / src0_type_size, (uint32_t)src0->nb[3] / src0_type_size,
+        (uint32_t)dst->ne[0], (uint32_t)dst->ne[1], (uint32_t)dst->ne[2],(uint32_t)dst->ne[3],
+        sf0, sf1, sf2, sf3,
+    }, dryrun);
+}
+
+static void ggml_vk_scale(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, ggml_tensor * dst, bool dryrun = false) {
+    float * op_params = (float *)dst->op_params;
+    const uint32_t src0_type_size = ggml_type_size(src0->type);
+    const uint32_t dst_type_size = ggml_type_size(dst->type);
+
+    ggml_vk_op_f32<vk_op_unary_push_constants>(ctx, subctx, src0, nullptr, nullptr, dst, GGML_OP_SCALE, {
+        (uint32_t)ggml_nelements(src0),
+        (uint32_t)src0->ne[0], (uint32_t)src0->ne[1], (uint32_t)src0->ne[2], (uint32_t)src0->ne[3], (uint32_t)src0->nb[0] / src0_type_size, (uint32_t)src0->nb[1] / src0_type_size, (uint32_t)src0->nb[2] / src0_type_size, (uint32_t)src0->nb[3] / src0_type_size,
+        (uint32_t) dst->ne[0], (uint32_t) dst->ne[1], (uint32_t) dst->ne[2], (uint32_t) dst->ne[3], (uint32_t) dst->nb[0] /  dst_type_size, (uint32_t) dst->nb[1] /  dst_type_size, (uint32_t) dst->nb[2] /  dst_type_size, (uint32_t) dst->nb[3] /  dst_type_size,
+        0,
+        op_params[0], 0.0f
+    }, dryrun);
+}
+
+static void ggml_vk_sqr(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, ggml_tensor * dst, bool dryrun = false) {
+    const uint32_t src0_type_size = ggml_type_size(src0->type);
+    const uint32_t dst_type_size = ggml_type_size(dst->type);
+
+    ggml_vk_op_f32<vk_op_unary_push_constants>(ctx, subctx, src0, nullptr, nullptr, dst, GGML_OP_SQR, {
+        (uint32_t)ggml_nelements(src0),
+        (uint32_t)src0->ne[0], (uint32_t)src0->ne[1], (uint32_t)src0->ne[2], (uint32_t)src0->ne[3], (uint32_t)src0->nb[0] / src0_type_size, (uint32_t)src0->nb[1] / src0_type_size, (uint32_t)src0->nb[2] / src0_type_size, (uint32_t)src0->nb[3] / src0_type_size,
+        (uint32_t) dst->ne[0], (uint32_t) dst->ne[1], (uint32_t) dst->ne[2], (uint32_t) dst->ne[3], (uint32_t) dst->nb[0] /  dst_type_size, (uint32_t) dst->nb[1] /  dst_type_size, (uint32_t) dst->nb[2] /  dst_type_size, (uint32_t) dst->nb[3] /  dst_type_size,
+        0,
+        0.0f, 0.0f,
+    }, dryrun);
+}
+
+static void ggml_vk_sin(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, ggml_tensor * dst, bool dryrun = false) {
+    const uint32_t src0_type_size = ggml_type_size(src0->type);
+    const uint32_t dst_type_size = ggml_type_size(dst->type);
+
+    ggml_vk_op_f32<vk_op_unary_push_constants>(ctx, subctx, src0, nullptr, nullptr, dst, GGML_OP_SIN, {
+        (uint32_t)ggml_nelements(src0),
+        (uint32_t)src0->ne[0], (uint32_t)src0->ne[1], (uint32_t)src0->ne[2], (uint32_t)src0->ne[3], (uint32_t)src0->nb[0] / src0_type_size, (uint32_t)src0->nb[1] / src0_type_size, (uint32_t)src0->nb[2] / src0_type_size, (uint32_t)src0->nb[3] / src0_type_size,
+        (uint32_t) dst->ne[0], (uint32_t) dst->ne[1], (uint32_t) dst->ne[2], (uint32_t) dst->ne[3], (uint32_t) dst->nb[0] /  dst_type_size, (uint32_t) dst->nb[1] /  dst_type_size, (uint32_t) dst->nb[2] /  dst_type_size, (uint32_t) dst->nb[3] /  dst_type_size,
+        0,
+        0.0f, 0.0f,
+    }, dryrun);
+}
+
+static void ggml_vk_cos(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, ggml_tensor * dst, bool dryrun = false) {
+    const uint32_t src0_type_size = ggml_type_size(src0->type);
+    const uint32_t dst_type_size = ggml_type_size(dst->type);
+
+    ggml_vk_op_f32<vk_op_unary_push_constants>(ctx, subctx, src0, nullptr, nullptr, dst, GGML_OP_COS, {
+        (uint32_t)ggml_nelements(src0),
+        (uint32_t)src0->ne[0], (uint32_t)src0->ne[1], (uint32_t)src0->ne[2], (uint32_t)src0->ne[3], (uint32_t)src0->nb[0] / src0_type_size, (uint32_t)src0->nb[1] / src0_type_size, (uint32_t)src0->nb[2] / src0_type_size, (uint32_t)src0->nb[3] / src0_type_size,
+        (uint32_t) dst->ne[0], (uint32_t) dst->ne[1], (uint32_t) dst->ne[2], (uint32_t) dst->ne[3], (uint32_t) dst->nb[0] /  dst_type_size, (uint32_t) dst->nb[1] /  dst_type_size, (uint32_t) dst->nb[2] /  dst_type_size, (uint32_t) dst->nb[3] /  dst_type_size,
+        0,
+        0.0f, 0.0f,
+    }, dryrun);
+}
+
+static void ggml_vk_clamp(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, ggml_tensor * dst, bool dryrun = false) {
+    float * op_params = (float *)dst->op_params;
+    const uint32_t src0_type_size = ggml_type_size(src0->type);
+    const uint32_t dst_type_size = ggml_type_size(dst->type);
+
+    ggml_vk_op_f32<vk_op_unary_push_constants>(ctx, subctx, src0, nullptr, nullptr, dst, GGML_OP_CLAMP, {
+        (uint32_t)ggml_nelements(src0),
+        (uint32_t)src0->ne[0], (uint32_t)src0->ne[1], (uint32_t)src0->ne[2], (uint32_t)src0->ne[3], (uint32_t)src0->nb[0] / src0_type_size, (uint32_t)src0->nb[1] / src0_type_size, (uint32_t)src0->nb[2] / src0_type_size, (uint32_t)src0->nb[3] / src0_type_size,
+        (uint32_t) dst->ne[0], (uint32_t) dst->ne[1], (uint32_t) dst->ne[2], (uint32_t) dst->ne[3], (uint32_t) dst->nb[0] /  dst_type_size, (uint32_t) dst->nb[1] /  dst_type_size, (uint32_t) dst->nb[2] /  dst_type_size, (uint32_t) dst->nb[3] /  dst_type_size,
+        0,
+        op_params[0], op_params[1],
+    }, dryrun);
+}
+
+static void ggml_vk_pad(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, ggml_tensor * dst, bool dryrun = false) {
+    const uint32_t src0_type_size = ggml_type_size(src0->type);
+    const uint32_t dst_type_size = ggml_type_size(dst->type);
+
+    ggml_vk_op_f32<vk_op_unary_push_constants>(ctx, subctx, src0, nullptr, nullptr, dst, GGML_OP_PAD, {
+        (uint32_t)ggml_nelements(dst),
+        (uint32_t)src0->ne[0], (uint32_t)src0->ne[1], (uint32_t)src0->ne[2], (uint32_t)src0->ne[3], (uint32_t)src0->nb[0] / src0_type_size, (uint32_t)src0->nb[1] / src0_type_size, (uint32_t)src0->nb[2] / src0_type_size, (uint32_t)src0->nb[3] / src0_type_size,
+        (uint32_t) dst->ne[0], (uint32_t) dst->ne[1], (uint32_t) dst->ne[2], (uint32_t) dst->ne[3], (uint32_t) dst->nb[0] /  dst_type_size, (uint32_t) dst->nb[1] /  dst_type_size, (uint32_t) dst->nb[2] /  dst_type_size, (uint32_t) dst->nb[3] /  dst_type_size,
+        0,
+        0.0f, 0.0f,
+    }, dryrun);
+}
+
+static void ggml_vk_repeat(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, ggml_tensor * dst, bool dryrun = false) {
+    const uint32_t src0_type_size = ggml_type_size(src0->type);
+    const uint32_t dst_type_size = ggml_type_size(dst->type);
+
+    ggml_vk_op_f32<vk_op_unary_push_constants>(ctx, subctx, src0, nullptr, nullptr, dst, GGML_OP_REPEAT, {
+        (uint32_t)ggml_nelements(dst),
+        (uint32_t)src0->ne[0], (uint32_t)src0->ne[1], (uint32_t)src0->ne[2], (uint32_t)src0->ne[3], (uint32_t)src0->nb[0] / src0_type_size, (uint32_t)src0->nb[1] / src0_type_size, (uint32_t)src0->nb[2] / src0_type_size, (uint32_t)src0->nb[3] / src0_type_size,
+        (uint32_t) dst->ne[0], (uint32_t) dst->ne[1], (uint32_t) dst->ne[2], (uint32_t) dst->ne[3], (uint32_t) dst->nb[0] /  dst_type_size, (uint32_t) dst->nb[1] /  dst_type_size, (uint32_t) dst->nb[2] /  dst_type_size, (uint32_t) dst->nb[3] /  dst_type_size,
+        0,
+        0.0f, 0.0f,
+    }, dryrun);
+}
+
+static void ggml_vk_cpy(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, ggml_tensor * dst, bool dryrun = false) {
+    const uint32_t src0_type_size = ggml_type_size(src0->type);
+    const uint32_t dst_type_size = ggml_type_size(dst->type);
+    const uint32_t d_offset = ((vk_tensor_offset(dst) + dst->view_offs) % ctx->device->properties.limits.minStorageBufferOffsetAlignment) / dst_type_size;
+
+    ggml_vk_op_f32<vk_op_unary_push_constants>(ctx, subctx, src0, nullptr, nullptr, dst, GGML_OP_CPY, {
+        (uint32_t)ggml_nelements(src0),
+        (uint32_t)src0->ne[0], (uint32_t)src0->ne[1], (uint32_t)src0->ne[2], (uint32_t)src0->ne[3], (uint32_t)src0->nb[0] / src0_type_size, (uint32_t)src0->nb[1] / src0_type_size, (uint32_t)src0->nb[2] / src0_type_size, (uint32_t)src0->nb[3] / src0_type_size,
+        (uint32_t) dst->ne[0], (uint32_t) dst->ne[1], (uint32_t) dst->ne[2], (uint32_t) dst->ne[3], (uint32_t) dst->nb[0] /  dst_type_size, (uint32_t) dst->nb[1] /  dst_type_size, (uint32_t) dst->nb[2] /  dst_type_size, (uint32_t) dst->nb[3] /  dst_type_size,
+        d_offset,
+        0.0f, 0.0f,
+    }, dryrun);
+}
+
+static void ggml_vk_norm(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, ggml_tensor * dst, bool dryrun = false) {
+    float * op_params = (float *)dst->op_params;
+
+    ggml_vk_op_f32<vk_op_push_constants>(ctx, subctx, src0, nullptr, nullptr, dst, GGML_OP_NORM, { (uint32_t)src0->ne[0], (uint32_t)src0->ne[1], op_params[0], 0.0f }, dryrun);
+}
+
+static void ggml_vk_group_norm(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, ggml_tensor * dst, bool dryrun = false) {
+    const int * int_op_params = (const int *)dst->op_params;
+    const float * float_op_params = (const float *)dst->op_params;
+
+    const uint32_t num_groups = int_op_params[0];
+    const float eps = float_op_params[1];
+    const uint32_t group_size = src0->ne[0] * src0->ne[1] * ((src0->ne[2] + num_groups - 1) / num_groups);
+
+    ggml_vk_op_f32<vk_op_push_constants>(ctx, subctx, src0, nullptr, nullptr, dst, GGML_OP_GROUP_NORM, { group_size, 0, eps, 0.0f }, dryrun);
+}
+
+static void ggml_vk_rms_norm(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, ggml_tensor * dst, bool dryrun = false) {
+    float * op_params = (float *)dst->op_params;
+    ggml_vk_op_f32<vk_op_push_constants>(ctx, subctx, src0, nullptr, nullptr, dst, GGML_OP_RMS_NORM, { (uint32_t)src0->ne[0], (uint32_t)src0->ne[1], op_params[0], 0.0f }, dryrun);
+}
+
+static void ggml_vk_unary(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, ggml_tensor * dst, bool dryrun = false) {
+    ggml_vk_op_f32<vk_op_push_constants>(ctx, subctx, src0, nullptr, nullptr, dst, GGML_OP_UNARY, { (uint32_t)ggml_nelements(src0), 0, 0.0f, 0.0f }, dryrun);
+}
+
+static void ggml_vk_diag_mask_inf(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, ggml_tensor * dst, bool dryrun = false) {
+    int32_t * op_params = (int32_t *)dst->op_params;
+    ggml_vk_op_f32<vk_op_diag_mask_push_constants>(ctx, subctx, src0, nullptr, nullptr, dst, GGML_OP_DIAG_MASK_INF, { (uint32_t)src0->ne[0], (uint32_t)src0->ne[1], op_params[0] }, dryrun);
+}
+
+static void ggml_vk_soft_max(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, bool dryrun = false) {
+    float * op_params = (float *)dst->op_params;
+
+    float scale = op_params[0];
+    float max_bias = op_params[1];
+
+    const uint32_t ncols =   (uint32_t)src0->ne[0];
+    const uint32_t nrows_x = (uint32_t)ggml_nrows(src0);
+    const uint32_t nrows_y = (uint32_t)src0->ne[1];
+
+    const uint32_t n_head_kv   = nrows_x/nrows_y;
+    const uint32_t n_head_log2 = 1u << (uint32_t) floorf(log2f((float) n_head_kv));
+
+    const float m0 = powf(2.0f, -(max_bias       ) / n_head_log2);
+    const float m1 = powf(2.0f, -(max_bias / 2.0f) / n_head_log2);
+
+    ggml_vk_op_f32<vk_op_soft_max_push_constants>(ctx, subctx, src0, src1, nullptr, dst, GGML_OP_SOFT_MAX, {
+        ncols,
+        src1 != nullptr ? nrows_y : (uint32_t)0,
+        scale, max_bias,
+        m0, m1,
+        n_head_log2,
+    }, dryrun);
+}
+
+static void ggml_vk_rope(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, const ggml_tensor * src1, const ggml_tensor * src2, ggml_tensor * dst, bool dryrun = false) {
+    const int n_dims        = ((int32_t *) dst->op_params)[1];
+    // const int mode          = ((int32_t *) dst->op_params)[2];
+    // const int n_ctx         = ((int32_t *) dst->op_params)[3];
+    const int n_ctx_orig    = ((int32_t *) dst->op_params)[4];
+    const float freq_base   = ((float *)   dst->op_params)[5];
+    const float freq_scale  = ((float *)   dst->op_params)[6];
+    const float ext_factor  = ((float *)   dst->op_params)[7];
+    const float attn_factor = ((float *)   dst->op_params)[8];
+    const float beta_fast   = ((float *)   dst->op_params)[9];
+    const float beta_slow   = ((float *)   dst->op_params)[10];
+
+    float corr_dims[2];
+    ggml_rope_yarn_corr_dims(n_dims, n_ctx_orig, freq_base, beta_fast, beta_slow, corr_dims);
+
+    const float theta_scale = powf(freq_base, -2.0f/n_dims);
+
+    ggml_vk_op_f32<vk_op_rope_push_constants>(ctx, subctx, src0, src1, src2, dst, GGML_OP_ROPE, {
+        (uint32_t)src0->ne[0], (uint32_t)n_dims, freq_scale, (uint32_t)src0->ne[1],
+        freq_base, ext_factor, attn_factor, {corr_dims[0], corr_dims[1]}, theta_scale,
+        src2 != nullptr,
+    }, dryrun);
+}
+
+static void ggml_vk_argsort(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, ggml_tensor * dst, bool dryrun = false) {
+    int32_t * op_params = (int32_t *)dst->op_params;
+
+    uint32_t ncols = src0->ne[0];
+
+    uint32_t ncols_pad = 1;
+    while (ncols_pad < ncols) {
+        ncols_pad *= 2;
+    }
+
+    GGML_ASSERT(ncols_pad <= 1024);
+
+    ggml_vk_op_f32<vk_op_argsort_push_constants>(ctx, subctx, src0, nullptr, nullptr, dst, GGML_OP_ARGSORT, {
+        ncols,
+        ncols_pad,
+        op_params[0],
+    }, dryrun);
+}
+
+static void ggml_vk_sum_rows(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, ggml_tensor * dst, bool dryrun = false) {
+    ggml_vk_op_f32<vk_op_push_constants>(ctx, subctx, src0, nullptr, nullptr, dst, GGML_OP_SUM_ROWS, { (uint32_t)src0->ne[0], 0, 0.0f, 0.0f }, dryrun);
+}
+
+static void ggml_vk_im2col(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, bool dryrun = false) {
+    const int32_t s0 = dst->op_params[0];
+    const int32_t s1 = dst->op_params[1];
+    const int32_t p0 = dst->op_params[2];
+    const int32_t p1 = dst->op_params[3];
+    const int32_t d0 = dst->op_params[4];
+    const int32_t d1 = dst->op_params[5];
+
+    const bool is_2D = dst->op_params[6] == 1;
+
+    const uint32_t IC = src1->ne[is_2D ? 2 : 1];
+    const uint32_t IH = is_2D ? src1->ne[1] : 1;
+    const uint32_t IW =         src1->ne[0];
+
+    const uint32_t KH = is_2D ? src0->ne[1] : 1;
+    const uint32_t KW =         src0->ne[0];
+
+    const uint32_t OH = is_2D ? dst->ne[2] : 1;
+    const uint32_t OW =         dst->ne[1];
+
+    const uint32_t offset_delta = src1->nb[is_2D ? 2 : 1] / 4; // nb is byte offset, src is type float32
+    const uint32_t batch_offset = src1->nb[3] / 4; // nb is byte offset, src is type float32
+
+    const uint32_t pelements = OW * KW * KH;
+
+    ggml_vk_op_f32<vk_op_im2col_push_constants>(ctx, subctx, src0, src1, nullptr, dst, GGML_OP_IM2COL, {
+        batch_offset, offset_delta,
+        IC, IW, IH, OW, OH, KW, KH,
+        pelements,
+        IC * KH * KW,
+        s0, s1, p0, p1, d0, d1,
+    }, dryrun);
+}
+
+static void ggml_vk_timestep_embedding(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, ggml_tensor * dst, bool dryrun = false) {
+    const uint32_t dim = dst->op_params[0];
+    const uint32_t max_period = dst->op_params[1];
+    const uint32_t nb1 = dst->nb[1] / ggml_type_size(dst->type);
+
+    ggml_vk_op_f32<vk_op_timestep_embedding_push_constants>(ctx, subctx, src0, nullptr, nullptr, dst, GGML_OP_TIMESTEP_EMBEDDING, {
+        nb1, dim, max_period,
+    }, dryrun);
+}
+
+static void ggml_vk_leaky_relu(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, ggml_tensor * dst, bool dryrun = false) {
+    const float * op_params = (const float *)dst->op_params;
+    ggml_vk_op_f32<vk_op_push_constants>(ctx, subctx, src0, nullptr, nullptr, dst, GGML_OP_LEAKY_RELU, { (uint32_t)ggml_nelements(src0), 0, op_params[0], 0.0f }, dryrun);
+}
+
+#ifdef GGML_VULKAN_RUN_TESTS
+static void ggml_vk_print_matrix_area(const void * data, ggml_type type, int ne0, int ne1, int i0, int i1, int i2) {
+    if (type != GGML_TYPE_F32 && type != GGML_TYPE_F16) {
+        return;
+    }
+    i0 = std::max(i0, 5);
+    i1 = std::max(i1, 5);
+    i2 = std::max(i2, 0);
+    fprintf(stderr, "         ");
+    for (int idx1 = i1 - 5; idx1 < i1 + 5; idx1++) {
+        fprintf(stderr, "%7d ", idx1);
+    }
+    fprintf(stderr, "\n");
+    for (int idx0 = i0 - 5; idx0 < i0 + 5; idx0++) {
+        fprintf(stderr, "%7d: ", idx0);
+        for (int idx1 = i1 - 5; idx1 < i1 + 5; idx1++) {
+            if (idx0 >= 0 && idx0 < ne0 && idx1 >= 0 && idx1 < ne1) {
+                float val;
+                if (type == GGML_TYPE_F32) {
+                    val = *((const float *) data + i2*ne1*ne0 + idx1*ne0 + idx0);
+                } else if (type == GGML_TYPE_F16) {
+                    val = ggml_fp16_to_fp32(*((const ggml_fp16_t *) data + i2*ne1*ne0 + idx1*ne0 + idx0));
+                } else {
+                    GGML_ABORT("fatal error");
+                }
+                fprintf(stderr, "% 7.2f ", val);
+            } else {
+                fprintf(stderr, "        ");
+            }
+        }
+        fprintf(stderr, "\n");
+    }
+}
+
+template <typename X_TYPE, typename Y_TYPE>
+static void ggml_vk_test_matmul(ggml_backend_vk_context * ctx, size_t m, size_t n, size_t k, size_t batch, size_t num_it, int split_k, int shader_size) {
+    VK_LOG_DEBUG("ggml_vk_test_matmul(" << m << ", " << n << ", " << k << ", " << batch << ", " << num_it << ", " << split_k << ", " << shader_size << ")");
+    const size_t x_ne = m * k * batch;
+    const size_t y_ne = k * n * batch;
+    const size_t d_ne = m * n * batch;
+
+    vk_pipeline p;
+    std::string shname;
+    if (shader_size == 0) {
+        if (std::is_same<float, X_TYPE>() && std::is_same<float, Y_TYPE>()) {
+            p = ctx->device->pipeline_matmul_f32->a_s;
+            shname = "F32_ALIGNED_S";
+        } else if (std::is_same<float, X_TYPE>() && std::is_same<ggml_fp16_t, Y_TYPE>()) {
+            p = ctx->device->pipeline_matmul_f32_f16->a_s;
+            shname = "F32_F16_ALIGNED_S";
+        } else if (std::is_same<ggml_fp16_t, X_TYPE>() && std::is_same<float, Y_TYPE>()) {
+            p = ctx->device->pipeline_matmul_f16_f32->a_s;
+            shname = "F16_F32_ALIGNED_S";
+        } else if (std::is_same<ggml_fp16_t, X_TYPE>() && std::is_same<ggml_fp16_t, Y_TYPE>()) {
+            p = ctx->device->pipeline_matmul_f16->a_s;
+            shname = "F16_ALIGNED_S";
+        } else {
+            GGML_ABORT("fatal error");
+        }
+    } else if (shader_size == 1) {
+        if (std::is_same<float, X_TYPE>() && std::is_same<float, Y_TYPE>()) {
+            p = ctx->device->pipeline_matmul_f32->a_m;
+            shname = "F32_ALIGNED_M";
+        } else if (std::is_same<float, X_TYPE>() && std::is_same<ggml_fp16_t, Y_TYPE>()) {
+            p = ctx->device->pipeline_matmul_f32_f16->a_m;
+            shname = "F32_F16_ALIGNED_M";
+        } else if (std::is_same<ggml_fp16_t, X_TYPE>() && std::is_same<float, Y_TYPE>()) {
+            p = ctx->device->pipeline_matmul_f16_f32->a_m;
+            shname = "F16_F32_ALIGNED_M";
+        } else if (std::is_same<ggml_fp16_t, X_TYPE>() && std::is_same<ggml_fp16_t, Y_TYPE>()) {
+            p = ctx->device->pipeline_matmul_f16->a_m;
+            shname = "F16_ALIGNED_M";
+        } else {
+            GGML_ABORT("fatal error");
+        }
+    } else if (shader_size == 2) {
+        if (std::is_same<float, X_TYPE>() && std::is_same<float, Y_TYPE>()) {
+            p = ctx->device->pipeline_matmul_f32->a_l;
+            shname = "F32_ALIGNED_L";
+        } else if (std::is_same<float, X_TYPE>() && std::is_same<ggml_fp16_t, Y_TYPE>()) {
+            p = ctx->device->pipeline_matmul_f32_f16->a_l;
+            shname = "F32_F16_ALIGNED_L";
+        } else if (std::is_same<ggml_fp16_t, X_TYPE>() && std::is_same<float, Y_TYPE>()) {
+            p = ctx->device->pipeline_matmul_f16_f32->a_l;
+            shname = "F16_F32_ALIGNED_L";
+        } else if (std::is_same<ggml_fp16_t, X_TYPE>() && std::is_same<ggml_fp16_t, Y_TYPE>()) {
+            p = ctx->device->pipeline_matmul_f16->a_l;
+            shname = "F16_ALIGNED_L";
+        } else {
+            GGML_ABORT("fatal error");
+        }
+    } else {
+        GGML_ASSERT(0);
+    }
+
+    const size_t kpad = ggml_vk_align_size(k, p->align);
+
+    if (k != kpad) {
+        if (shader_size == 0) {
+            if (std::is_same<float, X_TYPE>() && std::is_same<float, Y_TYPE>()) {
+                p = ctx->device->pipeline_matmul_f32->s;
+                shname = "F32_S";
+            } else if (std::is_same<float, X_TYPE>() && std::is_same<ggml_fp16_t, Y_TYPE>()) {
+                p = ctx->device->pipeline_matmul_f32_f16->s;
+                shname = "F32_F16_S";
+            } else if (std::is_same<ggml_fp16_t, X_TYPE>() && std::is_same<float, Y_TYPE>()) {
+                p = ctx->device->pipeline_matmul_f16_f32->s;
+                shname = "F16_F32_S";
+            } else if (std::is_same<ggml_fp16_t, X_TYPE>() && std::is_same<ggml_fp16_t, Y_TYPE>()) {
+                p = ctx->device->pipeline_matmul_f16->s;
+                shname = "F16_S";
+            }
+        } else if (shader_size == 1) {
+            if (std::is_same<float, X_TYPE>() && std::is_same<float, Y_TYPE>()) {
+                p = ctx->device->pipeline_matmul_f32->m;
+                shname = "F32_M";
+            } else if (std::is_same<float, X_TYPE>() && std::is_same<ggml_fp16_t, Y_TYPE>()) {
+                p = ctx->device->pipeline_matmul_f32_f16->m;
+                shname = "F32_F16_M";
+            } else if (std::is_same<ggml_fp16_t, X_TYPE>() && std::is_same<float, Y_TYPE>()) {
+                p = ctx->device->pipeline_matmul_f16_f32->m;
+                shname = "F16_F32_M";
+            } else if (std::is_same<ggml_fp16_t, X_TYPE>() && std::is_same<ggml_fp16_t, Y_TYPE>()) {
+                p = ctx->device->pipeline_matmul_f16->m;
+                shname = "F16_M";
+            }
+        } else if (shader_size == 2) {
+            if (std::is_same<float, X_TYPE>() && std::is_same<float, Y_TYPE>()) {
+                p = ctx->device->pipeline_matmul_f32->l;
+                shname = "F32_L";
+            } else if (std::is_same<float, X_TYPE>() && std::is_same<ggml_fp16_t, Y_TYPE>()) {
+                p = ctx->device->pipeline_matmul_f32_f16->l;
+                shname = "F32_F16_L";
+            } else if (std::is_same<ggml_fp16_t, X_TYPE>() && std::is_same<float, Y_TYPE>()) {
+                p = ctx->device->pipeline_matmul_f16_f32->l;
+                shname = "F16_F32_L";
+            } else if (std::is_same<ggml_fp16_t, X_TYPE>() && std::is_same<ggml_fp16_t, Y_TYPE>()) {
+                p = ctx->device->pipeline_matmul_f16->l;
+                shname = "F16_L";
+            }
+        }
+    }
+
+    ggml_pipeline_request_descriptor_sets(ctx->device, p, num_it);
+    if (split_k > 1) {
+        ggml_pipeline_request_descriptor_sets(ctx->device, ctx->device->pipeline_matmul_split_k_reduce, num_it);
+
+        if (ctx->prealloc_split_k == nullptr || ctx->prealloc_split_k->size < sizeof(float) * d_ne * split_k) {
+            // Resize buffer
+            if (ctx->prealloc_split_k != nullptr) {
+                ggml_vk_destroy_buffer(ctx->prealloc_split_k);
+            }
+            ctx->prealloc_split_k = ggml_vk_create_buffer_check(ctx->device, sizeof(float) * d_ne * split_k, vk::MemoryPropertyFlagBits::eDeviceLocal);
+        }
+    }
+
+    ggml_pipeline_allocate_descriptor_sets(ctx->device);
+
+    vk_buffer d_X = ggml_vk_create_buffer_check(ctx->device, sizeof(X_TYPE) * x_ne, vk::MemoryPropertyFlagBits::eDeviceLocal);
+    vk_buffer d_Y = ggml_vk_create_buffer_check(ctx->device, sizeof(Y_TYPE) * y_ne, vk::MemoryPropertyFlagBits::eDeviceLocal);
+    vk_buffer d_D = ggml_vk_create_buffer_check(ctx->device, sizeof(float) * d_ne, vk::MemoryPropertyFlagBits::eDeviceLocal);
+
+    X_TYPE* x = (X_TYPE *) malloc(sizeof(X_TYPE) * x_ne);
+    Y_TYPE* y = (Y_TYPE *) malloc(sizeof(Y_TYPE) * y_ne);
+    float* d = (float *) malloc(sizeof(float) * d_ne);
+
+    for (size_t i = 0; i < x_ne; i++) {
+        if (std::is_same<float, X_TYPE>()) {
+            x[i] = (rand() / (float)RAND_MAX) * 2.0f - 1.0f;
+        } else if (std::is_same<ggml_fp16_t, X_TYPE>()) {
+            x[i] = ggml_fp32_to_fp16((rand() / (float)RAND_MAX) * 2.0f - 1.0f);
+        } else {
+            GGML_ABORT("fatal error");
+        }
+    }
+    for (size_t i = 0; i < y_ne; i++) {
+        if (std::is_same<float, Y_TYPE>()) {
+            // y[i] = (rand() / (float)RAND_MAX) * 2.0f - 1.0f;
+            y[i] = (i % k == i / k) ? 1.0f : 0.0f;
+        } else if (std::is_same<ggml_fp16_t, Y_TYPE>()) {
+            // y[i] = ggml_fp32_to_fp16((rand() / (float)RAND_MAX) * 2.0f - 1.0f);
+            y[i] = ggml_fp32_to_fp16((i % k == i / k) ? 1.0f : 0.0f);
+        } else {
+            GGML_ABORT("fatal error");
+        }
+    }
+
+    ggml_vk_buffer_write(d_X, 0, x, sizeof(X_TYPE) * k * m * batch);
+    ggml_vk_buffer_write(d_Y, 0, y, sizeof(Y_TYPE) * k * n * batch);
+
+    vk_context subctx = ggml_vk_create_context(ctx, ctx->device->compute_queue);
+    for (size_t i = 0; i < num_it; i++) {
+        ggml_vk_ctx_begin(ctx->device, subctx);
+        ggml_vk_matmul(
+            ctx, subctx, p, ggml_vk_subbuffer(d_X), ggml_vk_subbuffer(d_Y), ggml_vk_subbuffer(d_D), ggml_vk_subbuffer(ctx->prealloc_split_k),
+            m, n, k,
+            k, k, m, k*m, k*n, m*n,
+            split_k, batch, batch, batch, 1, 1
+        );
+        ggml_vk_ctx_end(subctx);
+    }
+
+    auto begin = std::chrono::high_resolution_clock::now();
+    ggml_vk_submit(subctx, ctx->fence);
+    VK_CHECK(ctx->device->device.waitForFences({ ctx->fence }, true, UINT64_MAX), "ggml_vk_test_matmul waitForFences");
+    ctx->device->device.resetFences({ ctx->fence });
+
+    auto end = std::chrono::high_resolution_clock::now();
+    double time = std::chrono::duration_cast<std::chrono::microseconds>(end-begin).count() / 1000.0;
+
+    // copy dst to host
+    ggml_vk_buffer_read(d_D, 0, d, sizeof(float) * d_ne);
+
+    float * d_chk = (float *) malloc(sizeof(float) * d_ne);
+
+    ggml_init_params iparams = {
+        /*.mem_size   =*/ 1024*1024*1024,
+        /*.mem_buffer =*/ NULL,
+        /*.no_alloc   =*/ true,
+    };
+
+    ggml_context * ggml_ctx = ggml_init(iparams);
+
+    ggml_type src0_type;
+    ggml_type src1_type;
+
+    if (std::is_same<float, X_TYPE>()) {
+        src0_type = GGML_TYPE_F32;
+    } else if (std::is_same<ggml_fp16_t, X_TYPE>()) {
+        src0_type = GGML_TYPE_F16;
+    } else {
+        GGML_ABORT("fatal error");
+    }
+    if (std::is_same<float, Y_TYPE>()) {
+        src1_type = GGML_TYPE_F32;
+    } else if (std::is_same<ggml_fp16_t, Y_TYPE>()) {
+        src1_type = GGML_TYPE_F16;
+    } else {
+        GGML_ABORT("fatal error");
+    }
+
+    ggml_tensor * src0_ggml = ggml_new_tensor_3d(ggml_ctx, src0_type, k, m, batch);
+    ggml_tensor * src1_ggml = ggml_new_tensor_3d(ggml_ctx, src1_type, k, n, batch);
+    ggml_tensor * tensor_ggml = ggml_mul_mat(ggml_ctx, src0_ggml, src1_ggml);
+
+    src0_ggml->data = x;
+    src1_ggml->data = y;
+    tensor_ggml->data = d_chk;
+
+    ggml_cgraph * cgraph = ggml_new_graph(ggml_ctx);
+    ggml_build_forward_expand(cgraph, tensor_ggml);
+
+    ggml_graph_compute_with_ctx(ggml_ctx, cgraph, 1);
+
+    ggml_free(ggml_ctx);
+
+    double avg_err = 0.0;
+    int first_err_n = -1;
+    int first_err_m = -1;
+    int first_err_b = -1;
+
+    for (size_t i = 0; i < m*n*batch; i++) {
+        double err = std::fabs(d[i] - d_chk[i]);
+        avg_err += err;
+
+        if (err > 0.05f && first_err_n == -1) {
+            first_err_b = i / (m * n);
+            first_err_n = (i % (m * n)) / m;
+            first_err_m = (i % (m * n)) % m;
+        }
+    }
+
+    avg_err /= m * n;
+
+    double tflops = 2.0*m*n*k*batch*num_it / (time / 1000.0) / (1000.0*1000.0*1000.0*1000.0);
+
+    std::cerr << "TEST " << shname << " m=" << m << " n=" << n << " k=" << k << " batch=" << batch << " split_k=" << split_k << " matmul " << time / num_it << "ms " << tflops << " TFLOPS avg_err=" << avg_err << std::endl;
+
+    if (avg_err > 0.1) {
+        std::cerr << "m = " << first_err_m << " n = " << first_err_n << " b = " << first_err_b << std::endl;
+        std::cerr << "Actual result: " << std::endl << std::endl;
+        ggml_vk_print_matrix_area(d, GGML_TYPE_F32, m, n, first_err_m, first_err_n, first_err_b);
+        std::cerr << std::endl;
+        ggml_vk_print_matrix_area(d, GGML_TYPE_F32, m, n, first_err_m, first_err_n + 15, first_err_b);
+        std::cerr << "Expected result: " << std::endl << std::endl;
+        ggml_vk_print_matrix_area(d_chk, GGML_TYPE_F32, m, n, first_err_m, first_err_n, first_err_b);
+
+        if (split_k > 1) {
+            float * split_k_buf = (float *) malloc(sizeof(float) * d_ne * split_k);
+            ggml_vk_buffer_read(ctx->prealloc_split_k, 0, split_k_buf, sizeof(float) * d_ne * split_k);
+
+            std::cerr << "d_buf0: " << std::endl << std::endl;
+            ggml_vk_print_matrix_area(split_k_buf, GGML_TYPE_F32, m, n, first_err_m, first_err_n, first_err_b);
+
+            std::cerr << "d_buf1: " << std::endl << std::endl;
+            ggml_vk_print_matrix_area(split_k_buf + d_ne, GGML_TYPE_F32, m, n, first_err_m, first_err_n, first_err_b);
+
+            std::cerr << "d_buf2: " << std::endl << std::endl;
+            ggml_vk_print_matrix_area(split_k_buf + 2 * d_ne, GGML_TYPE_F32, m, n, first_err_m, first_err_n, first_err_b);
+
+            std::cerr << "d_buf3: " << std::endl << std::endl;
+            ggml_vk_print_matrix_area(split_k_buf + 3 * d_ne, GGML_TYPE_F32, m, n, first_err_m, first_err_n, first_err_b);
+
+            free(split_k_buf);
+        }
+    }
+
+    free(d_chk);
+
+    ggml_vk_queue_cleanup(ctx->device, ctx->device->transfer_queue);
+    ggml_vk_queue_cleanup(ctx->device, ctx->device->compute_queue);
+
+    ggml_vk_destroy_buffer(d_X);
+    ggml_vk_destroy_buffer(d_Y);
+    ggml_vk_destroy_buffer(d_D);
+
+    ggml_pipeline_cleanup(p);
+    ggml_pipeline_cleanup(ctx->device->pipeline_matmul_split_k_reduce);
+
+    free(x);
+    free(y);
+    free(d);
+}
+
+static void ggml_vk_print_tensor_area(const ggml_tensor * tensor, int i0, int i1, int i2, int i3) {
+    if (tensor->type != GGML_TYPE_F32 && tensor->type != GGML_TYPE_F16) {
+        return;
+    }
+    i0 = std::max(i0, 5);
+    i1 = std::max(i1, 5);
+    i2 = std::max(i2, 0);
+    i3 = std::max(i3, 0);
+    fprintf(stderr, "         ");
+    for (int idx1 = i1 - 5; idx1 < i1 + 5; idx1++) {
+        fprintf(stderr, "%7d ", idx1);
+    }
+    fprintf(stderr, "\n");
+    for (int idx0 = i0 - 5; idx0 < i0 + 5; idx0++) {
+        fprintf(stderr, "%7d: ", idx0);
+        for (int idx1 = i1 - 5; idx1 < i1 + 5; idx1++) {
+            if (idx0 >= 0 && idx0 < tensor->ne[0] && idx1 >= 0 && idx1 < tensor->ne[1] && i2 >= 0 && i2 < tensor->ne[2] && i3 >= 0 && i3 < tensor->ne[3]) {
+                float val;
+                if (tensor->type == GGML_TYPE_F32) {
+                    val = *(float *) ((char *) tensor->data + i3*tensor->nb[3] + i2*tensor->nb[2] + idx1*tensor->nb[1] + idx0*tensor->nb[0]);
+                } else if (tensor->type == GGML_TYPE_F16) {
+                    val = ggml_fp16_to_fp32(*(ggml_fp16_t *) ((char *) tensor->data + i3*tensor->nb[3] + i2*tensor->nb[2] + idx1*tensor->nb[1] + idx0*tensor->nb[0]));
+                } else {
+                    GGML_ABORT("fatal error");
+                }
+                fprintf(stderr, "% 7.2f ", val);
+            } else {
+                fprintf(stderr, "        ");
+            }
+        }
+        fprintf(stderr, "\n");
+    }
+}
+
+static void ggml_vk_quantize_data(const float * from, void * to, size_t ne, ggml_type quant) {
+    ggml_quantize_chunk(quant, from, to, 0, 1, ne, nullptr);
+}
+
+static void ggml_vk_dequantize_data(const void * from, float * to, size_t ne, ggml_type quant) {
+    if (quant == GGML_TYPE_F32) {
+        memcpy(to, from, sizeof(float) * ne);
+        return;
+    }
+
+    ggml_type_traits_t tt = ggml_internal_get_type_traits(quant);
+
+    ggml_to_float_t dequant_fn = tt.to_float;
+
+    dequant_fn(from, to, ne);
+}
+
+static void ggml_vk_test_dequant(ggml_backend_vk_context * ctx, size_t ne, ggml_type quant) {
+    VK_LOG_DEBUG("ggml_vk_test_dequant(" << ne << ")");
+    const size_t x_sz = sizeof(float) * ne;
+    const size_t x_sz_f16 = sizeof(ggml_fp16_t) * ne;
+    const size_t qx_sz = ne * ggml_type_size(quant)/ggml_blck_size(quant);
+    float * x = (float *) malloc(x_sz);
+    void * qx = malloc(qx_sz);
+    vk_buffer qx_buf = ggml_vk_create_buffer_check(ctx->device, qx_sz, vk::MemoryPropertyFlagBits::eDeviceLocal);
+    vk_buffer x_buf = ggml_vk_create_buffer_check(ctx->device, x_sz_f16, vk::MemoryPropertyFlagBits::eDeviceLocal);
+    float * x_ref = (float *) malloc(x_sz);
+    ggml_fp16_t * x_chk = (ggml_fp16_t *) malloc(x_sz_f16);
+
+    for (size_t i = 0; i < ne; i++) {
+        x[i] = rand() / (float)RAND_MAX;
+    }
+
+    vk_pipeline p = ggml_vk_get_to_fp16(ctx, quant);
+
+    ggml_vk_quantize_data(x, qx, ne, quant);
+    ggml_vk_dequantize_data(qx, x_ref, ne, quant);
+
+    ggml_pipeline_request_descriptor_sets(ctx->device, p, 1);
+
+    ggml_pipeline_allocate_descriptor_sets(ctx->device);
+
+    ggml_vk_buffer_write(qx_buf, 0, qx, qx_sz);
+
+    vk_context subctx = ggml_vk_create_context(ctx, ctx->device->compute_queue);
+    ggml_vk_ctx_begin(ctx->device, subctx);
+    const std::vector<uint32_t> pc = { 1, (uint32_t)ne, (uint32_t)ne, (uint32_t)ne, (uint32_t)ne };
+    ggml_vk_dispatch_pipeline(ctx, subctx, p, { vk_subbuffer{ qx_buf, 0, qx_sz }, vk_subbuffer{ x_buf, 0, x_sz_f16 } }, pc.size() * sizeof(int), pc.data(), { (uint32_t)ne, 1, 1});
+    ggml_vk_ctx_end(subctx);
+
+    auto begin = std::chrono::high_resolution_clock::now();
+
+    ggml_vk_submit(subctx, ctx->fence);
+    VK_CHECK(ctx->device->device.waitForFences({ ctx->fence }, true, UINT64_MAX), "ggml_vk_test_dequant waitForFences");
+    ctx->device->device.resetFences({ ctx->fence });
+
+    auto end = std::chrono::high_resolution_clock::now();
+
+    double ms_dequant = std::chrono::duration_cast<std::chrono::microseconds>(end-begin).count() / 1000.0;
+    ggml_vk_buffer_read(x_buf, 0, x_chk, x_sz_f16);
+
+    int first_err = -1;
+
+    double avg_err = 0.0;
+    for (size_t i = 0; i < ne; i++) {
+        double error = std::fabs(x_ref[i] - ggml_fp16_to_fp32(x_chk[i]));
+        avg_err += error;
+
+        if (first_err < 0 && error > 0.05) {
+            first_err = i;
+        }
+    }
+
+    avg_err /= ne;
+
+    std::cerr << "TEST DEQUANT " << ggml_type_name(quant) << " time=" << ms_dequant << "ms avg_err=" << avg_err << std::endl;
+
+    if (avg_err > 0.1) {
+        std::cerr << "first_error = " << first_err << std::endl;
+        std::cerr << "Actual result: " << std::endl << std::endl;
+        for (int i = std::max(0, first_err - 5); i < std::min((int)ne, first_err + 5); i++) {
+            std::cerr << ggml_fp16_to_fp32(x_chk[i]) << ", ";
+        }
+        std::cerr << std::endl << "Expected result: " << std::endl << std::endl;
+        for (int i = std::max(0, first_err - 5); i < std::min((int)ne, first_err + 5); i++) {
+            std::cerr << x_ref[i] << ", ";
+        }
+        std::cerr << std::endl;
+    }
+
+    ggml_vk_destroy_buffer(x_buf);
+    ggml_vk_destroy_buffer(qx_buf);
+
+    free(x);
+    free(qx);
+    free(x_ref);
+    free(x_chk);
+}
+
+static void ggml_vk_test_dequant_matmul(ggml_backend_vk_context * ctx, size_t m, size_t n, size_t k, size_t batch, size_t num_it, size_t split_k, size_t shader_size, ggml_type quant) {
+    VK_LOG_DEBUG("ggml_vk_test_dequant_matmul(" << m << ", " << n << ", " << k << ", " << batch << ", " << num_it << ", " << split_k << ", " << ggml_type_name(quant) << ")");
+    const size_t x_ne = m * k * batch;
+    const size_t y_ne = k * n * batch;
+    const size_t d_ne = m * n * batch;
+
+    vk_pipeline p;
+    std::string shname;
+    if (shader_size == 0) {
+        p = ctx->device->pipeline_dequant_mul_mat_mat[quant]->a_s;
+        shname = std::string(ggml_type_name(quant)) + "_ALIGNED_S";
+    } else if (shader_size == 1) {
+        p = ctx->device->pipeline_dequant_mul_mat_mat[quant]->a_m;
+        shname = std::string(ggml_type_name(quant)) + "_ALIGNED_M";
+    } else if (shader_size == 2) {
+        p = ctx->device->pipeline_dequant_mul_mat_mat[quant]->a_l;
+        shname = std::string(ggml_type_name(quant)) + "_ALIGNED_L";
+    } else {
+        GGML_ASSERT(0);
+    }
+
+    const size_t kpad = ggml_vk_align_size(k, p->align);
+
+    if (k != kpad) {
+        if (shader_size == 0) {
+            p = ctx->device->pipeline_dequant_mul_mat_mat[quant]->s;
+            shname = std::string(ggml_type_name(quant)) + "_S";
+        } else if (shader_size == 1) {
+            p = ctx->device->pipeline_dequant_mul_mat_mat[quant]->m;
+            shname = std::string(ggml_type_name(quant)) + "_M";
+        } else if (shader_size == 2) {
+            p = ctx->device->pipeline_dequant_mul_mat_mat[quant]->l;
+            shname = std::string(ggml_type_name(quant)) + "_L";
+        } else {
+            GGML_ASSERT(0);
+        }
+    }
+
+    const size_t x_sz = sizeof(float) * x_ne;
+    const size_t y_sz = sizeof(float) * y_ne;
+    const size_t qx_sz = x_ne * ggml_type_size(quant)/ggml_blck_size(quant);
+    const size_t d_sz = sizeof(float) * d_ne;
+    float * x = (float *) malloc(x_sz);
+    float * y = (float *) malloc(y_sz);
+    void * qx = malloc(qx_sz);
+    vk_buffer qx_buf = ggml_vk_create_buffer_check(ctx->device, qx_sz, vk::MemoryPropertyFlagBits::eDeviceLocal);
+    vk_buffer y_buf = ggml_vk_create_buffer_check(ctx->device, y_sz, vk::MemoryPropertyFlagBits::eDeviceLocal);
+    vk_buffer d_buf = ggml_vk_create_buffer_check(ctx->device, d_sz, vk::MemoryPropertyFlagBits::eDeviceLocal);
+    float * d = (float *) malloc(d_sz);
+    float * d_chk = (float *) malloc(d_sz);
+
+    for (size_t i = 0; i < x_ne; i++) {
+        x[i] = (rand() / (float)RAND_MAX) * 2.0f - 1.0f;
+    }
+
+    ggml_vk_quantize_data(x, qx, x_ne, quant);
+
+    for (size_t i = 0; i < y_ne; i++) {
+        // y[i] = rand() / (float)RAND_MAX;
+        y[i] = (i % k == i / k) ? 1.0f : 0.0f;
+    }
+
+    ggml_pipeline_request_descriptor_sets(ctx->device, p, num_it);
+    if (split_k > 1) {
+        ggml_pipeline_request_descriptor_sets(ctx->device, ctx->device->pipeline_matmul_split_k_reduce, num_it);
+
+        if (ctx->prealloc_split_k == nullptr || ctx->prealloc_split_k->size < sizeof(float) * d_ne * split_k) {
+            // Resize buffer
+            if (ctx->prealloc_split_k != nullptr) {
+                ggml_vk_destroy_buffer(ctx->prealloc_split_k);
+            }
+            ctx->prealloc_split_k = ggml_vk_create_buffer_check(ctx->device, sizeof(float) * d_ne * split_k, vk::MemoryPropertyFlagBits::eDeviceLocal);
+        }
+    }
+
+    ggml_pipeline_allocate_descriptor_sets(ctx->device);
+
+    ggml_vk_buffer_write(qx_buf, 0, qx, qx_sz);
+    ggml_vk_buffer_write(y_buf, 0, y, y_sz);
+
+    vk_context subctx = ggml_vk_create_context(ctx, ctx->device->compute_queue);
+    for (size_t i = 0; i < num_it; i++) {
+        ggml_vk_ctx_begin(ctx->device, subctx);
+        ggml_vk_matmul(
+            ctx, subctx, p, ggml_vk_subbuffer(qx_buf), ggml_vk_subbuffer(y_buf), ggml_vk_subbuffer(d_buf), ggml_vk_subbuffer(ctx->prealloc_split_k),
+            m, n, k,
+            k, k, m, k*m, k*n, m*n,
+            split_k, batch, batch, batch, 1, 1
+        );
+        ggml_vk_ctx_end(subctx);
+    }
+
+    auto begin = std::chrono::high_resolution_clock::now();
+
+    ggml_vk_submit(subctx, ctx->fence);
+    VK_CHECK(ctx->device->device.waitForFences({ ctx->fence }, true, UINT64_MAX), "ggml_vk_test_dequant waitForFences");
+    ctx->device->device.resetFences({ ctx->fence });
+
+    auto end = std::chrono::high_resolution_clock::now();
+
+    double time_ms = std::chrono::duration_cast<std::chrono::microseconds>(end-begin).count() / 1000.0;
+    ggml_vk_buffer_read(d_buf, 0, d, d_sz);
+
+    ggml_init_params iparams = {
+        /*.mem_size   =*/ 1024*1024*1024,
+        /*.mem_buffer =*/ NULL,
+        /*.no_alloc   =*/ true,
+    };
+
+    ggml_context * ggml_ctx = ggml_init(iparams);
+
+    ggml_tensor * src0_ggml = ggml_new_tensor_3d(ggml_ctx, quant, k, m, batch);
+    ggml_tensor * src1_ggml = ggml_new_tensor_3d(ggml_ctx, GGML_TYPE_F32, k, n, batch);
+    ggml_tensor * tensor_ggml = ggml_mul_mat(ggml_ctx, src0_ggml, src1_ggml);
+
+    src0_ggml->data = qx;
+    src1_ggml->data = y;
+    tensor_ggml->data = d_chk;
+
+    ggml_cgraph * cgraph = ggml_new_graph(ggml_ctx);
+    ggml_build_forward_expand(cgraph, tensor_ggml);
+
+    ggml_graph_compute_with_ctx(ggml_ctx, cgraph, 1);
+
+    ggml_free(ggml_ctx);
+
+    double avg_err = 0.0;
+    int first_err_n = -1;
+    int first_err_m = -1;
+    int first_err_b = -1;
+
+    for (size_t i = 0; i < m*n*batch; i++) {
+        double err = std::fabs(d[i] - d_chk[i]);
+        avg_err += err;
+
+        if ((err > 0.05f || std::isnan(err)) && first_err_n == -1) {
+            first_err_b = i / (m * n);
+            first_err_n = (i % (m * n)) / m;
+            first_err_m = (i % (m * n)) % m;
+        }
+    }
+
+    avg_err /= m * n;
+
+    double tflops = 2.0*m*n*k*batch*num_it / (time_ms / 1000.0) / (1000.0*1000.0*1000.0*1000.0);
+
+    std::cerr << "TEST MMQ " << shname << " m=" << m << " n=" << n << " k=" << k << " batch=" << batch << " split_k=" << split_k << " matmul " << time_ms / num_it << "ms " << tflops << " TFLOPS avg_err=" << avg_err << std::endl;
+
+    if (avg_err > 0.01 || std::isnan(avg_err)) {
+        std::cerr << "m = " << first_err_m << " n = " << first_err_n << " b = " << first_err_b << std::endl;
+        std::cerr << "Actual result: " << std::endl << std::endl;
+        ggml_vk_print_matrix_area(d, GGML_TYPE_F32, m, n, first_err_m, first_err_n, first_err_b);
+        std::cerr << std::endl;
+        std::cerr << "Expected result: " << std::endl << std::endl;
+        ggml_vk_print_matrix_area(d_chk, GGML_TYPE_F32, m, n, first_err_m, first_err_n, first_err_b);
+
+        if (split_k > 1) {
+            float * split_k_buf = (float *) malloc(sizeof(float) * d_ne * split_k);
+            ggml_vk_buffer_read(ctx->prealloc_split_k, 0, split_k_buf, sizeof(float) * d_ne * split_k);
+
+            std::cerr << "d_buf0: " << std::endl << std::endl;
+            ggml_vk_print_matrix_area(split_k_buf, GGML_TYPE_F32, m, n, first_err_m, first_err_n, first_err_b);
+
+            std::cerr << "d_buf1: " << std::endl << std::endl;
+            ggml_vk_print_matrix_area(split_k_buf + d_ne, GGML_TYPE_F32, m, n, first_err_m, first_err_n, first_err_b);
+
+            std::cerr << "d_buf2: " << std::endl << std::endl;
+            ggml_vk_print_matrix_area(split_k_buf + 2 * d_ne, GGML_TYPE_F32, m, n, first_err_m, first_err_n, first_err_b);
+
+            std::cerr << "d_buf3: " << std::endl << std::endl;
+            ggml_vk_print_matrix_area(split_k_buf + 3 * d_ne, GGML_TYPE_F32, m, n, first_err_m, first_err_n, first_err_b);
+
+            free(split_k_buf);
+        }
+    }
+
+    ggml_vk_destroy_buffer(qx_buf);
+    ggml_vk_destroy_buffer(y_buf);
+    ggml_vk_destroy_buffer(d_buf);
+
+    free(x);
+    free(qx);
+    free(y);
+    free(d);
+    free(d_chk);
+}
+#endif
+
+static void ggml_vk_preallocate_buffers(ggml_backend_vk_context * ctx) {
+#if defined(GGML_VULKAN_RUN_TESTS)
+    ggml_vk_test_dequant(ctx, 7680, GGML_TYPE_F32);
+    ggml_vk_test_dequant(ctx, 7680, GGML_TYPE_Q4_0);
+    ggml_vk_test_dequant(ctx, 7680, GGML_TYPE_Q4_1);
+    ggml_vk_test_dequant(ctx, 7680, GGML_TYPE_Q5_0);
+    ggml_vk_test_dequant(ctx, 7680, GGML_TYPE_Q5_1);
+    ggml_vk_test_dequant(ctx, 7680, GGML_TYPE_Q8_0);
+    ggml_vk_test_dequant(ctx, 7680, GGML_TYPE_Q2_K);
+    ggml_vk_test_dequant(ctx, 7680, GGML_TYPE_Q3_K);
+    ggml_vk_test_dequant(ctx, 7680, GGML_TYPE_Q4_K);
+    ggml_vk_test_dequant(ctx, 7680, GGML_TYPE_Q5_K);
+    ggml_vk_test_dequant(ctx, 7680, GGML_TYPE_Q6_K);
+    ggml_vk_test_dequant(ctx, 7680, GGML_TYPE_IQ4_NL);
+
+    ggml_vk_test_matmul<ggml_fp16_t, ggml_fp16_t>(ctx, 512, 512, 100, 32, 100, 1, 2);
+
+    ggml_vk_test_matmul<float, float>(ctx, 128, 512, 512, 2, 100, 1, 0);
+    ggml_vk_test_matmul<float, float>(ctx, 128, 512, 512, 2, 100, 1, 1);
+    ggml_vk_test_matmul<float, float>(ctx, 128, 512, 512, 2, 100, 1, 2);
+    // ggml_vk_test_matmul<float, float>(ctx, 128, 512, 512, 2, 100, 4, 0);
+    // ggml_vk_test_matmul<float, float>(ctx, 128, 512, 512, 2, 100, 4, 1);
+    // ggml_vk_test_matmul<float, float>(ctx, 128, 512, 512, 2, 100, 4, 2);
+
+    ggml_vk_test_dequant_matmul(ctx, 128, 512, 512, 2, 100, 1, 0, GGML_TYPE_Q4_0);
+    ggml_vk_test_dequant_matmul(ctx, 128, 512, 512, 2, 100, 1, 1, GGML_TYPE_Q4_0);
+    ggml_vk_test_dequant_matmul(ctx, 128, 512, 512, 2, 100, 1, 2, GGML_TYPE_Q4_0);
+    // ggml_vk_test_dequant_matmul(ctx, 128, 512, 512, 2, 100, 4, 0, GGML_TYPE_Q4_0);
+    // ggml_vk_test_dequant_matmul(ctx, 128, 512, 512, 2, 100, 4, 1, GGML_TYPE_Q4_0);
+    // ggml_vk_test_dequant_matmul(ctx, 128, 512, 512, 2, 100, 4, 2, GGML_TYPE_Q4_0);
+
+    ggml_vk_test_dequant_matmul(ctx, 128, 512, 512, 2, 100, 1, 0, GGML_TYPE_Q4_1);
+    ggml_vk_test_dequant_matmul(ctx, 128, 512, 512, 2, 100, 1, 1, GGML_TYPE_Q4_1);
+    ggml_vk_test_dequant_matmul(ctx, 128, 512, 512, 2, 100, 1, 2, GGML_TYPE_Q4_1);
+    // ggml_vk_test_dequant_matmul(ctx, 128, 512, 512, 2, 100, 4, 0, GGML_TYPE_Q4_1);
+    // ggml_vk_test_dequant_matmul(ctx, 128, 512, 512, 2, 100, 4, 1, GGML_TYPE_Q4_1);
+    // ggml_vk_test_dequant_matmul(ctx, 128, 512, 512, 2, 100, 4, 2, GGML_TYPE_Q4_1);
+
+    ggml_vk_test_dequant_matmul(ctx, 128, 512, 512, 2, 100, 1, 0, GGML_TYPE_Q5_0);
+    ggml_vk_test_dequant_matmul(ctx, 128, 512, 512, 2, 100, 1, 1, GGML_TYPE_Q5_0);
+    ggml_vk_test_dequant_matmul(ctx, 128, 512, 512, 2, 100, 1, 2, GGML_TYPE_Q5_0);
+    // ggml_vk_test_dequant_matmul(ctx, 128, 512, 512, 2, 100, 4, 0, GGML_TYPE_Q5_0);
+    // ggml_vk_test_dequant_matmul(ctx, 128, 512, 512, 2, 100, 4, 1, GGML_TYPE_Q5_0);
+    // ggml_vk_test_dequant_matmul(ctx, 128, 512, 512, 2, 100, 4, 2, GGML_TYPE_Q5_0);
+
+    ggml_vk_test_dequant_matmul(ctx, 128, 512, 512, 2, 100, 1, 0, GGML_TYPE_Q5_1);
+    ggml_vk_test_dequant_matmul(ctx, 128, 512, 512, 2, 100, 1, 1, GGML_TYPE_Q5_1);
+    ggml_vk_test_dequant_matmul(ctx, 128, 512, 512, 2, 100, 1, 2, GGML_TYPE_Q5_1);
+    // ggml_vk_test_dequant_matmul(ctx, 128, 512, 512, 2, 100, 4, 0, GGML_TYPE_Q5_1);
+    // ggml_vk_test_dequant_matmul(ctx, 128, 512, 512, 2, 100, 4, 1, GGML_TYPE_Q5_1);
+    // ggml_vk_test_dequant_matmul(ctx, 128, 512, 512, 2, 100, 4, 2, GGML_TYPE_Q5_1);
+
+    ggml_vk_test_dequant_matmul(ctx, 128, 512, 512, 2, 100, 1, 0, GGML_TYPE_Q8_0);
+    ggml_vk_test_dequant_matmul(ctx, 128, 512, 512, 2, 100, 1, 1, GGML_TYPE_Q8_0);
+    ggml_vk_test_dequant_matmul(ctx, 128, 512, 512, 2, 100, 1, 2, GGML_TYPE_Q8_0);
+    // ggml_vk_test_dequant_matmul(ctx, 128, 512, 512, 2, 100, 4, 0, GGML_TYPE_Q8_0);
+    // ggml_vk_test_dequant_matmul(ctx, 128, 512, 512, 2, 100, 4, 1, GGML_TYPE_Q8_0);
+    // ggml_vk_test_dequant_matmul(ctx, 128, 512, 512, 2, 100, 4, 2, GGML_TYPE_Q8_0);
+
+    ggml_vk_test_dequant_matmul(ctx, 128, 512, 512, 2, 100, 1, 0, GGML_TYPE_Q2_K);
+    ggml_vk_test_dequant_matmul(ctx, 128, 512, 512, 2, 100, 1, 1, GGML_TYPE_Q2_K);
+    ggml_vk_test_dequant_matmul(ctx, 128, 512, 512, 2, 100, 1, 2, GGML_TYPE_Q2_K);
+    // ggml_vk_test_dequant_matmul(ctx, 128, 512, 512, 2, 100, 4, 0, GGML_TYPE_Q2_K);
+    // ggml_vk_test_dequant_matmul(ctx, 128, 512, 512, 2, 100, 4, 1, GGML_TYPE_Q2_K);
+    // ggml_vk_test_dequant_matmul(ctx, 128, 512, 512, 2, 100, 4, 2, GGML_TYPE_Q2_K);
+
+    ggml_vk_test_dequant_matmul(ctx, 128, 512, 512, 2, 100, 1, 0, GGML_TYPE_Q3_K);
+    ggml_vk_test_dequant_matmul(ctx, 128, 512, 512, 2, 100, 1, 1, GGML_TYPE_Q3_K);
+    ggml_vk_test_dequant_matmul(ctx, 128, 512, 512, 2, 100, 1, 2, GGML_TYPE_Q3_K);
+    // ggml_vk_test_dequant_matmul(ctx, 128, 512, 512, 2, 100, 4, 0, GGML_TYPE_Q3_K);
+    // ggml_vk_test_dequant_matmul(ctx, 128, 512, 512, 2, 100, 4, 1, GGML_TYPE_Q3_K);
+    // ggml_vk_test_dequant_matmul(ctx, 128, 512, 512, 2, 100, 4, 2, GGML_TYPE_Q3_K);
+
+    ggml_vk_test_dequant_matmul(ctx, 128, 512, 512, 2, 100, 1, 0, GGML_TYPE_Q4_K);
+    ggml_vk_test_dequant_matmul(ctx, 128, 512, 512, 2, 100, 1, 1, GGML_TYPE_Q4_K);
+    ggml_vk_test_dequant_matmul(ctx, 128, 512, 512, 2, 100, 1, 2, GGML_TYPE_Q4_K);
+    // ggml_vk_test_dequant_matmul(ctx, 128, 512, 512, 2, 100, 4, 0, GGML_TYPE_Q4_K);
+    // ggml_vk_test_dequant_matmul(ctx, 128, 512, 512, 2, 100, 4, 1, GGML_TYPE_Q4_K);
+    // ggml_vk_test_dequant_matmul(ctx, 128, 512, 512, 2, 100, 4, 2, GGML_TYPE_Q4_K);
+
+    ggml_vk_test_dequant_matmul(ctx, 128, 512, 512, 2, 100, 1, 0, GGML_TYPE_Q5_K);
+    ggml_vk_test_dequant_matmul(ctx, 128, 512, 512, 2, 100, 1, 1, GGML_TYPE_Q5_K);
+    ggml_vk_test_dequant_matmul(ctx, 128, 512, 512, 2, 100, 1, 2, GGML_TYPE_Q5_K);
+    // ggml_vk_test_dequant_matmul(ctx, 128, 512, 512, 2, 100, 4, 0, GGML_TYPE_Q5_K);
+    // ggml_vk_test_dequant_matmul(ctx, 128, 512, 512, 2, 100, 4, 1, GGML_TYPE_Q5_K);
+    // ggml_vk_test_dequant_matmul(ctx, 128, 512, 512, 2, 100, 4, 2, GGML_TYPE_Q5_K);
+
+    ggml_vk_test_dequant_matmul(ctx, 128, 512, 512, 2, 100, 1, 0, GGML_TYPE_Q6_K);
+    ggml_vk_test_dequant_matmul(ctx, 128, 512, 512, 2, 100, 1, 1, GGML_TYPE_Q6_K);
+    ggml_vk_test_dequant_matmul(ctx, 128, 512, 512, 2, 100, 1, 2, GGML_TYPE_Q6_K);
+    // ggml_vk_test_dequant_matmul(ctx, 128, 512, 512, 2, 100, 4, 0, GGML_TYPE_Q6_K);
+    // ggml_vk_test_dequant_matmul(ctx, 128, 512, 512, 2, 100, 4, 1, GGML_TYPE_Q6_K);
+    // ggml_vk_test_dequant_matmul(ctx, 128, 512, 512, 2, 100, 4, 2, GGML_TYPE_Q6_K);
+
+    ggml_vk_test_dequant_matmul(ctx, 128, 512, 512, 2, 100, 1, 0, GGML_TYPE_IQ4_NL);
+    ggml_vk_test_dequant_matmul(ctx, 128, 512, 512, 2, 100, 1, 1, GGML_TYPE_IQ4_NL);
+    ggml_vk_test_dequant_matmul(ctx, 128, 512, 512, 2, 100, 1, 2, GGML_TYPE_IQ4_NL);
+
+    std::cerr << std::endl;
+
+    const std::vector<size_t> vals {
+        8, 8, 8,
+        100, 46, 576,
+        623, 111, 128,
+        100, 46, 558,
+        512, 1, 256,
+        128, 110, 622,
+        511, 511, 127,
+        511, 511, 7,
+        511, 511, 17,
+        49, 49, 128,
+        128, 49, 49,
+        4096, 49, 4096,
+        11008, 49, 4096,
+        4096, 49, 11008,
+        32000, 49, 4096,
+        512, 512, 128,
+        128, 512, 512,
+        4096, 512, 4096,
+        11008, 512, 4096,
+        4096, 512, 11008,
+        32000, 512, 4096,
+    };
+    const size_t num_it = 1;
+    for (size_t i = 0; i < vals.size(); i += 3) {
+        ggml_vk_test_matmul<ggml_fp16_t, float>(ctx, vals[i], vals[i + 1], vals[i + 2], 2, num_it, 1, 0);
+        ggml_vk_test_matmul<ggml_fp16_t, float>(ctx, vals[i], vals[i + 1], vals[i + 2], 2, num_it, 1, 1);
+        ggml_vk_test_matmul<ggml_fp16_t, float>(ctx, vals[i], vals[i + 1], vals[i + 2], 2, num_it, 1, 2);
+        // ggml_vk_test_matmul<ggml_fp16_t, float>(ctx, vals[i], vals[i + 1], vals[i + 2], 2, num_it, 4, 0);
+        // ggml_vk_test_matmul<ggml_fp16_t, float>(ctx, vals[i], vals[i + 1], vals[i + 2], 2, num_it, 4, 1);
+        // ggml_vk_test_matmul<ggml_fp16_t, float>(ctx, vals[i], vals[i + 1], vals[i + 2], 2, num_it, 4, 2);
+        std::cerr << std::endl;
+    }
+
+    GGML_ABORT("fatal error");
+#endif
+
+    if (ctx->prealloc_x == nullptr || (ctx->prealloc_size_x > 0 && ctx->prealloc_x->size < ctx->prealloc_size_x)) {
+        VK_LOG_MEMORY("ggml_vk_preallocate_buffers(x_size: " << ctx->prealloc_size_x << ")");
+        // Resize buffer
+        if (ctx->prealloc_x != nullptr) {
+            ggml_vk_destroy_buffer(ctx->prealloc_x);
+        }
+        ctx->prealloc_x = ggml_vk_create_buffer_device(ctx->device, ctx->prealloc_size_x);
+    }
+    if (ctx->prealloc_y == nullptr || (ctx->prealloc_size_y > 0 && ctx->prealloc_y->size < ctx->prealloc_size_y)) {
+        VK_LOG_MEMORY("ggml_vk_preallocate_buffers(y_size: " << ctx->prealloc_size_y << ")");
+        // Resize buffer
+        if (ctx->prealloc_y != nullptr) {
+            ggml_vk_destroy_buffer(ctx->prealloc_y);
+        }
+        ctx->prealloc_y = ggml_vk_create_buffer_device(ctx->device, ctx->prealloc_size_y);
+    }
+    if (ctx->prealloc_split_k == nullptr || (ctx->prealloc_size_split_k > 0 && ctx->prealloc_split_k->size < ctx->prealloc_size_split_k)) {
+        VK_LOG_MEMORY("ggml_vk_preallocate_buffers(split_k_size: " << ctx->prealloc_size_split_k << ")");
+        // Resize buffer
+        if (ctx->prealloc_split_k != nullptr) {
+            ggml_vk_destroy_buffer(ctx->prealloc_split_k);
+        }
+        ctx->prealloc_split_k = ggml_vk_create_buffer_device(ctx->device, ctx->prealloc_size_split_k);
+    }
+}
+
+static bool ggml_vk_compute_forward(ggml_backend_vk_context* ctx, ggml_tensor* tensor, int tensor_idx, bool use_fence);
+
+// Returns true if node has enqueued work into the queue, false otherwise
+// If submit is true the current all operations queued so far are being submitted to Vulkan to overlap cmdlist creation and GPU execution.
+static bool ggml_vk_build_graph(ggml_backend_vk_context * ctx, ggml_tensor * node, int node_idx, ggml_tensor *node_begin, int node_idx_begin, bool dryrun, bool last_node, bool submit){
+    if (ggml_is_empty(node) || !node->buffer) {
+        return false;
+    }
+
+    VK_LOG_DEBUG("ggml_vk_build_graph(" << node << ", " << ggml_op_name(node->op) << ")");
+    ctx->semaphore_idx = 0;
+
+    const ggml_tensor * src0 = node->src[0];
+    const ggml_tensor * src1 = node->src[1];
+    const ggml_tensor * src2 = node->src[2];
+
+    switch (node->op) {
+    // Return on empty ops to avoid generating a compute_ctx and setting exit_tensor
+    case GGML_OP_RESHAPE:
+    case GGML_OP_VIEW:
+    case GGML_OP_PERMUTE:
+    case GGML_OP_TRANSPOSE:
+    case GGML_OP_NONE:
+        return false;
+    case GGML_OP_UNARY:
+        switch (ggml_get_unary_op(node)) {
+        case GGML_UNARY_OP_SILU:
+        case GGML_UNARY_OP_GELU:
+        case GGML_UNARY_OP_GELU_QUICK:
+        case GGML_UNARY_OP_RELU:
+        case GGML_UNARY_OP_TANH:
+            break;
+        default:
+            return false;
+        }
+        break;
+    case GGML_OP_REPEAT:
+    case GGML_OP_GET_ROWS:
+    case GGML_OP_ADD:
+    case GGML_OP_ACC:
+    case GGML_OP_MUL:
+    case GGML_OP_DIV:
+    case GGML_OP_CONCAT:
+    case GGML_OP_UPSCALE:
+    case GGML_OP_SCALE:
+    case GGML_OP_SQR:
+    case GGML_OP_SIN:
+    case GGML_OP_COS:
+    case GGML_OP_CLAMP:
+    case GGML_OP_PAD:
+    case GGML_OP_CPY:
+    case GGML_OP_CONT:
+    case GGML_OP_DUP:
+    case GGML_OP_NORM:
+    case GGML_OP_GROUP_NORM:
+    case GGML_OP_RMS_NORM:
+    case GGML_OP_DIAG_MASK_INF:
+    case GGML_OP_SOFT_MAX:
+    case GGML_OP_ROPE:
+    case GGML_OP_MUL_MAT:
+    case GGML_OP_MUL_MAT_ID:
+    case GGML_OP_ARGSORT:
+    case GGML_OP_SUM_ROWS:
+    case GGML_OP_IM2COL:
+    case GGML_OP_TIMESTEP_EMBEDDING:
+    case GGML_OP_LEAKY_RELU:
+        break;
+    default:
+        std::cerr << "ggml_vulkan: Error: Missing op: " << ggml_op_name(node->op) << std::endl;
+        GGML_ABORT("fatal error");
+        return false;
+    }
+
+    vk_context compute_ctx;
+
+    if (!dryrun) {
+        if (ctx->compute_ctx.expired()) {
+            compute_ctx = ggml_vk_create_context(ctx, ctx->device->compute_queue);
+            ctx->compute_ctx = compute_ctx;
+            ggml_vk_ctx_begin(ctx->device, compute_ctx);
+        } else {
+            compute_ctx = ctx->compute_ctx.lock();
+        }
+    }
+
+    switch (node->op) {
+    case GGML_OP_REPEAT:
+        ggml_vk_repeat(ctx, compute_ctx, src0, node, dryrun);
+
+        break;
+    case GGML_OP_ACC:
+        ggml_vk_acc(ctx, compute_ctx, src0, src1, node, dryrun);
+
+        break;
+    case GGML_OP_GET_ROWS:
+        ggml_vk_get_rows(ctx, compute_ctx, src0, src1, node, dryrun);
+
+        break;
+    case GGML_OP_ADD:
+        ggml_vk_add(ctx, compute_ctx, src0, src1, node, dryrun);
+
+        break;
+    case GGML_OP_MUL:
+        ggml_vk_mul(ctx, compute_ctx, src0, src1, node, dryrun);
+
+        break;
+    case GGML_OP_DIV:
+        ggml_vk_div(ctx, compute_ctx, src0, src1, node, dryrun);
+
+        break;
+    case GGML_OP_CONCAT:
+        ggml_vk_concat(ctx, compute_ctx, src0, src1, node, dryrun);
+
+        break;
+    case GGML_OP_UPSCALE:
+        ggml_vk_upscale(ctx, compute_ctx, src0, node, dryrun);
+
+        break;
+    case GGML_OP_SCALE:
+        ggml_vk_scale(ctx, compute_ctx, src0, node, dryrun);
+
+        break;
+    case GGML_OP_SQR:
+        ggml_vk_sqr(ctx, compute_ctx, src0, node, dryrun);
+
+        break;
+    case GGML_OP_SIN:
+        ggml_vk_sin(ctx, compute_ctx, src0, node, dryrun);
+
+        break;
+    case GGML_OP_COS:
+        ggml_vk_cos(ctx, compute_ctx, src0, node, dryrun);
+
+        break;
+    case GGML_OP_CLAMP:
+        ggml_vk_clamp(ctx, compute_ctx, src0, node, dryrun);
+
+        break;
+    case GGML_OP_PAD:
+        ggml_vk_pad(ctx, compute_ctx, src0, node, dryrun);
+
+        break;
+    case GGML_OP_CPY:
+    case GGML_OP_CONT:
+    case GGML_OP_DUP:
+        ggml_vk_cpy(ctx, compute_ctx, src0, node, dryrun);
+
+        break;
+    case GGML_OP_NORM:
+        ggml_vk_norm(ctx, compute_ctx, src0, node, dryrun);
+
+        break;
+    case GGML_OP_GROUP_NORM:
+        ggml_vk_group_norm(ctx, compute_ctx, src0, node, dryrun);
+
+        break;
+    case GGML_OP_RMS_NORM:
+        ggml_vk_rms_norm(ctx, compute_ctx, src0, node, dryrun);
+
+        break;
+    case GGML_OP_UNARY:
+        switch (ggml_get_unary_op(node)) {
+        case GGML_UNARY_OP_SILU:
+        case GGML_UNARY_OP_GELU:
+        case GGML_UNARY_OP_GELU_QUICK:
+        case GGML_UNARY_OP_RELU:
+        case GGML_UNARY_OP_TANH:
+            ggml_vk_unary(ctx, compute_ctx, src0, node, dryrun);
+            break;
+        default:
+            return false;
+        }
+        break;
+    case GGML_OP_DIAG_MASK_INF:
+        ggml_vk_diag_mask_inf(ctx, compute_ctx, src0, node, dryrun);
+
+        break;
+    case GGML_OP_SOFT_MAX:
+        ggml_vk_soft_max(ctx, compute_ctx, src0, src1, node, dryrun);
+
+        break;
+    case GGML_OP_ROPE:
+        ggml_vk_rope(ctx, compute_ctx, src0, src1, src2, node, dryrun);
+
+        break;
+    case GGML_OP_ARGSORT:
+        ggml_vk_argsort(ctx, compute_ctx, src0, node, dryrun);
+
+        break;
+    case GGML_OP_SUM_ROWS:
+        ggml_vk_sum_rows(ctx, compute_ctx, src0, node, dryrun);
+
+        break;
+    case GGML_OP_IM2COL:
+        ggml_vk_im2col(ctx, compute_ctx, src0, src1, node, dryrun);
+
+        break;
+    case GGML_OP_TIMESTEP_EMBEDDING:
+        ggml_vk_timestep_embedding(ctx, compute_ctx, src0, node, dryrun);
+
+        break;
+    case GGML_OP_LEAKY_RELU:
+        ggml_vk_leaky_relu(ctx, compute_ctx, src0, node, dryrun);
+
+        break;
+    case GGML_OP_MUL_MAT:
+        ggml_vk_mul_mat(ctx, compute_ctx, src0, src1, node, dryrun);
+
+        break;
+    case GGML_OP_MUL_MAT_ID:
+        ggml_vk_mul_mat_id(ctx, compute_ctx, src0, src1, src2, node, dryrun);
+
+        break;
+    default:
+        return false;
+    }
+
+    if (dryrun) {
+        return false;
+    }
+
+    ctx->tensor_ctxs[node_idx] = compute_ctx;
+
+#if defined(GGML_VULKAN_CHECK_RESULTS) || defined(GGML_VULKAN_PERF)
+    // Force context reset on each node so that each tensor ends up in its own context
+    // and can be run and compared to its CPU equivalent separately
+    last_node = true;
+#endif
+
+    if (submit || last_node) {
+        ggml_vk_ctx_end(compute_ctx);
+
+        // TODO probably it'd be better to pass a exit_node flag to ggml_vk_compute_forward
+        if (last_node) {
+            compute_ctx->exit_tensor_idx = node_idx_begin;
+        }
+        else {
+            compute_ctx->exit_tensor_idx = -1;
+        }
+
+        ctx->compute_ctx.reset();
+
+        bool ok = ggml_vk_compute_forward(ctx, node_begin, node_idx_begin, false);
+        if (!ok) {
+            if (node->op == GGML_OP_UNARY) {
+                std::cerr << __func__ << ": error: op not supported UNARY " << node->name << " (" << ggml_unary_op_name(static_cast<ggml_unary_op>(node->op_params[0])) << ")" << std::endl;
+            }
+            else {
+                std::cerr << __func__ << ": error: op not supported " << node->name << " (" << ggml_op_name(node->op) << ")" << std::endl;
+            }
+        }
+
+    }
+    return true;
+}
+
+static bool ggml_vk_compute_forward(ggml_backend_vk_context * ctx, ggml_tensor * tensor, int tensor_idx, bool use_fence = true){
+    ggml_backend_buffer * buf = nullptr;
+
+    switch (tensor->op) {
+    case GGML_OP_ADD:
+    case GGML_OP_ACC:
+    case GGML_OP_GET_ROWS:
+    case GGML_OP_MUL:
+    case GGML_OP_DIV:
+    case GGML_OP_CONCAT:
+    case GGML_OP_UPSCALE:
+    case GGML_OP_SCALE:
+    case GGML_OP_SQR:
+    case GGML_OP_SIN:
+    case GGML_OP_COS:
+    case GGML_OP_CLAMP:
+    case GGML_OP_PAD:
+    case GGML_OP_CPY:
+    case GGML_OP_CONT:
+    case GGML_OP_DUP:
+    case GGML_OP_NORM:
+    case GGML_OP_GROUP_NORM:
+    case GGML_OP_RMS_NORM:
+    case GGML_OP_DIAG_MASK_INF:
+    case GGML_OP_SOFT_MAX:
+    case GGML_OP_ROPE:
+    case GGML_OP_RESHAPE:
+    case GGML_OP_VIEW:
+    case GGML_OP_PERMUTE:
+    case GGML_OP_TRANSPOSE:
+    case GGML_OP_NONE:
+    case GGML_OP_ARGSORT:
+    case GGML_OP_SUM_ROWS:
+    case GGML_OP_IM2COL:
+    case GGML_OP_TIMESTEP_EMBEDDING:
+    case GGML_OP_LEAKY_RELU:
+    case GGML_OP_REPEAT:
+        buf = tensor->buffer;
+
+        break;
+    case GGML_OP_UNARY:
+        switch (ggml_get_unary_op(tensor)) {
+        case GGML_UNARY_OP_SILU:
+        case GGML_UNARY_OP_GELU:
+        case GGML_UNARY_OP_GELU_QUICK:
+        case GGML_UNARY_OP_RELU:
+        case GGML_UNARY_OP_TANH:
+            buf = tensor->buffer;
+            break;
+        default:
+            return false;
+        }
+        break;
+    case GGML_OP_MUL_MAT:
+    case GGML_OP_MUL_MAT_ID:
+        buf = tensor->buffer;
+
+        break;
+    default:
+        return false;
+    }
+
+    if (buf == nullptr) {
+        return false;
+    }
+
+    VK_LOG_DEBUG("ggml_vk_compute_forward(" << tensor << ", name=" << tensor->name << ", op=" << ggml_op_name(tensor->op) << ", type=" << tensor->type << ", ne0=" << tensor->ne[0] << ", ne1=" << tensor->ne[1] << ", ne2=" << tensor->ne[2] << ", ne3=" << tensor->ne[3] << ", nb0=" << tensor->nb[0] << ", nb1=" << tensor->nb[1] << ", nb2=" << tensor->nb[2] << ", nb3=" << tensor->nb[3] << ", view_src=" << tensor->view_src << ", view_offs=" << tensor->view_offs << ")");
+
+    vk_context subctx = ctx->tensor_ctxs[tensor_idx].lock();
+
+    // always wait for the GPU work to be done for the last submit
+    if (tensor_idx == subctx->exit_tensor_idx) {
+        use_fence = true;
+    }
+
+    // Only run if ctx hasn't been submitted yet
+    if (!subctx->seqs.empty()) {
+#ifdef GGML_VULKAN_CHECK_RESULTS
+        ggml_vk_check_results_0(tensor);
+        use_fence = true;
+#endif
+
+        // Do staging buffer copies
+        for (auto& cpy : subctx->in_memcpys) {
+            memcpy(cpy.dst, cpy.src, cpy.n);
+        }
+
+        ggml_vk_submit(subctx, use_fence ? ctx->fence : vk::Fence{});
+
+        if (use_fence) {
+            VK_CHECK(ctx->device->device.waitForFences({ ctx->fence }, true, UINT64_MAX), "ggml_vk_compute_forward waitForFences");
+
+            ctx->device->device.resetFences({ ctx->fence });
+        }
+#ifdef GGML_VULKAN_CHECK_RESULTS
+        ggml_vk_check_results_1(tensor);
+#endif
+    }
+
+    if (tensor_idx == subctx->exit_tensor_idx) {
+        // Do staging buffer copies
+        for (auto& cpy : subctx->out_memcpys) {
+            memcpy(cpy.dst, cpy.src, cpy.n);
+        }
+        subctx->in_memcpys.clear();
+        subctx->out_memcpys.clear();
+    }
+
+    return true;
+}
+
+// Clean up after graph processing is done
+static void ggml_vk_graph_cleanup(ggml_backend_vk_context * ctx) {
+    VK_LOG_DEBUG("ggml_vk_graph_cleanup()");
+    for (auto& buffer : ctx->gc.temp_buffers) {
+        ggml_vk_pool_free(ctx, buffer);
+    }
+    ctx->gc.temp_buffers.clear();
+
+    for (auto& dsr : ctx->device->pipeline_descriptor_set_requirements) {
+        vk_pipeline_ref plr = ctx->device->pipelines[dsr.first];
+
+        if (plr.expired()) {
+            continue;
+        }
+
+        vk_pipeline pl = plr.lock();
+        ggml_pipeline_cleanup(pl);
+    }
+
+    ggml_vk_queue_cleanup(ctx->device, ctx->device->compute_queue);
+    ggml_vk_queue_cleanup(ctx->device, ctx->device->transfer_queue);
+
+    for (size_t i = 0; i < ctx->gc.semaphores.size(); i++) {
+        ctx->device->device.destroySemaphore({ ctx->gc.semaphores[i].s });
+    }
+    ctx->gc.semaphores.clear();
+
+    for (size_t i = 0; i < ctx->gc.tl_semaphores.size(); i++) {
+        ctx->device->device.destroySemaphore({ ctx->gc.tl_semaphores[i].s });
+    }
+    ctx->gc.tl_semaphores.clear();
+    ctx->semaphore_idx = 0;
+
+    ctx->event_idx = 0;
+
+    for (auto& event : ctx->gc.events) {
+        ctx->device->device.resetEvent(event);
+    }
+
+    ctx->tensor_ctxs.clear();
+    ctx->gc.contexts.clear();
+    ctx->device->pipeline_descriptor_set_requirements.clear();
+}
+
+// Clean up on backend free
+static void ggml_vk_cleanup(ggml_backend_vk_context * ctx) {
+    VK_LOG_DEBUG("ggml_vk_cleanup(" << ctx->name << ")");
+    ggml_vk_graph_cleanup(ctx);
+
+    ggml_vk_destroy_buffer(ctx->prealloc_x);
+    ggml_vk_destroy_buffer(ctx->prealloc_y);
+    ggml_vk_destroy_buffer(ctx->prealloc_split_k);
+
+    for (auto& buffer : ctx->buffer_pool) {
+        ggml_vk_destroy_buffer(buffer);
+    }
+
+    ctx->prealloc_size_x = 0;
+    ctx->prealloc_size_y = 0;
+    ctx->prealloc_size_split_k = 0;
+
+    for (auto& event : ctx->gc.events) {
+        ctx->device->device.destroyEvent(event);
+    }
+    ctx->gc.events.clear();
+
+    ctx->device->device.destroyFence(ctx->fence);
+}
+
+static int ggml_vk_get_device_count() {
+    ggml_vk_instance_init();
+
+    return vk_instance.device_indices.size();
+}
+
+static void ggml_vk_get_device_description(int device, char * description, size_t description_size) {
+    ggml_vk_instance_init();
+
+    std::vector<vk::PhysicalDevice> devices = vk_instance.instance.enumeratePhysicalDevices();
+
+    vk::PhysicalDeviceProperties props;
+    devices[device].getProperties(&props);
+
+    snprintf(description, description_size, "%s", props.deviceName.data());
+}
+
+// backend interface
+
+#define UNUSED GGML_UNUSED
+
+// device backend
+
+static const char * ggml_backend_vk_buffer_get_name(ggml_backend_buffer_t buffer) {
+    ggml_backend_vk_buffer_context * ctx = (ggml_backend_vk_buffer_context *)buffer->context;
+    return ctx->name.c_str();
+}
+
+static bool ggml_backend_buffer_is_vk(ggml_backend_buffer_t buffer) {
+    return buffer->iface.get_name == ggml_backend_vk_buffer_get_name;
+}
+
+static void ggml_backend_vk_buffer_free_buffer(ggml_backend_buffer_t buffer) {
+    VK_LOG_MEMORY("ggml_backend_vk_buffer_free_buffer()");
+    ggml_backend_vk_buffer_context * ctx = (ggml_backend_vk_buffer_context *)buffer->context;
+    ggml_vk_destroy_buffer(ctx->dev_buffer);
+    delete ctx;
+}
+
+static void * ggml_backend_vk_buffer_get_base(ggml_backend_buffer_t buffer) {
+    return vk_ptr_base;
+
+    UNUSED(buffer);
+}
+
+static void ggml_backend_vk_buffer_init_tensor(ggml_backend_buffer_t buffer, ggml_tensor * tensor) {
+    VK_LOG_DEBUG("ggml_backend_vk_buffer_init_tensor(" << buffer << " (" << buffer->context << "), " << tensor << ")");
+    if (tensor->view_src != nullptr) {
+        GGML_ASSERT(tensor->view_src->buffer->buft == buffer->buft);
+    }
+}
+
+static void ggml_backend_vk_buffer_set_tensor(ggml_backend_buffer_t buffer, ggml_tensor * tensor, const void * data, size_t offset, size_t size) {
+    VK_LOG_DEBUG("ggml_backend_vk_buffer_set_tensor(" << buffer << ", " << tensor << ", " << data << ", " << offset << ", " << size << ")");
+    ggml_backend_vk_buffer_context * buf_ctx = (ggml_backend_vk_buffer_context *)buffer->context;
+    vk_buffer buf = buf_ctx->dev_buffer;
+
+    ggml_vk_buffer_write(buf, vk_tensor_offset(tensor) + tensor->view_offs + offset, data, size);
+}
+
+static void ggml_backend_vk_buffer_get_tensor(ggml_backend_buffer_t buffer, const ggml_tensor * tensor, void * data, size_t offset, size_t size) {
+    VK_LOG_DEBUG("ggml_backend_vk_buffer_get_tensor(" << buffer << ", " << tensor << ", " << data << ", " << offset << ", " << size << ")");
+    ggml_backend_vk_buffer_context * buf_ctx = (ggml_backend_vk_buffer_context *)buffer->context;
+
+    vk_buffer buf = buf_ctx->dev_buffer;
+
+    ggml_vk_buffer_read(buf, vk_tensor_offset(tensor) + tensor->view_offs + offset, data, size);
+}
+
+static bool ggml_backend_vk_buffer_cpy_tensor(ggml_backend_buffer_t buffer, const ggml_tensor * src, ggml_tensor * dst) {
+    if (ggml_backend_buffer_is_vk(src->buffer)) {
+        ggml_backend_vk_buffer_context * src_buf_ctx = (ggml_backend_vk_buffer_context *)src->buffer->context;
+        ggml_backend_vk_buffer_context * dst_buf_ctx = (ggml_backend_vk_buffer_context *)dst->buffer->context;
+
+        vk_buffer src_buf = src_buf_ctx->dev_buffer;
+        vk_buffer dst_buf = dst_buf_ctx->dev_buffer;
+
+        ggml_vk_buffer_copy(dst_buf, vk_tensor_offset(dst) + dst->view_offs, src_buf, vk_tensor_offset(src) + src->view_offs, ggml_nbytes(src));
+
+        return true;
+    }
+    return false;
+
+    UNUSED(buffer);
+}
+
+static void ggml_backend_vk_buffer_clear(ggml_backend_buffer_t buffer, uint8_t value) {
+    ggml_backend_vk_buffer_context * ctx = (ggml_backend_vk_buffer_context *)buffer->context;
+
+    ggml_vk_buffer_memset(ctx->dev_buffer, 0, value, buffer->size);
+}
+
+static ggml_backend_buffer_i ggml_backend_vk_buffer_interface = {
+    /* .get_name        = */ ggml_backend_vk_buffer_get_name,
+    /* .free_buffer     = */ ggml_backend_vk_buffer_free_buffer,
+    /* .get_base        = */ ggml_backend_vk_buffer_get_base,
+    /* .init_tensor     = */ ggml_backend_vk_buffer_init_tensor,
+    /* .memset_tensor   = */ NULL,
+    /* .set_tensor      = */ ggml_backend_vk_buffer_set_tensor,
+    /* .get_tensor      = */ ggml_backend_vk_buffer_get_tensor,
+    /* .cpy_tensor      = */ ggml_backend_vk_buffer_cpy_tensor,
+    /* .clear           = */ ggml_backend_vk_buffer_clear,
+    /* .reset           = */ NULL,
+};
+
+// vk buffer type
+static const char * ggml_backend_vk_buffer_type_name(ggml_backend_buffer_type_t buft) {
+    ggml_backend_vk_buffer_type_context * ctx = (ggml_backend_vk_buffer_type_context *)buft->context;
+
+    return ctx->name.c_str();
+}
+
+static ggml_backend_buffer_t ggml_backend_vk_buffer_type_alloc_buffer(ggml_backend_buffer_type_t buft, size_t size) {
+    VK_LOG_MEMORY("ggml_backend_vk_buffer_type_alloc_buffer(" << size << ")");
+    ggml_backend_vk_buffer_type_context * ctx = (ggml_backend_vk_buffer_type_context *) buft->context;
+
+    vk_buffer dev_buffer = nullptr;
+    try {
+        dev_buffer = ggml_vk_create_buffer_device(ctx->device, size);
+    } catch (const vk::SystemError& e) {
+        return nullptr;
+    }
+
+    ggml_backend_vk_buffer_context * bufctx = new ggml_backend_vk_buffer_context(ctx->device, std::move(dev_buffer), ctx->name);
+
+    return ggml_backend_buffer_init(buft, ggml_backend_vk_buffer_interface, bufctx, size);
+}
+
+static size_t ggml_backend_vk_buffer_type_get_alignment(ggml_backend_buffer_type_t buft) {
+    ggml_backend_vk_buffer_type_context * ctx = (ggml_backend_vk_buffer_type_context *) buft->context;
+    return ctx->device->properties.limits.minStorageBufferOffsetAlignment;
+}
+
+static size_t ggml_backend_vk_buffer_type_get_max_size(ggml_backend_buffer_type_t buft) {
+    ggml_backend_vk_buffer_type_context * ctx = (ggml_backend_vk_buffer_type_context *) buft->context;
+    return ctx->device->max_memory_allocation_size;
+}
+
+static size_t ggml_backend_vk_buffer_type_get_alloc_size(ggml_backend_buffer_type_t buft, const ggml_tensor * tensor) {
+    return ggml_nbytes(tensor);
+
+    UNUSED(buft);
+}
+
+ggml_backend_buffer_type_t ggml_backend_vk_buffer_type(size_t dev_num) {
+    ggml_vk_instance_init();
+
+    VK_LOG_DEBUG("ggml_backend_vk_buffer_type(" << dev_num << ")");
+
+    vk_device dev = ggml_vk_get_device(dev_num);
+
+    return &dev->buffer_type;
+}
+
+// host buffer type
+
+static const char * ggml_backend_vk_host_buffer_type_name(ggml_backend_buffer_type_t buft) {
+    return GGML_VK_NAME "_Host";
+
+    UNUSED(buft);
+}
+
+static const char * ggml_backend_vk_host_buffer_name(ggml_backend_buffer_t buffer) {
+    return GGML_VK_NAME "_Host";
+
+    UNUSED(buffer);
+}
+
+static void ggml_backend_vk_host_buffer_free_buffer(ggml_backend_buffer_t buffer) {
+    VK_LOG_MEMORY("ggml_backend_vk_host_buffer_free_buffer()");
+    ggml_vk_host_free(vk_instance.devices[0], buffer->context);
+}
+
+static ggml_backend_buffer_t ggml_backend_vk_host_buffer_type_alloc_buffer(ggml_backend_buffer_type_t buft, size_t size) {
+    VK_LOG_MEMORY("ggml_backend_vk_host_buffer_type_alloc_buffer(" << size << ")");
+
+    size += 32;  // Behave like the CPU buffer type
+    void * ptr = nullptr;
+    try {
+        ptr = ggml_vk_host_malloc(vk_instance.devices[0], size);
+    } catch (vk::SystemError& e) {
+        std::cerr << "ggml_vulkan: Failed to allocate pinned memory." << std::endl;
+        std::cerr << "ggml_vulkan: " << e.what() << std::endl;
+        // fallback to cpu buffer
+        return ggml_backend_buft_alloc_buffer(ggml_backend_cpu_buffer_type(), size);
+    }
+
+    ggml_backend_buffer_t buffer = ggml_backend_cpu_buffer_from_ptr(ptr, size);
+    buffer->buft = buft;
+    buffer->iface.get_name = ggml_backend_vk_host_buffer_name;
+    buffer->iface.free_buffer = ggml_backend_vk_host_buffer_free_buffer;
+
+    return buffer;
+
+    UNUSED(buft);
+}
+
+static size_t ggml_backend_vk_host_buffer_type_get_alignment(ggml_backend_buffer_type_t buft) {
+    return vk_instance.devices[0]->properties.limits.minMemoryMapAlignment;
+
+    UNUSED(buft);
+}
+
+// Should be changed to return device-specific host buffer type
+// but that probably requires changes in llama.cpp
+ggml_backend_buffer_type_t ggml_backend_vk_host_buffer_type() {
+    static struct ggml_backend_buffer_type ggml_backend_vk_buffer_type_host = {
+        /* .iface    = */ {
+            /* .get_name         = */ ggml_backend_vk_host_buffer_type_name,
+            /* .alloc_buffer     = */ ggml_backend_vk_host_buffer_type_alloc_buffer,
+            /* .get_alignment    = */ ggml_backend_vk_host_buffer_type_get_alignment,
+            /* .get_max_size     = */ NULL, // defaults to SIZE_MAX
+            /* .get_alloc_size   = */ ggml_backend_cpu_buffer_type()->iface.get_alloc_size,
+            /* .is_host          = */ ggml_backend_cpu_buffer_type()->iface.is_host,
+        },
+        /* .device   = */ nullptr,
+        /* .context  = */ nullptr,
+    };
+
+    // Make sure device 0 is initialized
+    ggml_vk_instance_init();
+    ggml_vk_get_device(0);
+
+    return &ggml_backend_vk_buffer_type_host;
+}
+
+
+// backend
+
+static const char * ggml_backend_vk_name(ggml_backend_t backend) {
+    ggml_backend_vk_context * ctx = (ggml_backend_vk_context *)backend->context;
+
+    return ctx->name.c_str();
+}
+
+static void ggml_backend_vk_free(ggml_backend_t backend) {
+    ggml_backend_vk_context * ctx = (ggml_backend_vk_context *)backend->context;
+    VK_LOG_DEBUG("ggml_backend_vk_free(" << ctx->name << ")");
+
+    ggml_vk_cleanup(ctx);
+
+    delete ctx;
+    delete backend;
+}
+
+static ggml_backend_buffer_type_t ggml_backend_vk_get_default_buffer_type(ggml_backend_t backend) {
+    ggml_backend_vk_context * ctx = (ggml_backend_vk_context *)backend->context;
+
+    return &ctx->device->buffer_type;
+}
+
+static void ggml_backend_vk_set_tensor_async(ggml_backend_t backend, ggml_tensor * tensor, const void * data, size_t offset, size_t size) {
+    VK_LOG_DEBUG("ggml_backend_vk_set_tensor_async(" << size << ")");
+    ggml_backend_vk_context * ctx = (ggml_backend_vk_context *)backend->context;
+    GGML_ASSERT((tensor->buffer->buft == ggml_backend_vk_get_default_buffer_type(backend) || tensor->buffer->buft == ggml_backend_vk_host_buffer_type()) && "unsupported buffer type");
+
+    ggml_backend_vk_buffer_context * buf_ctx = (ggml_backend_vk_buffer_context *)tensor->buffer->context;
+
+    vk_context transfer_ctx;
+
+    if (ctx->transfer_ctx.expired()) {
+        // Initialize new transfer context
+        transfer_ctx = ggml_vk_create_context(ctx, ctx->device->transfer_queue);
+        ctx->transfer_ctx = transfer_ctx;
+        ggml_vk_ctx_begin(ctx->device, transfer_ctx);
+    } else {
+        transfer_ctx = ctx->transfer_ctx.lock();
+    }
+
+    vk_buffer buf = buf_ctx->dev_buffer;
+
+    ggml_vk_buffer_write_async(transfer_ctx, buf, vk_tensor_offset(tensor) + tensor->view_offs + offset, data, size);
+}
+
+static void ggml_backend_vk_get_tensor_async(ggml_backend_t backend, const ggml_tensor * tensor, void * data, size_t offset, size_t size) {
+    VK_LOG_DEBUG("ggml_backend_vk_get_tensor_async(" << size << ")");
+    ggml_backend_vk_context * ctx = (ggml_backend_vk_context *)backend->context;
+    GGML_ASSERT((tensor->buffer->buft == ggml_backend_vk_get_default_buffer_type(backend) || tensor->buffer->buft == ggml_backend_vk_host_buffer_type()) && "unsupported buffer type");
+
+    ggml_backend_vk_buffer_context * buf_ctx = (ggml_backend_vk_buffer_context *)tensor->buffer->context;
+
+    vk_context transfer_ctx;
+
+    if (ctx->transfer_ctx.expired()) {
+        // Initialize new transfer context
+        transfer_ctx = ggml_vk_create_context(ctx, ctx->device->transfer_queue);
+        ctx->transfer_ctx = transfer_ctx;
+        ggml_vk_ctx_begin(ctx->device, transfer_ctx);
+    } else {
+        transfer_ctx = ctx->transfer_ctx.lock();
+    }
+
+    vk_buffer buf = buf_ctx->dev_buffer;
+
+    ggml_vk_buffer_read_async(transfer_ctx, buf, vk_tensor_offset(tensor) + tensor->view_offs + offset, data, size);
+}
+
+static bool ggml_backend_vk_cpy_tensor_async(ggml_backend_t backend, const ggml_tensor * src, ggml_tensor * dst) {
+    VK_LOG_DEBUG("ggml_backend_vk_cpy_tensor_async()");
+    ggml_backend_vk_context * ctx = (ggml_backend_vk_context *)backend->context;
+    if ((dst->buffer->buft == ggml_backend_vk_get_default_buffer_type(backend) || dst->buffer->buft == ggml_backend_vk_host_buffer_type()) && ggml_backend_buffer_is_vk(src->buffer)) {
+        ggml_backend_vk_buffer_context * src_buf_ctx = (ggml_backend_vk_buffer_context *)src->buffer->context;
+        ggml_backend_vk_buffer_context * dst_buf_ctx = (ggml_backend_vk_buffer_context *)dst->buffer->context;
+
+        vk_context transfer_ctx;
+
+        if (ctx->transfer_ctx.expired()) {
+            // Initialize new transfer context
+            transfer_ctx = ggml_vk_create_context(ctx, ctx->device->transfer_queue);
+            ctx->transfer_ctx = transfer_ctx;
+            ggml_vk_ctx_begin(ctx->device, transfer_ctx);
+        } else {
+            transfer_ctx = ctx->transfer_ctx.lock();
+        }
+
+        vk_buffer src_buf = src_buf_ctx->dev_buffer;
+        vk_buffer dst_buf = dst_buf_ctx->dev_buffer;
+
+        ggml_vk_buffer_copy_async(transfer_ctx, dst_buf, vk_tensor_offset(dst) + dst->view_offs, src_buf, vk_tensor_offset(src) + src->view_offs, ggml_nbytes(src));
+        return true;
+    }
+
+    return false;
+}
+
+static void ggml_backend_vk_synchronize(ggml_backend_t backend) {
+    VK_LOG_DEBUG("ggml_backend_vk_synchronize()");
+    ggml_backend_vk_context * ctx = (ggml_backend_vk_context *)backend->context;
+    if(ctx->transfer_ctx.expired()) {
+        return;
+    }
+
+    vk_context transfer_ctx = ctx->transfer_ctx.lock();
+
+    ggml_vk_ctx_end(transfer_ctx);
+
+    for (auto& cpy : transfer_ctx->in_memcpys) {
+        memcpy(cpy.dst, cpy.src, cpy.n);
+    }
+
+    ggml_vk_submit(transfer_ctx, ctx->fence);
+    VK_CHECK(ctx->device->device.waitForFences({ ctx->fence }, true, UINT64_MAX), "ggml_backend_vk_synchronize waitForFences");
+    ctx->device->device.resetFences({ ctx->fence });
+
+    for (auto& cpy : transfer_ctx->out_memcpys) {
+        memcpy(cpy.dst, cpy.src, cpy.n);
+    }
+
+    ctx->transfer_ctx.reset();
+}
+
+static bool ggml_vk_is_empty(ggml_tensor * node) {
+    return ggml_is_empty(node) || node->op == GGML_OP_NONE || node->op == GGML_OP_RESHAPE || node->op == GGML_OP_TRANSPOSE || node->op == GGML_OP_VIEW || node->op == GGML_OP_PERMUTE;
+}
+
+static ggml_status ggml_backend_vk_graph_compute(ggml_backend_t backend, ggml_cgraph * cgraph) {
+    VK_LOG_DEBUG("ggml_backend_vk_graph_compute(" << cgraph->n_nodes << " nodes)");
+    ggml_backend_vk_context * ctx = (ggml_backend_vk_context *)backend->context;
+
+    for (int i = 0; i < cgraph->n_nodes; i++) {
+        ggml_vk_build_graph(ctx, cgraph->nodes[i], i, nullptr, 0, true, false, false);
+    }
+    ggml_vk_preallocate_buffers(ctx);
+    ggml_pipeline_allocate_descriptor_sets(ctx->device);
+
+    int last_node = cgraph->n_nodes - 1;
+
+    // If the last op in the cgraph isn't backend GPU, the command buffer doesn't get closed properly
+    while (last_node > 0 && ggml_vk_is_empty(cgraph->nodes[last_node])) {
+        last_node -= 1;
+    }
+
+    // Reserve tensor context space for all nodes
+    ctx->tensor_ctxs.resize(cgraph->n_nodes);
+
+    bool first_node_in_batch = true; // true if next node will be first node in a batch
+    int submit_node_idx = 0; // index to first node in a batch
+
+    // submit work every submit_count node to overlap CPU cmdbuffer generation with GPU execution
+    constexpr int submit_count = 100;
+    int submitted_nodes = 0;
+    for (int i = 0; i < cgraph->n_nodes; i++) {
+        if (first_node_in_batch) {
+            submit_node_idx = i;
+        }
+
+        bool submit = (submitted_nodes >= submit_count) || (i == last_node);
+
+
+        bool enqueued = ggml_vk_build_graph(ctx, cgraph->nodes[i], i, cgraph->nodes[submit_node_idx], submit_node_idx, false, i == last_node, submit);
+
+        if (enqueued) {
+            ++submitted_nodes;
+
+#ifndef GGML_VULKAN_CHECK_RESULTS
+            if (first_node_in_batch) {
+                first_node_in_batch = false;
+            }
+#endif
+        }
+
+        if (submit) {
+            first_node_in_batch = true;
+            submitted_nodes = 0;
+        }
+    }
+
+#ifdef GGML_VULKAN_PERF
+    ctx->device->perf_logger->print_timings();
+#endif
+
+    ggml_vk_graph_cleanup(ctx);
+
+    return GGML_STATUS_SUCCESS;
+
+    UNUSED(backend);
+}
+
+static bool ggml_backend_vk_supports_op(ggml_backend_t backend, const ggml_tensor * op) {
+    // ggml_backend_vk_context * ctx = (ggml_backend_vk_context *) backend->context;
+
+    switch (op->op) {
+        case GGML_OP_UNARY:
+            switch (ggml_get_unary_op(op)) {
+                case GGML_UNARY_OP_GELU:
+                case GGML_UNARY_OP_GELU_QUICK:
+                case GGML_UNARY_OP_SILU:
+                case GGML_UNARY_OP_RELU:
+                case GGML_UNARY_OP_TANH:
+                    return ggml_is_contiguous(op->src[0]);
+                default:
+                    return false;
+            }
+            break;
+        case GGML_OP_MUL_MAT:
+        case GGML_OP_MUL_MAT_ID:
+            {
+                switch (op->src[0]->type) {
+                    case GGML_TYPE_F32:
+                    case GGML_TYPE_F16:
+                    case GGML_TYPE_Q4_0:
+                    case GGML_TYPE_Q4_1:
+                    case GGML_TYPE_Q5_0:
+                    case GGML_TYPE_Q5_1:
+                    case GGML_TYPE_Q8_0:
+                    case GGML_TYPE_Q2_K:
+                    case GGML_TYPE_Q3_K:
+                    case GGML_TYPE_Q4_K:
+                    case GGML_TYPE_Q5_K:
+                    case GGML_TYPE_Q6_K:
+                    case GGML_TYPE_IQ4_NL:
+                        break;
+                    default:
+                        return false;
+                }
+                struct ggml_tensor * a;
+                struct ggml_tensor * b;
+                if (op->op == GGML_OP_MUL_MAT) {
+                    a = op->src[0];
+                    b = op->src[1];
+                } else {
+                    a = op->src[2];
+                    b = op->src[1];
+                }
+                if (a->ne[3] != b->ne[3]) {
+                    return false;
+                }
+                return true;
+            } break;
+        case GGML_OP_GET_ROWS:
+            {
+                switch (op->src[0]->type) {
+                    case GGML_TYPE_F32:
+                    case GGML_TYPE_F16:
+                    case GGML_TYPE_Q4_0:
+                    case GGML_TYPE_Q4_1:
+                    case GGML_TYPE_Q5_0:
+                    case GGML_TYPE_Q5_1:
+                    case GGML_TYPE_Q8_0:
+                    case GGML_TYPE_IQ4_NL:
+                        return true;
+                    default:
+                        return false;
+                }
+            } break;
+        case GGML_OP_CONT:
+        case GGML_OP_CPY:
+        case GGML_OP_DUP:
+            {
+                ggml_type src0_type = op->src[0]->type;
+                ggml_type src1_type = op->src[1] != nullptr ? op->src[1]->type : src0_type;
+                if (src0_type == GGML_TYPE_F32 && src1_type == GGML_TYPE_F32) {
+                    return true;
+                }
+                if (src0_type == GGML_TYPE_F32 && src1_type == GGML_TYPE_F16) {
+                    return true;
+                }
+                if (src0_type == GGML_TYPE_F16 && src1_type == GGML_TYPE_F16) {
+                    return true;
+                }
+                return false;
+            } break;
+        case GGML_OP_REPEAT:
+            return ggml_type_size(op->type) == sizeof(float) && ggml_type_size(op->src[0]->type) == sizeof(float);
+        case GGML_OP_ROPE:
+            return ggml_is_contiguous(op->src[0]);
+        case GGML_OP_NONE:
+        case GGML_OP_RESHAPE:
+        case GGML_OP_VIEW:
+        case GGML_OP_PERMUTE:
+        case GGML_OP_TRANSPOSE:
+        case GGML_OP_NORM:
+        case GGML_OP_GROUP_NORM:
+        case GGML_OP_RMS_NORM:
+        case GGML_OP_ADD:
+        case GGML_OP_ACC:
+        case GGML_OP_MUL:
+        case GGML_OP_DIV:
+        case GGML_OP_CONCAT:
+        case GGML_OP_UPSCALE:
+        case GGML_OP_SCALE:
+        case GGML_OP_SQR:
+        case GGML_OP_SIN:
+        case GGML_OP_COS:
+        case GGML_OP_CLAMP:
+        case GGML_OP_PAD:
+        case GGML_OP_DIAG_MASK_INF:
+        case GGML_OP_SOFT_MAX:
+        case GGML_OP_ARGSORT:
+        case GGML_OP_SUM_ROWS:
+        case GGML_OP_IM2COL:
+        case GGML_OP_TIMESTEP_EMBEDDING:
+        case GGML_OP_LEAKY_RELU:
+            return true;
+        default:
+            return false;
+    }
+
+    UNUSED(backend);
+}
+
+static bool ggml_backend_vk_offload_op(ggml_backend_t backend, const ggml_tensor * op) {
+    const int min_batch_size = 32;
+
+    return (op->ne[1] >= min_batch_size && op->op != GGML_OP_GET_ROWS) ||
+           (op->ne[2] >= min_batch_size && op->op == GGML_OP_MUL_MAT_ID);
+
+    UNUSED(backend);
+}
+
+static bool ggml_backend_vk_supports_buft(ggml_backend_t backend, ggml_backend_buffer_type_t buft) {
+    if (buft->iface.get_name != ggml_backend_vk_buffer_type_name) {
+        return false;
+    }
+
+    ggml_backend_vk_buffer_type_context * buft_ctx = (ggml_backend_vk_buffer_type_context *)buft->context;
+    ggml_backend_vk_context * ctx = (ggml_backend_vk_context *)backend->context;
+
+    return buft_ctx->device == ctx->device;
+}
+
+// TODO: enable async and synchronize
+static ggml_backend_i ggml_backend_vk_interface = {
+    /* .get_name                = */ ggml_backend_vk_name,
+    /* .free                    = */ ggml_backend_vk_free,
+    /* .get_default_buffer_type = */ ggml_backend_vk_get_default_buffer_type,
+    /* .set_tensor_async        = */ NULL,  // ggml_backend_vk_set_tensor_async,
+    /* .get_tensor_async        = */ NULL,  // ggml_backend_vk_get_tensor_async,
+    /* .cpy_tensor_async        = */ NULL,  // ggml_backend_vk_cpy_tensor_async,
+    /* .synchronize             = */ NULL,  // ggml_backend_vk_synchronize,
+    /* .graph_plan_create       = */ NULL,
+    /* .graph_plan_free         = */ NULL,
+    /* .graph_plan_update       = */ NULL,
+    /* .graph_plan_compute      = */ NULL,
+    /* .graph_compute           = */ ggml_backend_vk_graph_compute,
+    /* .supports_op             = */ ggml_backend_vk_supports_op,
+    /* .supports_buft           = */ ggml_backend_vk_supports_buft,
+    /* .offload_op              = */ ggml_backend_vk_offload_op,
+    /* .event_record            = */ NULL,
+    /* .event_wait              = */ NULL,
+};
+
+static ggml_guid_t ggml_backend_vk_guid() {
+    static ggml_guid guid = { 0xb8, 0xf7, 0x4f, 0x86, 0x40, 0x3c, 0xe1, 0x02, 0x91, 0xc8, 0xdd, 0xe9, 0x02, 0x3f, 0xc0, 0x2b };
+    return &guid;
+}
+
+ggml_backend_t ggml_backend_vk_init(size_t dev_num) {
+    VK_LOG_DEBUG("ggml_backend_vk_init(" << dev_num << ")");
+
+    ggml_backend_vk_context * ctx = new ggml_backend_vk_context;
+    ggml_vk_init(ctx, dev_num);
+
+    ggml_backend_t vk_backend = new ggml_backend {
+        /* .guid      = */ ggml_backend_vk_guid(),
+        /* .interface = */ ggml_backend_vk_interface,
+        /* .device    = */ nullptr,
+        /* .context   = */ ctx,
+    };
+
+    return vk_backend;
+}
+
+bool ggml_backend_is_vk(ggml_backend_t backend) {
+    return backend != NULL && ggml_guid_matches(backend->guid, ggml_backend_vk_guid());
+}
+
+int ggml_backend_vk_get_device_count() {
+    return ggml_vk_get_device_count();
+}
+
+void ggml_backend_vk_get_device_description(int device, char * description, size_t description_size) {
+    ggml_vk_get_device_description(device, description, description_size);
+}
+
+void ggml_backend_vk_get_device_memory(int device, size_t * free, size_t * total) {
+    GGML_ASSERT(device < (int) vk_instance.device_indices.size());
+
+    vk::PhysicalDevice vkdev = vk_instance.instance.enumeratePhysicalDevices()[vk_instance.device_indices[device]];
+
+    vk::PhysicalDeviceMemoryProperties memprops = vkdev.getMemoryProperties();
+
+    for (const vk::MemoryHeap& heap : memprops.memoryHeaps) {
+        if (heap.flags & vk::MemoryHeapFlagBits::eDeviceLocal) {
+            *total = heap.size;
+            *free = heap.size;
+            break;
+        }
+    }
+}
+
+// Extension availability
+static bool ggml_vk_instance_validation_ext_available(const std::vector<vk::ExtensionProperties>& instance_extensions) {
+#ifdef GGML_VULKAN_VALIDATE
+    bool portability_enumeration_ext = false;
+    // Check for portability enumeration extension for MoltenVK support
+    for (const auto& properties : instance_extensions) {
+        if (strcmp("VK_KHR_portability_enumeration", properties.extensionName) == 0) {
+            return true;
+        }
+    }
+    if (!portability_enumeration_ext) {
+        std::cerr << "ggml_vulkan: WARNING: Instance extension VK_KHR_portability_enumeration not found." << std::endl;
+    }
+#endif
+    return false;
+
+    UNUSED(instance_extensions);
+}
+static bool ggml_vk_instance_portability_enumeration_ext_available(const std::vector<vk::ExtensionProperties>& instance_extensions) {
+#ifdef __APPLE__
+    bool portability_enumeration_ext = false;
+    // Check for portability enumeration extension for MoltenVK support
+    for (const auto& properties : instance_extensions) {
+        if (strcmp("VK_KHR_portability_enumeration", properties.extensionName) == 0) {
+            return true;
+        }
+    }
+    if (!portability_enumeration_ext) {
+        std::cerr << "ggml_vulkan: WARNING: Instance extension VK_KHR_portability_enumeration not found." << std::endl;
+    }
+#endif
+    return false;
+
+    UNUSED(instance_extensions);
+}
+
+// checks
+
+#ifdef GGML_VULKAN_CHECK_RESULTS
+static void ggml_vk_print_graph_origin(const ggml_tensor * tensor, std::vector<const ggml_tensor *>& done, int level = 0) {
+    if (std::find(done.begin(), done.end(), tensor) != done.end() || level > 10) {
+        return;
+    }
+    for (int j = 0; j < level; j++) {
+        std::cerr << " ";
+    }
+    std::cerr << ggml_op_name(tensor->op) << " gpu=" << (tensor->extra != nullptr) << std::endl;
+
+    done.push_back(tensor);
+
+    for (int i = 0; i < GGML_MAX_SRC; i++) {
+        if (tensor->src[i] != nullptr) {
+            ggml_vk_print_graph_origin(tensor->src[i], done, level + 1);
+        }
+    }
+}
+
+static void ggml_vk_print_tensor_area(const ggml_tensor * tensor, const void * data, int i0, int i1, int i2, int i3) {
+    if (tensor->type != GGML_TYPE_F32 && tensor->type != GGML_TYPE_F16 && tensor->type != GGML_TYPE_I32) {
+        return;
+    }
+    i0 = std::max(i0, 5);
+    i1 = std::max(i1, 5);
+    i2 = std::max(i2, 0);
+    i3 = std::max(i3, 0);
+    fprintf(stderr, "         ");
+    for (int idx1 = i1 - 5; idx1 < i1 + 5; idx1++) {
+        fprintf(stderr, "%7d ", idx1);
+    }
+    fprintf(stderr, "\n");
+    for (int idx0 = i0 - 5; idx0 < i0 + 5; idx0++) {
+        fprintf(stderr, "%7d: ", idx0);
+        for (int idx1 = i1 - 5; idx1 < i1 + 5; idx1++) {
+            if (idx0 >= 0 && idx0 < tensor->ne[0] && idx1 >= 0 && idx1 < tensor->ne[1] && i2 >= 0 && i2 < tensor->ne[2] && i3 >= 0 && i3 < tensor->ne[3]) {
+                float val;
+                if (tensor->type == GGML_TYPE_F32) {
+                    val = *(const float *) ((const char *) data + i3*tensor->nb[3] + i2*tensor->nb[2] + idx1*tensor->nb[1] + idx0*tensor->nb[0]);
+                } else if (tensor->type == GGML_TYPE_F16) {
+                    val = ggml_fp16_to_fp32(*(const ggml_fp16_t *) ((const char *) data + i3*tensor->nb[3] + i2*tensor->nb[2] + idx1*tensor->nb[1] + idx0*tensor->nb[0]));
+                } else if (tensor->type == GGML_TYPE_I32) {
+                    val = *(const int32_t *) ((const char *) data + i3*tensor->nb[3] + i2*tensor->nb[2] + idx1*tensor->nb[1] + idx0*tensor->nb[0]);
+                } else {
+                    GGML_ABORT("fatal error");
+                }
+                fprintf(stderr, "% 7.2f ", val);
+            } else {
+                fprintf(stderr, "        ");
+            }
+        }
+        fprintf(stderr, "\n");
+    }
+}
+
+static void ggml_vk_print_tensor(const ggml_tensor * tensor, const char * name) {
+    void * tensor_data = tensor->data;
+
+    const bool is_gpu = tensor->buffer != nullptr && ggml_backend_buffer_is_vk(tensor->buffer);
+
+    if (is_gpu) {
+        const size_t tensor_size = ggml_nbytes(tensor);
+        tensor_data = malloc(tensor_size);
+
+        ggml_backend_vk_buffer_context * buf_ctx = (ggml_backend_vk_buffer_context *)tensor->buffer->context;
+
+        vk_buffer buffer_gpu = buf_ctx->dev_buffer;
+        ggml_vk_buffer_read(buffer_gpu, vk_tensor_offset(tensor) + tensor->view_offs, tensor_data, tensor_size);
+    }
+
+    std::cerr << "TENSOR CHECK " << name << " (" << tensor->name << "): " << ggml_op_name(tensor->op) << std::endl;
+    std::cerr << "tensor=" << tensor << " tensor->type: " << ggml_type_name(tensor->type) << " ne0=" << tensor->ne[0] << " nb0=" << tensor->nb[0] << " ne1=" << tensor->ne[1] << " nb1=" << tensor->nb[1] << " ne2=" << tensor->ne[2] << " nb2=" << tensor->nb[2] << " ne3=" << tensor->ne[3] << " nb3=" << tensor->nb[3] << std::endl;
+    if (tensor->src[0] != nullptr) {
+        std::cerr << "tensor->src[0]=" << tensor->src[0] << " name=" << tensor->src[0]->name << " op=" << ggml_op_name(tensor->src[0]->op) << " type=" << ggml_type_name(tensor->src[0]->type) << " ne0=" << tensor->src[0]->ne[0] << " nb0=" << tensor->src[0]->nb[0] << " ne1=" << tensor->src[0]->ne[1] << " nb1=" << tensor->src[0]->nb[1] << " ne2=" << tensor->src[0]->ne[2] << " nb2=" << tensor->src[0]->nb[2] << " ne3=" << tensor->src[0]->ne[3] << " nb3=" << tensor->src[0]->nb[3] << std::endl;
+    }
+    if (tensor->src[1] != nullptr) {
+        std::cerr << "tensor->src[1]=" << tensor->src[1] << " name=" << tensor->src[1]->name << " op=" << ggml_op_name(tensor->src[1]->op) << " type=" << ggml_type_name(tensor->src[1]->type) << " ne0=" << tensor->src[1]->ne[0] << " nb0=" << tensor->src[1]->nb[0] << " ne1=" << tensor->src[1]->ne[1] << " nb1=" << tensor->src[1]->nb[1] << " ne2=" << tensor->src[1]->ne[2] << " nb2=" << tensor->src[1]->nb[2] << " ne3=" << tensor->src[1]->ne[3] << " nb3=" << tensor->src[1]->nb[3] << std::endl;
+    }
+    std::cerr << std::endl << "Result:" << std::endl;
+    ggml_vk_print_tensor_area(tensor, tensor_data, 5, 5, 0, 0);
+    std::cerr << std::endl;
+    std::vector<const ggml_tensor *> done;
+    ggml_vk_print_graph_origin(tensor, done);
+
+    if (is_gpu) {
+        free(tensor_data);
+    }
+}
+
+void * comp_result;
+size_t comp_size;
+size_t comp_nb[GGML_MAX_DIMS];
+size_t check_counter = 0;
+static void ggml_vk_check_results_0(ggml_tensor * tensor) {
+    if (tensor->op == GGML_OP_TRANSPOSE) {
+        return;
+    }
+
+    check_counter++;
+    if (!(vk_output_tensor > 0 && vk_output_tensor == check_counter) && check_counter <= vk_skip_checks) {
+        return;
+    }
+
+    VK_LOG_DEBUG("ggml_vk_check_results_0(" << tensor->name << ")");
+
+    ggml_tensor * src0 = tensor->src[0];
+    ggml_tensor * src1 = tensor->src[1];
+    ggml_tensor * src2 = tensor->src[2];
+
+    struct ggml_init_params iparams = {
+        /*.mem_size   =*/ 2ul*1024ul*1024ul*1024ul,
+        /*.mem_buffer =*/ NULL,
+        /*.no_alloc   =*/ false,
+    };
+
+    struct ggml_context * ggml_ctx = ggml_init(iparams);
+
+    struct ggml_tensor * src0_clone = nullptr;
+    struct ggml_tensor * src1_clone = nullptr;
+    struct ggml_tensor * src2_clone = nullptr;
+    struct ggml_tensor * tensor_clone = nullptr;
+
+    size_t src0_size;
+    size_t src1_size;
+    size_t src2_size;
+
+    void * src0_buffer = nullptr;
+    void * src1_buffer = nullptr;
+    void * src2_buffer = nullptr;
+
+    if (src0 != nullptr) {
+        src0_clone = ggml_dup_tensor(ggml_ctx, src0);
+
+        src0_size = ggml_nbytes(src0);
+
+        src0_buffer = malloc(src0_size);
+        src0_clone->data = src0_buffer;
+        if (ggml_backend_buffer_is_host(src0->buffer)) {
+            memcpy(src0_clone->data, src0->data, src0_size);
+            memcpy(src0_clone->nb, src0->nb, sizeof(size_t) * GGML_MAX_DIMS);
+        } else if (ggml_backend_buffer_is_vk(src0->buffer)) {
+            ggml_backend_vk_buffer_context * buf_ctx = (ggml_backend_vk_buffer_context *)src0->buffer->context;
+            vk_buffer& buffer_gpu = buf_ctx->dev_buffer;
+            uint64_t offset = vk_tensor_offset(src0) + src0->view_offs;
+            if (!ggml_is_contiguous(src0) && ggml_vk_dim01_contiguous(src0)) {
+                for (int i3 = 0; i3 < src0->ne[3]; i3++) {
+                    for (int i2 = 0; i2 < src0->ne[2]; i2++) {
+                        const int idx = i3*src0->ne[2] + i2;
+                        ggml_vk_buffer_read(buffer_gpu, offset + idx * src0->nb[2], ((char *)src0_clone->data + idx * src0_clone->nb[2]), src0->ne[1] * src0->nb[1]);
+                    }
+                }
+
+                src0_clone->nb[0] = src0->nb[0];
+                src0_clone->nb[1] = src0->nb[1];
+                for (int i = 2; i < GGML_MAX_DIMS; i++) {
+                    src0_clone->nb[i] = src0_clone->nb[i - 1]*src0_clone->ne[i - 1];
+                }
+            } else {
+                if (offset + src0_size >= buffer_gpu->size) {
+                    src0_size = buffer_gpu->size - offset;
+                }
+                ggml_vk_buffer_read(buffer_gpu, offset, src0_clone->data, src0_size);
+                memcpy(src0_clone->nb, src0->nb, sizeof(size_t) * GGML_MAX_DIMS);
+            }
+        } else {
+            GGML_ABORT("fatal error");
+        }
+
+        if (vk_output_tensor > 0 && vk_output_tensor == check_counter) {
+            ggml_vk_print_tensor(src0, "src0");
+        }
+    }
+    if (src1 != nullptr) {
+        src1_clone = ggml_dup_tensor(ggml_ctx, src1);
+
+        src1_size = ggml_nbytes(src1);
+
+        src1_buffer = malloc(src1_size);
+        src1_clone->data = src1_buffer;
+        if (ggml_backend_buffer_is_host(src1->buffer)) {
+            memcpy(src1_clone->data, src1->data, src1_size);
+            memcpy(src1_clone->nb, src1->nb, sizeof(size_t) * GGML_MAX_DIMS);
+        } else if (ggml_backend_buffer_is_vk(src1->buffer)) {
+            ggml_backend_vk_buffer_context * buf_ctx = (ggml_backend_vk_buffer_context *)src1->buffer->context;
+            vk_buffer& buffer_gpu = buf_ctx->dev_buffer;
+            uint64_t offset = vk_tensor_offset(src1) + src1->view_offs;
+            if (!ggml_is_contiguous(src1) && ggml_vk_dim01_contiguous(src1)) {
+                for (int i3 = 0; i3 < src1->ne[3]; i3++) {
+                    for (int i2 = 0; i2 < src1->ne[2]; i2++) {
+                        const int idx = i3*src1->ne[2] + i2;
+                        ggml_vk_buffer_read(buffer_gpu, offset + idx * src1->nb[2], ((char *)src1_clone->data + idx * src1_clone->nb[2]), src1->ne[1] * src1->nb[1]);
+                    }
+                }
+
+                src1_clone->nb[0] = src1->nb[0];
+                src1_clone->nb[1] = src1->nb[1];
+                for (int i = 2; i < GGML_MAX_DIMS; i++) {
+                    src1_clone->nb[i] = src1_clone->nb[i - 1]*src1_clone->ne[i - 1];
+                }
+            } else {
+                if (offset + src1_size >= buffer_gpu->size) {
+                    src1_size = buffer_gpu->size - offset;
+                }
+                ggml_vk_buffer_read(buffer_gpu, offset, src1_clone->data, src1_size);
+                memcpy(src1_clone->nb, src1->nb, sizeof(size_t) * GGML_MAX_DIMS);
+            }
+        } else {
+            GGML_ABORT("fatal error");
+        }
+
+        if (vk_output_tensor > 0 && vk_output_tensor == check_counter) {
+            ggml_vk_print_tensor(src1, "src1");
+        }
+    }
+    if (src2 != nullptr) {
+        src2_clone = ggml_dup_tensor(ggml_ctx, src2);
+
+        src2_size = ggml_nbytes(src2);
+
+        src2_buffer = malloc(src2_size);
+        src2_clone->data = src2_buffer;
+        if (ggml_backend_buffer_is_host(src2->buffer)) {
+            memcpy(src2_clone->data, src2->data, src2_size);
+            memcpy(src2_clone->nb, src2->nb, sizeof(size_t) * GGML_MAX_DIMS);
+        } else if (ggml_backend_buffer_is_vk(src2->buffer)) {
+            ggml_backend_vk_buffer_context * buf_ctx = (ggml_backend_vk_buffer_context *)src2->buffer->context;
+            vk_buffer& buffer_gpu = buf_ctx->dev_buffer;
+            uint64_t offset = vk_tensor_offset(src2) + src2->view_offs;
+            if (!ggml_is_contiguous(src2) && ggml_vk_dim01_contiguous(src2)) {
+                for (int i3 = 0; i3 < src2->ne[3]; i3++) {
+                    for (int i2 = 0; i2 < src2->ne[2]; i2++) {
+                        const int idx = i3*src2->ne[2] + i2;
+                        ggml_vk_buffer_read(buffer_gpu, offset + idx * src2->nb[2], ((char *)src2_clone->data + idx * src2_clone->nb[2]), src2->ne[1] * src2->nb[1]);
+                    }
+                }
+
+                src2_clone->nb[0] = src2->nb[0];
+                src2_clone->nb[1] = src2->nb[1];
+                for (int i = 2; i < GGML_MAX_DIMS; i++) {
+                    src2_clone->nb[i] = src2_clone->nb[i - 1]*src2_clone->ne[i - 1];
+                }
+            } else {
+                if (offset + src2_size >= buffer_gpu->size) {
+                    src2_size = buffer_gpu->size - offset;
+                }
+                ggml_vk_buffer_read(buffer_gpu, offset, src2_clone->data, src2_size);
+                memcpy(src2_clone->nb, src2->nb, sizeof(size_t) * GGML_MAX_DIMS);
+            }
+        } else {
+            GGML_ABORT("fatal error");
+        }
+
+        if (vk_output_tensor > 0 && vk_output_tensor == check_counter) {
+            ggml_vk_print_tensor(src2, "src2");
+        }
+    }
+
+    if (tensor->op == GGML_OP_MUL_MAT) {
+        tensor_clone = ggml_mul_mat(ggml_ctx, src0_clone, src1_clone);
+    } else if (tensor->op == GGML_OP_MUL_MAT_ID) {
+        tensor_clone = ggml_mul_mat_id(ggml_ctx, src0_clone, src1_clone, src2_clone);
+    } else if (tensor->op == GGML_OP_MUL) {
+        tensor_clone = ggml_mul(ggml_ctx, src0_clone, src1_clone);
+    } else if (tensor->op == GGML_OP_DIV) {
+        tensor_clone = ggml_div(ggml_ctx, src0_clone, src1_clone);
+    } else if (tensor->op == GGML_OP_CONCAT) {
+        tensor_clone = ggml_concat(ggml_ctx, src0_clone, src1_clone, *(int *)tensor->op_params);
+    } else if (tensor->op == GGML_OP_UPSCALE) {
+        tensor_clone = ggml_upscale_ext(ggml_ctx, src0_clone, tensor->ne[0], tensor->ne[1], tensor->ne[2], tensor->ne[3]);
+    } else if (tensor->op == GGML_OP_SCALE) {
+        tensor_clone = ggml_scale(ggml_ctx, src0_clone, ((float *)tensor->op_params)[0]);
+    } else if (tensor->op == GGML_OP_SQR) {
+        tensor_clone = ggml_sqr(ggml_ctx, src0_clone);
+    } else if (tensor->op == GGML_OP_SIN) {
+        tensor_clone = ggml_sin(ggml_ctx, src0_clone);
+    } else if (tensor->op == GGML_OP_COS) {
+        tensor_clone = ggml_cos(ggml_ctx, src0_clone);
+    } else if (tensor->op == GGML_OP_CLAMP) {
+        tensor_clone = ggml_clamp(ggml_ctx, src0_clone, ((float *)tensor->op_params)[0], ((float *)tensor->op_params)[1]);
+    } else if (tensor->op == GGML_OP_PAD) {
+        tensor_clone = ggml_pad(ggml_ctx, src0_clone, tensor->ne[0] - src0_clone->ne[0], tensor->ne[1] - src0_clone->ne[1], tensor->ne[2] - src0_clone->ne[2], tensor->ne[3] - src0_clone->ne[3]);
+    } else if (tensor->op == GGML_OP_REPEAT) {
+        tensor_clone = ggml_repeat(ggml_ctx, src0_clone, tensor);
+    } else if (tensor->op == GGML_OP_ADD) {
+        tensor_clone = ggml_add(ggml_ctx, src0_clone, src1_clone);
+    } else if (tensor->op == GGML_OP_ACC) {
+        tensor_clone = ggml_acc(ggml_ctx, src0_clone, src1_clone, tensor->op_params[0], tensor->op_params[1], tensor->op_params[2], tensor->op_params[3]);
+    } else if (tensor->op == GGML_OP_NORM) {
+        tensor_clone = ggml_norm(ggml_ctx, src0_clone, *(float *)tensor->op_params);
+    } else if (tensor->op == GGML_OP_GROUP_NORM) {
+        tensor_clone = ggml_group_norm(ggml_ctx, src0_clone, *(int *)tensor->op_params, ((float *)tensor->op_params)[1]);
+    } else if (tensor->op == GGML_OP_RMS_NORM) {
+        tensor_clone = ggml_rms_norm(ggml_ctx, src0_clone, *(float *)tensor->op_params);
+    } else if (tensor->op == GGML_OP_SOFT_MAX) {
+        if (src1 != nullptr) {
+            tensor_clone = ggml_soft_max_ext(ggml_ctx, src0_clone, src1_clone, ((float *)tensor->op_params)[0], ((float *)tensor->op_params)[1]);
+        } else {
+            tensor_clone = ggml_soft_max(ggml_ctx, src0_clone);
+        }
+    } else if (tensor->op == GGML_OP_DIAG_MASK_INF) {
+        tensor_clone = ggml_diag_mask_inf(ggml_ctx, src0_clone, *(int *)tensor->op_params);
+    } else if (tensor->op == GGML_OP_ROPE) {
+        const int n_dims      = ((int32_t *) tensor->op_params)[1];
+        const int mode        = ((int32_t *) tensor->op_params)[2];
+        //const int n_ctx_ggml       = ((int32_t *) tensor->op_params)[3];
+        const int n_ctx_orig_ggml  = ((int32_t *) tensor->op_params)[4];
+        const float freq_base       = ((float *) tensor->op_params)[5];
+        const float freq_scale      = ((float *) tensor->op_params)[6];
+        const float ext_factor      = ((float *) tensor->op_params)[7];
+        const float attn_factor     = ((float *) tensor->op_params)[8];
+        const float beta_fast       = ((float *) tensor->op_params)[9];
+        const float beta_slow       = ((float *) tensor->op_params)[10];
+        tensor_clone = ggml_rope_ext(ggml_ctx, src0_clone, src1_clone, src2_clone, n_dims, mode, n_ctx_orig_ggml, freq_base, freq_scale, ext_factor, attn_factor, beta_fast, beta_slow);
+    } else if (tensor->op == GGML_OP_UNARY) {
+        switch (ggml_get_unary_op(tensor)) {
+        case GGML_UNARY_OP_SILU:
+            tensor_clone = ggml_silu(ggml_ctx, src0_clone);
+            break;
+        case GGML_UNARY_OP_GELU:
+            tensor_clone = ggml_gelu(ggml_ctx, src0_clone);
+            break;
+        case GGML_UNARY_OP_GELU_QUICK:
+            tensor_clone = ggml_gelu_quick(ggml_ctx, src0_clone);
+            break;
+        case GGML_UNARY_OP_RELU:
+            tensor_clone = ggml_relu(ggml_ctx, src0_clone);
+            break;
+        case GGML_UNARY_OP_TANH:
+            tensor_clone = ggml_tanh(ggml_ctx, src0_clone);
+            break;
+        default:
+            std::cerr << "Missing vk_check_results OP: " << ggml_op_name(tensor->op) << std::endl;
+            GGML_ABORT("fatal error");
+        }
+    } else if (tensor->op == GGML_OP_CPY || tensor->op == GGML_OP_DUP) {
+        if (src1 == nullptr) {
+            tensor_clone = ggml_dup(ggml_ctx, src0_clone);
+            tensor_clone->type = tensor->type;
+        } else {
+            tensor_clone = ggml_cpy(ggml_ctx, src0_clone, src1_clone);
+        }
+    } else if (tensor->op == GGML_OP_CONT) {
+        tensor_clone = ggml_cont_4d(ggml_ctx, src0_clone, tensor->ne[0], tensor->ne[1], tensor->ne[2], tensor->ne[3]);
+    } else if (tensor->op == GGML_OP_RESHAPE) {
+        tensor_clone = ggml_reshape_4d(ggml_ctx, src0_clone, tensor->ne[0], tensor->ne[1], tensor->ne[2], tensor->ne[3]);
+    } else if (tensor->op == GGML_OP_VIEW) {
+        tensor_clone = ggml_view_4d(ggml_ctx, src0_clone, tensor->ne[0], tensor->ne[1], tensor->ne[2], tensor->ne[3], tensor->nb[1], tensor->nb[2], tensor->nb[3], ((int32_t *) tensor->op_params)[0]);
+    } else if (tensor->op == GGML_OP_PERMUTE) {
+        int32_t * params = (int32_t *)tensor->op_params;
+        tensor_clone = ggml_permute(ggml_ctx, src0_clone, params[0], params[1], params[2], params[3]);
+    } else if (tensor->op == GGML_OP_TRANSPOSE) {
+        tensor_clone = ggml_transpose(ggml_ctx, src0_clone);
+    } else if (tensor->op == GGML_OP_GET_ROWS) {
+        tensor_clone = ggml_get_rows(ggml_ctx, src0_clone, src1_clone);
+    } else if (tensor->op == GGML_OP_ARGSORT) {
+        tensor_clone = ggml_argsort(ggml_ctx, src0_clone, (ggml_sort_order) *(int *)tensor->op_params);
+    } else if (tensor->op == GGML_OP_SUM_ROWS) {
+        tensor_clone = ggml_sum_rows(ggml_ctx, src0_clone);
+    } else if (tensor->op == GGML_OP_IM2COL) {
+        const int32_t s0 = tensor->op_params[0];
+        const int32_t s1 = tensor->op_params[1];
+        const int32_t p0 = tensor->op_params[2];
+        const int32_t p1 = tensor->op_params[3];
+        const int32_t d0 = tensor->op_params[4];
+        const int32_t d1 = tensor->op_params[5];
+
+        const bool is_2D = tensor->op_params[6] == 1;
+        tensor_clone = ggml_im2col(ggml_ctx, src0_clone, src1_clone, s0, s1, p0, p1, d0, d1, is_2D, tensor->type);
+    } else if (tensor->op == GGML_OP_TIMESTEP_EMBEDDING) {
+        const int32_t dim = tensor->op_params[0];
+        const int32_t max_period = tensor->op_params[1];
+        tensor_clone = ggml_timestep_embedding(ggml_ctx, src0_clone, dim, max_period);
+    } else if (tensor->op == GGML_OP_LEAKY_RELU) {
+        const float * op_params = (const float *)tensor->op_params;
+        tensor_clone = ggml_leaky_relu(ggml_ctx, src0_clone, op_params[0], false);
+    } else {
+        std::cerr << "Missing vk_check_results OP: " << ggml_op_name(tensor->op) << std::endl;
+        GGML_ABORT("fatal error");
+    }
+
+    ggml_cgraph * cgraph = ggml_new_graph(ggml_ctx);
+    ggml_build_forward_expand(cgraph, tensor_clone);
+
+    ggml_graph_compute_with_ctx(ggml_ctx, cgraph, 8);
+
+    if (vk_output_tensor > 0 && vk_output_tensor == check_counter) {
+        ggml_vk_print_tensor(tensor_clone, "tensor_clone");
+    }
+
+    comp_size = ggml_nbytes(tensor_clone);
+
+    comp_result = malloc(comp_size);
+    memcpy(comp_result, tensor_clone->data, comp_size);
+    memcpy(comp_nb, tensor_clone->nb, sizeof(size_t) * GGML_MAX_DIMS);
+
+    if (src0 != nullptr) {
+        free(src0_buffer);
+    }
+    if (src1 != nullptr) {
+        free(src1_buffer);
+    }
+
+    ggml_free(ggml_ctx);
+
+    VK_LOG_DEBUG("END ggml_vk_check_results_0(" << tensor->name << ")");
+}
+
+static void ggml_vk_check_results_1(ggml_tensor * tensor) {
+    if (tensor->op == GGML_OP_TRANSPOSE) {
+        return;
+    }
+    if (!(vk_output_tensor > 0 && vk_output_tensor == check_counter) && check_counter <= vk_skip_checks) {
+        return;
+    }
+
+    VK_LOG_DEBUG("ggml_vk_check_results_1(" << tensor->name << ")");
+
+    ggml_tensor * src0 = tensor->src[0];
+    ggml_tensor * src1 = tensor->src[1];
+    ggml_tensor * src2 = tensor->src[2];
+
+    void * tensor_data = tensor->data;
+
+    if (ggml_backend_buffer_is_vk(tensor->buffer)) {
+        size_t tensor_size = ggml_nbytes(tensor);
+        tensor_data = malloc(tensor_size);
+
+        ggml_backend_vk_buffer_context * buf_ctx = (ggml_backend_vk_buffer_context *)tensor->buffer->context;
+
+        vk_buffer& buffer_gpu = buf_ctx->dev_buffer;
+        uint64_t offset = vk_tensor_offset(tensor) + tensor->view_offs;
+        if (offset + tensor_size >= buffer_gpu->size) {
+            tensor_size = buffer_gpu->size - offset;
+        }
+
+        ggml_vk_buffer_read(buffer_gpu, offset, tensor_data, tensor_size);
+    }
+
+    float first_error_result = -1.0f;
+    float first_error_correct = -1.0f;
+    std::array<int, 4> first_error = { -1, -1, -1, -1 };
+    double avg_err = 0.0;
+    size_t counter = 0;
+
+    for (int i3 = 0; i3 < tensor->ne[3]; i3++) {
+        for (int i2 = 0; i2 < tensor->ne[2]; i2++) {
+            for (int i1 = 0; i1 < tensor->ne[1]; i1++) {
+                for (int i0 = 0; i0 < tensor->ne[0]; i0++) {
+                    const bool buffer_size_fit = i3*comp_nb[3] + i2*comp_nb[2] + i1*comp_nb[1] + i0*comp_nb[0] < comp_size;
+                    float correct = 0.0f;
+                    float result = 0.0f;
+
+                    if (buffer_size_fit) {
+                        if (tensor->type == GGML_TYPE_F32) {
+                            correct = *(float *) ((char *) comp_result + i3*comp_nb[3] + i2*comp_nb[2] + i1*comp_nb[1] + i0*comp_nb[0]);
+                            result  = *(float *) ((char *) tensor_data + i3*tensor->nb[3] + i2*tensor->nb[2] + i1*tensor->nb[1] + i0*tensor->nb[0]);
+                        } else if (tensor->type == GGML_TYPE_F16) {
+                            correct = ggml_fp16_to_fp32(*(ggml_fp16_t *) ((char *) comp_result + i3*comp_nb[3] + i2*comp_nb[2] + i1*comp_nb[1] + i0*comp_nb[0]));
+                            result  = ggml_fp16_to_fp32(*(ggml_fp16_t *) ((char *) tensor_data + i3*tensor->nb[3] + i2*tensor->nb[2] + i1*tensor->nb[1] + i0*tensor->nb[0]));
+                        } else if (tensor->type == GGML_TYPE_I32) {
+                            correct = *(int32_t *) ((char *) comp_result + i3*comp_nb[3] + i2*comp_nb[2] + i1*comp_nb[1] + i0*comp_nb[0]);
+                            result  = *(int32_t *) ((char *) tensor_data + i3*tensor->nb[3] + i2*tensor->nb[2] + i1*tensor->nb[1] + i0*tensor->nb[0]);
+                        } else {
+                            std::cerr << "Results check not implemented for type " << ggml_type_name(tensor->type) << std::endl;
+                        }
+                    } else {
+                        std::cerr << "Missing debug code for type " << ggml_type_name(tensor->type) << std::endl;
+                        GGML_ABORT("fatal error");
+                    }
+
+                    if ((std::isnan(correct) != std::isnan(result)) || (std::isinf(correct) != std::isinf(result)) || !buffer_size_fit) {
+                        std::cerr << "ERROR: Invalid value in " << ggml_op_name(tensor->op) << " i3=" << i3 << " i2=" << i2 << " i1=" << i1 << " i0=" << i0 << " result=" << result << " correct=" << correct << " avg_err=" << (avg_err / counter) << std::endl;
+                        std::cerr << "tensor=" << tensor << " tensor->name=" << tensor->name << " tensor->type: " << ggml_type_name(tensor->type) << " ne0=" << tensor->ne[0] << " nb0=" << tensor->nb[0] << " ne1=" << tensor->ne[1] << " nb1=" << tensor->nb[1] << " ne2=" << tensor->ne[2] << " nb2=" << tensor->nb[2] << " ne3=" << tensor->ne[3] << " nb3=" << tensor->nb[3] << " offset=" << tensor->view_offs << std::endl;
+                        if (src0 != nullptr) {
+                            std::cerr << "src0=" << src0 << " src0->name=" << src0->name << " op=" << ggml_op_name(src0->op) << " type=" << ggml_type_name(src0->type) << " ne0=" << src0->ne[0] << " nb0=" << src0->nb[0] << " ne1=" << src0->ne[1] << " nb1=" << src0->nb[1] << " ne2=" << src0->ne[2] << " nb2=" << src0->nb[2] << " ne3=" << src0->ne[3] << " nb3=" << src0->nb[3] << " offset=" << src0->view_offs << std::endl;
+                        }
+                        if (src1 != nullptr) {
+                            std::cerr << "src1=" << src1 << " src1->name=" << src1->name << " op=" << ggml_op_name(src1->op) << " type=" << ggml_type_name(src1->type) << " ne0=" << src1->ne[0] << " nb0=" << src1->nb[0] << " ne1=" << src1->ne[1] << " nb1=" << src1->nb[1] << " ne2=" << src1->ne[2] << " nb2=" << src1->nb[2] << " ne3=" << src1->ne[3] << " nb3=" << src1->nb[3] << " offset=" << src1->view_offs << std::endl;
+                        }
+                        if (src2 != nullptr) {
+                            std::cerr << "src2=" << src2 << " src2->name=" << src2->name << " op=" << ggml_op_name(src2->op) << " type=" << ggml_type_name(src2->type) << " ne0=" << src2->ne[0] << " nb0=" << src2->nb[0] << " ne1=" << src2->ne[1] << " nb1=" << src2->nb[1] << " ne2=" << src2->ne[2] << " nb2=" << src2->nb[2] << " ne3=" << src2->ne[3] << " nb3=" << src2->nb[3] << " offset=" << src2->view_offs << std::endl;
+                        }
+                        std::cerr << "First error: result=" << first_error_result << " correct=" << first_error_correct  << " i3=" << first_error[3] << " i2=" << first_error[2] << " i1=" << first_error[1] << " i0=" << first_error[0] << std::endl;
+                        std::cerr << std::endl << "Result:" << std::endl;
+                        ggml_vk_print_tensor_area(tensor, tensor_data, i0, i1, i2, i3);
+                        std::cerr << std::endl << "Correct:" << std::endl;
+                        ggml_vk_print_tensor_area(tensor, comp_result, i0, i1, i2, i3);
+                        std::cerr << std::endl;
+                        std::vector<const ggml_tensor *> done;
+                        ggml_vk_print_graph_origin(tensor, done);
+                        GGML_ABORT("fatal error");
+                    }
+                    if (first_error[0] == -1 && std::fabs(correct - result) > 0.1f) {
+                        first_error[0] = i0;
+                        first_error[1] = i1;
+                        first_error[2] = i2;
+                        first_error[3] = i3;
+                        first_error_result = result;
+                        first_error_correct = correct;
+                    }
+
+                    // Special case, value is infinite, avoid NaN result in avg_err
+                    // NaN also appears in results, if both are nan error is 0
+                    if (!std::isinf(correct) && !std::isinf(result) && !std::isnan(correct) && !std::isnan(result)) {
+                        avg_err += std::fabs(correct - result);
+                    }
+                    counter++;
+                }
+            }
+        }
+    }
+
+    avg_err /= counter;
+
+    if (vk_output_tensor > 0 && vk_output_tensor == check_counter) {
+        std::cerr << "TENSOR CHECK: avg_err=" << avg_err << " in " << ggml_op_name(tensor->op) << " (check " << check_counter << ")" << std::endl;
+        std::cerr << "tensor=" << tensor << " tensor->name=" << tensor->name << " tensor->type: " << ggml_type_name(tensor->type) << " ne0=" << tensor->ne[0] << " nb0=" << tensor->nb[0] << " ne1=" << tensor->ne[1] << " nb1=" << tensor->nb[1] << " ne2=" << tensor->ne[2] << " nb2=" << tensor->nb[2] << " ne3=" << tensor->ne[3] << " nb3=" << tensor->nb[3] << " offset=" << tensor->view_offs << std::endl;
+        if (src0 != nullptr) {
+            std::cerr << "src0=" << src0 << " op=" << ggml_op_name(src0->op) << " type=" << ggml_type_name(src0->type) << " ne0=" << src0->ne[0] << " nb0=" << src0->nb[0] << " ne1=" << src0->ne[1] << " nb1=" << src0->nb[1] << " ne2=" << src0->ne[2] << " nb2=" << src0->nb[2] << " ne3=" << src0->ne[3] << " nb3=" << src0->nb[3] << " offset=" << src0->view_offs << std::endl;
+        }
+        if (src1 != nullptr) {
+            std::cerr << "src1=" << src1 << " op=" << ggml_op_name(src1->op) << " type=" << ggml_type_name(src1->type) << " ne0=" << src1->ne[0] << " nb0=" << src1->nb[0] << " ne1=" << src1->ne[1] << " nb1=" << src1->nb[1] << " ne2=" << src1->ne[2] << " nb2=" << src1->nb[2] << " ne3=" << src1->ne[3] << " nb3=" << src1->nb[3] << " offset=" << src1->view_offs << std::endl;
+        }
+        if (src2 != nullptr) {
+            std::cerr << "src2=" << src2 << " op=" << ggml_op_name(src2->op) << " type=" << ggml_type_name(src2->type) << " ne0=" << src2->ne[0] << " nb0=" << src2->nb[0] << " ne1=" << src2->ne[1] << " nb1=" << src2->nb[1] << " ne2=" << src2->ne[2] << " nb2=" << src2->nb[2] << " ne3=" << src2->ne[3] << " nb3=" << src2->nb[3] << " offset=" << src2->view_offs << std::endl;
+        }
+        std::cerr << "First error: result=" << first_error_result << " correct=" << first_error_correct  << " i3=" << first_error[3] << " i2=" << first_error[2] << " i1=" << first_error[1] << " i0=" << first_error[0] << std::endl;
+        std::cerr << std::endl << "Result:" << std::endl;
+        ggml_vk_print_tensor_area(tensor, tensor_data, 5, 5, 0, 0);
+        std::cerr << std::endl << "Correct:" << std::endl;
+        ggml_vk_print_tensor_area(tensor, comp_result, 5, 5, 0, 0);
+        std::cerr << std::endl;
+        std::vector<const ggml_tensor *> done;
+        ggml_vk_print_graph_origin(tensor, done);
+    }
+
+    if (avg_err > 0.05 || std::isnan(avg_err)) {
+        std::cerr << "ERROR: avg_err=" << avg_err << " in " << ggml_op_name(tensor->op) << " (check " << check_counter << ")" << std::endl;
+        std::cerr << "tensor=" << tensor << " tensor->name=" << tensor->name << " tensor->type: " << ggml_type_name(tensor->type) << " ne0=" << tensor->ne[0] << " nb0=" << tensor->nb[0] << " ne1=" << tensor->ne[1] << " nb1=" << tensor->nb[1] << " ne2=" << tensor->ne[2] << " nb2=" << tensor->nb[2] << " ne3=" << tensor->ne[3] << " nb3=" << tensor->nb[3] << " offset=" << tensor->view_offs << std::endl;
+        if (src0 != nullptr) {
+            std::cerr << "src0=" << src0 << " op=" << ggml_op_name(src0->op) << " type=" << ggml_type_name(src0->type) << " ne0=" << src0->ne[0] << " nb0=" << src0->nb[0] << " ne1=" << src0->ne[1] << " nb1=" << src0->nb[1] << " ne2=" << src0->ne[2] << " nb2=" << src0->nb[2] << " ne3=" << src0->ne[3] << " nb3=" << src0->nb[3] << " offset=" << src0->view_offs << std::endl;
+        }
+        if (src1 != nullptr) {
+            std::cerr << "src1=" << src1 << " op=" << ggml_op_name(src1->op) << " type=" << ggml_type_name(src1->type) << " ne0=" << src1->ne[0] << " nb0=" << src1->nb[0] << " ne1=" << src1->ne[1] << " nb1=" << src1->nb[1] << " ne2=" << src1->ne[2] << " nb2=" << src1->nb[2] << " ne3=" << src1->ne[3] << " nb3=" << src1->nb[3] << " offset=" << src1->view_offs << std::endl;
+        }
+        if (src2 != nullptr) {
+            std::cerr << "src2=" << src2 << " op=" << ggml_op_name(src2->op) << " type=" << ggml_type_name(src2->type) << " ne0=" << src2->ne[0] << " nb0=" << src2->nb[0] << " ne1=" << src2->ne[1] << " nb1=" << src2->nb[1] << " ne2=" << src2->ne[2] << " nb2=" << src2->nb[2] << " ne3=" << src2->ne[3] << " nb3=" << src2->nb[3] << " offset=" << src2->view_offs << std::endl;
+        }
+        std::cerr << "First error: result=" << first_error_result << " correct=" << first_error_correct  << " i3=" << first_error[3] << " i2=" << first_error[2] << " i1=" << first_error[1] << " i0=" << first_error[0] << std::endl;
+        std::cerr << std::endl << "Result:" << std::endl;
+        ggml_vk_print_tensor_area(tensor, tensor_data, first_error[0], first_error[1], first_error[2], first_error[3]);
+        std::cerr << std::endl << "Correct:" << std::endl;
+        ggml_vk_print_tensor_area(tensor, comp_result, first_error[0], first_error[1], first_error[2], first_error[3]);
+        std::cerr << std::endl;
+        std::vector<const ggml_tensor *> done;
+        ggml_vk_print_graph_origin(tensor, done);
+        GGML_ABORT("fatal error");
+    } else {
+        std::cerr << check_counter << " " << tensor->name << " op=" << ggml_op_name(tensor->op) << " avg_err=" << avg_err << std::endl;
+    }
+
+    free(comp_result);
+    comp_result = nullptr;
+    comp_size = 0;
+
+    if (ggml_backend_buffer_is_vk(tensor->buffer)) {
+        free(tensor_data);
+    }
+
+    VK_LOG_DEBUG("END ggml_vk_check_results_1(" << tensor->name << ")");
+}
+#endif
diff --git a/src/whisper_cpp/ggml.c b/src/whisper_cpp/ggml/src/ggml.c
similarity index 61%
rename from src/whisper_cpp/ggml.c
rename to src/whisper_cpp/ggml/src/ggml.c
index 45a37c60..03b832d0 100644
--- a/src/whisper_cpp/ggml.c
+++ b/src/whisper_cpp/ggml/src/ggml.c
@@ -1,9 +1,12 @@
-#include "R.h"
 #define _CRT_SECURE_NO_DEPRECATE // Disables ridiculous "unsafe" warnings on Windows
 #define _USE_MATH_DEFINES // For M_PI on MSVC
 
+#include "ggml-backend.h"
 #include "ggml-impl.h"
+#include "ggml-cpu-impl.h"
 #include "ggml-quants.h"
+#include "ggml.h"
+#include "ggml-aarch64.h"
 
 #if defined(_MSC_VER) || defined(__MINGW32__)
 #include <malloc.h> // using malloc.h with MSC/MINGW
@@ -24,11 +27,26 @@
 #include <limits.h>
 #include <stdarg.h>
 #include <signal.h>
+#if defined(__gnu_linux__)
+#include <syscall.h>
+#endif
+
+#ifdef GGML_USE_OPENMP
+#include <omp.h>
+#endif
 
 #ifdef GGML_USE_METAL
 #include <unistd.h>
 #endif
 
+#if defined(__ARM_FEATURE_SVE) || defined(__ARM_FEATURE_MATMUL_INT8)
+#undef GGML_USE_LLAMAFILE
+#endif
+
+#ifdef GGML_USE_LLAMAFILE
+#include <llamafile/sgemm.h>
+#endif
+
 #if defined(_MSC_VER)
 // disable "possible loss of data" to avoid hundreds of casts
 // we should just be careful :)
@@ -37,27 +55,89 @@
 // disable POSIX deprecation warnings
 // these functions are never going away, anyway
 #pragma warning(disable: 4996)
+
+// unreachable code because of multiple instances of code after GGML_ABORT
+#pragma warning(disable: 4702)
 #endif
 
+// Note: once we move threading into a separate C++ file
+// will use std::hardware_destructive_interference_size instead of hardcoding it here
+// and we'll use C++ attribute syntax.
+#define GGML_CACHE_LINE  64
+
+#if defined(__clang__) || defined(__GNUC__)
+#define GGML_CACHE_ALIGN __attribute__((aligned(GGML_CACHE_LINE)))
+#endif
+
+#if defined(__has_feature)
+#if __has_feature(thread_sanitizer)
+#define GGML_TSAN_ENABLED 1
+#endif
+#else  // __has_feature
+#if defined(__SANITIZE_THREAD__)
+#define GGML_TSAN_ENABLED 1
+#endif
+#endif // __has_feature
+
 #if defined(_WIN32)
 
+#define WIN32_LEAN_AND_MEAN
+#ifndef NOMINMAX
+    #define NOMINMAX
+#endif
 #include <windows.h>
 
+#if !defined(__clang__)
+#define GGML_CACHE_ALIGN __declspec(align(GGML_CACHE_LINE))
+
 typedef volatile LONG atomic_int;
 typedef atomic_int atomic_bool;
+typedef atomic_int atomic_flag;
+
+#define ATOMIC_FLAG_INIT 0
+
+typedef enum {
+    memory_order_relaxed,
+    memory_order_consume,
+    memory_order_acquire,
+    memory_order_release,
+    memory_order_acq_rel,
+    memory_order_seq_cst
+} memory_order;
 
 static void atomic_store(atomic_int * ptr, LONG val) {
     InterlockedExchange(ptr, val);
 }
+static void atomic_store_explicit(atomic_int * ptr, LONG val, memory_order mo) {
+    // TODO: add support for explicit memory order
+    InterlockedExchange(ptr, val);
+}
 static LONG atomic_load(atomic_int * ptr) {
     return InterlockedCompareExchange(ptr, 0, 0);
 }
+static LONG atomic_load_explicit(atomic_int * ptr, memory_order mo) {
+    // TODO: add support for explicit memory order
+    return InterlockedCompareExchange(ptr, 0, 0);
+}
 static LONG atomic_fetch_add(atomic_int * ptr, LONG inc) {
     return InterlockedExchangeAdd(ptr, inc);
 }
-static LONG atomic_fetch_sub(atomic_int * ptr, LONG dec) {
-    return atomic_fetch_add(ptr, -(dec));
+static LONG atomic_fetch_add_explicit(atomic_int * ptr, LONG inc, memory_order mo) {
+    // TODO: add support for explicit memory order
+    return InterlockedExchangeAdd(ptr, inc);
+}
+static atomic_bool atomic_flag_test_and_set(atomic_flag * ptr) {
+    return InterlockedExchange(ptr, 1);
 }
+static void atomic_flag_clear(atomic_flag * ptr) {
+    InterlockedExchange(ptr, 0);
+}
+static void atomic_thread_fence(memory_order mo) {
+    MemoryBarrier();
+}
+#else // clang
+#include <stdatomic.h>
+#endif
 
 typedef HANDLE pthread_t;
 
@@ -86,8 +166,13 @@ static int sched_yield (void) {
     return 0;
 }
 #else
+
 #include <pthread.h>
 #include <stdatomic.h>
+#include <sched.h>
+#if defined(__FreeBSD__)
+#include <pthread_np.h>
+#endif
 
 typedef void * thread_ret_t;
 
@@ -97,6 +182,8 @@ typedef void * thread_ret_t;
 
 #endif
 
+typedef pthread_t ggml_thread_t;
+
 #ifdef GGML_USE_CPU_HBM
 #include <hbwmalloc.h>
 #endif
@@ -110,47 +197,119 @@ typedef void * thread_ret_t;
 
 #include <sys/wait.h>
 
-void ggml_print_backtrace(void) {
-    /*
-    #include <execinfo.h>
-    #include <dlfcn.h>
+#if defined(__ANDROID__)
+#include <unwind.h>
+#include <dlfcn.h>
+#include <stdio.h>
 
-    void * trace[100];
+struct backtrace_state {
+    void ** current;
+    void ** end;
+};
 
-    int nptrs = backtrace(trace, sizeof(trace)/sizeof(trace[0]));
+static _Unwind_Reason_Code unwind_callback(struct _Unwind_Context* context, void* arg) {
+    struct backtrace_state * state = (struct backtrace_state *)arg;
+    uintptr_t pc = _Unwind_GetIP(context);
+    if (pc) {
+        if (state->current == state->end) {
+            return _URC_END_OF_STACK;
+        } else {
+            *state->current++ = (void*)pc;
+        }
+    }
+    return _URC_NO_REASON;
+}
 
+static void ggml_print_backtrace_symbols(void) {
+    const int max = 100;
+    void* buffer[max];
+
+    struct backtrace_state state = {buffer, buffer + max};
+    _Unwind_Backtrace(unwind_callback, &state);
+
+    int count = state.current - buffer;
+
+    for (int idx = 0; idx < count; ++idx) {
+        const void * addr = buffer[idx];
+        const char * symbol = "";
+
+        Dl_info info;
+        if (dladdr(addr, &info) && info.dli_sname) {
+            symbol = info.dli_sname;
+        }
+
+        fprintf(stderr, "%d: %p %s\n", idx, addr, symbol);
+    }
+}
+#elif defined(__linux__) && defined(__GLIBC__)
+#include <execinfo.h>
+static void ggml_print_backtrace_symbols(void) {
+    void * trace[100];
+    int nptrs = backtrace(trace, sizeof(trace)/sizeof(trace[0]));
     backtrace_symbols_fd(trace, nptrs, STDERR_FILENO);
-    */
+}
+#else
+static void ggml_print_backtrace_symbols(void) {
+    // platform not supported
+}
+#endif
 
-    // backtrack_symbols does not show line numbers, use gdb instead
+static void ggml_print_backtrace(void) {
     char attach[32];
     snprintf(attach, sizeof(attach), "attach %d", getpid());
     int pid = fork();
     if (pid == 0) {
+        // try gdb
         execlp("gdb", "gdb", "--batch",
             "-ex", "set style enabled on",
             "-ex", attach,
             "-ex", "bt -frame-info source-and-location",
             "-ex", "detach",
             "-ex", "quit",
-            NULL);
+            (char *) NULL);
+        // try lldb
+        execlp("lldb", "lldb", "--batch",
+            "-o", "bt",
+            "-o", "quit",
+            "-p", attach,
+            (char *) NULL);
+        exit(EXIT_FAILURE);
     } else {
-        waitpid(pid, NULL, 0);
+        int wstatus;
+        waitpid(pid, &wstatus, 0);
+        if (WIFEXITED(wstatus)) {
+            if (WEXITSTATUS(wstatus) == EXIT_FAILURE) {
+                // gdb failed, fallback to backtrace_symbols
+                ggml_print_backtrace_symbols();
+            }
+        }
     }
 }
 #else
-void ggml_print_backtrace(void) {
+static void ggml_print_backtrace(void) {
     // platform not supported
 }
 #endif
 
-/*#define GGML_PERF*/
+void ggml_abort(const char * file, int line, const char * fmt, ...) {
+    fflush(stdout);
+
+    fprintf(stderr, "%s:%d: ", file, line);
+
+    va_list args;
+    va_start(args, fmt);
+    vfprintf(stderr, fmt, args);
+    va_end(args);
+
+    fprintf(stderr, "\n");
+
+    ggml_print_backtrace();
+    abort();
+}
+
 #define GGML_DEBUG 0
 #define GGML_GELU_FP16
 #define GGML_GELU_QUICK_FP16
-#define GGML_SILU_FP16
-// #define GGML_CROSS_ENTROPY_EXP_FP16
-// #define GGML_FLASH_ATTN_EXP_FP16
 
 #define GGML_SOFT_MAX_UNROLL 4
 #define GGML_VEC_DOT_UNROLL  2
@@ -160,26 +319,63 @@ void ggml_print_backtrace(void) {
 // logging
 //
 
+struct ggml_logger_state {
+    ggml_log_callback log_callback;
+    void * log_callback_user_data;
+};
+static struct ggml_logger_state g_logger_state = {ggml_log_callback_default, NULL};
+
+static void ggml_log_internal_v(enum ggml_log_level level, const char * format, va_list args) {
+    if (format == NULL)
+        return;
+    va_list args_copy;
+    va_copy(args_copy, args);
+    char buffer[128];
+    int len = vsnprintf(buffer, 128, format, args);
+    if (len < 128) {
+        g_logger_state.log_callback(level, buffer, g_logger_state.log_callback_user_data);
+    } else {
+        char * buffer2 = (char *) calloc(len + 1, sizeof(char));
+        vsnprintf(buffer2, len + 1, format, args_copy);
+        buffer2[len] = 0;
+        g_logger_state.log_callback(level, buffer2, g_logger_state.log_callback_user_data);
+        free(buffer2);
+    }
+    va_end(args_copy);
+}
+
+void ggml_log_internal(enum ggml_log_level level, const char * format, ...) {
+    va_list args;
+    va_start(args, format);
+    ggml_log_internal_v(level, format, args);
+    va_end(args);
+}
+
+void ggml_log_callback_default(enum ggml_log_level level, const char * text, void * user_data) {
+    (void) level;
+    (void) user_data;
+    fputs(text, stderr);
+    fflush(stderr);
+}
+
 #if (GGML_DEBUG >= 1)
-#define GGML_PRINT_DEBUG(...) Rprintf(__VA_ARGS__)
+#define GGML_PRINT_DEBUG(...) GGML_LOG_DEBUG(__VA_ARGS__)
 #else
 #define GGML_PRINT_DEBUG(...)
 #endif
 
 #if (GGML_DEBUG >= 5)
-#define GGML_PRINT_DEBUG_5(...) Rprintf(__VA_ARGS__)
+#define GGML_PRINT_DEBUG_5(...) GGML_LOG_DEBUG(__VA_ARGS__)
 #else
 #define GGML_PRINT_DEBUG_5(...)
 #endif
 
 #if (GGML_DEBUG >= 10)
-#define GGML_PRINT_DEBUG_10(...) Rprintf(__VA_ARGS__)
+#define GGML_PRINT_DEBUG_10(...) GGML_LOG_DEBUG(__VA_ARGS__)
 #else
 #define GGML_PRINT_DEBUG_10(...)
 #endif
 
-#define GGML_PRINT(...) Rprintf(__VA_ARGS__)
-
 //
 // end of logging block
 //
@@ -196,7 +392,7 @@ void ggml_print_backtrace(void) {
 #else
 inline static void * ggml_aligned_malloc(size_t size) {
     if (size == 0) {
-        GGML_PRINT("WARNING: Behavior may be unexpected when allocating 0 bytes for ggml_aligned_malloc!\n");
+        GGML_LOG_WARN("Behavior may be unexpected when allocating 0 bytes for ggml_aligned_malloc!\n");
         return NULL;
     }
     void * aligned_memory = NULL;
@@ -218,7 +414,8 @@ inline static void * ggml_aligned_malloc(size_t size) {
                 error_desc = "insufficient memory";
                 break;
         }
-        GGML_PRINT("%s: %s (attempted to allocate %6.2f MB)\n", __func__, error_desc, size/(1024.0*1024.0));
+        GGML_LOG_ERROR("%s: %s (attempted to allocate %6.2f MB)\n", __func__, error_desc, size/(1024.0*1024.0));
+        GGML_ABORT("fatal error");
         return NULL;
     }
     return aligned_memory;
@@ -231,24 +428,43 @@ inline static void * ggml_aligned_malloc(size_t size) {
 #endif
 #endif
 
+inline static void * ggml_malloc(size_t size) {
+    if (size == 0) {
+        GGML_LOG_WARN("Behavior may be unexpected when allocating 0 bytes for ggml_malloc!\n");
+        return NULL;
+    }
+    void * result = malloc(size);
+    if (result == NULL) {
+        GGML_LOG_ERROR("%s: failed to allocate %6.2f MB\n", __func__, size/(1024.0*1024.0));
+        GGML_ABORT("fatal error");
+    }
+    return result;
+}
+
+// calloc
+inline static void * ggml_calloc(size_t num, size_t size) {
+    if (num == 0 || size == 0) {
+        GGML_LOG_WARN("Behavior may be unexpected when allocating 0 bytes for ggml_calloc!\n");
+        return NULL;
+    }
+    void * result = calloc(num, size);
+    if (result == NULL) {
+        GGML_LOG_ERROR("%s: failed to allocate %6.2f MB\n", __func__, size/(1024.0*1024.0));
+        GGML_ABORT("fatal error");
+    }
+    return result;
+}
+
+#define GGML_MALLOC(size)      ggml_malloc(size)
+#define GGML_CALLOC(num, size) ggml_calloc(num, size)
+
+#define GGML_FREE(ptr) free(ptr)
+
 #define UNUSED GGML_UNUSED
-#define SWAP(x, y, T) do { T SWAP = x; x = y; y = SWAP; } while (0)
+#define SWAP(x, y, T) do { T SWAP = x; (x) = y; (y) = SWAP; } while (0)
 
 #if defined(GGML_USE_ACCELERATE)
 #include <Accelerate/Accelerate.h>
-#if defined(GGML_USE_CLBLAST) // allow usage of CLBlast alongside Accelerate functions
-#include "ggml-opencl.h"
-#endif
-#elif defined(GGML_USE_OPENBLAS)
-#if defined(GGML_BLAS_USE_MKL)
-#include <mkl.h>
-#else
-#include <cblas.h>
-#endif
-#elif defined(GGML_USE_CUBLAS)
-#include "ggml-cuda.h"
-#elif defined(GGML_USE_CLBLAST)
-#include "ggml-opencl.h"
 #endif
 
 // floating point type used to accumulate sums
@@ -270,33 +486,57 @@ static ggml_fp16_t ggml_table_gelu_f16[1 << 16];
 // precomputed quick gelu table for f16 (128 KB)
 static ggml_fp16_t ggml_table_gelu_quick_f16[1 << 16];
 
-// precomputed silu table for f16 (128 KB)
-static ggml_fp16_t ggml_table_silu_f16[1 << 16];
-
-// precomputed exp table for f16 (128 KB)
-static ggml_fp16_t ggml_table_exp_f16[1 << 16];
-
 // precomputed f32 table for f16 (256 KB) (ggml-impl.h)
 float ggml_table_f32_f16[1 << 16];
 
-// note: do not use these inside ggml.c
-// these are meant to be used via the ggml.h API
+#if defined(__ARM_ARCH)
+struct ggml_arm_arch_features_type {
+    int has_neon;
+    int has_i8mm;
+    int has_sve;
+    int sve_cnt;
+} ggml_arm_arch_features = {-1, -1, -1, 0};
+#endif
+
+const char * ggml_status_to_string(enum ggml_status status) {
+    switch (status) {
+        case GGML_STATUS_ALLOC_FAILED: return "GGML status: error (failed to allocate memory)";
+        case GGML_STATUS_FAILED:       return "GGML status: error (operation failed)";
+        case GGML_STATUS_SUCCESS:      return "GGML status: success";
+        case GGML_STATUS_ABORTED:      return "GGML status: warning (operation aborted)";
+    }
+
+    return "GGML status: unknown";
+}
+
 float ggml_fp16_to_fp32(ggml_fp16_t x) {
-    return (float) GGML_FP16_TO_FP32(x);
+#define ggml_fp16_to_fp32 do_not_use__ggml_fp16_to_fp32__in_ggml
+    return GGML_FP16_TO_FP32(x);
 }
 
 ggml_fp16_t ggml_fp32_to_fp16(float x) {
+#define ggml_fp32_to_fp16 do_not_use__ggml_fp32_to_fp16__in_ggml
     return GGML_FP32_TO_FP16(x);
 }
 
-void ggml_fp16_to_fp32_row(const ggml_fp16_t * x, float * y, int n) {
-    for (int i = 0; i < n; i++) {
+float ggml_bf16_to_fp32(ggml_bf16_t x) {
+#define ggml_bf16_to_fp32 do_not_use__ggml_bf16_to_fp32__in_ggml
+    return GGML_BF16_TO_FP32(x);  // it just left shifts
+}
+
+ggml_bf16_t ggml_fp32_to_bf16(float x) {
+#define ggml_fp32_to_bf16 do_not_use__ggml_fp32_to_bf16__in_ggml
+    return GGML_FP32_TO_BF16(x);
+}
+
+void ggml_fp16_to_fp32_row(const ggml_fp16_t * x, float * y, int64_t n) {
+    for (int64_t i = 0; i < n; i++) {
         y[i] = GGML_FP16_TO_FP32(x[i]);
     }
 }
 
-void ggml_fp32_to_fp16_row(const float * x, ggml_fp16_t * y, int n) {
-    int i = 0;
+void ggml_fp32_to_fp16_row(const float * x, ggml_fp16_t * y, int64_t n) {
+    int64_t i = 0;
 #if defined(__F16C__)
     for (; i + 7 < n; i += 8) {
         __m256 x_vec = _mm256_loadu_ps(x + i);
@@ -314,6 +554,60 @@ void ggml_fp32_to_fp16_row(const float * x, ggml_fp16_t * y, int n) {
     }
 }
 
+void ggml_bf16_to_fp32_row(const ggml_bf16_t * x, float * y, int64_t n) {
+    int64_t i = 0;
+#if defined(__AVX512F__)
+    for (; i + 16 <= n; i += 16) {
+        _mm512_storeu_ps(y + i,
+                         _mm512_castsi512_ps(
+                             _mm512_slli_epi32(
+                                 _mm512_cvtepu16_epi32(
+                                     _mm256_loadu_si256(
+                                         (const __m256i *)(x + i))),
+                                 16)));
+    }
+#elif defined(__AVX2__)
+    for (; i + 8 <= n; i += 8) {
+        _mm256_storeu_ps(y + i,
+                         _mm256_castsi256_ps(
+                             _mm256_slli_epi32(
+                                 _mm256_cvtepu16_epi32(
+                                     _mm_loadu_si128(
+                                         (const __m128i *)(x + i))),
+                                 16)));
+    }
+#endif
+    for (; i < n; i++) {
+        y[i] = GGML_BF16_TO_FP32(x[i]);
+    }
+}
+
+void ggml_fp32_to_bf16_row_ref(const float * x, ggml_bf16_t * y, int64_t n) {
+    for (int i = 0; i < n; i++) {
+        y[i] = ggml_compute_fp32_to_bf16(x[i]);
+    }
+}
+
+void ggml_fp32_to_bf16_row(const float * x, ggml_bf16_t * y, int64_t n) {
+  int i = 0;
+#if defined(__AVX512BF16__)
+  // subnormals are flushed to zero on this platform
+  for (; i + 32 <= n; i += 32) {
+        _mm512_storeu_si512(
+            (__m512i *)(y + i),
+            m512i(_mm512_cvtne2ps_pbh(_mm512_loadu_ps(x + i + 16),
+                                _mm512_loadu_ps(x + i))));
+  }
+#endif
+    for (; i < n; i++) {
+        y[i] = GGML_FP32_TO_BF16(x[i]);
+    }
+}
+
+bool ggml_guid_matches(ggml_guid_t guid_a, ggml_guid_t guid_b) {
+    return memcmp(guid_a, guid_b, sizeof(ggml_guid)) == 0;
+}
+
 //
 // timing
 //
@@ -364,17 +658,56 @@ int64_t ggml_cycles_per_ms(void) {
     return CLOCKS_PER_SEC/1000;
 }
 
-#ifdef GGML_PERF
-#define ggml_perf_time_ms()       ggml_time_ms()
-#define ggml_perf_time_us()       ggml_time_us()
-#define ggml_perf_cycles()        ggml_cycles()
-#define ggml_perf_cycles_per_ms() ggml_cycles_per_ms()
+//
+// cross-platform UTF-8 file paths
+//
+
+#ifdef _WIN32
+static wchar_t * ggml_mbstowcs(const char * mbs) {
+    int wlen = MultiByteToWideChar(CP_UTF8, 0, mbs, -1, NULL, 0);
+    if (!wlen) {
+        errno = EINVAL;
+        return NULL;
+    }
+
+    wchar_t * wbuf = GGML_MALLOC(wlen * sizeof(wchar_t));
+    wlen = MultiByteToWideChar(CP_UTF8, 0, mbs, -1, wbuf, wlen);
+    if (!wlen) {
+        GGML_FREE(wbuf);
+        errno = EINVAL;
+        return NULL;
+    }
+
+    return wbuf;
+}
+#endif
+
+FILE * ggml_fopen(const char * fname, const char * mode) {
+#ifdef _WIN32
+    FILE * file = NULL;
+
+    // convert fname (UTF-8)
+    wchar_t * wfname = ggml_mbstowcs(fname);
+    if (wfname) {
+        // convert mode (ANSI)
+        wchar_t * wmode = GGML_MALLOC((strlen(mode) + 1) * sizeof(wchar_t));
+        wchar_t * wmode_p = wmode;
+        do {
+            *wmode_p++ = (wchar_t)*mode;
+        } while (*mode++);
+
+        // open file
+        file = _wfopen(wfname, wmode);
+
+        GGML_FREE(wfname);
+        GGML_FREE(wmode);
+    }
+
+    return file;
 #else
-#define ggml_perf_time_ms()       0
-#define ggml_perf_time_us()       0
-#define ggml_perf_cycles()        0
-#define ggml_perf_cycles_per_ms() 0
+    return fopen(fname, mode);
 #endif
+}
 
 //
 // cache line
@@ -392,8 +725,9 @@ int64_t ggml_cycles_per_ms(void) {
 
 static const size_t CACHE_LINE_SIZE_F32 = CACHE_LINE_SIZE/sizeof(float);
 
-static void ggml_vec_dot_f32(const int n, float * restrict s, const float * restrict x, const float * restrict y);
-static void ggml_vec_dot_f16(const int n, float * restrict s, ggml_fp16_t * restrict x, ggml_fp16_t * restrict y);
+static void ggml_vec_dot_f32(int n, float * restrict s, size_t bs, const float * restrict x, size_t bx, const float * restrict y, size_t by, int nrc);
+static void ggml_vec_dot_f16(int n, float * restrict s, size_t bs, ggml_fp16_t * restrict x, size_t bx, ggml_fp16_t * restrict y, size_t by, int nrc);
+static void ggml_vec_dot_bf16(int n, float * restrict s, size_t bs, ggml_bf16_t * restrict x, size_t bx, ggml_bf16_t * restrict y, size_t by, int nrc);
 
 static const ggml_type_traits_t type_traits[GGML_TYPE_COUNT] = {
     [GGML_TYPE_I8] = {
@@ -414,6 +748,19 @@ static const ggml_type_traits_t type_traits[GGML_TYPE_COUNT] = {
         .type_size                = sizeof(int32_t),
         .is_quantized             = false,
     },
+    [GGML_TYPE_I64] = {
+        .type_name                = "i64",
+        .blck_size                = 1,
+        .type_size                = sizeof(int64_t),
+        .is_quantized             = false,
+    },
+    [GGML_TYPE_F64] = {
+        .type_name                = "f64",
+        .blck_size                = 1,
+        .type_size                = sizeof(double),
+        .is_quantized             = false,
+        .nrows                    = 1,
+    },
     [GGML_TYPE_F32] = {
         .type_name                = "f32",
         .blck_size                = 1,
@@ -421,6 +768,7 @@ static const ggml_type_traits_t type_traits[GGML_TYPE_COUNT] = {
         .is_quantized             = false,
         .vec_dot                  = (ggml_vec_dot_t) ggml_vec_dot_f32,
         .vec_dot_type             = GGML_TYPE_F32,
+        .nrows                    = 1,
     },
     [GGML_TYPE_F16] = {
         .type_name                = "f16",
@@ -429,9 +777,10 @@ static const ggml_type_traits_t type_traits[GGML_TYPE_COUNT] = {
         .is_quantized             = false,
         .to_float                 = (ggml_to_float_t) ggml_fp16_to_fp32_row,
         .from_float               = (ggml_from_float_t) ggml_fp32_to_fp16_row,
-        .from_float_reference     = (ggml_from_float_t) ggml_fp32_to_fp16_row,
+        .from_float_ref           = (ggml_from_float_t) ggml_fp32_to_fp16_row,
         .vec_dot                  = (ggml_vec_dot_t) ggml_vec_dot_f16,
         .vec_dot_type             = GGML_TYPE_F16,
+        .nrows                    = 1,
     },
     [GGML_TYPE_Q4_0] = {
         .type_name                = "q4_0",
@@ -440,9 +789,14 @@ static const ggml_type_traits_t type_traits[GGML_TYPE_COUNT] = {
         .is_quantized             = true,
         .to_float                 = (ggml_to_float_t) dequantize_row_q4_0,
         .from_float               = quantize_row_q4_0,
-        .from_float_reference     = (ggml_from_float_t) quantize_row_q4_0_reference,
+        .from_float_ref           = (ggml_from_float_t) quantize_row_q4_0_ref,
         .vec_dot                  = ggml_vec_dot_q4_0_q8_0,
         .vec_dot_type             = GGML_TYPE_Q8_0,
+#if defined (__ARM_FEATURE_MATMUL_INT8)
+        .nrows                    = 2,
+#else
+        .nrows                    = 1,
+#endif
     },
     [GGML_TYPE_Q4_1] = {
         .type_name                = "q4_1",
@@ -451,9 +805,14 @@ static const ggml_type_traits_t type_traits[GGML_TYPE_COUNT] = {
         .is_quantized             = true,
         .to_float                 = (ggml_to_float_t) dequantize_row_q4_1,
         .from_float               = quantize_row_q4_1,
-        .from_float_reference     = (ggml_from_float_t) quantize_row_q4_1_reference,
+        .from_float_ref           = (ggml_from_float_t) quantize_row_q4_1_ref,
         .vec_dot                  = ggml_vec_dot_q4_1_q8_1,
         .vec_dot_type             = GGML_TYPE_Q8_1,
+#if defined (__ARM_FEATURE_MATMUL_INT8)
+        .nrows                    = 2,
+#else
+        .nrows                    = 1,
+#endif
     },
     [4] = { // GGML_TYPE_Q4_2
         .type_name                = "DEPRECATED",
@@ -462,9 +821,10 @@ static const ggml_type_traits_t type_traits[GGML_TYPE_COUNT] = {
         .is_quantized             = false,
         .to_float                 = NULL,
         .from_float               = NULL,
-        .from_float_reference     = NULL,
+        .from_float_ref           = NULL,
         .vec_dot                  = NULL,
         .vec_dot_type             = GGML_TYPE_COUNT,
+        .nrows                    = 1,
     },
     [5] = { // GGML_TYPE_Q4_3
         .type_name                = "DEPRECATED",
@@ -473,9 +833,10 @@ static const ggml_type_traits_t type_traits[GGML_TYPE_COUNT] = {
         .is_quantized             = false,
         .to_float                 = NULL,
         .from_float               = NULL,
-        .from_float_reference     = NULL,
+        .from_float_ref           = NULL,
         .vec_dot                  = NULL,
         .vec_dot_type             = GGML_TYPE_COUNT,
+        .nrows                    = 1,
     },
     [GGML_TYPE_Q5_0] = {
         .type_name                = "q5_0",
@@ -484,9 +845,10 @@ static const ggml_type_traits_t type_traits[GGML_TYPE_COUNT] = {
         .is_quantized             = true,
         .to_float                 = (ggml_to_float_t) dequantize_row_q5_0,
         .from_float               = quantize_row_q5_0,
-        .from_float_reference     = (ggml_from_float_t) quantize_row_q5_0_reference,
+        .from_float_ref           = (ggml_from_float_t) quantize_row_q5_0_ref,
         .vec_dot                  = ggml_vec_dot_q5_0_q8_0,
         .vec_dot_type             = GGML_TYPE_Q8_0,
+        .nrows                    = 1,
     },
     [GGML_TYPE_Q5_1] = {
         .type_name                = "q5_1",
@@ -495,9 +857,10 @@ static const ggml_type_traits_t type_traits[GGML_TYPE_COUNT] = {
         .is_quantized             = true,
         .to_float                 = (ggml_to_float_t) dequantize_row_q5_1,
         .from_float               = quantize_row_q5_1,
-        .from_float_reference     = (ggml_from_float_t) quantize_row_q5_1_reference,
+        .from_float_ref           = (ggml_from_float_t) quantize_row_q5_1_ref,
         .vec_dot                  = ggml_vec_dot_q5_1_q8_1,
         .vec_dot_type             = GGML_TYPE_Q8_1,
+        .nrows                    = 1,
     },
     [GGML_TYPE_Q8_0] = {
         .type_name                = "q8_0",
@@ -506,9 +869,15 @@ static const ggml_type_traits_t type_traits[GGML_TYPE_COUNT] = {
         .is_quantized             = true,
         .to_float                 = (ggml_to_float_t) dequantize_row_q8_0,
         .from_float               = quantize_row_q8_0,
-        .from_float_reference     = (ggml_from_float_t) quantize_row_q8_0_reference,
+        .from_float_ref           = (ggml_from_float_t) quantize_row_q8_0_ref,
+        .from_float_to_mat        = quantize_mat_q8_0,
         .vec_dot                  = ggml_vec_dot_q8_0_q8_0,
         .vec_dot_type             = GGML_TYPE_Q8_0,
+#if defined (__ARM_FEATURE_MATMUL_INT8)
+        .nrows                    = 2,
+#else
+        .nrows                    = 1,
+#endif
     },
     [GGML_TYPE_Q8_1] = {
         .type_name                = "q8_1",
@@ -516,8 +885,9 @@ static const ggml_type_traits_t type_traits[GGML_TYPE_COUNT] = {
         .type_size                = sizeof(block_q8_1),
         .is_quantized             = true,
         .from_float               = quantize_row_q8_1,
-        .from_float_reference     = (ggml_from_float_t) quantize_row_q8_1_reference,
+        .from_float_ref           = (ggml_from_float_t) quantize_row_q8_1_ref,
         .vec_dot_type             = GGML_TYPE_Q8_1,
+        .nrows                    = 1,
     },
     [GGML_TYPE_Q2_K] = {
         .type_name                = "q2_K",
@@ -526,9 +896,10 @@ static const ggml_type_traits_t type_traits[GGML_TYPE_COUNT] = {
         .is_quantized             = true,
         .to_float                 = (ggml_to_float_t) dequantize_row_q2_K,
         .from_float               = quantize_row_q2_K,
-        .from_float_reference     = (ggml_from_float_t) quantize_row_q2_K_reference,
+        .from_float_ref           = (ggml_from_float_t) quantize_row_q2_K_ref,
         .vec_dot                  = ggml_vec_dot_q2_K_q8_K,
         .vec_dot_type             = GGML_TYPE_Q8_K,
+        .nrows                    = 1,
     },
     [GGML_TYPE_Q3_K] = {
         .type_name                = "q3_K",
@@ -537,9 +908,10 @@ static const ggml_type_traits_t type_traits[GGML_TYPE_COUNT] = {
         .is_quantized             = true,
         .to_float                 = (ggml_to_float_t) dequantize_row_q3_K,
         .from_float               = quantize_row_q3_K,
-        .from_float_reference     = (ggml_from_float_t) quantize_row_q3_K_reference,
+        .from_float_ref           = (ggml_from_float_t) quantize_row_q3_K_ref,
         .vec_dot                  = ggml_vec_dot_q3_K_q8_K,
         .vec_dot_type             = GGML_TYPE_Q8_K,
+        .nrows                    = 1,
     },
     [GGML_TYPE_Q4_K] = {
         .type_name                = "q4_K",
@@ -548,9 +920,10 @@ static const ggml_type_traits_t type_traits[GGML_TYPE_COUNT] = {
         .is_quantized             = true,
         .to_float                 = (ggml_to_float_t) dequantize_row_q4_K,
         .from_float               = quantize_row_q4_K,
-        .from_float_reference     = (ggml_from_float_t) quantize_row_q4_K_reference,
+        .from_float_ref           = (ggml_from_float_t) quantize_row_q4_K_ref,
         .vec_dot                  = ggml_vec_dot_q4_K_q8_K,
         .vec_dot_type             = GGML_TYPE_Q8_K,
+        .nrows                    = 1,
     },
     [GGML_TYPE_Q5_K] = {
         .type_name                = "q5_K",
@@ -559,9 +932,10 @@ static const ggml_type_traits_t type_traits[GGML_TYPE_COUNT] = {
         .is_quantized             = true,
         .to_float                 = (ggml_to_float_t) dequantize_row_q5_K,
         .from_float               = quantize_row_q5_K,
-        .from_float_reference     = (ggml_from_float_t) quantize_row_q5_K_reference,
+        .from_float_ref           = (ggml_from_float_t) quantize_row_q5_K_ref,
         .vec_dot                  = ggml_vec_dot_q5_K_q8_K,
         .vec_dot_type             = GGML_TYPE_Q8_K,
+        .nrows                    = 1,
     },
     [GGML_TYPE_Q6_K] = {
         .type_name                = "q6_K",
@@ -570,9 +944,118 @@ static const ggml_type_traits_t type_traits[GGML_TYPE_COUNT] = {
         .is_quantized             = true,
         .to_float                 = (ggml_to_float_t) dequantize_row_q6_K,
         .from_float               = quantize_row_q6_K,
-        .from_float_reference     = (ggml_from_float_t) quantize_row_q6_K_reference,
+        .from_float_ref           = (ggml_from_float_t) quantize_row_q6_K_ref,
         .vec_dot                  = ggml_vec_dot_q6_K_q8_K,
         .vec_dot_type             = GGML_TYPE_Q8_K,
+        .nrows                    = 1,
+    },
+    [GGML_TYPE_IQ2_XXS] = {
+        .type_name                = "iq2_xxs",
+        .blck_size                = QK_K,
+        .type_size                = sizeof(block_iq2_xxs),
+        .is_quantized             = true,
+        .to_float                 = (ggml_to_float_t) dequantize_row_iq2_xxs,
+        .from_float               = NULL,
+        .from_float_ref           = NULL,
+        .vec_dot                  = ggml_vec_dot_iq2_xxs_q8_K,
+        .vec_dot_type             = GGML_TYPE_Q8_K,
+        .nrows                    = 1,
+    },
+    [GGML_TYPE_IQ2_XS] = {
+        .type_name                = "iq2_xs",
+        .blck_size                = QK_K,
+        .type_size                = sizeof(block_iq2_xs),
+        .is_quantized             = true,
+        .to_float                 = (ggml_to_float_t) dequantize_row_iq2_xs,
+        .from_float               = NULL,
+        .from_float_ref           = NULL,
+        .vec_dot                  = ggml_vec_dot_iq2_xs_q8_K,
+        .vec_dot_type             = GGML_TYPE_Q8_K,
+        .nrows                    = 1,
+    },
+    [GGML_TYPE_IQ3_XXS] = {
+        .type_name                = "iq3_xxs",
+        .blck_size                = QK_K,
+        .type_size                = sizeof(block_iq3_xxs),
+        .is_quantized             = true,
+        .to_float                 = (ggml_to_float_t) dequantize_row_iq3_xxs,
+        .from_float               = quantize_row_iq3_xxs,
+        .from_float_ref           = (ggml_from_float_t)quantize_row_iq3_xxs_ref,
+        .vec_dot                  = ggml_vec_dot_iq3_xxs_q8_K,
+        .vec_dot_type             = GGML_TYPE_Q8_K,
+        .nrows                    = 1,
+    },
+    [GGML_TYPE_IQ3_S] = {
+        .type_name                = "iq3_s",
+        .blck_size                = QK_K,
+        .type_size                = sizeof(block_iq3_s),
+        .is_quantized             = true,
+        .to_float                 = (ggml_to_float_t) dequantize_row_iq3_s,
+        .from_float               = quantize_row_iq3_s,
+        .from_float_ref           = (ggml_from_float_t)quantize_row_iq3_s_ref,
+        .vec_dot                  = ggml_vec_dot_iq3_s_q8_K,
+        .vec_dot_type             = GGML_TYPE_Q8_K,
+        .nrows                    = 1,
+    },
+    [GGML_TYPE_IQ2_S] = {
+        .type_name                = "iq2_s",
+        .blck_size                = QK_K,
+        .type_size                = sizeof(block_iq2_s),
+        .is_quantized             = true,
+        .to_float                 = (ggml_to_float_t) dequantize_row_iq2_s,
+        .from_float               = quantize_row_iq2_s,
+        .from_float_ref           = (ggml_from_float_t)quantize_row_iq2_s_ref,
+        .vec_dot                  = ggml_vec_dot_iq2_s_q8_K,
+        .vec_dot_type             = GGML_TYPE_Q8_K,
+        .nrows                    = 1,
+    },
+    [GGML_TYPE_IQ1_S] = {
+        .type_name                = "iq1_s",
+        .blck_size                = QK_K,
+        .type_size                = sizeof(block_iq1_s),
+        .is_quantized             = true,
+        .to_float                 = (ggml_to_float_t) dequantize_row_iq1_s,
+        .from_float               = NULL,
+        .from_float_ref           = NULL,
+        .vec_dot                  = ggml_vec_dot_iq1_s_q8_K,
+        .vec_dot_type             = GGML_TYPE_Q8_K,
+        .nrows                    = 1,
+    },
+    [GGML_TYPE_IQ1_M] = {
+        .type_name                = "iq1_m",
+        .blck_size                = QK_K,
+        .type_size                = sizeof(block_iq1_m),
+        .is_quantized             = true,
+        .to_float                 = (ggml_to_float_t) dequantize_row_iq1_m,
+        .from_float               = NULL,
+        .from_float_ref           = NULL,
+        .vec_dot                  = ggml_vec_dot_iq1_m_q8_K,
+        .vec_dot_type             = GGML_TYPE_Q8_K,
+        .nrows                    = 1,
+    },
+    [GGML_TYPE_IQ4_NL] = {
+        .type_name                = "iq4_nl",
+        .blck_size                = QK4_NL,
+        .type_size                = sizeof(block_iq4_nl),
+        .is_quantized             = true,
+        .to_float                 = (ggml_to_float_t) dequantize_row_iq4_nl,
+        .from_float               = quantize_row_iq4_nl,
+        .from_float_ref           = (ggml_from_float_t)quantize_row_iq4_nl_ref,
+        .vec_dot                  = ggml_vec_dot_iq4_nl_q8_0,
+        .vec_dot_type             = GGML_TYPE_Q8_0,
+        .nrows                    = 1,
+    },
+    [GGML_TYPE_IQ4_XS] = {
+        .type_name                = "iq4_xs",
+        .blck_size                = QK_K,
+        .type_size                = sizeof(block_iq4_xs),
+        .is_quantized             = true,
+        .to_float                 = (ggml_to_float_t) dequantize_row_iq4_xs,
+        .from_float               = quantize_row_iq4_xs,
+        .from_float_ref           = (ggml_from_float_t)quantize_row_iq4_xs_ref,
+        .vec_dot                  = ggml_vec_dot_iq4_xs_q8_K,
+        .vec_dot_type             = GGML_TYPE_Q8_K,
+        .nrows                    = 1,
     },
     [GGML_TYPE_Q8_K] = {
         .type_name                = "q8_K",
@@ -580,7 +1063,91 @@ static const ggml_type_traits_t type_traits[GGML_TYPE_COUNT] = {
         .type_size                = sizeof(block_q8_K),
         .is_quantized             = true,
         .from_float               = quantize_row_q8_K,
-    }
+    },
+    [GGML_TYPE_BF16] = {
+        .type_name                = "bf16",
+        .blck_size                = 1,
+        .type_size                = sizeof(ggml_bf16_t),
+        .is_quantized             = false,
+        .to_float                 = (ggml_to_float_t) ggml_bf16_to_fp32_row,
+        .from_float               = (ggml_from_float_t) ggml_fp32_to_bf16_row,
+        .from_float_ref           = (ggml_from_float_t) ggml_fp32_to_bf16_row_ref,
+        .vec_dot                  = (ggml_vec_dot_t) ggml_vec_dot_bf16,
+        .vec_dot_type             = GGML_TYPE_BF16,
+        .nrows                    = 1,
+    },
+    [GGML_TYPE_Q4_0_4_4] = {
+        .type_name                = "q4_0_4x4",
+        .blck_size                = QK4_0,
+        .blck_size_interleave     = 4,
+        .type_size                = sizeof(block_q4_0),
+        .is_quantized             = true,
+        .to_float                 = NULL,
+        .from_float               = NULL,
+        .from_float_ref           = NULL,
+        .vec_dot                  = NULL,
+        .vec_dot_type             = GGML_TYPE_Q8_0,
+        .nrows                    = 1,
+        .ncols                    = 4,
+        .gemv                     = ggml_gemv_q4_0_4x4_q8_0,
+        .gemm                     = ggml_gemm_q4_0_4x4_q8_0,
+    },
+    [GGML_TYPE_Q4_0_4_8] = {
+        .type_name                = "q4_0_4x8",
+        .blck_size                = QK4_0,
+        .blck_size_interleave     = 8,
+        .type_size                = sizeof(block_q4_0),
+        .is_quantized             = true,
+        .to_float                 = NULL,
+        .from_float               = NULL,
+        .from_float_ref           = NULL,
+        .vec_dot                  = NULL,
+        .vec_dot_type             = GGML_TYPE_Q8_0,
+        .nrows                    = 1,
+        .ncols                    = 4,
+        .gemv                     = ggml_gemv_q4_0_4x8_q8_0,
+        .gemm                     = ggml_gemm_q4_0_4x8_q8_0,
+    },
+    [GGML_TYPE_Q4_0_8_8] = {
+        .type_name                = "q4_0_8x8",
+        .blck_size                = QK4_0,
+        .blck_size_interleave     = 8,
+        .type_size                = sizeof(block_q4_0),
+        .is_quantized             = true,
+        .to_float                 = NULL,
+        .from_float               = NULL,
+        .from_float_ref           = NULL,
+        .vec_dot                  = NULL,
+        .vec_dot_type             = GGML_TYPE_Q8_0,
+        .nrows                    = 1,
+        .ncols                    = 8,
+        .gemv                     = ggml_gemv_q4_0_8x8_q8_0,
+        .gemm                     = ggml_gemm_q4_0_8x8_q8_0,
+    },
+    [GGML_TYPE_TQ1_0] = {
+        .type_name                = "tq1_0",
+        .blck_size                = QK_K,
+        .type_size                = sizeof(block_tq1_0),
+        .is_quantized             = true,
+        .to_float                 = (ggml_to_float_t) dequantize_row_tq1_0,
+        .from_float               = quantize_row_tq1_0,
+        .from_float_ref           = (ggml_from_float_t) quantize_row_tq1_0_ref,
+        .vec_dot                  = ggml_vec_dot_tq1_0_q8_K,
+        .vec_dot_type             = GGML_TYPE_Q8_K,
+        .nrows                    = 1,
+    },
+    [GGML_TYPE_TQ2_0] = {
+        .type_name                = "tq2_0",
+        .blck_size                = QK_K,
+        .type_size                = sizeof(block_tq2_0),
+        .is_quantized             = true,
+        .to_float                 = (ggml_to_float_t) dequantize_row_tq2_0,
+        .from_float               = quantize_row_tq2_0,
+        .from_float_ref           = (ggml_from_float_t) quantize_row_tq2_0_ref,
+        .vec_dot                  = ggml_vec_dot_tq2_0_q8_K,
+        .vec_dot_type             = GGML_TYPE_Q8_K,
+        .nrows                    = 1,
+    },
 };
 
 // For internal test use
@@ -593,18 +1160,6 @@ ggml_type_traits_t ggml_internal_get_type_traits(enum ggml_type type) {
 // simd mappings
 //
 
-#if defined(__ARM_NEON)
-#if !defined(__aarch64__)
-
-// 64-bit compatibility
-
-inline static float vaddvq_f32(float32x4_t v) {
-    return vgetq_lane_f32(v, 0) + vgetq_lane_f32(v, 1) + vgetq_lane_f32(v, 2) + vgetq_lane_f32(v, 3);
-}
-
-#endif
-#endif
-
 // we define a common set of C macros which map to specific intrinsics based on the current architecture
 // we then implement the fundamental computation operations below using only these macros
 // adding support for new architectures requires to define the corresponding SIMD macros
@@ -634,21 +1189,21 @@ inline static float vaddvq_f32(float32x4_t v) {
 #define GGML_F32x4_ADD          vaddq_f32
 #define GGML_F32x4_MUL          vmulq_f32
 #define GGML_F32x4_REDUCE_ONE(x) vaddvq_f32(x)
-#define GGML_F32x4_REDUCE(res, x)              \
-{                                              \
-    int offset = GGML_F32_ARR >> 1;            \
-    for (int i = 0; i < offset; ++i) {         \
-        x[i] = vaddq_f32(x[i], x[offset+i]);   \
-    }                                          \
-    offset >>= 1;                              \
-    for (int i = 0; i < offset; ++i) {         \
-        x[i] = vaddq_f32(x[i], x[offset+i]);   \
-    }                                          \
-    offset >>= 1;                              \
-    for (int i = 0; i < offset; ++i) {         \
-        x[i] = vaddq_f32(x[i], x[offset+i]);   \
-    }                                          \
-    res = GGML_F32x4_REDUCE_ONE(x[0]);         \
+#define GGML_F32x4_REDUCE(res, x)                  \
+{                                                  \
+    int offset = GGML_F32_ARR >> 1;                \
+    for (int i = 0; i < offset; ++i) {             \
+        (x)[i] = vaddq_f32((x)[i], (x)[offset+i]); \
+    }                                              \
+    offset >>= 1;                                  \
+    for (int i = 0; i < offset; ++i) {             \
+        (x)[i] = vaddq_f32((x)[i], (x)[offset+i]); \
+    }                                              \
+    offset >>= 1;                                  \
+    for (int i = 0; i < offset; ++i) {             \
+        (x)[i] = vaddq_f32((x)[i], (x)[offset+i]); \
+    }                                              \
+    (res) = GGML_F32x4_REDUCE_ONE((x)[0]);         \
 }
 
 #define GGML_F32_VEC        GGML_F32x4
@@ -670,35 +1225,35 @@ inline static float vaddvq_f32(float32x4_t v) {
     #define GGML_F16x8              float16x8_t
     #define GGML_F16x8_ZERO         vdupq_n_f16(0.0f)
     #define GGML_F16x8_SET1(x)      vdupq_n_f16(x)
-    #define GGML_F16x8_LOAD         vld1q_f16
+    #define GGML_F16x8_LOAD(x)      vld1q_f16((const ggml_fp16_internal_t *)(x))
     #define GGML_F16x8_STORE        vst1q_f16
     #define GGML_F16x8_FMA(a, b, c) vfmaq_f16(a, b, c)
     #define GGML_F16x8_ADD          vaddq_f16
     #define GGML_F16x8_MUL          vmulq_f16
-    #define GGML_F16x8_REDUCE(res, x)                             \
-    do {                                                          \
-        int offset = GGML_F16_ARR >> 1;                           \
-        for (int i = 0; i < offset; ++i) {                        \
-            x[i] = vaddq_f16(x[i], x[offset+i]);                  \
-        }                                                         \
-        offset >>= 1;                                             \
-        for (int i = 0; i < offset; ++i) {                        \
-            x[i] = vaddq_f16(x[i], x[offset+i]);                  \
-        }                                                         \
-        offset >>= 1;                                             \
-        for (int i = 0; i < offset; ++i) {                        \
-            x[i] = vaddq_f16(x[i], x[offset+i]);                  \
-        }                                                         \
-        const float32x4_t t0 = vcvt_f32_f16(vget_low_f16 (x[0])); \
-        const float32x4_t t1 = vcvt_f32_f16(vget_high_f16(x[0])); \
-        res = (ggml_float) vaddvq_f32(vaddq_f32(t0, t1));         \
+    #define GGML_F16x8_REDUCE(res, x)                               \
+    do {                                                            \
+        int offset = GGML_F16_ARR >> 1;                             \
+        for (int i = 0; i < offset; ++i) {                          \
+            (x)[i] = vaddq_f16((x)[i], (x)[offset+i]);              \
+        }                                                           \
+        offset >>= 1;                                               \
+        for (int i = 0; i < offset; ++i) {                          \
+            (x)[i] = vaddq_f16((x)[i], (x)[offset+i]);              \
+        }                                                           \
+        offset >>= 1;                                               \
+        for (int i = 0; i < offset; ++i) {                          \
+            (x)[i] = vaddq_f16((x)[i], (x)[offset+i]);              \
+        }                                                           \
+        const float32x4_t t0 = vcvt_f32_f16(vget_low_f16 ((x)[0])); \
+        const float32x4_t t1 = vcvt_f32_f16(vget_high_f16((x)[0])); \
+        (res) = (ggml_float) vaddvq_f32(vaddq_f32(t0, t1));         \
     } while (0)
 
     #define GGML_F16_VEC                GGML_F16x8
     #define GGML_F16_VEC_ZERO           GGML_F16x8_ZERO
     #define GGML_F16_VEC_SET1           GGML_F16x8_SET1
     #define GGML_F16_VEC_LOAD(p, i)     GGML_F16x8_LOAD(p)
-    #define GGML_F16_VEC_STORE(p, r, i) GGML_F16x8_STORE(p, r[i])
+    #define GGML_F16_VEC_STORE(p, r, i) GGML_F16x8_STORE((ggml_fp16_internal_t *)(p), (r)[i])
     #define GGML_F16_VEC_FMA            GGML_F16x8_FMA
     #define GGML_F16_VEC_ADD            GGML_F16x8_ADD
     #define GGML_F16_VEC_MUL            GGML_F16x8_MUL
@@ -713,7 +1268,7 @@ inline static float vaddvq_f32(float32x4_t v) {
     #define GGML_F32Cx4              float32x4_t
     #define GGML_F32Cx4_ZERO         vdupq_n_f32(0.0f)
     #define GGML_F32Cx4_SET1(x)      vdupq_n_f32(x)
-    #define GGML_F32Cx4_LOAD(x)      vcvt_f32_f16(vld1_f16(x))
+    #define GGML_F32Cx4_LOAD(x)      vcvt_f32_f16(vld1_f16((const ggml_fp16_internal_t *)(x)))
     #define GGML_F32Cx4_STORE(x, y)  vst1_f16(x, vcvt_f16_f32(y))
     #define GGML_F32Cx4_FMA(a, b, c) vfmaq_f32(a, b, c)
     #define GGML_F32Cx4_ADD          vaddq_f32
@@ -724,13 +1279,108 @@ inline static float vaddvq_f32(float32x4_t v) {
     #define GGML_F16_VEC_ZERO           GGML_F32Cx4_ZERO
     #define GGML_F16_VEC_SET1           GGML_F32Cx4_SET1
     #define GGML_F16_VEC_LOAD(p, i)     GGML_F32Cx4_LOAD(p)
-    #define GGML_F16_VEC_STORE(p, r, i) GGML_F32Cx4_STORE(p, r[i])
+    #define GGML_F16_VEC_STORE(p, r, i) GGML_F32Cx4_STORE((ggml_fp16_internal_t *)(p), r[i])
     #define GGML_F16_VEC_FMA            GGML_F32Cx4_FMA
     #define GGML_F16_VEC_ADD            GGML_F32Cx4_ADD
     #define GGML_F16_VEC_MUL            GGML_F32Cx4_MUL
     #define GGML_F16_VEC_REDUCE         GGML_F32Cx4_REDUCE
 #endif
 
+#elif defined(__AVX512F__)
+
+#define GGML_SIMD
+
+// F32 AVX512
+
+#define GGML_F32_STEP 64
+#define GGML_F32_EPR  16
+
+#define GGML_F32x16         __m512
+#define GGML_F32x16_ZERO    _mm512_setzero_ps()
+#define GGML_F32x16_SET1(x) _mm512_set1_ps(x)
+#define GGML_F32x16_LOAD    _mm512_loadu_ps
+#define GGML_F32x16_STORE   _mm512_storeu_ps
+// _mm512_fmadd_ps is defined in AVX512F so no guard is required
+#define GGML_F32x16_FMA(a, b, c) _mm512_fmadd_ps(b, c, a)
+#define GGML_F32x16_ADD     _mm512_add_ps
+#define GGML_F32x16_MUL     _mm512_mul_ps
+#define GGML_F32x16_REDUCE(res, x)                                    \
+do {                                                                  \
+    int offset = GGML_F32_ARR >> 1;                                   \
+    for (int i = 0; i < offset; ++i) {                                \
+        x[i] = _mm512_add_ps(x[i], x[offset+i]);                      \
+    }                                                                 \
+    offset >>= 1;                                                     \
+    for (int i = 0; i < offset; ++i) {                                \
+        x[i] = _mm512_add_ps(x[i], x[offset+i]);                      \
+    }                                                                 \
+    offset >>= 1;                                                     \
+    for (int i = 0; i < offset; ++i) {                                \
+        x[i] = _mm512_add_ps(x[i], x[offset+i]);                      \
+    }                                                                 \
+    res = _mm512_reduce_add_ps(x[0]);                                 \
+} while (0)
+
+// TODO: is this optimal ?
+
+#define GGML_F32_VEC        GGML_F32x16
+#define GGML_F32_VEC_ZERO   GGML_F32x16_ZERO
+#define GGML_F32_VEC_SET1   GGML_F32x16_SET1
+#define GGML_F32_VEC_LOAD   GGML_F32x16_LOAD
+#define GGML_F32_VEC_STORE  GGML_F32x16_STORE
+#define GGML_F32_VEC_FMA    GGML_F32x16_FMA
+#define GGML_F32_VEC_ADD    GGML_F32x16_ADD
+#define GGML_F32_VEC_MUL    GGML_F32x16_MUL
+#define GGML_F32_VEC_REDUCE GGML_F32x16_REDUCE
+
+// F16 AVX512
+
+// F16 AVX
+
+#define GGML_F16_STEP 64
+#define GGML_F16_EPR  16
+
+// AVX512 has FP16 extension (AVX512_FP16) but I don't have it on my machine so I use FP32 instead
+
+#define GGML_F32Cx16             __m512
+#define GGML_F32Cx16_ZERO        _mm512_setzero_ps()
+#define GGML_F32Cx16_SET1(x)     _mm512_set1_ps(x)
+
+// unlike  _mm256_cvt intrinsics that require F16C, _mm512_cvt is defined in AVX512F
+// so F16C guard isn't required
+#define GGML_F32Cx16_LOAD(x)     _mm512_cvtph_ps(_mm256_loadu_si256((const __m256i *)(x)))
+#define GGML_F32Cx16_STORE(x, y) _mm256_storeu_si256((__m256i *)(x), _mm512_cvtps_ph(y, 0))
+
+#define GGML_F32Cx16_FMA(a, b, c) _mm512_fmadd_ps(b, c, a)
+#define GGML_F32Cx16_ADD         _mm512_add_ps
+#define GGML_F32Cx16_MUL         _mm512_mul_ps
+#define GGML_F32Cx16_REDUCE(res, x)                               \
+do {                                                              \
+    int offset = GGML_F32_ARR >> 1;                               \
+    for (int i = 0; i < offset; ++i) {                            \
+        x[i] = _mm512_add_ps(x[i], x[offset+i]);                  \
+    }                                                             \
+    offset >>= 1;                                                 \
+    for (int i = 0; i < offset; ++i) {                            \
+        x[i] = _mm512_add_ps(x[i], x[offset+i]);                  \
+    }                                                             \
+    offset >>= 1;                                                 \
+    for (int i = 0; i < offset; ++i) {                            \
+        x[i] = _mm512_add_ps(x[i], x[offset+i]);                  \
+    }                                                             \
+    res = _mm512_reduce_add_ps(x[0]);                             \
+} while (0)
+
+#define GGML_F16_VEC                GGML_F32Cx16
+#define GGML_F16_VEC_ZERO           GGML_F32Cx16_ZERO
+#define GGML_F16_VEC_SET1           GGML_F32Cx16_SET1
+#define GGML_F16_VEC_LOAD(p, i)     GGML_F32Cx16_LOAD(p)
+#define GGML_F16_VEC_STORE(p, r, i) GGML_F32Cx16_STORE(p, r[i])
+#define GGML_F16_VEC_FMA            GGML_F32Cx16_FMA
+#define GGML_F16_VEC_ADD            GGML_F32Cx16_ADD
+#define GGML_F16_VEC_MUL            GGML_F32Cx16_MUL
+#define GGML_F16_VEC_REDUCE         GGML_F32Cx16_REDUCE
+
 #elif defined(__AVX__)
 
 #define GGML_SIMD
@@ -769,7 +1419,7 @@ do {                                                              \
     const __m128 t0 = _mm_add_ps(_mm256_castps256_ps128(x[0]),    \
                                  _mm256_extractf128_ps(x[0], 1)); \
     const __m128 t1 = _mm_hadd_ps(t0, t0);                        \
-    res = _mm_cvtss_f32(_mm_hadd_ps(t1, t1));                     \
+    res = (ggml_float) _mm_cvtss_f32(_mm_hadd_ps(t1, t1));        \
 } while (0)
 // TODO: is this optimal ?
 
@@ -796,7 +1446,7 @@ do {                                                              \
 
 #if defined(__F16C__)
 // the  _mm256_cvt intrinsics require F16C
-#define GGML_F32Cx8_LOAD(x)     _mm256_cvtph_ps(_mm_loadu_si128((__m128i *)(x)))
+#define GGML_F32Cx8_LOAD(x)     _mm256_cvtph_ps(_mm_loadu_si128((const __m128i *)(x)))
 #define GGML_F32Cx8_STORE(x, y) _mm_storeu_si128((__m128i *)(x), _mm256_cvtps_ph(y, 0))
 #else
 static inline __m256 __avx_f32cx8_load(ggml_fp16_t *x) {
@@ -889,6 +1539,8 @@ static inline void __avx_f32cx8_store(ggml_fp16_t *x, __m256 y) {
 #define GGML_F16_VEC_ZERO   GGML_F32x4_ZERO
 #define GGML_F16_VEC_SET1   GGML_F32x4_SET1
 #define GGML_F16_VEC_FMA    GGML_F32x4_FMA
+#define GGML_F16_VEC_ADD    GGML_F32x4_ADD
+#define GGML_F16_VEC_MUL    GGML_F32x4_MUL
 #define GGML_F16_VEC_REDUCE GGML_F32x4_REDUCE
 // Use vec_xl, not vec_ld, in case the load address is not aligned.
 #define GGML_F16_VEC_LOAD(p, i) (i & 0x1) ?                   \
@@ -1050,7 +1702,7 @@ inline static void __wasm_f16x4_store(ggml_fp16_t * p, v128_t x) {
         x[i] = _mm_add_ps(x[i], x[offset+i]);                     \
     }                                                             \
     const __m128 t0 = _mm_hadd_ps(x[0], x[0]);                    \
-    res = _mm_cvtss_f32(_mm_hadd_ps(t0, t0));                     \
+    res = (ggml_float) _mm_cvtss_f32(_mm_hadd_ps(t0, t0));        \
 }
 // TODO: is this optimal ?
 
@@ -1111,148 +1763,569 @@ static inline void __sse_f16x4_store(ggml_fp16_t *x, __m128 y) {
 #define GGML_F16_VEC_MUL             GGML_F32Cx4_MUL
 #define GGML_F16_VEC_REDUCE          GGML_F32Cx4_REDUCE
 
-#endif
+#elif defined(__loongarch_asx)
 
-// GGML_F32_ARR / GGML_F16_ARR
-//   number of registers to use per step
-#ifdef GGML_SIMD
-#define GGML_F32_ARR (GGML_F32_STEP/GGML_F32_EPR)
-#define GGML_F16_ARR (GGML_F16_STEP/GGML_F16_EPR)
-#endif
+#define GGML_SIMD
 
-//
-// fundamental operations
-//
+// F32 LASX
+#define GGML_F32_STEP 32
+#define GGML_F32_EPR  8
 
-inline static void ggml_vec_set_i8(const int n, int8_t * x, const int8_t v) { for (int i = 0; i < n; ++i) x[i] = v; }
+#define GGML_F32x8         __m256
+#define GGML_F32x8_ZERO    (__m256)__lasx_xvldi(0)
+#define GGML_F32x8_SET1(x) (__m256)__lasx_xvreplfr2vr_s((x))
+#define GGML_F32x8_LOAD(x) (__m256)__lasx_xvld((x), 0)
+#define GGML_F32x8_STORE(x,y)   __lasx_xvst((y), (x), 0)
+#define GGML_F32x8_FMA(a, b, c) __lasx_xvfmadd_s(b, c, a)
+#define GGML_F32x8_ADD     __lasx_xvfadd_s
+#define GGML_F32x8_MUL     __lasx_xvfmul_s
+#define GGML_F32x8_REDUCE(res, x)                                 \
+do {                                                              \
+    int offset = GGML_F32_ARR >> 1;                               \
+    for (int i = 0; i < offset; ++i) {                            \
+        x[i] = __lasx_xvfadd_s(x[i], x[offset+i]);                  \
+    }                                                             \
+    offset >>= 1;                                                 \
+    for (int i = 0; i < offset; ++i) {                            \
+        x[i] = __lasx_xvfadd_s(x[i], x[offset+i]);                  \
+    }                                                             \
+    offset >>= 1;                                                 \
+    for (int i = 0; i < offset; ++i) {                            \
+        x[i] = __lasx_xvfadd_s(x[i], x[offset+i]);                  \
+    }                                                             \
+    float *tmp_p = (float *)&x[0]; \
+    res = tmp_p[0] + tmp_p[1] + tmp_p[2] + tmp_p[3] + tmp_p[4] + tmp_p[5] + tmp_p[6] + tmp_p[7];  \
+} while (0)
+// TODO: is this optimal ?
 
-inline static void ggml_vec_set_i16(const int n, int16_t * x, const int16_t v) { for (int i = 0; i < n; ++i) x[i] = v; }
+#define GGML_F32_VEC        GGML_F32x8
+#define GGML_F32_VEC_ZERO   GGML_F32x8_ZERO
+#define GGML_F32_VEC_SET1   GGML_F32x8_SET1
+#define GGML_F32_VEC_LOAD   GGML_F32x8_LOAD
+#define GGML_F32_VEC_STORE  GGML_F32x8_STORE
+#define GGML_F32_VEC_FMA    GGML_F32x8_FMA
+#define GGML_F32_VEC_ADD    GGML_F32x8_ADD
+#define GGML_F32_VEC_MUL    GGML_F32x8_MUL
+#define GGML_F32_VEC_REDUCE GGML_F32x8_REDUCE
 
-inline static void ggml_vec_set_i32(const int n, int32_t * x, const int32_t v) { for (int i = 0; i < n; ++i) x[i] = v; }
+// F16 LASX
 
-inline static void ggml_vec_set_f16(const int n, ggml_fp16_t * x, const int32_t v) { for (int i = 0; i < n; ++i) x[i] = v; }
+#define GGML_F16_STEP 32
+#define GGML_F16_EPR  8
 
-inline static void ggml_vec_add_f32 (const int n, float * z, const float * x, const float * y) { for (int i = 0; i < n; ++i) z[i]  = x[i] + y[i]; }
-inline static void ggml_vec_add1_f32(const int n, float * z, const float * x, const float   v) { for (int i = 0; i < n; ++i) z[i]  = x[i] + v;    }
-inline static void ggml_vec_acc_f32 (const int n, float * y, const float * x)                  { for (int i = 0; i < n; ++i) y[i] += x[i];        }
-inline static void ggml_vec_acc1_f32(const int n, float * y, const float   v)                  { for (int i = 0; i < n; ++i) y[i] += v;           }
-inline static void ggml_vec_sub_f32 (const int n, float * z, const float * x, const float * y) { for (int i = 0; i < n; ++i) z[i]  = x[i] - y[i]; }
-inline static void ggml_vec_set_f32 (const int n, float * x, const float   v)                  { for (int i = 0; i < n; ++i) x[i]  = v;           }
-inline static void ggml_vec_cpy_f32 (const int n, float * y, const float * x)                  { for (int i = 0; i < n; ++i) y[i]  = x[i];        }
-inline static void ggml_vec_neg_f32 (const int n, float * y, const float * x)                  { for (int i = 0; i < n; ++i) y[i]  = -x[i];       }
-inline static void ggml_vec_mul_f32 (const int n, float * z, const float * x, const float * y) { for (int i = 0; i < n; ++i) z[i]  = x[i]*y[i];   }
-inline static void ggml_vec_div_f32 (const int n, float * z, const float * x, const float * y) { for (int i = 0; i < n; ++i) z[i]  = x[i]/y[i];   }
+// F16 arithmetic is not supported by AVX, so we use F32 instead
 
-static void ggml_vec_dot_f32(const int n, float * restrict s, const float * restrict x, const float * restrict y) {
-#ifdef GGML_SIMD
-    float sumf = 0.0f;
-    const int np = (n & ~(GGML_F32_STEP - 1));
+#define GGML_F32Cx8          __m256
+#define GGML_F32Cx8_ZERO    (__m256)__lasx_xvldi(0)
+#define GGML_F32Cx8_SET1(x) (__m256)__lasx_xvreplgr2vr_w((x))
 
-    GGML_F32_VEC sum[GGML_F32_ARR] = { GGML_F32_VEC_ZERO };
+static inline __m256 __lasx_f32cx8_load(const ggml_fp16_t * x) {
+    float tmp[8];
 
-    GGML_F32_VEC ax[GGML_F32_ARR];
-    GGML_F32_VEC ay[GGML_F32_ARR];
+    for (int i = 0; i < 8; i++) {
+        tmp[i] = GGML_FP16_TO_FP32(x[i]);
+    }
 
-    for (int i = 0; i < np; i += GGML_F32_STEP) {
-        for (int j = 0; j < GGML_F32_ARR; j++) {
-            ax[j] = GGML_F32_VEC_LOAD(x + i + j*GGML_F32_EPR);
-            ay[j] = GGML_F32_VEC_LOAD(y + i + j*GGML_F32_EPR);
+    return (__m256)__lasx_xvld(tmp, 0);
+}
+static inline void __lasx_f32cx8_store(ggml_fp16_t * x, __m256 y) {
+    float arr[8];
 
-            sum[j] = GGML_F32_VEC_FMA(sum[j], ax[j], ay[j]);
-        }
+    __lasx_xvst(y, arr, 0);
+
+    for (int i = 0; i < 8; i++) {
+        x[i] = GGML_FP32_TO_FP16(arr[i]);
     }
+}
+#define GGML_F32Cx8_LOAD(x)     __lasx_f32cx8_load(x)
+#define GGML_F32Cx8_STORE(x, y) __lasx_f32cx8_store(x, y)
 
-    // reduce sum0..sum3 to sum0
-    GGML_F32_VEC_REDUCE(sumf, sum);
+#define GGML_F32Cx8_FMA         GGML_F32x8_FMA
+#define GGML_F32Cx8_ADD         __lasx_xvfadd_s
+#define GGML_F32Cx8_MUL         __lasx_xvfmul_s
+#define GGML_F32Cx8_REDUCE      GGML_F32x8_REDUCE
 
-    // leftovers
-    for (int i = np; i < n; ++i) {
-        sumf += x[i]*y[i];
-    }
-#else
-    // scalar
-    ggml_float sumf = 0.0;
-    for (int i = 0; i < n; ++i) {
-        sumf += (ggml_float)(x[i]*y[i]);
-    }
-#endif
+#define GGML_F16_VEC                GGML_F32Cx8
+#define GGML_F16_VEC_ZERO           GGML_F32Cx8_ZERO
+#define GGML_F16_VEC_SET1           GGML_F32Cx8_SET1
+#define GGML_F16_VEC_LOAD(p, i)     GGML_F32Cx8_LOAD(p)
+#define GGML_F16_VEC_STORE(p, r, i) GGML_F32Cx8_STORE(p, r[i])
+#define GGML_F16_VEC_FMA            GGML_F32Cx8_FMA
+#define GGML_F16_VEC_ADD            GGML_F32Cx8_ADD
+#define GGML_F16_VEC_MUL            GGML_F32Cx8_MUL
+#define GGML_F16_VEC_REDUCE         GGML_F32Cx8_REDUCE
 
-    *s = sumf;
+#elif defined(__loongarch_sx)
+
+#define GGML_SIMD
+
+// F32 LSX
+
+#define GGML_F32_STEP 32
+#define GGML_F32_EPR  4
+
+#define GGML_F32x4         __m128
+#define GGML_F32x4_ZERO    __lsx_vldi(0)
+#define GGML_F32x4_SET1(x) __lsx_vinsgr2vr_w(__lsx_vldi(0),(x), 0)
+#define GGML_F32x4_LOAD(x) __lsx_vld((x), 0)
+#define GGML_F32x4_STORE((x),(y))   __lsx_vst((y), (x), 0)
+#define GGML_F32x4_FMA(a, b, c) __lsx_vfmadd_s(b, c, a)
+#define GGML_F32x4_ADD     __lsx_vfadd_s
+#define GGML_F32x4_MUL     __lsx_vfmul_s
+#define GGML_F32x4_REDUCE(res, x)                                 \
+{                                                                 \
+    int offset = GGML_F32_ARR >> 1;                               \
+    for (int i = 0; i < offset; ++i) {                            \
+        x[i] = __lsx_vfadd_s(x[i], x[offset+i]);                     \
+    }                                                             \
+    offset >>= 1;                                                 \
+    for (int i = 0; i < offset; ++i) {                            \
+        x[i] = __lsx_vfadd_s(x[i], x[offset+i]);                     \
+    }                                                             \
+    offset >>= 1;                                                 \
+    for (int i = 0; i < offset; ++i) {                            \
+        x[i] = __lsx_vfadd_s(x[i], x[offset+i]);                     \
+    }                                                             \
+    __m128i tmp = __lsx_vsrli_d((__m128i)x[0], 32); \
+    tmp = (__m128i)__lsx_vfadd_s((__m128)tmp, x[0]); \
+    tmp = __lsx_vpickev_w(__lsx_vldi(0), tmp); \
+    const __m128 t0 = __lsx_vshuf4i_w(tmp, 0x88); \
+    tmp = __lsx_vsrli_d((__m128i)t0, 32); \
+    tmp = (__m128i)__lsx_vfadd_s((__m128)tmp, t0); \
+    tmp = __lsx_vpickev_w(__lsx_vldi(0), tmp); \
+    res = (ggml_float) __lsx_vpickve2gr_w(__lsx_vshuf4i_w(tmp, 0x88), 0);        \
 }
 
-static void ggml_vec_dot_f16(const int n, float * restrict s, ggml_fp16_t * restrict x, ggml_fp16_t * restrict y) {
-    ggml_float sumf = 0.0;
+#define GGML_F32_VEC        GGML_F32x4
+#define GGML_F32_VEC_ZERO   GGML_F32x4_ZERO
+#define GGML_F32_VEC_SET1   GGML_F32x4_SET1
+#define GGML_F32_VEC_LOAD   GGML_F32x4_LOAD
+#define GGML_F32_VEC_STORE  GGML_F32x4_STORE
+#define GGML_F32_VEC_FMA    GGML_F32x4_FMA
+#define GGML_F32_VEC_ADD    GGML_F32x4_ADD
+#define GGML_F32_VEC_MUL    GGML_F32x4_MUL
+#define GGML_F32_VEC_REDUCE GGML_F32x4_REDUCE
 
-#if defined(GGML_SIMD)
-    const int np = (n & ~(GGML_F16_STEP - 1));
+// F16 LSX
 
-    GGML_F16_VEC sum[GGML_F16_ARR] = { GGML_F16_VEC_ZERO };
+#define GGML_F16_STEP 32
+#define GGML_F16_EPR  4
 
-    GGML_F16_VEC ax[GGML_F16_ARR];
-    GGML_F16_VEC ay[GGML_F16_ARR];
+static inline __m128 __lsx_f16x4_load(const ggml_fp16_t * x) {
+    float tmp[4];
 
-    for (int i = 0; i < np; i += GGML_F16_STEP) {
-        for (int j = 0; j < GGML_F16_ARR; j++) {
-            ax[j] = GGML_F16_VEC_LOAD(x + i + j*GGML_F16_EPR, j);
-            ay[j] = GGML_F16_VEC_LOAD(y + i + j*GGML_F16_EPR, j);
+    tmp[0] = GGML_FP16_TO_FP32(x[0]);
+    tmp[1] = GGML_FP16_TO_FP32(x[1]);
+    tmp[2] = GGML_FP16_TO_FP32(x[2]);
+    tmp[3] = GGML_FP16_TO_FP32(x[3]);
 
-            sum[j] = GGML_F16_VEC_FMA(sum[j], ax[j], ay[j]);
-        }
-    }
+    return __lsx_vld(tmp, 0);
+}
 
-    // reduce sum0..sum3 to sum0
-    GGML_F16_VEC_REDUCE(sumf, sum);
+static inline void __lsx_f16x4_store(ggml_fp16_t * x, __m128 y) {
+    float arr[4];
 
-    // leftovers
-    for (int i = np; i < n; ++i) {
-        sumf += (ggml_float)(GGML_FP16_TO_FP32(x[i])*GGML_FP16_TO_FP32(y[i]));
-    }
-#else
-    for (int i = 0; i < n; ++i) {
-        sumf += (ggml_float)(GGML_FP16_TO_FP32(x[i])*GGML_FP16_TO_FP32(y[i]));
-    }
-#endif
+    __lsx_vst(y, arr, 0);
 
-    *s = sumf;
+    x[0] = GGML_FP32_TO_FP16(arr[0]);
+    x[1] = GGML_FP32_TO_FP16(arr[1]);
+    x[2] = GGML_FP32_TO_FP16(arr[2]);
+    x[3] = GGML_FP32_TO_FP16(arr[3]);
 }
 
-// compute GGML_VEC_DOT_UNROLL dot products at once
-// xs - x row stride in bytes
-inline static void ggml_vec_dot_f16_unroll(const int n, const int xs, float * restrict s, void * restrict xv, ggml_fp16_t * restrict y) {
-    ggml_float sumf[GGML_VEC_DOT_UNROLL] = { 0.0 };
+#define GGML_F32Cx4             __m128
+#define GGML_F32Cx4_ZERO        __lsx_vldi(0)
+#define GGML_F32Cx4_SET1(x)     __lsx_vinsgr2vr_w(__lsx_vldi(0),(x), 0)
+#define GGML_F32Cx4_LOAD(x)     __lsx_f16x4_load(x)
+#define GGML_F32Cx4_STORE(x, y) __lsx_f16x4_store(x, y)
+#define GGML_F32Cx4_FMA         GGML_F32x4_FMA
+#define GGML_F32Cx4_ADD         __lsx_vfadd_s
+#define GGML_F32Cx4_MUL         __lsx_vfmul_s
+#define GGML_F32Cx4_REDUCE      GGML_F32x4_REDUCE
 
-    ggml_fp16_t * restrict x[GGML_VEC_DOT_UNROLL];
+#define GGML_F16_VEC                 GGML_F32Cx4
+#define GGML_F16_VEC_ZERO            GGML_F32Cx4_ZERO
+#define GGML_F16_VEC_SET1            GGML_F32Cx4_SET1
+#define GGML_F16_VEC_LOAD(p, i)      GGML_F32Cx4_LOAD(p)
+#define GGML_F16_VEC_STORE(p, r, i)  GGML_F32Cx4_STORE(p, r[i])
+#define GGML_F16_VEC_FMA             GGML_F32Cx4_FMA
+#define GGML_F16_VEC_ADD             GGML_F32Cx4_ADD
+#define GGML_F16_VEC_MUL             GGML_F32Cx4_MUL
+#define GGML_F16_VEC_REDUCE          GGML_F32Cx4_REDUCE
 
-    for (int i = 0; i < GGML_VEC_DOT_UNROLL; ++i) {
-        x[i] = (ggml_fp16_t *) ((char *) xv + i*xs);
-    }
+#endif
 
-#if defined(GGML_SIMD)
-    const int np = (n & ~(GGML_F16_STEP - 1));
+// GGML_F32_ARR / GGML_F16_ARR
+//   number of registers to use per step
+#ifdef GGML_SIMD
+#define GGML_F32_ARR (GGML_F32_STEP/GGML_F32_EPR)
+#define GGML_F16_ARR (GGML_F16_STEP/GGML_F16_EPR)
+#endif
 
-    GGML_F16_VEC sum[GGML_VEC_DOT_UNROLL][GGML_F16_ARR] = { { GGML_F16_VEC_ZERO } };
+//
+// ggml object
+//
 
-    GGML_F16_VEC ax[GGML_F16_ARR];
-    GGML_F16_VEC ay[GGML_F16_ARR];
+struct ggml_object {
+    size_t offs;
+    size_t size;
 
-    for (int i = 0; i < np; i += GGML_F16_STEP) {
-        for (int j = 0; j < GGML_F16_ARR; j++) {
-            ay[j] = GGML_F16_VEC_LOAD(y + i + j*GGML_F16_EPR, j);
+    struct ggml_object * next;
 
-            for (int k = 0; k < GGML_VEC_DOT_UNROLL; ++k) {
-                ax[j] = GGML_F16_VEC_LOAD(x[k] + i + j*GGML_F16_EPR, j);
+    enum ggml_object_type type;
 
-                sum[k][j] = GGML_F16_VEC_FMA(sum[k][j], ax[j], ay[j]);
-            }
-        }
-    }
+    char padding[4];
+};
 
-    // reduce sum0..sum3 to sum0
-    for (int k = 0; k < GGML_VEC_DOT_UNROLL; ++k) {
-        GGML_F16_VEC_REDUCE(sumf[k], sum[k]);
-    }
+static const size_t GGML_OBJECT_SIZE = sizeof(struct ggml_object);
 
-    // leftovers
+//
+// ggml context
+//
+
+struct ggml_context {
+    size_t mem_size;
+    void* mem_buffer;
+    bool   mem_buffer_owned;
+    bool   no_alloc;
+    bool   no_alloc_save; // this is used to save the no_alloc state when using scratch buffers
+
+    int    n_objects;
+
+    struct ggml_object * objects_begin;
+    struct ggml_object * objects_end;
+
+    struct ggml_scratch scratch;
+    struct ggml_scratch scratch_save;
+};
+
+struct ggml_context_container {
+    bool used;
+
+    struct ggml_context context;
+};
+
+//
+// Threading defs
+//
+
+typedef pthread_t          ggml_thread_t;
+
+#if defined(_WIN32)
+
+typedef CONDITION_VARIABLE ggml_cond_t;
+typedef SRWLOCK            ggml_mutex_t;
+
+#define ggml_mutex_init(m)   InitializeSRWLock(m)
+#define ggml_mutex_destroy(m)
+#define ggml_mutex_lock(m)   AcquireSRWLockExclusive(m)
+#define ggml_mutex_unlock(m) ReleaseSRWLockExclusive(m)
+#define ggml_mutex_lock_shared(m)   AcquireSRWLockShared(m)
+#define ggml_mutex_unlock_shared(m) ReleaseSRWLockShared(m)
+
+#define ggml_cond_init(c)    InitializeConditionVariable(c)
+#define ggml_cond_destroy(c)
+#define ggml_cond_wait(c, m) SleepConditionVariableSRW(c, m, INFINITE, CONDITION_VARIABLE_LOCKMODE_SHARED)
+#define ggml_cond_broadcast(c) WakeAllConditionVariable(c)
+
+#define ggml_thread_create pthread_create
+#define ggml_thread_join   pthread_join
+
+#else
+
+typedef pthread_cond_t     ggml_cond_t;
+typedef pthread_mutex_t    ggml_mutex_t;
+
+#define ggml_mutex_init(m)          pthread_mutex_init(m, NULL)
+#define ggml_mutex_destroy(m)       pthread_mutex_destroy(m)
+#define ggml_mutex_lock(m)          pthread_mutex_lock(m)
+#define ggml_mutex_unlock(m)        pthread_mutex_unlock(m)
+#define ggml_mutex_lock_shared(m)   pthread_mutex_lock(m)
+#define ggml_mutex_unlock_shared(m) pthread_mutex_unlock(m)
+
+#define ggml_lock_init(x)    UNUSED(x)
+#define ggml_lock_destroy(x) UNUSED(x)
+#if defined(__x86_64__) || (defined(_MSC_VER) && defined(_M_AMD64))
+#define ggml_lock_lock(x)    _mm_pause()
+#else
+#define ggml_lock_lock(x)    UNUSED(x)
+#endif
+#define ggml_lock_unlock(x)  UNUSED(x)
+
+#define GGML_LOCK_INITIALIZER 0
+#define ggml_cond_init(c)      pthread_cond_init(c, NULL)
+#define ggml_cond_destroy(c)   pthread_cond_destroy(c)
+#define ggml_cond_wait(c, m)   pthread_cond_wait(c, m)
+#define ggml_cond_broadcast(c) pthread_cond_broadcast(c)
+
+#define ggml_thread_create pthread_create
+#define ggml_thread_join   pthread_join
+
+#endif
+
+// Threadpool def
+struct ggml_threadpool {
+    ggml_mutex_t mutex;       // mutex for cond.var
+    ggml_cond_t  cond;        // cond.var for waiting for new work
+
+    struct ggml_cgraph * cgraph;
+    struct ggml_cplan  * cplan;
+
+    // synchronization primitives
+    atomic_int n_graph;       // incremented when there is work to be done (i.e each graph)
+    atomic_int GGML_CACHE_ALIGN n_barrier;
+    atomic_int GGML_CACHE_ALIGN n_barrier_passed;
+    atomic_int current_chunk; // currently processing chunk during Mat_Mul, shared between all the threads.
+
+    // these are atomic as an annotation for thread-sanitizer
+    atomic_bool stop;         // Used for stopping the threadpool altogether
+    atomic_bool pause;        // Used for pausing the threadpool or individual threads
+    atomic_bool abort;        // Used for aborting processing of a graph
+
+    struct ggml_compute_state * workers;   // per thread state
+    int          n_threads_max; // number of threads in the pool
+    atomic_int   n_threads_cur; // number of threads used in the current graph
+
+    int32_t      prio;        // Scheduling priority
+    uint32_t     poll;        // Polling level (0 - no polling)
+
+    enum ggml_status ec;
+};
+
+// Per-thread state
+struct ggml_compute_state {
+#ifndef GGML_USE_OPENMP
+    ggml_thread_t thrd;
+    bool cpumask[GGML_MAX_N_THREADS];
+    int  last_graph;
+    bool pending;
+#endif
+    struct ggml_threadpool * threadpool;
+    int ith;
+};
+
+struct ggml_compute_params {
+    // ith = thread index, nth = number of threads
+    int ith, nth;
+
+    // work buffer for all threads
+    size_t wsize;
+    void * wdata;
+
+    struct ggml_threadpool * threadpool;
+};
+
+//
+// fundamental operations
+//
+
+inline static void ggml_vec_set_i8(const int n, int8_t * x, const int8_t v) { for (int i = 0; i < n; ++i) x[i] = v; }
+
+inline static void ggml_vec_set_i16(const int n, int16_t * x, const int16_t v) { for (int i = 0; i < n; ++i) x[i] = v; }
+
+inline static void ggml_vec_set_i32(const int n, int32_t * x, const int32_t v) { for (int i = 0; i < n; ++i) x[i] = v; }
+
+inline static void ggml_vec_set_f16(const int n, ggml_fp16_t * x, const int32_t v) { for (int i = 0; i < n; ++i) x[i] = v; }
+
+inline static void ggml_vec_set_bf16(const int n, ggml_bf16_t * x, const ggml_bf16_t v) { for (int i = 0; i < n; ++i) x[i] = v; }
+
+inline static void ggml_vec_add_f32 (const int n, float * z, const float * x, const float * y) { for (int i = 0; i < n; ++i) z[i]  = x[i] + y[i]; }
+inline static void ggml_vec_add1_f32(const int n, float * z, const float * x, const float   v) { for (int i = 0; i < n; ++i) z[i]  = x[i] + v;    }
+inline static void ggml_vec_acc_f32 (const int n, float * y, const float * x)                  { for (int i = 0; i < n; ++i) y[i] += x[i];        }
+inline static void ggml_vec_acc1_f32(const int n, float * y, const float   v)                  { for (int i = 0; i < n; ++i) y[i] += v;           }
+inline static void ggml_vec_sub_f32 (const int n, float * z, const float * x, const float * y) { for (int i = 0; i < n; ++i) z[i]  = x[i] - y[i]; }
+inline static void ggml_vec_set_f32 (const int n, float * x, const float   v)                  { for (int i = 0; i < n; ++i) x[i]  = v;           }
+inline static void ggml_vec_cpy_f32 (const int n, float * y, const float * x)                  { for (int i = 0; i < n; ++i) y[i]  = x[i];        }
+inline static void ggml_vec_neg_f32 (const int n, float * y, const float * x)                  { for (int i = 0; i < n; ++i) y[i]  = -x[i];       }
+inline static void ggml_vec_mul_f32 (const int n, float * z, const float * x, const float * y) { for (int i = 0; i < n; ++i) z[i]  = x[i]*y[i];   }
+inline static void ggml_vec_div_f32 (const int n, float * z, const float * x, const float * y) { for (int i = 0; i < n; ++i) z[i]  = x[i]/y[i];   }
+
+static void ggml_vec_dot_f32(int n, float * restrict s, size_t bs, const float * restrict x, size_t bx, const float * restrict y, size_t by, int nrc) {
+   assert(nrc == 1);
+   UNUSED(nrc);
+   UNUSED(bx);
+   UNUSED(by);
+   UNUSED(bs);
+
+#if defined(GGML_SIMD)
+    float sumf = 0.0f;
+    const int np = (n & ~(GGML_F32_STEP - 1));
+
+    GGML_F32_VEC sum[GGML_F32_ARR] = { GGML_F32_VEC_ZERO };
+
+    GGML_F32_VEC ax[GGML_F32_ARR];
+    GGML_F32_VEC ay[GGML_F32_ARR];
+
+    for (int i = 0; i < np; i += GGML_F32_STEP) {
+        for (int j = 0; j < GGML_F32_ARR; j++) {
+            ax[j] = GGML_F32_VEC_LOAD(x + i + j*GGML_F32_EPR);
+            ay[j] = GGML_F32_VEC_LOAD(y + i + j*GGML_F32_EPR);
+
+            sum[j] = GGML_F32_VEC_FMA(sum[j], ax[j], ay[j]);
+        }
+    }
+
+    // reduce sum0..sum3 to sum0
+    GGML_F32_VEC_REDUCE(sumf, sum);
+
+    // leftovers
+    for (int i = np; i < n; ++i) {
+        sumf += x[i]*y[i];
+    }
+#else
+    // scalar
+    ggml_float sumf = 0.0;
+    for (int i = 0; i < n; ++i) {
+        sumf += (ggml_float)(x[i]*y[i]);
+    }
+#endif
+
+    *s = sumf;
+}
+
+static void ggml_vec_dot_bf16(int n, float * restrict s, size_t bs, ggml_bf16_t * restrict x, size_t bx, ggml_bf16_t * restrict y, size_t by, int nrc) {
+    assert(nrc == 1);
+    UNUSED(nrc);
+    UNUSED(bx);
+    UNUSED(by);
+    UNUSED(bs);
+    int i = 0;
+    ggml_float sumf = 0;
+
+#if defined(__AVX512BF16__)
+    __m512 c1 = _mm512_setzero_ps();
+    __m512 c2 = _mm512_setzero_ps();
+    for (; i + 64 <= n; i += 64) {
+        c1 = _mm512_dpbf16_ps(c1, m512bh(_mm512_loadu_si512((x + i))),
+                             m512bh(_mm512_loadu_si512((y + i))));
+        c2 = _mm512_dpbf16_ps(c2, m512bh(_mm512_loadu_si512((x + i + 32))),
+                             m512bh(_mm512_loadu_si512((y + i + 32))));
+    }
+    sumf += (ggml_float)_mm512_reduce_add_ps(c1);
+    sumf += (ggml_float)_mm512_reduce_add_ps(c2);
+
+#elif defined(__AVX512F__)
+#define LOAD(p) _mm512_castsi512_ps(_mm512_slli_epi32(_mm512_cvtepu16_epi32(_mm256_loadu_si256((const __m256i *)(p))), 16))
+    __m512 c1 = _mm512_setzero_ps();
+    __m512 c2 = _mm512_setzero_ps();
+    for (; i + 32 <= n; i += 32) {
+        c1 = _mm512_add_ps(_mm512_mul_ps(LOAD(x + i), LOAD(y + i)), c1);
+        c2 = _mm512_add_ps(_mm512_mul_ps(LOAD(x + i + 16), LOAD(y + i + 16)), c2);
+    }
+    sumf += (ggml_float)_mm512_reduce_add_ps(c1);
+    sumf += (ggml_float)_mm512_reduce_add_ps(c2);
+
+#undef LOAD
+#elif defined(__AVX2__)
+#define LOAD(p) _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_cvtepu16_epi32(_mm_loadu_si128((const __m128i *)(p))), 16))
+    __m256 c1 = _mm256_setzero_ps();
+    __m256 c2 = _mm256_setzero_ps();
+    __m256 c3 = _mm256_setzero_ps();
+    __m256 c4 = _mm256_setzero_ps();
+    for (; i + 32 <= n; i += 32) {
+        c1 = _mm256_add_ps(_mm256_mul_ps(LOAD(x + i), LOAD(y + i)), c1);
+        c2 = _mm256_add_ps(_mm256_mul_ps(LOAD(x + i + 8), LOAD(y + i + 8)), c2);
+        c3 = _mm256_add_ps(_mm256_mul_ps(LOAD(x + i + 16), LOAD(y + i + 16)), c3);
+        c4 = _mm256_add_ps(_mm256_mul_ps(LOAD(x + i + 24), LOAD(y + i + 24)), c4);
+    }
+    __m128 g;
+    c1 = _mm256_add_ps(_mm256_add_ps(c1, c3),
+                       _mm256_add_ps(c2, c4));
+    g = _mm_add_ps(_mm256_extractf128_ps(c1, 1),
+                   _mm256_castps256_ps128(c1));
+    g = _mm_add_ps(g, _mm_movehl_ps(g, g));
+    g = _mm_add_ss(g, _mm_movehdup_ps(g));
+    sumf += (ggml_float)_mm_cvtss_f32(g);
+
+#undef LOAD
+#endif
+
+    for (; i < n; ++i) {
+        sumf += (ggml_float)(GGML_BF16_TO_FP32(x[i]) *
+                             GGML_BF16_TO_FP32(y[i]));
+    }
+    *s = sumf;
+}
+
+static void ggml_vec_dot_f16(int n, float * restrict s, size_t bs, ggml_fp16_t * restrict x, size_t bx, ggml_fp16_t * restrict y, size_t by, int nrc) {
+    assert(nrc == 1);
+    UNUSED(nrc);
+    UNUSED(bx);
+    UNUSED(by);
+    UNUSED(bs);
+
+    ggml_float sumf = 0.0;
+
+#if defined(GGML_SIMD)
+    const int np = (n & ~(GGML_F16_STEP - 1));
+
+    GGML_F16_VEC sum[GGML_F16_ARR] = { GGML_F16_VEC_ZERO };
+
+    GGML_F16_VEC ax[GGML_F16_ARR];
+    GGML_F16_VEC ay[GGML_F16_ARR];
+
+    for (int i = 0; i < np; i += GGML_F16_STEP) {
+        for (int j = 0; j < GGML_F16_ARR; j++) {
+            ax[j] = GGML_F16_VEC_LOAD(x + i + j*GGML_F16_EPR, j);
+            ay[j] = GGML_F16_VEC_LOAD(y + i + j*GGML_F16_EPR, j);
+
+            sum[j] = GGML_F16_VEC_FMA(sum[j], ax[j], ay[j]);
+        }
+    }
+
+    // reduce sum0..sum3 to sum0
+    GGML_F16_VEC_REDUCE(sumf, sum);
+
+    // leftovers
+    for (int i = np; i < n; ++i) {
+        sumf += (ggml_float)(GGML_FP16_TO_FP32(x[i])*GGML_FP16_TO_FP32(y[i]));
+    }
+#else
+    for (int i = 0; i < n; ++i) {
+        sumf += (ggml_float)(GGML_FP16_TO_FP32(x[i])*GGML_FP16_TO_FP32(y[i]));
+    }
+#endif
+
+    *s = sumf;
+}
+
+// compute GGML_VEC_DOT_UNROLL dot products at once
+// xs - x row stride in bytes
+inline static void ggml_vec_dot_f16_unroll(const int n, const int xs, float * restrict s, void * restrict xv, ggml_fp16_t * restrict y) {
+    ggml_float sumf[GGML_VEC_DOT_UNROLL] = { 0.0 };
+
+    ggml_fp16_t * restrict x[GGML_VEC_DOT_UNROLL];
+
+    for (int i = 0; i < GGML_VEC_DOT_UNROLL; ++i) {
+        x[i] = (ggml_fp16_t *) ((char *) xv + i*xs);
+    }
+
+#if defined(GGML_SIMD)
+    const int np = (n & ~(GGML_F16_STEP - 1));
+
+    GGML_F16_VEC sum[GGML_VEC_DOT_UNROLL][GGML_F16_ARR] = { { GGML_F16_VEC_ZERO } };
+
+    GGML_F16_VEC ax[GGML_F16_ARR];
+    GGML_F16_VEC ay[GGML_F16_ARR];
+
+    for (int i = 0; i < np; i += GGML_F16_STEP) {
+        for (int j = 0; j < GGML_F16_ARR; j++) {
+            ay[j] = GGML_F16_VEC_LOAD(y + i + j*GGML_F16_EPR, j);
+
+            for (int k = 0; k < GGML_VEC_DOT_UNROLL; ++k) {
+                ax[j] = GGML_F16_VEC_LOAD(x[k] + i + j*GGML_F16_EPR, j);
+
+                sum[k][j] = GGML_F16_VEC_FMA(sum[k][j], ax[j], ay[j]);
+            }
+        }
+    }
+
+    // reduce sum0..sum3 to sum0
+    for (int k = 0; k < GGML_VEC_DOT_UNROLL; ++k) {
+        GGML_F16_VEC_REDUCE(sumf[k], sum[k]);
+    }
+
+    // leftovers
     for (int i = np; i < n; ++i) {
         for (int j = 0; j < GGML_VEC_DOT_UNROLL; ++j) {
             sumf[j] += (ggml_float)(GGML_FP16_TO_FP32(x[j][i])*GGML_FP16_TO_FP32(y[i]));
@@ -1302,6 +2375,37 @@ inline static void ggml_vec_mad_f32(const int n, float * restrict y, const float
 #endif
 }
 
+inline static void ggml_vec_mad_f16(const int n, ggml_fp16_t * restrict y, const ggml_fp16_t * restrict x, const float v) {
+#if defined(GGML_SIMD)
+    const int np = (n & ~(GGML_F16_STEP - 1));
+
+    GGML_F16_VEC vx = GGML_F16_VEC_SET1(v);
+
+    GGML_F16_VEC ax[GGML_F16_ARR];
+    GGML_F16_VEC ay[GGML_F16_ARR];
+
+    for (int i = 0; i < np; i += GGML_F16_STEP) {
+        for (int j = 0; j < GGML_F16_ARR; j++) {
+            ax[j] = GGML_F16_VEC_LOAD(x + i + j*GGML_F16_EPR, j);
+            ay[j] = GGML_F16_VEC_LOAD(y + i + j*GGML_F16_EPR, j);
+            ay[j] = GGML_F16_VEC_FMA(ay[j], ax[j], vx);
+
+            GGML_F16_VEC_STORE(y + i + j*GGML_F16_EPR, ay, j);
+        }
+    }
+
+    // leftovers
+    for (int i = np; i < n; ++i) {
+        y[i] = GGML_FP32_TO_FP16(GGML_FP16_TO_FP32(y[i]) + GGML_FP16_TO_FP32(x[i])*v);
+    }
+#else
+    // scalar
+    for (int i = 0; i < n; ++i) {
+        y[i] = GGML_FP32_TO_FP16(GGML_FP16_TO_FP32(y[i]) + GGML_FP16_TO_FP32(x[i])*v);
+    }
+#endif
+}
+
 // xs and vs are byte strides of x and v
 inline static void ggml_vec_mad_f32_unroll(const int n, const int xs, const int vs, float * restrict y, const float * restrict xv, const float * restrict vv) {
 
@@ -1386,23 +2490,59 @@ inline static void ggml_vec_scale_f32(const int n, float * y, const float   v) {
 #endif
 }
 
-inline static void ggml_vec_norm_f32 (const int n, float * s, const float * x) { ggml_vec_dot_f32(n, s, x, x); *s = sqrtf(*s);   }
-inline static void ggml_vec_sqr_f32  (const int n, float * y, const float * x) { for (int i = 0; i < n; ++i) y[i] = x[i]*x[i];   }
-inline static void ggml_vec_sqrt_f32 (const int n, float * y, const float * x) { for (int i = 0; i < n; ++i) y[i] = sqrtf(x[i]); }
-inline static void ggml_vec_log_f32  (const int n, float * y, const float * x) { for (int i = 0; i < n; ++i) y[i] = logf(x[i]);   }
-inline static void ggml_vec_abs_f32  (const int n, float * y, const float * x) { for (int i = 0; i < n; ++i) y[i] = fabsf(x[i]); }
-inline static void ggml_vec_sgn_f32  (const int n, float * y, const float * x) { for (int i = 0; i < n; ++i) y[i] = (x[i] > 0.f) ? 1.f : ((x[i] < 0.f) ? -1.f : 0.f); }
-inline static void ggml_vec_step_f32 (const int n, float * y, const float * x) { for (int i = 0; i < n; ++i) y[i] = (x[i] > 0.f) ? 1.f : 0.f; }
-inline static void ggml_vec_tanh_f32 (const int n, float * y, const float * x) { for (int i = 0; i < n; ++i) y[i] = tanhf(x[i]);  }
-inline static void ggml_vec_elu_f32  (const int n, float * y, const float * x) { for (int i = 0; i < n; ++i) y[i] = (x[i] > 0.f) ? x[i] : expf(x[i])-1; }
-inline static void ggml_vec_relu_f32 (const int n, float * y, const float * x) { for (int i = 0; i < n; ++i) y[i] = (x[i] > 0.f) ? x[i] : 0.f; }
-inline static void ggml_vec_leaky_relu_f32 (const int n, float * y, const float * x, const float ns) { for (int i = 0; i < n; ++i) y[i] = ((x[i] > 0.f) ? x[i] : 0.f) + ns * ((x[i] < 0.0f) ? x[i] : 0.f); }
+inline static void ggml_vec_scale_f16(const int n, ggml_fp16_t * y, const float v) {
+#if defined(GGML_SIMD)
+    const int np = (n & ~(GGML_F16_STEP - 1));
 
-static const float GELU_COEF_A     = 0.044715f;
-static const float GELU_QUICK_COEF = -1.702f;
-static const float SQRT_2_OVER_PI  = 0.79788456080286535587989211986876f;
+    GGML_F16_VEC vx = GGML_F16_VEC_SET1(v);
 
-inline static float ggml_gelu_f32(float x) {
+    GGML_F16_VEC ay[GGML_F16_ARR];
+
+    for (int i = 0; i < np; i += GGML_F16_STEP) {
+        for (int j = 0; j < GGML_F16_ARR; j++) {
+            ay[j] = GGML_F16_VEC_LOAD(y + i + j*GGML_F16_EPR, j);
+            ay[j] = GGML_F16_VEC_MUL(ay[j], vx);
+
+            GGML_F16_VEC_STORE(y + i + j*GGML_F16_EPR, ay, j);
+        }
+    }
+
+    // leftovers
+    for (int i = np; i < n; ++i) {
+        y[i] = GGML_FP32_TO_FP16(GGML_FP16_TO_FP32(y[i])*v);
+    }
+#else
+    // scalar
+    for (int i = 0; i < n; ++i) {
+        y[i] = GGML_FP32_TO_FP16(GGML_FP16_TO_FP32(y[i])*v);
+    }
+#endif
+}
+
+inline static void ggml_vec_norm_f32 (const int n, float * s, const float * x) { ggml_vec_dot_f32(n, s, 0, x, 0, x, 0, 1); *s = sqrtf(*s);   }
+inline static void ggml_vec_sqr_f32  (const int n, float * y, const float * x) { for (int i = 0; i < n; ++i) y[i] = x[i]*x[i];   }
+inline static void ggml_vec_sqrt_f32 (const int n, float * y, const float * x) { for (int i = 0; i < n; ++i) y[i] = sqrtf(x[i]); }
+inline static void ggml_vec_log_f32  (const int n, float * y, const float * x) { for (int i = 0; i < n; ++i) y[i] = logf(x[i]);  }
+inline static void ggml_vec_sin_f32  (const int n, float * y, const float * x) { for (int i = 0; i < n; ++i) y[i] = sinf(x[i]);  }
+inline static void ggml_vec_cos_f32  (const int n, float * y, const float * x) { for (int i = 0; i < n; ++i) y[i] = cosf(x[i]);  }
+inline static void ggml_vec_abs_f32  (const int n, float * y, const float * x) { for (int i = 0; i < n; ++i) y[i] = fabsf(x[i]); }
+inline static void ggml_vec_sgn_f32  (const int n, float * y, const float * x) { for (int i = 0; i < n; ++i) y[i] = (x[i] > 0.f) ? 1.f : ((x[i] < 0.f) ? -1.f : 0.f); }
+inline static void ggml_vec_step_f32 (const int n, float * y, const float * x) { for (int i = 0; i < n; ++i) y[i] = (x[i] > 0.f) ? 1.f : 0.f; }
+inline static void ggml_vec_tanh_f32 (const int n, float * y, const float * x) { for (int i = 0; i < n; ++i) y[i] = tanhf(x[i]);  }
+inline static void ggml_vec_elu_f32  (const int n, float * y, const float * x) { for (int i = 0; i < n; ++i) y[i] = (x[i] > 0.f) ? x[i] : expm1f(x[i]); }
+inline static void ggml_vec_relu_f32 (const int n, float * y, const float * x) { for (int i = 0; i < n; ++i) y[i] = (x[i] > 0.f) ? x[i] : 0.f; }
+inline static void ggml_vec_leaky_relu_f32 (const int n, float * y, const float * x, const float ns) { for (int i = 0; i < n; ++i) y[i] = ((x[i] > 0.f) ? x[i] : 0.f) + ns * ((x[i] < 0.0f) ? x[i] : 0.f); }
+inline static void ggml_vec_sigmoid_f32 (const int n, float * y, const float * x) { for (int i = 0; i < n; ++i) y[i] = 1.f / (1.f + expf(-x[i])); }
+// TODO: optimize performance
+inline static void ggml_vec_hardswish_f32 (const int n, float * y, const float * x) { for (int i = 0; i < n; ++i) y[i] = x[i] * fminf(1.0f, fmaxf(0.0f, (x[i] + 3.0f) / 6.0f)); }
+inline static void ggml_vec_hardsigmoid_f32 (const int n, float * y, const float * x) { for (int i = 0; i < n; ++i) y[i] = fminf(1.0f, fmaxf(0.0f, (x[i] + 3.0f) / 6.0f)); }
+inline static void ggml_vec_exp_f32 (const int n, float * y, const float * x) { for (int i = 0; i < n; ++i) y[i] = expf(x[i]); }
+
+static const float GELU_COEF_A     = 0.044715f;
+static const float GELU_QUICK_COEF = -1.702f;
+static const float SQRT_2_OVER_PI  = 0.79788456080286535587989211986876f;
+
+inline static float ggml_gelu_f32(float x) {
     return 0.5f*x*(1.0f + tanhf(SQRT_2_OVER_PI*x*(1.0f + GELU_COEF_A*x*x)));
 }
 
@@ -1417,9 +2557,15 @@ inline static void ggml_vec_gelu_f16(const int n, ggml_fp16_t * y, const ggml_fp
 inline static void ggml_vec_gelu_f32(const int n, float * y, const float * x) {
     uint16_t t;
     for (int i = 0; i < n; ++i) {
-        ggml_fp16_t fp16 = GGML_FP32_TO_FP16(x[i]);
-        memcpy(&t, &fp16, sizeof(uint16_t));
-        y[i] = GGML_FP16_TO_FP32(ggml_table_gelu_f16[t]);
+        if (x[i] <= -10.0f) {
+            y[i] = 0.0f;
+        } else if (x[i] >= 10.0f) {
+            y[i] = x[i];
+        } else {
+            ggml_fp16_t fp16 = GGML_FP32_TO_FP16(x[i]);
+            memcpy(&t, &fp16, sizeof(uint16_t));
+            y[i] = GGML_FP16_TO_FP32(ggml_table_gelu_f16[t]);
+        }
     }
 }
 #else
@@ -1463,52 +2609,304 @@ inline static float ggml_silu_f32(float x) {
     return x/(1.0f + expf(-x));
 }
 
-//inline static void ggml_vec_silu_f16(const int n, ggml_fp16_t * y, const ggml_fp16_t * x) {
-//    const uint16_t * i16 = (const uint16_t *) x;
-//    for (int i = 0; i < n; ++i) {
-//        y[i] = ggml_table_silu_f16[i16[i]];
-//    }
-//}
+#if __FINITE_MATH_ONLY__
+#error "some routines in ggml.c require non-finite math arithmetics -- pass -fno-finite-math-only to the compiler to fix"
+#error "ref: https://github.com/ggerganov/llama.cpp/pull/7154#issuecomment-2143844461"
+#endif
 
-#ifdef GGML_SILU_FP16
-inline static void ggml_vec_silu_f32(const int n, float * y, const float * x) {
-    uint16_t t;
-    for (int i = 0; i < n; ++i) {
-        ggml_fp16_t fp16 = GGML_FP32_TO_FP16(x[i]);
-        memcpy(&t, &fp16, sizeof(uint16_t));
-        y[i] = GGML_FP16_TO_FP32(ggml_table_silu_f16[t]);
-    }
-}
+#if defined(__ARM_NEON) && defined(__aarch64__)
+
+// adapted from arm limited optimized routine
+// the maximum error is 1.45358 plus 0.5 ulps
+// numbers above 88.38 will flush to infinity
+// numbers beneath -103.97 will flush to zero
+inline static float32x4_t ggml_v_expf(float32x4_t x) {
+    const float32x4_t r = vdupq_n_f32(0x1.8p23f);
+    const float32x4_t z = vfmaq_f32(r, x, vdupq_n_f32(0x1.715476p+0f));
+    const float32x4_t n = vsubq_f32(z, r);
+    const float32x4_t b = vfmsq_f32(vfmsq_f32(x, n, vdupq_n_f32(0x1.62e4p-1f)), n,
+                                    vdupq_n_f32(0x1.7f7d1cp-20f));
+    const uint32x4_t e = vshlq_n_u32(vreinterpretq_u32_f32(z), 23);
+    const float32x4_t k = vreinterpretq_f32_u32(vaddq_u32(e, vreinterpretq_u32_f32(vdupq_n_f32(1))));
+    const uint32x4_t c = vcagtq_f32(n, vdupq_n_f32(126));
+    const float32x4_t u = vmulq_f32(b, b);
+    const float32x4_t j = vfmaq_f32(
+        vmulq_f32(vdupq_n_f32(0x1.ffffecp-1f), b),
+        vfmaq_f32(vfmaq_f32(vdupq_n_f32(0x1.fffdb6p-2f), vdupq_n_f32(0x1.555e66p-3f), b),
+                  vfmaq_f32(vdupq_n_f32(0x1.573e2ep-5f), vdupq_n_f32(0x1.0e4020p-7f), b), u), u);
+    if (!vpaddd_u64(vreinterpretq_u64_u32(c)))
+        return vfmaq_f32(k, j, k);
+    const uint32x4_t d = vandq_u32(vclezq_f32(n), vdupq_n_u32(0x82000000));
+    const float32x4_t s1 = vreinterpretq_f32_u32(vaddq_u32(d, vdupq_n_u32(0x7f000000)));
+    const float32x4_t s2 = vreinterpretq_f32_u32(vsubq_u32(e, d));
+    return vbslq_f32(vcagtq_f32(n, vdupq_n_f32(192)), vmulq_f32(s1, s1),
+                     vbslq_f32(c, vmulq_f32(vfmaq_f32(s2, s2, j), s1), vfmaq_f32(k, k, j)));
+}
+
+// computes silu x/(1+exp(-x)) in single precision vector
+inline static float32x4_t ggml_v_silu(float32x4_t x) {
+    const float32x4_t one = vdupq_n_f32(1.0f);
+    const float32x4_t zero = vdupq_n_f32(0.0f);
+    const float32x4_t neg_x = vsubq_f32(zero, x);
+    const float32x4_t exp_neg_x = ggml_v_expf(neg_x);
+    const float32x4_t one_plus_exp_neg_x = vaddq_f32(one, exp_neg_x);
+    return vdivq_f32(x, one_plus_exp_neg_x);
+}
+
+#elif defined(__AVX512F__) && defined(__AVX512DQ__)
+
+// adapted from arm limited optimized routine
+// the maximum error is 1.45358 plus 0.5 ulps
+// numbers above 88.38 will flush to infinity
+// numbers beneath -103.97 will flush to zero
+inline static __m512 ggml_v_expf(__m512 x) {
+  const __m512 r = _mm512_set1_ps(0x1.8p23f);
+  const __m512 z = _mm512_fmadd_ps(x, _mm512_set1_ps(0x1.715476p+0f), r);
+  const __m512 n = _mm512_sub_ps(z, r);
+  const __m512 b =
+      _mm512_fnmadd_ps(n, _mm512_set1_ps(0x1.7f7d1cp-20f),
+                       _mm512_fnmadd_ps(n, _mm512_set1_ps(0x1.62e4p-1f), x));
+  const __mmask16 d =
+      _mm512_cmp_ps_mask(_mm512_abs_ps(n), _mm512_set1_ps(192), _CMP_GT_OQ);
+  const __m512 u = _mm512_mul_ps(b, b);
+  const __m512 j = _mm512_fmadd_ps(
+      _mm512_fmadd_ps(_mm512_fmadd_ps(_mm512_set1_ps(0x1.0e4020p-7f), b,
+                                      _mm512_set1_ps(0x1.573e2ep-5f)),
+                      u,
+                      _mm512_fmadd_ps(_mm512_set1_ps(0x1.555e66p-3f), b,
+                                      _mm512_set1_ps(0x1.fffdb6p-2f))),
+      u,
+      _mm512_fmadd_ps(_mm512_set1_ps(0x1.ffffecp-1f), b, _mm512_set1_ps(1.0F)));
+  const __m512 res = _mm512_scalef_ps(j, n);
+  if (_mm512_kortestz(d, d))
+    return res;
+  const __m512 zero = _mm512_setzero_ps();
+  const __m512 alt = _mm512_mask_blend_ps(
+      _mm512_cmp_ps_mask(n, zero, _CMP_LE_OQ), _mm512_set1_ps(INFINITY), zero);
+  return _mm512_mask_blend_ps(d, res, alt);
+}
+
+// computes silu x/(1+exp(-x)) in single precision vector
+inline static __m512 ggml_v_silu(__m512 x) {
+    const __m512 one = _mm512_set1_ps(1);
+    const __m512 zero = _mm512_setzero_ps();
+    const __m512 neg_x = _mm512_sub_ps(zero, x);
+    const __m512 exp_neg_x = ggml_v_expf(neg_x);
+    const __m512 one_plus_exp_neg_x = _mm512_add_ps(one, exp_neg_x);
+    return _mm512_div_ps(x, one_plus_exp_neg_x);
+}
+
+#elif defined(__AVX2__) && defined(__FMA__)
+
+// adapted from arm limited optimized routine
+// the maximum error is 1.45358 plus 0.5 ulps
+// numbers above 88.38 will flush to infinity
+// numbers beneath -103.97 will flush to zero
+inline static __m256 ggml_v_expf(__m256 x) {
+  const __m256 r = _mm256_set1_ps(0x1.8p23f);
+  const __m256 z = _mm256_fmadd_ps(x, _mm256_set1_ps(0x1.715476p+0f), r);
+  const __m256 n = _mm256_sub_ps(z, r);
+  const __m256 b = _mm256_fnmadd_ps(n, _mm256_set1_ps(0x1.7f7d1cp-20f),
+                                    _mm256_fnmadd_ps(n, _mm256_set1_ps(0x1.62e4p-1f), x));
+  const __m256i e = _mm256_slli_epi32(_mm256_castps_si256(z), 23);
+  const __m256 k = _mm256_castsi256_ps(
+      _mm256_add_epi32(e, _mm256_castps_si256(_mm256_set1_ps(1))));
+  const __m256i c = _mm256_castps_si256(
+      _mm256_cmp_ps(_mm256_andnot_ps(_mm256_set1_ps(-0.f), n),
+                    _mm256_set1_ps(126), _CMP_GT_OQ));
+  const __m256 u = _mm256_mul_ps(b, b);
+  const __m256 j = _mm256_fmadd_ps(_mm256_fmadd_ps(_mm256_fmadd_ps(_mm256_set1_ps(0x1.0e4020p-7f), b,
+                                                                   _mm256_set1_ps(0x1.573e2ep-5f)), u,
+                                                   _mm256_fmadd_ps(_mm256_set1_ps(0x1.555e66p-3f), b,
+                                                                   _mm256_set1_ps(0x1.fffdb6p-2f))),
+                                   u, _mm256_mul_ps(_mm256_set1_ps(0x1.ffffecp-1f), b));
+  if (!_mm256_movemask_ps(_mm256_castsi256_ps(c)))
+    return _mm256_fmadd_ps(j, k, k);
+  const __m256i g = _mm256_and_si256(
+      _mm256_castps_si256(_mm256_cmp_ps(n, _mm256_setzero_ps(), _CMP_LE_OQ)),
+      _mm256_set1_epi32(0x82000000u));
+  const __m256 s1 =
+      _mm256_castsi256_ps(_mm256_add_epi32(g, _mm256_set1_epi32(0x7f000000u)));
+  const __m256 s2 = _mm256_castsi256_ps(_mm256_sub_epi32(e, g));
+  const __m256i d = _mm256_castps_si256(
+      _mm256_cmp_ps(_mm256_andnot_ps(_mm256_set1_ps(-0.f), n),
+                    _mm256_set1_ps(192), _CMP_GT_OQ));
+  return _mm256_or_ps(
+      _mm256_and_ps(_mm256_castsi256_ps(d), _mm256_mul_ps(s1, s1)),
+      _mm256_andnot_ps(
+          _mm256_castsi256_ps(d),
+          _mm256_or_ps(
+              _mm256_and_ps(_mm256_castsi256_ps(c),
+                            _mm256_mul_ps(_mm256_fmadd_ps(s2, j, s2), s1)),
+              _mm256_andnot_ps(_mm256_castsi256_ps(c), _mm256_fmadd_ps(k, j, k)))));
+}
+
+// computes silu x/(1+exp(-x)) in single precision vector
+inline static __m256 ggml_v_silu(__m256 x) {
+    const __m256 one = _mm256_set1_ps(1);
+    const __m256 zero = _mm256_setzero_ps();
+    const __m256 neg_x = _mm256_sub_ps(zero, x);
+    const __m256 exp_neg_x = ggml_v_expf(neg_x);
+    const __m256 one_plus_exp_neg_x = _mm256_add_ps(one, exp_neg_x);
+    return _mm256_div_ps(x, one_plus_exp_neg_x);
+}
+
+#elif defined(__SSE2__) // __AVX2__ / __ARM_NEON
+
+#if defined(__FMA__)
+#define MADD128(x, y, z) _mm_fmadd_ps(x, y, z)
+#define NMADD128(x, y, z) _mm_fnmadd_ps(x, y, z)
 #else
-inline static void ggml_vec_silu_f32(const int n, float * y, const float * x) {
-    for (int i = 0; i < n; ++i) {
+#define MADD128(x, y, z) _mm_add_ps(_mm_mul_ps(x, y), z)
+#define NMADD128(x, y, z) _mm_sub_ps(z, _mm_mul_ps(x, y))
+#endif
+
+// adapted from arm limited optimized routine
+// the maximum error is 1.45358 plus 0.5 ulps
+// numbers above 88.38 will flush to infinity
+// numbers beneath -103.97 will flush to zero
+inline static __m128 ggml_v_expf(__m128 x) {
+    const __m128 r = _mm_set1_ps(0x1.8p23f);
+    const __m128 z = MADD128(x, _mm_set1_ps(0x1.715476p+0f), r);
+    const __m128 n = _mm_sub_ps(z, r);
+    const __m128 b =
+        NMADD128(n, _mm_set1_ps(0x1.7f7d1cp-20f), NMADD128(n, _mm_set1_ps(0x1.62e4p-1f), x));
+    const __m128i e = _mm_slli_epi32(_mm_castps_si128(z), 23);
+    const __m128 k = _mm_castsi128_ps(_mm_add_epi32(e, _mm_castps_si128(_mm_set1_ps(1))));
+    const __m128i c =
+        _mm_castps_si128(_mm_cmpgt_ps(_mm_andnot_ps(_mm_set1_ps(-0.f), n), _mm_set1_ps(126)));
+    const __m128 u = _mm_mul_ps(b, b);
+    const __m128 j =
+        MADD128(MADD128(MADD128(_mm_set1_ps(0x1.0e4020p-7f), b, _mm_set1_ps(0x1.573e2ep-5f)), u,
+                        MADD128(_mm_set1_ps(0x1.555e66p-3f), b, _mm_set1_ps(0x1.fffdb6p-2f))),
+                u, _mm_mul_ps(_mm_set1_ps(0x1.ffffecp-1f), b));
+    if (!_mm_movemask_epi8(c))
+        return MADD128(j, k, k);
+    const __m128i g = _mm_and_si128(_mm_castps_si128(_mm_cmple_ps(n, _mm_setzero_ps())),
+                                    _mm_set1_epi32(0x82000000u));
+    const __m128 s1 = _mm_castsi128_ps(_mm_add_epi32(g, _mm_set1_epi32(0x7f000000u)));
+    const __m128 s2 = _mm_castsi128_ps(_mm_sub_epi32(e, g));
+    const __m128i d =
+        _mm_castps_si128(_mm_cmpgt_ps(_mm_andnot_ps(_mm_set1_ps(-0.f), n), _mm_set1_ps(192)));
+    return _mm_or_ps(
+        _mm_and_ps(_mm_castsi128_ps(d), _mm_mul_ps(s1, s1)),
+        _mm_andnot_ps(_mm_castsi128_ps(d),
+                      _mm_or_ps(_mm_and_ps(_mm_castsi128_ps(c), _mm_mul_ps(MADD128(s2, j, s2), s1)),
+                                _mm_andnot_ps(_mm_castsi128_ps(c), MADD128(k, j, k)))));
+}
+
+// computes silu x/(1+exp(-x)) in single precision vector
+inline static __m128 ggml_v_silu(__m128 x) {
+    const __m128 one = _mm_set1_ps(1);
+    const __m128 zero = _mm_setzero_ps();
+    const __m128 neg_x = _mm_sub_ps(zero, x);
+    const __m128 exp_neg_x = ggml_v_expf(neg_x);
+    const __m128 one_plus_exp_neg_x = _mm_add_ps(one, exp_neg_x);
+    return _mm_div_ps(x, one_plus_exp_neg_x);
+}
+
+#endif // __ARM_NEON / __AVX2__ / __SSE2__
+
+static void ggml_vec_silu_f32(const int n, float * y, const float * x) {
+    int i = 0;
+#if defined(__AVX512F__) && defined(__AVX512DQ__)
+    for (; i + 15 < n; i += 16) {
+        _mm512_storeu_ps(y + i, ggml_v_silu(_mm512_loadu_ps(x + i)));
+    }
+#elif defined(__AVX2__) && defined(__FMA__)
+    for (; i + 7 < n; i += 8) {
+        _mm256_storeu_ps(y + i, ggml_v_silu(_mm256_loadu_ps(x + i)));
+    }
+#elif defined(__SSE2__)
+    for (; i + 3 < n; i += 4) {
+        _mm_storeu_ps(y + i, ggml_v_silu(_mm_loadu_ps(x + i)));
+    }
+#elif defined(__ARM_NEON) && defined(__aarch64__)
+    for (; i + 3 < n; i += 4) {
+        vst1q_f32(y + i, ggml_v_silu(vld1q_f32(x + i)));
+    }
+#endif
+    for (; i < n; ++i) {
         y[i] = ggml_silu_f32(x[i]);
     }
 }
+
+static ggml_float ggml_vec_soft_max_f32(const int n, float * y, const float * x, float max) {
+    int i = 0;
+    ggml_float sum = 0;
+#if defined(__AVX512F__) && defined(__AVX512DQ__)
+    for (; i + 15 < n; i += 16) {
+        __m512 val = ggml_v_expf(_mm512_sub_ps(_mm512_loadu_ps(x + i),
+                                               _mm512_set1_ps(max)));
+        _mm512_storeu_ps(y + i, val);
+        sum += (ggml_float)_mm512_reduce_add_ps(val);
+    }
+#elif defined(__AVX2__) && defined(__FMA__)
+    for (; i + 7 < n; i += 8) {
+        __m256 val = ggml_v_expf(_mm256_sub_ps(_mm256_loadu_ps(x + i),
+                                               _mm256_set1_ps(max)));
+        _mm256_storeu_ps(y + i, val);
+        __m128 val2 = _mm_add_ps(_mm256_extractf128_ps(val, 1),
+                                 _mm256_castps256_ps128(val));
+        val2 = _mm_add_ps(val2, _mm_movehl_ps(val2, val2));
+        val2 = _mm_add_ss(val2, _mm_movehdup_ps(val2));
+        sum += (ggml_float)_mm_cvtss_f32(val2);
+    }
+#elif defined(__SSE2__)
+    for (; i + 3 < n; i += 4) {
+        __m128 val = ggml_v_expf(_mm_sub_ps(_mm_loadu_ps(x + i),
+                                            _mm_set1_ps(max)));
+        _mm_storeu_ps(y + i, val);
+#if defined(__AVX__) || defined(__AVX2__) || defined(__AVX512F__)
+        val = _mm_add_ps(val, _mm_movehl_ps(val, val));
+        val = _mm_add_ss(val, _mm_movehdup_ps(val));
+#else
+        __m128 tmp = _mm_shuffle_ps(val, val, _MM_SHUFFLE(2, 3, 0, 1));
+        val = _mm_add_ps(val, tmp);
+        tmp = _mm_movehl_ps(tmp, val);
+        val = _mm_add_ss(val, tmp);
 #endif
+        sum += (ggml_float)_mm_cvtss_f32(val);
+    }
+#elif defined(__ARM_NEON) && defined(__aarch64__)
+    for (; i + 3 < n; i += 4) {
+        float32x4_t val = ggml_v_expf(vsubq_f32(vld1q_f32(x + i),
+                                                vdupq_n_f32(max)));
+        vst1q_f32(y + i, val);
+        sum += (ggml_float)vaddvq_f32(val);
+    }
+#endif
+    for (; i < n; ++i) {
+        float val = expf(x[i] - max);
+        sum += (ggml_float)val;
+        y[i] = val;
+    }
+    return sum;
+}
+
+static ggml_float ggml_vec_log_soft_max_f32(const int n, float * y, const float * x, float max) {
+    // log(soft_max) = log(soft_max_i / soft_max_sum) = log(soft_max_i) - log(soft_max_sum) = (logit_i - max) - log(soft_max_i)
+
+    int i = 0;
+    ggml_float sum = 0;
+    for (; i < n; ++i) {
+        float val = x[i] - max;
+        y[i] = val;
+        sum += (ggml_float)expf(val);
+    }
+    return sum = (ggml_float)logf(sum);
+}
 
 inline static float ggml_silu_backward_f32(float x, float dy) {
     const float s = 1.0f/(1.0f + expf(-x));
     return dy*s*(1.0f + x*(1.0f - s));
 }
 
-#ifdef GGML_SILU_FP16
-inline static void ggml_vec_silu_backward_f32(const int n, float * dx, const float * x, const float * dy) {
-    for (int i = 0; i < n; ++i) {
-        // we did not use x[i] to compute forward silu but its f16 equivalent
-        // take derivative at f16 of x[i]:
-        ggml_fp16_t fp16 = GGML_FP32_TO_FP16(x[i]);
-        float usedx = GGML_FP16_TO_FP32(fp16);
-        dx[i] = ggml_silu_backward_f32(usedx, dy[i]);
-    }
-}
-#else
 inline static void ggml_vec_silu_backward_f32(const int n, float * dx, const float * x, const float * dy) {
     for (int i = 0; i < n; ++i) {
         dx[i] = ggml_silu_backward_f32(x[i], dy[i]);
     }
 }
-#endif
 
 inline static void ggml_vec_sum_f32(const int n, float * s, const float * x) {
 #ifndef GGML_USE_ACCELERATE
@@ -1538,6 +2936,14 @@ inline static void ggml_vec_sum_f16_ggf(const int n, float * s, const ggml_fp16_
     *s = sum;
 }
 
+inline static void ggml_vec_sum_bf16_ggf(const int n, float * s, const ggml_bf16_t * x) {
+    float sum = 0.0f;
+    for (int i = 0; i < n; ++i) {
+        sum += GGML_BF16_TO_FP32(x[i]);
+    }
+    *s = sum;
+}
+
 inline static void ggml_vec_max_f32(const int n, float * s, const float * x) {
 #ifndef GGML_USE_ACCELERATE
     float max = -INFINITY;
@@ -1582,10 +2988,13 @@ static const char * GGML_OP_NAME[GGML_OP_COUNT] = {
     "SQR",
     "SQRT",
     "LOG",
+    "SIN",
+    "COS",
     "SUM",
     "SUM_ROWS",
     "MEAN",
     "ARGMAX",
+    "COUNT_EQUAL",
     "REPEAT",
     "REPEAT_BACK",
     "CONCAT",
@@ -1616,25 +3025,30 @@ static const char * GGML_OP_NAME[GGML_OP_COUNT] = {
     "SOFT_MAX_BACK",
     "ROPE",
     "ROPE_BACK",
-    "ALIBI",
     "CLAMP",
     "CONV_TRANSPOSE_1D",
     "IM2COL",
+    "IM2COL_BACK",
     "CONV_TRANSPOSE_2D",
     "POOL_1D",
     "POOL_2D",
+    "POOL_2D_BACK",
     "UPSCALE",
     "PAD",
+    "ARANGE",
+    "TIMESTEP_EMBEDDING",
     "ARGSORT",
     "LEAKY_RELU",
 
-    "FLASH_ATTN",
-    "FLASH_FF",
+    "FLASH_ATTN_EXT",
     "FLASH_ATTN_BACK",
+    "SSM_CONV",
+    "SSM_SCAN",
     "WIN_PART",
     "WIN_UNPART",
     "GET_REL_POS",
     "ADD_REL_POS",
+    "RWKV_WKV",
 
     "UNARY",
 
@@ -1651,9 +3065,10 @@ static const char * GGML_OP_NAME[GGML_OP_COUNT] = {
 
     "CROSS_ENTROPY_LOSS",
     "CROSS_ENTROPY_LOSS_BACK",
+    "OPT_STEP_ADAMW",
 };
 
-static_assert(GGML_OP_COUNT == 72, "GGML_OP_COUNT != 72");
+static_assert(GGML_OP_COUNT == 81, "GGML_OP_COUNT != 81");
 
 static const char * GGML_OP_SYMBOL[GGML_OP_COUNT] = {
     "none",
@@ -1668,10 +3083,13 @@ static const char * GGML_OP_SYMBOL[GGML_OP_COUNT] = {
     "x^2",
     "√x",
     "log(x)",
+    "sin(x)",
+    "cos(x)",
     "Σx",
     "Σx_k",
     "Σx/n",
     "argmax(x)",
+    "count_equal(x)",
     "repeat(x)",
     "repeat_back(x)",
     "concat(x, y)",
@@ -1702,25 +3120,30 @@ static const char * GGML_OP_SYMBOL[GGML_OP_COUNT] = {
     "soft_max_back(x)",
     "rope(x)",
     "rope_back(x)",
-    "alibi(x)",
     "clamp(x)",
     "conv_transpose_1d(x)",
     "im2col(x)",
+    "im2col_back(x)",
     "conv_transpose_2d(x)",
     "pool_1d(x)",
     "pool_2d(x)",
+    "pool_2d_back(x)",
     "upscale(x)",
     "pad(x)",
+    "arange(start, stop, step)",
+    "timestep_embedding(timesteps, dim, max_period)",
     "argsort(x)",
     "leaky_relu(x)",
 
-    "flash_attn(x)",
-    "flash_ff(x)",
+    "flash_attn_ext(x)",
     "flash_attn_back(x)",
+    "ssm_conv(x)",
+    "ssm_scan(x)",
     "win_part(x)",
     "win_unpart(x)",
     "get_rel_pos(x)",
     "add_rel_pos(x)",
+    "rwkv_wkv(k, v, r, tf, td, s)",
 
     "unary(x)",
 
@@ -1737,9 +3160,10 @@ static const char * GGML_OP_SYMBOL[GGML_OP_COUNT] = {
 
     "cross_entropy_loss(x,y)",
     "cross_entropy_loss_back(x,y)",
+    "adamw(x)",
 };
 
-static_assert(GGML_OP_COUNT == 72, "GGML_OP_COUNT != 72");
+static_assert(GGML_OP_COUNT == 81, "GGML_OP_COUNT != 81");
 
 static_assert(GGML_OP_POOL_COUNT == 2, "GGML_OP_POOL_COUNT != 2");
 
@@ -1752,78 +3176,33 @@ static const char * GGML_UNARY_OP_NAME[GGML_UNARY_OP_COUNT] = {
     "TANH",
     "ELU",
     "RELU",
+    "SIGMOID",
     "GELU",
     "GELU_QUICK",
     "SILU",
+    "HARDSWISH",
+    "HARDSIGMOID",
+    "EXP",
 };
 
-static_assert(GGML_UNARY_OP_COUNT == 10, "GGML_UNARY_OP_COUNT != 10");
+static_assert(GGML_UNARY_OP_COUNT == 14, "GGML_UNARY_OP_COUNT != 14");
 
 
 static_assert(sizeof(struct ggml_object)%GGML_MEM_ALIGN == 0, "ggml_object size must be a multiple of GGML_MEM_ALIGN");
 static_assert(sizeof(struct ggml_tensor)%GGML_MEM_ALIGN == 0, "ggml_tensor size must be a multiple of GGML_MEM_ALIGN");
 
-// WARN:
-// Mis-configuration can lead to problem that's hard to reason about:
-// * At best  it crash or talks nosense.
-// * At worst it talks slightly difference but hard to perceive.
-//
-// An op has to enable INIT or FINALIZE when any of it's branch needs that pass.
-// Take care about compile options (e.g., GGML_USE_xxx).
-static bool GGML_OP_HAS_INIT    [GGML_OP_COUNT] = { 0 };
-static bool GGML_OP_HAS_FINALIZE[GGML_OP_COUNT] = { 0 };
-
-static void ggml_setup_op_has_task_pass(void) {
-    {   // INIT
-        bool * p = GGML_OP_HAS_INIT;
-
-        p[GGML_OP_ACC                    ] = true;
-        p[GGML_OP_MUL_MAT                ] = true;
-        p[GGML_OP_MUL_MAT_ID             ] = true;
-        p[GGML_OP_OUT_PROD               ] = true;
-        p[GGML_OP_SET                    ] = true;
-        p[GGML_OP_GET_ROWS_BACK          ] = true;
-        p[GGML_OP_DIAG_MASK_INF          ] = true;
-        p[GGML_OP_DIAG_MASK_ZERO         ] = true;
-        p[GGML_OP_CONV_TRANSPOSE_1D      ] = true;
-        p[GGML_OP_CONV_TRANSPOSE_2D      ] = true;
-        p[GGML_OP_FLASH_ATTN_BACK        ] = true;
-        p[GGML_OP_CROSS_ENTROPY_LOSS     ] = true;
-        p[GGML_OP_ADD_REL_POS            ] = true;
-    }
-
-    {   // FINALIZE
-        bool * p = GGML_OP_HAS_FINALIZE;
-
-        p[GGML_OP_CROSS_ENTROPY_LOSS     ] = true;
-    }
+// Helpers for polling loops
+#if defined(__aarch64__) && ( defined(__clang__) || defined(__GNUC__) )
+static inline void ggml_thread_cpu_relax(void) {
+    __asm__ volatile("yield" ::: "memory");
 }
-
-//
-// ggml context
-//
-
-struct ggml_context {
-    size_t mem_size;
-    void * mem_buffer;
-    bool   mem_buffer_owned;
-    bool   no_alloc;
-    bool   no_alloc_save; // this is used to save the no_alloc state when using scratch buffers
-
-    int    n_objects;
-
-    struct ggml_object * objects_begin;
-    struct ggml_object * objects_end;
-
-    struct ggml_scratch scratch;
-    struct ggml_scratch scratch_save;
-};
-
-struct ggml_context_container {
-    bool used;
-
-    struct ggml_context context;
-};
+#elif defined(__x86_64__)
+static inline void ggml_thread_cpu_relax(void) {
+    _mm_pause();
+}
+#else
+static inline void ggml_thread_cpu_relax(void) {;}
+#endif
 
 //
 // NUMA support
@@ -1838,9 +3217,16 @@ struct ggml_numa_node {
 };
 
 struct ggml_numa_nodes {
+    enum ggml_numa_strategy numa_strategy;
     struct ggml_numa_node nodes[GGML_NUMA_MAX_NODES];
     uint32_t n_nodes;
     uint32_t total_cpus; // hardware threads on system
+    uint32_t current_node; // node on which main process is execting
+#if defined(__gnu_linux__)
+    cpu_set_t cpuset; // cpuset from numactl
+#else
+    uint32_t cpuset; // no NUMA support outside of Linux at this time. Use a portable datatype
+#endif
 };
 
 //
@@ -1854,38 +3240,94 @@ struct ggml_state {
 
 // global state
 static struct ggml_state g_state;
-static atomic_int g_state_barrier = 0;
+static atomic_flag g_state_critical = ATOMIC_FLAG_INIT;
 
-// barrier via spin lock
+// critical section via spin lock
 inline static void ggml_critical_section_start(void) {
-    int processing = atomic_fetch_add(&g_state_barrier, 1);
+    while (atomic_flag_test_and_set(&g_state_critical)) {
+        // spin
+        sched_yield();
+    }
+}
+
+static void ggml_barrier(struct ggml_threadpool * tp) {
+    int n_threads = atomic_load_explicit(&tp->n_threads_cur, memory_order_relaxed);
+    if (n_threads == 1) {
+        return;
+    }
+
+#ifdef GGML_USE_OPENMP
+    #pragma omp barrier
+#else
+    int n_passed = atomic_load_explicit(&tp->n_barrier_passed, memory_order_relaxed);
+
+    // enter barrier (full seq-cst fence)
+    int n_barrier = atomic_fetch_add_explicit(&tp->n_barrier, 1, memory_order_seq_cst);
+
+    if (n_barrier == (n_threads - 1)) {
+        // last thread
+        atomic_store_explicit(&tp->n_barrier, 0, memory_order_relaxed);
+
+        // exit barrier (fill seq-cst fence)
+        atomic_fetch_add_explicit(&tp->n_barrier_passed, 1, memory_order_seq_cst);
+        return;
+    }
 
-    while (processing > 0) {
-        // wait for other threads to finish
-        atomic_fetch_sub(&g_state_barrier, 1);
-        sched_yield(); // TODO: reconsider this
-        processing = atomic_fetch_add(&g_state_barrier, 1);
+    // wait for other threads
+    while (atomic_load_explicit(&tp->n_barrier_passed, memory_order_relaxed) == n_passed) {
+        ggml_thread_cpu_relax();
     }
+
+    // exit barrier (full seq-cst fence)
+    // TSAN doesn't support standalone fence yet, we use a dummy read-modify-write instead
+    #ifdef GGML_TSAN_ENABLED
+    atomic_fetch_add_explicit(&tp->n_barrier_passed, 0, memory_order_seq_cst);
+    #else
+    atomic_thread_fence(memory_order_seq_cst);
+    #endif
+#endif
 }
 
 // TODO: make this somehow automatically executed
 //       some sort of "sentry" mechanism
 inline static void ggml_critical_section_end(void) {
-    atomic_fetch_sub(&g_state_barrier, 1);
+    atomic_flag_clear(&g_state_critical);
+}
+
+#if defined(__gnu_linux__)
+static cpu_set_t ggml_get_numa_affinity(void) {
+    cpu_set_t cpuset;
+    pthread_t thread;
+    thread = pthread_self();
+    CPU_ZERO(&cpuset);
+    pthread_getaffinity_np(thread, sizeof(cpu_set_t), &cpuset);
+    return cpuset;
+}
+#else
+static uint32_t ggml_get_numa_affinity(void) {
+    return 0; // no NUMA support
 }
+#endif
 
-void ggml_numa_init(void) {
+void ggml_numa_init(enum ggml_numa_strategy numa_flag) {
     if (g_state.numa.n_nodes > 0) {
-        Rprintf("ggml_numa_init: NUMA already initialized\n");
+        fprintf(stderr, "ggml_numa_init: NUMA already initialized\n");
 
         return;
     }
 
-#ifdef __linux__
+#if defined(__gnu_linux__)
     struct stat st;
     char path[256];
     int rv;
 
+    // set numa scheme
+    g_state.numa.numa_strategy = numa_flag;
+
+    GGML_PRINT_DEBUG("numa strategy %u\n",g_state.numa.numa_strategy);
+
+    g_state.numa.cpuset = ggml_get_numa_affinity();
+
     // enumerate nodes
     while (g_state.numa.n_nodes < GGML_NUMA_MAX_NODES) {
         rv = snprintf(path, sizeof(path), "/sys/devices/system/node/node%u", g_state.numa.n_nodes);
@@ -1904,11 +3346,26 @@ void ggml_numa_init(void) {
 
     GGML_PRINT_DEBUG("found %u numa nodes, %u CPUs\n", g_state.numa.n_nodes, g_state.numa.total_cpus);
 
-    if (g_state.numa.n_nodes < 1 || g_state.numa.total_cpus < 1) {
+    // figure out which node we're on
+    uint current_cpu;
+    int getcpu_ret = 0;
+#if __GLIBC__ > 2 || (__GLIBC__ == 2 && __GLIBC_MINOR__ > 28) || defined(__COSMOPOLITAN__)
+    getcpu_ret = getcpu(&current_cpu, &g_state.numa.current_node);
+#else
+    // old glibc doesn't have a wrapper for this call. Fall back on direct syscall
+#   if !defined(SYS_getcpu) && defined(SYS_get_cpu)
+#       define SYS_getcpu SYS_get_cpu // some older glibc versions use this name
+#   endif
+    getcpu_ret = syscall(SYS_getcpu, &current_cpu, &g_state.numa.current_node);
+#endif
+
+    if (g_state.numa.n_nodes < 1 || g_state.numa.total_cpus < 1 || getcpu_ret != 0) {
         g_state.numa.n_nodes = 0;
         return;
     }
 
+    GGML_PRINT_DEBUG("found our process on numa node %u, CPU %u\n", g_state.numa.current_node, current_cpu);
+
     for (uint32_t n = 0; n < g_state.numa.n_nodes; ++n) {
         struct ggml_numa_node * node = &g_state.numa.nodes[n];
         GGML_PRINT_DEBUG("CPUs on node %u:", n);
@@ -1929,12 +3386,13 @@ void ggml_numa_init(void) {
         if (fptr != NULL) {
             char buf[42];
             if (fgets(buf, sizeof(buf), fptr) && strncmp(buf, "0\n", sizeof(buf)) != 0) {
-                GGML_PRINT("WARNING: /proc/sys/kernel/numa_balancing is enabled, this has been observed to impair performance\n");
+                GGML_LOG_WARN("/proc/sys/kernel/numa_balancing is enabled, this has been observed to impair performance\n");
             }
             fclose(fptr);
         }
     }
 #else
+    UNUSED(numa_flag);
     // TODO
 #endif
 }
@@ -1946,21 +3404,21 @@ bool ggml_is_numa(void) {
 ////////////////////////////////////////////////////////////////////////////////
 
 void ggml_print_object(const struct ggml_object * obj) {
-    GGML_PRINT(" - ggml_object: type = %d, offset = %zu, size = %zu, next = %p\n",
+    GGML_LOG_INFO(" - ggml_object: type = %d, offset = %zu, size = %zu, next = %p\n",
             obj->type, obj->offs, obj->size, (const void *) obj->next);
 }
 
 void ggml_print_objects(const struct ggml_context * ctx) {
     struct ggml_object * obj = ctx->objects_begin;
 
-    GGML_PRINT("%s: objects in context %p:\n", __func__, (const void *) ctx);
+    GGML_LOG_INFO("%s: objects in context %p:\n", __func__, (const void *) ctx);
 
     while (obj != NULL) {
         ggml_print_object(obj);
         obj = obj->next;
     }
 
-    GGML_PRINT("%s: --- end ---\n", __func__);
+    GGML_LOG_INFO("%s: --- end ---\n", __func__);
 }
 
 int64_t ggml_nelements(const struct ggml_tensor * tensor) {
@@ -1998,7 +3456,7 @@ size_t ggml_nbytes_pad(const struct ggml_tensor * tensor) {
     return GGML_PAD(ggml_nbytes(tensor), GGML_MEM_ALIGN);
 }
 
-int ggml_blck_size(enum ggml_type type) {
+int64_t ggml_blck_size(enum ggml_type type) {
     return type_traits[type].blck_size;
 }
 
@@ -2016,7 +3474,7 @@ double ggml_type_sizef(enum ggml_type type) {
 }
 
 const char * ggml_type_name(enum ggml_type type) {
-    return type_traits[type].type_name;
+    return type < GGML_TYPE_COUNT ? type_traits[type].type_name : "NONE";
 }
 
 bool ggml_is_quantized(enum ggml_type type) {
@@ -2040,9 +3498,7 @@ const char * ggml_op_desc(const struct ggml_tensor * t) {
         enum ggml_unary_op uop = ggml_get_unary_op(t);
         return ggml_unary_op_name(uop);
     }
-    else {
-        return ggml_op_name(t->op);
-    }
+    return ggml_op_name(t->op);
 }
 
 size_t ggml_element_size(const struct ggml_tensor * tensor) {
@@ -2102,6 +3558,7 @@ enum ggml_type ggml_ftype_to_ggml_type(enum ggml_ftype ftype) {
     switch (ftype) {
         case GGML_FTYPE_ALL_F32:              wtype = GGML_TYPE_F32;   break;
         case GGML_FTYPE_MOSTLY_F16:           wtype = GGML_TYPE_F16;   break;
+        case GGML_FTYPE_MOSTLY_BF16:          wtype = GGML_TYPE_BF16;  break;
         case GGML_FTYPE_MOSTLY_Q4_0:          wtype = GGML_TYPE_Q4_0;  break;
         case GGML_FTYPE_MOSTLY_Q4_1:          wtype = GGML_TYPE_Q4_1;  break;
         case GGML_FTYPE_MOSTLY_Q5_0:          wtype = GGML_TYPE_Q5_0;  break;
@@ -2112,6 +3569,18 @@ enum ggml_type ggml_ftype_to_ggml_type(enum ggml_ftype ftype) {
         case GGML_FTYPE_MOSTLY_Q4_K:          wtype = GGML_TYPE_Q4_K;  break;
         case GGML_FTYPE_MOSTLY_Q5_K:          wtype = GGML_TYPE_Q5_K;  break;
         case GGML_FTYPE_MOSTLY_Q6_K:          wtype = GGML_TYPE_Q6_K;  break;
+        case GGML_FTYPE_MOSTLY_IQ2_XXS:       wtype = GGML_TYPE_IQ2_XXS;  break;
+        case GGML_FTYPE_MOSTLY_IQ2_XS:        wtype = GGML_TYPE_IQ2_XS;   break;
+        case GGML_FTYPE_MOSTLY_IQ3_XXS:       wtype = GGML_TYPE_IQ3_XXS;  break;
+        case GGML_FTYPE_MOSTLY_IQ1_S:         wtype = GGML_TYPE_IQ1_S;    break;
+        case GGML_FTYPE_MOSTLY_IQ1_M:         wtype = GGML_TYPE_IQ1_M;    break;
+        case GGML_FTYPE_MOSTLY_IQ4_NL:        wtype = GGML_TYPE_IQ4_NL;   break;
+        case GGML_FTYPE_MOSTLY_IQ4_XS:        wtype = GGML_TYPE_IQ4_XS;   break;
+        case GGML_FTYPE_MOSTLY_IQ3_S:         wtype = GGML_TYPE_IQ3_S;    break;
+        case GGML_FTYPE_MOSTLY_IQ2_S:         wtype = GGML_TYPE_IQ2_S;    break;
+        case GGML_FTYPE_MOSTLY_Q4_0_4_4:      wtype = GGML_TYPE_Q4_0_4_4; break;
+        case GGML_FTYPE_MOSTLY_Q4_0_4_8:      wtype = GGML_TYPE_Q4_0_4_8; break;
+        case GGML_FTYPE_MOSTLY_Q4_0_8_8:      wtype = GGML_TYPE_Q4_0_8_8; break;
         case GGML_FTYPE_UNKNOWN:              wtype = GGML_TYPE_COUNT; break;
         case GGML_FTYPE_MOSTLY_Q4_1_SOME_F16: wtype = GGML_TYPE_COUNT; break;
     }
@@ -2129,23 +3598,42 @@ bool ggml_is_transposed(const struct ggml_tensor * tensor) {
     return tensor->nb[0] > tensor->nb[1];
 }
 
+static bool ggml_is_contiguous_n(const struct ggml_tensor * tensor, int n) {
+    size_t next_nb = ggml_type_size(tensor->type);
+    if (tensor->ne[0] != ggml_blck_size(tensor->type) && tensor->nb[0] != next_nb) {
+        return false;
+    }
+    next_nb *= tensor->ne[0]/ggml_blck_size(tensor->type);
+    for (int i = 1; i < GGML_MAX_DIMS; i++) {
+        if (tensor->ne[i] != 1) {
+            if (i > n) {
+                if (tensor->nb[i] != next_nb) {
+                    return false;
+                }
+                next_nb *= tensor->ne[i];
+            } else {
+                // this dimension does not need to be contiguous
+                next_nb = tensor->ne[i]*tensor->nb[i];
+            }
+        }
+    }
+    return true;
+}
+
 bool ggml_is_contiguous(const struct ggml_tensor * tensor) {
-    static_assert(GGML_MAX_DIMS == 4, "GGML_MAX_DIMS is not 4 - update this function");
+    return ggml_is_contiguous_0(tensor);
+}
 
-    return
-        tensor->nb[0] == ggml_type_size(tensor->type) &&
-        tensor->nb[1] == (tensor->nb[0]*tensor->ne[0])/ggml_blck_size(tensor->type) &&
-        tensor->nb[2] == tensor->nb[1]*tensor->ne[1] &&
-        tensor->nb[3] == tensor->nb[2]*tensor->ne[2];
+bool ggml_is_contiguous_0(const struct ggml_tensor * tensor) {
+    return ggml_is_contiguous_n(tensor, 0);
 }
 
-static inline bool ggml_is_contiguous_except_dim_1(const struct ggml_tensor * tensor) {
-    static_assert(GGML_MAX_DIMS == 4, "GGML_MAX_DIMS is not 4 - update this function");
+bool ggml_is_contiguous_1(const struct ggml_tensor * tensor) {
+    return ggml_is_contiguous_n(tensor, 1);
+}
 
-    return
-        tensor->nb[0] == ggml_type_size(tensor->type) &&
-        tensor->nb[2] == tensor->nb[1]*tensor->ne[1] &&
-        tensor->nb[3] == tensor->nb[2]*tensor->ne[2];
+bool ggml_is_contiguous_2(const struct ggml_tensor * tensor) {
+    return ggml_is_contiguous_n(tensor, 2);
 }
 
 bool ggml_is_permuted(const struct ggml_tensor * tensor) {
@@ -2163,21 +3651,41 @@ static inline bool ggml_is_padded_1d(const struct ggml_tensor * tensor) {
         tensor->nb[3] == tensor->nb[2]*tensor->ne[2];
 }
 
+bool ggml_is_empty(const struct ggml_tensor * tensor) {
+    for (int i = 0; i < GGML_MAX_DIMS; ++i) {
+        if (tensor->ne[i] == 0) {
+            // empty if any dimension has no elements
+            return true;
+        }
+    }
+    return false;
+}
+
 bool ggml_are_same_shape(const struct ggml_tensor * t0, const struct ggml_tensor * t1) {
     static_assert(GGML_MAX_DIMS == 4, "GGML_MAX_DIMS is not 4 - update this function");
 
     return
-        (t0->ne[0] == t1->ne[0] ) &&
-        (t0->ne[1] == t1->ne[1] ) &&
-        (t0->ne[2] == t1->ne[2] ) &&
-        (t0->ne[3] == t1->ne[3] );
+        (t0->ne[0] == t1->ne[0]) &&
+        (t0->ne[1] == t1->ne[1]) &&
+        (t0->ne[2] == t1->ne[2]) &&
+        (t0->ne[3] == t1->ne[3]);
 }
 
-// check if t1 can be represented as a repeatition of t0
-static inline bool ggml_can_repeat(const struct ggml_tensor * t0, const struct ggml_tensor * t1) {
+bool ggml_are_same_stride(const struct ggml_tensor * t0, const struct ggml_tensor * t1) {
     static_assert(GGML_MAX_DIMS == 4, "GGML_MAX_DIMS is not 4 - update this function");
 
     return
+        (t0->nb[0] == t1->nb[0]) &&
+        (t0->nb[1] == t1->nb[1]) &&
+        (t0->nb[2] == t1->nb[2]) &&
+        (t0->nb[3] == t1->nb[3]);
+}
+
+// check if t1 can be represented as a repeatition of t0
+bool ggml_can_repeat(const struct ggml_tensor * t0, const struct ggml_tensor * t1) {
+    static_assert(GGML_MAX_DIMS == 4, "GGML_MAX_DIMS is not 4 - update this function");
+
+    return ggml_is_empty(t0) ? ggml_is_empty(t1) :
         (t1->ne[0]%t0->ne[0] == 0) &&
         (t1->ne[1]%t0->ne[1] == 0) &&
         (t1->ne[2]%t0->ne[2] == 0) &&
@@ -2205,11 +3713,75 @@ static inline int ggml_up(int n, int m) {
 }
 
 // assert that pointer is aligned to GGML_MEM_ALIGN
-#define ggml_assert_aligned(ptr) \
+#define GGML_ASSERT_ALIGNED(ptr) \
     GGML_ASSERT(((uintptr_t) (ptr))%GGML_MEM_ALIGN == 0)
 
 ////////////////////////////////////////////////////////////////////////////////
 
+#if defined(__ARM_ARCH)
+
+#if defined(__linux__) && defined(__aarch64__)
+#include <sys/auxv.h>
+#elif defined(__APPLE__)
+#include <sys/sysctl.h>
+#endif
+
+#if !defined(HWCAP2_I8MM)
+#define HWCAP2_I8MM 0
+#endif
+
+static void ggml_init_arm_arch_features(void) {
+#if defined(__linux__) && defined(__aarch64__)
+    uint32_t hwcap = getauxval(AT_HWCAP);
+    uint32_t hwcap2 = getauxval(AT_HWCAP2);
+
+    ggml_arm_arch_features.has_neon = !!(hwcap & HWCAP_ASIMD);
+    ggml_arm_arch_features.has_i8mm = !!(hwcap2 & HWCAP2_I8MM);
+    ggml_arm_arch_features.has_sve  = !!(hwcap & HWCAP_SVE);
+
+#if defined(__ARM_FEATURE_SVE)
+    ggml_arm_arch_features.sve_cnt = PR_SVE_VL_LEN_MASK & prctl(PR_SVE_GET_VL);
+#endif
+#elif defined(__APPLE__)
+    int oldp = 0;
+    size_t size = sizeof(oldp);
+    if (sysctlbyname("hw.optional.AdvSIMD", &oldp, &size, NULL, 0) != 0) {
+        oldp = 0;
+    }
+    ggml_arm_arch_features.has_neon = oldp;
+
+    if (sysctlbyname("hw.optional.arm.FEAT_I8MM", &oldp, &size, NULL, 0) != 0) {
+        oldp = 0;
+    }
+    ggml_arm_arch_features.has_i8mm = oldp;
+
+    ggml_arm_arch_features.has_sve = 0;
+    ggml_arm_arch_features.sve_cnt = 0;
+#else
+// Run-time CPU feature detection not implemented for this platform, fallback to compile time
+#if defined(__ARM_NEON)
+    ggml_arm_arch_features.has_neon = 1;
+#else
+    ggml_arm_arch_features.has_neon = 0;
+#endif
+
+#if defined(__ARM_FEATURE_MATMUL_INT8)
+    ggml_arm_arch_features.has_i8mm = 1;
+#else
+    ggml_arm_arch_features.has_i8mm = 0;
+#endif
+
+#if defined(__ARM_FEATURE_SVE)
+    ggml_arm_arch_features.has_sve = 1;
+    ggml_arm_arch_features.sve_cnt = 16;
+#else
+    ggml_arm_arch_features.has_sve = 0;
+    ggml_arm_arch_features.sve_cnt = 0;
+#endif
+#endif
+}
+#endif
+
 struct ggml_context * ggml_init(struct ggml_init_params params) {
     // make this function thread safe
     ggml_critical_section_start();
@@ -2224,15 +3796,14 @@ struct ggml_context * ggml_init(struct ggml_init_params params) {
         {
             const uint64_t t_start = ggml_time_us(); UNUSED(t_start);
 
-            ggml_fp16_t ii;
             for (int i = 0; i < (1 << 16); ++i) {
-                uint16_t ui = i;
-                memcpy(&ii, &ui, sizeof(ii));
-                const float f = ggml_table_f32_f16[i] = GGML_COMPUTE_FP16_TO_FP32(ii);
+                union {
+                    uint16_t u16;
+                    ggml_fp16_t fp16;
+                } u = {i};
+                float f = ggml_table_f32_f16[i] = GGML_COMPUTE_FP16_TO_FP32(u.fp16);
                 ggml_table_gelu_f16[i] = GGML_FP32_TO_FP16(ggml_gelu_f32(f));
                 ggml_table_gelu_quick_f16[i] = GGML_FP32_TO_FP16(ggml_gelu_quick_f32(f));
-                ggml_table_silu_f16[i] = GGML_FP32_TO_FP16(ggml_silu_f32(f));
-                ggml_table_exp_f16[i]  = GGML_FP32_TO_FP16(expf(f));
             }
 
             const uint64_t t_end = ggml_time_us(); UNUSED(t_end);
@@ -2261,14 +3832,10 @@ struct ggml_context * ggml_init(struct ggml_init_params params) {
             GGML_PRINT_DEBUG("%s: g_state initialized in %f ms\n", __func__, (t_end - t_start)/1000.0f);
         }
 
-#if defined(GGML_USE_CUBLAS)
-        ggml_init_cublas();
-#elif defined(GGML_USE_CLBLAST)
-        ggml_cl_init();
+#if defined(__ARM_ARCH)
+        ggml_init_arm_arch_features();
 #endif
 
-        ggml_setup_op_has_task_pass();
-
         is_first_call = false;
     }
 
@@ -2315,7 +3882,7 @@ struct ggml_context * ggml_init(struct ggml_init_params params) {
 
     GGML_ASSERT(ctx->mem_buffer != NULL);
 
-    ggml_assert_aligned(ctx->mem_buffer);
+    GGML_ASSERT_ALIGNED(ctx->mem_buffer);
 
     GGML_PRINT_DEBUG("%s: context initialized\n", __func__);
 
@@ -2325,6 +3892,10 @@ struct ggml_context * ggml_init(struct ggml_init_params params) {
 }
 
 void ggml_free(struct ggml_context * ctx) {
+    if (ctx == NULL) {
+        return;
+    }
+
     // make this function thread safe
     ggml_critical_section_start();
 
@@ -2385,7 +3956,8 @@ size_t ggml_get_max_tensor_size(const struct ggml_context * ctx) {
     size_t max_size = 0;
 
     for (struct ggml_tensor * tensor = ggml_get_first_tensor(ctx); tensor != NULL; tensor = ggml_get_next_tensor(ctx, tensor)) {
-        max_size = MAX(max_size, ggml_nbytes(tensor));
+        size_t bytes = ggml_nbytes(tensor);
+        max_size = MAX(max_size, bytes);
     }
 
     return max_size;
@@ -2429,8 +4001,8 @@ static struct ggml_object * ggml_new_object(struct ggml_context * ctx, enum ggml
     struct ggml_object * const obj_new = (struct ggml_object *)(mem_buffer + cur_end);
 
     if (cur_end + size_needed + GGML_OBJECT_SIZE > ctx->mem_size) {
-        GGML_PRINT("%s: not enough space in the context's memory pool (needed %zu, available %zu)\n",
-                __func__, cur_end + size_needed, ctx->mem_size);
+        GGML_LOG_WARN("%s: not enough space in the context's memory pool (needed %zu, available %zu)\n",
+                __func__, cur_end + size_needed + GGML_OBJECT_SIZE, ctx->mem_size);
         assert(false);
         return NULL;
     }
@@ -2442,7 +4014,7 @@ static struct ggml_object * ggml_new_object(struct ggml_context * ctx, enum ggml
         .type = type,
     };
 
-    ggml_assert_aligned(mem_buffer + obj_new->offs);
+    GGML_ASSERT_ALIGNED(mem_buffer + obj_new->offs);
 
     if (obj_cur != NULL) {
         obj_cur->next = obj_new;
@@ -2466,7 +4038,8 @@ static struct ggml_tensor * ggml_new_tensor_impl(
         struct ggml_tensor  * view_src,
         size_t                view_offs) {
 
-    assert(n_dims >= 1 && n_dims <= GGML_MAX_DIMS);
+    GGML_ASSERT(type >= 0 && type < GGML_TYPE_COUNT);
+    GGML_ASSERT(n_dims >= 1 && n_dims <= GGML_MAX_DIMS);
 
     // find the base tensor and absolute offset
     if (view_src != NULL && view_src->view_src != NULL) {
@@ -2479,7 +4052,7 @@ static struct ggml_tensor * ggml_new_tensor_impl(
         data_size *= ne[i];
     }
 
-    GGML_ASSERT(view_src == NULL || data_size + view_offs <= ggml_nbytes(view_src));
+    GGML_ASSERT(view_src == NULL || data_size == 0 || data_size + view_offs <= ggml_nbytes(view_src));
 
     void * data = view_src != NULL ? view_src->data : NULL;
     if (data != NULL) {
@@ -2492,7 +4065,7 @@ static struct ggml_tensor * ggml_new_tensor_impl(
         if (ctx->scratch.data != NULL) {
             // allocate tensor data in the scratch buffer
             if (ctx->scratch.offs + data_size > ctx->scratch.size) {
-                GGML_PRINT("%s: not enough space in the scratch memory pool (needed %zu, available %zu)\n",
+                GGML_LOG_WARN("%s: not enough space in the scratch memory pool (needed %zu, available %zu)\n",
                         __func__, ctx->scratch.offs + data_size, ctx->scratch.size);
                 assert(false);
                 return NULL;
@@ -2507,36 +4080,44 @@ static struct ggml_tensor * ggml_new_tensor_impl(
         }
     }
 
-    struct ggml_object * const obj_new = ggml_new_object(ctx, GGML_OBJECT_TENSOR, GGML_TENSOR_SIZE + obj_alloc_size);
+    struct ggml_object * const obj_new = ggml_new_object(ctx, GGML_OBJECT_TYPE_TENSOR, GGML_TENSOR_SIZE + obj_alloc_size);
+    GGML_ASSERT(obj_new);
 
     // TODO: for recoverable errors, we would need to free the data allocated from the scratch buffer here
 
     struct ggml_tensor * const result = (struct ggml_tensor *)((char *)ctx->mem_buffer + obj_new->offs);
 
+#ifdef __clang__
+    // temporary until ggml_tensor::backend is removed
+    #pragma clang diagnostic push
+    #pragma clang diagnostic ignored "-Wdeprecated-declarations"
+#endif
+
     *result = (struct ggml_tensor) {
         /*.type         =*/ type,
-        /*.backend      =*/ GGML_BACKEND_CPU,
+        /*.backend      =*/ GGML_BACKEND_TYPE_CPU,
         /*.buffer       =*/ NULL,
         /*.ne           =*/ { 1, 1, 1, 1 },
         /*.nb           =*/ { 0, 0, 0, 0 },
         /*.op           =*/ GGML_OP_NONE,
         /*.op_params    =*/ { 0 },
-        /*.is_param     =*/ false,
+        /*.flags        =*/ 0,
         /*.grad         =*/ NULL,
         /*.src          =*/ { NULL },
-        /*.perf_runs    =*/ 0,
-        /*.perf_cycles  =*/ 0,
-        /*.perf_time_us =*/ 0,
         /*.view_src     =*/ view_src,
         /*.view_offs    =*/ view_offs,
         /*.data         =*/ obj_alloc_size > 0 ? (void *)(result + 1) : data,
         /*.name         =*/ { 0 },
         /*.extra        =*/ NULL,
-        /*.padding      =*/ { 0 },
+        ///*.padding      =*/ { 0 },
     };
 
+#ifdef __clang__
+    #pragma clang diagnostic pop
+#endif
+
     // TODO: this should not be needed as long as we don't rely on aligned SIMD loads
-    //ggml_assert_aligned(result->data);
+    //GGML_ASSERT_ALIGNED(result->data);
 
     for (int i = 0; i < n_dims; i++) {
         result->ne[i] = ne[i];
@@ -2637,13 +4218,31 @@ static int32_t ggml_get_op_params_i32(const struct ggml_tensor * tensor, uint32_
     return ((const int32_t *)(tensor->op_params))[i];
 }
 
+static float ggml_get_op_params_f32(const struct ggml_tensor * tensor, uint32_t i) {
+    assert(i < GGML_MAX_OP_PARAMS / sizeof(float));
+    return ((const float *)(tensor->op_params))[i];
+}
+
 static void ggml_set_op_params_i32(struct ggml_tensor * tensor, uint32_t i, int32_t value) {
     assert(i < GGML_MAX_OP_PARAMS / sizeof(int32_t));
     ((int32_t *)(tensor->op_params))[i] = value;
 }
 
+static void ggml_set_op_params_f32(struct ggml_tensor * tensor, uint32_t i, float value) {
+    assert(i < GGML_MAX_OP_PARAMS / sizeof(float));
+    ((float *)(tensor->op_params))[i] = value;
+}
+
 struct ggml_tensor * ggml_set_zero(struct ggml_tensor * tensor) {
-    memset(tensor->data, 0, ggml_nbytes(tensor));
+    if (ggml_is_empty(tensor)) {
+        return tensor;
+    }
+    if (tensor->buffer) {
+        ggml_backend_tensor_memset(tensor, 0, 0, ggml_nbytes(tensor));
+    } else {
+        GGML_ASSERT(tensor->data);
+        memset(tensor->data, 0, ggml_nbytes(tensor));
+    }
     return tensor;
 }
 
@@ -2683,6 +4282,13 @@ struct ggml_tensor * ggml_set_i32 (struct ggml_tensor * tensor, int32_t value) {
                     ggml_vec_set_f16(nc, (ggml_fp16_t *)(data + i*n1), GGML_FP32_TO_FP16(value));
                 }
             } break;
+        case GGML_TYPE_BF16:
+            {
+                assert(tensor->nb[0] == sizeof(ggml_fp16_t));
+                for (int i = 0; i < n; i++) {
+                    ggml_vec_set_bf16(nc, (ggml_bf16_t *)(data + i*n1), GGML_FP32_TO_BF16(value));
+                }
+            } break;
         case GGML_TYPE_F32:
             {
                 assert(tensor->nb[0] == sizeof(float));
@@ -2692,8 +4298,8 @@ struct ggml_tensor * ggml_set_i32 (struct ggml_tensor * tensor, int32_t value) {
             } break;
         default:
             {
-                GGML_ASSERT(false);
-            } break;
+                GGML_ABORT("fatal error");
+            }
     }
 
     return tensor;
@@ -2735,6 +4341,13 @@ struct ggml_tensor * ggml_set_f32(struct ggml_tensor * tensor, float value) {
                     ggml_vec_set_f16(nc, (ggml_fp16_t *)(data + i*n1), GGML_FP32_TO_FP16(value));
                 }
             } break;
+        case GGML_TYPE_BF16:
+            {
+                assert(tensor->nb[0] == sizeof(ggml_bf16_t));
+                for (int i = 0; i < n; i++) {
+                    ggml_vec_set_bf16(nc, (ggml_bf16_t *)(data + i*n1), GGML_FP32_TO_BF16(value));
+                }
+            } break;
         case GGML_TYPE_F32:
             {
                 assert(tensor->nb[0] == sizeof(float));
@@ -2744,8 +4357,8 @@ struct ggml_tensor * ggml_set_f32(struct ggml_tensor * tensor, float value) {
             } break;
         default:
             {
-                GGML_ASSERT(false);
-            } break;
+                GGML_ABORT("fatal error");
+            }
     }
 
     return tensor;
@@ -2802,6 +4415,11 @@ int32_t ggml_get_i32_1d(const struct ggml_tensor * tensor, int i) {
                 GGML_ASSERT(tensor->nb[0] == sizeof(ggml_fp16_t));
                 return GGML_FP16_TO_FP32(((ggml_fp16_t *)(tensor->data))[i]);
             }
+        case GGML_TYPE_BF16:
+            {
+                GGML_ASSERT(tensor->nb[0] == sizeof(ggml_bf16_t));
+                return GGML_BF16_TO_FP32(((ggml_bf16_t *)(tensor->data))[i]);
+            }
         case GGML_TYPE_F32:
             {
                 GGML_ASSERT(tensor->nb[0] == sizeof(float));
@@ -2809,11 +4427,9 @@ int32_t ggml_get_i32_1d(const struct ggml_tensor * tensor, int i) {
             }
         default:
             {
-                GGML_ASSERT(false);
+                GGML_ABORT("fatal error");
             }
     }
-
-    return 0.0f;
 }
 
 void ggml_set_i32_1d(const struct ggml_tensor * tensor, int i, int32_t value) {
@@ -2844,6 +4460,11 @@ void ggml_set_i32_1d(const struct ggml_tensor * tensor, int i, int32_t value) {
                 GGML_ASSERT(tensor->nb[0] == sizeof(ggml_fp16_t));
                 ((ggml_fp16_t *)(tensor->data))[i] = GGML_FP32_TO_FP16(value);
             } break;
+        case GGML_TYPE_BF16:
+            {
+                GGML_ASSERT(tensor->nb[0] == sizeof(ggml_bf16_t));
+                ((ggml_bf16_t *)(tensor->data))[i] = GGML_FP32_TO_BF16(value);
+            } break;
         case GGML_TYPE_F32:
             {
                 GGML_ASSERT(tensor->nb[0] == sizeof(float));
@@ -2851,8 +4472,8 @@ void ggml_set_i32_1d(const struct ggml_tensor * tensor, int i, int32_t value) {
             } break;
         default:
             {
-                GGML_ASSERT(false);
-            } break;
+                GGML_ABORT("fatal error");
+            }
     }
 }
 
@@ -2867,13 +4488,13 @@ int32_t ggml_get_i32_nd(const struct ggml_tensor * tensor, int i0, int i1, int i
             return ((int32_t *) data)[0];
         case GGML_TYPE_F16:
             return GGML_FP16_TO_FP32(((ggml_fp16_t *) data)[0]);
+        case GGML_TYPE_BF16:
+            return GGML_BF16_TO_FP32(((ggml_bf16_t *) data)[0]);
         case GGML_TYPE_F32:
             return ((float *) data)[0];
         default:
-            GGML_ASSERT(false);
+            GGML_ABORT("fatal error");
     }
-
-    return 0.0f;
 }
 
 void ggml_set_i32_nd(const struct ggml_tensor * tensor, int i0, int i1, int i2, int i3, int32_t value) {
@@ -2895,14 +4516,18 @@ void ggml_set_i32_nd(const struct ggml_tensor * tensor, int i0, int i1, int i2,
             {
                 ((ggml_fp16_t *)(data))[0] = GGML_FP32_TO_FP16(value);
             } break;
+        case GGML_TYPE_BF16:
+            {
+                ((ggml_bf16_t *)(data))[0] = GGML_FP32_TO_BF16(value);
+            } break;
         case GGML_TYPE_F32:
             {
                 ((float *)(data))[0] = value;
             } break;
         default:
             {
-                GGML_ASSERT(false);
-            } break;
+                GGML_ABORT("fatal error");
+            }
     }
 }
 
@@ -2915,36 +4540,33 @@ float ggml_get_f32_1d(const struct ggml_tensor * tensor, int i) {
     switch (tensor->type) {
         case GGML_TYPE_I8:
             {
-                GGML_ASSERT(tensor->nb[0] == sizeof(int8_t));
                 return ((int8_t *)(tensor->data))[i];
             }
         case GGML_TYPE_I16:
             {
-                GGML_ASSERT(tensor->nb[0] == sizeof(int16_t));
                 return ((int16_t *)(tensor->data))[i];
             }
         case GGML_TYPE_I32:
             {
-                GGML_ASSERT(tensor->nb[0] == sizeof(int32_t));
                 return ((int32_t *)(tensor->data))[i];
             }
         case GGML_TYPE_F16:
             {
-                GGML_ASSERT(tensor->nb[0] == sizeof(ggml_fp16_t));
                 return GGML_FP16_TO_FP32(((ggml_fp16_t *)(tensor->data))[i]);
             }
+        case GGML_TYPE_BF16:
+            {
+                return GGML_BF16_TO_FP32(((ggml_bf16_t *)(tensor->data))[i]);
+            }
         case GGML_TYPE_F32:
             {
-                GGML_ASSERT(tensor->nb[0] == sizeof(float));
                 return ((float *)(tensor->data))[i];
             }
         default:
             {
-                GGML_ASSERT(false);
+                GGML_ABORT("fatal error");
             }
     }
-
-    return 0.0f;
 }
 
 void ggml_set_f32_1d(const struct ggml_tensor * tensor, int i, float value) {
@@ -2957,33 +4579,32 @@ void ggml_set_f32_1d(const struct ggml_tensor * tensor, int i, float value) {
     switch (tensor->type) {
         case GGML_TYPE_I8:
             {
-                GGML_ASSERT(tensor->nb[0] == sizeof(int8_t));
                 ((int8_t *)(tensor->data))[i] = value;
             } break;
         case GGML_TYPE_I16:
             {
-                GGML_ASSERT(tensor->nb[0] == sizeof(int16_t));
                 ((int16_t *)(tensor->data))[i] = value;
             } break;
         case GGML_TYPE_I32:
             {
-                GGML_ASSERT(tensor->nb[0] == sizeof(int32_t));
                 ((int32_t *)(tensor->data))[i] = value;
             } break;
         case GGML_TYPE_F16:
             {
-                GGML_ASSERT(tensor->nb[0] == sizeof(ggml_fp16_t));
                 ((ggml_fp16_t *)(tensor->data))[i] = GGML_FP32_TO_FP16(value);
             } break;
+        case GGML_TYPE_BF16:
+            {
+                ((ggml_bf16_t *)(tensor->data))[i] = GGML_FP32_TO_BF16(value);
+            } break;
         case GGML_TYPE_F32:
             {
-                GGML_ASSERT(tensor->nb[0] == sizeof(float));
                 ((float *)(tensor->data))[i] = value;
             } break;
         default:
             {
-                GGML_ASSERT(false);
-            } break;
+                GGML_ABORT("fatal error");
+            }
     }
 }
 
@@ -2998,13 +4619,13 @@ float ggml_get_f32_nd(const struct ggml_tensor * tensor, int i0, int i1, int i2,
             return ((int32_t *) data)[0];
         case GGML_TYPE_F16:
             return GGML_FP16_TO_FP32(((ggml_fp16_t *) data)[0]);
+        case GGML_TYPE_BF16:
+            return GGML_BF16_TO_FP32(((ggml_bf16_t *) data)[0]);
         case GGML_TYPE_F32:
             return ((float *) data)[0];
         default:
-            GGML_ASSERT(false);
+            GGML_ABORT("fatal error");
     }
-
-    return 0.0f;
 }
 
 void ggml_set_f32_nd(const struct ggml_tensor * tensor, int i0, int i1, int i2, int i3, float value) {
@@ -3026,14 +4647,18 @@ void ggml_set_f32_nd(const struct ggml_tensor * tensor, int i0, int i1, int i2,
             {
                 ((ggml_fp16_t *)(data))[0] = GGML_FP32_TO_FP16(value);
             } break;
+        case GGML_TYPE_BF16:
+            {
+                ((ggml_bf16_t *)(data))[0] = GGML_FP32_TO_BF16(value);
+            } break;
         case GGML_TYPE_F32:
             {
                 ((float *)(data))[0] = value;
             } break;
         default:
             {
-                GGML_ASSERT(false);
-            } break;
+                GGML_ABORT("fatal error");
+            }
     }
 }
 
@@ -3056,8 +4681,11 @@ const char * ggml_get_name(const struct ggml_tensor * tensor) {
 }
 
 struct ggml_tensor * ggml_set_name(struct ggml_tensor * tensor, const char * name) {
-    strncpy(tensor->name, name, sizeof(tensor->name));
-    tensor->name[sizeof(tensor->name) - 1] = '\0';
+    size_t i;
+    for (i = 0; i < sizeof(tensor->name) - 1 && name[i] != '\0'; i++) {
+        tensor->name[i] = name[i];
+    }
+    tensor->name[i] = '\0';
     return tensor;
 }
 
@@ -3088,7 +4716,7 @@ struct ggml_tensor * ggml_get_first_tensor(const struct ggml_context * ctx) {
     char * const mem_buffer = ctx->mem_buffer;
 
     while (obj != NULL) {
-        if (obj->type == GGML_OBJECT_TENSOR) {
+        if (obj->type == GGML_OBJECT_TYPE_TENSOR) {
             return (struct ggml_tensor *)(mem_buffer + obj->offs);
         }
 
@@ -3105,7 +4733,7 @@ struct ggml_tensor * ggml_get_next_tensor(const struct ggml_context * ctx, struc
     char * const mem_buffer = ctx->mem_buffer;
 
     while (obj != NULL) {
-        if (obj->type == GGML_OBJECT_TENSOR) {
+        if (obj->type == GGML_OBJECT_TYPE_TENSOR) {
             return (struct ggml_tensor *)(mem_buffer + obj->offs);
         }
 
@@ -3121,7 +4749,7 @@ struct ggml_tensor * ggml_get_tensor(struct ggml_context * ctx, const char * nam
     char * const mem_buffer = ctx->mem_buffer;
 
     while (obj != NULL) {
-        if (obj->type == GGML_OBJECT_TENSOR) {
+        if (obj->type == GGML_OBJECT_TYPE_TENSOR) {
             struct ggml_tensor * cur = (struct ggml_tensor *)(mem_buffer + obj->offs);
             if (strcmp(cur->name, name) == 0) {
                 return cur;
@@ -3140,18 +4768,11 @@ struct ggml_tensor * ggml_get_tensor(struct ggml_context * ctx, const char * nam
 
 static struct ggml_tensor * ggml_dup_impl(
         struct ggml_context * ctx,
-        struct ggml_tensor * a,
-        bool inplace) {
-    bool is_node = false;
-
-    if (!inplace && (a->grad)) {
-        is_node = true;
-    }
-
+        struct ggml_tensor  * a,
+        bool                  inplace) {
     struct ggml_tensor * result = inplace ? ggml_view_tensor(ctx, a) : ggml_dup_tensor(ctx, a);
 
-    result->op   = GGML_OP_DUP;
-    result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL;
+    result->op     = GGML_OP_DUP;
     result->src[0] = a;
 
     return result;
@@ -3159,13 +4780,13 @@ static struct ggml_tensor * ggml_dup_impl(
 
 struct ggml_tensor * ggml_dup(
         struct ggml_context * ctx,
-        struct ggml_tensor * a) {
+        struct ggml_tensor  * a) {
     return ggml_dup_impl(ctx, a, false);
 }
 
 struct ggml_tensor * ggml_dup_inplace(
         struct ggml_context * ctx,
-        struct ggml_tensor * a) {
+        struct ggml_tensor  * a) {
     return ggml_dup_impl(ctx, a, true);
 }
 
@@ -3173,23 +4794,14 @@ struct ggml_tensor * ggml_dup_inplace(
 
 static struct ggml_tensor * ggml_add_impl(
         struct ggml_context * ctx,
-        struct ggml_tensor * a,
-        struct ggml_tensor * b,
-        bool inplace) {
+        struct ggml_tensor  * a,
+        struct ggml_tensor  * b,
+        bool                  inplace) {
     GGML_ASSERT(ggml_can_repeat(b, a));
 
-    bool is_node = false;
-
-    if (!inplace && (a->grad || b->grad)) {
-        // TODO: support backward pass for broadcasting
-        GGML_ASSERT(ggml_are_same_shape(a, b));
-        is_node = true;
-    }
-
     struct ggml_tensor * result = inplace ? ggml_view_tensor(ctx, a) : ggml_dup_tensor(ctx, a);
 
-    result->op   = GGML_OP_ADD;
-    result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL;
+    result->op     = GGML_OP_ADD;
     result->src[0] = a;
     result->src[1] = b;
 
@@ -3198,15 +4810,15 @@ static struct ggml_tensor * ggml_add_impl(
 
 struct ggml_tensor * ggml_add(
         struct ggml_context * ctx,
-        struct ggml_tensor * a,
-        struct ggml_tensor * b) {
+        struct ggml_tensor  * a,
+        struct ggml_tensor  * b) {
     return ggml_add_impl(ctx, a, b, false);
 }
 
 struct ggml_tensor * ggml_add_inplace(
         struct ggml_context * ctx,
-        struct ggml_tensor * a,
-        struct ggml_tensor * b) {
+        struct ggml_tensor  * a,
+        struct ggml_tensor  * b) {
     return ggml_add_impl(ctx, a, b, true);
 }
 
@@ -3214,26 +4826,21 @@ struct ggml_tensor * ggml_add_inplace(
 
 static struct ggml_tensor * ggml_add_cast_impl(
         struct ggml_context * ctx,
-        struct ggml_tensor * a,
-        struct ggml_tensor * b,
-        enum   ggml_type     type) {
+        struct ggml_tensor  * a,
+        struct ggml_tensor  * b,
+        enum   ggml_type      type) {
     // TODO: support less-strict constraint
     //       GGML_ASSERT(ggml_can_repeat(b, a));
     GGML_ASSERT(ggml_can_repeat_rows(b, a));
-    GGML_ASSERT(ggml_is_quantized(a->type) || a->type == GGML_TYPE_F16); // currently only supported for quantized input and f16
 
-    bool is_node = false;
-
-    if (a->grad || b->grad) {
-        // TODO: support backward pass for broadcasting
-        GGML_ASSERT(ggml_are_same_shape(a, b));
-        is_node = true;
-    }
+    // currently only supported for quantized input and f16
+    GGML_ASSERT(ggml_is_quantized(a->type) ||
+                a->type == GGML_TYPE_F16 ||
+                a->type == GGML_TYPE_BF16);
 
     struct ggml_tensor * result = ggml_new_tensor(ctx, type, GGML_MAX_DIMS, a->ne);
 
-    result->op   = GGML_OP_ADD;
-    result->grad = is_node ? ggml_new_tensor(ctx, GGML_TYPE_F32, GGML_MAX_DIMS, a->ne) : NULL;
+    result->op     = GGML_OP_ADD;
     result->src[0] = a;
     result->src[1] = b;
 
@@ -3242,9 +4849,9 @@ static struct ggml_tensor * ggml_add_cast_impl(
 
 struct ggml_tensor * ggml_add_cast(
         struct ggml_context * ctx,
-        struct ggml_tensor * a,
-        struct ggml_tensor * b,
-        enum   ggml_type     type) {
+        struct ggml_tensor  * a,
+        struct ggml_tensor  * b,
+        enum   ggml_type      type) {
     return ggml_add_cast_impl(ctx, a, b, type);
 }
 
@@ -3252,22 +4859,15 @@ struct ggml_tensor * ggml_add_cast(
 
 static struct ggml_tensor * ggml_add1_impl(
         struct ggml_context * ctx,
-        struct ggml_tensor * a,
-        struct ggml_tensor * b,
-        bool inplace) {
+        struct ggml_tensor  * a,
+        struct ggml_tensor  * b,
+        bool                  inplace) {
     GGML_ASSERT(ggml_is_scalar(b));
     GGML_ASSERT(ggml_is_padded_1d(a));
 
-    bool is_node = false;
-
-    if (a->grad || b->grad) {
-        is_node = true;
-    }
-
     struct ggml_tensor * result = inplace ? ggml_view_tensor(ctx, a) : ggml_dup_tensor(ctx, a);
 
-    result->op   = GGML_OP_ADD1;
-    result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL;
+    result->op     = GGML_OP_ADD1;
     result->src[0] = a;
     result->src[1] = b;
 
@@ -3276,15 +4876,15 @@ static struct ggml_tensor * ggml_add1_impl(
 
 struct ggml_tensor * ggml_add1(
         struct ggml_context * ctx,
-        struct ggml_tensor * a,
-        struct ggml_tensor * b) {
+        struct ggml_tensor  * a,
+        struct ggml_tensor  * b) {
     return ggml_add1_impl(ctx, a, b, false);
 }
 
 struct ggml_tensor * ggml_add1_inplace(
         struct ggml_context * ctx,
-        struct ggml_tensor * a,
-        struct ggml_tensor * b) {
+        struct ggml_tensor  * a,
+        struct ggml_tensor  * b) {
     return ggml_add1_impl(ctx, a, b, true);
 }
 
@@ -3292,31 +4892,24 @@ struct ggml_tensor * ggml_add1_inplace(
 
 static struct ggml_tensor * ggml_acc_impl(
         struct ggml_context * ctx,
-        struct ggml_tensor * a,
-        struct ggml_tensor * b,
-        size_t               nb1,
-        size_t               nb2,
-        size_t               nb3,
-        size_t               offset,
-        bool inplace) {
+        struct ggml_tensor  * a,
+        struct ggml_tensor  * b,
+        size_t                nb1,
+        size_t                nb2,
+        size_t                nb3,
+        size_t                offset,
+        bool                  inplace) {
     GGML_ASSERT(ggml_nelements(b) <= ggml_nelements(a));
     GGML_ASSERT(ggml_is_contiguous(a));
     GGML_ASSERT(a->type == GGML_TYPE_F32);
     GGML_ASSERT(b->type == GGML_TYPE_F32);
 
-    bool is_node = false;
-
-    if (!inplace && (a->grad || b->grad)) {
-        is_node = true;
-    }
-
     struct ggml_tensor * result = inplace ? ggml_view_tensor(ctx, a) : ggml_dup_tensor(ctx, a);
 
     int32_t params[] = { nb1, nb2, nb3, offset, inplace ? 1 : 0 };
     ggml_set_op_params(result, params, sizeof(params));
 
-    result->op   = GGML_OP_ACC;
-    result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL;
+    result->op     = GGML_OP_ACC;
     result->src[0] = a;
     result->src[1] = b;
 
@@ -3325,23 +4918,23 @@ static struct ggml_tensor * ggml_acc_impl(
 
 struct ggml_tensor * ggml_acc(
         struct ggml_context * ctx,
-        struct ggml_tensor * a,
-        struct ggml_tensor * b,
-        size_t               nb1,
-        size_t               nb2,
-        size_t               nb3,
-        size_t               offset) {
+        struct ggml_tensor  * a,
+        struct ggml_tensor  * b,
+        size_t                nb1,
+        size_t                nb2,
+        size_t                nb3,
+        size_t                offset) {
     return ggml_acc_impl(ctx, a, b, nb1, nb2, nb3, offset, false);
 }
 
 struct ggml_tensor * ggml_acc_inplace(
         struct ggml_context * ctx,
-        struct ggml_tensor * a,
-        struct ggml_tensor * b,
-        size_t               nb1,
-        size_t               nb2,
-        size_t               nb3,
-        size_t               offset) {
+        struct ggml_tensor  * a,
+        struct ggml_tensor  * b,
+        size_t                nb1,
+        size_t                nb2,
+        size_t                nb3,
+        size_t                offset) {
     return ggml_acc_impl(ctx, a, b, nb1, nb2, nb3, offset, true);
 }
 
@@ -3349,21 +4942,14 @@ struct ggml_tensor * ggml_acc_inplace(
 
 static struct ggml_tensor * ggml_sub_impl(
         struct ggml_context * ctx,
-        struct ggml_tensor * a,
-        struct ggml_tensor * b,
-        bool inplace) {
-    GGML_ASSERT(ggml_are_same_shape(a, b));
-
-    bool is_node = false;
-
-    if (!inplace && (a->grad || b->grad)) {
-        is_node = true;
-    }
+        struct ggml_tensor  * a,
+        struct ggml_tensor  * b,
+        bool                  inplace) {
+    GGML_ASSERT(ggml_can_repeat(b, a));
 
     struct ggml_tensor * result = inplace ? ggml_view_tensor(ctx, a) : ggml_dup_tensor(ctx, a);
 
-    result->op   = GGML_OP_SUB;
-    result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL;
+    result->op     = GGML_OP_SUB;
     result->src[0] = a;
     result->src[1] = b;
 
@@ -3372,15 +4958,15 @@ static struct ggml_tensor * ggml_sub_impl(
 
 struct ggml_tensor * ggml_sub(
         struct ggml_context * ctx,
-        struct ggml_tensor * a,
-        struct ggml_tensor * b) {
+        struct ggml_tensor  * a,
+        struct ggml_tensor  * b) {
     return ggml_sub_impl(ctx, a, b, false);
 }
 
 struct ggml_tensor * ggml_sub_inplace(
         struct ggml_context * ctx,
-        struct ggml_tensor * a,
-        struct ggml_tensor * b) {
+        struct ggml_tensor  * a,
+        struct ggml_tensor  * b) {
     return ggml_sub_impl(ctx, a, b, true);
 }
 
@@ -3388,27 +4974,14 @@ struct ggml_tensor * ggml_sub_inplace(
 
 static struct ggml_tensor * ggml_mul_impl(
         struct ggml_context * ctx,
-        struct ggml_tensor * a,
-        struct ggml_tensor * b,
-        bool inplace) {
+        struct ggml_tensor  * a,
+        struct ggml_tensor  * b,
+        bool                  inplace) {
     GGML_ASSERT(ggml_can_repeat(b, a));
 
-    bool is_node = false;
-
-    if (!inplace && (a->grad || b->grad)) {
-        // TODO: support backward pass for broadcasting
-        GGML_ASSERT(ggml_are_same_shape(a, b));
-        is_node = true;
-    }
-
-    if (inplace) {
-        GGML_ASSERT(!is_node);
-    }
-
     struct ggml_tensor * result = inplace ? ggml_view_tensor(ctx, a) : ggml_dup_tensor(ctx, a);
 
-    result->op   = GGML_OP_MUL;
-    result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL;
+    result->op     = GGML_OP_MUL;
     result->src[0] = a;
     result->src[1] = b;
 
@@ -3433,25 +5006,14 @@ struct ggml_tensor * ggml_mul_inplace(
 
 static struct ggml_tensor * ggml_div_impl(
         struct ggml_context * ctx,
-        struct ggml_tensor * a,
-        struct ggml_tensor * b,
-        bool inplace) {
+        struct ggml_tensor  * a,
+        struct ggml_tensor  * b,
+        bool                  inplace) {
     GGML_ASSERT(ggml_can_repeat(b, a));
 
-    bool is_node = false;
-
-    if (!inplace && (a->grad || b->grad)) {
-        is_node = true;
-    }
-
-    if (inplace) {
-        GGML_ASSERT(!is_node);
-    }
-
     struct ggml_tensor * result = inplace ? ggml_view_tensor(ctx, a) : ggml_dup_tensor(ctx, a);
 
-    result->op   = GGML_OP_DIV;
-    result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL;
+    result->op     = GGML_OP_DIV;
     result->src[0] = a;
     result->src[1] = b;
 
@@ -3476,18 +5038,11 @@ struct ggml_tensor * ggml_div_inplace(
 
 static struct ggml_tensor * ggml_sqr_impl(
         struct ggml_context * ctx,
-        struct ggml_tensor * a,
-        bool inplace) {
-    bool is_node = false;
-
-    if (!inplace && (a->grad)) {
-        is_node = true;
-    }
-
+        struct ggml_tensor  * a,
+        bool                  inplace) {
     struct ggml_tensor * result = inplace ? ggml_view_tensor(ctx, a) : ggml_dup_tensor(ctx, a);
 
-    result->op   = GGML_OP_SQR;
-    result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL;
+    result->op     = GGML_OP_SQR;
     result->src[0] = a;
 
     return result;
@@ -3509,18 +5064,11 @@ struct ggml_tensor * ggml_sqr_inplace(
 
 static struct ggml_tensor * ggml_sqrt_impl(
         struct ggml_context * ctx,
-        struct ggml_tensor * a,
-        bool inplace) {
-    bool is_node = false;
-
-    if (!inplace && (a->grad)) {
-        is_node = true;
-    }
-
+        struct ggml_tensor  * a,
+        bool                  inplace) {
     struct ggml_tensor * result = inplace ? ggml_view_tensor(ctx, a) : ggml_dup_tensor(ctx, a);
 
-    result->op   = GGML_OP_SQRT;
-    result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL;
+    result->op     = GGML_OP_SQRT;
     result->src[0] = a;
 
     return result;
@@ -3543,17 +5091,10 @@ struct ggml_tensor * ggml_sqrt_inplace(
 static struct ggml_tensor * ggml_log_impl(
         struct ggml_context * ctx,
         struct ggml_tensor  * a,
-        bool inplace) {
-    bool is_node = false;
-
-    if (!inplace && (a->grad)) {
-        is_node = true;
-    }
-
+        bool                  inplace) {
     struct ggml_tensor * result = inplace ? ggml_view_tensor(ctx, a) : ggml_dup_tensor(ctx, a);
 
-    result->op   = GGML_OP_LOG;
-    result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL;
+    result->op     = GGML_OP_LOG;
     result->src[0] = a;
 
     return result;
@@ -3571,21 +5112,66 @@ struct ggml_tensor * ggml_log_inplace(
     return ggml_log_impl(ctx, a, true);
 }
 
-// ggml_sum
+// ggml_sin
 
-struct ggml_tensor * ggml_sum(
+static struct ggml_tensor * ggml_sin_impl(
         struct ggml_context * ctx,
-        struct ggml_tensor * a) {
-    bool is_node = false;
+        struct ggml_tensor  * a,
+        bool                  inplace) {
+    struct ggml_tensor * result = inplace ? ggml_view_tensor(ctx, a) : ggml_dup_tensor(ctx, a);
 
-    if (a->grad) {
-        is_node = true;
-    }
+    result->op     = GGML_OP_SIN;
+    result->src[0] = a;
+
+    return result;
+}
+
+struct ggml_tensor * ggml_sin(
+        struct ggml_context * ctx,
+        struct ggml_tensor  * a) {
+    return ggml_sin_impl(ctx, a, false);
+}
+
+struct ggml_tensor * ggml_sin_inplace(
+        struct ggml_context * ctx,
+        struct ggml_tensor  * a) {
+    return ggml_sin_impl(ctx, a, true);
+}
+
+// ggml_cos
+
+static struct ggml_tensor * ggml_cos_impl(
+        struct ggml_context * ctx,
+        struct ggml_tensor  * a,
+        bool                  inplace) {
+    struct ggml_tensor * result = inplace ? ggml_view_tensor(ctx, a) : ggml_dup_tensor(ctx, a);
+
+    result->op     = GGML_OP_COS;
+    result->src[0] = a;
+
+    return result;
+}
+
+struct ggml_tensor * ggml_cos(
+        struct ggml_context * ctx,
+        struct ggml_tensor  * a) {
+    return ggml_cos_impl(ctx, a, false);
+}
 
+struct ggml_tensor * ggml_cos_inplace(
+        struct ggml_context * ctx,
+        struct ggml_tensor  * a) {
+    return ggml_cos_impl(ctx, a, true);
+}
+
+// ggml_sum
+
+struct ggml_tensor * ggml_sum(
+        struct ggml_context * ctx,
+        struct ggml_tensor  * a) {
     struct ggml_tensor * result = ggml_new_tensor_1d(ctx, a->type, 1);
 
-    result->op   = GGML_OP_SUM;
-    result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL;
+    result->op     = GGML_OP_SUM;
     result->src[0] = a;
 
     return result;
@@ -3595,13 +5181,7 @@ struct ggml_tensor * ggml_sum(
 
 struct ggml_tensor * ggml_sum_rows(
         struct ggml_context * ctx,
-        struct ggml_tensor * a) {
-    bool is_node = false;
-
-    if (a->grad) {
-        is_node = true;
-    }
-
+        struct ggml_tensor  * a) {
     int64_t ne[GGML_MAX_DIMS] = { 1 };
     for (int i = 1; i < GGML_MAX_DIMS; ++i) {
         ne[i] = a->ne[i];
@@ -3609,8 +5189,7 @@ struct ggml_tensor * ggml_sum_rows(
 
     struct ggml_tensor * result = ggml_new_tensor(ctx, a->type, GGML_MAX_DIMS, ne);
 
-    result->op   = GGML_OP_SUM_ROWS;
-    result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL;
+    result->op     = GGML_OP_SUM_ROWS;
     result->src[0] = a;
 
     return result;
@@ -3620,19 +5199,11 @@ struct ggml_tensor * ggml_sum_rows(
 
 struct ggml_tensor * ggml_mean(
         struct ggml_context * ctx,
-        struct ggml_tensor * a) {
-    bool is_node = false;
-
-    if (a->grad) {
-        GGML_ASSERT(false); // TODO: implement
-        is_node = true;
-    }
-
+        struct ggml_tensor  * a) {
     int64_t ne[4] = { 1, a->ne[1], a->ne[2], a->ne[3] };
     struct ggml_tensor * result = ggml_new_tensor(ctx, GGML_TYPE_F32, 4, ne);
 
-    result->op   = GGML_OP_MEAN;
-    result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL;
+    result->op     = GGML_OP_MEAN;
     result->src[0] = a;
 
     return result;
@@ -3642,20 +5213,30 @@ struct ggml_tensor * ggml_mean(
 
 struct ggml_tensor * ggml_argmax(
         struct ggml_context * ctx,
-        struct ggml_tensor * a) {
+        struct ggml_tensor  * a) {
     GGML_ASSERT(ggml_is_matrix(a));
-    bool is_node = false;
-
-    if (a->grad) {
-        GGML_ASSERT(false);
-        is_node = true;
-    }
 
     struct ggml_tensor * result = ggml_new_tensor_1d(ctx, GGML_TYPE_I32, a->ne[1]);
 
-    result->op   = GGML_OP_ARGMAX;
-    result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL;
+    result->op     = GGML_OP_ARGMAX;
+    result->src[0] = a;
+
+    return result;
+}
+
+// ggml_count_equal
+
+struct ggml_tensor * ggml_count_equal(
+        struct ggml_context * ctx,
+        struct ggml_tensor  * a,
+        struct ggml_tensor  * b) {
+    GGML_ASSERT(ggml_are_same_shape(a, b));
+
+    struct ggml_tensor * result = ggml_new_tensor_1d(ctx, GGML_TYPE_I64, 1);
+
+    result->op     = GGML_OP_COUNT_EQUAL;
     result->src[0] = a;
+    result->src[1] = b;
 
     return result;
 }
@@ -3664,20 +5245,13 @@ struct ggml_tensor * ggml_argmax(
 
 struct ggml_tensor * ggml_repeat(
         struct ggml_context * ctx,
-        struct ggml_tensor * a,
-        struct ggml_tensor * b) {
+        struct ggml_tensor  * a,
+        struct ggml_tensor  * b) {
     GGML_ASSERT(ggml_can_repeat(a, b));
 
-    bool is_node = false;
-
-    if (a->grad) {
-        is_node = true;
-    }
-
     struct ggml_tensor * result = ggml_new_tensor(ctx, a->type, GGML_MAX_DIMS, b->ne);
 
-    result->op   = GGML_OP_REPEAT;
-    result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL;
+    result->op     = GGML_OP_REPEAT;
     result->src[0] = a;
 
     return result;
@@ -3687,24 +5261,13 @@ struct ggml_tensor * ggml_repeat(
 
 struct ggml_tensor * ggml_repeat_back(
         struct ggml_context * ctx,
-        struct ggml_tensor * a,
-        struct ggml_tensor * b) {
+        struct ggml_tensor  * a,
+        struct ggml_tensor  * b) {
     GGML_ASSERT(ggml_can_repeat(b, a));
 
-    bool is_node = false;
-
-    if (a->grad) {
-        is_node = true;
-    }
-
-    if (ggml_are_same_shape(a, b) && !is_node) {
-        return a;
-    }
-
     struct ggml_tensor * result = ggml_new_tensor(ctx, a->type, GGML_MAX_DIMS, b->ne);
 
-    result->op   = GGML_OP_REPEAT_BACK;
-    result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL;
+    result->op     = GGML_OP_REPEAT_BACK;
     result->src[0] = a;
 
     return result;
@@ -3713,21 +5276,27 @@ struct ggml_tensor * ggml_repeat_back(
 // ggml_concat
 
 struct ggml_tensor * ggml_concat(
-    struct ggml_context* ctx,
-    struct ggml_tensor* a,
-    struct ggml_tensor* b) {
-    GGML_ASSERT(a->ne[0] == b->ne[0] && a->ne[1] == b->ne[1] && a->ne[3] == b->ne[3]);
-
-    bool is_node = false;
+    struct ggml_context * ctx,
+    struct ggml_tensor  * a,
+    struct ggml_tensor  * b,
+    int                   dim) {
+    GGML_ASSERT(dim >= 0 && dim < GGML_MAX_DIMS);
 
-    if (a->grad || b->grad) {
-        is_node = true;
+    int64_t ne[GGML_MAX_DIMS];
+    for (int d = 0; d < GGML_MAX_DIMS; ++d) {
+        if (d == dim) {
+            ne[d] = a->ne[d] + b->ne[d];
+            continue;
+        }
+        GGML_ASSERT(a->ne[d] == b->ne[d]);
+        ne[d] = a->ne[d];
     }
 
-    struct ggml_tensor * result = ggml_new_tensor_4d(ctx, a->type, a->ne[0], a->ne[1], a->ne[2] + b->ne[2], a->ne[3]);
+    struct ggml_tensor * result = ggml_new_tensor(ctx, a->type, GGML_MAX_DIMS, ne);
 
-    result->op = GGML_OP_CONCAT;
-    result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL;
+    ggml_set_op_params_i32(result, 0, dim);
+
+    result->op     = GGML_OP_CONCAT;
     result->src[0] = a;
     result->src[1] = b;
 
@@ -3836,23 +5405,33 @@ struct ggml_tensor * ggml_relu_inplace(
 
 struct ggml_tensor * ggml_leaky_relu(
         struct ggml_context * ctx,
-        struct ggml_tensor  * a, float negative_slope, bool inplace) {
-    bool is_node = false;
-
-    if (!inplace && (a->grad)) {
-        is_node = true;
-    }
-
+        struct ggml_tensor  * a,
+        float                 negative_slope,
+        bool                  inplace) {
     struct ggml_tensor * result = inplace ? ggml_view_tensor(ctx, a) : ggml_dup_tensor(ctx, a);
+
     ggml_set_op_params(result, &negative_slope, sizeof(negative_slope));
 
-    result->op   = GGML_OP_LEAKY_RELU;
-    result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL;
+    result->op     = GGML_OP_LEAKY_RELU;
     result->src[0] = a;
 
     return result;
 }
 
+// ggml_sigmoid
+
+struct ggml_tensor * ggml_sigmoid(
+        struct ggml_context * ctx,
+        struct ggml_tensor  * a) {
+    return ggml_unary(ctx, a, GGML_UNARY_OP_SIGMOID);
+}
+
+struct ggml_tensor * ggml_sigmoid_inplace(
+        struct ggml_context * ctx,
+        struct ggml_tensor  * a) {
+    return ggml_unary_inplace(ctx, a, GGML_UNARY_OP_SIGMOID);
+}
+
 // ggml_gelu
 
 struct ggml_tensor * ggml_gelu(
@@ -3901,43 +5480,57 @@ struct ggml_tensor * ggml_silu_back(
         struct ggml_context * ctx,
         struct ggml_tensor  * a,
         struct ggml_tensor  * b) {
-    bool is_node = false;
-
-    if (a->grad || b->grad) {
-        // TODO: implement backward
-        is_node = true;
-    }
-
     struct ggml_tensor * result = ggml_dup_tensor(ctx, a);
 
-    result->op   = GGML_OP_SILU_BACK;
-    result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL;
+    result->op     = GGML_OP_SILU_BACK;
     result->src[0] = a;
     result->src[1] = b;
 
     return result;
 }
 
+// ggml hardswish
+
+struct ggml_tensor * ggml_hardswish(
+        struct ggml_context * ctx,
+        struct ggml_tensor  * a) {
+    return ggml_unary(ctx, a, GGML_UNARY_OP_HARDSWISH);
+}
+
+// ggml hardsigmoid
+
+struct ggml_tensor * ggml_hardsigmoid(
+        struct ggml_context * ctx,
+        struct ggml_tensor  * a) {
+    return ggml_unary(ctx, a, GGML_UNARY_OP_HARDSIGMOID);
+}
+
+// ggml exp
+
+struct ggml_tensor * ggml_exp(
+        struct ggml_context * ctx,
+        struct ggml_tensor  * a) {
+    return ggml_unary(ctx, a, GGML_UNARY_OP_EXP);
+}
+
+struct ggml_tensor * ggml_exp_inplace(
+        struct ggml_context * ctx,
+        struct ggml_tensor  * a) {
+    return ggml_unary_inplace(ctx, a, GGML_UNARY_OP_EXP);
+}
+
 // ggml_norm
 
 static struct ggml_tensor * ggml_norm_impl(
         struct ggml_context * ctx,
         struct ggml_tensor  * a,
-        float eps,
-        bool inplace) {
-    bool is_node = false;
-
-    if (!inplace && (a->grad)) {
-        GGML_ASSERT(false); // TODO: implement backward
-        is_node = true;
-    }
-
+        float                 eps,
+        bool                  inplace) {
     struct ggml_tensor * result = inplace ? ggml_view_tensor(ctx, a) : ggml_dup_tensor(ctx, a);
 
     ggml_set_op_params(result, &eps, sizeof(eps));
 
-    result->op   = GGML_OP_NORM;
-    result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL;
+    result->op     = GGML_OP_NORM;
     result->src[0] = a;
 
     return result;
@@ -3946,14 +5539,14 @@ static struct ggml_tensor * ggml_norm_impl(
 struct ggml_tensor * ggml_norm(
         struct ggml_context * ctx,
         struct ggml_tensor  * a,
-        float eps) {
+        float                 eps) {
     return ggml_norm_impl(ctx, a, eps, false);
 }
 
 struct ggml_tensor * ggml_norm_inplace(
         struct ggml_context * ctx,
         struct ggml_tensor  * a,
-        float eps) {
+        float                 eps) {
     return ggml_norm_impl(ctx, a, eps, true);
 }
 
@@ -3962,20 +5555,13 @@ struct ggml_tensor * ggml_norm_inplace(
 static struct ggml_tensor * ggml_rms_norm_impl(
         struct ggml_context * ctx,
         struct ggml_tensor  * a,
-        float eps,
-        bool inplace) {
-    bool is_node = false;
-
-    if (!inplace && (a->grad)) {
-        is_node = true;
-    }
-
+        float                 eps,
+        bool                  inplace) {
     struct ggml_tensor * result = inplace ? ggml_view_tensor(ctx, a) : ggml_dup_tensor(ctx, a);
 
     ggml_set_op_params(result, &eps, sizeof(eps));
 
-    result->op   = GGML_OP_RMS_NORM;
-    result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL;
+    result->op     = GGML_OP_RMS_NORM;
     result->src[0] = a;
 
     return result;
@@ -3984,14 +5570,14 @@ static struct ggml_tensor * ggml_rms_norm_impl(
 struct ggml_tensor * ggml_rms_norm(
         struct ggml_context * ctx,
         struct ggml_tensor  * a,
-        float  eps) {
+        float                 eps) {
     return ggml_rms_norm_impl(ctx, a, eps, false);
 }
 
 struct ggml_tensor * ggml_rms_norm_inplace(
         struct ggml_context * ctx,
         struct ggml_tensor  * a,
-        float eps) {
+        float                 eps) {
     return ggml_rms_norm_impl(ctx, a, eps, true);
 }
 
@@ -4001,20 +5587,12 @@ struct ggml_tensor * ggml_rms_norm_back(
         struct ggml_context * ctx,
         struct ggml_tensor  * a,
         struct ggml_tensor  * b,
-        float  eps) {
-    bool is_node = false;
-
-    if (a->grad) {
-        // TODO: implement backward
-        is_node = true;
-    }
-
+        float                 eps) {
     struct ggml_tensor * result = ggml_dup_tensor(ctx, a);
 
     ggml_set_op_params(result, &eps, sizeof(eps));
 
-    result->op   = GGML_OP_RMS_NORM_BACK;
-    result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL;
+    result->op     = GGML_OP_RMS_NORM_BACK;
     result->src[0] = a;
     result->src[1] = b;
 
@@ -4024,40 +5602,36 @@ struct ggml_tensor * ggml_rms_norm_back(
 // ggml_group_norm
 
 static struct ggml_tensor * ggml_group_norm_impl(
-    struct ggml_context * ctx,
-    struct ggml_tensor * a,
-    int n_groups,
-    bool inplace) {
-
-    bool is_node = false;
-    if (!inplace && (a->grad)) {
-        GGML_ASSERT(false); // TODO: implement backward
-        is_node = true;
-    }
-
+        struct ggml_context * ctx,
+        struct ggml_tensor  * a,
+        int                   n_groups,
+        float                 eps,
+        bool                  inplace) {
     struct ggml_tensor * result = inplace ? ggml_view_tensor(ctx, a) : ggml_dup_tensor(ctx, a);
 
-    result->op_params[0] = n_groups;
+    ggml_set_op_params_i32(result, 0, n_groups);
+    ggml_set_op_params_f32(result, 1, eps);
 
-    result->op = GGML_OP_GROUP_NORM;
-    result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL;
+    result->op     = GGML_OP_GROUP_NORM;
     result->src[0] = a;
 
     return result;
 }
 
 struct ggml_tensor * ggml_group_norm(
-    struct ggml_context * ctx,
-    struct ggml_tensor * a,
-    int n_groups) {
-    return ggml_group_norm_impl(ctx, a, n_groups, false);
+        struct ggml_context * ctx,
+        struct ggml_tensor  * a,
+        int                   n_groups,
+        float                 eps) {
+    return ggml_group_norm_impl(ctx, a, n_groups, eps, false);
 }
 
 struct ggml_tensor * ggml_group_norm_inplace(
-    struct ggml_context * ctx,
-    struct ggml_tensor * a,
-    int n_groups) {
-    return ggml_group_norm_impl(ctx, a, n_groups, true);
+        struct ggml_context * ctx,
+        struct ggml_tensor  * a,
+        int                   n_groups,
+        float                 eps) {
+    return ggml_group_norm_impl(ctx, a, n_groups, eps, true);
 }
 
 // ggml_mul_mat
@@ -4069,17 +5643,10 @@ struct ggml_tensor * ggml_mul_mat(
     GGML_ASSERT(ggml_can_mul_mat(a, b));
     GGML_ASSERT(!ggml_is_transposed(a));
 
-    bool is_node = false;
-
-    if (a->grad || b->grad) {
-        is_node = true;
-    }
-
     const int64_t ne[4] = { a->ne[1], b->ne[1], b->ne[2], b->ne[3] };
     struct ggml_tensor * result = ggml_new_tensor(ctx, GGML_TYPE_F32, 4, ne);
 
-    result->op   = GGML_OP_MUL_MAT;
-    result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL;
+    result->op     = GGML_OP_MUL_MAT;
     result->src[0] = a;
     result->src[1] = b;
 
@@ -4089,6 +5656,8 @@ struct ggml_tensor * ggml_mul_mat(
 void ggml_mul_mat_set_prec(
         struct ggml_tensor * a,
         enum ggml_prec       prec) {
+    GGML_ASSERT(a->op == GGML_OP_MUL_MAT);
+
     const int32_t prec_i32 = (int32_t) prec;
 
     ggml_set_op_params_i32(a, 0, prec_i32);
@@ -4096,45 +5665,40 @@ void ggml_mul_mat_set_prec(
 
 // ggml_mul_mat_id
 
+/*
+    c = ggml_mul_mat_id(ctx, as, b, ids);
+
+    as  -> [cols, rows, n_expert]
+    ids -> [n_experts_used, n_tokens] (i32)
+    b   -> [cols, n_expert_used, n_tokens]
+    c   -> [rows, n_expert_used, n_tokens]
+
+    in b, n_experts_used can be broadcasted to match the n_expert_used of ids
+
+    c ~= as[:,:,i] @ b[:,i%r,t], i = ids[e,t] for all e,t in ids
+*/
 struct ggml_tensor * ggml_mul_mat_id(
         struct ggml_context * ctx,
-        struct ggml_tensor  * const as[],
-        int                   n_as,
-        struct ggml_tensor  * ids,
-        int                   id,
-        struct ggml_tensor  * b) {
-
+        struct ggml_tensor  * as,
+        struct ggml_tensor  * b,
+        struct ggml_tensor  * ids) {
+    GGML_ASSERT(!ggml_is_transposed(as));
     GGML_ASSERT(ids->type == GGML_TYPE_I32);
-    GGML_ASSERT(ids->ne[2] == 1 && ids->ne[3] == 1);
-    GGML_ASSERT(ids->ne[1] == b->ne[1]);
-    GGML_ASSERT(ids->ne[2] == b->ne[2] && ids->ne[3] == b->ne[3]);
-    GGML_ASSERT(n_as > 0 && n_as <= GGML_MAX_SRC - 2);
-    GGML_ASSERT(id >= 0 && id < ids->ne[0]);
-
-    bool is_node = false;
 
-    if (as[0]->grad || b->grad) {
-        is_node = true;
-    }
+    GGML_ASSERT(as->ne[3] == 1); // as is 3d (one matrix per expert)
+    GGML_ASSERT(b->ne[3] == 1); // b is 3d
+    GGML_ASSERT(ids->ne[2] == 1 && ids->ne[3] == 1); // ids is 2d
+    GGML_ASSERT(ids->ne[1] == b->ne[2]); // must have an expert list per b row
+    GGML_ASSERT(as->ne[0] == b->ne[0]); // can_mul_mat
+    GGML_ASSERT(ids->ne[0] % b->ne[1] == 0); // can broadcast
 
-    const int64_t ne[4] = { as[0]->ne[1], b->ne[1], b->ne[2], b->ne[3] };
+    const int64_t ne[4] = { as->ne[1], ids->ne[0], b->ne[2], 1 };
     struct ggml_tensor * result = ggml_new_tensor(ctx, GGML_TYPE_F32, 4, ne);
 
-    ggml_set_op_params_i32(result, 0, id);
-    ggml_set_op_params_i32(result, 1, n_as);
-
-    result->op   = GGML_OP_MUL_MAT_ID;
-    result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL;
-    result->src[0] = ids;
+    result->op     = GGML_OP_MUL_MAT_ID;
+    result->src[0] = as;
     result->src[1] = b;
-
-    for (int i = 0; i < n_as; i++) {
-        struct ggml_tensor * a = as[i];
-        GGML_ASSERT(ggml_are_same_shape(as[0], a));
-        GGML_ASSERT(ggml_can_mul_mat(a, b));
-        GGML_ASSERT(!ggml_is_transposed(a));
-        result->src[i + 2] = a;
-    }
+    result->src[2] = ids;
 
     return result;
 }
@@ -4148,18 +5712,11 @@ struct ggml_tensor * ggml_out_prod(
     GGML_ASSERT(ggml_can_out_prod(a, b));
     GGML_ASSERT(!ggml_is_transposed(a));
 
-    bool is_node = false;
-
-    if (a->grad || b->grad) {
-        is_node = true;
-    }
-
     // a is broadcastable to b for ne[2] and ne[3] -> use b->ne[2] and b->ne[3]
     const int64_t ne[4] = { a->ne[0], b->ne[0], b->ne[2], b->ne[3] };
     struct ggml_tensor * result = ggml_new_tensor(ctx, GGML_TYPE_F32, 4, ne);
 
-    result->op   = GGML_OP_OUT_PROD;
-    result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL;
+    result->op     = GGML_OP_OUT_PROD;
     result->src[0] = a;
     result->src[1] = b;
 
@@ -4172,21 +5729,14 @@ static struct ggml_tensor * ggml_scale_impl(
         struct ggml_context * ctx,
         struct ggml_tensor  * a,
         float                 s,
-        bool inplace) {
+        bool                  inplace) {
     GGML_ASSERT(ggml_is_padded_1d(a));
 
-    bool is_node = false;
-
-    if (a->grad) {
-        is_node = true;
-    }
-
     struct ggml_tensor * result = inplace ? ggml_view_tensor(ctx, a) : ggml_dup_tensor(ctx, a);
 
     ggml_set_op_params(result, &s, sizeof(s));
 
-    result->op   = GGML_OP_SCALE;
-    result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL;
+    result->op     = GGML_OP_SCALE;
     result->src[0] = a;
 
     return result;
@@ -4194,15 +5744,15 @@ static struct ggml_tensor * ggml_scale_impl(
 
 struct ggml_tensor * ggml_scale(
         struct ggml_context * ctx,
-        struct ggml_tensor * a,
-        float                s) {
+        struct ggml_tensor  * a,
+        float                 s) {
     return ggml_scale_impl(ctx, a, s, false);
 }
 
 struct ggml_tensor * ggml_scale_inplace(
         struct ggml_context * ctx,
-        struct ggml_tensor * a,
-        float                s) {
+        struct ggml_tensor  * a,
+        float                 s) {
     return ggml_scale_impl(ctx, a, s, true);
 }
 
@@ -4216,23 +5766,17 @@ static struct ggml_tensor * ggml_set_impl(
         size_t                nb2,
         size_t                nb3,
         size_t                offset,
-        bool inplace) {
+        bool                  inplace) {
     GGML_ASSERT(ggml_nelements(a) >= ggml_nelements(b));
 
-    bool is_node = false;
-
-    if (a->grad || b->grad) {
-        is_node = true;
-    }
-
     // make a view of the destination
     struct ggml_tensor * result = inplace ? ggml_view_tensor(ctx, a) : ggml_dup_tensor(ctx, a);
 
+    GGML_ASSERT(offset < (size_t)(1 << 30));
     int32_t params[] = { nb1, nb2, nb3, offset, inplace ? 1 : 0 };
     ggml_set_op_params(result, params, sizeof(params));
 
-    result->op   = GGML_OP_SET;
-    result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL;
+    result->op     = GGML_OP_SET;
     result->src[0] = a;
     result->src[1] = b;
 
@@ -4241,8 +5785,8 @@ static struct ggml_tensor * ggml_set_impl(
 
 struct ggml_tensor * ggml_set(
         struct ggml_context * ctx,
-        struct ggml_tensor *  a,
-        struct ggml_tensor *  b,
+        struct ggml_tensor  * a,
+        struct ggml_tensor  * b,
         size_t                nb1,
         size_t                nb2,
         size_t                nb3,
@@ -4252,8 +5796,8 @@ struct ggml_tensor * ggml_set(
 
 struct ggml_tensor * ggml_set_inplace(
         struct ggml_context * ctx,
-        struct ggml_tensor *  a,
-        struct ggml_tensor *  b,
+        struct ggml_tensor  * a,
+        struct ggml_tensor  * b,
         size_t                nb1,
         size_t                nb2,
         size_t                nb3,
@@ -4263,24 +5807,24 @@ struct ggml_tensor * ggml_set_inplace(
 
 struct ggml_tensor * ggml_set_1d(
         struct ggml_context * ctx,
-        struct ggml_tensor *  a,
-        struct ggml_tensor *  b,
+        struct ggml_tensor  * a,
+        struct ggml_tensor  * b,
         size_t                offset) {
     return ggml_set_impl(ctx, a, b, a->nb[1], a->nb[2], a->nb[3], offset, false);
 }
 
 struct ggml_tensor * ggml_set_1d_inplace(
         struct ggml_context * ctx,
-        struct ggml_tensor *  a,
-        struct ggml_tensor *  b,
+        struct ggml_tensor  * a,
+        struct ggml_tensor  * b,
         size_t                offset) {
     return ggml_set_impl(ctx, a, b, a->nb[1], a->nb[2], a->nb[3], offset, true);
 }
 
 struct ggml_tensor * ggml_set_2d(
         struct ggml_context * ctx,
-        struct ggml_tensor *  a,
-        struct ggml_tensor *  b,
+        struct ggml_tensor  * a,
+        struct ggml_tensor  * b,
         size_t                nb1,
         size_t                offset) {
     return ggml_set_impl(ctx, a, b, nb1, a->nb[2], a->nb[3], offset, false);
@@ -4288,8 +5832,8 @@ struct ggml_tensor * ggml_set_2d(
 
 struct ggml_tensor * ggml_set_2d_inplace(
         struct ggml_context * ctx,
-        struct ggml_tensor *  a,
-        struct ggml_tensor *  b,
+        struct ggml_tensor  * a,
+        struct ggml_tensor  * b,
         size_t                nb1,
         size_t                offset) {
     return ggml_set_impl(ctx, a, b, nb1, a->nb[2], a->nb[3], offset, true);
@@ -4300,16 +5844,9 @@ struct ggml_tensor * ggml_set_2d_inplace(
 static struct ggml_tensor * ggml_cpy_impl(
         struct ggml_context * ctx,
         struct ggml_tensor  * a,
-        struct ggml_tensor  * b,
-        bool inplace) {
+        struct ggml_tensor  * b) {
     GGML_ASSERT(ggml_nelements(a) == ggml_nelements(b));
 
-    bool is_node = false;
-
-    if (!inplace && (a->grad || b->grad)) {
-        is_node = true;
-    }
-
     // make a view of the destination
     struct ggml_tensor * result = ggml_view_tensor(ctx, b);
     if (strlen(b->name) > 0) {
@@ -4318,8 +5855,7 @@ static struct ggml_tensor * ggml_cpy_impl(
         ggml_format_name(result, "%s (copy)", a->name);
     }
 
-    result->op   = GGML_OP_CPY;
-    result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL;
+    result->op     = GGML_OP_CPY;
     result->src[0] = a;
     result->src[1] = b;
 
@@ -4330,33 +5866,32 @@ struct ggml_tensor * ggml_cpy(
         struct ggml_context * ctx,
         struct ggml_tensor * a,
         struct ggml_tensor * b) {
-    return ggml_cpy_impl(ctx, a, b, false);
+    return ggml_cpy_impl(ctx, a, b);
 }
 
-struct ggml_tensor * ggml_cpy_inplace(
+struct ggml_tensor * ggml_cast(
         struct ggml_context * ctx,
-        struct ggml_tensor * a,
-        struct ggml_tensor * b) {
-    return ggml_cpy_impl(ctx, a, b, true);
+        struct ggml_tensor  * a,
+        enum   ggml_type      type) {
+    struct ggml_tensor * result = ggml_new_tensor(ctx, type, GGML_MAX_DIMS, a->ne);
+    ggml_format_name(result, "%s (copy)", a->name);
+
+    result->op     = GGML_OP_CPY;
+    result->src[0] = a;
+    result->src[1] = result;
+
+    return result;
 }
 
 // ggml_cont
 
 static struct ggml_tensor * ggml_cont_impl(
         struct ggml_context * ctx,
-        struct ggml_tensor  * a,
-        bool inplace) {
-    bool is_node = false;
-
-    if (!inplace && a->grad) {
-        is_node = true;
-    }
-
-    struct ggml_tensor * result = inplace ? ggml_view_tensor(ctx, a) : ggml_dup_tensor(ctx, a);
+        struct ggml_tensor  * a) {
+    struct ggml_tensor * result = ggml_dup_tensor(ctx, a);
     ggml_format_name(result, "%s (cont)", a->name);
 
-    result->op   = GGML_OP_CONT;
-    result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL;
+    result->op     = GGML_OP_CONT;
     result->src[0] = a;
 
     return result;
@@ -4365,13 +5900,7 @@ static struct ggml_tensor * ggml_cont_impl(
 struct ggml_tensor * ggml_cont(
         struct ggml_context * ctx,
         struct ggml_tensor * a) {
-    return ggml_cont_impl(ctx, a, false);
-}
-
-struct ggml_tensor * ggml_cont_inplace(
-        struct ggml_context * ctx,
-        struct ggml_tensor * a) {
-    return ggml_cont_impl(ctx, a, true);
+    return ggml_cont_impl(ctx, a);
 }
 
 // make contiguous, with new shape
@@ -4408,13 +5937,10 @@ struct ggml_tensor * ggml_cont_4d(
         int64_t               ne3) {
     GGML_ASSERT(ggml_nelements(a) == (ne0*ne1*ne2*ne3));
 
-    bool is_node = false;
-
     struct ggml_tensor * result = ggml_new_tensor_4d(ctx, a->type, ne0, ne1, ne2, ne3);
     ggml_format_name(result, "%s (cont)", a->name);
 
-    result->op   = GGML_OP_CONT;
-    result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL;
+    result->op     = GGML_OP_CONT;
     result->src[0] = a;
 
     return result;
@@ -4430,22 +5956,10 @@ struct ggml_tensor * ggml_reshape(
     // as only the shape of b is relevant, and not its memory layout, b is allowed to be non contiguous.
     GGML_ASSERT(ggml_nelements(a) == ggml_nelements(b));
 
-    bool is_node = false;
-
-    if (a->grad) {
-        is_node = true;
-    }
-
-    if (b->grad) {
-        // gradient propagation is not supported
-        //GGML_ASSERT(false);
-    }
-
     struct ggml_tensor * result = ggml_new_tensor_impl(ctx, a->type, GGML_MAX_DIMS, b->ne, a, 0);
     ggml_format_name(result, "%s (reshaped)", a->name);
 
-    result->op   = GGML_OP_RESHAPE;
-    result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL;
+    result->op     = GGML_OP_RESHAPE;
     result->src[0] = a;
 
     return result;
@@ -4458,18 +5972,11 @@ struct ggml_tensor * ggml_reshape_1d(
     GGML_ASSERT(ggml_is_contiguous(a));
     GGML_ASSERT(ggml_nelements(a) == ne0);
 
-    bool is_node = false;
-
-    if (a->grad) {
-        is_node = true;
-    }
-
     const int64_t ne[1] = { ne0 };
     struct ggml_tensor * result = ggml_new_tensor_impl(ctx, a->type, 1, ne, a, 0);
     ggml_format_name(result, "%s (reshaped)", a->name);
 
-    result->op   = GGML_OP_RESHAPE;
-    result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL;
+    result->op     = GGML_OP_RESHAPE;
     result->src[0] = a;
 
     return result;
@@ -4483,18 +5990,11 @@ struct ggml_tensor * ggml_reshape_2d(
     GGML_ASSERT(ggml_is_contiguous(a));
     GGML_ASSERT(ggml_nelements(a) == ne0*ne1);
 
-    bool is_node = false;
-
-    if (a->grad) {
-        is_node = true;
-    }
-
     const int64_t ne[2] = { ne0, ne1 };
     struct ggml_tensor * result = ggml_new_tensor_impl(ctx, a->type, 2, ne, a, 0);
     ggml_format_name(result, "%s (reshaped)", a->name);
 
-    result->op   = GGML_OP_RESHAPE;
-    result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL;
+    result->op     = GGML_OP_RESHAPE;
     result->src[0] = a;
 
     return result;
@@ -4509,18 +6009,11 @@ struct ggml_tensor * ggml_reshape_3d(
     GGML_ASSERT(ggml_is_contiguous(a));
     GGML_ASSERT(ggml_nelements(a) == ne0*ne1*ne2);
 
-    bool is_node = false;
-
-    if (a->grad) {
-        is_node = true;
-    }
-
     const int64_t ne[3] = { ne0, ne1, ne2 };
     struct ggml_tensor * result = ggml_new_tensor_impl(ctx, a->type, 3, ne, a, 0);
     ggml_format_name(result, "%s (reshaped)", a->name);
 
-    result->op   = GGML_OP_RESHAPE;
-    result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL;
+    result->op     = GGML_OP_RESHAPE;
     result->src[0] = a;
 
     return result;
@@ -4536,18 +6029,11 @@ struct ggml_tensor * ggml_reshape_4d(
     GGML_ASSERT(ggml_is_contiguous(a));
     GGML_ASSERT(ggml_nelements(a) == ne0*ne1*ne2*ne3);
 
-    bool is_node = false;
-
-    if (a->grad) {
-        is_node = true;
-    }
-
     const int64_t ne[4] = { ne0, ne1, ne2, ne3 };
     struct ggml_tensor * result = ggml_new_tensor_impl(ctx, a->type, 4, ne, a, 0);
     ggml_format_name(result, "%s (reshaped)", a->name);
 
-    result->op   = GGML_OP_RESHAPE;
-    result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL;
+    result->op     = GGML_OP_RESHAPE;
     result->src[0] = a;
 
     return result;
@@ -4559,20 +6045,12 @@ static struct ggml_tensor * ggml_view_impl(
         int                   n_dims,
         const int64_t       * ne,
         size_t                offset) {
-
-    bool is_node = false;
-
-    if (a->grad) {
-        is_node = true;
-    }
-
     struct ggml_tensor * result = ggml_new_tensor_impl(ctx, a->type, n_dims, ne, a, offset);
     ggml_format_name(result, "%s (view)", a->name);
 
     ggml_set_op_params(result, &offset, sizeof(offset));
 
-    result->op   = GGML_OP_VIEW;
-    result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL;
+    result->op     = GGML_OP_VIEW;
     result->src[0] = a;
 
     return result;
@@ -4585,7 +6063,6 @@ struct ggml_tensor * ggml_view_1d(
         struct ggml_tensor  * a,
         int64_t               ne0,
         size_t                offset) {
-
     struct ggml_tensor * result = ggml_view_impl(ctx, a, 1, &ne0, offset);
 
     return result;
@@ -4600,7 +6077,6 @@ struct ggml_tensor * ggml_view_2d(
         int64_t               ne1,
         size_t                nb1,
         size_t                offset) {
-
     const int64_t ne[2] = { ne0, ne1 };
 
     struct ggml_tensor * result = ggml_view_impl(ctx, a, 2, ne, offset);
@@ -4623,7 +6099,6 @@ struct ggml_tensor * ggml_view_3d(
         size_t                nb1,
         size_t                nb2,
         size_t                offset) {
-
     const int64_t ne[3] = { ne0, ne1, ne2 };
 
     struct ggml_tensor * result = ggml_view_impl(ctx, a, 3, ne, offset);
@@ -4648,7 +6123,6 @@ struct ggml_tensor * ggml_view_4d(
         size_t                nb2,
         size_t                nb3,
         size_t                offset) {
-
     const int64_t ne[4] = { ne0, ne1, ne2, ne3 };
 
     struct ggml_tensor * result = ggml_view_impl(ctx, a, 4, ne, offset);
@@ -4681,12 +6155,6 @@ struct ggml_tensor * ggml_permute(
     GGML_ASSERT(axis1 != axis3);
     GGML_ASSERT(axis2 != axis3);
 
-    bool is_node = false;
-
-    if (a->grad) {
-        is_node = true;
-    }
-
     struct ggml_tensor * result = ggml_view_tensor(ctx, a);
     ggml_format_name(result, "%s (permuted)", a->name);
 
@@ -4713,8 +6181,7 @@ struct ggml_tensor * ggml_permute(
     result->nb[2] = nb[2];
     result->nb[3] = nb[3];
 
-    result->op   = GGML_OP_PERMUTE;
-    result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL;
+    result->op     = GGML_OP_PERMUTE;
     result->src[0] = a;
 
     int32_t params[] = { axis0, axis1, axis2, axis3 };
@@ -4728,12 +6195,6 @@ struct ggml_tensor * ggml_permute(
 struct ggml_tensor * ggml_transpose(
         struct ggml_context * ctx,
         struct ggml_tensor  * a) {
-    bool is_node = false;
-
-    if (a->grad) {
-        is_node = true;
-    }
-
     struct ggml_tensor * result = ggml_view_tensor(ctx, a);
     ggml_format_name(result, "%s (transposed)", a->name);
 
@@ -4743,8 +6204,7 @@ struct ggml_tensor * ggml_transpose(
     result->nb[0] = a->nb[1];
     result->nb[1] = a->nb[0];
 
-    result->op   = GGML_OP_TRANSPOSE;
-    result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL;
+    result->op     = GGML_OP_TRANSPOSE;
     result->src[0] = a;
 
     return result;
@@ -4760,12 +6220,6 @@ struct ggml_tensor * ggml_get_rows(
     GGML_ASSERT(b->ne[3] == 1);
     GGML_ASSERT(b->type == GGML_TYPE_I32);
 
-    bool is_node = false;
-
-    if (a->grad || b->grad) {
-        is_node = true;
-    }
-
     // TODO: implement non F32 return
     enum ggml_type type = GGML_TYPE_F32;
     if (a->type == GGML_TYPE_I32) {
@@ -4773,8 +6227,7 @@ struct ggml_tensor * ggml_get_rows(
     }
     struct ggml_tensor * result = ggml_new_tensor_4d(ctx, type, a->ne[0], b->ne[0], b->ne[1], b->ne[2]);
 
-    result->op   = GGML_OP_GET_ROWS;
-    result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL;
+    result->op     = GGML_OP_GET_ROWS;
     result->src[0] = a;
     result->src[1] = b;
 
@@ -4791,18 +6244,11 @@ struct ggml_tensor * ggml_get_rows_back(
     GGML_ASSERT(ggml_is_matrix(a) && ggml_is_vector(b) && b->type == GGML_TYPE_I32);
     GGML_ASSERT(ggml_is_matrix(c) && (a->ne[0] == c->ne[0]));
 
-    bool is_node = false;
-
-    if (a->grad || b->grad) {
-        is_node = true;
-    }
-
     // TODO: implement non F32 return
     //struct ggml_tensor * result = ggml_new_tensor_2d(ctx, a->type, a->ne[0], b->ne[0]);
     struct ggml_tensor * result = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, c->ne[0], c->ne[1]);
 
-    result->op   = GGML_OP_GET_ROWS_BACK;
-    result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL;
+    result->op     = GGML_OP_GET_ROWS_BACK;
     result->src[0] = a;
     result->src[1] = b;
 
@@ -4815,17 +6261,11 @@ struct ggml_tensor * ggml_diag(
         struct ggml_context * ctx,
         struct ggml_tensor  * a) {
     GGML_ASSERT(a->ne[1] == 1);
-    bool is_node = false;
-
-    if (a->grad) {
-        is_node = true;
-    }
 
     const int64_t ne[4] = { a->ne[0], a->ne[0], a->ne[2], a->ne[3] };
     struct ggml_tensor * result = ggml_new_tensor(ctx, a->type, 4, ne);
 
-    result->op   = GGML_OP_DIAG;
-    result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL;
+    result->op     = GGML_OP_DIAG;
     result->src[0] = a;
 
     return result;
@@ -4838,19 +6278,12 @@ static struct ggml_tensor * ggml_diag_mask_inf_impl(
         struct ggml_tensor  * a,
         int                   n_past,
         bool                  inplace) {
-    bool is_node = false;
-
-    if (a->grad) {
-        is_node = true;
-    }
-
     struct ggml_tensor * result = inplace ? ggml_view_tensor(ctx, a) : ggml_dup_tensor(ctx, a);
 
     int32_t params[] = { n_past };
     ggml_set_op_params(result, params, sizeof(params));
 
-    result->op   = GGML_OP_DIAG_MASK_INF;
-    result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL;
+    result->op     = GGML_OP_DIAG_MASK_INF;
     result->src[0] = a;
 
     return result;
@@ -4877,19 +6310,12 @@ static struct ggml_tensor * ggml_diag_mask_zero_impl(
         struct ggml_tensor  * a,
         int                   n_past,
         bool                  inplace) {
-    bool is_node = false;
-
-    if (a->grad) {
-        is_node = true;
-    }
-
     struct ggml_tensor * result = inplace ? ggml_view_tensor(ctx, a) : ggml_dup_tensor(ctx, a);
 
     int32_t params[] = { n_past };
     ggml_set_op_params(result, params, sizeof(params));
 
-    result->op   = GGML_OP_DIAG_MASK_ZERO;
-    result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL;
+    result->op     = GGML_OP_DIAG_MASK_ZERO;
     result->src[0] = a;
 
     return result;
@@ -4916,28 +6342,28 @@ static struct ggml_tensor * ggml_soft_max_impl(
         struct ggml_tensor  * a,
         struct ggml_tensor  * mask,
         float                 scale,
+        float                 max_bias,
         bool                  inplace) {
     GGML_ASSERT(ggml_is_contiguous(a));
+
     if (mask) {
+        GGML_ASSERT(mask->type == GGML_TYPE_F16 || mask->type == GGML_TYPE_F32);
         GGML_ASSERT(ggml_is_contiguous(mask));
-        GGML_ASSERT(mask->ne[2] == 1);
-        GGML_ASSERT(mask->ne[3] == 1);
-        GGML_ASSERT(ggml_can_repeat_rows(mask, a));
+        GGML_ASSERT(ggml_is_matrix(mask));
+        GGML_ASSERT(mask->ne[0] == a->ne[0]);
+        GGML_ASSERT(mask->ne[1] >= a->ne[1]);
     }
 
-    bool is_node = false;
-
-    if (a->grad) {
-        is_node = true;
+    if (max_bias > 0.0f) {
+        GGML_ASSERT(mask);
     }
 
     struct ggml_tensor * result = inplace ? ggml_view_tensor(ctx, a) : ggml_dup_tensor(ctx, a);
 
-    float params[] = { scale };
+    float params[] = { scale, max_bias };
     ggml_set_op_params(result, params, sizeof(params));
 
-    result->op   = GGML_OP_SOFT_MAX;
-    result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL;
+    result->op     = GGML_OP_SOFT_MAX;
     result->src[0] = a;
     result->src[1] = mask;
 
@@ -4947,21 +6373,22 @@ static struct ggml_tensor * ggml_soft_max_impl(
 struct ggml_tensor * ggml_soft_max(
         struct ggml_context * ctx,
         struct ggml_tensor  * a) {
-    return ggml_soft_max_impl(ctx, a, NULL, 1.0f, false);
+    return ggml_soft_max_impl(ctx, a, NULL, 1.0f, 0.0f, false);
 }
 
 struct ggml_tensor * ggml_soft_max_inplace(
         struct ggml_context * ctx,
         struct ggml_tensor  * a) {
-    return ggml_soft_max_impl(ctx, a, NULL, 1.0f, true);
+    return ggml_soft_max_impl(ctx, a, NULL, 1.0f, 0.0f, true);
 }
 
 struct ggml_tensor * ggml_soft_max_ext(
         struct ggml_context * ctx,
         struct ggml_tensor  * a,
         struct ggml_tensor  * mask,
-        float                 scale) {
-    return ggml_soft_max_impl(ctx, a, mask, scale, false);
+        float                 scale,
+        float                 max_bias) {
+    return ggml_soft_max_impl(ctx, a, mask, scale, max_bias, false);
 }
 
 // ggml_soft_max_back
@@ -4971,16 +6398,9 @@ static struct ggml_tensor * ggml_soft_max_back_impl(
         struct ggml_tensor  * a,
         struct ggml_tensor  * b,
         bool                  inplace) {
-    bool is_node = false;
-
-    if (a->grad || b->grad) {
-        is_node = true; // TODO : implement backward pass
-    }
-
     struct ggml_tensor * result = inplace ? ggml_view_tensor(ctx, a) : ggml_dup_tensor(ctx, a);
 
-    result->op   = GGML_OP_SOFT_MAX_BACK;
-    result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL;
+    result->op     = GGML_OP_SOFT_MAX_BACK;
     result->src[0] = a;
     result->src[1] = b;
 
@@ -5007,46 +6427,43 @@ static struct ggml_tensor * ggml_rope_impl(
         struct ggml_context * ctx,
         struct ggml_tensor  * a,
         struct ggml_tensor  * b,
+        struct ggml_tensor  * c,
         int                   n_dims,
         int                   mode,
-        int                   n_ctx,
-        int                   n_orig_ctx,
+        int                   n_ctx_orig,
         float                 freq_base,
         float                 freq_scale,
         float                 ext_factor,
         float                 attn_factor,
         float                 beta_fast,
         float                 beta_slow,
-        float                 xpos_base,
-        bool                  xpos_down,
         bool                  inplace) {
+    GGML_ASSERT((mode & 1) == 0 && "mode & 1 == 1 is no longer supported");
+
     GGML_ASSERT(ggml_is_vector(b));
     GGML_ASSERT(b->type == GGML_TYPE_I32);
     GGML_ASSERT(a->ne[2] == b->ne[0]);
 
-    bool is_node = false;
-
-    if (a->grad) {
-        is_node = true;
+    if (c) {
+        GGML_ASSERT(c->type == GGML_TYPE_F32);
+        GGML_ASSERT(c->ne[0] >= n_dims / 2);
     }
 
     struct ggml_tensor * result = inplace ? ggml_view_tensor(ctx, a) : ggml_dup_tensor(ctx, a);
 
-    int32_t params[13] = { /*n_past*/ 0, n_dims, mode, n_ctx, n_orig_ctx };
+    int32_t params[11] = { /*n_past*/ 0, n_dims, mode, /*n_ctx*/ 0, n_ctx_orig };
     memcpy(params +  5, &freq_base,    sizeof(float));
     memcpy(params +  6, &freq_scale,   sizeof(float));
     memcpy(params +  7, &ext_factor,   sizeof(float));
     memcpy(params +  8, &attn_factor,  sizeof(float));
     memcpy(params +  9, &beta_fast,    sizeof(float));
     memcpy(params + 10, &beta_slow,    sizeof(float));
-    memcpy(params + 11, &xpos_base,    sizeof(float));
-    memcpy(params + 12, &xpos_down,    sizeof(bool));
     ggml_set_op_params(result, params, sizeof(params));
 
-    result->op   = GGML_OP_ROPE;
-    result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL;
+    result->op     = GGML_OP_ROPE;
     result->src[0] = a;
     result->src[1] = b;
+    result->src[2] = c;
 
     return result;
 }
@@ -5056,10 +6473,9 @@ struct ggml_tensor * ggml_rope(
         struct ggml_tensor  * a,
         struct ggml_tensor  * b,
         int                   n_dims,
-        int                   mode,
-        int                   n_ctx) {
+        int                   mode) {
     return ggml_rope_impl(
-        ctx, a, b, n_dims, mode, n_ctx, 0, 10000.0f, 1.0f, 0.0f, 1.0f, 0.0f, 0.0f, 0.0f, false, false
+        ctx, a, b, NULL, n_dims, mode, 0, 10000.0f, 1.0f, 0.0f, 1.0f, 0.0f, 0.0f, false
     );
 }
 
@@ -5068,21 +6484,40 @@ struct ggml_tensor * ggml_rope_inplace(
         struct ggml_tensor  * a,
         struct ggml_tensor  * b,
         int                   n_dims,
+        int                   mode) {
+    return ggml_rope_impl(
+        ctx, a, b, NULL, n_dims, mode, 0, 10000.0f, 1.0f, 0.0f, 1.0f, 0.0f, 0.0f, true
+    );
+}
+
+struct ggml_tensor * ggml_rope_ext(
+        struct ggml_context * ctx,
+        struct ggml_tensor  * a,
+        struct ggml_tensor  * b,
+        struct ggml_tensor  * c,
+        int                   n_dims,
         int                   mode,
-        int                   n_ctx) {
+        int                   n_ctx_orig,
+        float                 freq_base,
+        float                 freq_scale,
+        float                 ext_factor,
+        float                 attn_factor,
+        float                 beta_fast,
+        float                 beta_slow) {
     return ggml_rope_impl(
-        ctx, a, b, n_dims, mode, n_ctx, 0, 10000.0f, 1.0f, 0.0f, 1.0f, 0.0f, 0.0f, 0.0f, false, true
+        ctx, a, b, c, n_dims, mode, n_ctx_orig, freq_base, freq_scale,
+        ext_factor, attn_factor, beta_fast, beta_slow, false
     );
 }
 
-struct ggml_tensor * ggml_rope_custom(
+struct ggml_tensor * ggml_rope_ext_inplace(
         struct ggml_context * ctx,
         struct ggml_tensor  * a,
         struct ggml_tensor  * b,
+        struct ggml_tensor  * c,
         int                   n_dims,
         int                   mode,
-        int                   n_ctx,
-        int                   n_orig_ctx,
+        int                   n_ctx_orig,
         float                 freq_base,
         float                 freq_scale,
         float                 ext_factor,
@@ -5090,19 +6525,18 @@ struct ggml_tensor * ggml_rope_custom(
         float                 beta_fast,
         float                 beta_slow) {
     return ggml_rope_impl(
-        ctx, a, b, n_dims, mode, n_ctx, n_orig_ctx, freq_base, freq_scale,
-        ext_factor, attn_factor, beta_fast, beta_slow, 0.0f, false, false
+        ctx, a, b, c, n_dims, mode, n_ctx_orig, freq_base, freq_scale,
+        ext_factor, attn_factor, beta_fast, beta_slow, true
     );
 }
 
-struct ggml_tensor * ggml_rope_custom_inplace(
+struct ggml_tensor * ggml_rope_custom(
         struct ggml_context * ctx,
         struct ggml_tensor  * a,
         struct ggml_tensor  * b,
         int                   n_dims,
         int                   mode,
-        int                   n_ctx,
-        int                   n_orig_ctx,
+        int                   n_ctx_orig,
         float                 freq_base,
         float                 freq_scale,
         float                 ext_factor,
@@ -5110,19 +6544,28 @@ struct ggml_tensor * ggml_rope_custom_inplace(
         float                 beta_fast,
         float                 beta_slow) {
     return ggml_rope_impl(
-        ctx, a, b, n_dims, mode, n_ctx, n_orig_ctx, freq_base, freq_scale,
-        ext_factor, attn_factor, beta_fast, beta_slow, 0.0f, false, true
+        ctx, a, b, NULL, n_dims, mode, n_ctx_orig, freq_base, freq_scale,
+        ext_factor, attn_factor, beta_fast, beta_slow, false
     );
 }
 
-struct ggml_tensor * ggml_rope_xpos_inplace(
+struct ggml_tensor * ggml_rope_custom_inplace(
         struct ggml_context * ctx,
         struct ggml_tensor  * a,
         struct ggml_tensor  * b,
         int                   n_dims,
-        float                 base,
-        bool                  down) {
-    return ggml_rope_impl(ctx, a, b, n_dims, 0, 0, 0, 10000.0f, 1.0f, 0.0f, 1.0f, 0.0f, 0.0f, base, down, true);
+        int                   mode,
+        int                   n_ctx_orig,
+        float                 freq_base,
+        float                 freq_scale,
+        float                 ext_factor,
+        float                 attn_factor,
+        float                 beta_fast,
+        float                 beta_slow) {
+    return ggml_rope_impl(
+        ctx, a, b, NULL, n_dims, mode, n_ctx_orig, freq_base, freq_scale,
+        ext_factor, attn_factor, beta_fast, beta_slow, true
+    );
 }
 
 // ggml_rope_back
@@ -5131,78 +6574,35 @@ struct ggml_tensor * ggml_rope_back(
         struct ggml_context * ctx,
         struct ggml_tensor  * a,
         struct ggml_tensor  * b,
+        struct ggml_tensor  * c,
         int                   n_dims,
         int                   mode,
-        int                   n_ctx,
-        int                   n_orig_ctx,
+        int                   n_ctx_orig,
         float                 freq_base,
         float                 freq_scale,
         float                 ext_factor,
         float                 attn_factor,
         float                 beta_fast,
-        float                 beta_slow,
-        float                 xpos_base,
-        bool                  xpos_down) {
+        float                 beta_slow) {
     GGML_ASSERT(ggml_is_vector(b));
     GGML_ASSERT(b->type == GGML_TYPE_I32);
     GGML_ASSERT(a->ne[2] == b->ne[0]);
 
-    GGML_ASSERT((mode & 4) == 0 && "ggml_rope_back() for ChatGLM not implemented yet");
-
-    bool is_node = false;
-
-    if (a->grad) {
-        is_node = false; // TODO: implement backward
-    }
-
     struct ggml_tensor * result = ggml_dup_tensor(ctx, a);
 
-    int32_t params[13] = { /*n_past*/ 0, n_dims, mode, n_ctx, n_orig_ctx };
+    int32_t params[11] = { /*n_past*/ 0, n_dims, mode, /*n_ctx*/ 0, n_ctx_orig };
     memcpy(params +  5, &freq_base,    sizeof(float));
     memcpy(params +  6, &freq_scale,   sizeof(float));
     memcpy(params +  7, &ext_factor,   sizeof(float));
     memcpy(params +  8, &attn_factor,  sizeof(float));
     memcpy(params +  9, &beta_fast,    sizeof(float));
     memcpy(params + 10, &beta_slow,    sizeof(float));
-    memcpy(params + 11, &xpos_base,    sizeof(float));
-    memcpy(params + 12, &xpos_down,    sizeof(bool));
     ggml_set_op_params(result, params, sizeof(params));
 
-    result->op   = GGML_OP_ROPE_BACK;
-    result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL;
+    result->op     = GGML_OP_ROPE_BACK;
     result->src[0] = a;
     result->src[1] = b;
-
-    return result;
-}
-
-// ggml_alibi
-
-struct ggml_tensor * ggml_alibi(
-        struct ggml_context * ctx,
-        struct ggml_tensor  * a,
-        int                   n_past,
-        int                   n_head,
-        float                 bias_max) {
-    GGML_ASSERT(n_past >= 0);
-    bool is_node = false;
-
-    if (a->grad) {
-        GGML_ASSERT(false); // TODO: implement backward
-        is_node = true;
-    }
-
-    // TODO: when implement backward, fix this:
-    //struct ggml_tensor * result = inplace ? ggml_view_tensor(ctx, a) : ggml_dup_tensor(ctx, a);
-    struct ggml_tensor * result = ggml_view_tensor(ctx, a);
-
-    int32_t op_params[3] = { n_past, n_head };
-    memcpy(op_params + 2, &bias_max, sizeof(float));
-    ggml_set_op_params(result, op_params, sizeof(op_params));
-
-    result->op   = GGML_OP_ALIBI;
-    result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL;
-    result->src[0] = a;
+    result->src[2] = c;
 
     return result;
 }
@@ -5214,21 +6614,13 @@ struct ggml_tensor * ggml_clamp(
         struct ggml_tensor  * a,
         float                 min,
         float                 max) {
-    bool is_node = false;
-
-    if (a->grad) {
-        GGML_ASSERT(false); // TODO: implement backward
-        is_node = true;
-    }
-
     // TODO: when implement backward, fix this:
     struct ggml_tensor * result = ggml_view_tensor(ctx, a);
 
     float params[] = { min, max };
     ggml_set_op_params(result, params, sizeof(params));
 
-    result->op   = GGML_OP_CLAMP;
-    result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL;
+    result->op     = GGML_OP_CLAMP;
     result->src[0] = a;
 
     return result;
@@ -5247,7 +6639,7 @@ GGML_API struct ggml_tensor * ggml_conv_1d(
         int                   s0,
         int                   p0,
         int                   d0) {
-    struct ggml_tensor * im2col = ggml_im2col(ctx, a, b, s0, 0, p0, 0, d0, 0, false); // [N, OL, IC * K]
+    struct ggml_tensor * im2col = ggml_im2col(ctx, a, b, s0, 0, p0, 0, d0, 0, false, GGML_TYPE_F16); // [N, OL, IC * K]
 
     struct ggml_tensor * result =
         ggml_mul_mat(ctx,
@@ -5290,13 +6682,6 @@ GGML_API struct ggml_tensor * ggml_conv_transpose_1d(
     GGML_ASSERT(p0 == 0);
     GGML_ASSERT(d0 == 1);
 
-    bool is_node = false;
-
-    if (a->grad || b->grad) {
-        GGML_ASSERT(false); // TODO: implement backward
-        is_node = true;
-    }
-
     const int64_t ne[4] = {
         ggml_calc_conv_transpose_1d_output_size(b->ne[0], a->ne[0], s0, 0 /*p0*/, 1 /*d0*/),
         a->ne[1], b->ne[2], 1,
@@ -5306,14 +6691,37 @@ GGML_API struct ggml_tensor * ggml_conv_transpose_1d(
     int32_t params[] = { s0, p0, d0 };
     ggml_set_op_params(result, params, sizeof(params));
 
-    result->op = GGML_OP_CONV_TRANSPOSE_1D;
-    result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL;
+    result->op     = GGML_OP_CONV_TRANSPOSE_1D;
     result->src[0] = a;
     result->src[1] = b;
 
     return result;
 }
 
+// ggml_conv_depthwise
+
+struct ggml_tensor * ggml_conv_depthwise_2d(
+        struct ggml_context * ctx,
+        struct ggml_tensor  * a,
+        struct ggml_tensor  * b,
+        int                   s0,
+        int                   s1,
+        int                   p0,
+        int                   p1,
+        int                   d0,
+        int                   d1) {
+    struct ggml_tensor * new_a = ggml_reshape_4d(ctx, a, a->ne[0], a->ne[1], 1, a->ne[2] * a->ne[3]);
+    struct ggml_tensor * im2col = ggml_im2col(ctx, new_a,
+                                        ggml_reshape_4d(ctx, b, b->ne[0], b->ne[1], 1, b->ne[2] * b->ne[3]),
+                                        s0, s1, p0, p1, d0, d1, true, GGML_TYPE_F16); // [N * IC, OH, OW, KH * KW]
+    struct ggml_tensor * new_b = ggml_reshape_4d(ctx, im2col, im2col->ne[0], im2col->ne[2] * im2col->ne[1], b->ne[2], b->ne[3]); // [N * IC, OH, OW, KH * KW] => [N, IC, OH * OW, KH * KW]
+
+    new_a = ggml_reshape_4d(ctx, new_a, (new_a->ne[0] * new_a->ne[1]), new_a->ne[2],  new_a->ne[3], 1);                       // [OC，1, KH, KW] => [1, OC, 1, KH * KW]
+    struct ggml_tensor * result = ggml_mul_mat(ctx, new_a, new_b);
+    result = ggml_reshape_4d(ctx, result, im2col->ne[1], im2col->ne[2], b->ne[2], b->ne[3]); // [N, OC, OH, OW]
+
+    return result;
+}
 // ggml_conv_2d
 
 // im2col: [N, IC, IH, IW] => [N, OH, OW, IC*KH*KW]
@@ -5321,32 +6729,30 @@ GGML_API struct ggml_tensor * ggml_conv_transpose_1d(
 // b: [N, IC, IH, IW]
 // result: [N, OH, OW, IC*KH*KW]
 struct ggml_tensor * ggml_im2col(
-    struct ggml_context * ctx,
-    struct ggml_tensor  * a,
-    struct ggml_tensor  * b,
-    int                  s0,
-    int                  s1,
-    int                  p0,
-    int                  p1,
-    int                  d0,
-    int                  d1,
-    bool                 is_2D) {
-
+        struct ggml_context * ctx,
+        struct ggml_tensor  * a,
+        struct ggml_tensor  * b,
+        int                   s0,
+        int                   s1,
+        int                   p0,
+        int                   p1,
+        int                   d0,
+        int                   d1,
+        bool                  is_2D,
+        enum ggml_type        dst_type) {
     if(is_2D) {
         GGML_ASSERT(a->ne[2] == b->ne[2]);
     } else {
         GGML_ASSERT(a->ne[1] == b->ne[1]);
-    }
-    bool is_node = false;
-
-    if (a->grad || b->grad) {
-        GGML_ASSERT(false); // TODO: implement backward
-        is_node = true;
+        GGML_ASSERT(b->ne[3] == 1);
     }
 
     const int64_t OH = is_2D ? ggml_calc_conv_output_size(b->ne[1], a->ne[1], s1, p1, d1) : 0;
     const int64_t OW =         ggml_calc_conv_output_size(b->ne[0], a->ne[0], s0, p0, d0);
 
+    GGML_ASSERT((!is_2D || OH > 0) && "b too small compared to a");
+    GGML_ASSERT((OW > 0)           && "b too small compared to a");
+
     const int64_t ne[4] = {
         is_2D ? (a->ne[2] * a->ne[1] * a->ne[0]) : a->ne[1] * a->ne[0],
         OW,
@@ -5354,12 +6760,34 @@ struct ggml_tensor * ggml_im2col(
         is_2D ?      b->ne[3] : 1,
     };
 
-    struct ggml_tensor * result = ggml_new_tensor(ctx, GGML_TYPE_F16, 4, ne);
+    struct ggml_tensor * result = ggml_new_tensor(ctx, dst_type, 4, ne);
     int32_t params[] = { s0, s1, p0, p1, d0, d1, (is_2D ? 1 : 0) };
     ggml_set_op_params(result, params, sizeof(params));
 
-    result->op = GGML_OP_IM2COL;
-    result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL;
+    result->op     = GGML_OP_IM2COL;
+    result->src[0] = a;
+    result->src[1] = b;
+
+    return result;
+}
+
+struct ggml_tensor * ggml_im2col_back(
+        struct ggml_context * ctx,
+        struct ggml_tensor  * a,
+        struct ggml_tensor  * b,
+        int64_t             * ne,
+        int                   s0,
+        int                   s1,
+        int                   p0,
+        int                   p1,
+        int                   d0,
+        int                   d1,
+        bool                  is_2D) {
+    struct ggml_tensor * result = ggml_new_tensor(ctx, GGML_TYPE_F32, 4, ne);
+    int32_t params[] = { s0, s1, p0, p1, d0, d1, (is_2D ? 1 : 0) };
+    ggml_set_op_params(result, params, sizeof(params));
+
+    result->op     = GGML_OP_IM2COL_BACK;
     result->src[0] = a;
     result->src[1] = b;
 
@@ -5373,25 +6801,28 @@ struct ggml_tensor * ggml_conv_2d(
         struct ggml_context * ctx,
         struct ggml_tensor  * a,
         struct ggml_tensor  * b,
-        int                  s0,
-        int                  s1,
-        int                  p0,
-        int                  p1,
-        int                  d0,
-        int                  d1) {
-    struct ggml_tensor * im2col = ggml_im2col(ctx, a, b, s0, s1, p0, p1, d0, d1, true); // [N, OH, OW, IC * KH * KW]
+        int                   s0,
+        int                   s1,
+        int                   p0,
+        int                   p1,
+        int                   d0,
+        int                   d1) {
+    struct ggml_tensor * im2col = ggml_im2col(ctx, a, b, s0, s1, p0, p1, d0, d1, true, a->type); // [N, OH, OW, IC * KH * KW]
 
     struct ggml_tensor * result =
         ggml_mul_mat(ctx,
                 ggml_reshape_2d(ctx, im2col, im2col->ne[0],  im2col->ne[3] * im2col->ne[2] * im2col->ne[1]), // [N, OH, OW, IC * KH * KW] => [N*OH*OW, IC * KH * KW]
                 ggml_reshape_2d(ctx, a, (a->ne[0] * a->ne[1] * a->ne[2]),  a->ne[3]));                       // [OC，IC, KH, KW] => [OC, IC * KH * KW]
 
-    result = ggml_reshape_4d(ctx, result, im2col->ne[1], im2col->ne[2], a->ne[3], im2col->ne[3]); // [N, OC, OH, OW]
+    result = ggml_reshape_4d(ctx, result, im2col->ne[1], im2col->ne[2], im2col->ne[3], a->ne[3]); // [OC, N, OH, OW]
+    result = ggml_cont(ctx, ggml_permute(ctx, result, 0, 1, 3, 2)); // [N, OC, OH, OW]
+
 
     return result;
 }
 
 // ggml_conv_2d_sk_p0
+
 struct ggml_tensor * ggml_conv_2d_sk_p0(
         struct ggml_context * ctx,
         struct ggml_tensor  * a,
@@ -5421,13 +6852,6 @@ struct ggml_tensor * ggml_conv_transpose_2d_p0(
         int                   stride) {
     GGML_ASSERT(a->ne[3] == b->ne[2]);
 
-    bool is_node = false;
-
-    if (a->grad || b->grad) {
-        GGML_ASSERT(false); // TODO: implement backward
-        is_node = true;
-    }
-
     const int64_t ne[4] = {
         ggml_calc_conv_transpose_output_size(b->ne[0], a->ne[0], stride, 0 /*p0*/),
         ggml_calc_conv_transpose_output_size(b->ne[1], a->ne[1], stride, 0 /*p1*/),
@@ -5438,8 +6862,7 @@ struct ggml_tensor * ggml_conv_transpose_2d_p0(
 
     ggml_set_op_params_i32(result, 0, stride);
 
-    result->op = GGML_OP_CONV_TRANSPOSE_2D;
-    result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL;
+    result->op     = GGML_OP_CONV_TRANSPOSE_2D;
     result->src[0] = a;
     result->src[1] = b;
 
@@ -5461,25 +6884,18 @@ struct ggml_tensor * ggml_pool_1d(
         int                   k0,
         int                   s0,
         int                   p0) {
-
-    bool is_node = false;
-
-    if (a->grad) {
-        GGML_ASSERT(false); // TODO: implement backward
-        is_node = true;
-    }
-
-    const int64_t ne[2] = {
+    const int64_t ne[4] = {
         ggml_calc_pool_output_size(a->ne[0], k0, s0, p0),
         a->ne[1],
+        a->ne[2],
+        a->ne[3],
     };
-    struct ggml_tensor * result = ggml_new_tensor(ctx, GGML_TYPE_F32, 2, ne);
+    struct ggml_tensor * result = ggml_new_tensor(ctx, GGML_TYPE_F32, 4, ne);
 
     int32_t params[] = { op, k0, s0, p0 };
     ggml_set_op_params(result, params, sizeof(params));
 
-    result->op = GGML_OP_POOL_1D;
-    result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL;
+    result->op     = GGML_OP_POOL_1D;
     result->src[0] = a;
 
     return result;
@@ -5497,27 +6913,44 @@ struct ggml_tensor * ggml_pool_2d(
         int                   s1,
         float                 p0,
         float                 p1) {
-
-    bool is_node = false;
-
-    if (a->grad) {
-        GGML_ASSERT(false); // TODO: implement backward
-        is_node = true;
-    }
-
-    const int64_t ne[3] = {
+    struct ggml_tensor * result;
+    const int64_t ne[4] = {
         ggml_calc_pool_output_size(a->ne[0], k0, s0, p0),
         ggml_calc_pool_output_size(a->ne[1], k1, s1, p1),
         a->ne[2],
+        a->ne[3],
     };
-    struct ggml_tensor * result = ggml_new_tensor(ctx, GGML_TYPE_F32, 3, ne);
+    result = ggml_new_tensor(ctx, GGML_TYPE_F32, 4, ne);
 
     int32_t params[] = { op, k0, k1, s0, s1, p0, p1 };
     ggml_set_op_params(result, params, sizeof(params));
 
-    result->op = GGML_OP_POOL_2D;
-    result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL;
+    result->op     = GGML_OP_POOL_2D;
+    result->src[0] = a;
+
+    return result;
+}
+
+struct ggml_tensor * ggml_pool_2d_back(
+        struct ggml_context * ctx,
+        struct ggml_tensor  * a,
+        struct ggml_tensor  * af,
+        enum ggml_op_pool     op,
+        int                   k0,
+        int                   k1,
+        int                   s0,
+        int                   s1,
+        float                 p0,
+        float                 p1) {
+    struct ggml_tensor * result;
+    result = ggml_new_tensor(ctx, GGML_TYPE_F32, 4, af->ne);
+
+    int32_t params[] = { op, k0, k1, s0, s1, p0, p1 };
+    ggml_set_op_params(result, params, sizeof(params));
+
+    result->op     = GGML_OP_POOL_2D_BACK;
     result->src[0] = a;
+    result->src[1] = af;
 
     return result;
 }
@@ -5525,74 +6958,119 @@ struct ggml_tensor * ggml_pool_2d(
 // ggml_upscale
 
 static struct ggml_tensor * ggml_upscale_impl(
-    struct ggml_context * ctx,
-    struct ggml_tensor * a,
-    int scale_factor) {
-    bool is_node = false;
-
-    if (a->grad) {
-        GGML_ASSERT(false); // TODO: implement backward
-        is_node = true;
-    }
+        struct ggml_context * ctx,
+        struct ggml_tensor  * a,
+        int                   ne0,
+        int                   ne1,
+        int                   ne2,
+        int                   ne3) {
+    GGML_ASSERT(a->ne[0] <= ne0);
+    GGML_ASSERT(a->ne[1] <= ne1);
+    GGML_ASSERT(a->ne[2] <= ne2);
+    GGML_ASSERT(a->ne[3] <= ne3);
 
-    struct ggml_tensor * result = ggml_new_tensor_4d(ctx, a->type,
-            a->ne[0] * scale_factor,
-            a->ne[1] * scale_factor,
-            a->ne[2], a->ne[3]);
+    struct ggml_tensor * result = ggml_new_tensor_4d(ctx, a->type, ne0, ne1, ne2, ne3);
 
-    result->op = GGML_OP_UPSCALE;
-    result->op_params[0] = scale_factor;
-    result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL;
+    result->op     = GGML_OP_UPSCALE;
     result->src[0] = a;
 
     return result;
 }
 
-struct ggml_tensor * ggml_pad(
-    struct ggml_context * ctx,
-    struct ggml_tensor  * a,
-    int p0, int p1, int p2, int p3) {
-    bool is_node = false;
+struct ggml_tensor * ggml_upscale(
+        struct ggml_context * ctx,
+        struct ggml_tensor  * a,
+        int                   scale_factor) {
+    return ggml_upscale_impl(ctx, a, a->ne[0] * scale_factor, a->ne[1] * scale_factor, a->ne[2], a->ne[3]);
+}
 
-    if (a->grad) {
-        GGML_ASSERT(false); // TODO: implement backward
-        is_node = true;
-    }
+struct ggml_tensor * ggml_upscale_ext(
+        struct ggml_context * ctx,
+        struct ggml_tensor  * a,
+        int                   ne0,
+        int                   ne1,
+        int                   ne2,
+        int                   ne3) {
+    return ggml_upscale_impl(ctx, a, ne0, ne1, ne2, ne3);
+}
 
+// ggml_pad
+
+struct ggml_tensor * ggml_pad(
+        struct ggml_context * ctx,
+        struct ggml_tensor  * a,
+        int                   p0,
+        int                   p1,
+        int                   p2,
+        int                   p3) {
     struct ggml_tensor * result = ggml_new_tensor_4d(ctx, a->type,
             a->ne[0] + p0,
             a->ne[1] + p1,
             a->ne[2] + p2,
             a->ne[3] + p3);
 
-    result->op = GGML_OP_PAD;
-    result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL;
+    result->op     = GGML_OP_PAD;
     result->src[0] = a;
 
     return result;
 }
 
-struct ggml_tensor * ggml_upscale(
-    struct ggml_context * ctx,
-    struct ggml_tensor * a,
-    int scale_factor) {
-    return ggml_upscale_impl(ctx, a, scale_factor);
+// ggml_arange
+
+struct ggml_tensor * ggml_arange(
+        struct ggml_context * ctx,
+        float                 start,
+        float                 stop,
+        float                 step) {
+    GGML_ASSERT(stop > start);
+
+    const int64_t steps = (int64_t) ceilf((stop - start) / step);
+
+    struct ggml_tensor * result = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, steps);
+
+    ggml_set_op_params_f32(result, 0, start);
+    ggml_set_op_params_f32(result, 1, stop);
+    ggml_set_op_params_f32(result, 2, step);
+
+    result->op = GGML_OP_ARANGE;
+
+    return result;
 }
 
-// ggml_argsort
+// ggml_timestep_embedding
 
-struct ggml_tensor * ggml_argsort(
+struct ggml_tensor * ggml_timestep_embedding(
         struct ggml_context * ctx,
-        struct ggml_tensor  * a,
-        enum ggml_sort_order  order) {
-    bool is_node = false;
+        struct ggml_tensor  * timesteps,
+        int                   dim,
+        int                   max_period) {
+    int actual_dim = dim;
+    if (dim % 2 != 0) {
+        actual_dim = dim + 1;
+    }
+
+    struct ggml_tensor * result = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, actual_dim, timesteps->ne[0]);
 
+    ggml_set_op_params_i32(result, 0, dim);
+    ggml_set_op_params_i32(result, 1, max_period);
+
+    result->op     = GGML_OP_TIMESTEP_EMBEDDING;
+    result->src[0] = timesteps;
+
+    return result;
+}
+
+// ggml_argsort
+
+struct ggml_tensor * ggml_argsort(
+        struct ggml_context  * ctx,
+        struct ggml_tensor   * a,
+        enum ggml_sort_order   order) {
     struct ggml_tensor * result = ggml_new_tensor(ctx, GGML_TYPE_I32, GGML_MAX_DIMS, a->ne);
 
     ggml_set_op_params_i32(result, 0, (int32_t) order);
 
-    result->op   = GGML_OP_ARGSORT;
-    result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL;
+    result->op     = GGML_OP_ARGSORT;
     result->src[0] = a;
 
     return result;
@@ -5606,7 +7084,7 @@ struct ggml_tensor * ggml_top_k(
         int                   k) {
     GGML_ASSERT(a->ne[0] >= k);
 
-    struct ggml_tensor * result = ggml_argsort(ctx, a, GGML_SORT_DESC);
+    struct ggml_tensor * result = ggml_argsort(ctx, a, GGML_SORT_ORDER_DESC);
 
     result = ggml_view_4d(ctx, result,
                 k, result->ne[1], result->ne[2], result->ne[3],
@@ -5616,68 +7094,60 @@ struct ggml_tensor * ggml_top_k(
     return result;
 }
 
-// ggml_flash_attn
+// ggml_flash_attn_ext
 
-struct ggml_tensor * ggml_flash_attn(
+struct ggml_tensor * ggml_flash_attn_ext(
         struct ggml_context * ctx,
         struct ggml_tensor  * q,
         struct ggml_tensor  * k,
         struct ggml_tensor  * v,
-        bool                  masked) {
+        struct ggml_tensor  * mask,
+        float                 scale,
+        float                 max_bias,
+        float                 logit_softcap) {
     GGML_ASSERT(ggml_can_mul_mat(k, q));
     // TODO: check if vT can be multiplied by (k*qT)
 
-    bool is_node = false;
+    if (mask) {
+        GGML_ASSERT(ggml_is_contiguous(mask));
+        GGML_ASSERT(mask->ne[2] == 1);
+        GGML_ASSERT(mask->ne[3] == 1);
+        GGML_ASSERT(mask->ne[1] >= GGML_PAD(q->ne[1], GGML_KQ_MASK_PAD) &&
+                "the Flash-Attention kernel requires the mask to be padded to GGML_KQ_MASK_PAD and at least n_queries big");
+        //GGML_ASSERT(ggml_can_repeat_rows(mask, qk));
+    }
 
-    if (q->grad || k->grad || v->grad) {
-        is_node = true;
+    if (max_bias > 0.0f) {
+        GGML_ASSERT(mask);
     }
 
-    //struct ggml_tensor * result = ggml_dup_tensor(ctx, q);
-    struct ggml_tensor * result = ggml_new_tensor(ctx, GGML_TYPE_F32, GGML_MAX_DIMS, q->ne);
+    bool is_node = false;
+
+    // permute(0, 2, 1, 3)
+    int64_t ne[4] = { q->ne[0], q->ne[2], q->ne[1], q->ne[3] };
+    struct ggml_tensor * result = ggml_new_tensor(ctx, GGML_TYPE_F32, 4, ne);
 
-    int32_t t = masked ? 1 : 0;
-    ggml_set_op_params(result, &t, sizeof(t));
+    float params[] = { scale, max_bias, logit_softcap };
+    ggml_set_op_params(result, params, sizeof(params));
 
-    result->op   = GGML_OP_FLASH_ATTN;
+    result->op   = GGML_OP_FLASH_ATTN_EXT;
     result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL;
     result->src[0] = q;
     result->src[1] = k;
     result->src[2] = v;
+    result->src[3] = mask;
 
     return result;
 }
 
-// ggml_flash_ff
-
-struct ggml_tensor * ggml_flash_ff(
-        struct ggml_context * ctx,
-        struct ggml_tensor  * a,
-        struct ggml_tensor  * b0,
-        struct ggml_tensor  * b1,
-        struct ggml_tensor  * c0,
-        struct ggml_tensor  * c1) {
-    GGML_ASSERT(ggml_can_mul_mat(b0, a));
-    // TODO: more checks
-
-    bool is_node = false;
-
-    if (a->grad || b0->grad || b1->grad || c0->grad || c1->grad) {
-        is_node = true;
-    }
-
-    //struct ggml_tensor * result = ggml_dup_tensor(ctx, a);
-    struct ggml_tensor * result = ggml_new_tensor(ctx, GGML_TYPE_F32, GGML_MAX_DIMS, a->ne);
+void ggml_flash_attn_ext_set_prec(
+        struct ggml_tensor * a,
+        enum ggml_prec       prec) {
+    GGML_ASSERT(a->op == GGML_OP_FLASH_ATTN_EXT);
 
-    result->op   = GGML_OP_FLASH_FF;
-    result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL;
-    result->src[0] = a;
-    result->src[1] = b0;
-    result->src[2] = b1;
-    result->src[3] = c0;
-    result->src[4] = c1;
+    const int32_t prec_i32 = (int32_t) prec;
 
-    return result;
+    ggml_set_op_params_i32(a, 3, prec_i32); // scale is on first pos, max_bias on second
 }
 
 // ggml_flash_attn_back
@@ -5689,6 +7159,8 @@ struct ggml_tensor * ggml_flash_attn_back(
         struct ggml_tensor  * v,
         struct ggml_tensor  * d,
         bool                  masked) {
+    GGML_ABORT("TODO: adapt to ggml_flash_attn_ext() changes");
+
     GGML_ASSERT(ggml_can_mul_mat(k, q));
     // TODO: check if vT can be multiplied by (k*qT)
 
@@ -5758,38 +7230,109 @@ struct ggml_tensor * ggml_flash_attn_back(
     return result;
 }
 
-// ggml_win_part
+// ggml_ssm_conv
 
-struct ggml_tensor * ggml_win_part(
+struct ggml_tensor * ggml_ssm_conv(
         struct ggml_context * ctx,
-        struct ggml_tensor  * a,
-        int                   w) {
-    GGML_ASSERT(a->ne[3] == 1);
-    GGML_ASSERT(a->type  == GGML_TYPE_F32);
-
-    bool is_node = false;
+        struct ggml_tensor  * sx,
+        struct ggml_tensor  * c) {
+    GGML_ASSERT(ggml_is_3d(sx));
+    GGML_ASSERT(ggml_is_matrix(c));
 
-    if (a->grad) {
-        GGML_ASSERT(false); // TODO: implement backward
-        is_node = true;
-    }
+    const int64_t d_conv  = c->ne[0];
+    const int64_t d_inner = c->ne[1];
+    const int64_t n_t     = sx->ne[0] - d_conv + 1; // tokens per sequence
+    const int64_t n_s     = sx->ne[2];
 
-    // padding
-    const int px = (w - a->ne[1]%w)%w;
-    const int py = (w - a->ne[2]%w)%w;
+    // TODO: maybe support other strides than 1?
+    GGML_ASSERT(sx->ne[0] == d_conv - 1 + n_t);
+    GGML_ASSERT(sx->ne[1] == d_inner);
+    GGML_ASSERT(n_t >= 0);
 
-    const int npx = (px + a->ne[1])/w;
-    const int npy = (py + a->ne[2])/w;
-    const int np  = npx*npy;
+    struct ggml_tensor * result = ggml_new_tensor_3d(ctx, GGML_TYPE_F32, d_inner, n_t, n_s);
 
-    const int64_t ne[4] = { a->ne[0], w, w, np, };
-    struct ggml_tensor * result = ggml_new_tensor(ctx, GGML_TYPE_F32, 4, ne);
+    result->op     = GGML_OP_SSM_CONV;
+    result->src[0] = sx;
+    result->src[1] = c;
+
+    return result;
+}
+
+// ggml_ssm_scan
+
+struct ggml_tensor * ggml_ssm_scan(
+        struct ggml_context * ctx,
+        struct ggml_tensor  * s,
+        struct ggml_tensor  * x,
+        struct ggml_tensor  * dt,
+        struct ggml_tensor  * A,
+        struct ggml_tensor  * B,
+        struct ggml_tensor  * C) {
+    GGML_ASSERT(ggml_is_contiguous(s));
+    GGML_ASSERT(ggml_is_contiguous(x));
+    GGML_ASSERT(ggml_is_contiguous(dt));
+    GGML_ASSERT(ggml_is_contiguous(A));
+    GGML_ASSERT(ggml_is_matrix(A));
+    GGML_ASSERT(ggml_is_3d(B));
+    GGML_ASSERT(ggml_is_3d(s));
+    GGML_ASSERT(B->nb[0] == ggml_type_size(B->type));
+    GGML_ASSERT(C->nb[0] == ggml_type_size(C->type));
+    GGML_ASSERT(ggml_are_same_shape(x, dt));
+    GGML_ASSERT(ggml_are_same_shape(B, C));
+
+    {
+        const int64_t d_state      = s->ne[0];
+        const int64_t d_inner      = s->ne[1];
+        const int64_t n_seq_tokens = x->ne[1];
+        const int64_t n_seqs       = x->ne[2];
+
+        GGML_ASSERT(s->ne[2] == n_seqs);
+        GGML_ASSERT(x->ne[0] == d_inner);
+        GGML_ASSERT(A->ne[0] == d_state);
+        GGML_ASSERT(A->ne[1] == d_inner);
+        GGML_ASSERT(B->ne[0] == d_state);
+        GGML_ASSERT(B->ne[1] == n_seq_tokens);
+        GGML_ASSERT(B->ne[2] == n_seqs);
+    }
+
+    // concatenated y + ssm_states
+    struct ggml_tensor * result = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, ggml_nelements(x) + ggml_nelements(s));
+
+    result->op   = GGML_OP_SSM_SCAN;
+    result->src[0] = s;
+    result->src[1] = x;
+    result->src[2] = dt;
+    result->src[3] = A;
+    result->src[4] = B;
+    result->src[5] = C;
+
+    return result;
+}
+
+// ggml_win_part
+
+struct ggml_tensor * ggml_win_part(
+        struct ggml_context * ctx,
+        struct ggml_tensor  * a,
+        int                   w) {
+    GGML_ASSERT(a->ne[3] == 1);
+    GGML_ASSERT(a->type  == GGML_TYPE_F32);
+
+    // padding
+    const int px = (w - a->ne[1]%w)%w;
+    const int py = (w - a->ne[2]%w)%w;
+
+    const int npx = (px + a->ne[1])/w;
+    const int npy = (py + a->ne[2])/w;
+    const int np  = npx*npy;
+
+    const int64_t ne[4] = { a->ne[0], w, w, np, };
+    struct ggml_tensor * result = ggml_new_tensor(ctx, GGML_TYPE_F32, 4, ne);
 
     int32_t params[] = { npx, npy, w };
     ggml_set_op_params(result, params, sizeof(params));
 
-    result->op   = GGML_OP_WIN_PART;
-    result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL;
+    result->op     = GGML_OP_WIN_PART;
     result->src[0] = a;
 
     return result;
@@ -5805,21 +7348,13 @@ struct ggml_tensor * ggml_win_unpart(
         int                   w) {
     GGML_ASSERT(a->type == GGML_TYPE_F32);
 
-    bool is_node = false;
-
-    if (a->grad) {
-        GGML_ASSERT(false); // TODO: implement backward
-        is_node = true;
-    }
-
     const int64_t ne[4] = { a->ne[0], w0, h0, 1, };
     struct ggml_tensor * result = ggml_new_tensor(ctx, GGML_TYPE_F32, 3, ne);
 
     int32_t params[] = { w };
     ggml_set_op_params(result, params, sizeof(params));
 
-    result->op   = GGML_OP_WIN_UNPART;
-    result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL;
+    result->op     = GGML_OP_WIN_UNPART;
     result->src[0] = a;
 
     return result;
@@ -5835,18 +7370,10 @@ struct ggml_tensor * ggml_get_rel_pos(
     GGML_ASSERT(qh == kh);
     GGML_ASSERT(2*MAX(qh, kh) - 1 == a->ne[1]);
 
-    bool is_node = false;
-
-    if (a->grad) {
-        GGML_ASSERT(false); // TODO: implement backward
-        is_node = true;
-    }
-
     const int64_t ne[4] = { a->ne[0], kh, qh, 1, };
     struct ggml_tensor * result = ggml_new_tensor(ctx, GGML_TYPE_F16, 3, ne);
 
-    result->op   = GGML_OP_GET_REL_POS;
-    result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL;
+    result->op     = GGML_OP_GET_REL_POS;
     result->src[0] = a;
 
     return result;
@@ -5870,17 +7397,10 @@ static struct ggml_tensor * ggml_add_rel_pos_impl(
     GGML_ASSERT(pw->ne[0]*pw->ne[0] == a->ne[0]);
     GGML_ASSERT(pw->ne[1]*pw->ne[2] == a->ne[1]);
 
-    bool is_node = false;
-
-    if (!inplace && (a->grad || pw->grad || ph->grad)) {
-        is_node = true;
-    }
-
     struct ggml_tensor * result = inplace ? ggml_view_tensor(ctx, a) : ggml_dup_tensor(ctx, a);
     ggml_set_op_params_i32(result, 0, inplace ? 1 : 0);
 
-    result->op   = GGML_OP_ADD_REL_POS;
-    result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL;
+    result->op     = GGML_OP_ADD_REL_POS;
     result->src[0] = a;
     result->src[1] = pw;
     result->src[2] = ph;
@@ -5904,25 +7424,65 @@ struct ggml_tensor * ggml_add_rel_pos_inplace(
     return ggml_add_rel_pos_impl(ctx, a, pw, ph, true);
 }
 
-// gmml_unary
+// ggml_rwkv_wkv
 
-static struct ggml_tensor * ggml_unary_impl(
+struct ggml_tensor * ggml_rwkv_wkv(
         struct ggml_context * ctx,
-        struct ggml_tensor * a,
-        enum ggml_unary_op op,
-        bool inplace) {
-    bool is_node = false;
-
-    if (!inplace && (a->grad)) {
-        is_node = true;
+        struct ggml_tensor  * k,
+        struct ggml_tensor  * v,
+        struct ggml_tensor  * r,
+        struct ggml_tensor  * tf,
+        struct ggml_tensor  * td,
+        struct ggml_tensor  * state) {
+    GGML_ASSERT(ggml_is_contiguous(k));
+    GGML_ASSERT(ggml_is_contiguous(v));
+    GGML_ASSERT(ggml_is_contiguous(r));
+    GGML_ASSERT(ggml_is_contiguous(tf));
+    GGML_ASSERT(ggml_is_contiguous(td));
+    GGML_ASSERT(ggml_is_contiguous(state));
+
+    const int64_t S = k->ne[0];
+    const int64_t H = k->ne[2];
+    const int64_t n_tokens = k->ne[3];
+    const int64_t n_seqs = state->ne[1];
+    {
+        GGML_ASSERT(k->ne[1] == 1);
+        GGML_ASSERT(v->ne[0] == 1 && v->ne[1] == S && v->ne[2] == H && v->ne[3] == n_tokens);
+        GGML_ASSERT(r->ne[0] == 1 && r->ne[1] == S && r->ne[2] == H && r->ne[3] == n_tokens);
+        // TODO: RWKV v4 and v5
+        GGML_ASSERT(td->ne[0] == 1 && td->ne[1] == S && td->ne[2] == H && td->ne[3] == n_tokens);
+        GGML_ASSERT(ggml_nelements(state) == S * S * H * n_seqs);
     }
 
+    // concat output and new_state
+    const int64_t ne[4] = { S * H, n_tokens + S * n_seqs, 1, 1 };
+    struct ggml_tensor * result = ggml_new_tensor(ctx, GGML_TYPE_F32, 4, ne);
+
+    result->op     = GGML_OP_RWKV_WKV;
+    result->src[0] = k;
+    result->src[1] = v;
+    result->src[2] = r;
+    result->src[3] = tf;
+    result->src[4] = td;
+    result->src[5] = state;
+
+    return result;
+}
+
+// ggml_unary
+
+static struct ggml_tensor * ggml_unary_impl(
+        struct ggml_context * ctx,
+        struct ggml_tensor  * a,
+        enum ggml_unary_op    op,
+        bool                  inplace) {
+    GGML_ASSERT(ggml_is_contiguous_1(a));
+
     struct ggml_tensor * result = inplace ? ggml_view_tensor(ctx, a) : ggml_dup_tensor(ctx, a);
 
     ggml_set_op_params_i32(result, 0, (int32_t) op);
 
-    result->op   = GGML_OP_UNARY;
-    result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL;
+    result->op     = GGML_OP_UNARY;
     result->src[0] = a;
 
     return result;
@@ -5931,14 +7491,14 @@ static struct ggml_tensor * ggml_unary_impl(
 struct ggml_tensor * ggml_unary(
         struct ggml_context * ctx,
         struct ggml_tensor  * a,
-        enum ggml_unary_op op) {
+        enum ggml_unary_op    op) {
     return ggml_unary_impl(ctx, a, op, false);
 }
 
 struct ggml_tensor * ggml_unary_inplace(
         struct ggml_context * ctx,
         struct ggml_tensor  * a,
-        enum ggml_unary_op op) {
+        enum ggml_unary_op    op) {
     return ggml_unary_impl(ctx, a, op, true);
 }
 
@@ -5947,20 +7507,13 @@ struct ggml_tensor * ggml_unary_inplace(
 static struct ggml_tensor * ggml_map_unary_impl_f32(
         struct ggml_context        * ctx,
         struct ggml_tensor         * a,
-        const  ggml_unary_op_f32_t fun,
-        bool   inplace) {
-    bool is_node = false;
-
-    if (!inplace && a->grad) {
-        is_node = true;
-    }
-
+        const  ggml_unary_op_f32_t   fun,
+        bool                         inplace) {
     struct ggml_tensor * result = inplace ? ggml_view_tensor(ctx, a) : ggml_dup_tensor(ctx, a);
 
     ggml_set_op_params(result, (const void *) &fun, sizeof(fun));
 
-    result->op = GGML_OP_MAP_UNARY;
-    result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL;
+    result->op     = GGML_OP_MAP_UNARY;
     result->src[0] = a;
 
     return result;
@@ -5969,14 +7522,14 @@ static struct ggml_tensor * ggml_map_unary_impl_f32(
 struct ggml_tensor * ggml_map_unary_f32(
         struct ggml_context        * ctx,
         struct ggml_tensor         * a,
-        const  ggml_unary_op_f32_t fun) {
+        const  ggml_unary_op_f32_t   fun) {
     return ggml_map_unary_impl_f32(ctx, a, fun, false);
 }
 
 struct ggml_tensor * ggml_map_unary_inplace_f32(
         struct ggml_context        * ctx,
         struct ggml_tensor         * a,
-        const  ggml_unary_op_f32_t fun) {
+        const  ggml_unary_op_f32_t   fun) {
     return ggml_map_unary_impl_f32(ctx, a, fun, true);
 }
 
@@ -5986,22 +7539,15 @@ static struct ggml_tensor * ggml_map_binary_impl_f32(
         struct ggml_context         * ctx,
         struct ggml_tensor          * a,
         struct ggml_tensor          * b,
-        const  ggml_binary_op_f32_t fun,
-        bool   inplace) {
+        const  ggml_binary_op_f32_t   fun,
+        bool                          inplace) {
     GGML_ASSERT(ggml_are_same_shape(a, b));
 
-    bool is_node = false;
-
-    if (!inplace && (a->grad || b->grad)) {
-        is_node = true;
-    }
-
     struct ggml_tensor * result = inplace ? ggml_view_tensor(ctx, a) : ggml_dup_tensor(ctx, a);
 
     ggml_set_op_params(result, (const void *) &fun, sizeof(fun));
 
-    result->op = GGML_OP_MAP_BINARY;
-    result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL;
+    result->op     = GGML_OP_MAP_BINARY;
     result->src[0] = a;
     result->src[1] = b;
 
@@ -6012,7 +7558,7 @@ struct ggml_tensor * ggml_map_binary_f32(
         struct ggml_context         * ctx,
         struct ggml_tensor          * a,
         struct ggml_tensor          * b,
-        const  ggml_binary_op_f32_t fun) {
+        const  ggml_binary_op_f32_t   fun) {
     return ggml_map_binary_impl_f32(ctx, a, b, fun, false);
 }
 
@@ -6020,7 +7566,7 @@ struct ggml_tensor * ggml_map_binary_inplace_f32(
         struct ggml_context         * ctx,
         struct ggml_tensor          * a,
         struct ggml_tensor          * b,
-        const  ggml_binary_op_f32_t fun) {
+        const  ggml_binary_op_f32_t   fun) {
     return ggml_map_binary_impl_f32(ctx, a, b, fun, true);
 }
 
@@ -6030,19 +7576,12 @@ static struct ggml_tensor * ggml_map_custom1_impl_f32(
         struct ggml_context          * ctx,
         struct ggml_tensor           * a,
         const  ggml_custom1_op_f32_t   fun,
-        bool   inplace) {
-    bool is_node = false;
-
-    if (!inplace && a->grad) {
-        is_node = true;
-    }
-
+        bool                           inplace) {
     struct ggml_tensor * result = inplace ? ggml_view_tensor(ctx, a) : ggml_dup_tensor(ctx, a);
 
     ggml_set_op_params(result, (const void *) &fun, sizeof(fun));
 
-    result->op = GGML_OP_MAP_CUSTOM1_F32;
-    result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL;
+    result->op     = GGML_OP_MAP_CUSTOM1_F32;
     result->src[0] = a;
 
     return result;
@@ -6069,19 +7608,12 @@ static struct ggml_tensor * ggml_map_custom2_impl_f32(
         struct ggml_tensor           * a,
         struct ggml_tensor           * b,
         const  ggml_custom2_op_f32_t   fun,
-        bool   inplace) {
-    bool is_node = false;
-
-    if (!inplace && (a->grad || b->grad)) {
-        is_node = true;
-    }
-
+        bool                           inplace) {
     struct ggml_tensor * result = inplace ? ggml_view_tensor(ctx, a) : ggml_dup_tensor(ctx, a);
 
     ggml_set_op_params(result, (const void *) &fun, sizeof(fun));
 
-    result->op = GGML_OP_MAP_CUSTOM2_F32;
-    result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL;
+    result->op     = GGML_OP_MAP_CUSTOM2_F32;
     result->src[0] = a;
     result->src[1] = b;
 
@@ -6112,19 +7644,12 @@ static struct ggml_tensor * ggml_map_custom3_impl_f32(
         struct ggml_tensor           * b,
         struct ggml_tensor           * c,
         const  ggml_custom3_op_f32_t   fun,
-        bool   inplace) {
-    bool is_node = false;
-
-    if (!inplace && (a->grad || b->grad || c->grad)) {
-        is_node = true;
-    }
-
+        bool                           inplace) {
     struct ggml_tensor * result = inplace ? ggml_view_tensor(ctx, a) : ggml_dup_tensor(ctx, a);
 
     ggml_set_op_params(result, (const void *) &fun, sizeof(fun));
 
-    result->op = GGML_OP_MAP_CUSTOM3_F32;
-    result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL;
+    result->op     = GGML_OP_MAP_CUSTOM3_F32;
     result->src[0] = a;
     result->src[1] = b;
     result->src[2] = c;
@@ -6152,26 +7677,20 @@ struct ggml_tensor * ggml_map_custom3_inplace_f32(
 
 // ggml_map_custom1
 struct ggml_map_custom1_op_params {
-    ggml_custom1_op_t fun;
-    int n_tasks;
-    void * userdata;
+    ggml_custom1_op_t  fun;
+    int                n_tasks;
+    void             * userdata;
 };
 
 static struct ggml_tensor * ggml_map_custom1_impl(
-        struct ggml_context          * ctx,
-        struct ggml_tensor           * a,
-        const  ggml_custom1_op_t       fun,
-        int                            n_tasks,
-        void                         * userdata,
-        bool                           inplace) {
+        struct ggml_context      * ctx,
+        struct ggml_tensor       * a,
+        const  ggml_custom1_op_t   fun,
+        int                        n_tasks,
+        void                     * userdata,
+        bool                       inplace) {
     GGML_ASSERT(n_tasks == GGML_N_TASKS_MAX || n_tasks > 0);
 
-    bool is_node = false;
-
-    if (!inplace && a->grad) {
-        is_node = true;
-    }
-
     struct ggml_tensor * result = inplace ? ggml_view_tensor(ctx, a) : ggml_dup_tensor(ctx, a);
 
     struct ggml_map_custom1_op_params params = {
@@ -6181,55 +7700,48 @@ static struct ggml_tensor * ggml_map_custom1_impl(
     };
     ggml_set_op_params(result, (const void *) &params, sizeof(params));
 
-    result->op = GGML_OP_MAP_CUSTOM1;
-    result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL;
+    result->op     = GGML_OP_MAP_CUSTOM1;
     result->src[0] = a;
 
     return result;
 }
 
 struct ggml_tensor * ggml_map_custom1(
-        struct ggml_context          * ctx,
-        struct ggml_tensor           * a,
-        const  ggml_custom1_op_t       fun,
-        int                            n_tasks,
-        void                         * userdata) {
+        struct ggml_context      * ctx,
+        struct ggml_tensor       * a,
+        const  ggml_custom1_op_t   fun,
+        int                        n_tasks,
+        void                     * userdata) {
     return ggml_map_custom1_impl(ctx, a, fun, n_tasks, userdata, false);
 }
 
 struct ggml_tensor * ggml_map_custom1_inplace(
-        struct ggml_context          * ctx,
-        struct ggml_tensor           * a,
-        const  ggml_custom1_op_t       fun,
-        int                            n_tasks,
-        void                         * userdata) {
+        struct ggml_context      * ctx,
+        struct ggml_tensor       * a,
+        const  ggml_custom1_op_t   fun,
+        int                        n_tasks,
+        void                     * userdata) {
     return ggml_map_custom1_impl(ctx, a, fun, n_tasks, userdata, true);
 }
 
 // ggml_map_custom2
 
 struct ggml_map_custom2_op_params {
-    ggml_custom2_op_t fun;
-    int n_tasks;
-    void * userdata;
+    ggml_custom2_op_t   fun;
+    int                 n_tasks;
+    void              * userdata;
 };
 
 static struct ggml_tensor * ggml_map_custom2_impl(
-        struct ggml_context          * ctx,
-        struct ggml_tensor           * a,
-        struct ggml_tensor           * b,
-        const  ggml_custom2_op_t       fun,
-        int                            n_tasks,
-        void                         * userdata,
-        bool                           inplace) {
+        struct ggml_context      * ctx,
+        struct ggml_tensor       * a,
+        struct ggml_tensor       * b,
+        const  ggml_custom2_op_t   fun,
+        int                        n_tasks,
+        void                     * userdata,
+        bool                       inplace) {
     GGML_ASSERT(n_tasks == GGML_N_TASKS_MAX || n_tasks > 0);
 
-    bool is_node = false;
-
-    if (!inplace && (a->grad || b->grad)) {
-        is_node = true;
-    }
-
     struct ggml_tensor * result = inplace ? ggml_view_tensor(ctx, a) : ggml_dup_tensor(ctx, a);
 
     struct ggml_map_custom2_op_params params = {
@@ -6239,8 +7751,7 @@ static struct ggml_tensor * ggml_map_custom2_impl(
     };
     ggml_set_op_params(result, (const void *) &params, sizeof(params));
 
-    result->op = GGML_OP_MAP_CUSTOM2;
-    result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL;
+    result->op     = GGML_OP_MAP_CUSTOM2;
     result->src[0] = a;
     result->src[1] = b;
 
@@ -6248,22 +7759,22 @@ static struct ggml_tensor * ggml_map_custom2_impl(
 }
 
 struct ggml_tensor * ggml_map_custom2(
-        struct ggml_context          * ctx,
-        struct ggml_tensor           * a,
-        struct ggml_tensor           * b,
-        const  ggml_custom2_op_t       fun,
-        int                            n_tasks,
-        void                         * userdata) {
+        struct ggml_context      * ctx,
+        struct ggml_tensor       * a,
+        struct ggml_tensor       * b,
+        const  ggml_custom2_op_t   fun,
+        int                        n_tasks,
+        void                     * userdata) {
     return ggml_map_custom2_impl(ctx, a, b, fun, n_tasks, userdata, false);
 }
 
 struct ggml_tensor * ggml_map_custom2_inplace(
-        struct ggml_context          * ctx,
-        struct ggml_tensor           * a,
-        struct ggml_tensor           * b,
-        const  ggml_custom2_op_t       fun,
-        int                            n_tasks,
-        void                         * userdata) {
+        struct ggml_context      * ctx,
+        struct ggml_tensor       * a,
+        struct ggml_tensor       * b,
+        const  ggml_custom2_op_t   fun,
+        int                        n_tasks,
+        void                     * userdata) {
     return ggml_map_custom2_impl(ctx, a, b, fun, n_tasks, userdata, true);
 }
 
@@ -6276,22 +7787,16 @@ struct ggml_map_custom3_op_params {
 };
 
 static struct ggml_tensor * ggml_map_custom3_impl(
-        struct ggml_context          * ctx,
-        struct ggml_tensor           * a,
-        struct ggml_tensor           * b,
-        struct ggml_tensor           * c,
-        const  ggml_custom3_op_t       fun,
-        int                            n_tasks,
-        void                         * userdata,
-        bool                           inplace) {
+        struct ggml_context      * ctx,
+        struct ggml_tensor       * a,
+        struct ggml_tensor       * b,
+        struct ggml_tensor       * c,
+        const  ggml_custom3_op_t   fun,
+        int                        n_tasks,
+        void                     * userdata,
+        bool                       inplace) {
     GGML_ASSERT(n_tasks == GGML_N_TASKS_MAX || n_tasks > 0);
 
-    bool is_node = false;
-
-    if (!inplace && (a->grad || b->grad || c->grad)) {
-        is_node = true;
-    }
-
     struct ggml_tensor * result = inplace ? ggml_view_tensor(ctx, a) : ggml_dup_tensor(ctx, a);
 
     struct ggml_map_custom3_op_params params = {
@@ -6301,8 +7806,7 @@ static struct ggml_tensor * ggml_map_custom3_impl(
     };
     ggml_set_op_params(result, (const void *) &params, sizeof(params));
 
-    result->op = GGML_OP_MAP_CUSTOM3;
-    result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL;
+    result->op     = GGML_OP_MAP_CUSTOM3;
     result->src[0] = a;
     result->src[1] = b;
     result->src[2] = c;
@@ -6311,44 +7815,38 @@ static struct ggml_tensor * ggml_map_custom3_impl(
 }
 
 struct ggml_tensor * ggml_map_custom3(
-        struct ggml_context          * ctx,
-        struct ggml_tensor           * a,
-        struct ggml_tensor           * b,
-        struct ggml_tensor           * c,
-        const  ggml_custom3_op_t       fun,
-        int                            n_tasks,
-        void                         * userdata) {
+        struct ggml_context      * ctx,
+        struct ggml_tensor       * a,
+        struct ggml_tensor       * b,
+        struct ggml_tensor       * c,
+        const  ggml_custom3_op_t   fun,
+        int                        n_tasks,
+        void                     * userdata) {
     return ggml_map_custom3_impl(ctx, a, b, c, fun, n_tasks, userdata, false);
 }
 
 struct ggml_tensor * ggml_map_custom3_inplace(
-        struct ggml_context          * ctx,
-        struct ggml_tensor           * a,
-        struct ggml_tensor           * b,
-        struct ggml_tensor           * c,
-        const  ggml_custom3_op_t       fun,
-        int                            n_tasks,
-        void                         * userdata) {
+        struct ggml_context      * ctx,
+        struct ggml_tensor       * a,
+        struct ggml_tensor       * b,
+        struct ggml_tensor       * c,
+        const  ggml_custom3_op_t   fun,
+        int                        n_tasks,
+        void                     * userdata) {
     return ggml_map_custom3_impl(ctx, a, b, c, fun, n_tasks, userdata, true);
 }
 
 // ggml_cross_entropy_loss
 
 struct ggml_tensor * ggml_cross_entropy_loss(
-        struct ggml_context         * ctx,
-        struct ggml_tensor          * a,
-        struct ggml_tensor          * b) {
+        struct ggml_context * ctx,
+        struct ggml_tensor  * a,
+        struct ggml_tensor  * b) {
     GGML_ASSERT(ggml_are_same_shape(a, b));
-    bool is_node = false;
-
-    if (a->grad || b->grad) {
-        is_node = true;
-    }
 
     struct ggml_tensor * result = ggml_new_tensor_1d(ctx, a->type, 1);
 
-    result->op   = GGML_OP_CROSS_ENTROPY_LOSS;
-    result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL;
+    result->op     = GGML_OP_CROSS_ENTROPY_LOSS;
     result->src[0] = a;
     result->src[1] = b;
 
@@ -6358,17 +7856,16 @@ struct ggml_tensor * ggml_cross_entropy_loss(
 // ggml_cross_entropy_loss_back
 
 struct ggml_tensor * ggml_cross_entropy_loss_back(
-        struct ggml_context         * ctx,
-        struct ggml_tensor          * a,
-        struct ggml_tensor          * b,
-        struct ggml_tensor          * c) {
+        struct ggml_context * ctx,
+        struct ggml_tensor  * a,
+        struct ggml_tensor  * b,
+        struct ggml_tensor  * c) {
     GGML_ASSERT(ggml_are_same_shape(a, b));
     GGML_ASSERT(ggml_is_scalar(c));
 
     struct ggml_tensor * result = ggml_dup_tensor(ctx, a);
 
-    result->op   = GGML_OP_CROSS_ENTROPY_LOSS_BACK;
-    result->grad = NULL;
+    result->op     = GGML_OP_CROSS_ENTROPY_LOSS_BACK;
     result->src[0] = a;
     result->src[1] = b;
     result->src[2] = c;
@@ -6376,34 +7873,59 @@ struct ggml_tensor * ggml_cross_entropy_loss_back(
     return result;
 }
 
-////////////////////////////////////////////////////////////////////////////////
+// opt_step_adamw
 
-void ggml_set_param(
+struct ggml_tensor * ggml_opt_step_adamw(
         struct ggml_context * ctx,
-        struct ggml_tensor * tensor) {
-    tensor->is_param = true;
+        struct ggml_tensor  * a,
+        struct ggml_tensor  * grad,
+        float                 alpha,
+        float                 beta1,
+        float                 beta2,
+        float                 eps,
+        float                 wd) {
+    GGML_ASSERT(a->flags & GGML_TENSOR_FLAG_PARAM);
+    GGML_ASSERT(ggml_are_same_shape(a, grad));
+    GGML_ASSERT(alpha >  0.0f);
+    GGML_ASSERT(beta1 >= 0.0f && beta1 <= 1.0f);
+    GGML_ASSERT(beta2 >= 0.0f && beta2 <= 1.0f);
+    GGML_ASSERT(eps   >= 0.0f);
+    GGML_ASSERT(wd    >= 0.0f && wd    <= 1.0f);
+
+    struct ggml_tensor * result = ggml_view_tensor(ctx, a);
 
-    GGML_ASSERT(tensor->grad == NULL);
-    tensor->grad = ggml_dup_tensor(ctx, tensor);
-    ggml_format_name(tensor->grad, "%s (grad)", tensor->name);
+    const int64_t iter = 1;
+    memcpy(&result->op_params[0], &iter, sizeof(int64_t));
+    ggml_set_op_params_f32(result, 2, alpha);
+    ggml_set_op_params_f32(result, 3, beta1);
+    ggml_set_op_params_f32(result, 4, beta2);
+    ggml_set_op_params_f32(result, 5, eps);
+    ggml_set_op_params_f32(result, 6, wd);
+
+    result->op     = GGML_OP_OPT_STEP_ADAMW;
+    result->src[0] = a;
+    result->src[1] = grad;
+    result->src[2] = ggml_dup_tensor(ctx, grad);
+    result->src[3] = ggml_dup_tensor(ctx, grad);
+
+    return result;
 }
 
+////////////////////////////////////////////////////////////////////////////////
+
 // ggml_compute_forward_dup
 
 static void ggml_compute_forward_dup_same_cont(
         const struct ggml_compute_params * params,
-        const struct ggml_tensor * src0,
         struct ggml_tensor * dst) {
+
+    const struct ggml_tensor * src0 = dst->src[0];
+
     GGML_ASSERT(ggml_nelements(dst) == ggml_nelements(src0));
     GGML_ASSERT(ggml_is_contiguous(dst) && ggml_is_contiguous(src0));
     GGML_ASSERT(src0->type == dst->type);
 
-    if (params->type == GGML_TASK_INIT || params->type == GGML_TASK_FINALIZE) {
-        return;
-    }
-
-    const size_t nb00 = src0->nb[0];
-    const size_t nb0 = dst->nb[0];
+    const size_t nb0 = ggml_type_size(src0->type);
 
     const int ith = params->ith; // thread index
     const int nth = params->nth; // number of threads
@@ -6417,31 +7939,24 @@ static void ggml_compute_forward_dup_same_cont(
     if (ie0 < ie1) {
         memcpy(
             ((char *)  dst->data + ie0*nb0),
-            ((char *) src0->data + ie0*nb00),
-            (ie1 - ie0) * ggml_type_size(src0->type));
+            ((char *) src0->data + ie0*nb0),
+            (ie1 - ie0) * nb0);
     }
-
 }
+
 static void ggml_compute_forward_dup_f16(
         const struct ggml_compute_params * params,
-        const struct ggml_tensor * src0,
         struct ggml_tensor * dst) {
-    GGML_ASSERT(ggml_nelements(dst) == ggml_nelements(src0));
 
-    if (params->type == GGML_TASK_INIT || params->type == GGML_TASK_FINALIZE) {
-        return;
-    }
+    const struct ggml_tensor * src0 = dst->src[0];
+
+    GGML_ASSERT(ggml_nelements(dst) == ggml_nelements(src0));
 
     GGML_TENSOR_UNARY_OP_LOCALS
 
     const int ith = params->ith; // thread index
     const int nth = params->nth; // number of threads
 
-    if (ggml_is_contiguous(src0) && ggml_is_contiguous(dst) && src0->type == dst->type) {
-        ggml_compute_forward_dup_same_cont(params, src0, dst);
-        return;
-    }
-
     // parallelize by rows
     const int nr = ne01;
     // number of rows per thread
@@ -6530,7 +8045,7 @@ static void ggml_compute_forward_dup_f16(
                     }
                 }
             } else {
-                GGML_ASSERT(false); // TODO: implement
+                GGML_ABORT("fatal error"); // TODO: implement
             }
         } else {
             //printf("%s: this is not optimal - fix me\n", __func__);
@@ -6572,7 +8087,7 @@ static void ggml_compute_forward_dup_f16(
                     }
                 }
             } else {
-                GGML_ASSERT(false); // TODO: implement
+                GGML_ABORT("fatal error"); // TODO: implement
             }
         }
         return;
@@ -6689,30 +8204,23 @@ static void ggml_compute_forward_dup_f16(
             }
         }
     } else {
-        GGML_ASSERT(false); // TODO: implement
+        GGML_ABORT("fatal error"); // TODO: implement
     }
 }
 
-static void ggml_compute_forward_dup_f32(
+static void ggml_compute_forward_dup_bf16(
         const struct ggml_compute_params * params,
-        const struct ggml_tensor * src0,
         struct ggml_tensor * dst) {
-    GGML_ASSERT(ggml_nelements(dst) == ggml_nelements(src0));
 
-    if (params->type == GGML_TASK_INIT || params->type == GGML_TASK_FINALIZE) {
-        return;
-    }
+    const struct ggml_tensor * src0 = dst->src[0];
+
+    GGML_ASSERT(ggml_nelements(dst) == ggml_nelements(src0));
 
     GGML_TENSOR_UNARY_OP_LOCALS
 
     const int ith = params->ith; // thread index
     const int nth = params->nth; // number of threads
 
-    if (ggml_is_contiguous(src0) && ggml_is_contiguous(dst) && src0->type == dst->type) {
-        ggml_compute_forward_dup_same_cont(params, src0, dst);
-        return;
-    }
-
     // parallelize by rows
     const int nr = ne01;
     // number of rows per thread
@@ -6739,10 +8247,11 @@ static void ggml_compute_forward_dup_f32(
         return;
     }
 
+    // TODO: add more special-case implementations for tensor shapes/strides that can benefit from memcpy
+
     if (ggml_is_contiguous(dst)) {
-        // TODO: simplify
-        if (nb00 == sizeof(float)) {
-            if (dst->type == GGML_TYPE_F32) {
+        if (nb00 == sizeof(ggml_bf16_t)) {
+            if (dst->type == GGML_TYPE_BF16) {
                 size_t id = 0;
                 const size_t rs = ne00 * nb00;
                 char * dst_ptr = (char *) dst->data;
@@ -6758,8 +8267,43 @@ static void ggml_compute_forward_dup_f32(
                         id += rs * (ne01 - ir1);
                     }
                 }
+            } else if (dst->type == GGML_TYPE_F16) {
+                size_t id = 0;
+                ggml_fp16_t * dst_ptr = (ggml_fp16_t *) dst->data;
+
+                for (int i03 = 0; i03 < ne03; i03++) {
+                    for (int i02 = 0; i02 < ne02; i02++) {
+                        id += ne00 * ir0;
+                        for (int i01 = ir0; i01 < ir1; i01++) {
+                            const ggml_bf16_t * src0_ptr = (ggml_bf16_t *) ((char *) src0->data + i01*nb01 + i02*nb02 + i03*nb03);
+                            for (int i00 = 0; i00 < ne00; i00++) {
+                                dst_ptr[id] = GGML_FP32_TO_FP16(GGML_BF16_TO_FP32(src0_ptr[i00]));
+                                id++;
+                            }
+                        }
+                        id += ne00 * (ne01 - ir1);
+                    }
+                }
+            } else if (dst->type == GGML_TYPE_F32) {
+                size_t id = 0;
+                float * dst_ptr = (float *) dst->data;
+
+                for (int i03 = 0; i03 < ne03; i03++) {
+                    for (int i02 = 0; i02 < ne02; i02++) {
+                        id += ne00 * ir0;
+                        for (int i01 = ir0; i01 < ir1; i01++) {
+                            const ggml_bf16_t * src0_ptr = (ggml_bf16_t *) ((char *) src0->data + i01*nb01 + i02*nb02 + i03*nb03);
+                            for (int i00 = 0; i00 < ne00; i00++) {
+                                dst_ptr[id] = GGML_BF16_TO_FP32(src0_ptr[i00]);
+                                id++;
+                            }
+                        }
+                        id += ne00 * (ne01 - ir1);
+                    }
+                }
             } else if (type_traits[dst->type].from_float) {
                 ggml_from_float_t const quantize_row_q = type_traits[dst->type].from_float;
+                float * src0_f32 = (float *) params->wdata + (ne00 + CACHE_LINE_SIZE_F32) * ith;
 
                 size_t id = 0;
                 size_t rs = nb0 * (ne00 / ggml_blck_size(dst->type));
@@ -6769,15 +8313,20 @@ static void ggml_compute_forward_dup_f32(
                     for (int i02 = 0; i02 < ne02; i02++) {
                         id += rs * ir0;
                         for (int i01 = ir0; i01 < ir1; i01++) {
-                            const float * src0_ptr = (float *) ((char *) src0->data + i01*nb01 + i02*nb02 + i03*nb03);
-                            quantize_row_q(src0_ptr, dst_ptr + id, ne00);
+                            const ggml_bf16_t * src0_ptr = (ggml_bf16_t *) ((char *) src0->data + i01*nb01 + i02*nb02 + i03*nb03);
+
+                            for (int i00 = 0; i00 < ne00; i00++) {
+                                src0_f32[i00] = GGML_BF16_TO_FP32(src0_ptr[i00]);
+                            }
+
+                            quantize_row_q(src0_f32, dst_ptr + id, ne00);
                             id += rs;
                         }
                         id += rs * (ne01 - ir1);
                     }
                 }
             } else {
-                GGML_ASSERT(false); // TODO: implement
+                GGML_ABORT("fatal error"); // TODO: implement
             }
         } else {
             //printf("%s: this is not optimal - fix me\n", __func__);
@@ -6791,7 +8340,25 @@ static void ggml_compute_forward_dup_f32(
                         id += ne00 * ir0;
                         for (int i01 = ir0; i01 < ir1; i01++) {
                             for (int i00 = 0; i00 < ne00; i00++) {
-                                const float * src0_ptr = (float *) ((char *) src0->data + i00*nb00 + i01*nb01 + i02*nb02 + i03*nb03);
+                                const ggml_bf16_t * src0_ptr = (ggml_bf16_t *) ((char *) src0->data + i00*nb00 + i01*nb01 + i02*nb02 + i03*nb03);
+
+                                dst_ptr[id] = GGML_BF16_TO_FP32(*src0_ptr);
+                                id++;
+                            }
+                        }
+                        id += ne00 * (ne01 - ir1);
+                    }
+                }
+            } else if (dst->type == GGML_TYPE_BF16) {
+                size_t id = 0;
+                ggml_bf16_t * dst_ptr = (ggml_bf16_t *) dst->data;
+
+                for (int i03 = 0; i03 < ne03; i03++) {
+                    for (int i02 = 0; i02 < ne02; i02++) {
+                        id += ne00 * ir0;
+                        for (int i01 = ir0; i01 < ir1; i01++) {
+                            for (int i00 = 0; i00 < ne00; i00++) {
+                                const ggml_bf16_t * src0_ptr = (ggml_bf16_t *) ((char *) src0->data + i00*nb00 + i01*nb01 + i02*nb02 + i03*nb03);
 
                                 dst_ptr[id] = *src0_ptr;
                                 id++;
@@ -6809,9 +8376,9 @@ static void ggml_compute_forward_dup_f32(
                         id += ne00 * ir0;
                         for (int i01 = ir0; i01 < ir1; i01++) {
                             for (int i00 = 0; i00 < ne00; i00++) {
-                                const float * src0_ptr = (float *) ((char *) src0->data + i00*nb00 + i01*nb01 + i02*nb02 + i03*nb03);
+                                const ggml_bf16_t * src0_ptr = (ggml_bf16_t *) ((char *) src0->data + i00*nb00 + i01*nb01 + i02*nb02 + i03*nb03);
 
-                                dst_ptr[id] = GGML_FP32_TO_FP16(*src0_ptr);
+                                dst_ptr[id] = GGML_FP32_TO_FP16(GGML_BF16_TO_FP32(*src0_ptr));
                                 id++;
                             }
                         }
@@ -6819,21 +8386,19 @@ static void ggml_compute_forward_dup_f32(
                     }
                 }
             } else {
-                GGML_ASSERT(false); // TODO: implement
+                GGML_ABORT("fatal error"); // TODO: implement
             }
         }
-
         return;
     }
 
     // dst counters
-
     int64_t i10 = 0;
     int64_t i11 = 0;
     int64_t i12 = 0;
     int64_t i13 = 0;
 
-    if (dst->type == GGML_TYPE_F32) {
+    if (dst->type == GGML_TYPE_BF16) {
         for (int64_t i03 = 0; i03 < ne03; i03++) {
             for (int64_t i02 = 0; i02 < ne02; i02++) {
                 i10 += ne00 * ir0;
@@ -6854,7 +8419,59 @@ static void ggml_compute_forward_dup_f32(
                         const char * src0_ptr = ((char *) src0->data + i00*nb00 + i01*nb01 + i02*nb02 + i03*nb03);
                               char * dst_ptr  = ((char *)  dst->data + i10*nb0  + i11*nb1  + i12*nb2  + i13*nb3);
 
-                        memcpy(dst_ptr, src0_ptr, sizeof(float));
+                        memcpy(dst_ptr, src0_ptr, sizeof(ggml_bf16_t));
+
+                        if (++i10 == ne00) {
+                            i10 = 0;
+                            if (++i11 == ne01) {
+                                i11 = 0;
+                                if (++i12 == ne02) {
+                                    i12 = 0;
+                                    if (++i13 == ne03) {
+                                        i13 = 0;
+                                    }
+                                }
+                            }
+                        }
+                    }
+                }
+                i10 += ne00 * (ne01 - ir1);
+                while (i10 >= ne0) {
+                    i10 -= ne0;
+                    if (++i11 == ne1) {
+                        i11 = 0;
+                        if (++i12 == ne2) {
+                            i12 = 0;
+                            if (++i13 == ne3) {
+                                i13 = 0;
+                            }
+                        }
+                    }
+                }
+            }
+        }
+    } else if (dst->type == GGML_TYPE_F16) {
+        for (int64_t i03 = 0; i03 < ne03; i03++) {
+            for (int64_t i02 = 0; i02 < ne02; i02++) {
+                i10 += ne00 * ir0;
+                while (i10 >= ne0) {
+                    i10 -= ne0;
+                    if (++i11 == ne1) {
+                        i11 = 0;
+                        if (++i12 == ne2) {
+                            i12 = 0;
+                            if (++i13 == ne3) {
+                                i13 = 0;
+                            }
+                        }
+                    }
+                }
+                for (int64_t i01 = ir0; i01 < ir1; i01++) {
+                    for (int64_t i00 = 0; i00 < ne00; i00++) {
+                        const char * src0_ptr = ((char *) src0->data + i00*nb00 + i01*nb01 + i02*nb02 + i03*nb03);
+                              char * dst_ptr  = ((char *)  dst->data + i10*nb0  + i11*nb1  + i12*nb2  + i13*nb3);
+
+                        *(ggml_fp16_t *) dst_ptr = GGML_FP32_TO_FP16(GGML_BF16_TO_FP32(*(const ggml_bf16_t *) src0_ptr));
 
                         if (++i10 == ne0) {
                             i10 = 0;
@@ -6885,7 +8502,7 @@ static void ggml_compute_forward_dup_f32(
                 }
             }
         }
-    } else if (dst->type == GGML_TYPE_F16) {
+    } else if (dst->type == GGML_TYPE_F32) {
         for (int64_t i03 = 0; i03 < ne03; i03++) {
             for (int64_t i02 = 0; i02 < ne02; i02++) {
                 i10 += ne00 * ir0;
@@ -6906,7 +8523,7 @@ static void ggml_compute_forward_dup_f32(
                         const char * src0_ptr = ((char *) src0->data + i00*nb00 + i01*nb01 + i02*nb02 + i03*nb03);
                               char * dst_ptr  = ((char *)  dst->data + i10*nb0  + i11*nb1  + i12*nb2  + i13*nb3);
 
-                        *(ggml_fp16_t *) dst_ptr = GGML_FP32_TO_FP16(*(const float *) src0_ptr);
+                        *(float *) dst_ptr = GGML_BF16_TO_FP32(*(const ggml_bf16_t *) src0_ptr);
 
                         if (++i10 == ne0) {
                             i10 = 0;
@@ -6938,34 +8555,23 @@ static void ggml_compute_forward_dup_f32(
             }
         }
     } else {
-        GGML_ASSERT(false); // TODO: implement
+        GGML_ABORT("fatal error"); // TODO: implement
     }
 }
 
-// A simplified version of ggml_compute_forward_dup that doesn't do float upcasting, and just plain old memcpy.
-static void ggml_compute_forward_dup_bytes(
+static void ggml_compute_forward_dup_f32(
         const struct ggml_compute_params * params,
-        const struct ggml_tensor * src0,
         struct ggml_tensor * dst) {
-    GGML_ASSERT(ggml_nelements(dst) == ggml_nelements(src0));
-    GGML_ASSERT(src0->type == dst->type);
 
-    if (params->type == GGML_TASK_INIT || params->type == GGML_TASK_FINALIZE) {
-        return;
-    }
+    const struct ggml_tensor * src0 = dst->src[0];
 
-    if (ggml_is_contiguous(src0) && ggml_is_contiguous(dst)) {
-        ggml_compute_forward_dup_same_cont(params, src0, dst);
-        return;
-    }
+    GGML_ASSERT(ggml_nelements(dst) == ggml_nelements(src0));
 
-    GGML_TENSOR_UNARY_OP_LOCALS;
+    GGML_TENSOR_UNARY_OP_LOCALS
 
-    const size_t type_size = ggml_type_size(src0->type);
     const int ith = params->ith; // thread index
     const int nth = params->nth; // number of threads
 
-
     // parallelize by rows
     const int nr = ne01;
     // number of rows per thread
@@ -6976,9 +8582,9 @@ static void ggml_compute_forward_dup_bytes(
 
     if (src0->type == dst->type &&
         ne00 == ne0 &&
-        nb00 == type_size && nb0 == type_size) {
+        nb00 == ggml_type_size(src0->type) && nb0 == ggml_type_size(dst->type)) {
         // copy by rows
-        const size_t rs = ne00 * type_size;
+        const size_t rs = ne00*nb00;
         for (int64_t i03 = 0; i03 < ne03; i03++) {
             for (int64_t i02 = 0; i02 < ne02; i02++) {
                 for (int64_t i01 = ir0; i01 < ir1; i01++) {
@@ -6993,39 +8599,104 @@ static void ggml_compute_forward_dup_bytes(
     }
 
     if (ggml_is_contiguous(dst)) {
-        size_t id = 0;
-        char * dst_ptr = (char *) dst->data;
-        const size_t rs = ne00 * type_size;
+        // TODO: simplify
+        if (nb00 == sizeof(float)) {
+            if (dst->type == GGML_TYPE_F32) {
+                size_t id = 0;
+                const size_t rs = ne00 * nb00;
+                char * dst_ptr = (char *) dst->data;
 
-        if (nb00 == type_size) {
-            // src0 is contigous on first dimension, copy by rows
-            for (int64_t i03 = 0; i03 < ne03; i03++) {
-                for (int64_t i02 = 0; i02 < ne02; i02++) {
-                    id += rs * ir0;
-                    for (int64_t i01 = ir0; i01 < ir1; i01++) {
-                        const char * src0_ptr = (char *) src0->data + i01*nb01 + i02*nb02 + i03*nb03;
-                        memcpy(dst_ptr + id, src0_ptr, rs);
-                        id += rs;
+                for (int i03 = 0; i03 < ne03; i03++) {
+                    for (int i02 = 0; i02 < ne02; i02++) {
+                        id += rs * ir0;
+                        for (int i01 = ir0; i01 < ir1; i01++) {
+                            const char * src0_ptr = (char *) src0->data + i01*nb01 + i02*nb02 + i03*nb03;
+                            memcpy(dst_ptr + id, src0_ptr, rs);
+                            id += rs;
+                        }
+                        id += rs * (ne01 - ir1);
+                    }
+                }
+            } else if (type_traits[dst->type].from_float) {
+                ggml_from_float_t const quantize_row_q = type_traits[dst->type].from_float;
+
+                size_t id = 0;
+                size_t rs = nb0 * (ne00 / ggml_blck_size(dst->type));
+                char * dst_ptr = (char *) dst->data;
+
+                for (int i03 = 0; i03 < ne03; i03++) {
+                    for (int i02 = 0; i02 < ne02; i02++) {
+                        id += rs * ir0;
+                        for (int i01 = ir0; i01 < ir1; i01++) {
+                            const float * src0_ptr = (float *) ((char *) src0->data + i01*nb01 + i02*nb02 + i03*nb03);
+                            quantize_row_q(src0_ptr, dst_ptr + id, ne00);
+                            id += rs;
+                        }
+                        id += rs * (ne01 - ir1);
                     }
-                    id += rs * (ne01 - ir1);
                 }
+            } else {
+                GGML_ABORT("fatal error"); // TODO: implement
             }
         } else {
             //printf("%s: this is not optimal - fix me\n", __func__);
 
-            for (int64_t i03 = 0; i03 < ne03; i03++) {
-                for (int64_t i02 = 0; i02 < ne02; i02++) {
-                    id += rs * ir0;
-                    for (int64_t i01 = ir0; i01 < ir1; i01++) {
-                        for (int64_t i00 = 0; i00 < ne00; i00++) {
-                            const char * src0_ptr = (char *) src0->data + i00*nb00 + i01*nb01 + i02*nb02 + i03*nb03;
-                            memcpy(dst_ptr + id, src0_ptr, type_size);
+            if (dst->type == GGML_TYPE_F32) {
+                size_t id = 0;
+                float * dst_ptr = (float *) dst->data;
 
-                            id += type_size;
+                for (int i03 = 0; i03 < ne03; i03++) {
+                    for (int i02 = 0; i02 < ne02; i02++) {
+                        id += ne00 * ir0;
+                        for (int i01 = ir0; i01 < ir1; i01++) {
+                            for (int i00 = 0; i00 < ne00; i00++) {
+                                const float * src0_ptr = (float *) ((char *) src0->data + i00*nb00 + i01*nb01 + i02*nb02 + i03*nb03);
+
+                                dst_ptr[id] = *src0_ptr;
+                                id++;
+                            }
                         }
+                        id += ne00 * (ne01 - ir1);
+                    }
+                }
+            } else if (dst->type == GGML_TYPE_F16) {
+                size_t id = 0;
+                ggml_fp16_t * dst_ptr = (ggml_fp16_t *) dst->data;
+
+                for (int i03 = 0; i03 < ne03; i03++) {
+                    for (int i02 = 0; i02 < ne02; i02++) {
+                        id += ne00 * ir0;
+                        for (int i01 = ir0; i01 < ir1; i01++) {
+                            for (int i00 = 0; i00 < ne00; i00++) {
+                                const float * src0_ptr = (float *) ((char *) src0->data + i00*nb00 + i01*nb01 + i02*nb02 + i03*nb03);
+
+                                dst_ptr[id] = GGML_FP32_TO_FP16(*src0_ptr);
+                                id++;
+                            }
+                        }
+                        id += ne00 * (ne01 - ir1);
+                    }
+                }
+            } else if (dst->type == GGML_TYPE_BF16) {
+                size_t id = 0;
+                ggml_bf16_t * dst_ptr = (ggml_bf16_t *) dst->data;
+
+                for (int i03 = 0; i03 < ne03; i03++) {
+                    for (int i02 = 0; i02 < ne02; i02++) {
+                        id += ne00 * ir0;
+                        for (int i01 = ir0; i01 < ir1; i01++) {
+                            for (int i00 = 0; i00 < ne00; i00++) {
+                                const float * src0_ptr = (float *) ((char *) src0->data + i00*nb00 + i01*nb01 + i02*nb02 + i03*nb03);
+
+                                dst_ptr[id] = GGML_FP32_TO_BF16(*src0_ptr);
+                                id++;
+                            }
+                        }
+                        id += ne00 * (ne01 - ir1);
                     }
-                    id += rs * (ne01 - ir1);
                 }
+            } else {
+                GGML_ABORT("fatal error"); // TODO: implement
             }
         }
 
@@ -7039,81 +8710,343 @@ static void ggml_compute_forward_dup_bytes(
     int64_t i12 = 0;
     int64_t i13 = 0;
 
-    for (int64_t i03 = 0; i03 < ne03; i03++) {
-        for (int64_t i02 = 0; i02 < ne02; i02++) {
-            i10 += ne00 * ir0;
-            while (i10 >= ne0) {
-                i10 -= ne0;
-                if (++i11 == ne1) {
-                    i11 = 0;
-                    if (++i12 == ne2) {
-                        i12 = 0;
-                        if (++i13 == ne3) {
-                            i13 = 0;
+    if (dst->type == GGML_TYPE_F32) {
+        for (int64_t i03 = 0; i03 < ne03; i03++) {
+            for (int64_t i02 = 0; i02 < ne02; i02++) {
+                i10 += ne00 * ir0;
+                while (i10 >= ne0) {
+                    i10 -= ne0;
+                    if (++i11 == ne1) {
+                        i11 = 0;
+                        if (++i12 == ne2) {
+                            i12 = 0;
+                            if (++i13 == ne3) {
+                                i13 = 0;
+                            }
                         }
                     }
                 }
-            }
-            for (int64_t i01 = ir0; i01 < ir1; i01++) {
-                for (int64_t i00 = 0; i00 < ne00; i00++) {
-                    const char * src0_ptr = ((char *) src0->data + i00*nb00 + i01*nb01 + i02*nb02 + i03*nb03);
-                          char * dst_ptr  = ((char *)  dst->data + i10*nb0  + i11*nb1  + i12*nb2  + i13*nb3);
+                for (int64_t i01 = ir0; i01 < ir1; i01++) {
+                    for (int64_t i00 = 0; i00 < ne00; i00++) {
+                        const char * src0_ptr = ((char *) src0->data + i00*nb00 + i01*nb01 + i02*nb02 + i03*nb03);
+                              char * dst_ptr  = ((char *)  dst->data + i10*nb0  + i11*nb1  + i12*nb2  + i13*nb3);
 
-                    memcpy(dst_ptr, src0_ptr, type_size);
+                        memcpy(dst_ptr, src0_ptr, sizeof(float));
 
-                    if (++i10 == ne0) {
-                        i10 = 0;
-                        if (++i11 == ne1) {
-                            i11 = 0;
-                            if (++i12 == ne2) {
-                                i12 = 0;
-                                if (++i13 == ne3) {
-                                    i13 = 0;
+                        if (++i10 == ne0) {
+                            i10 = 0;
+                            if (++i11 == ne1) {
+                                i11 = 0;
+                                if (++i12 == ne2) {
+                                    i12 = 0;
+                                    if (++i13 == ne3) {
+                                        i13 = 0;
+                                    }
                                 }
                             }
                         }
                     }
                 }
-            }
-            i10 += ne00 * (ne01 - ir1);
-            while (i10 >= ne0) {
-                i10 -= ne0;
-                if (++i11 == ne1) {
-                    i11 = 0;
-                    if (++i12 == ne2) {
-                        i12 = 0;
-                        if (++i13 == ne3) {
-                            i13 = 0;
+                i10 += ne00 * (ne01 - ir1);
+                while (i10 >= ne0) {
+                    i10 -= ne0;
+                    if (++i11 == ne1) {
+                        i11 = 0;
+                        if (++i12 == ne2) {
+                            i12 = 0;
+                            if (++i13 == ne3) {
+                                i13 = 0;
+                            }
                         }
                     }
                 }
             }
         }
-    }
-}
-
+    } else if (dst->type == GGML_TYPE_F16) {
+        for (int64_t i03 = 0; i03 < ne03; i03++) {
+            for (int64_t i02 = 0; i02 < ne02; i02++) {
+                i10 += ne00 * ir0;
+                while (i10 >= ne0) {
+                    i10 -= ne0;
+                    if (++i11 == ne1) {
+                        i11 = 0;
+                        if (++i12 == ne2) {
+                            i12 = 0;
+                            if (++i13 == ne3) {
+                                i13 = 0;
+                            }
+                        }
+                    }
+                }
+                for (int64_t i01 = ir0; i01 < ir1; i01++) {
+                    for (int64_t i00 = 0; i00 < ne00; i00++) {
+                        const char * src0_ptr = ((char *) src0->data + i00*nb00 + i01*nb01 + i02*nb02 + i03*nb03);
+                              char * dst_ptr  = ((char *)  dst->data + i10*nb0  + i11*nb1  + i12*nb2  + i13*nb3);
+
+                        *(ggml_fp16_t *) dst_ptr = GGML_FP32_TO_FP16(*(const float *) src0_ptr);
+
+                        if (++i10 == ne0) {
+                            i10 = 0;
+                            if (++i11 == ne1) {
+                                i11 = 0;
+                                if (++i12 == ne2) {
+                                    i12 = 0;
+                                    if (++i13 == ne3) {
+                                        i13 = 0;
+                                    }
+                                }
+                            }
+                        }
+                    }
+                }
+                i10 += ne00 * (ne01 - ir1);
+                while (i10 >= ne0) {
+                    i10 -= ne0;
+                    if (++i11 == ne1) {
+                        i11 = 0;
+                        if (++i12 == ne2) {
+                            i12 = 0;
+                            if (++i13 == ne3) {
+                                i13 = 0;
+                            }
+                        }
+                    }
+                }
+            }
+        }
+    } else if (dst->type == GGML_TYPE_BF16) {
+        for (int64_t i03 = 0; i03 < ne03; i03++) {
+            for (int64_t i02 = 0; i02 < ne02; i02++) {
+                i10 += ne00 * ir0;
+                while (i10 >= ne0) {
+                    i10 -= ne0;
+                    if (++i11 == ne1) {
+                        i11 = 0;
+                        if (++i12 == ne2) {
+                            i12 = 0;
+                            if (++i13 == ne3) {
+                                i13 = 0;
+                            }
+                        }
+                    }
+                }
+                for (int64_t i01 = ir0; i01 < ir1; i01++) {
+                    for (int64_t i00 = 0; i00 < ne00; i00++) {
+                        const char * src0_ptr = ((char *) src0->data + i00*nb00 + i01*nb01 + i02*nb02 + i03*nb03);
+                              char * dst_ptr  = ((char *)  dst->data + i10*nb0  + i11*nb1  + i12*nb2  + i13*nb3);
+
+                        *(ggml_bf16_t *) dst_ptr = GGML_FP32_TO_BF16(*(const float *) src0_ptr);
+
+                        if (++i10 == ne0) {
+                            i10 = 0;
+                            if (++i11 == ne1) {
+                                i11 = 0;
+                                if (++i12 == ne2) {
+                                    i12 = 0;
+                                    if (++i13 == ne3) {
+                                        i13 = 0;
+                                    }
+                                }
+                            }
+                        }
+                    }
+                }
+                i10 += ne00 * (ne01 - ir1);
+                while (i10 >= ne0) {
+                    i10 -= ne0;
+                    if (++i11 == ne1) {
+                        i11 = 0;
+                        if (++i12 == ne2) {
+                            i12 = 0;
+                            if (++i13 == ne3) {
+                                i13 = 0;
+                            }
+                        }
+                    }
+                }
+            }
+        }
+    } else {
+        GGML_ABORT("fatal error"); // TODO: implement
+    }
+}
+
+// A simplified version of ggml_compute_forward_dup that doesn't do float upcasting, and just plain old memcpy.
+static void ggml_compute_forward_dup_bytes(
+        const struct ggml_compute_params * params,
+        struct ggml_tensor * dst) {
+
+    const struct ggml_tensor * src0 = dst->src[0];
+
+    GGML_ASSERT(ggml_nelements(dst) == ggml_nelements(src0));
+    GGML_ASSERT(src0->type == dst->type);
+
+    GGML_TENSOR_UNARY_OP_LOCALS;
+
+    if (ggml_is_contiguous(src0) && ggml_is_contiguous(dst)) {
+        ggml_compute_forward_dup_same_cont(params, dst);
+        return;
+    }
+
+    const size_t type_size = ggml_type_size(src0->type);
+    const int ith = params->ith; // thread index
+    const int nth = params->nth; // number of threads
+
+
+    // parallelize by rows
+    const int nr = ne01;
+    // number of rows per thread
+    const int dr = (nr + nth - 1) / nth;
+    // row range for this thread
+    const int ir0 = dr * ith;
+    const int ir1 = MIN(ir0 + dr, nr);
+
+    if (src0->type == dst->type &&
+        ne00 == ne0 &&
+        nb00 == type_size && nb0 == type_size) {
+        // copy by rows
+        const size_t rs = ne00 * type_size;
+        for (int64_t i03 = 0; i03 < ne03; i03++) {
+            for (int64_t i02 = 0; i02 < ne02; i02++) {
+                for (int64_t i01 = ir0; i01 < ir1; i01++) {
+                    memcpy(
+                        ((char *)  dst->data + i01*nb1  + i02*nb2  + i03*nb3),
+                        ((char *) src0->data + i01*nb01 + i02*nb02 + i03*nb03),
+                        rs);
+                }
+            }
+        }
+        return;
+    }
+
+    if (ggml_is_contiguous(dst)) {
+        size_t id = 0;
+        char * dst_ptr = (char *) dst->data;
+        const size_t rs = ne00 * type_size;
+
+        if (nb00 == type_size) {
+            // src0 is contigous on first dimension, copy by rows
+            for (int64_t i03 = 0; i03 < ne03; i03++) {
+                for (int64_t i02 = 0; i02 < ne02; i02++) {
+                    id += rs * ir0;
+                    for (int64_t i01 = ir0; i01 < ir1; i01++) {
+                        const char * src0_ptr = (char *) src0->data + i01*nb01 + i02*nb02 + i03*nb03;
+                        memcpy(dst_ptr + id, src0_ptr, rs);
+                        id += rs;
+                    }
+                    id += rs * (ne01 - ir1);
+                }
+            }
+        } else {
+            //printf("%s: this is not optimal - fix me\n", __func__);
+
+            for (int64_t i03 = 0; i03 < ne03; i03++) {
+                for (int64_t i02 = 0; i02 < ne02; i02++) {
+                    id += rs * ir0;
+                    for (int64_t i01 = ir0; i01 < ir1; i01++) {
+                        for (int64_t i00 = 0; i00 < ne00; i00++) {
+                            const char * src0_ptr = (char *) src0->data + i00*nb00 + i01*nb01 + i02*nb02 + i03*nb03;
+                            memcpy(dst_ptr + id, src0_ptr, type_size);
+
+                            id += type_size;
+                        }
+                    }
+                    id += rs * (ne01 - ir1);
+                }
+            }
+        }
+
+        return;
+    }
+
+    // dst counters
+
+    int64_t i10 = 0;
+    int64_t i11 = 0;
+    int64_t i12 = 0;
+    int64_t i13 = 0;
+
+    for (int64_t i03 = 0; i03 < ne03; i03++) {
+        for (int64_t i02 = 0; i02 < ne02; i02++) {
+            i10 += ne00 * ir0;
+            while (i10 >= ne0) {
+                i10 -= ne0;
+                if (++i11 == ne1) {
+                    i11 = 0;
+                    if (++i12 == ne2) {
+                        i12 = 0;
+                        if (++i13 == ne3) {
+                            i13 = 0;
+                        }
+                    }
+                }
+            }
+            for (int64_t i01 = ir0; i01 < ir1; i01++) {
+                for (int64_t i00 = 0; i00 < ne00; i00++) {
+                    const char * src0_ptr = ((char *) src0->data + i00*nb00 + i01*nb01 + i02*nb02 + i03*nb03);
+                          char * dst_ptr  = ((char *)  dst->data + i10*nb0  + i11*nb1  + i12*nb2  + i13*nb3);
+
+                    memcpy(dst_ptr, src0_ptr, type_size);
+
+                    if (++i10 == ne0) {
+                        i10 = 0;
+                        if (++i11 == ne1) {
+                            i11 = 0;
+                            if (++i12 == ne2) {
+                                i12 = 0;
+                                if (++i13 == ne3) {
+                                    i13 = 0;
+                                }
+                            }
+                        }
+                    }
+                }
+            }
+            i10 += ne00 * (ne01 - ir1);
+            while (i10 >= ne0) {
+                i10 -= ne0;
+                if (++i11 == ne1) {
+                    i11 = 0;
+                    if (++i12 == ne2) {
+                        i12 = 0;
+                        if (++i13 == ne3) {
+                            i13 = 0;
+                        }
+                    }
+                }
+            }
+        }
+    }
+}
+
 static void ggml_compute_forward_dup(
         const struct ggml_compute_params * params,
-        const struct ggml_tensor * src0,
         struct ggml_tensor * dst) {
+
+    const struct ggml_tensor * src0 = dst->src[0];
+
     if (src0->type == dst->type) {
-        ggml_compute_forward_dup_bytes(params, src0, dst);
+        ggml_compute_forward_dup_bytes(params, dst);
         return;
     }
 
     switch (src0->type) {
         case GGML_TYPE_F16:
             {
-                ggml_compute_forward_dup_f16(params, src0, dst);
+                ggml_compute_forward_dup_f16(params, dst);
+            } break;
+        case GGML_TYPE_BF16:
+            {
+                ggml_compute_forward_dup_bf16(params, dst);
             } break;
         case GGML_TYPE_F32:
             {
-                ggml_compute_forward_dup_f32(params, src0, dst);
+                ggml_compute_forward_dup_f32(params, dst);
             } break;
         default:
             {
-                GGML_ASSERT(false);
-            } break;
+                GGML_ABORT("fatal error");
+            }
     }
 }
 
@@ -7121,14 +9054,12 @@ static void ggml_compute_forward_dup(
 
 static void ggml_compute_forward_add_f32(
         const struct ggml_compute_params * params,
-        const struct ggml_tensor * src0,
-        const struct ggml_tensor * src1,
         struct ggml_tensor * dst) {
-    GGML_ASSERT(ggml_can_repeat(src1, src0) && ggml_are_same_shape(src0, dst));
 
-    if (params->type == GGML_TASK_INIT || params->type == GGML_TASK_FINALIZE) {
-        return;
-    }
+    const struct ggml_tensor * src0 = dst->src[0];
+    const struct ggml_tensor * src1 = dst->src[1];
+
+    GGML_ASSERT(ggml_can_repeat(src1, src0) && ggml_are_same_shape(src0, dst));
 
     const int ith = params->ith;
     const int nth = params->nth;
@@ -7198,14 +9129,12 @@ static void ggml_compute_forward_add_f32(
 
 static void ggml_compute_forward_add_f16_f32(
         const struct ggml_compute_params * params,
-        const struct ggml_tensor * src0,
-        const struct ggml_tensor * src1,
         struct ggml_tensor * dst) {
-    GGML_ASSERT(ggml_are_same_shape(src0, src1) && ggml_are_same_shape(src0, dst));
 
-    if (params->type == GGML_TASK_INIT || params->type == GGML_TASK_FINALIZE) {
-        return;
-    }
+    const struct ggml_tensor * src0 = dst->src[0];
+    const struct ggml_tensor * src1 = dst->src[1];
+
+    GGML_ASSERT(ggml_are_same_shape(src0, src1) && ggml_are_same_shape(src0, dst));
 
     const int ith = params->ith;
     const int nth = params->nth;
@@ -7247,7 +9176,82 @@ static void ggml_compute_forward_add_f16_f32(
                 float *       src1_ptr = (float *)       ((char *) src1->data + i3*nb13 + i2*nb12 + i1*nb11);
 
                 for (int i = 0; i < ne0; i++) {
-                    dst_ptr[i] = GGML_FP32_TO_FP16(GGML_FP16_TO_FP32(src0_ptr[i]) + src1_ptr[i]);
+                    dst_ptr[i] = GGML_FP32_TO_FP16(GGML_FP16_TO_FP32(src0_ptr[i]) + src1_ptr[i]);
+                }
+            }
+        } else {
+            for (int ir = ir0; ir < ir1; ++ir) {
+                // src0, src1 and dst are same shape => same indices
+                const int i3 = ir/(ne2*ne1);
+                const int i2 = (ir - i3*ne2*ne1)/ne1;
+                const int i1 = (ir - i3*ne2*ne1 - i2*ne1);
+
+                float *       dst_ptr  = (float *)       ((char *) dst->data  + i3*nb3  + i2*nb2  + i1*nb1);
+                ggml_fp16_t * src0_ptr = (ggml_fp16_t *) ((char *) src0->data + i3*nb03 + i2*nb02 + i1*nb01);
+                float *       src1_ptr = (float *)       ((char *) src1->data + i3*nb13 + i2*nb12 + i1*nb11);
+
+                for (int i = 0; i < ne0; i++) {
+                    dst_ptr[i] = GGML_FP16_TO_FP32(src0_ptr[i]) + src1_ptr[i];
+                }
+            }
+        }
+    }
+    else {
+        // src1 is not contiguous
+        GGML_ABORT("fatal error");
+    }
+}
+
+static void ggml_compute_forward_add_bf16_f32(
+        const struct ggml_compute_params * params,
+        struct ggml_tensor * dst) {
+
+    const struct ggml_tensor * src0 = dst->src[0];
+    const struct ggml_tensor * src1 = dst->src[1];
+
+    GGML_ASSERT(ggml_are_same_shape(src0, src1) && ggml_are_same_shape(src0, dst));
+
+    const int ith = params->ith;
+    const int nth = params->nth;
+
+    const int nr  = ggml_nrows(src0);
+
+    GGML_TENSOR_BINARY_OP_LOCALS
+
+    GGML_ASSERT(src0->type == GGML_TYPE_BF16);
+    GGML_ASSERT(src1->type == GGML_TYPE_F32);
+
+    if (dst->type == GGML_TYPE_F32) {
+        GGML_ASSERT( nb0 == sizeof(float));
+    }
+    else {
+        GGML_ASSERT(dst->type  == GGML_TYPE_BF16);
+        GGML_ASSERT( nb0 == sizeof(ggml_bf16_t));
+    }
+
+    GGML_ASSERT(nb00 == sizeof(ggml_bf16_t));
+
+    // rows per thread
+    const int dr = (nr + nth - 1)/nth;
+
+    // row range for this thread
+    const int ir0 = dr*ith;
+    const int ir1 = MIN(ir0 + dr, nr);
+
+    if (nb10 == sizeof(float)) {
+        if (dst->type == GGML_TYPE_BF16) {
+            for (int ir = ir0; ir < ir1; ++ir) {
+                // src0, src1 and dst are same shape => same indices
+                const int i3 = ir/(ne2*ne1);
+                const int i2 = (ir - i3*ne2*ne1)/ne1;
+                const int i1 = (ir - i3*ne2*ne1 - i2*ne1);
+
+                ggml_bf16_t * dst_ptr  = (ggml_bf16_t *) ((char *) dst->data  + i3*nb3  + i2*nb2  + i1*nb1);
+                ggml_bf16_t * src0_ptr = (ggml_bf16_t *) ((char *) src0->data + i3*nb03 + i2*nb02 + i1*nb01);
+                float *       src1_ptr = (float *)       ((char *) src1->data + i3*nb13 + i2*nb12 + i1*nb11);
+
+                for (int i = 0; i < ne0; i++) {
+                    dst_ptr[i] = GGML_FP32_TO_BF16(GGML_BF16_TO_FP32(src0_ptr[i]) + src1_ptr[i]);
                 }
             }
         } else {
@@ -7258,31 +9262,29 @@ static void ggml_compute_forward_add_f16_f32(
                 const int i1 = (ir - i3*ne2*ne1 - i2*ne1);
 
                 float *       dst_ptr  = (float *)       ((char *) dst->data  + i3*nb3  + i2*nb2  + i1*nb1);
-                ggml_fp16_t * src0_ptr = (ggml_fp16_t *) ((char *) src0->data + i3*nb03 + i2*nb02 + i1*nb01);
+                ggml_bf16_t * src0_ptr = (ggml_bf16_t *) ((char *) src0->data + i3*nb03 + i2*nb02 + i1*nb01);
                 float *       src1_ptr = (float *)       ((char *) src1->data + i3*nb13 + i2*nb12 + i1*nb11);
 
                 for (int i = 0; i < ne0; i++) {
-                    dst_ptr[i] = GGML_FP16_TO_FP32(src0_ptr[i]) + src1_ptr[i];
+                    dst_ptr[i] = GGML_BF16_TO_FP32(src0_ptr[i]) + src1_ptr[i];
                 }
             }
         }
     }
     else {
         // src1 is not contiguous
-        GGML_ASSERT(false);
+        GGML_ABORT("fatal error");
     }
 }
 
 static void ggml_compute_forward_add_f16_f16(
         const struct ggml_compute_params * params,
-        const struct ggml_tensor * src0,
-        const struct ggml_tensor * src1,
         struct ggml_tensor * dst) {
-    GGML_ASSERT(ggml_are_same_shape(src0, src1) && ggml_are_same_shape(src0, dst));
 
-    if (params->type == GGML_TASK_INIT || params->type == GGML_TASK_FINALIZE) {
-        return;
-    }
+    const struct ggml_tensor * src0 = dst->src[0];
+    const struct ggml_tensor * src1 = dst->src[1];
+
+    GGML_ASSERT(ggml_are_same_shape(src0, src1) && ggml_are_same_shape(src0, dst));
 
     const int ith = params->ith;
     const int nth = params->nth;
@@ -7323,20 +9325,70 @@ static void ggml_compute_forward_add_f16_f16(
     }
     else {
         // src1 is not contiguous
-        GGML_ASSERT(false);
+        GGML_ABORT("fatal error");
     }
 }
 
-static void ggml_compute_forward_add_q_f32(
+static void ggml_compute_forward_add_bf16_bf16(
         const struct ggml_compute_params * params,
-        const struct ggml_tensor * src0,
-        const struct ggml_tensor * src1,
         struct ggml_tensor * dst) {
+
+    const struct ggml_tensor * src0 = dst->src[0];
+    const struct ggml_tensor * src1 = dst->src[1];
+
     GGML_ASSERT(ggml_are_same_shape(src0, src1) && ggml_are_same_shape(src0, dst));
 
-    if (params->type == GGML_TASK_INIT || params->type == GGML_TASK_FINALIZE) {
-        return;
+    const int ith = params->ith;
+    const int nth = params->nth;
+
+    const int nr  = ggml_nrows(src0);
+
+    GGML_TENSOR_BINARY_OP_LOCALS
+
+    GGML_ASSERT(src0->type == GGML_TYPE_BF16);
+    GGML_ASSERT(src1->type == GGML_TYPE_BF16);
+    GGML_ASSERT(dst->type  == GGML_TYPE_BF16);
+
+    GGML_ASSERT( nb0 == sizeof(ggml_bf16_t));
+    GGML_ASSERT(nb00 == sizeof(ggml_bf16_t));
+
+    // rows per thread
+    const int dr = (nr + nth - 1)/nth;
+
+    // row range for this thread
+    const int ir0 = dr*ith;
+    const int ir1 = MIN(ir0 + dr, nr);
+
+    if (nb10 == sizeof(ggml_bf16_t)) {
+        for (int ir = ir0; ir < ir1; ++ir) {
+            // src0, src1 and dst are same shape => same indices
+            const int i3 = ir/(ne2*ne1);
+            const int i2 = (ir - i3*ne2*ne1)/ne1;
+            const int i1 = (ir - i3*ne2*ne1 - i2*ne1);
+
+            ggml_bf16_t * dst_ptr  = (ggml_bf16_t *) ((char *) dst->data  + i3*nb3  + i2*nb2  + i1*nb1);
+            ggml_bf16_t * src0_ptr = (ggml_bf16_t *) ((char *) src0->data + i3*nb03 + i2*nb02 + i1*nb01);
+            ggml_bf16_t * src1_ptr = (ggml_bf16_t *) ((char *) src1->data + i3*nb13 + i2*nb12 + i1*nb11);
+
+            for (int i = 0; i < ne0; i++) {
+                dst_ptr[i] = GGML_FP32_TO_BF16(GGML_BF16_TO_FP32(src0_ptr[i]) + GGML_BF16_TO_FP32(src1_ptr[i]));
+            }
+        }
+    }
+    else {
+        // src1 is not contiguous
+        GGML_ABORT("fatal error");
     }
+}
+
+static void ggml_compute_forward_add_q_f32(
+        const struct ggml_compute_params * params,
+        struct ggml_tensor * dst) {
+
+    const struct ggml_tensor * src0 = dst->src[0];
+    const struct ggml_tensor * src1 = dst->src[1];
+
+    GGML_ASSERT(ggml_are_same_shape(src0, src1) && ggml_are_same_shape(src0, dst));
 
     const int nr  = ggml_nrows(src0);
 
@@ -7407,24 +9459,43 @@ static void ggml_compute_forward_add_q_f32(
 
 static void ggml_compute_forward_add(
         const struct ggml_compute_params * params,
-        const struct ggml_tensor * src0,
-        const struct ggml_tensor * src1,
         struct ggml_tensor * dst) {
+
+    const struct ggml_tensor * src0 = dst->src[0];
+    const struct ggml_tensor * src1 = dst->src[1];
+
     switch (src0->type) {
         case GGML_TYPE_F32:
             {
-                ggml_compute_forward_add_f32(params, src0, src1, dst);
+                if (src1->type == GGML_TYPE_F32) {
+                    ggml_compute_forward_add_f32(params, dst);
+                }
+                else {
+                    GGML_ABORT("fatal error");
+                }
             } break;
         case GGML_TYPE_F16:
             {
                 if (src1->type == GGML_TYPE_F16) {
-                    ggml_compute_forward_add_f16_f16(params, src0, src1, dst);
+                    ggml_compute_forward_add_f16_f16(params, dst);
                 }
                 else if (src1->type == GGML_TYPE_F32) {
-                    ggml_compute_forward_add_f16_f32(params, src0, src1, dst);
+                    ggml_compute_forward_add_f16_f32(params, dst);
                 }
                 else {
-                    GGML_ASSERT(false);
+                    GGML_ABORT("fatal error");
+                }
+            } break;
+        case GGML_TYPE_BF16:
+            {
+                if (src1->type == GGML_TYPE_BF16) {
+                    ggml_compute_forward_add_bf16_bf16(params, dst);
+                }
+                else if (src1->type == GGML_TYPE_F32) {
+                    ggml_compute_forward_add_bf16_f32(params, dst);
+                }
+                else {
+                    GGML_ABORT("fatal error");
                 }
             } break;
         case GGML_TYPE_Q4_0:
@@ -7437,13 +9508,27 @@ static void ggml_compute_forward_add(
         case GGML_TYPE_Q4_K:
         case GGML_TYPE_Q5_K:
         case GGML_TYPE_Q6_K:
-            {
-                ggml_compute_forward_add_q_f32(params, src0, src1, dst);
+        case GGML_TYPE_TQ1_0:
+        case GGML_TYPE_TQ2_0:
+        case GGML_TYPE_IQ2_XXS:
+        case GGML_TYPE_IQ2_XS:
+        case GGML_TYPE_IQ3_XXS:
+        case GGML_TYPE_IQ1_S:
+        case GGML_TYPE_IQ1_M:
+        case GGML_TYPE_IQ4_NL:
+        case GGML_TYPE_IQ4_XS:
+        case GGML_TYPE_IQ3_S:
+        case GGML_TYPE_IQ2_S:
+        case GGML_TYPE_Q4_0_4_4:
+        case GGML_TYPE_Q4_0_4_8:
+        case GGML_TYPE_Q4_0_8_8:
+            {
+                ggml_compute_forward_add_q_f32(params, dst);
             } break;
         default:
             {
-                GGML_ASSERT(false);
-            } break;
+                GGML_ABORT("fatal error");
+            }
     }
 }
 
@@ -7451,16 +9536,14 @@ static void ggml_compute_forward_add(
 
 static void ggml_compute_forward_add1_f32(
         const struct ggml_compute_params * params,
-        const struct ggml_tensor * src0,
-        const struct ggml_tensor * src1,
         struct ggml_tensor * dst) {
+
+    const struct ggml_tensor * src0 = dst->src[0];
+    const struct ggml_tensor * src1 = dst->src[1];
+
     GGML_ASSERT(ggml_are_same_shape(src0, dst));
     GGML_ASSERT(ggml_is_scalar(src1));
 
-    if (params->type == GGML_TASK_INIT || params->type == GGML_TASK_FINALIZE) {
-        return;
-    }
-
     const int ith = params->ith;
     const int nth = params->nth;
 
@@ -7503,16 +9586,14 @@ static void ggml_compute_forward_add1_f32(
 
 static void ggml_compute_forward_add1_f16_f32(
         const struct ggml_compute_params * params,
-        const struct ggml_tensor * src0,
-        const struct ggml_tensor * src1,
         struct ggml_tensor * dst) {
+
+    const struct ggml_tensor * src0 = dst->src[0];
+    const struct ggml_tensor * src1 = dst->src[1];
+
     GGML_ASSERT(ggml_are_same_shape(src0, dst));
     GGML_ASSERT(ggml_is_scalar(src1));
 
-    if (params->type == GGML_TASK_INIT || params->type == GGML_TASK_FINALIZE) {
-        return;
-    }
-
     // scalar to add
     const float v = *(float *) src1->data;
 
@@ -7553,16 +9634,14 @@ static void ggml_compute_forward_add1_f16_f32(
 
 static void ggml_compute_forward_add1_f16_f16(
         const struct ggml_compute_params * params,
-        const struct ggml_tensor * src0,
-        const struct ggml_tensor * src1,
         struct ggml_tensor * dst) {
+
+    const struct ggml_tensor * src0 = dst->src[0];
+    const struct ggml_tensor * src1 = dst->src[1];
+
     GGML_ASSERT(ggml_are_same_shape(src0, dst));
     GGML_ASSERT(ggml_is_scalar(src1));
 
-    if (params->type == GGML_TASK_INIT || params->type == GGML_TASK_FINALIZE) {
-        return;
-    }
-
     // scalar to add
     const float v = GGML_FP16_TO_FP32(*(ggml_fp16_t *) src1->data);
 
@@ -7603,16 +9682,14 @@ static void ggml_compute_forward_add1_f16_f16(
 
 static void ggml_compute_forward_add1_q_f32(
         const struct ggml_compute_params * params,
-        const struct ggml_tensor * src0,
-        const struct ggml_tensor * src1,
         struct ggml_tensor * dst) {
+
+    const struct ggml_tensor * src0 = dst->src[0];
+    const struct ggml_tensor * src1 = dst->src[1];
+
     GGML_ASSERT(ggml_are_same_shape(src0, dst));
     GGML_ASSERT(ggml_is_scalar(src1));
 
-    if (params->type == GGML_TASK_INIT || params->type == GGML_TASK_FINALIZE) {
-        return;
-    }
-
     // scalar to add
     const float v = *(float *) src1->data;
 
@@ -7668,26 +9745,136 @@ static void ggml_compute_forward_add1_q_f32(
     }
 }
 
+static void ggml_compute_forward_add1_bf16_f32(
+        const struct ggml_compute_params * params,
+        struct ggml_tensor * dst) {
+
+    const struct ggml_tensor * src0 = dst->src[0];
+    const struct ggml_tensor * src1 = dst->src[1];
+
+    GGML_ASSERT(ggml_are_same_shape(src0, dst));
+    GGML_ASSERT(ggml_is_scalar(src1));
+
+    // scalar to add
+    const float v = *(float *) src1->data;
+
+    const int ith = params->ith;
+    const int nth = params->nth;
+
+    const int nr  = ggml_nrows(src0);
+
+    GGML_TENSOR_UNARY_OP_LOCALS
+
+    GGML_ASSERT(src0->type == GGML_TYPE_BF16);
+    GGML_ASSERT(src1->type == GGML_TYPE_F32);
+    GGML_ASSERT(dst->type  == GGML_TYPE_BF16);
+
+    GGML_ASSERT( nb0 == sizeof(ggml_bf16_t));
+    GGML_ASSERT(nb00 == sizeof(ggml_bf16_t));
+
+    // rows per thread
+    const int dr = (nr + nth - 1)/nth;
+
+    // row range for this thread
+    const int ir0 = dr*ith;
+    const int ir1 = MIN(ir0 + dr, nr);
+
+    for (int ir = ir0; ir < ir1; ++ir) {
+        // src0 and dst are same shape => same indices
+        const int i3 = ir/(ne2*ne1);
+        const int i2 = (ir - i3*ne2*ne1)/ne1;
+        const int i1 = (ir - i3*ne2*ne1 - i2*ne1);
+
+        ggml_bf16_t * dst_ptr  = (ggml_bf16_t *) ((char *) dst->data  + i3*nb3  + i2*nb2  + i1*nb1 );
+        ggml_bf16_t * src0_ptr = (ggml_bf16_t *) ((char *) src0->data + i3*nb03 + i2*nb02 + i1*nb01);
+        for (int i = 0; i < ne0; i++) {
+            dst_ptr[i] = GGML_FP32_TO_BF16(GGML_BF16_TO_FP32(src0_ptr[i]) + v);
+        }
+    }
+}
+
+static void ggml_compute_forward_add1_bf16_bf16(
+        const struct ggml_compute_params * params,
+        struct ggml_tensor * dst) {
+
+    const struct ggml_tensor * src0 = dst->src[0];
+    const struct ggml_tensor * src1 = dst->src[1];
+
+    GGML_ASSERT(ggml_are_same_shape(src0, dst));
+    GGML_ASSERT(ggml_is_scalar(src1));
+
+    // scalar to add
+    const float v = GGML_BF16_TO_FP32(*(ggml_bf16_t *) src1->data);
+
+    const int ith = params->ith;
+    const int nth = params->nth;
+
+    const int nr  = ggml_nrows(src0);
+
+    GGML_TENSOR_UNARY_OP_LOCALS
+
+    GGML_ASSERT(src0->type == GGML_TYPE_BF16);
+    GGML_ASSERT(src1->type == GGML_TYPE_BF16);
+    GGML_ASSERT(dst->type  == GGML_TYPE_BF16);
+
+    GGML_ASSERT( nb0 == sizeof(ggml_bf16_t));
+    GGML_ASSERT(nb00 == sizeof(ggml_bf16_t));
+
+    // rows per thread
+    const int dr = (nr + nth - 1)/nth;
+
+    // row range for this thread
+    const int ir0 = dr*ith;
+    const int ir1 = MIN(ir0 + dr, nr);
+
+    for (int ir = ir0; ir < ir1; ++ir) {
+        // src0 and dst are same shape => same indices
+        const int i3 = ir/(ne2*ne1);
+        const int i2 = (ir - i3*ne2*ne1)/ne1;
+        const int i1 = (ir - i3*ne2*ne1 - i2*ne1);
+
+        ggml_bf16_t * dst_ptr  = (ggml_bf16_t *) ((char *) dst->data  + i3*nb3  + i2*nb2  + i1*nb1 );
+        ggml_bf16_t * src0_ptr = (ggml_bf16_t *) ((char *) src0->data + i3*nb03 + i2*nb02 + i1*nb01);
+        for (int i = 0; i < ne0; i++) {
+            dst_ptr[i] = GGML_FP32_TO_BF16(GGML_BF16_TO_FP32(src0_ptr[i]) + v);
+        }
+    }
+}
+
 static void ggml_compute_forward_add1(
         const struct ggml_compute_params * params,
-        const struct ggml_tensor * src0,
-        const struct ggml_tensor * src1,
         struct ggml_tensor * dst) {
+
+    const struct ggml_tensor * src0 = dst->src[0];
+    const struct ggml_tensor * src1 = dst->src[1];
+
     switch (src0->type) {
         case GGML_TYPE_F32:
             {
-                ggml_compute_forward_add1_f32(params, src0, src1, dst);
+                ggml_compute_forward_add1_f32(params, dst);
             } break;
         case GGML_TYPE_F16:
             {
                 if (src1->type == GGML_TYPE_F16) {
-                    ggml_compute_forward_add1_f16_f16(params, src0, src1, dst);
+                    ggml_compute_forward_add1_f16_f16(params, dst);
                 }
                 else if (src1->type == GGML_TYPE_F32) {
-                    ggml_compute_forward_add1_f16_f32(params, src0, src1, dst);
+                    ggml_compute_forward_add1_f16_f32(params, dst);
                 }
                 else {
-                    GGML_ASSERT(false);
+                    GGML_ABORT("fatal error");
+                }
+            } break;
+        case GGML_TYPE_BF16:
+            {
+                if (src1->type == GGML_TYPE_BF16) {
+                    ggml_compute_forward_add1_bf16_bf16(params, dst);
+                }
+                else if (src1->type == GGML_TYPE_F32) {
+                    ggml_compute_forward_add1_bf16_f32(params, dst);
+                }
+                else {
+                    GGML_ABORT("fatal error");
                 }
             } break;
         case GGML_TYPE_Q4_0:
@@ -7701,13 +9888,27 @@ static void ggml_compute_forward_add1(
         case GGML_TYPE_Q4_K:
         case GGML_TYPE_Q5_K:
         case GGML_TYPE_Q6_K:
-            {
-                ggml_compute_forward_add1_q_f32(params, src0, src1, dst);
+        case GGML_TYPE_TQ1_0:
+        case GGML_TYPE_TQ2_0:
+        case GGML_TYPE_IQ2_XXS:
+        case GGML_TYPE_IQ2_XS:
+        case GGML_TYPE_IQ3_XXS:
+        case GGML_TYPE_IQ1_S:
+        case GGML_TYPE_IQ1_M:
+        case GGML_TYPE_IQ4_NL:
+        case GGML_TYPE_IQ4_XS:
+        case GGML_TYPE_IQ3_S:
+        case GGML_TYPE_IQ2_S:
+        case GGML_TYPE_Q4_0_4_4:
+        case GGML_TYPE_Q4_0_4_8:
+        case GGML_TYPE_Q4_0_8_8:
+            {
+                ggml_compute_forward_add1_q_f32(params, dst);
             } break;
         default:
             {
-                GGML_ASSERT(false);
-            } break;
+                GGML_ABORT("fatal error");
+            }
     }
 }
 
@@ -7715,9 +9916,11 @@ static void ggml_compute_forward_add1(
 
 static void ggml_compute_forward_acc_f32(
         const struct ggml_compute_params * params,
-        const struct ggml_tensor * src0,
-        const struct ggml_tensor * src1,
         struct ggml_tensor * dst) {
+
+    const struct ggml_tensor * src0 = dst->src[0];
+    const struct ggml_tensor * src1 = dst->src[1];
+
     GGML_ASSERT(ggml_are_same_shape(src0, dst));
     GGML_ASSERT(ggml_is_contiguous(dst) && ggml_is_contiguous(src0));
 
@@ -7729,17 +9932,16 @@ static void ggml_compute_forward_acc_f32(
     size_t offset  = ((int32_t *) dst->op_params)[3];
     bool   inplace = (bool) ((int32_t *) dst->op_params)[4];
 
-    if (!inplace && (params->type == GGML_TASK_INIT)) {
-        // memcpy needs to be synchronized across threads to avoid race conditions.
-        // => do it in INIT phase
-        memcpy(
-            ((char *)  dst->data),
-            ((char *) src0->data),
-            ggml_nbytes(dst));
-    }
-
-    if (params->type == GGML_TASK_INIT || params->type == GGML_TASK_FINALIZE) {
-        return;
+    if (!inplace) {
+        if (params->ith == 0) {
+            // memcpy needs to be synchronized across threads to avoid race conditions.
+            // => do it in INIT phase
+            memcpy(
+                ((char *)  dst->data),
+                ((char *) src0->data),
+                ggml_nbytes(dst));
+        }
+        ggml_barrier(params->threadpool);
     }
 
     const int ith = params->ith;
@@ -7794,16 +9996,17 @@ static void ggml_compute_forward_acc_f32(
 
 static void ggml_compute_forward_acc(
         const struct ggml_compute_params * params,
-        const struct ggml_tensor * src0,
-        const struct ggml_tensor * src1,
         struct ggml_tensor * dst) {
 
+    const struct ggml_tensor * src0 = dst->src[0];
+
     switch (src0->type) {
         case GGML_TYPE_F32:
             {
-                ggml_compute_forward_acc_f32(params, src0, src1, dst);
+                ggml_compute_forward_acc_f32(params, dst);
             } break;
         case GGML_TYPE_F16:
+        case GGML_TYPE_BF16:
         case GGML_TYPE_Q4_0:
         case GGML_TYPE_Q4_1:
         case GGML_TYPE_Q5_0:
@@ -7815,10 +10018,24 @@ static void ggml_compute_forward_acc(
         case GGML_TYPE_Q4_K:
         case GGML_TYPE_Q5_K:
         case GGML_TYPE_Q6_K:
+        case GGML_TYPE_TQ1_0:
+        case GGML_TYPE_TQ2_0:
+        case GGML_TYPE_IQ2_XXS:
+        case GGML_TYPE_IQ2_XS:
+        case GGML_TYPE_IQ3_XXS:
+        case GGML_TYPE_IQ1_S:
+        case GGML_TYPE_IQ1_M:
+        case GGML_TYPE_IQ4_NL:
+        case GGML_TYPE_IQ4_XS:
+        case GGML_TYPE_IQ3_S:
+        case GGML_TYPE_IQ2_S:
+        case GGML_TYPE_Q4_0_4_4:
+        case GGML_TYPE_Q4_0_4_8:
+        case GGML_TYPE_Q4_0_8_8:
         default:
             {
-                GGML_ASSERT(false);
-            } break;
+                GGML_ABORT("fatal error");
+            }
     }
 }
 
@@ -7826,15 +10043,15 @@ static void ggml_compute_forward_acc(
 
 static void ggml_compute_forward_sub_f32(
         const struct ggml_compute_params * params,
-        const struct ggml_tensor * src0,
-        const struct ggml_tensor * src1,
         struct ggml_tensor * dst) {
-    assert(params->ith == 0);
-    assert(ggml_are_same_shape(src0, src1) && ggml_are_same_shape(src0, dst));
 
-    if (params->type == GGML_TASK_INIT || params->type == GGML_TASK_FINALIZE) {
-        return;
-    }
+    const struct ggml_tensor * src0 = dst->src[0];
+    const struct ggml_tensor * src1 = dst->src[1];
+
+    assert(ggml_can_repeat(src1, src0) && ggml_are_same_shape(src0, dst));
+
+    const int ith = params->ith;
+    const int nth = params->nth;
 
     const int nr  = ggml_nrows(src0);
 
@@ -7843,40 +10060,55 @@ static void ggml_compute_forward_sub_f32(
     GGML_ASSERT( nb0 == sizeof(float));
     GGML_ASSERT(nb00 == sizeof(float));
 
+    // rows per thread
+    const int dr = (nr + nth - 1)/nth;
+
+    // row range for this thread
+    const int ir0 = dr*ith;
+    const int ir1 = MIN(ir0 + dr, nr);
+
     if (nb10 == sizeof(float)) {
-        for (int ir = 0; ir < nr; ++ir) {
-            // src0, src1 and dst are same shape => same indices
-            const int i3 = ir/(ne2*ne1);
-            const int i2 = (ir - i3*ne2*ne1)/ne1;
-            const int i1 = (ir - i3*ne2*ne1 - i2*ne1);
+        for (int ir = ir0; ir < ir1; ++ir) {
+            // src1 is broadcastable across src0 and dst in i1, i2, i3
+            const int64_t i03 = ir/(ne02*ne01);
+            const int64_t i02 = (ir - i03*ne02*ne01)/ne01;
+            const int64_t i01 = (ir - i03*ne02*ne01 - i02*ne01);
+
+            const int64_t i13 = i03 % ne13;
+            const int64_t i12 = i02 % ne12;
+            const int64_t i11 = i01 % ne11;
+            const int64_t nr0 = ne00 / ne10;
+
+            float * dst_ptr  = (float *) ((char *) dst->data  + i03*nb3  + i02*nb2  + i01*nb1 );
+            float * src0_ptr = (float *) ((char *) src0->data + i03*nb03 + i02*nb02 + i01*nb01);
+            float * src1_ptr = (float *) ((char *) src1->data + i13*nb13 + i12*nb12 + i11*nb11);
 
+            for (int64_t r = 0; r < nr0; ++r) {
 #ifdef GGML_USE_ACCELERATE
-            vDSP_vsub(
-                    (float *) ((char *) src1->data + i3*nb13 + i2*nb12 + i1*nb11), 1,
-                    (float *) ((char *) src0->data + i3*nb03 + i2*nb02 + i1*nb01), 1,
-                    (float *) ((char *) dst->data  + i3*nb3  + i2*nb2  + i1*nb1 ), 1,
-                    ne0);
+                vDSP_vsub(src1_ptr, 1, src0_ptr + r*ne10, 1, dst_ptr + r*ne10, 1, ne10);
 #else
-            ggml_vec_sub_f32(ne0,
-                    (float *) ((char *) dst->data  + i3*nb3  + i2*nb2  + i1*nb1 ),
-                    (float *) ((char *) src0->data + i3*nb03 + i2*nb02 + i1*nb01),
-                    (float *) ((char *) src1->data + i3*nb13 + i2*nb12 + i1*nb11));
+                ggml_vec_sub_f32(ne10, dst_ptr + r*ne10, src0_ptr + r*ne10, src1_ptr);
 #endif
-                // }
-            // }
+            }
         }
     } else {
         // src1 is not contiguous
-        for (int ir = 0; ir < nr; ++ir) {
-            // src0, src1 and dst are same shape => same indices
-            const int i3 = ir/(ne2*ne1);
-            const int i2 = (ir - i3*ne2*ne1)/ne1;
-            const int i1 = (ir - i3*ne2*ne1 - i2*ne1);
+        for (int ir = ir0; ir < ir1; ++ir) {
+            // src1 is broadcastable across src0 and dst in i1, i2, i3
+            const int64_t i03 = ir/(ne02*ne01);
+            const int64_t i02 = (ir - i03*ne02*ne01)/ne01;
+            const int64_t i01 = (ir - i03*ne02*ne01 - i02*ne01);
+
+            const int64_t i13 = i03 % ne13;
+            const int64_t i12 = i02 % ne12;
+            const int64_t i11 = i01 % ne11;
+
+            float * dst_ptr  = (float *) ((char *) dst->data  + i03*nb3  + i02*nb2  + i01*nb1 );
+            float * src0_ptr = (float *) ((char *) src0->data + i03*nb03 + i02*nb02 + i01*nb01);
 
-            float * dst_ptr  = (float *) ((char *) dst->data  + i3*nb3  + i2*nb2  + i1*nb1 );
-            float * src0_ptr = (float *) ((char *) src0->data + i3*nb03 + i2*nb02 + i1*nb01);
-            for (int i0 = 0; i0 < ne0; i0++) {
-                float * src1_ptr = (float *) ((char *) src1->data + i3*nb13 + i2*nb12 + i1*nb11 + i0*nb10);
+            for (int64_t i0 = 0; i0 < ne0; ++i0) {
+                const int64_t i10 = i0 % ne10;
+                float * src1_ptr = (float *) ((char *) src1->data + i13*nb13 + i12*nb12 + i11*nb11 + i10*nb10);
 
                 dst_ptr[i0] = src0_ptr[i0] - *src1_ptr;
             }
@@ -7886,18 +10118,19 @@ static void ggml_compute_forward_sub_f32(
 
 static void ggml_compute_forward_sub(
         const struct ggml_compute_params * params,
-        const struct ggml_tensor * src0,
-        const struct ggml_tensor * src1,
         struct ggml_tensor * dst) {
+
+    const struct ggml_tensor * src0 = dst->src[0];
+
     switch (src0->type) {
         case GGML_TYPE_F32:
             {
-                ggml_compute_forward_sub_f32(params, src0, src1, dst);
+                ggml_compute_forward_sub_f32(params, dst);
             } break;
         default:
             {
-                GGML_ASSERT(false);
-            } break;
+                GGML_ABORT("fatal error");
+            }
     }
 }
 
@@ -7905,28 +10138,16 @@ static void ggml_compute_forward_sub(
 
 static void ggml_compute_forward_mul_f32(
         const struct ggml_compute_params * params,
-        const struct ggml_tensor * src0,
-        const struct ggml_tensor * src1,
         struct ggml_tensor * dst) {
+
+    const struct ggml_tensor * src0 = dst->src[0];
+    const struct ggml_tensor * src1 = dst->src[1];
+
     GGML_ASSERT(ggml_can_repeat(src1, src0) && ggml_are_same_shape(src0, dst));
 
-    if (params->type == GGML_TASK_INIT || params->type == GGML_TASK_FINALIZE) {
-        return;
-    }
     const int ith = params->ith;
     const int nth = params->nth;
 
-#ifdef GGML_USE_CLBLAST
-    if (src1->backend == GGML_BACKEND_GPU) {
-        // TODO: OpenCL kernel support full broadcast
-        GGML_ASSERT(ggml_can_repeat_rows(src1, src0));
-        if (ith == 0) {
-            ggml_cl_mul(src0, src1, dst);
-        }
-        return;
-    }
-#endif
-
     const int64_t nr = ggml_nrows(src0);
 
     GGML_TENSOR_BINARY_OP_LOCALS
@@ -7988,20 +10209,22 @@ static void ggml_compute_forward_mul_f32(
 
 static void ggml_compute_forward_mul(
         const struct ggml_compute_params * params,
-        const struct ggml_tensor * src0,
-        const struct ggml_tensor * src1,
         struct ggml_tensor * dst) {
+
+    const struct ggml_tensor * src0 = dst->src[0];
+    const struct ggml_tensor * src1 = dst->src[1];
+
     GGML_ASSERT(src1->type == GGML_TYPE_F32 && "only f32 src1 supported for now");
 
     switch (src0->type) {
         case GGML_TYPE_F32:
             {
-                ggml_compute_forward_mul_f32(params, src0, src1, dst);
+                ggml_compute_forward_mul_f32(params, dst);
             } break;
         default:
             {
-                GGML_ASSERT(false);
-            } break;
+                GGML_ABORT("fatal error");
+            }
     }
 }
 
@@ -8009,14 +10232,12 @@ static void ggml_compute_forward_mul(
 
 static void ggml_compute_forward_div_f32(
         const struct ggml_compute_params * params,
-        const struct ggml_tensor * src0,
-        const struct ggml_tensor * src1,
         struct ggml_tensor * dst) {
-    GGML_ASSERT(ggml_can_repeat(src1, src0) && ggml_are_same_shape(src0, dst));
 
-    if (params->type == GGML_TASK_INIT || params->type == GGML_TASK_FINALIZE) {
-        return;
-    }
+    const struct ggml_tensor * src0 = dst->src[0];
+    const struct ggml_tensor * src1 = dst->src[1];
+
+    GGML_ASSERT(ggml_can_repeat(src1, src0) && ggml_are_same_shape(src0, dst));
 
     const int ith = params->ith;
     const int nth = params->nth;
@@ -8082,18 +10303,19 @@ static void ggml_compute_forward_div_f32(
 
 static void ggml_compute_forward_div(
         const struct ggml_compute_params * params,
-        const struct ggml_tensor * src0,
-        const struct ggml_tensor * src1,
         struct ggml_tensor * dst) {
+
+    const struct ggml_tensor * src0 = dst->src[0];
+
     switch (src0->type) {
         case GGML_TYPE_F32:
             {
-                ggml_compute_forward_div_f32(params, src0, src1, dst);
+                ggml_compute_forward_div_f32(params, dst);
             } break;
         default:
             {
-                GGML_ASSERT(false);
-            } break;
+                GGML_ABORT("fatal error");
+            }
     }
 }
 
@@ -8101,15 +10323,16 @@ static void ggml_compute_forward_div(
 
 static void ggml_compute_forward_sqr_f32(
         const struct ggml_compute_params * params,
-        const struct ggml_tensor * src0,
         struct ggml_tensor * dst) {
-    assert(params->ith == 0);
-    assert(ggml_are_same_shape(src0, dst));
 
-    if (params->type == GGML_TASK_INIT || params->type == GGML_TASK_FINALIZE) {
+    const struct ggml_tensor * src0 = dst->src[0];
+
+    if (params->ith != 0) {
         return;
     }
 
+    assert(ggml_are_same_shape(src0, dst));
+
     const int n     = ggml_nrows(src0);
     const int nc    = src0->ne[0];
 
@@ -8125,17 +10348,19 @@ static void ggml_compute_forward_sqr_f32(
 
 static void ggml_compute_forward_sqr(
         const struct ggml_compute_params * params,
-        const struct ggml_tensor * src0,
         struct ggml_tensor * dst) {
+
+    const struct ggml_tensor * src0 = dst->src[0];
+
     switch (src0->type) {
         case GGML_TYPE_F32:
             {
-                ggml_compute_forward_sqr_f32(params, src0, dst);
+                ggml_compute_forward_sqr_f32(params, dst);
             } break;
         default:
             {
-                GGML_ASSERT(false);
-            } break;
+                GGML_ABORT("fatal error");
+            }
     }
 }
 
@@ -8143,15 +10368,16 @@ static void ggml_compute_forward_sqr(
 
 static void ggml_compute_forward_sqrt_f32(
         const struct ggml_compute_params * params,
-        const struct ggml_tensor * src0,
         struct ggml_tensor * dst) {
-    assert(params->ith == 0);
-    assert(ggml_are_same_shape(src0, dst));
 
-    if (params->type == GGML_TASK_INIT || params->type == GGML_TASK_FINALIZE) {
+    const struct ggml_tensor * src0 = dst->src[0];
+
+    if (params->ith != 0) {
         return;
     }
 
+    assert(ggml_are_same_shape(src0, dst));
+
     const int n  = ggml_nrows(src0);
     const int nc = src0->ne[0];
 
@@ -8167,17 +10393,19 @@ static void ggml_compute_forward_sqrt_f32(
 
 static void ggml_compute_forward_sqrt(
         const struct ggml_compute_params * params,
-        const struct ggml_tensor * src0,
         struct ggml_tensor * dst) {
+
+    const struct ggml_tensor * src0 = dst->src[0];
+
     switch (src0->type) {
         case GGML_TYPE_F32:
             {
-                ggml_compute_forward_sqrt_f32(params, src0, dst);
+                ggml_compute_forward_sqrt_f32(params, dst);
             } break;
         default:
             {
-                GGML_ASSERT(false);
-            } break;
+                GGML_ABORT("fatal error");
+            }
     }
 }
 
@@ -8185,15 +10413,16 @@ static void ggml_compute_forward_sqrt(
 
 static void ggml_compute_forward_log_f32(
         const struct ggml_compute_params * params,
-        const struct ggml_tensor * src0,
         struct ggml_tensor * dst) {
-    GGML_ASSERT(params->ith == 0);
-    GGML_ASSERT(ggml_are_same_shape(src0, dst));
 
-    if (params->type == GGML_TASK_INIT || params->type == GGML_TASK_FINALIZE) {
+    const struct ggml_tensor * src0 = dst->src[0];
+
+    if (params->ith != 0) {
         return;
     }
 
+    GGML_ASSERT(ggml_are_same_shape(src0, dst));
+
     const int n  = ggml_nrows(src0);
     const int nc = src0->ne[0];
 
@@ -8209,17 +10438,109 @@ static void ggml_compute_forward_log_f32(
 
 static void ggml_compute_forward_log(
         const struct ggml_compute_params * params,
-        const struct ggml_tensor * src0,
         struct ggml_tensor * dst) {
+
+    const struct ggml_tensor * src0 = dst->src[0];
+
+    switch (src0->type) {
+        case GGML_TYPE_F32:
+            {
+                ggml_compute_forward_log_f32(params, dst);
+            } break;
+        default:
+            {
+                GGML_ABORT("fatal error");
+            }
+    }
+}
+
+// ggml_compute_forward_sin
+
+static void ggml_compute_forward_sin_f32(
+        const struct ggml_compute_params * params,
+        struct ggml_tensor * dst) {
+
+    const struct ggml_tensor * src0 = dst->src[0];
+
+    if (params->ith != 0) {
+        return;
+    }
+
+    GGML_ASSERT(ggml_are_same_shape(src0, dst));
+
+    const int n  = ggml_nrows(src0);
+    const int nc = src0->ne[0];
+
+    GGML_ASSERT( dst->nb[0] == sizeof(float));
+    GGML_ASSERT(src0->nb[0] == sizeof(float));
+
+    for (int i = 0; i < n; i++) {
+        ggml_vec_sin_f32(nc,
+                (float *) ((char *) dst->data  + i*( dst->nb[1])),
+                (float *) ((char *) src0->data + i*(src0->nb[1])));
+    }
+}
+
+static void ggml_compute_forward_sin(
+        const struct ggml_compute_params * params,
+        struct ggml_tensor * dst) {
+
+    const struct ggml_tensor * src0 = dst->src[0];
+
     switch (src0->type) {
         case GGML_TYPE_F32:
             {
-                ggml_compute_forward_log_f32(params, src0, dst);
+                ggml_compute_forward_sin_f32(params, dst);
             } break;
         default:
             {
-                GGML_ASSERT(false);
+                GGML_ABORT("fatal error");
+            }
+    }
+}
+
+// ggml_compute_forward_cos
+
+static void ggml_compute_forward_cos_f32(
+        const struct ggml_compute_params * params,
+        struct ggml_tensor * dst) {
+
+    const struct ggml_tensor * src0 = dst->src[0];
+
+    if (params->ith != 0) {
+        return;
+    }
+
+    GGML_ASSERT(ggml_are_same_shape(src0, dst));
+
+    const int n  = ggml_nrows(src0);
+    const int nc = src0->ne[0];
+
+    GGML_ASSERT( dst->nb[0] == sizeof(float));
+    GGML_ASSERT(src0->nb[0] == sizeof(float));
+
+    for (int i = 0; i < n; i++) {
+        ggml_vec_cos_f32(nc,
+                (float *) ((char *) dst->data  + i*( dst->nb[1])),
+                (float *) ((char *) src0->data + i*(src0->nb[1])));
+    }
+}
+
+static void ggml_compute_forward_cos(
+        const struct ggml_compute_params * params,
+        struct ggml_tensor * dst) {
+
+    const struct ggml_tensor * src0 = dst->src[0];
+
+    switch (src0->type) {
+        case GGML_TYPE_F32:
+            {
+                ggml_compute_forward_cos_f32(params, dst);
             } break;
+        default:
+            {
+                GGML_ABORT("fatal error");
+            }
     }
 }
 
@@ -8227,12 +10548,11 @@ static void ggml_compute_forward_log(
 
 static void ggml_compute_forward_sum_f32(
         const struct ggml_compute_params * params,
-        const struct ggml_tensor * src0,
         struct ggml_tensor * dst) {
-    assert(params->ith == 0);
-    assert(ggml_is_scalar(dst));
 
-    if (params->type == GGML_TASK_INIT || params->type == GGML_TASK_FINALIZE) {
+    const struct ggml_tensor * src0 = dst->src[0];
+
+    if (params->ith != 0) {
         return;
     }
 
@@ -8260,15 +10580,16 @@ static void ggml_compute_forward_sum_f32(
 
 static void ggml_compute_forward_sum_f16(
     const struct ggml_compute_params * params,
-    const struct ggml_tensor * src0,
           struct ggml_tensor * dst) {
-    assert(params->ith == 0);
-    assert(ggml_is_scalar(dst));
 
-    if (params->type == GGML_TASK_INIT || params->type == GGML_TASK_FINALIZE) {
+    const struct ggml_tensor * src0 = dst->src[0];
+
+    if (params->ith != 0) {
         return;
     }
 
+    assert(ggml_is_scalar(dst));
+
     assert(src0->nb[0] == sizeof(ggml_fp16_t));
 
     GGML_TENSOR_LOCALS(int64_t, ne0, src0, ne)
@@ -8290,23 +10611,62 @@ static void ggml_compute_forward_sum_f16(
     ((ggml_fp16_t *) dst->data)[0] = GGML_FP32_TO_FP16(sum);
 }
 
+static void ggml_compute_forward_sum_bf16(
+    const struct ggml_compute_params * params,
+          struct ggml_tensor * dst) {
+
+    const struct ggml_tensor * src0 = dst->src[0];
+
+    if (params->ith != 0) {
+        return;
+    }
+
+    assert(ggml_is_scalar(dst));
+
+    assert(src0->nb[0] == sizeof(ggml_bf16_t));
+
+    GGML_TENSOR_LOCALS(int64_t, ne0, src0, ne)
+    GGML_TENSOR_LOCALS(size_t,  nb0, src0, nb)
+
+    float sum = 0;
+    float row_sum = 0;
+
+    for (int64_t i03 = 0; i03 < ne03; i03++) {
+        for (int64_t i02 = 0; i02 < ne02; i02++) {
+            for (int64_t i01 = 0; i01 < ne01; i01++) {
+                ggml_vec_sum_bf16_ggf(ne00,
+                    &row_sum,
+                    (ggml_bf16_t *) ((char *) src0->data + i01 * nb01 + i02 * nb02 + i03 * nb03));
+                sum += row_sum;
+            }
+        }
+    }
+    ((ggml_bf16_t *) dst->data)[0] = GGML_FP32_TO_BF16(sum);
+}
+
 static void ggml_compute_forward_sum(
         const struct ggml_compute_params * params,
-        const struct ggml_tensor * src0,
         struct ggml_tensor * dst) {
+
+    const struct ggml_tensor * src0 = dst->src[0];
+
     switch (src0->type) {
         case GGML_TYPE_F32:
             {
-                ggml_compute_forward_sum_f32(params, src0, dst);
+                ggml_compute_forward_sum_f32(params, dst);
             } break;
         case GGML_TYPE_F16:
             {
-                ggml_compute_forward_sum_f16(params, src0, dst);
+                ggml_compute_forward_sum_f16(params, dst);
             } break;
-        default:
+        case GGML_TYPE_BF16:
             {
-                GGML_ASSERT(false);
+                ggml_compute_forward_sum_bf16(params, dst);
             } break;
+        default:
+            {
+                GGML_ABORT("fatal error");
+            }
     }
 }
 
@@ -8314,11 +10674,11 @@ static void ggml_compute_forward_sum(
 
 static void ggml_compute_forward_sum_rows_f32(
         const struct ggml_compute_params * params,
-        const struct ggml_tensor * src0,
         struct ggml_tensor * dst) {
-    GGML_ASSERT(params->ith == 0);
 
-    if (params->type == GGML_TASK_INIT || params->type == GGML_TASK_FINALIZE) {
+    const struct ggml_tensor * src0 = dst->src[0];
+
+    if (params->ith != 0) {
         return;
     }
 
@@ -8347,17 +10707,19 @@ static void ggml_compute_forward_sum_rows_f32(
 
 static void ggml_compute_forward_sum_rows(
         const struct ggml_compute_params * params,
-        const struct ggml_tensor * src0,
         struct ggml_tensor * dst) {
+
+    const struct ggml_tensor * src0 = dst->src[0];
+
     switch (src0->type) {
         case GGML_TYPE_F32:
             {
-                ggml_compute_forward_sum_rows_f32(params, src0, dst);
+                ggml_compute_forward_sum_rows_f32(params, dst);
             } break;
         default:
             {
-                GGML_ASSERT(false);
-            } break;
+                GGML_ABORT("fatal error");
+            }
     }
 }
 
@@ -8365,11 +10727,11 @@ static void ggml_compute_forward_sum_rows(
 
 static void ggml_compute_forward_mean_f32(
         const struct ggml_compute_params * params,
-        const struct ggml_tensor * src0,
         struct ggml_tensor * dst) {
-    assert(params->ith == 0);
 
-    if (params->type == GGML_TASK_INIT || params->type == GGML_TASK_FINALIZE) {
+    const struct ggml_tensor * src0 = dst->src[0];
+
+    if (params->ith != 0) {
         return;
     }
 
@@ -8402,17 +10764,19 @@ static void ggml_compute_forward_mean_f32(
 
 static void ggml_compute_forward_mean(
         const struct ggml_compute_params * params,
-        const struct ggml_tensor * src0,
         struct ggml_tensor * dst) {
+
+    const struct ggml_tensor * src0 = dst->src[0];
+
     switch (src0->type) {
         case GGML_TYPE_F32:
             {
-                ggml_compute_forward_mean_f32(params, src0, dst);
+                ggml_compute_forward_mean_f32(params, dst);
             } break;
         default:
             {
-                GGML_ASSERT(false);
-            } break;
+                GGML_ABORT("fatal error");
+            }
     }
 }
 
@@ -8420,11 +10784,11 @@ static void ggml_compute_forward_mean(
 
 static void ggml_compute_forward_argmax_f32(
         const struct ggml_compute_params * params,
-        const struct ggml_tensor * src0,
         struct ggml_tensor * dst) {
-    assert(params->ith == 0);
 
-    if (params->type == GGML_TASK_INIT || params->type == GGML_TASK_FINALIZE) {
+    const struct ggml_tensor * src0 = dst->src[0];
+
+    if (params->ith != 0) {
         return;
     }
 
@@ -8448,17 +10812,99 @@ static void ggml_compute_forward_argmax_f32(
 
 static void ggml_compute_forward_argmax(
         const struct ggml_compute_params * params,
-        const struct ggml_tensor * src0,
         struct ggml_tensor * dst) {
+
+    const struct ggml_tensor * src0 = dst->src[0];
+
     switch (src0->type) {
         case GGML_TYPE_F32:
             {
-                ggml_compute_forward_argmax_f32(params, src0, dst);
+                ggml_compute_forward_argmax_f32(params, dst);
             } break;
         default:
             {
-                GGML_ASSERT(false);
+                GGML_ABORT("fatal error");
+            }
+    }
+}
+
+// ggml_compute_forward_count_equal
+
+static void ggml_compute_forward_count_equal_i32(
+        const struct ggml_compute_params * params,
+        struct ggml_tensor * dst) {
+
+    const struct ggml_tensor * src0 = dst->src[0];
+    const struct ggml_tensor * src1 = dst->src[1];
+
+    GGML_TENSOR_BINARY_OP_LOCALS;
+
+    GGML_ASSERT(src0->type == GGML_TYPE_I32);
+    GGML_ASSERT(src1->type == GGML_TYPE_I32);
+    GGML_ASSERT(ggml_are_same_shape(src0, src1));
+    GGML_ASSERT(ggml_is_scalar(dst));
+    GGML_ASSERT(dst->type == GGML_TYPE_I64);
+
+    const int64_t nr = ggml_nrows(src0);
+
+    const int ith = params->ith;
+    const int nth = params->nth;
+
+    int64_t * sums = (int64_t *) params->wdata;
+    int64_t sum_thread = 0;
+
+    // rows per thread
+    const int64_t dr = (nr + nth - 1)/nth;
+
+    // row range for this thread
+    const int64_t ir0 = dr*ith;
+    const int64_t ir1 = MIN(ir0 + dr, nr);
+
+    for (int64_t ir = ir0; ir < ir1; ++ir) {
+        const int64_t i03 =  ir                        / (ne02*ne01);
+        const int64_t i02 = (ir - i03*ne03)            /       ne01;
+        const int64_t i01 =  ir - i03*ne03 - i02*ne02;
+
+        const char * data0 = (const char *) src0->data + i03*nb03 + i02*nb02 + i01*nb01;
+        const char * data1 = (const char *) src1->data + i03*nb13 + i02*nb12 + i01*nb11;
+
+        for (int64_t i00 = 0; i00 < ne00; ++i00) {
+            const int32_t val0 = *((const int32_t *) (data0 + i00*nb00));
+            const int32_t val1 = *((const int32_t *) (data1 + i00*nb10));
+
+            sum_thread += val0 == val1;
+        }
+    }
+    if (ith != 0) {
+        sums[ith] = sum_thread;
+    }
+    ggml_barrier(params->threadpool);
+
+    if (ith != 0) {
+        return;
+    }
+
+    for (int ith_other = 1; ith_other < nth; ++ith_other) {
+        sum_thread += sums[ith_other];
+    }
+    *((int64_t *) dst->data) = sum_thread;
+}
+
+static void ggml_compute_forward_count_equal(
+        const struct ggml_compute_params * params,
+        struct ggml_tensor * dst) {
+
+    const struct ggml_tensor * src0 = dst->src[0];
+
+    switch (src0->type) {
+        case GGML_TYPE_I32:
+            {
+                ggml_compute_forward_count_equal_i32(params, dst);
             } break;
+        default:
+            {
+                GGML_ABORT("fatal error");
+            }
     }
 }
 
@@ -8466,15 +10912,16 @@ static void ggml_compute_forward_argmax(
 
 static void ggml_compute_forward_repeat_f32(
         const struct ggml_compute_params * params,
-        const struct ggml_tensor * src0,
         struct ggml_tensor * dst) {
-    GGML_ASSERT(params->ith == 0);
-    GGML_ASSERT(ggml_can_repeat(src0, dst));
 
-    if (params->type == GGML_TASK_INIT || params->type == GGML_TASK_FINALIZE) {
+    const struct ggml_tensor * src0 = dst->src[0];
+
+    if (params->ith != 0) {
         return;
     }
 
+    GGML_ASSERT(ggml_can_repeat(src0, dst));
+
     GGML_TENSOR_UNARY_OP_LOCALS
 
     // guaranteed to be an integer due to the check in ggml_can_repeat
@@ -8509,15 +10956,16 @@ static void ggml_compute_forward_repeat_f32(
 
 static void ggml_compute_forward_repeat_f16(
         const struct ggml_compute_params * params,
-        const struct ggml_tensor * src0,
         struct ggml_tensor * dst) {
-    GGML_ASSERT(params->ith == 0);
-    GGML_ASSERT(ggml_can_repeat(src0, dst));
 
-    if (params->type == GGML_TASK_INIT || params->type == GGML_TASK_FINALIZE) {
+    const struct ggml_tensor * src0 = dst->src[0];
+
+    if (params->ith != 0) {
         return;
     }
 
+    GGML_ASSERT(ggml_can_repeat(src0, dst));
+
     GGML_TENSOR_UNARY_OP_LOCALS
 
     // guaranteed to be an integer due to the check in ggml_can_repeat
@@ -8555,23 +11003,26 @@ static void ggml_compute_forward_repeat_f16(
 
 static void ggml_compute_forward_repeat(
         const struct ggml_compute_params * params,
-        const struct ggml_tensor * src0,
         struct ggml_tensor * dst) {
+
+    const struct ggml_tensor * src0 = dst->src[0];
+
     switch (src0->type) {
         case GGML_TYPE_F16:
+        case GGML_TYPE_BF16:
         case GGML_TYPE_I16:
             {
-                ggml_compute_forward_repeat_f16(params, src0, dst);
+                ggml_compute_forward_repeat_f16(params, dst);
             } break;
         case GGML_TYPE_F32:
         case GGML_TYPE_I32:
             {
-                ggml_compute_forward_repeat_f32(params, src0, dst);
+                ggml_compute_forward_repeat_f32(params, dst);
             } break;
         default:
             {
-                GGML_ASSERT(false);
-            } break;
+                GGML_ABORT("fatal error");
+            }
     }
 }
 
@@ -8579,15 +11030,16 @@ static void ggml_compute_forward_repeat(
 
 static void ggml_compute_forward_repeat_back_f32(
         const struct ggml_compute_params * params,
-        const struct ggml_tensor * src0,
         struct ggml_tensor * dst) {
-    GGML_ASSERT(params->ith == 0);
-    GGML_ASSERT(ggml_can_repeat(dst, src0));
 
-    if (params->type == GGML_TASK_INIT || params->type == GGML_TASK_FINALIZE) {
+    const struct ggml_tensor * src0 = dst->src[0];
+
+    if (params->ith != 0) {
         return;
     }
 
+    GGML_ASSERT(ggml_can_repeat(dst, src0));
+
     GGML_TENSOR_UNARY_OP_LOCALS
 
     // guaranteed to be an integer due to the check in ggml_can_repeat
@@ -8636,17 +11088,19 @@ static void ggml_compute_forward_repeat_back_f32(
 
 static void ggml_compute_forward_repeat_back(
         const struct ggml_compute_params * params,
-        const struct ggml_tensor * src0,
         struct ggml_tensor * dst) {
+
+    const struct ggml_tensor * src0 = dst->src[0];
+
     switch (src0->type) {
         case GGML_TYPE_F32:
             {
-                ggml_compute_forward_repeat_back_f32(params, src0, dst);
+                ggml_compute_forward_repeat_back_f32(params, dst);
             } break;
         default:
             {
-                GGML_ASSERT(false);
-            } break;
+                GGML_ABORT("fatal error");
+            }
     }
 }
 
@@ -8654,13 +11108,10 @@ static void ggml_compute_forward_repeat_back(
 
 static void ggml_compute_forward_concat_f32(
     const struct ggml_compute_params * params,
-    const struct ggml_tensor * src0,
-    const struct ggml_tensor * src1,
     struct ggml_tensor * dst) {
 
-    if (params->type == GGML_TASK_INIT || params->type == GGML_TASK_FINALIZE) {
-        return;
-    }
+    const struct ggml_tensor * src0 = dst->src[0];
+    const struct ggml_tensor * src1 = dst->src[1];
 
     GGML_ASSERT(src0->nb[0] == sizeof(float));
 
@@ -8669,31 +11120,29 @@ static void ggml_compute_forward_concat_f32(
 
     GGML_TENSOR_BINARY_OP_LOCALS
 
-    // TODO: support for transposed / permuted tensors
-    GGML_ASSERT(nb0  == sizeof(float));
-    GGML_ASSERT(nb00 == sizeof(float));
-    GGML_ASSERT(nb10 == sizeof(float));
+    const int32_t dim = ggml_get_op_params_i32(dst, 0);
+
+    GGML_ASSERT(dim >= 0 && dim < 4);
 
+    int64_t o[4] = {0, 0, 0, 0};
+    o[dim] = src0->ne[dim];
+
+    const float * x;
+
+    // TODO: smarter multi-theading
     for (int i3 = 0; i3 < ne3; i3++) {
         for (int i2 = ith; i2 < ne2; i2 += nth) {
-            if (i2 < ne02) { // src0
-                for (int i1 = 0; i1 < ne1; i1++) {
-                    for (int i0 = 0; i0 < ne0; i0++) {
-                        const float * x = (float *)((char *) src0->data + i0 * nb00 + i1 * nb01 + i2 * nb02 + i3 * nb03);
-
-                        float * y = (float *)((char *)dst->data + i0 * nb0 + i1 * nb1 + i2 * nb2 + i3 * nb3);
-                        *y = *x;
-                    }
-                }
-            } // src1
-            else {
-                for (int i1 = 0; i1 < ne1; i1++) {
-                    for (int i0 = 0; i0 < ne0; i0++) {
-                        const float * x = (float *)((char *) src1->data + i0 * nb10 + i1 * nb11 + (i2 - ne02) * nb12 + i3 * nb13);
-
-                        float * y = (float *)((char *)dst->data + i0 * nb0 + i1 * nb1 + i2 * nb2 + i3 * nb3);
-                        *y = *x;
+            for (int i1 = 0; i1 < ne1; i1++) {
+                for (int i0 = 0; i0 < ne0; i0++) {
+                    if (i0 < ne00 && i1 < ne01 && i2 < ne02 && i3 < ne03) {
+                        x = (const float *) ((const char *)src0->data + (i0       )*nb00 + (i1       )*nb01 + (i2       )*nb02 + (i3       )*nb03);
+                    } else {
+                        x = (const float *) ((const char *)src1->data + (i0 - o[0])*nb10 + (i1 - o[1])*nb11 + (i2 - o[2])*nb12 + (i3 - o[3])*nb13);
                     }
+
+                    float * y = (float *)((char *)dst->data + i0*nb0 + i1*nb1 + i2*nb2 + i3*nb3);
+
+                    *y = *x;
                 }
             }
         }
@@ -8701,20 +11150,21 @@ static void ggml_compute_forward_concat_f32(
 }
 
 static void ggml_compute_forward_concat(
-    const struct ggml_compute_params* params,
-    const struct ggml_tensor* src0,
-    const struct ggml_tensor* src1,
-    struct ggml_tensor* dst) {
+    const struct ggml_compute_params * params,
+    struct ggml_tensor * dst) {
+
+    const struct ggml_tensor * src0 = dst->src[0];
+
     switch (src0->type) {
         case GGML_TYPE_F32:
         case GGML_TYPE_I32:
             {
-                ggml_compute_forward_concat_f32(params, src0, src1, dst);
+                ggml_compute_forward_concat_f32(params, dst);
             } break;
         default:
             {
-                GGML_ASSERT(false);
-            } break;
+                GGML_ABORT("fatal error");
+            }
     }
 }
 
@@ -8722,21 +11172,21 @@ static void ggml_compute_forward_concat(
 
 static void ggml_compute_forward_abs_f32(
         const struct ggml_compute_params * params,
-        const struct ggml_tensor * src0,
         struct ggml_tensor * dst) {
-    assert(params->ith == 0);
-    assert(ggml_are_same_shape(src0, dst));
 
-    if (params->type == GGML_TASK_INIT || params->type == GGML_TASK_FINALIZE) {
+    const struct ggml_tensor * src0 = dst->src[0];
+
+    if (params->ith != 0) {
         return;
     }
 
+    assert(ggml_is_contiguous_1(src0));
+    assert(ggml_is_contiguous_1(dst));
+    assert(ggml_are_same_shape(src0, dst));
+
     const int n  = ggml_nrows(src0);
     const int nc = src0->ne[0];
 
-    assert(dst->nb[0]  == sizeof(float));
-    assert(src0->nb[0] == sizeof(float));
-
     for (int i = 0; i < n; i++) {
         ggml_vec_abs_f32(nc,
                 (float *) ((char *) dst->data  + i*( dst->nb[1])),
@@ -8746,17 +11196,19 @@ static void ggml_compute_forward_abs_f32(
 
 static void ggml_compute_forward_abs(
         const struct ggml_compute_params * params,
-        const struct ggml_tensor * src0,
         struct ggml_tensor * dst) {
+
+    const struct ggml_tensor * src0 = dst->src[0];
+
     switch (src0->type) {
         case GGML_TYPE_F32:
             {
-                ggml_compute_forward_abs_f32(params, src0, dst);
+                ggml_compute_forward_abs_f32(params, dst);
             } break;
         default:
             {
-                GGML_ASSERT(false);
-            } break;
+                GGML_ABORT("fatal error");
+            }
     }
 }
 
@@ -8764,21 +11216,21 @@ static void ggml_compute_forward_abs(
 
 static void ggml_compute_forward_sgn_f32(
         const struct ggml_compute_params * params,
-        const struct ggml_tensor * src0,
         struct ggml_tensor * dst) {
-    assert(params->ith == 0);
-    assert(ggml_are_same_shape(src0, dst));
 
-    if (params->type == GGML_TASK_INIT || params->type == GGML_TASK_FINALIZE) {
+    const struct ggml_tensor * src0 = dst->src[0];
+
+    if (params->ith != 0) {
         return;
     }
 
+    assert(ggml_is_contiguous_1(src0));
+    assert(ggml_is_contiguous_1(dst));
+    assert(ggml_are_same_shape(src0, dst));
+
     const int n  = ggml_nrows(src0);
     const int nc = src0->ne[0];
 
-    assert(dst->nb[0]  == sizeof(float));
-    assert(src0->nb[0] == sizeof(float));
-
     for (int i = 0; i < n; i++) {
         ggml_vec_sgn_f32(nc,
                 (float *) ((char *) dst->data  + i*( dst->nb[1])),
@@ -8788,17 +11240,19 @@ static void ggml_compute_forward_sgn_f32(
 
 static void ggml_compute_forward_sgn(
         const struct ggml_compute_params * params,
-        const struct ggml_tensor * src0,
         struct ggml_tensor * dst) {
+
+    const struct ggml_tensor * src0 = dst->src[0];
+
     switch (src0->type) {
         case GGML_TYPE_F32:
             {
-                ggml_compute_forward_sgn_f32(params, src0, dst);
+                ggml_compute_forward_sgn_f32(params, dst);
             } break;
         default:
             {
-                GGML_ASSERT(false);
-            } break;
+                GGML_ABORT("fatal error");
+            }
     }
 }
 
@@ -8806,21 +11260,21 @@ static void ggml_compute_forward_sgn(
 
 static void ggml_compute_forward_neg_f32(
         const struct ggml_compute_params * params,
-        const struct ggml_tensor * src0,
         struct ggml_tensor * dst) {
-    assert(params->ith == 0);
-    assert(ggml_are_same_shape(src0, dst));
 
-    if (params->type == GGML_TASK_INIT || params->type == GGML_TASK_FINALIZE) {
+    const struct ggml_tensor * src0 = dst->src[0];
+
+    if (params->ith != 0) {
         return;
     }
 
+    assert(ggml_is_contiguous_1(src0));
+    assert(ggml_is_contiguous_1(dst));
+    assert(ggml_are_same_shape(src0, dst));
+
     const int n  = ggml_nrows(src0);
     const int nc = src0->ne[0];
 
-    assert(dst->nb[0]  == sizeof(float));
-    assert(src0->nb[0] == sizeof(float));
-
     for (int i = 0; i < n; i++) {
         ggml_vec_neg_f32(nc,
                 (float *) ((char *) dst->data  + i*( dst->nb[1])),
@@ -8830,17 +11284,19 @@ static void ggml_compute_forward_neg_f32(
 
 static void ggml_compute_forward_neg(
         const struct ggml_compute_params * params,
-        const struct ggml_tensor * src0,
         struct ggml_tensor * dst) {
+
+    const struct ggml_tensor * src0 = dst->src[0];
+
     switch (src0->type) {
         case GGML_TYPE_F32:
             {
-                ggml_compute_forward_neg_f32(params, src0, dst);
+                ggml_compute_forward_neg_f32(params, dst);
             } break;
         default:
             {
-                GGML_ASSERT(false);
-            } break;
+                GGML_ABORT("fatal error");
+            }
     }
 }
 
@@ -8848,21 +11304,21 @@ static void ggml_compute_forward_neg(
 
 static void ggml_compute_forward_step_f32(
         const struct ggml_compute_params * params,
-        const struct ggml_tensor * src0,
         struct ggml_tensor * dst) {
-    assert(params->ith == 0);
-    assert(ggml_are_same_shape(src0, dst));
 
-    if (params->type == GGML_TASK_INIT || params->type == GGML_TASK_FINALIZE) {
+    const struct ggml_tensor * src0 = dst->src[0];
+
+    if (params->ith != 0) {
         return;
     }
 
+    assert(ggml_is_contiguous_1(src0));
+    assert(ggml_is_contiguous_1(dst));
+    assert(ggml_are_same_shape(src0, dst));
+
     const int n  = ggml_nrows(src0);
     const int nc = src0->ne[0];
 
-    assert(dst->nb[0]  == sizeof(float));
-    assert(src0->nb[0] == sizeof(float));
-
     for (int i = 0; i < n; i++) {
         ggml_vec_step_f32(nc,
                 (float *) ((char *) dst->data  + i*( dst->nb[1])),
@@ -8872,17 +11328,19 @@ static void ggml_compute_forward_step_f32(
 
 static void ggml_compute_forward_step(
         const struct ggml_compute_params * params,
-        const struct ggml_tensor * src0,
         struct ggml_tensor * dst) {
+
+    const struct ggml_tensor * src0 = dst->src[0];
+
     switch (src0->type) {
         case GGML_TYPE_F32:
             {
-                ggml_compute_forward_step_f32(params, src0, dst);
+                ggml_compute_forward_step_f32(params, dst);
             } break;
         default:
             {
-                GGML_ASSERT(false);
-            } break;
+                GGML_ABORT("fatal error");
+            }
     }
 }
 
@@ -8890,21 +11348,21 @@ static void ggml_compute_forward_step(
 
 static void ggml_compute_forward_tanh_f32(
         const struct ggml_compute_params * params,
-        const struct ggml_tensor * src0,
         struct ggml_tensor * dst) {
-    assert(params->ith == 0);
-    assert(ggml_are_same_shape(src0, dst));
 
-    if (params->type == GGML_TASK_INIT || params->type == GGML_TASK_FINALIZE) {
+    const struct ggml_tensor * src0 = dst->src[0];
+
+    if (params->ith != 0) {
         return;
     }
 
+    assert(ggml_is_contiguous_1(src0));
+    assert(ggml_is_contiguous_1(dst));
+    assert(ggml_are_same_shape(src0, dst));
+
     const int n  = ggml_nrows(src0);
     const int nc = src0->ne[0];
 
-    assert(dst->nb[0]  == sizeof(float));
-    assert(src0->nb[0] == sizeof(float));
-
     for (int i = 0; i < n; i++) {
         ggml_vec_tanh_f32(nc,
                 (float *) ((char *) dst->data  + i*( dst->nb[1])),
@@ -8914,17 +11372,19 @@ static void ggml_compute_forward_tanh_f32(
 
 static void ggml_compute_forward_tanh(
         const struct ggml_compute_params * params,
-        const struct ggml_tensor * src0,
         struct ggml_tensor * dst) {
+
+    const struct ggml_tensor * src0 = dst->src[0];
+
     switch (src0->type) {
         case GGML_TYPE_F32:
             {
-                ggml_compute_forward_tanh_f32(params, src0, dst);
+                ggml_compute_forward_tanh_f32(params, dst);
             } break;
         default:
             {
-                GGML_ASSERT(false);
-            } break;
+                GGML_ABORT("fatal error");
+            }
     }
 }
 
@@ -8932,21 +11392,21 @@ static void ggml_compute_forward_tanh(
 
 static void ggml_compute_forward_elu_f32(
         const struct ggml_compute_params * params,
-        const struct ggml_tensor * src0,
         struct ggml_tensor * dst) {
-    assert(params->ith == 0);
-    assert(ggml_are_same_shape(src0, dst));
 
-    if (params->type == GGML_TASK_INIT || params->type == GGML_TASK_FINALIZE) {
+    const struct ggml_tensor * src0 = dst->src[0];
+
+    if (params->ith != 0) {
         return;
     }
 
+    assert(ggml_is_contiguous_1(src0));
+    assert(ggml_is_contiguous_1(dst));
+    assert(ggml_are_same_shape(src0, dst));
+
     const int n  = ggml_nrows(src0);
     const int nc = src0->ne[0];
 
-    assert(dst->nb[0]  == sizeof(float));
-    assert(src0->nb[0] == sizeof(float));
-
     for (int i = 0; i < n; i++) {
         ggml_vec_elu_f32(nc,
                 (float *) ((char *) dst->data  + i*( dst->nb[1])),
@@ -8956,17 +11416,19 @@ static void ggml_compute_forward_elu_f32(
 
 static void ggml_compute_forward_elu(
         const struct ggml_compute_params * params,
-        const struct ggml_tensor * src0,
         struct ggml_tensor * dst) {
+
+    const struct ggml_tensor * src0 = dst->src[0];
+
     switch (src0->type) {
         case GGML_TYPE_F32:
             {
-                ggml_compute_forward_elu_f32(params, src0, dst);
+                ggml_compute_forward_elu_f32(params, dst);
             } break;
         default:
             {
-                GGML_ASSERT(false);
-            } break;
+                GGML_ABORT("fatal error");
+            }
     }
 }
 
@@ -8974,21 +11436,21 @@ static void ggml_compute_forward_elu(
 
 static void ggml_compute_forward_relu_f32(
         const struct ggml_compute_params * params,
-        const struct ggml_tensor * src0,
         struct ggml_tensor * dst) {
-    assert(params->ith == 0);
-    assert(ggml_are_same_shape(src0, dst));
 
-    if (params->type == GGML_TASK_INIT || params->type == GGML_TASK_FINALIZE) {
+    const struct ggml_tensor * src0 = dst->src[0];
+
+    if (params->ith != 0) {
         return;
     }
 
+    assert(ggml_is_contiguous_1(src0));
+    assert(ggml_is_contiguous_1(dst));
+    assert(ggml_are_same_shape(src0, dst));
+
     const int n  = ggml_nrows(src0);
     const int nc = src0->ne[0];
 
-    assert(dst->nb[0]  == sizeof(float));
-    assert(src0->nb[0] == sizeof(float));
-
     for (int i = 0; i < n; i++) {
         ggml_vec_relu_f32(nc,
                 (float *) ((char *) dst->data  + i*( dst->nb[1])),
@@ -8998,17 +11460,63 @@ static void ggml_compute_forward_relu_f32(
 
 static void ggml_compute_forward_relu(
         const struct ggml_compute_params * params,
-        const struct ggml_tensor * src0,
         struct ggml_tensor * dst) {
+
+    const struct ggml_tensor * src0 = dst->src[0];
+
     switch (src0->type) {
         case GGML_TYPE_F32:
             {
-                ggml_compute_forward_relu_f32(params, src0, dst);
+                ggml_compute_forward_relu_f32(params, dst);
             } break;
         default:
             {
-                GGML_ASSERT(false);
+                GGML_ABORT("fatal error");
+            }
+    }
+}
+
+// ggml_compute_forward_sigmoid
+
+static void ggml_compute_forward_sigmoid_f32(
+        const struct ggml_compute_params * params,
+        struct ggml_tensor * dst) {
+
+    const struct ggml_tensor * src0 = dst->src[0];
+
+    if (params->ith != 0) {
+        return;
+    }
+
+    assert(ggml_is_contiguous_1(src0));
+    assert(ggml_is_contiguous_1(dst));
+    assert(ggml_are_same_shape(src0, dst));
+
+    const int n  = ggml_nrows(src0);
+    const int nc = src0->ne[0];
+
+    for (int i = 0; i < n; i++) {
+        ggml_vec_sigmoid_f32(nc,
+                (float *) ((char *) dst->data  + i*( dst->nb[1])),
+                (float *) ((char *) src0->data + i*(src0->nb[1])));
+    }
+}
+
+static void ggml_compute_forward_sigmoid(
+        const struct ggml_compute_params * params,
+        struct ggml_tensor * dst) {
+
+    const struct ggml_tensor * src0 = dst->src[0];
+
+    switch (src0->type) {
+        case GGML_TYPE_F32:
+            {
+                ggml_compute_forward_sigmoid_f32(params, dst);
             } break;
+        default:
+            {
+                GGML_ABORT("fatal error");
+            }
     }
 }
 
@@ -9016,15 +11524,13 @@ static void ggml_compute_forward_relu(
 
 static void ggml_compute_forward_gelu_f32(
         const struct ggml_compute_params * params,
-        const struct ggml_tensor * src0,
         struct ggml_tensor * dst) {
-    GGML_ASSERT(ggml_is_contiguous_except_dim_1(src0));
-    GGML_ASSERT(ggml_is_contiguous_except_dim_1(dst));
-    GGML_ASSERT(ggml_are_same_shape(src0, dst));
 
-    if (params->type == GGML_TASK_INIT || params->type == GGML_TASK_FINALIZE) {
-        return;
-    }
+    const struct ggml_tensor * src0 = dst->src[0];
+
+    assert(ggml_is_contiguous_1(src0));
+    assert(ggml_is_contiguous_1(dst));
+    assert(ggml_are_same_shape(src0, dst));
 
     const int ith = params->ith;
     const int nth = params->nth;
@@ -9057,17 +11563,19 @@ static void ggml_compute_forward_gelu_f32(
 
 static void ggml_compute_forward_gelu(
         const struct ggml_compute_params * params,
-        const struct ggml_tensor * src0,
         struct ggml_tensor * dst) {
+
+    const struct ggml_tensor * src0 = dst->src[0];
+
     switch (src0->type) {
         case GGML_TYPE_F32:
             {
-                ggml_compute_forward_gelu_f32(params, src0, dst);
+                ggml_compute_forward_gelu_f32(params, dst);
             } break;
         default:
             {
-                GGML_ASSERT(false);
-            } break;
+                GGML_ABORT("fatal error");
+            }
     }
 }
 
@@ -9075,15 +11583,13 @@ static void ggml_compute_forward_gelu(
 
 static void ggml_compute_forward_gelu_quick_f32(
         const struct ggml_compute_params * params,
-        const struct ggml_tensor * src0,
         struct ggml_tensor * dst) {
-    GGML_ASSERT(ggml_is_contiguous_except_dim_1(src0));
-    GGML_ASSERT(ggml_is_contiguous_except_dim_1(dst));
-    GGML_ASSERT(ggml_are_same_shape(src0, dst));
 
-    if (params->type == GGML_TASK_INIT || params->type == GGML_TASK_FINALIZE) {
-        return;
-    }
+    const struct ggml_tensor * src0 = dst->src[0];
+
+    assert(ggml_is_contiguous_1(src0));
+    assert(ggml_is_contiguous_1(dst));
+    assert(ggml_are_same_shape(src0, dst));
 
     const int ith = params->ith;
     const int nth = params->nth;
@@ -9116,17 +11622,19 @@ static void ggml_compute_forward_gelu_quick_f32(
 
 static void ggml_compute_forward_gelu_quick(
         const struct ggml_compute_params * params,
-        const struct ggml_tensor * src0,
         struct ggml_tensor * dst) {
+
+    const struct ggml_tensor * src0 = dst->src[0];
+
     switch (src0->type) {
         case GGML_TYPE_F32:
             {
-                ggml_compute_forward_gelu_quick_f32(params, src0, dst);
+                ggml_compute_forward_gelu_quick_f32(params, dst);
             } break;
         default:
             {
-                GGML_ASSERT(false);
-            } break;
+                GGML_ABORT("fatal error");
+            }
     }
 }
 
@@ -9134,15 +11642,13 @@ static void ggml_compute_forward_gelu_quick(
 
 static void ggml_compute_forward_silu_f32(
         const struct ggml_compute_params * params,
-        const struct ggml_tensor * src0,
         struct ggml_tensor * dst) {
-    GGML_ASSERT(ggml_is_contiguous_except_dim_1(src0));
-    GGML_ASSERT(ggml_is_contiguous_except_dim_1(dst));
-    GGML_ASSERT(ggml_are_same_shape(src0, dst));
 
-    if (params->type == GGML_TASK_INIT || params->type == GGML_TASK_FINALIZE) {
-        return;
-    }
+    const struct ggml_tensor * src0 = dst->src[0];
+
+    assert(ggml_is_contiguous_1(src0));
+    assert(ggml_is_contiguous_1(dst));
+    assert(ggml_are_same_shape(src0, dst));
 
     const int ith = params->ith;
     const int nth = params->nth;
@@ -9175,32 +11681,37 @@ static void ggml_compute_forward_silu_f32(
 
 static void ggml_compute_forward_silu(
         const struct ggml_compute_params * params,
-        const struct ggml_tensor * src0,
         struct ggml_tensor * dst) {
+
+    const struct ggml_tensor * src0 = dst->src[0];
+
     switch (src0->type) {
         case GGML_TYPE_F32:
             {
-                ggml_compute_forward_silu_f32(params, src0, dst);
+                ggml_compute_forward_silu_f32(params, dst);
             } break;
         default:
             {
-                GGML_ASSERT(false);
-            } break;
+                GGML_ABORT("fatal error");
+            }
     }
 }
 // ggml_compute_forward_leaky_relu
 
 static void ggml_compute_forward_leaky_relu_f32(
         const struct ggml_compute_params * params,
-        const struct ggml_tensor * src0,
         struct ggml_tensor * dst) {
-    assert(params->ith == 0);
-    assert(ggml_are_same_shape(src0, dst));
 
-    if (params->type == GGML_TASK_INIT || params->type == GGML_TASK_FINALIZE) {
+    const struct ggml_tensor * src0 = dst->src[0];
+
+    if (params->ith != 0) {
         return;
     }
 
+    assert(ggml_is_contiguous_1(src0));
+    assert(ggml_is_contiguous_1(dst));
+    assert(ggml_are_same_shape(src0, dst));
+
     const int n  = ggml_nrows(src0);
     const int nc = src0->ne[0];
 
@@ -9219,17 +11730,19 @@ static void ggml_compute_forward_leaky_relu_f32(
 
 static void ggml_compute_forward_leaky_relu(
         const struct ggml_compute_params * params,
-        const struct ggml_tensor * src0,
         struct ggml_tensor * dst) {
+
+    const struct ggml_tensor * src0 = dst->src[0];
+
     switch (src0->type) {
         case GGML_TYPE_F32:
             {
-                ggml_compute_forward_leaky_relu_f32(params, src0, dst);
+                ggml_compute_forward_leaky_relu_f32(params, dst);
             } break;
         default:
             {
-                GGML_ASSERT(false);
-            } break;
+                GGML_ABORT("fatal error");
+            }
     }
 }
 
@@ -9237,18 +11750,16 @@ static void ggml_compute_forward_leaky_relu(
 
 static void ggml_compute_forward_silu_back_f32(
         const struct ggml_compute_params * params,
-        const struct ggml_tensor * src0,
-        const struct ggml_tensor * grad,
         struct ggml_tensor * dst) {
-    GGML_ASSERT(ggml_is_contiguous_except_dim_1(grad));
-    GGML_ASSERT(ggml_is_contiguous_except_dim_1(src0));
-    GGML_ASSERT(ggml_is_contiguous_except_dim_1(dst));
-    GGML_ASSERT(ggml_are_same_shape(src0, dst));
-    GGML_ASSERT(ggml_are_same_shape(src0, grad));
 
-    if (params->type == GGML_TASK_INIT || params->type == GGML_TASK_FINALIZE) {
-        return;
-    }
+    const struct ggml_tensor * src0 = dst->src[0];
+    const struct ggml_tensor * grad = dst->src[1];
+
+    assert(ggml_is_contiguous_1(grad));
+    assert(ggml_is_contiguous_1(src0));
+    assert(ggml_is_contiguous_1(dst));
+    assert(ggml_are_same_shape(src0, dst));
+    assert(ggml_are_same_shape(src0, grad));
 
     const int ith = params->ith;
     const int nth = params->nth;
@@ -9282,33 +11793,159 @@ static void ggml_compute_forward_silu_back_f32(
 
 static void ggml_compute_forward_silu_back(
         const struct ggml_compute_params * params,
-        const struct ggml_tensor * src0,
-        const struct ggml_tensor * grad,
         struct ggml_tensor * dst) {
+
+    const struct ggml_tensor * src0 = dst->src[0];
+
     switch (src0->type) {
         case GGML_TYPE_F32:
             {
-                ggml_compute_forward_silu_back_f32(params, src0, grad, dst);
+                ggml_compute_forward_silu_back_f32(params, dst);
             } break;
         default:
             {
-                GGML_ASSERT(false);
+                GGML_ABORT("fatal error");
+            }
+    }
+}
+
+
+static void ggml_compute_forward_hardswish_f32(
+        const struct ggml_compute_params * params,
+        struct ggml_tensor * dst) {
+
+    const struct ggml_tensor * src0 = dst->src[0];
+
+    if (params->ith != 0) {
+        return;
+    }
+
+    assert(ggml_is_contiguous_1(src0));
+    assert(ggml_is_contiguous_1(dst));
+    assert(ggml_are_same_shape(src0, dst));
+
+    const int n  = ggml_nrows(src0);
+    const int nc = src0->ne[0];
+
+    for (int i = 0; i < n; i++) {
+        ggml_vec_hardswish_f32(nc,
+                (float *) ((char *) dst->data  + i*( dst->nb[1])),
+                (float *) ((char *) src0->data + i*(src0->nb[1])));
+    }
+}
+static void ggml_compute_forward_hardswish(
+        const struct ggml_compute_params * params,
+        struct ggml_tensor * dst) {
+
+    const struct ggml_tensor * src0 = dst->src[0];
+
+    switch (src0->type) {
+        case GGML_TYPE_F32:
+            {
+                ggml_compute_forward_hardswish_f32(params, dst);
             } break;
+        default:
+            {
+                GGML_ABORT("fatal error");
+            }
     }
 }
 
-// ggml_compute_forward_norm
+static void ggml_compute_forward_hardsigmoid_f32(
+        const struct ggml_compute_params * params,
+        struct ggml_tensor * dst) {
 
-static void ggml_compute_forward_norm_f32(
+    const struct ggml_tensor * src0 = dst->src[0];
+
+    if (params->ith != 0) {
+        return;
+    }
+
+    assert(ggml_is_contiguous_1(src0));
+    assert(ggml_is_contiguous_1(dst));
+    assert(ggml_are_same_shape(src0, dst));
+
+    const int n  = ggml_nrows(src0);
+    const int nc = src0->ne[0];
+
+    for (int i = 0; i < n; i++) {
+        ggml_vec_hardsigmoid_f32(nc,
+                (float *) ((char *) dst->data  + i*( dst->nb[1])),
+                (float *) ((char *) src0->data + i*(src0->nb[1])));
+    }
+}
+
+static void ggml_compute_forward_hardsigmoid(
         const struct ggml_compute_params * params,
-        const struct ggml_tensor * src0,
         struct ggml_tensor * dst) {
-    GGML_ASSERT(ggml_are_same_shape(src0, dst));
 
-    if (params->type == GGML_TASK_INIT || params->type == GGML_TASK_FINALIZE) {
+    const struct ggml_tensor * src0 = dst->src[0];
+
+    switch (src0->type) {
+        case GGML_TYPE_F32:
+            {
+                ggml_compute_forward_hardsigmoid_f32(params, dst);
+            } break;
+        default:
+            {
+                GGML_ABORT("fatal error");
+            }
+    }
+}
+
+static void ggml_compute_forward_exp_f32(
+        const struct ggml_compute_params * params,
+        struct ggml_tensor * dst) {
+
+    const struct ggml_tensor * src0 = dst->src[0];
+
+    if (params->ith != 0) {
         return;
     }
 
+    assert(ggml_is_contiguous_1(src0));
+    assert(ggml_is_contiguous_1(dst));
+    assert(ggml_are_same_shape(src0, dst));
+
+    const int n  = ggml_nrows(src0);
+    const int nc = src0->ne[0];
+
+    for (int i = 0; i < n; i++) {
+        ggml_vec_exp_f32(nc,
+                (float *) ((char *) dst->data  + i*( dst->nb[1])),
+                (float *) ((char *) src0->data + i*(src0->nb[1])));
+    }
+}
+
+static void ggml_compute_forward_exp(
+        const struct ggml_compute_params * params,
+        struct ggml_tensor * dst) {
+
+    const struct ggml_tensor * src0 = dst->src[0];
+
+    switch (src0->type) {
+        case GGML_TYPE_F32:
+            {
+                ggml_compute_forward_exp_f32(params, dst);
+            } break;
+        default:
+            {
+                GGML_ABORT("fatal error");
+            }
+    }
+}
+
+
+// ggml_compute_forward_norm
+
+static void ggml_compute_forward_norm_f32(
+        const struct ggml_compute_params * params,
+        struct ggml_tensor * dst) {
+
+    const struct ggml_tensor * src0 = dst->src[0];
+
+    GGML_ASSERT(ggml_are_same_shape(src0, dst));
+
     GGML_ASSERT(src0->nb[0] == sizeof(float));
 
     const int ith = params->ith;
@@ -9354,17 +11991,19 @@ static void ggml_compute_forward_norm_f32(
 
 static void ggml_compute_forward_norm(
         const struct ggml_compute_params * params,
-        const struct ggml_tensor * src0,
         struct ggml_tensor * dst) {
+
+    const struct ggml_tensor * src0 = dst->src[0];
+
     switch (src0->type) {
         case GGML_TYPE_F32:
             {
-                ggml_compute_forward_norm_f32(params, src0, dst);
+                ggml_compute_forward_norm_f32(params, dst);
             } break;
         default:
             {
-                GGML_ASSERT(false);
-            } break;
+                GGML_ABORT("fatal error");
+            }
     }
 }
 
@@ -9372,13 +12011,11 @@ static void ggml_compute_forward_norm(
 
 static void ggml_compute_forward_rms_norm_f32(
         const struct ggml_compute_params * params,
-        const struct ggml_tensor * src0,
         struct ggml_tensor * dst) {
-    GGML_ASSERT(ggml_are_same_shape(src0, dst));
 
-    if (params->type == GGML_TASK_INIT || params->type == GGML_TASK_FINALIZE) {
-        return;
-    }
+    const struct ggml_tensor * src0 = dst->src[0];
+
+    GGML_ASSERT(ggml_are_same_shape(src0, dst));
 
     GGML_ASSERT(src0->nb[0] == sizeof(float));
 
@@ -9422,30 +12059,30 @@ static void ggml_compute_forward_rms_norm_f32(
 
 static void ggml_compute_forward_rms_norm(
         const struct ggml_compute_params * params,
-        const struct ggml_tensor * src0,
         struct ggml_tensor * dst) {
+
+    const struct ggml_tensor * src0 = dst->src[0];
+
     switch (src0->type) {
         case GGML_TYPE_F32:
             {
-                ggml_compute_forward_rms_norm_f32(params, src0, dst);
+                ggml_compute_forward_rms_norm_f32(params, dst);
             } break;
         default:
             {
-                GGML_ASSERT(false);
-            } break;
+                GGML_ABORT("fatal error");
+            }
     }
 }
 
 static void ggml_compute_forward_rms_norm_back_f32(
         const struct ggml_compute_params * params,
-        const struct ggml_tensor * src0,
-        const struct ggml_tensor * src1,
         struct ggml_tensor * dst) {
-    GGML_ASSERT(ggml_are_same_shape(src0, dst) && ggml_are_same_shape(src0, src1));
 
-    if (params->type == GGML_TASK_INIT || params->type == GGML_TASK_FINALIZE) {
-        return;
-    }
+    const struct ggml_tensor * src0 = dst->src[0];
+    const struct ggml_tensor * src1 = dst->src[1];
+
+    GGML_ASSERT(ggml_are_same_shape(src0, dst) && ggml_are_same_shape(src0, src1));
 
     GGML_ASSERT(src0->nb[0] == sizeof(float));
 
@@ -9595,18 +12232,19 @@ static void ggml_compute_forward_rms_norm_back_f32(
 
 static void ggml_compute_forward_rms_norm_back(
         const struct ggml_compute_params * params,
-        const struct ggml_tensor * src0,
-        const struct ggml_tensor * src1,
         struct ggml_tensor * dst) {
+
+    const struct ggml_tensor * src0 = dst->src[0];
+
     switch (src0->type) {
         case GGML_TYPE_F32:
             {
-                ggml_compute_forward_rms_norm_back_f32(params, src0, src1, dst);
+                ggml_compute_forward_rms_norm_back_f32(params, dst);
             } break;
         default:
             {
-                GGML_ASSERT(false);
-            } break;
+                GGML_ABORT("fatal error");
+            }
     }
 }
 
@@ -9614,13 +12252,11 @@ static void ggml_compute_forward_rms_norm_back(
 
 static void ggml_compute_forward_group_norm_f32(
     const struct ggml_compute_params * params,
-    const struct ggml_tensor * src0,
     struct ggml_tensor * dst) {
-    GGML_ASSERT(ggml_are_same_shape(src0, dst));
 
-    if (params->type == GGML_TASK_INIT || params->type == GGML_TASK_FINALIZE) {
-        return;
-    }
+    const struct ggml_tensor * src0 = dst->src[0];
+
+    GGML_ASSERT(ggml_are_same_shape(src0, dst));
 
     GGML_ASSERT(src0->nb[0] == sizeof(float));
 
@@ -9629,14 +12265,15 @@ static void ggml_compute_forward_group_norm_f32(
 
     GGML_TENSOR_UNARY_OP_LOCALS
 
-    const float eps = 1e-6f; // TODO: make this a parameter
-
     // TODO: optimize
 
+    float eps;
+    memcpy(&eps, dst->op_params + 1, sizeof(float));
+
     int n_channels = src0->ne[2];
     int n_groups = dst->op_params[0];
     int n_channels_per_group = (n_channels + n_groups - 1) / n_groups;
-    for (int i = ith; i < n_groups; i+=nth) {
+    for (int i = ith; i < n_groups; i += nth) {
         int start = i * n_channels_per_group;
         int end = start + n_channels_per_group;
         if (end > n_channels) {
@@ -9650,28 +12287,32 @@ static void ggml_compute_forward_group_norm_f32(
                 for (int64_t i01 = 0; i01 < ne01; i01++) {
                     const float * x = (float *)((char *) src0->data + i01 * nb01 + i02 * nb02 + i03 * nb03);
 
+                    ggml_float sumr = 0.0;
                     for (int64_t i00 = 0; i00 < ne00; i00++) {
-                        sum += (ggml_float)x[i00];
+                        sumr += (ggml_float)x[i00];
                     }
+                    sum += sumr;
                 }
             }
-            float mean = sum / (ne00 * ne01 * step);
-            ggml_float sum2 = 0.0;
+            const float mean = sum / (ne00 * ne01 * step);
 
+            ggml_float sum2 = 0.0;
             for (int64_t i02 = start; i02 < end; i02++) {
                 for (int64_t i01 = 0; i01 < ne01; i01++) {
                     const float * x = (float *)((char *) src0->data + i01 * nb01 + i02 * nb02 + i03 * nb03);
 
                     float * y = (float *)((char *) dst->data + i01 * nb1 + i02 * nb2 + i03 * nb3);
 
+                    ggml_float sumr = 0.0;
                     for (int64_t i00 = 0; i00 < ne00; i00++) {
                         float v = x[i00] - mean;
                         y[i00] = v;
-                        sum2 += (ggml_float)(v * v);
+                        sumr += (ggml_float)(v * v);
                     }
+                    sum2 += sumr;
                 }
             }
-            float variance = sum2 / (ne00 * ne01 * step);
+            const float variance = sum2 / (ne00 * ne01 * step);
             const float scale = 1.0f / sqrtf(variance + eps);
 
             for (int64_t i02 = start; i02 < end; i02++) {
@@ -9686,62 +12327,121 @@ static void ggml_compute_forward_group_norm_f32(
 
 static void ggml_compute_forward_group_norm(
     const struct ggml_compute_params * params,
-    const struct ggml_tensor * src0,
     struct ggml_tensor * dst) {
+
+    const struct ggml_tensor * src0 = dst->src[0];
+
     switch (src0->type) {
         case GGML_TYPE_F32:
             {
-                ggml_compute_forward_group_norm_f32(params, src0, dst);
+                ggml_compute_forward_group_norm_f32(params, dst);
             } break;
         default:
             {
-                GGML_ASSERT(false);
-            } break;
+                GGML_ABORT("fatal error");
+            }
     }
 }
 
 // ggml_compute_forward_mul_mat
 
-#if defined(GGML_USE_ACCELERATE) || defined(GGML_USE_OPENBLAS)
-// helper function to determine if it is better to use BLAS or not
-// for large matrices, BLAS is faster
-static bool ggml_compute_forward_mul_mat_use_blas(
-        const struct ggml_tensor * src0,
-        const struct ggml_tensor * src1,
-              struct ggml_tensor * dst) {
-    //const int64_t ne00 = src0->ne[0];
-    //const int64_t ne01 = src0->ne[1];
+static void ggml_compute_forward_mul_mat_one_chunk(
+    const struct ggml_compute_params * params,
+    struct ggml_tensor * dst,
+    const int64_t num_rows_per_vec_dot,
+    const int64_t ir0_start,
+    const int64_t ir0_end,
+    const int64_t ir1_start,
+    const int64_t ir1_end) {
 
-    const int64_t ne10 = src1->ne[0];
+    const struct ggml_tensor * src0 = dst->src[0];
+    const struct ggml_tensor * src1 = dst->src[1];
 
-    const int64_t ne0 = dst->ne[0];
-    const int64_t ne1 = dst->ne[1];
+    GGML_TENSOR_BINARY_OP_LOCALS
 
-    // NOTE: with GGML_OP_MUL_MAT_ID we don't want to go through the BLAS branch because it will dequantize (to_float)
-    //       all the experts for each batch element and the processing would become incredibly slow
-    // TODO: find the optimal values for these
-    if (dst->op != GGML_OP_MUL_MAT_ID &&
-        ggml_is_contiguous(src0) &&
-        ggml_is_contiguous(src1) &&
-      //src0->type == GGML_TYPE_F32 &&
-        src1->type == GGML_TYPE_F32 &&
-        (ne0 >= 32 && ne1 >= 32 && ne10 >= 32)) {
+    const enum ggml_type type = src0->type;
 
-        /*printf("BLAS: %d %d %d %d %d\n", ne0, ne1, ne10, ne00, ne01);*/
-        return true;
+    const bool src1_cont = ggml_is_contiguous(src1);
+
+    ggml_vec_dot_t const vec_dot      = type_traits[type].vec_dot;
+    enum ggml_type const vec_dot_type = type_traits[type].vec_dot_type;
+
+    // broadcast factors
+    const int64_t r2 = ne12 / ne02;
+    const int64_t r3 = ne13 / ne03;
+
+    //printf("ir0_start = %6lld, ir0_end = %6lld, ir1_start = %6lld, ir1_end = %6lld\n", ir0_start, ir0_end, ir1_start, ir1_end);
+
+    // threads with no work simply yield (not sure if it helps)
+    if (ir0_start >= ir0_end || ir1_start >= ir1_end) {
+        return;
     }
 
-    return false;
+    const void * wdata = (src1->type == vec_dot_type) ? src1->data : params->wdata;
+    const size_t row_size = ggml_row_size(vec_dot_type, ne10);
+
+    assert(ne12 % ne02 == 0);
+    assert(ne13 % ne03 == 0);
+
+    // block-tiling attempt
+    const int64_t blck_0 = 16;
+    const int64_t blck_1 = 16;
+
+    const size_t src1_col_stride = src1_cont || src1->type != vec_dot_type ? row_size : nb11;
+
+    // attempt to reduce false-sharing (does not seem to make a difference)
+    // 16 * 2, accounting for mmla kernels
+    float tmp[32];
+
+    for (int64_t iir1 = ir1_start; iir1 < ir1_end; iir1 += blck_1) {
+        for (int64_t iir0 = ir0_start; iir0 < ir0_end; iir0 += blck_0) {
+            for (int64_t ir1 = iir1; ir1 < iir1 + blck_1 && ir1 < ir1_end; ir1 += num_rows_per_vec_dot) {
+                const int64_t i13 = (ir1 / (ne12 * ne1));
+                const int64_t i12 = (ir1 - i13 * ne12 * ne1) / ne1;
+                const int64_t i11 = (ir1 - i13 * ne12 * ne1 - i12 * ne1);
+
+                // broadcast src0 into src1
+                const int64_t i03 = i13 / r3;
+                const int64_t i02 = i12 / r2;
+
+                const int64_t i1 = i11;
+                const int64_t i2 = i12;
+                const int64_t i3 = i13;
+
+                const char * src0_row = (const char*)src0->data + (0 + i02 * nb02 + i03 * nb03);
+
+                // desc: when src1 is not a contiguous memory block we have to calculate the offset using the strides
+                //       if it is, then we have either copied the data to params->wdata and made it contiguous or we are using
+                //       the original src1 data pointer, so we should index using the indices directly
+                // TODO: this is a bit of a hack, we should probably have a better way to handle this
+                const char * src1_col = (const char*)wdata +
+                    (src1_cont || src1->type != vec_dot_type
+                        ? (i11 + i12 * ne11 + i13 * ne12 * ne11) * row_size
+                        : (i11 * nb11 + i12 * nb12 + i13 * nb13));
+                float * dst_col = (float*)((char*)dst->data + (i1 * nb1 + i2 * nb2 + i3 * nb3));
+
+                //for (int64_t ir0 = iir0; ir0 < iir0 + blck_0 && ir0 < ir0_end; ++ir0) {
+                //    vec_dot(ne00, &dst_col[ir0], src0_row + ir0*nb01, src1_col);
+                //}
+
+                for (int64_t ir0 = iir0; ir0 < iir0 + blck_0 && ir0 < ir0_end; ir0 += num_rows_per_vec_dot) {
+                    vec_dot(ne00, &tmp[ir0 - iir0], (num_rows_per_vec_dot > 1 ? 16 : 0), src0_row + ir0 * nb01, (num_rows_per_vec_dot > 1 ? nb01 : 0), src1_col, (num_rows_per_vec_dot > 1 ? src1_col_stride : 0), num_rows_per_vec_dot);
+                }
+
+                for (int cn = 0; cn < num_rows_per_vec_dot; ++cn) {
+                    memcpy(&dst_col[iir0 + cn * nb1 / nb0], tmp + (cn * 16), (MIN(iir0 + blck_0, ir0_end) - iir0) * sizeof(float));
+                }
+            }
+        }
+    }
 }
-#endif
 
 static void ggml_compute_forward_mul_mat(
         const struct ggml_compute_params * params,
-        const struct ggml_tensor * src0,
-        const struct ggml_tensor * src1,
               struct ggml_tensor * dst) {
-    int64_t t0 = ggml_perf_time_us();
-    UNUSED(t0);
+
+    const struct ggml_tensor * src0 = dst->src[0];
+    const struct ggml_tensor * src1 = dst->src[1];
 
     GGML_TENSOR_BINARY_OP_LOCALS
 
@@ -9750,11 +12450,14 @@ static void ggml_compute_forward_mul_mat(
 
     const enum ggml_type type = src0->type;
 
-    const bool src1_cont = ggml_is_contiguous(src1);
-
-    ggml_vec_dot_t    const vec_dot               = type_traits[type].vec_dot;
-    enum ggml_type    const vec_dot_type          = type_traits[type].vec_dot_type;
-    ggml_from_float_t const from_float_to_vec_dot = type_traits[vec_dot_type].from_float;
+    enum ggml_type           const vec_dot_type         = type_traits[type].vec_dot_type;
+    ggml_from_float_t        const from_float           = type_traits[vec_dot_type].from_float;
+    ggml_from_float_to_mat_t const from_float_to_mat    = type_traits[vec_dot_type].from_float_to_mat;
+    int64_t                  const vec_dot_num_rows     = type_traits[type].nrows;
+    int64_t                  const matmul_num_cols      = type_traits[type].ncols;
+    int64_t                  const blck_size_interleave = type_traits[type].blck_size_interleave;
+    ggml_gemv_t              const gemv                 = type_traits[type].gemv;
+    ggml_gemm_t              const gemm                 = type_traits[type].gemm;
 
     GGML_ASSERT(ne0 == ne01);
     GGML_ASSERT(ne1 == ne11);
@@ -9771,180 +12474,180 @@ static void ggml_compute_forward_mul_mat(
     GGML_ASSERT(nb1 <= nb2);
     GGML_ASSERT(nb2 <= nb3);
 
-    // broadcast factors
-    const int64_t r2 = ne12/ne02;
-    const int64_t r3 = ne13/ne03;
-
     // nb01 >= nb00 - src0 is not transposed
     //   compute by src0 rows
 
-#if defined(GGML_USE_CLBLAST)
-    if (ggml_cl_can_mul_mat(src0, src1, dst)) {
-        if (params->ith == 0 && params->type == GGML_TASK_COMPUTE) {
-            ggml_cl_mul_mat(src0, src1, dst, params->wdata, params->wsize);
-        }
-        return;
-    }
-#endif
-
-#if defined(GGML_USE_ACCELERATE) || defined(GGML_USE_OPENBLAS)
-    if (ggml_compute_forward_mul_mat_use_blas(src0, src1, dst)) {
-        if (params->ith != 0) {
-            return;
-        }
-
-        if (params->type == GGML_TASK_INIT) {
-            return;
-        }
-
-        if (params->type == GGML_TASK_FINALIZE) {
-            return;
-        }
-
-        for (int64_t i13 = 0; i13 < ne13; i13++) {
-            for (int64_t i12 = 0; i12 < ne12; i12++) {
-                // broadcast src0 into src1 across 2nd,3rd dimension
-                const int64_t i03 = i13/r3;
-                const int64_t i02 = i12/r2;
-
-                const void  * x = (char *)            src0->data + i02*nb02 + i03*nb03;
-                const float * y = (float *) ((char *) src1->data + i12*nb12 + i13*nb13);
-                      float * d = (float *) ((char *)  dst->data + i12*nb2  + i13*nb3);
-
-                if (type != GGML_TYPE_F32) {
-                            float * const wdata    = params->wdata;
-                    ggml_to_float_t const to_float = type_traits[type].to_float;
-
-                    size_t id = 0;
-                    for (int64_t i01 = 0; i01 < ne01; ++i01) {
-                        to_float((const char *) x + i01*nb01, wdata + id, ne00);
-                        id += ne00;
-                    }
-
-                    assert(id*sizeof(float) <= params->wsize);
-                    x = wdata;
-                }
-
-                cblas_sgemm(CblasRowMajor, CblasNoTrans, CblasTrans,
-                          ne1, ne01, ne10,
-                         1.0f,    y, ne10,
-                                  x, ne00,
-                         0.0f,    d, ne01);
-            }
-        }
+#if GGML_USE_LLAMAFILE
+    // broadcast factors
+    const int64_t r2 = ne12 / ne02;
+    const int64_t r3 = ne13 / ne03;
 
-        //printf("CBLAS = %f ms, %d x %d x %d x %d\n", (ggml_perf_time_us() - t0)/1000.0, ne0, ne1, ne2, ne3);
+    const bool src1_cont = ggml_is_contiguous(src1);
 
+    if (src1_cont) {
+        for (int64_t i13 = 0; i13 < ne13; i13++)
+            for (int64_t i12 = 0; i12 < ne12; i12++)
+                if (!llamafile_sgemm(ne01, ne11, ne00/ggml_blck_size(src0->type),
+                                     (const char *)src0->data + i12/r2*nb02 + i13/r3*nb03,
+                                     nb01/ggml_type_size(src0->type),
+                                     (const char *)src1->data + i12*nb12 + i13*nb13,
+                                     nb11/ggml_type_size(src1->type),
+                                     (char *)dst->data + i12*nb2 + i13*nb3,
+                                     nb1/ggml_type_size(dst->type),
+                                     ith, nth,
+                                     src0->type,
+                                     src1->type,
+                                     dst->type))
+                    goto UseGgmlGemm1;
         return;
     }
+UseGgmlGemm1:;
 #endif
 
-    if (params->type == GGML_TASK_INIT) {
-        if (src1->type != vec_dot_type) {
-            char * wdata = params->wdata;
-            const size_t row_size = ggml_row_size(vec_dot_type, ne10);
-
-            assert(params->wsize >= ne11*ne12*ne13*row_size);
-            GGML_ASSERT(src1->type == GGML_TYPE_F32);
+    if (src1->type != vec_dot_type) {
+        char * wdata = params->wdata;
 
-            for (int64_t i13 = 0; i13 < ne13; ++i13) {
-                for (int64_t i12 = 0; i12 < ne12; ++i12) {
-                    for (int64_t i11 = 0; i11 < ne11; ++i11) {
-                        from_float_to_vec_dot((float *)((char *) src1->data + i13*nb13 + i12*nb12 + i11*nb11), (void *) wdata, ne10);
-                        wdata += row_size;
+        const size_t nbw1 = ggml_row_size(vec_dot_type, ne10);
+        const size_t nbw2 = nbw1*ne11;
+        const size_t nbw3 = nbw2*ne12;
+
+        assert(params->wsize >= ne13*nbw3);
+        GGML_ASSERT(src1->type == GGML_TYPE_F32);
+
+        for (int64_t i13 = 0; i13 < ne13; ++i13) {
+            for (int64_t i12 = 0; i12 < ne12; ++i12) {
+                int64_t i11_processed = 0;
+                if ((ggml_n_dims(src1) == 2) && from_float_to_mat && gemm) {
+                    for (int64_t i11 = ith * 4; i11 < ne11 - ne11 % 4; i11 += nth * 4) {
+                        from_float_to_mat((float *)((char *) src1->data + i13*nb13 + i12*nb12 + i11*nb11),
+                                          (void *)               (wdata + i13*nbw3 + i12*nbw2 + i11*nbw1),
+                                          4, ne10, blck_size_interleave);
                     }
+                    i11_processed = ne11 - ne11 % 4;
+                }
+                for (int64_t i11 = i11_processed + ith; i11 < ne11; i11 += nth) {
+                    from_float((float *)((char *) src1->data + i13*nb13 + i12*nb12 + i11*nb11),
+                           (void *)               (wdata + i13*nbw3 + i12*nbw2 + i11*nbw1),
+                           ne10);
                 }
             }
         }
-
-        return;
     }
 
-    if (params->type == GGML_TASK_FINALIZE) {
-        return;
+    if (ith == 0) {
+        // Every thread starts at ith, so the first unprocessed chunk is nth.  This save a bit of coordination right at the start.
+        atomic_store_explicit(&params->threadpool->current_chunk, nth, memory_order_relaxed);
     }
 
-    const void * wdata    = (src1->type == vec_dot_type) ? src1->data : params->wdata;
-    const size_t row_size = ggml_row_size(vec_dot_type, ne10);
-
-    const int64_t nr0 = ne01;          // src0 rows
-    const int64_t nr1 = ne1*ne12*ne13; // src1 rows
-
-    //printf("nr0 = %lld, nr1 = %lld\n", nr0, nr1);
-
-    // distribute the thread work across the inner or outer loop based on which one is larger
+    ggml_barrier(params->threadpool);
 
-    const int64_t nth0 = nr0 > nr1 ? nth : 1; // parallelize by src0 rows
-    const int64_t nth1 = nr0 > nr1 ? 1 : nth; // parallelize by src1 rows
+#if GGML_USE_LLAMAFILE
+    if (src1->type != vec_dot_type) {
+        const void* wdata = (src1->type == vec_dot_type) ? src1->data : params->wdata;
+        const size_t row_size = ggml_row_size(vec_dot_type, ne10);
 
-    const int64_t ith0 = ith % nth0;
-    const int64_t ith1 = ith / nth0;
+        for (int64_t i13 = 0; i13 < ne13; i13++)
+            for (int64_t i12 = 0; i12 < ne12; i12++)
+                if (!llamafile_sgemm(ne01, ne11, ne00/ggml_blck_size(src0->type),
+                                     (const char *)src0->data + i12/r2*nb02 + i13/r3*nb03,
+                                     nb01/ggml_type_size(src0->type),
+                                     (const char *)wdata + (i12*ne11 + i13*ne12*ne11)*row_size,
+                                     row_size/ggml_type_size(vec_dot_type),
+                                     (char *)dst->data + i12*nb2 + i13*nb3,
+                                     nb1/ggml_type_size(dst->type),
+                                     ith, nth,
+                                     src0->type,
+                                     vec_dot_type,
+                                     dst->type))
+                    goto UseGgmlGemm2;
+        return;
+    }
+UseGgmlGemm2:;
+#endif
 
-    const int64_t dr0 = (nr0 + nth0 - 1)/nth0;
-    const int64_t dr1 = (nr1 + nth1 - 1)/nth1;
+    // This is the size of the first dimension of the result, so we can iterate that way. (see the ASSERT above, these are the same numbers)
+    const int64_t nr0 = ne0;
 
-    const int64_t ir010 = dr0*ith0;
-    const int64_t ir011 = MIN(ir010 + dr0, nr0);
+    // This is the size of the rest of the dimensions of the result
+    const int64_t nr1 = ne1 * ne2 * ne3;
 
-    const int64_t ir110 = dr1*ith1;
-    const int64_t ir111 = MIN(ir110 + dr1, nr1);
+    // dot kernels can handle 1 row and col at a time, but mmla kernels can process 2 rows and cols
+    int64_t num_rows_per_vec_dot = vec_dot_num_rows;
+    // TODO: currently the mmla kernels support only even numbered rows/cols.
+    // this check can be removed once they are extended to support odd numbered rows/cols too
+    if ((nr0 % 2 != 0) || (ne11 % 2 != 0)) {
+        num_rows_per_vec_dot = 1;
+    }
 
-    //printf("ir010 = %6lld, ir011 = %6lld, ir110 = %6lld, ir111 = %6lld\n", ir010, ir011, ir110, ir111);
+    // Now select a reasonable chunk size.
+    int chunk_size = 16;
 
-    // threads with no work simply yield (not sure if it helps)
-    if (ir010 >= ir011 || ir110 >= ir111) {
-        sched_yield();
-        return;
+    // We need to step up the size if it's small
+    if (nr0 == 1 || nr1 == 1) {
+        chunk_size = 64;
     }
 
-    assert(ne12 % ne02 == 0);
-    assert(ne13 % ne03 == 0);
+    // distribute the work across the inner or outer loop based on which one is larger
+    // The number of chunks in the 0/1 dim.
+    // CEIL(nr0/chunk_size)
+    int64_t nchunk0 = (nr0 + chunk_size - 1) / chunk_size;
+    int64_t nchunk1 = (nr1 + chunk_size - 1) / chunk_size;
 
-    // block-tiling attempt
-    const int64_t blck_0 = 16;
-    const int64_t blck_1 = 16;
+    // If the chunking is poor for the number of threads on this setup, scrap the whole plan.  Re-chunk it by thread.
+    //   Also, chunking by thread was measured to have perform better on NUMA systems.  See https://github.com/ggerganov/llama.cpp/pull/6915
+    //   In theory, chunking should be just as useful on NUMA and non NUMA systems, but testing disagreed with that.
+    if (nchunk0 * nchunk1 < nth * 4 || ggml_is_numa()) {
+        // distribute the thread work across the inner or outer loop based on which one is larger
+        nchunk0 = nr0 > nr1 ? nth : 1; // parallelize by src0 rows
+        nchunk1 = nr0 > nr1 ? 1 : nth; // parallelize by src1 rows
+    }
 
-    // attempt to reduce false-sharing (does not seem to make a difference)
-    float tmp[16];
+    // The number of elements in each chunk
+    const int64_t dr0 = (nr0 + nchunk0 - 1) / nchunk0;
+    const int64_t dr1 = (nr1 + nchunk1 - 1) / nchunk1;
 
-    for (int64_t iir1 = ir110; iir1 < ir111; iir1 += blck_1) {
-        for (int64_t iir0 = ir010; iir0 < ir011; iir0 += blck_0) {
-            for (int64_t ir1 = iir1; ir1 < iir1 + blck_1 && ir1 < ir111; ++ir1) {
-                const int64_t i13 = (ir1/(ne12*ne1));
-                const int64_t i12 = (ir1 - i13*ne12*ne1)/ne1;
-                const int64_t i11 = (ir1 - i13*ne12*ne1 - i12*ne1);
+    if ((ggml_n_dims(src0) == 2) && gemv) {
+        const void * src1_wdata      = (src1->type == vec_dot_type) ? src1->data : params->wdata;
+        const size_t src1_col_stride = ggml_is_contiguous(src1) || src1->type != vec_dot_type ? ggml_row_size(vec_dot_type, ne10) : nb11;
+        int64_t src0_start = (ith * ne01) / nth;
+        int64_t src0_end   = ((ith + 1) * ne01) / nth;
+        src0_start = (src0_start % matmul_num_cols) ? src0_start + matmul_num_cols - (src0_start % matmul_num_cols): src0_start;
+        src0_end   = (src0_end   % matmul_num_cols) ? src0_end   + matmul_num_cols - (src0_end   % matmul_num_cols): src0_end;
+        if (src0_start >= src0_end) return;
 
-                // broadcast src0 into src1
-                const int64_t i03 = i13/r3;
-                const int64_t i02 = i12/r2;
+        // If there are more than three rows in src1, use gemm; otherwise, use gemv.
+        if (gemm && (ne11 > 3)) {
+            gemm(ne00, (float *)((char *) dst->data) + src0_start, ne01, (const char *) src0->data + src0_start * nb01,
+                 (const char *) src1_wdata, ne11 - ne11 % 4, src0_end - src0_start);
+        }
+        for (int iter = gemm ? ne11 - ne11 % 4 : 0; iter < ne11; iter++) {
+            gemv(ne00, (float *)((char *) dst->data + (iter * nb1)) + src0_start, ne01,
+                 (const char *) src0->data + src0_start * nb01, (const char *) src1_wdata + (src1_col_stride * iter), 1,
+                 src0_end - src0_start);
+        }
+        return;
+    }
 
-                const int64_t i1 = i11;
-                const int64_t i2 = i12;
-                const int64_t i3 = i13;
+    // The first chunk comes from our thread_id, the rest will get auto-assigned.
+    int current_chunk = ith;
 
-                const char * src0_row = (const char *) src0->data + (0 + i02*nb02 + i03*nb03);
+    while (current_chunk < nchunk0 * nchunk1) {
+        const int64_t ith0 = current_chunk % nchunk0;
+        const int64_t ith1 = current_chunk / nchunk0;
 
-                // desc: when src1 is not a contiguous memory block we have to calculate the offset using the strides
-                //       if it is, then we have either copied the data to params->wdata and made it contiguous or we are using
-                //       the original src1 data pointer, so we should index using the indices directly
-                // TODO: this is a bit of a hack, we should probably have a better way to handle this
-                const char * src1_col = (const char *) wdata +
-                    (src1_cont || src1->type != vec_dot_type
-                     ? (i11      + i12*ne11 + i13*ne12*ne11)*row_size
-                     : (i11*nb11 + i12*nb12 + i13*nb13));
+        const int64_t ir0_start = dr0 * ith0;
+        const int64_t ir0_end = MIN(ir0_start + dr0, nr0);
 
-                float * dst_col = (float *) ((char *) dst->data + (i1*nb1 + i2*nb2 + i3*nb3));
+        const int64_t ir1_start = dr1 * ith1;
+        const int64_t ir1_end = MIN(ir1_start + dr1, nr1);
 
-                //for (int64_t ir0 = iir0; ir0 < iir0 + blck_0 && ir0 < ir011; ++ir0) {
-                //    vec_dot(ne00, &dst_col[ir0], src0_row + ir0*nb01, src1_col);
-                //}
+        ggml_compute_forward_mul_mat_one_chunk(params, dst, num_rows_per_vec_dot, ir0_start, ir0_end, ir1_start, ir1_end);
 
-                for (int64_t ir0 = iir0; ir0 < iir0 + blck_0 && ir0 < ir011; ++ir0) {
-                    vec_dot(ne00, &tmp[ir0 - iir0], src0_row + ir0*nb01, src1_col);
-                }
-                memcpy(&dst_col[iir0], tmp, (MIN(iir0 + blck_0, ir011) - iir0)*sizeof(float));
-            }
+        if (nth >= nchunk0 * nchunk1) {
+            break;
         }
+
+        current_chunk = atomic_fetch_add_explicit(&params->threadpool->current_chunk, 1, memory_order_relaxed);
     }
 }
 
@@ -9952,11 +12655,11 @@ static void ggml_compute_forward_mul_mat(
 
 static void ggml_compute_forward_mul_mat_id(
         const struct ggml_compute_params * params,
-        const struct ggml_tensor * ids,
-        const struct ggml_tensor * src1,
               struct ggml_tensor * dst) {
 
-    const struct ggml_tensor * src0 = dst->src[2]; // only for GGML_TENSOR_BINARY_OP_LOCALS
+    const struct ggml_tensor * src0 = dst->src[0];
+    const struct ggml_tensor * src1 = dst->src[1];
+    const struct ggml_tensor * ids = dst->src[2];
 
     GGML_TENSOR_BINARY_OP_LOCALS
 
@@ -9967,14 +12670,11 @@ static void ggml_compute_forward_mul_mat_id(
 
     const bool src1_cont = ggml_is_contiguous(src1);
 
-    ggml_vec_dot_t    const vec_dot               = type_traits[type].vec_dot;
-    enum ggml_type    const vec_dot_type          = type_traits[type].vec_dot_type;
-    ggml_from_float_t const from_float_to_vec_dot = type_traits[vec_dot_type].from_float;
-
-    GGML_ASSERT(ne0 == ne01);
-    GGML_ASSERT(ne1 == ne11);
-    GGML_ASSERT(ne2 == ne12);
-    GGML_ASSERT(ne3 == ne13);
+    ggml_vec_dot_t    const vec_dot         = type_traits[type].vec_dot;
+    enum ggml_type    const vec_dot_type    = type_traits[type].vec_dot_type;
+    ggml_from_float_t const from_float      = type_traits[vec_dot_type].from_float;
+    int64_t           const matmul_num_cols = type_traits[type].ncols;
+    ggml_gemv_t       const gemv            = type_traits[type].gemv;
 
     // we don't support permuted src0 or src1
     GGML_ASSERT(nb00 == ggml_type_size(type));
@@ -9986,60 +12686,63 @@ static void ggml_compute_forward_mul_mat_id(
     GGML_ASSERT(nb1 <= nb2);
     GGML_ASSERT(nb2 <= nb3);
 
-    // broadcast factors
-    const int64_t r2 = ne12/ne02;
-    const int64_t r3 = ne13/ne03;
-
     // row groups
-    const int id   = ggml_get_op_params_i32(dst, 0);
-    const int n_as = ggml_get_op_params_i32(dst, 1);
+    const int n_ids = ids->ne[0]; // n_expert_used
+    const int n_as  = ne02;       // n_expert
 
     char * wdata_src1_end = (src1->type == vec_dot_type) ?
             (char *) params->wdata :
             (char *) params->wdata + GGML_PAD(ggml_row_size(vec_dot_type, ggml_nelements(src1)), sizeof(int64_t));
 
-    int64_t * matrix_row_counts = (int64_t *) (wdata_src1_end); // [n_as]
-    int64_t * matrix_rows       = matrix_row_counts + n_as;     // [n_as][ne11]
+    struct mmid_row_mapping {
+        int32_t i1;
+        int32_t i2;
+    };
 
-    #define MMID_MATRIX_ROW(row_id, i1) matrix_rows[(row_id)*ne11 + (i1)]
+    int64_t * matrix_row_counts = (int64_t *) (wdata_src1_end); // [n_as]
+    struct mmid_row_mapping * matrix_rows = (struct mmid_row_mapping *)(matrix_row_counts + n_as); // [n_as][ne11]
 
-   if (params->type == GGML_TASK_INIT) {
+    if (src1->type != vec_dot_type) {
         char * wdata = params->wdata;
-        if (src1->type != vec_dot_type) {
-            const size_t row_size = ggml_row_size(vec_dot_type, ne10);
 
-            assert(params->wsize >= ne11*ne12*ne13*row_size);
-            assert(src1->type == GGML_TYPE_F32);
+        const size_t nbw1 = ggml_row_size(vec_dot_type, ne10);
+        const size_t nbw2 = nbw1*ne11;
+        const size_t nbw3 = nbw2*ne12;
 
-            for (int64_t i13 = 0; i13 < ne13; ++i13) {
-                for (int64_t i12 = 0; i12 < ne12; ++i12) {
-                    for (int64_t i11 = 0; i11 < ne11; ++i11) {
-                        from_float_to_vec_dot((float *)((char *) src1->data + i13*nb13 + i12*nb12 + i11*nb11), (void *) wdata, ne10);
-                        wdata += row_size;
-                    }
+        assert(params->wsize >= ne13*nbw3);
+        GGML_ASSERT(src1->type == GGML_TYPE_F32);
+
+        for (int64_t i13 = 0; i13 < ne13; ++i13) {
+            for (int64_t i12 = 0; i12 < ne12; ++i12) {
+                for (int64_t i11 = ith; i11 < ne11; i11 += nth) {
+                    from_float((float *)((char *) src1->data + i13*nb13 + i12*nb12 + i11*nb11),
+                               (void *)               (wdata + i13*nbw3 + i12*nbw2 + i11*nbw1),
+                               ne10);
                 }
             }
         }
+    }
+
+#define MMID_MATRIX_ROW(row_id, i1) matrix_rows[(row_id)*ne12 + (i1)]
 
+    if (ith == 0) {
         // initialize matrix_row_counts
-        GGML_ASSERT(wdata == wdata_src1_end);
         memset(matrix_row_counts, 0, n_as*sizeof(int64_t));
 
         // group rows by src0 matrix
-        for (int64_t i01 = 0; i01 < ids->ne[1]; i01++) {
-            const int32_t row_id = *(const int32_t *) ((const char *) ids->data + i01*ids->nb[1] + id*ids->nb[0]);
+        for (int64_t iid1 = 0; iid1 < ids->ne[1]; ++iid1) {
+            for (int id = 0; id < n_ids; ++id) {
+                const int32_t i02 = *(const int32_t *) ((const char *) ids->data + iid1*ids->nb[1] + id*ids->nb[0]);
 
-            GGML_ASSERT(row_id >= 0 && row_id < n_as);
-            MMID_MATRIX_ROW(row_id, matrix_row_counts[row_id]) = i01;
-            matrix_row_counts[row_id] += 1;
-        }
+                assert(i02 >= 0 && i02 < n_as);
 
-        return;
+                MMID_MATRIX_ROW(i02, matrix_row_counts[i02]) = (struct mmid_row_mapping) {id, iid1};
+                matrix_row_counts[i02] += 1;
+            }
+        }
     }
 
-    if (params->type == GGML_TASK_FINALIZE) {
-        return;
-    }
+    ggml_barrier(params->threadpool);
 
     // compute each matrix multiplication in sequence
     for (int cur_a = 0; cur_a < n_as; ++cur_a) {
@@ -10049,15 +12752,41 @@ static void ggml_compute_forward_mul_mat_id(
             continue;
         }
 
-        const struct ggml_tensor * src0_cur = dst->src[cur_a + 2];
+        const char * src0_cur = (const char *) src0->data + cur_a*nb02;
 
         const void * wdata    = (src1->type == vec_dot_type) ? src1->data : params->wdata;
         const size_t row_size = ggml_row_size(vec_dot_type, ne10);
 
-        const int64_t nr0 = ne01;           // src0 rows
-        const int64_t nr1 = cne1*ne12*ne13; // src1 rows
+        const int64_t nr0 = ne01; // src0 rows
+        const int64_t nr1 = cne1; // src1 rows
 
-        //printf("nr0 = %lld, nr1 = %lld\n", nr0, nr1);
+        if (((ggml_n_dims(src0) - 1) == 2) && gemv) {
+            int64_t src0_cur_start = (ith * ne01) / nth;
+            int64_t src0_cur_end   = ((ith + 1) * ne01) / nth;
+            src0_cur_start = (src0_cur_start % matmul_num_cols) ? src0_cur_start + matmul_num_cols - (src0_cur_start % matmul_num_cols): src0_cur_start;
+            src0_cur_end   = (src0_cur_end % matmul_num_cols) ? src0_cur_end + matmul_num_cols - (src0_cur_end % matmul_num_cols): src0_cur_end;
+            if (src0_cur_start >= src0_cur_end) return;
+
+            for (int ir1 = 0; ir1 < nr1; ir1++) {
+                struct mmid_row_mapping row_mapping = MMID_MATRIX_ROW(cur_a, ir1);
+                const int id       = row_mapping.i1; // selected expert index
+
+                const int64_t  i11 = id % ne11;
+                const int64_t  i12 = row_mapping.i2; // row index in src1
+
+                const int64_t  i1 = id;  // selected expert index
+                const int64_t  i2 = i12; // row
+
+                const char * src1_col = (const char *) wdata +
+                    (src1_cont || src1->type != vec_dot_type
+                    ? (i11        + i12 * ne11) * row_size
+                    : (i11 * nb11 + i12 * nb12));
+
+                gemv(ne00, (float *)((char *) dst->data + (i1 * nb1 + i2 * nb2)) + src0_cur_start, ne01,
+                     (const char *) src0_cur + src0_cur_start * nb01, src1_col, 1, src0_cur_end - src0_cur_start);
+            }
+            continue;
+        }
 
         // distribute the thread work across the inner or outer loop based on which one is larger
 
@@ -10076,16 +12805,11 @@ static void ggml_compute_forward_mul_mat_id(
         const int64_t ir110 = dr1*ith1;
         const int64_t ir111 = MIN(ir110 + dr1, nr1);
 
-        //printf("ir010 = %6lld, ir011 = %6lld, ir110 = %6lld, ir111 = %6lld\n", ir010, ir011, ir110, ir111);
-
         // threads with no work simply yield (not sure if it helps)
-        if (ir010 >= ir011 || ir110 >= ir111) {
-            sched_yield();
-            continue;
-        }
-
-        assert(ne12 % ne02 == 0);
-        assert(ne13 % ne03 == 0);
+        //if (ir010 >= ir011 || ir110 >= ir111) {
+        //    sched_yield();
+        //    continue;
+        //}
 
         // block-tiling attempt
         const int64_t blck_0 = 16;
@@ -10097,20 +12821,16 @@ static void ggml_compute_forward_mul_mat_id(
         for (int64_t iir1 = ir110; iir1 < ir111; iir1 += blck_1) {
             for (int64_t iir0 = ir010; iir0 < ir011; iir0 += blck_0) {
                 for (int64_t ir1 = iir1; ir1 < iir1 + blck_1 && ir1 < ir111; ++ir1) {
-                    const int64_t  i13 = (ir1/(ne12*cne1)); // Note: currently, src1 is always a matrix
-                    const int64_t  i12 = (ir1 - i13*ne12*cne1)/cne1;
-                    const int64_t _i11 = (ir1 - i13*ne12*cne1 - i12*cne1);
-                    const int64_t  i11 = MMID_MATRIX_ROW(cur_a, _i11);
+                    const int64_t _i12 = ir1; // logical row index for this expert
 
-                    // broadcast src0 into src1
-                    const int64_t i03 = i13/r3;
-                    const int64_t i02 = i12/r2;
+                    struct mmid_row_mapping row_mapping = MMID_MATRIX_ROW(cur_a, _i12);
+                    const int id       = row_mapping.i1; // selected expert index
 
-                    const int64_t i1 = i11;
-                    const int64_t i2 = i12;
-                    const int64_t i3 = i13;
+                    const int64_t  i11 = id % ne11;
+                    const int64_t  i12 = row_mapping.i2; // row index in src1
 
-                    const char * src0_row = (const char *) src0_cur->data + (0 + i02*nb02 + i03*nb03);
+                    const int64_t  i1 = id;  // selected expert index
+                    const int64_t  i2 = i12; // row
 
                     // desc: when src1 is not a contiguous memory block we have to calculate the offset using the strides
                     //       if it is, then we have either copied the data to params->wdata and made it contiguous or we are using
@@ -10118,39 +12838,43 @@ static void ggml_compute_forward_mul_mat_id(
                     // TODO: this is a bit of a hack, we should probably have a better way to handle this
                     const char * src1_col = (const char *) wdata +
                         (src1_cont || src1->type != vec_dot_type
-                        ? (i11      + i12*ne11 + i13*ne12*ne11)*row_size
-                        : (i11*nb11 + i12*nb12 + i13*nb13));
+                        ? (i11      + i12*ne11)*row_size
+                        : (i11*nb11 + i12*nb12));
 
-                    float * dst_col = (float *) ((char *) dst->data + (i1*nb1 + i2*nb2 + i3*nb3));
+                    float * dst_col = (float *) ((char *) dst->data + (i1*nb1 + i2*nb2));
 
                     //for (int64_t ir0 = iir0; ir0 < iir0 + blck_0 && ir0 < ir011; ++ir0) {
                     //    vec_dot(ne00, &dst_col[ir0], src0_row + ir0*nb01, src1_col);
                     //}
 
                     for (int64_t ir0 = iir0; ir0 < iir0 + blck_0 && ir0 < ir011; ++ir0) {
-                        vec_dot(ne00, &tmp[ir0 - iir0], src0_row + ir0*nb01, src1_col);
+                        vec_dot(ne00, &tmp[ir0 - iir0], 0, src0_cur + ir0*nb01, 0, src1_col, 0, 1);
                     }
+
                     memcpy(&dst_col[iir0], tmp, (MIN(iir0 + blck_0, ir011) - iir0)*sizeof(float));
                 }
             }
         }
     }
 
-    #undef MMID_MATRIX_ROW
+#undef MMID_MATRIX_ROW
 }
 
 // ggml_compute_forward_out_prod
 
 static void ggml_compute_forward_out_prod_f32(
         const struct ggml_compute_params * params,
-        const struct ggml_tensor * src0,
-        const struct ggml_tensor * src1,
               struct ggml_tensor * dst) {
-    // int64_t t0 = ggml_perf_time_us();
-    // UNUSED(t0);
+
+    const struct ggml_tensor * src0 = dst->src[0];
+    const struct ggml_tensor * src1 = dst->src[1];
 
     GGML_TENSOR_BINARY_OP_LOCALS
 
+    GGML_ASSERT(dst->type == GGML_TYPE_F32);
+    GGML_ASSERT(src0->type == GGML_TYPE_F32);
+    GGML_ASSERT(src1->type == GGML_TYPE_F32);
+
     const int ith = params->ith;
     const int nth = params->nth;
 
@@ -10173,73 +12897,10 @@ static void ggml_compute_forward_out_prod_f32(
     // nb01 >= nb00 - src0 is not transposed
     //   compute by src0 rows
 
-    // TODO: #if defined(GGML_USE_CUBLAS) ggml_cuda_out_prod
-    // TODO: #if defined(GGML_USE_CLBLAST)
-
-#if defined(GGML_USE_ACCELERATE) || defined(GGML_USE_OPENBLAS)
-    bool use_blas = ggml_is_matrix(src0) &&
-        ggml_is_matrix(src1) &&
-        ggml_is_contiguous(src0) &&
-        (ggml_is_contiguous(src1) || ggml_is_transposed(src1));
-#endif
-
-    if (params->type == GGML_TASK_INIT) {
-#if defined(GGML_USE_ACCELERATE) || defined(GGML_USE_OPENBLAS) // gemm beta will zero dst
-        if (use_blas) {
-            return;
-        }
-#endif
+    if (ith == 0) {
         ggml_vec_set_f32(ne0*ne1*ne2*ne3, dst->data, 0);
-        return;
-    }
-
-    if (params->type == GGML_TASK_FINALIZE) {
-        return;
-    }
-
-#if defined(GGML_USE_ACCELERATE) || defined(GGML_USE_OPENBLAS)
-    if (use_blas) {
-        if (params->ith != 0) { // All threads other than the first do no work.
-            return;
-        }
-        // Arguments to ggml_compute_forward_out_prod (expressed as major,minor)
-        // src0: (k,n)
-        // src1: (k,m)
-        // dst:  (m,n)
-        //
-        // Arguments to sgemm (see https://github.com/Reference-LAPACK/lapack/blob/master/BLAS/SRC/sgemm.f)
-        // Also expressed as (major,minor)
-        // a: (m,k): so src1 transposed
-        // b: (k,n): so src0
-        // c: (m,n)
-        //
-        // However, if ggml_is_transposed(src1) is true, then
-        // src1->data already contains a transposed version, so sgemm mustn't
-        // transpose it further.
-
-        int n = src0->ne[0];
-        int k = src0->ne[1];
-        int m = src1->ne[0];
-
-        int transposeA, lda;
-
-        if (!ggml_is_transposed(src1)) {
-            transposeA = CblasTrans;
-            lda = m;
-        } else {
-            transposeA = CblasNoTrans;
-            lda = k;
-        }
-
-        float * a = (float *) ((char *) src1->data);
-        float * b = (float *) ((char *) src0->data);
-        float * c = (float *) ((char *) dst->data);
-
-        cblas_sgemm(CblasRowMajor, transposeA, CblasNoTrans, m, n, k, 1.0, a, lda, b, n, 0.0, c, n);
-
-        return;
     }
-#endif
+    ggml_barrier(params->threadpool);
 
     // dst[:,:,:,:] = 0
     // for i2,i3:
@@ -10315,28 +12976,14 @@ static void ggml_compute_forward_out_prod_f32(
             }
         }
     }
-
-    //int64_t t1 = ggml_perf_time_us();
-    //static int64_t acc = 0;
-    //acc += t1 - t0;
-    //if (t1 - t0 > 10) {
-    //    Rprintf("\n");
-    //    Rprintf("ne00 = %5d, ne01 = %5d, ne02 = %5d, ne03 = %5d\n", ne00, ne01, ne02, ne03);
-    //    Rprintf("nb00 = %5d, nb01 = %5d, nb02 = %5d, nb03 = %5d\n", nb00, nb01, nb02, nb03);
-    //    Rprintf("ne10 = %5d, ne11 = %5d, ne12 = %5d, ne13 = %5d\n", ne10, ne11, ne12, ne13);
-    //    Rprintf("nb10 = %5d, nb11 = %5d, nb12 = %5d, nb13 = %5d\n", nb10, nb11, nb12, nb13);
-
-    //    Rprintf("XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX task %d/%d: %d us, acc = %d\n", ith, nth, (int) (t1 - t0), (int) acc);
-    //}
 }
 
 static void ggml_compute_forward_out_prod_q_f32(
         const struct ggml_compute_params * params,
-        const struct ggml_tensor * src0,
-        const struct ggml_tensor * src1,
               struct ggml_tensor * dst) {
-    // int64_t t0 = ggml_perf_time_us();
-    // UNUSED(t0);
+
+    const struct ggml_tensor * src0 = dst->src[0];
+    const struct ggml_tensor * src1 = dst->src[1];
 
     GGML_TENSOR_BINARY_OP_LOCALS;
 
@@ -10368,17 +13015,10 @@ static void ggml_compute_forward_out_prod_q_f32(
     // nb01 >= nb00 - src0 is not transposed
     //   compute by src0 rows
 
-    // TODO: #if defined(GGML_USE_CUBLAS) ggml_cuda_out_prod
-    // TODO: #if defined(GGML_USE_ACCELERATE) || defined(GGML_USE_OPENBLAS) || defined(GGML_USE_CLBLAST)
-
-    if (params->type == GGML_TASK_INIT) {
+    if (ith == 0) {
         ggml_vec_set_f32(ne0*ne1*ne2*ne3, dst->data, 0);
-        return;
-    }
-
-    if (params->type == GGML_TASK_FINALIZE) {
-        return;
     }
+    ggml_barrier(params->threadpool);
 
     // parallelize by last three dimensions
 
@@ -10425,26 +13065,14 @@ static void ggml_compute_forward_out_prod_q_f32(
             ggml_vec_mad_f32(ne0, d, wdata, *s1);
         }
     }
-
-    //int64_t t1 = ggml_perf_time_us();
-    //static int64_t acc = 0;
-    //acc += t1 - t0;
-    //if (t1 - t0 > 10) {
-    //    Rprintf("\n");
-    //    Rprintf("ne00 = %5d, ne01 = %5d, ne02 = %5d, ne03 = %5d\n", ne00, ne01, ne02, ne03);
-    //    Rprintf("nb00 = %5d, nb01 = %5d, nb02 = %5d, nb03 = %5d\n", nb00, nb01, nb02, nb03);
-    //    Rprintf("ne10 = %5d, ne11 = %5d, ne12 = %5d, ne13 = %5d\n", ne10, ne11, ne12, ne13);
-    //    Rprintf("nb10 = %5d, nb11 = %5d, nb12 = %5d, nb13 = %5d\n", nb10, nb11, nb12, nb13);
-
-    //    Rprintf("XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX task %d/%d: %d us, acc = %d\n", ith, nth, (int) (t1 - t0), (int) acc);
-    //}
 }
 
 static void ggml_compute_forward_out_prod(
         const struct ggml_compute_params * params,
-        const struct ggml_tensor * src0,
-        const struct ggml_tensor * src1,
         struct ggml_tensor * dst) {
+
+    const struct ggml_tensor * src0 = dst->src[0];
+
     switch (src0->type) {
         case GGML_TYPE_Q4_0:
         case GGML_TYPE_Q4_1:
@@ -10456,22 +13084,36 @@ static void ggml_compute_forward_out_prod(
         case GGML_TYPE_Q4_K:
         case GGML_TYPE_Q5_K:
         case GGML_TYPE_Q6_K:
-            {
-                ggml_compute_forward_out_prod_q_f32(params, src0, src1, dst);
+        case GGML_TYPE_TQ1_0:
+        case GGML_TYPE_TQ2_0:
+        case GGML_TYPE_IQ2_XXS:
+        case GGML_TYPE_IQ2_XS:
+        case GGML_TYPE_IQ3_XXS:
+        case GGML_TYPE_IQ1_S:
+        case GGML_TYPE_IQ1_M:
+        case GGML_TYPE_IQ4_NL:
+        case GGML_TYPE_IQ4_XS:
+        case GGML_TYPE_IQ3_S:
+        case GGML_TYPE_IQ2_S:
+        case GGML_TYPE_Q4_0_4_4:
+        case GGML_TYPE_Q4_0_4_8:
+        case GGML_TYPE_Q4_0_8_8:
+            {
+                ggml_compute_forward_out_prod_q_f32(params, dst);
             } break;
         case GGML_TYPE_F16:
             {
-                GGML_ASSERT(false); // todo
-                // ggml_compute_forward_out_prod_f16_f32(params, src0, src1, dst);
-            } break;
+                GGML_ABORT("fatal error"); // todo
+                // ggml_compute_forward_out_prod_f16_f32(params, dst);
+            }
         case GGML_TYPE_F32:
             {
-                ggml_compute_forward_out_prod_f32(params, src0, src1, dst);
+                ggml_compute_forward_out_prod_f32(params, dst);
             } break;
         default:
             {
-                GGML_ASSERT(false);
-            } break;
+                GGML_ABORT("fatal error");
+            }
     }
 }
 
@@ -10479,16 +13121,14 @@ static void ggml_compute_forward_out_prod(
 
 static void ggml_compute_forward_scale_f32(
         const struct ggml_compute_params * params,
-        const struct ggml_tensor * src0,
         struct ggml_tensor * dst) {
+
+    const struct ggml_tensor * src0 = dst->src[0];
+
     GGML_ASSERT(ggml_is_contiguous(src0));
     GGML_ASSERT(ggml_is_contiguous(dst));
     GGML_ASSERT(ggml_are_same_shape(src0, dst));
 
-    if (params->type == GGML_TASK_INIT || params->type == GGML_TASK_FINALIZE) {
-        return;
-    }
-
     // scale factor
     float v;
     memcpy(&v, dst->op_params, sizeof(float));
@@ -10521,17 +13161,19 @@ static void ggml_compute_forward_scale_f32(
 
 static void ggml_compute_forward_scale(
         const struct ggml_compute_params * params,
-        const struct ggml_tensor * src0,
         struct ggml_tensor * dst) {
+
+    const struct ggml_tensor * src0 = dst->src[0];
+
     switch (src0->type) {
         case GGML_TYPE_F32:
             {
-                ggml_compute_forward_scale_f32(params, src0, dst);
+                ggml_compute_forward_scale_f32(params, dst);
             } break;
         default:
             {
-                GGML_ASSERT(false);
-            } break;
+                GGML_ABORT("fatal error");
+            }
     }
 }
 
@@ -10539,9 +13181,11 @@ static void ggml_compute_forward_scale(
 
 static void ggml_compute_forward_set_f32(
         const struct ggml_compute_params * params,
-        const struct ggml_tensor * src0,
-        const struct ggml_tensor * src1,
         struct ggml_tensor * dst) {
+
+    const struct ggml_tensor * src0 = dst->src[0];
+    const struct ggml_tensor * src1 = dst->src[1];
+
     GGML_ASSERT(ggml_are_same_shape(src0, dst));
     GGML_ASSERT(ggml_is_contiguous(dst) && ggml_is_contiguous(src0));
 
@@ -10553,17 +13197,16 @@ static void ggml_compute_forward_set_f32(
     size_t offset  = ((int32_t *) dst->op_params)[3];
     bool   inplace = (bool) ((int32_t *) dst->op_params)[4];
 
-    if (!inplace && (params->type == GGML_TASK_INIT)) {
-        // memcpy needs to be synchronized across threads to avoid race conditions.
-        // => do it in INIT phase
-        memcpy(
-            ((char *)  dst->data),
-            ((char *) src0->data),
-            ggml_nbytes(dst));
-    }
-
-    if (params->type == GGML_TASK_INIT || params->type == GGML_TASK_FINALIZE) {
-        return;
+    if (!inplace) {
+        if (params->ith == 0) {
+            // memcpy needs to be synchronized across threads to avoid race conditions.
+            // => do it in INIT phase
+            memcpy(
+                ((char *)  dst->data),
+                ((char *) src0->data),
+                ggml_nbytes(dst));
+        }
+        ggml_barrier(params->threadpool);
     }
 
     const int ith = params->ith;
@@ -10609,16 +13252,17 @@ static void ggml_compute_forward_set_f32(
 
 static void ggml_compute_forward_set(
         const struct ggml_compute_params * params,
-        const struct ggml_tensor * src0,
-        const struct ggml_tensor * src1,
         struct ggml_tensor * dst) {
 
+    const struct ggml_tensor * src0 = dst->src[0];
+
     switch (src0->type) {
         case GGML_TYPE_F32:
             {
-                ggml_compute_forward_set_f32(params, src0, src1, dst);
+                ggml_compute_forward_set_f32(params, dst);
             } break;
         case GGML_TYPE_F16:
+        case GGML_TYPE_BF16:
         case GGML_TYPE_Q4_0:
         case GGML_TYPE_Q4_1:
         case GGML_TYPE_Q5_0:
@@ -10630,10 +13274,24 @@ static void ggml_compute_forward_set(
         case GGML_TYPE_Q4_K:
         case GGML_TYPE_Q5_K:
         case GGML_TYPE_Q6_K:
+        case GGML_TYPE_TQ1_0:
+        case GGML_TYPE_TQ2_0:
+        case GGML_TYPE_IQ2_XXS:
+        case GGML_TYPE_IQ2_XS:
+        case GGML_TYPE_IQ3_XXS:
+        case GGML_TYPE_IQ1_S:
+        case GGML_TYPE_IQ1_M:
+        case GGML_TYPE_IQ4_NL:
+        case GGML_TYPE_IQ4_XS:
+        case GGML_TYPE_IQ3_S:
+        case GGML_TYPE_IQ2_S:
+        case GGML_TYPE_Q4_0_4_4:
+        case GGML_TYPE_Q4_0_4_8:
+        case GGML_TYPE_Q4_0_8_8:
         default:
             {
-                GGML_ASSERT(false);
-            } break;
+                GGML_ABORT("fatal error");
+            }
     }
 }
 
@@ -10641,29 +13299,25 @@ static void ggml_compute_forward_set(
 
 static void ggml_compute_forward_cpy(
         const struct ggml_compute_params * params,
-        const struct ggml_tensor * src0,
         struct ggml_tensor * dst) {
-    ggml_compute_forward_dup(params, src0, dst);
+    ggml_compute_forward_dup(params, dst);
 }
 
 // ggml_compute_forward_cont
 
 static void ggml_compute_forward_cont(
         const struct ggml_compute_params * params,
-        const struct ggml_tensor * src0,
         struct ggml_tensor * dst) {
-    ggml_compute_forward_dup(params, src0, dst);
+    ggml_compute_forward_dup(params, dst);
 }
 
 // ggml_compute_forward_reshape
 
 static void ggml_compute_forward_reshape(
         const struct ggml_compute_params * params,
-        const struct ggml_tensor * src0,
         struct ggml_tensor * dst) {
     // NOP
     UNUSED(params);
-    UNUSED(src0);
     UNUSED(dst);
 }
 
@@ -10671,49 +13325,45 @@ static void ggml_compute_forward_reshape(
 
 static void ggml_compute_forward_view(
         const struct ggml_compute_params * params,
-        const struct ggml_tensor * src0) {
+        const struct ggml_tensor * dst) {
     // NOP
     UNUSED(params);
-    UNUSED(src0);
+    UNUSED(dst);
 }
 
 // ggml_compute_forward_permute
 
 static void ggml_compute_forward_permute(
         const struct ggml_compute_params * params,
-        const struct ggml_tensor * src0) {
+        const struct ggml_tensor * dst) {
     // NOP
     UNUSED(params);
-    UNUSED(src0);
+    UNUSED(dst);
 }
 
 // ggml_compute_forward_transpose
 
 static void ggml_compute_forward_transpose(
         const struct ggml_compute_params * params,
-        const struct ggml_tensor * src0) {
+        const struct ggml_tensor * dst) {
     // NOP
     UNUSED(params);
-    UNUSED(src0);
+    UNUSED(dst);
 }
 
 // ggml_compute_forward_get_rows
 
 static void ggml_compute_forward_get_rows_q(
         const struct ggml_compute_params * params,
-        const struct ggml_tensor * src0,
-        const struct ggml_tensor * src1,
               struct ggml_tensor * dst) {
-    assert(params->ith == 0);
 
-    if (params->type == GGML_TASK_INIT || params->type == GGML_TASK_FINALIZE) {
-        return;
-    }
+    const struct ggml_tensor * src0 = dst->src[0];
+    const struct ggml_tensor * src1 = dst->src[1];
 
     GGML_TENSOR_BINARY_OP_LOCALS
 
     const int64_t nc = ne00;
-    const int64_t nr = ggml_nelements(src1); GGML_UNUSED(nr);
+    const int64_t nr = ggml_nelements(src1);
 
     const enum ggml_type type = src0->type;
     ggml_to_float_t const dequantize_row_q = type_traits[type].to_float;
@@ -10723,95 +13373,159 @@ static void ggml_compute_forward_get_rows_q(
     assert(nb00 == ggml_type_size(type));
     assert(ggml_nrows(dst) == nr);
 
-    // TODO: multi-thread
-    for (int64_t i12 = 0; i12 < ne12; ++i12) {
-        for (int64_t i11 = 0; i11 < ne11; ++i11) {
-            for (int64_t i10 = 0; i10 < ne10; ++i10) {
-                const int64_t i01 = *(int32_t *) ((char *) src1->data + i10*nb10 + i11*nb11 + i12*nb12);
+    const int ith = params->ith;
+    const int nth = params->nth;
 
-                dequantize_row_q(
-                        (const void *) ((char *) src0->data + i01*nb01 + i11*nb02 + i12*nb03),
-                             (float *) ((char *)  dst->data + i10*nb1  + i11*nb2  + i12*nb3), nc);
-            }
-        }
+    // rows per thread
+    const int dr = (nr + nth - 1)/nth;
+
+    // row range for this thread
+    const int ir0 = dr*ith;
+    const int ir1 = MIN(ir0 + dr, nr);
+
+    for (int64_t i = ir0; i < ir1; ++i) {
+        const int64_t i12 = i/(ne11*ne10);
+        const int64_t i11 = (i - i12*ne11*ne10)/ne10;
+        const int64_t i10 = (i - i12*ne11*ne10 - i11*ne10);
+        const int64_t i01 = *(int32_t *) ((char *) src1->data + i10*nb10 + i11*nb11 + i12*nb12);
+
+        GGML_ASSERT(i01 >= 0 && i01 < ne01);
+
+        dequantize_row_q(
+                (const void *) ((char *) src0->data + i01*nb01 + i11*nb02 + i12*nb03),
+                     (float *) ((char *)  dst->data + i10*nb1  + i11*nb2  + i12*nb3), nc);
     }
 }
 
 static void ggml_compute_forward_get_rows_f16(
         const struct ggml_compute_params * params,
-        const struct ggml_tensor * src0,
-        const struct ggml_tensor * src1,
               struct ggml_tensor * dst) {
-    assert(params->ith == 0);
 
-    if (params->type == GGML_TASK_INIT || params->type == GGML_TASK_FINALIZE) {
-        return;
-    }
+    const struct ggml_tensor * src0 = dst->src[0];
+    const struct ggml_tensor * src1 = dst->src[1];
 
     GGML_TENSOR_BINARY_OP_LOCALS
 
     const int64_t nc = ne00;
-    const int64_t nr = ggml_nelements(src1); GGML_UNUSED(nr);
+    const int64_t nr = ggml_nelements(src1);
 
     assert(ne0  == nc);
     assert(ne02 == ne11);
     assert(nb00 == sizeof(ggml_fp16_t));
     assert(ggml_nrows(dst) == nr);
 
-    // TODO: multi-thread
-    for (int64_t i12 = 0; i12 < ne12; ++i12) {
-        for (int64_t i11 = 0; i11 < ne11; ++i11) {
-            for (int64_t i10 = 0; i10 < ne10; ++i10) {
-                const int64_t i01 = *(int32_t *) ((char *) src1->data + i10*nb10 + i11*nb11 + i12*nb12);
+    const int ith = params->ith;
+    const int nth = params->nth;
+
+    // rows per thread
+    const int dr = (nr + nth - 1)/nth;
 
-                ggml_fp16_to_fp32_row(
-                        (const void *) ((char *) src0->data + i01*nb01 + i11*nb02 + i12*nb03),
-                             (float *) ((char *)  dst->data + i10*nb1  + i11*nb2  + i12*nb3), nc);
-            }
-        }
+    // row range for this thread
+    const int ir0 = dr*ith;
+    const int ir1 = MIN(ir0 + dr, nr);
+
+    for (int64_t i = ir0; i < ir1; ++i) {
+        const int64_t i12 = i/(ne11*ne10);
+        const int64_t i11 = (i - i12*ne11*ne10)/ne10;
+        const int64_t i10 = (i - i12*ne11*ne10 - i11*ne10);
+        const int64_t i01 = *(int32_t *) ((char *) src1->data + i10*nb10 + i11*nb11 + i12*nb12);
+
+        GGML_ASSERT(i01 >= 0 && i01 < ne01);
+
+        ggml_fp16_to_fp32_row(
+                (const void *) ((char *) src0->data + i01*nb01 + i11*nb02 + i12*nb03),
+                     (float *) ((char *)  dst->data + i10*nb1  + i11*nb2  + i12*nb3), nc);
     }
 }
 
-static void ggml_compute_forward_get_rows_f32(
+static void ggml_compute_forward_get_rows_bf16(
         const struct ggml_compute_params * params,
-        const struct ggml_tensor * src0,
-        const struct ggml_tensor * src1,
               struct ggml_tensor * dst) {
-    assert(params->ith == 0);
 
-    if (params->type == GGML_TASK_INIT || params->type == GGML_TASK_FINALIZE) {
-        return;
+    const struct ggml_tensor * src0 = dst->src[0];
+    const struct ggml_tensor * src1 = dst->src[1];
+
+    GGML_TENSOR_BINARY_OP_LOCALS
+
+    const int64_t nc = ne00;
+    const int64_t nr = ggml_nelements(src1);
+
+    assert(ne0  == nc);
+    assert(ne02 == ne11);
+    assert(nb00 == sizeof(ggml_bf16_t));
+    assert(ggml_nrows(dst) == nr);
+
+    const int ith = params->ith;
+    const int nth = params->nth;
+
+    // rows per thread
+    const int dr = (nr + nth - 1)/nth;
+
+    // row range for this thread
+    const int ir0 = dr*ith;
+    const int ir1 = MIN(ir0 + dr, nr);
+
+    for (int64_t i = ir0; i < ir1; ++i) {
+        const int64_t i12 = i/(ne11*ne10);
+        const int64_t i11 = (i - i12*ne11*ne10)/ne10;
+        const int64_t i10 = (i - i12*ne11*ne10 - i11*ne10);
+        const int64_t i01 = *(int32_t *) ((char *) src1->data + i10*nb10 + i11*nb11 + i12*nb12);
+
+        GGML_ASSERT(i01 >= 0 && i01 < ne01);
+
+        ggml_bf16_to_fp32_row(
+                (const void *) ((char *) src0->data + i01*nb01 + i11*nb02 + i12*nb03),
+                     (float *) ((char *)  dst->data + i10*nb1  + i11*nb2  + i12*nb3), nc);
     }
+}
+
+static void ggml_compute_forward_get_rows_f32(
+        const struct ggml_compute_params * params,
+              struct ggml_tensor * dst) {
+
+    const struct ggml_tensor * src0 = dst->src[0];
+    const struct ggml_tensor * src1 = dst->src[1];
 
     GGML_TENSOR_BINARY_OP_LOCALS
 
     const int64_t nc = ne00;
-    const int64_t nr = ggml_nelements(src1); GGML_UNUSED(nr);
+    const int64_t nr = ggml_nelements(src1);
 
     assert(ne0  == nc);
     assert(ne02 == ne11);
     assert(nb00 == sizeof(float));
     assert(ggml_nrows(dst) == nr);
 
-    // TODO: multi-thread
-    for (int64_t i12 = 0; i12 < ne12; ++i12) {
-        for (int64_t i11 = 0; i11 < ne11; ++i11) {
-            for (int64_t i10 = 0; i10 < ne10; ++i10) {
-                const int64_t i01 = *(int32_t *) ((char *) src1->data + i10*nb10 + i11*nb11 + i12*nb12);
+    const int ith = params->ith;
+    const int nth = params->nth;
 
-                ggml_vec_cpy_f32(nc,
-                        (float *) ((char *)  dst->data + i10*nb1  + i11*nb2  + i12*nb3),
-                        (float *) ((char *) src0->data + i01*nb01 + i11*nb02 + i12*nb03));
-            }
-        }
+    // rows per thread
+    const int dr = (nr + nth - 1)/nth;
+
+    // row range for this thread
+    const int ir0 = dr*ith;
+    const int ir1 = MIN(ir0 + dr, nr);
+
+    for (int64_t i = ir0; i < ir1; ++i) {
+        const int64_t i12 = i/(ne11*ne10);
+        const int64_t i11 = (i - i12*ne11*ne10)/ne10;
+        const int64_t i10 = (i - i12*ne11*ne10 - i11*ne10);
+        const int64_t i01 = *(int32_t *) ((char *) src1->data + i10*nb10 + i11*nb11 + i12*nb12);
+
+        GGML_ASSERT(i01 >= 0 && i01 < ne01);
+
+        ggml_vec_cpy_f32(nc,
+                (float *) ((char *)  dst->data + i10*nb1  + i11*nb2  + i12*nb3),
+                (float *) ((char *) src0->data + i01*nb01 + i11*nb02 + i12*nb03));
     }
 }
 
 static void ggml_compute_forward_get_rows(
         const struct ggml_compute_params * params,
-        const struct ggml_tensor * src0,
-        const struct ggml_tensor * src1,
         struct ggml_tensor * dst) {
+
+    const struct ggml_tensor * src0 = dst->src[0];
+
     switch (src0->type) {
         case GGML_TYPE_Q4_0:
         case GGML_TYPE_Q4_1:
@@ -10824,22 +13538,40 @@ static void ggml_compute_forward_get_rows(
         case GGML_TYPE_Q4_K:
         case GGML_TYPE_Q5_K:
         case GGML_TYPE_Q6_K:
-            {
-                ggml_compute_forward_get_rows_q(params, src0, src1, dst);
+        case GGML_TYPE_TQ1_0:
+        case GGML_TYPE_TQ2_0:
+        case GGML_TYPE_IQ2_XXS:
+        case GGML_TYPE_IQ2_XS:
+        case GGML_TYPE_IQ3_XXS:
+        case GGML_TYPE_IQ1_S:
+        case GGML_TYPE_IQ1_M:
+        case GGML_TYPE_IQ4_NL:
+        case GGML_TYPE_IQ4_XS:
+        case GGML_TYPE_IQ3_S:
+        case GGML_TYPE_IQ2_S:
+        case GGML_TYPE_Q4_0_4_4:
+        case GGML_TYPE_Q4_0_4_8:
+        case GGML_TYPE_Q4_0_8_8:
+            {
+                ggml_compute_forward_get_rows_q(params, dst);
             } break;
         case GGML_TYPE_F16:
             {
-                ggml_compute_forward_get_rows_f16(params, src0, src1, dst);
+                ggml_compute_forward_get_rows_f16(params, dst);
+            } break;
+        case GGML_TYPE_BF16:
+            {
+                ggml_compute_forward_get_rows_bf16(params, dst);
             } break;
         case GGML_TYPE_F32:
         case GGML_TYPE_I32:
             {
-                ggml_compute_forward_get_rows_f32(params, src0, src1, dst);
+                ggml_compute_forward_get_rows_f32(params, dst);
             } break;
         default:
             {
-                GGML_ASSERT(false);
-            } break;
+                GGML_ABORT("fatal error");
+            }
     }
 
     //static bool first = true;
@@ -10850,14 +13582,14 @@ static void ggml_compute_forward_get_rows(
     //    for (int k = 0; k < dst->ne[1]; ++k) {
     //        for (int j = 0; j < dst->ne[0]/16; ++j) {
     //            for (int i = 0; i < 16; ++i) {
-    //                Rprintf("%8.4f ", ((float *) dst->data)[k*dst->ne[0] + j*16 + i]);
+    //                printf("%8.4f ", ((float *) dst->data)[k*dst->ne[0] + j*16 + i]);
     //            }
-    //            Rprintf("\n");
+    //            printf("\n");
     //        }
-    //        Rprintf("\n");
+    //        printf("\n");
     //    }
-    //    Rprintf("\n");
-    //    Rf_error("whispercpp error");
+    //    printf("\n");
+    //    exit(0);
     //}
 }
 
@@ -10865,22 +13597,21 @@ static void ggml_compute_forward_get_rows(
 
 static void ggml_compute_forward_get_rows_back_f32_f16(
         const struct ggml_compute_params * params,
-        const struct ggml_tensor * src0,
-        const struct ggml_tensor * src1,
               struct ggml_tensor * dst) {
-    GGML_ASSERT(params->ith == 0);
-    GGML_ASSERT(ggml_is_contiguous(dst));
 
-    // ggml_compute_forward_dup_same_cont(params, opt0, dst);
-
-    if (params->type == GGML_TASK_INIT) {
-        memset(dst->data, 0, ggml_nbytes(dst));
-    }
+    const struct ggml_tensor * src0 = dst->src[0];
+    const struct ggml_tensor * src1 = dst->src[1];
 
-    if (params->type == GGML_TASK_INIT || params->type == GGML_TASK_FINALIZE) {
+    if (params->ith != 0) {
         return;
     }
 
+    GGML_ASSERT(ggml_is_contiguous(dst));
+
+    // ggml_compute_forward_dup_same_cont(params, opt0, dst);
+
+    memset(dst->data, 0, ggml_nbytes(dst));
+
     const int nc = src0->ne[0];
     const int nr = ggml_nelements(src1);
 
@@ -10899,22 +13630,21 @@ static void ggml_compute_forward_get_rows_back_f32_f16(
 
 static void ggml_compute_forward_get_rows_back_f32(
         const struct ggml_compute_params * params,
-        const struct ggml_tensor * src0,
-        const struct ggml_tensor * src1,
               struct ggml_tensor * dst) {
-    GGML_ASSERT(params->ith == 0);
-    GGML_ASSERT(ggml_is_contiguous(dst));
-
-    // ggml_compute_forward_dup_same_cont(params, opt0, dst);
 
-    if (params->type == GGML_TASK_INIT) {
-        memset(dst->data, 0, ggml_nbytes(dst));
-    }
+    const struct ggml_tensor * src0 = dst->src[0];
+    const struct ggml_tensor * src1 = dst->src[1];
 
-    if (params->type == GGML_TASK_INIT || params->type == GGML_TASK_FINALIZE) {
+    if (params->ith != 0) {
         return;
     }
 
+    GGML_ASSERT(ggml_is_contiguous(dst));
+
+    // ggml_compute_forward_dup_same_cont(params, opt0, dst);
+
+    memset(dst->data, 0, ggml_nbytes(dst));
+
     const int nc = src0->ne[0];
     const int nr = ggml_nelements(src1);
 
@@ -10933,22 +13663,23 @@ static void ggml_compute_forward_get_rows_back_f32(
 
 static void ggml_compute_forward_get_rows_back(
         const struct ggml_compute_params * params,
-        const struct ggml_tensor * src0,
-        const struct ggml_tensor * src1,
         struct ggml_tensor * dst) {
+
+    const struct ggml_tensor * src0 = dst->src[0];
+
     switch (src0->type) {
         case GGML_TYPE_F16:
             {
-                ggml_compute_forward_get_rows_back_f32_f16(params, src0, src1, dst);
+                ggml_compute_forward_get_rows_back_f32_f16(params, dst);
             } break;
         case GGML_TYPE_F32:
             {
-                ggml_compute_forward_get_rows_back_f32(params, src0, src1, dst);
+                ggml_compute_forward_get_rows_back_f32(params, dst);
             } break;
         default:
             {
-                GGML_ASSERT(false);
-            } break;
+                GGML_ABORT("fatal error");
+            }
     }
 
     //static bool first = true;
@@ -10959,14 +13690,14 @@ static void ggml_compute_forward_get_rows_back(
     //    for (int k = 0; k < dst->ne[1]; ++k) {
     //        for (int j = 0; j < dst->ne[0]/16; ++j) {
     //            for (int i = 0; i < 16; ++i) {
-    //                Rprintf("%8.4f ", ((float *) dst->data)[k*dst->ne[0] + j*16 + i]);
+    //                printf("%8.4f ", ((float *) dst->data)[k*dst->ne[0] + j*16 + i]);
     //            }
-    //            Rprintf("\n");
+    //            printf("\n");
     //        }
-    //        Rprintf("\n");
+    //        printf("\n");
     //    }
-    //    Rprintf("\n");
-    //    Rf_error("whispercpp error");
+    //    printf("\n");
+    //    exit(0);
     //}
 }
 
@@ -10974,11 +13705,11 @@ static void ggml_compute_forward_get_rows_back(
 
 static void ggml_compute_forward_diag_f32(
         const struct ggml_compute_params * params,
-        const struct ggml_tensor * src0,
         struct ggml_tensor * dst) {
-    GGML_ASSERT(params->ith == 0);
 
-    if (params->type == GGML_TASK_INIT || params->type == GGML_TASK_FINALIZE) {
+    const struct ggml_tensor * src0 = dst->src[0];
+
+    if (params->ith != 0) {
         return;
     }
 
@@ -11014,17 +13745,19 @@ static void ggml_compute_forward_diag_f32(
 
 static void ggml_compute_forward_diag(
         const struct ggml_compute_params * params,
-        const struct ggml_tensor * src0,
         struct ggml_tensor * dst) {
+
+    const struct ggml_tensor * src0 = dst->src[0];
+
     switch (src0->type) {
         case GGML_TYPE_F32:
             {
-                ggml_compute_forward_diag_f32(params, src0, dst);
+                ggml_compute_forward_diag_f32(params, dst);
             } break;
         default:
             {
-                GGML_ASSERT(false);
-            } break;
+                GGML_ABORT("fatal error");
+            }
     }
 }
 
@@ -11032,10 +13765,11 @@ static void ggml_compute_forward_diag(
 
 static void ggml_compute_forward_diag_mask_f32(
         const struct ggml_compute_params * params,
-        const struct ggml_tensor * src0,
         struct ggml_tensor * dst,
         const float value) {
 
+    const struct ggml_tensor * src0 = dst->src[0];
+
     const int ith = params->ith;
     const int nth = params->nth;
 
@@ -11044,19 +13778,18 @@ static void ggml_compute_forward_diag_mask_f32(
 
     GGML_ASSERT(n_past >= 0);
 
-    if (!inplace && (params->type == GGML_TASK_INIT)) {
-        // memcpy needs to be synchronized across threads to avoid race conditions.
-        // => do it in INIT phase
-        GGML_ASSERT(ggml_nelements(dst) == ggml_nelements(src0));
-        GGML_ASSERT(ggml_is_contiguous(dst) && ggml_is_contiguous(src0));
-        memcpy(
-            ((char *)  dst->data),
-            ((char *) src0->data),
-            ggml_nbytes(dst));
-    }
-
-    if (params->type == GGML_TASK_INIT || params->type == GGML_TASK_FINALIZE) {
-        return;
+    if (!inplace) {
+        if (ith == 0) {
+            // memcpy needs to be synchronized across threads to avoid race conditions.
+            // => do it in INIT phase
+            GGML_ASSERT(ggml_nelements(dst) == ggml_nelements(src0));
+            GGML_ASSERT(ggml_is_contiguous(dst) && ggml_is_contiguous(src0));
+            memcpy(
+                ((char *)  dst->data),
+                ((char *) src0->data),
+                ggml_nbytes(dst));
+        }
+        ggml_barrier(params->threadpool);
     }
 
     // TODO: handle transposed/permuted matrices
@@ -11082,33 +13815,37 @@ static void ggml_compute_forward_diag_mask_f32(
 
 static void ggml_compute_forward_diag_mask_inf(
         const struct ggml_compute_params * params,
-        const struct ggml_tensor * src0,
         struct ggml_tensor * dst) {
+
+    const struct ggml_tensor * src0 = dst->src[0];
+
     switch (src0->type) {
         case GGML_TYPE_F32:
             {
-                ggml_compute_forward_diag_mask_f32(params, src0, dst, -INFINITY);
+                ggml_compute_forward_diag_mask_f32(params, dst, -INFINITY);
             } break;
         default:
             {
-                GGML_ASSERT(false);
-            } break;
+                GGML_ABORT("fatal error");
+            }
     }
 }
 
 static void ggml_compute_forward_diag_mask_zero(
         const struct ggml_compute_params * params,
-        const struct ggml_tensor * src0,
         struct ggml_tensor * dst) {
+
+    const struct ggml_tensor * src0 = dst->src[0];
+
     switch (src0->type) {
         case GGML_TYPE_F32:
             {
-                ggml_compute_forward_diag_mask_f32(params, src0, dst, 0);
+                ggml_compute_forward_diag_mask_f32(params, dst, 0);
             } break;
         default:
             {
-                GGML_ASSERT(false);
-            } break;
+                GGML_ABORT("fatal error");
+            }
     }
 }
 
@@ -11116,25 +13853,36 @@ static void ggml_compute_forward_diag_mask_zero(
 
 static void ggml_compute_forward_soft_max_f32(
         const struct ggml_compute_params * params,
-        const struct ggml_tensor * src0,
-        const struct ggml_tensor * src1,
               struct ggml_tensor * dst) {
+
+    const struct ggml_tensor * src0 = dst->src[0];
+    const struct ggml_tensor * src1 = dst->src[1];
+
     assert(ggml_is_contiguous(dst));
     assert(ggml_are_same_shape(src0, dst));
 
-    if (params->type == GGML_TASK_INIT || params->type == GGML_TASK_FINALIZE) {
-        return;
-    }
+    float scale    = 1.0f;
+    float max_bias = 0.0f;
 
-    float scale = 1.0f;
-    memcpy(&scale, (float *) dst->op_params + 0, sizeof(float));
+    memcpy(&scale,    (float *) dst->op_params + 0, sizeof(float));
+    memcpy(&max_bias, (float *) dst->op_params + 1, sizeof(float));
 
     // TODO: handle transposed/permuted matrices
 
     const int ith = params->ith;
     const int nth = params->nth;
 
-    const int64_t ne11 = src1 ? src1->ne[1] : 1;
+    GGML_TENSOR_UNARY_OP_LOCALS
+
+    //const int64_t ne11 = src1 ? src1->ne[1] : 1;
+
+    // TODO: is this supposed to be ceil instead of floor?
+    //       https://huggingface.co/mosaicml/mpt-7b/blob/main/attention.py#L370
+    const uint32_t n_head      = ne02;
+    const uint32_t n_head_log2 = 1u << (uint32_t) floor(log2(n_head));
+
+    const float m0 = powf(2.0f, -(max_bias       ) / n_head_log2);
+    const float m1 = powf(2.0f, -(max_bias / 2.0f) / n_head_log2);
 
     const int nc = src0->ne[0];
     const int nr = ggml_nrows(src0);
@@ -11148,17 +13896,32 @@ static void ggml_compute_forward_soft_max_f32(
 
     float * wp = (float *) params->wdata + (nc + CACHE_LINE_SIZE_F32) * ith;
 
+    const bool use_f16 = (src1 && src1->type == GGML_TYPE_F16);
+
     for (int i1 = ir0; i1 < ir1; i1++) {
+        // ALiBi
+        const uint32_t h = (i1/ne01)%ne02; // head
+        const float slope = (max_bias > 0.0f) ? h < n_head_log2 ? powf(m0, h + 1) : powf(m1, 2*(h - n_head_log2) + 1) : 1.0f;
+
         float * sp = (float *)((char *) src0->data + i1*src0->nb[1]);
         float * dp = (float *)((char *)  dst->data +  i1*dst->nb[1]);
 
         // broadcast the mask across rows
-        float * mp = src1 ? (float *)((char *) src1->data + (i1%ne11)*src1->nb[1]) : NULL;
+        ggml_fp16_t * mp_f16 = src1 ? (ggml_fp16_t *)((char *) src1->data) + (i1%ne01)*ne00 : NULL;
+        float       * mp_f32 = src1 ? (float       *)((char *) src1->data) + (i1%ne01)*ne00 : NULL;
 
         ggml_vec_cpy_f32  (nc, wp, sp);
         ggml_vec_scale_f32(nc, wp, scale);
-        if (mp) {
-            ggml_vec_acc_f32(nc, wp, mp);
+        if (mp_f32) {
+            if (use_f16) {
+                for (int i = 0; i < nc; ++i) {
+                    wp[i] += slope*GGML_FP16_TO_FP32(mp_f16[i]);
+                }
+            } else {
+                for (int i = 0; i < nc; ++i) {
+                    wp[i] += slope*mp_f32[i];
+                }
+            }
         }
 
 #ifndef NDEBUG
@@ -11171,22 +13934,7 @@ static void ggml_compute_forward_soft_max_f32(
         float max = -INFINITY;
         ggml_vec_max_f32(nc, &max, wp);
 
-        ggml_float sum = 0.0;
-
-        uint16_t scvt;
-        for (int i = 0; i < nc; i++) {
-            if (wp[i] == -INFINITY) {
-                dp[i] = 0.0f;
-            } else {
-                // const float val = (wp[i] == -INFINITY) ? 0.0 : exp(wp[i] - max);
-                ggml_fp16_t s = GGML_FP32_TO_FP16(wp[i] - max);
-                memcpy(&scvt, &s, sizeof(scvt));
-                const float val = GGML_FP16_TO_FP32(ggml_table_exp_f16[scvt]);
-                sum += (ggml_float)val;
-                dp[i] = val;
-            }
-        }
-
+        ggml_float sum = ggml_vec_soft_max_f32(nc, dp, wp, max);
         assert(sum > 0.0);
 
         sum = 1.0/sum;
@@ -11203,38 +13951,38 @@ static void ggml_compute_forward_soft_max_f32(
 
 static void ggml_compute_forward_soft_max(
         const struct ggml_compute_params * params,
-        const struct ggml_tensor * src0,
-        const struct ggml_tensor * src1,
               struct ggml_tensor * dst) {
+
+    const struct ggml_tensor * src0 = dst->src[0];
+
     switch (src0->type) {
         case GGML_TYPE_F32:
             {
-                ggml_compute_forward_soft_max_f32(params, src0, src1, dst);
+                ggml_compute_forward_soft_max_f32(params, dst);
             } break;
         default:
             {
-                GGML_ASSERT(false);
-            } break;
+                GGML_ABORT("fatal error");
+            }
     }
 }
 
+
 // ggml_compute_forward_soft_max_back
 
 static void ggml_compute_forward_soft_max_back_f32(
         const struct ggml_compute_params * params,
-        const struct ggml_tensor * src0,
-        const struct ggml_tensor * src1,
         struct ggml_tensor * dst) {
+
+    const struct ggml_tensor * src0 = dst->src[0];
+    const struct ggml_tensor * src1 = dst->src[1];
+
     GGML_ASSERT(ggml_is_contiguous(src0));
     GGML_ASSERT(ggml_is_contiguous(src1));
     GGML_ASSERT(ggml_is_contiguous(dst));
     GGML_ASSERT(ggml_are_same_shape(src0, dst));
     GGML_ASSERT(ggml_are_same_shape(src1, dst));
 
-    if (params->type == GGML_TASK_INIT || params->type == GGML_TASK_FINALIZE) {
-        return;
-    }
-
     // TODO: handle transposed/permuted matrices
 
     const int ith = params->ith;
@@ -11283,7 +14031,7 @@ static void ggml_compute_forward_soft_max_back_f32(
 
         // linear runtime, no additional memory
         float dot_y_dy = 0;
-        ggml_vec_dot_f32 (nc, &dot_y_dy, y, dy);
+        ggml_vec_dot_f32 (nc, &dot_y_dy, 0, y, 0, dy, 0, 1);
         ggml_vec_cpy_f32 (nc, dx, dy);
         ggml_vec_acc1_f32(nc, dx, -dot_y_dy);
         ggml_vec_mul_f32 (nc, dx, dx, y);
@@ -11294,180 +14042,24 @@ static void ggml_compute_forward_soft_max_back_f32(
             assert(!isinf(dx[i]));
         }
 #endif
-    }
-}
-
-static void ggml_compute_forward_soft_max_back(
-        const struct ggml_compute_params * params,
-        const struct ggml_tensor * src0,
-        const struct ggml_tensor * src1,
-        struct ggml_tensor * dst) {
-    switch (src0->type) {
-        case GGML_TYPE_F32:
-            {
-                ggml_compute_forward_soft_max_back_f32(params, src0, src1, dst);
-            } break;
-        default:
-            {
-                GGML_ASSERT(false);
-            } break;
-    }
-}
-
-// ggml_compute_forward_alibi
-
-static void ggml_compute_forward_alibi_f32(
-        const struct ggml_compute_params * params,
-        const struct ggml_tensor * src0,
-        struct ggml_tensor * dst) {
-    assert(params->ith == 0);
-
-    if (params->type == GGML_TASK_INIT || params->type == GGML_TASK_FINALIZE) {
-        return;
-    }
-
-    //const int n_past = ((int32_t *) dst->op_params)[0];
-    const int n_head = ((int32_t *) dst->op_params)[1];
-    float max_bias;
-    memcpy(&max_bias, (int32_t *) dst->op_params + 2, sizeof(float));
-
-    const int64_t ne0 = src0->ne[0]; // all_seq_len = n_past + ne1
-    const int64_t ne1 = src0->ne[1]; // seq_len_without_past
-    const int64_t ne2 = src0->ne[2]; // n_head -> this is k
-    //const int64_t ne3 = src0->ne[3]; // 1 -> bsz
-
-    const int64_t n  = ggml_nrows(src0);
-    const int64_t ne2_ne3 = n/ne1; // ne2*ne3
-
-    const size_t nb0 = src0->nb[0];
-    const size_t nb1 = src0->nb[1];
-    const size_t nb2 = src0->nb[2];
-    //const int nb3 = src0->nb[3];
-
-    GGML_ASSERT(nb0 == sizeof(float));
-    GGML_ASSERT(n_head == ne2);
-
-    // add alibi to src0 (KQ_scaled)
-    const int n_heads_log2_floor = 1 << (int) floor(log2(n_head));
-
-    const float m0 = powf(2.0f, -(max_bias) / n_heads_log2_floor);
-    const float m1 = powf(2.0f, -(max_bias / 2.0f) / n_heads_log2_floor);
-
-    for (int64_t i = 0; i < ne0; i++) {
-        for (int64_t j = 0; j < ne1; j++) {
-            for (int64_t k = 0; k < ne2_ne3; k++) {
-                float * const src = (float *)((char *) src0->data + i*nb0 + j*nb1 + k*nb2);
-                float *      pdst = (float *)((char *)  dst->data + i*nb0 + j*nb1 + k*nb2);
-
-                // TODO: k*nb2 or k*nb3
-
-                float m_k;
-
-                if (k < n_heads_log2_floor) {
-                    m_k = powf(m0, k + 1);
-                } else {
-                    m_k = powf(m1, 2 * (k - n_heads_log2_floor) + 1);
-                }
-
-                pdst[0] = i * m_k + src[0];
-            }
-        }
-    }
-}
-
-static void ggml_compute_forward_alibi_f16(
-        const struct ggml_compute_params * params,
-        const struct ggml_tensor * src0,
-        struct ggml_tensor * dst) {
-    assert(params->ith == 0);
-
-    if (params->type == GGML_TASK_INIT || params->type == GGML_TASK_FINALIZE) {
-        return;
-    }
-
-    //const int n_past = ((int32_t *) dst->op_params)[0];
-    const int n_head = ((int32_t *) dst->op_params)[1];
-    float max_bias;
-    memcpy(&max_bias, (int32_t *) dst->op_params + 2, sizeof(float));
-
-    const int ne0 = src0->ne[0]; // all_seq_len = n_past + ne1
-    const int ne1 = src0->ne[1]; // seq_len_without_past
-    const int ne2 = src0->ne[2]; // n_head -> this is k
-    //const int ne3 = src0->ne[3]; // 1 -> bsz
-
-    const int n  = ggml_nrows(src0);
-    const int ne2_ne3 = n/ne1; // ne2*ne3
-
-    const int nb0 = src0->nb[0];
-    const int nb1 = src0->nb[1];
-    const int nb2 = src0->nb[2];
-    //const int nb3 = src0->nb[3];
-
-    GGML_ASSERT(nb0 == sizeof(ggml_fp16_t));
-    //GGML_ASSERT(ne1 + n_past == ne0); (void) n_past;
-    GGML_ASSERT(n_head == ne2);
-
-    // add alibi to src0 (KQ_scaled)
-    const int n_heads_log2_floor = 1 << (int) floor(log2(n_head));
-
-    const float m0 = powf(2.0f, -(max_bias) / n_heads_log2_floor);
-    const float m1 = powf(2.0f, -(max_bias / 2.0f) / n_heads_log2_floor);
-
-    for (int i = 0; i < ne0; i++) {
-        for (int j = 0; j < ne1; j++) {
-            for (int k = 0; k < ne2_ne3; k++) {
-                ggml_fp16_t * const src  = (ggml_fp16_t *)((char *) src0->data + i*nb0 + j*nb1 + k*nb2);
-                      float *      pdst  =       (float *)((char *)  dst->data + i*nb0 + j*nb1 + k*nb2);
-
-                // TODO: k*nb2 or k*nb3
-
-                float m_k;
-
-                if (k < n_heads_log2_floor) {
-                    m_k = powf(m0, k + 1);
-                } else {
-                    m_k = powf(m1, 2 * (k - n_heads_log2_floor) + 1);
-                }
-
-                // we return F32
-                pdst[0] = i * m_k + GGML_FP16_TO_FP32(src[0]);
-            }
-        }
-    }
-}
-
-static void ggml_compute_forward_alibi(
-        const struct ggml_compute_params * params,
-        const struct ggml_tensor * src0,
-        struct ggml_tensor * dst) {
-    switch (src0->type) {
-        case GGML_TYPE_F16:
-            {
-                ggml_compute_forward_alibi_f16(params, src0, dst);
-            } break;
-        case GGML_TYPE_F32:
-            {
-                ggml_compute_forward_alibi_f32(params, src0, dst);
-            } break;
-        case GGML_TYPE_Q4_0:
-        case GGML_TYPE_Q4_1:
-        case GGML_TYPE_Q5_0:
-        case GGML_TYPE_Q5_1:
-        case GGML_TYPE_Q8_0:
-        case GGML_TYPE_Q8_1:
-        case GGML_TYPE_Q2_K:
-        case GGML_TYPE_Q3_K:
-        case GGML_TYPE_Q4_K:
-        case GGML_TYPE_Q5_K:
-        case GGML_TYPE_Q6_K:
-        case GGML_TYPE_Q8_K:
-        case GGML_TYPE_I8:
-        case GGML_TYPE_I16:
-        case GGML_TYPE_I32:
-        case GGML_TYPE_COUNT:
+    }
+}
+
+static void ggml_compute_forward_soft_max_back(
+        const struct ggml_compute_params * params,
+        struct ggml_tensor * dst) {
+
+    const struct ggml_tensor * src0 = dst->src[0];
+
+    switch (src0->type) {
+        case GGML_TYPE_F32:
             {
-                GGML_ASSERT(false);
+                ggml_compute_forward_soft_max_back_f32(params, dst);
             } break;
+        default:
+            {
+                GGML_ABORT("fatal error");
+            }
     }
 }
 
@@ -11475,11 +14067,11 @@ static void ggml_compute_forward_alibi(
 
 static void ggml_compute_forward_clamp_f32(
         const struct ggml_compute_params * params,
-        const struct ggml_tensor * src0,
         struct ggml_tensor * dst) {
-    assert(params->ith == 0);
 
-    if (params->type == GGML_TASK_INIT || params->type == GGML_TASK_FINALIZE) {
+    const struct ggml_tensor * src0 = dst->src[0];
+
+    if (params->ith != 0) {
         return;
     }
 
@@ -11515,14 +14107,17 @@ static void ggml_compute_forward_clamp_f32(
 
 static void ggml_compute_forward_clamp(
         const struct ggml_compute_params * params,
-        const struct ggml_tensor * src0,
         struct ggml_tensor * dst) {
+
+    const struct ggml_tensor * src0 = dst->src[0];
+
     switch (src0->type) {
         case GGML_TYPE_F32:
             {
-                ggml_compute_forward_clamp_f32(params, src0, dst);
+                ggml_compute_forward_clamp_f32(params, dst);
             } break;
         case GGML_TYPE_F16:
+        case GGML_TYPE_BF16:
         case GGML_TYPE_Q4_0:
         case GGML_TYPE_Q4_1:
         case GGML_TYPE_Q5_0:
@@ -11534,14 +14129,30 @@ static void ggml_compute_forward_clamp(
         case GGML_TYPE_Q4_K:
         case GGML_TYPE_Q5_K:
         case GGML_TYPE_Q6_K:
+        case GGML_TYPE_TQ1_0:
+        case GGML_TYPE_TQ2_0:
+        case GGML_TYPE_IQ2_XXS:
+        case GGML_TYPE_IQ2_XS:
+        case GGML_TYPE_IQ3_XXS:
+        case GGML_TYPE_IQ1_S:
+        case GGML_TYPE_IQ1_M:
+        case GGML_TYPE_IQ4_NL:
+        case GGML_TYPE_IQ4_XS:
+        case GGML_TYPE_IQ3_S:
+        case GGML_TYPE_IQ2_S:
         case GGML_TYPE_Q8_K:
+        case GGML_TYPE_Q4_0_4_4:
+        case GGML_TYPE_Q4_0_4_8:
+        case GGML_TYPE_Q4_0_8_8:
         case GGML_TYPE_I8:
         case GGML_TYPE_I16:
         case GGML_TYPE_I32:
+        case GGML_TYPE_I64:
+        case GGML_TYPE_F64:
         case GGML_TYPE_COUNT:
             {
-                GGML_ASSERT(false);
-            } break;
+                GGML_ABORT("fatal error");
+            }
     }
 }
 
@@ -11556,8 +14167,7 @@ static float rope_yarn_ramp(const float low, const float high, const int i0) {
 // MIT licensed. Copyright (c) 2023 Jeffrey Quesnelle and Bowen Peng.
 static void rope_yarn(
     float theta_extrap, float freq_scale, float corr_dims[2], int64_t i0, float ext_factor, float mscale,
-    float * cos_theta, float * sin_theta
-) {
+    float * cos_theta, float * sin_theta) {
     // Get n-d rotational scaling corrected for extrapolation
     float theta_interp = freq_scale * theta_extrap;
     float theta = theta_interp;
@@ -11574,39 +14184,52 @@ static void rope_yarn(
 
 // Apparently solving `n_rot = 2pi * x * base^((2 * max_pos_emb) / n_dims)` for x, we get
 // `corr_dim(n_rot) = n_dims * log(max_pos_emb / (n_rot * 2pi)) / (2 * log(base))`
-static float ggml_rope_yarn_corr_dim(int n_dims, int n_orig_ctx, float n_rot, float base) {
-    return n_dims * logf(n_orig_ctx / (n_rot * 2 * (float)M_PI)) / (2 * logf(base));
+static float ggml_rope_yarn_corr_dim(int n_dims, int n_ctx_orig, float n_rot, float base) {
+    return n_dims * logf(n_ctx_orig / (n_rot * 2 * (float)M_PI)) / (2 * logf(base));
+}
+
+static void ggml_rope_cache_init(
+     float theta_base, float freq_scale, const float * freq_factors, float corr_dims[2], int64_t ne0, float ext_factor, float mscale,
+     float * cache, float sin_sign, float theta_scale) {
+    // ref: https://github.com/jquesnelle/yarn/blob/master/scaled_rope/LlamaYaRNScaledRotaryEmbedding.py
+    float theta = theta_base;
+    for (int64_t i0 = 0; i0 < ne0; i0 += 2) {
+        const float ff = freq_factors ? freq_factors[i0/2] : 1.0f;
+        rope_yarn(
+            theta/ff, freq_scale, corr_dims, i0, ext_factor, mscale, &cache[i0 + 0], &cache[i0 + 1]
+        );
+        cache[i0 + 1] *= sin_sign;
+
+        theta *= theta_scale;
+    }
 }
 
 void ggml_rope_yarn_corr_dims(
-    int n_dims, int n_orig_ctx, float freq_base, float beta_fast, float beta_slow, float dims[2]
+    int n_dims, int n_ctx_orig, float freq_base, float beta_fast, float beta_slow, float dims[2]
 ) {
     // start and end correction dims
-    dims[0] = MAX(0,         floorf(ggml_rope_yarn_corr_dim(n_dims, n_orig_ctx, beta_fast, freq_base)));
-    dims[1] = MIN(n_dims - 1, ceilf(ggml_rope_yarn_corr_dim(n_dims, n_orig_ctx, beta_slow, freq_base)));
+    float start = floorf(ggml_rope_yarn_corr_dim(n_dims, n_ctx_orig, beta_fast, freq_base));
+    float end   =  ceilf(ggml_rope_yarn_corr_dim(n_dims, n_ctx_orig, beta_slow, freq_base));
+    dims[0] = MAX(0, start);
+    dims[1] = MIN(n_dims - 1, end);
 }
 
 static void ggml_compute_forward_rope_f32(
         const struct ggml_compute_params * params,
-        const struct ggml_tensor * src0,
-        const struct ggml_tensor * src1,
         struct ggml_tensor * dst,
         const bool forward) {
-    if (params->type == GGML_TASK_INIT || params->type == GGML_TASK_FINALIZE) {
-        return;
-    }
 
-    float freq_base, freq_scale, ext_factor, attn_factor, beta_fast, beta_slow;
+    const struct ggml_tensor * src0 = dst->src[0];
+    const struct ggml_tensor * src1 = dst->src[1];
+    const struct ggml_tensor * src2 = dst->src[2];
 
-    // these two only relevant for xPos RoPE:
-    float xpos_base;
-    bool  xpos_down;
+    float freq_base, freq_scale, ext_factor, attn_factor, beta_fast, beta_slow;
 
     //const int n_past     = ((int32_t *) dst->op_params)[0];
     const int n_dims     = ((int32_t *) dst->op_params)[1];
     const int mode       = ((int32_t *) dst->op_params)[2];
-    const int n_ctx      = ((int32_t *) dst->op_params)[3];
-    const int n_orig_ctx = ((int32_t *) dst->op_params)[4];
+    //const int n_ctx      = ((int32_t *) dst->op_params)[3];
+    const int n_ctx_orig = ((int32_t *) dst->op_params)[4];
 
     memcpy(&freq_base,   (int32_t *) dst->op_params +  5, sizeof(float));
     memcpy(&freq_scale,  (int32_t *) dst->op_params +  6, sizeof(float));
@@ -11614,8 +14237,6 @@ static void ggml_compute_forward_rope_f32(
     memcpy(&attn_factor, (int32_t *) dst->op_params +  8, sizeof(float));
     memcpy(&beta_fast,   (int32_t *) dst->op_params +  9, sizeof(float));
     memcpy(&beta_slow,   (int32_t *) dst->op_params + 10, sizeof(float));
-    memcpy(&xpos_base,   (int32_t *) dst->op_params + 11, sizeof(float));
-    memcpy(&xpos_down,   (int32_t *) dst->op_params + 12, sizeof(bool));
 
     GGML_TENSOR_UNARY_OP_LOCALS
 
@@ -11643,12 +14264,18 @@ static void ggml_compute_forward_rope_f32(
     int ir = 0;
 
     const float theta_scale = powf(freq_base, -2.0f/n_dims);
-    const float inv_ndims = -1.f/n_dims;
+
     float corr_dims[2];
-    ggml_rope_yarn_corr_dims(n_dims, n_orig_ctx, freq_base, beta_fast, beta_slow, corr_dims);
+    ggml_rope_yarn_corr_dims(n_dims, n_ctx_orig, freq_base, beta_fast, beta_slow, corr_dims);
 
-    const bool is_neox = mode & 2;
-    const bool is_glm  = mode & 4;
+    const bool is_neox = mode & GGML_ROPE_TYPE_NEOX;
+
+    const float * freq_factors = NULL;
+    if (src2 != NULL) {
+        GGML_ASSERT(src2->type == GGML_TYPE_F32);
+        GGML_ASSERT(src2->ne[0] >= n_dims / 2);
+        freq_factors = (const float *) src2->data;
+    }
 
     // backward process uses inverse rotation by cos and sin.
     // cos and sin build a rotation matrix, where the inverse is the transpose.
@@ -11660,50 +14287,18 @@ static void ggml_compute_forward_rope_f32(
     for (int64_t i3 = 0; i3 < ne3; i3++) {
         for (int64_t i2 = 0; i2 < ne2; i2++) {
             const int64_t p = pos[i2];
+
+            float * cache = (float *) params->wdata + (ne0 + CACHE_LINE_SIZE_F32)*ith;
+            ggml_rope_cache_init(p, freq_scale, freq_factors, corr_dims, ne0, ext_factor, attn_factor, cache, sin_sign, theta_scale);
+
             for (int64_t i1 = 0; i1 < ne1; i1++) {
                 if (ir++ < ir0) continue;
                 if (ir   > ir1) break;
 
-                float theta_base = (float)p;
-
-                if (is_glm) {
-                    theta_base = MIN(p, n_ctx - 2);
-                    float block_theta = MAX(p - (n_ctx - 2), 0);
-                    for (int64_t i0 = 0; i0 < ne0 / 4; i0++) {
-                        const float cos_theta = cosf(theta_base);
-                        const float sin_theta = sinf(theta_base) * sin_sign;
-                        const float cos_block_theta = cosf(block_theta);
-                        const float sin_block_theta = sinf(block_theta) * sin_sign;
-
-                        theta_base *= theta_scale;
-                        block_theta *= theta_scale;
-
-                        const float * const src = (float *)((char *) src0->data + i3*nb03 + i2*nb02 + i1*nb01 + i0*nb00);
-                              float * dst_data  = (float *)((char *)  dst->data +  i3*nb3 + i2*nb2  + i1*nb1  + i0*nb0);
-
-                        const float x0 = src[0];
-                        const float x1 = src[n_dims/2];
-                        const float x2 = src[n_dims];
-                        const float x3 = src[n_dims/2*3];
-
-                        dst_data[0]          = x0*cos_theta - x1*sin_theta;
-                        dst_data[n_dims/2]   = x0*sin_theta + x1*cos_theta;
-                        dst_data[n_dims]     = x2*cos_block_theta - x3*sin_block_theta;
-                        dst_data[n_dims/2*3] = x2*sin_block_theta + x3*cos_block_theta;
-                    }
-                } else if (!is_neox) {
-                    for (int64_t i0 = 0; i0 < ne0; i0 += 2) {
-                        float cos_theta, sin_theta;
-                        rope_yarn(
-                            theta_base, freq_scale, corr_dims, i0, ext_factor, attn_factor, &cos_theta, &sin_theta
-                        );
-                        sin_theta *= sin_sign;
-
-                        // zeta scaling for xPos only:
-                        float zeta = xpos_base != 0.0f ? powf((i0 + 0.4f * ne0) / (1.4f * ne0), p / xpos_base) : 1.0f;
-                        if (xpos_down) zeta = 1.0f / zeta;
-
-                        theta_base *= theta_scale;
+                if (!is_neox) {
+                    for (int64_t i0 = 0; i0 < n_dims; i0 += 2) {
+                        const float cos_theta = cache[i0 + 0];
+                        const float sin_theta = cache[i0 + 1];
 
                         const float * const src = (float *)((char *) src0->data + i3*nb03 + i2*nb02 + i1*nb01 + i0*nb00);
                               float * dst_data  = (float *)((char *)  dst->data + i3*nb3  + i2*nb2  + i1*nb1  + i0*nb0);
@@ -11711,73 +14306,56 @@ static void ggml_compute_forward_rope_f32(
                         const float x0 = src[0];
                         const float x1 = src[1];
 
-                        dst_data[0] = x0*cos_theta*zeta - x1*sin_theta*zeta;
-                        dst_data[1] = x0*sin_theta*zeta + x1*cos_theta*zeta;
+                        dst_data[0] = x0*cos_theta - x1*sin_theta;
+                        dst_data[1] = x0*sin_theta + x1*cos_theta;
                     }
                 } else {
-                    // TODO: this might be wrong for ne0 != n_dims - need double check
-                    //       it seems we have to rope just the first n_dims elements and do nothing with the rest
-                    // ref:  https://github.com/ml-explore/mlx/blob/dc2edc762c797e3b8de50b1dad4dc0a131691033/benchmarks/python/llama_jax_bench.py#L11-L26
-                    theta_base *= freq_scale;
-                    for (int64_t ic = 0; ic < ne0; ic += 2) {
-                        if (ic < n_dims) {
-                            const int64_t ib = 0;
-
-                            // simplified from `(ib * n_dims + ic) * inv_ndims`
-                            float cur_rot = inv_ndims * ic - ib;
+                    for (int64_t i0 = 0; i0 < n_dims; i0 += 2) {
+                        const int64_t ic = i0/2;
 
-                            float cos_theta, sin_theta;
-                            rope_yarn(
-                                theta_base, freq_scale, corr_dims, cur_rot, ext_factor, attn_factor,
-                                &cos_theta, &sin_theta
-                            );
-                            sin_theta *= sin_sign;
+                        const float cos_theta = cache[i0 + 0];
+                        const float sin_theta = cache[i0 + 1];
 
-                            theta_base *= theta_scale;
+                        const float * const src = (float *)((char *) src0->data + i3*nb03 + i2*nb02 + i1*nb01 + ic*nb00);
+                        float * dst_data  = (float *)((char *)  dst->data + i3*nb3  + i2*nb2  + i1*nb1  + ic*nb0);
 
-                            const int64_t i0 = ib*n_dims + ic/2;
-
-                            const float * const src = (float *)((char *) src0->data + i3*nb03 + i2*nb02 + i1*nb01 + i0*nb00);
-                                  float * dst_data  = (float *)((char *)  dst->data + i3*nb3  + i2*nb2  + i1*nb1  + i0*nb0);
-
-                            const float x0 = src[0];
-                            const float x1 = src[n_dims/2];
+                        const float x0 = src[0];
+                        const float x1 = src[n_dims/2];
 
-                            dst_data[0]        = x0*cos_theta - x1*sin_theta;
-                            dst_data[n_dims/2] = x0*sin_theta + x1*cos_theta;
-                        } else {
-                            const int64_t i0 = ic;
+                        dst_data[0]        = x0*cos_theta - x1*sin_theta;
+                        dst_data[n_dims/2] = x0*sin_theta + x1*cos_theta;
+                    }
+                }
 
-                            const float * const src = (float *)((char *) src0->data + i3*nb03 + i2*nb02 + i1*nb01 + i0*nb00);
-                                  float * dst_data  = (float *)((char *)  dst->data + i3*nb3  + i2*nb2  + i1*nb1  + i0*nb0);
+                for (int64_t i0 = n_dims; i0 < ne0; i0 += 2) {
+                    const float * const src = (float *)((char *) src0->data + i3*nb03 + i2*nb02 + i1*nb01 + i0*nb00);
+                    float * dst_data  = (float *)((char *)  dst->data + i3*nb3  + i2*nb2  + i1*nb1  + i0*nb0);
 
-                            dst_data[0] = src[0];
-                            dst_data[1] = src[1];
-                        }
-                    }
+                    dst_data[0] = src[0];
+                    dst_data[1] = src[1];
                 }
             }
         }
     }
 }
 
+// TODO: deduplicate f16/f32 code
 static void ggml_compute_forward_rope_f16(
         const struct ggml_compute_params * params,
-        const struct ggml_tensor * src0,
-        const struct ggml_tensor * src1,
         struct ggml_tensor * dst,
         const bool forward) {
-    if (params->type == GGML_TASK_INIT || params->type == GGML_TASK_FINALIZE) {
-        return;
-    }
+
+    const struct ggml_tensor * src0 = dst->src[0];
+    const struct ggml_tensor * src1 = dst->src[1];
+    const struct ggml_tensor * src2 = dst->src[2];
 
     float freq_base, freq_scale, ext_factor, attn_factor, beta_fast, beta_slow;
 
     //const int n_past     = ((int32_t *) dst->op_params)[0];
     const int n_dims     = ((int32_t *) dst->op_params)[1];
     const int mode       = ((int32_t *) dst->op_params)[2];
-    const int n_ctx      = ((int32_t *) dst->op_params)[3];
-    const int n_orig_ctx = ((int32_t *) dst->op_params)[4];
+    //const int n_ctx      = ((int32_t *) dst->op_params)[3];
+    const int n_ctx_orig = ((int32_t *) dst->op_params)[4];
     memcpy(&freq_base,   (int32_t *) dst->op_params +  5, sizeof(float));
     memcpy(&freq_scale,  (int32_t *) dst->op_params +  6, sizeof(float));
     memcpy(&ext_factor,  (int32_t *) dst->op_params +  7, sizeof(float));
@@ -11811,12 +14389,18 @@ static void ggml_compute_forward_rope_f16(
     int ir = 0;
 
     const float theta_scale = powf(freq_base, -2.0f/n_dims);
-    const float inv_ndims = -1.f/n_dims;
+
     float corr_dims[2];
-    ggml_rope_yarn_corr_dims(n_dims, n_orig_ctx, freq_base, beta_fast, beta_slow, corr_dims);
+    ggml_rope_yarn_corr_dims(n_dims, n_ctx_orig, freq_base, beta_fast, beta_slow, corr_dims);
+
+    const bool is_neox = mode & GGML_ROPE_TYPE_NEOX;
 
-    const bool is_neox = mode & 2;
-    const bool is_glm  = mode & 4;
+    const float * freq_factors = NULL;
+    if (src2 != NULL) {
+        GGML_ASSERT(src2->type == GGML_TYPE_F32);
+        GGML_ASSERT(src2->ne[0] >= n_dims / 2);
+        freq_factors = (const float *) src2->data;
+    }
 
     // backward process uses inverse rotation by cos and sin.
     // cos and sin build a rotation matrix, where the inverse is the transpose.
@@ -11828,46 +14412,18 @@ static void ggml_compute_forward_rope_f16(
     for (int64_t i3 = 0; i3 < ne3; i3++) {
         for (int64_t i2 = 0; i2 < ne2; i2++) {
             const int64_t p = pos[i2];
+
+            float * cache = (float *) params->wdata + (ne0 + CACHE_LINE_SIZE_F32)*ith;
+            ggml_rope_cache_init(p, freq_scale, freq_factors, corr_dims, ne0, ext_factor, attn_factor, cache, sin_sign, theta_scale);
+
             for (int64_t i1 = 0; i1 < ne1; i1++) {
                 if (ir++ < ir0) continue;
                 if (ir   > ir1) break;
 
-                float theta_base = (float)p;
-
-                if (is_glm) {
-                    theta_base = MIN(p, n_ctx - 2);
-                    float block_theta = MAX(p - (n_ctx - 2), 0);
-                    for (int64_t i0 = 0; i0 < ne0 / 4; i0++) {
-                        const float cos_theta = cosf(theta_base);
-                        const float sin_theta = sinf(theta_base) * sin_sign;
-                        const float cos_block_theta = cosf(block_theta);
-                        const float sin_block_theta = sinf(block_theta) * sin_sign;
-
-                        theta_base *= theta_scale;
-                        block_theta *= theta_scale;
-
-                        const ggml_fp16_t * const src = (ggml_fp16_t *)((char *) src0->data + i3*nb03 + i2*nb02 + i1*nb01 + i0*nb00);
-                              ggml_fp16_t * dst_data  = (ggml_fp16_t *)((char *)  dst->data +  i3*nb3 + i2*nb2  + i1*nb1  + i0*nb0);
-
-                        const float x0 = GGML_FP16_TO_FP32(src[0]);
-                        const float x1 = GGML_FP16_TO_FP32(src[n_dims/2]);
-                        const float x2 = GGML_FP16_TO_FP32(src[n_dims]);
-                        const float x3 = GGML_FP16_TO_FP32(src[n_dims/2*3]);
-
-                        dst_data[0]          = GGML_FP32_TO_FP16(x0*cos_theta - x1*sin_theta);
-                        dst_data[n_dims/2]   = GGML_FP32_TO_FP16(x0*sin_theta + x1*cos_theta);
-                        dst_data[n_dims]     = GGML_FP32_TO_FP16(x2*cos_block_theta - x3*sin_block_theta);
-                        dst_data[n_dims/2*3] = GGML_FP32_TO_FP16(x2*sin_block_theta + x3*cos_block_theta);
-                    }
-                } else if (!is_neox) {
-                    for (int64_t i0 = 0; i0 < ne0; i0 += 2) {
-                        float cos_theta, sin_theta;
-                        rope_yarn(
-                            theta_base, freq_scale, corr_dims, i0, ext_factor, attn_factor, &cos_theta, &sin_theta
-                        );
-                        sin_theta *= sin_sign;
-
-                        theta_base *= theta_scale;
+                if (!is_neox) {
+                    for (int64_t i0 = 0; i0 < n_dims; i0 += 2) {
+                        const float cos_theta = cache[i0 + 0];
+                        const float sin_theta = cache[i0 + 1];
 
                         const ggml_fp16_t * const src = (ggml_fp16_t *)((char *) src0->data + i3*nb03 + i2*nb02 + i1*nb01 + i0*nb00);
                               ggml_fp16_t * dst_data  = (ggml_fp16_t *)((char *)  dst->data + i3*nb3  + i2*nb2  + i1*nb1  + i0*nb0);
@@ -11879,46 +14435,29 @@ static void ggml_compute_forward_rope_f16(
                         dst_data[1] = GGML_FP32_TO_FP16(x0*sin_theta + x1*cos_theta);
                     }
                 } else {
-                    // TODO: this might be wrong for ne0 != n_dims - need double check
-                    //       it seems we have to rope just the first n_dims elements and do nothing with the rest
-                    // ref:  https://github.com/ml-explore/mlx/blob/dc2edc762c797e3b8de50b1dad4dc0a131691033/benchmarks/python/llama_jax_bench.py#L11-L26
-                    theta_base *= freq_scale;
-                    for (int64_t ic = 0; ic < ne0; ic += 2) {
-                        if (ic < n_dims) {
-                            const int64_t ib = 0;
-
-                            // simplified from `(ib * n_dims + ic) * inv_ndims`
-                            float cur_rot = inv_ndims * ic - ib;
+                    for (int64_t i0 = 0; i0 < n_dims; i0 += 2) {
+                        const int64_t ic = i0/2;
 
-                            float cos_theta, sin_theta;
-                            rope_yarn(
-                                theta_base, freq_scale, corr_dims, cur_rot, ext_factor, attn_factor,
-                                &cos_theta, &sin_theta
-                            );
-                            sin_theta *= sin_sign;
+                        const float cos_theta = cache[i0 + 0];
+                        const float sin_theta = cache[i0 + 1];
 
-                            theta_base *= theta_scale;
+                        const ggml_fp16_t * const src = (ggml_fp16_t *)((char *) src0->data + i3*nb03 + i2*nb02 + i1*nb01 + ic*nb00);
+                        ggml_fp16_t * dst_data  = (ggml_fp16_t *)((char *)  dst->data + i3*nb3  + i2*nb2  + i1*nb1  + ic*nb0);
 
-                            const int64_t i0 = ib*n_dims + ic/2;
-
-                            const ggml_fp16_t * const src = (ggml_fp16_t *)((char *) src0->data + i3*nb03 + i2*nb02 + i1*nb01 + i0*nb00);
-                                  ggml_fp16_t * dst_data  = (ggml_fp16_t *)((char *)  dst->data + i3*nb3  + i2*nb2  + i1*nb1  + i0*nb0);
-
-                            const float x0 = GGML_FP16_TO_FP32(src[0]);
-                            const float x1 = GGML_FP16_TO_FP32(src[n_dims/2]);
+                        const float x0 = GGML_FP16_TO_FP32(src[0]);
+                        const float x1 = GGML_FP16_TO_FP32(src[n_dims/2]);
 
-                            dst_data[0]        = GGML_FP32_TO_FP16(x0*cos_theta - x1*sin_theta);
-                            dst_data[n_dims/2] = GGML_FP32_TO_FP16(x0*sin_theta + x1*cos_theta);
-                        } else {
-                            const int64_t i0 = ic;
+                        dst_data[0]        = GGML_FP32_TO_FP16(x0*cos_theta - x1*sin_theta);
+                        dst_data[n_dims/2] = GGML_FP32_TO_FP16(x0*sin_theta + x1*cos_theta);
+                    }
+                }
 
-                            const ggml_fp16_t * const src = (ggml_fp16_t *)((char *) src0->data + i3*nb03 + i2*nb02 + i1*nb01 + i0*nb00);
-                                  ggml_fp16_t * dst_data  = (ggml_fp16_t *)((char *)  dst->data + i3*nb3  + i2*nb2  + i1*nb1  + i0*nb0);
+                for (int64_t i0 = n_dims; i0 < ne0; i0 += 2) {
+                    const ggml_fp16_t * const src = (ggml_fp16_t *)((char *) src0->data + i3*nb03 + i2*nb02 + i1*nb01 + i0*nb00);
+                    ggml_fp16_t * dst_data  = (ggml_fp16_t *)((char *)  dst->data + i3*nb3  + i2*nb2  + i1*nb1  + i0*nb0);
 
-                            dst_data[0] = src[0];
-                            dst_data[1] = src[1];
-                        }
-                    }
+                    dst_data[0] = src[0];
+                    dst_data[1] = src[1];
                 }
             }
         }
@@ -11927,22 +14466,23 @@ static void ggml_compute_forward_rope_f16(
 
 static void ggml_compute_forward_rope(
         const struct ggml_compute_params * params,
-        const struct ggml_tensor * src0,
-        const struct ggml_tensor * src1,
         struct ggml_tensor * dst) {
+
+    const struct ggml_tensor * src0 = dst->src[0];
+
     switch (src0->type) {
         case GGML_TYPE_F16:
             {
-                ggml_compute_forward_rope_f16(params, src0, src1, dst, true);
+                ggml_compute_forward_rope_f16(params, dst, true);
             } break;
         case GGML_TYPE_F32:
             {
-                ggml_compute_forward_rope_f32(params, src0, src1, dst, true);
+                ggml_compute_forward_rope_f32(params, dst, true);
             } break;
         default:
             {
-                GGML_ASSERT(false);
-            } break;
+                GGML_ABORT("fatal error");
+            }
     }
 }
 
@@ -11950,22 +14490,23 @@ static void ggml_compute_forward_rope(
 
 static void ggml_compute_forward_rope_back(
         const struct ggml_compute_params * params,
-        const struct ggml_tensor * src0,
-        const struct ggml_tensor * src1,
         struct ggml_tensor * dst) {
+
+    const struct ggml_tensor * src0 = dst->src[0];
+
     switch (src0->type) {
         case GGML_TYPE_F16:
             {
-                ggml_compute_forward_rope_f16(params, src0, src1, dst, false);
+                ggml_compute_forward_rope_f16(params, dst, false);
             } break;
         case GGML_TYPE_F32:
             {
-                ggml_compute_forward_rope_f32(params, src0, src1, dst, false);
+                ggml_compute_forward_rope_f32(params, dst, false);
             } break;
         default:
             {
-                GGML_ASSERT(false);
-            } break;
+                GGML_ABORT("fatal error");
+            }
     }
 }
 
@@ -11973,16 +14514,15 @@ static void ggml_compute_forward_rope_back(
 
 static void ggml_compute_forward_conv_transpose_1d_f16_f32(
         const struct ggml_compute_params * params,
-        const struct ggml_tensor * src0,
-        const struct ggml_tensor * src1,
               struct ggml_tensor * dst) {
+
+    const struct ggml_tensor * src0 = dst->src[0];
+    const struct ggml_tensor * src1 = dst->src[1];
+
     GGML_ASSERT(src0->type == GGML_TYPE_F16);
     GGML_ASSERT(src1->type == GGML_TYPE_F32);
     GGML_ASSERT( dst->type == GGML_TYPE_F32);
 
-    int64_t t0 = ggml_perf_time_us();
-    UNUSED(t0);
-
     GGML_TENSOR_BINARY_OP_LOCALS
 
     const int ith = params->ith;
@@ -11993,7 +14533,7 @@ static void ggml_compute_forward_conv_transpose_1d_f16_f32(
     GGML_ASSERT(nb00 == sizeof(ggml_fp16_t));
     GGML_ASSERT(nb10 == sizeof(float));
 
-    if (params->type == GGML_TASK_INIT) {
+    if (ith == 0) {
         memset(params->wdata, 0, params->wsize);
 
         // permute kernel data (src0) from (K x Cout x Cin) to (Cin x K x Cout)
@@ -12026,13 +14566,8 @@ static void ggml_compute_forward_conv_transpose_1d_f16_f32(
 
         // need to zero dst since we are accumulating into it
         memset(dst->data, 0, ggml_nbytes(dst));
-
-        return;
-    }
-
-    if (params->type == GGML_TASK_FINALIZE) {
-        return;
     }
+    ggml_barrier(params->threadpool);
 
     const int32_t s0 = ((const int32_t*)(dst->op_params))[0];
 
@@ -12056,9 +14591,9 @@ static void ggml_compute_forward_conv_transpose_1d_f16_f32(
             const int i1n = i10*ne11;
             for (int i00 = 0; i00 < ne00; i00++) {
                 float v = 0;
-                ggml_vec_dot_f16(ne02, &v,
-                        (ggml_fp16_t *)    wdata_src + i1n,
-                        (ggml_fp16_t *) wdata_kernel + i00*ne02);
+                ggml_vec_dot_f16(ne02, &v, 0,
+                        (ggml_fp16_t *)    wdata_src + i1n, 0,
+                        (ggml_fp16_t *) wdata_kernel + i00*ne02, 0, 1);
                 dst_data[i10*s0 + i00] += v;
             }
         }
@@ -12067,16 +14602,15 @@ static void ggml_compute_forward_conv_transpose_1d_f16_f32(
 
 static void ggml_compute_forward_conv_transpose_1d_f32(
         const struct ggml_compute_params * params,
-        const struct ggml_tensor * src0,
-        const struct ggml_tensor * src1,
               struct ggml_tensor * dst) {
+
+    const struct ggml_tensor * src0 = dst->src[0];
+    const struct ggml_tensor * src1 = dst->src[1];
+
     GGML_ASSERT(src0->type == GGML_TYPE_F32);
     GGML_ASSERT(src1->type == GGML_TYPE_F32);
     GGML_ASSERT( dst->type == GGML_TYPE_F32);
 
-    int64_t t0 = ggml_perf_time_us();
-    UNUSED(t0);
-
     GGML_TENSOR_BINARY_OP_LOCALS
 
     const int ith = params->ith;
@@ -12087,7 +14621,7 @@ static void ggml_compute_forward_conv_transpose_1d_f32(
     GGML_ASSERT(nb00 == sizeof(float));
     GGML_ASSERT(nb10 == sizeof(float));
 
-    if (params->type == GGML_TASK_INIT) {
+    if (ith == 0) {
         memset(params->wdata, 0, params->wsize);
 
         // prepare kernel data (src0) from (K x Cout x Cin) to (Cin x K x Cout)
@@ -12120,13 +14654,8 @@ static void ggml_compute_forward_conv_transpose_1d_f32(
 
         // need to zero dst since we are accumulating into it
         memset(dst->data, 0, ggml_nbytes(dst));
-
-        return;
-    }
-
-    if (params->type == GGML_TASK_FINALIZE) {
-        return;
     }
+    ggml_barrier(params->threadpool);
 
     const int32_t s0 = ((const int32_t*)(dst->op_params))[0];
 
@@ -12150,9 +14679,9 @@ static void ggml_compute_forward_conv_transpose_1d_f32(
             const int i1n = i10*ne11;
             for (int i00 = 0; i00 < ne00; i00++) {
                 float v = 0;
-                ggml_vec_dot_f32(ne02, &v,
-                        wdata_src + i1n,
-                        wdata_kernel + i00*ne02);
+                ggml_vec_dot_f32(ne02, &v, 0,
+                        wdata_src + i1n, 0,
+                        wdata_kernel + i00*ne02, 0, 1);
                 dst_data[i10*s0 + i00] += v;
             }
         }
@@ -12161,40 +14690,117 @@ static void ggml_compute_forward_conv_transpose_1d_f32(
 
 static void ggml_compute_forward_conv_transpose_1d(
         const struct ggml_compute_params * params,
-        const struct ggml_tensor * src0,
-        const struct ggml_tensor * src1,
               struct ggml_tensor * dst) {
+
+    const struct ggml_tensor * src0 = dst->src[0];
+
     switch (src0->type) {
         case GGML_TYPE_F16:
             {
-                ggml_compute_forward_conv_transpose_1d_f16_f32(params, src0, src1, dst);
+                ggml_compute_forward_conv_transpose_1d_f16_f32(params, dst);
             } break;
         case GGML_TYPE_F32:
             {
-                ggml_compute_forward_conv_transpose_1d_f32(params, src0, src1, dst);
+                ggml_compute_forward_conv_transpose_1d_f32(params, dst);
             } break;
         default:
             {
-                GGML_ASSERT(false);
-            } break;
+                GGML_ABORT("fatal error");
+            }
+    }
+}
+
+// ggml_compute_forward_im2col_f32
+// src0: kernel [OC, IC, KH, KW]
+// src1: image [N, IC, IH, IW]
+// dst:  result [N, OH, OW, IC*KH*KW]
+static void ggml_compute_forward_im2col_f32(
+        const struct ggml_compute_params * params,
+              struct ggml_tensor * dst) {
+
+    const struct ggml_tensor * src0 = dst->src[0];
+    const struct ggml_tensor * src1 = dst->src[1];
+
+    GGML_ASSERT(src1->type == GGML_TYPE_F32);
+    GGML_ASSERT( dst->type == GGML_TYPE_F32);
+
+    GGML_TENSOR_BINARY_OP_LOCALS;
+
+    const int32_t s0 = ((const int32_t *)(dst->op_params))[0];
+    const int32_t s1 = ((const int32_t *)(dst->op_params))[1];
+    const int32_t p0 = ((const int32_t *)(dst->op_params))[2];
+    const int32_t p1 = ((const int32_t *)(dst->op_params))[3];
+    const int32_t d0 = ((const int32_t *)(dst->op_params))[4];
+    const int32_t d1 = ((const int32_t *)(dst->op_params))[5];
+    const bool is_2D = ((const int32_t *)(dst->op_params))[6] == 1;
+
+    const int ith = params->ith;
+    const int nth = params->nth;
+
+    const int64_t N  = is_2D ? ne13 : ne12;
+    const int64_t IC = is_2D ? ne12 : ne11;
+    const int64_t IH = is_2D ? ne11 : 1;
+    const int64_t IW = ne10;
+
+    const int64_t KH = is_2D ? ne01 : 1;
+    const int64_t KW = ne00;
+
+    const int64_t OH = is_2D ? ne2 : 1;
+    const int64_t OW = ne1;
+
+    int ofs0 = is_2D ? nb13 : nb12;
+    int ofs1 = is_2D ? nb12 : nb11;
+
+    GGML_ASSERT(nb10 == sizeof(float));
+
+    // im2col: [N, IC, IH, IW] => [N, OH, OW, IC*KH*KW]
+    {
+        float * const wdata = (float *) dst->data;
+
+        for (int64_t in = 0; in < N; in++) {
+            for (int64_t ioh = 0; ioh < OH; ioh++) { // 1
+                for (int64_t iow = 0; iow < OW; iow++) {
+                    for (int64_t iic = ith; iic < IC; iic += nth) {
+
+                        // micro kernel
+                        float * dst_data = wdata + (in*OH*OW + ioh*OW + iow)*(IC*KH*KW); // [IC, KH, KW]
+                        const float * const src_data = (float *)((char *) src1->data + in*ofs0 + iic*ofs1); // [IH, IW]
+
+                        for (int64_t ikh = 0; ikh < KH; ikh++) {  // 1
+                            for (int64_t ikw = 0; ikw < KW; ikw++) {
+                                const int64_t iiw = iow*s0 + ikw*d0 - p0;
+                                const int64_t iih = ioh*s1 + ikh*d1 - p1;
+
+                                if (iih < 0 || iih >= IH || iiw < 0 || iiw >= IW) {
+                                    dst_data[iic*(KH*KW) + ikh*KW + ikw] = 0;
+                                } else {
+                                    dst_data[iic*(KH*KW) + ikh*KW + ikw] = (src_data[iih*IW + iiw]);
+                                }
+                            }
+                        }
+                    }
+                }
+            }
+        }
     }
 }
 
+
+// ggml_compute_forward_im2col_f16
 // src0: kernel [OC, IC, KH, KW]
 // src1: image [N, IC, IH, IW]
 // dst:  result [N, OH, OW, IC*KH*KW]
 static void ggml_compute_forward_im2col_f16(
         const struct ggml_compute_params * params,
-        const struct ggml_tensor * src0,
-        const struct ggml_tensor * src1,
               struct ggml_tensor * dst) {
+
+    const struct ggml_tensor * src0 = dst->src[0];
+    const struct ggml_tensor * src1 = dst->src[1];
+
     GGML_ASSERT(src0->type == GGML_TYPE_F16);
     GGML_ASSERT(src1->type == GGML_TYPE_F32);
     GGML_ASSERT( dst->type == GGML_TYPE_F16);
 
-    int64_t t0 = ggml_perf_time_us();
-    UNUSED(t0);
-
     GGML_TENSOR_BINARY_OP_LOCALS;
 
     const int32_t s0 = ((const int32_t *)(dst->op_params))[0];
@@ -12225,14 +14831,6 @@ static void ggml_compute_forward_im2col_f16(
     GGML_ASSERT(nb00 == sizeof(ggml_fp16_t));
     GGML_ASSERT(nb10 == sizeof(float));
 
-    if (params->type == GGML_TASK_INIT) {
-        return;
-    }
-
-    if (params->type == GGML_TASK_FINALIZE) {
-        return;
-    }
-
     // im2col: [N, IC, IH, IW] => [N, OH, OW, IC*KH*KW]
     {
         ggml_fp16_t * const wdata = (ggml_fp16_t *) dst->data;
@@ -12265,41 +14863,132 @@ static void ggml_compute_forward_im2col_f16(
     }
 }
 
-static void ggml_compute_forward_im2col(
-        const struct ggml_compute_params * params,
-        const struct ggml_tensor * src0,
-        const struct ggml_tensor * src1,
-              struct ggml_tensor * dst) {
-    switch (src0->type) {
-        case GGML_TYPE_F16:
-            {
-                ggml_compute_forward_im2col_f16(params, src0, src1, dst);
-            } break;
-        case GGML_TYPE_F32:
-            {
-                GGML_ASSERT(false);
-            } break;
-        default:
-            {
-                GGML_ASSERT(false);
-            } break;
-    }
-}
-
+static void ggml_compute_forward_im2col(
+        const struct ggml_compute_params * params,
+              struct ggml_tensor * dst) {
+    switch (dst->type) {
+        case GGML_TYPE_F16:
+            {
+                ggml_compute_forward_im2col_f16(params, dst);
+            } break;
+        case GGML_TYPE_F32:
+            {
+                ggml_compute_forward_im2col_f32(params, dst);
+            } break;
+        default:
+            {
+                GGML_ABORT("fatal error");
+            }
+    }
+}
+
+// ggml_compute_forward_im2col_back_f32
+
+static void ggml_compute_forward_im2col_back_f32(
+        const struct ggml_compute_params * params,
+              struct ggml_tensor * dst) {
+
+    const struct ggml_tensor * src0 = dst->src[0];
+    const struct ggml_tensor * src1 = dst->src[1];
+
+    GGML_ASSERT(src1->type == GGML_TYPE_F32);
+    GGML_ASSERT( dst->type == GGML_TYPE_F32);
+
+    GGML_TENSOR_BINARY_OP_LOCALS;
+
+    const int32_t s0 = ((const int32_t *)(dst->op_params))[0];
+    const int32_t s1 = ((const int32_t *)(dst->op_params))[1];
+    const int32_t p0 = ((const int32_t *)(dst->op_params))[2];
+    const int32_t p1 = ((const int32_t *)(dst->op_params))[3];
+    const int32_t d0 = ((const int32_t *)(dst->op_params))[4];
+    const int32_t d1 = ((const int32_t *)(dst->op_params))[5];
+    const bool is_2D = ((const int32_t *)(dst->op_params))[6] == 1;
+
+    const int ith = params->ith;
+    const int nth = params->nth;
+
+    const int64_t N  = is_2D ? ne3 : ne2;
+    const int64_t IC = is_2D ? ne2 : ne1;
+    const int64_t IH = is_2D ? ne1 : 1;
+    const int64_t IW = ne0;
+
+    const int64_t KH = is_2D ? ne01 : 1;
+    const int64_t KW = ne00;
+
+    const int64_t OH = is_2D ? ne12 : 1;
+    const int64_t OW = ne11;
+
+    int ofs0 = is_2D ? nb3 : nb2;
+    int ofs1 = is_2D ? nb2 : nb1;
+
+    GGML_ASSERT(nb0  == sizeof(float));
+
+    // im2col: [N, IC, IH, IW] => [N, OH, OW, IC*KH*KW]
+    {
+        float * const wdata = (float *) dst->data;
+
+        for (int64_t in = 0; in < N; in++) {
+            for (int64_t iic = ith; iic < IC; iic += nth) {
+                for (int64_t iih = 0; iih < IH; iih++) {
+                    for (int64_t iiw = 0; iiw < IW; iiw++) {
+
+                        // micro kernel
+                        float grad = 0.0f;
+                        for (int64_t ikh = 0; ikh < KH; ikh++) {
+                            for (int64_t ikw = 0; ikw < KW; ikw++) {
+                                // For s0 > 1 some values were skipped over in the forward pass.
+                                // These values have tmpw % s0 != 0 and need to be skipped in the backwards pass as well.
+                                const int64_t tmpw = (iiw + p0 - ikw*d0);
+                                if (tmpw % s0 != 0) {
+                                    continue;
+                                }
+                                const int64_t iow = tmpw / s0;
+
+                                // Equivalent logic as above except for s1.
+                                int64_t ioh;
+                                if (is_2D) {
+                                    const int64_t tmph = iih + p1 - ikh*d1;
+
+                                    if (tmph % s1 != 0) {
+                                        continue;
+                                    }
+
+                                    ioh = tmph / s1;
+                                } else {
+                                    ioh = 0;
+                                }
+
+                                if (iow < 0 || iow >= OW || ioh < 0 || ioh >= OH) {
+                                    continue;
+                                }
+
+                                const float * const src_data = (const float *) src1->data
+                                    + (in*OH*OW + ioh*OW + iow)*(IC*KH*KW); // [IC, KH, KW]
+                                grad += src_data[iic*(KH*KW) + ikh*KW + ikw];
+                            }
+                        }
+                        float * dst_data = (float *)((char *) wdata + (in*ofs0 + iic*ofs1)); // [IH, IW]
+                        dst_data[iih*IW + iiw] = grad;
+                    }
+                }
+            }
+        }
+    }
+}
+
 // ggml_compute_forward_conv_transpose_2d
 
 static void ggml_compute_forward_conv_transpose_2d(
         const struct ggml_compute_params * params,
-        const struct ggml_tensor * src0,
-        const struct ggml_tensor * src1,
               struct ggml_tensor * dst) {
+
+    const struct ggml_tensor * src0 = dst->src[0];
+    const struct ggml_tensor * src1 = dst->src[1];
+
     GGML_ASSERT(src0->type == GGML_TYPE_F16);
     GGML_ASSERT(src1->type == GGML_TYPE_F32);
     GGML_ASSERT( dst->type == GGML_TYPE_F32);
 
-    int64_t t0 = ggml_perf_time_us();
-    UNUSED(t0);
-
     GGML_TENSOR_BINARY_OP_LOCALS
 
     const int ith = params->ith;
@@ -12310,7 +14999,7 @@ static void ggml_compute_forward_conv_transpose_2d(
     GGML_ASSERT(nb00 == sizeof(ggml_fp16_t));
     GGML_ASSERT(nb10 == sizeof(float));
 
-    if (params->type == GGML_TASK_INIT) {
+    if (ith == 0) {
         memset(params->wdata, 0, params->wsize);
 
         // permute kernel data (src0) from (Kw x Kh x Cout x Cin) to (Cin x Kw x Kh x Cout)
@@ -12345,13 +15034,8 @@ static void ggml_compute_forward_conv_transpose_2d(
         }
 
         memset(dst->data, 0, ggml_nbytes(dst));
-
-        return;
-    }
-
-    if (params->type == GGML_TASK_FINALIZE) {
-        return;
     }
+    ggml_barrier(params->threadpool);
 
     const int32_t stride = ggml_get_op_params_i32(dst, 0);
 
@@ -12377,9 +15061,9 @@ static void ggml_compute_forward_conv_transpose_2d(
                 for (int i01 = 0; i01 < ne01; i01++) {
                     for (int i00 = 0; i00 < ne00; i00++) {
                         float v = 0;
-                        ggml_vec_dot_f16(ne03, &v,
-                                wdata_src + i1n,
-                                wdata_kernel + i01*ne00*ne03 + i00*ne03);
+                        ggml_vec_dot_f16(ne03, &v, 0,
+                                wdata_src + i1n, 0,
+                                wdata_kernel + i01*ne00*ne03 + i00*ne03, 0, 1);
                         dst_data[(i11*stride + i01)*ne0 + i10*stride + i00] += v;
                     }
                 }
@@ -12393,13 +15077,14 @@ static void ggml_compute_forward_conv_transpose_2d(
 static void ggml_compute_forward_pool_1d_sk_p0(
         const struct ggml_compute_params * params,
         const enum ggml_op_pool op,
-        const struct ggml_tensor * src,
         const int k,
         struct ggml_tensor * dst) {
-    assert(src->type == GGML_TYPE_F32);
-    assert(params->ith == 0);
 
-    if (params->type == GGML_TASK_INIT || params->type == GGML_TASK_FINALIZE) {
+    const struct ggml_tensor * src = dst->src[0];
+
+    assert(src->type == GGML_TYPE_F32 || src->type == GGML_TYPE_F16);
+
+    if (params->ith != 0) {
         return;
     }
 
@@ -12410,28 +15095,27 @@ static void ggml_compute_forward_pool_1d_sk_p0(
     const int64_t rs = dst->ne[0];
 
     while (cdata < data_end) {
-        const float * const srow = (const float *)cdata;
-
+        const void * srow = (const void *)cdata;
         int j = 0;
-
         for (int64_t i = 0; i < rs; ++i) {
             switch (op) {
                 case GGML_OP_POOL_AVG:   drow[i] = 0;        break;
                 case GGML_OP_POOL_MAX:   drow[i] = -FLT_MAX; break;
-                case GGML_OP_POOL_COUNT: GGML_ASSERT(false); break;
+                case GGML_OP_POOL_COUNT: GGML_ABORT("fatal error");
             }
             for (int ki = 0; ki < k; ++ki) {
+                const float srow_j = (src->type == GGML_TYPE_F32) ? ((const float*)srow)[j] : GGML_FP16_TO_FP32(((const ggml_fp16_t*)srow)[j]);
                 switch (op) {
-                    case GGML_OP_POOL_AVG:                          drow[i] += srow[j]; break;
-                    case GGML_OP_POOL_MAX:   if (srow[j] > drow[i]) drow[i]  = srow[j]; break;
-                    case GGML_OP_POOL_COUNT:                        GGML_ASSERT(false); break;
+                    case GGML_OP_POOL_AVG:                         drow[i] += srow_j; break;
+                    case GGML_OP_POOL_MAX:   if (srow_j > drow[i]) drow[i]  = srow_j; break;
+                    case GGML_OP_POOL_COUNT:                       GGML_ABORT("fatal error");
                 }
                 ++j;
             }
             switch (op) {
                 case GGML_OP_POOL_AVG:         drow[i] /= k; break;
                 case GGML_OP_POOL_MAX:                       break;
-                case GGML_OP_POOL_COUNT: GGML_ASSERT(false); break;
+                case GGML_OP_POOL_COUNT: GGML_ABORT("fatal error");
             }
         }
 
@@ -12444,7 +15128,6 @@ static void ggml_compute_forward_pool_1d_sk_p0(
 
 static void ggml_compute_forward_pool_1d(
         const struct ggml_compute_params * params,
-        const struct ggml_tensor * src0,
               struct ggml_tensor * dst) {
 
     const int32_t * opts = (const int32_t *)dst->op_params;
@@ -12455,19 +15138,20 @@ static void ggml_compute_forward_pool_1d(
     GGML_ASSERT(p0 == 0); // padding not supported
     GGML_ASSERT(k0 == s0); // only s = k supported
 
-    ggml_compute_forward_pool_1d_sk_p0(params, op, src0, k0, dst);
+    ggml_compute_forward_pool_1d_sk_p0(params, op, k0, dst);
 }
 
 // ggml_compute_forward_pool_2d
 
 static void ggml_compute_forward_pool_2d(
         const struct ggml_compute_params * params,
-        const struct ggml_tensor * src,
         struct ggml_tensor * dst) {
-    assert(src->type == GGML_TYPE_F32);
-    assert(params->ith == 0);
 
-    if (params->type == GGML_TASK_INIT || params->type == GGML_TASK_FINALIZE) {
+    const struct ggml_tensor * src = dst->src[0];
+
+    assert(src->type == GGML_TYPE_F32 || src->type == GGML_TYPE_F16);
+
+    if (params->ith != 0) {
         return;
     }
 
@@ -12500,7 +15184,7 @@ static void ggml_compute_forward_pool_2d(
                 switch (op) {
                     case GGML_OP_POOL_AVG:     *out = 0;        break;
                     case GGML_OP_POOL_MAX:     *out = -FLT_MAX; break;
-                    case GGML_OP_POOL_COUNT: GGML_ASSERT(false); break;
+                    case GGML_OP_POOL_COUNT: GGML_ABORT("fatal error");
                 }
 
                 const int ix = offset0 + ox * s0;
@@ -12508,21 +15192,22 @@ static void ggml_compute_forward_pool_2d(
 
                 for (int ky = 0; ky < k1; ++ky) {
                     if (iy + ky < 0 || iy + ky >= src->ne[1]) continue;
-                    const float * const srow = (const float *)(cdata + src->nb[1] * (iy + ky));
+                    const void * srow = (const void *)(cdata + src->nb[1] * (iy + ky));
                     for (int kx = 0; kx < k0; ++kx) {
                         int j = ix + kx;
                         if (j < 0 || j >= src->ne[0]) continue;
+                        const float srow_j = (src->type == GGML_TYPE_F32) ? ((const float*)srow)[j] : GGML_FP16_TO_FP32(((const ggml_fp16_t*)srow)[j]);
                         switch (op) {
-                            case GGML_OP_POOL_AVG:                     *out += srow[j]; break;
-                            case GGML_OP_POOL_MAX: if (srow[j] > *out) *out  = srow[j]; break;
-                            case GGML_OP_POOL_COUNT:                GGML_ASSERT(false); break;
+                            case GGML_OP_POOL_AVG:                     *out += srow_j; break;
+                            case GGML_OP_POOL_MAX: if (srow_j > *out)  *out  = srow_j; break;
+                            case GGML_OP_POOL_COUNT:               GGML_ABORT("fatal error");
                         }
                     }
                 }
                 switch (op) {
                     case GGML_OP_POOL_AVG:           *out /= ka; break;
                     case GGML_OP_POOL_MAX:                       break;
-                    case GGML_OP_POOL_COUNT: GGML_ASSERT(false); break;
+                    case GGML_OP_POOL_COUNT: GGML_ABORT("fatal error");
                 }
             }
         }
@@ -12532,675 +15217,435 @@ static void ggml_compute_forward_pool_2d(
     }
 }
 
-// ggml_compute_forward_upscale
-
-static void ggml_compute_forward_upscale_f32(
-    const struct ggml_compute_params * params,
-    const struct ggml_tensor * src0,
-    struct ggml_tensor * dst) {
-
-    if (params->type == GGML_TASK_INIT || params->type == GGML_TASK_FINALIZE) {
-        return;
-    }
-
-    GGML_ASSERT(src0->nb[0] == sizeof(float));
-
-    const int ith = params->ith;
-    const int nth = params->nth;
-
-    GGML_TENSOR_UNARY_OP_LOCALS
-
-    const int scale_factor = dst->op_params[0];
-
-    // TODO: optimize
-
-    for (int64_t i3 = 0; i3 < ne3; i3++) {
-        const int64_t i03 = i3;
-        for (int64_t i2 = ith; i2 < ne2; i2 += nth) {
-            const int64_t i02 = i2;
-            for (int64_t i1 = 0; i1 < ne1; i1++) {
-                const int64_t i01 = i1 / scale_factor;
-                for (int64_t i0 = 0; i0 < ne0; i0++) {
-                    const int64_t i00 = i0 / scale_factor;
-
-                    const float * x = (float *)((char *) src0->data + i00*nb00 + i01*nb01 + i02*nb02 + i03*nb03);
-                          float * y = (float *)((char *)  dst->data +  i0*nb0  +  i1*nb1  +  i2*nb2  +  i3*nb3);
-
-                    *y = *x;
-                }
-            }
-        }
-    }
-}
-
-static void ggml_compute_forward_upscale(
-    const struct ggml_compute_params * params,
-    const struct ggml_tensor * src0,
-    struct ggml_tensor * dst) {
-    switch (src0->type) {
-        case GGML_TYPE_F32:
-            {
-                ggml_compute_forward_upscale_f32(params, src0, dst);
-            } break;
-        default:
-            {
-                GGML_ASSERT(false);
-            } break;
-    }
-}
-
-// ggml_compute_forward_pad
-
-static void ggml_compute_forward_pad_f32(
-    const struct ggml_compute_params * params,
-    const struct ggml_tensor * src0,
-          struct ggml_tensor * dst) {
-
-    if (params->type == GGML_TASK_INIT || params->type == GGML_TASK_FINALIZE) {
-        return;
-    }
-
-    GGML_ASSERT(src0->nb[0] == sizeof(float));
-    GGML_ASSERT( dst->nb[0] == sizeof(float));
-
-    const int ith = params->ith;
-    const int nth = params->nth;
-
-    GGML_TENSOR_UNARY_OP_LOCALS
-
-    float * dst_ptr = (float *) dst->data;
-
-    // TODO: optimize
-
-    for (int64_t i2 = 0; i2 < ne2; ++i2) {
-        for (int64_t i1 = ith; i1 < ne1; i1 += nth) {
-            for (int64_t i0 = 0; i0 < ne0; ++i0) {
-                for (int64_t i3 = 0; i3 < ne3; ++i3) {
-                    const int64_t dst_idx = i3*(ne0*ne1*ne2) + i2*(ne0*ne1) + i1*ne0 + i0;
-
-                    const float * src_ptr = (const float *)((char *) src0->data + i3*nb03 + i2*nb02 + i1*nb01 + i0*nb00);
-
-                    if (i0 < ne00 && i1 < ne01 && i2 < ne02 && i3 < ne03) {
-                        dst_ptr[dst_idx] = *src_ptr;
-                    } else {
-                        dst_ptr[dst_idx] = 0;
-                    }
-                }
-            }
-        }
-    }
-}
-
-static void ggml_compute_forward_pad(
-    const struct ggml_compute_params * params,
-    const struct ggml_tensor * src0,
-    struct ggml_tensor * dst) {
-    switch (src0->type) {
-        case GGML_TYPE_F32:
-            {
-                ggml_compute_forward_pad_f32(params, src0, dst);
-            } break;
-        default:
-            {
-                GGML_ASSERT(false);
-            } break;
-    }
-}
-
-// ggml_compute_forward_argsort
-
-static void ggml_compute_forward_argsort_f32(
-    const struct ggml_compute_params * params,
-    const struct ggml_tensor * src0,
-    struct ggml_tensor * dst) {
-
-    if (params->type == GGML_TASK_INIT || params->type == GGML_TASK_FINALIZE) {
-        return;
-    }
-
-    GGML_TENSOR_UNARY_OP_LOCALS
-
-    GGML_ASSERT(nb0 == sizeof(float));
-
-    const int ith = params->ith;
-    const int nth = params->nth;
-
-    const int64_t nr = ggml_nrows(src0);
-
-    enum ggml_sort_order order = (enum ggml_sort_order) ggml_get_op_params_i32(dst, 0);
-
-    for (int64_t i = ith; i < nr; i += nth) {
-        int32_t * dst_data = (int32_t *)((char *) dst->data + i*nb1);
-        const float * src_data = (float *)((char *) src0->data + i*nb01);
-
-        for (int64_t j = 0; j < ne0; j++) {
-            dst_data[j] = j;
-        }
-
-        // C doesn't have a functional sort, so we do a bubble sort instead
-        for (int64_t j = 0; j < ne0; j++) {
-            for (int64_t k = j + 1; k < ne0; k++) {
-                if ((order == GGML_SORT_ASC && src_data[dst_data[j]] > src_data[dst_data[k]]) ||
-                    (order == GGML_SORT_DESC && src_data[dst_data[j]] < src_data[dst_data[k]])) {
-                    int32_t tmp = dst_data[j];
-                    dst_data[j] = dst_data[k];
-                    dst_data[k] = tmp;
-                }
-            }
-        }
-    }
-}
-
-static void ggml_compute_forward_argsort(
-    const struct ggml_compute_params * params,
-    const struct ggml_tensor * src0,
-    struct ggml_tensor * dst) {
-
-    switch (src0->type) {
-        case GGML_TYPE_F32:
-            {
-                ggml_compute_forward_argsort_f32(params, src0, dst);
-            } break;
-        default:
-            {
-                GGML_ASSERT(false);
-            } break;
-    }
-}
-
-// ggml_compute_forward_flash_attn
+// ggml_compute_forward_pool_2d_back
 
-static void ggml_compute_forward_flash_attn_f32(
+static void ggml_compute_forward_pool_2d_back(
         const struct ggml_compute_params * params,
-        const struct ggml_tensor * q,
-        const struct ggml_tensor * k,
-        const struct ggml_tensor * v,
-        const bool masked,
         struct ggml_tensor * dst) {
-    int64_t t0 = ggml_perf_time_us();
-    UNUSED(t0);
-
-    GGML_TENSOR_LOCALS(int64_t, neq, q,   ne)
-    GGML_TENSOR_LOCALS(size_t,  nbq, q,   nb)
-    GGML_TENSOR_LOCALS(int64_t, nek, k,   ne)
-    GGML_TENSOR_LOCALS(size_t,  nbk, k,   nb)
-    GGML_TENSOR_LOCALS(int64_t, nev, v,   ne)
-    GGML_TENSOR_LOCALS(size_t,  nbv, v,   nb)
-    GGML_TENSOR_LOCALS(int64_t, ne,  dst, ne)
-    GGML_TENSOR_LOCALS(size_t,  nb,  dst, nb)
 
-    const int ith = params->ith;
-    const int nth = params->nth;
+    const struct ggml_tensor * src  = dst->src[0];
+    const struct ggml_tensor * dstf = dst->src[1]; // forward tensor of dst
 
-    const int64_t D = neq0;
-    const int64_t N = neq1;
-    const int64_t P = nek1 - N;
-    const int64_t M = P + N;
+    assert(dst->type == GGML_TYPE_F32 || dst->type == GGML_TYPE_F16);
 
-    const int Mup = ggml_up(M, GGML_SOFT_MAX_UNROLL);
+    if (params->ith != 0) {
+        return;
+    }
 
-    GGML_ASSERT(ne0 == D);
-    GGML_ASSERT(ne1 == N);
-    GGML_ASSERT(P >= 0);
+    const int32_t * opts = (const int32_t *)dst->op_params;
+    enum ggml_op_pool op = opts[0];
+    const int k0 = opts[1];
+    const int k1 = opts[2];
+    const int s0 = opts[3];
+    const int s1 = opts[4];
+    const int p0 = opts[5];
+    const int p1 = opts[6];
 
-    GGML_ASSERT(nbq0 == sizeof(float));
-    GGML_ASSERT(nbk0 == sizeof(float));
-    GGML_ASSERT(nbv0 == sizeof(float));
+    char       * cdata  = (char       *) dst->data;
+    const char * cdataf = (const char *) dstf->data;
+    const char * const data_end = cdata + ggml_nbytes(dst);
 
-    GGML_ASSERT(neq0 == D);
-    GGML_ASSERT(nek0 == D);
-    GGML_ASSERT(nev1 == D);
+    GGML_ASSERT(params->ith == 0);
+    memset(cdata, 0, ggml_nbytes(dst));
 
-    GGML_ASSERT(neq1 == N);
-    GGML_ASSERT(nek1 == N + P);
-    GGML_ASSERT(nev1 == D);
+    const int64_t px = src->ne[0];
+    const int64_t py = src->ne[1];
+    const int64_t pa = px * py;
 
-    // dst cannot be transposed or permuted
-    GGML_ASSERT(nb0 == sizeof(float));
-    GGML_ASSERT(nb0 <= nb1);
-    GGML_ASSERT(nb1 <= nb2);
-    GGML_ASSERT(nb2 <= nb3);
+    const float * splane = (const float *) src->data;
 
-    if (params->type == GGML_TASK_INIT) {
-        return;
-    }
+    const int ka = k0 * k1;
+    const int offset0 = -p0;
+    const int offset1 = -p1;
 
-    if (params->type == GGML_TASK_FINALIZE) {
-        return;
-    }
+    while (cdata < data_end) {
+        for (int oy = 0; oy < py; ++oy) {
+            const float * const srow = splane + oy * px;
+            for (int ox = 0; ox < px; ++ox) {
+                const float grad0 = srow[ox];
 
-    // parallelize by q rows using ggml_vec_dot_f32
+                const int ix = offset0 + ox * s0;
+                const int iy = offset1 + oy * s1;
 
-    // total rows in q
-    const int nr = neq1*neq2*neq3;
+                if (op == GGML_OP_POOL_MAX) {
+                    float maxval = -FLT_MAX;
+                    int kxmax = -1;
+                    int kymax = -1;
 
-    // rows per thread
-    const int dr = (nr + nth - 1)/nth;
+                    for (int ky = 0; ky < k1; ++ky) {
+                        if (iy + ky < 0 || iy + ky >= dst->ne[1]) {
+                            continue;
+                        }
+                        const void * drowf = (const void *)(cdataf + dst->nb[1] * (iy + ky));
+                        for (int kx = 0; kx < k0; ++kx) {
+                            int j = ix + kx;
+                            if (j < 0 || j >= dst->ne[0]) {
+                                continue;
+                            }
 
-    // row range for this thread
-    const int ir0 = dr*ith;
-    const int ir1 = MIN(ir0 + dr, nr);
+                            const float val = dst->type == GGML_TYPE_F32 ?
+                                ((const float *) drowf)[j] : GGML_FP16_TO_FP32(((const ggml_fp16_t *) drowf)[j]);
+                            if (val <= maxval) {
+                                continue;
+                            }
 
-    const float scale = 1.0f/sqrtf(D);
+                            maxval = val;
+                            kxmax = kx;
+                            kymax = ky;
+                        }
+                    }
 
-    //printf("P=%d N=%d D=%d ir0=%d ir1=%d scale = %f\n", P, N, D, ir0, ir1, scale);
+                    if (kxmax == -1 || kymax == -1) {
+                        continue;
+                    }
 
-    for (int ir = ir0; ir < ir1; ++ir) {
-        // q indices
-        const int iq3 = ir/(neq2*neq1);
-        const int iq2 = (ir - iq3*neq2*neq1)/neq1;
-        const int iq1 = (ir - iq3*neq2*neq1 - iq2*neq1);
+                    void * drow = (void *)(cdata + dst->nb[1] * (iy + kymax));
+                    const int j = ix + kxmax;
+                    if (dst->type == GGML_TYPE_F32) {
+                        ((float *) drow)[j] += grad0;
+                    } else {
+                        ((ggml_fp16_t *) drow)[j] = GGML_FP32_TO_FP16(grad0 + GGML_FP16_TO_FP32(((const ggml_fp16_t *) drow)[j]));
+                    }
+                } else if (op == GGML_OP_POOL_AVG) {
+                    const float grad = grad0 / ka;
 
-        float * S = (float *) params->wdata + ith*(Mup + CACHE_LINE_SIZE_F32);
+                    for (int ky = 0; ky < k1; ++ky) {
+                        if (iy + ky < 0 || iy + ky >= dst->ne[1]) {
+                            continue;
+                        }
+                        void * drow = (void *)(cdata + dst->nb[1] * (iy + ky));
+                        for (int kx = 0; kx < k0; ++kx) {
+                            int j = ix + kx;
+                            if (j < 0 || j >= dst->ne[0]) {
+                                continue;
+                            }
 
-        for (int i = M; i < Mup; ++i) {
-            S[i] = -INFINITY;
+                            if (dst->type == GGML_TYPE_F32) {
+                                ((float *) drow)[j] += grad;
+                            } else {
+                                ((ggml_fp16_t *) drow)[j] += GGML_FP32_TO_FP16(grad);
+                            }
+                        }
+                    }
+                } else {
+                    GGML_ASSERT(false);
+                }
+            }
         }
 
-        const int64_t masked_begin = masked ? (P + iq1 + 1) : M;
-        for (int64_t ic = 0; ic < masked_begin; ++ic) {
-            // k indices
-            const int ik3 = iq3;
-            const int ik2 = iq2 % nek2;
-            const int ik1 = ic;
-
-            // S indices
-            const int i1 = ik1;
+        cdata  += dst->nb[2];
+        cdataf += dst->nb[2];
+        splane += pa;
+    }
+}
 
-            ggml_vec_dot_f32(neq0,
-                    S + i1,
-                    (float *) ((char *) k->data + (ik1*nbk1 + ik2*nbk2 + ik3*nbk3)),
-                    (float *) ((char *) q->data + (iq1*nbq1 + iq2*nbq2 + iq3*nbq3)));
-        }
+// ggml_compute_forward_upscale
 
-        // scale
-        ggml_vec_scale_f32(masked_begin, S, scale);
+static void ggml_compute_forward_upscale_f32(
+    const struct ggml_compute_params * params,
+    struct ggml_tensor * dst) {
 
-        for (int64_t i = masked_begin; i < M; i++) {
-            S[i] = -INFINITY;
-        }
+    const struct ggml_tensor * src0 = dst->src[0];
 
-        // softmax
-        // exclude known -INF S[..] values from max and loop
-        // dont forget to set their SW values to zero
-        {
-            float max = -INFINITY;
-            ggml_vec_max_f32(masked_begin, &max, S);
+    GGML_ASSERT(src0->type == GGML_TYPE_F32);
 
-            ggml_float sum = 0.0;
-            {
-#ifdef GGML_SOFT_MAX_ACCELERATE
-                max = -max;
-                vDSP_vsadd(S, 1, &max, S, 1, Mup);
-                vvexpf(S, S, &Mup);
-                ggml_vec_sum_f32(Mup, &sum, S);
-#else
-                uint16_t   scvt[GGML_SOFT_MAX_UNROLL]; UNUSED(scvt);
-                ggml_float sump[GGML_SOFT_MAX_UNROLL] = { 0.0 };
+    const int ith = params->ith;
+    const int nth = params->nth;
 
-                for (int i = 0; i < Mup; i += GGML_SOFT_MAX_UNROLL) {
-                    if (i >= masked_begin) {
-                        break;
-                    }
-                    float * SS = S + i;
+    GGML_TENSOR_UNARY_OP_LOCALS
 
-                    for (int j = 0; j < GGML_SOFT_MAX_UNROLL; ++j) {
-                        if (i + j >= masked_begin) {
-                            break;
-                        } else if (SS[j] == -INFINITY) {
-                            SS[j] = 0.0f;
-                        } else {
-#ifndef GGML_FLASH_ATTN_EXP_FP16
-                            const float val = expf(SS[j] - max);
-#else
-                            ggml_fp16_t s = GGML_FP32_TO_FP16(SS[j] - max);
-                            memcpy(&scvt[j], &s, sizeof(uint16_t));
-                            const float val = GGML_FP16_TO_FP32(ggml_table_exp_f16[scvt[j]]);
-#endif
-                            sump[j] += (ggml_float)val;
-                            SS[j] = val;
-                        }
-                    }
-                }
+    const float sf0 = (float)ne0/src0->ne[0];
+    const float sf1 = (float)ne1/src0->ne[1];
+    const float sf2 = (float)ne2/src0->ne[2];
+    const float sf3 = (float)ne3/src0->ne[3];
 
-                for (int i = 0; i < GGML_SOFT_MAX_UNROLL; i++) {
-                    sum += sump[i];
-                }
-#endif
-            }
+    // TODO: optimize
 
-            assert(sum > 0.0);
+    for (int64_t i3 = 0; i3 < ne3; i3++) {
+        const int64_t i03 = i3 / sf3;
+        for (int64_t i2 = ith; i2 < ne2; i2 += nth) {
+            const int64_t i02 = i2 / sf2;
+            for (int64_t i1 = 0; i1 < ne1; i1++) {
+                const int64_t i01 = i1 / sf1;
+                for (int64_t i0 = 0; i0 < ne0; i0++) {
+                    const int64_t i00 = i0 / sf0;
 
-            sum = 1.0/sum;
-            ggml_vec_scale_f32(masked_begin, S, sum);
+                    const float * x = (float *)((char *) src0->data + i00*nb00 + i01*nb01 + i02*nb02 + i03*nb03);
+                          float * y = (float *)((char *)  dst->data +  i0*nb0  +  i1*nb1  +  i2*nb2  +  i3*nb3);
 
-#ifndef NDEBUG
-            for (int i = 0; i < masked_begin; ++i) {
-                assert(!isnan(S[i]));
-                assert(!isinf(S[i]));
+                    *y = *x;
+                }
             }
-#endif
         }
+    }
+}
 
-        for (int64_t ic = 0; ic < nev1; ++ic) {
-            // dst indices
-            const int i1 = iq1;
-            const int i2 = iq2;
-            const int i3 = iq3;
+static void ggml_compute_forward_upscale(
+    const struct ggml_compute_params * params,
+    struct ggml_tensor * dst) {
 
-            // v indices
-            const int iv2 = iq2 % nev2;
-            const int iv3 = iq3;
+    const struct ggml_tensor * src0 = dst->src[0];
 
-            ggml_vec_dot_f32(masked_begin,
-                    (float *) ((char *) dst->data + (ic*nb0 + i1*nb1  + i2*nb2   + i3*nb3)),
-                    (float *) ((char *) v->data   + (         ic*nbv1 + iv2*nbv2 + iv3*nbv3)),
-                    S);
-        }
+    switch (src0->type) {
+        case GGML_TYPE_F32:
+            {
+                ggml_compute_forward_upscale_f32(params, dst);
+            } break;
+        default:
+            {
+                GGML_ABORT("fatal error");
+            }
     }
 }
 
-static void ggml_compute_forward_flash_attn_f16(
-        const struct ggml_compute_params * params,
-        const struct ggml_tensor * q,
-        const struct ggml_tensor * k,
-        const struct ggml_tensor * v,
-        const bool masked,
-        struct ggml_tensor * dst) {
-    int64_t t0 = ggml_perf_time_us();
-    UNUSED(t0);
 
-    GGML_TENSOR_LOCALS(int64_t, neq, q,   ne)
-    GGML_TENSOR_LOCALS(size_t,  nbq, q,   nb)
-    GGML_TENSOR_LOCALS(int64_t, nek, k,   ne)
-    GGML_TENSOR_LOCALS(size_t,  nbk, k,   nb)
-    GGML_TENSOR_LOCALS(int64_t, nev, v,   ne)
-    GGML_TENSOR_LOCALS(size_t,  nbv, v,   nb)
-    GGML_TENSOR_LOCALS(int64_t, ne,  dst, ne)
-    GGML_TENSOR_LOCALS(size_t,  nb,  dst, nb)
+// ggml_compute_forward_pad
+
+static void ggml_compute_forward_pad_f32(
+    const struct ggml_compute_params * params,
+          struct ggml_tensor * dst) {
+
+    const struct ggml_tensor * src0 = dst->src[0];
+
+    GGML_ASSERT(src0->nb[0] == sizeof(float));
+    GGML_ASSERT( dst->nb[0] == sizeof(float));
 
     const int ith = params->ith;
     const int nth = params->nth;
 
-    const int64_t D = neq0;
-    const int64_t N = neq1;
-    const int64_t P = nek1 - N;
-    const int64_t M = P + N;
+    GGML_TENSOR_UNARY_OP_LOCALS
 
-    const int Mup = ggml_up(M, GGML_SOFT_MAX_UNROLL);
+    float * dst_ptr = (float *) dst->data;
 
-    GGML_ASSERT(ne0 == D);
-    GGML_ASSERT(ne1 == N);
-    GGML_ASSERT(P >= 0);
+    // TODO: optimize
 
-    GGML_ASSERT(nbq0 == sizeof(ggml_fp16_t));
-    GGML_ASSERT(nbk0 == sizeof(ggml_fp16_t));
-    GGML_ASSERT(nbv0 == sizeof(ggml_fp16_t));
+    for (int64_t i2 = 0; i2 < ne2; ++i2) {
+        for (int64_t i1 = ith; i1 < ne1; i1 += nth) {
+            for (int64_t i0 = 0; i0 < ne0; ++i0) {
+                for (int64_t i3 = 0; i3 < ne3; ++i3) {
+                    const int64_t dst_idx = i3*(ne0*ne1*ne2) + i2*(ne0*ne1) + i1*ne0 + i0;
 
-    GGML_ASSERT(neq0 == D);
-    GGML_ASSERT(nek0 == D);
-    GGML_ASSERT(nev1 == D);
+                    const float * src_ptr = (const float *)((char *) src0->data + i3*nb03 + i2*nb02 + i1*nb01 + i0*nb00);
 
-    GGML_ASSERT(neq1 == N);
-    GGML_ASSERT(nek1 == N + P);
-    GGML_ASSERT(nev1 == D);
+                    if (i0 < ne00 && i1 < ne01 && i2 < ne02 && i3 < ne03) {
+                        dst_ptr[dst_idx] = *src_ptr;
+                    } else {
+                        dst_ptr[dst_idx] = 0;
+                    }
+                }
+            }
+        }
+    }
+}
 
-    // dst cannot be transposed or permuted
-    GGML_ASSERT(nb0 == sizeof(float));
-    GGML_ASSERT(nb0 <= nb1);
-    GGML_ASSERT(nb1 <= nb2);
-    GGML_ASSERT(nb2 <= nb3);
+static void ggml_compute_forward_pad(
+    const struct ggml_compute_params * params,
+    struct ggml_tensor * dst) {
 
-    if (params->type == GGML_TASK_INIT) {
-        return;
-    }
+    const struct ggml_tensor * src0 = dst->src[0];
 
-    if (params->type == GGML_TASK_FINALIZE) {
-        return;
+    switch (src0->type) {
+        case GGML_TYPE_F32:
+            {
+                ggml_compute_forward_pad_f32(params, dst);
+            } break;
+        default:
+            {
+                GGML_ABORT("fatal error");
+            }
     }
+}
 
-    // parallelize by q rows using ggml_vec_dot_f32
 
-    // total rows in q
-    const int nr = neq1*neq2*neq3;
+// ggml_compute_forward_arange
 
-    // rows per thread
-    const int dr = (nr + nth - 1)/nth;
+static void ggml_compute_forward_arange_f32(
+    const struct ggml_compute_params * params,
+    struct ggml_tensor * dst) {
 
-    // row range for this thread
-    const int ir0 = dr*ith;
-    const int ir1 = MIN(ir0 + dr, nr);
+    GGML_ASSERT(dst->nb[0] == sizeof(float));
 
-    const float scale = 1.0f/sqrtf(D);
+    const int ith = params->ith;
+    const int nth = params->nth;
 
-    //printf("P=%d N=%d D=%d ir0=%d ir1=%d scale = %f\n", P, N, D, ir0, ir1, scale);
+    const float start = ggml_get_op_params_f32(dst, 0);
+    const float stop  = ggml_get_op_params_f32(dst, 1);
+    const float step  = ggml_get_op_params_f32(dst, 2);
 
-    for (int ir = ir0; ir < ir1; ++ir) {
-        // q indices
-        const int iq3 = ir/(neq2*neq1);
-        const int iq2 = (ir - iq3*neq2*neq1)/neq1;
-        const int iq1 = (ir - iq3*neq2*neq1 - iq2*neq1);
+    const int64_t steps = (int64_t) ceilf((stop - start) / step);
 
-        float * S = (float *) params->wdata + ith*(2*Mup + CACHE_LINE_SIZE_F32);
+    GGML_ASSERT(ggml_nelements(dst) == steps);
 
-        for (int i = M; i < Mup; ++i) {
-            S[i] = -INFINITY;
-        }
+    for (int64_t i = ith; i < steps; i+= nth) {
+        float value = start + step * i;
+        ((float *)dst->data)[i] = value;
+    }
+}
 
-        if (GGML_VEC_DOT_UNROLL > 2 || nek1 % GGML_VEC_DOT_UNROLL != 0) {
-            for (int64_t ic = 0; ic < nek1; ++ic) {
-                // k indices
-                const int ik3 = iq3;
-                const int ik2 = iq2 % nek2;
-                const int ik1 = ic;
+static void ggml_compute_forward_arange(
+    const struct ggml_compute_params * params,
+    struct ggml_tensor * dst) {
+    switch (dst->type) {
+        case GGML_TYPE_F32:
+            {
+                ggml_compute_forward_arange_f32(params, dst);
+            } break;
+        default:
+            {
+                GGML_ABORT("fatal error");
+            }
+    }
+}
 
-                // S indices
-                const int i1 = ik1;
+static void ggml_compute_forward_timestep_embedding_f32(
+    const struct ggml_compute_params * params,
+    struct ggml_tensor * dst) {
 
-                ggml_vec_dot_f16(neq0,
-                        S + i1,
-                        (ggml_fp16_t *) ((char *) k->data + (ik1*nbk1 + ik2*nbk2 + ik3*nbk3)),
-                        (ggml_fp16_t *) ((char *) q->data + (iq1*nbq1 + iq2*nbq2 + iq3*nbq3)));
-            }
-        } else {
-            for (int64_t ic = 0; ic < nek1; ic += GGML_VEC_DOT_UNROLL) {
-                // k indices
-                const int ik3 = iq3;
-                const int ik2 = iq2 % nek2;
-                const int ik1 = ic;
+    const struct ggml_tensor * src0 = dst->src[0];
 
-                // S indices
-                const int i1 = ik1;
+    GGML_ASSERT(src0->nb[0] == sizeof(float));
 
-                ggml_vec_dot_f16_unroll(neq0, nbk1,
-                        S + i1,
-                        ((char *) k->data + (ik1*nbk1 + ik2*nbk2 + ik3*nbk3)),
-                        (ggml_fp16_t *) ((char *) q->data + (iq1*nbq1 + iq2*nbq2 + iq3*nbq3)));
-            }
-        }
+    const int ith = params->ith;
+    const int nth = params->nth;
 
-        // scale
-        ggml_vec_scale_f32(nek1, S, scale);
+    GGML_TENSOR_UNARY_OP_LOCALS
 
-        if (masked) {
-            for (int64_t i = P; i < M; i++) {
-                if (i > P + iq1) {
-                    S[i] = -INFINITY;
-                }
-            }
-        }
+    const int dim = ggml_get_op_params_i32(dst, 0);
+    const int max_period = ggml_get_op_params_i32(dst, 1);
 
-        // softmax
-        // todo: exclude known -INF S[..] values from max and loop, assuming their results to be zero.
-        // dont forget to set their S values to zero
-        {
-            float max = -INFINITY;
-            ggml_vec_max_f32(M, &max, S);
+    int half = dim / 2;
 
-            ggml_float sum = 0.0;
-            {
-#ifdef GGML_SOFT_MAX_ACCELERATE
-                max = -max;
-                vDSP_vsadd(S, 1, &max, S, 1, Mup);
-                vvexpf(S, S, &Mup);
-                ggml_vec_sum_f32(Mup, &sum, S);
-#else
-                uint16_t   scvt[GGML_SOFT_MAX_UNROLL];
-                ggml_float sump[GGML_SOFT_MAX_UNROLL] = { 0.0 };
+    for (int64_t i = 0; i < ne00; i++) {
+        float * embed_data = (float *)((char *)  dst->data +  i*nb1);
+        for (int64_t j = ith; j < half; j += nth) {
+            float timestep = ((float *)src0->data)[i];
+            float freq = (float)expf(-logf(max_period) * j / half);
+            float arg = timestep * freq;
+            embed_data[j] = cosf(arg);
+            embed_data[j + half] = sinf(arg);
+        }
+        if (dim % 2 != 0 && ith == 0) {
+            embed_data[dim] = 0.f;
+        }
+    }
+}
 
-                for (int i = 0; i < Mup; i += GGML_SOFT_MAX_UNROLL) {
-                    float * SS = S + i;
+static void ggml_compute_forward_timestep_embedding(
+    const struct ggml_compute_params * params,
+    struct ggml_tensor * dst) {
 
-                    for (int j = 0; j < GGML_SOFT_MAX_UNROLL; ++j) {
-                        if (SS[j] == -INFINITY) {
-                            SS[j] = 0.0f;
-                        } else {
-                            ggml_fp16_t s = GGML_FP32_TO_FP16(SS[j] - max);
-                            memcpy(&scvt[j], &s, sizeof(uint16_t));
-                            const float val = GGML_FP16_TO_FP32(ggml_table_exp_f16[scvt[j]]);
-                            sump[j] += (ggml_float)val;
-                            SS[j] = val;
-                        }
-                    }
-                }
+    const struct ggml_tensor * src0 = dst->src[0];
 
-                for (int i = 0; i < GGML_SOFT_MAX_UNROLL; i++) {
-                    sum += sump[i];
-                }
-#endif
+    switch (src0->type) {
+        case GGML_TYPE_F32:
+            {
+                ggml_compute_forward_timestep_embedding_f32(params, dst);
+            } break;
+        default:
+            {
+                GGML_ABORT("fatal error");
             }
+    }
+}
 
-            assert(sum > 0.0);
+// ggml_compute_forward_argsort
 
-            sum = 1.0/sum;
-            ggml_vec_scale_f32(M, S, sum);
+static void ggml_compute_forward_argsort_f32(
+    const struct ggml_compute_params * params,
+    struct ggml_tensor * dst) {
 
-#ifndef NDEBUG
-            for (int i = 0; i < M; ++i) {
-                assert(!isnan(S[i]));
-                assert(!isinf(S[i]));
-            }
-#endif
-        }
+    const struct ggml_tensor * src0 = dst->src[0];
 
-        ggml_fp16_t * S16 = (ggml_fp16_t *) ((float *) params->wdata + ith*(2*Mup + CACHE_LINE_SIZE_F32) + Mup);
+    GGML_TENSOR_UNARY_OP_LOCALS
 
-        for (int64_t i = 0; i < M; i++) {
-            S16[i] = GGML_FP32_TO_FP16(S[i]);
-        }
+    GGML_ASSERT(nb0 == sizeof(float));
 
-        // todo: exclude known zero S[..] values from dot (reducing nev0 and increasing begin of v and S16).
-        if (GGML_VEC_DOT_UNROLL == 1 || (nev1 % GGML_VEC_DOT_UNROLL != 0)) {
-            for (int64_t ic = 0; ic < nev1; ++ic) {
-                // dst indices
-                const int i1 = iq1;
-                const int i2 = iq2;
-                const int i3 = iq3;
+    const int ith = params->ith;
+    const int nth = params->nth;
 
-                // v indices
-                const int iv2 = iq2 % nev2;
-                const int iv3 = iq3;
+    const int64_t nr = ggml_nrows(src0);
 
-                ggml_vec_dot_f16(nev0,
-                        (float *)       ((char *) dst->data + (ic*nb0 + i1*nb1  + i2*nb2   + i3*nb3)),
-                        (ggml_fp16_t *) ((char *) v->data   + (         ic*nbv1 + iv2*nbv2 + iv3*nbv3)),
-                        S16);
-            }
-        } else {
-            for (int64_t ic = 0; ic < nev1; ic += GGML_VEC_DOT_UNROLL) {
-                // dst indices
-                const int i1 = iq1;
-                const int i2 = iq2;
-                const int i3 = iq3;
+    enum ggml_sort_order order = (enum ggml_sort_order) ggml_get_op_params_i32(dst, 0);
+
+    for (int64_t i = ith; i < nr; i += nth) {
+        int32_t * dst_data = (int32_t *)((char *) dst->data + i*nb1);
+        const float * src_data = (float *)((char *) src0->data + i*nb01);
 
-                // v indices
-                const int iv2 = iq2 % nev2;
-                const int iv3 = iq3;
+        for (int64_t j = 0; j < ne0; j++) {
+            dst_data[j] = j;
+        }
 
-                ggml_vec_dot_f16_unroll(nev0, nbv1,
-                        (float *) ((char *) dst->data + (ic*nb0 + i1*nb1  + i2*nb2   + i3*nb3)),
-                        ((char *)             v->data + (         ic*nbv1 + iv2*nbv2 + iv3*nbv3)),
-                        S16);
+        // C doesn't have a functional sort, so we do a bubble sort instead
+        for (int64_t j = 0; j < ne0; j++) {
+            for (int64_t k = j + 1; k < ne0; k++) {
+                if ((order == GGML_SORT_ORDER_ASC  && src_data[dst_data[j]] > src_data[dst_data[k]]) ||
+                    (order == GGML_SORT_ORDER_DESC && src_data[dst_data[j]] < src_data[dst_data[k]])) {
+                    int32_t tmp = dst_data[j];
+                    dst_data[j] = dst_data[k];
+                    dst_data[k] = tmp;
+                }
             }
         }
     }
 }
 
-static void ggml_compute_forward_flash_attn(
-        const struct ggml_compute_params * params,
-        const struct ggml_tensor * q,
-        const struct ggml_tensor * k,
-        const struct ggml_tensor * v,
-        const bool masked,
-        struct ggml_tensor * dst) {
-    switch (q->type) {
-        case GGML_TYPE_F16:
-            {
-                ggml_compute_forward_flash_attn_f16(params, q, k, v, masked, dst);
-            } break;
+static void ggml_compute_forward_argsort(
+    const struct ggml_compute_params * params,
+    struct ggml_tensor * dst) {
+
+    const struct ggml_tensor * src0 = dst->src[0];
+
+    switch (src0->type) {
         case GGML_TYPE_F32:
             {
-                ggml_compute_forward_flash_attn_f32(params, q, k, v, masked, dst);
+                ggml_compute_forward_argsort_f32(params, dst);
             } break;
         default:
             {
-                GGML_ASSERT(false);
-            } break;
+                GGML_ABORT("fatal error");
+            }
     }
 }
 
-// ggml_compute_forward_flash_ff
+// ggml_compute_forward_flash_attn_ext
 
-static void ggml_compute_forward_flash_ff_f16(
+static void ggml_compute_forward_flash_attn_ext_f16(
         const struct ggml_compute_params * params,
-        const struct ggml_tensor * a,  // F16
-        const struct ggml_tensor * b0, // F16 fc_w
-        const struct ggml_tensor * b1, // F32 fc_b
-        const struct ggml_tensor * c0, // F16 proj_w
-        const struct ggml_tensor * c1, // F32 proj_b
+        const struct ggml_tensor * q,
+        const struct ggml_tensor * k,
+        const struct ggml_tensor * v,
+        const struct ggml_tensor * mask,
         struct ggml_tensor * dst) {
-    int64_t t0 = ggml_perf_time_us();
-    UNUSED(t0);
-
-    GGML_TENSOR_LOCALS(int64_t, nea,  a,   ne)
-    GGML_TENSOR_LOCALS(size_t,  nba,  a,   nb)
-    GGML_TENSOR_LOCALS(int64_t, neb0, b0,  ne)
-    GGML_TENSOR_LOCALS(size_t,  nbb0, b0,  nb)
-    GGML_TENSOR_LOCALS(int64_t, neb1, b1,  ne)
-    GGML_TENSOR_LOCALS(size_t,  nbb1, b1,  nb)
-    GGML_TENSOR_LOCALS(int64_t, nec0, c0,  ne)
-    GGML_TENSOR_LOCALS(size_t,  nbc0, c0,  nb)
-    GGML_TENSOR_LOCALS(int64_t, nec1, c1,  ne)
-    GGML_TENSOR_LOCALS(size_t,  nbc1, c1,  nb)
-    GGML_TENSOR_LOCALS(int64_t, ne,   dst, ne)
-    GGML_TENSOR_LOCALS(size_t,  nb,   dst, nb)
+
+    GGML_TENSOR_LOCALS(int64_t, neq, q,   ne)
+    GGML_TENSOR_LOCALS(size_t,  nbq, q,   nb)
+    GGML_TENSOR_LOCALS(int64_t, nek, k,   ne)
+    GGML_TENSOR_LOCALS(size_t,  nbk, k,   nb)
+    GGML_TENSOR_LOCALS(int64_t, nev, v,   ne)
+    GGML_TENSOR_LOCALS(size_t,  nbv, v,   nb)
+    GGML_TENSOR_LOCALS(int64_t, ne,  dst, ne)
+    GGML_TENSOR_LOCALS(size_t,  nb,  dst, nb)
 
     const int ith = params->ith;
     const int nth = params->nth;
 
-    const int64_t D = nea0;
-    //const int64_t N = nea1;
-    const int64_t M = neb01;
+    const int64_t D = neq0;
+    const int64_t N = neq1;
 
-    GGML_ASSERT(ne0 == nea0);
-    GGML_ASSERT(ne1 == nea1);
-    GGML_ASSERT(ne2 == nea2);
+    GGML_ASSERT(ne0 == D);
+    GGML_ASSERT(ne2 == N);
 
-    GGML_ASSERT(nba0  == sizeof(ggml_fp16_t));
-    GGML_ASSERT(nbb00 == sizeof(ggml_fp16_t));
-    GGML_ASSERT(nbb10 == sizeof(float));
-    GGML_ASSERT(nbc00 == sizeof(ggml_fp16_t));
-    GGML_ASSERT(nbc10 == sizeof(float));
+    // input tensor rows must be contiguous
+    GGML_ASSERT(nbq0 == ggml_type_size(q->type));
+    GGML_ASSERT(nbk0 == ggml_type_size(k->type));
+    GGML_ASSERT(nbv0 == ggml_type_size(v->type));
 
-    GGML_ASSERT(neb00 == D);
-    GGML_ASSERT(neb01 == M);
-    GGML_ASSERT(neb10 == M);
-    GGML_ASSERT(neb11 == 1);
+    GGML_ASSERT(neq0 == D);
+    GGML_ASSERT(nek0 == D);
+    GGML_ASSERT(nev0 == D);
 
-    GGML_ASSERT(nec00 == M);
-    GGML_ASSERT(nec01 == D);
-    GGML_ASSERT(nec10 == D);
-    GGML_ASSERT(nec11 == 1);
+    GGML_ASSERT(neq1 == N);
+    GGML_ASSERT(nev0 == D);
 
     // dst cannot be transposed or permuted
     GGML_ASSERT(nb0 == sizeof(float));
@@ -13208,18 +15653,17 @@ static void ggml_compute_forward_flash_ff_f16(
     GGML_ASSERT(nb1 <= nb2);
     GGML_ASSERT(nb2 <= nb3);
 
-    if (params->type == GGML_TASK_INIT) {
-        return;
-    }
+    // broadcast factors
+    const int64_t rk2 = neq2/nek2;
+    const int64_t rk3 = neq3/nek3;
 
-    if (params->type == GGML_TASK_FINALIZE) {
-        return;
-    }
+    const int64_t rv2 = neq2/nev2;
+    const int64_t rv3 = neq3/nev3;
 
-    // parallelize by a rows using ggml_vec_dot_f32
+    // parallelize by q rows using ggml_vec_dot_f32
 
-    // total rows in a
-    const int nr = nea1*nea2*nea3;
+    // total rows in q
+    const int nr = neq1*neq2*neq3;
 
     // rows per thread
     const int dr = (nr + nth - 1)/nth;
@@ -13228,83 +15672,173 @@ static void ggml_compute_forward_flash_ff_f16(
     const int ir0 = dr*ith;
     const int ir1 = MIN(ir0 + dr, nr);
 
+    float scale         = 1.0f;
+    float max_bias      = 0.0f;
+    float logit_softcap = 0.0f;
+
+    memcpy(&scale,         (float *) dst->op_params + 0, sizeof(float));
+    memcpy(&max_bias,      (float *) dst->op_params + 1, sizeof(float));
+    memcpy(&logit_softcap, (float *) dst->op_params + 2, sizeof(float));
+
+    if (logit_softcap != 0) {
+        scale /= logit_softcap;
+    }
+
+    const uint32_t n_head      = neq2;
+    const uint32_t n_head_log2 = 1u << (uint32_t) floor(log2(n_head));
+
+    const float m0 = powf(2.0f, -(max_bias       ) / n_head_log2);
+    const float m1 = powf(2.0f, -(max_bias / 2.0f) / n_head_log2);
+
+    enum ggml_type    const k_vec_dot_type = type_traits[k->type].vec_dot_type;
+    ggml_from_float_t const q_to_vec_dot   = type_traits[k_vec_dot_type].from_float;
+    ggml_vec_dot_t    const kq_vec_dot     = type_traits[k->type].vec_dot;
+    ggml_to_float_t   const v_to_float     = type_traits[v->type].to_float;
+
+    // loop over n_batch and n_head
     for (int ir = ir0; ir < ir1; ++ir) {
-        // a indices
-        const int ia3 = ir/(nea2*nea1);
-        const int ia2 = (ir - ia3*nea2*nea1)/nea1;
-        const int ia1 = (ir - ia3*nea2*nea1 - ia2*nea1);
+        // q indices
+        const int iq3 = ir/(neq2*neq1);
+        const int iq2 = (ir - iq3*neq2*neq1)/neq1;
+        const int iq1 = (ir - iq3*neq2*neq1 - iq2*neq1);
 
-        float * S = (float *) params->wdata + ith*(2*M + CACHE_LINE_SIZE_F32);
+        const uint32_t h = iq2; // head index
+        const float slope = (max_bias > 0.0f) ? h < n_head_log2 ? powf(m0, h + 1) : powf(m1, 2*(h - n_head_log2) + 1) : 1.0f;
 
-        for (int64_t ic = 0; ic < neb01; ++ic) {
-            // b0 indices
-            const int ib03 = ia3;
-            const int ib02 = ia2;
-            const int ib01 = ic;
+        float S = 0.0f;      // sum
+        float M = -INFINITY; // maximum KQ value
 
-            // S indices
-            const int i1 = ib01;
+        float       * VKQ32 = (float       *) params->wdata + ith*(3*D + CACHE_LINE_SIZE_F32); // FP32 VKQ accumulator
+        float       * V32   =                 (VKQ32 + 1*D); // (temporary) FP32 V buffer
+        ggml_fp16_t * VKQ16 = (ggml_fp16_t *) (VKQ32 + 1*D); // (temporary) FP16 VKQ accumulator
+        ggml_fp16_t * Q_q   = (ggml_fp16_t *) (VKQ32 + 2*D); // (temporary) buffer for Q converted to quantized/FP16
 
-            ggml_vec_dot_f16(nea0,
-                    S + i1,
-                    (ggml_fp16_t *) ((char *) b0->data + (ib01*nbb01 + ib02*nbb02 + ib03*nbb03)),
-                    (ggml_fp16_t *) ((char *)  a->data + ( ia1*nba1  +  ia2*nba2  +  ia3*nba3)));
+        if (v->type == GGML_TYPE_F16) {
+            memset(VKQ16, 0, D*sizeof(ggml_fp16_t));
+        } else {
+            memset(VKQ32, 0, D*sizeof(float));
         }
 
-        ggml_vec_add_f32(neb01, S, S, (float *) b1->data);
-        //ggml_vec_gelu_f32(neb01, S, S);
+        const ggml_fp16_t * mp = mask ? (ggml_fp16_t *)((char *) mask->data + iq1*mask->nb[1]) : NULL;
 
-        ggml_fp16_t * S16 = (ggml_fp16_t *) ((float *) params->wdata + ith*(2*M + CACHE_LINE_SIZE_F32) + M);
+        // k indices
+        const int ik3 = iq3 / rk3;
+        const int ik2 = iq2 / rk2;
 
-        for (int64_t i = 0; i < M; i++) {
-            S16[i] = GGML_FP32_TO_FP16(S[i]);
-        }
+        // v indices
+        const int iv3 = iq3 / rv3;
+        const int iv2 = iq2 / rv2;
 
-        ggml_vec_gelu_f16(neb01, S16, S16);
+        const float * pq = (const float *) ((char *) q->data + (iq1*nbq1 + iq2*nbq2 + iq3*nbq3));
+        q_to_vec_dot(pq, Q_q, D);
 
-        {
-            // dst indices
-            const int i1 = ia1;
-            const int i2 = ia2;
-            const int i3 = ia3;
+        // online softmax / attention
+        // loop over n_kv and n_head_kv
+        // ref: https://arxiv.org/pdf/2112.05682.pdf
+        for (int64_t ic = 0; ic < nek1; ++ic) {
+            const float mv = mp ? slope*GGML_FP16_TO_FP32(mp[ic]) : 0.0f;
+            if (mv == -INFINITY) {
+                continue;
+            }
+
+            float s; // KQ value
+
+            const char * k_data = (const char *) k->data + ( ic*nbk1 + ik2*nbk2 + ik3*nbk3);
+            kq_vec_dot(D, &s, 0, k_data, 0, Q_q, 0, 1);
+
+            s = s*scale; // scale KQ value
+
+            if (logit_softcap != 0.0f) {
+                s = logit_softcap*tanhf(s);
+            }
+
+            s += mv; // apply mask
+
+            const float Mold = M;
+
+            float ms = 1.0f; // upon new higher max val, scale VKQ and KQ sum with this value
+            float vs = 1.0f; // post-softmax KQ value, expf(s - M)
+
+            const char * v_data = ((const char *) v->data + (ic*nbv1 + iv2*nbv2 + iv3*nbv3));
 
-            for (int64_t ic = 0; ic < nec01; ++ic) {
+            if (v->type == GGML_TYPE_F16) {
+                if (s > M) {
+                    // s is new maximum, ms < 1.0f, vs == expf(s - s) == 1.0f
+                    M = s;
+                    ms = expf(Mold - M);
 
-                ggml_vec_dot_f16(neb01,
-                        (float *)       ((char *) dst->data + (ic*nb0 + i1*nb1   + i2*nb2   + i3*nb3)),
-                        (ggml_fp16_t *) ((char *) c0->data  + (         ic*nbc01 + i2*nbc02 + i3*nbc03)),
-                        S16);
+                    // V = V*expf(Mold - M)
+                    ggml_vec_scale_f16(D, VKQ16, ms);
+                } else {
+                    // no new maximum, ms == 1.0f, vs != 1.0f
+                    vs = expf(s - M);
+                }
+
+                // V += v*expf(s - M)
+                ggml_vec_mad_f16(D, VKQ16, (const ggml_fp16_t *) v_data, vs);
+            } else {
+                if (s > M) {
+                    // s is new maximum, ms < 1.0f, vs == expf(s - s) == 1.0f
+                    M = s;
+                    ms = expf(Mold - M);
+
+                    // V = V*expf(Mold - M)
+                    ggml_vec_scale_f32(D, VKQ32, ms);
+                } else {
+                    // no new maximum, ms == 1.0f, vs != 1.0f
+                    vs = expf(s - M);
+                }
+
+                v_to_float(v_data, V32, D);
+
+                // V += v*expf(s - M)
+                ggml_vec_mad_f32(D, VKQ32, V32, vs);
             }
 
-            ggml_vec_add_f32(nec01,
-                    (float *) ((char *) dst->data + (i1*nb1 + i2*nb2 + i3*nb3)),
-                    (float *) ((char *) dst->data + (i1*nb1 + i2*nb2 + i3*nb3)),
-                    (float *) c1->data);
+            S = S*ms + vs; // scale and increment sum with partial sum
+        }
+
+        if (v->type == GGML_TYPE_F16) {
+            for (int64_t d = 0; d < D; ++d) {
+                VKQ32[d] = GGML_FP16_TO_FP32(VKQ16[d]);
+            }
         }
+
+        // V /= S
+        const float S_inv = 1.0f/S;
+        ggml_vec_scale_f32(D, VKQ32, S_inv);
+
+        // dst indices
+        const int i1 = iq1;
+        const int i2 = iq2;
+        const int i3 = iq3;
+
+        // original
+        //memcpy((char *) dst->data + (i1*nb1 + i2*nb2 + i3*nb3), V, nev0*sizeof(float));
+
+        // permute(0, 2, 1, 3)
+        memcpy((char *) dst->data + (i3*ne2*ne1 + i2 + i1*ne1)*nb1, VKQ32, nb1);
     }
 }
 
-static void ggml_compute_forward_flash_ff(
+static void ggml_compute_forward_flash_attn_ext(
         const struct ggml_compute_params * params,
-        const struct ggml_tensor * a,
-        const struct ggml_tensor * b0,
-        const struct ggml_tensor * b1,
-        const struct ggml_tensor * c0,
-        const struct ggml_tensor * c1,
+        const struct ggml_tensor * q,
+        const struct ggml_tensor * k,
+        const struct ggml_tensor * v,
+        const struct ggml_tensor * mask,
         struct ggml_tensor * dst) {
-    switch (b0->type) {
-        case GGML_TYPE_F16:
-            {
-                ggml_compute_forward_flash_ff_f16(params, a, b0, b1, c0, c1, dst);
-            } break;
-        case GGML_TYPE_F32:
+    switch (dst->op_params[3]) {
+        case GGML_PREC_DEFAULT:
+        case GGML_PREC_F32:
             {
-                GGML_ASSERT(false); // TODO
+                // uses F32 accumulators
+                ggml_compute_forward_flash_attn_ext_f16(params, q, k, v, mask, dst);
             } break;
         default:
             {
-                GGML_ASSERT(false);
-            } break;
+                GGML_ABORT("fatal error");
+            }
     }
 }
 
@@ -13312,14 +15846,13 @@ static void ggml_compute_forward_flash_ff(
 
 static void ggml_compute_forward_flash_attn_back_f32(
         const struct ggml_compute_params * params,
-        const struct ggml_tensor * q,
-        const struct ggml_tensor * k,
-        const struct ggml_tensor * v,
-        const struct ggml_tensor * d,
         const bool masked,
               struct ggml_tensor * dst) {
-    int64_t t0 = ggml_perf_time_us();
-    UNUSED(t0);
+
+    const struct ggml_tensor * q = dst->src[0];
+    const struct ggml_tensor * k = dst->src[1];
+    const struct ggml_tensor * v = dst->src[2];
+    const struct ggml_tensor * d = dst->src[3];
 
     GGML_TENSOR_LOCALS(int64_t, neq, q,   ne)
     GGML_TENSOR_LOCALS(size_t,  nbq, q,   nb)
@@ -13367,16 +15900,10 @@ static void ggml_compute_forward_flash_attn_back_f32(
     GGML_ASSERT(nb1 <= nb2);
     GGML_ASSERT(nb2 <= nb3);
 
-    if (params->type == GGML_TASK_INIT) {
-        if (ith == 0) {
-            memset(dst->data, 0, nb0*ne0*ne1*ne2*ne3);
-        }
-        return;
-    }
-
-    if (params->type == GGML_TASK_FINALIZE) {
-        return;
+    if (ith == 0) {
+        memset(dst->data, 0, nb0*ne0*ne1*ne2*ne3);
     }
+    ggml_barrier(params->threadpool);
 
     const int64_t elem_q = ggml_nelements(q);
     const int64_t elem_k = ggml_nelements(k);
@@ -13460,9 +15987,9 @@ static void ggml_compute_forward_flash_attn_back_f32(
                     const int i1 = ik1;
 
                     ggml_vec_dot_f32(neq0,
-                            S + i1,
-                            (float *) ((char *) k->data + (ik1*nbk1 + ik2*nbk2 + ik3*nbk3)),
-                            (float *) ((char *) q->data + (iq1*nbq1 + iq2*nbq2 + iq3*nbq3)));
+                            S + i1, 0,
+                            (float *) ((char *) k->data + (ik1*nbk1 + ik2*nbk2 + ik3*nbk3)), 0,
+                            (float *) ((char *) q->data + (iq1*nbq1 + iq2*nbq2 + iq3*nbq3)), 0, 1);
                 }
 
                 // scale
@@ -13487,38 +16014,7 @@ static void ggml_compute_forward_flash_attn_back_f32(
                         vvexpf(SM, SM, &Mup);
                         ggml_vec_sum_f32(Mup, &sum, SM);
 #else
-                        uint16_t   scvt[GGML_SOFT_MAX_UNROLL]; UNUSED(scvt);
-                        ggml_float sump[GGML_SOFT_MAX_UNROLL] = { 0.0 };
-
-                        for (int i = 0; i < Mup; i += GGML_SOFT_MAX_UNROLL) {
-                            if (i >= masked_begin) {
-                                break;
-                            }
-                            float * SR =  S + i;
-                            float * SW = SM + i;
-
-                            for (int j = 0; j < GGML_SOFT_MAX_UNROLL; ++j) {
-                                if (i + j >= masked_begin) {
-                                    break;
-                                } else if (SR[j] == -INFINITY) {
-                                    SW[j] = 0.0f;
-                                } else {
-#ifndef GGML_FLASH_ATTN_EXP_FP16
-                                    const float val = expf(SR[j] - max);
-#else
-                                    ggml_fp16_t s = GGML_FP32_TO_FP16(SR[j] - max);
-                                    memcpy(&scvt[j], &s, sizeof(uint16_t));
-                                    const float val = GGML_FP16_TO_FP32(ggml_table_exp_f16[scvt[j]]);
-#endif
-                                    sump[j] += (ggml_float)val;
-                                    SW[j] = val;
-                                }
-                            }
-                        }
-
-                        for (int i = 0; i < GGML_SOFT_MAX_UNROLL; i++) {
-                            sum += sump[i];
-                        }
+                        sum = ggml_vec_soft_max_f32(Mup, SM, S, max);
 #endif
                     }
 
@@ -13607,7 +16103,7 @@ static void ggml_compute_forward_flash_attn_back_f32(
 
                 // S = SM * (S - dot(SM, S))
                 float dot_SM_gradSM = 0;
-                ggml_vec_dot_f32 (masked_begin, &dot_SM_gradSM, SM, S);
+                ggml_vec_dot_f32 (masked_begin, &dot_SM_gradSM, 0, SM, 0, S, 0, 1);
                 ggml_vec_acc1_f32(M, S, -dot_SM_gradSM);
                 ggml_vec_mul_f32 (masked_begin, S, S, SM);
 
@@ -13665,21 +16161,182 @@ static void ggml_compute_forward_flash_attn_back_f32(
 
 static void ggml_compute_forward_flash_attn_back(
         const struct ggml_compute_params * params,
-        const struct ggml_tensor * q,
-        const struct ggml_tensor * k,
-        const struct ggml_tensor * v,
-        const struct ggml_tensor * d,
         const bool masked,
         struct ggml_tensor * dst) {
+
+    const struct ggml_tensor * q = dst->src[0];
+
     switch (q->type) {
         case GGML_TYPE_F32:
             {
-                ggml_compute_forward_flash_attn_back_f32(params, q, k, v, d, masked, dst);
+                ggml_compute_forward_flash_attn_back_f32(params, masked, dst);
+            } break;
+        default:
+            {
+                GGML_ABORT("fatal error");
+            }
+    }
+}
+
+// ggml_compute_forward_ssm_conv
+
+static void ggml_compute_forward_ssm_conv_f32(
+        const struct ggml_compute_params * params,
+        struct ggml_tensor * dst) {
+    const struct ggml_tensor * src0 = dst->src[0]; // conv_x
+    const struct ggml_tensor * src1 = dst->src[1]; // conv1d.weight
+
+    const int ith = params->ith;
+    const int nth = params->nth;
+
+    const int nc  = src1->ne[0]; // d_conv
+    const int ncs = src0->ne[0]; // d_conv - 1 + n_t
+    const int nr  = src0->ne[1]; // d_inner
+    const int n_t =  dst->ne[1]; // tokens per sequence
+    const int n_s =  dst->ne[2]; // number of sequences in the batch
+
+    GGML_ASSERT( dst->ne[0] == nr);
+    GGML_ASSERT(src0->nb[0] == sizeof(float));
+    GGML_ASSERT(src1->nb[0] == sizeof(float));
+    GGML_ASSERT(src0->nb[1] == src0->ne[0]*sizeof(float));
+
+    // rows per thread
+    const int dr = (nr + nth - 1)/nth;
+
+    // row range for this thread
+    const int ir0 = dr*ith;
+    const int ir1 = MIN(ir0 + dr, nr);
+    const int ir  = ir1 - ir0;
+
+    for (int i3 = 0; i3 < n_s; ++i3) {
+        for (int i2 = 0; i2 < n_t; ++i2) {
+            // {d_conv - 1 + n_t, d_inner, n_seqs}
+            // sliding window
+            const float * s = (const float *) ((const char *) src0->data + ir0*(src0->nb[1]) + i2*(src0->nb[0]) + i3*(src0->nb[2])); // {d_conv, d_inner, n_s}
+            const float * c = (const float *) ((const char *) src1->data + ir0*(src1->nb[1])); // {d_conv, d_inner}
+            float * x = (float *) ((char *) dst->data + ir0*(dst->nb[0]) + i2*(dst->nb[1]) + i3*(dst->nb[2])); // {d_inner, n_t, n_s}
+
+            // TODO: transpose the output for smaller strides for big batches?
+            // d_inner
+            for (int i1 = 0; i1 < ir; ++i1) {
+                // rowwise dot product
+                // NOTE: not using ggml_vec_dot_f32, because its sum is in double precision
+                float sumf = 0.0f;
+
+                // d_conv
+                for (int i0 = 0; i0 < nc; ++i0) {
+                    sumf += s[i0 + i1*ncs] * c[i0 + i1*nc];
+                }
+                x[i1] = sumf;
+            }
+        }
+    }
+}
+
+static void ggml_compute_forward_ssm_conv(
+        const struct ggml_compute_params * params,
+        struct ggml_tensor * dst) {
+    switch (dst->src[0]->type) {
+        case GGML_TYPE_F32:
+            {
+                ggml_compute_forward_ssm_conv_f32(params, dst);
             } break;
         default:
             {
-                GGML_ASSERT(false);
+                GGML_ABORT("fatal error");
+            }
+    }
+}
+
+// ggml_compute_forward_ssm_scan
+
+static void ggml_compute_forward_ssm_scan_f32(
+        const struct ggml_compute_params * params,
+        struct ggml_tensor * dst) {
+    const struct ggml_tensor * src0 = dst->src[0]; // s
+    const struct ggml_tensor * src1 = dst->src[1]; // x
+    const struct ggml_tensor * src2 = dst->src[2]; // dt
+    const struct ggml_tensor * src3 = dst->src[3]; // A
+    const struct ggml_tensor * src4 = dst->src[4]; // B
+    const struct ggml_tensor * src5 = dst->src[5]; // C
+
+    const int ith = params->ith;
+    const int nth = params->nth;
+
+    const int64_t nc  = src0->ne[0]; // d_state
+    const int64_t nr  = src0->ne[1]; // d_inner
+    const int64_t n_t = src1->ne[1]; // number of tokens per sequence
+    const int64_t n_s = src0->ne[2]; // number of sequences in the batch
+
+    GGML_ASSERT(ggml_nelements(src1) + ggml_nelements(src0) == ggml_nelements(dst));
+    GGML_ASSERT(src0->nb[0] == sizeof(float));
+    GGML_ASSERT(src1->nb[0] == sizeof(float));
+    GGML_ASSERT(src2->nb[0] == sizeof(float));
+    GGML_ASSERT(src3->nb[0] == sizeof(float));
+    GGML_ASSERT(src4->nb[0] == sizeof(float));
+    GGML_ASSERT(src5->nb[0] == sizeof(float));
+    // required for the dot product between s and C
+    GGML_ASSERT(src0->nb[1] == src0->ne[0]*sizeof(float));
+    // required for per-sequence offsets for states
+    GGML_ASSERT(src0->nb[2] == src0->ne[0]*src0->ne[1]*sizeof(float));
+    // required to get correct offset for state destination (i.e. src1->nb[3])
+    GGML_ASSERT(src1->nb[3] == src1->ne[0]*src1->ne[1]*src1->ne[2]*sizeof(float));
+
+    // rows per thread
+    const int dr = (nr + nth - 1)/nth;
+
+    // row range for this thread
+    const int ir0 = dr*ith;
+    const int ir1 = MIN(ir0 + dr, nr);
+    const int ir  = ir1 - ir0;
+
+    for (int i3 = 0; i3 < n_s; ++i3) {
+        for (int i2 = 0; i2 < n_t; ++i2) {
+            const float * s0 = (const float *) ((const char *) src0->data + ir0*(src0->nb[1]) + i3*(src0->nb[2])); // {d_state, d_inner, n_s}
+            const float * x  = (const float *) ((const char *) src1->data + ir0*(src1->nb[0]) + i2*(src1->nb[1]) + i3*(src1->nb[2])); // {d_inner, n_t, n_s}
+            const float * dt = (const float *) ((const char *) src2->data + ir0*(src2->nb[0]) + i2*(src2->nb[1]) + i3*(src2->nb[2])); // {d_inner, n_t, n_s}
+            const float * A  = (const float *) ((const char *) src3->data + ir0*(src3->nb[1])); // {d_state, d_inner}
+            const float * B  = (const float *) ((const char *) src4->data +  i2*(src4->nb[1]) + i3*(src4->nb[2])); // {d_state, n_t, n_s}
+            const float * C  = (const float *) ((const char *) src5->data +  i2*(src5->nb[1]) + i3*(src5->nb[2])); // {d_state, n_t, n_s}
+                  float * y  = (      float *) ((      char *) dst->data  + ir0*(src1->nb[0]) + i2*(src1->nb[1]) + i3*(src1->nb[2])); // {d_inner, n_t, n_s}
+                  float * s  = (      float *) ((      char *) dst->data  + ir0*(src0->nb[1]) + i3*(src0->nb[2]) +     src1->nb[3]);  // {d_state, d_inner, n_s}
+
+            // use the output as the source for the next token-wise iterations
+            if (i2 > 0) { s0 = s; }
+
+            // d_inner
+            for (int i1 = 0; i1 < ir; ++i1) {
+                // ref: https://github.com/state-spaces/mamba/blob/34076d664838588a3c97727b263478ab9f621a07/mamba_ssm/ops/triton/selective_state_update.py#L78
+                float dt_soft_plus = dt[i1] <= 20.0f ? log1pf(expf(dt[i1])) : dt[i1];
+                float x_dt = x[i1] * dt_soft_plus;
+                float sumf = 0.0f;
+                // d_state
+                for (int i0 = 0; i0 < nc; ++i0) {
+                    int i = i0 + i1*nc;
+                    // state = prev_state * dA + dB * x
+                    float state = (s0[i] * expf(dt_soft_plus * A[i])) + (B[i0] * x_dt);
+                    // y = rowwise_dotprod(state, C)
+                    sumf += state * C[i0];
+                    s[i] = state;
+                }
+                y[i1] = sumf;
+            }
+        }
+    }
+}
+
+static void ggml_compute_forward_ssm_scan(
+        const struct ggml_compute_params * params,
+        struct ggml_tensor * dst) {
+    switch (dst->src[0]->type) {
+        case GGML_TYPE_F32:
+            {
+                ggml_compute_forward_ssm_scan_f32(params, dst);
             } break;
+        default:
+            {
+                GGML_ABORT("fatal error");
+            }
     }
 }
 
@@ -13687,11 +16344,10 @@ static void ggml_compute_forward_flash_attn_back(
 
 static void ggml_compute_forward_win_part_f32(
         const struct ggml_compute_params * params,
-        const struct ggml_tensor * src0,
         struct ggml_tensor * dst) {
-    if (params->type == GGML_TASK_INIT || params->type == GGML_TASK_FINALIZE) {
-        return;
-    }
+    UNUSED(params);
+
+    const struct ggml_tensor * src0 = dst->src[0];
 
     GGML_TENSOR_LOCALS(int64_t, ne0, src0, ne)
     GGML_TENSOR_LOCALS(int64_t, ne,  dst,  ne)
@@ -13731,17 +16387,19 @@ static void ggml_compute_forward_win_part_f32(
 
 static void ggml_compute_forward_win_part(
         const struct ggml_compute_params * params,
-        const struct ggml_tensor * src0,
         struct ggml_tensor * dst) {
+
+    const struct ggml_tensor * src0 = dst->src[0];
+
     switch (src0->type) {
         case GGML_TYPE_F32:
             {
-                ggml_compute_forward_win_part_f32(params, src0, dst);
+                ggml_compute_forward_win_part_f32(params, dst);
             } break;
         default:
             {
-                GGML_ASSERT(false);
-            } break;
+                GGML_ABORT("fatal error");
+            }
     }
 }
 
@@ -13749,11 +16407,10 @@ static void ggml_compute_forward_win_part(
 
 static void ggml_compute_forward_win_unpart_f32(
         const struct ggml_compute_params * params,
-        const struct ggml_tensor * src0,
         struct ggml_tensor * dst) {
-    if (params->type == GGML_TASK_INIT || params->type == GGML_TASK_FINALIZE) {
-        return;
-    }
+    UNUSED(params);
+
+    const struct ggml_tensor * src0 = dst->src[0];
 
     GGML_TENSOR_LOCALS(int64_t, ne0, src0, ne)
     GGML_TENSOR_LOCALS(int64_t, ne,  dst,  ne)
@@ -13791,17 +16448,19 @@ static void ggml_compute_forward_win_unpart_f32(
 
 static void ggml_compute_forward_win_unpart(
         const struct ggml_compute_params * params,
-        const struct ggml_tensor * src0,
         struct ggml_tensor * dst) {
+
+    const struct ggml_tensor * src0 = dst->src[0];
+
     switch (src0->type) {
         case GGML_TYPE_F32:
             {
-                ggml_compute_forward_win_unpart_f32(params, src0, dst);
+                ggml_compute_forward_win_unpart_f32(params, dst);
             } break;
         default:
             {
-                GGML_ASSERT(false);
-            } break;
+                GGML_ABORT("fatal error");
+            }
     }
 }
 
@@ -13809,55 +16468,71 @@ static void ggml_compute_forward_win_unpart(
 
 static void ggml_compute_forward_unary(
         const struct ggml_compute_params * params,
-        const struct ggml_tensor * src0,
         struct ggml_tensor * dst) {
+
     const enum ggml_unary_op op = ggml_get_unary_op(dst);
 
     switch (op) {
         case GGML_UNARY_OP_ABS:
             {
-                ggml_compute_forward_abs(params, src0, dst);
+                ggml_compute_forward_abs(params, dst);
             } break;
         case GGML_UNARY_OP_SGN:
             {
-                ggml_compute_forward_sgn(params, src0, dst);
+                ggml_compute_forward_sgn(params, dst);
             } break;
         case GGML_UNARY_OP_NEG:
             {
-                ggml_compute_forward_neg(params, src0, dst);
+                ggml_compute_forward_neg(params, dst);
             } break;
         case GGML_UNARY_OP_STEP:
             {
-                ggml_compute_forward_step(params, src0, dst);
+                ggml_compute_forward_step(params, dst);
             } break;
         case GGML_UNARY_OP_TANH:
             {
-                ggml_compute_forward_tanh(params, src0, dst);
+                ggml_compute_forward_tanh(params, dst);
             } break;
         case GGML_UNARY_OP_ELU:
             {
-                ggml_compute_forward_elu(params, src0, dst);
+                ggml_compute_forward_elu(params, dst);
             } break;
         case GGML_UNARY_OP_RELU:
             {
-                ggml_compute_forward_relu(params, src0, dst);
+                ggml_compute_forward_relu(params, dst);
+            } break;
+        case GGML_UNARY_OP_SIGMOID:
+            {
+                ggml_compute_forward_sigmoid(params, dst);
             } break;
         case GGML_UNARY_OP_GELU:
             {
-                ggml_compute_forward_gelu(params, src0, dst);
+                ggml_compute_forward_gelu(params, dst);
             } break;
         case GGML_UNARY_OP_GELU_QUICK:
             {
-                ggml_compute_forward_gelu_quick(params, src0, dst);
+                ggml_compute_forward_gelu_quick(params, dst);
             } break;
         case GGML_UNARY_OP_SILU:
             {
-                ggml_compute_forward_silu(params, src0, dst);
+                ggml_compute_forward_silu(params, dst);
             } break;
-        default:
+        case GGML_UNARY_OP_HARDSWISH:
+            {
+                ggml_compute_forward_hardswish(params, dst);
+            } break;
+        case GGML_UNARY_OP_HARDSIGMOID:
+            {
+                ggml_compute_forward_hardsigmoid(params, dst);
+            } break;
+        case GGML_UNARY_OP_EXP:
             {
-                GGML_ASSERT(false);
+                ggml_compute_forward_exp(params, dst);
             } break;
+        default:
+            {
+                GGML_ABORT("fatal error");
+            }
     }
 }
 
@@ -13865,11 +16540,10 @@ static void ggml_compute_forward_unary(
 
 static void ggml_compute_forward_get_rel_pos_f16(
         const struct ggml_compute_params * params,
-        const struct ggml_tensor * src0,
         struct ggml_tensor * dst) {
-    if (params->type == GGML_TASK_INIT || params->type == GGML_TASK_FINALIZE) {
-        return;
-    }
+    UNUSED(params);
+
+    const struct ggml_tensor * src0 = dst->src[0];
 
     // ref: https://github.com/facebookresearch/segment-anything/blob/main/segment_anything/modeling/image_encoder.py#L292-L322
 
@@ -13892,17 +16566,20 @@ static void ggml_compute_forward_get_rel_pos_f16(
 
 static void ggml_compute_forward_get_rel_pos(
         const struct ggml_compute_params * params,
-        const struct ggml_tensor * src0,
         struct ggml_tensor * dst) {
+
+    const struct ggml_tensor * src0 = dst->src[0];
+
     switch (src0->type) {
         case GGML_TYPE_F16:
+        case GGML_TYPE_BF16:
             {
-                ggml_compute_forward_get_rel_pos_f16(params, src0, dst);
+                ggml_compute_forward_get_rel_pos_f16(params, dst);
             } break;
         default:
             {
-                GGML_ASSERT(false);
-            } break;
+                GGML_ABORT("fatal error");
+            }
     }
 }
 
@@ -13910,23 +16587,19 @@ static void ggml_compute_forward_get_rel_pos(
 
 static void ggml_compute_forward_add_rel_pos_f32(
         const struct ggml_compute_params * params,
-        const struct ggml_tensor * src0,
-        const struct ggml_tensor * src1,
-        const struct ggml_tensor * src2,
         struct ggml_tensor * dst) {
 
+    const struct ggml_tensor * src0 = dst->src[0];
+    const struct ggml_tensor * src1 = dst->src[1];
+    const struct ggml_tensor * src2 = dst->src[2];
+
     const bool inplace = (bool) ((int32_t *) dst->op_params)[0];
-    if (!inplace && params->type == GGML_TASK_INIT) {
-        memcpy((char *) dst->data, (char *) src0->data, ggml_nbytes(dst));
-        return;
-    }
-    if (params->type == GGML_TASK_INIT || params->type == GGML_TASK_FINALIZE) {
-        return;
+    if (!inplace) {
+        if (params->ith == 0) {
+            memcpy((char *) dst->data, (char *) src0->data, ggml_nbytes(dst));
+        }
+        ggml_barrier(params->threadpool);
     }
-
-    int64_t t0 = ggml_perf_time_us();
-    UNUSED(t0);
-
     // ref: https://github.com/facebookresearch/segment-anything/blob/main/segment_anything/modeling/image_encoder.py#L357-L359
 
     float * src1_data = (float *) src1->data;
@@ -13975,19 +16648,109 @@ static void ggml_compute_forward_add_rel_pos_f32(
 
 static void ggml_compute_forward_add_rel_pos(
         const struct ggml_compute_params * params,
-        const struct ggml_tensor * src0,
-        const struct ggml_tensor * src1,
-        const struct ggml_tensor * src2,
         struct ggml_tensor * dst) {
+
+    const struct ggml_tensor * src0 = dst->src[0];
+
     switch (src0->type) {
         case GGML_TYPE_F32:
             {
-                ggml_compute_forward_add_rel_pos_f32(params, src0, src1, src2, dst);
+                ggml_compute_forward_add_rel_pos_f32(params, dst);
             } break;
         default:
             {
-                GGML_ASSERT(false);
+                GGML_ABORT("fatal error");
+            }
+    }
+}
+
+// ggml_compute_forward_rwkv_wkv
+
+static void ggml_compute_forward_rwkv_wkv_f32(
+        const struct ggml_compute_params * params,
+        struct ggml_tensor * dst) {
+    const size_t T = dst->src[1]->ne[3];
+    const size_t C = dst->ne[0];
+    const size_t H = dst->src[1]->ne[2];
+    const size_t n_seqs = dst->src[5]->ne[1];
+
+    float * dst_data = (float *) dst->data;
+    float * state = ((float *) dst->data) + C * T;
+
+    if (params->ith != 0) {
+        return;
+    }
+
+    memset(dst_data, 0, T * C * sizeof(float));
+
+    float * k =          (float *) dst->src[0]->data;
+    float * v =          (float *) dst->src[1]->data;
+    float * r =          (float *) dst->src[2]->data;
+    float * time_faaaa = (float *) dst->src[3]->data;
+    float * time_decay = (float *) dst->src[4]->data;
+
+    size_t t_stride = H * (C / H);
+
+    size_t h_stride = C / H;
+    size_t h_stride_2d = (C / H) * (C / H);
+
+    // basically fused operations:
+    // dst = r @ (time_faaaa * (k @ v) + state),
+    // state = time_decay * state + (k @ v),
+    // recursive through each token
+    for (size_t t = 0; t < T; t++) {
+        size_t t_offset = t * t_stride;
+        size_t state_offset = (C / H) * C * (t / (T / n_seqs));
+        float * state_cur = state + state_offset;
+        float * state_prev = t % (T / n_seqs) ? state_cur : (float*)dst->src[5]->data + state_offset;
+
+        for (size_t h = 0; h < H; h++) {
+            size_t h_offset = h * h_stride;
+            size_t t_h_offset = t_offset + h_offset;
+            size_t h_2d_offset = h * h_stride_2d;
+
+            for (size_t i = 0; i < C / H; i++) {
+                size_t t_h_i_offset = t_h_offset + i;
+                size_t h_i_offset = h_offset + i;
+                size_t h_2d_i_offset = h_2d_offset + i * h_stride;
+
+                float k_val = k[t_h_i_offset];
+                float r_val = r[t_h_i_offset];
+                float time_faaaa_val = time_faaaa[h_i_offset];
+                // RWKV v6: different time_decay for each token.
+                float time_decay_val = time_decay[t_h_i_offset];
+
+                for (size_t j = 0; j < C / H; j ++) {
+                    size_t t_h_j_offset = t_h_offset + j;
+                    size_t h_2d_i_j_offset = h_2d_i_offset + j;
+
+                    float v_val = v[t_h_j_offset];
+                    float kv_val = v_val * k_val;
+                    float prev_state_val = state_prev[h_2d_i_j_offset];
+                    float temp_val = kv_val * time_faaaa_val + prev_state_val;
+                    dst_data[t_h_j_offset] += temp_val * r_val;
+                    state_cur[h_2d_i_j_offset] = prev_state_val * time_decay_val + kv_val;
+                }
+            }
+        }
+    }
+}
+
+static void ggml_compute_forward_rwkv_wkv(
+        const struct ggml_compute_params * params,
+        struct ggml_tensor * dst) {
+
+    const struct ggml_tensor * src0 = dst->src[0];
+
+    switch (src0->type) {
+        case GGML_TYPE_F32:
+            {
+                ggml_compute_forward_rwkv_wkv_f32(params, dst);
             } break;
+        default:
+            {
+                GGML_ABORT("fatal error");
+            }
     }
 }
 
@@ -13995,21 +16758,22 @@ static void ggml_compute_forward_add_rel_pos(
 
 static void ggml_compute_forward_map_unary_f32(
         const struct ggml_compute_params * params,
-        const struct ggml_tensor * src0,
         struct ggml_tensor * dst,
         const ggml_unary_op_f32_t fun) {
-    GGML_ASSERT(ggml_are_same_shape(src0, dst));
 
-    if (params->type == GGML_TASK_INIT || params->type == GGML_TASK_FINALIZE) {
+    const struct ggml_tensor * src0 = dst->src[0];
+
+    if (params->ith != 0) {
         return;
     }
 
+    assert(ggml_is_contiguous_1(src0));
+    assert(ggml_is_contiguous_1(dst));
+    assert(ggml_are_same_shape(src0, dst));
+
     const int n  = ggml_nrows(src0);
     const int nc = src0->ne[0];
 
-    assert( dst->nb[0] == sizeof(float));
-    assert(src0->nb[0] == sizeof(float));
-
     for (int i = 0; i < n; i++) {
         fun(nc,
                 (float *) ((char *) dst->data  + i*( dst->nb[1])),
@@ -14019,18 +16783,20 @@ static void ggml_compute_forward_map_unary_f32(
 
 static void ggml_compute_forward_map_unary(
         const struct ggml_compute_params * params,
-        const struct ggml_tensor * src0,
         struct ggml_tensor * dst,
         const ggml_unary_op_f32_t fun) {
+
+    const struct ggml_tensor * src0 = dst->src[0];
+
     switch (src0->type) {
         case GGML_TYPE_F32:
             {
-                ggml_compute_forward_map_unary_f32(params, src0, dst, fun);
+                ggml_compute_forward_map_unary_f32(params, dst, fun);
             } break;
         default:
             {
-                GGML_ASSERT(false);
-            } break;
+                GGML_ABORT("fatal error");
+            }
     }
 }
 
@@ -14038,24 +16804,24 @@ static void ggml_compute_forward_map_unary(
 
 static void ggml_compute_forward_map_binary_f32(
         const struct ggml_compute_params * params,
-        const struct ggml_tensor * src0,
-        const struct ggml_tensor * src1,
         struct ggml_tensor * dst,
         const ggml_binary_op_f32_t fun) {
-    assert(params->ith == 0);
-    assert(ggml_are_same_shape(src0, src1) && ggml_are_same_shape(src0, dst));
 
-    if (params->type == GGML_TASK_INIT || params->type == GGML_TASK_FINALIZE) {
+    const struct ggml_tensor * src0 = dst->src[0];
+    const struct ggml_tensor * src1 = dst->src[1];
+
+    if (params->ith != 0) {
         return;
     }
 
+    assert(ggml_is_contiguous_1(src0));
+    assert(ggml_is_contiguous_1(src1));
+    assert(ggml_is_contiguous_1(dst));
+    assert(ggml_are_same_shape(src0, src1) && ggml_are_same_shape(src0, dst));
+
     const int n  = ggml_nrows(src0);
     const int nc = src0->ne[0];
 
-    assert( dst->nb[0] == sizeof(float));
-    assert(src0->nb[0] == sizeof(float));
-    assert(src1->nb[0] == sizeof(float));
-
     for (int i = 0; i < n; i++) {
         fun(nc,
                 (float *) ((char *) dst->data  + i*( dst->nb[1])),
@@ -14066,19 +16832,20 @@ static void ggml_compute_forward_map_binary_f32(
 
 static void ggml_compute_forward_map_binary(
         const struct ggml_compute_params * params,
-        const struct ggml_tensor * src0,
-        const struct ggml_tensor * src1,
         struct ggml_tensor * dst,
         const ggml_binary_op_f32_t fun) {
+
+    const struct ggml_tensor * src0 = dst->src[0];
+
     switch (src0->type) {
         case GGML_TYPE_F32:
             {
-                ggml_compute_forward_map_binary_f32(params, src0, src1, dst, fun);
+                ggml_compute_forward_map_binary_f32(params, dst, fun);
             } break;
         default:
             {
-                GGML_ASSERT(false);
-            } break;
+                GGML_ABORT("fatal error");
+            }
     }
 }
 
@@ -14086,12 +16853,12 @@ static void ggml_compute_forward_map_binary(
 
 static void ggml_compute_forward_map_custom1_f32(
         const struct ggml_compute_params * params,
-        const struct ggml_tensor * a,
         struct ggml_tensor * dst,
         const ggml_custom1_op_f32_t fun) {
-    assert(params->ith == 0);
 
-    if (params->type == GGML_TASK_INIT || params->type == GGML_TASK_FINALIZE) {
+    const struct ggml_tensor * a = dst->src[0];
+
+    if (params->ith != 0) {
         return;
     }
 
@@ -14102,13 +16869,13 @@ static void ggml_compute_forward_map_custom1_f32(
 
 static void ggml_compute_forward_map_custom2_f32(
         const struct ggml_compute_params * params,
-        const struct ggml_tensor * a,
-        const struct ggml_tensor * b,
         struct ggml_tensor * dst,
         const ggml_custom2_op_f32_t fun) {
-    assert(params->ith == 0);
 
-    if (params->type == GGML_TASK_INIT || params->type == GGML_TASK_FINALIZE) {
+    const struct ggml_tensor * a = dst->src[0];
+    const struct ggml_tensor * b = dst->src[1];
+
+    if (params->ith != 0) {
         return;
     }
 
@@ -14119,14 +16886,14 @@ static void ggml_compute_forward_map_custom2_f32(
 
 static void ggml_compute_forward_map_custom3_f32(
         const struct ggml_compute_params * params,
-        const struct ggml_tensor * a,
-        const struct ggml_tensor * b,
-        const struct ggml_tensor * c,
         struct ggml_tensor * dst,
         const ggml_custom3_op_f32_t fun) {
-    assert(params->ith == 0);
 
-    if (params->type == GGML_TASK_INIT || params->type == GGML_TASK_FINALIZE) {
+    const struct ggml_tensor * a = dst->src[0];
+    const struct ggml_tensor * b = dst->src[1];
+    const struct ggml_tensor * c = dst->src[1];
+
+    if (params->ith != 0) {
         return;
     }
 
@@ -14137,172 +16904,140 @@ static void ggml_compute_forward_map_custom3_f32(
 
 static void ggml_compute_forward_map_custom1(
         const struct ggml_compute_params * params,
-        const struct ggml_tensor * a,
               struct ggml_tensor * dst) {
-    if (params->type == GGML_TASK_INIT || params->type == GGML_TASK_FINALIZE) {
-        return;
-    }
 
-    struct ggml_map_custom1_op_params * p = (struct ggml_map_custom1_op_params *) dst->op_params;
+    const struct ggml_tensor * a = dst->src[0];
 
-    p->fun(dst, a, params->ith, params->nth, p->userdata);
+    struct ggml_map_custom1_op_params p;
+    memcpy(&p, dst->op_params, sizeof(p));
+
+    p.fun(dst, a, params->ith, params->nth, p.userdata);
 }
 
 // ggml_compute_forward_map_custom2
 
 static void ggml_compute_forward_map_custom2(
         const struct ggml_compute_params * params,
-        const struct ggml_tensor * a,
-        const struct ggml_tensor * b,
               struct ggml_tensor * dst) {
-    if (params->type == GGML_TASK_INIT || params->type == GGML_TASK_FINALIZE) {
-        return;
-    }
 
-    struct ggml_map_custom2_op_params * p = (struct ggml_map_custom2_op_params *) dst->op_params;
+    const struct ggml_tensor * a = dst->src[0];
+    const struct ggml_tensor * b = dst->src[1];
+
+    struct ggml_map_custom2_op_params p;
+    memcpy(&p, dst->op_params, sizeof(p));
 
-    p->fun(dst, a, b, params->ith, params->nth, p->userdata);
+    p.fun(dst, a, b, params->ith, params->nth, p.userdata);
 }
 
 // ggml_compute_forward_map_custom3
 
 static void ggml_compute_forward_map_custom3(
         const struct ggml_compute_params * params,
-        const struct ggml_tensor * a,
-        const struct ggml_tensor * b,
-        const struct ggml_tensor * c,
               struct ggml_tensor * dst) {
-    if (params->type == GGML_TASK_INIT || params->type == GGML_TASK_FINALIZE) {
-        return;
-    }
 
-    struct ggml_map_custom3_op_params * p = (struct ggml_map_custom3_op_params *) dst->op_params;
+    const struct ggml_tensor * a = dst->src[0];
+    const struct ggml_tensor * b = dst->src[1];
+    const struct ggml_tensor * c = dst->src[2];
 
-    p->fun(dst, a, b, c, params->ith, params->nth, p->userdata);
+    struct ggml_map_custom3_op_params p;
+    memcpy(&p, dst->op_params, sizeof(p));
+
+    p.fun(dst, a, b, c, params->ith, params->nth, p.userdata);
 }
 
 // ggml_compute_forward_cross_entropy_loss
 
 static void ggml_compute_forward_cross_entropy_loss_f32(
         const struct ggml_compute_params * params,
-        const struct ggml_tensor * src0,
-        const struct ggml_tensor * src1,
         struct ggml_tensor * dst) {
-    GGML_ASSERT(ggml_is_contiguous(src0));
-    GGML_ASSERT(ggml_is_contiguous(src1));
-    GGML_ASSERT(ggml_is_scalar(dst));
-    GGML_ASSERT(ggml_are_same_shape(src0, src1));
 
-    const int ith = params->ith;
-    const int nth = params->nth;
+    const struct ggml_tensor * src0 = dst->src[0];
+    const struct ggml_tensor * src1 = dst->src[1];
 
-    float * sums = (float *) params->wdata;
+    GGML_ASSERT(src0->type == GGML_TYPE_F32);
+    GGML_ASSERT(src1->type == GGML_TYPE_F32);
+    GGML_ASSERT(src0->nb[0] == ggml_type_size(src0->type));
+    GGML_ASSERT(src1->nb[0] == ggml_type_size(src1->type));
+    GGML_ASSERT(ggml_are_same_shape(src0, src1));
+    GGML_ASSERT(ggml_is_scalar(dst));
+    GGML_ASSERT(dst->type == GGML_TYPE_F32);
 
     // TODO: handle transposed/permuted matrices
-    const int nc = src0->ne[0];
-    const int nr = ggml_nrows(src0);
-
-    GGML_ASSERT(params->wsize >= sizeof(float) * (nth + nth * nc));
+    const int64_t nc = src0->ne[0];
+    const int64_t nr = ggml_nrows(src0);
 
-    if (params->type == GGML_TASK_INIT) {
-        if (ith == 0) {
-            memset(sums, 0, sizeof(float) * (nth + nth * nc));
-        }
-        return;
-    }
+    const int ith = params->ith;
+    const int nth = params->nth;
 
-    if (params->type == GGML_TASK_FINALIZE) {
-        if (ith == 0) {
-            float * dp = (float *) dst->data;
-            ggml_vec_sum_f32(nth, dp, sums);
-            dp[0] *= -1.0f / (float) nr;
-        }
-        return;
-    }
+    float * sums =  (float *) params->wdata;
+    float * st   = ((float *) params->wdata) + nth + ith*nc;
+    float sum_thread = 0.0f;
 
-    const double eps = 1e-9;
+    GGML_ASSERT(params->wsize >= sizeof(float) * (nth + nth * nc));
 
     // rows per thread
-    const int dr = (nr + nth - 1)/nth;
+    const int64_t dr = (nr + nth - 1)/nth;
 
     // row range for this thread
-    const int ir0 = dr*ith;
-    const int ir1 = MIN(ir0 + dr, nr);
+    const int64_t ir0 = dr*ith;
+    const int64_t ir1 = MIN(ir0 + dr, nr);
 
-    for (int i1 = ir0; i1 < ir1; i1++) {
-        float * s0 = (float *)((char *) src0->data + i1*src0->nb[1]);
-        float * s1 = (float *)((char *) src1->data + i1*src1->nb[1]);
-        float * st = ((float *) params->wdata) + nth + ith*nc;
+    for (int64_t i1 = ir0; i1 < ir1; ++i1) {
+        const float * s0 = (const float *)((const char *) src0->data + i1*src0->nb[1]);
+        const float * s1 = (const float *)((const char *) src1->data + i1*src1->nb[1]);
 
 #ifndef NDEBUG
-        for (int i = 0; i < nc; ++i) {
+        for (int64_t i = 0; i < nc; ++i) {
             //printf("p[%d] = %f\n", i, p[i]);
             assert(!isnan(s0[i]));
             assert(!isnan(s1[i]));
         }
 #endif
-        // soft_max
-        ggml_float sum = 0.0;
-        {
-            float max = -INFINITY;
-            ggml_vec_max_f32(nc, &max, s0);
 
-            uint16_t scvt; UNUSED(scvt);
-            for (int i = 0; i < nc; i++) {
-                if (s0[i] == -INFINITY) {
-                    st[i] = 0.0f;
-                } else {
-#ifndef GGML_CROSS_ENTROPY_EXP_FP16
-                    const float s = s0[i] - max;
-                    const float val = expf(s);
-#else
-                    ggml_fp16_t s = GGML_FP32_TO_FP16(s0[i] - max);
-                    memcpy(&scvt, &s, sizeof(scvt));
-                    const float val = GGML_FP16_TO_FP32(ggml_table_exp_f16[scvt]);
-#endif
-                    sum += (ggml_float)val;
-                    st[i] = val;
-                }
-            }
+        float max = -INFINITY;
+        ggml_vec_max_f32(nc, &max, s0);
+        const ggml_float sum_softmax = ggml_vec_log_soft_max_f32(nc, st, s0, max);
+        assert(sum_softmax >= 0.0);
 
-            assert(sum > 0.0);
-            // sum = 1.0/sum;
-        }
-        // avoid log(0) by rescaling from [0..1] to [eps..1]
-        sum = (1.0 - eps) / sum;
-        ggml_vec_scale_f32(nc, st, sum);
-        ggml_vec_add1_f32(nc, st, st, eps);
-        ggml_vec_log_f32(nc, st, st);
+        ggml_vec_add1_f32(nc, st, st, -sum_softmax);
         ggml_vec_mul_f32(nc, st, st, s1);
 
-        float st_sum = 0;
-        ggml_vec_sum_f32(nc, &st_sum, st);
-        sums[ith] += st_sum;
+        float sum_st = 0.0f;
+        ggml_vec_sum_f32(nc, &sum_st, st);
+        sum_thread += sum_st;
 
 #ifndef NDEBUG
-        for (int i = 0; i < nc; ++i) {
+        for (int64_t i = 0; i < nc; ++i) {
             assert(!isnan(st[i]));
             assert(!isinf(st[i]));
         }
 #endif
     }
+    sums[ith] = sum_thread;
+    ggml_barrier(params->threadpool);
 
+    if (ith == 0) {
+        float * dp = (float *) dst->data;
+        ggml_vec_sum_f32(nth, dp, sums);
+        dp[0] *= -1.0f / (float) nr;
+    }
 }
 
 static void ggml_compute_forward_cross_entropy_loss(
         const struct ggml_compute_params * params,
-        const struct ggml_tensor * src0,
-        const struct ggml_tensor * src1,
         struct ggml_tensor * dst) {
+
+    const struct ggml_tensor * src0 = dst->src[0];
+
     switch (src0->type) {
         case GGML_TYPE_F32:
             {
-                ggml_compute_forward_cross_entropy_loss_f32(params, src0, src1, dst);
+                ggml_compute_forward_cross_entropy_loss_f32(params, dst);
             } break;
         default:
             {
-                GGML_ASSERT(false);
-            } break;
+                GGML_ABORT("fatal error");
+            }
     }
 }
 
@@ -14310,10 +17045,12 @@ static void ggml_compute_forward_cross_entropy_loss(
 
 static void ggml_compute_forward_cross_entropy_loss_back_f32(
         const struct ggml_compute_params * params,
-        const struct ggml_tensor * src0,
-        const struct ggml_tensor * src1,
-        const struct ggml_tensor * opt0,
         struct ggml_tensor * dst) {
+
+    const struct ggml_tensor * src0 = dst->src[0];
+    const struct ggml_tensor * src1 = dst->src[1];
+    const struct ggml_tensor * opt0 = dst->src[2];
+
     GGML_ASSERT(ggml_is_contiguous(dst));
     GGML_ASSERT(ggml_is_contiguous(src0));
     GGML_ASSERT(ggml_is_contiguous(src1));
@@ -14323,12 +17060,6 @@ static void ggml_compute_forward_cross_entropy_loss_back_f32(
     const int64_t ith = params->ith;
     const int64_t nth = params->nth;
 
-    if (params->type == GGML_TASK_INIT || params->type == GGML_TASK_FINALIZE) {
-        return;
-    }
-
-    const double eps = 1e-9;
-
     // TODO: handle transposed/permuted matrices
     const int64_t nc = src0->ne[0];
     const int64_t nr = ggml_nrows(src0);
@@ -14340,7 +17071,7 @@ static void ggml_compute_forward_cross_entropy_loss_back_f32(
     const int64_t ir0 = dr*ith;
     const int64_t ir1 = MIN(ir0 + dr, nr);
 
-    float * d   = (float *) opt0->data;
+    const float d_by_nr = ((const float *) opt0->data)[0] / (float) nr;
 
     for (int64_t i1 = ir0; i1 < ir1; i1++) {
         float * ds0 = (float *)((char *) dst->data  + i1*dst->nb[1]);
@@ -14348,7 +17079,7 @@ static void ggml_compute_forward_cross_entropy_loss_back_f32(
         float * s1  = (float *)((char *) src1->data + i1*src1->nb[1]);
 
 #ifndef NDEBUG
-        for (int i = 0; i < nc; ++i) {
+        for (int64_t i = 0; i < nc; ++i) {
             //printf("p[%d] = %f\n", i, p[i]);
             assert(!isnan(s0[i]));
             assert(!isnan(s1[i]));
@@ -14356,41 +17087,18 @@ static void ggml_compute_forward_cross_entropy_loss_back_f32(
 #endif
 
         // soft_max
-        ggml_float sum = 0.0;
-        {
-            float max = -INFINITY;
-            ggml_vec_max_f32(nc, &max, s0);
-
-            uint16_t scvt; UNUSED(scvt);
-            for (int i = 0; i < nc; i++) {
-                if (s0[i] == -INFINITY) {
-                    ds0[i] = 0.0f;
-                } else {
-#ifndef GGML_CROSS_ENTROPY_EXP_FP16
-                    const float s = s0[i] - max;
-                    const float val = expf(s);
-#else
-                    ggml_fp16_t s = GGML_FP32_TO_FP16(s0[i] - max);
-                    memcpy(&scvt, &s, sizeof(scvt));
-                    const float val = GGML_FP16_TO_FP32(ggml_table_exp_f16[scvt]);
-#endif
-                    sum += (ggml_float)val;
-                    ds0[i] = val;
-                }
-            }
-
-            assert(sum > 0.0);
-            sum = (1.0 - eps)/sum;
-        }
+        float max = -INFINITY;
+        ggml_vec_max_f32(nc, &max, s0);
+        ggml_float sum = ggml_vec_soft_max_f32(nc, ds0, s0, max);
+        assert(sum > 0.0);
+        ggml_vec_scale_f32(nc, ds0, 1.0/sum);
 
         // grad(src0) = (softmax(src0) - src1) * grad(cross_entropy_loss(src0, src1)) / nr
-        ggml_vec_scale_f32(nc, ds0, sum);
-        ggml_vec_add1_f32(nc, ds0, ds0, eps);
         ggml_vec_sub_f32(nc, ds0, ds0, s1);
-        ggml_vec_scale_f32(nc, ds0, d[0] / (float) nr);
+        ggml_vec_scale_f32(nc, ds0, d_by_nr);
 
 #ifndef NDEBUG
-        for (int i = 0; i < nc; ++i) {
+        for (int64_t i = 0; i < nc; ++i) {
             assert(!isnan(ds0[i]));
             assert(!isinf(ds0[i]));
         }
@@ -14400,349 +17108,462 @@ static void ggml_compute_forward_cross_entropy_loss_back_f32(
 
 static void ggml_compute_forward_cross_entropy_loss_back(
         const struct ggml_compute_params * params,
-        const struct ggml_tensor * src0,
-        const struct ggml_tensor * src1,
-        const struct ggml_tensor * opt0,
         struct ggml_tensor * dst) {
+
+    const struct ggml_tensor * src0 = dst->src[0];
+
     switch (src0->type) {
         case GGML_TYPE_F32:
             {
-                ggml_compute_forward_cross_entropy_loss_back_f32(params, src0, src1, opt0, dst);
+                ggml_compute_forward_cross_entropy_loss_back_f32(params, dst);
             } break;
         default:
             {
-                GGML_ASSERT(false);
-            } break;
+                GGML_ABORT("fatal error");
+            }
     }
 }
 
-/////////////////////////////////
+static void ggml_compute_forward_opt_step_adamw_f32(
+        const struct ggml_compute_params * params,
+        struct ggml_tensor * dst) {
 
-static void ggml_compute_forward(struct ggml_compute_params * params, struct ggml_tensor * tensor) {
-    GGML_ASSERT(params);
+    const struct ggml_tensor * src0        = dst->src[0];
+    const struct ggml_tensor * src0_grad   = dst->src[1];
+    const struct ggml_tensor * src0_grad_m = dst->src[2];
+    const struct ggml_tensor * src0_grad_v = dst->src[3];
+    GGML_ASSERT(ggml_are_same_shape(src0, src0_grad));
+
+    const int ith = params->ith;
+    const int nth = params->nth;
+
+    const int nr  = ggml_nrows(src0);
+
+    GGML_TENSOR_UNARY_OP_LOCALS
+    GGML_ASSERT(nb00 == sizeof(float));
+
+    // rows per thread
+    const int dr = (nr + nth - 1)/nth;
+
+    // row range for this thread
+    const int ir0 = dr*ith;
+    const int ir1 = MIN(ir0 + dr, nr);
+
+    /* const float   gnorm = 1.0f; */
+    int64_t       iter;   memcpy(&iter, &dst->op_params[0], sizeof(int64_t));
+    const float   alpha = ggml_get_op_params_f32(dst, 2);
+    const float   beta1 = ggml_get_op_params_f32(dst, 3);
+    const float   beta2 = ggml_get_op_params_f32(dst, 4);
+    const float   eps   = ggml_get_op_params_f32(dst, 5);
+    const float   wd    = ggml_get_op_params_f32(dst, 6);
+
+    const float beta1h  = alpha/(1.0f - powf(beta1, iter));
+    const float beta2h  =  1.0f/(1.0f - powf(beta2, iter));
+
+    for (int ir = ir0; ir < ir1; ++ir) {
+        const int64_t i03 = ir/(ne02*ne01);
+        const int64_t i02 = (ir - i03*ne02*ne01)/ne01;
+        const int64_t i01 = (ir - i03*ne02*ne01 - i02*ne01);
 
-    if (tensor->op == GGML_OP_NONE) {
+        const size_t offset = i03*nb03 + i02*nb02 + i01*nb01;
+
+        float       * w = (float       *) ((char       *) src0->data        + offset); // weight
+        const float * g = (const float *) ((const char *) src0_grad->data   + offset); // grad
+        float       * m = (float       *) ((char       *) src0_grad_m->data + offset);
+        float       * v = (float       *) ((char       *) src0_grad_v->data + offset);
+
+        for (int i00 = 0; i00 < ne00; ++i00) {
+            m[i00] = m[i00]*beta1 +        g[i00]*(1.0f - beta1);
+            v[i00] = v[i00]*beta2 + g[i00]*g[i00]*(1.0f - beta2);
+
+            const float mh =       m[i00]*beta1h;
+            const float vh = sqrtf(v[i00]*beta2h) + eps;
+
+            // The weight decay is applied independently of the Adam momenta m and v.
+            // This is NOT equivalent to l2 regularization that adds w[i00]*w[i00] to the loss.
+            // See: https://arxiv.org/pdf/1711.05101v3.pdf
+            w[i00] = w[i00]*(1.0f - alpha*wd) - mh/vh;
+        }
+    }
+
+    ggml_barrier(params->threadpool);
+    if (ith != 0) {
         return;
     }
 
-#ifdef GGML_USE_CUBLAS
-    bool skip_cpu = ggml_cuda_compute_forward(params, tensor);
-    if (skip_cpu) {
+    iter++;
+    memcpy(&dst->op_params[0], &iter, sizeof(int64_t));
+}
+
+static void ggml_compute_forward_opt_step_adamw(
+        const struct ggml_compute_params * params,
+        struct ggml_tensor * dst) {
+
+    const struct ggml_tensor * src0 = dst->src[0];
+
+    switch (src0->type) {
+        case GGML_TYPE_F32:
+            {
+                ggml_compute_forward_opt_step_adamw_f32(params, dst);
+            } break;
+        default:
+            {
+                GGML_ABORT("fatal error");
+            }
+    }
+}
+/////////////////////////////////
+
+static void ggml_compute_forward(struct ggml_compute_params * params, struct ggml_tensor * tensor) {
+    GGML_ASSERT(params);
+
+    if (tensor->op == GGML_OP_NONE || ggml_is_empty(tensor)) {
         return;
     }
-    GGML_ASSERT(tensor->src[0] == NULL || tensor->src[0]->backend == GGML_BACKEND_CPU);
-    GGML_ASSERT(tensor->src[1] == NULL || tensor->src[1]->backend == GGML_BACKEND_CPU);
-#endif // GGML_USE_CUBLAS
 
     switch (tensor->op) {
         case GGML_OP_DUP:
             {
-                ggml_compute_forward_dup(params, tensor->src[0], tensor);
+                ggml_compute_forward_dup(params, tensor);
             } break;
         case GGML_OP_ADD:
             {
-                ggml_compute_forward_add(params, tensor->src[0], tensor->src[1], tensor);
+                ggml_compute_forward_add(params, tensor);
             } break;
         case GGML_OP_ADD1:
             {
-                ggml_compute_forward_add1(params, tensor->src[0], tensor->src[1], tensor);
+                ggml_compute_forward_add1(params, tensor);
             } break;
         case GGML_OP_ACC:
             {
-                ggml_compute_forward_acc(params, tensor->src[0], tensor->src[1], tensor);
+                ggml_compute_forward_acc(params, tensor);
             } break;
         case GGML_OP_SUB:
             {
-                ggml_compute_forward_sub(params, tensor->src[0], tensor->src[1], tensor);
+                ggml_compute_forward_sub(params, tensor);
             } break;
         case GGML_OP_MUL:
             {
-                ggml_compute_forward_mul(params, tensor->src[0], tensor->src[1], tensor);
+                ggml_compute_forward_mul(params, tensor);
             } break;
         case GGML_OP_DIV:
             {
-                ggml_compute_forward_div(params, tensor->src[0], tensor->src[1], tensor);
+                ggml_compute_forward_div(params, tensor);
             } break;
         case GGML_OP_SQR:
             {
-                ggml_compute_forward_sqr(params, tensor->src[0], tensor);
+                ggml_compute_forward_sqr(params, tensor);
             } break;
         case GGML_OP_SQRT:
             {
-                ggml_compute_forward_sqrt(params, tensor->src[0], tensor);
+                ggml_compute_forward_sqrt(params, tensor);
             } break;
         case GGML_OP_LOG:
             {
-                ggml_compute_forward_log(params, tensor->src[0], tensor);
+                ggml_compute_forward_log(params, tensor);
+            } break;
+        case GGML_OP_SIN:
+            {
+                ggml_compute_forward_sin(params, tensor);
+            } break;
+        case GGML_OP_COS:
+            {
+                ggml_compute_forward_cos(params, tensor);
             } break;
         case GGML_OP_SUM:
             {
-                ggml_compute_forward_sum(params, tensor->src[0], tensor);
+                ggml_compute_forward_sum(params, tensor);
             } break;
         case GGML_OP_SUM_ROWS:
             {
-                ggml_compute_forward_sum_rows(params, tensor->src[0], tensor);
+                ggml_compute_forward_sum_rows(params, tensor);
             } break;
         case GGML_OP_MEAN:
             {
-                ggml_compute_forward_mean(params, tensor->src[0], tensor);
+                ggml_compute_forward_mean(params, tensor);
             } break;
         case GGML_OP_ARGMAX:
             {
-                ggml_compute_forward_argmax(params, tensor->src[0], tensor);
+                ggml_compute_forward_argmax(params, tensor);
+            } break;
+        case GGML_OP_COUNT_EQUAL:
+            {
+                ggml_compute_forward_count_equal(params, tensor);
             } break;
         case GGML_OP_REPEAT:
             {
-                ggml_compute_forward_repeat(params, tensor->src[0], tensor);
+                ggml_compute_forward_repeat(params, tensor);
             } break;
         case GGML_OP_REPEAT_BACK:
             {
-                ggml_compute_forward_repeat_back(params, tensor->src[0], tensor);
+                ggml_compute_forward_repeat_back(params, tensor);
             } break;
         case GGML_OP_CONCAT:
             {
-                ggml_compute_forward_concat(params, tensor->src[0], tensor->src[1], tensor);
+                ggml_compute_forward_concat(params, tensor);
             } break;
         case GGML_OP_SILU_BACK:
             {
-                ggml_compute_forward_silu_back(params, tensor->src[0], tensor->src[1], tensor);
+                ggml_compute_forward_silu_back(params, tensor);
             } break;
         case GGML_OP_NORM:
             {
-                ggml_compute_forward_norm(params, tensor->src[0], tensor);
+                ggml_compute_forward_norm(params, tensor);
             } break;
         case GGML_OP_RMS_NORM:
             {
-                ggml_compute_forward_rms_norm(params, tensor->src[0], tensor);
+                ggml_compute_forward_rms_norm(params, tensor);
             } break;
         case GGML_OP_RMS_NORM_BACK:
             {
-                ggml_compute_forward_rms_norm_back(params, tensor->src[0], tensor->src[1], tensor);
+                ggml_compute_forward_rms_norm_back(params, tensor);
             } break;
         case GGML_OP_GROUP_NORM:
             {
-                ggml_compute_forward_group_norm(params, tensor->src[0], tensor);
+                ggml_compute_forward_group_norm(params, tensor);
             } break;
         case GGML_OP_MUL_MAT:
             {
-                ggml_compute_forward_mul_mat(params, tensor->src[0], tensor->src[1], tensor);
+                ggml_compute_forward_mul_mat(params, tensor);
             } break;
         case GGML_OP_MUL_MAT_ID:
             {
-                ggml_compute_forward_mul_mat_id(params, tensor->src[0], tensor->src[1], tensor);
+                ggml_compute_forward_mul_mat_id(params, tensor);
             } break;
         case GGML_OP_OUT_PROD:
             {
-                ggml_compute_forward_out_prod(params, tensor->src[0], tensor->src[1], tensor);
+                ggml_compute_forward_out_prod(params, tensor);
             } break;
         case GGML_OP_SCALE:
             {
-                ggml_compute_forward_scale(params, tensor->src[0], tensor);
+                ggml_compute_forward_scale(params, tensor);
             } break;
         case GGML_OP_SET:
             {
-                ggml_compute_forward_set(params, tensor->src[0], tensor->src[1], tensor);
+                ggml_compute_forward_set(params, tensor);
             } break;
         case GGML_OP_CPY:
             {
-                ggml_compute_forward_cpy(params, tensor->src[0], tensor);
+                ggml_compute_forward_cpy(params, tensor);
             } break;
         case GGML_OP_CONT:
             {
-                ggml_compute_forward_cont(params, tensor->src[0], tensor);
+                ggml_compute_forward_cont(params, tensor);
             } break;
         case GGML_OP_RESHAPE:
             {
-                ggml_compute_forward_reshape(params, tensor->src[0], tensor);
+                ggml_compute_forward_reshape(params, tensor);
             } break;
         case GGML_OP_VIEW:
             {
-                ggml_compute_forward_view(params, tensor->src[0]);
+                ggml_compute_forward_view(params, tensor);
             } break;
         case GGML_OP_PERMUTE:
             {
-                ggml_compute_forward_permute(params, tensor->src[0]);
+                ggml_compute_forward_permute(params, tensor);
             } break;
         case GGML_OP_TRANSPOSE:
             {
-                ggml_compute_forward_transpose(params, tensor->src[0]);
+                ggml_compute_forward_transpose(params, tensor);
             } break;
         case GGML_OP_GET_ROWS:
             {
-                ggml_compute_forward_get_rows(params, tensor->src[0], tensor->src[1], tensor);
+                ggml_compute_forward_get_rows(params, tensor);
             } break;
         case GGML_OP_GET_ROWS_BACK:
             {
-                ggml_compute_forward_get_rows_back(params, tensor->src[0], tensor->src[1], tensor);
+                ggml_compute_forward_get_rows_back(params, tensor);
             } break;
         case GGML_OP_DIAG:
             {
-                ggml_compute_forward_diag(params, tensor->src[0], tensor);
+                ggml_compute_forward_diag(params, tensor);
             } break;
         case GGML_OP_DIAG_MASK_INF:
             {
-                ggml_compute_forward_diag_mask_inf(params, tensor->src[0], tensor);
+                ggml_compute_forward_diag_mask_inf(params, tensor);
             } break;
         case GGML_OP_DIAG_MASK_ZERO:
             {
-                ggml_compute_forward_diag_mask_zero(params, tensor->src[0], tensor);
+                ggml_compute_forward_diag_mask_zero(params, tensor);
             } break;
         case GGML_OP_SOFT_MAX:
             {
-                ggml_compute_forward_soft_max(params, tensor->src[0], tensor->src[1], tensor);
+                ggml_compute_forward_soft_max(params, tensor);
             } break;
         case GGML_OP_SOFT_MAX_BACK:
             {
-                ggml_compute_forward_soft_max_back(params, tensor->src[0], tensor->src[1], tensor);
+                ggml_compute_forward_soft_max_back(params, tensor);
             } break;
         case GGML_OP_ROPE:
             {
-                ggml_compute_forward_rope(params, tensor->src[0], tensor->src[1], tensor);
+                ggml_compute_forward_rope(params, tensor);
             } break;
         case GGML_OP_ROPE_BACK:
             {
-                ggml_compute_forward_rope_back(params, tensor->src[0], tensor->src[1], tensor);
-            } break;
-        case GGML_OP_ALIBI:
-            {
-                ggml_compute_forward_alibi(params, tensor->src[0], tensor);
+                ggml_compute_forward_rope_back(params, tensor);
             } break;
         case GGML_OP_CLAMP:
             {
-                ggml_compute_forward_clamp(params, tensor->src[0], tensor);
+                ggml_compute_forward_clamp(params, tensor);
             } break;
         case GGML_OP_CONV_TRANSPOSE_1D:
             {
-                ggml_compute_forward_conv_transpose_1d(params, tensor->src[0], tensor->src[1], tensor);
+                ggml_compute_forward_conv_transpose_1d(params, tensor);
             } break;
         case GGML_OP_IM2COL:
             {
-                ggml_compute_forward_im2col(params, tensor->src[0], tensor->src[1], tensor);
+                ggml_compute_forward_im2col(params, tensor);
+            } break;
+        case GGML_OP_IM2COL_BACK:
+            {
+                ggml_compute_forward_im2col_back_f32(params, tensor);
             } break;
         case GGML_OP_CONV_TRANSPOSE_2D:
             {
-                ggml_compute_forward_conv_transpose_2d(params, tensor->src[0], tensor->src[1], tensor);
+                ggml_compute_forward_conv_transpose_2d(params, tensor);
             } break;
         case GGML_OP_POOL_1D:
             {
-                ggml_compute_forward_pool_1d(params, tensor->src[0], tensor);
+                ggml_compute_forward_pool_1d(params, tensor);
             } break;
         case GGML_OP_POOL_2D:
             {
-                ggml_compute_forward_pool_2d(params, tensor->src[0], tensor);
+                ggml_compute_forward_pool_2d(params, tensor);
+            } break;
+        case GGML_OP_POOL_2D_BACK:
+            {
+                ggml_compute_forward_pool_2d_back(params, tensor);
             } break;
         case GGML_OP_UPSCALE:
             {
-                ggml_compute_forward_upscale(params, tensor->src[0], tensor);
+                ggml_compute_forward_upscale(params, tensor);
             } break;
         case GGML_OP_PAD:
             {
-                ggml_compute_forward_pad(params, tensor->src[0], tensor);
+                ggml_compute_forward_pad(params, tensor);
             } break;
-        case GGML_OP_ARGSORT:
+        case GGML_OP_ARANGE:
             {
-                ggml_compute_forward_argsort(params, tensor->src[0], tensor);
+                ggml_compute_forward_arange(params, tensor);
             } break;
-        case GGML_OP_LEAKY_RELU:
+        case GGML_OP_TIMESTEP_EMBEDDING:
             {
-                ggml_compute_forward_leaky_relu(params, tensor->src[0], tensor);
+                ggml_compute_forward_timestep_embedding(params, tensor);
             } break;
-        case GGML_OP_FLASH_ATTN:
+        case GGML_OP_ARGSORT:
             {
-                const int32_t t = ggml_get_op_params_i32(tensor, 0);
-                GGML_ASSERT(t == 0 || t == 1);
-                const bool masked = t != 0;
-                ggml_compute_forward_flash_attn(params, tensor->src[0], tensor->src[1], tensor->src[2], masked, tensor);
+                ggml_compute_forward_argsort(params, tensor);
+            } break;
+        case GGML_OP_LEAKY_RELU:
+            {
+                ggml_compute_forward_leaky_relu(params, tensor);
             } break;
-        case GGML_OP_FLASH_FF:
+        case GGML_OP_FLASH_ATTN_EXT:
             {
-                ggml_compute_forward_flash_ff(params, tensor->src[0], tensor->src[1], tensor->src[2], tensor->src[3], tensor->src[4], tensor);
+                ggml_compute_forward_flash_attn_ext(params, tensor->src[0], tensor->src[1], tensor->src[2], tensor->src[3], tensor);
             } break;
         case GGML_OP_FLASH_ATTN_BACK:
             {
                 int32_t t = ggml_get_op_params_i32(tensor, 0);
                 GGML_ASSERT(t == 0 || t == 1);
                 bool masked = t != 0;
-                ggml_compute_forward_flash_attn_back(params, tensor->src[0], tensor->src[1], tensor->src[2], tensor->src[3], masked, tensor);
+                ggml_compute_forward_flash_attn_back(params, masked, tensor);
+            } break;
+        case GGML_OP_SSM_CONV:
+            {
+                ggml_compute_forward_ssm_conv(params, tensor);
+            } break;
+        case GGML_OP_SSM_SCAN:
+            {
+                ggml_compute_forward_ssm_scan(params, tensor);
             } break;
         case GGML_OP_WIN_PART:
             {
-                ggml_compute_forward_win_part(params, tensor->src[0], tensor);
+                ggml_compute_forward_win_part(params, tensor);
             } break;
         case GGML_OP_WIN_UNPART:
             {
-                ggml_compute_forward_win_unpart(params, tensor->src[0], tensor);
+                ggml_compute_forward_win_unpart(params, tensor);
             } break;
         case GGML_OP_UNARY:
             {
-                ggml_compute_forward_unary(params, tensor->src[0], tensor);
+                ggml_compute_forward_unary(params, tensor);
             } break;
         case GGML_OP_GET_REL_POS:
             {
-                ggml_compute_forward_get_rel_pos(params, tensor->src[0], tensor);
+                ggml_compute_forward_get_rel_pos(params, tensor);
             } break;
         case GGML_OP_ADD_REL_POS:
             {
-                ggml_compute_forward_add_rel_pos(params, tensor->src[0], tensor->src[1], tensor->src[2], tensor);
+                ggml_compute_forward_add_rel_pos(params, tensor);
+            } break;
+        case GGML_OP_RWKV_WKV:
+            {
+                ggml_compute_forward_rwkv_wkv(params, tensor);
             } break;
         case GGML_OP_MAP_UNARY:
             {
                 ggml_unary_op_f32_t fun;
                 memcpy(&fun, tensor->op_params, sizeof(fun));
-                ggml_compute_forward_map_unary(params, tensor->src[0], tensor, fun);
+                ggml_compute_forward_map_unary(params, tensor, fun);
             }
             break;
         case GGML_OP_MAP_BINARY:
             {
                 ggml_binary_op_f32_t fun;
                 memcpy(&fun, tensor->op_params, sizeof(fun));
-                ggml_compute_forward_map_binary(params, tensor->src[0], tensor->src[1], tensor, fun);
+                ggml_compute_forward_map_binary(params, tensor, fun);
             }
             break;
         case GGML_OP_MAP_CUSTOM1_F32:
             {
                 ggml_custom1_op_f32_t fun;
                 memcpy(&fun, tensor->op_params, sizeof(fun));
-                ggml_compute_forward_map_custom1_f32(params, tensor->src[0], tensor, fun);
+                ggml_compute_forward_map_custom1_f32(params, tensor, fun);
             }
             break;
         case GGML_OP_MAP_CUSTOM2_F32:
             {
                 ggml_custom2_op_f32_t fun;
                 memcpy(&fun, tensor->op_params, sizeof(fun));
-                ggml_compute_forward_map_custom2_f32(params, tensor->src[0], tensor->src[1], tensor, fun);
+                ggml_compute_forward_map_custom2_f32(params, tensor, fun);
             }
             break;
         case GGML_OP_MAP_CUSTOM3_F32:
             {
                 ggml_custom3_op_f32_t fun;
                 memcpy(&fun, tensor->op_params, sizeof(fun));
-                ggml_compute_forward_map_custom3_f32(params, tensor->src[0], tensor->src[1], tensor->src[2], tensor, fun);
+                ggml_compute_forward_map_custom3_f32(params, tensor, fun);
             }
             break;
         case GGML_OP_MAP_CUSTOM1:
             {
-                ggml_compute_forward_map_custom1(params, tensor->src[0], tensor);
+                ggml_compute_forward_map_custom1(params, tensor);
             }
             break;
         case GGML_OP_MAP_CUSTOM2:
             {
-                ggml_compute_forward_map_custom2(params, tensor->src[0], tensor->src[1], tensor);
+                ggml_compute_forward_map_custom2(params, tensor);
             }
             break;
         case GGML_OP_MAP_CUSTOM3:
             {
-                ggml_compute_forward_map_custom3(params, tensor->src[0], tensor->src[1], tensor->src[2], tensor);
+                ggml_compute_forward_map_custom3(params, tensor);
             }
             break;
         case GGML_OP_CROSS_ENTROPY_LOSS:
             {
-                ggml_compute_forward_cross_entropy_loss(params, tensor->src[0], tensor->src[1], tensor);
+                ggml_compute_forward_cross_entropy_loss(params, tensor);
             }
             break;
         case GGML_OP_CROSS_ENTROPY_LOSS_BACK:
             {
-                ggml_compute_forward_cross_entropy_loss_back(params, tensor->src[0], tensor->src[1], tensor->src[2], tensor);
+                ggml_compute_forward_cross_entropy_loss_back(params, tensor);
+            }
+            break;
+        case GGML_OP_OPT_STEP_ADAMW:
+            {
+                ggml_compute_forward_opt_step_adamw(params, tensor);
             }
             break;
         case GGML_OP_NONE:
@@ -14751,14 +17572,32 @@ static void ggml_compute_forward(struct ggml_compute_params * params, struct ggm
             } break;
         case GGML_OP_COUNT:
             {
-                GGML_ASSERT(false);
-            } break;
+                GGML_ABORT("fatal error");
+            }
     }
 }
 
 ////////////////////////////////////////////////////////////////////////////////
 
-static size_t ggml_hash_size(size_t min_sz) {
+struct ggml_hash_set ggml_hash_set_new(size_t size) {
+    size = ggml_hash_size(size);
+    struct ggml_hash_set result;
+    result.size = size;
+    result.keys = GGML_MALLOC(sizeof(struct ggml_tensor *) * size);
+    result.used = GGML_CALLOC(ggml_bitset_size(size), sizeof(ggml_bitset_t));
+    return result;
+}
+
+void ggml_hash_set_reset(struct ggml_hash_set * hash_set) {
+    memset(hash_set->used, 0, sizeof(ggml_bitset_t) * ggml_bitset_size(hash_set->size));
+}
+
+void ggml_hash_set_free(struct ggml_hash_set * hash_set) {
+    GGML_FREE(hash_set->used);
+    GGML_FREE(hash_set->keys);
+}
+
+size_t ggml_hash_size(size_t min_sz) {
     // next primes after powers of two
     static const size_t primes[] = {
         2, 3, 5, 11, 17, 37, 67, 131, 257, 521, 1031,
@@ -14769,7 +17608,7 @@ static size_t ggml_hash_size(size_t min_sz) {
     };
     static const size_t n_primes = sizeof(primes)/sizeof(primes[0]);
 
-    // find the smallest prime that is larger or equal to min_sz
+    // find the smallest prime that is larger or equal than min_sz
     size_t l = 0;
     size_t r = n_primes;
     while (l < r) {
@@ -14784,84 +17623,22 @@ static size_t ggml_hash_size(size_t min_sz) {
     return sz;
 }
 
-static size_t ggml_hash(const void * p) {
-    return (size_t)p;
-}
-
-size_t ggml_hash_find(const struct ggml_hash_set hash_set, struct ggml_tensor * key) {
-    size_t h = ggml_hash(key) % hash_set.size;
-
-    // linear probing
-    size_t i = h;
-    while (hash_set.keys[i] != NULL && hash_set.keys[i] != key) {
-        i = (i + 1) % hash_set.size;
-        if (i == h) {
-            // visited all hash table entries -> not found
-            return GGML_HASHTABLE_FULL;
-        }
-    }
-    return i;
-}
-
-bool ggml_hash_contains(struct ggml_hash_set hash_set, struct ggml_tensor * key) {
-    size_t i = ggml_hash_find(hash_set, key);
-    return i != GGML_HASHTABLE_FULL && hash_set.keys[i] == key;
-}
-
-size_t ggml_hash_insert(struct ggml_hash_set hash_set, struct ggml_tensor * key) {
-    size_t i = ggml_hash_find(hash_set, key);
-
-    GGML_ASSERT(i != GGML_HASHTABLE_FULL);
-
-    if (hash_set.keys[i] == key) {
-        return GGML_HASHTABLE_ALREADY_EXISTS;
-    }
-
-    // insert
-    GGML_ASSERT(hash_set.keys[i] == NULL);
-    hash_set.keys[i] = key;
-    return i;
-}
-
-size_t ggml_hash_find_or_insert(struct ggml_hash_set hash_set, struct ggml_tensor * key) {
-    size_t i = ggml_hash_find(hash_set, key);
-
-    GGML_ASSERT(i != GGML_HASHTABLE_FULL);
-
-    hash_set.keys[i] = key;
-    return i;
-}
-
-static struct ggml_hash_set ggml_hash_set_new(size_t size) {
-    size = ggml_hash_size(size);
-    struct ggml_hash_set result;
-    result.size = size;
-    result.keys = malloc(sizeof(struct ggml_tensor *) * size);
-    memset(result.keys, 0, sizeof(struct ggml_tensor *) * size);
-    return result;
-}
-
-static void ggml_hash_set_free(struct ggml_hash_set hash_set) {
-    free(hash_set.keys);
-}
-
 struct hash_map {
     struct ggml_hash_set set;
     struct ggml_tensor ** vals;
 };
 
 static struct hash_map * ggml_new_hash_map(size_t size) {
-    struct hash_map * result = malloc(sizeof(struct hash_map));
+    struct hash_map * result = GGML_MALLOC(sizeof(struct hash_map));
     result->set = ggml_hash_set_new(size);
-    result->vals = malloc(sizeof(struct ggml_tensor *) * result->set.size);
-    memset(result->vals, 0, sizeof(struct ggml_tensor *) * result->set.size);
+    result->vals = GGML_CALLOC(result->set.size, sizeof(struct ggml_tensor *));
     return result;
 }
 
 static void ggml_hash_map_free(struct hash_map * map) {
-    ggml_hash_set_free(map->set);
-    free(map->vals);
-    free(map);
+    ggml_hash_set_free(&map->set);
+    GGML_FREE(map->vals);
+    GGML_FREE(map);
 }
 
 // gradient checkpointing
@@ -14876,11 +17653,11 @@ static struct ggml_tensor * ggml_recompute_graph_node(
         return NULL;
     }
 
-    if (node->is_param) {
+    if (node->flags & GGML_TENSOR_FLAG_PARAM) {
         return node;
     }
 
-    if (!ggml_hash_contains(graph->visited_hash_table, node)) {
+    if (!ggml_hash_contains(&graph->visited_hash_set, node)) {
         return node;
     }
 
@@ -14895,8 +17672,8 @@ static struct ggml_tensor * ggml_recompute_graph_node(
         return node;
     }
 
-    size_t i = ggml_hash_find(replacements->set, node);
-    GGML_ASSERT(i != GGML_HASHTABLE_FULL); // assert that not full
+    size_t i = ggml_hash_find(&replacements->set, node);
+    GGML_ASSERT(i != GGML_HASHSET_FULL); // assert that not full
     if (replacements->set.keys[i] == node) {
         return replacements->vals[i];
     }
@@ -14910,7 +17687,7 @@ static struct ggml_tensor * ggml_recompute_graph_node(
 
     clone->op       = node->op;
     clone->grad     = node->grad;
-    clone->is_param = node->is_param;
+    clone->flags    = node->flags;
     clone->extra    = node->extra;
     for (int k = 0; k < GGML_MAX_DIMS; ++k) {
         clone->nb[k] = node->nb[k];
@@ -14943,7 +17720,7 @@ void ggml_build_backward_gradient_checkpointing(
         struct ggml_tensor  * * checkpoints,
         int                     n_checkpoints) {
     ggml_graph_cpy(gf, gb_tmp);
-    ggml_build_backward_expand(ctx, gf, gb_tmp, true);
+    ggml_build_backward_expand(ctx, gf, gb_tmp, false);
 
     if (n_checkpoints <= 0) {
         ggml_graph_cpy(gb_tmp, gb);
@@ -14954,8 +17731,8 @@ void ggml_build_backward_gradient_checkpointing(
 
     // insert checkpoints in replacements
     for (int i = 0; i < n_checkpoints; ++i) {
-        size_t k = ggml_hash_find(replacements->set, checkpoints[i]);
-        GGML_ASSERT(k != GGML_HASHTABLE_FULL); // assert that not full
+        size_t k = ggml_hash_find(&replacements->set, checkpoints[i]);
+        GGML_ASSERT(k != GGML_HASHSET_FULL); // assert that not full
         GGML_ASSERT(replacements->set.keys[k] == NULL); // assert that we don't overwrite
         replacements->set.keys[k] = checkpoints[i];
         replacements->vals[k]     = checkpoints[i];
@@ -14981,77 +17758,133 @@ void ggml_build_backward_gradient_checkpointing(
     ggml_hash_map_free(replacements);
 }
 
-// functions to change gradients considering the case that input a might be initial gradient with zero value
-
-static struct ggml_tensor * ggml_add_or_set(struct ggml_context * ctx, struct ggml_tensor * a, struct ggml_tensor * b, struct ggml_hash_set zero_table) {
+// utility functions to change gradients
+// if a is in acc_table, modify gradients in-place and mark result as gradient accumulator
+// else if a is in zero_table, replace a
+// else, just add/subtract/etc. the gradients
+
+static struct ggml_tensor * ggml_add_or_set(
+        struct ggml_context  * ctx,
+        struct ggml_tensor   * a,
+        struct ggml_tensor   * b,
+        struct ggml_hash_set * zero_table,
+        struct ggml_hash_set * acc_table) {
+    if (ggml_hash_contains(acc_table, a)) {
+        struct ggml_tensor * ret = ggml_add_impl(ctx, a, b, true);
+        const size_t insert_result = ggml_hash_insert(acc_table, ret);
+        GGML_ASSERT(insert_result != GGML_HASHSET_FULL);
+        GGML_ASSERT(insert_result != GGML_HASHSET_ALREADY_EXISTS);
+        return ret;
+    }
     if (ggml_hash_contains(zero_table, a)) {
         return b;
-    } else {
-        return ggml_add_impl(ctx, a, b, false);
     }
+    return ggml_add_impl(ctx, a, b, false);
 }
 
-static struct ggml_tensor * ggml_acc_or_set(struct ggml_context * ctx, struct ggml_tensor * a, struct ggml_tensor * b, size_t nb1, size_t nb2, size_t nb3, size_t offset, struct ggml_hash_set zero_table) {
+static struct ggml_tensor * ggml_acc_or_set(
+        struct ggml_context  * ctx,
+        struct ggml_tensor   * a,
+        struct ggml_tensor   * b,
+        const  size_t          nb1,
+        const  size_t          nb2,
+        const  size_t          nb3,
+        const  size_t          offset,
+        struct ggml_hash_set * zero_table,
+        struct ggml_hash_set * acc_table) {
+    if (ggml_hash_contains(acc_table, a)) {
+        struct ggml_tensor * ret = ggml_acc_impl(ctx, a, b, nb1, nb2, nb3, offset, true);
+        const size_t insert_result = ggml_hash_insert(acc_table, ret);
+        GGML_ASSERT(insert_result != GGML_HASHSET_FULL);
+        GGML_ASSERT(insert_result != GGML_HASHSET_ALREADY_EXISTS);
+        return ret;
+    }
     if (ggml_hash_contains(zero_table, a)) {
-        struct ggml_tensor * a_zero = ggml_scale(ctx, a, 0.0f);
+        struct ggml_tensor * a_zero = ggml_scale(ctx, a, 0.0f); // FIXME this is going to produce NaN if a contains inf/NaN
         return ggml_acc_impl(ctx, a_zero, b, nb1, nb2, nb3, offset, false);
-    } else {
-        return ggml_acc_impl(ctx, a, b, nb1, nb2, nb3, offset, false);
     }
+    return ggml_acc_impl(ctx, a, b, nb1, nb2, nb3, offset, false);
 }
 
-static struct ggml_tensor * ggml_add1_or_set(struct ggml_context * ctx, struct ggml_tensor * a, struct ggml_tensor * b, struct ggml_hash_set zero_table) {
+static struct ggml_tensor * ggml_add1_or_set(
+        struct ggml_context  * ctx,
+        struct ggml_tensor   * a,
+        struct ggml_tensor   * b,
+        struct ggml_hash_set * zero_table,
+        struct ggml_hash_set * acc_table) {
+    if (ggml_hash_contains(acc_table, a)) {
+        struct ggml_tensor * ret = ggml_add1_impl(ctx, a, b, true);
+        const size_t insert_result = ggml_hash_insert(acc_table, ret);
+        GGML_ASSERT(insert_result != GGML_HASHSET_FULL);
+        GGML_ASSERT(insert_result != GGML_HASHSET_ALREADY_EXISTS);
+        return ret;
+    }
     if (ggml_hash_contains(zero_table, a)) {
         return ggml_repeat(ctx, b, a);
-    } else {
-        return ggml_add1_impl(ctx, a, b, false);
     }
+    return ggml_add1_impl(ctx, a, b, false);
 }
 
-static struct ggml_tensor * ggml_sub_or_set(struct ggml_context * ctx, struct ggml_tensor * a, struct ggml_tensor * b, struct ggml_hash_set zero_table) {
+static struct ggml_tensor * ggml_sub_or_set(
+        struct ggml_context  * ctx,
+        struct ggml_tensor   * a,
+        struct ggml_tensor   * b,
+        struct ggml_hash_set * zero_table,
+        struct ggml_hash_set * acc_table) {
+    if (ggml_hash_contains(acc_table, a)) {
+        struct ggml_tensor * ret = ggml_sub_impl(ctx, a, b, true);
+        const size_t insert_result = ggml_hash_insert(acc_table, ret);
+        GGML_ASSERT(insert_result != GGML_HASHSET_FULL);
+        GGML_ASSERT(insert_result != GGML_HASHSET_ALREADY_EXISTS);
+        return ret;
+    }
     if (ggml_hash_contains(zero_table, a)) {
         return ggml_neg(ctx, b);
-    } else {
-        return ggml_sub_impl(ctx, a, b, false);
     }
+    return ggml_sub_impl(ctx, a, b, false);
 }
 
-static void ggml_compute_backward(struct ggml_context * ctx, struct ggml_tensor * tensor, struct ggml_hash_set zero_table) {
+static void ggml_compute_backward(struct ggml_context * ctx, struct ggml_tensor * tensor, struct ggml_hash_set * zero_table, struct ggml_hash_set * acc_table) {
     struct ggml_tensor * src0 = tensor->src[0];
     struct ggml_tensor * src1 = tensor->src[1];
+    struct ggml_tensor * src2 = tensor->src[2];
 
     switch (tensor->op) {
         case GGML_OP_DUP:
             {
                 if (src0->grad) {
-                    src0->grad = ggml_add_or_set(ctx, src0->grad, tensor->grad, zero_table);
+                    src0->grad = ggml_add_or_set(ctx, src0->grad, tensor->grad, zero_table, acc_table);
                 }
             } break;
         case GGML_OP_ADD:
             {
                 if (src0->grad) {
-                    src0->grad = ggml_add_or_set(ctx, src0->grad, tensor->grad, zero_table);
+                    src0->grad = ggml_add_or_set(ctx, src0->grad, tensor->grad, zero_table, acc_table);
                 }
                 if (src1->grad) {
-                    src1->grad = ggml_add_or_set(ctx, src1->grad, tensor->grad, zero_table);
+                    if (ggml_are_same_shape(src0, src1)) {
+                        src1->grad = ggml_add_or_set(ctx, src1->grad,                       tensor->grad,        zero_table, acc_table);
+                    } else {
+                        src1->grad = ggml_add_or_set(ctx, src1->grad, ggml_repeat_back(ctx, tensor->grad, src1), zero_table, acc_table);
+                    }
                 }
             } break;
         case GGML_OP_ADD1:
             {
                 if (src0->grad) {
-                    src0->grad = ggml_add_or_set(ctx, src0->grad, tensor->grad, zero_table);
+                    src0->grad = ggml_add_or_set(ctx, src0->grad, tensor->grad, zero_table, acc_table);
                 }
                 if (src1->grad) {
                     src1->grad = ggml_add_or_set(ctx,
                         src1->grad,
                         ggml_mean(ctx, tensor->grad), // TODO: should probably be sum instead of mean
-                        zero_table);
+                        zero_table, acc_table);
                 }
             } break;
         case GGML_OP_ACC:
             {
                 if (src0->grad) {
-                    src0->grad = ggml_add_or_set(ctx, src0->grad, tensor->grad, zero_table);
+                    src0->grad = ggml_add_or_set(ctx, src0->grad, tensor->grad, zero_table, acc_table);
                 }
                 if (src1->grad) {
                     const size_t nb1     = ((int32_t *) tensor->op_params)[0];
@@ -15073,16 +17906,16 @@ static void ggml_compute_backward(struct ggml_context * ctx, struct ggml_tensor
                             ggml_reshape(ctx,
                                 ggml_cont(ctx, tensor_grad_view),
                                 src1->grad),
-                            zero_table);
+                            zero_table, acc_table);
                 }
             } break;
         case GGML_OP_SUB:
             {
                 if (src0->grad) {
-                    src0->grad = ggml_add_or_set(ctx, src0->grad, tensor->grad, zero_table);
+                    src0->grad = ggml_add_or_set(ctx, src0->grad, tensor->grad, zero_table, acc_table);
                 }
                 if (src1->grad) {
-                    src1->grad = ggml_sub_or_set(ctx, src1->grad, tensor->grad, zero_table);
+                    src1->grad = ggml_sub_or_set(ctx, src1->grad, tensor->grad, zero_table, acc_table);
                 }
             } break;
         case GGML_OP_MUL:
@@ -15092,14 +17925,14 @@ static void ggml_compute_backward(struct ggml_context * ctx, struct ggml_tensor
                         ggml_add_or_set(ctx,
                                 src0->grad,
                                 ggml_mul(ctx, src1, tensor->grad),
-                                zero_table);
+                                zero_table, acc_table);
                 }
                 if (src1->grad) {
                     src1->grad =
                         ggml_add_or_set(ctx,
                                 src1->grad,
                                 ggml_mul(ctx, src0, tensor->grad),
-                                zero_table);
+                                zero_table, acc_table);
                 }
             } break;
         case GGML_OP_DIV:
@@ -15109,7 +17942,7 @@ static void ggml_compute_backward(struct ggml_context * ctx, struct ggml_tensor
                         ggml_add_or_set(ctx,
                                 src0->grad,
                                 ggml_div(ctx, tensor->grad, src1),
-                                zero_table);
+                                zero_table, acc_table);
                 }
                 if (src1->grad) {
                     src1->grad =
@@ -15118,7 +17951,7 @@ static void ggml_compute_backward(struct ggml_context * ctx, struct ggml_tensor
                                 ggml_mul(ctx,
                                     tensor->grad,
                                     ggml_div(ctx, tensor, src1)),
-                                zero_table);
+                                zero_table, acc_table);
                 }
             } break;
         case GGML_OP_SQR:
@@ -15130,7 +17963,7 @@ static void ggml_compute_backward(struct ggml_context * ctx, struct ggml_tensor
                                 ggml_scale(ctx,
                                     ggml_mul(ctx, src0, tensor->grad),
                                     2.0f),
-                                zero_table);
+                                zero_table, acc_table);
                 }
             } break;
         case GGML_OP_SQRT:
@@ -15144,19 +17977,43 @@ static void ggml_compute_backward(struct ggml_context * ctx, struct ggml_tensor
                                         tensor->grad,
                                         tensor),
                                     0.5f),
-                                zero_table);
+                                zero_table, acc_table);
                 }
             } break;
         case GGML_OP_LOG:
             {
                 if (src0->grad) {
                     src0->grad =
-                        ggml_add_or_set(ctx,
+                        ggml_add_or_set(ctx,
+                                src0->grad,
+                                ggml_div(ctx,
+                                    tensor->grad,
+                                    src0),
+                                zero_table, acc_table);
+                }
+            } break;
+        case GGML_OP_SIN:
+            {
+                if (src0->grad) {
+                    src0->grad =
+                        ggml_add_or_set(ctx,
+                                src0->grad,
+                                ggml_mul(ctx,
+                                    tensor->grad,
+                                    ggml_cos(ctx, src0)),
+                                zero_table, acc_table);
+                }
+            } break;
+        case GGML_OP_COS:
+            {
+                if (src0->grad) {
+                    src0->grad =
+                        ggml_sub_or_set(ctx,
                                 src0->grad,
-                                ggml_div(ctx,
+                                ggml_mul(ctx,
                                     tensor->grad,
-                                    src0),
-                                zero_table);
+                                    ggml_sin(ctx, src0)),
+                                zero_table, acc_table);
                 }
             } break;
         case GGML_OP_SUM:
@@ -15166,7 +18023,7 @@ static void ggml_compute_backward(struct ggml_context * ctx, struct ggml_tensor
                         ggml_add1_or_set(ctx,
                                 src0->grad,
                                 tensor->grad,
-                                zero_table);
+                                zero_table, acc_table);
                 }
             } break;
         case GGML_OP_SUM_ROWS:
@@ -15178,14 +18035,15 @@ static void ggml_compute_backward(struct ggml_context * ctx, struct ggml_tensor
                                 ggml_repeat(ctx,
                                     tensor->grad,
                                     src0->grad),
-                                zero_table);
+                                zero_table, acc_table);
                 }
             } break;
         case GGML_OP_MEAN:
         case GGML_OP_ARGMAX:
+        case GGML_OP_COUNT_EQUAL:
             {
-                GGML_ASSERT(false); // TODO: implement
-            } break;
+                GGML_ABORT("fatal error"); // TODO: implement
+            }
         case GGML_OP_REPEAT:
             {
                 // necessary for llama
@@ -15193,7 +18051,7 @@ static void ggml_compute_backward(struct ggml_context * ctx, struct ggml_tensor
                     src0->grad = ggml_add_or_set(ctx,
                             src0->grad,
                             ggml_repeat_back(ctx, tensor->grad, src0->grad),
-                            zero_table);
+                            zero_table, acc_table);
                 }
             } break;
         case GGML_OP_REPEAT_BACK:
@@ -15203,21 +18061,21 @@ static void ggml_compute_backward(struct ggml_context * ctx, struct ggml_tensor
                     src0->grad = ggml_add_or_set(ctx,
                             src0->grad,
                             ggml_repeat(ctx, tensor->grad, src0->grad),
-                            zero_table);
+                            zero_table, acc_table);
                 }
             } break;
         case GGML_OP_CONCAT:
             {
-                GGML_ASSERT(false); // TODO: implement
-            } break;
+                GGML_ABORT("fatal error"); // TODO: implement
+            }
         case GGML_OP_SILU_BACK:
             {
-                GGML_ASSERT(false); // TODO: not implemented
-            } break;
+                GGML_ABORT("fatal error"); // TODO: not implemented
+            }
         case GGML_OP_NORM:
             {
-                GGML_ASSERT(false); // TODO: not implemented
-            } break;
+                GGML_ABORT("fatal error"); // TODO: not implemented
+            }
         case GGML_OP_RMS_NORM:
             {
                 // necessary for llama
@@ -15228,17 +18086,17 @@ static void ggml_compute_backward(struct ggml_context * ctx, struct ggml_tensor
                     src0->grad = ggml_add_or_set(ctx,
                             src0->grad,
                             ggml_rms_norm_back(ctx, src0, tensor->grad, eps),
-                            zero_table);
+                            zero_table, acc_table);
                 }
             } break;
         case GGML_OP_RMS_NORM_BACK:
             {
-                GGML_ASSERT(false); // TODO: not implemented
-            } break;
+                GGML_ABORT("fatal error"); // TODO: not implemented
+            }
         case GGML_OP_GROUP_NORM:
             {
-                GGML_ASSERT(false); // TODO: not implemented
-            } break;
+                GGML_ABORT("fatal error"); // TODO: not implemented
+            }
         case GGML_OP_MUL_MAT:
             {
                 // https://cs231n.github.io/optimization-2/#staged
@@ -15276,7 +18134,7 @@ static void ggml_compute_backward(struct ggml_context * ctx, struct ggml_tensor
                         ggml_add_or_set(ctx,
                                 src0->grad, // [n,m,q1,r1]
                                 s1_tg,      // [n,m,q1,r1]
-                                zero_table);
+                                zero_table, acc_table);
                 }
                 if (src1->grad) {
                     src1->grad =
@@ -15294,17 +18152,17 @@ static void ggml_compute_backward(struct ggml_context * ctx, struct ggml_tensor
                                     src0,                           // [n,m,q1,r1]
                                     ggml_transpose(ctx,             // [p,m,qq,rr]
                                         tensor->grad)),             // [m,p,qq,rr]
-                                zero_table);
+                                zero_table, acc_table);
                 }
             } break;
         case GGML_OP_MUL_MAT_ID:
             {
-                GGML_ASSERT(false); // TODO: not implemented
-            } break;
+                GGML_ABORT("fatal error"); // TODO: not implemented
+            }
         case GGML_OP_OUT_PROD:
             {
-                GGML_ASSERT(false); // TODO: not implemented
-            } break;
+                GGML_ABORT("fatal error"); // TODO: not implemented
+            }
         case GGML_OP_SCALE:
             {
                 // necessary for llama
@@ -15316,7 +18174,7 @@ static void ggml_compute_backward(struct ggml_context * ctx, struct ggml_tensor
                         ggml_add_or_set(ctx,
                             src0->grad,
                             ggml_scale_impl(ctx, tensor->grad, s, false),
-                            zero_table);
+                            zero_table, acc_table);
                 }
             } break;
         case GGML_OP_SET:
@@ -15331,14 +18189,10 @@ static void ggml_compute_backward(struct ggml_context * ctx, struct ggml_tensor
                 if (src0->grad || src1->grad) {
                     GGML_ASSERT(src0->type == tensor->type);
                     GGML_ASSERT(tensor->grad->type == tensor->type);
-                    GGML_ASSERT(tensor->grad->type == src1->grad->type);
+                    GGML_ASSERT(!src1->grad || src1->grad->type == tensor->grad->type);
 
                     tensor_grad_view = ggml_view_4d(ctx,
-                        tensor->grad,
-                        src1->grad->ne[0],
-                        src1->grad->ne[1],
-                        src1->grad->ne[2],
-                        src1->grad->ne[3],
+                        tensor->grad, src1->ne[0], src1->ne[1], src1->ne[2], src1->ne[3],
                         nb1, nb2, nb3, offset);
                 }
 
@@ -15349,7 +18203,7 @@ static void ggml_compute_backward(struct ggml_context * ctx, struct ggml_tensor
                             tensor->grad,
                             ggml_neg(ctx, tensor_grad_view),
                             nb1, nb2, nb3, offset, false),
-                        zero_table);
+                        zero_table, acc_table);
                 }
 
                 if (src1->grad) {
@@ -15359,7 +18213,7 @@ static void ggml_compute_backward(struct ggml_context * ctx, struct ggml_tensor
                             ggml_reshape(ctx,
                                 ggml_cont(ctx, tensor_grad_view),
                                 src1->grad),
-                            zero_table);
+                            zero_table, acc_table);
                 }
             } break;
         case GGML_OP_CPY:
@@ -15370,7 +18224,7 @@ static void ggml_compute_backward(struct ggml_context * ctx, struct ggml_tensor
                 // tensor = src0 * 1 + src1 * 0
                 if (src0->grad) {
                     // dsrc0 = dtensor * 1
-                    src0->grad = ggml_add_or_set(ctx, src0->grad, tensor->grad, zero_table);
+                    src0->grad = ggml_add_or_set(ctx, src0->grad, tensor->grad, zero_table, acc_table);
                 }
                 if (src1->grad) {
                     // dsrc1 = dtensor * 0 -> noop
@@ -15382,7 +18236,7 @@ static void ggml_compute_backward(struct ggml_context * ctx, struct ggml_tensor
                 if (src0->grad) {
                     GGML_ASSERT(ggml_is_contiguous(src0->grad));
                     GGML_ASSERT(ggml_is_contiguous(tensor->grad));
-                    src0->grad = ggml_add_or_set(ctx, src0->grad, tensor->grad, zero_table);
+                    src0->grad = ggml_add_or_set(ctx, src0->grad, tensor->grad, zero_table, acc_table);
                 }
             } break;
         case GGML_OP_RESHAPE:
@@ -15396,7 +18250,7 @@ static void ggml_compute_backward(struct ggml_context * ctx, struct ggml_tensor
                                     ? tensor->grad
                                     : ggml_cont(ctx, tensor->grad),
                                 src0->grad),
-                        zero_table);
+                        zero_table, acc_table);
                 }
             } break;
         case GGML_OP_VIEW:
@@ -15407,9 +18261,9 @@ static void ggml_compute_backward(struct ggml_context * ctx, struct ggml_tensor
 
                     memcpy(&offset, tensor->op_params, sizeof(offset));
 
-                    size_t nb1     = tensor->nb[1];
-                    size_t nb2     = tensor->nb[2];
-                    size_t nb3     = tensor->nb[3];
+                    size_t nb1 = tensor->nb[1];
+                    size_t nb2 = tensor->nb[2];
+                    size_t nb3 = tensor->nb[3];
 
                     if (src0->type != src0->grad->type) {
                         // gradient is typically F32, but src0 could be other type
@@ -15425,7 +18279,7 @@ static void ggml_compute_backward(struct ggml_context * ctx, struct ggml_tensor
                         nb3 = (nb3 / n0) * ng;
                     }
 
-                    src0->grad = ggml_acc_or_set(ctx, src0->grad, tensor->grad, nb1, nb2, nb3, offset, zero_table);
+                    src0->grad = ggml_acc_or_set(ctx, src0->grad, tensor->grad, nb1, nb2, nb3, offset, zero_table, acc_table);
                 }
             } break;
         case GGML_OP_PERMUTE:
@@ -15450,7 +18304,7 @@ static void ggml_compute_backward(struct ggml_context * ctx, struct ggml_tensor
                                 axes_backward[1],
                                 axes_backward[2],
                                 axes_backward[3]),
-                            zero_table);
+                            zero_table, acc_table);
                 }
             } break;
         case GGML_OP_TRANSPOSE:
@@ -15460,7 +18314,7 @@ static void ggml_compute_backward(struct ggml_context * ctx, struct ggml_tensor
                     src0->grad =
                         ggml_add_or_set(ctx, src0->grad,
                             ggml_transpose(ctx, tensor->grad),
-                        zero_table);
+                        zero_table, acc_table);
                 }
             } break;
         case GGML_OP_GET_ROWS:
@@ -15472,7 +18326,7 @@ static void ggml_compute_backward(struct ggml_context * ctx, struct ggml_tensor
                             // last ggml_get_rows_back argument src0->grad is only
                             // necessary to setup correct output shape
                             ggml_get_rows_back(ctx, tensor->grad, src1, src0->grad),
-                        zero_table);
+                        zero_table, acc_table);
                 }
                 if (src1->grad) {
                     // noop
@@ -15480,12 +18334,12 @@ static void ggml_compute_backward(struct ggml_context * ctx, struct ggml_tensor
             } break;
         case GGML_OP_GET_ROWS_BACK:
             {
-                GGML_ASSERT(false); // TODO: not implemented
-            } break;
+                GGML_ABORT("fatal error"); // TODO: not implemented
+            }
         case GGML_OP_DIAG:
             {
-                GGML_ASSERT(false); // TODO: not implemented
-            } break;
+                GGML_ABORT("fatal error"); // TODO: not implemented
+            }
         case GGML_OP_DIAG_MASK_INF:
             {
                 // necessary for llama
@@ -15496,7 +18350,7 @@ static void ggml_compute_backward(struct ggml_context * ctx, struct ggml_tensor
                             /* ggml_diag_mask_inf_impl() shouldn't be here */
                             /* ref:  https://github.com/ggerganov/llama.cpp/pull/4203#discussion_r1412377992 */
                             ggml_diag_mask_zero_impl(ctx, tensor->grad, n_past, false),
-                        zero_table);
+                        zero_table, acc_table);
                 }
             } break;
         case GGML_OP_DIAG_MASK_ZERO:
@@ -15507,7 +18361,7 @@ static void ggml_compute_backward(struct ggml_context * ctx, struct ggml_tensor
                     src0->grad =
                         ggml_add_or_set(ctx, src0->grad,
                             ggml_diag_mask_zero_impl(ctx, tensor->grad, n_past, false),
-                        zero_table);
+                        zero_table, acc_table);
                 }
             } break;
         case GGML_OP_SOFT_MAX:
@@ -15517,14 +18371,14 @@ static void ggml_compute_backward(struct ggml_context * ctx, struct ggml_tensor
                     src0->grad =
                         ggml_add_or_set(ctx, src0->grad,
                             ggml_soft_max_back(ctx, tensor->grad, tensor),
-                        zero_table);
+                        zero_table, acc_table);
                 }
-
+                GGML_ASSERT((!src1 || !src1->grad) && "backward pass for softmax mask not implemented");
             } break;
         case GGML_OP_SOFT_MAX_BACK:
             {
-                GGML_ASSERT(false); // TODO: not implemented
-            } break;
+                GGML_ABORT("fatal error"); // TODO: not implemented
+            }
         case GGML_OP_ROPE:
             {
                 // necessary for llama
@@ -15532,9 +18386,9 @@ static void ggml_compute_backward(struct ggml_context * ctx, struct ggml_tensor
                     //const int n_past = ((int32_t *) tensor->op_params)[0];
                     const int n_dims     = ((int32_t *) tensor->op_params)[1];
                     const int mode       = ((int32_t *) tensor->op_params)[2];
-                    const int n_ctx      = ((int32_t *) tensor->op_params)[3];
-                    const int n_orig_ctx = ((int32_t *) tensor->op_params)[4];
-                    float freq_base, freq_scale, ext_factor, attn_factor, beta_fast, beta_slow, xpos_base, xpos_down;
+                    //const int n_ctx      = ((int32_t *) tensor->op_params)[3];
+                    const int n_ctx_orig = ((int32_t *) tensor->op_params)[4];
+                    float freq_base, freq_scale, ext_factor, attn_factor, beta_fast, beta_slow;
 
                     memcpy(&freq_base,   (int32_t *) tensor->op_params +  5, sizeof(float));
                     memcpy(&freq_scale,  (int32_t *) tensor->op_params +  6, sizeof(float));
@@ -15542,28 +18396,25 @@ static void ggml_compute_backward(struct ggml_context * ctx, struct ggml_tensor
                     memcpy(&attn_factor, (int32_t *) tensor->op_params +  8, sizeof(float));
                     memcpy(&beta_fast,   (int32_t *) tensor->op_params +  9, sizeof(float));
                     memcpy(&beta_slow,   (int32_t *) tensor->op_params + 10, sizeof(float));
-                    memcpy(&xpos_base,   (int32_t *) tensor->op_params + 11, sizeof(float));
-                    memcpy(&xpos_down,   (int32_t *) tensor->op_params + 12, sizeof(bool));
 
                     src0->grad = ggml_add_or_set(ctx,
                             src0->grad,
                             ggml_rope_back(ctx,
                                 tensor->grad,
                                 src1,
+                                src2,
                                 n_dims,
                                 mode,
-                                n_ctx,
-                                n_orig_ctx,
+                                n_ctx_orig,
                                 freq_base,
                                 freq_scale,
                                 ext_factor,
                                 attn_factor,
                                 beta_fast,
-                                beta_slow,
-                                xpos_base,
-                                xpos_down),
-                            zero_table);
+                                beta_slow),
+                            zero_table, acc_table);
                 }
+                GGML_ASSERT((!src2 || !src2->grad) && "gradients for freq factors not implemented");
             } break;
         case GGML_OP_ROPE_BACK:
             {
@@ -15571,9 +18422,9 @@ static void ggml_compute_backward(struct ggml_context * ctx, struct ggml_tensor
                     //const int n_past = ((int32_t *) tensor->op_params)[0];
                     const int n_dims     = ((int32_t *) tensor->op_params)[1];
                     const int mode       = ((int32_t *) tensor->op_params)[2];
-                    const int n_ctx      = ((int32_t *) tensor->op_params)[3];
-                    const int n_orig_ctx = ((int32_t *) tensor->op_params)[4];
-                    float freq_base, freq_scale, ext_factor, attn_factor, beta_fast, beta_slow, xpos_base, xpos_down;
+                    //const int n_ctx      = ((int32_t *) tensor->op_params)[3];
+                    const int n_ctx_orig = ((int32_t *) tensor->op_params)[4];
+                    float freq_base, freq_scale, ext_factor, attn_factor, beta_fast, beta_slow;
 
                     memcpy(&freq_base,   (int32_t *) tensor->op_params +  5, sizeof(float));
                     memcpy(&freq_scale,  (int32_t *) tensor->op_params +  6, sizeof(float));
@@ -15581,76 +18432,111 @@ static void ggml_compute_backward(struct ggml_context * ctx, struct ggml_tensor
                     memcpy(&attn_factor, (int32_t *) tensor->op_params +  8, sizeof(float));
                     memcpy(&beta_fast,   (int32_t *) tensor->op_params +  9, sizeof(float));
                     memcpy(&beta_slow,   (int32_t *) tensor->op_params + 10, sizeof(float));
-                    memcpy(&xpos_base,   (int32_t *) tensor->op_params + 11, sizeof(float));
-                    memcpy(&xpos_down,   (int32_t *) tensor->op_params + 12, sizeof(bool));
 
                     src0->grad = ggml_add_or_set(ctx,
                             src0->grad,
                             ggml_rope_impl(ctx,
                                 tensor->grad,
                                 src1,
+                                src2,
                                 n_dims,
                                 mode,
-                                n_ctx,
-                                n_orig_ctx,
+                                n_ctx_orig,
                                 freq_base,
                                 freq_scale,
                                 ext_factor,
                                 attn_factor,
                                 beta_fast,
                                 beta_slow,
-                                xpos_base,
-                                xpos_down,
                                 false),
-                            zero_table);
+                            zero_table, acc_table);
                 }
             } break;
-        case GGML_OP_ALIBI:
-            {
-                GGML_ASSERT(false); // TODO: not implemented
-            } break;
         case GGML_OP_CLAMP:
             {
-                GGML_ASSERT(false); // TODO: not implemented
-            } break;
+                GGML_ABORT("fatal error"); // TODO: not implemented
+            }
         case GGML_OP_CONV_TRANSPOSE_1D:
             {
-                GGML_ASSERT(false); // TODO: not implemented
-            } break;
+                GGML_ABORT("fatal error"); // TODO: not implemented
+            }
         case GGML_OP_IM2COL:
             {
-                GGML_ASSERT(false); // TODO: not implemented
+                if (src1->grad) {
+                    const int32_t s0    = ggml_get_op_params_i32(tensor, 0);
+                    const int32_t s1    = ggml_get_op_params_i32(tensor, 1);
+                    const int32_t p0    = ggml_get_op_params_i32(tensor, 2);
+                    const int32_t p1    = ggml_get_op_params_i32(tensor, 3);
+                    const int32_t d0    = ggml_get_op_params_i32(tensor, 4);
+                    const int32_t d1    = ggml_get_op_params_i32(tensor, 5);
+                    const bool    is_2D = ggml_get_op_params_i32(tensor, 6) == 1;
+
+                    src1->grad = ggml_add_or_set(ctx,
+                            src1->grad,
+                            ggml_im2col_back(ctx, src0, tensor->grad, src1->ne, s0, s1, p0, p1, d0, d1, is_2D),
+                            zero_table, acc_table);
+                }
             } break;
+        case GGML_OP_IM2COL_BACK:
+            {
+                GGML_ABORT("fatal error"); // TODO: not implemented
+            }
         case GGML_OP_CONV_TRANSPOSE_2D:
             {
-                GGML_ASSERT(false); // TODO: not implemented
-            } break;
+                GGML_ABORT("fatal error"); // TODO: not implemented
+            }
         case GGML_OP_POOL_1D:
             {
-                GGML_ASSERT(false); // TODO: not implemented
-            } break;
+                GGML_ABORT("fatal error"); // TODO: not implemented
+            }
         case GGML_OP_POOL_2D:
             {
-                GGML_ASSERT(false); // TODO: not implemented
+                if (src0->grad) {
+                    const enum ggml_op_pool op = ggml_get_op_params_i32(tensor, 0);
+                    const      int32_t      k0 = ggml_get_op_params_i32(tensor, 1);
+                    const      int32_t      k1 = ggml_get_op_params_i32(tensor, 2);
+                    const      int32_t      s0 = ggml_get_op_params_i32(tensor, 3);
+                    const      int32_t      s1 = ggml_get_op_params_i32(tensor, 4);
+                    const      int32_t      p0 = ggml_get_op_params_i32(tensor, 5);
+                    const      int32_t      p1 = ggml_get_op_params_i32(tensor, 6);
+
+                    src0->grad = ggml_add_or_set(ctx,
+                            src0->grad,
+                            ggml_pool_2d_back(ctx, tensor->grad, src0, op, k0, k1, s0, s1, p0, p1),
+                            zero_table, acc_table);
+                }
             } break;
+        case GGML_OP_POOL_2D_BACK:
+            {
+                GGML_ABORT("fatal error"); // TODO: not implemented
+            }
         case GGML_OP_UPSCALE:
             {
-                GGML_ASSERT(false); // TODO: not implemented
-            } break;
+                GGML_ABORT("fatal error"); // TODO: not implemented
+            }
         case GGML_OP_PAD:
             {
-                GGML_ASSERT(false); // TODO: not implemented
-            } break;
+                GGML_ABORT("fatal error"); // TODO: not implemented
+            }
+        case GGML_OP_ARANGE:
+            {
+                GGML_ABORT("fatal error"); // TODO: not implemented
+            }
+        case GGML_OP_TIMESTEP_EMBEDDING:
+            {
+                GGML_ABORT("fatal error"); // TODO: not implemented
+            }
         case GGML_OP_ARGSORT:
             {
-                GGML_ASSERT(false); // TODO: not implemented
-            } break;
+                GGML_ABORT("fatal error"); // TODO: not implemented
+            }
         case GGML_OP_LEAKY_RELU:
             {
-                GGML_ASSERT(false); // TODO: not implemented
-            } break;
-        case GGML_OP_FLASH_ATTN:
+                GGML_ABORT("fatal error"); // TODO: not implemented
+            }
+        case GGML_OP_FLASH_ATTN_EXT:
             {
+                GGML_ABORT("FA backward pass not adapted after rework");
                 struct ggml_tensor * flash_grad = NULL;
                 if (src0->grad || src1->grad || tensor->src[2]->grad) {
                     int32_t t = ggml_get_op_params_i32(tensor, 0);
@@ -15665,7 +18551,6 @@ static void ggml_compute_backward(struct ggml_context * ctx, struct ggml_tensor
                             masked);
                 }
 
-                struct ggml_tensor * src2 = tensor->src[2];
                 const int64_t elem_q = ggml_nelements(src0);
                 const int64_t elem_k = ggml_nelements(src1);
                 const int64_t elem_v = ggml_nelements(src2);
@@ -15684,7 +18569,7 @@ static void ggml_compute_backward(struct ggml_context * ctx, struct ggml_tensor
                     src0->grad = ggml_add_or_set(ctx,
                             src0->grad,
                             grad_q,
-                            zero_table);
+                            zero_table, acc_table);
                 }
                 if (src1->grad) {
                     struct ggml_tensor * view_k = ggml_view_1d(ctx, flash_grad, elem_k, offs_k);
@@ -15692,7 +18577,7 @@ static void ggml_compute_backward(struct ggml_context * ctx, struct ggml_tensor
                     src1->grad = ggml_add_or_set(ctx,
                             src1->grad,
                             grad_k,
-                            zero_table);
+                            zero_table, acc_table);
                 }
                 if (src2->grad) {
                     struct ggml_tensor * view_v = ggml_view_1d(ctx, flash_grad, elem_v, offs_v);
@@ -15700,17 +18585,18 @@ static void ggml_compute_backward(struct ggml_context * ctx, struct ggml_tensor
                     src2->grad = ggml_add_or_set(ctx,
                             src2->grad,
                             grad_v,
-                            zero_table);
+                            zero_table, acc_table);
                 }
             } break;
-        case GGML_OP_FLASH_FF:
-            {
-                GGML_ASSERT(false); // not supported
-            } break;
         case GGML_OP_FLASH_ATTN_BACK:
             {
-                GGML_ASSERT(false); // not supported
-            } break;
+                GGML_ABORT("fatal error"); // not supported
+            }
+        case GGML_OP_SSM_CONV:
+        case GGML_OP_SSM_SCAN:
+            {
+                GGML_ABORT("fatal error"); // TODO: not implemented
+            }
         case GGML_OP_WIN_PART:
         case GGML_OP_WIN_UNPART:
         case GGML_OP_UNARY:
@@ -15725,7 +18611,7 @@ static void ggml_compute_backward(struct ggml_context * ctx, struct ggml_tensor
                                             ggml_mul(ctx,
                                                 ggml_sgn(ctx, src0),
                                                 tensor->grad),
-                                            zero_table);
+                                            zero_table, acc_table);
                             }
                         } break;
                     case GGML_UNARY_OP_SGN:
@@ -15737,7 +18623,7 @@ static void ggml_compute_backward(struct ggml_context * ctx, struct ggml_tensor
                     case GGML_UNARY_OP_NEG:
                         {
                             if (src0->grad) {
-                                src0->grad = ggml_sub_or_set(ctx, src0->grad, tensor->grad, zero_table);
+                                src0->grad = ggml_sub_or_set(ctx, src0->grad, tensor->grad, zero_table, acc_table);
                             }
                         } break;
                     case GGML_UNARY_OP_STEP:
@@ -15748,12 +18634,12 @@ static void ggml_compute_backward(struct ggml_context * ctx, struct ggml_tensor
                         } break;
                     case GGML_UNARY_OP_TANH:
                         {
-                            GGML_ASSERT(false); // TODO: not implemented
-                        } break;
+                            GGML_ABORT("fatal error"); // TODO: not implemented
+                        }
                     case GGML_UNARY_OP_ELU:
                         {
-                            GGML_ASSERT(false); // TODO: not implemented
-                        } break;
+                            GGML_ABORT("fatal error"); // TODO: not implemented
+                        }
                     case GGML_UNARY_OP_RELU:
                         {
                             if (src0->grad) {
@@ -15762,17 +18648,21 @@ static void ggml_compute_backward(struct ggml_context * ctx, struct ggml_tensor
                                         ggml_mul(ctx,
                                             ggml_step(ctx, src0),
                                             tensor->grad),
-                                        zero_table);
+                                        zero_table, acc_table);
                             }
                         } break;
+                    case GGML_UNARY_OP_SIGMOID:
+                        {
+                            GGML_ABORT("fatal error"); // TODO: not implemented
+                        }
                     case GGML_UNARY_OP_GELU:
                         {
-                            GGML_ASSERT(false); // TODO: not implemented
-                        } break;
+                            GGML_ABORT("fatal error"); // TODO: not implemented
+                        }
                     case GGML_UNARY_OP_GELU_QUICK:
                         {
-                            GGML_ASSERT(false); // TODO: not implemented
-                        } break;
+                            GGML_ABORT("fatal error"); // TODO: not implemented
+                        }
                     case GGML_UNARY_OP_SILU:
                         {
                             // necessary for llama
@@ -15780,15 +18670,25 @@ static void ggml_compute_backward(struct ggml_context * ctx, struct ggml_tensor
                                 src0->grad = ggml_add_or_set(ctx,
                                         src0->grad,
                                         ggml_silu_back(ctx, src0, tensor->grad),
-                                        zero_table);
+                                        zero_table, acc_table);
+                            }
+                        } break;
+                    case GGML_UNARY_OP_EXP:
+                        {
+                            if (src0->grad) {
+                                src0->grad = ggml_add_or_set(ctx,
+                                        src0->grad,
+                                        ggml_mul(ctx, tensor, tensor->grad),
+                                        zero_table, acc_table);
                             }
                         } break;
                     default:
-                        GGML_ASSERT(false);
+                        GGML_ABORT("fatal error");
                 }
             } break;
         case GGML_OP_GET_REL_POS:
         case GGML_OP_ADD_REL_POS:
+        case GGML_OP_RWKV_WKV:
         case GGML_OP_MAP_UNARY:
         case GGML_OP_MAP_BINARY:
         case GGML_OP_MAP_CUSTOM1_F32:
@@ -15798,8 +18698,8 @@ static void ggml_compute_backward(struct ggml_context * ctx, struct ggml_tensor
         case GGML_OP_MAP_CUSTOM2:
         case GGML_OP_MAP_CUSTOM3:
             {
-                GGML_ASSERT(false); // not supported
-            } break;
+                GGML_ABORT("fatal error"); // not supported
+            }
         case GGML_OP_CROSS_ENTROPY_LOSS:
             {
                 if (src0->grad) {
@@ -15809,21 +18709,26 @@ static void ggml_compute_backward(struct ggml_context * ctx, struct ggml_tensor
                                     src0,
                                     src1,
                                     tensor->grad),
-                                zero_table);
+                                zero_table, acc_table);
                 }
+                GGML_ASSERT(!src1->grad && "backward pass for labels not implemented");
             } break;
         case GGML_OP_CROSS_ENTROPY_LOSS_BACK:
             {
-                GGML_ASSERT(false); // not supported
-            } break;
+                GGML_ABORT("fatal error"); // not supported
+            }
+        case GGML_OP_OPT_STEP_ADAMW:
+            {
+                GGML_ABORT("fatal error"); // not supported
+            }
         case GGML_OP_NONE:
             {
                 // nop
             } break;
         case GGML_OP_COUNT:
             {
-                GGML_ASSERT(false);
-            } break;
+                GGML_ABORT("fatal error");
+            }
     }
 
     for (int i = 0; i < GGML_MAX_SRC; ++i) {
@@ -15843,7 +18748,7 @@ static void ggml_visit_parents(struct ggml_cgraph * cgraph, struct ggml_tensor *
     }
 
     // check if already visited
-    if (ggml_hash_insert(cgraph->visited_hash_table, node) == GGML_HASHTABLE_ALREADY_EXISTS) {
+    if (ggml_hash_insert(&cgraph->visited_hash_set, node) == GGML_HASHSET_ALREADY_EXISTS) {
         return;
     }
 
@@ -15857,7 +18762,7 @@ static void ggml_visit_parents(struct ggml_cgraph * cgraph, struct ggml_tensor *
         }
     }
 
-    if (node->op == GGML_OP_NONE && node->grad == NULL) {
+    if (node->op == GGML_OP_NONE && !(node->flags & GGML_TENSOR_FLAG_PARAM)) {
         // reached a leaf node, not part of the gradient graph (e.g. a constant)
         GGML_ASSERT(cgraph->n_leafs < cgraph->size);
 
@@ -15875,9 +18780,6 @@ static void ggml_visit_parents(struct ggml_cgraph * cgraph, struct ggml_tensor *
         }
 
         cgraph->nodes[cgraph->n_nodes] = node;
-        if (cgraph->grads) {
-            cgraph->grads[cgraph->n_nodes] = node->grad;
-        }
         cgraph->n_nodes++;
     }
 }
@@ -15889,7 +18791,6 @@ static void ggml_build_forward_impl(struct ggml_cgraph * cgraph, struct ggml_ten
     }
 
     const int n0 = cgraph->n_nodes;
-    UNUSED(n0);
 
     ggml_visit_parents(cgraph, tensor);
 
@@ -15906,58 +18807,150 @@ void ggml_build_forward_expand(struct ggml_cgraph * cgraph, struct ggml_tensor *
     ggml_build_forward_impl(cgraph, tensor, true);
 }
 
-void ggml_build_backward_expand(struct ggml_context * ctx, struct ggml_cgraph * gf, struct ggml_cgraph * gb, bool keep) {
+void ggml_build_backward_expand(struct ggml_context * ctx, struct ggml_cgraph * gf, struct ggml_cgraph * gb, bool accumulate) {
     GGML_ASSERT(gf->n_nodes > 0);
+    GGML_ASSERT(gf->grads);
+
+    for (int i = 0; i < gf->n_nodes; ++i) {
+        struct ggml_tensor * node = gf->nodes[i];
+
+        if (node->type == GGML_TYPE_I32) {
+            continue;
+        }
+
+        bool needs_grad = node->flags & GGML_TENSOR_FLAG_PARAM;
+        bool ignore_src[GGML_MAX_SRC] = {false};
+        switch (node->op) {
+            // gradients in node->src[0] for one reason or another have no effect on output gradients
+            case GGML_OP_IM2COL:      // only used for its shape
+            case GGML_OP_IM2COL_BACK: // same as IM2COL
+                ignore_src[0] = true;
+                break;
+            case GGML_OP_UNARY: {
+                const enum ggml_unary_op uop = ggml_get_unary_op(node);
+                // SGN and STEP unary ops are piecewise constant
+                if (uop == GGML_UNARY_OP_SGN || uop == GGML_UNARY_OP_STEP) {
+                    ignore_src[0] = true;
+                }
+            } break;
 
-    // if we are keeping the gradient graph, we have to detach the gradient nodes from the original graph
-    if (keep) {
-        for (int i = 0; i < gf->n_nodes; i++) {
-            struct ggml_tensor * node = gf->nodes[i];
+            // gradients in node->src[1] for one reason or another have no effect on output gradients
+            case GGML_OP_CPY:           // gradients in CPY target  are irrelevant
+            case GGML_OP_GET_ROWS:      // row indices not differentiable
+            case GGML_OP_GET_ROWS_BACK: // same as for GET_ROWS
+            case GGML_OP_ROPE:          // positions not differentiable
+                ignore_src[1] = true;
+                break;
 
-            if (node->grad) {
-                node->grad = ggml_dup_tensor(ctx, node);
-                gf->grads[i] = node->grad;
+            default:
+                break;
+        }
+        for (int j = 0; j < GGML_MAX_SRC; ++j) {
+            if (!node->src[j] || !node->src[j]->grad || ignore_src[j]) {
+                continue;
             }
+            GGML_ASSERT(node->src[j]->type == GGML_TYPE_F32 || node->src[j]->type == GGML_TYPE_F16);
+            needs_grad = true;
+            break;
+        }
+        if (!needs_grad) {
+            continue;
         }
+
+        // inplace operations are currently not supported
+        GGML_ASSERT(!node->view_src || node->op == GGML_OP_CPY || node->op == GGML_OP_VIEW ||
+            node->op == GGML_OP_RESHAPE || node->op == GGML_OP_PERMUTE || node->op == GGML_OP_TRANSPOSE);
+
+        // create a new tensor with the same type and shape as the node and set it as grad
+        node->grad = ggml_dup_tensor(ctx, node);
     }
 
-    // remember original gradients which start with zero values
+    // keep tables of original gradients for replacement/accumulation logic
     struct ggml_hash_set zero_table = ggml_hash_set_new(gf->size);
+    struct ggml_hash_set acc_table  = ggml_hash_set_new(gf->size);
     for (int i = 0; i < gf->n_nodes; i++) {
-        if (gf->grads[i]) {
-            ggml_hash_insert(zero_table, gf->grads[i]);
+        struct ggml_tensor * node = gf->nodes[i];
+
+        if (node->grad) {
+            {
+                const size_t insert_result = ggml_hash_insert(&zero_table, node->grad);
+                GGML_ASSERT(insert_result != GGML_HASHSET_FULL);
+                GGML_ASSERT(insert_result != GGML_HASHSET_ALREADY_EXISTS);
+            }
+
+            // only gradients of trainable parameters should be accumulated
+            if (accumulate && (node->flags & GGML_TENSOR_FLAG_PARAM)) {
+                const size_t insert_result = ggml_hash_insert(&acc_table, node->grad);
+                GGML_ASSERT(insert_result != GGML_HASHSET_FULL);
+                GGML_ASSERT(insert_result != GGML_HASHSET_ALREADY_EXISTS);
+            }
         }
     }
 
     for (int i = gf->n_nodes - 1; i >= 0; i--) {
         struct ggml_tensor * node = gf->nodes[i];
 
-        // inplace operations to add gradients are not created by ggml_compute_backward
+        // inplace operations to add gradients are not created by ggml_compute_backward except for gradient accumulation
         // use allocator to automatically make inplace operations
         if (node->grad) {
-            ggml_compute_backward(ctx, node, zero_table);
+            ggml_compute_backward(ctx, node, &zero_table, &acc_table);
         }
     }
 
     for (int i = 0; i < gf->n_nodes; i++) {
         struct ggml_tensor * node = gf->nodes[i];
 
-        if (node->is_param) {
+        if (node->flags & GGML_TENSOR_FLAG_PARAM) {
             GGML_PRINT_DEBUG("%s: found root node %p\n", __func__, (void *) node);
             ggml_build_forward_expand(gb, node->grad);
         }
     }
 
-    ggml_hash_set_free(zero_table);
+    ggml_hash_set_free(&zero_table);
+    ggml_hash_set_free(&acc_table);
+}
+
+void ggml_build_opt_adamw(
+        struct ggml_context * ctx,
+        struct ggml_cgraph  * gf,
+        struct ggml_cgraph  * gb,
+        float                 alpha,
+        float                 beta1,
+        float                 beta2,
+        float                 eps,
+        float                 wd) {
+    for (int i = 0; i < gf->n_nodes; i++) {
+        struct ggml_tensor * node = gf->nodes[i];
+
+        if (node->flags & GGML_TENSOR_FLAG_PARAM) {
+            GGML_PRINT_DEBUG("%s: found root node %p\n", __func__, (void *) node);
+            struct ggml_tensor * opt_step = ggml_opt_step_adamw(ctx, node, node->grad, alpha, beta1, beta2, eps, wd);
+            ggml_build_forward_expand(gb, opt_step);
+        }
+    }
+}
+
+
+static void * incr_ptr_aligned(void ** p, size_t size, size_t align) {
+    void * ptr = *p;
+    ptr = (void *) GGML_PAD((uintptr_t) ptr, align);
+    *p = (void *) ((char *) ptr + size);
+    return ptr;
 }
 
 static size_t ggml_graph_nbytes(size_t size, bool grads) {
-    size_t nbytes = sizeof(struct ggml_cgraph);
-    nbytes += size * sizeof(struct ggml_tensor *) * 2; // leafs + nodes
+    size_t hash_size = ggml_hash_size(size * 2);
+    void * p = 0;
+    incr_ptr_aligned(&p, sizeof(struct ggml_cgraph), 1);
+    incr_ptr_aligned(&p, size * sizeof(struct ggml_tensor *), sizeof(struct ggml_tensor *)); // nodes
+    incr_ptr_aligned(&p, size * sizeof(struct ggml_tensor *), sizeof(struct ggml_tensor *)); // leafs
+    incr_ptr_aligned(&p, hash_size * sizeof(struct ggml_tensor *), sizeof(struct ggml_tensor *)); // hash keys
     if (grads) {
-        nbytes += size * sizeof(struct ggml_tensor *); // grads
+        incr_ptr_aligned(&p, size * sizeof(struct ggml_tensor *), sizeof(struct ggml_tensor *)); // grads
     }
-    nbytes += ggml_hash_size(size * 2) * sizeof(struct ggml_tensor *); // hash set
+    incr_ptr_aligned(&p, ggml_bitset_size(hash_size) * sizeof(ggml_bitset_t), sizeof(ggml_bitset_t));
+
+    size_t nbytes = (size_t) p;
     return nbytes;
 }
 
@@ -15971,22 +18964,22 @@ size_t ggml_graph_overhead(void) {
 
 struct ggml_cgraph * ggml_new_graph_custom(struct ggml_context * ctx, size_t size, bool grads) {
     const size_t obj_size = ggml_graph_nbytes(size, grads);
-    struct ggml_object * obj = ggml_new_object(ctx, GGML_OBJECT_GRAPH, obj_size);
+    struct ggml_object * obj = ggml_new_object(ctx, GGML_OBJECT_TYPE_GRAPH, obj_size);
     struct ggml_cgraph * cgraph = (struct ggml_cgraph *) ((char *) ctx->mem_buffer + obj->offs);
 
-    struct ggml_tensor ** data_start = (struct ggml_tensor **) (cgraph + 1);
-
+    // the size of the hash table is doubled since it needs to hold both nodes and leafs
     size_t hash_size = ggml_hash_size(size * 2);
-    struct ggml_tensor ** nodes_ptr = data_start;
-    struct ggml_tensor ** leafs_ptr = nodes_ptr + size;
-    struct ggml_tensor ** hash_keys_ptr = leafs_ptr + size;
-    struct ggml_tensor ** grads_ptr = grads ? hash_keys_ptr + hash_size : NULL;
 
-    // check that we allocated the correct amount of memory
-    assert(obj_size == (size_t) (
-        (grads ? (char *)(grads_ptr + size) : (char *)(hash_keys_ptr + hash_size)) - (char *)cgraph));
+    void * p = cgraph + 1;
 
-    memset(hash_keys_ptr, 0, hash_size * sizeof(struct ggml_tensor *));
+    struct ggml_tensor ** nodes_ptr = incr_ptr_aligned(&p, size * sizeof(struct ggml_tensor *), sizeof(struct ggml_tensor *));
+    struct ggml_tensor ** leafs_ptr = incr_ptr_aligned(&p, size * sizeof(struct ggml_tensor *), sizeof(struct ggml_tensor *));
+    struct ggml_tensor ** hash_keys_ptr = incr_ptr_aligned(&p, hash_size * sizeof(struct ggml_tensor *), sizeof(struct ggml_tensor *));
+    struct ggml_tensor ** grads_ptr = grads ? incr_ptr_aligned(&p, size * sizeof(struct ggml_tensor *), sizeof(struct ggml_tensor *)) : NULL;
+    ggml_bitset_t * hash_used = incr_ptr_aligned(&p, ggml_bitset_size(hash_size) * sizeof(ggml_bitset_t), sizeof(ggml_bitset_t));
+
+    // check that we allocated the correct amount of memory
+    assert(obj_size == (size_t)((char *)p - (char *)cgraph));
 
     *cgraph = (struct ggml_cgraph) {
         /*.size         =*/ size,
@@ -15995,13 +18988,12 @@ struct ggml_cgraph * ggml_new_graph_custom(struct ggml_context * ctx, size_t siz
         /*.nodes        =*/ nodes_ptr,
         /*.grads        =*/ grads_ptr,
         /*.leafs        =*/ leafs_ptr,
-        /*.hash_table   =*/ { hash_size, hash_keys_ptr },
+        /*.hash_table   =*/ { hash_size, hash_used, hash_keys_ptr },
         /*.order        =*/ GGML_CGRAPH_EVAL_ORDER_LEFT_TO_RIGHT,
-        /*.perf_runs    =*/ 0,
-        /*.perf_cycles  =*/ 0,
-        /*.perf_time_us =*/ 0,
     };
 
+    ggml_hash_set_reset(&cgraph->visited_hash_set);
+
     return cgraph;
 }
 
@@ -16017,11 +19009,8 @@ struct ggml_cgraph ggml_graph_view(struct ggml_cgraph * cgraph0, int i0, int i1)
         /*.nodes        =*/ cgraph0->nodes + i0,
         /*.grads        =*/ cgraph0->grads ? cgraph0->grads + i0 : NULL,
         /*.leafs        =*/ NULL,
-        /*.hash_table   =*/ { 0, NULL },
+        /*.hash_table   =*/ { 0, NULL, NULL },
         /*.order        =*/ cgraph0->order,
-        /*.perf_runs    =*/ 0,
-        /*.perf_cycles  =*/ 0,
-        /*.perf_time_us =*/ 0,
     };
 
     return cgraph;
@@ -16030,7 +19019,7 @@ struct ggml_cgraph ggml_graph_view(struct ggml_cgraph * cgraph0, int i0, int i1)
 void ggml_graph_cpy(struct ggml_cgraph * src, struct ggml_cgraph * dst) {
     GGML_ASSERT(dst->size >= src->n_leafs);
     GGML_ASSERT(dst->size >= src->n_nodes);
-    GGML_ASSERT(dst->visited_hash_table.size >= src->visited_hash_table.size);
+    GGML_ASSERT(dst->visited_hash_set.size >= src->visited_hash_set.size);
 
     dst->n_leafs = src->n_leafs;
     dst->n_nodes = src->n_nodes;
@@ -16051,9 +19040,10 @@ void ggml_graph_cpy(struct ggml_cgraph * src, struct ggml_cgraph * dst) {
         }
     }
 
-    for (size_t i = 0; i < src->visited_hash_table.size; ++i) {
-        if (src->visited_hash_table.keys[i]) {
-            ggml_hash_insert(dst->visited_hash_table, src->visited_hash_table.keys[i]);
+    for (size_t i = 0; i < src->visited_hash_set.size; ++i) {
+        // copy all hashset keys (tensors) that are in use
+        if (ggml_bitset_get(src->visited_hash_set.used, i)) {
+            ggml_hash_insert(&dst->visited_hash_set, src->visited_hash_set.keys[i]);
         }
     }
 }
@@ -16068,10 +19058,28 @@ void ggml_graph_reset(struct ggml_cgraph * cgraph) {
     GGML_ASSERT(cgraph->grads != NULL);
 
     for (int i = 0; i < cgraph->n_nodes; i++) {
-        struct ggml_tensor * grad = cgraph->grads[i];
+        struct ggml_tensor * node = cgraph->nodes[i];
+
+        // initial gradients of loss should be 1, 0 otherwise
+        if (node->grad) {
+            if (node->flags & GGML_TENSOR_FLAG_LOSS) {
+                GGML_ASSERT(node->grad->buffer);
+                GGML_ASSERT(node->type == GGML_TYPE_F32);
+                GGML_ASSERT(ggml_is_scalar(node));
+
+                const float onef = 1.0f;
+                ggml_backend_tensor_set(node->grad, &onef, 0, ggml_nbytes(node->grad));
+            } else {
+                ggml_set_zero(node->grad);
+            }
+        }
 
-        if (grad) {
-            ggml_set_zero(grad);
+        GGML_ASSERT(node);
+        if (node->op == GGML_OP_OPT_STEP_ADAMW) {
+            // set iteration to 1 and clear momenta
+            ggml_set_op_params_i32(node, 0, 1);
+            ggml_set_zero(node->src[2]);
+            ggml_set_zero(node->src[3]);
         }
     }
 }
@@ -16079,94 +19087,79 @@ void ggml_graph_reset(struct ggml_cgraph * cgraph) {
 void ggml_graph_clear(struct ggml_cgraph * cgraph) {
     cgraph->n_leafs = 0;
     cgraph->n_nodes = 0;
-    memset(cgraph->visited_hash_table.keys, 0, cgraph->visited_hash_table.size * sizeof(struct ggml_tensor *));
+    ggml_hash_set_reset(&cgraph->visited_hash_set);
 }
 
-//
-// thread data
-//
-// synchronization is done via busy loops
-// I tried using spin locks, but not sure how to use them correctly - the things I tried were slower than busy loops
-//
-
-#ifdef __APPLE__
-
-//#include <os/lock.h>
-//
-//typedef os_unfair_lock ggml_lock_t;
-//
-//#define ggml_lock_init(x)    UNUSED(x)
-//#define ggml_lock_destroy(x) UNUSED(x)
-//#define ggml_lock_lock       os_unfair_lock_lock
-//#define ggml_lock_unlock     os_unfair_lock_unlock
-//
-//#define GGML_LOCK_INITIALIZER OS_UNFAIR_LOCK_INIT
-
-typedef int ggml_lock_t;
-
-#define ggml_lock_init(x)    UNUSED(x)
-#define ggml_lock_destroy(x) UNUSED(x)
-#define ggml_lock_lock(x)    UNUSED(x)
-#define ggml_lock_unlock(x)  UNUSED(x)
-
-#define GGML_LOCK_INITIALIZER 0
-
-typedef pthread_t ggml_thread_t;
-
-#define ggml_thread_create pthread_create
-#define ggml_thread_join   pthread_join
-
-#else
-
-//typedef pthread_spinlock_t ggml_lock_t;
-
-//#define ggml_lock_init(x) pthread_spin_init(x, PTHREAD_PROCESS_PRIVATE)
-//#define ggml_lock_destroy pthread_spin_destroy
-//#define ggml_lock_lock    pthread_spin_lock
-//#define ggml_lock_unlock  pthread_spin_unlock
-
-typedef int ggml_lock_t;
+int ggml_graph_size(struct ggml_cgraph * cgraph) {
+    return cgraph->size;
+}
 
-#define ggml_lock_init(x)    UNUSED(x)
-#define ggml_lock_destroy(x) UNUSED(x)
-#if defined(__x86_64__) || (defined(_MSC_VER) && defined(_M_AMD64))
-#define ggml_lock_lock(x)    _mm_pause()
-#else
-#define ggml_lock_lock(x)    UNUSED(x)
-#endif
-#define ggml_lock_unlock(x)  UNUSED(x)
+struct ggml_tensor * ggml_graph_node(struct ggml_cgraph * cgraph, int i) {
+    if (i < 0) {
+        GGML_ASSERT(cgraph->n_nodes + i >= 0);
+        return cgraph->nodes[cgraph->n_nodes + i];
+    }
 
-#define GGML_LOCK_INITIALIZER 0
+    GGML_ASSERT(i < cgraph->n_nodes);
+    return cgraph->nodes[i];
+}
 
-typedef pthread_t ggml_thread_t;
+struct ggml_tensor ** ggml_graph_nodes(struct ggml_cgraph * cgraph) {
+    return cgraph->nodes;
+}
 
-#define ggml_thread_create pthread_create
-#define ggml_thread_join   pthread_join
+int ggml_graph_n_nodes(struct ggml_cgraph * cgraph) {
+    return cgraph->n_nodes;
+}
 
-#endif
+void ggml_graph_add_node(struct ggml_cgraph * cgraph, struct ggml_tensor * tensor) {
+    GGML_ASSERT(cgraph->size > cgraph->n_nodes);
+    cgraph->nodes[cgraph->n_nodes] = tensor;
+    cgraph->n_nodes++;
+}
 
 // Android's libc implementation "bionic" does not support setting affinity
-#if defined(__linux__) && !defined(__BIONIC__)
-static void set_numa_thread_affinity(int thread_n, int n_threads) {
+#if defined(__gnu_linux__)
+static void set_numa_thread_affinity(int thread_n) {
     if (!ggml_is_numa()) {
         return;
     }
 
-    // run thread on node_num thread_n / (threads per node)
-    const int node_num = thread_n / ((n_threads + g_state.numa.n_nodes - 1) / g_state.numa.n_nodes);
-    struct ggml_numa_node * node = &g_state.numa.nodes[node_num];
+    int node_num;
+    int rv;
     size_t setsize = CPU_ALLOC_SIZE(g_state.numa.total_cpus);
 
+    switch(g_state.numa.numa_strategy) {
+        case GGML_NUMA_STRATEGY_DISTRIBUTE:
+            // run thread on node_num thread_n / (threads per node)
+            node_num = thread_n % g_state.numa.n_nodes;
+            break;
+        case GGML_NUMA_STRATEGY_ISOLATE:
+            // run thread on current_node
+            node_num = g_state.numa.current_node;
+            break;
+        case GGML_NUMA_STRATEGY_NUMACTL:
+            // use the cpuset that numactl gave us
+            rv = pthread_setaffinity_np(pthread_self(), setsize, &g_state.numa.cpuset);
+            if (rv) {
+                fprintf(stderr, "warning: pthread_setaffinity_np() failed: %s\n",strerror(rv));
+            }
+            return;
+        default:
+            return;
+    }
+
+    struct ggml_numa_node * node = &g_state.numa.nodes[node_num];
+
     cpu_set_t * cpus = CPU_ALLOC(g_state.numa.total_cpus);
     CPU_ZERO_S(setsize, cpus);
     for (size_t i = 0; i < node->n_cpus; ++i) {
         CPU_SET_S(node->cpus[i], setsize, cpus);
     }
 
-    int rv = pthread_setaffinity_np(pthread_self(), setsize, cpus);
+    rv = pthread_setaffinity_np(pthread_self(), setsize, cpus);
     if (rv) {
-            Rprintf("warning: pthread_setaffinity_np() failed: %s\n",
-                    strerror(rv));
+            fprintf(stderr, "warning: pthread_setaffinity_np() failed: %s\n", strerror(rv));
     }
 
     CPU_FREE(cpus);
@@ -16187,8 +19180,7 @@ static void clear_numa_thread_affinity(void) {
 
     int rv = pthread_setaffinity_np(pthread_self(), setsize, cpus);
     if (rv) {
-        Rprintf("warning: pthread_setaffinity_np() failed: %s\n",
-            strerror(rv));
+        fprintf(stderr, "warning: pthread_setaffinity_np() failed: %s\n", strerror(rv));
     }
 
     CPU_FREE(cpus);
@@ -16196,48 +19188,23 @@ static void clear_numa_thread_affinity(void) {
 #else
 // TODO: Windows etc.
 // (the linux implementation may also work on BSD, someone should test)
-static void set_numa_thread_affinity(int thread_n, int n_threads) { UNUSED(thread_n); UNUSED(n_threads);  }
+static void set_numa_thread_affinity(int thread_n) { UNUSED(thread_n);  }
 static void clear_numa_thread_affinity(void) {}
 #endif
 
-struct ggml_compute_state_shared {
-    const struct ggml_cgraph * cgraph;
-    const struct ggml_cplan  * cplan;
-
-    int64_t perf_node_start_cycles;
-    int64_t perf_node_start_time_us;
-
-    const int n_threads;
-
-    // synchronization primitives
-    atomic_int n_active; // num active threads
-    atomic_int node_n;   // active graph node
-
-    bool (*abort_callback)(void * data); // abort ggml_graph_compute when true
-    void * abort_callback_data;
-};
-
-struct ggml_compute_state {
-    ggml_thread_t thrd;
-    int ith;
-    struct ggml_compute_state_shared * shared;
-};
-
-static void ggml_graph_compute_perf_stats_node(struct ggml_tensor * node, const struct ggml_compute_state_shared * st) {
-    int64_t cycles_cur  = ggml_perf_cycles()  - st->perf_node_start_cycles;
-    int64_t time_us_cur = ggml_perf_time_us() - st->perf_node_start_time_us;
-
-    node->perf_runs++;
-    node->perf_cycles  += cycles_cur;
-    node->perf_time_us += time_us_cur;
-}
-
 static int ggml_get_n_tasks(struct ggml_tensor * node, int n_threads) {
     int n_tasks = 0;
 
+    if (ggml_is_empty(node)) {
+        // no need to multi-thread a no-op
+        n_tasks = 1;
+        return n_tasks;
+    }
+
     switch (node->op) {
         case GGML_OP_CPY:
         case GGML_OP_DUP:
+        case GGML_OP_CONT:
         case GGML_OP_ADD:
         case GGML_OP_ADD1:
         case GGML_OP_ACC:
@@ -16248,10 +19215,19 @@ static int ggml_get_n_tasks(struct ggml_tensor * node, int n_threads) {
         case GGML_OP_SQR:
         case GGML_OP_SQRT:
         case GGML_OP_LOG:
+        case GGML_OP_SIN:
+        case GGML_OP_COS:
         case GGML_OP_SUM:
         case GGML_OP_SUM_ROWS:
         case GGML_OP_MEAN:
         case GGML_OP_ARGMAX:
+            {
+                n_tasks = 1;
+            } break;
+        case GGML_OP_COUNT_EQUAL:
+            {
+                n_tasks = n_threads;
+            } break;
         case GGML_OP_REPEAT:
         case GGML_OP_REPEAT_BACK:
         case GGML_OP_LEAKY_RELU:
@@ -16267,6 +19243,10 @@ static int ggml_get_n_tasks(struct ggml_tensor * node, int n_threads) {
                 case GGML_UNARY_OP_TANH:
                 case GGML_UNARY_OP_ELU:
                 case GGML_UNARY_OP_RELU:
+                case GGML_UNARY_OP_SIGMOID:
+                case GGML_UNARY_OP_HARDSWISH:
+                case GGML_UNARY_OP_HARDSIGMOID:
+                case GGML_UNARY_OP_EXP:
                     {
                         n_tasks = 1;
                     } break;
@@ -16278,65 +19258,36 @@ static int ggml_get_n_tasks(struct ggml_tensor * node, int n_threads) {
                         n_tasks = n_threads;
                     } break;
                 default:
-                    GGML_ASSERT(false);
+                    GGML_ABORT("fatal error");
             }
             break;
         case GGML_OP_SILU_BACK:
         case GGML_OP_MUL:
         case GGML_OP_DIV:
         case GGML_OP_NORM:
-        case GGML_OP_RMS_NORM:
-        case GGML_OP_RMS_NORM_BACK:
-        case GGML_OP_GROUP_NORM:
-        case GGML_OP_CONCAT:
-            {
-                n_tasks = n_threads;
-            } break;
-        case GGML_OP_MUL_MAT:
-            {
-                n_tasks = n_threads;
-
-                // TODO: use different scheduling for different matrix sizes
-                //const int nr0 = ggml_nrows(node->src[0]);
-                //const int nr1 = ggml_nrows(node->src[1]);
-
-                //n_tasks = MIN(n_threads, MAX(1, nr0/128));
-                //printf("nr0 = %8d, nr1 = %8d, nr0*nr1 = %8d, n_tasks%d\n", nr0, nr1, nr0*nr1, n_tasks);
-
-#if defined(GGML_USE_CUBLAS)
-                if (ggml_cuda_can_mul_mat(node->src[0], node->src[1], node)) {
-                    n_tasks = 1; // TODO: this actually is doing nothing
-                                 //       the threads are still spinning
-                }
-#elif defined(GGML_USE_CLBLAST)
-                if (ggml_cl_can_mul_mat(node->src[0], node->src[1], node)) {
-                    n_tasks = 1; // TODO: this actually is doing nothing
-                                 //       the threads are still spinning
-                }
-#endif
-#if defined(GGML_USE_ACCELERATE) || defined(GGML_USE_OPENBLAS)
-                if (ggml_compute_forward_mul_mat_use_blas(node->src[0], node->src[1], node)) {
-                    n_tasks = 1; // TODO: this actually is doing nothing
-                                 //       the threads are still spinning
-                }
-#endif
-            } break;
+        case GGML_OP_RMS_NORM:
+        case GGML_OP_RMS_NORM_BACK:
+        case GGML_OP_GROUP_NORM:
+        case GGML_OP_CONCAT:
+        case GGML_OP_MUL_MAT:
         case GGML_OP_MUL_MAT_ID:
+        case GGML_OP_OUT_PROD:
             {
                 n_tasks = n_threads;
             } break;
-        case GGML_OP_OUT_PROD:
+        case GGML_OP_GET_ROWS:
             {
-                n_tasks = n_threads;
+                // FIXME: get_rows can use additional threads, but the cost of launching additional threads
+                // decreases performance with GPU offloading
+                //n_tasks = n_threads;
+                n_tasks = 1;
             } break;
         case GGML_OP_SCALE:
         case GGML_OP_SET:
-        case GGML_OP_CONT:
         case GGML_OP_RESHAPE:
         case GGML_OP_VIEW:
         case GGML_OP_PERMUTE:
         case GGML_OP_TRANSPOSE:
-        case GGML_OP_GET_ROWS:
         case GGML_OP_GET_ROWS_BACK:
         case GGML_OP_DIAG:
             {
@@ -16351,62 +19302,43 @@ static int ggml_get_n_tasks(struct ggml_tensor * node, int n_threads) {
             {
                 n_tasks = n_threads;
             } break;
-        case GGML_OP_ALIBI:
-            {
-                n_tasks = 1; //TODO
-            } break;
         case GGML_OP_CLAMP:
             {
                 n_tasks = 1; //TODO
             } break;
         case GGML_OP_SOFT_MAX:
             {
-                n_tasks = MIN(MIN(4, n_threads), ggml_nrows(node->src[0]));
-            } break;
-        case GGML_OP_CONV_TRANSPOSE_1D:
-            {
-                n_tasks = n_threads;
+                n_tasks = MIN(n_threads, ggml_nrows(node->src[0]));
             } break;
         case GGML_OP_IM2COL:
-            {
-                n_tasks = n_threads;
-            } break;
+        case GGML_OP_IM2COL_BACK:
+        case GGML_OP_CONV_TRANSPOSE_1D:
         case GGML_OP_CONV_TRANSPOSE_2D:
             {
                 n_tasks = n_threads;
             } break;
         case GGML_OP_POOL_1D:
         case GGML_OP_POOL_2D:
+        case GGML_OP_POOL_2D_BACK:
             {
                 n_tasks = 1;
             } break;
         case GGML_OP_UPSCALE:
-            {
-                n_tasks = n_threads;
-            } break;
         case GGML_OP_PAD:
-            {
-                n_tasks = n_threads;
-            } break;
+        case GGML_OP_ARANGE:
+        case GGML_OP_TIMESTEP_EMBEDDING:
         case GGML_OP_ARGSORT:
-            {
-                n_tasks = n_threads;
-            } break;
-        case GGML_OP_FLASH_ATTN:
-            {
-                n_tasks = n_threads;
-            } break;
-        case GGML_OP_FLASH_FF:
-            {
-                n_tasks = n_threads;
-            } break;
+        case GGML_OP_FLASH_ATTN_EXT:
         case GGML_OP_FLASH_ATTN_BACK:
+        case GGML_OP_SSM_CONV:
+        case GGML_OP_SSM_SCAN:
             {
                 n_tasks = n_threads;
             } break;
         case GGML_OP_WIN_PART:
         case GGML_OP_WIN_UNPART:
         case GGML_OP_GET_REL_POS:
+        case GGML_OP_RWKV_WKV:
         case GGML_OP_MAP_UNARY:
         case GGML_OP_MAP_BINARY:
         case GGML_OP_MAP_CUSTOM1_F32:
@@ -16417,36 +19349,37 @@ static int ggml_get_n_tasks(struct ggml_tensor * node, int n_threads) {
             } break;
         case GGML_OP_MAP_CUSTOM1:
             {
-                struct ggml_map_custom1_op_params * p = (struct ggml_map_custom1_op_params *) node->op_params;
-                if (p->n_tasks == GGML_N_TASKS_MAX) {
+                struct ggml_map_custom1_op_params p;
+                memcpy(&p, node->op_params, sizeof(p));
+                if (p.n_tasks == GGML_N_TASKS_MAX) {
                     n_tasks = n_threads;
                 } else {
-                    n_tasks = MIN(p->n_tasks, n_threads);
+                    n_tasks = MIN(p.n_tasks, n_threads);
                 }
             } break;
         case GGML_OP_MAP_CUSTOM2:
             {
-                struct ggml_map_custom2_op_params * p = (struct ggml_map_custom2_op_params *) node->op_params;
-                if (p->n_tasks == GGML_N_TASKS_MAX) {
+                struct ggml_map_custom2_op_params p;
+                memcpy(&p, node->op_params, sizeof(p));
+                if (p.n_tasks == GGML_N_TASKS_MAX) {
                     n_tasks = n_threads;
                 } else {
-                    n_tasks = MIN(p->n_tasks, n_threads);
+                    n_tasks = MIN(p.n_tasks, n_threads);
                 }
             } break;
         case GGML_OP_MAP_CUSTOM3:
             {
-                struct ggml_map_custom3_op_params * p = (struct ggml_map_custom3_op_params *) node->op_params;
-                if (p->n_tasks == GGML_N_TASKS_MAX) {
+                struct ggml_map_custom3_op_params p;
+                memcpy(&p, node->op_params, sizeof(p));
+                if (p.n_tasks == GGML_N_TASKS_MAX) {
                     n_tasks = n_threads;
                 } else {
-                    n_tasks = MIN(p->n_tasks, n_threads);
+                    n_tasks = MIN(p.n_tasks, n_threads);
                 }
             } break;
         case GGML_OP_CROSS_ENTROPY_LOSS:
-            {
-                n_tasks = n_threads;
-            } break;
         case GGML_OP_CROSS_ENTROPY_LOSS_BACK:
+        case GGML_OP_OPT_STEP_ADAMW:
             {
                 n_tasks = n_threads;
             } break;
@@ -16456,18 +19389,18 @@ static int ggml_get_n_tasks(struct ggml_tensor * node, int n_threads) {
             } break;
         case GGML_OP_COUNT:
             {
-                GGML_ASSERT(false);
-            } break;
+                GGML_ABORT("fatal error");
+            }
         default:
             {
-                Rprintf("%s: op not implemented: ", __func__);
+                fprintf(stderr, "%s: op not implemented: ", __func__);
                 if (node->op < GGML_OP_COUNT) {
-                    Rprintf("%s\n", ggml_op_name(node->op));
+                    fprintf(stderr, "%s\n", ggml_op_name(node->op));
                 } else {
-                    Rprintf("%d\n", node->op);
+                    fprintf(stderr, "%d\n", node->op);
                 }
-                GGML_ASSERT(false);
-            } break;
+                GGML_ABORT("fatal error");
+            }
     }
 
     assert(n_tasks > 0);
@@ -16475,128 +19408,281 @@ static int ggml_get_n_tasks(struct ggml_tensor * node, int n_threads) {
     return n_tasks;
 }
 
-static thread_ret_t ggml_graph_compute_thread(void * data) {
-    struct ggml_compute_state * state = (struct ggml_compute_state *) data;
+static thread_ret_t ggml_graph_compute_secondary_thread(void* data);
 
-    const struct ggml_cgraph * cgraph = state->shared->cgraph;
-    const struct ggml_cplan  * cplan  = state->shared->cplan;
+#if defined(_WIN32)
+#include "windows.h"
+
+// TODO: support > 64 CPUs
+bool ggml_thread_apply_affinity(bool * mask) {
+    HANDLE    h = GetCurrentThread();
+    uint64_t  bitmask = 0ULL;
+
+    assert(GGML_MAX_N_THREADS >= 64);
+
+    for (int32_t i = 0; i < 8; i++) {
+        int32_t idx = i * 8;
+        uint8_t val = 0;
+        val |= mask[idx + 0] << 0;
+        val |= mask[idx + 1] << 1;
+        val |= mask[idx + 2] << 2;
+        val |= mask[idx + 3] << 3;
+        val |= mask[idx + 4] << 4;
+        val |= mask[idx + 5] << 5;
+        val |= mask[idx + 6] << 6;
+        val |= mask[idx + 7] << 7;
+        bitmask |= (uint64_t)val << idx;
+    }
+
+    for (int32_t i = 64; i < GGML_MAX_N_THREADS; i++) {
+        if (mask[i]) {
+            fprintf(stderr, "warn: setting thread-affinity for > 64 CPUs isn't supported on windows!\n");
+            break;
+        }
+    }
 
-    const int   n_threads   = state->shared->n_threads;
+    DWORD_PTR m = (DWORD_PTR)bitmask;
 
-    set_numa_thread_affinity(state->ith, n_threads);
+    m = SetThreadAffinityMask(h, m);
 
-    int node_n = -1;
+    return m != 0;
+}
 
-    while (true) {
-        if (cplan->abort_callback && cplan->abort_callback(cplan->abort_callback_data)) {
-            state->shared->node_n += 1;
-            return (thread_ret_t) GGML_EXIT_ABORTED;
-        }
-        if (atomic_fetch_sub(&state->shared->n_active, 1) == 1) {
-            // all other threads are finished and spinning
-            // do finalize and init here so we don't have synchronize again
-            struct ggml_compute_params params = {
-                /*.type  =*/ GGML_TASK_FINALIZE,
-                /*.ith   =*/ 0,
-                /*.nth   =*/ 0,
-                /*.wsize =*/ cplan->work_size,
-                /*.wdata =*/ cplan->work_data,
-            };
+static bool ggml_thread_apply_priority(int32_t prio) {
+    // Note that on Windows the Process Priority Class must be updated in order to set Thread priority.
+    // This is up to the applications.
+    DWORD p = THREAD_PRIORITY_NORMAL;
+    switch (prio) {
+        case GGML_SCHED_PRIO_NORMAL:   p = THREAD_PRIORITY_NORMAL;        break;
+        case GGML_SCHED_PRIO_MEDIUM:   p = THREAD_PRIORITY_ABOVE_NORMAL;  break;
+        case GGML_SCHED_PRIO_HIGH:     p = THREAD_PRIORITY_HIGHEST;       break;
+        case GGML_SCHED_PRIO_REALTIME: p = THREAD_PRIORITY_TIME_CRITICAL; break;
+    }
 
-            if (node_n != -1) {
-                /* FINALIZE */
-                struct ggml_tensor * node = cgraph->nodes[node_n];
-                if (GGML_OP_HAS_FINALIZE[node->op]) {
-                    params.nth = ggml_get_n_tasks(node, n_threads);
-                    ggml_compute_forward(&params, node);
-                }
-                ggml_graph_compute_perf_stats_node(node, state->shared);
-            }
+    if (prio == GGML_SCHED_PRIO_NORMAL) {
+        // Keep inherited policy/priority
+        return true;
+    }
 
-            // distribute new work or execute it direct if 1T
-            while (++node_n < cgraph->n_nodes) {
-                GGML_PRINT_DEBUG_5("%s: %d/%d\n", __func__, node_n, cgraph->n_nodes);
+    if (!SetThreadPriority(GetCurrentThread(), p)) {
+        fprintf(stderr, "warn: failed to set thread priority %d : (%d)\n", prio, (int) GetLastError());
+        return false;
+    }
 
-                struct ggml_tensor * node = cgraph->nodes[node_n];
-                const int n_tasks = ggml_get_n_tasks(node, n_threads);
+    return true;
+}
 
-                state->shared->perf_node_start_cycles  = ggml_perf_cycles();
-                state->shared->perf_node_start_time_us = ggml_perf_time_us();
+#elif defined(__APPLE__)
+#include <sys/types.h>
+#include <sys/resource.h>
 
-                params.nth = n_tasks;
+static bool ggml_thread_apply_affinity(const bool * mask) {
+    // Not supported on Apple platforms
+    UNUSED(mask);
+    return true;
+}
 
-                /* INIT */
-                if (GGML_OP_HAS_INIT[node->op]) {
-                    params.type = GGML_TASK_INIT;
-                    ggml_compute_forward(&params, node);
-                }
+static bool ggml_thread_apply_priority(int32_t prio) {
+    struct sched_param p;
+    int32_t policy = SCHED_OTHER;
+    switch (prio) {
+        case GGML_SCHED_PRIO_NORMAL:   policy = SCHED_OTHER; p.sched_priority = 0;  break;
+        case GGML_SCHED_PRIO_MEDIUM:   policy = SCHED_FIFO;  p.sched_priority = 40; break;
+        case GGML_SCHED_PRIO_HIGH:     policy = SCHED_FIFO;  p.sched_priority = 80; break;
+        case GGML_SCHED_PRIO_REALTIME: policy = SCHED_FIFO;  p.sched_priority = 90; break;
+    }
 
-                if (n_tasks == 1) {
-                    // TODO: maybe push node_n to the atomic but if other threads see n_tasks is 1,
-                    // they do something more efficient than spinning (?)
-                    params.type = GGML_TASK_COMPUTE;
-                    ggml_compute_forward(&params, node);
+    if (prio == GGML_SCHED_PRIO_NORMAL) {
+        // Keep inherited policy/priority
+        return true;
+    }
 
-                    if (GGML_OP_HAS_FINALIZE[node->op]) {
-                        params.type = GGML_TASK_FINALIZE;
-                        ggml_compute_forward(&params, node);
-                    }
+    int32_t err = pthread_setschedparam(pthread_self(), policy, &p);
+    if (err != 0) {
+        fprintf(stderr, "warn: failed to set thread priority %d : %s (%d)\n", prio, strerror(err), err);
+        return false;
+    }
 
-                    ggml_graph_compute_perf_stats_node(node, state->shared);
-                } else {
-                    break;
-                }
+    return true;
+}
 
-                if (cplan->abort_callback && cplan->abort_callback(cplan->abort_callback_data)) {
-                    break;
-                }
-            }
+#elif defined(__gnu_linux__)
+// TODO: this may not work on BSD, to be verified
 
-            atomic_store(&state->shared->n_active, n_threads);
-            atomic_store(&state->shared->node_n,   node_n);
-        } else {
-            // wait for other threads to finish
-            const int last = node_n;
-            while (true) {
-                // TODO: this sched_yield can have significant impact on the performance - either positive or negative
-                //       depending on the workload and the operating system.
-                //       since it is not clear what is the best approach, it should potentially become user-configurable
-                //       ref: https://github.com/ggerganov/ggml/issues/291
-#if defined(GGML_USE_ACCELERATE) || defined(GGML_USE_OPENBLAS)
-                sched_yield();
-#endif
+static bool ggml_thread_apply_affinity(const bool * mask) {
+    cpu_set_t cpuset;
+    int err;
 
-                node_n = atomic_load(&state->shared->node_n);
-                if (node_n != last) break;
-            };
+    CPU_ZERO(&cpuset);
+
+    for (uint32_t i = 0; i < GGML_MAX_N_THREADS; i++) {
+        if (mask[i]) {
+            GGML_PRINT_DEBUG("Thread %lx: adding %d to cpuset\n", pthread_self(), i);
+            CPU_SET(i, &cpuset);
         }
+    }
+
+#ifdef __ANDROID__
+    err = sched_setaffinity(0, sizeof(cpuset), &cpuset);
+    if (err < 0) {
+        err = errno;
+    }
+#else
+    err = pthread_setaffinity_np(pthread_self(), sizeof(cpuset), &cpuset);
+#endif
+    if (err != 0) {
+        fprintf(stderr, "warn: failed to set affinity mask 0x%llx : %s (%d)\n", (unsigned long long)mask, strerror(err), err);
+        return false;
+    }
 
-        // check if we should stop
-        if (node_n >= cgraph->n_nodes) break;
+    return true;
+}
 
-        /* COMPUTE */
-        struct ggml_tensor * node = cgraph->nodes[node_n];
-        const int n_tasks = ggml_get_n_tasks(node, n_threads);
+static bool ggml_thread_apply_priority(int32_t prio) {
+    struct sched_param p;
+    int32_t policy = SCHED_OTHER;
+    switch (prio) {
+        case GGML_SCHED_PRIO_NORMAL:   policy = SCHED_OTHER; p.sched_priority = 0;  break;
+        case GGML_SCHED_PRIO_MEDIUM:   policy = SCHED_FIFO;  p.sched_priority = 40; break;
+        case GGML_SCHED_PRIO_HIGH:     policy = SCHED_FIFO;  p.sched_priority = 80; break;
+        case GGML_SCHED_PRIO_REALTIME: policy = SCHED_FIFO;  p.sched_priority = 90; break;
+    }
 
-        struct ggml_compute_params params = {
-            /*.type  =*/ GGML_TASK_COMPUTE,
-            /*.ith   =*/ state->ith,
-            /*.nth   =*/ n_tasks,
-            /*.wsize =*/ cplan->work_size,
-            /*.wdata =*/ cplan->work_data,
-        };
+    if (prio == GGML_SCHED_PRIO_NORMAL) {
+        // Keep inherited policy/priority
+        return true;
+    }
+
+    int32_t err = pthread_setschedparam(pthread_self(), policy, &p);
+    if (err != 0) {
+        fprintf(stderr, "warn: failed to set thread priority %d : %s (%d)\n", prio, strerror(err), err);
+        return false;
+    }
+
+    return true;
+}
+
+#else // unsupported platforms
 
-        if (state->ith < n_tasks) {
-            ggml_compute_forward(&params, node);
+static bool ggml_thread_apply_affinity(const bool * mask) {
+    UNUSED(mask);
+    return true;
+}
+
+static bool ggml_thread_apply_priority(int32_t prio) {
+    UNUSED(prio);
+    return true;
+}
+
+#endif
+
+static bool ggml_thread_cpumask_is_valid(const bool * mask) {
+    for (int i = 0; i < GGML_MAX_N_THREADS; i++) {
+        if (mask[i]) { return true; }
+    }
+    return false;
+}
+
+static void ggml_thread_cpumask_next(const bool * global_mask, bool * local_mask, bool strict, int32_t* iter) {
+    if (!strict) {
+        memcpy(local_mask, global_mask, GGML_MAX_N_THREADS);
+        return;
+    } else {
+        memset(local_mask, 0, GGML_MAX_N_THREADS);
+        int32_t base_idx = *iter;
+        for (int32_t i = 0; i < GGML_MAX_N_THREADS; i++) {
+            int32_t idx = base_idx + i;
+            if (idx >= GGML_MAX_N_THREADS) {
+                // Just a cheaper modulo
+                idx -= GGML_MAX_N_THREADS;
+            }
+            if (global_mask[idx]) {
+                local_mask[idx] = 1;
+                *iter = idx + 1;
+                return;
+            }
         }
     }
+}
+
+void ggml_threadpool_free(struct ggml_threadpool* threadpool) {
+    if (!threadpool) return;
+
+#ifndef GGML_USE_OPENMP
+    struct ggml_compute_state* workers = threadpool->workers;
+    const int n_threads = threadpool->n_threads_max;
+
+    ggml_mutex_lock(&threadpool->mutex);
+
+    threadpool->stop = true;
+    threadpool->pause = false;
+
+    ggml_cond_broadcast(&threadpool->cond);
+    ggml_mutex_unlock(&threadpool->mutex);
+
+    for (int j = 1; j < n_threads; j++) {
+        int32_t rc = ggml_thread_join(workers[j].thrd, NULL);
+        GGML_ASSERT(rc == GGML_EXIT_SUCCESS || rc == GGML_EXIT_ABORTED);
+        UNUSED(rc);
+    }
+
+    ggml_mutex_destroy(&threadpool->mutex);
+    ggml_cond_destroy(&threadpool->cond);
+#endif // GGML_USE_OPENMP
+
+    GGML_ALIGNED_FREE(threadpool->workers);
+    GGML_ALIGNED_FREE(threadpool);
+}
+
+#ifndef GGML_USE_OPENMP
+// pause/resume must be called under mutex
+static void ggml_threadpool_pause_locked(struct ggml_threadpool * threadpool) {
+    GGML_PRINT_DEBUG("Pausing threadpool\n");
+    threadpool->pause = true;
+    ggml_cond_broadcast(&threadpool->cond);
+}
+
+static void ggml_threadpool_resume_locked(struct ggml_threadpool * threadpool) {
+    GGML_PRINT_DEBUG("Resuming threadpool\n");
+    threadpool->pause = false;
+    ggml_cond_broadcast(&threadpool->cond);
+}
+#endif
+
+void ggml_threadpool_pause(struct ggml_threadpool * threadpool) {
+#ifndef GGML_USE_OPENMP
+    ggml_mutex_lock(&threadpool->mutex);
+    if (!threadpool->pause) {
+       ggml_threadpool_pause_locked(threadpool);
+    }
+    ggml_mutex_unlock(&threadpool->mutex);
+#else
+    UNUSED(threadpool);
+#endif
+}
 
-    return GGML_EXIT_SUCCESS;
+void ggml_threadpool_resume(struct ggml_threadpool * threadpool) {
+#ifndef GGML_USE_OPENMP
+    ggml_mutex_lock(&threadpool->mutex);
+    if (threadpool->pause) {
+       ggml_threadpool_resume_locked(threadpool);
+    }
+    ggml_mutex_unlock(&threadpool->mutex);
+#else
+    UNUSED(threadpool);
+#endif
 }
 
-struct ggml_cplan ggml_graph_plan(struct ggml_cgraph * cgraph, int n_threads) {
+struct ggml_cplan ggml_graph_plan(
+          const struct ggml_cgraph * cgraph,
+                               int   n_threads,
+            struct ggml_threadpool * threadpool) {
+
+    if (threadpool == NULL) {
+        GGML_PRINT_DEBUG("Threadpool is not specified. Will create a disposable threadpool : n_threads %d\n", n_threads);
+    }
     if (n_threads <= 0) {
-        n_threads = GGML_DEFAULT_N_THREADS;
+        n_threads = threadpool ? threadpool->n_threads_max : GGML_DEFAULT_N_THREADS;
     }
 
     size_t work_size = 0;
@@ -16604,19 +19690,26 @@ struct ggml_cplan ggml_graph_plan(struct ggml_cgraph * cgraph, int n_threads) {
     struct ggml_cplan cplan;
     memset(&cplan, 0, sizeof(struct ggml_cplan));
 
+    int max_tasks = 1;
+
     // thread scheduling for the different operations + work buffer size estimation
     for (int i = 0; i < cgraph->n_nodes; i++) {
         struct ggml_tensor * node = cgraph->nodes[i];
 
         const int n_tasks = ggml_get_n_tasks(node, n_threads);
 
+        max_tasks = MAX(max_tasks, n_tasks);
+
         size_t cur = 0;
 
         switch (node->op) {
             case GGML_OP_CPY:
             case GGML_OP_DUP:
                 {
-                    if (ggml_is_quantized(node->type)) {
+                    if (ggml_is_quantized(node->type) ||
+                        // F16 -> BF16 and BF16 -> F16 copies go through intermediate F32
+                        (node->src[0]->type == GGML_TYPE_F16  && node->src[1] && node->src[1]->type == GGML_TYPE_BF16) ||
+                        (node->src[0]->type == GGML_TYPE_BF16 && node->src[1] && node->src[1]->type == GGML_TYPE_F16)) {
                         cur = ggml_type_size(GGML_TYPE_F32) * node->ne[0] * n_tasks;
                     }
                 } break;
@@ -16633,39 +19726,31 @@ struct ggml_cplan ggml_graph_plan(struct ggml_cgraph * cgraph, int n_threads) {
                         cur = ggml_type_size(GGML_TYPE_F32) * node->src[1]->ne[0] * n_tasks;
                     }
                 } break;
+            case GGML_OP_COUNT_EQUAL:
+                {
+                    cur = ggml_type_size(node->type)*n_tasks;
+                } break;
             case GGML_OP_MUL_MAT:
                 {
                     const enum ggml_type vec_dot_type = type_traits[node->src[0]->type].vec_dot_type;
 
-#if defined(GGML_USE_CLBLAST)
-                    if (ggml_cl_can_mul_mat(node->src[0], node->src[1], node)) {
-                        cur = ggml_cl_mul_mat_get_wsize(node->src[0], node->src[1], node);
-                    } else
-#endif
-#if defined(GGML_USE_ACCELERATE) || defined(GGML_USE_OPENBLAS)
-                    if (ggml_compute_forward_mul_mat_use_blas(node->src[0], node->src[1], node)) {
-                        if (node->src[0]->type != GGML_TYPE_F32) {
-                            // here we need memory just for single 2D matrix from src0
-                            cur = ggml_type_size(GGML_TYPE_F32)*(node->src[0]->ne[0]*node->src[0]->ne[1]);
-                        }
-                    } else
-#endif
                     if (node->src[1]->type != vec_dot_type) {
                         cur = ggml_row_size(vec_dot_type, ggml_nelements(node->src[1]));
                     }
                 } break;
             case GGML_OP_MUL_MAT_ID:
                 {
-                    const struct ggml_tensor * src0 = node->src[2];
+                    cur = 0;
+                    const struct ggml_tensor * src0 = node->src[0];
                     const struct ggml_tensor * src1 = node->src[1];
                     const enum ggml_type vec_dot_type = type_traits[src0->type].vec_dot_type;
                     if (src1->type != vec_dot_type) {
-                        cur = ggml_row_size(vec_dot_type, ggml_nelements(src1));
+                        cur += ggml_row_size(vec_dot_type, ggml_nelements(src1));
                     }
-                    const int n_as = ggml_get_op_params_i32(node, 1);
-                    cur = GGML_PAD(cur, sizeof(int64_t));        // align
+                    const int n_as = src0->ne[2];
+                    cur += GGML_PAD(cur, sizeof(int64_t));       // align
                     cur += n_as * sizeof(int64_t);               // matrix_row_counts
-                    cur += n_as * src1->ne[1] * sizeof(int64_t); // matrix_rows
+                    cur += n_as * src1->ne[2] * sizeof(int64_t); // matrix_rows
                 } break;
             case GGML_OP_OUT_PROD:
                 {
@@ -16674,6 +19759,7 @@ struct ggml_cplan ggml_graph_plan(struct ggml_cgraph * cgraph, int n_threads) {
                     }
                 } break;
             case GGML_OP_SOFT_MAX:
+            case GGML_OP_ROPE:
                 {
                     cur = ggml_type_size(GGML_TYPE_F32) * node->ne[0] * n_tasks;
                 } break;
@@ -16690,7 +19776,8 @@ struct ggml_cplan ggml_graph_plan(struct ggml_cgraph * cgraph, int n_threads) {
                     const int64_t ne10 = node->src[1]->ne[0];  // L
                     const int64_t ne11 = node->src[1]->ne[1];  // Cin
 
-                    if (node->src[0]->type == GGML_TYPE_F16 &&
+                    if ((node->src[0]->type == GGML_TYPE_F16 ||
+                         node->src[0]->type == GGML_TYPE_BF16) &&
                         node->src[1]->type == GGML_TYPE_F32) {
                         cur += sizeof(ggml_fp16_t)*ne00*ne01*ne02;
                         cur += sizeof(ggml_fp16_t)*ne10*ne11;
@@ -16699,7 +19786,7 @@ struct ggml_cplan ggml_graph_plan(struct ggml_cgraph * cgraph, int n_threads) {
                         cur += sizeof(float)*ne00*ne01*ne02;
                         cur += sizeof(float)*ne10*ne11;
                     } else {
-                        GGML_ASSERT(false);
+                        GGML_ABORT("fatal error");
                     }
                 } break;
             case GGML_OP_CONV_TRANSPOSE_2D:
@@ -16716,27 +19803,11 @@ struct ggml_cplan ggml_graph_plan(struct ggml_cgraph * cgraph, int n_threads) {
                     cur += sizeof(ggml_fp16_t)*ne00*ne01*ne02*ne03;
                     cur += sizeof(ggml_fp16_t)*ne10*ne11*ne12;
                 } break;
-            case GGML_OP_FLASH_ATTN:
+            case GGML_OP_FLASH_ATTN_EXT:
                 {
-                    const int64_t ne11 = ggml_up(node->src[1]->ne[1], GGML_SOFT_MAX_UNROLL);
+                    const int64_t ne00 = node->src[0]->ne[0]; // D
 
-                    if (node->src[1]->type == GGML_TYPE_F32) {
-                        cur  = sizeof(float)*ne11*n_tasks; // TODO: this can become (n_tasks-1)
-                        cur += sizeof(float)*ne11*n_tasks; // this is overestimated by x2
-                    } else if (node->src[1]->type == GGML_TYPE_F16) {
-                        cur  = sizeof(float)*ne11*n_tasks; // TODO: this can become (n_tasks-1)
-                        cur += sizeof(float)*ne11*n_tasks; // this is overestimated by x2
-                    }
-                } break;
-            case GGML_OP_FLASH_FF:
-                {
-                    if (node->src[1]->type == GGML_TYPE_F32) {
-                        cur  = sizeof(float)*node->src[1]->ne[1]*n_tasks; // TODO: this can become (n_tasks-1)
-                        cur += sizeof(float)*node->src[1]->ne[1]*n_tasks; // this is overestimated by x2
-                    } else if (node->src[1]->type == GGML_TYPE_F16) {
-                        cur  = sizeof(float)*node->src[1]->ne[1]*n_tasks; // TODO: this can become (n_tasks-1)
-                        cur += sizeof(float)*node->src[1]->ne[1]*n_tasks; // this is overestimated by x2
-                    }
+                    cur = 3*sizeof(float)*ne00*n_tasks; // 3x head size/thread
                 } break;
             case GGML_OP_FLASH_ATTN_BACK:
                 {
@@ -16749,6 +19820,9 @@ struct ggml_cplan ggml_graph_plan(struct ggml_cgraph * cgraph, int n_threads) {
                     } else if (node->src[1]->type == GGML_TYPE_F16) {
                         cur  = sizeof(float)*mxDn*n_tasks; // TODO: this can become (n_tasks-1)
                         cur += sizeof(float)*mxDn*n_tasks; // this is overestimated by x2
+                    } else if (node->src[1]->type == GGML_TYPE_BF16) {
+                        cur  = sizeof(float)*mxDn*n_tasks; // TODO: this can become (n_tasks-1)
+                        cur += sizeof(float)*mxDn*n_tasks; // this is overestimated by x2
                     }
                 } break;
 
@@ -16758,8 +19832,8 @@ struct ggml_cplan ggml_graph_plan(struct ggml_cgraph * cgraph, int n_threads) {
                 } break;
             case GGML_OP_COUNT:
                 {
-                    GGML_ASSERT(false);
-                } break;
+                    GGML_ABORT("fatal error");
+                }
             default:
                 break;
         }
@@ -16768,104 +19842,368 @@ struct ggml_cplan ggml_graph_plan(struct ggml_cgraph * cgraph, int n_threads) {
     }
 
     if (work_size > 0) {
-        work_size += CACHE_LINE_SIZE*(n_threads - 1);
+        work_size += CACHE_LINE_SIZE*(n_threads);
     }
 
-    cplan.n_threads = n_threads;
-    cplan.work_size = work_size;
-    cplan.work_data = NULL;
+    cplan.threadpool = threadpool;
+    cplan.n_threads  = MIN(max_tasks, n_threads);
+    cplan.work_size  = work_size;
+    cplan.work_data  = NULL;
 
     return cplan;
 }
 
-int ggml_graph_compute(struct ggml_cgraph * cgraph, struct ggml_cplan * cplan) {
+static thread_ret_t ggml_graph_compute_thread(void * data) {
+    struct ggml_compute_state * state = (struct ggml_compute_state *) data;
+    struct ggml_threadpool    * tp    = state->threadpool;
+
+    const struct ggml_cgraph * cgraph = tp->cgraph;
+    const struct ggml_cplan  * cplan  = tp->cplan;
+
+    set_numa_thread_affinity(state->ith);
+
+    struct ggml_compute_params params = {
+        /*.ith       =*/ state->ith,
+        /*.nth       =*/ atomic_load_explicit(&tp->n_threads_cur, memory_order_relaxed),
+        /*.wsize     =*/ cplan->work_size,
+        /*.wdata     =*/ cplan->work_data,
+        /*.threadpool=*/ tp,
+    };
+
+    for (int node_n = 0; node_n < cgraph->n_nodes && !tp->abort; node_n++) {
+        struct ggml_tensor * node = cgraph->nodes[node_n];
+
+        ggml_compute_forward(&params, node);
+
+        if (state->ith == 0 && cplan->abort_callback &&
+                cplan->abort_callback(cplan->abort_callback_data)) {
+            tp->abort = true;
+            tp->ec    = GGML_STATUS_ABORTED;
+        }
+
+        ggml_barrier(state->threadpool);
+    }
+
+    return 0;
+}
+
+#ifndef GGML_USE_OPENMP
+
+// check if thread is active
+static inline bool ggml_graph_compute_thread_active(struct ggml_compute_state * state) {
+    struct ggml_threadpool * threadpool = state->threadpool;
+    int n_threads = atomic_load_explicit(&threadpool->n_threads_cur, memory_order_relaxed);
+    return (state->ith < n_threads);
+}
+
+// check if thread is ready to proceed (exit from polling or sleeping)
+static inline bool ggml_graph_compute_thread_ready(struct ggml_compute_state * state) {
+    struct ggml_threadpool * threadpool = state->threadpool;
+
+    if (state->pending || threadpool->stop || threadpool->pause) { return true; }
+
+    // check for new graph/work
+    int new_graph = atomic_load_explicit(&threadpool->n_graph, memory_order_relaxed);
+    if (new_graph != state->last_graph) {
+        state->pending    = ggml_graph_compute_thread_active(state);
+        state->last_graph = new_graph;
+    }
+
+    return state->pending;
+}
+
+// sync thread state after polling
+static inline void ggml_graph_compute_thread_sync(struct ggml_compute_state * state) {
+    // TSAN doesn't support standalone fence yet, we use a dummy read-modify-write instead
+    #ifdef GGML_TSAN_ENABLED
+    atomic_fetch_add_explicit(&state->threadpool->n_graph, 0, memory_order_seq_cst);
+    #else
+    atomic_thread_fence(memory_order_seq_cst);
+    #endif
+    UNUSED(state);
+}
+
+static inline bool ggml_graph_compute_poll_for_work(struct ggml_compute_state * state) {
+    struct ggml_threadpool * threadpool = state->threadpool;
+
+    // Skip polling for unused threads
+    if (!ggml_graph_compute_thread_active(state)) {
+        return state->pending;
+    }
+
+    // This seems to make 0 ... 100 a decent range for polling level across modern processors.
+    // Perhaps, we can adjust it dynamically based on load and things.
+    const uint64_t n_rounds = 1024UL * 128 * threadpool->poll;
+
+    for (uint64_t i=0; !ggml_graph_compute_thread_ready(state) && i < n_rounds; i++) {
+        // No new work. Keep polling.
+        ggml_thread_cpu_relax();
+    }
+
+    return state->pending;
+}
+
+static inline bool ggml_graph_compute_check_for_work(struct ggml_compute_state * state) {
+    struct ggml_threadpool * threadpool = state->threadpool;
+
+    if (ggml_graph_compute_poll_for_work(state)) {
+        ggml_graph_compute_thread_sync(state);
+        return state->pending;
+    }
+
+    ggml_mutex_lock_shared(&threadpool->mutex);
+    while (!ggml_graph_compute_thread_ready(state)) {
+        // No new work. Wait for the signal.
+        GGML_PRINT_DEBUG("thread #%d waiting for work (sleeping)\n", state->ith);
+        ggml_cond_wait(&threadpool->cond, &threadpool->mutex);
+    }
+    ggml_mutex_unlock_shared(&threadpool->mutex);
+
+    return state->pending;
+}
+
+static thread_ret_t ggml_graph_compute_secondary_thread(void* data) {
+    struct ggml_compute_state * state = (struct ggml_compute_state *) data;
+    struct ggml_threadpool * threadpool = state->threadpool;
+
+    ggml_thread_apply_priority(threadpool->prio);
+    if (ggml_thread_cpumask_is_valid(state->cpumask)) {
+        ggml_thread_apply_affinity(state->cpumask);
+    }
+
+    while (true) {
+        // Check if we need to sleep
+        while (threadpool->pause) {
+            GGML_PRINT_DEBUG("thread #%d inside pause loop\n", state->ith);
+            ggml_mutex_lock_shared(&threadpool->mutex);
+            if (threadpool->pause) {
+                ggml_cond_wait(&threadpool->cond, &threadpool->mutex);
+            }
+            GGML_PRINT_DEBUG("thread #%d resuming after wait\n", state->ith);
+            ggml_mutex_unlock_shared(&threadpool->mutex);
+        }
+
+        // This needs to be checked for after the cond_wait
+        if (threadpool->stop) break;
+
+        // Check if there is new work
+        // The main thread is the only one that can dispatch new work
+
+        ggml_graph_compute_check_for_work(state);
+        if (state->pending) {
+            state->pending = false;
+
+            ggml_graph_compute_thread(state);
+        }
+    }
+
+    return (thread_ret_t) 0;
+}
+
+// Start processing new graph
+static void ggml_graph_compute_kickoff(struct ggml_threadpool * threadpool, int n_threads)
+{
+    // Always take the mutex here because the worker threads are doing hybrid poll/wait
+
+    ggml_mutex_lock(&threadpool->mutex);
+
+    GGML_PRINT_DEBUG("threadpool: n_threads_cur %d n_threads %d\n", threadpool->n_threads_cur, n_threads);
+
+    // Update the number of active threads
+    atomic_store_explicit(&threadpool->n_threads_cur, n_threads, memory_order_relaxed);
+
+    // Indicate the graph is ready to be processed
+    // We need the full seq-cst fence here because of the polling threads (used in thread_sync)
+    atomic_fetch_add_explicit(&threadpool->n_graph, 1, memory_order_seq_cst);
+
+    if (threadpool->pause) {
+       // Update main thread prio and affinity to match the threadpool settings
+       ggml_thread_apply_priority(threadpool->prio);
+       if (ggml_thread_cpumask_is_valid(threadpool->workers[0].cpumask)) {
+           ggml_thread_apply_affinity(threadpool->workers[0].cpumask);
+       }
+
+       // resume does cond broadcast
+       ggml_threadpool_resume_locked(threadpool);
+    } else {
+       ggml_cond_broadcast(&threadpool->cond);
+    }
+
+    ggml_mutex_unlock(&threadpool->mutex);
+}
+
+#endif // GGML_USE_OPENMP
+
+void ggml_threadpool_params_init(struct ggml_threadpool_params * p, int n_threads) {
+    p->n_threads  = n_threads;
+    p->prio       = 0;     // default priority (usually means normal or inherited)
+    p->poll       = 50;    // hybrid-polling enabled
+    p->strict_cpu = false; // no strict placement (all threads share same cpumask)
+    p->paused     = false; // threads are ready to go
+    memset(p->cpumask, 0, GGML_MAX_N_THREADS); // all-zero means use the default affinity (usually inherited)
+}
+
+struct ggml_threadpool_params ggml_threadpool_params_default(int n_threads) {
+    struct ggml_threadpool_params p;
+    ggml_threadpool_params_init(&p, n_threads);
+    return p;
+}
+
+bool ggml_threadpool_params_match(const struct ggml_threadpool_params * p0, const struct ggml_threadpool_params * p1) {
+    if (p0->n_threads      != p1->n_threads  )    return false;
+    if (p0->prio           != p1->prio       )    return false;
+    if (p0->poll           != p1->poll       )    return false;
+    if (p0->strict_cpu     != p1->strict_cpu )    return false;
+    return memcmp(p0->cpumask, p1->cpumask, GGML_MAX_N_THREADS) == 0;
+}
+
+static struct ggml_threadpool * ggml_threadpool_new_impl(
+    struct ggml_threadpool_params * tpp,
+               struct ggml_cgraph * cgraph,
+                struct ggml_cplan * cplan) {
+
+    struct ggml_threadpool * threadpool =
+        GGML_ALIGNED_MALLOC(sizeof(struct ggml_threadpool));
     {
-        GGML_ASSERT(cplan);
-        GGML_ASSERT(cplan->n_threads > 0);
+        threadpool->cgraph           = cgraph;
+        threadpool->cplan            = cplan;
+        threadpool->n_graph          = 0;
+        threadpool->n_barrier        = 0;
+        threadpool->n_barrier_passed = 0;
+        threadpool->current_chunk    = 0;
+        threadpool->stop             = false;
+        threadpool->pause            = tpp->paused;
+        threadpool->abort            = false;
+        threadpool->workers          = NULL;
+        threadpool->n_threads_max    = tpp->n_threads;
+        threadpool->n_threads_cur    = tpp->n_threads;
+        threadpool->poll             = tpp->poll;
+        threadpool->prio             = tpp->prio;
+        threadpool->ec               = GGML_STATUS_SUCCESS;
+    }
+
+    // Allocate and init workers state
+    const size_t workers_size = sizeof(struct ggml_compute_state) * tpp->n_threads;
+    struct ggml_compute_state * workers = GGML_ALIGNED_MALLOC(workers_size);
+
+    memset(workers, 0, workers_size);
+    for (int j = 0; j < tpp->n_threads; j++) {
+        workers[j].threadpool = threadpool;
+        workers[j].ith        = j;
+    }
+
+    threadpool->workers = workers;
+
+#ifndef GGML_USE_OPENMP
+    ggml_mutex_init(&threadpool->mutex);
+    ggml_cond_init(&threadpool->cond);
+
+    // Spin the threads for all workers, and update CPU placements.
+    // Place the main thread last (towards the higher numbered CPU cores).
+
+    int32_t cpumask_iter = 0;
+
+    for (int j = 1; j < tpp->n_threads; j++) {
+        ggml_thread_cpumask_next(tpp->cpumask, workers[j].cpumask, tpp->strict_cpu, &cpumask_iter);
+
+        int32_t rc = ggml_thread_create(&workers[j].thrd, NULL, ggml_graph_compute_secondary_thread, &workers[j]);
+        GGML_ASSERT(rc == 0);
+    }
+
+    ggml_thread_cpumask_next(tpp->cpumask, workers[0].cpumask, tpp->strict_cpu, &cpumask_iter);
+
+    if (!threadpool->pause) {
+        // Update main thread prio and affinity at the start, otherwise we'll do it in resume
+        ggml_thread_apply_priority(threadpool->prio);
+        if (ggml_thread_cpumask_is_valid(threadpool->workers[0].cpumask)) {
+            ggml_thread_apply_affinity(threadpool->workers[0].cpumask);
+        }
+    }
+#endif // GGML_USE_OPENMP
+
+    return threadpool;
+}
+
+struct ggml_threadpool * ggml_threadpool_new(struct ggml_threadpool_params * tpp) {
+    return ggml_threadpool_new_impl(tpp, NULL, NULL);
+}
+
+enum ggml_status ggml_graph_compute(struct ggml_cgraph * cgraph, struct ggml_cplan * cplan) {
+    GGML_ASSERT(cplan);
+    GGML_ASSERT(cplan->n_threads > 0);
+    GGML_ASSERT(cplan->work_size == 0 || cplan->work_data != NULL);
+
+    int n_threads                               = cplan->n_threads;
+    struct ggml_threadpool * threadpool = cplan->threadpool;
 
-        if (cplan->work_size > 0) {
-            GGML_ASSERT(cplan->work_data);
-        }
-    }
+    bool disposable_threadpool = false;
 
-    const int n_threads = cplan->n_threads;
+    if (threadpool == NULL) {
+        GGML_PRINT_DEBUG("Threadpool is not specified. Will create a disposable threadpool : n_threads %d\n", n_threads);
+        disposable_threadpool = true;
 
-    struct ggml_compute_state_shared state_shared = {
-        /*.cgraph                  =*/ cgraph,
-        /*.cgraph_plan             =*/ cplan,
-        /*.perf_node_start_cycles  =*/ 0,
-        /*.perf_node_start_time_us =*/ 0,
-        /*.n_threads               =*/ n_threads,
-        /*.n_active                =*/ n_threads,
-        /*.node_n                  =*/ -1,
-        /*.abort_callback          =*/ NULL,
-        /*.abort_callback_data     =*/ NULL,
-    };
-    struct ggml_compute_state * workers = alloca(sizeof(struct ggml_compute_state)*n_threads);
+        struct ggml_threadpool_params ttp = ggml_threadpool_params_default(n_threads);
+        threadpool = ggml_threadpool_new_impl(&ttp, cgraph, cplan);
+    } else {
+        // Reset some of the parameters that need resetting
+        // No worker threads should be accessing the parameters below at this stage
+        threadpool->cgraph           = cgraph;
+        threadpool->cplan            = cplan;
+        threadpool->current_chunk    = 0;
+        threadpool->abort            = false;
+        threadpool->ec               = GGML_STATUS_SUCCESS;
+    }
 
-    // create thread pool
+#ifdef GGML_USE_OPENMP
     if (n_threads > 1) {
-        for (int j = 1; j < n_threads; ++j) {
-            workers[j] = (struct ggml_compute_state) {
-                .thrd   = 0,
-                .ith = j,
-                .shared = &state_shared,
-            };
+        #pragma omp parallel num_threads(n_threads)
+        {
+            #pragma omp single
+            {
+                // update the number of threads from the actual number of threads that we got from OpenMP
+                n_threads = omp_get_num_threads();
+                atomic_store_explicit(&threadpool->n_threads_cur, n_threads, memory_order_relaxed);
+            }
 
-            const int rc = ggml_thread_create(&workers[j].thrd, NULL, ggml_graph_compute_thread, &workers[j]);
-            GGML_ASSERT(rc == 0);
-            UNUSED(rc);
+            ggml_graph_compute_thread(&threadpool->workers[omp_get_thread_num()]);
         }
+    } else {
+        atomic_store_explicit(&threadpool->n_threads_cur, 1, memory_order_relaxed);
+        ggml_graph_compute_thread(&threadpool->workers[0]);
+    }
+#else
+    if (n_threads > threadpool->n_threads_max) {
+        GGML_LOG_WARN("cplan requested more threads (%d) than available (%d)\n", n_threads, threadpool->n_threads_max);
+        n_threads = threadpool->n_threads_max;
     }
 
-    workers[0].ith = 0;
-    workers[0].shared = &state_shared;
-
-    const int64_t perf_start_cycles  = ggml_perf_cycles();
-    const int64_t perf_start_time_us = ggml_perf_time_us();
+    // Kick all threads to start the new graph
+    ggml_graph_compute_kickoff(threadpool, n_threads);
 
-    // this is a work thread too
-    int compute_status = (size_t) ggml_graph_compute_thread(&workers[0]);
+    // This is a work thread too
+    ggml_graph_compute_thread(&threadpool->workers[0]);
+#endif
 
     // don't leave affinity set on the main thread
     clear_numa_thread_affinity();
 
-    // join or kill thread pool
-    if (n_threads > 1) {
-        for (int j = 1; j < n_threads; j++) {
-            const int rc = ggml_thread_join(workers[j].thrd, NULL);
-            GGML_ASSERT(rc == 0);
-        }
-    }
-
-    // performance stats (graph)
-    {
-        int64_t perf_cycles_cur  = ggml_perf_cycles()  - perf_start_cycles;
-        int64_t perf_time_us_cur = ggml_perf_time_us() - perf_start_time_us;
-
-        cgraph->perf_runs++;
-        cgraph->perf_cycles  += perf_cycles_cur;
-        cgraph->perf_time_us += perf_time_us_cur;
+    enum ggml_status ret = threadpool->ec;
 
-        GGML_PRINT_DEBUG("%s: perf (%d) - cpu = %.3f / %.3f ms, wall = %.3f / %.3f ms\n",
-                __func__, cgraph->perf_runs,
-                (double) perf_cycles_cur      / (double) ggml_cycles_per_ms(),
-                (double) cgraph->perf_cycles  / (double) ggml_cycles_per_ms() / (double) cgraph->perf_runs,
-                (double) perf_time_us_cur     / 1000.0,
-                (double) cgraph->perf_time_us / 1000.0 / cgraph->perf_runs);
+    if (disposable_threadpool) {
+        ggml_threadpool_free(threadpool);
     }
 
-    return compute_status;
+    return ret;
 }
 
-void ggml_graph_compute_with_ctx(struct ggml_context * ctx, struct ggml_cgraph * cgraph, int n_threads) {
-    struct ggml_cplan cplan = ggml_graph_plan(cgraph, n_threads);
+enum ggml_status ggml_graph_compute_with_ctx(struct ggml_context * ctx, struct ggml_cgraph * cgraph, int n_threads) {
+    struct ggml_cplan cplan = ggml_graph_plan(cgraph, n_threads, NULL);
 
-    struct ggml_object * obj = ggml_new_object(ctx, GGML_OBJECT_WORK_BUFFER, cplan.work_size);
+    struct ggml_object * obj = ggml_new_object(ctx, GGML_OBJECT_TYPE_WORK_BUFFER, cplan.work_size);
 
     cplan.work_data = (uint8_t *)ctx->mem_buffer + obj->offs;
 
-    ggml_graph_compute(cgraph, &cplan);
+    return ggml_graph_compute(cgraph, &cplan);
 }
 
 struct ggml_tensor * ggml_graph_get_tensor(struct ggml_cgraph * cgraph, const char * name) {
@@ -16892,7 +20230,7 @@ static void ggml_graph_export_leaf(const struct ggml_tensor * tensor, FILE * fou
     const int64_t * ne = tensor->ne;
     const size_t  * nb = tensor->nb;
 
-    Rprintf("%-6s %-12s %8d %" PRId64 " %" PRId64 " %" PRId64 " %" PRId64 " %16zu %16zu %16zu %16zu %16p %32s\n",
+    fprintf(fout, "%-6s %-12s %8d %" PRId64 " %" PRId64 " %" PRId64 " %" PRId64 " %16zu %16zu %16zu %16zu %16p %32s\n",
             ggml_type_name(tensor->type),
             ggml_op_name  (tensor->op),
             ggml_n_dims(tensor),
@@ -16906,7 +20244,7 @@ static void ggml_graph_export_node(const struct ggml_tensor * tensor, const char
     const int64_t * ne = tensor->ne;
     const size_t  * nb = tensor->nb;
 
-    Rprintf("%-6s %-6s %-12s %8d %" PRId64 " %" PRId64 " %" PRId64 " %" PRId64 " %16zu %16zu %16zu %16zu %16p %32s\n",
+    fprintf(fout, "%-6s %-6s %-12s %8d %" PRId64 " %" PRId64 " %" PRId64 " %" PRId64 " %16zu %16zu %16zu %16zu %16p %32s\n",
             arg,
             ggml_type_name(tensor->type),
             ggml_op_name  (tensor->op),
@@ -16927,19 +20265,19 @@ void ggml_graph_export(const struct ggml_cgraph * cgraph, const char * fname) {
     }
 
     // print
-    FILE * fout = fopen(fname, "wb");
     {
+        FILE * fout = stdout;
 
-        Rprintf("\n");
-        Rprintf("%-16s %8x\n", "magic",        GGML_FILE_MAGIC);
-        Rprintf("%-16s %8d\n", "version",      GGML_FILE_VERSION);
-        Rprintf("%-16s %8d\n", "leafs",        cgraph->n_leafs);
-        Rprintf("%-16s %8d\n", "nodes",        cgraph->n_nodes);
-        Rprintf("%-16s %" PRIu64 "\n", "eval", size_eval);
+        fprintf(fout, "\n");
+        fprintf(fout, "%-16s %8x\n", "magic",        GGML_FILE_MAGIC);
+        fprintf(fout, "%-16s %8d\n", "version",      GGML_FILE_VERSION);
+        fprintf(fout, "%-16s %8d\n", "leafs",        cgraph->n_leafs);
+        fprintf(fout, "%-16s %8d\n", "nodes",        cgraph->n_nodes);
+        fprintf(fout, "%-16s %" PRIu64 "\n", "eval", size_eval);
 
         // header
-        Rprintf("\n");
-        Rprintf("%-6s %-12s %8s %8s %8s %8s %8s %16s %16s %16s %16s %16s %16s\n",
+        fprintf(fout, "\n");
+        fprintf(fout, "%-6s %-12s %8s %8s %8s %8s %8s %16s %16s %16s %16s %16s %16s\n",
                 "TYPE", "OP", "NDIMS", "NE0", "NE1", "NE2", "NE3", "NB0", "NB1", "NB2", "NB3", "DATA", "NAME");
 
         for (int i = 0; i < cgraph->n_leafs; ++i) {
@@ -16951,8 +20289,8 @@ void ggml_graph_export(const struct ggml_cgraph * cgraph, const char * fname) {
         }
 
         // header
-        Rprintf("\n");
-        Rprintf("%-6s %-6s %-12s %8s %8s %8s %8s %8s %16s %16s %16s %16s %8s %16s %16s\n",
+        fprintf(fout, "\n");
+        fprintf(fout, "%-6s %-6s %-12s %8s %8s %8s %8s %8s %16s %16s %16s %16s %8s %16s %16s\n",
                 "ARG", "TYPE", "OP", "NDIMS", "NE0", "NE1", "NE2", "NE3", "NB0", "NB1", "NB2", "NB3", "NTASKS", "DATA", "NAME");
 
         for (int i = 0; i < cgraph->n_nodes; ++i) {
@@ -16964,18 +20302,18 @@ void ggml_graph_export(const struct ggml_cgraph * cgraph, const char * fname) {
                 }
             }
 
-            Rprintf("\n");
+            fprintf(fout, "\n");
         }
 
-        Rprintf("\n");
+        fprintf(fout, "\n");
     }
 
     // write binary data
     {
-        //FILE * fout = fopen(fname, "wb");
+        FILE * fout = ggml_fopen(fname, "wb");
 
         if (!fout) {
-            Rprintf("%s: failed to open %s\n", __func__, fname);
+            fprintf(stderr, "%s: failed to open %s: %s\n", __func__, fname, strerror(errno));
             return;
         }
 
@@ -17000,9 +20338,11 @@ void ggml_graph_export(const struct ggml_cgraph * cgraph, const char * fname) {
 
                 const uint32_t type   = tensor->type;
                 const uint32_t op     = tensor->op;
+                const int32_t  flags  = tensor->flags;
 
                 fwrite(&type,   sizeof(uint32_t), 1, fout);
                 fwrite(&op,     sizeof(uint32_t), 1, fout);
+                fwrite(&flags,  sizeof(int32_t),  1, fout);
 
                 for (int j = 0; j < GGML_MAX_DIMS; ++j) {
                     const uint64_t ne = tensor->ne[j];
@@ -17032,9 +20372,11 @@ void ggml_graph_export(const struct ggml_cgraph * cgraph, const char * fname) {
 
                 const uint32_t type   = tensor->type;
                 const uint32_t op     = tensor->op;
+                const int32_t  flags  = tensor->flags;
 
                 fwrite(&type,   sizeof(uint32_t), 1, fout);
                 fwrite(&op,     sizeof(uint32_t), 1, fout);
+                fwrite(&flags,  sizeof(int32_t),  1, fout);
 
                 for (int j = 0; j < GGML_MAX_DIMS; ++j) {
                     const uint64_t ne = tensor->ne[j];
@@ -17080,7 +20422,7 @@ void ggml_graph_export(const struct ggml_cgraph * cgraph, const char * fname) {
                             }
 
                             if (idx == -1) {
-                                Rprintf("%s: failed to find tensor, arg = %d, node = %d\n", __func__, j, i);
+                                fprintf(stderr, "%s: failed to find tensor, arg = %d, node = %d\n", __func__, j, i);
                                 fclose(fout);
                                 return;
                             }
@@ -17093,12 +20435,19 @@ void ggml_graph_export(const struct ggml_cgraph * cgraph, const char * fname) {
                         }
                     }
                 }
+
+                // dump the data
+                // TODO: pad this to 32 byte boundary
+                if ((flags & GGML_TENSOR_FLAG_PARAM)) {
+                    const size_t size = ggml_nbytes(tensor);
+
+                    fwrite(tensor->data, sizeof(char), size, fout);
+                }
             }
         }
 
-        
+        fclose(fout);
     }
-    fclose(fout);
 }
 
 struct ggml_cgraph * ggml_graph_import(const char * fname, struct ggml_context ** ctx_data, struct ggml_context ** ctx_eval) {
@@ -17111,9 +20460,9 @@ struct ggml_cgraph * ggml_graph_import(const char * fname, struct ggml_context *
 
     // read file into data
     {
-        FILE * fin = fopen(fname, "rb");
+        FILE * fin = ggml_fopen(fname, "rb");
         if (!fin) {
-            Rprintf("%s: failed to open %s\n", __func__, fname);
+            fprintf(stderr, "%s: failed to open %s: %s\n", __func__, fname, strerror(errno));
             return result;
         }
 
@@ -17136,7 +20485,7 @@ struct ggml_cgraph * ggml_graph_import(const char * fname, struct ggml_context *
             *ctx_data = ggml_init(params);
 
             if (!*ctx_data) {
-                Rprintf("%s: failed to create ggml context\n", __func__);
+                fprintf(stderr, "%s: failed to create ggml context\n", __func__);
                 fclose(fin);
                 return result;
             }
@@ -17147,7 +20496,7 @@ struct ggml_cgraph * ggml_graph_import(const char * fname, struct ggml_context *
         {
             const size_t ret = fread(data->data, sizeof(char), fsize, fin);
             if (ret != fsize) {
-                Rprintf("%s: failed to read %s\n", __func__, fname);
+                fprintf(stderr, "%s: failed to read %s\n", __func__, fname);
                 fclose(fin);
                 return result;
             }
@@ -17163,14 +20512,14 @@ struct ggml_cgraph * ggml_graph_import(const char * fname, struct ggml_context *
         const uint32_t magic = *(const uint32_t *) ptr; ptr += sizeof(magic);
 
         if (magic != GGML_FILE_MAGIC) {
-            Rprintf("%s: invalid magic number, got %08x\n", __func__, magic);
+            fprintf(stderr, "%s: invalid magic number, got %08x\n", __func__, magic);
             return result;
         }
 
         const uint32_t version = *(const uint32_t *) ptr; ptr += sizeof(version);
 
         if (version != GGML_FILE_VERSION) {
-            Rprintf("%s: invalid version number\n", __func__);
+            fprintf(stderr, "%s: invalid version number\n", __func__);
             return result;
         }
 
@@ -17192,7 +20541,7 @@ struct ggml_cgraph * ggml_graph_import(const char * fname, struct ggml_context *
             *ctx_eval = ggml_init(params);
 
             if (!*ctx_eval) {
-                Rprintf("%s: failed to create ggml context\n", __func__);
+                fprintf(stderr, "%s: failed to create ggml context\n", __func__);
                 return result;
             }
         }
@@ -17207,10 +20556,12 @@ struct ggml_cgraph * ggml_graph_import(const char * fname, struct ggml_context *
         {
             uint32_t type;
             uint32_t op;
+            int32_t  flags;
 
             for (uint32_t i = 0; i < n_leafs; ++i) {
                 type   = *(const uint32_t *) ptr; ptr += sizeof(type);
                 op     = *(const uint32_t *) ptr; ptr += sizeof(op);
+                flags  = *(const int32_t  *) ptr; ptr += sizeof(flags);
 
                 int64_t ne[GGML_MAX_DIMS];
                 size_t  nb[GGML_MAX_DIMS];
@@ -17228,22 +20579,21 @@ struct ggml_cgraph * ggml_graph_import(const char * fname, struct ggml_context *
 
                 struct ggml_tensor * tensor = ggml_new_tensor(*ctx_eval, (enum ggml_type) type, GGML_MAX_DIMS, ne);
 
-                tensor->op = (enum ggml_op) op;
+                tensor->op    = (enum ggml_op) op;
+                tensor->flags = flags;
 
                 memcpy(tensor->name,      ptr, GGML_MAX_NAME);      ptr += GGML_MAX_NAME;
                 memcpy(tensor->op_params, ptr, GGML_MAX_OP_PARAMS); ptr += GGML_MAX_OP_PARAMS;
 
-                tensor->data = (void *) ptr;
-
                 for (int j = 0; j < GGML_MAX_DIMS; ++j) {
                     tensor->nb[j] = nb[j];
                 }
 
-                result->leafs[i] = tensor;
+                tensor->data = (void *) ptr; ptr += ggml_nbytes(tensor);
 
-                ptr += ggml_nbytes(tensor);
+                result->leafs[i] = tensor;
 
-                Rprintf("%s: loaded leaf %d: '%16s', %9zu bytes\n", __func__, i, tensor->name, ggml_nbytes(tensor));
+                fprintf(stderr, "%s: loaded leaf %u: '%16s', %9zu bytes\n", __func__, i, tensor->name, ggml_nbytes(tensor));
             }
         }
 
@@ -17253,10 +20603,12 @@ struct ggml_cgraph * ggml_graph_import(const char * fname, struct ggml_context *
         {
             uint32_t type;
             uint32_t op;
+            int32_t  flags;
 
             for (uint32_t i = 0; i < n_nodes; ++i) {
                 type   = *(const uint32_t *) ptr; ptr += sizeof(type);
                 op     = *(const uint32_t *) ptr; ptr += sizeof(op);
+                flags  = *(const int32_t  *) ptr; ptr += sizeof(flags);
 
                 enum ggml_op eop = (enum ggml_op) op;
 
@@ -17346,7 +20698,12 @@ struct ggml_cgraph * ggml_graph_import(const char * fname, struct ggml_context *
 
                 result->nodes[i] = tensor;
 
-                Rprintf("%s: loaded node %d: '%16s', %9zu bytes\n", __func__, i, tensor->name, ggml_nbytes(tensor));
+                // TODO tensor data is be duplicated due to ggml_new_tensor call above
+                if (flags & GGML_TENSOR_FLAG_PARAM) {
+                    tensor->data = (void *) ptr; ptr += ggml_nbytes(tensor);
+                }
+
+                fprintf(stderr, "%s: loaded node %u: '%16s', %9zu bytes\n", __func__, i, tensor->name, ggml_nbytes(tensor));
             }
         }
     }
@@ -17355,46 +20712,30 @@ struct ggml_cgraph * ggml_graph_import(const char * fname, struct ggml_context *
 }
 
 void ggml_graph_print(const struct ggml_cgraph * cgraph) {
-    int64_t perf_total_per_op_us[GGML_OP_COUNT] = {0};
-
-    GGML_PRINT("=== GRAPH ===\n");
+    GGML_LOG_INFO("=== GRAPH ===\n");
 
-    GGML_PRINT("n_nodes = %d\n", cgraph->n_nodes);
+    GGML_LOG_INFO("n_nodes = %d\n", cgraph->n_nodes);
     for (int i = 0; i < cgraph->n_nodes; i++) {
         struct ggml_tensor * node = cgraph->nodes[i];
 
-        perf_total_per_op_us[node->op] += MAX(1, node->perf_time_us);
-
-        GGML_PRINT(" - %3d: [ %5" PRId64 ", %5" PRId64 ", %5" PRId64 "] %16s %s (%3d) cpu = %7.3f / %7.3f ms, wall = %7.3f / %7.3f ms\n",
+        GGML_LOG_INFO(" - %3d: [ %5" PRId64 ", %5" PRId64 ", %5" PRId64 "] %16s %s\n",
                 i,
                 node->ne[0], node->ne[1], node->ne[2],
-                ggml_op_name(node->op), node->is_param ? "x" : node->grad ? "g" : " ", node->perf_runs,
-                (double) node->perf_cycles  / (double) ggml_cycles_per_ms(),
-                (double) node->perf_cycles  / (double) ggml_cycles_per_ms() / (double) node->perf_runs,
-                (double) node->perf_time_us / 1000.0,
-                (double) node->perf_time_us / 1000.0 / node->perf_runs);
+                ggml_op_name(node->op), (node->flags & GGML_TENSOR_FLAG_PARAM) ? "x" : node->grad ? "g" : " ");
     }
 
-    GGML_PRINT("n_leafs = %d\n", cgraph->n_leafs);
+    GGML_LOG_INFO("n_leafs = %d\n", cgraph->n_leafs);
     for (int i = 0; i < cgraph->n_leafs; i++) {
         struct ggml_tensor * node = cgraph->leafs[i];
 
-        GGML_PRINT(" - %3d: [ %5" PRId64 ", %5" PRId64 "] %8s %16s\n",
+        GGML_LOG_INFO(" - %3d: [ %5" PRId64 ", %5" PRId64 "] %8s %16s\n",
                 i,
                 node->ne[0], node->ne[1],
                 ggml_op_name(node->op),
                 ggml_get_name(node));
     }
 
-    for (int i = 0; i < GGML_OP_COUNT; i++) {
-        if (perf_total_per_op_us[i] == 0) {
-            continue;
-        }
-
-        GGML_PRINT("perf_total_per_op_us[%16s] = %7.3f ms\n", ggml_op_name(i), (double) perf_total_per_op_us[i] / 1000.0);
-    }
-
-    GGML_PRINT("========================================\n");
+    GGML_LOG_INFO("========================================\n");
 }
 
 // check if node is part of the graph
@@ -17447,12 +20788,12 @@ static void ggml_graph_dump_dot_leaf_edge(FILE * fp, struct ggml_tensor * node,
 void ggml_graph_dump_dot(const struct ggml_cgraph * gb, const struct ggml_cgraph * gf, const char * filename) {
     char color[16];
 
-    FILE * fp = fopen(filename, "w");
+    FILE * fp = ggml_fopen(filename, "w");
     GGML_ASSERT(fp);
 
     fprintf(fp, "digraph G {\n");
     fprintf(fp, "  newrank = true;\n");
-    fprintf(fp, "  rankdir = LR;\n");
+    fprintf(fp, "  rankdir = TB;\n");
 
     for (int i = 0; i < gb->n_nodes; i++) {
         struct ggml_tensor * node = gb->nodes[i];
@@ -17461,7 +20802,7 @@ void ggml_graph_dump_dot(const struct ggml_cgraph * gb, const struct ggml_cgraph
             continue;
         }
 
-        if (node->is_param) {
+        if (node->flags & GGML_TENSOR_FLAG_PARAM) {
             snprintf(color, sizeof(color), "yellow");
         } else if (node->grad) {
             if (ggml_graph_find(gf, node)) {
@@ -17514,13 +20855,15 @@ void ggml_graph_dump_dot(const struct ggml_cgraph * gb, const struct ggml_cgraph
         }
 
         fprintf(fp, "CONST %d [%" PRId64 ", %" PRId64 "]", i, node->ne[0], node->ne[1]);
-        if (ggml_nelements(node) < 5) {
+        if (ggml_nelements(node) < 5 && node->data != NULL) {
             fprintf(fp, " | (");
             for (int j = 0; j < ggml_nelements(node); j++) {
                 if (node->type == GGML_TYPE_I8 || node->type == GGML_TYPE_I16 || node->type == GGML_TYPE_I32) {
                     fprintf(fp, "%d", ggml_get_i32_1d(node, j));
                 }
-                else if (node->type == GGML_TYPE_F32 || node->type == GGML_TYPE_F16) {
+                else if (node->type == GGML_TYPE_F32 ||
+                         node->type == GGML_TYPE_F16 ||
+                         node->type == GGML_TYPE_BF16) {
                     fprintf(fp, "%.1e", (double)ggml_get_f32_1d(node, j));
                 }
                 else {
@@ -17563,7 +20906,7 @@ void ggml_graph_dump_dot(const struct ggml_cgraph * gb, const struct ggml_cgraph
 
     fclose(fp);
 
-    GGML_PRINT("%s: dot -Tpng %s -o %s.png && open %s.png\n", __func__, filename, filename, filename);
+    GGML_LOG_INFO("%s: dot -Tpng %s -o %s.png && open %s.png\n", __func__, filename, filename, filename);
 }
 
 ////////////////////////////////////////////////////////////////////////////////
@@ -17628,6 +20971,7 @@ static enum ggml_opt_result ggml_opt_adam(
         ggml_opt_callback callback,
         void * callback_data) {
     GGML_ASSERT(ggml_is_scalar(f));
+    GGML_ASSERT(f->type == GGML_TYPE_F32);
 
     // these will store the parameters we want to optimize
     struct ggml_tensor * ps[GGML_MAX_PARAMS];
@@ -17635,7 +20979,7 @@ static enum ggml_opt_result ggml_opt_adam(
     int np = 0;
     int64_t nx = 0;
     for (int i = 0; i < gf->n_nodes; ++i) {
-        if (gf->nodes[i]->is_param) {
+        if (gf->nodes[i]->flags & GGML_TENSOR_FLAG_PARAM) {
             GGML_PRINT_DEBUG("found param %d: grad->op = %d\n", np, gf->nodes[i]->grad->op);
 
             GGML_ASSERT(np < GGML_MAX_PARAMS);
@@ -17669,8 +21013,8 @@ static enum ggml_opt_result ggml_opt_adam(
 
     float * pf = params.past > 0 ? opt->adam.pf->data : NULL; // past function values
 
-    struct ggml_cplan cplan = ggml_graph_plan(gb, params.n_threads);
-    struct ggml_object * obj = ggml_new_object(ctx, GGML_OBJECT_WORK_BUFFER, cplan.work_size);
+    struct ggml_cplan cplan = ggml_graph_plan(gb, params.n_threads, NULL);
+    struct ggml_object * obj = ggml_new_object(ctx, GGML_OBJECT_TYPE_WORK_BUFFER, cplan.work_size);
     cplan.work_data = (uint8_t *)ctx->mem_buffer + obj->offs;
 
     bool cancel = false;
@@ -17682,7 +21026,7 @@ static enum ggml_opt_result ggml_opt_adam(
         if (callback) {
             callback(callback_data, accum_step, &sched, &cancel);
             if (cancel) {
-                return GGML_OPT_CANCEL;
+                return GGML_OPT_RESULT_CANCEL;
             }
         }
         // ggml_graph_reset  (gf);
@@ -17773,7 +21117,7 @@ static enum ggml_opt_result ggml_opt_adam(
             if (callback) {
                 callback(callback_data, accum_step, &sched, &cancel);
                 if (cancel) {
-                    return GGML_OPT_CANCEL;;
+                    return GGML_OPT_RESULT_CANCEL;;
                 }
             }
             // ggml_graph_reset  (gf);
@@ -17790,7 +21134,7 @@ static enum ggml_opt_result ggml_opt_adam(
         if (fabsf(fx - fx_prev[0])/fx < params.adam.eps_f) {
             GGML_PRINT_DEBUG("converged\n");
 
-            return GGML_OPT_OK;
+            return GGML_OPT_RESULT_OK;
         }
 
         // delta-based convergence test
@@ -17800,7 +21144,7 @@ static enum ggml_opt_result ggml_opt_adam(
                 const float rate = (pf[(iter0 + t)%params.past] - fx)/fx;
 
                 if (fabsf(rate) < params.delta) {
-                    return GGML_OPT_OK;
+                    return GGML_OPT_RESULT_OK;
                 }
             }
 
@@ -17816,7 +21160,7 @@ static enum ggml_opt_result ggml_opt_adam(
                 ++n_no_improvement[0];
 
                 if (n_no_improvement[0] >= params.max_no_improvement) {
-                    return GGML_OPT_OK;
+                    return GGML_OPT_RESULT_OK;
                 }
             }
         }
@@ -17834,7 +21178,7 @@ static enum ggml_opt_result ggml_opt_adam(
         }
     }
 
-    return GGML_OPT_DID_NOT_CONVERGE;
+    return GGML_OPT_RESULT_DID_NOT_CONVERGE;
 }
 
 //
@@ -17888,7 +21232,7 @@ static enum ggml_opt_result linesearch_backtracking(
     }
 
     // compute the initial gradient in the search direction
-    ggml_vec_dot_f32(nx, &dginit, g, d);
+    ggml_vec_dot_f32(nx, &dginit, 0, g, 0, d, 0, 1);
 
     // make sure that d points to a descent direction
     if (0 < dginit) {
@@ -17915,7 +21259,7 @@ static enum ggml_opt_result linesearch_backtracking(
                     float sched = 0;
                     callback(callback_data, accum_step, &sched, cancel);
                     if (*cancel) {
-                        return GGML_OPT_CANCEL;
+                        return GGML_OPT_RESULT_CANCEL;
                     }
                 }
                 // ggml_graph_reset  (gf);
@@ -17938,7 +21282,7 @@ static enum ggml_opt_result linesearch_backtracking(
                 return count;
             }
 
-            ggml_vec_dot_f32(nx, &dg, g, d);
+            ggml_vec_dot_f32(nx, &dg, 0, g, 0, d, 0, 1);
 
             // check the Wolfe condition
             if (dg < params->lbfgs.wolfe * dginit) {
@@ -17971,7 +21315,9 @@ static enum ggml_opt_result linesearch_backtracking(
         (*step) *= width;
     }
 
-    GGML_UNREACHABLE();
+    GGML_ABORT("line search failed");
+
+    //return GGML_LINESEARCH_FAIL;
 }
 
 static enum ggml_opt_result ggml_opt_lbfgs(
@@ -17986,7 +21332,7 @@ static enum ggml_opt_result ggml_opt_lbfgs(
     if (params.lbfgs.linesearch == GGML_LINESEARCH_BACKTRACKING_WOLFE ||
         params.lbfgs.linesearch == GGML_LINESEARCH_BACKTRACKING_STRONG_WOLFE) {
         if (params.lbfgs.wolfe <= params.lbfgs.ftol || 1.f <= params.lbfgs.wolfe) {
-            return GGML_OPT_INVALID_WOLFE;
+            return GGML_OPT_RESULT_INVALID_WOLFE;
         }
     }
 
@@ -17998,7 +21344,7 @@ static enum ggml_opt_result ggml_opt_lbfgs(
     int np = 0;
     int nx = 0;
     for (int i = 0; i < gf->n_nodes; ++i) {
-        if (gf->nodes[i]->is_param) {
+        if (gf->nodes[i]->flags & GGML_TENSOR_FLAG_PARAM) {
             GGML_PRINT_DEBUG("found param %d: grad->op = %d\n", np, gf->nodes[i]->grad->op);
 
             GGML_ASSERT(np < GGML_MAX_PARAMS);
@@ -18014,8 +21360,8 @@ static enum ggml_opt_result ggml_opt_lbfgs(
         opt->iter = iter;
     }
 
-    struct ggml_cplan cplan = ggml_graph_plan(gb, params.n_threads);
-    struct ggml_object * obj = ggml_new_object(ctx, GGML_OBJECT_WORK_BUFFER, cplan.work_size);
+    struct ggml_cplan cplan = ggml_graph_plan(gb, params.n_threads, NULL);
+    struct ggml_object * obj = ggml_new_object(ctx, GGML_OBJECT_TYPE_WORK_BUFFER, cplan.work_size);
     cplan.work_data = (uint8_t *)ctx->mem_buffer + obj->offs;
 
     float * x  = opt->lbfgs.x->data;  // current parameters
@@ -18056,7 +21402,7 @@ static enum ggml_opt_result ggml_opt_lbfgs(
                 float sched = 0;
                 callback(callback_data, accum_step, &sched, &cancel);
                 if (cancel) {
-                    return GGML_OPT_CANCEL;
+                    return GGML_OPT_RESULT_CANCEL;
                 }
             }
             // ggml_graph_reset  (gf);
@@ -18084,7 +21430,7 @@ static enum ggml_opt_result ggml_opt_lbfgs(
 
     // already optimized
     if (gnorm/xnorm <= params.lbfgs.eps) {
-        return GGML_OPT_OK;
+        return GGML_OPT_RESULT_OK;
     }
 
     if (opt->just_initialized) {
@@ -18129,7 +21475,7 @@ static enum ggml_opt_result ggml_opt_lbfgs(
         //       way to test and don't want to break something with so many changes lined up
         ls = linesearch_backtracking(&params, nx, x, &fx, g, d, step, xp, f, gb, &cplan, np, ps, &cancel, callback, callback_data);
         if (cancel) {
-            return GGML_OPT_CANCEL;
+            return GGML_OPT_RESULT_CANCEL;
         }
 
         if (ls < 0) {
@@ -18152,7 +21498,7 @@ static enum ggml_opt_result ggml_opt_lbfgs(
         }
         if (gnorm/xnorm <= params.lbfgs.eps) {
             // converged
-            return GGML_OPT_OK;
+            return GGML_OPT_RESULT_OK;
         }
 
         // delta-based convergence test
@@ -18162,7 +21508,7 @@ static enum ggml_opt_result ggml_opt_lbfgs(
                 const float rate = (pf[k[0]%params.past] - fx)/fx;
 
                 if (fabsf(rate) < params.delta) {
-                    return GGML_OPT_OK;
+                    return GGML_OPT_RESULT_OK;
                 }
             }
 
@@ -18178,14 +21524,14 @@ static enum ggml_opt_result ggml_opt_lbfgs(
                 n_no_improvement[0]++;
 
                 if (n_no_improvement[0] >= params.max_no_improvement) {
-                    return GGML_OPT_OK;
+                    return GGML_OPT_RESULT_OK;
                 }
             }
         }
 
         if (params.lbfgs.n_iter != 0 && params.lbfgs.n_iter < it + 1) {
             // reached the maximum number of iterations
-            return GGML_OPT_DID_NOT_CONVERGE;
+            return GGML_OPT_RESULT_DID_NOT_CONVERGE;
         }
 
         // update vectors s and y:
@@ -18199,8 +21545,8 @@ static enum ggml_opt_result ggml_opt_lbfgs(
         //     ys = y^t \cdot s    -> 1 / \rho.
         //     yy = y^t \cdot y.
         //
-        ggml_vec_dot_f32(nx, &ys, &lm_y[end[0]*nx], &lm_s[end[0]*nx]);
-        ggml_vec_dot_f32(nx, &yy, &lm_y[end[0]*nx], &lm_y[end[0]*nx]);
+        ggml_vec_dot_f32(nx, &ys, 0, &lm_y[end[0]*nx], 0, &lm_s[end[0]*nx], 0, 1);
+        ggml_vec_dot_f32(nx, &yy, 0, &lm_y[end[0]*nx], 0, &lm_y[end[0]*nx], 0, 1);
 
         lm_ys[end[0]] = ys;
 
@@ -18219,7 +21565,7 @@ static enum ggml_opt_result ggml_opt_lbfgs(
         for (int i = 0; i < bound; ++i) {
             j[0] = (j[0] + m - 1) % m;
             // \alpha_{j} = \rho_{j} s^{t}_{j} \cdot q_{k+1}
-            ggml_vec_dot_f32(nx, &lm_alpha[j[0]], &lm_s[j[0]*nx], d);
+            ggml_vec_dot_f32(nx, &lm_alpha[j[0]], 0, &lm_s[j[0]*nx], 0, d, 0, 1);
             lm_alpha[j[0]] /= lm_ys[j[0]];
             // q_{i} = q_{i+1} - \alpha_{i} y_{i}
             ggml_vec_mad_f32(nx, d, &lm_y[j[0]*nx], -lm_alpha[j[0]]);
@@ -18229,7 +21575,7 @@ static enum ggml_opt_result ggml_opt_lbfgs(
 
         for (int i = 0; i < bound; ++i) {
             // \beta_{j} = \rho_{j} y^t_{j} \cdot \gamma_{i}
-            ggml_vec_dot_f32(nx, &beta, &lm_y[j[0]*nx], d);
+            ggml_vec_dot_f32(nx, &beta, 0, &lm_y[j[0]*nx], 0, d, 0, 1);
             beta /= lm_ys[j[0]];
             // \gamma_{i+1} = \gamma_{i} + (\alpha_{j} - \beta_{j}) s_{j}
             ggml_vec_mad_f32(nx, d, &lm_s[j[0]*nx], lm_alpha[j[0]] - beta);
@@ -18239,17 +21585,19 @@ static enum ggml_opt_result ggml_opt_lbfgs(
         step[0] = 1.0;
     }
 
-    GGML_UNREACHABLE();
+    GGML_ABORT("lbfgs failed");
+
+    //return GGML_OPT_RESULT_DID_NOT_CONVERGE;
 }
 
 struct ggml_opt_params ggml_opt_default_params(enum ggml_opt_type type) {
     struct ggml_opt_params result;
 
     switch (type) {
-        case GGML_OPT_ADAM:
+        case GGML_OPT_TYPE_ADAM:
             {
                 result = (struct ggml_opt_params) {
-                    .type       = GGML_OPT_ADAM,
+                    .type       = GGML_OPT_TYPE_ADAM,
                     .graph_size = GGML_DEFAULT_GRAPH_SIZE,
                     .n_threads  = 1, // FIXME: GGML_DEFAULT_N_THREADS ?
                     .past       = 0,
@@ -18277,10 +21625,10 @@ struct ggml_opt_params ggml_opt_default_params(enum ggml_opt_type type) {
                     },
                 };
             } break;
-        case GGML_OPT_LBFGS:
+        case GGML_OPT_TYPE_LBFGS:
             {
                 result = (struct ggml_opt_params) {
-                    .type       = GGML_OPT_LBFGS,
+                    .type       = GGML_OPT_TYPE_LBFGS,
                     .graph_size = GGML_DEFAULT_GRAPH_SIZE,
                     .n_threads  = 1,
                     .past       = 0,
@@ -18325,12 +21673,12 @@ GGML_API void ggml_opt_init(
     opt->just_initialized = true;
     if (opt->ctx == NULL) {
         struct ggml_init_params ctx_opt_params;
-        if (opt->params.type == GGML_OPT_ADAM) {
+        if (opt->params.type == GGML_OPT_TYPE_ADAM) {
             ctx_opt_params.mem_size = GGML_MEM_ALIGN*3 + ggml_tensor_overhead()*3 + ggml_type_size(GGML_TYPE_F32)*nx*3;
             if (opt->params.past > 0) {
                 ctx_opt_params.mem_size += GGML_MEM_ALIGN + ggml_tensor_overhead() + ggml_type_size(GGML_TYPE_F32)*opt->params.past;
             }
-        } else if (opt->params.type == GGML_OPT_LBFGS) {
+        } else if (opt->params.type == GGML_OPT_TYPE_LBFGS) {
             ctx_opt_params.mem_size = GGML_MEM_ALIGN*9 + ggml_tensor_overhead()*9 + ggml_type_size(GGML_TYPE_F32)*(nx*5 + opt->params.lbfgs.m*2 + nx*opt->params.lbfgs.m*2);
             if (opt->params.past > 0) {
                 ctx_opt_params.mem_size += GGML_MEM_ALIGN + ggml_tensor_overhead() + ggml_type_size(GGML_TYPE_F32)*opt->params.past;
@@ -18342,7 +21690,7 @@ GGML_API void ggml_opt_init(
         opt->ctx = ggml_init(ctx_opt_params);
     }
     switch (opt->params.type) {
-        case GGML_OPT_ADAM:
+        case GGML_OPT_TYPE_ADAM:
             {
                 opt->adam.g  = ggml_new_tensor_1d(opt->ctx, GGML_TYPE_F32, nx);
                 opt->adam.m  = ggml_new_tensor_1d(opt->ctx, GGML_TYPE_F32, nx);
@@ -18356,7 +21704,7 @@ GGML_API void ggml_opt_init(
                     ggml_set_zero(opt->adam.pf);
                 }
             } break;
-        case GGML_OPT_LBFGS:
+        case GGML_OPT_TYPE_LBFGS:
             {
                 opt->lbfgs.x  = ggml_new_tensor_1d(opt->ctx, GGML_TYPE_F32, nx);
                 opt->lbfgs.xp = ggml_new_tensor_1d(opt->ctx, GGML_TYPE_F32, nx);
@@ -18400,13 +21748,13 @@ enum ggml_opt_result ggml_opt(
 
         ctx = ggml_init(params_ctx);
         if (ctx == NULL) {
-            return GGML_OPT_NO_CONTEXT;
+            return GGML_OPT_RESULT_NO_CONTEXT;
         }
 
         free_ctx = true;
     }
 
-    enum ggml_opt_result result = GGML_OPT_OK;
+    enum ggml_opt_result result = GGML_OPT_RESULT_OK;
 
     struct ggml_opt_context * opt = (struct ggml_opt_context *) alloca(sizeof(struct ggml_opt_context));
 
@@ -18430,7 +21778,7 @@ enum ggml_opt_result ggml_opt_resume(
     ggml_build_forward_expand(gf, f);
 
     struct ggml_cgraph * gb = ggml_graph_dup(ctx, gf);
-    ggml_build_backward_expand(ctx, gf, gb, true);
+    ggml_build_backward_expand(ctx, gf, gb, false);
 
     return ggml_opt_resume_g(ctx, opt, f, gf, gb, NULL, NULL);
 }
@@ -18444,15 +21792,17 @@ enum ggml_opt_result ggml_opt_resume_g(
         ggml_opt_callback callback,
         void * callback_data) {
 
+    GGML_ASSERT(f->grad && "ggml_set_param must be called for at least one ancestor");
+
     // build forward + backward compute graphs
-    enum ggml_opt_result result = GGML_OPT_OK;
+    enum ggml_opt_result result = GGML_OPT_RESULT_OK;
 
     switch (opt->params.type) {
-        case GGML_OPT_ADAM:
+        case GGML_OPT_TYPE_ADAM:
             {
                 result = ggml_opt_adam(ctx, opt, opt->params, f, gf, gb, callback, callback_data);
             } break;
-        case GGML_OPT_LBFGS:
+        case GGML_OPT_TYPE_LBFGS:
             {
                 result = ggml_opt_lbfgs(ctx, opt, opt->params, f, gf, gb, callback, callback_data);
             } break;
@@ -18473,211 +21823,137 @@ enum ggml_opt_result ggml_opt_resume_g(
 
 ////////////////////////////////////////////////////////////////////////////////
 
-size_t ggml_quantize_q4_0(const float * src, void * dst, int n, int k, int64_t * hist) {
-    assert(k % QK4_0 == 0);
-    const int nb = k / QK4_0;
-
-    for (int b = 0; b < n; b += k) {
-        block_q4_0 * restrict y = (block_q4_0 *) dst + b/QK4_0;
-
-        quantize_row_q4_0_reference(src + b, y, k);
-
-        for (int i = 0; i < nb; i++) {
-            for (int j = 0; j < QK4_0; j += 2) {
-                const uint8_t vi0 = y[i].qs[j/2] & 0x0F;
-                const uint8_t vi1 = y[i].qs[j/2] >> 4;
-
-                hist[vi0]++;
-                hist[vi1]++;
-            }
-        }
-    }
-
-    return (n/QK4_0*sizeof(block_q4_0));
+void ggml_set_input(struct ggml_tensor * tensor) {
+    tensor->flags |= GGML_TENSOR_FLAG_INPUT;
 }
 
-size_t ggml_quantize_q4_1(const float * src, void * dst, int n, int k, int64_t * hist) {
-    assert(k % QK4_1 == 0);
-    const int nb = k / QK4_1;
-
-    for (int b = 0; b < n; b += k) {
-        block_q4_1 * restrict y = (block_q4_1 *) dst + b/QK4_1;
-
-        quantize_row_q4_1_reference(src + b, y, k);
-
-        for (int i = 0; i < nb; i++) {
-            for (int j = 0; j < QK4_1; j += 2) {
-                const uint8_t vi0 = y[i].qs[j/2] & 0x0F;
-                const uint8_t vi1 = y[i].qs[j/2] >> 4;
-
-                hist[vi0]++;
-                hist[vi1]++;
-            }
-        }
-    }
-
-    return (n/QK4_1*sizeof(block_q4_1));
+void ggml_set_output(struct ggml_tensor * tensor) {
+    tensor->flags |= GGML_TENSOR_FLAG_OUTPUT;
 }
 
-size_t ggml_quantize_q5_0(const float * src, void * dst, int n, int k, int64_t * hist) {
-    assert(k % QK5_0 == 0);
-    const int nb = k / QK5_0;
-
-    for (int b = 0; b < n; b += k) {
-        block_q5_0 * restrict y = (block_q5_0 *)dst + b/QK5_0;
-
-        quantize_row_q5_0_reference(src + b, y, k);
+void ggml_set_param(struct ggml_context * ctx, struct ggml_tensor * tensor) {
+    GGML_UNUSED(ctx); // TODO: remove this parameter
+    tensor->flags |= GGML_TENSOR_FLAG_PARAM;
+}
 
-        for (int i = 0; i < nb; i++) {
-            uint32_t qh;
-            memcpy(&qh, &y[i].qh, sizeof(qh));
+void ggml_set_loss(struct ggml_tensor * tensor) {
+    GGML_ASSERT(ggml_is_scalar(tensor));
+    GGML_ASSERT(tensor->type == GGML_TYPE_F32);
+    tensor->flags |= GGML_TENSOR_FLAG_LOSS;
+}
 
-            for (int j = 0; j < QK5_0; j += 2) {
-                const uint8_t vh0 = ((qh & (1u << (j/2 + 0 ))) >> (j/2 + 0 )) << 4;
-                const uint8_t vh1 = ((qh & (1u << (j/2 + 16))) >> (j/2 + 12));
+////////////////////////////////////////////////////////////////////////////////
 
-                // cast to 16 bins
-                const uint8_t vi0 = ((y[i].qs[j/2] & 0x0F) | vh0) / 2;
-                const uint8_t vi1 = ((y[i].qs[j/2] >>   4) | vh1) / 2;
+void ggml_quantize_init(enum ggml_type type) {
+    ggml_critical_section_start();
 
-                hist[vi0]++;
-                hist[vi1]++;
-            }
-        }
+    switch (type) {
+        case GGML_TYPE_IQ2_XXS:
+        case GGML_TYPE_IQ2_XS:
+        case GGML_TYPE_IQ2_S:
+        case GGML_TYPE_IQ1_S:
+        case GGML_TYPE_IQ1_M:   iq2xs_init_impl(type); break;
+        case GGML_TYPE_IQ3_XXS: iq3xs_init_impl(256); break;
+        case GGML_TYPE_IQ3_S:   iq3xs_init_impl(512); break;
+        default: // nothing
+            break;
     }
 
-    return (n/QK5_0*sizeof(block_q5_0));
+    ggml_critical_section_end();
 }
 
-size_t ggml_quantize_q5_1(const float * src, void * dst, int n, int k, int64_t * hist) {
-    assert(k % QK5_1 == 0);
-    const int nb = k / QK5_1;
-
-    for (int b = 0; b < n; b += k) {
-        block_q5_1 * restrict y = (block_q5_1 *)dst + b/QK5_1;
-
-        quantize_row_q5_1_reference(src + b, y, k);
-
-        for (int i = 0; i < nb; i++) {
-            uint32_t qh;
-            memcpy(&qh, &y[i].qh, sizeof(qh));
-
-            for (int j = 0; j < QK5_1; j += 2) {
-                const uint8_t vh0 = ((qh & (1u << (j/2 + 0 ))) >> (j/2 + 0 )) << 4;
-                const uint8_t vh1 = ((qh & (1u << (j/2 + 16))) >> (j/2 + 12));
-
-                // cast to 16 bins
-                const uint8_t vi0 = ((y[i].qs[j/2] & 0x0F) | vh0) / 2;
-                const uint8_t vi1 = ((y[i].qs[j/2] >>   4) | vh1) / 2;
+void ggml_quantize_free(void) {
+    ggml_critical_section_start();
 
-                hist[vi0]++;
-                hist[vi1]++;
-            }
-        }
-    }
+    iq2xs_free_impl(GGML_TYPE_IQ2_XXS);
+    iq2xs_free_impl(GGML_TYPE_IQ2_XS);
+    iq2xs_free_impl(GGML_TYPE_IQ1_S);
+    iq3xs_free_impl(256);
 
-    return (n/QK5_1*sizeof(block_q5_1));
+    ggml_critical_section_end();
 }
 
-size_t ggml_quantize_q8_0(const float * src, void * dst, int n, int k, int64_t * hist) {
-    assert(k % QK8_0 == 0);
-    const int nb = k / QK8_0;
+bool ggml_quantize_requires_imatrix(enum ggml_type type) {
+    return
+        type == GGML_TYPE_IQ2_XXS ||
+        type == GGML_TYPE_IQ2_XS  ||
+        type == GGML_TYPE_IQ1_S;//   ||
+        //type == GGML_TYPE_IQ1_M;
+}
 
-    for (int b = 0; b < n; b += k) {
-        block_q8_0 * restrict y = (block_q8_0 *)dst + b/QK8_0;
+size_t ggml_quantize_chunk(
+        enum ggml_type   type,
+           const float * src,
+                  void * dst,
+               int64_t   start,
+               int64_t   nrows,
+               int64_t   n_per_row,
+           const float * imatrix) {
+    const int64_t n = (int64_t) nrows * n_per_row;
 
-        quantize_row_q8_0_reference(src + b, y, k);
+    if (ggml_quantize_requires_imatrix(type)) {
+        GGML_ASSERT(imatrix != NULL);
+    }
 
-        for (int i = 0; i < nb; i++) {
-            for (int j = 0; j < QK8_0; ++j) {
-                const int8_t vi = y[i].qs[j];
+    GGML_ASSERT(start % type_traits[type].blck_size == 0);
+    GGML_ASSERT(start % n_per_row == 0);
 
-                hist[vi/16 + 8]++;
-            }
-        }
-    }
+    ggml_quantize_init(type); // this is noop if already initialized
 
-    return (n/QK8_0*sizeof(block_q8_0));
-}
+    const size_t start_row = start / n_per_row;
+    const size_t row_size  = ggml_row_size(type, n_per_row);
 
-size_t ggml_quantize_chunk(enum ggml_type type, const float * src, void * dst, int start, int n, int64_t * hist) {
     size_t result = 0;
+
     switch (type) {
-        case GGML_TYPE_Q4_0:
-            {
-                GGML_ASSERT(start % QK4_0 == 0);
-                block_q4_0 * block = (block_q4_0*)dst + start / QK4_0;
-                result = ggml_quantize_q4_0(src + start, block, n, n, hist);
-            } break;
-        case GGML_TYPE_Q4_1:
-            {
-                GGML_ASSERT(start % QK4_1 == 0);
-                block_q4_1 * block = (block_q4_1*)dst + start / QK4_1;
-                result = ggml_quantize_q4_1(src + start, block, n, n, hist);
-            } break;
-        case GGML_TYPE_Q5_0:
-            {
-                GGML_ASSERT(start % QK5_0 == 0);
-                block_q5_0 * block = (block_q5_0*)dst + start / QK5_0;
-                result = ggml_quantize_q5_0(src + start, block, n, n, hist);
-            } break;
-        case GGML_TYPE_Q5_1:
-            {
-                GGML_ASSERT(start % QK5_1 == 0);
-                block_q5_1 * block = (block_q5_1*)dst + start / QK5_1;
-                result = ggml_quantize_q5_1(src + start, block, n, n, hist);
-            } break;
-        case GGML_TYPE_Q8_0:
-            {
-                GGML_ASSERT(start % QK8_0 == 0);
-                block_q8_0 * block = (block_q8_0*)dst + start / QK8_0;
-                result = ggml_quantize_q8_0(src + start, block, n, n, hist);
-            } break;
-        case GGML_TYPE_Q2_K:
-            {
-                GGML_ASSERT(start % QK_K == 0);
-                block_q2_K * block = (block_q2_K*)dst + start / QK_K;
-                result = ggml_quantize_q2_K(src + start, block, n, n, hist);
-            } break;
-        case GGML_TYPE_Q3_K:
-            {
-                GGML_ASSERT(start % QK_K == 0);
-                block_q3_K * block = (block_q3_K*)dst + start / QK_K;
-                result = ggml_quantize_q3_K(src + start, block, n, n, hist);
-            } break;
-        case GGML_TYPE_Q4_K:
-            {
-                GGML_ASSERT(start % QK_K == 0);
-                block_q4_K * block = (block_q4_K*)dst + start / QK_K;
-                result = ggml_quantize_q4_K(src + start, block, n, n, hist);
-            } break;
-        case GGML_TYPE_Q5_K:
-            {
-                GGML_ASSERT(start % QK_K == 0);
-                block_q5_K * block = (block_q5_K*)dst + start / QK_K;
-                result = ggml_quantize_q5_K(src + start, block, n, n, hist);
-            } break;
-        case GGML_TYPE_Q6_K:
-            {
-                GGML_ASSERT(start % QK_K == 0);
-                block_q6_K * block = (block_q6_K*)dst + start / QK_K;
-                result = ggml_quantize_q6_K(src + start, block, n, n, hist);
-            } break;
+        case GGML_TYPE_Q4_0:    result = quantize_q4_0(src + start, (char *) dst + start_row * row_size, nrows, n_per_row, imatrix); break;
+        case GGML_TYPE_Q4_1:    result = quantize_q4_1(src + start, (char *) dst + start_row * row_size, nrows, n_per_row, imatrix); break;
+        case GGML_TYPE_Q5_0:    result = quantize_q5_0(src + start, (char *) dst + start_row * row_size, nrows, n_per_row, imatrix); break;
+        case GGML_TYPE_Q5_1:    result = quantize_q5_1(src + start, (char *) dst + start_row * row_size, nrows, n_per_row, imatrix); break;
+        case GGML_TYPE_Q8_0:    result = quantize_q8_0(src + start, (char *) dst + start_row * row_size, nrows, n_per_row, imatrix); break;
+        case GGML_TYPE_Q2_K:    result = quantize_q2_K(src + start, (char *) dst + start_row * row_size, nrows, n_per_row, imatrix); break;
+        case GGML_TYPE_Q3_K:    result = quantize_q3_K(src + start, (char *) dst + start_row * row_size, nrows, n_per_row, imatrix); break;
+        case GGML_TYPE_Q4_K:    result = quantize_q4_K(src + start, (char *) dst + start_row * row_size, nrows, n_per_row, imatrix); break;
+        case GGML_TYPE_Q5_K:    result = quantize_q5_K(src + start, (char *) dst + start_row * row_size, nrows, n_per_row, imatrix); break;
+        case GGML_TYPE_Q6_K:    result = quantize_q6_K(src + start, (char *) dst + start_row * row_size, nrows, n_per_row, imatrix); break;
+        case GGML_TYPE_TQ1_0:   result = quantize_tq1_0(src + start, (char *) dst + start_row * row_size, nrows, n_per_row, imatrix); break;
+        case GGML_TYPE_TQ2_0:   result = quantize_tq2_0(src + start, (char *) dst + start_row * row_size, nrows, n_per_row, imatrix); break;
+        case GGML_TYPE_IQ2_XXS: result = quantize_iq2_xxs(src + start, (char *) dst + start_row * row_size, nrows, n_per_row, imatrix); break;
+        case GGML_TYPE_IQ2_XS:  result = quantize_iq2_xs (src + start, (char *) dst + start_row * row_size, nrows, n_per_row, imatrix); break;
+        case GGML_TYPE_IQ3_XXS: result = quantize_iq3_xxs(src + start, (char *) dst + start_row * row_size, nrows, n_per_row, imatrix); break;
+        case GGML_TYPE_IQ3_S:   result = quantize_iq3_s  (src + start, (char *) dst + start_row * row_size, nrows, n_per_row, imatrix); break;
+        case GGML_TYPE_IQ2_S:   result = quantize_iq2_s  (src + start, (char *) dst + start_row * row_size, nrows, n_per_row, imatrix); break;
+        case GGML_TYPE_IQ1_S:   result = quantize_iq1_s  (src + start, (char *) dst + start_row * row_size, nrows, n_per_row, imatrix); break;
+        case GGML_TYPE_IQ1_M:   result = quantize_iq1_m  (src + start, (char *) dst + start_row * row_size, nrows, n_per_row, imatrix); break;
+        case GGML_TYPE_IQ4_NL:  result = quantize_iq4_nl (src + start, (char *) dst + start_row * row_size, nrows, n_per_row, imatrix); break;
+        case GGML_TYPE_IQ4_XS:  result = quantize_iq4_xs (src + start, (char *) dst + start_row * row_size, nrows, n_per_row, imatrix); break;
+        case GGML_TYPE_Q4_0_4_4: result = quantize_q4_0_4x4(src + start, (char *) dst + start_row * row_size, nrows, n_per_row, imatrix); break;
+        case GGML_TYPE_Q4_0_4_8: result = quantize_q4_0_4x8(src + start, (char *) dst + start_row * row_size, nrows, n_per_row, imatrix); break;
+        case GGML_TYPE_Q4_0_8_8: result = quantize_q4_0_8x8(src + start, (char *) dst + start_row * row_size, nrows, n_per_row, imatrix); break;
         case GGML_TYPE_F16:
             {
-                int elemsize = sizeof(ggml_fp16_t);
+                size_t elemsize = sizeof(ggml_fp16_t);
                 ggml_fp32_to_fp16_row(src + start, (ggml_fp16_t *)dst + start, n);
                 result = n * elemsize;
             } break;
+        case GGML_TYPE_BF16:
+            {
+                size_t elemsize = sizeof(ggml_bf16_t);
+                ggml_fp32_to_bf16_row_ref(src + start, (ggml_bf16_t *)dst + start, n);
+                result = n * elemsize;
+            } break;
         case GGML_TYPE_F32:
             {
-                int elemsize = sizeof(float);
+                size_t elemsize = sizeof(float);
                 result = n * elemsize;
                 memcpy((uint8_t *)dst + start * elemsize, src + start, result);
             } break;
         default:
             assert(false);
     }
+
+    GGML_ASSERT(result == nrows * row_size);
+
     return result;
 }
 
@@ -18789,6 +22065,25 @@ struct gguf_context {
     void * data;
 };
 
+static size_t gguf_type_size(enum gguf_type type) {
+    GGML_ASSERT(0 <= type && type < GGUF_TYPE_COUNT);
+    return GGUF_TYPE_SIZE[type];
+}
+
+static void gguf_tensor_info_sanitize(struct gguf_tensor_info * info) {
+    GGML_ASSERT(info->n_dims <= GGML_MAX_DIMS);
+    GGML_ASSERT(0 <= info->type && info->type < GGML_TYPE_COUNT);
+
+    for (uint32_t i = 0; i < info->n_dims; ++i) {
+        GGML_ASSERT(info->ne[i] > 0);
+    }
+
+    // prevent overflow for total number of elements
+    GGML_ASSERT(INT64_MAX/info->ne[1] > info->ne[0]);
+    GGML_ASSERT(INT64_MAX/info->ne[2] > info->ne[0]*info->ne[1]);
+    GGML_ASSERT(INT64_MAX/info->ne[3] > info->ne[0]*info->ne[1]*info->ne[2]);
+}
+
 static bool gguf_fread_el(FILE * file, void * dst, size_t size, size_t * offset) {
     const size_t n = fread(dst, 1, size, file);
     *offset += n;
@@ -18801,14 +22096,49 @@ static bool gguf_fread_str(FILE * file, struct gguf_str * p, size_t * offset) {
 
     bool ok = true;
 
-    ok = ok && gguf_fread_el(file, &p->n,    sizeof(p->n), offset); p->data = calloc(p->n + 1, 1);
-    ok = ok && gguf_fread_el(file,  p->data, p->n,         offset);
+    ok = ok && gguf_fread_el(file, &p->n, sizeof(p->n), offset);
+
+    // early exit if string length is invalid, prevents from integer overflow
+    if (p->n == SIZE_MAX) {
+        fprintf(stderr, "%s: invalid string length (%" PRIu64 ")\n", __func__, p->n);
+        return false;
+    }
+
+    p->data = GGML_CALLOC(p->n + 1, 1);
+
+    ok = ok && gguf_fread_el(file,  p->data, p->n, offset);
 
     return ok;
 }
 
+static void gguf_free_kv(struct gguf_kv * kv) {
+    if (kv->key.data) {
+        GGML_FREE(kv->key.data);
+    }
+
+    if (kv->type == GGUF_TYPE_STRING) {
+        if (kv->value.str.data) {
+            GGML_FREE(kv->value.str.data);
+        }
+    }
+
+    if (kv->type == GGUF_TYPE_ARRAY) {
+        if (kv->value.arr.data) {
+            if (kv->value.arr.type == GGUF_TYPE_STRING) {
+                for (uint64_t j = 0; j < kv->value.arr.n; ++j) {
+                    struct gguf_str * str = &((struct gguf_str *) kv->value.arr.data)[j];
+                    if (str->data) {
+                        GGML_FREE(str->data);
+                    }
+                }
+            }
+            GGML_FREE(kv->value.arr.data);
+        }
+    }
+}
+
 struct gguf_context * gguf_init_empty(void) {
-    struct gguf_context * ctx = GGML_ALIGNED_MALLOC(sizeof(struct gguf_context));
+    struct gguf_context * ctx = GGML_CALLOC(1, sizeof(struct gguf_context));
 
     memcpy(ctx->header.magic, GGUF_MAGIC, sizeof(ctx->header.magic));
     ctx->header.version   = GGUF_VERSION;
@@ -18828,8 +22158,9 @@ struct gguf_context * gguf_init_empty(void) {
 }
 
 struct gguf_context * gguf_init_from_file(const char * fname, struct gguf_init_params params) {
-    FILE * file = fopen(fname, "rb");
+    FILE * file = ggml_fopen(fname, "rb");
     if (!file) {
+        fprintf(stderr, "%s: failed to open '%s': '%s'\n", __func__, fname, strerror(errno));
         return NULL;
     }
 
@@ -18844,7 +22175,7 @@ struct gguf_context * gguf_init_from_file(const char * fname, struct gguf_init_p
 
         for (uint32_t i = 0; i < sizeof(magic); i++) {
             if (magic[i] != GGUF_MAGIC[i]) {
-                Rprintf("%s: invalid magic characters '%c%c%c%c'\n", __func__, magic[0], magic[1], magic[2], magic[3]);
+                fprintf(stderr, "%s: invalid magic characters '%c%c%c%c'\n", __func__, magic[0], magic[1], magic[2], magic[3]);
                 fclose(file);
                 return NULL;
             }
@@ -18853,7 +22184,7 @@ struct gguf_context * gguf_init_from_file(const char * fname, struct gguf_init_p
 
     bool ok = true;
 
-    struct gguf_context * ctx = GGML_ALIGNED_MALLOC(sizeof(struct gguf_context));
+    struct gguf_context * ctx = GGML_CALLOC(1, sizeof(struct gguf_context));
 
     // read the header
     {
@@ -18868,14 +22199,20 @@ struct gguf_context * gguf_init_from_file(const char * fname, struct gguf_init_p
         ok = ok && gguf_fread_el(file, &ctx->header.n_kv,      sizeof(ctx->header.n_kv),      &offset);
 
         if (ctx->header.version == 1) {
-            Rprintf("%s: GGUFv1 is no longer supported. please use a more up-to-date version\n", __func__);
+            fprintf(stderr, "%s: GGUFv1 is no longer supported. please use a more up-to-date version\n", __func__);
             fclose(file);
             gguf_free(ctx);
             return NULL;
         }
 
+        // sanity-checks to prevent from integer/buffer overflows
+
+        ok = ok && (ctx->header.n_tensors < (SIZE_MAX/2)/sizeof(struct gguf_tensor_info));
+        ok = ok && (ctx->header.n_tensors < (SIZE_MAX/2)/ggml_tensor_overhead());
+        ok = ok && (ctx->header.n_kv      < (SIZE_MAX/2)/sizeof(struct gguf_kv));
+
         if (!ok) {
-            Rprintf("%s: failed to read header\n", __func__);
+            fprintf(stderr, "%s: failed to read header\n", __func__);
             fclose(file);
             gguf_free(ctx);
             return NULL;
@@ -18884,17 +22221,21 @@ struct gguf_context * gguf_init_from_file(const char * fname, struct gguf_init_p
 
     // read the kv pairs
     {
-        ctx->kv = malloc(ctx->header.n_kv * sizeof(struct gguf_kv));
+        const uint64_t n_kv = ctx->header.n_kv;
 
-        for (uint64_t i = 0; i < ctx->header.n_kv; ++i) {
+        // header.n_kv will hold the actual value of pairs that were successfully read in the loop below
+        ctx->header.n_kv = 0;
+        ctx->kv = GGML_CALLOC(n_kv, sizeof(struct gguf_kv));
+
+        for (uint64_t i = 0; i < n_kv; ++i) {
             struct gguf_kv * kv = &ctx->kv[i];
 
-            //Rprintf("%s: reading kv %d\n", __func__, i);
+            //fprintf(stderr, "%s: reading kv %d\n", __func__, i);
 
             ok = ok && gguf_fread_str(file, &kv->key,                    &offset);
             ok = ok && gguf_fread_el (file, &kv->type, sizeof(kv->type), &offset);
 
-            //Rprintf("%s: reading kv with key %s\n", __func__, kv->key.data);
+            //fprintf(stderr, "%s: reading kv with key %s\n", __func__, kv->key.data);
 
             switch (kv->type) {
                 case GGUF_TYPE_UINT8:   ok = ok && gguf_fread_el (file, &kv->value.uint8,   sizeof(kv->value.uint8),   &offset); break;
@@ -18912,7 +22253,7 @@ struct gguf_context * gguf_init_from_file(const char * fname, struct gguf_init_p
                 case GGUF_TYPE_ARRAY:
                     {
                         ok = ok && gguf_fread_el(file, &kv->value.arr.type, sizeof(kv->value.arr.type), &offset);
-                        ok = ok && gguf_fread_el(file, &kv->value.arr.n,    sizeof(kv->value.arr.n), &offset);
+                        ok = ok && gguf_fread_el(file, &kv->value.arr.n,    sizeof(kv->value.arr.n),    &offset);
 
                         switch (kv->value.arr.type) {
                             case GGUF_TYPE_UINT8:
@@ -18927,30 +22268,50 @@ struct gguf_context * gguf_init_from_file(const char * fname, struct gguf_init_p
                             case GGUF_TYPE_FLOAT64:
                             case GGUF_TYPE_BOOL:
                                 {
-                                    kv->value.arr.data = malloc(kv->value.arr.n * GGUF_TYPE_SIZE[kv->value.arr.type]);
-                                    ok = ok && gguf_fread_el(file, kv->value.arr.data, kv->value.arr.n * GGUF_TYPE_SIZE[kv->value.arr.type], &offset);
+                                    // prevent from integer overflow in the malloc below
+                                    if (kv->value.arr.n >= SIZE_MAX/gguf_type_size(kv->value.arr.type)) {
+                                        fprintf(stderr, "%s: array size is too large (%" PRIu64 ")\n", __func__, kv->value.arr.n);
+                                        fclose(file);
+                                        gguf_free(ctx);
+                                        return NULL;
+                                    }
+
+                                    kv->value.arr.data = GGML_CALLOC(kv->value.arr.n, gguf_type_size(kv->value.arr.type));
+
+                                    ok = ok && gguf_fread_el(file, kv->value.arr.data, kv->value.arr.n * gguf_type_size(kv->value.arr.type), &offset);
                                 } break;
                             case GGUF_TYPE_STRING:
                                 {
-                                    kv->value.arr.data = malloc(kv->value.arr.n * sizeof(struct gguf_str));
+                                    // prevent from integer overflow in the malloc below
+                                    if (kv->value.arr.n >= SIZE_MAX/sizeof(struct gguf_str)) {
+                                        fprintf(stderr, "%s: array size is too large (%" PRIu64 ")\n", __func__, kv->value.arr.n);
+                                        fclose(file);
+                                        gguf_free(ctx);
+                                        return NULL;
+                                    }
+
+                                    kv->value.arr.data = GGML_CALLOC(kv->value.arr.n, sizeof(struct gguf_str));
+
                                     for (uint64_t j = 0; j < kv->value.arr.n; ++j) {
                                         ok = ok && gguf_fread_str(file, &((struct gguf_str *) kv->value.arr.data)[j], &offset);
                                     }
                                 } break;
                             case GGUF_TYPE_ARRAY:
-                            case GGUF_TYPE_COUNT: GGML_ASSERT(false && "invalid type"); break;
+                            default: GGML_ABORT("invalid type");
                         }
                     } break;
-                case GGUF_TYPE_COUNT: GGML_ASSERT(false && "invalid type");
+                default: GGML_ABORT("invalid type");
             }
 
             if (!ok) {
                 break;
             }
+
+            ctx->header.n_kv++;
         }
 
         if (!ok) {
-            Rprintf("%s: failed to read key-value pairs\n", __func__);
+            fprintf(stderr, "%s: failed to read key-value pairs\n", __func__);
             fclose(file);
             gguf_free(ctx);
             return NULL;
@@ -18958,8 +22319,8 @@ struct gguf_context * gguf_init_from_file(const char * fname, struct gguf_init_p
     }
 
     // read the tensor infos
-    {
-        ctx->infos = malloc(ctx->header.n_tensors * sizeof(struct gguf_tensor_info));
+    if (ctx->header.n_tensors > 0) {
+        ctx->infos = GGML_CALLOC(ctx->header.n_tensors, sizeof(struct gguf_tensor_info));
 
         for (uint64_t i = 0; i < ctx->header.n_tensors; ++i) {
             struct gguf_tensor_info * info = &ctx->infos[i];
@@ -18970,14 +22331,29 @@ struct gguf_context * gguf_init_from_file(const char * fname, struct gguf_init_p
 
             ok = ok && gguf_fread_str(file, &info->name,                          &offset);
             ok = ok && gguf_fread_el (file, &info->n_dims, sizeof(info->n_dims),  &offset);
+
+            ok = ok && (info->n_dims <= GGML_MAX_DIMS);
+
             for (uint32_t j = 0; j < info->n_dims; ++j) {
                 ok = ok && gguf_fread_el(file, &info->ne[j], sizeof(info->ne[j]), &offset);
             }
+
             ok = ok && gguf_fread_el (file, &info->type,   sizeof(info->type),    &offset);
             ok = ok && gguf_fread_el (file, &info->offset, sizeof(info->offset),  &offset);
 
+            // TODO: return an error instead of crashing with GGML_ASSERT
+            gguf_tensor_info_sanitize(info);
+
+            // make sure there is no duplicated tensor names
+            for (uint64_t j = 0; j < i && ok; ++j) {
+                if (strcmp(info->name.data, ctx->infos[j].name.data) == 0) {
+                    fprintf(stderr, "%s: duplicated tensor name %s\n", __func__, info->name.data);
+                    ok = false;
+                }
+            }
+
             if (!ok) {
-                Rprintf("%s: failed to read tensor info\n", __func__);
+                fprintf(stderr, "%s: failed to read tensor info\n", __func__);
                 fclose(file);
                 gguf_free(ctx);
                 return NULL;
@@ -19017,9 +22393,9 @@ struct gguf_context * gguf_init_from_file(const char * fname, struct gguf_init_p
                 (int64_t) info->ne[2] *
                 (int64_t) info->ne[3];
 
-            if (ne % ggml_blck_size(info->type) != 0) {
-                Rprintf("%s: tensor '%s' number of elements (%" PRId64 ") is not a multiple of block size (%d)\n",
-                        __func__, info->name.data, ne, ggml_blck_size(info->type));
+            if (ggml_blck_size(info->type) == 0 || ne % ggml_blck_size(info->type) != 0) {
+                fprintf(stderr, "%s: tensor '%s' of type %d (%s) number of elements (%" PRId64 ") is not a multiple of block size (%" PRId64 ")\n",
+                        __func__, info->name.data, (int) info->type, ggml_type_name(info->type), ne, ggml_blck_size(info->type));
                 fclose(file);
                 gguf_free(ctx);
                 return NULL;
@@ -19050,6 +22426,12 @@ struct gguf_context * gguf_init_from_file(const char * fname, struct gguf_init_p
         };
 
         *params.ctx = ggml_init(pdata);
+        if (*params.ctx == NULL) {
+            fprintf(stderr, "%s: failed to initialize context\n", __func__);
+            fclose(file);
+            gguf_free(ctx);
+            return NULL;
+        }
 
         struct ggml_context * ctx_data = *params.ctx;
 
@@ -19064,7 +22446,7 @@ struct gguf_context * gguf_init_from_file(const char * fname, struct gguf_init_p
             ok = ok && gguf_fread_el(file, data->data, ctx->size, &offset);
 
             if (!ok) {
-                Rprintf("%s: failed to read tensor data\n", __func__);
+                fprintf(stderr, "%s: failed to read tensor data\n", __func__);
                 fclose(file);
                 ggml_free(ctx_data);
                 gguf_free(ctx);
@@ -19089,12 +22471,12 @@ struct gguf_context * gguf_init_from_file(const char * fname, struct gguf_init_p
 
             ok = ok && cur != NULL;
 
-            ggml_set_name(cur, ctx->infos[i].name.data);
-
             if (!ok) {
                 break;
             }
 
+            ggml_set_name(cur, ctx->infos[i].name.data);
+
             // point the data member to the appropriate location in the binary blob using the tensor infos
             if (!params.no_alloc) {
               //cur->data = (char *) data->data + ctx->infos[i].offset - ctx->offset; // offset from start of file
@@ -19103,7 +22485,7 @@ struct gguf_context * gguf_init_from_file(const char * fname, struct gguf_init_p
         }
 
         if (!ok) {
-            Rprintf("%s: failed to read the tensor data\n", __func__);
+            fprintf(stderr, "%s: failed to read the tensor data\n", __func__);
             fclose(file);
             ggml_free(ctx_data);
             gguf_free(ctx);
@@ -19125,50 +22507,26 @@ void gguf_free(struct gguf_context * ctx) {
 
     if (ctx->kv) {
         // free string memory - not great..
-        for (uint32_t i = 0; i < ctx->header.n_kv; ++i) {
-            struct gguf_kv * kv = &ctx->kv[i];
-
-            if (kv->key.data) {
-                free(kv->key.data);
-            }
-
-            if (kv->type == GGUF_TYPE_STRING) {
-                if (kv->value.str.data) {
-                    free(kv->value.str.data);
-                }
-            }
-
-            if (kv->type == GGUF_TYPE_ARRAY) {
-                if (kv->value.arr.data) {
-                    if (kv->value.arr.type == GGUF_TYPE_STRING) {
-                        for (uint32_t j = 0; j < kv->value.arr.n; ++j) {
-                            struct gguf_str * str = &((struct gguf_str *) kv->value.arr.data)[j];
-                            if (str->data) {
-                                free(str->data);
-                            }
-                        }
-                    }
-                    free(kv->value.arr.data);
-                }
-            }
+        for (uint64_t i = 0; i < ctx->header.n_kv; ++i) {
+            gguf_free_kv(&ctx->kv[i]);
         }
 
-        free(ctx->kv);
+        GGML_FREE(ctx->kv);
     }
 
     if (ctx->infos) {
-        for (uint32_t i = 0; i < ctx->header.n_tensors; ++i) {
+        for (uint64_t i = 0; i < ctx->header.n_tensors; ++i) {
             struct gguf_tensor_info * info = &ctx->infos[i];
 
             if (info->name.data) {
-                free(info->name.data);
+                GGML_FREE(info->name.data);
             }
         }
 
-        free(ctx->infos);
+        GGML_FREE(ctx->infos);
     }
 
-    GGML_ALIGNED_FREE(ctx);
+    GGML_FREE(ctx);
 }
 
 const char * gguf_type_name(enum gguf_type type) {
@@ -19375,6 +22733,19 @@ static int gguf_get_or_add_key(struct gguf_context * ctx, const char * key) {
     return n_kv;
 }
 
+void gguf_remove_key(struct gguf_context * ctx, const char * key) {
+    const int idx = gguf_find_key(ctx, key);
+    if (idx >= 0) {
+        const int n_kv = gguf_get_n_kv(ctx);
+        gguf_free_kv(&ctx->kv[idx]);
+        for (int i = idx; i < n_kv-1; ++i) {
+            ctx->kv[i] = ctx->kv[i+1];
+        }
+        ctx->kv = realloc(ctx->kv, (n_kv - 1) * sizeof(struct gguf_kv));
+        ctx->header.n_kv--;
+    }
+}
+
 void gguf_set_val_u8(struct gguf_context * ctx, const char * key, uint8_t val) {
     const int idx = gguf_get_or_add_key(ctx, key);
 
@@ -19466,8 +22837,8 @@ void gguf_set_arr_data(struct gguf_context * ctx, const char * key, enum gguf_ty
     ctx->kv[idx].type           = GGUF_TYPE_ARRAY;
     ctx->kv[idx].value.arr.type = type;
     ctx->kv[idx].value.arr.n    = n;
-    ctx->kv[idx].value.arr.data = malloc(n*GGUF_TYPE_SIZE[type]);
-    memcpy(ctx->kv[idx].value.arr.data, data, n*GGUF_TYPE_SIZE[type]);
+    ctx->kv[idx].value.arr.data = GGML_CALLOC(n, gguf_type_size(type));
+    memcpy(ctx->kv[idx].value.arr.data, data, n*gguf_type_size(type));
 }
 
 void gguf_set_arr_str(struct gguf_context * ctx, const char * key, const char ** data, int n) {
@@ -19476,7 +22847,7 @@ void gguf_set_arr_str(struct gguf_context * ctx, const char * key, const char **
     ctx->kv[idx].type           = GGUF_TYPE_ARRAY;
     ctx->kv[idx].value.arr.type = GGUF_TYPE_STRING;
     ctx->kv[idx].value.arr.n    = n;
-    ctx->kv[idx].value.arr.data = malloc(n*sizeof(struct gguf_str));
+    ctx->kv[idx].value.arr.data = GGML_CALLOC(n, sizeof(struct gguf_str));
     for (int i = 0; i < n; i++) {
         struct gguf_str * str = &((struct gguf_str *)ctx->kv[idx].value.arr.data)[i];
         str->n    = strlen(data[i]);
@@ -19503,19 +22874,19 @@ void gguf_set_kv(struct gguf_context * ctx, struct gguf_context * src) {
             case GGUF_TYPE_ARRAY:
                 {
                     if (src->kv[i].value.arr.type == GGUF_TYPE_STRING) {
-                        const char ** data = malloc(src->kv[i].value.arr.n*sizeof(char *));
+                        const char ** data = GGML_CALLOC(src->kv[i].value.arr.n, sizeof(char *));
                         for (uint32_t j = 0; j < src->kv[i].value.arr.n; j++) {
                             data[j] = ((struct gguf_str *)src->kv[i].value.arr.data)[j].data;
                         }
                         gguf_set_arr_str(ctx, src->kv[i].key.data, data, src->kv[i].value.arr.n);
-                        free((void *)data);
+                        GGML_FREE((void *)data);
                     } else if (src->kv[i].value.arr.type == GGUF_TYPE_ARRAY) {
-                        GGML_ASSERT(false && "nested arrays not supported");
+                        GGML_ABORT("nested arrays not supported");
                     } else {
                         gguf_set_arr_data(ctx, src->kv[i].key.data, src->kv[i].value.arr.type, src->kv[i].value.arr.data, src->kv[i].value.arr.n);
                     }
                 } break;
-            case GGUF_TYPE_COUNT:  GGML_ASSERT(false && "invalid type"); break;
+            default: GGML_ABORT("invalid type");
         }
     }
 }
@@ -19523,6 +22894,11 @@ void gguf_set_kv(struct gguf_context * ctx, struct gguf_context * src) {
 void gguf_add_tensor(
              struct gguf_context * ctx,
         const struct ggml_tensor * tensor) {
+    GGML_ASSERT(tensor);
+    if (gguf_find_tensor(ctx, tensor->name) != -1) {
+        GGML_ABORT("duplicated tensor name");
+    }
+
     const int idx = ctx->header.n_tensors;
     ctx->infos = realloc(ctx->infos, (idx + 1)*sizeof(struct gguf_tensor_info));
 
@@ -19553,7 +22929,7 @@ void gguf_add_tensor(
 void gguf_set_tensor_type(struct gguf_context * ctx, const char * name, enum ggml_type type) {
     const int idx = gguf_find_tensor(ctx, name);
     if (idx < 0) {
-        GGML_ASSERT(false && "tensor not found");
+        GGML_ABORT("tensor not found");
     }
 
     ctx->infos[idx].type = type;
@@ -19562,7 +22938,7 @@ void gguf_set_tensor_type(struct gguf_context * ctx, const char * name, enum ggm
 void gguf_set_tensor_data(struct gguf_context * ctx, const char * name, const void * data, size_t size) {
     const int idx = gguf_find_tensor(ctx, name);
     if (idx < 0) {
-        GGML_ASSERT(false && "tensor not found");
+        GGML_ABORT("tensor not found");
     }
 
     ctx->infos[idx].data = data;
@@ -19591,7 +22967,7 @@ struct gguf_buf {
 
 static struct gguf_buf gguf_buf_init(size_t size) {
     struct gguf_buf buf = {
-        /*buf.data   =*/ size == 0 ? NULL : malloc(size),
+        /*buf.data   =*/ size == 0 ? NULL : GGML_CALLOC(1, size),
         /*buf.size   =*/ size,
         /*buf.offset =*/ 0,
     };
@@ -19601,7 +22977,7 @@ static struct gguf_buf gguf_buf_init(size_t size) {
 
 static void gguf_buf_free(struct gguf_buf buf) {
     if (buf.data) {
-        free(buf.data);
+        GGML_FREE(buf.data);
     }
 }
 
@@ -19682,7 +23058,7 @@ static void gguf_write_to_buf(const struct gguf_context * ctx, struct gguf_buf *
                         case GGUF_TYPE_FLOAT64:
                         case GGUF_TYPE_BOOL:
                             {
-                                gguf_bwrite_el(buf, kv->value.arr.data, kv->value.arr.n * GGUF_TYPE_SIZE[kv->value.arr.type]);
+                                gguf_bwrite_el(buf, kv->value.arr.data, kv->value.arr.n * gguf_type_size(kv->value.arr.type));
                             } break;
                         case GGUF_TYPE_STRING:
                             {
@@ -19691,10 +23067,10 @@ static void gguf_write_to_buf(const struct gguf_context * ctx, struct gguf_buf *
                                 }
                             } break;
                         case GGUF_TYPE_ARRAY:
-                        case GGUF_TYPE_COUNT: GGML_ASSERT(false && "invalid type"); break;
+                        default: GGML_ABORT("invalid type");
                     }
                 } break;
-            case GGUF_TYPE_COUNT: GGML_ASSERT(false && "invalid type");
+            default: GGML_ABORT("invalid type");
         }
     }
 
@@ -19753,9 +23129,9 @@ static void gguf_write_to_buf(const struct gguf_context * ctx, struct gguf_buf *
 }
 
 void gguf_write_to_file(const struct gguf_context * ctx, const char * fname, bool only_meta) {
-    FILE * file = fopen(fname, "wb");
+    FILE * file = ggml_fopen(fname, "wb");
     if (!file) {
-        GGML_ASSERT(false && "failed to open file for writing");
+        GGML_ABORT("failed to open file for writing");
     }
 
     struct gguf_buf buf = gguf_buf_init(16*1024);
@@ -19838,6 +23214,14 @@ int ggml_cpu_has_avx512_vnni(void) {
 #endif
 }
 
+int ggml_cpu_has_avx512_bf16(void) {
+#if defined(__AVX512BF16__)
+    return 1;
+#else
+    return 0;
+#endif
+}
+
 int ggml_cpu_has_fma(void) {
 #if defined(__FMA__)
     return 1;
@@ -19847,8 +23231,16 @@ int ggml_cpu_has_fma(void) {
 }
 
 int ggml_cpu_has_neon(void) {
-#if defined(__ARM_NEON)
-    return 1;
+#if defined(__ARM_ARCH)
+    return ggml_arm_arch_features.has_neon;
+#else
+    return 0;
+#endif
+}
+
+int ggml_cpu_has_sve(void) {
+#if defined(__ARM_ARCH)
+    return ggml_arm_arch_features.has_sve;
 #else
     return 0;
 #endif
@@ -19862,6 +23254,14 @@ int ggml_cpu_has_arm_fma(void) {
 #endif
 }
 
+int ggml_cpu_has_riscv_v(void) {
+#if defined(__riscv_v_intrinsic)
+    return 1;
+#else
+    return 0;
+#endif
+}
+
 int ggml_cpu_has_metal(void) {
 #if defined(GGML_USE_METAL)
     return 1;
@@ -19895,23 +23295,63 @@ int ggml_cpu_has_wasm_simd(void) {
 }
 
 int ggml_cpu_has_blas(void) {
-#if defined(GGML_USE_ACCELERATE) || defined(GGML_USE_OPENBLAS) || defined(GGML_USE_CUBLAS) || defined(GGML_USE_CLBLAST)
+#if defined(GGML_USE_BLAS) || defined(GGML_USE_CUDA) || defined(GGML_USE_VULKAN) || defined(GGML_USE_SYCL)
+    return 1;
+#else
+    return 0;
+#endif
+}
+
+int ggml_cpu_has_cuda(void) {
+#if defined(GGML_USE_CUDA)
+    return 1;
+#else
+    return 0;
+#endif
+}
+
+int ggml_cpu_has_vulkan(void) {
+#if defined(GGML_USE_VULKAN)
+    return 1;
+#else
+    return 0;
+#endif
+}
+
+int ggml_cpu_has_kompute(void) {
+#if defined(GGML_USE_KOMPUTE)
+    return 1;
+#else
+    return 0;
+#endif
+}
+
+int ggml_cpu_has_sycl(void) {
+#if defined(GGML_USE_SYCL)
+    return 1;
+#else
+    return 0;
+#endif
+}
+
+int ggml_cpu_has_rpc(void) {
+#if defined(GGML_USE_RPC)
     return 1;
 #else
     return 0;
 #endif
 }
 
-int ggml_cpu_has_cublas(void) {
-#if defined(GGML_USE_CUBLAS)
+int ggml_cpu_has_cann(void) {
+#if defined(GGML_USE_CANN)
     return 1;
 #else
     return 0;
 #endif
 }
 
-int ggml_cpu_has_clblast(void) {
-#if defined(GGML_USE_CLBLAST)
+int ggml_cpu_has_llamafile(void) {
+#if defined(GGML_USE_LLAMAFILE)
     return 1;
 #else
     return 0;
@@ -19919,7 +23359,7 @@ int ggml_cpu_has_clblast(void) {
 }
 
 int ggml_cpu_has_gpublas(void) {
-    return ggml_cpu_has_cublas() || ggml_cpu_has_clblast();
+    return ggml_cpu_has_cuda() || ggml_cpu_has_vulkan() || ggml_cpu_has_kompute() || ggml_cpu_has_sycl();
 }
 
 int ggml_cpu_has_sse3(void) {
@@ -19946,4 +23386,24 @@ int ggml_cpu_has_vsx(void) {
 #endif
 }
 
+int ggml_cpu_has_matmul_int8(void) {
+#if defined(__ARM_ARCH)
+    return ggml_arm_arch_features.has_i8mm;
+#else
+    return 0;
+#endif
+}
+
+int ggml_cpu_get_sve_cnt(void) {
+#if defined(__ARM_ARCH)
+    return ggml_arm_arch_features.sve_cnt;
+#else
+    return 0;
+#endif
+}
+
+void ggml_log_set(ggml_log_callback log_callback, void * user_data) {
+    g_logger_state.log_callback = log_callback ? log_callback : ggml_log_callback_default;
+    g_logger_state.log_callback_user_data = user_data;
+}
 ////////////////////////////////////////////////////////////////////////////////
diff --git a/src/whisper_cpp/ggml/src/kompute-shaders/common.comp b/src/whisper_cpp/ggml/src/kompute-shaders/common.comp
new file mode 100644
index 00000000..62d62b02
--- /dev/null
+++ b/src/whisper_cpp/ggml/src/kompute-shaders/common.comp
@@ -0,0 +1,102 @@
+#extension GL_EXT_shader_16bit_storage: require
+#extension GL_EXT_shader_8bit_storage: require
+#extension GL_EXT_shader_explicit_arithmetic_types_float16: require
+#extension GL_EXT_shader_explicit_arithmetic_types_int8: require
+#extension GL_EXT_shader_explicit_arithmetic_types_int16: require
+#extension GL_EXT_control_flow_attributes: enable
+#extension GL_KHR_shader_subgroup_arithmetic : require
+#extension GL_EXT_debug_printf : enable
+
+#define QK4_0 32
+#define QK4_1 32
+
+#define GELU_COEF_A 0.044715
+#define SQRT_2_OVER_PI 0.79788456080286535587989211986876
+#define TWOPI_F 6.283185307179586f
+
+#define QK_K 256
+
+#define u8BufToU16(buf, idx) (((uint16_t(buf[idx + 1]) << 8)) | buf[idx])
+#define u8BufToFloat16(buf, idx) uint16BitsToHalf u8BufToU16(buf, idx)
+#define u8BufToU32(buf, idx) (((uint32_t u8BufToU16(buf, idx + 2) << 8 | buf[idx + 1]) << 8) | buf[idx])
+#define u8BufToFloat(buf, idx) uintBitsToFloat u8BufToU32(buf, idx)
+
+#define sizeof_block_q4_0 0x12
+struct block_q4_0 {
+    float16_t d;
+    uint8_t qs[QK4_0 / 2];
+};
+mat4 dequantize_q4_0(const block_q4_0 xb, uint il) {
+    const float d1 = il != 0 ? (xb.d / 16.f) : xb.d;
+    const float d2 = d1 / 256.f;
+    const float md = -8.f * xb.d;
+    const uint16_t mask0 = il != 0 ? uint16_t(0x00F0) : uint16_t(0x000F);
+    const uint16_t mask1 = mask0 << 8;
+
+    mat4 reg;
+    for (int i=0;i<8;i++) {
+        uint16_t b = (uint16_t(xb.qs[2 * i + 1]) << 8) | uint16_t(xb.qs[2 * i]);
+        reg[i/2][2*(i%2)+0] = d1 * (b & mask0) + md;
+        reg[i/2][2*(i%2)+1] = d2 * (b & mask1) + md;
+    }
+    return reg;
+}
+
+#define sizeof_block_q4_1 0x14
+struct block_q4_1 {
+    float16_t d;
+    float16_t m;
+    uint8_t qs[QK4_1 / 2];
+};
+mat4 dequantize_q4_1(const block_q4_1 xb, uint il) {
+    const float d1 = il != 0 ? (xb.d / 16.f) : xb.d;
+    const float d2 = d1 / 256.f;
+    const float  m = xb.m;
+    const uint16_t mask0 = il != 0 ? uint16_t(0x00F0) : uint16_t(0x000F);
+    const uint16_t mask1 = mask0 << 8;
+
+    mat4 reg;
+    for (int i=0;i<8;i++) {
+        uint16_t b = (uint16_t(xb.qs[2 * i + 1]) << 8) | uint16_t(xb.qs[2 * i]);
+        reg[i/2][2*(i%2)+0] = ((b & mask0) * d1) + m;
+        reg[i/2][2*(i%2)+1] = ((b & mask1) * d2) + m;
+    }
+    return reg;
+}
+
+#define sizeof_block_q6_k 210
+struct block_q6_k {
+    uint8_t ql[QK_K/2];      // quants, lower 4 bits
+    uint8_t qh[QK_K/4];      // quants, upper 2 bits
+    int8_t  scales[QK_K/16]; // scales, quantized with 8 bits
+    float16_t d;             // super-block scale
+};
+mat4 dequantize_q6_k(const block_q6_k xb, uint il) {
+    const float16_t d_all = xb.d;
+
+    const uint qlIndex = 64*(il/8) + 32*((il/2)&1) + 16*(il&1);
+    const uint qhIndex = 32*(il/8) + 16*(il&1);
+    float16_t sc = xb.scales[(il%2) + 2 * ((il/2))];
+    il = (il/2) & 3;
+
+    const uint16_t  kmask1 = il>1 ? uint16_t(il>2 ? 192 : 48) : uint16_t(il>0 ? 12 : 3);
+    const uint16_t  kmask2 = il>1 ? uint8_t(0xF0)             : uint8_t(0x0F);
+    const float16_t coef   = il>1 ? float16_t(1.f/16.f)       : float16_t(1.f);
+    const float16_t ml = float16_t(d_all * sc * 32.f);
+    const float16_t dl = float16_t(d_all * sc * coef);
+    mat4 reg;
+    for (int i = 0; i < 16; ++i) {
+        const float16_t q = (il&1) != 0 ? ((xb.ql[qlIndex + i] & kmask2) | ((xb.qh[qhIndex + i] & kmask1) << 2))
+                                        : ((xb.ql[qlIndex + i] & kmask2) | ((xb.qh[qhIndex + i] & kmask1) << 4));
+        reg[i/4][i%4] = dl * q - ml;
+    }
+    return reg;
+}
+
+
+#define QK8_0 32
+// struct block_q8_0 {
+//     float16_t d;         // delta
+//     int8_t    qs[QK8_0]; // quants
+// };
+#define sizeof_block_q8_0 34
diff --git a/src/whisper_cpp/ggml/src/kompute-shaders/op_add.comp b/src/whisper_cpp/ggml/src/kompute-shaders/op_add.comp
new file mode 100644
index 00000000..b7b76a79
--- /dev/null
+++ b/src/whisper_cpp/ggml/src/kompute-shaders/op_add.comp
@@ -0,0 +1,58 @@
+#version 450
+
+#include "common.comp"
+
+layout(local_size_x = 1024) in;
+
+layout(binding = 0) buffer restrict readonly tensorInA { float inA[]; };
+layout(binding = 1) buffer restrict readonly tensorInB { float inB[]; };
+layout(binding = 2) buffer restrict writeonly tensorOut { float out_[]; };
+
+layout(push_constant) uniform PushConstants {
+    uint inAOff;
+    uint inBOff;
+    uint outOff;
+    int ne00;
+    int nb00;
+    int nb01;
+    int nb02;
+    int nb03;
+    int ne10;
+    int ne11;
+    int ne12;
+    int ne13;
+    int nb10;
+    int nb11;
+    int nb12;
+    int nb13;
+    int ne0;
+    int nb0;
+    int nb1;
+    int nb2;
+    int nb3;
+  //int offs; // TODO: needed for GGML_OP_ACC, see metal code
+} pcs;
+
+// general-purpose kernel for addition of two tensors
+// pros: works for non-contiguous tensors, supports broadcast across dims 1, 2 and 3
+// cons: not very efficient
+void main() {
+    const uint i03 = gl_WorkGroupID.z;
+    const uint i02 = gl_WorkGroupID.y;
+    const uint i01 = gl_WorkGroupID.x;
+
+    const uint i13 = i03 % pcs.ne13;
+    const uint i12 = i02 % pcs.ne12;
+    const uint i11 = i01 % pcs.ne11;
+
+    int offs = 0; // TMP (see above)
+
+    uint src0_off = uint((i03*pcs.nb03 + i02*pcs.nb02 + i01*pcs.nb01 + offs) / 4);
+    uint src1_off = uint((i13*pcs.nb13 + i12*pcs.nb12 + i11*pcs.nb11       ) / 4);
+    uint dst_off  = uint((i03*pcs.nb3  + i02*pcs.nb2  + i01*pcs.nb1  + offs) / 4);
+
+    for (uint i0 = gl_LocalInvocationID.x; i0 < pcs.ne0; i0 += gl_WorkGroupSize.x) {
+        const uint i10 = i0 % pcs.ne10;
+        out_[pcs.outOff + dst_off + i0] = inA[pcs.inAOff + src0_off + i0] + inB[pcs.inBOff + src1_off + i10];
+    }
+}
diff --git a/src/whisper_cpp/ggml/src/kompute-shaders/op_addrow.comp b/src/whisper_cpp/ggml/src/kompute-shaders/op_addrow.comp
new file mode 100644
index 00000000..2376a6b8
--- /dev/null
+++ b/src/whisper_cpp/ggml/src/kompute-shaders/op_addrow.comp
@@ -0,0 +1,25 @@
+#version 450
+
+#include "common.comp"
+
+layout(local_size_x = 1) in;
+
+layout(binding = 0) buffer restrict readonly tensorInA { float inA[]; };
+layout(binding = 1) buffer restrict readonly tensorInB { float inB[]; };
+layout(binding = 2) buffer restrict writeonly tensorOut { float out_[]; };
+
+layout(push_constant) uniform PushConstants {
+    uint inAOff;
+    uint inBOff;
+    uint outOff;
+    uint row;
+} pcs;
+
+void main() {
+    const uint baseIndex = gl_WorkGroupID.x * 4;
+
+    for (uint x = 0; x < 4; x++) {
+        const uint i = baseIndex + x;
+        out_[i + pcs.outOff] = inA[i + pcs.inAOff] + inB[(i % pcs.row) + pcs.inBOff];
+    }
+}
diff --git a/src/whisper_cpp/ggml/src/kompute-shaders/op_cpy_f16_f16.comp b/src/whisper_cpp/ggml/src/kompute-shaders/op_cpy_f16_f16.comp
new file mode 100644
index 00000000..d57247d2
--- /dev/null
+++ b/src/whisper_cpp/ggml/src/kompute-shaders/op_cpy_f16_f16.comp
@@ -0,0 +1,52 @@
+#version 450
+
+#include "common.comp"
+
+#define IN_TYPE float16_t
+#define IN_TYPE_SIZE 2
+#define OUT_TYPE float16_t
+#define OUT_TYPE_SIZE 2
+
+layout(local_size_x = 1024) in;
+
+layout (binding = 0) readonly buffer tensorIn { IN_TYPE in_[]; };
+layout (binding = 1) writeonly buffer tensorOut { OUT_TYPE out_[]; };
+
+layout (push_constant) uniform parameter {
+    uint inOff;
+    uint outOff;
+    int ne00;
+    int ne01;
+    int ne02;
+    uint nb00;
+    uint nb01;
+    uint nb02;
+    uint nb03;
+    int ne0;
+    int ne1;
+    int ne2;
+    uint nb0;
+    uint nb1;
+    uint nb2;
+    uint nb3;
+} pcs;
+
+void main() {
+    const uint i03 = gl_WorkGroupID.z;
+    const uint i02 = gl_WorkGroupID.y;
+    const uint i01 = gl_WorkGroupID.x;
+
+    const int n = int(i03)*pcs.ne02*pcs.ne01*pcs.ne00 + int(i02)*pcs.ne01*pcs.ne00 + int(i01)*pcs.ne00;
+
+    const int i3 = n / (pcs.ne2*pcs.ne1*pcs.ne0);
+    const int i2 = (n - i3*pcs.ne2*pcs.ne1*pcs.ne0) / (pcs.ne1*pcs.ne0);
+    const int i1 = (n - i3*pcs.ne2*pcs.ne1*pcs.ne0 - i2*pcs.ne1*pcs.ne0) / pcs.ne0;
+    const int i0 = (n - i3*pcs.ne2*pcs.ne1*pcs.ne0 - i2*pcs.ne1*pcs.ne0 - i1*pcs.ne0);
+
+    const uint dst_data = (i3*pcs.nb3 + i2*pcs.nb2 + i1*pcs.nb1 + i0*pcs.nb0) / OUT_TYPE_SIZE + pcs.outOff; // Based from out_
+
+    for (uint i00 = gl_LocalInvocationID.x; i00 < pcs.ne00; i00 += gl_WorkGroupSize.x) {
+        const uint src = uint((i03*pcs.nb03 + i02*pcs.nb02 + i01*pcs.nb01 + i00*pcs.nb00) / IN_TYPE_SIZE) + pcs.inOff; // Based from in_
+        out_[dst_data+i00] = OUT_TYPE(in_[src]);
+    }
+}
diff --git a/src/whisper_cpp/ggml/src/kompute-shaders/op_cpy_f16_f32.comp b/src/whisper_cpp/ggml/src/kompute-shaders/op_cpy_f16_f32.comp
new file mode 100644
index 00000000..b568bcd7
--- /dev/null
+++ b/src/whisper_cpp/ggml/src/kompute-shaders/op_cpy_f16_f32.comp
@@ -0,0 +1,52 @@
+#version 450
+
+#include "common.comp"
+
+#define IN_TYPE float16_t
+#define IN_TYPE_SIZE 2
+#define OUT_TYPE float
+#define OUT_TYPE_SIZE 4
+
+layout(local_size_x = 1024) in;
+
+layout (binding = 0) readonly buffer tensorIn { IN_TYPE in_[]; };
+layout (binding = 1) writeonly buffer tensorOut { OUT_TYPE out_[]; };
+
+layout (push_constant) uniform parameter {
+    uint inOff;
+    uint outOff;
+    int ne00;
+    int ne01;
+    int ne02;
+    uint nb00;
+    uint nb01;
+    uint nb02;
+    uint nb03;
+    int ne0;
+    int ne1;
+    int ne2;
+    uint nb0;
+    uint nb1;
+    uint nb2;
+    uint nb3;
+} pcs;
+
+void main() {
+    const uint i03 = gl_WorkGroupID.z;
+    const uint i02 = gl_WorkGroupID.y;
+    const uint i01 = gl_WorkGroupID.x;
+
+    const int n = int(i03)*pcs.ne02*pcs.ne01*pcs.ne00 + int(i02)*pcs.ne01*pcs.ne00 + int(i01)*pcs.ne00;
+
+    const int i3 = n / (pcs.ne2*pcs.ne1*pcs.ne0);
+    const int i2 = (n - i3*pcs.ne2*pcs.ne1*pcs.ne0) / (pcs.ne1*pcs.ne0);
+    const int i1 = (n - i3*pcs.ne2*pcs.ne1*pcs.ne0 - i2*pcs.ne1*pcs.ne0) / pcs.ne0;
+    const int i0 = (n - i3*pcs.ne2*pcs.ne1*pcs.ne0 - i2*pcs.ne1*pcs.ne0 - i1*pcs.ne0);
+
+    const uint dst_data = (i3*pcs.nb3 + i2*pcs.nb2 + i1*pcs.nb1 + i0*pcs.nb0) / OUT_TYPE_SIZE + pcs.outOff; // Based from out_
+
+    for (uint i00 = gl_LocalInvocationID.x; i00 < pcs.ne00; i00 += gl_WorkGroupSize.x) {
+        const uint src = uint((i03*pcs.nb03 + i02*pcs.nb02 + i01*pcs.nb01 + i00*pcs.nb00) / IN_TYPE_SIZE) + pcs.inOff; // Based from in_
+        out_[dst_data+i00] = OUT_TYPE(in_[src]);
+    }
+}
diff --git a/src/whisper_cpp/ggml/src/kompute-shaders/op_cpy_f32_f16.comp b/src/whisper_cpp/ggml/src/kompute-shaders/op_cpy_f32_f16.comp
new file mode 100644
index 00000000..99b22834
--- /dev/null
+++ b/src/whisper_cpp/ggml/src/kompute-shaders/op_cpy_f32_f16.comp
@@ -0,0 +1,52 @@
+#version 450
+
+#include "common.comp"
+
+#define IN_TYPE float
+#define IN_TYPE_SIZE 4
+#define OUT_TYPE float16_t
+#define OUT_TYPE_SIZE 2
+
+layout(local_size_x = 1024) in;
+
+layout (binding = 0) readonly buffer tensorIn { IN_TYPE in_[]; };
+layout (binding = 1) writeonly buffer tensorOut { OUT_TYPE out_[]; };
+
+layout (push_constant) uniform parameter {
+    uint inOff;
+    uint outOff;
+    int ne00;
+    int ne01;
+    int ne02;
+    uint nb00;
+    uint nb01;
+    uint nb02;
+    uint nb03;
+    int ne0;
+    int ne1;
+    int ne2;
+    uint nb0;
+    uint nb1;
+    uint nb2;
+    uint nb3;
+} pcs;
+
+void main() {
+    const uint i03 = gl_WorkGroupID.z;
+    const uint i02 = gl_WorkGroupID.y;
+    const uint i01 = gl_WorkGroupID.x;
+
+    const int n = int(i03)*pcs.ne02*pcs.ne01*pcs.ne00 + int(i02)*pcs.ne01*pcs.ne00 + int(i01)*pcs.ne00;
+
+    const int i3 = n / (pcs.ne2*pcs.ne1*pcs.ne0);
+    const int i2 = (n - i3*pcs.ne2*pcs.ne1*pcs.ne0) / (pcs.ne1*pcs.ne0);
+    const int i1 = (n - i3*pcs.ne2*pcs.ne1*pcs.ne0 - i2*pcs.ne1*pcs.ne0) / pcs.ne0;
+    const int i0 = (n - i3*pcs.ne2*pcs.ne1*pcs.ne0 - i2*pcs.ne1*pcs.ne0 - i1*pcs.ne0);
+
+    const uint dst_data = (i3*pcs.nb3 + i2*pcs.nb2 + i1*pcs.nb1 + i0*pcs.nb0) / OUT_TYPE_SIZE + pcs.outOff; // Based from out_
+
+    for (uint i00 = gl_LocalInvocationID.x; i00 < pcs.ne00; i00 += gl_WorkGroupSize.x) {
+        const uint src = uint((i03*pcs.nb03 + i02*pcs.nb02 + i01*pcs.nb01 + i00*pcs.nb00) / IN_TYPE_SIZE) + pcs.inOff; // Based from in_
+        out_[dst_data+i00] = OUT_TYPE(in_[src]);
+    }
+}
diff --git a/src/whisper_cpp/ggml/src/kompute-shaders/op_cpy_f32_f32.comp b/src/whisper_cpp/ggml/src/kompute-shaders/op_cpy_f32_f32.comp
new file mode 100644
index 00000000..2fc99849
--- /dev/null
+++ b/src/whisper_cpp/ggml/src/kompute-shaders/op_cpy_f32_f32.comp
@@ -0,0 +1,52 @@
+#version 450
+
+#include "common.comp"
+
+#define IN_TYPE float
+#define IN_TYPE_SIZE 4
+#define OUT_TYPE float
+#define OUT_TYPE_SIZE 4
+
+layout(local_size_x = 1024) in;
+
+layout (binding = 0) readonly buffer tensorIn { IN_TYPE in_[]; };
+layout (binding = 1) writeonly buffer tensorOut { OUT_TYPE out_[]; };
+
+layout (push_constant) uniform parameter {
+    uint inOff;
+    uint outOff;
+    int ne00;
+    int ne01;
+    int ne02;
+    uint nb00;
+    uint nb01;
+    uint nb02;
+    uint nb03;
+    int ne0;
+    int ne1;
+    int ne2;
+    uint nb0;
+    uint nb1;
+    uint nb2;
+    uint nb3;
+} pcs;
+
+void main() {
+    const uint i03 = gl_WorkGroupID.z;
+    const uint i02 = gl_WorkGroupID.y;
+    const uint i01 = gl_WorkGroupID.x;
+
+    const int n = int(i03)*pcs.ne02*pcs.ne01*pcs.ne00 + int(i02)*pcs.ne01*pcs.ne00 + int(i01)*pcs.ne00;
+
+    const int i3 = n / (pcs.ne2*pcs.ne1*pcs.ne0);
+    const int i2 = (n - i3*pcs.ne2*pcs.ne1*pcs.ne0) / (pcs.ne1*pcs.ne0);
+    const int i1 = (n - i3*pcs.ne2*pcs.ne1*pcs.ne0 - i2*pcs.ne1*pcs.ne0) / pcs.ne0;
+    const int i0 = (n - i3*pcs.ne2*pcs.ne1*pcs.ne0 - i2*pcs.ne1*pcs.ne0 - i1*pcs.ne0);
+
+    const uint dst_data = (i3*pcs.nb3 + i2*pcs.nb2 + i1*pcs.nb1 + i0*pcs.nb0) / OUT_TYPE_SIZE + pcs.outOff; // Based from out_
+
+    for (uint i00 = gl_LocalInvocationID.x; i00 < pcs.ne00; i00 += gl_WorkGroupSize.x) {
+        const uint src = uint((i03*pcs.nb03 + i02*pcs.nb02 + i01*pcs.nb01 + i00*pcs.nb00) / IN_TYPE_SIZE) + pcs.inOff; // Based from in_
+        out_[dst_data+i00] = OUT_TYPE(in_[src]);
+    }
+}
diff --git a/src/whisper_cpp/ggml/src/kompute-shaders/op_diagmask.comp b/src/whisper_cpp/ggml/src/kompute-shaders/op_diagmask.comp
new file mode 100644
index 00000000..291c3fc1
--- /dev/null
+++ b/src/whisper_cpp/ggml/src/kompute-shaders/op_diagmask.comp
@@ -0,0 +1,30 @@
+#version 450
+
+#include "common.comp"
+
+layout(local_size_x = 1) in;
+
+layout(binding = 0) buffer restrict readonly tensorIn { float in_[]; };
+layout(binding = 1) buffer restrict writeonly tensorOut { float out_[]; };
+
+layout(push_constant) uniform PushConstants {
+    uint inOff;
+    uint outOff;
+    uint n_past;
+    int ne00;
+    int ne01;
+} pcs;
+
+void main() {
+    const uint i02 = gl_WorkGroupID.z;
+    const uint i01 = gl_WorkGroupID.y;
+    const uint i00 = gl_WorkGroupID.x;
+
+    const uint index = i02*pcs.ne01*pcs.ne00 + i01*pcs.ne00 + i00;
+
+    if (i00 > pcs.n_past + i01) {
+        out_[index + pcs.outOff] = uintBitsToFloat(0xFF800000);
+    } else {
+        out_[index + pcs.outOff] = in_[index + pcs.inOff];
+    }
+}
diff --git a/src/whisper_cpp/ggml/src/kompute-shaders/op_gelu.comp b/src/whisper_cpp/ggml/src/kompute-shaders/op_gelu.comp
new file mode 100644
index 00000000..9d8c5371
--- /dev/null
+++ b/src/whisper_cpp/ggml/src/kompute-shaders/op_gelu.comp
@@ -0,0 +1,22 @@
+#version 450
+
+#include "common.comp"
+
+layout(local_size_x = 1) in;
+
+layout(binding = 0) buffer restrict readonly tensorIn { float in_[]; };
+layout(binding = 1) buffer restrict writeonly tensorOut { float out_[]; };
+layout(push_constant) uniform PushConstants {
+    uint inOff;
+    uint outOff;
+} pcs;
+
+void main() {
+    const uint baseIndex = gl_WorkGroupID.x * 8;
+
+    for (uint x = 0; x < 8; x++) {
+        const uint i = baseIndex + x;
+        const float y = in_[i + pcs.inOff];
+        out_[i + pcs.outOff] = 0.5*y*(1.0 + tanh(clamp(SQRT_2_OVER_PI*y*(1.0 + GELU_COEF_A*y*y), -15.0, 15.0)));
+    }
+}
diff --git a/src/whisper_cpp/ggml/src/kompute-shaders/op_getrows.comp b/src/whisper_cpp/ggml/src/kompute-shaders/op_getrows.comp
new file mode 100644
index 00000000..1a5581b2
--- /dev/null
+++ b/src/whisper_cpp/ggml/src/kompute-shaders/op_getrows.comp
@@ -0,0 +1,17 @@
+void main() {
+    const uint i = gl_WorkGroupID.x;
+    const int r = inB[i + pcs.inBOff];
+
+    int z = 0;
+    for (uint ind = gl_LocalInvocationID.x; ind < pcs.ne00/16; ind += gl_WorkGroupSize.x) {
+        const uint inIndex = (r * pcs.nb01 + pcs.inAOff) + ind/NL * SIZE_OF_BLOCK;
+        const mat4 result = dequantize_block(inIndex, ind%NL);
+        for (uint j = 0; j < 4; ++j) {
+            for (uint k = 0; k < 4; ++k) {
+                const uint outIndex = i * pcs.nb1/BYTES_FOR_TYPE + pcs.outOff + z;
+                out_[outIndex] = result[j][k];
+                ++z;
+            }
+        }
+    }
+}
diff --git a/src/whisper_cpp/ggml/src/kompute-shaders/op_getrows_f16.comp b/src/whisper_cpp/ggml/src/kompute-shaders/op_getrows_f16.comp
new file mode 100644
index 00000000..48c93610
--- /dev/null
+++ b/src/whisper_cpp/ggml/src/kompute-shaders/op_getrows_f16.comp
@@ -0,0 +1,31 @@
+#version 450
+
+#include "common.comp"
+
+layout(local_size_x = 1) in;
+
+layout (binding = 0) readonly buffer tensorInA { float16_t inA[]; };
+layout (binding = 1) readonly buffer tensorInB { int inB[]; };
+layout (binding = 2) writeonly buffer tensorOut { float out_[]; };
+
+layout (push_constant) uniform parameter {
+    uint inAOff;
+    uint inBOff;
+    uint outOff;
+    int ne00;
+    int nb01;
+    int nb1;
+} pcs;
+
+void dequantize_row_f16(uint x /*Based from inA unaligned*/, uint y /*Based from out_*/, int k) {
+    for (int j = 0; j < k; j++) {
+        out_[y + j] = inA[x + j];
+    }
+}
+
+void main() {
+    const uint i = gl_WorkGroupID.x;
+    const int r = inB[i + pcs.inBOff];
+
+    dequantize_row_f16(r*pcs.nb01/2/*bytes for float16*/ + pcs.inAOff, i*pcs.nb1/4 + pcs.outOff, pcs.ne00);
+}
diff --git a/src/whisper_cpp/ggml/src/kompute-shaders/op_getrows_f32.comp b/src/whisper_cpp/ggml/src/kompute-shaders/op_getrows_f32.comp
new file mode 100644
index 00000000..9d7acdaf
--- /dev/null
+++ b/src/whisper_cpp/ggml/src/kompute-shaders/op_getrows_f32.comp
@@ -0,0 +1,31 @@
+#version 450
+
+#include "common.comp"
+
+layout(local_size_x = 1) in;
+
+layout (binding = 0) readonly buffer tensorInA { float inA[]; };
+layout (binding = 1) readonly buffer tensorInB { int inB[]; };
+layout (binding = 2) writeonly buffer tensorOut { float out_[]; };
+
+layout (push_constant) uniform parameter {
+    uint inAOff;
+    uint inBOff;
+    uint outOff;
+    int ne00;
+    int nb01;
+    int nb1;
+} pcs;
+
+void dequantize_row_f32(uint x /*Based from inA unaligned*/, uint y /*Based from out_*/, int k) {
+    for (int j = 0; j < k; j++) {
+        out_[y + j] = inA[x + j];
+    }
+}
+
+void main() {
+    const uint i = gl_WorkGroupID.x;
+    const int r = inB[i + pcs.inBOff];
+
+    dequantize_row_f32(r*pcs.nb01/4 + pcs.inAOff, i*pcs.nb1/4 + pcs.outOff, pcs.ne00);
+}
diff --git a/src/whisper_cpp/ggml/src/kompute-shaders/op_getrows_q4_0.comp b/src/whisper_cpp/ggml/src/kompute-shaders/op_getrows_q4_0.comp
new file mode 100644
index 00000000..32b2e891
--- /dev/null
+++ b/src/whisper_cpp/ggml/src/kompute-shaders/op_getrows_q4_0.comp
@@ -0,0 +1,38 @@
+#version 450
+
+#include "common.comp"
+
+#define NL 2
+#define BYTES_FOR_TYPE 4 /*bytes for float*/
+#define SIZE_OF_BLOCK sizeof_block_q4_0
+
+layout(local_size_x = 1) in;
+
+layout (binding = 0) readonly buffer tensorInA { uint8_t inA[]; };
+layout (binding = 1) readonly buffer tensorInB { int inB[]; };
+layout (binding = 2) writeonly buffer tensorOut { float out_[]; };
+
+layout (push_constant) uniform parameter {
+    uint inAOff;
+    uint inBOff;
+    uint outOff;
+    int ne00;
+    int nb01;
+    int nb1;
+} pcs;
+
+block_q4_0 get_unaligned_block_q4_0(uint index) {
+    block_q4_0 fres;
+    fres.d = u8BufToFloat16(inA, index);
+    [[unroll]] for (uint it = 0; it != QK4_0 / 2; it++) {
+        fres.qs[it] = inA[index+2+it];
+    }
+    return fres;
+}
+
+mat4 dequantize_block(uint index, uint il) {
+    const block_q4_0 block = get_unaligned_block_q4_0(index);
+    return dequantize_q4_0(block, il);
+}
+
+#include "op_getrows.comp"
diff --git a/src/whisper_cpp/ggml/src/kompute-shaders/op_getrows_q4_1.comp b/src/whisper_cpp/ggml/src/kompute-shaders/op_getrows_q4_1.comp
new file mode 100644
index 00000000..87f2fbe1
--- /dev/null
+++ b/src/whisper_cpp/ggml/src/kompute-shaders/op_getrows_q4_1.comp
@@ -0,0 +1,39 @@
+#version 450
+
+#include "common.comp"
+
+#define NL 2
+#define BYTES_FOR_TYPE 4 /*bytes for float*/
+#define SIZE_OF_BLOCK sizeof_block_q4_1
+
+layout(local_size_x = 1) in;
+
+layout (binding = 0) readonly buffer tensorInA { uint8_t inA[]; };
+layout (binding = 1) readonly buffer tensorInB { int inB[]; };
+layout (binding = 2) writeonly buffer tensorOut { float out_[]; };
+
+layout (push_constant) uniform parameter {
+    uint inAOff;
+    uint inBOff;
+    uint outOff;
+    int ne00;
+    int nb01;
+    int nb1;
+} pcs;
+
+block_q4_1 get_unaligned_block_q4_1(uint index) {
+    block_q4_1 fres;
+    fres.d = u8BufToFloat16(inA, index);
+    fres.m = u8BufToFloat16(inA, index+2);
+    [[unroll]] for (uint it = 0; it != QK4_1 / 2; it++) {
+        fres.qs[it] = inA[index+4+it];
+    }
+    return fres;
+}
+
+mat4 dequantize_block(uint index, uint il) {
+    const block_q4_1 block = get_unaligned_block_q4_1(index);
+    return dequantize_q4_1(block, il);
+}
+
+#include "op_getrows.comp"
diff --git a/src/whisper_cpp/ggml/src/kompute-shaders/op_getrows_q6_k.comp b/src/whisper_cpp/ggml/src/kompute-shaders/op_getrows_q6_k.comp
new file mode 100644
index 00000000..9ce3545d
--- /dev/null
+++ b/src/whisper_cpp/ggml/src/kompute-shaders/op_getrows_q6_k.comp
@@ -0,0 +1,44 @@
+#version 450
+
+#include "common.comp"
+
+#define NL 16
+#define BYTES_FOR_TYPE 4 /*bytes for float*/
+#define SIZE_OF_BLOCK sizeof_block_q6_k
+
+layout(local_size_x = 1) in;
+
+layout (binding = 0) readonly buffer tensorInA { uint8_t inA[]; };
+layout (binding = 1) readonly buffer tensorInB { int inB[]; };
+layout (binding = 2) writeonly buffer tensorOut { float out_[]; };
+
+layout (push_constant) uniform parameter {
+    uint inAOff;
+    uint inBOff;
+    uint outOff;
+    int ne00;
+    int nb01;
+    int nb1;
+} pcs;
+
+block_q6_k get_unaligned_block_q6_k(uint index) {
+    block_q6_k fres;
+    [[unroll]] for (uint it = 0; it != QK_K / 2; it++) {
+        fres.ql[it] = inA[index + it];
+    }
+    [[unroll]] for (uint it = 0; it != QK_K / 4; it++) {
+        fres.qh[it] = inA[index + QK_K/2 + it];
+    }
+    [[unroll]] for (uint it = 0; it != QK_K / 16; it++) {
+        fres.scales[it] = int8_t(inA[index + QK_K/2 + QK_K/4 + it]);
+    }
+    fres.d = u8BufToFloat16(inA, index + QK_K/2 + QK_K/4 + QK_K/16);
+    return fres;
+}
+
+mat4 dequantize_block(uint index, uint il) {
+    const block_q6_k block = get_unaligned_block_q6_k(index);
+    return dequantize_q6_k(block, il);
+}
+
+#include "op_getrows.comp"
diff --git a/src/whisper_cpp/ggml/src/kompute-shaders/op_mul.comp b/src/whisper_cpp/ggml/src/kompute-shaders/op_mul.comp
new file mode 100644
index 00000000..c92647c4
--- /dev/null
+++ b/src/whisper_cpp/ggml/src/kompute-shaders/op_mul.comp
@@ -0,0 +1,52 @@
+#version 450
+
+#include "common.comp"
+
+layout(local_size_x = 1024) in;
+
+layout(binding = 0) buffer restrict readonly tensorInA { float inA[]; };
+layout(binding = 1) buffer restrict readonly tensorInB { float inB[]; };
+layout(binding = 2) buffer restrict writeonly tensorOut { float out_[]; };
+
+layout(push_constant) uniform PushConstants {
+    uint inAOff;
+    uint inBOff;
+    uint outOff;
+    int ne00;
+    int nb00;
+    int nb01;
+    int nb02;
+    int nb03;
+    int ne10;
+    int ne11;
+    int ne12;
+    int ne13;
+    int nb10;
+    int nb11;
+    int nb12;
+    int nb13;
+    int ne0;
+    int nb0;
+    int nb1;
+    int nb2;
+    int nb3;
+} pcs;
+
+void main() {
+    const uint i03 = gl_WorkGroupID.z;
+    const uint i02 = gl_WorkGroupID.y;
+    const uint i01 = gl_WorkGroupID.x;
+
+    const uint i13 = i03 % pcs.ne13;
+    const uint i12 = i02 % pcs.ne12;
+    const uint i11 = i01 % pcs.ne11;
+
+    uint src0_off = uint((i03*pcs.nb03 + i02*pcs.nb02 + i01*pcs.nb01) / 4);
+    uint src1_off = uint((i13*pcs.nb13 + i12*pcs.nb12 + i11*pcs.nb11) / 4);
+    uint dst_off  = uint((i03*pcs.nb3  + i02*pcs.nb2  + i01*pcs.nb1)  / 4);
+
+    for (uint i0 = gl_LocalInvocationID.x; i0 < pcs.ne0; i0 += gl_WorkGroupSize.x) {
+        const uint i10 = i0 % pcs.ne10;
+        out_[pcs.outOff + dst_off + i0] = inA[pcs.inAOff + src0_off + i0] * inB[pcs.inBOff + src1_off + i10];
+    }
+}
diff --git a/src/whisper_cpp/ggml/src/kompute-shaders/op_mul_mat_f16.comp b/src/whisper_cpp/ggml/src/kompute-shaders/op_mul_mat_f16.comp
new file mode 100644
index 00000000..8f0a9031
--- /dev/null
+++ b/src/whisper_cpp/ggml/src/kompute-shaders/op_mul_mat_f16.comp
@@ -0,0 +1,67 @@
+#version 450
+
+#include "common.comp"
+
+#extension GL_KHR_shader_subgroup_arithmetic : require
+
+layout(local_size_x_id = 0) in;
+
+layout (binding = 0) readonly buffer tensorInA { float16_t inA[]; };
+layout (binding = 1) readonly buffer tensorInB { float inB[]; };
+layout (binding = 2) writeonly buffer tensorOut { float out_[]; };
+
+layout (push_constant) uniform parameter {
+    uint inAOff;
+    uint inBOff;
+    uint outOff;
+    int ne00;
+    int ne01;
+    int ne02;
+    uint nb00;
+    uint nb01;
+    uint nb02;
+    int ne10;
+    int ne11;
+    int ne12;
+    uint nb10;
+    uint nb11;
+    uint nb12;
+    int ne0;
+    int ne1;
+    uint r2;
+    uint r3;
+} pcs;
+
+#define N_F16_F32 4
+
+void main() {
+    const uint r0 = gl_WorkGroupID.x;
+    const uint rb = gl_WorkGroupID.y*N_F16_F32;
+    const uint im = gl_WorkGroupID.z;
+
+    const uint i12 = im%pcs.ne12;
+    const uint i13 = im/pcs.ne12;
+
+    const uint offset0 = r0*pcs.nb01 + (i12/pcs.r2)*pcs.nb02 + (i13/pcs.r3)*pcs.nb02*pcs.ne02;
+
+    const uint x = offset0 / 2 + pcs.inAOff; // Based from inA
+
+    for (uint row = 0; row < N_F16_F32; ++row) {
+        uint r1 = rb + row;
+        if (r1 >= pcs.ne11) {
+            break;
+        }
+
+        const uint y = (r1*pcs.nb11 + im*pcs.nb12) / 4 + pcs.inBOff; // Based from inB
+
+        float sumf = 0;
+        for (uint i = gl_SubgroupInvocationID.x; i < pcs.ne00; i += gl_SubgroupSize) {
+            sumf += float(inA[x+i]) * float(inB[y+i]);
+        }
+
+        const float all_sum = subgroupAdd(sumf);
+        if (subgroupElect()) {
+            out_[im*pcs.ne1*pcs.ne0 + r1*pcs.ne0 + r0 + pcs.outOff] = all_sum;
+        }
+    }
+}
diff --git a/src/whisper_cpp/ggml/src/kompute-shaders/op_mul_mat_mat_f32.comp b/src/whisper_cpp/ggml/src/kompute-shaders/op_mul_mat_mat_f32.comp
new file mode 100644
index 00000000..d1ca4ad6
--- /dev/null
+++ b/src/whisper_cpp/ggml/src/kompute-shaders/op_mul_mat_mat_f32.comp
@@ -0,0 +1,51 @@
+#version 450
+
+#include "common.comp"
+
+#extension GL_KHR_shader_subgroup_arithmetic : require
+#extension GL_EXT_debug_printf : enable
+
+// device subgroup size
+layout (local_size_x_id = 0) in;
+
+layout(binding = 0) readonly buffer tensorInA { float inA[]; };
+layout(binding = 1) readonly buffer tensorInB { float inB[]; };
+layout(binding = 2) writeonly buffer tensorOut { float out_[]; };
+
+layout(push_constant) uniform parameter {
+  uint inAOff;
+  uint inBOff;
+  uint outOff;
+  int ne00;
+  int ne01;
+  int ne02;
+  int ne11;
+  int ne12;
+  uint nb01;
+  uint nb02;
+  uint nb11;
+  uint nb12;
+  uint nb1;
+  uint nb2;
+}
+pcs;
+
+
+void main() {
+  uvec3 gid = gl_WorkGroupID;
+
+  uint bc_ab = pcs.ne12 > pcs.ne02 ? gid.z / (pcs.ne12 / pcs.ne02) : gid.z;
+  uint bc_ba = pcs.ne02 > pcs.ne12 ? gid.z / (pcs.ne02 / pcs.ne12) : gid.z;
+
+  const uint x = (gid.x*pcs.nb01 + bc_ab*pcs.nb02) / 4 + pcs.inAOff; // Based from inA
+  const uint y = (gid.y*pcs.nb11 + bc_ba*pcs.nb12) / 4 + pcs.inBOff; // based from inB
+  float sum = 0.0f;
+  for (uint i = gl_SubgroupInvocationID.x; i < pcs.ne00; i += gl_SubgroupSize) {
+      sum += float(inA[x+i]) * float(inB[y+i]);
+  }
+
+  const float all_sum = subgroupAdd(sum);
+  if (subgroupElect()) {
+    out_[gid.z*(pcs.nb2/4) + gid.y*(pcs.nb1/4) + gid.x + pcs.outOff] = all_sum;
+  }
+}
diff --git a/src/whisper_cpp/ggml/src/kompute-shaders/op_mul_mat_q4_0.comp b/src/whisper_cpp/ggml/src/kompute-shaders/op_mul_mat_q4_0.comp
new file mode 100644
index 00000000..b0cea8bb
--- /dev/null
+++ b/src/whisper_cpp/ggml/src/kompute-shaders/op_mul_mat_q4_0.comp
@@ -0,0 +1,33 @@
+#version 450
+
+#include "common.comp"
+
+#define BLOCKS_IN_QUANT QK4_0
+#define SIZE_OF_BLOCK sizeof_block_q4_0
+#define N_ROWS 4
+
+#include "op_mul_mv_q_n_pre.comp"
+
+// The q4_0 version of this function
+float block_q_n_dot_y(uint block_index, uint yb, uint il) {
+    vec2 acc = vec2(0.0, 0.0);
+    const uint index = (block_index) * SIZE_OF_BLOCK + pcs.inAOff;
+    float d = float(u8BufToFloat16(inA, index));
+    float sumy = 0.0f;
+    for (int i = 0; i < BLOCKS_IN_QUANT/4; i+=2) {
+        const uint16_t b = u8BufToU16(inA, index + 2 + il + i);
+
+        const float yl0 = inB[yb + i];
+        const float yl1 = inB[yb + i + 1];
+        const float yl8 = inB[yb + i + BLOCKS_IN_QUANT/2];
+        const float yl9 = inB[yb + i + BLOCKS_IN_QUANT/2 + 1];
+
+        sumy += yl0 + yl1 + yl8 + yl9;
+
+        acc[0] += yl0 * (b & 0x000F) + yl1 / 256.f * (b & 0x0F00);
+        acc[1] += yl8 / 16.f * (b & 0x00F0) + yl9 / 4096.f * (b & 0xF000);
+    }
+    return d * (sumy * -8.f + acc[0] + acc[1]);
+}
+
+#include "op_mul_mv_q_n.comp"
diff --git a/src/whisper_cpp/ggml/src/kompute-shaders/op_mul_mat_q4_1.comp b/src/whisper_cpp/ggml/src/kompute-shaders/op_mul_mat_q4_1.comp
new file mode 100644
index 00000000..8582c61a
--- /dev/null
+++ b/src/whisper_cpp/ggml/src/kompute-shaders/op_mul_mat_q4_1.comp
@@ -0,0 +1,35 @@
+#version 450
+
+#include "common.comp"
+
+#define BLOCKS_IN_QUANT QK4_1
+#define SIZE_OF_BLOCK sizeof_block_q4_1
+#define N_ROWS 4
+
+#include "op_mul_mv_q_n_pre.comp"
+
+// The q4_1 version of this function
+float block_q_n_dot_y(uint block_index, uint yb, uint il) {
+    vec2 acc = vec2(0.0, 0.0);
+    const uint index = (block_index) * SIZE_OF_BLOCK + pcs.inAOff;
+    float d = float(u8BufToFloat16(inA, index));
+    float m = float(u8BufToFloat16(inA, index+2));
+
+    float sumy = 0.0f;
+    for (int i = 0; i < BLOCKS_IN_QUANT/4; i+=2) {
+        const uint16_t b = u8BufToU16(inA, index + 4 + il + i);
+
+        const float yl0 = inB[yb + i];
+        const float yl1 = inB[yb + i + 1];
+        const float yl8 = inB[yb + i + BLOCKS_IN_QUANT/2];
+        const float yl9 = inB[yb + i + BLOCKS_IN_QUANT/2 + 1];
+
+        sumy += yl0 + yl1 + yl8 + yl9;
+
+        acc[0] += yl0 * (b & 0x000F) + yl1 / 256.f * (b & 0x0F00);
+        acc[1] += yl8 / 16.f * (b & 0x00F0) + yl9 / 4096.f * (b & 0xF000);
+    }
+    return d * (acc[0] + acc[1]) + sumy * m;
+}
+
+#include "op_mul_mv_q_n.comp"
diff --git a/src/whisper_cpp/ggml/src/kompute-shaders/op_mul_mat_q6_k.comp b/src/whisper_cpp/ggml/src/kompute-shaders/op_mul_mat_q6_k.comp
new file mode 100644
index 00000000..c9baebdf
--- /dev/null
+++ b/src/whisper_cpp/ggml/src/kompute-shaders/op_mul_mat_q6_k.comp
@@ -0,0 +1,94 @@
+#version 450
+
+#include "common.comp"
+
+#define SIZE_OF_BLOCK sizeof_block_q6_k
+
+layout(local_size_x_id = 0) in;
+layout(local_size_y_id = 1) in;
+layout(local_size_z = 1) in;
+
+layout (binding = 0) readonly buffer tensorInA { uint8_t inA[]; };
+layout (binding = 1) readonly buffer tensorInB { float inB[]; };
+layout (binding = 2) writeonly buffer tensorOut { float out_[]; };
+
+layout (push_constant) uniform parameter {
+    uint inAOff;
+    uint inBOff;
+    uint outOff;
+    int ne00;
+    int ne10;
+    int ne0;
+    int ne1;
+    int ne01;
+    int gqa;
+} pcs;
+
+void main() {
+    const uint8_t kmask1 = uint8_t(0x03);
+    const uint8_t kmask2 = uint8_t(0x0C);
+    const uint8_t kmask3 = uint8_t(0x30);
+    const uint8_t kmask4 = uint8_t(0xC0);
+
+    const uint nb = pcs.ne00/QK_K;
+
+    const uint r0 = gl_WorkGroupID.x;
+    const uint r1 = gl_WorkGroupID.y;
+    const uint r2 = gl_WorkGroupID.z;
+
+    const uint row = (r0 * gl_NumSubgroups + gl_SubgroupID);
+    const uint offset0 = r2/pcs.gqa*(nb*pcs.ne0);
+    const uint x = row * nb + offset0; // Based from inA without base offset
+    const uint yy = r1*pcs.ne10 + r2*pcs.ne00*pcs.ne1+pcs.inBOff; // Based from inB
+
+    float sumf = 0;
+
+    // bits of invocation ID for gl_SubgroupSize=32:
+    //  x   x   x   x   x
+    //  4   3   2   1   0
+    // (     tid     ) ix
+    //  ip (   il    )
+
+    const uint block_stride = gl_SubgroupSize / 16;         // number of blocks each subgroup processes
+    const uint tid  = gl_SubgroupInvocationID/block_stride; // first block_stride groups have tid=0
+    const uint ix   = gl_SubgroupInvocationID%block_stride; // first block is 0..block_stride-1
+    const uint ip   = tid/8;        // first or second half of block (0 or 1)
+    const uint il   = tid%8;        // each half has 8 parts, one per scale
+    const uint n    = 4;            // 4 scales at a time (and 4 sums)
+    const uint l0   = n*il;         // offset into half-block, 0..28
+    const uint is   = 8*ip + l0/16; // 0, 1, 8, 9
+
+    const uint y_offset = 128*ip + l0;
+    const uint q_offset_l = 64*ip + l0;
+    const uint q_offset_h = 32*ip + l0;
+
+    for (uint i = ix; i < nb; i += block_stride) {
+
+        const uint baseIndex = (x + i) * SIZE_OF_BLOCK + pcs.inAOff;
+
+        const uint qlIndex = q_offset_l;
+        const uint q2Index = qlIndex + QK_K/8;
+        const uint qhIndex = q_offset_h;
+        const uint y = yy + i * QK_K + y_offset;
+
+        float sums[4] = {0.0f, 0.0f, 0.0f, 0.0f};
+        for (uint l = 0; l < n; ++l) {
+            const uint8_t currentQ1 = inA[baseIndex + qlIndex + l];
+            const uint8_t currentQ2 = inA[baseIndex + q2Index + l];
+            const uint8_t currentQh = inA[baseIndex + QK_K/2 + qhIndex + l];
+
+            sums[0] += inB[y+l+ 0] * (int8_t((currentQ1 & 0xF) | ((currentQh & kmask1) << 4)) - 32);
+            sums[1] += inB[y+l+32] * (int8_t((currentQ2 & 0xF) | ((currentQh & kmask2) << 2)) - 32);
+            sums[2] += inB[y+l+64] * (int8_t((currentQ1  >> 4) | ((currentQh & kmask3) << 0)) - 32);
+            sums[3] += inB[y+l+96] * (int8_t((currentQ2  >> 4) | ((currentQh & kmask4) >> 2)) - 32);
+        }
+
+        float d = u8BufToFloat16(inA, baseIndex + QK_K/2 + QK_K/4 + QK_K/16);
+        sumf += d * (sums[0] * int8_t(inA[baseIndex + QK_K/2 + QK_K/4 + is]) + sums[1] * int8_t(inA[baseIndex + QK_K/2 + QK_K/4 + 2 + is]) + sums[2] * int8_t(inA[baseIndex + QK_K/2 + QK_K/4 + 4 + is]) + sums[3] * int8_t(inA[baseIndex + QK_K/2 + QK_K/4 + 6 + is]));
+    }
+
+    const float tot = subgroupAdd(sumf);
+    if (subgroupElect()) {
+        out_[r1*pcs.ne0 + r2*pcs.ne0*pcs.ne1 + row + pcs.outOff] = tot;
+    }
+}
diff --git a/src/whisper_cpp/ggml/src/kompute-shaders/op_mul_mat_q8_0.comp b/src/whisper_cpp/ggml/src/kompute-shaders/op_mul_mat_q8_0.comp
new file mode 100644
index 00000000..34d015e9
--- /dev/null
+++ b/src/whisper_cpp/ggml/src/kompute-shaders/op_mul_mat_q8_0.comp
@@ -0,0 +1,73 @@
+#version 450
+
+#include "common.comp"
+
+#include "op_mul_mv_q_n_pre.comp"
+
+#define SIZE_OF_D 2
+
+#define N_DST 4 // each SIMD group works on 4 rows
+#define N_SIMDGROUP 2 // number of SIMD groups in a thread group
+#define N_SIMDWIDTH 32 // assuming SIMD group size is 32
+
+#define NB_Q8_0 8
+
+void main() {
+    // NB: hack to make compatible with AMD GPUs that have a subgroup size of 64
+    if (gl_SubgroupInvocationID > 31)
+        return;
+
+    const int nr  = N_DST;
+    const int nsg = N_SIMDGROUP;
+    const int nw  = N_SIMDWIDTH;
+
+    const int nb = pcs.ne00/QK8_0;
+    const uint r0 = gl_WorkGroupID.x;
+    const uint r1 = gl_WorkGroupID.y;
+    const uint im = gl_WorkGroupID.z;
+
+    const uint first_row = (r0 * nsg + gl_SubgroupID) * nr;
+
+    const uint i12 = im%pcs.ne12;
+    const uint i13 = im/pcs.ne12;
+
+    const uint offset0 = first_row * nb + (i12/pcs.r2)*(nb*pcs.ne01) + (i13/pcs.r3)*(nb*pcs.ne01*pcs.ne02);
+
+    const uint x = offset0*sizeof_block_q8_0 + pcs.inAOff; // Based from inA
+    const uint y = r1*pcs.ne10 + im*pcs.ne00*pcs.ne1 + pcs.inBOff; // based from inB
+
+    float yl[NB_Q8_0];
+    float sumf[N_DST]={0.f, 0.f, 0.f, 0.f};
+
+    const uint ix = gl_SubgroupInvocationID.x/4;
+    const uint il = gl_SubgroupInvocationID.x%4;
+
+    uint yb = y + ix * QK8_0 + NB_Q8_0*il;
+
+    // each thread in a SIMD group deals with NB_Q8_0 quants at a time
+    for (uint ib = ix; ib < nb; ib += nw/4) {
+        for (int i = 0; i < NB_Q8_0; ++i) {
+            yl[i] = inB[yb + i];
+        }
+
+        for (int row = 0; row < nr; row++) {
+            const uint block_offset = (ib+row*nb) * sizeof_block_q8_0;
+            float sumq = 0.f;
+            for (int iq = 0; iq < NB_Q8_0; ++iq) {
+                const int8_t qs_iq = int8_t(inA[x + block_offset + SIZE_OF_D + NB_Q8_0*il + iq]);
+                sumq += qs_iq * yl[iq];
+            }
+            const float16_t d = u8BufToFloat16(inA, x + block_offset);
+            sumf[row] += sumq*d;
+        }
+
+        yb += NB_Q8_0 * nw;
+    }
+
+    for (int row = 0; row < nr; ++row) {
+        const float tot = subgroupAdd(sumf[row]);
+        if (subgroupElect() && first_row + row < pcs.ne01) {
+            out_[r1*pcs.ne0 + im*pcs.ne0*pcs.ne1 + first_row + row] = tot;
+        }
+    }
+}
diff --git a/src/whisper_cpp/ggml/src/kompute-shaders/op_mul_mv_q_n.comp b/src/whisper_cpp/ggml/src/kompute-shaders/op_mul_mv_q_n.comp
new file mode 100644
index 00000000..440b5ab2
--- /dev/null
+++ b/src/whisper_cpp/ggml/src/kompute-shaders/op_mul_mv_q_n.comp
@@ -0,0 +1,48 @@
+void main() {
+    // NB: hack to make compatible with AMD GPUs that have a subgroup size of 64
+    if (gl_SubgroupInvocationID > 31)
+        return;
+
+    const uint nb = uint(pcs.ne00/BLOCKS_IN_QUANT);
+
+    const uint r0 = gl_WorkGroupID.x;
+    const uint r1 = gl_WorkGroupID.y;
+    const uint im = gl_WorkGroupID.z;
+
+    const uint first_row = (r0 * gl_NumSubgroups + gl_SubgroupID) * N_ROWS;
+
+    const uint i12 = im%pcs.ne12;
+    const uint i13 = im/pcs.ne12;
+
+    const uint offset0 = first_row * nb + (i12/pcs.r2)*(nb*pcs.ne01) + (i13/pcs.r3)*(nb*pcs.ne01*pcs.ne02);
+
+    const uint x = offset0; // Based from inA without base offset
+    const uint y = r1*uint(pcs.ne10)+im*pcs.ne00*pcs.ne1+pcs.inBOff; // Based from inB
+
+    float sumf[N_ROWS] = {0.0f, 0.0f, 0.0f, 0.0f};
+
+    const uint ix = gl_SubgroupInvocationID/2;
+    const uint il = (BLOCKS_IN_QUANT/4)*(gl_SubgroupInvocationID%2);
+
+    uint yb = y + ix * BLOCKS_IN_QUANT + il;
+
+    //debugPrintfEXT("gl_NumSubgroups=%d, gl_SubgroupID=%d, gl_SubgroupInvocationID=%d, glSubgroupSize=%d, gl_WorkGroupSize.x=%d, gl_WorkGroupSize.y=%d, gl_WorkGroupSize.z=%d\n",
+    //    gl_NumSubgroups, gl_SubgroupID, gl_SubgroupInvocationID, gl_SubgroupSize,
+    //    gl_WorkGroupSize.x, gl_WorkGroupSize.y, gl_WorkGroupSize.z);
+
+    for (uint ib = ix; ib < nb; ib += 16) {
+        for (int row = 0; row < N_ROWS; row++) {
+            const uint block_index = x + ib + row * nb;
+            sumf[row] += block_q_n_dot_y(block_index, yb, il);
+        }
+
+        yb += BLOCKS_IN_QUANT * 16;
+    }
+
+    for (int row = 0; row < N_ROWS; ++row) {
+        const float tot = subgroupAdd(sumf[row]);
+        if (first_row + row < pcs.ne01 && subgroupElect()) {
+            out_[r1*pcs.ne0 + im*pcs.ne0*pcs.ne1 + first_row + row + pcs.outOff] = tot;
+        }
+    }
+}
diff --git a/src/whisper_cpp/ggml/src/kompute-shaders/op_mul_mv_q_n_pre.comp b/src/whisper_cpp/ggml/src/kompute-shaders/op_mul_mv_q_n_pre.comp
new file mode 100644
index 00000000..7912b09a
--- /dev/null
+++ b/src/whisper_cpp/ggml/src/kompute-shaders/op_mul_mv_q_n_pre.comp
@@ -0,0 +1,22 @@
+layout(local_size_x_id = 0) in;
+layout(local_size_y = 1) in;
+layout(local_size_z = 1) in;
+
+layout (binding = 0) readonly buffer tensorInA { uint8_t inA[]; };
+layout (binding = 1) readonly buffer tensorInB { float inB[]; };
+layout (binding = 2) writeonly buffer tensorOut { float out_[]; };
+
+layout (push_constant) uniform parameter {
+    uint inAOff;
+    uint inBOff;
+    uint outOff;
+    int  ne00;
+    int  ne01;
+    int  ne02;
+    int  ne10;
+    int  ne12;
+    int  ne0;
+    int  ne1;
+    uint r2;
+    uint r3;
+} pcs;
diff --git a/src/whisper_cpp/ggml/src/kompute-shaders/op_norm.comp b/src/whisper_cpp/ggml/src/kompute-shaders/op_norm.comp
new file mode 100644
index 00000000..ad0c3c01
--- /dev/null
+++ b/src/whisper_cpp/ggml/src/kompute-shaders/op_norm.comp
@@ -0,0 +1,84 @@
+#version 450
+
+#include "common.comp"
+
+layout(local_size_x = 256) in;
+
+layout(binding = 0) buffer restrict readonly tensorIn { float in_[]; };
+layout(binding = 1) buffer restrict tensorOut { float out_[]; };
+
+layout(push_constant) uniform PushConstants {
+    uint inOff;
+    uint outOff;
+    uint ne00;
+    uint nb01;
+    float eps;
+} pcs;
+
+shared float sum[gl_WorkGroupSize.x];
+
+void main() {
+    const uint x = (gl_WorkGroupID.x*pcs.nb01/4) + pcs.inOff; // Based from in_
+    // MEAN
+    // parallel sum
+    sum[gl_LocalInvocationID.x] = 0.0;
+    for (uint i00 = gl_LocalInvocationID.x; i00 < pcs.ne00; i00 += gl_WorkGroupSize.x) {
+        sum[gl_LocalInvocationID.x] += in_[x+i00];
+    }
+
+    // reduce
+    barrier();
+    memoryBarrierShared();
+    [[unroll]] for (uint i = gl_WorkGroupSize.x/2; i > 0; i /= 2) {
+        if (gl_LocalInvocationID.x < i) {
+            sum[gl_LocalInvocationID.x] += sum[gl_LocalInvocationID.x + i];
+        }
+        barrier();
+        memoryBarrierShared();
+    }
+
+    // broadcast
+    if (gl_LocalInvocationID.x == 0) {
+        sum[0] /= float(pcs.ne00);
+    }
+    barrier();
+    memoryBarrierShared();
+    const float mean = sum[0];
+
+    // recenter
+    const uint y = (gl_WorkGroupID.x*pcs.ne00) + pcs.outOff; // Based from out_
+    for (uint i00 = gl_LocalInvocationID.x; i00 < pcs.ne00; i00 += gl_WorkGroupSize.x) {
+        out_[y+i00] = in_[x+i00] - mean;
+    }
+
+    // VARIANCE
+    // parallel sum
+    sum[gl_LocalInvocationID.x] = 0.0;
+    for (uint i00 = gl_LocalInvocationID.x; i00 < pcs.ne00; i00 += gl_WorkGroupSize.x) {
+        sum[gl_LocalInvocationID.x] += out_[y+i00] * out_[y+i00];
+    }
+
+    // reduce
+    barrier();
+    memoryBarrierShared();
+    [[unroll]] for (uint i = gl_WorkGroupSize.x/2; i > 0; i /= 2) {
+        if (gl_LocalInvocationID.x < i) {
+            sum[gl_LocalInvocationID.x] += sum[gl_LocalInvocationID.x + i];
+        }
+        barrier();
+        memoryBarrierShared();
+    }
+
+    // broadcast
+    if (gl_LocalInvocationID.x == 0) {
+        sum[0] /= float(pcs.ne00);
+    }
+    barrier();
+    memoryBarrierShared();
+    const float variance = sum[0];
+
+    const float scale = 1.0f/sqrt(variance + pcs.eps);
+    for (uint i00 = gl_LocalInvocationID.x; i00 < pcs.ne00; i00 += gl_WorkGroupSize.x) {
+        out_[y+i00] *= scale;
+    }
+}
diff --git a/src/whisper_cpp/ggml/src/kompute-shaders/op_relu.comp b/src/whisper_cpp/ggml/src/kompute-shaders/op_relu.comp
new file mode 100644
index 00000000..52a601fe
--- /dev/null
+++ b/src/whisper_cpp/ggml/src/kompute-shaders/op_relu.comp
@@ -0,0 +1,21 @@
+#version 450
+
+#include "common.comp"
+
+layout(local_size_x = 1) in;
+
+layout(binding = 0) buffer restrict readonly tensorIn { float in_[]; };
+layout(binding = 1) buffer restrict writeonly tensorOut { float out_[]; };
+layout(push_constant) uniform PushConstants {
+    uint inOff;
+    uint outOff;
+} pcs;
+
+void main() {
+    const uint baseIndex = gl_WorkGroupID.x * 4;
+
+    for (uint x = 0; x < 4; x++) {
+        const uint i = baseIndex + x;
+        out_[i + pcs.outOff] = max(0.0, in_[i + pcs.inOff]);
+    }
+}
diff --git a/src/whisper_cpp/ggml/src/kompute-shaders/op_rmsnorm.comp b/src/whisper_cpp/ggml/src/kompute-shaders/op_rmsnorm.comp
new file mode 100644
index 00000000..da658c16
--- /dev/null
+++ b/src/whisper_cpp/ggml/src/kompute-shaders/op_rmsnorm.comp
@@ -0,0 +1,53 @@
+#version 450
+
+#include "common.comp"
+
+layout(local_size_x = 512) in;
+
+layout(binding = 0) buffer restrict readonly tensorIn { float in_[]; };
+layout(binding = 1) buffer restrict tensorOut { float out_[]; };
+
+layout(push_constant) uniform PushConstants {
+    uint inOff;
+    uint outOff;
+    uint ne00;
+    uint nb01;
+    float eps;
+} pcs;
+
+shared float sum[gl_WorkGroupSize.x];
+
+void main() {
+    const uint x = (gl_WorkGroupID.x*pcs.nb01/4) + pcs.inOff; // Based from in_
+
+    // parallel sum
+    sum[gl_LocalInvocationID.x] = 0.0;
+    for (uint i00 = gl_LocalInvocationID.x; i00 < pcs.ne00; i00 += gl_WorkGroupSize.x) {
+        sum[gl_LocalInvocationID.x] += in_[x+i00] * in_[x+i00];
+    }
+
+    // reduce
+    barrier();
+    memoryBarrierShared();
+    [[unroll]] for (uint i = gl_WorkGroupSize.x/2; i > 0; i /= 2) {
+        if (gl_LocalInvocationID.x < i) {
+            sum[gl_LocalInvocationID.x] += sum[gl_LocalInvocationID.x + i];
+        }
+        barrier();
+        memoryBarrierShared();
+    }
+
+    // broadcast
+    if (gl_LocalInvocationID.x == 0) {
+        sum[0] /= float(pcs.ne00);
+    }
+    barrier();
+    memoryBarrierShared();
+
+    const float scale = 1.0f/sqrt(sum[0] + pcs.eps);
+
+    const uint y = (gl_WorkGroupID.x*pcs.ne00) + pcs.outOff; // Based from out_
+    for (uint i00 = gl_LocalInvocationID.x; i00 < pcs.ne00; i00 += gl_WorkGroupSize.x) {
+        out_[y+i00] = in_[x+i00] * scale;
+    }
+}
diff --git a/src/whisper_cpp/ggml/src/kompute-shaders/op_rope_f16.comp b/src/whisper_cpp/ggml/src/kompute-shaders/op_rope_f16.comp
new file mode 100644
index 00000000..1a4058b3
--- /dev/null
+++ b/src/whisper_cpp/ggml/src/kompute-shaders/op_rope_f16.comp
@@ -0,0 +1,73 @@
+#version 450
+
+#include "rope_common.comp"
+
+layout(binding = 0) buffer restrict readonly  tensorInA { float16_t inA[]; };
+layout(binding = 1) buffer restrict readonly  tensorInB { int       inB[]; };
+layout(binding = 2) buffer restrict writeonly tensorOut { float16_t out_[]; };
+
+void main() {
+    const uint i3 = gl_WorkGroupID.z;
+    const uint i2 = gl_WorkGroupID.y;
+    const uint i1 = gl_WorkGroupID.x;
+
+    const bool is_neox = (pcs.mode & 2) != 0;
+
+    float corr_dims[2];
+    rope_yarn_corr_dims(pcs.n_dims, pcs.n_ctx_orig, pcs.freq_base, pcs.beta_fast, pcs.beta_slow, corr_dims);
+
+    const float theta_scale = pow(pcs.freq_base, -2.0/pcs.n_dims);
+
+    const int p = inB[pcs.inBOff + i2];
+
+    float theta = float(p);
+
+    if (!is_neox) {
+        for (uint i0 = 0; i0 < pcs.ne0; i0 += 2) {
+            float cos_theta, sin_theta;
+            rope_yarn(theta, pcs.freq_scale, corr_dims, i0, pcs.ext_factor, pcs.attn_factor, cos_theta, sin_theta);
+
+            theta *= theta_scale;
+
+            const uint src      = uint((i3*pcs.nb03 + i2*pcs.nb02 + i1*pcs.nb01 + i0*pcs.nb00) / 2) + pcs.inAOff; // Based from in
+            const uint dst_data = uint((i3*pcs.nb3  + i2*pcs.nb2  + i1*pcs.nb1  + i0*pcs.nb0)  / 2) + pcs.outOff; // Based from out_
+
+            const float x0 = float(inA[src]);
+            const float x1 = float(inA[src+1]);
+
+            out_[dst_data]   = float16_t(x0*cos_theta - x1*sin_theta);
+            out_[dst_data+1] = float16_t(x0*sin_theta + x1*cos_theta);
+        }
+    } else {
+        const float inv_ndims = -1.f/pcs.n_dims;
+        for (uint ic = 0; ic < pcs.n_dims; ic += 2) {
+            const uint cur_rot = ic;
+
+            float cos_theta, sin_theta;
+            rope_yarn(theta, pcs.freq_scale, corr_dims, cur_rot, pcs.ext_factor, pcs.attn_factor, cos_theta, sin_theta);
+
+            theta *= theta_scale;
+
+            const uint i0 = ic/2;
+
+            const uint src      = uint((i3*pcs.nb03 + i2*pcs.nb02 + i1*pcs.nb01 + i0*pcs.nb00) / 2) + pcs.inAOff; // Based from in
+            const uint dst_data = uint((i3*pcs.nb3  + i2*pcs.nb2  + i1*pcs.nb1  + i0*pcs.nb0)  / 2) + pcs.outOff; // Based from out_
+
+            const float x0 = float(inA[src]);
+            const float x1 = float(inA[src+pcs.n_dims/2]);
+
+            out_[dst_data]              = float16_t(x0*cos_theta - x1*sin_theta);
+            out_[dst_data+pcs.n_dims/2] = float16_t(x0*sin_theta + x1*cos_theta);
+        }
+
+        for (uint ic = pcs.n_dims; ic < pcs.ne0; ic += 2) {
+            const uint i0 = ic;
+
+            const uint src      = uint((i3*pcs.nb03 + i2*pcs.nb02 + i1*pcs.nb01 + i0*pcs.nb00) / 2) + pcs.inAOff; // Based from in
+            const uint dst_data = uint((i3*pcs.nb3  + i2*pcs.nb2  + i1*pcs.nb1  + i0*pcs.nb0)  / 2) + pcs.outOff; // Based from out_
+
+            out_[dst_data + 0] = inA[src + 0];
+            out_[dst_data + 1] = inA[src + 1];
+        }
+    }
+}
diff --git a/src/whisper_cpp/ggml/src/kompute-shaders/op_rope_f32.comp b/src/whisper_cpp/ggml/src/kompute-shaders/op_rope_f32.comp
new file mode 100644
index 00000000..65e03827
--- /dev/null
+++ b/src/whisper_cpp/ggml/src/kompute-shaders/op_rope_f32.comp
@@ -0,0 +1,73 @@
+#version 450
+
+#include "rope_common.comp"
+
+layout(binding = 0) buffer restrict readonly  tensorInA { float inA[]; };
+layout(binding = 1) buffer restrict readonly  tensorInB { int   inB[]; };
+layout(binding = 2) buffer restrict writeonly tensorOut { float out_[]; };
+
+void main() {
+    const uint i3 = gl_WorkGroupID.z;
+    const uint i2 = gl_WorkGroupID.y;
+    const uint i1 = gl_WorkGroupID.x;
+
+    const bool is_neox = (pcs.mode & 2) != 0;
+
+    float corr_dims[2];
+    rope_yarn_corr_dims(pcs.n_dims, pcs.n_ctx_orig, pcs.freq_base, pcs.beta_fast, pcs.beta_slow, corr_dims);
+
+    const float theta_scale = pow(pcs.freq_base, -2.0/pcs.n_dims);
+
+    const int p = inB[pcs.inBOff + i2];
+
+    float theta = float(p);
+
+    if (!is_neox) {
+        for (uint i0 = 0; i0 < pcs.ne0; i0 += 2) {
+            float cos_theta, sin_theta;
+            rope_yarn(theta, pcs.freq_scale, corr_dims, i0, pcs.ext_factor, pcs.attn_factor, cos_theta, sin_theta);
+
+            theta *= theta_scale;
+
+            const uint src      = uint((i3*pcs.nb03 + i2*pcs.nb02 + i1*pcs.nb01 + i0*pcs.nb00) / 4) + pcs.inAOff; // Based from in
+            const uint dst_data = uint((i3*pcs.nb3  + i2*pcs.nb2  + i1*pcs.nb1  + i0*pcs.nb0)  / 4) + pcs.outOff; // Based from out_
+
+            const float x0 = inA[src];
+            const float x1 = inA[src+1];
+
+            out_[dst_data]   = x0*cos_theta - x1*sin_theta;
+            out_[dst_data+1] = x0*sin_theta + x1*cos_theta;
+        }
+    } else {
+        const float inv_ndims = -1.f/pcs.n_dims;
+        for (uint ic = 0; ic < pcs.n_dims; ic += 2) {
+            const uint cur_rot = ic;
+
+            float cos_theta, sin_theta;
+            rope_yarn(theta, pcs.freq_scale, corr_dims, cur_rot, pcs.ext_factor, pcs.attn_factor, cos_theta, sin_theta);
+
+            theta *= theta_scale;
+
+            const uint i0 = ic/2;
+
+            const uint src      = uint((i3*pcs.nb03 + i2*pcs.nb02 + i1*pcs.nb01 + i0*pcs.nb00) / 4) + pcs.inAOff; // Based from in
+            const uint dst_data = uint((i3*pcs.nb3  + i2*pcs.nb2  + i1*pcs.nb1  + i0*pcs.nb0)  / 4) + pcs.outOff; // Based from out_
+
+            const float x0 = inA[src];
+            const float x1 = inA[src+pcs.n_dims/2];
+
+            out_[dst_data] = x0*cos_theta - x1*sin_theta;
+            out_[dst_data+pcs.n_dims/2] = x0*sin_theta + x1*cos_theta;
+        }
+
+        for (uint ic = pcs.n_dims; ic < pcs.ne0; ic += 2) {
+            const uint i0 = ic;
+
+            const uint src = uint((i3*pcs.nb03 + i2*pcs.nb02 + i1*pcs.nb01 + i0*pcs.nb00) / 4) + pcs.inAOff; // Based from in
+            const uint dst_data = uint((i3*pcs.nb3  + i2*pcs.nb2  + i1*pcs.nb1  + i0*pcs.nb0) / 4) + pcs.outOff; // Based from out_
+
+            out_[dst_data + 0] = inA[src + 0];
+            out_[dst_data + 1] = inA[src + 1];
+        }
+    }
+}
diff --git a/src/whisper_cpp/ggml/src/kompute-shaders/op_scale.comp b/src/whisper_cpp/ggml/src/kompute-shaders/op_scale.comp
new file mode 100644
index 00000000..bdae2673
--- /dev/null
+++ b/src/whisper_cpp/ggml/src/kompute-shaders/op_scale.comp
@@ -0,0 +1,19 @@
+#version 450
+
+#include "common.comp"
+
+layout(local_size_x = 1) in;
+
+layout(binding = 0) buffer restrict readonly tensorIn { float in_[]; };
+layout(binding = 1) buffer restrict writeonly tensorOut { float out_[]; };
+
+layout(push_constant) uniform PushConstants {
+    uint inOff;
+    uint outOff;
+    float scale;
+} pcs;
+
+void main() {
+    const uint i = gl_WorkGroupID.x;
+    out_[i + pcs.outOff] = in_[i + pcs.inOff] * pcs.scale;
+}
diff --git a/src/whisper_cpp/ggml/src/kompute-shaders/op_scale_8.comp b/src/whisper_cpp/ggml/src/kompute-shaders/op_scale_8.comp
new file mode 100644
index 00000000..ada69754
--- /dev/null
+++ b/src/whisper_cpp/ggml/src/kompute-shaders/op_scale_8.comp
@@ -0,0 +1,23 @@
+#version 450
+
+#include "common.comp"
+
+layout(local_size_x = 1) in;
+
+layout(binding = 0) buffer restrict readonly tensorIn { float in_[]; };
+layout(binding = 1) buffer restrict writeonly tensorOut { float out_[]; };
+
+layout(push_constant) uniform PushConstants {
+    uint inOff;
+    uint outOff;
+    float scale;
+} pcs;
+
+void main() {
+    const uint baseIndex = gl_WorkGroupID.x * 8;
+
+    for (uint x = 0; x < 8; x++) {
+        const uint i = baseIndex + x;
+        out_[i + pcs.outOff] = in_[i + pcs.inOff] * pcs.scale;
+    }
+}
diff --git a/src/whisper_cpp/ggml/src/kompute-shaders/op_silu.comp b/src/whisper_cpp/ggml/src/kompute-shaders/op_silu.comp
new file mode 100644
index 00000000..0fb8e4b7
--- /dev/null
+++ b/src/whisper_cpp/ggml/src/kompute-shaders/op_silu.comp
@@ -0,0 +1,22 @@
+#version 450
+
+#include "common.comp"
+
+layout(local_size_x = 1) in;
+
+layout(binding = 0) buffer restrict readonly tensorIn { float in_[]; };
+layout(binding = 1) buffer restrict writeonly tensorOut { float out_[]; };
+layout(push_constant) uniform PushConstants {
+    uint inOff;
+    uint outOff;
+} pcs;
+
+void main() {
+    const uint baseIndex = gl_WorkGroupID.x * 4;
+
+    for (uint x = 0; x < 4; x++) {
+        const uint i = baseIndex + x;
+        const float y = in_[i + pcs.inOff];
+        out_[i + pcs.outOff] = y / (1.0 + exp(-y));
+    }
+}
diff --git a/src/whisper_cpp/ggml/src/kompute-shaders/op_softmax.comp b/src/whisper_cpp/ggml/src/kompute-shaders/op_softmax.comp
new file mode 100644
index 00000000..7bc9176c
--- /dev/null
+++ b/src/whisper_cpp/ggml/src/kompute-shaders/op_softmax.comp
@@ -0,0 +1,56 @@
+// TODO: implement multi-simd softmax (llama.cpp commit e16b9fa4)
+
+#version 450
+
+#include "common.comp"
+
+layout(local_size_x_id = 0) in;
+
+layout(binding = 0) buffer restrict readonly tensorInA { float inA[]; };
+layout(binding = 1) buffer restrict readonly tensorInB { float inB[]; };
+layout(binding = 2) buffer restrict writeonly tensorOut { float out_[]; };
+
+layout(push_constant) uniform PushConstants {
+    uint inAOff;
+    uint inBOff;
+    uint outOff;
+    int ne00;
+    int ne01;
+    int ne02;
+    float scale;
+    int mask;
+} pcs;
+
+void main() {
+    if (gl_SubgroupInvocationID > 31)
+        return;
+
+    const uint i03 = gl_WorkGroupID.z;
+    const uint i02 = gl_WorkGroupID.y;
+    const uint i01 = gl_WorkGroupID.x;
+
+    const uint extra_off = i03*pcs.ne02*pcs.ne01*pcs.ne00 + i02*pcs.ne01*pcs.ne00 + i01*pcs.ne00;
+    const uint psrc0 = extra_off + pcs.inAOff; // Based from inA
+    const uint pmask = i01*pcs.ne00 + pcs.inBOff; // Based from inB
+    const uint pdst = extra_off + pcs.outOff; // Based from out_
+
+    // parallel max
+    float localMax = uintBitsToFloat(0xFF800000);
+    for (uint i00 = gl_SubgroupInvocationID.x; i00 < pcs.ne00; i00 += 32) {
+        localMax = max(localMax, inA[psrc0 + i00]*pcs.scale + (pcs.mask!=0 ? inB[pmask + i00] : 0.0f));
+    }
+    float max_ = subgroupMax(localMax);
+
+    // parallel sum
+    float localSum = 0.0f;
+    for (uint i00 = gl_SubgroupInvocationID.x; i00 < pcs.ne00; i00 += 32) {
+        const float exp_psrc0 = exp(inA[psrc0 + i00]*pcs.scale + (pcs.mask!=0 ? inB[pmask + i00] : 0.0f) - max_);
+        localSum += exp_psrc0;
+        out_[pdst + i00] = exp_psrc0;
+    }
+
+    const float sum = subgroupAdd(localSum);
+    for (uint i00 = gl_SubgroupInvocationID.x; i00 < pcs.ne00; i00 += 32) {
+        out_[pdst + i00] /= sum;
+    }
+}
diff --git a/src/whisper_cpp/ggml/src/kompute-shaders/rope_common.comp b/src/whisper_cpp/ggml/src/kompute-shaders/rope_common.comp
new file mode 100644
index 00000000..7b9394cb
--- /dev/null
+++ b/src/whisper_cpp/ggml/src/kompute-shaders/rope_common.comp
@@ -0,0 +1,67 @@
+#include "common.comp"
+
+// TODO: use a local size of 32 or more (Metal uses 1024)
+layout(local_size_x = 1) in;
+
+layout (push_constant) uniform parameter {
+    uint inAOff;
+    uint inBOff;
+    uint outOff;
+    int n_dims;
+    int mode;
+    int n_ctx_orig;
+    float freq_base;
+    float freq_scale;
+    float ext_factor;
+    float attn_factor;
+    float beta_fast;
+    float beta_slow;
+    uint nb00;
+    uint nb01;
+    uint nb02;
+    uint nb03;
+    int ne0;
+    uint nb0;
+    uint nb1;
+    uint nb2;
+    uint nb3;
+} pcs;
+
+float rope_yarn_ramp(const float low, const float high, const float i0) {
+    const float y = (i0 / 2 - low) / max(0.001f, high - low);
+    return 1.0f - min(1.0f, max(0.0f, y));
+}
+
+// YaRN algorithm based on LlamaYaRNScaledRotaryEmbedding.py from https://github.com/jquesnelle/yarn
+// MIT licensed. Copyright (c) 2023 Jeffrey Quesnelle and Bowen Peng.
+void rope_yarn(
+    float theta_extrap, float freq_scale, float corr_dims[2], float i0, float ext_factor, float mscale,
+    out float cos_theta, out float sin_theta
+) {
+    // Get n-d rotational scaling corrected for extrapolation
+    float theta_interp = freq_scale * theta_extrap;
+    float theta = theta_interp;
+    if (ext_factor != 0.0f) {
+        float ramp_mix = rope_yarn_ramp(corr_dims[0], corr_dims[1], i0) * ext_factor;
+        theta = theta_interp * (1 - ramp_mix) + theta_extrap * ramp_mix;
+
+        // Get n-d magnitude scaling corrected for interpolation
+        mscale *= 1.0f + 0.1f * log(1.0f / freq_scale);
+    }
+    cos_theta = cos(theta) * mscale;
+    sin_theta = sin(theta) * mscale;
+}
+
+// Apparently solving `n_rot = 2pi * x * base^((2 * max_pos_emb) / n_dims)` for x, we get
+// `corr_fac(n_rot) = n_dims * log(max_pos_emb / (n_rot * 2pi)) / (2 * log(base))`
+float rope_yarn_corr_factor(int n_dims, int n_ctx_orig, float n_rot, float base) {
+    return n_dims * log(n_ctx_orig / (n_rot * TWOPI_F)) / (2 * log(base));
+}
+
+void rope_yarn_corr_dims(
+    int n_dims, int n_ctx_orig, float freq_base, float beta_fast, float beta_slow, out float dims[2]
+) {
+    // start and end correction dims
+    dims[0] = max(0.0f,         floor(rope_yarn_corr_factor(n_dims, n_ctx_orig, beta_fast, freq_base)));
+    dims[1] = min(n_dims - 1.0f, ceil(rope_yarn_corr_factor(n_dims, n_ctx_orig, beta_slow, freq_base)));
+}
diff --git a/src/whisper_cpp/ggml/src/sgemm.cpp b/src/whisper_cpp/ggml/src/sgemm.cpp
new file mode 100644
index 00000000..6626ceb2
--- /dev/null
+++ b/src/whisper_cpp/ggml/src/sgemm.cpp
@@ -0,0 +1,1027 @@
+// Copyright 2024 Mozilla Foundation
+//
+// Permission is hereby granted, free of charge, to any person obtaining
+// a copy of this software and associated documentation files (the
+// "Software"), to deal in the Software without restriction, including
+// without limitation the rights to use, copy, modify, merge, publish,
+// distribute, sublicense, and/or sell copies of the Software, and to
+// permit persons to whom the Software is furnished to do so, subject to
+// the following conditions:
+//
+// The above copyright notice and this permission notice shall be
+// included in all copies or substantial portions of the Software.
+//
+// THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
+// EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
+// MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
+// NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
+// BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
+// ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
+// CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+// SOFTWARE.
+
+//
+//                   _   _          ___ _      _   ___
+//                  | |_(_)_ _ _  _| _ ) |    /_\ / __|
+//                  |  _| | ' \ || | _ \ |__ / _ \\__ \.
+//                   \__|_|_||_\_, |___/____/_/ \_\___/
+//                             |__/
+//
+//                    BASIC LINEAR ALGEBRA SUBPROGRAMS
+//
+//
+// This file implements multithreaded CPU matrix multiplication for the
+// common contiguous use case C = Aᵀ * B. These kernels are designed to
+// have excellent performance[1] for matrices that fit in the CPU cache
+// without imposing any overhead such as cache filling or malloc calls.
+//
+// This implementation does not guarantee any upper bound with rounding
+// errors, which grow along with k. Our goal's to maximally exploit the
+// hardware for performance, and then use whatever resources remain for
+// improving numerical accuracy.
+//
+// [1] J. Tunney, ‘LLaMA Now Goes Faster on CPUs’, Mar. 2024. [Online].
+//     Available: https://justine.lol/matmul/. [Accessed: 29-Mar-2024].
+
+#if defined(__GNUC__)
+#pragma GCC diagnostic ignored "-Wpedantic"
+#pragma GCC diagnostic ignored "-Wignored-attributes"
+#endif
+
+#include "sgemm.h"
+#include "ggml-impl.h"
+#include "ggml-quants.h"
+
+#ifdef _MSC_VER
+#define NOINLINE __declspec(noinline)
+#else
+#define NOINLINE __attribute__((__noinline__))
+#endif
+
+#if defined(__ARM_NEON) || defined(__AVX512F__)
+#define VECTOR_REGISTERS 32
+#else
+#define VECTOR_REGISTERS 16
+#endif
+
+#define MM256_SET_M128I(a, b) _mm256_insertf128_si256(_mm256_castsi128_si256(b), (a), 1)
+
+namespace {
+
+inline float unhalf(ggml_fp16_t d) {
+    return GGML_FP16_TO_FP32(d);
+}
+
+////////////////////////////////////////////////////////////////////////////////////////////////////
+// VECTORIZED ARITHMETIC OPERATIONS
+
+#if defined(__SSE__) || defined(__AVX__) || defined(__AVX2__) || defined(__AVX512F__)
+inline __m128 add(__m128 x, __m128 y) { return _mm_add_ps(x, y); }
+inline __m128 sub(__m128 x, __m128 y) { return _mm_sub_ps(x, y); }
+inline __m128 mul(__m128 x, __m128 y) { return _mm_mul_ps(x, y); }
+#endif  // __SSE__
+
+#if defined(__AVX__) || defined(__AVX2__) || defined(__AVX512F__)
+inline __m256 add(__m256 x, __m256 y) { return _mm256_add_ps(x, y); }
+inline __m256 sub(__m256 x, __m256 y) { return _mm256_sub_ps(x, y); }
+inline __m256 mul(__m256 x, __m256 y) { return _mm256_mul_ps(x, y); }
+#endif // __AVX__
+
+#if defined(__AVX512F__)
+inline __m512 add(__m512 x, __m512 y) { return _mm512_add_ps(x, y); }
+inline __m512 sub(__m512 x, __m512 y) { return _mm512_sub_ps(x, y); }
+inline __m512 mul(__m512 x, __m512 y) { return _mm512_mul_ps(x, y); }
+#endif // __AVX512F__
+
+#if defined(__ARM_NEON)
+inline float32x4_t add(float32x4_t x, float32x4_t y) { return vaddq_f32(x, y); }
+inline float32x4_t sub(float32x4_t x, float32x4_t y) { return vsubq_f32(x, y); }
+inline float32x4_t mul(float32x4_t x, float32x4_t y) { return vmulq_f32(x, y); }
+#endif // __ARM_NEON
+
+#if defined(__ARM_FEATURE_FP16_VECTOR_ARITHMETIC)
+inline float16x8_t add(float16x8_t x, float16x8_t y) { return vaddq_f16(x, y); }
+inline float16x8_t sub(float16x8_t x, float16x8_t y) { return vsubq_f16(x, y); }
+inline float16x8_t mul(float16x8_t x, float16x8_t y) { return vmulq_f16(x, y); }
+#endif // __ARM_FEATURE_FP16_VECTOR_ARITHMETIC
+
+////////////////////////////////////////////////////////////////////////////////////////////////////
+// VECTORIZED FUSED MULTIPLY ADD
+
+/**
+ * Computes a * b + c.
+ */
+template <typename T, typename U>
+inline U madd(T a, T b, U c) {
+    return add(mul(a, b), c);
+}
+
+#if defined(__FMA__)
+#if defined(__AVX__) || defined(__AVX2__) || defined(__AVX512F__)
+template <>
+inline __m256 madd(__m256 a, __m256 b, __m256 c) {
+    return _mm256_fmadd_ps(a, b, c);
+}
+#endif
+#if defined(__AVX512F__)
+template <>
+inline __m512 madd(__m512 a, __m512 b, __m512 c) {
+    return _mm512_fmadd_ps(a, b, c);
+}
+#endif
+#endif
+
+#if defined(__ARM_FEATURE_FMA)
+template <>
+inline float32x4_t madd(float32x4_t a, float32x4_t b, float32x4_t c) {
+    return vfmaq_f32(c, b, a);
+}
+#if defined(__ARM_FEATURE_FP16_VECTOR_ARITHMETIC) && !defined(_MSC_VER)
+template <>
+inline float16x8_t madd(float16x8_t a, float16x8_t b, float16x8_t c) {
+    return vfmaq_f16(c, b, a);
+}
+#endif
+#endif
+
+////////////////////////////////////////////////////////////////////////////////////////////////////
+// VECTORIZED HORIZONTAL SUM
+
+#if defined(__ARM_NEON)
+inline float hsum(float32x4_t x) {
+    return vaddvq_f32(x);
+}
+#endif // __ARM_NEON
+
+#if defined(__ARM_FEATURE_FP16_VECTOR_ARITHMETIC) && !defined(_MSC_VER)
+inline float hsum(float16x8_t x) {
+    return vaddvq_f32(vaddq_f32(vcvt_f32_f16(vget_low_f16(x)),
+                                vcvt_f32_f16(vget_high_f16(x))));
+}
+#endif // __ARM_FEATURE_FP16_VECTOR_ARITHMETIC
+
+#if defined(__SSE__) || defined(__AVX__) || defined(__AVX2__) || defined(__AVX512F__)
+inline float hsum(__m128 x) {
+#if defined(__AVX__) || defined(__AVX2__) || defined(__AVX512F__)
+    x = _mm_add_ps(x, _mm_movehl_ps(x, x));
+    x = _mm_add_ss(x, _mm_movehdup_ps(x));
+#else
+    __m128 t;
+    t = _mm_shuffle_ps(x, x, _MM_SHUFFLE(2, 3, 0, 1));
+    x = _mm_add_ps(x, t);
+    t = _mm_movehl_ps(t, x);
+    x = _mm_add_ss(x, t);
+#endif
+    return _mm_cvtss_f32(x);
+}
+#endif
+
+#if defined(__AVX__) || defined(__AVX2__) || defined(__AVX512F__)
+inline float hsum(__m256 x) {
+    return hsum(_mm_add_ps(_mm256_extractf128_ps(x, 1),
+                           _mm256_castps256_ps128(x)));
+}
+#endif // __AVX__
+
+#if defined(__AVX512F__)
+inline float hsum(__m512 x) {
+    return _mm512_reduce_add_ps(x);
+}
+#endif // __AVX512F__
+
+////////////////////////////////////////////////////////////////////////////////////////////////////
+// VECTORIZED MEMORY LOADING
+
+template <typename T, typename U> T load(const U *);
+
+#if defined(__ARM_NEON)
+template <> inline float32x4_t load(const float *p) {
+    return vld1q_f32(p);
+}
+#if !defined(_MSC_VER)
+template <> inline float16x8_t load(const ggml_fp16_t *p) {
+    return vld1q_f16((const float16_t *)p);
+}
+template <> inline float32x4_t load(const ggml_fp16_t *p) {
+    return vcvt_f32_f16(vld1_f16((const float16_t *)p));
+}
+#endif // _MSC_VER
+#endif // __ARM_NEON
+
+#if defined(__SSE__) || defined(__AVX__) || defined(__AVX2__) || defined(__AVX512F__)
+template <> inline __m128 load(const float *p) {
+    return _mm_loadu_ps(p);
+}
+#endif  // __SSE__
+
+#if defined(__AVX__) || defined(__AVX2__) || defined(__AVX512F__)
+template <> inline __m256 load(const float *p) {
+    return _mm256_loadu_ps(p);
+}
+#endif // __AVX__
+
+#if defined(__F16C__)
+template <> inline __m256 load(const ggml_fp16_t *p) {
+    return _mm256_cvtph_ps(_mm_loadu_si128((const __m128i *)p));
+}
+#endif // __F16C__
+
+#if defined(__AVX512F__)
+template <> inline __m512 load(const float *p) {
+    return _mm512_loadu_ps(p);
+}
+template <> inline __m512 load(const ggml_fp16_t *p) {
+    return _mm512_cvtph_ps(_mm256_loadu_si256((const __m256i *)p));
+}
+#endif // __AVX512F__
+
+////////////////////////////////////////////////////////////////////////////////////////////////////
+// FLOATING POINT MATRIX MULTIPLICATION
+
+template <int KN, typename D, typename V, typename TA, typename TB, typename TC>
+class tinyBLAS {
+  public:
+    tinyBLAS(int64_t k,
+             const TA *A, int64_t lda,
+             const TB *B, int64_t ldb,
+             TC *C, int64_t ldc,
+             int ith, int nth)
+        : A(A), B(B), C(C), k(k), lda(lda), ldb(ldb), ldc(ldc), ith(ith), nth(nth) {
+    }
+
+    void matmul(int64_t m, int64_t n) {
+        mnpack(0, m, 0, n);
+    }
+
+  private:
+    NOINLINE void mnpack(int64_t m0, int64_t m, int64_t n0, int64_t n) {
+        int64_t mc, nc, mp, np;
+        switch ((MIN(m - m0, 5) << 4) | MIN(n - n0, 5)) {
+#if VECTOR_REGISTERS == 32
+        case 0x55:
+            mc = 5;
+            nc = 5;
+            gemm<5, 5>(m0, m, n0, n);
+            break;
+        case 0x45:
+            mc = 4;
+            nc = 5;
+            gemm<4, 5>(m0, m, n0, n);
+            break;
+        case 0x54:
+            mc = 5;
+            nc = 4;
+            gemm<5, 4>(m0, m, n0, n);
+            break;
+        case 0x44:
+            mc = 4;
+            nc = 4;
+            gemm<4, 4>(m0, m, n0, n);
+            break;
+        case 0x53:
+            mc = 5;
+            nc = 3;
+            gemm<5, 3>(m0, m, n0, n);
+            break;
+        case 0x35:
+            mc = 3;
+            nc = 5;
+            gemm<3, 5>(m0, m, n0, n);
+            break;
+        case 0x43:
+            mc = 4;
+            nc = 3;
+            gemm<4, 3>(m0, m, n0, n);
+            break;
+#else
+        case 0x55:
+        case 0x54:
+        case 0x53:
+        case 0x45:
+        case 0x44:
+        case 0x43:
+            mc = 4;
+            nc = 3;
+            gemm<4, 3>(m0, m, n0, n);
+            break;
+        case 0x35:
+#endif
+        case 0x34:
+            mc = 3;
+            nc = 4;
+            gemm<3, 4>(m0, m, n0, n);
+            break;
+        case 0x52:
+            mc = 5;
+            nc = 2;
+            gemm<5, 2>(m0, m, n0, n);
+            break;
+        case 0x33:
+            mc = 3;
+            nc = 3;
+            gemm<3, 3>(m0, m, n0, n);
+            break;
+        case 0x25:
+            mc = 2;
+            nc = 5;
+            gemm<2, 5>(m0, m, n0, n);
+            break;
+        case 0x42:
+            mc = 4;
+            nc = 2;
+            gemm<4, 2>(m0, m, n0, n);
+            break;
+        case 0x24:
+            mc = 2;
+            nc = 4;
+            gemm<2, 4>(m0, m, n0, n);
+            break;
+        case 0x32:
+            mc = 3;
+            nc = 2;
+            gemm<3, 2>(m0, m, n0, n);
+            break;
+        case 0x23:
+            mc = 2;
+            nc = 3;
+            gemm<2, 3>(m0, m, n0, n);
+            break;
+        case 0x51:
+            mc = 5;
+            nc = 1;
+            gemm<5, 1>(m0, m, n0, n);
+            break;
+        case 0x41:
+            mc = 4;
+            nc = 1;
+            gemm<4, 1>(m0, m, n0, n);
+            break;
+        case 0x22:
+            mc = 2;
+            nc = 2;
+            gemm<2, 2>(m0, m, n0, n);
+            break;
+        case 0x15:
+            mc = 1;
+            nc = 5;
+            gemm<1, 5>(m0, m, n0, n);
+            break;
+        case 0x14:
+            mc = 1;
+            nc = 4;
+            gemm<1, 4>(m0, m, n0, n);
+            break;
+        case 0x31:
+            mc = 3;
+            nc = 1;
+            gemm<3, 1>(m0, m, n0, n);
+            break;
+        case 0x13:
+            mc = 1;
+            nc = 3;
+            gemm<1, 3>(m0, m, n0, n);
+            break;
+        case 0x21:
+            mc = 2;
+            nc = 1;
+            gemm<2, 1>(m0, m, n0, n);
+            break;
+        case 0x12:
+            mc = 1;
+            nc = 2;
+            gemm<1, 2>(m0, m, n0, n);
+            break;
+        case 0x11:
+            mc = 1;
+            nc = 1;
+            gemm<1, 1>(m0, m, n0, n);
+            break;
+        default:
+            return;
+        }
+        mp = m0 + (m - m0) / mc * mc;
+        np = n0 + (n - n0) / nc * nc;
+        mnpack(mp, m, n0, np);
+        mnpack(m0, m, np, n);
+    }
+
+    template <int RM, int RN>
+    NOINLINE void gemm(int64_t m0, int64_t m, int64_t n0, int64_t n) {
+        int64_t ytiles = (m - m0) / RM;
+        int64_t xtiles = (n - n0) / RN;
+        int64_t tiles = xtiles * ytiles;
+        int64_t duty = (tiles + nth - 1) / nth;
+        int64_t start = duty * ith;
+        int64_t end = start + duty;
+        if (end > tiles)
+            end = tiles;
+        for (int64_t job = start; job < end; ++job) {
+            int64_t ii = m0 + job / xtiles * RM;
+            int64_t jj = n0 + job % xtiles * RN;
+            D Cv[RN][RM] = {};
+            for (int64_t l = 0; l < k; l += KN)
+                for (int64_t j = 0; j < RN; ++j)
+                    for (int64_t i = 0; i < RM; ++i)
+                        Cv[j][i] = madd(load<V>(A + lda * (ii + i) + l),
+                                        load<V>(B + ldb * (jj + j) + l),
+                                        Cv[j][i]);
+            for (int64_t j = 0; j < RN; ++j)
+                for (int64_t i = 0; i < RM; ++i)
+                    C[ldc * (jj + j) + (ii + i)] = hsum(Cv[j][i]);
+        }
+    }
+
+    const TA *const A;
+    const TB *const B;
+    TC *const C;
+    const int64_t k;
+    const int64_t lda;
+    const int64_t ldb;
+    const int64_t ldc;
+    const int ith;
+    const int nth;
+};
+
+//////////////////////////////////////////////////////////////////////////////////////////
+// QUANT ZERO MATRIX MULTIPLICATION
+
+#if defined(__ARM_FEATURE_DOTPROD)
+template <typename TA>
+class tinyBLAS_Q0_ARM {
+  public:
+    tinyBLAS_Q0_ARM(int64_t k,
+                    const TA *A, int64_t lda,
+                    const block_q8_0 *B, int64_t ldb,
+                    float *C, int64_t ldc,
+                    int ith, int nth)
+        : A(A), B(B), C(C), k(k), lda(lda), ldb(ldb), ldc(ldc), ith(ith), nth(nth) {
+    }
+
+    void matmul(int64_t m, int64_t n) {
+        mnpack(0, m, 0, n);
+    }
+
+  private:
+    NOINLINE void mnpack(int64_t m0, int64_t m, int64_t n0, int64_t n) {
+        int64_t mc, nc, mp, np;
+        switch ((MIN(m - m0, 3) << 4) | MIN(n - n0, 3ll)) {
+        case 0x33:
+            mc = 3;
+            nc = 3;
+            gemm<3, 3>(m0, m, n0, n);
+            break;
+        case 0x32:
+            mc = 3;
+            nc = 2;
+            gemm<3, 2>(m0, m, n0, n);
+            break;
+        case 0x23:
+            mc = 2;
+            nc = 3;
+            gemm<2, 3>(m0, m, n0, n);
+            break;
+        case 0x22:
+            mc = 2;
+            nc = 2;
+            gemm<2, 2>(m0, m, n0, n);
+            break;
+        case 0x31:
+            mc = 3;
+            nc = 1;
+            gemm<3, 1>(m0, m, n0, n);
+            break;
+        case 0x13:
+            mc = 1;
+            nc = 3;
+            gemm<1, 3>(m0, m, n0, n);
+            break;
+        case 0x21:
+            mc = 2;
+            nc = 1;
+            gemm<2, 1>(m0, m, n0, n);
+            break;
+        case 0x12:
+            mc = 1;
+            nc = 2;
+            gemm<1, 2>(m0, m, n0, n);
+            break;
+        case 0x11:
+            mc = 1;
+            nc = 1;
+            gemm<1, 1>(m0, m, n0, n);
+            break;
+        default:
+            return;
+        }
+        mp = m0 + (m - m0) / mc * mc;
+        np = n0 + (n - n0) / nc * nc;
+        mnpack(mp, m, n0, np);
+        mnpack(m0, m, np, n);
+    }
+
+    template <int RM, int RN>
+    NOINLINE void gemm(int64_t m0, int64_t m, int64_t n0, int64_t n) {
+        int64_t ytiles = (m - m0) / RM;
+        int64_t xtiles = (n - n0) / RN;
+        int64_t tiles = xtiles * ytiles;
+        int64_t duty = (tiles + nth - 1) / nth;
+        int64_t start = duty * ith;
+        int64_t end = start + duty;
+        if (end > tiles)
+            end = tiles;
+        for (int64_t job = start; job < end; ++job) {
+            int64_t ii = m0 + job / xtiles * RM;
+            int64_t jj = n0 + job % xtiles * RN;
+            float32x4_t Cv[RN][RM] = {};
+            for (int64_t l = 0; l < k; ++l)
+                for (int64_t j = 0; j < RN; ++j)
+                    for (int64_t i = 0; i < RM; ++i)
+                        Cv[j][i] = vmlaq_n_f32(Cv[j][i],
+                                               vcvtq_f32_s32(vdotq_s32(
+                                                   vdotq_s32(vdupq_n_s32(0),
+                                                             load_lo(A + lda * (ii + i) + l),
+                                                             load_lo(B + ldb * (jj + j) + l)),
+                                                   load_hi(A + lda * (ii + i) + l),
+                                                   load_hi(B + ldb * (jj + j) + l))),
+                                               unhalf(A[lda * (ii + i) + l].d) *
+                                               unhalf(B[ldb * (jj + j) + l].d));
+            for (int64_t j = 0; j < RN; ++j)
+                for (int64_t i = 0; i < RM; ++i)
+                    C[ldc * (jj + j) + (ii + i)] = hsum(Cv[j][i]);
+        }
+    }
+
+    inline int8x16_t load_lo(const block_q8_0 *b) {
+        return vld1q_s8(b->qs);
+    }
+
+    inline int8x16_t load_hi(const block_q8_0 *b) {
+        return vld1q_s8(b->qs + 16);
+    }
+
+    inline int8x16_t load_lo(const block_q4_0 *b) {
+        return vsubq_s8(vreinterpretq_s8_u8(vandq_u8(vld1q_u8(b->qs),
+                                                     vdupq_n_u8(0x0f))),
+                        vdupq_n_s8(0x8));
+    }
+
+    inline int8x16_t load_hi(const block_q4_0 *b) {
+        return vsubq_s8(vreinterpretq_s8_u8(vshrq_n_u8(vld1q_u8(b->qs), 4)),
+                        vdupq_n_s8(0x8));
+    }
+
+    const TA *const A;
+    const block_q8_0 *const B;
+    float *const C;
+    const int64_t k;
+    const int64_t lda;
+    const int64_t ldb;
+    const int64_t ldc;
+    const int ith;
+    const int nth;
+};
+#endif // __ARM_FEATURE_DOTPROD
+
+#if defined(__AVX2__) || defined(__AVX512F__) || defined(__AVX__)
+template <typename TA, typename TB, typename TC>
+class tinyBLAS_Q0_AVX {
+  public:
+    tinyBLAS_Q0_AVX(int64_t k,
+                    const TA *A, int64_t lda,
+                    const TB *B, int64_t ldb,
+                    TC *C, int64_t ldc,
+                    int ith, int nth)
+        : A(A), B(B), C(C), k(k), lda(lda), ldb(ldb), ldc(ldc), ith(ith), nth(nth) {
+    }
+
+    void matmul(int64_t m, int64_t n) {
+        mnpack(0, m, 0, n);
+    }
+
+  private:
+    void mnpack(int64_t m0, int64_t m, int64_t n0, int64_t n) {
+        int64_t mc, nc, mp, np;
+        switch ((MIN(m - m0, 4) << 4) | MIN(n - n0, 4)) {
+#if VECTOR_REGISTERS == 32
+        case 0x44:
+            mc = 4;
+            nc = 4;
+            gemm<4, 4>(m0, m, n0, n);
+            break;
+        case 0x43:
+            mc = 4;
+            nc = 3;
+            gemm<4, 3>(m0, m, n0, n);
+            break;
+        case 0x34:
+            mc = 3;
+            nc = 4;
+            gemm<3, 4>(m0, m, n0, n);
+            break;
+        case 0x33:
+            mc = 3;
+            nc = 3;
+            gemm<3, 3>(m0, m, n0, n);
+            break;
+        case 0x42:
+            mc = 4;
+            nc = 2;
+            gemm<4, 2>(m0, m, n0, n);
+            break;
+        case 0x24:
+            mc = 2;
+            nc = 4;
+            gemm<2, 4>(m0, m, n0, n);
+            break;
+#else
+        case 0x44:
+        case 0x43:
+        case 0x42:
+            mc = 4;
+            nc = 2;
+            gemm<4, 2>(m0, m, n0, n);
+            break;
+        case 0x34:
+        case 0x24:
+            mc = 2;
+            nc = 4;
+            gemm<2, 4>(m0, m, n0, n);
+            break;
+        case 0x33:
+#endif
+        case 0x32:
+            mc = 3;
+            nc = 2;
+            gemm<3, 2>(m0, m, n0, n);
+            break;
+        case 0x23:
+            mc = 2;
+            nc = 3;
+            gemm<2, 3>(m0, m, n0, n);
+            break;
+        case 0x41:
+            mc = 4;
+            nc = 1;
+            gemm<4, 1>(m0, m, n0, n);
+            break;
+        case 0x22:
+            mc = 2;
+            nc = 2;
+            gemm<2, 2>(m0, m, n0, n);
+            break;
+        case 0x14:
+            mc = 1;
+            nc = 4;
+            gemm<1, 4>(m0, m, n0, n);
+            break;
+        case 0x31:
+            mc = 3;
+            nc = 1;
+            gemm<3, 1>(m0, m, n0, n);
+            break;
+        case 0x13:
+            mc = 1;
+            nc = 3;
+            gemm<1, 3>(m0, m, n0, n);
+            break;
+        case 0x21:
+            mc = 2;
+            nc = 1;
+            gemm<2, 1>(m0, m, n0, n);
+            break;
+        case 0x12:
+            mc = 1;
+            nc = 2;
+            gemm<1, 2>(m0, m, n0, n);
+            break;
+        case 0x11:
+            mc = 1;
+            nc = 1;
+            gemm<1, 1>(m0, m, n0, n);
+            break;
+        default:
+            return;
+        }
+        mp = m0 + (m - m0) / mc * mc;
+        np = n0 + (n - n0) / nc * nc;
+        mnpack(mp, m, n0, np);
+        mnpack(m0, m, np, n);
+    }
+
+    template <int RM, int RN>
+    NOINLINE void gemm(int64_t m0, int64_t m, int64_t n0, int64_t n) {
+        int64_t ytiles = (m - m0) / RM;
+        int64_t xtiles = (n - n0) / RN;
+        int64_t tiles = xtiles * ytiles;
+        int64_t duty = (tiles + nth - 1) / nth;
+        int64_t start = duty * ith;
+        int64_t end = start + duty;
+        if (end > tiles)
+            end = tiles;
+        for (int64_t job = start; job < end; ++job) {
+            int64_t ii = m0 + job / xtiles * RM;
+            int64_t jj = n0 + job % xtiles * RN;
+            __m256 Cv[RN][RM] = {};
+            for (int64_t l = 0; l < k; ++l)
+                for (int64_t j = 0; j < RN; ++j)
+                    for (int64_t i = 0; i < RM; ++i) {
+#if defined(__AVX2__)
+                        __m256 udTmp = updot(_mm256_sign_epi8(load(A + lda * (ii + i) + l),
+                                                              load(A + lda * (ii + i) + l)),
+                                             _mm256_sign_epi8(load(B + ldb * (jj + j) + l),
+                                                              load(A + lda * (ii + i) + l)));
+#else
+                        __m128i ali0 = load0(A + lda * (ii + i) + l);
+                        __m128i ali1 = load1(A + lda * (ii + i) + l);
+                        __m128i blj0 = load0(B + ldb * (jj + j) + l);
+                        __m128i blj1 = load1(B + ldb * (jj + j) + l);
+
+                        __m128i sepAA0 = _mm_sign_epi8(ali0, ali0);
+                        __m128i sepAA1 = _mm_sign_epi8(ali1, ali1);
+                        __m128i sepBA0 = _mm_sign_epi8(blj0, ali0);
+                        __m128i sepBA1 = _mm_sign_epi8(blj1, ali1);
+
+                        // updot
+                        const __m128i oneFill = _mm_set1_epi16(1);
+                        __m128i mad0 = _mm_maddubs_epi16(sepAA0, sepBA0);
+                        __m128i mad1 = _mm_maddubs_epi16(sepAA1, sepBA1);
+                        __m256 udTmp = _mm256_cvtepi32_ps(MM256_SET_M128I(_mm_madd_epi16(oneFill, mad1), _mm_madd_epi16(oneFill, mad0)));
+#endif
+                        Cv[j][i] = madd(_mm256_set1_ps(unhalf(A[lda * (ii + i) + l].d) *
+                                                       unhalf(B[ldb * (jj + j) + l].d)),
+                                                       udTmp,
+                                                       Cv[j][i]);
+                    }
+            for (int64_t j = 0; j < RN; ++j)
+                for (int64_t i = 0; i < RM; ++i)
+                    C[ldc * (jj + j) + (ii + i)] = hsum(Cv[j][i]);
+        }
+    }
+
+    inline __m256i load(const block_q8_0 *b) {
+        return _mm256_loadu_si256((const __m256i *)b->qs);
+    }
+
+    inline __m128i load0(const block_q8_0 *b) {
+        return _mm_loadu_si128((const __m128i *)b->qs);
+    }
+
+    inline __m128i load1(const block_q8_0 *b) {
+        return _mm_loadu_si128(((const __m128i *)b->qs) + 1);
+    }
+
+    inline __m256i load(const block_q4_0 *b) {
+        return _mm256_sub_epi8(denibble(b->qs), _mm256_set1_epi8(8));
+    }
+
+    inline __m128i load0(const block_q4_0 *b) {
+        const __m128i x = _mm_loadu_si128((const __m128i *)(b->qs));
+        return _mm_sub_epi8(_mm_and_si128(_mm_set1_epi8(15), x), _mm_set1_epi8(8));
+    }
+
+    inline __m128i load1(const block_q4_0 *b) {
+        const __m128i x = _mm_loadu_si128((const __m128i *)(b->qs));
+        return _mm_sub_epi8(_mm_and_si128(_mm_set1_epi8(15), _mm_srli_epi16(x, 4)), _mm_set1_epi8(8));
+    }
+
+    inline __m256 updot(__m256i u, __m256i s) {
+        __m256i res;
+#if defined(__AVXVNNI__) || (defined(__AVX512VNNI__) && defined(__AVX512VL__))
+        res = _mm256_dpbusd_epi32(_mm256_setzero_si256(), u, s);
+#else
+        res = _mm256_madd_epi16(_mm256_set1_epi16(1), _mm256_maddubs_epi16(u, s));
+#endif
+        return _mm256_cvtepi32_ps(res);
+    }
+
+    static inline __m256i denibble(const uint8_t *p) {
+        __m128i x = _mm_loadu_si128((const __m128i *)p);
+        return _mm256_and_si256(_mm256_set1_epi8(15),
+                                _mm256_insertf128_si256(_mm256_castsi128_si256(x),
+                                                        _mm_srli_epi16(x, 4), 1));
+    }
+
+    const TA *const A;
+    const TB *const B;
+    TC *const C;
+    const int64_t k;
+    const int64_t lda;
+    const int64_t ldb;
+    const int64_t ldc;
+    const int ith;
+    const int nth;
+};
+#endif // __AVX__
+
+} // namespace
+
+/**
+ * Performs optimized matrix multiplication on CPU.
+ *
+ * This subroutine may compute C = Aᵀ * B with column major ordering.
+ * Despite its name, this isn't a generalized implementation. Work is
+ * only performed when a handwritten kernel is written and available.
+ * Otherwise the caller should fall back to a general matmul routine.
+ *
+ * For example, for single-threaded single-precision GEMM you can say
+ *
+ *     llamafile_sgemm(m, n, k, A, lda, B, ldb, C, ldc,
+ *                     0, 1,
+ *                     GGML_TYPE_F32, GGML_TYPE_F32, GGML_TYPE_F32);
+ *
+ * @param m is rows in `A` and `C`
+ * @param n is cols in `B` and `C`
+ * @param k is cols in `A` and rows in `B`
+ * @param A is first input matrix (always transposed)
+ * @param lda is row stride of `A`
+ * @param B is second input matrix (never transposed)
+ * @param ldb is row stride of `B`
+ * @param C is input/output array of output matrices
+ * @param ldc is row stride of `C`
+ * @param ith is thread id (must be less than `nth`)
+ * @param nth is number of threads (must be greater than zero)
+ * @param Atype is GGML data type of `A`
+ * @param Btype is GGML data type of `B`
+ * @param Ctype is GGML data type of `C`
+ * @return true if this function was able to service the matmul request
+ */
+bool llamafile_sgemm(int64_t m, int64_t n, int64_t k, const void *A, int64_t lda, const void *B, int64_t ldb, void *C,
+                     int64_t ldc, int ith, int nth, int Atype, int Btype, int Ctype) {
+
+    assert(m >= 0);
+    assert(n >= 0);
+    assert(k >= 0);
+    assert(lda >= k);
+    assert(ldb >= k);
+    assert(ldc >= m);
+    assert(nth > 0);
+    assert(ith < nth);
+
+    if (Ctype != GGML_TYPE_F32)
+        return false;
+
+    switch (Atype) {
+
+    case GGML_TYPE_F32: {
+        if (Btype != GGML_TYPE_F32)
+            return false;
+#if defined(__AVX512F__)
+        if (k % 16)
+            return false;
+        tinyBLAS<16, __m512, __m512, float, float, float> tb{
+            k, (const float *)A, lda,
+            (const float *)B, ldb,
+            (float *)C, ldc,
+            ith, nth};
+        tb.matmul(m, n);
+        return true;
+#elif defined(__AVX__) || defined(__AVX2__)
+        if (k % 8)
+            return false;
+        tinyBLAS<8, __m256, __m256, float, float, float> tb{
+            k, (const float *)A, lda,
+            (const float *)B, ldb,
+            (float *)C, ldc,
+            ith, nth};
+        tb.matmul(m, n);
+        return true;
+#elif defined(__ARM_NEON)
+        if (n < 4)
+            return false;
+        if (k % 4)
+            return false;
+        tinyBLAS<4, float32x4_t, float32x4_t, float, float, float> tb{
+            k, (const float *)A, lda,
+            (const float *)B, ldb,
+            (float *)C, ldc,
+            ith, nth};
+        tb.matmul(m, n);
+        return true;
+#else
+        return false;
+#endif
+    }
+
+    case GGML_TYPE_F16: {
+#if defined(__AVX512F__)
+        if (k % 16)
+            return false;
+        if (Btype != GGML_TYPE_F32)
+            return false;
+        tinyBLAS<16, __m512, __m512, ggml_fp16_t, float, float> tb{
+            k, (const ggml_fp16_t *)A, lda,
+            (const float *)B, ldb,
+            (float *)C, ldc,
+            ith, nth};
+        tb.matmul(m, n);
+        return true;
+#elif (defined(__AVX__) || defined(__AVX2__)) && defined(__F16C__)
+        if (k % 8)
+            return false;
+        if (Btype != GGML_TYPE_F32)
+            return false;
+        tinyBLAS<8, __m256, __m256, ggml_fp16_t, float, float> tb{
+            k, (const ggml_fp16_t *)A, lda,
+            (const float *)B, ldb,
+            (float *)C, ldc,
+            ith, nth};
+        tb.matmul(m, n);
+        return true;
+#elif defined(__ARM_FEATURE_FP16_VECTOR_ARITHMETIC) && !defined(_MSC_VER)
+        if (n < 8)
+            return false;
+        if (k % 8)
+            return false;
+        if (Btype != GGML_TYPE_F16)
+            return false;
+        tinyBLAS<8, float16x8_t, float16x8_t, ggml_fp16_t, ggml_fp16_t, float> tb{
+            k, (const ggml_fp16_t *)A, lda,
+            (const ggml_fp16_t *)B, ldb,
+            (float *)C, ldc,
+            ith, nth};
+        tb.matmul(m, n);
+        return true;
+#elif defined(__ARM_NEON) && !defined(_MSC_VER)
+        if (k % 4)
+            return false;
+        if (Btype != GGML_TYPE_F32)
+            return false;
+        tinyBLAS<4, float32x4_t, float32x4_t, ggml_fp16_t, float, float> tb{
+            k, (const ggml_fp16_t *)A, lda,
+            (const float *)B, ldb,
+            (float *)C, ldc,
+            ith, nth};
+        tb.matmul(m, n);
+        return true;
+#else
+        return false;
+#endif
+    }
+
+    case GGML_TYPE_Q8_0: {
+        if (Btype != GGML_TYPE_Q8_0)
+           return false;
+#if defined(__AVX2__) || defined(__AVX512F__) || defined(__AVX__)
+        tinyBLAS_Q0_AVX<block_q8_0, block_q8_0, float> tb{
+            k, (const block_q8_0 *)A, lda,
+            (const block_q8_0 *)B, ldb,
+            (float *)C, ldc,
+            ith, nth};
+        tb.matmul(m, n);
+        return true;
+#elif defined(__ARM_FEATURE_DOTPROD)
+        tinyBLAS_Q0_ARM<block_q8_0> tb{
+            k, (const block_q8_0 *)A, lda,
+            (const block_q8_0 *)B, ldb,
+            (float *)C, ldc,
+            ith, nth};
+        tb.matmul(m, n);
+        return true;
+#else
+        return false;
+#endif
+    }
+
+    case GGML_TYPE_Q4_0: {
+        if (Btype != GGML_TYPE_Q8_0)
+            return false;
+#if defined(__AVX2__) || defined(__AVX512F__) || defined(__AVX__)
+        tinyBLAS_Q0_AVX<block_q4_0, block_q8_0, float> tb{
+            k, (const block_q4_0 *)A, lda,
+            (const block_q8_0 *)B, ldb,
+            (float *)C, ldc,
+            ith, nth};
+        tb.matmul(m, n);
+        return true;
+#elif defined(__ARM_FEATURE_DOTPROD)
+        tinyBLAS_Q0_ARM<block_q4_0> tb{
+            k, (const block_q4_0 *)A, lda,
+            (const block_q8_0 *)B, ldb,
+            (float *)C, ldc,
+            ith, nth};
+        tb.matmul(m, n);
+        return true;
+#else
+        return false;
+#endif
+    }
+
+    default:
+        return false;
+    }
+
+    (void)m;
+    (void)n;
+    (void)k;
+    (void)A;
+    (void)lda;
+    (void)B;
+    (void)ldb;
+    (void)C;
+    (void)ldc;
+    (void)ith;
+    (void)nth;
+    (void)Atype;
+    (void)Btype;
+    (void)Ctype;
+}
diff --git a/src/whisper_cpp/ggml/src/sgemm.h b/src/whisper_cpp/ggml/src/sgemm.h
new file mode 100644
index 00000000..caf6dd55
--- /dev/null
+++ b/src/whisper_cpp/ggml/src/sgemm.h
@@ -0,0 +1,14 @@
+#pragma once
+#include <stdint.h>
+#include <stdbool.h>
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+bool llamafile_sgemm(int64_t, int64_t, int64_t, const void *, int64_t,
+                     const void *, int64_t, void *, int64_t, int, int,
+                     int, int, int);
+
+#ifdef __cplusplus
+}
+#endif
diff --git a/src/whisper_cpp/ggml/src/vulkan-shaders/CMakeLists.txt b/src/whisper_cpp/ggml/src/vulkan-shaders/CMakeLists.txt
new file mode 100644
index 00000000..10075db3
--- /dev/null
+++ b/src/whisper_cpp/ggml/src/vulkan-shaders/CMakeLists.txt
@@ -0,0 +1,7 @@
+find_package (Threads REQUIRED)
+
+set(TARGET vulkan-shaders-gen)
+add_executable(${TARGET} vulkan-shaders-gen.cpp)
+install(TARGETS ${TARGET} RUNTIME)
+target_compile_features(${TARGET} PRIVATE cxx_std_11)
+target_link_libraries(vulkan-shaders-gen PUBLIC Threads::Threads)
diff --git a/src/whisper_cpp/ggml/src/vulkan-shaders/acc.comp b/src/whisper_cpp/ggml/src/vulkan-shaders/acc.comp
new file mode 100644
index 00000000..4c8739ef
--- /dev/null
+++ b/src/whisper_cpp/ggml/src/vulkan-shaders/acc.comp
@@ -0,0 +1,24 @@
+#version 450
+
+#include "types.comp"
+#include "generic_binary_head.comp"
+
+void main() {
+    const uint idx = gl_GlobalInvocationID.x;
+    if (idx >= p.ne) {
+        return;
+    }
+
+    const uint offset = p.param3;
+    const uint src1_i = idx - offset;
+    const uint oz = src1_i / p.nb02;
+    const uint oy = (src1_i - (oz * p.nb02)) / p.nb01;
+    const uint ox = src1_i % p.nb01;
+
+    if (ox < p.ne10 && oy < p.ne11 && oz < p.ne12) {
+        data_d[p.d_offset + dst_idx(idx)] = D_TYPE(FLOAT_TYPE(data_a[src0_idx(idx)]) + FLOAT_TYPE(data_b[ox + oy * p.ne10 + oz * p.ne10 * p.ne11]));
+    } else {
+        data_d[p.d_offset + dst_idx(idx)] = D_TYPE(FLOAT_TYPE(data_a[src0_idx(idx)]));
+    }
+}
+
diff --git a/src/whisper_cpp/ggml/src/vulkan-shaders/add.comp b/src/whisper_cpp/ggml/src/vulkan-shaders/add.comp
new file mode 100644
index 00000000..3974845d
--- /dev/null
+++ b/src/whisper_cpp/ggml/src/vulkan-shaders/add.comp
@@ -0,0 +1,14 @@
+#version 450
+
+#include "types.comp"
+#include "generic_binary_head.comp"
+
+void main() {
+    const uint idx = get_idx();
+
+    if (idx >= p.ne) {
+        return;
+    }
+
+    data_d[p.d_offset + dst_idx(idx)] = D_TYPE(FLOAT_TYPE(data_a[src0_idx(idx)]) + FLOAT_TYPE(data_b[src1_idx(idx)]));
+}
diff --git a/src/whisper_cpp/ggml/src/vulkan-shaders/argsort.comp b/src/whisper_cpp/ggml/src/vulkan-shaders/argsort.comp
new file mode 100644
index 00000000..d4fa45b1
--- /dev/null
+++ b/src/whisper_cpp/ggml/src/vulkan-shaders/argsort.comp
@@ -0,0 +1,69 @@
+#version 450
+
+#include "types.comp"
+
+#define BLOCK_SIZE 1024
+#define ASC 0
+
+layout(local_size_x = BLOCK_SIZE, local_size_y = 1, local_size_z = 1) in;
+
+layout (binding = 0) readonly buffer A {A_TYPE data_a[];};
+layout (binding = 1)          buffer D {int data_d[];};
+
+layout (push_constant) uniform parameter {
+    uint ncols;
+    uint ncols_pad;
+    uint order;
+} p;
+
+shared int dst_row[BLOCK_SIZE];
+
+void swap(uint idx0, uint idx1) {
+    int tmp = dst_row[idx0];
+    dst_row[idx0] = dst_row[idx1];
+    dst_row[idx1] = tmp;
+}
+
+void main() {
+    // bitonic sort
+    const int col = int(gl_LocalInvocationID.x);
+    const uint row = gl_WorkGroupID.y;
+
+    const uint row_offset = row * p.ncols;
+
+    // initialize indices
+    if (col < p.ncols_pad) {
+        dst_row[col] = col;
+    }
+    barrier();
+
+    for (uint k = 2; k <= p.ncols_pad; k *= 2) {
+        for (uint j = k / 2; j > 0; j /= 2) {
+            const uint ixj = col ^ j;
+            if (col < p.ncols_pad && ixj > col) {
+                if ((col & k) == 0) {
+                    if (dst_row[col] >= p.ncols ||
+                        (dst_row[ixj] < p.ncols && (p.order == ASC ?
+                            data_a[row_offset + dst_row[col]] > data_a[row_offset + dst_row[ixj]] :
+                            data_a[row_offset + dst_row[col]] < data_a[row_offset + dst_row[ixj]]))
+                    ) {
+                        swap(col, ixj);
+                    }
+                } else {
+                    if (dst_row[ixj] >= p.ncols ||
+                        (dst_row[col] < p.ncols && (p.order == ASC ?
+                            data_a[row_offset + dst_row[col]] < data_a[row_offset + dst_row[ixj]] :
+                            data_a[row_offset + dst_row[col]] > data_a[row_offset + dst_row[ixj]]))
+                    ) {
+                        swap(col, ixj);
+                    }
+                }
+            }
+            barrier();
+        }
+    }
+
+    if (col < p.ncols) {
+        data_d[row_offset + col] = dst_row[col];
+    }
+}
diff --git a/src/whisper_cpp/ggml/src/vulkan-shaders/clamp.comp b/src/whisper_cpp/ggml/src/vulkan-shaders/clamp.comp
new file mode 100644
index 00000000..7071302a
--- /dev/null
+++ b/src/whisper_cpp/ggml/src/vulkan-shaders/clamp.comp
@@ -0,0 +1,15 @@
+#version 450
+
+#include "types.comp"
+#include "generic_unary_head.comp"
+
+void main() {
+    const uint idx = get_idx();
+
+    if (idx >= p.ne) {
+        return;
+    }
+
+    const FLOAT_TYPE val = FLOAT_TYPE(data_a[src0_idx(idx)]);
+    data_d[p.d_offset + dst_idx(idx)] = D_TYPE(val < p.param1 ? p.param1 : (val > p.param2 ? p.param2 : val));
+}
diff --git a/src/whisper_cpp/ggml/src/vulkan-shaders/concat.comp b/src/whisper_cpp/ggml/src/vulkan-shaders/concat.comp
new file mode 100644
index 00000000..c23b6eb1
--- /dev/null
+++ b/src/whisper_cpp/ggml/src/vulkan-shaders/concat.comp
@@ -0,0 +1,39 @@
+#version 450
+
+#include "types.comp"
+#include "generic_binary_head.comp"
+
+void main() {
+    const uint idx = gl_GlobalInvocationID.z * 262144 + gl_GlobalInvocationID.y * 512 + gl_GlobalInvocationID.x;
+    const int dim = p.param3;
+
+    if (idx >= p.ne) {
+        return;
+    }
+
+    const uint i3 = idx / (p.ne22*p.ne21*p.ne20);
+    const uint i3_offset = i3 * p.ne22*p.ne21*p.ne20;
+    const uint i2 = (idx - i3_offset) / (p.ne21*p.ne20);
+    const uint i2_offset = i2*p.ne21*p.ne20;
+    const uint i1 = (idx - i3_offset - i2_offset) / p.ne20;
+    const uint i0 = idx - i3_offset - i2_offset - i1*p.ne20;
+
+    uint o[4] = {0, 0, 0, 0};
+    o[dim] = dim == 0 ? p.ne00 : (dim == 1 ? p.ne01 : (dim == 2 ? p.ne02 : p.ne03));
+
+    const uint src0_idx = i3*p.nb03 + i2*p.nb02 + i1*p.nb01 + i0*p.nb00;
+    const uint src1_idx = (i3 - o[3])*p.nb13 + (i2 - o[2])*p.nb12 + (i1 - o[1])*p.nb11 + (i0 - o[0])*p.nb10;
+    const uint dst_idx = i3*p.nb23 + i2*p.nb22 + i1*p.nb21 + i0*p.nb20;
+
+    const bool is_src0 = i0 < p.ne00 && i1 < p.ne01 && i2 < p.ne02 && i3 < p.ne03;
+
+#ifndef OPTIMIZATION_ERROR_WORKAROUND
+    data_d[p.d_offset + dst_idx] = D_TYPE(is_src0 ? data_a[src0_idx] : data_b[src1_idx]);
+#else
+    if (is_src0) {
+        data_d[p.d_offset + dst_idx] = data_a[src0_idx];
+    } else {
+        data_d[p.d_offset + dst_idx] = data_b[src1_idx];
+    }
+#endif
+}
diff --git a/src/whisper_cpp/ggml/src/vulkan-shaders/copy.comp b/src/whisper_cpp/ggml/src/vulkan-shaders/copy.comp
new file mode 100644
index 00000000..c26917c0
--- /dev/null
+++ b/src/whisper_cpp/ggml/src/vulkan-shaders/copy.comp
@@ -0,0 +1,18 @@
+#version 450
+
+#include "types.comp"
+#include "generic_unary_head.comp"
+
+void main() {
+    const uint idx = get_idx();
+
+    if (idx >= p.ne) {
+        return;
+    }
+
+#ifndef OPTIMIZATION_ERROR_WORKAROUND
+    data_d[p.d_offset + dst_idx(idx)] = D_TYPE(data_a[src0_idx(idx)]);
+#else
+    data_d[p.d_offset + dst_idx(idx)] = data_a[src0_idx(idx)];
+#endif
+}
diff --git a/src/whisper_cpp/ggml/src/vulkan-shaders/cos.comp b/src/whisper_cpp/ggml/src/vulkan-shaders/cos.comp
new file mode 100644
index 00000000..f9a858cb
--- /dev/null
+++ b/src/whisper_cpp/ggml/src/vulkan-shaders/cos.comp
@@ -0,0 +1,15 @@
+#version 450
+
+#include "types.comp"
+#include "generic_unary_head.comp"
+
+void main() {
+    const uint idx = get_idx();
+
+    if (idx >= p.ne) {
+        return;
+    }
+
+    const FLOAT_TYPE val = FLOAT_TYPE(data_a[src0_idx(idx)]);
+    data_d[p.d_offset + dst_idx(idx)] = D_TYPE(cos(val));
+}
diff --git a/src/whisper_cpp/ggml/src/vulkan-shaders/dequant_f32.comp b/src/whisper_cpp/ggml/src/vulkan-shaders/dequant_f32.comp
new file mode 100644
index 00000000..a4d3fca5
--- /dev/null
+++ b/src/whisper_cpp/ggml/src/vulkan-shaders/dequant_f32.comp
@@ -0,0 +1,20 @@
+#version 450
+
+#include "dequant_head.comp"
+
+layout(local_size_x = 256, local_size_y = 1, local_size_z = 1) in;
+
+layout (binding = 0) readonly buffer A {float data_a[];};
+layout (binding = 1) writeonly buffer D {D_TYPE data_b[];};
+
+void main() {
+    const uint i = gl_GlobalInvocationID.x * 16;
+
+    if (i >= p.nel) {
+        return;
+    }
+
+    [[unroll]] for (uint l = 0; l < 16; l++) {
+        data_b[i + l] = D_TYPE(data_a[i + l]);
+    }
+}
diff --git a/src/whisper_cpp/ggml/src/vulkan-shaders/dequant_funcs.comp b/src/whisper_cpp/ggml/src/vulkan-shaders/dequant_funcs.comp
new file mode 100644
index 00000000..d5b98973
--- /dev/null
+++ b/src/whisper_cpp/ggml/src/vulkan-shaders/dequant_funcs.comp
@@ -0,0 +1,68 @@
+#if !defined(DATA_A_F32) && !defined(DATA_A_F16)
+#extension GL_EXT_shader_explicit_arithmetic_types_int8 : require
+#endif
+
+#if defined(DATA_A_F32)
+vec2 dequantize(uint ib, uint iqs, uint a_offset) {
+    return vec2(data_a[a_offset + ib], data_a[a_offset + ib + 1]);
+}
+#endif
+
+#if defined(DATA_A_F16)
+vec2 dequantize(uint ib, uint iqs, uint a_offset) {
+    return vec2(data_a[a_offset + ib], data_a[a_offset + ib + 1]);
+}
+#endif
+
+#if defined(DATA_A_Q4_0)
+vec2 dequantize(uint ib, uint iqs, uint a_offset) {
+    const float d = float(data_a[a_offset + ib].d);
+    const uint vui = uint(data_a[a_offset + ib].qs[iqs]);
+    return (vec2(vui & 0xF, vui >> 4) - 8.0f) * d;
+}
+#endif
+
+#if defined(DATA_A_Q4_1)
+vec2 dequantize(uint ib, uint iqs, uint a_offset) {
+    const float d = float(data_a[a_offset + ib].d);
+    const float m = float(data_a[a_offset + ib].m);
+    const uint vui = uint(data_a[a_offset + ib].qs[iqs]);
+    return vec2(vui & 0xF, vui >> 4) * d + m;
+}
+#endif
+
+#if defined(DATA_A_Q5_0)
+vec2 dequantize(uint ib, uint iqs, uint a_offset) {
+    const float d = float(data_a[a_offset + ib].d);
+    const uint uint_qh = uint(data_a[a_offset + ib].qh[1]) << 16 | data_a[a_offset + ib].qh[0];
+    const ivec2 qh = ivec2(((uint_qh >> iqs) << 4) & 0x10, (uint_qh >> (iqs + 12)) & 0x10);
+    const uint vui = uint(data_a[a_offset + ib].qs[iqs]);
+    return (vec2((vui & 0xF) | qh.x, (vui >> 4) | qh.y) - 16.0f) * d;
+}
+#endif
+
+#if defined(DATA_A_Q5_1)
+vec2 dequantize(uint ib, uint iqs, uint a_offset) {
+    const float d = float(data_a[a_offset + ib].d);
+    const float m = float(data_a[a_offset + ib].m);
+    const uint uint_qh = data_a[a_offset + ib].qh;
+    const ivec2 qh = ivec2(((uint_qh >> iqs) << 4) & 0x10, (uint_qh >> (iqs + 12)) & 0x10);
+    const uint vui = uint(data_a[a_offset + ib].qs[iqs]);
+    return vec2((vui & 0xF) | qh.x, (vui >> 4) | qh.y) * d + m;
+}
+#endif
+
+#if defined(DATA_A_Q8_0)
+vec2 dequantize(uint ib, uint iqs, uint a_offset) {
+    const float d = float(data_a[a_offset + ib].d);
+    return vec2(int(data_a[a_offset + ib].qs[iqs]), int(data_a[a_offset + ib].qs[iqs + 1])) * d;
+}
+#endif
+
+#if defined(DATA_A_IQ4_NL)
+vec2 dequantize(uint ib, uint iqs, uint a_offset) {
+    const float d = float(data_a[a_offset + ib].d);
+    const uint vui = uint(data_a[a_offset + ib].qs[iqs]);
+    return vec2(kvalues_iq4nl[vui & 0xF], kvalues_iq4nl[vui >> 4]) * d;
+}
+#endif
diff --git a/src/whisper_cpp/ggml/src/vulkan-shaders/dequant_head.comp b/src/whisper_cpp/ggml/src/vulkan-shaders/dequant_head.comp
new file mode 100644
index 00000000..8d806435
--- /dev/null
+++ b/src/whisper_cpp/ggml/src/vulkan-shaders/dequant_head.comp
@@ -0,0 +1,13 @@
+#extension GL_EXT_control_flow_attributes : require
+#extension GL_EXT_shader_16bit_storage : require
+
+layout (push_constant) uniform parameter
+{
+    uint M;
+    uint K;
+    uint stride_a;
+    uint stride_b;
+    uint nel;
+} p;
+
+#include "types.comp"
diff --git a/src/whisper_cpp/ggml/src/vulkan-shaders/dequant_iq4_nl.comp b/src/whisper_cpp/ggml/src/vulkan-shaders/dequant_iq4_nl.comp
new file mode 100644
index 00000000..34ef3da3
--- /dev/null
+++ b/src/whisper_cpp/ggml/src/vulkan-shaders/dequant_iq4_nl.comp
@@ -0,0 +1,30 @@
+#version 450
+
+#include "dequant_head.comp"
+
+layout(local_size_x = 256, local_size_y = 1, local_size_z = 1) in;
+
+layout (binding = 0) readonly buffer A {block_iq4_nl data_a[];};
+layout (binding = 1) writeonly buffer D {D_TYPE data_b[];};
+
+void main() {
+    const uint i = gl_WorkGroupID.x * 4 + gl_LocalInvocationID.x / 64;
+
+    const uint tid = gl_LocalInvocationID.x % 64;
+    const uint il  = tid/32;
+    const uint ir  = tid%32;
+    const uint ib = 32*i + ir;
+    if (ib >= p.nel / 32) {
+        return;
+    }
+
+    const uint q_idx = 8*il;
+    const uint b_idx = 1024*i + 32*ir + q_idx;
+
+    const float d = float(data_a[ib].d);
+
+    [[unroll]] for (uint l = 0; l < 8; ++l) {
+        data_b[b_idx + l +  0] = D_TYPE(d * kvalues_iq4nl[data_a[ib].qs[q_idx + l] & 0xF]);
+        data_b[b_idx + l + 16] = D_TYPE(d * kvalues_iq4nl[data_a[ib].qs[q_idx + l] >>  4]);
+    }
+}
diff --git a/src/whisper_cpp/ggml/src/vulkan-shaders/dequant_q2_k.comp b/src/whisper_cpp/ggml/src/vulkan-shaders/dequant_q2_k.comp
new file mode 100644
index 00000000..157154af
--- /dev/null
+++ b/src/whisper_cpp/ggml/src/vulkan-shaders/dequant_q2_k.comp
@@ -0,0 +1,34 @@
+#version 450
+
+#include "dequant_head.comp"
+
+layout(local_size_x = 64, local_size_y = 1, local_size_z = 1) in;
+
+layout (binding = 0) readonly buffer A {A_TYPE data_a[];};
+layout (binding = 1) writeonly buffer D {D_TYPE data_b[];};
+
+void main() {
+    [[unroll]] for (uint wgy = 0; wgy < 256; wgy++) {
+        const uint i = gl_WorkGroupID.x * 256 + wgy;
+        if (i >= p.M * p.K / QUANT_K) {
+            return;
+        }
+
+        const uint tid = gl_LocalInvocationID.x;
+        const uint ip = tid / 32;
+        const uint il = tid - 32 * ip;
+        const uint is = 8 * ip + il / 16;
+
+        const uint y_idx = i * QUANT_K + 128 * ip + il;
+
+        const uint ql_idx = 32 * ip + il;
+        const uint8_t qs = data_a[i].qs[32 * ip + il];
+
+        FLOAT_TYPE dall = FLOAT_TYPE(data_a[i].d.x);
+        FLOAT_TYPE dmin = FLOAT_TYPE(data_a[i].d.y);
+        data_b[y_idx +  0] = D_TYPE(dall * FLOAT_TYPE((data_a[i].scales[is+0] & 0xF) * ((qs >> 0) & 3)) - dmin * FLOAT_TYPE(data_a[i].scales[is+0] >> 4));
+        data_b[y_idx + 32] = D_TYPE(dall * FLOAT_TYPE((data_a[i].scales[is+2] & 0xF) * ((qs >> 2) & 3)) - dmin * FLOAT_TYPE(data_a[i].scales[is+2] >> 4));
+        data_b[y_idx + 64] = D_TYPE(dall * FLOAT_TYPE((data_a[i].scales[is+4] & 0xF) * ((qs >> 4) & 3)) - dmin * FLOAT_TYPE(data_a[i].scales[is+4] >> 4));
+        data_b[y_idx + 96] = D_TYPE(dall * FLOAT_TYPE((data_a[i].scales[is+6] & 0xF) * ((qs >> 6) & 3)) - dmin * FLOAT_TYPE(data_a[i].scales[is+6] >> 4));
+    }
+}
diff --git a/src/whisper_cpp/ggml/src/vulkan-shaders/dequant_q3_k.comp b/src/whisper_cpp/ggml/src/vulkan-shaders/dequant_q3_k.comp
new file mode 100644
index 00000000..c17dd0d9
--- /dev/null
+++ b/src/whisper_cpp/ggml/src/vulkan-shaders/dequant_q3_k.comp
@@ -0,0 +1,42 @@
+#version 450
+
+#include "dequant_head.comp"
+
+layout(local_size_x = 64, local_size_y = 1, local_size_z = 1) in;
+
+layout (binding = 0) readonly buffer A {A_TYPE data_a[];};
+layout (binding = 1) writeonly buffer D {D_TYPE data_b[];};
+
+void main() {
+    [[unroll]] for (uint wgy = 0; wgy < 256; wgy++) {
+        const uint i = uint(gl_WorkGroupID.x * 256 + wgy);
+        if (i >= p.M * p.K / QUANT_K) {
+            return;
+        }
+
+        const uint r = gl_LocalInvocationID.x / 4;
+        const uint tid = r / 2;
+        const uint is0 = r % 2;
+        const uint l0 = 16 * is0 + 4 * (gl_LocalInvocationID.x % 4);
+        const uint n = tid / 4;
+        const uint j = tid - 4*n;
+
+        const uint8_t m = uint8_t(1 << (4*n + j));
+        const uint is = 8*n + 2*j + is0;
+        const uint shift = 2*j;
+
+        const int8_t us = int8_t(is <  4 ? (data_a[i].scales[is-0] & 0xF) | (((data_a[i].scales[is+8] >> 0) & 3) << 4) :
+                                 is <  8 ? (data_a[i].scales[is-0] & 0xF) | (((data_a[i].scales[is+4] >> 2) & 3) << 4) :
+                                 is < 12 ? (data_a[i].scales[is-8] >>  4) | (((data_a[i].scales[is+0] >> 4) & 3) << 4) :
+                                           (data_a[i].scales[is-8] >>  4) | (((data_a[i].scales[is-4] >> 6) & 3) << 4));
+        const FLOAT_TYPE d_all = FLOAT_TYPE(data_a[i].d);
+        const FLOAT_TYPE dl    = d_all * FLOAT_TYPE(us - 32);
+
+        const uint y_idx = i * QUANT_K + 128 * n + 32 * j;
+        const uint qs_idx = 32*n;
+
+        for (uint l = l0; l < l0 + 4; ++l) {
+            data_b[y_idx + l] = D_TYPE(dl * FLOAT_TYPE(int8_t((data_a[i].qs[qs_idx + l] >> shift) & 3) - (((data_a[i].hmask[l] & m) != 0) ? 0 : 4)));
+        }
+    }
+}
diff --git a/src/whisper_cpp/ggml/src/vulkan-shaders/dequant_q4_0.comp b/src/whisper_cpp/ggml/src/vulkan-shaders/dequant_q4_0.comp
new file mode 100644
index 00000000..40818532
--- /dev/null
+++ b/src/whisper_cpp/ggml/src/vulkan-shaders/dequant_q4_0.comp
@@ -0,0 +1,30 @@
+#version 450
+
+#include "dequant_head.comp"
+
+layout(local_size_x = 256, local_size_y = 1, local_size_z = 1) in;
+
+layout (binding = 0) readonly buffer A {block_q4_0 data_a[];};
+layout (binding = 1) writeonly buffer D {D_TYPE data_b[];};
+
+void main() {
+    const uint i = gl_WorkGroupID.x * 4 + gl_LocalInvocationID.x / 64;
+
+    const uint tid = gl_LocalInvocationID.x % 64;
+    const uint il  = tid/32;
+    const uint ir  = tid%32;
+    const uint ib = 32*i + ir;
+    if (ib >= p.nel / 32) {
+        return;
+    }
+
+    const uint q_idx = 8*il;
+    const uint b_idx = 1024*i + 32*ir + q_idx;
+
+    const float d = float(data_a[ib].d);
+
+    [[unroll]] for (uint l = 0; l < 8; ++l) {
+        data_b[b_idx + l +  0] = D_TYPE(d * ((data_a[ib].qs[q_idx + l] & 0xF) - 8.0f));
+        data_b[b_idx + l + 16] = D_TYPE(d * ((data_a[ib].qs[q_idx + l] >>  4) - 8.0f));
+    }
+}
diff --git a/src/whisper_cpp/ggml/src/vulkan-shaders/dequant_q4_1.comp b/src/whisper_cpp/ggml/src/vulkan-shaders/dequant_q4_1.comp
new file mode 100644
index 00000000..2f27eee6
--- /dev/null
+++ b/src/whisper_cpp/ggml/src/vulkan-shaders/dequant_q4_1.comp
@@ -0,0 +1,32 @@
+#version 450
+
+#include "dequant_head.comp"
+
+layout(local_size_x = 256, local_size_y = 1, local_size_z = 1) in;
+
+layout (binding = 0) readonly buffer A {block_q4_1 data_a[];};
+layout (binding = 1) writeonly buffer D {D_TYPE data_b[];};
+
+void main() {
+    const uint i = gl_WorkGroupID.x * 4 + gl_LocalInvocationID.x / 64;
+
+    const uint tid = gl_LocalInvocationID.x % 64;
+    const uint il  = tid/32;
+    const uint ir  = tid%32;
+    const uint ib = 32*i + ir;
+    if (ib >= p.nel / 32) {
+        return;
+    }
+
+    const uint b_idx = 1024*i + 32*ir + 8*il;
+
+    const float d = float(data_a[ib].d);
+    const float m = float(data_a[ib].m);
+
+    const uint q_idx = 8*il;
+
+    [[unroll]] for (uint l = 0; l < 8; ++l) {
+        data_b[b_idx + l +  0] = D_TYPE(d * (data_a[ib].qs[q_idx + l] & 0xF) + m);
+        data_b[b_idx + l + 16] = D_TYPE(d * (data_a[ib].qs[q_idx + l] >>  4) + m);
+    }
+}
diff --git a/src/whisper_cpp/ggml/src/vulkan-shaders/dequant_q4_k.comp b/src/whisper_cpp/ggml/src/vulkan-shaders/dequant_q4_k.comp
new file mode 100644
index 00000000..92acb754
--- /dev/null
+++ b/src/whisper_cpp/ggml/src/vulkan-shaders/dequant_q4_k.comp
@@ -0,0 +1,56 @@
+#version 450
+
+#include "dequant_head.comp"
+
+layout(local_size_x = 32, local_size_y = 1, local_size_z = 1) in;
+
+layout (binding = 0) readonly buffer A {A_TYPE data_a[];};
+layout (binding = 1) writeonly buffer D {D_TYPE data_b[];};
+
+void main() {
+    [[unroll]] for (uint wgy = 0; wgy < 256; wgy++) {
+        const uint i = gl_WorkGroupID.x * 256 + wgy;
+        if (i >= p.M * p.K / QUANT_K) {
+            return;
+        }
+
+        const uint tid = gl_LocalInvocationID.x;
+        const uint il = tid / 8;
+        const uint ir = tid % 8;
+        const uint is = 2 * il;
+        const uint n = 4;
+
+        const FLOAT_TYPE dall = FLOAT_TYPE(data_a[i].d.x);
+        const FLOAT_TYPE dmin = FLOAT_TYPE(data_a[i].d.y);
+
+        const uint y_idx = i * QUANT_K + 64 * il + n * ir;
+        const uint qs_idx = 32*il + n * ir;
+
+        uint8_t sc;
+        uint8_t m;
+        if (is < 4) {
+            sc = uint8_t(data_a[i].scales[is] & 63);
+            m  = uint8_t(data_a[i].scales[is + 4] & 63);
+        } else {
+            sc = uint8_t((data_a[i].scales[is + 4] & 0xF) | ((data_a[i].scales[is - 4] >> 6) << 4));
+            m  = uint8_t((data_a[i].scales[is + 4] >>  4) | ((data_a[i].scales[is    ] >> 6) << 4));
+        }
+        const FLOAT_TYPE d1 = dall * sc;
+        const FLOAT_TYPE m1 = dmin * m;
+
+        if (is < 4) {
+            sc = uint8_t(data_a[i].scales[is + 1] & 63);
+            m  = uint8_t(data_a[i].scales[is + 5] & 63);
+        } else {
+            sc = uint8_t((data_a[i].scales[is + 5] & 0xF) | ((data_a[i].scales[is - 3] >> 6) << 4));
+            m  = uint8_t((data_a[i].scales[is + 5] >>  4) | ((data_a[i].scales[is + 1] >> 6) << 4));
+        }
+        const FLOAT_TYPE d2 = dall * sc;
+        const FLOAT_TYPE m2 = dmin * m;
+
+        [[unroll]] for (uint l = 0; l < n; ++l) {
+            data_b[y_idx + l     ] = D_TYPE(d1 * FLOAT_TYPE(data_a[i].qs[qs_idx + l] & 0xF) - m1);
+            data_b[y_idx + l + 32] = D_TYPE(d2 * FLOAT_TYPE(data_a[i].qs[qs_idx + l] >>  4) - m2);
+        }
+    }
+}
diff --git a/src/whisper_cpp/ggml/src/vulkan-shaders/dequant_q5_0.comp b/src/whisper_cpp/ggml/src/vulkan-shaders/dequant_q5_0.comp
new file mode 100644
index 00000000..b20b8052
--- /dev/null
+++ b/src/whisper_cpp/ggml/src/vulkan-shaders/dequant_q5_0.comp
@@ -0,0 +1,34 @@
+#version 450
+
+#include "dequant_head.comp"
+
+layout(local_size_x = 256, local_size_y = 1, local_size_z = 1) in;
+
+layout (binding = 0) readonly buffer A {block_q5_0 data_a[];};
+layout (binding = 1) writeonly buffer D {D_TYPE data_b[];};
+
+void main() {
+    const uint i = gl_WorkGroupID.x * 4 + gl_LocalInvocationID.x / 64;
+
+    const uint tid = gl_LocalInvocationID.x % 64;
+    const uint il  = tid/32;
+    const uint ir  = tid%32;
+    const uint ib = 32*i + ir;
+    if (ib >= p.nel / 32) {
+        return;
+    }
+
+    const uint b_idx = 1024*i + 32*ir + 8*il;
+
+    const float d = float(data_a[ib].d);
+    const uint qh = uint(data_a[ib].qh[1]) << 16 | data_a[ib].qh[0];
+
+    const uint q_idx = 8*il;
+
+    [[unroll]] for (uint l = 0; l < 8; ++l) {
+        const uint iqs = q_idx + l;
+        const uint vui = uint(data_a[ib].qs[iqs]);
+        data_b[b_idx + l +  0] = D_TYPE(d * (((vui & 0xF) | (((qh >> iqs) << 4) & 0x10)) - 16.0f));
+        data_b[b_idx + l + 16] = D_TYPE(d * (((vui >>  4) | ((qh >> (iqs + 12)) & 0x10)) - 16.0f));
+    }
+}
diff --git a/src/whisper_cpp/ggml/src/vulkan-shaders/dequant_q5_1.comp b/src/whisper_cpp/ggml/src/vulkan-shaders/dequant_q5_1.comp
new file mode 100644
index 00000000..dc59fe3b
--- /dev/null
+++ b/src/whisper_cpp/ggml/src/vulkan-shaders/dequant_q5_1.comp
@@ -0,0 +1,35 @@
+#version 450
+
+#include "dequant_head.comp"
+
+layout(local_size_x = 256, local_size_y = 1, local_size_z = 1) in;
+
+layout (binding = 0) readonly buffer A {block_q5_1 data_a[];};
+layout (binding = 1) writeonly buffer D {D_TYPE data_b[];};
+
+void main() {
+    const uint i = gl_WorkGroupID.x * 4 + gl_LocalInvocationID.x / 64;
+
+    const uint tid = gl_LocalInvocationID.x % 64;
+    const uint il  = tid/32;
+    const uint ir  = tid%32;
+    const uint ib = 32*i + ir;
+    if (ib >= p.nel / 32) {
+        return;
+    }
+
+    const uint b_idx = 1024*i + 32*ir + 8*il;
+
+    const float d = float(data_a[ib].d);
+    const float m = float(data_a[ib].m);
+    const uint qh = data_a[ib].qh;
+
+    const uint q_idx = 8*il;
+
+    [[unroll]] for (uint l = 0; l < 8; ++l) {
+        const uint iqs = q_idx + l;
+        const uint vui = uint(data_a[ib].qs[iqs]);
+        data_b[b_idx + l +  0] = D_TYPE(d * (((vui & 0xF) | (((qh >> iqs) << 4) & 0x10))) + m);
+        data_b[b_idx + l + 16] = D_TYPE(d * (((vui >>  4) | ((qh >> (iqs + 12)) & 0x10))) + m);
+    }
+}
diff --git a/src/whisper_cpp/ggml/src/vulkan-shaders/dequant_q5_k.comp b/src/whisper_cpp/ggml/src/vulkan-shaders/dequant_q5_k.comp
new file mode 100644
index 00000000..f314a76d
--- /dev/null
+++ b/src/whisper_cpp/ggml/src/vulkan-shaders/dequant_q5_k.comp
@@ -0,0 +1,58 @@
+#version 450
+
+#include "dequant_head.comp"
+
+layout(local_size_x = 64, local_size_y = 1, local_size_z = 1) in;
+
+layout (binding = 0) readonly buffer A {A_TYPE data_a[];};
+layout (binding = 1) writeonly buffer D {D_TYPE data_b[];};
+
+void main() {
+    [[unroll]] for (uint wgy = 0; wgy < 256; wgy++) {
+        const uint i = gl_WorkGroupID.x * 256 + wgy;
+        if (i >= p.M * p.K / QUANT_K) {
+            return;
+        }
+
+        const uint tid = gl_LocalInvocationID.x;
+        const uint il = tid / 16;
+        const uint ir = tid % 16;
+        const uint is = 2 * il;
+
+        const FLOAT_TYPE dall = FLOAT_TYPE(data_a[i].d.x);
+        const FLOAT_TYPE dmin = FLOAT_TYPE(data_a[i].d.y);
+
+        const uint y_idx = i * QUANT_K + 64 * il + 2 * ir;
+        const uint qs_idx = 32*il + 2 * ir;
+        const uint qh_idx = 2 * ir;
+
+        uint8_t sc;
+        uint8_t m;
+        if (is < 4) {
+            sc = uint8_t(data_a[i].scales[is] & 63);
+            m  = uint8_t(data_a[i].scales[is + 4] & 63);
+        } else {
+            sc = uint8_t((data_a[i].scales[is + 4] & 0xF) | ((data_a[i].scales[is - 4] >> 6) << 4));
+            m  = uint8_t((data_a[i].scales[is + 4] >>  4) | ((data_a[i].scales[is    ] >> 6) << 4));
+        }
+        const FLOAT_TYPE d1 = dall * sc;
+        const FLOAT_TYPE m1 = dmin * m;
+
+        if (is < 4) {
+            sc = uint8_t(data_a[i].scales[is + 1] & 63);
+            m  = uint8_t(data_a[i].scales[is + 5] & 63);
+        } else {
+            sc = uint8_t((data_a[i].scales[is + 5] & 0xF) | ((data_a[i].scales[is - 3] >> 6) << 4));
+            m  = uint8_t((data_a[i].scales[is + 5] >>  4) | ((data_a[i].scales[is + 1] >> 6) << 4));
+        }
+        const FLOAT_TYPE d2 = dall * sc;
+        const FLOAT_TYPE m2 = dmin * m;
+
+        const uint8_t hm1 = uint8_t(1 << (2 * il    ));
+        const uint8_t hm2 = uint8_t(1 << (2 * il + 1));
+        data_b[y_idx     ] = D_TYPE(d1 * FLOAT_TYPE((data_a[i].qs[qs_idx    ] & 0xF) + (((data_a[i].qh[qh_idx    ] & hm1) != 0) ? 16 : 0)) - m1);
+        data_b[y_idx +  1] = D_TYPE(d1 * FLOAT_TYPE((data_a[i].qs[qs_idx + 1] & 0xF) + (((data_a[i].qh[qh_idx + 1] & hm1) != 0) ? 16 : 0)) - m1);
+        data_b[y_idx + 32] = D_TYPE(d2 * FLOAT_TYPE((data_a[i].qs[qs_idx    ]  >> 4) + (((data_a[i].qh[qh_idx    ] & hm2) != 0) ? 16 : 0)) - m2);
+        data_b[y_idx + 33] = D_TYPE(d2 * FLOAT_TYPE((data_a[i].qs[qs_idx + 1]  >> 4) + (((data_a[i].qh[qh_idx + 1] & hm2) != 0) ? 16 : 0)) - m2);
+    }
+}
diff --git a/src/whisper_cpp/ggml/src/vulkan-shaders/dequant_q6_k.comp b/src/whisper_cpp/ggml/src/vulkan-shaders/dequant_q6_k.comp
new file mode 100644
index 00000000..0b913175
--- /dev/null
+++ b/src/whisper_cpp/ggml/src/vulkan-shaders/dequant_q6_k.comp
@@ -0,0 +1,33 @@
+#version 450
+
+#include "dequant_head.comp"
+
+layout(local_size_x = 64, local_size_y = 1, local_size_z = 1) in;
+
+layout (binding = 0) readonly buffer A {A_TYPE data_a[];};
+layout (binding = 1) writeonly buffer D {D_TYPE data_b[];};
+
+void main() {
+    [[unroll]] for (uint wgy = 0; wgy < 256; wgy++) {
+        const uint i = gl_WorkGroupID.x * 256 + wgy;
+        if (i >= p.M * p.K / QUANT_K) {
+            return;
+        }
+        const uint tid = gl_LocalInvocationID.x;
+        const uint ip = tid / 32;
+        const uint il = tid - 32 * ip;
+        const uint is = 8 * ip + il / 16;
+
+        const uint y_idx = i * QUANT_K + 128 * ip + il;
+
+        const uint ql_idx = 64 * ip + il;
+        const uint8_t qh = data_a[i].qh[32 * ip + il];
+
+        const FLOAT_TYPE d = FLOAT_TYPE(data_a[i].d);
+
+        data_b[y_idx +  0] = D_TYPE(d * FLOAT_TYPE(data_a[i].scales[is + 0] * (int8_t((data_a[i].ql[ql_idx +  0] & 0xF) | (((qh >> 0) & 3) << 4)) - 32)));
+        data_b[y_idx + 32] = D_TYPE(d * FLOAT_TYPE(data_a[i].scales[is + 2] * (int8_t((data_a[i].ql[ql_idx + 32] & 0xF) | (((qh >> 2) & 3) << 4)) - 32)));
+        data_b[y_idx + 64] = D_TYPE(d * FLOAT_TYPE(data_a[i].scales[is + 4] * (int8_t((data_a[i].ql[ql_idx +  0] >>  4) | (((qh >> 4) & 3) << 4)) - 32)));
+        data_b[y_idx + 96] = D_TYPE(d * FLOAT_TYPE(data_a[i].scales[is + 6] * (int8_t((data_a[i].ql[ql_idx + 32] >>  4) | (((qh >> 6) & 3) << 4)) - 32)));
+    }
+}
diff --git a/src/whisper_cpp/ggml/src/vulkan-shaders/dequant_q8_0.comp b/src/whisper_cpp/ggml/src/vulkan-shaders/dequant_q8_0.comp
new file mode 100644
index 00000000..bd1344a8
--- /dev/null
+++ b/src/whisper_cpp/ggml/src/vulkan-shaders/dequant_q8_0.comp
@@ -0,0 +1,31 @@
+#version 450
+
+#include "dequant_head.comp"
+
+layout(local_size_x = 256, local_size_y = 1, local_size_z = 1) in;
+
+layout (binding = 0) readonly buffer A {block_q8_0 data_a[];};
+layout (binding = 1) writeonly buffer D {D_TYPE data_b[];};
+
+void main() {
+    const uint i = gl_WorkGroupID.x * 4 + gl_LocalInvocationID.x / 64;
+
+    const uint tid = gl_LocalInvocationID.x % 64;
+    const uint il  = tid/32;
+    const uint ir  = tid%32;
+    const uint ib = 32*i + ir;
+    if (ib >= p.nel / 32) {
+        return;
+    }
+
+    const uint b_idx = 1024*i + 32*ir + 16*il;
+
+    const float d = float(data_a[ib].d);
+
+    const uint q_idx = 16*il;
+
+    [[unroll]] for (uint l = 0; l < 16; l += 2) {
+        data_b[b_idx + l    ] = D_TYPE(d * data_a[ib].qs[q_idx + l    ]);
+        data_b[b_idx + l + 1] = D_TYPE(d * data_a[ib].qs[q_idx + l + 1]);
+    }
+}
diff --git a/src/whisper_cpp/ggml/src/vulkan-shaders/diag_mask_inf.comp b/src/whisper_cpp/ggml/src/vulkan-shaders/diag_mask_inf.comp
new file mode 100644
index 00000000..4e68742b
--- /dev/null
+++ b/src/whisper_cpp/ggml/src/vulkan-shaders/diag_mask_inf.comp
@@ -0,0 +1,34 @@
+#version 450
+
+#extension GL_EXT_shader_16bit_storage : require
+#extension GL_EXT_control_flow_attributes : enable
+
+layout (push_constant) uniform parameter
+{
+    uint ncols;
+    uint rows_per_channel;
+    uint n_past;
+} p;
+
+#include "types.comp"
+
+layout(local_size_x = 512, local_size_y = 1, local_size_z = 1) in;
+
+layout (binding = 0) readonly buffer X {A_TYPE data_a[];};
+layout (binding = 1) writeonly buffer D {D_TYPE data_d[];};
+
+void main() {
+    const uint col = gl_GlobalInvocationID.y;
+    const uint row = gl_GlobalInvocationID.x;
+
+    if (col >= p.ncols) {
+        return;
+    }
+
+    const uint i = row*p.ncols + col;
+    if (col > p.n_past + row % p.rows_per_channel) {
+        data_d[i] = D_TYPE(uintBitsToFloat(0xFF800000));
+    } else {
+        data_d[i] = D_TYPE(data_a[i]);
+    }
+}
diff --git a/src/whisper_cpp/ggml/src/vulkan-shaders/div.comp b/src/whisper_cpp/ggml/src/vulkan-shaders/div.comp
new file mode 100644
index 00000000..8cfce58b
--- /dev/null
+++ b/src/whisper_cpp/ggml/src/vulkan-shaders/div.comp
@@ -0,0 +1,14 @@
+#version 450
+
+#include "types.comp"
+#include "generic_binary_head.comp"
+
+void main() {
+    const uint idx = get_idx();
+
+    if (idx >= p.ne) {
+        return;
+    }
+
+    data_d[p.d_offset + dst_idx(idx)] = D_TYPE(FLOAT_TYPE(data_a[src0_idx(idx)]) / FLOAT_TYPE(data_b[src1_idx(idx)]));
+}
diff --git a/src/whisper_cpp/ggml/src/vulkan-shaders/gelu.comp b/src/whisper_cpp/ggml/src/vulkan-shaders/gelu.comp
new file mode 100644
index 00000000..4cc7a68c
--- /dev/null
+++ b/src/whisper_cpp/ggml/src/vulkan-shaders/gelu.comp
@@ -0,0 +1,25 @@
+#version 450
+
+#include "generic_head.comp"
+#include "types.comp"
+
+#extension GL_EXT_control_flow_attributes : enable
+
+layout(local_size_x = 512, local_size_y = 1, local_size_z = 1) in;
+
+layout (binding = 0) readonly buffer X {A_TYPE data_a[];};
+layout (binding = 1) writeonly buffer D {D_TYPE data_d[];};
+
+void main() {
+    const float GELU_COEF_A    = 0.044715f;
+    const float SQRT_2_OVER_PI = 0.79788456080286535587989211986876f;
+    const uint i = gl_GlobalInvocationID.z * 262144 + gl_GlobalInvocationID.y * 512 + gl_GlobalInvocationID.x;
+
+    if (i >= p.KX) {
+        return;
+    }
+
+    const float xi = float(data_a[i]);
+    const float val = SQRT_2_OVER_PI*xi*(1.0f + GELU_COEF_A*xi*xi);
+    data_d[i] = D_TYPE(0.5f*xi*(2.0f - 2.0f / (exp(2 * val) + 1)));
+}
diff --git a/src/whisper_cpp/ggml/src/vulkan-shaders/gelu_quick.comp b/src/whisper_cpp/ggml/src/vulkan-shaders/gelu_quick.comp
new file mode 100644
index 00000000..e6e6fcfd
--- /dev/null
+++ b/src/whisper_cpp/ggml/src/vulkan-shaders/gelu_quick.comp
@@ -0,0 +1,23 @@
+#version 450
+
+#include "generic_head.comp"
+#include "types.comp"
+
+#extension GL_EXT_control_flow_attributes : enable
+
+layout(local_size_x = 512, local_size_y = 1, local_size_z = 1) in;
+
+layout (binding = 0) readonly buffer X {A_TYPE data_a[];};
+layout (binding = 1) writeonly buffer D {D_TYPE data_d[];};
+
+void main() {
+    const float GELU_QUICK_COEF = -1.702f;
+    const uint i = gl_GlobalInvocationID.z * 262144 + gl_GlobalInvocationID.y * 512 + gl_GlobalInvocationID.x;
+
+    if (i >= p.KX) {
+        return;
+    }
+
+    const float x = float(data_a[i]);
+    data_d[i] = D_TYPE(x * (1.0f / (1.0f + exp(GELU_QUICK_COEF * x))));
+}
diff --git a/src/whisper_cpp/ggml/src/vulkan-shaders/generic_binary_head.comp b/src/whisper_cpp/ggml/src/vulkan-shaders/generic_binary_head.comp
new file mode 100644
index 00000000..b6beaff1
--- /dev/null
+++ b/src/whisper_cpp/ggml/src/vulkan-shaders/generic_binary_head.comp
@@ -0,0 +1,52 @@
+#extension GL_EXT_shader_16bit_storage : require
+
+layout (push_constant) uniform parameter
+{
+    uint ne;
+    uint ne00; uint ne01; uint ne02; uint ne03; uint nb00; uint nb01; uint nb02; uint nb03;
+    uint ne10; uint ne11; uint ne12; uint ne13; uint nb10; uint nb11; uint nb12; uint nb13;
+    uint ne20; uint ne21; uint ne22; uint ne23; uint nb20; uint nb21; uint nb22; uint nb23;
+    uint d_offset;
+    float param1; float param2; int param3;
+} p;
+
+layout(local_size_x = 512, local_size_y = 1, local_size_z = 1) in;
+
+layout (binding = 0) readonly buffer A {A_TYPE data_a[];};
+layout (binding = 1) readonly buffer B {B_TYPE data_b[];};
+layout (binding = 2) writeonly buffer D {D_TYPE data_d[];};
+
+uint get_idx() {
+    return gl_GlobalInvocationID.z * 262144 + gl_GlobalInvocationID.y * 512 + gl_GlobalInvocationID.x;
+}
+
+uint src0_idx(uint idx) {
+    const uint i03 = idx / (p.ne02*p.ne01*p.ne00);
+    const uint i03_offset = i03 * p.ne02*p.ne01*p.ne00;
+    const uint i02 = (idx - i03_offset) / (p.ne01*p.ne00);
+    const uint i02_offset = i02*p.ne01*p.ne00;
+    const uint i01 = (idx - i03_offset - i02_offset) / p.ne00;
+    const uint i00 = idx - i03_offset - i02_offset - i01*p.ne00;
+    return i03*p.nb03 + i02*p.nb02 + i01*p.nb01 + i00*p.nb00;
+}
+
+uint src1_idx(uint idx) {
+    const uint i03 = idx / (p.ne02*p.ne01*p.ne00);
+    const uint i03_offset = i03 * p.ne02*p.ne01*p.ne00;
+    const uint i02 = (idx - i03_offset) / (p.ne01*p.ne00);
+    const uint i02_offset = i02*p.ne01*p.ne00;
+    const uint i01 = (idx - i03_offset - i02_offset) / p.ne00;
+    const uint i00 = idx - i03_offset - i02_offset - i01*p.ne00;
+
+    return (i03 % p.ne13)*p.nb13 + (i02 % p.ne12)*p.nb12 + (i01 % p.ne11)*p.nb11 + (i00 % p.ne10)*p.nb10;
+}
+
+uint dst_idx(uint idx) {
+    const uint i23 = idx / (p.ne22*p.ne21*p.ne20);
+    const uint i23_offset = i23 * p.ne22*p.ne21*p.ne20;
+    const uint i22 = (idx - i23_offset) / (p.ne21*p.ne20);
+    const uint i22_offset = i22*p.ne21*p.ne20;
+    const uint i21 = (idx - i23_offset - i22_offset) / p.ne20;
+    const uint i20 = idx - i23_offset - i22_offset - i21*p.ne20;
+    return i23*p.nb23 + i22*p.nb22 + i21*p.nb21 + i20*p.nb20;
+}
diff --git a/src/whisper_cpp/ggml/src/vulkan-shaders/generic_head.comp b/src/whisper_cpp/ggml/src/vulkan-shaders/generic_head.comp
new file mode 100644
index 00000000..66e46ae6
--- /dev/null
+++ b/src/whisper_cpp/ggml/src/vulkan-shaders/generic_head.comp
@@ -0,0 +1,9 @@
+#extension GL_EXT_shader_16bit_storage : require
+
+layout (push_constant) uniform parameter
+{
+    uint KX;
+    uint KY;
+    float param1;
+    float param2;
+} p;
diff --git a/src/whisper_cpp/ggml/src/vulkan-shaders/generic_unary_head.comp b/src/whisper_cpp/ggml/src/vulkan-shaders/generic_unary_head.comp
new file mode 100644
index 00000000..eacdefc7
--- /dev/null
+++ b/src/whisper_cpp/ggml/src/vulkan-shaders/generic_unary_head.comp
@@ -0,0 +1,39 @@
+#extension GL_EXT_shader_16bit_storage : require
+
+layout (push_constant) uniform parameter
+{
+    uint ne;
+    uint ne00; uint ne01; uint ne02; uint ne03; uint nb00; uint nb01; uint nb02; uint nb03;
+    uint ne10; uint ne11; uint ne12; uint ne13; uint nb10; uint nb11; uint nb12; uint nb13;
+    uint d_offset;
+    float param1; float param2;
+} p;
+
+layout(local_size_x = 512, local_size_y = 1, local_size_z = 1) in;
+
+layout (binding = 0) readonly buffer A {A_TYPE data_a[];};
+layout (binding = 1) writeonly buffer D {D_TYPE data_d[];};
+
+uint get_idx() {
+    return gl_GlobalInvocationID.z * 262144 + gl_GlobalInvocationID.y * 512 + gl_GlobalInvocationID.x;
+}
+
+uint src0_idx(uint idx) {
+    const uint i03 = idx / (p.ne02*p.ne01*p.ne00);
+    const uint i03_offset = i03 * p.ne02*p.ne01*p.ne00;
+    const uint i02 = (idx - i03_offset) / (p.ne01*p.ne00);
+    const uint i02_offset = i02*p.ne01*p.ne00;
+    const uint i01 = (idx - i03_offset - i02_offset) / p.ne00;
+    const uint i00 = idx - i03_offset - i02_offset - i01*p.ne00;
+    return i03*p.nb03 + i02*p.nb02 + i01*p.nb01 + i00*p.nb00;
+}
+
+uint dst_idx(uint idx) {
+    const uint i13 = idx / (p.ne12*p.ne11*p.ne10);
+    const uint i13_offset = i13 * p.ne12*p.ne11*p.ne10;
+    const uint i12 = (idx - i13_offset) / (p.ne11*p.ne10);
+    const uint i12_offset = i12*p.ne11*p.ne10;
+    const uint i11 = (idx - i13_offset - i12_offset) / p.ne10;
+    const uint i10 = idx - i13_offset - i12_offset - i11*p.ne10;
+    return i13*p.nb13 + i12*p.nb12 + i11*p.nb11 + i10*p.nb10;
+}
diff --git a/src/whisper_cpp/ggml/src/vulkan-shaders/get_rows.comp b/src/whisper_cpp/ggml/src/vulkan-shaders/get_rows.comp
new file mode 100644
index 00000000..e9ff22ef
--- /dev/null
+++ b/src/whisper_cpp/ggml/src/vulkan-shaders/get_rows.comp
@@ -0,0 +1,26 @@
+#version 450
+
+#include "types.comp"
+#include "generic_binary_head.comp"
+
+void main() {
+    const uint i00 = gl_GlobalInvocationID.x;
+    const uint i10 = gl_GlobalInvocationID.y;
+    const uint i11 = (gl_GlobalInvocationID.z)/p.ne12;
+    const uint i12 = (gl_GlobalInvocationID.z)%p.ne12;
+
+    if (i00 >= p.ne00) {
+        return;
+    }
+
+    const uint i01 = data_b[i10*p.nb10 + i11*p.nb11 + i12*p.nb12];
+
+    const uint a_offset = i01*p.nb01 + i11*p.nb02 + i12*p.nb03;
+    const uint d_offset = i10*p.nb21 + i11*p.nb22 + i12*p.nb23;
+
+#ifndef OPTIMIZATION_ERROR_WORKAROUND
+    data_d[d_offset + i00] = D_TYPE(data_a[a_offset + i00]);
+#else
+    data_d[d_offset + i00] = data_a[a_offset + i00];
+#endif
+}
diff --git a/src/whisper_cpp/ggml/src/vulkan-shaders/get_rows_quant.comp b/src/whisper_cpp/ggml/src/vulkan-shaders/get_rows_quant.comp
new file mode 100644
index 00000000..53a9a96f
--- /dev/null
+++ b/src/whisper_cpp/ggml/src/vulkan-shaders/get_rows_quant.comp
@@ -0,0 +1,31 @@
+#version 450
+
+#include "types.comp"
+#include "generic_binary_head.comp"
+#include "dequant_funcs.comp"
+
+void main() {
+    const uint i00 = (gl_GlobalInvocationID.x)*2;
+    const uint i10 = gl_GlobalInvocationID.y;
+    const uint i11 = (gl_GlobalInvocationID.z)/p.ne12;
+    const uint i12 = (gl_GlobalInvocationID.z)%p.ne12;
+
+    if (i00 >= p.ne00) {
+        return;
+    }
+
+    const uint i01 = data_b[i10*p.nb10 + i11*p.nb11 + i12*p.nb12];
+
+    const uint a_offset = i01*p.nb01 + i11*p.nb02 + i12*p.nb03;
+    const uint d_offset = i10*p.nb21 + i11*p.nb22 + i12*p.nb23;
+
+    const uint ib = a_offset + i00/QUANT_K; // block index
+    const uint iqs = (i00%QUANT_K)/QUANT_R; // quant index
+    const uint iybs = i00 - i00%QUANT_K; // dst block start index
+    const uint y_offset = QUANT_R == 1 ? 1 : QUANT_K/2;
+
+    vec2 v = dequantize(ib, iqs, 0);
+
+    data_d[d_offset + iybs + iqs           ] = D_TYPE(v.x);
+    data_d[d_offset + iybs + iqs + y_offset] = D_TYPE(v.y);
+}
diff --git a/src/whisper_cpp/ggml/src/vulkan-shaders/group_norm.comp b/src/whisper_cpp/ggml/src/vulkan-shaders/group_norm.comp
new file mode 100644
index 00000000..5ad9b28d
--- /dev/null
+++ b/src/whisper_cpp/ggml/src/vulkan-shaders/group_norm.comp
@@ -0,0 +1,66 @@
+#version 450
+
+#include "generic_head.comp"
+#include "types.comp"
+
+#extension GL_EXT_control_flow_attributes : enable
+#define BLOCK_SIZE 512
+
+layout(local_size_x = BLOCK_SIZE, local_size_y = 1, local_size_z = 1) in;
+
+layout (binding = 0) readonly buffer X {A_TYPE data_a[];};
+layout (binding = 1) writeonly buffer D {D_TYPE data_d[];};
+
+shared float tmp[BLOCK_SIZE];
+
+void main() {
+    const uint group_size = p.KX;
+    const float eps = p.param1;
+
+    const uint tid = gl_LocalInvocationID.x;
+    const uint start = gl_WorkGroupID.x * group_size + tid;
+    const uint end = start + group_size;
+
+    tmp[tid] = 0.0f;
+
+    // Calculate mean
+    [[unroll]] for (uint col = start; col < end; col += BLOCK_SIZE) {
+        tmp[tid] += float(data_a[col]);
+    }
+
+    // tmp up partial tmps and write back result
+    barrier();
+    [[unroll]] for (int s = BLOCK_SIZE / 2; s > 0; s >>= 1) {
+        if (tid < s) {
+            tmp[tid] += tmp[tid + s];
+        }
+        barrier();
+    }
+
+    const float mean = tmp[0] / group_size;
+    barrier();
+    tmp[tid] = 0.0f;
+
+    // Calculate variance
+    [[unroll]] for (uint col = start; col < end; col += BLOCK_SIZE) {
+        const float xi = float(data_a[col]) - mean;
+        data_d[col] = D_TYPE(xi);
+        tmp[tid] += xi * xi;
+    }
+
+    // sum up partial sums and write back result
+    barrier();
+    [[unroll]] for (int s = BLOCK_SIZE / 2; s > 0; s >>= 1) {
+        if (tid < s) {
+            tmp[tid] += tmp[tid + s];
+        }
+        barrier();
+    }
+
+    const float variance = tmp[0] / group_size;
+    const float scale = inversesqrt(variance + eps);
+
+    [[unroll]] for (uint col = start; col < end; col += BLOCK_SIZE) {
+        data_d[col] *= D_TYPE(scale);
+    }
+}
diff --git a/src/whisper_cpp/ggml/src/vulkan-shaders/im2col.comp b/src/whisper_cpp/ggml/src/vulkan-shaders/im2col.comp
new file mode 100644
index 00000000..4d48610a
--- /dev/null
+++ b/src/whisper_cpp/ggml/src/vulkan-shaders/im2col.comp
@@ -0,0 +1,57 @@
+#version 450
+
+#extension GL_EXT_shader_16bit_storage : require
+
+layout (push_constant) uniform parameter
+{
+    uint batch_offset; uint offset_delta;
+    uint IC;
+    uint IW; uint IH;
+    uint OW; uint OH;
+    uint KW; uint KH;
+    uint pelements;
+    uint CHW;
+    int s0; int s1;
+    int p0; int p1;
+    int d0; int d1;
+} p;
+
+#include "types.comp"
+
+#define BLOCK_SIZE 256
+
+layout(local_size_x = BLOCK_SIZE, local_size_y = 1, local_size_z = 1) in;
+
+layout (binding = 0) readonly buffer X {A_TYPE data_a[];};
+layout (binding = 1) writeonly buffer D {D_TYPE data_d[];};
+
+void main() {
+    const uint i = gl_GlobalInvocationID.x;
+    if (i >= p.pelements) {
+        return;
+    }
+
+    const uint ksize = p.OW * (p.KH > 1 ? p.KW : 1);
+    const uint kx = i / ksize;
+    const uint kd = kx * ksize;
+    const uint ky = (i - kd) / p.OW;
+    const uint ix = i % p.OW;
+
+    const uint oh = gl_GlobalInvocationID.y;
+    const uint batch = gl_GlobalInvocationID.z / p.IC;
+    const uint ic = gl_GlobalInvocationID.z % p.IC;
+
+    const uint iiw = ix * p.s0 + kx * p.d0 - p.p0;
+    const uint iih = oh * p.s1 + ky * p.d1 - p.p1;
+
+    const uint offset_dst =
+        ((batch * p.OH + oh) * p.OW + ix) * p.CHW +
+        (ic * (p.KW * p.KH) + ky * p.KW + kx);
+
+    if (iih < 0 || iih >= p.IH || iiw < 0 || iiw >= p.IW) {
+        data_d[offset_dst] = D_TYPE(0.0f);
+    } else {
+        const uint offset_src = ic * p.offset_delta + batch * p.batch_offset;
+        data_d[offset_dst] = D_TYPE(data_a[offset_src + iih * p.IW + iiw]);
+    }
+}
diff --git a/src/whisper_cpp/ggml/src/vulkan-shaders/leaky_relu.comp b/src/whisper_cpp/ggml/src/vulkan-shaders/leaky_relu.comp
new file mode 100644
index 00000000..d90a99ae
--- /dev/null
+++ b/src/whisper_cpp/ggml/src/vulkan-shaders/leaky_relu.comp
@@ -0,0 +1,22 @@
+#version 450
+
+#include "generic_head.comp"
+#include "types.comp"
+
+#extension GL_EXT_control_flow_attributes : enable
+
+layout(local_size_x = 512, local_size_y = 1, local_size_z = 1) in;
+
+layout (binding = 0) readonly buffer X {A_TYPE data_a[];};
+layout (binding = 1) writeonly buffer D {D_TYPE data_d[];};
+
+void main() {
+    const uint i = gl_GlobalInvocationID.z * 262144 + gl_GlobalInvocationID.y * 512 + gl_GlobalInvocationID.x;
+
+    if (i >= p.KX) {
+        return;
+    }
+
+    const float val = float(data_a[i]);
+    data_d[i] = D_TYPE(max(val, 0.0f) + min(val, 0.0f) * p.param1);
+}
diff --git a/src/whisper_cpp/ggml/src/vulkan-shaders/mul.comp b/src/whisper_cpp/ggml/src/vulkan-shaders/mul.comp
new file mode 100644
index 00000000..bfb61c92
--- /dev/null
+++ b/src/whisper_cpp/ggml/src/vulkan-shaders/mul.comp
@@ -0,0 +1,14 @@
+#version 450
+
+#include "types.comp"
+#include "generic_binary_head.comp"
+
+void main() {
+    const uint idx = get_idx();
+
+    if (idx >= p.ne) {
+        return;
+    }
+
+    data_d[p.d_offset + dst_idx(idx)] = D_TYPE(FLOAT_TYPE(data_a[src0_idx(idx)]) * FLOAT_TYPE(data_b[src1_idx(idx)]));
+}
diff --git a/src/whisper_cpp/ggml/src/vulkan-shaders/mul_mat_split_k_reduce.comp b/src/whisper_cpp/ggml/src/vulkan-shaders/mul_mat_split_k_reduce.comp
new file mode 100644
index 00000000..825b9103
--- /dev/null
+++ b/src/whisper_cpp/ggml/src/vulkan-shaders/mul_mat_split_k_reduce.comp
@@ -0,0 +1,29 @@
+#version 450
+
+#extension GL_EXT_control_flow_attributes : enable
+
+layout(local_size_x = 256, local_size_y = 1, local_size_z = 1) in;
+
+layout (binding = 0) readonly buffer A {float data_a[];};
+layout (binding = 1) writeonly buffer D {float data_d[];};
+
+layout (push_constant) uniform parameter {
+    uint ne;
+    uint k_num;
+} p;
+
+void main() {
+    const uint idx = gl_GlobalInvocationID.x;
+
+    if (idx >= p.ne) {
+        return;
+    }
+
+    float result = 0.0f;
+
+    [[unroll]] for (uint i = 0; i < p.k_num; i++) {
+        result += data_a[i * p.ne + idx];
+    }
+
+    data_d[idx] = result;
+}
diff --git a/src/whisper_cpp/ggml/src/vulkan-shaders/mul_mat_vec.comp b/src/whisper_cpp/ggml/src/vulkan-shaders/mul_mat_vec.comp
new file mode 100644
index 00000000..d3ccba7f
--- /dev/null
+++ b/src/whisper_cpp/ggml/src/vulkan-shaders/mul_mat_vec.comp
@@ -0,0 +1,56 @@
+#version 450
+
+#ifdef FLOAT16
+#extension GL_EXT_shader_explicit_arithmetic_types_float16 : require
+#endif
+
+#include "mul_mat_vec_base.comp"
+
+layout(local_size_x_id = 0, local_size_y = 1, local_size_z = 1) in;
+
+layout (constant_id = 0) const uint BLOCK_SIZE = 32;
+
+shared FLOAT_TYPE tmp[BLOCK_SIZE];
+
+void main() {
+    const uint row = gl_WorkGroupID.x + gl_NumWorkGroups.x * gl_WorkGroupID.z;
+    const uint tid = gl_LocalInvocationID.x;
+
+    // There are not enough cols to use all threads
+    if (tid >= p.ncols) {
+        return;
+    }
+
+    const uint block_size = min(p.ncols, BLOCK_SIZE);
+
+    uint a_offset, b_offset, d_offset;
+    get_offsets(a_offset, b_offset, d_offset);
+
+    const uint y_offset = QUANT_R == 1 ? 1 : QUANT_K/2;
+
+    tmp[tid] = FLOAT_TYPE(0.0f);
+
+    [[unroll]] for (uint i = 0; i < p.ncols/block_size; i += 2) {
+        const uint col = i*block_size + 2*tid;
+        const uint ib = (row*p.ncols + col)/QUANT_K; // block index
+        const uint iqs = (col%QUANT_K)/QUANT_R; // quant index
+        const uint iybs = col - col%QUANT_K; // y block start index
+
+        vec2 v = dequantize(ib, iqs, a_offset / QUANT_K);
+
+        // matrix multiplication
+        tmp[tid] = fma(FLOAT_TYPE(v.x), FLOAT_TYPE(data_b[b_offset + iybs + iqs]), fma(FLOAT_TYPE(v.y), FLOAT_TYPE(data_b[b_offset + iybs + iqs + y_offset]), tmp[tid]));
+    }
+
+    // sum up partial sums and write back result
+    barrier();
+    [[unroll]] for (uint s = block_size/2; s > 0; s >>= 1) {
+        if (tid < s) {
+            tmp[tid] += tmp[tid + s];
+        }
+        barrier();
+    }
+    if (tid == 0) {
+        data_d[d_offset + row] = D_TYPE(tmp[0]);
+    }
+}
diff --git a/src/whisper_cpp/ggml/src/vulkan-shaders/mul_mat_vec_base.comp b/src/whisper_cpp/ggml/src/vulkan-shaders/mul_mat_vec_base.comp
new file mode 100644
index 00000000..5920bc93
--- /dev/null
+++ b/src/whisper_cpp/ggml/src/vulkan-shaders/mul_mat_vec_base.comp
@@ -0,0 +1,81 @@
+#extension GL_EXT_control_flow_attributes : enable
+#extension GL_EXT_shader_16bit_storage : require
+#extension GL_EXT_shader_8bit_storage : require
+
+#define K_QUANTS_PER_ITERATION 2
+
+#ifdef MUL_MAT_ID
+#define EXPERT_COUNT 8
+#endif
+
+#include "types.comp"
+
+layout (binding = 0) readonly buffer A {A_TYPE data_a[];};
+layout (binding = 1) readonly buffer B {B_TYPE data_b[];};
+layout (binding = 2) writeonly buffer D {D_TYPE data_d[];};
+#ifdef MUL_MAT_ID
+layout (binding = 3) readonly buffer IDS {int data_ids[];};
+#endif
+
+#include "dequant_funcs.comp"
+
+layout (push_constant) uniform parameter
+{
+    uint ncols;
+    uint stride_a;
+    uint stride_b;
+    uint stride_d;
+
+    uint batch_stride_a;
+    uint batch_stride_b;
+    uint batch_stride_d;
+
+#ifdef MUL_MAT_ID
+    uint nei0;
+    uint ne11;
+#else
+    uint ne02;
+    uint ne12;
+    uint broadcast2;
+    uint broadcast3;
+#endif
+} p;
+
+void get_offsets(out uint a_offset, out uint b_offset, out uint d_offset) {
+#ifdef MUL_MAT_ID
+    const uint expert_idx = gl_GlobalInvocationID.y;
+#else
+    const uint batch_idx = gl_GlobalInvocationID.y;
+#endif
+
+#ifndef MUL_MAT_ID
+    const uint i13 = batch_idx / p.ne12;
+    const uint i12 = batch_idx % p.ne12;
+
+    const uint i03 = i13 / p.broadcast3;
+    const uint i02 = i12 / p.broadcast2;
+
+    const uint batch_idx_a = i03 * p.ne02 + i02;
+#else
+    const uint expert_id = data_ids[expert_idx];
+#endif
+
+    a_offset =
+#ifdef MUL_MAT_ID
+            expert_id * p.batch_stride_a;
+#else
+            batch_idx_a * p.batch_stride_a;
+#endif
+    b_offset =
+#ifdef MUL_MAT_ID
+            (expert_idx % p.ne11) * p.stride_b;
+#else
+            batch_idx * p.batch_stride_b;
+#endif
+    d_offset =
+#ifdef MUL_MAT_ID
+            expert_idx * p.stride_d;
+#else
+            batch_idx * p.batch_stride_d;
+#endif
+}
diff --git a/src/whisper_cpp/ggml/src/vulkan-shaders/mul_mat_vec_nc.comp b/src/whisper_cpp/ggml/src/vulkan-shaders/mul_mat_vec_nc.comp
new file mode 100644
index 00000000..1cc4996d
--- /dev/null
+++ b/src/whisper_cpp/ggml/src/vulkan-shaders/mul_mat_vec_nc.comp
@@ -0,0 +1,71 @@
+#version 450
+
+#extension GL_EXT_control_flow_attributes : enable
+#extension GL_EXT_shader_16bit_storage : require
+
+#define BLOCK_SIZE 32
+#define FLOAT_TYPE float
+
+layout(local_size_x = BLOCK_SIZE, local_size_y = 1, local_size_z = 1) in;
+
+layout (binding = 0) readonly buffer A {A_TYPE data_a[];};
+layout (binding = 1) readonly buffer B {B_TYPE data_b[];};
+layout (binding = 2) writeonly buffer D {D_TYPE dst[];};
+
+layout (push_constant) uniform parameter
+{
+    uint ncols_x;
+    uint nrows_x;
+    uint row_stride_x;
+    uint channel_stride_x;
+    uint channel_x_divisor;
+    uint b_offset;
+    uint d_offset;
+} p;
+
+shared FLOAT_TYPE tmp[BLOCK_SIZE];
+
+void main() {
+    const uint tid       = gl_LocalInvocationID.x;
+    const uint row_x     = gl_GlobalInvocationID.y;
+    const uint channel   = gl_GlobalInvocationID.z;
+    const uint channel_x = channel / p.channel_x_divisor;
+
+    const uint nrows_y   = p.ncols_x;
+    const uint nrows_dst = p.nrows_x;
+    const uint row_dst   = row_x;
+
+    const uint idst = channel*nrows_dst + row_dst;
+
+    tmp[tid] = 0.0f;
+
+    for (uint col_x0 = 0; col_x0 < p.ncols_x; col_x0 += BLOCK_SIZE) {
+        const uint col_x = col_x0 + tid;
+
+        if (col_x >= p.ncols_x) {
+            break;
+        }
+
+        const uint row_y = col_x;
+
+        const uint ix = channel_x*p.channel_stride_x + row_x*p.row_stride_x + col_x;
+        const uint iy = channel*nrows_y + row_y;
+
+        const FLOAT_TYPE xi = FLOAT_TYPE(data_a[ix]);
+
+        tmp[tid] = fma(xi, FLOAT_TYPE(data_b[iy]), tmp[tid]);
+    }
+
+    // sum up partial sums and write back result
+    barrier();
+    [[unroll]] for (int s = BLOCK_SIZE / 2; s > 0; s >>= 1) {
+        if (tid < s) {
+            tmp[tid] += tmp[tid + s];
+        }
+        barrier();
+    }
+
+    if (tid == 0) {
+        dst[idst] = tmp[0];
+    }
+}
diff --git a/src/whisper_cpp/ggml/src/vulkan-shaders/mul_mat_vec_p021.comp b/src/whisper_cpp/ggml/src/vulkan-shaders/mul_mat_vec_p021.comp
new file mode 100644
index 00000000..9b443807
--- /dev/null
+++ b/src/whisper_cpp/ggml/src/vulkan-shaders/mul_mat_vec_p021.comp
@@ -0,0 +1,73 @@
+#version 450
+
+#extension GL_EXT_control_flow_attributes : enable
+#extension GL_EXT_shader_16bit_storage : require
+
+#define BLOCK_SIZE 32
+#define FLOAT_TYPE float
+
+layout(local_size_x = BLOCK_SIZE, local_size_y = 1, local_size_z = 1) in;
+
+layout (binding = 0) readonly buffer A {A_TYPE data_a[];};
+layout (binding = 1) readonly buffer B {B_TYPE data_b[];};
+layout (binding = 2) writeonly buffer D {D_TYPE dst[];};
+
+layout (push_constant) uniform parameter
+{
+    uint ncols_x;
+    uint nrows_x;
+    uint nchannels_x;
+    uint nchannels_y;
+    uint b_offset;
+    uint d_offset;
+} p;
+
+shared FLOAT_TYPE tmp[BLOCK_SIZE];
+
+void main() {
+    const uint tid = gl_LocalInvocationID.x;
+    const uint row_x = gl_GlobalInvocationID.y;
+    const uint channel = gl_GlobalInvocationID.z;
+    const uint channel_x = channel / (p.nchannels_y / p.nchannels_x);
+
+    const uint nrows_y = p.ncols_x;
+    const uint nrows_dst = p.nrows_x;
+    const uint row_dst = row_x;
+
+    tmp[tid] = FLOAT_TYPE(0.0f);
+
+    for (uint col_x0 = 0; col_x0 < p.ncols_x; col_x0 += BLOCK_SIZE) {
+        const uint col_x = col_x0 + tid;
+
+        if (col_x >= p.ncols_x) {
+            break;
+        }
+
+        // x is transposed and permuted
+        const uint ix = row_x*p.nchannels_x*p.ncols_x + channel_x*p.ncols_x + col_x;
+        const FLOAT_TYPE xi = FLOAT_TYPE(data_a[ix]);
+
+        const uint row_y = col_x;
+
+        // y is not transposed but permuted
+        const uint iy = channel*nrows_y + row_y;
+
+        tmp[tid] = fma(xi, FLOAT_TYPE(data_b[iy]), tmp[tid]);
+    }
+
+    // dst is not transposed and not permuted
+    const uint idst = channel*nrows_dst + row_dst;
+
+    // sum up partial sums and write back result
+    barrier();
+    [[unroll]] for (int s = BLOCK_SIZE / 2; s > 0; s >>= 1) {
+        if (tid < s) {
+            tmp[tid] += tmp[tid + s];
+        }
+        barrier();
+    }
+
+    if (tid == 0) {
+        dst[idst] = tmp[0];
+    }
+}
diff --git a/src/whisper_cpp/ggml/src/vulkan-shaders/mul_mat_vec_q2_k.comp b/src/whisper_cpp/ggml/src/vulkan-shaders/mul_mat_vec_q2_k.comp
new file mode 100644
index 00000000..ec8eadcd
--- /dev/null
+++ b/src/whisper_cpp/ggml/src/vulkan-shaders/mul_mat_vec_q2_k.comp
@@ -0,0 +1,74 @@
+#version 450
+
+#include "mul_mat_vec_base.comp"
+
+layout(local_size_x = 32, local_size_y = 1, local_size_z = 1) in;
+
+shared FLOAT_TYPE tmp[32];
+
+void main() {
+    const uint row = gl_WorkGroupID.x + gl_NumWorkGroups.x * gl_WorkGroupID.z;
+
+    uint a_offset, b_offset, d_offset;
+    get_offsets(a_offset, b_offset, d_offset);
+
+    const uint num_blocks_per_row = p.ncols / QUANT_K;
+    const uint ib0 = a_offset / QUANT_K + row*num_blocks_per_row;
+
+    const uint tid = gl_LocalInvocationID.x/K_QUANTS_PER_ITERATION;  // 0...31 or 0...16
+    const uint ix  = gl_LocalInvocationID.x%K_QUANTS_PER_ITERATION;  // 0 or 0, 1
+
+    const uint step = 16/K_QUANTS_PER_ITERATION;            // 16 or 8
+
+    const uint v_im = tid/step;                             // 0 or 1. 0 computes 0..., 1 computes 128...
+    const uint v_in = tid - step*v_im;                      // 0...15 or 0...7
+
+    const uint l0 = K_QUANTS_PER_ITERATION*v_in;            // 0...15
+    const uint q_offset = 32*v_im + l0;
+    const uint s_offset = 8*v_im;
+    const uint y_offset = 128*v_im + l0;
+
+    tmp[16 * ix + tid] = FLOAT_TYPE(0.0); // partial sum for thread in warp
+
+    [[unroll]] for (uint i = ix; i < num_blocks_per_row; i += K_QUANTS_PER_ITERATION) {
+        const uint y_idx = i * QUANT_K + y_offset;
+
+        const FLOAT_TYPE dall = FLOAT_TYPE(data_a[ib0 + i].d.x);
+        const FLOAT_TYPE dmin = FLOAT_TYPE(data_a[ib0 + i].d.y);
+
+        FLOAT_TYPE sum1 = FLOAT_TYPE(0.0);
+        FLOAT_TYPE sum2 = FLOAT_TYPE(0.0);
+        for (int l = 0; l < K_QUANTS_PER_ITERATION; ++l) {
+            sum1 = fma(FLOAT_TYPE(data_b[b_offset + y_idx + l +  0]), FLOAT_TYPE(data_a[ib0 + i].scales[s_offset + 0] & 0xF) * FLOAT_TYPE((data_a[ib0 + i].qs[q_offset + l + 0] >> 0) & 3),
+                   fma(FLOAT_TYPE(data_b[b_offset + y_idx + l + 16]), FLOAT_TYPE(data_a[ib0 + i].scales[s_offset + 1] & 0xF) * FLOAT_TYPE((data_a[ib0 + i].qs[q_offset + l +16] >> 0) & 3),
+                   fma(FLOAT_TYPE(data_b[b_offset + y_idx + l + 32]), FLOAT_TYPE(data_a[ib0 + i].scales[s_offset + 2] & 0xF) * FLOAT_TYPE((data_a[ib0 + i].qs[q_offset + l + 0] >> 2) & 3),
+                   fma(FLOAT_TYPE(data_b[b_offset + y_idx + l + 48]), FLOAT_TYPE(data_a[ib0 + i].scales[s_offset + 3] & 0xF) * FLOAT_TYPE((data_a[ib0 + i].qs[q_offset + l +16] >> 2) & 3),
+                   fma(FLOAT_TYPE(data_b[b_offset + y_idx + l + 64]), FLOAT_TYPE(data_a[ib0 + i].scales[s_offset + 4] & 0xF) * FLOAT_TYPE((data_a[ib0 + i].qs[q_offset + l + 0] >> 4) & 3),
+                   fma(FLOAT_TYPE(data_b[b_offset + y_idx + l + 80]), FLOAT_TYPE(data_a[ib0 + i].scales[s_offset + 5] & 0xF) * FLOAT_TYPE((data_a[ib0 + i].qs[q_offset + l +16] >> 4) & 3),
+                   fma(FLOAT_TYPE(data_b[b_offset + y_idx + l + 96]), FLOAT_TYPE(data_a[ib0 + i].scales[s_offset + 6] & 0xF) * FLOAT_TYPE((data_a[ib0 + i].qs[q_offset + l + 0] >> 6) & 3),
+                   fma(FLOAT_TYPE(data_b[b_offset + y_idx + l +112]), FLOAT_TYPE(data_a[ib0 + i].scales[s_offset + 7] & 0xF) * FLOAT_TYPE((data_a[ib0 + i].qs[q_offset + l +16] >> 6) & 3), sum1))))))));
+            sum2 = fma(FLOAT_TYPE(data_b[b_offset + y_idx + l +  0]), FLOAT_TYPE((data_a[ib0 + i].scales[s_offset + 0] >> 4) & 0xF),
+                   fma(FLOAT_TYPE(data_b[b_offset + y_idx + l + 16]), FLOAT_TYPE((data_a[ib0 + i].scales[s_offset + 1] >> 4) & 0xF),
+                   fma(FLOAT_TYPE(data_b[b_offset + y_idx + l + 32]), FLOAT_TYPE((data_a[ib0 + i].scales[s_offset + 2] >> 4) & 0xF),
+                   fma(FLOAT_TYPE(data_b[b_offset + y_idx + l + 48]), FLOAT_TYPE((data_a[ib0 + i].scales[s_offset + 3] >> 4) & 0xF),
+                   fma(FLOAT_TYPE(data_b[b_offset + y_idx + l + 64]), FLOAT_TYPE((data_a[ib0 + i].scales[s_offset + 4] >> 4) & 0xF),
+                   fma(FLOAT_TYPE(data_b[b_offset + y_idx + l + 80]), FLOAT_TYPE((data_a[ib0 + i].scales[s_offset + 5] >> 4) & 0xF),
+                   fma(FLOAT_TYPE(data_b[b_offset + y_idx + l + 96]), FLOAT_TYPE((data_a[ib0 + i].scales[s_offset + 6] >> 4) & 0xF),
+                   fma(FLOAT_TYPE(data_b[b_offset + y_idx + l +112]), FLOAT_TYPE((data_a[ib0 + i].scales[s_offset + 7] >> 4) & 0xF), sum2))))))));
+        }
+        const uint tmp_idx = 16 * ix + tid;
+        tmp[tmp_idx] = fma(dall, sum1, fma(-dmin, sum2, tmp[tmp_idx]));
+    }
+
+    // sum up partial sums and write back result
+    barrier();
+    [[unroll]] for (uint s = 16; s > 0; s >>= 1) {
+        if (tid < s) {
+            tmp[tid] += tmp[tid + s];
+        }
+        barrier();
+    }
+    if (tid == 0) {
+        data_d[d_offset + row] = D_TYPE(tmp[0]);
+    }
+}
diff --git a/src/whisper_cpp/ggml/src/vulkan-shaders/mul_mat_vec_q3_k.comp b/src/whisper_cpp/ggml/src/vulkan-shaders/mul_mat_vec_q3_k.comp
new file mode 100644
index 00000000..3ca4ad85
--- /dev/null
+++ b/src/whisper_cpp/ggml/src/vulkan-shaders/mul_mat_vec_q3_k.comp
@@ -0,0 +1,67 @@
+#version 450
+
+#include "mul_mat_vec_base.comp"
+
+layout(local_size_x = 32, local_size_y = 1, local_size_z = 1) in;
+
+shared FLOAT_TYPE tmp[32];
+
+void main() {
+    const uint row = gl_WorkGroupID.x + gl_NumWorkGroups.x * gl_WorkGroupID.z;
+
+    uint a_offset, b_offset, d_offset;
+    get_offsets(a_offset, b_offset, d_offset);
+
+    const uint num_blocks_per_row = p.ncols / QUANT_K;
+    const uint ib0 = a_offset / QUANT_K + row*num_blocks_per_row;
+
+    const uint tid = gl_LocalInvocationID.x/K_QUANTS_PER_ITERATION;  // 0...31 or 0...16
+    const uint ix  = gl_LocalInvocationID.x%K_QUANTS_PER_ITERATION;  // 0 or 0, 1
+
+    const uint step = 16/K_QUANTS_PER_ITERATION;            // 16 or 8
+
+    const uint v_im = tid/step;                             // 0 or 1. 0 computes 0..., 1 computes 128...
+    const uint v_in = tid - step*v_im;                      // 0...15 or 0...7
+
+    const uint8_t m = uint8_t(1 << (4 * v_im));
+
+    const uint l0 = K_QUANTS_PER_ITERATION*v_in;            // 0...15
+    const uint q_offset = 32*v_im + l0;
+    const uint y_offset = 128*v_im + l0;
+
+    tmp[16 * ix + tid] = FLOAT_TYPE(0.0); // partial sum for thread in warp
+
+    const uint s_shift = 4 * v_im;
+
+    [[unroll]] for (uint i = ix; i < num_blocks_per_row; i += K_QUANTS_PER_ITERATION) {
+        const uint y_idx = i * QUANT_K + y_offset;
+
+        const FLOAT_TYPE d = FLOAT_TYPE(data_a[ib0 + i].d);
+
+        FLOAT_TYPE sum = FLOAT_TYPE(0.0);
+        for (int l = 0; l < K_QUANTS_PER_ITERATION; ++l) {
+            sum = fma(FLOAT_TYPE(data_b[b_offset + y_idx + l +  0]) * FLOAT_TYPE(int8_t(((data_a[ib0 + i].scales[0] >> s_shift) & 0xF) | ((data_a[ib0 + i].scales[ 8] >> (s_shift + 0) & 0x3) << 4)) - 32), FLOAT_TYPE(((data_a[ib0 + i].qs[q_offset + l   ]     ) & 3) - (((data_a[ib0 + i].hmask[l0 + l   ] & (m << 0)) != 0) ? 0 : 4)),
+                  fma(FLOAT_TYPE(data_b[b_offset + y_idx + l + 32]) * FLOAT_TYPE(int8_t(((data_a[ib0 + i].scales[2] >> s_shift) & 0xF) | ((data_a[ib0 + i].scales[10] >> (s_shift + 0) & 0x3) << 4)) - 32), FLOAT_TYPE(((data_a[ib0 + i].qs[q_offset + l   ] >> 2) & 3) - (((data_a[ib0 + i].hmask[l0 + l   ] & (m << 1)) != 0) ? 0 : 4)),
+                  fma(FLOAT_TYPE(data_b[b_offset + y_idx + l + 64]) * FLOAT_TYPE(int8_t(((data_a[ib0 + i].scales[4] >> s_shift) & 0xF) | ((data_a[ib0 + i].scales[ 8] >> (s_shift + 2) & 0x3) << 4)) - 32), FLOAT_TYPE(((data_a[ib0 + i].qs[q_offset + l   ] >> 4) & 3) - (((data_a[ib0 + i].hmask[l0 + l   ] & (m << 2)) != 0) ? 0 : 4)),
+                  fma(FLOAT_TYPE(data_b[b_offset + y_idx + l + 96]) * FLOAT_TYPE(int8_t(((data_a[ib0 + i].scales[6] >> s_shift) & 0xF) | ((data_a[ib0 + i].scales[10] >> (s_shift + 2) & 0x3) << 4)) - 32), FLOAT_TYPE(((data_a[ib0 + i].qs[q_offset + l   ] >> 6) & 3) - (((data_a[ib0 + i].hmask[l0 + l   ] & (m << 3)) != 0) ? 0 : 4)),
+                  fma(FLOAT_TYPE(data_b[b_offset + y_idx + l + 16]) * FLOAT_TYPE(int8_t(((data_a[ib0 + i].scales[1] >> s_shift) & 0xF) | ((data_a[ib0 + i].scales[ 9] >> (s_shift + 0) & 0x3) << 4)) - 32), FLOAT_TYPE(((data_a[ib0 + i].qs[q_offset + l+16]     ) & 3) - (((data_a[ib0 + i].hmask[l0 + l+16] & (m << 0)) != 0) ? 0 : 4)),
+                  fma(FLOAT_TYPE(data_b[b_offset + y_idx + l + 48]) * FLOAT_TYPE(int8_t(((data_a[ib0 + i].scales[3] >> s_shift) & 0xF) | ((data_a[ib0 + i].scales[11] >> (s_shift + 0) & 0x3) << 4)) - 32), FLOAT_TYPE(((data_a[ib0 + i].qs[q_offset + l+16] >> 2) & 3) - (((data_a[ib0 + i].hmask[l0 + l+16] & (m << 1)) != 0) ? 0 : 4)),
+                  fma(FLOAT_TYPE(data_b[b_offset + y_idx + l + 80]) * FLOAT_TYPE(int8_t(((data_a[ib0 + i].scales[5] >> s_shift) & 0xF) | ((data_a[ib0 + i].scales[ 9] >> (s_shift + 2) & 0x3) << 4)) - 32), FLOAT_TYPE(((data_a[ib0 + i].qs[q_offset + l+16] >> 4) & 3) - (((data_a[ib0 + i].hmask[l0 + l+16] & (m << 2)) != 0) ? 0 : 4)),
+                  fma(FLOAT_TYPE(data_b[b_offset + y_idx + l +112]) * FLOAT_TYPE(int8_t(((data_a[ib0 + i].scales[7] >> s_shift) & 0xF) | ((data_a[ib0 + i].scales[11] >> (s_shift + 2) & 0x3) << 4)) - 32), FLOAT_TYPE(((data_a[ib0 + i].qs[q_offset + l+16] >> 6) & 3) - (((data_a[ib0 + i].hmask[l0 + l+16] & (m << 3)) != 0) ? 0 : 4)), sum))))))));
+        }
+        const uint tmp_idx = 16 * ix + tid;
+        tmp[tmp_idx] = fma(d, sum, tmp[tmp_idx]);
+    }
+
+    // sum up partial sums and write back result
+    barrier();
+    [[unroll]] for (uint s = 16; s > 0; s >>= 1) {
+        if (tid < s) {
+            tmp[tid] += tmp[tid + s];
+        }
+        barrier();
+    }
+    if (tid == 0) {
+        data_d[d_offset + row] = D_TYPE(tmp[0]);
+    }
+}
diff --git a/src/whisper_cpp/ggml/src/vulkan-shaders/mul_mat_vec_q4_k.comp b/src/whisper_cpp/ggml/src/vulkan-shaders/mul_mat_vec_q4_k.comp
new file mode 100644
index 00000000..d91e00e1
--- /dev/null
+++ b/src/whisper_cpp/ggml/src/vulkan-shaders/mul_mat_vec_q4_k.comp
@@ -0,0 +1,118 @@
+#version 450
+
+#include "mul_mat_vec_base.comp"
+
+layout(local_size_x = 32, local_size_y = 1, local_size_z = 1) in;
+
+shared FLOAT_TYPE tmp[32];
+
+void main() {
+    const uint row = gl_WorkGroupID.x + gl_NumWorkGroups.x * gl_WorkGroupID.z;
+
+    uint a_offset, b_offset, d_offset;
+    get_offsets(a_offset, b_offset, d_offset);
+
+    const uint num_blocks_per_row = p.ncols / QUANT_K;
+    const uint ib0 = a_offset / QUANT_K + row*num_blocks_per_row;
+
+    const uint tid = gl_LocalInvocationID.x/K_QUANTS_PER_ITERATION;  // 0...31 or 0...16
+    const uint ix  = gl_LocalInvocationID.x%K_QUANTS_PER_ITERATION;  // 0 or 0, 1
+
+    const uint step = 8/K_QUANTS_PER_ITERATION;             // 8 or 4
+
+    const uint il = tid/step;                               // 0...3
+    const uint ir = tid - step*il;                          // 0...7 or 0...3
+    const uint n =  2 * K_QUANTS_PER_ITERATION;             // 2 or 4
+
+    const uint v_im = il / 2;  // 0 or 1. 0 computes 0,32 + 128,160, 1 computes 64,96 + 192,224
+    const uint v_in = il % 2;
+
+    const uint l0 = n * (2 * ir + v_in);            // 0...15
+    const uint q_offset = 32*v_im + l0;
+    const uint y_offset = 64*v_im + l0;
+
+    tmp[16 * ix + tid] = FLOAT_TYPE(0.0); // partial sum for thread in warp
+
+    [[unroll]] for (uint i = ix; i < num_blocks_per_row; i += K_QUANTS_PER_ITERATION) {
+        const uint y1_idx = i * QUANT_K + y_offset;
+        const uint y2_idx = y1_idx + 128;
+
+        const FLOAT_TYPE dall = FLOAT_TYPE(data_a[ib0 + i].d.x);
+        const FLOAT_TYPE dmin = FLOAT_TYPE(data_a[ib0 + i].d.y);
+
+        const uint8_t sc0 = uint8_t(  data_a[ib0 + i].scales[v_im * 2    ]       & 0x3f);
+        const uint8_t sc1 = uint8_t(  data_a[ib0 + i].scales[v_im * 2 + 1]       & 0x3f);
+        const uint8_t sc2 = uint8_t(  data_a[ib0 + i].scales[v_im * 2 + 4]       & 0x3f);
+        const uint8_t sc3 = uint8_t(  data_a[ib0 + i].scales[v_im * 2 + 5]       & 0x3f);
+        const uint8_t sc4 = uint8_t(( data_a[ib0 + i].scales[v_im * 2 + 8]       & 0x0f) | ((data_a[ib0 + i].scales[v_im * 2    ] & 0xc0) >> 2));
+        const uint8_t sc5 = uint8_t(( data_a[ib0 + i].scales[v_im * 2 + 9]       & 0x0f) | ((data_a[ib0 + i].scales[v_im * 2 + 1] & 0xc0) >> 2));
+        const uint8_t sc6 = uint8_t(((data_a[ib0 + i].scales[v_im * 2 + 8] >> 4) & 0x0f) | ((data_a[ib0 + i].scales[v_im * 2 + 4] & 0xc0) >> 2));
+        const uint8_t sc7 = uint8_t(((data_a[ib0 + i].scales[v_im * 2 + 9] >> 4) & 0x0f) | ((data_a[ib0 + i].scales[v_im * 2 + 5] & 0xc0) >> 2));
+
+#if K_QUANTS_PER_ITERATION == 2
+        const uint8_t q4_0  = uint8_t(data_a[ib0 + i].qs[q_offset     ] & 0xf);
+        const uint8_t q4_1  = uint8_t(data_a[ib0 + i].qs[q_offset +  1] & 0xf);
+        const uint8_t q4_2  = uint8_t(data_a[ib0 + i].qs[q_offset +  2] & 0xf);
+        const uint8_t q4_3  = uint8_t(data_a[ib0 + i].qs[q_offset +  3] & 0xf);
+        const uint8_t q4_4  = uint8_t(data_a[ib0 + i].qs[q_offset     ]  >> 4);
+        const uint8_t q4_5  = uint8_t(data_a[ib0 + i].qs[q_offset +  1]  >> 4);
+        const uint8_t q4_6  = uint8_t(data_a[ib0 + i].qs[q_offset +  2]  >> 4);
+        const uint8_t q4_7  = uint8_t(data_a[ib0 + i].qs[q_offset +  3]  >> 4);
+        const uint8_t q4_8  = uint8_t(data_a[ib0 + i].qs[q_offset + 64] & 0xf);
+        const uint8_t q4_9  = uint8_t(data_a[ib0 + i].qs[q_offset + 65] & 0xf);
+        const uint8_t q4_10 = uint8_t(data_a[ib0 + i].qs[q_offset + 66] & 0xf);
+        const uint8_t q4_11 = uint8_t(data_a[ib0 + i].qs[q_offset + 67] & 0xf);
+        const uint8_t q4_12 = uint8_t(data_a[ib0 + i].qs[q_offset + 64]  >> 4);
+        const uint8_t q4_13 = uint8_t(data_a[ib0 + i].qs[q_offset + 65]  >> 4);
+        const uint8_t q4_14 = uint8_t(data_a[ib0 + i].qs[q_offset + 66]  >> 4);
+        const uint8_t q4_15 = uint8_t(data_a[ib0 + i].qs[q_offset + 67]  >> 4);
+
+        const FLOAT_TYPE sx = fma(FLOAT_TYPE(data_b[b_offset + y1_idx]),      q4_0,  fma(FLOAT_TYPE(data_b[b_offset + y1_idx + 1]),  q4_1,  fma(FLOAT_TYPE(data_b[b_offset + y1_idx + 2]),  q4_2,  FLOAT_TYPE(data_b[b_offset + y1_idx + 3]) *  q4_3)));
+        const FLOAT_TYPE sy = fma(FLOAT_TYPE(data_b[b_offset + y1_idx + 32]), q4_4,  fma(FLOAT_TYPE(data_b[b_offset + y1_idx + 33]), q4_5,  fma(FLOAT_TYPE(data_b[b_offset + y1_idx + 34]), q4_6,  FLOAT_TYPE(data_b[b_offset + y1_idx + 35]) * q4_7)));
+        const FLOAT_TYPE sz = fma(FLOAT_TYPE(data_b[b_offset + y2_idx]),      q4_8,  fma(FLOAT_TYPE(data_b[b_offset + y2_idx + 1]),  q4_9,  fma(FLOAT_TYPE(data_b[b_offset + y2_idx + 2]),  q4_10, FLOAT_TYPE(data_b[b_offset + y2_idx + 3]) *  q4_11)));
+        const FLOAT_TYPE sw = fma(FLOAT_TYPE(data_b[b_offset + y2_idx + 32]), q4_12, fma(FLOAT_TYPE(data_b[b_offset + y2_idx + 33]), q4_13, fma(FLOAT_TYPE(data_b[b_offset + y2_idx + 34]), q4_14, FLOAT_TYPE(data_b[b_offset + y2_idx + 35]) * q4_15)));
+        const FLOAT_TYPE smin =
+            fma(FLOAT_TYPE(data_b[b_offset + y1_idx    ]), sc2, fma(FLOAT_TYPE(data_b[b_offset + y1_idx + 32]), sc3, fma(FLOAT_TYPE(data_b[b_offset + y2_idx    ]), sc6, fma(FLOAT_TYPE(data_b[b_offset + y2_idx + 32]), sc7,
+            fma(FLOAT_TYPE(data_b[b_offset + y1_idx + 1]), sc2, fma(FLOAT_TYPE(data_b[b_offset + y1_idx + 33]), sc3, fma(FLOAT_TYPE(data_b[b_offset + y2_idx + 1]), sc6, fma(FLOAT_TYPE(data_b[b_offset + y2_idx + 33]), sc7,
+            fma(FLOAT_TYPE(data_b[b_offset + y1_idx + 2]), sc2, fma(FLOAT_TYPE(data_b[b_offset + y1_idx + 34]), sc3, fma(FLOAT_TYPE(data_b[b_offset + y2_idx + 2]), sc6, fma(FLOAT_TYPE(data_b[b_offset + y2_idx + 34]), sc7,
+            fma(FLOAT_TYPE(data_b[b_offset + y1_idx + 3]), sc2, fma(FLOAT_TYPE(data_b[b_offset + y1_idx + 35]), sc3, fma(FLOAT_TYPE(data_b[b_offset + y2_idx + 3]), sc6,     FLOAT_TYPE(data_b[b_offset + y2_idx + 35]) * sc7)))))))))))))));
+        const uint tmp_idx = 16 * ix + tid;
+        tmp[tmp_idx] = fma(dall, fma(sx, sc0, fma(sy, sc1, fma(sz, sc4, sw * sc5))), fma(-dmin, smin, tmp[tmp_idx]));
+#else
+        const uint8_t q4_0 = uint8_t(data_a[ib0 + i].qs[q_offset     ] & 0xf);
+        const uint8_t q4_1 = uint8_t(data_a[ib0 + i].qs[q_offset +  1] & 0xf);
+        const uint8_t q4_2 = uint8_t(data_a[ib0 + i].qs[q_offset     ]  >> 4);
+        const uint8_t q4_3 = uint8_t(data_a[ib0 + i].qs[q_offset +  1]  >> 4);
+        const uint8_t q4_4 = uint8_t(data_a[ib0 + i].qs[q_offset + 64] & 0xf);
+        const uint8_t q4_5 = uint8_t(data_a[ib0 + i].qs[q_offset + 65] & 0xf);
+        const uint8_t q4_6 = uint8_t(data_a[ib0 + i].qs[q_offset + 64]  >> 4);
+        const uint8_t q4_7 = uint8_t(data_a[ib0 + i].qs[q_offset + 65]  >> 4);
+
+        const FLOAT_TYPE sx = fma(FLOAT_TYPE(data_b[b_offset + y1_idx     ]), q4_0, FLOAT_TYPE(data_b[b_offset + y1_idx +  1]) * q4_1);
+        const FLOAT_TYPE sy = fma(FLOAT_TYPE(data_b[b_offset + y1_idx + 32]), q4_2, FLOAT_TYPE(data_b[b_offset + y1_idx + 33]) * q4_3);
+        const FLOAT_TYPE sz = fma(FLOAT_TYPE(data_b[b_offset + y2_idx     ]), q4_4, FLOAT_TYPE(data_b[b_offset + y2_idx +  1]) * q4_5);
+        const FLOAT_TYPE sw = fma(FLOAT_TYPE(data_b[b_offset + y2_idx + 32]), q4_6, FLOAT_TYPE(data_b[b_offset + y2_idx + 33]) * q4_7);
+        const FLOAT_TYPE smin =
+            fma(FLOAT_TYPE(data_b[b_offset + y1_idx    ]), sc2, fma(FLOAT_TYPE(data_b[b_offset + y1_idx + 32]), sc3, fma(FLOAT_TYPE(data_b[b_offset + y2_idx    ]), sc6, fma(FLOAT_TYPE(data_b[b_offset + y2_idx + 32]), sc7,
+          + fma(FLOAT_TYPE(data_b[b_offset + y1_idx + 1]), sc2, fma(FLOAT_TYPE(data_b[b_offset + y1_idx + 33]), sc3, fma(FLOAT_TYPE(data_b[b_offset + y2_idx + 1]), sc6, FLOAT_TYPE(data_b[b_offset + y2_idx + 33]) * sc7)))))));
+
+        tmp[16 * ix + tid] += FLOAT_TYPE(dall * (sx * FLOAT_TYPE(data_a[ib0 + i].scales[v_im] & 0x3f) + sy * FLOAT_TYPE(data_a[ib0 + i].scales[v_im + 1] & 0x3f) +
+                        sz * FLOAT_TYPE((data_a[ib0 + i].scales[v_im + 4] & 0x0f) | ((data_a[ib0 + i].scales[v_im] & 0xc0) >> 2)) + sw * FLOAT_TYPE((data_a[ib0 + i].scales[v_im + 5] & 0x0f) | ((data_a[ib0 + i].scales[v_im + 1] & 0xc0) >> 2))) - dmin * smin);
+        const uint tmp_idx = 16 * ix + tid;
+        tmp[tmp_idx] = fma(dall, (fma(sx, FLOAT_TYPE(data_a[ib0 + i].scales[v_im] & 0x3f), fma(sy, FLOAT_TYPE(data_a[ib0 + i].scales[v_im + 1] & 0x3f),
+                       fma(sz, FLOAT_TYPE((data_a[ib0 + i].scales[v_im + 4] & 0x0f) | ((data_a[ib0 + i].scales[v_im] & 0xc0) >> 2)), fma(sw, FLOAT_TYPE((data_a[ib0 + i].scales[v_im + 5] & 0x0f) | ((data_a[ib0 + i].scales[v_im + 1] & 0xc0) >> 2))))))), fma(-dmin, smin, tmp[tmp_idx]));
+#endif
+    }
+
+    // sum up partial sums and write back result
+    barrier();
+    [[unroll]] for (uint s = 16; s > 0; s >>= 1) {
+        if (tid < s) {
+            tmp[tid] += tmp[tid + s];
+        }
+        barrier();
+    }
+    if (tid == 0) {
+        data_d[d_offset + row] = D_TYPE(tmp[0]);
+    }
+}
diff --git a/src/whisper_cpp/ggml/src/vulkan-shaders/mul_mat_vec_q5_k.comp b/src/whisper_cpp/ggml/src/vulkan-shaders/mul_mat_vec_q5_k.comp
new file mode 100644
index 00000000..2306785a
--- /dev/null
+++ b/src/whisper_cpp/ggml/src/vulkan-shaders/mul_mat_vec_q5_k.comp
@@ -0,0 +1,109 @@
+#version 450
+
+#include "mul_mat_vec_base.comp"
+
+layout(local_size_x = 32, local_size_y = 1, local_size_z = 1) in;
+
+shared FLOAT_TYPE tmp[32];
+
+void main() {
+    const uint row = gl_WorkGroupID.x + gl_NumWorkGroups.x * gl_WorkGroupID.z;
+
+    uint a_offset, b_offset, d_offset;
+    get_offsets(a_offset, b_offset, d_offset);
+
+    const uint num_blocks_per_row = p.ncols / QUANT_K;
+    const uint ib0 = a_offset / QUANT_K + row*num_blocks_per_row;
+
+    const uint tid = gl_LocalInvocationID.x/2;  // 0...31 or 0...16
+    const uint ix  = gl_LocalInvocationID.x%2;  // 0 or 0, 1
+
+    const uint il = tid/4;                           // 0...3
+    const uint ir = tid - 4*il;                      // 0...7 or 0...3
+
+    const uint v_im = il / 2;  // 0 or 1. 0 computes 0,32 + 128,160, 1 computes 64,96 + 192,224
+    const uint v_in = il % 2;
+
+    const uint l0 = 4*ir + 2*v_in;                   // 0...15
+    const uint q_offset = 32*v_im + l0;
+    const uint y_offset = 64*v_im + l0;
+
+    const uint8_t hm1 = uint8_t(1 << (2*v_im));
+    const uint8_t hm2 = uint8_t(hm1 << 4);
+
+    tmp[16 * ix + tid] = FLOAT_TYPE(0.0); // partial sum for thread in warp
+
+    [[unroll]] for (uint i = ix; i < num_blocks_per_row; i += 2) {
+        const uint y1_idx = i * QUANT_K + y_offset;
+        const uint y2_idx = y1_idx + 128;
+
+        const FLOAT_TYPE dall = FLOAT_TYPE(data_a[ib0 + i].d.x);
+        const FLOAT_TYPE dmin = FLOAT_TYPE(data_a[ib0 + i].d.y);
+
+        const uint8_t sc0 = uint8_t(  data_a[ib0 + i].scales[v_im * 2    ]       & 0x3f);
+        const uint8_t sc1 = uint8_t(  data_a[ib0 + i].scales[v_im * 2 + 1]       & 0x3f);
+        const uint8_t sc2 = uint8_t(  data_a[ib0 + i].scales[v_im * 2 + 4]       & 0x3f);
+        const uint8_t sc3 = uint8_t(  data_a[ib0 + i].scales[v_im * 2 + 5]       & 0x3f);
+        const uint8_t sc4 = uint8_t(( data_a[ib0 + i].scales[v_im * 2 + 8]       & 0x0f) | ((data_a[ib0 + i].scales[v_im * 2    ] & 0xc0) >> 2));
+        const uint8_t sc5 = uint8_t(( data_a[ib0 + i].scales[v_im * 2 + 9]       & 0x0f) | ((data_a[ib0 + i].scales[v_im * 2 + 1] & 0xc0) >> 2));
+        const uint8_t sc6 = uint8_t(((data_a[ib0 + i].scales[v_im * 2 + 8] >> 4) & 0x0f) | ((data_a[ib0 + i].scales[v_im * 2 + 4] & 0xc0) >> 2));
+        const uint8_t sc7 = uint8_t(((data_a[ib0 + i].scales[v_im * 2 + 9] >> 4) & 0x0f) | ((data_a[ib0 + i].scales[v_im * 2 + 5] & 0xc0) >> 2));
+
+        const uint8_t q4_0  = uint8_t(data_a[ib0 + i].qs[q_offset     ] & 0xf);
+        const uint8_t q4_1  = uint8_t(data_a[ib0 + i].qs[q_offset +  1] & 0xf);
+        const uint8_t q4_2  = uint8_t(data_a[ib0 + i].qs[q_offset + 16] & 0xf);
+        const uint8_t q4_3  = uint8_t(data_a[ib0 + i].qs[q_offset + 17] & 0xf);
+        const uint8_t q4_4  = uint8_t(data_a[ib0 + i].qs[q_offset     ]  >> 4);
+        const uint8_t q4_5  = uint8_t(data_a[ib0 + i].qs[q_offset +  1]  >> 4);
+        const uint8_t q4_6  = uint8_t(data_a[ib0 + i].qs[q_offset + 16]  >> 4);
+        const uint8_t q4_7  = uint8_t(data_a[ib0 + i].qs[q_offset + 17]  >> 4);
+        const uint8_t q4_8  = uint8_t(data_a[ib0 + i].qs[q_offset + 64] & 0xf);
+        const uint8_t q4_9  = uint8_t(data_a[ib0 + i].qs[q_offset + 65] & 0xf);
+        const uint8_t q4_10 = uint8_t(data_a[ib0 + i].qs[q_offset + 80] & 0xf);
+        const uint8_t q4_11 = uint8_t(data_a[ib0 + i].qs[q_offset + 81] & 0xf);
+        const uint8_t q4_12 = uint8_t(data_a[ib0 + i].qs[q_offset + 64]  >> 4);
+        const uint8_t q4_13 = uint8_t(data_a[ib0 + i].qs[q_offset + 65]  >> 4);
+        const uint8_t q4_14 = uint8_t(data_a[ib0 + i].qs[q_offset + 80]  >> 4);
+        const uint8_t q4_15 = uint8_t(data_a[ib0 + i].qs[q_offset + 81]  >> 4);
+
+        const FLOAT_TYPE sx =
+          fma(FLOAT_TYPE(data_b[b_offset + y1_idx     ]), (q4_0 + (((data_a[ib0 + i].qh[l0     ] & hm1) != 0) ? 16 : 0)),
+          fma(FLOAT_TYPE(data_b[b_offset + y1_idx +  1]), (q4_1 + (((data_a[ib0 + i].qh[l0 +  1] & hm1) != 0) ? 16 : 0)),
+          fma(FLOAT_TYPE(data_b[b_offset + y1_idx + 16]), (q4_2 + (((data_a[ib0 + i].qh[l0 + 16] & hm1) != 0) ? 16 : 0)),
+             FLOAT_TYPE(data_b[b_offset + y1_idx + 17]) * (q4_3 + (((data_a[ib0 + i].qh[l0 + 17] & hm1) != 0) ? 16 : 0)))));
+        const FLOAT_TYPE sy =
+          fma(FLOAT_TYPE(data_b[b_offset + y1_idx + 32]), (q4_4 + (((data_a[ib0 + i].qh[l0     ] & (hm1 << 1)) != 0) ? 16 : 0)),
+          fma(FLOAT_TYPE(data_b[b_offset + y1_idx + 33]), (q4_5 + (((data_a[ib0 + i].qh[l0 +  1] & (hm1 << 1)) != 0) ? 16 : 0)),
+          fma(FLOAT_TYPE(data_b[b_offset + y1_idx + 48]), (q4_6 + (((data_a[ib0 + i].qh[l0 + 16] & (hm1 << 1)) != 0) ? 16 : 0)),
+             FLOAT_TYPE(data_b[b_offset + y1_idx + 49]) * (q4_7 + (((data_a[ib0 + i].qh[l0 + 17] & (hm1 << 1)) != 0) ? 16 : 0)))));
+        const FLOAT_TYPE sz =
+          fma(FLOAT_TYPE(data_b[b_offset + y2_idx     ]), (q4_8  + (((data_a[ib0 + i].qh[l0     ] & hm2) != 0) ? 16 : 0)),
+          fma(FLOAT_TYPE(data_b[b_offset + y2_idx +  1]), (q4_9  + (((data_a[ib0 + i].qh[l0 +  1] & hm2) != 0) ? 16 : 0)),
+          fma(FLOAT_TYPE(data_b[b_offset + y2_idx + 16]), (q4_10 + (((data_a[ib0 + i].qh[l0 + 16] & hm2) != 0) ? 16 : 0)),
+             FLOAT_TYPE(data_b[b_offset + y2_idx + 17]) * (q4_11 + (((data_a[ib0 + i].qh[l0 + 17] & hm2) != 0) ? 16 : 0)))));
+        const FLOAT_TYPE sw =
+          fma(FLOAT_TYPE(data_b[b_offset + y2_idx + 32]), (q4_12 + (((data_a[ib0 + i].qh[l0     ] & (hm2 << 1)) != 0) ? 16 : 0)),
+          fma(FLOAT_TYPE(data_b[b_offset + y2_idx + 33]), (q4_13 + (((data_a[ib0 + i].qh[l0 +  1] & (hm2 << 1)) != 0) ? 16 : 0)),
+          fma(FLOAT_TYPE(data_b[b_offset + y2_idx + 48]), (q4_14 + (((data_a[ib0 + i].qh[l0 + 16] & (hm2 << 1)) != 0) ? 16 : 0)),
+             FLOAT_TYPE(data_b[b_offset + y2_idx + 49]) * (q4_15 + (((data_a[ib0 + i].qh[l0 + 17] & (hm2 << 1)) != 0) ? 16 : 0)))));
+        const FLOAT_TYPE smin =
+          fma(FLOAT_TYPE(data_b[b_offset + y1_idx     ]) + FLOAT_TYPE(data_b[b_offset + y1_idx + 1 ]) + FLOAT_TYPE(data_b[b_offset + y1_idx + 16]) + FLOAT_TYPE(data_b[b_offset + y1_idx + 17]), sc2,
+          fma(FLOAT_TYPE(data_b[b_offset + y1_idx + 32]) + FLOAT_TYPE(data_b[b_offset + y1_idx + 33]) + FLOAT_TYPE(data_b[b_offset + y1_idx + 48]) + FLOAT_TYPE(data_b[b_offset + y1_idx + 49]), sc3,
+          fma(FLOAT_TYPE(data_b[b_offset + y2_idx     ]) + FLOAT_TYPE(data_b[b_offset + y2_idx + 1 ]) + FLOAT_TYPE(data_b[b_offset + y2_idx + 16]) + FLOAT_TYPE(data_b[b_offset + y2_idx + 17]), sc6,
+              (FLOAT_TYPE(data_b[b_offset + y2_idx + 32]) + FLOAT_TYPE(data_b[b_offset + y2_idx + 33]) + FLOAT_TYPE(data_b[b_offset + y2_idx + 48]) + FLOAT_TYPE(data_b[b_offset + y2_idx + 49])) * sc7)));
+        const uint tmp_idx = 16 * ix + tid;
+        tmp[tmp_idx] = fma(dall, fma(sx, sc0, fma(sy, sc1, fma(sz, sc4, sw * sc5))), fma(-dmin, smin, tmp[tmp_idx]));
+    }
+
+    // sum up partial sums and write back result
+    barrier();
+    [[unroll]] for (uint s = 16; s > 0; s >>= 1) {
+        if (tid < s) {
+            tmp[tid] += tmp[tid + s];
+        }
+        barrier();
+    }
+    if (tid == 0) {
+        data_d[d_offset + row] = D_TYPE(tmp[0]);
+    }
+}
diff --git a/src/whisper_cpp/ggml/src/vulkan-shaders/mul_mat_vec_q6_k.comp b/src/whisper_cpp/ggml/src/vulkan-shaders/mul_mat_vec_q6_k.comp
new file mode 100644
index 00000000..95c286ee
--- /dev/null
+++ b/src/whisper_cpp/ggml/src/vulkan-shaders/mul_mat_vec_q6_k.comp
@@ -0,0 +1,79 @@
+#version 450
+
+#include "mul_mat_vec_base.comp"
+
+layout(local_size_x = 32, local_size_y = 1, local_size_z = 1) in;
+
+shared FLOAT_TYPE tmp[32];
+
+void main() {
+    const uint row = gl_WorkGroupID.x + gl_NumWorkGroups.x * gl_WorkGroupID.z;
+
+    uint a_offset, b_offset, d_offset;
+    get_offsets(a_offset, b_offset, d_offset);
+
+    const uint num_blocks_per_row = p.ncols / QUANT_K;
+    const uint ib0 = a_offset / QUANT_K + row*num_blocks_per_row;
+
+    const uint tid = gl_LocalInvocationID.x/K_QUANTS_PER_ITERATION;  // 0...31 or 0...16
+    const uint ix  = gl_LocalInvocationID.x%K_QUANTS_PER_ITERATION;  // 0 or 0, 1
+
+    const uint step = 16/K_QUANTS_PER_ITERATION;            // 16 or 8
+
+    const uint v_im = tid/step;                             // 0 or 1. 0 computes 0..., 1 computes 128...
+    const uint v_in = tid - step*v_im;                      // 0...15 or 0...7
+
+#if K_QUANTS_PER_ITERATION == 1
+    const uint l0 = v_in;                                   // 0...15
+    const uint is = 0;
+#else
+    const uint l0 = 4 * v_in;                               // 0, 4, 8, ..., 28
+    const uint is = v_in / 4;
+#endif
+
+    const uint ql_offset = 64*v_im + l0;
+    const uint qh_offset = 32*v_im + l0;
+    const uint s_offset  =  8*v_im + is;
+    const uint y_offset = 128*v_im + l0;
+
+    tmp[16 * ix + tid] = FLOAT_TYPE(0.0); // partial sum for thread in warp
+
+    [[unroll]] for (uint i = ix; i < num_blocks_per_row; i += K_QUANTS_PER_ITERATION) {
+        const uint y_idx   = i * QUANT_K + y_offset;
+
+        const FLOAT_TYPE d = FLOAT_TYPE(data_a[ib0 + i].d);
+
+#if K_QUANTS_PER_ITERATION == 1
+        const uint tmp_idx = 16 * ix + tid;
+        tmp[tmp_idx] = fma(FLOAT_TYPE(data_b[b_offset + y_idx +  0]) * FLOAT_TYPE(data_a[ib0 + i].scales[s_offset + 0]) * d, FLOAT_TYPE(int8_t((data_a[ib0 + i].ql[ql_offset +  0] & 0xF) | ((data_a[ib0 + i].qh[qh_offset +  0] & 0x03) << 4)) - 32),
+                       fma(FLOAT_TYPE(data_b[b_offset + y_idx + 16]) * FLOAT_TYPE(data_a[ib0 + i].scales[s_offset + 1]) * d, FLOAT_TYPE(int8_t((data_a[ib0 + i].ql[ql_offset + 16] & 0xF) | ((data_a[ib0 + i].qh[qh_offset + 16] & 0x03) << 4)) - 32),
+                       fma(FLOAT_TYPE(data_b[b_offset + y_idx + 32]) * FLOAT_TYPE(data_a[ib0 + i].scales[s_offset + 2]) * d, FLOAT_TYPE(int8_t((data_a[ib0 + i].ql[ql_offset + 32] & 0xF) | ((data_a[ib0 + i].qh[qh_offset +  0] & 0x0c) << 2)) - 32),
+                       fma(FLOAT_TYPE(data_b[b_offset + y_idx + 48]) * FLOAT_TYPE(data_a[ib0 + i].scales[s_offset + 3]) * d, FLOAT_TYPE(int8_t((data_a[ib0 + i].ql[ql_offset + 48] & 0xF) | ((data_a[ib0 + i].qh[qh_offset + 16] & 0x0c) << 2)) - 32),
+                       fma(FLOAT_TYPE(data_b[b_offset + y_idx + 64]) * FLOAT_TYPE(data_a[ib0 + i].scales[s_offset + 4]) * d, FLOAT_TYPE(int8_t((data_a[ib0 + i].ql[ql_offset +  0]  >> 4) | ((data_a[ib0 + i].qh[qh_offset +  0] & 0x30) >> 0)) - 32),
+                       fma(FLOAT_TYPE(data_b[b_offset + y_idx + 80]) * FLOAT_TYPE(data_a[ib0 + i].scales[s_offset + 5]) * d, FLOAT_TYPE(int8_t((data_a[ib0 + i].ql[ql_offset + 16]  >> 4) | ((data_a[ib0 + i].qh[qh_offset + 16] & 0x30) >> 0)) - 32),
+                       fma(FLOAT_TYPE(data_b[b_offset + y_idx + 96]) * FLOAT_TYPE(data_a[ib0 + i].scales[s_offset + 6]) * d, FLOAT_TYPE(int8_t((data_a[ib0 + i].ql[ql_offset + 32]  >> 4) | ((data_a[ib0 + i].qh[qh_offset +  0] & 0xc0) >> 2)) - 32),
+                       fma(FLOAT_TYPE(data_b[b_offset + y_idx +112]) * FLOAT_TYPE(data_a[ib0 + i].scales[s_offset + 7]) * d, FLOAT_TYPE(int8_t((data_a[ib0 + i].ql[ql_offset + 48]  >> 4) | ((data_a[ib0 + i].qh[qh_offset + 16] & 0xc0) >> 2)) - 32), tmp[tmp_idx]))))))));
+#else
+        FLOAT_TYPE sum = FLOAT_TYPE(0.0);
+        [[unroll]] for (int l = 0; l < 4; ++l) {
+            sum = fma(FLOAT_TYPE(data_b[b_offset + y_idx + l+ 0]) * FLOAT_TYPE(data_a[ib0 + i].scales[s_offset + 0]) * d, FLOAT_TYPE(int8_t((data_a[ib0 + i].ql[ql_offset + l+ 0] & 0xF) | (((data_a[ib0 + i].qh[qh_offset + l] >> 0) & 3) << 4)) - 32),
+                  fma(FLOAT_TYPE(data_b[b_offset + y_idx + l+32]) * FLOAT_TYPE(data_a[ib0 + i].scales[s_offset + 2]) * d, FLOAT_TYPE(int8_t((data_a[ib0 + i].ql[ql_offset + l+32] & 0xF) | (((data_a[ib0 + i].qh[qh_offset + l] >> 2) & 3) << 4)) - 32),
+                  fma(FLOAT_TYPE(data_b[b_offset + y_idx + l+64]) * FLOAT_TYPE(data_a[ib0 + i].scales[s_offset + 4]) * d, FLOAT_TYPE(int8_t((data_a[ib0 + i].ql[ql_offset + l+ 0]  >> 4) | (((data_a[ib0 + i].qh[qh_offset + l] >> 4) & 3) << 4)) - 32),
+                  fma(FLOAT_TYPE(data_b[b_offset + y_idx + l+96]) * FLOAT_TYPE(data_a[ib0 + i].scales[s_offset + 6]) * d, FLOAT_TYPE(int8_t((data_a[ib0 + i].ql[ql_offset + l+32]  >> 4) | (((data_a[ib0 + i].qh[qh_offset + l] >> 6) & 3) << 4)) - 32), sum))));
+        }
+        tmp[16 * ix + tid] += sum;
+#endif
+    }
+
+    // sum up partial sums and write back result
+    barrier();
+    [[unroll]] for (uint s = 16; s > 0; s >>= 1) {
+        if (tid < s) {
+            tmp[tid] += tmp[tid + s];
+       }
+        barrier();
+    }
+    if (tid == 0) {
+        data_d[d_offset + row] = D_TYPE(tmp[0]);
+    }
+}
diff --git a/src/whisper_cpp/ggml/src/vulkan-shaders/mul_mm.comp b/src/whisper_cpp/ggml/src/vulkan-shaders/mul_mm.comp
new file mode 100644
index 00000000..fffdd181
--- /dev/null
+++ b/src/whisper_cpp/ggml/src/vulkan-shaders/mul_mm.comp
@@ -0,0 +1,508 @@
+#version 450
+
+#extension GL_EXT_control_flow_attributes : enable
+#extension GL_EXT_shader_16bit_storage : require
+
+#ifdef FLOAT16
+#extension GL_EXT_shader_explicit_arithmetic_types_float16 : require
+#endif
+
+#ifdef MUL_MAT_ID
+#extension GL_EXT_shader_explicit_arithmetic_types_int16 : require
+#endif
+
+#include "types.comp"
+
+#ifndef LOAD_VEC_A
+#define LOAD_VEC_A 1
+#endif
+#ifndef LOAD_VEC_B
+#define LOAD_VEC_B 1
+#endif
+
+layout(local_size_x_id = 0, local_size_y = 1, local_size_z = 1) in;
+
+layout (binding = 0) readonly buffer A {A_TYPE data_a[];};
+layout (binding = 1) readonly buffer B {B_TYPE data_b[];};
+layout (binding = 2) writeonly buffer D {D_TYPE data_d[];};
+
+#ifdef MUL_MAT_ID
+layout (binding = 3) readonly buffer IDS {int data_ids[];};
+#endif
+
+layout (push_constant) uniform parameter
+{
+    uint M;
+    uint N;
+    uint K;
+    uint stride_a;
+    uint stride_b;
+    uint stride_d;
+
+    uint batch_stride_a;
+    uint batch_stride_b;
+    uint batch_stride_d;
+
+#ifdef MUL_MAT_ID
+    uint nei0;
+    uint nei1;
+    uint nbi1;
+    uint ne11;
+#else
+    uint k_split;
+    uint ne02;
+    uint ne12;
+    uint broadcast2;
+    uint broadcast3;
+#endif
+} p;
+
+layout (constant_id = 1) const uint BM = 64;
+layout (constant_id = 2) const uint BN = 64;
+layout (constant_id = 3) const uint BK = 16;  // Assumed to be 32 if working with a quant
+layout (constant_id = 4) const uint WM = 32;
+layout (constant_id = 5) const uint WN = 32;
+layout (constant_id = 6) const uint WMITER = 2;
+layout (constant_id = 7) const uint TM = 4;
+layout (constant_id = 8) const uint TN = 2;
+layout (constant_id = 9) const uint WARP = 32;
+
+shared FLOAT_TYPE buf_a[BM * (BK+1)];
+shared FLOAT_TYPE buf_b[BN * (BK+1)];
+
+#ifdef MUL_MAT_ID
+shared u16vec2 row_ids[3072];
+#endif
+
+void main() {
+#ifdef MUL_MAT_ID
+    const uint expert_idx = gl_GlobalInvocationID.z;
+#else
+    const uint batch_idx = gl_GlobalInvocationID.z;
+
+    const uint i13 = batch_idx / p.ne12;
+    const uint i12 = batch_idx % p.ne12;
+
+    const uint i03 = i13 / p.broadcast3;
+    const uint i02 = i12 / p.broadcast2;
+
+    const uint batch_idx_a = i03 * p.ne02 + i02;
+#endif
+
+    const uint blocks_m = (p.M + BM - 1) / BM;
+    const uint ir = gl_WorkGroupID.x % blocks_m;
+    const uint ik = gl_WorkGroupID.x / blocks_m;
+    const uint ic = gl_WorkGroupID.y;
+
+    const uint warp_i = gl_LocalInvocationID.x / WARP;
+    const uint warp_r = warp_i % (BM / WM);
+    const uint warp_c = warp_i / (BM / WM);
+
+    const uint WNITER = (WM * WN) / (WARP * TM * TN * WMITER);
+    const uint WSUBM = WM / WMITER;
+    const uint WSUBN = WN / WNITER;
+
+    const uint tiw = gl_LocalInvocationID.x % WARP;
+    const uint tiwr = tiw % (WSUBM / TM);
+    const uint tiwc = tiw / (WSUBM / TM);
+
+    const uint loadr_a = gl_LocalInvocationID.x % (BK / LOAD_VEC_A);
+    const uint loadc_a = gl_LocalInvocationID.x / (BK / LOAD_VEC_A);
+    const uint loadr_b = gl_LocalInvocationID.x % (BK / LOAD_VEC_B);
+    const uint loadc_b = gl_LocalInvocationID.x / (BK / LOAD_VEC_B);
+
+    const uint loadstride_a = gl_WorkGroupSize.x * LOAD_VEC_A / BK;
+    const uint loadstride_b = gl_WorkGroupSize.x * LOAD_VEC_B / BK;
+
+#ifdef MUL_MAT_ID
+    uint _ne1 = 0;
+    for (uint ii1 = 0; ii1 < p.nei1; ii1++) {
+        for (uint ii0 = 0; ii0 < p.nei0; ii0++) {
+            if (data_ids[ii1*p.nbi1 + ii0] == expert_idx) {
+                row_ids[_ne1] = u16vec2(ii0, ii1);
+                _ne1++;
+            }
+        }
+    }
+
+    barrier();
+
+    // Workgroup has no work
+    if (ic * BN >= _ne1) return;
+#endif
+
+#ifdef MUL_MAT_ID
+    const uint start_k = 0;
+    const uint end_k = p.K;
+#else
+    const uint start_k = ik * p.k_split;
+    const uint end_k = min(p.K, (ik + 1) * p.k_split);
+#endif
+
+    uint pos_a = (
+#ifdef MUL_MAT_ID
+        expert_idx * p.batch_stride_a +
+#else
+        batch_idx_a * p.batch_stride_a +
+#endif
+        ir * BM * p.stride_a + start_k) / LOAD_VEC_A;
+#ifdef MUL_MAT_ID
+    uint pos_b = 0;
+#else
+    uint pos_b = (batch_idx * p.batch_stride_b + ic * BN * p.stride_b + start_k) / LOAD_VEC_B;
+#endif
+
+    float sums[WMITER * TM * WNITER * TN];
+    FLOAT_TYPE cache_a[WMITER * TM];
+    FLOAT_TYPE cache_b[WNITER * TN];
+
+    [[unroll]] for (uint i = 0; i < WMITER*TM*WNITER*TN; i++) {
+        sums[i] = 0.0f;
+    }
+
+    [[unroll]] for (uint block = start_k; block < end_k; block += BK) {
+        [[unroll]] for (uint l = 0; l < BM; l += loadstride_a) {
+
+#if defined(DATA_A_F32) || defined(DATA_A_F16)
+#if LOAD_VEC_A == 8
+            const uint idx = pos_a + (loadc_a + l) * p.stride_a / LOAD_VEC_A + loadr_a;
+            const uint buf_idx = (loadc_a + l) * (BK+1) + loadr_a * LOAD_VEC_A;
+            buf_a[buf_idx    ] = FLOAT_TYPE(data_a[idx][0].x);
+            buf_a[buf_idx + 1] = FLOAT_TYPE(data_a[idx][0].y);
+            buf_a[buf_idx + 2] = FLOAT_TYPE(data_a[idx][0].z);
+            buf_a[buf_idx + 3] = FLOAT_TYPE(data_a[idx][0].w);
+            buf_a[buf_idx + 4] = FLOAT_TYPE(data_a[idx][1].x);
+            buf_a[buf_idx + 5] = FLOAT_TYPE(data_a[idx][1].y);
+            buf_a[buf_idx + 6] = FLOAT_TYPE(data_a[idx][1].z);
+            buf_a[buf_idx + 7] = FLOAT_TYPE(data_a[idx][1].w);
+#elif LOAD_VEC_A == 4
+            const uint idx = pos_a + (loadc_a + l) * p.stride_a / LOAD_VEC_A + loadr_a;
+            const uint buf_idx = (loadc_a + l) * (BK+1) + loadr_a * LOAD_VEC_A;
+            buf_a[buf_idx    ] = FLOAT_TYPE(data_a[idx].x);
+            buf_a[buf_idx + 1] = FLOAT_TYPE(data_a[idx].y);
+            buf_a[buf_idx + 2] = FLOAT_TYPE(data_a[idx].z);
+            buf_a[buf_idx + 3] = FLOAT_TYPE(data_a[idx].w);
+#else
+            if (ir * BM + loadc_a + l < p.M && block + loadr_a < end_k) {
+                buf_a[(loadc_a + l) * (BK+1) + loadr_a] = FLOAT_TYPE(data_a[pos_a + (loadc_a + l) * p.stride_a + loadr_a]);
+            } else {
+                buf_a[(loadc_a + l) * (BK+1) + loadr_a] = FLOAT_TYPE(0.0f);
+            }
+#endif
+#elif defined(DATA_A_Q4_0)
+            const uint idx = pos_a + (loadc_a + l) * p.stride_a / LOAD_VEC_A + loadr_a;
+            const uint buf_idx = (loadc_a + l) * (BK+1) + loadr_a;
+
+            const uint ib = idx / 16;
+            const uint iqs = idx & 0xF;
+
+            const float d = float(data_a[ib].d);
+            const uint vui = uint(data_a[ib].qs[iqs]);
+            const vec2 v = (vec2(vui & 0xF, vui >> 4) - 8.0f) * d;
+
+            buf_a[buf_idx     ] = FLOAT_TYPE(v.x);
+            buf_a[buf_idx + 16] = FLOAT_TYPE(v.y);
+#elif defined(DATA_A_Q4_1)
+            const uint idx = pos_a + (loadc_a + l) * p.stride_a / LOAD_VEC_A + loadr_a;
+            const uint buf_idx = (loadc_a + l) * (BK+1) + loadr_a;
+
+            const uint ib = idx / 16;
+            const uint iqs = idx & 0xF;
+
+            const float d = float(data_a[ib].d);
+            const float m = float(data_a[ib].m);
+            const uint vui = uint(data_a[ib].qs[iqs]);
+            const vec2 v = vec2(vui & 0xF, vui >> 4) * d + m;
+
+            buf_a[buf_idx     ] = FLOAT_TYPE(v.x);
+            buf_a[buf_idx + 16] = FLOAT_TYPE(v.y);
+#elif defined(DATA_A_Q5_0)
+            const uint idx = pos_a + (loadc_a + l) * p.stride_a / LOAD_VEC_A + loadr_a;
+            const uint buf_idx = (loadc_a + l) * (BK+1) + loadr_a;
+
+            const uint ib = idx / 16;
+            const uint iqs = idx & 0xF;
+
+            const float d = float(data_a[ib].d);
+            const uint uint_qh = uint(data_a[ib].qh[1]) << 16 | data_a[ib].qh[0];
+            const ivec2 qh = ivec2(((uint_qh >> iqs) << 4) & 0x10, (uint_qh >> (iqs + 12)) & 0x10);
+            const uint vui = uint(data_a[ib].qs[iqs]);
+            const vec2 v = (vec2((vui & 0xF) | qh.x, (vui >> 4) | qh.y) - 16.0f) * d;
+
+            buf_a[buf_idx     ] = FLOAT_TYPE(v.x);
+            buf_a[buf_idx + 16] = FLOAT_TYPE(v.y);
+#elif defined(DATA_A_Q5_1)
+            const uint idx = pos_a + (loadc_a + l) * p.stride_a / LOAD_VEC_A + loadr_a;
+            const uint buf_idx = (loadc_a + l) * (BK+1) + loadr_a;
+
+            const uint ib = idx / 16;
+            const uint iqs = idx & 0xF;
+
+            const float d = float(data_a[ib].d);
+            const float m = float(data_a[ib].m);
+            const uint uint_qh = data_a[ib].qh;
+            const ivec2 qh = ivec2(((uint_qh >> iqs) << 4) & 0x10, (uint_qh >> (iqs + 12)) & 0x10);
+            const uint vui = uint(data_a[ib].qs[iqs]);
+            const vec2 v = vec2((vui & 0xF) | qh.x, (vui >> 4) | qh.y) * d + m;
+
+            buf_a[buf_idx     ] = FLOAT_TYPE(v.x);
+            buf_a[buf_idx + 16] = FLOAT_TYPE(v.y);
+#elif defined(DATA_A_Q8_0)
+            const uint idx = pos_a + (loadc_a + l) * p.stride_a / LOAD_VEC_A + loadr_a;
+            const uint buf_idx = (loadc_a + l) * (BK+1) + loadr_a * LOAD_VEC_A;
+
+            const uint ib = idx / 16;
+            const uint iqs = (idx & 0xF) * 2;
+
+            const float d = float(data_a[ib].d);
+            const vec2 v = vec2(int(data_a[ib].qs[iqs]), int(data_a[ib].qs[iqs + 1])) * d;
+
+            buf_a[buf_idx    ] = FLOAT_TYPE(v.x);
+            buf_a[buf_idx + 1] = FLOAT_TYPE(v.y);
+#elif defined(DATA_A_Q2_K)
+            const uint idx = pos_a + (loadc_a + l) * p.stride_a / LOAD_VEC_A + loadr_a;
+            const uint buf_idx = (loadc_a + l) * (BK+1) + loadr_a * LOAD_VEC_A;
+
+            const uint ib = idx / 128;                         // 2 values per idx
+            const uint iqs = idx % 128;                        // 0..127
+
+            const uint qsi = (iqs / 64) * 32 + (iqs % 16) * 2; // 0,2,4..30
+            const uint scalesi = iqs / 8;                      // 0..15
+            const uint qsshift = ((iqs % 64) / 16) * 2;        // 0,2,4,6
+
+            const uvec2 qs = uvec2(data_a[ib].qs[qsi], data_a[ib].qs[qsi + 1]);
+            const uint scales = data_a[ib].scales[scalesi];
+            const vec2 d = vec2(data_a[ib].d);
+
+            const vec2 v = d.x * float(scales & 0xF) * vec2((qs >> qsshift) & 3) - d.y * float(scales >> 4);
+
+            buf_a[buf_idx    ] = FLOAT_TYPE(v.x);
+            buf_a[buf_idx + 1] = FLOAT_TYPE(v.y);
+#elif defined(DATA_A_Q3_K)
+            const uint idx = pos_a + (loadc_a + l) * p.stride_a / LOAD_VEC_A + loadr_a;
+            const uint buf_idx = (loadc_a + l) * (BK+1) + loadr_a * LOAD_VEC_A;
+
+            const uint ib = idx / 128;                   // 2 values per idx
+            const uint iqs = idx % 128;                  // 0..127
+
+            const uint n = iqs / 64;                     // 0,1
+            const uint qsi = n * 32 + (iqs % 16) * 2;    // 0,2,4..62
+            const uint hmi =          (iqs % 16) * 2;    // 0,2,4..30
+            const uint j = (iqs % 64) / 4;               // 0..3
+            const uint is = iqs / 8;                     // 0..15
+            const uint halfsplit = ((iqs % 64) / 16);    // 0,1,2,3
+            const uint qsshift = halfsplit * 2;          // 0,2,4,6
+            const uint m = 1 << (4 * n + halfsplit);     // 1,2,4,8,16,32,64,128
+
+            const int8_t us = int8_t(is <  4 ? (data_a[ib].scales[is-0] & 0xF) | (((data_a[ib].scales[is+8] >> 0) & 3) << 4) :
+                                    is <  8 ? (data_a[ib].scales[is-0] & 0xF) | (((data_a[ib].scales[is+4] >> 2) & 3) << 4) :
+                                    is < 12 ? (data_a[ib].scales[is-8] >>  4) | (((data_a[ib].scales[is+0] >> 4) & 3) << 4) :
+                                            (data_a[ib].scales[is-8] >>  4) | (((data_a[ib].scales[is-4] >> 6) & 3) << 4));
+            const float dl = float(data_a[ib].d) * float(us - 32);
+
+            buf_a[buf_idx    ] = FLOAT_TYPE(dl * float(int8_t((data_a[ib].qs[qsi    ] >> qsshift) & 3) - (((data_a[ib].hmask[hmi    ] & m) != 0) ? 0 : 4)));
+            buf_a[buf_idx + 1] = FLOAT_TYPE(dl * float(int8_t((data_a[ib].qs[qsi + 1] >> qsshift) & 3) - (((data_a[ib].hmask[hmi + 1] & m) != 0) ? 0 : 4)));
+#elif defined(DATA_A_Q4_K)
+            const uint idx = pos_a + (loadc_a + l) * p.stride_a / LOAD_VEC_A + loadr_a;
+            const uint buf_idx = (loadc_a + l) * (BK+1) + loadr_a * LOAD_VEC_A;
+
+            const uint ib = idx / 128;                 // 2 values per idx
+            const uint iqs = idx % 128;                // 0..127
+
+            const uint n = iqs / 32;                   // 0,1,2,3
+            const uint b = (iqs % 32) / 16;            // 0,1
+            const uint is = 2 * n + b;                 // 0..7
+            const uint qsi = n * 32 + (iqs % 16) * 2;  // 0,2,4..126
+
+            const vec2 loadd = vec2(data_a[ib].d);
+
+            uint8_t sc;
+            uint8_t mbyte;
+            if (is < 4) {
+                sc    = uint8_t(data_a[ib].scales[is    ] & 63);
+                mbyte = uint8_t(data_a[ib].scales[is + 4] & 63);
+            } else {
+                sc    = uint8_t((data_a[ib].scales[is + 4] & 0xF) | ((data_a[ib].scales[is - 4] >> 6) << 4));
+                mbyte = uint8_t((data_a[ib].scales[is + 4] >>  4) | ((data_a[ib].scales[is    ] >> 6) << 4));
+            }
+            const float d = loadd.x * sc;
+            const float m = -loadd.y * mbyte;
+
+            buf_a[buf_idx    ] = FLOAT_TYPE(fma(d, float((data_a[ib].qs[qsi    ] >> (b * 4)) & 0xF), m));
+            buf_a[buf_idx + 1] = FLOAT_TYPE(fma(d, float((data_a[ib].qs[qsi + 1] >> (b * 4)) & 0xF), m));
+#elif defined(DATA_A_Q5_K)
+            const uint idx = pos_a + (loadc_a + l) * p.stride_a / LOAD_VEC_A + loadr_a;
+            const uint buf_idx = (loadc_a + l) * (BK+1) + loadr_a * LOAD_VEC_A;
+
+            const uint ib = idx / 128;                 // 2 values per idx
+            const uint iqs = idx % 128;                // 0..127
+
+            const uint n = iqs / 32;                   // 0,1,2,3
+            const uint b = (iqs % 32) / 16;            // 0,1
+            const uint is = 2 * n + b;                 // 0..7
+            const uint qsi = n * 32 + (iqs % 16) * 2;  // 0,2,4..126
+            const uint qhi = (iqs % 16) * 2;           // 0,2,4..30
+
+            const uint8_t hm = uint8_t(1 << (iqs / 16));
+
+            const vec2 loadd = vec2(data_a[ib].d);
+
+            uint8_t sc;
+            uint8_t mbyte;
+            if (is < 4) {
+                sc    = uint8_t(data_a[ib].scales[is    ] & 63);
+                mbyte = uint8_t(data_a[ib].scales[is + 4] & 63);
+            } else {
+                sc    = uint8_t((data_a[ib].scales[is + 4] & 0xF) | ((data_a[ib].scales[is - 4] >> 6) << 4));
+                mbyte = uint8_t((data_a[ib].scales[is + 4] >>  4) | ((data_a[ib].scales[is    ] >> 6) << 4));
+            }
+            const float d = loadd.x * sc;
+            const float m = -loadd.y * mbyte;
+
+            buf_a[buf_idx    ] = FLOAT_TYPE(fma(d, float((data_a[ib].qs[qsi    ] >> (b * 4)) & 0xF) + float((data_a[ib].qh[qhi    ] & hm) != 0 ? 16 : 0), m));
+            buf_a[buf_idx + 1] = FLOAT_TYPE(fma(d, float((data_a[ib].qs[qsi + 1] >> (b * 4)) & 0xF) + float((data_a[ib].qh[qhi + 1] & hm) != 0 ? 16 : 0), m));
+#elif defined(DATA_A_Q6_K)
+            const uint idx = pos_a + (loadc_a + l) * p.stride_a / LOAD_VEC_A + loadr_a;
+            const uint buf_idx = (loadc_a + l) * (BK+1) + loadr_a * LOAD_VEC_A;
+
+            const uint ib = idx / 128;                  // 2 values per idx
+            const uint iqs = idx % 128;                 // 0..127
+
+            const uint n = iqs / 64;                    // 0,1
+            const uint b = (iqs % 64) / 32;             // 0,1
+            const uint is_b = (iqs % 16) / 8;           // 0,1
+            const uint qhshift = ((iqs % 64) / 16) * 2; // 0,2,4,6
+            const uint is = 8 * n + qhshift + is_b;     // 0..15
+            const uint qsi = n * 64 + (iqs % 32) * 2;   // 0,2,4..126
+            const uint qhi = n * 32 + (iqs % 16) * 2;   // 0,2,4..62
+
+            const float dscale = float(data_a[ib].d) * float(data_a[ib].scales[is]);
+
+            buf_a[buf_idx    ] = FLOAT_TYPE(dscale * float(int8_t(((data_a[ib].ql[qsi    ] >> (b * 4)) & 0xF) | (((data_a[ib].qh[qhi    ] >> qhshift) & 3) << 4)) - 32));
+            buf_a[buf_idx + 1] = FLOAT_TYPE(dscale * float(int8_t(((data_a[ib].ql[qsi + 1] >> (b * 4)) & 0xF) | (((data_a[ib].qh[qhi + 1] >> qhshift) & 3) << 4)) - 32));
+#elif defined(DATA_A_IQ4_NL)
+            const uint idx = pos_a + (loadc_a + l) * p.stride_a / LOAD_VEC_A + loadr_a;
+            const uint buf_idx = (loadc_a + l) * (BK+1) + loadr_a;
+
+            const uint ib = idx / 16;
+            const uint iqs = idx & 0xF;
+
+            const float d = float(data_a[ib].d);
+            const uint vui = uint(data_a[ib].qs[iqs]);
+            const vec2 v = vec2(kvalues_iq4nl[vui & 0xF], kvalues_iq4nl[vui >> 4]) * d;
+
+            buf_a[buf_idx     ] = FLOAT_TYPE(v.x);
+            buf_a[buf_idx + 16] = FLOAT_TYPE(v.y);
+#endif
+        }
+        [[unroll]] for (uint l = 0; l < BN; l += loadstride_b) {
+#if LOAD_VEC_B == 8
+#ifdef MUL_MAT_ID
+            const u16vec2 row_idx = row_ids[ic * BN + loadc_b + l];
+            const uint idx = pos_b + row_idx.y * p.batch_stride_b / LOAD_VEC_B + (row_idx.x % p.ne11) * p.stride_b / LOAD_VEC_B + loadr_b;
+#else
+            const uint idx = pos_b + (loadc_b + l) * p.stride_b / LOAD_VEC_B + loadr_b;
+#endif
+            const uint buf_idx = (loadc_b + l) * (BK+1) + loadr_b * LOAD_VEC_B;
+            buf_b[buf_idx + 0] = FLOAT_TYPE(data_b[idx][0].x);
+            buf_b[buf_idx + 1] = FLOAT_TYPE(data_b[idx][0].y);
+            buf_b[buf_idx + 2] = FLOAT_TYPE(data_b[idx][0].z);
+            buf_b[buf_idx + 3] = FLOAT_TYPE(data_b[idx][0].w);
+            buf_b[buf_idx + 4] = FLOAT_TYPE(data_b[idx][1].x);
+            buf_b[buf_idx + 5] = FLOAT_TYPE(data_b[idx][1].y);
+            buf_b[buf_idx + 6] = FLOAT_TYPE(data_b[idx][1].z);
+            buf_b[buf_idx + 7] = FLOAT_TYPE(data_b[idx][1].w);
+#elif LOAD_VEC_B == 4
+#ifdef MUL_MAT_ID
+            const u16vec2 row_idx = row_ids[ic * BN + loadc_b + l];
+            const uint idx = pos_b + row_idx.y * p.batch_stride_b / LOAD_VEC_B + (row_idx.x % p.ne11) * p.stride_b / LOAD_VEC_B + loadr_b;
+#else
+            const uint idx = pos_b + (loadc_b + l) * p.stride_b / LOAD_VEC_B + loadr_b;
+#endif
+            const uint buf_idx = (loadc_b + l) * (BK+1) + loadr_b * LOAD_VEC_B;
+            buf_b[buf_idx + 0] = FLOAT_TYPE(data_b[idx].x);
+            buf_b[buf_idx + 1] = FLOAT_TYPE(data_b[idx].y);
+            buf_b[buf_idx + 2] = FLOAT_TYPE(data_b[idx].z);
+            buf_b[buf_idx + 3] = FLOAT_TYPE(data_b[idx].w);
+#elif !MUL_MAT_ID
+            if (ic * BN + loadc_b + l < p.N && block + loadr_b < end_k) {
+                buf_b[(loadc_b + l) * (BK+1) + loadr_b] = FLOAT_TYPE(data_b[pos_b + (loadc_b + l) * p.stride_b + loadr_b]);
+            } else {
+                buf_b[(loadc_b + l) * (BK+1) + loadr_b] = FLOAT_TYPE(0.0f);
+            }
+#else
+            const uint row_i = ic * BN + loadc_b + l;
+            if (row_i < _ne1) {
+                const u16vec2 row_idx = row_ids[row_i];
+                buf_b[(loadc_b + l) * (BK+1) + loadr_b] = FLOAT_TYPE(data_b[pos_b + row_idx.y * p.batch_stride_b + (row_idx.x % p.ne11) * p.stride_b + loadr_b]);
+            } else {
+                buf_b[(loadc_b + l) * (BK+1) + loadr_b] = FLOAT_TYPE(0.0f);
+            }
+#endif
+        }
+
+        barrier();
+
+        pos_a += BK / LOAD_VEC_A;
+        pos_b += BK / LOAD_VEC_B;
+
+        for (uint i = 0; i < BK; i++) {
+            // Load from shared into cache
+            [[unroll]] for (uint wsir = 0; wsir < WMITER; wsir++) {
+                [[unroll]] for (uint j = 0; j < TM; j++) {
+                    cache_a[wsir * TM + j] = buf_a[(warp_r * WM + wsir * WSUBM + tiwr * TM + j) * (BK+1) + i];
+                }
+            }
+            [[unroll]] for (uint wsic = 0; wsic < WNITER; wsic++) {
+                [[unroll]] for (uint j = 0; j < TN; j++) {
+                    cache_b[wsic * TN + j] = buf_b[(warp_c * WN + wsic * WSUBN + tiwc * TN + j) * (BK+1) + i];
+                }
+            }
+
+            [[unroll]] for (uint wsic = 0; wsic < WNITER; wsic++) {
+                [[unroll]] for (uint wsir = 0; wsir < WMITER; wsir++) {
+                    [[unroll]] for (uint cc = 0; cc < TN; cc++) {
+                        [[unroll]] for (uint cr = 0; cr < TM; cr++) {
+                            const uint sums_idx = (wsic * TN + cc) * (WMITER * TM) + wsir * TM + cr;
+                            sums[sums_idx] = fma(float(cache_a[wsir * TM + cr]), float(cache_b[wsic * TN + cc]), sums[sums_idx]);
+                        }
+                    }
+                }
+            }
+        }
+
+        barrier();
+    }
+
+    const uint dr = ir * BM + warp_r * WM;
+    const uint dc = ic * BN + warp_c * WN;
+
+#ifndef MUL_MAT_ID
+    const uint offsets = batch_idx * p.batch_stride_d + ik * p.batch_stride_d * gl_NumWorkGroups.z;
+#endif
+
+    [[unroll]] for (uint wsic = 0; wsic < WNITER; wsic++) {
+        [[unroll]] for (uint wsir = 0; wsir < WMITER; wsir++) {
+
+            const uint dr_warp = dr + wsir * WSUBM + tiwr * TM;
+            const uint dc_warp = dc + wsic * WSUBN + tiwc * TN;
+            [[unroll]] for (uint cc = 0; cc < TN; cc++) {
+#ifdef MUL_MAT_ID
+                const uint row_i = dc_warp + cc;
+                if (row_i >= _ne1) break;
+
+                const u16vec2 row_idx = row_ids[row_i];
+#endif
+                [[unroll]] for (uint cr = 0; cr < TM; cr++) {
+#ifdef MUL_MAT_ID
+                    data_d[row_idx.y * p.batch_stride_d + row_idx.x * p.stride_d + dr_warp + cr] = D_TYPE(sums[(wsic * TN + cc) * (WMITER * TM) + wsir * TM + cr]);
+#else
+                    if (dr_warp + cr < p.M && dc_warp + cc < p.N) {
+                        data_d[offsets + (dc_warp + cc) * p.stride_d + dr_warp + cr] = D_TYPE(sums[(wsic * TN + cc) * (WMITER * TM) + wsir * TM + cr]);
+                    }
+#endif
+                }
+            }
+        }
+    }
+}
diff --git a/src/whisper_cpp/ggml/src/vulkan-shaders/norm.comp b/src/whisper_cpp/ggml/src/vulkan-shaders/norm.comp
new file mode 100644
index 00000000..6627a50b
--- /dev/null
+++ b/src/whisper_cpp/ggml/src/vulkan-shaders/norm.comp
@@ -0,0 +1,44 @@
+#version 450
+
+#include "generic_head.comp"
+#include "types.comp"
+
+#extension GL_EXT_control_flow_attributes : enable
+#define BLOCK_SIZE 512
+
+layout(local_size_x = BLOCK_SIZE, local_size_y = 1, local_size_z = 1) in;
+
+layout (binding = 0) readonly buffer X {A_TYPE data_a[];};
+layout (binding = 1) writeonly buffer D {D_TYPE data_d[];};
+
+shared vec2 sum[BLOCK_SIZE];
+
+void main() {
+    const uint row = gl_WorkGroupID.z * 262144 + gl_WorkGroupID.y * 512 + gl_WorkGroupID.x;
+    const uint tid = gl_LocalInvocationID.x;
+
+    sum[tid] = vec2(0.0f, 0.0f);
+
+    [[unroll]] for (uint col = tid; col < p.KX; col += BLOCK_SIZE) {
+        const float xi = float(data_a[row*p.KX + col]);
+        sum[tid].x += xi;
+        sum[tid].y += xi * xi;
+    }
+
+    // sum up partial sums and write back result
+    barrier();
+    [[unroll]] for (int s = BLOCK_SIZE / 2; s > 0; s >>= 1) {
+        if (tid < s) {
+            sum[tid] += sum[tid + s];
+        }
+        barrier();
+    }
+
+    const float mean = sum[0].x / p.KX;
+    const float var = sum[0].y / p.KX - mean * mean;
+    const float inv_std = inversesqrt(var + p.param1);
+
+    [[unroll]] for (uint col = tid; col < p.KX; col += BLOCK_SIZE) {
+        data_d[row*p.KX + col] = D_TYPE((float(data_a[row*p.KX + col]) - mean) * inv_std);
+    }
+}
diff --git a/src/whisper_cpp/ggml/src/vulkan-shaders/pad.comp b/src/whisper_cpp/ggml/src/vulkan-shaders/pad.comp
new file mode 100644
index 00000000..a465cd52
--- /dev/null
+++ b/src/whisper_cpp/ggml/src/vulkan-shaders/pad.comp
@@ -0,0 +1,26 @@
+#version 450
+
+#include "types.comp"
+#include "generic_unary_head.comp"
+
+void main() {
+    const uint idx = gl_GlobalInvocationID.z * 262144 + gl_GlobalInvocationID.y * 512 + gl_GlobalInvocationID.x;
+
+    if (idx >= p.ne) {
+        return;
+    }
+
+    const uint i3 = idx / (p.ne12*p.ne11*p.ne10);
+    const uint i3_offset = i3 * p.ne12*p.ne11*p.ne10;
+    const uint i2 = (idx - i3_offset) / (p.ne11*p.ne10);
+    const uint i2_offset = i2*p.ne11*p.ne10;
+    const uint i1 = (idx - i3_offset - i2_offset) / p.ne10;
+    const uint i0 = idx - i3_offset - i2_offset - i1*p.ne10;
+
+    const uint src0_idx = i3*p.nb03 + i2*p.nb02 + i1*p.nb01 + i0*p.nb00;
+    const uint dst_idx = i3*p.nb13 + i2*p.nb12 + i1*p.nb11 + i0*p.nb10;
+
+    const bool is_src0 = i0 < p.ne00 && i1 < p.ne01 && i2 < p.ne02 && i3 < p.ne03;
+
+    data_d[p.d_offset + dst_idx] = D_TYPE(is_src0 ? data_a[src0_idx] : 0.0f);
+}
diff --git a/src/whisper_cpp/ggml/src/vulkan-shaders/relu.comp b/src/whisper_cpp/ggml/src/vulkan-shaders/relu.comp
new file mode 100644
index 00000000..52a19b62
--- /dev/null
+++ b/src/whisper_cpp/ggml/src/vulkan-shaders/relu.comp
@@ -0,0 +1,21 @@
+#version 450
+
+#include "generic_head.comp"
+#include "types.comp"
+
+#extension GL_EXT_control_flow_attributes : enable
+
+layout(local_size_x = 512, local_size_y = 1, local_size_z = 1) in;
+
+layout (binding = 0) readonly buffer X {A_TYPE data_a[];};
+layout (binding = 1) writeonly buffer D {D_TYPE data_d[];};
+
+void main() {
+    const uint i = gl_GlobalInvocationID.z * 262144 + gl_GlobalInvocationID.y * 512 + gl_GlobalInvocationID.x;
+
+    if (i >= p.KX) {
+        return;
+    }
+
+    data_d[i] = max(float(data_a[i]), 0);
+}
diff --git a/src/whisper_cpp/ggml/src/vulkan-shaders/repeat.comp b/src/whisper_cpp/ggml/src/vulkan-shaders/repeat.comp
new file mode 100644
index 00000000..a86af87e
--- /dev/null
+++ b/src/whisper_cpp/ggml/src/vulkan-shaders/repeat.comp
@@ -0,0 +1,24 @@
+#version 450
+
+#include "types.comp"
+#include "generic_unary_head.comp"
+
+uint src0_idx_mod(uint idx) {
+    const uint i13 = idx / (p.ne12*p.ne11*p.ne10);
+    const uint i13_offset = i13 * p.ne12*p.ne11*p.ne10;
+    const uint i12 = (idx - i13_offset) / (p.ne11*p.ne10);
+    const uint i12_offset = i12*p.ne11*p.ne10;
+    const uint i11 = (idx - i13_offset - i12_offset) / p.ne10;
+    const uint i10 = idx - i13_offset - i12_offset - i11*p.ne10;
+    return (i13 % p.ne03)*p.nb03 + (i12 % p.ne02)*p.nb02 + (i11 % p.ne01)*p.nb01 + (i10 % p.ne00)*p.nb00;
+}
+
+void main() {
+    const uint idx = get_idx();
+
+    if (idx >= p.ne) {
+        return;
+    }
+
+    data_d[p.d_offset + dst_idx(idx)] = D_TYPE(data_a[src0_idx_mod(idx)]);
+}
diff --git a/src/whisper_cpp/ggml/src/vulkan-shaders/rms_norm.comp b/src/whisper_cpp/ggml/src/vulkan-shaders/rms_norm.comp
new file mode 100644
index 00000000..b554400b
--- /dev/null
+++ b/src/whisper_cpp/ggml/src/vulkan-shaders/rms_norm.comp
@@ -0,0 +1,42 @@
+#version 450
+
+#include "generic_head.comp"
+#include "types.comp"
+
+#extension GL_EXT_control_flow_attributes : enable
+#define BLOCK_SIZE 512
+
+layout(local_size_x = BLOCK_SIZE, local_size_y = 1, local_size_z = 1) in;
+
+layout (binding = 0) readonly buffer X {A_TYPE data_a[];};
+layout (binding = 1) writeonly buffer D {D_TYPE data_d[];};
+
+shared FLOAT_TYPE sum[BLOCK_SIZE];
+
+void main() {
+    const uint row = gl_WorkGroupID.z * 262144 + gl_WorkGroupID.y * 512 + gl_WorkGroupID.x;
+    const uint tid = gl_LocalInvocationID.x;
+
+    sum[tid] = FLOAT_TYPE(0.0f); // partial sum for thread in warp
+
+    [[unroll]] for (uint col = tid; col < p.KX; col += BLOCK_SIZE) {
+        const FLOAT_TYPE xi = FLOAT_TYPE(data_a[row*p.KX + col]);
+        sum[tid] += xi * xi;
+    }
+
+    // sum up partial sums and write back result
+    barrier();
+    [[unroll]] for (int s = BLOCK_SIZE / 2; s > 0; s >>= 1) {
+        if (tid < s) {
+            sum[tid] += sum[tid + s];
+        }
+        barrier();
+    }
+
+    const FLOAT_TYPE mean = sum[0] / FLOAT_TYPE(p.KX);
+    const FLOAT_TYPE scale = inversesqrt(mean + FLOAT_TYPE(p.param1));
+
+    [[unroll]] for (uint col = tid; col < p.KX; col += BLOCK_SIZE) {
+        data_d[row*p.KX + col] = D_TYPE(scale * FLOAT_TYPE(data_a[row*p.KX + col]));
+    }
+}
diff --git a/src/whisper_cpp/ggml/src/vulkan-shaders/rope_head.comp b/src/whisper_cpp/ggml/src/vulkan-shaders/rope_head.comp
new file mode 100644
index 00000000..ea895422
--- /dev/null
+++ b/src/whisper_cpp/ggml/src/vulkan-shaders/rope_head.comp
@@ -0,0 +1,44 @@
+#include "types.comp"
+
+#extension GL_EXT_shader_16bit_storage : require
+
+layout(local_size_x = 1, local_size_y = 256, local_size_z = 1) in;
+
+layout (binding = 0) readonly buffer X {A_TYPE data_a[];};
+layout (binding = 1) readonly buffer Y {int data_pos[];};
+layout (binding = 2) readonly buffer Z {float data_ff[];};
+layout (binding = 3) writeonly buffer D {D_TYPE data_d[];};
+
+layout (push_constant) uniform parameter {
+    uint ncols;
+    uint n_dims;
+    float freq_scale;
+    uint p_delta_rows;
+    float freq_base;
+    float ext_factor;
+    float attn_factor;
+    float corr_dims[2];
+    float theta_scale;
+    uint has_ff;
+} p;
+
+float rope_yarn_ramp(const float low, const float high, const uint i0) {
+    const float y = (i0 / 2 - low) / max(0.001f, high - low);
+    return 1.0f - min(1.0f, max(0.0f, y));
+}
+
+void rope_yarn(const float theta_extrap, const uint i0, out float cos_theta, out float sin_theta) {
+    float mscale = p.attn_factor;
+    // Get n-d rotational scaling corrected for extrapolation
+    float theta_interp = p.freq_scale * theta_extrap;
+    float theta = theta_interp;
+    if (p.ext_factor != 0.0f) {
+        float ramp_mix = rope_yarn_ramp(p.corr_dims[0], p.corr_dims[1], i0) * p.ext_factor;
+        theta = theta_interp * (1 - ramp_mix) + theta_extrap * ramp_mix;
+
+        // Get n-d magnitude scaling corrected for interpolation
+        mscale *= 1.0f + 0.1f * log(1.0f / p.freq_scale);
+    }
+    cos_theta = cos(theta) * mscale;
+    sin_theta = sin(theta) * mscale;
+}
diff --git a/src/whisper_cpp/ggml/src/vulkan-shaders/rope_neox.comp b/src/whisper_cpp/ggml/src/vulkan-shaders/rope_neox.comp
new file mode 100644
index 00000000..83b46b69
--- /dev/null
+++ b/src/whisper_cpp/ggml/src/vulkan-shaders/rope_neox.comp
@@ -0,0 +1,37 @@
+#version 450
+
+#include "rope_head.comp"
+
+void main() {
+    const uint col = gl_GlobalInvocationID.y * 2;
+    const uint row = gl_GlobalInvocationID.x;
+
+    if (col >= p.ncols) {
+        return;
+    }
+
+    if (col >= p.n_dims) {
+        const uint i = row*p.ncols + col;
+
+        data_d[i + 0] = data_a[i + 0];
+        data_d[i + 1] = data_a[i + 1];
+
+        return;
+    }
+
+    const uint i  = row*p.ncols + col/2;
+    const uint i2 = row/p.p_delta_rows;
+
+    const float theta_base = data_pos[i2] * pow(p.theta_scale, col/2.0f);
+
+    const float freq_factor = p.has_ff != 0 ? data_ff[col/2] : 1.0f;
+
+    float cos_theta, sin_theta;
+    rope_yarn(theta_base / freq_factor, col, cos_theta, sin_theta);
+
+    const float x0 = float(data_a[i + 0]);
+    const float x1 = float(data_a[i + p.n_dims/2]);
+
+    data_d[i + 0]        = D_TYPE(x0*cos_theta - x1*sin_theta);
+    data_d[i + p.n_dims/2] = D_TYPE(x0*sin_theta + x1*cos_theta);
+}
diff --git a/src/whisper_cpp/ggml/src/vulkan-shaders/rope_norm.comp b/src/whisper_cpp/ggml/src/vulkan-shaders/rope_norm.comp
new file mode 100644
index 00000000..e416ad93
--- /dev/null
+++ b/src/whisper_cpp/ggml/src/vulkan-shaders/rope_norm.comp
@@ -0,0 +1,37 @@
+#version 450
+
+#include "rope_head.comp"
+
+void main() {
+    const uint col = gl_GlobalInvocationID.y * 2;
+    const uint row = gl_GlobalInvocationID.x;
+
+    if (col >= p.ncols) {
+        return;
+    }
+
+    if (col >= p.n_dims) {
+        const uint i = row*p.ncols + col;
+
+        data_d[i + 0] = data_a[i + 0];
+        data_d[i + 1] = data_a[i + 1];
+
+        return;
+    }
+
+    const uint i = row*p.ncols + col;
+    const uint i2 = row/p.p_delta_rows;
+
+    const float theta_base = data_pos[i2] * pow(p.theta_scale, col/2.0f);
+
+    const float freq_factor = p.has_ff != 0 ? data_ff[col/2] : 1.0f;
+
+    float cos_theta, sin_theta;
+    rope_yarn(theta_base / freq_factor, col, cos_theta, sin_theta);
+
+    const float x0 = float(data_a[i + 0]);
+    const float x1 = float(data_a[i + 1]);
+
+    data_d[i + 0] = D_TYPE(x0*cos_theta - x1*sin_theta);
+    data_d[i + 1] = D_TYPE(x0*sin_theta + x1*cos_theta);
+}
diff --git a/src/whisper_cpp/ggml/src/vulkan-shaders/scale.comp b/src/whisper_cpp/ggml/src/vulkan-shaders/scale.comp
new file mode 100644
index 00000000..5cd2f668
--- /dev/null
+++ b/src/whisper_cpp/ggml/src/vulkan-shaders/scale.comp
@@ -0,0 +1,14 @@
+#version 450
+
+#include "types.comp"
+#include "generic_unary_head.comp"
+
+void main() {
+    const uint idx = get_idx();
+
+    if (idx >= p.ne) {
+        return;
+    }
+
+    data_d[p.d_offset + dst_idx(idx)] = D_TYPE(FLOAT_TYPE(data_a[src0_idx(idx)]) * FLOAT_TYPE(p.param1));
+}
diff --git a/src/whisper_cpp/ggml/src/vulkan-shaders/silu.comp b/src/whisper_cpp/ggml/src/vulkan-shaders/silu.comp
new file mode 100644
index 00000000..4d36f88e
--- /dev/null
+++ b/src/whisper_cpp/ggml/src/vulkan-shaders/silu.comp
@@ -0,0 +1,22 @@
+#version 450
+
+#include "generic_head.comp"
+#include "types.comp"
+
+#extension GL_EXT_control_flow_attributes : enable
+
+layout(local_size_x = 512, local_size_y = 1, local_size_z = 1) in;
+
+layout (binding = 0) readonly buffer X {A_TYPE data_a[];};
+layout (binding = 1) writeonly buffer D {D_TYPE data_d[];};
+
+void main() {
+    const uint i = gl_GlobalInvocationID.z * 262144 + gl_GlobalInvocationID.y * 512 + gl_GlobalInvocationID.x;
+
+    if (i >= p.KX) {
+        return;
+    }
+
+    const float xi = float(data_a[i]);
+    data_d[i] = D_TYPE(xi / (1.0f + exp(-xi)));
+}
diff --git a/src/whisper_cpp/ggml/src/vulkan-shaders/sin.comp b/src/whisper_cpp/ggml/src/vulkan-shaders/sin.comp
new file mode 100644
index 00000000..7faf9be9
--- /dev/null
+++ b/src/whisper_cpp/ggml/src/vulkan-shaders/sin.comp
@@ -0,0 +1,15 @@
+#version 450
+
+#include "types.comp"
+#include "generic_unary_head.comp"
+
+void main() {
+    const uint idx = get_idx();
+
+    if (idx >= p.ne) {
+        return;
+    }
+
+    const FLOAT_TYPE val = FLOAT_TYPE(data_a[src0_idx(idx)]);
+    data_d[p.d_offset + dst_idx(idx)] = D_TYPE(sin(val));
+}
diff --git a/src/whisper_cpp/ggml/src/vulkan-shaders/soft_max.comp b/src/whisper_cpp/ggml/src/vulkan-shaders/soft_max.comp
new file mode 100644
index 00000000..0bd51eca
--- /dev/null
+++ b/src/whisper_cpp/ggml/src/vulkan-shaders/soft_max.comp
@@ -0,0 +1,106 @@
+#version 450
+
+#extension GL_EXT_shader_16bit_storage : require
+
+layout (push_constant) uniform parameter
+{
+    uint KX;
+    uint KY;
+    float scale;
+    float max_bias;
+    float m0;
+    float m1;
+    uint n_head_log2;
+} p;
+
+#include "types.comp"
+
+#extension GL_EXT_control_flow_attributes : enable
+#define BLOCK_SIZE 512
+
+layout(local_size_x = BLOCK_SIZE, local_size_y = 1, local_size_z = 1) in;
+
+layout (binding = 0) readonly buffer X {A_TYPE data_a[];};
+layout (binding = 1) readonly buffer Y {B_TYPE data_b[];};
+layout (binding = 2) buffer D {D_TYPE data_d[];};
+
+shared FLOAT_TYPE vals[BLOCK_SIZE];
+
+void main() {
+    const uint tid = gl_LocalInvocationID.x;
+    const uint rowx = gl_WorkGroupID.z * 262144 + gl_WorkGroupID.y * 512 + gl_WorkGroupID.x;
+    const uint rowy = rowx % p.KY;
+
+    float slope = 1.0f;
+
+    // ALiBi
+    if (p.max_bias > 0.0f) {
+        const uint h = rowx/p.KY; // head index
+
+        const float base = h < p.n_head_log2 ? p.m0 : p.m1;
+        const uint   exp  = h < p.n_head_log2 ? h + 1 : 2*(h - p.n_head_log2) + 1;
+
+        slope = pow(base, exp);
+    }
+
+    // Find max
+    FLOAT_TYPE max_val = uintBitsToFloat(0xFF800000);
+
+    [[unroll]] for (uint col0 = 0; col0 < p.KX; col0 += BLOCK_SIZE) {
+        const uint col = col0 + tid;
+
+        if (col >= p.KX) {
+            break;
+        }
+
+        max_val = max(max_val, FLOAT_TYPE(data_a[rowx * p.KX + col]) * p.scale + (p.KY > 0 ? slope * FLOAT_TYPE(data_b[rowy * p.KX + col]) : FLOAT_TYPE(0.0f)));
+    }
+    vals[tid] = max_val;
+
+    barrier();
+    [[unroll]] for (int s = BLOCK_SIZE / 2; s > 0; s >>= 1) {
+        if (tid < s) {
+            vals[tid] = max(vals[tid], vals[tid + s]);
+        }
+        barrier();
+    }
+
+    max_val = vals[0];
+    barrier();
+
+    // Sum up values
+    vals[tid] = FLOAT_TYPE(0.0f);
+
+    [[unroll]] for (uint col0 = 0; col0 < p.KX; col0 += BLOCK_SIZE) {
+        const uint col = col0 + tid;
+
+        if (col >= p.KX) {
+            break;
+        }
+
+        const uint i = rowx * p.KX + col;
+        const FLOAT_TYPE val = exp(FLOAT_TYPE(data_a[i]) * p.scale + (p.KY > 0 ? slope * FLOAT_TYPE(data_b[rowy * p.KX + col]) : FLOAT_TYPE(0.0f)) - max_val);
+        vals[tid] += val;
+        data_d[i] = D_TYPE(val);
+    }
+
+    barrier();
+    [[unroll]] for (int s = BLOCK_SIZE / 2; s > 0; s >>= 1) {
+        if (tid < s) {
+            vals[tid] += vals[tid + s];
+        }
+        barrier();
+    }
+
+    const D_TYPE divisor = D_TYPE(vals[0]);
+
+    [[unroll]] for (uint col0 = 0; col0 < p.KX; col0 += BLOCK_SIZE) {
+        const uint col = col0 + tid;
+
+        if (col >= p.KX) {
+            break;
+        }
+
+        data_d[rowx*p.KX + col] /= divisor;
+    }
+}
diff --git a/src/whisper_cpp/ggml/src/vulkan-shaders/square.comp b/src/whisper_cpp/ggml/src/vulkan-shaders/square.comp
new file mode 100644
index 00000000..1fa118c9
--- /dev/null
+++ b/src/whisper_cpp/ggml/src/vulkan-shaders/square.comp
@@ -0,0 +1,15 @@
+#version 450
+
+#include "types.comp"
+#include "generic_unary_head.comp"
+
+void main() {
+    const uint idx = get_idx();
+
+    if (idx >= p.ne) {
+        return;
+    }
+
+    const FLOAT_TYPE val = FLOAT_TYPE(data_a[src0_idx(idx)]);
+    data_d[p.d_offset + dst_idx(idx)] = D_TYPE(val * val);
+}
diff --git a/src/whisper_cpp/ggml/src/vulkan-shaders/sum_rows.comp b/src/whisper_cpp/ggml/src/vulkan-shaders/sum_rows.comp
new file mode 100644
index 00000000..961e5ffa
--- /dev/null
+++ b/src/whisper_cpp/ggml/src/vulkan-shaders/sum_rows.comp
@@ -0,0 +1,37 @@
+#version 450
+
+#include "generic_head.comp"
+#include "types.comp"
+
+#extension GL_EXT_control_flow_attributes : enable
+layout(local_size_x_id = 0, local_size_y = 1, local_size_z = 1) in;
+
+layout (binding = 0) readonly buffer A {A_TYPE data_a[];};
+layout (binding = 1) writeonly buffer D {D_TYPE data_d[];};
+
+layout (constant_id = 0) const uint BLOCK_SIZE = 32;
+
+shared FLOAT_TYPE tmp[BLOCK_SIZE];
+
+void main() {
+    const uint row = gl_WorkGroupID.z * 262144 + gl_WorkGroupID.y * 512 + gl_WorkGroupID.x;
+    const uint col = gl_LocalInvocationID.x;
+
+    tmp[col] = FLOAT_TYPE(0.0f);
+
+    for (uint i = col; i < p.KX; i += BLOCK_SIZE) {
+        tmp[col] += FLOAT_TYPE(data_a[row*p.KX + i]);
+    }
+
+    barrier();
+    [[unroll]] for (int s = int(BLOCK_SIZE) / 2; s > 0; s >>= 1) {
+        if (col < s) {
+            tmp[col] += tmp[col + s];
+        }
+        barrier();
+    }
+
+    if (col == 0) {
+        data_d[row] = D_TYPE(tmp[0]);
+    }
+}
diff --git a/src/whisper_cpp/ggml/src/vulkan-shaders/tanh.comp b/src/whisper_cpp/ggml/src/vulkan-shaders/tanh.comp
new file mode 100644
index 00000000..74630dc7
--- /dev/null
+++ b/src/whisper_cpp/ggml/src/vulkan-shaders/tanh.comp
@@ -0,0 +1,21 @@
+#version 450
+
+#include "generic_head.comp"
+#include "types.comp"
+
+#extension GL_EXT_control_flow_attributes : enable
+
+layout(local_size_x = 512, local_size_y = 1, local_size_z = 1) in;
+
+layout (binding = 0) readonly buffer X {A_TYPE data_a[];};
+layout (binding = 1) writeonly buffer D {D_TYPE data_d[];};
+
+void main() {
+    const uint i = gl_GlobalInvocationID.z * 262144 + gl_GlobalInvocationID.y * 512 + gl_GlobalInvocationID.x;
+
+    if (i >= p.KX) {
+        return;
+    }
+
+    data_d[i] = D_TYPE(tanh(data_a[i]));
+}
diff --git a/src/whisper_cpp/ggml/src/vulkan-shaders/timestep_embedding.comp b/src/whisper_cpp/ggml/src/vulkan-shaders/timestep_embedding.comp
new file mode 100644
index 00000000..79e065a9
--- /dev/null
+++ b/src/whisper_cpp/ggml/src/vulkan-shaders/timestep_embedding.comp
@@ -0,0 +1,41 @@
+#version 450
+
+#extension GL_EXT_shader_16bit_storage : require
+
+layout (push_constant) uniform parameter
+{
+    uint nb1;
+    uint dim;
+    uint max_period;
+} p;
+
+#include "types.comp"
+
+#extension GL_EXT_control_flow_attributes : enable
+#define BLOCK_SIZE 256
+
+layout(local_size_x = BLOCK_SIZE, local_size_y = 1, local_size_z = 1) in;
+
+layout (binding = 0) readonly buffer X {A_TYPE data_a[];};
+layout (binding = 1) writeonly buffer D {D_TYPE data_d[];};
+
+void main() {
+    const uint i = gl_WorkGroupID.y;
+    const uint j = gl_GlobalInvocationID.x;
+    const uint d_offset = i * p.nb1;
+
+    if (p.dim % 2 != 0 && j == ((p.dim + 1) / 2)) {
+        data_d[d_offset + p.dim] = 0.f;
+    }
+
+    const uint half_dim = p.dim / 2;
+    if (j >= half_dim) {
+        return;
+    }
+
+    const float timestep = float(data_a[i]);
+    const float freq = float(exp(-log(p.max_period) * j / half_dim));
+    const float arg = timestep * freq;
+    data_d[d_offset + j] = D_TYPE(cos(arg));
+    data_d[d_offset + j + half_dim] = D_TYPE(sin(arg));
+}
diff --git a/src/whisper_cpp/ggml/src/vulkan-shaders/types.comp b/src/whisper_cpp/ggml/src/vulkan-shaders/types.comp
new file mode 100644
index 00000000..21dce72f
--- /dev/null
+++ b/src/whisper_cpp/ggml/src/vulkan-shaders/types.comp
@@ -0,0 +1,200 @@
+#if !defined(DATA_A_F32) && !defined(DATA_A_F16)
+#extension GL_EXT_shader_explicit_arithmetic_types_int8 : require
+#endif
+
+#if defined(DATA_A_F32)
+#define QUANT_K 1
+#define QUANT_R 1
+
+#if !defined(LOAD_VEC_A) || LOAD_VEC_A == 1
+#define A_TYPE float
+#elif LOAD_VEC_A == 4
+#define A_TYPE vec4
+#elif LOAD_VEC_A == 8
+#define A_TYPE mat2x4
+#endif
+#endif
+
+#if defined(DATA_A_F16)
+#define QUANT_K 1
+#define QUANT_R 1
+
+#if !defined(LOAD_VEC_A) || LOAD_VEC_A == 1
+#define A_TYPE float16_t
+#elif LOAD_VEC_A == 4
+#define A_TYPE f16vec4
+#elif LOAD_VEC_A == 8
+#define A_TYPE f16mat2x4
+#endif
+#endif
+
+#if defined(DATA_A_Q4_0)
+#extension GL_EXT_shader_16bit_storage : require
+#define QUANT_K 32
+#define QUANT_R 2
+
+struct block_q4_0
+{
+    float16_t d;
+    uint8_t qs[16];
+};
+
+#define A_TYPE block_q4_0
+#endif
+
+#if defined(DATA_A_Q4_1)
+#extension GL_EXT_shader_16bit_storage : require
+#define QUANT_K 32
+#define QUANT_R 2
+
+struct block_q4_1
+{
+    float16_t d;
+    float16_t m;
+    uint8_t qs[16];
+};
+
+#define A_TYPE block_q4_1
+#endif
+
+#if defined(DATA_A_Q5_0)
+#extension GL_EXT_shader_16bit_storage : require
+#extension GL_EXT_shader_explicit_arithmetic_types_int16 : require
+#define QUANT_K 32
+#define QUANT_R 2
+
+struct block_q5_0
+{
+    float16_t d;
+    uint16_t qh[2];
+    uint8_t qs[16];
+};
+
+#define A_TYPE block_q5_0
+#endif
+
+#if defined(DATA_A_Q5_1)
+#extension GL_EXT_shader_16bit_storage : require
+#extension GL_EXT_shader_explicit_arithmetic_types_int16 : require
+#define QUANT_K 32
+#define QUANT_R 2
+
+struct block_q5_1
+{
+    float16_t d;
+    float16_t m;
+    uint qh;
+    uint8_t qs[16];
+};
+
+#define A_TYPE block_q5_1
+#endif
+
+#if defined(DATA_A_Q8_0)
+#extension GL_EXT_shader_16bit_storage : require
+#define QUANT_K 32
+#define QUANT_R 1
+
+struct block_q8_0
+{
+    float16_t d;
+    int8_t qs[32];
+};
+
+#define A_TYPE block_q8_0
+#endif
+
+// K-quants
+#if defined(DATA_A_Q2_K)
+#extension GL_EXT_shader_16bit_storage : require
+#define QUANT_K 256
+
+struct block_q2_K
+{
+    uint8_t scales[QUANT_K/16];
+    uint8_t qs[QUANT_K/4];
+    f16vec2 d;
+};
+
+#define A_TYPE block_q2_K
+#endif
+
+#if defined(DATA_A_Q3_K)
+#extension GL_EXT_shader_16bit_storage : require
+#define QUANT_K 256
+
+struct block_q3_K
+{
+    uint8_t hmask[QUANT_K/8];
+    uint8_t qs[QUANT_K/4];
+    uint8_t scales[12];
+    float16_t d;
+};
+
+#define A_TYPE block_q3_K
+#endif
+
+#if defined(DATA_A_Q4_K)
+#extension GL_EXT_shader_16bit_storage : require
+#define QUANT_K 256
+
+struct block_q4_K
+{
+    f16vec2 d;
+    uint8_t scales[3*QUANT_K/64];
+    uint8_t qs[QUANT_K/2];
+};
+
+#define A_TYPE block_q4_K
+#endif
+
+#if defined(DATA_A_Q5_K)
+#extension GL_EXT_shader_16bit_storage : require
+#define QUANT_K 256
+
+struct block_q5_K
+{
+    f16vec2 d;
+    uint8_t scales[12];
+    uint8_t qh[QUANT_K/8];
+    uint8_t qs[QUANT_K/2];
+};
+
+#define A_TYPE block_q5_K
+#endif
+
+#if defined(DATA_A_Q6_K)
+#extension GL_EXT_shader_16bit_storage : require
+#define QUANT_K 256
+
+struct block_q6_K
+{
+    uint8_t ql[QUANT_K/2];
+    uint8_t qh[QUANT_K/4];
+    int8_t scales[QUANT_K/16];
+    float16_t d;
+};
+
+#define A_TYPE block_q6_K
+#endif
+
+// IQuants
+
+#if defined(DATA_A_IQ4_NL)
+#extension GL_EXT_shader_16bit_storage : require
+#define QUANT_K 32
+#define QUANT_R 2
+
+struct block_iq4_nl
+{
+    float16_t d;
+    uint8_t qs[QUANT_K/2];
+};
+
+#define A_TYPE block_iq4_nl
+
+const int8_t kvalues_iq4nl[16] = {
+    int8_t(-127), int8_t(-104), int8_t(-83), int8_t(-65), int8_t(-49), int8_t(-35), int8_t(-22), int8_t(-10),
+    int8_t(1), int8_t(13), int8_t(25), int8_t(38), int8_t(53), int8_t(69), int8_t(89), int8_t(113)
+};
+#endif
diff --git a/src/whisper_cpp/ggml/src/vulkan-shaders/upscale.comp b/src/whisper_cpp/ggml/src/vulkan-shaders/upscale.comp
new file mode 100644
index 00000000..511a086e
--- /dev/null
+++ b/src/whisper_cpp/ggml/src/vulkan-shaders/upscale.comp
@@ -0,0 +1,36 @@
+#version 450
+
+layout (push_constant) uniform parameter
+{
+    uint ne; uint d_offset;
+    uint nb00; uint nb01; uint nb02; uint nb03;
+    uint ne10; uint ne11; uint ne12; uint ne13;
+    float sf0; float sf1; float sf2; float sf3;
+} p;
+
+#include "types.comp"
+
+layout(local_size_x = 512, local_size_y = 1, local_size_z = 1) in;
+
+layout (binding = 0) readonly buffer A {A_TYPE data_a[];};
+layout (binding = 1) writeonly buffer D {D_TYPE data_d[];};
+
+void main() {
+    const uint idx = gl_GlobalInvocationID.z * 262144 + gl_GlobalInvocationID.y * 512 + gl_GlobalInvocationID.x;
+
+    if (idx >= p.ne) {
+        return;
+    }
+
+    const uint i10 = idx % p.ne10;
+    const uint i11 = (idx / p.ne10) % p.ne11;
+    const uint i12 = (idx / (p.ne10 * p.ne11)) % p.ne12;
+    const uint i13 = (idx / (p.ne10 * p.ne11 * p.ne12)) % p.ne13;
+
+    const uint i00 = uint(i10 / p.sf0);
+    const uint i01 = uint(i11 / p.sf1);
+    const uint i02 = uint(i12 / p.sf2);
+    const uint i03 = uint(i13 / p.sf3);
+
+    data_d[p.d_offset + idx] = D_TYPE(data_a[i03 * p.nb03 + i02 * p.nb02 + i01 * p.nb01 + i00 * p.nb00]);
+}
diff --git a/src/whisper_cpp/ggml/src/vulkan-shaders/vulkan-shaders-gen.cpp b/src/whisper_cpp/ggml/src/vulkan-shaders/vulkan-shaders-gen.cpp
new file mode 100644
index 00000000..1bd1b6f6
--- /dev/null
+++ b/src/whisper_cpp/ggml/src/vulkan-shaders/vulkan-shaders-gen.cpp
@@ -0,0 +1,600 @@
+
+
+#include <iostream>
+#include <fstream>
+#include <sstream>
+#include <string>
+#include <stdexcept>
+#include <array>
+#include <vector>
+#include <map>
+#include <thread>
+#include <mutex>
+#include <future>
+#include <queue>
+#include <condition_variable>
+#include <cstdio>
+#include <cstring>
+#include <cstdlib>
+#include <sys/stat.h>
+#include <sys/types.h>
+
+#ifdef _WIN32
+    #include <windows.h>
+    #include <direct.h> // For _mkdir on Windows
+    #include <algorithm> // For std::replace on w64devkit
+#else
+    #include <unistd.h>
+    #include <sys/wait.h>
+    #include <fcntl.h>
+#endif
+
+#define ASYNCIO_CONCURRENCY 64
+
+std::mutex lock;
+std::vector<std::pair<std::string, std::string>> shader_fnames;
+
+std::string GLSLC = "glslc";
+std::string input_dir = "vulkan-shaders";
+std::string output_dir = "/tmp";
+std::string target_hpp = "ggml-vulkan-shaders.hpp";
+std::string target_cpp = "ggml-vulkan-shaders.cpp";
+bool no_clean = false;
+
+const std::vector<std::string> type_names = {
+    "f32",
+    "f16",
+    "q4_0",
+    "q4_1",
+    "q5_0",
+    "q5_1",
+    "q8_0",
+    "q2_k",
+    "q3_k",
+    "q4_k",
+    "q5_k",
+    "q6_k",
+    "iq4_nl"
+};
+
+void execute_command(const std::string& command, std::string& stdout_str, std::string& stderr_str) {
+#ifdef _WIN32
+    HANDLE stdout_read, stdout_write;
+    HANDLE stderr_read, stderr_write;
+    SECURITY_ATTRIBUTES sa = { sizeof(SECURITY_ATTRIBUTES), NULL, TRUE };
+
+    if (!CreatePipe(&stdout_read, &stdout_write, &sa, 0) ||
+        !SetHandleInformation(stdout_read, HANDLE_FLAG_INHERIT, 0)) {
+        throw std::runtime_error("Failed to create stdout pipe");
+    }
+
+    if (!CreatePipe(&stderr_read, &stderr_write, &sa, 0) ||
+        !SetHandleInformation(stderr_read, HANDLE_FLAG_INHERIT, 0)) {
+        throw std::runtime_error("Failed to create stderr pipe");
+    }
+
+    PROCESS_INFORMATION pi;
+    STARTUPINFOA si = { sizeof(STARTUPINFOA) };
+    si.dwFlags = STARTF_USESTDHANDLES;
+    si.hStdOutput = stdout_write;
+    si.hStdError = stderr_write;
+
+    std::vector<char> cmd(command.begin(), command.end());
+    cmd.push_back('\0');
+
+    if (!CreateProcessA(NULL, cmd.data(), NULL, NULL, TRUE, 0, NULL, NULL, &si, &pi)) {
+        throw std::runtime_error("Failed to create process");
+    }
+
+    CloseHandle(stdout_write);
+    CloseHandle(stderr_write);
+
+    std::array<char, 128> buffer;
+    DWORD bytes_read;
+
+    while (ReadFile(stdout_read, buffer.data(), buffer.size(), &bytes_read, NULL) && bytes_read > 0) {
+        stdout_str.append(buffer.data(), bytes_read);
+    }
+
+    while (ReadFile(stderr_read, buffer.data(), buffer.size(), &bytes_read, NULL) && bytes_read > 0) {
+        stderr_str.append(buffer.data(), bytes_read);
+    }
+
+    CloseHandle(stdout_read);
+    CloseHandle(stderr_read);
+    WaitForSingleObject(pi.hProcess, INFINITE);
+    CloseHandle(pi.hProcess);
+    CloseHandle(pi.hThread);
+#else
+int stdout_pipe[2];
+    int stderr_pipe[2];
+
+    if (pipe(stdout_pipe) != 0 || pipe(stderr_pipe) != 0) {
+        throw std::runtime_error("Failed to create pipes");
+    }
+
+    pid_t pid = fork();
+    if (pid < 0) {
+        throw std::runtime_error("Failed to fork process");
+    }
+
+    if (pid == 0) {
+        close(stdout_pipe[0]);
+        close(stderr_pipe[0]);
+        dup2(stdout_pipe[1], STDOUT_FILENO);
+        dup2(stderr_pipe[1], STDERR_FILENO);
+        close(stdout_pipe[1]);
+        close(stderr_pipe[1]);
+        execl("/bin/sh", "sh", "-c", command.c_str(), (char*) nullptr);
+        _exit(EXIT_FAILURE);
+    } else {
+        close(stdout_pipe[1]);
+        close(stderr_pipe[1]);
+
+        std::array<char, 128> buffer;
+        ssize_t bytes_read;
+
+        while ((bytes_read = read(stdout_pipe[0], buffer.data(), buffer.size())) > 0) {
+            stdout_str.append(buffer.data(), bytes_read);
+        }
+
+        while ((bytes_read = read(stderr_pipe[0], buffer.data(), buffer.size())) > 0) {
+            stderr_str.append(buffer.data(), bytes_read);
+        }
+
+        close(stdout_pipe[0]);
+        close(stderr_pipe[0]);
+        waitpid(pid, nullptr, 0);
+    }
+#endif
+}
+
+bool directory_exists(const std::string& path) {
+    struct stat info;
+    if (stat(path.c_str(), &info) != 0) {
+        return false; // Path doesn't exist or can't be accessed
+    }
+    return (info.st_mode & S_IFDIR) != 0; // Check if it is a directory
+}
+
+bool create_directory(const std::string& path) {
+#ifdef _WIN32
+    return _mkdir(path.c_str()) == 0 || errno == EEXIST; // EEXIST means the directory already exists
+#else
+    return mkdir(path.c_str(), 0755) == 0 || errno == EEXIST; // 0755 is the directory permissions
+#endif
+}
+
+std::string to_uppercase(const std::string& input) {
+    std::string result = input;
+    for (char& c : result) {
+        c = std::toupper(c);
+    }
+    return result;
+}
+
+bool string_ends_with(const std::string& str, const std::string& suffix) {
+    if (suffix.size() > str.size()) {
+        return false;
+    }
+    return std::equal(suffix.rbegin(), suffix.rend(), str.rbegin());
+}
+
+static const char path_separator = '/';
+
+std::string join_paths(const std::string& path1, const std::string& path2) {
+    return path1 + path_separator + path2;
+}
+
+std::string basename(const std::string &path) {
+    return path.substr(path.find_last_of("/\\") + 1);
+}
+
+void string_to_spv(const std::string& _name, const std::string& in_fname, const std::map<std::string, std::string>& defines, bool fp16 = true) {
+    std::string name = _name + (fp16 ? "" : "_fp32");
+    std::string out_fname = join_paths(output_dir, name + ".spv");
+    std::string in_path = join_paths(input_dir, in_fname);
+
+    #ifdef _WIN32
+        std::vector<std::string> cmd = {GLSLC, "-fshader-stage=compute", "--target-env=vulkan1.2", "-O", "\"" + in_path + "\"", "-o", "\"" + out_fname + "\""};
+    #else
+        std::vector<std::string> cmd = {GLSLC, "-fshader-stage=compute", "--target-env=vulkan1.2", "-O", in_path, "-o",  out_fname};
+    #endif
+
+    #ifdef GGML_VULKAN_SHADER_DEBUG_INFO
+        cmd.push_back("-g");
+    #endif
+
+    for (const auto& define : defines) {
+        cmd.push_back("-D" + define.first + "=" + define.second);
+    }
+
+    std::string command;
+    for (const auto& part : cmd) {
+        command += part + " ";
+    }
+
+    std::string stdout_str, stderr_str;
+    try {
+        // std::cout << "Executing command: ";
+        // for (const auto& part : cmd) {
+        //     std::cout << part << " ";
+        // }
+        // std::cout << std::endl;
+
+        execute_command(command, stdout_str, stderr_str);
+        if (!stderr_str.empty()) {
+            std::cerr << "cannot compile " << name << "\n\n" << command << "\n\n" << stderr_str << std::endl;
+            return;
+        }
+
+        std::lock_guard<std::mutex> guard(lock);
+        shader_fnames.push_back(std::make_pair(name, out_fname));
+    } catch (const std::exception& e) {
+        std::cerr << "Error executing command for " << name << ": " << e.what() << std::endl;
+    }
+}
+
+std::map<std::string, std::string> merge_maps(const std::map<std::string, std::string>& a, const std::map<std::string, std::string>& b) {
+    std::map<std::string, std::string> result = a;
+    result.insert(b.begin(), b.end());
+    return result;
+}
+
+void matmul_shaders(std::vector<std::future<void>>& tasks, bool fp16, bool matmul_id) {
+    std::string load_vec = fp16 ? "8" : "4";
+    std::string aligned_b_type_f32 = fp16 ? "mat2x4" : "vec4";
+    std::string aligned_b_type_f16 = fp16 ? "f16mat2x4" : "f16vec4";
+
+    std::map<std::string, std::string> base_dict = {{"FLOAT_TYPE", fp16 ? "float16_t" : "float"}};
+    std::string shader_name = "matmul";
+
+    if (matmul_id) {
+        base_dict["MUL_MAT_ID"] = "1";
+        shader_name = "matmul_id";
+    }
+
+    if (fp16) {
+        base_dict["FLOAT16"] = "1";
+    }
+
+    // Shaders with f16 B_TYPE
+    tasks.push_back(std::async(std::launch::async, [=] {
+        string_to_spv(shader_name + "_f32_f16", "mul_mm.comp", merge_maps(base_dict, {{"DATA_A_F32", "1"}, {"B_TYPE", "float16_t"}, {"D_TYPE", "float"}}), fp16);
+    }));
+    tasks.push_back(std::async(std::launch::async, [=] {
+        string_to_spv(shader_name + "_f32_f16_aligned", "mul_mm.comp", merge_maps(base_dict, {{"DATA_A_F32", "1"}, {"LOAD_VEC_A", load_vec}, {"LOAD_VEC_B", load_vec}, {"B_TYPE", aligned_b_type_f16}, {"D_TYPE", "float"}}), fp16);
+    }));
+
+    tasks.push_back(std::async(std::launch::async, [=] {
+        string_to_spv(shader_name + "_f16", "mul_mm.comp", merge_maps(base_dict, {{"DATA_A_F16", "1"}, {"B_TYPE", "float16_t"}, {"D_TYPE", "float"}}), fp16);
+    }));
+    tasks.push_back(std::async(std::launch::async, [=] {
+        string_to_spv(shader_name + "_f16_aligned", "mul_mm.comp", merge_maps(base_dict, {{"DATA_A_F16", "1"}, {"LOAD_VEC_A", load_vec}, {"LOAD_VEC_B", load_vec}, {"B_TYPE", aligned_b_type_f16}, {"D_TYPE", "float"}}), fp16);
+    }));
+
+    for (const auto& tname : type_names) {
+        std::string data_a_key = "DATA_A_" + to_uppercase(tname);
+        // For unaligned, load one at a time for f32/f16, or two at a time for quants
+        std::string load_vec_a_unaligned = (tname == "f32" || tname == "f16") ? "1" : "2";
+        // For aligned matmul loads
+        std::string load_vec_a = (tname == "f32" || tname == "f16") ? load_vec : "2";
+        tasks.push_back(std::async(std::launch::async, [=] {
+            string_to_spv(shader_name + "_" + tname + "_f32", "mul_mm.comp", merge_maps(base_dict, {{data_a_key, "1"}, {"LOAD_VEC_A", load_vec_a_unaligned}, {"B_TYPE", "float"}, {"D_TYPE", "float"}}), fp16);
+        }));
+        tasks.push_back(std::async(std::launch::async, [=] {
+            string_to_spv(shader_name + "_" + tname + "_f32_aligned", "mul_mm.comp", merge_maps(base_dict, {{data_a_key, "1"}, {"LOAD_VEC_A", load_vec_a}, {"LOAD_VEC_B", load_vec}, {"B_TYPE", aligned_b_type_f32}, {"D_TYPE", "float"}}), fp16);
+        }));
+    }
+}
+
+void process_shaders(std::vector<std::future<void>>& tasks) {
+    std::cout << "ggml_vulkan: Generating and compiling shaders to SPIR-V" << std::endl;
+    std::map<std::string, std::string> base_dict = {{"FLOAT_TYPE", "float"}};
+
+    for (const auto& fp16 : {false, true}) {
+        matmul_shaders(tasks, fp16, false);
+        matmul_shaders(tasks, fp16, true);
+    }
+
+    for (const auto& tname : type_names) {
+        // mul mat vec
+        std::string data_a_key = "DATA_A_" + to_uppercase(tname);
+        std::string shader = (string_ends_with(tname, "_k")) ? "mul_mat_vec_" + tname + ".comp" : "mul_mat_vec.comp";
+
+        tasks.push_back(std::async(std::launch::async, [=] {
+            string_to_spv("mul_mat_vec_" + tname + "_f32_f32", shader, merge_maps(base_dict, {{data_a_key, "1"}, {"B_TYPE", "float"}, {"D_TYPE", "float"}}));
+        }));
+        tasks.push_back(std::async(std::launch::async, [=] {
+            string_to_spv("mul_mat_vec_" + tname + "_f16_f32", shader, merge_maps(base_dict, {{data_a_key, "1"}, {"B_TYPE", "float16_t"}, {"D_TYPE", "float"}}));
+        }));
+
+        tasks.push_back(std::async(std::launch::async, [=] {
+            string_to_spv("mul_mat_vec_id_" + tname + "_f32", shader, merge_maps(base_dict, {{"MUL_MAT_ID", "1"}, {data_a_key, "1"}, {"B_TYPE", "float"}, {"D_TYPE", "float"}}));
+        }));
+
+        // Dequant shaders
+        if (tname != "f16") {
+            tasks.push_back(std::async(std::launch::async, [=] {
+                string_to_spv("dequant_" + tname, "dequant_" + tname + ".comp", merge_maps(base_dict, {{data_a_key, "1"}, {"D_TYPE", "float16_t"}}));
+            }));
+        }
+
+        if (!string_ends_with(tname, "_k")) {
+            shader = (tname == "f32" || tname == "f16") ? "get_rows.comp" : "get_rows_quant.comp";
+
+            if (tname == "f16") {
+                tasks.push_back(std::async(std::launch::async, [=] {
+                    string_to_spv("get_rows_" + tname, shader, {{data_a_key, "1"}, {"B_TYPE", "int"}, {"D_TYPE", "float16_t"}, {"OPTIMIZATION_ERROR_WORKAROUND", "1"}});
+                }));
+            } else {
+                tasks.push_back(std::async(std::launch::async, [=] {
+                    string_to_spv("get_rows_" + tname, shader, {{data_a_key, "1"}, {"B_TYPE", "int"}, {"D_TYPE", "float16_t"}});
+                }));
+            }
+            tasks.push_back(std::async(std::launch::async, [=] {
+                string_to_spv("get_rows_" + tname + "_f32", shader, {{data_a_key, "1"}, {"B_TYPE", "int"}, {"D_TYPE", "float"}});
+            }));
+        }
+    }
+
+    tasks.push_back(std::async(std::launch::async, [] {
+        string_to_spv("mul_mat_vec_p021_f16_f32", "mul_mat_vec_p021.comp", {{"A_TYPE", "float16_t"}, {"B_TYPE", "float"}, {"D_TYPE", "float"}});
+    }));
+    tasks.push_back(std::async(std::launch::async, [] {
+        string_to_spv("mul_mat_vec_nc_f16_f32", "mul_mat_vec_nc.comp", {{"A_TYPE", "float16_t"}, {"B_TYPE", "float"}, {"D_TYPE", "float"}});
+    }));
+
+    // Norms
+    tasks.push_back(std::async(std::launch::async, [=] {
+        string_to_spv("norm_f32", "norm.comp", merge_maps(base_dict, {{"A_TYPE", "float"}, {"D_TYPE", "float"}}));
+    }));
+    tasks.push_back(std::async(std::launch::async, [=] {
+        string_to_spv("group_norm_f32", "group_norm.comp", merge_maps(base_dict, {{"A_TYPE", "float"}, {"D_TYPE", "float"}}));
+    }));
+    tasks.push_back(std::async(std::launch::async, [=] {
+        string_to_spv("rms_norm_f32", "rms_norm.comp", merge_maps(base_dict, {{"A_TYPE", "float"}, {"D_TYPE", "float"}}));
+    }));
+
+    tasks.push_back(std::async(std::launch::async, [] {
+        string_to_spv("cpy_f32_f32", "copy.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}});
+    }));
+    tasks.push_back(std::async(std::launch::async, [] {
+        string_to_spv("cpy_f32_f16", "copy.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float16_t"}});
+    }));
+    tasks.push_back(std::async(std::launch::async, [] {
+        string_to_spv("cpy_f16_f16", "copy.comp", {{"A_TYPE", "float16_t"}, {"D_TYPE", "float16_t"}, {"OPTIMIZATION_ERROR_WORKAROUND", "1"}});
+    }));
+
+    tasks.push_back(std::async(std::launch::async, [] {
+        string_to_spv("add_f32", "add.comp", {{"A_TYPE", "float"}, {"B_TYPE", "float"}, {"D_TYPE", "float"}, {"FLOAT_TYPE", "float"}});
+    }));
+    tasks.push_back(std::async(std::launch::async, [] {
+        string_to_spv("add_f16_f32_f16", "add.comp", {{"A_TYPE", "float16_t"}, {"B_TYPE", "float"}, {"D_TYPE", "float16_t"}, {"FLOAT_TYPE", "float"}});
+    }));
+
+    tasks.push_back(std::async(std::launch::async, [] {
+        string_to_spv("acc_f32", "acc.comp", {{"A_TYPE", "float"}, {"B_TYPE", "float"}, {"D_TYPE", "float"}, {"FLOAT_TYPE", "float"}});
+    }));
+
+    tasks.push_back(std::async(std::launch::async, [] {
+        string_to_spv("split_k_reduce", "mul_mat_split_k_reduce.comp", {});
+    }));
+
+    tasks.push_back(std::async(std::launch::async, [] {
+        string_to_spv("mul_f32", "mul.comp", {{"A_TYPE", "float"}, {"B_TYPE", "float"}, {"D_TYPE", "float"}, {"FLOAT_TYPE", "float"}});
+    }));
+
+    tasks.push_back(std::async(std::launch::async, [] {
+        string_to_spv("div_f32", "div.comp", {{"A_TYPE", "float"}, {"B_TYPE", "float"}, {"D_TYPE", "float"}, {"FLOAT_TYPE", "float"}});
+    }));
+
+    tasks.push_back(std::async(std::launch::async, [] {
+        string_to_spv("repeat_f32", "repeat.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}});
+    }));
+
+    tasks.push_back(std::async(std::launch::async, [] {
+        string_to_spv("scale_f32", "scale.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}, {"FLOAT_TYPE", "float"}});
+    }));
+
+    tasks.push_back(std::async(std::launch::async, [] {
+        string_to_spv("sqr_f32", "square.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}, {"FLOAT_TYPE", "float"}});
+    }));
+
+    tasks.push_back(std::async(std::launch::async, [] {
+        string_to_spv("sin_f32", "sin.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}, {"FLOAT_TYPE", "float"}});
+    }));
+
+    tasks.push_back(std::async(std::launch::async, [] {
+        string_to_spv("cos_f32", "cos.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}, {"FLOAT_TYPE", "float"}});
+    }));
+
+    tasks.push_back(std::async(std::launch::async, [] {
+        string_to_spv("clamp_f32", "clamp.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}, {"FLOAT_TYPE", "float"}});
+    }));
+
+    tasks.push_back(std::async(std::launch::async, [] {
+        string_to_spv("pad_f32", "pad.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}});
+    }));
+
+    tasks.push_back(std::async(std::launch::async, [] {
+        string_to_spv("concat_f32", "concat.comp", {{"A_TYPE", "float"}, {"B_TYPE", "float"}, {"D_TYPE", "float"}});
+    }));
+    tasks.push_back(std::async(std::launch::async, [] {
+        string_to_spv("concat_f16", "concat.comp", {{"A_TYPE", "float16_t"}, {"B_TYPE", "float16_t"}, {"D_TYPE", "float16_t"}, {"OPTIMIZATION_ERROR_WORKAROUND", "1"}});
+    }));
+    tasks.push_back(std::async(std::launch::async, [] {
+        string_to_spv("concat_i32", "concat.comp", {{"A_TYPE", "int"}, {"B_TYPE", "int"}, {"D_TYPE", "int"}});
+    }));
+
+    tasks.push_back(std::async(std::launch::async, [] {
+        string_to_spv("upscale_f32", "upscale.comp", {{"A_TYPE", "float"}, {"B_TYPE", "float"}, {"D_TYPE", "float"}});
+    }));
+
+    tasks.push_back(std::async(std::launch::async, [] {
+        string_to_spv("gelu_f32", "gelu.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}});
+    }));
+    tasks.push_back(std::async(std::launch::async, [] {
+        string_to_spv("gelu_quick_f32", "gelu_quick.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}});
+    }));
+    tasks.push_back(std::async(std::launch::async, [] {
+        string_to_spv("silu_f32", "silu.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}});
+    }));
+    tasks.push_back(std::async(std::launch::async, [] {
+        string_to_spv("relu_f32", "relu.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}});
+    }));
+    tasks.push_back(std::async(std::launch::async, [] {
+        string_to_spv("leaky_relu_f32", "leaky_relu.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}});
+    }));
+    tasks.push_back(std::async(std::launch::async, [] {
+        string_to_spv("tanh_f32", "tanh.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}});
+    }));
+
+    tasks.push_back(std::async(std::launch::async, [] {
+        string_to_spv("diag_mask_inf_f32", "diag_mask_inf.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}});
+    }));
+
+    tasks.push_back(std::async(std::launch::async, [=] {
+        string_to_spv("soft_max_f32", "soft_max.comp", merge_maps(base_dict, {{"A_TYPE", "float"}, {"B_TYPE", "float"}, {"D_TYPE", "float"}}));
+    }));
+    tasks.push_back(std::async(std::launch::async, [=] {
+        string_to_spv("soft_max_f32_f16", "soft_max.comp", merge_maps(base_dict, {{"A_TYPE", "float"}, {"B_TYPE", "float16_t"}, {"D_TYPE", "float"}}));
+    }));
+
+    tasks.push_back(std::async(std::launch::async, [] {
+        string_to_spv("rope_norm_f32", "rope_norm.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}});
+    }));
+    tasks.push_back(std::async(std::launch::async, [] {
+        string_to_spv("rope_norm_f16", "rope_norm.comp", {{"A_TYPE", "float16_t"}, {"D_TYPE", "float16_t"}});
+    }));
+
+    tasks.push_back(std::async(std::launch::async, [] {
+        string_to_spv("rope_neox_f32", "rope_neox.comp", {{"A_TYPE", "float"}, {"D_TYPE", "float"}});
+    }));
+    tasks.push_back(std::async(std::launch::async, [] {
+        string_to_spv("rope_neox_f16", "rope_neox.comp", {{"A_TYPE", "float16_t"}, {"D_TYPE", "float16_t"}});
+    }));
+
+    tasks.push_back(std::async(std::launch::async, [] {
+        string_to_spv("argsort_f32", "argsort.comp", {{"A_TYPE", "float"}});
+    }));
+
+    tasks.push_back(std::async(std::launch::async, [=] {
+        string_to_spv("sum_rows_f32", "sum_rows.comp", merge_maps(base_dict, {{"A_TYPE", "float"}, {"D_TYPE", "float"}}));
+    }));
+
+    tasks.push_back(std::async(std::launch::async, [=] {
+        string_to_spv("im2col_f32", "im2col.comp", merge_maps(base_dict, {{"A_TYPE", "float"}, {"D_TYPE", "float"}}));
+    }));
+    tasks.push_back(std::async(std::launch::async, [=] {
+        string_to_spv("im2col_f32_f16", "im2col.comp", merge_maps(base_dict, {{"A_TYPE", "float"}, {"D_TYPE", "float16_t"}}));
+    }));
+
+    tasks.push_back(std::async(std::launch::async, [=] {
+        string_to_spv("timestep_embedding_f32", "timestep_embedding.comp", merge_maps(base_dict, {{"A_TYPE", "float"}, {"D_TYPE", "float"}}));
+    }));
+}
+
+void write_output_files() {
+    FILE* hdr = fopen(target_hpp.c_str(), "w");
+    FILE* src = fopen(target_cpp.c_str(), "w");
+
+    fprintf(hdr, "#include <cstdint>\n\n");
+    fprintf(src, "#include \"%s\"\n\n", basename(target_hpp).c_str());
+
+    for (const auto& pair : shader_fnames) {
+        const std::string& name = pair.first;
+        #ifdef _WIN32
+            std::string path = pair.second;
+            std::replace(path.begin(), path.end(), '/', '\\' );
+        #else
+            const std::string& path = pair.second;
+        #endif
+
+        FILE* spv = fopen(path.c_str(), "rb");
+        if (!spv) {
+            std::cerr << "Error opening SPIR-V file: " << path << " (" << strerror(errno) << ")\n";
+            continue;
+        }
+
+        fseek(spv, 0, SEEK_END);
+        size_t size = ftell(spv);
+        fseek(spv, 0, SEEK_SET);
+
+        std::vector<unsigned char> data(size);
+        size_t read_size = fread(data.data(), 1, size, spv);
+        fclose(spv);
+        if (read_size != size) {
+            std::cerr << "Error reading SPIR-V file: " << path << " (" << strerror(errno) << ")\n";
+            continue;
+        }
+
+        fprintf(hdr, "extern unsigned char %s_data[%zu];\n", name.c_str(), size);
+        fprintf(hdr, "const uint64_t %s_len = %zu;\n\n", name.c_str(), size);
+
+        fprintf(src, "unsigned char %s_data[%zu] = {\n", name.c_str(), size);
+        for (size_t i = 0; i < size; ++i) {
+            fprintf(src, "0x%02x,", data[i]);
+            if ((i + 1) % 12 == 0) fprintf(src, "\n");
+        }
+        fprintf(src, "\n};\n\n");
+
+        if (!no_clean) {
+            std::remove(path.c_str());
+        }
+    }
+
+    fclose(hdr);
+    fclose(src);
+}
+
+int main(int argc, char** argv) {
+    std::map<std::string, std::string> args;
+    for (int i = 1; i < argc; i += 2) {
+        if (i + 1 < argc) {
+            args[argv[i]] = argv[i + 1];
+        }
+    }
+
+    if (args.find("--glslc") != args.end()) {
+        GLSLC = args["--glslc"]; // Path to glslc
+    }
+    if (args.find("--input-dir") != args.end()) {
+        input_dir = args["--input-dir"]; // Directory containing shader sources
+    }
+    if (args.find("--output-dir") != args.end()) {
+        output_dir = args["--output-dir"]; // Directory for containing SPIR-V output
+    }
+    if (args.find("--target-hpp") != args.end()) {
+        target_hpp = args["--target-hpp"]; // Path to generated header file
+    }
+    if (args.find("--target-cpp") != args.end()) {
+        target_cpp = args["--target-cpp"]; // Path to generated cpp file
+    }
+    if (args.find("--no-clean") != args.end()) {
+        no_clean = true; // Keep temporary SPIR-V files in output-dir after build
+    }
+
+    if (!directory_exists(input_dir)) {
+        std::cerr << "\"" << input_dir << "\" must be a valid directory containing shader sources" << std::endl;
+        return EXIT_FAILURE;
+    }
+
+    if (!directory_exists(output_dir)) {
+        if (!create_directory(output_dir)) {
+            std::cerr << "Error creating output directory: " << output_dir << "\n";
+            return EXIT_FAILURE;
+        }
+    }
+
+    std::vector<std::future<void>> tasks;
+    process_shaders(tasks);
+
+    for (auto& task : tasks) {
+        task.get();
+    }
+
+    write_output_files();
+
+    return EXIT_SUCCESS;
+}
diff --git a/src/whisper_cpp/whisper.h b/src/whisper_cpp/include/whisper.h
similarity index 92%
rename from src/whisper_cpp/whisper.h
rename to src/whisper_cpp/include/whisper.h
index 35257a16..375c1ebc 100644
--- a/src/whisper_cpp/whisper.h
+++ b/src/whisper_cpp/include/whisper.h
@@ -1,4 +1,3 @@
-#include "R.h"
 #ifndef WHISPER_H
 #define WHISPER_H
 
@@ -56,14 +55,14 @@ extern "C" {
     //     struct whisper_context * ctx = whisper_init_from_file_with_params("/path/to/ggml-base.en.bin", cparams);
     //
     //     if (whisper_full(ctx, wparams, pcmf32.data(), pcmf32.size()) != 0) {
-    //         Rprintf("failed to process audio\n");
+    //         fprintf(stderr, "failed to process audio\n");
     //         return 7;
     //     }
     //
     //     const int n_segments = whisper_full_n_segments(ctx);
     //     for (int i = 0; i < n_segments; ++i) {
     //         const char * text = whisper_full_get_segment_text(ctx, i);
-    //         Rprintf("%s", text);
+    //         printf("%s", text);
     //     }
     //
     //     whisper_free(ctx);
@@ -85,8 +84,46 @@ extern "C" {
     typedef int32_t whisper_token;
     typedef int32_t whisper_seq_id;
 
+    enum whisper_alignment_heads_preset {
+        WHISPER_AHEADS_NONE,
+        WHISPER_AHEADS_N_TOP_MOST,  // All heads from the N-top-most text-layers
+        WHISPER_AHEADS_CUSTOM,
+        WHISPER_AHEADS_TINY_EN,
+        WHISPER_AHEADS_TINY,
+        WHISPER_AHEADS_BASE_EN,
+        WHISPER_AHEADS_BASE,
+        WHISPER_AHEADS_SMALL_EN,
+        WHISPER_AHEADS_SMALL,
+        WHISPER_AHEADS_MEDIUM_EN,
+        WHISPER_AHEADS_MEDIUM,
+        WHISPER_AHEADS_LARGE_V1,
+        WHISPER_AHEADS_LARGE_V2,
+        WHISPER_AHEADS_LARGE_V3,
+    };
+
+    typedef struct whisper_ahead {
+        int n_text_layer;
+        int n_head;
+    } whisper_ahead;
+
+    typedef struct whisper_aheads {
+        size_t n_heads;
+        const whisper_ahead * heads;
+    } whisper_aheads;
+
     struct whisper_context_params {
         bool  use_gpu;
+        bool  flash_attn;
+        int   gpu_device;  // CUDA device
+
+        // [EXPERIMENTAL] Token-level timestamps with DTW
+        bool dtw_token_timestamps;
+        enum whisper_alignment_heads_preset dtw_aheads_preset;
+
+        int dtw_n_top;
+        struct whisper_aheads dtw_aheads;
+
+        size_t dtw_mem_size; // TODO: remove
     };
 
     typedef struct whisper_token_data {
@@ -103,6 +140,11 @@ extern "C" {
         int64_t t0;        // start time of the token
         int64_t t1;        //   end time of the token
 
+        // [EXPERIMENTAL] Token-level timestamps with DTW
+        // do not use if you haven't computed token-level timestamps with dtw
+        // Roughly corresponds to the moment in audio in which the token was output
+        int64_t t_dtw;
+
         float vlen;        // voice length of the token
     } whisper_token_data;
 
@@ -224,22 +266,6 @@ extern "C" {
                                int   n_samples,
                                int   n_threads);
 
-    // Convert RAW PCM audio to log mel spectrogram but applies a Phase Vocoder to speed up the audio x2.
-    // The resulting spectrogram is stored inside the default state of the provided whisper context.
-    // Returns 0 on success
-    WHISPER_API int whisper_pcm_to_mel_phase_vocoder(
-        struct whisper_context * ctx,
-                   const float * samples,
-                           int   n_samples,
-                           int   n_threads);
-
-    WHISPER_API int whisper_pcm_to_mel_phase_vocoder_with_state(
-        struct whisper_context * ctx,
-          struct whisper_state * state,
-                   const float * samples,
-                           int   n_samples,
-                           int   n_threads);
-
     // This can be used to set a custom log mel spectrogram inside the default state of the provided whisper context.
     // Use this instead of whisper_pcm_to_mel() if you want to provide your own log mel spectrogram.
     // n_mel must be 80
@@ -296,7 +322,7 @@ extern "C" {
     // Convert the provided text into tokens.
     // The tokens pointer must be large enough to hold the resulting tokens.
     // Returns the number of tokens on success, no more than n_max_tokens
-    // Returns -1 on failure
+    // Returns a negative number on failure - the number of tokens that would have been returned
     // TODO: not sure if correct
     WHISPER_API int whisper_tokenize(
             struct whisper_context * ctx,
@@ -304,8 +330,12 @@ extern "C" {
                      whisper_token * tokens,
                                int   n_max_tokens);
 
+    // Return the number of tokens in the provided text
+    // Equivalent to: -whisper_tokenize(ctx, text, NULL, 0)
+    int whisper_token_count(struct whisper_context * ctx, const char * text);
+
     // Largest language id (i.e. number of available languages - 1)
-    WHISPER_API int whisper_lang_max_id();
+    WHISPER_API int whisper_lang_max_id(void);
 
     // Return the id of the specified language, returns -1 if not found
     // Examples:
@@ -412,11 +442,6 @@ extern "C" {
     // If it returns false, the computation is aborted
     typedef bool (*whisper_encoder_begin_callback)(struct whisper_context * ctx, struct whisper_state * state, void * user_data);
 
-    // Abort callback
-    // If not NULL, called before ggml computation
-    // If it returns true, the computation is aborted
-    typedef bool (*whisper_abort_callback)(void * user_data);
-
     // Logits filter callback
     // Can be used to modify the logits before sampling
     // If not NULL, called after applying temperature to logits
@@ -458,15 +483,19 @@ extern "C" {
 
         // [EXPERIMENTAL] speed-up techniques
         // note: these can significantly reduce the quality of the output
-        bool speed_up;          // speed-up the audio by 2x using Phase Vocoder
         bool debug_mode;        // enable debug_mode provides extra info (eg. Dump log_mel)
         int  audio_ctx;         // overwrite the audio context size (0 = use default)
 
         // [EXPERIMENTAL] [TDRZ] tinydiarize
         bool tdrz_enable;       // enable tinydiarize speaker turn detection
 
+        // A regular expression that matches tokens to suppress
+        const char * suppress_regex;
+
         // tokens to provide to the whisper decoder as initial prompt
         // these are prepended to any existing text context from a previous call
+        // use whisper_tokenize() to convert text to tokens
+        // maximum of whisper_n_text_ctx()/2 tokens are used (typically 224)
         const char * initial_prompt;
         const whisper_token * prompt_tokens;
         int prompt_n_tokens;
@@ -513,7 +542,7 @@ extern "C" {
         void * encoder_begin_callback_user_data;
 
         // called each time before ggml computation starts
-        whisper_abort_callback abort_callback;
+        ggml_abort_callback abort_callback;
         void * abort_callback_user_data;
 
         // called by each decoder to filter obtained logits
@@ -527,10 +556,10 @@ extern "C" {
     };
 
     // NOTE: this function allocates memory, and it is the responsibility of the caller to free the pointer - see whisper_free_context_params & whisper_free_params()
-    WHISPER_API struct whisper_context_params * whisper_context_default_params_by_ref();
-    WHISPER_API struct whisper_context_params whisper_context_default_params(void);
+    WHISPER_API struct whisper_context_params * whisper_context_default_params_by_ref(void);
+    WHISPER_API struct whisper_context_params   whisper_context_default_params       (void);
     WHISPER_API struct whisper_full_params * whisper_full_default_params_by_ref(enum whisper_sampling_strategy strategy);
-    WHISPER_API struct whisper_full_params whisper_full_default_params(enum whisper_sampling_strategy strategy);
+    WHISPER_API struct whisper_full_params   whisper_full_default_params       (enum whisper_sampling_strategy strategy);
 
     // Run the entire model: PCM -> log mel spectrogram -> encoder -> decoder -> text
     // Not thread safe for same context
diff --git a/src/whisper_cpp/spm-headers/ggml-alloc.h b/src/whisper_cpp/spm-headers/ggml-alloc.h
new file mode 100644
index 00000000..0361ffc3
--- /dev/null
+++ b/src/whisper_cpp/spm-headers/ggml-alloc.h
@@ -0,0 +1 @@
+../ggml/include/ggml-alloc.h
\ No newline at end of file
diff --git a/src/whisper_cpp/spm-headers/ggml-backend.h b/src/whisper_cpp/spm-headers/ggml-backend.h
new file mode 100644
index 00000000..7295f0f0
--- /dev/null
+++ b/src/whisper_cpp/spm-headers/ggml-backend.h
@@ -0,0 +1 @@
+../ggml/include/ggml-backend.h
\ No newline at end of file
diff --git a/src/whisper_cpp/spm-headers/ggml-metal.h b/src/whisper_cpp/spm-headers/ggml-metal.h
new file mode 100644
index 00000000..aefad5fa
--- /dev/null
+++ b/src/whisper_cpp/spm-headers/ggml-metal.h
@@ -0,0 +1 @@
+../ggml/include/ggml-metal.h
\ No newline at end of file
diff --git a/src/whisper_cpp/spm-headers/ggml.h b/src/whisper_cpp/spm-headers/ggml.h
new file mode 100644
index 00000000..0bdfeacb
--- /dev/null
+++ b/src/whisper_cpp/spm-headers/ggml.h
@@ -0,0 +1 @@
+../ggml/include/ggml.h
\ No newline at end of file
diff --git a/src/whisper_cpp/spm-headers/whisper.h b/src/whisper_cpp/spm-headers/whisper.h
new file mode 100644
index 00000000..0e0e6f22
--- /dev/null
+++ b/src/whisper_cpp/spm-headers/whisper.h
@@ -0,0 +1 @@
+../include/whisper.h
\ No newline at end of file
diff --git a/src/whisper_cpp/src/CMakeLists.txt b/src/whisper_cpp/src/CMakeLists.txt
new file mode 100644
index 00000000..ff54d645
--- /dev/null
+++ b/src/whisper_cpp/src/CMakeLists.txt
@@ -0,0 +1,118 @@
+# TODO: should not use this
+if (WIN32)
+    add_compile_definitions(_CRT_SECURE_NO_WARNINGS)
+
+    if (BUILD_SHARED_LIBS)
+        set(CMAKE_WINDOWS_EXPORT_ALL_SYMBOLS ON)
+    endif()
+endif()
+
+if (WHISPER_COREML)
+    find_library(FOUNDATION_FRAMEWORK Foundation)
+    find_library(COREML_FRAMEWORK CoreML)
+
+    if (COREML_FRAMEWORK)
+        message(STATUS "CoreML framework found")
+
+        set(WHISPER_EXTRA_FLAGS ${WHISPER_EXTRA_FLAGS} -DWHISPER_USE_COREML)
+    else()
+        message(FATAL_ERROR "CoreML framework not found")
+    endif()
+
+    if (WHISPER_COREML_ALLOW_FALLBACK)
+        set(WHISPER_EXTRA_FLAGS ${WHISPER_EXTRA_FLAGS} -DWHISPER_COREML_ALLOW_FALLBACK)
+    endif()
+endif()
+
+if (WHISPER_OPENVINO)
+    find_package(OpenVINO REQUIRED COMPONENTS Runtime)
+endif()
+
+#
+# libraries
+#
+
+# whisper.coreml
+
+if (WHISPER_COREML)
+    set(TARGET whisper.coreml)
+
+    add_library(${TARGET}
+        coreml/whisper-encoder.h
+        coreml/whisper-encoder.mm
+        coreml/whisper-encoder-impl.h
+        coreml/whisper-encoder-impl.m
+        )
+
+    include(DefaultTargetOptions)
+
+    target_include_directories(${TARGET} PUBLIC
+        .
+        )
+
+    target_link_libraries(${TARGET} PRIVATE ${FOUNDATION_FRAMEWORK} ${COREML_FRAMEWORK})
+
+    set_target_properties(${TARGET} PROPERTIES
+        COMPILE_FLAGS "-fobjc-arc"
+        XCODE_ATTRIBUTE_CLANG_ENABLE_OBJC_ARC YES
+        )
+    set_target_properties(${TARGET} PROPERTIES FOLDER "libs")
+endif()
+
+if (WHISPER_OPENVINO)
+    set(TARGET whisper.openvino)
+
+    add_library(${TARGET} OBJECT
+        openvino/whisper-openvino-encoder.h
+        openvino/whisper-openvino-encoder.cpp
+        )
+
+    target_include_directories(${TARGET} PUBLIC
+        .
+        )
+
+    set_property(TARGET ${TARGET} PROPERTY POSITION_INDEPENDENT_CODE ON)
+    set(WHISPER_EXTRA_FLAGS ${WHISPER_EXTRA_FLAGS} -DWHISPER_USE_OPENVINO)
+
+    target_link_libraries(${TARGET} PRIVATE ggml openvino::runtime)
+    set_target_properties(${TARGET} PROPERTIES FOLDER "libs")
+endif()
+
+# whisper
+
+add_library(whisper
+            ../include/whisper.h
+            whisper.cpp
+            )
+
+# Set the version numbers
+set_target_properties(whisper PROPERTIES
+    VERSION ${PROJECT_VERSION}
+    SOVERSION ${SOVERSION}
+)
+
+target_include_directories(whisper PUBLIC . ../include)
+target_compile_features   (whisper PUBLIC cxx_std_11) # don't bump
+
+if (WHISPER_EXTRA_FLAGS)
+    target_compile_options(whisper PRIVATE ${WHISPER_EXTRA_FLAGS})
+endif()
+
+target_link_libraries(whisper PUBLIC ggml)
+
+if (WHISPER_COREML)
+    target_link_libraries(whisper PRIVATE whisper.coreml)
+endif()
+
+if (WHISPER_OPENVINO)
+    target_link_libraries(whisper PRIVATE whisper.openvino)
+endif()
+
+if (WHISPER_MKL)
+    target_link_libraries(whisper PRIVATE MKL::MKL)
+endif()
+
+if (BUILD_SHARED_LIBS)
+    set_target_properties(whisper PROPERTIES POSITION_INDEPENDENT_CODE ON)
+    target_compile_definitions(whisper PRIVATE WHISPER_SHARED WHISPER_BUILD)
+endif()
diff --git a/src/whisper_cpp/coreml/whisper-decoder-impl.h b/src/whisper_cpp/src/coreml/whisper-decoder-impl.h
similarity index 100%
rename from src/whisper_cpp/coreml/whisper-decoder-impl.h
rename to src/whisper_cpp/src/coreml/whisper-decoder-impl.h
diff --git a/src/whisper_cpp/coreml/whisper-decoder-impl.m b/src/whisper_cpp/src/coreml/whisper-decoder-impl.m
similarity index 100%
rename from src/whisper_cpp/coreml/whisper-decoder-impl.m
rename to src/whisper_cpp/src/coreml/whisper-decoder-impl.m
diff --git a/src/whisper_cpp/coreml/whisper-encoder-impl.h b/src/whisper_cpp/src/coreml/whisper-encoder-impl.h
similarity index 100%
rename from src/whisper_cpp/coreml/whisper-encoder-impl.h
rename to src/whisper_cpp/src/coreml/whisper-encoder-impl.h
diff --git a/src/whisper_cpp/coreml/whisper-encoder-impl.m b/src/whisper_cpp/src/coreml/whisper-encoder-impl.m
similarity index 100%
rename from src/whisper_cpp/coreml/whisper-encoder-impl.m
rename to src/whisper_cpp/src/coreml/whisper-encoder-impl.m
diff --git a/src/whisper_cpp/coreml/whisper-encoder.h b/src/whisper_cpp/src/coreml/whisper-encoder.h
similarity index 100%
rename from src/whisper_cpp/coreml/whisper-encoder.h
rename to src/whisper_cpp/src/coreml/whisper-encoder.h
diff --git a/src/whisper_cpp/coreml/whisper-encoder.mm b/src/whisper_cpp/src/coreml/whisper-encoder.mm
similarity index 100%
rename from src/whisper_cpp/coreml/whisper-encoder.mm
rename to src/whisper_cpp/src/coreml/whisper-encoder.mm
diff --git a/src/whisper_cpp/src/ggml-cpu-impl.h b/src/whisper_cpp/src/ggml-cpu-impl.h
new file mode 100644
index 00000000..5b45155b
--- /dev/null
+++ b/src/whisper_cpp/src/ggml-cpu-impl.h
@@ -0,0 +1,614 @@
+#pragma once
+
+// GGML CPU internal header
+
+#include "ggml.h"
+#include "ggml-impl.h"
+#include <stdlib.h> // load `stdlib.h` before other headers to work around MinGW bug: https://sourceforge.net/p/mingw-w64/bugs/192/
+//#include <stddef.h>
+#include <stdbool.h>
+#include <string.h> // memcpy
+#include <math.h>   // fabsf
+
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+#if defined(_MSC_VER)
+
+#define m512bh(p) p
+#define m512i(p) p
+
+#else
+
+#define m512bh(p) (__m512bh)(p)
+#define m512i(p) (__m512i)(p)
+
+#endif
+
+/**
+ * Converts brain16 to float32.
+ *
+ * The bfloat16 floating point format has the following structure:
+ *
+ *       ┌sign
+ *       │
+ *       │   ┌exponent
+ *       │   │
+ *       │   │      ┌mantissa
+ *       │   │      │
+ *       │┌──┴───┐┌─┴───┐
+ *     0b0000000000000000 brain16
+ *
+ * Since bf16 has the same number of exponent bits as a 32bit float,
+ * encoding and decoding numbers becomes relatively straightforward.
+ *
+ *       ┌sign
+ *       │
+ *       │   ┌exponent
+ *       │   │
+ *       │   │      ┌mantissa
+ *       │   │      │
+ *       │┌──┴───┐┌─┴───────────────────┐
+ *     0b00000000000000000000000000000000 IEEE binary32
+ *
+ * For comparison, the standard fp16 format has fewer exponent bits.
+ *
+ *       ┌sign
+ *       │
+ *       │  ┌exponent
+ *       │  │
+ *       │  │    ┌mantissa
+ *       │  │    │
+ *       │┌─┴─┐┌─┴──────┐
+ *     0b0000000000000000 IEEE binary16
+ *
+ * @see IEEE 754-2008
+ */
+static inline float ggml_compute_bf16_to_fp32(ggml_bf16_t h) {
+    union {
+        float f;
+        uint32_t i;
+    } u;
+    u.i = (uint32_t)h.bits << 16;
+    return u.f;
+}
+
+/**
+ * Converts float32 to brain16.
+ *
+ * This is binary identical with Google Brain float conversion.
+ * Floats shall round to nearest even, and NANs shall be quiet.
+ * Subnormals aren't flushed to zero, except perhaps when used.
+ * This code should vectorize nicely if using modern compilers.
+ */
+static inline ggml_bf16_t ggml_compute_fp32_to_bf16(float s) {
+    ggml_bf16_t h;
+    union {
+        float f;
+        uint32_t i;
+    } u;
+    u.f = s;
+    if ((u.i & 0x7fffffff) > 0x7f800000) { /* nan */
+        h.bits = (u.i >> 16) | 64; /* force to quiet */
+        return h;
+    }
+    h.bits = (u.i + (0x7fff + ((u.i >> 16) & 1))) >> 16;
+    return h;
+}
+
+#define GGML_FP32_TO_BF16(x) ggml_compute_fp32_to_bf16(x)
+#define GGML_BF16_TO_FP32(x) ggml_compute_bf16_to_fp32(x)
+
+// __FMA__ and __F16C__ are not defined in MSVC, however they are implied with AVX2/AVX512
+#if defined(_MSC_VER) && (defined(__AVX2__) || defined(__AVX512F__))
+#ifndef __FMA__
+#define __FMA__
+#endif
+#ifndef __F16C__
+#define __F16C__
+#endif
+#endif
+
+// __SSE3__ and __SSSE3__ are not defined in MSVC, but SSE3/SSSE3 are present when AVX/AVX2/AVX512 are available
+#if defined(_MSC_VER) && (defined(__AVX__) || defined(__AVX2__) || defined(__AVX512F__))
+#ifndef __SSE3__
+#define __SSE3__
+#endif
+#ifndef __SSSE3__
+#define __SSSE3__
+#endif
+#endif
+
+#if defined(__ARM_FEATURE_SVE)
+#include <arm_sve.h>
+#include <sys/prctl.h>
+#endif
+
+// 16-bit float
+// on Arm, we use __fp16
+// on x86, we use uint16_t
+#if defined(__ARM_NEON)
+
+// if YCM cannot find <arm_neon.h>, make a symbolic link to it, for example:
+//
+//   $ ln -sfn /Library/Developer/CommandLineTools/usr/lib/clang/13.1.6/include/arm_neon.h ./src/
+//
+#include <arm_neon.h>
+
+#ifdef _MSC_VER
+
+typedef uint16_t ggml_fp16_internal_t;
+
+#define ggml_vld1q_u32(w,x,y,z) { ((w) + ((uint64_t)(x) << 32)), ((y) + ((uint64_t)(z) << 32)) }
+
+#else
+
+typedef __fp16 ggml_fp16_internal_t;
+
+#define ggml_vld1q_u32(w,x,y,z) { (w), (x), (y), (z) }
+
+#endif // _MSC_VER
+
+#if !defined(__aarch64__)
+
+// 32-bit ARM compatibility
+
+// vaddlvq_s16
+// vpaddq_s16
+// vpaddq_s32
+// vaddvq_s32
+// vaddvq_f32
+// vmaxvq_f32
+// vcvtnq_s32_f32
+// vzip1_u8
+// vzip2_u8
+
+inline static int32_t vaddlvq_s16(int16x8_t v) {
+    int32x4_t v0 = vreinterpretq_s32_s64(vpaddlq_s32(vpaddlq_s16(v)));
+    return vgetq_lane_s32(v0, 0) + vgetq_lane_s32(v0, 2);
+}
+
+inline static int16x8_t vpaddq_s16(int16x8_t a, int16x8_t b) {
+    int16x4_t a0 = vpadd_s16(vget_low_s16(a), vget_high_s16(a));
+    int16x4_t b0 = vpadd_s16(vget_low_s16(b), vget_high_s16(b));
+    return vcombine_s16(a0, b0);
+}
+
+inline static int32x4_t vpaddq_s32(int32x4_t a, int32x4_t b) {
+    int32x2_t a0 = vpadd_s32(vget_low_s32(a), vget_high_s32(a));
+    int32x2_t b0 = vpadd_s32(vget_low_s32(b), vget_high_s32(b));
+    return vcombine_s32(a0, b0);
+}
+
+inline static int32_t vaddvq_s32(int32x4_t v) {
+    return vgetq_lane_s32(v, 0) + vgetq_lane_s32(v, 1) + vgetq_lane_s32(v, 2) + vgetq_lane_s32(v, 3);
+}
+
+inline static float vaddvq_f32(float32x4_t v) {
+    return vgetq_lane_f32(v, 0) + vgetq_lane_f32(v, 1) + vgetq_lane_f32(v, 2) + vgetq_lane_f32(v, 3);
+}
+
+inline static float vmaxvq_f32(float32x4_t v) {
+    return
+        MAX(MAX(vgetq_lane_f32(v, 0), vgetq_lane_f32(v, 1)),
+            MAX(vgetq_lane_f32(v, 2), vgetq_lane_f32(v, 3)));
+}
+
+inline static int32x4_t vcvtnq_s32_f32(float32x4_t v) {
+    int32x4_t res;
+
+    res[0] = roundf(vgetq_lane_f32(v, 0));
+    res[1] = roundf(vgetq_lane_f32(v, 1));
+    res[2] = roundf(vgetq_lane_f32(v, 2));
+    res[3] = roundf(vgetq_lane_f32(v, 3));
+
+    return res;
+}
+
+inline static uint8x8_t vzip1_u8(uint8x8_t a, uint8x8_t b) {
+    uint8x8_t res;
+
+    res[0] = a[0]; res[1] = b[0];
+    res[2] = a[1]; res[3] = b[1];
+    res[4] = a[2]; res[5] = b[2];
+    res[6] = a[3]; res[7] = b[3];
+
+    return res;
+}
+
+inline static uint8x8_t vzip2_u8(uint8x8_t a, uint8x8_t b) {
+    uint8x8_t res;
+
+    res[0] = a[4]; res[1] = b[4];
+    res[2] = a[5]; res[3] = b[5];
+    res[4] = a[6]; res[5] = b[6];
+    res[6] = a[7]; res[7] = b[7];
+
+    return res;
+}
+
+// vld1q_s16_x2
+// vld1q_u8_x2
+// vld1q_u8_x4
+// vld1q_s8_x2
+// vld1q_s8_x4
+// TODO: double-check these work correctly
+
+typedef struct ggml_int16x8x2_t {
+    int16x8_t val[2];
+} ggml_int16x8x2_t;
+
+inline static ggml_int16x8x2_t ggml_vld1q_s16_x2(const int16_t * ptr) {
+    ggml_int16x8x2_t res;
+
+    res.val[0] = vld1q_s16(ptr + 0);
+    res.val[1] = vld1q_s16(ptr + 8);
+
+    return res;
+}
+
+typedef struct ggml_uint8x16x2_t {
+    uint8x16_t val[2];
+} ggml_uint8x16x2_t;
+
+inline static ggml_uint8x16x2_t ggml_vld1q_u8_x2(const uint8_t * ptr) {
+    ggml_uint8x16x2_t res;
+
+    res.val[0] = vld1q_u8(ptr + 0);
+    res.val[1] = vld1q_u8(ptr + 16);
+
+    return res;
+}
+
+typedef struct ggml_uint8x16x4_t {
+    uint8x16_t val[4];
+} ggml_uint8x16x4_t;
+
+inline static ggml_uint8x16x4_t ggml_vld1q_u8_x4(const uint8_t * ptr) {
+    ggml_uint8x16x4_t res;
+
+    res.val[0] = vld1q_u8(ptr + 0);
+    res.val[1] = vld1q_u8(ptr + 16);
+    res.val[2] = vld1q_u8(ptr + 32);
+    res.val[3] = vld1q_u8(ptr + 48);
+
+    return res;
+}
+
+typedef struct ggml_int8x16x2_t {
+    int8x16_t val[2];
+} ggml_int8x16x2_t;
+
+inline static ggml_int8x16x2_t ggml_vld1q_s8_x2(const int8_t * ptr) {
+    ggml_int8x16x2_t res;
+
+    res.val[0] = vld1q_s8(ptr + 0);
+    res.val[1] = vld1q_s8(ptr + 16);
+
+    return res;
+}
+
+typedef struct ggml_int8x16x4_t {
+    int8x16_t val[4];
+} ggml_int8x16x4_t;
+
+inline static ggml_int8x16x4_t ggml_vld1q_s8_x4(const int8_t * ptr) {
+    ggml_int8x16x4_t res;
+
+    res.val[0] = vld1q_s8(ptr + 0);
+    res.val[1] = vld1q_s8(ptr + 16);
+    res.val[2] = vld1q_s8(ptr + 32);
+    res.val[3] = vld1q_s8(ptr + 48);
+
+    return res;
+}
+
+// NOTE: not tested
+inline static int8x16_t ggml_vqtbl1q_s8(int8x16_t a, uint8x16_t b) {
+    int8x16_t res;
+
+    res[ 0] = a[b[ 0]];
+    res[ 1] = a[b[ 1]];
+    res[ 2] = a[b[ 2]];
+    res[ 3] = a[b[ 3]];
+    res[ 4] = a[b[ 4]];
+    res[ 5] = a[b[ 5]];
+    res[ 6] = a[b[ 6]];
+    res[ 7] = a[b[ 7]];
+    res[ 8] = a[b[ 8]];
+    res[ 9] = a[b[ 9]];
+    res[10] = a[b[10]];
+    res[11] = a[b[11]];
+    res[12] = a[b[12]];
+    res[13] = a[b[13]];
+    res[14] = a[b[14]];
+    res[15] = a[b[15]];
+
+    return res;
+}
+
+// NOTE: not tested
+inline static uint8x16_t ggml_vqtbl1q_u8(uint8x16_t a, uint8x16_t b) {
+    uint8x16_t res;
+
+    res[ 0] = a[b[ 0]];
+    res[ 1] = a[b[ 1]];
+    res[ 2] = a[b[ 2]];
+    res[ 3] = a[b[ 3]];
+    res[ 4] = a[b[ 4]];
+    res[ 5] = a[b[ 5]];
+    res[ 6] = a[b[ 6]];
+    res[ 7] = a[b[ 7]];
+    res[ 8] = a[b[ 8]];
+    res[ 9] = a[b[ 9]];
+    res[10] = a[b[10]];
+    res[11] = a[b[11]];
+    res[12] = a[b[12]];
+    res[13] = a[b[13]];
+    res[14] = a[b[14]];
+    res[15] = a[b[15]];
+
+    return res;
+}
+
+#else
+
+#define ggml_int16x8x2_t  int16x8x2_t
+#define ggml_uint8x16x2_t uint8x16x2_t
+#define ggml_uint8x16x4_t uint8x16x4_t
+#define ggml_int8x16x2_t  int8x16x2_t
+#define ggml_int8x16x4_t  int8x16x4_t
+
+#define ggml_vld1q_s16_x2 vld1q_s16_x2
+#define ggml_vld1q_u8_x2  vld1q_u8_x2
+#define ggml_vld1q_u8_x4  vld1q_u8_x4
+#define ggml_vld1q_s8_x2  vld1q_s8_x2
+#define ggml_vld1q_s8_x4  vld1q_s8_x4
+#define ggml_vqtbl1q_s8   vqtbl1q_s8
+#define ggml_vqtbl1q_u8   vqtbl1q_u8
+
+#endif // !defined(__aarch64__)
+
+#if !defined(__ARM_FEATURE_DOTPROD)
+
+inline static int32x4_t ggml_vdotq_s32(int32x4_t acc, int8x16_t a, int8x16_t b) {
+    const int16x8_t p0 = vmull_s8(vget_low_s8 (a), vget_low_s8 (b));
+    const int16x8_t p1 = vmull_s8(vget_high_s8(a), vget_high_s8(b));
+
+    return vaddq_s32(acc, vaddq_s32(vpaddlq_s16(p0), vpaddlq_s16(p1)));
+}
+
+#else
+
+#define ggml_vdotq_s32(a, b, c) vdotq_s32(a, b, c)
+
+#endif // !defined(__ARM_FEATURE_DOTPROD)
+
+#endif // defined(__ARM_NEON)
+
+#if defined(__ARM_NEON) && !defined(_MSC_VER)
+
+#define GGML_COMPUTE_FP16_TO_FP32(x) ggml_compute_fp16_to_fp32(x)
+#define GGML_COMPUTE_FP32_TO_FP16(x) ggml_compute_fp32_to_fp16(x)
+
+#define GGML_FP16_TO_FP32(x) ggml_compute_fp16_to_fp32(x)
+
+static inline float ggml_compute_fp16_to_fp32(ggml_fp16_t h) {
+    ggml_fp16_internal_t tmp;
+    memcpy(&tmp, &h, sizeof(ggml_fp16_t));
+    return (float)tmp;
+}
+
+static inline ggml_fp16_t ggml_compute_fp32_to_fp16(float f) {
+    ggml_fp16_t res;
+    ggml_fp16_internal_t tmp = f;
+    memcpy(&res, &tmp, sizeof(ggml_fp16_t));
+    return res;
+}
+
+#else
+
+#ifdef __wasm_simd128__
+#include <wasm_simd128.h>
+#else
+#ifdef __POWER9_VECTOR__
+#include <altivec.h>
+#undef bool
+#define bool _Bool
+#else
+#if defined(_MSC_VER) || defined(__MINGW32__)
+#include <intrin.h>
+#else
+#if defined(__AVX__) || defined(__AVX2__) || defined(__AVX512F__) || defined(__SSSE3__) || defined(__SSE3__) || defined(__SSE__)
+#if !defined(__riscv)
+#include <immintrin.h>
+#endif
+#endif
+#endif
+#endif
+#endif
+
+#ifdef __riscv_v_intrinsic
+#include <riscv_vector.h>
+#endif
+
+#if defined(__loongarch64)
+#if defined(__loongarch_asx)
+#include <lasxintrin.h>
+#endif
+#if defined(__loongarch_sx)
+#include <lsxintrin.h>
+#endif
+#endif
+
+#if defined(__loongarch_asx)
+
+typedef union {
+    int32_t i;
+    float f;
+} ft_union;
+
+/* float type data load instructions */
+static __m128 __lsx_vreplfr2vr_s(float val) {
+    ft_union fi_tmpval = {.f = val};
+    return (__m128)__lsx_vreplgr2vr_w(fi_tmpval.i);
+}
+
+static __m256 __lasx_xvreplfr2vr_s(float val) {
+    ft_union fi_tmpval = {.f = val};
+    return (__m256)__lasx_xvreplgr2vr_w(fi_tmpval.i);
+}
+#endif
+
+#ifdef __F16C__
+
+#ifdef _MSC_VER
+#define GGML_COMPUTE_FP16_TO_FP32(x) _mm_cvtss_f32(_mm_cvtph_ps(_mm_cvtsi32_si128(x)))
+#define GGML_COMPUTE_FP32_TO_FP16(x) _mm_extract_epi16(_mm_cvtps_ph(_mm_set_ss(x), 0), 0)
+#else
+#define GGML_COMPUTE_FP16_TO_FP32(x) _cvtsh_ss(x)
+#define GGML_COMPUTE_FP32_TO_FP16(x) _cvtss_sh(x, 0)
+#endif
+
+#elif defined(__POWER9_VECTOR__)
+
+#define GGML_COMPUTE_FP16_TO_FP32(x) ggml_compute_fp16_to_fp32(x)
+#define GGML_COMPUTE_FP32_TO_FP16(x) ggml_compute_fp32_to_fp16(x)
+/* the inline asm below is about 12% faster than the lookup method */
+#define GGML_FP16_TO_FP32(x) GGML_COMPUTE_FP16_TO_FP32(x)
+#define GGML_FP32_TO_FP16(x) GGML_COMPUTE_FP32_TO_FP16(x)
+
+static inline float ggml_compute_fp16_to_fp32(ggml_fp16_t h) {
+    register float f;
+    register double d;
+    __asm__(
+        "mtfprd %0,%2\n"
+        "xscvhpdp %0,%0\n"
+        "frsp %1,%0\n" :
+        /* temp */ "=d"(d),
+        /* out */  "=f"(f):
+        /* in */   "r"(h));
+    return f;
+}
+
+static inline ggml_fp16_t ggml_compute_fp32_to_fp16(float f) {
+    register double d;
+    register ggml_fp16_t r;
+    __asm__( /* xscvdphp can work on double or single precision */
+        "xscvdphp %0,%2\n"
+        "mffprd %1,%0\n" :
+        /* temp */ "=d"(d),
+        /* out */  "=r"(r):
+        /* in */   "f"(f));
+    return r;
+}
+
+#else
+
+// FP16 <-> FP32
+// ref: https://github.com/Maratyszcza/FP16
+
+static inline float fp32_from_bits(uint32_t w) {
+    union {
+        uint32_t as_bits;
+        float as_value;
+    } fp32;
+    fp32.as_bits = w;
+    return fp32.as_value;
+}
+
+static inline uint32_t fp32_to_bits(float f) {
+    union {
+        float as_value;
+        uint32_t as_bits;
+    } fp32;
+    fp32.as_value = f;
+    return fp32.as_bits;
+}
+
+static inline float ggml_compute_fp16_to_fp32(ggml_fp16_t h) {
+    const uint32_t w = (uint32_t) h << 16;
+    const uint32_t sign = w & UINT32_C(0x80000000);
+    const uint32_t two_w = w + w;
+
+    const uint32_t exp_offset = UINT32_C(0xE0) << 23;
+#if defined(__STDC_VERSION__) && (__STDC_VERSION__ >= 199901L) || defined(__GNUC__) && !defined(__STRICT_ANSI__)
+    const float exp_scale = 0x1.0p-112f;
+#else
+    const float exp_scale = fp32_from_bits(UINT32_C(0x7800000));
+#endif
+    const float normalized_value = fp32_from_bits((two_w >> 4) + exp_offset) * exp_scale;
+
+    const uint32_t magic_mask = UINT32_C(126) << 23;
+    const float magic_bias = 0.5f;
+    const float denormalized_value = fp32_from_bits((two_w >> 17) | magic_mask) - magic_bias;
+
+    const uint32_t denormalized_cutoff = UINT32_C(1) << 27;
+    const uint32_t result = sign |
+        (two_w < denormalized_cutoff ? fp32_to_bits(denormalized_value) : fp32_to_bits(normalized_value));
+    return fp32_from_bits(result);
+}
+
+static inline ggml_fp16_t ggml_compute_fp32_to_fp16(float f) {
+#if defined(__STDC_VERSION__) && (__STDC_VERSION__ >= 199901L) || defined(__GNUC__) && !defined(__STRICT_ANSI__)
+    const float scale_to_inf = 0x1.0p+112f;
+    const float scale_to_zero = 0x1.0p-110f;
+#else
+    const float scale_to_inf = fp32_from_bits(UINT32_C(0x77800000));
+    const float scale_to_zero = fp32_from_bits(UINT32_C(0x08800000));
+#endif
+    float base = (fabsf(f) * scale_to_inf) * scale_to_zero;
+
+    const uint32_t w = fp32_to_bits(f);
+    const uint32_t shl1_w = w + w;
+    const uint32_t sign = w & UINT32_C(0x80000000);
+    uint32_t bias = shl1_w & UINT32_C(0xFF000000);
+    if (bias < UINT32_C(0x71000000)) {
+        bias = UINT32_C(0x71000000);
+    }
+
+    base = fp32_from_bits((bias >> 1) + UINT32_C(0x07800000)) + base;
+    const uint32_t bits = fp32_to_bits(base);
+    const uint32_t exp_bits = (bits >> 13) & UINT32_C(0x00007C00);
+    const uint32_t mantissa_bits = bits & UINT32_C(0x00000FFF);
+    const uint32_t nonsign = exp_bits + mantissa_bits;
+    return (sign >> 16) | (shl1_w > UINT32_C(0xFF000000) ? UINT16_C(0x7E00) : nonsign);
+}
+
+#define GGML_COMPUTE_FP16_TO_FP32(x) ggml_compute_fp16_to_fp32(x)
+#define GGML_COMPUTE_FP32_TO_FP16(x) ggml_compute_fp32_to_fp16(x)
+
+#endif // __F16C__
+
+#endif // defined(__ARM_NEON) && (!defined(__MSC_VER)
+
+#ifdef __ARM_FEATURE_SVE
+#include <arm_sve.h>
+#endif // __ARM_FEATURE_SVE
+
+// precomputed f32 table for f16 (256 KB)
+// defined in ggml.c, initialized in ggml_init()
+extern float ggml_table_f32_f16[1 << 16];
+
+// On ARM NEON, it's quicker to directly convert x -> x instead of calling into ggml_lookup_fp16_to_fp32,
+// so we define GGML_FP16_TO_FP32 and GGML_FP32_TO_FP16 elsewhere for NEON.
+// This is also true for POWER9.
+#if !defined(GGML_FP16_TO_FP32)
+inline static float ggml_lookup_fp16_to_fp32(ggml_fp16_t f) {
+    uint16_t s;
+    memcpy(&s, &f, sizeof(uint16_t));
+    return ggml_table_f32_f16[s];
+}
+
+#define GGML_FP16_TO_FP32(x) ggml_lookup_fp16_to_fp32(x)
+#endif
+
+#if !defined(GGML_FP32_TO_FP16)
+#define GGML_FP32_TO_FP16(x) GGML_COMPUTE_FP32_TO_FP16(x)
+#endif
+
+#ifdef __cplusplus
+}
+#endif
diff --git a/src/whisper_cpp/src/openvino/whisper-openvino-encoder.cpp b/src/whisper_cpp/src/openvino/whisper-openvino-encoder.cpp
new file mode 100644
index 00000000..4d9ce122
--- /dev/null
+++ b/src/whisper_cpp/src/openvino/whisper-openvino-encoder.cpp
@@ -0,0 +1,108 @@
+#include "openvino/whisper-openvino-encoder.h"
+#include "ggml.h"
+#include <openvino/openvino.hpp>
+#include <iostream>
+
+struct whisper_openvino_context {
+    ov::InferRequest inferRequest;
+};
+
+struct whisper_openvino_context * whisper_openvino_init(const char* path_model,
+    const char* device,
+    const char* cache_dir)
+{
+    if (!path_model || !device) {
+        fprintf(stderr, "%s: path_model and/or device is null\n", __func__);
+        return nullptr;
+    }
+
+    fprintf(stderr, "%s: path_model = %s, device = %s, cache_dir = %s\n",
+        __func__, path_model, device, cache_dir ? cache_dir : "(not set)");
+
+	whisper_openvino_context *context = new whisper_openvino_context;
+    try {
+        ov::Core core;
+
+        if (cache_dir) {
+            // enables caching of device-specific 'blobs' during core.compile_model
+            // routine. This speeds up calls to compile_model for successive runs.
+            core.set_property(ov::cache_dir(cache_dir));
+        }
+
+        //Read the OpenVINO encoder IR (.xml/.bin) from disk, producing an ov::Model object.
+        std::shared_ptr<ov::Model> model = core.read_model(path_model);
+
+        // Produce a compiled-model object, given the device ("CPU", "GPU", etc.)
+        auto compiledModel = core.compile_model(model, device);
+
+        // From the compiled model object, create an infer request. This is the thing that we
+        //  we will use later on to trigger inference execution.
+        context->inferRequest = compiledModel.create_infer_request();
+    }
+    catch (const std::exception& error) {
+        std::cout << "in openvino encoder compile routine: exception: " << error.what() << std::endl;
+        delete context;
+        context = nullptr;
+    }
+
+    return context;
+}
+
+void whisper_openvino_free(struct whisper_openvino_context * ctx) {
+    if( ctx ) {
+        delete ctx;
+    }
+}
+
+int whisper_openvino_encode(
+    whisper_openvino_context* ctx,
+    ggml_tensor* mel,
+    ggml_tensor* out) {
+
+    if (!ctx || !mel || !out) {
+        fprintf(stderr, "%s: Error! ctx / mel / out is null\n", __func__);
+        return 0;
+    }
+
+    if (ggml_n_dims(mel) != 2) {
+        fprintf(stderr, "%s: Error! mel ggml_tensor expected to have n_dims=2, but it has n_dims=%d\n",
+            __func__, ggml_n_dims(mel));
+        return 0;
+    }
+
+    if (ggml_n_dims(out) != 2) {
+        fprintf(stderr, "%s: Error! out ggml_tensor expected to have n_dims=2, but it has n_dims=%d\n",
+            __func__, ggml_n_dims(out));
+        return 0;
+    }
+
+    try {
+
+        //wrap the passed-in mel ggml_tensor as an OpenVINO Tensor object, and set as input tensor to infer request
+        {
+            // note, we populate shape & stride dimensions in opposite order from how they are listed in ne / nb arrays
+            ov::Shape input_shape = { 1, (unsigned long long)mel->ne[1], (unsigned long long)mel->ne[0] };
+            ov::Strides input_strides = { mel->nb[2], mel->nb[1], mel->nb[0] };
+            ov::Tensor input_tensor(ov::element::f32, input_shape, mel->data, input_strides);
+            ctx->inferRequest.set_input_tensor(input_tensor);
+        }
+
+        //wrap the passed-in out ggml_tensor as an OpenVINO Tensor object, and set as output tensor to infer request
+        {
+            // note, we populate shape & stride dimensions in opposite order from how they are listed in ne / nb arrays
+            ov::Shape output_shape = { 1, (unsigned long long)out->ne[1], (unsigned long long)out->ne[0] };
+            ov::Strides output_strides = { out->nb[2], out->nb[1], out->nb[0] };
+            ov::Tensor out_tensor(ov::element::f32, output_shape, out->data, output_strides);
+            ctx->inferRequest.set_output_tensor(out_tensor);
+        }
+
+        //run inference
+        ctx->inferRequest.infer();
+    }
+    catch (const std::exception& error) {
+        std::cout << "in openvino encode inference execution routine: exception: " << error.what() << std::endl;
+        return 0;
+    }
+
+    return 1;
+}
diff --git a/src/whisper_cpp/src/openvino/whisper-openvino-encoder.h b/src/whisper_cpp/src/openvino/whisper-openvino-encoder.h
new file mode 100644
index 00000000..7c2f6dfc
--- /dev/null
+++ b/src/whisper_cpp/src/openvino/whisper-openvino-encoder.h
@@ -0,0 +1,31 @@
+// Wrapper of the OpenVINO Whisper Encoder model
+//
+
+#if __cplusplus
+extern "C" {
+#endif
+
+struct whisper_openvino_context;
+
+// initialize openvino encoder, given path to model xml, device ("CPU", "GPU", etc.), and
+// path to cache_dir. Returns null upon failure.
+struct whisper_openvino_context * whisper_openvino_init(const char * path_model,
+                                                        const char * device,
+                                                        const char * cache_dir);
+
+// clean up a ctx previously returned from whisper_openvino_init()
+void whisper_openvino_free(struct whisper_openvino_context * ctx);
+
+struct ggml_tensor;
+
+// Perform encode using OpenVINO.
+// Returns 1 on success
+// Returns 0 on failure
+int whisper_openvino_encode(
+    whisper_openvino_context* ctx,
+    ggml_tensor* mel,
+    ggml_tensor* out);
+
+#if __cplusplus
+}
+#endif
diff --git a/src/whisper_cpp/whisper.cpp b/src/whisper_cpp/src/whisper.cpp
similarity index 77%
rename from src/whisper_cpp/whisper.cpp
rename to src/whisper_cpp/src/whisper.cpp
index ab066c1a..31c6861e 100644
--- a/src/whisper_cpp/whisper.cpp
+++ b/src/whisper_cpp/src/whisper.cpp
@@ -1,4 +1,3 @@
-#include <Rcpp.h>
 #include "whisper.h"
 
 #ifdef WHISPER_USE_COREML
@@ -9,14 +8,30 @@
 #include "ggml-metal.h"
 #endif
 
-#ifdef GGML_USE_CUBLAS
+#ifdef GGML_USE_CUDA
 #include "ggml-cuda.h"
 #endif
 
+#ifdef GGML_USE_SYCL
+#include "ggml-sycl.h"
+#endif
+
+#ifdef GGML_USE_VULKAN
+#include "ggml-vulkan.h"
+#endif
+
+#ifdef GGML_USE_BLAS
+#include "ggml-blas.h"
+#endif
+
 #ifdef WHISPER_USE_OPENVINO
 #include "openvino/whisper-openvino-encoder.h"
 #endif
 
+#ifdef GGML_USE_CANN
+#include "ggml-cann.h"
+#endif
+
 #include "ggml.h"
 #include "ggml-alloc.h"
 #include "ggml-backend.h"
@@ -38,6 +53,7 @@
 #include <regex>
 #include <random>
 #include <functional>
+#include <codecvt>
 
 #if defined(_MSC_VER)
 #pragma warning(disable: 4244 4267) // possible loss of data
@@ -128,7 +144,7 @@ static void whisper_log_callback_default(ggml_log_level level, const char * text
 #define WHISPER_LOG_INFO(...)  whisper_log_internal(GGML_LOG_LEVEL_INFO , __VA_ARGS__)
 
 // define this to enable verbose trace logging - useful for debugging purposes
-// #define WHISPER_DEBUG
+//#define WHISPER_DEBUG
 
 #if defined(WHISPER_DEBUG)
 #define WHISPER_LOG_DEBUG(...) whisper_log_internal(GGML_LOG_LEVEL_DEBUG, __VA_ARGS__)
@@ -140,12 +156,10 @@ static void whisper_log_callback_default(ggml_log_level level, const char * text
     do { \
         if (!(x)) { \
             WHISPER_LOG_ERROR("WHISPER_ASSERT: %s:%d: %s\n", __FILE__, __LINE__, #x); \
-            Rcpp::stop("whispercpp error"); \
+            abort(); \
         } \
     } while (0)
 
-//#define WHISPER_USE_FLASH_ATTN
-//#define WHISPER_USE_FLASH_FF
 #define WHISPER_MAX_DECODERS 8
 #define WHISPER_MAX_NODES 4096
 
@@ -157,11 +171,11 @@ static bool ggml_graph_compute_helper(
           struct ggml_cgraph * graph,
         std::vector<uint8_t> & buf,
                          int   n_threads,
-      whisper_abort_callback   abort_callback,
+         ggml_abort_callback   abort_callback,
                         void * abort_callback_data) {
-    struct ggml_cplan plan = ggml_graph_plan(graph, n_threads);
+    struct ggml_cplan plan = ggml_graph_plan(graph, n_threads, nullptr);
 
-    plan.abort_callback = abort_callback;
+    plan.abort_callback      = abort_callback;
     plan.abort_callback_data = abort_callback_data;
 
     if (plan.work_size > 0) {
@@ -173,18 +187,25 @@ static bool ggml_graph_compute_helper(
 }
 
 static bool ggml_graph_compute_helper(
-       struct ggml_backend * backend,
+      ggml_backend_sched_t   sched,
         struct ggml_cgraph * graph,
                        int   n_threads) {
-    if (ggml_backend_is_cpu(backend)) {
-        ggml_backend_cpu_set_n_threads(backend, n_threads);
-    }
-#ifdef GGML_USE_METAL
-    if (ggml_backend_is_metal(backend)) {
-        ggml_backend_metal_set_n_cb(backend, n_threads);
-    }
+
+    for (int i = 0; i < ggml_backend_sched_get_n_backends(sched); ++i) {
+        ggml_backend_t backend = ggml_backend_sched_get_backend(sched, i);
+        if (ggml_backend_is_cpu(backend)) {
+            ggml_backend_cpu_set_n_threads(backend, n_threads);
+        }
+#ifdef GGML_USE_BLAS
+        if (ggml_backend_is_blas(backend)) {
+            ggml_backend_blas_set_n_threads(backend, n_threads);
+        }
 #endif
-    return ggml_backend_graph_compute(backend, graph);
+    }
+
+    bool t = ggml_backend_sched_graph_compute(sched, graph) == GGML_STATUS_SUCCESS;
+    ggml_backend_sched_reset(sched);
+    return t;
 }
 
 // faster matrix multiplications for tensors that do not have dimension 0 divisible by "pad"
@@ -348,6 +369,35 @@ static const std::map<std::string, std::pair<int, std::string>> g_lang = {
     { "yue", { 99,  "cantonese",      } },
 };
 
+// [EXPERIMENTAL] Token-level timestamps with DTW
+static const whisper_ahead g_aheads_tiny_en[]   = { {1, 0}, {2, 0}, {2, 5}, {3, 0}, {3, 1}, {3, 2}, {3, 3}, {3, 4} };
+static const whisper_ahead g_aheads_tiny[]      = { {2, 2}, {3, 0}, {3, 2}, {3, 3}, {3, 4}, {3, 5} };
+static const whisper_ahead g_aheads_base_en[]   = { {3, 3}, {4, 7}, {5, 1}, {5, 5}, {5, 7} };
+static const whisper_ahead g_aheads_base[]      = { {3, 1}, {4, 2}, {4, 3}, {4, 7}, {5, 1}, {5, 2}, {5, 4}, {5, 6} };
+static const whisper_ahead g_aheads_small_en[]  = { {6, 6}, {7, 0}, {7, 3}, {7, 8}, {8, 2}, {8, 5}, {8, 7}, {9, 0}, {9, 4}, {9, 8}, {9, 10}, {10, 0}, {10, 1}, {10, 2}, {10, 3}, {10, 6}, {10, 11}, {11, 2}, {11, 4} };
+static const whisper_ahead g_aheads_small[]     = { {5, 3}, {5, 9}, {8, 0}, {8, 4}, {8, 7}, {8, 8}, {9, 0}, {9, 7}, {9, 9}, {10, 5} };
+static const whisper_ahead g_aheads_medium_en[] = { {11, 4}, {14, 1}, {14, 12}, {14, 14}, {15, 4}, {16, 0}, {16, 4}, {16, 9}, {17, 12}, {17, 14}, {18, 7}, {18, 10}, {18, 15}, {20, 0}, {20, 3}, {20, 9}, {20, 14}, {21, 12} };
+static const whisper_ahead g_aheads_medium[]    = { {13, 15}, {15, 4}, {15, 15}, {16, 1}, {20, 0}, {23, 4} };
+static const whisper_ahead g_aheads_large_v1[]  = { {9, 19}, {11, 2}, {11, 4}, {11, 17}, {22, 7}, {22, 11}, {22, 17}, {23, 2}, {23, 15} };
+static const whisper_ahead g_aheads_large_v2[]  = { {10, 12}, {13, 17}, {16, 11}, {16, 12}, {16, 13}, {17, 15}, {17, 16}, {18, 4}, {18, 11}, {18, 19}, {19, 11}, {21, 2}, {21, 3}, {22, 3}, {22, 9}, {22, 12}, {23, 5}, {23, 7}, {23, 13}, {25, 5}, {26, 1}, {26, 12}, {27, 15} };
+static const whisper_ahead g_aheads_large_v3[]  = { {7, 0}, {10, 17}, {12, 18}, {13, 12}, {16, 1}, {17, 14}, {19, 11}, {21, 4}, {24, 1}, {25, 6} };
+
+static const std::map<whisper_alignment_heads_preset, whisper_aheads> g_aheads {
+    { WHISPER_AHEADS_TINY_EN,   {  8, g_aheads_tiny_en   } },
+    { WHISPER_AHEADS_TINY,      {  6, g_aheads_tiny      } },
+    { WHISPER_AHEADS_BASE_EN,   {  5, g_aheads_base_en   } },
+    { WHISPER_AHEADS_BASE,      {  8, g_aheads_base      } },
+    { WHISPER_AHEADS_SMALL_EN,  { 19, g_aheads_small_en  } },
+    { WHISPER_AHEADS_SMALL,     { 10, g_aheads_small     } },
+    { WHISPER_AHEADS_MEDIUM_EN, { 18, g_aheads_medium_en } },
+    { WHISPER_AHEADS_MEDIUM,    {  6, g_aheads_medium    } },
+    { WHISPER_AHEADS_LARGE_V1,  {  9, g_aheads_large_v1  } },
+    { WHISPER_AHEADS_LARGE_V2,  { 23, g_aheads_large_v2  } },
+    { WHISPER_AHEADS_LARGE_V3,  { 10, g_aheads_large_v3  } },
+};
+
+static std::vector<uint32_t> get_alignment_heads_by_layer(const whisper_context_params & cparams, int il, int32_t n_text_layer, int32_t n_head);
+
 struct whisper_mel {
     int n_len;
     int n_len_org;
@@ -410,7 +460,7 @@ struct whisper_batch {
 
     whisper_token  *  token;
     whisper_pos    *  pos;
-    int32_t        *  n_seq_id;
+    int32_t        *  n_seq_id; // always 1, here for consistency with llama.cpp
     whisper_seq_id ** seq_id;   // null terminated
     int8_t         *  logits;
 };
@@ -470,54 +520,42 @@ struct whisper_pair {
     whisper_pair() : first(A()), second(B()) {}
 };
 
-// ggml_allocr wrapper for whisper usage
-struct whisper_allocr {
-    ggml_allocr * alloc = nullptr;
+// ggml_backend_sched wrapper for whisper usage
+struct whisper_sched {
+    ggml_backend_sched_t sched = nullptr;
 
     std::vector<uint8_t> meta;
-
-    ggml_backend_buffer_t buffer;
 };
 
-static size_t whisper_allocr_size(struct whisper_allocr & allocr) {
-    return allocr.meta.size() + ggml_allocr_max_size(allocr.alloc);
+static size_t whisper_sched_size(struct whisper_sched & allocr) {
+    size_t size = allocr.meta.size();
+    for (int i = 0; i < ggml_backend_sched_get_n_backends(allocr.sched); ++i) {
+        ggml_backend_t backend = ggml_backend_sched_get_backend(allocr.sched, i);
+        size += ggml_backend_sched_get_buffer_size(allocr.sched, backend);
+    }
+    return size;
 }
 
 // measure the memory usage of a graph and prepare the allocr's internal data buffer
-static void whisper_allocr_graph_init(struct whisper_allocr & allocr, ggml_backend_t backend, std::function<struct ggml_cgraph *()> && get_graph) {
-    auto & alloc = allocr.alloc;
+static bool whisper_sched_graph_init(struct whisper_sched & allocr, std::vector<ggml_backend_t> backends, std::function<struct ggml_cgraph *()> && get_graph) {
+    auto & sched = allocr.sched;
     auto & meta  = allocr.meta;
 
-    alloc = ggml_allocr_new_measure_from_backend(backend);
+    sched = ggml_backend_sched_new(backends.data(), nullptr, backends.size(), WHISPER_MAX_NODES, false);
 
     meta.resize(ggml_tensor_overhead()*WHISPER_MAX_NODES + ggml_graph_overhead());
 
-    ggml_allocr_alloc_graph(alloc, get_graph());
-}
-
-static void whisper_allocr_graph_realloc(struct whisper_allocr & allocr, ggml_backend_t backend) {
-    if (allocr.alloc == nullptr) {
-        // this can be null if we use external encoder like CoreML or OpenVINO
-        return;
+    // since there are dependencies between the different graphs,
+    // we need to allocate them instead of only reserving to get the correct compute buffer size
+    if (!ggml_backend_sched_alloc_graph(sched, get_graph())) {
+        // failed to allocate the compute buffer
+        WHISPER_LOG_ERROR("%s: failed to allocate the compute buffer\n", __func__);
+        return false;
     }
 
-    auto & alloc  = allocr.alloc;
-    auto & buffer = allocr.buffer;
-
-    size_t size = ggml_allocr_max_size(alloc);
+    ggml_backend_sched_reset(sched);
 
-    ggml_allocr_free(alloc);
-
-    buffer = ggml_backend_alloc_buffer(backend, size);
-    alloc = ggml_allocr_new_from_buffer(buffer);
-}
-
-static void whisper_allocr_free(struct whisper_allocr & allocr) {
-    if (allocr.alloc) {
-        ggml_allocr_free(allocr.alloc);
-        ggml_backend_buffer_free(allocr.buffer);
-        allocr.alloc = nullptr;
-    }
+    return true;
 }
 
 // medium
@@ -659,9 +697,9 @@ struct whisper_kv_cache {
     struct ggml_tensor * k;
     struct ggml_tensor * v;
 
-    struct ggml_context * ctx;
+    struct ggml_context * ctx = nullptr;
 
-    ggml_backend_buffer_t buffer;
+    ggml_backend_buffer_t buffer = nullptr;
 };
 
 struct whisper_model {
@@ -699,10 +737,10 @@ struct whisper_model {
     std::vector<whisper_layer_decoder> layers_decoder;
 
     // ggml context that contains all the meta information about the model tensors
-    struct ggml_context * ctx;
+    struct ggml_context * ctx = nullptr;
 
     // the model backend data is read-only and can be shared between processors
-    struct ggml_backend_buffer * buffer;
+    ggml_backend_buffer_t buffer = nullptr;
 
     // tensors
     int n_loaded;
@@ -767,6 +805,13 @@ struct whisper_decoder {
     mutable std::mt19937 rng; // used for sampling at t > 0.0
 };
 
+// [EXPERIMENTAL] Token-level timestamps with DTW
+struct whisper_aheads_masks {
+    std::vector<struct ggml_tensor *> m;    // One mask per text layer.
+    struct ggml_context * ctx = nullptr;
+    ggml_backend_buffer_t buffer = nullptr;
+};
+
 struct whisper_state {
     int64_t t_sample_us = 0;
     int64_t t_encode_us = 0;
@@ -783,6 +828,9 @@ struct whisper_state {
     int32_t n_fail_p = 0; // number of logprob threshold failures
     int32_t n_fail_h = 0; // number of entropy threshold failures
 
+    // number of decoders for which we have constructed the KV cache
+    int32_t kv_self_n_dec = 0;
+
     // unified self-attention KV cache for all decoders
     whisper_kv_cache kv_self;
 
@@ -790,21 +838,22 @@ struct whisper_state {
     // shared between all decoders
     whisper_kv_cache kv_cross;
 
+    // padded buffer for flash-attention
+    whisper_kv_cache kv_pad;
+
     whisper_mel mel;
 
     whisper_batch batch;
 
     whisper_decoder decoders[WHISPER_MAX_DECODERS];
 
-    ggml_backend_t backend = nullptr;
+    std::vector<ggml_backend_t> backends;
 
-    // ggml-alloc:
     // - stores meta info about the intermediate tensors into the `meta` buffers
-    // - stores the actual tensor data into the `data` buffers
-    whisper_allocr alloc_conv;
-    whisper_allocr alloc_encode;
-    whisper_allocr alloc_cross;
-    whisper_allocr alloc_decode;
+    whisper_sched sched_conv;
+    whisper_sched sched_encode;
+    whisper_sched sched_cross;
+    whisper_sched sched_decode;
 
     // result of the encoder
     struct ggml_tensor * embd_conv = nullptr;
@@ -840,6 +889,11 @@ struct whisper_state {
 
     std::vector<float> energy; // PCM signal energy
 
+    // [EXPERIMENTAL] Token-level timestamps with DTW
+    whisper_aheads_masks aheads_masks;
+    ggml_tensor * aheads_cross_QKs = nullptr;
+    std::vector<float> aheads_cross_QKs_data;
+
     // [EXPERIMENTAL] speed-up techniques
     int32_t exp_n_audio_ctx = 0; // 0 - use default
 };
@@ -858,8 +912,6 @@ struct whisper_context {
 
     whisper_state * state = nullptr;
 
-    ggml_backend_t backend = nullptr;
-
     std::string path_model; // populated by whisper_init_from_file_with_params()
 };
 
@@ -877,15 +929,13 @@ static void read_safe(whisper_model_loader * loader, T & dest) {
     BYTESWAP_VALUE(dest);
 }
 
-static bool kv_cache_init(
-        const struct whisper_hparams & hparams,
+static bool whisper_kv_cache_init(
              struct whisper_kv_cache & cache,
                       ggml_backend_t   backend,
                            ggml_type   wtype,
+                             int64_t   n_text_state,
+                             int64_t   n_text_layer,
                                  int   n_ctx) {
-    const int64_t n_text_state = hparams.n_text_state;
-    const int64_t n_text_layer = hparams.n_text_layer;
-
     const int64_t n_mem      = n_text_layer*n_ctx;
     const int64_t n_elements = n_text_state*n_mem;
 
@@ -904,36 +954,28 @@ static bool kv_cache_init(
     cache.ctx = ggml_init(params);
 
     if (!cache.ctx) {
-        WHISPER_LOG_ERROR("%s: failed to allocate memory for kv cache\n", __func__);
+        WHISPER_LOG_ERROR("%s: failed to allocate memory for the kv cache context\n", __func__);
         return false;
     }
 
     cache.k = ggml_new_tensor_1d(cache.ctx, wtype, n_elements);
     cache.v = ggml_new_tensor_1d(cache.ctx, wtype, n_elements);
 
-    const size_t mem_bytes = ggml_nbytes(cache.k) + ggml_nbytes(cache.v);
-
-    cache.buffer = ggml_backend_alloc_buffer(backend, mem_bytes);
-
-    // allocate the tensors into the backend buffer
-    {
-        ggml_allocr * alloc = ggml_allocr_new_from_buffer(cache.buffer);
-
-        ggml_allocr_alloc(alloc, cache.k);
-        ggml_allocr_alloc(alloc, cache.v);
-
-        ggml_allocr_free(alloc);
+    cache.buffer = ggml_backend_alloc_ctx_tensors(cache.ctx, backend);
+    if (!cache.buffer) {
+        WHISPER_LOG_ERROR("%s: failed to allocate memory for the kv cache\n", __func__);
+        return false;
     }
 
+    ggml_backend_buffer_clear(cache.buffer, 0);
+
     return true;
 }
 
-static void kv_cache_free(struct whisper_kv_cache & cache) {
-    if (cache.ctx) {
-        ggml_free(cache.ctx);
-        ggml_backend_buffer_free(cache.buffer);
-        cache.ctx = nullptr;
-    }
+static void whisper_kv_cache_free(struct whisper_kv_cache & cache) {
+    ggml_free(cache.ctx);
+    ggml_backend_buffer_free(cache.buffer);
+    cache.ctx = nullptr;
 }
 
 static bool whisper_kv_cache_find_slot(
@@ -1004,6 +1046,8 @@ static void whisper_kv_cache_clear(struct whisper_kv_cache & cache) {
         cache.cells[i].seq_id.clear();
     }
     cache.head = 0;
+
+    ggml_backend_buffer_clear(cache.buffer, 0);
 }
 
 static void whisper_kv_cache_seq_rm(
@@ -1054,15 +1098,167 @@ static void whisper_kv_cache_seq_cp(
     }
 }
 
-static ggml_backend_t whisper_backend_init(const whisper_context_params & params) {
-    ggml_backend_t backend_gpu = NULL;
+static uint32_t whisper_kv_cache_get_padding(const struct whisper_context & wctx) {
+    if (!wctx.params.flash_attn || !wctx.params.use_gpu) {
+        return 1u;
+    }
+
+#ifdef GGML_USE_METAL
+    if (wctx.params.use_gpu) {
+        return 32u;
+    }
+#endif
+
+#ifdef GGML_USE_CUDA
+    if (wctx.params.use_gpu) {
+        return 256u;
+    }
+#endif
+
+    return 1u;
+}
+
+// [EXPERIMENTAL] Token-level timestamps with DTW
+static bool aheads_masks_init(
+        const whisper_context_params & cparams,
+               const whisper_hparams & hparams,
+         struct whisper_aheads_masks & aheads_masks,
+                      ggml_backend_t   backend) {
+
+    const int32_t n_text_layer = hparams.n_text_layer;
+    const int32_t n_head = hparams.n_text_head;
+
+    // Sanity checks
+    if (cparams.dtw_aheads_preset == WHISPER_AHEADS_NONE) {
+        WHISPER_LOG_ERROR("%s: dtw_aheads_preset should be != DTW_AHEADS_NONE\n", __func__);
+        return false;
+    } else if (cparams.dtw_aheads_preset == WHISPER_AHEADS_N_TOP_MOST) {
+        if (cparams.dtw_n_top > n_text_layer || cparams.dtw_n_top <= 0) {
+            WHISPER_LOG_ERROR("%s: dtw_n_top must be between %d and %d for this model.", __func__, 1, n_text_layer);
+            return false;
+        }
+    } else {
+        const auto aheads = cparams.dtw_aheads_preset == WHISPER_AHEADS_CUSTOM ? cparams.dtw_aheads : g_aheads.at(cparams.dtw_aheads_preset);
+        if (cparams.dtw_aheads_preset == WHISPER_AHEADS_CUSTOM) {
+            if (aheads.n_heads == 0) {
+                WHISPER_LOG_ERROR("%s: dtw_aheads.n_heads should be > 0", __func__);
+                return false;
+            }
+            if (aheads.heads == NULL) {
+                WHISPER_LOG_ERROR("%s: dtw_aheads.heads unset", __func__);
+                return false;
+            }
+        }
+        for (size_t i = 0; i < aheads.n_heads; ++i) {
+            if (aheads.heads[i].n_text_layer >= n_text_layer) {
+                WHISPER_LOG_ERROR("%s: tried to set alignment head on text layer %d, but model only has %d text layers", __func__, aheads.heads[i].n_text_layer + 1, n_text_layer);
+                return false;
+            }
+            if (aheads.heads[i].n_text_layer < 0) {
+                WHISPER_LOG_ERROR("%s: tried to set alignment head on text layer < 0", __func__);
+                return false;
+            }
+            if (aheads.heads[i].n_head >= n_head) {
+                WHISPER_LOG_ERROR("%s: tried to set alignment head on head %d, but model only has %d heads", __func__, aheads.heads[i].n_head + 1, n_head);
+                return false;
+            }
+            if (aheads.heads[i].n_head < 0) {
+                WHISPER_LOG_ERROR("%s: tried to set alignment head on head < 0", __func__);
+                return false;
+            }
+        }
+    }
+
+    struct ggml_init_params params = {
+        /*.mem_size   =*/ (size_t) static_cast<size_t>(n_text_layer)*ggml_tensor_overhead(),
+        /*.mem_buffer =*/ nullptr,
+        /*.no_alloc   =*/ true,
+    };
+
+    aheads_masks.ctx = ggml_init(params);
+
+    if (!aheads_masks.ctx) {
+        WHISPER_LOG_ERROR("%s: failed to allocate memory for the aheads_masks context\n", __func__);
+        return false;
+    }
+
+    for (int64_t il = 0; il < n_text_layer; ++il) {
+        auto aheads = get_alignment_heads_by_layer(cparams, il, n_text_layer, n_head);
+        if (!aheads.empty()) {
+            aheads_masks.m.push_back(ggml_new_tensor_2d(aheads_masks.ctx, GGML_TYPE_F32, n_head, aheads.size()));
+        } else {
+            aheads_masks.m.push_back(nullptr);
+        }
+    }
+
+    aheads_masks.buffer = ggml_backend_alloc_ctx_tensors(aheads_masks.ctx, backend);
+    if (!aheads_masks.buffer) {
+        WHISPER_LOG_ERROR("%s: failed to allocate memory for aheads_masks\n", __func__);
+        return false;
+    }
+
+    // Set data on mask tensors
+    // Since this must be backend agnostic, we write our desired values on mask_data,
+    // and send it to backend with ggml_backend_tensor_set.
+    // Each mask in N_HEADS*N_ALIGNMENT_HEADS, one per text layer containing alignment
+    // heads. Each row of the mask "marks" one alignment head. E.g. if some text layer
+    // has a total of 10 heads and of those, heads 0,5,6 are alignment heads, the mask
+    // should read:
+    // 1 0 0 0 0 0 0 0 0 0
+    // 0 0 0 0 0 1 0 0 0 0
+    // 0 0 0 0 0 0 1 0 0 0
+    std::vector<float> mask_data;
+    for (int64_t il = 0; il < n_text_layer; ++il) {
+        if (aheads_masks.m[il] != nullptr) {
+            auto aheads = get_alignment_heads_by_layer(cparams, il, n_text_layer, n_head);
+
+            size_t data_size = aheads_masks.m[il]->ne[0] * aheads_masks.m[il]->ne[1];
+            size_t data_size_bytes = data_size * sizeof(float);
+            mask_data.resize(data_size);
+
+            std::fill(mask_data.begin(), mask_data.end(), 0);
+            for (size_t ih = 0; ih < aheads.size(); ++ih) {
+                size_t pos = (aheads[ih] + (ih * aheads_masks.m[il]->ne[0]));
+                mask_data[pos] = 1.0f;
+            }
+
+            ggml_backend_tensor_set(aheads_masks.m[il], mask_data.data(), 0, data_size_bytes);
+        }
+    }
+
+    if (aheads_masks.m.empty()) {
+        WHISPER_LOG_ERROR("%s: \n", __func__);
+        return false;
+    }
+
+    return true;
+}
+
+static void aheads_masks_free(struct whisper_aheads_masks & aheads_masks) {
+    ggml_free(aheads_masks.ctx);
+    ggml_backend_buffer_free(aheads_masks.buffer);
+    aheads_masks.ctx = nullptr;
+}
 
-    // initialize the backends
-#ifdef GGML_USE_CUBLAS
-    if (params.use_gpu && ggml_cublas_loaded()) {
+static size_t aheads_masks_nbytes(struct whisper_aheads_masks & aheads_masks) {
+    size_t size = 0;
+    for (size_t i = 0; i < aheads_masks.m.size(); ++i) {
+        if (aheads_masks.m[i] != nullptr)
+            size += ggml_nbytes(aheads_masks.m[i]);
+    }
+    return size;
+}
+
+static ggml_backend_t whisper_backend_init_gpu(const whisper_context_params & params) {
+    ggml_backend_t result = NULL;
+
+    ggml_log_set(g_state.log_callback, g_state.log_callback_user_data);
+
+#ifdef GGML_USE_CUDA
+    if (params.use_gpu) {
         WHISPER_LOG_INFO("%s: using CUDA backend\n", __func__);
-        backend_gpu = ggml_backend_cuda_init(0);
-        if (!backend_gpu) {
+        result = ggml_backend_cuda_init(params.gpu_device);
+        if (!result) {
             WHISPER_LOG_ERROR("%s: ggml_backend_cuda_init() failed\n", __func__);
         }
     }
@@ -1071,22 +1267,108 @@ static ggml_backend_t whisper_backend_init(const whisper_context_params & params
 #ifdef GGML_USE_METAL
     if (params.use_gpu) {
         WHISPER_LOG_INFO("%s: using Metal backend\n", __func__);
-        ggml_metal_log_set_callback(whisper_log_callback_default, nullptr);
-        backend_gpu = ggml_backend_metal_init();
-        if (!backend_gpu) {
+        result = ggml_backend_metal_init();
+        if (!result) {
             WHISPER_LOG_ERROR("%s: ggml_backend_metal_init() failed\n", __func__);
-        } else if (!ggml_backend_metal_supports_family(backend_gpu, 7)) {
+        } else if (!ggml_backend_metal_supports_family(result, 7)) {
             WHISPER_LOG_ERROR("%s: Metal GPU does not support family 7 - falling back to CPU\n", __func__);
-            ggml_backend_free(backend_gpu);
-            backend_gpu = NULL;
+            ggml_backend_free(result);
+            result = NULL;
+        }
+    }
+#endif
+
+#ifdef GGML_USE_SYCL
+    if (params.use_gpu) {
+        WHISPER_LOG_INFO("%s: using SYCL backend\n", __func__);
+        result = ggml_backend_sycl_init(params.gpu_device);
+        if (!result) {
+            WHISPER_LOG_ERROR("%s: ggml_backend_sycl_init() failed\n", __func__);
+        }
+    }
+#endif
+
+#ifdef GGML_USE_VULKAN
+    if (params.use_gpu) {
+        WHISPER_LOG_INFO("%s: using Vulkan backend\n", __func__);
+        result = ggml_backend_vk_init(params.gpu_device);
+        if (!result) {
+            WHISPER_LOG_ERROR("%s: ggml_backend_vk_init() failed\n", __func__);
         }
     }
 #endif
 
+#ifdef GGML_USE_CANN
+    if (params.use_gpu) {
+        WHISPER_LOG_INFO("%s: using CANN backend\n", __func__);
+        result = ggml_backend_cann_init(params.gpu_device);
+        if (!result) {
+            WHISPER_LOG_ERROR("%s: ggml_backend_cann_init() failed\n", __func__);
+        }
+    }
+#endif
+
+    GGML_UNUSED(params);
+
+    return result;
+}
+
+static std::vector<ggml_backend_t> whisper_backend_init(const whisper_context_params & params) {
+    std::vector<ggml_backend_t> result;
+
+    ggml_backend_t backend_gpu = whisper_backend_init_gpu(params);
+
     if (backend_gpu) {
-        return backend_gpu;
+        result.push_back(backend_gpu);
+    }
+
+#ifdef GGML_USE_BLAS
+    {
+        WHISPER_LOG_INFO("%s: using BLAS backend\n", __func__);
+        ggml_backend_t backend_blas = ggml_backend_blas_init();
+        if (!backend_blas) {
+            WHISPER_LOG_ERROR("%s: ggml_backend_blas_init() failed\n", __func__);
+        } else {
+            result.push_back(backend_blas);
+        }
     }
-    return ggml_backend_cpu_init();
+#endif
+
+    GGML_UNUSED(params);
+
+    result.push_back(ggml_backend_cpu_init());
+
+    return result;
+}
+
+static ggml_backend_buffer_type_t whisper_default_buffer_type(const whisper_context_params & params) {
+    ggml_backend_buffer_type_t result = nullptr;
+
+    params.use_gpu || (result = ggml_backend_cpu_buffer_type());
+
+#ifdef GGML_USE_CUDA
+    result || (result = ggml_backend_cuda_buffer_type(params.gpu_device));
+#endif
+
+#ifdef GGML_USE_METAL
+    result || (result = ggml_backend_metal_buffer_type());
+#endif
+
+#ifdef GGML_USE_SYCL
+    result || (result = ggml_backend_sycl_buffer_type(params.gpu_device));
+#endif
+
+#ifdef GGML_USE_VULKAN
+    result || (result = ggml_backend_vk_buffer_type(params.gpu_device));
+#endif
+
+#ifdef GGML_USE_CANN
+    result || (result == ggml_backend_cann_buffer_type(params.gpu_device));
+#endif
+
+    result || (result = ggml_backend_cpu_buffer_type());
+
+    return result;
 }
 
 // load the model from a ggml file
@@ -1513,29 +1795,16 @@ static bool whisper_model_load(struct whisper_model_loader * loader, whisper_con
         }
     }
 
-    wctx.backend = whisper_backend_init(wctx.params);
-
-    {
-        size_t size_main = 0;
-
-        for (const auto & t : model.tensors) {
-            size_main += ggml_nbytes(t.second) + ggml_tensor_overhead();
-        }
-
-        model.buffer = ggml_backend_alloc_buffer(wctx.backend, size_main);
-
-        WHISPER_LOG_INFO("%s: %8s buffer size = %8.2f MB\n", __func__, ggml_backend_name(wctx.backend), size_main / 1e6);
-    }
-
-    ggml_allocr * alloc = ggml_allocr_new_from_buffer(model.buffer);
-
     // allocate tensors in the backend buffers
-    {
-        for (const auto & t : model.tensors) {
-            ggml_allocr_alloc(alloc, t.second);
-        }
+    model.buffer = ggml_backend_alloc_ctx_tensors_from_buft(model.ctx, whisper_default_buffer_type(wctx.params));
+    if (!model.buffer) {
+        WHISPER_LOG_ERROR("%s: failed to allocate memory for the model\n", __func__);
+        return false;
     }
 
+    size_t size_main = ggml_backend_buffer_get_size(model.buffer);
+    WHISPER_LOG_INFO("%s: %8s total size = %8.2f MB\n", __func__, ggml_backend_buffer_name(model.buffer), size_main / 1e6);
+
     // load weights
     {
         size_t total_size = 0;
@@ -1597,15 +1866,11 @@ static bool whisper_model_load(struct whisper_model_loader * loader, whisper_con
                 return false;
             }
 
-            ggml_backend_t backend = wctx.backend;
+            //ggml_backend_t backend = wctx.backend;
 
             //printf("%s: [%5.5s] %s\n", __func__, ggml_backend_name(backend), name.c_str());
 
-            if ((ggml_backend_is_cpu(backend)
-#ifdef GGML_USE_METAL
-                || ggml_backend_is_metal(backend)
-#endif
-                )) {
+            if (ggml_backend_buffer_is_host(model.buffer)) {
                 // for the CPU and Metal backend, we can read directly into the tensor
                 loader->read(loader->context, tensor->data, ggml_nbytes(tensor));
                 BYTESWAP_TENSOR(tensor);
@@ -1633,7 +1898,7 @@ static bool whisper_model_load(struct whisper_model_loader * loader, whisper_con
         }
     }
 
-    ggml_allocr_free(alloc);
+    ggml_backend_buffer_set_usage(model.buffer, GGML_BACKEND_BUFFER_USAGE_WEIGHTS);
 
     wctx.t_load_us = ggml_time_us() - t_start_us;
 
@@ -1660,10 +1925,8 @@ static bool whisper_encode_external(const whisper_state & wstate) {
 
 static struct ggml_cgraph * whisper_build_graph_conv(
         whisper_context & wctx,
-          whisper_state & wstate,
-              const int   mel_offset) {
+          whisper_state & wstate) {
     const auto & model   = wctx.model;
-    const auto & mel_inp = wstate.mel;
     const auto & hparams = model.hparams;
 
     const int n_ctx   = wstate.exp_n_audio_ctx > 0 ? wstate.exp_n_audio_ctx : hparams.n_audio_ctx;
@@ -1672,8 +1935,8 @@ static struct ggml_cgraph * whisper_build_graph_conv(
     const int n_mels = hparams.n_mels;
 
     struct ggml_init_params params = {
-        /*.mem_size   =*/ wstate.alloc_conv.meta.size(),
-        /*.mem_buffer =*/ wstate.alloc_conv.meta.data(),
+        /*.mem_size   =*/ wstate.sched_conv.meta.size(),
+        /*.mem_buffer =*/ wstate.sched_conv.meta.data(),
         /*.no_alloc   =*/ true,
     };
 
@@ -1681,31 +1944,9 @@ static struct ggml_cgraph * whisper_build_graph_conv(
 
     ggml_cgraph * gf = ggml_new_graph(ctx0);
 
-    ggml_allocr * alloc = wstate.alloc_conv.alloc;
-
     struct ggml_tensor * mel = ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, 2*n_ctx, n_mels);
-    ggml_allocr_alloc(alloc, mel);
-
-    assert(mel->type == GGML_TYPE_F32);
-    if (!ggml_allocr_is_measure(alloc)) {
-        assert(mel_inp.n_mel == n_mels);
-
-        wstate.inp_mel.resize(ggml_nelements(mel));
-
-        float * dst = wstate.inp_mel.data();
-        memset(dst, 0, ggml_nbytes(mel));
-
-        const int i0 = std::min(mel_offset,           mel_inp.n_len);
-        const int i1 = std::min(mel_offset + 2*n_ctx, mel_inp.n_len);
-
-        for (int j = 0; j < mel_inp.n_mel; ++j) {
-            for (int i = i0; i < i1; ++i) {
-                dst[j*2*n_ctx + (i - i0)] = mel_inp.data[j*mel_inp.n_len + i];
-            }
-        }
-
-        ggml_backend_tensor_set(mel, wstate.inp_mel.data(), 0, ggml_nelements(mel)*sizeof(float));
-    }
+    ggml_set_name(mel, "mel");
+    ggml_set_input(mel);
 
     struct ggml_tensor * cur = nullptr;
 
@@ -1726,27 +1967,17 @@ static struct ggml_cgraph * whisper_build_graph_conv(
         ggml_set_name(cur, "embd_conv");
         wstate.embd_conv = cur;
     } else {
-#ifdef WHISPER_USE_COREML
-        cur = ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, n_state, n_ctx);
-        ggml_allocr_alloc(alloc, cur);
+        ggml_build_forward_expand(gf, mel);
 
-        if (!ggml_allocr_is_measure(alloc)) {
-            whisper_coreml_encode(wstate.ctx_coreml, mel->ne[0], mel->ne[1], (float *) mel->data, (float *) cur->data);
-        }
-#endif
-#ifdef WHISPER_USE_OPENVINO
         cur = ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, n_state, n_ctx);
-        ggml_allocr_alloc(alloc, cur);
-
-        if (!ggml_allocr_is_measure(alloc)) {
-            whisper_openvino_encode(wstate.ctx_openvino, mel, cur);
-        }
-#endif
+        ggml_set_input(cur); // the external encoder will write into this tensor
 
         ggml_set_name(cur, "embd_enc");
         wstate.embd_enc = cur;
     }
 
+    ggml_set_output(cur);
+
     ggml_build_forward_expand(gf, cur);
 
     ggml_free(ctx0);
@@ -1765,9 +1996,17 @@ static struct ggml_cgraph * whisper_build_graph_encoder(
     const int n_head  = hparams.n_audio_head;
     const int n_layer = hparams.n_audio_layer;
 
+    const int n_state_head = n_state/n_head;
+
+    auto & kv_pad = wstate.kv_pad;
+
+    WHISPER_ASSERT(!!kv_pad.ctx);
+
+    const int n_ctx_pad = GGML_PAD(n_ctx, 256);
+
     struct ggml_init_params params = {
-        /*.mem_size   =*/ wstate.alloc_encode.meta.size(),
-        /*.mem_buffer =*/ wstate.alloc_encode.meta.data(),
+        /*.mem_size   =*/ wstate.sched_encode.meta.size(),
+        /*.mem_buffer =*/ wstate.sched_encode.meta.data(),
         /*.no_alloc   =*/ true,
     };
 
@@ -1775,17 +2014,9 @@ static struct ggml_cgraph * whisper_build_graph_encoder(
 
     ggml_cgraph * gf = ggml_new_graph_custom(ctx0, WHISPER_MAX_NODES, false);
 
-    //ggml_allocr * alloc = wstate.alloc_encode.alloc;
-
-    //struct ggml_tensor * cur = ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, n_ctx, n_state);
-    //ggml_allocr_alloc(alloc, cur);
-
-    //if (!ggml_allocr_is_measure(alloc)) {
-    //    ggml_backend_tensor_copy(wstate.embd_conv, cur);
-    //}
     struct ggml_tensor * cur = ggml_view_tensor(ctx0, wstate.embd_conv);
 
-    const float KQscale = 1.0f/sqrtf(float(n_state)/n_head);
+    const float KQscale = 1.0f/sqrtf(float(n_state_head));
 
     // ===================================================================
     // NOTE: experimenting with partial evaluation of the encoder (ignore)
@@ -1835,14 +2066,14 @@ static struct ggml_cgraph * whisper_build_graph_encoder(
 
             Qcur = ggml_add(ctx0, Qcur, layer.attn_q_b);
 
-            //Qcur = ggml_scale(ctx0, Qcur, pow(float(n_state)/n_head, -0.25));
+            //Qcur = ggml_scale(ctx0, Qcur, pow(float(n_state_head), -0.25));
 
             // note: no bias for Key
             struct ggml_tensor * Kcur = ggml_mul_mat(ctx0,
                     layer.attn_k_w,
                     cur);
 
-            //Kcur = ggml_scale(ctx0, Kcur, pow(float(n_state)/n_head, -0.25));
+            //Kcur = ggml_scale(ctx0, Kcur, pow(float(n_state_head), -0.25));
 
             struct ggml_tensor * Vcur = ggml_mul_mat(ctx0,
                     layer.attn_v_w,
@@ -1852,70 +2083,60 @@ static struct ggml_cgraph * whisper_build_graph_encoder(
 
             // ------
 
-#ifdef WHISPER_USE_FLASH_ATTN
             struct ggml_tensor * Q =
                 ggml_permute(ctx0,
-                        ggml_cpy(ctx0,
-                            Qcur,
-                            ggml_new_tensor_3d(ctx0, wctx.itype, n_state/n_head, n_head, n_ctx)),
+                        ggml_reshape_3d(ctx0, Qcur, n_state_head, n_head, n_ctx),
                         0, 2, 1, 3);
 
-            struct ggml_tensor * K =
-                ggml_permute(ctx0,
-                        ggml_cpy(ctx0,
-                            Kcur,
-                            ggml_new_tensor_3d(ctx0, wctx.itype, n_state/n_head, n_head, n_ctx)),
-                        0, 2, 1, 3);
+            if (wctx.params.flash_attn) {
+                ggml_build_forward_expand(gf, ggml_cpy(ctx0, Kcur, ggml_view_1d(ctx0, kv_pad.k, n_ctx*n_state, 0)));
+                ggml_build_forward_expand(gf, ggml_cpy(ctx0, Vcur, ggml_view_1d(ctx0, kv_pad.v, n_ctx*n_state, 0)));
 
-            struct ggml_tensor * V =
-                ggml_cpy(ctx0,
-                        ggml_permute(ctx0,
-                            ggml_reshape_3d(ctx0,
-                                Vcur,
-                                n_state/n_head, n_head, n_ctx),
-                            1, 2, 0, 3),
-                        ggml_new_tensor_3d(ctx0, wctx.itype, n_ctx, n_state/n_head, n_head));
+                struct ggml_tensor * K =
+                    ggml_view_3d(ctx0, kv_pad.k,
+                            n_state_head, n_ctx_pad, n_head,
+                            ggml_element_size(kv_pad.k)*n_state,
+                            ggml_element_size(kv_pad.k)*n_state_head,
+                            0);
 
-            struct ggml_tensor * KQV = ggml_flash_attn(ctx0, Q, K, V, false);
-#else
-            struct ggml_tensor * Q =
-                ggml_permute(ctx0,
-                        ggml_cpy(ctx0,
-                            Qcur,
-                            ggml_new_tensor_3d(ctx0, GGML_TYPE_F32, n_state/n_head, n_head, n_ctx)),
-                        0, 2, 1, 3);
+                struct ggml_tensor * V =
+                    ggml_view_3d(ctx0, kv_pad.v,
+                            n_state_head, n_ctx_pad, n_head,
+                            ggml_element_size(kv_pad.v)*n_state,
+                            ggml_element_size(kv_pad.v)*n_state_head,
+                            0);
 
-            struct ggml_tensor * K =
-                ggml_permute(ctx0,
-                        ggml_cpy(ctx0,
-                            Kcur,
-                            ggml_new_tensor_3d(ctx0, wctx.itype, n_state/n_head, n_head, n_ctx)),
-                        0, 2, 1, 3);
+                cur = ggml_flash_attn_ext(ctx0, Q, K, V, nullptr, KQscale, 0.0f, 0.0f);
 
-            // K * Q
-            struct ggml_tensor * KQ = ggml_mul_mat(ctx0, K, Q);
+                cur = ggml_reshape_2d(ctx0, cur, n_state, n_ctx);
+            } else {
+                struct ggml_tensor * K =
+                    ggml_permute(ctx0,
+                            ggml_cast(ctx0,
+                                ggml_reshape_3d(ctx0, Kcur, n_state_head, n_head, n_ctx),
+                                wctx.itype),
+                            0, 2, 1, 3);
 
-            struct ggml_tensor * KQ_scaled = ggml_scale(ctx0, KQ, KQscale);
+                // K * Q
+                struct ggml_tensor * KQ = ggml_mul_mat(ctx0, K, Q);
 
-            struct ggml_tensor * KQ_soft_max = ggml_soft_max(ctx0, KQ_scaled);
+                struct ggml_tensor * KQ_soft_max = ggml_soft_max_ext(ctx0, KQ, nullptr, KQscale, 0.0f);
 
-            struct ggml_tensor * V =
-                ggml_cpy(ctx0,
-                        ggml_permute(ctx0,
-                            ggml_reshape_3d(ctx0,
-                                Vcur,
-                                n_state/n_head, n_head, n_ctx),
-                            1, 2, 0, 3),
-                        ggml_new_tensor_3d(ctx0, wctx.itype, n_ctx, n_state/n_head, n_head)
-                        );
+                struct ggml_tensor * V =
+                    ggml_cast(ctx0,
+                            ggml_permute(ctx0,
+                                ggml_reshape_3d(ctx0,
+                                    Vcur,
+                                    n_state_head, n_head, n_ctx),
+                                1, 2, 0, 3),
+                            wctx.itype);
 
-            struct ggml_tensor * KQV = ggml_mul_mat(ctx0, V, KQ_soft_max);
-#endif
-            struct ggml_tensor * KQV_merged = ggml_permute(ctx0, KQV, 0, 2, 1, 3);
+                struct ggml_tensor * KQV = ggml_mul_mat(ctx0, V, KQ_soft_max);
 
-            cur = ggml_cpy(ctx0,
-                    KQV_merged,
-                    ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, n_state, n_ctx));
+                struct ggml_tensor * KQV_merged = ggml_permute(ctx0, KQV, 0, 2, 1, 3);
+
+                cur = ggml_cont_2d(ctx0, KQV_merged, n_state, n_ctx);
+            }
         }
 
         // projection
@@ -1944,11 +2165,6 @@ static struct ggml_cgraph * whisper_build_graph_encoder(
                         layer.mlp_ln_b);
             }
 
-#ifdef WHISPER_USE_FLASH_FF
-            cur = ggml_flash_ff(ctx0,
-                    ggml_cpy(ctx0, cur, ggml_new_tensor_2d(ctx0, wstate.itype, n_state, n_ctx)),
-                    layer.mlp_0_w, layer.mlp_0_b, layer.mlp_1_w, layer.mlp_1_b);
-#else
             // fully connected
             cur = ggml_mul_mat(ctx0,
                     layer.mlp_0_w,
@@ -1965,7 +2181,6 @@ static struct ggml_cgraph * whisper_build_graph_encoder(
                     cur);
 
             cur = ggml_add(ctx0, cur, layer.mlp_1_b);
-#endif
         }
 
         inpL = ggml_add(ctx0, cur, inpFF);
@@ -2014,9 +2229,13 @@ static struct ggml_cgraph * whisper_build_graph_cross(
     const int n_state = hparams.n_audio_state;
     const int n_head  = hparams.n_audio_head;
 
+    const int n_state_head = n_state/n_head;
+
+    const int n_ctx_pad = GGML_PAD(n_ctx, 256);
+
     struct ggml_init_params params = {
-        /*.mem_size   =*/ wstate.alloc_cross.meta.size(),
-        /*.mem_buffer =*/ wstate.alloc_cross.meta.data(),
+        /*.mem_size   =*/ wstate.sched_cross.meta.size(),
+        /*.mem_buffer =*/ wstate.sched_cross.meta.data(),
         /*.no_alloc   =*/ true,
     };
 
@@ -2024,28 +2243,20 @@ static struct ggml_cgraph * whisper_build_graph_cross(
 
     ggml_cgraph * gf = ggml_new_graph(ctx0);
 
-    //ggml_allocr * alloc = wstate.alloc_cross.alloc;
-
-    //struct ggml_tensor * cur = ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, n_state, n_ctx);
-    //ggml_allocr_alloc(alloc, cur);
-
-    //if (!ggml_allocr_is_measure(alloc)) {
-    //    ggml_backend_tensor_copy(wstate.embd_enc, cur);
-    //}
     struct ggml_tensor * cur = ggml_view_tensor(ctx0, wstate.embd_enc);
 
-    const float  Kscale = pow(float(n_state) / n_head, -0.25);
+    const float  Kscale = pow(float(n_state_head), -0.25);
 
     for (int il = 0; il < model.hparams.n_text_layer; ++il) {
         auto & layer = model.layers_decoder[il];
 
-        struct ggml_tensor* Kcross = ggml_mul_mat(ctx0,
+        struct ggml_tensor * Kcross = ggml_mul_mat(ctx0,
                 layer.cross_attn_k_w,
                 cur);
 
         Kcross = ggml_scale(ctx0, Kcross, Kscale);
 
-        struct ggml_tensor* Vcross = ggml_mul_mat(ctx0,
+        struct ggml_tensor * Vcross = ggml_mul_mat(ctx0,
                 layer.cross_attn_v_w,
                 cur);
 
@@ -2053,15 +2264,25 @@ static struct ggml_cgraph * whisper_build_graph_cross(
                     Vcross,
                     layer.cross_attn_v_b);
 
-        Vcross = ggml_transpose(ctx0, ggml_reshape_2d(ctx0, Vcross, n_state, n_ctx));
+        struct ggml_tensor * k;
+        struct ggml_tensor * v;
+
+        if (wctx.params.flash_attn) {
+            k = ggml_view_1d(ctx0, wstate.kv_cross.k, n_state*n_ctx,
+                    (ggml_element_size(wstate.kv_cross.k)*n_state)*(il*n_ctx_pad));
 
-        struct ggml_tensor * k = ggml_view_1d(ctx0, wstate.kv_cross.k,
-                n_state*n_ctx,
-                (ggml_element_size(wstate.kv_cross.k)*n_state)*(il*n_ctx));
+            v = ggml_view_1d(ctx0, wstate.kv_cross.v, n_state*n_ctx,
+                    (ggml_element_size(wstate.kv_cross.v)*n_state)*(il*n_ctx_pad));
+        } else {
+            Vcross = ggml_transpose(ctx0, ggml_reshape_2d(ctx0, Vcross, n_state, n_ctx));
 
-        struct ggml_tensor * v = ggml_view_2d(ctx0, wstate.kv_cross.v, n_ctx, n_state,
-                (   n_ctx)*ggml_element_size(wstate.kv_cross.v),
-                (il*n_ctx)*ggml_element_size(wstate.kv_cross.v)*n_state);
+            k = ggml_view_1d(ctx0, wstate.kv_cross.k, n_state*n_ctx,
+                    (ggml_element_size(wstate.kv_cross.k)*n_state)*(il*n_ctx));
+
+            v = ggml_view_2d(ctx0, wstate.kv_cross.v, n_ctx, n_state,
+                    (   n_ctx)*ggml_element_size(wstate.kv_cross.v),
+                    (il*n_ctx)*ggml_element_size(wstate.kv_cross.v)*n_state);
+        }
 
         ggml_build_forward_expand(gf, ggml_cpy(ctx0, Kcross, k));
         ggml_build_forward_expand(gf, ggml_cpy(ctx0, Vcross, v));
@@ -2089,53 +2310,89 @@ static bool whisper_encode_internal(
           whisper_state & wstate,
               const int   mel_offset,
               const int   n_threads,
- whisper_abort_callback   abort_callback,
+    ggml_abort_callback   abort_callback,
                    void * abort_callback_data) {
     const int64_t t_start_us = ggml_time_us();
 
     // conv
     {
-        auto & alloc = wstate.alloc_conv.alloc;
+        auto & sched = wstate.sched_conv.sched;
+
+        ggml_cgraph * gf = whisper_build_graph_conv(wctx, wstate);
+
+        if (!ggml_backend_sched_alloc_graph(sched, gf)) {
+            // should never happen as we pre-allocate the memory
+            return false;
+        }
+
+        struct ggml_tensor * mel = ggml_graph_get_tensor(gf, "mel");
+
+        // set the input
+        {
+            const auto & mel_inp = wstate.mel;
+            const int n_ctx      = wstate.exp_n_audio_ctx > 0 ? wstate.exp_n_audio_ctx : wctx.model.hparams.n_audio_ctx;
+
+            assert(mel->type == GGML_TYPE_F32);
+            assert(mel_inp.n_mel == wctx.model.hparams.n_mels);
 
-        ggml_allocr_reset(alloc);
+            wstate.inp_mel.resize(ggml_nelements(mel));
 
-        ggml_cgraph * gf = whisper_build_graph_conv(wctx, wstate, mel_offset);
+            float * dst = wstate.inp_mel.data();
+            memset(dst, 0, ggml_nbytes(mel));
 
-        ggml_allocr_alloc_graph(alloc, gf);
+            const int i0 = std::min(mel_offset,           mel_inp.n_len);
+            const int i1 = std::min(mel_offset + 2*n_ctx, mel_inp.n_len);
+
+            for (int j = 0; j < mel_inp.n_mel; ++j) {
+                for (int i = i0; i < i1; ++i) {
+                    dst[j*2*n_ctx + (i - i0)] = mel_inp.data[j*mel_inp.n_len + i];
+                }
+            }
+
+            ggml_backend_tensor_set(mel, wstate.inp_mel.data(), 0, ggml_nelements(mel)*sizeof(float));
+        }
 
         if (!whisper_encode_external(wstate)) {
-            if (!ggml_graph_compute_helper(wstate.backend, gf, n_threads)) {
+            if (!ggml_graph_compute_helper(sched, gf, n_threads)) {
                 return false;
             }
+        } else {
+#if defined(WHISPER_USE_COREML)
+            whisper_coreml_encode(wstate.ctx_coreml, mel->ne[0], mel->ne[1], (float *) mel->data, (float *) wstate.embd_enc->data);
+#elif defined(WHISPER_USE_OPENVINO)
+            whisper_openvino_encode(wstate.ctx_openvino, mel, wstate.embd_enc);
+#endif
         }
     }
 
     // encoder
     if (!whisper_encode_external(wstate)) {
-        auto & alloc = wstate.alloc_encode.alloc;
-
-        ggml_allocr_reset(alloc);
+        auto & sched = wstate.sched_encode.sched;
 
         ggml_cgraph * gf = whisper_build_graph_encoder(wctx, wstate);
 
-        ggml_allocr_alloc_graph(alloc, gf);
+        if (!ggml_backend_sched_alloc_graph(sched, gf)) {
+            // should never happen as we pre-allocate the memory
+            return false;
+        }
 
-        if (!ggml_graph_compute_helper(wstate.backend, gf, n_threads)) {
+        if (!ggml_graph_compute_helper(sched, gf, n_threads)) {
             return false;
         }
     }
 
     // cross
     {
-        auto & alloc = wstate.alloc_cross.alloc;
-
-        ggml_allocr_reset(alloc);
+        auto & sched = wstate.sched_cross.sched;
 
         ggml_cgraph * gf = whisper_build_graph_cross(wctx, wstate);
 
-        ggml_allocr_alloc_graph(alloc, gf);
+        if (!ggml_backend_sched_alloc_graph(sched, gf)) {
+            // should never happen as we pre-allocate the memory
+            return false;
+        }
 
-        if (!ggml_graph_compute_helper(wstate.backend, gf, n_threads)) {
+        if (!ggml_graph_compute_helper(sched, gf, n_threads)) {
             return false;
         }
     }
@@ -2149,7 +2406,9 @@ static bool whisper_encode_internal(
 static struct ggml_cgraph * whisper_build_graph_decoder(
          whisper_context & wctx,
          whisper_state   & wstate,
-     const whisper_batch & batch) {
+     const whisper_batch & batch,
+                    bool   save_alignment_heads_QKs,
+                    bool   worst_case) {
     const auto & model   = wctx.model;
     const auto & hparams = model.hparams;
 
@@ -2157,24 +2416,26 @@ static struct ggml_cgraph * whisper_build_graph_decoder(
 
     WHISPER_ASSERT(!!kv_self.ctx);
 
-    ggml_allocr * alloc = wstate.alloc_decode.alloc;
-
     const int n_ctx   = kv_self.size;
     const int n_state = hparams.n_text_state;
     const int n_head  = hparams.n_text_head;
     const int n_layer = hparams.n_text_layer;
 
+    const int n_state_head = n_state/n_head;
+
     const int n_tokens    = batch.n_tokens;
     const int n_audio_ctx = wstate.exp_n_audio_ctx > 0 ? wstate.exp_n_audio_ctx : hparams.n_audio_ctx;
 
-    const int32_t n_kv     = ggml_allocr_is_measure(alloc) ? n_ctx            : kv_self.n;
-    const int32_t kv_head  = ggml_allocr_is_measure(alloc) ? n_ctx - n_tokens : kv_self.head;
+    const int n_audio_ctx_pad = GGML_PAD(n_audio_ctx, 256);
+
+    const int32_t n_kv    = worst_case ? n_ctx            : kv_self.n;
+    const int32_t kv_head = worst_case ? n_ctx - n_tokens : kv_self.head;
 
     //WHISPER_LOG_DEBUG("%s: n_past = %d, n_tokens = %d, n_audio_ctx = %d, n_ctx = %d\n", __func__, n_past, n_tokens, n_audio_ctx, n_ctx);
 
     struct ggml_init_params params = {
-        /*.mem_size   =*/ wstate.alloc_decode.meta.size(),
-        /*.mem_buffer =*/ wstate.alloc_decode.meta.data(),
+        /*.mem_size   =*/ wstate.sched_decode.meta.size(),
+        /*.mem_buffer =*/ wstate.sched_decode.meta.data(),
         /*.no_alloc   =*/ true,
     };
 
@@ -2183,48 +2444,20 @@ static struct ggml_cgraph * whisper_build_graph_decoder(
     ggml_cgraph * gf = ggml_new_graph_custom(ctx0, WHISPER_MAX_NODES, false);
 
     struct ggml_tensor * embd = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, n_tokens);
-    ggml_allocr_alloc(alloc, embd);
-
-    if (!ggml_allocr_is_measure(alloc)) {
-        ggml_backend_tensor_set(embd, batch.token, 0, n_tokens*ggml_element_size(embd));
-    }
+    ggml_set_name(embd, "embd");
+    ggml_set_input(embd);
 
     struct ggml_tensor * position = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, n_tokens);
-    ggml_allocr_alloc(alloc, position);
-
-    if (!ggml_allocr_is_measure(alloc)) {
-        for (int i = 0; i < n_tokens; ++i) {
-            const int32_t val = batch.pos[i];
-            ggml_backend_tensor_set(position, &val, i*sizeof(int32_t), sizeof(int32_t));
-        }
-    }
+    ggml_set_name(position, "position");
+    ggml_set_input(position);
 
-    const float KQscale = pow(float(n_state)/n_head, -0.25);
+    const float KQscale = pow(float(n_state_head), -0.25);
 
-    struct ggml_tensor * KQ_mask = ggml_new_tensor_3d(ctx0, GGML_TYPE_F32, n_kv, n_tokens, 1);
-    ggml_allocr_alloc(alloc, KQ_mask);
+    struct ggml_tensor * KQ_mask = ggml_new_tensor_3d(ctx0, GGML_TYPE_F32, n_kv, GGML_PAD(n_tokens, GGML_KQ_MASK_PAD), 1);
+    ggml_set_name(KQ_mask, "KQ_mask");
+    ggml_set_input(KQ_mask);
 
-    if (!ggml_allocr_is_measure(alloc)) {
-        wstate.inp_mask.resize(n_kv*n_tokens);
-
-        float * data = wstate.inp_mask.data();
-        memset(data, 0, ggml_nbytes(KQ_mask));
-
-        for (int h = 0; h < 1; ++h) {
-            for (int j = 0; j < n_tokens; ++j) {
-                const whisper_pos    pos    = batch.pos[j];
-                const whisper_seq_id seq_id = batch.seq_id[j][0];
-
-                for (int i = 0; i < n_kv; ++i) {
-                    if (!kv_self.cells[i].has_seq_id(seq_id) || kv_self.cells[i].pos > pos) {
-                        data[h*(n_kv*n_tokens) + j*n_kv + i] = -INFINITY;
-                    }
-                }
-            }
-        }
-
-        ggml_backend_tensor_set(KQ_mask, wstate.inp_mask.data(), 0, ggml_nelements(KQ_mask)*sizeof(float));
-    }
+    struct ggml_tensor * KQ_mask_f16 = ggml_cast(ctx0, KQ_mask, GGML_TYPE_F16);
 
     // token encoding + position encoding
     struct ggml_tensor * cur =
@@ -2234,6 +2467,9 @@ static struct ggml_cgraph * whisper_build_graph_decoder(
 
     struct ggml_tensor * inpL = cur;
 
+    // [EXPERIMENTAL] Token-level timestamps with DTW
+    struct ggml_tensor * aheads_cross_QKs = nullptr;
+
     for (int il = 0; il < n_layer; ++il) {
         const auto & layer = model.layers_decoder[il];
 
@@ -2278,12 +2514,25 @@ static struct ggml_cgraph * whisper_build_graph_decoder(
                             Vcur,
                             layer.attn_v_b);
 
-                Vcur = ggml_transpose(ctx0, ggml_reshape_2d(ctx0, Vcur, n_state, n_tokens));
+                struct ggml_tensor * k;
+                struct ggml_tensor * v;
 
-                struct ggml_tensor * k = ggml_view_1d(ctx0, kv_self.k, n_tokens*n_state, (ggml_element_size(kv_self.k)*n_state)*(il*n_ctx + kv_head));
-                struct ggml_tensor * v = ggml_view_2d(ctx0, kv_self.v, n_tokens, n_state,
-                        (   n_ctx)*ggml_element_size(kv_self.v),
-                        (il*n_ctx)*ggml_element_size(kv_self.v)*n_state + kv_head*ggml_element_size(kv_self.v));
+                if (wctx.params.flash_attn) {
+                    k = ggml_view_1d(ctx0, kv_self.k, n_tokens*n_state,
+                            (ggml_element_size(kv_self.k)*n_state)*(il*n_ctx + kv_head));
+
+                    v = ggml_view_1d(ctx0, kv_self.v, n_tokens*n_state,
+                            (ggml_element_size(kv_self.v)*n_state)*(il*n_ctx + kv_head));
+                } else {
+                    Vcur = ggml_transpose(ctx0, ggml_reshape_2d(ctx0, Vcur, n_state, n_tokens));
+
+                    k = ggml_view_1d(ctx0, kv_self.k, n_tokens*n_state,
+                            (ggml_element_size(kv_self.k)*n_state)*(il*n_ctx + kv_head));
+
+                    v = ggml_view_2d(ctx0, kv_self.v, n_tokens, n_state,
+                            (   n_ctx)*ggml_element_size(kv_self.v),
+                            (il*n_ctx)*ggml_element_size(kv_self.v)*n_state + kv_head*ggml_element_size(kv_self.v));
+                }
 
                 ggml_build_forward_expand(gf, ggml_cpy(ctx0, Kcur, k));
                 ggml_build_forward_expand(gf, ggml_cpy(ctx0, Vcur, v));
@@ -2293,40 +2542,46 @@ static struct ggml_cgraph * whisper_build_graph_decoder(
 
             struct ggml_tensor * Q =
                 ggml_permute(ctx0,
-                        ggml_reshape_3d(ctx0, Qcur, n_state/n_head, n_head, n_tokens),
+                        ggml_reshape_3d(ctx0, Qcur, n_state_head, n_head, n_tokens),
                         0, 2, 1, 3);
 
             struct ggml_tensor * K =
                 ggml_view_3d(ctx0, kv_self.k,
-                        n_state/n_head, n_kv, n_head,
+                        n_state_head, n_kv, n_head,
                         ggml_element_size(kv_self.k)*n_state,
-                        ggml_element_size(kv_self.k)*n_state/n_head,
+                        ggml_element_size(kv_self.k)*n_state_head,
                         ggml_element_size(kv_self.k)*n_state*n_ctx*il);
 
-            // K * Q
-            struct ggml_tensor * KQ = ggml_mul_mat(ctx0, K, Q);
+            if (wctx.params.flash_attn) {
+                struct ggml_tensor * V =
+                    ggml_view_3d(ctx0, kv_self.v,
+                            n_state_head, n_kv, n_head,
+                            ggml_element_size(kv_self.v)*n_state,
+                            ggml_element_size(kv_self.v)*n_state_head,
+                            ggml_element_size(kv_self.v)*n_state*n_ctx*il);
 
-            //struct ggml_tensor * KQ_scaled = ggml_scale(ctx0, KQ, KQ_scale);
+                cur = ggml_flash_attn_ext(ctx0, Q, K, V, KQ_mask_f16, 1.0f, 0.0f, 0.0f);
 
-            //struct ggml_tensor * KQ_masked = ggml_diag_mask_inf(ctx0, KQ, n_past);
-            struct ggml_tensor * KQ_masked = ggml_add(ctx0, KQ, KQ_mask);
+                cur = ggml_reshape_2d(ctx0, cur, n_state, n_tokens);
+            } else {
+                // K * Q
+                struct ggml_tensor * KQ = ggml_mul_mat(ctx0, K, Q);
 
-            struct ggml_tensor * KQ_soft_max = ggml_soft_max(ctx0, KQ_masked);
+                struct ggml_tensor * KQ_soft_max = ggml_soft_max_ext(ctx0, KQ, KQ_mask, 1.0f, 0.0f);
 
-            struct ggml_tensor * V =
-                ggml_view_3d(ctx0, kv_self.v,
-                        n_kv, n_state/n_head, n_head,
-                        n_ctx*ggml_element_size(kv_self.v),
-                        n_ctx*ggml_element_size(kv_self.v)*n_state/n_head,
-                        n_ctx*ggml_element_size(kv_self.v)*n_state*il);
+                struct ggml_tensor * V =
+                    ggml_view_3d(ctx0, kv_self.v,
+                            n_kv, n_state_head, n_head,
+                            n_ctx*ggml_element_size(kv_self.v),
+                            n_ctx*ggml_element_size(kv_self.v)*n_state_head,
+                            n_ctx*ggml_element_size(kv_self.v)*n_state*il);
 
-            struct ggml_tensor * KQV = ggml_mul_mat(ctx0, V, KQ_soft_max);
+                struct ggml_tensor * KQV = ggml_mul_mat(ctx0, V, KQ_soft_max);
 
-            struct ggml_tensor * KQV_merged = ggml_permute(ctx0, KQV, 0, 2, 1, 3);
+                struct ggml_tensor * KQV_merged = ggml_permute(ctx0, KQV, 0, 2, 1, 3);
 
-            cur = ggml_cpy(ctx0,
-                    KQV_merged,
-                    ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, n_state, n_tokens));
+                cur = ggml_cont_2d(ctx0, KQV_merged, n_state, n_tokens);
+            }
         }
 
         // projection
@@ -2365,62 +2620,75 @@ static struct ggml_cgraph * whisper_build_graph_decoder(
                         Qcur,
                         layer.cross_attn_q_b);
 
-            Qcur = ggml_scale(ctx0, Qcur, KQscale);
-
-            // Kcross is already scaled
-            struct ggml_tensor * Kcross =
-                ggml_view_3d(ctx0, wstate.kv_cross.k,
-                        n_state/n_head, n_audio_ctx, n_head,
-                        ggml_element_size(wstate.kv_cross.k)*n_state,
-                        ggml_element_size(wstate.kv_cross.k)*n_state/n_head,
-                        ggml_element_size(wstate.kv_cross.k)*n_state*n_audio_ctx*il);
-
-            //struct ggml_tensor * Vcross =
-            //    ggml_reshape_3d(ctx0,
-            //            ggml_view_1d(ctx0, wstate.kv_cross.v, n_audio_ctx*n_state, il*n_audio_ctx*ggml_element_size(wstate.kv_cross.v)*n_state),
-            //            n_state/n_head, n_head, n_audio_ctx);
-
-            //struct ggml_tensor * V_trans =
-            //    ggml_cpy(ctx0,
-            //            ggml_permute(ctx0, Vcross, 1, 2, 0, 3),
-            //            ggml_new_tensor_3d(ctx0, Vcross->type, n_audio_ctx, n_state/n_head, n_head));
-
-            struct ggml_tensor * V =
-                ggml_view_3d(ctx0, wstate.kv_cross.v,
-                        n_audio_ctx, n_state/n_head, n_head,
-                        n_audio_ctx*ggml_element_size(wstate.kv_cross.v),
-                        n_audio_ctx*ggml_element_size(wstate.kv_cross.v)*n_state/n_head,
-                        n_audio_ctx*ggml_element_size(wstate.kv_cross.v)*n_state*il);
-
-            // ------
-
             struct ggml_tensor * Q =
                 ggml_permute(ctx0,
-                        ggml_reshape_3d(ctx0, Qcur, n_state/n_head, n_head, n_tokens),
+                        ggml_reshape_3d(ctx0, Qcur, n_state_head, n_head, n_tokens),
                         0, 2, 1, 3);
 
-            // K * Q
-            struct ggml_tensor * KQ = ggml_mul_mat(ctx0, Kcross, Q);
+            if (wctx.params.flash_attn) {
+                struct ggml_tensor * Kcross =
+                    ggml_view_3d(ctx0, wstate.kv_cross.k,
+                            n_state_head, n_audio_ctx_pad, n_head,
+                            ggml_element_size(wstate.kv_cross.k)*n_state,
+                            ggml_element_size(wstate.kv_cross.k)*n_state_head,
+                            ggml_element_size(wstate.kv_cross.k)*n_state*n_audio_ctx_pad*il);
 
-            //struct ggml_tensor * KQ_scaled =
-            //    ggml_scale(ctx0,
-            //            KQ,
-            //            ggml_new_f32(ctx0, 1.0f/sqrt(float(n_state)/n_head))
-            //            );
+                struct ggml_tensor * Vcross =
+                    ggml_view_3d(ctx0, wstate.kv_cross.v,
+                            n_state_head, n_audio_ctx_pad, n_head,
+                            ggml_element_size(wstate.kv_cross.v)*n_state,
+                            ggml_element_size(wstate.kv_cross.v)*n_state_head,
+                            ggml_element_size(wstate.kv_cross.v)*n_state*n_audio_ctx_pad*il);
 
-            // no masking for cross-attention
-            //struct ggml_tensor * KQ_masked = ggml_diag_mask_inf(ctx0, KQ_scaled, n_past);
+                cur = ggml_flash_attn_ext(ctx0, Q, Kcross, Vcross, nullptr, KQscale, 0.0f, 0.0f);
 
-            struct ggml_tensor * KQ_soft_max = ggml_soft_max(ctx0, KQ);
+                cur = ggml_reshape_2d(ctx0, cur, n_state, n_tokens);
+            } else {
+                struct ggml_tensor * Kcross =
+                    ggml_view_3d(ctx0, wstate.kv_cross.k,
+                            n_state_head, n_audio_ctx, n_head,
+                            ggml_element_size(wstate.kv_cross.k)*n_state,
+                            ggml_element_size(wstate.kv_cross.k)*n_state_head,
+                            ggml_element_size(wstate.kv_cross.k)*n_state*n_audio_ctx*il);
+
+                struct ggml_tensor * Vcross =
+                    ggml_view_3d(ctx0, wstate.kv_cross.v,
+                            n_audio_ctx, n_state_head, n_head,
+                            n_audio_ctx*ggml_element_size(wstate.kv_cross.v),
+                            n_audio_ctx*ggml_element_size(wstate.kv_cross.v)*n_state_head,
+                            n_audio_ctx*ggml_element_size(wstate.kv_cross.v)*n_state*il);
+
+                // ------
+
+                // K * Q
+                struct ggml_tensor * KQ = ggml_mul_mat(ctx0, Kcross, Q);
+
+                struct ggml_tensor * KQ_soft_max = ggml_soft_max_ext(ctx0, KQ, nullptr, KQscale, 0.0f);
+
+                // [EXPERIMENTAL] Token-level timestamps with DTW
+                if (wctx.params.dtw_token_timestamps) {
+                    if (wstate.aheads_masks.m[il] != nullptr) {
+                        struct ggml_tensor * aheads_KQs = ggml_reshape_2d(ctx0, KQ_soft_max, KQ_soft_max->ne[0] * KQ_soft_max->ne[1], KQ_soft_max->ne[2]);
+                        aheads_KQs = ggml_transpose(ctx0, aheads_KQs);
+                        aheads_KQs = ggml_cont(ctx0, aheads_KQs);
+                        aheads_KQs = ggml_mul_mat(ctx0, wstate.aheads_masks.m[il], aheads_KQs);
+                        aheads_KQs = ggml_transpose(ctx0, aheads_KQs);
+                        aheads_KQs = ggml_cont(ctx0, aheads_KQs);
+                        aheads_KQs = ggml_reshape_3d(ctx0, aheads_KQs, KQ_soft_max->ne[0], KQ_soft_max->ne[1], wstate.aheads_masks.m[il]->ne[1]);
+                        if (aheads_cross_QKs == NULL) {
+                            aheads_cross_QKs = aheads_KQs;
+                        } else {
+                            aheads_cross_QKs = ggml_concat(ctx0, aheads_cross_QKs, aheads_KQs, 2);
+                        }
+                    }
+                }
 
-            struct ggml_tensor * KQV = ggml_mul_mat(ctx0, V, KQ_soft_max);
+                struct ggml_tensor * KQV = ggml_mul_mat(ctx0, Vcross, KQ_soft_max);
 
-            struct ggml_tensor * KQV_merged = ggml_permute(ctx0, KQV, 0, 2, 1, 3);
+                struct ggml_tensor * KQV_merged = ggml_permute(ctx0, KQV, 0, 2, 1, 3);
 
-            // cur = KQV_merged.contiguous().view(n_state, n_tokens)
-            cur = ggml_cpy(ctx0,
-                    KQV_merged,
-                    ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, n_state, n_tokens));
+                cur = ggml_cont_2d(ctx0, KQV_merged, n_state, n_tokens);
+            }
         }
 
         // projection
@@ -2498,6 +2766,16 @@ static struct ggml_cgraph * whisper_build_graph_decoder(
 
     struct ggml_tensor * logits = ggml_mul_mat(ctx0, model.d_te, cur);
 
+    // [EXPERIMENTAL] Token-level timestamps with DTW
+    if (wctx.params.dtw_token_timestamps && aheads_cross_QKs != nullptr) {
+        aheads_cross_QKs = ggml_transpose(ctx0, aheads_cross_QKs);
+        aheads_cross_QKs = ggml_cont(ctx0, aheads_cross_QKs);
+        if (save_alignment_heads_QKs) {
+            ggml_build_forward_expand(gf, aheads_cross_QKs);
+            wstate.aheads_cross_QKs = aheads_cross_QKs;
+        }
+    }
+
     ggml_build_forward_expand(gf, logits);
 
     ggml_free(ctx0);
@@ -2520,7 +2798,8 @@ static bool whisper_decode_internal(
           whisper_state & wstate,
     const whisper_batch & batch,
               const int   n_threads,
- whisper_abort_callback   abort_callback,
+                   bool   save_alignment_heads_QKs,
+    ggml_abort_callback   abort_callback,
                    void * abort_callback_data) {
     const int64_t t_start_us = ggml_time_us();
 
@@ -2542,24 +2821,75 @@ static bool whisper_decode_internal(
             return false;
         }
 
-        kv_self.n = whisper_kv_cache_cell_max(kv_self);
+        const uint32_t pad = whisper_kv_cache_get_padding(wctx);
+        kv_self.n = std::min(kv_self.size, std::max(pad, GGML_PAD(whisper_kv_cache_cell_max(kv_self), pad)));
+
         //kv_self.n = std::min((int32_t) hparams.n_text_ctx, std::max(32, whisper_kv_cache_cell_max(kv_self)));
         //printf("n_tokens = %5d, kv_self.head = %5d, kv_self.n = %5d, seq_id = %5d\n", batch.n_tokens, kv_self.head, kv_self.n, batch.seq_id[0][0]);
     }
 
     // decoder
     {
-        auto & alloc = wstate.alloc_decode.alloc;
+        auto & sched = wstate.sched_decode.sched;
+
+        ggml_cgraph * gf = whisper_build_graph_decoder(wctx, wstate, batch, save_alignment_heads_QKs, false);
+
+        if (!ggml_backend_sched_alloc_graph(sched, gf)) {
+            // should never happen as we pre-allocate the memory
+            return false;
+        }
+
+        // set the inputs
+        {
+            struct ggml_tensor * embd = ggml_graph_get_tensor(gf, "embd");
+            ggml_backend_tensor_set(embd, batch.token, 0, n_tokens*ggml_element_size(embd));
+        }
+
+        {
+            struct ggml_tensor * position = ggml_graph_get_tensor(gf, "position");
+            for (int i = 0; i < n_tokens; ++i) {
+                const int32_t val = batch.pos[i];
+                ggml_backend_tensor_set(position, &val, i*sizeof(int32_t), sizeof(int32_t));
+            }
+        }
+
+        {
+            struct ggml_tensor * KQ_mask = ggml_graph_get_tensor(gf, "KQ_mask");
+
+            auto & kv_self = wstate.kv_self;
+
+            const int32_t n_kv = kv_self.n;
+
+            wstate.inp_mask.resize(ggml_nelements(KQ_mask));
+
+            float * data = wstate.inp_mask.data();
+            memset(data, 0, ggml_nbytes(KQ_mask));
+
+            for (int h = 0; h < 1; ++h) {
+                for (int j = 0; j < n_tokens; ++j) {
+                    const whisper_pos    pos    = batch.pos[j];
+                    const whisper_seq_id seq_id = batch.seq_id[j][0];
 
-        ggml_allocr_reset(alloc);
+                    for (int i = 0; i < n_kv; ++i) {
+                        if (!kv_self.cells[i].has_seq_id(seq_id) || kv_self.cells[i].pos > pos) {
+                            data[h*(n_kv*n_tokens) + j*n_kv + i] = -INFINITY;
+                        }
+                    }
+                }
 
-        ggml_cgraph * gf = whisper_build_graph_decoder(wctx, wstate, batch);
+                for (int i = n_tokens; i < GGML_PAD(n_tokens, GGML_KQ_MASK_PAD); ++i) {
+                    for (int j = 0; j < n_kv; ++j) {
+                        data[h*(n_kv*n_tokens) + i*n_kv + j] = -INFINITY;
+                    }
+                }
+            }
 
-        ggml_allocr_alloc_graph(alloc, gf);
+            ggml_backend_tensor_set(KQ_mask, wstate.inp_mask.data(), 0, ggml_nelements(KQ_mask)*sizeof(float));
+        }
 
-        logits = gf->nodes[gf->n_nodes - 1];
+        logits = ggml_graph_node(gf, -1);
 
-        if (!ggml_graph_compute_helper(wstate.backend, gf, n_threads)) {
+        if (!ggml_graph_compute_helper(sched, gf, n_threads)) {
             return false;
         }
     }
@@ -2613,29 +2943,47 @@ static std::string to_timestamp(int64_t t, bool comma = false) {
 }
 
 #define SIN_COS_N_COUNT WHISPER_N_FFT
-static float sin_vals[SIN_COS_N_COUNT];
-static float cos_vals[SIN_COS_N_COUNT];
+namespace {
+struct whisper_global_cache {
+    // In FFT, we frequently use sine and cosine operations with the same values.
+    // We can use precalculated values to speed up the process.
+    float sin_vals[SIN_COS_N_COUNT];
+    float cos_vals[SIN_COS_N_COUNT];
+
+    // Hann window (Use cosf to eliminate difference)
+    // ref: https://pytorch.org/docs/stable/generated/torch.hann_window.html
+    // ref: https://github.com/openai/whisper/blob/main/whisper/audio.py#L147
+    float hann_window[WHISPER_N_FFT];
+
+    whisper_global_cache() {
+        fill_sin_cos_table();
+        fill_hann_window(sizeof(hann_window)/sizeof(hann_window[0]), true, hann_window);
+    }
+
+    void fill_sin_cos_table() {
+        for (int i = 0; i < SIN_COS_N_COUNT; i++) {
+            double theta = (2 * M_PI * i) / SIN_COS_N_COUNT;
+            sin_vals[i] = sinf(theta);
+            cos_vals[i] = cosf(theta);
+        }
+    }
 
-// In FFT, we frequently use sine and cosine operations with the same values.
-// We can use precalculated values to speed up the process.
-static void fill_sin_cos_table() {
-    static bool is_filled = false;
-    if (is_filled) return;
-    for (int i = 0; i < SIN_COS_N_COUNT; i++) {
-        double theta = (2*M_PI*i)/SIN_COS_N_COUNT;
-        sin_vals[i] = sinf(theta);
-        cos_vals[i] = cosf(theta);
+    void fill_hann_window(int length, bool periodic, float * output) {
+        int offset = -1;
+        if (periodic) {
+            offset = 0;
+        }
+        for (int i = 0; i < length; i++) {
+            output[i] = 0.5 * (1.0 - cosf((2.0 * M_PI * i) / (length + offset)));
+        }
     }
-    is_filled = true;
+} global_cache;
 }
 
 // naive Discrete Fourier Transform
 // input is real-valued
 // output is complex-valued
-static void dft(const std::vector<float> & in, std::vector<float> & out) {
-    int N = in.size();
-
-    out.resize(N*2);
+static void dft(const float* in, int N, float* out) {
     const int sin_cos_step = SIN_COS_N_COUNT / N;
 
     for (int k = 0; k < N; k++) {
@@ -2644,8 +2992,8 @@ static void dft(const std::vector<float> & in, std::vector<float> & out) {
 
         for (int n = 0; n < N; n++) {
             int idx = (k * n * sin_cos_step) % (SIN_COS_N_COUNT); // t = 2*M_PI*k*n/N
-            re += in[n]*cos_vals[idx]; // cos(t)
-            im -= in[n]*sin_vals[idx]; // sin(t)
+            re += in[n]*global_cache.cos_vals[idx]; // cos(t)
+            im -= in[n]*global_cache.sin_vals[idx]; // sin(t)
         }
 
         out[k*2 + 0] = re;
@@ -2657,47 +3005,38 @@ static void dft(const std::vector<float> & in, std::vector<float> & out) {
 // poor man's implementation - use something better
 // input is real-valued
 // output is complex-valued
-static void fft(const std::vector<float> & in, std::vector<float> & out) {
-    out.resize(in.size()*2);
-
-    int N = in.size();
-
+static void fft(float* in, int N, float* out) {
     if (N == 1) {
         out[0] = in[0];
         out[1] = 0;
         return;
     }
 
-    if (N%2 == 1) {
-        dft(in, out);
+    const int half_N = N / 2;
+    if (N - half_N*2 == 1) {
+        dft(in, N, out);
         return;
     }
 
-    std::vector<float> even;
-    std::vector<float> odd;
-
-    even.reserve(N/2);
-    odd.reserve(N/2);
-
-    for (int i = 0; i < N; i++) {
-        if (i % 2 == 0) {
-            even.push_back(in[i]);
-        } else {
-            odd.push_back(in[i]);
-        }
+    float* even = in + N;
+    for (int i = 0; i < half_N; ++i) {
+        even[i]= in[2*i];
     }
+    float* even_fft = out + 2 * N;
+    fft(even, half_N, even_fft);
 
-    std::vector<float> even_fft;
-    std::vector<float> odd_fft;
-
-    fft(even, even_fft);
-    fft(odd, odd_fft);
+    float* odd = even;
+    for (int i = 0; i < half_N; ++i) {
+        odd[i] = in[2*i + 1];
+    }
+    float* odd_fft = even_fft + N;
+    fft(odd, half_N, odd_fft);
 
     const int sin_cos_step = SIN_COS_N_COUNT / N;
-    for (int k = 0; k < N/2; k++) {
+    for (int k = 0; k < half_N; k++) {
         int idx = k * sin_cos_step; // t = 2*M_PI*k/N
-        float re = cos_vals[idx]; // cos(t)
-        float im = -sin_vals[idx]; // sin(t)
+        float re = global_cache.cos_vals[idx]; // cos(t)
+        float im = -global_cache.sin_vals[idx]; // sin(t)
 
         float re_odd = odd_fft[2*k + 0];
         float im_odd = odd_fft[2*k + 1];
@@ -2705,61 +3044,49 @@ static void fft(const std::vector<float> & in, std::vector<float> & out) {
         out[2*k + 0] = even_fft[2*k + 0] + re*re_odd - im*im_odd;
         out[2*k + 1] = even_fft[2*k + 1] + re*im_odd + im*re_odd;
 
-        out[2*(k + N/2) + 0] = even_fft[2*k + 0] - re*re_odd + im*im_odd;
-        out[2*(k + N/2) + 1] = even_fft[2*k + 1] - re*im_odd - im*re_odd;
-    }
-}
-
-static bool hann_window(int length, bool periodic, std::vector<float> & output) {
-    if (output.size() < static_cast<size_t>(length)) {
-        output.resize(length);
+        out[2*(k + half_N) + 0] = even_fft[2*k + 0] - re*re_odd + im*im_odd;
+        out[2*(k + half_N) + 1] = even_fft[2*k + 1] - re*im_odd - im*re_odd;
     }
-    int offset = -1;
-    if (periodic) {
-        offset = 0;
-    }
-    for (int i = 0; i < length; i++) {
-        output[i] = 0.5*(1.0 - cosf((2.0*M_PI*i)/(length + offset)));
-    }
-
-    return true;
 }
 
-static void log_mel_spectrogram_worker_thread(int ith, const std::vector<float> & hann, const std::vector<float> & samples,
+static void log_mel_spectrogram_worker_thread(int ith, const float * hann, const std::vector<float> & samples,
                                               int n_samples, int frame_size, int frame_step, int n_threads,
                                               const whisper_filters & filters, whisper_mel & mel) {
-    std::vector<float> fft_in(frame_size, 0.0);
-    std::vector<float> fft_out(2 * frame_step);
-    // make sure n_fft == 1 + (WHISPER_N_FFT / 2), bin_0 to bin_nyquist
-    int n_fft = 1 + (frame_size / 2);
+    std::vector<float> fft_in(frame_size * 2, 0.0);
+    std::vector<float> fft_out(frame_size * 2 * 2 * 2);
+
+    int n_fft = filters.n_fft;
     int i = ith;
 
+    // make sure n_fft == 1 + (WHISPER_N_FFT / 2), bin_0 to bin_nyquist
+    assert(n_fft == 1 + (frame_size / 2));
+
     // calculate FFT only when fft_in are not all zero
     for (; i < std::min(n_samples / frame_step + 1, mel.n_len); i += n_threads) {
         const int offset = i * frame_step;
 
-        // apply Hanning window (~10% faster)
+        // apply Hann window (~10% faster)
         for (int j = 0; j < std::min(frame_size, n_samples - offset); j++) {
             fft_in[j] = hann[j] * samples[offset + j];
         }
+
         // fill the rest with zeros
         if (n_samples - offset < frame_size) {
             std::fill(fft_in.begin() + (n_samples - offset), fft_in.end(), 0.0);
         }
 
         // FFT
-        fft(fft_in, fft_out);
+        fft(fft_in.data(), frame_size, fft_out.data());
 
         // Calculate modulus^2 of complex numbers
         // Use pow(fft_out[2 * j + 0], 2) + pow(fft_out[2 * j + 1], 2) causes inference quality problem? Interesting.
-        for (int j = 0; j < frame_size; j++) {
+        for (int j = 0; j < n_fft; j++) {
             fft_out[j] = (fft_out[2 * j + 0] * fft_out[2 * j + 0] + fft_out[2 * j + 1] * fft_out[2 * j + 1]);
         }
 
         // mel spectrogram
         for (int j = 0; j < mel.n_mel; j++) {
             double sum = 0.0;
-
             // unroll loop (suggested by GH user @lunixbochs)
             int k = 0;
             for (k = 0; k < n_fft - 3; k += 4) {
@@ -2769,14 +3096,11 @@ static void log_mel_spectrogram_worker_thread(int ith, const std::vector<float>
                         fft_out[k + 2] * filters.data[j * n_fft + k + 2] +
                         fft_out[k + 3] * filters.data[j * n_fft + k + 3];
             }
-
             // handle n_fft remainder
             for (; k < n_fft; k++) {
                 sum += fft_out[k] * filters.data[j * n_fft + k];
             }
-
             sum = log10(std::max(sum, 1e-10));
-
             mel.data[j * mel.n_len + i] = sum;
         }
     }
@@ -2805,12 +3129,9 @@ static bool log_mel_spectrogram(
               whisper_mel & mel) {
     const int64_t t_start_us = ggml_time_us();
 
-    // Hanning window (Use cosf to eliminate difference)
-    // ref: https://pytorch.org/docs/stable/generated/torch.hann_window.html
-    // ref: https://github.com/openai/whisper/blob/main/whisper/audio.py#L147
-    std::vector<float> hann;
-    hann_window(frame_size, true, hann);
-
+    // Hann window
+    WHISPER_ASSERT(frame_size == WHISPER_N_FFT && "Unsupported frame_size");
+    const float * hann = global_cache.hann_window;
 
     // Calculate the length of padding
     int64_t stage_1_pad = WHISPER_SAMPLE_RATE * 30;
@@ -2835,12 +3156,11 @@ static bool log_mel_spectrogram(
     mel.n_len_org = 1 + (n_samples + stage_2_pad - frame_size) / frame_step;
     mel.data.resize(mel.n_mel * mel.n_len);
 
-
     {
         std::vector<std::thread> workers(n_threads - 1);
         for (int iw = 0; iw < n_threads - 1; ++iw) {
             workers[iw] = std::thread(
-                    log_mel_spectrogram_worker_thread, iw + 1, std::cref(hann), samples_padded,
+                    log_mel_spectrogram_worker_thread, iw + 1, hann, samples_padded,
                     n_samples + stage_2_pad, frame_size, frame_step, n_threads,
                     std::cref(filters), std::ref(mel));
         }
@@ -3000,19 +3320,24 @@ static std::string whisper_openvino_get_path_cache(std::string path_bin) {
 #endif
 
 struct whisper_state * whisper_init_state(whisper_context * ctx) {
-    fill_sin_cos_table();
-
     whisper_state * state = new whisper_state;
 
-    state->backend = whisper_backend_init(ctx->params);
-
-    // at this point, we don't know yet how many decoders will be used, so we overallocate 3x ctx
-    // in theory, there can be a case where this is not enough, but in practice it should always be enough
-    const int factor = 3;
+    state->backends = whisper_backend_init(ctx->params);
+    if (state->backends.empty()) {
+        WHISPER_LOG_ERROR("%s: whisper_backend_init() failed\n", __func__);
+        whisper_free_state(state);
+        return nullptr;
+    }
 
-    if (!kv_cache_init(ctx->model.hparams, state->kv_self, ctx->backend, ctx->itype, factor*ctx->model.hparams.n_text_ctx)) {
-        WHISPER_LOG_ERROR("%s: kv_cache_init() failed for self-attention cache\n", __func__);
-        delete state;
+    // at this point, we don't know yet how many decoders will be used
+    // later during decoding, if more decoders are used, we will recreate the KV cache respectively
+    state->kv_self_n_dec = 1;
+    if (!whisper_kv_cache_init(state->kv_self, state->backends[0], ctx->itype,
+                ctx->model.hparams.n_text_state,
+                ctx->model.hparams.n_text_layer,
+                GGML_PAD(ctx->model.hparams.n_text_ctx, 256))) {
+        WHISPER_LOG_ERROR("%s: whisper_kv_cache_init() failed for self-attention cache\n", __func__);
+        whisper_free_state(state);
         return nullptr;
     }
 
@@ -3021,9 +3346,12 @@ struct whisper_state * whisper_init_state(whisper_context * ctx) {
         WHISPER_LOG_INFO("%s: kv self size  = %7.2f MB\n", __func__, memory_size / 1e6);
     }
 
-    if (!kv_cache_init(ctx->model.hparams, state->kv_cross, ctx->backend, ctx->itype, ctx->model.hparams.n_audio_ctx)) {
-        WHISPER_LOG_ERROR("%s: kv_cache_init() failed for cross-attention cache\n", __func__);
-        delete state;
+    if (!whisper_kv_cache_init(state->kv_cross, state->backends[0], ctx->itype,
+                ctx->model.hparams.n_text_state,
+                ctx->model.hparams.n_text_layer,
+                GGML_PAD(ctx->model.hparams.n_audio_ctx, 256))) {
+        WHISPER_LOG_ERROR("%s: whisper_kv_cache_init() failed for cross-attention cache\n", __func__);
+        whisper_free_state(state);
         return nullptr;
     }
 
@@ -3032,6 +3360,31 @@ struct whisper_state * whisper_init_state(whisper_context * ctx) {
         WHISPER_LOG_INFO("%s: kv cross size = %7.2f MB\n", __func__, memory_size / 1e6);
     }
 
+    if (!whisper_kv_cache_init(state->kv_pad, state->backends[0], ctx->itype,
+                ctx->model.hparams.n_audio_state,
+                1,
+                GGML_PAD(ctx->model.hparams.n_audio_ctx, 256))) {
+        WHISPER_LOG_ERROR("%s: whisper_kv_cache_init() failed for self-attention cache\n", __func__);
+        whisper_free_state(state);
+        return nullptr;
+    }
+
+    {
+        const size_t memory_size = ggml_nbytes(state->kv_pad.k) + ggml_nbytes(state->kv_pad.v);
+        WHISPER_LOG_INFO("%s: kv pad  size  = %7.2f MB\n", __func__, memory_size / 1e6);
+    }
+
+    // [EXPERIMENTAL] Token-level timestamps with DTW
+    if (ctx->params.dtw_token_timestamps) {
+        if (!aheads_masks_init(ctx->params, ctx->model.hparams, state->aheads_masks, state->backends[0])) {
+            WHISPER_LOG_ERROR("%s: aheads_masks_init() failed for alignment heads masks\n", __func__);
+            whisper_free_state(state);
+            return nullptr;
+        }
+        const size_t memory_size = aheads_masks_nbytes(state->aheads_masks);
+        WHISPER_LOG_INFO("%s: alignment heads masks size = %ld B\n", __func__, memory_size);
+    }
+
 #ifdef WHISPER_USE_COREML
     const auto path_coreml = whisper_get_coreml_path_encoder(ctx->path_model);
 
@@ -3042,7 +3395,7 @@ struct whisper_state * whisper_init_state(whisper_context * ctx) {
     if (!state->ctx_coreml) {
         WHISPER_LOG_ERROR("%s: failed to load Core ML model from '%s'\n", __func__, path_coreml.c_str());
 #ifndef WHISPER_COREML_ALLOW_FALLBACK
-        delete state;
+        whisper_free_state(state);
         return nullptr;
 #endif
     } else {
@@ -3066,37 +3419,55 @@ struct whisper_state * whisper_init_state(whisper_context * ctx) {
 
     // conv allocator
     {
-        whisper_allocr_graph_init(state->alloc_conv, ctx->backend,
+        bool ok = whisper_sched_graph_init(state->sched_conv, state->backends,
                 [&]() {
-                    return whisper_build_graph_conv(*ctx, *state, 0);
+                    return whisper_build_graph_conv(*ctx, *state);
                 });
 
-        WHISPER_LOG_INFO("%s: compute buffer (conv)   = %7.2f MB\n", __func__, whisper_allocr_size(state->alloc_conv) / 1e6);
+        if (!ok) {
+            WHISPER_LOG_ERROR("%s: failed to init conv allocator\n", __func__);
+            whisper_free_state(state);
+            return nullptr;
+        }
+
+        WHISPER_LOG_INFO("%s: compute buffer (conv)   = %7.2f MB\n", __func__, whisper_sched_size(state->sched_conv) / 1e6);
     }
 
     // encoder allocator
     if (!whisper_encode_external(*state)) {
-        whisper_allocr_graph_init(state->alloc_encode, ctx->backend,
+        bool ok = whisper_sched_graph_init(state->sched_encode, state->backends,
                 [&]() {
                     return whisper_build_graph_encoder(*ctx, *state);
                 });
 
-        WHISPER_LOG_INFO("%s: compute buffer (encode) = %7.2f MB\n", __func__, whisper_allocr_size(state->alloc_encode) / 1e6);
+        if (!ok) {
+            WHISPER_LOG_ERROR("%s: failed to init encoder allocator\n", __func__);
+            whisper_free_state(state);
+            return nullptr;
+        }
+
+        WHISPER_LOG_INFO("%s: compute buffer (encode) = %7.2f MB\n", __func__, whisper_sched_size(state->sched_encode) / 1e6);
     }
 
     // cross allocator
     {
-        whisper_allocr_graph_init(state->alloc_cross, ctx->backend,
+        bool ok = whisper_sched_graph_init(state->sched_cross, state->backends,
                 [&]() {
                     return whisper_build_graph_cross(*ctx, *state);
                 });
 
-        WHISPER_LOG_INFO("%s: compute buffer (cross)  = %7.2f MB\n", __func__, whisper_allocr_size(state->alloc_cross) / 1e6);
+        if (!ok) {
+            WHISPER_LOG_ERROR("%s: failed to init cross allocator\n", __func__);
+            whisper_free_state(state);
+            return nullptr;
+        }
+
+        WHISPER_LOG_INFO("%s: compute buffer (cross)  = %7.2f MB\n", __func__, whisper_sched_size(state->sched_cross) / 1e6);
     }
 
     // decoder allocator
     {
-        whisper_allocr_graph_init(state->alloc_decode, ctx->backend,
+        bool ok = whisper_sched_graph_init(state->sched_decode, state->backends,
                 [&]() {
                     const auto & hparams = ctx->model.hparams;
 
@@ -3106,16 +3477,17 @@ struct whisper_state * whisper_init_state(whisper_context * ctx) {
 
                     whisper_batch_prep_legacy(state->batch, nullptr, n_tokens, n_past, 0);
 
-                    return whisper_build_graph_decoder(*ctx, *state, state->batch);
+                    return whisper_build_graph_decoder(*ctx, *state, state->batch, ctx->params.dtw_token_timestamps, true);
                 });
 
-        WHISPER_LOG_INFO("%s: compute buffer (decode) = %7.2f MB\n", __func__, whisper_allocr_size(state->alloc_decode) / 1e6);
-    }
+        if (!ok) {
+            WHISPER_LOG_ERROR("%s: failed to init decoder allocator\n", __func__);
+            whisper_free_state(state);
+            return nullptr;
+        }
 
-    whisper_allocr_graph_realloc(state->alloc_conv,   ctx->backend);
-    whisper_allocr_graph_realloc(state->alloc_encode, ctx->backend);
-    whisper_allocr_graph_realloc(state->alloc_cross,  ctx->backend);
-    whisper_allocr_graph_realloc(state->alloc_decode, ctx->backend);
+        WHISPER_LOG_INFO("%s: compute buffer (decode) = %7.2f MB\n", __func__, whisper_sched_size(state->sched_decode) / 1e6);
+    }
 
     return state;
 }
@@ -3171,15 +3543,32 @@ int whisper_ctx_init_openvino_encoder(
 
 struct whisper_context_params whisper_context_default_params() {
     struct whisper_context_params result = {
-        /*.use_gpu    =*/ true,
+        /*.use_gpu              =*/ true,
+        /*.flash_attn           =*/ false,
+        /*.gpu_device           =*/ 0,
+
+        /*.dtw_token_timestamps =*/ false,
+        /*.dtw_aheads_preset    =*/ WHISPER_AHEADS_NONE,
+        /*.dtw_n_top            =*/ -1,
+        /*.dtw_aheads           =*/ {
+            /*.n_heads          =*/ 0,
+            /*.heads            =*/ NULL,
+        },
+        /*.dtw_mem_size         =*/ 1024*1024*128,
     };
     return result;
 }
 
 struct whisper_context * whisper_init_from_file_with_params_no_state(const char * path_model, struct whisper_context_params params) {
     WHISPER_LOG_INFO("%s: loading model from '%s'\n", __func__, path_model);
-
+#ifdef _MSC_VER
+    // Convert UTF-8 path to wide string (UTF-16) for Windows, resolving character encoding issues.
+    std::wstring_convert<std::codecvt_utf8<wchar_t>> converter;
+    std::wstring path_model_wide = converter.from_bytes(path_model);
+    auto fin = std::ifstream(path_model_wide, std::ios::binary);
+#else
     auto fin = std::ifstream(path_model, std::ios::binary);
+#endif
     if (!fin) {
         WHISPER_LOG_ERROR("%s: failed to open '%s'\n", __func__, path_model);
         return nullptr;
@@ -3254,6 +3643,16 @@ struct whisper_context * whisper_init_from_buffer_with_params_no_state(void * bu
 struct whisper_context * whisper_init_with_params_no_state(struct whisper_model_loader * loader, struct whisper_context_params params) {
     ggml_time_init();
 
+    if (params.flash_attn && params.dtw_token_timestamps) {
+        WHISPER_LOG_WARN("%s: dtw_token_timestamps is not supported with flash_attn - disabling\n", __func__);
+        params.dtw_token_timestamps = false;
+    }
+
+    WHISPER_LOG_INFO("%s: use gpu    = %d\n", __func__, params.use_gpu);
+    WHISPER_LOG_INFO("%s: flash attn = %d\n", __func__, params.flash_attn);
+    WHISPER_LOG_INFO("%s: gpu_device = %d\n", __func__, params.gpu_device);
+    WHISPER_LOG_INFO("%s: dtw        = %d\n", __func__, params.dtw_token_timestamps);
+
     whisper_context * ctx = new whisper_context;
     ctx->params = params;
 
@@ -3338,11 +3737,11 @@ struct whisper_context * whisper_init_no_state(struct whisper_model_loader * loa
     return whisper_init_with_params_no_state(loader, whisper_context_default_params());
 }
 
-void whisper_free_state(struct whisper_state * state)
-{
+void whisper_free_state(struct whisper_state * state) {
     if (state) {
-        kv_cache_free(state->kv_self);
-        kv_cache_free(state->kv_cross);
+        whisper_kv_cache_free(state->kv_self);
+        whisper_kv_cache_free(state->kv_cross);
+        whisper_kv_cache_free(state->kv_pad);
 
 #ifdef WHISPER_USE_COREML
         if (state->ctx_coreml != nullptr) {
@@ -3360,12 +3759,17 @@ void whisper_free_state(struct whisper_state * state)
 
         whisper_batch_free(state->batch);
 
-        whisper_allocr_free(state->alloc_conv);
-        whisper_allocr_free(state->alloc_encode);
-        whisper_allocr_free(state->alloc_cross);
-        whisper_allocr_free(state->alloc_decode);
+        ggml_backend_sched_free(state->sched_conv.sched);
+        ggml_backend_sched_free(state->sched_encode.sched);
+        ggml_backend_sched_free(state->sched_cross.sched);
+        ggml_backend_sched_free(state->sched_decode.sched);
 
-        ggml_backend_free(state->backend);
+        for (auto & backend : state->backends) {
+            ggml_backend_free(backend);
+        }
+
+        // [EXPERIMENTAL] Token-level timestamps with DTW
+        aheads_masks_free(state->aheads_masks);
 
         delete state;
     }
@@ -3373,18 +3777,12 @@ void whisper_free_state(struct whisper_state * state)
 
 void whisper_free(struct whisper_context * ctx) {
     if (ctx) {
-        if (ctx->model.ctx) {
-            ggml_free(ctx->model.ctx);
-        }
+        ggml_free(ctx->model.ctx);
 
-        if (ctx->model.buffer) {
-            ggml_backend_buffer_free(ctx->model.buffer);
-        }
+        ggml_backend_buffer_free(ctx->model.buffer);
 
         whisper_free_state(ctx->state);
 
-        ggml_backend_free(ctx->backend);
-
         delete ctx;
     }
 }
@@ -3414,30 +3812,6 @@ int whisper_pcm_to_mel(struct whisper_context * ctx, const float * samples, int
     return whisper_pcm_to_mel_with_state(ctx, ctx->state, samples, n_samples, n_threads);
 }
 
-// same as whisper_pcm_to_mel, but applies a Phase Vocoder to speed up the audio x2 (PV without phase lock is not good)
-int whisper_pcm_to_mel_phase_vocoder_with_state(struct whisper_context * ctx, struct whisper_state * state, const float * samples, int n_samples, int n_threads) {
-    if (!log_mel_spectrogram(*state, samples, n_samples, WHISPER_SAMPLE_RATE, 2 * WHISPER_N_FFT, 2 * WHISPER_HOP_LENGTH, ctx->model.filters.n_mel, n_threads, ctx->model.filters, false, state->mel)) {
-        WHISPER_LOG_ERROR("%s: failed to compute mel spectrogram\n", __func__);
-        return -1;
-    }
-
-    return 0;
-}
-
-// same as whisper_pcm_to_mel, but applies a Phase Vocoder to speed up the audio x2 (PV without phase lock is not good)
-int whisper_pcm_to_mel_phase_vocoder(struct whisper_context * ctx, const float * samples, int n_samples, int n_threads) {
-    return whisper_pcm_to_mel_phase_vocoder_with_state(ctx, ctx->state, samples, n_samples, n_threads);
-}
-
-// same as whisper_pcm_to_mel, but applies WSOLA to speed up the audio x2
-// TODO
-
-// same as whisper_pcm_to_mel, but applies HPTSM to speed up the audio x2
-// TODO
-
-// same as whisper_pcm_to_mel, but applies PV (with phase lock) to speed up the audio x2
-// TODO
-
 int whisper_set_mel_with_state(
         struct whisper_context * ctx,
           struct whisper_state * state,
@@ -3490,7 +3864,7 @@ int whisper_decode_with_state(struct whisper_context * ctx, struct whisper_state
 
     whisper_kv_cache_seq_rm(state->kv_self, 0, n_past, -1);
 
-    if (!whisper_decode_internal(*ctx, *state, state->batch, n_threads, nullptr, nullptr)) {
+    if (!whisper_decode_internal(*ctx, *state, state->batch, n_threads, false, nullptr, nullptr)) {
         WHISPER_LOG_ERROR("%s: failed to eval\n", __func__);
         return 1;
     }
@@ -3512,7 +3886,7 @@ int whisper_tokenize(struct whisper_context * ctx, const char * text, whisper_to
 
     if (n_max_tokens < (int) res.size()) {
         WHISPER_LOG_ERROR("%s: too many resulting tokens: %d (max %d)\n", __func__, (int) res.size(), n_max_tokens);
-        return -1;
+        return -(int) res.size();
     }
 
     for (int i = 0; i < (int) res.size(); i++) {
@@ -3522,7 +3896,11 @@ int whisper_tokenize(struct whisper_context * ctx, const char * text, whisper_to
     return res.size();
 }
 
-int whisper_lang_max_id() {
+int whisper_token_count(struct whisper_context * ctx, const char * text) {
+    return -whisper_tokenize(ctx, text, NULL, 0);
+}
+
+int whisper_lang_max_id(void) {
     auto max_id = 0;
     for (const auto & kv : g_lang) {
         max_id = std::max(max_id, kv.second.first);
@@ -3866,10 +4244,10 @@ const char * whisper_print_system_info(void) {
     s += "SSE3 = "      + std::to_string(ggml_cpu_has_sse3())      + " | ";
     s += "SSSE3 = "     + std::to_string(ggml_cpu_has_ssse3())     + " | ";
     s += "VSX = "       + std::to_string(ggml_cpu_has_vsx())       + " | ";
-    s += "CUDA = "      + std::to_string(ggml_cpu_has_cublas())    + " | ";
+    s += "CUDA = "      + std::to_string(ggml_cpu_has_cuda())      + " | ";
     s += "COREML = "    + std::to_string(whisper_has_coreml())     + " | ";
     s += "OPENVINO = "  + std::to_string(whisper_has_openvino())   + " | ";
-
+    s += "CANN = "      + std::to_string(ggml_cpu_has_cann())             ;
     return s.c_str();
 }
 
@@ -3879,7 +4257,7 @@ const char * whisper_print_system_info(void) {
 
 // Decodes a UTF-8 string which may end in an incomplete sequence. Adds a terminating 0 for use as
 // pointer. If an invalid sequence is encountered, returns `whisper_partial_utf8.n_remain == -1`.
-std::pair<std::vector<uint32_t>, whisper_partial_utf8> decode_utf8(
+static std::pair<std::vector<uint32_t>, whisper_partial_utf8> decode_utf8(
         const char         * src,
         whisper_partial_utf8   partial_start) {
     static const int      lookup[] = { 1, 1, 1, 1, 1, 1, 1, 1, 0, 0, 0, 0, 2, 2, 3, 4 };
@@ -4260,7 +4638,7 @@ static void whisper_suppress_invalid_grammar(
     //if (!allow_eot) {
     //    logits[eot] -= params.grammar_penalty;
     //}
-    //Rprintf("Allowed: (%zu tokens)\n", size - rejects.size());
+    //fprintf(stderr, "Allowed: (%zu tokens)\n", size - rejects.size());
 }
 
 static void whisper_grammar_accept_token(whisper_context & ctx, whisper_grammar & grammar, whisper_token token) {
@@ -4268,15 +4646,15 @@ static void whisper_grammar_accept_token(whisper_context & ctx, whisper_grammar
         return;
     }
 
-    //Rprintf("Accept: '%s'\n", ctx.vocab.id_to_token[token].c_str());
+    //fprintf(stderr, "Accept: '%s'\n", ctx.vocab.id_to_token[token].c_str());
 
     const std::string & text = ctx.vocab.id_to_token[token];
 
     if (text.rfind("[_", 0) == 0) {
-        // Rprintf(" (skipped)\n");
+        // fprintf(stderr, " (skipped)\n");
         return;
     }
-    // Rprintf("\n");
+    // fprintf(stderr, "\n");
 
     // Note terminating 0 in decoded string
     const auto   decoded     = decode_utf8(text.c_str(), grammar.partial_utf8);
@@ -4293,7 +4671,7 @@ static void whisper_grammar_accept_token(whisper_context & ctx, whisper_grammar
 
 ////////////////////////////////////////////////////////////////////////////
 
-struct whisper_context_params * whisper_context_default_params_by_ref() {
+struct whisper_context_params * whisper_context_default_params_by_ref(void) {
     struct whisper_context_params params = whisper_context_default_params();
 
     struct whisper_context_params* result = new whisper_context_params();
@@ -4334,12 +4712,13 @@ struct whisper_full_params whisper_full_default_params(enum whisper_sampling_str
         /*.split_on_word     =*/ false,
         /*.max_tokens        =*/ 0,
 
-        /*.speed_up          =*/ false,
         /*.debug_mode        =*/ false,
         /*.audio_ctx         =*/ 0,
 
         /*.tdrz_enable       =*/ false,
 
+        /* suppress_regex    =*/ nullptr,
+
         /*.initial_prompt    =*/ nullptr,
         /*.prompt_tokens     =*/ nullptr,
         /*.prompt_n_tokens   =*/ 0,
@@ -4425,6 +4804,17 @@ static inline bool should_split_on_word(const char * txt, bool split_on_word) {
     return txt[0] == ' ';
 }
 
+static void whisper_exp_compute_token_level_timestamps_dtw(
+            struct whisper_context * ctx,
+              struct whisper_state * state,
+        struct whisper_full_params   params,
+                               int   i_segment,
+                            size_t   n_segments,
+                               int   seek,
+                               int   n_frames,
+                               int   medfilt_width,
+                               int   n_threads);
+
 // wrap the last segment to max_len characters
 // returns the number of new segments
 static int whisper_wrap_segment(struct whisper_context & ctx, struct whisper_state & state, int max_len, bool split_on_word) {
@@ -4572,6 +4962,17 @@ static void whisper_process_logits(
             params.logits_filter_callback(&ctx, &state, tokens_cur.data(), tokens_cur.size(), logits.data(), params.logits_filter_callback_user_data);
         }
 
+        // suppress any tokens matching a regular expression
+        // ref: https://github.com/openai/whisper/discussions/1041
+        if (params.suppress_regex != nullptr) {
+            std::regex re(params.suppress_regex);
+            for (std::pair<whisper_vocab::token, whisper_vocab::id> token_id : vocab.token_to_id) {
+                if (std::regex_match(token_id.first, re)) {
+                    logits[token_id.second] = -INFINITY;
+                }
+            }
+        }
+
         // suppress non-speech tokens
         // ref: https://github.com/openai/whisper/blob/7858aa9c08d98f75575035ecd6481f462d66ca27/whisper/tokenizer.py#L224-L253
         if (params.suppress_non_speech_tokens) {
@@ -4728,7 +5129,7 @@ static void whisper_process_logits(
     //    const auto prob    = probs[i];
     //    const auto logit   = logits[i];
     //    const auto logprob = logprobs[i];
-    //    Rprintf("%16s : prob=%9.5f logit=%9.5f logprob=%9.5f\n", token.c_str(), prob, logit, logprob);
+    //    printf("%16s : prob=%9.5f logit=%9.5f logprob=%9.5f\n", token.c_str(), prob, logit, logprob);
     //}
 
     // print sorted
@@ -4748,10 +5149,10 @@ static void whisper_process_logits(
             const auto prob    = pairs[i].first;
             const auto logit   = logits[pairs[i].second];
             const auto logprob = logprobs[pairs[i].second];
-            Rprintf("%16s : id=%6d prob=%9.5f logit=%9.5f logprob=%9.5f '%s'\n", token.c_str(), pairs[i].second, prob, logit, logprob, token.c_str());
+            printf("%16s : id=%6d prob=%9.5f logit=%9.5f logprob=%9.5f '%s'\n", token.c_str(), pairs[i].second, prob, logit, logprob, token.c_str());
         }
 
-        Rprintf("----------------\n");
+        printf("----------------\n");
     }
 
     // "And", "and", " And", " and"
@@ -4775,12 +5176,25 @@ static void whisper_process_logits(
 #endif
 }
 
+static bool whisper_sequence_tokens_equal(const whisper_sequence & a, const whisper_sequence & b) {
+    if (a.tokens.size() != b.tokens.size()) {
+        return false;
+    }
+    // sequences are more likely to diverge at the end
+    for (int i = a.tokens.size() - 1; i >= 0; i--) {
+        if (a.tokens[i].id != b.tokens[i].id) {
+            return false;
+        }
+    }
+    return true;
+}
+
 static whisper_token_data whisper_sample_token(
             whisper_context & ctx,
       const whisper_decoder & decoder,
                        bool   best) {
     whisper_token_data result = {
-        0, 0, 0.0f, 0.0f, 0.0f, 0.0f, -1, -1, 0.0f,
+        0, 0, 0.0f, 0.0f, 0.0f, 0.0f, -1, -1, -1, 0.0f,
     };
 
     const auto & vocab = ctx.vocab;
@@ -4898,7 +5312,7 @@ static std::vector<whisper_token_data> whisper_sample_token_topk(
         const auto id = dist(decoder.rng);
         //printf("XXX %d %d %f %f %f %f\n", id, tid, probs[id], logprobs[id], pt, ptsum);
 
-        result.push_back({ id, tid, probs[id], logprobs[id], pt, ptsum, -1, -1, 0.0f, });
+        result.push_back({ id, tid, probs[id], logprobs[id], pt, ptsum, -1, -1, -1, 0.0f, });
 
         if (result[i].id >= vocab.token_beg) {
             result[i].tid = result[i].id;
@@ -4971,15 +5385,9 @@ int whisper_full_with_state(
 
     if (n_samples > 0) {
         // compute log mel spectrogram
-        if (params.speed_up) {
-            // TODO: Replace PV with more advanced algorithm
+        if (whisper_pcm_to_mel_with_state(ctx, state, samples, n_samples, params.n_threads) != 0) {
             WHISPER_LOG_ERROR("%s: failed to compute log mel spectrogram\n", __func__);
-            return -1;
-        } else {
-            if (whisper_pcm_to_mel_with_state(ctx, state, samples, n_samples, params.n_threads) != 0) {
-                WHISPER_LOG_ERROR("%s: failed to compute log mel spectrogram\n", __func__);
-                return -2;
-            }
+            return -2;
         }
     }
 
@@ -5016,8 +5424,8 @@ int whisper_full_with_state(
     // if length of spectrogram is less than 1.0s (100 frames), then return
     // basically don't process anything that is less than 1.0s
     // see issue #39: https://github.com/ggerganov/whisper.cpp/issues/39
-    if (seek_end < seek_start + (params.speed_up ? 50 : 100)) {
-        WHISPER_LOG_DEBUG("%s: input is too short - %d ms < 1000 ms\n", __func__, (seek_end - seek_start)*10);
+    if (seek_end < seek_start + 100) {
+        WHISPER_LOG_WARN("%s: input is too short - %d ms < 1000 ms. consider padding the input audio with silence\n", __func__, (seek_end - seek_start)*10);
         return 0;
     }
 
@@ -5080,7 +5488,12 @@ int whisper_full_with_state(
         // initial prompt
         if (!params.prompt_tokens && params.initial_prompt) {
             prompt_tokens.resize(1024);
-            prompt_tokens.resize(whisper_tokenize(ctx, params.initial_prompt, prompt_tokens.data(), prompt_tokens.size()));
+            int n_needed = whisper_tokenize(ctx, params.initial_prompt, prompt_tokens.data(), prompt_tokens.size());
+            if (n_needed < 0) {
+                prompt_tokens.resize(-n_needed);
+                n_needed = whisper_tokenize(ctx, params.initial_prompt, prompt_tokens.data(), prompt_tokens.size());
+            }
+            prompt_tokens.resize(n_needed);
             params.prompt_tokens   = prompt_tokens.data();
             params.prompt_n_tokens = prompt_tokens.size();
         }
@@ -5116,11 +5529,11 @@ int whisper_full_with_state(
         }
     }
 
-    // distilled models require the "no_timestamps" token
+    // first release distilled models require the "no_timestamps" token
     {
-        const bool is_distil = ctx->model.hparams.n_text_layer == 2;
+        const bool is_distil = ctx->model.hparams.n_text_layer == 2 && ctx->model.hparams.n_vocab != 51866;
         if (is_distil && !params.no_timestamps) {
-            WHISPER_LOG_WARN("%s: using distilled model - forcing no_timestamps\n", __func__);
+            WHISPER_LOG_WARN("%s: using first release distilled models - forcing no_timestamps\n", __func__);
             params.no_timestamps = true;
         }
     }
@@ -5256,13 +5669,34 @@ int whisper_full_with_state(
                 }
                 WHISPER_LOG_DEBUG("\n\n");
 
+                // recreate the KV cache if the number of decoders has changed
+                if (state->kv_self_n_dec < n_decoders_cur) {
+                    WHISPER_LOG_DEBUG("%s: recreating KV cache: n_decoders_cur = %d\n", __func__, n_decoders_cur);
+
+                    whisper_kv_cache_free(state->kv_self);
+
+                    // overallocate to workaround KV cache fragmentation issues
+                    const int factor = n_decoders_cur > 1 ? n_decoders_cur + 2 : 1;
+
+                    if (!whisper_kv_cache_init(state->kv_self, state->backends[0], ctx->itype,
+                                ctx->model.hparams.n_text_state,
+                                ctx->model.hparams.n_text_layer,
+                                GGML_PAD(ctx->model.hparams.n_text_ctx, 256)*factor)) {
+                        WHISPER_LOG_ERROR("%s: whisper_kv_cache_init() failed for self-attention cache\n", __func__);
+                        whisper_free_state(state);
+                        return -7;
+                    }
+
+                    state->kv_self_n_dec = n_decoders_cur;
+                }
+
                 whisper_kv_cache_clear(state->kv_self);
 
                 whisper_batch_prep_legacy(state->batch, prompt.data(), prompt.size(), 0, 0);
 
-                if (!whisper_decode_internal(*ctx, *state, state->batch, params.n_threads, params.abort_callback, params.abort_callback_user_data)) {
+                if (!whisper_decode_internal(*ctx, *state, state->batch, params.n_threads, false, params.abort_callback, params.abort_callback_user_data)) {
                     WHISPER_LOG_ERROR("%s: failed to decode\n", __func__);
-                    return -7;
+                    return -8;
                 }
 
                 {
@@ -5373,7 +5807,10 @@ int whisper_full_with_state(
                             beam_candidates.begin(),
                             beam_candidates.end(),
                             [](const beam_candidate & a, const beam_candidate & b) {
-                        return a.sequence.sum_logprobs_all > b.sequence.sum_logprobs_all;
+                        if (a.sequence.sum_logprobs_all != b.sequence.sum_logprobs_all) {
+                            return a.sequence.sum_logprobs_all > b.sequence.sum_logprobs_all;
+                        }
+                        return a.decoder_idx < b.decoder_idx;
                     });
 
                     uint32_t cur_c = 0;
@@ -5391,7 +5828,7 @@ int whisper_full_with_state(
 
                         auto & cur = beam_candidates[cur_c++];
 
-                        while (beam_candidates.size() > cur_c && beam_candidates[cur_c].sequence.sum_logprobs_all == cur.sequence.sum_logprobs_all && i > 0) {
+                        while (beam_candidates.size() > cur_c && whisper_sequence_tokens_equal(beam_candidates[cur_c].sequence, cur.sequence) && i > 0) {
                             ++cur_c;
                         }
 
@@ -5470,7 +5907,7 @@ int whisper_full_with_state(
                            (params.max_tokens > 0 && i >= params.max_tokens) || // max tokens per segment reached
                            (has_ts && seek + seek_delta + 100 >= seek_end)      // end of audio reached
                            ) {
-                            if (result_len == 0) {
+                            if (result_len == 0 && !params.no_timestamps) {
                                 if (seek + seek_delta + 100 >= seek_end) {
                                     result_len = i + 1;
                                 } else {
@@ -5480,7 +5917,7 @@ int whisper_full_with_state(
                                 }
                             }
 
-                            if (params.single_segment) {
+                            if (params.single_segment || params.no_timestamps) {
                                 result_len = i + 1;
                                 seek_delta = 100*WHISPER_CHUNK_SIZE;
                             }
@@ -5557,9 +5994,9 @@ int whisper_full_with_state(
 
                     assert(batch.n_tokens > 0);
 
-                    if (!whisper_decode_internal(*ctx, *state, state->batch, params.n_threads, params.abort_callback, params.abort_callback_user_data)) {
+                    if (!whisper_decode_internal(*ctx, *state, state->batch, params.n_threads, false, params.abort_callback, params.abort_callback_user_data)) {
                         WHISPER_LOG_ERROR("%s: failed to decode\n", __func__);
-                        return -8;
+                        return -9;
                     }
 
                     const int64_t t_start_sample_us = ggml_time_us();
@@ -5680,6 +6117,9 @@ int whisper_full_with_state(
 
             const auto & tokens_cur = best_decoder.sequence.tokens;
 
+            // [EXPERIMENTAL] Token-level timestamps with DTW
+            const auto n_segments_before = state->result_all.size();
+
             //WHISPER_LOG_DEBUG("prompt_init.size() = %d, prompt.size() = %d, result_len = %d, seek_delta = %d\n", prompt_init.size(), prompt.size(), result_len, seek_delta);
 
             // update prompt_past
@@ -5717,15 +6157,15 @@ int whisper_full_with_state(
                         const auto t1 = seek + 2*(tokens_cur[i].tid - whisper_token_beg(ctx));
 
                         if (!text.empty()) {
-                            const auto tt0 = params.speed_up ? 2*t0 : t0;
-                            const auto tt1 = params.speed_up ? 2*t1 : t1;
+                            const auto tt0 = t0;
+                            const auto tt1 = t1;
 
                             if (params.print_realtime) {
                                 if (params.print_timestamps) {
-                                    Rprintf("[%s --> %s]  %s\n", to_timestamp(tt0).c_str(), to_timestamp(tt1).c_str(), text.c_str());
+                                    printf("[%s --> %s]  %s\n", to_timestamp(tt0).c_str(), to_timestamp(tt1).c_str(), text.c_str());
                                 } else {
-                                    Rprintf("%s", text.c_str());
-                                    Rcpp::checkUserInterrupt();
+                                    printf("%s", text.c_str());
+                                    fflush(stdout);
                                 }
                             }
 
@@ -5764,15 +6204,15 @@ int whisper_full_with_state(
                 if (!text.empty()) {
                     const auto t1 = seek + seek_delta;
 
-                    const auto tt0 = params.speed_up ? 2*t0 : t0;
-                    const auto tt1 = params.speed_up ? 2*t1 : t1;
+                    const auto tt0 = t0;
+                    const auto tt1 = t1;
 
                     if (params.print_realtime) {
                         if (params.print_timestamps) {
-                            Rprintf("[%s --> %s]  %s\n", to_timestamp(tt0).c_str(), to_timestamp(tt1).c_str(), text.c_str());
+                            printf("[%s --> %s]  %s\n", to_timestamp(tt0).c_str(), to_timestamp(tt1).c_str(), text.c_str());
                         } else {
-                            Rprintf("%s", text.c_str());
-                            Rcpp::checkUserInterrupt();
+                            printf("%s", text.c_str());
+                            fflush(stdout);
                         }
                     }
 
@@ -5797,6 +6237,17 @@ int whisper_full_with_state(
                 }
             }
 
+            // FIXME: will timestamp offsets be correct?
+            // [EXPERIMENTAL] Token-level timestamps with DTW
+            {
+                const auto n_segments = state->result_all.size() - n_segments_before;
+                if (ctx->params.dtw_token_timestamps && n_segments) {
+                    const int n_frames = std::min(std::min(WHISPER_CHUNK_SIZE * 100, seek_delta), seek_end - seek);
+                    whisper_exp_compute_token_level_timestamps_dtw(
+                            ctx, state, params, result_all.size() - n_segments, n_segments, seek, n_frames, 7, params.n_threads);
+                }
+            }
+
             // update audio window
             seek += seek_delta;
 
@@ -6026,7 +6477,7 @@ float whisper_full_get_token_p(struct whisper_context * ctx, int i_segment, int
 //
 
 WHISPER_API int whisper_bench_memcpy(int n_threads) {
-    Rcpp::Rcout << whisper_bench_memcpy_str(n_threads);
+    fputs(whisper_bench_memcpy_str(n_threads), stderr);
     return 0;
 }
 
@@ -6065,7 +6516,7 @@ WHISPER_API const char * whisper_bench_memcpy_str(int n_threads) {
 
             tsum += (t1 - t0)*1e-6;
 
-            src[((int) floor(R::runif(0, 32767))) % size] = ((int) floor(R::runif(0, 32767))) % 256;
+            src[rand() % size] = rand() % 256;
         }
 
         snprintf(strbuf, sizeof(strbuf), "memcpy: %7.2f GB/s (heat-up)\n", (double) (n*size)/(tsum*1e9));
@@ -6100,7 +6551,7 @@ WHISPER_API const char * whisper_bench_memcpy_str(int n_threads) {
 
             tsum += (t1 - t0)*1e-6;
 
-            src[((int) floor(R::runif(0, 32767))) % size] = ((int) floor(R::runif(0, 32767))) % 256;
+            src[rand() % size] = rand() % 256;
         }
 
         snprintf(strbuf, sizeof(strbuf), "memcpy: %7.2f GB/s ( 1 thread)\n", (double) (n*size)/(tsum*1e9));
@@ -6134,7 +6585,7 @@ WHISPER_API const char * whisper_bench_memcpy_str(int n_threads) {
             for (size_t i = 0; i < n; i++) {
                 memcpy(dst + i0, src + i0, i1 - i0);
 
-                src[i0 + ((int) floor(R::runif(0, 32767))) % (i1 - i0)] = ((int) floor(R::runif(0, 32767))) % 256;
+                src[i0 + rand() % (i1 - i0)] = rand() % 256;
             };
         };
 
@@ -6174,7 +6625,7 @@ WHISPER_API const char * whisper_bench_memcpy_str(int n_threads) {
 }
 
 WHISPER_API int whisper_bench_ggml_mul_mat(int n_threads) {
-    Rcpp::Rcout << whisper_bench_ggml_mul_mat_str(n_threads);
+    fputs(whisper_bench_ggml_mul_mat_str(n_threads), stderr);
     return 0;
 }
 
@@ -6590,7 +7041,7 @@ static void whisper_exp_compute_token_level_timestamps(
     //for (int j = 0; j < n; ++j) {
     //    const auto & token = tokens[j];
     //    const auto tt = token.pt > thold_pt && token.ptsum > 0.01 ? whisper_token_to_str(&ctx, token.tid) : "[?]";
-    //    Rprintf("%s: %10s %6.3f %6.3f %6.3f %6.3f %5d %5d '%s'\n", __func__,
+    //    printf("%s: %10s %6.3f %6.3f %6.3f %6.3f %5d %5d '%s'\n", __func__,
     //            tt, token.p, token.pt, token.ptsum, token.vlen, (int) token.t0, (int) token.t1, whisper_token_to_str(&ctx, token.id));
 
     //    if (tokens[j].id >= whisper_token_eot(&ctx)) {
@@ -6599,6 +7050,322 @@ static void whisper_exp_compute_token_level_timestamps(
     //}
 }
 
+//
+// token level timestamps - dtw version
+//
+
+// n_text_layer -> total text layers on model
+// n_head -> total heads per text layer on model
+static std::vector<uint32_t> get_alignment_heads_by_layer(const whisper_context_params & cparams, int il, int n_text_layer, int n_head) {
+    std::vector<uint32_t> ret;
+    if (cparams.dtw_aheads_preset == WHISPER_AHEADS_NONE) {
+        return ret;
+    } else if (cparams.dtw_aheads_preset == WHISPER_AHEADS_N_TOP_MOST) {
+        if (il >= n_text_layer - cparams.dtw_n_top) {
+            for (int32_t i = 0; i < n_head; ++i) {
+                ret.push_back(i);
+            }
+        }
+    } else {
+        const auto aheads = cparams.dtw_aheads_preset == WHISPER_AHEADS_CUSTOM ? cparams.dtw_aheads : g_aheads.at(cparams.dtw_aheads_preset);
+        for (size_t i = 0; i < aheads.n_heads; ++i) {
+            if (aheads.heads[i].n_text_layer == il) {
+                ret.push_back(aheads.heads[i].n_head);
+            }
+        }
+    }
+    return ret;
+}
+
+// dtw + backtrace to return found path
+// based on
+// https://github.com/openai/whisper/blob/main/whisper/timing.py#L83
+static ggml_tensor * dtw_and_backtrace(ggml_context * ctx, ggml_tensor * x) {
+    WHISPER_ASSERT(ggml_n_dims(x) == 2);
+
+    int64_t N = x->ne[0];
+    int64_t M = x->ne[1];
+    struct ggml_tensor * cost = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, N + 1, M + 1);
+    struct ggml_tensor * trace = ggml_new_tensor_2d(ctx, GGML_TYPE_I32, N + 1, M + 1);
+
+    cost = ggml_set_f32(cost, INFINITY);
+    trace = ggml_set_f32(trace, -1);
+    ggml_set_f32_nd(cost, 0, 0, 0, 0, 0.0);
+
+    // dtw
+    // supposedly can be optmized by computing diagonals in parallel ?
+    // Not sure it is worth it since x will be GENERATED_TOKENS*1500 size at most.
+    for (int64_t j = 1; j < M + 1; ++j) {
+        for (int64_t i = 1; i < N + 1; ++i) {
+            float c0 = ggml_get_f32_nd(cost, i - 1, j - 1, 0, 0);
+            float c1 = ggml_get_f32_nd(cost, i - 1, j, 0, 0);
+            float c2 = ggml_get_f32_nd(cost, i, j - 1, 0, 0);
+
+            float c;
+            int32_t t;
+            if (c0 < c1 && c0 < c2) {
+                c = c0;
+                t = 0;
+            } else if (c1 < c0 && c1 < c2) {
+                c = c1;
+                t = 1;
+            } else {
+                c = c2;
+                t = 2;
+            }
+
+            c = ggml_get_f32_nd(x, i - 1, j - 1, 0, 0) + c;
+            ggml_set_f32_nd(cost, i, j, 0, 0, c);
+            ggml_set_i32_nd(trace, i, j, 0, 0, t);
+        }
+    }
+
+    // Backtrace
+    const int64_t BT_MAX_ROWS = N + M - 1;
+    struct ggml_tensor * bt = ggml_new_tensor_2d(ctx, GGML_TYPE_I32, BT_MAX_ROWS, 2);
+    // trace[0, :] = 2;
+    for (int64_t i = 0; i < M + 1; ++i)
+        ggml_set_i32_nd(trace, 0, i, 0, 0, 2);
+    //trace[:, 0] = 1;
+    for (int64_t i = 0; i < N + 1; ++i)
+        ggml_set_i32_nd(trace, i, 0, 0, 0, 1);
+    int bt_row_idx = BT_MAX_ROWS - 1;
+    int64_t i = N;
+    int64_t j = M;
+    while (i > 0 || j > 0) {
+        ggml_set_i32_nd(bt, bt_row_idx, 0, 0, 0, i - 1);
+        ggml_set_i32_nd(bt, bt_row_idx, 1, 0, 0, j - 1);
+        --bt_row_idx;
+
+        int32_t t = ggml_get_i32_nd(trace, i, j, 0, 0);
+        if (t == 0) {
+            --i;
+            --j;
+        } else if (t == 1) {
+            --i;
+        } else if (t == 2) {
+            --j;
+        } else {
+            WHISPER_ASSERT(0);
+        }
+    }
+
+    // FIXME: manual clip/transpose might not be the most efficient way? (e.g. use ggml funcs)
+    // Clip + transpose
+    // This might not be entirely necessary for our case, but leaving it for now so output matrix
+    // is identical to dtw on openAI timing.py
+    const int64_t result_n_cols = BT_MAX_ROWS-bt_row_idx-1;
+    ggml_tensor * r = ggml_new_tensor_2d(ctx, GGML_TYPE_I32, 2, result_n_cols);
+    for (int64_t i = 0; i < 2; ++i) {
+        for (int64_t j = 0; j < result_n_cols; ++j) {
+            int32_t v = ggml_get_i32_nd(bt, j+bt_row_idx+1, i, 0, 0);
+            ggml_set_i32_nd(r, i, j, 0, 0, v);
+        }
+    }
+
+    return r;
+}
+
+struct median_filter_user_data {
+    int filter_width;
+};
+
+static void median_filter(struct ggml_tensor * dst , const struct ggml_tensor * a, int ith, int /*nth*/, void * userdata) {
+    if (ith != 0) {
+        return;
+    }
+    int filter_width = ((median_filter_user_data *) userdata)->filter_width;
+    WHISPER_ASSERT(filter_width < a->ne[2]);
+    WHISPER_ASSERT(filter_width % 2);
+    WHISPER_ASSERT(ggml_n_dims(a) == 3);
+    WHISPER_ASSERT(a->type == GGML_TYPE_F32);
+
+    std::vector<float> filter;
+    filter.reserve(filter_width);
+    for (int64_t i = 0; i < a->ne[0]; ++i) {
+        for (int64_t j = 0; j < a->ne[1]; ++j) {
+            for (int64_t k = 0; k < a->ne[2]; ++k) {
+                for (int64_t off = -filter_width/2; off <= filter_width/2; ++off) {
+                    // "reflect" padding
+                    int64_t idx = k + off;
+                    if (idx < 0) {
+                        idx = -idx;
+                    } else if (idx >= a->ne[2]) {
+                        idx = 2*(a->ne[2] - 1) - idx;
+                    }
+
+                    filter.push_back(ggml_get_f32_nd(a, i, j, idx, 0));
+                }
+                std::sort(filter.begin(), filter.end());
+                const float v = filter[filter.size()/2];
+                ggml_set_f32_nd(dst, i, j, k, 0, v);
+                filter.clear();
+            }
+        }
+    }
+}
+
+static void whisper_exp_compute_token_level_timestamps_dtw(
+            struct whisper_context * ctx,
+              struct whisper_state * state,
+        struct whisper_full_params   params,
+                               int   i_segment,
+                            size_t   n_segments,
+                               int   seek,
+                               int   n_frames,
+                               int   medfilt_width,
+                               int   n_threads)
+{
+    const int n_audio_ctx = state->exp_n_audio_ctx > 0 ? state->exp_n_audio_ctx : ctx->model.hparams.n_audio_ctx;
+    WHISPER_ASSERT(medfilt_width % 2);
+    WHISPER_ASSERT(n_frames <= n_audio_ctx * 2);
+    WHISPER_ASSERT(ctx->params.dtw_aheads_preset != WHISPER_AHEADS_NONE);
+
+    // FIXME: Allocating mem everytime we call this func
+    // Our ggml buffer should be pre-allocated somewhere during init and reused
+    // when we call this function
+    struct ggml_init_params gparams = {
+        /*.mem_size   =*/ ctx->params.dtw_mem_size,
+        /*.mem_buffer =*/ NULL,
+        /*.no_alloc   =*/ false,
+    };
+    struct ggml_context * gctx = ggml_init(gparams);
+
+    // Build token sequence that will be passed to decoder
+    // sot + [lang] + text result + eot
+    std::vector<whisper_token> tokens = { whisper_token_sot(ctx), };
+    if (whisper_is_multilingual(ctx)) {
+        const int lang_id = whisper_lang_id(params.language);
+        state->lang_id = lang_id;
+        tokens.push_back(whisper_token_lang(ctx, lang_id));
+    }
+    const size_t sot_sequence_length = tokens.size();
+    tokens.push_back(whisper_token_not(ctx));
+    for (size_t i = i_segment; i < i_segment + n_segments; ++i) {
+        auto & segment = state->result_all[i];
+        for (auto &t: segment.tokens) {
+            // Only text tokens
+            if (t.id < whisper_token_eot(ctx)) {
+                tokens.push_back(t.id);
+            }
+        }
+    }
+    tokens.push_back(whisper_token_eot(ctx));
+
+    // Get result tokens, pass then along to decoder to get cross attention QKs
+    // used in timestamping
+    // Decoder already returns only alignment head QKs, already concatenated in
+    // one tensor.
+    whisper_kv_cache_clear(state->kv_self);
+    whisper_batch_prep_legacy(state->batch, tokens.data(), tokens.size(), 0, 0);
+    whisper_kv_cache_seq_rm(state->kv_self, 0, 0, -1);
+    if (!whisper_decode_internal(*ctx, *state, state->batch, n_threads, true, nullptr, nullptr)) {
+        WHISPER_LOG_INFO("DECODER FAILED\n");
+        WHISPER_ASSERT(0);
+    }
+    WHISPER_ASSERT(state->aheads_cross_QKs != nullptr);
+
+    const auto n_audio_tokens = n_frames/2;
+    WHISPER_ASSERT(state->aheads_cross_QKs != NULL);
+    WHISPER_ASSERT(n_audio_tokens <= state->aheads_cross_QKs->ne[1]);
+    const auto n_tokens = state->aheads_cross_QKs->ne[0];
+    const auto n_heads = state->aheads_cross_QKs->ne[2];
+
+    // Copy data from decoder buffer to a local CPU tensor, discarding unused audio
+    // tokens (i.e. discarding rows at the end of tensor)
+    // IN: Tensor with N_TOKENS*audio_ctx*N_ALIGNMENT_HEADS dims
+    // OUT: Tensor with N_TOKENS*N_AUDIO_TOKENS*N_ALIGNMENT_HEADS dims
+    WHISPER_ASSERT(state->aheads_cross_QKs->type == GGML_TYPE_F32);
+    WHISPER_ASSERT(ggml_is_contiguous(state->aheads_cross_QKs));
+    ggml_tensor * w = ggml_new_tensor_3d(gctx, GGML_TYPE_F32, n_tokens, n_audio_tokens, n_heads);
+    auto & data = state->aheads_cross_QKs_data;
+    data.resize(n_tokens * n_audio_ctx * n_heads);
+    ggml_backend_tensor_get(state->aheads_cross_QKs, data.data(), 0, sizeof(float) * n_tokens * n_audio_ctx * n_heads);
+    for (int k = 0; k < n_heads; ++k) {
+        for (int j = 0; j < n_audio_tokens; ++j) {
+            memcpy(
+                (char *) w->data + j * w->nb[1] + k * w->nb[2],
+                data.data() + j * n_tokens + k * n_tokens * n_audio_ctx,
+                n_tokens * sizeof(float)
+            );
+        }
+    }
+
+    // Normalize - in original OpenAI code, this is done over dim=-2. In this case,
+    // we already permuted N_TOKENS dimension to columns on last loop, becase ggml_norm
+    // operates over columns. Afterwards, permute to a shape that facilitates mean
+    // operation (after median filter)
+    // IN: Tensor with N_TOKENS*N_AUDIO_TOKENS*N_ALIGNMENT_HEADS dims
+    // OUT: Tensor with N_ALIGNMENT_HEADS*N_TOKENS*N_AUDIO_TOKENS dims
+    w = ggml_norm(gctx, w, 1e-9f);
+    w = ggml_permute(gctx, ggml_permute(gctx, w, 2, 1, 0 ,3), 0, 2, 1, 3);
+
+    // Pass median filter - this is done over AUDIO_TOKENS dimension.
+    // IN: Tensor with N_ALIGNMENT_HEADS*N_TOKENS*N_AUDIO_TOKENS dims
+    // OUT: Same dims
+    median_filter_user_data mf_user_data = {medfilt_width};
+    w = ggml_map_custom1(gctx, w, median_filter, 1, &mf_user_data);
+
+    // Take mean over columns, scale by -1, reshape to 2D tensor, remove SOT sequence and EOT
+    // IN: Tensor with N_ALIGNMENT_HEADS*N_TOKENS*N_AUDIO_TOKENS dims
+    // OUT: Tensor with N_TOKENS*N_AUDIO_TOKENS dims
+    w = ggml_mean(gctx, w);
+    w = ggml_scale(gctx, w, -1.0);
+    w = ggml_reshape_2d(gctx, w, w->ne[1], w->ne[2]);
+
+    // Remove SOT sequence and EOT
+    // Out dimension is (N_TOKENS-sot_sequence_length-1)*N_AUDIO_TOKENS
+    w = ggml_view_2d(gctx, w, w->ne[0] - sot_sequence_length - 1, w->ne[1], w->nb[1], sot_sequence_length * w->nb[0]);
+
+    // Compute
+    struct ggml_cgraph * gf = ggml_new_graph(gctx);
+    ggml_build_forward_expand(gf, w);
+    ggml_graph_compute_with_ctx(gctx, gf, n_threads);
+
+    ggml_tensor * alignment = dtw_and_backtrace(gctx, w);
+
+    // Place timestamps on segments
+    int32_t last_v = 0;
+    auto seg_i = state->result_all.begin() + i_segment;
+    auto tok_i = seg_i->tokens.begin();
+    for (int i = 0; i < alignment->ne[1]; ++i) {
+        int32_t v = ggml_get_i32_nd(alignment, 0, i, 0, 0);
+        if (v != last_v) {
+            int32_t time_index = ggml_get_i32_nd(alignment, 1, i, 0, 0);
+            int64_t timestamp = (time_index * 2) + seek; // Each index on DTW result = 20mS audio
+            last_v = v;
+
+            // Skip non-text tokens
+            while (!(tok_i->id < whisper_token_eot(ctx))) {
+                ++tok_i;
+                if (tok_i == seg_i->tokens.end()) {
+                    ++seg_i;
+                    tok_i = seg_i->tokens.begin();
+                }
+            }
+
+            tok_i->t_dtw = timestamp;
+            ++tok_i;
+            if (tok_i == seg_i->tokens.end()) {
+                ++seg_i;
+                tok_i = seg_i->tokens.begin();
+            }
+        }
+    }
+
+    // Print DTW timestamps
+    /*for (size_t i = i_segment; i < i_segment + n_segments; ++i) {
+        auto & segment = state->result_all[i];
+        for (auto &t: segment.tokens) {
+            const char * tok = whisper_token_to_str(ctx, t.id);
+            fprintf(stderr, "|%s|(%.2f) ", tok, (float)t.t_dtw/100);
+        }
+        fprintf(stderr, "\n");
+    }*/
+
+    ggml_free(gctx);
+}
+
 void whisper_log_set(ggml_log_callback log_callback, void * user_data) {
     g_state.log_callback = log_callback ? log_callback : whisper_log_callback_default;
     g_state.log_callback_user_data = user_data;
@@ -6625,6 +7392,6 @@ static void whisper_log_internal(ggml_log_level level, const char * format, ...)
 static void whisper_log_callback_default(ggml_log_level level, const char * text, void * user_data) {
     (void) level;
     (void) user_data;
-    Rcpp::Rcout << text;
-    Rcpp::checkUserInterrupt();
+    fputs(text, stderr);
+    fflush(stderr);
 }